[www.gfxmad.me] 9814335479 probabilit
TRANSCRIPT
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
1/90
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
2/90
This page intentionally left blankThis page intentionally left blank
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
3/90
N E W J E R S E Y L O N D O N S I N G A P O R E B E I J I N G S H A N G H A I H O N G K O N G T A I P E I C H E N N A I
World Scientifc
Narahari PrabhuCornell University, USA
TOPICS INPROBABILITY
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
4/90
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN-13 978-981-4335-47-8
ISBN-10 981-4335-47-9
Typeset by Stallion Press
Email: [email protected]
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
Copyright 2011 by World Scientific Publishing Co. Pte. Ltd.
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Printed in Singapore.
TOPICS IN PROBABILITY
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
5/90
Now Ive understood
Times magic play:
Beating his drum he rolls out the show,
Shows different images
And then gathers them in again
Kabir (14501518)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
6/90
This page intentionally left blankThis page intentionally left blank
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
7/90
CONTENTS
Preface ix
Abbreviations xi
1. Probability Distributions 11.1. Elementary Properties . . . . . . . . . . . . . . . . . 1
1.2. Convolutions . . . . . . . . . . . . . . . . . . . . . . . 4
1.3. Moments . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4. Convergence Properties . . . . . . . . . . . . . . . . . 8
2. Characteristic Functions 11
2.1. Regularity Properties . . . . . . . . . . . . . . . . . . 112.2. Uniqueness and Inversion . . . . . . . . . . . . . . . . 15
2.3. Convergence Properties . . . . . . . . . . . . . . . . . 17
2.3.1. Convergence of types . . . . . . . . . . . . . . 19
2.4. A Criterion for c.f.s . . . . . . . . . . . . . . . . . . . 21
2.5. Problems for Solution . . . . . . . . . . . . . . . . . . 24
3. Analytic Characteristic Functions 27
3.1. Definition and Properties . . . . . . . . . . . . . . . . 27
3.2. Moments . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3. The Moment Problem . . . . . . . . . . . . . . . . . . 31
3.4. Problems for Solution . . . . . . . . . . . . . . . . . . 40
vii
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
8/90
viii Topics in Probability
4. Infinitely Divisible Distributions 43
4.1. Elementary Properties . . . . . . . . . . . . . . . . . 43
4.2. Feller Measures . . . . . . . . . . . . . . . . . . . . . 464.3. Characterization of Infinitely Divisible
Distributions . . . . . . . . . . . . . . . . . . . . . . . 50
4.4. Special Cases of Infinitely Divisible Distributions . . 54
4.5. Levy Processes . . . . . . . . . . . . . . . . . . . . . . 57
4.6. Stable Distributions . . . . . . . . . . . . . . . . . . . 58
4.7. Problems for Solution . . . . . . . . . . . . . . . . . . 66
5. Self-Decomposable Distributions; Triangular Arrays 69
5.1. Self-Decomposable Distributions . . . . . . . . . . . . 69
5.2. Triangular Arrays . . . . . . . . . . . . . . . . . . . . 72
5.3. Problems for Solution . . . . . . . . . . . . . . . . . . 78
Bibliography 79
Index 81
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
9/90
PREFACE
In this monograph we treat some topics that have been of some
importance and interest in probability theory. These include, in
particular, analytic characteristic functions, the moment problem,
infinitely divisible and self-decomposable distributions.
We begin with a review of the measure-theoretical foundations
of probability distributions (Chapter 1) and characteristic functions
(Chapter 2).
In many important special cases the domain of characteristic func-
tions can be extended to a strip surrounding the imaginary axis
of the complex plane, leading to analytic characteristic functions.
It turns out that distributions that have analytic characteristic func-
tions are uniquely determined by their moments. This is the essence
of the moment problem. The pioneering work in this area is due toC. C. Heyde. This is treated in Chapter 3.
Infinitely divisible distributions are investigated in Chapter 4. The
final Chapter 5 is concerned with self-decomposable distributions and
triangular arrays. The coverage of these topics as given by Feller in
his 1971 book is comparatively modern (as opposed to classical) but
is still somewhat diffused. We give a more compact treatment.
N. U. Prabhu
Ithaca, New York
January 2010
ix
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
10/90
This page intentionally left blankThis page intentionally left blank
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
11/90
ABBREVIATIONS
Term Abbreviation
characteristic function c.f.
distribution function d.f
if and only if iff
Laplace transform L.T.
probability generating function p.g.f
random variable r.v.
Terminology: We write xd
= y if the r.v.s x, y have the same distribution.
xi
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
12/90
Chapter 1
Probability Distributions
1.1. Elementary Properties
A function F on the real line is called a probability distribution
function if it satisfies the following conditions:
(i) F is non-decreasing: F(x + h) F(x) for h > 0;(ii) F is right-continuous: F(x+) = F(x);
(iii) F() = 0, F() 1.
We shall say that F is proper if F() = 1, and F is defective other-wise.
Every probability distribution induces an assignment of proba-
bilities to all Borel sets on the real line, thus yielding a proba-bility measure P. In particular, for an interval I = (a, b] we have
P{I} = F(b) F(a). We shall use the same letter F both for thepoint function and the corresponding set function, and write F{I}instead of P{I}. In particular
F(x) = F{(, x]}.
We shall refer to F as a probability distribution, or simply a distri-
bution.
A point x is an atom if it carries positive probability (weight). It is
a point of increaseiffF{I} > 0 for every open interval I containing x.
1
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
13/90
2 Topics in Probability
A distribution F is concentrated on the set A if F(Ac) = 0, where
Ac is the complement of A. It is atomic if it is concentrated on the
set of its atoms. A distribution without atoms is continuous.As a special case of the atomic distribution we have the arithmetic
distribution which is concentrated on the set {k(k = 0, 1, 2, . . .)}for some > 0. The largest with this property is called the span
of F.
A distribution is singular if it is concentrated on a set of Lebesgue
measure zero. Theorem 1.1 (below) shows that an atomic distribution
is singular, but there exist singular distributions which are continu-
ous.
A distribution F is absolutely continuous if there exists a function
f such that
F(A) =
A
f(x)dx.
If there exists a second function g with the above property, then it is
clear that f = g almost everywhere, that is, except possibly on a setof Lebesgue measure zero. We have F(x) = f(x) almost everywhere;
f is called the density of F.
Theorem 1.1. A probability distribution has at most countably
many atoms.
Proof. Suppose F has n atoms x1, x2, . . . , xn in I = (a, b] witha < x1 < x2 < < xn b and weights p(xk) = F{xk}. Thenn
k=1
p(xk) F{I}.
This shows that the number of atoms with weights > 1n is at most
equal to n. Let
Dn = {x : p(x) > 1/n};
then the set Dn has at most n points. Therefore the set D = Dn isat most countable.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
14/90
Probability Distributions 3
Theorem 1.2 (Jordan decomposition). A probability distribu-
tion F can be represented in the form
F = pFa + qFc (1.1)
where p 0, q 0, p + q = 1, Fa, Fc are both distributions, Fa beingatomic and Fc continuous.
Proof. Let {xn, n 1} be the atoms and p =
p(xn), q = 1 p.If p = 0 or if p = 1, the theorem is trivially true. Let us assume that
0 < p < 1 and for < x < define the two functions
Fa(x) =1
p
xnx
p(xn), Fc(x) =1
q[F(x) pFa(x)]. (1.2)
Here Fa is a distribution because it satisfies the conditions (i)(iii)
above. For Fc we find that for h > 0
q[Fc(x + h) Fc(x)] = F(x + h) F(x) x
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
15/90
4 Topics in Probability
Proof. By the Lebesgue decomposition theorem on measures we
can express F as
F = aFs + bFac, (1.5)
where a 0, b 0, a + b = 1, Fs is a singular distribution andFac is an absolutely continuous distribution. Applying Theorem 1.2
to Fs we find that Fs = p1Fa + q1Fsc, where p1 0, q1 0, p1 +q1 = 1. Writing p = ap1, q = aq1, r = b we arrive at the desired
result (1.4).
Remark. Although it is possible to study distribution functionsand measures without reference to random variables (r.v.) as we have
done above, it is convenient to start with the definition
F(x) = P{X x}where X is a random variable defined on an appropriate sample
space.
1.2. Convolutions
Let F1, F2 be distributions and F be defined by
F(x) =
F1(x y)dF2(y) (1.6)
where the integral obviously exists. We call F the convolution of F1and F2 and write F = F1 F2. Clearly F1 F2 = F2 F1.Theorem 1.4. The function F is a distribution.
Proof. For h > 0 we have
F(x + h) F(x) =
[F1(x y + h) F1(x y)]dF2(y) 0
(1.7)
so that F is non-decreasing. As h 0,F1(x y + h) F1(x y) F1(x y+) F1(x y) = 0;
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
16/90
Probability Distributions 5
since
|F1(x y + h) F1(x y)| 2,
2dF2(y) = 2,
the right side of (1.7) tends to 0 by the dominated convergence theo-
rem. Therefore F(x+)F(x) = 0, so that F is right-continuous. SinceF1() = 1 the dominated convergence theorem gives F() = 1 .Similarly F() = 0. Therefore F is a distribution.
Theorem 1.5. If F1 is continuous, so is F. If F1 is absolutely
continuous, so is F.
Proof. We have seen in Theorem 1.4 that the right-continuity of
F1 implies the right-continuity of F. Similarly the left-continuity of
F1 implies that of F. It follows that if F1 is continuous, so is F.
Next let F1 be absolutely continuous, so there exists a function
f1 such that
F1(x) =
x
f1(u)du.
Then
F(x) =
dF2(y)
x
f1(u y)du
=
x
f1(u y)dF2(y) duso that F is absolutely continuous, with density
f(x) =
f1(x y)dF2(y). (1.8)
Remarks.1. If X1, X2 are independent random variables with distributions
F1, F2, then the convolution F = F1 F2 is the distribution of their
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
17/90
6 Topics in Probability
sum X1 + X2. For
F(z) = P{
X1
+ X2
z}
= x+yz
dF1
(x)dF2
(y)
=
dF2(y)
zy
dF1(x) =
F1(z y)dF2(y).
However, it should be noted that dependent random variables X1, X2may have the property that the distribution of their sum is given by
the convolution of their distributions.
2. The converse of Theorem 1.5 is false. In fact two singular distri-
butions may have a convolution which is absolutely continuous.
3. The conjugate of any distribution F is defined as the distribution
F, where
F(x) = 1 F(x).If F is the distribution of the random variable X, then F is the
distribution of
X. The distribution F is symmetric if F = F.
4. Given any distribution F, we can symmetrize it by defining thedistribution F, where
F = F F .It is seen that F is a symmetric distribution. It is the distribution
of the difference X1 X2, where X1, X2 are independent variableswith the same distribution F.
1.3. Moments
The moment of order > 0 of a distribution F is defined by
=
xdF(x)
provided that the integral converges absolutely, that is,
=
|x|dF(x) < ;
is called the absolute moment of order . Let 0 < < . Then
for |x| 1 we have |x| 1, while for |x| > 1 we have |x| |x|.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
18/90
Probability Distributions 7
Thus we can write |x| |x| + 1 for all x and so
|x|dF(x)
(1 + |x|)dF(x) = 1 +
|x|dF(x).
This shows that the existence of the moment of order implies the
existence of all moments of order < .
Theorem 1.6. The moment of a distribution F exists iff
x1[1
F(x) + F(
x)] (1.9)
is integrable over (0, ).
Proof. For t > 0 an integration by parts yields the relation
tt
|x|dF(x) = t[1 F(t) + F(t)]
+ t0
x1[1 F(x) + F(x)]dx. (1.10)From this we find thatt
t|x|dF(x)
t0
x1[1 F(x) + F(x)]dx
so that if (1.9) is integrable over (0, ), (and therefore ) exists.Conversely, if exists, then since
|x|>t|x|dF(x) > |t|[1 F(t) + F(t)]
the first term on the right side of (1.10) vanishes as t and theintegral there converges as t .
Theorem 1.7. Let
(t) =
|x|tdF(x) <
for t in some interval I. Then log (t) is a convex function of t I.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
19/90
8 Topics in Probability
Proof. Let a 0, b 0, a + b = 1. Then for two functions 1, 2we have the Holder inequality
|1(x)2(x)|dF(x)
|1(x)|1/adF(x)
a
|2(x)|1/bdF(x)
b
provided that the integrals exist. In this put 1(x) = xat1 , 2(x) =
xbt2 , where t1, t2
I. Then
(at1 + bt2) (t1)a(t2)b (1.11)or taking logarithms,
log (at1 + bt2) a log (t1) + b log (t2)which establishes the convexity property of log .
Corollary 1.1 (Lyapunovs inequality). Under the hypothesis of
Theorem 1.7, 1
t
t is non-decreasing for t I.
Proof. Let , I and choose a = /, t1 = , b = 1 a, t2 = 0.Then (1.11) reduces to
/ ( )where we have written t = (t).
1.4. Convergence Properties
We say that I is an interval of continuity of a distribution F if I is
open and its end points are not atoms ofF. The whole line (, )
is considered to be an interval of continuity.Let {Fn, n 1} be a sequence of proper distributions. We saythat the sequence converges to F if
Fn{I} F{I} (1.12)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
20/90
Probability Distributions 9
for every bounded interval of continuity of F. If (1.12) holds for
every (bounded or unbounded) interval of continuity of F, then the
convergence is said to be proper, and otherwise improper. Properconvergence implies in particular that F() = 1.Examples
1. Let Fn be uniform in (n, n). Then for every bounded intervalcontained in (n, n) we have
Fn{I} =I
dx
2n=
|I|2n
0 as n
where |I| is the length of I. This shows that the convergence isimproper.
2. Let Fn be concentrated on { 1n , n} with weight 1/2 at each atom.Then for every bounded interval I we have
Fn{I} 0 or 1/2according as I does not or does contain the origin. Therefore the
limit F is such that it has an atom at the origin, with weight 1/2.Clearly F is not a proper distribution.
3. Let Fn be the convolution of a proper distribution F with the
normal distribution with mean zero and variance n2. Thus
Fn(x) =
F(x y) n2
e(1/2)n2y2dy
= F(x y/n)
12 e
(1/2)y2
dy.
For finite a, b we haveba
dFn(x) =
[F(b y/n) F(a y/n)] 12
e(1/2)y2
dy
F(b) F(a) as n
by the dominated convergence theorem. Ifa, b are points of continuityof we can write
Fn{(a, b)} F{(a, b)} (1.13)so that the sequence {Fn} converges properly to F.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
21/90
10 Topics in Probability
If X is a random variable with the distribution F and Yn is an
independent variable with the above normal distribution, then we
know that Fn is the distribution of the sum X + Yn. As n , itis obvious that the distribution of this sum converges to that of X.
This justifies the definition of convergence which requires (1.13) to
hold only for points of continuity a, b.
Theorem 1.8 (Selection theorem). Every sequence {Fn} of dis-tributions contains a subsequence {Fnk , k 1} which converges(properly or improperly) to a limit F.
Theorem 1.9. A sequence{Fn} of proper distributions converges toF iff
u(x)dFn(x)
u(x)dF(x) (1.14)
for every function u which is bounded, continuous and vanishing at
. If the convergence is proper, then (1.14) holds for every boundedcontinuous function u.
The proofs of these two theorems are omitted.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
22/90
Chapter 2
Characteristic Functions
2.1. Regularity Properties
Let F be a probability distribution. Then its characteristic function
(c.f.) is defined by
() =
eixdF(x) (2.1)
where i =1, real. This integral exists, since
|eix|dF(x) =
dF(x) = 1. (2.2)
Theorem 2.1. A c.f. has the following properties:
(a) (0) = 1 and |()| 1 for all .(b) () = (), and is also a c.f.(c) Re is also a c.f.
Proof. (a) We have
(0) =
dF(x) = 1, |()|
|eix|dF(x) = 1.
11
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
23/90
12 Topics in Probability
(b) () = e
ixF(dx) = (). Moreover, let F(x) =1
F(
x). Then
eixF{dx} =
eixF{dx} =
eixF{dx}.
Thus () is the c.f. of F, which is a distribution.(c) Re = 12 +
12 = c.f. of
12F +
12 F, which is a distribution.
Theorem 2.2. If 1, 2 are c.f.s, so is their product 12.
Proof. Let 1, 2 be the c.f.s of F1, F2 respectively and considerthe convolution
F(x) =
F1(x y)dF2(y).
We know that F is a distribution. Its c.f. is given by
() =
eixdF(x) =
eix
dF1(x y)dF2(y)
=
eiydF2(y)
ei(xy)dF1(x y)
= 1()2().
Thus the product 12 is the c.f. of the convolution F1 F2. Corollary 2.1. If is a c.f., so is ||2.
Proof. We can write ||2
= , where is a c.f. byTheorem 2.1(b).
Theorem 2.3. A distribution F is arithmetic iff there exists a real
0 = 0 such that (0) = 1.Proof. (i) Suppose that the distribution is concentrated on {k, > 0, k = 0, 1, 2, . . .} with the weight pk at k. Then the c.f.is given by
() =
pkeik.
Clearly (2/) = 1.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
24/90
Characteristic Functions 13
(ii) Conversely, let (0) = 1 for 0 = 0. This gives
(1 e
i0x
)dF(x) = 0.
Therefore
(1 cos 0x)dF(x) = 0
which shows that the points of increase of F are among 2k0 (k =
0, 1, 2, . . .). Thus the distribution is arithmetic. Corollary 2.2. If () = 1 for all , then the distribution is con-
centrated at the origin.
Remarks.
1. If F is the distribution of a random variable, then we can write
() = E(eiX)
so that the c.f. is the expected value of eiX
. We have () =E(eiX), so that () is the c.f. of the random variable X. Thisis Theorem 2.1(b).
2. If X1, X2 are two independent random variables with c.f.s 1, 2,
then
1()2() = E[ei(X1+X2)]
so that the product 12 is the c.f. of the sum X1 + X2. This is only
a special case of Theorem 2.2, since the convolution F1 F2 is notnecessarily defined for independent random variables.
3. If is the c.f. of the random variable X, then ||2 is the c.f. ofthe symmetrized variable X1 X2, where X1, X2 are independentvariables with the same distribution as X.
Theorem 2.4. (a) is uniformly continuous.
(b) If the n-th moment exists, then the n-th derivative exists and isa continuous function given by
(n)() =
eix(ix)ndF(x). (2.3)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
25/90
14 Topics in Probability
(c) If the n-th moment exists, then admits the expansion
() = 1 +
n1
n(i)n
n! + 0(n
) ( 0). (2.4)Proof. (a) We have
( + h) () =
eix(eihx 1)dF(x) (2.5)
so that
|( + h) ()|
|eihx 1|dF(x)
2
| sin(hx/2)|dF(x).
Now xB
| sin(hx/2)|dF(x) xB
dF(x) <
by taking A, B large, whileBA
| sin(hx/2)|dF(x) BA
dF(x) < .
since | sin(hx/2)| < for h small. Therefore |( +h)()| 0as h 0, which proves uniform continuity.
(b) We shall prove (2.3) for n = 1, the proof being similar for n > 1.
We can write (2.5) as
( + h) ()h
=
eix eihx 1
hdF(x). (2.5)
Here eix e
ihx 1h
eihx 1h
|x|
and
|x|dF(x) <
by hypothesis. Moreover (eihx 1)/h ix as h 0. Thereforeletting h 0 in (2.5) we obtain by the dominated convergence
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
26/90
Characteristic Functions 15
theorem that
( + h) ()h
ixeixdF(x)
as required. Clearly, this limit is continuous.
(c) We have
eix =n
n=0
(ix)n
n!+ o(nxn) ( 0)
so that
eixdF(x) = 1 +
nn=1
(i)n
n!n +
o(nxn)dF(x),
where the last term on the right side is seen to be o(n).
Remark . The converse of (b) is not always true: thus () may
exist, but the mean may not. A partial converse is the following:
Suppose that (n)() exists. Ifn is even, then the first n moments
exist, while if n is odd, the first n 1 moments exist.2.2. Uniqueness and Inversion
Theorem 2.5 (uniqueness). Distinct distributions have distinct
c.f.s.
Proof. Let F have the c.f. , so that
() =
eixdF(x).
We have for a > 0
a2
e1
2a22iy()d
=
a2
e1
2a22iy
eixdF(x)
=
dF(x)
ei(xy)
a2
e1
2a22d,
the inversion of integrals being clearly justified. The last integral is
the c.f. (evaluated at x y) of the normal distribution with mean 0
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
27/90
16 Topics in Probability
and variance a2, and therefore equals e(xy)2/2a2 . We therefore
obtain the identity
1
2
e
1
2a22iy()d =
12a
e1
2a2(yx)2dF(x)
(2.6)
for all a > 0. We note that the right side of (2.6) is the density of the
convolution F Na, where Na is the normal distribution with mean0 and variance a2. Now if G is a second distribution with the c.f. ,
it follows from (2.6) that F Na = G Na. Letting a 0+ we findthat F G as required. Theorem 2.6 (inversion). (a) If the distribution F has c.f. and
|()/| is integrable, then for h > 0
F(x + h) F(x) = 12
eix 1 eih
i()d. (2.7)
(b) If || is integrable, then F has a bounded continuous density fgiven by
f(x) =1
2
eix()d. (2.8)
Proof. (b) From (2.6) we find that the density fa of Fa = F Nais given by
fa(x) = 12
e1
2a2
2
ix()d. (2.9)
Here the integrand is bounded by |()|, which is integrable byhypothesis. Moreover, as a 0+, the integrand eix(). There-fore by the dominated convergence theorem as a 0+,
fa(x) 12
eix()d = f(x) (say).
Clearly, f is bounded and continuous. Now for every bounded
interval I we have
Fa{I} =I
fa(x)dx.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
28/90
Characteristic Functions 17
Letting a 0+ in this we obtain
F{I} = I
f(x)dx
if I is an interval of continuity of F. This shows that f is the density
of F, as required.
(a) Consider the uniform distribution with density
uh(x) =1
h
for
h < x < 0, and = 0 elsewhere.
Its convolution with F has the density
fh(x) =
uh(x y)dF(y) =x+hx
1
hdF(y) =
F(x + h) F(x)h
and c.f.
h() = ()
eixuh(x)dx = () 1 eih
ih.
By (b) we therefore obtain
F(x + h) F(x)h
=1
2
eix() 1 eih
ihd
provided that |()(1 eih)/i| is integrable. This conditionreduces to condition that |()/| is integrable.
2.3. Convergence Properties
Theorem 2.7 (continuity theorem). A sequence
{Fn
}of distri-
butions converges properly to a distribution F iff the sequence {n}of their c.f.s converges to , which is continuous at the origin. In
this case is the c.f. of F.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
29/90
18 Topics in Probability
Proof. (i) If{Fn} converges properly to F, then
u(x)dFn(x)
u(x)dF(x)
for every continuous and bounded function u. For u(x) = eix
it follows that n() () where is the c.f. of F. FromTheorem 2.4(a) we know that is uniformly continuous.
(ii) Conversely suppose that n() (), where is continuous atthe origin. By the selection theorem there exists a subsequence{Fnk , k 1} which converges to F, a possibly defective distri-bution. Using (2.6) we have
a2
eiy1
2a22nk()d =
e1
2a2(yx)2dFnk(x).
Letting k in this we obtaina2
eiy
1
2a22()d =
e
1
2a2(yx)2dF(x)
F() F(). (2.10)
Writing the first expression in (2.10) as
1
2
ei(y/a)1
22(/a)d (2.11)
and applying the dominated convergence theorem we find that (2.11)
converges to (0) = 1 as a . By (2.10) it follows that F() F() 1, which gives F() = 0, F() = 1, so that F is proper.By (i) is the c.f. of F, and by the uniqueness theorem F is unique.
Thus every subsequence {Fnk} converges to F.
Theorem 2.8 (weak law of large numbers). Let {Xn, n 1}be a sequence of independent random variables with a common dis-
tribution and finite mean . Let Sn = X1 + X2 + + Xn (n 1).Then as n , Sn/n in probability.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
30/90
Characteristic Functions 19
Proof. Let be the c.f. of Xn. The c.f. of Sn/n is then
E(ei(Sn/n)
) = (/n)n
= [1 + i(/n) + 0(1/n)]n
ei
as n . Here ei is the c.f. of a distribution concentrated at thepoint . By the continuity theorem it follows that the distribution of
Sn/n converges to this degenerate distribution.
Theorem 2.9 (central limit theorem). Let {Xn, n 1} be asequence of independent random variables with a common distribu-
tion and
E(Xn) = , Var(Xn) = 2
(both being finite). Let Sn = X1 + X2 + + Xn (n 1). Then asn , the distribution of (Snn)/
n converges to the standard
normal.
Proof. The random variables (Xn)/ have mean zero and vari-ance unity. Let their common c.f. be . Then the c.f. of (Sn n)/
n is
(/
n)n = [1 2/2n + 0(1/n)]n e 122
where the limit is the c.f. of the standard normal distribution. The
desired result follows by the continuity theorem.
Remark . In Theorem 2.7 the convergence of n is uniformwith respect to in [, ].
2.3.1. Convergence of types
Two distributions F and G are said to be of the same type if
G(x) = F(ax + b) (2.12)
with a > 0, b real.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
31/90
20 Topics in Probability
Theorem 2.10. If for a sequence {Fn} of distributions we haveFn(nx + n)
G(x), Fn(anx + bn)
H(x) (2.13)
for all points of continuity, with n > 0, an > 0, and G and H are
non-degenerate distributions, then
nan
a, n bnan
b and G(x) = H(ax + b) (2.14)
(0 < a < , |b| < ).
Proof. Let Hn(x) = Fn(anx + bn). Then we are given thatHn(x) H(x) and also Hn(nx + n) = Fn(nx + n) G(x),where
n =nan
, n =n bn
an. (2.15)
With the obvious notations we are given that
n() (), n() ein/nn(/n) ()uniformly in . Let {nk} be a subsequence of{n} suchthat nk a (0 a ). Let a = , then
|()| = lim |nk()| = lim |nk(/nk)| = |(0)| = 1uniformly in [, ], so that is degenerate, which is not true. If
a = 0, then|()| = lim |nk()| = lim |nk(nk)| = |(0)| = 1,
so that is degenerate, which is not true. So 0 < a < . Now
ei(nk /nk ) =nk()
nk() ()
()
so that nk/nk
a limit b/a (say). Also
() = ei(b/a)(/a). (2.16)
It remains to prove the uniqueness of the limit a. Suppose there are
two subsequences of {n} converging to a and a, and assume that
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
32/90
Characteristic Functions 21
a < a. Then the corresponding subsequences of{bn} converge to b, b(say) From (2.16) we obtain
ei(b/a)(/a) = ei(b/a)(/a)
and hence |(/a)| = |(/a)| or|()| = |(a/a)| = |(a2/a2)| = = |(an/an)| = |(0)| = 1.This means that is degenerate, which is not true. So a a.
Similarly a a. Therefore a = a, as required. Since we have
proved (2.16), the theorem is completely proved.
2.4. A Criterion for c.f.s
A function f of a real variable is said to be non-negative definite in
(, ) if for all real numbers 1, 2, . . . , n and complex numbersa1, a2, . . . , an
nr,s=1
f(r s)aras 0. (2.17)
For such a function the following properties hold.
(a) f(0) 0. If in (2.17) we put n = 2, 1 = , 2 = 0, a1 = a, a2 = 1we obtain
f(0)(1 + |a|2) + f()a + f()a 0. (2.18)When = 0 and a = 1 this reduces to f(0) 0.
(b) f() = f(). We see from (2.18) that f()a + f()a is real.This gives f() = f().
(c) |f()| f(0). In (2.18) let us choose a = f() where is real.Then
f(0) + 2|f()|2 + 2|f()|2f(0) 0.
This is true for all , so |f()|4
|f()|2
[f(0)]2
or |f()| f(0), asrequired.
Theorem 2.11. A function of a real variable is the c.f. of a dis-
tribution iff it is continuous and non-negative definite.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
33/90
22 Topics in Probability
Proof. (i) Suppose is a c.f.; that is,
() = e
ix
dF(x)
where F is a distribution. By Theorem 2.4(a), is continuous.
Moreover,
nr,s=1
(r s)aras
=n
r,s=1
aras
ei(rs)xdF(x)
=
n1
areirn
n1
aseisx
dF(x)
=
n
i
ar
eirx2
dF(x)
0
which shows that is non-negative definite.
(ii) Conversely, let be continuous and non-negative definite. Then
considering the integral as the limit of a sum we find that
0
0ei(
)x( )dd 0 (2.19)
for > 0. Now consider
P(x) =1
0
0
e()x( )dd
=
eisx(s)ds (2.20)
where
(t) =
1 |t|
(t) for |t| 0 for |t|
.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
34/90
Characteristic Functions 23
From (2.20) we obtain
(t) =
1
2
1 |
| eitP()d=
1
2
(s)ds
1 ||
e(ts)d
=1
2
4sin2 12 (s t)(s t)2 (s)ds (t) as .
On the account of (2.19), is a c.f., and is continuous at
the origin. By the continuity theorem is a c.f. Again
(t) (t) as and since is continuous at the origin it follows that is a c.f.
as was to be proved.
Remark. This last result is essentially a theorem due to S. Bochner.
Remark on Theorem 2.7. If a sequence {Fn} of distributions con-verges properly to a distribution F, then the sequence {n} of theirc.f.s converges to , which is the c.f. of F and the convergence is
uniform in every finite interval.
Proof. Let A < 0, B > 0 be points of continuity of F. We have
n() () =
eixFn{dx}
eixF{dx}
=
xB
eixFn{dx} xB
eixF{dx}
+
BA
eixFn{dx} BA
eixF{dx}
= I1 + I2 + I3 (say).
We have
I3 =
BA
eixFn{dx} BA
eixF{dx}
= {eix[Fn(x) F(x)]}BA iBA
eix[Fn(x) F(x)]dx
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
35/90
24 Topics in Probability
and so
|I3
|=
|Fn(B)
F(B)
|+
|Fn(A)
F(A)
|+ ||
BA
|Fn(x) F(x)|dx.
Given > 0 we can make
|Fn(B) F(B)| < /9, |Fn(A) F(A)| < /9for n sufficiently large. Also, since |Fn(x) F(x)| 2 and Fn(x) F(x) at points of continuity of F, we have for || <
||BA
|Fn(x) F(x)|dx BA
|Fn(x) F(x)|dx < /9.
Thus
|I3| < /3.
Also for A, B sufficiently large
|I1| xB
eixFn{dx} 1 Fn(B) + Fn(A) < 13
|I2| xB
eixFn{dx} 1 Fn(B) Fn(A) < 13 .
The results follow from the last three inequalities.
2.5. Problems for Solution
1. Consider the family of distributions with densities fa(1 a 1) given by
fa(x) = f(x)[1 + a sin(2 log x)]
where f(x) is the log-normal density
f(x) =12
x1e1/2(log x)2
for x > 0.
= 0 for x 0.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
36/90
Characteristic Functions 25
Show that fa has exactly the same moments as f. (Thus
the log-normal distribution is not uniquely determined by its
moments).2. Let {pk, k 0} be a probability distribution, and {Fn, n 0} a
sequence of distributions. Show thatn=0
pnFn(x)
is also a distribution.
3. Show that () = e(e||1) is a c.f., and find the corresponding
density.
4. A distribution is concentrated on {2, 3, . . .} with weightspk =
c
k2 log |k|(k = 2, 3, . . .)
where c is such that the distribution is proper. Find its c.f. and
show that exists but the mean does not.
5. Show that the function () = e||
( > 2) is not a c.f.
6. If a c.f. is such that ()2 = (c) for some constant c, and
the variance is finite, show that is the c.f. of the normal distri-
bution.
7. A degenerate c.f. is factorized in the form = 12, where 1and 2 are c.f.s. Show that 1 and 2 are both degenerate.
8. If the sequence of c.f.s {n} converges to a c.f. and n 0,show that n(n)
n(0).
9. If{n} is a sequence of c.f.s such that n() 1 for < < ,then n() 1 for all .
10. A sequence of distributions {Fn} converges properly to a non-degenerate distribution F. Prove that the sequence {Fn(anx +bn)} converges to a distribution degenerate at the origin iffan and bn = 0(an).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
37/90
Chapter 3
Analytic CharacteristicFunctions
3.1. Definition and Properties
Let F be a probability distribution and consider the transform
() =
exdF(x) (3.1)
for = + i, where , are real and i =1. This certainly
exists for = i. SinceBA
exdF(x)
BA
exdF(x), (3.2)
() exists if e
xdF(x) is finite. Clearly, the integrals
0
exdF(x),0
exdF(x) (3.3)
converge for < 0, > 0 respectively. Suppose there exist numbers
, (0 < , ) such that the first integral in (3.3) convergesfor < and the second for > , then
exdF(x) < for < < . (3.4)
In this case () converges in the strip < < of the complexplane, and we say (in view of Theorem 3.1 below) that F has an
analytic c.f. . If = = the c.f. is said to be entire (analytic onthe whole complex plane).
27
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
38/90
28 Topics in Probability
The following examples show that a distribution need not have
an analytic c.f. and also that there are distributions with entire
c.f.s. The conditions under which an analytic c.f. exists are statedin Theorem 3.5.
Examples
Distribution c.f. Regions of existence
Binomial: f(n, k) =
n
k
pkqnk (q+ pe)n whole plane
Normal: f(x) = 12 e
1
2x2
e1
2
2
whole plane
Cauchy: f(x) =1
1
1 + x2e|| = 0
Gamma: f(x) = exx1
()
1
<
Laplace: f(x) =1
2e|x| (1 2)1 1 < < 1
Poisson: f(k) = ek
k!e(e
1) whole plane
Theorem 3.1. The c.f. is analytic in the interior of the strip of
its convergence.
Proof. Let
I =( + h)
()
h xe
xdF(x)
where the integral converges in the interior of the strip of conver-
gence, since for > 0,
xexdF(x)
|x|exdF(x)
e|x|+xdF(x)
and the last integral is finite for + < < . We have
I =
ex
ehx 1 hxh
dF(x)
=
ex(h(x2/2!) + h2x3/3! + )dF(x).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
39/90
Analytic Characteristic Functions 29
Therefore
|I|
ex
|h||x|2
(1 + |hx|/1! + |hx|2
/2! + )dF(x)
|h|
ex+|x|+|h||x|dF(x) <
in the interior of the strip of convergence. As |h| 0 the last expres-sion tends to zero, so
( + h)
()
h xe
xdF(x).
Thus () exists for in the interior of the strip, which means that() is analytic there.
Theorem 3.2. The c.f. is uniformly continuous along vertical
lines that belong to the strip of convergence.
Proof. We have
|( + i1) ( + i2)| =
ex(ei1x ei2x)dF(x)
ex|ei(12)x 1|dF(x)
= 2
ex
|sin(1
2)(x/2)
|dF(x).
Since the integrand is uniformly bounded by ex and approaches 0
as 1 2, uniformly continuity follows.
Theorem 3.3. An analytic c.f. is uniquely determined by its values
on the imaginary axis.
Proof. (i) is the c.f. discussed in Chapter 2 and the result fol-lows by the uniqueness theorem of that section.
Theorem 3.4. The function log () is convex in the interior of the
strip of convergence.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
40/90
30 Topics in Probability
Proof. We have
d2
d2 log () =
()()
()2
()2
and by the Schwarz inequality
()2 =
xexdF(x)
2=
e12x xe 12xdF(x)
2
exdF(x)
x2exdF(x) = ()().
Therefore d2d2
log () 0, which shows that log () is convex. Corollary 3.1. If F has an analytic c.f. and (0) = 0, then ()is minimal at = 0. If is an entire function, then () as , unless F is degenerate.
3.2. Moments
Recall that
n =
xndF(x), n
|x|ndF(x)
have been defined as the ordinary moment and absolute moment of
order n respectively. If F has an analytic c.f. , then n = (n)(0),
and
() =0
nn
n!,
the series being convergent in || < = min(, ). The converse isstated in the following theorem.
Theorem 3.5. If all moments of F exist and the series
nn
n! has
a nonzero radius of convergence , then exists in||
< , and
inside the circle || < ,
() =
0
nn
n!.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
41/90
Analytic Characteristic Functions 31
Proof. We first consider the series
nn
n! and show that it also
converges in || < . From Lyapunovs inequality
1nn
1n+1
n+1
we obtain
lim sup
1nn
n= lim sup
12n2n
2n= lim sup
12n2n
2n lim sup |n|
1n
n.
Also, since |n| n we have
lim sup |n|1n
n lim sup
1
nn
n.
Therefore
limsup|n| 1n
n= lim sup
1nn
n
which shows that the series
n
n
n! has radius of convergence . For
arbitrary A > 0 we have
>0
n||nn!
||n
n!
AA
|x|ndF(x) =AA
e|x|dF(x)
for || < . So
AA
exdF(x)
AA
e|x|dF(x) <
for || < . Since A is arbitrary, this implies that () converges inthe strip || < .
3.3. The Moment Problem
The family of distributions given by
F(x) = kx e|
y
|
{1 + sin(|y|
tan )}dyfor 1 1, 0 < < 1 has the same moments of all orders. Thisraises the question: under what conditions is a distribution uniquely
determined by its moments?
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
42/90
32 Topics in Probability
Theorem 3.6. If F has an analytic c.f. then it is uniquely deter-
mined by its moments.
Proof. If F has an analytic c.f., then the series
nnn converges
in || < = min(, ) and () is given by this series there. Ifthere is a second d.f. G with the same moments n, then by Theo-
rem 3.5, G has an analytic c.f. (), and () is also given by that
series in || < . Therefore () = () in the strip || < andhence F = G.
The cumulant generating function
The principal value of log () is called the cumulant generating func-
tion K(). It exists at least on the imaginary axis between = 0 and
the first zero of (i). The cumulant of order r is defined by
Kr = i1r
d
d
rlog (i)
=0
.
This exists if, and only if, r exists; Kr can be expressed in terms ofr. We have
K(i) =0
Kr(i)r
r!
whenever the series converges.
Theorem 3.7. Let () = 1()2(), where (), 1(), 2() arec.f.s. If () is analytic in < < , so are 1() and 2().Proof. We have (with the obvious notations)
exdF(x) =
exdF1(x)
exdF2(x),
and since () is convergent, so are 1() and 2().
Theorem 3.8 (Cramer). If X1 and X2 are independent r.v. such
that their sum X = X1 + X2 has a normal distribution, then X1,
X2 have normal distributions (including the degenerate case of the
normal with zero variance).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
43/90
Analytic Characteristic Functions 33
Proof. Assume without loss of generality that E(X1) = E(X2) = 0 .
Then E(X) = 0. Assume further that E(X2) = 1. Let 1(), 2()
be the c.f.s of X1 and X2. Then we have
1()2() = e122 . (3.5)
Since the right side of (3.5) is an entire function without zeros, so are
1() and 2(). By the convexity property (Theorem 3.4) we have
1() 1, 2() 1 as moves away from zero. Then (3.5) gives
e122 = 1()2()
1()
|1()
|. (3.6)
Similarly |2()| e 122 . Therefore
e122 |1()| |1()2()| = e 12Re(2) = e 12 (22),
so that
|()| e 122 . (3.7)From (3.6) and (3.7) we obtain
12||2 1
22 log |1()| 1
22 1
2||2,
or, setting K1() = log 1(),
|Re K1()| 12||2. (3.8)
From a strengthened version of Liouvilles theorem (see Lemma 3.1)
it follows that K1() = a1 + a22. Similarly K2() = b1 +
b22.
Theorem 3.9 (Raikov). If X1 and X2 are independent r.v. such
that their sum X = X1 + X2 has a Poisson distribution, then X1,
X2 have also Poisson distributions.
Proof. The points of increase of X are k = 0, 1, 2, . . . , so all points
of increase 1 and 2 of X1 and X2 are such that 1 + 2 = some k,
and moreover the first points of increase of X1 and X2 are and where is some finite number. Without loss of generality we take
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
44/90
34 Topics in Probability
= = 0, so that X1 and X2 have k = 0, 1, 2, . . . as the onlypossible points of increase. Their c.f.s are then of the form
1() =0
akek , 2() =
0
bkek (3.9)
with a0, b0 > 0, ak, bk 0 (k 1) and
ak =
bk = 1. Let z = e
and 1() = f1(z), 2() = f2(z). We have
f1(z)f2(z) = e(z1). (3.10)
Therefore
a0bk + a1bk1 + + akb0 = ek
k!(k = 0, 1, . . .), (3.11)
which gives
ak 1b0
ek
k!, |f1(z)| 1
b0e(|z|1). (3.12)
Similarly |f2(z)| 1a0 e(|z|1). Hence1
a0e(|z|1)|f1(z)| |f1(z)f2(z)| = e(u1)
where u = Re (z). This gives
|f1(z)| a0e(|z|u) a0e2|z|. (3.13)From (3.12) and (3.13), noting that a0b0 = e
we find that
2
|z
| log
|f1(z)
| log a0
2
|z
|,
or setting K1(z) = log f1(z), and log a0 = 1 < 0,|Re K1(z) + 1| 2|z|. (3.14)
Proceeding as in the proof of Theorem 3.8, we obtain 1+K1(z) = cz,
where c is a constant. Since f1 = 1, K1(1) = 0, so c = 1and f1(z) = e
1(z1), which is the transform of the Poissondistribution.
Theorem 3.10 (Marcinkiewicz). Suppose a distribution has a
c.f. () such that (i) = eP(i), where P is a polynomial. Then
(i) () = eP() in the whole plane, and (ii) is the c.f. of a normal
distribution (so that P() = + 2 with 0).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
45/90
Analytic Characteristic Functions 35
Proof. Part (i) is obvious. For Part (ii) let
P() =
n1
akk, n finite, ak real (cumulants).
From |()| () we obtain |eP()| eP() or eReP() eP().Therefore Re P() P(). Put = rei, so that = r cos , =r sin . Then
anrn cos n + an1rn1 cos(n 1) + anrn cosn + an1rn1 cos(n1) +
Suppose an = 0. Dividing both sides of this inequality by rn andletting r we obtain an cos n an cosn . Putting = 2n weobtain
an 0 an cosn 2n
,
so an 0 for n 2. Similarly, putting = 2n we find that
an an cosn 2n
,
and since cosn 2n < 1 for n > 2 we obtain an 0. Thereforean = 0 for n > 2, P() = a1 + a2
2, and () is the c.f. of a nor-
mal distribution, the case a2 = 0 being the degenerate case of zero
variance.
Theorem 3.11 (Bernstein). Let X1 and X2 be independent r.v.
with unit variances. Then if
Y1 = X1 + X2, Y2 = X1 X2 (3.15)are independent, all four r.v. X1, X2, Y1, Y2 are normal.
This is a special case of the next theorem (with n = 2, a1 =
b1 = a2 = 1, b2 = 1). For a more general result see [Feller (1971),pp. 7780, 525526]. He considers the linear transformation Y1 =
a11X1 + a12X2, Y2 = a21X1 + a22X2 with || = 0, where is the
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
46/90
36 Topics in Probability
determinant
=a11 a12
a21 a22
.
If a11a21 + a12a22 = 0 then the transformation represents a rotation.
Thus (3.15) is a rotation.
Theorem 3.12 (Skitovic). Let X1, X2, . . . , X n be n independent
r.v. such that the linear forms
L1 = a1X1 + a2X2 + + anXn,L2 = b1X1 + b2X2 + + bnXn, (ai = 0, bi = 0),
are independent. Then all the (n + 2) r.v. are normal.
Proof. We shall first assume that (i) the ratios ai/bi are all distinct,
and (ii) all moments of X1, X2, . . . , X n exist. Then for , real we
have (with obvious notations)
()L1+L2 = ()L1
()L2
so thatni=1
()(ai+bi)Xi
=
ni=1
()aiXi
ni=1
()biXi
.
Taking logarithms of both sides and expanding in powers of we
obtainni=1
K(ai+bi)Xir =
ni=1
KaiXir +
ni=1
KbiXir
orni=1
K(Xi)r {(ai + bi)r (ai)r (bi)r} = 0
for all r 1. This can be written asni=1
K(Xi)r
r1s=1
r
s
(ai)
s(bi)rs = 0
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
47/90
Analytic Characteristic Functions 37
for all r 1 and all , . Henceni=1
asi brsi K(Xi)r = 0 (s = 1, 2, . . . , r 1, r 1),
Let r n + 1. Then for s = 1, 2, . . . , n, i = 1, 2, . . . n we can writethe above equations as
Arr = 0 (3.16)
where Ar
= aSi
brsi 1 s, i n and r is the column vector withelements K(X1)r , K(X2)r , . . . , K (Xn)r . Since
|Ar| = (a1a2 an)(b1b2 bn)r1j>i
(cj ci) = 0,
the only solution of (3.16) is r = 0. Therefore
K(Xi)r = 0 for r
n + 1, i = 1, 2, . . . , n . (3.17)
Thus all cumulants of Xi of order n + 1 vanish, and K(Xi)()reduces to a polynomial of degree at most n. By the theorem of
Marcinkiewicz, each Xi has a normal distribution. Hence L1 and L2have normal distributions.
Next suppose that some of the ai/bi are the same. For example,
let a1/b1 = a2/b2, and let Y1 = a1X1 + a2X2. Then
L1 = Y1 + a3X3 + + anXn,L2 =
b1a1
Y1 + b3X3 + + bnXn.
Repeat this process till all the ai/bi are distinct. Then by what has
just proved, the Yi are normal. By Cramers theorem the Xi are
normal.
Finally it remains to prove that the moments of Xi exist. Thisfollows from the fact that L1 and L2 have finite moments of all orders.
To prove this, we note that since ai = 0, bi = 0 we can take a, c > 0such that |ai|, |bi| c > 0. Also, let us standardize the ai and bi sothat |ai| 1, |bi| 1. Now if|L| = |a1X1+a2X2 + + anXn| nM,
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
48/90
38 Topics in Probability
then at least one |Xi| M. Therefore
P{|L1| nM} ni=1
P{|Xi| M}. (3.18)
Further, if c|Xi| nM and |Xj | < M for all j = i, then |L1| M,|L2| M. Thus
P{|L1| M, |L2| M} P
|Xi| nMc
j=1
P{|Xj | < M}
P
|Xi| nMc
nj=1
P{|Xj | < M}.
Summing this over i = 1, 2, . . . , n we obtain, using (3.18),
n P{|L1| M, |L2| M} P
|L1| n2M
c
nj=1
P{|Xj | < M}.
Since L1 and L2 are independent, this gives
P|L1| n2Mc
P{|L1| M} n
P{|L2| M}n1 P{|Xj | < M}
0 (3.19)
as M . We can write (3.19) as follows. Choose n2/c = > 1.Then
P
{|L1
| M
}P{|L1| M} 0 as M . (3.20)By a known result (Lemma 3.2), L1, and similarly L2, has finite
moments of all orders.
Lemma 3.1 (see [Hille (1962)]). Iff() is an entire function and
|Ref()| c||2, then f() = a1 + a22.
Proof. We have f() =0 ann, the series being convergent on
the whole plane. Here
an =n!
2i
||r
f()
n+1d (n = 0, 1, 2, . . .). (3.21)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
49/90
Analytic Characteristic Functions 39
Also, since there are no negative powers
0 =
n!
2i||r f()
n
1
d (n = 1, 2, . . .). (3.22)
From (3.21) we obtain
an =n!
2i
20
f(rei)
rn+1ei(n+1)reiid
or
anrn
=
n!
220 f(re
i
)ein
d (n = 0, 1, . . .). (3.23)
Similarly from (3.22) we obtain
0 =n!
2
20
f(rei)eind
or
0 =
n!
220 f(re
i
)ein
d (n = 1, 2, . . .). (3.24)
From (3.23) and (3.24) we obtain
anrn =
n!
20
Ref(rei)eind (n 1).
Therefore
|an|rn n!20
ck2d = 2cn!r2
or
|an| 2cn!rn2
0 as r for n > 2.
This gives f() = a0 + a1 + a22.
Lemma 3.2 (see [Loeve (1963)]). For > 1 if1 F(x) + F(x)
1 F(x) + F(x) 0 as x
then F has moments of all orders.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
50/90
40 Topics in Probability
Proof. Given > 0 choose A so large that for x > A
1
F(x) + F(
x)
1 F(x) + F(x) < and 1 F(A) + F(A) < .Then for any positive integer r,
1 F(rA) + F(rA)1 F(A) + F(A) =
rs=1
1 F(sA) + F(sA)1 F(s1A) + F(s1A) <
r
so that
1 F(rA) + F(rA) < r+1.Therefore
1 F(x) + F(x) < r+1
for x > rA. Now
A nx
n
1
[1 F(x) + F(x)]dx=
r=0
r+1ArA
nxn1[1 F(x) + F()]dx
0, showthat F is uniquely determined by its moments.
2. Show that the distribution whose density is given by
f(x) =
12
e|x| for x > 0
0 for x 0does not have an analytic c.f.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
51/90
Analytic Characteristic Functions 41
3. Proof of Bernsteins theorem. Introduce a change of scale so
that Y1 =12
(X1 + X2), Y2 =12
(X1 X2). Then prove that
K(Y1)s =
1
2
s
1sK(X1)s + 1sK(X2)s
,
K(Y2)s =
1
2
s 1sK(X1)s + (1)sK(X2)s
,
and similarly for K(X1)s , K
(X2)s in terms of K
(Y1)s , K
(Y2)s . Hence
show thatK(Xi)s 12s
2K(X1)s + 2K(X2)s (i = 1, 2).
This gives K(Xi)s = 0 for s > 2, i = 1, 2.
4. If X1, X2 are independent and there exists one rotation (X1,
X2) (Y1, Y2) such that Y1, Y2 are also independent, then showthat Y1, Y2 are independent for every rotation.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
52/90
Chapter 4
Infinitely DivisibleDistributions
4.1. Elementary Properties
A distribution and its c.f. are called infinitely divisible if for each
positive integer n there exists a c.f. n such that
() = n()n
. (4.1)
It is proved below (Corollary 4.1) that if is infinitely divisible, then
() = 0. Defining 1/n as the principal branch of the n-th root,we see that the above definition implies that 1/n is a c.f. for every
n 1.Examples
(1) A distribution concentrated at a single point is infinitely divisi-
ble, since for it we have
() = eia = (eia/n)n
where a is a real constant.
(2) The Cauchy density f(x) = a [a2+(x)2]1 (a > 0) has () =
eia||. The relation (4.1) holds with n() = ei/na||/n.Therefore the Cauchy density is infinitely divisible.
(3) The normal density with mean m and variance 2 has c.f. () =
eim1
222 = (eim/n
1
2
2
n 2
)n. Thus the normal distribution is
infinitely divisible.
43
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
53/90
44 Topics in Probability
(4) The gamma distribution (including the exponential) is infinitely
divisible, since its c.f. is
() = (1 i/) =
(1 i/)/nn .
The discrete counterparts, the negative binomial and geometric
distributions are also infinitely divisible.
(5) Let N be a random variable with the (simple) Poisson distribu-
tion ek/k!(k = 0, 1, 2, . . .). Its c.f. is given by
() = e(ei1),
which is clearly infinitely divisible. Now let {Xk} be a sequenceof independent random variables with a common c.f. and let
these be independent ofN. Then the sum X1+ X2 + + XNbhas the c.f.
() = eib+[()1],
which is the compound Poisson. Clearly, this is also infinitely
divisible.
Lemma 4.1. Let {n} be a sequence of c.f.s. Then nn contin-uous iff n(n 1) with continuous. In this case = e.Theorem 4.1. A c.f. is infinitely divisible iff there exists a
sequence {n} of c.f.s such that nn .
Proof. If is infinitely divisible, then by definition there exists ac.f. n such that
nn = (n 1). Therefore the condition is necessary.
Conversely, let nn . Then by Lemma 4.1, n[n() 1] =log . Now for t > 0,
ent[n()1] et() as n .Here the expression on the left side is the c.f. of the compound Poisson
distribution and the right side is a continuous function. Therefore foreach t > 0, et is a c.f. and
= e = (e/n)n,
which shows that is infinitely divisible.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
54/90
Infinitely Divisible Distributions 45
Corollary 4.1. If is infinitely divisible, = 0.This was proved in the course of the proof of Theorem 4.1.
Corollary 4.2. If is infinitely divisible, so is ()a for each a > 0.
Proof. We have a = ea = (aa/n)n.
Proof of Lemma 4.1. (i) Suppose n(n 1) which is contin-uous. Then n 1 and the convergence is uniform in [, ].Therefore |1 n()| < 12 for [, ] and n > N. Thus log nexists for [, ], and n > N, and is continuous and bounded.Now
log n = log[1 + (n 1)]
= (n 1) 12
(n 1)2 + 13
(n 1)3 = (n 1)[1 + o(1)]
and therefore
n log n = n(n 1)[1 + o(1)] or nn = e.
(ii) Suppose nn . We shall first prove that has no zeros. Itsuffices to prove that |n|2n ||2 implies ||2 > 0. Assume that thissymmetrization has been carried out, so that nn with n 0, 0. Since is continuous with (0) = 1, there exists an interval[, ] in which does not vanish and therefore log exists and is
bounded. Therefore log n exists and is bounded for [, ] andn > N, so n log n log . Thus log n 0 or n 1. As in (i),n(n 1) log = .Theorem 4.2. If {n} is a sequence of infinitely divisible c.f.s andn which is continuous, then is an infinitely divisible c.f.
Proof. Since n is infinitely divisible, 1/nn is a c.f. Since
1/nnn continuous,
is an infinitely divisible c.f. by Theorem 4.1.
Theorem 4.3 (De Finetti). A distribution is infinitely divisible iff
it is the limit of compound Poisson distributions.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
55/90
46 Topics in Probability
Proof. If n is the c.f. of a compound Poisson distribution, and
n which is continuous, then by Theorem 4.2, is an infinitelydivisible c.f. Conversely, let be an infinitely divisible c.f. Then byTheorem 4.1 there exists a sequence {n} of c.f.s such that nn .By Lemma 4.1
en[n()1] e = .Here en[n()1] is the c.f. of a compound Poisson distribution.
4.2. Feller Measures
A measure M is said to be a Feller measure if M{I} < for everyfinite interval I, and the integrals
M+(x) =
x
1
y2M{dy}, M(x) =
x+
1
y2M{dy} (4.2)
converge for all x > 0.
Examples
(1) A finite measure M is a Feller measure, since|y|>x
1
y2M{dy} 1
x2[M{(, x)} + M{(x, )}].
(2) The Lebesgue measure is a Feller measure, since|y|>x
1
y2dy =
2
x(x > 0).
(3) Let F be a distribution measure and M{dx} = x2F{dx}. ThenM is a Feller measure with
M+(x) = 1 F(x), M(x) = F(x+).
Theorem 4.4. Let M be a Feller measure, b a real constant and
() = ib +
eix 1 i sin xx2
M{dx} (4.3)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
56/90
Infinitely Divisible Distributions 47
(the integral being convergent). Then corresponding to a given there
is only one measure M and one constant b.
Proof. Consider
() = () 12h
hh
( + s)ds (h > 0). (4.4)
We have
() =
eix{dx} (4.5)
where
{dx} =
1 sin hxhx
1
x2M{dx} (4.6)
and it is easily verified that is a finite measure. Therefore ()determines uniquely, so M uniquely. Since b = Im (1), the con-
stant b is uniquely determined.
Convergence of Feller measures. Let {Mn} be a sequence ofFeller measures. We say that Mn converges properly to a Feller mea-
sure M if Mn{I} M{I} for all finite intervals I of continuity ofM, and
M+n (x) M+(x), Mn (x) M(x) (4.7)
at all points x of continuity of M. In this case we write Mn M.Examples
(1) Let Mn{dx} = nx2Fn{dx} where Fn is a distribution measurewith weights 12 at each of the points 1n . Then
Mn{I} = Inx2Fn{dx} = n
1
n 1
2+
1
n 1
2 = 1if { 1
n, 1
n} I. Also M+n (x) = Mn (x) = 0 for x > 1n .
Therefore Mn M where M is a distribution measure concen-trated at the origin. Clearly, M is a Feller measure.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
57/90
48 Topics in Probability
(2) Let Fn be a distribution measure with Cauchy density1 n1+n2x2
and consider Mn
{dx
}= nx2Fn
{dx
}. We have
Mn{(a, b)} =ba
n2x2
1 + n2x2dx |b a|,
M+n (x) =
x
n2
1 + n2y2dy
x
dy
y2,
Mn (x) = x
n2
1 + n2y2dy
x
dy
y2.
Therefore Mn M where M is the Lebesgue measure.
Theorem 4.5. Let {Mn} be a sequence of Feller measures, {bn} asequence of real constants and
n() = ibn +
eix 1 i sin xx2
Mn{dx}. (4.8)
Then n continuous iff there exists a Feller measure M and areal constant b such that Mn M and bn b. In this case
() = ib +
eix 1 i sin xx2
M{dx}. (4.9)
Proof. As suggested by (4.4)(4.6) let
n{dx} = K(x)Mn{dx}, where K(x) = x21 sin hxhx
(4.10)
n = n{(, )} < . (4.11)Then
Mn{dx} =1
nn{dx} (4.12)
is a distribution measure. We can write
n() = ibn + n
eix 1 i sin xx2
K(x)1M n {dx}.(4.13)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
58/90
Infinitely Divisible Distributions 49
(i) Let Mn M and bn b. Then
n =
K(x)M{dx} > 0and
M n M, where M{dx} =1
K(x)M{dx}.
Therefore from (4.13) we find that
n()
ib +
eix 1 i sin x
x2
K(x)1M
{dx
}= ().
(ii) Conversely, let n() () continuous. Then with n(),with () defined as in (4.4), n
() (); that is,
eixn{dx} (). (4.14)
In particularn = n{(, )} (0).
If(0) = 0, then n{I} and Mn{I} tend to 0 for every finite intervalI and by (i) () = ib with b = lim bn. We have thus proved the
required results in this case. Let = (0) > 0. Then (4.14) can bewritten as
n e
ix
Mn{dx} ().Therefore M n M where M is the distribution measure corre-sponding to the c.f. ()/(0). Thus
n
eix 1 i sin xx2
K(x)1M n {dx}
eix 1 i sin xx2
K(x)1M
{dx
},
(the integrand being a bounded continuous function), and bn b.Clearly,
M{dx} = K(x)1M{dx}
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
59/90
50 Topics in Probability
is a Feller measure and
() = ib +
eix
1
i sin x
x2 M{dx}as required.
4.3. Characterization of Infinitely Divisible
Distributions
Theorem 4.6. A distribution is infinitely divisible iff its c.f. is ofthe form = e, with
() = ib +
eix 1 i sin xx2
M{dx}, (4.15)
M being a Feller measure, and b a real constant.
Proof. (i) Let = e with given by (4.15). We can write
() = ib 12
2M{0} + lim0+
() (4.16)
where
() =
|x
|>
eix 1 i sin xx2
M{dx}
= i + c
|x|>
(eix 1)G{dx}
with
cx2G{dx} = M{dx} for |x| > , and
= |x|>sin x
M{dx}x2
,
c being determined so that G is a distribution measure. Let
denote the c.f. of G; then
e() = ei+c[()1]
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
60/90
Infinitely Divisible Distributions 51
is the c.f. of a compound Poisson distribution. As 0, 0, where
0() =
|x|>0
eix 1 i sin xx2
M{dx}
is clearly a continuous function. By Theorem 4.3, e0 is an
infinitely divisible c.f. Now we can write
e() = eib1
22M{0} e0(),
so that is the product of e0() and the c.f. of a normal distri-
bution. Therefore is infinitely divisible.
(ii) Conversely, let be an infinitely divisible c.f. Then by Theo-
rem 4.3. is the limit of a sequence of compound Poisson c.f.s.
That is,
ecn[n()1in] ()or
cn
(eix 1 iwn)Fn{dx} log ()
where cn > 0, n is real and Fn is the distribution measure cor-
responding to the c.f. n. We can write this as
eix 1 i sin xx2
Mn{dx}
+ icn
sin xFn{dx} n
log ()where Mn{dx} = cnx2Fn{dx}. Clearly, Mn is a Feller measure.By Theorem 4.5 it follows that
Mn M and cn
sin xFn{dx} n
b
where M is a Feller measure, b a real constant and
log () = ib +
eix 1 i sin xx2
M{dx}.
This proves that = e, with given by (4.15).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
61/90
52 Topics in Probability
Remarks.
(a) The centering function sin x is such that
eix 1 i sin xx2
M{dx}
is real. Other possible centering functions are
(i) (x) =x
1 + x2
and
(ii) (x) =
a for x < a,|x| for a x a,a for x > a with a > 0.
.
(b) The measure (Levy measure) is defined as follows: {0} = 0and {dx} = x2M{dx} for x = 0. We have
min(1, x
2
){dx} < ,as can be easily verified. The measure K{dx} = (1+x2)1M{dx}is seen to be a finite measure. This was used by Khintchine.
(c) The spectral function H is defined as follows:
H(x) =
x
M{dy}y2
for x > 0
x+
M{dy}
y2for x < 0,
H being undefined at x = 0. We can then write
() = ib 12
w22 +
0+
[eix 1 i(x)]dH(x)
+ 0 [eix 1 i(x)]dH(x),where the centering function is usually (x) = x(1 + x2)1. Thisis the so-called LevyKhintchine representation. Here H is non-
decreasing in (, 0) and (0, ), with H() = 0, H() = 0.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
62/90
Infinitely Divisible Distributions 53
Also, for each > 0
|x|
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
63/90
54 Topics in Probability
or
P(s) = (q0n + q1ns + q2ns2 +
)n.
In particular qn0n = P(0) = p0. If p0 = 0, then q0n = 0 and P(s) =
sn(q1n + q2ns + q3ns2 + )n. This implies that p0 = p1 = p2 = =
pn1 = 0 for each n 1, which is absurd. Therefore p0 > 0. It followsthat P(s) > 0 and therefore P(s)1/n 1 for 0 s 1. Now
log P(s) log P(0)
log P(0)=
log n
P(s)P(0)
log n1
P(0)
n
P(s)P(0) 1
n1
P(0) 1
=n
P(s) nP(0)1 nP(0) = Qn(s) Qn(0)1 Qn(0) .
Thus
Qn(s) Qn(0)1
Qn(0)
log P(s) log P(0)
log P(0)
.
Here the left side is seen to be a p.g.f. By the continuity theorem
the limit is the generating function of a non-negative sequence {fj}.Thus
log P(s) log P(0) log P(0) =
1
fjsf = F(s) (say).
Putting s = 1 we find that F(1) = 1. Putting =
log P(0) > 0 we
obtain
P(s) = e[1F(s)]
which is equivalent to (4.18).
4.4. Special Cases of Infinitely Divisible
Distributions
(A) Let the measure M be concentrated at the origin, with weight
2 > 0. Then (4.15) gives () = eib1
222 , which is the c.f. of
the normal distribution.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
64/90
Infinitely Divisible Distributions 55
(B) Let M be concentrated at h(= 0) with weight h2. Then
() = e
ir+(eih
1)
, r = b sin h.Thus is the c.f. of the random variable hN + r, where N has
the (simple) Poisson distribution ek/k! (k = 0, 1, 2, . . .).(C) Let M{dx} = x2G{dx} where G is the distribution measure
with the c.f. . Clearly, M is a Feller measure and
() = ei+[()1], = b
sin x G{dx}.
We thus obtain the c.f. of a compound Poisson distribution.
(D) Let M be concentrated on (0, ) with density exx(x > 0).It is easily verified that M is a Feller measure. We have
0
eix 1x2
M{dx} = 0
e(i)x exx
{dx}
= log
i = log
1 i
.
Choosing
b =
0
sin x
xexdx <
we find that
() =
1 i
,This is the c.f. of the gamma density exx1/().
(E) Stable distributions. These are characterized by the measure M,
where
M{(y, x)} = C(px2 + qy2) (x > 0, y > 0)where C > 0, p 0, q 0, p + q = 1, 0 < 2. If = 2,M is concentrated at the origin, and the distribution is the
normal, as discussed in (A). Let 0 < < 2, and denote by
the corresponding expression . In evaluating it we choose
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
65/90
56 Topics in Probability
an appropriate centering function (x) depending on . This
changes the constant b and we obtain
() = i +
eix 1 i(x)x2
M{dx}
where
= b +
(x) sin xx2
M{dx} (|r| < )
and
(x) =
sin x if = 1
0 if 0 < < 1
x if 1 < < 2.
Substituting for M we find that
() = i + c(2 )[pI() + qI()]
where
I() =
0
eix 1 i(x)x+1
dx.
Evaluating the integral I we find that
() = i c|w|
1 + i ||(||, )
where c > 0, || 1 and
(||, ) =
tan
2if = 1
2
log
|w
|if = 1.
In Sec. 4.6 we shall discuss the detailed properties of stable dis-
tributions. We note that when = 0 and = 1 we obtain
() = i c||, so that is the c.f. of the Cauchy distribution.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
66/90
Infinitely Divisible Distributions 57
4.5. Levy Processes
We say a stochastic process {X(t), t 0} has stationary independentincrements if it satisfies the following properties:
(i) For 0 t1 < t2 < < tn(n 2) the random variablesX(t1), X(t2) X(t1), X(t3) X(t2), . . . , X (tn) X(tn1)
are independent.
(ii) The distribution of the increment X(tp)X(tp1) depends onlyon the difference tp tp1.For such a process we can take X(0) 0 without loss ofgenerality. For ifX(0) 0, then the process Y(t) = X(t)X(0)has stationary independent increments, and Y(0) = 0.
If we write
X(t) =
nk=1
X
k
nt
X
k 1
nt
(4.19)
then X(t) is seen to be the sum of n independent random vari-ables all of which are distributed as X(t/n). Thus a process with
stationary independent increments is the generalization to con-
tinuous time of sums of independent and identically distributed
random variables.
A Levy process is a process with stationary independent incre-
ments that satisfies the following additional conditions:
(iii) X(t) is continuous in probability. That is, for each > 0P{|X(t)| > } 0 as t 0. (4.20)
(iv) There exist left and right limits X(t) and X(t+) and we assumethat X(t) is right-continuous: that is, X(t+) = X(t).
Theorem 4.9. The c.f. of a Levy process is given by E[eiX(t)] =
et(), where is given by Theorem 4.6.
Proof. Let 1() = E[eiX(t)]. From (4.19) we find that t() =
[t/n()]n, so for each t > 0, t is infinitely divisible and t = e
t .
Also from the relation X(t+s) d= X(t)+X(s) we obtain the functional
equation t+s = t + s. On account of (4.20), t 0 as t 0, so
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
67/90
58 Topics in Probability
we must have t() = t1(). Thus t() = et() with = 1 in
the required form.
Special cases: Each of the special cases of infinitely divisible dis-
tributions discussed in Sec. 4.4 leads to a Levy process with c.f.
t() = et() and in the prescribed form. Thus for appropriate
choices of the measure M we obtain the Brownian motion, simple
and compound Poisson processes, gamma process and stable pro-
cesses (including the Cauchy process).
A Levy process with non-decreasing sample functions is called a
subordinator. Thus the simple Poisson process and gamma processare subordinators.
4.6. Stable Distributions
A distribution and its c.f. are called stable if for every positive integer
n there exist real numbers cn > 0, dn such that
()n = (cn)eidn . (4.21)
If X, X1, X2, . . . are independent random variables with the c.f. ,
then the above definition is equivalent to
X1 + X2 + + Xn d= cnX + dn. (4.22)Examples
(A) IfX has a distribution concentrated at a single point, then (4.22)
is satisfied with cn = n, dn = 0. Thus a degenerate distribution
is (trivially) stable. We shall ignore this from our consideration.
(B) If X has the Cauchy density f(x) = a [a2 + (x r)2]1 (a > 0),
then () = eira||. The relation (4.21) holds with cn = n,dn = 0. Thus the Cauchy distribution is stable.
(C) If X has a normal density with mean m and variance 2, then
(22) holds with cn = n and dn = m(n cn). Thus the normaldistribution is stable.
The concept of stable distributions is due to Levy (1924), who
gave a second definition (see Problem 11).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
68/90
Infinitely Divisible Distributions 59
Theorem 4.10. Stable distributions are infinitely divisible.
Proof. The relation (4.21) can be written as
() =
cn
ei
dnncn
n= n()
n
where n is clearly a c.f. By definition is infinitely divisible.
Domains of attraction. Let {Xk, k 1} be a sequence of inde-pendent random variables with a common distribution F, and Sn =
X1 + X2 + + Xn (n 1). We say that F belongs to the domainof attraction of a distribution G if there exist real constants an > 0,
bn such that the normed sum (Sn bn)/an converges in distributionto G.
It is clear that a stable distribution G belongs to its own domain
of attraction, with an = cn, bn = dn. Conversely, we shall prove
below that the only non-empty domains of attraction are those of
stable distributions.
Theorem 4.11. If the normed sum (Sn bn)/an converges in dis-tribution to a limit, then
(i) as n , an , an+1/an 1 and (bn+1 bn)/an b with|b| < , and
(ii) the limit distribution is stable.
Proof. (i) With the obvious notation we are given that
[(/an)eibn/nan ]n () (4.23)
uniformly in [, ]. By Lemma 4.1 we conclude that
n[(/an)eibn/nan 1] ()
where = e. Therefore
n() = (/an)eibn/nan 1.
Let {ank} be a subsequence of{an} such that ank a (0 a ).
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
69/90
60 Topics in Probability
If 0 < a < , then
1 = lim|(/a
nk)|
=|(/a)
|,
while if a = 0, then
1 = lim|nkank| = |()|.Both implications here would mean that is degenerate, which is
not true. Hence a = and an . From (4.23) we have
an+1
n+1eibn+1/an+1 (),
which can be written as
an+1
neibn+1/an+1 (), (4.24)
since (/an+1)
1. By Theorem 2.10 it follows from (4.23) and
(4.24) that an+1/an 1 and (bn+1 bn)/an b.(ii) For fixed m 1 we have
an
mneimbn/an =
an
neibn/an
m m().
Again by Theorem 2.10 it follows that amn/an cm, (bmn mbn)/an
dm, where cm > 0 and dm is real, while
() = m
cm
eidm/cm
or
m() = (cm)eidm .
This shows that is stable.
Theorem 4.12. A c.f. is stable iff = e, with
() = i c||
1 + i
||(||, )
(4.25)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
70/90
Infinitely Divisible Distributions 61
where is real, c > 0, 0 < 2, || 1 and
(||, ) = tan
2 if = 12
log|| if = 1.
(4.26)
Here is called the characteristic exponent of .
Proof. (i) Suppose is given by (4.25) and (4.26). Then for a > 0
we have
a() (a1/) = i(a a1/)ac||i || [(||, ) (a
1/||, )]
=
i(a a1/) if = 1
i
2c
a log a if = 1.
This shows that is stable.(ii) Conversely, let be stable. Then by Theorem 4.11 it possesses
a domain of attraction; that is, there exists a c.f. and real constants
an > 0, bn such that as n [(/an)e
ibn ]n ().Therefore by Lemma 4.1,
n[(/an)eibn 1] ()where = e. Let F be the distribution corresponding to . We first
consider the case where F is symmetric; then bn = 0. Let Mn{dx} =nx2F{andx}. Then by Theorem 4.5 it follows that there exists aFeller measure M and a constant b such that
() = ib +
eix 1 i sin xx2
M
{dx
}. (4.27)
Let
U(x) =
xx
y2F{dy} (x > 0). (4.28)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
71/90
62 Topics in Probability
Then
Mn{(x, x)} =n
a2n U(anx) M{(x, x)} (4.29a)
n[1 F(anx)] =x
y2Mn{dy} M+(x) (4.29b)
nF(anx) =x+
y2Mn{dy} M(x). (4.29c)
By Theorem 4.11 we know that an
, an+1/an
1. Therefore
U(x) varies regularly at infinity and M{(x, x)} = Cx2 whereC > 0, 0 < 2. If = 2 the measure M is concentrated at theorigin. If 0 < < 2 the measure M is absolutely continuous.
In the case where F is unsymmetric we have
n[1 F(anx + anbn)] M+(x), nF(anx + anbn) M(x)
and an analogous modification of (4.29a). However it is easily seenthat bn 0, and so these results are fully equivalent to (4.29).Considering (4.29b) we see that either M+(x) 0 o r 1 F(x)varies regularly at infinity and M+(x) = Ax. Similarly F(x) and1 F(x) + F(x) vary regularly at infinity and the exponent isthe same for both M+ and M. Clearly 0 < 2.
If M+ and M vanish identically, then clearly M is concentrated
at the origin. Conversely of M has an atom at the origin, then asymmetrization argument shows that M is concentrated at the ori-
gin, and M+, M vanish identically. Accordingly, when < 2 themeasure M is uniquely determined by its density, which is propor-
tional to |x|1. For each interval (y, x) containing the origin wetherefore obtain
M{
(
y, x)}
= C(px2
+ qy2
) (4.30)
where p + q = 1. For = 2, M is concentrated at the origin. For
0 < < 2 we have already shown in Sec. 4.4 that the measure (4.30)
yields the required expression (4.25) for .
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
72/90
Infinitely Divisible Distributions 63
Corollary 4.3. IfG is the stable distribution with the characteristic
exponent , then as x
x[1 G(x)] Cp 2
, xG(x) Cq2
. (4.31)
Proof. Clearly, G belongs to its own domain of attraction with
the norming constants an = n1/. For 0 < < 2, choosing n1/x = t
in (4.29b) we find that t[1 G(t)] Cp2 as t . For = 2,G is the normal distribution and for it we have a stronger result,
namely, x[1
G(x)]
0 as x
.
Theorem 4.13. (i) All stable distributions are absolutely continu-
ous.
(ii) Let 0 < < 2. Then moments of order < exist, while
moments of order > do not.
Proof. (i) We have |()| = ec||, with c > 0. Since the functionis integrable over (
,
), the result (i) follows by Theorem 2.6(b).
(ii) For t > 0 an integration by parts givestt
|x|F{dx} = t[1 F(t) + F(t)]
+
t0
x1[1 F(x) + F(x)]dx
t
0
x1[1
F(x) + F(
x)]dx.
If < , this last integral converges as t . Since by Corollary 4.3we have x[1 F(x) + F(x)] M for x > t where t is large. It fol-lows that the absolute moment (and therefore the ordinary moment)
of order < is finite. Conversely if the absolute moment of order
> exists, then for > 0 we have
> |x|>t |x|F
{dx
}> t[1
F(t) + F(
t)]
or t[1 F(t) + F(t)] < t 0 as t , which is a contra-diction. Therefore absolute moments of order > do not exist.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
73/90
64 Topics in Probability
Remarks
(1) From the proof of Theorem 4.12 it is clear that()a = (ca)e
ida
for all a > 0, and the functions ca and da are given by
(i) ca = a1/ with 0 < 2, and
(ii) da =
(a a1/) if = 1(2c/)a log a if = 1.
(2) If in the definition (4.21), dn = 0, then the distribution is calledstrictly stable. However, the distinction between strict and weak
stability matters only when = 1, because when = 1 we cantake dn = 0 without loss of generality. To prove this we note that
dn = (n n1/) for = 1, and consider the c.f.() = ()ei.
We have
()n = ()nein = (cn)ei(dnn)
= (cn)ei(cn+dnn) = (cn)
which shows that is strictly stable.
(3) Let = 1 and assume that = 0. Then we can write
() = a|| for > 0, and a|| for < 0 (4.32)where a is a complex constant. Choosing a scale so that |a| = 1we can write a = ei
2, where tan 2 = tan
2 . Since || 1 it
follows that
|| if 0 < < 1, and || 2 if 1 < < 2.(4.33)
Theorem 4.14. Let = 1 and let the c.f. of a stable distribution beexpressed in the form
() = e||ei/2 (4.34)
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
74/90
Infinitely Divisible Distributions 65
where in 1 the upper sign prevails for > 0 and the lower sign for < 0. Let the corresponding density be denoted by f(x; , ). Then
f(x; , ) = f(x; , ) for x > 0. (4.35)For x > 0 and 0 < < 1,
f(x; , ) =1
x
k=1
(k + 1)
k!(x)k sin k
2( ) (4.36)
and for x > 0 and 1 < < 2
f(x; , ) =1
x
k=1
(k1 + 1)k!
(x)k sin k2
( ). (4.37)
Corollary 4.4. A stable distribution is concentrated on (0, ) if0 < < 1, = and on (, 0) if 0 < < 1, = .Proofs are omitted.
Theorem 4.15. (a) A distribution F belongs to the domain ofattraction of the normal distribution iff
U(x) =
xx
y2F{dy} (4.38)
varies slowly.
(b) A distribution F belongs to the domain of attraction of a stable
distribution with characteristic exponent < 2 iff
1 F(x) + F(x) xL(x) (x ) (4.39)and
1 F(x)1 F(x) + F(x) p,
F(x)1 F(x) + F(x) q (4.40)
where p 0, q 0 and p +q = 1, Here L is a slowly varying function
on (0, ); that is, for each x > 0L(tx)
L(t) 1 as t . (4.41)
The proof is omitted.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
75/90
66 Topics in Probability
Theorem 4.16. Let F be a proper distribution concentrated on
(0, ) and Fn the n-fold convolution of F with itself: If Fn(anx) G(x), where G is a non-degenerate distribution, then G = G, thestable distribution concentrated on (0, ), with exponent (0 < 0) or() = ec
, so that G is the stable distribution with exponent .
Here 0 < < 1 since G is non-degenerate.(ii) Conversely, let 1 F(t) tL(t)/(1 )(t ). Thisgives 1 F() L(1/)( 0+). Let us choose constants an sothat n[1 F(an)] c/(1 ) for 0 < c < . Then as n ,
nan L(an) =an L(an)
[1 F(an)](1 ) n[1 F(an)](1 ) c
and also
nan L(an/) = nan L(an)
L(an/)L(an)
c.
Therefore 1 F(/an) c/n andF(/an)n = [1 c/n + o(1/n)]n ec .
This shows that Fn(anx) G(x).
4.7. Problems for Solution
1. Show that if F and G are infinitely divisible distributions so is
their convolution F G.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
76/90
Infinitely Divisible Distributions 67
2. If is an infinitely divisible c.f., prove that || is also an infinitelydivisible c.f.
3. Show that the uniform distribution is not infinitely divisible.More generally, a distribution concentrated on a finite interval is
not infinitely divisible, unless it is concentrated at a point.
4. Let 0 < rj < 1 and
rj < . Prove that for arbitrary aj theinfinite product
() =
j=11 rj
1
rjeiaj
converges, and represents an infinitely divisible c.f.
5. Let X =
1 Xk/k where the random variables Xk are inde-
pendent and have the common density 12e|x|. Show that X is
infinitely divisible, and find the associated Feller measure.
6. Let P be an infinitely divisible p.g.f. and the c.f. of an arbitrary
distribution. Show that P() is an infinitely divisible c.f.
7. If 0
a < b < 1 and is a c.f., then show that
1 b1 a
1 a1 b
is an infinitely divisible c.f.
8. Prove that a probability distribution with a completely monotone
density is infinitely divisible.
9. Mixtures of exponential (geometric) distributions. Let
f(x) =n
k=1
pkkekx
where pk > 0,
pk = 1 and for definiteness 0 < 1 < 2 < < n. Show that the density f(x) is infinitely divisible. (Similarly
a mixture of geometric distributions is infinitely divisible.) By a
limit argument prove that the density
f(x) =0
exG(d),
where G is a distribution concentrated on (0, ), is infinitelydivisible.
-
7/29/2019 [Www.gfxmad.me] 9814335479 Probabilit
77/90
68 Topics in Probability
10. If X, Y are two independent random variables such that X > 0
and Y has an exponential density, then prove that XY is
infinitely divisible.11. Show that a c.f. is stable if and only if given c > 0, c > 0
there exist constants c > 0, d such that
(c)(c) = (c)eid .
12. Let the c.f. be given by log () = 2
2k(cos 2k 1).
Show that ()n = (n) for n = 2, 4, 8, . . . , () is infinitely
divisible, but not stable.
13. If ()2 = (c) and the variance is finite, show that () is
stable (in fact normal).
14. If ()2 = (a) and ()3 = (b) with a > 0, b > 0, show
that () is stable.
15. If F and G are stable with the same exponent , so is their
convolution F G.16. If X, Y are independent random variables such that X is stable
with exponent , while Y is positive and stable with exponent(< 1), show that XY1/ is stable with exponent .
17. The Holtsmark distribution. Suppose that n stars are dis-
tributed in the interval (n, n) on the real line, their locationsdi(i = 1, 2, . . . , n) being independent r.v. with a