Edwards

Graduate Texts in Mathematics
64
Editorial Board
F. W. Gehring
P.R. Halmos
Managing Editor
c.c. Moore
R.E. Edwards
Fourier Series
A Modem Introduction
Volume 1
Second Edition
Springer-Verlag
;fC f ~ ..- f. }f~ .",. iJ
R.E. Edwards
Institute for Advanced Studies
The Australian National University
Canberra, A.C.T. 2600
Australia
Editorial Board
P.R. Halmos F. W. Gehring c.c. Moore

MalUlging Editor University of Michigan University of California
Indiana University Department of Mathematics at Berkeley
Department of Mathematics Ann Arbor, Michigan 48104 Department of Mathematics
Bloomington, Indiana 47401 USA Berkeley, California 94720
USA USA
AMS Subject Classification: 42-01
Library of Congress Cataloging in Publication Data
Edwards, Robert E
Fourier series, a modem introduction.
(Graduate texts in mathematics; 64-

Bibliography: v. I, p.
Includes index.
I. Fourier series I. Title. II. Series .
QA404.E25 1979 515'.2433 79-11932
ISBN-13 : 978-1-4612-6210-7 e-ISBN-13 : 978-1-4612-6208-4

DOl: 10.1007/978-1-4612-6208-4
The fU"St edition was published in 1967 by Holt, Rinehart and Winston, Inc.
All rights reserved.
No part of this book may be translated or reproduced

in any fonn without written pennission from Springer-Verlag.
1967, 1979 by Springer-Verlag New York, Inc.
This reprint has been authorized by Springer-Verlag (8erlinfHeidelberg/New York) for sale in
the People's Republic of China only and not for export therefrom.
Reprinted in China by Beijing World Publishing Corporation, 2003
PREFACE TO REVISED EDITION
(OF VOLUMED
There are a multitude of minor corrections. In addition there are a

few substantial changes in and supplements to the exposition, including some
proofs. (Such is the case, for example, with Sections 7.1-7.3.)
Professor Katznelson's book [Kz] is recommended as a companion text.
Many references to Mathematical Review8 have been inserted. None of these
is essential to an understanding of the main text, and all may be ignored on a
first reading. There is an already very large and rapidly increasing literature,
and a preliminary glance at a review (often more rapidly accessible than the
corresponding original paper) may help more ambitious readers to decide
which research papers they wish to study. The list of such references is not
claimed to be complete.
I am grateful to Professor Goes for correspondence which has led to a
number of additions and improvements. My warmest thanks go to my friend
and colleague Dr. Jeff Sanders for his help with the revision.
Finally, my wife earns my gratitude for her help in preparing the revised
typescript.
R.E.E.
CANBERRA., January 1979
v
PREFACE
The principal aim in writing this book has been to provide an intro-
duction, barely more, to some aspects of Fourier series and related topics
in which a liberal use is made of modem techniques and which guides the
reader toward some of the problems of current interest in harmonic analysis
generally. The use of modem concepts and techniques is, in fact, as wide-
spread as is deemed to be compatible with the desire that the book shall
be useful to senior undergraduates and beginning graduate students, for
whom it may perhaps serve as preparation for Rudin's Harmonic Analysis
on Groups and the promised second volume of Hewitt and Ross's Abstract
Harmonic Analysis.
The emphasis on modem techniques and outlook has affected not only
the type of arguments favored, but also to a considerable extent the choice
of material. Above all, it has led to a minimal treatment of pointwise con-
vergence and summability: as is argued in Chapter 1, Fourier series are not
necessarily seen in their best or most natural role through pointwise-tinted
spectacles. Moreover, the famous treatises by Zygmund and by Baryon
trigonometric series cover these aspects in great detail, wl:tile leaving some
gaps in the presentation of the modern viewpoint; the same is true of the
more elementary account given by Tolstov. Likewise, and again for reasons
discussed in Chapter 1, trigonometric series in general form no part of the
program attempted.
A considerable amount of space has been devoted to matters that cannot
in a book of this size and scope receive detailed treatment. Among such
material, much of which appears in small print, appear comments on diverse
specialized topics (capacity, spectral synthesis sets, Helson sets, and 80
forth), as well as remarks on extensions of results to more general groups.
The object in including such material is, in the first case, to say enough for
the reader to gain some idea of the meaning and significance of the problems
involved, and to provide a guide to further reading; and in the second case,
to provide some sort of "cultural" background stressing a unity that
underlies apparently diverse fields. It cannot be over-emphasized that the
book is perforce introductory in all such matters.
The demands made in terms of the reader's active cooperation increase
vii
viii PREFACE
fairly steadily with the chapter numbers, and although the book is surely
best regarded as a whole, Volume I is self-contained, is easier than Volume II,
and might be used as the basis of a short course. In such a short course, it
would be feasible to omit Chapter 9 and Section 10.6.
As to specific requirements made of the reader, the primary and essential
item is a fair degree of familiarity with Lebesgue integration to at least
the extent described in Williamson's introductory book Lebesgue Integration.
Occasionally somewhat more is needed, in which case reference is made to
Appendix C, to Hewitt and Stromberg's Real and Abstract Analysis, or to
Asplund and Bungart's A First Course in Integration. In addition, the
reader needs to know what metric spaces and normed linear spaces are, and
to have some knowledge of the rudiments of point-set topology. The remain-
ing results in functional analysis (category arguments, uniform boundedness
principles, the closed graph, open mapping, and Hahn-Banach theorems)
are dealt with in Appendixes A and B. The basic terminology of linear
algebra is used, but no result of any depth is assumed.
Exercises appear at the end of each chapter, the more difficult ones being
provided with hints to their solutions.
The bibliography, which refers to both book and periodical literature,
contains many suggestions for further reading in almost all relevant directions
and also a sample of relevant research papers that have appeared since the
publication of the works by Zygmund, Bary, and Rudin already cited.
Occasionally, the text contains references to reviews of periodical literature.
My first acknowledgment is to thank Professors Hanna Neumann and
Edwin Hewitt for encouragement to begin the book, the former also for the
opportunity to tryout early drafts of Volume I on undergraduate students
in the School of General Studies of the Australian National University, and
the latter also for continued encouragement and advice. My thanks are due
also to the aforesaid students for corrections to the early drafts.
In respect to the technical side of composition, I am extremely grateful
to my colleague, Dr. Garth Gaudry, who read the entire typescript (apart
from last-minute changes) with meticulous care, made innumerable valuable
suggestions and vital corrections, and frequently dragged me from the
brink of disaster. Beside this, the compilation of Sections 13.7 and 13.8
and Subsection 13.9.1 is due entirely to him. Since, however, we did not
always agree on minor points of presentation, I alone must take the blame
for shortcomings of this nature. To him I extend my warmest thanks.
My thanks are offered to Mrs. Avis Debnam, Mrs. K. Sumeghy, and Mrs.
Gail Liddell for their joint labors on the typescript.
Finally, I am deeply in debt to my wife for all her help with the proof-
reading and her unfailing encouragement.
R.E.E.
CA.NBERRA, 1967
CONTENTS
Chapter 1 TRIGONOMETRIC SERIES

AND FOURIER SERIES 1
1.1 The Genesis of Trigonometric Series and Fourier Series 1

1.2 Pointwise Representation of Functions by Trigonometric
Series 3
1.3 New Ideas about Representation 7
Exercises 10
Chapter 2 GROUP ST.RUCTURE

A.lIm FOURIER SERIES 14
2.1 Periodic Functions 14

2.2 Translates of Functions. Characters and Exponentials.
The Invariant Integral 16
2.3 Fourier Coefficients and Their Elementary Properties 30
2.4 The Uniqueness Theorem and the Density
of Trigonometric Polynomials 40
2.5 Remarks on the Dual Problems 43
Exercises 46
Chapter 3 CONVOLUTIONS OF FUNCTIONS
3.1 Definition and First Properties of Convolution

3.2 Approximate Identities for Convolution
3.3 The Group Algebra Concept
3.4 The Dual Concepts
Exercises
ix
x CONTENTS
Chapter 4 HOMOMORPHISMS OF CONVOLUTION

ALGEBRAS 69
4.1 Complex Homomorphisms and Fourier Coefficients 69

4.2 Homomorphisms of the Group Algebra 72
Exercises 76
Chapter 5 THE DIRICHLET AND FEJER KERNELS.

CESARO SUMMABILITY 78
5.1 The Dirichlet and Fejer Kernels 78

5.2 The Localization Principle 81
5.3 Remarks concerning Summability 82
Exercises 85
Chapter 6 CEsARO SUMMABILITY OF FOURIER SERIES

AND ITS CONSEQUENCES 87
6.1 Uniform and Mean Summability 87

6.2 Applications and Corollaries of 6.1.1 90
6.3 More about Pointwise Summability 94
6.4 Pointwise Summability Almost Everywhere 95
6.5 Approximation by Trigonometric Polynomials 99
6.6 General Comments on Summability of Fourier Series 102
6.7 Remarks on the Dual Aspects 103
Exercises 104
Chapter 7 SOME SPECIAL SERIES AND THEIR

APPLICATIONS 109
7.1 Some Preliminaries 109

7.2 Pointwise Convergence of the Series (C) and (S) 114
7.3 The Series (C) and (S) as Fourier Series 117
7.4 Application to A(Z) 124
7.5 Application to Factorization Problems 124
Exercises 128
CONTENTS xi
Chapter 8 FOURIER SERIES IN L2 130
8.1 A Minimal Property 131
8.2 Mean Convergence of Fourier Series in V. Parseval's

Formula 131
8.3 The Riesz-Fischer Theorem 132
8.4 Factorization Problems Again 134
8.5 More about Mean Moduli of Continuity 135
8.6 Concerning Subsequences of 8 N! 137
8.7 A(Z) Once Again 139

Exercises 142
Chapter 9 POSITIVE DEFINITE FUNCTIONS

AND BOCHNER'S THEOREM 148
9.1 Mise-en-Scene 148

9.2 Toward the Bochner Theorem 149
9.3 An Alternative Proof of the Parseval Formula 152
9.4 Other Versions of the Bochner Theorem 152

Exercises 153
Chapter 10 POINTWISE CONVERGENCE

OF FOURIER SERIES 155
10.1 Functions of Bounded Variation and Jordan's Test 156
10.2 Remarks on Other Criteria for Convergence; Dini's Test 159
10.3 The Divergence of Fourier Series 160
10.4 The Order of Magnitude of 8 N!. Pointwise Convergence

Almost Everywhere 166
10.5 More about the Parseval Formula 171
10.6 Functions with Absolutely Convergent Fourier Series 173
Exercises 180
xii CONTENTS
Appendix A METRIC SPACES AND BAIRE'S THEOREM 187
A.I Some Definitions 187
A.2 Baire's Category Theorem 187
A.3 Corollary 188
A.4 Lower Semicontinuous Functions 188
A.5 A Lemma 189
Appendix B CONCERNING TOPOLOGICAL LINEAR SPACES 191
B.I Preliminary Definitions 191
B.2 Uniform Boundedness Principles 194
B.3 Open Mapping and Closed Graph Theorems 195
B.4 The Weak Compacity Principle 197
B.5 The Hahn-Banach Theorem 199
Appendix C THE DUAL OF LP (1 ~ P < co); WEAK

SEQUENTIAL COMPLETENESS OF Ll 201
C.I The Dual ofLP (I ~ P < (0) 201

C.2 Weak Sequential Completeness of Ll 202
Appendix D A WEAK FORM OF RUNGE'S THEOREM 205
Bibliography 207
Research Publications 213
Symbols 21a
Index 221
CHAPTER 1
Trigonometric Series and Fourier Series
1.1 The Genesis of Tri~onometric Series and Fourier Series

1.1.1. The Be~innin~s. D. Bernoulli, D'Alembert, Lagrange, and Euler,
from about 1740 onward, were led by problems in mathematical physics to
consider and discuss heatedly the possibility of representing a more or less
arbitrary function f with period 21T as the sum of a trigonometric series of the
form
2: (a" cos nx + b" sin nx),

IX>
Y2a o + (1.1.1)
,,=1
or of the formally equivalent series in its so-called "complex" form
(1.1.1*)
in which, on writing bo = 0, the coefficients c" are given by the formulae

(n = 0, 1,2, ... ).
This discussion sparked off one of the crises in the development of analysis.
Fourier announced his belief in the possibility of such a representation in
1811. His book Theorie Analytique de la Chaleur, which was published in
1822, contains many particulaJ," instances of such representations and makes
widespread heuristic use of trigonometric expansions. As a result, Fourier's
name is customarily attached to the following prescription for the coefficients
a", b", and c,,:
a" = -1
1T
f"
_"
f(x) cos nx dx, b" = -1 fn f(x)osin nx dx,
1T -n
(1.1.2)
c" =~
21T
fn
_II
f(x)e-'''X dx, (1.1.2*)
the a" and b" being now universally known as the "real," and the ell as the
"complex," Fourier coefficients of the function f (which is tacitly assumed to
be integrable over (-1T, 1T. The formulae (1.1.2) were, however, known
earlier to Euler and Lagrange.
1
2 TRIGONOMETRIC SERIES AND FOURIER SERIES
The grounds for adopting Fourier's prescription, which assigns a definite

trigonometric series to each function j that is integrable over (-1T, 1T), will
be scrutinized more closely in 1.2.3. The series (l.l.I) and (1.1.1 *), with the
coefficients prescribed by (1.l.2) and (1.1.2*), respectively, thereby assigned
to j are termed the "real" and "complex" Fourier series of j, respectively.
During the period 1823-1827, both Poisson and Cauchy constructed proofs
of the representation of restricted types of functions j by their Fourier series,
but they imposed conditions which were soon shown to be unnecessarily
stringent.
It seems fair to credit Dirichlet with the beginning of the rigorous study of
Fourier series in 1829, and with the closely related concept of function in
1837. Both topics have been pursued with great vigor ever since, in spite of
more than one crisis no less serious than that which engaged the attentions
of Bernoulli, Euler, d' Alembert, and others and which related to the pre-
vailing concept of functions and their representation by trigonometric series.
(Cantor's work in set theory, which led ultimately to another major crisis,
had its origins in the study of trigonometric series.)
1.1.2. The rigorous developments just mentioned showed in due course

that there are subtle differences between trigonometric series which converge
at all points and Fourier series of functions which are integrable over (-1T, 1T),
even though there may be no obvious clue to this difference. For example, the
trigonometric series
converges everywhere; but, as will be seen in Exercise 7.7 and again in 10.1.6,
it is not the Fourier series of any function that is (Lebesgue-)integrable over
( -1T, 1T).
The theory of trigonometric series in general has come to involve itself
with many questions that simply do not arise for Fourier series. For the
express purpose of attacking such questions, many techniques have been
evolved which are largely irrelevant to the study of Fourier series. It thus
comes about that Fourier series may in fact be studied quite effectively
without reference to general trigonometric series, and this is the course to be
adopted in this book.
The remaining sections of this chapter are devoted to showing that, while
Fourier series have their limitations, general trigonometric series have others
no less serious; and that there a.re well-defined senses a.nd contexts in which
Fourier series are the natural and distinguished tools for representing functions
in useful ways. Any reader who is prepared to accept without question the
restriction of attention to Fourier series can pass from 1.1.3 to the exercises
at the end of this chapter.
[1.2] POINTWISE REPRESENTATION OF FUNCTIONS 3
1.1.3. The Orthogonality Relations. Before embarking upon the

discussion promised in the last paragraph, it is necessary to record some 'facts
that provide the heuristic basis for the Fourier formulae (1.1.2) and (1.1.2*)
and for whatever grounds there are for according a special role to Fourier
series.
These facts, which result from straightforward and elementary calcula-
tions, are expressed in the following so-called orthogonality relations satisfied
by the circular and complex exponential functions:
2~ L =
(m # n, m ~ 0, n 0)
~ {~
~
00' 00' nx dx (m = n > 0),

(m = n = 0)
LL = ~ {r.
(m # n, m ~ 0, n ~ 0)
'in ,in nx dx (m=n>O),
(m = n = 0)
(1.1.3)
-2
1
1T
I"_"
cos mx sin nx dx = 0,
-1 I"_" e
21T
1m % e -1,,% dx = {O1 (m # n)
(m = n);
in these relations m and n denote integers, and the interval [ -1T, 1T] may be
replaced by any other interval of length 21T.
1.2 Pointwise Representation of Functions by Tri~onometric

Series
1.2.1. Pointwise Representation. The general theory of trigonometric
series was inaugurated by Riemann in 1854, since when it has been pursued
with vigor and to the great enrichment of analysis as a whole. For modern
accounts of the general theory, see [Zl]' Chapter IX and [Ba 1 2 ], Chapters
XII-XV.
From the beginning a basic problem was that of representing a more or
less arbitrary given function! defined on a period-interval I (say the interval
[ -1T, 1T]) as the sum of at least one trigonometric series (1.1.1), together with a
discussion of the uniqueness of this representation.
A moment's thought will make it clear that the content of this problem
depends on the interpretation assigned to the verb "to represent" or, what
comes to much the same thing, to the term "sum" as applied to an infinite
series. Initially, the verb was taken to mean the pointwise convergence ofthe
series at all points of the period interval to the given function f. With the
passage of time this interpretation underwent modification in at least two
ways. In the first place, the demand for convergence of the series to! at aU
points of the period-interval I was relaxed to convergence at almost all

points of that interval. In the second place, convergence of the series to f at
all or almost all points was weakened to the demand that the series be
summable to f by one of several possible methods, again at all or almost all
points. For the purposes of the present discussion it will suffice to speak of
just one such summability method, that known after Cesaro, which consists
of replacing the partial dums
so(x) = 12 ao ,
L (a" cos nx + b" sin nx)
N
SN(X) = 12ao + (N = 1,2, .. ) (1.2.1)
,,=1
of the series (1.1.1) by their arithmetic means

So + ... + SN (N = 0, I, 2, ... ).
aN = N + I (1.2.2)
Thus we shall say that the series (1.1.1) is summable at a point x to the
function f if and only if
lim aN(x) = f(x).
N_oo
It will be convenient to group all these interpretations of the verb "to

represent" under the heading of pointwise representation (everywhere or
almost everywhere, by convergence or by summability, as the case may be)
of the functionf by the series (1.1.1).
In terms of these admittedly rather crude definitions we can essay a
bird's-eye view of the state of affairs in the realm of pointwise representation,
and in particular we can attempt to describe the place occupied by Fourier
series in the general picture.
1.2.2. Limitations of Pointwise Representation. Although it is

undeniably of great intrinsic interest to know that a certain function, or
each member of a given class of functions, admits a pointwise representation
by some trigonometric series, it must be pointed out without delay that this
type of representation leaves much to be desired on the grounds of utility. A
mode of representation can be judged to be successful or otherwise useful as
a tool in subsequent investigations by estimating what standard analytical
operations applied to the represented function can, via the representation,
be expressed with reasonable simplicity in terms of the expansion coefficients
a" and b". This is, after all, one of the main reasons for seeking a representation
in series form. Now it is a sad fact that pointwise representations are in
themselves not very useful in this sense; they are simply too weak to justify
the termwise application of standard analytical procedures.
Another inherent defect is that a pointwise representation at almost a.ll
points of I is never unique. This is 80 because, a.s was established by Men'shov
[1.2] POINTWISE REPRESENTATION OF FUNCTIONS 5
in 1916, there exist trigonometric series which converge to zero almost every-
where and which nevertheless have at least one nonvanishing coefficient;
see 12.12.8. (That this can happen came as a considerable surprise to the
mathematical community.)
1.2.3. The Role of the Ortho~onality Relations. The a priori grounds

for expecting the Fourier series of an integrable function f to effect a point-
wise representation off (or, indeed, to effect a representation in any reasonable
sense) rest on the orthogonality relations (1.1.3). It is indeed a simple
consequence of these relations that, if there exists any trigonometric series
(1.1.1) which representsfin the pointwise sense, and iffurthermore the 8N (or
the aN) converge dominatedly (see [W], p. 60) to f, then the series (1.1.1)
must be the Fourier series of f. However, the second conditional clause
prevents any very wide-sweeping conclusions being drawn at the outset.
As will be seen in due course, the requirements expressed by the second
conditional clause are fulfilled by the Fourier series of sufficiently smooth
functions f (for instance, for those functions f that are continuous and of
bounded variation). But, alas, the desired extra condition simply does not
obtain for more general functions of types we wish to consider in this book.
True, a greater degree of success results if convergence is replaced by summa-
bility (see 1.2.4). But in either case the investigation of this extra condition
itself carries one well into Fourier-series lore. This means that this would-be
simple and satisfying explanation for according a dominating role to Fourier
series can scarcely be maintained at the out8et for functions of the type we
aim to study.
1.2.4. Fourier Series and Pointwise Representations. What has been

said in 1.2.3 indicates that Fourier series can be expected to have but
limited success in the pointwise representation problem. Let us tabulate a
little specific evidence.
The Fourier series of a periodic function f which is continuous and of
bounded variation converges boundedly at all points to that function. The
Fourier series of a periodic continuous function may, on the contrary,
diverge at infinitely many points; even the pointwise convergence almost
everywhere of the Fourier series of a general continuous function remained
in doubt until 1966 (see 10.4.5), although it had been established much
earlier and much more simply that certain fixed subsequences of the sequence
of partial sums of the Fourier series of any such function is almost everywhere
convergent to that function (the details will appear in Section 8.6). The
Fourier series of an integrable function may diverge at all points.
If ordinary convergence be replaced by summability, the situation
improves. The Fourier series of a periodic continuous function is uniformly
summable to that function. The Fourier series of any periodic integrable

function is sum mabIe at almost all points to that function, but in this case
neither the 8 N nor the UN need be dominated.
1.2.5. Trigonometric Series and Pointwise Representations. Having

reviewed a few of the limitations of Fourier series vis a.-vis the problem of
pointwise representation, we should indicate what success is attainable by
using trigonometric series in general.
In 1915 both Lusin and Privalov established the existence of a pointwise
representation almost everywhere by summability methods of any functionf
which is measurable and finite almost everywhere. For 25 years doubts
lingered as to whether summability could here be replaced by ordinary
convergence, the question being resolved affirmatively by Men'shov in 1940.
This result was sharpened in 1952 by Bary, who showed that, if the function
f is measurable and finite almost everywhere on the interval I, there exists a
continuous function F such that F'(x) = f(x) at almost all points of I, and
such that the series obtained by termwise differentiation of the Fourier
series of F converges at almost all points x of I to f(x). Meanwhile Men'shov
had in 1950 shown also that to any measurablef (which may be infinite on a
set of positive measure) corresponds at least one trigonometric series (1.1.1)
whose partial sums SN have the property that lim N _ oo SN = / in measure on
I. This means that one can write 8 N = UN + VN' where UN and V N are finite-
valued almost everywhere, limN _ '" uN(x) = f(x) at almost all points x of I,
and where, for any fixed e > 0, the set of points x of I for which IVN(X)l > e
has a measure which tends to zero as N -00. (The stated condition on the
V N is equivalent to the demand that
lim
N-",
J"
_"
IVNI dx = 0;
1 + IVNI
and the circuitous phrasing is necessary because f may take infinite values
on a set of positive measure.) This sense of representation is weaker than
pointwise representation. For more details see [Ba2]' Chapter XV.
These theorems of Men'shov and Bary lie very deep and represent enormous
achievements. However, as has been indicated at the end of 1.2.2, the
representations whose existence they postulate are by no means unique.
Cantor succeeded in showing that a representation at all points by a
convergent trigonometric series is necessarily unique, if it exists at all.
Unfortunately, only relatively few functions f admit such a representation:
for instance, there are continuous periodic functions f that admit no such
representation. (This follows on combining a theorem due to du Bois-Reymond
and Lebesgue, which appears on p. 202 of [Bad, with results about Fourier
series dealt with in Chapter 10 of this book.) It is indeed the case that, in a
sense, "most" continuous functions admit no representation of this sort.
[1.3] NEW IDEAS ABOUT REPRESENTATION 7
1.2.6. Summary. It can thus be said in summary that pointwise repre-

sentations are subject to inherent limitations as analytical tools, and that
Fourier series can be accorded a distinguished role in respect of this type of
representation only for functions of a type more restricted than one might
hope to handle.
This being so, it is natural to experiment by varying the meaning assigned
to the verb "to represent" in the hope of finding a more operationally effective
meaning and of installing Fourier series in a more dominating role.
RJore embarking on this program, it is perhaps of interest to add that a
similar choice prevails in the interpretation of differentiation (which in fact
has connections with the representation problem). The pointwise everywhere
or almost everywhere interpretation of the derivative, if deprived of any
further qualification, is also not entirely effective operationally. A new
interpntation is possible and leads to distributional concepts; Chapter 12 is
devoted to this topic.
1.3 New Ideas about Representation

1.3.1. Plan of Action. In the preceding section we have recounted some
of the difficulties in the way of according a unique position to Fourier series
on the grounds of their behavior in relation to the traditionally phrased
problem of representing functions by trigonometric series. We have also
indicated the shortcomings of this type of representation.
To this it may be added that in cases where the mathematical model of a
physical problem suggests the use of expansions in trigonometric series,
pointwise representations frequently do not correspond very closely to the
physical realities.
Faced with all this, we propose to consider new meanings for the verb "to
represent" that are in complete accord with modern trends, and which will in
due course be seen to justify fully a concentration on Fourier series as a
representational device.
1.3.2. Different Senses of Convergence and Representation. In

recent times analysts have become accustomed to, and adept at working in
diverse fields with, other meanings for the verb "to represent," most of
which (and all of which we shall have occasion to consider) are tantamount
to novel ways in which a series of functions may be said to converge. Such
ideas are indeed the concrete beginnings of general topology and the theory
of topological linear spaces.
Thus encouraged, we contemplate some possible relationships between an
integrable function! on ( -TT, TT) and a trigonometric series (1.1.1) or (1.1.1*)
expressed by each of equations (A) to (D) below.
For this purpose we write again
L (an cos nx + bn sin nx),

N
SO(X) = Y2a o , s,v(x) = Y2a o +

n=l
so that
s,v(x) = L cne1n:r,
Inl .. N
(1.3.1)
and also
The relationships referred to are (compare 6.1.1, 6.2.6, 12.5.3, and 12.10.1):
(A) ~~~ f" If(x) - uN(x)1 dx = 0;
(B) lim
N-CX)
J"
-n:
If(x) - SN(X)jP dx = 0;
(C) lim sUP:r If(x) - UN(X) I = 0;

N-a>
(D) ~~~ i:" u(x)s,v(x) dx = i:" u(x)f(x) dx
for each indefinitely differentiable periodic function u.

If anyone of these relations holds for a given f and a given trigonometric
series, one may say that the trigonometric series represents f in the corre-
sponding sense: in case (A) it would be usual to say that the trigonometric
series isCesaro-summable in mean with exponent (or index) 1 to f; in case (B)
that the trigonometric series is convergent in mean with exponent (or index) p
to f; in case (0) that the trigonometric series is uniformly Cesaro-summable to
f; and in case (D) that the trigonometric series is distributionally convergent
to f.
1.3.3. The Role of Fourier Series. It is genuinely simple to verify that,

given f, there is at most one trigonometric series for which anyone of relations
(A) to (D) is true, and that this only contender is the Fourier series of f (see
the argument in 6.1.3). Moreover, it is true that the relations do hold if the
trigonometric series is the Fourier series off, provided in case (B) that either
1 < p < 00 and f E V' or p = 1 and f log+ IfI E L1 (see 8.2.1, 12.10.1, and
12.10.2); and in case (0) that f is continuous and periodic. (The symbols
L1 and V' here denote the sets of measurable functions f on (-TT, TT) such
that IfI and IfI P, respectively, are Lebesgue-integrable over (-TT, TT). A
tiny modification to this definition is explained in detail in 2.2.4 and will be
adopted thereafter in this book.}
[1.3] NEW IDEAS ABOUT REPRESENTATION 9
Each of the relations (A) to (D) can, therefore, be used to characterize the
Fourier series of f under the stated conditions, and each provides some
justification for singling out the Fourier series for study. (There are, by the
way, numerous other relationships that might be added to the list.)
It turns out that the weakest relationship (D) is suggestive of fruitful
generalizations of the concept of Fourier series of such a type that the
distinction between Fourier series and trigonometric series largely disap-
pears. It suggests in fact the introduction of so-called distributions or
generalized functions in the manner first done by L. Schwartz [Sl.2]' It will
then appear that any trigonometric series in which Cn = O(lnl") for some
k may be regarded as the Fourier series of a distribution, to which this
series is distributionally convergent. These matters will be dealt with in
Chapter 12.
1.3.4. Summary. The substance of Section 1.2 and 1.3.3 summarizes the
justification for subsequent concentration of attention on Fourier series in
particular, at least insofar as reference is restricted to harmonic analysis in
its classical setting. We shall soon em bark on a program that will include at
appropriate points a verification of each of the unproved statements upon
which this justification is based. As for trigonometric series in general, we
shall do no more than pause occasionally to mention a few of the simpler
results that demand no special techniques.
A bird's-eye view of-many of the topics to be discussed at some length in
this book is provided by the survey article G. Weiss [1].
1.3.5. Fourier Series and General Groups. There are still other reasons
in favor of the chosen policy which are based upon recent trends in analysis.
Harmonic analysis has not remained tied to the study of Fourier series of
periodic functions of a real variable; in particular it is now quite clear tha.t
Fourier-series theory has its analogue for functions defined on compact Abelian
groups (and even, to some extent, on still more genera.l groups); see, for
example, [HR], [Re], [E l ]. While the level at which this book is written
precludes a detailed treatment of such extensions, we shall make frequent
reference to modern developments. However regrettable it may seem, it is a
fact that these developments cluster arOlind the extension of precisely those
portions of the cla.ssical theory which do not depend upon the deeper properties
of p~intwise convergence and summability, and that a detailed treatment of
the analogue for compact groups of the theory of general trigonometric series
appears to lie in the future. Moreover, the portions of the classical theory that
have so far been extended appear to be those most natural for handling those
problems which are currently the center of attention in general harmonic
analysis. Of course, these prevailing features may well change with the pa.ssa.ge
of time. While they prevail, however, they add support to the view that it is
reasonable to accord some autonomy to a theory in which the modes of
representation mentioned in 1.3.2 take precedence over that of pointwise
representation.
EXERCISES
1.1. Establish the formulae
( ) -_ ~
D NX "" 170% _ sin (N + %)x
e - . u '
1701.; N SIn /2X
FN(x) = (N + 1)-1 [Do(x) + ... + DN(x)]
= (N + 1)-1 [Sin ~(N + 1)x]2
sm%x
for N ~ 0 an integer and x ~ 0 modulo 217, where the equality signs im-
mediately following DN(x) and F N(X) are intended as definitions for all real x.
1.2. Prove that if p and q are integers and p < q, and if x ~ 0 modulo 217,
then
I L:
P.; .. .;q
eln:rl ~ Icosec %xl
By using partial summation (see 7.1.2 and [HJ, p. 97 if.) deduce that if
cp ~ CP+l ~ ~ Cq ~ 0, then, for x ~ 0 modulo 217,
L:
p.;n.;q
cneln%1 ~ cp Icosec %xl
1.3. Assume that Cn ~ Cn + 1 and lim .. _<Xl Cn = O. Show that the series
is convergent for x ~ 0 modulo 217', and that the convergence is uniform on

any compact set of real numbers x which contains no number =0 modulo 217'.
1.4. Assume that Cn ~ C"+ 1 ~ 0 and nc.. ~ A. Show that
1L:
N
Cn sin nxl ~ A(lT + 1).
70=1
Hints: One may assume 0 < x < 17. Put m = min (N, [IT/XJ) and split
the sum into 2:i' + 2:~ + 1> an empty sum being counted zero. Estimate the
partial sums separately, using Exercise 1.2 for 2~+l'
1.5. Assume that the Cn are as in Exercise 1.4. Show that the series
2::'= 1 c.. sin nx is boundedly convergent, and that the sum function is con-
tinuous, except perhaps at the points x =
0 modulo 217'. (More general
results will appear in Chapter 7.)
EXERCISES 11
1.6. Compute the complex Fourier coefficients of the following functions,

each defined by the prescribed formula over [ -71',71') and defined elsewhere
so as to have period 271':
(1) f(x) = x;
(2) f(x) = Isin xl ;
(3) f(x) = x for -71' ~ X ~ 0, f(x) = 0 for 0 < x < 71'.
1.7. By a trigonometric polynomial is meant a function f admitting at
least one expression of the form
f(x) = 2: cneilll:,
Inl .. N
where the Cn are f-dependent complex numbers.

(1) Use the orthogonality relations to show that, if f is a trigonometric
polynomial, then
j(n) == -1
271' J"
_II
f(x)e-inl: dx
vanishes for all but a finite number of integers nand thatf(x) = Lnezj(n)einl:.
Show also that
1
271' J"
_II If(x)i2 dx = .
nfz If(nW
whenever f is a trigonometric polynomial. (This is a special case of Parseval's
formula, to which we shall return in Chapter 8 and Section 10.5; see also
Remark 6.2.7.)
A trigonometric polynomial f such that j(n) = 0 for Inl > N is said to be
of degree at most N.
(2) Verify that the set TN of trigonometric polynomials of degree at most
N forms a complex linear space of dimension 2N + 1 with respect to point-
wise operations, and that if f E TN' then also Re f E TN and 1m f E TN'
(3) Show that if f E TN' f =ft 0, then f admits at most 2N zeros (counted
according to multiplicity) in the interval [0,271') (or in any interval congruent
modulo 271' to this one).
1.8. (Steckin's lemma) Suppose f E TN is real-valued, and that
Ilfll", == sup If(x)1 = M = f(xo)

%
Prove that
f(xo + y) ~ M cos Ny for Iyl ~ ; .
Hints: Put g(y) = f(x o + y) - M cos Ny. Assuming the assertion false,
we choose Yo so that IYol < 71'/N and g(yo) < O. We assume 0 < YO<71'/N;
otherwise the subsequent argument proceeds with the interval [-271',0) in
place of [0,211-)' By examining closely the signs of g at the points k.,,/N

(k = 0, 1,2, .. " 2N), show that g admits at least 2N + I zeros in [0, 2.,,). A
contradiction results from Exercise 1.7.
1. 9. (Bernstein's inequality) Prove that if I E TN, then 11f' I '" ~ N 11/11",
(the notation being as in the preceding exercise).
Hints: It suffices, by Exercises 1.7 and 1.10, to prove the inequality for
real-valued I E TN' Iff'(xo) = m == 11f'11", (which can be arranged by changing
I into -I if necessary) and 111 = 11/11"" Exercise 1.8 givesf'(xo + y) ~ m C08
Ny for Iyl ~ ."IN. Integrate this inequality.
Notes: Many other proofs are known; the above, due to Steckin, is perhaps
the simplest. For a proof based upon interpolation methods, see [Z2]' p. 11.
More general results, also due to Bernstein, apply to entire functions of order
one and exponential type; see [Z2], p. 277.
See also the approach in [Kz], p. 17; W. R. Bloom [1], [2]; MR 51 # 1239;
52 ## 6288, 11446; 53 # 11289; 54 # 829.
The inequality has also been extended in an entirely different way by
Privalov, who showed that if J = (a', b') and J = (a, b) are any two sub-
intervals of [-11,11] satisfying a < a' < b' < b, then there exists a number
c(I, J) such that
sup II'(x) 1~ c(J, J)N . sup 11(x)1
.%61 xeJ
for any 1 E TN' It is furthermore established that similarly (but perhaps with a
different value for c(I, J)) one has
for any 1 E TN and any p satisfying 1 ~ P < 00. Both inequalities are also valid
when I = J = [ -11,11] and c(J, J) = 1, the first reducing to that of Bernstein
and the second being in this case due to Zygmund. For more details, see [B~],
pp. 458-462. See also [L2]' Chapter 3.
1.10. Suppose that E is a complex linear space of complex-valued

functions on a given set (pointwise operations), that E = Eo + iEo where
Eo is the set of real-valued functions in E, that l is a complex-linear functional
on E which is real-valued on Eo, and that p is a seminorm on E (see Appendix
B.1.2). Suppose finally that p(x) ~ p(y) whenever x, y E E and Ixl ~ Iyl,
and that Il(x)1 ~ p(x) for x E Eo. Prove that Il(x)1 ~ p(x) for x E E.
Hints: Write x = a + ib with a, b E Eo and l(x) = r(a + ifJ) with r ~ 0,
a and fJ real, and a2 + fJ2 = 1. Then
Il(x)1 = r = (a - ifJ)l(x) = l[(a - ifJ)(a + ib)];

expanding and taking real parts: Il(x)1 = l(aa + fJb) ~p(aa + fJb), and 80
forth.
EXERCISES 13
1.11. Prove that, if a trigonometric polynomial f is real-valued and

nonnegative, then f = Igl2 for some trigonometric polynomial g (Fejer
and F. Riesz).
Hints: Suppose f(x) == Llnl .. NCne 1nz and consider first the case in which
f(x) > 0 for all x. Assume (without loss of generality) that c_ N i' 0 and
examine the polynomial P(z) = ZN Llnl .. NCnZn. Observe that P(z) = Z2N P(Z-I)
andf(x) = e- 1Nz P(e IZ ). Verify that the zeros of P are of the form aI' a2 , ,
and all, ail, ... , where 0 < larl < I, and factorize P accordingly.
In case one knows merely thatf ~ 0, apply the above to thefk = f + Ilk
(k = I, 2, ... ) and use a limiting argument.
Remarks. The theorem does not extend in the expected way to other
groups; see [R], 8.4.5.
CHAPTER 2
Group Structure and Fourier Series
The aim of the first two sections of this chapter is to show how and to
what extent the topological group structure of the set R of real numbers,
and of some of its subgroups and quotient groups, lead to the study of
periodic functions, the complex exponential functions, and the problem of
expansions in trigonometric series in general and Fourier expansions in
particular. In the remaining sections of this chapter we shall begin the study
of Fourier coefficients in some detail.
In pursuing the aims of Sections 2.1 and 2.2 we are led to refer to fairly
general topological groups and to constructs related to them. It is hoped
thus to convey a very rough idea of how the classical theory of Fourier series
fits into contemporary developments in parts of analysis, and to prepare the
reader for a later perception of genuine structural unity underlying obvious
similarities. It is of course not expected, nor is it necessary for an under-
standing of subsequent developments in this book, that the reader should
at this stage stop to gather the details concerning topological groups and the
related concepts to be spoken of (duality, invariant integrals, and so on);
this writer will indeed venture the opinion that the return to a detailed study
of generalizations is best made after some familiarity with special cases has
been attained. On the other hand the reader will, it is hoped, gain from the
realization that the classical theory is tributary to a broader stream, and will
in due course want to try his hand at exploring the latter with the help of the
references cited in this chapter.
2.1 Periodic Functions

For any reasonable interpretation of the term "represent" (see Chapter 1),
any function of a real variable which is to be globally representable by a
trigonometric series must admit 21T as a period, or must do this after suitable
correction on a null set. Insofar as such correction does not alter the Fourier
series of the function, we may and will assume that all functions of a real
variable have period 21T. (Representation over a restricted range by so-called
"half-ra.nge series" does not in any way conflict with this convention.)
14
(2.1] PERIODIC FUNCTIONS 15
2.1.1. The Groups Rand T. The set R of real numbers, taken with
addition as its law of composition and with its usual metric topology, is an
example of an Abelian topological group. This means that it is first an
Abelian group, and second a topological space, and that moreover the
algebra and the topology are so related that the mapping (x, y) - x - y is
continuous from R x R into R. If one drops the demand that the group
structure be Abelian, one has here the concept of a topological group in
general; see [B], pp. 98 ff., and/or [HR], Chapter II. Hereinafter the term
"group" will always mean "locally compact group whose topology satisfies
the Hausdorff separation axiom." This particular topological group R is
locally compact but noncompact. We wish to focus attention, not so much on
R, as on quotient groups thereof.
It is a simple matter to show that the only closed subgroups of R, other
than {O} and R itself, are those consisting of all integer multiples of some
nonzero positive number (see Exercise 2.1). Which of these is selected is
largely immaterial: we choose that one which is formed of all integer multiples
of 21T and which is hereinafter denoted by 21TZ (Z denoting the additive
group of integers).
Let us form the quotient group Rj21TZ = T and denote by p the natural
projection of R onto T, which assigns to x E R the coset x = x + 21TZ
containing x. The group T is made into a topological group by endowing it
with the so-called quotient topology. In concrete terms, this means that the
open sets in T are precisely the sets p( U) where U is open in R. Even more
concretely put, the quotient topology on T is that defined by the metric
d(x, iI) = inf {Ix - y + 2n1Tl: n E Z}.
Another way of looking at T is to recognize that the mapping x _ exp(ix)
is an isomorphism of T onto the multiplicative group of complex numbers
having unit absolute value. In this isomorphism, the quotient topology
corresponds to that induced on the unit circumference in the complex plane
by the usual metric topology on the latter. In view of this, the group T is
often referred to as the circle group or the one-dimensional torus group.
Perhaps the most essential difference between Rand T is that the latter is
compact. Were we to attempt to apply to R the subsequent considerations
concerning T, we should be led to Fourier integrals in place of Fourier series;
almost a.ll the additional difficulties thereby encountered would stem from
the fact that R is noncompact.
2.1.2. Periodic Functions. Iff is a periodic function on R (by "periodic"

we shall always mean "with period 21T "), there is just one function j on T
such that f = j 0 p. (Notice that we shall never speak of so-called "many-
valued functions.") Conversely, every periodic functionf on R is expressible
in this way. Moreover, in this one-to-one correspondence f ~ j, continuous
16 GROUP STRUCTURE AND FOURIER SERIES
!'S correspond to continuous j's. It will in fact be the case that this corre.
spondence preserves every structure relevant to our purpose, and we shall
before long ask the reader to make a mental identification of / and f.
It is also perfectly legitimate to regard functions on the circle group as
functions of the complex variable z = el:r having unit absolute value, but we
shall make no systematic use of this notation.
2.1.3. Role of the Group Structure. As we shall see in Section 2.2, the
topological group structure of T is inextricably bound up with the genesis
and study of Fourier series. Indeed, it will slowly emerge that many of the
most fundamental aspects of this study depend almost exclusively on the
fact that T is a compact Abelian group. It will be seen, too, that the Lebesgue
integral itself is determined (up to a nonzero constant factor of proportionality)
by the topological group structure.
To this basic ingredient may be added, for the sake of richness and
refinement, more specialized structures and concepts-the concepts of
bounded variation and differentiability for functions, for example. In line
with the remarks in 2.1.2, we say that a function / on T is of class Ck (=k
times continuously differentiable, or indefinitely differentiable if k = CX),
or is of bounded variation, if and only if the function / 0 p on R has the
corresponding property on some one (and therefore every) interval in R of
length 27T.
2.2 Translates of Functions. Characters and Exponentials. The

Invariant Integral
2.2.1. Translates and Characters. We pose the question: What are the
fundamental reasons for considering expansions in terms of cosines and sines
cos AX and sin Ax, or, equivalently, in terms of the complex exponentials eIA:r?
The historical answer, which is also the one based on applications, might
be that these functions are the eigenfunctions of particularly simple linear
differential operators. The restriction of the continuous parameter A to the
discrete range 27TZ reflects periodic boundary conditions.
There is, however, another and even more fundamental explanation,
which hinges only on the topological group structure of Rand T. Let us look
into this.
The simplest and most obvious way in which the group structure can be
used in handling functions is via the translation operators T" (a = a group
element) acting on functions according to the rule
T"f(x) =/(x - a).
Attention paid to the T a is justified in retrospect, because most of the linear
operators featuring in harmonic analysis prove to be limits in some sense of
[2.2] TRANSLATES OF FUNCTIONS 17
linear combinations of translation operators (see, for example, 3.1.9 and

Chapter 16).
To fix ideas, we visualize the Ta as acting on the linear space C = C(R) or
C(T) of continuous, complex-valued functions on R or T as the case may be.
(Almost all we have to say would remain true on replacing C = C(R) or C(T)
by various other function spaces over R or T.)
If f E C, then Taf E C. Each T a is indeed an automorphism of the linear
space C. To this we add for future reference the relations
To = I, (2.2.1)
where I denotes the identity automorphism of C.
In general, and certainly for the groups Rand T here considered, the space
C is infinite-dimensional and the problem of analyzing the behavior of the
operators T a on C is a complicated one. However, elementary linear algebra
(and, even more so, suitable forms of the simultaneous spectral resolution
theorem) encourage one to hope for simplification if one can "reduce" the
problem by finding linear subspaces V of C which are invariant in the sense
that T a(V) c V for all group elements a. For brevity we term such a V an
invariant 8ub8pace. The hope would lie in decomposing C into some sort of
(possibly infinite) direct sum of invariant subspaces V1 , V2 , " , each VI
being as small as possible. The T a could then be examined on each VI
separately.
In this way one is led to consider the existence of minimal invariant sub-
spaces V of C, "minimal" meaning that V contains properly no invariant sub-
space other than {O}. Now it is evident that a onedimensional invariant
subspace V (if any such there be) is certainly minimal; and that such a
subspace V is generated or spanned by a function f which is a simultaneous
eigenvector of the Ta (if any such functions exist). So, without more ado, we
seek such functions. (For nonAbelian groups in general there would not
exist anyone-dimensional invariant subspaces-one would have to be content
with seeking finite-dimensional ones, which in fact exist in abundance for
compact groups; for noncompact, non-Abelian groups, the situation is even
more complicated.)
Given fEe, denote by V, the smallest invariant subspace containing f,
that is, the set of all finite linear combinations of translates Taf of f. We
seek functions f such that dim V, = l. Clearly, therefore, f "# 0 and to each
group element a corresponds a complex scalar x( -a) such that
Taf = x( -a)f.
This signifies that
f(x - a) = x( -a) f(x) (2.2.2)
for all pairs (a, x). If x = 0, f( -a) = f(O) x( -a), which shows in particular
that X is continuous and that X -# O. On the other hand, (2.2.1) and (2.2.2)
yield the functional equation
X(a + b) = X(a) X(b). (2.2.3)
As a matter of definition, a complex-valued function X -# 0 that satisfies

(2.2.3) is termed a character of the group in question. It follows at once that X
is nonvanishing, X(O) = 1, and X( -a) = X(a) -1. We shall have occasion to
consider only characters that are continuous. If a character X is bounded,
then (2.2.3) shows that Ix(a)1 = 1 for all group elements a, so that X defines a
homomorphism of the group into the multiplicative group of complex
numbers of absolute value 1.
Returning to (2.2.2), we may say that the function X appearing there is a
continuous character. Moreover, since (2.2.2) gives f( -a) = X( -a)f(O) for
all a, it follows that e = f(O) is nonzero, and f = ex is thus a nonzero scalar
multiple of the character x.
Let us next determine explicitly all the continuous characters of R and of
T. Concerning characters that are not assumed to be continuous, see Exercise
3.19.
Supposing that X is a continuous character of R, we integrate the relation
(2.2.3) with respect to b, over an interval (0, h), to get
f x(a + b) db = X(a)' f X(b) db.
Since X is continuous and X(O) = 1, h may be chosen and fixed so that the
factor
J: X(b) db
is nonzero. Moreover,
Jo
r x(a + b) db = Jara+ll X(e) de.
ll
Again since X is continuous, this last expression is a differentiable function

of a. It follows that X is differentiable. Knowing this, we find that (2.2.3)
yields
+ h)
'()
Xa=lm
l'
1l .... 0
x(a
I.
L
- x(a)
1 x(h) - x(O) ()
=lm
1l .... 0
h 'x a ,
so that X satisfies the differential equation
X' = ik X, (2.2.4)
where k = -iX'(O). The only solution of (2.2.4) taking the value 1 at the
origin is
x(x) = etkx (2.2.5)
Evidently, whatever the complex number k, (2.2.5) defines a continuous
character of R. This character is bounded, if and onJy if k is real.
To determine the continuous characters of T, it is merely necessary to add
the demand that X have period 27T. This signifies that k E Z.
To sum up, we find that
(1) The continuous (and so necessarily bounded) characters of T are in
one-to-one correspondence with Z, the character corresponding to 11, E Z
being (derived by passage to the quotient from) the function
(2.2.6)
Corresponding to 11, = 0 is the character eo, which is the constant function 1;
this is usually termed the principal character.
(2) The one-dimensional invariant subspaces of CrT) are precisely the
subspaces V" = {"e,,: " a scalar}, where n ranges over Z.
(3) The problem of harmonic analysis in respect to CrT) (and similarly in
respect of other function spaces) may be suggestively but perhaps oversimply
described as that of expressing C( T) as some sort of direct sum of the
subspaces V" (11, E Z). This task falls into two parts:
(a) GivenfE CrT), it is required to determine the corresponding "com-
ponents " off lying in the various subspaces V". This is, strictly speaking, the
problem of harmonic (or spectral) analysis and is, in the case of compact
Abelian groups anyway, relatively simple. The said components are just the
functions J(n)e", where
1(11,) = 2~ ff(x)en(x) dx.
It will appear in Chapter 11 that the component j(n)e n is nonzero, if and only
if Vn n V of: {O}, where V is the closed invariant subspace generated by f.
(b) The study of the formula
f = L J(n)e
nEZ
n,
which it is hoped will reconstitute f from its harmonic components. This may
be described as the problem of the harmonic (or spectral) synthesis of f. It
presents what is by far the more difficult part of the program and embraces,
of course, the question of representing f by a trigonometric series. It must
be stressed that such a series representation is indeed generally impossible in
C, if one dema.nds pointwise convergence. The study of the sense in which
the synthesis is valid (which will vary from one function space to another)
is an essential part of the problem before us; see the remarks in 10.3.6 and
Section 16.8.
In connection with (1) above, it is interesting to obscrve that the group
structure of Z corresponds, when Z is used to label the characters en, to point-
wise multiplication of characters. Moreover, thc corresponding "dual topology"
on Z is that for which the relation n -- no signifies that
en(x) ->- eno (x)
uniformly for x E T, and turns out to be just the discrete topology on Z (having
a base of neighborhoods of 0 E Z comprising the one set {O}). This is a general
feature: the bounded continuous characters of any given group may them-
selves be formed into a group under pointwise multiplication, termed the dual
or character group of the given group, and topologized in such a way that
(speaking informally) a sequence or net (xtl of characters converges to the
character X if and only if lim, X,(x) = X(x) uniformly for x E K, and this for
each compact subset K of the original group. Up to this point, everything is
largely a matter of observation and definition. The interesting and decidedly
nontrivial fact is that, by way of justification of the term "duality," the
character group of the character group turns out to be (isomorphic with) the
original group. This duality is profound and is fundamental in general harmonic
analysis, but to develop the ideas in any generality would take us much too
far afield. Suffice it to say that locally compact Abelian groups run around in
mutually dual pairs-such as (R, R) and (T, Z)-either member of such 'l pair
being isomorphic with the dual of the other: this is the so-called Pontryagin
duality law, for more details of which the reader is referred to [B), Chapter 11,
and [RR], Chapter VI. Our main concern will always be harmonic analysis on
the group T, but we shall from time to time cast fleeting glances at the dual
problems concerning harmonic analysis on the group Z, which is always to be
thought of as being endowed with its discrete topology. To the reader we issue
a standing invitation to reflect on the possible analogues for Z of results
established in the text for T. As a start, he might verify that, in conformity
with the Pontryagin duality law, the character group of Z can be identified
with T in the manner suggested by (2.2.6); that is, to each bounded (necessarily
continuous) character ~ of Z corresponds exactly one x E T such that ~(n) =
en (x) for n E Z and that the initial topology on T corresponds exactly to its
dual topology under the association x ~ ~:r (see Exercise 2.3).
Studies of harmonic analysis on each of the groups T and Z form, when
taken together, a useful forerunner to that of general harmonic analysis. This
is partly because they illustrate separately a number of the difficulties that
one encounters in an intermingled state when one moves along to harmonic
analysis on general groups. Actually, the next degree of complexity is repre-
sented by the group R (the additive group of real numbers with its usual
topology). In T, R, and Z one has, so to speak, the natural building bricks
from which quite general locally compact Abelian groups may be built up_ It
is known, for example, that any such compactly generated group is isomorphic
with a product Ra x Zb X F, where a and b are nonnegative integers and F is
a compact Abelian group (see [RR], Theorem (9.8)); moreover, F is isomorphic

to a closed subgroup of a possibly infinite product of copies of T; and there
exist arbitrarily small closed subgroups H of F such that FIH is isomorphic
with Tc x Fo, where c is a nonnegative integer and Fo is a finite group (see
[RR], Theorem (24.7)). These facts are quoted merely in order that the reader
may get some idea of how the limited program attempted by this book fits
into the scheme of general harmonic analysis; they will never be used hereafter
in this book.
The reader would do well to peruse the survey article G. Weiss [1], which
deals with both classical and modern aspects of the subject.
Before temporarily leaving the present topics we should indicate that in

Chapter 11 the theory of Fourier series obtained up to that point will be
used to classify all the closed in variant subspaces of C( T) (and of certain other
function spaces). The theory will also show to what extent C(T) (and these
other function spaces) can be decomposed into a direct sum of one-dimensional
(and therefore minimal) invariant subspaces.
2.2.2. The Invariant Inte~ral. Let us momentarily broaden the outlook

by considering a locally compact topological group G (see 2.1.1); for the
moment we do not assume that G is Abelian. Owing to this we must be careful
to specify that our concern will lie with the left translation operators T ..
defined by TaJ(x) = f( -a + x). If G is Abelian, this agrees with the notation
introduced in 2.2.1; in the contrary case one must distinguish these T .. from
the right translation operators f(x) - f(x - a).
Denote by Cc(G) the linear space of complex-valued continuous functions
f on G, each of which vanishes outside some f-dependent compact subset of
G. Evidently, Cc(G) is a linear subspace of C(G). If G is compact (for example,
if G = Rj27TZ), Cc(G) and C(G) are identical.
A fundamental and cardinal fact underlying all general harmonic analysis
is the existence of a linear functional 1 on Cc(G) which is
(I) positive, in the sense that l(f) > 0 if f '" 0 is a nonnegative real-
valued function in Cc(G); and
(2) left (translation) invariant, in the sense that
I(T..J) = l(f)
for all f E Cc(G) and all a E G.
It is also a fact that, apart from a positive factor of proportionality, there
is only one such functional. Any such functional 1 is called a left invariant
(or left Haar) integral on G. (Similar remarks apply to right Haar integra18.)
It is known that the left invariant integral can in all cases be extended to
more general functions in such a way as to preserve the basic, crucial, and
pleasant properties of the Lebesgue integral of functions of one real variable.
The details of this extension are to be found in a.ny one of several references,
2'2 GROUP STRUCTURE AND FOURIER SERIES
for example: [HR], Chapters III and IV; [HS], Chapter 3; [B], Chapters
8-10; [E], Chapter 4. However, an intelligent reading of the present book
will demand no more than a knowledge of the results of this extension for
functions of one real variable; it is of little importance which of several
possible approaches to the Lebesgue integral has been followed. More details
about what we shall need to assume appear in 2.2.4.
The choice of Cc(G) in place of C(G) as the initial domain of definition of I
comes about in the following way. It is quite easy to see at the outset that,
whenever Gis noncompact, there cannot be any invariant integral I for which
IU) is finite-valued for all nonnegative real-valued 1 E C(G) (or even for all
nonnegative real-valued 1 E C(G) which tend to zero at infinity). In other
words, the "integrability" of a function will demand quite severe restrictions
on the "average smallness" of the function at infinity. One v~ry simple and,
as it turns out, entirely effective way of imposing a priori such a restriction
on 1 is to demand that it shall vanish outside some compact subset of G.
(Of course, it turns out ultimately that this condition is not necessary for
integrability.)
It is not too much to say that the inauguration of modern harmonic
analysis on groups had to await the discovery, by Haar in 1933, of the
existence of a left invariant integral on any locally com pact group G satisfying
the second countability axiom. Subsequent developments, including the
removal of all countaobility restrictions on G, have been due to Weil, Kakutani,
H. Cartan, von Neumann, and many others. The interested reader may also
wish to consult a recent note by Bredon [1]. See also MR 39 # 7066.
On considering some familiar groups, old friends appear in a new light.
For example, if G = Rn, the characteristic properties (1) and (2) show that an
invariant integral is
I(f) = J-.. Jl(xl'" " x,,) dXl ... dx,,,

a Lebesgue (or Riemann) integral extended over any hypercube outside
which 1 vanishes.
Again, if G is R/27TZ, an invariant integral is
IU) = LJ1 0 p(x) dx, (2.2.7)
a Lebesgue (or Riemann) integral extended over any interval of R of length

27T. Here we have chosen the disposable proportionality factor so as to
arrange that 1(1) = 1 (a choice that is possible for compact groups and for
those only).
The essential uniqueness of the invariant integral for the circle group
T (and likewise for R") can be established by quite simple and down-to-earth
arguments, as follows. We handle functions on T as if they were periodic
functions on R (see 2.1.2). Let us first note that any invariant integral I has
the property that
II(!)I ~ I(l} . sup If I ; (2.2.8)
this follows from property (1) and the linearity of I. Now, iffis continuously
differentiable,
lim Taf - f = -1'
a-O a
holds uniformly. Consequently (2.2.8) and property (2) combine to show that
l(j') = 0 for any continuously differentiable f. Next, if g is continuous and
periodic and satisfies
[2"
Jo g(x) dx = 0,
then g = 1', where
f(x) = 50" g(t) dt
is continuously differentiable and periodic. Thus I(g) = 0 for such g. Finally,
choose any nonnegative continuous periodic ho such that
fIt ho(x) dx = 1.
Given any continuous periodic h, we apply what precedes to the function g
defined by
g(x) = h(x) - ho(x) Jo[2" h(t) dt,
which is continuous and periodic, and satisfies
f" g(x) dx = O.
The result of this application, namely, the conclusion I(g) = 0, signifies

exactly that
[2"
I(h) = I(h o) Jo h(x) dx,
showing that I differs from the expression appearing on the right-hand side
of (2.2.7) by the oonstant faotor 217I(h o) > O. This oompletes the verification
of the essential uniqueness of the invariant integral on T.
Armed with this uniqueness property, it is simple to deduce other invariance
properties of the integral. The elementary properties of the Riemann integral
show that, if f is a oontinuous periodic funotion on R, and if k E Z and
k =1= O,then
1 [211 1 [2"
217 Jo f(kx) dx = 217 Jo f(x) dx. (2.2.9)
This can be established by using the identification afforded by (2.2.7) and

invoking the uniqueness property of the invariant integral. In doing this, we
may replace T by any compact group G and the mapping x -+ kx by any
continuous group homomorphism t of G onto itself. Let I be the invariant
integral on G, normalized so that 1(1) = 1. We will show that
I(f 0 t) = I(f) (2.2.10)
for all continuous functions I on G, from which (2.2.9) will follow by special-
ization. It is to be observed that, since k E Z and k =F 0, the mapping
t: x -+ (kxy is a continuous homomorphism of T onto itself.
To prove (2.2.10), we consider the new functional l' defined by
1'(j) = I(j 0 t); (2.2.11)
this definition is effective since, owing to the continuity of t, lot is con-
tinuous whenever I is continuous. Since t maps G onto G, it is clear that I'
enjoys property (1) of invariant integrals. Since also
T ",(1 0 t)(x) = 1 t(x - a) = I[t(x -
0 a = I[t(x) - t(a
= Tc(a,j[t(x)] = (Tt(a,j) 0 t(x) ,
owing to the fact that t is a group homomorphism, we have Ta(lo t)

= (TI(al/) 0 t. We again use the assumption that t maps G onto G; then the
translation invariance of I shows that l' is also translation-invariant. By
uniqueness, therefore, there is a number c such that 1'(j) = C I(j) for all
continuous functions I. Choosing 1= 1, we find that (2.2.11) gives 1'(1)
= 1(1 0 t) = 1(1) = 1. Hence c = 1 and l' is identical with I. This is just
what (2.2.10) asserts.
The dual situation. Let us turn momentarily from the circle group T
to the dual group Z. There is no lasting mystery about the invariant integral
on Z; apart from a disposable constant of proportionality, it must be ex-
pressed by summation:
I(t/ = 2 t/>(n),
nsZ
(2.2.7)
at least for those functions t/> on Z whose support {n E Z : t/>(n) =F O} is

finite. (The compact subsets of a discrete space, such as Z, are exactly the
finite subsets thereof.)
The linear space offunctions t/> on Z having finite support is, however, too
narrow to accommodate fully effective operation, and it is desirable that the
invariant integral be extended to other functions. No problem arises for
those functions t/> for which the series in (2.2.7) is absolutely convergent:
this is the space usually denoted by tl(Z) and is the exact analogue, for the
group Z, of the space Ll of Lebesgue integrable functions on T.
To go still further, it will be necessary to interpret the right-hand side of

(2.2.7*) according to one of a number of conventions. For example, the
conditional convergence of the series will by convention always mean the
existence of a finite limit for the sequence of symmetric partial sums
L cp(n)
Inl .. N
when N _ 00. A yet more general interpretation which will playa funda-
mental role in the sequel lies in interpreting the right. hand side of (2.2.7*) as
the limit as N -00, when it exists, of the arithmetic means of the first N + 1
symmetric partial sums. This arithmetic mean is expressible in the form
L (1 - ).n+1 1) >(n),
Inl .. N
and this process of attaching a generalized sum to the series in (2.2.7*) is

known as summation by Cesaro means of the first order. As applied to Fourier
series, the method will be studied in some detail in Chapters 5 and 6. Yet
other summability methods are known to be useful and effective, though
we shall not dwell on them to any length in this book (see Section 6.6).
2.2.3. The Ortho~onality Relations. It is interesting to note at this

point that the orthogonality relations (1.1.3), which have been seen to be at
the basis of the formation of Fourier series, flow inevitably from the defining
properties of the invariant integral.
Suppose here that G is any compact topological group and that I is that
left invariant integral on G for which 1(1) = 1. Consider any nonprincipal
continuous (and therefore bounded) character X of G and choose any a E G
suoh that X( -a) # 1. Then, by (2.2.3) and property (2) of 2.2.2,
I(X) = I(TaX) = I[X( -a) X] = X( -a) . l(x)
showing that I(X) = O. Applying this to the product X = Xl X2 of two con-

tinuous characters Xl and Xll, we obtain the orthogonality relations
if Xl = X2
(2.2.12)
otherwise.
In view of (2.2.6) and (2.2.7), these relations reduce, when G = T, to the

relations (1.1.3), which are now seen in their true relationship to the group
structure of T.
There are other orthogonality relations pertaining to irreducible unitary
representations of compact topological groups that reduce to (2.2.12) when
the representations are one-dimensional (see 2.2.1); to discuss these would
take us too far afield, and is in any case irrelevant to our main theme.
26 GROUP STRUCTTJRE AND FOURIER SERIES
The dual orthogonality relations. In view of (2.2.7*), any would-be

orthogonality relations for the discrete group Z would read somewhat as follows:
Le
nEZ
lnx eln~ =1 or 0
according as the real numbers x and yare, or are not, congruent modulo 21T.
There is, however, no way of making sense of this relation which is based upon
applying a summability method to the series on the left for individual values
of x and y. On using concepts to be introduced in Chapter 12, it is nevertheless
the case that, for a fixed y, the series converges distributionally to a certain
distribution (or generalized function) known as the Dirac measure placed at the
point y. This latter entity does, in a sense, vanish on the open set of points
x ;f:. y modulo 21T, but there is no reasonable way of attaching to it a numerical
value at points x == y modulo 21T.
There is therefore a residual and irreducible asymmetry separating the
mutually dual situations; this is, in the last analysis, because of the profound
topological differences between the "smooth" compact group G and the
discrete noncompact group Z.
2.2.4. V' and Other Function Spaces. It has been remarked in 2.2.2
that the invariant integral can in all cases be extended to functions more
general than those in Cc(G). For G = T, in which case the invariant integral
has been identified in (2.2.7), the extension involved is that from the Riemann
to the Lebesgue integral; for the dual group Z, several stages in the extension
have already been mentioned at the end of 2.2.2. It is essential for a smooth
and satisfactory development of Fourier theory that advantage be taken of
this extension. Broadly and figuratively speaking, the Lebesgue theory of
integration is that which is necessary and sufficient for the major portion of
contemporary analysis; integration theories for functions on more general
sets and spaces almost invariably share the characteristic basic properties of
the Lebesgue theory. However, in certain special connections involving
functions of a real variable, more elaborate theories have proved useful. We
shall have neither occasion nor space for more than a passing reference in
12.8.2(3) to some such theories. (Others, mainly designed to handle integration
strictly as an antiderivation process, will receive no mention at all in this
book.)
We shall therefore assume that the reader is familiar with the definition
and basic properties of the Lebesgue integral of a function of one real
variable. With but relatively few exceptions, some of which are dealt with in
Appendix C, all the results we shall need will be found in the brief account in
[W]. For the exceptional points the reader is referred to [HS], [AB], or [E],
Chapter 4, or to anyone of the several excellent accounts of integration
theory now available. In making use of these sources of results about the
Lebesgue integral of functions of one real variable, it will be agreed that a
functionf on T is measurable (or integrable) if the associated periodic function

fop is Lebesgue measurable (or integrable) over some one-and hence
every-interval of length 21T.
Having reached this stage we shall drop the notational distinction between
f and fop-in other words, we shall not distinguish between a periodic
function on R and the corresponding function on the circle group.
It will be convenient to introduce some notations for the function spaces
that will appear constantly in the following pages.
If k is an integer,k ~ 0, CIc = Ck(T) will denote the set of complex-valued
functions with period 21T and with k continuous derivatives, and C'" =
C"'(T) = n {Ck: k = 1,2, ... }. For brevity, C is written in place of Co.
For any real number p > 0 we denote by LP = LP( T) the set of periodic
complex-valued measurable functions f such that
(2.2.13)
is finite, the integral being extended over any interval of length 21T; compare
[W], p. 68, [AB], p. 215, or [HS], p. 188. In addition, L'" = L"'(T) denotes the
set of essentially bounded periodic complex-valued measurable functions,
that is, of periodic complex-valued measurable functions f for which
IIfll", == ess. sup If(x)1 (2.2.14)
is finite, the essential supremum being taken relative to any interval of x-
values of length 21T.
To be perfectly accurate, we shall frequently use V' (0 < p ~ 00) to
denote the set of equivalence cla8se8 of the appropriate type, two functions
going into the same class if and only if they agree almost everywhere (a.e.).
Since we shall not always signal which viewpoint is being adopted, the reader
is warned to be on his guard and to be prepared to devote a little thought to
deciding which interpretation is appropriate. The Fourier series of a function
depends only on the class determined by that function.
Each of Ck (k an integer ~O, or 00) and V' (0 < p ~ 00) is a linear space;
in view of preceding remarks, the reader should check the truth of this when
LP is regarded as a set of equivalence classes.
When 1 ~ p ~ 00, I lip is a norm on LP if the latter is considered as a. set
of equivalence classes of functions (but only a seminorm if LP is viewed as a
set of individual functions); see Appendix B.1.2 for an explanation of the
terminology. This statement is virtually the content of Minkowski 's inequality,
which asserts that f + g E LP and
(2.2.15)
whenever 1 ~ p ~ 00 and f, g E LP. For a proof of Minkowski's inequality,
see p. 68 of[W], or p. 146 of [HLP], or Section 4.11 of [E], or [AB], p. 218, or
[HS], pp. 191-192. The assertion is false if 0 < p < 1 (see [HLP], loco cit.),
but it is then true that Ilf - gil: is a metric on Lp qua set of equivalence
classes (or a semimetric if 1P is considered as a set of individual functions).
For 0 < p ~ 00, 1P is complete for the metric Ilf - gllp if p ~ 1, or for the
metric Ilf - gil: if 0 < p < 1; the former case is dealt with in [W], Theorem
4.5a, and the same argument adapts readily to the case 0 < p < 1; alterna-
tively, see [HS], p. 192, or [AB], p. 220.
To complete the picture, on C" (k an integer ~ 0) we introduce the norm
Ilfll(,,) = sup II DhJll", , (2.2.16)
O"i h"i "
here and subsequently D is the symbol of derivation.

On C'" we introduce the metric Ilf - gll(",), where
<Xi 2-"llfll(k).
Ilfll(",) = ,,~o 1 + IIfll(,,)'
(2.2.17)
Despite the notation, Ilfll(",) is not a norm. Then Ck is complete for the metric
Ilf - gll(k) whenever k = 0,1,2, . ',00 (the reader should supply a proof of
this).
With their appropriate metric topologies, all these spaces are topological
linear spaces (see Appendix B.l.1), that is (compare 2.l.1 in relation to
topological groups), the linear space operations (j, g) -+ f - g and (A, f)
-+ A f (A a complex scalar) are continuous. Further details concerning C'"
appear in Section 12.l.
There will be constant use for one or more links in the chain of inclusion
relations
C'" c ... C C"+l C C" c ... c Co = C c L'" c LP C Lq, (2.2.18)
where k is an integer ~O and where 00 > p > q > O. What is more, each
inclusion map of one term of this sequence into any other lying to its right
is continuous. The only nontrivial portion of this last assertion depends on
the inequality
O<q<p~oo, (2.2.19)
The estimate (2.2.19) is itself a consequence of Holder's inequality, which
asserts that if 1 ~ p ~ 00, and if p' denotes the conjugate exponent (or index)
defined by IIp + IIp' = 1 (supplemented by the convention that p' = 00
if p = 1 and p' = 1 if P = 00), then f . g E Ll and
(2.2.20)
whenever f E Lp and g E Lp. A proof of Holder's inequality will be found on
pp. 72-73 of [Ka], the assumption there made concerning continuity of f
and g being unnecessary; see also Section 4.11 of [E], or [HS], pp. 190-191, or
[AB], p. 217. An extended discussion of both the Minkowski and Holder
inequalities is undertaken in Chapter VI of [HLP], but this is unnecessarily

elaborate for our purposes.
Each of the spaces V' (0 < p ::;; 00) and Ck (k = 0,1,2", ',00) is trans-
lation-invariant, as also are the appropriate metrics or norms. If E denotes
anyone of these topological linear spaces other than L"", and if 1 E E, the
mapping a ~ Tal is continuous from R (or from R/2rr Z) into E. (For the
case E = Ll a proof appears on p. 67 of [WJ; this proof is readily adaptable
to the case E = LP whenever 0 < p < 00. In the remaining admissible
cases the result is almost evident in view of the well-known result that a
continuous complex-valued function on a compact metric space is uniformly
continuous. Regarding the excluded case p = 00, see Exercise 3.5.) In all
cases the mapping 1 ~ Tal is, for any fixed a E R, a continuous endomorphism
of E; moreover
if 1 E V' and 0 < p ::;; 00, and

IITaIIICk) = IIJllck)
if 1 E C" and k = 0, 1,2, . " 00.
Convergence in the sense of the metric on LP (0 < p ::;; 00) will be termed
convergence in V' or convergence in mean with index (or exponent) p. We note
also that convergence in C in the sense of the norm I II"" is equivalent to
uniform convergence.
2.2.5. The Dual Concepts. In view of (2.2.7*), the natural analogues,

for the group Z, of the spaces LP introduced above, are the spaces IP = I"(Z)
of complex-valued functions", on Z such that
11"'111' == { 2: 1",(n)JPp'"
neZ
ifO<p<oo,
or
11"'11", == sup
nFZ
''''(n)1 if p = 00,
is finite.
In addition to these, we occasionally wish to refer to the subspace Co = co(Z)
of I""(Z) formed of those", for which
lim "'(n) = O.
Inl-""
Each of Co and II' (1 ::;; p ::;; 00) is a Banach space; if 0 < p < 1, II' is a
complete metric space.
In lieu of (2.2.18) and (2.2.19) one has the relations
(2.2.18*)
and
(2.2.19*)
for 0 < q < p < 00. (Notice that 11"'l q ::;; 1 implies 1",(n)1 ::;; 1 for all nEZ,
hence 1"'(n)IP ::;; 1",(n)lq for all nEZ, hence L 1""1' ::;; L ''''Iq ::;; 1.)
The Holder and Minkowski inequalities suffer no change in form other than
the obvious replacement of integrals by the appropriate sums. Proofs for
the case of finite sums appear on pp. 67-72 of [Ka]; for our purposes, which
involve infinite sums, transparent limiting processes constitute the final step;
see [HS], p. 194. A much more elaborate account appears in Chapter II of
[HLP].
Concerning notation, we shall sometimes denote a function cP on Z in the
sequential form: (CPn)nez; this is sometimes a convenience and is in any case
in accord with tradition. There is, however, a nonvanishing chance of con-
fusion with the convention according to which (CPn)"e z might also denote a
two-way infinite sequence of functions on Z. The context will in all cases
dispel initial doubts on this score.
2.3 Fourier Coefficients and Their Elementary Properties

Except in the discussion of certain specific examples, we shall use
systematically the so-called "complex" Fourier coefficients. Indeed, the
substance of Sections 2.1 and 2.2 constitutes ample indication that the
exponentials elnX playa much more fundamental role than do their real and
imaginary parts separately. Not until Chapter 12 is reached shall we consider
the Fourier coefficients and Fourier series of anything more general than
integrable functions.
For JELl we adopt in general work the systematic notation
j(n) = ..!..JJ(x)e- lnX dx for all n E Z (2.3.1)

21T
for the (complex) Fourier coefficients ofJ. The integral in (2.3.1) extends over
any interval of length 21T. The symbol j naturally denotes the function
n-'o j(n) defined on Z; it is a two-way infinite sequence. Throughout this
section we shall establish some of the simplest properties of the Fourier
transformation J -'0 j. In order to a void confusion with the Fourier series of
measures and distributions introduced in Chapter 12, a series of the type
L.nez j(n)e 1nX with JELl will be termed a Fourier-Lebesgue series.
The reader will notice that (2.3.1) makes no sense for a general J E V', if
o < p < 1. At no time shall we contemplate in detail any such extension,
though the methods of Chapter 12 would permit us to make one sort of
extension to restricted nonintegrable functions; see the example in 12.5.8.
Before beginning to display the elementary properties of the Fourier
transformation, we introduce the following notations:
The symbol D will be that of derivation, as applied to functions of a real
variable. There will be no occasion, until we reach Chapter 12, to apply D to
any functions that are not absolutely continuous; for absolutely continuous
functions, the new interpretation of D introduced in Chapter 12 is in agree-
[2.3] FOURIER COEFFICIENTS, THEIR ELEMENTARY PROPERTIES 31
ment with the traditional one referred to here. For any complex-valued
function I, J denotes the complex-conjugate function. For any function I
defined on any group (on R, T, or Z in particular), j denotes the function
t---+/(-t), andf* the function t---+/(-t); thusf* = (J)~ = (j)-. Accord-
ingly D and I ---+ I are linear, whereas I ---+ J and I ---+ f* are conjugate-linear.
2.3.1. Themappingj---+jislinear.Moreover, (J),,' = (/)*and(f*),,' = (/)-.

Proof. The first statement is clear, integration being a linear process on
the integrand. Of the remaining two assertions it will suffice to indicate the
proof of the first, thus
for all n E Z.
Note: In 3.1.1 and 4.1.2 we shall add some most important complements
to the first assertion in 2.3.1.
2.3.2. For each I ELI and nEZ, 1/(n)1 ~ 11/111'

Proof. By [W], Theorem 3Ac, we have
.
I/(n)1 1
~ 21T f I/(x)e-I""'1 dx 1
= 21T f I/(x)1 dx
Illk
==
Note: If we write II/II", = sup {1/(n)1 : nEZ}, 2.3.2 is equivalent to the
inequality II/II", ~ 1IIIk This estimate, although well-nigh trivial, is the best
possible in the sense that /(0) = IIIIII whenever I is real and nonnegative.
On the other hand, for general real- or complex-valued functions I the
relationship between II/II", and 11/111 is complicated; see Exercise 8.8 and
Subsections 11.3.1 and 11.4.14. In particular, there exist functions IE LI
for which IIIIII > 0 and the ratio 11/11",/11/111 is arbitrarily small. Hosts of
examples of this phenomenon can be constructed by using the results of
Chapter 15. A simpler example is provided by the so-called Dirichlet kernel
( ) -_ L
D NX '" I"" _ sin (N + 1/2)x.
e -. ,
I"I"N smx/2
if I = D N , llill", = 1 and yet 11/111 is (as will be seen in 5.1.1) asymptotic to

(4/1T 2 ) log N as N ---+00, so that the said ratio is in this case 1T2 J(410g N) and
can be made as small as we wish by choosing N sufficiently large. Moreover,
it has been shown (D. J. Newman [2]) that for each positive integer N there
exists a trigonometric polynomial
2: elle
N
f(x) = 1nz
II~O
such that lenl = 1 (n = 0, 1, ... , N) (and therefore II/II., = 1) and IIflll

> e, where e is a suitable absolute constant.
Nl/2 -
Were this type of phenomenon absent, the theory of Fourier series would
be much simpler and much less intriguing than it in fact is.
2.3.3. (To.f)"'(n) = e-Ina/(n) for n E Z andf ELl.

Proof. It is easy to verify that, if g(x) = f(x)e- 'n %, then
T aU(x) = elno. To.f(x) e- In %.
Integrating this relation and using translation-invariance of the integral, we
f f
obtain
.
f(n) = 217
1 g(x) dx = 217
1 T aU(x) dx
= elno. 2~ f To.f(x) e- In % dx
= elno. (To.ft'(n) ,
which is equivalent to the stated result.
Note: On being asked for a proof of 2.3.3, the reader's first reaction might
be to apply the usual formula for change of variable in the integrals involved:
this procedure is, of course, perfectly legitimate. But we prefer to phrase the
device in terms of the characteristic in variance property of the integral
(see 2.2.2).
2.3.4. Suppose that f is absolutely continuous, and let Df denote any

integrable function equal almost everywhere to the derivative of f. Then
(Df)"(n) = in /(n) for all n E Z.
Proof. That the derivative off exists almost everywhere and is integrable
follows from [W], Section 6.3, Exercises 15 and 16 on p. 111, and Theorem
5.2g. The formula for partial integration ([W], Theorem 5.4a) then yields
(Df)"(n) = ~ (211 Df(x). e- In dx

217 Jo
%
= -1 [f(x) e-In%]~" + -1 1211 f(x) e- In % in dx

217 217 0
= /(n) in,
which completes the proof.
Remarks. All that is required ofDj is that the partial integration formula
f" Dj' U dx = - f" j Du dx
shall hold for all periodic, indefinitely differentiable functions u. This means
that the preceding interpretation of Dj for absolutely continuous functions f
will accord with the generalized concept of derivation introduced in Chapter
12. The reader is reminded, however, that the result is not generally true for
functions j possessing almost everywhere an integrable derivative: it is in
addition necessary that j be equal to the indefinite integral of this derivatwe,
which is ensured by (and indeed equivalent to) absolute continuity.
2.3.5. Suppose that I is absolutely continuous, and that its derivative Df

is equal almost everywhere to an absolutely continuous function. Then
j(n) = O(I/n2) as Inl-oo, so that the Fo~rier series of I is absolutely and
uniformly convergent.
Proof. The result expressed in 2.3.4 may now be applied to DI in place
of I, showing that
for n i= o. The desired majorization follows upon using 2.3.2.

Remarks. (1) Much stronger results will be noted in Section 10.6.
(2) On using 2.3.8 (which could quite well be established immediately
following 2.3.4), the 0 appearing in 2.3.5 could be replaced by o. Notice that
the hypotheses of 2.3.5 are amply fulfilled whenever I E C2.
(3) The hypotheses of 2.3.5 ensure that the Fourier series of I is indeed
convergent to I(x) at all points, though we are not in a position to prove this
just yet; see 2.4.3.
(4) In a similar way, 2.3.4 shows that (n) j = O(I/lnl) (and, with 2.3.8, that
j(n) = o(l/Inl)) whenever I is absolutely continuous. The next result asserts
that the former majorization is in fact true for any I of bounded variation.
For a periodic function I we define the total variation V(f) to be the
supremum of all sums
m
L I/(x
k-l
k) -/(Xk-lli
withrespecttoallsequences(xk)~aosuchthatxo < Xl < ... < xm ~ Xo +271.

Then I is of bounded variation if and only if V(f) < 00 in which case we
shall write IE BV; compare [W), p. 105; [HS), p. 266; and [AB), p. 256.
Evidently, in taking the supremum above one may assume that the X k fall
into any preassigned interval of length 21T.
2.3.6. If I is of bounded variation, then
. 1
In'/(n) I ~ 271 V(f) for all n E Z.
Proof. Granted a knowledge of Riemann-Stieltjes integrals ([HSJ,

Section 8; [ABJ, Chapter 8), one may write
/(n) = Lfn I(x) d [t~~:]

for n :F 0, and apply partial integration for such integrals. Since we do not
wish to make explicit use of properties of Riemann-Stieltjes integrals, we
shall adopt a more pedestrian approach.
Suppose first that I is continuous. Put, for n :F 0, Il(x) = rln:r/( -in). It is
then easy to verify that, given e > 0, one has for n :F
1
L
m
I/(n) - -2 I(x/c)[u(x/c) - ll(x/C-l)JI ~ e
7T /c= 1

for all sufficiently fine partitions = Xo < Xl < ... < Xm = 27T of the interval
[0, 27TJ. Denoting by L: the sum appearing above, and applying partial
summation, we obtain
L = [f(27T) - L [f(X/C+l) -
m-l
l(xl)JIl(O) - l(x/c)JIl(x/c).
/C=l
By continuity (and periodicity) of I, the first summand on the right will not
exceed e in absolute value, provided the partition is sufficiently fine. Thus
1 m-l
1/(n)1 ~ e + 2: + 27T /C~l II(x/C+l) - I(x/c) I IIl(x/c) I
~ (1 + 27T1) e
1
+ 27T V(f) Tni'
1
since IIl(x)1 ~ l/Inl. Letting e ~ 0, we obtain
1/(n)1 ~ V(f)
27Tlnl
for n :F 0, which is equivalent to the stated result.
Suppose finally that I is merely of bounded variation. We shall obtain the
desired result in this case by approximating! by a suitable sequence of
continuous functions Ir of bounded variation. Perhaps the simplest choice is
I,(x) = f
r:r
:r+(1/')
I(t) dt = r
(lir
Jo I(x + t) dt.
i
Whatever the increasing finite sequence (x/c) of points of [0, 27T], one has
ll,
L/c I!r(X/c) - Ir(x/C-l)1 ~ r
0
L/c I!(x/c + t) - !(X/C-l + t)1 dt,
which, since the integrand never exceeds V(f), is majorized by V(f). Thus
V(f,) ~ V(f) for all r. By what is already established, therefore, we have
Ii, (n)1 ~ V(fr) ~ V(f) (2.3.2)
r 27Tl n l '" 27Tl n l
for n :F 0 and all r.
Beside this, a simple computation shows that
J, n
A ( )
= exp (in) ) sin (n/2r)
- . fA( n'
r 2r n/2r
so that
lim Ilr(n)! = Il(n)!.
r-oo
Combining this with (2.3.2), we are led to

A V(f)
If(n)1 ~ 21Tlnl
for n #- 0, and the proof is complete.

Remarks. (1) The converse of 2.3.6 is false: there exist continuous
functions f for which len) = o(I/lnj) as Inl- 00 and yet f BV.
(2) The estimate in 2.3.6, namely, thatl(n) = O(I/In!) as Inl-oo, cannot
be improved, even if it be assumed that I is continuous as well as being of
bounded variation. In other words, there exist continuous functions f of
bounded variation such that len) #- o(I/ln!) as Inl-oo. For a proof, see
[Ba 1 ], pp. 210-211; or Exercise 12.44. In view of 2.3.4 and 2.3.8, any such
function I fails to be absolutely continuous.
Incidentally, it is known (after Wiener) that a function I of bounded
variation is continuous if and only if
l",
lim N ~ Inf(n)!
A
= 0;
N_oo I"I"N
see Exercise 8.13.

(3) At the expense of replacing the factor (21T) -1 by 1 in 2.3.6,. there is a
very neat proof due to Taibleson [1]. Thus, if n E Z and n #- 0, write ak =
2k1T1 nl- 1 for k E {O, 1,2, .. " Inl}. Denote by g the step function which is
equal to I(ak) on (a k - 1 , a k) for k E {l, ... , In!}. Then, since
i 2l1(k+llll1l- 1
2kl1l"l-l
e-I":t dx = 0,
it follows that
21Tll(n) I = If" (I(x) - g(x e-I":t dx I

where V" is the total variation off on the interval [ak _ 1> akJ. Since V 1 + ...
+ V" ~ V(f) and a" - ak-l = 27Tlnl-1,
27Tlj(nli ~ V(f).27Tl nl- 1
and hence
Inj(nli ~ V(f).
See also M. and S.I. Izumi [1].
2.3.7. Define the mean modulus of continuity of f with exponent (or index)
1 by
Then, if fELl,
.
If(n)1 1 wd (TT)
~ 2 n (n E Z, n - 0).
Proof. By definition
.
f(n) = 2TT
1 If(x)e-1,,;r dx,
and by 2.3.3
Subtracting and dividing by two, we obtain

j(n) = %(f - T -nlnf)"'(n) ,
whence the result follows on applying 2.3.2.
2.3.8. (Riemann-Lebesgue lemma) For any integrable f one has

lim j(n) = O.
1"1_ 00
Proof. This follows immediately from 2.3.7 and the fact ([W], Theorem
4.3c) that wd(a) -+ 0 as a -+ O.
Remarks. (1) The Riemann-Lebesgue lemma is so fundamental that it
is worth pointing out another method of proof (which indeed lies behind the
proof of Williamson's Theorem 4.3c just cited and used). Suppose we denote
by E the set of integrable functions f for which the statement of the lemma
holds. Then 2.3.2 shows that E is a closed subset ofLl (relative to the topology
defined by the norm II Ill)' It is otherwise evident that E is a linear subspace
of Ll. To prove the lemma it therefor~ suffices to show that E contains a.
set of functions, say S, the finite linear combinations of which are dense in U.
There are many such sets S which may be indicated. Examples are: (i) the
set of characteristic functions of intervals [a, b] (0 < a < b < 271"), extended
by periodicity. The finite linear combinations of these are dense in Ll (as is
shown in [W], Theorem 4.3a); and each such function is directly verifiable
to have a Fourier transform satisfying the lemma (assurance on this point
also comes from 2.3.6). (ii) the set C"'; see [W], Theorem 4.3b and 2.3.4.
(2) It is worth pointing out that 2.3.4 to 2.3.8 are all essentially concerned
with restrictions on the rate of decay ofi(n) as Inl ~OO. The indications are
clearly that the smoother the function j, the more rapid this decay. This
conclusion will receive further reinforcement as we progress; some extreme
instances are covered by Exercises 2.7 and 2.8.
2.3.9. Introduction of A(Z). The preceding results and remarks might raise
hopes that the membership ofjto various function spaces (such as C or V' for
various values of p) might be decidable solely by inspection of the rate of
decay ofi, or at any rate by examining Iii. However, while there are many
criteria of this sort that are either sufficient or necessary, with the sole
exception of the case of L2 (dealt with in Chapter 8), there are no known
necessary and sufficient conditions of this type. Moreover, it will appear in
Chapters 12 and 14 that there definitely cannot be any such complete
characterization involving only the values of Iil. The few necessary and
sufficient conditions that are known are of a much more complicated sort
and are unfortunately extremely difficult to apply in specific instances; see
2.3.10. Much remains to be discovered in this direction.
To make things more specific, let us consider Ll itself. If we denote by
co( Z) the linear space of complex- valued functions (two-way infinite sequences)
<fo on Z for which lim,n, ... ",<fo(n) = 0 and equip it with the norm
11<fo11", = sup {I<fo(n)1 : n E Z} (2.3.3)

i
(see 2.2.5), we have learned so far that j ~ is a continuous linear mapping
of Ll into co(Z). Denote by A(Z) the range of this mapping. The question
is: Given in advance a <fo E co(Z), how ca'n one determine whether or not
<fo E A(Z)? No effective and general method is known for doing this.
i
Although we know that tends to zero at infinity for each JELl, the rate
of decay can be arbitrarily slow. For example, given any <fo E co(Z), one may
choose positive integers Nl < N2 < ... so that
for
Then
2: 2<fo(Nk) cos Nkx

00
j(x) =
k=l
is a continuous function for which

for n = N" (k = 1,2, .). Furthermore we shall see in Section 7.4 that
(again for any assigned q, E co(Z a function IE Ll can be chosen so
that
l(n) ~ Iq,(n)l
for all n E Z.
Again, although the sequence q, defined by
for Inl ~ 2
otherwise
belongs to A(Z), the sequence q,l defined by
for Inl ~ 2
otherwise
has not this property (see Exercise 7.7 and 10.1.6). This shows that an orderly,
and therefore seemingly harmless, change of sign can destroy membership
of A(Z).
This (or any other similar) example shows incidentally that A(Z) is a
proper subset of co(Z); it also shows that 1<p1 may belong to A(Z) while <p fails
to do so. There is an entirely different, and tYPically modern, approach to
the proper inclusion relation A( Z) ~ co( Z) which shows a little more,
namely, that A(Z) is in fact a meager (that is, first category; see Appendix
A.l) subset of co(Z).
To see this, we must observe that Ll and co(Z) are Banach spaces when
endowed with the norms defined in (2.2.13) and (2.3.3), respectively, and
that T: 1-1 is a continuous linear operator mapping Ll into co(Z) whose
range is A(Z) (see 2.3.1 and 2.3.2). If, contrary to our assertion, A(Z) were
nonmeager in co(Z), the open mapping theorem (Appendix B.3.2) would
entail that T is an open map of Ll onto co(Z). Assuming for the moment the
uniqueness theorem (2.4.1), this would imply the existence of a number
B > 0 such that
11/111 ~ B' 11111.., (2.3.4)
for eachl ELl. However, (2.3.4) can be negatived, the resulting contradiction
thus establishing our assertion. For example, if I = D N , as in Exercise 1.1, a
direct computation, which will be carried out in detail in 5.1.1, shows that
11/111 == IIDNlll "" .".42 log N
for large N. Since in this case 11111.., = 1, it is plain that a contradiction of

(2.3.4) results whenever N is sufficiently large.
[2.3] FOURIER COEFFICIENTS, THEIR ELEME~TARY PROPERTIES 39
At this point see also Exercise 9.8.

In addition to the linear space structure of co(Z), one may consider its
structure as an algebra under pointwise operations. It is then natural to ask
whether A(Z) is a subalgebra (as well as a linear subspace) of co(Z). This
leads us to seek a way of combining integrable functions-a sort of multiplica-
tion-which corresponds to pointwise multiplication of their Fourier trans-
forms. We shall consider this question and its ramifications in Chapter 3,
returning in Chapter 4 to the consideration of A(Z) and the Fourier trans-
formation in this enriched setting. Further results will appear in 11.4.13 and
11.4.16.
2.3.10. Criteria for Membership of A(Z). Simply as a matter of interest

(for we shall make no subsequent use of these results), we sample a few of the
known criteria for a given sequence 4> = (4)n)nez to be the sequence of Fourier
coefficients of a function belonging to certain prescribed function spaces.
Further results of this sort will appear in Section 8.7, 10.6.3(1), and (2),
12.7.5, 12.7.6, 12.7.9(2), and Exercise 12.50. If the reader will ponder these
conditions, he will soon be convinced of the difficulty of applying them in
specific instances.
(1) In order that rP shall belong to A(Z), it is necessary and sufficient that,
having chosen any p satisfying 0 nUT(n)
r-+oo nez
= 0
for any sequence (u T );"., 1 of trigonometric polynomials satisfying
lim IluTllp = o.
T- '"
An equivalent condition is that to each e > 0 shall correspond a number

k(e) ;;il: 0 such that
I 2: rPnu(n)1 ~ e' Iluil", + k(e)' [Iullp (2.3.5)

nez
for all trigonometric polynomials u; see R. E. Edwards [1] and Ryan [1].
The case p = 2 of this result is due to Sall'm; sec [Bad, pp; 239-240.
(2) In order that 4> shall be the sequence of Fourier coefficients of a con-
tinuous function, it is necessary and sufficient that to each e > 0 shall corre-
spond a number k(e) ;;il: 0 and a finite subset F. of Z such that
I L rPnu(n) I ~ e' Ilulll + k(e)' sup lu(n)1 (2.3.6)

nez neF
for all trigonometric polynomials u; see R. E. Edwards [1].

(3) It will appear in 13.5.1 that, if rP = j for some f E LP, where 1 E t P', that is,
(2.3.7)
p' being defined by lip + lip' = 1. It is known (Rooney [1]) that if cfo satisfies
(2.3.7), then it is the sequence of Fourier coefficients of some function in
V(l 2) if and only if

suP. (v + l)P-1 2:
maO
IM . m(4)W < 00, (2.3.8)
the supremum being taken as 1/ ranges over all nonnegative integers, where
and
an,Y.m = YCm i1o
tm(l- t)-m e n1nl dt
for nEZ, II E Z, m E Z, v ;l!: 0, and 0 E> m E> v.

(4) A further sufficient condition for membership of A(Z) will be discussed
at some length in Section 8.7.
(5) The behavior of A(Z) wider permutations of Z is discussed by Kahane
[4].
2.4 The Uniqueness Theorem and the Density of Trigonometric

Polynomials
In this section we shall establish the uniqueness theorem, which asserts
that a function is determined almost everywhere by its Fourier transform,
and certain consequences thereof concerning approximation by trigonometric
polynomials.
2.4.1. {I} HfE e and/ = 0, thenf = O.

{2} HfE Ll and/ = 0, thenf = 0 a.e.
Proof. Statement (I) is, of course, a special case of (2). We shall prove it
first and deduce statement (2) from it.
By 2.3.1, we may in all cases assume that f is real valued. Moreover, since
/ = 0 entails (Taf}"" = 0 for all a {by 2.3.3}, it will suffice to show that if
fEe and
~fftdX=O
271.
for all trigonometric polynomials t, (2.4.1)
then
f(O) = O. (2.4.2)
We will in fact show that the negation of {2.4.2} implies the negation of(2.4.1}.
H (2.4.2) is false, we may (by changing f into - f if necessary) assume that
f(O} = c > 0 and then choose 8 > 0 so that
f(x} ~ Yzc for Ixl ~ 8 (2.4.3)
[2.4] THE UNIQUENESS THEOREM, THE DENSITY OF POLYNOMIALS 41
To construct a trigonometric polynomial t violating (2.4.1), write

to(x) = 1 + cos x - cos 8
and then t = toN, where the large positive integer N will be chosen later. It
is plain that t is indeed a trigonometric polynomial. It is also clear that
It(x)1 ~ 1 for 8 ~ Ixl ~ 17, t(x) ~ 0
(2.4.4)
t(x) ~ qN for Ixl ~ %8,
where q = 1 + cos %8 - cos 8 > 1. By (2.4.3) and (2.4.4) we have
2117 fit dx ~ 21 Jr
17 Ixl-' 6
It dx - 21 11/11> (217 - 28)
17
~ Yz c
12 r
17 Jlxl-.6
t dx - 11111>
~ 4c
17
r
Jlxl-. %6
t dx - 11/11>
C
~ 4TT. 8 qN - 11/11>,
which is positive provided we choose Nlarger than log (417 11/11",/c8)/log q, thus
negating (2.4.1). This proves statement (1).
Now assume that I is as in statement (2). Define
F(x) = c + LX I(y) dy,

where the number c is chosen so as to make .1'(0) vanish. Since 1(0) = 0, F is
periodic. Now F is absolutely continuous and DF = I a.e. ([W], Theorem
5.2g). By 2.3.4, the choice of c, and the main hypothesis on I, it follows that
.1' = O. Thus statement (1) shows that F = 0, and so 1= 0 a.e., as alleged.
Remark. The uniqueness theorem for trigonometric polynomialB is a
direct consequence of the orthogonality relations (1.1.3) and is covered by
Exercise 1.7 (1).
We proceed to deduce from 2.4.1 two rather special and purely provisional
results concerning the recapture of a function from its Fourier series; more
satisfactory results of this nature will appear in Chapters 6 and 10. As was
pointed out in 2.2.1, these results concern harmonic synthesis on the circle
group.
2.4.2. If IE Ll has a Fourier series that is dominatedly convergent almost

everywhere, then
I(x) = L: j(n)e
neZ
1nX a.e.
Proof. Let g be defined almost everywhere to be the sum of the Fourier

series of I wherever the latter converges, and to be, say, 0 elsewhere. By
dominated convergence, g ELl. By the same token, one has for any m E Z,
.
g(m) = -If g(x)e- lmx dx = L.
"".f(n) - If e1nx e- lmx dx
217 n"Z 217
=j(m),
the last step by the orthogonality relations. By the uniqueness theorem

(2.4.1), f = g a.e.
2.4.3. If f is a continuous function with a uniformly convergent Fourier

series, then
f(x) = j(n)e1n% 2:
n"Z
everywhere.
Proof. Using the notation of the proof of 2.4.2, the sum function g is
now everywhere defined and is continuous thanks to uniform convergence of
the series. Also, since the range of integration involved is a bounded interval,
uniform convergence entails dominated convergence. So 2.4.2 entails that f
and g agree almost everywhere. But, since both are continuous, this in tum
implies agreement everywhere.
A further and very important deduction from 2.4.1 is the following density
theorem; it too will be refined later.
2.4.4. The set T of all trigonometric polynomials is everywhere dense in

each ofthe Banach spaces C, V' (1 ::;; p < (0), that is, givenf E C (respectively
f E LP) and e > 0, there exists t E T such that
IIf - til"" ::;; e (respectively Ilf - tll p ::;; e). (2.4.5)
Proof. (1) First take the case of C. Given e > 0, first choose g E C2 such
that
Ilf - gil"" ::;; Y2e. (2.4.6)
This may be done by choosing a sufficiently small positive a and setting
gl(X) = a- 1 r+ 1l
f(y) dy, g(x)=a-1J%
(%+11
gl(y)dy.
By 2.3.5 and 2.4.3,
the series being convergent in C (that is, uniformly convergent). One may
therefore choose N so large that
IIg - 2:
Inl .. N
g(n)enll", ::;; Y2e. (2.4.7)
[2.5] REMARKS ON THE DUAL PROBLEMS 43
Then (2.4.6) and (2.4.7) combine to yield (2.4.5) with
t = L
l"I<iN
g(n)e".
(2) This case follows from (I), the fact that C is everywhere dense in V'
(compare [W], Theorem 4.4e), and the inequality Ilk II" ~ Ilkllco. Thus, given
! E V' and e > 0, first choose gEe so that II! - gil" ~ Y2e, and then [by (I)] a
t E T so that Ilg - tll co ~ Y2e. Then, a fortiori, Ilg - til" ~ Y2e and so
Remarks. (I) The assertion in 2.4.4 is false for p = 00. (Why?)

(2) It is possible to deduce 2.4.4 from 2.4.1 via the Hahn-Banach theorem
(see Appendix B.5) and results about the topological duals of the spaces C
and L" given in Chapter 12 and Appendix C, respectively. One of the con-
sequences of the Hahn-Banach theorem is, in fact, that uniqueness theorems
and density theorems consort in "dual pairs," so to speak.
(3) When combined with 2.3.2, 2.4.4 leads to an independent proof of
2.3.8; see Exercise 2.9.
(4) Assuming that part of 2.4.4 which refers to the space C, one can derive
2.4.1; see Exercise 2.10. Thus, 2.4.1 and 2.4.4 are equivalent, a fact which
illustrates the substance of Remark (2) immediately above.
(5) Other proofs of2.4.1 and 2.4.4 will appear in 5.1.2 and6.I.I, where more
refined versions of 2.4.4 are considered; see also Section 6.2, where some
applications are mentioned.
2.5 Remarks on the Dual Problems

2.5.1. Definition of the Fourier Transform. If we are given a function
<pon Z, it is naturalto attempt to define its Fourier transform ~ as the function
on T given by
(2.5.1)
in comparing this with (2.3.1), the reader will observe a change from e- inx to
ei"x, which is made purely on the grounds of subsequent convenience.
Although (2.5.1) makes excellent sense whenever <p E (l(Z), in which case ~ is
evidently a continuous function on T satisfying
(2.5.2)
(compare 2.3.2), it is plain that complications arise if, for example, 1. (In the case of the group T
no analogous complications appeared, because of the compactness of T.)
One has in fact to contemplate conditional convergence and summability,
perhaps merely for almost all x, as was heralded by the remarks at the end of
2.2.2 concerning the interpretation of the invariant integral on Z.
As a matter of fact, and as will appear in the course of Chapters 8, 12, and
13, it is often more effective to replace considerations of pointwise convergence
(everywhere or almost everywhere) of the series on the right of (2.5.1) by
that of convergence, either in one of the spaces Lq or distributionally, of the
symmetric partial sums
~N(X) == L .p(n)e1n %.
Inl"N
On the other hand, for functions .p which are structurally special, the results
of Chapter 7 yield pointwise convergence (at least almost everywhere) of the
series defining ~, although even here there is no assurance that the function ~
so defined almost everywhere will belong to Ll. Further very special results
of this sort, applying to cases in which .p is known to be of the form J for
some fELl, are contained in Chapters 5, 6, and 10.
2.5.2. The Uniqueness Theorem. Related difficulties arise in connection

with the appropriate uniqueness theorem, at least if pointwise convergence or
summabiIity is envisaged and unless severe a priori restrictions are imposed
upon .p. [For instance, there are no difficulties if we assume that the series
on the right of (2.5.1) is dominatedly convergent almost everywhere; but
this can only be the case if.p belongs to A(Z), a criterion extremely difficult
to verify at the outset.] If convergence is meant in the distributional sense,
the uniqueness theorem presents no difficulties and is implicit in results to
be obtained in Chapter 12. Dogged insistence on the pointwise interpretation
of convergence leads right to the heart of the Riemann-inspired theory of
general trigonometric series, including a number of problems (such as those
concerning the characterization of the so called sets of multiplicity and sets of
uniqueness) of great difficulty and whose delicacy is such that they "fall
through the mesh" imposed by requirements of distributional convergence or
convergence in some space Lq. Two relatively very simple results concerning
pointwise convergence appear in Exercises 2.13 and 2.14.
2.5.3. The Space A. Dual to A(Z) is A = A(T): this is the linear space of
functions on T of the form ~ obtained when .p ranges over {l(Z). Equiva-
lently, A consists precisely of those continuous functions f on T such that
JE {l. As in the case of A(Z), so with A: there is no complete solution to the
problem of characterizing directly in terms of their functional values the
elements of A. We shall return to the consideration of A in Sections 10.6 and
12.11, where partial results will be obtained and used.
45
2.5.4. The Dual of 2.4.4: Almost Periodicity. In view of the closing

remarks in 2.2.1 and Exercise 2.3, it is natural to attach the label trigonometric
polynomial on Z to each function on Z which is a finite linear combination of
characters e a : n ~ elan of Z, the parameter a ranging over T (or, equivalently,
over R).
This being so, the dual of 2.4.4 is concerned with the characterization of
those complex valued functions on Z that are uniform limits of trigonometric
polynomials on Z. The investigation of this problem cannot be undertaken
here. The functions which are approximable in the stated fashion are precisely
the socalled almost periodic functions on Z.
Almost periodicity is a concept applying to functions on arbitrary groups
and has behind it a vast literature. For the case of the group R, see [Bes]; for
more general groups, see [Mk]; [HR], Section 18; [We], Chapitre VII; [Lo],
Chapter VIII.
On a compact group, such as T, all continuous functions are almost periodic;
this fact explains the form taken by 2.4.4.
On the noncompact group Z, the only member of co(Z) that is almost periodic
is the zero element.
EXERCISES
2.1. Let S be a closed subgroup of R distinct from {OJ and R. Show that
there exists a number d > 0 such that S consists precisely of all integer
multiples of d.
Hint: Consider the infimum of all positive members of S.
2.2. Let x be a real number such that X/7T is irrational. Show that the set
{e lnz : n E Z} is everywhere dense in the unit circumference in the complex
plane.
Remarks. A stronger result will appear in Exercise 2.15. The stated result is
a special case of Kronecker's theorem, for which see [HW], 'Chapter XXIII,
especially p. 370. There is a general grouptheoretic formulation of this theorem
which is discussed in [HR], pp. 431-432, 435-436, and which asserts in particular
that any character X of R/21TZ, continuous or not, can be approximated
arbitrarily closely on any preassigned finite subset of R/21TZ by a suitably
chosen continuous character en'
Hint: Show tha.t el .,,: =1= el u if m, n E Z and m =1= n. Deduce that the said
set has I as a limiting point.
2.3. Let, be a bounded character of the group Z. Show that there exists
exactly one x E T such that {{n) = e lnz == {z{n) for all n E Z.
Assume that Z is endowed with its discrete topology; verify that the dual
topology on T (when it is regarded as the dual of Z) is identical with its initial
topology (as defined in 2.1.1).
2.4. Determine the continuous linear functionals I on C = C( T) which are

"relatively invariant" in t,he sense that there exists a function il on T such
that
I(Taf) = il(a)"I(f)
for all a E T and aIl/E C.
Hint: Show first that, if I '# 0, then il(a) = elY,, for some n E Z. Then
consider the functional J defined by J(f) = I(Ln "I).
2.5. Consider the finite group Z'" = ZjmZ, where m is a positive integer,
taken with the discrete topology. What are the invariant integrals on Z",?
What are the characters of Z",? Discuss the Fourier theory for this group.
2.6. Let t be a continuous homomorphism of R/2TTZ onto itself. Show that
there exists k E Z such that t maps the coset x + 2TTZ into the coset kx + 2TTZ.
Hint: Consider the character of T which carries x + 2TTZ into efn %,
so getting a homomorphism s of Z into itself for which e1nl(%+211Z) = els(n)(%+211Z).
Then examine s.
2.7. Suppose that I ELl. Show that
j(n) = O(e-Blnl)
for some B > 0 if and only ifI is equal almost everywhere to a function which
is analytic in a horizontal strip 11m zl < 8 for some 8 > O.
Hint: For the "if" part, apply Cauchy's theorem for a suitable rectangle
to the integral defining j(n).
2.8. Suppose I E Coo and pt.~
(k = 1,2, ... ).
Show that for n '# 0
Show also that if

Mk ~ const Rkr(ak + 1) (k = 1,2, .. ) (1)
for some R > 0, a > 0, then
Ij(n)1 ~ const In11/2a" exp [ _ ('~I) 1/a] (n E Z, n '# 0); (2)
and that if (2) holds, then

Mk ~ const Rlcr(ak + 1)(1 + k a - 1/2 ) (k = 1,2, .. ). (3)
Hint: Make use of Stirling's formula describing the behavior of r(z) for
large positive values of z.
Remarks. Functions f for which an inequality of the type (1) holds
(the constant possibly depending uponj) form the simplest types of what are
termed q'Ua8ianalytic cla8se8 of functions; the case a = 1 corresponds to the
analytic functions. The relationships between such classes and those defined by
means of inequalities involving the Fourier coefficients-like (2). for example---
EXERCISES 47
have been studied in great detail. See, for example, [M], especially pp. 78-79,
138-139.
2.9. Give a proof of 2.3.8 based solely on 2.4.4 and 2.3.2.
2.10. Derive 2.4.1 directly from that part of 2.4.4 which refers to the
space C.
2.11. Deduce from 2.4.1 the following uniqueness theorem: If! is defined
and integrable over (-7T, 7T), and if
f" !{X)XN dx =0 (N = 0, 1, 2, ... ),

then! = 0 a.e. on (-7T, 7T).
Note: On the basis of the HahnBanach theorem (see Appendix B.5), this
assertion implies results about approximation by ordinary polynomials akin to
Weierstrass' theorem (see 6.2.2).
2.12. Show that if! is defined and integrable over ( -7T, 7T), and if
f" !{x)e1aNo: dx =0 (N = 1,2, .. ),

where (aN)N'=l is a sequence of complex numbers having at least one (finite)
limiting point, then! = 0 a.e. on (-7T, 7T).
Hint: We may assume without loss of generality that aN =F 0, aN -r O.
Consider J~,,!{x)eIZX dx as a function of the complex variable z.
2.13. (Lusin.Denjoy theorem) Write the trigonometric series Lne z cne1no:
in its real form Y2a o + L:~l (an cos nx + bn sin nx), and suppose that
L L
~ ~
Icne inx + c_ne-inxi == laol + Ian cos nx + bn sin nxl

n=O n=l
converges for x E E, where E is measurable and the Lebesgue measure m{E)
of E is positive. Prove that
that is, that

2: (Ianl + IbnD <
~
ex).
nal
Hints: Assume without loss of generality that the an and bn are rea1-
valued and put an = Tn cos 8n , bn = Tn sin 8n , where r n ~ 0 and 8n is real.
Use Egorov's theorem to justify termwise integration of the series
2:
~
Tn ICOS (nx - 8n)1

n=l
over some set Eo with m{Eo) > 0, and observe that
f Icos {nx - 8n )1 dx ~ f cos 2 (nx - 8n ) dx

JEo JEo
= Y2m{Eo) + o{l).
Note: For a simple generalization of this result, see [KS], p. 84, TMoreme II.
The Lusin-Denjoy theorem has prompted numerous more elaborate investiga-
tions of the absolute convergence of trigonometric series: see [Zl], Chapter VI;
[Ba:!], Chapter IX; [KS], Chapitre VII.
2.14. (Cantor-Lebesgue theorem) As indicated at the end of 2.2.2, a

trigonometric series L:"e z e"e'"'' is said to converge for a particular value of x
if and only if
exists finitely for that value of x. Show that if this is true for each point x
belonging to a measurable set E having positive Lebesgue measure, then
liml"I_'" e" = O.
Hints: As in the hints for the preceding exercise, reduce the problem to
the case in which L:r" cos (nx - 8,,) is uniformly convergent for x E Eo, where
m(Eo) > O. Were the assertion to be false, there would exist integers
n l < n2 < . . . so that
cos (n,.x - 8"k) - 0
uniformly for x E Eo. Consider the integrals
f cos 2 (n,.x - 8"k) dx.

JEo
Notea: Cantor considered the case in which E is a nondegenerate interval.
Steinhaus produced examples of series L neZ Cn elnx for which Cn ~ 0 and yet
the series diverges everywhere. One such example is
L
00
(logn)-l.cosn(x -loglogn);
.. -3
see [Bal]' p. 176.
2.15. (Equidistributed sequences) Let Xo be real and such that xO/fT is

irrational. Suppose g is a periodic function with the property that to ea.ch
e > 0 correspond continuous periodic functions'll. and v such that 'II. ~ g ~ v
and
2~ f (v - '11.) dx < e.
Show that
L g(nxo) = g(O).
N
lim N-1 (I)
N ... ca ,,-1
Deduce that if I is a subinterval of [0, 2fT), then N-l times the number of
points x o, 2xo, ... , Nxo which belong modulo 2fT to I converges, as N _ 00,
to (2fT) -1 times the length of I, that is, that the points nxo (n = I, 2, ... )
are equidistributed modulo 2fT. Observe that this result implies that of
Exercise 2.2. See also [Ba 2 ], p. 473.
Hints: First prove (I) for continuous periodic g by using 2.4.4.
EXERCISES 49
Note8: The idea of equidistributed sequences is due to Wey!. Several results

will be found in [PS], Band I, pp. 70-74.
For extensions to general groups, see [HR], pp. 432, 437-438 and Rubel [1].
2.16. (Fejer's lemma) Suppose that 1 ~ P ~ 00, that IEV', and that
g E V", where lip + lip' = 1. Prove that
f
2Irr I(x)g(nx) dx -* !(O)g(O)
for n E Z and.lnl-*oo.
Hints: Assume first that p > 1 and use 2.4.4 to approximate g in Lp' by
trigonometric polynomials; then use 2.3.8. If p = 1, so that p' = 00, approxi-
mate I in L1 by continuous functions.
Note: The result actually remains true if the restriction n E Z is replaced
by n E R.
2.17. Let (e);:'z1 be any sequence of positive numbers converging to zero.

By extracting a suitable subsequence (E nk ) and considering the series
2:
co
Enk exp (in "x) ,
k=1
show that there exist continuous functions I such that
lim sup wd(rr/n) > O.

n- co en
Hint: Use 2.3.7.

2.18. Prove that any nonnegative continuous function I is the uniform
limit of functions Ig12, where g denotes a trigonometric polynomial.
Formulate and prove an analogous result for functions I in LP (1 ~ P < (0).
Hint: See Exercise l.ll.
2.19. Prove that, for any finite set FeZ and any e > 0, a trigonometric
polynomial I exists such that
o ~ !(n) ~ 1 for all n E Z, !(n) = 1 for all n E F,

111111 ~ 1 + e.
For which sets F does the result remain true when e = 01
Hints: Suppose r is a positive integer such that F c [-r, r] and choose a
large positive integer N. Consider
I(x) = [(2N + 1)-1 L

Inl .. N
e1nx] . [ L
Inl"N+r
e1nx ]
= u(x) v(x)
and use Exercise 1.7(1) and the Cauchy. Schwarz inequality.

CHAPTER 3
Convolutions of Functions
3.1 Definition and First Properties of Convolution

At the end of 2.3.9 we posed the problem of finding a binary o]?eration on
integrable functions that would correspond to pointwise multiplication of
their Fourier transforms. To attempt directness by trying to define the result,
say I * g, of applying this operation to functions I, g E Ll by requiring that
(f * gt' = j. g is not very effective, because we do not know how to charac-
terize A(Z) in such a way that it is clear that it is closed under pointwise
multiplication. A more useful clue is provided by the orthogonality relations
combined with the special properties of characters.
Suppose that we write en for the function x --+ etnx (n E Z). For m and n in
Z, the orthogonality relations (1.1.3) show that
1
21'1'
f em(x - y)en(y) dy = {en(X)
0
ifm = n
otherwise.
Accordingly, if we define 1* g by
1* g(x) = 211'1' f
I(x - y)g(y) dy, (3.1.1)
then it appears that em * en has as its Fourier transform the pointwise product
of the Fourier transforms em and ~n' Since each of 1* g and J. g is evidently
bilinear in the pair (j, g), the desired relation will obtain for functions I and g
which are trigonometric polynomials, that is, finite linear combinations of
the en' It thus appears that (3.1.1) constitutes a hopeful starting point. We
proceed to the details forthwith.
Suppose that I and g belong to Ll. Then the Fubini Tonelli theorem is
applicable (see [W], Theorems 4.2b, 4.2c, .and 4.2d; [HS], pp. 384-386, 396;
[AB], pp. 154-155) and shows that the integrand appearing on the right of
(3.1.1) is, for almost all x, an integrable function of y, so that (3.1.1) effectively
defines 1* g(x) for almost all x; moreover the function so defined almost
everywhere is measurable and
(3.1.2)
In particular, I * g ELI.
50
[3.1] DEFINITION AND FIRST PROPERTIES OF CONVOLUTION 51
From these arguments it also appears that
If * g(x)i ~ If I * Igi (x) a.e. (3.1.3)

As a consequence of invariance of the integral it appears that at any point
x for which f * g(x) exists, g * f(x) exists and has the same value. Thus
(3.1.4)
Let us now compute the Fourier coefficients of f * g, using en route the
Fubini-Tonelli theorem and the relation e - in;r = e - in(;r - II) e - inll:
(J * g)"'(n) I*
= 2~ f g(x)e-1n;r dx
= 2~ I I e- 1nz {211T f(X - y)g(y) dY} dx
= (2~r II
= (2~rI{I
e-1n;r f(x - y)g(y) d(x, y)
f(x - y)e- 1n(Z-1I) g(y)e- 1nll dX} dy
by two appeals to the Fubini-Tonelli theorem
= !1T I g(y)e- 1nll {2~ I f(x - y)e- 1n(Z-II) dX} dy
I
= 21T I .
g(y)e- 1nll {f(n)} dy
by translation invariance of the inner integral

= j(n)g(n).
Thus we have the desired relationship:
(J * g)"'(n) = j(n) g(n) for all n E Z. (3.1.5)
Convolution is associative, that is, (J * g) * h = f * (g * h) for f, g, hELl. A
direct verification is possible, using the Fubini-Tonelli theorem. Alt.ernatively,
one may appeal to (3.1.5), to the uniqueness theorem of 2.4.1, and to the
evident associativity of pointwise multiplication of the transformsj, g, and Ii.
Remarks. The definition and the above properties of convolution may
be formulated and established in another way, thereby making appeal to no
more than the most primitive form of the Fubini theorem applying to
continuous integrands (which somebody aptly christened "Fubinito").
One begins by defining f * g for f, gEe by (3.1.1). Fubinito then yields
(3.1.2) to (3.1.5) for such! and g. It is furthermore evident that (J, g) ~ f * g
is bilinear from C x C into CeLl.
The inequality (3.1.2), which expresses the continuity of the bilinear map
52 CONVOLUTIONS OF FUNCTIONS
(J, g) ~ 1* g from C x C (with the topology induced by that on V x Ll)

into V, ensures that this mapping can be uniquely extended continuously
into a bilinear mapping from Ll x Ll into Ll. Specifically, if I, g E Ll, 1* g
will be the Ll-limit of (Jt * gil, where (Jt) and (gt) are sequences extracted
from C and converging tol and g in Ll, respectively. [The inequality (3.1.2)
ensures that (Jt * gil is Cauchy, and therefore convergent, in Ll; and that the
limit does not depend on the chosen sequences but only on I and g.] This
mode of extension ensures that (3.1.2), (3.1.4), and (3.1.5) continue to hold,
the last by virtue of 2.3.2.
It remains to verify that (3.1.1) holds almost everywhere for general I and
g in Ll. Since both sides of this inequality are bilinear in (J, g), and since (as
the reader should pause to prove) any real-valued hELl is equal a.e. to the
difference of two nonnegative integrable functions hl and h2' each equal to
the limit of a monotone increasing sequence of nonnegative continuous
functions, it may and will be assumed that sequences (Jt) and (gt) may be

chosen from C so that ~ II t I a.e. and ~ gt t g a.e. Then the monotone
convergence theorem shows that
II * gi(X) t f
21rr I(x - y)g(y) dy.
Since (Ji * gil converges in Ll to 1* g, it follows that (3.1.1) holds for almost
all x. This shows in particular that y -+ f(x - y)g(y) is integrable for almost
all x whenever I and g are nonnegative functions in Ll. IfI and g are replaced
by If I and Igl, it is seen that the same is true whenever f and g belong to Ll.
Once (3.1.1) is established for general I, g E Ll, (3.1.3) results immediately.
[The reader should note that it is not the case that every nonnegative
integrable function h is equal a.e. to the limit of an t sequence (h n ) of
continuous nonnegative functions. As a counterexample, take h to be defined
on [0, 2rr] as the characteristic function of the complement, relative to [0, 2rr],
of a closed, nowhere dense set K ; [0, 2rr] having positive measure.]
Similar techniques are applicable in connection with 3.1.4 to 3.1.6.
At this point it is convenient to summarize what little we do know so far
about convolution and to mention a few questions that arise, which will
guide some of the subsequent developments.
3.1.1. Some Problems. The convolution I * g of two functions I and g

chosen from Ll being defined by (3.1.1), the mapping (J, g) ~ 1* g is an
associative and commutative bilinear mapping of Ll x Ll into Ll; this
mapping is continuous by virtue of (3.1.2). In current terminology (which
will be explained at greater length in Section 11.4), Ll forms a commutative
complex Banach algebra under convolution. As will be seen subsequently in
(c), however, Ll possesses no identity (or unit) element relative to convolution.
By 2.3.1, 2.3.2, and (3.1.5), the Fourier transformation I ~ J is a con-
tinuous homomorphism of the convolution algebra Ll into the algebra co(Z),
the latter being taken with pointwise operations as in 2.3.9. In particular,

for each n E Z the mapping
(3.l.6)
is a continuous homomorphism of L1 onto the complex field (the latter being

regarded as an algebra over itself).
A few questions with an algebraic flavor arise quite naturally at this point.
The following selection is typical and significant.
(a) Is the mapping 1-+ j actually an isomorphism of L1 into co(Z)? This is
equivalent to asking whether the relationsj(n) = 0 (n E Z) entail that I = 0
a.e., and has thus been answered affirmatively in 2.4.l.
(b) Are there any continuous homomorphisms of L1 onto the complex
numbers distinct from the Yn (n E Z)? In 4.1.2 we shall see that the response is
negative, thereby providing a very satisfactory explanation of the funda-
mental nature of the Fourier transformation.
(c) From 2.3.8 and (3.1.5) it appears at once that L1 contains no identity
for convolution, that is, no element e such that e * I = I for all I E L1. [Were
such an element e to exist, one would infer from (3.1.5) that e(n) = I for all
nEZ, a possibility which is ruled out by 2.3.8.] In view of this, one is
prompted to ask whether every IE L1 can be factored into a convolution
product 11 * 12 with 11 and 12 in L1. An affirmative answer was published
relatively recently (Walter Rudin and P. J. Cohen), though the result was
known to Salem and Zygmund somewhat earlier; see [Z1]' p. 378. Cohen's
method is a most elegant one, applying to a general class of Banach algebras.
We shall return to this and similar problems in Section 7.5; see also 11.4.18(6).
(d) Which elements e of L1 are idempotent, that is, satisfy e * e = e?
Plainly, every trigonometric polynomial of the form
(a finite sum)
is idempotent. From 2.3.8 and 2.4.1 it appears that these are indeed the only
idempotents in L1.
(e) Since em * en = 0 if m i: n, it is clear that L1 is not an integral domain
(that is, that it possesses an abundance of zero divisors). Which subalgebras
of L1 possess no zero divisors? Some light will be shed upon this in 11.3.9.
(f) Can one classify or describe the closed subalgebras of V? This appears
to be an extremely difficult problem. It is easy to see that, if one takes a
sequence (Sk)k': 1 of pairwise disjoint and finite subsets of Z, then the set of
IE Ll, such that j takes an (f-dependent) constant value on Sk (k = 1,2,)
and j = 0 on Z\Uk'= 1 Sk> is a closed subalgebra of Ll. It was natural to
hope that all closed subalgebras of L1 would prove to be of this type. How-
ever, J.-P. Kahane has recently disproved this conjecture. His results, as
well as some of the simpler aspects of the study of closed subalgebras, will be
discussed in Section 11.3.
(g) Conforming with normal algebraic terminology, an ideal in L1 will

mean a linear subspace I of L1 having the property that I * gEl whenever
I E I and g ELI; a closed ideal will mean a subset that is both an ideal and a
closed subset of L1.
This being so, can one effectively describe all the closed ideals in L11
In Section 11.2 we shall uncover an affirmative answer to this and a number
of similar questions for the case in which the underlying group is T, noticing
at the same time, however, that for many groups of interest the analogous
questions have no known answer that is completely satisfactory. Meanwhile,
see Exercises 3.2 to 3.5.
The reader may be struck by the fact that the questions in (f) and (g) are
not phrased in purely algebraic terms, inasmuch as we speak of Clo8ed sub
algebras and ideals. This topological restriction is customary when one is
dealing with infinite dimensional algebras. To seek to classify all (not
necessarily closed) subalgebras or ideals is both rather unnatural and over
ambitious. Topological restrictions compensate in some measure for the
infinite dimensionality and are natural just because of this feature.
As has been indicated, problem (f) is not yet solved completely. Beside this
we shall in 12.7.4 encounter problems (b), (d), and (g) in a new setting, the
algebra L1 being replaced by a larger one; and in Section 16.8 problem (d) will
appear for still larger convolution algebras. The analogues of (b) and (g) for
some of these enlarged algebras have not yet received complete solutions. It
may also be added that, for underlying groups of types markedly different
from T, the L1 version of problem. (d) is a good deal more difficult; see
Chapter 3 of [R] and also Rudin [3], Rudin and Schneider [1], Rider [1].
With these traces of ignorance left showing, we turn to some simple analyt
ical properties of the convolution process that will playa fundamental role
in subsequent developments.
A start is made with two properties that stem directly from the invariance
of the integral; the proofs are left for the reader to provide.
3.1.2. Ta(f*g) = Tal*g =1* Tag
3.1.3. Tal * Tbg r= Ta+bl * g.
3.1.4. Suppose that I ~ P ~ 00 and that p' is the conjugate exponent (or
index), defined by lip + lip' = I (p' = I if P = 00 and p' = 00 if P = 1);
If IE LP and g E LP', then I * g is defined everywhere, is continuous, and
111* glloo ~ 11111,,' Ilgll,,
Proof. Holder's inequality shows that the function y -+ I (x - y) g( y) is
in this case integrable for each x, so that 1* g(x) is defined for all x, and that
11* g(x}1 = I LJ I(x - y)g(y) dYI
~ 111(x - y)ll" '1Igll,,,

where in the first factor I(x - y) is regarded as a function of y. By translation-

invariance, 11/(x - y)llp = 11/11p; whence the stated inequality. To show that
1* g is continuous, we may by symmetry suppose that p < 00 and then use
3.1.1, 3.1.2, and the inequality just established to obtain
I Ta(J * g) -I*gll", = IITa/*g -I*gll",

= II(Tal - J) * gil",
~ IITa/-fllp'lIglip"
Finally, II Taf - Ilip -0 as a - 0 whenever I E LP and p < 00 (see 2.2.4).
Remarks. (1) The preceding result contains the first hint that convolu-
tion is a smoothing process. The next two results develop this theme by
showing that, if 1 ELl, then 1* g shares with g a number of smoothness
properties. Further developments along these lines must be deferred until
Chapter 12; see especially 12.6.2, 12.7.2, and 12.7.3.
(2) There are valid converses of 3.1.4 which are in essence contained in
12.8.4 and 16.3.5.
3.1.5. (1) If 1 E V and if g E CIc, or is of bounded variation, or is absolutely

continuous, then 1 * g has the same property. Moreover, in the first case
one has
(3.1.7)
for any integer m ~ 0 not exceeding k.
(2) The formula (3.1.7) holds for m = 1 whenever 1 E V and g is absolutely
continuous.
Proof. (1) We shall deal with the assertion involving CIc, leaving the
remaining (similar) arguments to be provided by the reader. In dealing with
this selected case, it will suffice to show that f * g E Cl if g E Cl and that
D(J * g) = f * Dg; the rest will follow by induction on m.
Now if a -=1= 0
a-1[f* g(x + a) - 1* g(x)] = ~ fl(Y) g(x +a - y) - g(x - y) dy.
2" a
Since g E Cl, the cofactor of f( y) in the integrand tends, as a _ 0, to
Dg(x - y); and, as the mean value theorem shows, the convergence is
uniform with respect to y (and to x). It follows from general convergence
theorems ([W], Theorem 4.1 b, for example) that f * g has a derivative equal to
1* Dg, this last function being continuous by virtue of the case p = lof3.1.4.
(2) Suppose that IE Ll and that g is absolutely continuous. We will first
show that f * g is absolutely continuous.
For any two real numbers a and b we have
11 * g(b) - 1 * g(a)1 ~ Lf 11(y)llg(b - y) - g(a - y)1 dy. (3.1.8)

Since g is absolutely continuous, to any e > 0 corresponds a number

o = O(e) > 0 such that
L Ig(b k) -
r
g(ak)1 ~ e
k=l
for any sequence ([ak> bkm=l of nonoverlapping intervals [a k , bk ] for which

2:L 1 (b k - ak ) ~ o. But then, under the same conditions on these intervals,
L Ig(b k -
r
y) - g(a k - y)1 ~ e
k=l
for all y and so (3.1.8) shows that
L II*g(bk) -I*g(ak)1
r
~ IIIlIloe
k=l
This shows that 1 * g is absolutely continuous.

That (3.1.7) holds with m = 1 is now most easily seen by applying 2.3.4
to each side and referring to (3.1.5).
3.1.6. HIELlandgEV(I ~p ~CX)),thenI*gEV'and
III*gllp ~ IIIlIlolIglip
Proofo For any h E U the Fubini-Tonelli theorem gives
12~ J 1 * g(x)h(x) dxl ~ L J Ih(x)1 {L J If(y)g(x - y)1 dY} dx
= LJII(Y)I {LJlh(x)g(x - y)1 dX}dY.

By Holder's inequality the inner integral is majorized by IIhllp IIglip Hence
12~JI*g(X)h(x)dxl ~ 111111 IIhll p o IIgllp
The converse of Holder's inequality (see Exercise 3.6) now goes to show that
1 * g E LP and 111 * gllp ~ Ilflll Ilgllp, as alleged.
Remarkso (1) The argument can be made less sophisticated by assuming
first that 1 and g are continuous. One may then assume that h, too, is con-
tinuous. The required versions of the Fubini-Tonelli theorem and the Holder
inequality and its converse then become simpler. This leads to the stated
result for 1 and g continuous. In general, we may assume that p < co, since
otherwise the result is contained in 3.1.4, and then approximate 1 and g in Ll
and in LP, respectively, by continuous functions I .. and g" (n = 1,2, .. ).
By (3.1.2), I .. * g" -+ 1 * g in LI, and so a subsequence converges almost
everywhere. By applying the result already established for continuous 1 and
g to the terms of such a subsequence, and making use of Fatou's lemma ([W],
Theorem 4.Id) on the way, the desired result appears.
(2) The assertion in 3.1.6 can be improved in several ways; see 12.7.3 and
Exercise 13.5. A result that combines and extends 3.1.4 and 3.1.6 will be
obtained in Section 13.6.
3.1.7. Other Convolution Algebras. From 3.1.5 it appears in particular

that each C" is an associative and commutative complex algebra under *, and
that the same is true of the set of functions of bounded variation, and the
set of absolutely continuous functions. Each of these, save Coo, can in fact be
made into a commutative complex Banach algebra under *. (COO is not
normable, but it is a good complete, metrisable, associative, and commutative
complex topological algebra under *.) Similarly, by 3.1.6, V' (1 ~ P ~ (0) is
an associative and commutative complex Banach algebra. Being subalgebras
of Ll (in the purely algebraic sense) which contain all trigonometric poly-
nomials, none of these algebras possess an identity element; se~ 3.1.1(c).
Beside this, 3.1.6 shows that LP (1 ~ P ~ (0) can be regarded as a module
over the ring Ll (* being both the ring product in Ll and the module product
between elements of Ll and elements of LP).
3.1.8. Convolution and Translation. Both 3.1.2 and 3.1.3 hint at close
connections between translation operators T a and convolution. This will be
borne out as we progress (see especially Sections 16.2 and 16.3). Meanwhile
here is a basic result in this direction.
3.1.9. Let JELl and let E denote anyone of the normed spaces C or LP
(1 ~ P < (0). If gEE, thenJ * g is the limit in E of finite linear combinations
of translates of g.
Proof. Let gEE be given. Denote by Vg the closed linear subspace of E
generated by the translates TaU of g, that is, the closure in E of the set of all
finite linear combinations of elements TaU. Denote also by 8 the set of JELl
such that J * g E VII' It has to be shown that 8 = Ll. Now it is evident that
8 is a linear subspace of Ll; and from 3.1.6 it follows that 8 is closed in Ll. It
will therefore suffice to show that 8 contains a subset 80 such that the finite
linear combinations of elements of 8 0 are dense in Ll.
If E = C, a convenient choice of 8 0 is C (see [W], Theorem 4.3b). We will
leave to the reader the task of showing that in fact J * g is the uniform
limit of finite linear combinations of translates of g, whenever J .and g are
continuous. (Hint: Approximate the integral definingJ * g by Riemann sums,
using uniform continuity of the functions involved.)
We pass on to the remaining cases.
Suppose, then, that E = L" (1 ~ P < (0). In this case a convenient choice
of 80 is the set of functions J which coincide on [0, 21T] with the characteristic
function of an interval I = [a, b], where 0 < a < b < 21T, and which are
defined elsewhere by periodicity (compare [W], Theorem 4.3a). In this case
we partition I by a finite number of subintervals lie whose lengths Illel are

majorized by a number 3 to be chosen shortly. Choose and fix a point ale in
each lie' We then have
say, and so Minkowski's inequality yields
II!* y -
1
-2
'TT1e
L: Illel' T"kyllp ~ 2-1 L: IIhlell
'TT
p (3.1.9)
Next, using Holder's inequality and the Fubini-Tonelli theorem,
IIhlell: = 2~f{lL [g(x - y) - g(x - ale)]dyIP}dx
~ 2~ f {I lie Ipip' , L Ig(x - y) - g(x - ale)jP dy} dx

= IlleI PIP' ik {(2~) flg(x - y) - g(x - ale)jPdx}dy
= IlleIP'P', f IITlig - T"kgll: dy. (3.1.10)

J lk
Now, given e > 0, we can choose 3 > 0 so small that

IITlig - T"4 g ll: ~ eP
for all YElle (see 2.2.4). Then (3.1.10) shows that
II hlell: ~ IlleIP'P', Illel ' eP = IllelP' eP,
since pIp' + 1 = p, and therefore

Ilhlell p ~ Illel' e. (3.1.11)
Combining (3~1.9) with (3.1.11), we obtain
Since
[3.2] APPROXIMATE IDENTITIES FOR CONVOLUTION 59
is a finite linear combination of translates of g, this shows that f. g E Vg and

the proof is complete.
Remarks. For a complement to 3.1.9, see 3.2.3. See also Exercise 3.7.
An alternative proof of3.1.9 will appear in part (2)'ofthe proof of 11.1.2.
3.1.W. Characterization of Convolution. Several of the results

appearing in this section, taken singly or in combination, have converses
which are interesting in that they virtually characterize convolution as a
linear or bilinear process in terms of such basic concepts as the invariant
integral, the standard function spaces, and the translation operators.
For example, 3.1.2 and 3.1.4 combine to show that, if f E LP' is given, then
the mapping U: g ~ f g is a continuous linear operator from LP into C
which commutes with translations (that is, TaU = UTa for all a E Rj21TZ).
As will be seen in 16.3.5, the converse is also true.
Again, in Subsection 16.3.11 we shall comment on converses of the result
appearing in Subsection 3.1.9.
To take a third example, it has appeared that the mapping B: (f, g) ~ f g
is bilinear, has various continuity and positivity properties, and is related to
translation in such a way that
It will appear in Subsection 16.3.12 that these properties go a long way

toward characterizing convolution as a bilinear operator.
3.2 Approximate Identities for Convolution

In 3.1.1(c) it has been remarked that Ll contains no identity element for .
The same is true of the smaller .-algebras, Ck and LP (1 ~ p ~ co). This
being so, we are going to consider and seek the next best thing, namely, a
so-called "approximate identity."
3.2.1. By an approximate identity (for convolution) we shall mean a sequence

(K n ):= 1 of elements of Ll such that
(3.2.1)
lim 21 JKn(X) dx = 1, (3.2.2)

R- co 'It'
and
lim
n_CXl
r
)6<1%1<"
\Kn(x)\ dx = 0 (3.2.3)
for any fixed 8 satisfying 0 < 8 < 1T.

It will be seen in Chapter 5, and it is fundamental to our work, that there

exist approximate identities (X n );:'= 1 in which each X" is a trigonometric
polynomial.
More immediate examples stem from the observation that any sequence
(X,,) of nonnegative integrable functions satisfying (3.2.2) and
lim r
"... "" J6 "Ixl""
X,,(x) dx =0
for each fixed 8 satisfying 0 < 8 < 1T, constitutes an approximate identity.
Thus, we might take for Xn (n = 1,2, .. ) the function which is defined on
[ -1T, 1T) to be ?Tn times the characteristic function ofthe interval [ -lin, lin],
and defined elsewhere by periodicity.
The name "approximate identity" is justified by the following result.
3.2.2. Let (X n );:'= 1 be an approximate identity.

Then
lim
" ... 00
IIX" *f - !II"" = 0 (f E C);
lim IIDm(Xn *f) - D"'fll"" = 0 (f E Ck)
provided m is an integer ~0 not exceeding k; and

lim IIX n *f - flip =0 (f ELp),
"... ""
provided I ~ P < 00.
Proof. Inasmuch as Dm(x... * f) = X" * Dmf whenever Xn E V and f E Cm
(see 3.1.5), the second statement will follow from the first. The first follows
from the uniform continuity of f in the following way.
We have
Kn *f(x) - f(x) Lf Kn(Y) dy = 2~ f K,,(y)[f(x - y) - f(x)] dy.

Putting
gives
(3.2.4)
say. Being assigned any e > 0, choose and fix 8 satisfying 0 < 8 ~ ?T, so
that IIT],.j-fll", ~ dar Iyl ~ 8.
Then
(3.2.5)
[3.2] APPROXIMATE IDENTITIES FOR CONVOLUTION 61
The first integral is, by virtue of (3.2.1), majorized by
eO 2~ IIK"I dy ~ Me, (3.2.6)
where M is independent of n. Since also IITlIf - fll .. ~ 211fll .. for all y, the
second integral is majorized by
211fll .. 21 f IK,,(y)\ dy. (3.2.7)

11' J6o;; 1111 .. "
Keeping e and 8 fixed, we find that (3.2.3) to (3.2.7) show that

IIKn.f - anlll .. ~ Me.
".....
lim sup
Since 8 is an arbitrary positive number, and since lim a" = 1 by (3.2.2), it

follows that
lim IIKn.f - III .. = 0
".....
whenever f is continuous.
To prove the third statement, given f E V' and e > 0, first choose pee
such that Ilf - PilI' ~ e. By 3.1.6 and (3.2.1),
IIK".f - K" .ftll p ~ IIK"lll 8 ~ Me, (3.2.8)
where M is independent of n. By what has been established, there exists
no = no (e) so that
for n > no.
A fortiori, then,
for n > no,
and therefore
11 K " .ft - fill' ~ 28 for n > no. (3.2.9)
Hence, by (3.2.8) and (3.2.9),
IIK".f - fill' ~ Me + 28 forn > no,

which proves the third statement and completes the proof.
3.2.3. Let E denote anyone of the normed spaces C or Lp (1 ~ p < (0).

Since each T" is a continuous endomorphism of E, 3.2.2 shows that T "(K,,. f)
- T"ffor eachfe E. Also, by 3.1.2, T"(K,, .f) = T"K" .f. It thus appears
that T "f is the limit in E of convolutions k f with k eLl. This complements
3.1.9.
The two results taken together show that, given feE, the closed linear
subspace of E generated by all translates of f is identical with the closure in
E of the set of all convolutions k f with k ranging over Ll; if E is regarded
as a *-module over Ll, this latter set is just the closed submodule of E
generated by I. Further results of this type will appear in Section 11.1.
3.2.4. Approximate Identities and the Dirac &-Function. The first

part of 3.2.2 shows in particular that
lim KII */(0) = 1(0),
""'"
or, what is equivalent after replacing I by],
lim 21 fKII(x)/(x) dx
A_CD 'Ir
= 1(0)
for each continuous I. (Actually, scrutiny of the proof would show that this
holds for any IE L'" such that I is continuous at 0.) This means that the
sequence (Kn) is of the type which is often said to converge (in some un-
specified sense) to the so-called Dirac S-Iunction. Complete precision will be
attained in terms of the ideas to be studied in Chapter 12; see especially
12.2.3 and 12.3.2(3).
3.2.5. Approximate Identities and Summation Factors. Insofar as

the study of Fourier series is concerned, one of the main effects of using
a.pproximate identities is the insertion of "summation factors" into series
which, in their unadorned state, will in general diverge. The summation
factors are, indeed, just the Fourier transforms J(n, which, as the results of
3.2.4 show, have the property that
lim J(n(m) = 1 (m E Z).
Specific examples appear in Sections 5.1 and 6.6.
3.3 The Group Algebra Concept

3.3.1. The Classical Concept. In the classical and purely algebraic
theory of a finite group G, additional flexibility was sought by introduc-
ing the so-called group algebra (or group ring) " of G. This was defined,
after choosing a field K of scalars, as the set of all formal (finite) linear
combinations
1= I(x)'x L
:reG
of group elements x E G with coefficients I(x) E K. The algebraic operations
are as follows:
al = L [al(x)) x a E K,
:reG
I +g= L [f(x) + g(x)] x,
:reG
Ig = L [ L I(x - y)g(y)] x.
:reG lIeG
(3.3] THE GROUP ALGEBRA CONCEPT 63
Taking stock of the fact that

I(f) = L: /(x)
;rliia
is an invariant integral on G (see 2.2.2 and Exercise 2.5), a little thought will
show that the group algebra may very well be pictured as the algebra of
K-valued functions on G, the linear space operations being point-wise and
the product being convolution; the sum
L: /(x -
Ilea
y)g(y)
corresponds exactly to the integral
2~ J f(x - y)g(y) dy
used in Section 3.1 to define the convolution of two functions on the group
T.
It is no part of our purpose to carry forward the study of the group algebra
of a finite group (see, for example, [Bo] and [vW]): the concept is mentioned
merely because it is the forerunner of one that holds an important place in
the modern .developments in harmonic analysis (see 3.3.2). For purposes of
subsequent comparison it is to be noted that the study of the group algebra
of a finite group leads ultimately to a good deal of information about the
structure of the underlying group, albeit only when combined with the study
of representations of the group.
3.3.2. The Modern Concept. On turning to infinite groups, and

specifically to locally compact topological groups G, there are various ways in
which the group algebra concept can be extended. It is customary in these
conditions to assume that the ground field K is the complex field. Nonetheless,
considerable freedom of choice remains, especially when G is compact. For
G = T, for instance, one might attach the term ... group algebra" to any
one of I? (1 ~ P ~ 00), or Ck (0 ~ Ie ~ 00), or again the measure algebra
M introduced in Chapter 12. The favorite contender for the title is Ll, mainly
perhaps because L1 remains a convolution algebra even when the underlying
group is noncompact. (This last property is shared by M, but M is generally
far more mysterious than is Ll; see 12.7.4.)
There are, of course, important differences between the group algebra" of
a finite group and anyone of Lp or Ck. Thus (1) " possesses an identity
element (to wit, the function taking the value 1 at x = 0 and the value 0
elsewhere; compare 3.1.1(c; (2) " is a finite-dimensional linear space, which
is evidently not the case with I? or Ck. The difference pointed out in (2) means
that the study of the group algebra in its modern guise is as much one of
analysis as of algebra.
Again because of these differences, it is much more difficult to establish

clear-cut relations between the properties of a group algebra and those of the
underlying group; see the remarks in 4.2.7. The fact is that the modern
approach lays more emphasis on the structure of function spaces carried by a
group, and less on the underlying group itself.
3.4 The Dual Concepts

There is no trouble involved in framing the definition of the convolution
t/J * rp of two functions on Z, provided that these functions are suitably
restricted in their behavior at infinity; see Exercise 3.15.
The same is not true ofthe analogue of (3.1.5), however. The analogue reads
(3.4.1)
and the remarks in Section 2.5 suffice to point up some of the difficulties
encountered in establishing (3.4.1). The only simple case is that in which both
t/J and rp belong to 11, in which case cP * rp belongs to f1, too [compare (3.1.2)].
Our more immediate concern in the sequel will lie with cases of (3.4.1) in
which cP and rp have the forms j and g, respectively, where i and g are suitably
restricted functions on G = T: in this case the formula (3.4.1) appears
essentially in the disguise of versions of the Parseval formula to be discussed
in Chapters 8 and 10. See also 12.6.9 and 12.11.3.
EXERCISES
3.1. For i ELl let

SNi = 2:
Inl"N
j(n)e n ,
and suppose that DN is defined as in Exercise 1.1. Verify that sNi = DN *i

and that
8Ni * g = 8N(i * g) = i *SNg
for i, g ELl. Deduce that sNi is the limit in Ll of linear combinations of
transla tes of f.
3.2. Suppose that 1 ~ P ~ 00 and define
I = {f E V' : IlsNillp = 0(1) as N ~ oo},
J = {fEV: lim Iii - sNillp = O}.
N-""
Verify that I and J are submodules of V (considered as a *-module over Ll)
and that J c I. Show that I and J are everywhere dense in LP if P < 00. If
p = 00, is J dense in L""? Give reasons for your answer.
EXERCISES 65
3.3. Take p = 1 in the preceding exercise. Show that I and ,J are non-
closed in Ll.
Hint: Assume the result of the computations in 5.1.1 and use the uniform
boundedness principle as stated in Appendix B.2.1 or in B.2.2.
3.4. (1) Let E be the set off ELl such thatf(x) = f( -x) a.e. Is E an ideal
in L1? Is it a subalgebra of L1? Is E closed in L1? Give reasons for your
answers.
(2) Prove that W1 (f * g) ~ II! III . w1g for f, g E Ll. Deduce that the set
1= {fE Ll: wd(a) = O(laI112) as a~ O}
is an ideal in Ll. Is I closed in Ll? Give reasons.
Hint for (2): Observe that I is everywhere dense in L1.
3.5. Suppose that f E La> is such that the function a ~ Taf is continuous
from R into L'" for the normed topology on the latter space (that is, that
IITaf-fII",~O as a~O). Prove thatfis equal almost everywhere to a.
continuous function. (The converse is true and almost trivial.)
Hints: Take an approximate identity (K n );:'=l in Ll and consider the
functions fn = Kn *f. Show that the fn are equicontinuous and uniformly
bounded. Let (xdt": 1 be a sequence that is everywhere dense in (0, 21T) and
pick strictly increasing sequences of natural numbers (nk(i))k'=l so that
(nk (1+ 1)k'= 1 is a subsequence of (nk(f))k'= 1 and limk~ fn<(O(x,) exists finitely
<Xl
for each i. Deduce that there exists g E C such that, if n k = nk(k) (the
"diagonal subsequence"), then fnk ~ g uniformly. Use 3.2.2 to compare /
and g.
Remarks. This is the special case, for the group R/21fZ, of a result due to
D. A. Edwards [1] for general groups. An analogous and older result for Radon
measures (see Chapter 12) is the work of Plessner and Raikov. Both types of
result are treated in R. E. Edwards [2]. See also Exercises 11.22 and 12.23. The
existence of a uniformly convergent subsequence of (in) is a special case of
Ascoli's theorem; see [E], Section 0.4.
3.6. (Converse of Holder's inequality) Suppose that 1 ~ P ~ 00 and that
/ E Ll is such that
(1)
for each g E C, m being a number independent of g. Prove that/ E V", where

lip + lip' = 1, and that
II/lip. ~ m.
Hint: Show that (1) continues to hold for g E Leo.
Remark. It is even enough to assume that (1) holds for g E Coo. For
another variant, see 13.1.5.
3.7. Does 3.1.9 remain valid if therein one takes E = Ck? Justify your
answer.
3.8. Let (K,,);'ml be any approximate identity. Show that, if p > 1, then
compare Exercise 7.5.

3.9. Let H" denote the set of 1 E L" such that j(n) = 0 for n < 0 (here
1 ~ P ~ (0). Show that H" is a closed ideal in L".
Remarks. The study of the so-called Hardy spaces H" is an elaborate

procedure having close conncctions with complex analytic function theory, the
connection being explained by the remark that each f E H" may be regarded as
the boundary values (for Izl = 1 and in a sense that can, and must" be made
precise) of the functionf# defined for complex z satisfying Izl < 1 by the power
series Ln>oi(n)zn. (In terms of the functions f#, the definition of H" can be
extended to cases where 0 < p < 1.) This book contains no attempt to discuss
the subject systematically, although sidelong glances are thrown in that
direction in Exercises 6.15, 8.15, 11.8, 11.10 and in Section 12,9 and 12.10.3.
For detailed accounts of the subject, sec [Ho]; [Hel]; [dBR]; [R], Chapter 8;
[Zd, Chapter VII; [Kz], pp. 81 ff.; and [Ba2], pp. 70-93. For a survey of the
abstract theory, see Srinivasan and Wang [1] and the references cited there and
also MR 37 # 1982; 55 ## 989, 990.
3.10. Let (KN)N=l be a sequence of nonnegative integrable functions

such that limN_",KN(n) = 1 (n E Z). Show that
lim KN
N-",
*1 = 1
uniformly for each continuous J. Deduce that (KN)N= 1 is an approximate
identity (see 3.2.1).
3.11. Assume that a is a real number such that a/TT is irrational and that
f is a measurable complex-valued function such that Tal = f a.e. Show
that I = const a.e. (Recall that all functions considered have period 211'.)
3.12. Suppose 1 ~ P ~ 00. Show that the convolution algebra L" has no
nonzero generalized (or topological) nilpotents, that is, elements 1 such that
inf" 111*"11/'" = 0,
where 1*1 = 1 and 1*(1<+ 1) = 1 *1*" (k = 1,2, .. ). See 11.4.18(1).
3.13. Assume that K E Ll and IIKlll < 1. Given gEL" (where 1 ~ P ~ (0),
show that the equation
has the unique solution
f = g + L'" K*" * g,
,,=1
EXERCISES 67
and that this solution satisfies
Remarks. An obvious necessary condition in order that the given equation

be soluble for each g E V' is that Jt(n) :f. 1 (n E Z). That this condition is
sufficient lies somewhat deeper. It is a corollary of 11.4.13.
\Ve remark in passing that convolution inequalities of the formJ - K J ~ 0
have come to play a role in certain parts of morlern potential theory and
harmonic analysis. In this connection K, which has hitherto denoted an
integrable function, frequently denotes a measure (see 12.2.3). Such inequalities
will play no role in this book, but the interested reader should consult Essen [IJ
and the references cited there.
3.14. Suppose that (cn)nez is a sequence such that
neZ
2: Ic,.j(n)i < 00
for each fELl. Show that Lnezlcnl

< 00.
Hint: Use the uniform boundedness principle in Appendix B.2.1.
Note: The hypothesis should not be confused with the demand that
LcJ(n) be merely convergent for each fELl; see 10.5.1.
Similar arguments will establish a more general lemma of Bosanquet and
Kestelman, which asserts that if the functions Uk (k = 1,2, ... ) are such that
JUk E L1 and '2.:=1 If JUk dxl < a) whenever JELl, then '2.:=1 IUkl E L"'.
3.15. Consider sequences (= complex-valued functions on Z) 4>,.p, .. '.

Frame a definition of the convolution 4> * .p which will be such that (fg)'" =
j * g for trigonometric polynomials f and g.
Using the notations introduced in 2.2.5, discuss the dual aspects of the
results in Section 3.1.
Note: Further discussion of the relation (fg)'" = j. g appears later in
the guise of the Parseval formula; see Chapters 8 [especially (8.2.5)] and 10.
The change of the underlying group from T to Z has led to some interesting
problems concerning convolution over noncompact groups, which have been
studied in a sequence of papers by Rajagopalan and Zelazko, a useful sum-
mary of which appears in Math. Rev. 32 # 2506. See also Exercise 4.6 below;
[HR], (38.26) and (38.27); MR 34 # 1868, 8213; 35 # 7136; 37 # 4509; Gaudet
and Gamlen [1).
3.16. Let S be a set of real numbers whose interior measure
m.(S) == sup {m(F) : F closed, F c S}
is positive. Show that the set of differences St == S - S == {x - y : XES,
YES} contains a neighborhood of O.
Hints: By taking a suitable closed subset of some translate of S, it may
be assumed that S is measurable, is contained in (0, Y27T), and that m(S) > O.
LetJbe the periodic function that coincides on [-17', 17'} with the characteristic
function of S. Apply 3.1.4 to J * j. .
Notes: The result is due to Steinhaus (Sur les distances des points des
ensembles de mesure positive. Fund. Math. 1 (1920), 93-104). It could be for-
mulated entirely in terms of the group T; in fact, the result is valid for any
locally compact group whatsoever, the proof being a simple adaptation of tha.t
which is proposed above. A corollary is that a subgroup is either open or has
zero interior measure.
For some generalizations, see Ray and Lahiri [I], Mueller [I], and the
references cited there.
There is an analogous result due to Banach, Kuratowski, and Pettis applying
to topological groups that are not necessarily locally compact, measure-
theoretic concepts being repla.ced by category; see [K], p. 211 and (for special
cases) [Zd, p. 250, Example 2. For a converse, see MR 51 # 6286.
3.17. Let (an}:'=l be a sequence of real numbers and suppose that
lim n _ '" exp (ianx) exists for each x belonging to a set of real numbers having
positive interior measure (see the preceding exercise). Prove that lim n _ co an
exists finitely.
Hints: The set S of points of convergence of the sequence (exp (ianx)) is
evidently a subgroup of R. Use Exercise 3.16 to conclude that S = R, so that
g(x) = lim n _ '" exp (ianx) exists for all real x. By integration theory, therefore,
lim
n-aJ
r J(x} exp (ianx) dx = JRr J(x}g(x} dx
JB
for every function J which is Lebesgue-integrable over R. Deduce first that
(a n):'= 1 is bounded (an adaptation of 2.3.8 will be needed here), and then (by
choosing J suitably, or by a compactness argument) that this sequence is
convergent.
3.18. Let (Cn}:'=l be a sequence of complex numbers and (an):'=l a
sequence of real numbers. Suppose that lim n _ co Cn exp (ianx) exists for each
x belonging to a set of real numbers having positive interior measure (see
Exercise 3.16). Show that (i) (c n ) is convergent to some complex number and
(ii) if lim n _ '" Cn i= 0, then (ex n ) is convergent in R.
3.19. Let X be a measurable character of T. Prove that X is continuous.
Hints: Use Exercise 3.16 to show that X is bounded. Then establish
continuity of X by adapting the reasoning' used in 2.2.1 to show that a
continuous character is differentiable.
Alternatively, see [HR), p. 346, where the result is stated and proved in a
more general form.
Remark. Despite the stated result, there exist characters of T which are
both hounded and nonmeasurahle.
CHAPTER 4
Homomorphisms of Convolution
Algebras
In this brief chapter we introduce the reader to two problems typical of the
current outlook on harmonic analysis. The first problem, which will be solved
in detail in Section 4.1, arises on choosing anyone of the convolution algebras
E mentioned in Subsection 3.3.2 and seeking to determine all the homo-
morphisms I' of E into the complex field. The answer highlights the funda-
mental im portance of the Fourier transformation in relation to group structure.
The second problem is concerned with the (self-) homomorphisms of E
(that is, the homomorphisms of E into itself). Of the available choices of E,
only the cases L2 and L1 are fully solved. The former case is easy and of
relatively little interest (compare Exercise 8.1). The case E = L1 is, on the
contrary, comparatively very complex, and we shall be able only to indicate
how the solution of the complex homomorphism problem allows a useful
reduction to be made, and to indicate the solution for this case.
An incomplete account of the homomorphism problem is inserted at this
early stage because it has been learned in 3.1.9 and 3.1.10 that convolution
is related in a very basic way to the group structure, and because, granted
the ensuing fundamental role of convolution, the homomorphism problem
begs for recognition without delay. This problem has, in fact, proved to be
one focus of interest in contemporary work.
4.1 Complex Homomorphisms and Fourier Coefficients

4.1.1. We have to consider nontrivial complex homomorphisms I' of the
convolution algebra L1 into the complex field. In other words, y is a linear
functional on the complex linear space L1 which is not the zero functional
and which satisfies
1'(/ * g) = 1'(/) y(g) (4.1.1)
for f, g ELl.
lt will appear in 11.4.9 and 11.4.12 that such a homomorphism I' is
necessarily continuous on L1, that is, that
11'(/)1 ~ const II fill' (4.1.2)

but we shall temporarily assume explicitly that I' is continuous.
69
70 HmlOl\IORPHIS~IS OF CONVOLUTION ALGEBRAS
4.1.2. Let Y be any nontrivial continuous complex homomorphism of Ll.

Then there exists a unique It E Z such that Y = Yn' where [see (3.1.6)]
Yn(J) = j(n)
for fE Ll.
Proof. The uniqueness of It is clear.
We shall offer two proofs of the existence of n, the relative merits of which
will be weighed in 4.1.3.
First Proof. Define Cn = y(e n ) for n E Z. Since y is continuous and
nontrivial, the density theorem of 2.4.4 entails that Cn i: 0 for at least one
integer n. Since also en * en' = en or 0 according as n' is, or is not, equal to n,
an application of Y shows that Cn Cn' = Cn or 0 according as n' is, or is not,
equal to n. Thus Cn , is equal to 1 or to 0 according as n' is, or is not, equal
to n. Linearity of Y shows then that y(J) = j(n) for all trigonometric
polynomialsf. Continuity of y, together with the density theorem, accordingly
show that y(f) = j(n) for allfE Ll.
Second Proof. In view of the continuity of y and the fact that C is
everywhere dense in V, it will suffice to show that, for some nEZ, the
formula
y(J) = j(n)
holds for each continuous f.
Now, again since y is continuous, we can choose and fix a continuous fo
such that y(Jo) is nonzero. Consider the function X defined on Rj21TZ by
( ) _ y(T xfo}. (4.1.3)

Xx - y(Jo)
Evidently, X(O) = 1. By (4.1.2) and the fact that
which tends to zero with x - y, we see that X is continuous. Moreover, 3.1.3

and (4.1.3) combine to show that
x(x + y) = X(x)X( y) .
Consequently (see 2.2.1) there exists n E Z such that
(4.1.4)
Now take any continuousf. Then (compare the proof of 3.1.9), given any
B > 0, fo * f is uniformly approximated to within
B by any sum of the form
[4.1] COMPLEX HOl'tlOMORPHIS}IS AND FOURIER COEFFICIENTS 71
where 0 = Xo < Xl < ... < xm = 271 is any partition of[O, 271] whose "mesh"
max"(x,, - X"-l) is sufficiently small (depending upon '/0' and/). A fortiori,
for any such partition. Therefore, if we apply y and use (4.1.2), we conclude
that
Using (4.1.1), dividing through by y(/o). and using (4.1.3) and (4.1.4), we
find that
(4.1.5)
for all partitions of sufficiently small mesh. But, as the mesh of the partition
tends to zero, the sum appearing on the left of (4.1.5) converges (since / is
continuous) to the integral
f -
1 /(x)e-1n:r dx = /(n).
271
It follows that y(/) = j(n), and the proof is finished.
4.1.3. Comments on the Precedin~ Proofs. (1) It is very easy to see

that the first proof of 4.1.2 adapts readily to the case in which V is replaced
by any subset E ofLl fulfilling the following four conditions:
(a) E is an algebra under convolution;

(b) E is a topological space;
(c) a set S of integers exists such that en E E for n E S, while the linear
combinations of these en are everywhere dense in E;
(d) for each n E S, the function/ _ j(n) is continuous on E.
The conclusion is then that each nontrivial continuous complex homo-

morphism y ofE is ofthe form y(/) = j(n) for all/ E E and some 'Y-dependent
neS.
Moreover, when the substance of Chapter 12 has been absorbed the reader
will see that in the above it is unnecessary to assume that the elements of E
are integrable functions: they may be permitted to be distributions. Use
will be made of this remark in Section 16.6.
(2) If the reader scrutinizes carefully the second proof, he will see that it
also stands with only verbal changes when Ll is replaced by Ck or by
LP (1 ~ P < 00).
72 HOMOMORPHISMS OF CONVOLUTION ALGEBRAS
In the case of LOO, one change is necessary owing to the fact that C is not
dense in LOO. However, (4.1.1) and the fact thatf*gEC whenever fEL""
and gEL"', ensures that Y, being by hypothesis not identically zero on L"",
cannot vanish identically on C. Hence fo can be chosen as before. The
preceding argument shows that y{f) = j(n) = Yn{f) for all f E C. But then,
if gEL"', we have (since fo * g E C)
y{fo * g) = Yn{fo * g),
that is,
y{fo)y(g) = Yn{fo)Yn(g)
Since y{fo) = Yn{fo) # 0, so y(g) = Yn(g) for all gEL"'. The conclusion is thus
valid for LOO also.
The second proof is clearly more complicated and laborious than the first.
It has been included, because it can be adapted to cases in which the under-
lying group is noncompact (the groups Z and R, for example). In such cases
the first proof breaks down completely because V then contains no con-
tinuous characters at all. Moreover, the second proof bypasses the density
theorem and relies on fewer facts concerning harmonic analysis.
4.1.4. An analogous and much more difficult problem arises when Ll is

replaced by the measure algebra M introduced in Chapter 12; see especially
the remarks in 12.7.4.
4.1.5. A full exploitation of the results obtained in 4.1.2 and 4.1.3 depends
on the Gelfand theory of complex commutative Banach algebras. Were we
seeking to develop harmonic analysis on a general group, it would at this
point be advantageous to embark on the Gelfand theory and reap the fruits
of its application. As it is, however, we shall defer this sowing and harvesting
until Section 11.4.
4.2 Homomorphisms of the Group Al~ebra
4.2.1. Statement of the Problem. In Section 3.3 we have remarked that

each of V' (1 ::;; p :::; (0) and Ck (0 ::;; k < 00) is a possible analogue of the
group algebra of a finite group, and that Ll is the favored contender for this
title. Exhibiting no prejudice for the moment, we let E denote any Qne of
these group algebras.
A problem exerting a natural appeal and depending for its solution (as far
as this is known at present) on harmonic analysis, is that of determining as
explicitly as possible the homomorphisms of E (into itself). By such a
homomorphism we shall mean a continuous linear mapping T of E into
itself with the property that
T{f* g) = Tf* Tg (4.2.1)
[4.2] HOMOMORPHISMS OF THE GROUP ALGEBRA 73
for f, gEE. (By using the closed graph theorem as indicated in 4.2.3, together
with the necessary continuity of complex homomorphisms of E, one could
show that any homomorphism of E into itself is necessarily continuous.)
In Exercises 4.2, 4.3, and 4.8, the reader's attention is directed to the
simplest and most obvious such homomorphisms, namely those of the type
Tf (x) = e'n", f (kx) ,
where n, k E Z and k i: O. Each of these particular mappings T defines a

homomorphism of each of the group algebras E envisaged. This feature is
nontypical inasmuch as there are homomorphisms T of Coo (or of L2) which
are not extendible into homomorphisms of Ll; see Exercise 8.l.
This homomorphism problem has been posed for general groups. For the
case in which E is either Ll or the measure algebra M (see Chapter 12), most
of what is currently known is presented in detail in (R), Chapter 4. All the
results of this nature are relatively recent.
Our aim is confined to indicating how a knowledge of the complex homo-
morphisms ofE, discussed in Section 4.1, permits a small step forward in the
shape of representing the problem in a different and more tractable form,
and to stating the solution.
4.2.2. Reformulation of the Problem. Let T be a homomorphism of

E. For each n E Z the mapping
f -* (Tf('(n) (4.2.2)
is a continuous complex homomorphism of E. So, in accordance with 4.1.2

and 4.1.3, this mapping (4.2.2) is either trivial (that is, identically zero) or is
of the type Yn' for some n' E Z.
Denote by Y the set of n E Z for which the mapping (4.2.2) is nontrivial on
E. We then have
(Tf('(n) = j(n')
for n E Y and fEE. This' relationship entails that n' is uniquely determined
by n E Y, so that a mapping
a: Y -* Z
is obtained for which a(n) = n'. Thus

(Tf)"'(n) = j a(n)
0 (4.2.3)
for n E Y and fEE.
For n E Z\ Y, (Tf)"'(n) = 0 for all fEE. The Riemann-Lebesgue lemma
2.3.8 suggests that we write accordingly a(n) = 00 for n E Z\Y, interpreting
j( (0) as O. If this be done, a may be regarded as a mapping of Z into Z u {oo}
and (4.2.3) holds for n E Z andf E E.
74 HOMOMORPIDSMS OF CONVOLUTION ALGEBRAS
The mapping a must have the property that
"L/(a(ne" =
"eZ fteY
"L I(a(n))e" E E (4.2.4)
whenever fEE. The reader will notice that (4.2.4) is intended to mean simply
that there exists gEE (unique by 2.4.1) such that g = loa; no statement is
intended concerning the convergence of the series in (4.2.4). (The series do in
fact converge distributionally, as will appear in Chapter 12; no use is made of
this fact.)
4.2.3. The next step is to show that, if a is a mapping of Z into Z v {<Xl}

having the property expressed in (4.2.4), then the mapping T of E into itself
defined by
Tf = I(a(ne"
"eZ
"L (4.2.5)
is a continuous homomorphism of E into itself. The uniqueness theorem 2.4.1

is being invoked once again in order to reach assurance that there is precisely
one gEE for which g = loa, this g being Tf.
To begin with, this same uniqueness theorem shows at once that T is
linear and satisfies (4.2.1) as a consequence ofthe relations (>.J) 0 a = >..(1 0 a),
(/1 + /2) 0 a = (/1 0 a) + (/2 0 a), and (/d2) 0 a = (/1 0 a) (/2 0 a) holding
for any scalar>.. and any fl' f2 E E. It therefore remains to show that T is
continuous, to achieve which end we shall invoke the closed graph theorem
(see Appendix B.3.3). This invocation is permissible, since E is in all cases
either a Banach space or a Frechet space (see 2.2.4).
In order to prove continuity of T it is sufficient (according to the said
closed graph theorem) to show that the assumptions .
lim
Ie ......
lie = 0 in E, lim Tile = g in E
Ie ......
(4.2.6)
imply the conclusion

g = O. (4.2.7)
But, in all cases here envisaged, limle .... 00 lie = 0 in E entails that limle .... 00 A = 0
pointwise on Z. Similarly the second clause of (4.2.6) entails thatlim/c ..... (TI/c)"
= gpointwise on Z. Since (Tltr' = l/c 0 a, a(Z) c Z v {<Xl}, andA(<Xl) = 0, it
follows that g = 0 and so (by the uniqueness theorem once again) the
conclusion (4.2.7) is obtained. This completes the proof that T is continuous.
4.2.4. To sum up, we find that the homomorphisms T of E into itself

correspond to the mappings a : Z - Z v {<Xl} having the property expressed
in (4.2.4), the correspondence T +-+ a being specified by (4.2.5).
At this point the reader may care to try Exercises 4.4 and 4.5.
[4.2] HOMOMORPHISMS OF THE GROUP ALGEBRA 75
4.2.5. The determination of the mappings a of Z into Z u {co} which have

the property expressed in (4.2.4) is (except when E = L2; compare Exercise
8.1) lengthy and difficult. For the case in which E = Ll, the solution was
given by Rudin in 1956 and is described in 4.2.6; the solution of the dual
problem, in which the group R/27TZ is replaced by Z, will be mentioned in
10.6.3(3). The analogous problem for general groups (hence, in particular, the
dual problem for the group' Z) and homomorphisms of V into the measure
algebra M was solved in generality by P. J. Cohen [2] in 1960, although
partial solutions had been discovered earlier bi various writers (see the
remarks in 12.7.4). Details and more references appear in [R], Chapter 4,
and [Kah], Capitulos IV-VII. See also Exercises 12.49 and 13.6.
4.2.6. Statement of the Solution. Rudin's solution of the problem of

homomorphisms of L1 into itself (for the group R/27TZ) is expressed in terms
of the associated mappings a and is as follows.
In order that a mapping a of Z into Z u {co} shall yield a homomorphism T
of L1 into itself via the formula (4.2.5), it is necessary and sufficient that
there exist an integer q > 0 and a niapping f3 of Z into itself with the following
properties:
(1) If A l' ... , Aq denote the residue classes of Z modulo q, then Y = a -1( Z)
is of the form S1 U .. U S" where each Sf is either finite or is contained in
some A J from which it differs by a finite set, and where the Sf are pairwise
disjoint;
(2) f3(n + q) - f3(n) (n E Z);
(3) f3(n + q) + f3(n - q) = 2f3(n) (n E Z);
(4) a(n) = f3(n) for all save at most a finite number of n E Y.
The conditions (1), (3), and (4) are necessary and sufficient in order that a;
shall yield a homomorphism of L1 into the measure algebra M introduced in
Chapter 12.
The reader will experience no difficulty in verifying that condition (3)
signifies the existence of integers Uh and V h (h = 0, 1" . " q - 1) such that
f3(kq + h) = uhk + vh (4.2.8)
for (k, h) E Z x {O,-l, .. " q - I}; and that the conjunction of conditions (2)
and (3) signifies that (4.2.8) holds and that in addition U h - 0 for h =
0,1, .,q - 1.
4.2.7. Other Problems; Multipliers. A specialized question which has

received a good deal of attention is this: To what extent does the existence of
an automorphism T of L1 entail the existence of an automorphism of the
underlying group Rj21TZ? Or, in a more general form: To what extent does the
existence of an isomorphism T of L1(G 1) onto L1(G 2 ) entail the existence of an
isomorphism of G 1 onto G2 1 Partial solutions of these problems are known;
see [R], Theorems 4.7.1, 4.7.2, and Section 4.7.7, and 16.7.1 below.
76 HOMOMORPHISMS OF CONVOLUTION ALGEBRAS
Analogous problems concerning homomorphisms and isomorphisms of

V(G 1 ) onto V'(G 2 ) have been posed. The RudinCohen method does not appear
to extend to values of p i= 1. Another method of attack on the isomorphism
problem rests upon a preliminary study of the socalled multipliers of an algebra
LP(G), that is, the continuous endomorph isms U of LP(G) which commute with
translations (so that UTa = TaU for all a E G). The connection between the
two problems stems from the fact that, if T is an isomorphism of V(G 1 ) onto
LP(G 2 ), thenrthe formula U 1 = T-IU 2 T sets up a onetoone correspondence
between the multipliers of V(Gd and those of LP(G 2 ). From there on the hope
is that connections between the multiplier algebras can be translated into ones
between the underlying groups.
In Chapter 16 we shall study multipliers per se, turning aside in 16.7.1 in
order to sununarize how information about multipliers of special types bears
upon isomorphism problems.
Regarding homomorpbisms of LP, where 1 < P < 00, see also 15.3.6.
For a guide to further reading, see MR 41 # 4141; 53 # 8781; 54 ## 3296,
5746.
EXERCISES
4.1. Without using 4.1.2, show that if y is a nontrivial complex homo-

morphism of Ll, and if
M = y-l({O}) == {JELl: y{f) = O},
then M is an ideal in V such that
(1) M =F Ll;
(2) M is maximal, that is, there exists no ideal I in Ll, distinct from M
and from Ll, such that Mel;
(3) There is an identity modulo M, that is, an element e of Ll such that
e *f - f E M for every fELl (M is accordingly termed regular or modular).
Remark. It follows from 11.4.9 that the above relation between modular
maximal ideals and complex homomorphisms is reversible and remains in
force in a more general setting.
4.2. Let mE Z, k E Z, k =F O. Define
Tf(x) = e'm::: f(kx).
Verify that T defines a continuous homomorphism of E (= LP or C) into
itself. When is T(E) = E?
Hint: See 2.2.2.
4.3. T being as in the preceding exercise, what is the corresponding map
IX (as introduced in 4.2.2)1
4.4. Suppose that T and IX are as in 4.2.1 and 4.2.2. Show that T is
one-to-one and T(E) contains all trigonometric polynomials if and only if
y = Z and IX maps Z one-toone onto itself.
EXERCISES 77
4.5. Suppose that T is a continuous homomorphism of Lt' (1 ~ P ~ 00)

into itself and that a is the corresponding map of Z into Z u {oo} (see 4.2.2).
Show that
(1) For any m E Z, a-1({m}) is finite;
(2) A number k ~ 0 exists such that for all m E Z
L:
"Ea- 1({m
e.. ll" ~ k.
Hint: Consider the T-image of em'

4.6. /"(Z) is defined for 1 ~ P ~ 00 as in 2.2.5. Verify that /l(Z) is an
algebra under convolution. Has the algebra /l(Z) an identity element1
Is t"(Z) an algebra under convolution, if p > 11 (See the Note attached to
Exercise 3.15.)
4.7. Find all the continuous complex homomorphisms of the algebra
11(Z). (See 11.4.17.)
4.8. Determine all complex-valued functions h on R/21TZ having the
property that the operator T: f -- hf is a homomorphism of T into Ll, each
of the latter being considered as a convolution algebra.
4.9. In the general theory developed in Chapter 4 of [R], affine maps play
a prominent role.
Let a be a map of Z into itself which is affine in the sense that
a(n + n' - nW) = a(n) + a(n') - a(nW)
whenever n, n', nW E Z. Show that a defines, via formula (4.2.3), a homo-
morphism T of Ll into itself, and find an explicit closed formula expressing
Tf in terms off.
CHAPTER 5
The Dirichlet and Fejer Kernels. Cesaro

Summability
In this chapter we introduce the so-called Dirichlet and Fejer kernels and
their elementary properties. These kernels are basic in the study of pointwise
convergence and summability, respectively, of Fourier series. From their
properties we shall derive the localization principle, together with alternative
proofs of the uniqueness and approximation theorems of 2.4.1 and 2.4.4.
Included in the text and in some of the attached exercises are a few
properties of Cesaro summability as applied to general series, most of which
will later be applied in the case of Fourier series.
5.1 The Dirichlet and Fejer Kernels

In this section we define these kernels and state a few of their basic
properties that are crucial in the study of pointwise convergence and Cesaro
summability of Fourier series.
A few words about terminology and notation are required. The functions
DN and FN introduced in (5.1.2) and (5.1.7), and herein called the Dirichlet
and Fejer kernels, respectively, are precisely twice the functions to which
these names are customarily attached; compare, for example, [Ba l ], pp. 85
and 133-134 and [Zd, pp. 49 and 88, where KN is written in place of %F N'
This choice of nomenclature has been made in order that the convolution
expressions in (5.1.1) and (5.1.6) shall be valid.
If N is a nonnegative integer, the Nth symmetric partial sum of the
Fourier series of! is
8N!(X) = L
Inl';;N
j(n)e 1nz
When we speak of the convergence of the Fourier series of! we shall always
mean the convergence of these symmetric partial sums; see the end of 2.2.2.
Inserting the integral expression for j(n) we find that
8N!(X) = 2~ f!(y)D.v(X - y) dy = DN *!(x), (5.1.1)

78
[5.1] THE DIRICHLET A...~D FEJER KERNELS 79
where (see Exercise 1.1) one has for xO (mod 21T)
DN(x) = 2:
Inl';N
e1nx = sin (l!
sm
-:- Yz)x;
Y2x
(5.1.2)
if x is congruent to 0 (mod 21T), DN(X) has the value 2N + I, which is the

value obtained by continuous extension of the expression on the extreme
right in (5.1.2). The function D N , or sometimes the sequence (DN)'N=o, will be
spokm of as the Dirichlet kernel.
It is to be observed that DN is a trigonometric polynomial of degree N
which is even in x and satisfies
(5.1.3)
Furthermore,
(0 < 8 ~ Ixl ~ 1T). (5.1.4)
The Nth Cesaro sum (see the end of 2.2.2) of the Fourier series of f is
the arithmetic mean of the first N + 1 terms of the sequence of symmetric
partial sums thereof, namely,
(5.1.5)
These are also spoken of as the (C, I)-means of the Fourier series off' .. "c"
for Cesaro and the" 1" indicating first order arithmetic means. Using (5.1.1)
and some elementary calculations, we find that
UN!(X) = L f!(Y)FN(X - y) dy = FN *!(x)

= 2: (1 - N_In_I
Inl';N +1
)J(n)e 1nX , (5.1.6)
where the functions (see Exercise 1.1)

F ( ) = Do(x) + ... + D.v(x)
N X N + 1
= 2: (1 - N_In_I
Inl.;N +1
)e 1nX
_ [sin Yz(N + I)x/sin YzX]2

(5.1. 7)
- N +1
constitute the so-called Fejer kernel (because Fejer was the first to consider
systematically the Cesaro summability of Fourier series); when x == 0 (mod
21T), the final expression (5.1.7) is to be interpreted as N + 1 by continuous
extension.
Notice that F N is an even trigonometric polynomial of degree N and that
(5.1.8)
80 THE DIRICHLET AND FEJER KERNELS. CEsARO SUMMABILITY
Moreover,
(0 < 8 ~ Ixl ~ rr). (5.1.9)
5.1.1. Concerning IIDNIIt- Since FN ~ 0, the relations (5.1.8) and (5.1.9)

suffice to show that (FN)NzO is an approximate identity (see 3.2.1). This is
not true of the D N , for which (3.2.1) fails (as will be seen forthwith). This
failure is the root cause of most troubles concerning the convergence of
Fourier series, and also the reason why summability is often effective when
convergence is not.
We shall show now that (3.2.1) fails for (DN)N=O, since in fact
(5.1.10)
as N -00. Indeed, (5.1.2) gives
IIDNl11 =~ I"
2rr -11
ISin (~ + %)Xl dx
sm %x
2 i,,/2lsin
= -rr 0 . + l)YI d y
(2N
smy
(tt
pu mg y = 11 )
72 X
= ~ f/2lsin(2Ny + l)YI dy + 0(1),

since (siny)-l - y-1 is bounded on (0, rr/2). Putting t =(2N + l)y, the
remaining integral becomes
~ r Yo (2N+1)1I Isin tl dt = ~ ~ (Yo(k+1)11 Isin tl dt
rrJo t rrkm OJYok1l t
= ~ ~ (Yon Uk(S) ds
rr k~O Jo %krr +s
on putting t = % krr + sand uk(s) = sin s or cos s according as k is even or
odd. Now, for k = 1,2, ... and 0 ~ s ~ rr,
1 1 rr 4
o """"~ - --
%krr %krr +s ~ (%krr)2
= -2,
rrk
and L.k= 1 liP is convergent. So
I DNl11 = -2 L
2N
1/( %krr) 5" /2
uk(s) ds + 0(1)
rr k= 1 0
=-
2
rr
L -2
2N
krr
1t=1
+ 0(1).
Since L.~~1 11k = log N + 0(1), (5.1.10) is established.
[5.2] THE LOCALIZATION PRINCIPLE 81
5.1.2. New Proofs of 2.4.1 and 2.4.4. It is worth noting that 3.2.2 and
the opening sentence of 5.1.1 combine to yield independent proofs of 2.4.1
and 2.4.4.
Consider, for example, 2.4.1. If j = 0, then (5.1.6) shows that FN *f = 0
for all N. Since, as has been remarked in 5.1.1, the FN form an approximate
identity in V, 3.2.2 shows thatf, the mean limit in V (and, iff is continuous,
the uniform limit) of F N * f as N -+ 00, is zero almost everywhere (or
everywhere).
The deduction of 2.4.4 follows similarly from (5.1.6) and 3.2.2, the former
showing that FN * f is a trigonometric polynomial for all N and all fELl.
The same arguments and sources yield results that may be interpreted in
terms of the Cesaro summability of Fourier series, but we defer this develop-
ment until Chapter 6.
5.2 The Localization Principle

The formulae collected in 5.1 can be used to show that the convergence or
summability at a point x of the Fourier series of a function f depends solely
on the behavior of f in the immediate neighborhood of x. A generalized
version of this so-called localization principle reads as follows.
5.2.1. If f and g are integrable functions and if, for a given point x, the
function
y-+f(Y) - g(y)
y-x
is integrable over some neighborhood of the point x, then
(5.2.1)
and
(5.2.2)
Proof. Since both h -+ SNh and h -+ uNh are linear operators, we may
assume without loss of generality that g = 0 throughout. Again, since
(5.1.1), (5.1.6), and 3.1.2 show that
we may take x = O.
Now, by (5.1.1),
SNf(O) = 217
1 I"
-ll f( y)D N( y) dy
= ~I"
217 -ll
{!(y) }'Sin(N
sm Y2y
+ Y2)ydy. (5.2.3)
82 THE DIRICHLET AND FEJER KERNELS jESARO SU~L\[ABILITY
If f (y)/y is integrable over some neighborhood of 0, it is easily seen that

f(y)/sin YzY is integrable over [-77, 77], since for any small IX > 0 the
ratio y/sin 'l2Y is bounded for 0 < I yl :::;; IX and cosec Yzy is bounded for
IX :::;; I yl :::;; 77. Equation (5.2.1) [with IJ = x = 0] then follows from the
Riemann-Lebesgue lemma 2.3.8 applied to the right-hand side of (5.2.3).
The least laborious way of arriving at (5.2.2) is to infer it from (5.2.1) on
the basis of the simple and general result in 5.3.1.
5.2.2. Evidently, the hypotheses of 5.2.1 hold whenever f and g agree

throughout some neighborhood of x, which is the situation to which the
localization principle refers.
5.2.3. Taking g to be a constant function 8, in which case

(N = 0, 1,2, ... ),
we see that 5.2.1 already implies that

lim 8Nf(x) = lim uNf(x) = 8 (5.2.4)
N N
whenever [f( y) - 8]/( y - x) is integrable over some neighborhood of y = x.

This is notably the case whenever f'(x) exists and 8 = f(x).
This, the first of our results on pointwise convergence and summability of
Fourier series, will be supplemented by numerous other criteria in Chapters
6 and 10; it should be compared with Dini's test in 10.2.3. Crude though it is,
it suffices to cover many specific instances. It shows that the Fourier series
of most functions which arise in problems of applied mathematics converge
to these functions at most points.
5.2.4. Remarks_ (i) One cannot in 5.2.2 replace "some neighborhood of

x" by "some set of positive measure containing x"; see [Bad, p. 465.
(ii) The localization principle (5.2.1) breaks down badly for groups Tm
(m > 1), even for continuous functions f and g. The weaker principle (5.2.2)
retains validity, at least for bounded functions f and g. See [Z2]' pp. 304-305.
For certain other groups, see MR 37 ## 1527, 5330.
5.3 Remarks concerning Summability

Although summability theory is a highly developed field of activity, our
concern rests almost entirely and solely in the use of Cesaro's method in
connection with Fourier series. Even here, moreover, its principal merit is
simply that it succeeds, in situations where ordinary convergence fails, in
recapturing at almost all points a function from its Fourier series. Our
remarks about summability are therefore very few.
[5.3] REMARKS CONCERNING SUMMABILITY 83
Consider a two-way infinite series Lnez Cn and define the partial sums
and the Cesaro (or arithmetic) means

So + ... + SN
UN = N +1
We recall (see the end of 2.2.2) that the given series is said to be Cesaro (or
(C, 1i-) summable to sum s if and only if
[There is a Cesaro method (C, a) for every a i= -1, - 2, .. " but we shall
make n,) use of this concept; see, for example [Zl]' p. 76.]
Consider as an example the series Lneze1nx. Equation (5.1.2) shows that
this series converges (to a finite sum) for no real values of x. On the other
hand, (5.1.7) shows that this same series is Cesaro-sum mabie to 0 for all real
x tc. 0 (mod 211"). It will appear in Chapters 6 and 10 that this very special
example is surprisingly significant in relation to the behavior of Fourier
series in general.
Turning to generalities, we shall first verify that the Cesaro method of
summability is stronger than, and consistent with, ordinary convergence.
5.3.1. If SN~ s, then also UN~ a (as N ~oo in each case).

Proof. Since one would expect the arithmetic means aN to behave more
regularly than the SN' this statement should occasion no surprise. In view of
the identity
(so - s) + ... + (SN - s)
N +1 = UN - S,
we may in the proof assume that S = O. Then, given e > 0, determine

No = No(e) so that ISNI ~ e for N > No. For N > No one has accordingly
So + ... + SNo SNo + 1 + ... + aN

UN = N + 1 + N + 1 '
so that
where}J = SUPN ISNI < 00. Letting N ~oo, we infer that
lim sup IUNI ~ e.
Since e is arbitrarily small, the result follows.

84 THE DIRICHLET AND FEJER KERNELS. CEsARO SUl\l~[ABILITY
5.3.2. W~ have in 5.3.1 tacitly assumed that the limit s is finite. However,
if we look at real-valued sequences SN' the result extends to properly divergent
ones. For example, the reader will verify easily that the preceding proof
adapts easily to show that, if lim SN = 00, then also lim aN = 00. More
generally, indeed, one has in all cases
5.3.3. Converse Results. The unrestricted converse of 5.3.1 is false.

Thus (5.1.2) shows that the series
2: e
neZ
ln %
diverges properly to 00 if x == O(moo 217), and is boundedly divergent for all

other values of x. On the other hand (5.1.7) shows that. the series is Cesaro-
summable to 0 for any x not congruent to 0 mod 217; if x == O(mod 217),
aN ~OO, in accord with the remarks in 5.3.2.
An easy partial converse is contained in the next result.
5.3.4. If the SN are real and increasing (for example, if en ~ 0), then Cesaro
summability implies convergence.
Proof. There is no loss of generality in assuming that SN ~ O. Then
Hence the. SN are bounded above and convergence follows.

The result also follows from 5.3.2.
5.3.5. Tauberian Theorems. There are more subtle partial converses of

5.3.1, both for Cesaro and for other summability methods of importance. In
these the positivity of the Cn (assumed in 5.3.4) is replaced by other conditions.
Since the first results of this type (applying to Abel summability) were
established by Tauber, such theorems have come to be known as Tauberian
theorems. A little more about one source of such theorems will be found in
11.2.4.
A simple such theorem (due to Hardy; see Exercise 5.8) states that if
Cn = O(I/lnJ), then Cesaro summability implies convergence. This is worth
mentioning here because it could be used in conjunction with 2.3.6 to infer
the convergence of the Fourier series of functions of bounded variation from
the Cesaro summability thereof (yet to be established in Chapter 6). Not much
economy would result from this approach, however, and we shall give a
direct proof of convergence in due course (see 10.1.4).
EXERCISES 85
EXERCISES
5.1. Suppose that fELl satisfies

wd(a) = o(lai) asa_O,
the notation being as in 2.3.7. Prove that f is equal ",lmost everywhere to a.
constant.
Hint: Consider first the case wheref is a trigonometric polynomial. Then
reduce the general case to this by using a suitable approximate identity.
5.2. Suppose that fELl satisfies
asa_O
for some a > O. Let N be any integer such that N a > 1. Prove that f*N is
equal almost everywhere to a continuous function, where f*N = f * ... * f
(N factors). (Compare with Exercise 8.4.)
Note: In the following exercises the notation is as in Section 5.3.
5.3. Show that if limN ... ",aN exists finitely, then 8N = o(N) as N _00.
5.4. . Give detailed proofs of the statements in 5.3.2.
5.5. Show that tlN == 8N - aN is expressible as
tlN = (N + 1)-1 L
I"I",N
Inlc".
Show that if
M == L InIP-
neZ
1 Ic"IP < 00
for some p ~ 1, and if limN'" ",uN = 8, then lim N ... ..,8N = 8.

Hint: Use Holder's inequality for series to show that
where A is an absolute constant. Notice that lim supN ... ..,I~NI is unaffected if,
for any k, en is redefined to be 0 for Inl :s;; k.
5.6. (I) Suppose that en _ 0 as Inl-oo, and that e" = 0, except perha.ps
for n = 0, n" (k = 1,2, .), where I :s;; n 1 < n 2 < ... and inf" nk+ljn"
== q > 1. Show that if lim.v",,,,uN = 8, then lim N ... ..,8 N = 8.
(2) Show also that the same conclusion is valid whenever en = o(Ijn).
Hint: Define tlN as in Exercise 5.5, and write
u" = sup (IC"kl, ILn/.l)j

then show that for n" :s;; N < n" + 1
L u,qr-".
Ie
I~NI :s;; 2
ral
Show that this last expression tends to zero as k _ 00.

86 THE DIRICHLET AND FEJER KERNELS. CESARO SUMMABILITY
Remark. The result (1) allows a slight generalization to the case in

which the hypotheses inf nl<+ l/n/c > 1 is replaced by
1 (nm1) '.
2:-=0-
n/c
/c>m
see [Bad, pp. 178-181.

5.7. Define, for N, k = 1,2, . "
UN./C == k-l(SN + SN+l + ... + SN+k-l)'
Verify that
and that
UN./C = SN + ""
L... (1 - Inl k- N) cn
N<)n) .. N+/C
Deduce from the former relation that if k = kN ~oo with N in such a way
that NlkN remains bounded, and if limN_",uN = s, then limN_",uN./C = s.
5.8. Suppose that Cn = O(l/lnl) as Inl ~oo and that limN_",uN = s.
Show that limN_",sN = s. (This is Hardy's theorem referred to in 5.3.5.)
Hint: Use the second relation established in the preceding exercise to
show that
for a suitable constant A. Choose k = kN suitably and employ the final

assertion in the preceding exercise.
5.9. Suppose that 1 ELl satisfies
as Z ~ zo, for some e > O. Prove that
2: J(n)e1nZo = 1(zo)'
nEZ
Hint: Use 5.2.3.
CHAPTER 6
Cesaro Summability of Fourier Series

and Its Consequences
6.1 Uniform and Mean Summability

From 3.2.2 and the fact that the Fejer kernels F N form an approximate
identity (see 5.1.1), we infer at once the following basic results about uniform
and mean summability of Fourier series.
6.1.1. If I E C", and if m is a nonnegative integer not exceeding k, then

lim IIDm(f - aNf)II", =
N-oo
o.
If I E V, where 1 ~ p < 00, then
6.1.2. The case k = 0 of 6.1.1~s especially significant. It asserts that the

Fourier series of any continuous function is uniformly Cesaro-summable to
that function. This, together with several other results of a similar nature
dealt with in this chapter, were given by Fejer in 1904. Since it was already
known by then that the Fourier series of a continuous function may diverge
at certain points, Fejer's result can be expected to have helped analysts to
breathe more freely once again. If the Cesaro method did nothing more than
this, it would amply justify its existence; as we shall see, it actually does a
good deal more.
The proof of 6.1.1 works equally well for many function spaces H other than
C" and V' (1 ~ P < (0), including at least all so-called homogeneous Banach
spaces over T. These are by definition the linear subspaces ofL1(T) which are
endowed with a norm II liB under which it is a Banach space and such that:
(i) 11/111 ~ 11/1lBforall/EH;

(ii) if IE H and a E T, then TalE Hand IITalllB = II/IIB;
(iii) lim a_ o II Tal - IIIB = 0 for aU/E H.
(See [Kz], Chapter I, 2.10 and 2.11.)
87
88 CEsARO SUM:\-[ABILITY OF FOURIER SERIES
It follows from Exercises 3.1 and 3.2 that the second part of 6.1.1 is not
true, if uNf is replaced by 8 Nf; the first part also fails. In both cases the
failure is a consequence of the fact that
itself a corollary of (5.1.10). (For the case k = 0 of 6.U, see also 10.3 below.)
For a homogeneous Banach space B over T satisfying IlenfllB = IIfllB for
all n E Z and all fEB, it is known (see [Kz], p. 49) that
lim
N-J
IIf - 8 NfJIB = 0 for all fEB
if and only ifB admits conjugation; that is to say, if and only iff E B implies
JE B, J
where the conjugate function is defined as in 12.8.1. [This last is
equivalent to demanding that, for every fEB, there is a function JE B such
that
j(n) = -i.sgn n.j(n) for all n E Z.]

The space LP (I < P < co) admits conjugation (see 12.9.3), but neither C
nor Ll admits conjugation (see 12.8.3-12.8.5).
6.1.3. Characterization of Fourier Series. At this point we can charac-

terize the Fourier series of a given fELl among all trigonometric series,
thereby fulfilling a prediction made in 1.3. Thus the case p = 1 of 6.1.1 shows
that the uNf, the Cesaro means of the Fourier series of f, converge in mean
in Ll to f. On the other hand, the Fourier series of f is the only trigonometric
series with this property.
Suppose indeed that LCnefn:& is a trigonometric series whose Cesaro means
converge in mean in Ll to f. Then, by 2.3.2, one has

lim uN(n) = j(n).
N- 00
However,
UN(n) = (1 - )~ l)c n
if InJ ~ N and is zero otherwise, so that lim",_oou",(n) = Cn' Thus Cn = j(n)

for all nEZ, and the trigonometric series in question is the Fourier series off
6.1.4. Behavior of IIf - uNf lip as N ~co. Although we know from 6.1.1
that, as N ~CO,
[6.1] UNIFORM AND MEAN SUMMABILITY 89
if 1 E V and 1 :::; P < 00, and that

111 - aNfiI.., = 0(1)
if 1 E C, these assertions cannot be improved to the extent of replacing 0(1)
by O(eN) for any fixed sequence eN -+ O.
More precisely, it is impossible to find a sequence (K N )N=1 of integrable
functions and a sequence (e.V)N=1 of positive numbers such that
(1) lim inf eN = 0
N-oo
and either
(2) 111 - KN *Illp = O(eN) for some p satisfying 1 :::; p < 00 and each
lEV,
or
(2a) 111 - KN *11100 = O(eN) for eachiE C.
On the other hand, it will be seen in Section 6.5 and Exercises 6.6 to 6.9
that integrable functions K N and positive numbers eN -+ 0 can be chosen so
that (2a) is valid for functions 1 satisfying additional smoothness conditions.
We will exhibit a proof of the impossibility of satisfying (1) and (2); an
exactly similar argument may be used to establish the impossibility of
satisfying (1) and (2a). This proof shows, by using functional analytic
techniques, that the assumption that (1) and (2) can be fulfilled leads to a
contradiction of 2.3.8. Before reading the following proof, the reader is urged
to look at Appendices A and B.1.
Suppose then that (1) and (2) are satisfied by some choice of (K N)N=1 and
(eN)N=1' We shall study the function q defined on V by the formula
q(n = sup {e;1 111 - KN*Illp: N = 1,2,..}.
According to (2), q is finite-valued on V, and from this it is almost evident
that q is a seminorm on V (see Appendix B.I.l). For each N, an application
of 3.1.6 shows that the function
I-+e;1 111 - KN *Illp
is continuous on LP. Hence (see Appendix A.4) q is a lower semicontinuous
sl'minorm on V. The crucial step is to apply Appendix B.2.1(1), using
therein E = V and Pk = q for k = 1,2" .. ; since LP is complete (see 2.2.4),
this application is justified and leads to the conclusion that q is continuous on
Lp. This signifies the existence of a constant c such that q(f) :::; c 111 lip for
all 1 E v'. In other words, we have
(6.1.1)
for 1 E V and N = 1,2" . '. If herein we choose to take 1= en, we shall

obtain thereby the relation
(6.1.2)
90 CEsARO SUMMABILITY OF FOURIER SERIES
for all n E Z and N = 1,2, . '. Therefore, in view of (1), it will appear from
(6.1.2) that
I)(N(n) - 11 ~ %
for all n E Z and infinitely many N. For any such N, it thus appears that
lim inf IKN(n)i ~ %> 0,
Inl-'"
which contradicts 2.3.8. This contradiction terminates the proof.

Further illustrations of this type of proof will be discussed in detail in
Section 10.3.
6.2 Applications and Corollaries of 6.1.1

6.2.1. Comments on 6.1.1. Before pressing on to some refinements of6.1.1
contained in Sections 6.3 and 6.4, we shall make a number of deductions
from 6.1.1, each of which is of considerable importance.
Since it is evident that 6.1.1 refines the trigonometric polynomial approxi-
mation theorem of 2.4.4 to the extent of specifying an algorithm for the
construction of trigonometric polynomials approximating a given function,
it may well suggest two lines of thought, namely:
(1) Real analysis contains another famous approximation theorem, to wit,
that of Weierstrass. This refers to the approximation, uniformly on a compact
real interval I, of continuous functions on I by ordinary polynomials. In
6.2.2 and 6.2.4 we shall show how this theorem is deducible from 6.1.1 and
indicate a common source of both theorems.
(2) What can be said in general about approximation by trigonometric
polynomials~ For a givenj, how good an approximation is afforded by UNj in
comparison with other trigonometric polynomials of degree at most N? We
shall come round to a brief discussion of these matters in Section 6.5.
Meanwhile, 6.2.5 to 6.2.8 will be concerned with some deductions from 6.1.1
more directly concerned with Fourier series.
6.2.2. The Weierstrass Polynomial Approximation Theorem. This
asserts that. iff is a (not necessarily periodic) continuolls function on a compact
interval [a, b] of the line, then f is uniformly approximable on [a, bJ by
(ordinary) polynomial functions.
In proving this on the basis of 6.1.1, we may without loss of generality
assume that [a. b] is [ - 1T, 1T]' Then a constant c may be chosen so that f - ex
takes the same value at -1T as at 1T, and can therefore be extended into a
periodic continuous function. It is evidently sufficient to show that this
modified function is uniformly approximable on [-1T, 1T] by polynomials. Thus
we may assume from the outset that f is periodic and continuous.
Given any e > 0 we choose N so large that
[6.2] APPLICATIONS AND COROLLARIES OF 6.l.l 91
The trigonometric polynomial
uNf(x) = L en e1n%,
Inl';'"
can in turn be uniformly approximated on [ - 1T, 1T] to within % e by ordinary

polynomials, to do which it suffices to replace each exponential e1n % by a
sufficiently large number of terms of its Taylor expansion about the origin.
the latter series converging uniformly on any compact set. The result is a
polynomial function P such that
uniformly for Ixl ~ 1T. But then
I(J - I
P)(x) ~ e
uniformly for Ixl ~ 1T, and Weierstrass' theorem is established.
6.2.3. Other Proofs of Weierstrass' Theorem. There are many other

proofs of Weierstrass' theorem, both "classical" and "abstract-modern" in
flavor; the latter are certainly the more enlightening.
It was M. H. Stone who, in 1937, first undertook an abstract analysis of the
status of Weierstrass' theorem and its close relatives. His work and subsequent
developments laid bare the anatomy of the situation and have resulted in very
general approximation theorems concerning closed subalgebras of the Banach
algebra (with pointwise operations) of continuous functions on any compact
Hausdorff space; these algebras will be encountered again in 11.4.1. Both 2.4.4
and \Veierstrass' theorem are contained as very special cases within this
scheme. For a recent survey, see Stone's article "A generalization of Weier-
strass' approximation theorem" appearing on pp. 30-87 of [SMA]; see also [E],
Section 4.10, [HS], pp. 94-98, and [L 2 ], Chapter 1.
6.2.4. Bernstein Polynomials. In just the same way that 6.1.1 includes
and refines 2.4.4, a famous theorem of Bernstein includes and refines the
Weierstrass theorem in 6:2.2. Bernstein's theorem asserts that, if f is a con-
tinuous function on [0, 1], then the associated so-called Bermtein polynomial8
..,
B..,f(x) = 2: f(nfN) ..,Cn xn(l - x)N-n (N = 0, 1, 2, ... )
n=O
conv('rgc to f uniformly on [0, 1].

Th('rc is a very ('xt('nsivc literature dealing with Bernstein polynomials; for
a start, the interested reader should consult [Ka], pp. 52-59, [Ld, and [L2 ],
Chapter 1.
We now turn to some deductions from 6.l.1 more closely connected with
Fourier series.
6.2.5. Suppose that f, g E L1 and that the Fourier series of g is essentially

boundedly convergent almost everywhere, that is, that SUPN IlsNgl[., < 00
and that limN_",sNg(x) exists for almost all x. Then
2 II
7T
f(x)g(x)dx= Lf(n)"g(-n),
neZ
"'. A
the series being convergent.

Proof. By hypothesis we have
[SNg(X)[ ~ M a.e.,
M being a number independent of N and of x, and limN_oosNg(x) exists almost
everywhere. By 5.3.1 and 6.1.1, limN_oosNg(x) = g(x) almost everywhere.
Consequently,
2~ I f(x)g(x) dx = lim 21 If(X)SNg(X) dx

N-aJ 11'
lim
N-oo
2:
Inl",N
j( -n) g(n),
passage to the limit under the integral sign being justified by Lebesgue's
theorem ([W], Theorem 4.lb). From this the stated results follow.
6.2.6. Remarks. (1) By 2.3.5, the hypotheses on g are certainly fulfilled

whenever g E C2. Consequently, 6.2.5 justifies the characterization of the
Fourier series of f E L1 among all trigonometric series mentioned in (D) of
1.3.2.
(2) The conclusion of 6.2.5 may be derived from the mere assumption that
SUPN [[sNgl[ '" < 00. Thus, Appendix B.4.1 may be used to show that there is
at least one subsequence (SNJ]) which converges weakly jn LOO. Moreover, if
(SNJ]) is any subsequence converging to h weakly in 1''', and if U E T, then
2~ I U(X)SN,p(X) dx = Lf u(x)g(x) dx
for all large k; hence
2~ f u(x)h(x) dx = 2~ f u(x)g(x) dx
and so h = g a.e. It follows that SNg _ g weakly in Loo. In view of Appendix

C.I, this is equivalent to the desired conclusion.
6.2.7. Remark. The formula appearing in 6.2.5 is one variant of the

so-called Parse val formula, a prototype version of which has appeared in
Exercise 1.7, and to which we shall return in Sections 8.2 and 10.5 with
different hypotheses onf and g.
It is to be observed that the series appearing in 6.2.5 is not convergent for
allf E L1 and all continuous g; see Exercise 10.7. On the other hand, 6.1.1 is
[6.2] APPLICATIONS AND COROLLARIES OF 6.1.1 93
easily seen to imply that the series .is Cesaro-summable to (lj21T) f fg dx

whenever J E V and gEL"'; see Exercise 6.2.
One very special but important case of 6.2.5 demands closer examination,
since it leads to the conclusion that, no matter how badly the Fourier series
of JELl may behave in respect of pointwise convergence (see 10.3.4), yet
nevertheless one may always legitimately integrate term wise this Fourier
series.
6.2.8. If J E L\ then
f b
a
J(x) dx = 2: j(n)
neZ
elnb
:-
~n
e 1na
,
the term corresponding to n = 0 being understood to meanj(O)(b - a).

Proof. Owing to the meaning assigned to the term corresponding to
n = 0, it is sufficient to establish the formula for the case in which a = 0
and 0 < b < 21T. Let y be the function equal to 1 on the interval [0, b), zero
elsewhere on [0, 21T), and extended so as to have period 21T. A direct computa-
tion shows that
e -Inb - 1 e - Vzibn sin %bn
g(n) = . = , (6.2.1)
-21T~n Tm
the right-hand side being understood to mean bJ21T when n = O. Consequently

one finds after some reduction that
b ~ sin n(x - b) - sin nx

8N Y()
x = - + L.. .
21T n=l Tm
Reference to Exercise 1.5 confirms that this series is boundedly convergent

for all x. Also, by 5.2.3, the limit of 8 NY(X) is g(x) provided 0 < x < 21T and
x i= b, since g is constant on some neighborhood of each such point x. Hence
([W]. Theorem 4.lb)
[211
JobJ(x) dx = J211
0
J(x)g(x) dx = lim Jo J(X)' 8NY(X) dx
N-oo
2: g(n) [1l1l J(x)e1n:z: dx

Inl .. N Jo
2:
Inl .. N
g(n)' 21Tj( -n)
2:
Inl .. N
j(n)' 21Tg( -n).
Using (6.2.1), the desired result appears.

Remark. This result (and even a little more) will be obtained on the
basis of more general theorems in 10.1.5.
94 CESARO SUl\1l\IABILITY OF FOURIER SERIES
6.3 More about Pointwise Summability

We shall here deal with some refinements of the first half of 6.1.1.
It is convenient to introduce some notation. Given a point x and a number
8, we write
f~( y) == f~( y, x) = Y2[f(x + y) + f(x - y) - 28], (6.3.1)
in terms of which (5.1.6) and (5.1.8) lead to
(6.3.2)
It is our aim to give simple conditions sufficient to ensure that the expression
(6.3.2) tends to zero as N -,)000. In studying this we shall not assume from the
outset that 8 is the" right" value, namely f (x). Indeed, it is the behavior of
n in the neighborhood of y = 0 (that is, of f in the neighborhood of x) which
is significant, and the prime feature is not the value ofn at o (that is, of f
at x) but rather the limiting behavior of f~ near 0 (that is, of f near x). In
this section we consider the simplest case, in which we assume outright the
existence of the limit off~(y) as y-,)o +0.
6.3.1. Suppose that fELl and that

f(x + 0) +f(x - 0) == lim [f(x
]1- +0
+ y) +f(x - y)] (6.3.3)
exists finitely. Then

lim uNf(x) = Y2[f(x + 0) + f(x - 0)). (6.3.4)
N-",
The limit in (6.3.4) is attained uniformly on any set on which the limit in
(6.3.3) is attained uniformly.
Proof. The formula (6.3.2) will be applied, taking therein
8 = Y2[f(x + 0) + f(x - 0)).
The hypotheses signify that n(

y) -,)0 0 as y -,)0 + O. Given any e > 0, there
exists therefore a number 0 > 0 such that If~(y)1 ~ e for 0 ~ y ~ o.
This number 0 will depend upon x and e, but can be chosen uniformly with
respect to x when the latter varies in any set on which the limit in (6.3.3) is
attained uniformly. The integral in (6.3.2) is then expressed as a sum
say. Then
[6.4] POINTWISE SUM)IABILITY ALMOST EVERYWHERE 95
by (5.1.8). (Notice that we are here using positivity of F 1'1: this is not available
for DN and the substitution of IDNI for DN would vitiate the argument
irreparably.) On the other hand, the use of (5.1.9) leads to
If t: and S are held fixed while N is allowed to tend to infinity, it is seen that
lim sup IUNf(x) -

N-oo
81 :s;; t: + lim sup 1121 =
N-~
t:.
Since t: is a freely chosen positive number, the desired conclusion now follows.
6.4 Pointwise Summability Almost Everywhere

So far we have depended solely on the fact that the Fejer kernels form an
approximate identity. By using somewhat more special properties, combined
with a fundamental theorem of Lebesgue related to the differentiability of
indefinite integrals, we can establish the pointwise summability almost
everywhere of the Fourier series of any integrable function. Before proving
this result, we shall review the auxiliary requirements.
6.4.1. Auxiliary Inequalities. Concerning F 1'1 we observe two inequalities.

Temporarily using A, B to denote absolute constants, the first inequality
reads
(6.4.1)
which is easily established by examining separately the ranges 0 :s;; Y :s;; TT/ N
and TT/N :s;; Y :s;; TT; in the first interval, F N( y) is majorized by a multiple of
N, and in the second by a multiple of N-ly-2. This inequality shows in
turn that
o :s;; FN(y) :s;; F~(yl :s;; ~ (0 < y :s;; TTl. (6.4.2)
We shall also need the following consequence of (6.4.1):
(6.4.3)
6.4.2. Regarding the theorem of Lebesgue it suffices to recall that for any
integrable f, periodic or not, and almost all x it is true that
f If(x + y) - f(x)1 dy = o(h)
f If(x - y) - f(x)1 dy = o(h)
as h ~ +o. For a proof, see [HS], p. 276.

Using the notation introduced in Section 6.3, it follows that for almost
all values of x one has
f In( y)1 dy = o(h) (h~ +0), (6.4.4)
when s is taken to be f (x). Such points x are usually termed Lebesgue points
of f and the set of such points the Lebesgue set of f.
Now we can state and prove the main theorems of this section.
6.4.3. Iff E Ll, and if for a given x and s one has
50" If~(y)1 dy = o(h) (h~ +0), (6.4.4)

then
lim uNf(x) = s. (6.4.5)
N-J
Proof. We start again from the formula (6.3.2), namely,
Assuming that (6.4.4) is satisfied, we suppose that e > 0 is given and choose
8 > 0 so that
o~ J(h) == f In(y)1 dy ~ eh (0 ~ h ~ 8). (6.4.6)
By (6.4.1),
1-1 J6 n(y)FN(y) dYI ~ -I J6 If!(y)IF~(y) dy.

~ 0 ~ 0
Partial integration and (6.4.6) show that this is majorized by
~-lJ(8)F~(8) _ ~-1 [6 J(y)Fr(Y) dy

Jo 6
~ ~-le8 F~(8) + ~-le 50 yIFt'(y)1 dy

~ YzA~-le + B~-le
[6.4] POINTWISE SUMMABILITY ALMOST EVERYWHERE 97
by (6.4.2) and (6.4.3). Thus
I.! (6 f~(y)FN(Y) dyl ~ %A7T- le + B7T- l e,

7T Jo
and therefore
laNf(x) - 81 ~ 7T- l (%A + B)e + 1-7T1 fn6 f~(y)FN(y)dyl (6.4.7)
Now, for any S > 0, the inequality (5.1.9) shows that the integral appearing
in (6.4.7) tends to zero as N ---+00. So, keeping e and S fixed, we infer from
(6.4.7) that
lim sup laNf(x) - 81 ~ 7T- l (%A + B)e.
N->
Finally, letting e ---+ 0, (6.4.5) follows.
6.4.4. If f E L1, then

lim aNf(x) = f(x) (6.4.8)
N-co
holds for almost all x.

Proof. This is immediate on combining the closing remark in 6.4.2
with 6.4.3.
6.4.5. If a trigonometric series 2:c nefn % is Cesarosummable almost every-

where to a sum f(x), and if this series is a Fourier-Lebesgue series, then
f E L1 and the series is the Fourier series of f.
Proof. By hypothesis the series is the Fourier series of a function
g ELl. By 6.4.4, the Cesaro means aN of the given series converge almost
everywhere to g. Hence g = f a.e., .and the stated results follow.
6.4.6. Remark. In connection with 6.4.5,it must be remarked that a

trigonometric series 2:cne1n:r may well be CesiLro-summable almost everywhere
to an integrable sum and yet fail to be a Fourier-Lebesgue series. Thus the
series 1 + 22::'=1 cos nx is Cesaro-summable to 0 for every x :t: 0 (mod 211)
[as appears from (7.1.1) and (7.1.2)], but 2.3.8 shows that it is not a Fourier-
Lebesgue series. (It is, however, the Fourier-Stieltjes series of the Dirac
measure e; see 12.2.3 and 12.5.10.)
6.4.7. The Majorant Function a*f. It is worth observing that, if g E Lao,

then (5.1.8) gives at once for all x the inequality
(6.4.9)
In view of (6.4.9) it is interesting to consider the majorant function
a*f(x) == sup la.vf(x)1 (~ 00). (6.4.10)
N
Naturally, u*J is a nonnegative measurable (in fact, lower semi continuous)

function. By using special properties of the F N it may be shown (see [Zd, pp.
154-156; Edwards and Hewitt [1], Theorem 3.1) that this majorant satisfies
the following integral inequalities:
.
Ilu*J lip ~ ApllJ lip if J E LP and p > 1, (6.4.11)
Ilu*Jllp ~ ApllJIIl if JELl and 0 <p < 1, (6.4.12)
Ilu*J III ~~ f IJllog+ IJI dx + B

if J log + IJ I E Ll . (6.4.13)
In these inequalities A and B denote absolute constants, Ap depends upon p

only, and log+t = log (max (1, t)).
Concerning the analogous assertions applying to the majorant
s*J(x) == sup
N
ISNJ(x)1 (6.4.14)
of the ordinary partial sums of the Fourier series of J, see 10.3.5 and 10.4.5.
6.4.S. The Estimate (6.4.9). The relation (6.4.9), valid for g E V", is close
to being the best possible. Indeed, if we are given any set E of measure zero,
there exists a nonnegative function f belonging to LP for every p < 00 and
such that
(x E E). (6.4.15)
To verify this we may assume that E lies in (0,21T) and choose numbers
a" > 0 and c" ~ 0 (k = I, 2, ... ) so that
(6.4.16)
for every p < 00; for example, c" = log k, a" = k - 2. Since E has measure
zero, we may choose open sets E" such that E c E" c (0, 21T) and having
measure m(E,,) E; ale' Let fIe denote the characteristic function of E,,, extended
by periodicity, and put
Then f ELP for every p < 00, as follows from the second clause of (6.4.16) and
Beppo Levi's theorem ([W], Theorem 4.1e). Observe also thatf is nonnegative
and lower semicontinuous. According to 6.3.1,
lim uNf,,(x) = 1
N~"
for all x E E" and, a fortiori, for x E E. Since FN is nonnegative, one has for all
x and all k
[6.5] APPROXIMATION BY TRIGONOMETRIC POLYNOMIALS 99
and so for aU x E E and all k

lim inf ClNf(x} ~ c".
N~'"
This, together with the first clause of (6.4.16), yields (6.4.15).

See also Exercise 6.18.
6.4.9. Remarks on Conver~ence. By combining Exercises 5.5 and 5.6
with 6.4.4 one can specify some simple conditions onf e Ll sufficient to ensure
that
a.e.; (6.4.17)
see Exercises 6.12 and 6.13, and also 10.1.1, 10.1.4, 10.2.1, and 10.2.3. As will
appear in 10.3.4, the relation (6.4.17) is hopelessly over-optimistic for general
feLl.
6.5 Approximation by Tri~onometric Polynomials

In this section we shall take a look at the issues enumerated in 6.2.1(2).
This involves an excursion around the fringes of the general question of
approximation by trigonometric polynomials; for major inroads into the
general theory the reader must turn to the relevant portions of [L 2 ], [Z],
[Ba], [Ti], and [BK].
6.5.1. The Functionals PN and EN' For definiteness we shall work within
the Banach space C of continuous functions, but the reader will scarcely
need to be told that each question posed in this setting has some sort of
analogue for the case in which C is replaced by V' (or, indeed, by anyone
of a number of other quite natural function spaces).
In order to examine the questions raised in vague terms in 6.2.1(2), we
introduce two sequences of functionals PN and EN (N = 0,1,2, .. ) defined
on C in the following way:
(6.5.1)
where TN denotes the set of trigonometric polynomials of degree at most N.
Plainly, the relative magnitude of PHf and EHf provides a sensible measure
of just how good UHf is as an approximant to f, when compared with other
elements of T H'
It is evident that
EN+d:!O; EHf " PH!.
and that PN! = 0 if and only if! is a constant function. Moreover (see
Exercise 6.5), the infimum EH! is actually an a.ssumed minimum; as a
consequence it follows that E Hf = 0 if and only if! e T H'
According to 6.1.1, PH! = 0(1) as N -00 for each f e C; yet, by 6.1.4, the
rela.tion PNf = O(eN) is false for some f e C (in fact. for a nonmeager set of
IE C) whenever E,y = 0(1) is given. It is also easily shown (see Exercise 6.10)
that the relation p,yl = o(ljN) holds if and only if I is a constant. This
simple result is noteworthy when it is compared with Exercises 6.8 and 6.9;
it then shows conclusively that, for sufficiently smooth nonconstant con-
tinuous functions I, aNf is far from being the optimal approximant to f
among all elements of TN. If f is very smooth, sNf is a decidedly better
approximant to I than is aNf. Crudely speaking, the great advantage of aNf
is to be seen for general.continuous functions f.
After these preliminary remarks we now proceed to establish some
improved estimates for PNI in case I satisfies certain Lipschitz (or Holder)
conditions. This latter type of condition is for our immediate purposes best
expressed in terms of a modified modulus of continuity, namely,
compare with (8.5.1) and the definition of wd in 2.3.7. Applying (6.3.1) and
choosing s = I(x), we see that
so that (6;3.2) yields
(6.5.2)
Since (6.4.1) shows that

A
FN(y) ~ Ny2
for a suitable absolute constant A, (6.5.2) entails that
PN f ~
AN IllN
7T 0
r\
Ua:
I( Y)dY + ~
N
7T
I"
lIN
O",/(y) dy
Y
2 (6.5.3)
Further progress will be facilitated by appeal to the following simple result,

the proof of which will be left to the reader.
6.5.2. Suppose that a and fJ are nonnegative functions defined on some

interval (0, c), where c > 0, and integrable over (c', c) for each c' satisfying
o < c' < c. Suppose further that a( y) = o[.B( y)] as y -. +0, and that
f fJ(y) dy-.oo as t-. +0.
Then also
f a( y) dy = O[f fJ( y) dy] as t-. +0.
Here now is the main result of this section.

[6.5] APPROXIMATION BY TRIGONOMETRIC POLYNOMIALS 101
6.5.3. Suppose that the function w is defined, nonnegative, and increasing

on some interval (0, e), where e > 0, while y-aw(y) is decreasing on the same
interval for some choice of a satisfying < a < 1. Suppose further that f is
continuous and satisfies
D.",f(y) = O[w(y)] (respectively,o[w(y)]) as y~ +0. (6.5.4)
Then
as N ~ ct) (6.5.5)
Proof. If e < 1T, and if we extend w by setting w( y) = w(e) for e < y ~ 1T,
this extended function satisfies the required conditions on the interval
(0, 1T]. Thus, we may as well assume from the outset that c = 1T.
Now, without appeal to 6.5.2. it follows from (6.5.4) that
{liN (liN
N Jo D.",f( y) dy = 0 (respectively, 0) (N . Jo w( y) dy)
= 0(respectively, 0) (N'N-l.w(~))
= 0 (respectively, 0) (w(~)), (6.5.6)
the second step being justified since w is increasing.

To handle the second term on the right of (6.5.3), we apply 6.5.2, taking
a( y) = D.",f( y)/y2 and f3( y) = w( y)/y2. Unless w is identically vanishing
(in which case f is constant and nothing remains to be proved), the decreasing

character of w( y)/ya for some a satisfying < a < 1 ensures that f3 satisfies
the hypotheses of 6.5.2. We thus obtain
N-l. fit D.",f(Y,} dy = N-l. 0 (respectively, 0) (fit w(y)2 dy )

liN y liN Y
= 0 (respectively,o)[N-lf y-a w (y)ya- 2 dy]
lt
liN
= 0(respectively,o) [N-l. (N-l)-a w(~)

(" y a - 2 dY] ;
JIIH
the last step is justified since y-a w ( y) is decreasing. Evaluating the remaining
integral and simplifying, one obtains
N-l. L~H D.Jf~) dy = 0 (respectively, 0) [w(~)]. (6.5.7)
It now remains but to combine (6.5.3), (6.5.6), and (6.5.7) in order to derive
(6.5.5).
6.5.4. Remarks (1). The majorization given in 6.5.3 is. at least for certain
natural choices of w, the best poBBible (see [Bad. p. 206). Somewhat similar
results are known for functions! which satisfy a mean Lipschitz condition of
the type IITJ - flit = O(iala) as lal->- 0; see [Zd, p. 117. See also MR 53
# 6203.
(2) Inasmuch as EN! "" PN!, Subsection 6.5.3 and its analogues yield majorants
for EN! for restricted functions!. Superior results are obtainable by estimating,
not
PN! = Ilf - UN! II "',
but rather II! - 'TN!II"" where
'TN! = 2U2N-d - uN-deT2N - l ;
some calculations of this nature are proposed in Exercises 6.6 to 6.9. More
elaborate results appear in Timan [1]. One might also use to the same end
the socalled Jackson polynomials J N * f, where
J N(X) = sin
eN ( .
%N'X)'
1; ,
sm ZX
where N' = [%N] + 1 and where the number eN is chosen to make IIJNlll = 1;
see [L 2 ], pp. 55-56. See also W. R. Bloom [3], [4].
6.5.5. Converse Results. It should also be mentioned that many results

in the reverse direction are known: given the possibility of approximating f
with a given degree of accuracy by trigonometric polynomials of degree at
most N, one can infer smoothness properties of f. The earliest such results
appear to be the work of Bernstein (1912) and of de la Vallee Poussin (1919);
since then the subject has been studied vigorously (see [L 2 ], Chapter 4, and
[BK], pp. 45-59 and the references cited there). A very special instance of this
type of result appears in Exercise 6.10. A crucial role in these investigations is
Bernstein's inequality (see Exercise 1.9). the basic reason being that this
inequality combines with the fir.st mean value theorem to yield an estimate for
the modulus of continuity, IITat - til",. of a trigonometric polynomial t in
terms of Iltll", and the degree Nof t.
A sample result asserts that if fee is such that ENf = O(N-a), where
o < IX "" 1, then
O(lala ) ifa < 1,
{
IIT"j-jll",= O(lalloglal- l ) if a = 1;
this is very close to being a converse to a special case of 6.5.3. Other results
infer the existence of several continuous derivatives off. together with estimates
of their iterated differences; see [L 2 ], loco cit., and [BK], pp. 45-57. 72-88. See
also MR 54 ## 832, 13433; 55 # 960; and Zamansky [1]. [2].
6.6 General Comments on Summability of Fourier Series

Cesaro summability has so far received all our attention, but we should
mention in passing the Abel (or Abel.Poisson) method, which is an equal
favorite. See also Exercise 6.14 for yet another method of great importance
in the theory of trigonometric aeries.
[6.7] REMARKS ON THE DUAL ASPECTS 103
The Abel means of the Fourier series of! are the functions
Arf(x) = Lr
neZ
lnl j(n)e tnx
= P T *!(x), (6.6.1)
where the continuous parameter r satisfies 0 ~ r < 1 and where P T is the

so-called Poisson kernel
1 - r2
PT(x) = 1 _ 2r cos x + r2
encountered in the Poisson representation formula for harmonic functions.
Abel summability of the Fourier series of! refers to the limiting behavior of
the means AT! as r --+ 1 (from below).
All thlJ results proved in this chapter about Cesaro summability of Fourier
series remain true in respect of their Abel summability. (The reader is urged
to verify this statement as an extended exercise.) In a few regards (which
will nowhere concern us), the Abel method is slightly to be preferred, partly
because of its evidently closer connections with complex variable theory;
for details, see [ZtJ, Chapters III and VII; [Z2], Chapter XIV. See also
Exercise 6.16.
Even for quite general (locally compact Abelian or compact) groups, there is
indeed an unlimited number of summability methods, each expressible in terms
of a limiting process lim K, * j, and each just as effective as the Cesaro or Abel
methods in respect of normconvergence in C or in LP( 1 ~ P < co). In each
case, (K,) is a sequence or net of kernels, usually forming an approximate
identity in Ll and composed of very wellbehaved functions. The fun begins
when. one wishes to examine the associated problem of pointwise almost every-
where summability for discontinuous functions, concerning which surprisingly
little is yet known (except for the circle group T, R, and their finite products).
Some progress is reported by Stein [1] for the case of compact groups and
spaces V' with 1 ~ P ~ !; lIomewhat weaker results with a wider range of
applicability are discussed by Edwards and Hewitt [1]. These results apply in
fact to sequences of m-operators of type (LP, V) (1 ~ P < co); see 16.2.7 and
16.2.8.
6.7 Remarks on the Dual Aspects

The results in this chapter and in Chapter lO to follow can both be regarded
as investigations bearing upon how one may interpret, in a pointwise sense,
the" inversion formula"
! = (jr" (6.7.1)
where! is a given integrable function on G = T. In this chapter we have

concentrated on the case in which the second Fourier transformation in
(6.7.1) is interpreted via pointwise summability methods; in Chapter 10 the

interpretation is via pointwise convergence (see the discussion at the end
of 2.2.2).
In Chapters 8, 12, and 13 it will appear that the same inversion formula
can be interpreted by using mean and distributional convergence.
There remains still the dual inversion formula, to wit,
(6.7.2)
where now 4> is a given function on Z. Here again, the difficulties concentrate
around the definition of ~; see Section 2.5. As usual, the adoption of point-
wise convergence or summability as the means of defining ~ raises problems
(except in the transparent case in which 4> Ell). These thorny features largely
evaporate if mean or distributional convergence is used, as will be done in
8.3.3. 12.5.4, and 13.5.1(2).
EXERCISES
6.1. Suppose that f E LP (1 ~ P ~ 00) and that 2:c"el"x is a trigonometric

series with partial sums 8 N whose Cesaro means are aN. Prove that if
then
This is an analogue, for mean convergence, of 5.3.1.

6.2. Suppose that fELl and gEL"'. Show that
-.!..ffg
21T
dX = lim L
N_oo l"I<iN
(1 - N 1n1+ l)i{n)y(-n).
6.3. Write, for f E L1,
m(S) = ess. sup If(x)l. m = lim m{S).

Ixl .. 6 610
Prove that
6.4. Is
( _1)"e 1nx
~ (1 + Ini)log (2 + n 2 )
EXERCISES 105
the Fourier series of a bounded (measurable) function? Give reasons for

your answer.
Note: The next five exercises form a sequence concerned with approxima-
tion by trigonometric polynomials; they should be attempted in order of
enumeration.
6.5. For IE C write, as in (6.5.1),
ENI = inf III - til <Xl ,
the infimum being taken with respect to all trigonometric polynomials t of
order at most N. Prove that this infimum is assumed.
6.6. Assuming the formula
uN-d(x) = -I fOO I(x + y) 2 sin2 YzNy' NdY2

~ -<Xl Y
for lEe (see [Zl]' p. 92), verify that for N ~ I
TNI == 2U2N-d - uN-d
is given by
TNI(x) =:;;:2f<Xl
0
[ f ( x + Ny) + f ( x Y)] k(y)
- N ~. dy
where
h(y) = sin 2 y - sin 2 Yzy = Y2(cOS y - cos 2y).
6.7. The notation being as in Exercise 6.6, define
(i = 1,2, .. ; y > 0).
Verify that
(I) fa<Xl IHt(y)i dy < 00 (i = 0, 1,2,),
(2) Hi(O) = 0 if i > 1 is odd.

Hints: For (I), check that Ht(y) = O(y-2) as y-+oo. For (2), apply
the results of Exercise 6.6 taking in turn f (x) = 1 and I (x) = cos x. Taking
x = 0 yields
-4f<Xl H o( y) dy = 1,
1T 0
~ fa<Xl cos (1)' Ho(Y) dy = 1.
Integrate the second relation by parts and use the first to obtain
fa<Xl cos (1) .H (y) dy

2 = O.
Now let N -+00 to conclude that H 3 (0) = O. Similarly,

H5(0) = f" H 4 (y) dy = 0,
and so forth.
106 cEsARO SUMMABILITY OF FOURIER SERIES
6.8. Suppose thatl E C" (k an integer;?; 0) and that N is a positive integer.

Show that
Ihd - III", ~ A,,' IID"III",' N-",
and deduce that
ENI ~ A~' IID"III", . N-",
where A" and A~ depend only on k.
Hints: From Exercise 6.6,
TNI(x) - I(x) = ~ Jo<O [f( x + ;) + I( x - ;) - 2/(X)]'~(~2dY.

Apply partial integration k times and use the results of Exercise 6.7.
6.9. Suppose that I E C" (k an integer ~ 0) and show that
IITNI - III", ~ B,,' Qoo(D"J)(~)' N-'''.

Deduce that
ENI ~ B~' Q<o(DkJ)(~) N-".
Here Bk and B~ depend only upon k, and
Qpg(S) = sup IITa!1- gllp (1 ~ p ~ 00).
lal" 6
Hints: Consider 16(x) = (2S) -1 f~61(x + y) dy where S > O. Show that
IID1<+ 11611", ~ (2S)-1 Q",(DkJ)(20) ~ 0- 1 Q",(D"J)(o) ,

and
IIDk(f - 16)11", ~ Q",(DkJ)(O).
Since Tid = TNI6 + TN(f - 16)' the desired results follow on using Exercise
6.8 and taking 0 = 27T/N.
6.10. Suppose that IE Ll and 11/ - uN/Ill = O(N-l) as N _00. Show
that I is equal almost everywhere to a constant.
Show also that if IE Ll and
E N(1)j = inf {III - till: t E TN}'
then
Ij(n)i ~ El!l-l / (n = 1, 2, ... ).
Remark. The first part of this exercise asserts that the sequence of
operators / _ uNI (N = 1,2, .. ) on C is "saturated by the function
~(N) = N-l"; for this concept, see [L 2 ], pp. 98-~02. See also MR 36 # 5605.
6.11. Let Lnezcne1n.r be a trigonometric series and
UN = L
Inl"N
(1 - Inl/N + 1)cne 1n.r
its Nth Cesaro mean. Show that the given series is a Fourier-Lebesgue series
if and only if
lim IluN - UN' III = O.
N.N'-.:o
(Compare the results in 12.7.5 and 12.7.6.)

EXERCISES 107
6.12. (Fatou's theorem) Show that if fELl and
N-l 2:
Inl<iN
Inj(n)I = 0(1) as N --+00,
then limN_.,sNf(x) = f(x) for almost all x; and that if, furthermore, f is
continuous, then limN_",sNf = f uniformly.
H?:nt: Use Exercise 5.5 and the results appea~ing in 6.1.1 and 6.4.4.
Note: One can show, by using the same method in combination with
8.3.1, that if Cn = o( I/Ini) as Inl --+ co, then the trigonometric series Lnezcnelnr
is convergent for almost all x.
6.13. Suppose that fE Ll and that j(n) = 0 save perhaps for n = 0,
nt, n 2 ,' " where 0 < nl < n 2 < ... and inf nk+1/nk > I. Prove that
limN_",sNf(x) = f(x) for almost all x; and that if, in addition,J is continuous,
then limN_",sNf = f uniformly.
Hint: Use Exercise 5.6, and Subsections 6.1.1 and 6.4.4.
Note: These so-called lacunary series will receive further attention in
Chapter 15; see especially the remarks in Section 15.6. The result, like that in
Exercise 5.6, admits some generalization.
6.14. Let the function rN (N = 1, 2, ... ) be defined to be equal to l7N in
[ -liN, liN), to 0 in [-17, -liN) and in (liN, 17), and be defined elsewhere so
as to be periodic. Put RN = rN * rN. Verify that (RN);=l is an approximate
identity. Compute flN and deduce that
uniformly for each continuous f, provided sin 0/0 is interpreted as L

Note: This is a case of Riemann's method of summability, which is of
fundamental importance in the general theory of trigonometric series; see
[Zl]' Chapter IX and [Ba l ), p. 192.
6.15. Show that a necessary and sufficient condition that a continuous
functionf be of the form} = F 0 el , where ei(x) = elr and F is defined and
continuous on the closed unit disk in the cOD:lplex plane and holomorphic in
the interior of this disk, is that j(n) = 0 for n E Z and n < 0 (compare
Exercise 3.9).
6.16. Suppose that fELl and define
F(x) = f:/(y) dy.
Prove that, if x is a point for which the symmetric derivative
D F( X) _=1l'm F(x + e)2 - F(x - e)

.. 0 e
exists finitely, then

lim Arf(x) = DsF(x) ,
T-1-0
Arf being defined as in Section 6.6.

Remark. The relation D.F(x) = 1(x) holds at all points x of the Lebesgue
set of 1 (see 6.4.2), but it may well hold at other points as well. Compare the
result with 6.4.3.
Hints: Assume, without loss of generality, that F(l7) = 0. Integrate by
parts in the formula (6.6.1) to obtain
Ar1(x) = Lf" (2siny)-1{F(x + y) - F(x - y)}Qr(y)dy,
where Qr( y) = - sin y P~( y). Verify that Qr ~ 0, IIQr !Il = r, and that
lim
r-1 - 0 6"
sup Qr( y) =
Iyl""

for any 8 E (0, 17]. Mimic the proof of 6.4.3.
6.17. Sometimes (see 6.4.7, 12.9.9, 13.8.3, and 13.10, Exercises 13.21 to
13.23; [Zd, pp. 170-175; [Z21, pp. 116-119, 158) one wishes to consider the
set L ~ of measurable functions 1 such that <1>( I1 I) E 11, <1> being a fixed non-
negative function of a suitable type, and one wants to know that the set T
of trigonometric polynomials is contained and everywhere dense in L~ in
the sense that, given any 1 E L~ and e > 0, there exists t E T such that
N(1f - til == 2~ f <1>(11 - til dx < e. (*)
Suppose that <1> is a nonnegative real-valued function defined on [0, (0)

having the following properties:
(1) <1> is increasing and <1>(s) _
as s - +0;
(2) <1>(s) ~ As for large s > 0;
(3) <1>(s + s') ~ B{s + s' + <1>(s) + ~(s')} for s, s' ~ 0.
Here A and B denote positive numbers. Prove that L~ is a linear subspace
ofLl, that L~ => Lee, and that the approximation (*) above is always possible.
6.18. Let (YN);=1 be any sequence of positive numbers such that "YN
= o(N) as N _00. Use the uniform boundedness principle (Appendix B.2.1)
in order to prove the existence of a nonnegative function 1 E Ll such that
lim sup ClNf(O) = 00. (1)

N-ee YN
Is there any sense in which this result is the best possible?

Can you construct explicitly nonnegative functions 1 E L1 that satisfy (1),
for certain specific choices of (YN) ... for example, when YN = N/{log (N + 2W
and e > 01
CHAPTER 7
Some Special Series

and Their Applications
In this chapter we assemble a few results about two special types of series,
namely,
Yz ao + 2:"" an cos nx = 2:
n=1 neZ
cne'n:r , (0)
2:"" an sin nx = 2:
n= 1 neZ
cne'n:r, (8)
where Cn = (1/2i) sgn n alnl' We shall assume throughout that the an are
real-valued, and write SN and aN for the Nth partial sum and the Nth Cesaro
mean, respectively, of whichever series happens to be under discussion.
The series (C) and (8) are examples of so called conjugate series, a topic to
which we return in Section 12.8.
A primary concern will be the determination of conditions under which
these series are Fourier-Lebesgue series. While the results are rather special,
inasmuch as they assume heavy restrictions on the sequence (an), they
frequently play an important role in handling questions of general significance
(as, for example, in Section 7.5).
For an extended study of more special series, see [Zd, Chapter V.
The results we shall obtain can be easily recast into statements about the
definition and nature of ~ (see Section 2.5) for rather special functions cp on Z.
7.1 Some Preliminaries

7.1.1. Modified Kernels. It will be helpful to list a few formulae, some
of which have been encountered in Section 5.1.
N
Yz + 2: cos nx = Yz 2:
na 1 Inl" N
e'n % = Yz DN(x)
sin (N + Yz)x
(7.1.1)
= 2sinYzx '
109
110 SOME SPECIAL SERIES AND THEIR APPLICATIONS
F ( ) = Do(x) + ... + DN(x)

N x N +1
= (N + l)_l[Sin ~(N + l)x]2. (7.1.2)
sm Y2x
Running parallel to (7.1.1) is a formula serving to introduce the so-called
Dirichlet conjugate kernel:
11 D ( ) _ ~. _ cos y:! x - cos (N + y:!)x.

(7.1.3)
72 N x - L. sm nx - 2 . 11 '
n= 1 sIn 72 x
the symbol - marks the passage to the so-called conjugate series, a topic
dealt with in Section 12.8 in some generality.
The modified kernels defined by
(7.1.4)
and
. 1 - cos Nx
Y:! D N#(X) = Y:! DN(X) - Y:! sm N x = 2t Yl
an 2X
(7.1.5)
play useful, if transient, roles. Notice that DN# is even, and that DN# is odd.
Moreover,
(0 < x < 1T), (7.1.6)
and
(0 < x < 1T). (7.1.7)
From 5.1.1 and the defining formulae above we infer that
(7.1.8)
(7.1.9)
as N -HO. It may be shown similarly that
(7.1.10)
(7.1.11)
as N ---+ 00; see Exercise 12.20 and [Zd, pp. 49, 67.
7.1.2. Convex Sequences. In the remainder of this section we deal with

some matters concerning sequences (an) to play the role of coefficients in the
series (0) and (8). We shall assume that the sequence is indexed from n = 0
[7.1] SmIE PRELIMINARIES III
[as is the case with (e)]; only minor changes are needed to take care of
sequences indexed from n = 1 [as is the case with (8)].
For any real-valued sequence (an):'=o we define the sequence of differences
~an = an - a n+ 1 In particular, then, (an) is decreasing (in the wide sense),
if and only if ~a" ~ O. The sequence of second differences is defined by
~ 2 a" = ~a" - ~an + 1 .

The sequence (an) is said to be convex, if and only if ~2an ~ 0 ~ 0; quasi-
convex, if and only if
2: (n + 1)1 ~2a,,1
<Xl
< 00;
,,=0
of bounded variation (BV) if and only if
2: l~a,,1
<Xl
< 00.
,,=0
It is worth explaining at once that the intervention of the difference
sequences is explained by frequent use of the technique of partial summation,
which has already been used without special comment in earlier chapters;
see also [H], pp. 97 ff. Given two sequences (a,,) and (b,,), the formula for
partial summation reads
2:
p~,,~q
a"b" = aqBq - apBp_l + L
p~,,~q-l
~a,,' B,,; (7.1.12)
in this formula it is understood that p ~ q, and that
where r is any fixed integer satisfying r ~ p and such that a" and btl are
defined for n ~ r; an empty sum (that is, a sum extending over a range that
is empty) is always understood to have the value zero. Repetition of the
technique introduces the second differences ~2an'
The following simple result about convex sequences helps to illuminate
7.3.1 to follow.
7.1.3. (1) If (a,,):'=o is convex and bounded, then it is decreasing,

lim n' ~a" = 0, (7.1.13)
and
L'" (n + 1) .~2a" = ao -
,,=0
lim a".
n-co
(7.1.14)
(2) If (a,,) is quasiconvex and bounded, then (a,,) is BV and (n.~a,,) is

bounded.
(3) If (an) is quasiconvex and convergent (to a finite limit), then (an) is
BVand (7.1.13) is true.
Proof. (1) Convexity signifies that (&I n ) is decreasing. If &J", = C < 0,
then ~an ~ c for n ~ m, in which case an __ 00 as n -- 00, contrary to the
assumed boundedness of (an). So ~an ~ 0, (an) is decreasing, and therefore
boundedness implies that a = lim an exists finitely. Now
(7.1.15)
which we now know to be a series of nonnegative terms converging to
ao - a. Since also ~(~an) ~ 0, it follows easily that n.~an __ 0, which is
(7.1.13). Finally, (7.1.14) follows on applying partial summation to (7.1.15)
and using (7.1.13).
(2) We have
(7.1.16)
[This may be proved by induction on n, thus: It is trivially true for n = o.
If it is true for n, then
= (am - a",+n+l) + ~a"'+n+1

= ~2am + 2~2a"'+1 + ... + n~2a"'+n_l
+ (n + l)~a"'+n + ~a"'+n+l
= ~2am + 2~2am+l + ... + n~2am+n_l
+ (n + l){~am+n+1 + ~2am+n} + ~am+n+1

= (~2am + 2~2am+l + ... + (n + 1)~2am+n)
+ (n + 2)~am+n+l;
that is, the equality holds with n replaced throughout by n + 1. Induction

does the rest.]
Taking m = 0 in (7.1.16), we infer that
I(n + l)~anl ~ lao - an+11 + 1~2aol + 21~2all + ... + nl6. 2an_ll,
(7.1.17)
which is bounded whenever (an) is bounded and quasiconvex.
Next, interchanging (an) and (b n) in (7.1.12) and then taking an = sgn bn,
it-appears that
q q
2:
n=O
Ibnl ~ (q + l)/bq l + 2: nl6.b l
n~O
n
[7.1] SOME PRELIMINARIES 113
In this, take b" = aa", to conclude that
2: laa,,1 ~ (q + 1)laa l + 2: nla a"l,

q q
q 2
,,=0 n=O
which is bounded with respect to q [by (7.1.17)] whenever (a,,) is quasiconvex

and bounded.
This completes the proof of (2).
(3) In view of (2), it remains only to show that (7.1.13) is true. But this
follows in taking in turn m = nand m = n + 1 in (7.1.16), leading to
I(n + l)aa2 ,,1 ~ la" - a2"+ll + Ia 2 a" I + 2\a2a1l+ll + ... + nla2a2"_ll

2,,-1
~ la" - a 2 "+ll + 2:
/cal
I(k + l)a2a/c1
and similarly
2: (k + 1)la a/cl
2"
I(n + 2)aa2"u1 ~ la"+l - a211+21 + 2
/czn+l
In Section 7.4 we shall need to know that there exist sequences (an) which
are positive and convex, and which tend to zero arbitrarily slowly. How
such sequences may be constructed will appear from the next two results.
7.1.4. Let a > 1 and suppose that (Nk)r=l is a strictly increasing sequence
of positive integers such that
N2 ~ [1 + Yz(a - 1)-l]Nl'
(7.1.18)
(k + I)Nl<+l ~ 2kN/c - (k - I)N/C-l (k = 2,3, ... ).
If (an) is the sequence defined so that ao == a, an = 11k for n = N /c (k
= 1,2, .. ), and so as to be linear for values of n satisfying 0 < n < Nl or
Nk < n < N/c+l(k = 1,2, ... ),then(an)ispositive,decreasestozero,andis
convex.
Proof. It is evident that (an) is positive and decreases to zero. Conditions
(7.1.18) express the convexity of (an), :which amounts to saying that the
negatives of the slopes of the line segments, obtained by joining neighbors
in the sequence of points of the plane having coordinates
(0, a), (Nl' I},., (N/c, D, ",

form a decreasing sequence.
7.1.5. Let (cn):'=o be any complex sequence such that
lim c"
n-oo
= o.
There exists a sequence (a,,):'.o which is positive, convex, decreasing to zero,

and such that
Ic,,1 ~ a" (n = 0, 1, 2, ... ). (7.1.19)
Proof. First choose NI > 0 so that Ic,,1 ~ Y2 for n ~ N 1 , and then
a > 1 so la.rge that the line segment in the pla.ne joining the points (0, a) and
(NI' 1) lies above all the points
(0, ICol),, (NI - 1, IcN1 - 1 1).
Then, as is easily seen, the integers N 2, N 3, . may be chosen inductively so
as to increase strictly, to sa.tisfy (7.1.18), and to be such that
1
Ic,,1 ~ k +1 for n ~ N".
The sequence (a,,) constructed in 7.1.4 will then sa.tisfy all requirements.
7.2 Pointwise Convergence of the Series (0) and (S)

7.2.1. Suppose that a" = 0(1) and ~:'-o l~a,,1 < co. Then
(1) (8) is convergent every where;
(2) (0) and (8) are uniformly convergent for 8 ~ Ixl ~ 'tT for any 8 > O.
Proof. Consider (8). By (7.1.12) and (7.1.3),
L:
P~"~Q
a"sin na: = aQ Y2D Q(x) - ap ' Y2Dp _ I (x) + L:
P~"~Q
~a". Y2D,,(x)
and so, if 0 < Ixl ~ 'tT,

I L:
P~"~Q
a" sin na:1 ~ 1a QIllsin Y2 xl + lapl/lsin Y~I + L:
P~"~Q
l~a"l/lsin Y2 xl
The statements concerning (S) follow from this and the hypotheses. Those
about (0) are proved simila.rly, using (7.1.1) in pla.ce of (7.1.3).
Conditions for the uniform convergence of (8) are not so superficial.
7.2.~. Suppose that a" t o. Then
(1) the series (S) is uniformly convergent, if and only if M" = 0(1);
(2) the series (8) is bounded1y convergent if and only if nan = 0(1);
(3) the series (8) is the Fourier series' of a continuous function if and only
if na ll _ 0;
(4) the series (8) is the Fourier series of a function in La> if and only if
M" = 0(1).
Proof. (1) Suppose first that (8) is uniformly convergent. Putting
x = 'tT/2N we have
i
[YsN)+ I
all sin nx ~ sin (i) .aN' iI,
(YoN] + I
whence uniform convergence is seen to imply that nan _ O.

[7.2] POINTWISE CONVERGENCE OF THE SERIES (C) AND (S) ll5
Next suppose that, conversely, naIl _ O. Write b" = sup">"na,,, so that

b,,_O aslc_oo. Let 0 < x ~ 'IT and define N = N% to be that integer for
which
_'IT_ < X ~~.
N+ 1 N
Decompose the remainder:
L a" sin nx = L
,,>m m.. ,,<m+N
+ L
">,,,+N
= R + R'.
Then
IRI = I L
m",,<m+N
a"sinnxl ~ X'
"'''"<,,,+N
L naIl
~ xN bIll ~ 'lTb",. (7.2.1)

On the other hand, using partial summation in conjunction with (7.1.7), we
find that
IR'I = I L
n>",+N
da,,' Y2 1),,(x) - a m +NY21)",+N-l(x)1
~ 2a",+N'IT
x
~ 2(N + l)a"'+N ~ 2b",. (7.2.2)
The combination of (7.2.1) and (7.2.2) yields

IL an sin nxl ~ ('IT + 2}b""
n>'"
and uniform convergence is seen to obtain.
(2) The proof is an obvious adaptation of that of (1), the sole difference
being that now b" = O(l).
(3) If nan _ 0, (1) asserts that the series (8) is uniformly convergent and
is, therefore, the Fourier series of its continuous sum function.
Conversely suppose that (8) is the Fourier series of a continuous functionf.
Then, by 6.1.1, aN - f uniformly. This shows that f(O) = O. So, by uniform
convergence,
(7.2.3)
Since sin t ~ 2t/'IT for 0 < t < Y2 'IT, we have
aN ( ; ) ~ L (I __
n .. YoN"
a
N + 1
n ). ~ 7Tn.
'IT N
So (7.2.3) entails
N-l 2: na,,_O,
" .. YoN
hence
N-1al'/.NJ L n_ 0,
" .. YoN
and so finally NaN _ O.
(4) The proof is entirely similar to that of (3), using (2) and the fact that
IluNII",
= 0(1), if 2::'.. 1 an sin nx is the Fourier series of the bounded
(measurable) function/; see (6.4.9).
Remark. In cases (3) and (4) the proofs show that any function, of
which (8) is the Fourier series, is equal almost everywhere to the sum
function of the series (8). This also follows from 6.4.5.
7.2.3. Define
a,,_ = sup {k- L1

n~m<"+k
miami: k {I, 2, .. }}.
E
~n = n L
m~"
I~aml
Then
(1) if an =0(1) and ~" = 0(1), (8) is boundedly convergent;
(2) if an =0(1) and ~n = 0(1), (8) is uniformly convergent.
Proof. We may and will suppose that 0 < x ~ 'TT. Denote by N = N:z;
the positive integer such that
'TT 'TT
-- <x~-'
N+ 1 N
Then, if mE {I, 2, ... },
L an sin nx =
n~m
L
m~n<m+N
+ L
m~m+N
= R+ R'.
Here,
IRI ~ L
m~"<m+"
la"lonx
~ x L
m~n<m+N
nlanl = NxN-l L
m~"<m+N
nlanl
~ Nxoa m ~ 'TTa m (7.2.4)
and [using partial summation and noting that necessarily an = 0(1))
IR'I = IL"~m+"
~n Y2l>,,(x) - am+NoY2l>m+N-1(X) I
~ L
n~m+N
l~anl'TTX-1 + lam+Nl o'TTX- 1
~ (lam+NI + L
n~m+N
l~a"I)(N + 1).
Since
a" ~ nlanl for all n E {I, 2, .},
[7.3] THE SERIES (C) AND (8) AS FOURIER SERIES 117
therefore
Hence
IR'I ~ IXm+N + (m + N) 2:
n;om+N
I~anl
~ IXm+N + {JUN. (7.2.5)

Both statements (1) and (2) follow on combining (7.2.4) and (7.2.5).
7.3 The Series (0) and (S) as Fourier Series

The aim is to establish analogues of cases (3) and (4) of 7.2.2 under weaker
hypotheses on the coefficients an. We begin with the series (0).
7.3.1. Assume that
(7.3.1)
2: (n + 1)1~2anl <
ao
00. (7.3.2)
n=O
Then, for the series (0),
2: (n + 1)~2an' Y2 F n(x)
co
8N(X) - !(x) = (7.3.3)
n=O
pointwise for x t=. 0 (mod 27T),! E L1 and {OJ is the FS ofJ. The hypotheses are
satisfied if an t 0 and (an) is convex (see 7.1.3) in which case I ~ o.
Proof. Two applications of partial summation yield, via the formulas in
7.1.1, the equality
N-2
8N(X) = 2:
n=O
(n + 1)~2an'Y2Fn(x) + Y2N~aN_lFN_l(X) + Y~NDN(X).
(7.3.4)
By 7.1.3(3), n~an - 0, and so pointwise convergence for x t=. 0 (mod 27T) is
clear from (7.1.1), (7.1.2) and (7.3.4). Since also IIFnill = 1, the series
2: (n + 1)~2an' Y2 F n
co
(7.3.5)
n=O
converges in L1 to J. [We are not here asserting that 8NI in L1; see 7.3.2(1).]
It remains to show that (0) is the FS off. We may and will assume without
loss of generality that a o == o. Consider
2: n 2: a~ sin na.
ao co
g{x) = -1 an sin nx = (7.3.6)
n=1 n-1
By 7.2.3, this series is uniformly convergent, since
a: = sup L1 ~
mla:1 = sup L1 ~
laml
k ~"<m<"+k k ~"<m<"+k
~ sup
m;lo"
laml-O,
aa*" = aa"
-n + n.
-..+
(I
1 -
n - --
n
I)
+1 ,
and so
recall that maam_ 0 and am - O.

Now, if
we have DUN = 8N and so, by (7.3.4),
= f.
o
Z (n + l)a~a"o%F,,(y)ay + %NaaN_l of.z F
ZN-~
,,-0 0
N _ 1 (y)ay
+ Y~N LZ
DN(y) ay.
By convergence in Ll of (7.3.5), the first term converges to f(y) ay; since J:

naa" _ 0, the second term tends to 0; since a" _ 0, the third term tends to O.
f:
So
g(z) = f(y) ay.
Hence, by 2.3.4,/(n) = ing(n) for all n E Z. Also, by uniform convergence of

the series in (7.3.6), g(n) = (2in)-la,,,, for all nonzero n E Z. So /(n) = Y~,,,,
for nonzero n E Z and/(O) = 0, showing that the Fourier series off is indeed
the series (0).
(7.3] THE SERIES (0) AND (S) AS FOURIER SERIES 119
7.3.2. Supplements to 7.3.1.

(1) If (all) is quasiconvex, then SN - J in L1 if and only if
alllogn = 0(1). (7.3.7)
Proof. Since (all) is quasiconvex, the series (7.3.5) is convergent in V.
Hence, by (7.3.4), SN _ J in L1 if and only if
(7.3.7')
in Ll.
If (7.3.7) is true, then all = 0(1). Then 7.1.3(3) shows that n.dall = 0(1).
Hence (7.3.7') follows from (7.1.8) and the equation IIFN~l = 1, valid for
Ne{O,I,2, ... }.
Conversely, if (7.3.7') is true, then
YzN.AaN_l1'N(n) + Y~NDN(n)-O asN_oo,

uniformly for n e Z. Taking first n = 0 and then n = YzN or Yz(N - 1)
according as N is even or odd, it is easy to conclude that
N.AaN_1-O as N -00.
Then, since IIFN- 1~l = 1 for all N e {I, 2, .. }, (7.3.7') entails aNDN - 0 in
L' and so, by (7.1.8), that (7.3.7) is true.
(2) If (all) is quasiconvex and
alllog n = 0(1) (7.3.8)
then (7.3.4), 7.1.3(3) and (7.1.8) combine to show that
IlsNlb = 0(1).
From this one may conclude (see 12.5.2 and 12.7.5) that (0) is the Fourier
(-Stieltjes) series of a measure.
See also Teljakowski [2]; MR 48 # 794; 52 # 1480Sa,b; 54 # 13436.
(3) It is easily shown (see Exercise 7.5) that, if p > 1,
(7.3.9)
where (lIp) + (lIp') = 1. From this and (7.3.4) it appears that the con-
ditions
n1-<l/p)a ll = 0(1), (7.3.10)
n 2 -(1/p) Aa ll = 0(1) (7.3.11)
.
Ln2-(1/p)IA2alll < 00 (7.3.12)
11'"0
together suffice to ensure that IlsNllp is bounded with respect to N, in which

case it may be shown (see 12.7.6) that (O) is the Fourier series of a function
inll'.
For the case in which an t 0, a sharper result appears in 7.3.5(2}.
(4) If (an) is quasiconvex and an = 0(1}, (O) is the Fourier series of the
function I ELl; it converges in Ll to I, if and only if (7.3;7) holds; and
if and only if (7.3.8) holds.

The first assertion here is included in 7.3.1; the second is covered by (I)
above; and the third is proved almost exactly as was (I) above.
These results are due essentially to W. H. Young and A. N. Kolmogorov;
see [Ball, Chapter I, 30; [Ba2]' Chapter X, 2; [Z], Chapter V, Theorem (1.12).
(5) If an t 0, 7.2.1 shows that each of (O) and (8) is convergent for x 0
(mod 217-). In either case, the series is a Fo~rier series, if and only if its sum
function belongs to Ll. (This requires proof; see, for example, [Ba 2 ], p. 199.)
(6) If an t 0, the proof of7.3:1 can be modified in the following way (kindly
suggested to me by Professor G. Goes). By 7.2.1(2}, the series (O) converges
. uniformly for 8 ~ Ixl ~ 'IT, for any 8 > O. The sum function f is thus con-
tinuous except perhaps at points x == 0 (mod 2'IT). The equation (7.3.4) and
the estimates [see (7.1.1) and (7.1.2)]
I%N ~aN_lFN-l(x)1 ~ %1 ~aN_11(sin %x} -2,

I%aNDN(x) I ~ %aNlsin %xl- l
combine with the assumption an tOto show that (7.3.13) holds for 0 < Ixl ~
'IT. Defining
L: n -la
N
gN(X) = n sin nx,
n=l
one has for 0 < a~ x ~ 'IT
and, by uniformity of the convergence of 8 N to Ion [8, 'IT], it follows that
g(x) - g( 8) = i% I(y) dy
for 0 < 8 ~ x ~ 'IT. By (7.3.3) and (7.3.2),feLl. Since also g is continuous
and g(O) = 0, it follows that
g(x) = 1% I(y) dy
[7.3] THE SERIES (C) AND (S) AS FOURIER SERIES 121
for O:s;; x :s;; 17. A similar argument yields the same equation for
-17 :s;; X :s;; O. From this point on, the argument proceeds as before.
Moreover, if an t 0, one can even dispense to some extent with (7.3.2),
though one cannot now conclude that f ELI. However, it will still be true
that
g(x) - g(xo) = I'" f(y) dy

"'0
for -17 :s;; Xo :s;; x < 0 and for 0 < Xo :s;; x :s;; 17. Because of this one may
integrate by parts in the formula
n -Ian = 2/17 fa" g(x) sin nx dx

lim 2/17i" g(x) sin nx dx
6-0 6
to conclude that
an = lim2/TTi" f(x) cos nxdx;

6-0 6
recall that g is continuous and g(O) = 0, and that we are assuming (as before)
that ao = o. In other words, (0) is the series
L /(n)e
neZ
ln ""
where now
/(n) = lim 1/217

6-0
r
J6 ",,,,,,,,,
I(x)e -In'" dx
is a Cauchy principal value. Thus (0) is, in this wider sense, still the Fourier
series of f.
We next turn to a few analogous results for the sine series (8).
7.3.3. Suppose that an t 0 a.nd write
L a"sinnx,
co
f(x) =
n=1
the series being everywhere convergent by 7.2.1. Then IE Ll if and only if
L 2an
co
n=1
<00, (7.3.13)
in which case (8) is the Fourier series of I and

lim
N-oo
Ils.v -/111 = O. (7.3.14)
Proof. The reader will verify easily that, under the stated conditions on
(a,,), (7.3.13) is equivalent to
L:'" Aa" log n <

,,=1
00. (7.3.13')
By partial summation,
L: Aa" % D,,(x) + %aN DN(x)

N-1
8.... (X) =
,,=1
~ L: Aa % D,,(x) = f(x}.
co
n ' (7.3.15)
,,=1
If we replace herein D" by D,.# we obtain a function
L: Aa" %D,,#(x}
co
f#(x} = (7.3.16)
n=l
which differs from f by
L:
'"
,,=1
Aa,,' %sin nx,
which is continuous since, by virtue of the relation a" t 0, one has
L: IAa,,1 <
CD
00.
,,=1
Thus fELl if and only if f# ELl.
On the other hand, since D,,# is odd and is nonnegative on (0,11) [see
(7.1.5)], f# E Ll if and only if
L Aa,,' IID"#IIl <

ao
00.
,,=1
This requirement is equivalent to (7.3.13'), in view of (7.1.11).
Finally, assuming (7.3.13') to hold, we have
a" log n= log n' L: Aa k ~ L: Aa k log k,

rc .. " rc .. "
which tends to zero as n~oo. Hence (7.1.10) and (7.3.15) show that 8N-f
in mean in Ll as N ~oo.
7.3.4. Suppose that a" +0 and that

(7.3.17)
Then 2:'=1 a" sin nx, although everywhere convergent, is not a Fourier-
Lebesgue series.
Proof. This follows immediately on combining 5.3.1,6.4.5, and 7.3.3.
[7.3] THE SERIES (C) AND (8) AS FOURIER SERIES 123
7.3.5 Supplements.
(i) 7.3.3 can be generalized ..For example, Teljakowskii [1] has proved that,
if (an) is quasiconvex and an ~ 0, then (8) is a Fourier series, if and only if
Also, Kano and Uchiyama [1] have proved that, if (an) is quasiconvex and
bounded, then (8) converges in Ll, if and only if
L: n-Ilanl <
II)
00 and anlogn~O,
n=1
while
if and only if
L: n-Ilanl <
II)
00 and anlog n = 0(1).

n=1
These writers show also that there exists a nonnegative quasiconvex sequence
(an) such that an ~ 0, both (0) and (8) are Fourier series, and yet both (0)
and (8) diverge in LI.
See also TeljakowskIi [2] MR 52 ## 14805, 14819; MR 55 #8673.
(ii) In 7.3.2(3) we have indicated conditions sufficient to ensure that the
sum function! ofthe cosine series (0) shall belong to IJ'. For the case in which
+
an 0, a sharper necessary and sufficient condition is known. Denoting by g
the sum function of the sine series (8), and assuming that 1 < P < 00, it is
known that the following conditions are equivalent:
(1) f E IJ';
(2) g E IJ';
(3) the sequence (an):=I' assumed to decrease monotonelyto zero, satisfies
the condition
Despite appearances, this result is considerably deeper than that men-

tioned in 7.3.2(3). Its proof depends on Theorem 12.9.1 and an inequality of
Hardy; see [Z2], p. 129 or [Ba2]' p. 207.
For further results, see Aljancic [1] and the references cited there.
Hardy [1] proved that, if (8) is the Fourier series of some f E IJ' (p ~ 1),
then so too is the series
L: (Ta)n sin nx,

II)
(7.3.18)
n=1
124 SOME SPECIAL SERIES A..~D THEIR APPLICATIONS
where
(Ta)" = n-1(a1 + ... + an) for n E {I, 2, ... }.
On the other hand, G. and S. Goes [1] proved that if (an) is BV and an ~ 0,
then (7.3.18) is a Fourier series, if and only if
L: n-1lanl
co
< 00;
,,=1
See also MR 55 # 8673.

(iii) See also [Boa] and Boas [2].
7.4 Application to A(Z)

We can now substantiate a statement made in 2.3.9 concerning the
existence of integrable functions whose Fourier coefficients tend to zero
arbitrarily slowly.
Let c/> E co(Z) be given and define
A" = max [I c/>(n) I , Ic/>( -n)1l (n = 0, 1,2, ... ).
Then A,,~O as n~co and we may, as in 7.1.5, construct a sequence (a,,)
which is convex and satisfies an t 0 and
(n = 0, 1,2, ... ).
Consider the functionf figuring in the proof of 7.3.1. Since (a,,) is convex, f is
nonnegative and integrable. The Fourier series off is Y2 a o + 'L.;'= 1 an COS nx,
SO that j(n) = Y2alnl for all n E Z. Hence
/(n) ~ Ic/>(n) I for all n E Z.
See also [HR], (32.47).
7.5 Application to Factorization Problems

In this section it will be shown how 7.3.1 aids in the solution of a number
of problems concerning the possibility of factorizations f = g * h with f, g,
and h in specified function spaces. Included in the discussion will be the
Salem-Zygmund-Rudin-Cohen result mentioned in 3.1.1(c). The approach
will be "classical" (as opposed to the Banach algebra-based technique
introduced by Cohen [4] for certain such problems; see 11.4.18(6)), being a
modified form of the arguments published by Rudin [1].
7.5.1. Let E denote anyone of LP (1 ~ P < co) or Ck (0 ~ k < co). Then

E = Ll * E.
[7.5] APPLICATION TO FACTORIZATION PROBLEMS 125
Proof. Since, by 3.1.5 and 3.1.6, Ll * E c E, it suffices to show that

E c Ll * E. In doing this we shall use the simplest results about Cesaro
summability; numerous other summability methods would serve equally
well, however.
Suppose that / E E. The aim is to prove the existence of gEE a.nd hELl
such that
(7.5.1)
The starting point in the construction of g and h is the remark that, by

6.1.1,
(7.5.2)
where 11'11 denotes the appropriate norm (II-I\p or 11'II(k in E. As a consequence

of (7.5.2) one may, as it is easy to verify, choose nonnegative numbers aN
(N = 1, 2, ... ) so that
(7.5.3)
L'" aNII! -
Nal
FN *!II < 00, (7.5.4)
and
(7.5.5)
Now (7.5.4) and the completeness of E together ensure that
'"
g =/+ L
N=l
aN(j- FN*f) (7.5.6)
belongs to E, the series being convergent in E. This in tum entails that
g = J. [1 + L
'"
aN(1 - PN)); (7.5.7)
N=l
the reader will observe that, for a fixed nEZ, 1 - PN(n) = O(ljN) for
large N, so that (7.5.5) ensures the pointwise convergence on Z of the series
appearing in (7.5.7).
In view of 2.4.1, (3.1.5), and (7.5.7), to establish (7.5.1) for some hE Lt, it
suffices to show that the sequence (a")"EZ defined by
(7.5.8)
is a Fourier-Lebesgue sequence. It is at this point that 7.3.1 comes to our aid.

Noting that a" = a\II\' 7.3.1 shows that we have only to verify the following
two points:
(1) all t 0 as n too;
(2) (a"),, .. o is convex.
As to (1), we have
1 __ n
ifO~n..;;N,
PN(n) = {0 N +1
ifn > N,
so that 0 ~ PN(n) ~ 1 and PN(n) = 0 for n > N. Therefore, by (7.5.3),
asn~oo.
Moreover b"+1 ~ b", since aN ~ 0, and PN(n + 1) ~ PN(n). Thus (1) is

satisfied.
Turning to (2), the convexity of (a"),, .. 0 is equivalent to the condition
(7.5.9)
Now it turns out that
a.nd so, since aN ~ 0 and bll ..;; bn+ 1 ~ bn+ 2' it is clear that (7.5.9) holds.
Thus (2) is satisfied a.nd the proof is complete.
7.5.2. Remarks. (1) The excluded case k = 00 of 7.5.1 is easily dealt

with separately: indeed, on using 12.1.1 it is easily seen that Coo = Coo * Coo.
On the other hand, the excluded case p = 00 is 3 false assertion (as follows
from 3.1.4).
As with 6.1.1, the method of proof may be adapted to apply to homo
geneous Banach spaces over T; see [KzJ, p. 61, Exercise 1.
(2) Extensions of parts of 7.5.1 (other than the case E = Ll) to more
general groups have been examined by Hewitt [IJ and by Curtis and Figs.-
Talamanca [1] jointly. (Rudin's original argument applies to locally
Euclidean groups.) Hewitt exploits Cohen's method, working within a
general Banach algebra setting; see 11.4.18(6).
7.5.3. Having now shown that V = Ll * Ll, it is natural to ask whether

LP = LP * LP for p > 1. That the answer is negative will be established for
p = 2 in Section 8.4, and for general p > 1 in Exercise 13.20 and again in
15.3.4.
[7.5] APPLICATION TO FACTORIZATION PROBLEMS 127
7.5.4. Supplements to 7.5.1. (1) Concerning the functions g and h appearing

in (7.5.1) and constructed in the proof of 7.5.1, two comments deserve to be
recorded.
In the first place, the sum appearing in (7.5.4) may plainly be made inferior
to any preassigned I) > 0, in which case (7.5.6) shows that
Ilg -111 < I). (7.5.10)
In the second place, having arranged that the sequence (a,,),,>o decreases to
zero and is convex, equation (7.3.4) shows that the function h is nonnegative,
so that Ilhlll = ao and therefore
(7.5.11)
(2) With but little further effort 7.5.1 may be sharpened to the following
extent.
Suppose that El is a acompact subset of E (by which it is meant that El is
contained in the union of a countable sequence AI (i = 1, 2, ... ) of compact
subsets of E). Then there exists an hE V, which is nonnegative and satisfies
(7.5.11), with the property that to each. 1 EEl CJorresponds agE E such that
(7.5.1) holds. (The choice of h may be made the same for all 1 E Ed
The basis of the proof of this extension is the remark that, since AI is a
compact subset of E, the numbers
tend to zero as N ---+ 00 for each fixed i. (This in turn depends on the fact that.
in any metric space, any compact set can be covered by a finite number of
balls of arbitrarily small radius; this remark is used in conjunction with the
observation that 111 - F N .111 ~ 2111 I for all N and all 1 E E.) This being so,
the numbers aN ;;. 0 are chosen to satisfy (7.5.3) and (7.5.5), while in place of
(7.5.4) we impose (as we may) the demand that
( 7.5.4')
for each i = 1,2,. The proof then proceeds exactly as before.

(3) Other factorization theorems may be derived from 7.5.1. For example,
starting from the relation C = C Ll, one may infer that
Cl = C. AC, (7.5.12)
where AC denotes the space of (periodic) absolutely continuous functions; see

Exercise 7.10.
7.5.5 For further reading, see 11.4.18(6) below; [HR], (32.14) ff.; [DW];
MR 34 ## 4817, 4818; MR 46 # 2355; MR 52 # 6327; MR 53 ## 8782,8789;
MR 54 ## 843, 8151.
EXERCISES
7.1. Show that if '2:'=0 e" is convergent, and if '2:'=0 IDaA,,1 < 00, then
'2:'=0 A"e" is convergent.
7.2. Prove the converse of the result in Exercise 7.1, that is, that if
'2:'=0 A"e" converges whenever '2:'=0 e" converges, then '2:'=0 IDaA,,1 < 00.
Hint: Show that the hypothesis means that '2:'=0 DaA" 8" is convergent
for any sequence 8" _ O.
7,3. Prove that, if '2:'= 0 e,,/log (2 + n) is convergent, then '2:'=0 e,,/(l + n)'"
is convergent for any a > O.
7.4. Write 8" = '2~=0 ele and ale = (n + 1)-1(80 + 81 + ... + 8,,). If we
are given that (a,,) is a bounded sequence and that 8" = o(log n) as n-oo,
prove that '2:'=0 e,,/log (2 + n) and '2:'=0 e,,/(l + n)'" are convergent for
a > O.
Hint: Write A" = 1/log (2 + n) and apply partial summation twice in
succession to the series '2:'= 0 A"e"
7.5. Verify that, if p > 1,
as N _00, where l/p + l/p' = 1 and where A,. and B,. are positive and
independent of N. [Compare (7.3.9).]
7.6. Suppose that a(x) is defined and continuous for x ~ 0, and that the
second derivative a"(x) exists and is nonnegative for x > O. Show that the
sequence (a(n:'=o is convex.
7.7. Show that '2:'=2 cos nx/log n is the Fourier series of an integrable
function! ~ 0, but that '2:'=2 sin nx/log n is not a Fourier-Lebesgue series at
all. (This example will gain in significance in 12.8.3.)
7.S. Using 7.3.2(3), show that if 0 < IX ~ 1, then '2:'=1 cos nx/n'" is the
Fourier series of a nonnegative function! E V' provided 1 ~ P < (1 - a)-l,
and that then
lim
N..... II! - i:
"=1
cos ,"nxll
n ,.
= O.
What can be said if a > 11

7.9. Using Exercise 7.8, show that, if 0 < a ~ 1, then '2"ez,,, .. oU(n)/lnl'"
is convergent for each g E Lq with q > a- 1.
For further results in this direction, see 10.4.3, Exercises 10.4 to 10.6,
and Exercise 13.1.
7.10. Construct a. proof of (7.5.12), and deduce that
Cm = cm -1 * AC = C * AC * ... * AC
~
m (acton
for m = 1,2,.
EXERCISES 129
7.11. Prove that [with the notation used in Remark (4) following 2.3.5 and
7.5.4(3)]
(1) AC = Ll * AC;
(2) AC = L1 * BV.
Note: From (2) we see that Ll * BV is a proper subset ofBV. In proving
that L1 * BV c AC, it is useful to note that V (J) = I Df III for each trigono-
metrio polynomial f. (Actually, V(J) = IIDflll is true for each absolutely
continuous f; see [Na], p. 259.)
CHAPTER 8
Fourier Series in L2
It will be shown in this chapter that the problem of mean convergence of

Fourier series in L2 has a complete and simple solution. The abstract
foundation for this situation lies in the fact that L2 is a Hilbert space with
the inner (or scalar) product
(j, g) = 211T f
jg dx, (S.l)
and that moreover the functions en defined by
(n E Z) (S.2)
form an orthonormal base in V. This last means that the family (en) is
orthonormal, in the sense that
(m,nEZ), (S.3)
and that
a.e. (S.4)
Indeed, (S.3) is simply a restatement of the orthogonality relations, and

the implication (S.4) is a special case of the uniqueness theorem 2.4.1. As
Hilbert space theory shows, these two facts imply that each f E V has a
convergent expansion
f = 2: (j, ell)e,.;
neZ
(S.5)
see, for example, [E], Corollary 1.12.5, or [HS], pp. 245-246, or [AB], pp.
239-240.
Despite this ready-made solution to the problem before us, we shall not
assume a knowledge of Hilbert space and will give all the necessary details
pertaining to the present situation. For general orthogonal expansions, see
[KSt], Kapitel III.
With the exception of Section S.6, what little we have to say about point-
wise convergence is included in Chapter 10.
130
[8.2] MEAN CONVERGENCE OF FOURIER SERIES IN L' 131
8.1 A Minimal Property

We make a start by showing that, for a given function! E L2, the sequence
of partial sums SN! of the Fourier series of I possesses a minimal property
which already serves to distinguish the Fourier series of! among all trigono-
met.ric series (see the discussion in Section 1.2).
Denote by TN the linear space of all trigonometric polynomials of degree
at most N, that is, all linear combinations
(8.Ll)
of the functions en for which Inl ~ N; see Exercise 1.7.
8.1.1. For a given I E L2 one has
for every t E TN different from SN!'

Proof. A perfectly straightforward calculation, based upon the ortho-
gonality relations (8.3), leads to the identity
II! - tl122 = 1111122 + L:

l"I<iN
Ian - i(n)J2 - L:
l"I<iN
li(n)12 (8.1.2)
for an arbitrary t E TN, given by (8.1.1). The right-hand side of (8.1.2)

plainly has a strict minimum which is attained for the choice an = i(n)
(Inl ~ N) and for no other, and this minimum value is III - sN!112',
8.2 Mean Conver~ence of Fourier Series in L2. Parseval's

Formula
Using 8.1.1 and the results in Chapter 6, it is a simple matter to establish
mean convergence of Fourier series in L2.
8.2.1. If! E L2, then
lim II! -
N-QI)
8N ! 112 = lim II! - UN!112
N_co
=0 (8.2.1)
and
(8.2.2)
Furthermore, given any e > 0 there exists a finite subset Fo = Fo(e) of Z

such that for every finite subset F of Z satisfying F :::> F 0 one has
II! - L: i(n)e,,112 ~ e.
neF
(8.2.3)
132 FOURIER SERIES IN L2
Proof. Since aNl E TN' 8.1.1 entails that
111 - 8 .... 1112 ~ 111 - aNl112

On the other hand, the case p = 2 of 6.1.1 shows that
lim
N->
111 - aNl112 = 0,
and (8.2.1) is thereby established. Taking t = 8N l in (8.1.2) yields
111 - 8 Nl1122 = 1111122 - L

Inl"N
jj(n)j2;
if we let N _00, (8.2.2) follows.
Finally, putting SF = LneF/(n)enforanyfinitesetF c Z,andg =1- 8 F,
we have g E L2 and
g(n) = 0 (n E F), g(n) = j(n) (n E Z\F).
An application of (8.2.2) to g yields
111 - sJ112 2 = L
neZIF
Ij(n)i2.
Since L Ij(n)12 is a convergent series of nonnegative terms, (8.2.3) follows

at once.
Remarks. Equation (8.2.2) is one form of the Parsevallormula. Valid
extensions of (8.2.2) will appear in Sections 13.5 and 13.11. If one takes
1, g E L2, replaces 1 by 1 + Ag in (8.2.2), and varies the scalar A, one infers
directly the so-called polarized version of the formula, which reads as follows.
8.2.2. If 1, g E L2, then
(j, g) == 21 jl(x)g(x) dx =
7r
L j(n)g(n) ,
neZ
(8.2.4)
the series being absolutely convergent. In view of 2.3.1, it is equivalent to

assert that
21 jl{x)g(x) dx
7r
= L j(n)g(-n),
neZ
(8.2.5)
the series being absolutely convergent, for 1, g E L2.
8.3 The Riesz-Fischer Theorem

In view of the fact that
L Ij(n)i2 <
neZ
00
for each 1 E L2, it is especially satisfying to be able to state and prove an

unqualified converse.
[8.3] THE RIESZFISCHER THEORE~I 133
8.3.1. (Riesz-Fischer theorem) Let (cn)nez be any sequence of complex

numbers such that
(8.3.1 )
Then there exists an f E L2, uniquely determined almost everywhere. such

that j(n) = Cn (n E Z).
Proof. If we define
the orthogonality relations (8.3) yield for 111 < N the equality
1 8M - 8.'11112 2 = L
M<lnl<;!V
Ic l2
n
which, by (8.3.1), tends to zero as 111, N -+00. By completeness of L2 ([W),

Theorem 4.5a), there exists an essentially unique f E L2 such that
lim
lV-X)
IIf - 8N !12 = O. (8.3.2)
Moreover (8.3.2) entails (Cauchy-Schwarz inequality) that j(n) = lim N _ oo

sN{n) for each n E Z. Since sN(n) is cn or 0 according as Inl ~ N or Inl > N,
it follows that j(n) = Cn for all n E Z.
8.3.2. Remarks. The result 8.3.1 is known to be the best possible in the
following sense: given any sequence (cn)nez for which
it is possible to choose the signs in such a way that the series L cne tnx is
not the Fourier series of- any integrable function (nor even the Fourier-
Stieltjes series of any measure, as defined in 12.5.2). Such questions will
receive more attention in Chapter 14; see especially 14.3.5 and 14.3.6.
It can also be shown that the relation Lnez Ij(n)j2 < 00, valid for fE L2,
cannot be much improved even for continuous fun.ctions f. For example (see
Exercise 8.9), given any positive function w on Z such that lim 1nl _ oo w(n)
= 00, there exist continuous functions f such that
lim sup Iw(n)j(n)1 = 00 (8.3.3)

n-oo
and
neZ
L w(n)jj(nW = 00. (8.3.4)
Moreover, there exist (Carleman) continuous functions f such that
) Ij(n)1 2 -S = 00 for any given e > 0, (8.3.5)

,{ez
and even (Banach) such that
L IJ(n)12-e. = 00
neZ
for suitable sequences en ~ 0 (see Exercise 15.13). A specific example of

(8.3.5) is
~ eicnlolrn
l(x) = L. ein %
n=2 nYz(log n)B '
where,B > 1 and c > 0; see [Zl]' p. 200, Theorem (4.11); this example is quite
different in nature from those indicated in Exercise 15.13. See also [Ba l ],
p.337.
For some valid extensions of 8.3.1, see Sections 13.5 and 13.11. Regarding
extensions of 8.2.1 and 8.3.1 to general groups, see 13.5.2.
8.3.3. Dual Version of 8.3.1. As has been hinted at in Section 2.5, 8.3.1
can be recast into the form of a dual result. It says in effect that, if t/J E 12 ,
then the trigonometric polynomials
converge in mean in L2 to a function 1 = ~ on T such that = t/J. ThisJ

includes a possible interpretation of the inversion formula spoken ofin Section
6.7, and in particular attaches a good meaning to the Fourier transform ~
whenever t/J E t2 (see Section 2.5).
See also Sections 13.5 and 13.11.
8.4 Factorization Problems Again

It has been stated in 3.1.1(c) and proved in 7.5.1 that there are no prime
elements in the *-algebra L1, that is, that every 1 E Ll can be factorized in at
least one way as 1 = 11 * 12 with 11,12 ELl. In 7.5.3 we alleged that the
analogous assertion for V (p > 1) is false; see also Exercise 13.20. We can
now verify this when p = 2.
Indeed, if p = 2 the Parseval formula (8.2.2) and the Riesz-Fischer
theorem 8.3.1 combine to show that any 1 E L2 for which ~nez IJ(n>l = 00
is a prime element of L2 and (see Exercise 8.2) that these are the only primes
in L2.
In view of this it is natural to ask whether it is true that any nonprime
element of V (p > 1) is expressible as a finite convolution product of prime
elements of LP. An affirmative answer for p = 2 is given in Exercise 8.3.
[8.5] MORE ABOUT MEAN MODULI OF CONTINUITY 135
8.5 More about Mean Moduli of Continuity

Extending the notation introduced in 2.3.7, we define the mean modulus
oJ continuity oJ J with exponent p > 0 by
(8.5.1)
This may be regarded as defined for any measurable f, setting wpf(a) = 00

if TaJ - J does not belong to LP. Translation invariance of the integral
shows that
(8.5.2)
For p ~ lone has also by Minkowski's inequality
WpJ(a + b) I
= Tauf - flip ~ IITa+d - Tbfllp + IITbf - flip
= wpf(a) + wl,j(b) , (8.5.3)
again by invariance of the integral. Also, by (2.2.19),
(0 < p < q). (8.5.4)
It has been seen in Exercises 5.1 and 5.2 that restrictions on the rate of
decrease of wd(a) as a -+ 0 bear upon smoothness properties ofJ. It will now
be seen, by using the results of 8.3, that further results of this nature may be
expressed in terms of wd and w2f. Similar and much more elaborate results
will be mentioned at the end of 10.4.6.; see also 10.6.2.
8.5.1. If JELl and
then f E L2 and
(8.5.5)
Proof. By 2.3.7 and (8.5.2) we have
,j(n), ~ Yzwd (,:,)
for nEZ, n oF O. So (8.2.2) gives
which is equivalent to (8.5.5).

8.5.2. Suppose that f E L2 and that a > O. Then
\ 2: Inj(n)12 ~ [w 2f (a)]2
'TT Inl""la a
~ 2:
nez
Inj(n)12. (8.5.6)
Proof. By 2.3.3 and (8.2.2),
[wd(a))2 = 2:le lna - 112 !j(n)i2. (8.5.7)

Furthermore,
le lna - 11 = 21sin 72nal;
and
21sin 72 nal ~ 2 'TT~ 172 nal
while
21sin 72nal ~ 2172 nal = lnal
for all n and all a. Insert these estimates into (8.5.7): (8.5.6) emerges after
division bya 2
8.5.3. Concernin~ Absolute Continuity. On the basis of 8.5.2 we can

establish a number of interesting conditions, each necessary and sufficient in
order that a given L2 function shall be equal almost everywhere to an
absolutely continuous function whose derivative (existing pointwise almost
everywhere) belongs to L2.
8.5.4. Suppose thatf E L2. Then the following four conditions are equivalent:
(1) after correction on a null set, f is absolutely continuous and Df E L2;
(2) Lnez Inj(n)J2 < 00;
(3) lima.:.oa-1(T -af - f) exists in mean in 12;
(4) w2f(a)/a = 0(1) as a-+O.
If anyone of these conditions is fulfilled, the limit mentioned in (3) is DJ.
Proof. That (1) implies (2) follows directly from 2.3.4 and (8.2.2).
Assuming (2), 8.3.1 ensures that there exist g E L2 such that g(n) = inj(n)
for all n E Z. Then 6.2.8 shows that
L j(n)(elnb -
nez
elna ) = [b g(x) dx
Ja
(8.5.8)
for all a and b. Moreover (2) entails that Lnez Ij(n)1 < 00, and 2.4.2 shows
that, after correcting f on a null set,
f(x) = 2: j(n)eln:r:
neZ
for all x. So (8.5.8) and Lebesgue's theorem on the derivation of integrals
([W), Theorem 5.2g) combine to show that f is absolutely continuous and
[8.6] CONCERNING SUBSEQUENCES OF 8NI 137
Df = g almost everywhere. Thus (1) is satisfied and the equivalence of (1)

and (2) is established.
Again assuming (2), and using the above notation, Parseval's formula
(8.2.2) gives
Ila-1(T -af - f) - gl122 = L: la-1(e
neZ
IM - I)J(n) - inJ(n)i2
= L: la-1(e
neZ
'na - 1) - inI211(n)i2. (8.5.9)
Now
lim a-1(e 1na - 1) - in = 0 (n E Z),
a-O
and (see the proof of 8.5.2)

la-1(e 'na - 1)1 ~ const Inl.
These facts, combined with (2) and (8.5.9), show that
Iia -l(T -af - f) - ul122 ~ 0
as a ~ O. Thus (2) implies (3).
Since w2f(a) = liT -af - fl12' it is evident that (3) implies (4).
The first inequality in 8.5.2 shows that (4) implies (2).
We now know that
(1) .;:> (2) => (3) => (4) => (2),
so that the proof is complete.
8.5.5. Remarks. (1) By using the uniform boundedness principles

(Appendix B.2.1 and B.2.2) it could be shown that a fifth equivalent condition
is obtained on apparently weakening (3) to the demand that the said limit
exists weakly in L2, that is, (see Appendix C.l), that
lim fa-1(T -af - f)' h dx

a-O
exists finitely for each hE L2.

(2) Yet another equivalent condition is the existence of U E L2 such that
lim fa-1(T -af - f)' u dx = fuu dx

a-O
for each u E C>; see Exercise 12.24.
8.6 Concerning Subsequences of 8 N f

As will be described in 10.4.5, Carleson announced in 1966 a proof of the
almost everywhere pointwise convergence of 8 N f for each f E L2. Despite this,
we shall here illustrate in detail the use of the Parseval formula in proving a
138 FOURIER SERIES IN LI
much earlier result due to Kolmogorov (1925). This theorem of Kolmogorov,

which asserts the existence of specifiable, rapidly increasing sequences
(N /C);'-l of positive integers having the property that lim/c ... ",8N,.! = f point-
wise almost everywhere for each f E L2, prompted a number of similar investi-
gations which are not superseded by Carleson's theorems; see lO.4.6.
We begin with a definition. A sequence (N/C);'=l of positive integers is
termed a Hadamard 8equence if
q == infN/C+l > l'

Ic NIc '
such sequences have appeared already in Exercises 5.6 and 6.13, and they
will be encountered again in Chapter 15.
8.6.1. (Kolmogorov) Let (N/C);'_l be a Hadamard sequence of positive

integers. If f E L2, we have 8N,.!(X) -+ f(x) for almost all x.
Proof. Recall from 6.4.4 that aNf""f a.lmost everywhere. Observe also
tha.t if (gr):"-l is any sequence of nonnegative integrable functions, and if
i
r=1
fgr < 00,
then ([W), Theorem 4.1e) ~;o..1 gr is integrable, hence finite. valued almost
everywhere, and so gr -+ 0 almost everywhere. In view of these two remarks,
it will be sufficient to show that
...
S = L IlaNJ - BN"JII22 < 00. (8.6.1)
"'-1
Now, by (8.2.2),
~aNJ - BN"fll!!2 = (N", + 1)-2 L n2IJ(n)12,

I"I<ION"
80 that
L IlaN,.! - BN,.!!!2!! ~ L Nji2 L n2Ij(n)12
/C<lOP /C<lOP 1"1 <10 N"
== Sp, (8.6.2)
say. Putting U" =n Il IJ(n)12, we may write
with the understanding that No be read &8 O. On the other hand, since (N/c) is a
Hadamard sequence, it is easily seen that
L N m -2 ~ O'Nj-lI,
m;.j
[8.7] A(Z) ONCE AGAIN 139
where C is a number depending only upon q. Hence
(8.6.3)
Moreover,
So, by (8.6.3) and (8.2.2),
8 p EO C' L li(k)l:! EO C' Ilf1122.

Ilel",N"
From this estimate and the obvious relation

8 = sup8p ,
p
(8.6.1) follows and the proof is complete. More general results of similar nature
are proved in [Z2], Chapter XV; see especially p. 231.
8.7 A(Z) Once Again

In this section we shall use 8.2.1 and 8.3.1 in a manner due to Hirschman
so as to obtain a sufficient condition for a complex. valued function", on
Z to belong to A(Z); see 2.3.10(4). The result actually given by Hirschman
([1], Lemma 3a) appears as 8.7.3 and will be used in Section 16.4.
For leE Z, TIe~ will denote the translated function n_~(n - Ie) on Z
(compare the notation introduced in 2.2.1 for functions defined on groups).
8.7.1. Suppose that", E co(Z) and that
L:
aD
8 == 2-YamIIT(2")~ - ~112 < 00. (8.7.1)

01=0
Then there exists a function 1 E Ll such that j = ~ and
111111 ~ %8. (8.7.2)

Proof. Let m be a nonnegative integer and define k =2
m According to
8.3.1, (8.7.1) shows first that there exists a function fie E L2 such that jle =
Tie" - " and
in V 3, where
BU = L {,,(n - k) - "(n)}e,,.
1"lsN
So (see the proof of Theorem 4.58 of [WJ) one can choose sequences (N,(Ie');"..1
(k = 2,4, ... ) of integers such that (N~Ic+1::'1 is 8 subsequence of (N,(Ie'):"_1
and
pointwise almost everywhere. On taking the "diagonal subsequence" (Nr );o..l

= (Nr(rl):"al' one will have
for almost all x and all k = 1,2,.. Since; is known to belong to co(Z), this
last relation may be written
for almost all x and all k = 2, 4, .... This relation makes it plain that there
exists a measurable periodic function I such that
a.e., (8.7.3)
and 8.2.1 shows that
Let 1m denote the interval defined by the inequalities
so that U:=o 1m = (-.", .,,). For x Elm'
so that (8.7.4) entails that
J,r.. I/(x)12 dx =EO ."IIT/c; - ;112 .

2 (8.7.5)
From the Cauchy-Schwarz inequality and (8.7.5) it may be inferred that
J,r.. I/(x)1 dx =EO .,,2-%mIIT/c; - ;112,
and, summing over m, (8.7.1) yields
1111It =EO %8,

which is (8.7.2). In particular, I ELl as a consequence of (8.7.1).
Finally, for all nEZ,
;(n - k) - ;(n) = 2~ fl/c(X)e- 1nz dx,
which, by (8.7.3), is equal toJ(n - k) - J(n). If we let m_oo and use 2.3.8
and the hypothesis that; e co(Z), it appears that; = J. The proof is complete.
[8.7] A(Z) ONCE AGAIN 141
8.7.2. The statement of Hirschman's lemma is made in terms of the

expression
,-1
V/I(4)) = sup { 2: 14>(nle +l) - 4>(nIeWPI/I,
Ie-O
where f3 is a positive real number and the supremum is taken with respect to
all strictly increasing finite sequences (nle)~ ~ 0 of integers. The supremum
may, of course, be 00.
It is evident that V 1(4)) = Ila4> I 1, where the difference operator a is
defined as in 7.1.2. Also, if 4> is bounded,
for any f3 > 1.

Less evident is Hirschman's estimate which follows.
8.7.3. Suppose that
14>(n) I ~ K(1 + InD-21a (ne Z) (8.7.6)
for some IX > 2, and that V8 (4)) < 00 for some f3 satisfying 1 ~ f3 < 2. Then,
for k = 1, 2, 3, .. "
\I T 1e4> - 4>1\2 ~. A a 8 V 8(4))BCa-2l12Ca-8lKaC2-8)12(a-/l l kCa - 2 )(2Ca-Bl. (8.7.7)
In particular, there exists an 1 e L1 such that I = 4> and

111111 ~ A~.8V8(4))B(a-2l12(a-BlKa(2-B)J2(a-B).
Proof. Holder's inequa.lity for sums gives
II TIe~ - ~1122 :s;; { L: 1(n) - ~(n + k)I,,}<a-2)/(a-B).

"EZ
{ L: 1(n) - (n + k)laf~-8)/<a-B). (8.7.8)
"EZ
Beside this,
Ie-I
L:
"eZ
1(n) - 4>(n + kW = 2 L:
1-0 mEZ
I~(mk + j) -4> (mk + j - k)I'
(8.7.9)
and, by (8.7.6),
L: 14>(n) - 4>(n + kW:s;; L: A,,{I,(n)la + 14>(n + k)I"}

"ez "eZ
:s;; 2A"K" L: (1 + In l>-2. (8.7.10)
"EZ
The estimate (8.7.7) follows at once on combining (8.7.8), (8.7.9), and (8.7.10).
The final statement is a consequence of (8.7.7) and 8.7.1.
EXERCISES
8.1. Let a be any permutation of Z. Define Tf for f E L2 by
Tf = 2: /(a(n))e".
"EZ
Verify that T is an isometric isomorphism of the convolution algebra L2 onto
itself.
Construct permutations a of Z such that T is not the restriction to V~ of
any homomorphism of the convolution algebra Ll into itself.
Hint: For the second part refer to conditions (3) and (4) of 4.2.6, and
show how to construct permutations a of Z with the property that, for any
integer q > 0, the relation,
a(n + q) + a(n - q) = 2a(n)
is false for infinitely many n E Z.

8.1. Show that an element f of L2 is a prime element of L2 if and only if
~"EZ 1/(n)1 = 00. (See 8.4.)
8.3. Show that any nonprime element of L2, say f, is e:1.press:ible as the
product of two prime elements of L2.
Hint: Reduce the problem to showing that, if 2: Ic,,1 < 00, then one can
write c" = a"b", where 2: la,J~ < 00, 2: Ib,,1 2 < 00, 2: la,,1 = 2: Ib,,1 = 00.
8.4. Suppose that fELl and that wd(a) = O(lala) as a_O for some
a > O. Show that f*N E L~ for any integer N for which Na. > %. (Compare
this result with that of Exercise 5.2.)
8.5. Let k E Ll and let Tf = k * f. Show that T is a continuous endo-
morphism of L~ such that
IIfll", is defined as in 2.2.5, f

being a function on Z.
What are the eigenvalues anl eigenvectors of T? Can it ever happen that
T(L2) = L21 (give reasons for your answer). Under what conditions is T(L2)
everywhere dense in L21
8.6. The notation being as in Exercise 8.5, consider the same questions
when T - AI replaces T, ,\ being a complex number and 1 the identity
endomorphism of L2.
8.7. Let a = (a.")"EZ be a sequence such that a. E 11' for some p > 0 (the
notation being as in 2.2.5). Prove that
lim
k~'"
lIakllll/k = lIa I ""
where ak is the pointwise product of k sequences each identical with a..

EXERCISES 143
8.8. The spectral radiU8 formula for L2 (see Exercise 3.12 and Subsection
11.4.14). Show that, iff E VI', then
lim IIf*k1121/k = Ilill",.

k-oo
Hint: Use Exercise 8.7 in conjunction with the Parseval formula.

8.9. Let w be a positive function on Z such that limlnl_""w(n) = 00.
Prove the existence of continuous functions f satisfying (8.3.3) and (8.3.4).
Hint: Consider L.k=l w(nk)-Yzeln k %, where the integers n k increase
sufficiently rapidly.
8.10. The wave equation
with boundary conditions u(O, t) = U(21T, t) and initial conditions
au
u(X, 0) = f(x), at (x, 0) = g(x) ,
is to be considered in the following interpretation:

(1) for each t > 0, UI E Cl, ul(O) = u l (21T) = 0;
(2) Du, is absolutely continuous and D 2u I E L2 for each t > 0;
(3) 'ILl == L2 - lim._oe-l(ut+t - u l ) and
iii == L2 - lim ... oe-l(ul+B - UI) exist for each t > 0;
(4) iiI = D2Ut as elements of L2 for each t > 0;
(5) L2 - lim t _ +ou t = f and L2 - lim t _ + OUt = g, f and g being given
elements of V.
Give a rigorous discussion of (a) conditions under which a solution exists,
and (b) the uniqueness of this solution.
8.11. WritingsN(x) = L.lnl"wne1nx, show that ifc n = O(lnl-l)as Inl~oo,
then (SN(XN=l converges almost everywhere and that the limit function
belongs to L2.
8.11. Let (Pn)nez be a sequence of nonnegative numbers such that
2: Pn =
"eZ
00 (1)
and
2: p,,2 <
"eZ
00. (2)
Show that there exists a function f E V such that

(i) li(n)1 = o(p,,) for nEZ, Inl ~oo;
(ii) f is essentially unbounded on every nondegenerate interval.
Hints: Write N = {n E Z : p" oF OJ. Consider the linear space E of
functions f E V such that i(Z\N) = {OJ and i(n) = o(Pn) as Inl ~oo.
Introduce into E the norm

IlfilE = sup Pn -llj(n)1
neN
and verify that E is thereby made into a Banach space.

Assuming the assertion to be false, apply Appendix B.2.1 to show that
Ilfll", ~ const IlfilE. (3)
Take any g E V and consider the linear functional defined Qn E by
A(f) = LJf(-x)g(X)dX.
Use (3) and Appendix B.5.1 to show that there exists ex E tl(Z) such that
A(f) = L ex(n)Pn-1!(n)
neN
(fEE); (4)
in doing this you will need to verify that each continuous linear functional on
co(Z) (see 2.2.5) is expressible as
q, --+ L ex(n)q,(n)
neZ
for some ex E tl(Z). Deduce a contradiction of (1) from (4), using Exercise
3.14 on the way.
Remarks. The condition (2) is not essential: it is included merely to
shorten the proof somewhat. One can in any case show that f may be chosen
to satisfy (i) and (ii) and to belong to V for every p < <Xl; for details, see
Edwards [3].
8.13. Suppose that f E L2. By applying. the Parseval formula to the
function
show that
8r n~ 1!(n)12 Si n 2(;;) = LJ k~ljI(X + kr1T) -f(X + (k ~ 1)17)1 2 dx

~ o.",f (~) . V(f), (1)
whereo.",f(S) =sup{IITaf-fll",: lal ~ S}.Deducethatiffisofbolmded

variation, then it is continuous if and only if
lim r
r- co
L 1!(n)j2 sin (n217)r
neZ
2 = O. (2)
Hint: Observe that if If(~ + 0) - f(~ - 0)1 = d > 0, then, for all
large r and almost all x, at least one term in the sum appearing on the right-
hand side of (1) contributes an amount not less than (d/3)2,
EXERCISES 145
Remarks. It can be shown without too much difficulty that (2) is

equivalent to
N-'"
lim N-l L
Inl .. N
Inj(n)1 = 0; (3)
see, for example, [Bad, pp. 214-215. In this final form the criterion of
continuity for functions of bounded variation is due to Wiener. It follows in
particular tha t any function f of bounded variation, for which j (n) = o( I / Inil,
is continuous. Compare with the remarks following 2.3.6. See also MR 38
# 487; 44 # 7220.
S.14. (I) Suppose that f, g, k e L2. Prove the following extension of
Appolonius' identity:
Ilk - fl12 + Ilk - gl12 = %llf - gl12 + 211k - %U + g)112, (1)

where we have written 11'11 in place of 11'112'
(2) Let M be a closed convex subset of L2. (Convexity of M means that
exf + (I - ex)g e M whenever f, geM and 0 < ex < 1.) Show that there
exists a unique fo eM such that
Ilfoll = inf {Ilf II : f eM}.
Hints: The proof of (I) rests on direct calculation.
For (2), choose fn e M (n = 1, 2, ... ) so that
Ilfnll == Sn ~ S == inf{llfll :feM}.
Apply (1) to conclude that the sequence Un) is Cauchy in L2. Consider the
limit fo of this sequence. Prove uniqueness by another application of (1).
Remark. The result in (2) may be termed the "projection principle"; it
has numerous interesting applications to problems in concrete analysis, for a
discussion of some of which see Exercise 8.15 and [E], pp. 99 fr.
S.lS. Suppose that f e H2 (see Exercise 3.9) does not vanish almost
everywhere (that is, that Ilfll == IIfl12 > 0). Prove thatf(x):I= 0 for almost
all x.
Hints: Assume (without loss of generality) that itO) :1= o. Let M be the
smallest closed convex set in L2 containing all functionsf' t, where t denotes a
trigonometric polynomial belonging to H2 and i(O) = 1. Take geM, mini-
mizing the distance from 0 of elements of M; see Exercise 8.14. By comparing
Ilgll with Ilg + Ae"o;"gll, where A is any scalar and n = 1,2,, deduce that
Igl is equal almost everywhere to some constant, c. Check that c cannot be O.
Observe finally that, if f vanished on a set S of positive measure, then the
same would be true of g.
Remark. It can be shown that in fact
2~ flOg Ifl dx > -00.

For this, as well as many other related results, see [Hel] and [Ho]. Compare
also Exercise 15.17.
8.16. The so-called isoperimetric problem is that of determining, among all
closed plane curves with given perimeter (here conveniently chosen to have
the value 27T), that (or those) enclosing the greatest possible area.
Assuming sufficient smoothness on the part of the admissible curves, and
expressing these curves in terms of arc length s as parameter, the problem
can be formulated analytically as follows: among all pairs (j, g) of real-valued,
absolutely continuous periodic functions I and g such that DI, Dg E L2 and
(Df)2 + (Dg)2 = 1 a.e.,
determine those for which

(2"
A = Yz Jo (jDg - gDf) ds
is a maximum.
Use the Parseval formula to solve this version of the problem.
Notes: For a discussion of the isoperimetric problem for plane polygons,
see [Ka], pp. 27-34. Chapter VII of [HLP] contains a brief account of the
classical approach to this and some similar problems by the methods of the
calculus of variations. Two other problems of this sort that have had a pro-
found influence on mathematics are the Dirichlet problem and the Plateau
problem; see [CH] and the references cited there, [E], Section 5.13, and [Am].
Variational methods are applicable to the study of many linear functional
equations, in which connection they often suffice to yield useful information
about the eigenvalue distributions in cases where the equation is not con-
veniently soluble explicitly; see, for example, [So], Chapter II. The existence
theorems appropriate to many variational problems are crucial and difficult:
they represent the concrete origins of, and the initial motivation for, the modern
study of compactness in function spaces and of numerous other functional
analytic techniques. Exercise 8.14 above includes a simple variational principle,
and Exercise 8.15 illustrates a concrete application of this.
8.17. The reader is reminded that f F dx is defined (possibly 00) for any
nonnegative measurable function F on T to be lim k ... ", f inf (F, k) dx; and
that 2"&#,, is defined (possibly 00) for any nonnegative function (c,,),,&Z
on Z to be lim k ... ", 21"I<kC",
Verify that, with these definitions, the Parseval formula
21
7T
JIII2 dx = L li(n)!2
"eZ
holds for any I E Ll.

8.18. Suppose that IE Ll and that S denotes the set of real numbers a
such that T 41 - IE L2. Show that S is a subgroup of R.
EXERCISES 147
Define
for a E R, where (compare Exercise 8.17) 00 112 is taken to mean 00. Prove
that N is lower semi continuous (see Appendix A.4).
By using Exercise 3.16 and Appendix A.5 show that, if S either has positive
interior measure or is nonmeager, then there exist positive numbers 8 and A
such that
2: 1/(nW sin
neZ
2 Y2na ~ A
whenever lal ~ 8. Conclude that f E L2.

Note: It is possible to reformulate the result so as to apply when f is
assumed merely to be a distribution (see Chapter 12). Also, it is possible to
replace L2 throughout by other spaces V'; the proof then becomes somewhat
more complicated. The first consideration of this type of result appears to be
due to de Bruijn [1], [2]. See also the work of Kemperman (Math. Rev. 20,
1123), Carroll (Math. Rev. 28 #5137, 30 #1126, 2101), and a paper by the
present writer entitled "Differences of functions and measures" (to appear
J. Austr. Math. Soc.).
CHAPTER 9
Positive Definite Functions

and Bochner's Theorem
9.1 Mise-en-Scene
Continuous (and not necessarily periodic) positive definite functions of a

real variable were seemingly first studied by Bochner who, by using the
existing theory of Fourier integrals, established for them a fundamental
representation theorem now known by his name and which is the analogue
for the group R. of 9.2.8. These positive definite functions were not
seen in their true perspective until some ten or fifteen years had elapsed.
Then, as a result of the birth and growth of the theory of commutative
Banach algebras and the applications of this theory to harmonic analysis on
locally compact Abelian groups (see 11.4.18(3)), the central position of the
Bochner theorem came to be appreciated. The developments in this direction
were due largely to the Russian mathematicians Gelfand and Raikov, who
enlarged still more the role played by positive definite functions by noticing
their intimate relationship with the theory of representations of (not
necessarily Abelian) locally compact groups. A similar path was hewn,
independently, almost simultaneously, and from a slightly different point of
view, by the French mathematicians H. Cartan and Godement; see [B], pp.
220 ff. It is now true to say that a considerable portion of our function-
theoretical knowledge of locally compact groups rests upon a study of
positive definite functions on such groups.
It turns out that continuity plus positive definiteness of an integrable
functionf is one of the very few known conditions which (a) ensures that the
transform j is integrable, and (b) is expressible solely in terms of the
topological group structure. For the group T, the result is contained in
9.2.8; for the group Z, see 12.13.3. Because of this, the positive definite
concept is useful as a primitive tool in the development of harmonic analysis,
rather than as an afterthought to the latter (which is how it appeared
initially).
It is also true to say that positive definiteness of an integrable function f
is equivalent to the demand that its transform j be nonnegative. (For the
group T, see 9.2.4; the case of the group Z is covered by 12.13.2.) Granted
148
[9.2) TOWARD THE BOCHNER THEORE~f 149
the importance ofthe Fourier transformation, this fact adds to the significance
of the positive definite concept.
In this brief chapter we shall deal rapidly with a part of the theory of
positive definite functions on the group T, and observe how it might be
used as an alternative approach to the 12 theory and the Parse val formula.
(see Chapter 8). The present situation, dealing as it does with a compact
Abelian group, is technically much simpler than that involving a general
locally compact group.
Another special case of the general setting is the dual one, in which the
underlying group is Z rather than T. There is a perfect Bochner theorem
applying to his case, but its discussion must be deferred until Section 12.13.
For remarks concerning the theory in a more general setting, see Section
9.4.
9.2 Toward the Bochner Theorem

We begin by framing our definition of positive definite functions. (Bochner'S
original definition was different; see 9.2.7.)
9.2.1. A function fELl is said to be positive definite if and only if
f * u * u*(O) = 4~2 ff
f(x - y)u(x)u( y) dx dy ;:.: 0 (9.2.1)
for each continuous function u. (The reader is reminded that u* denotes the
function x-+ u( -x); see the start of Section 2.3.)
9.2.2. The set of positive definite functions does not, of course, f.arm a
linear space. It is true, however, that any linear combination with real
nonnegative coefficients of positive definite functions is again positive
definite.
It is easily verified that any continuous character is positive definite; the
same is therefore true of any trigonometric polynomial with nonnegative
coefficients.
Further examples appear in 9.2.5.
9.2.3. A function fELl is positive definite if and only if (9.2.1) holds for
each trigonometric polynomial u.
Proof. The necessity is plain. Suppose, conversely, that (9.2.1) holds for
each trigonometric polynomial u. If u is any continuous function, choose (as
is possible by 6.1.1) a sequence (Un);'=l of trigonometric polynomials con-
verging uniformly tou. Then, by 3.1.4 and 3.1.6, f * Un * u! -+ f * u * u*
uniformly. Since, by hypothesis, f * un * u~(O) ;:.: 0, it follows that
f * u * u*(O) ;:.: 0, showing that the hypothesis of 9.2.1 is fulfilled.
150 POSITIVE DEFINITE FUNCTIONS AND BOCHNER'S THEOREM
9.2.4. A function IE L1 is positive definite if and only if j ~ 0, so that in

particular I = 1* almost everywhere.
Proof. If I is positive definite, and jf in (9.2.1) we take u = en, where
en(x) = elnx , we obtain exactly the inequality j(n) ~ O. If, on the other hand,
j ~ 0, and if
is a trigonometric polynomial (the range of summation being thus a finite

subset of Z), then
1* u * u*(O) = 2: cmc" I * em * en*(0)

m.n
= 2: cmc SmJ(n)
m.Il
n '
= 2:lcn I 2j(n) ~ 0,
n
and 9.2.3 shows that I is positive definite.
9.2.5. By using the criterion 9.2.4 it is very simple to verify that for any
g E L2 the continuous function
I (x) = g * g*(x) = 27T If g(x + y)g(- y) dy

is positive definite. The same conclusion also follows readily from 9.2.6 to
follow. For the converse, see 9.2.10.
9.2.6. A continuous function I is positive definite if and only if

k
2: I (xm - xn)imz" ~ 0 (9.2.2)
m.n=l
holds for any finite sequence (x!J)~ = 1 of real numbers and any finite sequence
(zn)~ = 1 of complex numbers.
Proof. We leave it to the reader to show that (9.2.2) implies (9.2.1),
remarking merely that it suffices to approximate the integral appearing in
(9.2.1) by Riemann sums
k
2: I(x m - xn)u(xm)u(xn)(xm - x m- 1 )(xn - x n- 1 )
m.n=l
It will thus be seen that (9.2.2) implies that I is positive definite.

Conversely suppose that I is continuous and positive definite. By 9.2.4,
j ~ O. Hence uN/is a trigonometric polynomial with nonnegative coefficients,
and it is immediately verifiable that (9.2.2) holds when! is replaced by UNf.
It now suffices to let N -HI) and use 6.1.1, in order to be led to (9.2.2) itself.
[9.2] TOWARD THE BOCHNER THEORE}I 151
9.2.7. In 9.2.6 we have the exact analogue of Bochner's original definition

of continuous positive definite functions on the additive group R of real
numbers. It is not suitable for the study of discontinuous functions, which is
why we have substituted 9.2.l.
W,' can now establish the form of Bochner's theorem appropriate to the
group T.
9.2.8. (Bochner) Suppose that / ELl is positive definite, and that / is

essentially bounded on some neighborhood of the origin. Then Lnezj(n) < 00
and
/(x) = j(n)e1nX 2:
a.e.,
neZ
(9.2.3)
so that/is equal almost everywhere to a continuous positive definite function.

If / is continuous, equality holds everywhere in (9.2.3).
Proof. In the first place we have
UN/tO) = 2~ fJ(X)FN(X) dx.

Now, if we suppose that I/(x)j ~ m 00) for almost all x satisfying Ixl ~ a
for some a > 0, it follows that
m
IUN/(O)I ~ -2 fa FN(X) dx + -2
I f 1/(x)1 dx' (N + 1)-1 cosec 2 Y2a.
7T _ a 7T a.;; Ixl ';;1t
This shows that

sup IUN/(O)I < 00. (9.2.4)
N
On the other hand,
UN/tO) = 2:
Inl .. N
j(n) (I - )nl+ I)' (9.2.5)
This, together with (9.2.4) and the fact thatj ~ 0, shows that LnEzj(n) < 00.
The equality (9.2.3) is now a consequence of 2.4.2. Further, if/ is known to be
continuous, we may appeal to 2.4.3 to infer that (9.2.3) holds everywhere.
9.2.9. We notice two corollaries of 9.2.8 and its proof.

(I) If / ELl is positive definite, and if I/(x)i ~ m almost everywhere on
some neighborhood of 0, then the same inequality holds almost everywhere.
Proof. The inequality preceding (9.2.4) shows indeed that
lim sup IUN/(O)I ~ m.
N-<Xl
This, combined with 9.2.4 and (9.2.5), leads at once to
2: j(n) =
neZ
lim
N-<Xl Inl.;;N
2: j(n) (I - Nlnll)
+
~ m;
then (9.2.3) and 9.2.4 show that 1/(x)1 ~ m for almost all x.
152 POSITIVE DEFINITE FUNCTIONS AND BOCHNER'S THEORE;\f
(2) If I is a continuous positive definite function, then I/(x)1 ~ 1(0) for

all x.
Proof. This follows from (9.2.3), which now holds for all x, and 9.2.4.
(It is also easily deducible from 9.2.6.)
9.2.10. The converse of 9.2.5 can now be established with ease: any
continuous positive definite function I is expressible in the form I = g * g*
for some g E V.
In fact, by 9.2.4 and 9.2.8, j ~ 0 and
2: j(n) < 00.
neZ
From 8.3.1 it therefore follows that a function g E L2 exists such that g = j1l2,
in which case 2.3.1, (3.1.5), and 2.4.1 combine to show that I = g * g*.
The reader should compare the conjunction of 9.2.5 and its converse (just
established) with the criterion of M. Riesz mentioned in 10.6.2(4), and
applying to functions with absolutely convergent Fourier series.
9.3 An Alternative Proof of the Parseval Formula

One can combine 9.2.5 and 9.2.8 so as to yield very rapidly a proof of the
Parseval formula, on which all the L2 theory of Chapter 8 may then be
founded.
Thus, if g E L2, and if we apply to the continuous positive definite function
I = g * g* the formula (9.2.3) and take x = 0, it appears that
1(0)= 2:j(n). (9.3.1)
neZ
Since
1(0) = Lflg(x)j2 dx
and
j(n) = Ig(n)12,
(9.3.1) reads
21n flg(x)12 dx = 2: Ig(n)j2,

neZ
which is the Parseval formula.

In many developments of harmonic analysis on locally compact Abelian
groups, the above procedure is precisely that by which the Parseval formula
and the L2-theory are approached. See [B], p. 235 ff.
9.4 Other Versions of the Bochner Theorem

We have mentioned in Section 9.1 that the concept of positive definite
function extends altogether naturally to quite general groups ,and that a great
EXERCISES 153
deal of work has been done in this direction. Here we insert a few biblio-
graphical indications for the benefit of the interested reader with an ample
supply of ambition and energy.
What would appear to be perhaps the "natural" extension of the Bochner
theorem is the perfect version applying to any locally compact Abelian group.
For the details, see [B], pp. 220 ff: [R], p. 19; [N], pp. 404 ff.; [We], Chapitre VI;
[E], Sections 10.3 and 10.4; Bucy a.nd Maltese [1]; [Ph]. The third reference
displays the use of Banach-algebra techniques; see the remarks in 11.4.18(1).
Infinitesimal fragments of such extensions of the Bochner theorem appear in
Section 12.13 and in Exercises 12.34 and 12.35.
Slightly less complete versions of the theorem have been worked out for
non-Abelian locally compact groups; see [N], pp. 394 ff. and Godement [1],
especially pp. 50-53.
In pursuit of still more generality, Ito and M. G. Krein have devised formula-
tions of the theorem applying to functions on sets and spaces that are not
groups. For an example, see [N], pp. 427-428; for a brief survey, see [Hew],
pp. 145-149 and the references cited there.
Owing in part to impulses transmitted by mathematical physicists with
interests in the quantum theory of fields, special attention has been paid to
forms of the Bochner. theorem and of harmonic analysis in general applying
to Abelian nonlocally compact topological groups. A special, but particularly
relevant, such group is the underlying additive group of an infinite-dimensional
Hilbert space, for which case the reader should consult [G] and the references
there, and [GV] , Chapter IV. Possibly the Hilbert space example, although
especially significant, is too special to be genuinely typical-a state of affairs
due perhaps to the existing variety of possible methods of approach tending to
obscure the underlying essentials. Be that as it may, more general cases have
been examined with some success. For a nonlocally compact Abelian group
there may exist no genuine invariant (=Haa.r) integral having all the customary
properties (as described in, for example, Section 15 of[HR]). However Shah [1]
has shown that considerable progress can be made if there exists a suitable
stand-in for the missing invariant integral. Such a stand-in does exist for
Hilbert space and this is probably the underlying reason for success in this case.
For further developments, see MR 37 ## 1893, 5610, 5611; 40 # 6224;
51 # 13582; 52 # 14870; 55 ## 3678, 3679.
EXERCISES
9.1. Suppose that fELl and that (fn):= 1 is a sequence of integrable

functions such that limn_oo!n = J on Z. Given that each fn is positive
definite, show that f is positive definite.
9.2. Suppose that f is continuous and positive definite and that
,\ == liminfa- 2[2f(O) -f(a) -f(-a)] <00.

a-O
154 POSITIVE DEFINITE FUNCTIONS AND BOCHNER'S THEOREM
Prove that
L (1 + n2)j(n) ~ /(0) + A.
neZ
9.3. Given that / is positive definite and that / E COO(U) (or is analytic
on U), for some open neighborhood U of 0, show that / is Coo (or analytic)
everywhere.
Hint: Use Exercise 9.2 to majorize the sums Lnez InI2kJ(n) for k
= 0,1,2,.
9.4. Prove that the (pointwise) product of two positive definite functions
in Leo is again positive definite.
9.5. A positive definite function/ E Leo is said to be minimal if, whenever
gEL'" is positive definite and such that / - g is positive definite, g is equal
almost everywhere to a scalar multiple of /. Verify that the only minimal
functions are the nonnegative scalar multiples of the continuous characters
e1nx
Note: This characterization of the continuous characters is important in
the Cp..:tanGodement account mentioned in Section 9.!.
9.6. Suppose that/ E Lao is positive definite. Show that If/is equal almost
everywhere to a positive definite function in LOO, if and only if / is equal
almost everywhere to a positive multiple of a continuous character e1nx
9.7. Show that the linear subspace of C generated by the set of continuous
positive definite functions is precisely the set A of continuous functions /
such that
L IJ(n)1
neZ
< 00.
(The space A will be encountered again in Section 10.6,11.4.17, and Section

12.11.)
9.S. It is known (see 10.6.3) that there exist functions g E Ll such that
1!l1
is not the Fourier transform of any Llfunction. Give examples of this
phenomenon for the case in which Ll is replaced throughout by C and by Lao.
CHAPTER 10
Po}ntwise Convergence
of Fourier Series
In this chapter we shall deal rather summarily with some positive and
negative results about the pointwise convergence of Fourier series. The
reasons lor not according this topic a fuller treatment are discussed in
Chapter 1. The reader who is particularly attracted by these aspects may
consult [Z], especially Chapters II, VIII, and XIII; [Ba], especially Chapters
I, III-V, VII, IX; [HaR], especially Chapter IV; [I], pp. 23 If., pp. 103 If.;
[A]; and the work of Carleson mentioned in 10.4.5.
Our account is particularly terse in relation to the many known sufficient
conditions for convergence at a particular point. Out of a veritable multitude
of such results, increasing almost daily, we shall in fact prove only the very
familiar criteria associated with the names of Jordan and Dini, respectively.
These are perhaps the most useful aids to handling the functions that occur
naturally and with appreciable frequency.
On the other hand, partly in order to reinforce the remarks in Chapter 1
that bear upon the difficulties of characterizing Fourier series directly in
terms of pointwise convergence, and partly to exhibit some of the chara-
teristic functional analytic techniques of modern analysis, we shall devote
quite a large fraction of the chapter to telling part of the story of the mis-
behavior of Fourier series in relation to pointwise convergence.
Although our main concern throughout this book is with Fourier series of
functions on T, it is worth noting at this point that, rather late in the his-
torical development of harmonic analysis, it came to be recognized that the
classically natural groups T and its finite powers are" difficult" in respect of
questions about pointwise convergence of Fourier series. More precisely,
there are compact Abelian groups which are, from a classical viewpoint,
rather bizarre, and yet relative to which Fourier series behave in a fashion
simpler and more civilized than they do when the underlying group is T. The
simplest (infinite) instance of such a group is the so-called Cantor group rc,
which will be discussed in Chapter 14 as an aid in the study of Fourier series
of functions on T. It turns out that, if f is any continuous complex-valued
function on rc, then the Fourier series of f converges uniformly to f.
Concerning the dual aspect of the problems handled in this chapter, see
the remarks in Section 6.7.
155
156 POINTWISE CONVERGENCE OF FOURIER SERIES
The reader is again reminded that, in the absence of any statement to the
contrary, the convergence of a numerical series L:nezcn is defined to signify
the existence of a finite limit
A similar convention applies if the cn are elements of anyone of the topological

linear spaces of functions, measures, or distributions to be encountered in
due course. If the Cn denote arbitrary nonnegative real numbers, the above
definition of the sum L:nezCn makes it equal to
sup
F
L
neF
Cn'
where the supremum is taken with respect to all finite subsets F of Z. In

this case, if the series is not convergent, the set of such finite partial sums is
unbounded above and, by universal convention, one then writes
Lc
neZ
n = 00.
The sum of a series of nonnegative real numbers is thus assigned a finite

value if and only if it is convergent.
10.1 Functions of Bounded Variation and Jordan's Test

10.1.1. (Jordan's test) If fE Ll is of bounded variation on some neighbor-
hood of a point x, then
lim 8 N f(x) = %[f(x + 0)+ f(x - 0)].
N->
Proof. Since ([W], Lemma 6.4b) the real and imaginary parts of fare
each expressible, throughout some neighborhood of x, as the difference of
two monotone functions, both limits f(x + 0) == lim" +of (x + a) and
f(x - 0) == lim" +of (x - a) exist finitely.
It is in any case enough to handle the case in which f is real-valued, which
we assume henceforth. Now
8N f(x) = ;1T f f(x - y)D N ( y) dy

1 ("
= 21T Jo [f(x + y) + f(x - y)]DN(Y) dy,
and it will therefore suffice to show that
lim -2
1
N-> 1T
i"0
g( y)D N( y) dy = %g( + 0) (10.1.1)
for any real-valued function 9 which is integrable over (0, 1T) and of bounded
variation on some right-hand neighborhood [0, S] of 0, where 0 < [j < 1T. In
doing this we may, without loss of generality, assume that 9 is increasing on
[10.1] FUNCTIONS OF BOUNDED VARIATION AND JORDAN'S TEST 157
[0, S] and that g( +0) = O. [The latter reduction is possible in view of the
obvious fact that (10.1.1) holds when g is a constant; see (5.1.3).] This being
so, the second mean value theorem of the integral calculus gives
~ (" g(y)DN(Y) dy
Jo
=~ r + ~217 J6("
Jo
5: r
21T 21T
= 2~ g(S - 0) DN( y) dy + 2~ g( y)DN( y) dy
for some ~ satisfying 0 ~ ~ ~ S. This ~ may depend upon N. Thus
12~f: g(Y)DN(Y)dyl ~ 2~g(S - 0)15: DN(Y)dyl+ 12~r g(Y)DN(Y)dyl

(10.1.2)
Now
I ~ J(6~ D N( y) dyl
21T
= I! J
1T
(6 sin (N .+ %)y dYI ~
~ 2 sm%y 1T ~
I! Jr 6
Y
sin (N + %)y dyl
+ -1 f" ly-1 - %cosec %yl dy ~ - 11 i(N+%J6 sintdtl

--
1T ~ 1T (N + %)~ t
+ -1
1T
J" Iy
0
-1 - %cosec %yl dy.
Since y-l - %cosec %y is integrable over (0, 1T), 10.1.2 to follow shows that
(10.1.3)
where A is independent of e, S, and N.

Reverting to (10.1.2), and assuming that e > 0 is assigned, we first fix
S > 0 so small that Ag(S) ~ e [which is possible since fl( +0) = 0], so
obtaining
(10.1.4)
Since g(y) cosec %y is integrable over (S, 1T), we may allow N to tend to
infinity and apply 2.3.8 to the second term. on the right in (10.1.4) and 80
conClude that
lim sup r 21
N-aJ ..". Jor" g( y)DN( y) dyl ~ e.
Since e is freely chosen, (lO.l.1) is thus established. See also Izumi [1].
10.1.2. The integral f: t- 1 sin t de is bounded for all real values of a and b,
that is,
sup
4. bell
lIb t-1sintdt\ <
4
00.
Proof. This is left as Exercise 10.1 for the reader.
10.1.3. As has been indicated in 5.3.5, by using 2.3.6, 5.2.1 and a Tauberian
theorem of Hardy (see Exercise 5.8), 10.1.1 could be inferred directly from
6.3.1.
An examination of the proof of 10.1.1 leads to the following global version
thereof.
10.1.4. If I is of bounded variation, then

lim sNI(x) = Y2[f(x + 0) + I(x - 0)]
N .... '"
for all x, and the convergence is bounded:

ISNI(x)1 ~ const (11111"" + VU, (10.1.5)
where VU) denotes the total variation of lover any interval of length 217.
Proof. Only the inequality has to be proved. A perusal of the proof of
10.1.1, and a glance at the way in which a function of bounded variation can
be expressed as the difference of two increasing functions ([W], p. 105), show
that it will suffice to prove that
12~ s: g(y)DN(Y) dy I~ const sup {Ig(y)i : 0 ~ y ~ 17} (10.1.6)
for any increasing function g on [0,17] such that g( +0) = O. Now the second
mean value theorem of the integral calculus gives for such g the relation
for some a in [0, 17], so that (10.1.6) is an immediate consequence of (10.1.3).

Remarks. (1) A different proof is easily derived from Exercise 10.12.
(2) For functions f which are both continuous and of bounded variation,
it is true that limN .... ""sNI = I unIformly. See Exercises 10.13 and 10.14.
10.1.5. If IE Lt, then L.... o J(n)ei"%/n is uniformly convergent for all x.
r
Proof. The function g defined by
g(x) = I(y) dy - J(O)x
is periodic and absolutely continuous, and its Fourier series is
L (in)-lJ(n)e inX ;
" .. 0
see 2.3.4. The result follows on applying Remark (2) following 10.1.4.
[10.2] REMARKS ON OTHER CRITERIA FOR CONVERGENCE 159
10.1.6. Remarks. (1) The proof of 10.1.5 can be extended to show that
L.n .. o{l.(n)/n is convergent for any measure fL (see Sections 12.2 and 12.5).
(2) From 10.1.5 it follows that L.:'=2 sin nx/log n is not a Fourier-Lebesgue
series. (Nor, by Remark (1) immediately above, is it even a so-called
Fourier-Stieltjes series; see 12.5.2.) See also 7.3.4 and Exercise 7.7.
10.2 Remarks on Other Criteria for Conver~ence; Dini's Test

We adopt the notation introduced in Section 6.3, writing in particular
f~(y) == f~(y, x) = %[f(x + y) + f(x - y) - 28]. (10.2.1)
Then, parallel to (6.3.2), we have
(10.2.2)
Sincef!( y) cosec %y is integrable over (8, 17") for any 8 satisfying 0 < 8 ~ 17",
2.3.8 .shows that
(0 < 8 ~ 17"). (10.2.3)
An immediate corollary of this is the following statement.
10.2.1. Iff ELl, in order that

lim 8 N f(x) = 8, (10.2.4)
it is necessary and sufficient that for some 8 satisfying 0 < 8 ~ 17" it is true
that
lim -1 f6 n(y)DN(y) dy = O. (10.2.5)
N- 00 '1T 0
Beside this we may infer
10.2.2. In order that (10.2.4) be true, it is sufficient that to any e > 0 shall
correspond a number 8(e) satisfying 0 < 8(e) ~ 17" and a positive integer
N(e) such that
1 ;.1 Jo(6() f~( y)DN( y) dy I ~ e for N ~ N(e). (10.2.6)
Proof. In this case

160 POINTWISE .CONVERGENCEOF FOURIER SERIES
for a.ll N ~ N(e), whence: us~ of 2.3.8 yields from (10.2.2)

lim sup Is~f(x) -sl ~ e.
N .. ..,
On letting e_ 0, we obtain '(10.2.4).
10.2.3. (Dini's test) In order that (10.2.4) be true, it is sufficient that for
,I:
same 8 satisfying 0 < 8'~:1rdne has
In(~)1 dy < 00. (10.2.7)
(Compare this with 5.2.3.)

Proof. Since y - 1 - y:! cosec y:! y is bounded over (0, 8), (10.2.7) and
2.3.8 combine to yield (10.2.5). The result therefore follows from 10.2.1.
10.2.4. Remarks. (1) If (1.0.2.7) holds for any value of sat all,and iff
has at x at \vorst a 'jump;discontinuity, tl~e'value ?f s must be Y:!U(x + 0)
+ f(x - 0. .
(2) It is evident that (10.2.7) is fulfilled, with 8 = f(x), if, for example,
l(x + y) - f(x) = 0(1 ylti) for some ex > o.
10.3 The Divergence of Fourier Series

In this section we shall assemble a few results concerning the: pointwise
4.ivergence of the Fourier series, of fun9ti9I),S of various types. For many
more details the reader is referred to[Z~], Chapter VIII and [Bad, Chapters
~&nc~ Y. Concerning,mean c9nverg~nre in.L\ seeEx~,ci~e,IP.2 and 12.10.2.
, On~ niay seek to s~pport ~ state~ent of the type: "The Fourier series of a
continuous function may diverge" in one of at least two ways. Either one
may try to construct in as' explicit a:manri.:eraspossible a specific continuous
function with the desired property; or one may use, a reductio ad absurdum
argument by showing that the hypothesis, that no such functions exist, l~ads
to a contradiction of what is already known. In the former case one has (if the
alleged construction is su,cc~sful) II; cOll,structive proofof tile,statement; in
the latter case one '~as (as!'luming 'that no euofs are ,made on the way) an
exi~tential proof of 'the statement, so-called because one 'has shown that
functions of the specified type must exist without prescribing any way of
finding one. In 6.1,4, we h,ave already encountereda:nexample of the second
type of argument. Both types of proof have their appeal. A constructive
proof is usually more satisfying, but it is often ruled out by h,lCk of,enough
detailed information about certain elements of the proposed construction.
This is frequently the case in' ~bstract an~lysis, and this is where the existential
type of proof often comes to the rescue.
[10.3] THE DIVERGENCE OF FOURIER SERIES 161
A further illustration is provided by the assertion: "There exist integrable

functions whose Fourier transforms tend to zero arbitrarily slowly at
infinity." For the group T, a constructive proof of a strong form of this
assertion is included in Section 7.4 and depends on many considerations
peculiar to this group. There is a meaningful analogous assertion for any
nondiscrete locally compact Abelian group. This analogue can be painlessly
established (see Exercise 10.22) by an existential proof using abstract
arguments; by expending more effort and using the properties of Sidon sets
(see the introductory remarks to Chapter 15), more or less constructive
proofs may also be furnished.
We now proceed to illustrate both types of proof in connection with state-
ments about the divergence of Fourier series.
10.3.1. Fejer's Example. We begin with a construction due to Fejer

leading to continuous functions whose Fourier series diverge at a given
point, which may without loss of generality be taken to be the origin.
If p and q are integers satisfying p ~ q ~ 1, let tp,q denote the trigono-
metric polynomial defined by
t (x) = cos (p - q)x + ... + cos (p - l)x _ cos (p + l)x _ ...

p,q q 1 1
cos (p + q)x
q
L k-
q
= 2 sinpx 1 sin kx. (10.3.1)

k=l
In view of Exercise 1.4, the t p q are uniformly bounded.
Suppose now that (Pk)k'= 1 and (q,Jk'= 1 are sequences of integers satisfying
(10.3.2)
Suppose further that (ee,,)k'= 1 is any sequence of complex numbers such that
(10.3.3)
liminfleekllogqk > O. (10.3.4)

k-oo
One might take, for example,
Consider the function
f(x) = L'"
/(=1
Ct.k tpk,qk(X). (10.3.5)
Because of (10.3.3) the series converges uniformly, so that f is continuous.

Due to uniform convergence, the Fourier coefficients of f may be calculated

by termwise integration of the series appearing in (10.3.5). One finds in this
way that
\SPk +qJ(O) - spJ(O}i
= \ak\ll- 1 cos (Pk + 1)0 + ... + q;;l cos (Pk + qk)O\
qk
= lak\ L: r-
r=1
1
"'" \ak\log qk. (10.3.6)

In the course of this calculation, the relations (10.3.2) are used in ensuring
that the various trigonometric polynomials lpk.qk have no "overlapping"
harmonics. The relations (10.3.6) and (10.3.4) shows that the sequence
(sNf(O)); = 1 is not convergent, that is, that the Fourier series of f is not
convergent at the origin. Indeed, if we choose the Ctk so that
lim sup lakllog qk = 00,
k~l
it follows from (10.3.6) that the partial sums of the Fourier series of fare
unbounded at the origin.
Numerous variations may be played on the preceding construction; see
[Z1]' pp. 299-300; [Bad, 45; [Kz], p. 51, Proof B; Edwards and Price [1].
10.3.2. Existential Proofs. In this and the following subsection we shall

use the uniform boundedness principle as the basis of existential proofs of
the statement asserting the possible divergence of the Fourier series of
continuous functions.
The aim of this subsection is as follows. Let (xk)k'= 1 be any sequence of real
numbers and (PN); = 1 any sequence of positive real numbers such that
(10.3.7)
C denotes the Banach space of continuous (periodic) functions; see 2.2.4.

Our claim is that for each fEe, save perhaps those of a meager (= first
category; see Appendix A.l) subset of C, it is the case that
lim sup \sNf(Xk)1 = 00 (k = 1,2,). (10.3.8)

N~'" PN log (N + 1)
Since C is not a meager subset of itself (Appendix A.3), this shows that
continuous functions f certainly exist for which (10.3.8) is true.
Proof. We shall apply the result stated in Appendix B.2.1, taking for the
Frechet space featuring therein the Banach space C, and defining
ISNf(xk)1
Pk(f) = !~~ PN log (N + 1)
This Pk has all the properties demanded, lower semicontinuity being a

consequence of the fact that
is evidently a continuous linear functional on C. The set of f E C satisfying

(10.3.8) is exactly the complement, relative to C, of the set
S = {fE C : infpk(f) < oJ}.
k
It therefore suffices to show that S is meager.

Now, if S were nonmeager, Appendix B.2.1 would entail that for some k
there is a constant c such that
for all f E C. This would signify that
(10.3.9)
Appealing to transI'ation invariance of the integral, combined with the

converse of Holder's inequality (see Exercise 3.6), it would appear from
(10.3.9) that
or, by translation invariance of the integral once again, that

(10.3.10)
But (10.3.7) and (10.3.10) flatly contradict (5.1.10), which says that
4
IIDNlll '" "2 log N
TT
(N -00).
TtJ.is contradiction establishes the desired result. See also [Kz], p. 51, Proof A.
10.3.3. Let (Xk):'= 1 and (PN)'; = 1 be as in 10.3.2. There exists a meagre

subset S of C with the following property: iff E C\S, one has
(xEE), (10.3.11)
where E is a (possibly j-dependent), set which is everywhere dense and whose

complement is meager (so that E is nonmeager and therefore uncountable),
and such that X k E E for all k.
Proof. We may suppose from the outset that the points Xk are every-
where dense in [ -TT, TT] and that all values of x considered lie in this same
interval. For S we choose the set specified in the proof if 10.3.2, where it is
shown that S is meager. Take f E C\S and define
w(x) = sup ISNf(x)1

N;'l PN log (N + 1)'
so that w(XI<) = 00 for k = 1,2, . '. If E is defined to be the set of points

x E [ -TT, TT] for which w(x) = 00, then E contains every xI< and is therefore
everywhere dense in [-TT, TT]. Moreover, if
Er = {x E [-TT, TT] : w(x) > r} (r = 1,2,),
then Er is open relative to [-TT, TT] (since w is evidently lower semicontinuous),

Er contains all the XI<, and is thus everywhere dense in [-TT, TT], and
E = n:,= 1 E r Taking complements relative to [ -TT, TT], the complement of
E is the union of the complements of the E r Each of the latter is closed and
nowhere dense, so that the complement of E is meager. By the category
theorem (Appendix A.3), E must therefore be nonmeager; and since [ -TT, TT]
is nondiscrete, this entails that E is uncountable.
10.304. Further Results. Evidently, 10.3.3 entails the existence of

continuous functions f whose Fourier series diverge on sets E that are un-
countable, nonmeager, and everywhere dense. (The existence of continuous
functions whose Fourier series diverge at a specified point was known to du
Bois-Reymond in 1872.)
The examples in 10.3.2 and 10.3.3 are aU such that ISNf(x) I is unbounded
with respect to N for certain values of x. It can, however, happen that (sNf(x))';= 1
is boundedly divergent for each point x in a set with the power of the continuum,
and this for suitable continuous functions f; for an example, sec [Bad, p. 348.
It results from the work of Carleson, described in 1004.5, that the Fourier
series of any continuous function converges pointwise almost everywhere to
that function.
Men'shov showed in 1947 that there exists a continuous functionf such that
any subsequence of (8Nfl: = 1 diverges at some point. He also established the
curious fact that any fEe can be decomposed into a sum f 1 + f 2' where fl E C
and some subsequence of (sNfl):=l is uniformly convergent (i = 1,2).
There are very few simple operations on (say) continuous functions that
preserve convergence of the Fourier series. Thus, for example, there exists at
least one fEe having a uniformly convergent Fourier series, while the Fourier
series of f2 diverges on a set having the power of the continuum; and also a
similar f such that the Fourier series of If I diverges at some points ([Bad, p.
350; p. 360, Problem 14). For further results of this type, see Kahane and
Katznelson [1].
In 1926 Kolmogorov showed that there exist integrable functions f whose
Fourier series diverge everywhere. The proof, which is more or less con-
structive but complicated and difficult, is given in [Zd, pp. 310-314 and
[Bad, pp. 455-464; [HaR], Theorem 79, is an earlier construction of Kolmogorov
designed to produce integrable functions whose Fourier series diverge almost
everywhere. That such functions exist has been established by Stein ([1],
Theorem 6) on the basis of a general theorem which is cited in 16.2.8 and which
provides a powerful general approach to many existence theorems of this type.
Granted the existence of one such function, it is relatively simple to deduce
that they exist in abundance; see Exercise 10.21. See also Chen [1] and M. and
S.-1. Izumi [2].
Regarding similar results when the operators 8N are replaced by something
similar but more general, see Exercises 10.23 and 10.24.
For Kolmogorov's functionj, as for those in 10.3.2 and 10.3.3, it is the case
that 8Nj(X) is unbounded. In 1936 Marcinkiewicz showed that integrable
functions j exist for which 8Nj(X) is boundedly divergent for almost all x (see
[Zd, p. 308, and [Bad, pp. 430-443), but it is apparently still unknown whether
"almost all" can here be replaced by "all."
There is an elegant discussion of divergence for homogeneous Banach spaces
(including the Kolmogorov example cited above) in [Kz], pp. 55-61; see also
[Moz], Appendix C.
See also MR 33 # 6266; 35 # 3349; 39 # 7342; 41 # 8906; 51 # 13578.
10.3.5. Majorants for the sNf. The example given by Kolmogorov and
mentioned in 10.3.4 shows that, if we write
s*f(x) == sup ISNf(x)1 (~oo),
N
theri there exist integrable functions f such that
s*f(x) = 00
for all x. The readel' should compare this with the results quoted in 6.4.7
and relating to the majorant a*f of the Cesaro means aNf. See also 10.4.5
and Exercise 10.2, and 13.10 below.
In spite of this disconcerting situation, there are some remarkable assertions
of a type which say that, iff is real-valued and if the sNf are bounded below
in a suitable sense, then they are also bounded above in a corresponding
sense. A typical statement of this type affirms that if fEll is real-valued
and such that infN8.vf Ell, then SUPNSNf E LP for any p satisfying 0 < p < 1.
For other similar results and the relevant details, see [Z2]' pp. 173-175.
10.3.6. Topolo~ical Bases of Tri~onometric Polynomials. Consider a

Banach space E and a sequence (an)nez of elements of E. This sequence is said
to form a topological base for E if to each j E E corresponds precisely one
sequence (an)nez of scalars such that
(10.3.12)
the limit being taken with respect to the norm on E. A similar definition applies
to sequences (an):=l' (The reader will recall that a not necessarily countable
family (an) of elemcnts of E forms an algebraic, or Hamel, base for E if each
Je E is expressible uniquely as a finite sum of terms a"a", the coefficients a..

beingf-dependent scalars.) For more details about topological bases in general,
see, for example, [E], Section 6.S. The existence of topological bases has been
settled only for a number of particular (albeit especially important) Banach
spaces; but see MR 35 # 700.
Let us now specialize by 8.S8uming that E is one of the spaces LP, where
1 E; p E; co, or C, and a" = 6 .., the function :l:- 6 1'''':. It will appear in 12.10.1
that (6"),,ez is a topological base for LP whenever 1 < P < co. On the contrary,
however, this sequence is not a topological base for anyone of the spaces Ll,
L"', or C; see Exercise 10.15.
Taking the case of C, the question arises as to whether there exists any
topological base (t n ):'=l for C in which each t .. is a trigonometric polynomial;
and, if so, what can be said about the degree d n of tn. It has emerged (see
[Ball, p. 360, Problem 16) that one cannot have d n E; n for all n; and yet
(loc:. cit., Problem 17) that there exists a topological base (t n ):'= 1 for C comprised
of trigonometric polynomials for which d n = o(n2+1) for every II > O. As yet
no conditions are known concerning the degrees d" which are necessary and
sufficient to ensure the existence of a topological base for C formed by
trigonometric polynomials tn of degree dn'
In 2.2.1 we mentioned the question of the possibility of decomposing C into a
direct sum of minimal translation-invariant subspaces. In view of 2.2.1(2), such
a decomposition is p088ible if and only if (6 n )"ez is a topological base for C.
Accordingly, by Exercise 10.15, no such decomposition is possible for C. By the
sam' token, the analogous proc'dure is impossible for Ll and Lx;, too. On t.he
ot,her hand. t.hE' fact that, thl' en do form a topolo~ical base in LP (1 < P < co),
ensures t.hat, t.he analogous (lecomposition is possibll' in ('ach of the'se spaces.
lOA The Order of Magnitude of SNj. Pointwise Convergence

Almost Everywhere
This section begins with a result to the effect that the misbehavior of
Fourier series shown in 10.3.2 and 10.3.3 to occur for certain continuous
functions is, in a sense, the worst that can happen even for arbitrary
integrable functio~s. Knowing'this, it is possible to infer the convergence
everywhere or almost everywhere of certain series simply related to Fourier-
Lebesgue series (see 10.4.3 and 10.4.4). This in turn leads to the consideration
of conditions on a sequence (cn)nez sufficient to ensure the convergence almost
everywhere of the trigonometric series
Within the circle of ideas thus suggested appear certain problems concerning
the pointwise convergence of Fourier series which, while remaining intrac-
table for a very long while, provided the incentive for a great deal of
fruitful research; see 10.4.2 and 10.4.4 to 10.4.6.
[lOA] THE ORDER OF MAGNITUDE OF 8Ni 167
10.4.1. Iff ELl, then the relation

8N f(x) = o(log N) (10.4.1)
holds
(1) whenever f(x
finitely, and
++0) f(x - 0) == limlll o[f(x y) ++ f(x - y)] exists
(2) for almost all x in any case.

Proof. (1) Suppose that the said limit exists, and let 8 denote its value.
By (10.2.2),
18Nf(x) - 81 = I~ f n(y)DN(y) dyl

~1~fl+I~JJ'
where S > 0 will be chosen in a moment.
By hypothesis, n(
y) --* 0 as y t O. So, given e > 0, we may choose and
fix S > 0 so small that In(y)1 ~ e for 0 ~ y ~ S. Then
I~fl ~ e'~fIDN(Y)ldy ~ AelogN,
by 5.1.1. A denoting an absolute constant. Having fixed S, 2.3.8 shows that

(1/1T) J~ = 0(1) as N --*00. One thus obtains
18Nf(x) - 81 ~ Ae log N +e
for all sufficiently large N, which implies (10.4.1).
(2) In the general case, we proceed as before save that the range of
integration (0, 1T) is divided into (0, 1T/N) and (1T/N, 1T). Using the estimates
IDN(y)\ ~ AN, \ D",(y) 1 ~ A/y, where A is again a positive absolute
constant, we find that
A -118",f(x) - 81 ~ ~ ("IN
1T Jo
In(y)IN dy +~
1T
f"
nlN
In(y)\ dy
Y
(10.4.2)
The choice of the number 8 is so far immaterial. Defining, as in (6.4.6),
J(h) = f In(y)1 dy,

we have
Jl~~NJ(;).
Moreover, by partial integration,
Thus, by (10.4.2),
A -II SN f()
x - SI ::;; -1 Nl(7T)
- + -1(7T) If" l(y)dy
2 + - --2-' (10.4.3)
7T N 7T 7T "IN Y
According to 6.4.2, if we take S =f(x), then l(y) = o(y) as y t 0 for

almost all x. For any such x, (10.4.3) shows that
ISNf(x) - f(x)1 = 0(1) + o( r" dy ) = o(log N),

J1IIN y
as required.
10.4.2. Remarks. It is apparently unknown whether 10.4.1 is the best-

possible result of its kind. However, it is known (Stein [1], Theorem 6;
Carleson [1]) that, if EN t 0 as N t 00, then there exists at least one fELl
for which the relation
sNf(x) - sN'f(x) = O{EN_N' log (N - N')} (N > N' + 1, N'-H:O)
is false for almost all x; and that there exists at least one f E V for which the
relation
sNf(x) = O(eN log log N) (N -+(0)
is false for almost all x.
10.4.3. Iff ELI and ex > 0, each of the series
2: j(n)e inZ , 2: j{n)e inX

neZ log (2 + In!) neZ (I + InW
is convergent
(1) wherever f(x + 0) + f(x - 0) exists finitely, and
(2) for almost all x, in any case.
Proof. This follows from 6.3.1,6.4.4, 10.4.1, and Exercise 7.4.
10.4.4. The Case p > l. With somewhat more effort it is possible to

improve (10.4.1) for the case in which f E LP for some p > l. For example, it
was proved long ago by Littlewood and Paley that
if f E LP, 1 ~ P ~ 2 (10.4.4)
for almost all x. Yet, for an equally long time, no success attended attempts to
establish (10.4.4) for the case where p > 2, or even to show that (10.4.4) holds
for fEe and p > 2; see [Z2]' pp. 161-162, 166-167. The present position will be
outlined in Subsection 10.4.5.
It is relatively simple to show on the basis of 10.4.3 that the trigonometric
series
(10.4.5)
[10.4] THE ORDER OF MAGNITUDE OF BN! 169
converges almost everywhere whenever
2:
neZ
Icn l2 log2 (1 + In/) < 00; (10.4.6)
this result, due to Hardy, is the forerunner of more elaborate investigations

mentioned immediately below and in the next two subsections.
On the basis of (10.4.4) it can be shown ([Z2], p. 170) that
2: !(n)e1nX
neZ [log (2 + Inl )]l/P
converges almost everywhere whenever f E LP and 1 ~ P ~ 2; and that
(10.4.5) converges almost everywhere provided
2:
neZ
Icn l2 log (1 + In/) < 00. (10.4.7)
This last result is due to Kolmogorov, and Seliverstov, and Plessner; see [Z2],
p. 163 or [Bad, p. 363.
On the other hand, it is known (see [Bad, p. 483, Problem 1) that there exist
functions fEe such that
2:
neZ
1/(n)1 2 Iog (1 + In/) < 00
and yet the Fourier series off diverges at infinitely many points.
10.4.5. Lusin's Problem and Carleson's Theorems. The results

mentioned in Subsections 10.4.1 to 10.4.4 were by-products of a prolonged
and arduous study of a problem posed by Lusin in 1915, namely: does the
condition
(10.4.8)
suffice to ensure the convergence almost everywhere of the series (10.4.5)?

In other words, does the Fourier series of any function in L2 converge almost
everywhere? It is natural to ask more generally: does the Fourier series
of any function belonging to V' for some p > 1 converge almost everywhere?
Over a period of fifty years, an enormous amount of effort was expended
on efforts to solve these problems. Some description of the situation prevailing
until 1966 appears in [Z2]' pp. 165-166 and in the relevant portions of [Li).
In 1966 Carleson [1] announced an affirmative answer to Lusin's question,
together with major improvements on (10.4.4). His results, the proofs of
which are far too long and complicated to be given here, are as follows:
(1) ifJ(log+ IJ1l1+6 ELI for some S > 0, then
8... J(X) = o(log log N)

for almost all x;
(2) if 1 E V' for some p > 1, then

8N I(x) = o(log log log N)
for almost all x;
(3) if 1 e L2, then
lim
N .....
8 N I(x) = I(x)
for almost all x.
Subsequently, R. A. Hunt proved that in (3) the exponent 2 may be re-
placed by any exponent p > 1. For proofs, see (Ga] and/or (Moz], MR 49
# 5676. See also MR 52 # 6300 for an illuminating discussion.
We here remark merely that, granted the convergence almost everywhere
of the Fourier series of each 1 E V' for any fixed p satisfying 1 < P ~ 2, a
theorem of Stein stated in Subsection 16.2.8 entails that the maximal
opera.tor
8*:/-+8*1
defined in Subsection 10.3.5 is of weak type (p, p) on V', that is, there exists a
number AI' such that
m({xe[0,211): 8*/(x) > A}) ~ ApA-Pll/ilpP (10.4.9)
for each A > 0 a.nd eachl E LP; and that to each numberq satisfying 0 < q <p
corresponds a number Ap,q such that
(10.4.10)
for each 1 E V', The numbers Ap and ApQare independent off. In (10.4.9), m
denotes Lebesgue measure. Concerning these matters, see Subsection
13.10.2.
Beside t.his, if 1 A}) ~ (const 1I/IIp/A)Q
for each number A > 0 and each 1E LP, it is a relatively easy task to deduce
that limN ..... SNI(x) = I(x) for almost all x whenever 1e LPj see Exercise
13.26.
10.4.6. Sets of Dlver.tence. What has been said about consequences of
conditions like (10.4.6) and (10.4.7) will explain the intense interest that has
been shown in the following type of problem. Take a weight 8equence W defined
on the nonnegative integers and such that WeN) increases to infinity with N.
H we suppose that
~ Ic"I!3W(lnj) < 00, (10.4.11)
"EZ
what can be said regarding the size of the set E of points of divergence of the
aeries
[10.5] MORE ABOUT THE PARSEVAL FORMULA 171
It has turned out that, at least when W(N) increases more rapidly than log N
(compare (10.4.7)), new concepts of the thinness of sets, more refined than, that
provided by vanishing of the Lebesgue measure, come into play. Owing to a
strong formal similarity with definitions in potential theory, the appropriate
refined set functions are called capacities; see Section 12.12. The general type of
conclusion to be drawn from (10.4.11) is that E is a set of zero capacity, where
the particular type of capacity referred to depends on the weight function W.
Several results of this sort are known and due to Beurling, Salem, and Zygmund
and Temko (1940 onward). For details, see [Bad, pp. 398-414; [KS], Chapitre IV.
As has been seen in Section 8.6, if (10.4.5) is the Fourier series of some
function f E L2, then BNJ --+ f almost everywhere for any Hadamard sequence
(N Ie);:= 1 of positive integers (that is, any sequence of positive integers for
which inf N1<+l/N" > 1). Also, by the Kolmogorov-Seliverstov-Plessner
theorem, if (10.4.7) holds, then BNf--+f almost everywhere. To Salem we owe
a general investigation of such assertions, based upon hypotheses of the type
(10.4.11). This study produced further conditions on the sequence (N");:=l
sufficient to ensure that (10.4.11) entails that BNkf --+ f almost everywhere. One
such condition is that
for some A > 0; the case W(N) = log N includes the Kolmogorov.Seliverstov-
Plessner result. Perhaps even more remarkable is the fact that Salem obtained
conditions on the N" sufficient to ensure that BN.! --+ f almost everywhere for
any preassigned fELl: the condition reads
where wdis defined as in 2.3.7. The details are presented in [Bad, pp. 389-397.
10.5 More about the Parseval Formula

Certain cases of the Parseval formula have already been discussed in 6.2.5
and Section 8.2 and we now extend the discussion a little. In all the cases we
have to consider it is immaterial whether we take the (polarized) formula
to be (8.2.4) or (8.2.5). For definiteness, we choose the latter. Thus we shall
be concerned with the formula
21
~
f
j(x)g(x) dx = ". .
L.., j(n)g( -n).
neZ
(10.5.1)
10.5.1. Ifj ELI and g E BV, then (10.5.1) holds, the series being convergent.
Proof. We have seen in 10.1.4 that
SNg(X) - Y2{g(x + 0) + g(x - O)}
boundedly. Moreover, the right-hand side here is equal to y(x) save perhaps
for a countable (and therefore null) set of x. So ([W], Theorem 4.lb) we have
lim 21
211r flY dx = N~CIO T1'
JI' 8 NY dx
= lim
N
2: i( -n)g(n) ,
Inl .. N
which is equivalent to (10.5.1).

Remark. As was remarked to me by Professor Goes (private corre-
spondence) (10.5.1) holds for all I e Ll, if and only if
sup ~8NUII '" < 00;
N
cf. Remark 6.2.6(2) above.
10.5.2. If Ie Ll and Y e Leo, the series 'i.nEzi(n)g( -n) is Cesaro-summable

to (1/27T) JIy dx.
Proof. This proof is similar to that of 10.5.1, the sole difference being
that 8 NY is replaced by aNY and that appeal is made to 6.4.4 and 6.4.7 in place
of 10.1.4. More simply, an appeal to 6.1.1 suffices; see 10.5.3.
10.5.3. If 1 < P < 00 andieV', yeV", then the series 'i.nEzi(n)g(-n) is

Cesaro-summable to (1/27T) JIy dx.
Proof. Once again the same method is used, but now we appeal to the
fact that aNY _ Y in mean in V" (see 6.1.1) coupled with Holder's inequality,
which shows that
10.5.4. Remark. It is actually the case that, under the hypotheses of

10.5.3, the series 'i.nEzi(n)g( -n) is convergent. This is so because, if leV'
and 1 < p < 00, then 8NI converges in mean in LP to I (which assertion is
false if p = 1 or (0). The proof of this result will be given later, in Section
12.10.
10.5.5. Neaative Results. (1) It is not true that the series 'i.1Iezi(n)g( -n)
is convergent whenever I eLl and ye C; see Exercise 10.7.
(2) By 10.5.4 we know that 'i.nEzi(n)g( -n) is convergent whenever
1 < P < ooandleLP,yeLP';andifp = p' = 2,theseriesisevenabsolutely
convergent (see 8.2.2). It can be shown that for no value of p :f: 2 does
'i.nEzi(n)g( -n) converge absolutely for alII e LP and all yeLP'. [Notice that
(1) immediately above rules out the possibility that absolute convergence
might obtain for all Ie Ll and all y e L"'.]
[10.6] ABSOLUTELY CONVERGENT FOURIER SERIES 173
10.6 Functions with Absolutely Convergent Fourier Series

10.6.1. The Space A. As has been noted in 2.5.3, the entity associated
with the group T which is analogous and dual to the space A(Z) introduced
in 2.3.9 is the space A = A(R/2-rrZ) offunctions on T of the form ~, where tP
ranges over (l(Z). It is trivial to assert that A consists exactly of those con-
tinuous functions I on T such that
IIIIIA == 2: V(n)1
neZ
< 00; (10.6.1)
see also the characterization afforded by Exercise 9.7.

(1) It is left as a simple exercise for the reader to verify that A is a Banach
space under pointwise linear operations and the norm defined in (10.6.1);
that A is also an algebra under pointwise multiplication; and that
(10.6.2)
see Exercise 10.16. This means that A, taken with pointwise operations and
the norm (10.6.1), forms a commutative complex Banach algebra with the
constant function 1 as its identity element. To this aspect we shall return in
11.4.1 and 11.4.17.
(2) That A i.s a proper subset of C is seen from 10.3.1, or by applying 7.2.2(1)
to the sum function of the series
IX)
'" sm nx
n~2 (n log n)'
the sum function of this series is even absolutely continuous (see 12.8.3(2))
and does not belong to A. See also Exercise 10.18.
Incidentally, since the sum function I is absolutely continuous (hence of
bounded variation), it ensues from the Remark following 10.6.2(1) below that
I satisfies no Lipschitz condition of order a > O.
(3) It will appear in 12.11.3 that the problem of determining all the
continuous linear functionals on A leads to a significant class of distributions.
10.6.2. The Classical Approach. The classical approach to A (for which

the reader is referred to [Zl]' Chapter VI; [Ba2], Chapter IX; [KS], Chapitre X;
[Kah 2], especially Chapters I, II; and [I], pp. 66 if.) has in the main concen-
trated attention on seeking conditions on an individual function I which are
sufficient or necessary to ensure that lEA. Problems concerning the algebraic-
topological structure of A belong to the modern approach mentioned in
10.6.3. For both aspects [Kah 2] is the most recent and perhaps the most
incisive account.
Of the classical results, we shall handle in detail only two, contenting our-
selves with brief references to the many others.
We begin with some calculations. By (8.5.7), if f E L2,
[w2f(a)J2 = .2 4 sin
neZ
2 %na'IJ(n)12
and so
[w2f(2a)]2 = 4 .2 sin2 naIJ(n)l2.
neZ
Integrating with respect to a over [0, 71'/N], where N E {I, 2, ... },
4 .2 (J,"IN sin 2 na 00) li(n)\2 :!":; 71'N-1[02f(271'N-1)]2.

neZ 0
Now, if 1 :!":; 8 :!":; N-11nl < s + 1, then

J,o"IN sin 2 na 00 = In l- 1 10
(,,,'N-l,,
sin2 t dt
~ (s + l)-lN-l f" sin 2 tdt
= (s + I)-lN-1 S I" sin2 t dt
= (8 + 1)-lN-1 S 7T /2
~ (4N)-171'.
Hence
y~ == .2
Inl;oN
IJ(n)[2:!":; [02f(27TN-1)]2. (10.6.3)
Moreover
.2 li(n)1 = li(O)1 + ,,=.2 (IJ(n)1 + li( -n)\)

00
IIil11 ==
neZ 1
ao n
= li(O)1 + .2 2: n- 1<1J(n)1 + IJ( -n)l)
n= 1 N= 1
= li(O)\ + 2:
'"
N=ln;oN
.2 n -l(!i(n)1 + li( -n)j)
.2 .2 n- 12V'<IJ(n)1 2 + li(-n)1 2)V2
00
:!":; li(O)1 +
r
n=ln;oN
:!":; li(O)1 + ~l C&N 2n -2 r"(,~ (V(n)12 + V( _n)12)
= li(O)1 + i
N=l
2N- y.( L
\n\;oN
!1(X)1 2 )
L N-%r,v.
<Xl
= li(O)1 +2
N=l
By (10.6.3), therefore
Ilnl ~ 1i(0) I + 2 2:"" N-Y.D.d(2TTN-l). (10.6.4)

N=l
Next, supposing J to be of bounded variation and a to satisfy 0 < a ~
2TT/il, one has
2: IJ(x + lea) -
4N
J(x + (k - l)a)12 ~ D.",J(2TTN-l). V(f).
k=l
Integrating with respect to x and using translation invariance of the integral,
and so
D.d(2TTN-l)2 ~ (8TTN)-1D.",J(2TTN-l). V(f).
Hence, (10.6.4) yields
IIJlll ~ IJ(O)I + 2(8TT)-Y.V(f) 2:"" N-ID.",J(2TTN-l)y'. (10.6.5)

N=l
From (10.6.4) and (10.6.5) one can read off as corollaries a number of
results.
(I) If J is of bounded variation and
2:
"" (D.",J(2- TT))Y. <
k=l
k 00, (10.6.6)
then j E (1; in particular, if f is also continuous, then f E A. Notice that

(10.6.6) holds whenever D.",J(a) = O(lala) as a _ 0 for some a > O.
This result is due to Zygmund; see [Kah 2 ], p. 13. See also Exercise 10.17.
Proof. Apply (10.6.5), noticing that
2:'"
N=2
N-1D.""f(2TTN-1)y' = 2:
""
k=22k-l"N<2k
2: N-1D.",f(2TTN-1)y'
~ L"" 2k-1.(2k-l)-1D.""J(TT2-k+2)Y.
k=2
= L
'"
k=2
D.",J(2- k + TT)Yz 2
(2) IfJE L2 and
L'" N- YzD. 2J(2TTN-l) < 00, (10.6.7)

N=l
thenJE (1. Notice that (10.6.7) holds whenever
L"" 2k/2D.d(2 -k TT) < 00,

k=l
and that both criteria are fulfilled whenever

D.2f(a) = O(jala) as a -- 0 for some IX > Y2.
These results are due to Bernstein and Szasz; see [Ba 2 ], pp. 154-155 and
[Kah 2 ], pp. 13-14. See also Exercises 10.25 and 13.2.
Proof. This follows from (10.6.4), just as (1) followed from (10.6.5).
Remarks and further results.
The example cited in 10.6.1(2) shows that conditions like (10.6.6), bearing
upon moduli of continuity off, cannot be entirely suppressed in the hypotheses
of (1). For somewhat similar results, see Hirschman [1], Lemmas 2d and 3c
(the latter applying to the dual situation) and also Boas [1].
There is a converse to (2): it is known that there exist continuous functions
f such that D.2f(a) = O( lal Yo) as a -- 0 and which nonetheless do not belong to
A. See [Zd, pp. 240-243; [KS], p. 129; [Kah 2 ], pp. 14-45. See also Mitjagin
[1], Yadov and Goyal [1J.
There is also a similar converse to (1); see [Kah 2 ], p. 16.
It is known that no condition involving only D. oof can be at once necessary
and sufficient in order that f be equal a.e. to a member of A.
Further reading: MR 34 ## 2790,3197,4797; 35 # 7081; 39 ## 1912, 6015;
40 # 7729.
(3) Other sufficient conditions for membership of f to A involve the
numbers
cf. Section 6.5, where E~oolJ is written as ENf, and Exercise 6.10. Among such
results we note that of Steckin, which asserts that, if f E L2, then i E tl
whenever
2: N- YoE<J1 <
00
<Xl;
N=l
see [Ba 2 ], p. 155. Some of Steckin's results apply to general orthogonal

expansions and thus yield criteria in order that
2: li(n)1
00
< <Xl
neS
for preassigned subsets S of Z. See also Yadov [1] and Zuk [1J.
(4) It follows easily from 8.2.1 and 8.3.1 that f E A if and only if it can be
expressed in the form f = u * v with u, v E L2. This criterion was first noted
by M. Riesz; unfortunately, it is difficult to apply in specific cases which are
not already decidable in more evident ways.
Somewhat similar results have been given by M. and S.-I. Izumi [4] and
others.
[10.6] ABSOLUTELY CONVERGE~T FOURIER SERIES 177
(5) Another necessary and sufficient condition, discovered by Steckin in

1951, asserts that, if! E L2, then! E {l if and only if
2: N-Y.,.E ! <
<Xl
N 00,
N=l
where
the infimum being taken as S ranges over all (not necessarily harmonic)
trigonometric polynomials of the form
2: cn,exp (iAnx) ,
N
S(x) =
n=l
where the Cn are complex numbers and the An are distinct real numbers; see
[Ba2]' p. 186; [Kah 2], p. 10. Compare this with the sufficient condition given
in (3), noting that evidently
.E2N +2! ~ .E2N +d ~ E<J>J.

(6) Wiener showed that a function on T, which agrees ona neighborhood
of each point x with some (possibly x dependent) element of A, itself belongs
to A: in brief, a function which belongs locally to A belongs globally to A.
We shall not give the original proof of this (for which see [Zd, p. 245; [Ba 2],
p. 188; (Kah 2 ], p. 11) but rather a proof based upon Banach algebra theory;
see Exercise 11.19.
A simple sufficient condition for local membership of A maybe derived
from Exercise 13.3.
(7) If one introduces prematurely the conjugate function J,defined in
Section 12.8, one can state the remarkable result of Hardy and Littlewood
asserting that, if both! and J are of bounded variation, then
L
neZ
Jj(n)1 < 00.
Fora proof, see (Zl], pp. 242 and 287. Compare this with F. and M. Riesz'
theorem mentioned in 12.8.5(4). See also Exercise 12.19.
(8) Let E denote a closed subset of T. Almost every question so far posed
in relation to A can also be posed in relation to A(E), the set of restrictions
to E of elements of A. For such variants we must refer the reader to [KS],
Chapitre X, especially pp. 130 ff and [Kah 2], p. 19.
It is of course evident that in all cases A(E) is a subset of C(E), the space
of all continuous complex. valued functions on E. Much less evident is the
fact that there are infinite closed sets E such that A(E) exhausts C(E): such
sets E are termed" Helson sets," concerning which a little more will be said
in Section 15.7; see especially Subsection 15.7.3. See also [Kah 2 ], Chapitres
III, IV, IX; MR 40 ## 630,7731; 43 # 6660; 55 ## 966,970.
10.6.3. The Modern Approach. (1) In more recent times emphasis has
been placed on structural properties of A as a whole. Thus, one of the problems
on which attention has been centered is the following: under what conditions
upon the function F, defined on some subset D of the complex plane, is it
true that Fo/EAwhenever/EAand/(R) =/(T) c D?
As a matter of fact, some aspects of this problem were first considered by
Levy as long ago as 1934; and even prior t.o this Wiener had obtained a
special case of Levy's result. Wiener showed that 1- 1 E A whenever I E A
and I is nonvanishing, while Levy showed that a function F has the property
mentioned in the preceding paragraph whenever D is open and F is analytic
at each point of D. The modern approach to these problems differs from the
original (for which see [Zl], pp. 245-247; [Ba 2 ], pp. 186-194; [Kah 2], pp. 57,
58) in the methods utilized, namely, the theory of Banach algebras. We shall
deal with these theorems by use of this technique in 11.4.17; the dual results
appear in 11.4.13 and 11.4.16.
In 1958 Katznelson discussed the necessity of Levy'S sufficient condition,
and variants and analogues of t.he Levy-Katznelson results have been
examined for cases in which the underlying group T is replaced by a more
general group. Katznelson, Helson, Rudin, and Kahane have all contributed
to these problems; see Herz [1], Rudin [2], [5], [R], Chapter 6; [Kah], Capitulos
IV to VI; [Kah 2], Chapitre VI; [Kz], Chapter VIII. Here we mention merely
the versions ofthe Levy-Katznelson results appropriate to the groups T and
Z, namely:
(a) If F is defined on [-1, 1] and F 0 I E A whenever I E A and I(T)
c [-1,1], then F is analytic on [-1,1].
(b) If F is defined on [ -1, 1] and F 0 4> E A(Z) whenever 4> E A(Z) and
4>(Z) c [ -1, 1], then F is analytic at 0 and F(O) = O.
Of these statements, (a) is due to Kat.znelson and (b) to Helson and Kahane
jointly. The assertion that F is analytic on [-1, 1] (respectively at 0)
signifies that F is extendible into a function analytic on some open subset of
the complex plane containing [-1, 1] (respectively 0).
From (a) it follows as a corollary that there exist functions I E A such that
III does not belong to A; the analogous assertion with A{Z) in place of A
follows likewise from (b). (Both of these corollaries were established a little
earlier by Kahane [2], [3].) On the other hand, Beurling has shown that if
I E A is such that 1/( n)1 ~ Cn (n = 0,1,2" .. ), where Cn .j. 0 anrl
"2.:'=0 Cn < 00, then Ifl E A; see [I], p. 78.
(2) There are analogous problems in which A(Z) is replaced by AP(Z)
== {/ : IE LP}. The case in which 2 ~ p ~ 00 is solved by Rider 12], who
shows (among other things) that a complex-valued function F on the complex
plane has the property that F 0 4> E AP(Z) whenever 4> E AP(Z) if and only if
it has the form
F(z) = az + bE + IzI2!P'c(z) ,
where a and b are complex numbers and the function c is bounded on a

neighborhood of the origin. (For the "if" assertion, see Exercise 13.24.) See
also Rudin [4].
The same problem has been studied for yet other important algebras; see
the o;\nd of 11.4.17 below.
(3) This is a convenient place to comment further on the dual form of the
problem stated in 4.2.5, namely: which maps 41) of T into itself have the
property that f 0 41) E A whenever f E A? By dualizing the substance of Chapter
4, it may be seen that these maps 41) correspond to the homomorphisms of the
convolution algebra tl(Z). Insofar as this can be and has been subsumed under
the study of the L1 homomorphism problem for general groups, the appropriate
reference is again [R]. Chapter 4. For the particular group we have in mind, an
independent and more direct solution is due in part to Leibenson and in part
to Kahane (1954-56); an account of Kahane's approach will be found in
Capitulo III of [Kah]; see also [Kahll], p. 86 and Chapitre IX. A mapping 41)
having the stated property may be said to define a permi88ibk change of
variable (relative to A), and the Leibenson-Kahane result asserts that the
permissible changes of variable are precisely those defined by maps 41) having
the form 41): i -.. (nz + a)', where n E Z and a E R.
In Kahane's approach to this problem one first thinks of 41), which must
obviously be continuous, as a (periodic) map from R into T. For each x E H,
41)(x) E T and el~") is uniquely defined. In x -+ el~") one has a continuous map
of H into the multiplicative group of complex numbers having unit absolute
value. A simple argument (using local branches of the logarithm) shows that
there exists a continuous real-valued function", on H such that 81~") el~"), =
80 that 41)(x) = ("'(x))', the coset modulo T containing the real number "'(x).
Since 41) is periodic, '" must have the property that
identically in x, n being some integer. The crucial point of this transformation

is that", is a complex- (actually real-) valued function on H, which 41) is not.
However, '" is not necessarily periodic and we make one further change of
focus to take care of this. namely, we look at the periodic real-valued function
"'0 defined by
By using the fact that f 0 '" =f 0 41) E A whenever j E A, it is relatively easy

to deduce that
(c) SIIPkEZ!le1k'-o\\A = SIIPkEZ!!e1k'-\\A < (Xl
and t.hat
(Ii) "'0 E A;
for (c), sec j.~x(rcisc 10.19; thc proof of (d) dcpends upon and is an easy corollary
of the reRlllt. stat.('<i in 10.fI.2(6).
Thc crllx of Kahanc's ar~lIment. is the difficult deduction from (c) and (d)
t.hat "'0is a constant.
(4) j.'or a stllliy of isomorphisms and homomorphisms of the algebras A(E)
defined in Subsection 10.6.2(8), see de Leeuw and Katznelson [1], McGehee [1],
and [Kah 2 ], Chapitre IX. (In the terminology introduced in Chapter 11,
A(E) is isomorphic to the quotient algebra of f1(Z) modulo the ideal IB
composed of those .p E (l(Z) such that $ vanishes on E.)
For further reading, consult MR 34 ## 3226, 3365, 3366, 4793, 4796; 35
## 670, 3373; 36 ## 1924,6880; 37 ## 6672, 6690; 41 # 744; 49 ## 5723, 7700;
50 ## 895, 5338; 51 ## 1289, 6627; 52 ## 8777, 8798, 8805, 14816; 53 ## 8791,
8792; 54 ## 856, 858, 8160, 8163.
EXERCISES
10.1. Prove the statement in 10.1.2, namely
sup {lIb t -1 sin t dt I:

a E R, b E R} < 00.
10.2. Let (fl,v)'; = 1 be a sequence of positive numbers converging to zero.

Show that there exists an 1 E Ll such that
Verify that nonetheless

Ils,vll1l = o(log N) as N ---;. 00
for any 1 ELl.

Hints: For the first part, adapt the type of proof used in 10.3.2. For the
second part, use the FubiniTonelli theorem to show that
and use the fact that I T -J - fill ---;. 0 with y.

10.3. By examining the proof of 10.4.1, show that the relation (10.4.1)
holds uniformly with respect to x when f is continuous.
10.4. Show that if f is continuous, then the series
L j(n)e tnx
neZ log (2 + InJ)
is uniformly convergent.
Hints: Use the preceding exercise and examine the proof of 10.4.3.
Alternatively, put S(f) = Lnezj(n)/log (2 + In!> and S,v(f) = L'n, . . ,vj(n)/
log (2 + In!). Verify that SN (f) ---;. S(f) for each fEe and use the uniform
EXERCISES 181
boundedness principle (Appendix B.2.1(2)) to deduce that the SN are equi-

continuous on C. Conclude that SN(J) ~ S(J) uniformly whenfranges over a
relatively compact subset of C, and hence that SN(T xf) ~ S(Txf) uniformly
in x for a givenfE C.
10.5. Prove that Lnezj(n)g(n)/iog (2 + In!) is convergent whenever
f E C and g ELI. Deduce that there exists a number c such that
" g(e)en
II L
Inl"N
log (2 + In!) III ~ c IIgl11
for all N and all g E Ll (c being independent of N and g), and that the series
Lnezg(n)en/iog (2 + In!) is convergent in mean in Ll for each g E L1.
Conclude finally that the series Lnezj(n)g(n)/Iog (2 + In!) is convergent
whenever fE Loo and g ELI.
Hints: For the first statement use the uniform boundedness principle
(Appendix B.2.1(2)) and Exercise 3.6. Deduce from this that the set of g E L1,
for which Lnezg(n)en/iog (2 + In!) is convergent in mean in L1, is closed in L1.
10.6. Prove that if (An);'=o is a sequence such that L;'=o I~Anl < 00, then
the series LnezAlnlg(n)en/log (2 + In!) is convergent in mean in L1 whenever
gEV.
Hints: Use the preceding exercise, together with an adaptation of
Exercise 7.l.
10.7. Show that the series Lnez j(n)g( -n) diverges for suitably chosen
f E L1 and all g E C. (Compare with 10.5.5(1).)
Hint: Argue by contradiction, using the uniform boundedness principle
and Exercise 10.2.
10.8. Let j be the periodic function such that j(x) = Y2(17' - x) for
o ~ x < 27T. Show that the Fourier series ofj is L;'= 1 sin nxjn. Verify that for
o ~ x ~ 17',
and deduce that
(1) lim {SNj(X)

N-oo
+ Y2x - l IN
0
+ Y.'x sin t dt
--} =
t
ouniformly forO ~ x ~ 17';
(2) lim SNj (x) = j (x) for 0 < x < 17';

N-oo

(3) I1m . (a)
SNJ N = la sintdt
--
t
for a > 0;
1" sintdt
N-oo ~ 0
sup
(4) I1m SNJ
. ()
x ~ --
N-oo,x-+O 0 t
>Y217'=j(+O).
Notes: Conclusion (4) shows that the sequence (SNj) offunctions exhibits
the so-called Gibbs phenomenon on right-hand neighborhoods of zero. A
similar situation prevails on left-hand neighborhoods of zero. This exhibition

is typical of the sequence (SN!) whenever f is of bounded variation and has
jump discontinuities; see [Zl]' pp. 61-62. The Gibbs phenomenon at a point
Xo of the sequence (SN!) is a feature of the nonuniformity of the convergence
on neighborhoods of a point of discontinuity.
For many more details, plus a most interesting survey of the history of the
Gibbs phenomenon, see Hewitt and Hewitt [1].
10.9. Discuss the following suggested procedure, and in particular frame
hypotheses sufficient to justify the steps: take a function F on R and form
the periodic function
f(x) = L F(x + 217k).
lceZ
Then
.
f(n) 1
= -217 f'"_ a> F(y)e- lnll dy (n E Z),
and so, sincef(O) = 2.nezi(n}, we obtain Poisson's summation formula:
L F(2kl7)" = 2"117 L fa>a> F(y)e-lnlldy.

lceZ neZ -
See also [Kz], p. 129; MR 36 # 4265; 54 # 5734.

10.10. Justify the use of Poisson's summation formula (Exercise 10.9)
in case F( y) = (a 2 + y2) -1 (a real and nonzero) and so deduce that
for such values of a. Conclude that

a>
L k-
2
2 =~.
1c=1 6
10.11. Justify the use of Pois,'!on's summation formula (Exercise 10.9) in

case F( y).= exp (_a 2y 2) (a real and nonzero), and so deduce that
for real s > O.

Note: This is a famous transformation formula for one of the so-called
theta functions; see [Be], p. 11.
10.12. Denote by CBV the linear space of (periodic) functions that are
continuous and of bounded variation. Show that CBV is a Banach space
when endowed with the norm
II! I IlfII", + V(f).

EXERCISES 183
Using the uniform boundedness principle (Appendix B.2.1) and the result
(included in 10.1.4) that SUPN ISNI(O)I < 00 for IE CBV, deduce inequality
(10.1.5), that is, the existence of a number m ~ 0, independent of I, such
that
for IE CBV.
Remark. I am grateful to Professor G. Goes for the remark that, by
7.2.2(2), the series h(x) = 2 2;:'= 1 n -1 sin nx is boundedly convergent. Iff is of
bounded variation, integration by parts yields
and so
10.13. Prove that limN~",sNI = I uniformly for each IE CBV.

Hint: Reexamine the proof of 10.1.1.
Alternatively, use Exercises 5.5 and 8.13 and Theorem 6.1.1. See also
MR 50 #lO657.
10.14. Is it true to assert that limN~",sNI = j, in the sense of the norm on
CBV defined in Exercise 10.12. whenever I E CBV?
Is it true to assert that the trigonometric polynomials are everywhere
dense in the space CBV (relative to the norm defined in Exercise 10.12)?
Full justification is required for your answers. You may assume and use
the fact that there exist continuous functions of bounded variation which are
not absolutely continuous; see Remark (2) following 2.3.6.
10.15. Prove that the sequence (en)nEz is not a topological base for any
one of the spaces Ll, L"', and C.
Hint: Assuming the contrary, determine the form of the coefficients an in
the associated expansion (lO.3.12).
10.16. Verify in detail the statements made in 10.6.1(1) concerning A.
10.17. Prove that, if I is of bounded variation and
O",I(a) = O((log lal-1)-P)
as a -'>' 0 for some f3 < 2, then j E (1.

Hint: Use (10.6.5).
10.18. (1) Suppose that LnEZlcnl < 00 and that nC n f= 0(1) as Inl-'>'oo.
Show that I(x) = LnEZCneinx belongs to A and is not absolutely continuous.
(2) Use Exercise 3.14 to prove that there exist absolutely continuous
functions which do not belong to A (compare 10.6.1(2) and Exercise 10.25).
10.19. Suppose </> is as in 10.6.3(3), namely, a real-valued function on R
such that
</>(x + 211) - </>(x) = 2M
for some integer n and all x E R, and having the property that f 0 .p E A
whenever f E A. Prove that
sup Ilell<<bII A < 00.
keZ
Hint: Consider the linear mapping T of A into itself defined by Tf = f 0 .p.

Show that T has a closed graph and apply Appendix B.3.3.
10.20. (S. Saks) Let E be a Banach space and (SN)~=l a sequence of
continuous linear operators from E into V', where 0 a
such that lim sUPN~",ISNf(x)i < <X) for almost all x E A/.
Prove the following two statements:
(I) if 0 < e < a, there exists a set A .such that m(A) > a - e and
lim sup N~ '" IS Nf (x)1 < <X) for almost all x E A for each fEE;
(2) there exists a set A and a meager subset Eo of E such that
lim sup ISNf(x)1 < <X) for almost all x E A
N~'"
whenever fEE, and
lim sup lSNf(x)1 = 00 for almost all x f/:A

N-<Xl
whenever f E E\E o
Hint8: See [KSt], p. 25 or [GP], p. 153.
Remark. The results are true if E is a Frechet space; and Lp can be
replaced by more general spaces of functions.
10.21. Assuming Kolmogorov's theorem (asserting the existence of at
least one integrable function whose Fourier series diverges almost every-
where), deduce that the set of functions fELl, whose Fourier srrirs diverge
almost everywhrre, is a comeagrr subset of V (that is, is a subset of Ll whose
complement is meager and which is therefore itself nonmeager).
Hint: Use the preceding exercise ..
10.22. Employ the uniform boundedness principle (see Appendix B.2.1)
to prove the following statement: if 4> is a nonnegative function on Z such
that the set N = {n E Z : 4>(n) > O} is infinite and
lim inf 4>(n) = 0,

neN.lnl- GO
then there exist functions fELl such that
lim sup 1/(n)1 = 00.

neN.lnl- <Xl 4>(n)
(See the opening remarks in Section 10.3.)

EXERCISES 185
10.23. Let U be a linear operator from T onto TN such that UI = I for

I E TN (so that U is a linear projection of T onto TN)' Prove that
217T J
UTal(x + a) da = sNI(x)
for IE T.
Defining
and similarly for IlsNIIp,p, deduce that
IIUlll.l ~ IlsNI11.1' IIUII",.", ~ IlsNII",.",

10.24. Let E denote C or L\ and let (N ,,)';= 1 be a sequence of positive
integers such that sup" N" = 00. Suppose that U" is a linear projection from
E onto TNk which is continuous from E into E. Prove that there exists an
lEE such that
lim sup IJU"I liE = 00. (I)
k-'"
Remarks. In case E = C, it is not known whether in all cases lEe and
x E T exist such that lim sup"_,,, IU"I{x)J = 00, but this is easily deduced
from (1) if each U" commutes with translations. In case E = L1, it is likewise
in doubt whether IE L1 exists such that lim suP,,_oo IU d(x)1 = <X) for almost
all x (or even for those x in some set of positive measure). See 10.3.4 for the
case in which N" = k and Uk = SIC'
Hints: Use the preceding exercise and (5.1.10) in combination with the
uniform boundedness principle (Appendix B.2.2).
10.25. Suppose that I is absolutely continuous and that DI E L2. Prove
that I E A and that
II/IIA ~ Ij(O)1 + B!IDlk
where B is an absolute constant.
Remark. Stronger results appear in Exercises 13.2 and 13.3; compare
also Exercise 10.18.
Hint: Examine the proof of 8.5.4 and use a similar method.
10.26. Suppose that 0 < e < 7Tj2 and let Ve be the continuous nonnegative
even periodic function that vanishes outside [- e, e] (mod 27T), is linear on
[0, e], and satisfies Ilv el11 = Ij21T. Verify that v. E A and IIveilA = e- 1.
Define Ue = 4ev2e - eVe and verify that Ue = 1 on [- e, e], that U.
vanishes outside [-2e, 2e] (mod 27T), and that IluellA ~ 3.
Suppose thatlE A satisfies 1(0) = O. Show that lime_ollu./IIA = O.
Remark. This construction will find a use in connection with spectral
synthesis sets; see Exercise 12.52 and compare [R], Theorem 2.6.4.
Hints: For the second part, show first that it is enough to deal with the
case in which lET. In this case, estimate the V-norm of usl and the
L<Xl- norm of D{uef), and then use the preceding exercise.
APPENDIX A
Metric Spaces and Baire's Theorem
It is assumed that the reader is familiar with the definition and simple
properties of a metric space and its metric topology. The aim of this appendix
is to introduce the concepts of meager and nonmeager sets, to prove Baire's
"category theorem," and to give some corollaries thereof (some of which
form the basis of results given in Appendix B). See also [K], pp. 200-203, or
[HS], p. 68.
A.l Some Definitions

Let E be a metric space (or any topological space). A subset A of E is said
to be:
(1) nowhere dense (or nondense) if the closure A contains no interior points;
(2) everywhere dense if A = E;
(3) meager (or of jir8t category) it is expressible as the union of countably
many nowhere dense sets;
(4) nonmeager (or of 8econd category) if it is not meager;
(5) comeager if its complement is meager.
The reader will notice that a set A is nowhere dense if and only if the open
set E\A is everywhere dense.
A.2 Baire's Category Theorem

If E is a complete metric space, then
(1) the intersection U = n:=l
Un of a sequence of everywhere dense
open subsets Un of E is everywhere dense;
(2) a meager subset ofE has no interior points (or, equivalently, a comeager
subset is everywhere dense).
Proof. (1) Let us first show that U is nonvoid. For any x E E and any
number e > 0 we write
B(x, e) = {y E E : d( y, x) < e}
and
B(x, e) = {y E E : d( y, x) ~ e}
187
188 METRIC SPACES AND BAIRE'S THEOREM
where d denotes the metric on E. (To allay possible confusion resulting from
the notation, we point out that E{x, e), although closed in the topology
defined by d, is in general not the closure of B(x, e).)
We may assume that E is nonvoid. The same is therefore true of U 1 (since
U l is everywhere dense). Choose freely Xl E U l and then el > 0 so that
Since U 2 is everywhere dense, U2 meets B(X1' ell Choose X2 E U2 n B(Xl> el)

and then e2 > 0 so that
e:a < Yz,

which is possible since B{Xl' el) n U2 is open. Proceeding thus, we obtain
numbers en satisfying 0 < en < lin and points Xn E E such that
(A.2.1)
If n > m we have Xn E B(xn' en) c B{xm' em), by (A.2.I), so that d(xn' xm)
 m ~ 1, so
X E ..8(xm' em) for all m, and (A.2.1) shows then that x E U. Thus U is nonvoid.
Take now any closed ball E = B(xo, S) in E. It is easy to verify that
Un n E is everywhere dense in the complete metric space B (a subspace of E).
So, by what we have proved, n:=1 (B nUn) isnonvoid, that is, n:=lU n = U
meets E. This being so for any B, U is everywhere dense in E. This proves (1).
(2) Let M be any meager subset of E. Then we can write M = U.7'= lAn,
where An is nowhere dense. So Un = E\An is everywhere dense and open.
By (1), n:=lU n is everywhere dense. That is, E\U:=lA n is everywhere
dense, so that U:= IAn has no interior points. The same is therefore true of
Me U::lAn.
A.3 Corollary
If E is a complete metric space, it is nonmeager in itself.
A.4 Lower Semicontinuous Functions

Let E be a metric space (or a topological space). A function f on E with
values in (-00, 00] is said to be lower semicontinuous if and only if for each
real number ex the set {x E E : f(x) > ex} is an open subset of E.
The reader will verify that the upper envelope of an arbitrary family of
lower semicontinuous functions is again lower semicontinuous.
[A.5] A LEMMA 189
A.5 A Lemma
Let E be a metric space (or a topological space) and (f1)leI an arbitrary
family of lower semi continuous functions on E. If
sup fl(x) < 00
leI
holds for each x in some nonmeager subset 8 of E, then there exists a number
m < 00 and a nonvoid open subset U of E such that
SUpft(X) ~ m (x E U).
leI
Proof. For each natural number m let

8 m = {x E E : f(x) ~ m},
where f is the upper envelope of the k Since f is lower semicontinuous, 8 m is
closed in E. Also, 8 = U~=18m. Since 8 is nonmeager, 8 m must fail to be
nowhere dense for some m. For this m, 8 m contains a nonvoid open set U.
APPENDIX B
Concerning Topological Linear Spaces
B.I Preliminary Definitions

B.I.1. All linear spaces involved are over the real or complex field of
scalars, the scalar field being denoted by <1>.
By a topological linear space we mean a linear space E, together with a
designated topology on E relative to which the functions (x, y) ~ x + y and
(A, x) ~ Ax are continuous from E x E into E and from <1> x E into E,
respectively. See [E], Chapter I.
The specific results about topological linear spaces needed in the main text
refer solely to a particular type of such space, namely, those classified as
Frechet spaces (see [E], Chapter 6). Our definition of these will be made in
terms of seminorms.
B.1.2. By a seminorm (or prenorm) on a linear space E is meant a function p

from E into [0, co) having the following properties:
p(x + y) ~ p(x) + p( y), p(Ax) = IAlp(x)
for x, y E E and A E <1>.
A norm is a seminorm p for which p(x) > 0 whenever x # O. Norms will
usually be denoted by 11'11.
B.1.3. Frechet Spaces. A Frechet space is a topological linear space E

satisfying the following conditions:
(a) There is a finite or denumerably infinite family (Pk) of seminorms on E
which define the topology of E in the sense that the sets
{x E E : Pk{X) < e for all k E J} , (B.Ll)
obtained when e ranges over all positive numbers and J over all finite sets of
indices k, constitute a base (or fundamental system) of neighborhoods of 0 for
the topology of E.
(b) The topology of E is Hausdorff, that is (what is easily seen to be
equivalent), x = 0 is the only element of E for which Pk{X) = 0 for all
indices k.
191
192 CONCERNING TOPOLOGICAL LINEAR SPACES
(c) E is complete in the sense that to any sequence (x,,):'= 1 of points of E

for which
lim Pk(Xm - x,,) = 0
min_IX)
for each index k, corresponds an x E E to which the sequence (x,,) converges,

that is, for which
for each index k.

We then speak of (Pk) as a defining family of seminorms for the Frechet
space E.
B.l.4. Remarks. (1) For a given Frechet space there are many different
defining families of seminorms.
(2) Given a linear space E and a countable family (Pk) of seminorms on E
such that (b) and (c) are fulfilled, we can topologize E in just one way as a
Frechet space for which (Pk) is a defining family of seminorms. Namely, we
agree (as a matter of definition) that the sets (B.l.I) shall constitute a base of
neighborhoods of 0; and that, for any Xo E E, the images of these sets under
the translation x -+ x + Xo shall constitute a base of neighborhoods of Xo.
The properties of seminorms ensure that in this way one does indeed obtain a
topological linear space which satisfies conditions (a) to (c).
(3) If E is a Frechet space, it can be made into a complete metric space
whose topology is identical with the initial topology on E, and this in several
ways. One way is to define the metric
d(x, y) = "" k- 2 Pk(X - Y)

f 1 + Pk(X - Y)
(4) In a Frechet space one can always choose the defining family (Pk) so
that the index set is the set of positive integers and so that Pl ~ P2 ~ ....
(By repeating seminorms we may suppose that the original set of indices is
the set of natural numbers, define new seminorms q" = sup {Pk : 1 ~ k ~ h},
and take the (q,.) as the desired defining family.] If this be done, a neighbor-
hood base at 0 is comprised of the sets {x E E : Pk(X) < E} when k and E > 0
vary; a neighborhood base at 0 is also obtained if E is restricted to range over
any sequence of positive numbers tending to 0; or again, if the strict
inequalities Pk(X) < E are replaced throughout by Pk(X) ~ E.
(5) The product of two Frechet spaces, or a closed linear subspace of a
Frechet space, is a Frechet space. If (Pk) and (qh) are defining families for
Frechet spaces E and F, respectively, the semi norms rkh(x, y) = Pk(X) + q,,( y)
constitute a defining family for E x F.
B.l.5. Banach Spaces. A Banach space is a Fnlchet space possessing a

defining family comprising just one element, which must of necessity be a
[B.l] PRELIMINARY DEFINITIONS 193
norm. Expressed in another way, a Banach space is a normed linear space

that is complete for it,s norm.
The reader will note that the scalar field <1> is itself a Banach space, the
norm being equal to the absolute value.
B.1.6. Bounded Sets. Let E be a topological linear space and A a subset

of E. A is said to be bounded if and only if to each neighborhood U of 0
in E corresponds a positive scalar A such that A c AU (the set of multiples
by A of all elements of U). Except when E is a Banach space, this concept of
boundedness is different from metric boundedness (which is equivalent to
finiteness of the diameter).
If E is a Frechet space and (Pk) a defining family of seminorms for E, the
set AcE is bounded if and only if sup {Pk(X) : x E A} < 00 for each index k;
if E is a Banach space, the condition is merely that sup {llxll : x E A} < 00.
B.1.7. The Dual Space. If E is a topological linear space we denote by

E' the linear space of all continuous linear functionals on E (the algebraic
operations in E' being "pointwise"). E' is termed the (topological) dual of E;
some authors use the terms" adjoint" or "conjugate" where we use the term
"dual."
A sequence Un);: = 1 of elements of E' is said to converge weakly in E' to
fEE' if and only if limn~.,fn(x) = f(x) for each x E E.
If E is a Banach space, E' is also a Banach space for the so-called dual norm
Ilf I = sup {If(x)1 : x E E, Ilxll ~ I};
the proof of completeness of E' is exactly like that in the special case dealt
with in 12.7.1.
B.1.8. Quotient Spaces and Quotient Norms. Let E be a linear space

and L a linear subspace of E. The quotient space Ej L, whose elements are
cosets x + L, is defined in purely algebraic terms and is turned into a linear
space by defining
(x + L) + (y + L) = (x + y) + L,
A(x + L) = (Ax) + L
for x, y E E and AE <1>. Denote by 4> the quotient map x ~ x + L of E onto

Ej L; 4> is linear.
If E is a topological linear space, one can make Ej L into a topological
linear space (again spoken of as the quotient space) by taking a base of
neighborhoods of zero in Ej L to be formed on the sets 4>( U), where U ranges
over a base of neighborhoods of zero in E. The quotient space Ej L is Hausdorff
if and only if L is closed in E.
A special case of importance is that in which E is normed, in which case the

aforesaid quotient topology of EjL is derivable from the quotient seminorm
Ilx + LII = inf IIx

lIeL
+ yll
If L is closed in E, this quotient seminorm is actually a norm on Ej L and is
termed the quotient norm on EjL.
If E is a Banach space and L is closed in E, Ej L is a Banach space (for its
quotient norm); compare Subsection 11.4.7.
For more details see, for example, [E], Sections 1.8.5 and 1.10.5.
B.2 Uniform Boundedness Principles

We require two such principles stemming from Baire's theorem (see
Section A.2) as a common source. The first concerns seminorms and linear
functionals, and the second refers to linear operators.
B.2.1. (I) Let E be a Frechet space and Pk (k = 1,2, .. ) a lower semi-

continuous function on E with values in [0, 00] such that
for x, y E E and A E (these conditions on Pk being fulfilled whenever Pk is a

lower semi continuous seminorm on E). If infk Pk{X) < 00 for each x in a
nonmeager subset S of E (in particular, if infk Pk{X) < 00 for each x E E), then
there exists an index k such that Pk is finite valued and continuous on E.
{2} Let E be a Frechet space and (f)leI an arbitrary family of continuous
linear functionals on E. If
sup ~1/1{x)l : i E J} < 00 (B.2.1)

for each x in a nonmeager subset S of E (in particular, if (B.2.I) holds for
each x E E), then the II are equicontinuous on E, that is, to each e >
corresponds a neighborhood U ofO in E such that x E U entails supd/l{x)l < e.
IfE is a Banach space, the conclusion reads simply suplerll/,1I < 00.
Proof. (I) In interpreting the hypotheses on Pk we agree that a + 00
= 00 + 00 = 00 for any real a ~ 0, that 0'00 = 0, and that a'OO = 00 for
any real a > 0. For k, r = 1,2" . " let Sk.r denote the set of x E E such that
Pk{X) ~ r. Since Pk is lower semicontinuous, each Sk.r is closed in E. Plainly,
S c U:'r=ISk,r so that, since S is nonmeager by hypothesis (see A.3 for the
particular case in which S = E), some Sk.r has interior points. Thus there exists

Xo E E and a neighborhood U of in E such that Pk{X) ~ r for x E Xo + U.
Then the identity x = %(xo + x) - %(xo - x) combines with the prop-
erties. of Pk to show that Pk{X) ~ r for x E U. Consequently, given e > 0,

we have Pk{X) ~ e provided that x E U1 == {r- 1 e)U. Now U 1 is again a
neighborhood of in E, because the function x -* (re -1)X is continuous from
[B.3] OPEN MAPPING AND CLOSED GRAPH THEORE;\IS 195
E into E (see the axioms of a topological linear space in B.l.l). This shows
that PIc is finite and continuous on E and so completes the proof of (1).
(2) This follows immediately from (1) on defining PIc = p for all k, where
pix) = sup {ift(x)i : iEI}.

An important corollary of B.2.l is the following statement about families
of continuous linear operators.
B.2.2. Banach-Steinhaus Theorem. Let E and F be Frechet spaces

and (Tt)tel an arbitrary family of continuous linear operators from E into F.
Suppose that, for each x in a nonmeager subset S of E (for example, for each
x E E), the set {Ttx : i E I} is a bounded subset of F. Then the T t are
equicontinuous on E, that is, to each neighborhood V of 0 in F corresponds a
neighborhood U of 0 in E such that TtU c V for all i E 1.
Proof. We may assume that the given neighborhood V is of the form
V = {y E F : q( y) ~ e}, where q is some continuous seminorm on F (being,
for example, a member of some defining family for F; see B.l.4(4)). Now
apply B.2.l(l) to the situation in which
PIc = sup
tel
iqi 0 Tt (k=I,2,.).
Remark. In the proof of B.2.2 no essential use is made of the fact that
F is a Frechet space: all that is necessary is that F be a topological linear space
whose topology can be defined in terms of continuous seminorms. In other
words, B.2.2 extends to the case in which F belongs to the category of so-
called "locally convex" topological linear spaces. For further developments,
see [Jj:], Chapter 7.
B.3 Open Mapping and Closed Graph Theorems

B.3.l. Some Definitions. Let E and F be Frechet spaces and T a linear
operator from E into F. The graph of T is the subset of E x F comprising
those pairs (x, y) in which y = Tx. The operator T is said to be closed (or to
have a closed graph) if and only if its graph is a closed subset of the product
space E x F. This signifies that, if lim n _ <XlXn = 0 in E and lim n _ <Xl TX n = y in
F, then necessarily y = O.
Evidently, if T is continuous, then it is closed. One of the two theorems
we aim to prove asserts the converse of this. This will be deduced from
another rather surprising result involving the concept of an open mapping.
If, as before, T is a linear operator from E into F, we shall say that T is
open if TU is an open subset ofF whenever U is an open subset ofE. Because
of linearity, this signifies exactly that T U is a neighborhood of 0 in F when-
ever U is a neighborhood of 0 in E.
B.3.2. Open Mapping Theorem. Let E and F be Frechet spaces and T

a closed linear operator from E into F such that TE is nonmeagcr in F. Then
T is open; in particular, TE = F. (Notice that the hypotheses are fulfilled
whenever T is continuous and TE = F; see Section A.3.)
Proof. It has to be shown that, U being any neighborhood of 0 in E. TU
is a neighborhood of 0 in F. Now [see B.1.4(4)] we can choose an increasing
defining family (Pk);'= 1 for E so that U contains the set {x E E : Pl(X) ~ I}.
Let (q")h=l be a defining family for F, also increasing. We will show that there
exists an index h1 and numbers r > 0 and 1 > 0 such that each Y E F
satisfying q"l (y) ~ 1 is expressible as y = Tx for some x E E satisfying
Pl(X) ~ r. This is clearly enough to establish the required result.
To do this, write Uk = {x E E : Pk(X) ~ k- 2 } for k = 1,2, . '. Since
E = U;'= l(nU k)' TE c U;'= IT(nU k)' Then. TE being nonmeager by hy-
pothesis, an n exists such that the closure of T(nU k) = n' TU k contains
interior points. This entails, as is easily verified, that T Uk is a neighborhood
of 0 in F. Thus ~here exist an index hk and a number k > 0 such that T Uk
contains the set Vk = {y E F : q"k (y) ~ k}' One may assume without loss of
generality that hk ;;: k and that lim k _ "'k = O.
Suppose now that y E V 1 Since TU 1 :::> V 1 , Xl E U 1 may be chosen so that
q"2(Y - Tx 1) ~ 2' Then y - TX1 E V 2 and, accordingly, since 1'U 2 :::> V 2 ,
x 2 E U 2 may be chosen so that Q"3'( y - TX1 - Tx 2 ) ~ Ea. Proceeding in this
way, we obtain points Xn of E so that
The completeness of E entails that
x = lim (Xl + X 2 + ... + xn)

n-'"
exists in E, since the first relation in (B.3'!) ensures that 2n Pk(X n ) < 00 and
therefore that the sequence (Xl + ... + x n );'= 1 is a Cauchy sequence in E.
The second inequality in (B.3.I) shows that for n ;;: h
so that TX1 + ... + TX n = T(X1 + ... + xn) __ yin F. Since T is closed, it

follows that y = Tx. Finally,
L P1(Xk)
n
P1(X) = lim P1(X 1 + ... + xn) ~ lim inf
n-CJJ n-CXl k=l
~ lim inf
n_co
L Pk(Xk) ~ L k-
n
k=l
'"
k=l
2 = r,
say, by (B.3.I) again. The proof is thus complete.

[BA] THE WEAK COMPACITY PRINCIPLE 197
B.3.3. Closed Graph Theorem. Let E and F be Frechet spaces. Any

closed linear operator from E into F is continuous.
Proof. Let G be the graph of T. Then G is a closed linear subspace of
E x F and is therefore (see B.1.4(5)) itself a Frechet space. Define the linear
operator S from G into E by S(x, Tx) = x for x E E. It is clear that S is
continuous and SG = E. So, by B.3.2, S is open. This entails that T is
continuous from E into F: indeed, if V is any neighborhood of 0 in F, the set
W of pairs (x, Tx) for which x E E and Tx E V, is a neighborhood of 0 in G,
so that SW must be a neighborhood V of 0 in E; but x E V entails that
Tx E V, so that TV c V and T is thereby seen to be continuous.
B.4 The Weak Compacity Principle

B.4.1. Let E be a separable Frechet space (that is, a Frechet space in which
there exists a countable, everywhere dense subset) and (j,,)~= 1 a sequence of
continuous linear functionals on E such that
lim sup if,,(x)i <

,,_<Xl 00
for each x in a nonmeager subset S of E. Then there exists a subsequence

(j,,);:'=l which converges weakly in E' to somefEE'. (See B.1.7.)
Proof. From B.2.1 it follows that the f" are equicontinuous, and in
particular that sUPnifn(x)i < 00 for each x E E.
Now choose and enumerate as (x!c);:'= 1 a countable everywhere dense subset
of E. The numerical sequence (jn(Xd)~=l being bounded, a subsequence
(j",(n)~= 1 may be extracted so that limn_",f",(n)(x 1 ) exists finitely. (Here, CLl
is a strictly increasing map of the set N of natural numbers into itself.)
Again, the numerical sequence (f",(n)(X2))~= 1 being bounded, a subsequence
(f"'''2(n)~= 1 may be extracted so that (j"'''2(n)(x 2);:'= 1 is convergent to a finite
limit. (We are writing CL1CL2 for the composite map CLl 0 CL2') Proceeding in
this way, we obtain iterated subsequences CLV CLICL2' so that
exists finitely for 1 ~ j ~ i andi = 1,2, '. We now take the "diagonal
subsequence" {3 defined by
This has the crucial property that ({3(n:'= 1 is a subsequence of (CLl CLj(n:'= 1
so that
lim f8(n)(XI) (B.4.1)
,,_<Xl
exists finitely for each i.
Take now any x E E and any e > O. By equicontinuity, there is a neighbor-

hood U of 0 such that I/n( y)1 ~ e/4 for all n and all Y E U. Next choose i so
that x - XI e U (possible since the x, are everywhere dense in E). We
then have
!18(n)(X) - 18(n')(X)/ ~ Ifu(n)(x) - IU(n)(x,ll + IIU(n)(xl - lucn)(x,)1

+ 1/',(n)(x I) - IU(n)(x,>I
= 1f8Cn)(X - XI)! + 1f8(n')(X - x,)1 + 1f8cn)(XI) - I/lCn,)(XI)!
uniformly in nand n'. The existence of the limit (B.4.1) then shows that
provided n, n' > note). We thus infer that
I(x) = lim 16(n)(x) (B.4.2)

n_ac
exists finitely for each x E E. It is plain that I is a linear functional on E.

That lEE' (that is, is continuous) follows from the fact that (with the above
notation) II ( y) I ~ e/4 for Y E U. Thus lEE' and the definition (B.4.2)
ensures that lim n _ ac I 8(n) = I weakly in E'. So (/8(n:31 is a subsequence of
the type whose existence is asserted.
B.4.2. We remark that there is an analogue of B.4.1 which is valid for any
(not necessarily separable) Frechet space E, and indeed in a still wider
context.
Suppose that E is any topological linear space and that (/n):.1 is any
equicontinuous sequence of linear functionals on E. Although it may not be
possible to extract a subsequence of (/n) which converges weakly in E', the
following statement is true.
There exists a continuous linear functional Ion E with the property that,
corresponding to any given e > 0, any finite subset {Xl" .. , x r } ofE, and any
integer no, there is an integer n > no for which
(i = 1,2, .. ,r).
(Such an I is nothing other than a limiting point of the given sequence in

relation to the so-called weak topology on E' generated by E; see [E), pp.
88-89.)
[B.51 THE HAHN BANACH THEOREM 199
B.5 The Hahn-Banach Theorem

The only aspects of this many-headed theorem that are used in this book
are stated in B.5.l to B.5.3. For further discussion of the theorem and its
applications, see [E], Chapter 2.
B.5.l. Let E be a linear space, p a seminorm on E, L a linear subspace of

E, and fo a linear functional defined on L and satisfying
(x E L). (B.5.I)
Then there exists a linear functional f on E such that
f(x) = fo(x) (x E L), (B.5.2)

If(x)1 ~ p(x) (x E E). (B.5.3)
Proof. This may be taken verbatim from pp. 53-55 of [E], but note the
misprint on the last line of p. 54, where "~ " should read "~."
B.5.2. Let E be a Frechet space (or, indeed, any locally convex topological
linear space; see the Remark following Subsection B.2.2), A any nonvoid
subset of E, and Xo an element of E. Then Xo is the limit in E of finite linear
combinations of elements of A if and only if the following condition is
fulfilled: iff is any continuous linear functional E (that is, iff EE') such that
f(A) = {O}, thenf(xo) = O.
Proof. The" only if" assertion is trivial in view of the linearity and
continuity of J.
Suppose, conversely, that the condition is fulfilled. Let Lo denote the
closed linear subspace of E generated by A and suppose, if possible, that
Xo ~ Lo Since Lo is closed, and since the topology of E is defined by a family
of continuous semi norms , there is a continuous seminorm p on E such that
p(.y - x o) > I for all Y E L o, and hence also
p( y + xo) = p( - y - xo) > 1 (B.5.4)
Let L = Lo + <IIxo and define the linear functional fo on L by the formula

fo( Y + ,\xo) = A (y E L o, A E <II). The~, for x = y + ,\xo E L, (B.5.4) gives
for A =F 0
Ifo(x)1 = IAI ~ IAI'p(~ + xo) =p(y + Axo) =p(x);

and the same inequality is trivially valid for A = O. Appeal to B.5.1 shows
that fo can be extended into a linear functional f on E such that (B.5.3) is
true, which, since p is continuous, entails that f is continuous. On the other
hand, it is evident that fo, and therefore f too, vanishes on Lo, a fortiori
vanishes on A, while f(x o} = fo(xo} = 1 =f O. This contradicts the main

hypothesis.
B.5.3. Let E be a normed linear space and E' its dual (see B.1.7). If Xo E E,
there exists fEE' such that
Ilfll ~ 1
and
f(x o} = Ilxoll
Proof. Let L be the linear subspace of E generated by Xo. Define fa on
L by foU'.x o} = "llxoll (" E <1>). It now suffices to apply B.5.l, taking p(x} =
Ilxll: any extension f of fo of the type specified in B.5.l satisfies all
requirements.
APPENDIX C
The Dual of LP (I ~ P <(0);

Weak Sequential Completeness of Ll
The aim of this appendix is to give a proof of two results used in the
main text.
C.l The Dual of LP (1 ~ P < 00 )

Theorem. Let 1 ~ p < 00 and let F be any continuous linear functional
on LP. Then there exists an essentially unique function g ELp (lip + lip' = 1)
such that
(C.1)
for all f E V'. For any such function fI one has
Ilgllp = I FII == sup {I F(f)1 : f E LP, Ilf lip ~ I}. (C.2)
Proof. T~at any g E LP' satisfying (C.l) also satisfies (C.2), follows at
once from Holder's inequality and its converse (see Exercise 3.6). So we
confine our attention to the proof of the existence of g [its essential uniqueness
being a corollary of (C.2)]. This proof will be based upon use of the Radon-
Nikodym theorem (see [W], Chapter 6; [HS], Section 19; [AB], p. 406).
For this purpose we consider Borel subsets E of the interval X = (0,21T)
(compare [W], p. 93). The characteristic function of E may be extended by
periodicity, the result being denoted by XE and being a member of LP. The
number
(C.3)
is thus well-defined. Since F is linear, if is (finitely) additive. If we can show

that if is countably additive, it will follow that if is a complex Borel measure
on X ([W], p. 95; [HS], p. 329). Now, if E is the union of disjoint Borel sets
En (n = 1,2, .), then
co
XE = 2: XE
nal
n,
201
202 THE DUAL OF L'(l ~ p <CXl)
the series being convergent in V' since p < 00. (Here is the major reason for
breakdown of the theorem for p = 00.) The continuity of F, together with its
linearity, shows that therefore
v(E) = "" 2: v(En) ,

n=l
that is, that v is indeed countably additive.

Again since p < 00, the continuity of F shows that v is absolutely con-
tinuous relative to the restriction p. of Lebesgue measure to X (see [W], p. 98;
[HS], p. 312).
At this stage the Radon-Nikodym theorem ([W], Theorem 6.3d; [HS], p.
315) may be invoked to ensure the existence of an integrable function g on
X (which may be extended by periodicity) such that
v(E) = F(XE) = 2~ Is gdx = 2~ f XEUdx

for all Borel sets E c X. The linearity of F then shows that
F(f) = L ffg dX
holds for allf which are finite linear combinations of functions XE' Knowing
this, it is easy to conclude that g E V" and that
compare Exercise 3.6. Holder's inequality shows then that the difference
Fo(f) = F(f) - L fh dx
is a continuous linear functional on V, which vanishes for all f in the every-

where dense subset of V' formed of the finite linear 'Combinations of charac-
teristic functions of Borel sets. Fo must therefore vanish identically, that is,
(C. 1) holds for all f E V'.
Note: For 1 < P < 00 a different proof is possible; see [HS], pp. 222 ff.,
[AB]. p. 246. The proof given above is that used ([HS]. p. 351), to deal
with the case p = 1.
C.2 Weak Sequential Completeness of Ll

Although 12.3.10(2) is false whenp = 1, there is a sort of partial substitute
that sometimes will save the day.
[C.2] WEAK SEQUENTIAL COMPLETENESS OF Ll 203
Theorem. Suppose that (fili"'= 1 is a sequence of integrable functions

forming a weak Cauchy sequence in Ll, that is, for which
li~ L fkl dX
exists finitely for each I] E L"'. Then there exists a function i ELI to which
(f;) is weakly convergent, that is, which is such that
li~ Lfiil] dx = Lfil] dx

for earh I] E L''' .
The proof of this result is rather lengthy and the reader is referred to [E),
p.275.
APPENDIX D
A Weak Form of Runge's Theorem
We aim to employ the Hahn-Banach theorem (Appendix B.5.2) to prove a

simplified version of Runge's theorem adequate for the purposes of 12.9.8(3).
In what follows we denote by A the complex plane, by Ll the compactified
complex plane (that is, the Riemann sphere), by n a nonvoid open subset of
A, and by K a nonvoid compact subset of n. It will be assumed that
(1) A\K is connected (or, equivalently, that K is simply connected), and
tha.t
(2) there lies in n\K a smooth closed path r such that the Cauchy integral
formula
r
I(z) = ~ I(t) dt (D.1)
2m Jr t - z
holds for each z E K and each function I which is holomorphic on n.

We do not intend discussing in detail conditions that guarantee (2) (see
the Remarks below); suffice it to say here that the condition is evidently
fulfilled in the conditions prevailing in 12.9.8.
Theorem. With the above notations and hypotheses, each function I

which is holomorphic on n is the limit, uniformly on K, of polynomials in z.
(As is customary, the symbol "z" is used to denote a complex number and
also the natural complex coordinate function on A.)
Proof. Denote by C(K) the complex linear space of continuous, complex-
valued functions on K, into which we introduce the norm
IIIII = sup {1/(z)1 : zeK}.

Also, let H denote the subspace of C(K) formed of those I e C(K) which are
restrictions to K of functions holomorphic on n. According to B.5.1 and
B.5.2, our task will be finished as soon as it is shown that any continuous
linear funotional F on C(K) with the property that
F(u,,) = 0 (n = 0,.1, 2, ... ), (D.2)

205
206 A WEAK FORM OF RUNGE'S THEOREM
where u" is the restriction to K of the function z ~ z", satisfies also

FU) = 0 (D.3)
for any feB.
For any t e 4\K, let ft denote that element of C(K) obtained by restricting
to K the function z ~ (t - Z)-I. If fe B, the formula (D.l) may be applied
to show thatfis the limit in C(K) (that is, uniformly on K) of linear combina
tions of functions It with t e r; for this it suffices to approximate the integral
appearing in (D.I) by Riemann sums. Thus (D.3) is implied by (and is actually
equivalent to) the assertion
(t e r). (D.4)
In order to show, finally, that (D.2) implies (D.4), we examine the function
4>(t) = FU,) , (D.5)
defined for t e 4\K. Using the continuity of F on C(K), it is easy to verify

that 4> is holomorphic at all points of 4\K: we leave this as a simple exercise
for the reader. Moreover, if ItI is sufficiently large,
ao z"
(t - Z)-1 = ~ -,
L, t,,+1
,,~O
the series converging uniformly for z e K, 80 that continuity of F shows that
L: F(u,,)t-"-I,
ao
4>(t) = (D.6)
n-O
again for sufficiently large Itl. Now (D.6) shows first that 4> can be extended
to 3:\K in such a way as to be holomorphic on a neighborhood of 00 (simply by
setting 4>(00) = 0); and second, in conjunction with (D.2), that 4> vanishes
on a neighborhood of 00. Since, by (1), 4\K is connected, the same is true of
a\K = (4\K) U {oo}, and it follows that 4> must vanish throughout the whole
of 4\K. In particular, (D.4) is true. This completes the proof.
Remarks. For a different approach to Runge'S theorem and related
questions, see [He], pp. 149-153, and the references cited there. The theorem
has ramifications and analogues extending to Riemann surfaces (see, for
example, the relevant portions of Behnke and Stein's "Theorie der analy.
tischen Funktionen einer komplexen Veranderlichen") and to functions of
several complex variables or on complex analytic manifolds (see Gunning
and Rossi's "Analytic Functions of Several Complex Variables," Chapter I,
Section F, and Chapter VII).
There are also real variable analogues of Runge's theorem, applicable to
solutions of linear partial differential equations; see [E], p. 396, and the
references cited there.
Bibliography
Background reading in general topology, functional analysis, and integration

theory.
[AB] ASPLUND, E. AND BUNGART, L. A Fir8t Cour8e in Integration. Holt,

Rinehart and Winston, Inc., New York (1966).
DUNFORD, N. AND SCHWARTZ, J. T. Linear Operator8, Parts I, II.
Interscience Publishers, Inc., New York (1958, 1963).
[E) EDWARDS, R. E. Functional Analysis: Theory and Applications.
Holt, Rinehart and Winston, Inc., New York (1965).
[GP] GOFFMAN, C. AND PEDRICK, G. Fir8t Cour8e in Functional Analysis.
Prentice-Hall, Inc., Englewood Cliffs, New Jersey (1965).
[HS] HEwrrr, E. AND STROMBERG, K. Real and Ab8tract AnalY8is. Springer-
Verlag, Berlin (1965).
[K] KELLEY, J. L. General Topology. D. Van Nostrand Company, Inc.,
Princeton, New Jersey (1955).
[Na] NATANSON, 1. P. ThetJry of Functions of a Real Variable. Frederick

Ungar Publishing Co., New York (1965).
[W] WILLIAMSON, J. H. Lebesgue Integration. Holt, Rinehart and Winston,

Inc., New York (1962).
References to Fourier series and other specialized topics.
[A] ALEXITS, G. Contergence Problems of Orthogonal Series. Pergamon

Press, Inc., New York (1961).
[Am] ALMGREN, F. J. Plateau'8 Problem: an Invitation to Vanfold Geometry.

W. A. Benjamin, Inc., New York (1966).
[B) BACHMAN, G. Elements of Ab8tract Harmonic Analysis. Academic

BARY, N. A Treatise on Trigonometric Series, Vols.l and 2. Pergamon

207
208 BIBLIOGRAPHY
[Be] BELLMAN, R. A Brief Introduction to Theta Functions. Holt, Rinehart

and Winston, Inc., New York (1961).
[Bes] BESICOVITCH, A. S. Almost Periodic Functions. Cambridge University
Press, New York (1932).
[BK] BUTZER, P. L. AND KOREVAAR, J. (editors). On Approximation

Theory. Birkhauser Verlag, Basel und Stuttgart (1964).
[Bo] BOERNER, H. Representations of Groups with Special Consideration
for the Needs of Modern Physics. North Holland Publishing Co. and
Interscience Publishers, Inc., New York (1963).
[Boa] BOAS, R. P. Jr. Integrability Theoremsfor Trigonometric Transformll.

Ergebnisse der Mathematik und ihrer Grenzgebiete, Bd. 38. Springer-
Verlag, New York (1967).
[CR] COURANT, R. AND HILBERT, D. Methods of Mathematical Physics,

Vol. II. Interscience Publishers, Inc., New York (1962).
[dBR] DE BRANGES, L. AND ROVNYAK,.J. Square Summable Power Seriu.
Holt, Rinehart and Winston, Inc., New York (1966).
[DW] DORAN, R. S. AND WICHMANN, J. Approximate Identities and Factor-

ization in Banach Modules. To appear Springer-Verlag Lecture Notes
in Mathematics Series.
EDWARDS, R. E. Integration and Harmonic Analysis on Compacl

Groups. London Math. Soc. Lecture Note Series 8, Cambridge
University Press, New York (1972).
[G) GROSS, L. Harmonic Analysis on Hilbert Space. Mem. Amer. Math.

Soc., No. 46, Amer. Math. Soc., Providence (1963).
[Ga] GARSIA, A. M. Topics in Almost Everywhere Convergence. Lectures in

Advanced Mathematics, 4. Markham Publishing Co., Chicago (1970).
[GP] GOFFMAN, C. AND PEDRICK, G. First Course in Functional Analysis.

Prentice-Hall, Inc., Englewood Cliffs, New Jersey (1965).
[GV] GELFAND, I. M. AND VlLENKIN, N. YA. Generalized Functions, Vol. 4

Applications of Harmonic Analysis. Academic Press, Inc., New York
(1964).
[R] HIRsCHMAN, I. I., JR. Infinite Series. Holt, Rinehart and Winston,
Inc., New York (1962).
[Ra] HARDY, G. H. Divergent Series. Oxford University Press, New York
(1949).
[HaR] HARDY, G. H. AND ROGOSINSKI, W. Fourier Series. Cambridge
BIBLIOGRAPHY 209
[He] HEINS, M. T0pic8 in Complex Function Theory. Holt, Rinehart and

Winston, Inc., New York (1962).
[Hel] HELSON, H. Luture8 on Invariant Subapace8. Academic Press, Inc.,
New York (1964).
[Hew] HEWITT, E. A Survey oj Ab8tract Harmonic AnalyBiB. Survey. in
Applied MathematicB, IV. John Wiley & Sons, Inc., New York
(1958), 107-168.
[Hi] HIRSCHMAN, I. I., JR. (editor). StudieB in Real and Complex AnalyBiB.
StudieB in MathematicB, Vol. 3. The Math. Association of America;
PrenticeHall, Inc., Englewood Cliffs, New Jersey (1965).
[HLP] HARDY, G. H., LITTLEWOOD, J. E. AND P6LYA, G. InequalitieB.
Cambridge University Press, New York (1934).
[Ho] HOFFMAN, K. Ba~h Spacea oj Analytic FUfICtionB. PrenticeHall,

Inc., Englewood Cliffs, New Jersey (1962).
[HR] HEWITT, E. AND Ross, K. A. Ab8tract Harmonic AnalyBiB, I, II.
Springer.Verlag, Berlin (1963, 1970).
[HW] HARDY, G. H. AND WRIGHT, E. M. An Introduction to the Theory oj
Number8. Oxford University Press, New York (1938).
[I] Izum, S.I. Introduction to the Theory oj Fourier SerieB. Institute of

Mathematics, Academia Sinica, Taipei (1965).
[Ka] KAzAmNOFF, N. D. Analytic Inequalitiu. Holt, Rinehart and
[Ka.h] KAHANE, J.P. Algebraa de Convolucion de S'UCe8i0ne8, Funcionu 11
MedidaIJ Sumablu. Oursos y Seminarios de Matematica, Fasc. 6.
Univ. de Buenos Aires (1961).
KAHANE, J.P. Ucture8 on Mean Periodic FunctionB. Tata Institute
of Fundamental Research Lectures on Mathematics and Physics,
No. 15. Bombay (1959).
[Ka.h:a] KAHANE, J .P. SerieB de Fourier abBolument convergente8. Ergebnisse

der Mathematik und ihrer Grenzgebiete, Bd. 50. Springer.Verlag,
Berlin-Heidelberg-New York (1970).
[KS] KAHANE, J ..P. AND SALEM, R. Enaemblu ParJaila et SerieB Trigona.

m&riquu. Hermann & Cie, Paris (1963).
[KSt] KAczMARZ, S. AND STEINHAUS, H. Theorie der Orthogonal,."ihen.

Chelsea Publishing Company, New York (1951).
[Kz] KATZNELSON, Y. Introduction to Harmonic AnalyBiB. John Wiley &

Sons, Inc., New York (1968).
210 BIBLIOGRAPHY
LORENTZ, G. G. Bern8tein Polynomials. University of Toronto Press

(1955).
[L:a] LORENTZ, G. G. Approximation oj Functions. Holt, Rinehart and
[Li] LITTLEWOOD, J. E. Some Problema in Real and CompTa: Analy8i8.

Lecture notes, Univ. of Wisconsin, Madison (1962).
[Lo] Looms, L. H. An Introduction to Ab8tract Harmonic Analy8i8. D.
Van Nostrand Company, Inc., Princeton, New Jersey (1953).
[M] MANDELBROJT, M. S6riu de Fourier et Cla8BeB QuaBianalytiquea de

Fonctions. Gauthier.Villars, Paris (1935).
[Mk] MAAK, W. FaBtperiodiBche Fun1ctionen. Springer.Verlag, Berlin
(1950).
[Moz] MOZZOCHI, C. P. On the pointwiBe convergence oj Fourier SerieB. Lec-

ture Notes in Mathematics, 199. Springer.Verlag, Berlin-Heidelberg-
New York (1971).
[N] NAIMA.BX, M. A. Normed Ri7llJ8. P. Noordhoff, N. V., Groningen,

Netherlands (1959).
[P] PITT, H. R. Tauberian Theorema. Oxford University Press, New

York (1958).
[Ph] PHELPS, R. R. Lecturu on Choquet'8 Theorem. D. Van Nostrand
Company, Inc., Princeton, New Jersey (1966).
[PS] P6LYA, G . .AND SZEGD, G. AuJgaben und LehreBdtze aUB der Analy8i8,
Bd. I, II Springer-Verlag, Berlin (1925).
[R] RUDIN, W. Fourier AnalY8iB on GroupB. Interscience Publishers, Inc.,
New York (1962).
[Re] REITER, H. ClaBBical Harmonic Analy8i8 and Locally Compact GrOUpB.
Oxford Mathematical Monographs. Oxford University Press, New
York (1968).
[Ri] RICKABT, C. E. Banach AlgebraB. D. Van Nostrand Company, Inc.,
Princeton, New Jersey (1960).
[SMA] Studiu in Modern Analy8i8, Vol. I. (edited by R. C. Buck) The Math.
Assoc. of America, PrenticeHall, Inc., Englewood Cliffs, New Jersey
(1962).
[So] SOBOLEV, S. L. ApplicationB oj Functional AnalyBiB in Mathematical
PhyBic8. Amer. Math. Soc. Translations of Math. Monographs, Vol.
VII, Providence (1963).
[T] TONELLI, L. Serie Trigonometriche. Bologna (1928).
BIBLIOGRAPHY 211
[Ti) TIMAN, A. F. Theory of Approximation of Functiona of a Real Variable.

The Macmillan Company, New York (1963).
[TV) TOLSTOV, G. P. FO'Urier Seriu. Prentice-Hall, Inc., Englewood Cliffs,

New Jersey (1962).
[vW) VAN DER WAERDEN, B. L. Moderne Algebra. Springer-Verlag, Berlin
(1937).
[We) WEIL, A. L'InUgration da718 lea GrO'Upe8 Topologiquu et Sell Applica-
tio7l8. Hermann & Cie, Pari! (19&1).
[Wi) WIENER, N. The FO'Urier Integral and Certain of Ita Applicatio7l8.
Cambridge University Press, New York (1933).
[Zl.:Z) ZYGMUND, A. Trigonometrical Seriu, Vola. 1 and 11. Cambridge
Research Publications
The following list makes no attempt at completeM88" it i8 rather a 8upplement

ro the bibliographies of 8uch 8tandard references as [Z], [Ba], and [R].
ALJANCIC, S.
[1] On the integral moduli of continuity in Lp (1 < P < C() of Fourier series
with monotone coefficients. Proc. Amer. Math. Soc. 17 (1966), 287-294.
BLOOM, W. R.
[1] Bernstein's inequality for locally compact Abelian groups. J. Austr. Math.
Soc. XVII(I) (1974), 88-102.
[2] A converse of Bernstein's inequality for locally compact Abelian groups.
Bull. Austr. Math. Soc. 9(2) (1973), 291-298.
[3] Jackson's theorem for locally compact Abelian groups. ibid. 10(1) (1974),
59-66.
[4] Jackson's theorem for finite products and homomorphic images of locally
compact Abelian groups. ibid. 12(2} (1975), 301-309.
BOAS, R. P. JR.
[1] Beurling's test for absolute convergence of Fourier series. Bull. Amer.
Math. Soc. 66 (1960), 24-27.
[2] Fourier Series with positive coefficients. J. Math. Anal. Appl. 17 (1967),
463-483.
BRED ON, G. E.
[1] A new treatment of the Haar integral. Mich. Math. J. 10 (1963), 365-373.
Buoy, R. S. AND MALTESE, G.

[1] Extreme positive definite functions and Choquet's representation theorem.
J. Math. Anal. Appl. 12 (1965), 371-377.
BUTZER, P. L. AND NESSEL, R. J.

[1] Ober eine verallgemeinerung eines satzes von de la Vallee POU88in.
International Serie8 in Numerical Mathematic8, 5, pp. 45-48. Birkhauser
Verlag (1964).
CARLESON, L.
[1] On convergence and growth of partial sums of Fourier series. Acta Math.
116 (1966), 135-157.
213
214 RESEARCH PUBLICATIONS
Cm:N, Y. M.
[1] An ahnoet everywhere divergent Fourier series of the class L(log + log + L )1-
J. London Math. Soo. 44 (1969), 643-654.
COllEN, P. J.
[1] Factorization in group algebras. Duke Math. J. 18 (1959), 199-205.
[2] On homomorphisms of group algebras. Amer. J. Math. 81 (1960), 213-226.
[3] On a conjecture of Littlewood and idempotent measures. Amer. J. Math. 81
(1960), 191-212.
[4] A note on constructive methods in Banach algebras. Proo. Amer. Math.
Soo. 11 (1961), 159-163.
CuRTIS, P. C. JR. AND FIG1.TALAHANCA, .A.

[1] Factorization theorems for Banach algebras. Function Algebras, pp. 169-
185. Scott, Foresman and Company, Chicago (1966).
DE BBUIJN, N. G.
[1] Functions whose differences belong to a given class. Nieu1J} Arch. Wilkunae
13 (1951), 194-218.
[2] A difference property for Riemann integrable functions and for some
similar classes of functions. Nederl. Akad. WUensch. Proo. Ser. A55 (1952),
145-151.
DE LEEUW, K. AND KATZNELSON, Y.
[1] On certain homomorphisms of quotients of group algebras. [Brad J. Math.
I (1964), 120-126.
EDWARDS, D. A.
[1] On translates of L> functions. J. London Math. Soo. 38 (1961), 431-432.
EDWARDS, R. E.
[1] Criteria for Fourier transforms. J. Amer. Math. Soc. 7 (1967), 239-246.
[2] Translates of L> functions and of bounded measures. J. Amer. Math. Soo.
IV (1964), 403-409.
[3] Boundedness principles and Fourier theory. To appear Pacific J. Math.
EDWARDS, R. E. AND HEwrrr, E.

[1] Pointwise limits for sequences of convolution operators. Acta Math. III
(1965), 181-218.
EDWARDS,R. E. AND PRICE, J. F.

[1] A naively constructive approach to boundedness principles, with applica.
tions to harmonic analysis. L'Ens. math. XVI, fasc. 3-4 (1970), 255--296.
EssiN, M.
[1] Studies on a convolution inequality. Ark. Mat. 5 (9) (1963), 113-152.
GAUDET, R. J. AND G..un.EN J. L. B.

[1] An elementary proof of part of a classical conjecture. Bull. Austr. Math.
Soo.3 (1970), 289-291.
RESEARCH PUBLICATIONS 215
GoDEMENT, R.
[1] Les fonctions de type positif et 1& theorie des groupes. TraM. Amer. Math.
Soc. 63 (1948), 1-84.
GOES, G. AND GoES, S.
[1] Sequences of bounded variation and sequences of Fourier coefficients. I.
Math. Z. 118 (1970), 93-102.
HARDY, G.H.
[1] Notes on some points in the integral calculus, LXVI, The arithmetic mean
of a Fourier constant. MU8enger oj Math. 58 (1928), 50-52.
HEBz, C. S.
[1] Fonctions operant sur certains semi-groupes. O. R. Acad. Sci. Paris 155
(1962), 2046-2048.
HEWl'rr, E.
[1] The ranges of certain convolution operators. Math. Scani/,. 15 (1964),
147-155.
HEWIT.r, E. AND HEWITT, R. E.
[1) The Gibbs-Wilbraham phenomenon: an episode in Fourier analysis. To
appear The Mathematical Intelligencer.
HmscHMAN, I. I., JR.
[1] On multiplier transformations. Duke Math. J. 16 (1959), 221-242.
IZUMI, M. AND IZUMI, S.-I.
[1] Fourier series of functions of bounded variation. Proc. Japan Acad. 44 (6)
(1968), 415-417.
[2] Divergence of Fourier series. Bull_ A'UBtr. Math. Soc. 8 (1973), 289-304.
[3] On absolute convergence of Fourier series. Ark. JOr Mat. 7(12) (1967),
177-184.
[4] Absolute Convergence of Fourier series of convolution functions. J. Approz.
Theory 1 (1968), 103-109.
KAHANE, J.-P.
[1] Idempotents and closed subalgebras of A(Z). Function Algebras, pp. 198-
207. Scott, Foresman and Company, Chicago (1966).
[2] Sur certaines classes de mea de Fourier absolument convergentes. J.
Math. PUreB et Appl. 35 (1956), 249-258.
[3] Sur un probleme de Littlewood. Nederl. Akad. Wetensch. Proc. Ser. AGO =
Indag. Math. It (1957), 268-271.
[4] Sur les rtSarrangements des suites de coefficients de Fourier-Lebesgue.
O. R. Acad. Sci. Paris. Ser. A-B 265 (1967), A310-A312.
KAHANE, J.-P. AND KATZNELBON, Y.
[1] Sur les series de Fourier uniformement convergentes. O. R. Acad. Sci. Paris
181 (1965), 3025-3028.
KANo, T. AND UOHIYAMA, S.
[1] On the convergence in L of some special Fourier series. Proc. Japan Acad.
53 A(2) (1977), 72--77.
216 RESEARC.H PUBLICATIONS
MCGEHEE, O. C.
(1] Certain isomorphisms between quotients of group algebras. Notices Amer.
Math. Soc. 13 (1966), 475.
MEN'SHOV, D. E.
[1] On convergence in measure of trigonometric series. (In Russian) Trudy Mat.
Inst. im V. A. Steklova 32 (1950), 3-97.
MITJAGIN, B. S.
[1] On the absolute convergence of the series of Fourier coefficients. (In
Russian) Dokl. 4kad. Nauk SSSR 157 (1964), 1047-1050.
MUELLER, M. J.
[1] Three results for locally compact groups connected with the Haar measure
density theorem. Proe. Amer. Math. Soc. 16 (1965), 1414-1416.
NEWMAN, D. J.
[1] The non existence of projections from LI to HI. Proc. Amer. Math. Soc. 12
(1961), 98-99.
[2] An LI extremal problem for polynomials. Proc. Amer. Math. Soc. 16 (1965),
1287-1290.
RAY, K. C. AND LAHIRI, B. K.
[1] An extension of a theorem of Steinhaus. Bull. Calcutta Math. Soc. 56 (1964),
29-31.
RIDER, D.
[1] Central idempotents in group algebras. To appear Bull. Amer. Math. Soc.
[2] Transformations of Fourier coefficients. To appear Padlu; J. Math.
ROONEY, P. G.
[1] On the representation of sequences as Fourier coefficients. Proc. Amer.
Math. Soc. 11 (1960), 762-768.
RUBEL, L. A.
[1] Uniform distributions and densities on locally compact Abelian groups.
Symposia on Theor. Physics and Math., Vol. 9,183-193. Plenum Press, New
York (1969).
RUDIN, W.
[1] Representation of functions by convolutions. J. Math. Meeh. 7 (1958),
103-116.
[2] A strong converse of the WienerLevy theorem. Canad. J. Math. 14 (1962),
694-701.
[3] Idempotents in group algebras. Bull. Amer. Math. Soc. 69 (1963), 224-227.
[4] Some theorems on Fourier coefficients. Proc. Amer. Math. Soc. 10 (1959),
855-859.
[5] Positive definite sequences and absolutely monotonic functions. Duke
Math. J. 26 (1959), 617-622.
RUDIN, W. AND SCHNEIDER, H.
(1] Idempotents in group rings. Duke Math. J. 31 (1964), 585-602.
RESEARCH PUBLICATIONS 217
RYAN, R.
[I] Fourier transforms of certain classes of integrable functions. TraM. Amer.
Math. Soc. 105 (1962), 102-111.
SHAH, T S.
[I] Positive definite functions on commutative topological groups. Scientia
Sinica (5)XIV (1965), 653-665.
SRINIVASAN, T. P. AND WANG, J.-K.
[1] Weak Dirichlet Algebras. Function Algebras, pp. 216-249. Scott, Foresman
and Company, Chicago (1966).
STEIN, E. M.
[1] On limits of sequences of operators. Ann. Math. 7' (1961), 14:0-170.
TAIBLESON, M.
[1] Fourier coefficients of functions of bounded variation. Proc. Amer. Math.
Soc. 18 (1967), 766.
TELJAKowsKir, S. A.
[1] Some estimates for trigonometric series with quasiconvex coefficients
(Russian). Mat. Sbornik 63 (105) (1964), 426-444 = A.M.S. Translations (2)
86 (1970), I09-13l.
[2] On the question of convergence of Fourier series in the L-metric (Russian).
Mat. Zametki 1 (1967), 91-98. MR 34 # 8067.
TIMAN, M. F.
[I] Best approximation of a function and linear methods of summing Fourier
series. (In Russian) lzv. Akad. Nauk SSSR Ser. Mat. 29 (1965), 587-604.
WEISS, G.
[I] Harmonic Analysis. Studi68 in Mathematics, Vol. 3, pp. 124-178. The
Math. Assoc. of America; PrenticeHall, Inc., Englewood Cliffs, New
Jersey (1965).
YADAV, B. S.
[1] On a theorem of Titchmarsh. Publ. Inse. Math. (Beograd) (N.S.) 2(18)
(1962), 87-92 (1963).
YADAV, B. S. AND GOYAL, O. P.
[1] On the absolute convergence of Fourier series. J. MaharajaSayajirao Unit}.
Baroda 10 (1961), 29-34.
ZAlIlANSKY, M.
[1] Approximation et Analyse Harmonique, I. Bull. Sci. math., 2" serie, 101
(1977), 3-70.
[2] Approximation et Analyse Harmonique, II. Bull. Sci. math., 2" serie, 101
(1977), 149-180.
ZUK, V. V.
[1] On the absolute convergence of Fourier series. (In Russian) Dokl. Akad.
Nauk SSSR 180 (1965), 519-522.
Index
Numerals in boldface type refer to the exercises.
Abel kernel (Abel-Poisson kernel), 103 Category, first (meager), 187

Abel means, 103 Category, second (nonmeager), 187
Abel summability, 103,6.16 Category theorem, 187
Absolutely continuous function, 30 ff., Cesaro means, 8, 79, 82, 87
136, 137 Cesaro summability, 82 ff., 5.6-5.8,
Absolutely convergent Fourier series, 87 ff.
173 ff. in mean, 87
Affine map, 4.9 pointwise, 94 ff.
Algebra, convolution, 52, 56, 57 Change of variable in A, 191
Algebraic (Hamel) base, 165 Character, 18,3.19
Almost periodic function, 45 principal, 19
Approximate identity, 59 ff. Character group, 20, 2.3
Approximation by trigonometric poly- Circle group, 15
nomials, 42, 99 ff., 6.5-6.10, 6.17, Closed graph, 195
131 theorem, 197
Closed linear. operator, 195
Comeager, 187
Banach algebra, 52, 57, 72, 126, 173,
Compacity principle, 197
10.16
Complex homomorphism, 69 ff., 4.1,
Banach space, 192
4.7
Banach-Steinhaus theorem, 195
Conjugate Dirichlet kernel, 110
Base, algebraic (Hamel), 165
Conjugate exponent, 28
Base, topological, 165
Conjugate series, 109, 110
Bases of trigonometric polynomials,
Continuity, modulus of, 36, 99, 135
165, 166
Convergence,
Bernstein polynomials, 91
in mean, 29, 131
Bernstein's inequality, 1.9, 101
in measure, 6
Bernstein's theorem, 91
weak (in dual space), 193
Bochner's theorem, 151
Convex sequence, 111
Bosanquet-Kestelman lemma, 3.14
Convolution,
Bounded variation, 33, 34
characterization of, 59
of functions, 50 ff.
Cantor-Lebesgue theorem, 2.14 of sequences, 64, 3.15
Cantor group, 153 Convolution algebra, 52, 56, 57
Capacity, 171 Coset (of reals modulo 211'), 15
Carleson's theorem, 169 Cosine series, 114 ff., 117 ff.
219 - 221
222 INDEX
Defining family (of seminorms), 192 Generalized function (distribution), 9

Dense, everywhere, 187 Generalized (topological) nilpotents,
Dense, nowhere, 187 3.12
Density theorem, 42 Gibbs phenomenon, 10.8
Derivative, pointwise, 7, 30, 33 Graph, 195
Derivative, symmetric, 6.16 Group, 15
Diagonal subsequence, 3.5, 140, 197 character, 20,2.3
Dini's test, 159 circle, 15
Dirac afunction, 62 of integers, 15
Dirichlet kernel, 5.1, 79 locally compact Abelian, 15
conjugate, 110 quotient, 15
modified, 11 0 of real numbers, 15
Distribution (generalized function), 9 topological, 15
Divergence of Fourier series, 160 ff. Group algebra, 62 ff.
Fejer's example, 161
Dual of co(Z), 8.12
Haa.r (invariant) integral, 21 ff.
Dual of V', 201
Hadamai-d sequence, 138
Dual space, 193
Hahn-Banach theorem, 199
Duality, 20
Hamel (8J.gebraic) base, 165
Duality law (Pontryagin), 20
Ha.rdy spaces, 3.9, 8.15
Hardy's theorem, 84,5.8,169
Equidistributed sequence, 2.15
Harmonic (spectral) analysis, 19
Equivalence class,
Harmonic (spectral) synthesis, 19
(coset) of reals modulo 2"" 15
Hilbert space, 130
of functions, 27
Holder (Lipschitz) condition, 100
HOlder's inequality, 28, 30
Factorization problems, 53, 124, 134
converse of, 3.6
Fatou's theorem, 6.12
Homogeneous Banach space, 87
Fejer kernel, 79
Homomorphism, 72 ff., 4.2-4.5, 4.8,
Fejer's lemma, 2.16
4.9,8.1,179
Fourier coefficients, 1, 30 ff.
complex, 69 ff., 4.1, 4.7
Fourier-Lebesgue series, 30
problem, 72 ff.
Fourier series, 2
Fourier transformation, 39, 52
Frechetspace, 191 Ideal, 54, 3.4
Function, closed, 54, 3.4
absolutely continuous, 32, 136 maximal, 4.1
almost periodic, 45 regular (modular), 4.1
of bounded variation, 33, 34, 8.13, Idempotents, 53
156, 157 Inequality,
generalized, 9 Bernstein's, 1.9, 102
integrable, 22, 27 Holder's, 28, 30
lower semicontinuous, 188 converse of, 3.6
measurable, 27 Minkowski's, 27, 30
periodic, 15 Integral,
positive definite, 149 Haar (invariant), 21 ff.
quasianalytic, 2.8 Lebesgue, 22, 26
INDEX 223
Integral-eont. Measure algebra, 72, 73, 75

relatively invariant, 2.4 Measure, interior, 3.16
Riemann, 22, 26 Minimal (positive definite function),
RiemannStieltjes, 34 9.5
Interior measure, 3.16 Minkowski's inequality, 27, 30
Invariant (Haar) integral, 21 ff. Modular ideal, 4.1
Invariant subspace, 17 Module, 57,3.2
Inversion formula, 103, 134 Modulus of continuity, 36, 100, 135
Isomorphism problem, 75 Multiplier, 76
Isoperimetric problem, 8.16
Nonmeager, 187
Jackson polynomial, 102 Norm, 191
Jordan's test, 149 V'" 27
11', 29
quotient, 195
Kernel, Nowhere dense, 187
Abel.Poisson, 103
conjugate, Dirichlet, 110
Dirichlet, 5.1, 79 Open linear operator, 195
Fejer,79 Open mapping theorem, 196
modified Dirichlet, 110 Order of magnitude,
Kolmogorov Seliverstov Plessner theo of 8Ni, 166 ff., 10.2
rem, 169, 171 of aNi, 97 ff., 6.18
Kolmogorov's theorem, 138, 171,10.21 Orthogonality relations, 3, 25, 26
Orthonormal base, 130
Lacunary series, 5.6, 6.13

Lebesgue integral, 22, 26 Parseval formula, 1.7,131 ff., 8.17,152,
Lebesgue point, 96 171 ff.
Lebesgue set, 96 Partial summation, III
Levy's theorem, 178 Periodic function, 15
Lipschitz (HOlder) condition, 100 Poisson (Abel.Poisson) kernel, 103
Localization principle, 81 ff. Poisson's summation formula, 10.9-
Locally compact Abelian group, 15 10.11
Locally convex space, 195 Pointwise convergence of Fourier series,
Lower semicontinuous function, 188 33, 6.12, 6.13, 130 if.
Lusin.Denjoy theorem, 2.13 Pointwise derivative, 7, 30, 33
Pointwise summability of Fourier
series, 87, 94 ff., 103, 6.14, 6.16
Majorant, Pontryagin duality law, 20
for 8Ni, 165 Polynomial,
for aNi, 97 Bernstein, 91
Maximal ideal, 4.1 Jackson, 102
Meager, 187 trigonometric, 1.7, 44
Mean convergence, 29 Positive definite function, 149
of Fourier series in L'J, 131 minimal, 9.5
224 INDEX
Principal character, 19 Summability,

Projection principle, 8.14 Abel (Abel-Poisson), 103, 6.16
Cesaro, 82 ff., 5.6-5.8, 87 ff.
Riemann, 6.14
Quasianalytic function, 2.8 Summation formula, Poisson's, 10.9-
Quotient group, 15 10.11
Quotient map, 193
Quotient norm, 194
Quotient seminorm, 194
Tauberian theorem, 84
Quotient space, 193
Topological base, 165
Quotient topology, 15, 193
Topological group, 15
Topological linear space, 191
Topology, quotient, 15, 193
Regular ideal, 4.1
Total variation, 33
Relatively invariant integral, 2.4
Translates of functions, 16 ff.
Representation by trigonometric series,
Translation operators, 16
3 ff.
Trigonometric polynomial, 1.7
Riemann integral, 22, 26
on Z, 44
Riemann-Lebesgue lemma, 36
Trigonometric polynomials,
Riemann-Stieltjes integral, 34
bases of, 165, 166
Riemann summability, 6.14
density of, 42
Riesz-Fischer theorem, 133
Trigonometric series, 1 ff., 2.13, 2.14
Runge's theorem, 205
Saturated sequence, 6.10 Uniform boundedness principle, 194,

Seminorm, 191 195
quotient, 194 Uniqueness theorem, 40
Sequences, convolution of, 64, 3.15
Series, Variation,
Fourier, 2 bounded, 33, 34
Fourier-Lebesgue, 30 total, 33
trigonometric, 1 if., 2.13, 2.14
Set of differences, 3.16
Sine series, 114, 115 ff.
Wave equation, 8.10
Space,
Weak compacity principle, 197
Banach,192
Weak convergence in dual space, 193
Frechet, 191
Weak sequential completeness, 202
locally convex, 195
Weierstrass' approximation theorem,
topological linear, 191
90
Spectral (harmonic) analysis, 19
Weierstrass-Stone theorem, 91
Spectral (harmonic) synthesis, 19
Wiener's theorem, 8.13, 177
Spectral radius formula, 8.8
Steinhaus' theorem, 3.16
Subalgebras, 53
Submodule, 3.2 Zero divisors, 53
Graduate Texts in Mathematics
Soft and hard cover editions are available for each volume up to vol. 14, hard cover
only from Vol. 15
1 TAKEUTI/ZARING. Introduction to Axiomatic Set Theory.

2 OXTOBY. Measure and Category.
3 SCHAEFFER. Topological Vector Spaces.
4 HILTON/STAMMBACH. A Course in Homological Algebra.
5 MACLANE. Categories for the Working Mathematician.
6 HUGHES/ PIPER. Projective Planes.
7 SERRE. A course in Arithmetic.
8 T AKEUTI/ ZARING. Axiomatic Set Theory.
9 HUMPHREYS. Introduction to Lie Algebras and Representation Theory.
2nd printing, revised.
10 COHEN. A Course in Simple Homotopy Theory.
11 CONWAY. Functions of One Complex Variable. 2nd ed.
12 BEALS. Advanced Mathematical Analysis, xi, 230 pages. 1973.
13 ANDERSON/ FULLER. Rings and Categories of Modules.
14 GOLUBITSKY / GUILLEMIN. Stable Mappings and Their Singularities.
15 BERBERIAN. Lectures in Functional Analysis and Operator Theory.
16 WINTER. The Structure of Fields.
17 ROSENBLATT. Random Processes. 2nd ed.
18 HALMOS. Measure Theory.
19 HALMOS. A Hilbert Space Problem Book.
20 HUSEMOLLER. Fibre Bundles. 2nd ed.
21 H'tJMPHREYS. Linear Algebraic Groups.
22 BARNES/ MACK. An Algebraic Introduction to Mathematical Logic.
23 GREUB. Linear Algebra. 4th ed.
24 HOLMES. Geometric Functional Analysis and Its Applications.
25 HEWITT/ STROMBERG. Real and Abstract Analysis. 4th printing.
26 MANES. Algebraic Theories.
27 KELLEY. General Topology.
28 ZARISKI/ SAMUEL. Commutative Algebra 1.
29 ZARISKI/ SAMUEL. Commutative Algebra II.
30 JACOBSON. Lectures in Abstract Algebra I: Basic Concepts.
31 JACOBSON. Lectures in Abstract Algebra II: Linear Algebra.
32 JACOBSON. Lectures in Abstract Algebra III: Theory of Fields and
Galois Theory.
33 HIRSCH. Differential Topology.
34 SPITZER. Principles of Random Walk. 2nd ed.
35 WERMER. Banach Algebras and Several Complex Variables. 2nd ed.
36 KELLEy/NAMIOKA. Linear Topological Spaces.
37 MONK. Mathematical Logic.
38 GRAUERT/ FRITZSCHE. Several Complex Variables.
39 ARVESON. An Invitation to C*-Algebras.
40 KEMENY / SNELL/ KNAPP Denumerable Markov Chains. 2nd ed.
41 ApOSTOL. Modular Functions and Dirichlet Series in Number Theory.
42 SEERE. Linear Representations of Finite Groups.
43 GILLMAN/ JERISON. Rings of Continuous Functions.
44 KENDIG. Elementary Algebraic Geometry.
45 LOEVE. Probability Theory. 4th ed. Vol. 1.
46 LoEVE. Probability Theory. 4th ed. Vol. 2.
47 MOISE. Geometric Topology in Dimensions 2 and 3.
48 SACHS/WU. General Relativity for Mathematicians.
49 GRUENBERG/WEIR. Linear Geometry. 2nd ed.
50 EDWARDS. Fermat's Last Theorem.
51 KLINGENBERG. A Course in Differential Geometry.
52 HARTSHORNE. Algebraic Geometry.
53 MANIN. A Course in Mathematical Logic.
54 GRAVER/WATKINS. Combinatorics with Emphasis on the Theory of
Graphs.
55 BROWN/PEARCY. Introduction to Operator Theory. Vol. 1: Elements
of Functional Analysis.
56 MASSEY. Algebraic Topology: An Introduction.
57 CROWELL/ Fox. Introduction to Knot Theory.
58 KOBLITz.p-adic Numbers, p-adic Analysis, and Zeta-Functions.
59 LANG. Cyclotomic Fields.
60 ARNOLD. Mathematical Methods in Classical Mechanics.
61 WHITEHEAD. Elements of Homotopy Theory.
62 KARGAPOLOV / MERZWAKOV. Fundamentals of the Theory of Groups.
63 BOLLOBAS. Graph Theory-An Introductory Course.
64 EDWARDS. Fourier Series, Volume 1. second edition.
65 WELLS. Differential Analysis on Complex Manifolds.
66 WATERHOUSE. Introduction to Affine Group Schemes.
67 SERRE. Local Fields.
68 WEIDMANN. Linear Operators in Hilbert Spaces.

Edwards

Uploaded by

Copyright:

Available Formats

Edwards

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Edwards

Uploaded by

Copyright:

Available Formats

Graduate Texts in Mathematics

P.R. Halmos F. W. Gehring c.c. Moore

AMS Subject Classification: 42-01

Library of Congress Cataloging in Publication Data

(Graduate texts in mathematics; 64-

ISBN-13 : 978-1-4612-6210-7 e-ISBN-13 : 978-1-4612-6208-4

All rights reserved.

No part of this book may be translated or reproduced

1967, 1979 by Springer-Verlag New York, Inc.

There are a multitude of minor corrections. In addition there are a

Chapter 1 TRIGONOMETRIC SERIES

1.1 The Genesis of Trigonometric Series and Fourier Series 1

Chapter 2 GROUP ST.RUCTURE

2.1 Periodic Functions 14

Chapter 3 CONVOLUTIONS OF FUNCTIONS

3.1 Definition and First Properties of Convolution

Chapter 4 HOMOMORPHISMS OF CONVOLUTION

4.1 Complex Homomorphisms and Fourier Coefficients 69

Chapter 5 THE DIRICHLET AND FEJER KERNELS.

5.1 The Dirichlet and Fejer Kernels 78

Chapter 6 CEsARO SUMMABILITY OF FOURIER SERIES

6.1 Uniform and Mean Summability 87

Chapter 7 SOME SPECIAL SERIES AND THEIR

7.1 Some Preliminaries 109

Chapter 8 FOURIER SERIES IN L2 130

8.1 A Minimal Property 131

8.2 Mean Convergence of Fourier Series in V. Parseval's

8.4 Factorization Problems Again 134

8.5 More about Mean Moduli of Continuity 135

8.6 Concerning Subsequences of 8 N! 137

8.7 A(Z) Once Again 139

Chapter 9 POSITIVE DEFINITE FUNCTIONS

9.1 Mise-en-Scene 148

9.3 An Alternative Proof of the Parseval Formula 152

9.4 Other Versions of the Bochner Theorem 152

Chapter 10 POINTWISE CONVERGENCE

10.1 Functions of Bounded Variation and Jordan's Test 156

10.2 Remarks on Other Criteria for Convergence; Dini's Test 159

10.3 The Divergence of Fourier Series 160

10.4 The Order of Magnitude of 8 N!. Pointwise Convergence

10.5 More about the Parseval Formula 171

10.6 Functions with Absolutely Convergent Fourier Series 173

Appendix A METRIC SPACES AND BAIRE'S THEOREM 187

A.I Some Definitions 187

A.2 Baire's Category Theorem 187

A.3 Corollary 188

A.4 Lower Semicontinuous Functions 188

A.5 A Lemma 189

Appendix B CONCERNING TOPOLOGICAL LINEAR SPACES 191

B.I Preliminary Definitions 191

B.2 Uniform Boundedness Principles 194

B.3 Open Mapping and Closed Graph Theorems 195

B.4 The Weak Compacity Principle 197

B.5 The Hahn-Banach Theorem 199

Appendix C THE DUAL OF LP (1 ~ P < co); WEAK

C.I The Dual ofLP (I ~ P < (0) 201

Appendix D A WEAK FORM OF RUNGE'S THEOREM 205

Trigonometric Series and Fourier Series

1.1 The Genesis of Tri~onometric Series and Fourier Series

2: (a" cos nx + b" sin nx),