Pee

ee SS
~

pee tere
nce aee:

ey
ed
Beg
Pe
SYN"
@

ene
Feet
 ANNI

ea~ | -[ea]mo —op)

ST. JOHN FISHER

COLLEGE
 Digitized by the Internet Archive
in 2022 with funding from
Kahle/Austin Foundation

https://archive.org/details/characteristicfuO000luka
 a

mae
CHARACTERISTIC
FUNCTIONS
 GRIFFIN BOOKS OF COGNATE INTEREST

Applications of characteristic functions .. E, LUKACS and R. G. LAHA

Discusses applications of characteristic functions to mathe-
matical statistics, using for the most part the methods of
analytical probability theory (No. 14 in Griffin’s Statistical
Monographs and Courses).
An introduction to infinitely many variates .. ne E. A. ROBINSON
Green’s function methods in probability theory a .. J. KEILSON

The advanced theory of statistics (3 volumes)

M. G. KENDALL and A. STUART
Fundamentals of statistical reasoning . . ae .. M. H. QUENOUILLE
A course in theoretical statistics aA ae ore N. A. RAHMAN

Combinatorial chance .. = F. N. DAVID and D. E. BARTON

Games, gods and gambling Bh ib es Le .. EF. N. DAVID

Complete list of books on mathematics, probability theory

and statistics available on request from the Publishers.
Cisizelio-4OM
Mhoaiboy Mie
FUNCTIONS

EUGENE “LUKACS
Professor of Mathematics
The Catholic University of America
Washington, D.C.

Second Edition
Revised and enlarged

13.5
we
he

GRIFFIN '#%° LONDON

 CHARLES GRIFFIN & COMPANY LIMITED
42 DRURY LANE, LONDON, W.C.2
Copyright © 1970
All rights reserved

No part of this publication may be reproduced, stored in a retrieval

system, or transmitted, in any form or by any means, electronic,
mechanical, photocopying, recording or otherwise, without the prior
permission of the Publishers, as above named.

First published in 1960

Second edition 1970

Large Medium 8vo, x + 350 pages

SBN 85264 170 2

Made and printed in Great Britain by

Butler & Tanner Ltd., Frome and London
 PREFACE TO SECOND EDITION

The first edition of this book appeared in the series “‘Griffin’s Statistical
Monographs and Courses’. The general aims and the arrangement of the
book remain unchanged in this second edition; however, it seemed desir-
able to expand the treatment so as to include not only recent developments
but also certain parts of the theory of characteristic functions which were
known when the first edition was written but for various reasons had to be
omitted.
The original section on stable distributions contained a number of
results which were stated without proof. The part of Chapter 5 which
deals with stable distributions has now been greatly extended; and material
-on series expansions and on the analytic properties of stable densities has
been added, together with detailed proofs. Asymptotic expansions of stable
densities and the demonstration of the unimodality of stable distributions
have also been included. Chapter 9 is entirely new and deals with infinitely
divisible distributions which have no indecomposable factor. Most theor-
ems concerning this class have been obtained during the last few years; the
first highly significant results were announced, mostly without proofs, by
Yu. V. Linnik in Doklady when the manuscript of the first edition was
almost completed. It was therefore only possible to state some of the results
in the first edition, but the present edition gives a detailed account of these
developments. Chapter 10 contains new material on a-decompositions,
while Chapter 11 treats of boundary characteristic functions; the latter
concept, introduced by J. Marcinkiewicz, has received little attention
until quite recently. The topics listed above do not exhaust the additions;
there are a number of new sections such as 3.7 (infinite convolutions),
4.4 (non-negative definite functions), and 4.5 (unimodal distributions), as
well as many changes and much supplementary matter throughout the
book.
I have received many communications from readers and friends which
I greatly appreciated and have used in this revision. I am particularly
indebted to Dr. R. Cuppens who read the manuscript of the second
edition with great care and offered many valuable suggestions.
Thanks are also due to Dr. R. Mureika, Dr. B. Ramachandran, Dr.
V. K. Rohatgi, to my wife who helped me with the proof-reading, and
Vv
vi PREFACE TO SECOND EDITION

to Mrs. A. Miller and Mrs. P. Spathelf for the accurate and speedy typing
of the manuscript.
Finally, I wish to express my appreciation to the Publishers for meeting
my wishes concerning the appearance of the book.

EUGENE LUKACS
WASHINGTON, D.C.
November 1969
 PREFACE TO FIRST EDITION

Characteristic functions and generating functions were originally developed

as a tool for the solution of problems in Probability Theory and admit
many important applications in this branch of mathematics as well as in
Mathematical Statistics. In the present monograph we study characteristic
functions for their intrinsic mathematical interest and are not concerned
with their possible applications. This statement could be taken as an
admission that this book does not properly fit into a series of statistical
monographs and courses. However, a mathematical statistician needs a
good background in pure mathematics, and the methods discussed in a
considerable part of this monograph should be contained in his tool chest.
The literature of the subject is widely scattered over many periodicals,
and a substantial part of the important results can be found in Russian
and French journals. This part of the literature is often not easily accessible
and its use is made even more difficult by the language barrier. A reason-
ably complete account of the subject is only possible if the foreign litera-
ture is adequately covered. The author hopes that such a coverage is
provided in this monograph and that this feature will increase its usefulness
for students of the subject.
Certain parts of the theory, primarily those discussed in sections 8.2
and 8.5, are the areas in which major advances are being made at present.
Important new results had been announced without proof shortly before
the manuscript was completed. It was only possible to mention some of
these advances (for example theorem 8.2.7) in order to indicate the direc-
tion in which research is proceeding at present.
Most of the theorems discussed in this book are given with detailed
proofs; occasionally it was possible only to state some results and to refer
the reader to the literature for the proofs. The monograph does not contain
a complete bibliography; we give selected references for readers who wish
either to find the proofs not given in the book or who desire to study some
background material. The book can therefore not be used to decide ques-
tions of priority. The author to whom a particular result is due is often
not even mentioned. If it is customary to associate the name of a particular
mathematician with a theorem, then this is done, but the paper in which
this result was published is not necessarily listed in the References. An
vil
Viil PREFACE TO FIRST EDITION

article is included in the list of references only if it contains material which

is mentioned in the monograph but not discussed in sufficient detail.
I should like to express my thanks to Dr. R. G. Laha, Dr. I. R. Savage
and Dr. S. Vajda for reading the manuscript. ‘The author is particularly
indebted to Dr. R. G. Laha for valuable comments and advice. Thanks
are also due to the Catholic University for providing clerical help; to Miss
I. Christensen, Dr. R. G. Laha, Mrs. E. C. Lukacs, and Mr. B. Ramachan-
dran who read the proofs; and to Mrs. K. Winner who typed the manu-
script. The book was written with the partial support of the National
Science Foundation under grant NSF—G—4220 which is gratefully acknow-
ledged here.
EuGENE LuKacs
WASHINGTON, D.C.
December 1958
 CONTENTS

Preface to second edition

Preface to first edition
1 INTRODUCTION
al Distribution Sekt ae
hs Examples of distribution functions ..
3 The method of integral transforms . .
1.4 Moments
2 PRELIMINARY STUDY OF CHARACTERISTIC FUNCTIONS
Dell Elementary properties of characteristic functions ..
2.2 Lebesgue decomposition of characteristic functions
2:3 Characteristic functions and moments
2.4 The second characteristic
3 FUNDAMENTAL PROPERTIES OF CHARACTERISTIC FUNCTIONS
The uniqueness theorem
Inversion formulae
The convolution theorem
Limits of distribution functions
The theorems of Helly
The continuity theorem
Infinite convolutions ..
4 CRITERIA FOR CHARACTERISTIC FUNCTIONS
Necessary conditions ..
Necessary and sufficient Beat ant
Sufficient conditions ..
Supplementary remarks Se aisnon- pe eure deGnice faction:
Unimodal distributions
An essential property of Shafacteriatie flncHons
5 FACTORIZATION PROBLEMS—INFINITELY DIVISIBLE CHARACTERISTIC
FUNCTIONS
Preliminary remarks on Perle cations ee
Definition of infinitely divisible characteristic fncions os
Elementary properties of infinitely divisible characteristic Pike tions
Construction of infinitely divisible characteristic functions
Canonical representations
A limit theorem :
Characteristic Rincaons of Mable Aigiributions
Frequency functions of stable distributions
Asymptotic expansions and ica Pee at ticns GE erable
densities
5.10 Unimodality OF erable Peermunonae
5.11 Self-decomposable distributions
1x
x CONTENTS

6 FACTORIZATION PROBLEMS—GENERAL THEOREMS FROM THE ARITHMETIC

OF DISTRIBUTION FUNCTIONS .. a we ae 166
6.1 Some notations and lemmas. . : : : 166
6.2 General decomposition theorems .. a ye ae ae 169
6.3. Indecomposable characteristic functions 180
7 ANALYTIC CHARACTERISTIC FUNCTIONS 191
7.1 The strip of regularity and the integral fepresentation 191
7.2 Analytic characteristic functions and their distribution factors 197
7.3 Criteria for analytic characteristic functions 210
7.4 Periodic analytic characteristic functions 225
7.5 Analytic characteristic functions as solutions of eer differential
equations 227
8 FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS : 236
8.1 Properties of the factors of an analytic characteristic function 236
8.2 Factorization of certain entire characteristic functions we 243
8.3 Determination of certain entire characteristic functions by proper-
ties of their factors : 253
8.4 Infinitely divisible analytic Ghercteristic functions 258
9 INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS WITHOUT INDECOM-
POSABLE FACTORS 262
9.1 The class Y : 262
9.2 A sufficient condition for membership of a 266
9.3 A necessary condition for membership of J 280
9.4 Infinitely divisible characteristic functions with Rounded Poiccor
spectrum : 281
9.5 ‘Theorems concerning een factorizations 290
10 -DECOMPOSITIONS 292
10.1 General theorems on «-decompositions of analyeea Gharsctemcc
functions : aH 292
10.2 Special results concerning «- decompoaitians 301
11 BOUNDARY CHARACTERISTIC FUNCTIONS 306
11.1 The integral representation ‘ 306
11.2 Infinitely divisible boundary charicienene fancuans 310
12 MIXTURES OF DISTRIBUTION FUNCTIONS AND TRANSFORMATIONS OF
CHARACTERISTIC FUNCTIONS S15
12.1 Mixtures of distribution functions : 315
12.2 ‘Transformations of characteristic Ginetions Lvs)jakeZ)

APPENDICES
Appendix A: The notations O and o
Appendix B: Schwarz’s inequality .. ‘
Appendix C: Weierstrass’ approximation theokeen
Appendix D: Order and type of entire functions. .
Appendix E: Proof of lemmas needed in Chapter 9
Appendix F: Schwarz’s reflection principle
LIST OF EXAMPLES OF CHARACTERISTIC FUNCTIONS

REFERENCES WwW
Ww
Oo
WW
WwW
WOWWwWNhd
WOPOOm
WY
ds
W Im

INDEX £
 L INTRODUCTION

It is now universally recognized that probability theory is a branch of

measure theory. Thus probability is defined as a bounded and normed
measure. An important part of probability theory is devoted to the study of
finite and measurable functions, the random variables. The probability
that the value of a random variable belongs to a given set is of interest;
this probability is a set function and is called the ‘‘probability function” of
the random variable. It is known that this set function determines a point
function—the distribution function of the random variable. Many of the
most important problems concerning random variables can be expressed
in terms of distribution functions. This part of probability theory can be
studied independently of its measure-theoretic foundation and of theorems
of purely measure-theoretic character. The methods of classical analysis
provide an efficient approach to problems of this sort.
The present monograph deals with one of the most powerful tools used
for the investigation of distribution functions, namely characteristic
functions. Before defining characteristic functions it is necessary to
summarize briefly some important concepts.

1.1 Distribution functions

In this section, in which we discuss some of the basic properties of
distribution functions, we will confine ourselves mostly to the listing of
definitions and theorems. Proofs will be only given if they are short, other-
wise the reader is referred to standard texts such as the books of Cramér
(1946) and Loéve (1963). The lack of proofs in this chapter should not
impede the reader in using this monograph. In fact, it will be possible to
follow the discussion in the subsequent chapters if one takes for granted the
statements made in this introduction. Familiarity with proofs which are
omitted is not required.
For the sake of brevity we introduce the following notation. Let F(x) be
a function of the real variable x; we then write:
F(+ c0) = lim F(x)
F(— 00) = lim F(x)
(13151) F(x+0) = limF(x +h)
P(x—0) = lim F(x—h).
2 CHARACTERISTIC FUNCTIONS

Here h | 0 means that h tends to zero from above, assuming only positive
values.
A function F(x) of a real variable x is called a distribution function if it
satisfies the following three conditions :—
(i) F(x) is non-decreasing, that is F(x+h) > F(x) if h > 0;
(ii) F(«) is right-continuous, i.e. F(v+0) = F(x);
(iii) F(+ 00) = 1, F(—o) =0.
Thus distribution functions are by definition bounded and monotone and
many of their basic properties follow from this fact. For instance a distri-
bution function can have only discontinuities of the first kind. A point «
is called a discontinuity point of the distribution function F(x) if F(«+0) =
F(x) #4F(x—0); if F(x) = F(«—0) then z is called a continuity point of
F(x). The quantity
(1.1.2) p, = F(a+0)—F(x—0) = F(x) —F(«—0)
is called the saltus (jump) of F(x) at the point «. The saltus of a distribution
function is positive at its discontinuity points and zero at its continuity
points. An interval is called a continuity interval for the distribution
function F(x) if both its endpoints are continuity points of F(x). A
point x is called a point of increase of the distribution function F(x) if
F(x+e)—F(«—«) >0 for any e > 0.
Let F(x) be a distribution function and k be a positive integer. Denote
by D, the set of discontinuity points of F(«) with saltus contained in the
half open interval

(isi al
k+1' Rk)’
this set contains at most k+1 points. The set of all discontinuity points of
F(x) is the union of all sets D,(k = 1,2,...) and is therefore at most
enumerable. We have therefore proved

Theorem 1.1.1. The set of discontinuity points of a distribution function is at

most enumerable.
Remark. It follows that the set of continuity points of a distribution
function is always dense in the set of real numbers. However, it can happen
that the set of discontinuity points is also dense. Let for example {r;}
be the enumerable but dense set of all rational numbers and assign to
r,(k = 1, 2,...) the saltus p, = 2-*. Then F(x) = D> p, is a distribution
Ty, <e
function whose discontinuity points form a dense set.
Let {x,} be the set of discontinuity points of the distribution function

(*) See E. W. Hobson (1927), pp. 301, 318.

INTRODUCTION 3

F(x) and write p,, = F(x,+0) — F(x,—0) for the saltus of F(x) at
the point x,. We define the function
ee eee
the summation is here extended over all discontinuity points not exceeding
x. The function (x) increases only by jumps, at points of the sequence
{x,}, and is constant in every closed interval not containing an x,. Such a
function is called a “‘step-function’’. The saltus of ®(x) at x, is p, . We form
also a second function
(1.1.4) y(x) = F(x)— D(x).
The function (x) is continuous, while ®(«) is only continuous from the
right; both functions are non-decreasing and satisfy the relations
O(— 0) = y(— 0) = 0,7 @(+ 0) =a, <1, y(+o)=b<1.
It is therefore possible to define two distribution functions by normalizing
@(x) and y(x).
The functions F(x) = (1/a,)®(x) and F(x) = (1/6)y(x) are both distri-
bution functions; F,(x) is a step function while F(x) is continuous for all x.
According to (1.1.4) we have
(1.1.5) F(x) = O(x)+y(x)
and therefore also
(1.1.6) F(x) = a,F{x)+bF(x) (4,20, b20, a+ = 1).
The decomposition (1.1.5) and therefore also the decomposition (1.1.6) are
unique. Suppose that there are two decompositions
F(x) = O(x)+y(x) = O,(x)+ y,(x).
Then
D(x)— D,(x) = yx(~)— (a).
The right-hand side is a continuous function while the left-hand side is
the difference of two step functions. Therefore both sides must vanish
identically and we obtain
Theorem1.1.2. Every distribution function F(x) can be decomposed according
to
F(x) = a, F{x)+bF,(x).
Here F(x) and F(x) are both distribution functions. F(x) is continuous for
all x while F(x) is a step function. The coefficients a, and b satisfy the relations
0<a,<1,0
<b <1,a,+5=1.
F(x) is called the discontinuous (discrete) part, and F(x) is called the
continuous part of F(x). We note an immediate consequence of theorem
1.1.1: There exists an enumerable set D such that | Gl (x) sole Here
D
4 CHARACTERISTIC FUNCTIONS

and in the following the integral is a Lebesgue-Stieltjes integral. The

function F,(x) is continuous but has not necessarily a derivative at all
points; however, every distribution function has a derivative almost every-
where (in the sense of Lebesgue measure). A further decomposition can be
obtained by means of more powerful tools of analysis. Applying Lebesgue’s
decomposition theorem [see Loéve (1963)] one can show that it is possible
to determine two distribution functions F,,(x) and F(x) such that
(1.1.7) F(x) = by Fu(x)
+b. F.(x)
where by = 0, b, > 0, b,+5, = Ite

The function F,,(«) can be represented as the integral of its derivative,

F(*) = ii Fy) dy. Moreover | dF (x) = 0 if Nisaset of Lebesgue

—o N

measure zero. Such a distribution function is said to be absolutely con-

tinuous. The distribution function F,(x) is a continuous function whose
derivative is almost everywhere equal to zero; moreover there exists a set

N of Lebesgue measure zero such that | dF (x) = 1. A distribution

N
function with this property is called a singular distribution. Combining
(1.1.7) and (1.1.6) we obtain the following

Theorem 1.1.3. Every distribution function F(x) can be decomposed uniquely

according to
(1.1.8) F(x) = a, F(x) + a, Fy.(x) + a3 F(x).
Here F(x), F(x), F(«) are three distribution functions. The functions
F(x) and F(x) are both continuous; however F,,.(x) 1s absolutely continuous
while F(x) 1s singular and F(x) 1s a step function. The coefficients ay, az, az
satisfy the relations a, > 0, a, > 0, a3 > 0, a,;+a,+a, = 1.
The distribution functions F(x), F,,(x), F(x) are called the discrete, the
absolutely continuous, and the singular parts, respectively, of F(x).
A distribution function is said to be pure if one of the coefficients in the
representation (1.1.8) equals 1. For pure distributions we will use the
expressions discrete distribution if a, = 1, absolutely continuous distri-
bution if a, = 1, and singular distribution if a, = 1. In applications one
almost inevitably encounters either discrete or absolutely continuous
distributions. Singular distributions are interesting from a theoretical
viewpoint but hardly ever occur in practical work. This is the reason why
statistical texts frequently refer to absolutely continuous distributions as
continuous distributions and seemingly ignore the existence of singular
distributions.
INTRODUCTION 5
Let F(x) be an absolutely continuous distribution function. Then

(1.1.9) F(x)= | F'(y)dy

where the integral is supposed to be a Lebesgue integral. The derivative
p(x) = F'(x) is called the frequency function) of the distribution F(x).
From the definition of a distribution function it follows that every function
p(x) which satisfies the conditions
p(x) > 0 for all x,
tots1 %
( % | paiax = 1

is the frequency function of an absolutely continuous distribution function

which is given by (1.1.9).

1.2 Examples of distribution functions

In this section we list a number of important distributions which occur
frequently in practical work. For the sake of brevity we refrain from
describing mathematical models which lead to these distributions.
(1) Discrete distributions. ‘The simplest discrete distribution function is
a function which has a single saltus of magnitude 1 at the origin. We
denote this function by
0 for x < 0
(1.2.1) e(x) = i: for x10.
Let & be an arbitrary fixed real number; the function «(x—§) is then a
distribution function. It has a single discontinuity point at x = £ and the
magnitude of the saltus at this point is 1. The distributions e(x—§&) are
called ‘degenerate’ (or “improper’’) distributions.
The most general purely discrete distribution function is determined by
the location of its discontinuity points and by the magnitude of the corres-
ponding jumps. Let {€;} be a sequence which contains all discontinuity
points of F(«) and let p; be the saltus at the point &;. The corresponding
distribution function is then given by
(1.2.2) F(x) = sD) Dex€,).
j
Here the &; are real numbers and the ; satisfy the relations
A228) Veh aya ae
In Table 1 we present some discrete distribution functions. The Pascal
distribution as well as the geometric distribution are particular cases of the
(*) The expression “probability density function” or “density function” is often used
for ‘‘frequency function”.
6 CHARACTERISTIC FUNCTIONS

negative binomial distribution. The Pascal distribution is obtained if 7 is

a positive integer, the geometric distribution if r = 1.
In the examples given in Table 1 the discontinuity points are consecutive
non-negative integers.
Table 1
Discrete distributions

Discontinuity Sal Conditions on

Name point & altus at ¢5 parameters

Binomial ca) ("\o ght n positive integer

j
fa 0 heir hea 0<p<i,q=1-p

Hypergeometric | &; =7 (ees) N, M, n positive integers

UNG eel
p= Olle, (*) hte
min (M, n) n N>m

Geometric by Een 5 py 0<p<i1,qg=1-p

J = > ? ie hae ,

ad. inf.

Pascal a : =) ae r positive integer

cartes > os ( q)’ Ora p< gi Pp
ad. inf.

Negative fj =j (Gyr y rreal,r > 0

binomial Ve ure 2 . p j q 0<p<i,g=41=—p
ad. inf.

Poisson i) ow) A>0

f= OA) Dens
ad. inf.
Cin

A discrete distribution is called a “‘lattice distribution’”’ if its discon-

tinuity points form a (proper or improper) subset of a sequence of equi-
distant points. We sometimes refer to the discontinuity points of such
a distribution as lattice points. They have the form a+jd (a, d constant,
d > 0;j integer). The constant d is called a “‘span’’ of the lattice distri-
bution. The distributions of Table 1 are all lattice distributions which have
the origin as a lattice point.
(II) Absolutely continuous distributions. ‘These distributions are deter-
mined by their frequency function p(x). The distribution function can be
obtained by integration:

(1.24) F(x) = i ie) dy.

INTRODUCTION ii

The function p(y) satisfies the relations (1.1.10) which correspond to

(1.2.3).
In Table 2 we list a few frequency functions.
The normal distribution is often called the Gaussian distribution. In
the French literature it is referred to as the “law of Gauss—Laplace” or
“the law of de Moivre—Laplace”’ or the “second law of Laplace’. This last
terminology is used to distinguish it from the “first law of Laplace’”’ which
is called the ‘Laplace distribution” in Table 2.

Table 2

Frequency functions

Name of : Conditions on
Distribution Frequency function p(x) parameters

Rectangular distribution 1/2r if |x-al< a, r real

(Uniform distribution) 0 if |x—al| > r>0

Triangular distribution (r—|x—al)/r? if |x-—al <r a, r real

(Simpson’s distribution) 0 if |x—-al>r r>0

Laplace distribution 4e-|2| or more generally

1
Jo exP [-—|x—p|/o] Lt, o real
= oe <= hy <= Co) a>0

Normal distribution (1/+/2m)e-**/2 or more generally

1 le 2“| a real
Ripe sabe ere x
—-o<x< 0 o>0

Student’s distribution 1 I[(@+1)/2] x2] —(n+1)/2

(with n degrees of Van U(n/2) 1+= | n > 1 integer
freedom) =e
<a KE OO

Cauchy distribution 1/[7( + x*)] or more generally oe

0 at 0, » real
[02+ («— 1)?I Colle) Ee ae Sel
Ga
Gamma distribution ee for go> @ 0, A real

0 fOr 0) (PSs Os 2\ 0)

5 ae Ds Os et reel al
Beta distribution F@
To) GIS A[ M69) p,q re
form Or—ax<—o tl
0 otherwise p>0,q>0
8 CHARACTERISTIC FUNCTIONS

Two particular cases of the gamma distribution are of interest: for

= 1 one obtains the exponential distribution, while the choice 0 = 3,
A = n/2 (n integer) yields the chi-square distribution with m degrees of
freedom which is of great importance in mathematical statistics. We note
also that the Cauchy distribution is a particular case (n = 1) of Student’s
distribution. The Beta distribution with parameters p = q = } is called
the Arc sine distribution.

(III) Singular distributions. We give only one example of a singular

distribution; this example is closely related to Cantor’s ternary set defined
over the closed unit interval V = [0,1]. The Cantor set T is constructed by
means of a step-by-step procedure. In the first step we remove from V the
open interval (4, 2). In the second step two open intervals (4, 2) and (4, $
are removed from V. From each of the remaining 4 closed intervals
the middle (open) interval of length (4)? is deleted in the third step and
this process is continued indefinitely. Let each of the (k—1) numbers
C1, Co) «++ Cy-y assume either the value 0 or the value 1; we denote by
Ace,20,...,2e,, the open interval with initial point
k=1 2 1
uC; eee
j=l1 33 : 3%

and terminal point

Our procedure consists then in removing in the kth step the 2'-! open
intervals A,, 9......,,-, of length (4)*. The Cantor set T is obtained from
the closed interval V by removing a denumerable number of open intervals.
It is easily seen that the points of T can be characterized in the following
manner. Write each number x, 0 < x < 1, in the triadic system
% = A,/3+a,/3?+ ... +4,/3°+...;3
the set T consists of all numbers x which can be written in at least one way
in the form of a triadic fraction whose digits are only zeros or twos. The
points of T therefore have triadic expansions
2€,/3+2¢,/32+...+2¢,/3"+...
WNEEE.C15:Co,.s 2.5 Cnr s.- «1a eeaquence of “0 sands 12
We introduce the function
eee ye!
Peifit= ga? 2

atx ifxe Ay,

INTRODUCTION 9
Finally we define the function
0 ix <"0)
g(x) ifmeV—-T
TS F(x) =
ee kes is
pees
Oe ras
er
1 iho 1.
It is easily seen that F(x) is a continuous distribution function; moreover its
derivative is zero at all points of V—T while it is not differentiable in the
points of 7. The points of T are the only points of increase of F(x) and
form a set of measure zero. Thus F(x) is a singular distribution function.
The singular distribution which we have just discussed has the property
that it is constant in the open intervals A... o,,_,. This is, however, by no
means typical for singular distribution functions. R. Salem (1943) gave
a simple example of a singular distribution function which is strictly
increasing.

1.3 The method of integral transforms

In this section we introduce a procedure which is very useful in the study
of distribution functions. It is frequently advisable to consider, instead
of a distribution function F(x), expressions which are derived from this
function. These expressions are usually defined as integral transforms
using a suitable kernel K(¢, x) which is a function of the variable x and
contains also a parameter ¢. The parameter ¢ can be either discrete—for
instance integer-valued—or can have a continuous domain of variation.
The integral transform is defined by

(1.3.1) I K(t, x) dF(x),

provided that the integral (1.3.1) exists as a Lebesgue-Stieltjes integral.
The conditions for the existence of this integral are of course always of
great interest.
We now list a few possibilities for the choice of K(t, x) which are useful
in the study of distribution functions.
(A) K(t, x) = x°
(B) K(t, x) = |x|"
(C) K(t, x) = « = x(x—1)...(x—t+1) with x = 1.
In the preceding three examples the parameter ¢ is restricted to non-
negative integers.
(D) K(t,x) = e*
10 CHARACTERISTIC FUNCTIONS

(EZ) KE, x) = 90
(F) Ke x) = &* wheres = 4/—1.
In cases (D), (E), (F') the parameter is a real-valued and continuous
variable.
To emphasize the discrete character of the parameter in cases (A), (B)
and (C) we write k(k = 0, 1, 2, .. .) instead of ¢t. The transforms (A), (B)
and (C) then transform the distribution function F(x) into sequences
(provided that the integrals exist). We call

Ceayt eae |” gk dF(x)

the algebraic moment of order k of F(x), or more briefly the kth moment
of F(x). Similarly

evans ell eee { \x|* dF(x)

is called the kth absolute moment of F(x). We do not intend to use the

kernel (C); it leads to the factorial moments | x dF (x).

The kernels (D), (EZ) and (F) transform the distribution function F(x)
into functions of the real variable t. The function

(13.4) M(t) = |” dF (x)

is obtained by means of the kernel (D) and is called the moment generating
function of F(x). The kernel (#) is used only when F(x) is a purely dis-
continuous distribution which has all its jumps at non-negative integer
values of the variable x. In this case we get the probability generating
function

aes) Spee |as” #dF(x) =ph& tp,

where

Pj z 0, >: P; = il.
j=0
Here p; is the saltus of F(x) at the point x = 7 (j non-negative integer).
Probability generating functions were introduced by Laplace; we will use
these functions only rarely (in Section 6.3) and mention them here mainly
because they were the first integral transforms systematically used in
probability theory.
Finally we substitute the kernel (F) into (1.3.1); this yields

(133-6) CG) ie ot dF (x),

INTRODUCTION thal

This transform is called the characteristic function) of the distribution

function F(x), and its study is the object of this monograph.
It is known [Cramér (1946) p. 70] that every bounded and measurable
function is integrable with respect to any distribution function over
(— 0c, +00). This assures the existence of the characteristic function
(1.3.6) for every distribution function. We introduced the probability
generating function only for a special class of lattice distributions. For this
class P(t) exists, provided |¢| < 1. The existence of the moment generating
function is not assured for all distributions. We say that the moment
generating function M(t) of a distribution exists if the integral (1.3.4) is
convergent for all values of ¢ belonging to an interval (finite or infinite)
which contains the origin. The existence of M(#) must be established
before it is used. Suppose for the moment that we consider a distribution
function for which M(t) exists. We will see later [Section 7.1] that it is then
connected to the characteristic function by the relation
(1:3:7) © M(t) =f (2).
We conclude this introductory chapter by discussing in somewhat greater
detail some properties of algebraic and absolute moments which will be
needed later.

1.4 Moments
We first note that moments are defined as Lebesgue-Stieltjes integrals.
It is known that for these integrals the concepts of integrability and
absolute integrability are equivalent [see Loéve (1963) pp. 119 ff.]. We
apply this general property of the Lebesgue-Stieltjes integral to the ex-
pressions defining the moments and formulate it as

Theorem 1.4.1. The algebraic moment of order k of a distribution function

F(x) exists if, and only if, its absolute moment of order k exists.
From the properties of the integral and from (1.3.2) and (1.3.3) we see
immediately that for any positive integer k
(1.4.1) Xo, = Bor

tera < |%ax-1| < Bora

while « = £, = 1.

Theorem 1.4.2. Let F(x) be a distribution function and suppose that tts
moment a, of order k exists. Then the moments «, and B, exist for all orders
Souk:

(*) Formula (1.3.6) indicates that characteristic functions are the Fourier transforms of
distribution functions.
12 CHARACTERISTIC FUNCTIONS

According to Theorem 1.4.1, 8, = | |«|* dF*(x) exists. Clearly

Be > | |e| >1

Ix|tdF(x) > | |u| >1
lx|°
dF(x) if s<b.
Since the integral | |x|dF(«) is always finite, it follows that 6, and
|| <1

a, exist for alls < R.

The moments of a distribution function do not always exist; we give now
an example for a distribution which has moments only up to a certain
order.
The function
ra) = {1 i for Gx cal
laa" Sforaix Sl
is for m > 0 an absolutely continuous distribution function. An elementary
computation shows that a, = B, = m/(m—k)ifk < m, while the moments
of order equal or greater than m do not exist.
As another example we mention Student’s distribution. This distribution
has moments up to order (n—1). For n = 1 this shows that the Cauchy
distribution does not possess any moments.
It is of interest to know whether a given sequence of real numbers can be
the sequence of the moments of a distribution. The discussion of this
difficult problem is beyond the scope of this monograph, and we refer the
reader to the book of J. A. Shohat-J. D. Tamarkin (1943). However, we
remark that a moment sequence does not necessarily determine a distri-
bution function uniquely. It may happen that two different distributions
have the same set of moments. As an illustration we give the following
example. The function

(1.4.2) p(x) = {C exp [—x* cos mz] if x > 0

0 iu w= 0
where
_a(cos ze)
~ Td/p)
and 0) < mw < 4isa frequency function. The moments of the corresponding
distribution are

ee Sida (cos jut). 7? (n = Onze .)-

Let

(1.44) (x)= ‘ee[1 + sin (= sin wn)] exp [—st ens jae] if x0
0 if x < 0.
 INTRODUCTION 13

It is known [see Pélya-Szegé (1925), vol. 1, pp. 114 and 286] that for
O<uw<4t
(1.4.5) | x” sin (x" sin ye) exp [—x" cos px]dx = 0 (n= 0,1,...).
0

This shows that p,(x) is also a frequency function and is different from
P(x), but that
SS i OD s(X) ax ae |.x" Do(x) dx

for all non-negative integers 7.

The existence of the moments of a distribution function depends on
the behaviour of this function at infinity. We give here, without proof, a
sufficient condition for the existence of moments.

Theorem 1.4.3. Let F(x) be a distribution and assume that for some integer k
1—F(x)+F(—x) = O(x-*) aS X —> 00.
Then the moments of any order s < k exist.
- For the meaning of the symbol O we refer the reader to Appendix A.
A proof of theorem 1.4.3 may be found in Cramér (1946).
Table 3
Moments

Name of distribution Formula for moments «,; and Br

Rectangular distribution =a; _ FF» eo ip

with a = 0 PYG = > Gok = OE 41? a ped

Laplace distribution Cran =O, haya S (DME Bre SS 1a

Exponential distribution an = Pr = O-* R!

Gamma distribution On = Be = O-* AA+1)...(A+R—1)

21)!
Normal distribution 3-1 = O, X%y = as = Ba;
with 2 = O0,16 —' 1
1
Box +) 1 = ——
AOE 2-1)!
( )

Student’s distribution Z ay eee EB on a CIS We?

(n degrees of freedom) aor ny 2? (n—2)(n—4).. . (n—20)
(2u < n)
te 2ene-V2 (v—1)! TI@+1)/2]
Poo-1 = (n—1)(n—3)... (n—20+1) T'(4) T(n/2)
(2u-1 < n)
14 CHARACTERISTIC FUNCTIONS

We give several inequalities for absolute moments which supplement the

relations (1.4.1). Suppose that the kth moment of a distribution exists, then
(1.4.6) Bia < Bra Bee
This inequality follows almost immediately from Schwarz’s inequality [see
Appendix B or Cramér (1946)]. By elementary operations we obtain finally
from (1.4.6) the inequalities
(L479 BIO erp
or
(14.8) pr < Bi < Bi <... < By".
For a rather wide class of discrete distributions one can obtain recurrence
relations for the moments. This class was discussed by Noack (1950) and
contains many important distribution functions such as the binomial,
Poisson and negative binomial distributions. Recurrence formulae for the
moments of singular distributions (especially for the distribution function
described in Section 1.2) were given by G. C. Evans (1957). On the pre-
ceding page we list in Table 3 formulae for the moments of a few common
absolutely continuous distributions.
 Ze PRELIMINARY STUDY OF CHARACTERISTIC
FUNCTIONS

2.1 Elementary properties of characteristic functions

In the preceding chapter we denoted distribution functions by capital
letters, as F(x), and the characteristic function of F(x) by the corresponding
small letter, as f(t). We adhere to this notation throughout this monograph;
if subscripts are used on the symbol for a distribution function then the
same subscript is attached to its characteristic function.
We defined the characteristic function f(t) of a distribution function
F(x) by
G36 /@.— |epe®
dF(sx).
The properties of characteristic functions, stated in theorem 2.1.1, follow
immediately from this formula.

Theorem 2.1.1. Let F(x) be a distribution function with characteristic

function f(t). Then
(i) f(0) =
(i) \f@| <1
(iti) f(—2) = FO.
We use here the horizontal bar atop of f(t) to denote the complex conjugate
of f(z).

Theorem 2.1.2. Every characteristic function is uniformly continuous on the

whole real line.
It follows from (1.3.6) that

fe +h)-f| < [le 1 aF(w) = 2 | |sin (wh/2)| a(x)

so that
dere ay
<2)
xh
sy =are
— +2 | sin 5: dF (xn+o[- sin a
cae?
dF (x).

We note that a right-hand side of this inequality is independent of ¢; itis

possible to make the first and the third integral on the right arbitrarily
small by choosing A > 0, B > 0 sufficiently large. Moreover, the second
16 CHARACTERISTIC FUNCTIONS

integral on the right can be made arbitrarily small by selecting h > 0

sufficiently small, so that the statement is proved.
Let F(x) be a distribution function and let a and b be two real numbers
and suppose that a > 0. Then

20 ~Gay= i
is also a distribution function.
We say that two distribution functions F and G belong to the same type
if they are connected by relation (2.1.1) (with a > 0).
More generally, we consider a distribution function F(«) and two real
numbers a and 6 and suppose only that a 4 0. We define

r(*—*) if a>0
(iia) "Gey 2 if ,
as ae) if a<0.
Then G(x) is also a distribution function. A simple change of the variable
under the integral defining g(t) shows that
(212) 5 elt) =e (ae):
Fora = —1andb = 0 wesee that g(t) = f(—?) is a characteristic function
whenever f(t) is a characteristic function.
Let a,,...,@, be m real numbers such that

a; 0; > a,=1
ia
and let F,(x),....,. (2) be iene functions. Then

G(x) = z a; F(x)
is also a distribution function; the corresponding characteristic function is

a(t) = 3 afi.
Theorem 2.1.3. Suppose that the real numbers a,,...,Q, satisfy the
conditions
n

a; 2 0, ms i 1
j=1
and that f,(t),..., f,(t) are characteristic functions. Then
n

g(t) = aa a; fi{t)
is also a characteristic function.
As a particular case we obtain the following corollary.
PRELIMINARY STUDY a7
Corollary to theorem 2.1.3. Let f(t) be a characteristic function; then f(t) as
well as Re f(t) are characteristic functions. Here

Re f(é) = 4[f(t)+ f()] = i cos tx dF (x)

is the real part of f(t).
We showed in the preceding chapter that any purely discrete distribution
can be written in the form
(122.2) F(x) = & pie(x—§)
J

where the &; are real while the p; satisfy the relations
(1.2.3) p20, Dp,=1.
j
The Lebesgue-Stieltjes integral with respect to the distribution function
(1.2.2) reduces to a sum, so that the characteristic function f(t) of F(x)
becomes
(2.1.3) f(t) = Spe.
If, in particular, F(x) is a lattice distribution then we can write
(2.1.4) &;=atyd
where a and d are real numbers. The characteristic function of a lattice
distribution therefore has the form

(215) fle) = Epc,

where the p; satisfy (1.2.3). It is immediately seen that

AG) nt eta | 254.

Every lattice distribution F(x) therefore has the following property: there
exists a real number ¢, 4 0 such that the modulus of its characteristic
function f(#) assumes the value 1 for t = tp. We show next that this
property characterizes lattice distributions. Suppose that the characteristic
function f(t) of a distribution function F(x) has this property. We assume
therefore that there exists a fj # 0 such that |f(t))| = 1. This means that
f (to) = e** for some real & or
{a ev dix) =e,
It is then easily seen that F(x) satisfies the relation

(2.1.6) | [1—cos t.(x—6)] dF (x) = 0.

Since the function 1—cos to(x—€) is continuous and non-negative, (2.1.6)
18 CHARACTERISTIC FUNCTIONS

can hold only if F(x) is a purely discontinuous distribution whose dis-

continuity points are contained in the set of zeros of the function 1—cos
to(«—&). The discontinuity points of F(«) have then necessarily the form
£+(2m/t,)s (s integer) so that F(x) is a lattice distribution. We therefore
have the following result.

Theorem 2.1.4. A characteristic function f(t) is the characteristic function

of a lattice distribution if, and only if, there exists a real ty # 0 such that
[f(to)| = 1.
Table 4 (*)
Characteristic functions

Name of distribution function F(x) Characteristic function

f(t) =(- ett? d(x)

Se)

Degenerate distribution «(x — &) eité

Binomial distribution (q+p ett)”

Negative binomial distribution x iat ita

Poisson distribution exp {A(e*—1)}

Rectangular distribution Ge =
7

Laplace distribution aes

1+??

See. : ort
Normal distribution exp [ue 5) |

e7ltl
Cauchy distribution or more generally
exp [iut—0|t]].

Gamma distribution (1£. 4 “A

iF\(p, +4, it) ea

Beta distribution(t) (ee)
T(p+4q) x T(p +)
Mp) Motaqt+prGg+i ie

(*) The conditions on the parameters are stated in Table 2 for discrete distributions
and in Table 3 for absolutely continuous distributions. We denote here e4 by exp [A].
(t) 171 (@, 6, 2) is the confluent hypergeometric function.
PRELIMINARY STUDY 19

Theorem 2.1.4 implies that |f(z)| < 1 almost everywhere, provided that
f(#) is the characteristic function of a non-degenerate distribution. Suppose
that a < 0, then [f(¢)]* cannot be a characteristic function. This remark
leads to the following corollary.
Corollary 1 to theorem 2.1.4. The only characteristic functions whose
reciprocals are also characteristic functions belong to degenerate distributions.
We obtain easily from theorem 2.1.4 the following corollary:
Corollary 2 to theorem 2.1.4. If a characteristic function f(t) has the
property that for two incommensurable real values t, and t, the relations
[f(t)| = |f(t2)| = 1 hold, then |f(t)| = 1.
On the preceding page we listed in tabular form the characteristic
functions of some of the distributions given in Tables 1 and 2.
In Section 1.2 we constructed [see (II1)] a singular distribution function.
It can be shown that the characteristic function of this distribution is given
by
(2.1.7) f(t) = e/? lim TI cos z
n—>o j=1

For details and further examples of characteristic functions of singular

distributions we refer the reader to papers by B. Jessen—A. Wintner (1935),
A. Wintner (1936), R. Kershner (1936), and the thesis of M. Girault (1954).

2.2 Lebesgue decomposition of characteristic functions

The decomposition theorem 1.1.3 induces immediately a decomposition
of the characteristic function. Every characteristic function f(t) can be
written in the form
(2.2.1) f(t) = & falt) +42 fac(t) + 4 f.(2)
with a, > 0, a, > 0, a, > 0 and a,+a,+a; = 1. Here f,(t), f,,(t) and f,(¢)
respectively are characteristic functions of (purely) discrete, absolutely
continuous and singular distributions. For a pure distribution one of the
coefficients @,, a2, a3 is equal to 1 while the other two are zero. We add a
few remarks concerning characteristic functions of a pure type.
(A) If a, = 1 then f(2) is the characteristic function of a (purely) discrete
distribution. It follows from (2.1.3) that f(¢) is almost periodic™) so that
lim sup |f(¢)| = 1.
|t|—>o0
(B) If a, = 1 then f(t) belongs to an absolutely continuous distribution. It
follows from the Riemann—Lebesgue lemma [see Titchmarsh (1939) p. 403]
that
lim f(é) = 0.
|t|—>00

(*) See H. Bohr (1932), (1947).

20 CHARACTERISTIC FUNCTIONS

(C) If a; = 1 then f(2) is the characteristic function of a singular distribu-

tion. In this case f(t) does not necessarily go to zero as |z| tends to oo. Thus
L = lim sup |f(é)|
|t]—>00

may be any number between zero and 1. In fact, examples are known where
L assumes the value zero [Girault (1954)]. An example) of a characteristic
function of a singular distribution for which L = 1 was communicated to
the author by A. Wintner. Another example is due to C. G. Esseen (1944).
Singular distributions for which L equals 1 or 0, or a number between 0 and
1, were given by L. Schwartz (1951).
The behaviour of the characteristic function at infinity permits therefore
some inference concerning its type. For instance, if
lim sup [f()| = 0,
|t]—>
00

then f(t) belongs to a continuous distribution.

2.3 Characteristic functions and moments

There is a close connection between characteristic functions and moments.
In order to discuss this relation we introduce the following notation.
Let h(y) be an arbitrary function; we define its first (central) difference
with respect to an increment ¢ by
Arh(y) = Mh(y) = h(y+t)—h(y—t)
and the higher differences by
Agri h(y) = AA, Hy),
for k = 1, 2,.... It can be shown by induction that

(2.3.1) Aih(y) = d (-1) (i)hl y +(n—2k)¢].

k=0
In particular, for the function h(y) = e*” we have
(2.3.2) Abe te eH Ser) re 27 cin xt],
Theorem 2.3.1. Let f(t) be the characteristic function of a distribution
function F(x), and let
Ax f(0)
(2t)?*

eo

(*) Let f(t) =

II cos (t/n!); it follows from a result of B. Jessen—A. Wintner (1935)
ee
[theorem 11] that f(z) belongs to a purely singular distribution. Moreover, for integer
oe)

k it is easily seen that 1—f(27k!) = O( 2 k!#/n!*) = o(1) as k-> ©, so that

n=k+1 a
ae sup |f(t)| = 1. (For the meaning of the symbols O and o see Appendix A.)
0
PRELIMINARY STUDY 21

be the (2k)th (central) difference quotient of f(t) at the origin. Assume that
(As f(0)
M= Lienging (28)
Then the (2k)th moment a, of F(x) exists, as do all the moments «, of order
s < 2k. Moreover the derivatives f(t) exist for allt and fors = 1,2,...,2k
and

POO) =? \- we aye. (gant o2, 2, ZR)

so that
ees ers FEO);
From the assumptions of the theorem it is seen that there exists a finite
constant M such that

(2.3.3) 2 inf (Ad

lim aef(0)
It follows from (2.3.2) that

Ad f(y) = |” ¢M*(2i sin xt)

dF(x).
The difference quotient at the origin is then
Ai, f(0) | (= oF
= F(x).
(22) Wa arpara
We see therefore from (2.3.3) that

M = lim inf (= es) dF (x)

oo 1 2k

+0 —0 t
and hence that
b
sin xt 2k b
M > lim inf ( ) dF (x) = | x dE'(x)
t>0

for any finite a and 6. It follows then that the (2k)th moment

He fe x dF (x)
exists and that M > «»,. Let s be a positive integer such that s < 2k; then
it follows from theorem 1.4.2 that the moments f, and «, exist for s = 1,
eens oR
From the existence of the moments «, (s = 1, 2,...,2k) we see imme-
oO

diately that | x* e dF'(«) exists and converges absolutely and uniformly

—o
for all real t and s < 2k. It follows from a well-known theorem [see for
instance Cramér (1946), pp. 67-68] that all derivatives up to order 2k exist
22 CHARACTERISTIC FUNCTIONS

and are obtained by differentiating under the integral sign. This proves
theorem 2.3.1.
If a characteristic function has a derivative of even order, then the con-
ditions of theorem 2.3.1 are satisfied and we obtain

Corollary 1 to theorem 2.3.1. If the characteristic function of a distribution

F(x) has a derivative of order k at t = 0, then all the moments of F(x) up to
order k exist if k is even, or up to order k — 1 if k is odd.
The following example, due to A. Zygmund (1947), shows that the
result cannot be a
betes <
j=2 J”
ae
Then

Poa) = C71 Ssa1 lele—) +l +a)

is a distribution function re ies function is
ces cos jt
1) z J? log7.
j
It can be shown that f’(t) exists and is continuous for all values of ¢, in
particular for t = 0. However, the first moment of F(x) is infinite.

Corollary 2 to theorem 2.3.1. If the moment «, of order s of a distribution

function F(x) exists then the characteristic function f(t) of F(x) can be differ-
entiated s times and
Pt) = 2 la x! 8 dF(x).
Corollary 2 follows immediately from the argument used in the last part of
the proof of theorem 2.3.1.
Theorem 2.3.1 also yields another result:

Theorem 2.3.2. Let f(t) be the characteristic function of a distribution F(x)

and assume that for an infinite sequence of even integers {2n,}
re eng (U))
M,, k = limeaeinf (2t)?"*
|—*

is finite (but not necessarily bounded) fork = 1,2,... Then all moments «, of
the distribution function F(x) exist and f(t) can be differentiated for real t any
number of times, with
fot) = 38 iB xt dF(x).
Corollary to theorem 2.3.2. Let f(t) be the characteristic function of F(x); if
all the derivatives off(t) exist at the origin then all the moments of F(x) exist.
PRELIMINARY STUDY 23
It is worthwhile to note that a characteristic function may be nowhere
differentiable. As an example we mention the Weierstrass function

ite= ee ioe
etd ;

This is the characteristic function of the purely discrete distribution

>

Px 3 on
e(x-—5*).

Let us assume that the first n moments of a distribution function F(x)

exist and denote by f(t) its characteristic function. Then f(t) has the
Maclaurin expansion

fea FO
E H+ Rl
where (*)

R(t) =LO)
“ +o(1) ast +0.

It follows from theorem 2.3.1 that

{O= 5
X= ib + ole") as t—>0.

This connection is eee described in the following manner:

Theorem 2.3.3. Let F(x) be a distribution function and assume that the nth
moment of F(x) exists. Then the characteristic function f(t) of F(x) admits the
expansion

(234) ft) =14+De(it+o() as 10.

Conversely, suppose that the characteristic function f(t) of a distribution F(x)
has an expansion (2.3.4). Then the distribution function F(x) has moments up
to the order n if n is even, but only up to the order (n—1) tf nis odd. Moreover
Cp 0/7! for; = 1,-2.50 of nm ts even but only forj = 1,2,...,(n—1) ifn
is odd.
We have already established the first part of the theorem. To prove its
second part we compute A} f(0) according to formula (2.3.1) using the
assumption (2.3.4). This yields an expansion in powers of ¢; it is easily seen
that the constant term of this expansion vanishes and one obtains

(*) The remainder term used here is a modification of Lagrange’s form for the re-
mainder. This follows from a seldom used form of the remainder term [see Hardy (1963)
lols 290]
B
24 CHARACTERISTIC FUNCTIONS

(2.3.5) AL f(0) = Sic,

Ajt +o(t") gaa
j=1
0 if j+n is odd
where A; =
> (—1)' (j,)e0—2hy if j +n is even.
k=0

We can prove by induction that

0 a See |
(2.3.6) we LF (i) i ve 1)"n! if s = n.
From (2.3.5) and (2.3.6) one sees easily that A; f(0) = 7"c,, 2” 2" n! + o(t”) so
that
(237 SSO) = ¢,n!+0(t”) ast
> 0.
(2t)"
If m is even we can conclude therefore from theorem 2.3.1 that the moment
of order n of F(x) exists. If m is odd we see that the validity of (2.3.4)
implies the possibility of an expansion

ft) =1 + Se,(it)) +0(t"-2) ast 30.

The argument just used proves therefore that the moment of order (n— 1)
of F(x) exists. The last part of the theorem follows immediately from
(2:3,7)):
An expression for the remainder term in formula (2.3.5) was given by
E. J. G. Pitman (1961).
To illustrate the situation described by theorem 2.3.3 we give an example,
due to A. Wintner (1947).
It is easily seen that the function
0 if |x |< 2
(2:3.8) 5, sp(e)== C Teen eee)
x? log |x|
is a frequency function, provided that C is determined so as to make

ie p(x)de = 1. From
aa
i eae = log log A—log log2

it follows that the distribution determined by (2.3.8) does not have any
moments. The characteristic function f(t) of p(«) is given by the integral

fi) =20 |2 x? logx

cos tx
 PRELIMINARY STUDY 25
Then
HaiiOley| flotaks ai fageericssk (e 1—cos tx
20) 2 x logx de x? log x it x* log x iH
so that 1 — f(t) is a real, non-negative and even function of t. For any real z
one has 0 < 1—cosz < Min (2, z?) so that [1—/f(t)] has as a majorant a

oft deta ome

constant multiple of
hag doe 2 Whdx
; iowe fi dil iogs O(—t/log t) = o(t) ast 0.

Thus f(#) = 1+ 0(t) admits an expansion of the form (2.3.4) with n = 1,

c, = 0, even though the first moment does not exist.
If the moments «, of F(x) exist for all orders n and if
lim sup (|, |/n!)¥"= L
is finite, then the characteristic function f(t) of F(x) is regular at the origin
and has the power series expansion

fi) = Bay
and p = 1/L is the radius of convergence of this series.
It is possible to define symmetric moments

(2.3.9) (s2),= tim|" x F(a),

Symmetric moments may exist for distributions which do not possess
moments; the connection between the existence of symmetric moments
and symmetric Ath derivatives
Ai — f(0)
lim
to (22)"
of the characteristic function was investigated by A. Zygmund (1947). We
mention here only the simplest case:

Theorem 2.3.4. Suppose that the characteristic function f(t) of a distribution

function F(x) satisfies the “smoothness condition”A }f(0) = o(t) as t > 0, then
a necessary and sufficient condition for the existence of f'(0) 1s the existence of
the symmetric moment of order 1, and

(sx), = lim [" 2dF (x) = if (0).

Zygmund’s “smoothness condition” is expressed in terms of the
characteristic functions. E. J. G. Pitman (1956) replaced it by a condition
on the distribution function.
26 CHARACTERISTIC FUNCTIONS

R. P. Boas (1967) studied the related problem of the behaviour of a

distribution function whose characteristic function satisfies a Lipschitz
condition of order « (0 < « < 1).
It is also possible to express the absolute moments of a distribution
function F(x) in terms of its characteristic function f(z). To do this we use
the well-known fact that
i} A sin ux +1 ifu>0
(2.3.10) * Jim —— dx = sgnu = 40 if a =20
aaa as -1 ifu<0
Since absolute moments of even order are identical with the algebraic
moments of the same order, we have only to consider absolute moments of
odd order. Let 7 be an odd integ-r, then

age |OV dROy

3
|Momiganuiae
(eo) A .
| u’ tim \ eta ds|dF (u)
TJ —@ Acad —-A XxX

(ca) A y

Beh |— CO) |— 74 uf Xx de dF (u)

TU A> co

ie {i “|
ae if
‘
Ue —
:Wwe Fi
|
| — ies

ss Sve) 28 & ee

= 55), Fre @-Car (aS

—

so that

2311) Bases] LM mayne.

2.4 The second characteristic
We have shown that every characteristic function f(t) is continuous and
that f(0) = 1. Therefore there exists a neighbourhood of the origin in
which f(t) is different from zero; let |t| < A be this neighbourhood. The
function ¢(t) = log f(é) can be defined uniquely for |t| < A, provided we
understand by log f(z) the principal branch of the logarithm of the charac-
teristic function, i.e. that determination of log f(t) which is continuous and
vanishes at ¢ = 0. The function ¢(t) is called the second characteristic of
the distribution function F(x). We will consistently denote the second
characteristic by the Greek letter corresponding to the small letter used for
the characteristic function.
Let usagain assume that the first 7 moments of F(«) exist. We obtain then
from the Maclaurin series for log (1 +z) and from (2.3.4) the development
PRELIMINARY STUDY 21

(2.4.1) #(t) = DF) +04 i") ast > 0).

The coefficients «,; of this expansion are called the cumulants (or semi-
invariants) of F(x). Clearly
(2.4.2) Kp = 1-16 (0),
On account of the relation (2.4.1) one sometimes calls the function 4(¢)
the cumulant generating function of F(«). This terminology is however
somewhat awkward since ¢(¢) exists (in |¢| < A) even if cumulants and
moments do not exist. For this reason we prefer the name ‘second
characteristic’ used in the French literature.
The relations between cumulants and moments can be found easily by
means of Faa di Bruno’s formula [see Jordan (1950), pp. 33-34]. This
formula gives an explicit expression for the pth derivative of a function of
a function. Suppose that the moment «, of F(«) of order m exists; then we
get

(243) (k-Iip!
= y (1)! i! (Ry!)°.. ty! (Rg!)
and
$4 p!} A is
Catz Yi ery a mn)
a
for p = 1, 2,...,”. The summation is extended over all partitions of p
which satisfy
ty)tig+...+1, =k
1, ky t+igkgt+...+1,k, = p-
 3. FUNDAMENTAL PROPERTIES OF
CHARACTERISTIC FUNCTIONS

In Sections 3.1-3.6 we discuss the most significant theorems which des-

cribe the connection between characteristic functions and distribution
functions. These properties account for the importance of characteristic
functions in the theory of probability.

3.1 The uniqueness theorem

Theorem 3.1.1. Two distribution functions F(x) and F,(x) are identical 1f,
and only if, their characteristic functions f,(t) and f,(t) are identical.
We see immediately from (1.3.6) that the identity of the distribution
functions F(x) and F(x) implies the identity of their characteristic
functions; therefore we must only prove the converse proposition. Suppose
therefore that the characteristic functions of the distributions F(x) and
F(x) satisfy the relation
3.1.1) fi) =f2(2).
We denote F, (x)
— F,,(*) by B(x) and write (3.1.1) in the form

(3.1.2) | eap(x) ik
The function f(x) is the difference of two monotone increasing functions
and is therefore a function of bounded variation. Moreover we see from
(3.1.2) that B(x) satisfies the relation

(3.1.3) |mo dp(x) = 0

for all functions h(x) = e“” where t is an arbitrary real constant. Therefore
(3.1.3) also holds for any trigonometric polynomial

(3, 4)" Tii(2) = oyeanen

where 4 is an arbitrary real constant. The relation (3.1.3) is therefore also

valid for any function which is the uniform limit of trigonometric poly-
nomials (3.1.4).
We conclude then from Weierstrass’ approximation theorem [see
Appendix C] that (3.1.3) holds if h(x) is a continuous periodic function.
Let g(x) be a continuous function which vanishes outside a fixed bounded
interval J and choose m > 0 so large that the half open interval (—m, m]
FUNDAMENTAL PROPERTIES 29
contains J. We then define /,,(«) as a continuous periodic function of
period 2m such that h,, (x) = g(x) for —m < x < m. Then (3.1.3) holds
for the function h,,(x). Since B(*) is a function of bounded variation
it is possible to choose m so large that the variation of B(x) for |x| > m be-

comes arbitrarily small; the integral |: h,,(x) dB(x) therefore approaches

| &(x) dB(x) as m tends to infinity. Hence

[7 (0) 4500) = fFa(e) a5(~) = 0

for every continuous function which vanishes outside a fixed interval J. It
follows easily from the uniform boundedness of g(x) that

[s) ap(a) = 0,
provided that a and 6 are continuity points of §(«) and that g(x) is contin-
uous for a < x < b. But then f(x) must be constant on the set of its con-
tinuity points so that F(x) and F(x) must agree in all continuity points
and are therefore identical.
We would emphasize here that two characteristic functions f,(¢) and
Jo (t) must agree for all values of t in order to assure that the corresponding
distribution functions F(x) and F(x) should be identical. This require-
ment can only be weakened in a trivial way: one could suppose that the
functions f, (t) and f, (¢) agree for t-values which form a set which is dense
on the positive real axis. It follows then from theorem 2.1.1 (condition 111)
that they must agree for a set dense on the whole real axis, and one can
conclude from theorem 2.1.2 that f,(t) = f,(¢). Agreement over a finite
interval is, in general, not sufficient for the identity of the correspond-
ing distribution functions. In fact, it is not difficult to construct a pair of
characteristic functions which belong to different distributions and which
agree over a finite interval. It is also possible to show that a pair of different
characteristic functions can agree everywhere with the exception of two
symmetrically located intervals. We will give these examples in Section 4.3.
Let F(x) be an arbitrary distribution function. It is then easily seen that
the function 1— F(—x—0) is also a distribution function. This function is
called the conjugate distribution of F(x) and is denoted by
(3.1.5) F(x) = 1-F(—x—0).
Let f(t) be the characteristic function of the distribution F(x); an elemen-
tary computation shows that the characteristic function of the conjugate
distribution F(x) is

(3.1.6) |- ef”dF (x) = f(—1) = fb.

30 CHARACTERISTIC FUNCTIONS

A distribution function is said to be symmetric if it is equal to its conjugate.

The following characterization of symmetric distributions is easily
established.

Theorem 3.1.2. A distribution function is symmetric if, and only if, its
characteristic function 1s real and even.
The necessity of the condition follows from (3.1.6) while the sufficiency
is a consequence of the uniqueness theorem.
Moreover, we see from (3.1.6) that

(inh aaiOX= i cos tx dF (x).

This formula can be used to establish the following property of sym-
metric distributions.

Theorem 3.1.3. Let F(x) be a symmetric distribution with characteristic

function f(t) and suppose that the moments «a; (j = 1,2,...,2k) of F(x)
exist. Then

f2-(t) = (-1) ivesMGI LO Eh) deme eee

f(t) = (= 02d Gos\txdi(2) aE ake

The theorem is a consequence of formula (3.1.7) and the corollary to
theorem 2.3.1.
Corollary to theorem 3.1.3. Let F(x) be a symmetric distribution with char-
acteristic function f(t) and suppose that the moments «;(j = 1,..., 2k) of
F(x) exist. Then
OP tie pS) (fe aleect)

pO
8(-tVlay F128
(2j-1) t . ;

2(-1)+1aa lim iv
[® 024
4, Sim?
a(tx/2 dF (x)
(23) y= (23)
(iii) lim ro J wo)=
t>0 t t0
—1)i+1 ;
( sah e) (PH 1; 27.5 ee

The corollary follows immediately from theorem 3.1.3.

3.2. Inversion formulae

Theorem 3.1.1 of the preceding section establishes a one-to-one cor-
respondence between characteristic functions and distribution functions.
However, theorem 3.1.1 does not give a method for the determination of
FUNDAMENTAL PROPERTIES 31

the distribution function belonging to a given characteristic function. The

theorems discussed in the present section deal with this problem.

Theorem 3.2.1 (the inversion theorem). Let f(t) be the characteristic function
of the distribution function F(x). Then
r Lae (My .
(322-1) F(a+h)— F(a) = lim : | it e
= f(t) at ?
f( )
Posey Lit —T

provided that a and a+h (with h > 0) are continuity points of F(x).
For the proof of the inversion theorem we need the following well-
known lemma.

Lemma 3.2.1. The integral | (sin y)/y dy is bounded for x > 0 and
0
approaches x/2 as x tends to infinity.
The first part of the lemma is easily proved by dividing the range of
integration into segments of length z; the second statement is obtained by
contour integration. (*)
Let
Des 13 re

TENG oF 3) Fp nee OX
From this definition and from lemma 3.2.1 it is seen that A(h, T) is
bounded for all # and all T and that
hed =A)
while
Lifh>0
(3.2.2) lim Ah, T)=<{ Oifhk =0
eh thee:
We now introduce the integral
1 (2 emia e-itlath)
Ue Le a7 f(t) dt.

We substitute here for f(t) = i e” dF(x) and note that the absolute

value of the integrand does not exceed h. Hence the order of integration
may be reversed and one obtains
1 fo} Tt e7 tla- 0) == e7 tlath- 2%) |
ah dt |dF (x).
Jr 2a |= 00) |—? it (*)

(*) See Titchmarsh (1937), Section 3.122.

RyA CHARACTERISTIC FUNCTIONS

We replace the exponentials in the inner integral by trigonometric func-

tions and obtain
CO are ailsohpee t(«—a)—sin t(xn—a—h) at|dF (x).
IU 0 t
We write
g(x, T) = A(x—a, T)—A(x—a—h, T)
and conclude from the boundedness of A(x, T) and from (3.2.2) that
g(x, T) is also bounded for all « and all T and that
O if x.<74
sifx=a
(32:4) lime (x, Ps ol if a eta
Tc °
4ifx = ath
Oifx > ath.
Moreover, (3.2.3) can be written as

Gas y. = I a(x, T) dF (2).

Let ¢ > 0 be an arbitrary positive number. We assumed that a and a+h
are continuity points of F(x) and we have shown that g(x, T) is bounded.
Therefore it is possible to select a 6 > 0 which is so small that the three
inequalities

(L) jf as T)aF Qe)

a+6
Se

G.) fats 7) are)

a+h+6
Sy

(I) |kay |- dF (x) <6

a+6

hold simultaneously.
Moreover, we conclude from (3.2.4) that T can be chosen so large that
the relations

(y) [fle T)aR@)- |" dre)

ath—6 a+h—6
<S is
a+6 a+6
an

G) [fees Tar@j4[[" 66, TAF)

a—6o | i)
<eé
—« ath+6
are both satisfied. We decompose the range of integration of the integral
(3.2.5) into the five intervals (— 0, a—d), (a—6, a+6) (a+6, at+h—6),
(a+h—6, a+h+6) and (a+h+0, o) and obtain, using (1,), (I,) and (I;),

i) ommealedBa |is o(x, T) dF (x) ee

FUNDAMENTAL PROPERTIES 33
We see then from (I,), (I,) and (1,) that
at+h
Jn—| dF) ae

if Tis sufficiently large. This is equivalent to (3.2.1), so that the proof of

the inversion theorem is completed.
An inversion formula expressing F(x) [instead of its increment
F(x +h) —F (x)] in terms of an integral involving the characteristic function
of F(x) was given by J. Gil-Pelaez (1951).
We remark that the uniqueness theorem could easily have been obtained
from the inversion theorem. Such an approach would have shortened the
presentation; however, it is of some methodological interest to distinguish
between these two theorems.
The inversion formula can be written in a more symmetrical form by
putting a = x—6 and h = 20. In this way we obtain

Ge Fre
by Fig 8) Sli | sin16 we £(4)dt
Wy .

To tJ—T

provided that x—6d and x+6 are continuity points of F(x). The last
formula can also be written as
F(x+0)—F(x—6) . 1 (7 sindd
= lim — e-H F(t) dt.
26 T->0 21 —T t6

Let us now assume that f(t) is absolutely integrable over (— 00, + 00),
the integrand is then dominated by the absolutely integrable function f(t)
and converges to e~ f(f) as 6 tends to zero. Hence one may go to the limit
under the integral sign; one sees that F’(x) exists for all x and one
obtains the following result.

Theorem 3.2.2. If a characteristic function f(t) is absolutely integrable over

(— «, +) then the corresponding distribution function F(x) is absolutely
continuous and the formula

pe) = fr’ =Fa@y=

1 |tat oe
— | e* FO)
e ™ f(t) dt

expresses its density function p(x) in terms of the characteristic function. The
density function p(x) 1s continuous.
We have already proved the absolute continuity of the distribution
corresponding to f(t) and must still show that p(x) is continuous. We see
easily that
|p(x+h)—p(a)| <=[ [sin (eh/2)|1 FL ae
a |pea [Sin (th/2) FO Lat
34 CHARACTERISTIC FUNCTIONS

We choose A so large that the second integral becomes arbitrarily small and
can then make the first integral as small as we wish by selecting h sufficiently
small. This completes the proof of the theorem.
We note that this inversion formula is here not derived for all absolutely
continuous distributions but only for absolutely continuous distribution
functions which have absolutely integrable characteristic functions. We
will give later [p. 85] examples of characteristic functions which belong to
absolutely continuous distributions but which are not absolutely integrable.
However, we will see that under certain conditions the inversion formula is
still valid.
It is also possible to derive an inversion formula which is valid for
arbitrary absolutely continuous distributions, even if they are not absolutely
integrable. As the proof of this formula requires results of the next section,
we give it in Section 3.3 (see corollary 3 to theorem 3.3.2).
Let again f(t) be the characteristic function of an arbitrary distribution
F(x). We consider the integral

CD Tie sa (ake fo dt.

We write in (3.2.7) f(t) = | e dF (z) and see easily that the order of

the two integrations may be exchanged, so that

ae | sin 7A)

sin T'y sin Ty

= d,F (y+ | ;
la 1B es ey wi>n Ty EMG Sd
Here h is a positive number to be chosen later.
Let ¢ > 0 be an arbitrary positive number and denote the saltus of F(x)
at the point x by p,, that is p, = F(x)—F(x—0). It is then possible to
choose a value fy > 0 which is so small that the inequality
(3.2.8) | en deF+8)~Be <E
holds.
The integral
sin Ty
dF x
be Ty e (y+ )

converges to zero as T tends to infinity, since the integrand is dominated

by +1 and converges to zero as T > o. We can therefore choose T so
large that

629)
o29 | |[.i, Re Fts)| <e
sin Ty
d,F x)|
|
< €.
FUNDAMENTAL PROPERTIES 35
Let now h, be a number such that
(3.2.10) Age heei0,
It follows from (3.2.8) that
Pu-ێ < |
|u| <Ay
aF(y+a)< | lul<No
dyF(y+a) < pete
so that

J pa, OF O92
|u| <hy
<e

(3.2:11) while

0< | dy Biche), 28
hy <|y|<ho

We are still free to choose /,, subject only to the restriction (3.2.10). We
select h, so small that
e sin Ty
Ty
for |y| < h,. According to the law of the mean there exists a real 6(|0| < 1)
such that
sin Ty sin T6h, |
d,F =
fie Ty 2 2 ty) Toh, lul<hy getty)

=(1-0)| dy F(w+y)
lul<hy
where 0 < 6, < 1. From this we see immediately that
| sin ED ACR = | d, F(x+y)—py—O2¢
<a Ty lvl< hy
with 0 < 0, < 1. Hence we conclude from (3.2.11) that
sin Ty
(Bed, LZ ) |lul<tn Ty d,F
y (x +y) = P Palas 22.é

Since | iy < 1, we see from the second formula (3.2.11) that

sin Ty |
d,F < d,F ee.
(vee Tyee Oma) th <Wiche (ye)
We combine the last inequality with (3.2.9) and (3.2.12) and obtain
(322513) |Ir—pe| < 5e
if T is sufficiently large. Thus we have proved the following statement:
Theorem 3.2.3. Let f(t) be an arbitrary characteristic function. For every
real x the limit
== NI
ile 7 fia
ae
a limp _,¢ We)
36 CHARACTERISTIC FUNCTIONS

exists and is equal to the saltus of the distribution function of f(t) at the
point x.

Corollary 1 to theorem 3.2.3. The characteristic function of a continuous

(singular or absolutely continuous) distribution cannot be an almost periodic
function.
Let f(t) be the characteristic function of a continuous distribution. We
see immediately from theorem 3.2.3 that
sal a
(3.2.14) lim 57)_,e-# F(t)
dt = 0
for all real x. We give an indirect proof for the corollary and suppose that
f(t) is almost periodic. Formula (3.2.14) means then that all Fourier
coefficients of f(t) vanish; we conclude from the uniqueness theorem
for almost periodic functions [H. Bohr (1932), (1947), English edition,
p. 60] that f(z) = 0, which contradicts the assumption that f(t) is a
characteristic function.

Corollary 2 to theorem 3.2.3. A distribution function is purely discrete if,

and only if, its characteristic function is almost periodic.
The necessity of the condition follows from our remarks in Section 2.2.
To prove the sufficiency we note that an almost periodic characteristic
function f(t) necessarily satisfies the relation
lim sup |f(¢)| =
|t]—>co

in view of corollary 1 to theorem 3.2.3 f(#) must then belong to a purely

discrete distribution.

3.3. The convolution theorem

We consider next two distribution functions F,(«) and F(x) and their
characteristic functions f,(¢) and f,(t) respectively. We form the function

(3.3.1) F(z) = ie F,(2—x) dF, (x).

The function F, (z—«) is bounded so that the integral (3.3.1) exists. More-
over it is easily seen that F(z) is a distribution function; this follows from
the fact that the necessary passages to the limit can be carried out under the
integral sign.
We wish to determine the characteristic function f(t) of F(z). Clearly,

f(t) = le tt dF (2).
b
We consider first the integral | e dF (2) (where a and b are finite) and
FUNDAMENTAL PROPERTIES 3¥
write it as a limit of generalized Darboux sums.*) We use a sequence of
subdivisions {z‘} of decreasing modulus
Ce eee ee SR oe
of the closed interval [a, b]. In this manner we get
b n
| exp (izt)dF(z) = lim © exp (itz) LF (z,;) — F(2")].
a no j=1

In view of (3.3.1) this can be written as

| exp (izt) dF (z)

b

= lim > exp [it(25) — x)] LF, (2,0, —x) — F, (2 —x)] exp (ttx) dF, (x).
noo —o j=1

From this it follows that

b co b=—@
| edi (3) = | il et” dF, (») |e dF. (x).
We take the limit as a > — «© and b > o and obtain
(33.2) ff) =AOf().
Suppose conversely that a characteristic function f(t) is the product
(3.3.2) of two characteristic functions. According to the uniqueness theorem
relation (3.3.1) holds between the corresponding distribution functions.
Formula (3.3.1) defines an operation between distribution functions; it
indicates how a new distribution F can be obtained from two given distri-
butions F, and F;. This operation is called convolution (sometimes com-
position or Faltung) and F is called the convolution of F, and F, and is
written as a symbolic product
F = Ral dee

It is seen from (3.3.2) that the convolution is a commutative and associative

operation.
We summarize our results as

Theorem 3.3.1 (Convolution theorem). A distribution function F ts the con-

volution of two distributions F', and F., that ts

FQ) = | Fi@-a)dFa(e) 2 i F,(2z—2)dF,(x) = F.*F,

if, and only if, the corresponding characteristic functions satisfy the relation
ft) =A@®A@:
Hence the genuine multiplication of the characteristic functions and the

(*) See Cramér (1946), pp. 62, 72.

38 CHARACTERISTIC FUNCTIONS

symbolic multiplication of the distribution functions correspond to each

other uniquely.
The following corollaries follow almost immediately from the convolu-
tion theorem.

Corollary 1 to theorem 3.3.1. The product of two characteristic functions 1s

a characteristic function.

Corollary 2 to theorem 3.3.1. If f(t) is a characteristic function, |f(t) |? 1s

also a characteristic function.
This follows from theorem 2.1.1 and formula (3.1.6).
We mention some properties of the convolution operation which follow
easily from its definition.

Theorem 3.3.2. Let F = F,*F, be the convolution of two distributions F,

and F,. If one of the components of F is a continuous distribution, then the
symbolic product ts also a continuous distribution. If one of the components of
F is absolutely continuous then F is also absolutely continuous.
Remark 1. Let F = F,*F, be the convolution of two distributions F’,
and F’, and suppose that F is a discrete distribution. ‘Then both components
F, and F, are also discrete distributions.
Remark 2. It is, however, not possible to conclude from the assump-
tions (i) F = F,*F,, (ii) F is absolutely continuous, that at least one of the
distribution functions Ff, and fF, is absolutely continuous. This will be
seen from a representation of the rectangular distribution as the convolu-
tion of two singular distributions. This example will be given on page 189.

Corollary 1 to theorem 3.3.2. If f(t) ts the characteristic function of a

continuous (respectively absolutely continuous) distribution then |f(t) |? also
belongs to a continuous (respectively absolutely continuous) distribution.

Corollary 2 to theorem 3.3.2. Let F(x) and F(x) be two absolutely con-
tinuous distribution functions and denote by p,(x) and p(x) their frequency
functions. Let F = F,*F, and p(x) = F'(x«) be the density of F(x); then

p(2)-= es.Pi (2—%) po(x) dx.

We are now in a position to derive an inversion formula which is valid

for arbitrary frequency functions.
Let F(x) be an arbitrary, absolutely continuous distribution function and
denote its frequency function by p(x) = F’(x). We note that the charac-
teristic function f(t) of F(x) is not necessarily absolutely integrable. Let
G(x) be an absolutely continuous distribution function and write g(#) for
FUNDAMENTAL PROPERTIES 39

its characteristic function and q(x) = G’ («) for its frequency function.
Suppose that G(x) satisfies the following conditions:
(1) g(t) is absolutely integrable over (— 00, ©),
(it) g(x) = O(x-?) as |x| > o.
It follows from theorem 3.2.2 that
1 foo)

q(x) = 5 le en g(t) dt.

>) —itx

For any T > 0 the function gr(t) = g(t/T) is an absolutely integrable

characteristic function; the corresponding frequency function is
qr (x) = Tq(Tx).
We consider the function

G33) hn(t) = f@er( = f0¢(4).

We see from corollary 1 to theorem 3.3.1 that hr(t) is a characteristic
function which is absolutely integrable over (— «, 00). It belongs there-
fore to an absolutely continuous distribution function H7(x), and the
density of this distribution can be determined from theorem 3.2.2 and is
given by
Baa Hi (a= = |. ot e(7)f(tdt.
On the other hand we see from corollary 2 to theorem 3.3.2 that

Hy (x) = [_o(2-2) q(y) dy.

Therefore we see that

15(0)-00)1 = ||[2(x-2)-re ]aoro

Let a be a positive number to be chosen later, and write

oe i.<a (x5) na q(y) dy

as lee tle=m) (9) dy

I; = p(x) lind q(y) dy,
then
(3.3.5) |Hp(x)—p(x)| < ht+le+Js.
We write
(3.3.6) @(x,h) = ene |p(w+v) — p(x) |
40 CHARACTERISTIC FUNCTIONS

and

63%) = alex os or
It follows from property (ii) of g(y) that for y sufficiently large,
q(y) < Ay~*.
Therefore we have for sufficiently large values of a

(3.3.7b) peal (2-2) ay = “ae

as
and
A
(3:3.7¢€) Is < “(0

We put
a=T
and conclude from (3.3.7a), (3.3.7b), (3.3.7c) and (3.3.5) that
(3.3.8) |H.(x)—p(x)| < w(x, T-”*)+ AT 4+ 2Ap(x)T-*.
Let x be a continuity point of p(x), then limw(x, T~1/%) = 0 and we see
To
from (3.3.8) that
linn F(x) = pia).
T—>0

It follows therefore from (3.3.4) that

p(x)= eyaa
1 (itt
4 e7 ite (t.
=,
f(t) dt.

We have therefore obtained the following result.

Corollary 3 to theorem 3.3.2. Let g(t) be a characteristic function which ts
absolutely integrable over (— 00, 0) and suppose that the corresponding fre-
quency function q(x) satisfies the condition q(x) = O(x-*) as |x| > oo. Let
f (t) be the characteristic function of an arbitrary absolutely continuous distri-
bution function F(x); the frequency function p(x) of F(x) 1s then given by

D(x) = 1) fia = ue Ae
n)ft) dt,
T-—> 21

provided x is a continuity point of p(x).

Remark. 'This result could also be formulated by stating that the inver-
sion formula of theorem 3.2.2 remains valid for characteristic functions
which are not absolutely integrable if the integration in the inversion for-
mula is considered in the sense of a summability method with sum-
mability factor g(t/T). If we use g(t) = 1—|t| for |¢| < 1 and g(t) = 0
for |t| > 1, we obtain
P(x) = A {_e* (.-1) f(t) dt,
 FUNDAMENTAL PROPERTIES 4]
that is, the integral is taken in the sense of (C, 1)-summability. For
g(t) = «”, the integration is in the sense of Abel summability.
We consider next the convolutions of two purely discrete distributions.

Theorem 3.3.3. Let F(x) and F,(%) be two purely discrete distributions
with discontinuity points {£,) and {y,} respectively. Then F = F,*F, is also
purely discrete and the discontinuity points of F are the points of the sequence
16,+1,3. Morewer, let a, be the saltus of F, at £, and bh, be the saltus of
fyi Meand suppose that é is a discontinuity point of F. The saltus of F at é
is
% 4,03
$,49,~6

The characteristic functions f,(t) and f,(t) of F(x) and F,(x) respec-
tively are
Sf (t) = par 4, exp (ité,)

Silt) = Xb,
7
exp (ity,).
' The characteristic function f(t) of F(x) is, according to the convolution
theorem,
33.9) f(t) =f) falt) = EB a,b, exp (iE, +n,)].
Theorem 3.3.3 follows immediately from this formula.
Suppose that F(x) and F,(x) are two purely discrete distribution
functions and that at least one of these has infinitely many discontinuity
points. It follows then from (3.3.9) that F(x) = F,(x)* F,(x) also has
infinitely many discontinuity points.
Assume next that each of the purely discrete distributions F,(x) and
F,(%) has only finitely many discontinuity points, then (3.3.9) becomes
nN Hh

(3.3.10) f(it)/=flt) f(t) = y=l

X pma a,b, exp [it(E,+7,)]
where n and m are the numbers of discontinuity points of F(x) and
F(x) respectively. The last formula indicates that F = F,* F, also has a
finite number, say N, of discontinuity points. Let us consider the set
fé,4+n,}(v =1,...,n;p =1,...,m). Then it is no restriction
to assume
that the £,, £,,...,&, and 1, Nx ---» Ym are arranged in increasing order.
The set of discontinuity points of F(x) contains at most nm elements,
while it is seen immediately that the n+m—1 numbers £,+7, &.+7,
eg bat Irs Ent Nay ++ +s Ent are distinct. We formulate this result in
the following manner.
Corollary to theorem 3.3.3. Let F = F,*F, be the convolution of two
purely discrete distributions F', and F,. The distribution function F(x) has
42 CHARACTERISTIC FUNCTIONS

a finite number of discontinuity points if, and only if, each of the functions F,
and F, has finitely many discontinuity points. Denote by N,n and m the
number of discontinuity points of F(x), F(x) and F,(x); then
n+m—1< N < nm.
Next let F(x) be an arbitrary distribution function. If / (x) has a dis-
continuity at the point x = &; with saltus p, then the conjugate distribution
function F(x) of F (x) has a discontinuity at x = —é, with the same saltus
p; According to theorem 3.3.3 the convolution F*f has the saltus
> p;? at the point x = 0. On the other hand one can determine the saltus

of F * F at x = 0 from theorem 3.2.3 and we obtain the following result.

Theorem 3.3.4. Let F(x) be a distribution function and f(t) its characteristic
function. Then
in 2"gallop NG iptoideeae
Misery
1 T
= OSE Bs
where the p; are the saltus of F(x) and where the summation on the right 1s
to be taken over all discontinuity points of F(x).
We conclude this section by stating the connection between the existence
of the moments of a convolution and the moments of its components.

Theorem 3.3.5. Let F, and F, be two distributions and suppose that the
moments of order k exist for F, as well as for F,. The same is then true for
Weed foil dee
This follows easily from the elementary inequality
[x+y |P < 2h-7(|x|F + |y|*).
3.4 Limits of distribution functions
In this section we study sequences of distribution functions and their
limits and we introduce a specific definition for the convergence of such
sequences. In order to motivate this definition we consider first two
examples.
Example 1. Let
0 if x<—n

E(x) a if -n<x<n

1 Wve
(n = 1,2,...) be a sequence of rectangular distribution functions. This
sequence converges for all x and
lima (2) = 4.
no
FUNDAMENTAL PROPERTIES 43

We note that the limiting function of this sequence of distribution func-

tions is not a distribution function.
Example 2. Let
n «©

F(x) = om |, xP (—m8y"/2)ay (mir 12," ...)

be a sequence of normal distributions. It is easily seen that

lim F,,(x) = lim —— |edz = 4) if w= 0

nreo nro V2 J — 10h fake 20.
If we look at the graphs of the functions F,,(x) we might expect in-
tuitively that the sequence F,,(«) should converge to the degenerate
distribution «(«). This agrees also with the fact that
linv,, (2) Of x < O'and lint .f,,(x) = 1 if x*> 0.
Nn—> oo n—> oO

However, we have
lim F,,(0) = 4 while <(0) = 1.
n—> co

We observe therefore that it seems to be too restrictive to require that

a sequence of distribution functions should converge at all points to a
limiting distribution function. Example 2 suggests that exceptions should
be permitted for the discontinuity points of the limiting distribution.
Moreover, we see from Example 1 that a sequence of distribution functions
may converge at all points but that the limiting function is not necessarily
a distribution function. In view of the situation revealed by these two
examples the following definitions seem to be appropriate.
A sequence of functions {h,,(x)} is said to converge weakly to a limiting
function h(x) if
lim h, (x) = h(x)
n—> oo

for all continuity points x of h(x). We write then

Tam 2, (x)= h(x),

that is, we use the symbol “Lim” for weak convergence to distinguish it
from “‘lim”’ used for ordinary convergence.
Using this terminology we introduce the following definition.
A sequence {F,,(x)} of distribution functions is called a convergent
sequence if there exists a non-decreasing function (x) such that
iamyls, (caer (x):
n— oo

We note (see Example 1) that the weak limit of a sequence of distri-

bution functions is not necessarily a distribution function. However, the
44 CHARACTERISTIC FUNCTIONS

(weak) limit is always a bounded and non-decreasing function. We are

primarily interested in obtaining a necessary and sufficient condition for
the weak convergence of a sequence of distribution functions to a
limiting distribution. In order to obtain this condition we need some
results which are also of independent interest. These will be given in the
next section.

3.5 The theorems of Helly

We first prove the following lemma:
Lemma 3.5.1. Let {F,(x)} be a sequence of non-decreasing functions of the
real variable x and let D be a set which is dense on the real line. Suppose that
the sequence {F,, (x)} converges to some function F (x) in all points of the set D;
then
Lim 7, (4) =F (e).

Let x be an arbitrary continuity point of F(x) and choose two points

wéD, x 6D 80 that x “a Yea ee hen, (¢) crit (a) ate) nence
lim F,,(x*’) < lim inf F,,(x) < lim sup F,,(x) < lim F,,(x’’).
From the assumption of the lemma we conclude that

Since D is dense on the real line we have

F(x—0) < lim inf F(x) < lim sup F,,(x) < F(*+0).

From this relation one immediately obtains the lemma.

Theorem 3.5.1 (Helly’s First Theorem). Every sequence {F,,(x)} of uniformly

bounded non-decreasing functions contains a subsequence {F,,,(x)} which con-
verges weakly to some non-decreasing bounded function F (x).
The theorem is proved by the standard diagonal method and uses the
fact that the set of rational numbers is enumerable and can be arranged
in a sequence {r;}.
We form first the sequence {F,(7,)}. This is a bounded sequence of real
numbers and has therefore at least one accumulation point. Thus it is
possible to select a convergent subsequence {F,,,,(7,)}. Let
lim. F;.,,(71), =: (7).

In the second step we consider the sequence of functions {F,,,(x)}. We

select again from the bounded sequence of real numbers {F,.,(r2)} a con-
vergent subsequence {F,, (r,)} and write
limi tela: Ora)
FUNDAMENTAL PROPERTIES 45
The sequence of functions {/',,,(x)} is a subsequence of the original
sequence {F,,(«)} which converges for x = r,; and x = r,. We continue
this procedure and obtain a sequence of subsequences of {F’, («)}
Fgh) saly (BoP) (2)5. 0:0 «0 Heigl
4) yo.»
F (*), Foo (%), Fos (#), -- + Fon (x), ---

tie (x), Pics (x), Hes (x), MR fae (x), .

ey rer <o, a Oe Ole) ek ie 6) a (ef Benne |e Hele ce) oe ls). e :6

These sequences are selected in such a manner that each sequence is a

subsequence of the preceding sequence; moreover the mth sequence
Le (x), Eas (x), Ney (x), Sei 29 es (x), LD

converges at the first m rational points 71, 72, ... , 7m and we denote
fim ree DF) a ol oe oleae tite

We form the diagonal sequence {F,,,,(«)} and conclude that

lim F,, (72) = O(r;)
for all rational arguments r,. The functions {F,,,,(«)} are non-decreasing
and uniformly bounded, so that the function ®(r,), defined for all rational
values of the argument, is also bounded and non-decreasing. We introduce
now
F (x) = g.1.b.®(7,).
Th>&

The function F(x) is defined for all real x and agrees with ®(x) for rational
values of the argument. The function F'(«) is bounded and non-decreasing
and we see from lemma 3.5.1 that
(3.5.1) Lim F,,,(x) = F(x),
a.e-d,
The limiting function F'(«) is not necessarily right continuous; however
it is always possible to change the values of F(x) at its discontinuity points
in such a manner that it becomes right continuous. Such a change obviously
does not affect the validity of the relation (3.5.1).

Theorem 3.5.2 (Helly’s Second Theorem). Let f(x) be a continuous function

and assume that {F,,(x)} is a sequence of uniformly bounded, non-decreasing
functions which converge weakly to some function F(x) at all points of a
continuity interval [a, b| of F(x), then

lim [f(9)dP4(a) = |fC) AP,

46 CHARACTERISTIC FUNCTIONS
Since f(*) is continuous it is possible to construct a subdivision
Dey <a < So. = ey —rb of [apd] whichis’so tine xhag
(35.2) [f(®)—f(w)| <e© forxy <x < ayy
Here ¢ is an arbitrary positive constant. Moreover, the subdivision points
can be selected so that they are all continuity points of F(x). Hence
| F,(*;)—F («;) |can be made arbitrarily small if k is sufficiently large. Let
M = max f(x) and choose K so large that
axvxd

(3.5.3) |Fr (x,) —F( x;)| < MN

é€
fork > K.
We define a step function f, (x) in the interval [a, 5] by
(354) f(x) = f(x) forx,<* <a
and see from (3.5.2) that |f(~)— f,(x)| < «. Clearly we have

(3.5.5) | F(x) dF (x)- |f(x) dF, (s)

b b

b b
< |fear (o)—[ f(a) dF (@)
7) b
+ |
| I,(x) dF (x)— | Ff,(x) dF, (x)
b b
+
| f(x) dF, (x) — | f(x) dF (x)
We rewrite the first term on the right of (3.5.5) and see that

|UG)- £00] aF() <e| dF(s) = Ce,

In the same manner we get an estimate for the last term in (3.5.5)

| f(x) dF, (x)— | f(x) dF, () < Ce

b b

the existence of the constant C, is assured by the assumption of uniform

boundedness of the F',(«). Finally it follows from (3.5.3) and (3.5.4) that

[ fO)aF@)-
|" fear. e)
=| E fOMF OFM S/d) Fl=)
§ Me
<N. [Mae a ~~e
The last three inequalities and (3.5.5) yield the estimate
(3.5.6) | f(x) a(x) —| f(x) aF,(s)|< 2(1+Cye=Ce ifk > K.
FUNDAMENTAL PROPERTIES 47
But this means that

lim | f(x) dF,(x) = | f(x) dF (x),

: b b

which is the statement of the theorem.

Corollary to theorem 3.5.2 (extension of Helly’s Second Theorem). Let aed
be continuous and bounded in the infinite interval —~o < x < © and let
{F.(*)} be a sequence of non-decreasing, uniformly bounded functions which
converges weakly to some function F(x). Suppose that
lim F,(— 00) = F(—) and lim F,(+ 0) = F(+0),
ko —> 00
then
lima|© f(x) dP (x).b |iff(x)
ko
dF(x).
To prove the corollary we consider the three expressions

Ti=\|- feedrecs)—|" fear (e) — co

Ja= |[f(s)aP.(s)- |"f(a)dP@)

Ja= Jflo)arrs)— |”feyar(w)
where a < 0) < b.
Clearly

[- f@)ar(a)-|" f@)dF@)) < it JetIs

Since f(x) is bounded there exists a constant M > 0 such that
| f(~)| < M.
Let « > 0; as aconsequence of the uniform boundedness of the sequence
{F,,(x)} it is possible to determine |a| and |6| so large that J; < ¢ and
J; < &. From theorem 3.4.2 we see that there exists a K so large that
Nee Celork 2s, Pheretore

Hoe fe) dFy(x)— | f(a) 4F (x)

<(C+2)e fork > K

so that the corollary is proved.

A statement, analogous to the corollary, holds if the range of the integra-
tion is a semi-infinite interval.

3.6 The continuity theorem

In this section we derive necessary and sufficient conditions for the weak
convergence of a sequence of distribution functions to a limiting distribu-
tion. The theorems of the preceding section will serve as tools; we need,
48 CHARACTERISTIC FUNCTIONS

however, one more lemma which we deduce next from the inversion formula
and the convolution theorem.

Lemma 3.6.1. Let F (x) be a distribution function with characteristic function

f(t), then

[Fo)dy-[" Fo)dy=2(" SSA paar — cos ht

for any real positive h.

We denote by R(x) the uniform distribution over the interval (— a, +a);
the characteristic function (see Table 4) of R(x) is then

r(t)= sin ta

Let F(x) be an arbitrary distribution function and consider the distribution

Hd) = (x) RG clearly:
a eta

(3.6.1) H (x). = a} _.Fle-s)ay = a) Fede

while
sin ta
WO) Oa
We apply (3.2.6) and obtain

Hea)
(x + a) Bea)
—H(«—a) = tin1”
A
li esin tate.
ee 58a f(t)dt e
Using (3.6.1) we see that
T_1 —cos 2at
ie[F(x+v)—F(x—v)]dv = limair
— aera e7 tte
S(t) dt.
To I

The function (1 — cos 2az)/t? is absolutely integrable over (— 00, + 0) so

that the integral on the right converges absolutely. We write h = 2a and go
to the limit so that

| [F@+0)-Fee—2) dv = hi ey ~ite £(2) de.

We finally put x = 0 and transform the integral on the left and obtain the
formula of the lemma.
We now proceed to the main theorem of this section.

Theorem 3.6.1 (Continuity theorem). Let {F,,(x)} be a sequence of distribu-

tion functions and denote by {f,,(t)} the sequence of the corresponding charac-
teristic functions. The sequence {F,,(x)\ converges weakly to a distribution
function F(x) of, and only if, the sequence {f,,(t)} converges for every t to a
FUNDAMENTAL PROPERTIES 49
function f(t) which is continuous at t = 0. The limiting function is then the
characteristic function of F (x).
This theorem indicates that the one-to-one correspondence between
distribution functions and characteristic functions is continuous.
The necessity of the condition follows immediately from the extension
of Helly’s second theorem (corollary to theorem 3.5.2). To prove the
sufficiency we assume that the sequence f,(¢) converges for all ¢ to a func-
tion f(¢) which is continuous at t = 0. Let {F,,(x)} be the sequence of
distribution functions corresponding to the sequence {f,(t)} of charac-
teristic functions. According to Helly’s first theorem we can select a
subsequence {F,,, (x)}such that
Em. (x) =F (x)
k—>oo

where F(x) is a non-decreasing and bounded function which is continuous

to the right. Since the F,,, (x) are distribution functions we conclude that the
limiting function F(x) satisfies the inequality 0 < F(x) < 1. In order to
show that F(x) is a distribution function we must only show that
F(o)—F(-— o) = 1.
We apply lemma 3.6.1 to the functions F,,, («) and find that

| F(ay-| Facoddy = =(" SE pee

b . 1(° 1-—cosht

It is easily seen that the passage to the limit, k — oo, can be carried out
under the integral signs so that
1—cosy
‘| Fove-;| Fore ==| 7 f(y/h) dy.
The expression on the left of this equation tends to F'(co)—F(— 0) as
h > oo. Since we assumed that f(t) is continuous at ¢ = 0 we see that
lim f(t/h) = f(0) = lim f,(0) = 1.
Moreover, it is again permissible to carry out the passage to the limit under
the integral sign so that
1—cosy
F(«)—F(—0) = = are
dy.

It is well known) that

°“1—cosy qt
i, ye ao

(*) We integrate by parts and see that

eo CO -
Unies, =. | sin es
ae A eng
For the last integral see Titchmarsh (1937), Section 3.122.
50 CHARACTERISTIC FUNCTIONS

so that F'(co)—F(— oo) = 1. The limiting function F(x) of the subse-

quence F,,, (x) is therefore a distribution function. The argument just used
applies to every convergent subsequence of {F,,(x)}. It follows then from
the uniqueness theorem that every convergent subsequence of {F,,(x)}
converges weakly to the same limiting distribution F(x). This means
however that
Lim (x) =F (x):
n—>

Corollary 1 to theorem 3.6.1. If a sequence {f,(t)} of characteristic functions

converges to a characteristic function f(t) then the convergence is uniform in
every finite t-interval |—T, T].
Denote by F,,(x) and F(x) the distribution function of f,(t) and f(t)
respectively. Then

(3.6.2) PAU Rw AGIES I.e” dF, (x) — ife dF (x)

+[1—F,(6)+ F,(@]+[1-—F(6)+F(@)].
Let « be an arbitrary positive number and select for a and 5 two continuity
points of F(x), taking |a | and b so large that
1—F(b)+F(a) < «.
Since
Lim FF, (4) =F (x)

we have, for sufficiently large values of n,

1—F,,(6)+ F(a) < 1—F(b)+F(a)t+e < 2e.
The inequality (3.6.2) can then be written as

6.63) |f()-fO1 <|[ ede ()= fear) + 3e.

We wish to estimate the difference of the two integrals on the right of

(3.6.3) for values of ¢ from the interval [— 7, T]. To do this we subdivide
the interval [a, b] by means of the subdivision points
CX
Kym ag. a en
It is here no restriction to assume that all subdivision points are continuity
points of /(«) and that

m= max (X,—Xz_3) < €/T.

1<k<Nn
FUNDAMENTAL PROPERTIES LW
We note that

| et dF, (x) | et dF (x)

ta) b

< SI||maet dF, (x) — \ dF (| |

a

k=1
> {|ne (eit—ei) dF, (x) — (i? (citr»—eit ar(s)} .
Te-2
It is easily seen that for |t| < T
x
|; (eit _ es) dF, (x) < mT |; GF, (a) t< e| : dF, (x);
epe-a

this inequality remains valid if F,,(«) is replaced by F'(«). It follows then

from the preceding relation that

ipdF, (x)— Nhe dF (x) <> | dE (3) Vo dF (x) -+-2e.

The sum on the right-hand side of this inequality will not exceed 3¢ if is
sufficiently large, hence (3.6.3) becomes
|fn()—-F@)| < be
for all t e[—T, T], provided that n is chosen sufficiently large. This is the
statement of the corollary.
Corollary 2 to theorem 3.6.1. Let {f,(t)} be a sequence of characteristic
functions and suppose that this sequence converges for all values of t to a limit
function f (t). Assume that f (t) is continuous at t = 0; then f(t) 1s also a char-
acteristic function.
We remark that the limit of a sequence of characteristic functions is
necessarily continuous for all ¢ if it is continuous for the particular value
t = 0. The continuity at t = 0 is, however, an essential requirement of
theorem 3.6.1 and cannot be relaxed. As an illustration we consider once
more the sequence of rectangular distributions, discussed in Example 1 of
Section 3.4. We saw there that the distribution functions F’, (x) converge
for all values of x; however,
lim F,,(x) = 4
n—> oo

so that the limiting function is not a distribution. The characteristic

function f, (t) of F, (x) is
sin tn
In (t) = tn ’
therefore
: iy fort =.0
Boni) 3 if for t 4 0
so that the limiting function is not continuous for ¢ = 0.
 52 CHARACTERISTIC FUNCTIONS

The following corollary is a generalization of the continuity theorem. It

applies to bounded non-decreasing functions and will be used in Chapter 5.

Corollary 3 to theorem 3.6.1. Let {F,,(*)} be a sequence of bounded non-

decreasing functions such that F,,(— ©) = 0 and denote by

fa) = |eaFy()
their Fourier—Stieltjes transforms. The sequence {F,,(x)} converges weakly to
a bounded, non-decreasing function F(x) and
lim{Fy co)—Fy(— 00)}= F(co)—F(—0)
if, and only if, the sequence {f,,(t)} converges to a function f(t) which 1s
continuous at t = 0.
We prove first the sufficiency of the condition and write V,, = f,,(0) =

ir dF,,(y) for the total variation of F,, (x). Then

lim -Va=9f(0):
Consider first the case f(0) 4 0. Then the sequence of distribution func-
tions F’,,(x)/V,, converges weakly (according to theorem 3.6.1) to a distri-
bution H(x) and
Lim F,(x) = H(a)(0).
If (0) = 0, then
Lim F,, (x) = 0
no

so that the corollary holds in this case also. The necessity of the condition
is an immediate consequence of theorem 3.6.1.

Corollary 4 to theorem 3.6.1. Let {f,(t)} be a sequence of characteristic

functions and suppose that it converges uniformly to a limiting function f(t) in
every finite t-interval [—T, T]. Then the function f(t) is continuous at the
point t = 0.
To prove the corollary, we note that
(3.6.4) |f-FO] < |fO-fO|+1AO-f.O)1+1 f.0)—F(0)|-
Let « be an arbitrary positive number; we first choose 7 so large that

IfO-fO|<5
and for |t|
<1

|fn(0)-
FO) <5.
FUNDAMENTAL PROPERTIES 53
Then we choose ¢ so small that

| fn (t)—
fn(0)| < _
so that

If@)—F(0)| < e.
The last inequality proves the statement of the corollary.
We are now in a position to modify somewhat the statement of the
continuity theorem.

Theorem 3.6.2 (second version of the continuity theorem). Let {F,,(x)} be a

sequence of distribution functions and denote by {f,(t)} the sequence of
the corresponding characteristic functions. The sequence {F,,(x)} converges
weakly to a distribution function F(x) if, and only if, the sequence {f,(t)}
converges uniformly to a limiting function f(t) in every finite t-interval
[-T, +T]. The limiting function f(t) is then the characteristic function of

The statement of theorem 3.6.2 follows immediately from the continuity

theorem and from the corollaries 1 and 4.
We conclude this section with two remarks concerning the weak con-
vergence of a sequence of distribution functions to a limiting distribution
function.

Remark 1. It is possible that a sequence {F,,(«)} of distribution func-

tions converges weakly to a limit F(x) but that the moments of the F, (x)
do not converge to the moments of /'(x). Let, for instance, F(x) be given

byt («) :R,, (x) + (1- ~)e(a)where R,, (x) is the uniform distribution

over the interval [0, 7]. Since the characteristic function of F,, (x) is

f,() = =1 - dn
nit
1
+(1-2)
n
we see that Lim F,,(«) = (x). Moreover, it is easily seen that the moment
n> coo

p= | sar)(x) ==25
(oo) nk-1
o a= cs k dF —..

Therefore lim «”) = 00, while the moments of the limiting distribution
n>o

é(x) are all zero.

Remark 2. The weak convergence of a sequence {F,,(«)} of absolutely
continuous distribution functions to an absolutely continuous distribution
54 CHARACTERISTIC FUNCTIONS

function F(«) does not imply the convergence of the density functions
F; (x) to F’(«). Define, for example, F’,, (x) by
0 we
12
F(x) = 42 — if 0a

1 if Weal
Then F,,(x) converges weakly to the uniform distribution over [0, 1] but
Pn(x) = 1—cos 2anx does not converge to the rectangular density.
The continuity theorem suggests the expectation that two distribution
functions whose characteristic functions do not differ much are close to
each other in some sense. This idea was stated in a precise way by C. G.
Esseen (1944) [see also B. V. Gnedenko-A. N. Kolmogorov (1954),
pp. 196 ff.]. Esseen obtained the following results:

Theorem 3.6.3. Let A, T and « be arbitrary positive constants, F(x) a non-

decreasing function and G(x) a real function of bounded variation. Let f(t)
and g(t) be the Fourier—Stieltjes transforms of F(x) and G(x) respectively.
Suppose that
@ F(= 0) = Gee) Fo) = Gere)
(ii) i F(x)—G(x)|dx < 00
(iii) G'(x) exists for all xand |G'(x)| < A
Guy [7 HO=a dt =e

Then to every k > 1 there corresponds a finite, positive c(k) depending only
on k, such that

| F(x) —G(x)| < hot o(k) a

Theorem 3.6.4. Let A, T and « be arbitrary positive constants, F (x) a non-
decreasing purely discrete function, and G(x) a real function of bounded
variation. Let f(t) and g(t) be the Fourter—Stieltjes transforms of F(x) and
G(x) respectively. Suppose that
() F(-«) = G(— 0); F(+.0) = G(+«)
Gi) (fee|F(x)—G(x)|dx < 0
gay {7 =a a
t
FUNDAMENTAL PROPERTIES 55

(iv) the functions F(x) and G(x) have discontinuities only at the points
x,(v = 0, +1, £2,...54%,,, > x,) and there exists a constant
L > 0 such that Inf (x,,,—x*,) > L
Gy) |G. (@) he for alle Az.x, (Ui.
O sil, th2, aresa
Then to every number k > 1 there correspond two finite positive numbers
¢,(R) and c,(k), depending only on k, such that

|F(x)—G(x)|<
< ketei
provided that TL > c,(k).
The limit theorems of probability theory give approximations for the
distributions of normalized sums of random variables. Esseen’s theorems
provide an important tool for the study of the error terms of these approxi-
mations.

3.7. Infinite convolutions)

In this section we define and study convergent infinite convolutions and
their characteristic functions. It is not our intention to present here an
exhaustive discussion. We wish to give only a few results which will permit
the construction of certain interesting examples.
In Section 3.3 we defined the convolution of two distributions and
regarded it as a symbolic multiplication of distribution functions. It is
clearly possible to extend this operation to more than two factors, and it is
easily seen that the convolution of a finite number of distribution functions
is a commutative and associative operation. Therefore the convolution of a
finite number of distribution functions determines uniquely a distribution
which is independent of the order in which the factors are taken.
Before defining convergent infinite convolutions we introduce a suitable
notation and define certain sets which are useful in studying convolutions.
The concepts and notations introduced are also convenient in connection
with finite convolutions.
Let F,, F,,..., F, ben distribution functions. We write

3.71) TI* F(x) =the = Fithet...*F,

k=1 i=1
for their convolution and use the symbol []* to denote convolution pro-
ducts. The convolution of a finite number of distribution functions is again
(*) This section deals with a special topic; familiarity with this subject is not required
for an understanding of the rest of the book.
We deal in Section 3.7 with certain sets of real numbers and use the customary set-
theoretic notations. We designate sets by Latin italic capitals and their elements by lower-
case Latin italic letters. Thus x €_A means that the number x belongs to A; A © B (or
B 2 A) means that A is contained in B (or B contains A); A is the closure; Aciis the com-
plement of A. The empty set is denoted by ©.
c
56 CHARACTERISTIC FUNCTIONS

a distribution function; the characteristic function of (3.7.1) is, according

n

to the convolution theorem, given by the product [] f,(4) of the corres-

k=1

ponding characteristic functions.

In this section we shall need several lemmas which state simple properties
of characteristic functions and of sequences of characteristic functions.

Lemma 3.7.1. Let F(x) be a distribution function and f (t) its characteristic
function. Then Re [1—f(t)] > 4 Re [1-f(22)].

Since Re [1 —/f(é)] = ie (1—cos tx) dF («) we obtain the statement of

the lemma from the elementary relation,

. Lk 5 oy ee tx :
1—cos tx = 2 sin? 5 > 2 sin? — cos? — = 4 sin? tx = }(1— cos 2tx).
2 Z
Lemma 3.7.2. Let {F,,(«)} be a sequence of distribution functions and
{fn (t)} the corresponding sequence of characteristic functions. Suppose that
limf, (t) = 1 for |t| < to, then Lim F(x) = e(x).

It follows from lemma 3.7.1 that the relation lim f,,(#) = 1 is also valid

in the interval |t| < 2t) and therefore—by iteration—for all real t, so that
we can conclude that Lim F,,(x) = «(x).
N—>0

Lemma 3.7.3. Let {F,,(«)} and {G,,(«)} be two sequences of distribution

functions and suppose that there exists a distribution F(x) such that
Lim F,, (x) = F(x) and Lim (F,,* G,) = F; then Lim G,,(x) = e(x).

It follows from the assumptions of the lemma and from the con-
tinuity theorem that
(72a) limf,(t) = f(
and

(3.7.2b) lim f.()gn(0) =f.

Since f(#) is a characteristic function we know that f(0) = 1, and we con-
clude from (3.7.2a) that there necessarily exists a neighbourhood |t| < tf,
of the origin such that |f,,(¢)| > 4 for m sufficiently large and |t| < fp.
Using (3.7.2a) and (3.7.2b), it is then seen easily that lim g,(¢) = 1 for

|t| < t). The statement follows immediately from lemma 3.7.2.
Let F(x) be a distribution function; the spectrum Sp of F(x) is the
FUNDAMENTAL PROPERTIES 57
set of all points of increase of F(x). It is easily seen that Sp is closed, not

empty, and that | a(x) =. 1,

Sr
We give a second, similar definition. The point spectrum Dy of a distri-
bution function F(x) is the set of all its discontinuity points. Clearly, Dp is
an at most denumerable (proper or improper) subset of Sy. Dr need not be
closed, and can be empty; in fact Dy = @ if, and only if, F is a continuous
distribution.
Let A and B be two, not necessarily disjoint, sets on the real line. We
define the vectorial sum A (+) B of the sets A and B as the set of all real
numbers x which can be written in at least one way in the form x = a+,
where a € A and 6 € B. We agree to say that A(+) B = @ if either A or
B is empty. If both A and B are one-point sets, A = {a} and B = {5},
then the vectorial addition is equivalent to ordinary addition. If the set B
contains only the single point 6, B = {5}, then A(+) B = A(+) {5} is
obtained from the set A by the translation 6. The vectorial addition of sets
is commutative and associative. The vectorial sum of two closed sets is not
necessarily closed; however, if A and B are closed and bounded sets then
A (+) B is also closed.
Remark. 'The vectorial sum of two sets must not be confused with their
set-theoretic union.
We shall need in Chapter 9 a notation for the vectorial sum of n identical
summands A, and write (1)A = A and (n)A = (n—1)A(+)A for
te Dee Bente nen

Lemma 3.7.4. Let F,(x) and F,(x) be two distribution functions and let
F = F,* F, be their convolution. Then Sp = Sp (+) Sr, while
Dp = Dr, (+) Dr..

Let ye Sp, and z€ Spr, and put w = y+z. We select an arbitrary

positive number 6 and conclude from the assumptions of the lemma that

F(w+6)—F(w—6) = het F, (w-+-6—n)— F,(w—6—u)] dF, (u)

> [F,(w—2+28)—F,(w—2—-28)] fi dF,(u)
= [F,(y+26)—F,(y—26)] [F.(z+6)—F2(z—8)] > 0.
Therefore w = y+z2¢ Sy so that Sp (+) Sr, < Sr. Since Sr is a closed
set, we have also Sr, (+) Sr, S Sr.
We still have to prove the relation Sr, (+) Sr, 2 Sr. We give an in-
direct proof and assume that there exists a point w which belongs to Sy but
58 CHARACTERISTIC FUNCTIONS

not to Sr (+) Sr,. Since w € Sp we have for any 6 > 0

F(w+6)—F(w—6) = | [F,(w+6—u)—F,(w—6—u)] dF,(u) > 0.

But this is only possible if there exists a point u) € Spr, such that

Put vy = w—uy; then vp € Sp, and we Spr, (+) Sr, which is a contra-
diction, so that the first statement of the lemma is proved. In a similar way
one can also prove the statement concerning the point spectrum.
We also need another lemma which is of some independent interest:

Lemma 3.7.5. Let F(x) be a distribution function with characteristic

function f(t). Then t? [ aE (x) < 3|1-—f(#)| for |¢| < :where r 1s an
arbitrary positive number.

Since Re [1—f(é)] = lis (1—cos tx) dF (x) < |1—f(t)] we see im-
mediately that ae.
| (1 —cos tx) dF(x) < |1—f(2)|.
The statement of the lemma then follows easily from the fact that
x? < 3(1—cosx) for |x| < 1.
Let {F,,}, k = 1,2,..., bean infinite sequence of distribution functions.
In a purely formal manner we can introduce the infinite convolution(t)

1OD
laa = Yahi) 1deNes epee. bal ahaa’ °
k=1
of the distributions of the sequence. In order to give this infinite con-
volution a definite meaning, we form for each positive integer 7 the finite
convolution
n

P,,(x) = T[* Fi(*).

k=1
ive)

The infinite convolution T][* /, is said to be convergent if there exists a

k=1

distribution function F(x) such that Lim P,,(«) = F(x). We write then

(3.73) F(x)
=TT*Fe (8) (= Lim P,(»),
The characteristic function of P,, (x) is the finite product p, (¢) = III(t),
k=1
(t) In dealing with convolutions it is often convenient to omit the variable and to write
F instead of F(x).
FUNDAMENTAL PROPERTIES 59
and we conclude from theorem 3.6.2 that the necessary and sufficient con-
dition for the convergence of the infinite convolution (3.7.3) is that the
sequence p,(t) should converge uniformly in every finite ¢-interval to a
limit f(¢). This limit is then the characteristic function of F(x) and is given

by the infinite product f(t) = TJ fi(t).

k=1

Let [] f(t) be the characteristic function of a convergent convolution.

k=1
ioe)

We next show that J] f(t) is uniformly convergent in the sense of the

k=1
theory of infinite products.(+) Since lim p,(t) = f(t) is a characteristic
7

function (namely of the convergent infinite convolution), there exists an

interval |t| < t) such that f(z), and therefore also the factors f,,(t), do not
vanish for |t| < to. The infinite product is then uniformly convergent (in
the sense of the theory of infinite products) in this interval. It follows that
lim f,,(t) = 1 uniformly in |¢| < to, and we see from lemma 3.7.2 that
k>o
lim f,(¢t) = 1 uniformly in every fixed bounded t-interval. There exists
therefore, for every bounded interval, an N such that f,,(¢) does not vanish
on the interval if > N; this means that the infinite product which
represents f(t) is uniformly convergent.
We now derive criteria for the convergence of infinite convolutions.

Theorem 3.7.1. The infinite convolution F(x) = [[* F;, ts convergent if,
k=1
n+p
and only if, Lim [[*F, = (x).
n>o k=n+1
n+p n
Werwite.G,, =a) [ecl).and be =—)|.(* 2, so that 2,9 G,) = Da...
k=n+1 k=!
To prove the necessity of the condition, we assume that the infinite
convolution is convergent and note that Lim P,,, = Lim P, = F. We
n—> 0 n> 0

conclude from lemma 3.7.3 that Lim G,,,(x) = e(x). Conversely, if

n+p
Lim G,,,,(*) = e(«) then g,,(t) = J] f(t) converges uniformly to 1 in
N—> oo k=n+1

every finite ¢-interval, so that f(t) = [] fi. (¢) also converges uniformly in
k=1
every finite interval.

(t) See E. C. Titchmarsh (1939), A. I. Markushevich, vol. I (1965).

60 CHARACTERISTIC FUNCTIONS

Theorem 3.7.2. Let F = [[* F;, be a convergent infinite convolution. Then

k=1
foe)

the infinite convolution R,, = [[* F;, 1s also convergent and

k=n+1
Lim R,,(*) = e(x).

The convergence of R, lee immediately from theorem 3.7.1;

moreover, Ff = P,,* R, (whereagain P,, = II* F,) and Lim P,,(x) = F(x),
so that necessarily Lim R,,(x) = e(x). ai ms

We next derive a sufficient condition for the convergence of an infinite

convolution.
Theorem 3.7.3. Let {F,,} be a sequence of distribution functions and assume
that the second moment a), of F, exists and that the first moment «,,, of F,
ive)

is zero. Suppose that the sum > a; converges; the infinite convolution
k=1
co

[[* ¥: ts then convergent.

k=1

It follows from the assumptions and from ‘Taylor’s theorem that the
characteristic function f,(t) of F(x) can be written in the form

(3.7.4) fe(t) = 1+ 5Fe(Oxt) (|| < 1).

Let F(x) be a distribution function with finite moments of second order;
then its characteristic function satisfies the inequality

If Ol< | *4F(.
We see therefore from (3.7.4) that o
|1—fu(t)| < gta,
and conclude that the infinite product llJ, (t) is uniformly convergent in

every finite ¢-interval, so that the Bib tea: is proved.

Corollary to theorem 3.7.3. Let {F,} be a sequence of distribution functions

and assume that the integral M? = | |x |°dF,(x) exists for some & such

that 0 < 6 < 1. Suppose further that 3) M? < «; then the infinite con-
k=1
(oa)

volution [|* F’, 1s convergent.

k=1
FUNDAMENTAL PROPERTIES 61

We note that for 0 < 6 < 1 the inequality

Hs
sing
||
holds for real z. Therefore
tx
|e"*—1| — 2
a EE ARE AREMe
so that

lf,(t)—1| = (= 1) dF () < 2}t|? M2.

We conclude in the same way, as in the proof of theorem 3.7.3, that the
ioe)

convergence of the series >) M? implies the convergence of the infinite

k=1
convolution.
We next give a necessary and sufficient condition for the convergence
of an infinite convolution.

Theorem 3.7.4. Let {F;,} be a sequence of distribution functions and assume

that the first moment «,,, of F;, is zero. Suppose further that the spectra Sp,
ee)

are uniformly bounded. The infinite convolution [[* F’, 1s convergent if, and
b=1
only if, the series byOy 7, 1S convergent.(T)
The alata? faethe condition follows from theorem 3.7.3, so that we
need prove only that it is necessary.
Let F = [J*, be a convergent convolution of distributions which
k=1
satisfy the conditions of the theorem; write F*, for the distribution function
conjugate to F,,, and put G, = F,*F,. Let «$,, and «$,, be the first and
second moment respectively of G,. Then
a, = 0
while
(3.7.5) OSp = 2hepe

We conclude from the convergence of []*F, that []* Ff, and therefore
b=1 k=1
c

also [[* G; are convergent. Let f,(¢) and g,(¢) be the characteristic func-
k=1
tions of F(x) and G,(«) respectively. The convergence of the infinite

(+) The uniform boundedness of the spectra ensures the existence of the second
moments.
62 CHARACTERISTIC FUNCTIONS
co

convolution [T[*G, implies the convergence of the infinite product

k=1
Ilx(t). Since g,(t) = |f,(¢) |? is real and positive, we see that the product
k=1
TIg,(t) is absolutely convergent. Therefore the series 5) |1—g,(t¢)| < ©.
k=1 jaa
It then follows from lemma 3.7.5 that

Ge) | #dG,(x) < 00

k=1J/-r

for any real r. Since the spectra Sy, are, by assumption, uniformly bounded,
this is also true for the spectra Sj, of the conjugate distributions, and we
conclude from lemma 3.7.4 that the spectra Sg, are also uniformly bounded.
Therefore there exists a value r such that G;,(x) is constant for |x| > r and
all k. It follows from (3.7.6) that

>> 5% < 0.
k=1
In view of (3.7.5) we have then
(oo)

> Ho ie < ®,
k=1
so that the condition of the theorem is necessary.
We also need some properties of the spectra of convergent infinite con-
volutions. For this purpose it is convenient to introduce the following
terminology.
The closed limit inferior (+) of a sequence {A,,} of sets is the set of all
points x which have the property that every neighbourhood of x contains at
least one point of almost all sets A, (i.e. all sets A,, with n sufficiently large).
We write Li A, for the closed limit inferior of the sequence {A,}. The
statement x € Li A, means therefore that there exists a sequence of points
{x,} such that x, € A, and lim x, = x. We note that

Lid, = LiA, = LiA,

is a closed set.

oc

Theorem 3.7.5. Let F = [[*F;, be a convergent infinite convolution; then

k=1
Sp= Li[Sr (+)s. (ale Lie
For the proof of this theorem we need the following lemma:

(t) F. Hausdorff (1927), pp. 146, 147, uses the term “‘untere abgeschlossene limes”’.
See also C. Kuratowski (1952), p. 241. This is the closure of the inferior limit of the
sequence as defined by P. Halmos (1950).
 FUNDAMENTAL PROPERTIES 63
Lemma 3.7.6. Let G,,(x) (n = 1, 2,...) and G(x) be distribution functions
and suppose that Lim G,,(x) = G(x). Then Sy < Li S¢,.
Let « € S, and select h so that x+h and x—h are continuity points of
G(x). Then
G(x+h)—G(x—h) = lim [G,(*«+h)—G,(x—h)] > 0
so that G,,(x+h)—G,,(x—h) > 0 for sufficiently large n, say n > N, that is
x ES, and therefore x € Li Sg, so that the lemma is proved.
We proceed to prove theorem 3.7.5. Let F = [T]*F, be a convergent
k=1
n foe)

infinite convolution. We write again P,,z = T]*F, k and R, n = * F.k so

k=1 k=
k=n+1
that F = P,,* R, and Lim P,, = F, while—according to theorem 3.7.2—
Lim R, = «(x). It follows from lemma 3.7.6 that Sr © Li Sp, so that we
have only to show that Sr > Li S,,. Let x) ¢LiSp», and let 7 be an
arbitrary positive number. The open interval (* )—1, x )+) then contains
points of all the sets Sp,, provided 7 is sufficiently large, say n > N,. This
means that
(3.7.7a) Py(%o+)—Pa(Xo—7) > 0 (n > Nj).
Since Lim R,,(x) = e(x), there exists an integer N. such that
n—>co

(2/7D) me Ra Gh)
ha (— 9)i 0" fornia oNg
We select n > Ny = max (N,, N,) and note that F = P,,* R,, so that

P9421)
—F(y—2n) = [Pa ¥o+2-9)Pa¥o~20-3)] Ra (3)
> [Pa (o+3n)—Pa(so~3n)] |"dRa(3)
> [Px (*o+m)—Pn(%o—7)|[Ra(+7)—Rn(—7)] > 95
this follows immediately from (3.7.7a) and (3.7.7b). Therefore x) € Li Sp,
implies that x) € Sr, so that the theorem is proved.
We mention without proof two interesting results concerning infinite
convolutions.

Theorem 3.7.6. Let F = [[*F; be a convergent infinite convolution and

k=1
denote by p;, the maximum jump (saltus) of the distribution function F(x). The
point spectrum Dp = © if, and only if, the infinite product [| px diverges to
k=1
zero.

This result is due to P. Lévy (1931).

64 CHARACTERISTIC FUNCTIONS

Theorem 3.7.7. Let F = [[* Fy be a convergent infinite convolution of

k=1

purely discrete distribution functions F,. Then F is pure, that ts, F is either
purely discrete or purely singular or purely absolutely continuous.
For the proof we refer the reader to B. Jessen—A. Wintner (1935)
[theorem 35] or Wintner (1947) [No. 148].
We next discuss a particular case, i.e. the purely discrete distribution
function B(x) which has two discontinuity points atx = +1andx« = —1
and a saltus of 4 at each of these points,
B(x) = $o(x+1)+e(x—1)].
The corresponding characteristic function is
b(é)"= ‘cos t.
Let {r,,} be a sequence of positive numbers. In the following we shall use
the sequence of distribution functions

Ca eG) (2).
The infinite convolution

re)
is called a symmetric Bernoulli convolution. Let «1;, %;, be the first- and
second-order moments, respectively, of F’, (x). It is easily seen that «,, = 0
and «,, = 7%, while the characteristic function f,(¢) = cos7,t. The
peg Sr, = Dr, consists of the two points 7, and —7,. Suppose now
that

@:7.9) y X we,
k=1
Then the conditions of theorem 3.7.3 are satisfied and

(3.7.10) F(x) = TT Fe(@)i.ri2(*)

is a convergent infinite convolution. Condition (3.7.9) is therefore sufficient
to assure the convergence of asymmetric Bernoulli convolution. Conversely,
if we suppose that a symmetric Bernoulli convolution is convergent, then
we can conclude from theorem 3.7.1 that Lim F’, (x) = e(«); hence lim f, (t)
k>co k>o
= lim cos r,t = 1, so that lim 7, = 0. The spectra Sp, are therefore uni-
k—> co k—>00
formly bounded, and we see from theorem 3.7.4 that condition (3.7.9) is
also necessary for the convergence of the symmetric Bernoulli convolution
(3.7.10). We next study its spectrum Sy. A point x is in the spectrum Sz if,
and only if, it is possible to choose for every n the sign of 7, in such a way
FUNDAMENTAL PROPERTIES 65

that x becomes the sum of a convergent series of the form }) +7, In the
n=1

case where > fi. < ©, Say A = > Yn, this means that Sp ¢ [—A, A]
and that ae:aud +A are, Fespecnvely the smallest and greatest values
contained in Sp. If > r, 1s divergent, then it is possible to represent any
real number x as nett of a conditionally convergent series of the form
C= ss+7,, so that Sp is the whole real line. We next show that the point
=
spectrum of F is empty, Le.
(ll Dr "O:

Ceteagain® P= Il*F. and write -G,, = P37 Tl* F,), so that

fe * ICL a ay

F(x) = ["_ G,(-y)

dF,(9).
It then follows from (3.7.8) that
(3.7.12) F(x) = $[G,(*—71r.)+G,(x+7,)]-
Let s(y) and s,(y) be the saltus of F(x) and of G,,(«), respectively, at the
point x = y. It follows from (3.7.12) that
5(¥) = 315n(¥—7n)
+Sn(¥ +1)
and it is easily seen that
(3.7.13) s(w+2r,)+5(x—2r,) 2 s(x).
We give an indirect proof for (3.7.11) and assume therefore that there
exists a point x, € Dy. Then s(x) > 0, and it is possible to choose a
positive integer p such that
1
5(X5) > -.
P
We next determine p positive integers 1,, m,,...,%, such that r,, >
Vee n= Pag | WiSieis always possible since the 7, tend to zero.
The 2p numbers x9+7n, (f = 1,2,...,p) are then distinct and we see
from (3.7.13) that
S(¢9+ 2Fn,) + 5(%o—2%n,) = S(%o) (fF = 1,2,---,D)-
Hence

3 [s(t 21m) + 5(¢o—2rn,)] > psa) > 1.

j=
But this is impossible; hence Dr = O.
66 CHARACTERISTIC FUNCTIONS

This result is also a consequence of theorem 3.7.6, but we preferred to

give here a direct proof. We also note that we can conclude from theorem
3.7.7 that a convergent symmetric Bernoulli convolution is either purely
singular or (purely) absolutely continuous.
We summarize these results in the following statement:

Theorem 3.7.8. The necessary and sufficient condition for the convergence of
&
the symmetric Bernoulli convolution F(x) = T1*( )is the convergence of
k=1 Vr,
foe)

the sum > r2. The characteristic function of F (x) 1s then given by the infinite
k=1

product f(t) = [T[ cos7,t. The spectrum Sp is a bounded set if the sertes
k=1
ee) co

DX 7, converges, but is the whole real line if ¥ 1, diverges. The point spectrum
k=1 k=1
of an infinite symmetric Bernoulli convolution 1s always empty, and F (x) 1s
either purely singular or absolutely continuous.
We next consider briefly the case where the 7, form a geometric series,
Tr == He.

Corollary to theorem 3.7.8. The function f(t) = [J cos a*t is the charac-
k=1
teristic function of a convergent symmetric Bernoulli convolution

F(x) = TI" B(xa~)

ioe)

k=1

if, and only if,0 <a <1.

We finally mention, without proof, another interesting result concerning
symmetric Bernoulli convolutions with bounded spectrum. We consider a
convolution

ei) Fal ie (2),

Vr
foe) co

where the series >) 7, converges, and write p, = >) 7, (2 =.0,1,2,...)

kK=1 k=n+1
for the remainder of this series.

Theorem 3.7.9. Suppose that 7, > pn (or equivalently pp > 2pn4) for all
n. Then
L( Spr) = 2 lim/2"7,..
> eo

Here Sy is the spectrum of the convolution (3.7.14), while L(Sp) is the

Lebesgue measure of Sp.
FUNDAMENTAL PROPERTIES 67

For the proof of theorem 3.7.9 we refer the reader to R. Kershner-

A. Wintner (1935) or to A. Wintner (1947). Theorem 3.7.9 permits us to
decide whether a convolution of the form (3.7.9) is singular. We consider
a particular case.

Corollary to theorem 3.7.9. Let 0 < a < 4, then the symmetric Bernoulli
convolution F(x) = II" B(xa-*) ts purely singular.

We conclude bene seksby giving a few examples.

Cy) Let 7, = i then f (7) = TIcos (

4 is the characteristic function

mentioned in the footnote to page 20.

(II) Let 7, = k-* (k = 3, 4,...); then f(é) = T].<os(k-"#) belongs to a

singular distribution with a bounded spectrum, and it can be shown that
lim sup |f(¢)| > 0.
|tl>oo

(III) If a@ is a rational number, 0 < a < 4, but not the reciprocal of an

integer, then R. Borsbocy (1936) has shown that the singular characteristic
function f(t)= I cos (a* t) satisfies a relation |f(t)|= O[(log |t])~]=
o(1) as |t| > co (i> 0); therefore we have in this case asae by (e) 10:

(Vili, = 3 we obtain, except for a factor ¢”, fe characteristic

function (2.1.7).
Ss t
(VY) lfr, = 25 then f{(¢) =.]] cos an We can then show by an elementary
n=1

in ¢ ee
computation (given in section 6.3) that f(t) = = =/7(¢)) 1 Isiis the

characteristic function of the rectangular distribution and is therefore

absolutely continuous and has the interval [—1, + 1] as its spectrum.
(VI) For 7, = 4-” we obtain the singular characteristic function
2 t
= pa
ne enV ee 3)
which we shall need later.
Additional examples can be found in B. Jessen—A. Wintner (1935),
R. Kershner (1936), A. Wintner (1947) and A. Wintner (1938). The first
of these references also contains examples which show that the spectrum
of a singular as well as of an absolutely continuous symmetric Bernoulli
convolution can be bounded or the full real line.
 4 CRITERIA FOR CHARACTERISTIC FUNCTIONS
It is frequently of interest to decide whether a given complex-valued
function of a real variable is, or is not, the characteristic function of some
probability distribution. The inversion formulae provide a method to
answer this question. While this approach is theoretically always possible
it is often not practicable. We therefore develop in Sections 4.1-4.4
criteria which can be used to decide whether a given function is a charac-
teristic function. Section 4.5 deals with an essential property of charac-
teristic functions; knowledge of this property helps us to realize that the
characteristic functions are practically a unique tool for simplifying the
analytical treatment of certain probability problems.

4.1 Necessary conditions

Theorems 2.1.1 and 2.1.2 assert that every characteristic function f(t)
satisfies the relations |f(t)| <f(0) = 1 and f(—#) = f(t), and is uni-
formly continuous. Thus, these theorems already provide useful necessary
conditions which a function must satisfy in order to be a characteristic
function. However, these conditions are not sufficient. To show this we
first prove the following theorem and use it to construct an example.

Theorem 4.1.1. The only characteristic function which has the form
f(t) = 1+0(t?) as t +0 is the function f(t) = 1.
Let f(t) be a characteristic function and assume that f(t) = 1+ 0(t?)
as t >0Q. It follows then from theorem 2.3.3 that «, = a, = 0. Since
eo | x? dF (x) [where F(x) is the distribution function corresponding
to f(t)] it is seen that F(«) must be constant over every interval which
does not contain the point x = 0; that is F(x) = e(x) and f(t) = 1.
The function f(#) = e~" satisfies the conditions of the theorems 2.1.1
and 2.1.2, but e~* = 1+0(t?). Therefore we see from theorem 4.1.1 that
e~" is not a characteristic function. This example shows that the neces-
sary conditions stated in theorems 2.1.1 and 2.1.2 are not sufficient.

Corollary to theorem 4.1.1. Let w(t) = o(t) as t +0 and suppose

that w(—t) = —w(t). Then the only characteristic function of the form
f(t) = 1+ w(t) +o(t?), as t > 0, ts the function
f(t) = 1.
We know [corollary 2 to theorem 3.3.1] that f(t)f(—?) is a charac-
teristic function. Under the assumptions of the corollary f(t)f(—2)
CRITERIA 69
= [1+ w(t) + 0(t?)][1 — w(t) + o(t?)] = 1+0(t?), hence according to theo-
rem 4.1.1 |f(é)|? = f(f(—t) = 1 or f(t) = e (a real). Therefore
f(t) = 1+1at—}a*t?+0(t?) and this has the form 1+w(¢)+0(t?) (with
w(t) = o(t)) only if a = 0 so that f(t) = 1.
We next discuss an inequality which every characteristic function must
satisfy and which can therefore also be regarded as a necessary condition
for characteristic functions. This inequality is also of some independent
interest.
Let F(x) be a distribution function and f(t) its characteristic function.
Then the inequality

(4.1.1) | x? dF (x) < 3u-®[1—Re f(u)]

|x|<1/u
is valid for u > 0.
We note that

1—Re f(u) = ie (1—cos ux)

dF(x)
nt? 7 ox 11u?
= ——| 1——— }dF (x) > — | 2dF (x),
ieee ? ( ie ) (*) 24 woe iz)

so that (4.1.1) holds.

We have, according to lemma 3.7.1,
(4.1.2) Re [1— f(@)] > + Re [1— f(22)].
By induction we obtain easily the following condition:

Theorem 4.1.2. Let n be a non-negative integer; then the inequality

Re [1 f(] > gaRe [1- f(20)

is satisfied for every characteristic function.
We now consider a characteristic function which has the property that
(4.1.3) Pie te te Db.
From theorem 4.1.2, applied to the function |f(t) |?, we see that

1-[f@lt
> gll-1f2 OP.
Let t be a fixed value such that |t| < B, and choose n so that

ze=e 2x ae
Ue hie»
Then—according to (4.1.3)—
. 1 t?
bi(2tt) tae Aa while rT = Ap
70 CHARACTERISTIC FUNCTIONS

Hence we obtain from theorem 4.1.2 the inequality

t?
= 2 Nes 4/2

A slight modification yields the following corollary:

Corollary to theorem 4.1.2. Let f(t) be a characteristic function which
satisfies (4.1.3); then
If@| < 1- t?(1—A?)
SS
for all
t such that0 < |t| < B.
We next discuss another important property of characteristic functions.
Theorem 4.1.3. Let f(t) be a characteristic function and let N be an arbitrary
positive integer. Denote by
ie os 2
S= > 2 Site) shen
j=1 k=1

where t,, t.,..., tn are arbitrary real and &,, €,,.. ., &n are arbitrary complex
numbers. Then S is real and non-negative for any choice of N, t,, ta, . . -, tn,
&, bs, OO OF} En.
Let F(x) be the distribution function which corresponds to f(t). ‘Then
N WN . foe)
=, es All exp [7(t;—t,,)x]
dF (x)
j=1 k=1
© N d NGS .
| (> E, ete) > E,,ent dF (x)

—o \j=1 k=1

|” |S eet |’d(x).
N
I
—o |j=1
The last expression is real and non-negative.

4.2 Necessary and sufficient conditions

The property of characteristic functions which is described by the last
theorem suggests the introduction of a concept which is useful in formu-
lating necessary and sufficient conditions for characteristic functions.
A complex-valued function f(t) of the real variable t¢ is said to be non-
negative definite for — 0 < t < +o if the following two conditions are
satisfied:
(i) f(z) is continuous;
(ii) for any positive integer N and any real t,, ..., ty and any
complex &,, ..., vythe sum
N N
= ya YS (G—t)8; &,
j=1 k=1
is real and non-negative.
We establish next a few properties of non-negative definite functions.
CRITERIA 71

Theorem 4.2.1, Let f(t) be non-negative definite. Then

(a) f(0) ts real and f(0)>
et (oi) = 7(2)
(c) If@I1 < F(0).
Proof of (a). Put N = 1, t; = 0, &, = 1; then (ii) implies (a).
Proof of (b). Put N = 2, t, = 0, t, = t and choose arbitrary £, and é,.
Then S = f(0)[|& [?+ || "1+ f(—H6 6.4 f(t )£,&,. It follows therefore
from (a) and (ii) that f(—Z)&, + f(f)é26, is real for any &, and é. We
write f(—t) — a, +78, f(t) = % +1Bo, Gite = y +10. Then

(%1 + 2B1)(y +10) + (a+ 1B2)(y — 10)

is real, so that (8,+/2)y+(«,—%)d = 0 for any y and 6. This is only
possible if 6, +8, = 0 and «,—«, = 0, so that (b) is satisfied.
The property (b) is sometimes expressed by stating that the function
Ff(¢) is “Hermitian’’.
Proof of. (c).. We again put NV = 2,4, = 0, i, = ¢ but choose £, = f(#),
— |f(#)|. Using (ii) and (b) we see that
S = 2f0)|fOP-21
fF 2 0
In the case where |f(¢)| > 0 this inequality immediately yields (c). Since
[by (a)] f(0) > 0, relation (c) holds in a trivial way if |f(t) |= 0.
We can now formulate a criterion for characteristic functions.

Theorem 4.2.2 (Bochner’s theorem). A complex-valued function of a real

variable t is a characteristic function tf, and only tf,
(i) f(t) ts non-negative definite
(ii) f(0) =
The necessity of the conditions is established by theorems 2.1.1 and
4.1.3 so that we need prove only their sufficiency. We therefore assume
that f(t) is a non-negative definite function and choose positive integers
and N and a real number x and put ¢; = j/n, &; = e~**. It follows then
from the definition of non-negative definiteness that

wy() =4,0 SF") exp [1-H] 2 0

a

forall «, ‘The nebeeeSek =r occurs in N—|r| terms of this sum;

here 7 is an integer between —N+1 and N—1. We collect these terms
and write
I7|
(4.2.1) Bx(x) = i (- ED)fener > 0.
Therefore
[ge G"(x)dx = 2n (1=) f ()
Ae CHARACTERISTIC FUNCTIONS

or

(4.2.2) (1=) if(*)be= [5 (0) 1

We now introduce the function
0 forx < —7

(4.2.3) FP) = [- Beoar for —7 <4 # <2

1 forx >2
then (4.2.2) can be written as

(4.2.4) (a) #() - |) éary Gy

We see from (4.2.1) and (4.2.3) that F% (x) is a non-decreasing, right
continuous function and conclude from (4.2.4) and assumption (ii) that
| dF (x) = f(0) = 1 so that FY (x) is a distribution function. We
consider next, for a fixed n, the sequence {F'“) (x)}of distribution functions.
According to Helly’s first theorem there exists a subsequence
{F% (x)} (with lim N, = 0)
ko

which converges weakly to a non-decreasing, bounded function F™ (x).

This limiting function is a distribution function since for any N and
any « > 0 we have FY (—a-«) = 0 while FO (a+<e) = 1. We see from
(4.2.4) that

flsin) = Yin (1-21) f(oym) = him [oe dre

Applying Helly’s second theorem we obtain

(4.2.5) — f(s/n) = |"ist dB (x)

for all integers s. We cereus the distribution functions
Fe (2) =i (0/2):
The points of increase of the distribution function F,,(x) are all located
in the interval (—nz, +n); the characteristic function of F,, (x) is
rn
mm

dF, (x) = | eitny dF (y).

ie (t) = | e

It follows from (4.2.5) that for any integer k

(4.2.6) f,(k/n) = f(k/n).

Let ¢ be an arbitrary real number. It is then possible to determine a
sequence k = k(t, m) such that
CRITERIA 73
(4.2.7) 0 < t—(k/n) < (1/n).
We write 0 = t—(k/n); then 0 < 6 < (1/n) and
mm ; : |
| fu(t)— fa (R/m)| = | elton (ef — 1) dF (2)|< |ma |e*—1
n
.

—mNn
| dF, (x).
We apply Schwarz’s inequality (see Appendix B) to the last integral and
obtain

|fa(t)— falk/n)| < E [cos ox) aF, )|i

—7TMNn

2 [2|2 (1 —cos 62) dF™ | He

Since nO < 1 we have 1—cos nfz < 1—cos z for |z| < a; hence

|fa(@)—falk/n)| <2 {" (Leos 2) dF (ay)

= {2[1—Ref,(1/m)]}}¥? = {2[1—Re f(1/n)]}?.
By assumption, f(t) is a continuous function such that f(0) = 1, and the
right-hand member of this inequality can be made arbitrarily small by
selecting n sufficiently large; this means that
(4.2.8) lim |f, (t)—fn(k/n)| = 0.

Moreover, we see from (4.2.6) and (4.2.7) that

(4.2.9) #(He= lim f(k/n)-= liu -f;(k/n).

Since
lim f(t) = lim {[fn(4)— fn(k/m)] + fa (R/m)}
we see from (4.2.8) and (4.2.9) that for all ¢
linayy,(2) afte):

The continuous function f(t) is therefore the limit of the sequence {f, ()}
of characteristic functions and is therefore (corollary 2 to theorem 3.6.1)
also a characteristic function. This completes the proof of Bochner’s
theorem.
We derive next another criterion which is due to H. Cramér.

Theorem 4.2.3 (Cramér’s criterion). A bounded and continuous function

f(t) is a characteristic function tf, and only if,
(i) f(0) = 1
Gee |. |_f(t—w) exp [ix(t—u)] dt du
is real and non-negative for all real x and for all A > 0.
74 CHARACTERISTIC FUNCTIONS

We first prove the necessity of the condition by substituting

f(e—u) = |" exp ¢-u)y] 49)

into (ii). Since the inversion of the order of integration is permissible we
obtain easily
s —cos A(x
we) 2] © oa ci
This shows that y(«, A) is always real and non-negative. To prove the
sufficiency of the condition we assume (i) and (1i), so that

(4.2510) p(x Zul i= an [f(u-v)é"""” dudv> 0.

In the last integral we introduce new variables

t=u-—v
Pao
This change of variables transforms the original region of integration
into a parallelogram. One diagonal of this parallelogram is located on the
z-axis and decomposes it into two triangles. The function p(x, A) is then
computed by adding the two integrals taken over these triangular regions.
Thus
p(x, A) = 5|ief(t) et ieas|iSiP f(det ||* ds]dt.
We introduce the function

ce if |t| < A
(4.2.11) f4() =
0 otherwise
and can write (4.2.10) in the form

es, A)= |e fa (at > 0

Let a

Jos A) = 52| (1 Note, ayers

then

MURA = i (1aah |[oat f, (2)at|dx.

The order of the integrations may be inverted; moreover a simple
computation shows that
i evroe(t a) | eee 1—cos [(u+ t)By
WE B Buu+t)?
CRITERIA 75
Therefore

Ja (t, A) = :\-.Fee ta (t) dt.

Here we introduce a new variable v by t = (v/B)—u and obtain

EDO aOe Ne :|S eS ja 2—u) dv.

From the well-known relation)

|@ 0
1—cosv
2
d I
eS)

and (4.2.12) it follows that

lim Jz (u, A) = fa (—x).
Boo

The function which is defined as

<B
mt — 2)pe,A) for|x|
and 0 outside the interval (—B, B) is non-negative and bounded. If it is
multiplied by a suitable normalizing constant Cz it becomes a frequency
function. Therefore Cz Jz (u, A) is a characteristic function for any B and
A, We then conclude from corollary 2 to the continuity theorem that
lim Coo (uy A) = Cfa(—n)
is a characteristic function. Since f4 (0) = 1, we see that C = 1 and that
fa(u) is a characteristic function for any A. We apply once more the
continuity theorem and see finally that
f(@) = lim fa(t)
is a characteristic function. Thus the sufficiency of the condition is
established.
Theorem 4.2.3 is a particular case of more general results derived by
Cramér (1939). In this paper he also gave conditions for the possibility
of representing more general classes of functions by Fourier integrals.
We derive next a condition for absolutely continuous distributions
which will later be extended and will yield a general criterion.

Theorem 4.2.4. The complex-valued function f (t) of the real variable t is the
characteristic function of an absolutely continuous distribution tf, and only Uv,
it admits the representation

() f= | s¢+6) sao
(*) See footnote on page 49.
76 CHARACTERISTIC FUNCTIONS

where g(0) is a complex-valued function of the real variable 0 such that

Gi) ["_1e(@ 1249 = 1

is satisfied.
Before proving this theorem we summarize some known results con-
cerning quadratically integrable functions. For the definitions used in this
summary as well as for the statements of the theorems quoted we refer the
reader to R. E. A. C. Paley-N. Wiener (1934).
Let d(x) be a function which is quadratically integrable over (— 00, + 00);
according to Plancherel’s theorem ¢(x) has a Fourier transform (1)
which is also quadratically integrable over (— 00, + 00) and which can be
written as

(42.13) ®(u)=Lim. leya ion)dis

i] A

Here the symbol

li.m.
A->«

denotes the limit in the mean as A tends to infinity. It is known that

d(x)= Lite 125 ie e* O(u) du

wa Woe
or
A
A(x)= Bie
aS ee:
ae[osé”* O( —v)
dz.

Hence ¢(x) is the Fourier transform of h(v) = ©(—v). It is easily seen also
from (4.2.13) that ®(w) is the Fourier transform of ¢(—y). According to
Plancherel’s theorem we have the equation

(4.2.14) [fe|p(x) |2de = lee |@(u) |2du.

From Parseval’s theorem we get the following relations

[oboe tedu = |"iy) May) dy

(4.2.15) = JieO(—y) O(—x+y) dy

[20bert du= |" (9)aT Day,

To prove the necessity of the conditions of theorem 4.2.4 we consider
an arbitrary frequency function p(x). Then ¢(x) = V‘p(x) is quadratically
CRITERIA viel

integrable over (— 00, 00) and we write g(u) for its Fourier transform. From
the first of the equations (4.2.15) we obtain

[om p(w) du i | e(-p)e(- #43)dy.

Letting x = —t we get

(4.2.16) | e™p(u)du = |"g(e+y)g(—y)dy.

The function ¢(«) = V p(x) is real valued; we see therefore from (4.2.13)
that its Fourier transform g(w) is Hermitian. Since the integral on the left
of (4.2.16) is the characteristic function of the distribution whose frequency
function is p(x) we see that (i) is satisfied while (ii) follows then from
f(0) = 1. This completes the proof of the necessity of the conditions (i)
and (il).
To prove their sufficiency we suppose that the function f(t) admits the
representation (i) by means of a function g(6) which satisfies (ii). Then
g(9) is quadratically integrable over (— 00, + 00) and has a Fourier trans-
form which we denote by G(u). We see from (ii) and (4.2.14) that
\- | G(u) |?du = 1
so that |G(x) |? is the frequency function of some absolutely continuous
distribution. It follows then from the second equation (4.2.15) that

[= [G(w)|rerdu = |" e()aCe Fy)dy,

hence the integral on the right is the characteristic function of an absolutely
continuous distribution. Applying theorem 2.1.3 we see that its complex
conjugate

f= | se+aoday
is also a characteristic function. This completes the proof of the theorem.
We next derive a condition which is not restricted to absolutely con-
tinuous distributions.
Theorem 4.2.5 (Khinchine’s criterion). The complex-valued function f(t) of
the real variable t is a characteristic function if, and only if, there exists a
sequence {g,,(0)} of complex-valued functions of the real variable 0 satisfying

() | |gn(0)[249 = 1
such that the relation

(i) (2) = Tim | gu (t+) B5(0) a

holds uniformly in every finite t-interval.
78 CHARACTERISTIC FUNCTIONS

To prove the necessity of these conditions we must only note that every
distribution function is the limit of a sequence of absolutely continuous
distributions. We apply the preceding theorem to the characteristic
functions of these approximating distributions and see that our conditions
are satisfied.
To prove the sufficiency of the conditions we assume that the sequence
{g,(0)} is given and that these functions satisfy (i). According to theorem
4.2.4 the functions

falt) = |© Bu(t+9) 8(0)a0

are characteristic functions and condition (ii) states that
F(t) = lim f(t)
N—> co

is the uniform limit of a sequence of characteristic functions in every

finite ¢-interval.
The sufficiency of the condition follows immediately from the second
version of the continuity theorem.
We conclude this section by giving a necessary and sufficient condition
which an even function must satisfy in order to be a characteristic function.
This condition involves the Hermite polynomials which are defined by the
relation
d®
G27) TH Glee “ne
tere
It is easy to see that H,,(«) is a polynomial of degree n in x.
We first derive an estimate for the polynomials H,, (x).

Lemma 4.2.1. Let H,,(x) be the Hermite polynomial of degree n, then

Ed) |e nere ee!

ay 14 Hee:

We differentiate the relation

1 ie
a
a ex — y?/2+iyxdy =en eee"2
n times and obtain, in view of (4.2.17),
ie a ae
(4.2.18) He) = | (zy)” exp [—y?/2
+ tyx] dy.
V2n J-©
Therefore

|H,(*)| < &?

(eo) CWP) Yoho}

is ly pe"?dy = oe. | tre "/? dt

V2 WA 0
CRITERIA 79
n/2 27/2 oo
= 2 ie | y"-D/?2 e-"dy = or
Vx 0 awa asap (*5)
2
so that the lemma is proved.

Theorem 4.2.6 (theorem of M. Mathias). Let f(t) be a real, even, continuous

function which ts absolutely integrable over (— «©, + 0). Write for p > 0
OT eae an Ee ae

cn(P) = (= 1)" | f(ps)e-*?Hy(3) de,

The function f(t) is a characteristic function if, and only if,
(1) f(0) =
(it) Con(p) > 0
oral n= (1, 2... .,.and for allp > 0.
We prove first that the condition of theorem 4.2.6 is necessary and
assume that f(t) is non-negative definite. It follows from the definition of
Con(p) and from hee: (4.2.18) that

TD aie
ie (px) i exp (ixy—y*/2) dy|dx.
Since f(px) is absolutely integrable, the order of integration may be
reversed and we see that

cold) =a= | ymervay |”slpayet as}dy,

Since f(x) is non-negative definite, the integral in the braces is non-
negative and therefore c.,(p) > 0
For the proof of the sufficiency of the condition we need two lemmas.

Lemma 4.2.2. The function (—1)" e~*/? Hyn, (x) is non-negative definite.
It follows from (4.2.18) that
1 (oo)

(4.2.19 aa!)
= 1 \"en? “17,Oi(x) == =|
Fyre’! 2n exp p[—y?/2 +tyx]
[—y?2/2+iyx] dy
dy.

Theright-hand side of (4.2.19) is the Fourier transform of the bounded, non-

negative function (27)~/? ye~”/?; this is integrable over (— 00, + 00).
It follows from Bochner’s theorem that its Fourier transform is non-
negative definite.
Lemma 4.2.3. Suppose that ¢(«) is a real-valued, continuous and bounded
function which is absolutely integrable over (— «©, + 0) and put
e {- $(x)e-"2H,(x)de (n =0,1,2,...).
80 CHARACTERISTIC FUNCTIONS

Let r be a real number 0 < r < 1, and write

s a7 Hs
Son os n!\Te
Then

CG.) = Teas
J |. $(x) exp |---| dex
(x—ry)?

and

To prove the lemma we use first (4.2.18) and rewrite a, as

(4:22.20) ie Sage lie(x) eae (it)” exp (ttx— t?/2) at|dx

—© V 2x J-©
so that
(4.2.21) G(r, y)
tT | est aay fess La ie ;
ae = at ie sie {Gt(ctr)” H,, (y) exp [tx— t2/2] dt dx.

For the sake of brevity, we write

(4.2.22)

In(*, 9,7) = fn (x) = efIre (ctr)” H,, (y) exp [ttx— t?/2] dt,

then

(4.2.23) GG, y)= ae Ro Tu (x) dx.

We use lemma 4.2.1 to estimate the integral on the right-hand side of

(4.2.rs

5 pe (ctr)” H,, (y) exp [etx —t?2/2] a

<
yr ev/2 Qn/2 (

n! V2n V7
;
ioe)

|
—o
It \@e db te rere)
re

azI'(n+1)
D)
Tha

We apply Stirling’s formula and get

1 oats
oe ie (itr)" Hy (y) exp [ite—12/2] dt = orev") asn > o,
CRITERIA 81
Combining this with (4.2.22), and noting that ¢(«) is bounded, we see that
(4.2.24) | fn(%) |= |d(x) |ofr" "”) = ofr") asn > o.
Clearly, {f,(x)} is a sequence of absolutely integrable, continuous
functions, and it follows from (4.2.24) that 3F,(x) is absolutely and uni-
formly convergent and that the summation and integration can be inter-
changed in (4.2.23), so that

Gt09)=ae |” [ZA00]
or, in view of (4.2.22),
G(r, y)
1 Pa
a foe ‘igsP(xeee nl (ctr)" H,, (y) exp (itx
— t?/2) dt }dx.
JU . —o

We write

(4:2.25)
1
~g,(i)= ips (itr)" H,,() exp (it«
— t?/2)
;

so that

4226)
Donen G = = o{s
{pe |” » earhas
it (athde.
We see from (4.2.25) and from lemma 4.2.1 that

le g,(t)dt
re)
= | _len() ae < CA
, ziceae D
Ee Whe
= o(r"e"/")_ asn > 0.
It then follows easily that

so that

(4.2.27) . ‘i
x é exp (tx —t?/2) |3(ttr)"
_ &n (b) di se ie ere)Hn (Y) dt.

We use (4.2.18) to compute the sum on the right of (4.2.27) and obtain
ei(tiry®Ha(y) 1 5 (itr)"
try
Gale if (iv)" exp [—v?/2+ ivy] dv.
n=0 n!} AGE es
=0 nl

It can be shown that the summation and integration can be interchanged,

and we obtain by an elementary computation
82 CHARACTERISTIC FUNCTIONS

(422.23)
© (ar Hety) tog?
is exp
=,” + ivy
[—(v?—y?)/2 — trv] dv
a a ce
= exp [2?7?/2—zytr].
We combine (4.2.27) and (4.2.28) and get
oO ro 1 0
a ge be(t)
t dt a
pm
ee
elie [itx
rtayr
—t?/2+t27?/2—tytr]
#2 242 ry
dtP

We then see easily that

4.2.29 = |)(* eet = : (x—ry)?

|-o eI.
(122) Pa gee = 8 oe
We conclude from (4.2.26) and (4.2.29) that
1 eo
(SUE i) Sr= [seen oa al
x—ry)?

This is the first part of the statement; the relation

eae a,
follows immediately from (4.2.30) and the extension of Helly’s second
theorem, so that lemma 4.2.3 is proved.
To prove that the condition of theorem 4.2.6 is sufficient, we assume that
Con (P) 2 0

for every p > 0 and apply lemma 4.2.3 to the function (x) = f( px).
Since f(x) is, by assumption, a real and even function, we have

Gent = [ f((px) es (ate = OF 1)

while
Ayn = (- Lye Con (Pp),

and we conclude from lemma 4.2.3 that

lim ¥ (= 1)" Con (p)r?”
Hon (¥)
r>1 A=0 (2n)! V2 = f (py).
Therefore

(4231) in
ee
5 ConET(PI (1M
2/9
Hal3)= fonderP
2

Since the function (—1)"e~””? H,,(y) is, according to lemma 4.2.2, non-
negative definite, we conclude from (4.2.31) that f( py) e7%”? is also non-
negative definite. But then the same is true for f(y) exp (—y2/2p?);
moreover, we see from assumption (ii) of the theorem that this is a charac-
teristic function. We let p -> oo and conclude from the continuity theorem
CRITERIA 83
that f(y) is a characteristic function. The following corollary is an im-
mediate consequence of the preceding reasoning.

Corollary to theorem 4.2.6. If condition (ii) of the theorem is not satisfied

for allp > 0 but iffor somepy > 0 we have Con(po) > Oforn = 0,1,2,...
ad inf., then f(y) exp (—y?/2po) ts a characteristic function.

4.3 Sufficient conditions

The necessary and sufficient conditions discussed in the preceding
section are often not readily applicable. In the present section we will give
a very convenient and useful sufficient condition.

Theorem 4.3.1 (Pélya’s condition). Let f(t) be a real-valued and continuous

function which is defined for all real t and which satisfies the following con-
ditions:
(i) FO) = 1
(ii) f(—9 = f(0
(iii) f(t) convex) for t > 0
(iv) lim f(t) = 0.
to

Then f (t) is the characteristic function of an absolutely continuous distribution

Since f(t) is a convex function it has everywhere a right-hand derivative

which we denote by f(t). The function f’(t) is non-decreasing for t > 0.
It follows from (iv) that f(t) < 0 for ¢ > 0 and that
lim f(t) = 0.
t—>00

It is easily seen that the integral | e “* f(t) dé exists for all « 4 0.

We write

(43.1) p(s) = 5 |oo

= 1 a —tex

We see from (ii) and (4.3.1) that

1 co

(4.3.2) p(x) = i |.f(t) cos tx dt.

(*) A function f(t) is said to be convex for t > 0 if

cies) < iat

for all t; > 0, t2 > 0. For a survey of the properties of convex functions we refer the
reader to G. H. Hardy-J. E. Littlewood-G. Pélya (1934), 70-72, 91-96.
84 CHARACTERISTIC FUNCTIONS

The conditions of Fourier’s inversion theorem'* are satisfied and we

obtain

oye fo eee) de
It follows from (1) that |R p(x) dx = 1 and the proof of theorem 4.3.1
is completed as soon as we show that p(x) is non-negative.
Integrating by parts and writing g(t) = —/f’(t) we get
1 ice)

(4.3:3)i
Roe 0 wp) = ee
— |.g(t)
t) sin
sin xxt dt

where g(t) is a non-increasing, non-negative function for ¢ > 0 while

lim g(t) = 0.
to

Then
n/n ioe) s

= J|0 |S
P(x) = TX (-1)'e(+)| sin tx dt.
j=0 x
Let x > 0; the series

»? (= 1)'e(2+)
j=0 x
is an alternating series whose terms are non-increasing in absolute value;
since the first term of the series is non-negative one sees that the integrand
is non-negative. Thus p(x) > 0 for x > 0. Formula (4.3.2) indicates that
p(x) is an even function of x so that p(x) > 0 if « # 0. Therefore p(x) is a
frequency function and f(t) is the characteristic function of the absolutely
continuous distribution F(x) = | p(y) ad.
We will occasionally call functions which satisfy the conditions of
theorem 4.3.1 Pélya-type characteristic functions.
From the preceding proof it is clear that the frequency function p(«) of
a Pdlya-type characteristic function f(t) can always be obtained by means
of the Fourier inversion formula (4.3.1), even if the condition of theorem
3.2.2 that f(t) should be absolutely integrable is not satisfied.

(*) We use here the following theorem due to Pringsheim [Titchmarsh, (1937), p. 16]:
If the function f(t) is non-increasing over (0, ©) and if it is integrable over every finite
interval (0, a) (where a > 0) and if lim f(t) = 0 then the inversion formula
to

$f (¢+0)+ f(t—0)] = (2/7) | cos tu | f(y) cos yu dy |

0 0

holds for any positive ¢. A short proof of Pringsheim’s theorem can be found in M. Riesz—
A. E. Livingstone (1955).
CRITERIA 85

We list next a few Polya-type characteristic functions.

(4.3.4a) f(t) = e~"
1
(4.3.4b) 1) isa]

Lit) dor Of<alelraed

(4.3.4c) Tita s 1
41t|
—— fior |¢| =
pol

1—|t| for |¢| <1

(4.3.44) f(t) = {
0 for'|z (271.
Using the inversion formula (4.3.1) we see easily that (4.3.4a) is the
characteristic function of the Cauchy distribution. The characteristic
functions (4.3.4a) and (4.3.4d) are absolutely integrable; however (4.3.4b)
and (4.3.4c) are examples of characteristic functions of absolutely con-
tinuous distributions which are not absolutely integrable (see page 34).
The corresponding frequency functions can nevertheless be computed by
means of formula (4.3.1) but lead to higher transcendental functions. The
frequency function of the characteristic function (4.3.4d) is the function
asae yi| ip

Poélya’s condition permits us to construct examples which help us to get

a better insight into the assumptions of the uniqueness theorem.'*)

Example 1. Let f(t) be any Pdlya-type characteristic function whose

right-hand derivative f’(t) is strictly increasing for t > 0. Replace an
arbitrarily small arc of the right-hand side of f(t) by a chord and change
the left-hand side symmetrically. In this manner one obtains a new func-
tion f, (t) which also satisfies the conditions of theorem 4.3.1. Thus f, (¢)
is a Pélya-type characteristic function which agrees with f(t) everywhere,
except on two symmetrically located arbitrarily small intervals. As a con-
sequence of the uniqueness theorem /, (¢) and f(t) belong to two different
distributions.

Example 2. Let f, (t) be the characteristic function (4.3.4c) while f, (#)

is the function (4.3.4d). These are examples of two characteristic functions
which agree over a finite interval but belong to different distributions.
Pélya’s condition can be used to derive another sufficient condition
which is applicable to certain periodic functions.

(*) See page 29.

86 CHARACTERISTIC FUNCTIONS
Theorem 4.3.2. Let f(t) be a real-valued function which satisfies the follow-
ing conditions'*) :
(i)f(0)=1
(ii) f(—4) = f()
(iii) f(t) is convex and continuous in the interval (0, r)
(iv) f(t) is periodic with period 2r
(v)f(r) = 0,f(®)> 0 in(0,7].
Then f(t) is the characteristic function of a lattice distribution.
We consider the function f, (t) defined by

Aas) ences:
cada 2)t)
=
{0 ele) = 7.
Clearly, f, (¢) satisfies the conditions of Pélya’s theorem and is therefore a
characteristic function. Hence it follows from (4.3.1) that

i e- f,(t)dt > 0
for all x. Combining this with (4.3.5) we obtain

(4.3.6) | f(t) cos txdt > 0

while rca (11) implies that

(4.3.7) Is f(t) sin txdt = 0.

We subenetron = an/r (n integer) in (4.3.6) and (4.3.7) and see that

An, = f(t) cos edt > 0

(4.3.8) hse
tae . nn
B,, ~ [Fo Se tdt = 0.

The quantities A, and B,, are the Fourier coefficients of the function f(t).
It follows from Dirichlet’s conditions [Titchmarsh (1939)] that f(t) is
equal(t) to its Fourier series in the interval (—r, +7). On account of the
periodicity of f(¢) one has then
A fos)

(43.9) f()=52+
DA,cost n=1

(*) P, Lévy (1961) showed that condition (iii) can be replaced by (iii’): the functiong (2),
which equals f(t) in some interval (0, 7) and is zero for t > 1, is a characteristic function.
(t) D. Dugué (1955), (1957b) investigated the Fourier series of a characteristic function
and showed that a characteristic function is, in a certain interval, equal to the sum of its
Fourier series. L. Schmetterer (1965) supplemented these results and showed that a
similar statement is true if the trigonometric system is replaced by certain orthogonal
systems.
CRITERIA 87

for any real value of ¢. Formula (4.3.9) indicates that f(z) is the charac-
teristic function of a lattice distribution whose lattice points are the points
est7) (i= Ol oD)
The functions f(¢) and f, (t) discussed in this proof are also examples of
characteristic functions which agree over a finite interval. The first ex-
ample of this kind is due to Khinchine.
Extensions of Pédlya’s condition can be found in Girault (1954) and in
Dugué (1957b). These authors also obtained some interesting results
concerning Polya-type characteristic functions [D. Dugué—M. Girault
(1955)].
We discuss here only one of their theorems.
Theorem 4.3.3. A characteristic function is a Pdélya-type characteristic
function if, and only if, it can be represented in the form
Pag du
10) f(t) = | a(<)dF (x)
0

for t > 0 and f(t) = f(—t) for t < 0. Here

1—|t| if|t| <1
4.3.11 R(t) =
( ) ) {o ieee
while F(x) 1s a distribution function such that F(0) = 0.
t
We note that k(t) is the characteristic function (4.3.4d) so that a(*)

is also a Polya-type characteristic function, and we see easily that f(Z), as

given by (4.3.10), satisfies the conditions of theorem 4.3.1 and is therefore
a Pélya-type characteristic function. We prove next that every Polya-type
characteristic function admits a representation (4.3.10). Let f(t) be a
Pélya-type characteristic function; we mentioned above that f(t) has
everywhere a right-hand derivative f’(t) which is non-decreasing for ¢ > 0.
We note that f(x)—«f’(«) is the ordinate at the origin of the tangent
of the curve y = f(t) taken to the right of the point ¢ = x. Therefore
f (x) —xf'(«) tends to zero as x > oo, We use integration by parts to show
i
that

[ (-Sat-sey taro = -1f°5[OS3 ] ae.

The symbol = stands here for the right-hand derivative. It follows

immediately that
(4.3.12) i (1-‘)an — f(x) +2f'(#)] = f(i).
We see that
D
F(x) = 1— f(x) +9f"(a)
88 CHARACTERISTIC FUNCTIONS

is a distribution function and introduce F(x) and the function (*) into
x
formula (4.3.12) and obtain the desired result.
The decision whether a given function f(t) is a characteristic function
can sometimes be made by means of the results derived in earlier chapters.
The continuity theorem is frequently useful in this connection; we con-
sider next a simple example. Let
18 nl
Iie cosht e+e *
The function cosh ¢ is an entire function which has zeros at the points
in(2j —1)/2. Applying Weierstrass’ theorem on the factorization of entire
(integral) functions we get

cosh ¢ = ek E ue |
j=1 (2; —1)? 2?
so that

f()
t) ==Tes)
' g(t where
TT h 2,()
ihe)=[1+ 5"
ee de s |fs,
iy ei ee
Let I(t) = 1/(1+¢#?) be the characteristic function of the Laplace distribu-
tion, then
2
; = /f{ ===

is also a characteristic function. Then [corollary 1 to theorem 3.3.1]

h,,(t) ak TTsi(0

is also a characteristic function. Finally we conclude from the continuity

theorem that
f(0)= limhy(0
is also a characteristic function. In a similar way one can show that the
reciprocal of an entire function of order 1 which has only purely imaginary
zeros and which equals 1 at the origin is always a characteristic function.

4.4 Supplementary remarks concerning non-negative definite

functions
In the preceding section we saw (Example 2 on page 85) that two differ-
ent characteristic functions can agree over a finite interval. This observa-
tion motivates the introduction of a new concept, namely of functions
which are non-negative definite on a finite interval.
CRITERIA 89

A complex-valued function f(t) of a real variable ¢ is said to be non-

negative definite over the interval (— A, A) if
(i) f(t) is continuous in (— A, A);
(ii) for any positive integer N and any real numbers 4, fa, .. ., ty
such that |t;| < A (j = 1, 2,..., N) and any complex num-
bers £,, &,,..., €y the sum
N N .
XD f(ti—te)EsSe
§=1 k=1
is real and non-negative.
We denote the set of functions which are non-negative definite over
(—A, A) by Fa and write 7, for the set of functions which are non-
negative definite over (— 00, 00).
Bochner’s theorem and the example mentioned above suggest several
problems. The first of these is a characterization of the class Pa, the second
deals with the possibility of extrapolating a function non-negative definite
on (— A, A) to the whole real line. Finally, one is interested in conditions
for the uniqueness of this extension. A number of authors, M. Krein
(1940), (1943), D. A. Raikov (1940), E. J. Akutowicz (1959), (1960), and
P. Lévy (1961), have investigated these problems and obtained interesting
solutions. The tools used in these investigations exceed the scope of the
methods employed in this monograph. Therefore we only list here some
of the results, without proofs. Omission of these proofs will not cause any
difficulty in reading the book since the present section is only loosely
connected with the rest of the monograph.

Theorem 4.4.1 (Krein’s theorem). A function f(t), defined on a finite or

infinite interval (— A, A), belongs to P4 if, and only if, 1t admits the repre-
sentation

(44.1) f= |"eaF(a) (It < 4)

where F(x) is a non-decreasing function of bounded variation.
Theorem 4.4.1 is due to M. Krein (1940); an elementary proof was
given by D. A. Raikov (1940). In the case where A is infinite, Krein’s
theorem reduces to Bochner’s theorem. The integral representation (4.4.1)
is unique if A = oo; however, for a finite interval F(x) is in general not
uniquely determined by f(z). This means that a function f(t) which is
non-negative definite over a finite interval may admit several different non-
negative definite extensions to the full real line. Conditions for the unique-
ness of the extension can also be found in Krein’s paper. To formulate
these we introduce the following terminology.
Let f(t) € Pa and denote the set of non-decreasing functions of bounded
90 CHARACTERISTIC FUNCTIONS
variation which admit the representation (4.4.1) for f(t) by V;. We norm
the functions of V, so that F(— 0) = 0 while
LG) =|f@-0)er
a) 2:
We say that the extension (extrapolation) of f € Ay is unique if V, contains
only one element; otherwise we say that the extension of f is indeterminate.
Let Ba be the set of all entire functions g(z) of the complex variable
= = t+zy such that
(i) —ao<t<o
sup |a(é)| < ©
(ii) lim sup r-t log M(r) < A,

where M(r) = max |g(z)|. This means that g(z) is an entire function of
|z| <r

exponential type not exceeding A and is bounded on the real axis. Let
£€Sa, fe Pa and Fe V,. We define a functional ®, in the space Ba
by the formula

®/(g)= | aOdF@.
It can be shown that ®, depends only on f and g but is independent of F.
We are now in a position to formulate Krein’s result.

Theorem 4.4.2. The extension of the function f € P4 is unique tf there exists

a non-negative function g € Ba, g #0 such that O,( g) = 0.
Krein also gave a necessary and sufficient condition for the indetermin-
ateness of the extension of a function which is non-negative definite in a
finite interval.

Theorem 4.4.3. The extension of the function f € 74 ts indeterminate if, and

only tf, the following two conditions are satisfied :
(1) For every function t(x) of bounded variation for which

a(x) = 3[t(x+0)+ t(x—0)] # const (0 < x < A)

the inequality

|.|.fs dre ar(y) > 0

A A

holds.
(ii) The series
2
D An [ eb, (x) dx
A >

<"@
n=1 v0

for allrealt(—2o<t< o),

CRITERIA 91

Here {¢,(x)} ts a complete orthonormal system of fundamental solutions of

the integral equation

Ha) = 4) fle-y) Hy) dy

A

and {ji,,} is the corresponding sequence of eigenvalues.

We conclude this section with a remark concerning a generalization of
Bochner’s theorem. A non-negative definite function is by definition con-
tinuous. F. Riesz (1933) replaced the assumption of continuity by measur-
ability and showed that the representation of such a function by the
Fourier—Stieltjes integral of a bounded non-decreasing function is valid
almost everywhere. M. Crum (1956) carried out a detailed investigation
and obtained the following result.

Theorem 4.4.4. Let f(t) be a complex-valued function of the real variable t

(— 0 < t < «) which satisfies the following conditions:
(1) f(t) ts measurable;
(11) for any positive integer N and any real t,, tz, ..., ty and any
complex &,, €,..., €n the relation
N N o
S MG 1) ese 2 0
j=1 k=1

holds.
Then f(t) admits the decomposition f(t) = $(t)+y(t), where

we) = |" ear),

while w(t) = 0 almost everywhere and satisfies condition (11). Here F(x) is a
bounded and non-decreasing function.
Conditions generalizing Cramér’s criterion in a similar way were given
by G. Letta (1963).

4.5 Unimodal distributions

A distribution function F(x) is said to be unimodal if there exists at
least one value x = a such that F(x) is convex for x < a and concave for
% > a. The point x = a is called a vertex of F(x).
As examples of unimodal distributions we mention the normal distribu-
tion and the Cauchy distribution.
In this section we study properties of unimodal distributions and derive
criteria which assure that a characteristic function belongs to a unimodal
distribution. For the sake of simplicity we shall often assume that the
vertex is the point a = 0.
92 CHARACTERISTIC FUNCTIONS

Theorem 4.5.1. A distribution function F(x) is unimodal with vertex at

x = 0 af, and only if, its characteristic function f(t) can be represented as

A) =1)ewdu (0 <4 < 0)

where g(u) is a characteristic function.
Theorem 4.5.1 is due to A. Ya. Khinchine; for its proof we need two
lemmas:

Lemma 4.5.1. The relations

(i) lim »| 4G(y) _ 9 lim +| 4G(Y)

_6
+0 ©

aw—>-+ co x y t—>— © ey,

(ii) lim « f LOONIE) ee eg

x +00

2—>—0 — 0 y 2—->+0 x y

hold for an arbitrary distribution function G (x).

For x > 0 we have the inequalities
0< =| oni) [-acw.
+ 00

a a
The last integral tends to zero as x tends to infinity, and we obtain the first
equation in (i); the second equation in (i) is derived in the same way. Let
x —> +0 and note that
qa |,
+o oe Va ieee +o geet
« bd J Ney
; Vy x [rape
< i dG(y)+ Vx hes dG(y)
= G(Vx)—G(«)+ Vx(1—G(Vx)).
The expression on the right-hand side tends to zero as x > +0, so that
the second equation in (ii) follows. 'The first equation in (ii) is proved in
the same way.

Lemma 4.5.2. If a distribution function F(x) is convex in (— 0,0) then

there exists a function p(u) which is non-decreasing and integrable on (— 0, 0)

such that F(x) = | p(u) du. Similarly, if a distribution function F (x) ts

concave in (0, + 0) then it admits a representation F(x) = 1-{ q(u) du

x

where q(u) is non-increasing and integrable on (0, + 0).

The first statement of the lemma follows from a well-known property
of convex functions [see A. Zygmund (1952), 4.141]; the second statement
is an immediate consequence of the first part of the lemma.
CRITERIA 93
We now proceed to the proof of theorem 4.5.1 and show first that the
condition is sufficient. Suppose therefore that a characteristic function
f(t) can be represented in the form

(4.5.1) J@= feta

where g(u) is the characteristic function of some distribution G(x). We

introduce the function

le |-|" ac) dy forx <0

(4.5.2) F(x) = —a

ireHe a) dy+G(+0) for x > 0.

We rewrite (4.5.2), using integration by parts and lemma 4.5.1, and obtain

G(x)—x | — for x < 0

(4.5.3) E(x) =
G(a)+e | eee for x > 0.

It follows immediately from (4.5.2) that F'(«) is non-decreasing for x < 0

and also for x > 0. We also conclude from (4.5.3) and lemma 4.5.1 that
F(+ 0) = 1 while F(— ©) = 0; moreover we see in the same way that
F(+0) = G(+0) and F(—0) = G(—0). The function F(x) is therefore a
distribution function. Since F(x) is the integral of a non-decreasing
function in (— 0,0) and is the integral of a non-increasing function in
(0, 0) it is easily seen that (x) is unimodal with vertex at x = 0.
In order to complete the proof of the sufficiency of the condition of
theorem 4.5.1 we must show that f(t), as given by (4.5.1), is the charac-
teristic function of the distribution function F(x). According to (4.5.2)
we have
[-_ etaF (x) = — fe SO) des P(+0)- F(-0)

vai le znlk
We note that F(+0)—F(—0) = G(+0)—G(-—0) and using integration
by parts and lemma 4.5.1, we see after some elementary computations that

(4.5.4) [-_éar() = | ee
TF BiG a).
We note that
ite t
pak = “I é™ du;
Lodo
94 CHARACTERISTIC FUNCTIONS

we substitute this into (4.5.4) and obtain

[etary = 1 few du = 70,

so that the condition of the theorem is sufficient. We still have to show that
(4.5.1) is necessary and assume therefore that /'(«) is a unimodal distribu-
tion. We see then from lemma 4.5.2 that there exists a non-decreasing
function p(u), defined for u <0, and a non-increasing function g(u),
defined for u > 0, such that

ls p(u)du forx < 0

(4:5;5). Poe ne ae co)
i | q(u)du forx > 0.

Let

= | udp(u) forx <0

(45:6) G@)= i
1+| udgq(u) forx > 0.

Since p(x) is non-negative and non-decreasing in (— 00, 0) we see that

[ew du < xp(x) < 0 for x < 0, so that

(4.5.7) lim xp(x) = 0.

a—>—0

Moreover it follows from (4.5.6) that

G(x) = —xp(x) + F(x)
for x < 0; we conclude from (4.5.7) that
(4.5.8a) G(—0)= F(-0).
Similarly we show that
(4.5.8b) G(+0) = F(+0)
so that G(+0) > G(—0). We see from (4.5.6) that G(x) is non-decreasing
in (— o, —0) and (+0, + «), so that G(x) is a distribution function.)
Let
g(t) = | et dG (x)
be the characteristic function of G(x); then

| eda = [> <—*2cq).

1tx

(*) It might be necessary to modify G(x) at its discontinuity points to make it right
continuous.
CRITERIA 95
We substitute for G(x) the expression from (4.5.6) and integrate by parts
and obtain, using (4.5.8a) and (4.5.8b),

:| e(w)du = lise'* p(x) dx + F(+0)—F(-—0)-— [eats dx,

or, in view of (4.5.5)

*{ete du = [ear (a)= f(t),

so that the theorem is proved.

It is worthwhile to remark that P. Medgyessy (1963) has shown that it is
possible to derive theorem 4.5.1 from theorem 4.3.3 and also theorem
4.3.3 from theorem 4.5.1, so that these two theorems are equivalent. Paul
Lévy (1962) gave some interesting extensions of theorems 4.3.3 and 4.5.1.
The next theorem gives a sufficient condition which assures that a
characteristic function belongs to a unimodal distribution.

Theorem 4.5.2. Let f(t) be a continuous, real-valued and even function of

the real variable t such that f(0) = 1 and let A(z) be a function of the com-
plex variable z(z = t+iy = re”; t, y, and 6 real) which satisfies the follow-
ing conditions:

(i) A(z) is regular in the region D = {s:1 Sy ay <1) Tea}

where €, and €, can be arbitrarily small;
(ii) |A(z)| = O(1) as|z| +0;
(iii) |A(z)| = O(|z|~°) as |z| > (6 > 1);
(iv) Im A(ty) < 0 fory > 0;
(Vv) f(t) = AG). for E> 0:
Then f(t) is the characteristic function of a symmetrical unimodal and
absolutely continuous distribution function.
It follows from our assumptions that f(t) is absolutely integrable over
(— 0, + 0). Therefore

Pe) = [oe fod = =|- cos ix f(t)dt

is a continuous, real-valued and even function of x. It follows from the

preceding ee that

(4.5.9) p(x)= Re iPete f(t)

dt= “Re [”elt A(t)dt.
In order to compute the last integral in (4.5.9) we consider the function
y(z)= e** A(z) (z complex, « > 0) along the contour I’ which consists of
96 CHARACTERISTIC FUNCTIONS

the segment [7, R] of the positive real axis, the circular arc

Cp = {2:2 =Re’%,0<¢5 a

the segment [iR, ir] of the positive imaginary axis, and the circular arc

C= {2:2 = re; >¢2 of.

Since (z) is regular in ©, we can apply Cauchy’s theorem and see that

|ve) de |e4@) dz = 0
or

(4.5.10)
R r
| ev! A(t) dt + | y(z) dz+i | ee” A(ty) dy + | y(z) dz = 0.
r Cr R Cr

It is easily seen that

lim | y(z)dz = lim | v(z)dz = 0,

r—>0 JC, R>o J Op

so that one obtains from (4.5.10)

lief A(t) dt =i Iee- A(iy) dy.

0 0

We conclude from the last equation and (4.5.9) that

(oe)

(4.5.11) p(x) = ae iFe~* [Im A(ty)] dy.

It follows from condition (iv) and the assumption that f(f) is an even
function, that p(x) > O for all real x. Since f(0) = 1 we conclude that
p(x) is a frequency function whose characteristic function is f(t). To com-
plete the proof we must show that the corresponding distribution is uni-
modal. Let x, > x, > 0; it follows immediately from (4.5.11) that
P(%1) < P(%2) < p(0),
so that p(x) is a decreasing function for x > 0. Since p(x) is symmetric we
see that it has a unique maximum at x = 0.
We give next an application of theorems 4.5.1 and 4.5.2 and show that
certain functions are characteristic functions of unimodal distributions.
This result will be used in the next chapter.

Theorem 4.5.3. Let « be a real number such that 0 < « < 2; then the

function f(t) = T+ ]ef is the characteristic function of a unimodal distri-

bution.
CRITERIA 97

In proving the theorem we must distinguish three cases.

(a) 0 < « < 1. In this case we introduce the function
1+(1—«)|z|*
45:12
( ) t) = —~____.,
&( ) (1 ny |t \*)2

It is easily seen that for t > 0

at*—2
ao La") + 2(1 + 20?)
e*+ (1—02)*] > 0
The function g(t) therefore satisfies Pélya’s condition (theorem 4.3.1) and
is thus a characteristic function. It follows also from (4.5.12) that

BO = 9ft)+tf(@) fort >°0; ‘therefore’ f(t) = :io du and we con-

clude from theorem 4.5.1 that f(t) is the characteristic function of a

unimodal distribution.
(b) 1 < a < 2. In this case we introduce the function A(z) = eee
z
of the complex variable z and see that A(z) satisfies the conditions of
theorem 4.5.2, and we conclude from theorem 4.5.2 that f(t) is the
characteristic function of a unimodal distribution.
(c) a = 2. In this case the frequency function of f(t) is p(x) = de"
which has a unique maximum at x = 0.
1
The fact that f(t) = iste is a characteristic function was established
by Yu. V. Linnik (1953), and the unimodality of the corresponding dis-
tribution was shown by R. G. Laha.
Theorem 4.5.4. Let {F,,(x)} be a sequence of unimodal distribution functions
and suppose that the distributions F,,(x«) converge weakly to a distribution
function F(x); then F(x) is also unimodal.
Let a, be a vertex of F,,(x) anda = lim a, We consider first the case
where |a| < o and select a subsequence my such that ay Gy,= a. Let
X1, X, and (x,+.x,)/2 be continuity points of F(x) afte Fhe! x, <a and
%_ <a. For k sufficiently large one has x, < Gp, X, < Gy, Since F,,, (x)
is assumed to be unimodal we see that
Fr, (1) + Fr, (tg)>2Fn (44s),
We go with & to the limit and obtain

(4.5.13a) -F(x,)
+F(x.) > ee
98 CHARACTERISTIC FUNCTIONS

We assumed that x, and x, are continuity points of F(«); however, in-

equality (4.5.13a) is also true for arbitrary points. This follows from the
right-continuity of F'(«) and the fact that an arbitrary point can be approxi-
mated by a sequence of continuity points. In a similar way we obtain for
points x, and x,, such that x; > a, x, > a, the inequality

(4.5.13b) -F(x;)
+F(x) < pr eiees),
Z
It follows from equations (4.5.13a) and (4.5.13b) that F(x) is unimodal.
Finally we note that |a| is necessarily finite, otherwise /(x) would be
either convex or concave for all x. This is not possible, since a distribution
function is monotone and bounded.

Theorem 4.5.5. The convolution of two symmetric and unimodal distribu-

tions is symmetric and unimodal.
For the proof of the theorem we need the following lemma.

Lemma 4.5.3. Let F(x) be a unimodal distribution function. Then there

exists a sequence {F,,(x)} of absolutely continuous unimodal distribution
functions such that Lim F,,(x) = F(x) and F,,(x) is absolutely continuous.

We first note that if Lim F,,(x) = F(x) and Lim F,,,,(x) = F,(x),
n—> oo mo

there exists a sequence such that Lim F,,,,,,(x) = F(x). It is therefore

k>o
sufficient to consider distribution functions /'(«) whose derivative is a step
function. In this case there exist absolutely continuous functions p,, (x)
which do not decrease on (— 00, 0) and do not increase on (0, 00), such that

tim |" p(y = | B'day.

We write
1 e

F,, (x) = ee paya [play

and see that Lim F(x) = F(x), so that the lemma is proved.
n>

We proceed to prove the theorem. Let F(x) and F(x) be two sym-
metric unimodal distributions and denote their convolution by F(x),
F(x) = F,(x)*F, (x). Clearly F(x) is also symmetric, so that we have only
to show that it is unimodal. It is no restriction to assume that F(x) and
F,(x) are twice differentiable; this follows from theorem 4.5.4 and the
fact that—according to lemma 4.5.3—we can approximate F(x) and
F(x) by two sequences {F,,,(x)} and {F,,,(x)} respectively of twice
differentiable, symmetric, and unimodal distribution functions.
CRITERIA 99

If F,(«) and F(x) are twice differentiable, we see that

F(x) = | FY («—1) Fi (t)dt = | F' (x—t)
Fi(t)dt

(4.5.14) F(x) = | C@-9- Fy (+9) Fi (t)dt.

Since F(t) and F(t) are both unimodal we conclude that
(4.5.15) Pot) aot fort = U.
(4.5.16) fa atonal < 0 boas 244)
Fi(x—)—Fi(x+t) 20 ifx > 0.
It follows from (4.5.14), (4.5.15) and (4.5.16) that F(x) > 0 if x < 0
but F'’’(x) < 0 if x > 0. This completes the proof of the theorem.
Remark. ‘The assumption of theorem 4.5.5 that F,(«) and F,(«) are
symmetric is essential. The convolution of two unimodal distribution
functions is in general not unimodal. K. L. Chung (1953) gave an example
of an absolutely continuous distribution function F(x) which has its vertex
at x = 0 and which has the property that the density function of /* F has
two maxima so that /’*F is not unimodal.
I. A. Ibragimov (1956a) calls a distribution function strongly unimodal
if its convolution with any unimodal distribution is unimodal. He obtained
the following result:

Theorem 4.5.6. A (non-degenerate) unimodal distribution function F (x) 1s

strongly unimodal if, and only if, F(x) is continuous and tf log F(x) 1s con-
cave on the set of points at which neither the right-hand nor the left-hand
derivative of F(x) vanishes.

4.6 An essential property of characteristic functions

We have already mentioned that every distribution function has a
characteristic function and have discussed in Chapter 3 some other very
important theorems concerning characteristic functions, such as the
uniqueness theorem, the convolution theorem, and the continuity theorem.
The great usefulness and importance of characteristic functions in prob-
ability theory is largely explained by the fact that these properties make
them a very convenient tool for the solution of many problems. ‘The
present section deals with the question whether there are any other integral
transforms of distribution functions which have these properties.
Let G(x) be a distribution function and consider its integral transform
by means of the kernel K (s, x), that is

CNS De \@_K(s,x)dG(x),
100 CHARACTERISTIC FUNCTIONS

In the following we denote by G(x), G,(«) distribution functions and by

G1(S), G(s) their respective transforms (4.6.1).
In this section we show that the uniqueness and convolution properties
essentially determine the kernel. The following theorem gives a precise
formulation of this statement.

Theorem 4.6.1. Suppose that a kernel K(s, x) satisfies the following con-
ditions:
(I) K(s, x) is a complex-valued function defined for all values of
the real variables s and x and is bounded and measurable in x.
(II) (Uniqueness property): g1(s) = G2(s) if, and only if, G(x)
= G,(x).
(III) (Convolution property): If
G(s) G,4Gy= | G, (x—2)dG,(2)
then g(s) = gx (5) g2(s).
Then K(s, x) has the form
K(s, x) = pitAls

where A(s) is a real-valued function of s such that the values assumed by

|A(s)| form a set which is dense on (0, + 0). The converse statement is also
true.
We see from assumption (1) that every distribution function G(x) has a
transform given by (4.6.1). We write assumption (III) in terms of the
kernel and obtain

[KG x) dy | G@—) aGa(u)

= [KG t) dG, (t) | (s, u) dG, (u)

= | | K(s, t)K(s, u) dG,(t)dG, (u).

On the other hand

[KG x) de ee G(x—u) dG, (u)

a | | K(s, t+) dG, (1) dG, (u)
so that

(4.6.2) |- [KG t+) dG, (t)dG',(u)

£ |; |f K(s, t)K(s, u) dG, (1) dG, (u)
CRITERIA 101
holds for every pair of distributions G,(x) and G,(x). Let & and 7 be
arbitrary real numbers and put
G = e(x—
(4.6.3) { sity nss pee
G2 (x) = a[e(x) +e(x—6)].
Substituting (4.6.3) into (4.6.2) we get
(4.6.4) K(s,n)+K(s,n+é) = K(s,7)[K(s, 0)+K(s, &)]
for any real and 7. We obtain in particular for € = 0
2K (s,n) = 2K(s, 0) K(s, 7).
Therefore (4.6.4) reduces to
(4.6.5) K(s,9+8) = K(s,7)K(s, 8).
It is known [see for instance Hahn—Rosenthal (1948), pages 116-118]
that every measurable solution of the functional equation
v(n+&) = y(n) v(é)
has the form y(£) = e® where C is a constant. Since K(s, x) is by assump-
tion (I) measurable in x, every solution of (4.6.5) is of the form
OO eal (Six) 2,
Let p(s) = B(s)+72A(s); then |K(s, x)| = e*8®. Since K(s, x) is bounded
we have B(s) = 0. The kernel therefore has the form
(4.6.7) Kig aes eee
The transform (4.6.1) of a distribution G(x) is therefore

(4.6.8) g(s) = {.- 49 dG (x)

while the characteristic function g(t) of G(x) is

(4.6.9) g(t) = | dG (x).

It follows from (4.6.8) and (4.6.9) that
(4.6.10) gfA(s)] = 9(9).
We show next by an indirect proof that |A(s)| must assume all values
of a set dense in (0, + 00). Suppose tentatively that |A(s) |omits an arbit-
rary interval J = (a, a+h) on (0, + ©) and denote by J’ = (—a, —a—h)
the interval which is symmetric to J with respect to the origin. It is
then possible to construct two Pdlya-type characteristic functions g;(t)
and g,(t) which agree everywhere except on J and J’, The two correspond-
ing transforms (4.6.1) are g;(s) = g;[A(s)] (7 = 1, 2). Since |A(s)| does
not assume values of J we see that g,(s) and g,(s) agree for all values s but
belong to different distribution functions in contradiction to the unique-
ness assumption (II).
102 CHARACTERISTIC FUNCTIONS

We still have to prove the converse statement. Suppose that the kernel
is given by (4.6.7), then it is immediately seen that (I) holds. The proof of
(II) can be carried out in the customary manner with the aid of Weierstrass’
approximation theorem. Finally it is easy to show that (II1) is also satisfied.
We see therefore that an integral transform (4.6.1) which is defined for
every distribution function and for which the uniqueness and the con-
volution theorems hold, is obtained from the characteristic function by a
simple change of the variable. We note that we have arrived at this con-
clusion without using the continuity theorem.
This fact can be used [see E. Lukacs (1964)] to obtain a more general
characterization of the transform (4.6.8) which also uses the continuity
theorem but considers a linear mapping of the space of distribution
functions onto a set of bounded continuous functions instead of the in-
tegral transform (4.6.1).
 » FACTORIZATION PROBLEMS—INFINITELY
DIVISUBLE, CHARACTERISTIC FUNCTIONS

5.1 Preliminary remarks on factorizations

We showed in Chapter 3 that the product of two characteristic functions
is always a characteristic function. It is therefore obvious that some
characteristic functions can be written as products of two or more charac-
teristic functions.
Every characteristic function f(t) can be written as the product of the
two characteristic functions f,(t) = e”” (m real) and f,(t) = f(t)e7™.
We say that the representation of a characteristic function as the product
of two characteristic functions is trivial if one of the factors has the form
e. In order to avoid trivial product representations, we introduce the
following definition.
A characteristic function f(t) is said to be decomposable if it can be
written in the form
G11) fOM=AODAM
where f,(t) and f,(¢) are both characteristic functions of non-degenerate
distributions. We then say that f, (¢) and f,(t) are factors of f(t).
A characteristic function which admits only trivial product representa-
tions is called indecomposable.
We show next that there exist indecomposable characteristic functions.

Theorem 5.1.1. Let F(x) be a purely discrete distribution function which has
only two discontinuity points. Then its characteristic function is indecomposable.
We see from the corollary to theorem 3.3.3 that the components of F(x)
are necessarily purely discrete distributions with a finite number of dis-
continuity points. The inequality, given in this corollary, indicates that at
least one of the components must be a degenerate distribution. This
proves our assertion.
The factorization of a characteristic function into indecomposable
factors is somewhat similar to the factorization of integers into prime
factors. This is the reason why the theory of the decomposition of charac-
teristic functions is often called the arithmetic of distribution functions.
However, this analogy does not go very far; as an illustration we give an
example which shows that the factorization of a characteristic function
into indecomposable factors is not always unique.
104 CHARACTERISTIC FUNCTIONS
5
Example. Let f(t) = 45 e“ and write f,(#)= #(1+e"%+e™),
j=0
f(t) = A +e), git) = Sl tet+e™), gat) = a1 +e).
It follows from theorem 2.1.3 that the functions f,(z), f(t), £1 (4), £2 (4)
and f(t) are characteristic functions. Moreover it is easily seen that
f(t)= fi) fe(t)= £1 (2) 22(t). We conclude from theorem 5.1.1 that f.(¢)
and g,(t) are indecomposable. It follows from theorem 3.3.3 that a fac-
torization of g,(¢) must necessarily have the form
(5.1.2) gi (é) = [pe +(1—p)e™] [gem
+(1—g) ec]
where
(52143) 0 op fh BelenFl:
As a consequence of (5.1.2) p and q must satisfy the relations
PY = (15) a= DU a) tal Oy ae
which are incompatible with (5.1.3). Thus! g,(@) is indecomposable, and
since f,(t) = g,(2t) we see that f(t) is also indecomposable.
We give a second example which emphasizes another difference between
the arithmetic of distribution functions and the factorization of integers.
This example is due to B. V. Gnedenko.
Let f(t) be a real-valued periodic function with period 2 which is
defined by f,(t) = 1—|#| in the interval |¢| < 1. According to theorem
4.3.2 the function f, (tf) is the characteristic function of a lattice distribu-
tion. Let further
1—|t| for|¢| <1
fil) = 3 for |t| > 1.
Clearly f,(t) is the Pdlya-type characteristic function A(t) defined by
(4.3.11), and f,(¢) agrees with /f,(¢) in the interval |t| < 1. According to
a remark of A. Ya. Khinchine, this example shows that it is possible to
find two different characteristic functions f,(¢) and f,(t):
(5.1.4) fOM=AOh® =hOh().
This fact—sometimes called the Khinchine phenomenon—shows that the
cancellation law is not valid in the arithmetic of distribution functions.
It is known that the quotient of two characteristic functions is in general
not a characteristic function (an example is given on p. 194).
We see from (5.1.4) that the quotient of two characteristic functions [in
our example: f(t)/f,(¢)] need not be uniquely defined even in cases when
it is a characteristic function. Formula (5.1.4) indicates that there might
be a connection between the possibility of a factorization of type (5.1.4)
and the fact that one of the factors vanishes outside an interval. The
possibility of constructing characteristic functions which admit factoriza-
tions of the form (5.1.4) was investigated by T. Kawata (1940), and we
now give some of his results.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 105

Polya’s theorem shows that it is possible to construct characteristic

functions vanishing outside a finite interval. In the following we present
a different method for the construction of such functions.

Theorem 5.1.2. Let 6(u) be a positive, non-decreasing function defined on

(0, 00) such that

(5.1.5) ee
ee en
and let b be an arbitrary (but fixed) positive number. Then there exists a dis-
tribution function F (x) which satisfies for every a the relation
(5.1.6) F(—x+a)—F(—x—-a) = Ofexp [—6(x)]} (as * > oo)
and whose characteristic function f(t) vanishes for \t| > b.
For the proof of this theorem we need the following lemma which is due
to A. Ingham (1936) and N. Levinson (1936), (1938), and which we state
here without proof.

Lemma 5.1.1. Let 6(u) be a positive, non-decreasing function which satisfies

(5.1.5) and let b be an arbitrary, fixed positive number. Then there exists a
non-null function G(x) such that
(5:1:7) G(x) = Ofexp [—6(|x|)]} (as |x| > oo)
which has the property that its :ourier transform

g(u) = =sh G(x) eS ax

vanishes for |u| > b.

To prove theorem 5.1.2 we consider 6(2u) instead of 0(u). The function
6(2u) has the same properties as 6(u), so that we can apply lemma 5.1.1,
replacing b by 5/2. We put

f=) eee Fax,

where

A= |"|g()|ax.
According to theorem 4.2.4 f(t) is a characteristic function, and it follows
from lemma 5.1.1 that f(t) = 0 for |¢| > 5. Using the inversion formula,
Parseval’s theorem, and relation (5.1.7), one sees by means of a simple
computation that (5.1.6) is satisfied.

Theorem 5.1.3. Let 0(u) be a positive non-decreasing function which satisfies

(5.1.5). Then there exists a distribution function F(x) whose characteristic
106 CHARACTERISTIC FUNCTIONS

function f(t) admits a factorization of the form (5.1.4). Moreover F(x)

satisfies the condition (5.1.6) for a > 0.

We consider again 6(2u) instead of 6(u) and put b = 5in theorem 5.1.2.
Then there exists a distribution function F(x) whose characteristic func-
tion f, (t) vanishes for |¢| > a. The function f, (¢) is constructed using the
function g(x) of lemma 5.1.1 and is given by
1 ice) (oe)

6.18). 4A0= a | s@aeriar (4= ie Le(x) Pde).

We see also that
(5.1.9) g(x) =0 for |x| > :
and that
F,(—*x+a)—F,(—x—a) = Ofexp [—6(|2x|)]} as |x| > o.
It follows from (5.1.8) and (5.1.9) that
1 n/2 eo
(6.1510) A@= A | o(x)
e(x +t) dx.
—n/2

We see from (5.1.9) and (5.1.10) that

(Sota t) f(z) = fi(—7) = 0:
We define a function f,(¢) by requiring that f,(¢) is periodic with period
2x and coincides with f,(t) for |t| < a. It follows from (5.1.11) that f(t)
is a continuous function of t.
Let {c,} be the sequence of Fourier coefficients of f,(¢); then
1 nm j 1 n/2 del
1 au :
ee SO |)
= — —int
dt an | € iS[sae as |dt
a os | wie mn

| art
Sea | pnje | ae ay |an.
—7/2 Saetie

It follows from (5.1.9) that

1 n/2 ’ j 2/2 em :

Oo =
Beran | se —1
ne
dx |" a0)e
Ann —imy
dy
so that
1 1/2 : 2

Therefore
(5 5i1:2) Cee 0
and
foe)

he (t) =] Py Ch en
n=— ©
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 107

Since f,(0) = f2(0) = S c, = 1, we see that f,(Z) is the characteristic

function of a lattice distribution whose discontinuity points are contained
in a set of integers. The saltus of F,(«) at the point x = n equals c,(n = 0,
t1, +2,...). Clearly f,(#) is not identical with f,(¢), and
f() = TAO? = AOA,
so that the first part of the statement is proved. Let F(x) be the distribu-
tion function corresponding to f(#); the statement that F(x) satisfies con-
dition (5.1.6) is obtained by a somewhat lengthy but straightforward
computation.
T. Kawata also obtained a condition which assures that a factorization
of the form (5.1.4) is not possible.

Theorem 5.1.4. Let F(x) be a distribution function and let 0(u) be a positive,
non-decreasing function defined in (0, ©) such that

| ae
Oe, = (0;
1

Suppose that for some a > 0 the relation (5.1.6) holds and that the charac-
teristic function f(t) of F(x) admits the factorization
fi) = AOA):
Then f,(t) is uniquely determined by f(t) and f,(t).
For the proof we refer to Kawata (1940).

5.2 Definition of infinitely divisible characteristic functions

In this section we define infinitely divisible characteristic functions and
distribution functions and also give some simple examples. The concept
of infinite divisibility is very important in probability theory, particularly
in the study of limit theorems. Since the discussion of limit theorems
is beyond the scope of this monograph, we will not be able to reveal
here the full significance of infinitely divisible characteristic functions.
However, the analytic properties of this class of characteristic functions are
of independent interest and will be studied in this chapter.
A characteristic function f(t) is said to be infinitely divisible, if for every
positive integer 7, it is the mth power of some characteristic function.
This means that there exists for every positive integer 7 a characteristic
function f,,(¢), such that
(5.2.1) f= [A(OF. |
The function f,,(¢) is uniquely determined by f(t), f,(4) = [f(®]*",
provided that one selects for the mth root the principal branch.) The
(*) For this determination fn (t) is continuous and fn (0) = 1. It is defined in a neigh-
bourhood of the origin in which /(¢) does not vanish. We shall see that f(t) 4 0 for all ¢.
108 CHARACTERISTIC FUNCTIONS

distribution functions which correspond to infinitely divisible charac-

teristic functions are called infinitely divisible distributions.
Alternatively one could start by defining infinitely divisible distribu-
tions as distributions which can be written—for every positive integer n—
as the n-fold convolution of some distribution function. It is obvious that
this approach is equivalent to the one we used; we mention it here because
it is sometimes convenient to express infinite divisibility in terms of
distribution functions.
We give next a few examples of infinitely divisible distributions. In all
these examples f,(¢) has the same functional form as f(t) but contains
different parameters. In these cases we see immediately that f(t) is in-
finitely divisible.
Examples of infinitely divisible characteristic functions
(a) The Degenerate distribution
f(t) = et, sis(2) — eils/mt

(5) The Poisson distribution

f= ep Be}, fl) = exp {Zit}.

Anes

(c) The Negative Binomial distribution

f() = {pl -gety7¥, a(t) = {p[l—ge] 1}.
(d) The Normal distribution

A) = exp {inte}, = exp {iteFh.

(e) The Cauchy distribution
f(t) = exp {ut—6|t]}, ity exp {iS i}
(f) The Gamma distribution
f@) =0-@/)", int) = [1-@t/)-™.
5.3 Elementary properties of infinitely divisible characteristic
functions
We establish first a simple but rather important property of infinitely
divisible characteristic functions.
Theorem 5.3.1. An infinitely divisible characteristic function has no real
zeros.
Let f(t) be an infinitely divisible characteristic function; then |f, (t) |?
=| f(t)|?” is a characteristic function for any positive integer n. We
consider
g(t) = lim |fu(|* = lime |FO)
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 109

The function g(t) can assume only the two values 0 or 1 since g(t) = 1
whenever f(t) # 0, while g(t) = 0 for all t for which f(¢) = 0. The func-
tion f(t) is continuous and f(0) = 1, therefore f(¢) 4 0 in a certain
neighbourhood of the origin. In the same neighbourhood g(t) = 1, thus
g(z) is continuous at t = 0 and is, as a limit of characteristic functions, also
a characteristic function. But then it must be continuous everywhere and
we see that g(t) = 1. This means that f(t) 4 0 for all real ¢.
The characteristic function of a purely discrete distribution with two
discontinuity points (theorem 5.1.1) is indecomposable and therefore
a fortiori not infinitely divisible. But such a function, for instance
f(t) = 3(2+e"), need not have real zeros. This example indicates that the
converse of theorem 5.3.1 is not true; a characteristic function which has
no real zeros is not necessarily infinitely divisible.
Theorem 5.3.1 can be used to show that a given distribution is not in-
finitely divisible. Consider for example the rectangular distribution; its
characteristic function is

f() =e tr

Since f(t) has real zeros it cannot be infinitely divisible.

We next discuss two theorems which permit us to assert that a given
characteristic function is infinitely divisible.

Theorem 5.3.2.) The product of a finite number of infinitely divisible

characteristic functions ts infinitely divisible.
It is sufficient to prove the theorem for the case of two factors. Suppose
therefore that f(z) and g(t) are infinitely divisible characteristic functions.
Then there exist for any positive integer m two characteristic functions
Jaeand: ga(z) such/rthat yf(=f, )" ‘andi 2(t) =.[ga(2)]|".. Then
h(t) = g(t) f(t) = [gn(t) f,(2)]” so that A(t) is also infinitely divisible.
As an example we consider the Laplace distribution which has the
characteristic function f(t) = 1/(1+¢?). We can write
1 1
IO = aie
as the product of characteristic functions of two Gamma distributions
with parameters 6 = —1, A= 1 and 6 = +1, A= 1 respectively. We
know already that the Gamma distribution is infinitely divisible and con-
clude therefore that the Laplace distribution is also infinitely divisible.

(*) We give later [formula (5.5.12)] an example which shows that the converse statement
is not true.
 110 CHARACTERISTIC FUNCTIONS

Corollary to theorem 5.3.2. Let f(t) be an infinitely divisible characteristic

function, then |f (t)| is also an infinitely divisible characteristic function.
It is immediately seen that f(—f) is an infinitely divisible characteristic
function whenever f(t) is infinitely divisible. It follows then from theorem
5.3.2 that |f(£)|? is also infinitely divisible. This means that
TAME Seana ktate
is a characteristic function for any positive integer m. But this implies the
statement of the corollary.
We note that the result of this corollary cannot be improved since it is
not possible to assert that |f(t)| is a characteristic function whenever
f(t) is a characteristic function. Let for example f(t) = $(1+7e"); then
|f(t) |? = d(50+7e-*+ 7e"). We show by means of an indirect proof that
|f(¢)| is not a characteristic function. According to the corollary to
- theorem 3.3.3 it must have the form | f(t) | = ae“+(1—a)e”. Therefore
a should satisfy the relations a? = (1—a)? = 44, 2a(1—a) = 22. Since
these relations are inconsistent we conclude that |f(¢)| cannot be a
characteristic function.

Theorem 5.3.3. A characteristic function which is the limit of a sequence of

infinitely divisible characteristic functions is infinitely divisible.
Let f(t) be a sequence of infinitely divisible characteristic functions
and suppose that this sequence converges to a characteristic function f(t),
so that
(5.5.1) Fy= tim fd)
k—->oo

is continuous.
Let 2 be an arbitrary positive integer. Then |f (t) |? and |f(t) |? are real
characteristic functions and
lim | f (2) |!" = | f@) |?
ko

It follows from the continuity theorem that |f(t) |?/” is a characteristic

function; hence |f(é)|? is infinitely divisible and therefore has (theorem
5.3.1) no real zero, so that it is possible to define its mth root

(53.22) fold) = L/P! = exp {log flo}.

We write also

(53.20) A (0) = 7(} = exp {Log soo}.

It follows from (5.3.1), (5.3.2a) and (5.3.2b) that
(533) lim {2 = f.
where f,,(¢) is continuous at ¢ = 0.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 111

The characteristic functions f (t) are by assumption infinitely divisible,

so that the f,(t) are also characteristic functions. We conclude from
(5.3.3) and the continuity theorem that f,,(t) is also a characteristic func-
tion. Equation (5.3.2a) then indicates that f(¢) is infinitely divisible.

Corollary to theorem 5.3.3. Let f(t) be an infinitely divisible characteristic

function; then [f (t)]* ts also a characteristic function for any real, positive «.
The converse is also true.
If f(t) is infinitely divisible then the statement follows from the defining
property for rational « and is obtained from the continuity theorem for
arbitrary positive «. The converse is trivial.

Remark. A similar argument can be used to show that infinitely

divisible characteristic functions could have been defined in a slightly
different manner. A characteristic function is infinitely divisible if, and
only if, there exists a sequence of positive integers m, which tends to in-
finity and is such that for any & the function f(z) is the (m,)th power of some
characteristic function f;,(t).
Theorems 5.3.2 and 5.3.3 are closure theorems since they indicate that
the family of infinitely divisible characteristic functions is closed under
certain operations.

5.4 Construction of infinitely divisible characteristic functions

In this section we discuss two methods for the construction of infinitely
divisible characteristic functions. These methods give some interesting in-
formation concerning the structure of infinitely divisible distributions.
We first prove a lemma which is of some independent interest.

Lemma 5.4.1. Let g(t) be an arbitrary characteristic function and suppose

that p is a positive real number. Then f(t) = exp {p[g(t)—1]} ts an infinitely
divisible characteristic function.
Let 7 be a positive integer such that n > p. Then

falt) = (L-(oind]
+(o/npa(nyr = {14PBO—NY
is also a characteristic function. We see then from the continuity theorem

" f(t) = lim f.(0) = exp fple()—10}

is also a characteristic function. The function [f(t)]* (« > 0) satisfies the
conditions of the lemma and is also a characteristic function. We conclude
from the corollary to theorem 5.3.3 that f(t) is infinitely divisible.
112 CHARACTERISTIC FUNCTIONS

We use this lemma to prove the following theorem:

Theorem 5.4.1 (De Finetti’s theorem). A characteristic function 1s infinitely
divisible if, and only if, it has the form
IOs ane EXP {(Pml
gm(t)—1]}

where the Pm are positive real numbers while the g(t) are characteristic
functions.
The sufficiency of the condition of the theorem follows immediately
from lemma 5.4.1 and from the continuity theorem. We show next that
the condition is necessary and assume that f(t) is infinitely divisible. It
follows from the corollary to theorem 5.3.3 and from lemma 5.4.1 that

Fi@)s= exp . lf @y—m} is, for any real positive «, a characteristic

O
function. Since
f(0) = lim f,(0)
we see that f(t) can be represented in the above form with
Pm =m and gp(t) = ORs

Theorem 5.4.2. The limit of a sequence of finite products of Potsson-type

characteristic functions is infinitely divisible. The converse is also true.
Every infinitely divisible characteristic function can be written as the limit of
a sequence of finite products of Poisson-type characteristic functions.
The first part of the theorem is a consequence of the closure theorems.
To prove the second part we assume that f(¢) is infinitely divisible;
according to De Finetti’s theorem it can be represented as
OD) od ee ee eet,
where the g,,(¢) are the characteristic functions of some distributions
G,,(x), so that

iy = { e* dG, (x).
Then we see that

(5.4.2) Palla =1)= limp, A (e*— 1) dG,, (x).

A—>0o —A

We wish to approximate the integral by Darboux sums and therefore intro-

duce the subdivision
A=, Oy Sidp< eile Ky esa = FA
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 113

and write c, = p,[G, (ax) — G, (a,_,)]. Then

Du | (e"—1) dG, (x) = lim > ¢,(e%*—1)

A * N *

—A N—->o k=1

so that

(Sais seep E || 8D) dG, (0) = lim IIexp [ce(e#*—1)].

The function (5.4.3) is the limit of a finite product of Poisson-type
characteristic functions and we see from (5.4.1) and (5.4.2) that f(#) also has
this property.
The theorem can be used to show that a given characteristic function is
infinitely divisible. As an example we consider the characteristic function
ep at
f(t) = p—eé (p > 1).

It is easily seen that

A = S(1-(/pyl pyre,
and we note that this is the characteristic function of the geometric
distribution listed in Table 1. We expand

log f(t) = log (1~5)—log(1; ;)

into a series and see that

f(t)= =Thesp
3
{5
— (e*—] (e\
Litet®
—1)
We can apply theorem 5.4.2 and see that f(t) is infinitely divisible.

5.5 Canonical representations

The results of the preceding section can be used to deduce explicit
formulae for infinitely divisible characteristic functions. For their derivation
we need several auxiliary theorems.

Lemma 5.5.1. Let a be a real constant and let 6(x) be a real-valued, bounded
and non-decreasing function of the real variable x such that 6(— 00) = 0.
Suppose that a function f(t) of the real variable t admits the representation
; ii G io 1tx Jae
(6.5.1) og f(t) = ita ae Misty ener d0(x).

The integrand is defined for x = 0 by continuity, and is therefore equal to

—t2/2 if « = 0. Then f(t) is an infinitely divisible characteristic function.
Moreover, the constant a and the function 6(«) are uniquely determined by f(t).
114 CHARACTERISTIC FUNCTIONS

Let t belong to an arbitrary fixed interval; then it is seen that the inte-
grand of (5.5.1) is bounded and continuous in x, so that the integral exists
for all values of t. We first prove by repeated applications of the con-
tinuity theorem and of the closure theorems that f(t) is an infinitely
divisible characteristic function. Let 0 < ¢ < 1 and define

(5.5.2) Ljh=2 laren (<

ito 1 sala
ae) ae (x).

The function J,(t) is continuous at ¢ = 0; we can write it as a limit of

Darboux sums S,,,(¢) where
14+x}
See 3 Pa(el 1) ~ int] with a, = 2 [0(x;) — O(x;, =1)I,
Ni,

Me = .[A(%1.) — O(%,1)].

Sm(t) is the second characteristic of a product of Poisson type character-

istic functions. J,(¢) is the limit of these functions and therefore the
logarithm of an infinitely divisible characteristic function. Let now
(5.5.3) I(t) = lim J, (t)
, 2

‘ lise (1 a
146%? —xe a),
Clearly J, (t) is continuous at ¢ = 0; we conclude again from the continuity
theorem that exp [/,(¢)] is a characteristic function and then from theorem
5.3.3 that it is infinitely divisible. Finally it follows from (5.5.1) that

log f(t) = I(t) +ita—F [0(+0)—0(—0)]

The last equation shows that f(t) is the product of the infinitely divisible
characteristic function exp [I,(t)] and the characteristic function of a
normal distribution, so that f(t) is also infinitely divisible.
We show next that the constant a and the function 0(x) are uniquely
determined by (5.5.1). We write ¢(t) = log f(t) for the second character-
istic and see easily that
(5.5.4) o(t)-3[P(t+h)+¢4(t—h)]
s ib, e (1—cos xh) ta).
We now introduce the function

A(t) = ie|are) ee dh.

INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 115

Since the integrand in (5.5.4) is bounded, we can integrate with respect to h

under the integral sign and obtain
ioe) S 2

A(t) = | eile (1-2) EEN

x2
ots
We next introduce the function

A(x) = ae (1-2) =P ably)

Then

ax) = |" y (1-22) "arg

=e Il +?

while

Oe | ef dA(x).
It is easily seen that there exist two positive constants c, and cy such that
sin
1
_)heey 2
Z
ce
ee fad
The function A(x) is therefore non-decreasing and bounded; moreover
A(— co) = 0. We conclude then that A(x)/A( co) agrees with a distribution
function at all continuity points of A(x). Hence A(x) is uniquely deter-
mined by its Fourier transform A(t). Since A(t) is defined in terms of f(t),
we see that A(x) is uniquely determined by f(t). The fact that the constant
ais also determined by f(¢) is a consequence of (5.5.1).

Lemma 5.5.2. Let {¢,(t)} be a sequence of functions and suppose that

Fn(t) = exp [¢,(t)] ts determined by some constant a, and some function
6,(x) according to (5.5.1). Assume that the sequence >,,(t) converges to some
function $(t) which its continuous at t = 0. Then there exists a constant a and
a bounded and non-decreasing function 0(x) such that
(i) lima, =a
n>

(i) Lim 0,(x) = (3)

Gi) lim | dB, (x) = | d0(x).
i)

—@

The function 6(x), together with a, determines f(t) = exp [¢(t)] according
to (6.1);
The functions f,(t) are characteristic functions. Therefore f(t) = e*®
is also a characteristic function (by the continuity theorem) and ¢(t) is
everywhere continuous.
116 CHARACTERISTIC FUNCTIONS

We use the same notation as before and write

ni ('pnt fH eae fn
Ch ee (Ane SP dal)
so that
ee | ed, (x).
From the continuity of ¢(t) we conclude that the sequence 4, () converges
to a continuous function. We apply corollary 3 to the continuity theorem
(see page 52) to show that the sequence A,(x) converges weakly to a
bounded non-decreasing function A(x) and that

lim | Aaa) 3 ie dN(x).

N—> 00 —o

We have
: Sify \a ay?
a = A
Pn (x) (eae y Ly ae)
and conclude from Helly’s second theorem that

Lim 6,(3) = [| (1-S82) "2, ancy)=o(a) ay.

We write
Oe [° (#- Pin ite Jos db, (2)
Ix
and once more use Helly’s second theorem to show that
limels (1) = WA)
where
1 = [° (e- inte d0(x).
+x
From the convergence of the ¢,, (¢) and the J, (¢) it follows that the sequence
{a,} must also converge and that ¢(t) is determined by
a= lima,
no

and 6(x) according to formula (5.5.1).

Lemma 5.5.3. Let f(t) be an infinitely divisible characteristic function.

Then there exists a sequence of functions $,(t) which have the form (5.5.1)
such that
lim 44(0) = $(0) = log ().
 INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 117
We have, asn > «0,
nL f(t]
—1} = nfe/m* —1} = n{(1/n) d(t)+0(1/n)} = $(t)+0(1)
so that
(55.5) (t) = lim nf{[f(]”"—1}.
Since, by assumption, f(t) is infinitely divisible, [f(t)]!/" is a characteristic
function. Denote the corresponding distribution function by F,, (x); then

(55.6) mfff@P"—1} =n] (1) dF, (2)

x itx
= mF | —— aF,, (a)+n| Cox 1 =] dF, (x).
—~o 1+? + 42

If we write

(5.5.7) 0,(x) =n ls a dF’, (y)

. ae : itx \1+x?
fn (t) — anit | (-1-;**,) ps d6,, (x)

then we see from (5.5.5) and (5.5.6) that ¢,, (t) has the form (5.5.1) and that
lim $,(t) = $(2).
The three lemmas permit the derivation of the desired canonical repre-
sentations for infinitely divisible characteristic functions.
Let now f(t) be an infinitely divisible characteristic function. Since
f(t) cannot vanish, the function ¢(t) = log f(t) is defined for all values of
t. According to lemma 5.5.3 there exists a sequence of functions ¢, (t)
which has the following property: each ¢,,(t) has the form (5.5.1) and the
sequence ¢, (t) converges to ¢(t) as m tends to infinity. It follows from
lemma 5.5.2 that 4(t) also has the form (5.5.1). If we combine this with the
result of lemma 5.5.1 we obtain the following theorem.

Theorem 5.5.1 (the Lévy—Khinchine canonical representation). The function

f(t) ts an infinitely divisible characteristic function if, and only if, it can be
written in the canonical form
; ed fee tix) \ lee?
(555-1) log f(t) = itat| (« eit | “2
d6(x)

where a is real and where 6(x) is a non-decreasing and bounded function such
that 0(— 0) = 0. The integrand is defined for x = 0 by continuity to be equal
to —(t?/2). The representation (5.5.1) 1s unique.
118 CHARACTERISTIC FUNCTIONS

Remark 1. It can happen that a characteristic function f(t) admits a

representation of the form (5.5.1) with a function 6(«) which is not a
bounded non-decreasing function, but a function of bounded variation.
Such a characteristic function f(t) cannot be infinitely divisible.
Remark 2. A proof of the Lévy-Khinchine representation by means
of Choquet’s theorem was given by 5. Johansen (1966). The same paper
contains also a characterization of the logarithm of an infinitely divisible
characteristic function which is similar to Bochner’s theorem.
The canonical representation given by the last theorem can be somewhat
modified. We define two functions, M (uw) and N (u) and a constant o7 by
writing
Lx?
M (u) = \ ms d0(x) foru <0

He
(5.5.8) N(u) = -| d(x) foru > 0

o? = 0(+0)—6(—0).
The functions M (u) and N (u) are non-decreasing in the intervals (— 00, 0)
and (0, + 00) respectively and M(— oo) = N(+ «) = 0. For every finite
0 €
é > 0, the integrals | u? dM (u) and | u*dN (u) are finite. Conversely,
—8 0
any two functions M (uz) and N (u) and any constant o? satisfying these con-
ditions determine, by (5.5.8) and (5.5.1), an infinitely divisible character-
istic function. We have therefore obtained a second canonical form.

Theorem 5.5.2 (the Lévy canonical representation). The function f(t) is an

infinitely divisible characteristic function tf, and only tf, it can be written in the
form
: o “ie itu
(5.5.9) log f(t) = ata atl («
sspike: ,)
aM (u)

itu a=
+ (6 Tea)
os aN

where M(u), N(u) and o? satisfy the following conditions:

(i) M(u) and N(u) are non-decreasing in the intervals (— «, 0)
and (0, + ©) respectively.
(i) (— &) SEN Gacy =)
(iii) The integrals | u2dM (u) and | u2dN(u) are finite for every
0 €

—€é 0
eye;
(iv) The constant o is real and non-negative.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 119
The representation (5.5.9) is unique.
The canonical representations (5.5.1) and (5.5.9) are generalizations of
a representation, due to Kolmogorov, which is valid only for the character-
istic functions of infinitely divisible distributions with finite variance.

Theorem 5.5.3 (the Kolmogorov canonical representation). The function f(t)

is the characteristic function of an infinitely divisible distribution with finite
second moment tf, and only if, it can be written in the form
4 eee: Dans,
(5.5.10) log f(t) = ict+ | (e*— 1 —itx) ase

where c 1s a real constant while K (u) is a non-decreasing and bounded function

such that K(— ©) = 0. The representation is unique.
The integrand (e“”—1—itx)/x? is defined for x = 0 to be equal to
—(t?/2).
Let f(t) be an infinitely divisible characteristic function and suppose that
the second moment of its distribution function exists and is finite. Then
f(t), and therefore also 4(t) = log f(t), can be differentiated twice. We form
the second central difference quotient
Ak b(0)
(2h)?
and conclude that
ew LAL d(0)
ee
Badal ti
ee |a
1 = a ae,

We note that f(t) admits a representation (5.5.1) and use (5.5.11) to show

that the integral ib (1+ x?) d0(x) is finite. Then | x d6(x) is also finite.

We write K(x) = [ (1+?) d0(y) and c = at| y d0(y) and ob-

tain (5.5.10).
Conversely, suppose that the function 4(¢) = log f(t) admits a repre-
sentation (5.5.10). Then
Lh
O(x) = | (y)
ee Il el We

satisfies the conditions of theorem 5.5.1, so that f(¢) is an infinitely divisible

characteristic function. Moreover, it is easily seen that (5.5.10) may be
differentiated twice under the integral sign, so that the second moment of
the distribution function corresponding to f(t) exists. The uniqueness of
the representation is an immediate consequence of the uniqueness of the
representation (5.5.1).
E
120 CHARACTERISTIC FUNCTIONS

As an illustration we determine the canonical representation for a given

infinitely divisible characteristic function. The procedure repeats the steps
of the proofs of lemmas 5.5.3 and 5.5.2. We consider as an example the
Gamma distribution
A—1 e7 ¥ f
Rae wa Jot dy forx
> 0,

0 for x= 0)
The corresponding characteristic function is

f= (1-4)
where 6 and / are two positive parameters. It follows from the form of
f(t) that it is an infinitely divisible characteristic function, so that [f(z)]/”
is also a characteristic function. We denote the corresponding distribution
function by F,, (x); clearly F,,(«) is also a Gamma distribution and
g4/n)
| yan ye Horn S40
F(x), = <T(A/n) 0
0 for x < 0.
Substituting this into (5.5.7) we get
Q4/n x
| yete dy for ae
6,(x) = <T(A/n) Jolt+y
0 for x < 0
and
no?/* (oo) i| "
ae /n p—9y
ee Didi) I, ey ee
We note that

lim =" lim 4

Blo fan ae poem (A/N) [(a/n)
= lim A 7
aes T[(a/n) +1
1]
and obtain from lemma 5.5.2
x

: 6,,(x”) =
6(x) = Lim A |cays
ye tg ye tora"
0 for x < 0
while
(oe) e~
as lim, v=o | a
n—> oo 0 l+y

It is then easy to compute the other canonical representations.

Table 5 lists the canonical representations of some of the more common
infinitely divisible characteristic functions.
 0 < x 40F | 0 < * Joy 0<4
o=y | | Peis 0<¢-l=5
(4—*) yby K+) (4—*)? 4b Ghee
Kx 4 o<¢
< eas |
;
del y Et y+ Uses Pa ate nab — yermourg
0 > *I0F0) 5 ate a S* 0 ae <Ga 0 > * 1030 2 ite ss gee SATIeBON]
|
oe) ||
ORS
|
(L—* Ph Xe (T— HY, ) aX | (1-*)2 (Z/¥) 7/¥| [1 —3)x] dx uosstog
a | | ae; yeor 3
0 | Eo 0 0 ; 0 3 $n2 | oyerouaseq
a] qissod (oe | ea | , 0< 86
uolejussotdat ON 0 f) 0} Clot xuet ore iZe) 0 |a19—2 Ayonesd
| = = iz = * <
} , a. eS
QlS & TOF | | 0 < * Jox |
F
0 | &
|
| fp Ht fx 0<*‘0 <8
op me2
el | | |
|
x
n
||
|
| c+
0
| 096g c+
0 |
|
0 > * 1030 0/X) xP cs 0 jap
| fig-?
tt \y| > xs0yg 4 [e
dpe game v-(9/#—1) BUTTE)
29-9
| Poe Pane 2 We ’
| | | 0 < 29 ‘ear a
(x)>,0| W | 0 0 | a (x)> ,0 nd | [.7 .0%—131] dxo [ewI0oN
. eee =. = | a oy
(*) 9 9 o<n:
o>” |
2 | D | (*)@
uo1jeyUeserdal | ()N |) | uoneyussoider uorjouny wornqrystp
AOIOSOUWO | uoljeuesorder ag] | suTYyoUTYyy]—AAg’T o4siIojovreYy jo owen
SUOTIOUNJ OjSIIDJOVILYO OIQISIAIP A[o}TUYUT JO suOT}e}UaSeIdeI [eOTUOULD)

§9192
122 CHARACTERISTIC FUNCTIONS

We conclude this section with a few remarks concerning the factori-

zation of infinitely divisible characteristic functions.

Theorem 5.5.4. Let f(t) be an infinitely divisible characteristic function and

suppose that it can be decomposed into two infinitely divisible factors,
f() =A@Mf(f). Then f(t) and f,(t) determine f,(t) uniquely.
The theorem follows immediately from the uniqueness of the canonical
representation; it shows that the cancellation law holds if we restrict the
decompositions to infinitely divisible factors.
However, these are not the only possible decompositions; infinitely
divisible characteristic functions can have factors which are not themselves
infinitely divisible. We give the following example.
Let a and 5 be two positive real numbers and write v = a+zb. It can be
shown that the function
[1 + (t/v)][1 + (ct/d)]
65.12)
BelZ
10 = Travel
je
Goll — : ; Sa

is a characteristic function *) if
(5.5.13) b> 2av2.
Then f(—1?) is also a characteristic function, as is

s()
=FOA—1
= pon
= FOF = rae
—
1 Dos

We will show later (theorem 8.4.1) that f(t) and therefore also f(—1) are
not infinitely divisible. The product g(t) = f(t) f(—12) is the characteristic
function of the Laplace distribution which is known to be infinitely divi-
sible (see page 109). The function f(t), determined by (5.5.12) and (5.5.13),
has the following interesting property: f(¢) is a-characteristic function but
is not infinitely divisible, however |f(z)|? and therefore also |f(t)| are
infinitely divisible characteristic functions. Thus, the infinitely divisible
characteristic function |f(t) |? admits two decompositions,
IFO? = [FIFO] = FOF(-O.-
The first decomposition has two infinitely divisible factors while the
factors of the second decomposition are not infinitely divisible. This
example shows that two different characteristic functions, namely f(z) and
| f(t) |, can have the same absolute value.
The next example(f) presents an even more surprising phenomenon by
(*) To show this, one expands f(£) into partial fractions and computes
1 ioe)

= | Cmsicay(t) dt
2a
— 0
by integrating the expansion term by term. It is not difficult to show that the resulting
expression is non-negative if (5.5.13) is satisfied.
(t) Due to W. Feller.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 123
showing that two different real characteristic functions may have the same
square.
Let f(t) be a real periodic function with period 2 which is defined by
putting f(t) = 1—|z| for |t| < 1. The function f(t) satisfies the conditions
of theorem 4.3.2 and is therefore a characteristic function. We consider also
the Pélya-type characteristic function (4.3.11), that is,

zoe 1—|¢| RE feel<

0 foretell.
It follows from (4.3.9) that

(5.5.14) f(t)= — Selene

n=1

+1 1
where A, = | f(t) cos natdt = 2 | (1-1) cos natdt > 0
al! ’ 0

sodeg Alot ewer ee

n=1

Clearly f(t) is the characteristic function of a lattice distribution F(x)

which has a jump of magnitude 4 at x = 0 and jumps 4A,,/2 at the points
n=kn (k = +1, +2,...). We introduce now a second lattice distri-
bution H(«) which has saltus 0 at x = 0 and saltus A, at the points
n= kn (k = +1, +2,...). The corresponding characteristic function is

fit) > A,, cos nat.

n=1
It is easily seen that
A(t) = 2[f(@)—3]
so that

(5.5.15) g(t) = (3) 7 2[4()-3

is also a characteristic function. It follows from (5.5.15) that g(t) is periodic
with period 4 and that g(t) = 1—|¢| for |t| < 2. It is easily seen [for
instance by considering the graphs of f(¢) and g(t)] that |f(¢)| = |g(2)}-
Since f(#) and g(t) are both real-valued functions this means that
[f(ol* = [sO)*
We next give another
| aeons:
example which shows that an infinitely divisible
characteristic function may have an indecomposable factor.
Let p and g be two positive real numbers such that
p>q>O0 and ptq=1l.
The function
gi(t) = pte"
124 CHARACTERISTIC FUNCTIONS

is then—according to theorem 5.1.1—an indecomposable characteristic

function. We write
yi(t) = log
gi(t)= log
p+log [1+(q/p)e"]
(ye
= log oneP pager (2)en
It is then easily seen that

n= 3 Se 1). ay
Let

6) = 3 ha(2) tee
= 1 Shs it(2n—1)

= Swe f\. Qnit 4

¥2(t) = ‘sa
2» On (2 [e 1],

then
71 (t) = $(t)—y2 (2).
The functions f(t) = e* andg,(¢) = e” are infinitely divisible character-
istic functions; moreover,
F(t) = 81(#)82 (0).
The infinitely divisible characteristic function f(t) therefore has an in-
decomposable factor g; (¢).
We conclude this section by mentioning certain investigations con-
cerning the ‘‘Lebesgue properties” (absolute continuity, singularity, dis-
creteness) of infinitely divisible distributions. P. Hartman and A. Wintner
(1942) proved that an infinitely divisible characteristic function belongs to
a pure distribution if the function 6(«) in its Lévy—Khinchine canonical
representation is discrete. These authors also gave examples of the three
possible pure types of infinitely divisible characteristic functions. The
existence of infinitely divisible distribution functions of all these types
suggests the problem of finding conditions on the function 0(x) of the
Lévy—Khinchine canonical representation [respectively on o?, M(u), N(u)
of the Lévy canonical representation] which assure that the corresponding
distribution function belongs to a specified type.
J. R. Blum and M. Rosenblatt (1959) obtained the following result in
this direction.
Theorem 5.5.5. Let F(x) be an infinitely divisible distribution with character-
istic function f (t) and let 6(x) be the function in its Lévy—Khinchine canonical
representation. Then
(i) F(x) ts discrete if, and only if,
a
leera (*) < © and if (x) is purely discrete.
 INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 14s)
(1) F(x) ts a mixture if, and only if,
gy
i f 52 (x) < © while 0(x) is not purely discrete.

(it) F(x) ts continuous) if, and only if,

eal
(es x2 d6(x) = OO,

Theorem 5.5.5 gives a satisfactory criterion for the discreteness of an

infinitely divisible distribution but does not permit us to distinguish be-
tween purely singular and purely absolutely continuous distributions.
H. G. Tucker (1962) supplemented this result by giving a sufficient
condition which assures that an infinitely divisible distribution is ab-
solutely continuous.

Theorem 5.5.6. Let F(x) be an infinitely divisible distribution with char-

acteristic function f(t) and let 6(x) be the function in its Lévy—Khinchine
canonical representation. Then F(x) is absolutely continuous if at least one of
the following two conditions 1s satisfied :
(1) 6(x) is not continuous at x = 0, or
nea |eet oe
(ii) fr a d6,,(x) = 0.

The function 6,,(x) is the absolutely continuous component of 6(x). We

roe) —0 co
write here and in the following | = | a :
—o — © +0

A similar sufficient condition was given by M. Fisz and V. S. Varadarajan

(1963) who used the Lévy canonical representation.
In a subsequent paper H. G. Tucker (1964) gave sufficient conditions
which assure that a discrete 0(x) such that 0(+0)—6(—0) = 0 [or alter-
natively the discrete functions M(u) and N(w) defined by (5.5.8) for the
Lévy canonical representation] produce the characteristic function of a
purely singular infinitely divisible distribution function. These sufficient
conditions are not satisfied for an example given by P. Hartman—
A. Wintner (1942) of discrete functions M() and N(u) which produce
a purely singular distribution function.
A necessary and sufficient condition for the absolute continuity of an
infinitely divisible distribution was also given by H. G. Tucker (1965), and
this we now state. For the formulation of this condition it is convenient to

(*) i.e, absolutely continuous, or continuous singular, or a mixture of an absolutely

continuous and a singular component.
126 CHARACTERISTIC FUNCTIONS

write (5.5.8) in a slightly different form. We put

G(u) =
mn fi ee
d0(x) for u < 0

N(u) = aes1 ie*-db(x) for u > 0.

The Lévy canonical representation is then given by

(5.5.16) log f(t) = ita—ottt/24-f (em =aa dG (u).

We also introduce the following notation. Let G,,(u), G,(u) and G,(u) be
the absolutely continuous, the singular, and the discrete component of
G (u), respectively, and write F(x) for the infinitely divisible distribution
function which is obtained if G (wu) is replaced by G,(u) (¢ = ac, s, d) in
(5.5.16).
We can now formulate Tucker’s necessary and sufficient condition:

Theorem 5.5.7. Let F(x) be an infinitely divisible distribution function with

characteristic function given by (5.5.16). A necessary and sufficient condition
that F(x) be absolutely continuous is that at least one of the following five
conditions holds:

() |" dG.) = ©;
(ii) o? > O [2.e. O(u) not continuous at u = 0);
(iii) F4(x) is absolutely continuous;
(iv) F*(x) 2s absolutely continuous ;
(v) F%(x) is singular, F*(x) is continuous but not absolutely
continuous, while F** F* is absolutely continuous.
Remark. The theorem does not state that each of the conditions (i) to
(v) is necessary, but it states that at least one of them is necessary. Each of
these conditions is sufficient for the absolute continuity of F(x).

5.6 A limit theorem

We have shown (theorem 5.3.3) that a characteristic function which is
the limit of a sequence of infinitely divisible characteristic functions is also
infinitely divisible. In the present section we show that under certain con-
ditions the limit of a sequence of characteristic functions is infinitely
divisible, even if the elements of the sequence are not infinitely divisible
characteristic functions.
We consider in the following an infinite sequence of finite sets of
characteristic functions. Such a system {f,;(t)} ((=1,2,..., Rn;
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 127

n = 1,2,..., ad. inf.) can be arranged in a two-dimensional array:

Fu, fis), ..- > Fx, ();
JNO AOS Seem ale
-
seh nce dl Pee era ee ;
Oe Gis’ es (ee Jeuwuie jel 2) 6 o70\) .% >

Fn (t), fn (t), oi seat Sten (t);

wh 54! Fey. 6h) ei a a, coe 4) os Neti tel. &. rey 6)

We form the (finite) products

fall)
=TTfusld
kn

of the functions in each row of the scheme (5.6.1) and wish to investigate
their limits. As usual, we denote by F,,;(«) the distribution function
which corresponds to f,,;(t). The following theorem contains a very im-
portant result.

mnecovem 5.0.1.0 Ler 47.,(t), (j=l ite » kart = 1,2,..., ad. inf.) be
a system of characteristic functions and suppose that, for all t,
(5.6.2) lim [ sup |f,;(#)—1]] = 0.
n—>o 1<j<kyn

Denote by g,,(t) the characteristic function determined by

kn ce

(5.6.3) “log
g, (ft).= 3 ites | (e" — 1) dF ,,;(*+ ‘n)}
j=l —
where
srt | x dF,(x)
al<t
and where t > 0 is a constant. The necessary and sufficient condition for
the convergence of the sequence of characteristic functions
kn
(5.6.4) f,(t) = Il hat)
rol
to a characteristic function f(t) is that the sequence g,,(t) converge to a limit.
Then the limits of the sequences f,,(t) and g,,(t) coincide.
The characteristic functions g,(t) are infinitely divisible. This can be
seen by writing them in the canonical form (5.5.1) with

Ox)
@)= =3SP dayton
(0pyadooten)
Kn Hy] ye

hs oe
a=)j=1 Eel -o gts
1l+x?
003)|

or by noting that the g,(¢) are finite products of limits of Poisson-type

characteristic functions. Theorem 5.6.1 indicates that it is possible to
128 CHARACTERISTIC FUNCTIONS

replace the investigation of the limit of a system of arbitrary characteristic

functions [subject to the restriction (5.6.2)] by the investigation of the
limit of a sequence of infinitely divisible characteristic functions. This cir-
cumstance explains the great importance of theorem 5.6.1 in connection
with the study of limit distributions for sums of independent random
variables.
We do not intend to discuss limit theorems in this monograph and
will therefore not be in a position to appreciate the full significance of
this theorem. For its proof we refer the reader to B. V. Gnedenko-
A. N. Kolmogorov (1954), p. 112, where this result can be found in its
proper context.
In connection with our investigation of factorization problems we will
use a corollary to theorem 5.6.1.

Corollary to theorem 5.6.1. Let f(t) be a characteristic function and suppose

that f(t) admits a sequence of decompositions

f= ILfu (mun bgeZ see)

where the fy; (t)(j = 1,2,...,Rn3n = 1,2,...) forma system of character-

istic functions which satisfy (5.6.2). Then f(t) is infinitely divisible.
The corollary follows immediately from theorem 5.6.1 if we observe
that {,(t) =f @):

5.7. Characteristic functions of stable distributions

In this section we discuss a class of infinitely divisible distribution
functions, the so-called stable distributions. Stable distributions and their
characteristic functions are important in connection with certain limit
theorems and were originally introduced in this context. Our study of
these distributions is motivated by the fact that the class of stable character-
istic functions is of independent interest and occurs also in some problems
not related to limit theorems.
A distribution function F(x) is said to be stable if to every 5b, > 0,
b, > 0, and real c,, c, there corresponds a positive number 6 and a real

con a8) =a)

number c such that the relation

holds.
The characteristic function of a stable distribution is called a stable
characteristic function.
Equation (5.7.1) is not so much a property of an individual distribution
function F(x) but is rather a characteristic of the type to which F(x)
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 129

belongs. It would therefore be more appropriate to say that a distribution

belongs to a stable type if its type is closed with respect to convolutions.
The defining relation (5.7.1) can be expressed in terms of characteristic
functions as
(5.7.2) f(bst) f(bet) = f(bs)e"
where y = c—¢,—Cy.
Let 5, b,,..., 5) be n positive real numbers; it follows then from
(5.7.2) that
F(t) F (Bat). -F(,t) = ft e”"
where y’ is some real number, while 5’ is a positive number. If we put
b; = 1(j = 1,...,m) and write b, for the corresponding value of b’, then
we get
[F())” = fat) ev
or

#0) = {4(5-) exw [-2]}

The last formula implies the following result:

Theorem 5.7.1. A stable characteristic function is always infinitely divisible.

We see therefore that a stable characteristic function has no real zeros.
We can take logarithms in (5.7.2) and express this equation in terms of the
second characteristic ¢(t). We obtain
(5.7.3) f(b, t)+(b2t) = p(bt) +77.
Since ¢(#) is the logarithm of an infinitely divisible characteristic function,
we can write it in the canonical form (5.5.1) as
: Taal xe ttxee\ 14-0?
¢(t) = ita+| G niin ge d0(x).
It follows that
steal
p(bt) = iatb+ |" (¢ —-1 ofasa aA
115-853) B-8y2 (b-1y).

Since the function z/(1 oe is bounded, we see that the integral

aHiSnares SeeAGfe ———

rey d0(b-}
vp)
exists. We write

= ba+(1—6) |We
d0(b-1y)
and obtain, by means of an elementary computation,

6.74) $(01)
=itr
f
|3 (1, ity
;
2) \1+5b-*y?
pa a1),
130 CHARACTERISTIC FUNCTIONS

We introduce again the functions

; :
M(u) = ee TP ally) where u < 0
ce) 2
N(u) = | 1+9"46(y)
where u > 0
u
and write o? = 6(+0)—6(—0). With this notation we obtain from (5.7.3)
and (5.7.4) the relation
: bc? me Ae it
ita,——5 itil (e152 ,)aM 1y)
2 A.
P iy 1 1 _ b30
+[" (¢ 1 Tay? dN(6; 1y)+1ta,, mor 2
ayes ity
2) fp (< ity
Ps) 7
ty _ | — dM (b; 1 ity | dN (b; 1
— 0

‘i Be: ey er ity
= a a (#1 2 Wetty)
— oo

+ | (e" te
+0
2 ave Ly) + apt,
uty = ,

From the uniqueness of the canonical representation we see that

(5.7.5a) ~ o%(b?—b'—83) = 0
(5.7.5b) M(d-1y) =M(bp41y+M(6,'y) ify < 0
(5.7.5€) N (by) = N (by *y) NN (65 ty) ka 0-
We first determine the function M(u) (u < 0). Let B,, .,...,8, ben
positive real numbers; it follows from (5.7.5b) that there exists a positive
number 8 = B(;, Bo, ..., B,) such that
M(B,y)+M(Boy)+ ... +M(Bny) = M (by).
We substitute here 8; = 1.(7 =.1,.....,,%) and write B(1, 1,12), |) 4e
and see that
nM (y) = M(yA,)
or
(1/n)M(y) = M(y/A,).
Here y < 0 and A, > 0. Using this reasoning we see that to every positive
rational number 7 = m/n (m, n positive integers) there corresponds a
positive real number A = A(r) = A,,/A, such that
(5.7.6) rM(y)= M(Ay) (y <0).
The function A = A(r) is defined for all rational r > 0; we show next
that A(r) is non-increasing for rational values of the argument. Let
r, and r, be two rational numbers and suppose that 7, < r,. Since
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 134.

M(u) > 0 we see that r,M(u) <7r,M(u) or, according to (5.7.6),

M[A(r,)u] < M[A(r.)u]. Since M(u) is non- decreasing and u < 0 we
conclude that A(r,) > A(r,). By the same reasoning we can show that
A(r) is strictly decreasing, provided that M(u) 4 0. Tet us suppose from
now on that M(u) 4 0.
We now define a function for all positive real values of « by means of
Sries if x is a positive rational number
eet) ee. A(r) if x > 0 is irrational.
It follows from this definition that B(x) is non-increasing and it is easy to
show that B(x) is strictly decreasing. Let now x be an arbitrary positive
real number; then there exist two sequences {r,} and {r)} of rational
numbers such that the 7, approach x from below while the 7), tend to x
from above. Since 7, < x < r, we have B(r,) > B(x) > B(r,) and hence
yB(r,) < yB(x) < »Bi(r}) for any y < 0. Since M(u) is non-decreasing we
see that
M[yB(r,)] < M[yB(x)] < M[yB(r,)].
It follows from (5.7.7) and (5.7.6) that
r,M(y) < M[yB@)] < 1.M().
We let v tend to infinity and see that for every real positive x there exists
a B(x) > 0 such that
(5.7.8) xM(y)= M[yBix)] (y < 0).
Since the function M(u) is non-decreasing and has the property that
M(— ) = 0, we see that
B(0),= 0, Bl) =1, Bio) =0.
The strictly decreasing function z = B(«) has an inverse function x = (2).
This function is defined for z > 0 and is single-valued and non-negative.
We rewrite (5.7.8) in terms of f(z) and see that to every real z > 0 there
corresponds a f(z) > 0 such that
(5.7.9) B(2)M(y) = M(yz)
is satisfied. Let m,(y) and m,(y) be two “Evens of (5.7.9) and suppose
that m,(y) 4 0. We put
ms
(¥)
LO my (¥)
and see that
m,(2y) _ B(x)ms(y) _ ma(y) _
m,(2y) -—Bz)m(y) = ma (y) = m(y).
m(zy) =

This indicates that the quotient of two solutions of (5.7.9) is a constant.

Moreover m,(y) = |y|~* is a solution and f(z) = |2|~*. Therefore the
132 CHARACTERISTIC FUNCTIONS

general solution of (5.7.9) has the form

M(y) = Cily|™.
Since M(— 00) = 0 we must have «, > 0 and since M(y) is non-decreas-
ing we see that C, > 0. We know (theorem 5.5.2) that the integral
u? dM (u) is finite; this permits the conclusion «, < 2. We have there-
==
fore found that
(5.7.10) May = Cy lal (Ce Oe ete
The solution (5.7.10) includes the case M(y) = 0, since we admitted the
possibility that C, = 0. We substitute (5.7.10) into (5.7.5b) and see that
(5.7.11a) C, [b* —b¢ —dF] = 0.
The function N(u) can be determined from relation (5.7.5c) in the same
way in which M(x) was found. One obtains
(5.7.12) Niue Caus (Cy 2 OF 70 < on 292. a)
and notes that
(5.7.11b) C, [(b* — bf — dF] = 0.
We show next that o? 4 0 implies C,} = C, = 0 so that M(u) = 0 and
N(u) = 0. We conclude from (5.7.5a) that o? 4 0 implies b?— 57-62 = 0.
Since™ a; < 2,/\a, < 2; we infer from (5.7:11a) and (5:7.:11b) that
C, = C, = 0. If, on the other hand, M(x) [or N(u)] is not identically zero,
then C, > 0 [or C, > 0]. We put 5, = b, = 1 and conclude from (5.7.11a)
[or (5.7.11b)] that b* (or 6%) = 2. Then necessarily b? 4 2 so that it follows
from (5.7.5a) that o? = 0.
We finally show that «, = a. Suppose that C, > 0, C, > 0, and put
again b, = b,= 1; it follows from (5.7.11a) and (5.7.11b) that d% = 2 = 5%
so thato7j'="ag!
We have therefore determined the canonical representation (in Lévy’s
form) of stable distributions and summarize our result.

Theorem 5.7.2. The characteristic function of a stable distribution has the

canonical representation

(5.7.13) log f(t) = ita— 2S04 |-o (#=1- 14+?

me )am (a
S uA i)
+]. (« t ikea dN(u)

where either
o #0 and M(u)=0, N(u) =0
or
o=0, M(u)=C,lul-* (u< 0), N(u) = —C,u-* (u > 0).
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 133
The parameters are here subject to the restrictions
(Eo eesae Ge OU, Gy 2 Ur Cy OC, > Uy
Conversely, any characteristic function of the form (5.7.13) is stable.
The last statement of the theorem is easily verified by elementary com-
putations. The parameter « is called the exponent of the stable distribution.
It is possible to obtain an explicit formula for the second characteristic
of stable distributions by evaluating the integrals
: :
(5.7.14a) | (em He =) be 14+? 7 guile

and

(5.7.14b) |. (« a)
* itu ao)
zi
which occur in their canonical representation. The computations are
carried out separately for the three cases0 < « < 1ja = landl < « < 2.
We first consider the case 0 < a < 1. It is then easily seen that the
integrals
iC u du d re udu
IE a AN ee
so 1-u? |u)** o ltutyit
are finite. Therefore one can rewrite (5.7.13) in the form
e : du
(Ste) log f(t) = ita’ +aC, ie (e — 1) ae

a OCs ie(e" — hee

yerh

We suppose first that t > 0; changing the variables of integration in

(5.7.15), we get
(5.7.16) log
{(ft) =

ita’
ta’ + at a Meai"
2 (e-”—w —1) oe
dv C I (e”—1)
iw eel

Let I’ be the contour consisting of the segment [r, R] of the real axis, the

arc x = Re'* (0<< 5)of the circle with radius R around the origin,

the segment [7R, ir] of the imaginary axis, and the arc z = re? é >¢2> 0)

of the circle with radius 7 around the origin. It follows from Cauchy’s
theorem that
: dz
|e al) ite = 0
134 CHARACTERISTIC FUNCTIONS

Moreover it is easily seen that the integrals over the circular arcs tend to
zero asr 0, or as R +0. Therefore

[@-p
ah dv
= pa
™L@)
where

L, (a) =| (er) nts<0),

Similarly
- —w dv i700,
i,(e —1) ate = e771, (a).
It follows from (5.7.16) that
Ae ee TU CoC TUM
log f(¢) = ita’ +t’ aL, («)(C,+C,) COs [1a Cae, tan 5 if

Considering the Hermitian property of characteristic functions and

writing

ce = —aL,(«)(C,+
C,) cos >0

we see that for 0 < « < 1 and every ¢,

(5.7.17a) log f(t) = ita’'—c|t (1+ip a tan a

where c > 0 and |f| < 1.

We next consider the case 1 < a < 2. It is easily seen that
{° ur du _ i ue ae
oltueut J-oltu?lult
By changing the constant a, we can rewrite (5.7.13) in the form
log f(t) =
ita
Chel? +C,a |
y 1 —itu) ju
——
aaa +C,a
du i5 (ei
i — (tyes
=H?

or for t >-0;
(O:/sL0) log f(t) =

ripe dk
ita’ +t od ‘Crap2 (e —w —1+ie)
* dv i
Sit Con | iv
(e"—1—iv) dv
=}.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 135

We integrate the function (e~*—1 +7) rl along the contour [' and,

repeating the argument used above, we see that

ce
It eet?
(ers 1 +1v) eri = eee Te («)

while
5
ip(e ay | —iv) Cree
— =
dv en 1. (a)

where
3, dv
L,(«) = |.(e s1 +2) “13 > 0:

We then see from (5.7.18) that for ¢ > 0

log f(t) = zta”’ —ct* (1

+18 tg S

where c = —a(C,+C,) L.(«) cos > 0 while 6 = ae For? <0,

1 2

f(t) can be determined by means of the Hermitian property of character-

istic functions and we have for 1 < « < 2

(5.7.17b) log f(t) = ita’’—c|t |*(1+B tg a

with c > 0, |Bj < 1

We still have to discuss the case a = 1. We note that (see page 49)
° 1—cosy Tt
dy = =
i) Ve a)
and use this to compute the integrals in (5.7.13). We see that for t > 0
|
*(«7 itu )4 ifcos tu—1 {.(si poss
tu
i SIL +7 sin fu—
0 1+u?/ u? 0 dhe 1+u?

Eaie EeeP aaalim i Bo, ir

eso le of aut
pees |
e u(1+u?)
; sin v
= —St+ilim{—t | dv
2 e—0 Fe)

AA aleee 1 )ao]
id nen ce Gr ae
It is easily seen that
(as_ an) 4 i.(4 1 )ax wre
im a: ;
a3 fet ext = eeopdisie \i02y oih(Lex?)
136 CHARACTERISTIC FUNCTIONS

moreover,
EU es et
’ sin v / dv
lim | ——dv = lim | — = logt
v2 vy
so that

|bade (afe —1- dee

}
)d 4 = Wht
nitlog t+ Ait
Since the two integrals in (5.7.14) are complex conjugates, we see that
: itu 7
itu \duzune
- — itu
itu \ du
=
fo(¢ =e a) mc haps
= 5 tit log t— Ait
so that for t > 0

log f(t) = ita” —(C,+C,) 5t+(C,—C,)it log t.

It follows from the Hermitian property that for « = 1 and all real ¢

(5.7.17c) log f(t) = ita’ —clt Ht+p log ici}

[t|
Ci- Co
Here
¢ = (C1+.C,) 5 and
B=
Ci+C,.
We have therefore obtained the following result:

Theorem 5.7.3. A characteristic function f(t) 1s stable if, and only if, its
second characteristic has the form)
(5.7.19) $(t) = log f(t) = iat—<c|t cers oot | )}
where the constants c, B, « satisfy the conditionsc > 0,|B| < 1,0 < « < 2,
while ais a real number. The function (| t|,«) is given by
_ ftan (7«/2) ie 441
o(lé|,«) = tore)log |t| if« = 1.
We note that «(|z|, 2) = 0, so that one obtains the normal distribution
for¢ = 2.
We remark that P. Lévy (1937a) used the term stable distribution to
describe a somewhat narrower class. He used instead of (5.7.1) the equation

(5.7.20) F(z) *F(2) far(?)

(*) We follow in our notation B. V. Gnedenko—A. N. Kolmogorov (1954) and Loéve
(1963). This differs from the notation used by other authors who follow Lévy (1937a)
and assign the opposite sign to 8 in the canonical form (5.7.19).
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 137

as the defining relation. P. Lévy [(1937a), p. 208] called the distribu-

tions defined by (5.7.1) quasi-stable distributions. We adopt here the
terminology used by B. V. Gnedenko—A. N. Kolmogorov (1954) and we
will call the sub-class defined by (5.7.20) the stable distributions in the
restricted sense. The characteristic functions of these distributions can be
determined by an argument similar to the one which we used in deriving
the representation (5.7.19). The only essential difference between the two
classes occurs if « = 1. In this case (5.7.20) yields only the characteristic
function of the Cauchy distribution, log f(t) = —c|t |+dat, which corres-
ponds to the case « = 1, 8 = 0 in (5.7.19). Ifa # 1 then the characteristic
function of the class defined by (5.7.20) is obtained by putting a = 0 in
(5c7.19).
It is sometimes convenient to modify the representation (5.7.19) and to
write the characteristic function of stable distributions with exponent « # 1
in a different form. We show that the second characteristic of a stable
distribution with exponent « # 1 is also given by the formula)
Bus
(527.21) o(t) = iat—A\t |*exp {-iif]a

Here « is the exponent of the stable distribution, while 2 > 0 and y are the
parameters to be determined. Comparing (5.7.19) and (5.7.21), we obtain
the relations
c <a cos is
(5.7.22a)
B =amy —cot ariieet
7 tan es
7°

Formula (5.7.22a) gives the parameters c and f in terms of «, A and y. We

can also obtain expressions for A and y as functions of «, c and #. For this
we introduce a quantity A, defined by the equation

ae . cos ec5
Ago COs Tia = a
cos?
2
Then

cos = A-! cos a

(5.7.22b) Aa A(cosa =i

sen A = sgn (1—«).

(*) If « = 2, we put y = Oso that formula (5.7.21) is also valid in this case.
 138 CHARACTERISTIC FUNCTIONS

The last relation in (5.7.22b) follows from the inequalities c > 0, A > 0.
Using the relation |f| < 1 one can conclude that |y| < «if0 <a < 1,
while | y| < 2—a if 1 < a < 2. We write
K(«) = 1-—|1-a|
and see that
ly] < K(@).
We note that K(x) = aif 0 < « < 1, while K(«) < wif 1 < « < 2. If we
put y = K(a)é we obtain the representation
t aK
(5.7.23) f(t) = iat—y|t|* exp {15
teh a
whereA > 0,0 <0 5 2,0:4 1, ol <a. se constants A and ¢ are scale
factors, and by a suitable choice i the variable they can be made equal to 1.

5.8 Frequency functions of stable distributions

Let f(t) be the characteristic function of a stable distribution. It follows
from (5.7.19) that |f(é)| = exp [—c|z|*]. It is easily seen that f(z) is
absolutely integrable over (— «©, 00) and we then obtain from theorem
3.2.2 the following result:

Theorem 5.8.1. All stable distributions are absolutely continuous.

In this section we study the analytical properties of the frequency
functions of stable distributions and shall refer to these as stable frequency
functions or stable densities. We assume first that « 4 1 (« is the exponent
of the stable distribution) and defer the investigation of the case « = 1. We
denote the frequency function of the stable distribution with parameters
a, a, y, A by pPa(x; «, y, A) and write p(x; «, y, A) for po (x; a, y, A). These
functions can be determined by means of the inversion formula (theorem
3.2.2), and we obtain from (5.7.21)
Aei ger
(5.8.1a) — pa(wi a, y, 4) = On |.exp [—itx+ita—di*e~*""] dt
Lite . e
+57 lr exp [1tx —ita —At* e””/?] dt

or
(« #1)
i| (ee)

(5.8.1b) Pa(%; &, Y, A) = = Re | exp [—itx +ita—At* e~*™/?] dt.

0

The following relations follow immediately from (5.8.1a):

(5.8.2a) DalR5. CsA) = Pea; ey, A)
(5.8.26). "p(s, 7, A) — An PA en, cee)
(5.8.2c) p(x;a, y, 4) = p(—x; a, —y, A)
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 139
Equations (5.8.2a) and (5.8.2b) indicate that it is sufficient to study the
frequency function only for the values a = 0, 4 = 1 of the parameters.
For the sake of brevity we shall write p,,,(«) for p(x; «, y, 1). We say that
Pu, (x) is the standardized density and refer to y as its second parameter.
Explicit expressions for stable frequency functions in terms of element-
ary functions are known only ina few isolated cases. We obtain from (5.7.21)
for « = 2, y = 0 the characteristic function of the normal distribution,
while (5.7.19) yields for « = 1, 8 = 0 the characteristic function of the
Cauchy distribution. In addition to these distributions, only the stable
frequency corresponding to « = 4, y = }is known to admit representation
by a simple formula involving elementary functions (see p. 143). In view of
this situation it is of interest to obtain series expansions for stable densities.
We first consider the case where 0 < « < 1 and assume also that x > 0
(in view of (5.8.2c) the condition x > 0 is not a serious restriction). We see
from (5.8.1b) that

(5:33) p(x)" = aie | exp [—itx—t*e~*"””] dt.

7 0
Let
(5.8.4) g(z) = exp [—ixe—2* exp (—zy/2)],
where z is a complex variable. Denote the arc of the circle with centre at
= = 0 and radius p which is located in the fourth quadrant by

C2 {pei 3 ate of
and consider a closed contour I consisting of the segment [—7zr, —7R] of
the imaginary axis (r < R), the arc Cp, the segment [R, 7] of the real
axis, and the arc C,. According to Cauchy’s theorem we have

(5.8.5a) | p(2z)dz = 0.
rm
We next consider the integral along the arc C,,
e
4id]dd
[(C,) = | g(2)dz = ir | exp [—ixr e# — 1% elle—i/2.
.

Cr —2/2

Then
2 : muy
|Z(C,)7 |<7 |, exp E sin ¢—7* cos (+)2 dd.

It is then easily seen that

(5.8.5b) lim 1(C,) = 0.
r—>0
We next show that also
(5.8.5c) lim (Cz) = 0.
R->o
140 CHARACTERISTIC FUNCTIONS
Since |y| < « < 1 it is always possible to find an e > 0 and a ¢, such
that
1
Fe os > jar
Z |
holds for 0 < ¢ < ¢. We have

[L(Cz)| < R Ihexp bexR sin

¢ —R* cos (40+
sy) dd

ar a le exp E xR sin ¢ — R* cos (40+”)]dd,

and it is easily seen that each term in this inequality tends to zero as R
goes to infinity, so that (5.8.5c) holds.
We see therefore from (5.8.5a), (5.8.5b) and (5.8.5c) that

| exp [—ixt—t*e~™/?] dt --i| exp {-s-yexp [Sv +ay|bay

0 0

It follows from (5.8.3) that

1 ne :
Pox (*) = “Re{ ~i | exp {-9-9" exp |-So+9] ba}
A 0 y
1 a Pore
+ Re{ - | e— exp |— Serer ai}.
TUX 0 x
We expand exp[—7*x-*e~*"t/2] into a series and note that it is pos-
sible) to pic. the integration and the summation and get
ee T(ak +1) sin [ka(y + «)/2],
Pay (%)=
=z
70d k=1
provided that x > 0 and 0 < « < 1.
Using formula (5.8.2c) we obtain an expansion for x < 0, and it is then
possible to obtain a formula which is valid for x > 0 and x < 0, namely
(5.8.6)
a8) =2 3IEEE|sinCy 0.2 arg x)|eI
een = NF, < 0 and arg x = 0 forx > 0.
We consider next the case where 1 < « < 2 and assume again that
x > 0. We choose the following contour for integrating the function g(z)
defined by equation (5.8.4). In the case when y < 0 the contour consists of

the straight-line segment Eexp (lel) R exp (-27N)), the cir-

o
(*) See E. W. Hobson (1927) vol. 2, p. 306.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 141

cular arc zg = Re with oe < ¢ < 0, the segment [R, r] of the real

axis, and the circular arc x = re“ withO > ¢ > — aly|. If »y> 0 we use
255
the contour which consists of the line aaa [r, R] of the real axis, the

circular arc z = Re* withO <¢ < a the line segment

a

[Rew (52) re (Ze)

and the circular arc = re‘* with Se > ¢ > 0. It is easily seen that in both
4

cases the integrals taken over the circular arcs tend to zero as r +0, or as
R + o. It follows from Cauchy’s theorem that
(5.8.7)
(es) _ ty iy po tad
| exp |-ist—1e | at =e | exp |~ im e% | du
0 0

or
oo) mer ity po 1 iny 1-1
| exp |-iat— te | a Son: rs [ exp |- inset Jee ds
0 40

(the last expression is obtained from (5.8.7) by introducing a new variable

iat
s = u* in the integral on the right-hand side). We expand exp |— ia |

into a series and see, as before, that the order of integration and sum-
mation may be exchanged. In this way we obtain

| exp [—ixt—t*e
iny

*] dt =
Ge)ube pate+ oe Ane
0 20
We see then from (5.8.3) that for x > 0
k
a peas Ne) afpt) op
TUX p=1

Using relation (5.8.2c) one obtains a similar formula for x < 0. We

summarize these results in the following statement:
Theorem 5.8.2. The stable frequencies admit the following. representation by
convergent series.
142 CHARACTERISTIC FUNCTIONS
TO Sse,
(5.8.8a)
1 2 (-1)
1T@k+1) . pal _ 2a )] =
Pexy (%) are 6 Sap sin 5 y+o z arew, WiC ene ys

while for 1 < « < 2,

(5.8.8b)

1 2 1 l(t!) k 2a
..o)=— Ss (] Diriget ae EC arg x)
|Teale
i oY k! 2a 1
holds.
The expansions of stable frequency functions into convergent series were
obtained independently by H. Bergstrém (1952) and W. Feller (1952).
Several interesting properties of stable frequency functions follow from
theorem 5.8.2.
We assume that « # 1 and select |y| = « in the representation (5.7.21)
[this corresponds to the choice of |f| = 1 in formula (5.7.20)]. It follows
from (5.8.8a) that
p,,(%)="0 if BES Ol 0 — oe
and also
Pay (*) = 0) 1f x <0, y = «4,0 <@ <1.
To formulate this result we introduce the following terminology which
will also be useful later. We say that a distribution function F(x) is
bounded to the left and that a is its left extremity; in symbols, a = lext [F'],
if for any e > 0 we have F(a—e) = 0 while F(a+e) > 0. Similarly we
say that F(x) is bounded to the right, and that b is its right extremity; in
symbols, 6 = rext[F'], if F(b—e) < 1 for any positive ¢ while F(b) = 1.
Distributions which are bounded either to the right or to the left are called
one-sided distributions, distributions which are bounded both to the right
and to the left are called finite distributions. Our preceding result can now
be formulated in the following manner.

Theorem 5.8.3. The stable distribution functions with exponent0 < « < 1
and parameter |y| = « are one-sided distributions. They are bounded to the
right (with rext [F]= 0)ify = — «and bounded to the left (with lext [F] = 0)
ify = +4.
Remark. It is not possible to apply a similar reasoning to formula
(5.8.8b) since we have always |y| < 2—a < ainthe case when 1 < « < 2.
V. M. Zolotarev (1954) as well as P. Medgyessy (1956) obtained dif-
ferential equations for stable frequency functions with rational exponents.
V. M. Zolotarev (1956) also derived a number of relations between stable
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 143
distribution functions (density functions). A simple relation of this type
is equation (5.8.2c).
Theorem 5.8.2 can also be used to express a stable density with exponent
a greater than 1 in terms of a density with exponent 1 /«.
Let « be the exponent of a standardized stable density with second
parameter y and suppose that 2 > « > 1. Using formulae (5.8.8a) and
(5.8.8b), we derive easily the following result:

Theorem 5.8.4. Let «* = 1/a« and y* = eed 1:

then peed Cp eee ies Tipe(er

%)
weras>
O and 1)<a< 2.
It is easily seen that |y*| < «* so that p,.,.(x) is indeed a stable fre-
quency function.
Theorem 5.8.4 is due to V. M. Zolotarev (1954) who gave a different
proof which did not use the series expansions of theorem 5.8.2. He also
obtained a similar relation for stable distribution functions [V. M. Zolo-
tarev (1956)].
A particular case is of some interest. Leta = 2,y = 0;thena* = y* = }
and the corresponding density is, according to theorem 5.8.3, bounded to

the left. Since p2)(x) = e~"/4 we obtain from theorem 5.8.4 the
2Vx
stable density with parameters « = $, y = 3, namely
0 if x < 0

(5.8.9) ps, 3 (*) = 1 73/2 p14) if » >

2Vx
The frequency function (5.8.9) can also be obtained directly from the
series expansion (5.8.8a); it was derived by P. Lévy (1939) by a different
method.(t)
Apart from the normal distribution, the Cauchy distribution and the
distribution given by (5.8.9), no stable distributions are known whose
frequency functions are elementary functions. However, V. M. Zolotarev
(1954) expressed the standardized(§) frequency function of stable laws
for certain combinations of the parameters « and # in terms of higher
transcendental functions. These combinations of the parameters are
(@=%, B=1), @=% B=1), «@=% 6=0), «@=%, @=1),
(x = 4, f arbitrary).

(t+) B. V. Gnedenko-A. N. Kolmogorov (1954) mention that this frequency function

was also found by N. V. Smirnov.
(§) i.e. those obtained by putting a = 0, c = 1 in (5.7.14).
144 CHARACTERISTIC FUNCTIONS

We study next the analytical properties of stable densities and see from
(5.8.8a) and (5.8.8b) that they have the form

-®,(x=) “forte 0
(5.8.10) p,,(x) = . (0 <a <1)
—®,(|x|~*) forx <0
TUX

and

1
oy Et O*) for x > 0
(5.311) - p,(a)je= 1 (1 <¢e <2)
net 2All) for x <0;

where ®,(z) = imast, ¥,(2) = 5 bes, (= 1,2) with

k=1

a Tepe
gaTak +1) sin | ly+(- 1)/- ta}

be Be vy
one +1]sin {i ly+(- 1)/- ta}.

Since J! = (—1)*b)?, we see that V’,(|x|) = Y,(x) and can rewrite

(5.8.11) in the form

(5.8.1la) p,(#) = =H) (Le< ome 72),

where V(x) = x-1 P(x).

Using Stirling’s formula, one sees easily that
lim sup |a |}/* = lim sup |b® |}/* = 0
ko k>o

so that the functions ®,(z), ®,(z) and ‘(z) are entire functions. We can
also determine 6s order and type of these functions.
Let 0(2)= DSc;,* be an entire function. It is then known (see Appendix
D) that the iden p and type zt of 6(z) can be expressed in terms of the
coefficients c, of 6(z) and are given by
: klogk
5.8.12a = lim‘sup/=———=—
( ad Paes P fog fe, |=
and

(GAAlb). pm iene
ep ka
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 145

respectively. We substitute in these formulae for the c, the expressions for

the aj) and see after a simple computation (again involving Stirling’s for-
mula) that ®, (x) and ®,(z) are entire functions of finite order p = 1/(1—«)
and type t = (1—«)*/"~®, Similarly we see that Y’(z) is an entire function
of order p = «/(x—1) and type («—1)a~%/@-»,
The determinations of the order and type of the entire functions ®,(z)
and ‘Y’(z) were carried out for the functions associated with the densities
Pay (x), that is for the case A = 1. Similar results can also be obtained if
A # 1. Expansions analogous to (5.8.8a) and (5.8.8b) can be derived easily;
let aj’ (A) and by (A) be the coefficients in these series. It follows from
(5.8.2b) that af) (2) = a? A’, while bY (4) = bY 2-*/*, We conclude from
(5.8.12a) and (5.8.12b) that the order of the entire functions associated
with p(x; «,y,/4) is given by p=(l1—«)-1 if 0O<a<1 but by
p = ao(a—1)-1if 1 <a < 2. The type t = 4° (1—a)“"—® for0 <a < 1,
but t = 4-*/*(@—1)a-*/@-) in the case where 1 < « < 2. It follows
easily that the function ‘Y’(z) is also an entire function of order p= a(a—1)-4
and type t = A~*/*(a—1)q—~V/@-2,
We still have to consider the case « = 1. Since the representation
(5.7.21) is not valid for « = 1 we must use the canonical form (5.7.19) as
our starting-point. If 6 = 0 we have f(t) = exp(—c|t|); this is the
characteristic function of the Cauchy distribution, and the corresponding
frequency function) is

eed Aa n(x? +c)”

This is a rational function with poles at the points x = +7c and is therefore
regular for all real x. The radius of convergence of the Taylor series of
p(x; 1, 0, c) around the point x = 0 is equal to c.
We study next the case 8 # 0; in view of the fact that a relation similar
to (5.8.2c) is also valid for « = 1, it is no restriction to assume that B > 0.
Using the reasoning which yielded the expression (5.8.1b), we see that
ri Z
(5.8.13) p(x; 1, B, c) = GeRe | exp {-ite—ci[1 fe 1 log i)}ae
I 0
We write
; pas
(3) = —1xz—CZ— zo log 2,

where z is a complex variable, and consider again the closed contour I’ used
in deriving (5.8.8a). As in the earlier discussion we show that

(5.8.14a) lim |__ exp [e(@)] dz = 0.

(*) The use of the notation p(x; 1, 0, c) cannot create any confusion since the symbol
p(x; «, y, A), introduced on page 138 and based on the representation (5.7.2b), is not
defined for « = 1.
146 CHARACTERISTIC FUNCTIONS
Subdividing the range 0 < ¢ < x of integration along C, at a sufficiently
small ¢9, one can also show that

(5.8.14b) lim | exp [e(2)] dz = 0.

R>o / Cp

The assumption that 8 > 0 is needed in deriving (5.8.14b). One con-

cludes finally from Cauchy’s theorem, (5.8.14a), (5.8.14b) and (5.8.13)
that

(5.8.15) 0 Plvsels ese )e= °lie[sin (1+ £)ct] exp ‘os 2p ct log t >dt.
MA

We wish to study the analytic character of p(x; 1, 6, c) in the case where

B # 0. Without loss of generality we can put c = 1; for the sake of brevity
we write p,(x) instead of p(x; 1, B, 1). We expand e~” in (5.8.15) and ex-
change the order of summation and integration. In this way we obtain

p(x) = i el" ‘‘[sin (1 + £)t] exp (- a t log ) dt

4

or
1 fo)

(5.8.16) p(x) = - ae xt
where
aie wk. o 2
(5.8.16a) a, = ( = |. t*[sin (1+ )¢] exp (-2 t log :dt.

Let 7, = (2/26)n, where 7 > 1 may be chosen arbitrarily large, and put
t, = exp (7). We write the integral in (5.8.16a) as the sum of three inte-
grals J,, J, and Jz, taken over the intervals (0, 1), (1, ¢,) and (¢,, 00) res-
pectively, and estimate J,, J, and J3. Since max [—t log t] = e—! we see
that

(817) gc, [C=en ea)

Whi< k+ me
and
Prag
(5.8176) Wifal = (k + 1)
(the estimate for J, follows from the fact that t log ¢ > 0 for ¢ > 1).
We have

Js = eet* exp (-* t log t)sin(1+ 8)t] dt

th
so that
. 4
eg i wexp(—7 rtoge) at
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 147
In view of our choice of ¢,, we get

Jal < |
ive) foe)

Pert dic | tk e—™ dt:

hy 0

therefore
(5.8.17c) ate ne eR!
It follows from (4.8.17a), (5.8.17b), (5.8.17c) and (5.8.16a) that

Jal cot
Cyr el aie |

Puree.
or
(5.8.18) |a,| < n-*¥-1[1+0(1)] (ask > o).
We see therefore that
|a, |" < n-4[1+0(1)].
Since 7 can be arbitrarily large, we conclude that
lim sup |a, |!/* = 0
ko

so that p,(x) is an entire function.

We rewrite (5.8.18) in the form
(5.8.19) |a,| = 6.n-*-1[1+0(1)]
where 6, is a real number such that 0 < 6, < 1. Since p,(x) cannot be a
polynomial, there exists necessarily a subsequence 0,, of the 0, such that
6,, > 0. In order to simplify the notation we write in the following 6;
instead of 6,,. Using (5.8.19), we see that
j log j j log j .
log |a;/-2 (J+1)log
4—log 6, + (1) = O(logj)
hence

ko log |ay, ia »

Therefore p,(«) is [see (5.8.12a)] an entire function of infinite order.

We summarize these results in the following statements.

Theorem 5.8.5. The frequency function of a stable distribution with character-

istic exponent « < 1 has the form

1
a P, (%-*). efor x = 0
P(X% A= 4 4
a, 2 (lel) fon can),
148 CHARACTERISTIC FUNCTIONS

where ,(z) and ®,(z) are entire functions of order p = (1—«)~1 and type
pone (1 ws Caphone

The frequency function of a stable distribution with exponent « > 1 is an

entire function of order p = «(a—1)-1 and type t = A~*/* (a—1)a~#/%—®,
Theorem 5.8.6. Stable densities with exponent « = 1 are entire functions of
infinite order if B # 0 but are rational functions if 8 = 0. In this case they
have poles at the points ic and —1c.

5.9 Asymptotic expansions and integral representations of

stable densities
It is sometimes convenient to have asymptotic expansions of stable
density functions. In Section 5.10 we will also use a representation of the
derivative of a stable density by an integral. In the present section we
derive some of these formulae.
As a first example we derive an asymptotic formula for stable densities
Puy (x), with exponent 1 < « < 2, which is valid for large positive values of
x. We see from (5.8.1b) and (5.8.7) that

= Reem (FE) |, ee| ten

Poy (*) = ssRe {expae |.exp |—ixuexp\

We introduce a new variable by putting

(32) x] ay
> ]—¥ du?.

“= tx—1 e~ t/ (2a)

and get
(5.931) Poy (x) =

= Re {exp a ) ieexp E t exp Ce

3e= 2 =P a “er |ar}

According to Taylor’s formula we have

: at ; {HMA yan 1)
exp |—t arent] — 2 zl =) hg eve re 2 pe 4.0
=0 : (n+1)! °
where |6| < 1. For the sake of brevity we write dy) = = («+y—1) and

l= | i exp (—t e*) dt.

0

We then obtain from (5.9.1) the expression

622) Pay (*) = ne, in(y—1)] =2 a1 (—1)'x

1 Re {exp [Ae] oes
alkg—ink/2 J

xn 1)
9 ——_— ;
(n+1)! zl
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 149
In order to determine the integral J, we change the path of integration
from the positive real axis to the line t = ue~**, where 0 < u < o. To
show that this is permissible we consider the function
g(z) = 2 exp [—2 e*]
and the circular arc. = {z: 2 = re}, —d, < d < 0, and conclude that
lim | &(z) dz = 0 and also lim | &(3)dz = 0. In this way we see that
reo JT roo JT
I, = exp [—if)ak—igy] T(ak+1).
We substitute this into (5.9.2) and obtain the asymptotic formula
329.3)
1S (=I TRH)
Pap(®) = a 0 k!
|sin (a9)|aot + Ofereitny
asx > ooand! < « < 2.
We compare formula (5.9.3) with (5.8.6) and see that the series in
(5.8.6) is convergent for 0 < « < 1 but is still useful as an asymptotic
seriesit Tt. <2:
It is sometimes of interest to have asymptotic expansions as x tends to
zero. We treat as an example the case where 0 < a < 1 while x > 0. In
Section 5.8 we had

(5.8.1b) == Re ike~ # exp'(—i*e-™/2) dt.

,,(x%) =

We again use Taylor’s formula and write

ey ite n (—ix)F gntlynrl
(5.9.4) e __
= h+O~
Ie
Gain (|6| << 11)5
We write
Ja | t® exp (—t%e-*””) dt
0

and obtain from (5.9.4) and (5.8.1b) the equation

Pope) =2Re{
& SG j40 Jun}
We compute J, by changing the path of integration and justify this
change by applying Cauchy’s theorem. We choose the line
= uexp(iy/2a), 0< u< o,
as the new path and see easily that

i peat p(s) exp [imp(k+ 1)/(2a)],

150 CHARACTERISTIC FUNCTIONS

so that
at k+1
(5.9.5). SPay(X)
= ay
in “e+ +a|
AB fr ak!

fce) |
ia 40
iis (n+1)! :
This is an asymptotic formula (for small x) if 0 < « < 1, and it can be
shown that the series (5.9.5) is convergent if 1 < « < 2. Formula (5.9.5)
is due to H. Bergstrém (1952). The asymptotic behaviour of stable density
functions was also studied by Yu. V. Linnik (1954) and by A. V. Skorohod
(1954). A. V. Skorohod (1954) and I. A. Ibragimov-Yu. V. Linnik
(1965) also gave comprehensive surveys of these formulae.
In the same way one can derive formulae for the derivatives of stable
frequencies. As an example) we mention
1 (- par)

(5.9.6) Pin) = = Ds + w cos [Fe 1)(241) +22

2a
B("ie ) |
a
gn erempmee
We note that (5.9.6) can be obtained from (5.9.5) by formal differentiation.
In Section 5.10 we shall need also the representation of the derivative
Pan (x) of a stable density p,,,(«) by an integral taken over a finite interval,
and we now derive the following result.

Theorem 5.9.1. Let 0 < « < 1; then for x > 0,

(5.9.7) Pia(s)=1=9”,2/0- |"5(4) exp [—x*/*- alg] dp

where a(p) = @ iesn L- and
sin sin a

OTA Grg eter

while p..,(x) = 0 for x < 0.
The last statement follows from theorem 5.8.3, so that we have only to
prove (5.9.7).

(*) We write here and in the following px, (x) for the derivative of pyy(x) with respect
to the variable x.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS Sl
We differentiate (5.8.1a) to get an expression for p/,,(«) which is similar
to (5.8.1b) and see a

O98)" p,,(x)==*Re ie(—it) exp [—itx—t*e~*”/?] dt,

provided that « #4 = We introduce the new variable t = vx/“-) and

obtain
CORO
=x2/@-)) Re \-(—iv) exp {—x¥@—) [iv + v%e-*/?]\ du

or

(5.9.8b) p(x) = eae Re | (tv) exp {x%/@—) [in—y% e¥/2]}

do,

According to the soa ot»of theorem 5.9.1 we have 0 < « < 1; in this
case |y| < «. We substitute y = « in (5.9.8b) and obtain

Daalk)= =a oY Re | (iv) exp {x*/@—) [iv

— v*e/2]} do,
0

Let
o(z) = iz—2%e"™/2 (z complex)
and put
h(z) = iLel (a— Vig exp [x a/e—1) of 2)]

so that

Pia(8) = Re | h(e) dz
0

where the integral is to be taken along the positive real axis. Our next aim is
the computation of the expression Re | h(z)dz. This will be greatly
0
facilitated by showing that the path of integration can be replaced by a
curve along which the function g(z) is real. It is easy to determine such a
curve. Let z = pe’*; then
Im g(z) = p sin (9+3)— p* sin C 2h

Clearly Im g(z) = 0 if z = pe’, where

sin (443) gore
(5.9.9) pp = p($) = mee)
152 CHARACTERISTIC FUNCTIONS

We note that (+3) = © while lim p(¢) = «/"-® and denote the
o>— 1/2

path z = p(d)e*, —n/2 < 6 < m/2, by I’. The curve I has the point
3, = —ta/"-% of the imaginary axis as its initial point and intersects the
real axis in the point 2) = [sin az/2]”"-™ < 1. Let 2, be the point of
intersection of I’ with the circle of radius 7 and centre at the origin. We
denote the arc of the circle z = ne located in the first quadrant and having
the points z = n and z, as endpoints by QO, and we write I’, for the part of
I‘ located between the points 2, and z,. Let
Coe {ete =vet Usa 7/2}
be the arc of the circle with centre at zg = 0 and radius 7 which is located
in the fourth quadrant. We consider the contour K which consists of the
arc C,, the segment A, = [—77, —ia!/"-®] of the imaginary axis, the arcs
I, and O,, and the segment [n, 7] of the real axis. It follows from Cauchy’s
theorem that

| We)de = 0.
K

so that

(5.9.10) | Ma)de+| We)ds+| ne)de+| h(e)de = |"We)dz

Cr Ar In Qn T

It is easily seen that

(5.9.11a) lim | A(z)dz

=0.
—0 Cr

We show next that

(5.9.11b). <lim- ts h(2) dz =. 0),

n>o J Qn

Jue
Since
m/2
<n{ h(ne'*) |db
0

we see that

ola
4
h(z) d <
n/2
~ x2/@-D 2 | exp {-=e” ey
E sin ¢ +n* cos
I 0 o(¢rr5)
|
We select a ¢y such that
am m(1— 2)
0 < dy < min(
2 2
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 153

and conclude that

1 = 7%
\yh(z)dz| < % yO) 42d, exp {-e» n* cos o(
bo sy

1 an | cos o(4ae1 |
oat).n
Te ase
. exp | —nx 1) sin f| 1+ Ie a S72) n dp.

This means that

|| h(s) dz =o(1) asn > 0,

Q
so that (5.9.11b) is proved. We finally note that g(—zy) is real for real y;
therefore
(5.9.11¢) Re| h(2)dz = 0.
A,

It follows from (5.9.10), (5.9.11a), (5.9.11b) and (5.9.11c) that

6.9.12) _p/,(x) = Re | ae) dz = Re | h(x) de.

0 r,,
In view of the definition of the contour [’ we know that
gle” p($)] = Re gle p(4)],
so that

(5.9.13) gle] = o(4) cos (443) —[e(d)F cos o(+3).

We substitute into (5.9.13) for p(f) the expression given in (5.9.9) and see,
after an elementary computation, that

sin o($43) ar i -a($+7) |

(5.9.14) gle o(4)] = —
med)
7
ede)
= -4+5)
where the function a(¢) is defined in the statement of theorem 5.9.1. Since
z = p(¢)e* on the contour [' we see easily that
Re (izdz) = —[pp’ sin 26+ p? cos 24] db
and obtain from (5.9.9) the expression

Re (izds) = [o(H)}°B(4+5) ds,

where
2« cos ad sin 6—2 cos ¢ sin ad
Bid) = eres cos $+ cos 2¢.
154 CHARACTERISTIC FUNCTIONS

It follows from (5.9.12) and (5.9.14) that

(5.9.15) Pia(x)= : :
¥e—0 "] (42) |"BG) exp (20 a9),
A simple computation yields the expression
2x cos ¢ sin(1—a)b_
(5.9116) 0 Bey = (1—«) sin a :
We put
(5.9.17)
W(4) = E(s7 *)|* B(d) - E =f2/12) (Dey Toa :3 ap i}

and obtain the statement of theorem 5.9.1. The following alternative

expression for B(¢) is easily obtained from (5.9.16):

(5.9.16a) BU) = 3 a aedsinC7

ad
i}
and this will also be used.

Corollary 1 to theorem 5.9.1. The function a(¢), defined in theorem 5.9.1,

is strictly increasing in the interval [0, zt].

Let, for ¢ fixed, y(«) = « cot af —cot ¢. Clearly y(1) = 0 while es< 0,
a

so that
a cot ap > cotd
for 0 < a < 1. Moreover, it is easily seen that
d
as 1
i log a(¢) eee
(aa {a?(ae cot ad uh—cot 6+(1—«)?
=e? cot (1—«)d}
=

> (1—«) cot (1—a)d—cotd > 0,

which proves the corollary.

Corollary 2 to theorem 5.9.1. The function b(¢), in the statement of theorem

5.9.1, has exactly one change of sign in the interval [0, x].
In view of (5.9.17) and (5.9.16a) it is sufficient to show that

le
a sin (2—a
sin ad
1

has exactly one change of sign in [0, z]. We note that u(0) = 1—« while
u(m) = —(1+«)/(1—«) < 0, so that at least one change of sign occurs in
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 155

the interval. An elementary cae cashows that

uQ) =grag)
where
v(¢) = (1—«) sin 26—sin 2(1 —«)d
so that
vu(¢) = 2(1—«)[cos 24—cos 2(1—«)d].
It is then easily seen that there exists a unique value 4, such that
R80, <9
while v'(¢o) = 0, and we get $y = 2/(2—«). It follows that v(¢) has
exactly one minimum inside [0, ], so that v(¢) and therefore also u(¢) has
at most one change of sign in this interval. The statement of the corollary
follows immediately from (5.9.17).
It is also necessary to derive a result similar to theorem 5.9.1 for the case
where the exponent « > 1.

Theorem 5.9.2. Let 1 < « < 2; then for x > 0,

r9N3) ep, 4(a)a taiaeeY \.b (0) exp {x*/= "a, (0)} 40,

while for x < 0,

(5.9.18a) pa a(s) = +] )/-” |""45(0) exp (|x! a4(0)}4,

where
—sin 0]/°—» sin («—1)6° i 6 ie sin («—1)0
ZN Ne E a0 | aah, Owe Wes sin «0 sin a6
and
—sin 6 nie 2a cos
6 sin aa
gO Fed a («—1)
sin «
fee
sin 0 ie - cos 6 sin («—1)6_i}
Hs sin «6 («—1) sin «6 ;
In order to simplify the notation we write p’(«) instead of p,, (x) in the
proof of theorem 5.9.2. We consider first the case x > 0 and substitute
y = «—2 into (5.9. oeand see that
p (x) = x2/(—-1) Re if(iz) exp Me D 9(z)} dz

where Bees izp27 ee".

156 CHARACTERISTIC FUNCTIONS

We determine first, as in the proof of theorem 5.9.1, a curve along which

2(2) is real. It is easily seen that the curve I',, given by
’ f — cos Wa) ap ot I
z= pe* with p = p($)= = (2-7 <4<3)
p+
sin o( 3)

satisfies this requirement. Moreover, (3) = 0 while (2) = 00, Let

%, = ne'*” be the point of intersection of I’, and the circle of radius m with
centre at the origin and write C,, for the arc of this circle which is located
in the first quadrant between the real axis and the point z,. It is not
difficult to show that

|), 2)exp (227 e(2)} ds|= o(1)

as n —> oo, and we conclude from Cauchy’s theorem that

[-G2) exp oP g(a)} dz

n/2

3 |n/a—n/2 (ipp' — p®) e*™* exp [x*/“-) o(pe*)] dp = 0.

We use the fact that g[p(¢) e*] is real and introduce the new variable
0= $45 TU
and obtain after a somewhat tedious but quite elementary
computation formula (5.9.18).
We consider next the case x < 0 and see from (5.8.2c) that
Po) = DP (|x |)(x =a” |x | < 0).
We put y = «—2 and write again p’(x) for p,,,.(«) and see from (5.9.8a)
that

(5.9.19) p(s) = =| |" Re |*(is) exp {—|x[/-Yg(s)}dz

where
2) = 22",
Again it is easily seen that Im g(p e*) = 0 if s = pe* is on the curve I,
defined by
s, TU 1/(a@—1)
sin(¢ot 54
p(y) = | (-Z<9<% 3)
sin o(
4 + 5) Z ome
1/(a—1)
so that g(z) is real along [’,. We have of-*) oS (*) ; (2-*) yn 5y
ed
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 157

so that [, goes through the point —i«~/“-» of the imaginary axis and

approaches asymptotically the line z = r exp |i(2—)], (Q<r<o.

a

We also note that g(—zy) = y—y* for real and positive y, so that the inte-
gral in (5.9.19) is purely imaginary along the negative imaginary axis. Let
%, = ne'* be the point of intersection of I’, and the circle of radius n with
centre at the origin, and let C,, be the arc of this circle located in the first
quadrant between the real axis and the point z,. Then

[_ Ge)exp {-[xf/P¢(e)}de = o0(1) (n > o).

We consider the contour which consists of the segment [0, —ia~“°~] of

the imaginary axis, the arc of I’, between the origin and z,, the arc C,, and
the segment [n, 0] of the real axis. We apply Cauchy’s theorem and let n
tend to infinity and conclude that

Re {[esp l-le Petal] det FG) exp [-[x =" g(a)] de}
= 0.
It follows that

p(x) = |x? | (is) exp [—[x MP g(a)] ds,

Ts
Using the argument which we employed before, we obtain (5.9.18a).

Corollary 1 to theorem 5.9.2. Let 1 < «<2; then p,,-2(*) <0 for
x > 0.
To prove the lemma we need only show that b, (6) < 0 form/a <0 <2.
Since sin «f < 0 in this interval it is sufficient to show that
y(0) = («—1) sin «6 —2« cos 6 sin («—1)0 > 0
for z/a < 9 < a. A simple computation shows that
¥(0) = asin (2—a)6—sin af > 0.

Corollary 2 to theorem 5.9.2. The function a,(0) which occurs in (5.9.18a) is

strictly increasing in the interval [0, x/«].

Corollary 3 to theorem 5.9.2. The function b,(6) which occurs in (5.9.18a)

has exactly one change of sign in the interval [0, x/«].
The proof of the last two corollaries is analogous to the proof of corol-
laries 1 and 2 to theorem 5.9.1.
Theorems 5.9.1 and 5.9.2 expressed the derivatives of stable frequency
functions in terms of certain integrals. V. M. Zolotarev (1964) derived
somewhat similar representations for stable distribution functions. In the
158 CHARACTERISTIC FUNCTIONS

next theorem we present his result; we use here the notation of formula
(5.7.23) for stable characteristic functions. The characteristic function of a
stable distribution with exponent « 4 1 and parameters a = 0,4 = 1,
y = K(a)6 is then given by

log f(t; a, 6) = —|t|* exp {~i=

We write F(x; «, 6) for the corresponding distribution function.

Theorem 5.9.3. Ifa #1 and x > 0 then

J-a)+= |" exp (-Vees d)}ds, fa <t |

1 a/2 }

' ?
EF (x;.a, 0) =
i ; ,
1-- p if a >
(a)/20 wea {
TU J —ndK a (%, $)}

where
nu
ML a/(1—a)
(@)|
—1)¢+ 5 aKee
cos | e
s (xp+5 0K Si e =
a
=1) | ae Neera
=X ace(Oe
seam Z
VAS (Zs, f)

If « £ 1 but x = 0 one has

F (034, 6) = $[1—8K (w/a.
In the case where x < 0 one obtains the corresponding representations
from the above formulae and the relation
l—F(—x; «, 6) =o (x; a, — 6).
For the proof we refer the reader to Zolotarev’s paper, quoted above. This
paper also contains a similar formula for « = 1.

5.10 Unimodality of stable distributions

In this section we prove the following theorem which is due to I. A.
Ibragimov-K. E. Czernin.

Theorem 5.10.1. All stable distributions are unimodal.

The proof is carried out in several steps. We first study symmetric stable
distributions (i.e. y = 0); then asymmetric stable distributions with ex-
treme values of the parameter y [i.e., |y| = K(«) or equivalently |6| = 1,
respectively |8 | = 1]; and finally arbitrary stable distributions.

Lemma 5.10.1. All symmetric stable distributions are unimodal.

Theorem 4.5.3 states that the function

f) = rape OO <* <2)

INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 159

is the characteristic function of a symmetric unimodal distribution.

According to theorem 4.5.5 this is also true for the function

falt) = fay = (ty

We conclude finally from theorem 4.5.4 that
ff) = lim fr(t) = exp (—| 217)
is also the characteristic function of a symmetric unimodal distribution, sc
that the lemma is proved.

Lemma 5.10.2. Stable distributions with |y| = K(«) and «#1 are
unimodal.
We consider first the case « < 1. The function 6(¢), which occurs in the
statement of theorem 5.9.1, has, according to corollary 2 to theorem
5.9.1, exactly one zero in the interval (0, z). Let o be this zero and write
formula (5.9.7) in the form

Pals)==90/024 {”0(6) exp [—s/—a(6)] ds

7,
- (6) exp [=a alg] ap}.
Let x be an arbitrary zero of the function p,, (x); we differentiate p,,(x)
with respect to x and see easily from corollary 1 to theorem 5.9.1 that
(5.10.1)

Pale)<5 55 a8? ala) {J HA) exp [38aH AB}=

We see from (5.10.1) that there exists only one value x, in the interval
(0, 00) at which p;,,(«) becomes zero. The corresponding distribution is
therefore unimodal. We have treated the case y = K(«) = «; the validity
of the lemma for y = —« follows from the relation (5.8.2c).
In the case where 2 > « > 1 we use theorem 5.9.2 and its corollaries
instead of theorem 5.9.1. Corollary 1 of theorem 5.9.2 shows that (when
y = —K(«) = «—2), p,,-2(x) is decreasing and never vanishes for x > 0.
The argument given for the case « < 1 can again be applied if one replaces
corollaries 1 and 2 to theorem 5.9.1 by corollaries 2 and 3 to theorem
5.9.2. If y = 2—« the result follows again from (5.8.2c).
For the discussion of the case where |y| < K(«) we need two lemmas.
The first expresses a stable density with parameters («, y) in terms of
densities with parameters («, 0) and («, «), while the second deals with a
transformation of x?),,, (x).
160 CHARACTERISTIC FUNCTIONS

Lemma 5.10.3. Let a <1 and0 < y < a; then

Pee paced:
1/a 1/a
sin (2 —y) sin sie
where a= | ——— and b=
. :: . TO
sin 5 4 sin >

Since log f,,,(¢) = —|t|* exp Eele][2

al we can easily verify that

(5.102) pole) = |”pal“) baal)

We note (see theorem 5.9.1) that p,,(u) = 0 for u < 0 and differentiate
(5.10.2) with respect to x and obtain the statement of the lemma.

Lemma 5.10.4. The function A(t, 0) obtained by substituting x = e~*,

Le), , : : Shu’ :
y = —o into x*p,,,(x) ts a harmonic function in the strip —020 < t < ©,
I
mK (x)
|o| < seats
We use formula (5.9.8) and put v = tx and obtain
Ls
X? Pay (x) = — Re |0 (iv) exp [iv—v*x~*2] dy
so that
(5.10.3) A(z, 0) = = Re [°(Ge) exp [ove] ao.
tt)

An elementary computation shows that

Aon 0
or i do2
so that the lemma is proved.
In our discussion we consider first the case 0 < « < 1 and assume also
that 0 < y < «. We conclude from the unimodality of p,,(«) [lemma
5.10.1] and lemma 5.10.3 that p;,,(x) > 0 for x < 0. We denote the smallest
zero ofp,,,(x) by x) = x9(y). Clearly x)(0) = 0 while x,(y) > 0.
We consider the strip 0 < x < «0,0 < y < «and denote by © the set
of all points (x, y) in this strip for which x > xo(y) and p,,,(*) > 0. In
order to prove that p,,(«) is unimodal, we must show that the set © is
empty. Let 9 be the closure of 9. The set D is obviously bounded. The
2a
mappingx =e ",y = pat takes the strip0 < x < 0,0 < y < ainto the
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 161

: I
strip —-o<t< es , and a set 9, corresponds in this mapping
to S. The function x? p..,(*) is transformed by this mapping into the
harmonic function A(t, o). Suppose now that the set 9 is not empty. It
follows then from the definition of 9 that the function A(z, o) vanishes on
the boundary of 9,. Since A(t, c) is harmonic we have then necessarily
A(t, «) = 0 for (z, c) € 9, . In view of the definition of 9 this is impossible,
so that 9 is necessarily empty. This proves that stable distributions with
0 < « < landy > 0 are unimodal; formula (5.8.2c) shows that the state-
ment is also true for y < 0. We consider next the case 1 < « < 2. We
conclude from (5.8.8b) that yp,,,(0) > 0 for y # 0 and use the mapping of
lemma 5.10.4 in the same way as in the case where « < 1 to show that the
stable distributions corresponding to parameter values 1 < « < 2,y #0,
|y| # K(«) are also unimodal.
We still have to consider the case « = 1. We write f(t|«, Cy, C,) for
stable characteristic functions in the Lévy canonical representation (see
theorem 5.7.2) and have

(5.10.4) = Cya
log f(t|a, Cy, C,)
0
(#1 itu ) du

i itt21 itu Pe
+C,« |. (« 1 Tout) wr

for 0 < « < 2. We see from formula (5.10.4) that

1
(5.10.5) lim s(t|1 ae C;; C;)== f(t) 1, Cy, Ce).

Since we have already shown that stable distributions with exponent « 4 1

are unimodal, we conclude from (5.10.5) and from theorem 4.4.4 that stable
distributions with exponent « = 1 arealso unimodal, so that theorem 5.10.1
is completely proved.

5.11 Self-decomposable distributions

We defined stable distributions by means of the functional equation
(5.7.2). It is possible to introduce other classes of characteristic functions
in a somewhat similar manner. As an example we mention the characteristic
functions f(t) which obey the relation
(5.11.1) ff =f(a)t
for every c (0 < c < 1), where f,(t) is some characteristic function. The
functions (5.11.1) were introduced by P. Lévy and A. Ya. Khinchine
[see Lévy (1937a), p. 192, or second edition (1954), p. 195], and one
calls this family of characteristic functions the class of self-decomposable
162 CHARACTERISTIC FUNCTIONS

characteristic functions®) [Loéve (1955)]. In the present section we dis-

cuss some properties of this class.

Theorem 5.11.1. All self-decomposable characteristic functions are infinitely

divisible.
We first show that a function which satisfies (5.11.1) never vanishes. We
give an indirect proof and assume that f(t) has zeros. Then there exists a
ty such that f(t) = 0 while f(¢) 4 0 for |t| < tf). It follows from (5.11.1)
that f,(t)) = 0 while f,(¢) # 0 for |t| < to. We see from theorem 4.1.2,

i-|o(9) |> 1-1)? =

putting n = 1 and ¢ = ¢,/2, that

; t to/2) . , : shes
Since sah4 = f(to/2) is continuous in c, we obtain a contradiction by
2 Ff(cto/2)
choosing c sufficiently close to 1, so that the functions f(t) and f, (¢) never
vanish.
To prove the theorem we note that

AOS fo-n(= ‘)=

and see that

Cte) Sa
Since f (¢) is continuous and never vanishes, we conclude that dictdealt)aol
uniformly in k (1 < k < n) and in every finite ¢-interval. It then follows
from the corollary to theorem 5.6.1 that f(t) is infinitely divisible.

Corollary to theorem 5.11.1. Let f(t) be a self-decomposable characteristic

function; then f,(t) ts infinitely divisible.
Let m < n; we rewrite (5.11.2) in the form

IIfu-val !t)TLfaunal : = f(d).

: : : m
Suppose that m increases as m increases in such a manner that — +c
n

(*) Some authors [e.g. B. V. Gnedenko and A. N. Kolmogorov (1954) and others] refer
to this family as the “‘L-class’’. We do not use this terminology in order to avoid confusion
with the #-class introduced in Chapter 9.
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 163
asn -> 00. The first factor then tends to f(ct) while the second factor tends
to a characteristic function f,(t), which by the argument used in the proof
of the theorem is infinitely divisible.
Our next aim is the determination of the canonical representation of
self-decomposable characteristic functions f(t).
Since f(t) is infinitely divisible, we can write it in the Lévy canonical
form (theorem 5.5.2) and see that

(5.11.3) log f(t) = ita—ott/2+{

. 2/ ¥
Gitu -1- {am
itu

es | itu )

+] (6 : 1+4u? ze),
where o?, M(u) and N(uw) satisfy the conditions of theorem 5.5.2.
Substituting tc for ¢ in (5.11.3), we obtain after a simple change of the
variable of integration
Salad Lr itt
(5.11.4) log f(ct) = ita, —o8c812/2+ | (4#—1- de )an(*)
1+? é
le: a |_
itu dN
uet
+ ( eo) (“)
where 0 < c < 1 and
od (1—c?)u8
dM (u) +e |. cal 'N (1).
pane oF |—o(1 + ¢?u?)(1 +4?) +0(1+¢?u?)(1 44?)
We see from (5.11.3) and (5.11.4) that

ein eth ene

F(ct) , |
ee)
ol(Ose
BCE G)
where a, = a—a,. According to the corollary of theorem 5.11.1,
f(t) = f(t)/F(ct)
is infinitely divisible; therefore (5.11.5) is its canonical representation and
we conclude from the uniqueness of this representation that M (u) — M( u/c)
and N(u)—N (u/c) must be non-decreasing. Therefore
M (u,)—M (u,/c) < M(u.)—M (u2/c)
Pane, rear anes < N(v,)—N(v2/c)
whenever
(Sill 7) Cea and 20 <2; 1<,05):
164 CHARACTERISTIC FUNCTIONS

Conversely, if the inequalities (5.11.6) hold for every c (0 < c < 1) then
F(t)/f(ct) is the characteristic function of an infinitely divisible distribu-
tion, so that f(z) is self-decomposable. Suppose now that M(u) and N(u)
satisfy (5.11.6) for all u,, u:, v,, v, for which (5.11.7) holds. We then have
(5.11.8) © N(u,/c)—N(u,/c) < N(ue)—N(u,).
Leta < bandh > Oand choose aand bso thatc = e*~°. We put v, = e**",
v, = e* so that v,/c = e’t", v,/c = e’. It then follows from (5.11.8) that
N(e?t")— N(e’) < N(e**")— N(e*). We write N(e’) = A(v); then
A(b+h)—A(b) < A(a+h)—A(a)
or if we put b = a+h = x (say)
A(x) > 4[A(x+h)+ A(x—h)].
The function A(x) is therefore concave and has everywhere finite left-
hand and right-hand derivatives. The right-hand derivative never exceeds
the left-hand derivative, and both are non-increasing as x increases. Since
A(v) = N(e’) we have
AX(v) = & N(e").
Putting u = e’, we see that wN’(u) is a non-increasing function. In
exactly the same way one shows that uM ’(w) is non-increasing.
Suppose conversely that the functions M (u) and N(u) have the property
that uM '(u) and uN’(w) are non-increasing and that 0 < c < 1. Then

“m’(*) 2>uM'(u) foru <0

c

“wy(*) <= uN) “forac> 0:

c c
From these inequalities we obtain (5.11.6) by integration, so that the
infinitely divisible distributions determined by the functions M(u) and
N(u) are self-decomposable. We have therefore obtained the following
result:
Theorem 5.11.2. An infinitely divisible characteristic function 1s self-
decomposable if, and only if, the functions M(u) and N(u) in its Lévy
canonical representation have left- and right-hand derivatives everywhere and
if the function uM'(u) is non-increasing for u < 0 while uN'(u) is non-
increasing for u > 0. Here M'(u) and N’(u) denote either the right or left
derivatives, possibly different ones at different points.

Corollary to theorem 5.11.2. All stable characteristic functions are self-

decomposable.
The corollary follows immediately from theorems 5.11.2 and 5.7.2.
For some time it was believed that all distribution functions of self-
INFINITELY DIVISIBLE CHARACTERISTIC FUNCTIONS 165

decomposable characteristic functions are unimodal. K. L. Chung showed

in Appendix IJ of his translation of Gnedenko—Kolmogorov (1954) that the
proof given there for this statement is not valid. I. A. Ibragimov (1957)
gave an example intended to show that there exist self-decomposable
distributions which are not unimodal. However, T. C. Sun (1967) pointed
out that Ibragimov’s construction contained an error, so that the question
of the unimodality of this class is still open. The only known result at
present is due to A. Wintner (1956) who proved that all symmetric self-
decomposable distributions are unimodal. L. Kubik (1961/62, 1962/63)
studied certain analogies which exist between the family of infinitely
divisible characteristic functions and the class of self-decomposable
characteristic functions. He characterized the latter class of functions in a
manner which is similar to the way in which theorem 5.4.2 characterizes
infinitely divisible characteristic functions.
We finally mention the semi-stable distributions introduced by P. Lévy
(1937a). They are defined by means of the functional equation
Hat) = @°4(t) (¢ 40,9 # 1)
for the second characteristic.
V. M. Zolotarev (1963) investigated the smoothness properties (absolute
continuity, differentiability, analyticity) of self-decomposable distribution
functions. These properties depend on the functions M and N and on
the presence or absence of a normal component.
 6 FACTORIZATION PROBLEMS—GENERAL
THEOREMS FROM THE ARITHMETIC OF
DISTRIBUTION FUNCTIONS

In the preceding chapter we discussed a number of examples which indi-

cated that the analogy between the factorization of integers and the de-
composition of characteristic functions is rather limited. While a great
number of remarkable decompositions of characteristic functions is known,
we have only few general results, and one has the impression that the
arithmetic of distribution functions has not yet reached a final stage in its
development. In this chapter we present the most important general
theorems concerning the factorization of characteristic functions, and this
treatment will be supplemented by Chapters 8 and 9. The separation is
justified by the different tools used: in the present chapter we deal with
problems which can be handled without using the theory of functions of a
complex variable, while complex variable methods are essential in deriving
the results discussed in Chapters 8 and 9.

6.1 Some notations and lemmas

For the investigation of the general factorization theorems we need
certain lemmas which we discuss in this section.

Lemma 6.1.1. Let f(t) be the characteristic function of a symmetric distribu-

tion, then
1— f(2t) < 4[11—f(@]
for any real t.
Since the characteristic function of a symmetric distribution is real
(theorem 3.1.2), the assertion of lemma 6.1.1 follows immediately from
theorem 4.1.2.

Corollary to lemma 6.1.1. Let f(t) be a characteristic function and suppose

that |f(t) |= 1 in some neighbourhood |t| < 6 of the origin. Then f(t) is the
characteristic function of a degenerate distribution.
To prove the corollary we apply repeatedly the lemma to the function
| f(t) |? and see that |f(t)| = 1 in every finite interval.
We next introduce an operation which is applicable to any characteristic
function. Let f(t) be an arbitrary characteristic function; then there exists
FACTORIZATION PROBLEMS—GENERAL THEOREMS 167

a real number a such that

Vtt) |2.0 for 0"< t <a!
For a fixed a satisfying this relation we define

(6.1.1) NyLfO] = No(N) = ~ Jlog | f(0 Ia

The following properties of this operator are easily established:
(i) N,(f) > 0
(il) N.(e") = 0
(ili) If f(®) = f,(#) fa(0) then N,(f) = Na (f+ Nalfy)
(iv) NaC) > | =| F00| at
(v) N.(f) = 0 if, and only if, f(t) is the characteristic function of
a degenerate distribution.
Properties (i), (ii) and (iii) follow immediately from (6.1.1); (iv) is a
consequence of the inequality
—log
|f@)| = —log [1-(1-|f@])] = 1-|F@|
while (v) is easily obtained from (ii) and from (iv).
The quantity N,(f) is a measure of the departure of the distribution
belonging to f(t) from a degenerate distribution. We will refer to N,(f)
as the N,-value of f(t).
The main object of this section is the proof of the following lemma:

Lemma 6.1.2. Let {F,,(x)} be a sequence of distribution functions and denote

by {f,,(t)} the corresponding sequence of characteristic functions. Suppose that
x = 0 ts a median of F,,(x) (n = 1,2,...) and that there exists a real
a > 0 such that
(6.1.2) lisnglV.(7,.)== 0;

then
lim F(x) = «(x).

We say that the point x = m is a median of the distribution function

F(x) if the inequalities F(m—e) < 3, F(m+e) > 4 hold for any e > 0.
The assumptions of the lemma imply that for m sufficiently large
fn(t) # 0 for 0 < t < a. Using (iv) we see that
[ -[f0 la < 2) [1-40 lat < 2N,(f,)-
Moreover it follows from lemma 6.1.1 that

[sae< 8 |1-| (0 Pate

["a-lme lla = 2 ["E-|A@0
168 CHARACTERISTIC FUNCTIONS

We combine the last two inequalities and conclude from assumption

(6.1.2) of the lemma that

tim js[1-| f(t) |*]ae = 0.

It is then easily seen that for m sufficiently large f,(t) 4 0 for0 < t < 2a,
so the argument which we used can be repeated. In this way we see that,
for every T > 0,

(6.1.3) lim {i[1—| f.(é) |"1at = 0.

We denote by F(x) = 1—F,,(—x—0) the conjugate distribution of
F(x) and write
(6.1.4) ~.B(x) = Bye) * Fels)
for the symmetric distribution whose characteristic function is |f, (¢) |?.
We denote by
Dae =| “24 ly
1 ay

(x) agile’
the standardized normal distribution and consider the distribution defined
by
(6.1.5) G(x) F., (x) * B(x)
whose characteristic function is

8n(t) = e**| f(t) |*.

From the inversion formula we see that
1el ay t
GG aG (a) --{ singee- "2 |F(t)
|?dt.
Since G,,(x) is a symmetric distribution this can be written as

G.()-a
==1 | sin twBealon pyre
eel sin. tx ety Eee
sin tx | siais on
— |. ; dt+ [|fn (4)|?—1] at.
We write

I, (x) = _{=
sinf w/e f,(2) |2— 1] at
and see that
G,,(x) = O(x)+J, (x).
Since for any T > 0
eee
TU
#727 | f(t) |?— 1]dt| < - ||
a
[=| |e
FACTORIZATION PROBLEMS—GENERAL THEOREMS 169
we conclude from (6.1.3) that
lint J, (x) = 0
so that
(6.1.6) Iam G, (x) = O(2).

It follows from the continuity theorem that

lim gy (8) = 27 lim [fy (2) |? = erh
so that
bir fi(eyier t
and therefore
(6.1.7) Lim F,,(x) = e(x).

We write (6.1.4) in the form

Pye) = f° Paw—y)
dP(9)
and get

Ps) > | Fa(e—y)4Fa(y) > Fae)Fa(0)

= Pi (eel F,(—2— 05);
where ¢ is an arbitrary positive number. Since by assumption x = 0 is a
median of F,, (x) we see that
F(x) > $F (x—2).
We conclude from the last relation and (6.1.7) that for any x < 0
(6.1.8) lim F(x) = 0.
On the other hand, we see by a similar reasoning that

1-F,(x) > [" [1-F,@—9)]

dF,(9)> W-Fe+6)]
so that for any x > 0
(03129); (lim (x) =ol.
Formulae (6.1.9) and (6.1.8) imply the assertion of the lemma.

6.2 General decomposition theorems

In this section we discuss three general theorems concerning the factori-
zation of distribution functions and characteristic functions. The first two
of these theorems are due to A. Ya. Khinchine, the last is due to H. Cramér.
170 CHARACTERISTIC FUNCTIONS

Theorem 6.2.1. Every characteristic function can be represented as the

product of at most two characteristic functions which have the following
property: one does not have any indecomposable factors while the other is the
convergent product of a finite or denumerable sequence of indecomposable
factors.
Let f(t) be an arbitrary characteristic function and denote the corre-
sponding distribution function by F(«). Since f(t) is continuous and
f(0) = 1, there exists a real a such that f(t) 4 0 if |t| < a; in the follow-
ing we fix such a value a and write N,(f) = «.
If f(t) does not have any indecomposable factors then the theorem
holds. We suppose therefore that f(t) has indecomposable factors. ‘Then
it is possible that f(t) has a prime factor) p,(t) such that N,(p,) > «/2;
it follows then from (iii) [see p. 167] that one can write
fi) = AOA
where N,(p,) > «/2 while N.(f,) < «/2. In this case we repeat the
procedure with f, (¢) but use «/4 instead of «/2 as the lower bound for the
N,-value of its prime factor. If f(t) has an indecomposable factor whose
N,-value exceeds «/4, then one obtains a decomposition

ft) =A) aOA®

where N,(p;) > «/4 (gj = 1, 2)(t) while every indecomposable factor g(t)
of f,(t) has the property that N,(g) < «/4. In the case where no indecom-
posable factor with N,-value greater than «/2 exists, we search for prime
factors p(t) which satisfy the relation N,(p) > «/4. In the case where such
factors exist one obtains a decomposition into at most four factors
f(t) = Pi(t) - - - Pn, (t) fo (2)
where 1 < n, < 3. Here the p,(t) are indecomposable factors and satisfy
the inequality N,(p;) > «/4, while every prime factor g(t) of f,(¢) has the
property that N,(g) < «/4.
We repeat this procedure and see that f(t) can be decomposed in the
following manner:
(6.2.1) f(t) = Pilt)Po(t)-- - Pm A) Se (2)
where the ;(t) are indecomposable factors such that N,(p;) > «/2*
(Vj = 1,2,..., m3 1 < m, < 2*—1) and where every prime factor g(t) of
f(t) has the property that Na(g) < «/2*.
It can happen that for some k > 1 the characteristic function f, (¢) has no
indecomposable factors. ‘Then our process terminates and we see that the
theorem holds. We must therefore prove the theorem only in the case

(*) We use the terms “prime factor” and indecomposable factor synonymously.
(t) Note that according to our construction Na(p,) > «/2.
FACTORIZATION PROBLEMS—GENERAL THEOREMS ‘Weal

where the factorization process does not terminate; the factors p; (t) then
form an infinite sequence. Since

D Ne(es) < Na(f)

we see that the series

is convergent, so that the sum

k+m
> Na(?3)
j=k+1
converges to zero as k tends to infinity; this convergence is uniform in
m(m > 0). We now apply lemma 6.1.2 and see that there exist real numbers
A,r, such that

lim e“4.,. TT Pe (2)= 1

uniformly in every finite f-interval |t| < T and v’ > v. We write
Px(t) = px(t) exp [¢a,(¢)] and see that

FOO Vast AS on (ii Bay() FO) oadenr.crico

k=v

where B,,,,(t) assumes only integer values. The left-hand side of (6.2.2) is
continuous, moreover B,,,(0) = 0; hence B,,,(¢) = 0 for sufficiently
large v and we have

tA,y+ & ,(t) = o(1) as v>oo.

It is no restriction) to Rae that w,(1) = 0 so that A,,, = o(1) and

> @,(t) = 0(1) as’) © =="eo.

k=v
We see therefore that

(6:2:3) lim TIPs (2)= 1

v0 k=v
uniformly in |¢| < Tandv’ > w.
The infinite product

Ip. (t)
is then convergent; let v(t) be its limit. It follows from (6.2.3) that

v(t)= lim 1p.

k>o j=

(*) This can be seen if one multiplies each px (t) by exp [—itwx (1)].
172 CHARACTERISTIC FUNCTIONS

where the convergence is uniform in every finite ¢-interval. We see then

from the second version of the continuity theorem (theorem 3.6.2) that
v(t) is a characteristic function. Let e > 0 be an arbitrary positive number;
according to (6.2.3) we have
k+m

II #;(@)-1 <e (m>0,|t|

< T)
j=kt1
if k is sufficiently large. Since
(0) = frsm(®) TD Ps(0
Ne+m

IFN

one can conclude easily that

Ife()—frrm(t)| < € (m > 0, |t| < T).
This means that the sequence {f;,(£)} also converges to a characteristic
function. Denote
w(t) = lim f,(8)-
k—>co
It follows from (6.2.1) that
f(t) = v(t) u(t).
The function u(t) has no indecomposable factor; this follows easily from
the fact that each indecomposable factor of u(#) must be a prime factor of
fi. (t) for all k. But such a factor cannot exist since the N,-values of the
prime factors of f,,(¢) tend to zero as k goes to infinity. This completes the
proof of theorem 6.2.1.
The second decomposition theorem supplements the preceding result
by characterizing the distributions which have no indecomposable factor.
Theorem 6.2.2.%) A characteristic function which has no indecomposable
factor is infinitely divisible.
Let f(t) be a characteristic function which has no indecomposable
factors and denote by D
(D) f(t) =A@fe4)--- fal)
an arbitrary decomposition of f(t) where all f;(¢) are characteristic func-
tions. Suppose that @ is a positive number such that f(t) 4 0 if |t| < a;
we write then
v(D) = max N,(f;)
1<j<n
and denote by » the greatest lower bound of the »(D) for all possible
decompositions D of f(z).
We first derive a lemma:
Lemma 6.2.1. Let f(t) be a characteristic function which has no indecom-
posable factor, then vy = 0.
(*) The converse of theorem 6.2.2 is not true. This will be shown later.
FACTORIZATION PROBLEMS—GENERAL THEOREMS 173
It follows from the definition of y that there exists a sequence of decom-
positions {D,,}, say
(Dn) f(t) a Sua (t) fo (t) ssa Fax, (t) (n = 1,2,.. .)
for which »(D,,) converges to y so that
Pree t), eee ere (hess 1. 2. 1c).
Let f{”(t) be the factor of D, for which »(D,) = N,(f”) and write
fs”
(#) for the product of all other factors of D,,. Then
eaN Af) See lin (2 = "1,2, - 3.)
(6.2.4)
f() =f Of"O.
Let FY” (x), FY («) and F(x) be the distribution functions corresponding
to f\”, ff and f respectively; it is no restriction) to assume that x = 0 is
a median of FY" (x). According to Helly’s first theorem the sequence F”’ (x)
contains a convergent subsequence, it is only a simplification of our nota-
tion if we assume that the sequence F'” (x) itself is a convergent sequence.
We prove next that the limit of this sequence is necessarily a distribution
function; this is established if we can show that for every 7 > 0 and suffi-
ciently large values of a > 0 and n one has
F™(—a) <7 while F(a) > 1-7.
We carry the proof indirectly and assume that one of these inequalities,
for instance the second, is not satisfied for a > 0 and arbitrarily large n.
Then for any b > 0

1—F(b) [- U-FY 6-y14Fe'O)

> |"—1 LF e-y]4F Py)
ZL Py Oia
Fe (—1)) 2 an:
This however contradicts the assumption that F(x) is a distribution
function. The proof of the inequality F{”(—a) < » for sufficiently large
a > 0 and nis carried in a similar manner and so we have shown that
Vm F(x) == 1 (2)
is a distribution function. We write f, (t) for the corresponding character-
istic function. We consider next the sequence FY” («); it also contains a
convergent subsequence and we use again the notation Fy” («) for this
convergent subsequence. Let
(2) pa eimele (ax)5

(*) This can always be accomplished by a translation which does not affect the Na-values.
174 CHARACTERISTIC FUNCTIONS
we show that F(x) is also a distribution function. We have for any
a>OQandb>0

1-F(a) > |o [1 —F” (a—y)] dF2(y)

> [1 F{” (—d)][1— FY (a+8)].
For sufficiently large 6 we have
Lim F”(—6) = F,(-6) < 4

and therefore
1—FY (a+b) < 2[1-—F(a)].
This indicates that the left-hand member of this inequality tends to zero
as a increases; in a similar manner one can show that F{(«) tends to
zero as x goes to — oo provided that 7 is sufficiently large. Thus F(x) is a
distribution function and
f(t) = lim ff?(0)
is its characteristic function. It follows from (6.2.4) that
f(t) = AOA
while
NG.) = J:
We next show by an indirect proof that » < 4N,(f). Let us therefore
suppose that »y > 4N,(f); it follows from (iii) (page 167) that N,(f2) < ».
If we decompose f, and f, once more, we obtain a decomposition
(D*) f= 81 82 83 &4
which has the property that »(D*) < y. But this contradicts the definition
of v, so we can conclude that
vy < $N,(f).
Therefore there exists a decomposition D of f(t) such that
oD) < 2Na(f)-
Each factor of D is a characteristic function without indecomposable
factors and we can apply the result to a factor of D and see that
vy < 4o(D) < 4N,(/).
We iterate this procedure to obtain the statement of the lemma.
We proceed to prove theorem 6.2.2. Let {e,,} be a sequence of decreasing
positive numbers such that
lim ¢, =.0;
Nc

It follows from lemma 6.2.1 that there exists, for each n, a decomposition
(D,) f(t) = Las (t) fn (t) Sys ae (t)
FACTORIZATION PROBLEMS—GENERAL THEOREMS 175
such that

SA Pel) < Ey
0b p = lo, oe. eee while
lim Re. =)00,

Using lemma 6.1.2 we conclude that it is possible to find constants

Geet = 1h any kapiti= 122 205)csuch ithat
ty {Fn(t) exp [tte5]} = 1
where the pavement is uniform for 1 <j < k, and in every finite
interval |t|< T.
We write
Frag(t) = Png(#)exp [to,,; (2)]
25)
6.2.5
y= ivesp et
where w,, ;(0) = w(0) = 0.
Then
lim p,,;(t) = 1 and Hee [w,,;(¢)+ta,,;] = 0

(uniformly in |t| < T and for 1 < k,,). We see then that
lim [o,,,j (l)
ale
pene
ped o

hence also
lim [,;(t)—
te,;(1)]= 0
Writing
AGH 1.)
gn,5 (t) ae Pe (t) exp it k ra On, j (1)

we see that

(6.2.6) Ora) ng (t)—tw,;(1) + ail

exp {iEB

and conclude that

(6.2.7) lim g,;(t) = 1;

the convergence here is uniform in |¢| < T and for j = 1,2,..., Rp.
From the relation

f= juThas(®
kn

and from (6.2.5) and (6.2.6) we can easily see that

fl) = Teas
176 CHARACTERISTIC FUNCTIONS

It follows from (6.2.7) that the assumptions of the corollary to theorem

5.6.1 are satisfied; we finally conclude from this corollary that f(t) is
infinitely divisible.
The first two theorems discussed in this section show that there are three
possibilities for the product representation of an arbitrary characteristic
function f(t):
(1) f(t) has no indecomposable factor (in this case it is neces-
sarily infinitely divisible) ;
(II) f(t) is the product of a finite or denumerable sequence of
indecomposable factors;
(III) f(z) is the product of two characteristic functions f,(t) and
f(t) where f, (t) has no indecomposable factors, while f,(t) is
the finite or denumerable product of indecomposable factors.
The decomposition is in general not unique; this is illustrated by the
example discussed in Section 5.1 (p. 104) and also by the multiple factor-
ization of the characteristic function g(t), defined by (5.5.12). The first of
these examples refers to a purely discrete distribution, while the function
(5.5.12) is the characteristic function of an absolutely continuous distri-
bution. In the next section we will find some further examples of multiple
factorizations of absolutely continuous distributions.
The converse of theorem 6.2.2 is not true. We have already given (page
124, Section 5.5) an example of an infinitely divisible characteristic
function which is the product of an indecomposable characteristic function
and an infinitely divisible characteristic function. The characteristic
function

fi) = 2-5 @>)

of the geometric distribution which we considered at the end of Section
5.4 permits an even more remarkable factorization. It is easily seen that
p-l og (ee) p™ +e

pe ko 14+p”-
We see therefore that the characteristic function of the geometric distri-
bution is infinitely divisible but admits nevertheless a representation as a
product of an enumerable sequence of indecomposable characteristic
functions.
Our next theorem indicates that the existence of infinitely divisible
distributions with indecomposable factors is not a rare occurrence.

Theorem 6.2.3. An infinitely divisible characteristic function g(t) whose

Lévy canonical representation is determined by the constants a = o = 0 and
FACTORIZATION PROBLEMS—GENERAL THEOREMS L7y
by the functions
Eiie
Nw) = {i c) pte she
and M(u) = 0
0 otherwise
always has an indecomposable factor. Here k > 0 and c >0 are arbitrary
real constants.
The characteristic function g(t) is then given by
Cc

log g(t) =k | G —1- as) du.

0 1+?
Let ¢ be a real number such that 0 < e¢ < } and introduce the functions
—€ if (4—e)e < u < (4+8)c
(6.2.8a) 0, (4) = f elsewhere in (0, c)
0 outside (0, c)
(6.2.8b) aa (u) = {1 awe if (t-e)e ($— < u < ($+4+8)c
3
0 if |u—c/2| > ec
WEES tu
DBS)
(6.2.8c) tuog
ct iyg;(t) =k=k |
ae(em 1- 1 a Na, oy (ayaa Gij== 1h) 1,2).
Clearly k[a, (u)+«,(u)] = N’(u) so that
(6.2.9) log g(t) = log g, (t) + log g,(t).
The function g,(t) is, according to the representation theorem 5.5.2, the
characteristic function of an infinitely divisible distribution. We show next
that g,(z) is also a characteristic function provided that ¢ is sufficiently
small.
We define a sequence of functions 6,(x) by means of the recurrence
relations
B1(x) = ka, (x)

Ba(s) = | Baa) Bx(0) dt =| Bnale—

ha(Oat.
We conclude from these relations and from (6.2.8a) that 6,,(x) = 0 if
either x < 0 or x > nc and note also that
Vent) ese ie Coat:
We remark that £,,(«) is the m-fold convolution of £, (x) with itself so that
ifep, (x) dx = al iEe' x, (0) as |"
The series : :
E(/n!)i,(6)
is absolutely and uniformly convergent; therefore

(6.2.10) ineite pa oh ()|ie 2 exp E |.ote ce (2) as |ar

178 CHARACTERISTIC FUNCTIONS

We write
Cc Cc x

A= k [.a,(x)dx and n»=k |.er (x) dx

and obtain from (6.2.8c) and (6.2.10)

ey ee
ent) = [1+ lsCe (= 72.08) ds |exp (—A—itn).
Let
x

Gi(x) = {etn + | [2 “iao+n| ay},

then

AO= \\a el dG), (x).

In order to show that g, (¢) is a characteristic function we must prove that
G(x) is a distribution function. Clearly G,(— ©) = 0, while we see from
(6.2.8c) that G,(+ 0) = g,(0) = 1. We must still show that G,(x) is
non-decreasing; we do this by proving that
4a
2 n! Bn (*)
is non-negative for all x, provided that ¢ is chosen sufficiently small.
We first remark that £,(x) is non-negative except in the interval
(4-«e)c < x < ($+8€)c; moreover it is easily seen that f,(x) tends to
k?(c—|c—«x|) uniformly in the interval 0 < x < 2c ase tends to zero. One
can also show that £.(«) and /3(x) as well as
1 1
Bx () +5) Ba(*) + 37Bs(*)

are non-negative for all real x if ¢ is sufficiently small. Let ¢ = &) be such
a value. For n > 4 we can rewrite the relations defining the £,,(x) in the
form

Rony Pat) Ba( a (=o

and see that f, (x) > O for all x and all nm > 2 if e = &.
Therefore
eal
canta ae
and G,(x) is the distribution function whose characteristic function is
£1 (t). We show next by means of an indirect proof that g, (f) is not
infinitely divisible. Let us then assume tentatively that g,(¢) is an
infinitely divisible characteristic function. Then g,(t) admits a Lévy
FACTORIZATION PROBLEMS—GENERAL THEOREMS 179
canonical representation with some non-decreasing function N, (2). The
derivative N,(t) of N,(t) exists almost everywhere and is non-negative.
It follows from the uniqueness of the canonical representation that the
relation
N'(x) = k = N,(x)+ha.(x) > kag (x)
is valid almost everywhere in the interval (0,c); but this contradicts
(6.2.8b). Hence g, (¢) cannot be infinitely divisible. It follows from theorem
6.2.2 that g,(¢) must have an indecomposable factor, and we see from
(6.2.9) that this is also true for g(t), so that theorem 6.2.3 is proved.
In exactly the same way in which we proved theorem 6.2.3 it can be
shown that an infinitely divisible characteristic function whose Lévy
canonical representation is determined by the constants a = o = 0 and
by the functions N(u) = 0 and M(u) = k(u+c) for —c < u < 0 and
M(u) = 0 for u < —c or u > 0 always has an indecomposable factor. We
can therefore reformulate our result:

Corollary 1 to theorem 6.2.3. Let f(t) be an infinitely divisible characteristic

function and suppose that the functions M(u) and N(u) which occur in its
canonical representation satisfy the following condition: there exist two
positive constants k and c such that at least one of the relations M'(u) > k
almost everywhere in (—c,0) or N'(u) > k almost everywhere in (0, c) holds.
Then f(t) has an indecomposable factor.
We consider only the case where the condition is satisfied in the interval
(0, c). Then f(t) has an infinitely divisible factor of the form required by
theorem 6.2.3, so that the corollary is established.
Let f(t) now be a characteristic function which satisfies the conditions of
this corollary. The function f(t) can be written as an infinite product

f= TELM.
Each factor [f(¢)]? * also satisfies the conditions of the corollary, so that we
obtain the following result:

Corollary 2 to theorem 6.2.3. Suppose that the infinitely divisible character-

istic function f(t) satisfies the condition of the corollary 1, then it is divisible by
the product of an infinite sequence of indecomposable characteristic functions.
We finally remark that the conditions of corollary 1 are satisfied by the
Gamma distribution and also by all non-normal stable distributions.
Generalizations of theorem 6.2.3 were given by R. Shimizu (1964)
[see also B. Ramachandran (1967)]. We mention here only one of his results
to which we shall refer later.

Theorem 6.2.4. An infinitely divisible characteristic function g(t) whose Lévy

180 CHARACTERISTIC FUNCTIONS

canonical representation is determined by the constants a = o = 0 and the

functions ‘
k(b—c) for0O<u<b
N(u) = fram jor baa <¢ ‘and Min) =0
0 joruse
always has an indecomposable factor. Here k > Q is an arbitrary constant,
while the constant c satisfies the inequality 0 < 2b < c¢.
The proof is similar to the proof of theorem 6.2.3. We select first a point
d such that 2b < d < cand a number « > 0 so that
2b < d(1—e) < d(lt+e) < c.
We define
—e if dl—«)<u< d(l+e)
1
0 (1) -| elsewhere in (6, c)
0 outside (8, c)

aa(u) = {l+e0
if dl—e)
<u < d(1+e)
otherwise.
The functions «,(u) and «,(u) determine again—according to formula
(6.2.8c)—two functions g,(¢) and g,(¢). The function g,(¢) is an infinitely
divisible characteristic function, while it can be shown that g,(t) is, for
sufficiently small ¢, a characteristic function but not infinitely divisible.

6.3 Indecomposable characteristic functions

We have already given several examples of indecomposable character-
istic functions, and we showed at the end of the last section that a rather
wide class of infinitely divisible characteristic functions has indecompos-
able factors. This, however, is almost the only general theorem concerning
indecomposable characteristic functions which we know at present. There
is no general method for finding the prime factors of a given characteristic
function; our knowledge consists mostly of interesting special examples.
In the present section we will make a few general remarks about prime
factors and also list a number of remarkable decompositions.
We consider next an arbitrary distribution function F(«) (which is not
assumed to be of a pure type) and suppose that it is the convolution of two
distribution functions F, (x) and F(x):

F(x).= leeF, (x—1) dF (2).

Suppose that € is a point of increase of F',(«) and 7 a point of increase of
F, (x); it is then easy to see that £ + 7 is a point of increase of F(x). Similarly,
if € is a discontinuity point of F(x) and y a discontinuity point of F, (x)
FACTORIZATION PROBLEMS—GENERAL THEOREMS 181
then +7 is a discontinuity point of F(x).(+) It can also be shown that
every discontinuity point ¢ of F («) can be written in the form ¢ = +7
where & is a discontinuity point of F, (x) while 7 is a discontinuity point of
F, (x).
Let {&;} and {n;} be the (finite or enumerable) sets of discontinuity
points of F,(«) and F(x) respectively and denote’ the discontinuity
points of F(x) = F,(x)*F,(x) by {¢,}. Since the elements of the set {¢,!
can be written in the form {€;+7,} we see that every difference £;— &, must
occur among the differences of ¢;—¢;, at least as many times as there are
different 7-values.
From these considerations we obtain easily the following results:
(1) Suppose that all the differences €;—¢;, between the discontinuity
points of a purely discrete function F(x) are different. Then F(«) is an
indecomposable distribution function.(§)
(2) Suppose that the purely discrete distribution function F'(«) has at
least n? discontinuity points and that it is not possible to find m pairs of
discontinuity points which differ by the same number. Then F(x) is
indecomposable.
(3) A finite distribution which has two discontinuity points, one at each
extremity, is always indecomposable.
The last statement follows from the earlier remarks and from the
relations
lext [F',* F,] = lext [F,]+lext [F,]
rext [F,* F,] = rext [F,]+rext [F,]
which hold for the convolution of two finite distributions Ff, and Fy.
In Chapter 3 (corollary to theorem 3.3.3) we showed that the number
N of discontinuity points of a purely discrete distribution function has
lower and upper bounds which are determined by the numbers of dis-
continuities of its factors. It can be shown that this corollary is also true if
the distribution is not purely discrete, the same limits (n+m—1 < N < nm)
are valid also in the general case. We then obtain by induction the following
result:
(4) A purely discrete distribution which has exactly m+ 1 discontinuity
points has at most m indecomposable factors. This maximum can only be
attained if the discontinuity points are the (n+ 1) consecutive terms of an
arithmetic series.

(t) These statements follow from the inequality

F(E+n+ha)—F(€+9—M) > [Fi (E+h2—he) —Fi (E—Ia + Ry) MP 2 (9 +e) — Fa (n —Ai)].-
Here
ley SO, tig
> ©, Te SO, ky =O
(§) We say that the distribution function F(x) is indecomposable if its characteristic
function is indecomposable.
182 CHARACTERISTIC FUNCTIONS

We consider next a purely discrete distribution whose discontinuity

points are the consecutive terms of a finite arithmetic series. It is then no
restriction(*) to assume that this series consists of the integers 0, 1, 2,..., 7.
In studying this class of distributions it is more convenient to use the
probability generating function than the characteristic function. Let a), be
the saltus of the distribution at the point k (k = 0,1,...,m); the prob-
ability generating function is then the polynomial

(63:1) enaP(g) =a See eee eles tee

k=0 k=0

The substitution y = e” transforms P(y) into the characteristic function

f(t) = P(e); the corresponding distribution function is given by

F(@)'= 2 a, e(x— k).

Each decomposition of f(t) corresponds to a factorization of the generating

function P(y) into the product of polynomials with non-negative co-
efficients. If no such decomposition exists then f(t) is indecomposable. ‘The
number of factors of F(x) reaches its possible maximum if, and only if, the
generating function has 7 real, negative zeros. We assume next that all
coefficients of the polynomial (6.3.1) are equal and write
1 nm—1 1—y"
(Ge pa
) oaila Scream rl a
for the generating functions of this sub-class. Since
Pad) = 2) Pa = Ea) Pl)
we see that the distributions of this class admit multiple decompositions,
provided that the index m is not a prime number. Using the present
notation we could rewrite the example at the end of Section 5.1 in the
form P,(y) = Poy) Ps(y*) = Ps(y) Po(y?). If
Ni PP py. Pe (Oy FO. Oo, en)
is the decomposition of m into prime factors, then one can obtain in this
way
h!
Oy! oe! ... a!
different decompositions of P,(y) into prime factors. M. Krasner—
B. Ranulac (1937) have shown that these are the only decompositions of

The problem of decomposition into prime factors is therefore com-

(*) Since the decomposition properties of distribution functions are invariant under
linear transformations.
FACTORIZATION PROBLEMS—GENERAL THEOREMS 183

pletely solved) for the family of distributions with equal and equally-
spaced jumps.
The indecomposable distributions which we have so far studied all had
a finite set of discontinuity points. We show now that an indecomposable
distribution can have an enumerable set of discontinuity points and can
also be absolutely continuous or purely singular.
Let {p,} be the sequence of prime numbers and suppose that the
distribution function F(x) has its discontinuity points at &, = log p,
(v = 1, 2,...). It is then clear that the differences between discontinuity
points are all different, so that F(x) is necessarily indecomposable.
The following lemma will be used in our construction of a characteristic
function which belongs to an absolutely continuous, indecomposable
distribution function.

Lemma 6.3.1. Let p(x) be a frequency function which has a normal com-
ponent, then p(0) > 0.
If p(x) has a normal component then it can be written in the form

P(*) = oO 27
iy
—o
exp [| dF(y)
where F'(y) is a distribution function. Hence

(0) = = [exp [-“S3" ]arn) > 0

and the lemma is proved.
Let now

(6.3.2) f(t) = (1—-#)e"*? = eo)

Since

we see that
1 oa it ae
A={===
yomlh 2exp
chica fle s
| itx—— | dx.

It follows then that f(t) is the characteristic function of the density

1
P(x) = AVG x27 0/2,

(*) According to a remark by J. Hadamard, which is appended to the paper by Krasner—

Ranulac, this problem was also solved independently and simultaneously by A. Liénard
and D. A. Raikov.
G
184 CHARACTERISTIC FUNCTIONS

Since p(0) = 0 we see from lemma 6.3.1 that f(¢) has no normal component
and therefore also no component of the form
(1 —#?) exp (—o?#?/2) (0? < 1).
The only possible factors) have either the form (1—t)e~*/” or the form
(1+t)e-*/. These factors do not satisfy the necessary conditions of
theorem 2.1.1 and cannot therefore be characteristic functions. Hence f(t)
is indecomposable.
We have thus demonstrated that an absolutely continuous distribution
function can be indecomposable. We add a few remarks which refer to the
presence of normal components.
Since the characteristic function (6.3.2) is indecomposable we see that
h(t) = (1-42) e#
is also an indecomposable characteristic function. We consider next the
characteristic function
(6.3.3). g(t) = [AQ]? = (1-8? +44) 6"?
and show that it has a normal component.
This is the case if
£1(t) = (1—#2+ }*) exp (— A?#2/2)
is a characteristic function for some A such that |A| < 1. Let
Loe
Pix) = ae te ew ga(iat;

it is easily seen that

ata) ea) 11-2daeod 8

(6.3.4) pilx) = 1 3 aa
1/x\4 ‘ 3 Ee
1

The function p, (x) is a frequency function if, and only if, the polynomial

per( Bore) 3 i 3

is non-negative for all real z. It is easy to see that this is the case if A? > 3.
Moreover, for A® = ¢ we see that p,[+(3+/2)/4] = 0 so that g,(t) can
have no normal component. It follows then that for 42 = 3 the characteris-
tic function
x(t) = (1—£2+ de!)exp (—342/8)
is indecomposable.
We have just shown that the product of two characteristic functions

(*) This follows from theorem 8.1.2.

FACTORIZATION PROBLEMS—GENERAL THEOREMS 185

without normal components can have a normal component and have also
obtained the factorization
g(t) = (1—t84+ Jute"? = (1 —42tye"/4
= [(1 = £2/2)%e—#/8] Pa

where the factors (1—4¢?)e~*/* and (1—22/2)2e—**/8 are indecomposable

characteristic functions.
Our next example will show that the Poisson distribution has a similar
property. The product of two characteristic functions which have no
Poissonian factor may have a Poissonian component.
Bet 4 >. 1;

and
A2—1 \12
y2 (t) = e&
—? :

Clearly f,(t) is the characteristic function of an indecomposable law, and

it is not difficult to show that f,(¢) is also a characteristic function. More-
over we can write

Alt) =-Az exp [2 (-191

falt) =VO-F) exp [Ecl
Then

is a characteristic function which has the Poisson factor

exp {(1/4)(e"—1)}
while neither f,(t) nor f,(t) have Poisson factors. This is trivial for the
indecomposable characteristic function f,(t). If, on the other hand, f.(¢)
would have a Poissonian factor exp [u(e“—1)] (wu > 0), then

fa(t) exp [—w(e"—1)] = e*V(1—2-*) exp E ert ia

would be a characteristic function; if we expand the right-hand side of this
equation according to powers of e we then see that the coefficient of e” is
negative, so that f,(t) exp [—(e“—1)] cannot be a characteristic function.
This shows that f,(¢) has no Poissonian component.
The characteristic function (6.3.2) is an example of a characteristic
function of an absolutely continuous, unbounded distribution function
which is indecomposable.
186 CHARACTERISTIC FUNCTIONS

We next derive a theorem which enables us to construct certain interest-

ing examples of indecomposable distribution functions.

Theorem 6.3.1. Let F(x) be a distribution function and n be an integer.

Write
(6.3.5) pa = F(n+1-0)—F(n—-0)
and define the distribution function F(x) by

(6.3.6) F(x)= Dd pne(x—n)

and introduce for n, for which p, > 0, distribution functions F(x) by

0 if xe <0

(6.3.7) F(x) = 5 em) F 0) 70 261

1 af eS Me
Suppose that
(a) Papi = Ofor all k;
(b) Po Ze 0;
(c) F(x) 1s tndecomposable ;
(d) the distribution functions F(x) have no common, non-degenerate
factor.
Then the distribution function F (x) is indecomposable.
We give an indirect proof and assume that /'(x) admits a decomposition
(6.3.8) if Get ii.
where G and H are both non-degenerate distributions. We introduce the
quantities
Qn = G(n+1—0)—G(n—0) and », = H(n+1-—0)—H(n—-0)
and define the distribution functions G(x) and G,,(x) [respectively H (x)
and H,,(x)] corresponding to G(x) [respectively H(x)] by replacing, in
formulae (6.3.6) and (6.3.7), p, and F by q, and G [respectively r, and H].
In view of (b), we can assume without loss of generality that
(6.3.9) Go 20 Range ta e0:
We next show that
(6.3.10) F=G*H.
It follows from (a) and (6.3.8) that

pa FQRAI—D)—FCRD)= yal dG(x) dH(y).

2k <at+y<2k+2
FACTORIZATION PROBLEMS—GENERAL THEOREMS 187

On the other hand it is easily seen that

DS amta=
m+n=2k
|| G(s) aH(y)
2k <ae+y<2k+2
so that
(6.3.11) Poa = » Im Tn
m+n=2k

and (6.3.10) follows. We conclude from (c) that one of the factors in
(6.3.10) must be degenerate. Let H be this factor. We see from (6.3.9) that
(6.3.12) Fo= 137, =0 for, 40
and
(693.13). Pen = Cap
In view of (6.3.12) we have
(6.3.14a) r, = H(1—0)—A(—-0) = 1.
Moreover we see from (6.3.13) and assumption (a) that q.,_, = 0 or
G(2k—0)—G(2k—1—0) = 0;
hence
(6.3.14b) G(2k—y—0) = G(2k-0) for0 <y <1.
We next show that for a k such that p., > 0 the relation
(6:3.15) Fg =, Ga, ©
holds. We have
Gy, * H = | Gy.(x—y) dH(y)
or, using (6.3.14a) and (6.3.14b),
1
Gy Hia _| [G(«—y+2k)—G(2k—y—0)]dH(y
2k 0

1
a9 [F («+2k)—F(2k-0)] = Fy,(x),
2k

so that (6.3.15) holds. We conclude from assumption (d) that H (x) is

degenerate and the statement of the theorem is proved.
We use the theorem to construct examples of indecomposable distribu-
tions which are absolutely continuous or purely singular.
Let A and B be real numbers such that 0 < A < 1, 0 < B < 1 and
A/B is irrational. The function
0 if 0 Sox
Ky 2 ik a4
E(sya¢ A/2 fA<x<2
[AB+(2—A)(x—2)]/(2B) if2<x%<2+B
1 fx 22+8B
188 CHARACTERISTIC FUNCTIONS
is then an absolutely continuous distribution. We have
Po = A/2, Pp = 1-A/2, pp = 0 (GF #0, 2).
F(x) = Zetx)+(1-4) e(x—2), so that conditions (a), (b) and (c) of

theorem 6.3.1 are satisfied. We also have

0 ifx<0 0 ifx <0

Fy(x) = - if0<«<A and F,(x)= 3 if0<x<B

1: MittA-<iy 1 wt Box:
The corresponding characteristic functions are
_ ita/2 sin At/2 _ iB/2sin Bt/2
Te (t) é At/2 and f2(t) é Bt/2

respectively. Since A/B is irrational we see that f)(¢) and f,(¢) have no
common zeros, and therefore also no common factor, so that condition
(d) is also satisfied. Therefore /'(x) is indecomposable. ‘This example has
already been given by P. Lévy (1952), and theorem 6.3.1 is a modification
of his construction.
We now construct an indecomposable, purely singular distribution. Let
Pees is
pie [J cos a
j=1

be the characteristic function of the singular distribution constructed in

Chapter 2 over Cantor’s ternary set [see pp. 8 to 9 and formula (2.1.7)]. We
write C(x) for the corresponding distribution and put
CA (t) = c(At)
and write C4(x) for the distribution function of c4(t). Let again A and B
be two real numbers such that 0 < A < 1, 0 < B <1 and A/B irra-
tional. Let p, and p, be two positive numbers such that p)»+p, = 1. Then
it is easily seen that
F(x) = poCa(x)+p2Ca(x—2)
is a purely singular distribution function. Moreover it satisfies the condi-
tions of theorem 6.3.1 and is therefore indecomposable.
We next discuss several factorizations of the rectangular distributions.
The characteristic function of the rectangular distribution over the range
fel eal as
==
j=
sin t

Clearly
sint 2sin(#/2) cos (t/2) _ sin (t/2)
cos (t/2).
t t Zz c
FACTORIZATION PROBLEMS—GENERAL THEOREMS 189
We iterate this procedure and get

Hi)=e ([ ]ooWhe Oe)

n) t/2”
e=1
Since
sin (t/2")
ae ae
cohol} So Malice
we see that r(t) can be written as the product of an infinite sequence of
indecomposable factors:

(SRG) pers sia = [J cos—.

We know that the rectangular distribution is not infinitely divisible;

formula (6.3.16) indicates that it is possible to decompose a characteristic
function which is not infinitely divisible into a product of infinitely many
indecomposable factors. The decomposition (6.3.16) is not unique. It is
easy to verify that
sin¢ _ sin (t/3) |?cos (2t/3)+ |
(5D Faded Ui; 3 .
We repeat the procedure which led to (6.3.16) and obtain
(6.3.17) eye sin t - |?cos oe Aha nes “as
k=1

We have therefore obtained two different representations of the charac-

teristic function 7(t) as an infinite product of indecomposable factors.
We give next another interesting decomposition of the rectangular
distribution. Let
eS t
A@®= Liens real
(6.3.19) x f
f2(t) = [[ cos aa
k=1

It follows from the corollary to theorem 3.7.9 that f,(f) and f,(¢) are
characteristic functions of convergent symmetric Bernoulli convolutions.
Moreover we conclude from the corollary to theorem 3.7.10 that f,(¢) as
well as f,(t) are characteristic functions of purely singular distributions.
The product

FAOKACE Mes peer

and we see from (6.3.16) that

(6.3.20) (5) ET AOVACY

190 CHARACTERISTIC FUNCTIONS

Formulae (6.3.19) and (6.3.20) show that it is possible to represent the

rectangular distribution as the convolution of two purely singular distribu-
tions.
Several authors have raised the question whether it is possible to repre-
sent the rectangular distribution as a convolution of two absolutely con-
tinuous distributions. T. Lewis (1967) showed that this is not possible,
and that the convolution of two absolutely continuous distributions cannot
be a rectangular distribution.
We also note that r(t) is the characteristic function of a finite distribution ;
we will show later [corollary 3 to theorem 8.4.1] that the characteristic
function of a finite distribution is always the product of a finite or of a
denumerable sequence of indecomposable factors.
Using the examples of decompositions discussed so far, we are in a
position to summarize the Lebesgue properties of convolutions:
(A) The convolution of two discrete distributions is always discrete.
(B) A convolution which contains one absolutely continuous com-
ponent is absolutely continuous.
(C) The convolution of a discrete and a singular distribution is
always singular.
(D) The convolution of two singular distributions is continuous.
It is either purely singular or purely absolutely continuous, or
a mixture of a singular and an absolutely continuous com-
ponent.
Remarks. Statements (A) and (B) follow from theorem 3.3.2. (C) is a
consequence of some results contained in a paper by H. Tucker (1965). In
connection with (D) it is interesting to note that H. Tucker constructed an
infinitely divisible distribution F(x), produced by a purely discrete function
6(x) in the Lévy—Khinchine representation, such that F is purely singular
while F * F is absolutely continuous. Theorems 3.7.9 and 3.7.10 can also
be used to construct examples of purely singular distributions whose con-
volution is singular. It would be interesting to have a necessary and
sufficient condition which assures that the convolution of two singular
distributions is absolutely continuous. No such condition is known at
present.
 7 ANALYTIC CHARACTERISTIC FUNCTIONS

We now introduce the class of analytic characteristic functions. This class

includes many characteristic functions which are important in probability
theory and in mathematical statistics. In the present chapter we consider
the general theory, and in Chapters 8 and 9 we study factorization problems.
In the following sections we will denote real variables by t and y and
a complex variable by z = t+zy. We introduce the following definition.
A characteristic function f(t) is said to be an analytic characteristic
function if there exists a function A(z) of the complex variable z which is
regular in the circle |z| < p (p > 0) and a constant A > 0 such that
A(t) = f(t) for |t| < A. We can express this in an informal manner by say-
ing that an analytic characteristic function is a characteristic function which
coincides with a regular analytic function in some neighbourhood of the
origin in the complex z-plane.
As examples of distributions with analytic characteristic functions we
mention: the Normal distribution, the Gamma distribution, and the
Poisson distribution. Stable distributions with exponent « < 2 provide
examples of characteristic functions which are not analytic.

7.1 The strip of regularity and the integral representation

From now on we assume that f(t) is an analytic characteristic function.
We know (corollary to theorem 2.3.2) that all moments of the correspond-
ing distribution exist and that it admits a Maclaurin expansion
co gk
(Ait) f(z) =>) ie for || < p (z complex)
k=0 Ft
where p > 0 is the radius of convergence of the series.
We write for the even part of f(z)
fo(2) = af(2)+F(—2)]
and for the odd part of f(z)
fi(2) = alf(2)-F(-2)]
Then the two series

(7.1.2)
192 CHARACTERISTIC FUNCTIONS

also converge in circles about the origin. We denote the radii of convergence
of these series by p, and p,.
From the inequality
x 2k-1 < $(0* + -)

we see that
Con—1
hile < Bora < sl“or 2k) +————— hex—2 |
|.
ch) (2k—1)! ~ (2k—-1)! "2 (ary! | )+ OR)!
Here and in the following we write again «,, and ,, for the algebraic and
absolute moments of order R.
We conclude from (7.1.3) that py > p» > p and also that the series
(ce) ok

Pa kl
converges for |z| < po.
Let & be a real number and denote the radius of convergence of the
Taylor series of fy (2) [respectively of f, (z)] around é by py(é) [respectively
pi(é)]. According to corollary 2 to theorem 2.3.1 we have
biicces (é)| Se Oak and hioa (é) | = ieee

so that
Pol(&) 2 po(0) = po > p and p,(é) > pi(0) = pi > p.
The Taylor series of f, (z) and of f, (z) around é therefore converge in circles
of radii at least equal to p. The same is therefore true for the expansion of
f(z) around &, so that f(z) is regular at least in the strip |Im (z)| < p.
We have already mentioned that the series
x Bre Ie

Bini?
converges for |y| < p. Clearly
205, ‘ 0 |v ‘i A
Pers
— lea S ianekdeeiyae _|# k PaF() = |e |ye| dF (x)
for any A and |y| < p. Therefore the integral

| elt! dF (x)
exists for |y | < p, hence the integral
| e” dF (x)
is convergent whenever |e*| < e”!, where z = t+zy. This means that
the integral
[" e* dF (x)
ANALYTIC CHARACTERISTIC FUNCTIONS 193
is convergent for any ¢ and |y| < p. This integral is a regular function in
its strip of convergence and agrees with f(z) for real z, therefore it must
agree with f(z) also for complex values z = t+1y, provided that |y| < p.

The integral | e“*dF (x) converges in a strip —« < Im(z) < +f

where « > p, 6 > p and is regular inside this strip. We write

fikib= | et dF (x)+| e dF (x) = L,(z)+ L,(2).
ioe) 0

0 109)

The functions #,(z) and ¥,(z) are Laplace integrals, convergent in the
half-planes y > —« and y < f respectively. Let z = iw; then w = —iz =
—it+y and L, (iw) -| e—“*dF (x) = O(w) is convergent for Re (w) =
0
y> —«a.
It is known that the Laplace transform g(s) = | e—“ dG (t) of a mono-
0
tonic function G(t) has a singularity at the real point of its axis of con-
vergence. For a proof we refer the reader to D. V. Widder (1946) (p. 58,
theorem 5b). Since F(x) is non-decreasing we can apply this result to
@(w) and we conclude that —« is a singular point of O(w). Thus —7« is a
singular point of f(z). In the same way it is also seen that 7f is a singular
point of f(z). We summarize these results and obtain the following theorem:
Theorem7.1.1. If a characteristic function f () ts regular in a neighbourhood
of the origin, then it is also regular in a horizontal strip and can be represented
in this strip by a Fourier integral. This strip is either the whole plane, or it has
one or two horizontal boundary lines. The purely imaginary points on the
boundary of the strip of regularity (if this strip 1s not the whole plane) are
singular points of f(2).
From theorem 7.1.1 we obtain immediately the following result:
Corollary to theorem 7.1.1. A necessary condition that a function, analytic
in some neighbourhood of the origin, be a characteristic function is that in
either half-plane the singularity nearest to the real axis be located on the
imaginary axis.
The corollary can sometimes be used to decide whether a given function
could be a characteristic function. We illustrate this by an example. Let

fit) = 3) (0-55) (-s) -

p= [e-t)
and

with
a > b > 0.
194 CHARACTERISTIC FUNCTIONS

It is easy to see that both these functions are analytic characteristic

functions. Their quotient
_ fi)
tS a
satisfies the elementary necessary conditions for characteristic functions,
f(—t) =f, |(| < f() = 1 for real t. However the condition of the
corollary to theorem 7.1.1 is violated since f(t) has no singularity on the
imaginary axis while it has a pair of conjugate complex poles +5—za.
Therefore f(t) cannot be a characteristic function.
Suppose that a distribution function /'(«) has an analytic characteristic
function f(z) whose strip of regularity is
—a < Im(z) < B.
It is then possible to express « and # in terms of F(x). Using classical
results [see Widder (1946) pp. 42 ff.] concerning the abscissa of convergence
of a Laplace integral, one obtains
log [1 eel) awe
a = —lim cup {
c—> CO x

Let f(z) be an analytic characteristic function; according to theorem

7.1.1 it can be represented as the Fourier integral

f@)= |" ear (x) [—a < Im(2) < fg].

Therefore

f(a)= Ff ae [xem ar(e)

We write zg= a+zy where a and y are real and where —« < y < f, then

| f"(at+iy)| < ee panied Iie).

li f= "2e (Re= 0S 12, eso) 1s ant cornea then this becomes

PF (a+iy)| < | steak(x)= |f@)

so that pe
max |2 (a+iy)| =| (6))
We have ieee eenes the following result:

Theorem 7.1.2. Let f(z) be an analytic characteristic function. Then |f(z) |

attains its maximum along any horizontal line contained in the interior of its
strip of regularity on the imaginary axis. The derivatives f™ (z) of even order
of f(z) have the same property.
ANALYTIC CHARACTERISTIC FUNCTIONS 195
The relation
|f(E+ay)| < f(y)
is very important in the theory of analytic characteristic functions. We say
that functions satisfying this inequality have the “ridge property” and we
refer to them as “ridge functions”.
Corollary to theorem 7.1.2. An analytic characteristic function has no zeros
on the segment of the imaginary axis inside its strip of regularity. The zeros
and the singular points of an analytic characteristic function are located
symmetrically with respect to the imaginary axis.
The first part of the corollary follows immediately from theorem 7.1.2;
we obtain the statement about the location of the zeros and of the singulari-
ties of the characteristic function f(z) if we observe that the functional
relation
(714) f= f(-2)
holds not only in the strip of convergence of the Fourier integral but in the
entire domain of regularity of f(z).
It is in general possible to continue an analytic characteristic function
beyond the strip of regularity. However, an analytic characteristic function
may have a natural boundary. I. V. Ostrovskii (1966) showed that the
boundary of any region which satisfies certain conditions can be the natural
boundary of an analytic characteristic function. These conditions are con-
sequences of the ridge property and of the relation (7.1.4) f(z) = f(—2)
which all analytic characteristic functions satisfy.
The strip of regularity of an analytic characteristic function f(z) can also
be the whole z-plane; in this case f(z) is an entire (integral) function. We
next derive a result concerning the order) of an entire function.
Let f(z) be an entire characteristic function. We denote by
M(rsf) = max |f(@)|
the maximum modulus of f(z) in the circle |z| < r. This value is assumed
on the perimeter of this circle and we see from theorem 7.1.2 that
M(r;f) = max [f(ir), f(—ir)]. We formulate this result in a slightly
different way.

Lemma 7.1.1. If f(z) ts an entire characteristic function, then

[f(ir)+ f(-m)] = Mf) > af )+f(-7)).
If f(z) is an entire characteristic function, then the integral representa-
tion is valid in the whole plane so that

Ck). 2 | cosh rx dF (x).

(*) See Appendix D.
196 CHARACTERISTIC FUNCTIONS

It follows from (7.1.5) and the lemma that

M(rsf) > | cosh rx dF (#) = dette”) | dF (x)
la]| 26 |a| >6

> al dF (x)
|x| >6

where 6 is an arbitrary positive constant. We write « = | dF (x) > 0

|x| >6

and note that « = 0 for every 6 > 0 if, and only if, F(«) = e(x) where
e(x) denotes the degenerate distribution. From the preceding inequality
we obtain
(7.1.6) M(r;f) > tue”.
Since the order p of an entire function f(z) is (see Appendix D)

(7.1.7) p=i:LOR log

eeelog M(r;
eaef)

we see from (7.1.6) that p > 1, provided that « > 0, and have proved the
following theorem:

Theorem 7.1.3. The order of an entire characteristic function which does

not reduce to a constant 1s at least equal to 1.
We will resume the study of entire characteristic functions in the next
section, but proceed now to discuss some convexity properties of analytic
characteristic functions. We introduced in Section 1.4 the moment
generating function

(7.1.8) M(y)= |p>ev dF (x).

This function is defined for distributions for which the integral (7.1.8)
exists for all |y| < R, where R is some positive constant. If f(z) is an
analytic characteristic function, then we can take R = min(a, #) and see
that its moment generating function exists and that M(y) = f(—iy). We
see from (7.1.8) that M(y) is real and positive while M'(y) > 0. Therefore
f(—1) is real and positive and is a convex function of y for all y for which
the representation by a Fourier integral is valid. Moreover M(y) = f(—zy)
is strictly convex unless f(z) = 1.
We next discuss a less trivial convexity property of analytic character-
istic functions and their derivatives of even order. Let f(z) be an analytic
characteristic function which has —« < Im(z) < £ as its strip of regu-
larity; according to theorem 7.1.2
(7.1.9) |fP (t+2y)| < [Ff ).
Suppose that for some yy such that —a < yy < B we have f™ (iy,) = 0.
It follows from (7.1.9) that f° (z) = 0, so that f(z), and therefore also
ANALYTIC CHARACTERISTIC FUNCTIONS 197
f(t) for real t, is a polynomial of degree 2k. Since f(t) is bounded, this is
only possible if k = 0 and f(t) = 1. We see therefore that the derivative
f° (x) of even order of an analytic characteristic function has no zeros
on the segment of the imaginary axis contained in its strip of regularity
unless f(z) = 1. Each point zy of this segment has a neighbourhood in
which f(*) (z) is regular and different from zero. In this neighbourhood
log f" (z) is defined and regular. We can therefore write f (2) (x) = e®,
where g;,(2) is regular in every point of the line segment —« < Im (z) < f.
Let A;,(t, y) and B,(t,y) be the real and imaginary parts respectively of
&.(2), so that
F (2) = exp [Ar(t,y) +7B,(4,y)]-
Then A,(#,y) = log |f(z)| while B,(é,y) = arg f(z). We can
therefore write (7.1.9) as A,(t,v) < A,(0,y) and see that for fixed
y (—« < y < ) the function A, (¢, y) has a maximum at t = 0 so that
0A, (t, 2) 0°A,,(t,
at gigvu 0, me edu at2
Since g;(z) is regular at the point z = zy we see from the Cauchy-—
Riemann equations and from Laplace’s equation that
OB, (t, | ea) 07A,(t, a
oy t=0 ‘ dy? t ll o
We therefore obtain the following result:

Theorem 7.1.4. Let f(z) be an analytic characteristic function which has

the strip —a« < Im(z) = y < f as its strip of regularity and let k be a
non-negative integer. Then arg f*"’ (iy) = B,(0,¥) is independent of y
while log |f'* (iy)|= A;,(0,y) is convex for ally such that —« < y < B.
Moreover, for all such y, the function f (iy) is real-valued, non-negative and
convex and ts even strictly convex, except in the case where f(z) = 1.

7.2 Analytic characteristic functions and their distribution

functions
In this section we deal with the relation between the properties of an
analytic characteristic function and its distribution function. As a first step
we will derive a necessary and sufficient condition which a distribution
function must satisfy in order that its characteristic function be an analytic
characteristic function.
In the preceding section we have seen that the moment generating
function (7.1.8) exists for every distribution function F’(«) which has an
analytic characteristic function. The converse statement follows from the
argument used in the proof of theorem 7.1.1. Thus distribution functions
F(«) which have analytic characteristic functions are exactly those for
198 CHARACTERISTIC FUNCTIONS

which the moment generating function exists. This remark leads to a rather
obvious criterion:
The distribution function F («) has an analytic characteristic function if,
and only if, the following two conditions are satisfied:
(i) the distribution F'(«) has moments «, of all orders k
(ii) lim sup [|o,|/R!]" = 1/p is finite.
ko

These conditions are equivalent to the statement that the series

CP ap
fle)
= & aye
foe)

k=0 k .

represents a function which is regular in the circle |z| < p(p > 0) and
which agrees with the characteristic function of F'(«) for real values of z.
We note that it is easy to construct distributions which have moments
of all orders but do not have analytic characteristic functions. Such a
distribution then necessarily violates condition (ii). As an example we
mention the distribution whose frequency function is
aie i exp (—|1/x|) for x« > 0
0 for x < 0.
The corresponding moments are easily computed. We find «, = (2k+1)!
so that
eae Pah oie pi 1 k+1

Hence all moments exist but do not satisfy (ii).

This distribution has another interesting property. Although its moment
generating function does not exist, it is completely determined by its
moments. It is known [see Shohat—T'amarkin (1943), theorem 1.11] that a
distribution function is completely determined by the sequence {a,} of its
moments, if the sum >) a, /°" diverges. Since a, = (2k+1)! < (2k+1)*
k=1
it is easily seen that this condition is satisfied.
Our next result is more useful and is directly applicable to the distri-
bution function F(x).
Theorem 7.2.1. The characteristic function f(t) of a distribution function
F(x) 1s an analytic characteristic function if, and only if, there exists a
positive constant R such that the relation
(7.2.1) 1—F(«x)+F(—x) = O(e-") asx —>o
holds for all positive r < R. The strip of regularity of f(z) then contains the
strip |Im (z)| < R.
Remark. ‘The positive constant R may be infinite. In this case (7.2.1)
holds for all positive real r and f(z) is an entire function.
ANALYTIC CHARACTERISTIC FUNCTIONS 199
We first note that (7.2.1) is equivalent to the simultaneous validity of
the relations
(i) 1—F (x) = O(e"") asx > oo
(i) RF (=x) = O(er™) eras ix 00,
We prove first the sufficiency of these conditions. Let y be a real number
such that |y| < Rand choose a positive r so that |y| < r < R. Letk bea
positive integer; then

| ele
dF(x)< elk —F(k—1)].
k

k-1

Using condition (1), we see that 1—F(k—1) = O(e-"*—) as k > 0, so

that there exists a constant C, such that 1—F(k—1) < C,e~™ for suffi-
ciently large k, say for k > K,, so that
k
| evedi'(x) = Cie al “fork > Ky.
k-1

We choose an integer K > K, and real numbers a > K,b > 0. Then
a+b o k
| elu dF (x) <) | ell" dF (x)
a k=K Jk-1
Se es CG eo tlw

ic esac
The last expression can be made arbitrarily small by choosing K suffi-
a+b
ciently large; therefore the integral | el"l® dF’ (x) can be made arbitrarily
small—no matter what 5 is—by choosing a sufficiently large. Thus the
integral
|jelle F(x) = |:elvel dF (x)
exists and is finite.
We also have
—k+1
| eluel dF (x) < elk F(—k+1),
—k
and we conclude from (ii) that there exists a constant C; such that
F(—k+1) < C,e~™ for sufficiently large k, say for k > K,. We choose
now an integer K’ > K, and two real numbers c > K’, d > 0. We see
then that
(ex eo tlw K’
|" ghtl dF (x) < =e oa
—c—a

and apply an argument, similar to the one used above, to show that the
integral | elv"ldF(«) exists and is finite. Combining this with our
0 . . . .

— 0
200 CHARACTERISTIC FUNCTIONS
(oe)

earlier result we see that ell dF'(x), and therefore also | e”” dF (x),
exists and is finite for all y such that |y| < R. Let z = t+7zy, then the
integral f(z) = | e dF (x) is also convergent for any ¢ and |y| < R
and represents a regular analytic function, so that the sufficiency of our
condition is established.
We next prove that the condition (7.2.1) [or (1) and (ii)] is necessary,
and suppose that the characteristic function f(z) = | e dF (x) is an
—o
analytic characteristic function whose strip of regularity is the strip
—a < Im(z) < f. Let R = min (a, f) and let x > 0 be a real number;
then the two integrals

Iso dF(u) and We e”” dF (u)

x

exist and are finite for all |y| < R. We choose a number 7 < R and let
r, be such that r < 7, < R. Then there exists a constant C such that

Gs i eo"
dF(u)> &*[1—F(x)] > 0
or
0 < [1—F(@)le* << Ce=™™,
Since 7, > 7, the expression on the right of the last inequality goes to zero
as x tends to infinity, so that 1— F(x) = O(e~™) as x > oo. In the same
way we see that (ii) is satisfied.
If F(x) is a finite distribution then 1—F(x)+F(—«x) = 0 for sufh-
ciently large x, so that (7.2.1) is satisfied for all positive r.Every finite
distribution has therefore an entire characteristic function. However, it is
possible to make more precise statements concerning finite distributions.
These are closely related to properties of one-sided distributions with
analytic characteristic functions which we shall discuss first.

Theorem7.2.2. Let F(x) be adistribution function with an analytic character-

istic function. F(x) 1s bounded to the left [respectively to the right] if, and
only tf, its characteristic function 1s regular in the upper [respectively lower]
half-plane and if there exists a finite positive constant c such that
f(z) |< &
provided that Im (z) > 0 [respectively Im (z) < 0].
We first show that our condition is necessary and we suppose that(*)

a = lext [F] and consider the integral lese“ dF (x) where z = t+iy. In
a

(*) The symbol lext is defined in Section 5.8, page 142,

ANALYTIC CHARACTERISTIC FUNCTIONS 201

the case where a > 0 it is clear that this integral is regular in the half-plane
y = Im(z) > 0. If a < 0 wewrite
ioe) 0 co
| ert dF (x) = | eit dF (x) atte | eit dF (x) :

a a 0

the first integral is an entire function while the second is regular in the
upper half-plane. We assume that f(z) is an analytic characteristic function.
Therefore f(z) is regular in a horizontal strip containing the real axis in its
interior and admits in this strip the integral representation

f(s) = | e** dF (x).

fo.@)

— oo

Since f(z) is regular in the upper half-plane we see that the region of
validity of this representation contains the half-plane Im (z) > 0 in its
interior. Therefore

Lf@)|=|["_ ear(a) < \)Oe NMED)

for y = Im(z) > 0 and we can show easily that

(4.2.2) f(z) |< e7™ < etl < ela,
provided that y > 0.
We next prove the sufficiency of the condition and assume that f(z) is
an analytic characteristic function which is regular in the upper half-plane
and which satisfies the inequality
If(z)| <e" (c > 0),
provided that Im (z) > 0. Let here z = zy(y > 0); then it can be shown
that
J
h = —l\im sup - log f(zy)
yoo y

is finite. Let ¢ be an arbitrary positive number and let x, and x, be two

real numbers such that x, < x, and h—x, = 2. We then see from the
representation of f(z) as a Fourier integral that

Ai |hee-"
dF(x) > li e-
dF (x) > e~" [F (xy)
@.
—F (x,)].
The definition of h implies that —h+e > y—1 log f(zy) or that
Ty) z enue) pam eo Uit2 Fe) |

Therefore e~" > F(x,)—F(«,) > 0 for arbitrary « > 0 and sufh-
ciently large y. This, however, is only possible if F'(«,)—F'(x,) = 0 when-
ever x, < x, = h—2e; this means that F(x) is bounded to the left and
that
lext [F] 2 A.
202 CHARACTERISTIC FUNCTIONS

This completes the proof concerning distributions which are bounded

from the left. However, we see from (7.2.2) that log f(ty)< —ay, hence
the inequality h > a = lext [F] also holds and we have hf = a or
1
(72-3) lext [F] = — lim sup ; log
f(zy).
yo

We can therefore state

Corollary to thereom 7.2.2. If F(x) is a distribution function which ts

bounded to the left and has an analytic characteristic function, then
1
(23) lext [F] = —lim Sls log f(zy);
yo co

if F(x) is bounded to the right and has an analytic characteristic function,

then

(7:24) rext [F'] = lim sup log f(—7y).

yo

The proof has been given only for distributions which are bounded to
the left; the proof of the statements concerning distributions which are
bounded from the right is quite similar and is therefore omitted.

Remark.. Let k(x) be an arbitrary convex function such that k(0) = 0;

it is then easily seen that k(x) /x is a non-decreasing function for x > 0. Let
f(z) be a characteristic function which is regular in the half-plane
—« < Im(z) =y («a > 0). According to theorem 7.1.4, the function
log f(iy) is a convex function of y for y > 0, and we conclude that
[log f(iy)]/y is a non-decreasing function of y. Therefore it is possible to
replace in the formula for the left extremity the lim sup by lim, and to
write instead of (7.2.3)

(7.2.3a) Text eta lim log f (zy).

yoo

Using a similar argument one obtains

ited
(7.24) rext (Jf| == lim ; log f(—7y).
yoo

These limits will of course be infinite if the distribution function F (x) is

unbounded either to the left or to the right.
If the distribution (x) is finite, then we can combine the results of
theorem 7.2.2 and its corollary and obtain the following statement:

Theorem7.2.3. Let F(x) be a non-degenerate and finite distribution function.

The characteristic function f(t) of F(x) is then an entire function of ex-
ponential type t > 0 and of order 1 which has infinitely many zeros. Con-
ANALYTIC CHARACTERISTIC FUNCTIONS 203
versely, an entire characteristic function of exponential type t > 0 and of order
1 belongs always to a finite distribution. Moreover the two extremities of
F (x) are given by (7.2.3) and (7.2.4).
Leta = lext [F'],6 = rext [F], c = max(|a|,|5|). We see then from
theorem 7.2.2 that |f(z) | < e#!, so that M(r; f) < e”. It follows from
theorem 7.1.3 that f(z) is an entire function of order 1 of exponential
type not exceeding c. According to Hadamard’s factorization theorem)
f(s) = G(z)e” where G(z) is the canonical product formed with the
zeros of f(z). Since |f(t)| < 1 for real t, we see that G(z) cannot be a
polynomial and must therefore have infinitely many zcros. The second
statement of theorem 7.2.3 as well as the formulae for the extremities
follow immediately from theorem 7.2.2.
Remark 1. A one-sided distribution does not necessarily have an analytic
characteristic function. As examples we mention the stable distributions
with exponent 0 < a < 1 and 6 = +1 which were treated in theorem
5.8.3. A specific example was given by formula (5.8.9).
Remark 2. A one-sided distribution may have an entire characteristic
function, and the order of this function can exceed 1. As an example we
mention the characteristic function which is obtained by truncating the
standardized normal distribution at the point zero.
Remark 3. An interesting modification of the continuity theorem for
characteristic functions of one-sided distributions was established by
A. Zygmund (1951). He showed that in the case of one-sided distributions
the condition that the sequence of characteristic functions should converge
over every finite interval can be replaced by convergence over a fixed
interval around ¢ = 0. For a precise statement as well as for the proof we
refer the reader to A. Zygmund’s paper.

Remark 4. The characteristic function of a finite distribution necessarily

has infinitely many zeros. These need not, however, be real [Example:
f(t) = (p+ae)].
Remark 5. Entire characteristic functions of order 1 and maximal type
do not belong to finite distributions.
Remark 6. If B(x) is a non-decreasing function of bounded variation
such that its spectrum Sz is contained in the finite interval [a, 6] then its
Fourier-Stieltjes transform
OV = |a e'* dB(x) = | e* dB(x)
(*) See Copson (1935), pp. 174-175.
204 CHARACTERISTIC FUNCTIONS

is an entire function of order 1 and exponential type not exceeding

max (|a |,||).
Remark 7. If a distribution function /(«) has an entire characteristic
function of order 1 and exponential type 7 then
erie text] < rextif wes.
A similar statement is true if /'(x) is a function of bounded variation whose
Fourier—Stieltjes transform is an entire function of order 1 and exponential
type T.
We next study the order and type of entire characteristic functions. For
this purpose we need three lemmas:

Lemma 7.2.1. Let « > 0 and k > 0; then the integral

I(z)= | exp (ize

— kx! **) dx
0
lo
is an entire function of order p = 1+1/« and type t = Rv at gate k

We expand the factor e”* into a power series, and since the order of
integration and summation may be exchanged we see that

AEDES. DD) Ga) | wo

eel oy
n=0 n!\ 0

We introduce the new variable y = kx't* in the integral and obtain

co

I(z).= 2 2
n=0
where
Satay fiSia | 1
eco (;*) (L+a)Rerare”
We use Stirling’s formula and the expressions (D4) and (D5) of Appendix
D and obtain the statement of the lemma.
The order and type of entire characteristic functions depends on the
“tail behaviour” of the corresponding distribution function. In order to
study this behaviour it is convenient to introduce three functions. Let F(x)
be a distribution function; we write for x > 0
T(x) = 1—F(«)+F(-x)
log [T'(x)]™
(W2Sye Te) Scares (a > 0)
ry = eben
We note that T(x) depends on the positive parameter «.
ANALYTIC CHARACTERISTIC FUNCTIONS 205
Lemma 7.2.2. Let F(«) be a distribution function and « > 0 and k > 0.
Suppose that there exists an x) > 0 such that
(a) exp (— katt?)
for x > Xo. Then F(x) has an entire characteristic function f(z) which is
either of order equal to 1+ «-1 and type t < a[k¥/*(1+.«)!+1/*]-! or of order
less than 1+a7},
Let A > x, andr > 0; we see then (integrating by parts) that

[-e°dF (x) = -|" ed [1—F(x)]

= e®[1—F(x,)]—e4 [1—F(A)]+r if [1 —F (x)] e”* dx.

We let A tend to infinity and conclude from the assumption of the lemma
that
|” ef dP (x) < e™[1—F(x,)]+r |i exp (rx —kx't*) dx.
Xo

Since lic e dF (x) < e” F(x), we have

(7.2.6) lia
we
e?dF(x) < entr |0 exp (rx — kx!)dee,
Similarly one can show that

(24) ie é7 di (x) ae r ie exp (rx —kx'**) dx.

Since M(r, 4‘s max [f(ir), f (—77)] _ see from (7.2.6) and (7.2.7) that
M(r, f) < e™4r |;exp (re Lshilt?) de,
The statement of lemma 7.2.2 follows easily from lemma 7.2.1 and from the
last inequality.
Lemma 7.2.3. Let F(x) be a distribution function with characteristic
function f (t) and let 2 > 0, u > O be two constants. Suppose that there exists
a constant R such that
Mir; f) < exp [4'***]
jorr 2 R. Then
lim inf T, (x) > 1+1/u
and >
T(x) < exp (—2't*"~*)
for any ¢ > 0 and sufficiently large x.
We see from (7.1.6) that forx > 0 andr >R
T(x) < 2 exp (—rx+4r'*),
206 CHARACTERISTIC FUNCTIONS
We put x > x) = 2R“ andr = ($x)! so that r > R; then we get
T(x)< 2 exp [—(2—2) (3x)'*"7]
and we conclude from formula (7.2.5) that
lim inf T(x) > 1+1/p.

For any e > 0 and sufficiently large x, one then has

1
2 (*) > 14+ 7-—e.
T.(x)

Using again (7.2.5), one obtains the statement of the lemma.

We study next entire characteristic functions of order greater than 1.

Theorem 7.2.4. The distribution function F(x) has an entire characteristic

function f (2) of order 1+%-1(« > 0) and of intermediate type t if, and only
if, the following two conditions are satisfied :
a Sie (at7*)s
(i) es Le) (ito

(ied (2) =50

foriallixe+0:
We first prove that the condition is necessary, and assume that f(z) is an
entire function of order 1+«~1 and finite type t > 0. Clearly (ii) is
necessary, since T(x) = 0 means that F(x) is finite, so that f(z) would
have order 1. Moreover it is possible to find for any ¢ > 0 avalue R = R(e)
such that
(7.2.8) M(r; fs) < exp [(t+6) yr rea

provided that r > R. It follows from (7.1.6) and (7.2.8) that

(7.2.9) I(x)-< 2exp [—rx+(r+2¢) r'**7]
forv
> 00 sa ok.
= R
Let a be an arbitrary positive number and let x > Me lI rs)3 Qa.
|B
Q Seas
beam

r= axe: thenr > R:ioefollows from (7.2.9) that

a yee —log 2+x***[a—(7+¢) alt@ z)

log.

so that
+ 0(1)
T(x) > a—(t+e)a't*"
as x — oo. Therefore
(7.2.10) lim inf T,(«) > a—(r+8)at=",
ANALYTIC CHARACTERISTIC FUNCTIONS 207
This relation holds for any a > 0 and in particular for that value of a
which maximizes the right-hand side of (7.2.10), that is for
a= {a/[(t+e)(1+«)]}*.
We substitute this value of a into (7.2.10) and get
~1)0
Hm. inf 7, (x) 2 cae :

Since e > 0 is arbitrary we see that

CA eae he
(7e2.0)) OP elinv ints 7 (2c)
(l+a)'**
We show next by means of an indirect proof that the inequality sign cannot
hold in (7.2.11). Suppose therefore that
ie? or)
aa wee T(x) ‘eae

(cx<);
Then it is possible to find a k > such that 7,(x) > k for x

sufficiently large. Using (7.2.5), we see that T(x) < exp (—kx'**) for suffi-
ciently large x and conclude from lemma 7.2.2 that
(Ez I2y" “MG: 7) s exp ((z’-+e)r'**]
for any e > 0 and sufficiently larger, where t’ = [k¥*(1+a)t’*]-1a < 7.
Since the order of f(z) is, by assumption, 1+«~1, it follows from (7.2.12)
that the type of f(z) is at most equal to t’, hence less than t. This contra-
dicts the assumptions of the theorem; therefore the inequality sign cannot
hold in (7.2.11), so that the necessity of (1) is established.
We still have to show that conditions (i) and (ii) are sufficient. Clearly
(ii) implies that f(z) is not a function of exponential type and that T) (x) is
defined for x > 0. Let
(72213 ees Riss (are t)*(Lo) oi.
In view of (i), there exists a value x, = x,(k) such that 7,(x) > k for
x > x,. It follows from (7.2.5) that
(7.2.14) Lie) i<eexp hx Bo)
for x > x,, and we conclude from lemma 7.2.2 that f(z) is an entire func-
tion whose order p and type 1’ are such that either
(7.2.15) p=i1+a-! and 7’ < a/[k¥*(1+a)'t4]
or
(7.2.16) p< 1l4+oa-4
We show next that (7.2.16) cannot hold. We give an indirect proof and
assume therefore tentatively the validity of (7.2.16). It is then possible to
208 CHARACTERISTIC FUNCTIONS

find a number y > « such that

M(r;f) < exp (r'*™)
for sufficiently large r, say r > R. It follows from this inequality and
lemma 7.2.3 that for any ¢ > 0
T(x) < exp (—2'""~*),
provided that x is taken sufficiently large. We again apply lemma 7.2.2 and
see that f(z) is an entire function whose order p cannot exceed 1 + (y—e)~1.
Since ¢ is arbitrary, we see that
ps 1l+y> < 1l+ac*.
But then
M(r;f) < exp(r'"””),
and we see from lemma 7.2.3 that
(7.2.17) liminf T,(x) > 1+y > 1+a.
We also see from (7.2.14) and (7.2.5) that
(7.2.18) Da) ae
Since « and 7 are finite and positive, (i) implies that ee all T, (x)is finite
and positive. Equation (7.2.18) indicates that this is only. possible if
(7.2.19) lim inf T,(*) = 1+«.

Relation (7.2.17), derived under the tentative assumption (7.2.16), con-

tradicts (7.2.19), so that necessarily (7.2.15) is valid and p = 1+a71, as
stated in the theorem. Since & is only subject to condition (7.2.13) but is
otherwise arbitrary, we deduce easily from (7.2.15) that
(7.2220) OWT
We show, again by an indirect proof, that the inequality sign in (7.2.20)
leads to a contradiction. Suppose therefore that t’ < +. Then there exists
at’ such that t’ < t” < t and
M(r; f) < exp (tr! **")
for sufficiently large 7. The last inequality has the same form as (7.2.8).
We use the reasoning which led from (7.2.8) to (7.2.11) and see that
Sood (7See ery
Si dE; (x) 2 (1 Taira (L+a)it=’

This contradicts assumption (i) of the theorem, so that t’ = t and the

proof is completed.
It is also possible to derive conditions which assure that a distribution
function has an entire characteristic function of a given order greater than
1 and of intermediate but unspecified type, or of maximal or minimal type.
ANALYTIC CHARACTERISTIC FUNCTIONS 209
Theorem 7.2.5. The distribution function F(x) has an entire characteristic
function f(z) of order 1+a-1 (a > 0) if, and only if, the following two con-
ditions are satisfied:
(i) lim inf T, (x) = 1+«
(ii) T(x) >0 forallx > 0.
In view of theorem 7.2.4 it is clear that the conditions are necessary.
To prove that they are sufficient, we note that (i) implies that T(x) <
exp (—x't*~*) for any ¢ > 0 and sufficiently large x. Using the argument
which we employed in the proof of theorem 7.2.4, we can show that f(z)
has order 1+a7-1.

Theorem 7.2.6. The distribution function F(x) has an entire characteristic

function of order 1+a-1 (« > 0) and of minimal type if, and only if, the
following three conditions are satisfied:
(i) lim inf T(x) = 1+

(eT (x) 0 for'allx > 0

(iii) lim T, (x) exists and lim T,(x) = + 00.

Theorem 7.2.7. The distribution function F(x) has an entire characteristic

function of order 1+a-1 (a > 0) and of maximal type if, and only if, the
following three conditions are satisfied:
(i) lim inf T,(*) = 1+0
(ii) T(x) > 0 forallx >0
(iii) lim inf 7, (x) = 0.

It is also possible to obtain results concerning distributions whose

characteristic functions are entire functions of order 1. The method of proof
is similar to that used in proving theorem 7.2.4. We therefore list here only
the relevant results.
Theorem 7.2.8. The distribution function F(x) has an entire characteristic
function of order 1 and maximal type tf, and only 7f,
(i) T(x) > 0 forallx > 0
(ii) lim T(x) exists and lim T,(x) = + 0.

Theorem 7.2.9. An entire function of order 1 and minimal type cannot be a

characteristic function.
The last theorem is only a reformulation of a result from the theory of
entire functions which asserts that a non-constant entire function of at
210 CHARACTERISTIC FUNCTIONS

most first order and minimal type cannot be bounded on some line [see
B. Ya. Levin (1964), p. 51]. For a detailed proof of theorems 7.2.5 to
7.2.9 we refer to B. Ramachandran (1962).
Order and type of entire functions provide means of studying their
growth. This study can be refined by introducing proximate orders and
types with respect to proximate orders [see e.g. Levin (1964), pp. 31 ff.].
It is also possible to investigate the behaviour of characteristic functions
having given proximate orders. For these studies we refer the reader to
H. J. Rossberg (1966), (1967a), (1967b).
We finally remark that there exist entire characteristic functions of
infinite order. Let f(t) be an arbitrary characteristic function; it follows
from lemma 5.4.1 that exp [f(¢)—1] is also a characteristic function. We
define the sequence of functions
(7.2.21) Vee a= F(t)

Fm(t) = exp [fin-y()—-1] (# = 2,3,.-.)

and see, again using lemma 5.4.1, that all functions of the sequence f,,) (t)
are characteristic functions. Suppose now that f(¢) is an entire characteris-
tic function of finite order, then the functions /(,(¢) are entire functions of
infinite order ifn > 1. Asan example we mention the sequence of functions
which starts with f(,)(£) = e’; this yields a sequence of entire characteristic
functions of infinite order of more and more rapid growth. The second
function in this sequence is the characteristic function of the Poisson
distribution.

7.3 Criteria for analytic characteristic functions

In Chapter 4 we discussed various criteria for characteristic functions.
We have seen that the necessary and sufficient conditions developed by
Bochner, Cramér and Khinchine are not easily applicable. It is therefore
desirable to derive less general results which are applied more readily.
These results are usually restricted to certain classes of functions; in this
connection the problem arises whether it is possible to characterize those
functions which are regular in a (complex) neighbourhood of the origin
and are characteristic functions. This problem is still unsolved, but a
number of results, giving sometimes necessary and sometimes sufficient
conditions for analytic functions to be characteristic functions, were found.
The present section deals with these criteria. In some instances we only
mention conditions and give appropriate references, but a very important
criterion for a class of entire characteristic functions will be studied in
detail.
We note first that some of the results treated in Section 7.1 can be re-
garded as criteria for analytic characteristic functions. Thus theorems
7.1.1, 7.1.2 and their corollaries give necessary conditions which a function
ANALYTIC CHARACTERISTIC FUNCTIONS 211

regular in a neighbourhood of the origin must satisfy in order to be a

characteristic function. The same is true of theorem 7.1.4 or of P. Lévy’s
result (theorem 7.1.3) that a non-constant entire function of finite order
must have at least order 1 and must have infinitely many zeros if its order is
equal to 1.
It is easy to establish a condition similar to the one listed in theorem
4.1.2. Let f(z) be an analytic characteristic function; then

Re [f (ty) — f(t+iy)] = ies e~”(1 —cos tx)

dF (x)

24 \- ee sin ixal (x) = 4 \ é, (1 — cos 21x)

di (x).
Therefore
Re [f(ty)— f(t+ty)] > dRe [f(y)
— f(2t+)].
We iterate this procedure and obtain the following result:

Theorem 7.3.1. Let n be a non-negative integer; the inequality

Re [f(vy)— f(¢+zy)] > 4™ Re [f(ty)
— f(2"t+9)]
is then satisfied for every analytic characteristic function f(z), provided that
the point z = ty 1s in the interior of the strip of regularity of f(z).
The ridge property gives an upper bound for the values of an analytic
characteristic function along a line which is parallel to the real axis and is
located in the interior of the strip of regularity. It is also possible to derive
a lower bound for the values of an analytic characteristic function along
the imaginary axis in the strip.
Let f(z) be an analytic characteristic function which has the strip
—a«a < Im(z) < f as its strip of regularity (« > 0, 8 > 0, zs = t+1y).
Since

t= i(2) = |ier[1 —e-”/2] dF (x)

we conclude from Schwarz’s inequality that

ten)" < (ise[1 —2e-™/2

+e~'] dF (x)
or

(2) ]< #@) (-2 <<)

We apply this inequality repeatedly and see that for any positive integer
nN,

(-e<y<).
.

(73.1) fly) > ep

212 CHARACTERISTIC FUNCTIONS

It follows from the corollary to theorem 7.1.2 that log f(z) is defined in
a region which contains the segment —a« < Imz < f. It is easily seen
that in this region

log f(z) = > = (ey,

j=1 J !
where x; is the cumulant of order j. We then have
fe wy eo (= 1) Kj :

We combine the last equation with the inequality (7.3.1), and letting m tend
to infinity we see that
log f(y) > —ye
or

F (ty) > exp (—yx1)-

We therefore obtain the following result:

Theorem 7.3.2. Let f(z) be an analytic characteristic function which has the
strip —« <Im(z)=y <8 as its strip of regularity. Then f(ty) >
exp (—yx,), provided that —« < y < B. Here x, = 1-1f'(0) is the cumulant
of order 1 (first moment) of the distribution corresponding to f(z).
Theorem 7.3.2 is also a necessary condition which an analytic characteristic
function must satisfy. It will be used in Chapter 8.
We next discuss a sufficient condition which is applicable to certain
analytic functions.
Let 6 be a real number; the function [1—(it/6)]-1 is always a charac-
teristic function (it belongs either to an exponential distribution or to the
conjugate of such a distribution). Since the product of two characteristic
functions is always a characteristic function, we see that the reciprocal of a
polynomial which has only purely imaginary roots is always a characteristic
function. It follows from the continuity theorem that the reciprocals of
canonical products) of genus zero or 1 which have only purely imaginary
roots are characteristic functions. Necessary conditions for rational
functions were given by E. Lukacs—O. Szasz (1954a) and for reciprocals of
polynomials by K. Takano (1951); the last paper also contains a necessary
and sufficient condition for the reciprocals of polynomials of degree not
exceeding 3. Sufficient conditions for a special class of rational functions
are given by E. Lukacs—O. Szdsz (1954b) and by A. Zemanian (1959),
oGd):
The most important result concerning criteria for analytic characteristic
functions refers to a class of entire functions which we now introduce.
(®) See Titchmarsh (1939), p. 250.
ANALYTIC CHARACTERISTIC FUNCTIONS 213
It is convenient to adopt the following notation for iterated exponential
functions:
€,(3) = &, e,(z) = e*, ..., &,(%) = exp [e,-1(2)].
Our object is to derive the following theorem:

Theorem 7.3.3. Let

P(t) = Hyae

be a polynomial of degree m > 2 and denote by

Fa(t) = Knen[P(0)]
where K,,' = e,(€o). Then f, (t) cannot be a characteristic function.
The constant K,, is determined by the fact that every characteristic
function equals 1 for t = 0.
A particular case of theorem 7.3.3 is of great interest and is quite often
useful. If we put 2 = 1 in theorem 7.3.3, then we obtain the following
corollary:

Corollary to theorem 7.3.3 (Theorem of Marcinkiewicz). Let P,(t) be a

polynomial of degree m > 2 and denote by f(t) = exp [P,,(t)]. Then f(t)
cannot be a characteristic function.
The corollary to theorem 7.3.3 was first given by J. Marcinkiewicz
(1938) and is therefore frequently called the Theorem of Marcinkiewicz.
It is often useful and has been applied by many authors in studies con-
cerning the characterization of the normal distribution. Marcinkiewicz
derived his result in a different manner; he obtained it as a special case of
a more general theorem (see theorem 7.3.4) which will be discussed later.
We introduce first the following notation which will be used throughout
the proof of theorem 7.3.3. Let

$,(2) = yyCent

denote a polynomial of degree m without constant term and with c,, 4 0.

The coefficients ¢,, Cy,...,€m are arbitrary, real or complex, numbers.
Define the real functions «,(t, y) and /,(t, y) as the real and imaginary
parts, respectively, of ¢,(z) so that ¢,(z) = «,(¢,¥)+7/:(¢,y) where
2 = t+ty. Moreover define A, (t, y) = a(t, y)—% (0, y).
We write c, = «, +78, (v = 1,2,..., m; «,, 8, real) and obtain for the
polynomial 4, (z) the expression

(73.2) dle) = ¥ (ti e+Hy’

214 CHARACTERISTIC FUNCTIONS

We expand the powers of (¢+zy) according to the binomial theorem and get

evr =ryt SO)

Se) 5 3 (eee) J
and

@rayrt = (ty {2 ys(ESE)

HE Co Ge)
We note that the expressions in the braces are functions of t/y and see that
they contain certain polynomials in ¢?/y?. It is convenient to introduce
formally these polynomials. We denote by

(7.323) Vig)= & (3) é)'-1 for s 2 2 and V,(é) = 1

734 W.(é {(s-+1)/2] Ss oe [(s—1)/2] 5

ewe s => — a 4 k

( ) ©) = (a1) *) = (2x1) (S)

fors 22h:
The symbol [x] denotes here, as usual, the greatest integer contained in x.
In this way we obtain

G35)
coor care fli-tir(Jen (2)
(t-+2y)"-*
v2 20 y? y 20 4?

=oage nee f ae
We (S) +i 1-4ie rae (S)]
i?

We write in the following

(7.3.6) Yoo = an! You-1 = Bev-1» Say = Boys Ooy—1 = Sgy—4

and obtain from (7.3.5) and (7.3.6) by means of elementary computations

C327) («,+72B,) (¢+72y)s
2
= 4(—1)*t2),,. i=,
=a 6 Fs
=| 1 )s/21 pee W. ea); 8
{( ) ue y? y? (4) y s y? y

(rp. y Sw, (Z) 4¢—nyms, [1-4gPiy v4)»

t t2 t?
ae y* fie anys 4
10.5sel Z3 oes Seer
ANALYTIC CHARACTERISTIC FUNCTIONS 215
‘The last forrnula permits the computation of «, (t, y) and of f(t, y); one
obtains immediately

(7.3.8) o%4(t,y)= é {
\(- 1)s+0/21,n[1-— V, (4)
ae wt
Noa 3 t7\)
+(—1)” ; W, (4)
(= 1) 6,— whl s
and

739) falty) = 3 (ayy, £w, (4)

s=1 . y yy?

: 2 (/t
+(- 1)#/ 16, [~ y2 V, (4) |}

Since A, (t, y) = 4,(t, y)—%,(0, y) we obtain from (7.3.8)

(7.3.10)
2 2
A, (t,y)= 3 {(-1
Date es ae v,(5)+(-1)9,2 W, (4)

We ee a new variable = (t*/y*) (£ > 0) and write

(7.3.11) fA,(é) = (— 1/7, 6V, (6) + (— 1)" 5,8"? W,(6)

LB,(6) = (— 1) y, 6? W,(6) + (— 1/6,
[1—EV, (1;
with this notation(+) we have

AlyVéy¥] = 5 A,()y"
(7.3.12) ee
Alyvé x1 = & B,(E)9".
For the proof of theorem 7.3.3 we need several lemmas and formulate
next the following two statements concerning the coefficients of the highest
power of y in the polynomials A, [yVé, y] and pf, [yVé, y].

Lemma 7.3.1. Let m = 4; then it is possible to find a real number E,, > 0
such that Am (Em) > 0 while Bu(&m) # 0.

Lemma 7.3.2. Let m = 3 and y, # 0; then there exists a E,, > 0 such that
y3A,(&) < 0 and B,(&,) # 0. If m = 3 and y, = 0 then there exists a
E, > 0 such that A,(&,) > 0 and B,(£;) # 0.
In order to prove these statements we study the polynomials V,(é) and
W,(£) and show that they can be expressed in terms of Chebyshev poly-
nomials or trigonometric functions of an auxiliary variable.
(+) €1/2 is here and in the following the positive square root of &.
H
216 CHARACTERISTIC FUNCTIONS
We consider the expression (1+74/€)’, where s is a positive integer and
£ > 0, and set ¢ = arc tan v/é with |¢| < 2/2. Then
(7.3:13) (1+72/é)* = (1+ 4)? (cos sf +7 sin sd).
For s > 2 we expand (1+71/é)* according to the binomial theorem and
obtain
(7.3.14) (14 iv/é)* = 1-EV,(8) +iVEW, (6).
We note that (1+ &)”? = (cos ¢)-* and get from (7.3.13) and (7.3.14)

1—£V,(£) = (cos s$)/(cos )* = (1+) T, ( )

L
(7 - ou 5)

VEW,(E) = (sin sd)/(cos $)° = (146)? Uy (Gaze)

where T,(x) = cos (s arc cos x) and
sin (s arc cos x)
U,.,(#) =
V(1—~?)
are the Chebyshev polynomials of the first and of the second kind re-
spectively.
We introduce for the sake of brevity the notation
y=(- 1)m— 2/21 Yin
(7.3.16) ‘ — (-1)14,,
and express the functions A,,(&) and B,,(&) in terms of the variable ¢. If
we write
€ = tan’?¢
(73:17) {ew= A,, (tan? ¢)
D(?) = Bn (tan?9),
then we get from (7.3.11), (7.3.15) and (7.3.16)
73.18 Orie [ 4 sin mh
(73.18) C() =7 [1-40
sin mp cos baa
noe md
Tae 9 D(¢) =
( ) ie. (cos yet (cos p)”
We next prove lemma 7.3.1 by showing that it is always possible to find
a d, such that C(¢,) > 0 while D(¢,) 4 0. We give the following rules for
the selection of do:
(I) If y > 0 and 6 > 0, select ¢) so that 2/2m < dy < x/m
while tan md, 4 — 6/y.
(II) If y > 0 and 6 < 0, select dy so that x/m < do < 52/4m
while tan m¢) # —6/y.
ANALYTIC CHARACTERISTIC FUNCTIONS iy.

(III) If y= 0 and 6 > 0, select ¢y so that 2/2m < $y < x/m.

(IV) If y = 0 and 6 < 0, select dy so that x/m < dy < 52/4m.
(V) Ify < 0, select a value 4) which satisfies the following three
conditions:
It 2a
(a) m < Po < ee

(b) tan md, # —d/y,

(c) h(do) = y(cos™ dy—cos mo) +6 sin mhy > 0.
We must still show that it is possible to select do in case (V) so that
condition (c) is satisfied. We first observe that h(¢) = y(cos” ¢—cos m¢)
+6 sin md is a continuous function and that h(22/m) > 0. Hence the
function h(¢) is positive in some neighbourhood of ¢ = 2z/m, so that a
selection in accordance with (c) is possible.
The assumption c,, # 0 implies that y and 6 cannot vanish simul-
taneously, so that our selection rule covers all possibilities. Using this fact
as well as the assumption m > 4, it is easily seen that the value 4) whose
selection we have just described satisfies the conditions C(¢,) > 0 and
D(¢o) # 0. But then it follows from (7.3.17) that €,, = tan? ¢g satisfies the
assertion of lemma 7.3.1.
We next prove lemma 7.3.2. We see from (7.3.11) that
A;(&') = —3y36'—d3V/(E')(3—€)
B3(E’) = —ysV(E')(3 —) —63(1 — 3é’).
If vy; # 0 and 363 < 0 we choose & > 3, and if y,6, > 0 we choose
Or. 3, li vy, = Oand 0, = O}we select €,%> 3;:11.6, < 0 we select
0 < & < 3. Obviously it is possible to select &, in agreement with this rule
so that B,(&) 4 0. This completes the proof of lemma 7.3.2.
In the following we assume that m > 3 and choose é,, and &, in accord-
ance with lemma 7.3.1 and 7.3.2 respectively. We write
(7.3.20) A, = A, (En)> B, = B, (Em)
and obtain from (7.3.12)
m—1

(7.3.21) AilyV(Em) 9] = Am y+ % Ay y”.

Now let m > 4; then
A, [yV(Em)) 9] = Amy [1+o0(1)] as y > 0.
We see from lemma 7.3.1 that A, > 0, so that A,[yV/(ém), ¥] is positive
for sufficiently large positive values of +.
We consider next the case m = 3 and write e = sgn y = y/|y|. We
choose « so that ey, < 0. Then
AilyV(§3), 9] = Ase| y |?+429?
+Are| | = Ase y [1 +0(1)]
as | y | > 00.
218 CHARACTERISTIC FUNCTIONS

We know from lemma 7.3.2 that eA; > 0, so that A,[yV<&, y] becomes
positive if the sign of y is opposite to the sign of ys and if |y |is sufficiently
large. We summarize our findings in the following statement:

Lemma 7.3.3A. Let m > 3 and suppose that one or the other of the following
two conditions 1s satisfied :
(i) m>30rm=3 and y, = 8, = 0
(ii) m=3 and ys = Bp, 4 0.
Then there exists a Em > 0 and an A > 0 such that

Ai [yVEm ¥] = Aly |" [1+ 0(1)]

where the estimate holdsin case
(i)as y > 00, but incase (ii)
as(—sgnys3)y > ©.
Then there exists also a value Y = Y(m) such that
A, [yV Em Beall
provided that in case (i) y > Y while in case (ii) one must require
(—sgn ys)y 2 Y.
The chief instrument in the proof of theorem 7.3.3 is the following
lemma:

Lemma 7.3.3. Let 0 be an arbitrary real number. If m > 3, then it is

possible to find real numbers E,, > 0 and y* such that for t* = y*VE,, and
some integer g, the relations
A, (t*, y*) > 0, 62 (t*,9*)—2¢,4 = @
are satisfied.
To prove lemma 7.3.3 we must study the function f(t, y). We choose
again €,, in accordance with lemmas 7.3.1 and 7.3.2 respectively and con-
sider the polynomial

(7.3.22) Bly) = PilyVEm 9] = % By 9"

Here B, is given by (7.3.20). Let Y be the number determined by lemma
7.3.3A; since B, # 0 we conclude from (7.3.22) that
B(y) = Bn y™[1+o(1)] as |y| > oo.
This means that B(y) is monotone if y is sufficiently large. We can there-
fore find a Yy > Y such that B(y) is monotone for |y| > Yo. In view of
lemma 7.3.3A it is always possible to find a real yg such that |yo | > Yo and
A,[yov) Em Vol > 0. Let 6 be an arbitrary real number, then there exists
an integer g such that
O+2ng < Bly) < 94+2a(g+1).
ANALYTIC CHARACTERISTIC FUNCTIONS 219
We consider from now on only such values of y for which yy, > 0 and
|y| > |yo|. For such values of y, B(y) is either monotone increasing or
monotone decreasing. In the first case we can find a real number y, such
that B(y,) = 6+2(g+ 1); in the second case there exists a y, for which
B(y1) = 9+ 2gx. Since |y,| > |yo| and y, yo > 0, we see from lemma
7.3.3A that A, [y,V En, 1] > 0, while B(y,)—0 = By [91V Em ¥1]—9 is an
integer multiple of 27. 'To complete the proof of lemma 7.3.3 we need only
put y* = y, and ¢* = y, Vé,,. We are now ready to prove theorem 7.3.1.
Let
P(t)= Set
v=0

be a polynomial of degree m > 2 (cm 4 0) and K,! = e, (co). We carry an

indirect proof of the theorem and suppose therefore that
fn(t) = Kyen[PO]
is a characteristic function. The function f,, (¢) agrees for real values of z
with the function A(z) = K,,e, [P(z)] so that it is an entire characteristic
function. From now on we consider this characteristic function also for
complex values of the argument z = ¢t+7y and apply the ridge property
(theorem 7.1.2) of analytic characteristic functions. This theorem indicates
that necessarily
fr(t+ty)
CIR Ee ae are
oe Ui = |_| < l

for all real ¢ and y.

We now introduce the functions
(7.3.24) als) Keer EG) ee = 1) 20)
where K, "= fe, (c,)|—! and note that 7,(0) = 1(v = 1, 2,..., ) and that
f(z) = exp [¢1()]. We obtain easily from definition (7.3.24) of the
functions f,(z) the recursion formula
(7.3.25) f(a) = exp {Kz4[f-(@)-11} (@ =2,-- +5).
We now introduce the functions
(S252) Rd (cl) fe) — 1] 9(o— 253, oe, 2).
and write «,(t, y) for the real part, f,(t¢, y) for the imaginary part of
¢, (2), so that
(7.0.29b) __-¢,(2) = a, (, vy)+78,(2, y) (v = 1,2,...,7)
and
(73-256) is(s) = .exp [o, (2)1 a =-1, 2.0... 57).
We set
(7.3.26) Ke exp (p11) \(@ = 72, is 7):
220 CHARACTERISTIC FUNCTIONS
Since K>1, = €,2;(¢o) we see’that K>4'= exp (K,2,) or
exp (p,+7/,) = €3(p»_1+tA,_1)s
therefore
(7.3.27) — p tia, = exp (p,_1+2/,_1) + 22,70
where g, is an integer. It follows from (7.3.27) that
(7.3.28) o = OXp (p,-3) SiN A544 22,7.
We combine (7.3.25a), (7.3.25c) and (7.3.26) to get
bo (2) = exp [po1+tya] {exp [a(t ¥) + Boa ay—15
U =i Le3,. sah):
We separate real and imaginary parts in the last formula and obtain re-
cursion formulae for «,(¢, y) and 6, (t, y):
(7.3.29a) a» (t, y) = exp [py1+%y_ (4, y)] cos [A,-1
+ Boa (t, ¥)] —
= EXD (p24). COS Apiy, (0 25 Oye 10 50?)
(7.3.29b) Bp(t, v) = exp [ps4 +oy-1 (4, y)] sin [Ava +P, Y)| —
—€XP (fy-1) SIN Ap (OC = 25 0, eee Bs
We now define the functions
A, (i;V) = 4,(¢, y)—o,(0, y), (& = 1, 2,. +. , 2)
and see from (7.3.29a) that
(7:3:30) +» A,(t, y) = {exp [4,4 9))cos Dea pa y))=
— cos [Ay_1 + By-1(0, y)]} exp [Pv-1 + %-1 (0; ¥)]
(U2. 25 ee
We apply lemma 7.3.3 and select 0 = —A,. Then it is possible to find a pair
of real numbers ¢*, y* such that
(7.3.31) A(t vy" oe, 0
while
(73:32) Br (t*, y*) +4, = 22,2 (g, integer).
We show next that a similar relation holds for all functions ,(t, y,)
namely
(7.3.33) 8, (t*,.y*) Ap = 2¢,7
where t*, y* are the numbers determined according to lemma 7.3.3 and
used in (7.3.31) and (7.3.32); A, is given by (7.3.28) and g, is an integer.
We prove (7.3.33) by induction. Formula (7.3.32) indicates that (7.3.33)
is valid for v = 1; we suppose now that it holds for all subscripts inferior
to v. We then have in particular
Bea (i, y*) +Ay 4 = OL Saeee TG.

Substituting this into (7.3.29b) and using (7.3.28) we see that

By (t*, y*) = —exp (p,_1) sinA,_. = —A,+2,2.
Thus (7.3.33) is generally valid.
ANALYTIC CHARACTERISTIC FUNCTIONS 221
We see from (7.3.30) and (7.3.33) that
A, (t*, y*) = {exp [4,1 (¢*, y*)]—cos [4,1 + Bos (0, »*)]}
x exp [py_1+%1(0, y*)]
for v = 2,3,...,m. From this formula we see that the relation
A, (¢*, y*) > 0
implies that
A, (ty) s>2 0.
We can therefore conclude from (7.3.31) that
(7.3.34) WA rcs ag pl 4
We defined earlier the function R(t, y) as

Ri,Mty)= In(t+y)
|S,
Lira)
it follows from (7.3.25b) and (7.3.25c) that
Rit, y) = exp {A, (t, y)}-
We have therefore determined a pair of real numbers ¢*, y* such that
(753.3) Pte’ oe eal
But this contradicts (7.3.23) which must be satisfied if f,,(t) is a charac-
teristic function. This contradiction completes the proof of theorem 7.3.3
since it shows that f,, (¢) cannot be a characteristic function if m > 2.
In the case where m < 2 the iterated exponentials f,(¢) = e,[P(t)] can
be characteristic functions. The function f, (t) = exp [—a,t?+7a,t] where
a, and a, are both real, a, > 0, is a characteristic function (of a normal
or of a degenerate distribution). It follows from the recursion formula
(7.3.25) and from lemma 5.4.1 that f,(z) as defined by (7.3.24) is a charac-
teristic function for all values of v.
We note that for m = 1 and c, = 1 we obtain for nm = 2 a Poisson
distribution and for n = 3 a Neyman type A distribution,
Fa(t) = exp {u[exp (A(e"—1))— 1]}.
We have already mentioned that Marcinkiewicz derived a particular case
of theorem 7.3.3 in a different manner. He obtained it as a special case of
a more general theorem which gives a necessary condition which an entire
function of finite order must satisfy if it is a characteristic function. We
now state this theorem of Marcinkiewicz.

Theorem 7.3.4. An entire function of finite order p > 2 whose exponent of

convergence p, ts less than p cannot be a characteristic function.
In the proof of theorem 7.3.4 we use a number of theorems from the
theory of functions of a complex variable. The results needed may be
found, for instance, in Copson (1935), pp. 165-175.
222) CHARACTERISTIC FUNCTIONS

Let f(z) be an entire function of finite order p. By Hadamard’s factoriza-

tion theorem we can write f(z) in the form
(7.3.36) Jie) = Geiexo |H@)]
where G(z) is the canonical product of the zeros of f(z) and where H(z)
is a polynomial of degree m < p. We denote by p, the exponent of con-
vergence of the zeros of f(z); it is easily seen that p = max (p,, m). If
p: < p then necessarily p = m. It is known that the order of a canonical
product equals its exponent of convergence. Let G(z) be a canonical
product of order p,; then for any e > 0 the modulus |G(z)| < exp[]z|"*‘],
provided that |z| is sufficiently large. We will also use the following
result which is due to E. Borel.
If G(z) is a canonical product of order p, and if ¢ is an arbitrary positive
number, then there exists an infinite number of circles of arbitrarily large
radius on each of which the inequality
|G(z)| > exp(—|z eri)

holds.
Let z = i+iy and denote by r = |z2| = Vtt+y?. We see then that
there exist arbitrarily large values of r such that
|G(t+iy)| > exp (—r"**).
On the other hand we know that for arbitrary « > 0 and sufficiently large y
|G) | < exp (7****).
We combine the last two inequalities and see that there exists an increasing
sequence {r7,} of positive real numbers such that
lim: (7; = 7eo
k—>o

which has the property that for arbitrary « > 0 and sufficiently large k
G(t+1y)
13-07 Ise (5g) > expla)
provided that t?+ y? = 72.
We consider next f,(z) = exp [H(s)] and write
(7.3.38) Ra(t,y) = jexp [H(t+iy)—H |
so that

(73.39) R(ty) = HS) = Riley) Balt

| 1

We give an indirect proof for theorem 7.3.4 and assume therefore that
f(%) is an entire characteristic function of order p > 2 and suppose that
the exponent of convergence p, of the zeros of f(z) is less than p, p; < p.
We again apply theorem 7.1.2 and see that necessarily
(340). shy aad
for all real ¢ and y.
ANALYTIC CHARACTERISTIC FUNCTIONS HES
Since p; < p we have p = m, where m is the degree of the polynomial
H(z). We see also that H(0) = 0 (since f(0) = 1) and use the notation of
the preceding proof and write

i(z) = H(z) = & (a, +78,)2°

so that ais
(7.3.41) —Ry(t, y) = exp [A,(¢, 9).
We see then from (7.3.37), (7.3.39) and (7.3.41) that there exists an infinite
sequence of indefinitely increasing positive numbers r; such that for an
arbitrary e > 0
(7.3.42) R(t, y) > exp [—2r27*+
A, (t, y)],
provided that k is sufficiently large and that t2+? = rz.
We next define an infinite sequence of points (¢;, y;,) in the z-plane. In
order to be able to apply lemma 7.3.3A we subject these points to the
following restrictions:
(i) te = VeWV Em
(ii) |t.t+ivn| = re
(iti) ifm > 3 or m = 3 while y, = 8; = 0, theny, > 0
(iv) if m = 3 and y, = f; # 0, then (—sgn y3)4, > 0.
From (i) and (ii) it is seen that all these points are located in the same
quadrant and that |y,| = r./V/(1+émn). We deduce from lemma 7.3.3A
that
(7.3.43) Ax(ty yx) = Aly, (™[1+0(1)] ask > co
where A > 0. ?
We denote by 2% = A(1+é,,)-”? and obtain from (7.3.42) and (7.3.43)
R(t Yu) > exp {—2rh7°+ Ur? [1+0(1)]} ask > o.
Since by assumption p = m > p,, we can choose the arbitrary positive
quantity « so that py +e < m. We conclude from the last inequality that
R( try Vx) > exp {U7 [1+o0(1)]} ask > oo.
Since 2 > 0 we can determine k so large that R(t,, y;,) > 1.'This, however,
contradicts (7.3.40) and we see therefore that f(z) cannot be a charac-
teristic function and have thus completed the proof of theorem 7.3.4.
In conclusion we mention, without proof, another theorem of this type.
Theorem 7.3.5. Let
Nefel (NE Daa!
v=0

be a polynomial of degree m. The function

f(t) = exp [A, (4-1) +A, (e~*—-1) + Pn (#)]
is a characteristic function if, and only if, 2, > 0, A, > 0, m < 2 and if
P,(t) = a,(it)—a,t®? where a, and a, are real and a, > 0.
224 CHARACTERISTIC FUNCTIONS

This theorem contains again as a special case the theorem of Marcin-

kiewicz (corollary to theorem 7.3.3). Marcinkiewicz’s theorem is obtained
by putting 2, = A, = 0. For the proof, which is similar to the demonstra-
tion of theorem 7.3.3, the reader is referred to Lukacs (1958).
Several authors have discussed related necessary conditions for entire or
meromorphic functions to be characteristic functions. These conditions
can all be considered to be extensions of Marcinkiewicz’s theorem.
I. F. Christensen (1962) studied functions of the form
F(t) = Kng) en[Pm(2)],
where g(t) is a characteristic function subject to certain restrictions.
R. Cairoli (1964) investigated similar problems for meromorphic functions
of finite order. H. D. Miller (1967) studied entire functions of the form
go(t) f {exp [P(2)]} or f {exp [P(d)]}, where g(t) and f(z) are entire functions
while P(t) is a polynomial. The method in all these cases is similar to that
used in proving theorem 7.3.2; the principal tool is the ridge property
(7.3.44) | f(t+iy)| < f)
which is valid for all y if f(z) is an entire characteristic function.
Far-reaching generalizations of Marcinkiewicz’s theorem were obtained
by I. V. Ostrovskii (1963). His work is based on a careful study of entire
functions which belong to families characterized by the following in-
equalities:
(7.3.45a) |f(tt+iy)| < M(ly|;f) (-o<ty < o)
(7.3.45b) Re f(t+iy) < M(|y|;f) (—© <,y < ©)
It is easily seen that the class described by (7.3.45b) contains the class
described by (7.3.45a) which in turn is wider than the family of ridge
functions. The basic results of Ostrovskii’s paper are theorems on entire
functions belonging to these classes. These theorems are interesting on
account of their applicability to the theory of characteristic functions. The
reasoning which yields these results on entire functions is tedious, and the
discussion would exceed the scope of this monograph. We therefore list,
as a lemma, only one of Ostrovskii’s results and also indicate its application.

Lemma 7.3.4. Let A(w) and b(z) be entire functions and suppose that A(w)
does not reduce to a constant. Let f(z) = A[b(z)]. If the function f (2) satisfies
(7.3.45a) then b(z) is either a polynomial of degree not exceeding 2 or an entire
function of not less than order 1 and of normal type.
We deduce from the lemma the following results concerning characteris-
tic functions:

(*) I, F. Christensen and R. Cairoli also considered functions which are not entire.
However, they had to assume that these functions are regular in a half-plane which con-
tains the origin in its interior. Inequality (7.3.44) is then valid in this half-plane.
ANALYTIC CHARACTERISTIC FUNCTIONS 225
Theorem 7.3.5. Suppose that an entire characteristic function f(t) is the
superposition of two functions A(w) and b(z), that is, f(z) = A[b(z)]. Then
b(z) ts etther a polynomial of degree not exceeding 2 or an entire function of not
less than order 1 and of normal type.
Remark. 'Theorem 7.3.2, and therefore also Marcinkiewicz’s theorem,
are particular cases of theorem 7.3.5.

Corollary 1 to theorem 7.3.5. Let b(z) be an entire function of order 1 and

minimal type; then f(z) = exp [b(z)] cannot be a characteristic function.

Corollary 2 to theorem 7.3.5. Let b(z) be an entire function of order less than
1, then f(z) = exp [b(z)] cannot be a characteristic function.
Corollary 1 answers a question raised by Yu. V. Linnik, while corollary 2
solves a problem posed by D. Dugué.
I. V. Ostrovskii’s paper also contains a generalization of theorem 7.3.4.

7.4 Periodic analytic characteristic functions

The characteristic function of a lattice distribution which has the origin
as a lattice point has the form
(7.4.1) F(t Deke

where 7 is a real number and

(By 2 0, Dr Pe = 1.

Let k be an arbitrary integer; it follows from (7.4.1) that w = 2k/r is

a period of f(t). We see therefore that a characteristic function can be
a periodic function; however, a periodic characteristic function is not
necessarily analytic. The characteristic functions discussed in theorem
4.3.2 are examples of periodic characteristic functions which are not
analytic.
In the present section we discuss briefly the properties of analytic
characteristic functions f(z) which are single-valued and periodic. We
consider first the case where f(z) has a purely imaginary period w = iy
(n real); it is then no restriction to assume that 7 > 0.
We wish to avoid the discussion of trivial cases and suppose therefore
that f(z) #1. Let —« < Im(z) < f, (« > 0,8 > 0), be the strip of
regularity of f(z); we first show that necessarily 7 > min (a, f). We give
an indirect proof and assume that 7 < min («, f). The points 2; = ™,
2, = —%in are then in the interior of the strip of regularity of f(z) and
it follows from theorem 7.1.4 that

(7.4.2) f(0) < ee

226 CHARACTERISTIC FUNCTIONS

On the other hand it follows from the periodicity of f(z) that f(0) =
f(y) = f(—-in) = 1 so that f(0) = [f(m)+f(—m)]/2 = 1 in contra-
diction with (7.4.2). The indirect proof is therefore completed and we have
always 7 > min («, 6). But the equality sign would imply that the origin
is a singular point of f(z) so that always 7 > min (a, 6). But then at least
one of the inequalities 7 > « or 7 > f holds. If 7 > « [respectively 7 > ]
then i(7j—«) [respectively —7(j —)] is a singular point of f(z) located in the
upper [respectively lower] half-plane. Therefore y—« > 6, and we have
established the following result:

Theorem 7.4.1. If a non-constant analytic characteristic function has a

purely imaginary period w = 1n (n > 0), then this period ts at least equal to
the width of the strip of regularity of f(z), that is |w| = > «+f.
We consider next the case where f(z) has a complex period w = +7.
The case € = 0 (purely imaginary period) has just been treated, so that we
may assume & # 0. Using (7.1.4) and the assumption that m is a period of
f(z), we conclude easily that @ and — @ are also periods. Therefore 2 and
2nt are also periods of f(z) so that
(7.4.3) f(2é) = 1.
We conclude then from theorem 2.1.4 that f(z) is the characteristic function
of a lattice distribution whose lattice points are the points where 1 — cos 2&x
vanishes. Therefore f(z) is given by
(744) f(z) = E_ p.exp (dens/£)
where “ae
C25) ne
If 77 ="0, "then 7 (2) 4 bitie periodic and has a real period &, so that
f(0) = f(&) = 1, and we see by the same argument that it can be written
as
(TehO\eanFeV oh Ps exp (2nizs/E)
where the p, satisfy again (7.4.5). If 4 #4 0 then f(z) is given by (7.4.4) and
is a doubly periodic function which necessarily has a real and also a purely
imaginary period. We summarize this in the following manner:
Theorem 7.4.2. An analytic characteristic function which is single-valued
and simply periodic has either a real or a purely imaginary period. The period
1s real tf, and only if, the characteristic function belongs to a lattice distribution
which has the origin as a lattice point.
Let f(z) be an entire characteristic function which does not reduce to a
constant and assume that it is periodic. From theorem 7.4.1 we see that it
ANALYTIC CHARACTERISTIC FUNCTIONS 227

cannot be doubly periodic, and we can conclude that it must have a real
period and have the form (7.4.6).

Theorem 7.4.3. If a characteristic function ts an entire periodic function then

it 1s necessarily the characteristic function of a lattice distribution which has
the origin as a lattice point.
It is easy to give examples of analytic characteristic functions which are
periodic. We mention the Poisson distribution whose characteristic function
has the real period 27; the distribution with frequency function p(x) =
[2 cosh (xx /2)]|~! has the characteristic function f(z) = 1/(cosh z) which
is regular in the strip |Im (z)| < 2/2 and which has the purely imaginary
period 27. A doubly periodic characteristic function was constructed by
M. Girault (1955) who showed that the elliptic function
© 1 —f2n-1] 4 Rem-1¢iz
f(z) —_ II 1+22"-1 1 — R-1¢
n=—0

is a characteristic function. This function has the real period 2, the purely
imaginary period 47 log k and the strip of regularity |Im (z)| < |log k|.
In conclusion we remark that one could regard theorems 7.4.1, 7.4.2
and 7.4.3 also as conditions which a single-valued, periodic analytic
function must satisfy in order to be a characteristic function.

7.5 Analytic characteristic functions as solutions of certain

differential equations
Regression problems lead sometimes to a differential equation for the
characteristic function. After all solutions of this equation are found, one
has to determine those which can be characteristic functions. This is often
the most difficult part of the problem and it is therefore desirable to find
general properties of characteristic functions which satisfy certain differ-
ential equations.
In the present section we discuss a result due to A. A. Zinger and
Yu. V. Linnik, which is of great theoretical interest.
We write f(t) for the derivative of order s of f(t) and consider the
differential equation
(TSA) DA, gir tpl)... fo O) = eff Ol”
The A,_,, are real constants while the sum is here taken over all non-
negative integers j1, Jo, .- - 5 Jn Which satisfy the condition
(lay efile.
tie = Gf;
2 0; S =15%.., 2).
We assume that at least one coefficient with j,+jo+...+j, = m is
different from zero and denote by m the order of this differential equation.
228 CHARACTERISTIC FUNCTIONS

We adjoin to the differential equation (7.5.1) the polynomial

(7.5.2). SAM a hel = * yA

© (Si5c0s0Sh)

The first summation is here to be extended over all permutations

(51, Sa, . ++) Sn) Of the numbers (1, 2,..., 7); the second summation over
all integers j,,..., J, satisfying (7.5.1a).
The differential equation (7.5.1) is said to be positive definite if its
adjoint polynomial (7.5.2) is non-negative. We can now state the result of
A. A. Zinger and Yu. V. Linnik.

Theorem 7.5.1. Suppose that the function f (t) is, in a certain neighbourhood
of the origin, a solution of the positive definite differential equation (7.5.1) and
assume that m > n—1. If the solution f(t) 1s a characteristic function then
it is necessarily an entire function.
We state first a lemma, which uses only some of the assumptions of
theorem 7.5.1 and which therefore yields less information concerning the
solutions of (7.5.1).

Lemma 7.5.1. Suppose that the characteristic function f(t) 1s, in a certain
neighbourhood of the origin, a solution of the positive definite equation (7.5.1).
Then f(t) has derivatives of all orders at the origin.
Lemma 7.5.1 is certainly true if the distribution function F(x) of f(¢)
is a finite distribution [see theorem 7.2.3]. We therefore assume in the
following that for all x > 0
(7.5.3) F(—x)+1—F (x) > 0.
We remark that the assumptions of the lemma imply that f(z) can be
differentiated at least m times. Moreover, m is necessarily an even number
if A(x,, Xg,..., X,) is non-negative. Since f(t) is the characteristic function
of F(x), we know that

(7.5.4) fo@=i|” setear(a) HEMI

In view of (7.5.2) and (7.5.4) we can write (7.5.1) in the form
(7:5.5)

le _ ie Alotyy «<4 Hp) exp [it(oty +. atq)] dF (x,) .. .dF (609)

vee c ie rae in exp [7f(x,+...+%,)|dF (x)
...dF(x,).

We give an indirect proof for the lemma and assume therefore that f (¢)
has only a finite number of derivatives. Then there exists an even integer
ANALYTIC CHARACTERISTIC FUNCTIONS 229

2p such that f(t) can be differentiated at the origin 2p times but not
(2p+2) times. Clearly, one has
2p: > m > 2.
The function on the right of (7.5.5) can then be differentiated 2p—m-+2
times. Since A(x,,..., x,) is non-negative we can conclude from Fatou’s
lemma [see Titchmarsh (1939) p. 346] that

| oe | See oe Nahi t fy) EXD (tlXeno Hk)

x dF (x)... dF (xn)
=¢ | ys | (#4... s+%_)?-™T* exp [2¢(4,+...+%,)]
AE 8s) aes CE (Xp)
or, putting ¢ = 0,

(7.5.6) lip As Alotyy « « « 5 Sq) (20y-E. «boty PP?

al (Wa) ee Reda (4)
4 c| | (pte et ag)? ™H2dF (0). OF (ot).
The differential equation (7.5.1) has by assumption the order m so that
the polynomial A(x,,...,x,) contains the mth power of at least one
variable. It is evidently no restriction to assume that x, is this variable; one
can then write
(75.7) A Ks es Wa) = PAG Madar, Xela ta (hae 22) xP
ae cee MA BT tern
Since A(x,,...,%,) is non-negative we conclude that Ag(x:,..., Xn) is
also a non-negative polynomial.
It is always possible to find a bounded region Q,_, in the (n—1)-
dimensional space R,,_, of the variables (x,,..., %,) such that

(7.5.8a) | dF (1,)
dF(xs)... dF (tq) = « > 0
while
(7.5.8b) MINAS (Mayes hy Ka) SC 1
Q UR ps
This follows from (7.5.3) and the fact that the equation A 9(%2,...,%n) = 0
determines an algebraic surface in R,_,. We use here, and in the following,
the symbols C,, C,, ... to denote arbitrary positive constants.
We see from (7.5.7) that it is possible to find a sufficiently large C, > 0
such that for |x4 | = Gy and:(x,,..) “a) 2,4, the’ rélations
Ali, chnls Ma) ce Cady:
oe)
459
eee rte > C,|x,|
hold.
230 CHARACTERISTIC FUNCTIONS

Let Q,, be the set of all the points (x1,...,%») of the 2-dimensional
space which satisfy the condition
(75410) [%.| > Cy “and aimee.
Since m is even and A(x,, x2,...,X,) > 0 we conclude from (7.5.6) that

(7.5.11) | Alig «ony Xa) Barbs vet 8a)? TdF On) eda) <K
Q n

where

Kee oye see [ tee tag)

dF(a)... dF(ty)
—o v —-O

is a (finite) positive constant. Substituting (7.5.9) into (7.5.11), we see that

COT | amnestatC) koopa HN = cs | xy?dB(x1) < K.

Qn || >Ce
This inequality indicates that the moment of order 2p+2 of F(x) exists;
this is in contradiction with the assumption concerning p, so that the
indirect proof of lemma 7.5.1 is completed.
We proceed now to prove theorem 7.5.1. As a first step we show that
f(t) is an analytic characteristic function. We need the following lemma.

Lemma 7.5.2. Let G be a positive integer; then e > G°/G!.

To prove the lemma we note that
A 0 Gi GG
er = ar = GA:

Let N and m be two positive integers; according to lemma 7.5.2 we have

(2N+ mye eN +m

(2N+m)!
or
(7.5.12) cat Es
+m) < é.
}/en
[(2N+ m) |
We again use the region ,,_, introduced in the proof of lemma 7.5.1
and write

(he I |g Xg+. «+ Hy |’dF (a)... OF (%,).

It is then possible to find a positive number b such that
G25:13)" eds 0) Oo juan,NG
We consider also the set of those points (xj, %2,..., Xp») of the n-dimen-
sional space which satisfy the relations
(7.5.14) 4 | SE wand Ge aaens ex, ie Qos.
This set is bounded, therefore there exists a positive constant C, such that
ANALYTIC CHARACTERISTIC FUNCTIONS 231

(7.5.15) 1= | ll (]oey
[agt..|—
tay
|e] <C, Qn-1

x dF (x%_)...dF (*) G(x) = 0,0." 0.

The constant C, was introduced in the proof of lemma 7.5.1. We write

p= {" |xltar(x)
for the absolute moment of order k of F(x). Clearly it is possible to find
a sufficiently large positive constant M, such that the inequalities
(7.5.16) B, < k! Mi exp {[1-(-1)]/2}
hold fork =. 00... m=.
We next prove the following lemma:

Lemma 7.5.3. Suppose that the conditions of theorem 7.5.1 are satisfied
and that M, ts a positive constant such that
M5" C,0, 4(C,/ May" + ele — 1) < 4,
where C,= c/C; while by, b, Cs and C, were defined earlier.
Let M = max (M,, M,, Cg, 2be) and assume that
(7.5.17) Bu < Rk! M* exp {{1-—(—1)*]/2}
jor k= 0, 1,2). «, 2N-++-m—2. Then
Bexam< (2N-+m)! M24,
We differentiate equation (7.5.5) 2N times with respect to ¢, and putting
= (0 we obtain
(7.5.18) ee ‘3 i) Albets Hay « « + Hq) (yes bmg)
SH (41) ae dl (X,,)
beg {hee . \pae(ey. tay)?"
dF (x1)... dF (2).
It follows from the first relation (7.5.9) and from the obvious inequality
[Hy +X%e+...+%,| 2 l= [ahs te |
that
(7.5.19)
| aa Hi (ie (Bal ae eee 2 dP(x). .dF(x,)| dF (x)
Pelikan = Oud
where
(7.5.19a) j={ mi (U UTLEN LINER
PT O
232 CHARACTERISTIC FUNCTIONS

We add the integral J, defined in (7.5.15), to both sides of (7.5.19) and see
easily that

0 < (7) -1) Basramnads < CoI+E

>

=0
It follows from this inet and (7.5. He that
(7.5.20)

[boBaysn—2NB, Bavsmes| < Cr J4 bo CRon + Se3 Pall(77) Buxsm-sbo

j=

We estimate next the expressions on the right of (7.5.20).

We see from (7.5.13) and (7.5.17) that

© (PYfaxeem-sbs<ME eby3(Nm
—j)1 (N/M).
We note that
ON+m-})! fea i Cee ee ee

1
< 3Q2N+m)!
so that
(7521s en Bon am_1b; < (2N-+m)! M2¥+” eb, (@/f 1).
j=2
We turn now to the expression (7.5.19a) and see that

I< Ba
jit iin sen Jul
Bo
We again use (7.5.17) to show that
J SAN IME Ao
where o is the number of terms in the multinomial expansion of
(x+x,+...+%,)?%. It is not difficult to show that the number of terms in
a homogeneous polynomial of degree p in variables cannot exceed
(?+n— 1)
a—1 /’
using this fact we conclude that
(7.5.22) J <(2N+n—-1)! MW e"/(n—1)!
It follows then from (7.5.20), (7.5.21) and (7.5.22) that
|Box+m—2Nb,
by*Bay ama| < (2N+n—1)! MPNeC, by */(n—1)!
+ C2Nt™ + (2N +m)! M2N+ e(e?/@— 1).
According to the assumptions of theorem 7.5.1 we have
n—-l<m
ANALYTIC CHARACTERISTIC FUNCTIONS 233
so that
| Ban +m —2Nb, b> Bay +m |
P< (2N+ m)! M2N +m CY ha en Ca pet af (Co/M )2N +m + e(er/t — 1)}.

In view of the definition of M we then have

(7.5.23) |Ba-im—2Nb1 5 * Baw im—1| < 3(2N-+m)! MeN +,
For the further discussion of (7.5.23) we consider two mutually exclusive
possibilities described by the inequalities
(7.5.24) 2Nb,b)*
Ban +m—1 < tBowsm
(7.5.25) 2Nb, by * Bon +m—1 oe 2Pon+m-
We examine first the case where (7.5.24) holds. Then
Pan +m —2Nb, by * Ban m1 z 4Bonam .

We see then from (7.5.23) that

(7.5.26) Banam < (2N+m)! M2N+m,
We consider next the second case and assume that (7.5.25) holds. It is
known [see (1.4.7)] that
(7.5.27) Peniet= (Caan)
0
We substitute this into (7.5.25) and see easily that
Bon +m =< (4b) oe
or, using (7.5.12),
Bansm < (2N+m)!(2be)2+™ < (2N+m)! M2N+™,
It follows that (7.5.26) is also valid in the second case, so that lemma 7.5.3
is proved.
We show next that condition (7.5.17) holds for any positive integer k.
We establish this fact by induction; in view of lemma 7.5.3 it is only
necessary to show that condition (7.5.17) holds also for k = 2N+m-—1.
We substitute the expression (7.5.26) into (7.5.27) and see that
. (2N +m)! VAN +e-1
Bon +m-1
OO
< Banam)er re + m) !]/2N +m)
[(2N
It follows from (7.5.12) that
Benam—1 < e(2N+m—1)! M2N+m-1,
Thus condition (7.5.17) holds for k = 2N+m-—1 and therefore also for

(g<t(<)a
all positive integers k. We have then

and conclude that f(¢) is an analytic characteristic function which is regular

at least in the strip |Im(z)| < 1/M. We write as usual f(z) for the
function of the complex argument z = t+1y (t, y real) which agrees with
the characteristic function f(¢) on the real axis.
234 CHARACTERISTIC FUNCTIONS

We complete the proof of the theorem by showing that f(z) is an entire

function. This is accomplished by proving that the integral

(7.5.28) ie e” dF (x)
exists and is finite for arbitrary real y. We give an indirect proof and
suppose that the least upper bound 7 of all |y| for which the integral
(7.5.28) exists is finite. Then

FS
ive. Ife
oll OO

We now select a real yy > 0 such that

1
(75.29) N—aq < Yo < 7

Since f(z) is regular in the strip |Im (z)| < 7, the relation (7.5.5) is also
valid if we replace the real variable ¢ by the complex argument z = t+z7y
with |y| < 7. We do this and differentiate the new relation 2N times with
respect to z and then put z = —7yy. In this way we obtain the equation

(7.5.30) a me ie AGES SN Gap aeyeN

x exp [Vo(% +... +%,)] dF (x,) ...dF (xp)

=c| ran (yt... +3) exp [yo(ay +t... +2t,)] dF (x) ...dF (2%).
We divide both sides of (7.5.30) by A” where

ne | et AF (x)
and introduce the distribution function

G53. Gone “ie e dF (2).

We see then from (7.5.30) that G(x) satisfies a relation which corresponds
to the equation (7.5.18) for F(x). We conclude as before that the charac-
teristic function g(t) of G(x) is regular at least in the strip |Im (z)| < 1/M
so that

|= e"* dG(x)

exists and is finite if |u| < 1/M.

We see from (7.5.29) that it is always possible to select a real 2, such that
ANALYTIC CHARACTERISTIC FUNCTIONS LEAN

Then the integral

[exp [ewer yea] dr (s) = AY” exp (woe) dG)

exists and is finite. In a similar manner one can show also that the integral

[exp [= (uty) F(a)

exists and is finite. In view of the definition of 7 this is impossible, so that
the proof of theorem 7.5.1 is completed.
A. A. Zinger and Yu. V. Linnik (1957) also give in their paper further
conditions on the polynomial A(x,, x.,...,,) and on the solution f(¢)
which ensure that the only positive definite solutions of the equation
(7.5.1) are the characteristic functions of normal distributions.
 8 FACTORIZATION OF ANALYTIC
CHARACTERISTIC FUNCTIONS

In Chapter 6 we dealt with the factorization of distribution functions and

of characteristic functions and derived several general theorems. In the
present chapter we restrict ourselves to the study of decompositions of
analytic characteristic functions. This specialization permits us to obtain
further results by applying the tools furnished by the theory of functions
of a complex variable.

8.1 Properties of the factors of an analytic characteristic

function
Let f(z) be an analytic characteristic function which has the strip
=e < Im(z) <6 “@ > OF 6S 0)
as its strip of regularity. Suppose that f(z) is decomposable and has the two
non-degenerate characteristic functions f, (¢) and f(t) as factors. Then
811) f= AOK®
for real t; the corresponding distribution functions then satisfy the relation

(8.1.2) F()= |" Fie-y)dPi(y) = | F@-y) dFA(9).

Let A > 0, B > 0 and &, > &, be four real numbers; it follows from
(8.1.2) that

(8.1.3) H(E5) ies) | FAG-9)-F 6-91 4Pa(9).

We choose a fixed real number v such that —« < vw < f; since f(z) is an
analytic characteristic function we know that the integral
Me e”* dF (x)
exists and is finite and that
|"dF (x) > ise°* dF (x)
where a and b (b > a) are ane real Autnbele We next consider the integral
[e* dF’ (x) and represent it as the limit of Darboux sums. We construct
a sequence of subdivisions of the interval [a, b] by defining
b-—
x = at-“(j-1) {f= 1,2,...,(2*+1) and n= 1,2,..,}
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 237
so that
(8.1.4) Moree cae ee Oforl rake O08»: (27+ 1).
We can then write
b Qn

(8.1.5) [ed F (x)= lim ¥ exp(ox)LF (2) —F

a n—>o j=1
Qn

= lim & exp (ox? LF(21) -FP).

n—>o j=1

We denote by
(xi — y)] ifooo
hy n(¥3 v) = ‘ies(vx?) [Fy Cree —y)—-F;,

exp (ox{),)[F (42, -—y)-—Fi(xj?—y)] ifvo <0

Pease and by
for j a
Qn

En(¥3 0) = j=B hin(¥i 2).

We see then from (8.1.3) and (8.1.5) that
b B
(8.1.6) | e dF (x) > lim | ga(¥; 0) dF;(9).
a n>o /J—A

Using (8.1.4) together with the inequality

At < Qt < yietD
we see that
hia (¥3 v) < hoinaa (V3 0) + host (y3 2)
so that
En(¥3%) < Snir (¥3 2).
From the definition of the functions g, (¥) it follows that they are Darboux
sums and that
b—

Lio V5@ | Tents) dF, (2).

n> 0 a—y

We then apply to (8.1.6) the monotone convergence theorem [Loéve (1955),

p. 124] and conclude that

[ ear(x)> | [limgs(yi2)]dFa(9)
b B

or

6 > | dF (x) > | ev dF (x) > | en | dF, (2)|dF,(y).

co b B b-¥

—o a —A a—y

We note that

| &* dF,(2) > | o” dF, (2)

b-¥ b—B

a-y a+A
238 CHARACTERISTIC FUNCTIONS

so that

|‘_e°dF(s) > fSu dF, (| ||ee dF, (2).

The integral on the left of this inequality is finite and independent of
a, b, A, B. Carrying out the necessary passages to the limit, we see that the
integrals

ie e dF, (x) and

(8.1.7) =o
|| evdF (9)
exist and are finite and that

(8.1.8) | et dF (x) > | e”" dF (») | eo”dF,(y).

Here v is a real number such that —« < v < f, so that the integrals
(8.1.7) exist for all such v. But then the integrals

fie) = |e dFy(w) and f,(2)= [°c aFs()

exist and are finite for all complex z such that —« < Im(z) < f, and we
see that f,(z) and f.(z) are analytic characteristic functions whose strip of
regularity is at least the strip of f(z). Moreover equation (8.1.1), which
holds for real ¢, is also valid (by analytic continuation) in the entire strip of
regularity of f(z). We summarize this result as

Theorem 8.1.1. Let f(z) be an analytic characteristic function which has the
strip —a < Im(z) < f as its strip of regularity. Then any factor f,(z) of
f (2) ts also an analytic characteristic function which is regular, at least in the
strip of regularity of f (2).
We now turn back to inequality (8.1.8). There exist two real numbers
a, and a, such that 0 < F,(a,) while 1 > F,(a,). Then

|” & dF, (x) > &*[1—Fy(a)] if v> 0

| e dF .(y) 2 a
=o | edF,(x) > F,(a,) fv < 0.
Let C-1! = min [F,(a,), 1—F,(a@,)] and a = max [|a,]|, |a, |]; we then see
that

| e dF, (x) < Cel |"aE (x),

Corollary to theorem 8.1.1. Let f(z) be a decomposable analytic charac-
teristic function with strip of regularity —« < Im(z) < B and suppose that
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 239

fi (2) ts a factor of f(z). Then there exist positive constants C and a such that
fi(—iv) < Cel f(—-iv)
for all v satisfying —« < v < f.
We next consider an important particular case and suppose that f(z) is
an entire characteristic function.
Theorem 8.1.2. Every factor f,() of an entire characteristic function f (2)
is an entire characteristic function. The order of the factors of an entire
characteristic function f(z) cannot exceed the order of f(z).
The first part of this statement follows immediately from theorem 8.1.1.
The second part is a consequence of the relation M(r; f,) < ce" M(r; f)
which is easily obtained from the corollary and from the equation M(r; f)
= max [f(zr), f(—7r)] which was derived in Section 7.1.
Corollary to theorem 8.1.2. Let f(z) be an entire characteristic function of
order p > 1 and type t and suppose that f, (2) is a factor off(z). If the order
of f:() ts also p, then the type t, of f(z) cannot exceed the type t of f(z),
ty &
The statement of the corollary is obtained in the same way as the state-
ment of the theorem, using the definition of the type given in Appendix D.
Remark 1. 'The statement of the corollary does not hold if either p = 1
or if py < p, where p, is the order of f,(z).
Remark 2. Let f(z) be an entire characteristic function without zeros so
that f(z) = e*” [4(z) entire, zs= t+iy]. Then every factor f, (z) of f(z) is
also an entire characteristic function without zeros and therefore has the
form f,(z) = e* where ¢,(z) is an entire function.
We conclude this section by deriving a property of entire characteristic
functions without zeros.
Theorem 8.1.3. Let f(z) be an entire characteristic function without zeros
which has a factor f,(z). The entire functions ¢(z) = log f(z) and ¢,(2) =
log f,(s) then satisfy the relation M(r; ¢,) < 6rM(r+1; ¢)+Cr(r+1),
where C is a positive constant.
For the proof of the theorem we need two lemmas.)
Lemma 8.1.1. Let f(z) be a function which ts regular in a region G, let
Zo = ty tity, be an interior point of G, and let A be the distance between 24
and the boundary of G. Then
ia ote pe’ +(z—&)
(Si1.9) f(z) = =|, u(ty+p cos0,yo+p sin 8) ee etre

(*) Let w = é+%n be a complex number and let /(w) be a function which is regular in
a certain region. We write then u(£, 7) for the real part of f(w).
240 CHARACTERISTIC FUNCTIONS

Moreover,
1 27
'
(8.1.9a) — f'(%0) = ae | u(tyt+p cos 0, yotp sin 0) edd.
p Jo
Here 2, is an interior point of the region G; z is a point in the interior of the
circle with centre x, and radius p such that |z—2 | < p < A, while
By = Im [f(2o)].
The representation (8.1.9) is known as Schwarz’s formula; for its proof
see Markushevich (1965) [vol. 2, p. 151]. If we differentiate (8.1.9) with
respect to z and put z = 2, we obtain (8.1.9a).

Lemma 8.1.2. Let f(z) be an entire characteristic function [z = t+ ty; t, y

real], then there exists a positive constant M = M,, which depends on f but
is independent of yy, such that log f(ty) > —M|y\|.
According to theorem 7.3.2 we have log f(iy) > —/yx,, where x, = if'(0).
The statement of the lemma follows from the fact that —yx, > —| «|||
if x, # 0, so that in this case M = |x,|. If x, = 0, M is an arbitrary
positive number.
We proceed to the proof of theorem 8.1.3 and write

sito) {“%9)
»¥) = Re [$(é+2y)]
ere tate y) = Re [di (¢+4)]
It follows easily from theorem 7.1.2 that
(8.1.11) 0 <m(0,9)-m(t, 9) < W(0, y)—ult, ») < 2M (734),
where 7 = |¢+7y|.
Since f,(z) is a factor of f(z) there exists an entire characteristic func-
tion f,(z) without zeros such that f(z) = fi (2) f.(z). We write 4,(z)=
log f,(z) and u,(t, y) = Re [¢,(t+zy)], so that
(3.1.12) u(t, y) = u(t, y)—Uue(t, y).
We conclude from lemma 8.1.2 that there exist positive constants M/, and
My, such that
log f(y) > —Ms|y| (7 = 1,2)
and note that
u;(0, y) = log |fi(zy)| = log f;(%).
Hence
(8.1.13) 4;(0, 9) > —M;|y| (7 = 1, 2).
It is also easily seen that
(8.1.14) u(t, y) = log |f(t, )| < M(r; 4),
where r = (¢?+?)””, It follows from (8.1.12), (8.1.13) and (8.1.14) that
—M,;|y| < 4 (0, y) = u(0, y)—u2 (0, y) < u(0, y)+Ma|y|
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 241
or
—Mi|y| < (0,y) < M(y;
¢)+ Maly.
Therefore there exists a positive constant C such that
(8.1.15) | (0, »)| < M(y3 4)+C |y].
Clearly
[m(t,9)|< [ur (0, y)| +12 (0, »)— m4(9) |
or, using (8.1.11) and (8.1.15),
|v (t, y)| < M(y; 4)+C | y|+2M(r;4).
Since |y| < r we have
(8.1.16) [u(t )| < 3M(7; ¢)+C |y|.
We apply formula (8.1.9a) of lemma 8.1.1 to 4,(z). Since ¢,(z) is an
entire function we may put p = 1 and we write also z, ¢ and y instead of
Zo, ty and yo, respectively. We obtain
; 1 27 ' ‘

$,(2) = =| u,(t+cos 0, y+sin 6) e~* do.

0

It follows from (8.1.16) that

|b: (z)| < 6M(r+1; 6)+2C(r+1).
Since

If(z)| = |= [ (20) do| < |2| max |¢%(20)|

0 0<v<1

we see that
ld.(z)| < 6rM (r+1; ¢)+2Cr(r+1).
This is the estimate given in theorem 8.1.3.

Corollary 1 to theorem 8.1.3. Let f(z) be an entire characteristic function

without zeros which has a factor f,(z) and write ¢(z) = log f(z), $1(8) =
log f(z). The order p, of 6, (z) cannot exceed the order p of ¢(z). Moreover, if
p1 = p then the type of 4,(z2) cannot exceed the type of (2).
The statement of the corollary follows immediately from the theorem
and the definitions of order and type of an entire function given in Appen-
dix D.
Remark. ‘The estimate of theorem 8.1.3 can sometimes be improved,
namely if it is possible to find for u,(0, y) and u,(0, y)—m(t, y) better
bounds than those of formulae (8.1.11) and (8.1.15).
Suppose that we have
(8.1.11a) O< 4,(0, y)—u (4) < Alé,y)
(8.1.15a) — |#4(0,)| < Bly)
242 CHARACTERISTIC FUNCTIONS

where A(t, y) and B(y) are non-decreasing functions. Repeating the pre-
vious argument we get
| (t,)| < A(t, y) + BY)
and
(8.1.17) [di(s) < 212/A+
1 y+1)42/2/Biy+l) (2 = t+9).
We give an example which we shall use in the next section. Let
f(z) = exp {A(e*—1) +ipz—ya*},
where yu is real, while y > 0 and A > 0. We suppose that f(z) admits the
decomposition,
f(2) = fil®)f2(2)-
The function f(z) is an entire characteristic function without zeros; we
write again u(t, y) = Re [log f(t+zy)] = Re [A(t+zy)] and use analogous
notations for the factors f,(z) and f,(z). Then
wae
(8.1.18) u(0, y)—u(t, y) = log f(y) p=\ 2Ae"* sim? atye,
f(t+1y)
so that we see from (8.1.11) that
(8.1.19) A(t, y) = 2A etl+ yit?.
According to lemma 8.1.2 there exist two positive constants M, and M,
such that
dx(ty) > —M,|y| (Rk = 1,2)
and we see that

(8.1.20a) m4 (0, y) = Re [41 ()] = d1 (ty) > -—Mily!-

According to our assumptions we have
$1 (ty) = log f(ty)—
$2(iy)
so that

(8.1.20b) (0,y)< Ae“ 1) — wy + yy?+ Maly |.

Hence
|#1 (0,y)|< Ae” + yy?+O(|y |);
so that
(8.1.21) By) = del + yy?2+ O(|y |).
We see therefore from (8.1.17), (8.1.19) and (8.1.21) that
(8.1.22) |$1(2)| = Of]2] exp [| Im (z) []+] 2/9} (12 |> &)
and have obtained the following result:

Corollary 2 to theorem 8.1.3. Let f(z) be the characteristic function of the

convolution of a normal and a Poisson distribution,
f(a) = exp [A(*—1)
+ine—ys%].
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 243
If f(z) = exp [¢,(z)] is a factor of f(z), then
$1 (2) = O{|2| exp [|Im (z) |]+]/2|9}
as |z| > ©.

8.2 Factorization of certain entire characteristic functions

Certain entire characteristic functions have interesting factorization
properties. We next prove an important theorem concerning the decom-
position of the normal distribution; this theorem was first conjectured by
P. Lévy and somewhat later proved by H. Cramer.

Theorem 8.2.1 (Cramér’s theorem). The characteristic function f(t)

= exp [tut—o*t?/2] of the normal distribution has only normal factors.
Moreover, if f(t) = fi(t)fe(t) with f(t) = exp [iu;t—o?t?/2] (j = 1, 2),
then f+, = mw and o?+0% = o?.
The function f(t) is an entire function without zeros; it follows then
from theorem 8.1.2 that the same is true for its factors and that the order of
these factors cannot exceed 2. Therefore f,(z) has the form
f, (2) = exp [g,(2)]
and it follows from Hadamard’s factorization theorem that g,(z) is
a polynomial of degree not exceeding 2. Let for real argument 1,
£:(t) = 4)+a,t+a,t?; since f(0) = 1 we see that a) = 0. From the rela-
tion g,(—2) = (2)
g, we conclude that a, = iu, is purely imaginary and that
a, is real. Since a characteristic function is bounded for real values of its
argument we deduce finally from |f; (¢)| = exp [a,2?] that a, < 0 and set
a, = —4o?. Thus f,(¢) = exp [ju,t—4o7t*] is the characteristic function
of a normal distribution. The same argument applies to f,(t) while the
relation between the parameters of f(t) and those of its factors is established
by elementary reasoning.
We discussed in Section 6.2, without giving any examples, characteristic
functions which have no indecomposable factors. Cramér’s theorem shows
that the characteristic function of the normal distribution belongs to the
class of characteristic functions without indecomposable factors. Our next
theorem indicates that the characteristic function of the Poisson distribution
also belongs to this class. The following factorization theorem was derived
by D. A. Raikov and is in some respects similar to Cramér’s theorem.
Theorem 8.2.2 (Raikov’s theorem). The characteristic function f(t)
= exp [A(e“—1)] of the Poisson distribution has only Poissonian factors.
Moreover if f(t) = fi(t)f2(t) with f;(t) = exp [A;(e*—1)] (j = 1, 2) then
Ata, =A
To prove the theorem we suppose that
(8.2.1) f(t) = exp [Ae"—1)] = AOA)
244 CHARACTERISTIC FUNCTIONS

is decomposed into two non-degenerate factors. Since the convolution of

a discrete and a continuous distribution is always continuous, we see that
fi(t) and f(t) are necessarily characteristic functions of discrete distri-
butions. The Poisson distribution f(t) has its discontinuity points at the
non-negative integers; it is then no restriction to assume that the discon-
tinuity points of f,(¢) and f,(é) are also non-negative integers. ‘Then

Ji) = Sraje? andes) = > bj e, with ase eb, 10;

v=0 v=0
oo ogee
> a = by = "1" “and where f(t) =e" erie
v=0 iMs8
0 v=0 VU:

Since f(z) is an entire function without zeros, the same is also true for
fi(t) and f,(t), so that these series also converge for arbitrary complex
values of the argument. We now introduce a new variable w = e”; this
transforms the characteristic functions f, (t), f(t) and f(z) into the gener-
ating functions g, (2), g,(w) and g(w) respectively. Here

g(a) = e* Sat, gi) = Saya, galt) = ES bya

co) K» ioe) co)

= ~ v= v=0

and g(w) = g,(w)g.(w). The coefficients of these power series satisfy the
equation

(8.2.2) a)b6,+4,b,_14+. . .+a,_1b,+ a,b, = ete

vi
and it follows from the non-negativity of the a, and b, and from the
relation
ao by — er sé 0

that

(8.23) a, < byte te.

Since g(w) = exp [A(w—1)] is an entire function, we conclude from (8.2.3)
that g,(w) is also an entire function, and we see that for real ¢
8(t) < by *g(2).
It is also easy to verify that
M(r;g) = by)M(r, £1)
so that the order of g,(w) cannot exceed the order of g(w). The function
g(w) is an entire function of order 1 without zeros; therefore g,(w) has
the same property. We conclude from Hadamard’s factorization theorem
that g;(w) has exactly the order 1. Since g,(1) = 1 we see that g,(w)
= exp [A,(w—1)] so that f(t) = exp [A, (e”—1)] is the characteristic func-
tion of a Poisson distribution. A similar argument applies to f,(¢) and it is
easily seen that A, > 0, A, > 0, A, +A, = A.
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 245

The following corollary follows almost immediately from Raikov’s

theorem.
Corollary to theorem 8.2.2. A Poisson-type distribution) has only Poisson
type factors.
One can summarize theorems 8.2.1 and 8.2.2 by introducing the follow-
ing definition. A family of characteristic functions (or distribution func-
tions) is said to be factor-closed if the factors of every element of the family
belong necessarily to the family. The preceding results mean that the nor-
mal family, as well as the family of Poisson type distributions, is factor-
closed. H. Teicher (1954) showed that a family which contains the binomial
distributions is factor-closed. For the binomial distributions this fact was
already noted by N. A. Sapogov (1951). In this connection we mention an
interesting result which describes another family of characteristic functions
which is factor-closed. Yu. V. Linnik (1957) derived the following
generalization of the theorems of Cramér and Raikov.
Theorem 8.2.3. Let
f(t) = exp {A(e*—1) + iut— 407 t?}
(u real, o? > 0,4 > 0) be the characteristic function of the convolution of a
normal and of a Poisson distribution. Suppose that f(t) has the decomposition
FQ) = AOA. Then
f;(t) = exp {A;(e—1)+iu,;t—40%2?} (f = 1, 2)
where A = 2,+A, and o? = 02+ 03.
We note that theorems 8.2.1 and 8.2.2 can be obtained from this result
as particular cases. However, the proof of theorem 8.2.3 requires more
powerful analytical tools than theorems 8.2.1 and 8.2.2. This is explained
by the fact that theorem 8.2.1 deals with an entire function of finite order
while theorem 8.2.2 treats the characteristic function of a lattice distri-
bution. Under the assumptions of theorem 8.2.3 both these advantages are
lost and the proof becomes much more complicated.
For the proof of theorem 8.2.3 we need certain results from the theory of
analytic functions which we state as lemmas.
Lemma 8.2.1. Let f(z) be a function which is regular in the angle
Q = {z:0 < |x| < «,« < arg(z) < B} and which satisfies the following
conditions :
(i) |f(z)| << M, exp (|2 |?) for z € D where p < x/(B—«).
(ii) |f(2)| << M on the lines z = xe'* and z = xe” forming the
boundary of 2.
Then |f(z)| < M for all z€Q.
(*) That is a distribution with characteristic function of the form
1)).
f(t) = exp [tje+ Netto—
246 CHARACTERISTIC FUNCTIONS

Lemma 8.2.1 is a special case of the Phragmén-Lindelof theorem; for

its proof we refer the reader to Titchmarsh (1939) [p. 176] or to Markushe-
vich (1965) [vol. 2, p. 214].

Lemma 8.2.2. Let f(z) be an entire periodic function with period T, such
that the inequality |f(z)| < Ke‘ (K and a are real and positive constants)
holds. Then
i(2)= > Cy EXP Fe ahh
k=-Tt rh

is a trigonometric polynomial with t = [a| T |/2a].

Lemma 8.2.2 is a consequence of the theorem [see Markushevich (1967),
vol. 3, p. 143] which states that a non-constant, periodic entire function of
exponential type is necessarily a trigonometric polynomial.
We proceed to prove theorem 8.2.3 and suppose that the characteristic
function
(8.2.4) f(t) = exp {a(e*—1)+iut— 407 t?}
admits a decomposition
(8.2.5) f(t) =A@fr(?).
We see from theorem 8.1.2 and the fact that f(t) is an entire character-
istic function without zeros that f,(t) and f,(¢) are entire characteristic
functions without zeros. Therefore
file) = exp [hs(2)] (i= 1,2)
where the ¢,(z) are entire functions which are real for z = ty
(— 0 < y < o,y real) and have the property that 4,(0) = 0. We see
from corollary 2 to theorem 8.1.3 that
(8.2.6) $i (2) = Of/2| exp [|Im(z)|]+[e?} (|| > ).
Up to this point we have considered u, (t, y) = Re [¢,(z)] as a function
of the real variables ¢ and y. For the completion of the proof it is necessary
to fix y and to continue u, (¢, y) into the complex plane. It will be convenient
to introduce the function
(8.2.7) g(3) = ¢,(—22).
We note that g(z) is an entire function which is real for real z. Therefore
g(2) admits an expansion

&2) = L a2"
k=0
where the coefficients a, are real. Let z = t+7y (t, y real) then
Re [g(z)] = Re[g(¢+zy)] = 3[g¢+iy)+a(t-v)].
On the other hand, we see from (8.2.7) that
Re [g(z)] = Re [¢1(—72)] = Re [$1 (y—7t)] = u(y, —2)
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 247
so that

(8.2.8) um (y, —t) = d[g(tt+

iy) +e(t-9)].
The right-hand side of equation (8.2.8) is, for fixed y, an entire function of
t and can be continued into the complex plane. In the following we write
u,(y, w) if we consider also complex values of the second variable. Since
g(w+ty) = b1(y—iw)
g(w—ty) = $1(—-y—m),
we see from (8.2.8) and (8.2.6) that
(8.2.9) u(y, w) = O(|wleel+ ||?) (|w| > 0).
We now introduce the function
K(w) = u,(0, w) —u, (22, w).
Since u,(0, w) and uw, (2m, w) are entire functions, we see that K(w) is an
entire function. It follows from (8.1.11) and (8.1.18) that
(8.2.10a) K(w) = O(1) if Im(w)=0 and |w] > o,
while one sees from (8.2.9) that
‘(8.2.10b) K(w) = O(j/w|?) af Re(w)=0 and .|w] > o.
Moreover, one has for all w
(8.2.10c) K(w) = Ofexp (|w|*/”)] as |w| > oo.
We use the last three estimates to prove the following statement.

Lemma 8.2.3. The function K (w) reduces to a constant.

We consider the function
Hw) = K(e)(eo+ 1),
This function is analytic in the half-plane Re (w) > 0; in view of the
estimates (8.2.10a), (8.2.10b) and (8.2.10c), it satisfies the conditions) of
lemma 8.2.1 in each of the angles =, < arg (w) < OandO < arg(w) < A

We conclude from lemma 8.2.1 that for Re (w) > 0,

(3.2.11) K(@) = O(| |) as e0'| > 90.
We use the function
B,|w| = K(w)(w—1),
which is analytic in the half-plane Re (w) < 0, to show in the same way
that (8.2.11) is valid also for Re(w) < 0. Therefore the entire function
K (w) satisfies (8.2.11) for all w, so that it is necessarily a polynomial of
degree not exceeding 3. We conclude from the estimate (8.2.10a) that
K(w) is necessarily a constant.
(*) with p = 3/2, B—a = 7/2.
I
248 CHARACTERISTIC FUNCTIONS
We are now ready to complete the proof of theorem 8.2.3. It follows
from the definition of the function K(w) and from (8.2.8) that
—2K(—w) = g(w+2m1)+g(w—2at) —2¢(w).
Since, according to lemma 8.2.5, the function K(w) is a constant, we see
that g(z) satisfies the relation
o(z+2ni)+9(2—2n1)—2g(z) = c,
where c is a constant. We put

(8.2.12) —gi(z) = g(z)-g(2—2ni)

-
and see from the preceding equation that g, (z) is periodic with period 2z7.
Moreover we see from (8.2.6) that
£1(2) = Ofexp (3|2|/2)] as |z| > o.
The function g(z) satisfies the conditions of lemma 8.2.2 (with T = 221,
a = 3/2 = t), and applying it we get
£:(2) = Ap+4,¢+A,e~,
where A, A,, A, are constants. We see from (8.2.6) that g(z) = O(| 2 |°) as
Re (2) > — 00; the same is therefore true for g,(z), so that A, = 0 and
(8.213), og7(2)'= Agta,e:
In view of (8.2.12) and (8.2.13) we have
&(2)—g(2-—21) = Byo+ B 2+ Bie
with By, B, and B, constant. We put
BotiB, 2B,
82(2) = (2) Oni Ark aed
Repeating the reasoning which led to (8.2.13), we see that
§2(2) = Cot Cie,
where C, and C, are constants. Using the definition of g.(z) we conclude
that
(8.2.14) (2) = Dot Di z+ D.2?+D;,¢+D,2é
where the coefficients D; (j = 0, 1, 2, 3, 4) are constants. These constants
are real, since g(z) is real for real z. We see from (8.2.7) that (0) = 0;
therefore,

We put 2 = y+7t and separate the real and imaginary parts in (8.2.14) and
obtain
u(y, t) = Dyy+D,(y?—t?)+D,(e" cos t—1)+ Dye” (y cos tt sin 2).
It follows from the estimate (8.1.20b) that D, = 0. Therefore
u(y, 0)—u,(y, t) = D.t?+2D,e" sin? :
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 249

This expression must be non-negative for all real y and t. If we put t = x

and let y tend to + co we see that D; > 0, and if we put t = a and let y
tend to — oo we see that D, > 0. Therefore
g(z) = D,z+D,27+D;(e—1) (D, > 0, D; > 0).
If we write D, = ,, Dz = y = 402, Ds = A, we see that 4,(z) = g(—iz)
has the form
$1 (2) = ty 2-07 27/2+A, (e*—1)
so that the theorem is proved.
Some of the factorization theorems for analytic characteristic functions
admit interesting generalizations which we discuss in Chapter 9.
The results of Raikov’s theorem can also be extended in another direc-
tion. P. Lévy (1937b) and D. A. Raikov (1938) studied the multiplicative
structure of finite convolutions of Poisson type distributions and obtained
a number of interesting results. We now introduce certain notations which
will be used in formulating these results.
Let 2 > 0 be a real constant; we denote by

(2218) FER ee! Se)

x=o k!
the distribution function of the Poisson distribution with parameter A. We
write therefore F'[(x—y)/o; A] for distributions of the type of (8.2.15);
clearly the characteristic function of Poisson type distributions is f(t)
= exp [tyt+A(e""—1)] where 4 > 0, o > 0 and y are real numbers.
Let o, < o, < ... < o, ben positive numbers; wewrite A(oj, 02, .. ., On)
for the set of real numbers which can be represented in the form
£191
+ £202,+.--+£nOny
where the 9, 22,.-.,£n ate arbitrary non-negative integers such that
2it+2o+-.-+2n > 0.’ The set A(o,, o:,..., 0,) has no finite accumulation
point; it is therefore possible to arrange its elements in an increasing
sequence
ites ORIN Re gs An

We say that the m numbers 04, 2, ..., 0, are rationally independent if no

relation
(8.2.16) 110, +12,0,+..-+%nOn = 0
holds where the 71, 73, . . . , 7, are rational numbers such that
[7i[+|re|+...+|7n| > 0.
In other words a linear relation (8.2.16) with rational coefficients between
the o,, ..., 6, can only hold if all the coefficients are zero.
We now state Raikov’s results.
250 CHARACTERISTIC FUNCTIONS

Theorem 8.2.4. Let 1, Y2).++5n be n arbitrary real numbers, and let

Aqy Ay «+ + An be n non-negative numbers while o1, 62, ... On are n positive
numbers. The characteristic function of the distribution

F(x) = F(°ares 1) FE be 1) eer Oo

=e)
nr

then has only factors of the form

exp ee py on, (e"—1)]
where the A,, are the elements of Kee 2+) On) and where y and «4, are real
numbers.
We note that the factors of F(x) are not necessarily convolutions of
Poisson type distributions since the coefficient «,, may be negative. We will
give later an example of such a characteristic function.

Theorem 8.2.5. Let 61, 02,..+,Gn be n rationally independent positive

numbers. The distribution

F(x) = fale i)x ee in)ae Ape i.)

O71 02 On

then has only factors of the same form, namely

x—o x—0» x—6
F,
iy (x) = F (a sf m)* F (i 3 be Jen F( 5, ee a 1m )

where 06; 20 while0 < uw; < A;(j = 1,...,n).

Paul Lévy (1938b) has shown that theorems 8.2.3 and 8.2.4 are valid
even if the numbers oj, 02, ..., 0, have arbitrary signs.

Theorem 8.2.6. Let 04, 02,..., 6, be n positive numbers which satisfy the
condition
Onc O nes, 05 uly. PO
The distribution

F(x) = le i)* eee in)aie ae in)

O71 02 On

has only components of the same form, namely

x—0, x—Oy x—On
P(e) = F( 1s a)* F(*; ane aeZ 1)
2 n

where 0; = 0 while) < pw; < 4, = 1, 2,%..,

7).
For the proof of the last three theorems the reader is referred to the
paper by Raikov (1938).(+)
(t) We prove in the next chapter two theorems [theorems 9.4.2 and 9.4.4] which are
generalizations of Raikov’s theorems 8.2.6 and 8.2.5 respectively.
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS PSII

In his remarkable paper P. Lévy (1937b) studied convolutions of

Poisson type distributions of a somewhat more specialized character. He
considered real polynomials P(x) and the entire functions exp [P(x)]. If
the coefficients of P(«) are all non-negative, then the coefficients in the
Taylor series expansion of exp [P(x)] about the origin are also non-negative
so that M(t) = exp [P(e’)— P(1)] is a moment generating function. How-
ever, M(t) can be a generating function even if P(«) has negative co-
efficients. P. Lévy derived a necessary and sufficient condition which the
polynomial P(x) must satisfy in order that exp [P(x)—P(1)] should be
a probability generating function. The generating functions studied by
P. Lévy belong to distributions of the form

R(=34,] *F, (254) oo, *F,(£52,]

O71 0% On

where the positive numbers oj, 03, ... , G, are all integers.
We conclude this section with the discussion of an example which was
studied by D. A. Raikov as well as by P. Lévy.
We consider the polynomial
(8.2.17) . P(x) = 1+ax—fx?+
cx*+ dxt
of degree four. The numbers a, f, c,d are assumed to be positive. We
compute [P(x)]? and [P(x)]* and see that it is possible to determine the
coefficients a, f, c,d in such a way that [P(x)]? and [P(x)]® have no
negative coefficients.(t) Then [P(x)]* = {[P(x)]?}* and [P(x)]*+1 =
[P («)]}*{[P (x)]?}*-1 also have non-negative coefficients for k > 1. We
form
o [Pp j
(8.2.18) exp [P(x)|= >, aie
=0 t

and see that under these conditions only the quadratic term in exp [P(x)]
can have a negative coefficient. The coefficient of the quadratic term of
exp [P(x)] is easily determined; it is [(a?/2)—f]e. If we suppose that
2

(8219) pet
then exp[P(x)] has only non-negative coefficients. The function
exp [P(x)—P(1)] is then a generating function, so that
(3.2.20) p(t) = exp {ae*—Be™+ce™+ de” —a—c—d+p}
is a characteristic function. The function (8.2.20) cannot be an infinitely
divisible characteristic function. To show this we note that the coefh-
cients of the linear and of the quadratic term of the polynomial in the
exponent of [g(t)]'/” are a/n and f/n respectively. The condition corres-
ponding to (8.2.19) will therefore be violated for large m so that [g(t)]”"
(t) These conditions are for instance satisfied if a= ¢ = d and f = }.
Z5z CHARACTERISTIC FUNCTIONS

cannot be a characteristic function if m is chosen sufficiently large. The

function g(¢) can be used to construct an interesting decomposition. ‘The
characteristic function of the distribution
EF(x)= F (£34)°F (4230) F (4x, a)
is then
f(t) = exp {ac#+ce"+de—a—c—d}.
This convolution of three Poisson type distributions also admits the
factorization
f(t) = g(t) exp [B(e"—1)].
We conclude from theorem 6.2.2 that g(¢) must have an indecomposable
factor and see that a convolution of three Poisson type distributions can
have indecomposable factors. Since every factor of g(t) is also a factor of
f(t) we conclude from theorem 8.2.4 and from a result of P. Lévy(f) that
there exist indecomposable characteristic functions of the form (8.2.20).
P. Lévy (1937b) considered polynomials of the form
Py (%) = @x-+Bxr—yx* a bx* cx
and
P(x) = a'x —Bx*+yx3+b’x4
and showed that it is possible to determine the coefficients in such a way
that
f(t) = exp {P,(e")—P,(1)}
and
a(t) = exp {P2(e")—P(1)}
are both indecomposable characteristic functions. Therefore f(t) =
fi (2) fo(2) is the characteristic function of a convolution of three Poisson
type distributions and provides an example of the factorization of an
infinitely divisible characteristic function into two indecomposable factors.
We conclude this section by listing several theorems which indicate that
certain functions can be characteristic functions, provided a parameter is
suitably chosen. These theorems are somewhat similar to theorem 6.2.3
since they can also be used to prove the existence of infinitely divisible
characteristic functions having an indecomposable factor.
Theorem 8.2.7. Let « = p/q be a rational number and suppose that the
integers p and q are relatively prime and that 1 < p < q. For given positive
numbers i, A, and y it 1s possible to select a sufficiently small positive number v
so that
fi(t) = exp {—yt? +, (e*— 1) +A, (6-1) —r(e#/4—-1)}
is a characteristic function.
(t) P. Lévy (1937b) has shown that a function of the form exp [P(«)—P(1)] (P(x) a
polynomial) cannot be a generating function unless a term with negative coefficient is
preceded by one term and followed by at least two terms with positive coefficients.
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 253

Theorem 8.2.8. Let « be an irrational number, 0 < « < 1. For given positive
numbers A,, A, and y it is possible to select sufficiently small positive numbers v
and » so that the function
fa(t) = exp {—yt? +, (e*—1) +4, (1) —9(e—1)}
is a characteristic function.

Theorem 8.2.9. Let G(u) be a function which is continuous and non-decreasing

in the interval [b,, b,| and suppose that G(b,)—G(b,) > 0 and let y be a
positive constant. Then it is possible to select sufficiently small positive numbers
y and 7 so that the function
ba
fa(®) = exp {—yt+ |b*(e—1)dG(u)—(e"—1)}
1

ts a characteristic function.
Theorems 8.2.7, 8.2.8 and 8.2.9 are due to Yu. V. Linnik; for their proof
we refer the reader to Chapter 8 of Linnik (1964).
Remark. We see from the Remark 1 following theorem 5.5.1 that the
characteristic functions f, (¢), f,(¢) and f;(¢) are not infinitely divisible and
therefore have indecomposable factors.
We mention here another open problem of the arithmetic of distribution
functions. It is known that many infinitely divisible characteristic functions
have indecomposable factors ;however, it is only possible to determine these
factors in a few cases. It would be interesting to study methods which
would permit the determination of indecomposable factors of infinitely
divisible characteristic functions.

8.3. Determination of certain entire characteristic functions

by properties of their factors
In studying factorizations we disregard the trivial degenerate factors. It
is therefore convenient to introduce the following terminology. We say that
two characteristic functions f, (¢) and f,(¢) are equivalent and write
fit) ~ fat)
if f,(t) = e f,(t) where a is a real number. Similarly we say that the
second characteristics ¢, (t) = log f, (¢) and ¢,(t) = log f,(¢) are equivalent
(in symbols 4; (¢) ~ ¢2(#)) if
$,(t) = ait+ $.(?).
With this notation we can express the fact that two characteristic func-
tions f,(t) and f,(t) belong to distributions of the same type(t) by stating
that there exists a constant o > 0 such that f,(t) ~ fo (t/o).
In this section we show that certain entire characteristic functions can be

(+) For the sake of brevity we will say that f,(t) and f,(t) are of the same type.
254 CHARACTERISTIC FUNCTIONS

characterized by properties of their factors. We derive first a theorem

which is the converse of Cramér’s theorem.

Theorem 8.3.1. Let f(t) be a decomposable characteristic function and

suppose that all factors of f(t) are of the type off(t). Then f(t) 1s the charac-
teristic function of a normal distribution.
We prove first that f(z) is infinitely divisible. Let
f(t) = AOA)
be a decomposition of f(z). It follows then from the assumptions of the
theorem that there exist two positive constants c, and c, such that
(8.3.1) f(t) ~ flat) f(C22).
We apply the same decomposition to each factor on the right-hand side of
(8.3.1) and see that
F(t) ~ FOF (Cr Cat) fF(Cr Cot) f(t)
so that [f(c,c.t)]? is a factor of f(t). According to the assumption of the
theorem there exists then a positive constant c; such that
[f(cr¢at)]? ~ f (cst)
or

fe ~[4(222)]-
€1 C2 :

But this means that f(t) is the square of a characteristic function. We

repeat this argument to show that for any positive integer k, f(t) is the
2*th power of some characteristic function. It follows (see the remark
following the corollary to theorem 5.3.3) that f(¢) is infinitely divisible.
Therefore f(t) has the canonical representation
itx eee
log f(t) ~ |* (e-1- eres
(8.3.2) re d6(x).

We show next, by means of an indirect proof, that the only point of in-
crease of 6(x) is the point x = 0. Let us therefore assume tentatively that
x = a # (isa point of increase of 6(x). Select ¢ > 0 so that |a| > «, then
the three numbers a, a—« and a+<« have the same sign. We introduce the
function
0 ifx < a—é
H, (x) = {te -H0— ifa—-e<x<ate
3[0(a+¢e)—O(a—e)] ifate < x.
The functions H,(«) and 6(«)—H,(«) are then both bounded and non-
decreasing functions; therefore

h,(t) ~ exp fo (#-1- 1+?

ae ) anc]
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 255

is a factor of f(t); hence there exists a c > 0 such that

htt) ~ #(*)
or

fy G1 --he a) te dH, (x)

yD t

1+ x?
a eae itx/c\ 1+%x?
nG vr ne)ye Me):
By a simple transformation of the integral on the right-hand side of this
relation we see that

7 ail dteeciate pees 1+c?y?

f-(¢ 1 a) y2 cB(1 +2) di(cy).

It follows that

H,(x) = C+ fisae d0(cy)

£ ] 24,2

where C is a constant. The function H,(«) increases only in the interval

[a—e, a+e], therefore 6(x) grows only in [c(a—e), c(a+e)]. Since a is a
point of increase of 6(x), it must lie in this interval, ice.,
c(a—e) < a < c(ate)
so that
a a 5
Pg a iii
ate —Eé
(8.3.3)
a a :
— <_¢e— ifa< 0.
a-eé aA+€é

We see from (8.3.3) that c tends to 1 as e + 0); at the same time the interval
which contains all the points of increase of 6(x) shrinks to the point x = a;
hence x = a is the only point of increase of 6(x), and 6(x) has the form
O(x) = de(x—a) (A> 0).
Then
log f(t) ~ re (4-1)

and it is easy to show that f(t) cannot ate a proper decomposition (8.3.1).
This shows that a ~ 0 leads to a contradiction with the assumptions of our
theorem, so that a = 0 is necessarily the only point of increase of 6(x). This
means that f(t) is the characteristic function of a normal distribution, so that
theorem 8.3.1 is proved.
256 CHARACTERISTIC FUNCTIONS

It follows from Cramér’s theorem (theorem 8.2.1) that a normal charac-

teristic function has only factors of its own type. Theorem 8.3.1 is therefore
the converse of Cramér’s theorem and we obtain immediately the following
characterization of the normal distribution.
Corollary to theorem 8.3.1. The decomposable characteristic function f(t) is
the characteristic function of a normal distribution if, and only if, all factors of
I (t) are of the type of f(t).
Our next theorem gives a common property of normal distributions,
Poisson type distributions and their conjugates.
Theorem 8.3.2. Suppose that the characteristic function f(t) has an infinite
set of non-equivalent factors and assume that f(t) has the following property:
iff(t) and f,(t) are any two factors of f(t), then either f,(t) ts a factor of
f2(t) or f,(t) is a factor of f,(t). Then f(t) 1s the characteristic function of
either the normal distribution or of a Poisson-type distribution or of the con-
jugate to a Poisson-type distribution.
For the proof of theorem 8.3.2 we need the following lemma.
Lemma 8.3.1. If a characteristic function f(t) is divisible by an arbitrary
integer power of a characteristic function g(t), then g(t) belongs necessarily to
a degenerate distribution.
If the conditions of the lemma are satisfied, then
F(t) = [g@}"An(t) (a = 1,2,..-)
where h,, (t) is some characteristic function. Therefore
(8.3.4) If] = le@I"lam@] < le" @ = 1,2,.-.).
We now show that the assumption that g(t) is non-degenerate leads to a
contradiction. It follows from the corollary to lemma 6.1.1 that there exists
a 6 > Osuch that |g(t)| < 1 for 0 < t < 6. We choose such a ¢ and let x
tend to infinity in (8.3.4) and see that |f(t) |can be made arbitrarily small,
provided 0 < ¢ < 6. This contradicts the fact that f(t), as a characteristic
function, is continuous at f = 0 and f(0) = 1, so that the lemma is proved.
We proceed to the proof of theorem 8.3.2 and show first that the
characteristic function f(t) has no indecomposable factors. We give an
indirect proof and assume therefore tentatively that f,(¢) is an indecom-
posable factor of f(¢). According to the assumptions of the theorem, every
other factor g(t) of f(z) is divisible by some power of f,(¢). We see then
from lemma 8.3.1 that there exists a highest power of f, (¢) which is a factor
of g(t). Let n be the exponent of this power, so that g(t) = [f,(t)]"A(2).
The factor h(t) is not divisible by f,(¢); it follows from the assumptions of
the theorem that A(t) must be a factor of f, (#), but since f; (¢) is indecom-
posable, h(t) is necessarily degenerate, so that
(8.3.5) g(¢) ~ [A@I".
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 257
If the characteristic function is non-degenerate it can, according to lemma
8.3.1, not be divisible by arbitrarily large powers of f,(t). On the other
hand we see from (8.3.5) that every factor of f(t) is equivalent to some
power of f, (#), so that f(t) can have only a finite number of non-equivalent
factors. This contradicts the assumption of the theorem, so we must con-
clude that f(t) has no indecomposable factors. According to theorem 6.2.2
f (é) is then infinitely divisible. Therefore we can write f(t) in the canonical
form
1tx )1+ x?
(8.3.6) f(t) ~ exp (1, 5" 02)|.

We show next—again by means of an indirect proof—that 6(x) has only

a single point of increase. Let us therefore assume tentatively that this is
not true and that 6(x) has the points a, and a, as points of increase. We
select e > 0 so that a,+¢ < a,—e and construct two functions
0 if x < aj—e
(8.3.7) H;(«) = <6(«)—6(a;—«) ifa;—e < x < ajte
O(a;+e)—O(a;—«) ifajte <x
for j = 1, 2. The functions 6(«)—H;(«) (j = 1,2) are non-decreasing
and bounded, so that the functions

Ate) = exp [["(4-1-2

14x? ,)
ay] = 1.)
are characteristic functions. Moreover f,(t) as well as f,(t) are factors of
f(t). We conclude then from the conditions of the theorem that one of
these factors must divide the other. If f,(¢) would be a factor of f,(t), then

(8.3.8) i5 = exp {|u (a ia =a) a ealr, (Nadi. cn}

would be a factor of f(¢) and therefore an infinitely divisible characteristic
function. However we see from (8.3.7) that H,(«)—H,(x) is not mono-
tone, so that the expression (8.3.8) cannot represent an infinitely divisible
characteristic function. Therefore f,(¢) cannot be a factor of f,(¢). In the
same way we can rule out the possibility that f, (¢) is a factor of f(t) and
therefore obtain a contradiction with the assumptions of the theorem. ‘This
contradiction shows that 0(x) has exactly one point of increase. Let x = a
be this point. If a = 0, we see from (8.3.6) that f(t) is the characteristic
function of a normal distribution; if a > 0 then f(t) is the characteristic
function of a Poisson-type distribution; if a < 0 then f(¢) is the conjugate
of a Poisson-type characteristic function.
This completes the proof of theorem 8.3.2. Its converse is trivial, so that
it can provide a characterization of the family of all distributions which
belong to the type of the normal, the Poisson or the conjugate Poisson
distribution.
258 CHARACTERISTIC FUNCTIONS

We finally mention a result due to I. A. Ibragimov (1956b) which gives

a characterization of the normal distributions. He considered the class &
of infinitely divisible distribution functions (x) which have the following
property: if F(x) eG and if the convolution F * H = Q is infinitely
divisible then H is infinitely divisible. Ibragimov showed that the class &
coincides with the family of all normal distributions.

8.4 Infinitely divisible analytic characteristic functions

In this section we discuss analytic characteristic functions which are
infinitely divisible. We have seen earlier that an infinitely divisible charac-
teristic function does not vanish for real values of its argument and we now
extend this remark.
Theorem 8.4.1. Let f(z) be an analytic characteristic function and suppose
that it ts infinitely divisible. Then f(z) has no zeros in the interior of tts strip of
regularity.
Since f(z) is an infinitely divisible characteristic function, [f(¢)]' is a
characteristic function for any positive integer m and is also a factor of
f(t). According to theorem 8.1.1 the function [f(z)]’” is an analytic
characteristic function which is regular at least in the strip of regularity of
f(z). If f(z) should have a zero at some point 2, inside this strip, then
[f(z)]’" would have a singularity at the point z, for sufficiently large n,
which is impossible.
The statement of theorem 8.4.1 cannot be improved. This is shown by
constructing an analytic characteristic function of an infinitely divisible
distribution which has zeros on the boundary of its strip of regularity.
Let a > 0, b > 0 be two real numbers and put w = a+7b. It is easy to
show that

t) =
(1 —it/w)(1 —7t/w)
FO) (1 —1t/a)?
is an analytic characteristic function which is regular in the half-plane
Im (z) > —a and which has two zeros —iw and —iw on the boundary of
this region. Moreover it admits the representation
Cale tu
log f(t)
t) = mit |me(“a1 Rae
= mit+ 2 )dN
(u)

where

ee 2 | e- (1 — cos bx)(1
4+x2) dx
0

and
N(u) = -2 | e~“(1—cos bt)t-1 dt.
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 259
According to P. Lévy’s representation theorem (theorem 5.5.2) f(t) is
infinitely divisible and therefore provides the desired example.

Corollary 1 to theorem 8.4.1. An infinitely divisible entire characteristic

function has no zeros.
P. Lévy (1938a) raised the question whether an entire characteristic
function without zeros is infinitely divisible and solved it [P. Lévy (1937c)]
by constructing an example of an entire characteristic function without
zeros which is not infinitely divisible. The characteristic function (8.2.20)
is such an example. Moreover our argument in Section 8.2 indicates that it
is possible to determine the coefficients in such a manner that (8.2.20)
represents an entire and indecomposable characteristic function without
zeros.

Corollary 2 to theorem 8.4.1. The characteristic function of a finite distri-

bution cannot be infinitely divisible.
The corollary follows immediately from theorem 7.2.3 and corollary 1
to theorem 8.4.1.

Corollary 3 to theorem 8.4.1. The characteristic function f(t) of a finite

distribution 1s always the product of a finite or denumerable number of in-
decomposable factors.
From corollary 2 and from theorem 6.2.2 we conclude that f(¢) must
have indecomposable factors. It is also easily seen that it can have no
infinitely divisible factors since all its non-degenerate factors are entire
functions of order 1 and cannot therefore be infinitely divisible.
The preceding theorem and its corollaries can be regarded as necessary
conditions which an analytic characteristic function must satisfy in order
to be infinitely divisible. For instance it follows from theorem 8.4.1 that the
characteristic function determined by formula (5.5.12) is not infinitely
divisible. We used this fact—without proving it—in an example discussed
in Section 5.5. We now give another application of the theorems discussed
in the present section.
Let f(z) be an infinitely divisible analytic characteristic function which
has the strip of regularity —« < Im(z) < £. Then f(¢) has no factor of
the form g(t) = pe +(1—p)e” for which
4, < eprlee(l=2) — 5
Seip
It is easily seen that the existence of such a factor would produce a contra-
diction with theorem 8.4.1.
We have already remarked that the second characteristic ¢(t) = In f(¢)
is defined for every characteristic function in a (real) neighbourhood of the
200 CHARACTERISTIC FUNCTIONS

origin, Let f(s) be an analytic characteristic function which has the strip
—a< Im (s:) < 6 as its strip of regularity and suppose that f(s) has a
verogy in this strip, Then we can only state that 4(s) is defined and regular
in the interior of the circle whose radius is min (a, 8, |89 !). If f(s) has no
zeros in the interior of the strip, then 4(s) can be continued analytically in
the strip, 'This is the case iff(s) is an infinitely divisible analytic character-
istic function. In this case we obtain the following important result.
Theorem 84.2. Let f(s) be an analytic characteristic function and suppose
that it is infinitely divisible, Then the canonical representation

(84.1) log f(s) = text | (e*— 1 — sw)aX — (w)

v=® wu
ts valid in the interior of the strip of regularity off(s). Here ¢ ts a real constant
and K(u) ts @ non-decreasing bounded function such that K(— co) = 0 and
i aK (u) = K(+ 0) < ©,
We assume therefore that f(s) is an infinitely divisible analytic charac-
teristic function with —« < Im(s) < 9 as its strip of regularity. Then
the second moment of f(s) exists and f(s) admits, for real values of the
argument, the (Kolmogorov) representation (S.1). We see from the
remarks preceding the statement of the theorem that log f(s) is defined in
the strip of regularity, We denote by 4(s) = log ASs) that branch of the
function log #(s) which, for reals, is given by (S.4.1), The function ¢(s) is
regular in the strip —x < Im(s) < & Now let ¢ be real; we can apply
(8.4.1) and see that
A(t) =tim SETA SEA) 28)
R~D) he

or
— 1 —cos hw
"(?) = —lim ee . —~&K (w
$ “ moa |= & hh? a* 2 : ( )

so that

s"()=- |
eee th

&*aK(w). -

ve = ®

Since A(w) is bounded and non-decreasing we see that ¢"'(é)is, except

for a constant factor, a characteristic function. Moreover, it follows from
the analyticity of ¢(s) that 6"'(s) is, except for a constant factor, an analytic
characteristic function which is regular in the strip—« < Im(s} < 3.
The integral
hic
(842) -s"s)= | aK (w)
wv = @Q
FACTORIZATION OF ANALYTIC CHARACTERISTIC FUNCTIONS 261

converges in this strip and is a regular function of z. Let ¢ be a complex

number such that —a« < Im(¢) < f and select «’ > 0, 8’ > 0 so that
—a < —a’ < Im(C) < f’ < B. Then the integrals

| e®“dK(u) and | e*'“ dK (u)

both exist and are finite. From this it follows easily that (8.4.2) can be
integrated under the integral sign from 0 to ¢ and we obtain

(843) -4')+4'0) = |
ro) ete
= dK(u).
—2 1u

By a similar argument one can show that it is permissible to integrate

the expression on the right-hand side of (8.4.3) under the integral sign
from 0 to z, provided that —« < Im(z) < f. In this manner one obtains
: ae endkiu
d(z) = [¢'(0)]z+ hoe (e™— 1 —72zu) a

Since ¢’(0) = ix,, where x,—the first cumulant—is real, we see that
(8.4.1) holds in the entire strip of regularity of f(z), so that theorem 8.4.2 is
established.
 9 INFINITELY DIVISIBLE CHARACTERISTIC
FUNCTIONS WITHOUT INDECOMPOSABLE
FACTORS
Khinchine’s theorem (theorem 6.2.2) indicates that a characteristic func-
tion which has no indecomposable factors is always infinitely divisible. We
have seen from theorem 6.2.3 that the converse is not true; i.e. an infinitely
divisible characteristic function can have indecomposable factors. This
situation suggests the investigation of the family of infinitely divisible
characteristic functions which do not admit indecomposable factors. This
family is usually denoted by J, and one of the most important problems of
the arithmetic of distribution functions is the study of the class J). The
present chapter deals with this topic.
It follows from theorems 8.2.1, 8.2.2 and 8.2.3 that the class J, contains
the Normal distribution,®) the Poisson distribution, and the convolution
of a Normal and a Poisson distribution. Theorems 8.2.5 and 8.2.6 provide
other examples of members of Jy, while an example given on page 251
shows that not all convolutions of Poisson-type distribution functions
belong to J).
A systematic study of the class J) was carried out by Yu. V. Linnik; his
work was first published in a series of papers in the Russian probability
journal and later presented in a monograph [Linnik (1964)]. These investi-
gations were continued by other authors.

9.1 The class #

We first introduce some terms which are convenient in the discussion of
these studies.
We have shown that every infinitely divisible characteristic function
f(t) can be written in the canonical form
(9:11)
logog f(t) = ita—yt?
f(t) ita—yt? + ie G
Ane 1 Ber
ae )dM (u)

+ [alta F are
. i eee (i

where a is real, y > 0, and where the functions M(u) and N(u) satisfy the
itu )

conditions listed in theorem 5.5.2.

(*) We have defined the class Jy) as a family of characteristic functions. We will also
speak of distribution functions belonging to I), meaning that the corresponding charac-
teristic function is in I.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 263

We shall call M(u) and N (u) the spectral functions of f (2), or of the
corresponding distribution function F'(«); more specifically, we refer to
M(u) [respectively N(u)] as the spectral function of the negative [res-
pectively positive] spectrum. The negative [respectively positive] Poisson
spectrum of the infinitely divisible characteristic function f(t), or of its
distribution function F(x), is the set of all points of increase of the function
M (u) [respectively N (u)]. We write Su and Sy for the negative and positive
Poisson spectrum respectively. We call the set Sy U Sy simply the Pois-
son spectrum of f(z) [or F(«)]. An infinitely divisible characteristic function
is said to have a bounded negative [positive] Poisson spectrum if there
exists a number d [respectively 5] such that

i= dM (u) = 0 UbeaN (ay = 0}.

The Poisson spectrum of f(t) is bounded if the positive as well as the nega-
tive Poisson spectrum is bounded. An infinitely divisible characteristic
function f(t) is said to have a finite spectrum if

(Gie2 slog 7) = ab yt 1) 2, A (e — 1);

yy A, (e"*—
j=1 j=1
where m and m are non-negative integers and y > 0, A; > 0, A_, > 0,
bt; > 0, »; > O. If either m or n is equal to zero then the corresponding
sum is omitted.
The infinitely divisible characteristic function f(t) is said to have a
denumerable Poisson spectrum if
oo it F
(9.1.3) log f(t) = ait? + ¥ a,(om—1- a )
j=1 1+4;
© Loe ity
; 1
ity;I )

+B dale ee

where A; > 0, A_; > 0, y = O and where the series

ey A; pe 2 A_; y
a 7 and
aca Maat See :
converge so that
(9.1.3b) dD AwiF+ D A”
My<é vy<e

tends to zero as ¢ approaches zero. The numbers »; and yu; are called the
Poisson frequencies of f(¢). The A; and 4_; are called the energy parameters
of the frequencies yz; and v; respectively. sf
We also introduce a class\*) Y of infinitely divisible characteristic
functions which has the following properties:
(i) The Poisson spectrum of a characteristic function fe # is
(*) This class should not be confused with the class of self-decomposable characteristic
functions (L-class) treated in Section 5.11.
264 CHARACTERISTIC FUNCTIONS

either finite or denumerable. Therefore f(t) admits the repre-

sentation
(9.1.4) ;
log f(t) =4at—yt? aS,3
+>) Ag ee— tomy| t
r=1m=— © ; 1 + in

where
ais teal, y.2 0,7,,,.2 00 — 152,77 = 0, ey Pc.e
Bina > 0; Umo< 9.
(ii) 3 y Age fine bine) “1
= < 0.
r=1 m=— 0
(iii) DS An, M2, tends to zero as e > 0),
|umr|<e

(iv) The quotients pijss5/fecr = 1, 230 = 0, 4 1) £2). % 3) are

natural numbers greater than 1.
It follows that
«0 Miia S Moa s Mays >
-+Mii2 > Moo > Mig > >>>
and
lim jim = 90. lim) fips = +0
m—>— 00 m—>-+oo
lim fyo=9, lim py. = — 0.
m—>— 0 m—-+co

In this section we derive the following property of characteristic func-

tions of the class #.
Theorem 9.1.1. Let f(t) € & and suppose that the energy parameters of f (t)
satisfy for some k > 0 the condition
(9.1.5) Ame = Olexp.<— kuz, a(S +o, ra 2).
then ¢(z) = log f(z) ts an entire function, so that the characteristic function
J (2) is an entire function without zeros. Moreover,
$(2z) = Of] 2]? exp [M(Im (2))"]} (as |2] > 0)
where N > 0 is a constant.
For the proof of theorem 9.1.1 we need several lemmas concerning
analytic functions. We now state these lemmas, but since their proofs are
not easily accessible in the literature we give them in Appendix E (the
motivation for this separation of the statements and of their proofs is our
wish to avoid disrupting the discussion of the theorems concerning the
characteristic functions of the class I).
Lemma 9.1.1. Suppose that the function N (u) 1s non-decreasing in the half
open interval 0 < u < a(a< ©) and that i u*dN(u) < oo. Then the
+0
integral

RO at Ce Maresa
(*) Condition (ii1) is a consequence of (ii).
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 265

converges absolutely and uniformly on any bounded set of the z-plane, so that
f(&) ts an entire function. Moreover the estimate
f(8) = O{|2|?(1 + exp [a Re (z)])}
holds as |x| > o.
Lemma 9.1.2A. Suppose that the function f(z) admits the representation

fle) = 3 dy exp (2?»)

p=0
2
dh

where the coefficients d, satisfy, for some k > 0, the relation

d, = Ofexp (—p*)] (p > ©)
while T > 0. Then f(z) ts an entire periodic function with period iT and
O{[Re (2)? if Re (z) > 0
(9.1.6) log
| f(z)| = { {[Re (2)]} : «)
O(1) if Re (z) < 0.
Lemma 9.1.2B. Suppose that the entire function f(z) is periodic with purely
imaginary period 1T and that
O{[Re (z)]?+log |z|} if Re(z) > 0
O(log | |) if Re (z) < 0.
Then f(z) admits the expansion

f(z) = Bd,
as
exp (Ps
p=0
2p
TE:
where the coefficients d, satisfy, for some k > 0, the relation
d,= Ofexp (—kp?)] asp > o.
Remark. 'heestimate (9.1.7) follows from (9.1.6). It would therefore be
possible to use (9.1.7) as a necessary and sufficient condition for the series
representation of f(z). In view of the later application of the lemmas it is
more convenient to present the necessary and the sufficient condition
separately.
We now proceed to the proof of theorem 9.1.1 and suppose that f(f) is
a characteristic function of the class ¥ which satisfies the condition
(9.1.5). Then 4(t) = log f(t) can be written in the form (9.1.4); we con-
sider first the positive Poisson spectrum of f(t) and write
= i Tn i t
Ss Ag (crm 1 — - )

Pa Dit fina
e tlaat = = tims t
oe An i )>, Neg hi
Bi DS An (1
te ma )
a ‘ (« 1 =i Has m=1 fine m=1 % 1 aie te

= $,+S,+S3 (say).
0
Let N(u) = — DS Anae(u—Mn»)- We can then write the first sum S, as
266 CHARACTERISTIC FUNCTIONS

an integral
She |+0 (om—1- 14+u?
sti )avo.
Hot ,

This integral has (if we put 7 = z) the form of the integral in lemma 9.1.1.
We see from (9.1.3b) that the conditions of the lemma are satisfied, and we
conclude that S, is an entire function. We can therefore consider Sj also
for complex values of the variable ¢ and obtain from lemma 9.1.1 the
estimate
Sy = O{|t? (1+ exp [o, Im (2)])}
as |t| > oo. Since f(t) e #, the quotients w,,,u;7 are integers greater
than 1, and we can write

S, = x d, exp (ty pt)

pe
where
de te ne € (Mma Mir Sm=1
Ama 1 P = bmabaa-
We see from (9.1.5) that for p = mw, 1/u,, the coefficient
dy = dn = Olexp (—kuna)] = O(— kets p*)-
We apply lemma 9.1.2A and we see that S, is an entire function of the
complex variable ¢ and that
S, = Ofexp [N(Im (#))?]} as |¢| > 00
for some N > 0. The third sum S; is a linear function.
We treat the negative Poisson spectrum in the same way and obtain the
estimate for ¢(z) stated in theorem 9.1.1.

9.2 A sufficient condition for membership of J,

The problem of characterizing the class J, has not been solved completely
at present. We only have some necessary and some sufficient conditions for
membership of this class. In this section we prove a sufficient condition;
a necessary condition will be given in Section 9.3.
Theorem 9.2.1. Let f(t) be a characteristic function of the class & whose
energy parameters and frequencies satisfy the following conditions:
(9.2.1) Ag = Ofexp [—A(u2,,)]] (m > +003 r = 1,2)
where k is a positive constant. Then f(t) € Io, that is f (t) has no indecomposable
factor.
Condition (9.2.1) is identical with (9.1.5) used in the proof of theorem
oa.
We now assume that the characteristic function f() can be factored
f) = AOA().
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 267
It follows from theorem 9.1.1 that f(t) is an entire characteristic function
without zeros, and we see from this fact and from theorem 8.1.2 that
fi (t) and f,(t) are entire characteristic functions without zeros.
We can therefore write f(t) = e*, f;(t) = e* (j = 1, 2), where ¢(2),
$1 (t) and ¢,(t) are entire functions; these functions can be continued
into the complex z-plane, x = t+iy. We introduce

&(2) = $1(-i8) = 61(y—it)

and
u(t, y) = Re [g(z)].
Then
|fi(—iz)| = |f(y) |= et
and we see from the ridge property of analytic characteristic functions
(p. 195) that
1<fAl-) . f(-)
lfi(y—2#)| — |f(y-2)|
or
(9.2.2) 0 < u(t,0)—u(t, vy) < ¢(—it)—Re [d(y—it)].
Applying the estimate of theorem 9.1.1, we conclude that
(9.2.3) 0 < u(t, 0)—u(t, y) = O[]2|? exp (NE2)] as |z| > oo.
We next derive two similar estimates. We see from (9.1.2) that

Mit) ~ Re[My—i] = 1242 ES

2

r=1
coo)

m=—-o
dno sina) pt 2

We substitute for y = 2 w,,' and obtain (using property (iv) given in the
definition of the #-class) that
s=1 2

(9.2.4) $(—#)—Re [4(2nuz'—-it)] =2 DY Aysetmt (sinoe :) Ms

2 7\2 27 \2
+2 > Thies eltmaat (sinPm )+y (=)
L fsa

We note that
2 s—2 ‘A
s—2 IU
= o(eMt)
. 7t

eae (sin =) Soe A, a)

bsa
aa Ms

as t > 00 and

: 4 m,2
>> Awe eeeat (sinPm2t < > Ang =
DIS CD Msa [Hamel <M

eel eetne = OL) ast oo.

lM mval > Meer
268 CHARACTERISTIC FUNCTIONS

We conclude from (9.2.2), (9.2.4) and the last two estimates that
(2.2)

0 < u(t, 0)—H(t, 248) < (Bay +0(0) (sin=) exp yaa
2

8,1

asot >.00(s = Oj lye 2;’..02)-

In the same way we obtain the estimate
(9.2.6) 0 < u(t, 0)—u(t, 2mp;,')

< (24,-.2+0(1)) (sin=)" exp (4, 122) (> - 2)

2

8,2

We next derive an estimate for f(z). It follows from theorem 9.1.1 and
from the definition of g(z) that
(9.2.7) $(—iz) = Off |? exp [N(Re (2))*]}.
Moreover we see from theorem (8.1.3) that
(9.2.8) M(r;g) < 6rM(r+1, ¢)+O(r?)
and obtain the following result:

Lemma 9.2.1. For all complex z (z = t+iy; t,y real), the estimate
g(z) = Of{|
2| exp [NV (Re (z))?]} (|2| > 0) holds. Here N is a positive
constant.
Let g be an integer (positive, negative or zero) and put
9.2.9) Sae(2) = g(2) eXP (—Mye2) (r= 1,2),
and write
tyr(t, ¥) = Re[ga-(¢+2y)] (¢, y real).
We shall need estimates for the expressions u,,(t, 0) —u,,,(t, 27,1). One
has
uit 0) Us, cut) =

= g(t) exp (—p,,t)—Re {g(t + 2niu;,") exp [— my, (t+ 2niuz,)]}

= {u(t, 0) a u(t, 2m5) } exp (~ Har t).
We apply the estimate (9.2.5) in the case where y = 1 [respectively (9.2.6)
for r = 2] and conclude that
(9.2.7a) u,1(t, 0)—u,1(t, 2muz) = O(1) ast >+0
(D277) Uys(t, 0)— Ugo (t, 2u,o) = O(1) as t > — ow.
Lemma 9.2.2. The functions g,,(2) (yr = 1, 2) can be written as sums
(9.2.10) Bar (®) = Ser’ (2)+8a9(2) (r = 1, 2).
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 269

The summands g'*) and g\,) ave entire functions which are real for real 2 and
which admit the estimates (as |z| > 00)
H(z) = ae exp [N(Re(z))*]} #f Re(z) > 0
pie ten if Re(2) < 0
far (2)=
O(\z |) ifRe (2) > 0
O{|z | exp [N(Re(2))"]} if Re(z) < 0.
Here N ts a positive constant.
It follows from the definition of the function g,,(z) that it is real for real
= and that
(9.2.11) gy,(2) = Of] 2? exp [N(Re (z))*]}.
Let 0 < a < bandH > 0 and consider the rectangle which has the points
a+1H, b+iH, b—iH and a—1H as vertices. We integrate the function
2° Sar (S)
2nil® (€ — 2)
along the contour of this rectangle. According to Cauchy’s theorem, this
integral is equal to g,,(z) if z is inside the rectangle, but equals zero if z
is outside the rectangle. It follows from the estimate (9.2.11) that the
integrals along the horizontal sides of the rectangle tend to zero as H tends
to 00, while a and 8 are fixed. Therefore
(9.2.12)
Bor (2)
2 a ee if a < Re(z) <b
Qni Joico C8(C—2) 2ni Ja-iw C3(E—z) ~~ )0 if Re(z) < a
oreehke (2) 2.0
We consider the function g‘*)(z) defined by
1+i0 Pe Clas
for Re (z) < 1.
Bar’ (2)= Ai, eek (Ga)
We see that the value of the integral defining g‘*) does not change if the
limits 1—zoo and 1+70 are replaced by b—ico and b+100 respectively,
where 6 > 1. This means that it is possible to extend g‘*)(z) into the half-
plane Re (z)> 1; therefore g‘*) (z) is an entire function. For values z with
Re (z) > 1 we select b > Re(z) and obtain the representation
.
2b b+%00
Sar(S)do_ 2% 5 |ro Sar (b+ ty) dy
(7213) Wy eka ahs ni ieio (8(C—2) 20) —w (b+ 1y)*(b+ty—2)
ee °

It follows easily that g‘t)(z) is real for real z. Combining (9.2.12) and
(9.2.13), we see that for Re (z) > 1
gi eine) ac
(9.2.14) Bos (2)= Ini Nae ees )
270 CHARACTERISTIC FUNCTIONS

Moreover
be Seas
Joo 9(Ode =OO) (lz bee
> &)
if |Re(z)—b| > 1. The estimates stated in the lemma for g‘t)(z) are
obtained from (9.2.14), (9.2.13) and (9.2.11).
We define
Bir (®) = Sar (®)— Sar’ (2)
and obtain easily the second estimate of the lemma.
We saw that the functions g‘*) (z) and g‘;)(z) are entire functions which
are real for real z. The Maclaurin expansions of these functions therefore
have real coefficients, so that git) («+zy) and g\t)(«—zy) [respectively
& 7’ (x+1y) and gi,’ (x—zy)] are complex conjugate for x and y real. Writing
Uae(*,¥) = Relgi (x+y), wa’(% ¥) = Re[gi,’ (x+y)]
we have
ais) {tele9) = Hate ro) +a ON
Use (% Y) = 318c9 (wt iy) +807 (x—-vy)].
The functions on the right-hand side of (9.2.15) are entire functions. One
can therefore consider the equations (9.2.15) as definitions of u{ («, y) and
u\,) (x, y) for complex x (and fixed y). We use the estimates of lemma
9.2.2 and see that
‘ Oe O{|x |> exp [N (Re (x))?]} if Re (x) > 0
(9.2.16a) uw, (*,y) = ees Ane (get

(=) (xe — fA) if Re(x) 210

(9.2.16b) Ug, (%, ¥) = ee > exp [N (Re (x))2]} if Re(x) < 0.

We introduce the functions

Kyi (%) = ugh?

(x, 0)— ugh (x, Qu)
Ky 2(*) = ugg)
(x, 0)— upg? (x, 2zuz;).
Clearly these are entire functions, and we see from (9.2.16a) and (9.2.16b)
that they admit the estimates
(0.2.17) K,,(x) = are 5 exp [NV (Re (x))?]}
2
a1
Re (x) > 0
O(| x |°) if Re (x) < 0
- :
(9.2.17b) Kyo (x) = te ) glee
O{|x|5 exp [N(Re (x))?]}_ if Re (x) < 0
as |x| > oo.
For the study of the functions K,,(x) (r = 1, 2) we need two analytical
results. The first of these can be derived from lemma 8.2.1.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 271

Lemma 9.2.3. Suppose that the function f(z) is regular in the half-plane
Re (z) > 0 and that it satisfies the conditions
(i) 1f@)| <Myla+1l° for Re (z) = 0
(i) |f(z)| < M,e*(z+
1) for Im (z) = 0
(ii) |f(z)| < Mz; exp [d(Re (z))?]|z+1|* for Re(z) > 0
where M,, M, and M, are positive constants, while a, b, c, d are non-negative
constants and c > a. Then
|#(2)| < Mi |s+1|* exp [6 Re (2)]
in the half-plane Re (z) > 0.
Lemma 9.2.4. Let f(z) be an entire function which satisfies the condition
|f(z) |< exp {& Re (z)+ O(log |2|)} [Re (z) > 0]
where k is some real constant. se that f(z) can be represented in the form

fle) = (ar+by2) exp (FE)

with T > 0 and where otssertes converges uniformly on every bounded set.
Then a; = bs—0 tory So = (RE (27)s*\.
The proofs of lemmas 9.2.3 and 9.2.4 are given in Sections E3 and E4
respectively of Appendix E.
We return to the investigation of the functions K,,(x) and prove the
following statement.
Lemma 9.2.5. The functions K,,(x) (r = 1, 2) are polynomials of degree not
exceeding 5.
Since u,,,(x, y) = ult (x, y)+ ul,’ (x, y) we see that
K gu (3)= [tga(50) — thy(65 2oeuz8)) [0 (2, 0)—ul? (a, Qari].
In view of (9.2.7a) and (9.2.16b) we obtain for real x the eimate as x tends
ton-t.c0,
(9.2.18) Ky,(x) = (|x|).
We see from (9.2.17a) that
K,(*) = O{|~|* exp [N(Re (x))"]}_ (|x| > ©)
in the half-plane Re («)> 0. The conditions of lemma 9.2.3 (with f=
a=c=5,b=0,d = N) are satisfied, so that (9.2.18) holds in the half-
plane Re (x)2 0. We see from (9.2.17a) that (9.2.18) holds also for
Re (x) < 0, so that (9.2.18) is valid in the entire x-plane. Let
2.2)

Kyi (*) = d ax?

j=0
be the Maclaurin series for K,,(«) and consider the function
HO Ka (62) = hay ee:
j=0
Ziz CHARACTERISTIC FUNCTIONS

Since (9.2.18) holds for all real or complex x, we conclude that

H(w) = Ofexp [5 Re (w)]}.
The function H(w) therefore satisfies the conditions of lemma 9.2.4 (with
f = H,T = 22,6; = 0,k = 5, = 5) sothata; = 0 forj > 5. Therefore
K,1(*) is a polynomial of degree not exceeding 5. The statement con-
cerning K,.(x) is proved in the same way.
The information concerning the K,,,(x) (r= 1,2) given by lemma
9.2.5 permits us to get more precise results on the atone git) (x) and
82 (*).

Lemma 9.2.6. The eb git) («) and gi.)(x) admit the expansions

(9.2.19a) gh’(*) = = (ayn

+Bf x) exp [Mg1
JX] +Sy. (*)

(9.2.19b) gia () = = (ajo

+Bi x)exp [Myo J%] + Sy2 (x)
where the real constants aieand b') satisfy the condition
(9.2.19) ai? |+|8!2| = Ofexp(—Aj)} (j + «37 = 1,2)
for some k > 0. The S,,(x) are polynomials of degree not exceeding 7 and
have real coefficients.
We prove only the statement concerning git) («), since the statement
concerning the second function is proved in the same way.
To simplify the notation we write in the proof of formula (9.2.19a)
h(x), A and S(«) instead of g/i) (x), 2ni/u,, and S,,(x) respectively. After
the completion of the proof we revert to the original notation.
It follows from the definition of the function K,, (x) and from (9.2.15)
that
Kya (%)= 80a (®)— 24800 (@ + 2octugs) +801 (#— 2a) }-
Since K,, (x) is a polynomial of at most fifth degree, we obtain (using our
simplified notation) the relation

h(x +A) —2h(x)+h(*2—A) = = cj.x!

where the c; are constants (which depend on the stpbr ee subscript q).
We eHgaee constants c; (also depending on q) such that the polynomial
P{x) = > c;x satisfies the equation
j=2
5
P(x+A)—2P(x)+P(x—2) =>} ¢;x!.
j=0
The function
(9.2.20) h,(x) = h(x)— P(x)
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 273
then satisfies the equation
hy(x+A)—2h,(x)+h,(x—A) = 0
so that the function
hy (x) = hy (x)—hy(x—A)
is periodic with period A. From the definitions of the functions h, («) and
h, (x) and from lemma 9.2.2 we see that h,(«) is an entire function and that
for large | x |
h rs ee |? exp [N(Re (x))"]}_ if Re (x) > 0
2) = VO x |) if Re (x) < 0.
The conditions of lemma 9.1.2B are satisfied, so that

(9.2.21) h(x) = = d; exp (Ug1J%)

j=
where d; = O[exp (—4j*)] for some k. (We write here again d; instead of
ay )

The function h,(«)—A-1xh,(«) is periodic with period 4. It is again

possible to apply lemma 9.1.2B and we get

(9.2.22) hy (x) AT why (8) = ¥ 6; exp (gfx)

j=0
where 6; = 6° = Ofexp (—j?)] for some k. We now return to the original
notation and see from (9.2.20), (9.2.21) and (9.2.22) that
foo) 7

Sor(#)= & (aa

+ba*) exp (Hai) + Bi€g2
j= j=
where
|an |+|os5 | = Ofexp (—%y*)]
for some k > 0. The statement of the lemma follows from the fact (estab-
lished in lemma 9.2.2) that g\t) (x) is real for real x. Let
¢,(3) = 8(2)—821’ (2) €XP (Har 2) — 822) (2) EXP (Hg22)-
Using relations (9.2.9) and (9.2.10), we obtain the following two repre-
sentations for ¢, (2):

(9.2.23)
€,(2) = Sra)
is
(2) EXP (Maa ®) S02’
ie
(2) XP (My22)-
ty (2) = 842 (2) exp (Mas 2) 8a (2) exp (Hoa 2).
We apply the estimates of lemma 9.2.2 to these representations and see
that
1,(2) = Of 25 exp [ag Re (2)]}_ if Re (2) > 0
/,(2)= Off exp [!4g2 Re (z)]} if Re (2) <<0
as |z| > 00.
274 CHARACTERISTIC FUNCTIONS

We introduce the expansions (9.2.19a) and (9.2.19b) into (9.2.23) and

see that

sls) = BE (p+ ad) exp (Ha,J2)+La(@)

(9.2.24) where
L,(2) = (2)
+Soa (2%) &XP (Maa 2) + Sy2 (2) EXP (Haz®)
wherecc?\= a? sand d2e=1 lam
As a consequence of (9.2.19c) we have the estimate
(9.2.24) Jae |+|d0| = Olexp(—fy*)] (7 > &).
We also note that L,(z) is an entire function which is real for real z. Using
the estimates for ¢,(z) we see that

(9.2.24b) LL, (2) = O{|s|? exp [741 Re (z)]} for Re(z) > 0
O{|2|? exp [M42 Re (z)]} for Re(z) < 0
as |z| —> oo.
We introduce the functions
(9.2.25) ce (3) = Bit(2) -; k (z) exp (Mas 2)

Nga ()= 822) (2)—Soo (#) EXP (Hq22):

It follows from (9.2. sea) and (9.2.19b) that

(9.2.26) Mp3) =S (Ci0

+ A.8) exp (Har)
so that

(9.2.27) o(2) = y h,,.(2)+L,(2)

where L,(z) is the function defined in (9.2.24). The functions h,,,(z) are
entire functions and are real for real z. Let z = t+1y (t, y real) and write
(9.2.28) Hyq(ty)) = Igy (t)— Hegy (t+ 9) + g(t).
The right-hand side of this equation is for fixed y an entire function of f.
The function H,,(t, y) can therefore be continued into the complex plane,
and we write H,,(x, y) for its analytical continuation. We consider the
function H,,,(x, y) for fixed real y and complex x and use the estimates
of lemma 9.2.2 and formulae (9.2.25) and (9.2.28) and see that for
|x| + 0,
(0.2.29a) Cee {oH# exp [IV(Re (2))?|} for Reia)i.0
O(| x |) for Re (x) < 0
Lk O(|x |") for Re (x) > 0
(9.2.29b) Hy» (*, y) = oe exp [N (Re (x))?]} for Re (x) < 0.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 275
Let

A,(*, y) = L,(*)—3[L,(x+
ty) +L, (x—1y)].
The function A, («, y) is, for fixed real y, an entire function of the com-
plex variable x and we see from (9.2.24b) that for |x| > oo
Of) el Texp [pir Reta if Re (x) > 0
ow Aa(%¥) = a x :as AeRe x if Re i,< 0.
For real ¢ and real y we have
Hh,(t,y) = Ig,(t) — Re [gg (t + iy)]
A, (t, y) = L,(t)— Re [L, (¢+7y)].
Using these formulae, as well as (9.2.27) and the relation
u(t, y) = Re [g(t+zy)],
we see easily that

(9.2.31) u(t, 0)—u(t, y) = > Hi,(t,y)+Aq(t,9)-

We see from (9.2.5) that
u(t, 0)—u(t, 2mu,3') = Olexp (4,-1.1 2)]
as t—> oo. In view of our earlier estimate for A,(¢, y) and formula
(9.2.29b) we see easily that for s > g+1
(9.2.32a) —Hyy(t, 2nyz3) = Olexp (4,100) (8 = 9+2, 943...)
and

(9.2.32b) Hy,(t2rd) = OL exp (M2)

as t > + o. We see from (9.2.29a), (9.2.32a) and (9.2.32b) that the
functions H,,(t, 2au,,') satisfy, for integer s > q+1, the conditions of
lemma 9.2.3 and we conclude that
(9.2.33a) Hy (%, 2mu,') = Of]x|?exp [M.-11 Re (x)]}
[Re (x) 2 0,s = q+1,¢+2,...]
and we see from (9.2.29a) that, for |x| > ©,
(92 33D)em oie, (xeon Olle| for Retx)-< 0,
By means of a simple computation we obtain from (9.2.28) and (9.2.26)
the representation
roe) ay \2
(9.2234) 3 (*)¥) = 2m (eo dt 2) (sinwe) exp (Mai)
ies

te yas sin (Ug iJ) EXP (Uy 1J*)-

j=2
276 CHARACTERISTIC FUNCTIONS

We substitute y = 2n/u,, (with s = g+1,q+2,...) into (9.2.34) and

obtain a series of the form treated in lemma 9.2.4 with coefficients
17 \2 CIA athit oath
a; = 2¢% (sinee) + a4 — sin Pad
(9.2.35) 3,1 Msa Mea

b, = 249 (sinMad
8,1

and T = 2n/u,,. Then = m,_1,/g1 and the coefficients a; and 6; vanish

ifj > MaMa (6 = @+1,94+2,...). It follows from (9.2.35) that
dG =0 for My raMat <3 < Peabea (5 = 94+1,¢4+2,...)
and therefore also
cf= 0 for wy aaMar <3 < Msabar (8 =9+1,9+2,...).
The coefficients c\) may be different from zero only if 7 belongs to the set
Ji = Wg+ps/PMarihe-1- (It follows from the assumptions of the theorem
that J, is a set of integers.) We show next that dj‘ = 0 even forj € J;. We
carry an indirect proof and assume that d“| 4 0. We put y = 2z/u,, in
(9.2.34) and see that

H(t, 2p) = 2d'21(1 + 0(1)) sin? “=Pia exp (u,_1412)

(t + oo).
This contradicts (9.2.32a) so that
(9.2.36): ©-do=0? “foray: =. 253,05
We next derive an inequality for the coefficients c\. Let 2,,,be the
energy paraacicrs of f(t). We show that for m = Ply pads .. and
J = [n—11 Pas. the relation

(O237y a een af foryp ee

Mar

is valid. We give an indirect proof for (9.2.37) and assume that for some
integer m > g+2 the inequality (9.2.37) does not hold. Let p = uy. 4a
be the corresponding subscript. Then
® ein ane
Hy (:,-: = 2c\® (sinua) (1+0(1)) exp (fim—y,2) ast > 0.
m,1 m,1

It follows then from (9.2.31), (9.2.30) and (9.2.29b) that

Z
0 < u(t, 0) =C =) = 268 (sinHac) (1+0(1)) exp (H,~12 4)
m,1 ‘m,1

asi —> + ©,
The last relation leads to a contradiction with (9.2.5), so that the validity
of (9.2.37) is established.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 277

The function H,, (x, y) can be treated in a similar manner. One obtains
an expansion corresponding to (9.2.34) with coefficients c and d'\2. One
can again show that
Ge) fOr pos 2, 31.%..
while the coefficients c\2 may be different from zero only if j belongs to the
set Jo = {Ug+n2/Ma2}p—1 and where
C2372) a0 < ee Oe ator SE
Hae
for m = q+2,q+3....
We write Gor = Aah 100 2D = fae) fa
and m = q+1,q+2,... and obtain from (9.2.26) and (9.2.27) the repre-
sentation
Pe foe)

(9.2.38) (2) = XL LD AM, exp (Un,2)+L,

(2)
t=1 m=a+1
where
(9.2.38a) OnerA = Ang (t= 4+ 1, ¢ Zak):
The representation is valid for each g = 0, +1, +2,.... We select two
arbitrary integers g, and q2, q, < q, and write (9.2.38) for g = g,andq = q,
and subtracting the two equations we obtain
(9.2.39)
SA Are] exp (na 2) = — Are
—Ans] exp (4n.2 2)
M=Q+1 mM=d+1
2 Qs

— r=1
XZ D Ant xP (Hme 2) —L,, (2) +L,, (2).
m=q+1

We write A(z) for the sum on the left of (9.2.39) and see from (9.2.38a)
and from condition (9.2.1) of theorem 9.2.1 that
A(z) = O(1) for Re(z) < 0.
To estimate A(z) for Re (z) > 0 we use again (9.2.38a), (9.2.1), (9.2.24b)
and the fact that w,,. < 0 and obtain for the four terms on the right of
(9.2.39) the estimates
O(1), Ofexp [Ha,1 Re (2)]}, Of 21? exp [aa Re (2)]}
and O{| z|? exp [m,,; Re (z)]} respectively. Therefore
A(z) = Of{|2|? exp [u,,, Re (z)]} for Re(z) > 0.
We see that the function A(z) satisfies the conditions of lemma 9.2.4
with k = m,,1,4; = 0 if jis not one of the integers {u,,1/M@ga}m—a+1 O7 = 0
for j= 0, +1, +2...; T = 2n/u,,,. Then w = 1 and we see that
—A = 0 for m = q.+1, q2+2,....
Ae)
Similarly one can prove that
Ag —Ane = 0 for m= q2+1, q2+2,....
278 CHARACTERISTIC FUNCTIONS

Since q, and q, are arbitrary we conclude that the 4'”, do not depend on q;
we therefore write /,,, instead of 4, and obtain the following representation
for g(2):

(9.2.40) &(2) = » aa dag exp (ae ae oe(2)

r=1 m=arl
where
(9.2.40a) O0<4,,< 24, (2 =0,41, +255)
and where this representation holds for all integers q.
We introduce the function

241)
fal
$a)= BB r=1 m=—c
Aa (emer —1 ine)
ee 1+ ,2,, ;

_ We repeat the reasoning used in the proof of theorem 9.1.1 and see that
@(z) is an entire function and that
(9.2.42) —$(z) = O{|2|* exp [M(Re (2))"]}
where N > 0 is a constant.
We introduce the function
L(z) = g(z)-4(z).
To prove theorem 9.2.1 we must show that
L(z) = f2*+Bz
where f is real while 7 > 0.
The first step in the proof is the demonstration that L(z) is a polynomial
of degree not exceeding 3.
We see from (9.2.42) and lemma 9.2.1 that there exists a positive con-
stant A such that
[L{2)| < Alz|® for Re{ze) =.
According to formulae (9.2.40) and (9.2.40a) we can represent L(z), for
each g-= OF 1) 2s as
2 fo) re:
eee My yee (14o ee )
TrT=1 m=a+t+l1

= Lit LetL, (say).

The sum 5), is a linear function of z and we have
di = O(/2|) as |2| > o.
We next use lemma 9.1.1 and estimate >), in the same way in which we
estimated S, in the proof of theorem 9.1.2, and obtain
Le = Of]? exp [m,1 Re (2)] }+ Of] 2? exp [4,2 Re(z)]} (|2] > 0).
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 279

Using these estimates for )}, and ¥, and the estimate (9.2.24b) for L,(z),
we conclude that

(9.2.43) Lz) = {O{|2|? exp [u,, Re(z)]} if Re(z) > 0

O{|2|’ exp [1,2 Re(z)]} if Re (z) < 0,
as |z| > oo.
We apply lemma 9.2.3 to the function L(z) (with M, = A, a = 3,
b = d = p,,, ¢ = 7) and see that
|L(2)| <Al2/ exp [44 Re(2)]_ if Re(s)>0
If one applies lemma 9.2.3 to the function L(— 2) (with M, = A, a = 3,
b = d = —f,9, ¢ = 7) one sees that
|L(2)|<Alz
|exp [digo
Re(2)] if Re(z) < 0
We note that the constant A in these estimates is independent of g and that
q can be chosen arbitrarily from the positive or negative integers. We
therefore let g tend to — oo, and since lim yw,, = 0 (r = 1,2) we finally
q>—-oO
obtain the estimate
|L(z)| < Alz|®
which is valid for all z, so that the entire function L(z) is necessarily a
polynomial of degree not exceeding 3. We note that L(z) is real for real z
and that L(0) = 0; therefore
L(z) = 623+ 722+ Ba,
where f, 7 and 6 are real constants. We had
u(t, y) = Re [g(¢+v)],
so that
u(t, 0)—u(t, y) = g(t)— Re [g(t+y)]
or since g(z) = 4(z)+L(z)
u(t, 0)—u(t, y) = 6(t)—Re [4(t+zy)] + L(t)—Re [L(t
+iy)].
It is easily seen that
L(t)—Re [L(t+iy)]= (36t+P)y?.
If one of the relations 6 = 0, 7 > O were not satisfied then we could find
at such that 36t+7 < 0. We fix such a value of ¢ and see from (9.2.41) that
(t)—Re [d(t+2y)] = o(y?) asy > o.
Hence, for such a fixed t and y > oo, we would get
u(t, 0)—u(t, y) = o(y2) + (3dt+F)y? > — .
This contradicts (9.2.2) and we see that necessarily 6 = 0, 7 > 0. There-
fore
ar Ming? en as
g(z) = >, ine in (¢ ~1-; fest.)+ 62-72",
280 CHARACTERISTIC FUNCTIONS

so that the characteristic function f,(t) belongs to the class #; this means
that every factor of f(t) is infinitely divisible, ie. f(t) € J) and theorem
9.2.1 is proved.
Theorem 9.2.1 gives a sufficient condition which assures that a charac-
teristic function f(t) of the class Y belongs to I). This condition can be
weakened if f(t) belongs to a lattice distribution. I. V. Ostrovskii (1964)
obtained the following result:

Theorem 9.2.2. Let f (t) be the characteristic function of a lattice distribution

with span & and suppose that
(i) f@) eZ
(ti) Ane = O(eXP [—28-*
|Hm |log (E-*
|Mme [)]) a8 mm > + 005
fa ie,
Then f (t) belongs to I.
For the proof we refer to Ostrovskii’s paper.
Remark. A. A. Goldberg-I. V. Ostrovskii (1967) constructed an ex-
ample which shows that there exist characteristic functions which belong
to the class but not to J). This example also indicates that for charac-
teristic functions of lattice distributions of span & = 1 the condition (i1) of
theorem 9.2.2 cannot be replaced by /,,, = O[exp (—|Mm,,|)] asm > + ©,
f=, 2.

9.3 A necessary condition for membership of I,

We now present a necessary condition which an infinitely divisible
characteristic function with Gaussian component) must satisfy in order
to belong to J).

Theorem 9.3.1. If an infinitely divisible characteristic function with Gaussian

component belongs to I, then it necessarily belongs to the class L.
To prove the theorem we assume that the characteristic function
f(t) € 1) and has a Gaussian component. We show first that f(¢) has a
finite or denumerable Poisson spectrum. We give an indirect proof and
assume therefore that the positive Poisson spectrum has a (non-constant)
continuous component,(+) so that
f) = fi) fa);
where

a(t) = exp{—re+ |“(e—1yancw},

1

(*) We say that an infinitely divisible characteristic function has a Gaussian com-
ponent if y > 0.
(t) This means that in the decomposition
N(u) = a, Na(u)t+a,Ne(u) (a, > 0, ag > 0, ay tay = 1)
of N(x) into a discrete and a continuous component, a, # 0 and Ne(u) is not constant.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 281

with a continuous spectral function N(u) for which N(b,) > N(b,).
According to theorem 8.2.9 it is possible to determine positive numbers »
and 7 so small that
bs
Js(t). = exp {= | (ei — 1) dN (u)— we"— 1)}
by
is a characteristic function. Since f(t) is not infinitely divisible it necessarily
has an indecomposable factor. Writing
Ff) = AOA) exp W(e"-1)},
we see that f(t) has an indecomposable factor. This contradicts the
assumption f(t) € Jo, so that (wu) cannot have a continuous component.
The same argument is used if the negative Poisson spectrum has a non-
constant continuous component. We see therefore that f(t) ¢ J, implies
that the spectrum of f(t) is either denumerable or finite. We can therefore
write log f(t) in the form
(9.3.1) log f(t) = ait—yt?+ > a (ew ve =e.)
a 1+;
roo) tS ere ity; )

5 3 16 ; 1492)’
where) 1; > 0,A_; > 0,y > 0. Ifthe spectrum is denumerable we assume
that the conditions (9.1.3a) and (9.1.3b) are satisfied. Suppose that the
positive Poisson spectrum contains at least two points, let w and u’ > wu be
two frequencies of the positive spectrum and let A and 2’ be the corres-
ponding energy parameters. We show next that the quotient « = w/w’ is
-a rational number. The characteristic function f(z) then has a factor
filt) = exp {—yt?-+A(e—1) +2’ (et#/4—1)}.
If « is irrational then it follows from theorem 8.2.8 that f, (¢), and therefore
also f(£), has an indecomposable factor. This contradiction shows that «
is necessarily rational, say « = p/q, where p and q are integers and can be
assumed to be relatively prime, p < g. We apply the reasoning used
before and use theorem 8.2.7 to conclude that f,(¢), and therefore also
f(t), has an indecomposable factor unless p = 1. The negative Poisson
spectrum is treated in a similar way, so that f(t) belongs to the class Y.
Remark. 'The presence of a Gaussian component is essential since y > 0
is necessary for the validity of theorems 8.2.7, 8.2.8 and 8.2.9.

9.4 Infinitely divisible characteristic functions with bounded

Poisson spectrum
In this section we study the factorization of infinitely divisible charac-
teristic functions with bounded Poisson spectrum. We also derive sufficient
(*) In order to use (9.3.1) in case of a denumerable as well as a finite spectrum we admit
the possibility that only a finite number of energy parameters is positive.
282 CHARACTERISTIC FUNCTIONS

conditions which assure that a characteristic function with bounded

spectrum belongs to J).
We use for the spectra the notations of Section 3.7 and have to supple-
ment these by introducing a convenient notation for the vectorial sum of
identical summands. We define the symbol (n)A recurrently by writing
(1)A = A and (n)A = (n—1)A(+)A for n = 2,3,.... We also write
(o)4= Udacs
We need the following lemma:

Lemma 9.4.1. Let A be a closed set on the real line which is contained in the
finite interval [a, b], where0 < a< b < ow. Then (o)A ts a closed set.
We note that (n)A c [na, nb], and since a > 0 any finite interval can
intersect at most a finite number of the sets (7)A. Let x € (00)A; then there
exists a sequence of points {x,,} in (co)A which converges to x. The interval
(x—1,«+1) contains therefore almost all elements of this sequence {x;}.
However, the interval (v—1, «+1) intersects only a finite number of the
sets (n)A. Therefore there exists at least one set (x)A which contains an
infinite subsequence of the {x,}. The set (x)A is closed(f) so that x € (n)A
and therefore also x € (00)A; hence (0)A is closed.

Theorem 9.4.1. Let f(t) be an infinitely divisible characteristic function

without normal component and which has a Poisson spectrum A such that
0< a= inf x < b= sup x < o. Then any factor f,(t) of f(t) has the
aEeA cEA
form
fi(t) = exp ives Mi(1) dN w)),
where N(u) ts a function of bounded variation which is non-decreasing in the
half open interval [a, 2a) and which has a spectrum Sy < [(«)A] ¢ [a, 8].
The constant y 1s real.
Without loss of generality we can assume that f(t) is given by

(941) “¥@)= exp ithe(1) dN, w)|,

where N,(u) is non-decreasing and has the spectrum Sy, = A. Then

f(t)= {1+ >

Stif ei" dN, OltJk }
where c = exp |-| aN,(u)]. Let F(x) be the distribution function
0

(t) We have seen in Section 3.7, p. 57, that the vectorial sum of two closed and bounded
sets is closed.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 283

corresponding to f(t); then

(9.4.2) F(x) = c{es) + S N¥ (x)/k I

k=1

Here Nj" denotes the k-fold convolution of N(x) with itself. We see
from lemma 3.7.4 that
Sw, = (B)Sx, = (RA = (A.
It follows from (9.4.1) that
(9.4.3) Sr =([(o)A] {0}
where {0} is the set containing only the point 0.
We assume now that f(/) admits a decomposition
f(t) = fi(t) fo(?).
Then
B(x) = F(x) * F(x),
where F',(x) and F(x) are the distribution functions corresponding to
fi (é) and f,(t) respectively. We see from (9.4.1) and from the assumption
concerning the spectrum A of N,(uz) that log f(t) is an entire function of
order 1 and type not exceeding b. Therefore f(t), and hence also f, (t), is an
entire characteristic function without zeros. Moreover, we conclude from
corollary 1 to theorem 8.1.3 that log f, (z) is an entire function of type not
exceeding b.
We see from lemma 3.7.4 that
Sr = Sr(+)Sr, > Sr (+)S-F,;
since [0, 00) > Sp we conclude that Sp, and Sp, are both bounded from
the left. It is no restriction(}) to assume that the infimum of Sz, is the
point 0. Then 0 € Sp, and 0 € Sz, so that
(94-4) Sp U Sp, © Sp.(+)Sp, © Sr.
We see from (9.4.3) that the point 0 is an isolated point of Sp. Let
a, < a and write V = (—a,,0) U (0, a,) for the union of the two open
intervals (—a,, 0) and (0, a,). We see from (9.4.3) that V does not contain
any points of S, therefore V < S‘,. We conclude from (9.4.4) that V < S%,
so that the point 0 is an isolated point of Sp. This means that /’, has a
discontinuity at the origin. Let d be the saltus of F, at 0. Then F,(x) =
de(x)+G(x), where d > 0 and where G(x) is a non-decreasing function of
bounded variation such that
(O45) (co) Ane [a.c).
The characteristic function f, (¢) of F’,(«) is therefore given by

ae aes iP dG(2).
(+) This can be shown by replacing F(x) by F,(x+-8) and F(x) by F,(x—8) where
5 = lext [F,]-
284 CHARACTERISTIC FUNCTIONS

We select a positive real number 7 so large that | e~™® dG(x) < d and
a

write G,(x) = laa dG(¥). Let t be real; then

(9.4.6) [| eta, (0 < i dG,(x) < d

and we see that (¢, 7 real)
$,(t-+in) = log f,(t-+in) = log {a+s et dG, ()}
or

(94:7) © oi ele lon ds = (—1)' ie é dG,(2) |'(kd)

Let G*" (x) be the k-fold convolution of the function G,,(x). We see from
(9.4.6) that the series > (—1)*-1 Gi" (x)/(kd*) converges for all x. We
k=1

write N, (x) = y (—1)*-1G® (x)/(kd*). The function N, (x) is a function

of bounded vanaaths Since Sg, = Sg we see from (9.4.5) that

Sry = (0), = U HSo< UH U wal = (A.

It follows from (9.4.7) and thedefinition ofthe ashe N,, (x) that
(D4.8)\~ Sb,(x tiny tos a8 |:e* AN, (x) (t, 7 reall).
We mentioned earlier that ¢,(z) is an entire function of exponential type
not exceeding b. According to remark 7 following theorem 7.2.3, the spec-
trum Sy, is contained in the finite interval [—, 5]. Since ¢,(z) is an entire
function, relation (9.4.8) also holds for complex values of 7; we substitute
t = t—in (t, 7 real) into (9.4.8) and write N(«) = lb e""dN,,(¥). In this
way we obtain .
(9:49)! tod; (2) log ikoN (x).
Here :
Sw = Sxq = [()A] 0 [=8, 6] = [(c0)A] 2 [a 8].
We see from (9.4.9) that

A(t) = exp[4.()] = 4 k=0

||“ete dN (a)|[Rl
0

The corresponding distribution function is given by

(9.4.10) F,(x) = a4(0)ipx Ne (x)/h .

CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 285

Let H(x) = 5 N*(x)/R!

sothat
(9.4.11) F,(x) = d{e(x)+ N(x)+ H(x)}
where

Sq U (R)Sw € [2a, «).

k=2

Let x, and x, be two points such that

0 ta we — 27,
it follows from (9.4.11) that 0 < F,(«2)—F,(«,) = d[N(x.)—N(x,)], so
that N(x) is non-decreasing in the interval [a, 2a). We see further from

(9.4.9) that d = exp |-|.an (x);hence ¢,(t) = fe(e*—1)dN(x) so

that the theorem is proved.
Theorem 9.4.1 can be used to derive interesting conditions which assure
that a characteristic function belongs to the class J.

Theorem 9.4.2. Let f(t) be an infinitely divisible characteristic function

without normal component, and suppose that its Poisson spectrum lies in the
closed interval [a, b] where 0 < a < b < 2a. Then f(t) belongs to I.
The theorem follows from theorem 9.4.1 in the case where b < 2a. We
must therefore only consider the case b = 2a. Let f,(t) be a factor of f(¢).
We saw in proving theorem 9.4.1 that f,(¢) is an entire characteristic
function without zeros; in view of the statement of theorem 9.4.1 this
factor has the form

f(z) = exp {ins- | (Ee —1)dN (u) + A(e#*— i)

[a,2a)

Here z = t+vy (i, y real) and N(u) is non-decreasing in the half open
interval [a, 2a) over which the integral is taken. We have therefore to
prove only that A > 0. We give an indirect proof and assume tentatively
that 2 < 0. An elementary computation shows that

a ~ ZuexD {| e~™ (cos tu—1)dN (u)+A e~™™ (cos 2at— i).

1\ly {a,2a)

We put here ¢ = fy = icand see that

filtotty) = exp {-etl a— | e4—™Y (cos ty %— 1) ana}.

fil) {a,2a)

We note that
lim e4—"' (cos tu—1)dN(u) = 0,
yo [a,2a)
286 CHARACTERISTIC FUNCTIONS

therefore
fi(tot+vy) = exp {—e>*"[A—o(1)]} as y — =o.
fi (yy)

If 2 < 0 this means that Aultot)| 5 1 for y < 0 and |y| sufficiently
f(y)
large. But this contradicts the ridge property, so that A > 0. Hence the
theorem also holds in the case where b = 2a.
Remark 1. The assumption that 0 < a and b < 2a is essential. If
a=0 or b > 2a then theorem 6.2.3 (respectively theorem 6.2.4) can be
used to construct counter-examples.
Remark 2. An analogous result can be obtained for infinitely divisible
characteristic functions without normal component with bounded negative
Poisson spectrum.
Theorem 9.4.2 has an interesting consequence which illustrates the
important role of the class I).

Theorem 9.4.3. Every infinitely divisible characteristic function can be

represented as a product of at most denumerably many factors belonging to I.
Let f(t) be an infinitely divisible characteristic function whose Lévy
canonical representation (theorem 5.5.2) is determined by the constants a
and o? and the functions M(u) and N(u). We introduce, for k = 0,
+1, +2,..., the functions
0 ifu << —2*+1
M,.(u) = «M(u)—M(-2**2) if —2*tl< uy < —2%
M(—2")—M(-—-2**1) if —2§ <u <0
and
0 if0 <u < 2
N;,(u) = 4N(u)—N (2*) if 2* <u < 2h
N (2*+1) — N (2) if 241 < y,
We write f,?(t) [f(t] for the infinitely divisible characteristic function
without normal component whose Lévy canonical representation is de-
termined by M,,(u) LN; (u)]. Then
foe)

TROP ERD AGTt7/2 aA

=— 0
ne) ie)
and see from theorems 9.4.2 and 8.2.1 that all factors in this representation
belong to J).
We introduced in Section 8.2 (p. 249) the notion of a finite set of
rationally independent numbers. For the next theorem we need an exten-
sion of this concept.
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 287

A set A of points on the real line is said to be a set with (rationally) in-
dependent points if every finite subset of A is a set of rationally independent
points.

Theorem 9.4.4. Let f(t) be an infinitely divisible characteristic function

without normal component and suppose that its Poisson spectrum A its
positive and forms a closed, bounded set with independent points. Then f(t)
belongs to I.
Let a = infx and b = sup x.
aed wed

As a consequence of our assumption that A is a set with independent

points we see easily that the sets (k)A (k = 1,2,...) are pairwise dis-
joint. Let 2 be the largest positive integer such that na < b; then
(9.4.12) [(k)A] 1 [a,b] =O fork > xn.
The characteristic function f(t) satisfies the conditions of theorem 9.4.1.
Any factor f, (t) of f(t) therefore has the form

fi (t) = exp {int eg(= 1) an(w)}

with
Sy < [00(A)] A [a, 8],
In view of (9.4.12) we can write

(493) sic | U @)4| A [a, 5].

k=1
Let N,,(u) be the restriction of N(u) to the set [(m)A] - [a, 5] (m = 1,
2,...,) and consider the Fourier—Stieltjes transforms of these functions
of bounded variation:

a(t) = | ett dN, (tt).

(m)A

Then

(9.4.14) iO 6 exp live 3 bn(t)]

We see from (9.4.14) that the function exp |x Pn (| is, except for a
n= 1

constant positive factor, a characteristic function. Therefore there exists

a non-decreasing function G(x) of bounded variation defined on [a, «]
such that
iFge dG(x)\i—="exp bp dm |.
288 CHARACTERISTIC FUNCTIONS

Hence
ip€ dGtx)earexp jer ibiekdN (|

- Zarmyvm (Aran)
=v
. (2)A i hae (); ry ({(n)A ae ():
or
(415) | edG(x)= ah...
so that
G(x) = SNE
* NE *... Ni/(Ry! Ro! ... Rn!)
The function Ti Ne* has as its spectrum the set

(,)A(-+) RMA +) - (4) ea)Mle)

A]= (Rat Dky b. «.-fky)A.
If x €(m)A, then
k,+2ka+...+nky, = m
and we conclude from (9.4.15) and the fact that the sets (k)A are disjoint
that
(9.4.16)
| 6 dG (a — > [bi (ys. [on OYRel eel):
(m)A kyt+2ka+...+nkn=m

We see therefore that the expression on the right of (9.4.16) is the co-
efficient of y” in the expansion of
(9.4.17) exp [yi (t) ty? b2(t) +... +9" on (4)].
It follows that (9.4.17) is, except for a constant positive factor, a characteris-
tic function, provided y > 0. We also see from (9.4.16) that

|edo(e) = 41(0,
so that ¢,(t)/¢,(0) is a characteristic function. Since, according to our
assumption, fi (¢) is a factor of f(t), there exists a characteristic function
f(t) such that
(9.4.18) fi®f®) =f)
and we can repeat the earlier reasoning and show that
(9.4.19) f,(t) = Cexp (dt+y,(4)+y2(t)+...+Yn(2)}.
Here C and 6 are constants while the functions y,, (#) have the form

», (= | eM dN (x) (m= 1,2,...,n),

(m)A a [a,b]
CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 289
where N is a function of bounded variation such that

Se < [a,b] 0 |U (nA.

m=1
We see from (9.4.18) that
log fi (t) + log fa(t) = log f(z)
or, in view of (9.4.14) and (9.4.19),
[Ps (2) + yr (4)] + [$2 (2) + Ya] +... + [$n (t)+¥n(A)] = log f(e).
The spectral function of f(t) has the set A as its spectrum, while 4,, (t) +
Wm (t) has (m)A as its spectrum. The sets (m)A are pairwise disjoint, and we
conclude easily that
(2420), Vn (1) — bn(2) (rt S 2,3, aca, ny:
As in the case of (9.4.17) we conclude from (9.4.20) that, except for a
constant factor,
exp {yvi(t) — y*b2(t)—-..—y"$n()}
is a characteristic function, provided y > 0. But then the functions
yy (tv) +? bp(tv) +... +y"by (iv)
Pr (10) —Y"Po(22) —. . "bn (12)
are convex functions of the real variable v if y > 0. This is only possible if
be (Pi Oe (Jie 2, ee tt):
Since the entire functions ¢;(t) are Fourier—Stieltjes transforms of func-
tions of bounded variation we see easily that the functions ¢;(t) (7 = 2,...,
m) reduce to constants. We put
(9.4.21) 3 (Ei = DAO) rep fo 32, sheen; It),
and substituting (9.4.21) into (9.4.14) we see that
fit) = Cexp [tyt+4i()+¢e.+...+¢]
or
fi(t) = Cy exp [t+¢, (2)].
Similarly we obtain from (9.4.21), (9.4.20) and (9.4.19)
fo(t) = C, exp [2dt+
y, (2)].
The statement of the theorem follows from the fact that the functions
¢,(t) and y(t) are, except for a positive constant factor, characteristic
functions.
Remark 1. Theorems 8.2.6 and 8.2.5 are particular cases of theorems
9.4.2 and 9.4.4 respectively.
Remark 2. Extensions of theorems 9.4.1, 9.4.2 and 9.4.3 can be found
in Cuppens (1968) and in I. V. Ostrovskii (1966).
290 CHARACTERISTIC FUNCTIONS

9.5 Theorems concerning certain factorizations

The methods used in the last three sections make it possible to derive
results concerning the possible factors of certain infinitely divisible distri-
butions.) We list in the following a typical result.

Theorem 9.5.1. Let f(t) be a characteristic function which admits the repre-
sentation ;

log f(t) = BE (et brit) exp (DE) +3 aw xP (inet) b+LEH

2 il oo

r=1 Pp=tr m=1

where the parameters occurring in this representation and the function L(z)
satisfy the following conditions :
(a) 5 > 0, Ey < 3

(b) n, and n, are integers such that —0 <n, < © (r = 1,2) and
M6, > M252;
(c) n, and nj, are integers, n, > n, and n, > 0 (r = 1,2) [if n, =
n—1
then the sum ) 1s omitted];
rr

(d). the coefficients a,,.and b,,. are real, 7. 2 ),\W, = 0 my a

(m = 1, 2,......). Moreover,
(di) 4, = 1,€, (r = 1, 2),
(dy) M4 17/Y%me 1s @ natural number greater than 1,
(d3) for some k > 0 we have i,,, = Olexp (—kv?2,,)];
(e) L(z) ts an entire function which is real for real z and satisfies the
estimate (as |x| > ©)
ie fotz|* exp [£| Re(s)]} if Re(z) > 0
O{|z | exp [£ Re (s)]} if Re(z) <0,
where &, and &;, are real numbers such that
max [72 §s,(%,—1)&1] < & < més,
and m,&, < & < min [(m,—1)é., 2, é,].
Let f, (t) oea factor of f (t). Then f, (t) has the AKG

log hi (t) = 3bee (4, + 6,,it) exp (7pé, t)+ oy Amp EXP (rar) }+ Li,

where thea,, a, b,,,are real constants and where the Me any satisfy the
inequality 0 < Ruy < Am,» The function L(z)is entire, real for realx,and the
estimates
z)]} if Re(z) > 0
de i 2 | exp [£, Re(z)]
O{| | exp [& Re( } if Re(z) < 0
hold for |z| > o.

(*) See Yu. V. Linnik (1964), Chapter 9, and I. V. Ostrovskii (1965).

CHARACTERISTIC FUNCTIONS WITHOUT INDECOMPOSABLE FACTORS 291

Corollary to theorem 9.5.1. Suppose that the conditions of theorem 9.5.1 are
satisfied and that
L(z) = (y2?+ Be+a)e”
with «, B, y and y real and n # 0, where
max [(m,—1)&,, maf] < 4 < min [(n,—1)&., 2, &].
Then the function L(z) which occurs in the representation of f,(t) is given by
L(z) = & with & > 0.
For the proof of theorem 9.5.1 and its corollary the reader is referred
to the paper by I. V. Ostrovskii (1964). In this paper Ostrovskii also con-
siders the case where the constant 7 = 0 or where L(z) = 0. The same
paper contains several theorems similar to theorem 9.5.1.
Theorem 9.5.1 and its corollary can be used to derive conditions which
assure that the convolution of Poisson-type characteristic functions belongs
tod @:

Theorem 9.5.2. Let

3
ie) = exp {DAR [exp (itmt)—1)} (Ate Opt hier.)
m=1

be the characteristic function of the convolution of three Poisson type distri-

butions.*) Suppose that one of the following four conditions 1s satisfied :
(1) fa < 0, ws > 0,0 < py < min (Ws, |41 ]),
(Gi) ee Oe 0; 0 > i, Sa max (—fiz, 15),
(ME Bie pa = MUN (2h, pe),
(iv) 0 > py > fe > max (2pz, fs).
Then every factor of f(t) is also a convolution of at most three Poisson-type
distributions, so that f(t) belongs to I.
We indicate the proof of theorem 9.5.2 in the case when condition (i) is
satisfied.
Wer choose" 4) =n, Ey = 1, m1, = 1, 2 = 1,7, = 0; n, = 1 and
Qyg = —(Ay+Ag+As) and put Do. = B, Ary = Ag, Avg = A, and A, = 0 (for
m > 1 andr = 1, 2). Let further L(z) = A, e“*. It is then easily seen that
the conditions of theorem 9.5.1 are satisfied and that f(£) is the convolution
of three Poisson-type characteristic functions. The statement of theorem
9.5.2 follows almost immediately.
In his paper I. V. Ostrovskii indicates a similar result for the convolution
of four Poisson-type distributions.

(*) It was shown in Section 8.2, p. 252, that such a convolution does not necessarily
belong to Ip.
 10 «-DECOMPOSITIONS

In this chapter we extend some factorization theorems for analytic char-

acteristic functions. The results presented are, strictly speaking, of an
analytical nature but are closely connected with the arithmetic of distri-
bution functions. Most of these studies were originally motivated by other,
more penetrating investigations of the theorems of Raikov and Crameér.
We say that a characteristic function f(t) admits a (finite) «-decomposi-
tion if there exist characteristic functions f, (2), f.(t), . . . , fn (t) and positive
numbers «,, %,..., %, such that the relation

f(t) = I [fs (el

holds either in an interval |t| < 6 on which f(¢) is different from zero or on
a sequence of points {¢,} such that lim ¢, = 0. The powers of the character-
k—>o
istic functions f;(¢) are defined by [f;(t)]” = exp [«, log f;(t)], where we
take for log f;(t) that branch of the logarithm for which log f;(0) = 0 and
which is continuous. Denumerable «-decompositions are defined in a
similar way: the finite product in the representation of the characteristic
function f(z) is replaced by an infinite product.

10.1 General theorems on o-decompositions of analytic

characteristic functions
The first theorem of this section is related to theorem 8.1.1.

Theorem 10.1.1. Let f(t), fo(t), ...,f<(t) be arbitrary characteristic func-

tions and let a4, %,..., % be positive real numbers. Suppose that f(t) is an
analytic characteristic function which has no zeros inside its strip of regularity
and that the relation

(014) TTH@l =f
holds in a neighbourhood of the origin. Then the functions f,(t)(j = 1,2,...,5)
are analytic characteristic functions and are regular at least in the strip of
regularity of f(t) and (10.1.1) is valid in this strip.
In the following we write, as usual, F';(«) and F(x) for the distribution
functions corresponding to f;(t) and f(t) respectively. We show first that
the theorem holds if the distribution functions F(x) (j = 1, 2,...,s) and
F (x) are symmetric.)
(*) We will use in the following the properties of symmetric distributions, mainly
theorems 3.1.2 and 3.1.3.
&-DECOMPOSITIONS 293
We note that it is no restriction to assume that
(10.1.2) EE LAL mel es ees
This can always be achieved by raising both sides of (10.1.1) to an integer
power.
We first prove) that the second moments of the distribution functions
PF x)(j = 1,2,..., 8) exist. The characteristic functions f;(¢) are real and

i,(t) = | cos tx dF; (x) = 1-2 { sin? 5dF,(x)

so that
f(t) < exp te [ sint a,

Let 4(¢) = log f(t) be the second characteristic of F(x); we see then from
(10.1.1) that
8 co E
OL; | sin? 5 fF, (x) < —$¢4(2)
—
j = ad
hence
c t 1
[sine Sa (x) = WO) Go 1,2,..-.5),
We write x; for the jth cumulant of f(t) and note that x,,_, = 0 (j = 1,
2,... ad inf.). It follows from the preceding inequality that
© sin? 4(ix QOL) AK
[pare < ae = are as t >0.

It follows from Fatou’s lemma [see Titchmarsh (1939), section 10.8.1]

that the second moments of the distributions F;(«) (j = 1,2,..., s) exist.
We show next by induction that the distributions F;(«) have finite
moments of all orders. We assume therefore that the distributions
F(x) (j = 1,2,...,s5) have moments of order 2k and show that this im-
plies the existence of the moments of order 2k+2. We differentiate
equation (10.1.1) 2k times and obtain on the left side a sum where each
term contains derivatives of the f;(¢). We arrange the terms on the left
side into three groups and write
(10.1.3) S,(t)+S,(é)+S3(é) = f° (2).
Here

(10.1.4 ) Sas1(t) ea f()

a)
ie) py
& es* f(t)

contains all the derivatives of order 2k while S(t) contains only derivatives
of even order not exceeding 2k—2; the summand S,(f) consists of all
(*) The proof was suggested by R. G. Laha.
294 CHARACTERISTIC FUNCTIONS

terms which contain a derivative of odd order. We note that each term of
S(t) necessarily contains two derivatives of odd order, so that [see state-
ment (i) of the corollary to theorem 3.1.3] S,(0) = 0. We remark further
that each term on the left side of (10.1.3) has, except for a constant co-
efficient, the form
y
ioy inp pasc
f(t) i ane [
Jan (8) J weedeat) Fam(t)
where each a; is one of the integers 1, 2,...,s5 and where the positive
integers 71, To)... +) %m and 14, Mo, ..., Mm satisfy the relation

(10.1.5) 3 m7) = 2k.

j=
We see easily from (10.1.3) that

(10.1.6) Moe ip=() Ss is SOM Osf2(0)

It follows from the corollary to theorem 3.1.3 and from the definition of
the functions S,(t) and S,(t) that
Su)
—— and S4(t)— $50)
—*-——=~
i ft
tend to finite limits as t goes to zero. Moreover, we conclude from the
fact that f(t) is an analytic characteristic function that
i TR (t) = Fike (0)

t2
t>0

exists and is finite. Hence this is also true for

linn Sy (t) = Si (0)
t—>0 t

We see from (10.1.4) that

1-0)
_ $f
$ , Ob PPO-
S,(t)—S,(0 (2)= LO)
{28 _ Spon
3 HO=
(t)— FO
r=) e a Pf,(t)
It is easy to see that the second sum on the right of this equation tends to
a finite limit as t goes to zero; therefore
+ (2k) (¢)__ f2k) (9) 8 Ee) Boyegtx
ene (t)
t2
fi ( ) = (=1912 3
=
wy f
is
yok
t2
dF, (x)
j=1

also has a finite limit as ¢ approaches zero. Then this is also true for each
summand on the right of this equation and we use again Fatou’s lemma to
conclude that the moment of order 2k+2 exists for the distribution func-
tion F;(x) (j = 1, 2,..., 5). This completes the induction.
We prove next that the functions f;(¢) are analytic characteristic func-
&-DECOMPOSITIONS 295
tions. We raise equation (10.1.1) to the power 2k and differentiate it 2k
times. We write g(t) = [f(t)]** and obtain
(10.1.7) S,*()+S,*()+S,*(2) = 2 (0).
Here S,*(t), S,*(¢) and S,*(¢) contain the same kind of terms which we
had in S,(t), S,(t) and S, (t) respectively; expressions (10.1.7) and (10.1.3)
differ only in the numerical values of the coefficients. This difference is due
to the fact that we raised (10.1.1) to the power 2k. We have then

S.*() =f) & 2ka, LO

Fi(t)
Let G(x) be the distribution function which belongs to g(t) and denote
the algebraic moments of order r of G(x), Fy(x), ..., F(x) by «,
eos, 0. , wespectively,-and. pute?,—.0 in <erneaata (10.1-7); Since
S,* (0) = 0 we obtain
(1050-8) Sy (0) 55"(0) = (S41) om
where

(10.1.9) S,*(0) = (-—1)F 2 2ka, a9).

We conclude from the fact ‘se we raised (10.1.1) to the power 2k that
S,*(0) = (—1}FC
where C is a positive constant. It follows then from (10.1.8) and (10.1.9)
that

(—1)*[S1* (0) + S3*(0)]= z ZRaja tC = ogy

so that
(10.1.10) Co een freee la ys ny SiR ly 20s «os
The number «® is the (2k)th moment of the distribution function
belonging to [f(t)]®*, that is, of a distribution function which depends on
k. It is therefore not possible to conclude from (10.1.10) and the fact that
f(t) is an analytic characteristic function that the power-series expansion
of f;(z) converges, at least in the circle of convergence of f(z). Our next
aim is to show that the f;(z) are analytic characteristic functions.
Let R be the radius of convergence of f(z); according to Cauchy’s
integral formula we have

02 = Fell
(0) a” 1M
= OH!
2
|LEI”
Nee we
where C is the circle |z | = R/2. Let M, = sup |f(z)|, then
C

0 = (Ze\UMs,
296 CHARACTERISTIC FUNCTIONS

where M, = 2M,/R, and we see from (10.1.10) that

ag) < (2R)l Wao (geet Vie. s):
It follows then that f;(z) = f;(t+7y) is an analytic characteristic function
which is regular at least in the strip |Im (z)| < My. One also sees easily
that the f;(z) have no zeros in this strip and that the relation (10.1.1) holds
in | Im (z)| < My}.
We introduce the functions
i(2) = fi(-is) = |"ed) (= 12-0049)
g(a) = f(—is) = | eaF (x)
where z = t+7y (t, y real). The integrals representing the functions g; (2)
converge at least in the circle |z| < My,‘ and the relation

(10.1.1) I Les(ayle = (2)

holds in this circle; the function g(z) is regular in the circle |z| < R.
In order to prove that the analytic characteristic functions f;(z) are
regular, at least in the strip of regularity of f(z), we must show that the
radius of convergence of the series expansion of f;(z) around the origin is
at least equal to R. We carry the proof indirectly and assume that at least
one of the series has a radius of convergence inferior to R. It is no restric-
tion to assume that f(z) has the smallest radius of convergence 7, < R.
Clearly 7, is also the radius of convergence of g,(z). We note that the
functions g,(z) have non-negative coefficients and conclude from Prings-
heim’s theorem™) that the point z = 7, is a singular point of g,(z). Let
A < r,/2 be a small positive number and put 7, = r,—A. We see from
(10.1.1a) that the relation

(10.1.1b) Ul[gi(ra+e)]” = g(ry+m)

is valid for sufficiently small |w |. The expansion of g; (7, +w) according to
powers of w has eee wae
coefficients and the coefficient of w” is
1
n! dw” 83 ere)
In order to obtain an estimate for this aoe we raise (10.1.1b) to the
power 1, differentiate m times, and put w = 0. We can again assume that
a; > 1 and therefore obtain a sum of positive terms and conclude that
d” d”
(10.1.11) —g(r,) >
rar: £3 (1+ ))
5 ida Uae OT
(*) See Titchmarsh (1939), p. 214, or Hille (1962), 1, p. 133, or Markushevich (1965),
1, p. 389.
a&-DECOMPOSITIONS 297
The function g(r, + w) is regular for |w| < R—r,.We putd = (R—1,)/2
and apply Cauchy’s integral to estimate the expression on the right of
(10.1.11). We see that

(10.1.12) 2 [g(r
+ w)]” mutt et =S A
[e(ra +)”
_ dy,
bs 2n1 wrt

Let

Ars) = oP |g(7, +2) |,

since (R+7,)/2 > r, +6 we see that
p74) < C(n)
where
C(n)= sup
lel =(R+1)/2
|a(z)|
is a positive function of 7, but does not depend on A.
It follows then from (10.1.12) that
d”

dw”
[s(r4+~)]” w=0
< ny a" @ ricer
o”

where C,(r,) is a function of 7, and is independent of A. The terms on the

left of (10.1.11) are all positive, so that
ad - Ate
tae [gi(7,+~)] bi <n! rs
as [Cy (r,)]".
The radius of convergence of the expansion of g;(r,+w) according to
powers of w is therefore not less than [C,(7,)]~! for any A € (0, 7,/2). But
the function g, has a singularity at the point z = 7,, so that the radius of
convergence of g,(r,+w) cannot exceed A. Select A so small that A <
[C,(r,)]-1. The assumption that 7, < R then leads to a contradiction. The
radius of convergence of f;(z) around the origin is therefore at least equal
to R, so that the characteristic functions f;(z) are regular, at least in the
strip of regularity of f(z). We have therefore established theorem 10.1.1
in the case where the distributions F(x), F',(x),..., /,(x) are symmetric,
and must now consider the general case.
For this discussion we need the following lemma which is of independent
interest.

Lemma 10.1.1. Let g(z) be an analytic characteristic function which has the
strip of regularity —« < Im(z) < B and choose a real n such that —« <
n < B. Then
298 CHARACTERISTIC FUNCTIONS

is also an analytic characteristic function which is regular in the strip

—a—n < Im(z) < B—7.
Let G(x) be the distribution function of g(t); the integral
Oe ie e-" dG(x)
exists and is finite and positive. We put

H(x) = 5) em ac(y)s
this is a distribution function and we see that
Lak iy 2 ‘iar
A(t) = C | Eee PAGhy = | eo" atl (x)

is its characteristic function. Moreover, it follows from the assumption that

2() is an analytic characteristic function that h(t) is also analytic and that
it has the strip of regularity —«—7 < Im(z) < B—7.
We now proceed with the proof of theorem 10.1.1 and suppose that the
conditions of the theorem are satisfied by the characteristic functions
T(t), fi), .- +5 f(t). It follows from equation (10.1.1) that

(10.1.13) TILi) fi(—l = FO F(—2)

is also valid. The characteristic functions [f(¢) f(—2)], [A() A(—2),.--,
[f:(t) fs(—t)] belong to symmetric distributions. We apply the result
which we obtained for symmetric distributions and see that the charac-
teristic function f;(t) f;(—7¢) is an analytic characteristic function and is
regular at least in the strip of regularity of f(t) f(—2). It follows then from
theorem 8.1.1 that f;(t) (j = 1, 2,..., 5) is also regular, at least in this
strip.
If f(t) is an entire characteristic function then this completes the proof
of the theorem, so that we need only consider the case where the strip of
regularity of f(z) has one or two horizontal boundary lines. We assume first
that f(z) is regular in the strip —a« < Im (sz) < 6 where « and # are both
finite. The strip of regularity of the analytic characteristic function
f(t) f(—t) is then the symmetric strip |Im (z)| < min («, 6) and it is no
restriction to assume that « < f.
We show this in the case where 6 < 3a; if 8 > 3a, the desired result can
be obtained by iteration.
We take 7) = (8—«)/2 and see from lemma 10.1.1 that
Ff(2+ m0)/F (eo)
is an analytic characteristic function which is regular in the strip |Im (z)| <
(6 +«)/2. Moreover it follows from (10.1.1) that
p= e f(t+ino)
=
iL fi(%o) F(%0)
a&-DECOMPOSITIONS 299
and we can conclude that the characteristic functions
fi(t +170)
Fi (™o)
(j = 1,2,...,s) are regularatleast inthe strip |Im (z)| < (8+«)/2. There-
fore f;(2+77) is also regular at least in this strip. We write w = x+in,
and see that the analytic characteristic functions f;(w) (j = 1,2,..., 5)
are regular at least in the strip —a < Im(w) < #. Theorem 10.1.1 is
therefore proved in the case where the strip of regularity of f(z) has two
horizontal boundary lines. We consider finally the case where f(z) is
regular in the half-plane Im (z) > —«. In this case we can start with the
strip (—«,«) and prove by induction that f;(z) is regular in the strip
—a < Im(z) < (2"—1)« for any n. Therefore f;(z) is regular in the half-
plane Im (z) > —« and the proof of the theorem is completed.
Theorem 10.1.1 was modified by Yu. V. Linnik [(1964) theorem 4.2.1]
in the following way.
Theorem 10.1.1a. Let f,(t), fo(t),..-, f(t) be arbitrary characteristic
Junctions and let «1, %,..., «; be positive real numbers. Let {t,,} be a sequence
of real numbers such that t, 4 0 and lim t, = 0. Suppose that f(z) ts an
k>o
analytic characteristic function which has no zeros in tts strip of regularity and
that the relation (10.1.1) holds at the points of the sequence {t,}. Then the
functions f;(t) (j = 1,2,..., 5) are analytic characteristic functions and are
regular, at least in the strip of regularity off (t), and the relation (10.1.1) holds
in this strip.
We mention next a theorem which shows that denumerable «-decom-
positions are possible.
Theorem 10.1.2. Let f;(t) be a sequence of characteristic functions and let
{a;+ be a sequence of positive numbers which are bounded away from zero
(i.e. there exists an x) > 0 such that a; > & for allj). Suppose that f(t) 1s an
analytic characteristic function which does not have any zeros in tts strip of
regularity and assume that the relation

(0114) TTP =f
holds in a real neighbourhood of the origin. Then the functions f;(t) are
analytic characteristic functions and are regular, at least in the strip of
regularity of f(z), and the relation (10.1.14) is valid in this strip.
This theorem also has a modification which corresponds to theorem
10si-la:

Theorem 10.1.2a. Let {f,(t)} be a sequence of characteristic functions and

let {x;} be a sequence of positive numbers which are bounded away from zero.
300 CHARACTERISTIC FUNCTIONS

Suppose that f (t) is an analytic characteristic function which does not have any
zeros in its strip of regularity. Let {t,} be a sequence of real numbers such that
t, £0 and lim t, = 0. Suppose that the relation (10.1.14) holds at the
ko
points t,; then the functions f,(t) are analytic characteristic functions and are
regular, at least in the strip of regularity of f(t), and the relation (10.1.14) is
valid in this strip.
The proofs of theorems 10.1.2 and 10.1.2a are very similar to the proof
of theorem 10.1.1 and are based on the same idea. One begins by sym-
metrizing the characteristic functions and proves first that the second
moments of the f;(t) exist and then, by induction, the existence of all
moments of the f;(t). This is done by dividing the (2k)th derivative of f(¢)
into three sums, exactly as in the proof of theorem 10.1.1. The proof of the
analyticity is also based on the idea used in the earlier proof and leads to a
majoration of the moments as in (10.1.1). The reasoning is, of course,
more complicated than in the case of the proof of theorem 10.1.1, since
infinite series occur and require a careful consideration of technical detail
to justify the necessary operations (such as the term-by-term differentiation
of the infinite series). The proof that relation (10.1.14) is valid in the strip
of regularity is obvious in the case of theorem 10.1.1 but calls fora more care-
ful discussion in the case of denumerable «-decompositions. In view of the
similarity of the demonstrations we omit here the proof of theorems 10.1.2
and 10.1.2a and refer instead to the book of B. Ramachandran (1967) where
the proofs are presented in full detail.
The preceding theorems of this section contain the assertion that the
characteristic function f(t) on the right-hand side of (10.1.1) and of
(10.1.14) does not vanish. It was shown by B. Ramachandran (1965) and
by R. Cuppens (1963b) that this restriction is superfluous [see also B.
Ramachandran (1967)].
Results which are in some respects similar to theorem 10.1.1a have been
obtained by several authors. We mention here a theorem which is due to
R. G. Laha (1960).
Theorem 10.1.3. Let {t,} (Rk = +1, +2,...) be a sequence of real numbers
such that t, > 0 while t_, = —t, for any k > 0 and lim t, = 0. Let f(t) be
k—>0o
a characteristic function and let y(z) be a function of the complex variable z
(z = t+1y; t, y real) which is regular in a circle about the origin. Suppose
that f(t.) = p(t.) for all k; then f(t) is an analytic characteristic function and
f (2) = v(z) in the strip of regularity.
For the proof we refer the reader to the paper quoted above. The follow-
ing particular case is sometimes of interest.
Corollary to theorem 10.1.3. Let f(t) be an even characteristic function and
let p(z) be a function of the complex variable z (xz = t+ty; t, y real) which is
&-DECOMPOSITIONS 301
regular in a neighbourhood of the origin and which is even for real values of the
argument. Suppose that {t,} is a sequence of real numbers such that lim t, = 0
k>o
and that f (t,) = y(t.) for all points t,. Then f(t) is an analytic characteristic
function and f(z) = »(z).

10.2 Special results concerning «-decompositions

In Section 10.1 we assumed only that the function subjected to an
a-decomposition is an analytic characteristic function. In the present
section we consider other assumptions concerning this function. We first
treat «-decompositions of entire characteristic functions and then study
the «-decomposition of the characteristic functions of lattice distributions
and of some infinitely divisible distributions.

Theorem 10.2.1. If the function f(t) of theorem 10.1.1 or of theorem 10.1.1a

is an entire function of finite order p, then each function f;(t) is an entire
characteristic function of order not exceeding p.
It follows immediately from theorem 10.1.1 that the f;(t) are entire
functions whenever f(t) is an entire characteristic function, so we have
only to prove the statement concerning the order of the functions f;(t).
We introduce the symmetric characteristic functions
a(t) = f)f(-2t) and git) = AOF(-4
and denote by a (j = 1,2,..., 5) the moment of order k of the distri-
bution corresponding to g;(t). Since g(z), 2, (2), ..., £s(%) are entire func-
tions, equation (10.1.13) holds for all complex z. We can therefore put
z = ly (y real) in (10.1.13) and get

(10.2.1) TLleo)" = e6).

We note that g(iy) and g; (zy) are real and that
She a ee 27
(10.2.2) ee (zy) = 2 (2k)! Y ean

It follows from (10.2.1) and (10.2.2) that

(10.2.3) [gy(in)* < ) (§= 1,2... 59).
We denote, as usual, by M(r; g;) and M(r; g) the maximum modulus of
the function g;(z) and g(z) respectively and see from theorem 7.1.2 that

(10.2.4) MG) By Bir) 8-17)

M(r;gi)= gi(ir) = g;(-7) (f = 1,2,..-,58).
It follows from (10.2.3) and (10.2.4) that
[M(r; gs)]” < M(x; 8).
302 CHARACTERISTIC FUNCTIONS

But this means that the order of g;(z) cannot exceed the order p of g(z).
We see finally from theorem 8.1.2 that the order of f;(z) cannot exceed p.
We next consider applications of these results.
Suppose that f(¢) = exp [jut—402?]. It follows then from theorem
10.2.1 that the functions f;(t) are entire functions of order not exceeding
2. Moreover, one sees from (10.1.1) that the f;(¢) are entire functions
without zeros and must therefore have the form
f(t) exp [enjt 44527) Gy = 2a 755).
We have therefore obtained the following theorem:

Theorem 10.2.2. Let f,(t), fo(t),..., fs(t) be arbitrary characteristic

functions and let a, %,...,%; be positive real numbers. Suppose that the
relation

I [Ai(@Q]” = exp [iut— 3072")

holds in a neighbourhood of the origin. Then the characteristic functions
ji(t) Gg = 1,2,..., 8) belong to normal distributions.
Theorem 10.2.2 is a generalization of theorem 8.2.1 (Cramér’s theorem).
This extension is due to A. A. Zinger and Yu. V. Linnik. D. Dugué (1957a),
(1957c) gave a different proof for theorems 10.1.1 and 10.2.2; he deduced
these results from properties of the products of positive powers of ab-
solutely monotonic functions.
We prove next an a-decomposition theorem for lattice distributions.
This theorem is an important tool in studying the «-decomposition of the
Poisson distribution but has also some independent interest.

Theorem 10.2.3. Let f,(t), fo(t),..., f(t) be arbitrary characteristic

functions and let a, %,..., %; be positive real numbers. Suppose that f(t)
is the characteristic function of a lattice distribution F(x) and that the relation

(10.25) TELA@I = f@
holds. The characteristic function f;(t) then belongs also to a lattice distribution
F(x). Moreover, if f(x) is an entire characteristic function without zeros
which belongs to a lattice distribution whose lattice points are the non-
negative integers, then each F(x) is a one-sided lattice distribution whose dis-
continuity points are contained in a set of the form u;+v(v = 0,1, 2,... ad
inf.) where
LX Xf; = 0.
en
The first assertion of the theorem is an immediate consequence of
theorem 2.1.4. If f(t) is a lattice distribution then there exists a real
&-DECOMPOSITIONS 303
to # 0 such that |f(t,)| = 1. But then necessarily |f;(¢.)| = 1 (j = 1,
2,...,5), so that f;(¢) also belongs to a lattice distribution.
To prove the second part of the statement we apply formula (7.2.3a) and
see that
(10.2.6) lext [F] = —lim (1/y) log f(ty) = 0
yoru

where

(10.2.7) ~(1/p) log f() = ¥ asl-(1/9) logf,

The f;(¢) are, according to theorem 10.2.1, entire characteristic functions
and
fs = —lim (1/y) log f,()
is the left extremity of F’;(x). We see from (10.2.6) and (10.2.7) that all yu;
are finite and that

GDi2:S eaebeSiac pli

j=l
Since F(x) has its discontinuity points at the non-negative integers we
havesy (ct) = lPand*therefore also /7;(27) |—"1 (f= 127s) lt is
then easy to see that the discontinuity points of F’;(x) are contained in the
set u;+¥, where vy runs through all non-negative integers.
The «-decomposition theorem for the Poisson distribution is an exten-
sion of Raikov’s theorem (theorem 8.2.2).

Theorem 10.2.4. Let f,(t), fo(t),...,fs(t) be arbitrary characteristic

functions and let a, %2,...,%5 be positive numbers. Suppose that the re-
lation

(10.2.9) TILA)" = exp Hle"-1)] @ > 0)

holds in a neighbourhood of the origin. Then
f(t) = exp [A;(e"—
1)+7; #)]
where 2; > 0 and mu; are real numbers.
To prove theorem 10.2.4 we set f(t) = exp [A(e”—1)] in theorem
10.2.3 and conclude that
(10.2.10) f(t) =e y pre

where sia

(GPA ee ae ayaa)
p? 20, Dp? =1
and inte

Xi %5 fly = 0.
j=1
304 CHARACTERISTIC FUNCTIONS

It follows from the last equation and the assumption (10.2.9) of the theorem
that
$ foe) Oj ;
(LOZAL ce wil: |Spo e*| = exp [A(e"—1)]
j=1 +7=0

holds for real t. We see from theorem 10.1.1 that the functions f;(t) are
entire characteristic functions and conclude from (10.2.11) that they have
no zeros. Relation (10.2.11) is therefore valid also for complex values of the
variable. We put w = e” (z complex) and write

83(w) = & pya”

and see easily that the g;(w) are entire functions without zeros. We can
then rewrite (10.2.11) in the form

(10.212) TE [g,(w)]* = exp [4(w—1)].

s

j=1
The power series for the functions g;(w) and also for the function
g(w) = exp [A(w—1)] have non-negative coefficients, so that
M(r; gi) = 8:(7)
and
M(r; g) = g(r).
Since for r > 1 the functions g;(r) and g(r) are increasing functions of 7,
we see that
&i(r) > (1) = 1
and conclude from (10.2.12) that
8i(r) = M(r5 85) < exp Ar—1)]
for r > 1. The function g;(w) is therefore an entire function of order not
exceeding 1. Since it has no zeros, we conclude from Hadamard’s factoriza-
tion theorem that it has the form
§3(w) = exp (A; w+ B;)
or, since g;(1) = 1,
&;(w) = exp [A;(w—1)].
We see finally from (10.2.9) that
f; (t) = exp [A, (e*—1) +ip,;t]
so that the theorem is proved.

Remark. 'The statements of theorems 10.2.2, 10.2.3 and 10.2.4 are also
valid if one replaces the assumption that (10.2.4) [respectively (10.2.5) or
(10.2.9)] is valid in a real interval containing the origin by the assumption
that these relations are satisfied in the points of a sequence {¢,} such that
&-DECOMPOSITIONS 305
t, # 0, while lim ¢, = 0. It is also possible to derive corresponding de-
k—>oo
numerable «-decomposition theorems.
Yu. V. Linnik (1959) obtained an «-decomposition theorem for infinitely
divisible characteristic functions.

Theorem 10.2.5. Let f,(t), fo(t),..-,f.(t) be arbitrary characteristic

functions and let ,, %,..., x; be positive numbers. Let {t,} be a sequence of
real numbers such that t,, # 0 and lim t, = 0. Suppose that f(t) is a charac-
k—>0
teristic function of the class which has a bounded Poisson spectrum and that
the relation

(10.2.13) I Lal” = f(e)

holds at all points of the sequence {t,'. Then the f;(t) belong also to and the
relation (10.2.13) holds for all t. The functions f;(t) are infinitely divisible
entire characteristic functions and their spectra are subsets of the spectrum of
f(2). |
For the proof we refer to Linnik (1959) or to the monograph of Linnik
(1964).
We also note that the proof of theorem 9.2.1 uses only the ridge property
of analytic characteristic functions. Using this fact, an «-decomposition
theorem corresponding to theorem 9.2.1 can be derived.
We conclude this section by mentioning a few additional results of the
type discussed in this chapter.
R. Cuppens (1963b) and B. Ramachandran (1965) have proved a de-
numerable «-decomposition theorem for the convolution of a binomial
and a Poisson distribution with the same span. R. G. Laha and E. Lukacs
(1962) considered the situation of theorem 10.1.1 but replaced the assump-
tion that f(t) is an analytic characteristic function by the premise that f(t)
has derivatives up to the order 2N and obtained a finite «-decomposition
theorem. A corresponding denumerable «-decomposition theorem is due
to R. Cuppens (1963a).
 11 BOUNDARY CHARACTERISTIC FUNCTIONS

In Chapter 7 we introduced analytic characteristic functions and studied

their properties. In the present chapter we deal with characteristic functions
which are boundary values of analytic functions.
We say that a characteristic function f(t) is the boundary value of an
analytic function if there exists a complex-valued function A(z) of the
complex variable z = t+zy which is regular in the rectangle
tere
ASO <oy pt
[respectively in the rectangle {|t| < A, —a < y < 0}] and has the
property that en A(t+iy) = f(t) for |t| < Aandy > 0 [respectively for
|t| < A and y© 0]. The class of characteristic functions just described
includes the class of analytic characteristic functions but is more extensive.
For the sake of brevity we will call ‘boundary characteristic functions”
those characteristic functions which are boundary values of analytic
functions without being analytic characteristic functions. If f(t) is a bound-
ary characteristic function then we can extend its definition to complex
values of the variable by writing f(z)= A(z).

11.1 The integral representation

In this section we derive a number of properties of boundary charac-
teristic functions which are similar to results for analytic characteristic
functions obtained in Chapter 7. We give first a necessary and sufficient
condition which a distribution function must satisfy in order that its
characteristic function be a boundary characteristic function.

Theorem 11.1.1. Let F(x) be a distribution function and f(t) be its charac-
teristic function. The function f(t) is the boundary value of an analytic
function A(z) (2 = t+ty; t, y real) which is regular in the rectangle

|t] < A,0 < y < bof, and only if, the integral | e—” dF (x) exists and is

finite for 0 < y < b but does not exist for y < 0.
We first prove that the condition of the theorem is sufficient and assume
therefore that

|fxe-¥* dF (x)
BOUNDARY CHARACTERISTIC FUNCTIONS 307
is finite for 0 < y < b. Let

nla) = |geedF(x)
(11.1.1) —e
2.(2) = ieeo dix):

The function g,(z) is regular for Im(z) < 6, while g,(z) is regular for
Im (z) > 0. Therefore

A(z) = &1(2)+82(2) = | dF (a)

is regular in the strip 0 < Im (sz) < 5; moreover,

lim A(é-+iy) = | dt dF (x) = f (i).

y}0 —2
To prove the necessity of the condition we assume that f(z) is a boundary
characteristic function. Let g,(z) and g.(z) again be given by (11.1.1), then
(11.1.2) f(t) = gi(t)+82(4).
The function f(z) isregularinarectangle, say in D, = {|t| < A,0< y < bd},
while g.(z) is regular in the upper half-plane. Therefore g, (z) is regular in
D,; on the other hand, it follows from the definition of g, (z) that it is regular
in the lower half-plane, so that g, (2) is regular in the rectangle D, which is
symmetric to D, with respect to the real axis. It follows then from Schwarz’s
symmetry principle (see Appendix E) that g, (2) is regular in the rectangle
D = {|t| < 6,| | < 5}. We first assume that F(0) #4 0 and consider the
function

&1 (t) 1 it
= ——— ear (x):
Miers >
This is an analytic characteristic function whose strip of regularity con-
tains the strip |Im (z)| < 5. Therefore

silo) = | er aria)
0

exists and is finite for |y| < 5. Since the integral | e "dF (x) exists for
0

y > 0, we see that | e ”"dF (x) exists and is finite for 0 < y < b, so
the necessity of the condition is proved if F(0) 4 0. In the case where
F(0) = Owe note that f(¢) = g,(t) and obtain the necessityof the condition.
A criterion analogous to the statement of theorem 11.1.1 holds for
characteristic functions which are boundary values of functions regular in
a rectangle {|t| < A, —a < y < 0} contained in the lower half-plane.
The argument used in the proof of theorem 11.1.1 indicates that a
308 CHARACTERISTIC FUNCTIONS

boundary characteristic function can be represented by a Fourier—

Stieltjes integral in a horizontal strip. Let 0 < y < f be the strip of
greatest width in which the boundary characteristic function f(z) admits
the representation
(iti2 8 HG) = |et dF (a) (0 < Im(e) < f).
The strip 0 < Im(z) < £, in which (11.1.2) is valid, is called the strip of
regularity of the boundary characteristic function f(z). The validity of the
representation (11.1.2) is the reason for the similarity of many properties
of boundary characteristic functions and of analytic characteristic functions.
The discussion of these properties is facilitated by the following lemma:

Lemma 11.1.1. Let f(z) be a boundary characteristic function with the strip
of regularity 0 < Im(z) < B and choose a real y such that 0 < 9 < f.
Then
f(z+™)
es
1 i Be
ee
h(z) =
nme:

is an analytic characteristic function which is regular in the strip —n <

Im (2) < B—7.
Lemma 11.1.1 is analogous to lemma 10.1.1 for analytic characteristic
functions and is proved in the same way. The distribution function
corresponding to h(z) is
(tia Ail a|." aF (9)
where C = f(im) and F(x) is the distribution function corresponding to
F(z).
Corollary 1 to theorem 11.1.1. The strip of regularity of a boundary
characteristic function f(z) has one or two horizontal boundary lines. One of
these 1s always the real axis. The purely imaginary points on the boundary are
singular points of f(z).

Corollary 2 to theorem 11.1.1. Boundary characteristic functions have the

ridge property; moreover the zeros and the singular points of boundary
characteristic functions are located symmetrically with respect to the imaginary
axis.

Corollary 3 to theorem 11.1.1. A boundary characteristic function has no

zeros on the segment of the imaginary axis located in its strip of regularity.

Corollary 4 to theorem 11.1.1. Let f(z) be a boundary characteristic

function whose strip of regularity is 0 < Im(z) = y < B. Then log f (iy) is
convex for0 < y < B.
BOUNDARY CHARACTERISTIC FUNCTIONS 309
Corollary 5 to theorem 11.1.1. Let F(x) be a distribution function which has
a boundary characteristic function with strip of regularity 0 < Im(z) < f.
Then

B = —lim sup log #)e

x00 x
Similarly, iff(z) has —« < Im(z) < 0 as its strip of regularity then
log [1 oe
«= —lim cup {
wo x
These corollaries follow immediately from lemma 11.1.1 and the corre-
sponding theorems of Chapter 7.

Theorem 11.1.2. Let F(x) be a distribution function and f (t) its characteristic
function. F (x) ts bounded to the left (or right) if, and only if, f (t) ts regular in
the upper (respectively lower) half-plane and if
| f(2)| < 2
for some c > 0 and Im(z) > 0 [respectively Im (z) < 0]. The extremity of
F(x) 1s given by lext [F] = —lim y-? log f(ty) [respectively rext LF] =

lim y~" log f(—7y)].

yon

The formulae for the extremities were derived in Chapter 7 under the
assumption that f(z) is an analytic characteristic function. We now see that
they are also valid if this restriction is dropped.
To prove theorem 11.1.2 we first note that the given distribution F(x)
and the distribution H(x) which is defined by (11.1.4) have the same
extremities. The statement follows immediately from lemma 11.1.1 and
from theorem 7.2.2 and its corollary.
Remark. A one-sided distribution function has either an analytic
characteristic function or a boundary characteristic function.
We next consider factorizations of boundary characteristic functions.

Lemma 11.1.2. Let F(x), F,(x) and F,(x) be distribution functions and
f(t), fi) and f, (t) their characteristic functions. Suppose that iW e~” dF (x)

exists and is finite for 0 <y <b. If F=F,*F, then eB integrals

lie e- dF;(x) (j = 1,2) exist and are finite for 0 < y < 6, and the
relation
|” ¢-"dF(x) = |"et dF (x) [ie war, (x)
holds for 0 < 4 SDs "
310 CHARACTERISTIC FUNCTIONS

The proof of the lemma is completely analogous to the proof of the

convolution theorem.
Theorem 11.1.3. Let f(t) be a boundary characteristic function with strip
of regularity 0 < Im(z) < B. Then any factor f,(t) of f(t) is regular, at
least in the strip of regularity of f(t).
Since f, (#) is a factor of f(z), there exists a characteristic function f, (2)
such that f(¢) = f(t) fo(t). Let F(x), F(x) and F(x) be the distribution
functions which correspond to f(x), f,(x) and f.(*) respectively. Then
F = F,* F,. We see from theorem 11.1.1 that the integral | endl)
is finite for 0 < Im(z) < f, and we conclude from lemma 11.1.2 that the
integrals | e7* dF (x)"(7'= 1, 2) are alsovfinite for 0 < Im (@) <5:
Therefore f;(z) (j = 1, 2) is also regular at least in this strip; moreover
we see from lemma 11.1.2 that the equation f(zy) = f, (zy) fe (ty) holds in
the strip of regularity of f(z). The relation
f(z) = fi(2) fa(2);
valid in the strip 0 < Im(z) < £, then follows by analytic continuation.
Theorem 11.1.3 is analogous to theorem 8.1.1 for analytic charac-
teristic functions; it shows that the factors of boundary characteristic
functions are either analytic characteristic functions or boundary charac-
teristic functions.
We conclude this section by mentioning another property of boundary
characteristic functions noted by A. Zygmund (1951).
Theorem 11.1.4. Let {F,,(x)} be a sequence of distribution functions and let
{fn(t)} be the corresponding sequence of characteristic functions. Suppose that
the f,,(t) are boundary characteristic functions and that they are regular in a
fixed strip (independent of n). The sequence of distribution functions converges
weakly to a distribution function F(x) 1f, and only if, the following two con-
ditions are satisfied: (i) the functions f,,(t) converge to a limiting function
J (t) in a fixed interval around the origin; (ii) f(t) 1s continuous at t = 0.

11.2 Infinitely divisible boundary characteristic functions

In this section we study boundary characteristic functions which are
infinitely divisible and one-sided infinitely divisible distributions. As an
example of such a characteristic function we mention the stable distribution
with parameter « = y = $, whose frequency function was given by formula
(5.8.9).
Lemma 11.2.1. An infinitely divisible boundary characteristic function has
no zeros in the interior of its strip of regularity.
The lemma corresponds to theorem 8.4.1 and is proved in the same way.
BOUNDARY CHARACTERISTIC FUNCTIONS mi

The next theorem shows that a canonical representation of infinitely

divisible distributions also holds in the strip of regularity of boundary
characteristic functions. This is similar to the statement of theorem 8.4.2.
Since a boundary characteristic function does not necessarily have a finite
second moment, the representation cannot be the Kolmogorov representa-
tion used in theorem 8.4.2. We will use as our starting-point the Lévy
canonical representation and will have to modify our reasoning somewhat.

Theorem 11.2.1. Let f(z) be an infinitely divisible boundary characteristic

function. Then the canonical representation
2 —0 .
(11.2.1) log f(2) = iza—Z 2+ | (aa )amen
— oo 1+4u?

oon ig’ 13u

+| (« 1 ips )ave
+0 ue
is valid in the interior of the strip of regularity of f(z). The constants a and
o* and the spectral functions M (u) and N(u) satisfy the conditions of theorem
Byo22.
We first assume that 0 < Im (z) < f is the strip of regularity of f(z). It
follows from lemma 11.2.1 that log f(z) is regular in 0 < Im(z) < # and
continuous in 0 < Im(z) < #. Since f(¢) is infinitely divisible it admits
the Lévy representation (11.2.1) for real t. We write for Im (z) > 0
izU
(11.2.2), (2) = le (e—1- sae) dN (u).
The function ¢, (z) is regular for Im (z) > 0 and continuous for Im (z) > 0.
We put for Im(z) < 0

Hieso, (2), = a (e-—1- — 0 u?

)am
the function ¢,(z) is regular for Im (z) < 0 and continuous for Im (z) < 0.
Let
(11.2.4) y(z) = log f(z)—1az— 4307 27— ¢, (2).
It follows from our assumptions that y(z) is regular for 0 < Im(z) < B
and continuous for 0 < Im(z) < f. Since f(t) is infinitely divisible we see
from its Lévy canonical representation that
y(t) = $2(t)
for real ¢. We conclude then from Schwarz’s reflection principle that (2) is
the analytic continuation of $2(z), so that ¢,(z) is regular in Im (z) < f.
Therefore ¢,' (z) is regular in Im (z) < f. For Im(z) < 0 we have from
(11.2.3) the integral representation
(2) = -|~_étutdM (wi),
312 CHARACTERISTIC FUNCTIONS

We introduce the non-decreasing function

TONE if o2dM(v) (u < 0)

and see that

AEC a |“ed L(u).

The function ¢’(z) is—except for a constant factor—an analytic charac-
teristic function with strip of regularity Im (z) < 6; therefore its repre-
sentation by a Fourier integral is valid in this strip. We then see easily that
the representation (11.2.3) can be extended and is valid for Im (z) < f.
Since y(z) = ¢2(z) in0 < Im(z) < f we see from the equations (11.2.2),
(11.2.3) and (11.2.4) that log f(z) admits the representation (11.2.1).
The proof follows the same lines if the strip of regularity of f(z) is given
by —o=< Im(2)— 0,
We turn now to the study of infinitely divisible one-sided distributions.
In this connection we need the following result:

Lemma 11.2.2. Let F(x) be a distribution function which is bounded from

the left (right) and let f(t) be its characteristic function. Then the factors of
F(t) belong also to distribution functions which are bounded from the left
(right).
The lemma follows easily from theorem 7.2.2 and lemma 11.1.1.

Theorem 11.2.2. Let f(t) be the characteristic function of an infinitely

divisible distribution function F (x). The distribution function F (x) 1s bounded
to the left if, and only if, the following three conditions are satisfied:
(i)tot==20)
(ii) M(u) zs constant for u < 0

Gi) [wan) Ass

Here o* 1s the constant, M(u) and N(u) are the spectral functions occurring in
the canonical representation (11.2.1) off(t). If conditions (i), (ii) and (iii) are
satisfied, then
EASE
lext [F] = a-| 142 tN@.
We first prove that the conditions are necessary.
(i) If o? > 0, we see from (11.2.1) that
F(t) = exp (— 2072”) g(Z)
where g(t) is the infinitely divisible distribution without normal factor
which is determined by the constant a and the spectral functions WM(u) and
BOUNDARY CHARACTERISTIC FUNCTIONS 313

N(u). This means that f(t) has a normal factor, in contradiction to lemma
t2Z.2:
(11) If M(u) is not constant for u < 0, then there exists a finite interval
[a, 6], a < b < 0 such that C = M(b)—M(a) > 0. The function

az) ==AXP |Ie(e"—1) am()|

is an entire characteristic function which is a factor of f(z). But g(z) cannot

satisfy the conditions of theorem 7.2.2, so that g(t) cannot be the charac-
teristic function of a distribution bounded to the left. This contradicts
lemma 11.2.2, so that the necessity of (ii) is proved.
For the proof of the necessity of (iii) we can therefore assume that (i)
and (ii) are valid, and we see from (11.2.1) that for y > 0

log f(ty) = -ay+ |(e-™—1+ yu) dN (u)—-y{ ie AO)

We write
fo (rote) an
Te-™ _14yu
R(y) = [. dN (u)

and get

(11.2.5) log fl) >-ay+yR(9)-y| ———dN (u)— { dN (u).

We give an indirect proof for (iii) and assume tentatively that

(11.2.6) { wan a
If we can show that (11.2.6) implies
(11.2.7) lim R(y) =
yoo
1 ,
then it follows from (11.2.5) that lim 7 log f(iy) = oo. In view of theorem
yr

11.1.2, this contradicts the assumption that f(z) belongs to a distribution

bounded from the left. The necessity of (iii) will therefore be established
as soon as we show that (11.2.7) follows from (11.2.6).
Let H(u) be defined by

Hani | eave) (u > 0).

Then H(u) is non-decreasing and we see from (11.2.6) that
(11.2.8) H(0) = -
314 CHARACTERISTIC FUNCTIONS

It is easily seen that (e-*—1+.)/x is positive and non-decreasing in

(0, + 00). Therefore
dH(u) > |= *)
Ur aU:

R(y) = (:oe
yu
31 2.
so that

R(y) 2 e> [xa)-a()].

It follows from (11.2.8) that lim R(y) = + 00, so that the necessity of (111)
yoo
is completely proved.
We prove next the sufficiency of the conditions (i), (ii) and (iii). We
assume that these conditions hold, and we obtain, for real t, the canonical
representation
itu
log f(t) = tat+ [(e- 1- i a) dN(u).
If we replace in the integral the real variable ¢ by the complex variable
z = t+ty then we obtain a function which is regular in the half-plane
y > 0 and continuous in y > 0. Therefore f(t) is either an analytic
characteristic function or a boundary characteristic function, and we see
from theorem 11.2.1 that the canonical representation is valid in the upper
half-plane. Therefore
(11.2.9) log f(zy) = -ay+ | (em oS
“Jan (y2 0).
We put

A) = |<
and see from (11.2.9) that
(11.2.10) — log f(y) = -ayty| >
qa dN (u)+yQ(y).
It follows easily from a aaa (iii) of theorem 11.2.2 that

\oa
ae— dN(u) < ©
and that
lim O(y) = 0.
yY> oo

Therefore we conclude from (11.2.10) that

‘al : ss
Next] c= Mas AUD, -a-| a aN),
aaa
*

so that the proof of the theorem is completed.

A theorem concerning infinitely divisible distributions which are
bounded to the right can be stated and proved in the same way.
 12 MIXTURES OF DISTRIBUTION FUNCTIONS
AND TRANSFORMATIONS OF CHARACTERISTIC
FUNCTIONS

In this chapter we discuss briefly certain integral transforms of distribu-

tion functions. These transformations can be used to construct new
characteristic functions from given characteristic functions.

12.1 Mixtures of distribution functions

Let G(x), G,(x),...,G,(x) be m distribution functions; we saw in
Section 2.1 that

F(x) = j=13% a;G;(x)

is also a distribution function, provided that a; > 0(j = 1,2,...,m) and
»S a; — if.
j=1
We can regard F(x) as a mixture of the distribution functions G,(x),...,
G,,(x) with weights a,,..., d,. In the present section we consider a more
general mixing procedure and apply it also to the corresponding character-
istic functions.
Let G(x, y) be a family of functions which has the following properties:
(i) for each value of y the function G(x, y) is a distribution function
in x;
(ii) G(x, y) is a measurable function of y.
The functions G(x, y) form a family of distribution functions which
depends on a parameter y. In the following we consider only families
{G(x, y)} of distribution functions which satisfy conditions (i) and (11). An
exhaustive discussion of mixtures of distributions was given by H. Robbins
1948).
: Let H(y) be an arbitrary distribution function; we form the expression

(12.1.1) F(x) = ee y) dH (y)

and see easily from the dominated convergence theorem [Loéve (1963)
pp. 124-127] that F(x) is a distribution function. The corresponding
characteristic function is then given by

(121.2) fi)= | a(t.9)4H (9)

316 CHARACTERISTIC FUNCTIONS

where

sty) = i. ed, G(x, ¥).

Mixtures of distributions occur in a great variety of practical applications
and were used by W. Feller (1943) to study contagious distributions.
Two particular cases of (12.1.1) are of some interest.
(1) If G(x, y) is a purely discrete distribution,
G(x, y) i x P,(y) e(x—€,),

then
P(x)= | pea (y)]e(o-8)
is also a discrete distribution. The distribution which is obtained for
om
Bayt
P(y)=e%=,
gist & ee
=v (v= 0,1,2,...)
is called the compound Poisson distribution.
(LIS
H(y) = & poe(y—me)
is a purely discrete distribution then (12.1.1) yields
F(x) = X pv G(x, m)-
We put here :

and let G(x, v) = [G(x)]’* be the v-fold convolution (+) of G(x)

with tse (Vi= 015 2522. )al hen se

F(x) = Se*~ [oy

»=0 U:

is called the generalized Poisson distribution. Its characteristic

function is f(t) = exp {A[ g(t)—1]}.
In the following we consider a slightly more general mixture of distribu-
tion functions. The process still has the form (12.1.1) but we relax the
assumptions concerning the weight function H(y) in so far as we do not
require that H(y) be a distribution function. We will only assume that
H(y) is a non-decreasing function whose total variation is equal to 1. We

(t) Thus G («, v) is defined by G(x, 1) = G(x) and the relations

G(x,v) = i G(x—y, v-1)dG(y)

forlOn—a2 73.
MIXTURES AND TRANSFORMATIONS 317
therefore admit the possibility that H(— 00) # 0 and make no assumptions
concerning the values of H(y) at its discontinuity points.

Theorem 12.1.1. The function

(12:13) ".<F (x) = | G(x, y)dH(y)

is a distribution function whenever {G(x, y)} ts a family of distribution
functions which has the properties (i) and (ii) if, and only if, the conditions
(a) H(y) ts non-decreasing

(b) |dH(y) = 1
are satisfied.
We note that the characteristic function of F (x) is

(1214) f= | a(t. 9)dH(y),

where g(t, y) ts the characteristic function of G(x, y).
The sufficiency of the conditions follows easily from the properties of the
family G(x, y) and from the dominated convergence theorem. To prove
that conditions (a) and (b) are necessary we specialize the family G(x, y).
Let 7, and 7, be two arbitrary real numbers such that 7, < 7, and select
G(x, y) = ee —1) ) ifm<ys
pt y Ne

a(x) ify <m or y> MH.

It is then easily seen that
F(x) = e(x—1)[H (m2)—A(m)] +e) — A (m2)
+ A(m)].
If F(x) is a distribution function then we must necessarily have
H(2)—H(m) 2 0
which proves (a). To prove (b) we select G(x, y) = e(x) for all y and see that

P(x) = e(x)|dH(y)
so that (b) must be satisfied.
We remark that condition (a) is necessary only to assure that the weight
function should produce a distribution function, whatever family, satis-
fying (i) and (ii), is used in (12.1.3).
We illustrate this by an example where a non-monotone weight function
is used to transform a suitably chosen family into a new distribution. Let
1
318 CHARACTERISTIC FUNCTIONS

be the weight function and consider the family of normal distributions

with mean y and standard deviation 2a, that is(T)
Bet ic
G(x, y) = 0 )

re ~ eqtarculelic)(e)-vente)]
Phen

This is an absolutely continuous function with derivative

1
©) = 2a enVQ)—1]
ye 6

x {exp [—("—1)*/8a*] + exp [—(w+1)?/8a] — /(2) exp [-x*/8a*]}.

The function F(x) is a distribution function whenever F’ (x) > 0 for all
x; it is easy to show that this condition is satisfied if
2 >
i
————-
Oe" 4 ogi
The characteristic function of F(x) is, according to (12.1.4), given by

f(t) = is |exp (—2a???).

12.2 Transformations of characteristic functions

We have already noted that the mixture of distribution functions in-
duces a mixture of the corresponding characteristic functions. We now
use the results of Section 12.1 to discuss certain transformations of
characteristic functions.

Theorem 12.2.1. Let {f,(t)} be an arbitrary sequence of characteristic

functions and {a,} be a sequence of real numbers. The necessary and sufficient
condition that
(12.24). f@)= » Ay f(t)
should be a characteristic function for every sequence of characteristic functions
is that
(12.22) a5: 20, .)) ae.
v=0

This follows from theorem 12.1.1 if we put

A(y) = x a, e(y—V).
(t) Here

Ox) = sr 3 eee ne rs
MIXTURES AND TRANSFORMATIONS 319

Next we let g(t) be an arbitrary characteristic function and write

fo (t) = [g(®]? for v = 0, 1, 2,.... We obtain immediately the following
corollary:

Corollary to theorem 12.2.1. Let g(t) be a characteristic function and let

A(z) be a function of the complex variable z which is regular in |z| < R,
where R > 1. The function A[ g(t)] is also a characteristic function if, and
only if, A(z) has a power series expansion about the origin with non-negative
coefficients and if A(1) = 1.
The corollary can also be derived directly from theorem 12.1.1 if we set
H(y) = Diao y = 2)
and G(x, y) = [G(x)]™ forv < y < v+1.
An interesting generalization of the corollary to theorem 12.2.1 has
been derived by C. S. Herz (1963) and also by A. G. Konheim-B. Weiss
(1965).

Theorem 12.2.2. Let A(z) be a function of the complex variable z which has
the property that f(t) = A[g(t)] ts a characteristic function whenever g(t) is
a characteristic function. Then A(z) can be represented by a series, convergent
for |z| < 1, which has the form
io.@) ice)

A(z) aa y v3 ann a” 3”
m=0 n= 0

where the a m,n are real and where a m,n 20 Gnd Siew Ci le
m=0 n=0

Clearly the function A(z) need not be regular; a simple example is the
function A(z) = |2|?. For the proof of theorem 12.2.2 we refer the reader
to the papers quoted above.
In a subsequent paper A. G. Konheim-B. Weiss (1968) investigated
transformations of non-negative definite functions into infinitely divisible
non-negative definite functions.({) They obtained the following result,
formulated here in terms of characteristic functions.

Theorem 12.2.2a. Let 0(z) be a complex-valued function of the complex

variable z. The function 6(z) has the property that f(t) = O[g(t)] is an
infinitely divisible characteristic function whenever g(t) is a characteristic
function if, and only tf, 6(2) = exp [cA(z)—1], where c ts a posttive constant,
while A(z) is the function defined in theorem 12.2.2.
The sufficiency of the condition follows from theorem 12.2.2 and De
Finetti’s theorem; for the proof of its necessity the reader is referred to the
paper quoted above.
(+) These authors worked in the more general framework of locally compact Abelian
groups having elements of arbitrarily high order.
M
320 CHARACTERISTIC. FUNCTIONS

Theorem 12.2.3. Let g(t) be an arbitrary characteristic function and let p be

a real number such that p > 1; then

he seen
is an infinitely divisible characteristic function.
Let be an arbitrary integer and put ©

H(9) = ¥ a2)
where
* = 4 —» (+n)\(1+20)...01+(—1)n]
ay = Po Wile: “a, = a (np)!

fork = 1,2,...and set G(x, y) = [G(x)]”* if

v < y < v+1. This shows
that

Ok ese
is a characteristic function for any positive integer 1; in other words, f(t)
is an infinitely divisible characteristic function. Theorem 12.2.3 follows
also from the corollary to theorem 12.2.1 if we put

i Qhakele
and understand that A(z) is the principal value of this power.
We next discuss a few additional transformations. Let V(«) be a non-
decreasing function of bounded variation defined on the interval [0, /]
and let

aa) 2 | e dV (x).
Then the function g(z)/g(1) satisfies the conditions of the corollary of
theorem 12.2.1. Suppose that f(¢) is a characteristic function; then
na) - Hf)
g(1)
is also a characteristic function.
The transformation given by theorem 12.2.1 was derived by using a
step function H(¥) as the weight function of a mixture. We next specialize
Hy) in a different manner and assume that H(y) is a non-decreasing
function such that H(0) = 0 while H(1) = 1. Moreover we suppose that
g(t, y) is a function of the product ty so that g(t, y) = g(ty) where g(u) is
some characteristic function. We then obtain from theorem 12.1.1 the
following result.
MIXTURES AND TRANSFORMATIONS 321

Theorem 12.2.4. Let g(u) be an arbitrary characteristic function and suppose

that H(¥) is a non-decreasing function such that H(y) = 0 if y < 0, while
Hf(y) = 1afy > 10 Then

f(t) = |e(ty)dH(y) = |au) aH (u/2

1 t

is also a characteristic function.

Theorem 12.2.4 can also be used to derive transformations of charac-
teristic functions which are of some interest. We give some applications of
theorem 12.2.4.

Corollary to theorem 12.2.4. Let g(u) be an arbitrary characteristic

function and p > 1 a real number. The function
t
(12.2.3) f(t) = I g(u)u?—*du
0

is then a characteristic function.

The corollary follows from theorem 12.2.4 by putting
O.gnitripe<0
Hy) = <9" i 0 ey <1 (pp 2.1)
L fifaesyt.
If we put in the corollary p = 1, we see that the function

(12.2.3a) f(t) = ao) Wi

is a characteristic function whenever g(w) is a characteristic function. The
transformations (12.2.3) and (12.2.3a), as well as some generalizations,
were investigated by M. Girault (1954) and by H. Loeffel (1956). A. Ya.
Khinchine was the first to study the transformation (12.2.3a), and his
results have been presented in Section 4.5. We apply the technique of
mixtures to construct operators which transform characteristic functions
into other characteristic functions.
Let F(x) be a distribution function which has a finite second moment
be = |‘i x? dF (x) and let f(¢) be the characteristic function of F(x). ‘Then

H(s) == |"9*dF@)
is also a distribution function. We use in the following H(«) as a weight
function for certain mixtures.
Gees oe
(a) Let g(u) = Gare: be the characteristic function of a rectangular
322 CHARACTERISTIC FUNCTIONS
distribution. ‘Then
. Iinfemer4
k(t) = | s(ex) dt4(x) = = lhe rg dF (x)
2dF

is, according to theorem 12.2.3, a characteristic function. A simple com-

putation shows that

tf” (0)
(b) Let g(u) = e’”; the same procedure shows that
h(t) = [= dF (x) = f i
is a characteristic function.
(c) We now modify our assumption and suppose that F(x) is bounded
to the left with lext ['] = 0 and that the first moment «, of (x) exists.
Then
0 ie ead
H,(x) =
=| ydF(y) ifx 20
a, J0

is a distribution function. If we use H,(x) as a weight function for

g(u) = é”, we see that
| ean, (@) = =| eit? x dF (x) = ‘
is also a characteristic function.
oe
(d) The function g(t) = g ; ; is the characteristic function of a rect-
7
angular distribution over the interval [0, 1]. For this distribution the left
extremity is zero and the mean equals 4, so that, according to (c),

B0.= ‘(@t)) 21 +ite*—e*

g (0) t
is a characteristic function. Therefore
Dev ah
Ee ee) saa5 a!
is also a characteristic function. We again use H(«x) as a weight function
and see that
R(t) = =F | Nerd (ye TO) xc it 1 —itx) dF (x)
so that

Rt) = 5am [f(¢)—1—2f"(0)]

is a characteristic function.
MIXTURES AND TRANSFORMATIONS 320

We summarize these results.

Theorem 12.2.5. Let T,, T,, Ts and T, be operators defined by

Qed Mew fO-f'0) teen,

(b) T: f(t) =
f'()
f"(0)
(c) T; f(t) =
fo)
f(0)
(2) TO = apr 2 UO-1-4 O
Suppose that the domains of these operators are characteristic functions f(t)
which satisfy the following conditions:
In cases (a), (b) and (d), the f(t) are characteristic functions whose distri-
butions have finite second moments; in case (c), the f(t) belong to distributions
with finite first moment and left extremity at the point 0.
These operators transform characteristic functions satisfying these con-
ditions into other characteristic functions.
Remark. If we apply an operator 7; (¢ = 1,2, 3,4) to an analytic
characteristic function contained in its domain, then the transformed
function is also in the domain of T;, so that its application can be iterated.
As an example we consider the characteristic function f(t) = e~"/*. The
functions
ral
le heeed .i e 5/2 = (=)!
-— H,,(t) e —t?/2
Hor
and

Ra (t) = 7; Te-le-#/2 — a) Aap (4) 1/2

Kor t

are then characteristic functions. Here H;,(«) is the Hermite polynomial of

degree k defined by the relation

H, (2) = = (e-*),
2 ae 2

while a, = (2k)!/2*k! is the (2k)th moment of the normal distribution

with mean zero and variance 1.
The characteristic functions h,,(t) are indecomposable, while the
h,_,(t) are decomposable and always have a normal factor. The functions
hy,(t) and hy,_,(t) belong to the class of entire characteristic functions of
order 2 which have only a finite number of zeros. The factorizations of
this class are completely known [see E. Lukacs (1967)].
324 CHARACTERISTIC FUNCTIONS

We mention two other transformations which are obtained as mixtures

of characteristic functions.

Theorem 12.2.6. Let F(x) be a distribution function with characteristic

function f(t). Suppose that F (x) is bounded to the left and that lext [F] > 0.
If exp [y(t)] is the characteristic function of an infinitely divisible distribution
and if 2 is a positive number, then g(t) = f[—zAy(t)] ts a characteristic
function.

Since f(u) = | é’ dF (y) we see that

0

sit) = fl—iay(d)] = |”exp [ayy] 4 (9).

According to our assumption, exp [Ayy(t)] is a characteristic function
for y > 0, so that g(t) is a mixture of characteristic functions, and the
statement follows from theorem 12.1.1.
A particular case of some interest is obtained by putting y(t) = h(t)—1
where H(t) is an arbitrary characteristic function. It follows from lemma
5.4.1 and theorem 12.2.6 that f(t) = f[i(1—A(2))] is a characteristic
function, provided that the distribution which belongs to f(t) satisfies the
conditions of theorem 12.2.6.

Theorem 12.2.7. Let f(Viz) be an analytic characteristic function and

suppose that the corresponding distribution function F(x) is bounded to the
left and that lext [F] = 0; then f (iz) is also a characteristic function.
Let F(x) be the distribution function of f(Viz). Then

pty) = | ew (2) are)

is the frequency function of a mixture of normal frequency functions. Let

a, = ibey' P*(y) dy
be the moments of this mixture. Then
”
Cop = 0

pen =2{° {°gelocyy*—_ expfag

while

oi He Ns (-* )ar(ayay.
eateryass Wtf OAS
We change the variable of integration by putting y = uV2x and see,
after an elementary computation, that
2k)!
(12.2.4) a = ee ry
MIXTURES AND TRANSFORMATIONS 325
where «, is the moment of order k of F(x). The characteristic function
f (Viz) has the expansion

f (Viz)= = ann 2°
so that

a) =} = 2k
k=0
Therefore
2 1)Fa
fis) = 5 a k tk

and we see from (12.2.4) that

Staee ae (=)» ae 2k
Psy & pea.
Since this is the characteristic function which belongs to the frequency
function p*(y), the statement is proved.
As an example we mention the characteristic function

AD a; ae.

D. Dugué (1966) has shown that this is the characteristic function of a

distribution which is closely related to the distribution function of Kolmo-
gorov’s statistic used extensively in the theory of non-parametric statistical
tests.
We conclude this chapter with the discussion of transformations which
are not the result of mixtures of distributions.
Let g(t) be an arbitrary characteristic function and denote its distribution
function by G(x). Then

h(u) = ie dy = [||nee dG (x) dy.

It is easily seen that the order of the two integrations can be exchanged,
so that

Helier 20)
oe) eit _]
= G(x):
We introduce the integral

(228) g)=-| [ead

so that |

6) = —f He)du= [=
326 CHARACTERISTIC FUNCTIONS

It is again possible to exchange the order of the integrations and one obtains
age dG
g(t) = | (e — 1 —itx) ae

The last formula agrees with the Kolmogorov canonical representation

(theorem 5.5.3) and we conclude that the function ¢(¢), as defined by
(12.2.5), is the logarithm of an infinitely divisible characteristic function
which has a finite second moment. We have therefore obtained the follow-
ing result:

Theorem 12.2.8. Let g(y) be an arbitrary characteristic function; then

(12.2.6) f() = exp {-| |“a()dy dul

is the characteristic function of an infinitely divisible distribution with finite
second moment.
As an example we consider the characteristic function g(t) = e~|"! of the
Cauchy distribution. The corresponding function (12.2.6) is then
(12.2.7) ff, (t) = exp (—|¢|+1-e-");
this is an infinitely divisible characteristic function with finite second
moment. The function f(t) = exp (e~'!—1) is also (lemma 5.4.1) an
infinitely divisible characteristic function, and we obtain from (12.2.7) the
relation
el! = f(t) fa(t).
This indicates that it is possible to decompose the Cauchy distribution in
such a way that both factors are infinitely divisible but do not belong to
stable distributions.
 APPENDIX A

The notations O and o

A.1 The notation O. Let f(x) and g(x) be two functions and assume that
g(x) is positive for sufficiently large x. We say that f(x) is at most of the
order of g(x) as x tends to infinity and write
f(s)
=OLg(2)] asx >
if there exists a value x, and a constant A > 0 such that
oy (x) |PPAa(x) tors ,:
Thus f(x) = O[ g(x)] means that the quotient |f(«) |/g(x) is bounded for
sufficiently large x.
Examples. »/x = O(x), «+1 = O(x), exp (itx) = O(1), exp (+/log x) =
O(x), 1/x? = O(«-*””), x sin x = O(x). In all these examples we have
taken for granted that the statement holds as x > oo.
We write f(x) = O(1) to express that f(x) is bounded as » increases.
We list a few rules for the use of these notations.
(1) f(x) = O[gi(*)], fo() = OLg2(*)] imply that f(x) +f2(x)
= Ofgi(*) +g2(%)].
(II) If @>0 is a constant then f(x) = O[ag(x)] implies
F(x) = O[g(*)].
(II) If f:(~) = Ofgi()] and f2 (x) = O[g2(x)] then f(x) fo(*) =
OL81 (*) 82 (*)].-
A.2 The symbol “‘o”. Let f(x) and g(x) be both defined and positive for
sufficiently large x. We say that f(x) is of smaller order than g(x) as
x —> co and write
f(x)
=ofg(a)] asx>
if

Examples. log x = o(x), x = o(x*/), ete:

We list a few properties of this symbol.
(I) f(x) = o[g(*)] implies f(x) = O[g(x)].
(II) If fi(*) = Olgi (*)] and f.(x) = o[g2(*)] then f,(*) f2(*) =
o[81(*) £2(*)].
We write f(x) = o(1) to indicate that f(x) tends to zero.
328 APPENDICES
The same symbols O and 0 are used if x does not tend to infinity but to
some finite value; it is also possible to use this notation if the variable
assumes only integer values as it tends to infinity. This cannot lead to any
misunderstanding since the context will always indicate the variable and
the limit which it approaches.

APPENDIX B

Schwarz’s inequality
We prove this inequality in the form in which we need it: namely as an
inequality which refers to Lebesgue-Stieltjes integrals with respect to a
distribution function.
Let F'(«) be a distribution and consider two real-valued functions g(x)
and h(«) and suppose that g?(x«) and h?(x) are both integrable with respect
to F(x) over (— 0, + 0). Then

[-_ luste)
+0h@]2ar(a)
is a non-negative quadratic form in the variables u and v, so that the
discriminant of this form is non-negative. This yields

[f° semayarey]’ < |" teeter) (" taeoltar@)

foe) 2 © (ce)

which is the desired inequality.

APPENDIX. C

Weierstrass’ approximation theorem

We need here only the trigonometric approximation theorem of
Weierstrass and introduce the following notation.
We denote by C;, the class of all continuous functions f(x), defined for
all real x, which are periodic with period L. We define trigonometric
polynomials T;,(x) of period 2% and degree n
n

T(x) = dX (% cos vx+B, sin vx).

v=0

It is sometimes convenient to write these in complex form as

rn .
Th) = aes
v=-Nn
APPENDICES 329
where the a, can easily be expressed in terms of the «, and By. if 7, (x)is
a trigonometric polynomial with period 2x then T;,(2x/L) is a trigono-
metric polynomial with period L.

Weierstrass’ (trigonometric) approximation theorem. Let f(x) € Cy,; then

for every ¢ > 0) there exists a trigonometric polynomial T,,(x) (of period 2)
such that |f(x)—T,(x*)| < « for all real x.
If f(x) eC, then f[Ly/(2x)] € C,, so that one obtains essentially the
same approximation theorem for the functions of C,; the approximating
trigonometric polynomials then have necessarily the period L.
A convenient proof of the theorem may be found in I. P. Natanson
(1955) [see § 2] or in N. I. Achieser (1956) [see § 22].

APPENDIX D
Order and type of entire functions

Let f(z)= Ssc, 2® be an entire function.

We denote Be
(D1) Mr f)= ee |f(2)|
the maximum ete tdoff(z) in the circle |z| < r. This value is assumed
on the perimeter of the circle.
The order p of an entire function f(z) is defined as

(D.2) p = lim sup log log M(r; f)

log r ;
One has 0 < p < o. Inthis monograph we are in general not interested in
functions of order inferior to 1; if p < o then we say that f(z) is an entire
function of finite order.
An entire function f(z) of finite order p is said to be of type rt if
log M(r;
(D3) lim sup EES! =

An entire function f(z) of finite order p is said to be of minimal type if

t = 0, of normal (or intermediate) type if 0 < t < ©, and of maximal
type if tr = ©.
Entire functions of order 1 and finite type (rt < ©) or of order inferior
to 1 are called entire functions of exponential type.
330 APPENDICES

Order and type of an entire function can be expressed in terms of its

coefficients; one has
It
(D.4) p = lim sup ner eS
ko

and,
11 <7) p<./00;

(D.5) To eylim sup R| c, |°”*.

CP kw

For the proof of these statements we refer the reader to E. Hille (1962)
[see pp. 182-188] or A. I. Markushevich (1965) [see vol. II, Chapter 9].

APPENDIX E

Proof of lemmas needed in Chapter 9

E.1_ Proof of lemma 9.1.1.
We note that

|(en 20) de = &—z-1

0

and introduce the function

It is easily seen that |k(z)| < 1 for|z| < 1. We write z = x+7y (x, y real)
and consider the case |z| > 1. Since
2 y
R(z) = | ”
(2—a)e du ia-*e* |ft) (z—x—iv)edv,
.
we see that

[R(2)| < Jel?{\(@—Dz|+e

|92} < 2641.
Therefore
|k(z)| < 1+2 exp [Re (z)] for all z.
Since
SU sus
(Ea 1) e mf —1 —_ rae| (zu)? 2R(z
R( US en

< 2ut(L+er®)(|2|2+]2))
a

we conclude that the assumption | u*dN(u) < oimplies the uniform

+0
APPENDICES 351
convergence of the integral defining f(s) on every bounded z-set. The
function f(z) is therefore an entire function. We see from (E.1.1) that

4] 2 [2(1-+e"Re®) | u2dN(u) if |2| > 1 and Re(z) > 0

I f(z) < a os
8|z|* | u2dN (u) if |2| > 1 and Re(z) <0
+0
The estimate of the lemma follows immediately.

E.2 Proof of lemmas 9.1.2A and 9.1.2B

= 2
Let f(z) = & d, exp es :)where the coefficients satisfy the estimate
p=0

(E.2.1) d, = O[exp (—kp?)] (p > ©).

Condition (E.2.1) ensures that the series for f(z) converges absolutely and
uniformly in every bounded set of the z-plane. Therefore f(z) is an entire
function which obviously has the period 77. We write x = Re (z) and see

1915 hen °R) of Son(wh)

that

= of (irs),2?[-H(P-Fa)J}
= oles (irs) el
2
The last estimate follows from the fact that ) exp |—a 2) |is
p=— ©
a continuous periodic function of x with real period kT/z and is therefore
bounded. If x = Re(z) < 0 we see that |f(z)| < & |d,| = O(1) and
p=0
lemma 9.1.2A is proved.
We proceed to the proof of lemma 9.1.2B.
We assume that the entire function f(z) is periodic with period 7T and
that the estimates (9.1.7) hold.
We expand f(zy) into a Fourier series,

f(y) = 7D d, exp (=P

2) (—0 <y < o)
where

(E:2.2) n\ 10) exp (782)

d, = ui (pas Oe Ly, eeres

Let ¢ be a complex variable and consider the entire function A(¢)=

f (it) exp (—2inpl/T). [For real ¢ this is the integrand in (E.2. 2).] We
332 APPENDICES
integrate A(¢) around the rectangle whose vertices are the four points
0, 20, 10+ T and T (6 real). According to Cauchy’s theorem this integral is
zero. Since the function A(¢) is periodic with period T, the integrals along
the vertical sides of the rectangle cancel, and we see that
1 foe 2011: ;
djs AI f(ty—9) exp | =? (y+) dy.
T Jo T
It follows that

(E.2.3) |d,| < exp Ex) Wert f(iy—9).

Suppose that 6 > 0; we see then from (9.1.7) that
Z
|d,| < exp Ee 6+ O(log a).
We let 6 tend to + 0 and see that d, = 0 forp < 0. If 6 < 0 we obtain
from (E.2.3) and (9.1.7) the estimate

i= | exp(= 0+.NO*)]|
We put 6 = ae and get the desired estimate

dy = O( — kp’)
here k
where = a
NT?

E.3 Proof of lemma 9.2.3

Let z = t+1y (t,y real); we show first that the function 6(z) =
(z+1)~°e~™ f(z) is bounded in the half-plane Re (z) > 0. We select a
point 29 = f)+72y> in the first quadrant; i.e. 29 E[(t, y): ¢ > 0, y > OJ.
Let ©, be the angle formed by the rays ¢ = [(t, y): y = 0, t > 0] and
dt
(= tcyiv= ae t2> 0}.We select ¢ > 0 so that 2) € 9,.

We put 0,(2) = 6(z) exp (dez?), then for z € 9, we have

|8.(2)| = |6(2) |exp (—2ety) < My exp (dt?—2ety) < M, exp (d] 2)
Moreover, |6,(z)| < |6(z)| < M, on the ray ¢ and |6,(z)| < M, on
the ray ¢,. We can therefore apply lemma 8.2.1 with 9 =9,, M=
max (M,, M3), B = : > oa = arctan a and p»=2)< 7/p.
2é
Then
|0,(2)['< "4 for z€9,
and
|(%) | < max (M,, M3) exp (2eto yo).
APPENDICES 333
We let « tend to zero and see that
|O(z>)| < max (M,, M,).
The point 29 is an arbitrary point of the first quadrant, so that 6(z) is
bounded in the closed first quadrant. In a similar way) one shows that
6(z) is also bounded in the closed fourth quadrant [(¢, y): t > 0, y < 0]
and therefore in the half-plane Re (z) > 0. We consider the function
0,(2) = (s+ 1)-*e-
f(z)= (w+ 1)*0(2).
On the boundary of the half-plane Re (z) > 0 we have
[61(2)| < My
while in the half-plane Re (z) > 0
6,(2) = O[|s+1|°4.
The assumptions of lemma 8.2.1 are again satisfied if we identify 9 with
the half-plane Re (z) > 0 and put f(z) = 0,(z), M= M,, « = —2/2,
B = x/2 and p = $. We see then that for Re (z) 2 0
[9:(z)| < Mi,
or
|f(z)| < Mi|2+1|? exp [6 Re (z)],
as stated in lemma 9.2.3.
E.4 Proof of lemma 9.2.4
Let f, (2) = f(2+7T) —f(2); the function f, (2) is entire and
(E41) |fia) < exp fh Re (z)+ O(log |#|)} [Re (2) > 0].
It follows from the representation of f(z) that

fli); = 3an(GTB, exp (“2

j=-—@
21}

According to the assumption of the lemma, this series converges uniformly

on the interval 0 < y < T and is therefore the Fourier series of its sum
f (ty). We repeat the reasoning which led from formula (E.2.2) of Appendix
E.2 to (E.2.3) and see that
208 :
(E.4.2) |iTb;| < exp (=) max |f,(iy—6)|.
0<y<T

For 6 < 0 we can apply the estimate (E.4.1) and get

|b;| < exp | 10+ O(log op).

We let 6 tend to — « and conclude that

(E473) b,=0 forj > w = [RT/(2n)].
) To cher this we consider instead of 0, (z) the function 0(z) exp (—1ez?).
334 APPENDICES

We next put

fale) = Fla)— 3e (as+b,2) exp (7)

2nyz
= S2 (a+8,2) exp (“E).
27%

In view of (E.4.3) we have

flo) = 3. aexp (A)

! = 2a}
(-w <y < @),
Moreover
|fo(2)| < exp {k Re (z)+O(log|2|)} [Re(z) > 0].
Using the same argument as before, we see that

jay] < exp (“Z)

2778
max |faliy—6)|.
0<y<T
:

The estimate for f,(z) yields, for 6 < 0, the inequality

2770
|a;| < exp [722° 0 + OUlog \0))|

We let 0 - — o and see that a; = 0 forj > w = [kT /(2x)]. This com-
pletes the proof of lemma 9.2.4.

APPENDIX Fi

Schwarz’s reflection principle

Let D, and D, be two domains such that D, 7 D, = © while D, aD,
is an interval y on the real axis. Let f, (2) be regular in D, and continuous
in D, U yand let f, (2) be regular in D, and continuous in D, U y. Suppose
that for ey
eG) meal NEI SE)
where the approach is from D, in the first and from D, in the second limit.
Then there exists a function f(z) regular in D, U D, U y which coincides
with f,(z) in D, and with f,(z) in D,.
For the proof we refer to Hille (1959), p. 184. In Chapter 11 we used for
D, the interior of a rectangle located in the upper half-plane and for D, the
interior of the rectangle which is located symmetrically to D, with respect
to the real axis.
 LISF OF EXAMPLES OF
CHARACTERISTIC FUNCTIONS®

Chapter |
and Page Description of example
Section

#2 6 Table of discrete d.
iL? 7 Table of absolutely continuous d.
ee 8 Cantor d.
iD 9 Purely singular, strictly increasing d.
1.4 12 d.f. having moments of order inferior to m but not of order m
or of higher order.
1.4 13 Table of moments.
1.4 14 Recurrence relations for moments of singular d. (ref. to)—
1.4 1D, NS two different d.f. having the same sequence of moments.
Drill 18 Table of c.f.
Del 19 c.f. of Cantor d.
Dp 19, 20 Behaviour of L = a sup |f(t) |for singular d.
tl—oo
Des 22 c.f. for which f’(0) exists but not the first moment.
2.3 23 c.f. which is nowhere differentiable.
Des 24 A c.f. which has an expansion f(t) = 1+0(t), although the
first moment does not exist.
3.6 53 The weak convergence of a sequence {Fy} of d.f. to a d.f.
F(x) does not imply the convergence of the sequence of
moments of the Fp to the moments of F.
3.6 54 The weak convergence of a sequence {Fn} of d.f. to a
limiting d. F(x) does not imply the convergence of the cor-
responding densities.
3.7 67 Symmetric Bernoulli convolutions exhibiting different be-
haviour of L = lim sup |f(¢)|. c.f. of singular d. with L > 0
Or ll = ©, [elo-co
4.3 84-5 An absolutely continuous d. may have a c.f. which is not
absolutely integrable.
4.3 85 The c.f. of two different d. can agree over a finite interval.
Eyal 104 Multiple factorizations of c.f.
Sail 104 Cancellation law invalid in arithmetic of d.f.
eS 109 A c.f. which has no real zeros need not be i.d.
5a3 110 The absolute value of a c.f. is not necessarily a c.f.
55 121 Table of canonical representations of c.f.
DED 122 Two different c.f. can have the same absolute value.
525 122-3 Two different c.f. can have the same square.

(*) Abbreviations used: c.f. = characteristic function(s), d. = distribution(s),

d.f. = distribution function(s), i.d. = infinitely divisible.
336 LIST OF EXAMPLES

Chapter
and Page Description of example
Section

Ded 122 |Ac.f. f(#) such that f(2) is not i.d. but |f(2) | is id.
D.9 122 | Ani.d. c.f. may be the product of two factors which are not
id.
BS) 123-4 |An id. c.f. which has an indecomposable factor (factor
explicitly given).
5.8 143 | A stable density with exponent « = 4.
6.2 176 | Ani.d. c.f. can be the product of a denumerable number of
indecomposable c.f.
6.2 {79 |i.d. c.f. which have indecomposable factors.
6.3 183-4 |c.f. of an absolutely continuous and unbounded indecom-
posable d.
6.3 184 | Product of two c.f., neither of which has a normal component,
can have a normal factor.
6.3 185 | Product of two c.f., neither of which has a Poissonian com-
ponent, can have a Poissonian factor.
6.3 188 | Construction of a c.f. which belongs to a finite and inde-
composable absolutely continuous d.
6.3 188 | Construction of a c.f. which belongs to a finite, purely singu-
lar, indecomposable d.
6.3 189 | Factorization of rectangular d. (it can be represented in two
ways as an infinite product of indecomposable factors).
6.3 189 | Factorization of rectangular d. into a product of two purely
singular d.
Teal 194 | Quotient of two c.f. need not be a c.f.
Wow 198 | A d. which has moments of all orders but has not an analytic
c.f. Nevertheless the sequence of moments determines this d.
completely.
Tae) 203 | A one-sided distribution may have an entire c.f. of order
greater than 1.
Te? 203 | The infinitely many zeros of the c.f. of a finite d. need not be
real.
7.3 212 | References to examples of rational c.f.
7.4 225 | A periodic c.f. which is not analytic.
74 227 | A doubly periodic c.f.
8.2 252 | Convolution of three Poisson type d. can have an indecom-
posable factor.
8.4 258 | 1.d. c.f. with zeros on the boundary of its strip of regularity.
8.4 259 | Entire c.f. without zeros which is not i.d.
92 280 | Reference to an example which shows that a c.f. can belong
to ¥ but not to Ip.
12.1 317 | Mixture using a weight function which is not monotone.
22, 320-3 |Examples of transformations of c.f.
WO 323 |Examples of transformations applied to exp (—t?/2) which
yield entire c.f. of order 2 with a finite number of zeros.
 REFERENCES

ACHIESER, N. I. (1956). Theory of Approximation. Transl. from Russian, New York,

:rederick Ungar. [Russian originally published by Ogiz, Moscow-—Leningrad
1947).]
AKUTOWICZ, E. J. (1959). On extrapolating a positive definite function from a
finite interval. Mathematica Scandinavica, 7, 157-169.
AKUTOWICZ, E. J. (1960). Sur approximation par certaines fonctions entiéres. Ann.
Scientifiques Ecole Normale Supérieure (3e série), 77, 281-301.
BERGSTROM, H. (1952). On some expansions of stable distributions. Arkiv for
Matematik, 2, 375-378.
BLUM, J.-ROSENBLATT, M. (1959). On the structure of infinitely divisible distribu-
tions, Pacific J. Math., 9, 1-7.
BOAS, R. P. (1967). Lipschitz behaviour and integrability of characteristic functions.
Ann. Math. Statist., 38, 32-36.
BOCHNER, S. (1932). Vorlesungen iiber Fourier’ sche Integrale. Leipzig, Akademische
Verlagsgesellschaft. Reprinted by Chelsea Publishing Co., New York, N.Y.
(1948). English transl.: Lectures on Fourier Integrals, Annals of Math. Studies,
No. 42, Princeton University Press, Princeton, N.J. (1959).
BOHR, H. (1932). Fastperiodische Funktionen. Ergebnisse d. Mathematik I/4,
Berlin, J. Springer. English transl. (1947): Almost periodic functions, Chelsea
Publishing Co., New York, N.Y.
CAIROLI, R. (1964). Sur les fonctions caractéristiques des lois de probabilité. Publ.
Inst. Statist. Univ. Paris, 13, 45-53.
CHRISTENSEN, I. F. (1962). Some further extensions of a theorem of Marcinkiewicz.
Pacific J. Math., 12, 59-67.
CHUNG, K. L. (1953). Sur les lois de probabilité unimodales. C.R. Acad. Sci. Paris,
236, 583-584.
COPSON, E. T. (1935). An Introduction to the Theory of Functions of a Complex
Variable. Oxford, Clarendon Press.
CRAMER, H. (1939). On the representation of a function by certain Fourier integrals.
Trans. Amer. Math. Soc., 46, 191-201.
CRAMER, H. (1964). Mathematical Methods of Statistics. Princeton University Press,
Princeton, N.J.
CRUM, M. M. (1956). On positive definite functions. Proc. London Math. Soc., (3),
6, 548-560.
CUPPENS, R. (1963a). Sur la décomposition d’une fonction 2q fois dérivable 4
Vorigine en produit infini des fonctions caractéristiques. C.R. Acad. Sct.
Paris, 256, 3806-3808.
CUPPENS, R. (1963b). Sur un théoréme de Mamay. C.R. Acad. Sci. Paris, 257,
586-588.
CUPPENS, R. (1969). On the decomposition of infinitely divisible probability laws
without normal factor. Pacific ¥. Math., 28, 61-76.
puGug, D. (1955). Sur l’approximation d’une fonction caractéristique par sa série
de Fourier. C.R. Acad. Sci. Paris, 240, 151-152.
338 REFERENCES
DUGUE, D. (1957a). Résultats sur les fonctions absolument monotones et applica-
tions a l’arithmétique des fonctions de type positif. C.R. Acad. Sci. Parts,
244, 715-717.
DUGUE, D. (1957b). Arithmétique des lois de probabilités. Mémorial des Sciences
Math., 137. Paris, Gauthier-Villars.
DUGUE, D. (1957c). Sur le théoréme de Lévy—Cramér. Publ. Inst. Statist. Univ.
Paris, 6, 213-225.
DUGUE, D. (1966). Sur les lois de Kolmogoroff et de von Mises. C.R. Acad. Sct.
Paris, 262, 999-1000.
DUGUE, D.—GIRAULT, M. (1955). Fonctions convexes de Polya. Publ. Inst. Statist.
Univ. Paris, 4, 3-10.
ESSEEN, C. G. (1944). Fourier analysis of distribution functions. Acta Mathematica,
77, 1-125.
EVANS, G. C. (1957). Calculation of moments for a Cantor—Vitali function. Herbert
Ellsworth Slought Memorial paper No. 6, Supplement to Amer. Math. Monthly,
64, No. 8, 22-27.
FELLER, W. (1943). On a general class of ‘‘contagious’’ distributions. Ann. Math.
Statist., 14, 389-400.
FELLER, W. (1952). On a generalization of Marcel Riesz’ potentials and the semi-
groups generated by them. Medd. Lunds Univ. Mat. Sem., Tome Supplémen-
taire Marcel Riesz, 73-81.
FISZ, M.-VARADARAJAN, V. S. (1963). A condition for absolute continuity of in-
finitely divisible distribution functions. Z. f. Wahrscheinlichkettstheorie, 1,
335-339.
GIL-PELAEZ, J. (1951). Note on the inversion theorem. Biometrika, 38, 481-482.
GIRAULT, M. (1954). Les fonctions caractéristiques et leurs transformations. Publ.
Inst. Statist. Univ. Paris, 4, 223-299.
GIRAULT, M. (1955). Analyticité et périodicité des fonctions caractéristiques. Publ.
Inst. Statist. Univ. Paris, 5, 91-94.
GNEDENKO, B. V.-KOLMOGOROV, A. N. (1954). Limit distributions for sums of inde-
pendent random variables. (Transl. from Russian by K. L. Chung), Cambridge,
Mass., Addison—Wesley Publishing Co.
GOLDBERG, A. A.—OSTROVSKH, I. V. (1967). An application of a theorem of W. K.
Hayman to a problem in the theory of the decomposition of probability laws.
Ukrain. Mat. Zurnal, 19, 104-106. [A. A. Toapnepr.-U. B. Ocrposcxuit.
IIpumenenue teopemn Y. HK. Xeiimana kK ofHOMy BOTpocy Teopuit pasomenuit
BePOATHOCTHEIX 3aKOHOB. YKpauncKuii Mar. /Rypuaa, 19 (1967), 104-106.]
HAHN, H.—-ROSENTHAL, A. (1948). Set Functions. Albuquerque, N. M., Univ. of
New Mexico Press.
HALMOS, P. R. (1950). Measure Theory. D. Van Nostrand Co., New York.
HARDY, G. H. (1963). A Course in Pure Mathematics (10th edn), Cambridge, Univ.
Press:
HARDY, G. H.-LITTLEWOOD, J. E.—POLYA, G. (1934). Inequalities. Cambridge, Univ.
Press.
HARTMAN, P.—WINTNER, A. (1942). On the infinitesimal generators of integral
convolutions. Amer. J. Math., 64, 273-298.
HAUSDORFF, F. (1927). Mengenlehre (2nd rev. edn). Berlin—Leipzig, Walter de
Gruyter & Co.
HERZ, C. S. (1963). Fonctions opérant sur les fonctions définies positives. Annales
de Il’Institut Fourier, 13, 161-180.
HILLE, E. (1959, 1962). Analytic Function Theory, vol. I (1959), vol. II (1962).
Ginn & Co., Boston, Mass.
REFERENCES 339
HOBSON, E. W. (1927). The theory of functions of a real variable, 1, 11. Cambridge,
Univ. Press; reprinted by Dover Publications, New York (1957).
IBRAGIMOV, I. A. (1956a). On the composition of unimodal distributions. Teoriya
veroyatnostet 1 ee primeneniya, 1, 283-288. English transl.: Theory of proba-
bility and its applications, 1, 255-260.
IBRAGIMOY, I. A. (1956b). A theorem in the theory of infinitely divisible laws.
Teoriya veroyatnostei i ee primeneniya, 1, 485-489. English transl.: Theory of
probability and its applications, 1, 440-444.
IBRAGIMOV, I. A. (1957). Remark on a probability distribution of class L. Teoriya
veroyatnostet 1 ee primeneniya, 2, 121-124. English transl.: Theory of probability
and its applications, 2, 117-119.
IBRAGIMOV, I. A.-LINNIK, YU. V. (1965). Independent and stationary dependent
variables. Moscow, Izdat. Nauka. [V. A. W6parumos-lO. B. JInuuur.
Hesabicumble WM cTalgonapHo cBsasaHHbe Berm. Mocxsa (1965), Isqar. Hayxa.]
INGHAM, A. I. (1936). A note on Fourier transforms. J. London Math. Soc., 9,
27-32.
JESSEN, B.—-WINTNER, A. (1935). Distribution functions and the Riemann Zeta
function. Trans. Amer. Math. Soc., 38, 48-88.
JOHANSEN, S. (1966). An application of extreme point methods to the representation
of infinitely divisible distributions. Z. f. Wahrscheinlichkettstheorie, 5, 304-316.
JORDAN, C. (1950). Calculus of Finite Differences. New York, Chelsea Publishing Co.
KAWATA, T. (1940). On the division of probability laws. Proc. Imp. Acad. Tokyo,
16, 249-254.
KERSHNER, R. (1936). On singular Fourier transforms. Amer. J. Math., 58, 450-453.
KERSHNER, R.—WINTNER, A. (1935). On symmetric Bernoulli convolutions. Amer. J.
Math., 57, 541-545.
KONHEIM, A. G.—WEISS, B. (1965). Functions which operate on characteristic func-
tions. Pacific J. Math., 15, 1279-1293.
KONHEIM, A. G.—WEISS, B. (1968). A note on functions which operate. Pacific J.
Math., 24, 297-302.
KRASNER, M.—RANULAC, B. (1937). Sur une propriété des polynémes de la division
du cercle. C.R. Acad. Sci. Paris, 204, 397-399.
KREIN, M. G. (1940). Sur le probléme du prolongement des fonctions hermitiennes
positives et continues. C.R. (Doklady) Acad. Sci. USSR, 26, 17-20.
KREIN, M. G. (1943). On the representation of functions by Fourier—Stieltjes
integrals. Ucenie Zapiski Kuibishevskogo Gosud. Pedag. 1 Ucitelskogo Inst.,
7, 123-148 [M. I. Kpeiiu. O Upencrannenmu gynxuuit unterpanama dypse
Cruutbeca. Yuentie Sanucrn Kyit6smmescroro Tocyz. Ilegarormueckoro n YauresbeKoro
VWuernryta, 7, 123-148.]
KUBIK, L. (1961/62). A characterization of the class L of probability distributions.
Studia Math., 21, 254-252.
KUBIK, L. (1962/63). Some analogies between the class of infinitely divisible dis-
tributions and the class L of distributions. Studia Math., 22, 197-209.
KURATOWSKI, C. (1952). Topologie, vol. I (3rd edn). Monografie Matematyczne
XX: Polska Akademia Nauk, Warszawa.
LAHA, R. G. (1960). On a property of positive definite functions. Bull. Amer. Math.
Soc., 66, 388-391.
LAHA, R. G.-LUKACS, E. (1962). On a factorization of characteristic functions which
have a finite number of derivatives at the origin. Publ. Inst. Statist. Univ.
Paris, 11, 221-224.
LETTA, G. (1963). Eine Bemerkung zur Kennzeichnung der charakteristischen
Funktionen. Z. f. Wahrscheinlichkeitstheorie, 2, 69-74.
340 REFERENCES
LEVIN, B. YA. (1964). Distribution of zeros of entire functions. Amer. Math. Soc.,
Providence, R.I. [Russian original published Moscow, 1956.]
LEVINSON, N. (1936). On a class of non-vanishing functions. Proc. London Math.
Soc., (2), 41, 393-407.
LEVINSON, N. (1938). A theorem relating to non-vanishing and analytic functions.
J. Math. Phys., 16, 185-190.
LEvy, P. (1931). Sur les séries dont les termes sont des variables éventuelles in-
dépendantes. Studia Math., 3, 119-155.
LEvy, P. (1937a). Théorie de l’addition des variables aléatoires. Paris, Gauthier-
Villars. [2nd edn, Paris, 1954.] ;
LEvy, P. (1937b). Sur les exponentielles de polynémes. Ann. Scientifiques Ecole
Normale Supérieure (3e série), 73, 231-292.
LEvy, P. (1937c). L’arithmétique des lois de probabilité et les produits finis des
lois de Poisson. C.R. Acad. Sci. Paris, 204, 944-946.
LEvy, P. (1938a) L’arithmétique des lois de probabilité. J. Math. Pures et Appl.,
103, 17-40.
LEvy, P. (1938b). L’arithmétique des lois de probabilité et les produits finis des
lois de Poisson. Actualités Scientifiques et Industrielles, No. 736 (Colloque de
Genéve ITI), 25-59. Hermann, Paris.
LEvy, P. (1939). Sur certains processus stochastiques homogénes. Compositio
Mathematica, 7, 283-339.
LEvy, P. (1952). Sur une classe des lois de probabilité indécomposables. C.R. Acad.
Sct. Paris, 235, 489-491.
LEvy, P. (1961). Quelques problémes non résolus de la théorie des fonctions
caractéristiques. Annali Mat. Pura e Appl. (ser. 4), 53, 315-332.
Livy, P. (1962). Extensions d’un théoréme de D. Dugué et M. Girault. Z. f.
Wahrscheinlichkettstheorie, 1, 159-173.
LEWIS, T. (1967). The factorization of the rectangular distribution. J. Applied Prob.,
4, 529-542.
LINNIK, YU. Vv. (1953). Linear forms and statistical criteria: I, Il. Ukrain Mat. Z.,
5, 207-243, 247-290. English transl. in Selected Translations Math. Statist.
and Prob., 3, 1-40, 41-90. Amer. Math. Soc., Providence, R.I. (1962).
LINNIK, YU. Vv. (1957). On the decomposition of the convolution of Gaussian and
Poissonian laws. Teoriya veroyatnostei i ee primeneniya, 2, 34-59. English
transl.: Theory of probability and its applications, 2, 31-57.
LINNIK, YU. Vv. (1959). On ‘‘«-factorizations”’ of infinitely divisible probabilistic
laws. Vestnik Leningrad Univ. (1959), 14-23. English transl. in Selected trans-
lations Math. Statist. and Prob., 2, 159-169. Amer. Math. Soc., Providence ,R.1.
LINNIK, YU. V. (1964). Decomposition of Probability Distributions. Oliver and Boyd,
Edinburgh—London. [Russian original published Leningrad, 1960.]
LOEFFEL, H. (1956). Beitrége zur Theorie der charakteristischen Funktionen. Mitt.
Verein. Schweiz. Versich. Math., 56, 337-381.
LOEVE, M. (1955). Probability Theory. D. Van Nostrand, New York (3rd edn, 1963).
LUKACS, E. (1958). Some extensions of a theorem of Marcinkiewicz. Pacific J.
Math., 8, 487-501.
LUKACS, E. (1964). A linear mapping of the space of distribution functions onto a
set of bounded continuous functions. Z. f. Wahrscheinlichkeitstheorie, 3, 1-6.
LUKACS, E. (1967). On the arithmetic properties of certain entire characteristic
functions. Proc. 5th Berkeley Symp. Math. Statist. and Prob., 2, Part I, 401-414.
Univ. of Calif. Press, Berkeley and Los Angeles.
LUKACS, E.—SZASZ, 0. (1954a). Certain Fourier transforms of distributions: II.
Canad. J. Math., 5, 186-189.
REFERENCES 341
LUKACS, E.—SZASZ, 0. (1954b). Non-negative trigonometric polynomials and certain
rational characteristic functions. J. Research Nat. Bureau Standards, 52,
153-160.
MARCINKIEWICZ, J. (1938). Sur une propriété de la loi de Gauss. Math. Zeitschr.,
44, 612-618. Reprinted in J. Marcinkiewicz, Collected Papers. Panstwowe
wydawnictwo Naukowe Warszawa, 1964.
MARKUSHEVICH, A. I. (1965-67). Theory of Functions of a Complex Variable, vol. I, 11
(1965), III (1967). Transl. from Russian by R. A. Silvermann. Englewood
Cliffs, N.Y., Prentice Hall.
MEDGYESSY, P. (1956). Partial differential equations for stable density functions
and their applications. Magyar Tud. Akad. Mat. Kutaté Intéz. Kézl., 1,
489-518.
MEDGYESSY, P. (1963). On the interconnection between the representation theorems
of characteristic functions of unimodal distributions and of convex charac-
teristic functions. Magyar Tud. Akad. Mat. Kutaté Intéz. Kézl., 8, 425-430.
MILLER, H. D. (1967). Generalization of a theorem of Marcinkiewicz. Pacific J.
Math., 20, 261-274.
NATANSON, I. P. (1955). Konstruktive Funktionentheorie. Berlin, Akademie Verlag.
[Russian original published by Ogiz, Moscow—Leningrad, 1951.]
NOACK, A. (1950). A class of variables with discrete distributions. Ann. Math.
Statist., 21, 127-132.
OSTROVSKII, I. V. (1963). Entire functions satisfying some special inequalities con
nected with the theory of characteristic functions of probability laws. Khar-
kovskogo Univ. Mat. Obshestva, 29, 145-168. English transl. in Selected
Translations Math. Statist. and Prob., 7, 203-234. Amer. Math. Soc., Provi-
dence, R.I.
OSTROVSKII, I. v. (1964). On the decomposition of infinitely divisible lattice laws.
Vestnik Leningrad Univ., 19, 51-60. [U. B. Ocrposcxnit. O pasnomennax
pelleTuaTEx OesrpanHuHo euuMbIx sakoHos. Bectuuk JlesuurpaycKkoro YuuBepcutera,
19, 51-60 (1964).]
OSTROVSKII, I. v. (1965). Some theorems on the decomposition of probability laws.
Trudi Mat. Inst. Steklova, 79 (1965), 198-235. English transl.: Proceedings of
the Steklov Institute of Mathematics, No. 79 (1965), 221-259. Amer. Math.
Soc., Providence, R.I. (1966).
OSTROVSKII, I. v. (1966). On the factorization of multidimensional infinitely
divisible probability laws without Gaussian component. Vestnik Khar-
kovskogo Gosud. Univ. (Ser. Meh.-Mat.), 32, 51-72. [M. B. Ocrposcxuit. O
pasOwKeHMAX MHOTMepHEIX BesrpaHMGHO J{eeMEX 3aKOHOB 6e3 rayccoBoll KOMMONeHTEI.
Becrumk Xapxoscxoro Tocyy. Yaus. Cep. Mex.-Mar., 32, 51-72.]
PALEY, R. E. A. C.-WIENER, N. (1934). Fourier transforms in the complex domain.
Amer. Math. Soc. Colloquium Publ. No. 19, Amer. Math. Soc., New York,
N.Y.
PITMAN, E. J. G. (1956). On the derivation of a characteristic function at the origin.
Ann. Math. Statist., 27, 1156-1160.
PITMAN, E. J. G. (1961). Some theorems on characteristic functions of probability
distributions. Proc. 4th Berkeley Symp. Math. Statist. and Prob., 2, 393-402.
Univ. of Calif. Press, Berkeley and Los Angeles.
POLYA, G.-SZEGO, G. (1925). Aufgaben und Lehrsdtze aus der Analysis, I. J. Springer,
Berlin. American Edition: Dover Publications, New York (1945).
RAIKOY, D. A. (1938). On the decomposition of Gauss and Poisson laws. Izvest.
Akad. Nauk SSSR (Ser. Mat.), 2, 91-124. [J]. A. Paiixos. O pasuosenmm 3aKon0B
Taycea m Iyaccona. Mspecrua Axag. Hayn. CCCP, Cep. Mar., 2, 91-124.]
342 REFERENCES
RAIKOV, D. A. (1940). On positive definite functions. Doklady Akad. Nauk SSSR,
26, 857-862. [J]. A. Paiixos. O nonoxuTembuo onpenenennEx PynKyuax. Jforsaqy
Axay. Hayx CCCP, 26, 857-862.]
RAMACHANDRAN, B. (1962). On the order and type of entire characteristic functions.
Ann. Math. Statist., 33, 1238-1255.
RAMACHANDRAN, B. (1965). An extension of a theorem of Mamay with applica-
tion. Sankhyd, A, 27, 303-310.
RAMACHANDRAN, B. (1967). Advanced Theory of Characteristic Functions. Statistical
Publishing Society, Calcutta.
RIESZ, F. (1933). Uber Sitze von Stone und Bochner. Acta Univ. Szeged, 6, 184-198.
RIESZ, M.—LIVINGSTONE, A. E. (1955). A short proof of a classical theorem in the
theory of Fourier integrals. Amer. Math. Monthly, 62, 434-437.
ROBBINS, H. (1948). Mixture of distributions, Ann. Math. Statist., 19, 360-369.
ROSSBERG, H. J. (1966). Wachstum und Nullstellenverteilung ganzer charak-
teristischer Funktionen. Monatsberichte Deutsch. Akad. Wiss. Berlin, 8,
275-286.
ROSSBERG, H. J. (19672). Der Zusammenhang zwischen einer ganzen charak-
teristischen Funktion einer verfeinerten Ordnung und ihrer Verteilungs-
funktion. Czechoslovak Math. J., 17(92), 317-334.
ROSSBERG, H. J. (1967b). Ganze charakteristische Funktionen mit vollkommen
regularem Wachstum. Czechoslovak Math. J., 17(92), 335-346.
SALEM, R. (1943). On some singular monotonic functions which are strictly in-
creasing. Trans. Amer. Math. Soc., 53, 427-439.
SAPOGOV, N. A. (1951). The stability problem for a theorem of Cramér. Izvest.
Akad. Nauk SSSR (Ser. Mat.), 15, 205-218. English transl. in Selected
Translations Math. Statist. and Prob., 1, 41-53. Amer. Math. Soc., Providence,
Ree
SCHMETTERER, L. (1965). Some theorems on the Fourier analysis of positive definite
functions. Proc. Amer. Math. Soc., 16, 1141-1146.
SCHWARTZ, L. (1941). Sur le module de la fonction caractéristique du calcul des
probabilités. C.R. Acad. Sci. Paris, 212, 418-421.
SHIMIZU, R. (1964). On the decomposition of infinitely divisible characteristic
functions with a continuous Poisson spectrum. Ann. Inst. Statist. Math., 16,
387-407.
SHOHAT, J. A.-TAMARKIN, J. D. (1943). The Problem of Moments. (Mathematical
Surveys, 1.) Amer. Math. Soc., New York, N.Y.
SKOROHOD, A. V. (1954). Asymptotic formulas for stable distribution laws. Doklady
Akad. Nauk SSSR (N. S.), 98, 731-734. English transl. in Selected Transla-
tions Math. Statist. and Prob., 1, 157-161. Amer. Math. Soc., Providence,
Re
SUN, T. C. (1967). A note on the unimodality of distribution functions of the class L.
Ann. Math. Statist., 38, 1296-1299.
TAKANO, K. (1951). Certain Fourier transforms of distributions. Tohoku Math. J.,
(2), 3, 306-315.
TEICHER, H. (1954). On the factorization of distributions. Ann. Math. Statist., 25,
769-774.
TITCHMARSH, E. C. (1937). Introduction to the Theory of Fourier Integrals. Oxford,
Clarendon Press.
TITCHMARSH, E. C. (1939). The Theory of Functions (2nd edn). Oxford, Univ.
Press.
TUCKER, H. G. (1962). Absolute continuity of infinitely divisible distributions.
Pacific J. Math., 12, 1125-1129.
REFERENCES 343
TUCKER, H. G. (1964). On continuous singular infinitely divisible distribution
functions. Ann. Math. Statist., 35, 330-335.
TUCKER, H. G. (1965). On a necessary and sufficient condition that an infinitely
divisible distribution be absolutely continuous. Trans. Amer. Math. Soc.,
118, 316-330.
WIDDER, D. v. (1946). The Laplace Transform. Princeton, Univ. Press.
WINTNER, A. (1936). On a class of Fourier transforms. Amer. J. Math., 58, 45-90.
WINTNER, A. (1938). Asymptotic distributions and infinite convolutions (plano-
graphed lecture notes). Edwards Brothers, Ann Arbor, Michigan.
WINTNER, A. (1947). The Fourier Transforms of Probability Distributions. Balti-
more, Md (published by the author).
WINTNER, A. (1956). Cauchy’s stable distributions and an ‘‘explicit formula’”’ of
Mellin. Amer. J. Math., 78, 819-861.
ZEMANIAN, A. H. (1959). On the pole and zero location of rational Laplace trans-
formations of non-negative functions. Proc. Amer. Math. Soc., 10, 868-872.
ZEMANIAN, A. H. (1961). On the pole and zero locations of rational Laplace trans-
formations of non-negative functions, II. Proc. Amer. Math. Soc., 12, 870-874.
ZINGER, A.—LINNIK, YU. V. (1957). On a class of differential equations and its
application to some questions of regression theory. Vestnik Leningrad Univ.,
12, 121-130. English transl. in Selected Translations Math. Statist. and Prob.,
3, 181-190. Amer. Math. Soc., Providence, R.I.
ZOLOTAREY, V. M. (1954). Expression of the density of a stable distribution with
exponent « greater than 1 by means of a density with exponent 1/«. Doklady
Akad. Nauk SSSR, 98, 735-738. English transl. in Selected Translations
Math. Statist. and Prob., 1, 163-167. Amer. Math. Soc., Providence, R.I.
ZOLOTAREV, V. M. (1956). On analytic properties of stable distribution laws. Vestnik
Leningrad Univ., 11, No. 1, 49-52. English transl. in Selected Translations
Math. Statist. and Prob., 1, 207-211. Amer. Math. Soc., Providence, R.I.
ZOLOTAREV, V. M. (1963). Analytical structure of the infinitely divisible laws of the
L-class. Litovskii Mat. Sbornik, 3, 123-124. [B. M. Sonorapes. Anammrugeckoe
cTpoeHue Gesrpanwuno fenembIx sakonos Kaacca L. JIwrosckuii Maremarnyecknit
Coopunk, 3, 123-124.]
ZOLOTAREY, V. M. (1964). On the representation of stable laws by integrals. Trudi
Mat. Inst. Steklova, 71 (1964), 46-50. English transl. in Selected Translations
Math. Statist. and Prob., 6, 84-88. Amer. Math. Soc., Providence, R.I.
ZYGMUND, A. (1947). A remark on characteristic functions. Ann. Math. Statist., 18,
272-276.
ZYGMUND, A. (1951). A remark on characteristic functions. Proc. 2nd Berkeley
Symp. Math. Statist. and Prob., 369-372. Berkeley, Univ. of California Press.
ZYGMUND, A. (1952). Trigonometrical Series (2nd edn, corrected reprint of 1st edn,
Monographie Mat. 5, Warszawa, 1935). Chelsea Publishing Co., New York,
INES
 INDEX

Abel summability, 41 boundary c.f., ridge property of, 308

absolute moment, 10, 25 — —, strip of regularity of, 308, 310
absolutely continuous, 4, 66 — value of analytic f., 306
——d., 4, 5, 38, 75 bounded d. (to left, right), 142
— monotone, 302 — to (left, right), 142
Acuieser, N. I., 329, 337 — variation, 320
adjoint polynomial, 228
AxuTowicz, E. J., 89, 337 (C,1)-summability, 41
algebraic moment, 10 CatRoui, R., 224, 337
almost periodic, 19, 36 cancellation law, 122
a-decomposition, 292, 301 canonical product, 222
—, denumerable, 292 — representation, 113 ff.
—, finite, 292 — —, Kolmogorov, 119
— for i.d.d., 305 — —, Lévy, 118, 176, 178-179
— — lattice d., 302 — —, Lévy-Khinchine, 117
—— normal d., 302 — — of stable c.f., 132
— — Poisson d., 303 — —, table of, 121
— of a.c.f., 292, 301, 302 Cantor’s ternary set, 8
analytic c.f., 191 ff. Cauchy d., 7, 8, 12, 18, 85, 91, 108, 121,
— —, boundary values of, 306 ISG ANS). SAS
— —, convexity property, 196, 197 Characteristic function, 11, 15
— —, factorization of, 236 — —, analytic, 191
— —, i.d., 258 ff. — —, boundary, 306 ff.
approximation theorem of Weierstrass, — —, decomposable, 103, 238, 254
28, 102, 328 — —, indecomposable, 103, 109, 179
arithmetic of d.f., 103, 104, 166, 182 ff., — —,i.d., 107, 108, 109, 121, 122
252 — —, periodic, 225
asymptotic expansions of stable densi- — —, table of, 18
ties, 148 — —, transformation of, 315, 318
characteristic, second (= second c.f.),
BERGSTROM, H., 142, 150, 337 26, 253
Bernoulli convolution, symmetric, 64, characterization, 254
66, 67 — of normal d., 254
Beta d., 7, 8, 18 —of normal, Poisson, conjugate Pois-
binomial d., 6, 14, 18 son d., 256
BiuM, J. R., 124, 337 Chebyshev polynomials, 216
Boas; Ra Ps, 25) 337 chi square d., 8
Bocuner, S., 71, 210, 337 Choquet’s theorem, 118
Bochner’s theorem, 71, 79 CHRISTENSEN, I. F., 224, 337
Bonr, H., 19, 36, 337 Cuunce, K. L., 99, 164, 337
BoreL, E., 222 closed limit inferior, 62
boundary c.f., 306 closure theorems, 111, 112
— —, extremities of, 309 complex conjugate, 15
— —, factorization of, 309 components of a d.f., 38

* The following abbreviations are used in the index: c.f. = characteristic function;
a.c.f. = analytic characteristic function; d = distribution; d.f. = distribution functions;
f. = function; id. = infinitely divisible; i.d.d. = infinitely divisible distribution.
INDEX 345
composition, 37 densities, stable, 148
concave, 91 density function, 5, 33
conjugate complex, 15 denumerable «-decomposition, 292
— d., 29, 30, 168 derivative, symmetric, 25
continuity interval, 2 differential equations, 227, 228
— point, 2, 29 — —, c.f. as solutions of, 228
— theorem, 47, 48, 88, 102 — —, positive definite, 228
— — for bounded non-decreasing func- Dirichlet conditions, 86
tions, 52 discontinuity point, 2
— —, second version of, 53 discontinuous part of a d.f., 3
continuous, 125 discrete part, 3, 4
—, absolutely, 4, 125 distribution:
— part of d.f., 3 Beta, 7, 8, 18
— singular, 125 binomial, 6, 14, 18
— (to the right), 2 Cauchy, 7, 8, 12, 18, 85, 91, 108,
convergence in the mean, 76 121, 137, 143, 326
—, radius of, 191 chi square, 8
—., uniform (of a sequence of c.f.), 50 compound Poisson, 316
—, weak, 43, 44, 48 degenerate, 5, 18, 19, 108, 121, 166
convergence theorem, 237, 315, 317 exponential, 8, 13
— —, dominated, 315, 317 Gamma, 7, 13, 18, 108, 109, 120,
— —, monotone, 237 ID ATO 198
convergent sequence of d.f., 42, 43 Gaussian (Gauss-Laplace), 7
convex, 83, 91 generalized Poisson, 316
— function, 202 geometric, 6, 113, 176
convexity properties of a.c.f., 196, 197 hypergeometric, 6
convolution, 37, 41 Laplace, 7, 13, 18, 88, 109, 122
— of normal and Poisson d., 245 negative binomial, 6, 14,18, 103,121
—, symmetric Bernoulli, 64, 66, 67 Neyman type A, 221
— theorem, 36 normal, 7, 13, 18, 91, 108, 121, 136,
—, v-fold, 316 143, 191, 213, 243, 254, 255
convolutions, infinite, 55-67 Pascal, 6
—, —, convergence criteria for, 59-61 Poisson, 6, 14, 18, 108, 121, 185,
—, —, convergence of, 59 191, 221, 227, 243
Corson, E. T., 203, 221, 337 rectangular, 7, 8, 109, 189
Craven... 4 tas i14, 21375675: Simpson’s, 7
169,210, 243, 292, 337 Student’s, 7, 12, 13
Cramér’s criterion, 73, 210 triangular, 7
— theorem, 243, 245, 254, 256, 302 uniform, 7
— —, converse of, 254 distribution:
criteria for c.f., 68, 210 bounded, 142, 201, 202
— — convergence of infinite convolu- conjugate, 29, 30, 168
tions, 59-61 finite, 142, 181, 200, 228, 259
Crum, M. M., 91, 337 i.d., 103
cumulant generating function, 26 lattice, 6, 17, 225, 245, 302
cumulants, 26 one-sided, 142, 200-202, 309
Cuppens, R., 289, 300, 305, 337 stable, 128 ff.
CzeErRnin, K. E., 158 symmetric, 30, 292
unimodal, 91-99
Darboux sums, 37, 112, 236, 237 distribution function:
decomposable c.f., 103, 238, 254 — —, absolutely continuous, 6, 39, 75
decomposition of d., 3 — —, discrete, 5, 6, 36, 41
— theorem, 169, 170 — —, indecomposable, 181, 183
— —, general, 169 — — of a.c.f., 197 ff.
De Finetti’s theorem, 112 — —, singular, 8, 9, 19
degenerate d., 5, 18, 19, 108, 121, 166 dominated convergence theorem, 315,
degree of freedom, 8 317
346 INDEX

DucuE, D., 86, 87, 302, 325, 337, 338 Fourier transforms, 11
freedom, degree of, 8
elliptic function, 227 frequency, Poisson, 263
energy parameter, 263 — functions of stable d., 138, 147
entire characteristic function, 195, 198, — —, stable, 138, 147
206, 209, 210, 239 function, Weierstrass’, 23
— — —, determination by properties of — of bounded variation, 320
factors, 253 ff.
— — —, factorization of, 238 Gamma d., 7, 13, 18, 108, 109, 120, 121,
— — — of infinite order, 210 179; 191
— —— —finite order and minimal Gauss—Laplace (law of), 7
type, 209 generating function, cumulant, 26
— — — — — — maximal type, 209 — —, moment, 10, 196, 197, 251
— — —, order of, 195, 198, 205-210 — —, probability, 10, 182
—— —, type of, 205-210 genus, 212
entire function, 195, 329 ff. geometric d., 6, 113, 176
equivalence (of c.f. or second c.f.), 253 GiL-PELAEz, J., 33, 338
EssEEN, C. G., 20, 54, 338 GirRAvuLT, M., 19) 20; 87,227, 321, 338
Evans, G. C., 14, 338 GNEDENKO, B. V., 54, 104, 128, 136,
exponent of convergence, 222 137, 143, 162, 165, 338
— — stable d., 133, 142, 147, 148 Go.pBerc, A. A., 280, 338
exponential d., 8, 13
— type, entire function of, 203 HapaMarp, J., 183
extension of factorization theorems for Hadamard’s factorization theorem, 203,
Eente, PAOD Taee 222, 243, 244
—-—non-negative definite functions, Haun, H., 101, 338
89-91 Hatmos, P. R., 62, 338
—_—— — — , indeterminate, 90 Harpy, G. H., 23, 83, 338
——— , unique, 90 HarTMANN, P., 124, 125, 338
extremities, formulae for, 202, 309 Hausporrtr, F., 62, 338
— of d.f. with boundary c.f., 309 Helly’s first theorem, 44, 49, 72, 173
extremity (left, right) of a d.f., 142 — second theorem, 45, 72, 82, 116
— — —,, extension of, 47, 49, 72, 116
Faa di Bruno’s formula, 27 Hermite polynomials, 78, 323
factor-closed family, 245 Hermitian, 71, 77
factor, indecomposable, 103, 170, 176 HERZ CHS. 19508
— of a.c.f., 236 Hiixz, E.; 330, 334, 338
——c.f., 103 HInNcIn, see KHINCHINE
factorization of a.c.f., 236 ff. Hosson, E. W., 2, 140, 339
— — boundary c.f., 310 hypergeometric d., 6
— — entire c.f. of finite order, 239
— — id. a.c.f., 258 ff. Io, 266, 280, 281 ff.
— problems, 103, 122, 167, 191, 236 ff. —, necessary conditions for member-
— theorem, Hadamard’s, 203, 222, 243, ship, 280, 281
244 —, sufficient conditions for member-
— —, Weierstrass’, 88 ship, 266, 281 ff.
Faltung, 37 Ipracimoy, I. A., 99, 150, 158, 165,
Fatou’s lemma, 229, 293, 294 258, 339
FELLER, W., 122, 142, 316, 338 imaginary part (Im), 192
Finetti’s theorem, 112 increase, point of, 2
finite «-decompositions, 292 indecomposable absolutely continuous
—d., 142, 181, 202, 228, 259 GiielS3
JOS, IM WAS. SSK: ein MU Oe) Tye), esl, WS)
Fourier coefficients, 36, 86 —iG,, Lol 182. 183
— integral, 193, 194 — factor, 103, 170, 176
— inversion theorem, 84 independent, rationally, 249, 287
— series, 86 infinite convolution, 55 ff.
INDEX 347
infinitely divisible a.c.f., 258 ff. lattice d., «-decomposition of, 302
— — boundary c.f., 310 ff. — points, 6, 225
—— c.f., definition, 107 L-class, 162
— — —, examples, 108 £-class, 162, 262-266, 280
— — —, factorization, 122 Lebesgue decomposition of c.f., 19
— — distribution, 108 — — theorem, 4
— — —, «-decomposition of, 299, 301 — measure, 66
— — —, one-sided, 312 — properties, 124, 190
—-—-—with bounded Poisson spec- left extremity, 142
trum, 281 ff. LRerrasiGy.9t 339
— — — with Gaussian component, 280, Levin, B. Ya., 210, 340
281 LEvINSON, N., 105, 340
IncHaM, A. I., 105, 339 Lévy, P., 63, 86, 89, 95, 136, 137, 143,
integrable, quadratically, 76 161, 165, 188, 211, 243, 245, 250,
integral function: see entire function 251, 252, 259, 340
integral representation, 191 Lévy canonical representation, 118, 121,
— — of a.c.f., 191-193 132, 163, 164, 176, 179
— — of boundary c.f., 306 Lévy—Khinchine canonical representa-
— — of stable densities, 148 tion, 117, 118, 121, 124, 125, 190
— transform, 9 Lewis, T., 190, 340
inversion formulae, 30 Li, 62
— — for absolutely integrable c.f., 33 Lrénarp, A., 183
— theorem, 31 Lim (weak limit), 43
iterated exponentials, 213 limit in the mean (l.i.m.), 76
— inferior, closed, 62
JESSEN, B., 19, 20, 64, 67, 339 — theorem, 107, 126-128
JOHANSEN, S., 118, 339 limits of d.f., 42 ff.
Jorpan, C., 27, 339 LINNIK, Yue Ve, 975 150) 2255227--228;
jump, 2 DB, PAD, 2, A, LY), 2D, SOP;
305, 340, 343
Kawata, T., 104, 107, 339 Lipschitz condition, 25
kernel, 10, 99 LITTLEwoop, J. E., 83, 338
KERSHNER, R., 19, 67, 339 LIvINGsTONE, A. E., 84, 342
KHINCHINE, A. YA., 87, 92, 104, 210 LoerFFEL, H., 321, 340
— phenomenon, 104 LokEve, M., 1, 4, i ISO, Ko, P37 SS,
Khinchine’s, criterion, 77, 210 340
— theorem, 262 ILGEENCS, 185, MOP, BIA, Pin. S05, S78.
Ko.tmocorov, A. N., 54, 128, 136, 137, 339, 340, 341
143, 162, 165, 326, 338
—, canonical representation, 119, 121, Maclaurin expansion, 23, 191
MaRCcINKIEWICZ, J., 213, 221, 224, 341
—, —— of ac.f., 260 —, theorem of, 213, 221, 224
KonuEmM, ASCE "419, 339 MarKUSHEVICH, A. I., 59, 240, 246, 296,
Krasner, M., 182, 183, 339 330, 341
KREIN, M. Ga 89, 339 Mathias, theorem of, 79
Krein’s theorem, 89 maximum modulus, 195, 301, 329
une, Ws MOS SOY mean, convergence in the, 76
KuratTowskl, K., 62, 339 Mepeyessy, P., 95, 142, 341
median, 167
Laua, R. G., 97, 293, 300, 305, 339 Muirter, H. D., 224, 341
LapPLace, P. S. DE, 10 mixtures of d., 315-318
Laplace d., 7, 13, 18, 88, 109, 122 modulus, maximum, 195, 301, 329
—, first law of, 7 Moivre—Laplace, law of, 7
— integral, 193 moment, 10, 11 ff., 13, 20 ff., 25, 42, 198
—, second law of, 7 —, algebraic, 10
— transform, 193 —, absolute, 10, 25, 26
lattice d., 6, 17, 225, 245, 302 —, factorial, 10
348 INDEX

moment generating function, 10, 196, PRINGSHEIM, A., 84

198, 251 probability density function, 5, 7
—, symmetric, 25 — generating function, 10, 182
moments, table of, 13 product, canonical, 222
monotone convergence theorem, 237 — representation, trivial, 103
pure d.f., 4,5
Nal f(O], Na{f], 167 —type, c.f. of, 19
Na-value, 167 purely discrete d.f., 17
Natanson, I. P., 329, 341
negative binomial d., 6, 14, 18, 108, 121 Quadratically integrable function, 76
Neyman type A d., 221 quasi-stable, 137
Noack, A., 14, 341
non-degenerate d., 202 radius of convergence, 191
non-negative definite function, 70, 88 Rarkov, D. A., 89, 183, 243, 245, 249,
— — — — on finite interval, 89 250, 251, 292, 341, 342
normal’ d., 7, 13, 18, 91, 108, 121, 136, Raikov’s theorem, 243, 245, 303
143, 191, 213, 235, 242, 243, 254-5 RAMACHANDRAN, B., 179, 210, 300, 305,
342
O, 0, 327 RANu.ac, B., 183, 339
one-sided d., 142, 200, 202, 309 rationally independent, 249
operators which transformc.f. into other — — points, set with, 287
CHOON S 216323 real part (Re), 17
order (of entire function), 196, 204, 239, rectangular d., 7, 13, 18, 109, 189
301, 329-330 reflection principle, Schwarz’s, 307,
order and type of entire function, 329- 311, 334
330 regular, 191, 193
Ostrovskil, I. V., 195, 224, 225, 280, regularity, strip of, 191, 193, 236, 238,
289, 290, 291, 341
ridge functions, 195, 224
Paey, R. E. A. C., 76, 341 — property, 195, 211, 224
Parseval’s theorem, 76, 105 — — of boundary c.f., 308
parts of a d.f., 4 Riemann-—Lebesgue lemma, 19
Pascal d., 6 Riesz, F., 91, 342
periodic c.f., 225 Riesz, M., 84, 342
Pitman, E. J. G., 24, 25, 341 right continuous, 2
Plancherel’s theorem, 76 — extremity, 142
point of increase, 2 Rossins, H., 315, 342
— spectrum of a d.f., 56, 57 ROSENBLATT, M., 124, 337
Poisson d., 6, 14, 18, 108, 121, 185, 191, ROSENTHAL, A., 101, 338
221, 227, 243 RossBeroc, H. J., 210, 342
— —, «-decomposition of, 303
— —, compound, 316 SALEM, R., 9, 342
— —, generalized, 316 saltus, 2, 35, 36
— frequencies, 263 Sapocoy, N. A., 245, 342
— spectrum, 263, 281 ff. SCHMETTERER, L., 86, 342
—-— (bounded, denumerable, finite, ScHwarzZ, L., 20, 342
negative, positive), 263 Schwarz’s inequality, 14, 328
—type d., 112 — reflection (symmetry) principle, 307,
Pozya, G., 13, 83, 338, 341 311, 334
— condition, 83, 85 second characteristic, 26, 253
— theorem, 105 self-decomposable, 161-165
— type c.f., 85, 87, 104 semi-invariant, 26
polynomial, Chebyshev, 216 semi-stable, 165
—, trigonometric, 28, 328 set a rationally independent points,
positive definite differential equation, 87
228 set-theoretic notations, 55
prime factor, 170, 180, 182, 183 sets, vectorial sum of, 57
INDEX 349
SHIMIzuU, R., 179, 342 closure, 111
SHouat, J. A., 12, 198, 342 continuity, 48, 53
Simpson’s d., 7 convolution, 37
singular, 4, 8, 19, 20, 36, 64 Cramér’s, 73, 243, 254
SKOROHOD, A. V., 150, 342 De Finetti’s, 112
Smirnov, N. V., 143 Fourier inversion, 84
spectral function, 263 Helly’s first, 44
spectrum of a d.f., 56-57 — second, 45
—, Poisson, 263 — —,, extension of, 47
—, —, bounded, 263, 281, 282 inversion, 31
—,—, denumerable, 263 Khinchine’s, 77
—, —, finite, 263 ee representation,
—, —, negative, 263
—, —, positive, 263 Krein’s, 89
stable c.f., 128 Lévy’s representation, 118
— densities, 148 Lévy—Khinchine’s representation,
—-—, asymptotic expansion of, 148- 117
150 limit, 126
——,, integral representation of, 148, Marcinkiewicz’s, 213, 221
150-158 Mathias’, 79
stable d., 128, 179, 191 Pélya’s, 83
— — in the restricted sense, 137 Raikov’s, 243
— frequency function, 138 uniqueness for c.f., 28
— — —,, order of, 147-148 Weierstrass’ approximation, 328
— — —,, type of, 147-148 TitcHMarsH, E. C., 19, 31, 49, 59, 84,
— type, 129 86, 212, 229, 246, 293, 296, 342
step function, 3 transformations of c.f., 315, 318
Stirling’s formula, 80, 144 triangular d., 7
strip of regularity of an a.c.f., 191, 193, trigonometric polynomial, 28, 328
236, 238 trivial product representation, 103
— — — — boundary c.f., 308, 310 Tucker, H. G., 125, 126, 190, 343
strongly unimodal, 99 type of c.f., 253
Student’s d., 7, 12, 13 — — entire c.f. of order 1, 202, 203
sum, vectorial (of sets), 57 ——d.f., 16
summability, Abel, 41 —, stable, 129
aS) (C,1), At
SUN, D.C. 165, 342 uniform convergence (of sequence of
symmetric Bernoulli convolution, 64— Ce) HC
—d.,7
— d., 30, 292 unimodal d., 91, 95
— kth derivative, 25 —, strongly, 99
— moment, 25 unimodality of stable d., 158
symmetry principle of Schwarz, 307, uniqueness theorem, 28 ff.
311, 334 — — for almost periodic f., 36
SzAsz, O., 212, 340, 341
S$zEG6, G., 13, 341 VARADARAJAN, V. S., 125, 338
vectorial sum (of sets), 57
table of canonical representations of vertex (of unimodal d.), 91
reel
yenin, 1/2)
— — discrete d., 6 weak convergence, 43, 44, 48
— — frequency f., 7 Weierstrass’ approximation theorem,
—-— moments, 13 28. 1023284529
Takano, K., 212, 342 — factorization theorem, 88
TAMARKIN, J. D., 12, 198, 342 — function, 23
TrrcHer, H., 245, 342 weight function, 316
theorem: Weiss, B., 319, 339
Bochner’s, 71, 79 Wiper, D. V., 193, 194, 343
350 INDEX

WIENER, N., 76, 341 ZINGER, A. A., 227, 228, 235, 302, 343
Wintner, A., 19, 20, 24, 64, 67, 124, ZoroTarev, V. M., 142, 143, 157, 165,
125, 165, 338, 339, 343 343
ZYGMUND, A., 22, 25, 92, 203, 310,
ZEMANIAN, A. H., 212, 343 343
 ST. JOHN FISHER COLLEGE LIBRARY
ois ae 1970 |
010101 000 |

aa

DATE DUE .

ILL: /20(7G 7 Miedo?

 eN\
ST. JOHN FISHER COLLEGE LIBRAR\
 ST. JOHN FISHER COLLEGE LIBRARY
QA273.6 .L85 1970
Lukacs, Eugene. cn 010101 000
SOO a ceeee
Characteristic functions.
0 1219 0043945 O