Graduate Texts in Mathematics
Editorial Board
95
F. W. Gehring P.R. Halmos (Managing Editor)
C. C. Moore
"Order out of chaos"
(Courtesy of Professor A. T. Fomenko of the Moscow State University)
A. N. Shiryayev
Probability
Translated by R. P. Boas
With 54 Illustrations
Springer Science+Business Media, LLC
A. N. Shiryayev
Steklov Mathematical Institute
Vavilova 42
GSP1 117333 Moscow
U.S.S.R.
Editorial Board
P.R. Halmos
Managing Editor
Department of
Mathematics
Indiana University
Bloomington, IN 47405
U.S.A.
R. P. Boas (Translator)
Department of Mathematics
Northwestern University
Evanston, IL 60201
U.S.A.
F. W. Gehring
C. C. Moore
Department of
Mathematics
University of Michigan
Ann Arbor, MI 48109
U.S.A.
Department of
Mathematics
University of California
at Berkeley
Berkeley, CA 94720
U.S.A.
AMS Classification: 600 I
Library of Congress Cataloging in Publication Data
Shiriaev, Al'bert Nikolaevich.
Probability.
(Graduate texts in mathematics; 95)
Translation of: Veroiiitnost'.
Bibliography: p.
Includes index.
I. Probabilities. I. Title. II. Series.
519
8314813
QA273.S54413 1984
Original Russian edition: Veroicltnost'. Moscow: Nauka, 1979.
This book is part of the Springer Series in Soviet Mathematics.
1984 by Springer Science+Business Media New York
Originally published by SpringerVerlag New York, Inc. in 1984
Softcover reprint of the hardcover 1st edition 1984
All rights reserved. No part of this book may be translated or reproduced in any
form without written permission from Springer Science+Business Media, LLC.
Typeset by Composition House Ltd., Salisbury, England.
9 8 7 6 54 32 1
ISBN 9781489900203
ISBN 9781489900180 (eBook)
DOI 10.1007/9781489900180
Preface
This textbook is based on a threesemester course of lectures given by the
author in recent years in the MechanicsMathematics Faculty of Moscow
State University and issued, in part, in mimeographed form under the title
Probability, Statistics, Stochastic Processes, I, II by the Moscow State
University Press.
We follow tradition by devoting the first part of the course (roughly one
semester) to the elementary theory of probability (Chapter I). This begins
with the construction of probabilistic models with finitely many outcomes
and introduces such fundamental probabilistic concepts as sample spaces,
events, probability, independence, random variables, expectation, correlation, conditional probabilities, and so on.
Many probabilistic and statistical regularities are effectively illustrated
even by the simplest random walk generated by Bernoulli trials. In this
connection we study both classical results (law of large numbers, local and
integral De Moivre and Laplace theorems) and more modern results (for
example, the arc sine law).
The first chapter concludes with a discussion of dependent random variables generated by martingales and by Markov chains.
Chapters IIIV form an expanded version of the second part of the course
(second semester). Here we present (Chapter II) Kolmogorov's generally
accepted axiomatization of probability theory and the mathematical methods
that constitute the tools of modern probability theory (aalgebras, measures
and their representations, the Lebesgue integral, random variables and
random elements, characteristic functions, conditional expectation with
respect to a aalgebra, Gaussian systems, and so on). Note that two measuretheoretical resultsCaratheodory's theorem on the extension of measures
and the RadonNikodym theoremare quoted without proof.
VI
Preface
The third chapter is devoted to problems about weak convergence of
probability distributions and the method of characteristic functions for
proving limit theorems. We introduce the concepts of relative compactness
and tightness of families of probability distributions, and prove (for the
real line) Prohorov's theorem on the equivalence of these concepts.
The same part of the course discusses properties "with probability 1"
for sequences and sums of independent random variables (Chapter IV). We
give proofs of the "zero or one laws" of Kolmogorov and of Hewitt and
Savage, tests for the convergence of series, and conditions for the strong law
of large numbers. The law of the iterated logarithm is stated for arbitrary
sequences of independent identically distributed random variables with
finite second moments, and proved under the assumption that the variables
have Gaussian distributions.
Finally, the third part of the book (Chapters VVIII) is devoted to random
processes with discrete parameters (random sequences). Chapters V and VI
are devoted to the theory of stationary random sequences, where "stationary" is interpreted either in the strict or the wide sense. The theory of random
sequences that are stationary in the strict sense is based on the ideas of
ergodic theory: measure preserving transformations, ergodicity, mixing, etc.
We reproduce a simple proof (by A. Garsia) of the maximal ergodic theorem;
this also lets us give a simple proof of the BirkhoffKhinchin ergodic theorem.
The discussion of sequences of random variables that are stationary in
the wide sense begins with a proof of the spectral representation of the
covariance fuction. Then we introduce orthogonal stochastic measures, and
integrals with respect to these, and establish the spectral representation of
the sequences themselves. We also discuss a number of statistical problems:
estimating the covariance function and the spectral density, extrapolation,
interpolation and filtering. The chapter includes material on the KalmanBucy filter and its generalizations.
The seventh chapter discusses the basic results of the theory of martingales
and related ideas. This material has only rarely been included in traditional
courses in probability theory. In the last chapter, which is devoted to Markov
chains, the greatest attention is given to problems on the asymptotic behavior
of Markov chains with countably many states.
Each section ends with problems of various kinds: some of them ask for
proofs of statements made but not proved in the text, some consist of
propositions that will be used later, some are intended to give additional
information about the circle of ideas that is under discussion, and finally,
some are simple exercises.
In designing the course and preparing this text, the author has used a
variety of sources on probability theory. The Historical and Bibliographical
Notes indicate both the historical sources of the results, and supplementary
references for the material under consideration.
The numbering system and form of references is the following. Each
section has its own enumeration of theorems, lemmas and formulas (with
Preface
Vll
no indication of chapter or section). For a reference to a result from a
different section of the same chapter, we use double numbering, with the
first number indicating the number of the section (thus (2.10) means formula
(10) of 2). For references to a different chapter we use triple numbering
(thus formula (11.4.3) means formula (3) of 4 of Chapter II). Works listed
in the References at the end of the book have the form [L n], where Lis a
letter and n is a numeral.
The author takes this opportunity to thank his teacher A. N. Kolmogorov,
and B. V. Gnedenko and Yu. V. Prohorov, from whom he learned probability
theory and under whose direction he had the opportunity of using it. For
discussions and advice, the author also thanks his colleagues in the Departments of Probability Theory and Mathematical Statistics at the Moscow
State University, and his colleagues in the Section on probability theory of the
Steklov Mathematical Institute of the Academy of Sciences of the U.S.S.R.
Moscow
Steklov Mathematical Institute
A. N.
SHIRYAYEV
Translator's acknowledgement. I am grateful both to the author and to
my colleague C. T. Ionescu Tulcea for advice about terminology.
R. P. B.
Contents
Introduction
CHAPTER I
Elementary Probability Theory
1. Probabilistic Model of an Experiment with a Finite Number of
Outcomes
2. Some Classical Models and Distributions
3. Conditional Probability. Independence
4. Random Variables and Their Properties
5. The Bernoulli Scheme. I. The Law of Large Numbers
6. The Bernoulli Scheme. II. Limit Theorems (Local,
De MoivreLaplace, Poisson)
~7. Estimating the Probability of Success in the Bernoulli Scheme
8. Conditional Probabilities and Mathematical Expectations with
Respect to Decompositions
9. Random Walk. I. Probabilities of Ruin and Mean Duration in
Coin Tossing
10. Random Walk. II. Reflection Principle. Arcsine Law
11. Martingales. Some Applications to the Random Walk
12. Markov Chains. Ergodic Theorem. Strong Markov Property
5
5
17
23
32
45
55
68
74
81
92
101
108
CHAPTER II
Mathematical Foundations of Probability Theory
1. Probabilistic Model for an Experiment with Infinitely Many
Outcomes. Kolmogorov's Axioms
2. Algebras and ualgebras. Measurable Spaces
3. Methods oflntroducing Probability Measures on Measurable Spaces
4. Random Variables. I.
129
129
137
149
164
Contents
5. Random Elements
6. Lebesgue Integral. Expectation
7. Conditional Probabilities and Conditional Expectations with
Respect to a aAlgebra
8. Random Variables. II.
9. Construction of a Process with Given FiniteDimensional
Distribution
10. Various Kinds of Convergence of Sequences of Random Variables
11. The Hilbert Space of Random Variables with Finite Second Moment
12. Characteristic Functions
13. Gaussian Systems
174
178
210
232
243
250
260
272
295
CHAPTER III
Convergence of Probability Measures. Central Limit
Theorem
1. Weak Convergence of Probability Measures and Distributions
2. Relative Compactness and Tightness of Families of Probability
Distributions
3. Proofs of Limit Theorems by the Method of Characteristic Functions
4. Central Limit Theorem for Sums of Independent Random Variables
5. Infinitely Divisible and Stable Distributions
6. Rapidity of Convergence in the Central Limit Theorem
7. Rapidity of Convergence in Poisson's Theorem
306
306
314
318
326
335
342
345
CHAPTER IV
Sequences and Sums of Independent Random Variables
1.
2.
3.
4.
ZeroorOne Laws
Convergence of Series
Strong Law of Large Numbers
Law of the Iterated Logarithm
354
354
359
363
370
CHAPTER V
Stationary (Strict Sense) Random Sequences and Ergodic
Theory
1. Stationary (Strict Sense) Random Sequences. MeasurePreserving
Transformations
2. Ergodicity and Mixing
3. Ergodic Theorems
376
376
379
381
CHAPTER VI
Stationary (Wide Sense) Random Sequences. L 2 Theory
1.
2.
3.
4.
Spectral Representation of the Covariance Function
Orthogonal Stochastic Measures and Stochastic Integrals
Spectral Representation of Stationary (Wide Sense) Sequences
Statistical Estimation of the Covariance Function and the Spectral
Density
387
387
395
401
412
Contents
5. Wold's Expansion
6. Extrapolation, Interpolation and Filtering
7. The KalmanBucy Filter and Its Generalizations
xi
418
425
436
CHAPTER VII
Sequences of Random Variables that Form Martingales
1. Definitions of Martingales and Related Concepts
2. Preservation of the Martingale Property Under Time Change at a
Random Time
3. Fundamental Inequalities
4. General Theorems on the Convergence of Submartingales and
Martingales
5. Sets of Convergence of Submartingales and Martingales
6. Absolute Continuity and Singularity of Probability Distributions
7. Asymptotics of the Probability of the Outcome of a Random Walk
with Curvilinear Boundary
8. Central Limit Theorem for Sums of Dependent Random Variables
446
446
456
464
476
483
492
504
509
CHAPTER VIII
Sequences of Random Variables that Form Markov Chains
I. Definitions and Basic Properties
2. Classification of the States of a Markov Chain in Terms of
Arithmetic Properties of the Transition Probabilities p!jl
3. Classification of the States of a Markov Chain in Terms of
Asymptotic Properties of the Probabilities p!i>
4. On the Existence of Limits and of Stationary Distributions
5. Examples
523
523
528
532
541
546
Historical and Bibliographical Notes
555
References
561
Index of Symbols
565
Index
569
Introduction
The subject matter of probability theory is the mathematical analysis of
random events, i.e. of those empirical phenomena whichunder certain
circumstancescan be described by saying that:
They do not have deterministic regularity (observations of them do not
yield the same outcome) whereas at the same time;
They possess some statistical regularity (indicated by the statistical
stability of their frequency).
We illustrate with the classical example of a "fair" toss of an "unbiased"
coin. It is clearly impossible to predict with certainty the outcome of each
toss. The results of successive experiments are very irregular (now "head,"
now "tail") and we seem to have no possibility of discovering any regularity
in such experiments. However, if we carry out a large number of" independent" experiments with an "unbiased" coin we can observe a very definite
statistical regularity, namely that "head" appears with a frequency that is
"close" to f.
Statistical stability of a frequency is very likely to suggest a hypothesis
about a possible quantitative estimate of the "randomness" of some event A
connected with the results of the experiments. With this starting point,
probability theory postulates that corresponding to an event A there is a
definite number P(A), called the probability of the event, whose intrinsic
property is that as the number of "independent" trials (experiments) increases the frequ,ency of event A is approximated by P(A).
Applied to our example, this means that it is natural to assign the probability ! to the event A that consists of obtaining "head" in a toss of an
"unbiased" coin.
Introduction
There is no difficulty in multiplying examples in which it is very easy to
obtain numerical values intuitively for the probabilities of one or another
event. However, these examples are all of a similar nature and involve (so far)
undefined concepts such as "fair" toss, "unbiased" coin, "independence,"
etc.
Having been invented to investigate the quantitative aspects of"randomness," probability theory, like every exact science, became such a science
only at the point when the concept of a probabilistic model had been clearly
formulated and axiomatized. In this connection it is natural for us to discuss,
although only briefly, the fundamental steps in the development of probability theory.
Probability theory, as a science, originated in the middle of the seventeenth
century with Pascal (16231662), Fermat (16011655) and Huygens
( 16291695). Although special calculations of probabilities in games of chance
had been made earlier, in the fifteenth and sixteenth centuries, by Italian
mathematicians (Cardano, Pacioli, Tartaglia, etc.), the first general methods
for solving such problems were apparently given in the famous correspondence between Pascal and Fermat, begun in 1654, and in the first book on
probability theory, De Ratiociniis in Aleae Ludo (On Calculations in Games of
Chance), published by Huygens in 1657. It was at this time that the fundamental concept of" mathematical expectation" was developed and theorems
on the addition and multiplication of probabilities were established.
The real history of probability theory begins with the work of James
Bernoulli (16541705), Ars Conjectandi (The Art of Guessing) published in
1713, in which he proved (quite rigorously) the first limit theorem of probability theory, the law of large numbers; and of De Moivre (16671754),
Miscellanea Analytica Supplementum (a rough translation might be The
Analytic Method or Analytic Miscellany, 1730), in which the central limit
theorem was stated and proved for the first time (for symmetric Bernoulli
trials).
Bernoulli was probably the first to realize the importance of considering
infinite sequences of random trials and to make a clear distinction between
the probability of an event and the frequency of its realization. De Moivre
deserves the credit for defining such concepts as independence, mathematical
expectation, and conditional probability.
In 1812 there appeared Laplace's (17491827) great treatise Theorie
Analytique des Probabilities (Analytic Theory of Probability) in which he
presented his own results in probability theory as well as those of his predecessors. In particular, he generalized De Moivre's theorem to the general
(unsymmetric) case of Bernoulli trials, and at the same time presented De
Moivre's results in a more complete form.
Laplace's most important contribution was the application of probabilistic methods to errors of observation. He formulated the idea of considering errors of observation as the cumulative results of adding a large number
of independent elementary errors. From this it followed that under rather
Introduction
general conditions the distribution of errors of observation must be at least
approximately normal.
The work of Poisson (17811840) and Gauss (17771855) belongs to the
same epoch in the development of probability theory, when the center of the
stage was held by limit theorems.
In contemporary probability theory we think of Poisson in connection
with the distribution and the process that bear his name. Gauss is credited
with originating the theory of errors and, in particular, with creating the
fundamental method of least squares.
The next important period in the development of probability theory is
connected with the names of P. L. Chebyshev (18211894), A. A. Markov
(18561922), and A. M. Lyapunov (18571918), who developed effective
methods for proving limit theorems for sums of independent but arbitrarily
distributed random variables.
The number of Chebyshev's publications in probability theory is not
largefour in allbut it would be hard to overestimate their role in probability theory and in the development of the classical Russian school of that
subject.
"On the methodological side, the revolution brought about by Chebyshev
was not only his insistence for the first time on complete rigor in the proofs of
limit theorems, ... but also, and principally, that Chebyshev always tried to
obtain precise estimates for the deviations from the limiting regularities that are
available for large but finite numbers of trials, in the form of inequalities that are
valid unconditionally for any number of trials."
(A. N.
KOLMOGOROV
[30])
Before Chebyshev the main interest in probability theory had been in the
calculation of the probabilities of random events. He, however, was the
first to realize clearly and exploit the full strength of the concepts of random
variables and their mathematical expectations.
The leading exponent of Chebyshev's ideas was his devoted student
Markov, to whom there belongs the indisputable credit of presenting his
teacher's results with complete clarity. Among Markov's own significant
contributions to probability theory were his pioneering investigations of
limit theorems for sums of independent random variables and the creation
of a new branch of probability theory, the theory of dependent random
variables that form what we now call a Markov chain.
" ... Markov's classical course in the calculus of probability and his original
papers, which are models of precision and clarity, contributed to the greatest
extent to the transformation of probability theory into one of the most significant
branches of mathematics and to a wide extension of the ideas and methods of
Chebyshev."
{S. N. BERNSTEIN [3])
To prove the central limit theorem of probability theory (the theorem
on convergence to the normal distribution), Chebyshev and Markov used
Introduction
what is known as the method of moments. With more general hypotheses
and a simpler method, the method of characteristic functions, the theorem
was obtained by Lyapunov. The subsequent development of the theory has
shown that the method of characteristic functions is a powerful analytic
tool for establishing the most diverse limit theorems.
The modern period in the development of probability theory begins with
its axiomatization. The first work in this direction was done by S. N. Bernstein (18801968), R. von Mises (18831953), and E. Borel (18711956).
A. N. Kolmogorov's book Foundations of the Theory of Probability appeared
in 1933. Here he presented the axiomatic theory that has become generally
accepted and is not only applicable to all the classical branches of probability
theory, but also provides a firm foundation for the development of new
branches that have arisen from questions in the sciences and involve infinitedimensional distributions.
The treatment in the present book is based on Kolmogorov's axiomatic
approach. However, to prevent formalities and logical subtleties from obscuring the intuitive ideas, our exposition begins with the elementary theory of
probability, whose elementariness is merely that in the corresponding
probabilistic models we consider only experiments with finitely many outcomes. Thereafter we present the foundations of probability theory in their
most general form.
The 1920s and '30s saw a rapid development of one of the new branches of
probability theory, the theory of stochastic processes, which studies families
of random variables that evolve with time. We have seen the creation of
theories of Markov processes, stationary processes, martingales, and limit
theorems for stochastic processes. Information theory is a recent addition.
The present book is principally concerned with stochastic processes with
discrete parameters: random sequences. However, the material presented
in the second chapter provides a solid foundation (particularly of a logical
nature) for thestudy of the general theory of stochastic processes.
It was also in the 1920s and '30s that mathematical statistics became a
separate mathematical discipline. In a certain sense mathematical statistics
deals with inverses of the problems of probability: If the basic aim of probability theory is to calculate the probabilities of complicated events under a
given probabilistic model, mathematical statistics sets itself the inverse
problem: to clarify the structure of probabilisticstatistical models by
means of observations of various complicated events.
Some of the problems and methods of mathematical statistics are also
discussed in this book. However, all that is presented in detail here is probability theory and the theory of stochastic processes with discrete parameters.
CHAPTER I
Elementary Probability Theory
1. Probabilistic Model of an Experiment with a
Finite Number of Outcomes
1. Let us consider an experiment of which all possible results are included
in a finite number of outcomes w 1, .. , wN. We do not need to know the
nature of these outcomes, only that there are a finite number N of them.
We call w 1,
wN elementary events, or sample points, and the finite set
Q = {wl, ... , roN},
the space of elementary events or the sample space.
The choice of the space of elementary events is the first step in formulating
a probabilistic model for an experiment. Let us consider some examples of
sample spaces.
ExAMPLE
1. For a single toss of a coin the sample space Q consists of two
points:
Q = {H, T},
where H ="head" and T ="tail". (We exclude possibilities like "the coin
stands on edge," "the coin disappears," etc.)
ExAMPLE
2. For n tosses of a coin the sample space is
Q = {w: w = (a 1 ,
an), ai = H or T}
and the general number N(Q) of outcomes is 2".
I. Elementary Probability Theory
3. First toss a coin. If it falls "head" then toss a die (with six faces
numbered 1, 2, 3, 4, 5, 6); if it falls "tail", toss the coin again. The sample
space for this experiment is
ExAMPLE
Q = {H1, H2, H3, H4, H5, H6, TH, TT}.
We now consider some more complicated examples involving the selection of n balls from an urn containing M distinguishable balls.
2. ExAMPLE 4 (Sampling with replacement). This is an experiment in which
after each step the selected ball is returned again. In this case each sample of
n balls can be presented in the form (a 1 , , an), where ai is the label of the
ball selected at the ith step. It is clear that in sampling with replacement
each ai can have any of the M values 1, 2, ... , M. The description of the
sample space depends in an essential way on whether we consider samples
like, for example, (4, 1, 2, 1) and (1, 4, 2, 1) as different or the same. It is
customary to distinguish two cases: ordered samples and unordered samples.
In the first case samples containing the same elements, but arranged
differently, are considered to be different. In the second case the order of
the elements is disregarded and the two samples are considered to be the
same. To emphasize which kind of sample we are considering, we use the
notation (a 1 , .. , an) for ordered samples and [a 1 , .. , an] for unordered
samples.
Thus for ordered samples the sample space has the form
Q
= {w: w = (a 1, , an), ai = 1, ... , M}
and the number of (different) outcomes is
N(Q) = Mn.
(1)
If, however, we consider unordered samples, then
Q
= {w: w = [a 1 , ... , an], ai = 1, ... , M}.
Clearly the number N(Q) of (different) unordered samples is smaller than
the number of ordered samples. Let us show that in the present case
N(Q) = CM+nl
(2)
where q = k !j[l! (k  I)!] is the number of combinations of I elements,
taken k at a time.
We prove this by induction. Let N(M, n) be the number of outcomes of
interest. It is clear that when k ~ M we have
N(k, 1) = k =
q.
I. Probabilistic Model of an Experiment with a Finite Number of Outcomes
Now suppose that N(k, n) = C~+n 1 fork~ M; we show that this formula
continues to hold when n is replaced by n + 1. For the unordered samples
[a 1 , ... , an+ 1] that we are considering, we may suppose that the elements
are arranged in nondecreasing order: a 1 ~ a 2 ~ ~ an. It is clear that the
number of unordered samples with a 1 = 1 is N(M, n), the number with
a 1 = 2 is N(M  1, n), etc. Consequently
N(M, n
+ 1) =
+ N(M  1, n) + + N(1, n)
= CM+n1 + CM1+n1 + ' c:
= (C~+1n CM+1n1) + (CM+\+n C~!1+n1)
+ ... + (C:!i c:) = C~_}n;
N(M, n)
here we have used the easily verified property
Ci 1 + Ci = Ci + 1
of the binomial coefficients.
5 (Sampling without replacement). Suppose that n ~ M and that
the selected balls are not returned. In this case we again consider two possibilities, namely ordered and unordered samples.
For ordered samples without replacement the sample space is
EXAMPLE
= {w: w = (a 1 , .. , an), ak =F a1, k =F l, ai = 1, ... , M},
and the number of elements of this set (called permutations) is M(M 1)
(M  n + 1). We denote this by (M)n or AM and call it "the number of
permutations of M things, n at a time").
For unordered samples (called combinations) the sample space
Q
= {w: w = [a 1 , .. , an], ak =F a1, k =F l, ai = 1, ... , M}
consists of
N(Q)
CM
(3)
elements. In fact, from each unordered sample [a 1 , .. , an] consisting of
distinct elements we can obtain n! ordered samples. Consequently
N(Q) n! = (M)n
and therefore
N(Q) = (M;n =
n.
CM.
The results on the numbers of samples of n from an urn with M balls are
presented in Table 1.
I. Elementary Probability Theory
Table 1
M"
c~+n1
With
replacement
(M).
eM
Without
replacement
Ordered
Unordered
For the case M = 3 and n
displayed in Table 2.
= 2,
~
e
the corresponding sample spaces are
ExAMPLE 6 (Distribution of objects in cells). We consider the structure of
the sample space in the problem of placing n objects (balls, etc.) in M cells
(boxes, etc.). For example, such problems arise in statistical physics in studying the distribution of n particles (which might be protons, electrons, ... )
among M states (which might be energy levels).
Let the cells be numbered 1, 2, ... , M, and suppose first that the objects
are distinguishable (numbered 1, 2, ... , n). Then a distribution of the n
objects among the M cells is completely described by an ordered set
(a 1, .. , an), where ai is the index of the cell containing object i. However,
if the objects are indistinguishable their distribution among the M cells
is completely determined by the unordered set [a 1, .. , an], where ai is the
index of the cell into which an object is put at the ith step.
Comparing this situation with Examples 4 and 5, we have the following
correspondences:
(ordered samples)++ (distinguishable objects),
(unordered samples)++ (indistinguishable objects),
Table 2
(1, 1) (1, 2) (1, 3)
(2, 1) (2, 2) (2, 3)
(3, 1) (3, 2) (3, 3)
[1, 1] [2, 2] [3, 3]
[1, 2] [1, 3]
[2, 3]
With
replacement
(1, 2) (1, 3)
(2, 1) (2, 3)
(3, 1) (3, 2)
[1, 2] [1, 3]
[2, 3]
Without
replacement
Ordered
Unordered
~
e
1. Probabilistic Model of an Experiment with a Finite Number of Outcomes
by which we mean that to an instance of an ordered (unordered) sample of
n balls from an urn containing M balls there corresponds (one and only one)
instance of distributing n distinguishable (indistinguishable) objects among
M cells.
In a similar sense we have the following correspondences:
) (a cell may receive any number)
h
.
(samp1mg
w1t rep1acement f b" t
,
o o ~ec s
.
.
(a cell may receive at most)
b" t
(samphng without replacement)one o ~ec
These correspondences generate others of the same kind:
an unordered sample in)
( sampling without
replacement
(indistinguishable objects in the
)
problem of distribution among cells
when each cell may receive at
most one object
etc.; so that we can use Examples 4 and 5 to describe the sample space for
the problem of distributing distinguishable or indistinguishable objects
among cells either with exclusion (a cell may receive at most one object) or
without exclusion (a cell may receive any number of objects).
Table 3 displays the distributions of two objects among three cells. For
distinguishable objects, we denote them by W (white) and B (black). For
indistinguishable objects, the presence of an object in a cell is indicated
bya +.
Table 3
lwlsl I llwlsl llwl 18 I I++ I I II I+ +I II I I++ I
lslwl II lwlsl II lwlsl
1+1+1 I 1+1 1+1
IBl lwll lslwll I lwlsl
I 1+1+1
lwlsl
lslwl
I8 1
I lwl Isl
I I lwlsl
lwl I lslwl
Distinguishable
objects
1+1+1
I 1+1 1+1
I 1+1+1
:::: =
0
0 
~2u
..<:
><
'"'
.9
"'
~
><
'"'
Indistinguishable
objects
I~
n
10
I. Elementary Probability Theory
Table 4
N(Q) in the problem of placing n objects in M cells
Distinguishable
objects
Indistinguishable
objects
Without exclusion
Mn
C:V+n1
With exclusion
{M)n
c:v
Ordered
samples
Unordered
samples
(MaxwellBoltzmann
statistics)
(BoseEinstein
statistics)
(FermiDirac
statistics)
With
replacement
Without
replacement
~
e
N(Q) in the problem of choosing n balls from an urn
containing M balls
The duality that we have observed between the two problems gives us
obvious way of finding the number of outcomes in the problem of placing
objects in cells. The results, which include the results in Table 1, are given in
Table 4.
In statistical physics one says that distinguishable (or indistinguishable,
respectively) particles that are not subject to the Pauli exclusion principlet
obey MaxwellBoltzmann statistics (or, respectively, BoseEinstein statistics). If, however, the particles are indistinguishable and are subject to the
exclusion principle, they obey FermiDirac statistics (see Table 4). For
example, electrons, protons and neutrons obey FermiDirac statistics.
Photons and pions obey BoseEinstein statistics. Distinguishable particles
that are subject to the exclusion principle do not occur in physics.
an
3. In addition to the concept of sample space we now need the fundamental
concept of event.
Experimenters are ordinarily interested, not in what particular outcome
occurs as the result of a trial, but in whether the outcome belongs to some
subset of the set of all possible outcomes. We shall describe as events all
subsets A c Q for which, under the conditions ofthe experiment, it is possible
to say either "the outcome wE A" or "the outcome w A."
t At most one particle in each cell. (Translator)
1. Probabilistic Model of an Experiment with a Finite Number of Outcomes
11
For example, let a coin be tossed three times. The sample space Q consists
of the eight points
Q = {HHH, HHT, ... , TTT}
and if we are able to observe (determine, measure, etc.) the results of all three
tosses, we say that the set
A = {HHH, HHT, HTH, THH}
is the event consisting of the appearance of at least two heads. If, however,
we can determine only the result of the first toss, this set A cannot be considered to be an event, since there is no way to give either a positive or negative
answer to the question of whether a specific outcome w belongs to A.
Starting from a given collection of sets that are events, we can form new
events by means of statements containing the logical connectives "or,"
"and," and "not," which correspond in the language of set theory to the
operations "union," "intersection," and "complement."
If A and B are sets, their union, denoted by A u B, is the set of points that
belong either to A or to B:
Au B ={wen: we A or weB}.
In the language of probability theory, Au B is the event consisting of the
realization either of A or of B.
The intersection of A and B, denoted by A n B, or by AB, is the set of
points that belong to both A and B:
An B = {weQ: we A and weB}.
The event A n B consists of the simultaneous realization of both A and B.
For example, if A = {HH, HT, TH} and B = {TT, TH, HT} then
Au B
= {HH, HT, TH, TT} ( =0),
A n B = {TH, HT}.
If A is a subset of Q, its complement, denoted by A, is the set of points of
that do not belong to A.
If B\A denotes the difference of B and A (i.e. the set of points that belong
to B but not to A) then A= Q\A. In the language of probability, A is
the event consisting of the nonrealization of A. For example, if A =
{HH, HT, TH} then A = {TT}, the event in which two successive tails occur.
The sets A and A have no points in common and consequently A n A is
empty. We denote the empty set by 0. In probability theory, 0 is called an
impossible event. The set n is naturally called the certain event.
When A and B are disjoint (AB = 0), the union A u B is called the
sum of A and B and written A + B.
If we consider a collection d 0 of sets A s Q we may use the settheoretic
operators u, n and \ to form a new collection of sets from the elements of
Q
12
I. Elementary Probability Theory
d 0 ; these sets are again events. If we adjoin the certain and impossible
events Q and 0 we obtain a collection d of sets which is an algebra, i.e. a
collection of subsets of Q for which
(1) Qed,
(2) if A Ed, BEd, the sets A u B, A n B, A\B also belong to d.
It follows from what we have said that it will be advisable to consider
collections of events that form algebras. In the future we shall consider only
such collections.
Here are some examples of algebras of events:
(a) {0, 0}, the collection consisting of Q and the empty set (we call this the
trivial algebra);
(b) {A, A, 0, 0}, the collection generated by A;
(c) d = {A: A 0}, the collection consisting of all the subsets of Q
(including the empty set 0).
It is easy to check that all these algebras of events can be obtained from the
following principle.
We say that a collection
of sets is a decomposition of 0, and call the Di the atoms of the decomposition,
if the Di are not empty, are pairwise disjoint, and their sum is 0:
D1
+ + Dn
= Q.
For example, if Q consists of three points,
different decompositions:
!'}t={Dd
!'}2
= {Dt, D2}
= {1, 2, 3}, there are five
with D 1 = {1, 2, 3};
with D 1
= {1, 2}, D2 = {3};
!'}3 = {Dt, D2}
with D 1 = {1, 3}, D2 = {2};
!'}4 = {Dt, D2}
with D 1 = {2,3},D 2 = {1};
!'}s
= {Dl, D2, D3}
with
D1
= {1}, D2 = {2}, D3 = {3}.
(For the general number of decompositions of a finite set see Problem 2.)
If we consider all unions of the sets in !'}, the resulting collection of sets,
together with the empty set, forms an algebra, called the algebra induced by
!'}, and denoted by ex(!'}). Thus the elements of ex(!'}) consist of the empty set
together with the sums of sets which are atoms of!'},
Thus if!'} is a decomposition, there is associated with it a specific algebra
fJi = ex(!'}).
The converse is also true. Let f!i be an algebra of subsets of a finite space
0. Then there is a unique decomposition !'} whose atoms are the elements of
I. Probabilistic Model of an Experiment with a Finite Number of Outcomes
13
flA, with flA = oc(~). In fact, let D E 14 and let D have the property that for
every B E flA the set D n B either coincides with D or is empty. Then this
collection of sets D forms a decomposition ~ with the required property
oc(~) = flA. In Example (a),~ is the trivial decomposition consisting of the
single set D1 = Q; in (b),~= {A, A}. The most finegrained decomposition
~, which consists of the singletons {roJ, roi E Q, induces the algebra in
Example (c), i.e. the algebra of all subsets of Q.
Let ~ 1 and ~ 2 be two decompositions. We say that ~ 2 is finer than ~ 1 ,
and write ~ 1 ~ ~ 2 , ifoc(~ 1 ) ~ oc(~ 2 ).
Let us show that if Q consists, as we assumed above, of a finite number of
points ro 1, ... , roN, then the number N(d) of sets in the collection d is
equal to 2N. In fact, every nonempty set A Ed can be represented as A =
{roi,, ... , rod, where roij E Q, 1 ~ k ~ N. With this set we associate the sequence of zeros and ones
(0, ... ' 0, 1, 0, ... ' 0, 1, ... ),
where there are ones in the positions i 1, ... , ik and zeros elsewhere. Then
for a given k the number of different sets A of the form {roi,, ... , roik} is the
same as the number of ways in which k ones (k indistinguishable objects)
can be placed in N positions (N cells). According to Table 4 (see the lower
righthand square) we see that this number is C~. Hence (counting the empty
set) we find that
N(d) = 1 + C~
+ .. + CZ
= (1
+ It
= 2N.
4. We have now taken the first two steps in defining a probabilistic model
of an experiment with a finite number of outcomes: we have selected a sample
space and a collection d of subsets, which form an algebra and are called
events. We now take the next step, to assign to each sample point (outcome)
roi E Qi, i = 1, ... , N, a weight. This is denoted by p(roi) and called the
probability of the outcome roi ; we assume that it has the following properties:
(a) 0 ~ p(roJ ~ 1 (nonnegativity),
(b) p(ro 1) + + p(roN) = 1 (normalization).
Starting from the given probabilities p(roJ of the outcomes roi> we define
the probability P(A) of any event A Ed by
(4)
{i:ro;eA}
Finally, we say that a triple
(Q, d, P),
where Q = {ro 1 ,
. ,
roN}, dis an algebra of subsets of Q and
p = {P(A); A Ed}
14
I. Elementary Probability Theory
defines (or assigns) a probabilistic model, or a probability space, of experiments
with a (finite) space n of outcomes and algebra d of events.
The following properties of probability follow from (4):
(5)
P(0) = 0,
P(Q)
P(A u B)= P(A)
(6)
1,
P(B) P(A n B).
(7)
In particular, if An B = 0, then
P(A
+ B) =
P(A)
P(B)
(8)
and
P(A) = 1  P(A).
(9)
5. In constructing a probabilistic model for a specific situation, the construction of the sample space n and the algebra d of events are ordinarily
not difficult. In elementary probability theory one usually takes the algebra
d to be the algebra of all subsets of n. Any difficulty that may arise is in
assigning probabilities to the sample points. In principle, the solution to this
problem lies outside the domain of probability theory, and we shall not
consider it in detail. We consider that our fundamental problem is not the
question of how to assign probabilities, but how to calculate the probabilities of complicated events (elements of d) from the probabilities of the
sample points.
It is clear from a mathematical point of view that for finite sample spaces
we can obtain all conceivable (finite) probability spaces by assigning nonnegative numbers p 1 , . , PN, satisfying the condition p 1 + + PN = 1, to
the outcomes w 1, ... , wN.
The validity of the assignments of the numbers p 1, , PN can, in specific
cases, be checked to a certain extent by using the law of large numbers
(which will be discussed later on). It states that in a long series of "independent" experiments, carried out under identical conditions, the frequencies
with which the elementary events appear are "close" to their probabilities.
In connection with the difficulty of assigning probabilities to outcomes,
we note that there are many actual situations in which for reasons of symmetry it seems reasonable to consider all conceivable outcomes as equally
probable. In such cases, if the sample space consists of points w 1, . . , wN,
with N < oo, we put
and consequently
P(A) = N(A)/N
(10)
I. Probabilistic Model of an Experiment with a Finite Number of Outcomes
15
for every event A Ed, where N(A) is the number of sample points in A.
This is called the classical method of assigning probabilities. It is clear that
in this case the calculation of P(A) reduces to calculating the number of
outcomes belonging to A. This is usually done by combinatorial methods,
so that combinatorics, applied to finite sets, plays a significant role in the
calculus of probabilities.
7 (Coincidence problem). Let an urn contain M balls numbered
1, 2, ... , M. We draw an ordered sample of size n with replacement. It is
clear that then
ExAMPLE
Q = {w: w = (a 1 ,
.. ,
a.), a;= 1, ... , M}
and N(Q) = M". Using the classical assignment of probabilities, we consider
the M" outcomes equally probable and ask for the probability of the event
A = {w: w = (a 1 ,
a.), a; "# aj, i "# j},
i.e., the event in which there is no repetition. Clearly N(A) = M(M  1) ...
(M  n + 1), and therefore
P(A)
(z~ = (1 k) (1 k) ... (1 n ~ 1).
(11)
This problem has the following striking interpretation. Suppose that
there are n students in a class. Let us suppose that each student's birthday
is on one of 365 days and that all days are equally probable. The question
is, what is the probability P. that there are at least two students in the class
whose birthdays coincide? If we interpret selection of birthdays as selection
of balls from an urn containing 365 balls, then by (11)
P.
1
(365).
365" .
The following table lists the values of P" for some values of n:
n
16
22
23
40
64
P.
0.016
0.284
0.476
0.507
0.891
0.997
It is interesting to note that (unexpectedly!) the size of class in which there
is probability! of finding at least two students with the same birthday is not
very large: only 23.
8 (Prizes in a lottery). Consider a lottery that is run in the following
way. There are M tickets numbered 1, 2, ... , M, of which n, numbered
1, ... , n, win prizes (M ~ 2n). You buy n tickets, and ask for the probability
(P, say) of winning at least one prize.
EXAMPLE
16
I. Elementary Probability Theory
Since the order in which the tickets are drawn plays no role in the presence
or absence of winners in your purchase, we may suppose that the sample space
has the form
n=
{w: w = [a 1,
By Table 1, N(Q) =
C~.
.. ,
an], ak =I= a1, k =I= l, ai
= 1, ... , M}.
Now let
A 0 = {w: w = [a 1 ,
an], ak =I= a1, k =I= l, ai
= n + 1, ... , M}
be the event that there is no winner in the set of tickets you bought. Again
by Table 1, N(A 0 ) = C~n Therefore
P(A ) = C~n = (M n)n
(M)n
'C~
0
and consequently
P=
).
n
n) ... (1 (1  _
1 P(A = 1 (1  ~)
Mn+1
M1
M
0)
If M = n2 and n 4 oo, then P(A 0 )
4
e 1 and
P+ 1 e 1 ~ 0.632.
The convergence is quite fast: for n = 10 the probability is already P = 0.670.
6.
PROBLEMS
1. Establish the following properties of the operators n and u:
AB = BA
A uB= BuA,
A u (B u C) = (A u B) u C,
A(B u C) = AB u AC,
Au A= A,
(commutativity),
A(BC) = (AB)C
A u (BC)
(associativity),
(distributivity),
(A u B)(A u C)
(idempotency).
AA = A
Show also that
Au B
=An B,
AB
=Au B.
2. Let Q contain N elements. Show that the number d(N) of different decompositions of
Q is given by the formula
(12)
(Hint: Show that
N1
d(N)
k=O
c~ 1 d(k),
where d(O)
1,
and then verify that the series in (12) satisfies the same recurrence relation.)
17
2. Some Classical Models and Distributions
3. For any finite collection of sets
A~>
... , A.,
+ + P(A.).
P(A 1 u u A.) :0::: P(A 1 )
4. Let A and B be events. Show that AB u BA is the event in which exactly one of A
and B occurs. Moreover,
P(AB u BA) = P(A)
+ P(B) 
2P(AB).
5. Let A 1 , .. , A. be events, and define S 0 , S 1 , , s. as follows: S 0 = 1,
s, = L P(Ak, (\ ... (\ Ak),
J,
where the sum is over the unordered subsets J, = [k 1, ... , k,] of {1, ... , n}.
Let Bm be the event in which each of the events A 1 , . , A. occurs exactly m times.
Show that
n
P(Bm) =
L (1)mc~s,.
In particular, form= 0
P(B 0 ) = 1  S 1
+ S2
s.
Show also that the probability that at least m of the events A 1 , , A. occur
simultaneously is
n
P(B 1 )
+ + P(B.) = I (1ymc~_ls,.
r=m
In particular, the probability that at least one of the events A 1, . , A. occurs is
2. Some Classical Models and Distributions
1. Binomial distribution. Let a coin be tossed n times and record the results
as an ordered set {a 1 , ... , an), where a; = 1 for a head ("success") and a;= 0
for a tail ("failure"). The sample space is
n=
{w: w = (al, ... , an), a;= 0, 1}.
To each sample point w = (a 1,
p(w)
... ,
an) we assign the probability
= {la,qnr.a,,
where the nonnegative numbers p and q satisfy p + q = 1. In the first place,
we verify that this assignment of the weights p(w) is consistent. It is enough
to show that
p(w) = 1.
We consider all outcomes w = (a 1 , , a.) for which
a; = k, where
k = 0, 1, ... , n. According to Table 4 (distribution of k indistinguishable
Lwen
L;
18
I. Elementary Probability Theory
ones inn places) the number of these outcomes is
n
L p(w) = L
roen
C=pkqnk
k=O
c:. Therefore
= (p + q)n = 1.
Thus the space 0 together with the collection Jil of all its subsets and the
probabilities P(A) = LroeA p(w), A E J/1, defines a probabilistic model. It is
natural to call this the probabilistic model for n tosses of a coin.
In the case n = 1, when the sample space contains just the two points
w = 1 ("success") and w = 0 ("failure"), it is natural to call p(l) = p the
probability of success. We shall see later that this model for n tosses of a
coin can be thought of as the result of n "independent" experiments with
probability p of success at each trial.
Let us consider the events
k = 0, 1, ... , n,
consisting of exactly k successes. It follows from what we said above that
(1)
LZ=o
and
P(Ak) = 1.
The set of probabilities (P(A 0 ), , P(An)) is called the binomial distribution (the number of successes in a sample of size n). This distribution plays an
extremely important role in probability theory since it arises in the most
diverse probabilistic models. We write Pik) = P(Ak), k = 0, 1, ... , n.
Figure 1 shows the binomial distribution in the case p = ! (symmetric coin)
for n = 5, 10, 20.
We now present a different model (in essence, equivalent to the preceding
one) which describes the random walk of a "particle."
Let the particle start at the origin, and after unit time let it take a unit
step upward or downward (Figure 2).
Consequently after n steps the particle can have moved at most n units
up or n units down. It is clear that each path w of the particle is completely
specified by a set (a 1, . , an), where a; = + 1 if the particle moves up at the
ith step, and a; =  1 if it moves down. Let us assign to each path w the
weight p(w) = pv<w>qnv<w>, where v(w) is the number of+ 1's in the sequence
w = (a 1 , ... , an), i.e. v(w) = [(a 1 + + an) + n]/2, and the nonnegative
numbers p and q satisfy p + q = 1.
Since Lroen p(w) = 1, the set of probabilities p(w) together with the space
0 of paths w = (a 1, .. , an) and its subsets define an acceptable probabilistic
model of the motion of the particle for n steps.
Let us ask the following question: What is the probability of the event Ak
that after n steps the particle is at a point with ordinate k? This condition
is satisfied by those paths w for which v(ro)  (n  v(w)) = k, i.e.
v(w)
n+k
2
= .
19
2. Some Classical Models and Distributions
P.(k)
P.(k)
0.3
0.3
0.2
0.2
0.1
0.1
n = 10
= 20
012345
I1
I
012345678910
""
P.(k)
0.3
0.2
0.1
I
012345678910" . . . . . . . "20
Figure 1. Graph of the binomial probabilities P.(k) for n = 5, 10, 20.
The number of such paths (see Table 4) is
P(Ak)
c~n+kJ/ 2 ,
and therefore
c~n+k112pln+kJ12qinkl!2.
Consequently the binomial distribution (P(A_n), ... , P(A 0 ), ... , P(A"))
can be said to describe the probability distribution for the position of the
particle after n steps.
Note that in the symmetric case (p = q = !) when the probabilities of
the individual paths are equal tor",
P(Ak) = qn+kl/2. r".
Let us investigate the asymptotic behavior of these probabilities for large n.
If the number of steps is 2n, it follows from the properties of the binomial
coefficients that the largest of the probabilities P(Ak), Ik I :::; 2n, is
P(A 0 ) =
qn 2 2 ".
Figure 2
20
I. Elementary Probability Theory
4 3 2 1
Figure 3. Beginning of the binomial distribution.
From Stirling's formula (see formula (6) in Section 4)
n! "'~ e"nn.t
Consequently
en
2n
= (2n)! "'22n,_l_
fo
(n!?
and therefore for large n
P(A 0 )
.
Fn
Figure 3 represents the beginning of the binomial distribution for 2n
steps of a random walk (in contrast to Figure 2, the time axis is now directed
upward).
2. Multinomial distribution. Generalizing the preceding model, we now
suppose that the sample space is
n=
{w: w = (al, ... ' an), a;= bl, ... ' b,},
where b 1, . , b, are given numbers. Let v;(w) be the number of elements of
w = (a 1 , .. , an) that are equal to b;, i = 1, ... , r, and define the probability
ofw by
p(w) =
where P; ~ 0 and p 1
+ + Pr
p~'(w) ... p;r(w),
1. Note that
where Cn(n 1 , ... , n,) is the number of (ordered) sequences (a 1 , ... , an) in
which b 1 occurs n 1 times, ... , b, occurs n, times. Since n 1 elements b1 can
t The notation} (n) 
g(n) means that f (n)/g(n)+ 1 as n+ oo.
21
2. Some Classical Models and Distributions
be distributed into n positions in C~' ways; n2 elements b2 into n  n 1
positions in c~:._n, ways, etc., we have
(n
... 1
n 1 )!
n 1 ! (n n 1 )! n 2 ! (n n 1  n 2 )!
n!
n!
Therefore
""
L.... P(w )
wen
n P~"
n!
""
P1'
L.
{"l;;::o, ... ,nr2!0,} n 1t nr1
P1
+ + Prt =
1,
n1 + +nr=n
and consequently we have defined an acceptable method of assigning
probabilities.
Let
Then
(2)
The set of probabilities
is called the multinomial (or polynomial) distribution.
We emphasize that both this distribution and its special case, the binomial
distribution, originate from problems about sampling with replacement.
3. The multidimensional hypergeometric distribution occurs in problems that
involve sampling without replacement.
Consider, for example, an urn containing M balls numbered 1, 2, ... , M,
where M 1 balls have the color b 1 , . . . , Mr balls have the color b., and
M 1 + + Mr = M. Suppose that we draw a sample of size n < M without
replacement. The sample space is
Q
= {w: w = (a 1, . , an), ak =I a1 , k =I l, ai = 1, ... , M}
and N(Q) = (M)". Let us suppose that the sample point~ are equiprobable,
and find the probability of the event Bn,, " in which n 1 balls have
color b 1 , .. , nr balls have color b., where n 1 + + nr = n. It is easy to
show that
22
I. Elementary Probability Theory
and therefore
P{B
)=
,.,.......
N(B
)
C" C"
"" =
Mt
N(Q)
CM Mr
(3)
The set of probabilities {P(B, .... , ...)} is called the multidimensional
hypergeometric distribution. When r = 2 it is simply called the hypergeometric
distribution because its "generating function" is a hypergeometric function.
The structure of the multidimensional hypergeometric distribution is
rather complicated. For example, the probability
P(B,.,,2) =
C" C"2
~M M2'
(4)
contains nine factorials. However, it is easily established that if M + oo
and M 1 + oo in such a way that MdM+ p (and therefore M 2 /M+ 1  p)
then
(5)
In other words, under the present hypotheses the hypergeometric distribution is approximated by the binomial; this is intuitively clear since
when M and M 1 are large (but finite), sampling without replacement ought
to give almost the same result as sampling with replacement.
Let us use (4) to find the probability of picking six "lucky" numbers in a lottery of the following kind (this is an abstract formulation of the
"sportloto," which is well known in Russia):
There are 49 balls numbered from 1 to 49; six of them are lucky (colored
red, say, whereas the rest are white). We draw a sample of six balls, without
replacement. The question is, What is the probability that all six of these
balls are lucky? Taking M = 49, M 1 = 6, n 1 = 6, n2 = 0, we see that the
event of interest, namely
ExAMPLE.
B 6 , 0 = {6 balls, all lucky}
has, by (4), probability
1
P(B6,o) = c6 ~ 7.2
49
10 8
4. The numbers n! increase extremely rapidly with n. For example,
10!
3,628,800,
15! = 1,307,674,368,000,
and 100! has 158 digits. Hence from either the theoretical or the computational point of view, it is important to know Stirling's formula,
n! =
~ (;)" exp( 1~n),
o < e, < 1,
(6)
23
3. Conditional Probability. Independence
whose proof can be found in most textbooks on mathematical analysis
(see also [69]).
5. PROBLEMS
1. Prove formula (5).
2. Show that for the multinomial distribution {P(A.,, ... , A.J} the maximum probability is attained at a point (k 1 , , k.) that satisfies the inequalities np;  1 <
k; :::;; (n + r  1)p;, i = 1, ... , r.
3. Onedimensional Ising model. Consider n particles located at the points 1, 2, ... , n.
Suppose that each particle is of one of two types, and that there are n 1 particles ofthe
first type and n 2 of the second (n 1 + n 2 = n). We suppose that all n! arrangements of
the particles are equally probable.
Construct a corresponding probabilistic model and find the probability of the
event A.(m 11 , m 12 , m 21 , m22 ) = {v 11 = m11 , , v22 = m22 }, where vii is the number
of particles of type i following particles of type j (i, j = 1, 2).
4. Prove the following inequalities by probabilistic reasoning:
L (C!)
k=O
n
L(1)"kC! =
cz.
c;:._ 1,
m ;:o: n
m(m 1)2mz,
m ;:o: 2.
+ 1,
k=O
n
L k(k
l)C~ =
k=O
3. Conditional Probability. Independence
1. The concept of probabilities of events lets us answer questions of the following kind: If there are M balls in an urn, M 1 white and M 2 black, what is
the probability P(A) of the event A that a selected ball is white? With the
classical approach, P(A) = M tfM.
The concept of conditional probability, which will be introduced below,
lets us answer questions of the following kind: What is the probability that
the second ball is white (event B) under the condition that the first ball was
also white (event A)? (We are thinking of sampling without replacement.)
It is natural to reason as follows: if the first ball is white, then at the
second step we have an urn containing M  1 balls, of which M 1  1 are
white and M 2 black; hence it seems reasonable to suppose that the (conditional) probability in question is (M 1  1)/(M 1).
24
I. Elementary Probability Theory
We now give a definition of conditional probability that is consistent
with our intuitive ideas.
Let (0, d, P) be a (finite) probability space and A an event (i.e. A Ed).
Definition 1. The conditional probability of event B assuming event A with
P(A) > 0 (denoted by P(BIA)) is
P(AB)
P(A) .
(1)
In the classical approach we have P(A) = N(A)/N(D.), P(AB) =
N(AB)/N(D.), and therefore
P(BIA) = N(AB).
N(A)
(2)
From Definition 1 we immediately get the following properties of conditional probability:
P(AIA)
= 1,
P(0IA) = 0,
P(BIA) = 1,
B2A,
It follows from these properties that for a given set A the conditional
probability P( I A) has the same properties on the space (D. n A, dnA),
where dnA= {B n A: BEd}, that the original probability PO has on
(D., d).
Note that
+ P(BIA) =
1;
+ P(BIA) =F
P(BIA) + P(BIA) =F
1,
P(BIA)
however in general
P(BIA)
1.
ExAMPLE 1. Consider a family with two children. We ask for the probability
that both children are boys, assuming
(a) that the older child is a boy;
(b) that at least one of the children is a boy.
25
3. Conditional Probability. Independence
The sample space is
Q
{BB, BG, GB, GG},
where BG means that the older child is a boy and the younger is a girl, etc.
Let us suppose that all sample points are equally probable:
P(BB) = P(BG) = P(GB) = P(GG) =
Let A be the event that the older child is a boy, and B, that the younger
child is a boy. Then A u B is the event that at least one child is a boy, and
AB is the event that both children are boys. In question (a) we want the
conditional probability P(AB IA), and in (b), the conditional probability
P(ABIA u B).
It is easy to see that
P(ABIA) = P(AB) =! =!
P(A)
2'
P(AB)
t 1
P(ABIA u B)= P(A u B)=!= "3
2. The simple but important formula (3), below, is called the formula for
total probability. It provides the basic means for calculating the probabilities of complicated events by using conditional probabilities.
Consider a decomposition~= {A 1, , An} with P(A;) > 0, i = 1, ... , n
(such a decomposition is often called a complete set of disjoint events). It
is clear that
B
+ +
BA 1
BAn
and therefore
n
P(B) =
P(BA;).
i= 1
But
P(BA;)
= P(BIA;)P(A;).
Hence we have the formula for total probability:
n
P(B)
= L
P(B I A;)P(A;).
(3)
i= 1
In particular, if 0 < P(A) < 1, then
P(B) = P(BIA)P(A)
+ P(BIA)P(A).
(4)
26
I. Elementary Probability Theory
2. An urn contains M balls, m of which are "lucky." We ask for the
probability that the second ball drawn is lucky (assuming that the result of
the first draw is unknown, that a sample of size 2 is drawn without replacement, and that all outcomes are equally probable). Let A be the event that
the first ball is lucky, B the event that the second is lucky. Then
ExAMPLE
m(m 1)
P(BIA) = P(BA) = M(M 1) = m 1
P(A)
m
M 1'
m(M m)
_
P(BA)
M(M 1)
m
P(B IA) = P(A) = M  m = M  1
and
P(B) = P(BIA)P(A)
+ P(BIA)P(A)
m1 m
m
Mm
=M1M+M1. M
m
M"
It is interesting to observe that P(A) is precisely mjM. Hence, when the
nature of the first ball is unknown, it does not affect the probability that the
second ball is lucky.
By the definition of conditional probability (with P(A) > 0),
P(AB) = P(BIA)P(A).
(5)
This formula, the multiplication formula for probabilities, can be generalized
(by induction) as follows: IfA 1, ... , An_ 1 are events with P(A 1 An_ 1) > 0,
then
P(A1 An)
(here A 1 An
= P(A1)P(A2I A1) P(An IA1 An1)
(6)
A 1 n A 2 n nAn)
3. Suppose that A and B are.events with P(A) > 0 and P(B) > 0. Then
along with (5) we have the parallel formula
P(AB) = P(A IB)P(B).
(7)
From (5) and (7) we obtain Bayes's formula
P(AIB)
= P(A)P(BIA)
P(B)
(8)
27
3. Conditional Probability. Independence
If the events A 1,
Bayes's theorem:
... ,
An form a decomposition of
n,
(3) and (8) imply
P(A;)P(B IA;)
(9)
P(A;IB) = LJ=l P(Aj)P(BIA).
In statistical applications, A 1, , An (A 1 + +An = Q) are often
called hypotheses, and P(Ai) is called the a priorit probability of Ai. The
conditional probability P(A; IB) is considered as the a posteriori probability
of A; after the occurrence of event B.
ExAMPLE
3. Let an urn contain two coins: A 1, a fair coin with probability
! of falling H; and A 2 , a biased coin with probability! of falling H. A coin is
drawn at random and tossed. Suppose that it falls head. We ask for the
probability that the fair coin was selected.
Let us construct the corresponding probabilistic model. Here it is natural
to take the sample space to be the set n = {A 1 H, A 1 T, A 2 H, A 2 T}, which
describes all possible outcomes of a selection and a toss (A 1 H means that
coin A 1 was selected and fell heads, etc.) The probabilities p(w) of the various
outcomes have to be assigned so that, according to the statement of the
problem,
P(Ad
= P(A 2 ) =!
and
P(HIA 2 )
=!.
With these assignments, the probabilities of the sample points are uniquely
determined :
P(A 2 T)
=!.
Then by Bayes's formula the probability in question is
P(A 1 )P(HIA 1 )
P(AliH) = P(Al)P(HIAl) + P(A2)P(HIA2)
S'
and therefore
P(A 2 1H)
4. In certain sense, the concept of independence, which we are now going to
introduce, plays a central role in probability theory: it is precisely this concept
that distinguishes probability theory from the general theory of measure
spaces.
t
A priori: before the experiment; a posteriori: after the experiment.
28
I. Elementary Probability Theory
If A and B are two events, it is natural to say that B is independent of A
if knowing that A has occurred has no effect on the probability of B. In other
words, "B is independent of A" if
P(BIA) = P(B)
(10)
(we are supposing that P(A) > 0).
Since
P(BIA) =
P(AB)
P(A) ,
it follows from (10) that
P(AB) = P(A)P(B).
(11)
In exactly the same way, if P(B) > 0 it is natural to say that" A is independent
of B" if
P(A IB) = P(A).
Hence we again obtain (11 ), which is symmetric in A and B and still makes
sense when the probabilities of these events are zero.
After these preliminaries, we introduce the following definition.
Definition 2. Events A and Bare called independent or statistically independent
(with respect to the probability P) if
P(AB)
= P(A)P(B).
In probability theory it is often convenient to consider not only independence of events (or sets) but also independence of collections of events (or
sets).
Accordingly, we introduce the following definition.
Definition 3. Two algebras d 1 and d 2 of events (or sets) are called independent or statistically independent (with respect to the probability P) if all pairs
of sets A 1 and A 2 , belonging respectively to d 1 and d 2 , are independent.
For example, let us consider the two algebras
where A 1 and A 2 are subsets of n. It is easy to verify that d 1 and d 2 are
independent if and only if A 1 and A 2 are independent. In fact, the independence of .91 1 and .91 2 means the independence of the 16 events A 1 and A 2 ,
A 1 and A2 , . , nand n. Consequently A 1 and A 2 are independent. Conversely, if A 1 and A 2 are independent, we have to show that the other 15
29
3. Conditional Probability. Independence
pairs of events are independent. Let us verify, for example, the independence
of A 1 and A2 We have
P(A 1 A 2 )
= P(A 1 )
= P(A 1 )
P(A 1 A 2 )
P(A 1 )P(A 2 )
= P(A 1) (1  P(A 2 )) = P(A 1 )P(A 2 ).
The independence of the other pairs is verified similarly.
5. The concept of independence of two sets or two algebras of sets can be
extended to any finite number of sets or algebras of sets.
Thus we say that the sets A 1, ... , An are collectively independent or
statistically independent (with respect to the probability P) if fork = 1, ... , n
and 1 ~ i 1 < i 2 < < ik ~ n
= P(A;,) P(A;J
P(A;, A;)
(12)
The algebras d 1 , ... , dn of sets are called independent or statistically
independent (with respect to the probability P) if all sets A 1 , .. , An belonging
respectively to d 1 , , dn are independent.
Note that pairwise independence of events does not imply their independence. In fact if, for example, n = {w 1 , w 2 , w 3 , w4 } and all outcomes are
equiprobable, it is easily verified that the events
are pairwise independent, whereas
P(ABC) =
i= (t) 3
= P(A)P(B)P(C).
Also note that if
P(ABC)
P(A)P(B)P(C)
for events A, B and C, it by no means follows that these events are pairwise
independent. In fact, let n consist of the 36 ordered pairs (i, j), where i, j =
1, 2, ... , 6 and all the pairs are equiprobable. Then if A = {(i,j):j = 1, 2 or 5},
B = {(i,j):j = 4, 5 or 6}, C = {(i,j): i + j = 9} we have
P(AB) =
i i= i =
P(A)P(B),
P(AC)
l6
i= / 8
= P(A)P(C),
P(BC)
= /2
i= / 8
P(B)P(C),
but also
P(ABC)
l6 =
P(A)P(B)P(C).
6. Let us consider in more detail, from the point of view of independence,
the classical model (Q, d, P) that was introduced in 2 and used as a basis
for the binomial distribution.
30
I. Elementary Probability Theory
In this model
n=
{w:
((J
= (a1, ... ' a.), a; =
d ={A: As Q}
0, 1},
and
(13)
Consider an event A s n. We say that this event depends on a trial at
time k if it is determined by the value ak alone. Examples of such events are
Let us consider the sequence of algebras d 1, d 2 , , d., where dk =
{Ak, Ab 0, Q} and show that under (13) these algebras are independent.
It is clear that
=p
(at, ... , ak1, Uk+
1, ... ,
an)
X q<n1)(a,++akt+ak+t++an)
= p
n1
c~1plq(n1)l
= p,
i=O
and a similar calculation shows that P(Ak) = q and that, for k # 1,
2
P(AkA 1) = p ,
P(AkA 1) = pq,
P(AkAt) = q .
It is easy to deduce from this that d k and d 1 are independent for k # I.
It can be shown in the same way that d 1 , d 2 , . , d. are independent.
This is the basis for saying that our model (Q, d, P) corresponds to "n
independent trials with two outcomes and probability p of success." James
Bernoulli was the first to study this model systematically, and established
the law of large numbers (5) for it. Accordingly, this model is also called
the Bernoulli scheme with two outcomes (success and failure) and probability
p of success.
A detailed study of the probability space for the Bernoulli scheme shows
that it has the structure of a direct product of probability spaces, defined
as follows.
Suppose that we are given a collection (0 1 , 86 1 , P 1), ... , (n., 86., P.) of
finite probability spaces. Form the space n = n1 X n2 X ... X n. of points
((J = (a1, ... ' a.), where a; E ni. Let d
= 861 ... 86. be the algebra of
the subsets of n that consists of sums of sets of the form
A= B1
B2
X ... X
B.
with B;E86;. Finally, for w = (a 1, ... ,a.) take p(w) = p 1(a 1)p.(a.) and
define P(A) for the set A = B 1 x B 2 x x B. by
P(A) =
31
3. Conditional Probability. Independence
It is easy to verify that P(O) = 1 and therefore the triple (0, d, P) defines
a probability space. This space is called the direct product of the probability
spaces (0 1 , fJI 1 , P 1 ), . , (O., fJI., P.).
We note an easily verified property of the direct product of probability
spaces: with respect to P, the events
A 1 = {ro: a 1 E Bd,
where Bi e
fJii,
... , A. =
{ro: a. E B.},
are independent. In the same way, the algebras of subsets ofO,
d
= {A 1 :A 1 = {ro:a 1 eBd,B 1 efJI 1 },
are independent.
It is clear from our construction that the Bernoulli scheme
(0, d, P) with 0 = {ro: w = (a 1,
d ={A: As; 0}
.. ,
a.), ai = 0 or 1}
p(ro) = pr,a,qnr.a,
and
can be thought of as the direct product of the probability spaces (Oi,
i = 1, 2, ... , n, where
ni = {O, 1},
Pl{1})
fJii
fJii,
Pi),
= {{O}, {1}, 0, OJ,
= p,
Pi({O})
= q.
7.PROBLEMS
1. Give examples to show that in general the equations
+ P(B!A) =
P(B!A) + P(B!A) =
P(B!A)
1,
1
are false.
2. An urn contains M balls, of which M 1 are white. Consider a sample of size n. Let Bi
be the event that the ball selected at the jth step is white, and Ak the event that a sample
of size n contains exactly k white balls. Show that
P(Bi!Ak) = k/n
both for sampling with replacement and for sampling without replacement.
3. Let A 1, ... , A. be independent events. Then
4. Let A~> ... , A. be independent events with P(A;)
that neither event occurs is
Po=
fl(l pJ
i=l
= P;
Then the probability P0
32
I. Elementary Probability Theory
5. Let A and B be independent events. In terms of P(A) and P(B), find the probabilities
of the events that exactly k, at least k, and at most k of A and B occur (k = 0, 1, 2).
6. Let event A be independent of itself, i.e. Jet A and A be independent. Show that
P(A) is either 0 or 1.
7. Let event A have P(A) = 0 or 1. Show that A and an arbitrary event B are independent.
8. Consider the electric circuit shown in Figure 4:
Figure 4
Each of the switches A, B, C, D, and E is independently open or closed with
probabilities p and q, respectively. Find the probability that a signal fed in at "input"
will be received at "output". If the signal is received, what is the conditional probability that E is open?
4. Random Variables and Their Properties
1. Let (Q, d, P) be a probabilistic model of an experiment with a finite
number of outcomes, N(Q) < oo, where d is the algebra of all subsets of
n. We observe that in the examples above, where we calculated the probabilities of various events A E d, the specific nature of the sample space n was
of no interest. We were interested only in numerical properties depending
on the sample points. For example, we were interested in the probability of
some number of successes in a series of n trials, in the probability distribution
for the number of objects in cells, etc.
The concept "random variable," which we now introduce (later it will
be given a more general form) serves to define quantities that are subject to
"measurement" in random experiments.
Definition 1. Any numerical function ~ = ~(w) defined on a (finite) sample
space n is called a (simple) random variable. (The reason for the term" simple"
random variable will become clear after the introduction of the general
concept of random variable in 4 of Chapter II.)
33
4. Random Variables and Their Properties
1. In the model of two tosses of a coin with sample space Q =
{HH, HT, TH, TT}, define a random variable~ = ~(w) by the table
ExAMPLE
OJ
HH
HT
TH
TT
(w)
Here, from its very definition, ~(w) is nothing but the number of heads in the
outcome w.
Another extremely simple example of a random variable is the indicator
(or characteristic function) of a set A E .s:1:
wheret
IA(w)
1, WEA,
= { 0, wA.
When experimenters are concerned with random variables that describe
observations, their main interest is in the probabilities with which the
random variables take various values. From this point of view they are
interested, not in the distribution of the probability P over (Q, d), but in
its distribution over the range of a random variable. Since we are considering
the case when Q contains only a finite number of points, the range X of
the random variable (is also finite. Let X = {x 1 , . . . , xm}, where the (different) numbers X 1, ... , Xm exhaust the values of
Let fi be the collection of all subsets of X, and let BE fi. We can also
interpret B as an event if the sample space is taken to be X, the set of values
of
On (X, fi), consider the probability P~() induced bye according to the
formula
e.
e.
P~(B)
= P{w: ~(w) E B},
BE fi.
It is clear that the values of this probability are completely determined by
the probabilities
P~(x;)
= P{w: ~(w) =X;},
The set of numbers {P~(x 1 },
bution of the random variable ~.
t The notation
... ,
P~(xm)}
X;
EX.
is called the probability distri
/(A) is also used. For frequently used properties of indicators see Problem 1.
34
I. Elementary Probability Theory
2. A random variable that takes the two values 1 and 0 with
probabilities p ("success") and q ("failure"), is called a Bernoullit random
variable. Clearly
EXAMPLE
0, 1.
X=
(1)
A binomial (or binomially distributed) random variable is a random
variable that takes the n + 1 values 0, 1, ... , n with probabilities
x = 0, 1, ... , n.
(2)
Note that here and in many subsequent examples we do not specify the
sample spaces (Q, d, P), but are interested only in the values of the random
variables and their probability distributions.
is completely
The probabilistic structure of the random variables
specified by the probability distributions {P~(xi), i = 1, ... , m}. The concept
of distribution function, which we now introduce, yields an equivalent
description of the probabilistic structure of the random variables.
Definition 2. Let x e R 1 The function
F~(x) =
P{w: e(w)
x}
is called the distribution function of the random variable
e.
Clearly
F~(x) =
{i: Xi
P~(xi)
$X}
and
P ~(xi) = F ~(xi)  F f.._x;  ),
where F ~(x) = limyix F ~(y).
If we suppose that x 1 < x 2 < < xm and put F ~(x 0 ) = 0, then
i = 1, ... , m.
The following diagrams (Figure 5) exhibit P~(x) and F~(x) for a binomial
random variable.
It follows immediately from Definition 2 that the distribution F~ = F~(x)
has the following properties:
(1) F ~( oo)
= 0, F ~( + oo) = 1;
(2) F ~(x) is continuous on the right (F ~(x+) = F ~(x)) and piecewise constant.
t We use the terms "Bernoulli, binomial, Poisson, Gaussian, ... , random variables" for what
are more usually called random variables with Bernoulli, binomial, Poisson, Gaussian, ... , distributions.
35
4. Random Variables and Their Properties
.....
q",_/_/tt~~~==+.
0
'f_: _;,.
F<(x)
I
I
~~
0
,.
Figure 5
Along with random variables it is often necessary to consider random
~ = (~ 1 , ... , ~,) whose components are random variables. For
example, when we considered the multinomial distribution we were dealing
with a random vector v = (v 1, ... , v,), where v; = v;(w) is the number of
elements equal to b;, i = 1, ... , r, in the sequence w = (a 1, .. , an).
The set of probabilities
vectors
P~(x 1 , ... ,
x,)
= P{w: ~ 1 (w) = x 1 ,
... , ~,(w)
= x,},
where X; EX;, the range of ~;, is called the probability distribution of the
random vector ~, and the function
where
X; E
R 1, is called the distribution function of the random vector ~
(~1 ... ' ~,).
For example, for the random vector v = (v 1 ,
... ,
v,) mentioned above,
(see (2.2)).
2. Let ~ 1 , . . . , ~r be a set of random variables with values in a (finite) set
X s; R 1 . Let X be the algebra of subsets of X.
36
I. Elementary Probability Theory
Definition 3. The random variables
(collectively independent) if
P{~ 1 =
for all x 1,
x 1 , ,
xd P{~.
~. = x,} = P{~ 1 =
= x,}
x, EX; or, equivalently, if
P{~1EB1,
for all B 1,
are said to be independent
~ 1 , .. , ~.
... ,~,EB,} =
P{~1EBdP{~,EB,}
B, E f!l".
We can get a very simple example of independent random variables
from the Bernoulli scheme. Let
n=
{w: (I)= (a1, ... ' an), a;= 0, 1},
p(w) = pr.a;qnr.a;
and ~;(w) = a; for w = (a 1, ... , an), i = 1, ... , n. Then the random variables
~ 1 , ~ 2 , , ~n are independent, as follows from the independence of the events
A 1 = {w: a 1 = 1}, ... , An= {w: an= 1},
which was established in 3.
3. We shall frequently encounter the problem of finding the probability
distributions of random variables that are functionsf(~ 1 , .. , ~.)of random
variables ~ 1 , , ~. For the present we consider only the determination
of the distribution of a sum ( = ~ + 11 of random variables.
If~ and '1 take values in the respective sets X = {x 1, ... , xk} and Y =
{y 1 , .. , y 1}, the random variable ( = ~ + '1 takes values in the set Z =
{z: z = X; + Yi i = 1, ... , k;j = 1, ... , /}.Then it is clear that
P,(z) = P{( = z} = Pg
+ 11
= z} =
P{~ =X;, '1 = Yi}.
{(i, j):x;+yJ=z)
The case of independent random variables
ant. In this case
and '1 is particularly import
and therefore
P,(z)
P~(x;)P~(yj)
{(i,j):x;+yj=z}
L P~(x;)P~(z x;)
i=1
(3)
for all z E Z, where in the last sum Pq(z  X;) is taken to be zero if z  X;~ Y.
For example, if ~ and '1 are independent Bernoulli random variables,
taking the values 1 and 0 with respective probabilities p and q, then Z =
{0, 1, 2} and
P,(O)
P,(l)
= P~(O)Pq(O) = q2 ,
= P~(O)Pq(i) + Pp)Pq(O) = 2pq,
P,(2) = P~(l)Pq(1) = p2
37
4. Random Variables and Their Properties
It is easy to show by induction that if ~ 1 ~ 2 , , ~n are independent
Bernoulli random variables with P{~; = 1} = p, P{~; = 0} = q, then the
random variable ( = ~ 1 + + ~n has the binomial distribution
k = 0, 1, ... , n.
(4)
4. We now turn to the important concept of the expectation, or mean value,
of a random variable.
Let (Q, d, P) be a (finite) probability space and ~ = ~(w) a random
variable with values in the set X= {x 1 , , xk}. If we put A;= {w: ~ = x;},
i = 1, ... , k, then ~ can evidently be represented as
k
~(w) =
X;l(A;),
(5)
i= 1
where the sets A 1, , Ak form a decomposition of Q (i.e., they are pairwise
disjoint and their sum is Q; see Subsection 3 of 1).
Let P; = P{ ~ = x;}. It is intuitively plausible that if we observe the values
of the random variable ~ in "n repetitions of identical experiments", the
value X; ought to be encountered about p;n times, i = 1, ... , k. Hence the
mean value calculated from the results of n experiments is roughly
1
 [np 1 x 1
+ + npkxk] =
I P;X;.
i=1
This discussion provides the motivation for the following definition.
Definition 4. The expectation t or mean value of the random variable
~ =
L~= 1 xJ(A;) is the number
k
E~ =
X;P(AJ
(6)
i= 1
Since A;= {w:
~(w) =X;}
and
P~(x;) =
P(A;), we have
E~ =
Recalling the definition
i= 1
x;Pix;).
ofF~= F~(x)
(7)
and writing
dF ix) = F~(x) F~(x ),
we obtain P ~(x;)
= dF ~(x;) and consequently
k
E~ =
x;dF~(x;).
(8)
i= 1
t Also known as mathematical expectation, or expected value, or (especially in physics) expectation value. (Translator)
38
I. Elementary Probability Theory
Before discussing the properties of the expectation, we remark that it is
often convenient to use another representation of the random variable ~'
namely
l
~(w) =
L xji(B),
j= 1
where B 1 + + B1 = n, but some of the xj may be repeated. In this case
E~ can be calculated from the formula L~= 1 xjP(B), which differs formally
from (5) because in (5) the X; are all different. In fact,
xjP(B)
=X;
(j: xj = x;)
P(B)
x;P(A;)
(j: xj = x;)
and therefore
l
i= 1
xjP(B)
i= 1
x;P(AJ
5. We list the basic properties of the expectation:
(1) If~ 2': 0 then E~ 2': 0.
(2) E(a~ + b1]) = aE~ + bE1], where a and bare constants.
(3) If~ 2': '7 then E~ 2': El'f.
(4) IE~I ~ El~l(5) If~ and '1 are independent, then E~l] = E~ El].
e E17
( 6) (E I~171) 2 ~ E
(7)
If~
(Cauchy Bun yakovskii inequality). t
= I(A) then E~ = P(A).
Properties (1) and (7) are evident. To prove (2), let
~
= IxJ(AJ,
I]=
IYi(B).
j
Then
i, j
i,j
L(ax; + byj)I(A; n Bi)
i,j
and
E(a~
+ bl]) = I (ax; + byi)P(A; n Bi)
i,j
= Lax; P(A;)
i
+ Lbyi P(Bi)
j
= al;x;P(A;)
i
+ b LYiP(Bi) =
j
aE~
+ bE1].
t Also known as the CauchySchwarz or Schwarz inequality. (Translator)
39
4. Random Variables and Their Properties
Property (3) follows from (1) and (2). Property (4) is evident, since
~~x;P(A;)I ;S; ~lxdP(A;) = Elel.
!Eel=
To prove (5) we note that
Ee11 =
E(~x;I(A;)) (~>il(Bi))
= E
L X;Yi(A; n B)= L X;YiP(A; n Bi)
i,j
i,j
L X;yjP(A;)P(Bj)
i,j
where we have used the property that for independent random variables the
events
A;= {w: e(w) = x;}
Bi = {w: 17(w) = Yi}
and
are independent: P(A; n Bi) = P(A;)P(Bi).
To prove property (6) we observe that
L xf l(A;),
'7 2
Lx~P(A;),
E'7 2 = LYJP(Bj).
LYJI(Bj)
j
and
Ee 2 =
Let Ee 2 > 0, E77 2 > 0. Put
e=~,
'7
= '7 .
for
Since 21~~1
+
we have 2EI~~I $ E~ 2 + E~ 2 = 2. Therefore
El~~~ $ 1 and (Eie'71) 2 $ Ee 2 E77 2
However, if, say, Ee 2 = 0, this means that
xfP(A;) = 0 and consequently the mean value of e is 0, and P{w: e(w) = 0} = 1. Therefore if at
least one of Ee2 or E77 2 is zero, it is evident that EIe111 = 0 and consequently
the CauchyBunyakovskii inequality still holds.
$ ~2
~2,
L;
Remark. Property (5) generalizes in an obvious way to any finite number of
random variables: if
el, ... ' e, are independent, then
The proof can be given in the same way as for the case r = 2, or by induction.
40
I. Elementary Probability Theory
EXAMPLE 3. Let be a Bernoulli random variable, taking the values 1 and 0
with probabilities p and q. Then
ee = 1 Pg = 1}
+ o. P{e = o}
= p.
4. Let e1, ... , en ben Bernoulli random variables with P{e; = 1}
P{e; = O} = q, p + q = 1. Then if
ExAMPLE
= p,
we find that
ESn = np.
This result can be obtained in a different way. It is easy to see that ES"
is not changed if we assume that the Bernoulli random variables 1, , en
are independent. With this assumption, we have according to (4)
k = 0, 1, ... , n.
Therefore
ESn =
kP(Sn = k) =
k=O
kC~pkqnk
k=O
n!
knk
= k~okk!(nk)!pq
= np
k= 1
(n 1)!
pk1q<n1)(k1)
(k 1)! ((n 1) (k 1))!
.
_
~
(n1)!
l(n1)1_
 np 1~0 l!((n 1) 1)! pq
 np.
However, the first method is more direct.
Li
xJ(A;), where A; = {ro: e(ro) =X;}, and
6. Let e =
function of e(w). If Bi = {ro: cp(e(w)) = yj}, then
cp(e(w))
qJ =
cp(e(w)) is a
= LYi(Bj),
j
and consequently
Ecp
LYiP(B) = LYiP'P(yj).
j
But it is also clear that
cp(e(w)) =
cp(xJI(A;).
(9)
41
4. Random Variables and Their Properties
Hence, as in (9), the expectation of the random variable <p =
calculated as
E<p( ~)
<p(~)
can be
= L <p(x;)P ~(x;).
i
7. The important notion of the variance of a random variable ~ indicates
the amount of scatter of the values
around E~.
of~
Definition 5. The variance (also called the dispersion) of the random variable
(denoted by V ~) is
V~ = E(~ E~) 2 .
The number
Since
CJ
= + jVf, is called the standard deviation.
we have
Clearly V~
V(a
0. It follows from the definition that
+ b~) =
where a and bare constants.
b 2 V~,
In particular, Va = 0, V(b~) = b 2 V~.
Let ~ and 11 be random variables. Then
V(~
+ rJ) =
=
E((~ E~)
Vi;+ V17
+ ('1
E1J)) 2
2E(( E()(Y/ E17).
Write
cov(~,
'1) =
E(~ E~)(IJ
 ErJ).
This number is called the covariance of~ and '1 If V ~ > 0 and V11 > 0, then
( ~ ) = cov(~, '1)
JV~VIJ
p ,IJ
is called the correlation coefficient of~ and '1 It is easy to show (see Problem
7 below) that if p(~, rJ) = 1, then~ and 11 are linearly dependent:
11 =a~+ b,
with a > 0 if p(~, rJ) = 1 and a < 0 if p(~, rJ) = 1.
We observe immediately that if~ and 11 are independent, so are
and rJ  ErJ. Consequently by Property (5) of expectations,
cov(~,
rJ) =
E(~
 EO E(IJ  ErJ)
= 0.
E~
42
I. Elementary Probability Theory
Using the notation that we introduced for covariance, we have
V(~
+ Yf)
= V~
+ VYf +
(10)
2cov(~, Yf);
if~ and Yf are independent, the variance of the sum
of the variances,
V(~ + Y/) = V~ + VYf.
+ Yf is equal to the sum
(11)
It follows from (10) that (11) is still valid under weaker hypotheses than
the independence of~ and Yf In fact, it is enough to suppose that ~ and Yf are
uncorrelated, i.e. cov(~, ry) = 0.
Remark. If ~ and Yf are uncorrelated, it does not follow in general that they
are independent. Here is a simple example. Let the random variable o: take
the values 0, n/2 and n with probability t. Then ~ = sin o: and Y/ = cos o: are
uncorrelated; however, they are not only stochastically dependent (i.e., not
independent with respect to the probability P):
P{~ = 1, Yf = 1} = 0 ~ = P{~ = 1}P{ry = 1},
but even functionally dependent: ~ 2 + Yf 2 = 1.
Properties (10) and (11) can be extended in the obvious way to any number of random variables:
(12)
In particular, if ,;; 1 ,
. ,
.;;n are pairwise independent (pairwise uncorrelated
is sufficient), then
(13)
5. If ~ is a Bernoulli random variable, taking the values 1 and 0
with probabilities p and q, then
ExAMPLE
V~
= E(~ E~)2 = (~ p)2 = (1  p)2 p
+ p2q =
pq.
It follows that if~ 1 ,
... , ~n are independent identically distributed Bernoulli
random variables, and sn = ~ 1 + ... + ~n, then
vsn =
npq.
(14)
8. Consider two random variables ~ and Yf. Suppose that only ~ can be observed. If~ and Yf are correlated, we may expect that knowing the value of~
allows us to make some inference about the values of the unobserved variable Yf
Any function}= f(~) of~ is called an estimator for Yf We say that an estimator f* = f*(~) is best in the meansquare sense if
E(Yf  /*(0) 2 = inf E(Yf  /(~)) 2
f
43
4. Random Variables and Their Properties
Let us show how to find a best estimator in the class of linear estimators
a + b~. We consider the function g(a, b) = E(17  (a + b~)) 2 . Differentiating g(a, b) with respect to a and b, we obtain
A.(~) =
og(a, b)=  2E['7 (a+
oa
og(a, b)
ob
=  2E[('7
b~)],
(a+ be))e],
whence, setting the derivatives equal to zero, we find that the best meansquare linear estimator is ..t*(~) = a* + b*~. where
a*
= E17  b*E~,
b*
= cov(~, 17)
v~
(15)
In other words,
(16)
The number E(17  ..1.*(~)) 2 is called the meansquare error of observation.
An easy calculation shows that it is equal to
Consequently, the larger (in absolute value) the correlation coefficient
17) between ~ and '1 the smaller the meansquare error of observation
tl*. In particular, if IP(e. '7)1 = 1 then tl* = 0 (cf. Problem 7). On the other
hand, if~ and '1 are uncorrelated (p(~, '7) = 0}, then A.*(~) = E'7, i.e. in the
absence of correlation between ~ and '1 the best estimate of '1 in terms of~ is
simply E17 (cf. Problem 4).
p(~.
9. PROBLEMS
1. Verify the following properties of indicators IA = IA(w):
JAB= IA .JB,
]AuB =]A+ ]B JAB
The indicator of Uf=t Ai is 1 n~=l (1 lA,), the indicator of Ui=1 Ai is
Df= 1 (1  I A,), and the indicator o(D= 1 Ai is D= 1 I A,.
where A 6. B is the symmetric difference of A and B, i.e. the set (A\B) u (B\A).
44
I. Elementary Probability Theory
2. Let ~ 1 , ,
~.be
independent random variables and
~min= min(~ 1 ,
.. ,
~max= max(~! ... ' ~.).
~.),
Show that
n
P{~min;;:: x} = TIP{~;;;:: x},
i=l
< x} = TIPg; < x}.
P{~max
i=l
3. Let ~ ~> ... , ~. be independent Bernoulli random variables such that
0}
P{~; =
P{~;
where ~ is a small
Show that
number,~
= 1} =A;~,
> 0, A; > 0.
P{~~ + + ~. =
P{~ 1
1 A;~,
+ + ~.
(J/;)~ + 0(~ 2),
1} =
> 1}
= 0(~ 2 ).
4. Show that inL 00 <a< 00 E(~  a) 2 is attained for a = E~ and consequently
inf
E(~  a) 2 = V~.
oo<a<oo
5.
Let~ be a random variable with distribution function
of F~(x), i.e. a point such that
F~(x)
and let me be a median
Show that
inf
El~
ro<a<oo
6. Let
P~(x)
P{~
= x} and
F~(x)
al
P(~::;;
= El~
m.,l.
x}. Show that
X 
b)
Pa~+b(x) = p~ ( a '
X
b)
Fa~+b(x) = F~ ( a
for a > 0 and  oo < b < oo. If y ;;:: 0, then
F~z(y)
Let~+ = max(~,
= F~
+Jy) F ~ JY) + P~( JY).
0). Then
X< 0,
0,
X> 0.
X=
45
5. The Bernoulli Scheme. I. The Law of Large Numbers
7.
Let~ and 'I be random variables with V~ > 0, Vf1 > 0, and let p = p(~, 'I) be their
correlation coefficient. Show that IpI ::;; 1. If IpI = 1, find constants a and b such
that 'I = a~ + b. Moreover, if p = 1, then
(and therefore a > 0), whereas if p
1, then
(and therefore a < 0).
8.
Let~ and 'I be random variables with
coefficient p = p(~, f/). Show that
E~
= Ef/ = 0,
E max<e, f/ 2 )::;; 1 +
9. Use the equation
(IndicatorofiV1
to deduce the formula P(B 0 )
=1
S1
A;)
V~
= Vf/ = 1 and correlation
Jl=Pl.
=bY
IA),
+ S 2 + s. from Problem 4 of 1.
10. Let ~~> ... , ~. be independent random variables, cp 1 = cp 1 (~ 1 , . , ~k) and cp 2 =
cp 2 ( ~k+ ~> ... , ~.), functions respectively of~ 1 , . , ~k and ~k+ 1 , ... , ~ . Show that the
random variables cp 1 and cp 2 are independent.
11. Show that the random variables~ 1 , ... ,
F~, ..... ~Jx 1 ,
~.are
x.)
independent if and only if
= F~,(x 1 )
for all Xt. . , x., where F~, ..... ~Jx 1 , .. , x.) =
P{~ 1
F~"(x.)
::;;
x 1, ... , ~.::;; x.}.
12. Show that the random variable ~ is independent of itself (i.e.,
pendent) if and only if~ = const.
13. Under what hypotheses
~independent?
14.
on~
are the random
variables~
and sin
and
are inde
and 'I be independent random variables and 'I # 0. Express the probabilities
of the events P{~'l ::;; z} and P{~/'1 ::;; z} in terms of the probabilities P~(x) and Pq(y).
Let~
5. The Bernoulli Scheme.
Large Numbers
I. The Law of
1. In accordance with the definitions given above, a triple
(Q, .szl, P)
.91
with
=
{w: w
{A: A c;; Q},
= (a 1 , . , a.), ai = 0, 1},
p(w)
pr.a;qnr.a;
is called a probabilistic model of n independent experiments with two outcomes, or a Bernoulli scheme.
46
I. Elementary Probability Theory
In this and the next section we study some limiting properties (in a sense
described below) for Bernoulli schemes. These are best expressed in terms of
random variables and of the probabilities of events connected with them.
by taking
= ai, i =
We introduce random variables 1 , .. ,
1, ... , n, where w = (a 1 , , a,). As we saw above, the Bernoulli variables
are independent and identically distributed:
e,
ei(w)
ei(w)
i = 1, ... , n.
ei
It is natural to think of as describing the result of an experiment at the
ith stage (or at time i).
Let us put S 0 (w) 0 and
k = 1, ... , n.
As we found above, ES,
np and consequently
s,
(1)
E=p.
In other words, the mean value of the frequency of "success", i.e. S,/n,
coincides with the probability p of success. Hence we are led to ask how much
the frequency Snfn of success differs from its probability p.
We first note that we cannot expect that, for a sufficiently small e > 0
and for sufficiently large n, the deviation of S,/n from p is less than e for all
w, i.e. that
wen.
(2)
In fact, when 0 < p < 1,
P{~" = 1} = P{~l = 1, ... , e, = 1} = p",
P{~" = o} = P{el = o, ... , e, = o} = q",
whence it follows that (2) is not satisfied for sufficiently small e > 0.
We observe, however, that when n is large the probabilities of the events
{S,/n = 1} and {S,/n = 0} are small. It is therefore natural to expect that the
total probability of the events for which I [S,(w)/n] Pi > e will also be
small when n is sufficiently large.
We shall accordingly try to estimate the probability of the event
{w: i[S,(w)/n] Pi > e}. For this purpose we need the following inequality,
which was discovered by Chebyshev.
47
I. The Law of Large Numbers
5. The Bernoulli Scheme.
Chebyshev's inequality. Let (Q, d, P) be a probability space and
nonnegative random variable. Then
~ = ~(ro)
(3)
for all e > 0.
PROOF.
We notice that
where I(A) is the indicator of A.
Then, by the properties of the expectation,
which establishes (3).
Corollary. If~ is any random variable, we have for e > 0,
P{l~l ~ e}
s;
El~l/e,
P{l~l ~ e} = P{~ 2 ~ e2 } s; eeje 2 ,
P{l~ E~l ~ e}
s; V~je 2
In the last of these inequalities, take~ =
Therefore
{I
s. p
p n
(4)
s.;n. Then using (4.14), we obtain
I }s ;ne s ;4ne ,
1
pq
~e
(5)
from which we see that for large n there is rather small probability that the
frequency S./n of success deviates from the probability p by more than e.
For n ~ 1 and 0 s; k s; n, write
Then
P{l s.n  P I~ e} =
{k:i(k/n) PI<!:<}
P.(k),
and we have actually shown that
{k: l(k/n) PI<!: e}
P.(k) s;
pq
2
ne
s;
1
4ne
2'
(6)
48
I. Elementary Probability Theory
P.(k)
I
np _ ____.
vnp + ne
~
np  ne
Figure 6
i.e. we have proved an inequality that could also have been obtained analytically, without using the probabilistic interpretation.
It is clear from (6) that
P,.(k)+ 0,
n+ oo.
(7)
{k:l(k/n) PI ~e)
We can clarify this graphically in the following way. Let us represent the
binomial distribution {P,.(k), 0 ~ k ~ n} as in Figure 6.
Then as n increases the graph spreads out and becomes flatter. At the same
time the sum of Pik), over k for which np  ne ~ k < np + ne, tends to 1.
Let us think of the sequence of random variables S0 , S 1 , . , S,. as the
path of a wandering particle. Then (7) has the following interpretation.
Let us draw lines from the origin of slopes kp, k(p + e), and k(p  e). Then
on the average the path follows the kp line, and for every e > 0 we can say that
when n is sufficiently large there is a large probability that the point S,.
specifying the position of the particle at time n lies in the interval
[n(p  e), n(p + e)]; see Figure 7.
We would like to write (7) in the following form:
n+ oo,
(8)
k(p +e)
s~
lkp
I
I
I
I
I
I
k(p e)
..........~,.,...      ++:
Figure 7
5. The Bernoulli Scheme.
I. The Law of Large Numbers
49
However, we must keep in mind that there is a delicate point involved
here. Indeed, the form (8) is really justified only if P is a probability on a
space (0, d) on which infinitely many sequences of independent Bernoulli
random variables ~ 1 , ~ 2 , .. , are defined. Such spaces can actually be
constructed and (8) can be justified in a completely rigorous probabilistic
sense (see Corollary 1 below, the end of 4, Chapter II, and Theorem 1, 9,
Chapter II). For the time being, if we want to attach a meaning to the analytic
statement (7), using the language of probability theory, we have proved only
the following.
Let (n<n>, .Jil<">, p<n>), n ;;;:: 1, be a sequence of Bernoulli schemes such that
n<n) = {w<n): w<n) = (a<;'>, .. ' a~">), al") = 0, 1},
.Jil<n> ={A: A n<n>},
p<">( w<">) =
pr.a~
q" !:.a\"'
and
Sk">(w<"l) = ~\">(w<">)
+ + ek">(w<">),
where, for n :s; 1, ~~nl, .. , ~~nl are sequences of independent identically
distributed Bernoulli random variables.
Then
n t oo. (9)
Statements like (7)(9) go by the name of James Bernoulli's law of large
numbers. We may remark that to be precise, Bernoulli's proof consisted in
establishing (7), which he did quite rigorously by using estimates for the
"tails" of the binomial probabilities P"(k) (for the values of k for which
l(k/n) pi ;;;:: e). A direct calculation of the sum of the tail probabilities of
the binomial distribution L{k:l<k/nJpl~l Pn(k) is rather difficult problem for
large n, and the resulting formulas are ill adapted for actual estimates of the
probability with which the frequencies Sn/n differ from p by less than e.
Important progress resulted from the discovery by De Moivre (for p = !)
and then by Laplace (for 0 < p < 1) of simple asymptotic formulas for Pn(k),
which led not only to new proofs of the law of large numbers but also to
more precise statements of both local and integral limit theorems, the essence
of which is that for large n and at least for k "' np,
and
50
I. Elementary Probability Theory
2. The next section will be devoted to precise statements and proofs of these
results. For the present we consider the question of the real meaning of the
law of large numbers, and of its empirical interpretation.
Let us carry out a large number, say N, of series of experiments, each of
which consists of "n independent trials with probability p of the event C of
interest." Let S~/n be the frequency of event C in the ith series and N, the
number of series in which the frequency deviates from p by less than e:
N, is the number of i's for which i(S~n) pi ~e. Then
N,/N "' P,
(10)
where P, = P{i(S!/n) Pi~ e}.
It is important to emphasize that an attempt to make (10) precise inevitably
involves the introduction of some probability measure, just as an estimate for
the deviation of Sn/n from p becomes possible only after the introduction of a
probability measure P.
3. Let us consider the estimate obtained above,
p{l sn pI~ e}=
n
{k:l<k/nJpl~l
Pn(k)
~ ~.
4ne
(11)
as an answer to the following question that is typical of mathematical
statistics: what is the least number n of observations that is guaranteed to
have (for arbitrary 0 < p < 1)
(12)
where ex is a given number (usually small)?
It follows from (11) that this number is the smallest integer n for which
1
e ex
(13)
~4
2
For example, if ex = 0.05 and e = 0.02, then 12 500 observations guarantee
that (12) will hold independently of the value of the unknown parameter p.
Later (Subsection 5, 6) we shall see that this number is much overstated;
this came about because Chebyshev's inequality provides only a very crude
upper bound for P{i(Sn/n) Pi ~ e}.
4. Let us write
{ I
Sn(w)  p
C(n, e) = w: n
I ~ e} .
From the law of large numbers that we proved, it follows that for every
e > 0 and for sufficiently large n, the probability of the set C(n, e) is close to
1. In this sense it is natural to call paths (realizations) w that are in C(n, e)
typical (or (n, e)typical).
5. The Bernoulli Scheme.
51
I. The Law of Large Numbers
We ask the following question: How many typical realizations are there,
and what is the weight p(w) of a typical realization?
For this purpose we first notice that the total number N(O.) of points is 2",
and that if p = 0 or 1, the set of typical paths C(n, e) contains only the single
path (0, 0, ... , 0) or (1, 1, ... , 1). However, if p =!,it is intuitively clear that
"almost all" paths (all except those of the form (0, 0, ... , 0) or (1, 1, ... , 1))
are typical and that consequently there should be about 2n of them.
It turns out that we can give a definitive answer to the question whenever
0 < p < 1; it will then appear that both the number of typical realizations
and the weights p(w) are determined by a function of p called the entropy.
In order to present the corresponding results in more depth, it will be
helpful to consider the somewhat more general scheme of Subsection 2 of
2 instead of the Bernoulli scheme itself.
Let (p 1 , p 2 , , p,) be a finite probability distribution, i.e. a set of nonnegative numbers satisfying p 1 + + p, = 1. The entropy of this distribution is
r
H =
p)npi,
(14)
i= 1
with 0 In 0 = 0. It is clear that H 2: 0, and H = 0 if and only if every pi,
with one exception, is zero. The function f(x) = x In x, 0 s x s 1, is
convex upward, so that, as know from the theory of convex functions,
f(xd
+ ~ + f(x,) s
!(x + ; + x}
1
Consequently
H = 
L' Pi In Pi s  r P1 + r + Pr In (P1 + r + Pr) = In r.
i= 1
In other words, the entropy attains its largest value for p 1 = = p, = 1/r
(see Figure 8 for H = H(p) in the case r = 2).
If we consider the probability distribution (p 1 , p 2 , , p,) as giving the
probabilities for the occurrence of events A 1 , A 2 , ... , A, say, then it is quite
clear that the "degree of indeterminancy" of an event will be different for
H(p)
Figure 8. The function H(p)
pIn p (1  p)ln(l  p).
52
I. Elementary Probability Theory
different distributions. If, for example, p 1 = 1, p 2 = = Pr = 0, it is clear
that this distribution does not admit any indeterminacy: we can say with
complete certainty that the result of the experiment will be A 1 . On the other
hand, if p 1 = = Pr = 1/r, the distribution has maximal indeterminacy,
in the sense that it is impossible to discover any preference for the occurrence
of one event rather than another.
Consequently it is important to have a quantitative measure of the indeterminacy of different probability distributions, so that we may compare
them in this respect. The entropy successfully provides such a measure of
indeterminacy; it plays an important role in statistical mechanics and in many
significant problems of coding and communication theory.
Suppose now that the sample space is
Q
= {w: w = (a 1 ,
.. ,
a.), ai
= 1, ... , r}
and that p(w) = p~1 (w) pv;<rol, where v,(w) is the number of occurrences of i
in the sequence w, and (p 1 , ... , Pr) is a probability distribution.
Fore> 0 and n = 1, 2, ... , let us put
{lvi(w)
n 
Pi I :2: e} ,
{ I
vlw)  Pi < e, 1 = 1, ... , r .
C(n, e) = w: n~
It is clear that
r
P(C(n, e)) :2: 1  i~t
P
and for sufficiently large n the probabilities P{l(v;(w)/n) p;l ;::: e} are
arbitrarily small when n is sufficiently large, by the law of large numbers
applied to the random variables
k(w)
= {01' ak:
, ak
~,
k = 1, ... , n.
r z
Hence for large n the probability of the event C(n, e) is close to 1. Thus, as in
the case n = 2, a path in C(n, e) can be said to be typical.
If all Pi > 0, then for every w E Q
p(w)
= exp{ n
(
vk~w) In Pk)}.
Consequently if w is a typical path, we have
v (w) Pk )
I L  ~kIn
n
r
k=t
 H ::::;; 
v (w) L I~kr
k=t
Pk In Pk ::::;;  e
L In Pk.
r
k=t
It follows that for typical paths the probability p(w) is close to eH andsince, by the law of large numbers, the typical paths "almost" exhaust Q
when n is largethe number of such paths must be of order e"H. These considerations lead up to the following proposition.
5. The Bernoulli Scheme.
53
I. The Law of Large Numbers
Theorem (Macmillan). Let P; > 0, i = 1, ... , rand 0 < s < 1. Then there is
an n0 = n0 (s; p 1 , , p,) such that for all n > n0
(a) en<HJ :::; N(C(n, sl)) :::; en<H+l;
(b) en(H+J :::; p(w) :::; en(Hl,
(c) P(C(n, s 1)) =
p(w)
4
wE
1,
C(n, s 1);
n+ oo,
roe C(n, <nl
where
s 1 is the smaller of sands/{ 2
kt/n Pk}
PROOF. Conclusion (c) follows from the law of large numbers. To establish
the other conclusions, we notice that if we C(n, s) then
k = 1, ... , r,
and therefore
vk In Pk} < exp{ n
:::; exp{ n(H  ts)}.
p(w) = exp{
L Pk ln Pk s1n L ln Pd
Similarly
p(w)
> exp{ n(H + ts)}.
Consequently (b) is now established.
Furthermore, since
P(C(n, s 1)) 2:: N(C(n, s 1 ))
we have
N ( C(n, S1 ))
:::;
min p(w),
coeC(n, <t)
P(C(n, sl))
1
_ n(H+(l/2JJ
.
( ) < n(H+(l/2JJ  e
mm pw
e
roeC(n, <tl
and similarly
N(C(n, s 1 )) 2:: P(C(n, s 1))
max p(w)
> P(C(n, s 1))en<H(1/ 2Jl.
meC(n, 1 )
Since P(C(n, s 1))+ 1, n+ oo, there is an n 1 such that P(C(n, s 1 ))
for n > n 1, and therefore
N(C(n, s 1 )) 2:: (1  s) exp{n(H  t)}
= exp{n(H 
s)
+ (tns + ln(l
 s))}.
> 1 s
54
I. Elementary Probability Theory
Let n2 be such that
+ ln(1
!ne
 e) > 0.
for n > n2 Then when n:;::: n0 = max(nto n2 ) we have
N(C(n, e1))
:;::: en<H
>.
This completes the proof of the theorem.
5. The law of large numbers for Bernoulli schemes lets us give a simple and
elegant proof of Weierstrass's theorem on the approximation of continuous
functions by polynomials.
Letf = f(p) be a continuous function on the interval [0, 1]. We introduce
the polynomials
which are called Bernstein polynomials after the inventor of this proof of
Weierstrass's theorem.
If ~ 1 , , ~" is a sequence of independent Bernoulli random variables
with P{~i = 1} = p, P{~i = 0} = q and S" = ~ 1 ++~"'then
Ef(~) =
Bip).
Since the function.{ = f(p), being continuous on [0, 1], is uniformly continuous, for every e > 0 we can find b > 0 such that I f(x)  f(y)l :::;; e
whenever lx yl :::;; b. It is also clear that the function is bounded: If(x)l :::;;
M < oo.
Using this and (5), we obtain
lf(p) Bn(p)l
=I Jo
: :;
[J(p)
f(~) JC~lqnk I
{k:i(k/n) PiS~)
I f(p)
{k:i(kfn)pi>~)
:::;; e + 2M
!(~) Jc~pkqnk
lj(p) J(~) IC~pkqnk
{k:i<kfn) Pi> ~J
C~lqnk :::;;
Hence
lim max
n+oo 0Sp:S1
I f(p)
 Bn(p)l = 0,
which is the conclusion of Weierstrass's theorem.
2M
M
e +  2 = e +  2
4n<5
2nb
6. The Bernoulli Scheme.
II. Limit Theorems (Local, De MoivreLaplace, Poisson)
55
6.PROBLEMS
1. Let ~and 1J be random variables with correlation coefficient p. Establish the following
twodimensional analog of Chebyshev's inequality:
1
P{l~ E~l ~ e~ or 111 EIJI ~
~ 2 (1 +
e
eJVtl}
.Ji="7).
(Hint: Use the result of Problem 8 of 4.)
2. Let f = f(x) be a nonnegative even function that is nondecreasing for positive x.
Then for a random variable ~ with I~(w) I ~ C,
EfW f(e) < P{l;: E;:l >c.}< Ef(~ E~).
f(C)
"'
"' f(e)
In particular, if f(x) = x 2 ,
Eee
C
3. Let ~ 1 , ,
~.be
v~
~ P{l~ E~l ~c.}~~
a sequence of independent random variables with
V~; ~
p{ I~ 1 +...n +~. _E(~ 1 +n... +~.)I> e} <ne__
C. Then
(15)
(With the same reservations as in (8), inequality (15) implies the validity of the law of
large numbers in more general contexts than Bernoulli schemes.)
4. "Let ~ 1 , . , ~.be independent Bernoulli random variables with P{e; = 1} = p > 0,
P{e; = 1} = 1  p. Derive the following inequality of Bernstein: there is a number
a > 0 such that
p{ I~ (2pwhere
s. =
~1
1)
I~ c.}~ 2eae2,
+ + ~. and e > 0.
6. The Bernoulli Scheme. II. Limit Theorems
(Local, De Moivre.:.Laplace, Poisson)
1. As in the preceding section, let
s. = el + ... + e.
Then
s.
E= p,
n
(1)
and by (4.14)
(2)
56
I. Elementary Probability Theory
It follows from (1) that Sn/n ~ p, where the equivalence symbol ~ has been
given a precise meaning in the law of large numbers in terms of an inequality
for P{I(Sn/n) Pi ~ e}. It is natural to suppose that, in a similar way, the
relation
(3)
which follows from (2), can also be given a precise probabilistic meaning
involving, for example, probabilities of the form
xeR\
or equivalently
(since ES. = np and VS.
If, as before, we write
= npq).
O~k~n,
for n
1, then
p{ IsnJ'iS,.
ESn
I~
x} =
Pn(k).
(4)
{k:j(knp)/JiiNI:5x}
We set the problem of finding convenient asymptotic formulas, as n + oo,
for Pn(k) and for their sum over the values of k that satisfy the condition on
the righthand side of (4).
The following result provides an answer not only for these values of k
(that is, for those satisfying Ik  np I = O(JnPq)) but also for those satisfying
Ik  np I = o(npq) 213
Local Limit Theorem. Let 0 < p < 1 ; then
P.(k) ~
uniformly fork such that
J2rffiM
e<knpl2!<2npql,
lk npl = o(npq) 2 13 , i.e. as n+
P.(k)
sup
{k:jknpj,;cp(n)}         
j2nnpq
where <p(n)
o(npq) 213
(5)
e (k np)2/(2npq)
oo
+
0,
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
57
The proof depends on Stirling's formula (2.6)
where R(n)
+
0 as n+ oo.
Then if n + oo, k
ck =
n
+
oo, n  k + oo, we have
n!
k!(n k)!
j2im e"n"
1 + R(n)
J2nk 2n(n  k) ekkk. e<nkl(n  k)nk (1
J2.. Hl _~) (m~ ~r
1 + e(n, k, n k)
+ R(k))(1 + R(n k))
where e = e(n, k, n k) is defined in an evident way and e+ 0 as n+ oo,
k + oo, n  k + oo.
Therefore
p J.k)
~ C,p'qo' ~
l{l 
1
k(
p)nk
k) (k)k(
k)" k (1
2nn 1  1 n
n n
n
+ e).
Write P= kjn. Then
Pn(k)
(p)k(1 _ p)nk
1 p
(1
J2nnfi(1  p)
1
exp{k ln
J2nnfi(1  p)
P
1
exp{n ln +
J2nnfi(1  p)
n P
~ + (n 
[~ ~
;:::===:======::;= exp{
J2nnfi(1  fi)
nH(p)}(1
e)
~} (1 + e)
k) ln 1 
1 P
(1  ~) ln 1  P]}
(1 + e)
P
1
+ e),
where
H(x) = xln
1x
p
+ (1 x)ln1.
We are considering values of k such that lk npl
sequently p p+ 0, n+ oo.
o(npq) 213 , and con
58
I. Elementary Probability Theory
Since, for 0 < x < 1,
1 X
X
H'(x) =In  I n   ,
1 p
p
1
1
H"(x) =  +   ,
1 X
X
H"'( )
X
= 
=~
+ H'(p)(p
(! + ~)
2 p
(p p) 2
+ (1
+ (p
if we write H(fi) in the form H(p
find that for sufficiently large n
H(p) = H(p)
X2
p)
X )2 '
 p)) and use Taylor's formula, we
+ !H"(p)(p
p) 2
+ O(IP pl 3 )
+ O(IP Pl 3 ).
Consequently
Pn(k)
~
2n
1
(p p) 2
exp{pq
J2nnfi(1  p)
+ nO(IP
pl 3 )} (1
+ s).
Notice that
_!!_ (p  p)z = _!!_ (~  p)z
2pq n
2pq
= (k  np)z
2npq
Therefore
l
Pn(k) =
e<knp)>J(Znpq)(l
+ s'(n, k, n
k)),
where
1 + s'(n, k, n k) = (1
+ s(n, k, n
k))exp{n O(IP Pl 3 )}
and, as is easily seen,
n + oo,
supls'(n, k, n k)l+ 0,
if the sup is taken over the values of k for which
lk npl:::;; cp(n),
This completes the proof.
cp(n)
= o(npq) 213
p(l  p)
p(l  p)
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
59
Corollary. The conclusion of the local limit theorem can be put in the folio~
equivalent form: For all x E R 1 such that x = o(npq) 116 , and for np + x.Jnpq
an integer from the set {0, 1, ... , n},
Pinp
+ x~) ,..., ~ ex2f2,
(7)
2nnpq
i.e. as n + oo,
sup
{x:lxl~l/l(n))
Pinp+ x~) _ 1
+ 0,
1
x2j2
;:==e
(8)
where t/J(n) = o(npq) 116.
With the reservations made in connection with formula (5.8), we can
reformulate these results in probabilistic language in the following way:
P{Sn = k} ,...,
p{ S
n np}
~
=x
~
1
e<knp>2f<2npq>,
"'~e
x2j2
lk npl
= o(npq) 213 , (9)
x = o(npq)lf6.
(10)
(In the last formula np + x~ is assumed to have one of the values
0, 1, ... , n.)
If we put tk = (k np)/~ and t:.tk = tk+ 1  tk = 1/~, the preceding formula assumes the form
Sn  np =
P{"==~
tk
tltk _ 12 12
,...,   e k '
(11)
fo
It is clear that t:.tk = 1/~+ 0 and the set of points {tk} as it were
"fills" the real line. It is natural to expect that (11) can be used to obtain the
integral formula
P{a < Sn 
np
b} ,. ., _1_ fb ex2Jz dx,
fo
 oo < a ~ b < oo.
Let us now give a precise statement.
2. For  oo < a ~ b < oo let
Pn(a, b] =
Pn(np
+ x~),
a<x~b
where the summation is over those x for which np
+ x~ is an integer.
60
I. Elementary Probability Theory
It follows from the local theorem (see also (11)) that for all tk defined by
k = np + tkfipq and satisfying Itk I ~ T < oo,
Pn(np
"PI.f)
+ tky'r.:=
2 /2
ll.tk
M: etk [1
v 2n
+ e(tk, n)],
(12)
where
n + oo.
sup le(tk, n)l+ 0,
(13)
itkiST
Consequently, if a and b are given so that  T
~ a ~
b ~ T, then
where
R~ll(a, b)=
ll.tk
el~/2
a<tksbjbi
R~2 >(a, b) =
_1_ rb ex2j2 dx,
j2n Ja
~tk
e(tk> n)   etki 2 .
fo
a<tkSb
From the standard properties of Riemann sums,
n + oo.
IR~1 >(a, b)l+ 0,
sup
(15)
TsasbsT
It also clear that
JR~2 >(a, b)l
sup
TsasbsT
sup Ie(tk> n)l
itkiST
x [
~ JT
y'
2n
T
ex 212 dx
sup
TsasbsT
JR~ll(a, b)JJ + 0,
(16)
where the convergence of the righthand side to zero follows from (15) and
from
_1_ fT ex2!2 dx < _1_ f"' ex2/2 dx = 1
Jbi
T
Jbi
oo
the value of the last integral being well known.
'
(17)
61
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
We write
<P(x) = 1
Jin
fx
oo
e 1212 dt.
Then it follows from (14)(16) that
n + oo.
IPn(a, b]  (<l>(b) <P(a))l+ 0,
sup
(18)
T:Sa:Sb:ST
We now show that this result holds for T = oo as well as for finite T. By
(17), corresponding to a given e > 0 we can find a finite T = T(e) such that
1
IT
forT
> 1
ex>f 2 dx
e.
(19)
According to (18), we can find anN such that for all n >Nand T = T(e)
we have
sup
T:Sa:Sb:ST
IPn(a, b]  (<P(b) <l>(a))l <
e.
(20)
It follows from this and (19) that
PnC T, T]
> 1
t e,
and consequently
Pn( oo, T]
+ Pn(T, oo)
te,
where Pn( 00, T] = lims!oo PnCS, T] and Pn(T, oo) = limstoo Pn(T, S].
Therefore for  oo ~ a ~  T < T ~ b ~ oo,
IPn(a, b] 
 1
foe
fb ex>/2 dx I
a
~ IPn( T, T] ~IT
+I
v 2n
Pn(a, T]
+
1 foo
+foe
~ !e
4
fo
Pn(oo, T]
ex 12
ex>j 2
dx
dx
I+ I
T
iT ex>f2
Pn(T, b]
fo s:
ex>f2
dx
+v~fT
+
1
1
1
1
+
+
+
4
2
8
8
dx ~ e
2n  oo
e
ex 212
e
dx
PnCT, oo)
e = e.
By using (18) it is now easy to see that Pn(a, b] tends uniformly to <l>(b)<D(a) for  oo ::::;; a < b ::::;; oo.
Thus we have proved the following theorem.
62
I. Elementary Probability Theory
De MoivreLaplace Integral Theorem. Let 0 < p < 1,
Pn(a, b] =
a<xsb
Pn(np
+ xJnpq),
Then
'P"(a,b]
sup
cosa<bsco
~iba ex 12 dx,+O,
2
y 2n
n+ oo.
(21)
With the same reservations as in (5.8), (21) can be stated in probabilistic
language in the following way:
sup
cosa<bsco
ESn
IP{a < snv~
vsn
::;; b} 
~ ib ex212 dx I+ 0,
v 2n
n + oo.
It follows at once from this formula that
P{A < S" :s; B} 
 np)  <II(AJnpq
 np)]
.jnpq
[<II(B
+
0,
(22)
as n+ oo, whenever  oo :s; A < B :s; oo.
A true die is tossed 12 000 times. We ask for the probability P that
the number of 6's lies in the interval (1800, 2100].
The required probability is
EXAMPLE.
P
(l)k(5)L:
c
12 ooo 1800<kS2100
6 6
k
12 000 k.
An exact calculation of this sum would obviously be rather difficult.
However, if we use the integral theorem we find that the probability P in
question is (n = 12 000, p = i, a = 1800, b = 2100)
2100 2000)
( 1800 2000)
<II 6
<II ( jt2 000. i. i <II jt2 000. i. i = (j6) <II(  2 j6)
~
<11(2.449)  <II( 4.898)
0.992,
where the values of <11(2.449) and <II( 4.898) were taken from tables of <P(x)
(this is the normal distribution function; see Subsection 6 below).
3. We have plotted a graph of Pn(np + xJnpq) (with x assumed such that
np +
is an integer) in Figure 9.
Then the local theorem says that when x = o(npq) 1 16 , the curve
(1/~)ex2 1 2 provides a close fit to Pn(np + x~). On the other hand
xJnM
theintegraltheoremsaysthatP"(a, b] = P{aj;lpq < S" np :s; bj;lpq} =
P{np + aj;lpq < Sn :s; np + bj;lpq} is closely approximated by the integral
(1/j2ic)J! ex2f2 dx.
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
P.(np
63
+ xJniq)
Figure9
We write
Then it follows from (21) that
sup
oo:S:x:Soo
n oo.
IFn(x)  <D(x)l  0,
(23)
It is natural to ask how rapid the approach to zero is in (21) and (23),
as n oo. We quote a result in this direction (a special case of the BerryEsseen theorem: see 6 in Chapter Ill):
sup
JFn(x)  <D(x)l ~
oo:Sx:Soo
+ q2
C:::
v npq
p2
(24)
It is important to recognize that the order of the estimate (1/.}niq)
cannot be improved; this means that the approximation of Fn(x) by <l>(x)
can be poor for values of p that are close to 0 or 1, even when n is large. This
suggests the question of whether there is a better method of approximation
for the probabilities of interest when p or q is small, something better than
the normal approximation given by the local and integral theorems. In this
connection we note that for p =!,say, the binomial distribution {Pn(~)} is
symmetric (Figure 10). However, for small p the binomial distribution is
asymmetric (Figure 10), and hence it is not reasonable to expect that the
normal approximation will be satisfactory.
P.(k)
P.(k)
0.3
p = !, n
0.2
= 10
'
0.1
2
0.3
0.2
1 ..,.
p = ;\, n
= 10
0.1
8
10
Figure 10
10
64
I. Elementary Probability Theory
4. It turns out that for small values of p the distribution known as the Poisson
distribution provides a good approximation to {P,(k)}.
Let
k_ 0 1
P,(k) =
,p q '
 ' ' ' n,
0,
k = n + 1, n + 2, . . ,
{ck knk
and suppose that pis a function p(n) of n.
Poisson's Theorem. Let p(n) + 0, n + oo, in such a way that np(n)+ A,
where A.> 0. Then fork= 1, 2, ... ,
P,(k)+ nk,
where
A.ke;.
nk=~,
n+ oo,
(25)
k = 0, 1, ....
(26)
The proof is extremely simple. Since p(n) = (A./n)
for a given k = 0, 1, ... and sufficiently large n,
P,(k) =
+ o(1/n) by hypothesis,
C~pkqnk
But
n(n l)(n k
=
1)[~ + o(~)r
n(n  1) ~~n  k
and
[ 1 ;:;A.
+ 1\A. + o( 1W+A.\
+ o ( ;:;l)]nk + e;.,
n+ oo,
n+ oo,
which establishes (25).
The set of numbers {nk, k = 0, 1, ... } defines the Poisson probability
distribution (nk ~ 0, Lk'=o nk = 1). Notice that all the (discrete) distributions
considered previously were concentrated at only a finite number of points.
The Poisson distribution is the first example that we have encountered of a
(discrete) distribution concentrated at a countable number of points.
The following result of Prohorov exhibits the rapidity with which P,(k)
converges tonk as n+ oo: if np(n) = A. > 0, then
(27)
(For the proof of (27), see 7 of Chapter III.)
65
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
5. Let us return to the De MoivreLaplace limit theorem, and show how it
implies the law of large numbers (with the same reservation that was made
in connection with (5.8)). Since
it is clear from (21) that when e > 0
P
{I s I }
___.!!_
p :::;: e  1
s..(iljpq
fo  ,,;nrpq
ex>J 2 dx+
0,
n + oo,
(28)
whence
n+ oo,
which is the conclusion of the law of large numbers.
From (28)
P{l SnpI : :;: e},...,
f'..fnipq
n
fo.,;nrpq
_1_
ex>/2
dx,
n+ oo,
(29)
whereas Chebyshev's inequality yielded only
It was shown at the end of 5 that Chebyshev's inequality yielded the estimate
1
n:<::42
e a
for the number of observations needed for the validity of the inequality
Thus withe= 0.02 and a= 0.05, 12 500 observations were needed. We can
now solve the same problem by using the approximation (29).
We define the number k(a) by
1 Jk(IX)

fo
Since e .J(il!Pq)
ex>f 2
dx = 1  a.
k(<X)
; : : 2e.jn, if we define n as the smallest integer satisfying
2eJn ;;::: k(a)
(30)
we find that
(31)
66
I. Elementary Probability Theory
We find from (30) that the smallest integer n satisfying
k 2 (rx)
n 2 4s2
guarantees that (31) is satisfied, and the accuracy of the approximation can
easily be established by using (24).
Taking s = 0.02, rx = 0.05, we find that in fact 2500 observations suffice,
rather than the 12 500 found by using Chebyshev's inequality. The values
of k(rx) have been tabulated. We quote a number of values of k(rx) for various
values of rx:
0(
k(a)
0.50
0.3173
0.10
0.05
0.0454
O.ol
0.0027
0.675
1.000
1.645
1.960
2.000
2.576
3.000
6. The function
<I>(x) =
_1_
fo
fx
et2/2
dt,
(32)
oo
which was introduced above and occurs in the De MoivreLaplace integral
theorem, plays an exceptionally important role in probability theory. It is
known as the normal or Gaussian distribution on the real line, with the
(normal or Gaussian) density
3 2 1
0.67I 1 1.96 2.58
Figure 11. Graph of the normal probability density cp(x).
6. The Bernoulli Scheme. II. Limit Theorems (Local, De MoivreLaplace, Poisson)
0.9
67
1
I
I
I
I
I
I
I
I
1111 I
3 2 1
111111
q1J
0.25111 I
0.5211/ I
o.67
0.841 I
1.28i
Figure 12. Graph of the normal distribution Cl>(x).
We have already encountered (discrete) distributions concentrated on a
finite or countable set of points. The normal distribution belongs to another
important class of distributions that arise in probability theory. We have
mentioned its exceptional role; this comes about, first of all, because under
rather general hypotheses, sums of a large number of independent random
variables (not necessarily Bernoulli variables) are closely approximated by
the normal distribution (4 of Chapter III). For the present we mention only
some of the simplest properties of cp(x) and cD(x), whose graphs are shown in
Figures 11 and 12.
The function cp(x) is a symmetric bellshaped curve, decreasing very
rapidly with increasing lxl: thus cp(1) = 0.24197, cp(2) = 0.053991, cp(3) =
0.004432, cp(4)
0.000134, cp(5)
0.000016. Its maximum is attained at
x = 0 and is equal to (2nr 112 ~ 0.399.
The curve cD(x) = (1/fo) J~ao e 1212 dt approximates 1 very rapidly
as x increases: C1>(1) = 0.841345, cD(2) = 0.977250, Cl>(3) = 0.998650, Cl>(4) =
0.999968, Cl>(4, 5) = 0.999997.
For tables of cp(x) and cD(x), as well as of other important functions that
are used in probability theory and mathematical statistics, see [A1].
7. PROBLEMS
1. Let n = 100, p = / 0 ,
120 , 130 , 1~, 150 Using tables (for example, those in [A1]) of the
binomial and Poisson distributions, compare the values of the probabilities
P{lO <
s,oo ~ 12},
P{33 < S100
35},
P{20 < S100
P{40 <
P{so < s,oo
22},
s,oo ~ 42},
~52}
with the corresponding values given by the normal and Poisson approximations.
68
I. Elementary Probability Theory
2. Letp = !andZ. = 2S. n(theexcessofl'soverO 'sinntrials).Showthat
suplfoP{Z2
j }  eP/ 4 "1> 0,
n>
oo.
3. Show that the rate of convergence in Poisson's theorem is given by
7. Estimating the Probability of Success
in the Bernoulli Scheme
1. In the Bernoulli scheme (Q, d, P) with
0, 1) }, d = A: A ~ Q },
n=
{w:w = (xb ... ' x.), X;=
p(w) = p'f.x;qn'f.x;,
we supposed that p (the probability of success) was known.
Let us now suppose that p is not known in advance and that we want to
determine it by observing the outcomes of experiments; or, what amounts
to the same thing, by observations of the random variables ~1> , ~.,where
~;(co) = x;. This is a typical problem of mathematical statistics, and can be
formulated in various ways. We shall consider two of the possible formulations: the problem of estimation and the problem of constructing confidence
intervals.
In the notation used in mathematical statistics, the unknown parameter
is denoted bye, assuming a priori that e belongs to the set 0 = [0, 1]. We
say that the set (Q, d, P8 ; 8 E 01 with p8(w) = er. x'(l  8)"r.x, is a probabilisticstatistical model (corresponding to" n independent trials" with probability of "success" E 0), and any function T, = T,(w) with values in 0 is
called an estimator.
If s. = ~ 1 + + ~. and T: = S./n, it follows from the law of large
numbers that
is consistent, in the sense that (s > 0)
T:
P8{/'J!'
8/
s}+ 0,
n+ oo.
Moreover, this estimator is unbiased: for every
E0 T: =
e,
(1)
e
(2)
where E8 is the expectation corresponding to the probability P8
The property of being unbiased is quite natural: it expresses the fact that
any reasonable estimate ought, at least "on the average," to lead to the
desired result. However, it is easy to see that
is not the only unbiased
estimator. For example, the same property is possessed by every estimator
T:
T,=
b1 x 1
+ + b"",
x
n
69
7. Estimating the Probability of Success in the Bernoulli Scheme
where b 1 + + bn = n. Moreover, the law of large numbers (1) is also
satisfied by such estimators (at least if Ib;l ~ K < oo; see Problem 2(b), 3,
Chapter Ill) and so these estimators T, are just as "good" as
In this connection there arises the question of how to compare different
unbiased estimators, and which of them to describe as best, or optimal.
With the same meaning of "estimator," it is natural to suppose that an
estimator is better, the smaller its deviation from the parameter that is being
estimated. On this basis, we call an estimator f;, efficient (in the class of unbiased estimators T,) if,
r:.
Oe9,
(3)
where V9 T, is the dispersion ofT,, i.e. E9(T, 0) 2
Let us show that the estimator
considered above, is efficient. We have
r:,
*_
(Sn) _ VSn _ n0(1 9
VoTnVoHence to establish that
2
0) _ 0(1  0)
(4)
r: is efficient, we have only to show that
. f v T.
()
n ~
0(1  0)
n
(5)
This is obvious for 0 = 0 or 1. Let 0 E (0, 1) and
Po(X;) = ox(l  0)1x;.
It is clear that
n
Po(w) =
01 Po(x;).
i=
Let us write
L 9(w) = In p6 (w).
Then
Lo(w)
= In 0. LX; + ln(l  O)L(l 
and
oL (w)
ae =
9
L(x;  0)
0(1  0)
Since
(jJ
and since Tn is unbiased,
(} =Eo T, = L T,(w)p (w).
9
(jJ
X;)
70
I. Elementary Probability Theory
After differentiating with respect to (), we find that
Therefore
1 = E{(1;. _ ()) oL;~w)]
and by the CauchyBunyakovskii inequality,
whence
1
(6)
Eo[T,  ()] ~ In(())'
where
I.(())=
[aL;~w)T
is known as Fisher's information.
From (6) we can obtain a special case of the RaoCramer inequality
for unbiased estimators T,:
(7)
In the present case
I (()) = E [oLo(w)]z =
n
()()
 O)]z
E [L(~;
8
()(1 _ ())
n()(l  ())
[()(1  ())JZ
n
()(1  ()) '
which also establishes (5), from which, as we already noticed, there follows
the efficiency of the unbiased estimator
= Sn/n for the unknown parameter e.
r:
r:
2. It is evident that, in considering
as a pointwise estimator for(), we have
introduced a certain amount of inaccuracy. It can even happen that the
numerical value of
calculated from observations of x~> ... , xn differs
rather severely from the true value e. Hence it would be advisable to determine the size of the error.
It would be too much to hope that r:( w) differs little from the true value
() for all sample points w. However, we know from the law of large numbers
r:
71
7. Estimating the Probability of Success in the Bernoulli Scheme
that for every {J > 0 and for sufficiently large n, the probability of the event
{I
w)I > {J} will be arbitrarily small.
By Chebyshev's inequality
o r:c
P {IO _ T*l
6
b} <
>
v6[J2r:
oon[J2 O)
and therefore, for every A > 0,
P9{1o r:1
~ AJ0(1;
O)} ~ 1 ;
If we take, for example, A = 3, then with P6 probability greater than 0.888
(1  (1/3 2 ) = ! ~ 0.8889) the event
10 r:1
3)0(1~
O)
will be realized, and a fortiori the event
10since 0(1  0) ~
Therefore
P6
r:1 ~ 3;:,
2y n
i.
{1 o r:1 ~ 2~} =
Po{
r:  2~ ~ o~ r: + 2~} ~ 0.8888.
In other words, we can say with probabilipr greater than 0.8888 that the exact
value of 0 is in the interval [T~  (3/2y'n), T~ + (3/2,fn)]. This statement
is sometimes written in the symbolic form
o~ r:
if.
c~ 88%),
where " ~ 88%" means "in more than 88% of all cases."
The interval [T:  (3/2.j1i),
+ (3/2.j1i)] is an example of what are
called confidence intervals for the unknown parameter.
r:
Definition. An interval of the form
where
t/1 1( w) and t/1 i
w) are functions of sample points, is called a corifidence
{J (or of significance level {J) if
interval of reliability 1 
P11 {1/1 1 (w)
for all 0 E 9.
t/1 2 (w)}
1 b.
72
I. Elementary Probability Theory
The preceding discussion shows that the interval
[ Tn* 
A Tn* + Jn
A J
Jn'
2
has reliability 1  (1/A 2 ). In point of fact, the reliability of this confidence
interval is considerably higher, since Chebyshev's inequality gives only
crude estimates of the probabilities of events.
To obtain more precise results we notice that
{w: /8 r:1 s ;.,)8(1: 8)} = {w: t/1 (T:, n):::; 8 s 1/Jz(T:, n)},
1
where t/1 1 = t/1 1(T:,n) and t/1 2 = t/1 2 (T:,n) are the roots of the quadratic
equation
(8  T!) 2
= ;.,z 8(1  8),
n
which describes an ellipse situated as shown in Figure 13.
Now let
Then by (6.24)
sup /F(J(x) <D(x)/
x
v n8(1  8)
Therefore if we know a priori that
0 < Ll ::::;; 8 ::::;; 1  Ll < 1,
where Ll is a constant, then
sup /F 0(x) <D(x)/
x
()
Figure 13
r:
Llv n
73
7. Estimating the Probability of Success in the Bernoulli Scheme
2:: (2<1>(A)  1) 
r:..
~vn
Let A* be the smallest A for which
(2<1>(A)  1) 
r:.
~vn
2:: 1  <5*,
where <5* is a given significance level. Putting <5 = <5*  (2/~Jn), we find
that A* satisfies the equation
For large n we may neglect the term 2/~Jn and assume that A* satisfies
<I>( A*) = 1 
~ <5*.
In particular, if A.* = 3 then c5* = 0.9973 .... Then with probability
approximately 0.9973
r:: 
3J8(1
~ 8) ~ () ~ r:: + 3J8(1 ~ 8)
(8)
or, after iterating and then suppressing terms of order O(n 3 14 ), we obtain
T,.*  3
T.*(1  T.*)
n
~ () ~
T,.*
+3
T.n*(1  T.n*)
(9)
Hence it follows that the confidence interval
[r: 2~, r: + 2~]
(10)
has (for large n) reliability 0.9973 (whereas Chebyshev's inequality only
provided reliability approximately 0.8889).
Thus we can make the following practical application. Let us carry out
a large number N of series of experiments, in each of which we estimate the
parameter () after n observations. Then in about 99.73% of the N cases, in
each series the estimate will differ from the true value of the parameter by
at most 3/2Jn. (On this topic see also the end of 5.)
74
I. Elementary Probability Theory
3. PROBLEMS
1. Let it be known a priori that 8 has a value in the set 8
estimator fore, taking values only in E>o.
c;; [0, 1]. Construct an unbiased
2. Under the hypotheses of the preceding problem, find an analog of the RaoCramer
inequality and discuss the problem of efficient estimators.
3. Under the hypotheses of the first problem, discuss the construction of confidence
intervals for e.
8. Conditional Probabilities and Mathematical
Expectations with Respect to Decompositions
1. Let (0, d, P) be a finite probability space and
a decomposition ofO (D; Ed, P(D;) > 0, i = 1, ... , k, and D 1 + + Dk =
0). Also let A be an event from d and P(AID;) the conditional probability of
A with respect to D;.
With a set of conditional probabilities {P(A ID;), i = 1, ... , k} we may
associate the random variable
n(w)
P(A ID;)lv,(w)
(1)
i= 1
(cf. (4.5)), that takes the values P(AID;) on the atoms of D;. To emphasize
that this random variable is associated specifically with the decomposition
fl2, we denote it by
P(Aifl2)
or
P(Aifl2)(w)
and call it the conditional probability of the event A with respect to the de
composition fl2.
This concept, as well as the more general concept of conditional probabilities with respect to aualgebra, which will be introduced later, plays an important role in probability theory, a role that will be developed progressively
as we proceed.
We mention some of the simplest properties of conditional probabilities:
P(A
+ B I fl2)
= P(A I fl2)
+ P(B Ifl2);
(2)
if fl2 is the trivial decomposition consisting of the single set 0 then
P(A I0)
P(A).
(3)
75
8. Conditional Probabilities and Mathematical Expectations
The definition of P(A i.@) as a random variable lets us speak of its expectation; by using this, we can write the formula (3.3) for total probability
in the following compact form:
EP(Ai.@) = P(A).
(4)
In fact, smce
P(AI.@)
L P(AID;)Iv;(m),
i= 1
then by the definition of expectation (see (4.5) and (4.6))
EP(Ai.@)
P(AID;)P(D;)
i= 1
P(AD;)
i= 1
= P(A).
Now let '1 = 1J(m) be a random variable that takes the values y 1, ... , Yk
with positive probabilities:
k
1J(m) =
L Yivim),
j= 1
where Di = {m: 1J(m) = y). The decomposition.@,= {D 1, ... , Dk} is called
the decomposition induced by 'I The conditional probability P(A !.@,) will
be denoted by P(AI'I) or P(AI'I)(m), and called the conditional probability
of A with respect to the random variable '1 We also denote by P(AI'I =yi)
the conditional probability P(AIDi), where Di = {m: 1J(m) = Yi}.
Similarly, if 1J 1, 1J 2 , , '1m are random variables and .@,,,, 2 , ... ,,'" is the
decomposition induced by '11> 1J 2 , , '1m with atoms
DY!.Y2 .. Ym
= {m: '11(m) = Y1 ' 'lm(m) = Ym},
then P(AID~ . ~, ..... ~~) will be denoted by P(AI17 1 , '7 2 , , '1m) and called the
conditional probability of A with respect to '11> 1] 2 , , '1m
1. Let ~ and 11 be independent identically distributed random variables, each taking the values 1 and 0 with probabilities p and q. For k =
0, 1, 2, let us find the conditional probability P(~ + '1 = ki'l) of the event
A = {m: ~ + '1 = k} with respect to '1
To do this, we first notice the following useful general fact: if ~ and 1J are
independent random variables with respective values x and y, then
ExAMPLE
P(~
+ 1J =
zi1J = y) = P(~
+y
= z).
(5)
In fact,
P(~ + = fi'l = y) = P(~ + '1 = z, '1 = y)
'1
P('l = y)
P(~
= P(~
+ y = Z,1J = y)
P('l = y)
+ y = z).
P(~
+ y = z)P(y = 'I)
P('l
= y)
76
I. Elementary Probability Theory
Using this formula for the case at hand, we find that
P(~
+ 17 =
+ 17 = ki11 = O)J{,=OJ(w)
+ P(~ + 11 = kl11 = 1)J{.,= 11(w)
= P(~ = k)I!,=O!(w) + P{~ = k
ki17) = P(~
1}J{,= 11(w).
Thus
(6)
or equivalently
q(1  11),
P(~
+ 11 = kl17) = { p(1
 17)
+ q17,
P11,
2. Let ~ =
~(w)
k = 0,
k = 1,
k = 2,
be a random variable with values in the set X = { x 1 ,
(7)
... ,
xn}:
L xJ4w),
j= 1
and let f!} = {D 1 , .. , Dk} be a decomposition. Just as we defined the expectation of~ with respect to the probabilities P(A),j = 1, ... , l.
l
Ee = L
xjP(Aj),
(8)
j= 1
it is now natural to define the conditional expectation of ~ with respect to f!}
by using the conditional probabilities P(Ail f!}), j = 1, ... , l. We denote
this expectation by EW f!}) or E(~l f!}) (w), and define it by the formula
l
E(~l f!}) =
L xiP(Ail f!}).
(9)
j= 1
According to this definition the conditional expectation E(~ If!}) (w) is a
random variable which, at all sample points w belonging to the same atom
Di> takes the same value
1 xiP(AiiD;). This observation shows that the
definition of E( ~I D;) could have been expressed differently. In fact, we could
first define E(~ ID;), the conditional expectation of~ with respect to D;, by
'LJ=
(10)
and then define
E(~lf!})(w)
L E(~ID;)Iv,(w)
i= 1
(see the diagram in Figure 14).
(11)
77
8. Conditional Probabilities and Mathematical Expectations
(8)
PO
E~
1(3.1)
(10)
P(ID)
E(~ID)
j(l)
P(1
1(11)
{9)
~)
E(~l ~)
Figure 14
It is also useful to notice that E(~ID) and EW ~)are independent of the
representation of~.
The following properties of conditional expectations follow immediately
from the definitions:
E(a~
+ b17i ~) =
aE(~I ~)
+ bE('71 ~),
a and b constants;
(12)
(13)
E(~l!l) = E~;
E(CI
~) =
C constant;
C,
(14)
(15)
The last equation shows, in particular, that properties of conditional probabilities can be deduced directly from properties of conditional expectations.
The following important property generalizes the formula for total
probability (5):
EE(~I ~)
(16)
E~.
For the proof, it is enough to notice that by (5)
EE(~I ~) = E
j= 1
j= 1
j= 1
L xiP(Ail ~) = L xiEP(Ail2)) = L xiP(Ai) =
E~.
Let f!} = {D 1 , . , Dk} be a decomposition and '7 = 17(w) a random
variable. We say that '7 is measurable with respect to this decomposition,
or ~measurable, if f!}~ ~ ~.i.e. '7 = 17(w) can be represented in the form
k
'l(W)=
L YJn,(w),
i= 1
where some Y; might be equal. In other words, a random variable is
measurable if and only if it takes constant values on the atoms of f!}.
2. If~ is the trivial decomposition,~ = {Q}, then '7 ts ~measur
able if and only if 17 C, where C is a constant. Every random variable
'7 is measurable with respect to f!}~.
ExAMPLE
78
I. Elementary Probability Theory
Suppose that the random variable 1J is
~measurable.
Then
(17)
and in particular
= 1J).
(E(1J I~")
To establish (17) we observe that if~=
xi/AJ' then
~1] =
IJ=
(18)
L L
XiYJAjD;
j= 1 i= 1
and therefore
k
L L xiy;P(AiDd~)
E(~1JI ~) =
j= 1 i= 1
L L
j=l i=l
L L
j= 1 i = 1
On the other hand, since
IJE(~I ge) =
=
[t
[t
P(AjD;IDm)lvm(w)
xiy;P(AiDdD;)lv;(w)
j= 1 i= 1
= lv; and lv; lvm = 0, i
YJv;(w)
(19)
xiy;P(AiiD;)lv;(w).
1 YJv/w)J
L L
Jb;
m=l
XjYi
lt
=I=
m, we obtain
1 xiP(Ail ge)J
Jmt Lt XjP(AjiDm)] lvJw)
L L Y;XiP(AiiD;) lv;(w),
i= 1 j= 1
which, with (19), establishes (17).
We shall establish another important property of conditional expectations.
Let ge 1 and ge 2 be two decompositions, with ~ 1 ~ ge 2 (ge 2 is "finer"
than ~ 1 ). Then
E[E(~I ~z)l
ge1J
= E(~l
gel).
For the proof, suppose that
gel = {Dll, , D1m},
Then if~ =
LJ= xi AJ' we have
1
E(~l gez) =
j= 1
xiP(Ail ge 2),
(20)
8. Conditional Probabilities and Mathematical Expectations
79
and it is sufficient to establish that
(21)
Since
P(Aj/ ~ 2 ) =
q=1
P(Aj/D 2 q)lv 2 . ,
we have
n
E[P(Aj/ ~2)/ ~ 1 ]
P(Aj /D 2 q)P(D 2 q/ ~d
q=1
p= 1
lv,p
q= 1
p=1
P(Aj/D 2 q)P(D 2 q/D 1 p)
lv,p
P(Aj/D 2 q)P(D 2 q/D 1 P)
{q:D2qSD1p}
p=1
lv,p P(Aj/D 1 p) = P(Aj/ ~ 1 ),
which establishes (21).
When ~ is induced by the random variables 11~> ... , 11k ( ~ = ~~, .... ~J,
the conditional expectation E(~I~~~ ..... ~J is denoted by E(~/11 1 , ,11k),
or E(( IIJ 1 , ... , l]k){w), and is called the conditional expectation of ( with
respect to 11 1 , .. , 11k
It follows immediately from the definition of E(~ /11) that if~ and 11 are
independent, then
E(~/11) = E~.
(22)
11
(23)
From (18) it also follows that
E(11 /11)
Property (22) admits the following generalization. Let ~ be independent
of ~(i.e. for each D; E ~the random variables~ and lv, are independent).
Then
E(~/ ~) = E~.
(24)
As a special case of (20) we obtain the following useful formula:
(25)
80
I. Elementary Probability Theory
EXAMPLE 3. Let us find E(~ + 11 111) for the random variables~ and 11 considered in Example 1. By (22) and (23),
This result can also be obtained by starting from (8):
E(~
+ 11111) =
EXAMPLE 4. Let
variables. Then
L kP(~ + 11 =
k=O
kll1) = p(1 11)
+ ql1 + 2pl1
+ 11
and 11 be independent and identically distributed random
(26)
In fact, if we assume for simplicity that ~and 11 take the values 1, 2, ... , m,
we find (1 ::; k ::; m, 2 ::; I ::; 2m)
P( ~ = k I~
+ '1
= I) =
P( ~ = k, '1 = I  k)
P( ~ = k, ~ + 11 = l)
P( ~ + 11 = l)
=
P( ~ + 11 = l)
P(~ =
k)P(11 = I  k)
+ 11 = l)
P( ~
=
P(IJ =
klc; +
P(11 = k)P(~ = 1  k)
P( ~
+ '1 =
l)
1J = [).
This establishes the first equation in (26). To prove the second, it is enough
to notice that
3. We have already noticed in 1 that to each decomposition ~ =
{D 1 , ... , Dk} of the finite set Q there corresponds an algebra a(~) of subsets
of Q. The converse is also true: every algebra !!J of subsets of the finite space
Q generates a decomposition ~ (!!J = a(~)). Consequently there is a onetoone correspondence between algebras and decompositions of a finite
space Q. This should be kept in mind in connection with the concept, which
will be introduced later, of conditional expectation with respect to the special
systems of sets called aalgebras.
For finite spaces, the concepts of algebra and aalgebra coincide. It will
turn out that if !!J is an algebra, the conditional expectation E( ~I !!J) of a
random variable ~ with respect to !!J (to be introduced in 7 of Chapter II)
simply coincides with EW ~), the expectation of ~ with respect to the decomposition ~ such that !!J = a(~). In this sense we can, in dealing with
finite spaces in the future, not distinguish between E(~I!!J) and EW ~),
understanding in each case that E(~I !!J) is simply defined to be E(~I ~).
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
4. PROBLEMS
1. Give an example of random variables
which
(Cf. (22).)
81
eand '1 which are not independent but for
2. The conditional variance of with respect to !!} is the random variable
Show that
3. Starting from (17), show that for every function f
= f('1) the conditional expectation
E(el'1) has the property
4. Let e and '1 be random variables. Show that inf1 E('1  f(e)) 2 is attained for J*(e) =
E('11 e). (Consequently, the best estimator for '1 in terms of in the meansquare sense,
is the conditional expectation E('11 e)).
e.
el .... ' e.
el .... 'e. are identically
e + + e, is the
5. Let
't' be independent random variables, where
distributed and r takes the values 1, 2, ... , n. Show that if S, =
sum of a random number of the random variables,
and
ES, = Er E~ 1 ,
6. Establish equation (24).
9. Random Walk. I. Probabilities of Ruin and
Mean Duration in Coin Tossing
1. The value of the limit theorems of 6 for Bernoulli schemes is not just
that they provide convenient formulas for calculating probabilities P(Sn = k)
and P(A < Sn ::;; B). They have the additional significance of being of a
universal nature, i.e. they remain useful not only for independent Bernoulli
random variables that have only two values, but also for variables of much
more general character. In this sense the Bernoulli scheme appears as the
simplest model, on the basis of which we can recognize many probabilistic
regularities which are inherent also in much more general models.
In this and the next section we shall discuss a number of new probabilistic
regularities, some of which are quite surprising. The ones that we discuss are
82
I. Elementary Probability Theory
again based on the Bernoulli scheme, although many results on the nature
of random oscillations remain valid for random walks of a more general
kind.
2. Consider the Bernoulli scheme (Q, d, P), where n = {(J): (J) = (x 1 ... ' Xn),
Xi= 1}, d consists of all subsets of Q, and p(w) = pv(w)qnv(wl, v(w) =
(L xi + n)/2. Let ~i(w) = xi, i = 1, ... , n. Then, as we know, the sequence
~ 1, . , ~n is a sequence of independent Bernoulli random variables,
P(~i =
1) = p,
P(~i
= 1) = q,
p+q=l.
Let us put S0 = 0, Sk = ~ 1 + + ~b 1 s k s n. The sequence S0 ,
S 1 , . , Sn can be considered as the path of the random motion of a particle
starting at zero. Here Sk+ 1 = Sk + ~k> i.e. if the particle has reached the
point Sk at time k, then at time k + 1 it is displaced either one unit up (with
probability p) or one unit down (with probability q).
Let A and B be integers, A s 0 s B. An interesting problem about this
random walk is to find the probability that after n steps the moving particle
has left the interval (A, B). It is also of interest to ask with what probability
the particle leaves (A, B) at A or at B.
That these are natural questions to ask becomes particularly clear if we
interpret them in terms of a gambling game. Consider two players (first
and second) who start with respective bankrolls (A) and B. If ~i = + 1,
we suppose that the second player pays one unit to the first; if~; = 1, the
first pays the second. Then Sk = ~ 1 + + ~k can be interpreted as the
amount won by the first player from the second (if Sk < 0, this is actually
the amount lost by the first player to the second) after k turns.
At the instant k s nat which for the first time Sk = B (Sk = A) the bankroll of the second (first) player is reduced to zero; in other words, that player
is ruined. (If k < n, we suppose that the game ends at time k, although the
random walk itself is well defined up to time n, inclusive.)
Before we turn to a precise formulation, let us introduce some notation.
Let x be an integer in the interval [A, B] and for 0 s k s n lets:= x + Sk,
r: = min{O
s ls
k: St =A orB},
(1)
where we agree to take r: = k if A < Sf < B for all 0 s l s k.
For each k in 0 s k s n and x E [A, B], the instant r:, called a stopping
time (see 11), is an integervalued random variable defined on the sample
space n (the dependence of r: on n is not explicitly indicated).
It is clear that for alll < k the set {w: r: = l} is the event that the random
walk {Sf, 0 s is k}, starting at time zero at the point x, leaves the interval
(A, B) at time l. It is also clear that when l s k the sets {w: r: = l, Sf= A}
and {w: r: = l, Sf = B} represent the events that the wandering particle
leaves the interval (A, B) at time l through A orB respectively.
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
For 0
83
n, we write
die =
L: {w: tic = 1, s~ = A},
O:S:ISk
&lie =
L:
{w:
O:S:!Sk
(2)
tic = 1, Sf = B},
and let
be the probabilities that the particle leaves (A, B), through A orB respectively,
during the time interval [0, k]. For these probabilities we can find recurrent
relations from which we can successively determine 1X 1(x), ... , 1Xn(x) and
Pt (x), ... , Pn(x).
Let, then, A < x <B. It is clear that 1X 0(x) = {J 0 (x) = 0. Now suppose
1 ~ k ~ n. Then by (8.5),
pk(x)
= P(&llc) = P(&llciS1 =
l)P(~ 1
= 1)
+ P(&llciS1 =X 1)P(~l = 1)
= pP(&IIc = X + 1) + qP(&IIcl S1 = X
1).
(3)
We now show that
P(&llciS1
=X+ 1) =
P(&llc~D,
To do this, we notice that
&lie=
{w:(x,x
&llc can be represented in the form
+ ~ 1 , ... ,x + ~ 1 + ... + ~k)EBk},
Blc is the set of paths of the form
(x, x + x 1 , ... , x + x 1 + xk)
with x 1 = 1, which during the time [0, k] first leave (A, B) at B (Figure 15).
where
A~
Figure 15. Example of a path from the set B';.
84
I. Elementary Probability Theory
We represent B~ in the form B~,x+ 1 + B~,x 1 , where B~x+ 1 and BV 1
are the paths in B~ for which x 1 = + 1 or x 1 = 1, respectively.
Notice that the paths (x, x + 1, x + 1 + x 2 , , x + 1 + x 2 + + xk)
in B~x+ 1 are in onetoone correspondence with the paths
(x
+ 1, x + 1 + x 2 , , x + 1 + x 2, , x + 1 + X 2 + + xk)
in B~~ ~. The same is true for the paths in B~x 1 Using these facts, together
with independence, the identical distribution of ~ 1 , ... , ~b and (8.6), we
obtain
P(Bi~ISf = X
+ 1)
= P(Bi~l~ 1 = 1)
= P{(x,x + ~t>x + ~ 1 + + ~k)eB~I~t = 1}
+ 1, x + 1 + ~ 2 , , x + 1 + ~2 + + ~k)eB~~D
P{(x + 1,x + 1 + ~1, .. ,x + 1 + ~1 + + ~kt)eB~~n
= P{(x
=
= P(Bi~~D
In the same way,
P(Bi~ISf
=X
1) = P(Bi~=D
Consequently, by (3) with x e (A, B) and k
Pk(x) = PPkt(x
n,
+ 1) + qflkt(x
1),
(4)
1:::;;, n.
(5)
where
{J1(B)
1,
{J1(A)
= 0,
Similarly
(6)
with
1X1(A) = 1,
1X1(B) =
0,
O~l~n.
Since tX 0 (x) = {J 0 (x) = 0, x e (A, B), these recurrent relations can (at least
in principle) be solved for the probabilities
1X1(x), ... , IX"(x)
and
{J 1(x), ... , fln(x).
Putting aside any explicit calculation of the probabilities, we ask for their
values for large n.
For this purpose we notice that since Bi~ _ 1 c: Bi~, k ~ n, we have
Pk 1(x) ~ {Jk(x) ~ 1. It is therefore natural to expect (and this is actually
the case; see Subsection 3) that for sufficiently large n the probability fln(x)
will be close to the solution {J(x) of the equation
{J(x) = p{J(x
1)
+ q{J(x 
1)
(7)
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
85
with the boundary conditions
f3(B) = 1,
f3(A) = 0,
(8)
that result from a formal approach to the limit in (4) and (5).
To solve the problem in (7) and (8), we first suppose that p =1= q. We see
easily that the equation has the two particular solutions a and b(qjpy, where
a and b are constants. Hence we look for a solution of the form
f3(x)
+ b(qjpy.
(9)
Taking account of (8), we find that for A ::;; x ::;; B
f3(x)
= (q/pY  (q/p)A.
(10)
(qjp)B _ (qjp)A
Let us show that this is the only solution of our problem. It is enough to
show that all solutions of the problem in (7) and (8) admit the representation (9).
Let P(x) be a solution with P(A) = 0, P(B) = 1. We can always find
constants aand b such that
a + b(qjp)A =
b(A),
a+
b(q/p)A+l = P(A
a+
b(q/p)A+ 2
+ 1).
Then it follows from (7) that
P(A
+ 2)
and generally
P(x) =
a + b(qjpy.
Consequently the solution (10) is the only solution of our problem.
A similar discussion shows that the only solution of
a(x)
= pa(x + 1) + qa(x  1),
XE(A, B)
(11)
with the boundary conditions
a(A) = 1,
a(B)
=0
(12)
A::;;x::;;B.
(13)
is given by the formula
(p/q)B  (qjpy
a(x) = (p/q)B _ (pjp)A'
If p = q =
respectively
t, the only solutions f3(x) and a(x) of (7), (8) and (11 ), ( 12) are
f3(x)
xA
= B A
(14)
and
Bx
a(x) = B A.
(15)
86
I. Elementary Probability Theory
We note that
a(x)
+ fJ(x) =
(16)
for 0 s; p s; 1.
We call a(x) and {J(x) the probabilities of ruin for the first and second
players, respectively (when the first player's bankroll is x  A, and the second
player's is B  x) under the assumption of infinitely many turns, which of
course presupposes an infinite sequence of independent Bernoulli random
variables ~b ~ 2 , . , where~;= + 1 is treated as a gain for the first player,
and ~; = 1 as a loss. The probability space (Q, d, P) considered at the
beginning of this section turns out to be too small to allow such an infinite
sequence of independent variables. We shall see later that such a sequence
can actually be constructed and that {J(x) and a(x) are in fact the probabilities
of ruin in an unbounded number of steps.
We now take up some corollaries of the preceding formulas.
If we take A = 0, 0 s; x s; B, then the definition of {J(x) implies that this
is the probability that a particle starting at x arrives at B before it reaches 0.
It follows from (10) and (14) (Figure 16) that
fJ(x) =
xjB, p = q
{ (qjpy  1
= !,
(qjp)B  1' p =f q.
(17)
Now let q > p, which means that the game is unfavorable for the first
player, whose limiting probability of being ruined, namely a = a(O), is given
by
(qjp)B  1
(q/pl  (qjp)A .
G(=~c
Next suppose that the rules of the game are changed: the original bankrolls
of the players are still (A) and B, but the payoff for each player is now !,
{J(x)
Figure 16. Graph of fl(x), the probability that a particle starting from x reaches B
before reaching 0.
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
87
rather than 1 as before. In other words, now let P(ei = !) = p, P(ei = !) =
q. In this case let us denote the limiting probability of ruin for the first player
by a. 112 Then
(qjp)2B _ 1
r:J.l/2 = (qjp)2B _ (qjp)2A'
and therefore
=
r:J.l/2
r:J.'
(qjp)B + 1
(qjp)B + (qjp)A > r:t.,
if q > p.
Hence we can draw the following conclusion:
if the game is unfavorable
to the .first player (i.e., q > p) then doubling the stake decreases the probability
of ruin.
3. We now turn to the question of how fast r:t.n(x) and Pn(x) approach their
limiting values a.(x) and P(x).
Let us suppose for simplicity that x = 0 and put
r:l.n = r:t.n(O),
It is clear that
Yn = P{A < Sk < B, 0
where {A < Sk < B, 0
n},
n} denotes the event
{A< Sk < B}.
O!>k!>n
Let n = rm, where r and m are integers and
+ ... + em,
em+ 1 + ... + e2m
el = el
e2
e, =
em(r1)+ 1
+ ... + erm
Then if C = IA I + B, it is easy to see that
{A
< Sk < B, 1 ~ k
and therefore, since ( 1 ,
rm} {1( 1 1 < C, ... , I(, I < C},
(,are independent and identically distributed,
Yn ~ P{ let I < C, ... , I(, I < C} =
We notice that
ciently large m,
Ve
0 P{l(d < C} =
= m[1  (p q) 2 ]. Hence, for 0
P{l( 1 1 < c}:::;
where 8 1 < 1, since
(P{I(tl < C})'.
ve 1
(18)
i=l
81,
~ C 2 if P{IC 1 1 ~ C}
= 1.
< p < 1 and suffi(19)
88
I. Elementary Probability Theory
If p = 0 or p = 1, then P{ i(d < C} = 0 for sufficiently large m, and
consequently (19) is satisfied for 0 ::::;; p ::::;; 1.
It follows from (18) and (19) that for sufficiently large n
Yn::::;; e",
where e = e~fm < 1.
According to (16), oc
+ f3
(20)
= 1. Therefore
(oc  OCn)
+ ({3  f3n)
Yn,
and since oc ~ ocn, f3 ~ f3n, we have
0 ::::;; oc  ocn ::::;; Yn ::::;; e",
0 ::::;; f3  /3n ::::;; Yn ::::;; e",
e < 1.
There are similar inequalities for the differences oc(x)  ocn(x) and f3(x) /3n(x).
4. We now consider the question of the mean duration of the random walk.
Let mk(x) = Er: be the expectation of the stopping time r:,k::::;; n. Proceeding as in the derivation of the recurrent relations for f3x(x), we find that,
for x E (A, B),
mk(x)
= Er: =
=
=
1 :s;l:s;k
1 :s;l:s;k
lP(r: = l)
l [pP(rk = ~~~1 = 1)
I [. [pP(r:~I
1 :s;l :s;k
(l
O:s;l:s;k1
= pmk 1(x
1)[pP(r:~f = l)
1)
11~1
= 1)]
= l  1) + qP(r:=I = l  1)]
+ qmk 1(x
[pP(r:~f = l)
O:s;l:s;k1
= pmk1(x
+ qP(rk =
+ qP(r:=:f
1)
+ qP(r:=:f
+ 1) + qmk 1(x
= l)]
1)
= l)]
+ 1.
Thus,forx E (A, B)andO::::;; k::::;; n, thefunctionsmk(x)satisfytherecurrent
relations
(21)
with m0 (x) = 0. From these equations together with the boundary conditions
mk(A)
mk(B)
= 0,
we can successively find m 1(x), ... , mn(x).
Since mk(x) ::::;; mk+ 1(x), the limit
m(x) = lim mix)
n+oo
(22)
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
89
exists, and by (21) it satisfies the equation
m(x)
+ 1 + pm(x + 1) + qm(x
 1)
(23)
with the boundary conditions
m(A) = m(B) = 0.
(24)
To solve this equation, we first suppose that
m(x) < oo,
(A, B).
(25)
Then if p =I= q there is a particular solution of the form a.j(q  p) and the
general solution (see (9)) can be written in the form
m(x) = Xqp
+ a + b (q)x
 .
p
Then by using the boundary conditions m(A) = m(B) = 0 we find that
1
m(x) =   (B{J(x)
pq
+ Aa.(x) 
x],
where fJ(x) and a.(x) are defined by (10) and (13). If p = q =
solution of (23) has the form
m(x)
+ bx 
(26)
!, the general
x 2,
and since m(A) = m(B) = 0 we have
m(x) = (B  x)(x  A).
(27)
It follows, in particular, that if the players start with equal bankrolls
(B = A), then
m(O) = B 2
If we take B = 10, and suppose that each turn takes a second, then the
(limiting) time to the ruin of one player is rather long: 100 seconds.
We obtained (26) and (27) under the assumption that m(x) < oo, x E (A, B).
Let us now show that in fact m(x) is finite for all x E (A, B). We consider only
the case x = 0; the general case can be analyzed similarly.
Let p = q =!.We introduce the random variableS<" defined in terms of
the sequence S0 , Sl> ... , Sn and the stopping time n = .~by the equation
n
s,n =
L Skl{tn=k)(w).
k=O
(28)
The descriptive meaning of S<n is clear: it is the position reached by the
random walk at the stopping time
Thus, if < n, then S," =A orB;
ifrn = n, then A~ S," ~B.
90
I. Elementary Probability Theory
Let us show that when p = q =
t.
ES,n = 0,
(29)
Es;n =Ern.
(30)
To establish the first equation we notice that
n
ES,n =
L E[SkJ{tn=k}(w)]
k=O
n
L E[SnJ{tn=k}(w)] + L E[(Sk Sn)Jitn=k}(w)]
k=O
= ESn
k=O
L E[(Sk Sn)II<n=k)(w)],
k=O
(31)
where we evidently have ESn = 0. Let us show that
n
L E[(Sk Sn)Jitn=k}(w)] = 0.
k=O
To do this, we notice that {tn > k} = {A < S 1 < B, ... , A < Sk < B}
when 0:::;; k < n. The event {A< S 1 < B, ... , A< Sk < B} can evidently
be written in the form
(32)
where Ak is a subset of { 1, + 1}k. In other words, this set is determined by
just the values of e1, ... , ek and does not depend on ek+t ... , en. Since
{tn = k} = {tn > k 1}\{tn > k},
this is also a set of the form (32). It then follows from the independence of
1 , .. , en and from Problem 9 of 4 that the random variables Sn  Sk and
I l<n=kl are independent, and therefore
E[(Sn Sk)II<n=kl] = E[Sn Sk] EJIn=kJ = 0.
Hence we have established (29).
We can prove (30) by the same method:
ES;" =
k=O
k=O
L ESfJ{tn=k} = L E([Sn + (Sk Sn)] 2I{tn=k})
n
L [ES;II<n=k) + 2ESn(Sk Sn)I 1,"=k)
k=O
= n
L (n 
k=O
k)P(tn = k) =
L kP(tn = k) = Et".
k=O
9. Random Walk. I. Probabilities of Ruin and Mean Duration in Coin Tossing
(p
Thus we have (29) and (30) when p = q =
+ q = 1) it can be shown similarly that
= (p
ES,n
E[S," r. E~ 1 ] 2
t.
91
For general p and q
(33)
 q) Ern,
V~ 1 Ern,
(34)
where E~ 1 = p q, V~ 1 = 1  (p q) 2
With the aid of the results obtained so far we can now show that
limn+ 00 mn(O) = m(O) < 00.
If p = q = !, then by (30)
(35)
If p =F q, then by (33),
E
r.:::;;
max( lA I, B)
(36)
Ipq I '
from which it is clear that m(O) < oo.
We also notice that when p = q =!
Ern=
es;n = A 2 (X.+ B 2 fJ. + E[S;I{A<Sn<Ba
and therefore
It follows from this and (20) that as n _... oo, Er. converges with exponential
rapidity to
m(O) = A 2 lX
BA
BA
+ B 2 {J = A 2    B 2   =
IABI.
There is a similar result when p =F q:
Er. _... m(0) _ ocA
+ fJB ,
pq
exponentially fast.
5. PROBLEMS
1. Establish the following generalizations of (33) and (34):
Es:. = x
+ (p
 q)Et~,
E[Stlf : E~tJ 2 = V~ 1 . Et:.
2. Investigate the limits of oc(x), P(x), and m(x) when the level A !
3. Let p = q =
 oo.
tin the Bernoulli scheme. What is the order of EIS.l for large n?
92
I. Elementary Probability Theory
4. Two players each toss their own symmetric coins, independently. Show that the
probability that each has the same number of heads after n tosses is r 2" D=o (C=) 2
Hence deduce the equation D=o (C=>2 = Cin
Let u. be the first time when the number of heads for the first player coincides with
the number of heads for the second player (if this happens within n tosses; un = n + 1
if there is no such time). Find Emin(u., n).
10. Random Walk. II. Reflection Principle.
Arcsine Law
1. As in the preceding section, we suppose that ~ 1 , ~ 2 , , ~ 2 " is a sequence
of independent identically distributed Bernoulli random variables with
= 1) = p,
+ + ~k
P(~j
sk =
~1
1~ k
2n;
We define
a2n
= min{1
:::; k :::; 2n:
sk = 0},
putting G'2n = 00 if sk =1: 0 for 1 ~ k :::; 2n.
The descriptive meaning of a 2 n is clear: it is the time of first return to
zero. Properties of this time are studied in the present section, where we
assume that the random walk is symmetric, i.e. p = q = !.
For 0 :::; k :::; n we write
u2k
= P(S 2 k = 0),
f2k
= P(a2n = 2k).
(1)
It is clear that u0 = 1 and
Our immediate aim is to show that for 1 :::; k :::; n the probability f 2 k is
given by
(2)
It is clear that
{a 2 n = 2k} = {S 1 =1: 0, S 2 =1: 0, ... , S 2 k 1 =1: 0, S 2 k = 0}
for 1 :::; k :::; n, and by symmetry
!2k
o, ... , s2k1 =1: o, s2k = O}
2P{s1 > o, ... , s2k1 > o, s2k = O}.
= P{S1
=
=1:
(3)
93
10. Random Walk. II. Reflection Principle. Arcsine Law
A sequence (S 0 , , Sk) is called a path of length k; we denote by Lk(A)
the number of paths of length k having some specified property A. Then
f2k
L2n<s1 >
o, ... , s2k1 > o, s2k = o,
and S2k+1 = a2k+1 ... ,S2n = a2k+1
= 2L2k(s1 >
+ + a2n)2 2"
o, ... , s2k1 > o, s2k = O) 2 2\
(4)
where the summation is over all sets (a 2k+ 1, ... , a2n) with a; = 1.
Consequently the determination of the probability f 2 k reduces to calculating the number of paths L 2k(S 1 > 0, ... , S 2k 1 > 0, S 2k = 0).
Lemma 1. Let a and b be nonnegative integers, a  b > 0 and k = a
Then
ab
Lk(S1>0, ... ,Sk1>0,Sk=ab)=kCk
+ b.
(5)
PRooF. In fact,
Lk(S 1 > 0, ... , Sk 1 > 0, Sk =a b)
= Lk(S 1 = 1, S2 > 0, ... , Sk 1 > 0, Sk =a b)
= Lk(S 1 = 1, Sk =a b)  Lk(S 1 = 1, Sk =a b;
and 3 i, 2 :::;; i :::;; k  1, such that S; :::;; 0).
(6)
In other words, the number of positive paths (S 1, S2, ... , Sk) that originate
at (1, 1) and terminate at (k, a b) is the same as the total number of paths
from (1, 1) to (k, a  b) after excluding the paths that touch or intersect the
time axis.*
We now notice that
Lk(S 1 = 1, Sk =a b; 3 i, 2 :::;; i:::;; k 1, such that S;:::;; 0)
=
Lk(S 1 = 1, Sk =a b),
(7)
i.e. the number of paths from IX = (1, 1) to f3 = (k, a  b), neither touching
nor intersecting the time axis, is equal to the total number of paths that
connect IX* = (1, 1) with {3. The proof of this statement, known as the
reflection principle, follows from the easily established onetoone correspondence between the paths A= (S 1, ... , Sa, Sa+ 1, ... , Sk) joining IX and
{3, and paths B = ( S 1, ... , Sa, Sa+~> ... , Sk)joining IX* and f3 (Figure 17);
a is the first point where A and B reach zero.
* A path (S 1 , , Sk) is called positive (or nonnegative) if all S1 > 0 (S1 ~ 0); a path is said to
touch the time axis if S1 ~ 0 or else S1 :::;; 0, for I s j s k, and there is an i, I :::;; i :::;; k, such
that S 1 = 0; and a path is said to intersect the time axis if there are two times i and j such that
0 and sj < 0.
s, >
94
I. Elementary Probability Theory
fJ
Figure 17. The reflection principle.
From (6) and (7) we find
Lk(S 1 > 0, ... , Sk 1 > 0, Sk =a b)
= Lk(S 1 = 1,Sk =a b) Lk(S 1 =
a1
= ck1 ck1 =
1,Sk =a b)
abca
k k,
which establishes (5).
Turning to the calculation ofj2 k, we find that by (4) and (5) (with a = k,
b = k 1),
f2k
o, ... , s2k1 > o,
= o) 2 2k
2L2k(sl >
2Lzk1(S1 > O, ... ,Szk 1 = 1)2 2k
s2k
t
1 ck
1 2k1 = 2k Uz(k1)
= 2 . rzk . 2k
Hence (2) is established.
We present an alternative proof of this formula, based on the following
observation. A straightforward verification shows that
1
2k Uz(k1) = u2(k1) 
(8)
Uzk>
At the same time, it is clear that
= 2k} = {a2n > 2(k 1)}\{a2n >
{a2n > 2!} = {S1 =I= 0, ... , S 21 =!= 0}
{azn
2k},
and therefore
{a2n = 2k} = {S 1 =I= 0, ... , S2(k 1) =I= 0}\{S 1 =I= 0, ... , S 2k =I= 0}.
Hence
fzk
P{S1 =I= 0, ... , Sz(k1) =I= 0}  P{S1 =I= 0, ... , S2k =I= 0},
95
10. Random Walk. II. Reflection Principle. Arcsine Law
Figure 18
and consequently, because of (8), in order to show that
it is enough to show only that
L2kCS1 0, ... , S2k 0)
f 2k = (1/2k)u 2 <kl)
= L2kCS2k = 0).
(9)
For this purpose we notice that evidently
L2k(S1 0, ... , S2k 0) = 2L2k(St > 0, ... , S2k > 0).
Hence to verify (9) we need only establish that
2L2k(St > 0, ... , S2k > 0) = L2k(S1
0, ... , S2k ~ 0)
(10)
and
(11)
Now (10) will be established if we show that we can establish a onetoone
correspondence between the paths A = (S~> ... , S 2 k) for which at least one
S; = 0, and the positive paths B = (S ~> ... , S 2k).
Let A = (S ~> ... , S2k) be a nonnegative path for which the first zero occurs
at the point a (i.e., Sa = 0). Let us construct the path, starting at (a, 2),
(Sa + 2, Sa+ 1 + 2, ... , S 2 k + 2) (indicated by the broken lines in Figure 18).
Then the path B = (S 1, .. , Sa_ 1 , Sa+ 2, ... , S 2 k + 2) is positive.
Conversely, let B = (S 1, .. , S 2k) be a positive path and b the last instant
at which Sb = 1 (Figure 19). Then the path
A = (S 1' . . . , Sb, Sb+ 1
Figure 19
2, ... , Sk  2)
96
I. Elementary Probability Theory
2k
m
Figure 20
is nonnegative. It follows from these constructions that there is a onetoone
correspondence between the positive paths and the nonnegative paths with
at least one Si = 0. Therefore formula (10) is established.
We now establish (11). From symmetry and (10) it is enough to show that
L2k(S1 > 0, ... , S 2k > 0) + L 2k(S 1 :2:: 0, ... , S 2k :2:: 0 and 3 i,
1 ~ i ~ 2k, such that S; = 0) = L 2 k(S 2 k = 0).
The set of paths (S 2 k = 0) can be represented as the sum of the two sets
"t' 1 and "t'2 , where "t' 1 contains the paths (S 0 , , S 2k) that have just one
minimum, and "t' 2 contains those for which the minimum is attained at at
least two points.
Let C 1 E "t' 1 (Figure 20) and let y be the minimum point. We put the path
C 1 = (S 0 , St. ... , S 2 k) in correspondence with the path Ct obtained in the
following way (Figure 21). We reflect (S 0 , S 1, , S1 ) around the vertical
line through the point l, and displace the resulting path to the right and
upward, thus releasing it from the point (2k, 0). Then we move the origin to
the point (l, m). The resulting path Ct will be positive.
In the same way, if C 2 e "t' 2 we can use the same device to put it into
correspondence with a nonnegative path C~.
(2k, 2m)
2k
Figure 21
97
10. Random Walk. II. Reflection Principle. Arcsine Law
Conversely, let Ct = (S 1 > 0, ... , S 2 k > 0) be a positive path with
S 2 k =2m (see Figure 21). We make it correspond to the path C 1 that is
obtained in the following way. Let p be the last point at which SP = m.
Reflect (SP, ... , S 2 m) with respect to the vertical line x = p and displace the
resulting path downward and to the left until its righthand end coincides
with the point (0, 0). Then we move the origin to the lefthand end of the
resulting path (this is just the path drawn in Figure 20). The resulting path
C 1 = (S 0 , . , S 2 k) has a minimum at S 2 k = 0. A similar construction
applied to paths (S 1 ~ 0, ... , S 2 k ~ 0 and 3 i, 1 ~ i ~ 2k, with S; = 0) leads
to paths for which there are at least two minima and S 2 k = 0. Hence we have
established a onetoone correspondence, which establishes (11).
Therefore we have established (9) and consequently also the formula
f2k
Uz(kl) 
= (1/2k)u2(kl)
U2k
By Stirling's formula
u2k =
k
c2k.
2k
k+
"'   ,
fo
00.
Therefore
k+
00.
Hence it follows that the expectation of the first time when zero is reached,
namely
Emin(a 2 n, 2n)
L 2kP(a
k=l
2n
= 2k) + 2nu 2 n
L u2(k1) + 2nu2n
k= 1
can be arbitrarily large.
In addition,
1 u 2 <k 1 > = oo, and consequently the limiting value of
the mean time for the walk to reach zero (in an unbounded number of steps)
is oo.
This property accounts for many of the unexpected properties of the
symmetric random walk that we have been discussing. For example, it
would be natural to suppose that after time 2n the number of zero net scores
in a game between two equally matched players (p = q = !), i.e. the number
of instants i at which S; = 0, would be proportional to 2n. However, in fact
the number of zeros has order
(see [F1]). Hence it follows, in particular,
that, contrary to intuition, the "typical" walk (S 0 , S 1 , . , S") does not have
a sinusoidal character (so that roughly half the time the particle would be
on the positive side and half the time on the negative side), but instead must
resemble a stretchedout wave. The precise formulation of this statement is
given by the arcsine law, which we proceed to investigate.
Lf=
fo
98
I. Elementary Probability Theory
2. Let P2 k, 2n be the probability that during the interval [0, 2n] the particle
spends 2k units of time on the positive side.*
Lemma 2. Let u0 = 1 and 0 ::::; k ::::; n. Then
(12)
PRooF. It was shown above thatf2k = u2<klJ u2 k. Let us show that
k
Uzk =
Since {S 2 k = 0}
!;;;;; {11 2"::::;
L f2r U2(kr)
r=
(13)
2k}, we have
{S 2k = 0} = {S 2k = 0} n {O'zn::::; 2k} =
1 ::s;l:s;k
{S2k = 0} n {0'2n = 21}.
Consequently
Uzk = P(S2k = 0) =
L
L
1 ::s;l::s;k
P(S2k = 0, O'zn = 21)
P(Szk = OIO'zk = 2l)P(O'zn = 2/).
1 :s;l::s;k
But
P(S 2k = OIO'zn = 21) = P(Szk = OIS1 # 0, ... , S211 # 0, S21 = 0)
+ (~21+1 + .. + ~zk) = O!S1 # 0, ... , S211
P(S21 + (~21+1 + + ~zk) = OISz, = 0)
P(~zt+ 1 + + ~2k = 0) = P(Sz<k1J = 0).
= P(S21
=
=
# 0, S21 = 0)
Therefore
U2k =
1 ::s;l::s;k
P(S2(k1J = O)P(O'zn = 21),
which establishes (13).
We turn now to the proof of (12). It is obviously true fork = 0 and k = n.
Now let 1 ::::; k ::::; n  1. If the particle is on the positive side for exactly 2k
instants, it must pass through zero. Let 2r be the time of first passage through
zero. There are two possibilities: either Sk ;;;:: 0, k ::::; 2r, or Sk ::::; 0, k ::::; 2r.
The number of paths of the first kind is easily seen to be
(21 22rf:2r) ' 22(nrJp 2(kr),2(nr)
1 22n ' J:2r' p 2(kr),2(nr)
= 2'
* We say that the particle is on the positive side in the interval [m  l, m] if one, at least, of the
values S,._ 1 and S,. is positive.
99
10. Random Walk. II. Reflection Principle. Arcsine Law
The corresponding number of paths of the second kind is
Consequently, for 1 s k
p2k,2n
n  1,
1
2 7 ~1 fzr p2(kr),2(nr) + 2 ,~/2r P2k,2(nr)
(14)
Let us suppose that P Zk, zm = u2 k u2 m _ Zk holds for m = 1, ... , n  1. Then
we find from (13) and (14) that
P2k,2n
= !uzn2k
r;1
r;1
L fzr llzk2r + !uzk L fzr llzn2r2k
This completes the proof of the lemma.
Now let y(2n) be the number of time units that the particle spends on the
positive axis in the interval [0, 2n]. Then, when x < 1,
1 y(2n)
P{  <   S
2n
{k, 1/2 < (2k/2n) s x)
P 2k, 2n
Since
ask+ oo, we have
p 2k,2n =
ll
2k
2(nk)
;====
_ k)'
njk(n
as k + oo and n  k + oo.
Therefore
P 2k, 2n
{k: 1/2 <(2k/2n)Sx}
{k: 1/2 <(2kj2n)~x)
__!,_
nn
[~ (1  ~)]n
112
+
0,
whence
{k: l/2<(2k/2n)Sn)
p 2k
2n '
1
n
IX
1/2
dt
+ 0,
jt(l  t)
But, by symmetry,
{k:k/nS 1/2)
Pzk,2n+!
n + oo.
n+
00,
100
I. Elementary Probability Theory
and
_1
n
Jx
112
dt
. vC.
1
= 2 arcsm
x  2
Jt(1  t)
n
Consequently we have proved the following theorem.
Theorem (Arcsine Law). The probability that the fraction of the time spent
by the particle on the positive side is at most x tends to 2n 1 arcsin .jX:
P lk, ln ~ 2n 1 arcsin
.jX.
(15)
{k:k/n!S:x}
We remark that the integrand p(t) in the integral
ix
dt
n o Jt(1  t)
represents aUshaped curve that tends to infinity as t
Hence it follows that, for large n,
P{o < y(2n)
< ~} > p{! < y(2n) <
2n2
2n
1
2
0 or 1.
+ ~}
'
i.e., it is more likely that the fraction of the time spent by the particle on the
positive side is close to zero or one, than to the intuitive value !.
Using a table of arcsines and noting that the convergence in (15) is indeed
quite rapid, we find that
p{Y~nn) :::;; 0.024} ~ 0.1,
P{y~:):::;; 0.1} ~ 0.2,
P{y~:) : :; 0.2} ~ 0.3,
P{y~:):::;; 0.65} ~ 0.6.
Hence if, say, n = 1000, then in about one case in ten, the particle spends
only 24 units of time on the positive axis and therefore spends the greatest
amount of time, 976 units, on the negative axis.
3.PR.OBLEMS
1. How fast does Emin(a 2., 2n)+ oo as n+ oo?
2. Lett.= min{1:::;; k:::;; n: sk= 1}, where we take t. = oo if Sk < 1 for 1:::;; k:::;; n.
What is the limit of Emin(t., n) as n+ oo for symmetric (p = q =!)and for unsymmetric (p of. q) walks?
101
II. Martingales. Some Applications to the Random Walk
11. Martingales. Some Applications to the
Random Walk
1. The Bernoulli random walk discussed above was generated by a sequence
~ 1 , ... , ~n of independent random variables. In this and the next section we
introduce two important classes of dependent random variables, those that
constitute martingales and Markov chains.
The theory of martingales will be developed in detail in Chapter VII.
Here we shall present only the essential definitions, prove a theorem on the
preservation of the martingale property for stopping times, and apply this
to deduce the "ballot theorem." In turn, the latter theorem will be used for
another proof of proposition (10.5), which was obtained above by applying
the reflection principle.
2. Let (Q, .91, P) be a finite probability space and 22 1 ~ 22 2
sequence of decompositions.
22n a
Definition 1. A sequence of random variables~ 1, ... , ~n is called a martingale
(with respect to the decomposition 22 1 ~ 22 2 ~ ~ 22n) if
(1) ~k is 22kmeasurable,
(2) E(~k+ 1l22k) = ~k' 1 :::;; k :::;; n  1.
In order to emphasize the system of decompositions with respect to which
the random variables form a martingale, we shall use the notation
(1)
where for the sake of simplicity we often do not mention explicitly that
1 :::;; k :::;; n.
When 22k is induced by ~ 1, ... , ~n' i.e.
instead of saying that ~ = (~k 22k) is a martingale, we simply say that the
sequence~ = (~k) is a martingale.
Here are some examples of martingales.
ExAMPLE
1. Let 111, ... , IJn be independent Bernoulli random variables with
P(IJk
sk
= 1) = P(IJk
= 1)
= t,
= 1J1 + ... + 1Jk and 22k = 22qJ, ... ,q.
We observe that the decompositions 22k have a simple structure:
102
I. Elementary Probability Theory
where
v+
= {w: 171 = +1},
~
v
= {w: 171 = 1},
{v++ v+ v+
'
'
'
v}
'
where
v+ + =
{w: 171 = + 1,172 = + 1}, ... ' v = {w: 'It= 1,172 = 1},
etc.
It is also easy to see that ~~~~k = ~s~o ... ,sk
Let us show that (Sb ~k) forms a martingale. In fact, Skis ~kmeasurable,
and by (8.12), (8.18) and (8.24),
E(Sk+tl~k)
= E(Sk + 1'fk+ti~k)
= E(Skl~k) + E(17Htl~k) = Sk + E1'fk+t = Sk.
If we put S0 = 0 and take D 0 = {0}, the trivial decomposition, then the
sequence (Sk,
~k)osksn
also forms a martingale.
2. Let 17 1, ... , 1'fn be independent Bernoulli random variables with
P(1]; = 1) = p, P(1]; = 1) = q. If p ::/: q, each of the !;lequences ~ = (~k)
with
ExAMPLE
where
sk =
'11 + ... + '1n
is a martingale.
EXAMPLE
3. Let 17 be a random variable, ~ 1
~n
and
(2)
Then the sequence ~ = (~k ~k) is a martingale. In fact, it is evident that
E(17 I~k) is ~kmeasurable, and by (8.20)
In this connection we notice that
by (8.20)
if~
(~k ~k)
is any martingale, then
~k = E(~k+tl~k) = E[E(~k+21~k+t)l~k]
= E(~k+21~k) = = E(~nl~k).
(3)
Consequently the set of martingales ~ = (~k ~k) is exhausted by the
martingales of the form (2). (We note that for infinite sequences ~ =
(~k ~k)k~t this is, in general, no longer the case; see Problem 7 in 1 of
Chapter VII.)
103
11. Martingales. Some Applications to the Random Walk
EXAMPLE 4. Let '1 ~> ... , '1 nbe a sequence of independent identically distributed
random Variables, Sk = '11 + + IJb and E1 = Esn' E2 = Esn.Sn~t' ,
En = Es".s"~ 1 , ... ,s, Let us show that the sequence~= (~b Ek) with
;:
sn
n
;:
sn1
n
S1=,~z=1 , ... ,;,k=
sn+1k
k'''''~n=S1
n+ 1 
is a martingale. In the first place, it is clear that Ek ~ Ek+ 1 and ~k is Ekmeasurable. Moreover, we have by symmetry, for j ::::; n  k + 1,
(4)
(compare (8.26)). Therefore
(n k + 1)E(1Jt1Ek) =
nk+ 1
j= 1
E(,.,jiEk) = E(Snk+11Ek) = snk+1
and consequently
;: =
Sk
Snk+l
n _ k+ 1
and it follows from Example 3 that~ =
=E(
(~b
'11
IE)
k ,
Ek) is a martingale.
ExAMPLE 5. Let 1J 1, ... , '1n be independent Bernoulli random variables with
P(IJ; = +1) = P(IJ; = 1) =
!,
Sk = '7 1 + + '1k Let A and B be integers, A < 0 <B. Then with 0 <A.<
(~k Ek) with Ek = Es,, ... ,sk and
n/2, the sequence~=
~k =(cos A.)k exp{iA.(sk
B; A)}
is a complex martingale (i.e., the real and imaginary parts of
martingales).
(5)
~k
3. It follows from the definition of a martingale that the expectation
the same for every k:
E~k = E~ 1 .
form
E~k
is
It turns out that this property persists if time k is replaced by a random
time.
In order to formulate this property we introduce the following definition.
Definition 2. A random variable T = r(w) that takes the values 1, 2, ... , n is
called a stopping time (with respect to a decomposition (Ek) 1 ~k~n E 1 ~
E 2 ~ ~ En) if, for k = 1, ... , n, the random variable /{t=kl(w) is Ekmeasurable.
If we consider !?)k as the decomposition induced by observations for k
steps (for example, Ek = E~ ...... ~k' the decomposition induced by the
104
I. Elementary Probability Theory
variables 17 1, .. , 1'/k), then the ~kmeasurability of Ift=kl(w) means that the
realization or nonrealization of the event {r = k} is determined only by
observations for k steps (and is independent of the "future").
If f!Jk = a(~k), then the ~kmeasurability of J{t=kl(w) is equivalent to the
assumption that
{r = k} E f!Jk.
(6)
We have already introduced specific examples of stopping times: the times
rk, (Jzn introduced in 9 and 10. Those times are special cases of stopping
times of the form
rA
= min{O < k:::;
(JA
= min{O:::;
n: ~k E A},
k:::; n: ~kEA},
(7)
which are the times (respectively the first time after zero and the first time)
for a sequence ~ 0 , ~ 1 , ... , ~n to attain a point of the set A.
4. Theorem 1. Let ~ = (~k ~k) 1 s;ks;n be a martingale and r a stopping time
with respect to the decomposition (~k) 1 s;ks;n Then
(8)
where
n
~t =
L ~kl{t=kl(w)
(9)
k=l
and
(10)
PRooF (compare the proof of (9.29)). Let D E ~ 1 . Using (3) and the properties
of conditional expectations, we find that
E(~ ID) = E(~Jn)
t
P(D)
= P(D) t~1E[~J{t=tl. In]
1
= P(D) E(~nln) = E(~niD),
105
ll. Martingales. Some Applications to the Random Walk
and consequently
E(e,l~t) = E(enl~t) = e1.
The equation Ee, = Ee 1 then follows in an obvious way.
This completes the proof of the theorem.
Corollary. For the martingale (Sk, ~k) 1 sksn of Example 1, and any stopping
timer (with respect to (~k)) we have the formulas
(11)
ES, = 0,
known as Wald's identities (cf. (9.29) and (9.30); see also Problem 1 and
Theorem 3 in 2 of Chapter VII).
5. Let us use Theorem 1 to establish the following proposition.
Theorem 2 (Ballot Theorem). Let q 1, ... , '1n be a sequence of independent
identically distributed random variables whose values are nonnegative integers,
Sk = q 1 + ... + 1'fk, 1 :::; k :::; n. Then
P{Sk < kforall k, 1:::; k:::; niSn} = (1
~r'
(12)
where a+ = max(a, 0).
PRooF. On the set {w: Sn ~ n} the formula is evident. We therefore prove
(12) for the sample points at which sn < n.
Let us consider the martingale =
~k)tsksn introduced in Example
4, With = Sn+1J(n + 1 k) and ~k = ~Sn+lkSn
We define
e (ek,
ek
r = min{1 :::; k ::5; n: ~k ~ 1},
taking ! = n on the set {ek < 1 for all k such that 1 ::5; k ::5; n} =
{maxl:s;l:s;n(SI/1) < 1}. It is clear that
S1 = 0 on this set, and
therefore
e, =en=
{ max S11 < 1} = {max
lSISn
lSISn
~~ < 1, S" < n}
{e, = 0}.
(13)
Now let us consider those outcomes for which simultaneously
max 1s 1 sn(S1/l) ~ 1 and Sn < n. Write a= n + 1  r. It is easy to see that
a= max{1 :::; k:::; n: Sk;;:::; k}
and therefore (since Sn < n) we have a < n, Sa ~ a, and Sa+ 1 < a + 1.
Consequently '1a+ 1 = Sa+ 1  Sa < (a + 1)  a = 1, i.e. '1a+ 1 = 0. Therefore a :::; sa = sa+ 1 < a + 1, and consequently sa = (J and
e,=
Sn+t =Sa=l.
n+1r
a
106
I. Elementary Probability Theory
Therefore
{ max~~~ 1,Sn < n} {~t = 1}.
(14)
1:51:5n
From (13) and (14) we find that
1
{ max S/
1 :51:5n
~ 1, S" < n} = {~t =
1} n {S" <
n}.
Therefore, on the set {S" < n}, we have
where the last equation follows because ~< takes only the two values 0 and 1.
Let us notice now that E(~tiSn) = E(~tl~ 1 ), and (by Theorem 1)
E(~tl~ 1 ) = ~ 1 = Sn/n. Consequently, on the set {S" < n} we have P{Sk < k
for all k such that 1 :::;; k :::;; n ISn} = 1  (Sn/n).
This completes the proof of the theorem.
We now apply this theorem to obtain a different proof of Lemma 1 of
10, and explain why it is called the ballot theorem.
Let ~ 1 , .. , ~"be independent Bernoulli random variables with
P(~ 1 = 1) = P(~; = 1) =
Sk
+ + ~k
and a, b nonnegative integers such that a  b > 0,
b = n. We are going to show that
= ~1
t,
ab
a+
P{S 1 > 0, ... , S" > OISn =a b} =b.
(15)
In fact, by symmetry,
P{S 1 > 0, ... , S" > OISn =a b}
= P{S 1 < 0, ... , S" < OISn = (a b)}
= P{S 1 + 1 < 1, ... , S" + n < niSn + n = n (a b)}
= P{1Jt < 1,. ., '11 + + 1Jn < ni1Jt + + 1Jn = n (aab
n
b)}
ab
a+ b'
where we have put 1Jk = ~k + 1 and applied (12).
Now formula (10.5) follows from (15) in an evident way; the formula was
also established in Lemma 1 of 10 by using the reflection principle.
II. Martingales. Some Applications to the Random Walk
107
Let us interpret ~i = + 1 as a vote for candidate A and ~i = 1 as a vote
for B. Then Sk is the difference between the numbers of votes cast for A and
B at the time when k votes have been recorded, and
P{S 1 > 0, ... , Sn > OISn =a b}
is the probability that A was always ahead of B, with the understanding that
A received a votes in all, B received b votes, and a  b > 0, a + b = n.
According to (15) this probability is (a  b)/n.
6. PROBLEMS
1. Let !?} 0 ~ !!} 1 ~~!!}.be a sequence of decompositions with !?} 0 = {Q}, and Jet
''lk be !?}kmeasurable variables, 1 ~ k ~ n. Show that the sequence~ = (~b !!}k) with
k
~k =
I= I
['11  E('ltl !?}II)]
is a martingale.
2. Let the random variables '1~o'1k satisfy E('1ki'11, ,'1kl) = 0. Show that the
sequence~ = (~k) 1 ,;k,;n with ~ 1 = 17 1 and
k
~k+ 1 =
i=l
'li+lli<'11 ... , 11J,
where fi are given functions, is a martingale.
3. Show that every martingale ~ = (~i> !!}k) has uncorrelated increments: if a< b <
c < d then
4. Let ~ = (~ 1 , ... , ~.) be a random sequence such that ~k is !?}kmeasurable
(!!} ~ !!} 2 ~ ~ !?}.). Show that a necessary and sufficient condition for this
sequence to be a martingale (with respect to the system (!!}k)) is that E~t = E~ 1 for
every stopping timeT (with respect to (!!}k)). (The phrase "for every stopping time"
can be replaced by "for every stopping time that assumes two values.")
5. Show that if~ =
(~k
!!}k) 1,;k,;n is a martingale and Tis a stopping time, then
for every k.
6. Let~= (~k !!}k) and '1 = ('lk !!}k) be two martingales, ~ 1 = 11 1 = 0. Show that
E~n'ln =
k=2
E(~k  ~k1)('1k  'lk1)
and in particular that
E~; =
k=2
E(~k  ~k1) 2
108
I. Elementary Probability Theory
7. Let 111, ... , Yfn be a sequence of independent identically distributed random variables
with E11; = 0. Show that the sequence~ = (~k) with
~k =
~
(.I
l=l
'7;)
kEYff,
_ exp A.('7 1 + + Yfk)
(E exp A.q 1)k
k 
is a martingale.
8. Let '7t, ... , Yfn be a sequence of independent identically distributed random variables
taking values in a finite set Y. Let f 0 (y) = P('7 1 = y), y E Y, and let f 1(y) be a nonnegative function with LyeY f 1(y) = 1. Show that the sequence ~ = (~k !0Z) with
~:J = D~"~k'
~k =
J; (111) .. . J; ('1d
fo('11) foCYfk)'
is a martingale. (The variables ~k, known as likelihood ratios, are extremely important
in mathematical statistics.)
12. Markov Chains. Ergodic Theorem.
Strong Markov Property
1. We have discussed the Bernoulli scheme with
Q
= {w: W = (xlo ... , Xn), X; = 0, 1},
where the probability p(w) of each outcome is given by
(1)
= pxqlx. With these hypotheses, the variables ~ 1 , ... , ~n
= X; are independent and identically distributed with
X= 0, 1.
P(~ 1 = x) = .. = P(~. = x) = p(x),
with p(x)
~;(w)
with
If we replace (1) by
p(w) = P1(x1) Pn(xn),
where P;(x) = pf(1  p;), 0 :::; p; :::; 1, the random variables ~ 1 ,
still independent, but in general are differently distributed:
... ,
~.
are
We now consider a generalization that leads to dependent random variables
that form what is known as a Markov chain.
Let us suppose that
n=
{w: w = (xo. xl, ... ' x.), X; EX},
109
12. Markov Chains. Ergodic Theorem. Strong Markov Property
where X is a finite set. Let there be given nonnegative functions p0 (x),
Pt(x, y), ... , Pn(x, y) such that
L Po(x) =
1,
L Pk(x, y) =
1,
XEX
k=1, ... ,n; yEX.
(2)
yeX
(3)
It is easily verified that Lroe!l p(w) = 1, and consequently the set of numbers
p(w) together with the space Q and the collection of its subsets defines a
probabilistic model, which it is usual to call a model of experiments that form
a Markov chain.
Let us introduce the random variables ~ 0 , ~ 1 , ... , ~n with ~;(w) =X;. A
simple calculation shows that
P(~ 0 = a) = p 0 (a),
P(~o
ao, ... , ~k
ak)
Po(ao)Pt(ao, at) Pk(ak1 ak).
(4)
We now establish the validity of the following fundamental property of
conditional probabilities:
P{~k+t
ak+tl~k
ak, ... , ~o
= ao} =
P{~k+t
ak+tl~k =ad
(5)
(under the assumption that P(~k = ak, ... , ~ 0 = a0 ) > 0).
By (4),
P{(k+t
ak+tl~k
P{~k+t
ak> ., ~o
= ao}
= ak+t' ., ~o = ao}
P{ ~k = ak> ... , ~o = a 0 }
Po(ao)Pt(ao, a1) Pk+ 1(ak, ak+ 1)
Po(ao) Pk(akl' ak)
(
)
Pk+t abak+t.
In a similar way we verify
(6)
which establishes (5).
Let !0i = !0~o ... , ~k be the decomposition induced by ~ 0 , . , ~k> and
P.Bi = a(!0i).
Then, in the notation introduced in 8, it follows from (5) that
(7)
or
I. Elementary Probability Theory
110
If we use the evident equation
P(AB I C)
= P(A I BC)P(B I C),
we find from (7) that
P{ ~n
an,., ~k+ 1 = ak+11.?4D
= Pgn =
an,, ~k+ 1 = ak+ll ~k}
(8)
or
P{~" =an, ... , ~k+1
= ak+1l~o, ... , ~d = P{~" =a.,, ~k+1 = ak+ll~k}.
(9)
This equation admits the following intuitive interpretation. Let us think
of ~k as the position of a particle "at present," (~ 0 , .. , ~k 1 ) as the "past,"
and (~k+ 1 ... , ~.)as the "future." Then (9) says that if the past and the present
are given, the future depends only on the present and is independent of how
the particle arrived at ~k' i.e. is independent of the past (~ 0 , ... , ~k 1 ).
Let F =(~.=a., ... , ~k+t = ak+t), N = gk =ad,
B
= {~k 1 =
ak1 , ~o
= ao}.
Then it follows from (9) that
P(FINB) = P(FIN),
from which we easily find that
P(FB IN) = P(F I N)P(B IN).
(10)
In other words, it follows from (7) that for a given present N, the future F
and the past B are independent. It is easily shown that the converse also
holds: if (10) holds for all k = 0, 1, ... , n  1, then (7) holds for every k = 0,
1, ... , n  1.
The property of the independence of future and past, or, what is the same
thing, the lack of dependence of the future on the past when the present is
given, is called the Markov property, and the corresponding sequence of
random variables ~ 0 , . , ~.is a Markov chain.
Consequently if the probabilities p(w) of the sample points are given by
(3), the sequence (~ 0 , ... , ~.)with ~;(w) =X; forms a Markov chain.
We give the following formal definition.
Definition. Let (Q, d, P) be a (finite) probability space and let~ = (~ 0 , . , ~.)
be a sequence of random variables with values in a (finite) set X. If (7) is
satisfied, the sequence~ = (~ 0 , ... , ~")is called a (finite) Markov chain.
The set X is called the phase space or state space of the chain. The set of
probabilities (Pn(x)), x EX, with p 0 (x) = P(~ 0 = x) is the initial distribution,
and the matrix IIPk(x,y)ll, x, yEX, with p(x,y) = Pgk = Yl~kt = x} is
the matrix of transition probabilities (from state x to state y) at time
k = 1, ... , n.
12. Markov Chains. Ergodic Theorem. Strong Markov Property
Ill
When the transition probabilities Pk(x, y) are independent of k, that is,
=
is called a homogeneous
0 , .. ,
Markov chain with transition matrix llp(x, y)ll
Pk(x, y) = p(x, y), the sequence
e (e
en)
Notice that the matrix ll(x, y)ll is stochastic: its elements are nonnegative
and the sum of the elements in each row is 1:
p(x, y) = 1, x EX.
We shall suppose that the phase space X is a finite set of integers
(X = {0, 1, ... , N}, X = {0, 1, ... , N}, etc.), and use the traditional
notation Pi = p0 (i) and Pii = p(i, j).
It is clear that the properties of homogeneous Markov chains completely
determine the initial distributions Pi and the transition probabilities Pii In
specific cases we describe the evolution of the chain, not by writing out the
matrix IIPiill explicitly, but by a (directed) graph whose vertices are the states
in X, and an arrow from state ito state j with the number Pii over it indicates
that it is possible to pass from point i to point j with probability Pii When
Pii = 0, the corresponding arrow is omitted.
LY
Pii
~i
j
EXAMPLE
1. Let X = {0, 1, 2} and
IIPijll
(P i).
3
The following graph corresponds to this matrix:
.l2
~a~_f).l3
o~2~
~
Here state 0 is said to be absorbing: if the particle gets into this state it remains
there, since p00 = 1. From state 1 the particle goes to the adjacent states 0
or 2 with equal probabilities; state 2 has the property that the particle remains
there with probability! and goes to state 0 with probability i
ExAMPLE
for
2. Let X= {0, 1, ... , N}, Po= 1, PNN = P(N)(N) = 1, and,
Iii< N,
p, j = i + 1,
{
Pii = q, j = i  1,
0 otherwise.
(11)
112
I. Elementary Probability Theory
The transitions corresponding to this chain can be presented graphically in
the following way (N = 3):
This chain corresponds to the twoplayer game discussed earlier, when each
player has a bankroll N and at each turn the first player wins + 1 from the
second with probability p, and loses (wins 1) with probability q. If we
think of state i as the amount won by the first player from the second, then
reaching state N or  N means the ruin of the second or first player, respectively.
In fact, if 17 1 , 17 2, ... , 'ln are independent Bernoulli random variables with
P('l; = + 1) = p, P(1]; = 1) = q, S 0 = 0 and Sk = 17 1 + .. + 'lk the
amounts won by the first player from the second, then the sequence S0 ,
S 1, , Sn is a Markov chain with p0 = 1 and transition matrix (11 ), since
P{Sk+1 =jiSk = ik,Sk1 = ik1, ... }
= P{Sk
= P{Sk
+ '7k+1
+ '7k+1
=jiSk = ik,Sk1 = ik1, ... }
=jiSk = ik} = P{'7k+1 =j ik}.
This Markov chain has a very simple structure:
0
n 1,
where 17 1 ,17 2 , .. , 'ln is a sequence of independent random variables.
The same considerations show that if ~ 0 , 17 1 , .. , 'ln are independent
random variables then the sequence ~ 0 , ~ 1 , , ~n with
0
k ~ n 1,
(12)
is also a Markov chain.
It is worth noting in this connection that a Markov chain constructed in
this way can be considered as a natural probabilistic analog of a (deterministic) sequence x = (x 0 , . , xn) generated by the recurrent equations
We now give another example of a Markov chain of the form (12); this
example arises in queueing theory.
EXAMPLE 3. At a taxi stand let taxis arrive at unit intervals of time (one at a
time). If no one is waiting at the stand, the taxi leaves immediately. Let 'lk be
the number of passengers who arrive at the stand at time k, and suppose that
1] 1 , . , 'ln are independent random variables. Let ~k be the length of the
113
12. Markov Chains. Ergodic Theorem. Strong Markov Property
waiting line at time k, ~ 0 = 0. Then if ~k = i, at the next time k
~k + 1 of the waiting line is equal to
j.=
{1Jk+ 1
i 1
+ 1Jk+ 1
+ 1 the length
if i = 0,
ifizl.
In other words,
0
n 1,
where a+ = max(a, 0), and therefore the sequence ~ = (~ 0 ,
Markov chain.
.. ,
~")
is a
4. This example comes from the theory of branching processes. A
branching process with discrete times is a sequence of random variables
~ 0 , ~ 1 , .. , ~"'where ~k is interpreted as the number of particles in existence
at time k, and the process of creation and annihilation of particles is as
follows: each particle, independently of the other particles and of the "prehistory" of the process, is transformed into j particles with probability pi,
j = 0, 1, ... , M.
We suppose that at the initial time there is just one particle, ~ 0 = 1. If at
time k there are ~k particles (numbered 1, 2, ... , ~k), then by assumption
~k+ 1 is given as a random sum of random variables,
EXAMPLE
~k+ 1
IJlk)
+ ... + 1]~~,
where 1Jlkl is the number of particles produced by particle number i. It is
clear that if ~k = 0 then ~k+ 1 = 0. If we suppose that all the random variables
IJ~k>, k 2 0, are independent of each other, we obtain
Pgk+1 = ik+1i~k = ik, ~k1 = ikt .} = Pgk+t = ik+ti~k = ik}
= P{IJ\kl + + IJ!:> = ik+ d.
It is evident from this that the sequence ~ 0 , ~ 1 ,
... , ~"is a Markov chain.
A particularly interesting case is that in which each particle either vanishes
with probability q or divides in two with probability p, p + q = 1. In this
case it is easy to calculate that
is given by the formula
Pii
{c
0
1,."12pi12qi j/2, ). = 0' ... , 2'I,
in all other cases.
2. Let ~ = ( ~k, Ill, IP') be a homogeneous Markov chain with st~rting vectors
(rows) Ill = (p;) and transition matrix Ill = IIPiill. It is clear that
Pii = P{~1 =j/~o = i} = ... = P{~n =j/~n1 = i}.
114
I. Elementary Probability Theory
We shall use the notation
for the probability of a transition from state i to state j in k steps, and
for the probability of finding the particle at point j at time k. Also let
Let us show that the transition probabilities pj'> satisfy the Kolmogorov
Chapman equation
\~ I) = " p\k)p(l!
P'1
l.J ra ~J'
(13)
or, in matrix form,
(14)
The proof is extremely simple: using the formula for total probability
and the Markov property, we obtain
P!~+IJ
= P(~k+l = j l~o =
i)
L P(~k+l = j, ~k = IXI~o = i)
~
L P(~k+l = j I~k = 1X)P(~k = IX I~0 = i) = L p~}P!~>.
~
The following two cases of (13) are particularly important:
the backward equation
1J = "P
P\~+
IJ
...., I~ p<l!
~}
(15)
and the forward equation
\~+ 1)
P IJ
= "p\k)p"
.
...., I~ ~}
(16)
(see Figures 22 and 23). The forward and backward equations can be written
in the following matrix forms
= IP<k>. IP,
(17)
IP<k+ tJ = IP. jp<k>.
(18)
IP<k+ 1J
12. Markov Chains. Ergodic Theorem. Strong Markov Property
115
I+ I
Figure 22. For the backward equation.
Similarly, we find for the (unconditional) probabilities p)k) that
'\' p(k)p(l)
PJ(_k +I) = ~
a:
'2.]'
(19)
or in matrix form
fl (k +I) = fl (k) I]J>(I).
In particular,
rn<k+ 1) =
rn<k)
IP
(forward equation) and
fl (k+ 1) = fl (1).
I]J>(k)
(backward equation). Since IP 0 ) = IP, fl (1) = fl, it follows from these equations
that
JP(k) =
!Pk,
Consequently for homogeneous Markov chains the kstep transitiOn
probabilities p~j) are the elements of the kth powers of the matrix IP, so that
many properties of such chains can be investigated by the methods of matrix
analysis.
k k
+1
Figure 23. For the forward equation.
116
I. Elementary Probability Theory
5. Consider a homogeneous Markov chain with the two states 0 and
1 and the matrix
EXAMPLE
IFD
= (Poo Pot)
Pu
Pto
It is easy to calculate that
p2
Po;(Poo + Pu))
Ptt + PotPto
= ( P~o + Po1P1o
Pto(Poo
and (by induction)
+ Pu)
1 (1  1
IFD" _
 2  Poo  Pu
Pu
1  Pu
Poo)
1  Poo
+ (Poo + Pu
 1)" ( 1  Poo
2  Poo  Pu
(1  Pu)
(1  Poo))
1 Pu
(under the hypothesis that IPoo + p 11  11 < 1).
Hence it is clear that if the elements of IFD satisfy Ip00 + p11  11 < 1 (in
particular, if all the transition probabilities Pii are positive), then as n+ oo
IFD"
+
1 (1  1
2  Poo  Pu
Pu
1  Pu
Poo)
1  Poo '
(20)
and therefore
l  p 11
lim (nl =
P,o
2
'
n
 Poo Pu
. Pil(n)
1tm
"
1  Poo
= ,''
2 Poo Pu
Consequently if IPoo + p 11  11 < 1, such a Markov chain exhibits
regular behavior of the following kind: the influence of the initial state on
the probability of finding the particle in one state or another eventually
becomes negligible (plj> approach limits ni, independent of i and forming a
probability distribution: n0 ~ 0, n 1 ~ 0, n0 + n 1 = 1); if also all Pii > 0
then n0 > 0 and n 1 > 0.
3. The following theorem describes a wide class of Markov chains that have
the property called ergodicity: the limits ni = limn Pii not only exist, are
independent of i, and form a probability distribution (ni ~ 0, Li ni = 1}, but
also ni > 0 for allj (such a distribution ni is said to be ergodic).
Theorem 1 (Ergodic Theorem). Let 1P
= IIPiill be the transition matrix of a
chain with a .finite state space X = {1, 2, ... , N}.
(a) If there is an n0 such that
min p\~ol
> 0'
lJ
i,j
(21)
117
12. Markov Chains. Ergodic Theorem. Strong Markov Property
then there are numbers n 1,
nN
such that
(22)
and
n oo
(23)
for every i e X.
if there are numbers n 1, . , nN satisfying (22) and (23), there
is an n0 such that (21) holds.
(c) 1he numbers (n 1 , ... , nN) satisfy the equations
(b) Conversely,
= 1, ... ,N.
j
PROOF.
(24)
(a) Let
m(n>
= min p!~>
)
I] '
M)"> = max Pl~P.
Since
1) = "P p<n)
P (~+
I)
L, I~ ~)'
(25)
we have
1 > =min p!'!+ 1 > =min" P p<n) >min" P min p<") = m(n>
m("+
)
I)
'
L, I~ ~) L, I~
~)
) '
eX
whence m}"> :::;; m)"+ >and similarly M}"> ~ M}"+ 1l. Consequently, to establish
(23) it will be enough to prove that
1
M(">  m("> .,... 0
1
Let e
n .,... oo, j = 1, ... , N.
'
min;, i Pljo> > 0. 'fhen
= "p(no>p<n)
= "L, (p(no)
 sp!">]p<")
P!'!o+n)
I)
L, I~
~)
I~
)~
~)
~
+ s "p(">p<n)
L,
)~
~)
(p(no)
 sp(">]p<n)
~
)~
~)
= "
L,
+ sp(~n)
11
But p!no)
 sp(">
> 0 ' therefore
la
Ja. 
>
m(n)."
[p!no)
 sp(">J
P!'!o+n)
I)
)
L,
I~
)~
+ sp(~n) =
))
m("l(1
 s)
)
and consequently
m!no+n)
>
m("l(1
 s)
1
1
+ sp(~n)
))
M(no+n>
< M(">(1
 s)
1
1
+ sp!~n>
))
In a similar way
Combining these inequalities, we obtain
M(no+n> _ m(no+n> < (M("> _ m(">). (1 _ s)
1
+ sp(~n)
Jl '
118
I. Elementary Probability Theory
and consequently
M<J<no+n> _ m<.kno+n> < (M<.n> _ m<.">)(1 _ e)k! 0
J
'
k+ 00.
Thus M~"tll  m~npl + 0 for some subsequence np, np + oo. But the
difference M~nl  m~n) is monotonic in n, and therefore M~nl  m~nl + 0,
n+ oo.
If we put ni = limn m~">, it follows from the preceding inequalities that
lp \~)IJ
7CI
< M(n)m<.n>
< (1  e)[n/no)1
J J
J

for n ~ n0 , that is, p!j> converges to its limit ni geometrically (i.e., as fast as a
geometric progression).
It is also clear that m~n> ~ m~no) ~ e > 0 for n ~ n0 , and therefore ni > 0.
(b) Inequality (21) follows from (23) and (25).
(c) Equation (24) follows from (23) and (25).
This completes the proof of the theorem.
4. Equations (24) play a major role in the theory of Markov chains. A
nonnegative solution (n 1 , , nN) satisfying I,. n~ = 1 is said to be a stationary
or invariant probability distribution for the Markov chain with transition
matrix IIPiill. The reason for this terminology is as follows.
Let us select an initial distribution (n 1, ... , nN) and take Pi= ni. Then
PJ(l)
= "" 1C
p . = 1C.J
...,~~}
~
and in general p)"> = ni. In other words, if we take (n 1, ... , nN) as the initial
distribution, this distribution is unchanged as time goes on, i.e. for any k
j = 1, ... ,N.
Moreover, with this initial distribution the Markov chain ~ =(~,Ill, IP) is
really stationary: the joint distribution of the vector (~k ~k+ ~> ... , ~k+ 1) is
independent of k for alll (assuming that k + 1 :::;; n).
Property (21) guarantees both the existence of limits ni = lim plj>, which
are independent of i, and the existence of an ergodic distribution, i.e. one
with ni > 0. The distribution (n 1 , . , nN) is also a stationary distribution.
Let us now show that the set (n 1 , .. , nN) is the only stationary distribution.
In fact, let (ft 1, . , ftN) be another stationary distribution. Then
and since P~1+ ni we have
iii
= L (ft~ n) = ni.
~
These problems will be investigated in detail in Chapter VIII for Markov
chains with countably many states as well as with finitely many states.
119
12. Markov Chains. Ergodic Theorem. Strong Markov Property
We note that a stationary probability distribution (even unique) may
exist for a nonergodic chain. In fact, if
then
[p>2n
(0 1)
1 0 '
[p>2n+ 1
(1 0)
0 1 '
and consequently the limits lim pjj> do not exist. At the same time, the
system
j
= 1, 2,
reduces to
of which the unique solution satisfying n 1 + n 2 = 1 is(!,!).
We also notice that for this example the system (24) has the form
no
noPoo
+ n1P1o
from which, by the condition n 0 = n 1 = 1, we find that the unique stationary
distribution (n 0 , n 1) coincides with the one obtained above:
no=
1 Pu
'
2  Poo  Pu
nl
1 Poo
2 Poo
P11
We now consider some corollaries of the ergodic theorem.
Let A be a set of states, A ~ X and
IA(x)
1, XEA,
= { 0, x :A.
Consider
which is the fraction of the time spent by the particle in the set A. Since
E[JA((k)i(o = i] = P((kEAI(o = i) = LPI,l(=p\kl(A)),
jeA
120
I. Elementary Probability Theory
we have
E[vA(n)l~o = i] =  1
and in particular
E[vw(n)l~o = i] = 
1
L pjkl(A)
k=o
n
+ 1 k=o
pjj>.
It is known from analysis (see also Lemma 1 in 3 of Chapter IV) that if
an+ a then (a 0 + + an)/(n + 1)+ a, n+ oo. Hence if pj'l+ ni, k+ oo,
then
where
nA =
L ni.
jeA
For ergodic chains one can in fact prove more, namely that the following
result holds for I A(~ 0 ), , I A( ~n), ....
Law of Large Numbers. If ~ 0 , ~ 1 , ... form a .finite ergodic Markov chain, then
n+ oo,
(26)
for every e > 0 and every initial distribution.
Before we undertake the proof, let us notice that we cannot apply the
results of 5 directly to IA(~ 0 ), ... , JA(~n), ... , since these variables are, in
general, dependent. However, the proof can be carried through along the
same lines as for independent variables if we again use Chebyshev's inequality, and apply the fact that for an ergodic chain with finitely many
states there is a number p, 0 < p < 1, such that
(27)
Let us consider states i and j (which might be the same) and show that,
fore> 0,
n+ oo.
By Chebyshev's inequality,
Hence we have only to show that
n+ oo.
(28)
121
12. Markov Chains. Ergodic Theorem. Strong Markov Property
A simple calculation shows that
1
(n
'\'
'\' m!~ I)
L.,
l}
'
k=o l=o
L.,
1)2
where
s = min(k, l) and t = lk 11.
By (27),
P!'!>
l}
= n:.J + e!'!>
l}'
Therefore
lml~ 1 >1:::; C1[p
+ P + Pk + /],
1
where C 1 is a constant. Consequently
,____,.,...,.. '\' '\' m<k. I) <
(n + 1)2 k~o 1~0 ii  (n
<
 (n
'\'
'\' [
+ 1)2 k~o 1~0 p
4C 1
2(n + 1)
2
+ 1) 1  p
k+
1]
8C 1
(n
+ 1)(1
 p)
+
n + oo.
Then (28) follows from this, and we obtain (26) in an obvious way.
5. In 9 we gave, for a random walk S 0 , S 1, ... generated by a Bernoulli
scheme, recurrent equations for the probability and the expectation of the
exit time at either boundary. We now derive similar equations for Markov
chains.
Let~= (~ 0 , , ~")be a Markov chain with transition matrix IIPiill and
phase space X = {0, 1, ... , N}. Let A and B be two integers,  N :::;
A :::; 0 :::; B :::; N, and x EX. Let f!lk+ 1 be the set of paths {x 0 , x 1 , .. , xk),
xi E X, that leave the interval (A, B) for the first time at the upper end, i.e.
leave (A, B) by going into the set (B, B + 1, ... , N).
, For A :::; x :::; B, put
f3k(x) = P{(~o .... , ~k) E fflk+ll~o = x}.
In order to find these probabilities (for the first exit of the Markov chain
from (A, B) through the upper boundary) we use the method that was
applied in the deduction of the backward equations.
122
I. Elementary Probability Theory
We have
{3k(x) = P{(eo, ... ' ek) E .?Jk+ 11 eo = X}
=
L Pxy. P{(eo, ... ' ek) E .?Jk+lleo =X, el =
y
y},
where, as is easily seen by using the Markov property and the homogeneity
of the chain,
P{(eo, ... ' ek)
.?Jk+lleo =X, el = y}
= P{(x, y, e2, ... ' ek) E .?Jk+lleo =X, el
= P{(y, e2, ... , ek) E BBkle1 = y}
= y}
= P{(y, eb ... ' ek dE .?lkl eo = y} = Pkl(y).
Therefore
y
for A < x < B and 1 :::;; k :::;; n. Moreover, it is clear that
Pix)= 1,
= B, B + 1, ... ' N,
and
X= N, ... ,A.
In a similar way we can find equations for ocix), the probabilities for first
exit from (A, B) through the lower boundary.
Let rk = min{O :::;; l :::;; k: (A, B)}, where rk = k if the set { } = 0.
Then the same method, applied to mix)= E(rkl eo = x), leads to the following recurrent equations:
e,
mk(x) = 1 +
L mkl (Y)Pxy
y
(here 1 :::;; k :::;; n, A < x < B). We define
x (A, B).
It is clear that if the transition matrix is given by (11) the equations for
ock(x), {3k(x) and mk(x) become the corresponding equations from 9, where
they were obtained by essentially the same method that was used here.
These equations have the most interesting applications in the limiting
case when the walk continues for an unbounded length of time. Just as in 9,
the corresponding equations can be obtained by a formal limiting process
(k +
CX) ).
By way of example, we consider the Markov chain with states {0, 1, ... , B}
and transition probabilities
Poo
= 1,
PBB =
1,
123
12. Markov Chains. Ergodic Theorem. Strong Markov Property
and
Pi > 0, ~ = ~ + 1,
pij = { ri,
J = z,
qi > 0, j = i  1,
for 1 ::;; i ::;; B  1, where Pi + qi + ri = 1.
For this chain, the corresponding graph is
a~~e
0
q1
P1
qB1 B 1 PB1
It is clear that states 0 and B are absorbing, whereas for every other state
i the particle stays there with probability ri, moves one step to the right with
probability Pi and to the left with probability qi.
Let us find a(x) = limkoo ak(x), the limit of the probability that a particle
starting at the point x arrives at state zero before reaching state B. Taking
limits ask+ oo in the equations for aix), we find that
when 0 < j < B, with the boundary conditions
a(O)
1,
a(B) = 0.
Since rj 1  qj pj, we have
pj(a(j + 1)  a(j)) = qj(a(j)  a(j  1))
and consequently
a(j
+ 1) 
a(j) = pj(a(1)  1),
where
Po= 1.
But
a(j
+ 1) 
1=
L (a(i + 1) 
a(i)).
i=O
Therefore
a(j
+ 1) 
1 = (a(1)  1)
L Pi
i=l
124
I. Elementary Probability Theory
Ifj = B  1, we have r:~.(j
+ 1) =
0, and therefore
r:~.(B) =
r:J.(l)=1='\'B1
L..i= 1 Pi
'
whence
an d
'\'B1
L..i=i Pi
'
L..i=1 Pi
(
r:I.J
0)
='\'B1
j = 1,
0'
B.
(This should be compared with the results of 9.)
Now let m(x) = limk mk(x), the limiting value of the average time taken
to arrive at one of the states 0 or B. Then m(O) = m(B) = 0,
m(x)
= 1 + L m(y)Pxy
y
and consequently for the example that we are considering,
m(j) = 1 + qimU  1)
+ rimU) + pimU + 1)
for j = 1, 2, ... , B  1. To find mU) we put
M(j) = m(j)  m(j 1),
j = 0, 1,. 0. 'B.
Then
piM(j + 1)
= qiM(j)
 1,
j = 1, ... , B 1,
and consequently we find that
M(j
+ 1) =
piM(1) Ri,
where
q1 .. qj
0
P1 Pi '
0
Therefore
j1
m(i) = mU) m(O) =
L M(i + 1)
i=O
j1
i=O
(pim(l) Ri) = m(l)
j1
j1
i=O
i=O
L Pi L Ri.
It remains only to determine m(1). But m(B) = 0, and therefore
'\'Bl R
m(1) = Ik=o1 i,
i=O Pi
and for 1 < j ::;; B,
( 0)
mJ =
j1
'\'~1R.
j1
'\' R
L.. = 0 I
'\'
i~o Pi. Lf;o1 Pi  i:o io
125
12. Markov Chains. Ergodic Theorem. Strong Markov Property
(This should be compared with the results in 9 for the case ri = 0, Pi = p,
qi
= q.)
6. In this subsection we consider a stronger version of the Markov property
(8), namely that it remains valid if time k is replaced by a random time (see
also Theorem 2). The significance of this, the strong Markov property, can
be illustrated in particular by the example of the derivation of the recurrent
relations (38), which play an important role in the classification of the states
of Markov chains (Chapter VIII).
Let ~ = (~ 1 , . . . , ~n) be a homogeneous Markov chain with transition
matrix IIPijll; let ~~ = (~f)oo;;k,;;n be a system of decompositions, ~f =
~ ~o ... ~k. Let ~f denote the algebra ri(~i) generated by the decomposition
~f.
We first put the Markov property (8) into a somewhat different form. Let
Let us show that then
E ~f.
= ak+llB n (~k = ak)}
= P{~n =an, ... , ~k+l =
Pgn =an,, ~k+l
ak+ll~k
= ak}
(29)
(assuming that P{B n (~k = ak)} > 0). In fact, B can be represented in the
form
B=
where
I* {~~ .... , ~k = at},
I* extends over some set (a~, ... ' an Consequently
P{~n =an, .. , ~k+l
P{(~n
= ak+llB n
(~k
= ak)}
=an, ... , ~k = ak) n B}
P{(~k = ak) n B}
I* P{(~n
= aw .. , ~k = ak) n (~ 0 = a6, ... , ~k =a%)}
P{(~k = ak) n B}
(30)
But, by the Markov property,
P{(~n
=an, ... , ~k = ak) n (~o = a~, ... , ~k =at)}
P{~n =an, ... , ~k+l = ak+ll~o =a~,, ~k =at}
x P{~ 0 =a~, ... , ~k =at} ifak =at,
0 if ak i= at,
P{~n =an, ... , ~k+l = ak+ll~k = ak}P{~o =a~, ... , ~k =at}
if ak =at,
0 if ak i= at,
={
pgn =an,., ~k+1
0
if ak =at,
if ak i= at.
= ak+ll~k = ak}P{(~k = ak) n B}
126
I. Elementary Probability Theory
Therefore the sum I* in (30) is equal to
P{ ~n = an, ... , ~k+ 1 = ak+ d ~k = adP{(~k = ak) n B},
This establishes (29).
Let r be a stopping time (with respect to the system D~ = (DDo 5 k 5 n; see
Definition 2 in 11).
Definition. We say that a set Bin the algebra~~ belongs to the system of sets
~; if, for each k, 0 ~ k ~ n,
B n {r
= k}
(31)
E ~f.
It is easily verified that the collection of such sets B forms an algebra
(called the algebra of events observed at timer).
Theorem 2. Let ~ = (~ 0 , , ~n) be a homogeneous Markov chain with
transition matrix IIPull, r a stopping time (with respect to 2fi~), BE~~ and
A = {w: r + l ~ n}. Then if P{A n B n (~, = a0 )} > 0, we have
P{~r+t
and if P{A n
(~,
=
=
a1,
P{~r+t
= a 1 IA n B n (~, = a 0 )}
= a1, , ~r+l = a 1 IA n (~, = a 0 )},
~r+l
(32)
= a0 )} > 0 then
For the sake of simplicity, we give the proof only for the case l = 1. Since
B n (r = k) E ~L we have, according to (29),
P{~r+l =
a 1, An B n (~, = a0 )}
k5nl
P{~k+l
= a1, ~k =
a0 , r
= k, B}
k5nl
k5nl
Paoal.
k5nl
P{~k = ao, r = k, B} =
Paoal.
P{A n B n (~,
= ao)},
which simultaneously establishes (32) and (33) (for (33) we have to take
B= Q).
Remark. When l
= 1, the strong Markov property (32), (33) is evidently
equivalent to the property that
(34)
127
12. Markov Chains. Ergodic Theorem. Strong Markov Property
for every C s;; X, where
Pao(C) =
L Paoa1
a1EC
In turn, (34) can be restated as follows: on the set A = {T
n 1},
(35)
which is a form of the strong Markov property that is commonly used in the
general theory of homogeneous Markov processes.
7. Let ~ = (~ 0 , . ,
matrix IIPiill, and let
~n)
be a homogeneous Markov chain with transition
!!7>
= P{ ~k = i, ~~ =F i, 1 ~ l ~ k  11 ~ 0 = i}
(36)
f!J>
= P{~k =j, ~~ #j,
1 ~ l ~ k 11~ 0 = i}
(37)
and
for i =F j be respectively the probability of first return to state i at time k and
the probability of first arrival at state j at time k.
Let us show that
n
P I}\~)
= i..J
"'
j\~>p<.~k)
k=1
I}
}}
'
where
p)~> =
1.
(38)
The intuitive meaning of the formula is clear: to go from state i to state j
inn steps, it is necessary to reach statej for the first time ink steps (1 ~ k ~ n)
and then to go from state j to state j in n  k steps. We now give a rigorous
derivation.
Let j be given and
' = min{l ~ k ~ n: ~k = j},
assuming that'! = n
+ 1 if {}
0. Then f!J> = P{T =
kl~o = i} and
P!~j> = P{~n =jl~o = i}
P{~n = j, '! = kl~o = i}
P{~r+nk=j,T=kl~o=i},
15k5n
15k5n
(39)
where the last equation follows because ~t+nk = ~n on the set {'! = k}.
Moreover, the set {' = k} = {' = k, ~r = j} for every k, 1 ~ k ~ n. Therefore
if P{~ 0 = i,' = k} > 0, it follows from Theorem 2 that
P{~r+nk =jl~o = i, '! = k} = P{~r+nk =jl~o = i, '! = k, ~t =j}
= P{~t+nk = j I~t = j} = ptfk>
128
I. Elementary Probability Theory
and by (37)
P!j> =
=
L P{~t+nk = jl~o = i,
7:
k= 1
= k}P{r = kl~o = i}
~ p(~k>:r(~)
L,
k=l
JJ
I] '
which establishes (38).
8.PROBLEMS
= (~ 0 , ... , ~.)be a Markov chain with values in X and f = f(x) (x EX) a
function. Will the sequence (f(~ 0 ), ... ,f(~.)) form a Markov chain? Will the
"reversed" sequence
1. Let ~
(~., ~. 1> ' ~o)
form a Markov chain?
2. Let IJll = IIP;)I, 1 :::; i,j:::; r, be a stochastic matrix and A. an eigenvalue of the matrix,
i.e. a root of the characteristic equation det IIIJll  A.E II = 0. Show that A.0 = 1 is an
eigenvalue and that all the other eigenvalues have moduli not exceeding 1. If all the
eigenvalues ..1. 1, , A., are distinct, then p~~l admits the representation
p~~l = ni
+ a;;{1)A.~ + + a;k)A.~,
where ni, a;;(!), ... , a;;{r) can be expressed in terms of the elements of IJll. (It follows
from this algebraic approach to the study of Markov chains that, in particular, when
I..1. 1 1< 1, ... , IA., I < 1, the limit lim Pl~l exists for every j and is independent of i.)
3.
Let~ = (~ 0 , . , ~.)be
tion matrix
IJll
a homogeneous Markov chain with state space X and transi
IIPxyll. Let
T<p(x) =
E[<p(~l)l~o = x]
(=
~ <p(y)Pxy)
Let the nonnegative function <p satisfy
T<p(x) = <p(x),
x EX.
Show that the sequence of random variables
is a martingale.
4. Let ~=(,.,Ill, IJll) and ~=(,.,Ill, IJll) be two Markov chains with different initial
distributions Ill = (p 1, , p,) and li = (ph ... , p,). Show that if min;.i Pii ::::: e > 0
then
L lftl"l p~"ll :::; 2(1 e)".
i=l
CHAPTER II
Mathematical Foundations of
Probability Theory
I. Probabilistic Model for an Experiment with
Infinitely Many Outcomes. Kolmogorov's Axioms
1. The models introduced in the preceding chapter enabled us to give a
probabilisticstatistical description of experiments with a finite number of
outcomes. For example, the triple (Q, d, P) with
and p(w) = pEaqnr.a, is a model for the experiment in which a coin is tossed
n times "independently" with probability p of falling head. In this model the
number N(Q) of outcomes, i.e. the number of points in n, is the finite
number 2n.
We now consider the problem of constructing a probabilistic model for
the experiment consisting of an infinite number of independent tosses of a
coin when at each step the probability offalling head is p.
It is natural to take the set of outcomes to be the set
n=
{co: w = (al, a2, .. .), ai = 0, 1},
i.e. the space of sequences w = (a 1, a 2 , . ) whose elements are 0 or 1.
What is the cardinality N(Q) of Q? It is well known that every number
a e [0, 1) has a unique binary expansion (containing an infinite number of
zeros)
(ai = 0, 1).
130
II. Mathematical Foundations of Probability Theory
Hence it is clear that there is a onetoone correspondence between the points
OJ ofQ and the points a of the set [0, 1), and therefore Q has the cardinality of
the continuum.
Consequently if we wish to construct a probabilistic model to describe
experiments like tossing a coin infinitely often, we must consider spaces Q
of a rather complicated nature.
We shall now try to see what probabilities ought reasonably to be assigned
(or assumed) in a model of infinitely many independent tosses of a fair coin
{p + q = !).
Since we may take Q to be the set [0, 1), our problem can be considered
as the problem of choosing points at random from this set. For reasons of
symmetry, it is clear that all outcomes ought to be equiprobable. But the
set [0, 1) is uncountable, and if we suppose that its probability is 1, then it
follows that the probability p(OJ) of each outcome certainly must equal
zero. However, this assignment of probabilities {p{OJ) = 0, OJ e [0, 1)) does
not lead very far. The fact is that we are ordinarily not interested in the
probability of one outcome or another, but in the probability that the result
of the experiment is in one or another specified set A of outcomes (an event).
In elementary probability theory we use the probabilities p(OJ) to find the
probability P(A) of the event A: P(A) =
p(OJ). In the present case, with
p(OJ) = 0, OJ E [0, 1), we cannot define, for example, the probability that a
point chosen at random from [0, 1) belongs to the set [0, !). At the same time,
it is intuitively clear that this probability should be!.
These remarks should suggest that in constructing probabilistic models for
uncountable spaces Q we must assign probabilities, not to individual outcomes but to subsets of Q. The same reasoning as in the first chapter shows
that the collection of sets to which probabilities are assigned must be closed
with respect to unions, intersections and complements. Here the following
definition is useful.
LroeA
Definition 1. Let Q be a set of points OJ. A system d of subsets of Q is called
an algebra if
(a) Qed,
(b) A, BEd=> Au BEd,
(c) Aed=>Aed
An Bed,
(Notice that in condition (b) it is sufficient to require only that either
A u B E d or that A n B E d, since A u B = A n Band A n B = A u B.)
The next definition is needed in formulating the concept of a probabilistic
model.
Definition 2. Let d be an algebra of subsets of Q. A set function J.l = J.l(A),
A e d, taking values in [0, oo ], is called a finitely additive measure defined
131
1. Probabilistic Model for an Experiment with Infinitely Many Outcomes
on .91 if
Jl(A
+ B) =
Jl(A)
+ Jl(B).
(1)
for every pair of disjoint sets A and B in .91.
A finitely additive measure Jl with Jl(Q) < oo is called finite, and when
Jl(Q) = 1 it is called a finitely additive probability measure, or a finitely
additive probability.
2. We now define a probabilistic model (in the extended sense).
Definition 3. An ordered triple (Q, .91, P), where
(a) Q is a set of points w;
(b) .91 is an algebra of subsets of Q;
(c) Pis a finitely additive probability on A,
is a probabilistic model in the extended sense.
It turns out, however, that this model is too broad to lead to a fruitful
mathematical theory. Consequently we must restrict both the class of subsets of Q that we consider, and the class of admissible probability measures.
Definition 4. A system ff of subsets of Q is a (1algebra if it is an algebra and
satisfies the following additional condition (stronger than (b) of Definition 1):
(b*) if An E ff, n = 1, 2, ... , then
(it is sufficient to require either that
u An
$' or that
n An
$').
Definition 5. The space Q together with a (1algebra ff of its subsets is a
measurable space, and is denoted by (Q, $').
Definition 6. A finitely additive measure Jl defined on an algebra .91 of subsets
of Q is countably additive (or (1additive), or simply a measure, if, for all
pairwise disjoint subsets A 1, A 2 , of A,
Jl(~l A") =
Jl
Jl(An>
A finitely additive measure Jl is said to be (1finite ifQ can be represented in
the form
n=l
with Jl(Qn) <
00,
n = 1, 2, ....
132
II. Mathematical Foundations of Probability Theory
If a countably additive measure P on the algebra A satisfies P(Q) = 1,
it is called a probability measure or a probability (defined on the sets that
belong to the algebra d).
Probability measures have the following properties.
If 0 is the empty set then
P(0) = 0.
If A, B E d then
P(A u B)= P(A)
If A, B E d and B
!:;;;;
+ P(B)
P(A n B).
A then
P(B) ::; P(A).
If A. E d, n = 1, 2, ... , and
U A. E d, then
The first three properties are evident. To establish the last one it is enough
to observe that
1 B., where B 1 =At> B.= A1 n n
1 A.=
A._ 1 n A., n ~ 2, B; n Bj = 0, i # j, and therefore
L:'=
L:'=
The next theorem, which has many applications, provides conditions
under which a finitely additive set function is actually countably additive.
Theorem. Let P be a finitely additive set function defined over the algebra d,
with P(Q) = 1. The following four conditions are equivalent:
(1) P is aadditive (Pis a probability);
(2) P is continuous from below, i.e. for any sets A 1 , A 2 , Ed such that
A. !:;;;; An+ 1 and
1 A. Ed,
U:'=
(3) Pis continuous from above, i.e. for any sets A 1 , A 2 ,
and
n:'=
A. E d,
such that A. 2 A.+ 1
133
I. Probabilistic Model for an Experiment with Infinitely Many Outcomes
(4) Pis continuous at 0, i.e. for any sets A1, A 2 , Ed such that An+ 1
and
1 An= 0,
n:=
lim P(An)
An
0.
Since
PROOF. (1) => (2).
00
UAn =
A1 + (Az \A1) + (A3 \Az) + ,
n=l
we have
PC91An)
P(A 1) + P(A 2 \A 1) + P(A 3\A 2 ) + ...
P(A 1) + P(A 2 )
P(A 1) + P(A 3) P(A 2 ) +
=lim P(An).
n
(2) => (3). Let n
1; then
The sequence {A 1\An}n> 1 of sets is nondecreasing (see the table in Subsection
3 below) and
00
00
U<Al\An)
n=l
A1\ nAn.
n=l
Then, by (2)
and therefore
n
P(Al) PC91(A1\An))
P(A1) P(A1)
+ P(Q1An)
P(A1) P(Al\0/n)
=
P(Q1An)
(3) => (4). Obvious.
(4) =>(1). Let Ab A 2 , .. Ed be pairwise disjoint and let
Then
L:'=
An Ed.
n=l
U A.
00
Af:,.B
A+B
A\B
0
AnB=0
An B (or AB)
AuB
A=!l\A
ff
A Eff
(J)
Notation
Table
union of the sets A 1 , A 2 ,
sum of sets, i.e. union of disjoint sets
difference of A and B, i.e. the set of points that belong to A
but not to B
symmetric difference of sets, i.e. (A \B) u (B\A)
outcome, sample point, elementary event
sample space; certain event
aalgebra of events
event (if wE A, we say that event A occurs)
event that A does not occur
event that either A or B occurs
element or point
set of points
aalgebra of subsets
set of points
complement of A, i.e. the set of points w that are not in A
union of A and B, i.e. the set of points w belonging either
to A or to B
intersection of A and B, i.e. the set of points w belonging to
both A and B
empty set
A and B are disjoint
event that at least one of A 1 , A 2 , occurs
event that A or B occurs, but not both
impossible event
events A and B are mutually exclusive, i.e. cannot occur
simultaneously
event that one of two mutually exclusive events occurs
event that A occurs and B does not
event that both A and B occur
Interpretation in probability theory
Settheoretic interpretation
;.
"'
...,
'<
(\)
<
l
=0
"'
.....,
...,
0
g"
P
"'=
.,e:..
r;
(\)
a::
.j;>.
v.>
A.
li~ 1 A.)
* i.o.
infinitely often.
(or lim inf A.)
limA.
lim A.
(or lim sup A.
or* {A. i.o.})
A li~ 1' A")
1A
(or
A.
(or
A. 1' A
n= 1
00
n A.
n=l
00
00
00
n= 1 k= n
00
un
the set
Ak
UAk
n= 1 k=n
00
the set
A 1 2 A 2 2 and A
=
n=1
00
n A"
n=l
U A.
the decreasing sequence of sets A" converges to A, i.e.
A 1 c;:; A 2 c;:; and A=
00
the increasing sequence of sets A. converges to A, I.e.
intersection of A~> A 2 ,
sum, i.e. union of pairwise disjoint sets A 1 , A 2 , .
A~>
A2,
occur
event that all the events A 1 , A 2 , occur with the possible
exception of a finite number of them
event that infinitely many of events A 1 , A 2 .. occur
the decreasing sequence of events converges to event A
the increasing sequence of events converges to event A
event that all the events
event that one of the mutually exclusive events A 1 , A 2 ,
occurs
,.,
Vo
! j.)
0
c:
;:;
'<
Ol
a::
=
.:;
::n
=
g.
:.
&
"
(1)
$....
Ol
0'
....
g.
0.
a::
0
r;
0
cr"
::,0
'e:
136
II. Mathematical Foundations of Probability Theory
and since L~n+ 1 A; !
0, n +
oo, we have
3. We can now formulate Kolmogorov's generally accepted axiom system,
which forms the basis for the concept of a probability space.
Fundamental Definition. An ordered triple (Q, fi', P) where
(a)
is a set of points w,
(b) fi' is a ualgebra of subsets of n,
(c) Pis a probability on fi',
is called a probabilistic model or a probability space. Here n is the sample space
or space of elementary events, the sets A in ff are events, and P(A) is the
probability of the event A.
It is clear from the definition that the axiomatic formulation of probability
theory is based on set theory and measure theory. Accordingly, it is useful to
have a table (pp. 134135) displaying the ways in which various concepts are
interpreted in the two theories. In the next two sections we shall give examples
of the measurable spaces that are most important for probability theory and
of how probabilities are assigned on them.
4.
PROBLEMS
n = {r: r E [0, 1]} be the set of rational points of [0, 1], d the algebra of sets
each of which is a finite sum of disjoint sets A of one of the forms {r: a < r < b},
{r: a::;; r < b}, {r: a< r::;; b}, {r: a::;; r::;; b}, and P(A) = b a. Show that P(A),
A Ed, is finitely additive set function but not countably additive.
1. Let
2. Let Q be a countable set and :!F the collection of all its subsets. Put JJ(A) = 0 if A is
finite and JJ(A) = oo if A is infinite. Show that the set function JJ is finitely additive
but not countably additive.
3. Let JJ be a finite measure on a ualgebra :!F, A. E :!F, n = 1, 2, ... , and A
(i.e., A = lim. A. = lim. A.). Show that JJ(A) = lim. JJ(A.).
4. Prove that P(A 6 B)
= P(A) + P(B) 
2P(A n B).
= lim. A.
137
2. Algebras and aAlgebras. Measurable Spaces
5. Show that the "distances" p 1(A, B) and p2(A, B) defined by
p 1(A, B)= P(A
~B),
P(A ~B)
{
piA, B) = :(A v B)
ifP(A v B)# 0,
if P(A v B)= 0
satisfy the triangle inequality.
6. Let f.1. be a finitely additive measure on an algebra d, and let the sets A 1 , A 2 , Ed
be pairwise disjoint and satisfy A =
1 f..I.(A;).
1 Ai Ed. Then f.J.(A) ~
Ir;
Ir;
7. Prove that
lim sup An= lim inf A.,
lim inf A. ~ lim sup A.,
lim inf An = lim sup An,
lim sup(A. v B.) = lim sup A. v lim sup B.,
lim sup A. n lim inf B. ~ lim sup( A. n B.) ~ lim sup A. n lim sup B.
If A. j A or A.
A, then
lim inf An= lim sup A.
8. Let {x.} be a sequence of numbers and A.= ( oo, xn). Show that x =lim sup x.
and A = lim sup A. are related in the following way: ( oo, x) ~ A ~ ( oo, x].
In other words, A is equal to either ( oo, x) or to ( oo, x].
9. Give an example to show that if a measure takes the value
general that countable additivity implies continuity at 0.
+ oo, it does not follow in
2. Algebras and aAlgebras. Measurable Spaces
1. Algebras and aalgebras are the components out of which probabilistic
models are constructed. We shall present some examples and a number of
results for these systems.
Let n be a sample space. Evidently each of the collections of sets
~* =
{0, Q},
~*
= {A: A:; Q}
is both an algebra and a aalgebra. In fact, ~* is trivial, the "poorest"
aalgebra, whereas ~* is the "richest" aalgebra, consisting of all subsets
ofn.
When n is a finite space, the aalgebra ~* is fully surveyable, and commonly serves as the system of events in the elementary theory. However, when
the space is uncountable the class ~* is much too large, since it is impossible
to define "probability" on such a system of sets in any consistent way.
If A :; Q, the system
~A = {A,
A, 0, Q}
138
II. Mathematical Foundations of Probability Theory
is another example of an algebra (and a aalgebra), the algebra (or aalgebra)
generated by A.
This system of sets is a special case of the systems generated by decompositions. In fact, let
~
= {D 1 , D 2 , . }
be a countable decomposition of Q into nonempty sets:
Q = D1
+ D2 + ;
D; n Di =
0,
i =F j.
Then the system d = a(~), formed by the sets that are unions of finite
numbers of elements of the decomposition, is an algebra.
The following lemma is particularly useful since it establishes the important
principle that there is a smallest algebra, or aalgebra, containing a given
collection of sets.
Lemma 1. Let tff be a collection of subsets of n. Then there are a smallest
algebra a(&) and a smallest aalgebra a(&) containing all the sets that are inS.
The class !F* of all subsets of Q is a aalgebra. Therefore there are at
least one algebra and one aalgebra containing S. We now define a(S)
(or a(S)) to consist of all sets that belong to every algebra (or aalgebra)
containing&. It is easy to verify that this system is an algebra (or aalgebra)
and indeed the smallest.
PROOF.
Remark. The algebra r:x(E) (or a(E), respectively) is often referred to as the
smallest algebra (or aalgebra) generated by S.
We often need to know what additional conditions will make an algebra,
or some other system of sets, into a aalgebra. We shall present several results
of this kind.
Definition 1. A collection .H of subsets of Q is a monotonic class if An E .A,
n = 1, 2, ... , together with An j A or An ! A, implies that A E .H.
Let tff be a system of sets. Let f.l(S) be the smallest monotonic class containing&. (The proof of the existence of this class is like the proof of Lemma 1.)
Lemma 2. A necessary and sufficient condition for an algebra d to be a
aalgebra is that it is a monotonic class.
PRooF. A aalgebra is evidently a monotonic class. Now let d be a monotonic
class and An Ed, n = 1, 2, .... It is clear that Bn =
1 A; Ed and
Bn Bn+ 1 . Consequently, by the definition of a monotonic class,
Bn j U~ 1 A; Ed. Similarly we could show that n~ 1 A; Ed.
Ui'=
By using this lemma, we can prove that, starting with an algebra d, we
can construct the aalgebra a(d) by means of monotonic limiting processes.
139
2. Algebras and aAlgebras. Measurable Spaces
Theorem 1. Let .91 be an algebra. Then
f.l( .91) = a(.91).
(1)
PRooF. By Lemma 2, f.l(d) ~ a(d). Hence it is enough to show that f.l(d)
is a aalgebra. But At = f.l(d) is a monotonic class, and therefore, by Lemma
2 again, it is enough to show that f.l(d) is an algebra.
Let A E .A; we show that A E .A. For this purpose, we shall apply a
principle that will often be used in the future, the principle of appropriate sets,
which we now illustrate.
Let
.Ji =
{B:Be.A,Be.A}
be the sets that have the property that concerns us. It is evident that
.91 ~ .Ji ~ At. Let us show that .Ji is a monotonic class.
Let Bn E .Ji; then Bn E .A, Bn E .A, and therefore
lim j Bn E .A,
Consequently
l Bn E .A, lim l Bn = lim j Bn E .A,
lim l Bn = lim j Bn E A,
lim j Bn = lim l Bn E .A,
and therefore .A is a monotonic class. But .1i ~ At and At is the smallest
monotonic class. Therefore .ii = At, and if A EAt = f.l(d), then we also
lim j Bn = lim
have A E .A, i.e. At is closed under the operation of taking complements.
Let us now show that At is closed under intersections.
Let A EAt and
AlA= {B:Be.A,A n Be A}.
From the equations
lim
l (A n Bn) = A n
lim
l Bn,
lim j (A n Bn) = A n lim j Bn
it follows that At A is a monotonic class.
Moreover, it is easily verified that
(A
AIB)<=> (B
AI A).
(2)
Now let A E .91; then since .91 is an algebra, for every B E .91 the set
A n B E .91 and therefore
,s;l ~ A(A ~AI.
But At A is a monotonic class (since lim i ABn = A lim i Bn and lim l ABn =
A lim l Bn), and At is the smallest monotonic class. Therefore AtA = At for
all A E .91. But then it follows from (2) that
(A
AI B) <=> (B
AIA
At).
140
II. Mathematical Foundations of Probability Theory
whenever A E .91 and BE .A. Consequently if A E .91 then
for every BE .A. Since A is any set in .91, it follows that
Therefore for every B E .A
.I{B
=.If,
i.e. if B E .A and C E .A then C n B E .A.
Thus .A is closed under complementation and intersection (and therefore
under unions). Consequently .A is an algebra, and the theorem is established.
Definition 2. Let Q be a space. A class~ of subsets ofQ is adsystem if
(a) Q E ~;
(b) A, B, E ~' A s B = B\A E
(c) An E ~'An SAn+ 1 = UAn
If~
~;
E
~.
is a collection of sets then
d(~)
denotes the smallest dsystem con
taining~.
Theorem 2. If the collection ~ of sets is closed under intersections, then
(3)
d(~) = a(~)
PROOF. Every aalgebra is adsystem, and consequently d(~) s::; a(~). Hence
if we prove that d(~) is closed under intersections, d(~) must be a aalgebra
and then, of course, the opposite inclusion a(~) s d( ~) is valid.
The proof once again uses the principle of appropriate sets.
Let
~ 1 = {BEd(~):BnAEd(~)forallAE~}.
If B E ~ then B n A E ~ for all A E ~ and therefore ~ s ~ 1 . But ~ 1 is a
dsystem. Hence d(~) s ~ 1 On the other hand, ~ 1 s d(~) by definition.
Consequently
Now let
~2
{BEd(~):
Bn A
Ed(~)
for all A
Ed(~)}.
Again it is easily verified that ~ 2 is adsystem. If B E ~'then by the definition
of ~ 1 we obtain that B n A Ed(~) for all A E ~ 1 = d(~). Consequently
~ s ~ 2 and d(~) s ~ 2 But d(~) 2 ~ 2 ; hence d(~) = ~ 2 , and therefore
141
2. Algebras and aAlgebras. Measurable Spaces
whenever A and Bare in d(G), the set A n B also belongs to d(G), i.e. d(G) is
closed under intersections.
This completes the proof of the theorem.
We next consider some measurable spaces (Q, :F) which are extremely
important for probability theory.
2. The measurable space (R, PA(R)). Let R = ( oo, oo) be the real line and
= {xeR:a
(a,b]
< x:::;; b}
for all a and b,  oo :::;; a < b < oo. The interval (a, oo] is taken to be (a, oo ).
(This convention is required if the complement of an interval ( oo, b] is
to be an interval of the same form, i.e. open on the left and closed on the
right.)
Let d be the system of subsets of R which are finite sums of disjoint
intervals of the form (a, b]:
n
d if A =
(ai, b;],
i= 1
n < oo.
It is easily verified that this system of sets, in which we also include the
empty set 0, is an algebra. However, it is not a aalgebra, since if An =
(0, 1  1/n] Ed, we have
An= (0, 1) =d.
Let PA(R) be the smallest aalgebra a(d) containing d. This aalgebra,
which plays an important role in analysis, is called the Borel algebra of subsets
of the real line, and its sets are called Borel sets.
Iff is the system of intervals f of the form (a, b], and a(f) is the smallest
aalgebra containing J, it is easily verified that a(J) is the Borel algebra.
In other words, we can obtain the Borel algebra from J without going
through the algebra d, since a(J) = a(iX(J)).
We observe that
Un
(a, b)=
U(a, b !],
n
n=l
n(a!, b],
{a}= n(a!, a].
[a,
b] =
n=l
n=l
a< b,
a< b,
Thus the Borel algebra contains not only intervals (a, b] but also the singletons {a} and all sets of the six forms
(a, b),
[a, b],
[a, b),
( oo, b),
( oo, b],
(a, oo).
(4)
142
II. Mathematical Foundations of Probability Theory
Let us also notice that the construction of 81(R) could have been based on
any of the six kinds of intervals instead of on (a, b], since all the minimal
aalgebras generated by systems of intervals of any of the forms (4) are the
same as 81(R).
Sometimes it is useful to deal with the aalgebra 81(R) of subsets of the
extended real lineR = [  oo, oo]. This is the smallest aalgebra generated by
intervals of the form
(a, b] = {x E R: a < x::::;; b},
 oo ::::;; a < b ::::;; oo,
where ( oo, b] is to stand for the set {x E R:  oo ::::;; x::::;; b}.
Remark 1. The measurable space (R, 81(R)) is often denoted by (R, 81) or
(R 1 , 81 1).
Remark 2. Let us introduce the metric
lx Yl
p 1(x, y) = 1 + lx Yl
on the real line R (this is equivalent to the usual metric Ix  y I) and let
810 (R) be the smallest aalgebra generated by the open sets SP(x 0 ) =
{x E R: p 1(x, x 0 ) < p}, p > 0, x 0 E R. Then 81 0 (R) = 81(R) (see Problem 7).
3. The measurable space (R", ?4(R")). Let R" = R x x R be the direct, or
Cartesian, product of n copies of the real line, i.e. the set of ordered ntuples
x = (x 1 , . , x,), where  oo < xk < oo, k = 1, ... , n. The set
where Ik = (ak, bk], i.e. the set {x E R": xk E Ik> k = 1, ... , n}, is called a
rectangle, and I k is a side of the rectangle. Let J be the set of all rectangles I.
The smallest aalgebra a(J) generated by the system J is the Borel algebra
of subsets of R" and is denoted by 81(R"). Let us show that we can arrive at
this Borel algebra by starting in a different way.
Instead ofthe rectangles I = I 1 x x I, let us consider the rectangles
B = B 1 x x B, with Borel sides (Bk is the Borel subset of the real line
that appears in the kth place in the direct product R x x R). The smallest
aalgebra containing all rectangles with Borel sides is denoted by
81(R) 81(R)
and called the direct product of the aalgebras 81(R). Let us show that in fact
143
2. Algebras and uAlgebras. Measurable Spaces
In other words, the smallest aalgebra generated by the rectangles I =
I 1 x x I" and the (broader) class of rectangles B = B 1 x x B" with
Borel sides are actually the same.
The proof depends on the following proposition.
n, and define
Lemma 3. Let C be a class of subsets ofO., let B
C n B = {A n B: A E C}.
(5)
u(C n B) = u(C) n B.
(6)
Then
PROOF. Since C u(C), we have
C n B u(C) n B.
(7)
But u(C) n B is a aalgebra; hence it follows from (7) that
u(C n B) u(C) n B.
To prove the conclusion in the opposite direction, we again use the
principle of appropriate sets.
Define
1:6'8 = {A E u(C): An BE u(C n B)}.
Since u( C) and u(C n B) are aalgebras, 1:6'8 is also a aalgebra, and evidently
C 1:6'8 u(C),
whence u(C) a(!ffi'8 ) = 1:6'8 u(C) and therefore u(C) = 1:6'8 . Therefore
A n BE u(C n B)
for every A u(C), and consequently u(C) n B u(cS' n B).
This completes the proof of the lemma.
Proof that PJ(R") and PJ PJ are the same. This is obvious for n = 1.
We now show that it is true for n = 2.
Since PJ(R 2 ) PJ PJ, it is enough to show that the Borel rectangle
B 1 x B 2 belongs to PJ(R 2 ).
Let R 2 = R 1 x R 2 , where R 1 and R 2 are the "first" and "second" real
lines, ~ 1 = PJ 1 x R 2 , ~ 2 = R 1 x PJ 2 , where PJ 1 x R 2 (or R 1 x PJ 2 ) is the
collectionofsetsoftheformB 1 x R 2 (orR 1 x B2 ),withB 1 EPJ 1 (orB 2 EPJ 2 ).
Also let f 1 and f 2 be the sets of intervals in R 1 and R 2 , and .11 = f 1 x R 2 ,
.J2 = R 1 x f 2 Then, by (6),
B 1 x B2
B1 n B2 E ~ 1 n
~ 2 = u(.J 1 ) n
= u(.J 1 n B2 )
= u(f 1
as was to be proved.
B2
x f
2 ),
u(.J 1 n
.J2 )
144
II. Mathematical Foundations of Probability Theory
The case of any n, n > 2, can be discussed in the same way.
Remark. Let BI 0 (R") be the smallest aalgebra generated by the open sets
Sp(x 0 ) = {x E R": Pn(x, x 0) < p},
x 0 E R",
p > 0,
in the metric
n
Pn(x, x 0 ) =
L rkp1(xk, x~).
k=1
where x = (xl> ... , xn), x 0 = (x~, ... , x~).
Then BI 0 (Rn) = BI(R") (Problem 7).
4. The measurable space (R 00 , BI(R 00 ) ) plays a significant role in probability
theory, since it is used as the basis for constructing probabilistic models of
experiments with infinitely many steps.
The space R 00 is the space of ordered sequences of numbers,
 oo < xk < oo, k
1, 2, ...
Let I k and Bk denote, respectively, the intervals (ak, bk] and the Borel subsets
of the kth line (with coordinate xk). We consider the cylinder sets
J(J 1
J(B 1
X X
X X
In)= {x:x = (x 1,x 2 , .),x 1 E1 1, ,XnEin},
(8)
Bn) = {x:x = (x 1, X2 ), x 1 E B 1, , Xn E Bn},
(9)
J(B")
{x: (x1, ... , Xn) E B"},
(10)
where B" is a Borel set in 91(R"). Each cylinder J{B 1 x x Bn), or J(B"),
can also be thought of as a cylinder with base in R"+ 1 , R"+ 2 , . , since
J(B1
X X
Bn)
J(B")
= J(B1 X
= J(B"+1),
Bn
R),
where B"+ 1 = B" x R.
It follows that both systems of cylinders J(B 1 x x Bn) and J(B")
are algebras. It is easy to verify that the unions of disjoint cylinders
J(J 1
X X
In)
also form an algebra. Let 91(R 00 ), 91 1(R 00 ) and f!I 2 (R 00 ) be the smallest
aalgebras containing all the sets (8), (9) or (10), respectively. (The aalgebra
91 1(R 00 ) is often denoted by f!I(R) 91(R) x .)It is clear that f!I(R 00 )!;;;;
f!I 1 (R 00)!;;;; BI 2 (R 00). As a matter of fact, all three aalgebras are the same.
To prove this, we put
CfJn =
{A E R": {x: (x1, ... , Xn) E A} E BI(R 00 ) }
for n = 1, 2, .... Let B" e f!I(R"). Then
B"
E CfJn !;;;;
f!I(R 00 ).
145
2. Algebras and aAlgebras. Measurable Spaces
But
t(J n
is a (Jalgebra, and therefore
consequently
P4z(R 00 )
PJ(R 00 ).
Thus PJ(R 00 ) = P4 1(R 00 ) = P4 z(R 00 ).
From now on we shall describe sets in PJ(R 00 ) as Borel sets (in R 00 ).
Remark. Let P4 0 (R 00 ) be the smallest (Jalgebra generated by the open sets
= {x E Roo: Poo(x, x 0 ) < p},
Sp(x 0 )
p > 0,
x 0 E R 00 ,
in the metric
Poo(x,
X 0)
00
L rkpl(xb xf),
k=l
where x = (xt>x 2 , ... ), x = (x?,x~, ... ). Then
(Problem 7).
Here are some examples of Borel sets in Roo:
P4(R 00 ) = P4 0 (R 00 )
(a) {xER 00 :supxn>a},
{x E Roo: inf Xn <a};
(b) {x E Roo: lim
Xn ~a},
{x E Roo: lim
Xn
>a},
where, as usual,
Urn Xn
inf sup Xm,
n
lim Xn = sup inf Xm;
n
m~n
m~n
(c) {x E Roo: xn +}, the set of x E Roo for which lim xn exists and is finite;
(d) {x E R 00 : lim Xn >a};
(e) {X E R 00 :
1 IXn I > a} ;
(f) {x E Roo:
1 xk = 0 for at least one n ;::: 1}.
L:'=
Lk=
To be convinced, for example, that sets in (a) belong to the system PJ(R 00 ),
it is enough to observe that
{x: sup
Xn
> a} =
U {x: Xn >
n
{x: inf Xn <a} =
a} E P4(R 00 ),
U{x: Xn <a} E P4(R
00
).
5. The measurable space (Rr, PJ(RT)), where Tis an arbitrary set. The space
RT is the collection of real functions x = (x 1) defined for t E Tt. IIi general
we shall be interested in the case when Tis an uncountable subset of the real
t We shall also use the notations x = (x,),.Rr and x = (x,), t ERr, for elements of Rr.
146
II. Mathematical Foundations of Probability Theory
line. For simplicity and definiteness we shall suppose for the present that
T = [0, oo).
We shall consider three types of cylinder sets
X . X
In)= {x: x,,
J 1,, ... ,tn(B 1
X X
Bn) =
EB 1, ... ,X1nEBn},
{x:X11
Ib
x,n E I 1},
J,,, ... ,tn(J 1
. '
J,,, ... ,r"(Bn) = {x: (x,,, ... , x,J E Bn},
(11)
(12)
(13)
where Ik is a set of the form (ak, bk], Bk is a Borel set on the line, and Bn is a
Borel set in Rn.
The set J,,, ... ,r" (I 1 x x In) is just the set of functions that, at times
t 1 , ,tn, "get through the windows" I~>In and at other times have
arbitrary values (Figure 24).
Let BI(RT), 91 1(RT) and 91 2 (RT) be the smallest aalgebras corresponding
respectively to the cylinder sets (11), (12) and (13). It is clear that
(14)
As a matter of fact, all three of these aalgebras are the same. Moreover, we
can give a complete description of the structure of their sets.
Theorem 3. LetT be any uncountable set. Then BI(RT) = 91 1(RT) = 91 2 (RT),
and every set A E P.l(RT) has the following structure: there are a countable set of
points tt> t 2 , ofT and a Borel set Bin fJI(R such that
00 )
(15)
PRooF. Let t! denote the collection of sets of the form (15) (for various aggregates (t 1 , t 2 , ..) and Borel sets Bin BI(R 00 )). If A1 , A 2 , ... Et! and the
corresponding aggregates are r< 1> = (t~ll, t~1 >, ), r< 2 >= (t~2 >, t~2 >, ), . ,
Of+,,+,, Figure 24
147
2. Algebras and aAlgebras. Measurable Spaces
then the set r<oo> =
representation
Uk r<k> can be taken as a basis, so that every A<il has a
A;= {x: (x,,,
Xr 2 ,
B;},
where B; is a set in one and the same aalgebra :?4(R 00 ), and r; E r<oo>.
Hence it follows that the system C is a aalgebra. Clearly this aalgebra
contains all cylinder sets of the form (1) and, since PA 2 (RT) is the smallest
aalgebra containing these sets, and since we have (14), we obtain
(16)
Let us consider a set A from C, represented in the form (15). For a given
aggregate (t 1 , t 2 , ), the same reasoning as for the space (R 00 , :?I(R 00 )) shows
that A is an element of the aalgebra generated by the cylinder sets ( 11 ). But
this aalgebra evidently belongs to the aalgebra :?I(RT); together with (16),
this established both conclusions of the theorem.
Thus every Borel set A in the aalgebra :?I(RT) is determined by restrictions
imposed on the functions x = (x1), t E T, on an at most countable set of points
t 1, t 2 , Hence it follows, in particular, that the sets
A 1 = {x: sup x 1 < C for all t
[0, 1]},
A 2 = {x: x 1 = 0 for at least one t
A 3 = {x:
X1
[0, 1]},
is continuous at a given point t0
[0, 1]},
which depend on the behavior of the function on an uncountable set of points,
cannot be Borel sets. And indeed none of' these three sets belongs to 84(RI 0 1l).
Let us establish this for A 1 . If A 1 E PA(R 10 1l), then by our theorem there
are a point (t?, t~, . .. ) and a set B 0 E :?I(R 00 ) such that
It is clear that the function y1 C  1 belongs to
(Yr? ... ) E B 0 . Now form the function
21
{c 1,
A~>
and consequently
t E (t?, t~, .. .),
= C + 1, t (t?, t~, .. .).
It is clear that
and consequently the function z = (z1) belongs to the set {x: (x 1?, ... } E B0 }.
But at the same time itisclearthatitdoesnot belong to the set {x: sup X 1 < C}.
This contradiction shows that A 1 84(R 10 1l).
148
II. Mathematical Foundations of Probability Theory
Since the sets A 1 , A 2 and A 3 are nonmeasurable with respect to the
0'algebra dl[R[o, 11) in the space of all functions x = (x,), t e [0, 1], it is
natural to consider a smaller class of functions for which these sets are
measurable. It is intuitively clear that this will be the case if we take the
intial space to be, for example, the space of continuous functions.
6. The measurable space (C, dl(C)). LetT = [0, 1] and let C be the space of
continuous functions x = (x,), 0 s:; t s:; 1. This is a metric space with the
metric p(x, y) = SUPteT lx, y,l. We introduce two 0'algebras in C:
dl(C) is the 0'algebra generated by the cylinder sets, and d6' 0 (C) is generated
by the open sets (open with respect to the metric p(x, y)). Let us show that in
fact these 0'algebras are the same: dl(C) = d6' 0 (C).
Let B = {x: x 10 < b} be a cylinder set. It is easy to see that this set is open.
Hence it follows that {x: x 11 < b1 , , x 1" < bn} E d6' 0(C), and therefore
dl(C) s;; d6'0 (C).
Conversely, consider a set B P = {y: y e S p(x 0 )} where x 0 is an element of C
and Sp(x 0 ) = {x e C: SUPreTix1  x?i < p} is an open ball with center at
x 0 Since the functions in C are continuous,
BP
{y E C:
ye SP(x0)} = {y E C: m~x IYr x?i < P}
=
n{y
C: IYtk x~l < p}
dl(C),
(17)
lk
where tk are the rational points of [0, 1]. Therefore d6' 0 (C) s;; dl(C).
The following example is fundamental.
7. The measurable space (D, dl(D)), where Dis the space offunctions x = (x1),
t e [0, 1], that are continuous on the right (x 1 = x,+ for all t < 1) and have
limits from the left (at every t > 0).
Just as for C, we can introduce a metric d(x, y) on D such that the 0'algebra
.16'0 (D) generated by the open sets will coincide with the 0'algebra dl(D)
generated by the cylinder sets. This metric d(x, y), which was introduced
by Skorohod, is defined as follows:
d(x, y) = inf{e
> 0:3 A. e A: sup lx, YA<t> +sup It A.(t)i s:; e},
t
(18)
where A is the set of strictly increasing functions A. = A.(t) that are continuous
on [0, 1] and have A.(O) = 0, A.(1) = 1.
8. The measurable space CflreT Or, IlreT ~). Along with the space
(RT, dl(RT)), which is the direct product ofT copies of the real line together
with the system of Borel sets, probability theory also uses the measurable
space COteT n" RteT :F,), which is defined in the following way.
149
3. Methods of Introducing Probability Measures on Measurable Spaces
Let T be any set of indices and (Q0 ~1) a measurable space, t E T. Let
il10 the set of functions w = (w 1), t E T, such that w 1 E il1 for each
OreT
Q =
tE T.
The collection of cylinder sets
.F1,, ... , 1.(B 1
X X
B.)= {w: W 11
B 1, , w 1
B.},
where B1; E ~r; is easily shown to be an algebra. The smallest aalgebra
T ~ 1 , and the measurable
containing all these cylinder sets is denoted by
space
Qi,
~1) is called the direct product of the measurable spaces
(Q0 ~ 1 ), t E T.
<0
9.
Plre
PI
PROBLEMS
1. Let ffl 1 and f!l 2 be aalgebras of subsets of n. Are the following systems of sets aalgebras?
ffl 1 n ffl 2
= {A: A e ffl 1 and A e ffl 2 },
ffl 1 u ffl 2 ={A: A ef!l 1 or A ef!l 2 }.
2. Let C!fi = { D 1 , D 2 , } be a countable decomposition of n and f!l =
a(~).
Are there
also only countably many sets in f!l?
3. Show that
f!l(R") f!l(R) = f!l(R"+ 1).
4. Prove that the sets (b)(f) (see Subsection 4) belong to f!l(R 00 ).
5. Prove that the sets A 2 and A 3 (see Subsection 5) do not belong to f!l(RIO, 11).
6. Prove that the function (15) actually defines a metric.
7. Prove that 96 0 (R")
96(R"), n ~ 1, and fli 0 (R"') = f.!I(R"').
8. Let C = C[O, oo) be the space of continuous functions x = (x,) defined for t
Show that with the metric
p(x, y) =
I z min[ sup lx, y,l, 1],
n= 1
0.
x,yeC,
Ostsn
this is a complete separable metric space and that the aalgebra f!l 0 (C) generated by
the open sets coincides with the aalgebra f!l(C) generated by the cylinder sets.
3. Methods of Introducing Probability Measures
on Measurable Spaces
1. The measurable space (R, r!4(R)). Let P = P(A) be a probability measure
defined on the Borel subsets A of the real line. Take A = ( oo, x] and put
F(x)
= P( oo, x],
ER.
(1)
150
II. Mathematical Foundations of Probability Theory
This function has the following properties:
(1) F(x) is nondecreasing;
(2) F( oo) = 0, F( + oo) = 1, where
F( oo) = lim F(x),
F( + oo) = lim F(x);
xf oo
X~ 00
(3) F(x) is continuous on the right and has a limit on the left at each x
R.
The first property is evident, and the other two follow from the continuity
properties of probability measures.
Definition 1. Every function F = F(x) satisfying conditions (1)(3) is called
a distribution function (on the real line R).
Thus to every probability measure P on (R, &b(R)) there corresponds (by
(1)) a distribution function. It turns out that the converse is also true.
Theorem 1. Let F = F(x) be a distribution function on the real lineR. There
exists a unique probability measure P on (R, &b(R)) such that
P(a, b] = F(b) F(a)
(2)
for all a, b,  oo :::;; a < b < oo.
PRooF. Let d be the algebra of the subsets A of R that are finite sums of
disjoint intervals of the form (a, b]:
A =
L (ak, bk].
k= 1
On these sets we define a set function P0 by putting
n
P 0 (A) =
L [F(bk) F(ak)],
A Ed.
(3)
k=l
This formula defines, evidently uniquely, a finitely additive set function on d.
Therefore if we show that this function is also countably additive on this
algebra, the existence and uniqueness of the required measure P on &b(R)
will follow immediately from a general result of measure theory (which we
quote without proof).
CarathOOdory's Theorem. Let n be a space, d an algebra of its subsets, and
FA = a{d) the smallest aalgebra containing d. Let J.l.o be a afinite measure on
(0, A). Then there is a unique measure J.l. on (Q, a(d)) which is an extension
of J.l.o, i.e. satisfies
J.l.(A) = J.l.o(A),
A Ed.
3. Methods of Introducing Probability Measures on Measurable Spaces
151
We are now to show that P 0 is countably additive on .91. By a theorem
from 1 it is enough to show that P 0 is continuous at 0. i.e. to verify that
LetA 1 , A 2 , .. be a sequence of sets from .91 with the property An! 0. Let
us suppose first that the sets An belong to a closed interval [ N, N], N < oo.
Since A is the sum of finitely many intervals of the form (a, b] and since
P0 (a', b)= F(b) F(a') t F(b) F(a)
P0 (a, b]
as a' ! a, because F(x) is continuous on the right, we can find, for every An,
a set Bn E .91 such that its closure [BnJ An and
Po(An) P 0 (Bn)
e 2n,
where e is a preassigned positive number.
[BnJ = 0. But the sets [Bn]
By hypothesis, nAn= 0 and therefore
are closed, and therefore there is a finite n0 = n0 (e) such that
(4)
n= 1
(In fact, [ N, N] is compact, and the collection of sets {[ N, N]\[BnJ}n~l
is an open covering of this compact set. By the HeineBore! theorem there
is a finite subcovering:
no
U([ N, N]\[Bn]) =
[  N,
N]
n=1
and therefore n~~ 1 [BnJ = 0).
Using (4) and the inclusions Ano Ano 1 A1o we obtain
Po(Ano)
Po(Ano\01 Bk)
= Po(Ano\01
Bk)
+ Po(01 Bk)
~ Po(9 (Ak\Bk))
1
no
no
k=1
k=1
L Po(Ak\Bk) ~ L e 2k ~e.
Therefore P 0 (An)! 0, n t oo.
We now abandon the assumption that An [  N, N] for some N. Take
an e > 0 and choose N so that P 0 [ N, N] > 1  e/2. Then, since
An = An n [ N, N]
+ An n
[ N, N],
~e have
P 0 (A")
=
~
P 0 (An[ N, N] + P 0 (An n [ N, N])
P 0 (An n [ N, N]) + e/2
152
II. Mathematical Foundations of Probability Theory
and, applying the preceding reasoning (replacing An by An n [ N, N]), we
find that P0 (An n [ N, N]) s:; e/2 for sufficiently large n. Hence once again
P 0 (An)! 0, n+ oo. This completes the proof ofthe theorem.
Thus there is a onetoone correspondence between probability measures
P on (R, fJI(R)) and distribution functions F on the real line R. The measure
P constructed from the function F is usually called the LebesgueStieltjes
probability measure corresponding to the distribution function F.
The case when
F(x) =
0, X< 0,
{ x, 0 s:; x s:; 1,
1, X> 1.
is particularly important. In this case the corresponding probability measure
(denoted by A.) is Lebesgue measure on [0, 1]. Clearly A.(a, b] = b a. In
other words, the Lebesgue measure of (a, b] (as well as of any ofthe intervals
(a, b), [a, b] or [a, b)) is simply its length b a.
Let
fJI([O, 1]) = {A n [0, 1]: A E fJI(R)}
be the collection of Borel subsets of [0, 1]. It is often necessary to consider,
besides these sets, the Lebesgue measurable subsets of [0, 1]. We say that a
set A [0, 1] belongs to ~([0, 1)] if there are Borel sets A and B such that
A A Band A.(B\A) = O.lt is easily verified that .si([O, 1]) is aualgebra.
It is known as the system of Lebesgue measurable subsets of [0, 1]. Clearly
fJI([O, 1]) ,sj([O, 1]).
The measure A., defined so far only for sets in .?1([0, 1]), extends in a
natural way to the system ,sj([O, 1]) of Lebesgue measurable sets. Specifically,
if Ae,sj([O, 1]) and A A B, where A and Be~([O, 1]) and A.(B\A) = 0,
we define A:(A) = A.(A). The set function A: = A:(A), A e ~([0, 1]), is easily
seen to be a probability measure on ([0, 1], ~([0, 1])). It is usually called
Lebesgue measure (on the system of Lebesguemeasurable sets).
Remark. This process of completing (or extending) a measure can be applied,
and is useful, in other situations. For example, let (Q, fF, P) be a probability
space. Let P be the collection of all the subsets A of Q for which there are
sets B 1 and B 2 of fF such that B 1 A B 2 and P(B 2 \B 1) = 0. The probability measure can be defined for sets A e #'Pin a natural way (by P(A) =
P(B 1)). The resulting probability space is the completion of (Q, fF, P) with
respect to P.
A probability measure such that P = fF is called complete, and the corresponding space (Q, fF, P) is a complete probability space.
The correspondence between probability measures P and distribution
functions F established by the equation P(a, b] = F(b) F(a) makes it
153
3. Methods of Introducing Probability Measures on Measurable Spaces
F(x)
I
I ~F(x3)
I
1
~F(xz)
I~F(x,)
x,
x2
XJ
Figure 25
possible to construct various probability measures by obtaining the corresponding distribution functions.
Discrete measures are measures P for which the corresponding distributions F = F(x) are piecewise constant (Figure 25), changing their values
at the points x 1 , x 2 , (L'1F(xJ > 0, where L'1F(x) = F(x) F(x ). In
this case the measure is concentrated at the points x 1 , x 2 , .. :
The set of numbers (p 1 , p2, .. .), where Pk = P( {xk}), is called a discrete
probability distribution and the corresponding distribution function F = F(x)
is called discrete.
We present a table of the commonest types of discrete probability distribution, with their names.
Table 1
Distribution
Probabilities Pk
Parameters
Discrete uniform
Bernoulli
Binomial
1/N, k = 1,2, ... ,N
Pt = p, Po= q
C~pkq\ k = 0, 1, ... , n
N = 1,2, ...
0 ::; p ::; 1, q = 1  p
0 ::; p ::; 1, q = 1  p,
n = 1,2, ...
Poisson
Geometric
Negative binomial
l 1p, k =
q=lP'l',
ekfk!, k
0, 1, .. .
0, 1, .. .
k = r,r + 1, ...
1>0
0 ::; p ::; 1, q = 1  p
0 ::; p ::; 1, q = 1  p,
r = 1, 2, ...
Absolutely continuous measures. These are measures for which the corresponding distribution functions are such that
F(x) = roof(t) dt,
(5)
154
II. Mathematical Foundations of Probability Theory
where f = f(t) are nonnegative functions and the integral is at first taken in
the Riemann sense, but later (see 6) in that of Lebesgue.
The function f = f(x), x E R, is the density of the distribution function
F = F(x) (or the density of the probability distribution, or simply the density)
and F = F(x) is called absolutely continuous.
It is clear that every nonnegative f = f(x) that is Riemann integrable and
such that J~ cxJ(x) dx = 1 defines a distribution function by (5). Table 2
presents some important examples of various kinds of densities f = f(x)
with their names and parameters (a density f(x) is taken to be zero for values
of x not listed in the table).
Table 2
Uniform on [a, b]
Normal or Gaussian
Parameters
Density
Distribution
lj(b  a),
a,beR; a<b
a ::::; x ::::; b
(27ta)li2e<xm)2;(2a2)'
X E
meR,a>O
xlexiP
Gamma
,
r(IX}P"
IX> 0, f3 > 0
x~O
r>O,s>O
Beta
Exponential (gamma
with IX = 1, f3 = 1/A.)
Bilateral exponential
Chisquared, x2
(gamma with a
IX = n/2, f3 = 2)
Student, t
Cauchy
x2) en+ 1)/2
r(!{n + 1)) (
, x
(nn)112r(nj2) 1 + ;;
(m/n)mf2
xm/21
B(m/2, n/2) (1
0
n(x
+ mxjn)<m+n)/Z
+ 02)
1, 2, ...
1, 2, ...
m, n
1, 2, ...
xER
Singular measures. These are measures whose distribution functions are
continuous but have all their points of increases on sets of zero Lebesgue
measure. We do not discuss this case in detail; we merely give an example of
such a function.
3. Methods of Introducing Probability Measures on Measurable Spaces
155
Figure 26
We consider the interval [0, 1] and construct F(x) by the following procedure originated by Cantor.
We divide [0, 1] into thirds and put (Figure 26)
F2(x) =
(t, 1),
z,1
X E
xe(i,~),
(~, !),
X= 0,
4'
4'
0,
X E
1, x=l
defining it in the intermediate intervals by linear interpolation.
Then we divide each of the intervals [0, and [i, 1] into three parts and
define the function (Figure 27) with its values at other points determined by
linear interpolation.
tJ
Figure 27
156
II. Mathematical Foundations of Probability Theory
Continuing this process, we construct a sequence of functions Fn(x),
n = 1, 2, ... , which converges to a nondecreasing continuous function F(x)
(the Cantor function), whose points of increase (xis a point of increase of F(x)
+ e)  F(x  e) > 0 for every e > 0) form a set of Lebesgue measure
zero. In fact, it is clear from the construction of F(x) that the total length of
the intervals (j, ~), (!, ~), (~, !), ... on which the function is constant is
if F(x
! + ~ + _i_ + ... = !
3
27
f (~)n = 1.
3 n=O 3
(6)
Let % be the set of points of increase of the Cantor function F(x). It
follows from (6) that A.(%) = 0. At the same time, if Jl. is the measure corresponding to the Cantor function F(x), we have JJ.(%) = 1. (We then say
that the measure is singular with respect to Lebesgue measure A..)
Without any further discussion of possible types of distribution functions,
we merely observe that in fact the three types that have been mentioned cover
all possibilities. More precisely, every distribution function can be represented
in the form p 1 F 1 + p2 F 2 + p 3 F 3 , where F 1 is discrete, F 2 is absolutely
continuous, and F 3 is singular, and P; are nonnegative numbers, p1 + p2 +
P3 = 1.
2. Theorem 1 establishes a onetoone correspondence between probability
measures on (R, Ei(R)) and distribution functions on R. An analysis of the
proof of the theorem shows that in fact a stronger theorem is true, one that in
particular lets us introduce Lebesgue measure on the real line.
Let Jl. be a ufinite measure on (Q, d), where d is an algebra of subsets of
n. It turns out that the conclusion of Caratheodory's theorem on the extension of a measure and an algebra d to a minimal ualgebra u(d) remains
valid with aufinite measure; this makes it possible to generalize Theorem 1.
A LebesgueStieltjes measure on (R, Ei(R)) is a (countably additive)
measure Jl. such that the measure JJ.(I) of every bounded interval I is finite.
A generalized distribution junction on the real line R is a nondecreasing
function G = G(x), with values on ( oo, oo), that is continuous on the right.
Theorem 1 can be generalized to the statement that the formula
JJ.(a, b] = G(b) G(a),
a< b,
again establishes a onetoone correspondence between LebesgueStieltjes
measures Jl. and generalized distribution functions G.
In fact, if G( + oo)  G( oo) < oo, the proof of Theorem 1 can be taken
over without any change, since this case reduces to the case when G( + oo) G( oo) = 1 and G( oo) = 0.
Now let G( + oo) G( oo) = oo. Put
G(x),
Gn(x) = { G(n)
G( n),
lxl ~ n,
n,
x = n.
X =
157
3. Methods of Introducing Probability Measures on Measurable Spaces
On the algebra d let us define a finitely additive measure JLo such that
JLo(a, b] = G(b)  G( a), and let Jln be the finitely additive measure previously
constructed (by Theorem 1) from GnCx).
Evidently Jln j JLo on d. Now let A 1, A 2 , be disjoint sets in d and
A
LAnE d. Then (Problem 6 of 1)
00
JLo(A) ~ L JLo(An).
n=1
2:.%
If
1 JLo(An) = oo then JLo(A) = L:'= 1 JLo(An). Let us suppose that
L JLo(An) < oo. Then
00
jt 0 (A)
= lim JLn(A) = lim
n
L JLnCAk).
k=1
By hypothesis, L JLo(An) < oo. Therefore
0
~ jt 0(A) k~/ 0 (Ak) = li~ L~1 (JLnCAk) JLo(Ak))] ~ 0,
since Jln ~ JLo.
Thus a afinite finitely additive measure JLo is countably additive on d,
and therefore (by Caratheodory's theorem) it can be extended to a countably
additive measure J1 on a(d).
The case G(x) = xis particularly important. The measure Acorresponding
to this generalized distribution function is Lebesgue measure on (R, P4(R)).
As for the interval [0, 1] of the real line, we can define the system fJ(R) by
writing A E Pi(R) if there are Borel sets A and B such that A ~ A ~ B,
A(B\A) = 0. Then Lebesgue measure 1 on P4(R) is defined by l(A) = A(A)
if A ~ A ~ B, A E Pi(R) and A(B\A) = 0.
3. The measurable space (W, P4(R"). Let us suppose, as for the real line, that
Pis a probability measure on (R", P4(W).
Let us write
FnCx1, ... , Xn)
P(( 00,
x1J X X (  00,
Xn]),
or, in a more compact form,
Fn(x)
P( 00,
x],
where X = (x 1, ... , Xn), ( 00, X] = ( 00, X1J X X
Let us introduce the difference operator Lla,. b,: R"
formula
Lla,,b,Fn(X1, ... 'xn)
(  00,
+
Xn].
R, defined by the
Fn(xl, ... ) X; b b;, X;+ 1 ...)
 FnCx~> ... , X; 1, ai, X;+ 1 ...)
158
II. Mathematical Foundations of Probability Theory
where ai :S bi. A simple calculation shows that
da 1b 1 da"b"Fn(X1 Xn)
P(a, b],
(7)
where (a, b] = (a 1, b1 ] x x (an, bJ. Hence it is clear, in particular, that
(in contrast to the onedimensional case) P(a, b] is in general not equal to
Fn(b) Fn(a).
Since P(a, b] ~ 0, it follows from (7) that
(8)
for arbitrary a = (a 1, , an), b = (b 1, ... , bn).
It also follows from the continuity of P that Fn(x 1, ... , xn) is continuous
on the right with respect to the variables collectively, i.e. if x<k> Lx, x<k> =
(k)) th en
( x(k)
1 , .. ,xn ,
k+
(9)
00.
It is also clear that
(10)
and
lim Fn(Xb ... , Xn) = 0,
(11)
x!y
if at least one coordinate of y is  oo.
Definition 2. An ndimensional distribution function (on Rn) is a function
F = F(x 1, , Xn) with properties (8)(11).
The following result can be established by the same reasoning as in
Theorem 1.
Theorem 2. Let F = Fn(x 1, ... , Xn) be a distribution function on Rn. Then there
is a unique probability measure P on (Rn, PJ(Rn)) such that
(12)
Here are some examples of ndimensional distribution functions.
Let F 1, , pn be onedimensional distribution functions (on R) and
F n(X 1 , Xn) = F 1(x 1) F"(xJ.
It is clear that this function is continuous on the right and satisfies (10) and
(11). It is also easy to verify that
da,b, ... danbJn(X1, ... , Xn)
n [Fk(bk) 
Consequently F n(X 1 , x,.) is a distribution function.
Fk(ak)] ~ 0.
3. Methods of Introducing Probability Measures on Measurable Spaces
159
The case when
xk
< 0,
0 ~ xk ~ 1,
xk > 1
is particularly important. In this case
The probability measure corresponding to this ndimensional distribution
function is ndimensional Lebesgue measure on [0, 1]".
Many ndimensional distribution functions appear in the form
Fn(xl, , Xn) =
f~
f~
J,.{tl, . , tn) dtl dtn,
where J,.(t 1, ... , tn) is a nonnegative function such that
f_
00
00
f_
00
00
J,.(t h
... ,
tn) dt 1 dtn = 1,
and the integrals are Riemann (more generally, Lebesgue) integrals. The
function f = J,.(t 1, ... , tn) is called the density of the ndimensional distribution function, the density of the ndimensional probability distribution,
or simply anndimensional density.
When n = 1, the function
_ 1 e (xm) 2 /(2a 2 )
f( x ) 
u~
xeR,
'
with u > 0 is the density of the (nondegenerate) Gaussian or normal distribution. There are natural analogs of this density when n > 1.
Let~= llriill be a nonnegative definite symmetric n x n matrix:
n
L:
i,j= 1
rip~).i ;;::: 0,
A.i e R, i = 1, ... , n,
When ~is a positive definite matrix, I ~I = det
is an inverse matrix A = llaiill.
IAI 112
J,.(x~o ... , Xn) = ( 2n)"12 exp{ !
> 0 and consequently there
L: aiixi m)(xi mi)},
(13)
where mi e R, i = 1, ... , n, has the property that its (Riemann) integral over
the whole space equals 1 (this will be proved in 13) and therefore, since it is
also positive, it is a density.
This function is the density of thendimensional (nondegenerate) Gaussian
or normal distribution (with vector mean m = {m 1, ... , mn) and covariance
matrix~ =A  1).
160
II. Mathematical Foundations of Probability Theory
Figure 28. Density of the twodimensional Gaussian distribution.
When n = 2 the density f 2 (x 1 , x 2 ) can be put in the form
1
(14)
where ui > 0, IpI < 1. (The meanings of the parameters mi, ui and p will be
explained in 8.)
Figure 28 indicates the form of the twodimensional Gaussian density.
Remark. As in the case n = 1, Theorem 2 can be generalized to (similarly
defined) LebesgueStieltjes measures on (R", PI(R")) and generalized
distribution functions on R". When the generalized distribution function
Gn(x1o ... , Xn) is x 1 Xn, the corresponding measure is Lebesgue measure
on the Borel sets of R". It clearly satisfies
n(bi n
A.(a, b) =
ai),
i=l
i.e. the Lebesgue measure of the "rectangle"
is its "content."
4. The measurable space (R 00 , 91(R 00 )). For the spaces R", n ~ 1, the probaability measures were constructed in the following way: first for elementary
sets (rectangles (a, b]), then, in a natural way, for sets A = (ai, bJ, and
finally, by using Caratheodory's theorem, for sets in PI(R").
A similar construction for probability measures also works for the space
(Roo, 91(R 00 )).
3. Methods of Introducing Probability Measures on Measurable Spaces
161
Let
BE 14(R"),
denote a cylinder set in Roo with base BE &I(R"). We see at once that it is
natural to take the cylinder sets as elementary sets in R 00 , with their probabilities defined by the probability measure on the sets of &I(R 00 ).
Let P be a probability measure on (R 00 , &I(R 00 )). For n = 1, 2, ... , we
take
BE &I(R").
(15)
The sequence of probability measures P t> P 2 , defined respectively on
(R, &I(R)), (R 2 , &I(R 2 )), , has the following evident consistency property:
for n = 1, 2, ... and B E &I(R"),
(16)
It is noteworthy that the converse also holds.
Theorem 3 (Kolmogorov's Theorem on the Extension of Measures in
(R 00 , &B(R 00 ))). Let P 1 , P 2 , be a sequence of probability measures on
(R, 14(R)), (R 2 , &B(R 2 )), , possessing the consistency property (16). Then
there is a unique probability measure P on (R 00 , &I(R 00 )) such that
BE &I(R").
(17)
for n = 1, 2, ....
E &B(R") and let Jn(B") be the cylinder with base B". We assign
the measure P(J.(B")) to this cylinder by taking P(J.(B")) = P.(B").
PRooF. Let B"
Let us show that, in virtue of the consistency condition, this definition is
consistent, i.e. the value of P(Jn(B")) is independent of the representation of
the set Jn(B"). In fact, let the same cylinder be represented in two way:
J.(B") = Jn+k(B"+k).
It follows that, if (x 1, ... , xn+k)
R"+k, we have
(18)
and therefore, by (16) and (18),
Pn(B") = pn+ l((xb ... 'Xn+ l):(xl, ... 'xn) E B")
= ... = pn+k((xl, ... , Xn+k):(xl, ... 'x.) E B")
= pn+k(B"+k).
Let Jii(R 00 ) denote the collection of all cylinder sets B" = J nCB"), B" e14(R"),
n = 1, 2, ....
162
II. Mathematical Foundations of Probability Theory
Now let B 1, ... , Bk be disjoint sets in d(R 00 ). We may suppose without loss
of generality that B; = .fiBi), i = 1, ... , k, for some n, where B~, ... , Bi: are
disjoint sets in BI(Rn). Then
i.e. the set function Pis finitely additive on the algebra d(R 00 ).
Let us show that P is "continuous at zero," i.e. if the sequence of sets
Bn ! 0, n + oo, then P(Bn) + 0, n + oo. Suppose the contrary, i.e. let
lim P(BJ = J > 0. We may suppose without loss of generality that {Bn}
has the form
We use the following property of probability measures Pn on (Rn, BI(Rn))
(see Problem 9): if Bn E BI(Rn), for a given J > 0 we can find a compact set
An E BI(Rn) such that An c;; Bn and
Therefore if
we have
Form the set
en = n;:= 1 Ak and let en be such that
en= {x: (x1, ... , xn) E Cn}.
Then, since the sets Bn decrease, we obtain
P(Bn\en) ~
k= 1
k= 1
L P(Bn\Ak) ~ L P(Bk\Ak) ~ J/2.
But by assumption limn P(Bn) = J > 0, and therefore limn P(en) ~ J/2 > 0.
Let us show that this contradicts the condition en! 0.
. c~n Th en (X1(n), ... , Xn(n)) E c n
A(n)  (X1(n), X2(n) , ... ) Ill
. tX
Let USC hOOSe a pom
for n ~ 1.
Let (nd be a subsequence of (n) such that x~nJ)+ x?, where x? is a point
incl. (Such a sequence exists since x~n) E cl and cl is compact.) Then select
a subsequence (n 2) of (n 1 ) such that (An 2 >, x~ 2 >)+ (x?, xg) E C 2. Similarly let
(x~nk), ... , x1nkl)+ (x?, ... , xZ) E ck. Finally form the diagonal sequence
(mk), where mk is the kth term of (nk) Then x!mkl+ x? as mk+ oo fori = 1, 2, ... ;
and(x?, xg, ...) E e"forn = 1, 2, ... , whichevidentlycont radictstheassumption that en! 0, n+ oo. This completes the proof of the theorem.
163
3. Methods of Introducing Probability Measures on Measurable Spaces
Remark. In the present case, the space Roo is a countable product of lines,
R 00 = R x R x . It is natural to ask whether Theorem 3 remains true if
(R 00 , ?#(R 00 )) is replaced by a direct product of measurable spaces (Q;, ~i),
i = 1, 2, ....
We may notice that in the preceding proof the only topological property
of the real line that was used was that every set in ?#(R") contains a compact
subset whose probability measure is arbitrarily close to the probability
measure of the whole set. It is known, however, that this is a property not only
of spaces (R", 1l(R"), but also of arbitrary complete separable metric spaces
with aalgebras generated by the open sets.
Consequently Theorem 3 remains valid if we suppose that P 1 , P 2 , is a
sequence of consistent probability measures on (Q 1 , ~ 1 ),
where (Qi, ~) are complete separable metric spaces with aalgebras
generated by open sets, and (R 00 , ?#(R 00 )) is replaced by
(Ql X Q2 X ". ", ~ {8) ~2 {8) "" ).
In 9 (Theorem 2) it will be shown that the result of Theorem 3 remains
valid for arbitrary measurable spaces (Qi, ~) if the measures Pn are concentrated in a particular way. However, Theorem 3 may fail in the general
case (without any hypotheses on the topological nature of the measurable
spaces or on the structure of the family of measures {P nD This is shown by
the following example.
Let us consider the space Q = (0, 1], which is evidently not complete, and
construct a sequence f71 ~ f72 ~ of aalgebras in the following way. For
n = 1, 2, ... , let
<fJn
0 < w < 1/n,
0, 1/n ~ W ~ 1,
(w) = {1,
rcn = {A E Q: A = { w: <fJn(w) E B}, BE ?#(R)}
and let ,; = a{CC 1 , .. , CC"} be the smallest aalgebra containing the sets
eel, ... ' rcn. Clearly~l ~ ~2 ~ Let ~ = u(U ,;) be the smallest
aalgebra containing all the ,;. Consider the measurable space (Q, ,;)
and define a probability measure Pn on it as follows:
Pn{w. (<p 1(w), ... , <fJn(w)) E B} 
{1O
1)
if(l, ... ,
h
.
ot erwzse,
B",
where B" = ?#(R"). It is easy to see that the family {Pn} is consistent: if
A E ,;then Pn+ 1 (A) = Pn(A). However, we claim that there is no probability
measure P on (Q, ~) such that its restriction PI,; (i.e., the measure P
164
II. Mathematical Foundations of Probability Theory
considered only on sets in g;,) coincides with Pn for n = 1, 2, .... In fact, let
us suppose that such a probability measure P exists. Then
P{w: q> 1(w) = = q>n(w) = 1} = Pn{w: q> 1(w) = = q>n(w) = 1} = 1
(19)
for n = 1, 2, .... But
{w: q> 1 (w) = = q>n(w) = 1} = (0, 1/n)
t 0,
which contradicts (19) and the hypothesis of countable additivity (and therefore continuity at the "zero" 0) of the set function P.
We now give an example of a probability measure on (R 00 , PJ(R 00 )). Let
F 1 (x), F 2 (x), ... be a sequence of onedimensional distribution functions.
Define the functions G(x) = F 1 (x), Gix 1 , x 2 ) = F 1 (x 1 )F ix 2 ), ... , and denote
the corresponding probability measures on (R, PJ(R)), (R 2 , PJ(R 2 )), . by
P 1 , P 2 , Then it follows from Theorem 3 that there is a measure P on
(R 00 , PJ(R 00 )) such that
P{x E R 00 : (x 1 , . . . , Xn) E B} = Pn(B),
BE PJ(Rn)
and, in particular,
P{xER 00 :X 1
~ab,Xn~an}
=F 1(a1)Fn(an).
Let us take Fi(x) to be a Bernoulli distribution,
0,
F;(x) = { q,
1,
X< 0,
0:::::; x < 1,
X;::: 1.
Then we can say that there is a probability measure P on the space Q of
sequencesofnumbersx = (x 1 , x 2 , . ),xi= Oorl,togetherwiththecralgebra
of its Borel subsets, such that
P{x . X 1 _ a 1' ' x n _a } _ pra,qnra,
n
This is precisely the result that was not available in the first chapter for
stating the law of large numbers in the form (1.5.8).
5. The measurable space (RT, PJ(RT)). LetT be a set of indices t E T and R 1 a
real line corresponding to the index t. We consider a finite unordered set
r = [t 1, . , tnJ of distinct indices ti, tiE T, n ;::: 1, and let Pr be a probability
measure on (Rr, PJ(Rt)), where Rr = R1, x x R1".
We say that the family {Pt} of probability measures, where r runs through
all finite unordered sets, is consistent if, for all sets r = [t 1 , . . . , tn] and
a = [s 1 , . . . , sk] such that a ~ r we have
Pcr{(x.,, ... , x.J:(x.,, ... , x.J E B} = Pt{(X11 ,
x 1J:(x.,, ... , X 5k) E B}
(20)
for every BE PJ(R").
3. Methods of Introducing Probability Measures on Measurable Spaces
165
Theorem 4 (Ko1mogorov's Theorem on the Extension of Measures in
(RT, &U(RT))). Let {P,} be a consistent family of probability measures on
(R', :J#(R')). Then there is a unique probability measure P on (RT, &U(RT)) such
that
P(f.(B)) = P.(B)
(21)
for all unordered sets< = [t 1 , , tnJ of different indices t; E T, BE &U(R') and
f.(B) = {x E RT: (X11 , , x,J E B}.
PROOF. Let the set BE &U(RT). By the theorem of 2 there is an at most countable setS= {s 1 ,s 2 , ... } s; T such that B = {x:(X 51 ,X 52 , )EB}, where
B E 8U(R 8 ), R 8 = R51 x R52 x . In other words, B = f 8 (B) is a cylinder
set with base B E 8U(R 8 ).
We can define a set function P on such cylinder sets by putting
P(f8 (B)) = P8 (B),
(22)
where P 8 is the probability measure whose existence is guaranteed by
Theorem 3. We claim that Pis in fact the measure whose existence is asserted
in the theorem. To establish this we first verify that the definition (22) is
consistent, i.e. that it leads to a unique value of P(B) for all possible representations of B; and second, that this set function is countably additive.
Let B = f 8 /B 1 ) and B = f 8 ,(B 2 ). It is clear that then B = fs 1 us/B 3 )
with some B 3 E 8U(R 81 us 2 ) ; therefore it is enough to show that if S s; S'
and B E 8U(R 8 ), then P 8 .(B') = P 8 (B), where
B'
= {(X
8 1 ,
X 82 , ):(X 81
X 82 ,
E B}
with S' = {s~, s~, .. .}, S = {s 1 , s 2 , }. But by the assumed consistency of
(20) this equation follows immediately from Theorem 3. This establishes that
the value of P(B) is independent of the representation of B.
To verify the countable additivity of P, let us suppose that {En} is a sequence
of pairwise disjoint sets in &U(RT). Then there is an at most countable set
S s; T such that Bn = f 8 (Bn) for all n 2 1, where Bn E 8U(R 8 ). Since P 8
is a probability measure, we have
P(L Bn) = P(L: fs(Bn)) = Ps(L Bn) = L Ps(Bn)
= L P(Is(Bn)) = L P(Bn).
Finally, property (21) follows immediately from the way in which P was
constructed.
This completes the proof.
Remark 1. We emphasize that Tis any set of indices. Hence, by the remark
after Theorem 3, the present theorem remains valid if we replace the real
lines R, by arbitrary complete separable metric spaces !11 (with ualgebras
generated by open sets).
166
II. Mathematical Foundations of Probability Theory
Remark 2. The original probability measures {Pr} were assumed defined
on unordered sets r = [t b ... , tn] of different indices. It is also possible to
start from a family of probability measures {Pr} where r runs through all
ordered sets r = (t 1 , . , tn) of different indices. In this case, in order to have
Theorem 4 hold we have to adjoin to (20) a further consistency condition:
where (i, ... , in) is an arbitrary permutation of (1, ... , n) and A 1, E ?.6(R 1J As
a necessary condition for the existence of P this follows from (21) (with
P11 ,, ... , 1JB) replaced by P<t,, ... ,1JB)).
From now on we shall assume that the sets r under consideration are
unordered. If Tis a subset of the real line (or some completely ordered set),
we may assume without loss of generality that the set r = [tb ... , tn]
satisfies t 1 < t 2 < < tn. Consequently it is enough to define "finitedimensional" probabilities only for sets r = [t 1 , ... , tnJ for which t 1 <
t2 < ... < tn.
Now consider the case T = [0, oo ). Then RT is the space of all real functions x = (x 1) 1 ;:o A fundamental example of a probability measure on
(RIO. ool, 86(RI0 ool)) is Wiener measure, constructed as follows.
Consider the family {cpr{yjx)} 1 ;:o of Gaussian densities (as functions of y
for fixed x):
yeR,
and for each r = [t 1 , , tnJ,
B
t1
< t 2 < < tn, and each set
11 X X
In,
construct the measure P.(B) according to the formula
PrCJ1
=
X X
In)
J(it ... J( IPt.(aliO)<pt
In
t.(a2ial) .. <p1"t"_,(anlanl)dal"dan
(24)
(integration in the Riemann sense). Now we define the set function P for each
cylinder set f 1, ... 1Jf 1 x x In) = {x E RT: X11 E 11, ... , X 1" E Jn} by taking
P(f11 ... tn(J 1
X X
Jn)) =
Pitt ...
1if1
X X
In).
The intuitive meaning of this method of assigning a measure to the cylinder
set ft, ... In(I 1 X ... X In) is as follows.
The set f 11 ... 1"(1 1 x x In) is the set of functions that at times t 1 , ... , tn
pass through the" windows" 1 1 , , In (see Figure 24 in 2). We shall interpret
3. Methods of Introducing Probability Measures on Measurable Spaces
167
<p 1k 1k_Jaklak_ 1) as the probability
tk  tk_ 1, arrives in a neighborhood
that a particle, starting at ak 1 at time
of ak. Then the product of densities that
appears in (24) describes a certain independence of the increments of the
displacements of the moving "particle" in the time intervals
The family of measures {Pr} constructed in this way is easily seen to be
consistent, and therefore can be extended to a measure on (RIO, ool, ?8(RI 0 ool)).
The measure so obtained plays an important role in probability theory. It
was introduced by N. Wiener and is known as Wiener measure.
6.
PROBLEMS
1. Let F(x)
P( oo, x]. Verify the following formulas:
P(a, b]
= F(b) F(a),
P(a, b)= F(b) F(a),
P[a, b] = F(b) F(a ),
P[a, b)= F(b) F(a ),
P{x} = F(x) F(x),
where F(x) = limy 1 x F(y).
2. Verify (7).
3. Prove Theorem 2.
4. Show that a distribution function F = F(x) on R has at most a countable set of
points of discontinuity. Does a corresponding result hold for distribution functions
on R"?
5. Show that ea:ch of the functions
G(x, y) =
1,
0,
G(x, y) = [x
X+
y;;?: 0,
< 0,
+y
+ y], the integral part of x + y,
is continuous on the right, and continuous in each argument, but is not a (generalized)
distribution function on R 2
6. Let fl. be the LebesgueStieltjes measure generated by a continuous distribution
function. Show that if the set A is at most countable, then fl.(A) = 0.
7. Let c be the cardinal number of the continuum. Show that the cardinal number of the
collection of Borel sets in R" is c, whereas that of the collection of Lebesgue measurable sets is
zc.
8. Let (Q, ', P) be a probability space and d an algebra of subsets of Q such that
u(d) = !F. Using the principle of appropriate sets, prove that for every e > 0 and
B E : there is a set A E d such that
P(A !:::,. B) :::; e.
168
II. Mathematical Foundations of Probability Theory
9. Let P be a probability measure on (R", !I(R")). Using Problem 8, show that, for
every e > 0 and B e !I(R"), there is a compact subset A of !I(R") such that A ~ B
and
P(B\A) ;s; e.
(This was used in the proof of Theorem 1.)
10. Verify the consistency of the measure defined by (21).
4. Random Variables. I
1. Let (0, ~) be a measurable space and let (R, BI(R)) be the real line with
the system BI(R) of Borel sets.
Definition 1. A real
function~
~(w)
function, or a random variable, if
defined on (0, F) is an
{w: ~(w) E B}
ff'
~measurable
(1)
for every Be BI(R); or, equivalently, if the inverse image
~ 1 (B)
= {w: ~(w) E B}
is a measurable set in 0.
When (0, ~ = (R", BI(R")), the &I(R")measurable functions are called
Borel functions.
The simplest example of a random variable is the indicator IA(w) of an
arbitrary (measurable) set A E ff'.
A random variable ~ that has a representation
~(w)
00
L x;IA;(w),
i= 1
(2)
where L Ai = 0, Ai E ~. is called discrete. If the sum in (2) is finite, the
random variable is called simple.
With the same interpretation as in 4 of Chapter I, we may say that a
random variable is a numerical property of an experiment, with a value
depending on "chance." Here the requirement (1) of measurability is fundamental, for the following reason. If a probability measure P is defined on
(0, ~), it then makes sense to speak of the probability of the event {~(w) E B}
that the value of the random variable belongs to a Borel set B.
We introduce the following definitions.
Definition 2. A probability measure P~ on (R, BI(R)) with
P~(B)
= P{w: ~(w) E B},
BE BI(R),
is called the probability distribution of~ on (R, BI(R)).
169
4. Random Variables. I
Definition 3. The function
F~(x)
= P(w:
::::; x},
~(w)
XER,
is called the distribution function of ~For a discrete random variable the measure P~ is concentrated on an at
most countable set and can be represented in the form
P ~(B)
p(xk),
(3)
{k: XkEB)
wherep(xk) = P{~ = xd = M~(xk).
The converse is evidently true: If P ~ is represented in the form (3) then ~
is a discrete random variable.
A random variable~ is called continuous if its distribution function F~(x)
is continuous for x E R.
A random variable~ is called absolutely continuous if there is a nonnegative
function f = Hx), called its density, such that
F~(x) = fJ~(y) dy,
XER,
(4)
(the integral can be taken in the Riemann sense, or more generally in that of
Lebesgue; see 6 below).
2. To establish that a function ~ = ~(w) is a random variable, we have to
verify property (1) for all sets BE~ The following lemma shows that the
class of such "test" sets can be considerably narrowed.
Lemma 1. Let g be a system of sets such that a( g) = l(R). A necessary and
sufficient condition that a function~ = ~(w) is ~measurable is that
(5)
for all E
g_
PRooF. The necessity is evident. To prove the sufficiency we again use the
principle of appropriate sets.
Let~ be the system of those Borel sets D in f!.6(R) for which C 1(D) E ~
The operation "form the inverse image" is easily shown to preserve the settheoretic operations of union, intersection and complement:
yB") = yC
cl( 0B")
0Cl(B"),
C 1(
(B"),
~ l(B") =
C l(B").
(6)
170
II. Mathematical Foundations of Probability Theory
It follows that f0 is a aalgebra. Therefore
and
a(S)
a(f0) = f0
PJ(R).
But a{E) = PJ(R) and consequently f0 = PJ(R).
Corollary. A necessary and sufficient condition for
variable is that
for every x
for every x
< x}
IF
x}
IF
{w:
~(w)
{w:
~(w) ~
= ~(w) to be a random
R, or that
R.
The proof is immediate, since each of the systems
S 1 = {x: x < c, c E R},
S2 =
{X:
c, c E R}
generates the aalgebra PJ(R): a{ 1 ) = a(E 2 ) = f!J(R) (see 2).
The following lemma makes it possible to construct random variables as
functions of other random variables.
Lemma 2. Let cp = cp(x) be a Bore/function and~= ~(w) a random variable.
Then the composition 17 = cp o (, i.e. the function 17(w) = cp(((w)), is also a
random variable.
The proof follows from the equations
{w:1](W)EB} = {w:cp(((w))EB} = {w:~(w)Ecp 1 (B)}EIF
(7)
for BE PJ(R), since cp 1(B) E PJ(R).
Therefore if~ is a random variable, so are, for examples,~,~+ = max((, 0),
= min(~' 0), and In since the functions x", X+' X and IX I are Borel
functions (Problem 4).
3. Starting from a given collection of random variables {~.},we can construct
new functions, for example,
1 I~k I, lim ~., lim ~., etc. Notice that in
general such functions take values on the extended real line R = [ oo, oo].
Hence it is advisable to extend the class of IFmeasurable functions somewhat
by allowing them to take the values oo.
Lk'=
171
4. Random Variables. I
Definition 4. A function ~ = ~(w) defined on (0, ) with values in R =
[ oo, oo] will be called an extended random variable if condition (1) is
satisfied for every Borel set BE PJ(R).
The following theorem, despite its simplicity, is the key to the construction
of the Lebesgue integral (6).
Theorem 1.
~ = ~(w) (extended ones included) there is a
sequence of simple random variables ~ 1 , ~ 2 , . , such that I~" I : : ; I~ I and
~n(w)+ ~(w), n+ oo,for all wE 0.
(b) If also ~(w) ~ 0, there is a sequence of simple random variables ~t> ~ 2 , . ,
such that ~n(w) i ~(w), n+ oo, for all wE 0.
(a) For every random variable
PRooF. We begin by proving the second statement. For n = 1, 2, ... , put
n2" k 1
~n(w) = k~1 ~Ik,n(w)
+ nlww>~n)(w),
where Ik,n is the indicator of the set {(k  1)/2" ::::;; ~(w) < k/2"}. It is easy
to verify that the sequence ~"(w) so constructed is such that ~"(w) i ~(w)
for all w E 0. The first statement follows from this if we merely observe that
~can be represented in the form~ = ~+  ~.This completes the proof of
the theorem.
We next show that the class of extended random variables is closed under
pointwise convergence. For this purpose, we note fir~t that if ~ 1 , ~ 2 , .. is a
sequence of extended random variables, then sup~"' inf ~"'lim~" and lim~"
are also random variables (possibly extended). This follows immediately from
{w:sup ~n > x} =
{w: inf ~n
< x} =
U{w: ~n > x} E,
n
U {w: ~n < x} E ,
n
and
lim ~n = inf sup ~m
n m2:.n
Theorem 2. Let
~(w)
~t> ~ 2 ,
lim ~n
= sup sup ~m.
n
m2:.n
be a sequence of extended random variables and
= lim ~n(w). Then ~(w) is also an extended random variable.
The prooffollows immediately from the remark above and the fact that
{w:
~(w)
< x}
= {w: lim ~n(w) < x}
= {w: lim ~n(w) =lim ~n(w)} 11 {lim ~n(w) < x}
<X} = {lim ~nCw) <X} E .
= 011 {lim ~n(w)
172
II. Mathematical Foundations of Probability Theory
4. We mention a few more properties of the simplest functions of random
variables considered on the measurable space (Q, /F) and possibly taking
values on the extended real lineR = [  oo, oo].t
If e and '1 are random variables, e + '1 e  '1 e'1. and e/'1 are also random
variables (assuming that they are defined, i.e. that no indeterminate forms like
oo  oo, oojoo, a/0 occur.
In fact, let {en} and {1'/n} be sequences of random variables converging to
e and '1 (see Theorem 1). Then
en
'1n + e '1.
en '1n +
en
~+
1
1'/n + J{.,n=O}(w)
~1'/.
e

'1
The functions on the lefthand sides of these relations are simple random
variables. Therefore, by Theorem 2, the limit functions e 17, e11 and e/rt
are also random variables.
5. Let be a random variable. Let us consider sets from :F of the form
{w: e(w) E B}, BE PA(R). It is easily verified that they form a aalgebra,
called the aalgebra generated by and denoted by ~If qJ is a Borel function, it follows from Lemma 2 that the function 17 = qJ o e
is also a random variable, and in fact ~measurable, i.e. such that
e.
{w: 17(w) E B} E ~.
BE PA(R)
(see (7)). It turns out that the converse is also true.
Theorem 3. Let '1 be a :F~measurable random variable. Then there is a Borel
jimction qJ such that '1 = qJ
i.e. 1J(W) = qJ(e(w)) for every WE Q.
0
e,
PROOF. Let Cl> be the class of ~measurable functions 1J = 1J(w) and &~
the class of ~measurable functions representable in the form qJ 0 where
qJ is a Borel function. It is clear that & s; Cl>~. The conclusion of the theorem
is that in fact&~ = Cl>~.
Let A E ~and 17(w) = IA(w). Let us show that 1J E &~.In fact, if A E ~
there is aBE PA(R) such that A = {w: e(w) E B}. Let
e.
( )_{1,0,
XBX
X EB,
X: B.
Then/A(w) = XB(e(w)) E <ll~.Henceitfo1Iowsthateverysimple~measurable
function
1 c)A 1(w), Ai E ~.also belongs to&~.
L7=
t We shall assume the usual conventions about arithmetic operations in R: if a e R then
oo = oo, a/ oo = 0; a oo = oo if a > 0, and a oo =  oo if a < 0; 0 ( oo) = 0,
00
+ 00 = 00,
 0 0  00
00.
173
4. Random Variables. I
Now let 17 be an arbitrary ~measurable function. By Theorem 1 there
is a sequence of simple ~measurable functions {IJn} such that IJn( w) + IJ( w ),
n+ 00, wEn. As we just showed, there are Borel functions <fJn = <fJn(x) such
that 1Jn(w) = <fJn(~(w)). Then <fJn(~(w))+ 1J(W), n + 00, wEn.
Let B denote the set {x E R: limn <pn(x) exists}. This is a Borel set. Therefore
<p(x) = {
lim <fJn(x),
X E B,
xB
0,
is also a Borel function (see Problem 7).
But then it is evident that 1J(W) =limn <fJn(~(w)) = <p(~(w)) for all wEn.
Consequently d>~ = Cl>~.
6. Let us consider a probability space (Q, :F, P), where the aalgebra ff is
generated by a finite or countably infinite decomposition!":= {Dt. D2 , . },
I D; = n, P(D;) > 0. Here we shall suppose that the D; are atoms with
respect toP, i.e. if A
D;, A E :F, then either P(A) = 0 or P(D;\A) = 0.
Lemma 3. Let ~ be an .?measurable function, where ff = cr(!":). Then
constant on the atoms of the decomposition, i.e. ~ has the representation
is
00
~(w)
xklnk(w)
(8)
(Pa.s.).t
k= 1
(The notation
"~
= 1J (Pa.s.)" means that
P(~
=f 17) = 0.)
PROOF. Let D be an atom of the decomposition with respect to P. Let us
show that ~is constant (Pa.s.), i.e. P{D n (~ =f canst)} = 0.
Let K = sup{x E R: P{D n (.; < x)} = 0}. Then
P{D n ( < K)} = p[
ryK
{wED;
~(w)
< r}J = 0,
ratwnal r
since if P{D n (~ < x)} = 0, then also P{D n (~ < y)} = 0 for ally :::;; x.
Let x > K; then P{D n (~ < x)} > 0 and therefore P{D n (~ ~ x)} = 0,
since D is an atom. Therefore
P{D n (~ > K)} = P
U
[ r>K
{wED:~~ r}J
= 0.
ratiOnal r
Thus
P {D n ( ~ > K)} = P {D n ( ~ < K)} = 0
and therefore P{D n (~ =f K)} = 0.
Then (8) follows in general since
the lemma.
t a.s.
almost surely.
D; = n. This completes the proof of
174
7.
II. Mathematical Foundations of Probability Theory
PROBLEMS
1. Show that the random variable
is continuous if and only if P(~ = x) = 0 for all
xeR.
2. If I I is :Fmeasurable, is it true that is also :F measurable?
3. Show that ~ = ~(w) is an extended random variable if and only if {w: e(w) E B}
for all Be BII(R).
4. Prove that x", x+
functions.
= max(x, 0),
x
= min(x, 0), and lxl =
x+
+ x
:F
are Borel
5. If and '1 are :!'measurable, then {w: '(w) = "'(w)} e :F.
6. Let ~ and '1 be random variables on (Q, :F), and A e :F. Then the function
C(w) = ~(w) IA
+ '1(w)Ii
is also a random variable.
... ,
7. Let e~o
~.be random variables and cp(x 1, ... , x.) a Borel function. Show that
cp(, 1(w), ... , ~.(w)) is also a random variable.
8. Let
eand '1 be random variables, both taking the values 1, 2, ... , N. Suppose that
:F. Show that there is a permutation (i 1, i2, ... , iN) of (1, 2, ... , N) such thl!t
{w:' = j} = {w: '1 = ij} for j = 1, 2, ... , N.
~ =
5. Random Elements
1. In addition to random variables, probability theory and its applications
involve random objects of more general kinds, for example random points,
vectors, functions, processes, fields, sets, measures, etc. In this connection it is
desirable to have the concept of a random object of any kind.
Definition 1. Let (0, ff') and (E, tff) be measurable spaces. We say that a
function X= X(m), defined on 0 and taking values in E, is ff'/tffmeasurable,
or is a random element (with values in E), if
{m: X(m) E B} E ff'
(1)
for every BE tff. Random elements (with values in E) are sometimes called
valued random variables.
Let us consider some special cases.
If (E, tff) = (R, f!I(R)), the definition of a random element is the same as the
definition of a random variable (4).
Let (E, cC) = (R", f!I(R")). Then a random element X(m) is a "random
point" in R". If nk is the projection of R" on the kth coordinate axis, X(m) can
175
5. Random Elements
be represented in the form
(2)
where ~k = nk o X.
It follows from (1) that
BE ~(R) we have
{w: ~k(w) E B}
~k
is an ordinary random variable. In fact, for
{w: ~ 1 (w) E R, ... , ~k 1 ER, ~k E B, ~k+ 1 E R, .. .}
R x B x R x .. x R)} E ~
= {w: X(w) E (R x .. x
since R x x R x B x R x x R E ~(R").
Definition 2. An ordered set (1J 1(w), ... , 1].(w)) ofrandom variables is called
an ndimensional random vector.
According to this definition, every random element X(w) with values in
R" is an ndimensional random vector. The converse is also true: every
random vector X(w) = (~ 1 (w), ... , ~n(w)) is a random element in R". In fact,
if Bk E ~(R), k = 1, ... , n, then
{w: X(w) E (B1 X ... X Bn)} =
n{w:
k=1
n
~k(w) E Bk} E $'.
But ~(W) is the smallest aalgebra containing the sets B 1 x x Bn.
Consequently we find immediately, by an evident generalization of Lemma 1
of4, that whenever BE ~(R"), the set {w: X(w) E B} belongs to$'.
Let (E, S) = (Z, B(Z)), where Z is the set of complex numbers x + iy,
x, y E R, and B(Z) is the smallest aalgebra containing the sets {z: z = x + iy,
a 1 < x :::; b 1 , a 2 < y :::; b2 }. It follows from the discussion above that a
complexvalued random variable Z(w) can be represented as Z(w) =
X(w) + iY(w), where X(w) and Y(w) are random variables. Hence we may
also call Z(w) a complex random variable.
Let (E, S) = (Rr, ~(RT)), where Tis a subset of the real line. In this case
every random element X = X(w) can evidently be represented as X= (~ 1) 1 e T
with ~~ = n 1 o X, and is called a random function with time domain T.
Definition 3. Let T be a subset of the real line. A set of random variables
X = (~ 1 ) 1 e Tis called a random processt with time domain T.
If T = {1, 2, ... } we call X= (~ 1 , ~ 2 , .. ) a random process with discrete
time, or a random sequence.
If T = [0, 1], ( oo, oo), [0, oo), ... , we call X= (~ 1) 1 er a random
process with continuous time.
tOr stochastic process (Translator).
176
II. Mathematical Foundations of Probability Theory
It is easy to show, by using the structure of the ualgebra BB(RT) (2) that
every random process X= (~ 1)reT (in the sense of Definition 3) is also a
random function on the space (Rr, &B(RT)).
Definition 4. Let X= (~ 1 )reT be a random process. For each given wE Q
the function (~ 1(w))1 e T is said to be a realization or a trajectory of the process,
corresponding to the outcome w.
The following definition is a natural generalization of Definition 2 of 4.
Definition 5. Let X = (~ 1)reT be a random process. The probability measure
Px on (Rr, BB(RT)) defined by
BE &B(Rr),
Px(B) = P{w: X(w) E B},
is called the probability distribution of X. The probabilities
P 1,, .. , 1"(B)
= P{w: (~ 1 ,,
... ,
~~J E B}
with t 1 < t 2 < < tn, t; E T, are called finitedimensional probabilities
(or probability distributions). The functions
Fr,, ... ,tn(Xto ... ,Xn)
with t 1 <
functions.
t2
= P{w:~r,::::; Xto ... ,~tn::::; Xn}
< < tn, t; E T, are called finitedimensional distribution
Let (E, C) = (C, 8B0 (C)), where C is the space of continuous functions
x = (x 1)reT on T = [0, 1] and BB 0 (C) is the ualgebra generated by the open
sets (2). We show that every random element X on (C, 8B 0 (C)) is also a
random process with continuous trajectories in the sense of Definition 3.
In fact, according to 2 the set A = {x E C: x 1 <a} is open in &B 0(C).
Therefore
= {w: X(w) E A} E $'.
{w: ~ 1 (w) <a}
On the other hand, let X= (~ 1(w))reT be a random process (in the sense
of Definition 3) whose trajectories are continuous functions for every
w E n. According to (2.14 ),
{x E C:
n{x E C: lxrk x~kl < p},
E Sp{x 0 )} =
lk
where
tk
are the rational points of [0, 1]. Therefore
{w: X(w) E Sp(X 0 w))}
n{w:
l~rk(w) ~~(w)l
< p} E ff,
lk
and therefore we also have {w: X(w) E B} E ff for every BE 8B 0(C).
Similar reasoning will show that every random element of the space
(D, BB 0 (D) can be considered as a random process with trajectories in the
space of functions with no discontinuities of the second kind; and conversely.
177
5. Random Elements
2. Let (Q, ff, P) be a probability space and (E 11 , &11) measurable spaces,
where IX belongs to an (arbitrary) set 21.
Definition 6. We say that the ff/&11measurable functions (Xa;(w)), IX e 21,
are independent (or collectively independent) if, for every finite set of
indices IXl, ' IX" the random elements Xa:t ... ' xtln are independent, i.e.
P(X111 e B111 , , X a." e Ba:J = P(X111 e B111 )
P(Xa." e B11J,
(3)
where B11 e Ca..
Let 21 = {1, 2, ... , n}, let
F~(xl,
~a:
... , Xn)
be random variables, let IX e 21 and let
Xn)
P(~1 ~ X1, ... , ~" ~
be the ndimensional distribution function of the random vector
~ = (~ 1 , , ~"). Let F~.(x;) be the distribution functions of the random
variables ~;, i = 1, ... , n.
Theorem. A necessary and sufficient condition for the random variables
~ 1 , ... , ~"to be independent is that
(4)
F~(xl, ... ,Xn) = F~ 1 (X1)F~n(Xn)
for all (x 1, . , x") e R".
PRooF. The necessity is evident. To prove the sufficiency we put
a = (a 1 , , a"), b = (b 1 , .. , b")'
P~(a, b] = P{w: a 1 < ~ 1 ~ bl> ... , a"< ~" ~ b"},
P,.(a;, b;] = P{a; < ~; ~ b;}.
Then
P~(a, b] =
n" [F~,(b;) F~,(a;)] = n" P~.(a;, b;]
i=1
i= 1
by (4) and (3.7), and therefore
P{~ 1 e It, ... ,~" E In} =
n" P{~;
(5)
I;},
P{~1 E B1, ~2 E I2, ... ' ~" E In} = P{~1 E B1}
n"
i= 1
where I; = (a;, b;].
We fix I 2 , , I" and show that
i=2
P{~; E I;}
(6)
for all B 1 e iJ(R). Let vH be the collection of sets in ii(R) for which (6)
holds. Then vH evidently contains the algebra d of sets consisting of sums of
disjoint intervals of the form I 1 = (a 1, b 1]. Hence d vH iJ(R). From
178
II. Mathematical Foundations of Probability Theory
the countable additivity (and therefore continuity) of probability measures it
also follows that .A is a monotonic class. Therefore (see Subsection 1 of 2)
JJ.(d) .A ~(R).
But JJ.(d) = u(.r;l) = ~(R) by Theorem 1 of 2. Therefore .A = ~(R).
Thus (6) is established. Now fix B 1 , 13 , ,1.; by the same method we
can establish (6) with / 2 replaced by the Borel set B2 Continuing in this
way, we can evidently arrive at the required equation,
where B; e
3.
~(R).
This completes the proof of the theorem.
PROBLEMS
1. Let 1, ,
only if
e. be discrete random variables. Show that they are independent if and
P(e1
= X1,
00
.'
e. = x.) = TI P(e; = X;)
i=l
for all real Xt. , x .
2. Carry out the proof that every random function (in the sense of Definition 1) is a
random process (in the sense of Definition 3) and conversely.
3. Let X 1 , , X. be random elements with values in (Et. S 1 ), ... , (E., s.), respectively.
In addition let (E~, t9''1 ), , (E~, t9'~) be measurable spaces and let gt. ... , g. be
8tfS'1, . ,t9'.Jt9'~measurable functions, respectively. Show that if X 1 , ... ,x. are
independent, the random elements g 1 X t. .. , g. X. are also independent.
6. Lebesgue Integral. Expectation
1. When (0, 1Ji', P) is a finite probability space and
random variable,
~ = ~(m)
is a simple
~(m) =
L xk!Ak(m),
k=l
(1)
the expectation E~ was defined in 4 of Chapter I. The same definition of the
expectation
of a simple random variable can be used for any probability
space (Q, 1Ji', P). That is, we define
ee
E~ =
L xkP(Ak).
k=l
(2)
179
6. Lebesgue Integral. Expectation
This definition is consistent (in the sense that E~ is independent of the
particular representation of~ in the form (1)), as can be shown just as for
finite probability spaces. The simplest properties of the expectation can be
established similarly (see Subsection 5 of 4 of Chapter 1).
In the present section we shall define and study the properties of the
expectation E~ of an arbitrary random variable. In the language of analysis,
E~ is merely the Lebesgue integral of the g;;measurable function ~ = ~(co)
with respect to the measure P. In addition to E~ we shall use the notation
~(co)P(dco) or So~ dP.
Let~ = ~(co) be a nonnegative random variable. We construct a sequence
of simple nonnegative random variables {~n} n > 1 such that ~nCco) j ~(co),
n + oo, for each co E n (see Theorem 1 in 4).
Since E~n ::;; E~n+ 1 (cf. Property 3) of Subsection 5, 4, Chapter 1), the
limit limn E~n exists, possibly with the value + oo.
So
Definition 1. The Lebesgue integral of the nonnegative random variable
~ = ~(co), or its expectation, is
(3)
n
To see that this definition is consistent, we need to show that the limit is
independent of the choice of the approximating sequence {~n}. In other
words, we need to show that if ~n j ~and 17m j ~'where {17m} is a sequence of
simple functions, then
(4)
m
Lemma 1. Let 17 and '" be simple random variables, n ~ 1, and
Then
(5)
n
PROOF.
Let e > 0 and
It is clear that An j Q and
~n = ~n]An
+ ~n]An
~ ~n]An ~ (17 e)JAn
Hence by the properties of the expectations of simple random variables we
find that
E~n ~
E(17 e)JA" = E17IA" eP(An)
= E17  E17Ix" eP(An) ~ E17  CP(An) e,
180
II. Mathematical Foundations of Probability Theory
where C = maxw '1(w). Since e is arbitrary, the required inequality (5) follows.
It follows from this lemma that limn E~n ; : : : limm E'7m and by symmetry
limm Ef1m ;?:: limn E~n which proves (4).
The following remark is often useful.
Remark 1. The expectation E~ of the nonnegative random variable ~ satisfies
E~
sup Es,
(6)
{seS:s~~}
where S = {s} is a set of simple random variables (Problem 1).
Thus the expectation is well defined for nonnegative random variables.
We now consider the general case.
Let~ be a random variable and~+ =max(~, 0), ~ = min(~, 0).
Definition 2. We say that the expectation E~ of the random variable
exists, or is defined, if at least one of E~ + and E~ is finite:
min(E~+, E~)
< oo.
In this case we define
The expectation Ee is also called the Lebesgue integral (ofthe function
respect to the probability measure P).
Definition 3. We say that the expectation of
EC < oo.
is finite if
E~+
ewith
< oo and
Since 1~1 = e+ ~.the finiteness ofEe, or IE~I < oo, is equivalent to
EI I < 00. (In this sense one says that the Lebesgue integral is absolutely
convergent.)
e,
Remark 2. In addition to the expectation E significant numerical characteristics of a random variable eare the number E~' (if defined) and E IeI', r > 0,
which are known as the moment of order r (or rth moment) and the absolute
moment of order r (or absolute rth moment) of e.
Remark 3. In the definition of the Lebesgue integral Jn ~(w)P(dw) given
above, we suppose that P was a probability measure (P(Q) = 1) and that
the $'measurable functions (random variables) ~ had values in
R = ( oo, oo ). Suppose now that J1. is any measure defined on a measurable
space (Q, ff) and possibly taking the value + oo, and that ~ = e(w) is an
$'measurable function with values in R = [ oo, oo] (an extended random
variable). In this case the Lebesgue integral Jn e(w)Jl(dw) is defined in the
181
6. Lebesgue Integral. Expectation
same way: first, for nonnegative simple ~ (by (2) with P replaced by Jl),
then for arbitrary nonnegative ~. and in general by the formula
provided that no indeterminacy of the form oo  oo arises.
A case that is particularly important for mathematical analysis is that in
which (Q, F) = (R, PJ(R)) and J1 is Lebesgue measure. In this case the
integralSR ~(X)Jl(dx) is writtenSR ~(x) dx, Or s~oo ~(x) dx, Or (L) s~oo ~(x) dx
to emphasize its difference from the Riemann integral (R) S~oo ~(x) dx. If the
measure J1 (LebesgueStieltjes) corresponds to a generalized distribution
function G = G(x), the integral SR ~(x)Jl(dx) is also called a LebesgueStieltjes integral and is denoted by (LS) SR ~(x)G(dx), a notation that
distinguishes it from the corresponding RiemannStieltjes integral
(RS)
L~(x)G(dx)
(see Subsection 10 below).
It will be clear from what follows (Property D) that if E~ is defined then
so is the expectation E( 0 A) for every A E !F. The notations E( ~; A) or SA ~ dP
are often used for E(OA) or its equivalent, Sn OA dP. The integral SA~ dP is
called the Lebesgue integral of~ with respect to P over the set A.
Similarly, we write SA~ dJ1 instead of Sn ~ IA dJ1 for an arbitrary measure
Jl. In particular, if J1 is an ndimensional LebesgueStieltjes measure, and
A = (all b 1 ] x x (a., b.], we use the notation
L(
d)l.
instead of
If JliS Lebesgue measure, we write simplydx 1 dx.instead of J1(dx 1 , .. , dx.).
2. Properties of the expectation E~ of the random variable
A. Let c be a constant and let E~ exist. Then
E(c~)
~.
exists and
E(c~) = cE~.
B. Let
::::; 1J; then
with the understanding that
if oo <
E~
then
oo < E17
and
E~::::;
E17
or
if E17 < oo then
E~
< oo
and
E~
::::; E1J.
182
C.
II. Mathematical Foundations of Probability Theory
IfE~
exists then
IE~ I~ E 1~1
D. If E~ exists then E(eJA) exists for each A E ?'; if E~ is finite, E(~IA) is
finite.
E. If~ and rf are nonnegative random variables, or such that EI~I < oo and
Elt71 < oo,then
E(~
+ t'f) = E~ + Et'f.
(See Problem 2 for a generalization.)
Let us establish AE.
A. This is obvious for simple random variables. Let ~
are simple random variables and c ~ 0. Then c~n j
~
c~
0, ~n j ~' where
and therefore
~n
In the general case we need to use the representation ~ = ~+  ~
and notice that (c~)+ = c~+, (c~) = cC when c ~ 0, whereas when
c < 0, (c~)+ = cC, (c~) = c~+.
B. If 0 ~ ~ ~ rf, then E~ and Et7 are defined and the inequality E~ ~ Et7
follows directly from (6). Now let E~ >  oo; then E~ < oo. If~ ~ rf,
we have ~+ ~ rf+ and ~ ~ rf. Therefore Et7 ~ EC < oo; consequently E11 is defined and E~
= E~+
 E~ :::;; E11+  E11
case when E17 < oo can be discussed similarly.
C. Since  I~I ~ ~ ~ I~I, Properties A and B imply
E11. The
i.e.IE~I ~ E 1~1
D. This follows from B and
E. Let ~ ~ 0, '7 ~ 0, and let {~n} and }'In} be sequences of simple functions
such that ~n i ~and 'In j t'f Then E(~n + 1'/n) = E~n + Et'fn and
and therefore E(~ + rJ) = E~ + ErJ. The case when E 1~1 < oo and
E I'11 < oo reduces to this if we use the facts that
and
183
6. Lebesgue Integral. Expectation
The following group of statements about expectations involve the notion
of" Palmost surely." We say that a property holds" Palmost surely" if there
is a set% E ~with P(%) = 0 such that the property holds for every point
w ofO.\%. Instead of" Palmost surely" we often say" Palmost everywhere"
or simply "almost surely" (a.s.) or "almost everywhere" (a.e.).
F.
If~
= 0 (a.s.) then
E~
= 0.
In fact, if ~ is a simple random variable, ~ =
xk/Ak(w) and xk =I= 0,
wehaveP(Ak) = ObyhypothesisandthereforeE~ = O.If~ ~ OandO~s~~.
where s is a simple random variable, then s = 0 (a.s.) and consequently
Es = 0 and E~ = sup(seS:s:>~J Es = 0. The general case follows from this by
means of the representation ~ = ~+  C and the facts that ~+ ~ I~ I,
C ~ 1~1, and 1~1 = 0 (a.s.).
G. If~= 1J (a.s.) and El~l < oo, then EIIJI < oo and E~ = E17 (see also
Problem 3).
In fact, let % = {w: ~=I= IJ}. Then P(%) = 0 and ~ = O.,v + 0.:f,
1J = 1Jl.,v + IJI.,v = 1Jl.,v + O.K.BypropertiesEandF,wehaveE~ = E~l.,v +
E~(v = E1Jl.:f. But E1Jl.,v = 0, and therefore E~ = E1Jl.K + E1Jl.,v = E17, by
Property E.
H.
Let~ ~
0 and
E~
= 0.
Then~
= 0 (a.s).
For the proof, let A= {w: ~(w) > 0}, An= {w: ~(w) ~ 1/n}. It is clear
that Ani A and 0 ~ ~ IA" ~ ~ IA. Hence, by Property B,
0 S EOAn S E~ = 0.
Consequently
and therefore P(An) = 0 for all n
P(A) = 0.
1. But P(A) = lim P(An) and therefore
I. Let~ and 11 be such that El~l < oo, El11l < oo and E(OA) ~ E(17/A) for
all A E ~.Then~ s 11 (a.s.).
In fact, let B = {w: ~(w) > 17(w)}. Then E(17/8 ) ~ E(0 8 ) ~ E(17/8 ) and
therefore E(0 8 ) = E(17/ 8 ). By Property E, we have E((~ 11)18 ) = 0 and by
Property H we have(~  17)/8 = 0 (a.s.), whence P(B) = 0.
~ be an extended random variable and E I~ I < oo. Then I~ I < oo
(a. s). In fact, let A= {wl~(w)l = oo} and P(A) > 0. Then El~l ~
E(I~IJA) = oo P(A) = oo, which contradicts the hypothesis El~l < oo.
(See also Problem 4.)
J. Let
3. Here we consider the fundamental theorems on taking limits under the
expectation sign (or the Lebesgue integral sign).
184
II. Mathematical Foundations of Probability Theory
Theorem 1 (On Monotone Convergence). Let IJ,
~' ~ 1 , ~ 2 ,
be random
variables.
(a) If~. :2:: 1J for all n :2:: 1, El] >  oo, and ~.
~' then
i E~.
< oo, and~. t ~,then
E~.
(b) If~. :::;; 1J for all n ;;::: 1, E17
E~.
t E~.
PRooF. (a) First suppose that IJ ;;::: 0. For each k ;;::: 1let {~i"l} n"' 1 be a sequence
of simple functions such that ~kl i ~b n+ oo. Put ~(nJ = max 1 ,k,;;N~k>.
Then
,(n1) :::;; '(n) = max
:::;; max
~kn)
~k
= ~ .
1,;;k,;;n
1,;;k,;;n
for 1 :::;; k :::;; n, we find by taking limits as n
+
oo that
for every k ; : : 1 and therefore~ = (.
The random variables (<> are simple and (<>
i (. Therefore
Es = E( =lim E(<n>:::;; limEs .
On the other hand, it is obvious,
:::;;
since~.
limEs. :::;;
~n+ 1
:::;;
~,that
E~.
Consequently limE~. = E~.
Now let IJ be any random variable with EIJ >  oo.
If EIJ = oo then E~. = E~ = oo by Property B, and our proposition is
proved. Let EIJ < oo. Then instead of EIJ >  oo we find E IIJ I < oo. It is
clear that 0 :::;; ~.  1J i ~  IJ for all w E Q. Therefore by what has been
established, E(~.  l])i E(~  IJ)and therefore (by Property E and Problem 2)
E~.  E~
i E~
 EIJ.
But E IIJ I < oo, and therefore E~. i E~, n + oo.
The proof of (b) follows from (a) if we replace the original variables by
their negatives.
Corollary. Let {IJ.}n;, 1 be a sequence of nonnegative random variables. Then
00
00
EIJn
E IJn =
n=1
n=1
185
6. Lebesgue Integral. Expectation
The proof follows from Property E (see also Problem 2), the monotone
convergence theorem, and the remark that
k
00
L tin j n=L tin
n=
1
k..
00.
Theorem 2 (Fatou's Lemma). Lett~. ~b ~ 2 , be random variables.
(a) If~.. ~ t1 for all n ~ 1 and Et~ >  oo, then
E lim~~~ :::;; limE~...
(b) If~ .. :::;; t1 for all n ~ 1 and Et~ < oo, then
limE~ ..
:::;; E lim~ ...
(c) If 1~ .. 1~ t1 for all n ~ 1 and Et~ < oo, then
E lim~~~~ limE~ .. ~ llm E~ .. ~ E llm ~...
PROOF.
(a) Let
Cn =
infm~n ~m;
(7)
then
lim~~~= lim inf ~m =lim C...
n
It is clear that C.. j lim
~~~
m~n
and C.. ~ tl for all n ;;::: 1. Then by Theorem 1
E lim ~~~ = E lim C.. = lim EC.. = lim EC.. ~ limE~ .. ,
n
which establishes (a). The second conclusion follows from the first. The third
is a corollary of the first two.
Theorem 3 (Lebesgue's Theorem on Dominated Convergence). Let t~, ~.
~ 1 , ~ 2 , be random variables such that 1~ .. 1:::;; 1], E17 < oo and~~~..~ (a.s.).
ThenEI~I < oo,
(8)
and
(9)
as n.. oo.
PROOF.
llm ~~~ =
Formula (7) is valid by Fatou's lemma. By hypothesis, lim~~~=
~ (a.s.). Therefore by Property G,
E lim ~~~ = limE~ .. = limE~ .. = E lim ~~~ = E~,
which establishes (8). It is also clear that I~ I ~ tl Hence EI~ I < oo.
Conclusion (9) can be proved in the same way if we observe that
1~..
" ~ 21].
186
II. Mathematical Foundations of Probability Theory
Corollary Let fl, e. e1 be random variables such that Ien I :=:;; f/, en + e
(a.s.) and E17P < oo for some p > 0. Then E 1e1P < oo and E1e eniP+ 0,
n+ oo.
0
For the proof, it is sufficient to observe that
1e1 :=:;; fl, I~ ~niP:=:;; <lei+ leni)P :=:;; (2f1)P.
The condition "I en I :=:;; f/, Ef/ < 00" that appears in Fatou's lemma and
the dominated convergence theorem, and ensures the validity of formulas
(7)(9), can be somewhat weakened. In order to be able to state the corresponding result (Theorem 4), we introduce the following definition.
Definition 4. A family
integrable if
sup
11
{en} 11 ~ 1
J{i~nl>c)
of random variables is said to be uniformly
c+
le,IP(dw)+0,
00,
(10)
or, in a different notation,
sup E[ Ie~~l Iu~"' >c}J + 0,
c+
00.
(11)
II
It is clear that if ell' n ;;::: 1, satisfy Iell I :=:; f/, Ef/ < oo, then the family
{ell} 11 ~ 1 is uniformly integrable.
Theorem 4. Let
Then
{~ 11 }n~ 1
be a uniformly integrable family ofrandom variables.
(a) E lim e11 :=:;; lim Ee11 :=:;; ITiii Ee, :=:; E ITiii eli.
(b) If in addition 11 + (a.s.) then~ is integrable and
e e
Ee,+ Ee,
Ele~~el+0,
n+ oo,
n+oo.
PRooF. (a) For every c > 0
(12)
By uniform integrability, for every e > 0 we can take c so large that
sup IE[e~~I 1~"<cj]l <e.
(13)
II
By Fatou's lemma,
lim E[e,I 1~"~ c}J ;;::: E[lim ~nl{~n~ c}J.
But
~,If~"~ cl
;;:::
e, and therefore
lim E[e,J!~"~clJ;;::: E[lim e,].
(14)
187
6. Lebesgue Integral. Expectation
From (12)(14) we obtain
Since e > 0 is arbitrary, it follows that lim E~n 2: E lim ~n The inequality
with upper limits, 1lm E~n :::;; E lim~"' is proved similarly.
Conclusion (b) can be deduced from (a) as in Theorem 3.
The deeper significance of the concept of uniform integrability is revealed
by the following theorem, which gives a necessary and sufficient condition
for taking limits under the expectation sign.
~n> ~and E~n < oo. Then E~n> E~ < oo
the family {~n}n> 1 is uniformly integrable.
Theorem 5. Let 0 :::;;
if and only if
PROOF. The sufficiency follows from conclusion (b) of Theorem 4. For the
proof of the necessity we consider the (at most countable) set
= {a:
P(~
= a) > 0}.
Then we have ~n/{~"<al> Or~<a) for each a~ A, and the family
{~n/{~n<a}}n~ 1
is uniformly integrable. Hence, by the sufficiency part of the theorem, we have
E~nlr~"<aJ> E~/~~"<aJ' a~ A, and therefore
a~
Take an e > 0 and choose a 0
N 0 so large that
A,
n> oo.
(15)
A so large that E ~/ (~ 2: ao) < ej2; then choose
E~nlr~"2:aoJ :::;; E~lr~~aoJ
+ e/2
for all n 2: N 0 , and consequently E~n/{~"~ao}:::;; e. Then choose a 1 2: a0 so
large that E~r~"2:a!}::::;; e for all n:::;; N 0 Then we have
supE~nlr~"~ad :::;;
n
e,
which establishes the uniform integrability of the family {~n} n 2: 1 of random
variables.
4. Let us notice some tests for uniform integrability.
We first observe that if {~n} is a family of uniformly integrable random
variables, then
sup EI~n I < oo.
n
(16)
188
II. Mathematical Foundations of Probability Theory
In fact, for a given ~> > 0 and sufficiently large c > 0
sup E l~nl =sup [E(I~n IIu~nl~cl)
n
supE(I~n IImnl~c}}
n
+ E(l~n IIu~nl<cl)]
+ supE(I~niimnl<c)) s G + c,
n
which establishes (16).
It turns out that (16) together with a condition of uniform continuity is
necessary and sufficient for uniform integrability.
Lemma 2. A necessary and sufficient condition for a family {~n}n~ 1 of random
variables to be uniformly integrable is that E I~n I, n ;;::: 1, are uniformly bounded
(i.e., (16) holds) and that E{ I~n IIA},n ;;::: 1, are uniformly absolutely continuous
(i.e. sup E{I ~n II A} + 0 when P(A) + 0).
PROOF.
Necessity. Condition (16) was verified above. Moreover,
E{l~niiA} = E{l~niiArl{l~nl~c}} + E{l~niiAn{i~nl<c}}
s E{l~niiu~nl~cJ} + cP(A).
Take c so large that supn E{I~II 11 ~"I"cJ}
~>/2.
sup E{I ~n II A} S
Then if P(A)
~>/2c,
(17)
we have
I>
by (17). This establishes the uniform absolute continuity.
Sufficiency. Let 1> > 0 and b > 0 be chosen so that P(A) < b implies that
E (I ~n II A) $ e, uniformly in n. Since
El~nl ~ EJ~nliu~nl~cl ~ cP{l~nl ~ c}
for every c > 0 (cf. Chebyshev's inequality), we have
sup
P{l~nl;;:::
c}
1
c
s sup E l~nl+ 0,
c+ 00,
and therefore, when c is sufficiently large, any set {I ~n I ;;::: c }, n ;;::: 1, can be
taken as A. Therefore sup E(l ~nl I 0 ~" 1 ~, 1 ) s ~>,which establishes the uniform
integrability. This completes the proof of the lemma.
The following proposition provides a simple sufficient condition for
uniform integrability.
Lemma 3. Let ~b ~ 2 , . be a sequence of integrable random variables and
G = G(t) a nonnegative increasing function, defined fort ~ 0, such that
lim G(t) = oo.
t+ 00
supE[G(I~ni)J
n
(18)
<
00.
Then the family {~n}n" 1 is uniformly integrable.
(19)
189
6. Lebesgue Integral. Expectation
PRooF. Let e > 0, M =sup" E[G(I~ni)J, a= Mfe. Take c so large that
G(t)/t 2 a for t 2 c. Then
E[l~nllll~"l"'ca:::;; ~E[G(I~nl) Iu~"l"'cj]:::;;;; = e
uniformly for n 2 1.
5. If ~ and Yf are independent simple random variables, we can show, as in
Subsection 5 of 4 of Chapter I, that E~IJ = E~ Ery. Let us now establish a
similar proposition in the general case (see also Problem 5).
Theorem 6. Let ~ and 1J be independent random variables with EI~ I < oo,
Elryl < oo. ThenEI~Yfl < ooand
E~ry
PRooF. First let
E~
Ery.
(20)
2 0, Yf 2 0. Put
00
00
~n=
I J{k/n:S~(ro)<(k+l)/n}
k;o n
I  J{k/n:S~(w)<(k+ 1)/n)
k;o n
Yfn
Then ~n :::;; ~' I~n  ~I :::;; lfn and Yfn :::;; Yf, IYfn  Yf I :::;; 1/n. Since E~ < oo and
Ery < oo, it follows from Lebesgue's dominated convergence theorem that
lim
Moreover,
since~
E~nYfn =
and
Yf
E~n
E~,
are independent,
kl
L 2
k,l"'o
L
EJ{k/n:S~<(k+l)/n}/{1/n:S~<(I+l)/n}
n
kl
2E/{k/n:S~<(k+1)/n} EJ{I/n;S;~<(l+l)/n} = E~n EYfn
k,l"'o n
Now notice that
1 1 ( 1)
+E[IYfnll~~niJ::;E~+E ry+
~o,
n~
oo.
Therefore E~ry =limn E~nYfn =lim E~n lim E11n = E~ Ery, and E~Yf < oo.
The general case reduces to this one if we use the representations
~=C C,Yf=Yf+ ry,~Yf=~+Yf+ Cry+ ~+Yf +Cry.Thiscompletes the proof.
6. The inequalities for expectations that we develop in this subsection are
regularly used both in probability theory and in analysis.
190
II. Mathematical Foundations of Probability Theory
Chebyshev's Inequality. Let
be a nonnegative random variable. Then for
every e > 0
E~
(21)
P(~ ~e):::;;.
The proof follows immediately from
E~ ~ E[~ /(~>elJ ~ E/(~~l = eP(~ ~e).
From (21) we can obtain the following variant of Chebyshev's inequality:
If ~ is any random variable then
P(~ ~e):::;;
ee
(22)
and
(23)
where V~ = E(~  E~) 2 is the variance of~.
The CauchyBunyakovskiilnequality. Let~ and rt satisfy E~ 2 < oo, Ert 2 < oo.
Then E I~1'/ I < oo and
(E I~1'/ 1) 2
:::;;
(24)
E~ 2 . Ert 2
PROOF.
Suppose that E~ 2 > 0, E17 2 > 0. Then, with ~
rt!JErli,
we find, since
~~~' if=
21 ~~I :::;; ~ 2 + ~ 2 , that
2E ~~~~ :::;; E~ 2 + E~ 2 = 2,
i.e. E I~~ I :::;; 1, which establishes (24).
ee
= 0, then ~ = 0 (a.s.) by Property I, and
On the other hand if, say,
then E~rt = 0 by Property F, i.e. (24) is still satisfied.
Jensen's Inequality. Let the Borel function g = g(x) be convex downward and
El~l
< oo. Then
g(E~)
:::;;
PRooF. If g = g(x) is convex downward, for each x 0
A.(x 0 ) such that
g(x)
for all x
g(x 0 )
(25)
Eg(~).
+ (x 
R there is a number
x 0 ) A.(x 0 )
R. Putting x = ~ and x 0 = E~' we find from (26) that
g(~) ~ g(E~)
and consequently Eg(~)
~ g(E~).
+ (~ 
E~) A.(E~),
(26)
191
6. Lebesgue Integral. Expectation
A whole series of useful inequalities can be derived from Jensen's inequality.
We obtain the following one as an example.
Lyapunov's Inequality. IfO < s < t,
(27)
To prove this, let r = tfs. Then, putting 11 = I~ I" and applying Jensen's
inequality to g(x) = IxI', we obtain IE17l' :::;;; EI'71', i.e.
(E I~ 1")'1" :::;;; E IeJ',
which establishes (27).
The following chain of inequalities among absolute moments in a consequence of Lyapunov's inequality:
(28)
HOlder's Inequality. Let 1 < p < oo, 1 < q < oo, and (1/p)
E I~ IP < oo and EI11lq < oo, then EI~'71 < oo and
EI~'71
: :; ; (E I~ IP) 11P(E I'11q)lfq.
+ (1/q) = 1.
If
(29)
If E I~ IP = 0 or E Irtlq = 0, (29) follows immediately as for the CauchyBunyakovskii inequality (which is the special case p = q = 2 of Holder's
inequality).
Now let E I~IP > 0, E lrtlq > 0 and
~ = (E I~ lp)lfp '
We apply the inequality
(30)
which holds for positive x, y, a, b and a
from the concavity of the logarithm:
In[ax
+ by]
~ a
In x
+b=
1, and follows immediately
+ b In y =
In xayh.
Then, putting x = ~P, y = ;jq, a = 1/p, b = 1/q, we find that
~;j :::;;; ! ~p + ! ;jq,
p
whence

E~q
This establishes (29).
:::;;;  E~P
p
+ 1 E;jq = 1 + 1 =
q
1.
192
II. Mathematical Foundations of Probability Theory
Minkowski's Inequality.JfE I~IP < oo, E1'71P < oo, 1 $ p < oo, then we have
E I~ + 'liP < oo and
(E I~ + '71P)l/p
(E I~ IP)lfp + (E I'71P)lip.
(31)
We begin by establishing the following inequality: if a, b > 0 and p
then
1,
(32)
In fact, consider the function F(x) = (a + x)P  2p 1(aP + xP). Then
F'(x) = p(a + x)pl 2P 1pxpl,
and since p
1, we have F'(a) = 0, F'(x) > 0 for x < a and F'(x) < 0 for
x > a. Therefore
F(b) $ max F(x) = F(a) = 0,
from which (32) follows.
According to this inequality,
and therefore if EI~ IP < oo and EI'71P < oo it follows that EI~ + 'liP < oo.
If p = 1, inequality (31) follows from (33).
Now suppose that p > 1. Take q > 1 so that (1/p)
(1/q)
1. Then
I~+ 'lip= I~+ '711~ + '71pl Sl~ll~ + '71pl + 1'711~ + '71pl.
(34)
Notice that (p  1)q = p. Consequently
E(l~
+ '7ipl)q = Ei~ +'liP< oo,
and therefore by Holder's inequality
E(l~ll~
+ '7ip1)
+ '7i(pl>q)liq
= (E I~ IP) 11P(E I~ + '7ip)lfq < 00.
~ (Ei~IP)liP(Ei~
In the same way,
E( I'1 II~ + '71pl) $ (E I'71P) 11P(E I~ + '7ip)lfq,
Consequently, by (34),
EI~ + 'lip
(E I~ + '71P)lfq((E I~ IP)l/p + (E I'71P) 11P).
IfE I~ + 'liP = 0, the desired inequality (31) is evident. Now let EI~
Then we obtain
(E I~ + '71P)l(l/q)
(E I~ IP)lfp + (E I'71P)lfp
from (35), and (31) follows since 1  (1/q)
= 1/p.
(35)
+ 'liP > 0.
193
6. Lebesgue Integral. Expectation
7.
Let~ be a random variable for which
the set function
O(A)
E~
is defined. Then, by Property D,
=L~ dP,
(36)
is well defined. Let us show that this function is countably additive.
First suppose that ~ is nonnegative. If A 1 , A 2 , ... are pairwise disjoint sets
from fF and A = LA", the corollary to Theorem 1 implies that
O(A) = E(~. IA) = E(~. IEAJ = E(:L ~. IAJ
=IE(~ IAJ = L O(An).
is an arbitrary random variable for which E~ is defined, the countable
additivity of O(A) follows from the representation
If~
(37)
where
together with the countable additivity for nonnegative random variables and
the fact that min(O+(n), ocn)) < 00.
Thus if E~ is defined, the set function 0 = O(A) is a signed measurea countably additive set function representable as 0 = 0 1  0 2 , where at
least one of the measures 0 1 and 0 2 is finite.
We now show that 0 = O(A) has the following important property of
absolute continuity with respect to P:
if
P(A) = 0
O(A) = 0
then
(A
ff)
(this property is denoted by the abbreviation 0 ~ P).
To prove the sufficiency we consider nonnegative random variables. If
~ =
1 xkiAk is a simple nonnegative random variable and P(A) = 0,
then
I;;=
O(A) = E(~ IA) =
xkP(Ak n A)= 0.
k=1
If { ~"} n:e: 1 is a sequence of nonnegative simple functions such that
then the theorem on monotone convergence shows that
~"
~ ~
0,
O(A) = E(~ IA) = lim E(~n IA) = 0,
since E(~" IA) = 0 for all n ~ 1 and A with P(A) = 0.
Thus the Lebesgue integral O(A) = fA~ dP, considered as a function of
sets A E ff, is a signed measure that is absolutely continuous with respect to
P (0 ~ P). It is quite remarkable that the converse is also valid.
194
II. Mathematical Foundations of Probability Theory
RadonNikodym Theorem. Let (0, ff) be a measurable space, 11 a afinite
measure, and A a signed measure (i.e., A = A1  A2 , where at least one of the
measures A1 and A2 is finite) which is absolutely continuous with respect to 11
Then there is an ffmeasurable.function f = .f(w) with values in R = [ oo, oo]
such that
A(A) = {f(w)l1(dw),
(38)
The function f(w) is unique up to sets of 11measure zero: if h = h(w) is
another ffmeasurable function such that A(A) = JA h(w)11(dw), A E ff, then
11{ w: f(w) =P h(w)} = 0.
If A is a measure, then f = f (w) has its values in R+ = [0, oo].
Remark. The function f = f(w) in the representation (38) is called the
Radon N ikodym derivative or the density of the measure A with respect to 11,
and denoted by dA/dl1 or (dA/d11)(w).
The RadonNikodym theorem, which we quote without proof, will play
a key role in the construction of conditional expectations (7).
8. If~=
Li'=
xJA; is a simple random variable,
Eg( ~)
= L: g(xi)P(Ai) = Lg(xi)AF ~(xJ
(39)
In other words, in order to calculate the expectation of a function of the
(simple) random variable~ it is unnecessary to know the probability measure
P completely; it is enough to know the probability distribution P ~ or, equivalently, the distribution function F ~of~The following important theorem generalizes this property.
Theorem 7 (Change of Variables in a Lebesgue Integral). Let (0, ff) and
(E, $)be measurable spaces and X = X(w) an ff/$measurable function with
values in E. Let P be a probability measure on (0, ff) and Px the probability
measure on (E, $)induced by X= X(w):
Px(A) = P{w: X(w)
Then
g(x)Px(dx) =
AE$.
A},
g(X(w))P(dw),
AE$,
Let A
(41)
xl(A)
for every $measurable function g = g(x), x E E (in the sense that
integral exists, the other is well defined, and the two are equal).
PROOF.
(40)
if one
$and g(x) = IB(x), where BE$. Then (41) becomes
Px(AB)
P(X 1(A) n
x
1(B)),
(42)
195
6. Lebesgue Integral. Expectation
x
1(B) =
which follows from (40) and the observation that x 1(A) n
1
x (A n B).
It follows from (42) that (41) is valid for nonnegative simple functions
g = g(x), and therefore, by the monotone convergence theorem, also for all
nonnegative iff measurable functions.
In the general case we need only represent gas g+  g. Then, since (41)
is valid for g+ and g, if(for example) fAg+(x)Px(dx) < oo, we have
x 1 (A)
g+(X(w))P(dw)
< oo
also, and therefore the existence of fA g(x)Px(dx) implies the existence of
g(X(w))P(dw).
fx'(AJ
Corollary. Let (E, $) = (R, BI(R)) and
let~ = ~(w) be a random variable with
probability distribution P~. Then if g = g(x) is a Borel function and either ofthe
integrals fA g(x)P~(dx) or f~'(AJ g(~(w))P(dw) exists, we have
f g(x)P~(dx) J
=
g(~(w))P(dw).
~I(A)
In particular, for A = R we obtain
Eg(~(w)) = Lg(~(w))P(dw) = Lg(x)P~(dx).
(43)
The measure P~ can be uniquely reconstructed from the distribution
function F~ (Theorem 1 of 3). Hence the Lebesgue integral fR g(x)P~(dx) is
often denoted by JR g(x)F~(dx) and called a LebesgueStieltjes integral
(with respect to the measure corresponding to the distribution function
F~(x)).
Let us consider the case when F ~(x) has a density f~(x), i.e. let
(44)
where f~ = f~(x) is a nonnegative Borel function and the integral is a Lebesgue
integral with respect to Lebesgue measure on the set ( oo, x] (see Remark 2
in Subsection 1). With the assumption of (44), formula (43) takes the form
Eg(~(w)) = J:00 g(x)f~(x) dx,
(45)
where the integral is the Lebesgue integral of the function g(x)Nx) with
respect to Lebesgue measure. In fact, if g(x) = Ia(x), BE BI(R), the formula
becomes
BE 14(R);
(46)
196
II. Mathematical Foundations of Probability Theory
its correctness follows from Theorem 1 of 3 and the formula
F~(b) F~(a) = ff~(x) dx.
In the general case, the proof is the same as for Theorem 7.
9. Let us consider the special case of measurable spaces (Q, !F) with a
measure J.l, where Q = Q 1 x Q 2, !F = !F1 !F2 , and J.l = f.lt x J.l 2 is the
direct product of measures f.lt and f.l 2 (i.e., the measure on !F such that
the existence of this measure follows from the proof of Theorem 8).
The following theorem plays the same role as the theorem on the reduction
of a double Riemann integral to an iterated integral.
Theorem 8 (Fubini's Theorem). Let ~ = ~(ro 1 , ro 2) be an !F 1 !F rmeasurable function, integrable with respect to the measure f.lt x J.l 2:
(47)
Then the integrals Jn 1 ~(rob ro 2)J.l 1(dro 1) and Jn2 ~(ro 1 , ro 2)J.lidro2)
(1) are defined for all ro 1 and ro 2;
(2) are respectively !F 2  and !F 1 measurable functions with
J.l2{ro2:
J.l1{ro1:
t 1 1~(rot, ro2)1f.lt(dro
1)
= oo} =
t2 1~(rot, ro2)IJ.l2(dro2) =
0,
(48)
oo} = 0
and (3)
l"lt xn2
~(ro1, ro2) d(J.l1
x J.l2) =
i [i ~(rot,
l"lt
!12
ro2)J.l2(dro2)JJ.lt(drol)
tJt 1 ~(rob ro2)J.lt(drot)]f.lidro2).
(49)
PRooF. We first show that ~w 1 (ro 2 ) = ~(ro 1 , ro 2) is !F rmeasurable with
respect to ro 2, for each ro 1 E Q 1.
Let FE !F 1 !F 2 and ~(rob ro 2) = I iro 1, ro 2). Let
197
6. Lebesgue Integral. Expectation
be the crosssection ofF at Wt. and let re'w,
show that re'ro, =!IF for every w 1.
If F = A x B, A E !IF 1, B E !IF 2 , then
(A
B)w,
= {FE !IF: Fw, E F 2 }. We must
B if w 1 E A,
= {0
if w 1 A.
Hence rectangles with measurable sides belong to re' "''. In addition, if
FE !IF, then (Fln, = F w,, and if {F"kd are sets in !IF, then
F")w, =
F':,,. It follows that re'w, = !F.
Now let ~(w 1 , w 2 ) :2: 0. Then, since the function ~(w 1 , w 2 ) is !F2 measurable
for each rob the integral 2 ~(w 1 , w 2 )f1z(dw 2 ) is defined. Let us show that this
integral is an ff1measurable function and
<U
Jn
Let us suppose that ~(w 1 , w 2 ) = IAx 8 (w 1 , w 2 ), A E !F1, BE .?F2 . Then since
IA xiw1, w 2 ) = IA(w 1 )1 8 (w 2 ), we have
and consequently the integral on the left of (51) is an !F1measurable function.
Now let ~(w 1 , w 2 ) = IF(w 1 , w 2 ), FE !IF = ff1 0 !F2 Let us show that the
integralf(w 1) =
2 I F(w 1, w 2 )f1 2 (dw 2 ) is !IF measurable. For this purpose we
put re' = {FE !F:f(w 1) is !!i'rmeasurable}. According to what has been
proved, the set A x B belongs to '6' (A E $'1, B E $'2 ) and therefore the algebra
d consisting of finite sums of disjoint sets of this form also belongs to re'. It
follows from the monotone convergence theorem that re' is a monotonic
class, re' = Jl(re'). Therefore, because of the inclusions d <;; re' c;; !IF and
Theorem 1 of 2, we have ff = a{d) = fl(d) <;; Jl(re') = re' <;; .?F, i.e. re' = !F.
Finally, if ~(wb w 2 ) is an arbitrary nonnegative.!Fmeasurable function,
the !F 1measurability of the integral 2 ~(w 1 , w 2 )f1z(dw) follows from the
monotone convergence theorem and Theorem 2 of 4.
Let us now show that the measure f1 = f1 1 x flz defined on !IF = ff2 0 ff2 ,
withtheproperty(J1 1 x J1 2 )(A x B)= f1 1(A)J1 2 (B),AE!!i'1,BEff2 ,actually
exists and is unique.
For FE ff we put
Jn
Jn
Jl(F)
LJL/Fw,(Wz)f1z(dwz)}1(dw1).
As we have shown, the inner integral is an ff1measurable function, and
consequently the set function Jl(F) is actually defined for F E !F. It is clear
198
II. Mathematical Foundations of Probability Theory
that ifF =A x B, then fl(A x B)= 11 1(A)11 2 (B). Now let {Fn} be disjoint
sets from .F. Then
/l(L Fn) = LJL/<1:1"")w 1(Wz)flidwz)]/ll(dwl)
=
i [i
i [i
L
Clt n
Cl2
Clt
Cl2
/F:J 1 (wz)/lz(dwz)]/ll(dwl)
/F:J, (wz)/lz(dwz)]/1 1(dw 1) =
L fl(P),
n
i.e.fl is a (afinite) measure on .F.
It follows from Caratheodory's theorem that this measure 11 is the unique
measure with the property that /l(A x B) = 11 1(A)/1 2 (B).
We can now establish (50). If ~(w 1 , w 2) = I Ax B(w 1, Wz), A E !#'1 , BE !#'2,
then
Clt xQ2
I Ax B(wl> w 2 )d(fl 1 x /lz) = (/1 1 x 11 2)(A x B),
(52)
and since 1Axiw 1, w 2) = IA(w 1)la(w 2), we have
LJL/Axiwl, w2)/12(dw2)]111(dwl)
=
L.[IA(wl) L/B(wl> w2)/12(dw2)]/11(dw 1) = /11(A)fliB).
(53)
But, by the definition of 11 1 x J12,
(J1 1
J1 2)(A x B) = J1 1 (A)J1 2 (B).
Hence it follows from (52) and (53) that (50) is valid for
IAxiwl, w2).
Now let
~(w 1 ,
~(w 1 ,
w2) =
w 2) = JF(w 1, w 2 ), FE .F. The set function
is evidently a afinite measure. It is also easily verified that the set function
v(F) = L.[L/iwl, Wz)J1 2 (dw 2 )]J11(dw 1)
is a afinite measure. It will be shown below that Aand v coincide on sets of
the form F = A x B, and therefore on the algebra !F. Hence it follows by
Caratbeodory's theorem that Aand v coincide for all F E .F.
199
6. Lebesgue Integral. Expectation
We turn now to the proof of the full conclusion of Fubini's theorem.
By (47),
~+(wb w2) d(f.ll
Jn,xn 2
f.12) <
00,
By what has already been proved, the
:1'1measurable function of w 1 and
C(wl, w2)d(J.11
f.12) <
00.
Jn,xn 2
integral Jn 2~+(w 1 , w 2)J.1idw 2) is
r [ r ~+(wl, W2)f.12(dw2)]f.11(dwl) = Jn,xn2
r ~+(wl, W2) d(f.ll
Jn, Jn2
f.12) <
an
00.
Consequently by Problem 4 (see also Property J in Subsection 2)
r ~+(wb w2)f.12(dw2) <
00
(f.lla.s.).
( CCw1, w2)J.11(dw1) < oo
CJ.11a.s.),
Jo2
In the same way
Jo2
and therefore
It is clear that, except on a set Jll of J.1 1measure zero,
(
Jn2
~(wl, w2)f.1idw2) = ( ~+(wl, w2)J.12(dw2)Jo2
( C(wb w2)J.12(dw2).
Jn2
(54)
Taking the integrals to be zero for w 1 E%, we may suppose that (54) holds
for all w E n 1. Then, integrating (54) with respect to f.1 1 and using (50), we
obtain
t,[t2 ~(wl, w2)f.12(dw2)]f.11(dw1)
t,[t2 ~+(w 1 , w2)J.1z(dw2)]f.11(dw 1)
t,[t
r
C(wb w2)f.12(dw2)}1(dw1)
~+(wl, W2)d(J.11
Jn,xn 2
 Jn,
r xn2 ~(Wl, W2) d(f.ll
f.12)
X
f.12)
200
II. Mathematical Foundations of Probability Theory
Similarly we can establish the first equation in (48) and the equation
Jn, xn2
e(wl, W2) d(Jl.l X J1.2) =
r [ r e(wl, W2)Jl.l(dwl)JJ1.2(dw2).
Jn2 Jn,
This completes the proof of the theorem.
Corollary. If Jn, Un2 Ie<wl, w2) IJ1.idw2)]Jl.l (dwl) <
Fubini's theorem is still valid.
oo,
the conclusion of
In fact, under this hypothesis (47) follows from (50), and consequently
the conclusions of Fubini's theorem hold.
Let (e, q) be a pair of random variables whose distribution has a
twodimensional density f~~(x, y), i.e.
EXAMPLE.
P((e, q) e B) =
{f~~(x, y) dx dy,
where f~~(x, y) is a nonnegative &6'(R 2 )measurable function, and the integral
is a Lebesgue integral with respect to twodimensional Lebesgue measure.
Let us show that the onedimensional distributions for and '1 have
densities f~(x) and f,(y), and furthermore
f~(x) = f_oo,J~~(x, y) dy
(55)
and
~(y) =
J:oo f~~(x, y) dx.
In fact, if A e as'(R), then by Fubini's theorem
P(eeA) = P((e,q)eA x R) =
{x/~~(x,y)dxdy = L[Lf~~(x,y)dy]dx.
This establishes both the existence of a density for the probability distribution
of and the first formula in (55). The second formula is established similarly.
According to the theorem in 5, a necessary and sufficient condition that
and '1 are independent is that
F~71(x,
y) =
F~(x)F,(y),
(x, y)
R 2
Let us show that when there is a twodimensional density
variables eand '1 are independent if and only if
h.~(x, y),
the
(56)
(where the equation is to be understood in the sense of holding almost
surely with respect to twodimensional Lebesgue measure).
201
6. Lebesgue Integral. Expectation
In fact, in (56) holds, then by Fubini's theorem
F
~~(x, y) =
=
J(oo,x)x(oo,y] f~~(x, y) dx dy = J(oo,x)x(oo,y] f~(x)f~(y) dx dy
(oo,x]
Nx) dx(J
(oo,y)
j,(y) dy) =
F~(x)F~(y)
and consequently ~ and 11 are independent.
Conversely, if they are independent and have a density h~(x, y), then again
by Fubini's theorem
foo,x)x(oo,y/~~(X, y) dx dy = (f_oo,x/~(x) dx )(foo.y]j,(y) dy)
= J
f~(x)fq(y) dx dy.
(oo,x]x(oo,y]
It follows that
f/~~(x, y) dx dy = f/~(x)fq(y) dx dy
for every BE fJI (R 2 ), and it is easily deduced from Property I that (56) holds.
10. In this subsection we discuss the relation between the Lebesgue and
Riemann integrals.
We first observe that the construction of the Lebesgue integral is inde
pendent of the measurable space (Q, ') on which the integrands are given.
On the other hand, the Riemann integral is not defined on abstract spaces in
general, and for Q = W it is defined sequentially: first for R 1 , and then
extended, with corresponding changes, to the case n > 1.
We emphasize that the constructions of the Riemann and Lebesgue
integrals are based on different ideas. The first step in the construction of the
Riemann integral is to group the points x E R 1 according to their distances
along the x axis. On the other hand, in Lebesgue's construction (for Q = R 1 )
the points x E R 1 are grouped according to a different principle: by the
distances between the values of the integrand. It is a consequence of these
different approaches that the Riemann approximating sums have limits only
for "mildly" discontinuous functions, whereas the Lebesgue sums converge
to limits for a much wider class of functions.
Let us recall the definition of the RiemannStieltjes integral. Let G = G(x)
be a generalized distribution function on R (see subsection 2 of 3) and 11 its
corresponding LebesgueStieltjes measure, and let g = g(x) be a bounded
function that vanishes outside [a, b].
202
II. Mathematical Foundations of Probability Theory
Consider a decomposition fJJ = {x 0 , , xn},
a = x0
< x 1 < <
= b,
Xn
of [a, b], and form the upper and lower sums
n
L = L iHG(x;+ 1) ~
L=
G(x;)],
i=1
L~;[G(xi+ 1 ) G(x;)]
i=l
where
g; =
g(y),
sup
inf
ff.i =
g(y).
Xi1 <y:s";Xi
Xi1 <y~Xi
Define simple functions g~(x) and fl~(x) by taking
on x;_ 1 < x :::;;
X;,
and define
g~(a) = fl~(a) =
g(a). It is clear that then
L = (LS) Jbg~(x)G(dx)
a
and
L = (LS) Jb fl~(x)G(dx).
a
Now let{&\} beasequenceofdecomposition ssuchthatfJJk s;: .o/lk+ 1 Then
and if Jg(x)J :::;; C we have, by the dominated convergence theorem,
lim
L=
(LS)
rb g(x)G(dx),
k+ oo
~k
lim
L = (LS) Jb fl(x)G(dx),
Ja
(57)
k+oo ~k
where g(x) = limk g~k(x), g(x) = limk fl~Jx).
H the limits limk r~k and limk I~k are finite and equal, and their common
value is independent of the sequence ofdecompositions {fJJk}, we say that g = g(x)
is RiemannStieltjes integrable, and the common value of the limits is denoted
by
(RS)
g(x)G(dx).
(58)
When G(x) = x, the integral is called a Riemann integral and denoted by
(R) fg(x) dx.
203
6. Lebesgue Integral. Expectation
J!
Now let (LS) g(x)G(dx) be the corresponding LebesgueStieltjes
integral (see Remark 2 in Subsection 2).
Theorem 9. If g = g(x) is continuous on [a, b], it is RiemannStieltjes integrable and
(RS)
g(x)G(dx) = (LS)
(59)
g(x)G(dx).
PRooF. Since g(x) is continuous, we have g(x) = g(x) = g(x). Hence by (57)
Consequently g = g(x) is RiemannStieltjes
= limk+oo
limk+oo
integral (again by (57)):
L&k
L&k
Let us consider in more detail the question of the correspondence between
the Riemann and Lebesgue integrals for the case of Lebesgue measure on the
lineR.
Theorem 10. Let g(x) be a bounded function on [a, b].
if and only if it is
continuous almost everywhere (with respect to Lebesgue measure A on
(a) The function g = g(x) is Riemann integrable on [a, b]
~([a, b])).
(b) If g = g(x) is Riemann integrable, it is Lebesgue integrable and
(60)
(R) fg(x) dx = (L) fg(x)A(dx).
PROOF. (a) Let g = g(x) be Riemann integrable. Then, by (57),
(L)
But g(x)
g(x)
S:g(x)A(dx) = S:g(x)A(dx).
(L)
g(x), and hence by Property H
g_(x) = g(x) = g(x)
(61)
(Aa.s.),
from which it is easy to see that g(x) is continuous almost everywhere (with
respect to 1).
Conversely, let g = g(x) be continuous almost everywhere (with respect
to A). Then (61) is satisfied and consequently g(x) differs from the (Borel)
measurable function g(x) only on a set JV with A:(%) = 0. But then
{x: g(x)
c} = {x: g(x)
= {x: g(x)
It is clear that the set {x: g(x)
~
~
c} n JV
c} n JV
+ {x: g(x)
+ {x: g(x)
~
~
c} n JV
c} n JV
c} n JV E &~([a, b]), and that
{x: g(x)
c} n JV
204
II. Mathematical Foundations of Probability Theory
is a subset of .Y having Lebesgue measure A: equal to zero and therefore also
belonging to ~([a, b)]. Therefore g(x) is &l([a, b])measurable and, as a
bounded function, is Lebesgue integrable. Therefore by Property G,
(L) fg(x)A:(dx) = (L) ff!(x)A:(dx) = (L) fg(x)A(dx),
which completes the proof of (a).
(b) If g = g(x) is Riemann integrable, then according to (a) it is continuous
(A:a.s.). It was shown above than then g(x) is Lebesgue integrable and its
Riemann and Lebesgue integrals are equal.
This completes the proof of the theorem.
Remark. Let J.t be a LebesgueStieltjes measure on ~([a, b]). Let ~Jt([a, b])
be the system consisting of those subsets A s; [a, b] for which there are sets
A and B in ~([a, b]) such that A s; A s; B and J.t(B\A) = 0. Let J.t be an
extension of J.t to &I i[a, b]) (Jl(A) = J.t(A) for A such that A s; A s; B and
J.t(B\A) = 0). Then the conclusion of the theorem remains valid if we
consider J1 instead of Lebesgue measure A, and the RiemannStieltjes and
LebesgueStieltjes measures with respect to J1 instead of the Riemann and
Lebesgue integrals.
11. In this part we present a useful theorem on integration by parts for the
LebesgueStieltjes integral.
Let two generalized distribution functions F = F(x) and G = G(x) be
given on (R, ~(R)).
Theorem 11. The following formulas are valid for all real a and b, a < b:
F(b)G(b) F(a)G(a) = fF(s )dG(s)
+f
G(s) dF(s),
(62)
or equivalently
F(b)G(b) F(a)G(a)
fF(s)dG(s)
+ L
+ fG(s)dF(s)
llF(s) llG(s),
(63)
a<s:Sb
where F(s) = limqs F(t), !!.F(s) = F(s) F(s ).
Remark l. Formula (62) can be written symbolically in "differential" form
d(FG) = F _ dG
+ GdF.
(64)
6. Lebesgue Integral. Expectation
205
Remark 2. The conclusion of the theorem remains valid for functions F and G
of bounded variation on [a, b]. (Every such function that is continuous on
the right and has limits on the left can be represented as the difference oftwo
monotone nondecreasing functions.)
PRooF. We first recall that in accordance with Subsection 1 an integral
()means J1a,bJ ().Then (see formula (2) in 3)
J:
(F(b) F(a))(G(b) G(a)) =
dF(s) fdG(t).
Let F x G denote the direct product of the measures corresponding to F and
G. Then by Fubini's theorem
(F(b)  F(a))(G(b)  G(a)) =
=
f
f
(a, b] x (a, b]
d(F x G)(s, t)
(a, b] x (a, b]
I{s~t)(s, t) d(F
(G(s)  G(a)) dF(s)
~~
G(s)dF(s)
G)(s, t)
J(a,
b]
x (a, b]
I 1sst)(s, t) d(F
G)(s, t)
(F(t)  F(a)) dG(t)
~~
F(s )dG(s) G(a)(F(b) F(a)) F(a)(G(b) G(a)),
(65)
where I A is the indicator of the set A.
Formula (62) follows immediately from (65). In turn, (63) follows from
(62) if we observe that
f(G(s) G(s )) dF(s) =
a<~sbLlG(s) LlF(s).
(66)
Corollary 1. If F(x) and G(x) are distribution functions, then
F(x)G(x) = raoF(s) dG(s)
+ fao G(s) dF(s).
(67)
If also
F(x) = faof(s) ds,
then
F(x)G(x)
f ao F(s) dG(s)
+f
ao G(s)f(s) ds.
(68)
206
II. Mathematical Foundations of Probability Theory
Corollary 2. Let ~ be a random variable with distribution junction F(x) and
El~l" < oo. Then
00
{
xn dF(x)
= n {ooxn 1[1 F(x)] dx,
(69)
o lxlndF(x) =  JooxndF(x) = n [xn 1 F(x)dx
oo
0
0
(70)
and
El~ln =
f_
00
00
1XIn dF(x) = n Lnxn 1[1 F(x)
+ F(x)] dx.
(71)
To prove (69) we observe that
fXJ xn dF(x) =
s:
xn d(1  F(x))
=  bn(1 F(b))
+ n J:xn 1(1 F(x)) dx.
(72)
Let us show that since E 1 ~I" < oo,
bn(1 F(b)
+ F( b))::; bnP(I~I
In fact,
EJ~I"
Cf.)
k~1
rk
Jk_
2 b) t 0.
lxJ"dF(x) <
(73)
CXJ
and therefore
L rk
lxl"dF(x)tO,
L rk
lxln dF(x) 2
n too.
k;o,b+ 1 Jk1
But
k;o,b+ 1 Jk1
bnP(I~I
2 b),
which establishes (73).
Taking the limit as b t oo in (72), we obtain (69).
Formula (70) is proved similarly, and (71) follows from (69) and (70).
12. Let A(t ), t 2 0, be a function oflocally bounded variation (i.e., of bounded
variation on each finite interval [a, b]), which is continuous on the right and
has limits on the left. Consider the equation
z, = 1 + J~z._ dA(s),
(74)
207
6. Lebesgue Integral. Expectation
which can be written in differential form as
Zo = 1.
dZ = Z_ dA,
(75)
The formula that we have proved for integration by parts lets us solve (74)
explicitly in the class of functions of bounded variation.
We introduce the function
&,(A)= eA(rJA(OJ
(1
O:s;s:s;r
+ L\A(s))e&A(J,
(76)
where L\A(s) = A(s) A(s) for s > 0, and L\A(O) = 0.
The function A(s), 0 :::;; s :::;; t, has bounded variation and therefore has at
most a countable number of discontinuities, and so the series Lo::s:s::s:riM(s)l
converges. It follows that
O:s;s:s;r
(1
+ M(s))e&A(s)
is a function of locally bounded variation.
If Ac(t) = A(t)  Lo:s:s:s;r L\A(s) is the continuous component of A(t),
we can rewrite (?6) in the form
&,(A)= eAC(I)AC(O)
n (1 + M(s)).
O:s;s:s;r
(77)
Let us write
F(t) = eA<<rJA<<o>,
G(t) =
fl
O:s;s:s;r
(1
+ L\A(s)).
Then by (62)
tS',(A) = F(t)G(t) = 1 +
J~F(s) dG(s) + LG(s) dF(s)
1 + 0 J;,,F(s)G(s )L\A(s) + {G(s )F(s) dAc(s)
1 + {s._(A) dA(s).
Therefore tS',(A), t ~ 0, is a (locally bounded) solution of(74). Let us show that
this is the only locally bounded solution.
Suppose that there are two such solutions and let Y = Y(t), t ~ 0, be their
difference. Then
Y(t)
f~ Y(s) dA(s).
Put
T = inf{t ~ 0: Y(t) =I= 0},
where we take T
= oo if Y(t) = 0 fort~ 0.
208
II. Mathematical Foundations of Probability Theory
Since A(t) is a function of locally bounded variation, there are two
generalized distribution functions A 1(t) and A 2 (t) such that A(t) = A 1(t) A 2 (t). If we suppose that T < oo, we can find a finite T' such that
Then it follows from the equation
Y(t) =
I:
Y(s) dA(s),
t "2! T,
that
sup! Y(t)l ~!sup I Y(t)l
tSt'
tST'
and since sup! Y(t)l < oo, we have Y(t) = 0 forT< t
the assumption that T < oo.
Thus we have proved the following theorem.
T', contradicting
Theorem 12. There is a unique locally bounded solution of(74), and it is given
by (76).
13.
PROBLEMS
1. Establish the representation (6).
2. Prove the following extension of Property E. Let and '7 be random variables for
which Ee and E17 are defined and the sum Ee + E17 is meaningful (does not have the
form oo  oo or  oo + oo ). Then
3. Generalize Property G by showing that if = '7 (a.s.) and Ee exists, then E17 exists and
Ee = E17.
4. Let be an extended random variable, p. a afinite measure, and Jn lei dp. <
Show that I I < oo (p.a.s.) (cf. Property J).
00.
5. Let p. be a afinite measure, and '7 extended random variables for which Ee and E17
are defined. If J.A dP ::::;; J.A '7 dP for all A E ',then
'7 (p.a.s.). (Cf. Property 1.)
e: : ;
6. Let and '7 be independent nonnegative random variables. Show that Eel'f = Ee E17.
7. Using Fatou's lemma, show that
8. Find an example to show that in general it is impossible to weaken the hypothesis
"I I : : ; l'f, El'f < 00" in the dominated convergence theorem.
e.
209
6. Lebesgue Integral. Expectation
9. Find an example to show that in general the hypothesis "~.
Fatou's lemma cannot be omitted.
10. Prove the following variants ofFatou's lemma. Let the family
variables be uniformly integrable and let E lim ~. exist. Then
IJ, EIJ >  oo" in
g;}."' 1 of random
IJ., n 2:: 1, where the family {~:}."' 1 is uniformly integrable and 11.
converges a.s. (or only in probabilitysee 10 below) to a random variable 71 Then
limE~. s E lim~.
Let~.
11. Dirichlet's function
d(x) = {1,
0,
x irr~tional,
x rational,
is defined on [0, 1], Lebesgue integrable, but not Riemann integrable. Why?
12. Find an example of a sequence of Riemann integrable functions {f.}.;;: 1, defined on
[0, 1], such that I f. I s 1, f.> f almost everywhere (with Lebesgue measure), but
f is not Riemann integrable.
13. Let (a;,i; i,j 2:: 1) be a sequence ofreal numbers such that Li.i lai,il < oo. Deduce
from Fubini's theorem that
(78)
14. Findanexampleofasequ ence(aii; i,j;;::: 1)forwhich Li.i laiil = oo and the equation
in (78) does not hold.
15. Starting from simple functions and using the theorem on taking limits under the
Lebesgue integral sign, prove the following result on integration by substitution.
Let h = h(y) be a nondecreasing continuously differentiable function on [a, b],
and let f(x) be (Lebesgue) integrable on [h(a), h(b)]. Then the functionf(h(y))h'(y)
is integrable on [a, b] and
h(b)
f(x) dx =
fb
f(h(y))h'(y) dy.
h(a)
16. Prove formula (70).
17. Let ~.
~ 1 , ~ 2 ,
be nonnegative integrable random variables such that E~n ..... E~ and
P(~  ~. > ~:)> 0 for every 1: > 0. Show that then E I~n  ~I> 0, n> oo.
18. Let~. ry, (and ~n Yin (n, n 2:: 1, be random variables such that
p
'In > IJ,
E(n> E(,
and the expectations E~, EIJ, E( are finite. Show that then E~n ..... E~ (Pratt's lemma).
If also 1fn s 0 :::; (n then E I~.  ~I > 0,
Deduce that if~. f. ~. EI~. I > EI~ I and EI~ I < oo, then E I~n  ~I __. o.
210
II. Mathematical Foundations of Probability Theory
7. Conditional Probabilities and Conditional
Expectations with Respect to a aAlgebra
1. Let (0, ~ P) be a probability space, and let A E ' be an event such that
P(A) > 0. As for finite probability spaces, the conditional probability of B with
respect to A (denoted by P(BIA)) means P(BA)/P(A), and the conditional
probability of B with respect to the finite or countable decomposition
~ =
{Db D 2 } with P(Di) > 0, i ~ 1 (denoted by P(BI~))is the random variable
equal to P(BIDi) for wE Di, i ~ 1:
P(BI~)
P(BIDi)Iv,(w).
i;:. 1
In a similar way, if~ is a random variable for which E~ is defined, the
conditional expectation of~ with respect to the event A with P(A) > O(denoted
by E(~IA)) is E(OA)/P(A) (cf. (1.8.10).
The random variable P(B I~) is evidently measurable with respect to the
aalgebra r = a(~), and is consequently also denoted by P(B Ir) (see 8 of
Chapter 1).
However, in probability theory we may have to consider conditional
probabilities with respect to events whose probabilities are zero.
Consider, for example, the following experiment. Let ~ be a random
variable that is uniformly distributed on [0, 1]. If~ = x, toss a coin for which
the probability of head is x, and the probability of tail is 1  x. Let v be the
number of heads inn independent tosses of this coin. What is the "conditional
probability P( v = k I~ = x) "? Since P( ~ = x) = 0, the conditional pro bability P(v = k I~ = x) is undefined, although it is intuitively plausible that
"it ought to be C~xk(1  xtk."
Let us now give a general definition of conditional expectation (and, in
particular, of conditional probability) with respect to a aalgebra r, r ~ ~
and compare it with the definition given in 8 of Chapter I for finite probability
spaces.
2. Let (0, ~ P) be a probability space, r a aalgebra, r ~ ' (r is a asubalgebra of ff), and~ = ~(w) a random variable. Recall that, according to
6, the expectation E~was defined in two stages: first for a nonnegative random
variable ~' then in the general case by
and only under the assumption that
A similar twostage construction is also used to define conditional expectations E( ~I r).
7. Conditional Probabilities and Expectations with Respect to a uAigcbra
211
Definition 1.
( 1) The conditional expectation of a nonnegative random variable ~ with
respect to the aalgebra t is a nonnegative extended random variable,
denoted by E(~ It) or E(~ It)(m), such that
(a) E(~l~) is t'measurable;
(b) for every A E t
L~ LE(~l~)
dP =
dP.
(1)
(2) The conditional expectation E(~lt), or E(~l~)(m), of any random
variable ~ with respect to the aalgebra t, is considered to be defined if
min(E(~+ 1~), E(~1~))
< oo,
Pa.s., and it is given by the formula
EceJ~)
=E(~+ It) E(~1~),
where, on the set (of probability zero) of sample points for which E(~+ 1t)
= E(C I~) = oo, the difference E(~+ I~)  E(C I~) is given an arbitrary
value, for example zero.
We begin by showing that, for nonnegative random variables,
actually exists. By (6.36) the set function
Q(A) =
L~ dP,
A Et,
E(~l~)
(2)
is a measure on (Q, ~). and is absolutely continuous with respect to P
(considered on (Q, ~). ~ ;; ff). Therefore (by the RadonNikodym theorem)
there is a nonnegative ~measurable extended random variable EG I~) such
that
Q(A)
= LE(~I~)dP.
(3)
Then (1) follows from (2) and (3).
Remark 1. In accordance with the RadonNikodym theorem, the conditional expectation E( ~It) is defined only up to sets of Pmeasure zero.
In other words, E(~ It) can be taken to be any t'measurable function f(m)
for which Q(A) = JAf(m) dP, A E t (a "variant" of the conditional expectation).
Let us observe that, in accordance with the remark on the RadonNikodym theorem,
(4)
212
II. Mathematical Foundations of Probability Theory
i.e. the conditional expectation is just the derivative of the Radon Nikodym
measure Q with respect to P (considered on (0, <;9')).
Remark 2. In connection with (1), we observe that we cannot in general put
E(~ 1<;9') = ~,since~ is not necessarily <;9'measurable.
Remark 3. Suppose that ~ is a random variable for which E~ does not exist.
Then E(~ I<;9')maybe definable as a <;9'measurable function for which (1) holds.
This is usually just what happens. Our definition E(~l<;9') = E(~+ 1<;9')E(C 1<;9') has the advantage that for the trivial ualgebra <;9' = {0, 0} it
reduces to the definition of E~ but does not presuppose the existence of E~.
(For example, if~ is a random variable withE~+ = oo, E~ = oo, and <;9' = !F.,
then E~ is not defined, but in terms of Definition 1, E(~ 1<;9') exists and is simply
~=~+c.
Remark 4. Let the random variable~ have a conditional expectation E(~ 1<;9')
with respect to the ualgebra <;9'. The conditional variance (denoted by V( ~I <;9')
or V(~ I<;9')(tX)) of~ is the random variable
(Cf. the definition of the conditional variance V(~ I~) of~ with respect to a
decomposition~. as given in Problem 2, 8, Chapter 1.)
Definition 2. Let B E :F. The conditional expectation E(I B I<;9') is denoted by
P(B I<;9'), or P(B I<;9')( w), and is called the conditional probability of the event B
with respect to the ualgebra <;9', <;9' s;;; :F.
It follows from Definitions 1 and 2 that, for a given B E !F., P(B 1<;9') is a
random variable such that
(a) P(B I<;9') is <;9'measurable,
(b)
for every A
P(A n B)= LP(BI<;9')dP
E
(5)
<;9'.
Definition 3. Let ~ be a random variable and <;9'~ the ualgebra generated by
a random element '1 Then E(~1<;9'n), if defined, means E(~l'1 or E(~IIJ)(w),
and is called the conditional expectation of~ with respect to '1
The conditional probability P(BI<;9'71) is denoted by P(BIIJ) or P(BIIJ)(w),
and is called the conditional probability of B with respect to '1
3. Let us show that the definition of E( ~I <;9') given here agrees with the definition of conditional expectation in 8 of Chapter I.
213
7. Conditional Probabilities and Expectations with Respect to a uAigebra
Let~ = {D 1 , D 2 , .. } be a finite or countable decomposition with atoms
D; with respect to the probability P (i.e. P(D;) > 0, and if A ~ D;, then
either P(A) = 0 or P(D;\A) = 0).
Theorem 1. If~ = a(~) and ~ is a random variable for which E~ is defined,
then
EW~)
E(~ID;)
(Pa.s. on D;)
EW~
E(OD.)
P(D;)
(Pa.s. on D;).
(6)
or equivalently
(The
notation"~ =
"~
'1 (Pa.s. on A)," or
= 17(A; Pa.s.)" means that P(A n
PRooF. According to Lemma 3 of 4, E(~l~)
constants. But
{~
= K;
::/: 17}) = 0.)
on D;, where K; are
whence
1
E(OD)
K; = P(D;) JD,~ dP = P(D;) = EWD;).
This completes the proof of the theorem.
Consequently the concept of the conditional expectation
E(~ 1~)
with
respect to a finite decomposition ~ = {Db ... , Dn}, as introduced in
Chapter I, is a special case of the concept of conditional expectation with
respect to the aalgebra ~ = a(D).
4. Properties of conditional expectations. (We shall suppose that the expectations are defined for all the random variables that we consider and that
~ ~ $'.)
A*. IfC is a constant and~= C (a.s.), then E(~l~) = C (a.s.).
B*. If~ ~ '7 (a.s.) then EW ~) ~ E('71 ~)(a.s.).
C*. IEW~)I ~ E(l~ll~) (a.s.).
D*. If a, bare constants and aE~ + bE17 is defined, then
E(a~
+ b1'fl ~)
aEW ~)
+ bE('71 ~)
E*. Let ff'* = { (/), Q} be the trivial aalgebra. Then
E(~lff'*) = E~
(a.s.).
(a.s.).
214
F*.
II. Mathematical Foundations of Probability Theory
E(~\~)
~(a.s.).
G*. E(E(ell)) =
E~.
H*. Ift 1 t 2 then
I*. If t 1
;2
t 2 then
J*. Let a random variable for which Ee is defined be independent of the
aalgebra t (i.e., independent of IB, Bet). Then
E(eit) = Ee
(a.s.).
K*. Let '7 be a t'measurable random variable, EI'71 < oo and EIe'7\ < oo.
Then
E(e'71t) = '7E(e\t)
(a.s.).
Let us establish these properties.
A*. A constant function is measurable with respect to t. Therefore we need
only verify that
{edP
But, by the
hypothesis~ =
= {cdP,
Aet.
C (a.s.) and Property G of 6, this equation
is obviously satisfied.
B*. If $ '7 (a.s.), then by Property B of 6
L~ f}
dP $
dP,
A E t,
and therefore
{E(e\l)dP $ {E('71t)dP,
Aet.
The required inequality now follows from Property I (6).
C*. This follows from the preceding property if we observe that 
e$I~\.
D*. If A e t then by Problem 2 of 6,
{ (ae + brl) dP = {ae dP + J}'1 dP =
+{
which establishes D*.
bE('71 t) dP
= {
{ aE(e\t) dP
[aE( ~It)
+ bE('7\ t)] dP,
IeI
215
7. Conditional Probabilities and Expectations with Respect to a aAlgebra
E*. This property follows from the remark that E~ is an "*measurable
function and the evident fact that if A = 0 or A = 0 then
F*.
Since~
if "measurable and
L~dP = L~dP,
we have EW F) = ~ (a.s.).
G*. This follows from E* and H* by taking
H*. Let A E I' 1 ; then
Since I' 1 s;:
I' 2 ,
AE;
I' 1 =
{0, 0} and I' 2 =
1'.
we have A E I' 2 and therefore
Consequently, when A E 1' 1,
LE(~ll'1) LE[E(~II'2)11'1]
dP =
dP
and by Property I (6) and Problem 5 (6)
E(~ll'1)
I*. If A
E ~ 1o
E[E(~II'2)11'1]
(a.s.).
then by the definition of E[E( ~I~ 2 ) I~ 1 ]
The function E( ~II' 2) is I' rmeasurable and, since I' 2 s;: I' to also
I' rmeasurable. It follows that E( ~II' 2) is a variant of the expectation
E[E(~II' 2 )11' 1 ], which proves Property 1*.
J*. Since E~ is a /'measurable function, we have only to verify that
dP =
LE~dP,
i.e. that E[ ~ I B] = E~ EJB. If EI~ I < oo, this follows immediately from
Theorem 6 of 6. The general case can be reduced to this by applying
Problem 6 of 6.
The proof of Property K* will be given a little later; it depends on conclusion (a) of the following theorem.
216
II. Mathematical Foundations of Probability Theory
Theorem 2 (On Taking Limits Under the Expectation Sign). Let
be a sequence of extended random variables.
{en}n~l
(a) If len I~ 17, E17 < oo and en. (a.s.), then
ECenl~). EW~
(a.s.)
ECienell~).0
(a.s.).
and
(b) If en
17, E17 >  oo and en j
e(a.s.), then
ECenl~)
(c) If en
(a.s.).
l E(el~)
(a.s.).
17, E17 >  oo, then
E(lim enl~)
(e) If en
Ecei~)
17, E17 < oo, and en l (a.s.), then
ECenl~)
(d) If en
~lim E(enl~)
(a.s.).
~ E(lim e"l~)
(a.s.).
17, E17 < oo, then
Ilm E(enl~)
(f) If en ~ 0 then
E(L en I~)=
L E(enl~)
(a.s.).
(a) Let Cn = SUPm>n Iem el. Since en__.. (a.s.), we have '" l 0
E~n and E~ are finite; therefore by Properties D*
and C* (a.s.)
PROOF.
(a.s.). The expectations
IECenl~) E(el~)l
Since E(Cn+ 1 1~)
= IECen
el~)l ~
E(len
ell~)~ E(Cnl~).
E(Cn I~) (a.s.), the limit h = lim" E(Cn I~) exists (a.s.). Then
0~ LhdP~ {ECCnl~)dP= {CndP.0,
n.oo,
where the last statement follows from the dominated convergence theorem,
since 0 ~ Cn ~ 217, E17 < oo. Consequently Jn h dP = 0 and then h = 0
(a.s.) by Property H.
(b) First let 11 = 0. Since ECenl~) ~ ECen+tl~) (a.s.) the limit C(m) =
lim" ECenl~) exists (a.s.). Then by the equation
Len dP = LE(enl~) dP,
AE
~.
and the theorem on monotone convergence,
LeaP= LeaP,
Consequently
Ae~.
e= C(a.s.) by Property I and Problem 5 of 6.
217
7. Conditional Probabilities and Expectations with Respect to a uAlgcbra
For the proof in the general case, we observe that 0:::; ~: j ~+,and by
what has been proved,
E(~: /~)
But 0 s ~;; s
C, E~ <
E(~+ /~)
(a.s.).
(7)
oo, and therefore by (a)
E(~;; /~)+
E(C /~),
which, with (7), proves (b).
Conclusion (c) follows from (b).
(d) Let (n = infm~n ~m; then (n j (, where ( = lim ~n According to (b),
E((n/~) i E((/~) (a.s.). Therefore (a.s.) E(lim ~n/~) = E((/~) =limn E((n/~)
= lim E((J ~) $ lim E(~n /~).
Conclusion (e) follows from (d).
(f) If ~n ~ 0, by Property D* we have
E(J ~kl~) = J E(~k/~)
1
(a.s.)
which, with (b), establishes the required result.
This completes the proof of the theorem.
We can now establish Property K*. Let '1 = 18 ,
AE~,
f ~'1
A
dP =
A,,B
~ dP =
A,,B
E(~/~) dP =
BE~
f l8 E(~/~)
A
dP =
Then, for every
f 1JE(~/~)
A
dP.
By the additivity of the Lebesgue integral, the equation
(8)
remains valid for the simple random variables 1J = L~=l ykl 8 k, BkE~.
Therefore, by Property I (6), we have
(9)
for these random variables.
Now let 1J be any ~measurable random variable with E 1171 < oo, and
let {17n} n~ 1 be a sequence of simple ~measurable random variables such that
1'7n I s 1] and 'ln + 11 Then by (9)
E(~17nl~)
= 17nE(~/~)
(a.s.).
It is clear that 1~'7n/ si~IJI, where E/~nl < oo. Therefore E(~17nl~)+
E(~17/~) (a.s.) by Property (a). In addition, since E I~ I < oo, we have E(~l~)
finite (a.s.) (see Property C* and Property J of 6). Therefore 17nEC~I~)+
17E( ~I~) (a.s.). (The hypothesis that E( ~I~) is finite, almost surely, is essential,
since, according to the footnote on p. 172, 0 oo = 0, but if 'ln = 1/n, 1J = 0,
we have 1/n oo
0 oo = 0.)
218
II. Mathematical Foundations of Probability Theory
5. Here we consider the more detailed structure of conditional expectations
E(el~,). which we also denote, as usual, by E(el'l).
Since E(el'l) is a ~.,measurable function, then by Theorem 3 of 4 (more
precisely, by its obvious modification for extended random variables) there
is a Borel function m = m(y) from R toR such that
m('l(ro))
for all
OJ E
n.
= E(el'l)(ro)
(10)
we denote this function m(y) by E(
eI, = y) and call it the
conditional expectation of with respect to the event {, = y}' or the conditional
expectation of under the condition that , = y.
Correspondingly we define
(11)
Ae~.,.
Therefore by Theorem 7 of 6 (on change of variable under the Lebesgue
integral sign)
m(17) dP =
J{ro:.,eB}
f m(y)P,(dy),
JB
BeBI(R),
where P., is the probability distribution of 'I Consequently m
Borel function such that
(12)
m(y) is a
edP= fm(y)dP,.
J{ro: 11eB}
(13)
JB
for every B e BI(R).
This remark shows that we can give a different definition of the conditional
expectation E(el '1 = y).
Definition 4. Let and '1 be random variables (possible, extended) and let
Ee be defined. The conditional expectation of the random variable under
the condition that 'I = y is any BI(R)measurable function m = m(y) for
which
J{ro:.,eB}
edP = JBf m(y)P.,(dy),
BeBI(R).
(14)
That such a function exists follows again from the RadonNikodym theorem
if we observe that the set function
Q(B) =
J{ro: 11eB}
edP
is a signed measure absolutely continuous with respect to the measure P,.
7. Conditional Probabilities and Expectations with Respect to a aAlgebra
219
Now suppose that m(y) is a conditional expectation in the sense of Definition 4. Then if we again apply the theorem on change of variable under the
Lebesgue integral sign, we obtain
J{ro:~eB}
~ dP =
rm(y)P~(dy) = J{ro:~eB}
r m(11)P~(dy),
JB
BE &H(R).
The function m(17) is ~~measurable, and the sets {w: 11 E B}, BE &H(R),
exhaust the subsets of ~~.
Hence it follows that m(IJ) is the expectation E(~ IIJ). Consequently if we
know E(~l11 = y) we can reconstruct E(~l17), and conversely from E(~l11) we
can find E(~ I11 = y).
From an intuitive point of view, the conditional expectation E(~ 1'1 = y)
is simpler and more natural than E(~ I17). However, E(~ I17), considered as a
~~measurable random variable, is more convenient to work with.
Observe that Properties A*K* above and the conclusions of Theorem 2
can easily be transferred to E(~l11 = y) (replacing "almost surely" by
"P~almost surely"). Thus, for example, Property K* transforms as follows:
if E 1~1 < oo and E llf(IJ)I < oo, where f = f(y) is a &H(R) measurable function, then
E(lf(11)111 = y) = f(y)E(~IIJ = y)
In addition (cf. Property J*),
E(~l11
if~
(P~a.s.).
(15)
and 11 are independent, then
= y) = E~
(P~a.s.).
We also observe that if BE &H(R 2 ) and ~and 11 are independent, then
(16)
and if cp = cp(x, y) is a .c?6'(R 2 )measurable function such that E Icp(~, 1J) I < oo,
then
To prove (16) we make the following observation. If B = B 1 x B 2 , the
validity of (16) will follow from
J{ro:~eA}
IB,xB2(~. 11)P(dw) =
J(yeA)
EIB,xBi~. y)P~(dy).
But the lefthand side is P{~ E Bt. 11 E An B 2 }, and the righthand side is
P( ~ E B 1)P(17 E A n B 2); their equality follows from the independence of ~
and 11 In the general case the proof depends on an application of Theorem 1,
2, on monotone classes (cf. the corresponding part of the proof of Fubini's
theorem).
Definition 5. The conditional probability of the event A E ~under the condition that 11 = y (notation: P(A I11 = y)) is E(J A I1J = y).
220
II. Mathematical Foundations of Probability Theory
It is clear that P(A 111 = y) can be defined as the .14(R)measurable function
such that
P(A n {IJEB}) =
ft(AIIJ
y)P~(dy),
BE .14(R).
(17)
6. Let us calculate some examples of conditional probabilities and conditional expectations.
EXAMPLE
Lk'=
P(IJ
1. Let 11 be a discrete random variable with P(IJ = Yk) > 0,
= Yk) = 1. Then
P(AI
11
Yk
{17 = yk})
P( _ ) ,
11 Yk
) = P(A n
For y{y 1 ,y 2 , ... } the conditional probability P(AIIJ = y) can be defined
in any way, for example as zero.
If~ is a random variable for which E~ exists, then
When y {y 1, y 2 , . } the conditional expectation E( ~ 111
in any way (for example, as zero).
= y) can be defined
2. Let ( ~, 11) be a pair of random variables whose distribution has a
density ~~~(x, y):
EXAMPLE
P{(~,IJ)EB} = f!~~(x,y)dxdy,
Let Hx) and UY) be the densities of the probability distribution of~ and 11
(see (6.46), (6.55) and (6.56).
Let us put
r ( I )  h~(x, y)
X y j,(y) '
(18)
J~l~
taking J~ 1 ~(x Iy) = 0 if f~(y) = 0.
Then
P(~ E CIIJ = y) = J/~ 1 ~(xly) dx,
C E .14(R),
i.e. !~ 1 ~(x ly) is the density of a conditional probability distribution.
(19)
7. Conditional Probabilities and Expectations with Respect to a aAlgebra
221
In fact, in order to prove (19) it is enough to verify (17) for BE PJ(R),
{~ E C}. By (6.43), (6.45} and Fubini's theorem,
L[J/~~~(xly)
dx
JP~(dy) = [L.h 1 ~(xly) dx Jf~(y) dy
=
J J~ 1 ~(xiy)~(y)
J f~~(x,
dx dy
CxB
y) dx dy
CxB
P{(~,
IJ) E C x B}
P{(~ E
C) n (IJ E B)},
which proves (17).
In a similar way we can show that if E~ exists, then
(20)
EXAMPLE
3. Let the length of time that a piece of apparatus will continue to
operate be described by a nonnegative random variable 11 = IJ(w) whose
distribution Fq(y) has a density j~(y) (naturally, F ~(y) = j~(y) = 0 for y < 0).
Find the conditional expectation E(IJ  a 111 2 a), i.e. the average time for
which the apparatus will continue to operate on the hypothesis that it has
already been operating for time a.
Let P(IJ 2 a) > 0. Then according to the definition (see Subsection 1) and
(6.45),
E(11  a I'1 > a)

E[(l'/  a)I 1paJ
P(l'/ 2 a)
Jn (1'/ 
a)I 1 ~>a)P(dw)
"=''=c'"''=='___:_____:__
P(l'/ 2 a)
_ J:' (y 
 s:
a)f~(y) dy
f~(y) dy
It is interesting to observe that if '1 is exponentially distributed, i.e.
A. .<y
>0
f,(y) = { e ' y '
~
0
y < 0,
(21)
then El] = E(IJIIJ 2 0) = 1/A. and E(IJ aiiJ 2 a)= 1/A. for every a> 0. In
other words, in this case the average time for which the apparatus continues
to operate, assuming that it has already operated for time a, is independent
of a and simply equals the average time El].
Under the assumption (21) we can find the conditional distribution
P(l'/ a :S:: xl11 2 a).
222
II. Mathematical Foundations of Probability Theory
We have
P(
11  a ~ x '1 ;;;:: a =
P(a ~ '1 ~ a +. . x)
P('7 :2:: a)
F,(a
[1 
=e
+ x) F,(a) + P('1 =a)
1  F,(a) + P(17 =a)
[1  e..ta]
1 [1  e .l.a]
e..t(a+x)] 
.l.a[1 _ .l.x]
;. e
e a
=1
e.l.x.
Therefore the conditional distribution P('1  a ~ xI 11 :2:: a) is the same
as the unconditional distribution P(17 ~ x). This remarkable property
is unique to the exponential distribution: there are no other distributions
thathavedensitiesandpossessthepropertyP('1 a~ xl'1 :2:: a)= P('1 ~ x),
a :2:: 0, 0 ~ x < oo.
EXAMPLE 4 (Buffon's needle). Suppose that we toss a needle of unit length
"at random" onto a pair of parallel straight lines, a unit distance apart, in
a plane. What is the probability that the needle will intersect at least one of the
lines?
To solve this problem we must first define what it means to toss the
needle "at random." Let be the distance from the midpoint of the needle to
the lefthand line. We shall suppose that~ is uniformly distributed on [0, 1],
and (see Figure 29) that the angle (} is uniformly distributed on [ n/2, n/2].
In addition, we shall assume that and fJ are independent.
Let A be the event that the needle intersects one of the lines. It is easy to
see that if
B = {(a, x): Ia I ~
1[
2'
x E [0, teas a] u [1  tcos a, 1]},
then A = {w: (fJ, e) E B}, and therefore the probability in question is
P(A)
= EIA(w) = Ela(fJ(w), e(w)).
Figure 29
7. Conditional Probabilities and Expectations with Respect to a aAlgebra
223
By Property G* and formula (16),
EJa(tl(w),
~(w)) =
E(E[Ja(tl(w),
~(w))ltl(w)])
= LE[I 8 (tl(w), ~(w))ltl(w)]P(dw)
=
tt/2
=1
_,12 E[I8 (tl(w), ~(w))ltl(w) = tx]P8(da)
J"
12
tt/2
Ela(a,
~(w))da = 1
J"
12
n/2
cos ada=,
2
n
where we have used the fact that
Ela(a, ~(w)) = P{~ E [0,
t cos a] u [1  t cos a]}= cos a.
Thus the probability that a "random" toss of the needle intersects one of
the lines is 2/n. This result could be used as the basis for an experimental
evaluation of n. In fact, let the needle be tossed N times independently.
Define ~ito be 1 if the needle intersects a line on the ith toss, and 0 otherwise.
Then by the law of large numbers (see, for example, (1.5.6))
P{l~ 1 + ; + ~N P(A)I > e}+ 0,
N+
oo.
for every e > 0.
In this sense the frequency satisfies
~1
+ ... + ~N
N
P(A) = ~2
and therefore
2N
~1!:.
~1
+ ... +
~N
This formula has actually been used for a statistical evaluation of n. In
1850, R. Wolf (an astronomer in Zurich) threw a needle 5000 times and
obtained the value 3.1596 for n. Apparently this problem was one of the first
applications (now known as Monte Carlo methods) of probabilisticstatistical regularities to numerical analysis.
7. If { ~n}n> 1 is a sequence of nonnegative random variables, then according
to conclusion (f) of Theorem 2,
In particular, if B 1 , B 2 , is a sequence of pairwise disjoint sets,
(22)
224
II. Mathematical Foundations of Probability Theory
It must be emphasized that this equation is satisfied only almost surely
and that consequently the conditional probability P(B \~)(w) cannot be
considered as a measure on B for given w. One might suppose that, except
for a set ..;V of measure zero, P( I~) (w) would still be a measure for w E Y.
However, in general this is not the case, for the following reason. Let
.K(B 1 , B 2 , ) be the set of sample points w such that the countable additivity
property (22) fails for these B 1, B 2 , ... Then the excluded set ..;Vis
(23)
where the union is taken over all B 1, B 2 , . in fJ'. Although the Pmeasure
of each set .K(B 1 , B 2 , . . ) is zero, the Pmeasure of ..;V can be different from
zero (because of an uncountable union in (23)). (Recall that the Lebesgue
measure of a single point is zero, but the measure of the set ..;V = [0, 1],
which is an uncountable sum of the individual points {x}, is 1).
However, it would be convenient if the conditional probability P(\ ~)(w)
were a measure for each wEn, since then, for example, the calculation of
conditional probabilities E(e \~)could be carried out (see Theorem 3 below)
in a simple way by averaging with respect to the measure P( I~)(w):
E(e\~) =
fa ~(w)P(dw\~)
(a.s.)
(cf. (1.8.10)).
We introduce the following definition.
Definition 6. A function P(w; B), defined for all wEn and BE ff, is a regular
conditional probability with respect to ~ if
(a) P(w; )is a probability measure on$' for every wEn;
(b) For each B E $' the function P( w; B), as a function of w, is a variant of the
conditional probability P(B\~)(w), i.e. P(w: B)= P(B\~)(w) (a.s.).
Theorem3. Let P(w; B) be a regular conditional probability with respect to
~ and let be an integrable random variable. Then
Ece\~)(w) = [ ~(w)P(w; dw)
Jn
PROOF.
If
e= I
B'
(a.s.).
(24)
BE ff, the required formula (24) becomes
P(B\~)(w) =
P(w; B) (a.s.),
which holds by Definition 6(b). Consequently (24) holds for simple functions.
7. Conditional Probabilities and Expectations with Respect to a aAlgebra
225
Now let ~ 0 and n i , where n are simple functions. Then by (b) of
Theorem 2 we have E(/~)(w) =limn E(nl~)(w) (a.s.). But since P(w; )
is a measure for every wE Q, we have
lim
n
E(n/~)(w) =lim
n
( kiJ)P(w; dw) = ( (w)P(w; dw)
Jn
Jn
by the monotone convergence theorem.
The general case reduces to this one if we use the representation =
e+ c.
This completes the proof.
Corollary. Let ~ = ~~' where rt is a random variable, and let the pair (, rt)
have a probability distribution with density f~lx, y). Let E /g() / < oo. Then
where f~ 1 ~(x/y) is the density of the conditional distribution (see (18)).
In order to be able to state the basic result on the existence of regular
conditional probabilities, we need the following definitions.
Definition 7. Let (E, r!) be a measurable space, X= X(w) a random element
with values in E, and ~ a asubalgebra of g;, A function Q(w; B), defined
for wEn and BE S is a regular conditional distribution of X with respect to
~if
(a) for each wEn the function Q(w; B) is a probability measure on (E, rff);
(b) for each B E C the function Q( w; B), as a function of w, is a variant of the
conditional probability P(X E B/ ~)(w), i.e.
Q(w; B)= P(X EB/~)(w)
(a.s.).
Definition 8. Let be a random variable. A function F = F(w; x), wE Q,
XE
R, is a regular distribution function for with respect to ~
if :
(a) F(w; x) is, for each wE Q, a distribution function on R;
(b) F(w; x) = P( :.:::; x /~)(w)(a.s.), for each x E R.
Theorem 4. A regular distribution function and a regular conditional distribution function always exist for the random variable
ewith respect to ~.
226
II. Mathematical Foundations of Probability Theory
PROOF. For each rational number r E R, define F,(OJ) = P(~ :::; ri~)(OJ),
where P(~:::; ri~)(OJ) = E(J 1 ~ 9li~)(OJ) is any variant of the conditional
probability, with respect to~. of the event {~:::; r}. Let {r;} be the set of
rational numbers in R. If r; < ri, Property B* implies that P(~ :::; r;i ~) :::;
P(~:::; rii~) (a.s.), and therefore if Aii = {OJ: F,/OJ) < F,,(OJ)}, A= UAii
we have P(A) = 0. In other words, the set of points OJ at which the distribution function F,(OJ), r E {r;}, fails to be monotonic has measure zero.
Now let
B; = {OJ: lim Fr,+( 1 /n)(OJ) = F,,(OJ)l,
n+ oo
00
B=
U B;.
i= 1
It is clear that /{~:Sr,+( 1 /n)}! /{~~ril n+ oo. Therefore, by (a) of Theorem 2,
F,,+( 1 /nl(OJ)+ F,,(OJ) (a.s.), and therefore the set Bon which continuity on the
right fails (with respect to the rational numbers) also has measure zero,
P(B) = 0.
In addition, let
c=
Then, since
P(C) = 0.
Now put
{OJ: lim FnCOJ) = 1} u {OJ: lim F nCOJ) >
n+ oo
n+ oo
g :::; n} i Q,
n+ oo, and {~ :::; n}!
F(W,. X )
0,
o}.
n+  oo, we have
= {lim F,(OJ), OJ A u B u C,
r!x
G(x),
wE
Au B u C,
where G(x) is any distribution function on R; we show that F(OJ; x) satisfies
the conditions of Definition 8.
Let Au B u C. Then it is clear that F(OJ; x) is a nondecreasing function ofx. Ifx < x' :::; r, then F(OJ; x) :::; F(OJ; x'):::; F(w; r) = F,(OJ)! F(w, x)
when r! x. Consequently F(OJ; x) is continuous on the right. Similarly
limx. 00 F(OJ; x) = 1, limx.oo F(OJ; x) = 0. Since F(OJ; x) = G(x) when
OJ E Au B u C, it follows that F(OJ; x) is a distribution function on R for
every OJ E Q, i.e. condition (a) of Definition 8 is satisfied.
By construction, P(~:::; r)I~)(OJ) = F,(OJ) = F(OJ; r). If r! x, we have
F(OJ; r)! F(OJ; x) for all OJ E Q by the continuity on the right that we just
established. But by conclusion (a) of Theorem 2, we have P(~ :::; ri ~)(OJ)+
P(~:::; xi~)(OJ) (a.s.). Therefore F(OJ; x) = P(~:::; xiG)(OJ) (a.s.), which
establishes condition (b) of Definition 8.
We now turn to the proof of the existence of a regular conditional distribution of~ with respect to ~Let F(OJ; x) be the function constructed above. Put
Q(OJ; B) = LF(OJ; dx),
227
7. Conditional Probabilities and Expectations with Respect to auAlgebra
where the integral is a LebesgueStieltjes integral. From the properties of
the integral (see 6, Subsection 7), it follows that Q(ro; B) is a measure on B
for each given wEn. To establish that Q(ro; B) is a variant of the conditional
probability P(~ E Bl~)(ro), we use the principle of appropriate sets.
Let ~ be the collection of sets B in gj(R) for which Q(ro; B)=
P(~EBI~)(ro) (a.s.). Since F(w;'x) = P(~ ~ xl~)(w) (a.s.), the system~
contains the sets B of the form B = ( oo, x], x E R. Therefore ~ also
contains the intervals of the form (a, b], and the algebra d consisting of finite
sums of disjoint sets of the form (a, b]. Then it follows from the continuity
properties of Q(w; B) (w fixed) and from conclusion (b) of Theorem 2 that~
is a monotone class, and since d ~ gj(R), we have, from Theorem 1
of2,
gj(R) = a(d)
a(~)= Jl(~)
gj(R),
whence ~ = gj(R).
This completes the proof of the theorem.
By using topological considerations we can extend the conclusion of
Theorem 4 on the existence of a regular conditional distribution to random
elements with values in what are known as Borel spaces. We need the following definition.
Definition 9. A measurable space (E, tf) is a Borel space if it is Borel equivalent
to a Borel subset of the real line, i.e. there is a onetoone mapping cp = cp(e):
(E, tf)
+
(R, gj(R)) such that
(1) cp(E) {cp(e): eeE} is a set in gj(R);
(2) cp is tS'measurable (cp 1(A) E I, A E cp(E) n gj(R)),
(3) cp 1 is BI(R)/tS'measurable (cp(B) E cp(E) n BI(R), BE Iff).
Theorem 5. Let X = X(w) be a random element with values in the Borel space
(E, S). Then there is a regular conditional distribution of X with respect to ~Let cp = cp(e) be the function in Definition 9. By (2), cp(X(ro)) is a
random variable. Hence, by Theorem 4, we can define the conditional
distribution Q(ro; A) of cp(X(ro)) with respect to r, A E cp(E) n BI(R).
We introduce the function Q(w; B) = Q(w; cp(B)), BE tf. By (3) of
Definition 9, qJ(B) E qJ(E) n BI(R) and consequently Q(w; B) is defined.
Evidently Q(w; B) is a measure on BE S for every ro. Now fix BE G. By the
onetoone character of the mapping cp = cp(e),
PROOF.
Q(w; B)= Q(w; cp(B)) = P{cp(X) E cp(B)I~} = P{X E Bl~}
(a.s.).
Therefore Q(w; B) is a regular conditional distribution of X with respect
to~.
This completes the proof of the theorem.
228
II. Mathematical Foundations of Probability Theory
Corollary .. Let X = X( w) be a random element with values in a complete separable metric space (E, 6"). Then there is a regular conditional distribution of X
with respect to t. In particular, such a distribution exists for the spaces
(R", ~(R")) and (R 00 , ~(R 00 )).
The proof follows from Theorem 5 and the well known topological result
that such spaces are Borel spaces.
8. The theory of conditional expectations developed above makes it possible
to give a generalization of Bayes's theorem; this has applications in statistics.
with
Recall that if f: = {A 1 , . . . , An} is a partition of the space
that
P(A;) > 0, Bayes's theorem (1.3.9) states
P(A; IB)
P(A;)P(B IA;)
Li=l P(A)P(BIA)
for every B with P(B) > 0. Therefore if e = 2:? = a ;I
1
.
A,
(25)
is a discrete random
variable then, according to (1.8.10),
E[g(O)IBJ
L?=: g(a;)P(A;)P(BIA;)'
Li=l P(A)P(BIA)
(26)
J~oo g(a)P(BIO = a)P9(da)
J~oo P(BIO = a)P 9(da)
(27)
or
E[g(O)IBJ
On the basis of the definition ofE[g(B)JB] given at the beginning of this
section, it is easy to establish that (27) holds for all events B with P(B) > 0,
random variables and functions g = g(a) with E Ig(O) I < oo.
We now consider an analog of (27) for conditional expectations
E[g( 8) It] with respect to a aalgebra t, t ~ ff.
Let
Q(B) = Lg(O)P(dw),
BEt.
(28)
Then by (4)
(29)
We also consider the aalgebra t 9 . Then, by (5),
P(B) =
Lt<B I
t9)
dP
(30)
or, by the formula for change of variable in Lebesgue integrals,
(31)
7. Conditional Probabilities and Expectations with Respect to a (jAlgebra
229
Since
we have
Q(B)
f_'Xl}(a)P(BIO = a)P 0(da).
Now suppose that the conditional probability P(B I0
admits the representation
P(BIO =a)= Lp(w; a)A(dw),
(32)
= a) is regular and
(33)
where p = p(w; a) is nonnegative and measurable in the two variables
jointly, and A is a afinite measure on (Q, <;).
Let E lg(O)I < oo. Let us show that (Pa.s.)
E[g(O)I<;] = J~oo g(a)p(w; a)Po(da)
J~ oo p(w; a)P 0(da)
(34)
(generalized Bayes theorem).
In proving (34) we shall need the following lemma.
Lemma. Let (Q, F) be a measurable space.
(a) Let 11 and A be afinite measures, and f
f(w) an ffmeasurablefunction.
Then
(35)
(in the sense that !feither integral exists, the other exists and they are equal).
(b) If v is a signed measure and fl, A are afinite measures v ~ fl, 11 ~ A, then
dv
dA
dv dfl
dfl. dA.
(A.a.s.)
(36)
(11a.s.)
(37)
and
dv
dfl
PROOF. (a) Since
I!( A)
dvldfl
dA. dA.
L(~~)
dA.,
I/;
I A,. The general case
(35) is evidently satisfied for simple functions f =
monotone converthe
and
f+
f
=
f
follows from the representation
gence theorem.
230
II. Mathematical Foundations of Probability Theory
(b) From (a) with f = dv/d/1 we obtain
Then v
A. and therefore
J dA.dv dA.,
v(A)
whence (36) follows since A is arbitrary, by Property I (6).
Property (37) follows from (36) and the remark that
dJ.l = 0} =
11 { w:dA.
dJ.l
{ro: d!lfd).
= 0) dA.
dA. = 0
(on the set {w: dJ1/dA. = 0} the righthand side of (37) can be defined arbitrarily, for example as zero). This completes the proof of the lemma.
To prove (34) we observe that by Fubini's theorem and (33),
Q(B)
L[J:
L[f_
P(B) =
00
(38)
g(a)p(w; a)P0(da)}(dw),
00
00
(39)
p(w; a)Po(da)}(dw).
Then by the lemma
dQ
dQjd).
dP = dPjd).
(Pa.s.).
Taking account of (38), (39) and (29), we have (34).
Remark. Formula (34) remains valid if we replace() by a random element
with values in some measurable space (E, C) (and replace integration over
R by integration over E).
Let us consider some special cases of (34).
Let the aalgebra r be generated by the random variable
Suppose that
P(~ E A I() = a) = {
q(x; a)A.(dx),
~' r
A E fJIJ(R),
r ~.
(40)
where q = q(x; a) is a nonnegative function, measurable with respect to both
variables jointly, and ). is a afinite measure on (R, fJIJ(R)). Then we obtain
E[g(())l~ = x] = J:?oo g(a)q(x; a)Po(da)
J:?oo q(x; a)Po(da)
(41)
7. Conditional Probabilities and Expectations with Respect to a uAlgebra
231
In particular, let((},~) be a pair of discrete random variables,(} = La;! A,,
xiB j" Then, taking A. to be the counting measure (A.( {x;}) = 1, i = 1, 2, ...)
we find from (40) that
~='I
(Compare (26).)
Now let (0, ~) be a pair of absolutely continuous measures with density
fo.~(a, x). Then by (19) the representation (40) applies with q(x; a)=
!~ 16 (xla) and Lebesgue measure A.. Therefore
E[g(O) I~ = x] = s~ <Xl g(a)f~lfJ(x la)fo(a) da
J~ oo J~ 1 o(x Ia)fo(a) da
9.
(43)
PROBLEMS
1. Let eand '1 be independent identically distributed random variables with Ee defined.
Show that
E(ele + '1) = E(IJie + '1) = e+IJ
2 (a.s.).
ee
2. Let 1, 2 , be independent identically distributed random variables with
E1e;1 < oo. Show that
s.
E(e11 s., s.+ 1 ...) =  (a.s.),
n
where s.
e + ... + e.
1
3. Suppose that the random elements (X, Y) are such that there is a regular distribution
Px(B) = P(YeBJX = x). Show that ifE Jg(X, Y)l < oo then
E[g(X, Y)IX = x] =
4. Let
g(x, y)Px(dy)
(Pxa.s.).
ebe a random variable with distribution function Fix). Show that
E<eJIX <
(assuming that F~(b)
F~(a)
J!x dFix)
e~b)= F~(b) Fia)
> 0).
5. Let g = g(x) be a convex Borel function with E Jg<e)J < oo. Show that Jensen's
inequality
holds for the conditional expectations.
232
II. Mathematical Foundations of Probability Theory
6. Show that a necessary and sufficient condition for the random variable and the
ualgebra f to be independent (i.e., the random variables.; and Ia(w) are independent for every Bet) is that E(g(e)lf) = Eg(e) for every Borel function g(x) with
Elg<e)l < oo.
e
JA edP, is ufinite.
be a nonnegative random variable and < a ualgebra, < s;; !F. Show that
E(el<) < oo (a.s.) if and only if the measure Q, defined on sets A e <by Q(A) =
7. Let
8. Random Variables. II
1. In the first chapter we introduced characteristics of simple random
variables, such as the variance, covariance, and correlation coefficient. These
extend similarly to the general case. Let (Q, !F, P) be a probability space and
~ = ~(w) a random variable for which E~ is defined.
The variance of ~ is
fo
The number q = +
is the standard deviation.
If~ is a random variable with a Gaussian (normal) density
fi( ) _
~
x 
1
foq
e
the parameters m and
[(xm)2]/2a2
> 0,
 oo < m < oo,
(1)
in (1) are very simple:
m
E~,
Hence the probability distribution ofthis random variable~. which we call
Gaussian, or normally distributed, is completely determined by its mean
value m and variance q 2 , (It is often convenient to write~ "' .;V (m, q 2 ).)
Now let(~, 17) be a pair of random variables. Their covariance is
(2)
(assuming that the expectations are defined).
lfcov(~, 17) = 0 we say that~ and '1 are uncorrelated.
If V ~ > 0 and V17 > 0, the number
(3)
is the correlation coefficient of ~ and '7
The properties of variance, covariance, and correlation coefficient were
investigated in 4 of Chapter I for simple random variables. In the general
case these properties can be stated in a completely analogous way.
233
8. Random Variables. II
Let e = <el, ... ' en) be a random vector whose components have finite
second moments. The covariance matrix of e is then x n matrix~= IIRiJII,
where Ril = cov(e;, ei). It is clear that ~ is symmetric. Moreover, it is nonnegative definite, i.e.
n
L1 Rw~)i ~ o
i,j=
for all A.i e R, i = 1, ... , n, since
The following lemma shows that the converse is also true.
Lemma. A necessary and sufficient condition that an n x n matrix
~ is the
covariance matrix of a vector e = <el, ... ' en) is that the matrix is symmetric
and nonnegative definite, or, equivalently, that there is an n x k matrix A
(1 ~ k ~ n) such that
where T denotes the transpose.
PROOF. We showed above that every covariance matrix is symmetric and
nonnegative definite.
Conversely, let~ be a matrix with these properties. We know from matrix
theory that corresponding to every symmetric nonnegative definite matrix ~
there is an orthogonal matrix (!) (i.e., (!)(!)T = E, the unit matrix) such that
(!)T~(!) =
where
D =
(d0
D,
'
0)
dn
is a diagonal matrix with nonnegative elements di, i = 1, ... , n.
It follows that
~ = (!)D(!)T = ((!)B)(BT(!)T),
where B is the diagonal matrix with elements bi = + .jd;, i = 1, ... , n.
Consequently if we put A = (!)B we have the required representation
~ = AATfor ~.
It is clear that every matrix AAT is symmetric and nonnegative definite.
Consequently we have only to show that ~ is the covariance matrix of some
random vector.
Let tit ti 2 , , tin be a sequence of independent normally distributed
random variables, ..(0, 1). (The existence of such a sequence follows, for
example, from Corollary 1 of Theorem 1, 9, and in principle could easily
234
II. Mathematical Foundations of Probability Theory
be derived from Theorem 2 of 3.) Then the random vector~ = A17 (vectors
are thought of as column vectors) has the required properties. In fact,
E~~T
= E(A1'/)(A1'/)T =A. E,,T. AT= AEAT = AAT.
(If ( = li(iill is a matrix whose elements are random variables, E( means the
matrix IIE~iill).
This completes the proof of the lemma.
We now turn our attention to the twodimensional Gaussian (normal)
density
(4)
characterized by the five parameters m1 , m2 , a 1 , a 2 and p (cf. (3.14)), where
Im1 I < oo, Im2 1 < oo, a 1 > 0, a 2 > 0, IpI < 1. An easy calculation identifies
these parameters :
m1 =
E~,
m2 = E1],
ai
= V~,
a~ = V1],
p = p(~, 1]).
In4ofChapter I we explained that if~ and 17 areuncorrelated(p(~, 17) = 0),
it does not follow that they are independent. However, if the pair (~, 17) is
Gaussian, it does follow that if ~ and 17 are uncorrelated then they are
independent.
In fact, if p = 0 in (4), then
But by (6.55) and (4),
frf._x) =
oo
~~~(x,
oo
y) dy =
foal
e[(xmd21J2at,
Consequently
~~~(x,
y) = f~(x) fiy),
from which it follows that~ and 17 are independent (see the end of Subsection
8 of 6).
235
8. Random Variables. II
2. A striking example of the utility of the concept of conditional expectation
(introduced in 7) is its application to the solution of the following problem
which is connected with estimation theory (cf. Subsection 8 of 4 of Chapter
1).
Let(~, '1) be a pair of random variables such that~ is observable but '1 is
not. We ask how the unobservable component '1 can be "estimated" from
the knowledge of observations of~To state the problem more precisely, we need to define the concept of an
estimator. Let qJ = qJ(x) be a Borel function. We call the random variable
qJ(~) an estimator of11 in terms of~. and E['1  ({J(~)] 2 the (mean square) error
of this estimator. An estimator qJ*(~) is called optimal (in the meansquare
sense) if
L\
=E['1 qJ*(~)Y = infE['1 ({J(~)Y,
qJ
(5)
where inf is taken over all Borel functions qJ = qJ(x).
Theorem 1. Let E11 2 < oo. Then there is an optimal estimator qJ* = qJ*(~)
and qJ*(x) can be taken to be thejimction
qJ*(x) = E('11 ~ = x).
(6)
PRooF. Without loss of generality we may consider only estimators tp(~)
for which EqJ 2 (~) < oo. Then if ({J(~) is such an estimator, and qJ*(~) = E('11 ~),
we have
E['1  ({J(~)] 2 = E[('1  qJ*(~)) + (({J*(~) _ ({J(~))]2
= E['7  qJ*(~)]2 + E[({J*(~)  ({J(~)]2
+ 2E[('1  qJ*(~))(({J*(~)  ({J(~))] ~ E[17  qJ*(~)] 2 ,
since E[qJ*(~)  qJ(~)] 2 ~ 0 and, by the properties of conditional expectations,
E[('1
qJ*(~))(({J*(~) ({J(~))]
= E{E[('1
qJ*(~))(({J*(~) ({J(~)])I~]}
= E{(({J*(~) ({J(~))E('1 qJ*(~)I~)} = 0.
This completes the proof of the theorem.
Remark. It is clear from the proof that the conclusion of the theorem is still
valid when ~ is not merely a random variable but any random element
with values in a measurable space (E, tf). We would then assume that
(/) = qJ(x) is an tS'/81(R)measurable function.
Let us consider the form of qJ*(x) on the hypothesis that(~, '1) is a Gaussian
pair with density given by (4).
236
II. Mathematical Foundations of Probability Theory
From (1), (4) and (7.10) we find that the density J, 1 ~(yJx) of the conditional
probability distribution is given by
f~ I~{y IX)
1
J2rc(1  p2 )u 2
e[(y
m(x))2]/[2a~( 1  p2)]'
(7)
where
(8)
Then by the Corollary of Theorem 3, 7,
(9)
and
V('71~
x)
=E[('7 E('71~ = x)) 1~ = x]
2
J_'XJoo (y
= u~{1 
m(x)) 2f~ 1 ~(yJx) dy
p 2 ).
Notice that the conditional variance V('71 ~
therefore
(10)
x) is independent of x and
(11)
Formulas (9) and (11) were obtained under the assumption that V~ > 0
and V17 > 0. However, if V~ > 0 and V17 = 0 they are still evidently valid.
Hence we have the following result (cf. (1.4.16) and (1.4.17)).
Theorem 2. Let (~, 17) be a Gaussian vector with
estimator of '7 in terms of~ is
V~
> 0. Then the optimal
(12)
and its error is
(13)
Remark. The curve y(x) = E('71 ~ = x) is the curve of regression of '7 on ~
or of '7 with respect to ~. In the Gaussian case E('71 ~ = x) = a + bx and
consequently the regression of '1 and ~ is linear. Hence it is not surprising
that the righthand sides of (12) and (13) agree with the corresponding parts
of (1.4.6) and (1.4.17) for the optimal linear estimator and its error.
237
8. Random Variables. II
Corollary. Let e 1 and e2 be independent Gaussian random variables with mean
zero and unit variance, and
Then E~
and if af
= E11 = 0, V~ = af
+ a~ > 0, then
E(
11
+a~, V11
= bf + bLcov(~, 1'/) = a,b, + a2 b 2 ,
I~)= a 1 b1 + a 2 b2 ~
af +a~
(14)
'
Ll = (a 1 b2  azbd 2
ai +a~
(15)
3. Let us consider the problem of determining the distribution functions of
random variables that are functions of other random variables.
Let~ be a random variable with distribution function F~(x) (and density
Nx), if it exists), let q> = q>(x) be a Borel function and 11 = q>(~). Letting
I Y = ( oo, y), we obtain
F~(y) =
P(ry::::;; y)
= P(q>(~)Eiy) = P(~Eq> 1 (/y)) =
F~(dx),
(16)
"' l(ly)
which expresses the distribution function F~(y) in terms of F~(x) and q>.
For example, if 11 = a~ + b, a > 0, we have
( y b) (y b)
F ~(y) = p ~ ::::;; a = F ~ a .
(17)
If 11 = ~ 2 , it is evident that F ~(y) = 0 for y < 0, while for y 2 0
F /Y) = P(
e ::;
y) = P( Jy ::::;; ~ ::::;; Jy)
= F~(Jy) F~( Jy) + P(~ = Jy).
(18)
We now turn to the problem of determining f~(y).
Let us suppose that the range of~ is a (finite or infinite) open interval
I = (a, b), and that the function q> = q>(x), with domain (a, b), is continuously
differentiable and either strictly increasing or strictly decreasing. We also
suppose that q>'(x) =f. 0, xEI. Let us write h(y) = q> 1(y) and suppose for
definiteness that cp(x) is strictly increasing. Then when y E q>(/),
= P(e ::::;; h(y)) =
h(y)
_ 00Nx) dx.
(19)
238
II. Mathematical Foundations of Probability Theory
By Problem 15 of 6,
h(y)
fy
f  OONx) dx =  oof;(h(z))h'(z) dz
(20)
and therefore
fq(y)
= f~;(h(y))h'(y).
(21)
Similarly, if cp(x) is strictly decreasing,
fq(y)
Nh(y))(( h'(y)).
Hence in either case
J~(y)
For example, if 1J = a~
= Nh(y)) Ih'(y) 1.
+ b, a #
0, we have
yb
h(y) =aand
If~ ~
JV (m, u 2) and 11
fq(y)
= el:,
(22)
fq(y)
1 (yb)
= j;f
!~;a
we find from (22) that
1
exp[{ $uy
ln(y/~)2],
2u
y > 0,
(23)
y :::;; 0,
with M =em.
A probability distribution with the density (23) is said to be lognormal
(logarithmically normal).
If cp = cp(x) is neither strictly increasing nor strictly decreasing, formula
(22) is inapplicable. However, the following generalization suffices for many
applications.
Let cp = cp(x) be defined on the set
1 [ak, bk], continuously differentiable and either strictly increasing or strictly decreasing on each open
interval Ik = (ab bk), and with cp'(x) # 0 for x E Ik. Let hk = hk(y) be the
inverse of cp(x) for x Elk. Then we have the following generalization of(22):
Lk=
fq(y) =
L f~;(hk(y))lh~(y)l Ivk(y),
(24)
k=1
where Dk is the domain of hk(y).
For example, if 11 = ~ 2 we can take I 1 = (  oo, 0), I 2
find that h1 (y) =
h 2(y) =
and therefcre
.JY,
.JY,
f~(y) = {2Jy [f~;(.jY) + flJY)],
0,
y > 0,
y :::;; 0.
= (0,
oo ), and
(25)
239
8. Random Variables. II
Wecanobservethatthisresultalsofollowsfrom(18),sinceP(~ =
In particular, if~ ,..., .;V (0, 1),
h2(y) =
{k
JY) = 0.
y > 0,
ey/2,
0,
(26)
0.
A straightforward calculation shows that
fi~l(y)
f +v'l~l (y) 
{f~(y) + f~( y), y > 0,
0,
y::;; 0.
{2y(f~(y2)
0,
+ f~( y2)),
y > 0,
y ~ 0.
(27)
(28)
4. We now consider functions of several random variables.
If ~ and '7 are random variables with joint distribution F ~,(x, y), and
<p = cp(x, y) is a Borel function, then if we put' = cp(~, 17) we see at once that
F~:,(z) =
{x, y: q>(x, y) :S z}
dF~,(x, y).
(29)
For example, if cp(x, y) = x + y, and~ and '7 are independent (and therefore F~,(x, y) = F~(x) F,(y)) then Fubini's theorem shows that
F,(z) =
{x,y:x+y:Sz)
JR2
dF~(x) dF,(y)
I{x+y:s;zJ(x, y) dF~(x) dF,(y)
J:oo dF~(x){J:,/{x+y:s;z)(x, y) dF,(y)} =
J:ooF,(z x)
dF~(x)
(30)
and similarly
F,(z) =
J:ooF~(z y) dF,(y).
(31)
IfF and G are distribution functions, the function
H(z) = J:00 F(z x) dG(x)
is denoted by F * G and called the convolution ofF and G.
Thus the distribution function F 1 of the sum of two independent random
variables ~ and '7 is the convolution of their distribution functions F ~ and F,:
F, = F~*F,.
It is clear that F~ * F, = F, * F~.
II. Mathematical Foundations of Probability Theory
240
Now suppose that the independent random variables ~ and '1 have
densities f~ and f~. Then we find from (31 ), with another application ofFubini's
theorem, that
F~(z) = J:oo [f~Yf~(u) du ]f~(y) dy
= s:oo
[f
ooh(u  y) du J.,(y) dy
whence
J{(z)
f_
00
/iz 
00
00
[f_0000j~(u 
y)J,(y) dy du,
(32)
y)J.,(y) dy,
and similarly
(33)
Let us see some examples of the use of these formulas.
Let ~to ~ 2 , , ~n be a sequence of independent identically distributed
random variables with the uniform density on [ 1, 1]:
f(x) =
{!.
0,
lxl ~ 1,
lxl > 1.
Then by (32) we have
f~, +~ 2 (x)
={
2
_...,...:.....:.,
I 2
IX~'
0,
lxl > 2,
lxl
4
(3  lxl) 2
,
16
!~, +~2+~3(x)
= 3  x2
1~
lxl
3,
0 ~ lxl ~ 1,
lxl > 3,
0,
and by induction
[(n+x)/21
_ { 2"( _ 1) 1
n
/~, + ... +~n(x)
0,
Now let~ ,..., .% (m 1,
(1)kC~(n
on and '1 ,..., .% (m
cp(x)
+ x 2k)"t, lxl
~ n,
k=O
lxl > n.
2,
o~). If we write
= _1_ ex2J2,
fo
241
8. Random Variables. II
then
and the formula
follows easily from (32).
Therefore the sum of two independent Gaussian random variables is again a
Gaussian random variable with mean m1
en
+ m2 and variance uf + u~.
Let 1, ,, be independent random variables each of which is normally
distributed with mean 0 and variance 1. Then it follows easily from (26) (by
induction)
t::+ . . ~
+"(x)
{=2n:;;;t2=:,.,..(n....,./2,.,..) x<nf2)lexf2,
0,
ef
x > 0,
X~
x;,
e;
(34)
0.
The variable
+ + is usually denoted by and its distribution
(with density (30)) is the x2 distribution ("chisquare distribution") with n
degrees of freedom (cf. Table 2 in 3).
If we write Xn =
+.JX!, it follows from (28) and (34) that
fx . (x) =
2x"lex2f2
{ 2"12r(n/2) , X ~ 0,
(35)
X< 0.
0,
The probability distribution with this density is the xdistribution (chidistribution) with n degrees of freedom.
Again let ~ and 11 be independent random variables with densities f~ and
J,. Then
F~~(z) =
JJ
f~(x)f~(y) dx dy,
{x,y:xy:Sz)
F~1 ~(z) =
JJ
f~x)j,(y) dx dy.
{x, y: x/y:Sz)
Hence we easily obtain
~~~(z)
oo (z)
dy
dx
foof~
Y J.,(y) IYT
= foooof~ (z)
X J~(x) TXT
(36)
and
(37)
242
II. Mathematical Foundations of Probability Theory
e;)!n,
en
Putting =eo and '1 = j(e~ + +
in (37), where eo. 1 , .. ,
are independent Gaussian random variables with mean 0 and variance
rJ 2 > 0, and using (35), we find that
f~o/[.j(l/nH~t+ ... +~~)](x)
I r(n; 1)
C
y'Ttn
( )
r~
1
(
(38)
x2)(n+ 1)/2
1+
The variable eo/[)(1/n)(ei + ... + e:)J is denoted by t, and its distribution
is the tdistribution, or Student's distribution, with n degrees of freedom (cf.
Table 2 in 3). Observe that this distribution is independent of rJ.
5.
PROBLEMS
1. Verify formulas (9), (10), (24), (27), (28), and (34)(38).
2. Let ~t> , ~ n ~ 2, be independent identically distributed random variables with
distribution function F(x) (and density f(x), if it exists), and let~= max(~t ... , ~.),
~ = min(~t ... , ~.), p = ~  ~Show that
F (y x) = {(F(y))" (F(y) F(x))",
~.~ '
(F(y))",
y > x,
y ~ x,
n(n 1)[F(y) F(x)]" 2 f(x)f(y),
f~.~(y, x) = { 0,
y > x,
<X,
 {n s~oo [F(y) F(y x)]"tf(y) dy, X~ 0,
~000
'
X< '
n(n 1) s~oo [F(y) F(y x)]" 2 f(y x)f(y) dy,
fP(x) = { 0,
X>
0,
X< 0
3. Let ~t and ~ 2 be independent Poisson random variables with respective parameters
At and A2 Show that ~t + ~ 2 has a Poisson distribution with parameter At + A2
4. Let mt = m2 = 0 in (4). Show that
5. The maximal correlation coefficient of ~ and Yf is p*(~, rt) = sup"" p(u<e), v<e)),
where the supremum is taken over the Borel functions u = u(x) and v = v(x) for
which the correlation coefficient p(u(~). v(e)) is defined. Show that' and Yf are independent if and only if p*(~. rt) = 0.
6. Let "t 1 , "t 2 , , "t be independent nonnegative identically distributed random variables with the exponential density
f(t) = AeAt,
0.
9. Construction of a Process with Given FiniteDimensional Distribution
Show that the distribution of r 1
243
+ + rk has the density
A_ktk1eA'
(k1)!'
0, 1 :o;; k :o;; n,
and that
P(r 1
+ + rk >
t) =
k1
(A.tY
i=O
l.
L eA.1 .
7. Let ~ ~ %(0, a 2 ). Show that, for every p ~ 1,
EI~ IP =
CPaP,
where
and f(s)
So exxs
dx is the gamma function. In particular, for each integer n ~ 1,
E~ 2 " = (2n  1)!! a 2 ".
9. Construction of a Process with Given
FiniteDimensional Distribution
1. Let
~ = ~(w)
(Q, $', P), and let
be a random variable defined on the probability space
F~(x) =
P{w:
~(w)::;
x}
be its distribution function. It is clear that F ~(x) is a distribution function
on the real line in the sense of Definition 1 of 3.
We now ask the following question. Let F = F(x) be a distribution function on R. Does there exist a random variable whose distribution function is
F(x)?
One reason for asking this question is as follows. Many statements in
probability theory begin, "Let~ be a random variable with the distribution
function F(x); then ... ". Consequently if a statement of this kind is to be
meaningful we need to be certain that the object under consideration actually
exists. Since to know a random variable we first have to know its domain
(Q, g;), and in order to speak of its distribution we need to have a probability
measure P on (Q, g;), a correct way of phrasing the question of the existence
of a random variable with a given distribution function F(x) is this:
Do there exist a probability space (Q, $', P) and a random variable~
on it, such that
P{w:
~(w)::;
x}
F(x)?
= ~(w)
244
II. Mathematical Foundations of Probability Theory
Let us show that the answer is positive, and essentially contained in
Theorem 1 of 1.
In fact, let us put
:F = PJ(R).
!l=R,
It follows from Theorem 1 of 1 that there is a probability measure P (and
only one) on (R, PJ(R)) for which P(a, b)] = F(b)  F(a), a < b.
Put c;(w) = w. Then
P{w: c;(w)
~ x}
= P{w: w
~ x}
= P( oo, x] = F(x).
Consequently we have constructed the required probability space and the
random variable on it.
2. Let us now ask a similar question for random processes.
Let X= (c;,),er be a random process (in the sense of Definition 3, 5)
defined on the probability space (Q, !F, P), with t E T R.
From a physical point of view, the most fundamental characteristic of a
random process is the set {F,,, ... ,1"(x 1 , ... , xn)} of its finitedimensional
distribution functions
defined for all sets tt. ... , tn with t 1 < t 2 < < tn.
We see from (1) that, for each set t 1 , .. , tn with t 1 < t 2 < < tn the
functions F 1,, ... ,1Jx 1, ... , xn) are ndimensional distribution functions (in
the sense of Definition 2, 3) and that the collection {F,,, ... ,1"(x 1, ... , Xn)}
has the following consistency property:
lim F,,, ... ,tn(xl, ... ' Xn)
F,,, ... ,tk, ... ,tn(xl, ... '
xk> ... ' Xn)
(2)
Xkf 00
where ~ indicates an omitted coordinate.
Now it is natural to ask the following question: under what conditions
can a given family {F 11 , ... , 1"(xt. ... , xn)} of distribution functions
F,,, ... ,1"(x 1 , . , xn) (in the sense of Definition 2, 3) be the family of finitedimensional distribution functions of a random process? It is quite remarkable that all such conditions are covered by the consistency condition (2).
Theorem 1 (Kolmogorov's Theorem on the Existence of a Process). Let
{F1,,. .. , 1"(x 1, ... , Xn)}, with t; E T R, t 1 < t 2 < < tn, n ~ 1, be a given
family of finitedimensional distribution functions, satisfying the consistency
condition (2). Then there are a probability space (Q, !F, P) and a random
process X = (c;,),er such that
(3)
9. Construction of a Process with Given FiniteDimensional Distribution
245
PRooF. Put
i.e. taken to be the space of real functions w = (w,),eT with the aalgebra
generated by the cylindrical sets.
LetT= [t 1 , ... , tn], t 1 < t 2 < < tn. Then by Theorem 2 of 3 we can
construct on the space (R", PA(R")) a unique probability measure Pr such that
It follows from the consistency condition (2) that the family {P r} is also
consistent (see (3.20)). According to Theorem 4 of 3 there is a probability
measure P on (RT, PA(RT)) such that
P{w: (w11 ,
w,J E B} = Pr(B)
for every set r = [t 1 , .. , tn], t 1 < < tn.
From this, it also follows that (4) is satisfied. Therefore the required
random process X= (~,(w)),eT can be taken to be the process defined by
tE T.
(5)
This completes the proof of the theorem.
Remark 1. The probability space (RT, PA(RT), P) that we have constructed
is called canonical, and the construction given by (5) is called the coordinate
method of constructing the process.
Remark 2. Let (Ea, ga) be complete separable metric spaces, where oc belongs
to some set mof indices. Let {Pr} be a set of consistent finitedimensional
distribution functions P" T = [IX 1, . , ocnJ on
(1% 1
X X
EIXn' rff/% 1 rff/XJ
Then there are a probability space (Q, !F, P) and a family of !F /rff ameasurable
functions (Xa(w))ae!ll such that
P{(Xa 1,
XaJ E B} = P,(B)
for all r = [oc 1 , . , ocnJ and BE ~fa Can
This result, which generalizes Theorem 1, follows from Theorem 4 of 3
if we put n =
Ea., !F = IIa tel% and XIX(w) = WI% for each Q) = w(wl%), 0( em.
Tia
Corollary 1. Let F 1(x), F 2 (x), ... be a sequence of onedimensional distribution
junctions. Then there exist a probability space (Q, ff, P) and a sequence of
independent random variables ~ 1 , ~ 2 , such that
P{w: ~;(w) ::; x} = F;(x).
(6)
246
II. Mathematical Foundations of Probability Theory
In particular, there is a probability space (Q, ~ P) on which an infinite
sequence of Bernoulli random variables is defined (in this connection see
Subsection 2 of 5 of Chapter 1). Notice that Q can be taken to be the space
Q = {ro: w = (a 1, a2 ,
),
ai = 0, 1}
(cf. also Theorem 2).
To establish the corollary it is enough to put F 1 , ... ,ix1 ,
F 1(x 1 ) Fn(Xn) and apply Theorem 1.
Xn) =
Corollary 2. Let T = [0, oo) and let {p(s, x; t, B} be a family of nonnegative
functions defined for s, t E T, t > s, x E R, BE Bl(R), and satisfying the following
conditions:
(a) p(s, x; t, B) is a propability measure on B for given s, x and t;
(b) for given s, t and B, the function p(s, x; t, B) is a Borel function of x;
(c) for 0 ::S; s < t < rand BE Bl(R), the KolmogorovChapman equation
p(s, x; r, B)
{p(s, x; t, dy)p(t, y; r, B)
(7)
is satisfied.
Also let 1t = n(B) be a probability measure on (R, Bl(R)). Then there are
a probability space (Q, ~ P) and a random process X = (~ 1)1 2: 0 defined on
it, such that
(8)
for 0 = t 0 < t 1 < < tn.
The process X so constructed is a Markov process with initial distribution
nand transition probabilities {p(s, x; t, B}.
Corollary 3. Let T = {0, 1, 2, ... } and let {Pk(x; B)} be a family of nonnegative functions defined for k ;;::: 1, x E R, BE PJ(R), and such that pk(x; B)
is a probability measure on B (for given k and x) and measurable in x (for
given k and B). In addition, let n = n(B) be a probability measure on (R, Bl(R)).
Then there is a probability space (Q, ~ P) with a family of random variables X= {~ 0 , ~ 1 , . } defined on it, such that
247
9. Construction of a Process with Given FiniteDimensional Distribution
3. In the situation of Corollary 1, there is a sequence of independent random
variables ~ 1 , ~ 2 , . . . whose onedimensional distribution functions are
F 1 , F 2 , , respectively.
Now let ( 1 , 8' 1 ), ( 2 , 8' 2 ), .. be complete separable metric spaces and
let P 1 , P 2 , ... be probability measures on them. Then it follows from Remark
2 that there are a probability space (Q, ff', P) and a sequence of independent
elements X1o X 2 , .. such that Xn is ~/c8'nmeasurable and P(XnEB) =
PiB), BE @"n
It turns out that this result remains valid when the spaces (En,@" n) are
arbitrary measurable spaces.
Theorem 2 (lonescu Tulcea's Theorem on Extending a Measure and the
Existence of a Random Sequence). Let (Qn, ff,), n = 1, 2, ... , be arbitrary
PI
measurable spaces and n =
nn, ~ = ff,. Suppose that a probability
measure P1 is given on (!1 1, ~1 ) and that, for every set (wt> ... , wn) E !1 1 x
... X nn,n ~ l,probabilitymeasuresP(wl, ... ' wn; )aregivenon(Qn+ I~+ 1).
Suppose that for every BE ff,+ 1 the functions P(wl> ... , wn; B) are Borel
fimctions on (w 1 , , wn) and let
1.
(9)
Then there is a unique probability measure P on (Q, ~) such that
for every n
such that
1, and there is a random sequence X= (X 1 (w), X 2 (w), ...)
where A; E c8';.
PROOF. The first step is to establish that for each n > 1 the set function Pn
defined by (9) on the rectangle A 1 x x An can be extended to the
aalgebra ~1 ff,.
For each n ~ 2 and BE ~1 ff, we put
nn
I B(fl!1, ... , Wn)P(w1, ... , Wn 1; dwn).
(12)
It is easily seen that when B = A 1 x x An the righthand side of (12)
is the same as the righthand side of (9). Moreover, when n = 2 it can be
248
II. Mathematical Foundations of Probability Theory
shown, just as in Theorem 8 of 6, that P 2 is a measure. Consequently it is
easily established by induction that Pn is a measure for all n ~ 2.
The next step is the same as in Kolmogorov's theorem on the extension of
a measure in (R 00 , 14(R 00 ) ) (Theorem 3, 3). Thus for every cylindrical set
Jn(B) = {roe 0: (ro 1, ... , ron) e B}, Be '1 !F,, we define the set
function P by
(13)
If we use (12) and the fact that P(ro 1, ... , rok; )are measures, it is easy to
establish that the definition (13) is consistent, in the sense that the value of
P(Jn(B)) is independent of the representation of the cylindrical set.
It follows that the set function P defined in (13) for cylindrical sets, and in
an obvious way on the algebra that contains all the cylindrical sets, is a
finitely additive measure on this algebra. It remains to verify its countable
additivity and apply Caratheodory's theorem.
In Theorem 3 of 3 the corresponding verification was based on the
property of (R", 14(Rn)) that for every Borel set B there is a compact set
A B whose probability measure is arbitrarily close to the measure of B.
In the present case this part of the proof needs to be modified in the following
way.
As in Theorem 3 of 3, let {Bn} n~ 1 be a sequence of cylindrical sets
Bn = {ro: (rol> ... ' ron) E Bn},
that decrease to the empty set
0. but have
lim P(Bn) > 0.
(14)
n+oo
For n > 1, we have from (12)
where
Since .Bn+1 Bn, we have Bn+1 Bn
on+1 and therefore
JBn+l(ro1, ... ' ron+1) ~ [Bn(ro,, ... ' ron)Inn+l(ron+1>
Hence the sequence {f~1 >(ro 1 )}n~ 1 decreases. Let J(l>(ro 1) = limn f~1 >(ro 1 ).
By the dominated convergence theorem
lim P(Bn) =lim
n
i f~ >(ro 1 )P 1(dro 1 ) i
n,
n,
J< 1>(ro 1)P 1(dro 1).
By hypothesis, limn P(Bn) > 0. It follows that there is an ro~ e B such that
> 0, since if ro 1 rl B 1 then f~1 >(ro 1 ) = 0 for n ~ 1.
j< 1 >(ro~)
9. Construction of a Process with Given FiniteDimensional Distribution
249
Moreover, for n > 2,
(15)
where
f~2 >(w 2 ) =
LP(w~,
l IB"(w~,w2,
w 2 ; dw 3 )
On
... ,wn)P(w~,w 2 ,
,wnbdwn).
We can establish, as for {f~l)(w 1 )}, that {f~2 >(w 2 )} is decreasing. Let
j< 2 >(w 2 ) = limn .... oo Jf>(w 2 ). Then it follows from (15) that
0<
J(l>(w~) = j j< 2 >(w 2 )P(w~; dw 2 ),
Jn2
and there is a point w~ E Q 2 such that f(2)(w~) > 0. Then (w~, w~) E B 2.
Continuing this process, we find a point (w~, ... , w~) E Bn for each n.
0
Consequently (w 01, ... , wn,
.. .) E n~
Bn, but by hypothesis we have n~
Bn = 0.
This contradiction shows that limn P(Bn) = 0.
Thus we have proved the part of the theorem about the existence of the
probability measure P. The other part follows from this by putting XnCw)
= Wn, n;;::: 1.
Corollary 1. Let (En, Cn)n<: 1 be any measurable spaces and (Pn)n<: 1 , measures
on them. Then there are a probability space (Q, ', P) and a family of independent random elements X 1,X 2, ... with values in (E 1,C 1), (E 2 ,C 2 ), .. ,
respectively, such that
P{w: Xiw)
B}
= Pn(B),
Corollary 2. Let E = {1, 2, ... }, and let {pk(x, y)} be a family of nonnegative
functions, k;;::: 1, x,yEE, such that LyeEPk(x;y) = 1, xEE, k;;::: 1. Also
letn = n(x)beaprobabilitydistributiononE(thatis,n(x);;::: O,LxeE n(x) = 1).
Then there are a probability space (Q, ', P) and a family X = { ~ 0 , ~ 1 ,
of random variables on it, such that
Pgo = Xo, ~1 = X1, , ~n = Xn} = n(xo)P1(xo, X1) Pn(Xn1 Xn)
(16)
(cf. (1.12.4)) for all x; E E and n ;;::: 1. We may take Q to be tl".e space
Q
= {w: w = (x 0 , x 1,
. ), X; E
E}.
A sequence X= g 0 , ~ 1 , ... } of random variables satisfying (16) is a
Markov chain with a countable set E of states, transition matrix {pk(x, y)}
and initial probability distribution n. (Cf. the definition in 12 of Chapter 1.)
250
4.
II. Mathematical Foundations of Probability Theory
PROBLEMS
1. Let 0 = [0, 1], let :F be the class of Borel subsets of [0, 1], and let P be Lebesgue
measure on [0, 1]. Show that the space (0, !F, P) is universal in the following sense.
For every distribution function F(x) on (0, !F, P) there is a random variable = ro)
such that its distribution function F ~x) = P(e ~ x) coincides with F(x). (Hint.
e(ro) = r 1(ro), 0 < ro < 1, where r 1(ro) = sup{x: F(x) < ro}, when 0 < ro < 1,
and e(O), W) can be chosen arbitrarily.)
e e(
2. Verify the consistency of the families of distributions in the corollaries to Theorems
1 and 2.
3. Deduce Corollary 2, Theorem 2, from Theorem 1.
10. Various Kinds of Convergence of Sequences
of Random Variables
1. Just as in analysis, in probability theory we need to use various kinds of
convergence of random variables. Four of these are particularly important:
in probability, with probability one, in mean of order p, in distribution.
First some definitions. Let~. ~ 1 , ~ 2 , be random variables defined on a
probability space (Q, !F, P).
Definition 1. The sequence
1,
~ 2 , .
of random variables converges in
probability to the random variable~ (notation: ~n E.~) if for every s > 0
P{l~n ~~
> s}+ 0,
n+
oo.
(1)
We have already encountered this convergence in connection with the
law of large numbers for a Bernoulli scheme, which stated that
n+
oo
(see 5 of Chapter 1). In analysis this is known as convergence in measure.
Definition 2. The sequence ~ 1 , ~ 2 , of random variables converges with
probability one (almost surely, almost everywhere) to the random variable
~if
P{w: ~n
+ ~} =
0,
(2)
i.e. if the set of sample points w for which ~n(w) does not converge to ~has
probability zero.
This convergence is denoted by ~" + ~ (Pa.s.), or ~n ~ ~ or ~n ~ ~.
10. Various Kinds of Convergence of Sequences of Random Variables
251
Definition 3. The sequence ~ 1 , ~ 2 , of random variables converges in
mean of order p, 0 < p < oo, to the random variable~ if
n+ oo.
(3)
~.
In analysis this is known as convergence in LP, and denoted by ~n
In the special case p = 2 it is called mean square convergence and denoted by
~ = l.i.m. ~n (for "limit in the mean").
Definition 4. The sequence ~ 1 , ~ 2 , . . . of random variables converges in
distribution to the random variable ~ (notation: ~n ~ ~) if
n+ oo,
(4)
for every bounded continuous function f = f(x). The reason for the
terminology is that, according to what will be proved in Chapter III, 1,
condition (4) is equivalent to the convergence of the distribution F~"(x) to
F~(x) at each point x of continuity ofF ~(x). This convergence is denoted by
F~" => F~.
We emphasize that the convergence of random variables in distribution
is defined only in terms of the convergence of their distribution functions.
Therefore it makes sense to discuss this mode of convergence even when the
random variables are defined on different probability spaces. This convergence will be studied in detail in Chapter III, where, in particular, we
shall explain why in the definition ofF~" => F ~ we require only convergence
at points of continuity of F~(x) and not at all x.
2. In solving problems of analysis on the convergence (in one sense or
another) of a given sequence of functions, it is useful to have the concept of a
fundamental sequence (or Cauchy sequence). We can introduce a similar
concept for each of the first three kinds of convergence of a sequence of
random variables.
Let us say that a sequence {~n}n~ 1 of random variables is fundamental in
probability, or with probability 1, or in mean of order, p, 0 < p < oo, if the
corresponding one of the following properties is satisfied: P{ I~n  ~I > e}
+ 0, as m, n+ oo for every e > 0; the sequence gn(w)}n~ 1 is fundamental
for almost all WE Q; the sequence {~n( W)} n ~ 1 is fundamental in U, i.e.
EI~n  ~m IP + 0 as n, m+ 00.
3. Theorem 1.
(a) A necessary and sufficient condition that
~n
P{~~~~~k ~I~ e}+ 0,
for every e > 0.
+
(Pa.s.) is that
n+ oo.
(5)
252
II. Mathematical Foundations of Probability Theory
(b) The sequence {~"} "~ 1 is fundamental with probability 1 if and only
n+ oo,
if
(6)
for every e > 0; or equivalently
~n+k  ~n I 2
P{sup I
k~O
PROOF.
n __.. w.
e}+ 0,
=n:;,
(a) Let A~= {w: l~n ~~ 2 e}, A'= IlniA~
+ 0 = U A' = U A
(7)
Uk~n A;;. Then
00
{w: ~n
11 m.
m=1
e~O
But
P(A') = lim
n
P( U A;;),
k~n
Hence (a) follows from the following chain of implications:
~" + ~} =
0 = P{w:
P(U
e>O
A')> P( UA1m)= 0
> P(A 1 1m) = 0,
m=1
m 2 1 > P(A') = 0, e > 0,
> PCY" A;;)+ 0,
n+
00
>
P(~~~~~k ~I 2 e)+ 0,
n + oo.
(b) Let
n U Bk,.
00
B' =
n= 1
k~n
l~n
Then {w: g"{w)}"~ 1 is notfundamental} = U,~ 0 B', and it can be shown
as in (a) that P{w: {~n(w)}n~ 1 is not fundamental}= 0>(6). The equivalence of (6) and (7) follows from the obvious inequalities
supJ~n+k ~nl:::; supl~n+k ~n+ll:::; 2 supl~n+k ~nl
This completes the proof of the theorem.
Corollary. Since
253
10. Various Kinds of Convergence of Sequences of Random Variables
a sufficient condition for
~n ~ ~
is that
00
L P{i~k ~I~ e} <
k= 1
(8)
00
is satisfied for every e > 0.
It is appropriate to observe at this point that the reasoning used in
obtaining (8) lets us establish the following simple but important result which
is essential in studying properties that are satisfied with probability 1.
Let A 1, A 2 , be a sequence of events in F. Let (see the table in 1)
{An i.o.} denote the event ITiiiAn that consists in the realization of infinitely
many of A 1 , Az, ....
BoreiCantelli Lemma.
(a) If'[. P(An) < 00 then P{An i.o.} = 0.
(b) If'[. P(An) = oo and A1, A 2 , are independent, then P{An i.o.} = 1.
PROOF.
(a) By definition
{An i.o.} = ITiii An=
rl
n= 1
U Ak.
k?!n
Consequently
P{An i.o.} = Pt01 kyn Ak} =lim P(Yn Ak)
and (a) follows.
(b) If A 1, A 2 ,
are independent, so are
we have
Con Ak)
and it is then easy to deduce that
PCQ Ak)
Since log(1  x)
x, 0
00
log
fl
k=n
li
lt
A1, A2 ,
Hence for N ~ n
P(Ak),
(9)
P(Ak).
x < 1,
00
[1  P(Ak)] =
~lim k~nP(Ak),
k=n
00
log[1  P(Ak)] ~ 
Consequently
for all n, and therefore P(An i.o.) = 1.
This completes the proof of the lemma.
k=n
P(Ak) =  oo.
254
II. Mathematical Foundations of Probability Theory
Corollary 1. If A~= {w: 1e,. e1;,:: 6} then (8) shows that L:'= 1 P(A,.) < oo,
6 > 0, and then by the BorelCantelli lemma we have P(A") = 0, 6 > 0, where
A = llm A~. Therefore
L P{lek el;,:: 6} < 00,6 > 0 ~ P(A") =
0, 6 > 0
~
P{w:
e,. +e)}
= 0,
as we already observed above.
Corollary 2. Let (6,.)11 2: 1 be a sequence of positive numbers such that
n ~ oo. If
L P{le,. el;,:: 6,.} < 00,
6 11 ~
0,
00
(10)
n=l
e,.  e
In fact, let A,. = {I
I ;,:: 6,.}. Then P(A,. i.o.) = 0 by the BorelCantelli lemma. This means that, for almost every w e Q, there is an N =
N(w) such that le,.(w) e(w)l ~ 6,. for n;,:: N(w). But 6,. ~ 0, and therefore
e,.(w) ~ e(w) for almost every WE Q.
4. Theorem 2. We have the following implications:
e,. ~ ~ ~ ~~~ f. e.
.;,. g .; = .;,. .f. .;,
(11)
p > 0,
(12)
~~~ f. ~ ~ ~~~ ~ ~.
(13)
PRooF. Statement (11) follows from comparing the definition of convergence
in probability with (5), and (12) follows from Chebyshev's inequality.
To prove (13), let f(x) be a continuous function, let If (x) I ~ c, let 6 > 0,
and let N be such that P(l~l > N) ~ 6/4c. Take b so that lf(x) f(y)l ~
6/2c for lxl < N and lx Yl ~b. Then (cf. the proof of Weierstrass's
theorem in Subsection 5, 5, Chapter I)
N)
+ E(lf(~,.) f(~)l; 1~..  ~I ~ b, 1~1 > N)
+ E(lf<e..) J<~>l; 1e..  e1 >b)
~ 6/2 + 6/2 + 2cP{I~..  ~I> <5}
= 6 + 2cP{I~ ..  ~I> <5}.
Elf(~,.) f(~)l = E(lf(~,.) f(~)l; 1~ ..
~I~
b,
1~1 ~
But P{le,. ~I> <5}.....,. 0, and hence E If(~,.) f(~)l ~ 26 for sufficiently
large n; since 6 > 0 is arbitrary, this establishes (13).
This completes the proof of the theorem.
We now present a number of examples which show, in particular, that the
converses of (11) and (12) are false in general.
255
10. Various Kinds of Convergence of Sequences of Random Variables
1 (~n ~ ~ f:> ~n ~ ~; ~n ~ ~
86([0, 1]), P = Lebesgue measure. Put
EXAMPLE
f:> ~n ~ ~). Let 0
A;=[~~]
n
n 'n '
= [0, 1],
f7 =
i = 1, 2, ... , n; n ;;::: 1.
Then the sequence
{ ).:1. ).:1
).:2. ).:1
).:3.
).:2
'>1 '>2 '>2 '>3 '>3 '>3 ..
of random variables converges both in probability and in mean of order
p > 0, but does not converge at any point wE [0, 1].
2 (~n ~ ~ => ~n ~ ~ =f ~n!:! ~, p > 0). Again let 0
86[0, 1], P = Lebesgue measure, and let
EXAMPLE
~n(w)
{e",
0,
0 :::;;
W :::;;
w > 1/n.
= [0, 1],
f7
1/n,
Then {~n} converges with probability 1 (and therefore in probability) to
zero, but
n+
oo,
for every p > 0.
3 (~n!:! ~ =f ~n ~~).Let {~n} be a sequence of independent random
variables with
ExAMPLE
P(~n
= 0) =
1  Pn
Then it is easy to show that
~n
+ 0 > Pn + 0,
n + oo,
(14)
~n
+ 0 > Pn + 0,
LP
n + oo,
(15)
00
~n ~ 0 =>
L Pn <
(16)
00.
n=1
LP
In particular, if Pn = 1/n then ~n + 0 for every p > 0, but ~n
+0.
a.s.
The following theorem singles out an interesting case when almost sure
convergence implies convergence in L 1
Theorem 3. Let (~n) be a sequence of nonnegative random variables such that
~n ~~and E~n+ E~ < 00. Then
EI~n 
~I + 0,
n+
oo.
(17)
256
II. Mathematical Foundations of Probability Theory
PROOF. We have
we have
E~n
< oo for sufficiently large n, and therefore for such n
E ~~ ~nl = E(~ ~n)/{~2!~nl
+ E(~n ~)J{~n>~}
+ E(~n ~).
2E(~ ~n)/{~2!~n}
But 0 ::;; (~  ~n)J 1 ~.,~"l ::;; ~ Therefore, by the dominated convergence
theorem, limn E(~ ~n)/ 1 ~.,~") = 0, which together with E~n+ E~ proves
(17).
Remark. The dominated convergence theorem also holds when almost sure
convergence is replaced by convergence in probability (see Problem 1).
Hence in Theorem 3 we may replace "~" ~ C' by "~" f. ~"
5. It is shown in analysis that every fundamental sequence (xn), xn E R, is
convergent (Cauchy criterion). Let us give a similar result for the convergence
of a sequence of random variables.
Theorem 4 (Cauchy Criterion for Almost Sure Convergence). A necessary and
sufficient condition for the sequence (~n)n;;, 1 of random variables to converge
with probability 1 (to a random variable~) is that it is fundamental with probability 1.
PROOF. If ~n ~ ~ then
k;;,n
l;;,n
k;::n
l<!n
whence the necessity follows.
Now let C~n)n;;,l be fundamental with probability 1. Let .;V = {w: (~n(w))
is not fundamental}. Then whenever w E Q \ .;V the sequence of numbers
(~"(w))n;;,l is fundamental and, by Cauchy's criterion for sequences of
numbers, lim ~n(w) exists. Let
~(w) ={lim ~n(w), wEil\%,
0,
WE%.
(18)
The function so defined is a random variable, and evidently ~" ~ ~.
This completes the proof.
Before considering the case of convergence in probability, let us establish
the following useful result.
Theorem 5. If the sequence(~") is fundamental (or convergent) in probability,
it contains a subsequence (~"k) that is fundamental (or convergent) with probability 1.
257
10. Various Kinds of Convergence of Sequences of Random Variables
PROOF. Let(~") be fundamental in probability. By Theorem 4, it is enough
to show that it contains a subsequence that converges almost surely.
Take n 1 = 1 and define nk inductively as the smallest n > nkl for which
P{l~, ~.I> rk}
for all s
n, t
< rk.
n. Then
L P{l~nk+l ~nkl > 2k} < L 2k < 00
k
and by the BorelCantelli lemma
P{l~nk+l ~nkl > 2k i.o.} = 0.
Hence
00
L ~~nk+l 
k=l
with probability 1.
Let .At= {m: I~nk+l  ~nkl
~(m) =
1 ~n,(m)
,;
~nkl <
00
oo }. Then if we put
I(~~:~,  ~nk(m)),
t=I
0,
mE 0\.K,
WE
JV,
we obtain ~nk ~ ~
If the original sequence converges in probability, then it is fundamental in
probability (see also (19)), and consequently this case reduces to the one
already considered.
This completes the proof of the theorem.
Theorem 6 (Cauchy Criterion for Convergence in Probability). A necessary
and sufficient condition for a sequence (~n)n<?: 1 ofrandom variables to converge in
probability is that it is fundamental in probability.
PROOF.
If ~n ~ ~ then
P{len ~ml ~ e} ~ P{l~n ~I~ e/2}
+ P{l~m
~I~ e/2}
(19)
and consequently (~n) is fundamental in probability.
Conversely, if(~") is fundamental in probability, by Theorem 5 there are
a subsequence (~nJ and a random variable~ such that ~nk ~~.But then
P{l~n ~I~ e} ~ P{l~n ~nkl ~ e/2}
+ P{l~nk
~I~ e/2},
from which it is clear that ~n ~ ~. This completes the proof.
Before discussing convergence in mean of order p, we make some observations about LP spaces.
258
II. Mathematical Foundations of Probability Theory
We denote by U = U(O., !F, P) the space ofrandom variables~=
with EI~ IP = Jn I~ IP dP < oo. Suppose that p ~ 1 and put
Wlp =
~(ro)
(E I~ IP) 11P.
It is clear that
ll~llp ~
0,
llc~IIP = lciii~IIP'
c constant,
(20)
(21)
and by Minkowski's inequality (6.31)
II~
+ ttllp :s; Wlp + llttllp
(22)
Hence, in accordance with the usual terminology of functional analysis, the
function IIIlP' defined on U and satisfying (20)(22), is (for p ~ 1) a semi
norm.
For it to be a norm, it must also satisfy
ll~llp
= 0 ~ ~ = 0.
(23)
This property is, of course, not satisfied, since according to Property H
(6) we can only say that ~ = 0 almost surely. However, if U means the space
whose elements are not random variables ~ with EI~ IP < oo, but equivalence
classes of random variables (~ is equivalent to q if ~ = q almost surely),
then 1111 becomes a norm, so that U is a normed linear space. If we select
from each equivalence class of random variables a single element, taking the
identically zero function as the representative of the class equivalent to it, we
obtain a space (also denoted by U) wh~ch is actually a normed linear space
of functions (rather than of equivalence classes).
It is a basic result of functional analysis that the spaces U, p ~ 1, are
complete, i.e. that every fundamental sequence has a limit. Let us state and
prove this in probabilistic language.
Theorem 7 (Cauchy Test for Convergence in Mean pth Power). A necessary
and sufficient condition that a sequence (~n)n ~ 1 of random variables in U
convergences in mean of order p to a random variable in LP is that the sequence
is fundamental in mean of order p.
PRooF. The necessity follows from Minkowski's inequality. Let C~n) be
fundamental Cll~n ~miiP+ 0, n, m+ oo). As in the proof of Theorem 5, we
select a subsequence (~nk) such that ~nk ~ ~'where~ is a random variable with
Wlp < oo.
Let n1 = 1 and define nk inductively as the smallest n > nk_ 1 for which
II~, ~.liP<
for all s
n, t
n. Let
2 2 k
259
I 0. Various Kinds of Convergence of Sequences of Random Variables
Then by Chebyshev's inequality
P(A ) <
k
E I;;
'onk+ 1 
2kr
2 2kr
;; I' < __
'onk
 2kr
2kr < 2k.

As in Theorem 5, we deduce that there is a random variable
;;
':,nk
such that
~;;
We now deduce that II~. ~liP+ 0 as n+ oo. To do this, we fix B > 0
and choose N = N(e) so that II~.  ~mil~ < dor all n ~ N, m ~ N. Then for
any fixed n ~ N, by Fatou's lemma,
E
1~. ~IP = E{ lim 1~. ~.kiP}= E{ lim 1~. ~.kiP}
~oo
~oo
Consequently E I~.  ~ IP + 0, n + oo. It is also clear that since~ = ( ~  ~.)
+ ~.we have E I~ IP < oo by Minkowski's inequality.
This completes the proof of the theorem.
Remark 1. In the terminology of functional analysis a complete normed
linear space is called a Banach space. Thus U, p ~ 1, is a Banach space.
Remark 2. If 0 < p < 1, the function WI P = (E I~ IPY 1P does not satisfy the
triangle inequality (22) and consequently is not a norm. Nevertheless the
space (of equivalence classes) LP, 0 < p < 1, is complete in the metric
d(~,IJ) = El~ IJIP.
Remark 3. Let L ro = L'XJ(n, :, P) be the space (of equivalence classes of)
random variables~= ~(w) for which Wloo < oo, where Wloo, the essential
supremum of~. is defined by
6.
=inf{O s c s oo:
Wloo
=ess
supl~l
The function
1111 oo
is a norm, and L oo is complete in this norm.
P(l~l
>c)= 0}.
PROBLEMS
1. Use Theorem 5 to show that almost sure convergence can be replaced by convergence in probability in Theorems 3 and 4 of 6.
2. Prove that L 00 is complete.
3. Show that if~.
4. Let ~. !'. ~.lin
f.
~and also~.
f. 11 then~ and 11 are equivalent (P(~
!'. IJ, and let ~ and 11 be equivalent. Show that
P{l~n
for every e > 0.
llnl ~ ~>}> 0,
n> oo,
# 17) = 0).
260
II. Mathematical Foundations of Probability Theory
5. Let ~. ~ ~, IJn ~ YJ. Show that a~.+ bYJ. ~a~+ bYJ (a, b constants), 1~.1 ~ 1~1,
~.'I.~ ~'1
6.
Let(~.  ~) 2 + 0. Show that~; + ~ 2
7. Show that if~.
bility:
!!. C, where C is a constant, then this sequence converges in proba~. !!. c => ~ .f. c.
8. Let (~.)., 1 have the property that
~. + 0 (Pa.s.).
I:'=
E I~. IP < oo for some p > 0. Show that
9. Let(~.)., 1 be a sequence of independent identically distributed random variables.
Show that
<Xl
El~ 1 1 < oo=
n=1
P{l~ 1 1 > en} < oo
10. Let ( ~.)." 1 be a sequence of random variables. Suppose that there are a random variable~ and a sequence {nd such that~ + ~ (Pa.s.) and max _, <l,;n I~~  ~ _,I+ 0
(Pa.s.) ask + oo. Show that then ~. + ~ (Pa.s.).
11. Let the dmetric on the set of random variables be defined by
d(!' )  E I~  'II
1 +I~ 111
"'' 11 and identify random variables that coincide almost surely. Show that convergence
in probability is equivalent to convergence in the dmetric.
12. Show that there is no metric on the set of random variables such that convergence
in that metric is equivalent to almost sure convergence.
11. The Hilbert Space of Random Variables with
Finite Second Moment
1. An important role among the Banach spaces U, p ~ 1, is played by the
space L 2 = L 2 (0, !F, P), the space of (equivalence classes of) random variables with finite second moments.
If~ and '1 E L 2, we put
(~, 1'/)
It is clear that if~' IJ, (
(a~
=E~l'/
L 2 then
+ bl'f, 0
= a(~,
0+
(~, ~) ~
and
(1)
b(IJ,
0,
a, bER,
261
11. The Hilbert Space of Random Variables with Finite Second Moment
Consequently(~, '1) is a
scalar product. The space L 2 is complete with
respect to the norm
(2)
induced by this scalar product (as was shown in 10). In accordance with the
terminology of functional analysis, a space with the scalar product (1) is a
Hilbert space.
Hilbert space methods are extensively used in probability theory to study
properties that depend only on the first two moments of random variables
(" L 2 theory"). Here we shall introduce the basic concepts and facts that will
be needed for an exposition of L 2 theory (Chapter VI).
2. Two random variables ~ and 11 in L 2 are said to be orthogonal ( ~ l_ 11)
'1) E~11 = 0. According to 8, ~and 11 are uncorrelated ifcov(~, '1) = 0,
i.e. if
if(~,
It follows that the properties of being orthogonal and of being uncorrelated
coincide for random variables with zero mean values.
A set M ~ L 2 is a system of orthogonal random variables if ~ l_ 11 for
every~. 11 EM(~ = 11).
If also II~ I = 1 for every ~EM, then M is an orthonormal system.
3. Let M = {'1t. ... , '1n} be an orthonormal system and~ any random variable in L 2 . Let us find, in the class oflinear estimators L7= 1 a;'1;, the best meansquare estimator for~ (cf. Subsection 2, 8).
A simple computation shows that
El~ ;t1 ~;'1;1 2 = 11~ it1 a;'1f = (~ ;t1 a;'1;, ~ ;t1 a;'1;)
it a;(~, + (t J1
= Wl2
= 11~11 2 
i= 1
Wl 2
where we used the equation
i= 1
n
i= 1
a;'1;)
a;(~,
'1;) + L af
= Wl 2
a;'1;,
'1;)
i= 1
n
1(~, '1;W
1<~. 11;)1 2 ,
+ L
i= 1
Ia; (~, '1;W
(3)
262
II. Mathematical Foundations of Probability Theory
Ii=
It is now clear that the infimum of E I~ 1 a;1Jd 2 over all real
at> ... , an is attained for a; = (~, IJ;), i = 1, ... , n.
Consequently the best (in the meansquare sense) estimator for~ in terms
of 1] 1, . , IJn is
n
~=
i= 1
c~, IJ;)'li
Here
~ = infEI~ it1 a;IJ;'2 = El~ ~12 = Wlz
(4)
J1 1(~,1];)12
(5)
(compare (1.4.17) and (8.13)).
Inequality (3) also implies Bessel's inequality: if M = {1] 1, 1] 2, ... }
an orthonormal system and ~ E L 2 , then
00
i= 1
1(~,
'1;)1 2
IS
Wl 2 ;
(6)
(~, IJ;)}I;.
(7)
and equality is attained if and only if
n
~ = l.i.m.
n
i= 1
The best linear estimator of~ is often denoted byE(~ I1] 1, . , IJn) and called
the conditional expectation (of~ with respect to 1] 1, .. , IJn) in the wide sense.
The reason for the terminology is as follows. If we consider all estimators
cp = cp(1J 1, .. , 1'/n) of ( in terms of 17 t> ... , 1'/n (where cp is a Borel function),
the best estimator will be cp* = E(( Ii'J 1, , 1'/n), i.e. the conditional expectation
of ( with respect to 1] 1 , , 1'/n (cf. Theorem 1, 8). Hence the best linear
estimator is, by analogy, denoted by E(~l1'f 1 , . . . , 1'/n) and called the conditional expectation in the wide sense. We note that if 'lt> ... , 1'/n form a
Gaussian system (see 13 below), then E((I1J 1 , ... ,1Jn) and E((I1J 1 , ... ,1Jn)
are the same.
Let us discuss the geometric meaning of~ = E(~ I'lll ... , IJn).
Let !E = !E{1] 1, . , IJn} denote the linear manifold spanned by the orthonormal system of random variables 1] 1 , ... , 'ln (i.e., the set of random variables of the form Lf= 1 a;IJ;, a; E R).
Then it follows from the preceding discussion that ~ admits the "orthogonal decomposition"
(8)
where ~ E !E and ~  ~ .l !E in the sense that ~  ~ .ill for every ll E !E.
It is natural to call ~ the projection of ( on !E (the element of !t' "closest"
to ~),and to say that~  ~is perpendicular to !E.
4. The concept of orthonormality of the random variables '1~> ... , 'ln makes it
easy to find the best linear estimator (the projection) ~ of ~ in terms of
11. The Hilbert Space of Random Variables with Finite Second Moment
263
'7 1, .. , 'ln The situation becomes complicated if we give up the hypothesis of
orthonormality. However, the case of arbitrary '7 1, ... , 'ln can in a certain
sense be reduced to the case of orthonormal random variables, as will be
shown below. We shall suppose for the sake of simplicity that all our random
variables have zero mean values.
We shall say that the random variables 'It> ... , 'In are linearly independent
if the equation
n
L a;'l; =
0 (Pa.s.)
i= 1
is satisfied only when all a; are zero.
Consider the covariance matrix
~ =E'7'7T
of the vector '7 = ('7 1, ... , 'In). It is symmetric and nonnegative definite,
and as noticed in 8, can be diagonalized by an orthogonal matrix l!J:
l!JT~l!J =
D,
where
0)
D=(d1 ..
0 dn
has nonnegative elements d;, the eigenvalues of ~' i.e. the zeros A. of the
characteristic equation det(~  A.E) = 0.
If '1 b ... , 'ln are linearly independent, the Gram determinant (det ~) is
not zero and therefore d; > 0. Let
and
(Ft.Jd,.
f3
= Bll!JT'l
0 )
(9)
Then the covariance matrix of f3 is
Ef3f3T = B1l!.JTE'7'7Tl!.JB1 =
and therefore f3 = (/3 1 ,
It is also clear that
B1l!JT~l!JB1
= E,
f3n) consists of uncorrelated random variables.
'I= (l!.JB)/3.
(10)
Consequently if '7 1, .. , 'ln are linearly independent there is an orthonormal
system such that (9) and (10) hold. Here
.!l'{'lb .. ' '7n} = .!l'{/31, "' /3n}.
This method of constructing an orthonormal system f3 to , Pn is frequently inconvenient. The reason is that if we think of '7; as the value of the
random sequence ('7 1 , .. , 'ln) at the instant i, the value /3; constructed above
264
II. Mathematical Foundations of Probability Theory
depends not only on the "past," (Yf 1 , , Y/;), but also on the "future,"
(Y/;+ t> , Yfn). The GramSchmidt orthogonalization process, described
below, does not have this defect, and moreover has the advantage that it can
be applied to an infinite sequence of linearly independent random variables
(i.e. to a sequence in which every finite set of the variables are linearly
independent).
Let Yft> Yf 2, .. be a sequence of linearly independent random variables in
L 2 We construct a sequence e1,e 2 , .. as follows. Let e1 = Yft!IIY/ 1 11. If
e1 , . , e"_ 1 have been selected so that they are orthonormal, then
e
Yfn ~n
(11)
IIYfn  ~nil'
where ~" is the projection of Yfn on the linear manifold !l'(e 1 ,
generated by
en_ 1 )
n1
~" =
L (Yfn, ek)ek.
(12)
k=1
Since Yft>Yfn are linearly independent and !l'{Yf 1 , .. ,Yfnd =
!l'{e 1 , . , en 1 }, we have IIYfn ~nil > 0 and consequently en is well defined.
By construction, lien II = 1 for n ~ 1, and it is clear that (en, ek) = 0 for
k < n. Hence the sequence e1, e2 , is orthonormal. Moreover, by (11),
where b" = II Yin  ~nil and ~n is defined by (12).
Now let Yf 1, , Yfn be any set of random variables (not necessarily linearly
llriill is the covariance matrix of
independent). Let det ~ = 0, where ~
(17 1, ... , Yfn), and let
rank~=r<n.
Then, from linear algebra, the quadratic form
n
Q(a) =
L riiaiai,
i,j= 1
has the property that there are n  r linearly independent vectors a(l>, ... ,
a(nr) such that Q(dil) = 0, i = 1, ... , n r.
But
Consequently
n
L a~>'1k =
k= 1
with probability 1.
0,
i = 1, ... , n r,
II. The Hilbert Space of Random Variables with Finite Second Moment
265
In other words, there are n  r linear relations among the variables
'11> ... , '7n Therefore if, for example, 17 1, ... , '7, are linearly independent, the
other variables '7r+ 1, ... , '7n can be expressed linearly in terms of them, and
consequently 2'{171> ... , '7n} = !l'{e 1, ... , e,}. Hence it is clear that we can
find r orthonormal random variables et> ... , e, such that 17 1, ... , '7n can be
expressed linearly in terms ofthem and 2'{17 1, ... , '7n} = !l'{et> ... , e,}.
5. Let '7 1, 17 2 , be a sequence of random variables in L 2 Let .!l =
.!l{'7t> 17 2 , } be the linear manifold spanned by 17 1, 17 2 , . ,i.e. the set of
random variables of the form
1 a; I'/;. n ;:::: 1, a; e R. Then .!l =
.!l{'7t> 17 2 , } denotes the closed linear manifold spanned by 17 1,17 2, ... ,
i.e. the set of random variables in !l' together with their meansquare limits.
We say that a set 17 1 , 17 2 , is a countable orthonormal basis (or a complete
orthonormal system) if:
Li'=
(a) '11> 17 2 , is an orthonormal system,
(b) .!l{'71 '12, .. .} = L 2.
A Hilbert space with a countable orthonormal basis is said to be separable.
By (b), for every~ eL 2 and a given e > 0 there are numbers a 1 , ... , an
such that
Then by (3)
Consequently every element of a separable Hilbert space L 2 can be represented as
00
~ =
L <~. '7;). '7;,
(13)
i= 1
or more precisely as
n
~ = l.i.m.
n
L (~, '7)'7;
i= 1
We infer from this and (3) that Parseval's equation holds:
Wl 2
L 1(~. '7;)1
00
i=1
2,
(14)
It is easy to show that the converse is also valid: if '7 1 , 17 2 , is an orthonormal system and either (13) or (14) is satisfied, then the system is a basis.
We now give some examples of separable Hilbert spaces and their bases.
266
II. Mathematical Foundations of Probability Theory
EXAMPLE 1.
Let
n = R, :F = &I(R), and let P be the Gaussian measure,
P( oo, a]
Let D
f oo
qJ(x) dx,
= d/dx and
H ( ) = ( 1)"D"qJ(x)
"x
qJ(x)
,
n ~ 0.
(15)
We find easily that
= xqJ(x),
DqJ(x)
D 2 qJ(x) = (x 2
D 3 qJ(x)
(16)
1)qJ(x),
(3x  x 3 )qJ(x),
It follows that H,.(x) are polynomials (the Hermite polynomials). From (15)
and (16) we find that
H 0 (x)
= 1,
H 1(x) = x,
Hz(x) = x 2
1,
HJ(x) = x 3
3x,
A simple calculation shows that
(Hm, H,.)
s:ooHm(x)H,.(x) dP
J:ooHm(x)H,.(x)qJ(X) dx
n! c5mn
where c5m,. is the Kronecker delta (0, if m =F n, and 1 if m = n). Hence if we
put
h( )
,.x
= H,.(x)
Jn'
the system of normalized Hermite polynomials {h,.(x)}n>o will be an orthonormal system. We know from functional analysis that if
lim
c!O
foo eclxl P(dx)
 oo
< oo,
(17)
the system {1, x, x 2 , } is complete in L 2 , i.e. every function e = e(x) in L 2
can be represented either as
1 ai'li(x), where 1'/i(x) = xi, or as a limit of
Li=
II. The Hilbert Space of Random Variables with Finite Second Moment
267
these functions (in the meansquare sense). If we apply the GramSchmidt
orthogonalization process to the sequence 111(x), 11ix), ... , with IJ;(x) = xi,
the resulting orthonormal system will be precisely the system of normalized
Hermite polynomials. In the present case, (17) is satisfied. Hence {hn(x)}n;;,o
is a basis and therefore every random variable ~ = ~(x) on this probability
space can be represented in the form
~(x)
= l.i.m.
n
(~, h;)h;(x).
(18)
i=O
2. Let !2 = {0, 1, 2, ... } and let P = {P 1 , P 2 ,
distribution
EXAMPLE
be the Poisson
= 0, 1, ... ; A > 0.
Put fl.J(x) = f(x)  f(x  1) (f(x) = 0, x < 0), and by analogy with (15)
define the PoissonCharlier polynomials
1,
I1 0 = 1.
(19)
Since
00
(ITm, ITn) =
x=O
ITm(x)ITn(x)Px = Cnbmn'
where en are positive constants, the system of normalized PoissonCharlier
polynomials {nn(x)}n;;,o' nn(x) = IInCx)/Jc:, is an orthonormal system, which
is a basis since it satisfies (17).
In this example we describe the Rademacher and Haar systems,
which are of interest in function theory as well as in probability theory.
Let n = [0, 1], ff = ~([0, 1]), and let P be Lebesgue measure. As we
mentioned in 1, every x E [0, 1] has a unique binary expansion
EXAMPLE3.
where x; = 0 or 1. To ensure uniqueness of the expansion, we agree to
consider only expansions containing an infinite number of zeros. Thus we
choose the first of the two expansions
110
011
=+++
.
2 22 23
2 2 2 2 2 3 .. =+++"
We define random variables ~ 1 (x), ~ 2 (x), ... by putting
268
II. Mathematical Foundations of Probability Theory
R 2 (x)
R 1(x)
I
I
I
I
I
I
I
1.!.
1.!.
4
1.!. IJ.
2
I
I
1
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
1
I
I
I
I
I
I
I
I
I
I
I
I
I
I
..._...I ..._..
Figure 30
Then for any numbers a;, equal to 0 or 1,
a1
P{ x. 2
1}
a.
a 1 a2
a.
a2
++ ++< x <+ ++2" 2"
22
2
2" 22
= p { X'. XE [ a21
1
1 ]}
a. +a. a 1 + += .
+ +2"
2" 2"
2"' 2
It follows immediately that~ 1 , ~ 2 , ... form a sequence of independent Bernoulli
random variables (Figure 30 shows the construction of ~ 1 = ~ 1 (x) and
~2
~z(x)).
If we now set R.(x) = 1  2~.(x), n ;;:: 1, it is easily verified that {R.}
(the Rademacher functions, Figure 31) are orthonormal:
ERnRm = fR.(x)Rm(x) dx = Dnm
Notice that (1, R.) ER. = 0. It follows that this system is not complete.
However, the Rademacher system can be used to construct the Haar
system, which also has a simple structure and is both orthonormal and
complete.
~(x)
~(x)
I~
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
r;
I
I
I
I
I
I
I
I
t i
Figure 31. Rademacher functions.
I
I
I
I
I
I
I
269
11. The Hilbert Space of Random Variables with Finite Second Moment
Again let Q = [0, 1) and IF= .16'([0, 1)). Put
H 1 (x)=1,
H 2 (x)
R 1 (x),
k 1
k
v:::;;
x < 2i'
if
n = 2i
+ k,
1 :::;; k :::;; 2i,j;;?: 1,
otherwise.
It is easy to see that H.(x) can also be written in the form
Hzm+l(x) = {
2m/2
0 <X < 2(m+ 1)
0,
2(m+l):::;;
otherwise,
2~ 12 ,
X~ 2m,
m = 1, 2, ... '
Hzm+j(x)=H zm+l(xj;. 1). j= 1, ... ,2m.
Figure 32 shows graphs of the first eight functions, to give an idea of the
structure of the Haar functions.
H 1 (x)
H 2 (x)
1
2
~ 1
I
I
H 5 (x)
H6(x)
.l
2
1
4
I
I
I
I
I
I
I I
L.J
2
H 4 (x)
21/2
21/2
~
I
I
21/2
I
I
;)_
I
I
I
I
I
I
I
I
4 I
I
I
I
I
I
I
I
I I
~I
;)_
2
,.,
I
I
I
I
I
I
I
I
.l
4
2''2
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
'+l
H 8 (x)
.ll
H?Cx)
I
I
I
2
I
I
I
I
I
I
I
I
I
H 3 (x)
r+i
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
.ll
I
I
I
I
I
I
I
I
I I
r+i
I
I
I
I
I
I
I
I
I
I
1
2
Figure 32. The Haar functions H 1(x), ... , H 8(x).
I
I
I
I
I
I
I
I
I
:I
11
I
4 I
I
I
I
I
I
I
I
I I
270
II. Mathematical Foundations of Probability Theory
It is easy to see that the Haar system is orthonormal. Moreover, it is
complete both in L 1 and in L 2 , i.e. iff= f(x) E IJ for p = 1 or 2, then
lf(x)  k~1 (f, Hk)Hk(x)IP dx+ 0,
n+ oo.
The system also has the property that
n
L (f, Hk)Hk(x)+ f(x),
n+ oo,
k=1
with probability 1 (with respect to Lebesgue measure).
In 4, Chapter VII, we shall prove these facts by deriving them from general
theorems on the convergence of martingales. This will, in particular, provide
a good illustration of the application of martingale methods to the theory of
functions.
6. If t] 1, ... , t'/n is a finite orthonormal system then, as was shown above, for
every random variable ~ E L 2 there is a random variable ~ in the linear manifold 2 = 2{tJ 1, ... , t'/n}, namely the projection of~ on f1, such that
II~
 ~II
= inf{ll~ ell: ( E ff{t'/1, , t'fn}}.
L7=
Here~=
1 (~, t'/;)t'/; This result has a natural generalization to the case
when t] 1, t] 2 , is a countable orthonormal system (not necessarily a basis).
In fact, we have the following result.
Theorem. Let
t] 1 , t] 2 ,
be an orthonormal system of random variables, and
L = L{tJ 1, t] 2 , } the closed linear manifold spanned by the system. Then
there is a unique element ~ E L such that
II~ ~II = inf{ll~  Cll: (
2}.
(20)
Moreover,
~ = l.i.m.
n
and ~  ~
L (~, t'/;)t'/;
(21)
i= 1
l_ (, ( E L.
Let d = inf{ll~ Cll: ( E 2} and choose a sequence ( 1, ( 2 , such
that II~  (nil +d. Let us show that this sequence is fundamental. A simple
calculation shows that
PROOF.
II("
(mll 2
= 211'n
~11 2 + 211'm ~11 2  411'"; 'm ~r
It is clear that ((n + (m)/2 E 2; consequently
therefore ll(n  (mll 2 + 0, n, m+ oo.
II[((" + (m)/2]
 ~11 2 ~ d 2 and
II. The Hilbert Space of Random Variables with Finite Second Moment
271
The space L 2 is complete (Theorem 7, 10). Hence there is an element ~
such that II(.  ~II + 0. But !lis closed, so~ E !l. Moreover, II(.  ~II + d,
and consequently II~  ~II = d, which establishes the existence of the required element.
Let us show that ~ is the only element of !l with the required property.
Let ~ E !l and let
II~  ~II = II~  ~II = d.
Then (by Problem 3)
II~+~ 2~11 2 + II~ ~11 2
211~ ~11 2 + 211~ ~11 2 = 4d 2
But
II~+~ 2~11 2
411tC~
+ ~) ~11 2 ~ 4d 2
Consequently II~  ~11 2 = 0. This establishes the uniqueness of the element
of !l that is closest to ~.
Now let us show that ~  ~ l_ (, ( E !l. By (20),
II~  ~  c(ll ~ II~  ~II
for every c E R. But
II~
 ~ c'll 2
 ~11 2 + c2 WI 2
II~
2(~  ~,cO.
Therefore
c2 ll'll 2
~ 2(~  ~,
c().
(22)
Take c = A(~  ~' (), AE R. Then we find from (22) that
(~ ~' 0 2 [A 2 II'II 2  2A.] ~ 0.
< 0 if A is a sufficiently small positive number. Con
We have A. 2 11'11 2  2,{
sequently(~  ~' 0 = 0, ( E I.
It remains only to prove (21).
The set!= i{111, 17 2 , . } is a closed subspace of L 2 and therefore a
Hilbert space (with the same scalar product). Now the system 17 1 ,17 2 , ...
is a basis for! (Problem 4), and consequently
n
L (~, 11k)11k
l.i.m.
(23)
But ~  ~ l_ l]k, k ~ 1, and therefore(~, I'Jk) = (~, '1k), k ~ 0. This, with (23)
establishes (21).
This completes the proof of the theorem.
Remark. As in the finitedimensional case, we say that ~ is the projection of
~ on L = L{rJ 1, 1] 2 , }, that ~ ~ is perpendicular to L, and that the
representation
is the orthogonal decomposition
+ (~
of~.
~)
272
II. Mathematical Foundations of Probability Theory
We also denote~ by E(el'7 1 , 17 2 , ) and call it the conditional expectation
in the wide sense (of with respect to '11> 17 2 , ). From the point of view of
estimating in terms of '71> '72'
the variable eis the optimal linear estimator, with error
ll
'
= e1e ~1 2 = 11e ~11 2 = 11e11 2
00
L l<e. '7i)l
i=l
which follows from (5) and (23).
7. PROBLEMS
1. Show that if~ = l.i.m.
2. Show that if
!I e. !I + lleli
e= l.i.m. e. and t7 = l.i.m. tfn then (e., tt.) + (e, tf).
3. Show that the norm
4. Let (
~.then
1 , , ~J
1111 has the parallelogram property
II~+
ttll 2 + II~ ttll 2
= 2(11~11 2
+ llttll 2).
be a family of orthogonal random variables. Show that they have the
Pythagorean property,
5. Let tf 1 , tf 2 , be an orthonormal system and !l' = !l'{tt 1, tf 2 , } the closed linear
manifold spanned by tt 1, tt 2 , . Show that the system is a basis for the (Hilbert)
~~
6. Let ~ 1 , ~ 2 , be a sequence of orthogonal random variables and s. = ~ 1 + . + ~ .
Show that if L,~ 1 E~~ < oo there is a random variable S with ES2 < oo such that
l.i.m. s. = S, i.e. liS. Sll 2 = E IS. Sl 2 + 0, n+ oo.
7. Show that in the space L 2 = L 2 ([ n, n], aJ([ n, n]) with Lebesgue measure p.
the system {(1/fo)eu", n = 0, 1, ...} is an orthonormal basis.
12. Characteristic Functions
1. The method of characteristic functions is one of the main tools of the
analytic theory of probability. This will appear very clearly in Chapter III
in the proofs of limit theorems and, in particular, in the proof of the central
limit theorem, which generalizes the De MoivreLaplace theorem. In the
present section we merely define characteristic functions and present their
basic properties.
First we make some general remarks.
Besides random variables which take real values, the theory of characteristic functions requires random variables that take complex values (see
Subsection 1 of 5).
273
12. Characteristic Functions
Many definitions and properties involving random variables can easily
be carried over to the complex case. For example, the expectation E( of a
complex random variable C= ~ + i17 will exist if the expectations E~ and
E17 exist. In this case we define E( = E~ + iE17. It is easy to deduce from the
definition of the independence of random elements (Definition 6, 5) that
the complex random variables ( 1 = ~ 1 + i17 1 and ( 2 = ~ 2 + i17 2 are independent if and only if the pairs (~ 1 , 17 1 ) and (~ 2 , 17 2 ) are independent; or,
equivalently, the aalgebras !l' ~ .. ~~ and !l' ~M 2 are independent.
Besides the space L 2 of real random variables with finite second moment,
we shall consider the Hilbert space of complex random variables C= ~ + i17
with EICI 2 < oo, where 1(1 2 = ~ 2 + 17 2 and the scalar product (( 1, ( 2 ) is
defined by EC 1 C2 , where C2 is the complex conjugate of(. The term "random
variable" will now be used for both real and complex random variables,
with a comment (when necessary) on which is intended.
Let us introduce some notation.
We consider a vector a E R" to be a column vector,
~CJ
and aT to be a row vector, aT = (a 1, . , an). If a andb E R"their scalar product
(a, b) is
1 a;b;. Clearly (a, b) = aTb.
If a E R" and ~ = llriill is ann by n matrix,
Li=
(~a,a) = aT~a =
L riiaiai.
(1)
i,j= 1
2. Definition 1. Let F = F(x 1 , , xn) be an ndimensional distribution
function in (R", PJJ(R")). Its characteristic function is
cp(t)
= ( ei<t,x> dF(x),
JR"
teR".
(2)
Definition 2. If~ = (~ 1 , , ~")is a random vector defined on the probability
space (Q, .fF, P) with values in R", its characteristic function is
C{);(t) = ( ei<t,x) dF;(x),
JR"
where F~ = F~(x 1 ,
teR",
xn) is the distribution function of the vector
(~1> ... ' ~n).
If F(x) has a density f = f(x) then
cp(t) =
r ei(t,x)f(x) dx.
JRn
(3)
~ =
274
II. Mathematical Foundations of Probability Theory
In other words, in this case the characteristic function is just the Fourier
transform of f(x).
It follows from (3) and Theorem 6. 7 (on change of variable in a Lebesgue
integral) that the characteristic function cp~(t) of a random vector can also
be defined by
(4)
tER".
We now present some basic properties of characteristic functions, stated
and proved for n = 1. Further important results for the general case will be
given as problems.
Let~ = ~(w) be a random variable, F~ = F~(x) its distribution function,
and
its characteristic function.
We see at once that if 11
cp~(t)
= a~ + b then
= Eeit~ = Eeit(a~+b) =
eitbEeiat~.
Therefore
(5)
sn
Moreover, if ~ t> ~ 2 , , ~" are independent random variables and
= ~1 + ... +~",then
n cp~lt).
n
CfJs.(t) =
(6)
i= 1
In fact,
= Eeu~,
... Eeit~.
j=1
cp~
(t),
where we have used the property that the expectation of a product of independent (bounded) random variables (either real or complex; see Theorem 6
of 6, and Problem 1) is equal to the product of their expectations.
Property (6) is the key to the proofs of limit theorems for sums of independent random variables by the method of characteristic functions (see 3,
Chapter III). In this connection we note that the distribution function F s"
is expressed in terms of the distribution functions of the individual terms in a
rather complicated way, namely F s" = F ~~ * * F ~" where * denotes
convolution (see 8, Subsection 4).
Here are some examples of characteristic functions.
1. Let ~ be a Bernoulli random variable with P(~
0) = q, p + q = 1, 1 > p > 0; then
ExAMPLE
P(~
cp~(t) = peit
+ q.
= 1) =
p,
275
12. Characteristic Functions
If ~ 1 ,
... , ~n
are independent identically distributed random variables like
~,then, writing T,. = (Sn  np)/~, we have
<r>rJt) = EeiTnt = eit0ifW[peitfv'iijiq + q]n
[peitJq7(npJ
Notice that it follows that as n
+
+ qeitv'p/(nq)y.
(7)
oo
sn np
(8)
T,.= ~.
Let ~ "' %(m, a 2 ), Im I < oo, a 2 > 0. Let us show that
EXAMPLE 2.
(9)
Let 1J =
(~
m)ja. Then 1J "' %(0, 1) and, since
<p~(t)
= eitm<p~(at)
by (5), it is enough to show that
<p~(t) =
er2;z.
(10)
We have
<p~(t)
= Ee 11 ~ = 1
J21r:
Joo e'.xex2j2 dx
1
00
Joo 1..e
~ (itxt x2jZ dx_ 1
~ .(itt
n x2j2 dx
.   1  Joo xe
J21r: oo n=O n!
n=O n! J21r: oo
_ 1
=
=
(it)2n (2n  1) I I =
(it)2n (2n)!
n=O (2n)!
n=O (2n)! 2nn!
f ( tZ)n.
~
2
n.
n=O
= et212,
where we have used the formula (see Problem 7 in 8)
EXAMPLE 3.
Let
be a Poisson random variable,
eAA_k
P(~ = k) =
Then
kl'
= 0, 1, ....
276
II. Mathematical Foundations of Probability Theory
3. As we observed in 9, Subsection 1, with every distribution function in
(R, fJI(R)) we can associate a random variable of which it is the distribution
function. Hence in discussing the properties of characteristic functions (in
the sense either of Definition 1 or Definition 2), we may consider only
characteristic functions ({J(t) = (/J(,(t) of random variables e = e(w).
Theorem 1. Let be a random variable with distribution function F = F(x) and
({J(t) =
its characteristic function. Then
qJ
Eeu~
has the following properties:
(1) I({J(t) I ~ ({J(O) = 1;
(2) ({J(t) is uniformly continuous fortE R;
(3) ({J(t) = ({J( t);
(4) ({J(t) is realvalued if and only ifF is symmetric <h dF(x) = f B dF(x)),
Be fJI(R), B = {x: x eB};
(5) if EI I" < oo for some n :2 1, then ({J<'>(t) exists for every r ~ n, and
(ix)'eitx dF(x),
({J(rl(t) =
({J(rl(O)
Ee' = .,,
(13)
~ (it) 2 E~'
() = L...
({Jt
.,
1
r=o r.
(12)
()
+ (it)"1 e.t,
n.
(14)
where le.(t)l :::; 3E 1e1 and e.(t)+ 0, t+ 0;
(6) if qJ< 2 >(0) exists and is .finite then Ee 2 " < oo;
(7) if E I I" < oo for all n :2 1 and
 . (Eiel")li
hm
"
=
< oo,
then
({J(t) =
n=O
(itr Ee.
n.
(15)
for allltl < R.
PRooF. Properties (1) and (3) are evident. Property (2) follows from the
inequality
+ h) 
IEeit(,(eih~  1) I :::; EIeih(,  11
and the dominated convergence theorem, according to which EIeih(.  11 + 0,
I({J(t
({J(t) I =
h+ 0.
Property (4). Let F be symmetric. Then if g(x) is a bounded odd Borel
function, we have fR g(x) dF(x) = 0 (observe that for simple odd functions
277
12. Characteristic Functions
this follows directly from the definition of the symmetry of F). Consequently
JR sin tx dF(x) = 0 and therefore
cp(t) = E cos
Conversely, let
be a real function. Then by (3)
cp~(t)
({J lt)
t~.
= cp~( t) = cplt) = cp~(t),
t E R.
Hence (as will be shown below in Theorem 2) the distribution functions
~ and ~ are the same, and therefore (by
Theorem 3.1)
F _~and F ~of the random variables 
P( ~ E B) = P( ~ E B) = P( ~ E B)
for every BE fJ(R).
Property (5). If E I~ In < oo, we have E I~ lr < oo for r
inequality (6.28).
Consider the difference quotient
+ h)
cp(t
 cp(t) _
ir~(eih~  Ee
h
n, by Lyapunov's
1) .
Since
i
eihx _
h
11
lxl,
and E1 ~ 1 < oo, it follows from the dominated convergence theorem that the
limit
lim
h~o
exists and equals
. lim (eih~ _
h~o
h
Eeu~
Eeu~(ei_h~__1)
1) =
iE(~e' 1 ~)
= i
Joo xe'.x dF(x).
oo
(16)
Hence q/(t) exists and
cp'(t) =
i(E~eil~) =
s:oo xeitx dF(x).
The existence of the derivatives cp(r)(t), 1 < r ~ n, and the validity of (12),
follow by induction.
Formula (13) follows immediately from (12). Let us now establish (14).
Since
.
e'Y
=cosy
+ l Slll y
n1 (iy)k
= k~O F
(iy)n
+7
[cos ely+
. .
l Sill
e2y]
278
II. Mathematical Foundations of Probability Theory
for real y, with I8 1 I :::; 1 and I8 2 1:::; 1, we have
e''~
n 1 (it~)k
(it~t
= k~O ;z! +;:;![COS 81(w)t~ + i Sln 82(w)t~]
(17)
and
(18)
where
en(t) =
P[~"(cos 8 1 (w)t~
+ i sin 8 2 (w)t~
 1)].
It is clear that Ic5n(t) I :::; 3E I~n 1. The theorem on dominated convergence
shows that en(t) + 0, t + 0.
Property (6). We give a proof by induction. Suppose first that cp"(O)
exists and is finite. Let us show that in that case E~ 2 < oo. By L'Hopital's
rule and Fatou's lemma,
qJ
~ [cp'(2h)  cp'(O)
2h
"(O) = 1.
h~~ 2
cp'(O)  cp'( 2h)]
2h
= lim 2cp'(2h) ~~cp'(  2h) = lim 4h12 [cp(2h) 2cp(O) + cp( 2h)]
h+0
h+0
= lim
Joo
h+0
lim
h+0
= 
(eihx ;heihx)2 dF(x)
00
hx)
oo (sin
hJoo
x 2 dF(x) :::; 
hx) 2x
Joo lim (sin
hoo
h+0
dF(x)
J:oo x 2 dF(x).
Therefore
J:oo x
dF(x):::; cp"(O) < oo.
Now let cp< 2 k+ 2 l(O) exist, finite, and let J~~ x 2k dF(x) < oo. If J~oo x 2kdF(x)
= 0, then J~oo x 2k+ 2 dF(x) = 0 also. Hence we may suppose that
J~ x 2k dF(x) > 0. Then, by Property (5),
oo
cp<2kl(t)
f_oooo (ix)2keitx dF(x)
and therefore
( 1)kcp(2kl(t)
where G(x)
J~oo u 2 k dF(u).
s:ooeitx dG(x),
279
12. Characteristic Functions
Consequently the function ( 1)ktp< 2 k>(t)G 1 ( oo) is the characteristic
function ofthe probability distribution G(x) G 1(oo) and by what we have
proved,
G 1(oo)
J:
00
x 2 dG(x) < oo.
But G 1(oo) > 0, and therefore
Property (7). Let 0 < t 0 < R. Then, by Stirling's formula we find that
[E I~ l"t0/n !] converges by Cauchy's test, and
Consequently the series
therefore the series L~ 0 [(it)'jr !]E~' converges for It I ::;; t 0 But by (14),
for n ~ 1,
('t)'
L _z_l
E~' + Rn(t),
n
tp(t) =
r=O
tp(t) =
r=O
(it)'
r!
E~'
for all It I < R. This completes the proof of the theorem.
Remark 1. By a method similar to that used for (14), we can establish that if
00 for some n ~ 1, then
EIeI" <
" ik(t  s)k
tp(t) = k~O
k!
where len(t
s)l ::;;
Joo
oo xkex
dF(x)
3E WI, and en(t s)
t
i"(t  st
n!
en(t s),
(19)
0 as t s t 0.
Remark 2. With reference to the condition that appears in Property (7),
see also Subsection 9, below, on the "uniqueness of the solution of the
moment problem."
4. The following theorem shows that the characteristic function is uniquely
determined by the distribution function.
280
II. Mathematical Foundations of Probability Theory
~.
+6
Figure 33
Theorem 2 (Uniqueness). Let F and G be distribution functions with the same
characteristic function, i.e.
(20)
for all t E R. Then F(x)
= G(x).
PROOF. Choose a and bE R, and B > 0, and consider the function f' = f'(x)
shown in Figure 33. We show that
(21)
Let n ~ 0 be large enough so that [a e, b + e] ; [ n, n], and let the
sequence {b"} be such that 1 ~ b" L0, n + oo. Like every continuous function
on [ n, n] that has equal values at the endpoints,!' = f'(x) can be uniformly
approximated by trigonometric polynomials (Weierstrass's theorem), i.e.
there is a finite sum
(22)
such that
sup lf'(x) f~(x)l:::;; b".
nsx:s;n
Let us extend the periodic function J,.(x) to all of R, and observe that
sup ln(x)l:::;; 2.
X
Then, since by (20)
(23)
281
12. Characteristic Functions
we have
f_
00
IJ:00 f'(x)dF(x)
00
f'(x)dG(x)l
= lfJ'dF f.f'dGI
~I
f.n
dF
f/~ dGI + 2<5n
~ If_0000 f~ dF f_0000 f~ dGI + 25n
+ 2F([ n, n]) + 2G([ n, n]),
(24)
where F(A) = SA dF(x), G(A) = SA dG(x). As n+ oo, the righthand side
of (24) tends to zero, and this establishes (21).
Ass+ 0, we have f'(x)+ I(a,b/x). It follows from (21) by the theorem on
distribution functions' being the same.
f_'xo
00
/(a,bJ(x) dF(x)
= f_'xoool(a,b](x) dG(x),
i.e. F(b)  F(a) = G(b)  G(a). Since a and b are arbitrary, it follows that
F(x) = G(x) for all x E R.
This completes the proof of the theorem.
5. The preceding theorem says that a distribution function F = F(x) is
uniquely determined by its characteristic function <p = <p(t ). The next theorem
gives an explicit representation ofF in terms of <p.
Theorem 3 (Inversion Formula). Let F = F(x) be a distribution function and
<p(t) = s:ooeitx dF(x)
its characteristic function.
(a) For pairs of points a and b (a< b) at which F
F(b) F(a) = lim 21
c+oo
fc
eita
c
1t
F(x) is continuous,
eitb <p(t) dt;
(25)
lt
(b) If s~oo l<fJ(t)l dt < oo, the distribution function F(x) has a density f(x),
F(x) = roof(y) dy
(26)
and
f(x) = 1
2n
J(X)
oo
.
etxq>(t)
dt.
(27)
II. Mathematical Foundations of Probability Theory
282
PROOF.
We first observe that if F(x) has density f(x) then
f_'>)oo eitj(x) dx,
cp(t) =
(28)
and (27) is just the Fourier transform of the (integrable) function cp(t).
Integrating both sides of (27) and applying Fubini's theorem, we obtain
F(b) F(a) =
ff(x)
dx =
2~ f [f_
00
f_oooo cp(t) [feitx dx
00
eitxcp(t) dt] dx
Jdt
21n
eita eitb
1 Joo
dt.
cp(t)
2
zt.
n _ 00
After these remarks, which to some extent clarify (25), we turn to the proof.
(a) We have
~c
= 2n
1
=2n:
1
2n
=
fc
eita eitb
zt
c
fc
c
eita _ eitb
.
zt
Joo
[fc
_ 00
c
cp(t) dt
[foo
eitx dF(x) dt
00
eita _ eitb
eitx dt dF(x)
.
zt
(29)
where we have put
'I'c(x) = 
2n
fc
c
eita  eitb .
eztx dt
.
zt
and applied Fubini's theorem, which is applicable in this case because
_ei_ta_e_itb. eitxl
l
it
and
fc
1:
leita eitbl
it
(b  a) dF(x)
=I Ja
rbeitx dxl < b a
~ 2c(b 
a) < oo.
283
12. Characteristic Functions
In addition,
n1
T
c X
= _!_
27C
1
=
2n
fc
c
sin t(x  a)  sin t(x  b)
dt
t
fc(xa)
c(xa)
sin v
1
  dv  2n
The function
g(s, t)
fc(xb)
c(xb)
sin u
  du.
(30)
sin v
dv
v
is uniformly continuous in s and t, and
g(s, t) + n
(31)
ass!  oo and t j oo. Hence there is a constant C such that I'Pc(x) I < C < oo
for all c and x. Moreover, it follows from (30) and (31) that
'l'c(x)+ 'l'(x),
c +
00,
where
0, X < a, X > b,
'P(x) = { !, x = a, x = b,
1, a< x <b.
Let ll be a measure on (R, f!J(R)) such that /l(a, b] = F(b) F(a). Then
if we apply the dominated convergence theorem and use the formulas of
Problem 1 of 3, we find that, as c+ oo,
<l>c =
f_
00
00
'Pc(x) dF(x)+
f_
00
00
'P(x) dF(x)
= {L(a, b)+ t!l{a) + t!l{b}
= F(b) F(a) + t[F(a) F(a) + F(b) F(b )]
= F(b) \F(b) _ F(a) \F(a) = F(b) _ F(a),
where the last equation holds for all points a and b of continuity of F(x).
Hence (25) is established.
(b) Let f~oo I<P(t)l dt < 00. Write
f(x) = 21
7C
foo e"xcp(t)
. dt.
oo
284
II. Mathematical Foundations of Probability Theory
It follows from the dominated convergence theorem that this is a continuous
function of x and therefore is integrable on [a, b]. Consequently we find,
applying Fubini's theorem again, that
1
=lim 2
rae 1C
Jr
eita eitb
.
zt
c
qJ(t) dt = F(b)  F(a)
for all points a and b of continuity of F(x).
Hence it follows that
F(x) =
aof(y) dy,
XER,
and since f(x) is continuous and F(x) is nondecreasing, f(x) is the density
of F(x).
This completes the proof of the theorem.
Corollary. The inversion formula (25) provides a second proof of Theorem 2.
Theorem 4. A necessary and sufficient condition for the components of the
<el, ... ,
en) to be independent is that its characteristic
random vector e =
function is the product of the characteristic functions of the components:
Eei(t1~1 + ... +tn~n)
fl
Eeitk~k,
k=1
PRooF. The necessity follows from Problem 1. To prove the sufficiency we
let F(x 1, ... , Xn) be the distribution function of the vector e = (e 1, , en)
andFk(x), the distribution functions of the ek, 1 ~ k ~ n. PutG = G(xl> ... , Xn)
= F 1 (x 1 ) F n(xn). Then, by Fubini's theorem, for all (t 1 , .. , tn) ERn,
fl
Eeitk~k
Eei<r1~1 + ... +tk~k>
k=1
JRn
ei(I!X!+"+tnXn)
dF(x 1
...
Xn)
Therefore by Theorem 2 (or rather, by its multidimensional analog; see
Problem 3) we have F = G, and consequently, by the theorem of 5, the
random variables eh ... ' en are independent.
285
12. Characteristic Functions
6. Theorem 1 gives us necessary conditions for a function to be a characteristic
function. Hence if qJ = ((J(t) fails to satisfy, for example, one of the first three
conclusions of the theorem, that function cannot be a characteristic function.
We quote without proof some results in the same direction.
BochnerKhinchin Theorem. Let qJ(t) be continuous, t E R, with qJ(O) = I. A
necessary and sufficient condition that ((J(t) is a characteristic function is that it
is positive semidefinite, i.e. that for all real t 1, .. , tn and all complex Ato ... , A.n,
n = 1, 2, ... ,
n
((J(t;  tj)A.)j ~
i,j= 1
The necessity of (32) is evident since if ((J(t)
i.tl
o.
(32)
f~ oo
eirx dF(x) then
((J(t; tj)A.J.j = J:JJ/keilkXI2 dF(x)
~ o.
The proof of the sufficiency of (32) is more difficult.
Polya's Theorem. Let a continuous even function qJ(t) satisfy qJ(t) ~ 0,
qJ(O) = 1, qJ(t) + 0 as t + oo and let qJ(t) be convex on 0 :s; t < oo. Then
((J(t) is a characteristic function.
This theorem provides a very convenient method of constructing characteristic functions. Examples are
((Jl (t)
e lrl,
()  {1  ltl,
cpzt
It I :s; 1,
ltl >
0,
1.
Another is the function cp 3 (t) drawn in Figure 34. On [a, a], the function
qJ 3(t) coincides with qJ 2(t). However, the corresponding distribution functions F 2 and F 3 are evidently different. This example shows that in general
two characteristic functions can be the same on a finite interval without their
distribution functions' being the same.
1
a
Figure 34
286
II. Mathematical Foundations of Probability Theory
Marcinkiewicz's Theorem. If a characteristic function cp(t) is of the form
exp &>(t), where &>(t) is a polynomial, then this polynomial is of degree at
most 2.
It follows, for example, that er is not a characteristic function.
7. The following theorem shows that a property of the characteristic
function of a random variable can lead to a nontrivial conclusion about the
nature of the random variable.
Theorem 5. Let
cp~(t)
be the characteristic function of the random variable
(a) If Icp~(t 0 ) I = 1 for some t 0 =I= 0, then
+ nh, h = 2n/t 0 , for some a, that is,
is concentrated at the points
00
n= oo
Pg =a+ nh}
(33)
1,
where a is a constant.
(b) If
1 for two different points t and
is degenerate:
lcp~(t)l = lcp~(~t)l =
irrational, then
P{~
~t,
where
is
=a} = 1,
where a is some number.
= 1, then~ is degenerate.
(c) If Icp~(t)l
PROOF.
Then
(a) If Icp~(t 0 )1 = 1, t 0 =I= 0, there is a number a such that cp(t 0 ) = eitoa.
1=
f_
00
00
COS t 0 (x
 a) dF(x) =>
f_
00
00
[1 
 a)] dF(x)
COS t 0 (x
= 0.
Since 1  cos t 0 (x  a) 2 0, it follows from property H (Subsection 2 of
6) that
1
= cos t 0 (~
which is equivalent to (33).
(b) It follows from lcp~(t)l
n= oo
p{~ =
a)
(Pa.s.),
= lcp~(~t)l = 1 and from (33) that
a + 2n
t
n} f
=
m= oo
p{~ =
b + 2n
~t
m}
= 1.
If ~ is not degenerate, there must be at least two pairs of common points:
a
2n
+ t n 1
2n
b +  m1
~t
'
2n
a+ n 2 = b
y
2n
+ m 2 ,
~t
287
12. Characteristic Functions
in the sets
{b + ~~ m, m = 0, 1, .. }
{a+ 2; n, n = 0, 1, .. } and
whence
2n
(n 1
t
n2 ) =
2n
at
~(m 1 
m2 ),
and this contradicts the assumption that a is irrational. Conclusion (c)
follows from (b).
This completes the proof of the theorem.
8.
Let~= (~b
... , ~k) be a random vector,
lfJ~(t)
= Eei(t, ~>,
t = (t 1
.. '
tk),
its characteristic function. Let us suppose that E I~; In < oo for some n ; : .: 1,
i = 1, ... , k. From the inequalities of Holder (6.29) and Lyapunov (6.27)
it follows that the (mixed) moments E(~1' ~;;k) exist for all nonnegative
v1, ... , vk such that v1 + + vk :::;; n.
As in Theorem 1, this implies the existence and continuity of the partial
derivatives
for v1 + + vk :::;; n. Then if we expand
we see that
l{J~(tl, ... ,tk)=
i'++vk
++vk:Sn v1 .... vk.
.. ,
tk) in a Taylor series,
m<~,,. .. ,vklt1'tkk+o(ltln),
m(v., ... ,vk) _ EJ::Vt
'>1
~
lfJ~(t 1 ,
(34)
)::Vk
'>k
is the mixed moment of order v = (v 1, ... , vk).
Now lfJ~(t 1 , ... , tk) is continuous, lfJ~(O, ... , 0) = 1, and consequently this
function is different from zero in some neighborhood It I < 8 of zero. In
this neighborhood the partial derivative
a+ ... +vk
Vk In l{J~(tb ... ' tk)
1 . . tk
at VI
exists and is continuous, where In z denotes the principal value of the
logarithm (if z = rei8 , we take In z to be In r + W). Hence we can expand
In lfJlt 1, .. ,, tk) by Taylor's formula,
iVl + + Vk
s~ klt'l' t);k + o(lt n, (35)
I
ln l{J~(t 1 tk) =
VI + ... + Vk :S n VI ! Vk !
288
II. Mathematical Foundations of Probability Theory
where the coefficients s~v vk) are the (mixed) semiinvariants or cumulants
of order v = v(v 1, .. , vk) of~= ~ 1 , .. , ~k
Observe that if ~ and '1 are independent, then
ln
cp~+,(t)
= ln
+ ln cp,(t),
cp~(t)
(36)
and therefore
(37)
(It is this property that gives rise to the term "semiinvariant" for s~v ... ,vk>.)
To simply the formulas and make (34) and (35) look "onedimensional,"
we introduce the following notation.
If v = (vh ... , vk) is a vector whose components are nonnegative integers,
we put
We also put s~v> = s~v ... ,vk>, m~v) =
Then (34) and (35) can be written
cpr,(t) =
ln cpr,(t) =
m~v ... ,vk>.
jlvl
lvlsn V.
ilvl
Lr
lvlsn V.
m~v)tv
s~v>tv
+ o(ltl"),
(38)
+ o(ltl").
(39)
The following theorem and its corollaries give formulas that connect
moments and semiinvariants.
Theorem6. Let ~=(~h~k) be a random vector with
i = 1, ... , k, n;;;:: 1. Then for v =(vi> ... , vk) such that Ivi:::;; n
(V) 
m~;
_.!..
"
 ).Ol++A<l=v
L.
' 1(1)f
q II.
v!
1(q)f'
/1.
nq
p=l
().(P))
'
El~d"<oo,
(40)
(41)
where LA<l++A<l=v indicates summation over all ordered sets of nonnegative
integral vectors A<P>, IA<P> I > 0, whose sum is v.
PROOF.
Since
cpr,(t) = exp(ln cpr,(t)),
if we expand the function exp by Taylor's formula and use (39), we obtain
cpr,(t) = 1 +
1(
iiAI
)q
L
I
L
s~A>tA + o(ltl").
11
q=l q., lSIAisn
n
11..
(42)
289
12. Characteristic Functions
Comparing terms in e on the righthand sides of (38) and (42), and using
1..1<1)1 + + IA.(qJI = 1..1< 1> + + A_<qll, we obtain (40).
Moreover,
In
L ~~~ m~"lt" + o( It 1")].
cp~(t) = ln[1 + 1:<>1.<1:;n
(43)
For small z we have the expansion
ln(1
+ z)
L n
q= 1
1)q1
q
zq
+ o(zry.
Using this in (43) and then comparing the coefficients oft;., with the corresponding coefficients on the righthand side of (38), we obtain (41).
Corollary 1. The following formulas connect moments and semiinvariants:
n[
X
(44)
(}.Ul)J'j
v.
)
m~  {rt)..(l)+ .. +rx).(X)=v} r1! rX! (A_(l)!)'' (A_(X)!yxj=l s~
(v) 
s~'l
L
{rt),(IJ + ... + rx;.,(x) = v}
Il [
<.<UlJ]'
v!
(1)q1(q 1)!
J,
m~
... (1(X)I.)'xJ._1
(1(1).1)''
rx'
I
rl ....
1\.
1\.
(45)
where 'L1,,;.(1J+ .. +rx.<Cxl=vJ denotes summation over all unordered sets of
different nonnegative integral vectors A_Ul, IA_W I > 0, and over all ordered sets of
positive integral numbers ri such that r 1A.< 1l + + rxA_(x) = v.
To establish (44) we suppose that among all the vectors A_(ll, ... , A_(q)
that occur in (40), there are r 1 equal to A_(itl, ... , rx equal to A_(ixl (ri > 0,
r1 + + rx = q), where all the A_(i.J are different. There are q !f(r 1! ... rx !) different sets of vectors, corresponding (except for order) with the set {A_(ll, ...
il(ql}). But if two sets, say, {il< 1>, . , il<ql} and {J(ll, ... , J<ql} differ only in order,
then n~= 1 s~).(Pl) = n~= 1 s~'J.(Pl). Hence if we identify sets that differ only in
order, we obtain (44) from (40).
Formula (45) can be deduced from (41) in a similar way.
Corollary 2. Let us consider the special case when v = (1, ... , 1). In this case
the moments m~v) E~ 1 ~k' and the corresponding semiinvariants, are
called simple.
Formulas connecting simple moments and simple semiinvariants can
be read off from the formulas given above. However, it is useful to have them
written in a different way.
For this purpose, we introduce the following notation.
Let~= (~I> ... , ~k) be a vector, and I~= {1, 2, ... , k} its set of indices.
If I c:;: I~, let ~~denote the vector consisting of the components of~ whose
290
II. Mathematical Foundations of Probability Theory
indices belong to I. Let x(J) be the vector {XI> ... , Xn} for which Xi = 1 if
i e I, and Xi = 0 if i I. These vectors are in onetoone correspondence with
the sets I s; I~. Hence we can write
In other words, m~(J) and s~(J) are simple moments and semiinvariants
of the subvector of
In accordance with the definition given on p. 12, a decomposition of
a set I is an unordered collection of disjoint nonempty sets I P such that
LPIP =I.
In terms of these definitions, we have the formulas
el e.
(46)
q
sp)=
(1)q 1(q1)!0mpp).
l:~=llp=I
(47)
p=l
where Lr~=lip=I denotes summation over all decompositions of I,
1 ~ q ~ N(l).
We shall derive (46) from (44). If v = x(I) and A.< 1 > + + A.<q> = v, then
A_<P> = x(IP), IPs; I, where the A_<P> are all different, A_<P>f = v! = 1, and every
unordered set {x(J 1 ), , x(Iq)} is in onetoone correspondence with the
decomposition I= L~=l IP. Consequently (46) follows from (44).
In a similar way, (47) follows from (35).
EXAMPLE 1. Let
e be. a
random variable (k = 1) and mn = m~n> = Ee",
sn =st. Then (40) and (41) imply the following formulas:
ml = sl,
+ si,
s 3 + 3s 1s 2 + s~,
s4 + 3s~ + 4sls3 + 6sis2 + st,
m2 = s2
m3 =
m4 =
(48)
and
s1
= m 1 = Ee,
s2 = m2 mi =
ve,
s 3 = m3  3m 1 m2 + 2mt
s4 = m4  3m~ 4m 1 m3 + 12mim2  6mt,
(49)
291
12. Characteristic Functions
ExAMPLE
2. Let ~ ,.... JV(m, a 2 ). Since, by (9),
In cp~(t) = itm 
t2(J2
T'
we have s 1 = m, s 2 = a 2 by (39), and all the semiinvariants, from the third
on, are zero: sn = 0, n 2 3.
We may observe that by Marcinkiewicz's theorem a function exp &'(t),
where .9' is a polynomial, can be a characteristic function only when the
degree of that polynomial is at most 2. It follows, in particular, that the
Gaussian distribution is the only distribution with the property that all its
semiinvariants snare zero from a certain index onward.
EXAMPLE
3. If
is a Poisson random variable with parameter A > 0, then
by (11)
In
cp~(t)
= .Jc(eit  1).
It follows that
(50)
for all n 2 1.
EXAMPLE
4.
Let~
(~ 1 , . , ~n)
m~(l)
be a random vector. Then
s~(l),
+ s~(l)s~(2),
mp, 2, 3) = s~(1, 2, 3) + s~(1, 2)s~(3) +
+ s~(l, 3)sp) +
+ s~(2, 3)sll) + s~(l)se(2)sl3)
m~(l,
2) =
s~(l,
2)
(51)
These formulas show that the simple moments can be expressed in terms
of the simple semiinvariants in a very symmetric way. If we put ~ 1 = ~ 2 =
~k, we then, of course, obtain (48).
The grouptheoretical origin of the coefficients in (48) becomes clear
from (51). It also follows from (51) that
se(l, 2) = me(l, 2) me(l)mp) =
E~ 1 ~ 2  E~ 1 E~ 2 ,
(52)
i.e., s/1, 2) is just the covariance of ~ 1 and ~ 2 .
9. Let ~ be a random variable with distribution function F = F(x) and
characteristic function cp(t). Let us suppose that all the moments mn = E~n,
n 2 1, exist.
It follows from Theorem 2 that a characteristic function uniquely determines a probability distribution. Let us now ask the following question
292
II. Mathematical Foundations of Probability Theory
(uniqueness for the moment problem): Do the moments {mn}n> 1 determine
the probability distribution?
More precisely, let F and G be distribution functions with the same
moments, i.e.
f_'Xloo xn dF(x)
f_oooo xn dG(x)
(53)
for all integers n ~ 0. The question is whether F and G must be the same.
In general, the answer is "no." To see this, consider the distribution F
with density
J(x) = {keax',
0,
> 0,
X:::; 0,
where a: > 0, 0 < A, < t, and k is determined by the condition
Write f3 = a: tan An and let g(x) = 0 for x :::; 0 and
+ t; sin(f3xA)],
g(x) = ke ax'[1
It is evident that g(x)
It; I <
1,
J0 f(x) dx =
1.
> 0.
0. Let us show that
(54)
for all integers n ~ 0.
For p > 0 and complex q with Re q
> 0, we have
ootp1eqt dt = r(p).
qP
Take p = (n
+ 1)/A, q
a:+ i/3, t
= xA. Then
r(~)
 a;<n+ 1)/\1
+ i tan A.n)<n+ 1)/A.
But
(1
+ i tan A.n)<n+ I)/A
= (cos An
+ i sin A.n)<n+ 1)/A(cos A.n)<n+ I)/A
= ein(n+ll(cos A.n)<n+l)/A
=cos n(n + 1) cos(A.n)<n+ IliA,
since sin n(n
+ 1) =
0.
(55)
293
12. Characteristic Functions
Hence righthand side of (55) is real and therefore (54) is valid for all
integral n 2:: 0. Now let G(x) be the distribution function with density g(x).
It follows from (54) that the distribution functions F and G have equal
moments, i.e. (53) holds for all integers n 2:: 0.
We now give some conditions that guarantee the uniqueness of the solution of the moment problem.
Theorem 7. Let F = F(x) be a distribution function and J.ln = s~ 00
IX I" dF(x).
If
J.l1/n
lim"< oo,
n~CX)
(56)
the moments {mn}n~ to where mn = J~ oo x" dF(x), determine the distribution
F = F(x) uniquely.
PRooF. It follows from (56) and conclusion (7) of Theorem 1 that there is a
t 0 > 0 such that, for alii t I ~ t 0 , the characteristic function
cp(t) = J:oo eirx dF(x)
can be represented in the form
00
(itl
cp(t) = k~o k! mk
and consequently the moments {mn}n> 1 uniquely determine the characteristic function cp(t) for It I ~ t 0 .
Take a points with lsi::;; t 0 /2. Then, as in the proof of (15), we deduce
from (56) that
({J
for
It sl
~ t0 ,
( ) = ~ ik(t  s)k <k>( )
t
k~O
k!
({J
where
cp<k>(s) = ik J:oo xkeix dF(x)
is uniquely determined by the moments {mn} n~ 1. Consequently the moments
determine cp(t) uniquely for It I ~ !to. Continuing this process, we see that
{mn}n~ 1 determines cp(t) uniquely for all t, and therefore also determines
F(x).
This completes the proof of the theorem.
Corollary 1. The moments completely determine the probability distribution
if it is concentrated on a finite interval.
294
II. Mathematical Foundations of Probability Theory
Corollary 2. A sufficient condition for the moment problem to have a unique
solution is that
(m )1/2n
llm
2n
2n
n+oo
<
(57)
00.
For the proof it is enough to observe that the odd moments can be
estimated in terms of the even ones, and then use (56).
EXAMPLE.
Let F(x) be the normal distribution function,
F(x) = _1_
fx
et2Jza2 dt.
oo
Then m 2 n+ 1 = 0, m 2 n = [(2n) !/2"n !]u 2 ", and it follows from (57) that these
are the moments only of the normal distribution.
Finally we state, without proof:
Carleman's test for the uniqueness of the moment problem.
(a) Let
{mn}n~ 1 be
the moments of a probability distribution, and let
00
L (mzn)1f2n =
n=O
00
Then they determine the probability distribution uniquely.
(b) If {mn}n~l are the moments of a distribution that is concentrated on
[0, oo ), then the solution will be unique if we require only that
00
L (mn)lf2n
n=O
00
10. Let F = F(x) and G = G(x) be distribution functions with characteristic functions f = f(t) and g = g(t), respectively. The following theorem,
which we give without proof, makes it possible to estimate how close F
and G are to each other (in the uniform metric) in terms of the closeness of
fandg.
Theorem (Esseen's Inequality). Let G(x) have derivative G'(x) with
supiG'(x)l ~C. Thenfor every T > 0
sup IF(x) G(x)l
x
~ ~ JT if(t) g(t)i dt +
n o
2T
4 sup IG'(x)l.
(58)
(This will be used in 6 of Chapter III to prove a theorem on the rapidity
of convergence in the central limit theorem.)
295
13. Gaussian Systems
11.
PROBLEMS
1. Let and , be independent random variables, f(x) = j~(x) + if2(x), g(x) = gl(x)
+ ig 2(x), where A(x) and gb) are Borel functions, k= 1, 2. Show that ifE I!WI< oo
and E lg(tf)l < oo, then
E lf{e)g(tf)l < oo
and
EJ(e)g(tf) = Ef(e) Eg(tf).
2. Let
e= (e
1 , ... ,
e.) and Ell ell"< oo, where WI =
cp~(t)
where t = (t 1,
+~.Show that
I ..:, E(t, e)k + e.(t)lltll",
k=Ok.
t.) and e.(t)> 0, t> 0.
3. Prove Theorem 2 for ndimensional distribution functions F = F.(x 1,
G.(x 1 ,
x.) and
x.).
4. LetF = F(x 1, ... , x.)beanndimensionaldistributionfunctionandcp = cp(t 1, .. , t.)
its characteristic function. Using the notation of (3.12), establish the inversion formula
(We are to suppose that (a, b] is an interval of continuity of P(a, b], i.e. fork= 1,
... , n the points ak, bk are points of continuity of the marginal distribution functions
Fk(xk) which are obtained from F(x~> ... , x.) by taking all the variables except
xk equal to + oo.)
5. Let cpk(t), k ~ 1, be a characteristic function, and let the nonnegative numbers A.k,
k ~ 1, satisfy I A.k = 1. Show that I A.kcpk(t) is a characteristic function.
6. If cp(t) is a characteristic function, are Re cp(t) and Im cp(t) characteristic functions?
7. Let cp 1, cp 2 and cp 3 be characteristic functions, and cp 1 cp 2 = cp 1 cp 3 Does it follow that
CfJ2 =
CfJ3?
8. Construct the characteristic functions of the distributions given in Tables 1 and 2
of~3.
9. Let be an integralvalued random variable and
Show that
P(e = k) = 1

2n
f" .
eu"cp~(t) dt,
x
cp~
(t) its characteristic function.
k = 0, 1, 2 ....
13. Gaussian Systems
1. Gaussian, or normal, distributions, random variables, processes, and
systems play an extremely important role in probability theory and in
mathematical statistics. This is explained in the first instance by the central
296
II. Mathematical Foundations of Probability Theory
limit theorem (4 of Chapter III and 8 of Chapter VII), of which the De
MoivreLaplace limit theorem is a special case (6, Chapter 1). According
to this theorem, the normal distribution is universal in the sense that the
distribution of the sum of a large number of random variables or random
vectors, subject to some not very restrictive conditions, is closely approximated by this distribution.
This is what provides a theoretical explanation of the "law of errors" of
applied statistics, which says that errors of measurement that result from
large numbers of independent "elementary" errors obey the normal distribution.
A multidimensional Gaussian distribution is specified by a small number
of parameters; this is a definite advantage in using it in the construction of
simple probabilistic models. Gaussian random variables have finite second
moments, and consequently they can be studied by Hilbert space methods.
Here it is important that in the Gaussian case "uncorrelated" is equivalent
to "independent," so that the results of L 2 theory can be significantly
strengthened.
2. Let us recall that (see 8) a random variable ~ = ~(m) is Gaussian, or
normally distributed, with parameters m and a 2 (~"' .K(m, a 2 )), lml < oo,
a 2 > 0, if its density f~(x) has the form
1"( ) _ _1_ (xm)2f2a2
x  ;;;= e
,
J~
....; 2na
(1)
+P.
where a =
As a! 0, the density f~(x) "converges to the afunction supported at
x = m." It is natural to say that ~ is normally distributed with mean m
and a 2 = 0 (~ "' .K(m, 0)) if~ has the property that P(~ = m) = 1.
We can, however, give a definition that applies both to the nondegenerate
(a 2 > 0) and the degenerate (a 2 = 0) cases. Let us consider the characteristic
function cp~(t) Eei 1 ~, t E R.
If P(~ = m) = 1, then evidently
eitm,
(2)
eitm(1/2)t2 a 2
(3)
cp~(t) =
whereas if~ "' .K(m, a 2 ), a 2 > 0,
cp~(t)
It is obvious that when a 2 = 0 the righthand sides of (2) and (3) are the
same. It follows, by Theorem 1 of 12, that the Gaussian random variable
with parameters m and a 2 (I m I < oo, a 2 ~ 0) must be the same as the random
variable whose characteristic function is given by (3). This is an illustration
of the "attraction of characteristic functions," a very useful technique in the
multidimensional case.
297
13. Gaussian Systems
Let
e= Ce1o ... , en) be a random vector and
lp~(t) =
t = (t 1 ,
Eei(t, ~.
t") E R",
(4)
its characteristic function (see Definition 2, 12).
Definidon 1. A random vector e= Ce 1, ... , en) is Gaussian, or normally
distributed, if its characteristic function has the form
lp~(t)
(5)
ei(t,m)(l/2)(1Rt,t),
where m = (m1o ... , mn), Imk I < oo and ~ = llrk1l is a symmetric nonnegative definite n X n matrix; we use the abbreviation
JV(m, ~).
e"'
This definition immediately makes us ask whether (5) is in fact a characteristic function. Let us show that it is.
First suppose that ~ is nonsingular. Then we can define the inverse
A = ~ 1 and the function
IAI 112
(6)
f(x) = (2n)"' 2 exp{ !(A(x m), (x m))},
where x = (x 1 ,
us show that
.. ,
xn) and
IAI = det A. This function is nonnegative. Let
r ei(t,x)f(x) dx =
JR"
ei(t,m)(l/2)(1Rt,t),
or equivalently that
I
=
n 
r ei(t,xm) (2n)nf2
IA1
JR"
112
e(1/2)(A(xm),(xm))
dx
= e(lf2)(1Rt,t)
(7)
Let us make the change of variable
x m
(!)u,
t = (!)v,
where (!) is an orthogonal matrix such that
and
is a diagonal matrix with d; ~ 0 (see the proof of the lemma in 8). Since
I~I = det ~ =F 0, we have d; > 0, i = 1, ... , n. Therefore
(8)
298
II. Mathematical Foundations of Probability Theory
Moreover (for notation, see Subsection 1, 12)
i(t, x  m) i(A(x  m), x  m)) = i(@v, (9u)  i(A(!)u, (9u)
= i((9v)T(9u i((9u?A((9u)
= ivTu iuT(9TA(9u
= ivTu iuTD 1u.
Together with (8) and (12.9), this yields
In= (2n)n 12 (d1 ... dn) 112
=
Il (2ndk)
k=l
112
JR"
exp(ivTu tuTD 1u) du
Joo exp(ivk uk oo
2udl ) duk =
k
Il exp( ivl dk)
k=l
= exp( ivTDv) = exp( tvT(9T[R(9v) = exp( itT!Rt) = exp( i(IRt, t)).
It also follows from (6) that
r f(x) dx =
JR"
1.
(9)
Therefore (5) is the characteristic function of a nondegenerate ndimensional Gaussian distribution (see Subsection 3, 3).
Now let IR be singular. Take 8 > 0 and consider the positive definite
IR + 8E. Then by what has been proved,
symmetric matrix IR'
<p'(t) = exp{i(t, m)  !(IR't, t)}
is a characteristic function:
<p'(t) = (
JR"
ei<t,xl
dF.(x),
where F.(x) = F.(xb ... , xn) is anndimensional distribution function.
As 8+ 0,
<p"(t)+ <p(t) = exp{i(t, m) i(IR't, t)}.
The limit function <p(t) is continuous at (0, ... , 0). Hence, by Theorem 1
and Problem 1 of 3 of Chapter III, it is a characteristic function.
We have therefore established Theorem 1.
3. Let us now discuss the significance of the vector m and the matrix
IR = llrk1il that appear in (5).
Since
299
13. Gaussian Systems
we find from (12.35) and the formulas that connect the moments and the
semiinvariants that
m1 
0 Ol sU
~

E;::':.1,
mk 
s<o.
0 0 ~
E;::':.k.
Similarly
r 11
v;::':.1
s<2,0, ... ,0)~
and generally
Consequently m is the meanvalue vector of ~ and IR is its covariance
matrix.
If IR is nonsingular, we can obtain this result in a different way. In fact,
in this case~ has a density f(x) given by (6).
A direct calculation shows that
E~k = Jxd(x) dx =
cov(~k, ~~) =
(11)
mk,
J<xk mk)(x 1  m1)f(x) dx
rkl
4. Let us discuss some properties of Gaussian vectors.
Theorem 1
if and only if they
are independent.
(b) A vector ~ = (~ 1 , .. , ~n) is Gaussian if and only if, for every vector
A.= (A.b ... , A.n), A.keR, the random variable(~, A.)= A. 1 ~ 1 + + A.n~n
has a Gaussian distribution.
(a) The components of a Gaussian vector are uncorrelated
PRooF. (a) If the components of~= (~ 1 , ... , ~n) are uncorrelated, it follows
from the form of the characteristic function cp~(t) that it is a product of
characteristic functions. Therefore, by Theorem 4 of 12, the components
are independent.
The converse is evident, since independence always implies lack of correlation.
(b) If~ is a Gaussian vector, it follows from (5) that
E exp{it(e1A.1
+ ... + enA.n)}
and consequently
exp{it(LA.kmk)
(Lrk,A.kA.l)},
tER,
300
II. Mathematical Foundations of Probability Theory
Conversely, to say that the random variable
is Gaussian means, in particular, that
(~,A.)
= ~ 1 A. 1
+ + ~nAn
Since A._t. ... , An are arbitrary it follows from Definition 1 that the vector
= (~ 1 , , ~n) is Gaussian.
This completes the proof of the theorem.
Remark. Let (0,
~) be a Gaussian vector with (} = (Ot. ... , (}k) and ~ =
If (} and ~ are uncorrelated, i.e. cov(O;, ~j) = 0, i = 1, ... , k;
j = 1, ... , l, they are independent.
(~ 1 , .. , ~k).
The proof is the same as for conclusion (a) of the theorem.
Let ~ = (~ 1 , .. , ~n) be a Gaussian vector; let us suppose, for simplicity,
that its meanvalue vector is zero. If rank ~ = r < n, then (as was shown in
11), there are n r linear relations connecting ~ 1 , , ~n We may then
suppose that, say, ~ 1 , ... , ~r are linearly independent, and the others can
be expressed linearly in terms of them. Hence all the basic properties of the
vector~= ~ 1 , ... , ~n are determined by the first r components (~ 1 , ... , ~,)
for which the corresponding covariance matrix is already known to be
nonsingular.
Thus we may suppose that the original vector~ = ( ~ 1, .. , ~n) had linearly
independent components and therefore that I ~I > 0.
Let (!) be an orthogonal matrix that diagonalizes ~.
(f)T~(f)
=D.
The diagonal elements of D are positive and therefore determine the inverse
matrix. Put B 2 = D and
Then it is easily verified that
i.e. the vector P= (p 1, ... , Pn) is a Gaussian vector with components that are
uncorrelated and therefore (Theorem 1) independent. Then if we write
A = (f)B we find that the original Gaussian vector ~ = (~ 1 , , ~n) can be
represented as
~ =
Ap,
(12)
where P = (p 1, , Pn) is a Gaussian vector with independent components,
pk "'..(0, 1). Hence we have the following result. Let~ = (e 1 , , ~n) be a
301
13. Gaussian Systems
vector with linearly independent components such that
E~k =
0, k = 1,
... , n. This vector is Gaussian if and only if there are independent Gaussian
variables {3 1 , , f:Jn, f:Jk "' %(0, 1), and a nonsingular matrix A of order n
such that~ = Af:J. Here IR = AAT is the covariance matrix of~.
If IIRI =F 0, then by the GramSchmidt method (see 11)
k = 1, ... , n,
where since 1: =
(~: 1 ,
(13)
... , ~:k) "' %(0, E) is a Gaussian vector,
~k =
k1
l1
(~k> Et)El,
(14)
(15)
and
(16)
2'{~1 , ~k} = 2'{1:1, , Ek}.
We see immediately from the orthogonal decomposition (13) that
~k = E(~k~~k1 .,~d.
(17)
From this, with (16) and (14), it follows that in the Gaussian case the conditional expectation E(~k I~k1> ... , ~ 1 ) is a linear function of (~1> ... , ~k 1 ):
k= 1
E(~k~~k1
~1) =
i= 1
a;~;
(18)
(This was proved in 8 for the case k = 2.)
Since, according to a remark made in Theorem 1 of 8, E(~k I~k 1 , ... , ~ 1 )
is an optimal estimator (in the meansquare sense) for ~k in terms of
~ 1 , . , ~k 1 , it follows from (18) that in the Gaussian case the optimal
estimator is linear.
We shall use these results in looking for optimal estimators of()= ( () 1, .. , ()k)
in terms of~ = (~ 1 , , ~ 1) under the hypothesis that((),~) is Gaussian. Let
m6
= E(J,
m~
= E~
be the columnvector mean values and
V66 =cov(O, 0) = llcov(O;, 0)11,
1 ~ i,j
V11~ = cov(O, ~) = llcov(O;, ~)II,
1 ~ i ~ k, 1 ~ j
V~~ = cov(~, ~) = llcov(~;. ~i) II,
1 ~ i,j
k,
~ l,
the covariance matrices. Let us suppose that V~~has an inverse. Then we have
the following theorem.
Theorem 2 (Theorem on Normal Correlation). For a Gaussian vector (0,
the optimal estimator E(O I~) of 0 in terms of~. and its error matrix
L\ = E[0  E(O I~)] [0  E(O( ~)]T
~),
302
II. Mathematical Foundations of Probability Theory
are given by the formulas
E(OI~) = m6
+ V 6 ~ V"i/(~
m~),
ll = V 6o Vo~V~ 1 (Vo~)T.
(19)
(20)
PRooF. Form the vector
11 = (0 m8 )  V 6 ~ V~ 1 {~ m~).
(21)
We can verify at once that Ert(~  m~)T = 0, i.e. 11 is not correlated with
(~  m~). But since (0, ~)is Gaussian, the vector (I'/,~) is also Gaussian. Hence
by the remark on Theorem 1, 11 and ~  m~ are independent. Therefore 11 and
~ are independent, and consequently E('11 ~) = Ert = 0. Therefore
E[O m6 1~] V 6 ~V~ 1 (~ m~) = 0.
which establishes (19).
To establish (20) we consider the conditional covariance
cov(O, 01~)
= E[(O E(OI~))(O E(OI~WieJ.
Since 0 E(Oie) = 17, and 11 and
(22)
eare independent, we find that
cov(O, Ole)= E(rtrtTie) = E11rtT
= V9 6
= V 68
+ Vi/V~~V~ 1 v:~ 2V 6 ~V~ 1 V~~V~ 1 v:~

V 6 ~ V~ 1 V:~.
Since cov(O, Ole) does not depend on "chance," we have
ll. = Ecov(O, Ole) =cov(O, Ole),
and this establishes (20).
Corollary. Let (0, e1, ... , en) be an (n
~> ... , ~n independent. Then
+ 1)dimensional Gaussian vector, with
(cf. (8.12) and (8.13)).
5. Let e 1 , e 2 , be a sequence of Gaussian random vectors that converge
in probability to e. Let us show that is also Gaussian.
In accordance with (a) of Theorem 1, it is enough to establish this only for
random variables.
303
13. Gaussian Systems
Let mn =
theorem
Een, u;
ven
Then by Lebesgue's dominated convergence
n>oo
n+ co
It follows from the existence of the limit on the lefthand side that there are
numbers m and u 2 such that
m =lim mn,
n>oo
Consequently
i.e.
e"' JV(m, u
2 ).
e2, .. .}
It follows, in particular, that the closed linear manifold 2(e1,
generated by the Gaussian variables 1 , 2 , .. (see 11, Subsection 5) consists
of Gaussian variables.
ee
6. We now turn to the concept of Gaussian systems in general.
Definition 2. A collection of random variables e= <e~). where IX belongs to
some index set m, is a Gaussian system if the random vector (e~,, ... , e~") is
Gaussian for every n ~ 1 and all indices IX1, . ' CX.n chosen from m.
Let us notice some properties of Gaussian systems.
(a) If
e=
IX' E
(b) If
(c)
<e~), IX Em,
is a Gaussian system, then every subsystem
m' s; m, is also Gaussian.
e~, IX Em,
e= <eiX),
e' =
<e~.),
are independent Gaussian variables, then the system
IX Em, is Gaussian.
(~~), oc Em, is a Gaussian
=
system, the closed linear manifold Y(e),
consisting of all variables of the form
1 ca,ea,, together with their
meansquare limits, forms a Gaussian system.
If~
Li=
Let us observe that the converse of (a) is false in general. For example,
let e1 and 17 1 be independent and e1"'%(0, 1), 17 1"' %(0, 1). Define the
system
<e ) =
11
{<eb 1111D
<e1, 11111)
if e1 ~
if e1 <
o,
o.
(23)
Then it is easily verified that and 11 are both Gaussian, but (e, 17) is not.
Let = (ea)aelll be a Gaussian system with meanvalue vector m = (ma),
oc Em, and covariance matrix IR = (r~p)a,fJelll where ma = Eea Then IR is
evidently symmetric (r rzfJ = r pa) and nonnegative definite in the sense that
for every vector c = (ca)ael!l with values in R 111, and only a finite number of
nonzero coordinates ca,
(~c, c)
=L rrzpCaCp ~ 0.
tx,{J
(24)
304
II. Mathematical Foundations of Probability Theory
We now ask the converse question. Suppose that we are given a parameter
set m: = {ex}, a vector m = (m11) 11 e!ll and a symmetric nonnegative definite
matrix IR = (r11p)11,pe!ll Do there exist a probability space (0, F. P) and a
Gaussian system of random variables~= (~ 11)11 e 111 on it, such that
E~~~ = m~~,
cov(~ 11 , ~ 11)
r<Z,fl
ex, pem:?
If we take a finite set ex 1 , , exn, then for the vector ffi = (m~~., ... , m~~J
and the matrix IR = (r1111), ex, P= ex 1, , exn, we can construct in Rn the
Gaussian distribution F 111 , ... , 11"(x 1 , .. , xn) with characteristic function
qJ(t) = exp{i(t, m)  !(IRt, t)},
(t .. t.,
'
t..J.
It is easily verified that the family
{F1Zt, ... ,1Zn(X1, ... ' Xn); (Xi Em:}
is consistent. Consequently by Kolmogorov's theorem (Theorem 1, 9,
and Remark 2 on this) the answer to our question is positive.
7. If m: = {1, 2, ... }, then in accordance with the terminology of 5 the
system of random variables~= (~ 11)11 e!IJ is a random sequence and is denoted
by ~ = (~ 1 , ~ 2 , ...). A Gaussian sequence is completely described by its
meanvalue vector m = (m 1 , m2 , ) and covariance matrix IR = llriill,
rii = cov(~i ~i). In particular, if rii = afbii then ~ = (~ 1 , ~ 2 , ) is a
Gaussian sequence of independent random variables with ~i "' .!V(mi> af),
i ~ 1.
When m: = [0, 1], [0, 00 ), ( 00, 00 ), ... , the system ~ = (e,), t Em:, is a
random process with continuous time.
Let us mention some examples of Gaussian random processes. If we take
their mean values to be zero, their probabilistic properties are completely
described by the covariance matrices llr.,ll We write r(s, t) instead of r.,
and call it the covariance function.
EXAMPLE
1. If T = [0, 00) and
r(s, t) = min(s, t),
(25)
the Gaussian process = (~,)r~O with this covariance function (see Problem
2) and ~ 0 = 0 is a Brownian motion or Wiener process.
Observe that this process has independent increments; that is, for arbitrary
t 1 < t 2 < < tn the random variables
~12  elt' ' ' ~In  ~In I
are independent. In fact, because the process is Gaussian it is enough to
verify only that the increments are uncorrelated. But if s < t < u < v then
E[~ 1
~.] [~v
~..] =
[r(t, v)  r(t, u~]  [r(s, v)  r(s, u)]
= (t  t)  (s  s) = 0.
305
13. Gaussian Systems
EXAMPLE
2. The process
e= (e
1),
0 ::; t ::; 1, with
eo := 0 and
r(s, t) = min(s, t)  st
(26)
is a conditional Wiener process (observe that since r(1, 1) = 0 we have
= o) = 1).
P<el
3. The process
EXAMPLE
e= (e
1), 
oo < t <
00,
with
r(s, t) = elt1
(27)
is a GaussMarkov process.
8.
PROBLEMS
1. Let
e1,e2, e3be independent Gaussian random variables, e; ~ .(0, 1). Show that
e1 + e2e3 ~ Y(o, 1).
Jt + e~
(In this case we encounter the interesting problem of describing the nonlinear
transformations of independent Gaussian variables 1 , .. ,
whose distributions
are still Gaussian.)
e.
2. Show that (25), (26) and (27) are nonnegative definite (and consequently are actually
covariance functions).
3. Let A be an m x n matrix. An n x m matrix A E9 is a pseudo inverse of A if there are
matrices U and V such that
Show that A E9 exists and is unique.
4. Show that (19) and (20) in the theorem on normal correlation remains valid when
v~~ is singular provided that v;/ is replaced by v~.
e)= (ll1o ... , llk; e1,... , e,) be a Gaussian vector with nonsingular matrix
5. Let (ll,
.1 = V1111
V~ v:~. Show that the distribution function
P(ll::; ale)= P(01::; a1, ... , llk::; akle)
has (Pa.s.) the density p(al, ... ' ak Ie) defined by
1.11/21
(2n)k12 exp{ t(a E(OI~W.1 1 (a E(OI~))}.
6. (S. N. Bernstein). Let~ and '1 be independent identically distributed random variables
with finite variances. Show that if + '7 and ~  '1 are independent, then and '1
are Gaussian.
CHAPTER III
Convergence of Probability Measures.
Central Limit Theorem
1. Weak Convergence of Probability Measures and
Distributions
1. Many important results of probability theory are formulated as limit
theorems. So, indeed, were James Bernoulli's law of large numbers, as well
as the De MoivreLaplace limit theorem, the theorems with which the true
theory of probability began.
In the present chapter we discuss two central aspects of limit theorems:
one is the concept of weak convergence; the other is the method of characteristic functions, one of the most powerful methods for proving and refining
limit theorems.
We begin by recalling the statement of the law of large numbers (Chapter
I, 5) for the Bernoulli scheme.
Let ~ 1 , ~ 2 , be a sequence of independent identically distributed
random variables with P(~i = 1) = p, P(~i = 0) = q, p + q = 1. In terms
of the concept of convergence in probability (Chapter II, 10), Bernoulli's
law of large numbers can be stated as follows:
n+ oo,
(1)
where Sn = ~ 1 + + ~n (It will be shown in Chapter IV that in fact we
have convergence with probability 1.)
We put
(2)
307
I. Weak Convergence of Probability Measures and Distributions
where F(x) is the distribution function of the degenerate random variable
~
p. Also let P nand P be the probability measures on (R, Pl(R)) corresponding to the distributions F n and F.
In accordance with Theorem 2 of 1 0, Chapter II, convergence in probability, Sn/n ~ p, implies convergence in distribution, Sn/n .!4 p, which means that
Ef(~) + Ef(p),
for every function!
ous functions on R.
Since
E~~) =
n+ oo,
(3)
f(x) belonging to the class C(R) of bounded continu
f(x)P.(dx),
(3) can be written in the form
L
L
f(x)Pn(dx)+
Ef(p)
L
L
f(x)P(dx),
{f(x)P(dx),
fE C(R),
(4)
or (in accordance with 6 of Chapter II) in the form
f(x) dF n(x) +
f(x) dF(x ),
C(R).
(5)
In analysis, (4) is called weak convergence (of Pn to P, n + oo) and written
Pn ~ P. It is also natural to call ( 5) weak convergence ofF n to F and denote
it by Fn ~F.
Thus we may say that in a Bernoulli scheme
sn
+ p => F"+ F.
(6)
It is also easy to see from (1) that, for the distribution functions defined
in (2),
F.(x)+ F(x),
n+ oo,
for all points x E R except for the single point x = p, where F(x) has a discontinuity.
This shows that weak convergence F"+ F does not imply pointwise
convergence of F"(x) to F(x), n+ oo, for all points x E R. However, it turns
out that, both for Bernoulli schemes and for arbitrary distribution functions,
weak convergence is equivalent (see Theorem 2 below) to "convergence
in general" in the sense of the following definition.
Definition 1. A sequence of distribution functions {F"}, defined on the real
line, converges in general to the distribution function F (notation: F" => F)
if as n+ oo
Fn(x)+ F(x),
xEPc(F),
where P c(F) is the set of points of continuity ofF = F(x).
308
III. Convergence of Probability Measures. Central Limit Theorem
For Bernoulli schemes, F = F(x) is degenerate, and it is easy to see
(see Problem 7 of 10, Chapter II) that
CFn=F)=(~~P)
Therefore, taking account of Theorem 2 below,
(~ ~ P) =(F n.!!+ F)<=> (F n=F)= (~ E. p)
(7)
and consequently the law of large numbers can be considered as a theorem
on the weak convergence of the distribution functions defined in (2).
Let us write
Fix)= P{
F(x)
Sn np
jnpq : :; x } ,
= 1 fx
fo
eu 212 du.
(8)
oo
The De MoivreLaplace theorem (6, Chapter I) states that Fn(x)+ F(x)
for all x E R, and consequently Fn =F. Since, as we have observed, weak
convergence Fn ~ F and convergence in general, Fn = F, are equivalent,
we may therefore say that the De MoivreLaplace theorem is also a theorem
on the weak convergence of the distribution functions defined by (8).
These examples justify the concept of weak convergence of probability
measures that will be introduced below in Definition 2. Although, on the
real line, weak convergence is equivalent to convergence in general of the
corresponding distribution functions, it is preferable to use weak convergence
from the beginning. This is because in the first place it is easier to work with,
and in the second place it remains useful in more general spaces than the
real line, and in particular for metric spaces, including the especially important spaces R", R 00 , C and D (see 3 of Chapter II).
2. Let (E, tff, p) be a metric space with metric p = p(x, y) and ualgebra tff
of Borel subsets generated by the open sets, and let P, Pt> P2 , ... be probability measures on (E, tff, p).
Definition 2. A sequence of probability measures {P n} converges weakly to the
probability measure P (notation: P n ~ P) if
f(x)P n(dx)
+
J(x)P(dx)
(9)
for every function f = f(x) in the class IC(E) of continuous bounded functions on E.
I. Weak Convergence of Probability Measures and Distributions
309
Definition 3. A sequence of probability measures {P n} converges in general
to the probability measure P (notation: Pn => P) if
(10)
for every set A of G for which
P(oA)
= 0.
(11)
(Here oA denotes the boundary of A:
oA
[A] n [A],
where [A] is the closure of A.)
The following fundamental theorem shows the eqmvalence of the concepts of weak convergence and convergence in general for probability
measures, and contains still another equivalent statement.
Theorem 1. The following statements are equivalent.
(I) Pn ~ P.
(II) lim Pn(A) ::::;; P(A), A closed.
(Ill) lim Pn(A) 2': P(A), A open.
(IV) Pn=>P.
PRooF. (I)=> (II). Let A be closed, f(x)
fix)=
IA(x) and
g(~ p(x, A)),
s > 0,
where
p(x, A)
= inf{p(x, y): yEA},
1,
g(t) = { 1  t,
0,
t::::;; 0,
0 ::::;; t ::::;; 1,
t 2': 1.
Let us also put
A,= {x: p(x, A) < s}
and observe that A, t A as s t 0.
Since.fe(x) is bounded, continuous, and satisfies
we have
which establishes the required implication.
310
III. Convergence of Probability Measures. Central Limit Theorem
The implications (II) => (Ill) and (Ill) => (II) become obvious if we take
the complements of the sets concerned.
(III)=> (IV). Let A 0 = A\oA be the interior, and [A] the closure, of A.
Then from (II), (III), and the hypothesis P(ilA) = 0, we have
lim Pn(A)
~lim
n
Pn([A])
P([A]) = P(A),
and therefore Pn(A) ~ P(A) for every A such that P(ilA) = 0.
(IV)~ (1). Letf = f(x) be a bounded continuous function with lf(x)l
M. We put
= {tER: P{x:f(x) = t} =I= 0}
and consider a decomposition 7k = (t 0 , t 1 , .. , tk) of [ M, M]:
D
 M =
t0
<
t1
< <
tk =
M,
~ 1,
with ti : D, i = 0, 1, ... , k. (Observe that D is at most countable since the
sets f 1 { t} are disjoint and P is finite.)
Let Bi = {x: t; ~ f(x) < t;+ d. Since f(x) is continuous and therefore
the set f 1(t;, ti+ 1 ) is open, we have oB; f 1 {t;} u f 1 {ti+ 1 }. The points
t;, t;+ 1 D; therefore P(ilB;) = 0 and, by (IV),
k1
k1
t;Pn(B;) ~
i=O
But
Lf(x)Pn(dx) Lf(x)P(dx)
I~
:t~ t;Pn(B;)
+ I:t~ t; Pn(B;) 
:t~ t; P(B;) I
Lf(x)Pn(dx)
t;P(B;) Lf(x)P(dx)
2 max
(t;+ 1
O:s;i:s;k1
whence, by (12), since the
7k (k
1) are arbitrary,
lim ( f(x)Pn(dx) = ( f(x)P(dx).
n
JE
+I :t~
~
(12)
t;P(B;).
t=O
JE
This completes the proof of the theorem.
t;)
1. Weak Convergence of Probability Measures and Distributions
311
Remark 1. The functions f(x) = IA(x) and fe(x) that appear in the proof
that (I) => (II) are respectively upper semicontinuous and uniformly continuous.
Hence it is easy to show that each of the conditions of the theorem is equivalent to one of the following:
(V) JEf(x)P.(x) dx+ JEf(x)P(dx) for all bounded uniformly continuous
f(x);
(VI) lim JE f(x)P .(dx) ::s;; JE f(x)P(dx) for all bounded f(x) that are upper
semicontinuous (Iimf(x.) ::s;; f(x), x.+ x);
(VII) lim JEf(x)P.(dx) ~ JEf(x)P(dx) for all bounded f(x) that are lower
semicontinuous (lim f(x.) ~ f(x), x.+ x).
n
Remark 2.Theorem 1 admits a natural generalization to the case when the
probability measures P and P. defined on (E, tS, p) are replaced by arbitrary
(not necessarily probability) finite measures J.L and J.L . For such measures
we can introduce weak convergence J.l.n ~ J.L and convergence in general
f.Ln => J.1. and, just as in Theorem 1, we can establish the equivalence of the
following conditions:
(I*)
(II*)
(III*)
(IV*)
J.l.n ~ J.L;
lim J.L.(A) ::s;; J.L(A), where A is closed and J.LiE) + J.L(E);
lim J.L.(A) ~ J.L(A), where A is open and J.L.(E)+ J.L(E);
J.l.n => J.L.
Each of these is equivalent to any of (V*), (VI*), and (VII*), which are
(V), (VI), and (VII) with P" and P replaced by J.l.n and J.L.
3. Let (R, BI(R)) be the real line with the system BI(R) of sets generated by
the Euclidean metric p(x, y) = lx  yl (compare Remark 2 of subsection 2
of2 of Chapter II). Let P and P "' n ;:::: 1, be probability measures on (R, aJ(R))
and let F and F., n ~ 1, be the corresponding distribution functions.
Theorem 2. The following conditions are equivalent:
(1)
(2)
(3)
(4)
P. ~ P,
P. => P,
F.~ F,
F.=> F.
PRooF. Since (2) <=> (1) <=> (3), it is enough to show that (2) <=> (4).
If P. => P, then in particular
P.( oo, x]+ P( oo, x]
for all x e R such that P{x} = 0. But this means that F.=> F.
Now let F.=> F. To prove that P. =>Pit is enough (by Theorem 1) to
show that lim. P.(A) ~ P(A) for every open set A.
If A is open, there is a countable collection of disjoint open intervals
11 , 12 , . (of the form (a, b)) such that A= 2,1:': 1 /k. Choose e > 0 and in
312
III. Convergence of Probability Measures. Central Limit Theorem
each interval Ik = (ak> bk) select a subinterval Ik = (ak, bk] such that ak,
b~ E Pc(F) and P(Ik) :::; P(J~) + e rk. (Since F(x) has at most countably
many discontinuities, such intervals Ik, k 2:: 1, certainly exist.) By Fatou's
lemma,
00
00
lim Pn(A) =lim
n
Pn(Ik) 2::
k=1
k=1
lim Pn{/k)
n
00
2::
lim Pn(Ik).
k= 1
But
Therefore
00
00
lim Pn(A) 2::

P(Ik) 2::
k= 1
(P(h)  e 2k) = P(A) e.
Since e > 0 is arbitrary, this shows that limn Pn(A) 2:: P(A) if A is open.
This completes the proof of the theorem.
4. Let (, S) be a measurable space. A collection X 0 () <::; S of subsets is
a determining class if whenever two probability measures P and 0 on (E, S)
satisfy
P(A)
= Q(A)
for all
A EX 0 ()
it follows that the measures are identical, i.e.
P(A) = Q(A)
for all
S.
If(, S, p) is a metric space, a collection X 1(E) <::; Iff is a convergencedetermining class if whenever probability measures P, P 1 , P 2 , ... satisfy
P n(A)
t
P(A)
for all
A EX 1 ()
with
P(oA) = 0
it follows that
for all A E E with
P(oA) = 0.
When (E, S) = (R, ~(R)), we can take a determining class X 0 (R) to be
the class of "elementary" sets X = {( oo, x], x E R} (Theorem 1, 3,
Chapter II). It follows from the equivalence of (2) and (4) of Theorem 2
that this class X is also a convergencedetermining class.
It is natural to ask about such determining classes in more general spaces.
For W, n 2:: 2, the class X of "elementary" sets of the form ( oo, x] =
( oo, x 1 ] x x ( oo, Xn], where x = (x 1, ... , xn) ERn, is both a determining class (Theorem 2, 3, Chapter II) and a convergencedetermining
class (Problem 2).
313
I. Weak Convergence of Probability Measures and Distributions
ljn
2/n
Figure 35
For Roo the cylindrical sets % 0 (R 00 ) are the "elementary" sets whose
probabilities uniquely determine the probabilities of the Borel sets (Theorem
3, 3, Chapter II). It turns out that in this case the class of cylindrical sets is
also the class of convergencedetermining sets (Problem 3). Therefore
%,(Roo)= %o(Roo).
We might expect that the cylindrical sets would still constitute determining classes in more general spaces. However, this is, in general, not the
case.
For example, consider the space (C, 140 (C), p) with the uniform metric
p (see subsection 6, 2, Chapter II). Let P be the probability measure concentrated on the element x = x(t) 0, 0 ~ t s 1, and let P., n;;::: 1, be the
probability measures each of which is concentrated on the element x
= x.(t)
shown in Figure 35. It is easy to see that P.(A)+ P(A) for all cylindrical
sets A with P(oA) = 0. But if we consider, for example, the set
A = { oc E C: loc(t) I s
t, 0 s t s
1} E 140( C),
then P(oA) = 0, P.(A) = 0, P(A) = 1 and consequently P.
P.
Therefore % 0 (C) = 140 (C) but % 0 (C) c % 1 (C) (with strict inclusion).
5.
PROBLEMS
1. Let us say that a function F = F(x), defined on R", is continuous at x E R" provided
that, for every e > 0, there is a~ > 0 such that IF(x)  F(y) I < e for ally E R" that
satisfy
x  be < y < x
+ ~e,
where e = (1, ... , 1) E R". Let us say that a sequence of distribution functions {F.}
converges in general to the distribution function F (F.=> F) if F.(x)+ F(x), for all
points x E R" where F = F(x) is continuous.
Show that the conclusion of Theorem 2 remains valid for R", n > 1. (See the remark
on Theorem 2.)
314
III. Convergence of Probability Measures. Central Limit Theorem
2. Show that the class :K of"elementary" sets in R" is a convergencedetermining class.
3. Let E be one ofthe spaces R"", Cor D. Let us say that a sequence {P.} of probability
measures (defined on the ualgebra 8 of Borel sets generated by the open sets) converges in general in the sense of finitedimensional distributions to the probability
measure P (notation: P. b P) if P .(A) + P(A), n + oo, for all cylindrical sets A with
P(oA) = 0.
For R"", show that
(P. b P) <>(P. => P).
4. Let F and G be distribution functions on the real line and let
L(F, G)= inf{h > 0: F(x h) h
G(x)
F(x +h)+ h}
be the Levy distance (between F and G). Show that convergence in general is equivalent to convergence in the Levy metric:
(F =>F)<> L(F F)+ 0.
5. Let F. => F and let F be continuous. Show that in this case F .(x) converges uniformly
to F(x):
sup IF.(x) F(x)l+ 0,
n+ oo.
6. Prove the statement in Remark 1 on Theorem 1.
7. Establish the equivalence of (1*)(IV*) as stated in Remark 2 on Theorem 1.
8. Show that P. ~ P if and only if every subsequence {P.} of {P.} contains a subsequence {P... } such that P... ~ P.
2. Relative Compactness and Tightness of Families
of Probability Distributions
1. If we are given a sequence of probability measures, then before we can
consider the question of its (weak) convergence to some probability measure,
we have of course to establish whether the sequence converges in general
to some measure, or has at least one convergent subsequence.
For example, the sequence {P.}, where P 2 = P, P 2 .+ 1 = Q, and P and Q
are different probability measures, is evidently not convergent, but has the
two convergent subsequences {P 2 ,} and {P 2 ,+ 1 }
It is easy to construct a sequence {P.} of probability measures P., n ~ 1,
that not only fails to converge, but contains no convergent subsequences at
all. All that we have to do is to take P., n ~ 1, to be concentrated at {n} (that
is, P.{n} = 1). In fact, since lim. P.(a, b] = Owhenevera < b, a limit measure
would have to be identically zero, contradicting the fact that 1 = P,(R) 0,
2. Relative Compactness and Tightness of Families of Probability Distributions
315
n+ oo. It is interesting to observe that in this example the corresponding
sequence {F"} of distribution functions,
Fn{x)
1,
x;::: n,
= { O, x < n,
is evidently convergent: for every x
R,
Fn{x)+ G(x)
=0.
However, the limit function G = G(x) is not a distribution function (in the
sense of Definition 1 of 3, Chapter II).
This instructive example shows that the space of distribution functions is
not compact. It also shows that if a sequence of distribution functions is to
converge to a limit that is also a distribution function, we must have some
hypothesis that will prevent mass from "escaping to infinity."
After these introductory remarks, which illustrate the kinds of difficulty
that can arise, we turn to the basic definitions.
2. Let us suppose that all measures are defined on the metric space (E, tff, p).
Definition 1. A family of probability measures~= {P"; a Em} is relatively
compact if every sequence of measures from ~contains a subsequence which
converges weakly to a probability measure.
We emphasize that in this definition the limit measure is to be a probability
measure, although it need not belong to the original class ~ (This is why the
word "relatively" appears in the definition.)
It is often far from simple to verify that a given family of probability
measures is relatively compact. Consequently it is desirable to have simple
and useable tests for this property. We need the following definitions.
Definition 2. A family of probability measures ~ = {Pa; rx Em} is tight if,
for every e > 0, there is a compact set K <;; E such that
supPa{E\K):::; e.
(1)
ae'll
Definition 3. A family of distribution functions F = {Fa; rx Em} defined on
R", n ;::: 1, is relatively compact (or tight) if the same property is possessed by
the family~= {P"; rx Em} of probability measures, where P"' is the measure
constructed from F"'.
3. The following result is fundamental for the study of weak convergence of
probability measures.
Theorem 1 (Prohorov's Theorem). Let ~ = {P"'; rx Em} be a family of
probability measures defined on a complete separable metric space (E, iff, p).
Then~ is relatively compact if and only if it is tight.
316
III. Convergence of Probability Measures. Central Limit Theorem
We shall give the proof only when the space is the real line. (The proof can
be carried over, almost unchanged, to arbitrary Euclidean spaces Rn, n ~ 2.
Then the theorem can be extended successively to R 00 , to acompact spaces;
and finally to general complete separable metric spaces, by reducing each
case to the preceding one.)
Necessity. Let the family r!J> = {Pa: !'f. E m:} of probability measures defined
on (R, 86(R)) be relatively compact but not tight. Then there is an e > 0 such
that for every compact K ~ R
sup Pa(R\K) > e,
and therefore, for each interval I = (a, b),
sup Pa{R\J) >e.
It follows that for every interval In= ( n, n), n
such that
1, there is a measure Pan
Pan(R\In) >e.
Since the original family r!J> is relatively compact, we can select from {PaJ n;;, 1
a subsequence {Pan) such that Pank ~ Q, where Q is a probability measure.
Then, by the equivalence of conditions (I) and (II) in Theorem 1 of 1, we
have
(2)
for every n ~ 1. But Q(R\In) t 0, n+ oo, and the left side of (2) exceeds
e > 0. This contradiction shows that relatively compact sets are tight.
To prove the sufficiency we need a general result (Helly's theorem) on the
sequential compactness of families of generalized distribution functions
(Subsection 2 of 3 of Chapter II).
Let J = { G} be the collection of generalized distribution functions
= G(x) that satisfy:
(1) G(x) is nondecreasing;
(2) 0 ~ G( oo ), G( + oo) ~ 1;
(3) G(x) is continuous on the right.
Then J clearly contains the class of distribution functions ' = {F}
for which F( oo) = 0 and F( + oo) = 1.
Theorem 2 (Helly's Theorem). The class J = {G} of generalized distribution
fimctions is sequentially compact, i.e. for every sequence {Gn} of functions
from J we can .find afunction G E Janda sequence {nk} ~ {n} such that
k+
00,
for every point x of the set P c( G) of points of continuity of G
= G(x ).
2. Relative Compactness and Tightness of Families of Probability Distributions
317
PROOF. Let T = {x 1 , x 2 , ... } be a countable dense subset of R. Since the
sequence of numbers {G.{x 1 }} is bounded, there is a subsequence N 1 =
{nl1>, n~l)' .. .} such that G.p){x 1) approaches a limit g1 as i+ oo. Then we
extract from N 1 a subsequence N 2 = {nl2 >, n~2 >, .. } such that G"12)(x 2 )
approaches a limit g 2 as i + oo; and so on.
'
On the set T ~ R we can define a function Gr(x) by
XiE
T,
and consider the "Cantor" diagonal sequence N = {nl1>, n~2 >, .. .}. Then, for
each xi E T, as m+ oo, we have
G.~)(xi)
+ Gr(xJ
Finally, let us define G = G(x) for all x E R by putting
G(x) = inf{Gr(y): yET, y > x}.
(3)
We claim that G = G(x) is the required function and G.~,.~)(x)+ G(x) at all
points x of continuity of G.
Since all the functions G. under consideration are nondecreasing, we have
Gnl:rl(x) ~ G.~,.m)(y) for all x andy that belong to T and satisfy the inequality
x ~ y. Hence for such x and y,
Gr(x) ~ Gr(y).
It follows from this and (3) that G = G(x) is nondecreasing.
Now let us show that it is continuous on the right. Let xk ! x and d =
Iimk G(xk). Clearly G(x) ~ d, and we have to show that actually G(x) = d.
Suppose the contrary, that is, let G(x) < d. It follows from (3) that there is
ayE T, x < y, such that Gr(Y) < d. But x < xk < y for sufficiently large k,
and therefore G(xk) ~ Gr(Y) < d and lim G(xk) < d, which contradicts
d = Iimk G(xk). Thus we have constructed a function G that belongs to .f.
We now establish that G.<ml{x 0 )+ G(x 0 ) for every x 0 E Pc(G).
If x 0 < y E T, then
m
lim Gnl,.ml(x 0 ) ~ lim Gnl,.ml(y) = Gr(y),
m
whence
lim G.~:r){x 0 ) ~ inf{Gr(y): y > x 0 , yET}= G(x 0 ).
(4)
On the other hand, let x 1 < y < x 0 , y E T. Then
G(x 1 ) ~ Gr(Y) =lim Gnl:/'l(y) =lim Gnl:rl(y) ~lim Gnl:/'l(x 0 ).
m
Hence if we let x 1 j x 0 we find that
G(xo _) ~ li,!,D G.~:rlx 0 ).
(5)
318
III. Convergence of Probability Measures. Central Limit Theorem
But if G(x 0  ) = G(x 0 ) we can infer from (4) and (5) that Gnllr>(x 0 ) ~ G(x 0 ),
m~ oo.
This completes the proof of the theorem.
We can now complete the proof of Theorem 1.
Sufficiency. Let the family & be tight and let {P"} be a sequence of probability measures from&. Let {F"} be the corresponding sequence of distribution functions.
By Helly's theorem, there are a subsequence {F "k} ;; { F"}and a generalized
distribution function G e J such that F"k(x) ~ G(x) for x e Pc(G). Let us
show that because & was assumed tight, the function G = G(x) is in fact a
genuine distribution function (G( oo) = 0, G( + oo) = 1).
Take e > 0, and let I = (a, b] be the interval for which
sup Pn(R\1) < e,
or, equivalently,
n~l.
Choose points a', b' e Pc(G) such that a' <a, b' >b. Then 1  e::;;; P"k(a, b]
::;;; P"k(a', b'] = F"k(b') F"k(a') ~ G(b') G(a'). It follows that G( + oo)G(oo) = 1, and since 0::;;; G(oo)::;;; G(+oo)::;;; 1, we have G(oo) = 0
andG(+oo) = 1.
Therefore the limit function G = G(x) is a distribution function and
F"k =>G. Together with Theorem 2 of 1 this shows that Pnk ~ Q, where Q
is the probability measure corresponding to the distribution function G.
This completes the proof of Theorem 1.
4.
PROBLEMS
1. Carry out the proofs of Theorems 1 and 2 for R", n ;;:: 2.
2. Let P be a Gaussian measure on the real line, with parameters m. and cr;, ex E 21.
Show that the family f1J = {P; ex E 21} is tight if and only if
1m. I S: a,
cr; S: b,
ex E 21.
3. Construct examples of tight and nontight families f1J = {P.; ex E 21} of probability
measures defined on (R"", di(R"")).
3. Proofs of Limit Theorems by the Method of
Characteristic Functions
1. The proofs of the first limit theorems of probability theorythe law of
large numbers, and the De MoivreLaplace and Poisson theorems for
Bernoulli schemeswere based on direct analysis of the limit functions ofthe
3. Proofs of Limit Theorems by the Method of Characteristic Functions
319
distributions F n, which are expressed rather simply in terms of binomial
probabilities. (In the Bernoulli scheme, we are adding random variables that
take only two values, so that in principle we can find Fn explicitly.) However,
it is practically impossible to apply a similar direct method to the study of
more complicated random variables.
The first step in proving limit theorems for sums of arbitrarily distributed
random variables was taken by Chebyshev. The inequality that he discovered,
and which is now known as Chebyshev's inequality, not only makes it
possible to give an elementary proof of James Bernoulli's law oflarge numbers,
but also lets us establish very general conditions for this law to hold, when
stated in the form
n+ oo,
every c > 0,
(1)
for sums Sn = ~ 1 + + ~n n ~ 1, of independent random variables. (See
Problem 2.)
Furthermore, Chebyshev created (and Markov perfected) the "moment
method" which made it possible to show that the conclusion of the De
MoivreLaplace theorem, written in the form
 ESn
P{Sn 1\iCi"
y
Y..Jn
+
M
y2n
fx e u>;2 du,
(2)
co
is "universal," in the sense that it is valid under very general hypotheses
concerning the nature of the random variables. For this reason it is known as
the central limit theorem of probability theory.
Somewhat later Lyapunov proposed a different method for proving the
central limit theorem, based on the idea (which goes back to Laplace) of the
characteristic function of a probability distribution. Subsequent developments have shown that Lyapunov's method of characteristic functions is
extremely effective for proving the most diverse limit theorems. Consequently
it has been extensively developed and widely applied.
In essence, the method is as follows.
2. We already know (Chapter II, 12) that there is a onetoone correspondence between distribution functions and characteristic functions. Hence we
can study the properties of distribution functions by using the corresponding
characteristic functions. It is a fortunate circumstance that weak convergence
Fn ~ F of distributions is equivalent to pointwise convergence q>n+ q> of
the corresponding characteristic functions. Moreover, we have the following
result, which provides the basic method of proving theorems on weak convergence for distributions on the real line.
320
III. Convergence of Probability Measures. Central Limit Theorem
Theorem 1 (Continuity Theorem). Let {Fn} be a sequence of distribution
functions F n = FnCx ), x E R, and let {cpn} be the corresponding sequence of
characteristic jimctions,
IPn(t) = s:ooeirx dFn(x),
t E R.
(1) If Fn ~ F, where F = F(x) is a distribution function, then cpn(t) 4 cp(t),
t E R, where cp( t) is the characteristic junction ofF = F( x ).
(2) If limn cpit) exists for each t E R and cp(t) = limn cpn(t) is continuous at
t = 0, then cp(t) is the characteristic function of a probability distribution
F = F(x), and
The proof of conclusion (1) is an immediate consequence of the definition
of weak convergence, applied to the functions Re eirx and Im eirx.
The proof of (2) requires some preliminary propositions.
Lemma 1. Let {P n} be a tight family of probability measures. Suppose that
every weakly convergent subsequence {Pn.} of {Pn} converges to the same
probability measure P. Then the whole sequence {P n} converges to P.
PROOF. Suppose that Pn
f = f(x) such that
+ P. Then there is a bounded continuous function
It follows that there exist e > 0 and an infinite sequence {n'}
that
I Lf(x)Pn(dx)
Lf(x)P(dx) I
~ e > 0.
{n} such
(3)
By Prohorov's theorem (2) we can select a subsequence {P n"} of {P n'} such
that Pn" ~ 0, where 0 is a probability measure.
By the hypotheses of the lemma, 0 = P, and therefore
Lf(x)Pn ..(dx)
4
Lf(x)P(dx),
which leads to a contradiction with (3). This completes the proof of the
lemma.
Lemma 2. Let {Pn} be a tight family of probability measures on (R, 86(R)).
A necessary and stif.ficient condition for the sequence {Pn} to converge weakly
to a probability measure is that for each t E R the limit limn cpnCt) exists, where
IPn( t) is the characteristic function of Pn:
IPn(t) = L eirxp n(dx).
3. Proofs of Limit Theorems by the Method of Characteristic Functions
321
PRooF. If {P n} is tight, by Prohorov's theorem there are a subsequence
{Pn} and a probability measure P such that Pn' ~ P. Suppose that the whole
sequence {P n} does not converge toP (P n !!;\+ P). Then, by Lemma 1, there are a
subsequence {Pn"} and a probability measure Q such that Pn" ~ Q, and
p = Q.
Now we use the existence of limn cpn(t) for each t E R. Then
lim
n'
r eitxpn.(dx) = lim r eitxpn"(dx)
JR
n"
JR
and therefore
But the characteristic function determines the distribution uniquely
(Theorem 2, 12, Chapter II). Hence P = Q, which contradicts the assumption
that Pn ~ P.
The converse part of the lemma follows immediately from the definition
of weak convergence.
The following lemma estimates the "tails" of a distribution function in
terms of the behavior of its characteristic function in a neighborhood of
zero.
Lemma 3. Let F = F(x) be a distribution function on the real line and let
cp = cp(t) be its characteristic function. Then there is a constant K > 0 such
that for every a > 0
PROOF. Since Re cp(t)
! Ja[l a o
fa [1 
dF(x) :::;; K
Jlxl<e:lja
= J':'
00
Re cp(t)] dt.
(4)
cos tx dF(x), we find by Fubini's theorem that
Re cp(t)] dt
=! Ja[Joo (1 COS tx) dF(x)] dt
a o oo
00 [~ J:(l
= f_00
=
COS
Joo (
_ 00
sin ax) dF(x)
1 ~ax
. ( 1  sm
2:: mf
JyJ<e:l
=1
tx) dt] dF(x)
lxl<e:lfa
y) l
dF(x),
Jaxl<e:l
dF(x)
III. Convergence of Probability Measures. Central Limit Theorem
322
where
y)
sin = 1  sm 1 :?:
. f (1  1 = m
K
IYI2!1
7,
so that (4) holds with K = 7. This establishes the lemma.
Proof of conclusion (2) of Theorem 1. Let cpit) + cp(t), n + oo, where
cp(t) is continuous at 0. Let us show that it follows that the family of probability measures {Pn} is tight, where Pn is the measure corresponding to Fn.
By (4) and the dominated convergence theorem,
Pn{R\(
~ ))} =
a a
(
J,xj;;,1ja
dFn(x) ::::; K
a
fa [1 
Re cpn(t)] dt
Kfa
[1  Re cp(t)] dt
+ a o
as n + oo.
Since, by hypothesis, cp(t) is continuous at 0 and cp(O)
there is an a > 0 such that
1, for every s > 0
for all n :?: 1. Consequently {P .} is tight, and by Lemma 2 there is a probability measure P such that
Hence
but also cpn(t)+ cp(t). Therefore cp(t) is the characteristic function of P.
This completes the proof of the theorem.
Corollary. Let {Fn} be a sequence of distribution functions and {cpn} the
corresponding sequence of characteristic functions. Also let F be a distribution
function and cp its characteristic function. Then F.~ F if and only if cpit)+
cp(t) for all t E R.
Remark. Let 1], 1J1o 1] 2 , be random variables and Fq. ~ Fq. In accordance
with the definition of 10 of Chapter II, we then say that the random variables
1] 1 , 1] 2 , converge to '7 in distribution, and write 11. 141].
Since this notation is selfexplanatory, we shall frequently use it instead
ofF~.~ F~ when stating limit theorems.
3. Proofs of Limit Theorems by the Method of Characteristic Functions
323
3. In the next section, Theorem 1 will be applied to prove the central limit
theorem for independent but not identically distributed random variables.
In the present section we shall merely apply the method of characteristic
functions to prove some simple limit theorems.
Theorem 2 (Law of LargeN umbers). Let ~ 1 , ~ 2 , . be a sequence ofindependent
identically distributed random variables with E I ~ 1 1 < oo, sn = ~ 1 + + ~n
and E~ 1 = m. Then Sn/n ~ m, that is,for every e > 0
n ~ oo.
PROOF. Let <p(t) = Ee;'~' and <fJs";n(t) = EeitSn!n. Since the random variables
are independent, we have
by (11.12.6). But according to (11.12.14)
<p(t) = 1 + itm
+ o(t),
t~o.
Therefore for each given t E R
n+
oo,
and therefore
The function <p(t) = eitm is continuous at 0 and is the characteristic function
of the degenerate probability distribution that is concentrated atm. Therefore
sn
n
~m,
and consequently (see Problem 7, 10, Chapter II)
This completes the proof of the theorem.
324
III. Convergence of Probability Measures. Central Limit Theorem
Theorem 3 (Central Limit Theorem for Independent Identically Distributed
Random Variables). Let 1 , 2 , be a sequence of independent identically
distributed (nondegenerate) random variables with Ee~ < oo and Sn =
+ + Then as n oo
ee
en
e1
xeR,
(5)
where
1
W(x) = ;;,::.
y 2n
PROOF.
Let
Ee
m,
Ve
fx
e" 212 du.
oo
a 2 and
cp(t) =
Eeit(~,
m>.
Then if we put
ES}
S cp"(t) = E exp{ it ~" ,
we find that
But by (11.12.14)
to.
Therefore
cpn(t)
[1  ;;~: + o(~)
r
e1 2 /2.
as n  oo for fixed t.
The function e 1212 is the characteristic function of a random variable
(denoted by %(0, 1)) with mean zero and unit variance. This, by Theorem 1,
also establishes (5). In accordance with the remark in Theorem 1, this can
also be written in the form
S~n J4 .ff(O, 1).
vsn
(6)
This completes the proof of the theorem.
The preceding two theorems have dealt with the behavior of the probabilities of (normalized and symmetrized) sums of independent and identically
distributed random variables. However, in order to state Poisson's theorem
(6, Chapter I) we have to use a more general model.
325
3. Proofs of Limit Theorems by the Method of Characteristic Functions
Let us suppose that for each n ~ 1 we are given a sequence of independent
random variables ~" 1 , , ~nn In other words, let there be given a triangular
array
(~11
~21> ~22
~31 ~32 ~33
of random variables, those in each row being independent. Put
sn = ~n1 + ... + ~nn
Theorem 4 (Poisson's Theorem). For each n ~ 1 let the independent random
variables ~" 1 , , ~nn be such that
P(~nk = 1) =
Pnk
+ qnk = 1. Suppose that
Pnk
max Pnk+ 0,
and
Ll:=
Pnk +
,( >
n+ oo,
0, n + oo. Then,Jor each m = 0, 1, ... ,
e).A.m
P(Sn = m) +  1 ,
n + oo.
(7)
m.
PRooF.
Since
for 1
n, by our assumptions we have
CfJsn(t) = EeitSn =
rr (pnkeit + qnk)
n
k=
f1 (1 + Pnieit 
1))+ exp{A.(eit  1)},
k=1
n+ oo.
The function cp(t) = exp{A.(eit  1)} is the characteristic function of the
Poisson distribution (II.l2.11), so that (7) is established.
If n(A.) denotes a Poisson random variable with parameter A., then (7) can
be written like (6), in the form
sn ~ n(A.).
This completes the proof of the theorem.
4.
PROBLEMS
1. Prove Theorem 1 for R", n
2.
2. Let eto ~ 2 , be a sequence of independent random variables with finite means
E 1.;.1 and variances V such that V ~ K < oo, where K is a constant. Use
Chebyshev's inequality to prove the law of large numbers (1).
e.
e.
326
III. Convergence of Probability Measures. Central Limit Theorem
3. Show, as a corollary to Theorem 1, that the family {cp.} is uniformly continuous and
that <Pn
4.
+
q> uniformly on every finite interval.
Let~., n ;;:::: 1, be random variables with characteristic functions q>~"(t), n :2:: 1. Show
that ~ .!4 0 if and only if q>~"(t) + 1, n + oo, in some neighborhood oft = 0.
5. Let X~> X 2 , be a sequence of independent random vectors (with values in Rk) with
mean zero and (finite) covariance matrix r. Show that
xl + + x .!!.. JV(O, r).
Jn
(Compare Theorem 3.)
4. Central Limit Theorem for Sums of Independent
Random Variables
1. Let us suppose that for every n
1 we have a sequence
of independent random variables with
E~nk
Lets.= ~nl
= 0,
+ + ~ ,
F.k(x) = P(~nk::; x),
cll(x) = (2n)li2 fooey2;zdy,
Theorem 1. A sufficient (and necessary) condition that
s. ~ %(0, 1)
is that
(A)
k=l
Jlxl>
lxiiF.k(x) <I>.k(x)ldx+0,
n + oo,
for every t: > 0.
This theorem implies, in particular, the traditional statement of the central
limit theorem under the Lindeberg condition.
Theorem 2. Suppose that the Lindeberg condition is satisfied, that is, that for
every t: > 0
(L)
k=l
then
s. ~ %(0, 1).
x 2 dF.k(x)+ 0,
Jlxl>e
n+oo;
4. Central Limit Theorem for Sums oflndependent Random Variables
327
Before proving these theorems (notice that Theorem 2 is an easy corollary
of Theorem 1) we discuss the significance of conditions (A) and (L).
Since
max E~;k ::; e2
1~k~n
L E(~;ki( I~nk I > e)),
k=1
it follows from the Lindeberg condition (L) that
max E~;k + 0,
n + oo.
(1)
1 ~k~n
From this it follows (by Chebyshev's inequality) that the random variables
are asymptotically infinitesimal (negligible in the limit), that is, that, for
every e > 0,
max P{l~nkl > e}+ 0,
1:!>k:!>n
n + oo.
(2)
Consequently we may say that Theorem 2 provides a condition for the
validity of the central limit theorem when the random variables that are
summed are asymptotically infinitesimal.
Limit theorems which depend on a condition of this type are known as
theorems with a classical formulation.
It is easy to give examples which satisfy neither the Lindeberg condition
nor the condition of being asymptotically infinitesimal, but where the central
limit theorem nevertheless holds. Here is a particularly simple example.
Let ~ 1 , ~ 2 , . be a sequence of independent normally distributed random
variables withE~"= 0, V~ 1 = 1, V~k = 2k 2 ,k 2:: 2.PutSn = ~n 1 + .. + ~nn
with
~nk = ~kl Jit1 v~i
It is easily verified (Problem 1) that neither the Lindeberg condition nor the
condition of being asymptotically infinitesimal is satisfied, although the
central limit theorem is evident since S" is normally distributed withES" = 0,
vsn = 1.
Later (Theorem 3) we shall show that the Lindeberg condition (L) implies
condition (A). Nevertheless we may say that Theorem 1 also covers "nonclassical" situations (in which no condition of being asymptotically infinitesimal is used). In this sense we say that (A) is an example of a nonclassical condition for the validity of the central limit theorem.
2. PROOF OF THEOREM 1. We shall not take the time to verify the necessity of
(A), but we prove its sufficiency.
Let
328
III. Convergence of Probability Measures. Central Limit Theorem
It follows from 12, Chapter II, that
By the corollary to Theorem 1 of 3, we have S" ~ %(0, 1) if and only if
fn(t) + ({J(t), n + oo, for all real t.
We have
fn(t)  ({J(t) =
k;l
k;l
ll fnk(t)  ll (/Jnk(t).
Since I f,.k(t) I :::;; 1 and I({Jnk(t) I :::;; 1, we obtain
I f,.(t)
 ({J(t) I =
I)J f,.k(t) 
il
(/Jnk(t)
:::;; ktllfnk(t) (/Jnk(t)l
=
kt f_oooo (eitx itx
I
kt11 s:ooeitx d(Fnk <I>nk)
+ tt2x2)d(Fnk
<l>nk)
I,
I
(3)
where we have used the equations
i
1, 2.
Let us integrate by parts (Theorem 11 of Chapter II, 6, and Remark 2
following it) in the integral
and then let a+
oo and b+ oo. Since we have
and
X+ CIJ,
we obtain
4. Central Limit Theorem for Sums of Independent Random Variables
329
From (3) and (4),
lfn(t) q>(t)l:::;; kt1 1t
:::;; tltl 3 c
s:}eitxr
(
k=l
+ 2t 2
k=l
Jlxi:St
1 itx)(Fnix) <l>nk(x))dxl
lxiiFnk(x) <l>nk(x)ldx
Jlxl>e
lxiiFnk(x) <l>nk(x)ldx
where we have used the inequality
(
Jlxi:St
lxiiFnk(x) <l>nk(x)ldx:::;; 2(I;k,
(6)
which is easily deduced from (71), 6, Chapter II.
Since c > 0 is arbitrary, it follows from (5) and (A) that fn(t)+ cp(t) as
n + oo, for all t E R.
This completes the proof of the theorem.
3. The following theorem gives the connections between (A) and (L).
Theorem 3.
(1) (L) =>(A).
(2) If max, s;ks;n E(;k+ 0, n+ oo, then (A)=> (L).
(1) We noticed above that (L) implies that max 1 s;ks;n (I;k+ 0.
Consequently, since
1 (I;k = 1, we obtain
PROOF.
Lk=
n + oo,
(7)
where the integration on the right is over IxI > c/Jmax 1 s;ks;n (I;k. This,
together with (L), shows that
I (
k=l
Jlxl>e
x 2 d[Fnk(x)
+ <l>nk(x)]+ 0,
n+ oo,
(8)
foreveryc > O.Nowfixe > Oandleth = h(x)beacontinuouslydifferentiable
even function such that lh(x)l:::;; x 2 , h'(x) sgn x ~ 0, h(x) = x 2 for lxl > 2c,
h(x) = 0 for lxl:::;; B, and lh'(x)l :::;; 4x fore< lxl:::;; 2c. Then, by (8),
I (
k=l
Jlxl>t
h(x) d[Fnk(x)
+ <l>nk(x)] + 0,
n + oo.
330
III. Convergence of Probability Measures. Central Limit Theorem
By integration by parts we then find that, as n + oo,
kt1
L~.h'(x)[(1 
Since h'(x) = 2x for IxI
Jlxl~2e
r
k=l
+ <~nk(x)] dx
~
+ (1
 <lnk(x))] dx
L~.h(x) d[Fnk + <Ink]+ 0,
= kt
ktl Lseh'(x)[Fnk(x)
Fnk(x))
= ktl Lseh(x)d[Fnk
+ <InkJ+ 0.
2e, we therefore have
<~nk(x) I dx+ 0,
I X IIFnk(x) 
n+ oo.
Consequently, since e > 0 is arbitrary, we obtain (L)::? (A).
(2) For the function h = h(x) introduced above, we obtain by (7) and the
condition max 1 sksn a~k+ 0 that
(9)
n+ oo.
Again integrating by parts, we find that when (A) is satisfied,
r h(x) d[Fnk <~nkJ I ~ I fx~eh(x) d[(1  Fnk) (1  <~nJJ I
I JJxJ~e
+ Ikt1 Ls eh(x) d[Fnk  <~nkJ I
k=1
k=1
~ ktl L>eh'(x)[(1 
Fnk) (1 
+ kt1 Lselh'(x)IIFnk
k=1
JJxJ~e
lh'(x)IIFnk
<~nk)] dx
<~nkldx
(10)
<~nkldx
~ 4 ktl JxJ ~lxiiFnk(x) <~nk(x)l dx+ 0.
It follows from (9) and (10) that
( ~2e
k= 1
Jlxl
x 2 dFnk(x)
( ~
k= 1
JJ:xJ
h(x) dFnk(x)+ 0,
n+oo;
that is, (L) is satisfied.
4. PRooF OF THEOREM 2. According to Theorem 3, Condition (L) implies
(A); hence Theorem 2 follows at once from Theorem 1.
4. Central Limit Theorem for Sums of Independent Random Variables
331
5. We mention some corollaries in which ~ 1 , ~ 2 , . is a sequence of independent random variables with finite second moments. Let mk = E~k>
af = v~k > o,sn = ~1 + ... + ~n = Lk=1 af,andletFk = Fk(x)bethe
distribution function of ~k.
v;
Corollary 1. Let the Lindeberg condition be satisfied: for every 8 > 0,
n + oo.
(11)
Then
(12)
Corollary 2. Let the Lyapunov condition be satisfied:
1
v;+a k';;1E lsk mki
f:
2H
+
0,
n+ oo.
(13)
Then the central limit theorem (12) holds.
It is enough to prove that the Lyapunov condition implies the Lindeberg
condition.
Let 8 > 0; then
El~k mki 2 H = f_
00
2:
2:
00
lx mki 2 H dFk(x)
J{x:lxmki~<Vn)
8V~ I
lx mki 2 H dFk(x)
J{x:lxmki~<Vn)
(x  mk) 2 dFix)
and therefore
Corollary 3. Let
~ 1 , ~ 2 , ... be independent identically distributed random
variables with m = E~ 1 and 0 < a 2 = V~ 1 < oo. Then
332
III. Convergence of Probability Measures. Central Limit Theorem
Therefore the Lindeberg condition is satisfied and consequently Theorem 3
of 3 follows from Theorem 2.
Corollary 4. Let
n 2 1,
ee
1,
be independent random variables such that,for all
2 , ..
lenl:::; K,
where K is a constant and V,.+ oo as n+ oo. Then by Chebyshev's inequality
{x: lxmkl ~eVn
lx mkl 2 dFk(x)
= E[(ek
mk)2 I!ek mk! 2 sV,.)]
:::; (2K) 2 P{Iek mkl 2 sV,.}:::; (2k) 2
(]2
2
Vk 2 + 0,
n
n+ oo.
Consequently the Lindeberg condition is again applicable, and the central limit
theorem holds.
6. We remarked (without proof) in Theorem 1 that condition (A) is also
necessary. The following (LindebergFeller) theorem shows that, with the
supplementary condition max 1 s ks n Ee;k + 0, the Lindeberg condition (L)
is also necessary.
Theorem 4. Let max 1 sksn Ee;k+ 0, n + oo. Then (L) is necessary and suf
.ficientfor the validity of the central limit theorem: Sn ~ %(0, 1).
The proof is based on the following proposition, which estimates the "tails~
of the variance in terms of the behavior of the characteristic function at the
origin (compare Lemma 3, 3, Chapter III).
Lemma. Let e be a random variable with distribution function F = F(x),
Ee = 0, ve = a 2 < oo. Thenforeacha > 0
Jlxl~1/a
x 2 dF(x) :::; _.;. [Re J(jUa)  1 + 3a 2a2 ].
a
PRooF. We have
Re f(t)  1
+ !a2 t 2
!a 2 t 2
!a 2 t 2
Taking t
Jlxl:51/a
Jlxl>1/a
2 !a 2 t 2
=
J:oo [1 
(!t 2
(14)
cos tx] dF(x)
[1  cos tx] dF(x)
[1  cos tx] dF(x)
!t 2
2a 2 )
Jlxl:>1/a
(
Jlxl>1/a
x 2 dF(x)  2a 2
x 2 dF(x).
aJU, we obtain (14), as required.
Jlxl>1/a
x 2 dF(x)
333
4. Central Limit Theorem for Sums of Independent Random Variables
PROOF OF THEOREM 4. The sufficiency was established in Theorem 2. We now
prove the necessity.
Let
0,
E~nk =
max a;k
+
n+ oo.
0,
(15)
1 o<;ko<;n
Since
n fnk(t)+
n
(16)
e(1/2)t2,
k=1
we can find, for a given t, a number no = no(t) such that
~ n0 (t) and consequently
nfnk(t)
n
In
nk=
J,k(t) > 0 for
L In J,k(t),
k= 1
k= 1
where the logarithms are well defined. Then since
Ifnk( t)  11 S a;k t 2 ,
we have, by (15),
Ikt1{In [1 + Unk(t) 
1)]  Unk(t)  1]}
S Llfnit)11 2
k=1
t4
s 
4
max
L u~k
k=1
u;k
1o<;kos;n
t4
2
0,
 4 max ank+
n+ oo.
1o<;ko<;n
Consequently, by using (16), we have
n
Re
L Unk(t) 
k=1
1]
+ !t 2 = L [Re fnk(t)
k=1
 1
+ !t 2 a;k] + 0.
In particular, if we take t = a~ we find that
n
k= 1
[Re fnk(a~)  1
+ 3a 2a;k]
and therefore, by (14), and for every
1::
+
0,
n+ oo,
= 1/a > 0,
kt1~xi;,;ex 2 dFnk(x) s B2kt1[Re f,k(a~) 
1 + 3a 2a ]+
which shows that the Lindeberg condition is satisfied.
00,
n+ oo,
334
III. Convergence of Probability Measures. Central Limit Theorem
7. The method that we used for Theorem 1 can be used to obtain a corresponding condition for the central limit theorem without assuming that the
second moments are finite.
For each n ~ 1 let
be independent random variables with
E~nk =
0,
Let g = g(x) be a bounded nonnegative even function with the following
properties: g(x) = x 2 for lxl ~ 1, minlxl<:l g(x) > 0, lg'(x)l ~canst.
Define f:.nk(g) by the equation
J oo g(x) dFnk(x) = Joo g(x) d<D(~)
t:..k(g)

00
00
Theorem 5. Let
and for each
> 0 let
Then
The proof is left to the reader (Problem 4); it can be carried out along the
same lines as the proof of Theorem 1.
8.
PROBLEMS
~ 1 , ~ 2 , . be a sequence of independent normally distributed random variables
with E~k = 0, k ~ 1, and V~ 1 = 1, V~k = 2k 2 , k ~ 2. Show that gd does not satisfy
the Linde berg condition and also is not asymptotically infinitesimal.
1. Let
2. Prove (4).
3. Let ~ 1 , ~ 2 , . be a sequence of independent identically distributed random variables
with E~ 1 = 0, E~f = 1. Show that
4. Prove Theorem 5. (Hint: use the method of the proof of Theorem 1, applying integration by parts twice in the integral (eitn  itx + !t 2x 2) d(Fnk  <l>.k).)
J!
335
5. Infinitely Divisible and Stable Distributions
5. Infinitely Divisible and Stable Distributions
1. In stating Poisson's theorem in 3 we found it necessary to use a triangular
array, supposing that for each n ~ 1 there was a sequence of independent
random variables {en k}, 1 ~ k ~ n.
Put
T,.
en, 1 +
+ en,n
n ~ 1.
(1)
The idea of an infinitely divisible distribution arises in the following problem: how can we determine all the distributions that can be expressed as
limits of sequences of distributions of random variables T,., n ~ 1?
Generally speaking, the problem of limit distributions is indeterminate
k = 0,
in such great generality. Indeed, if is a random variable and 1 =
1 < k ~ n, then T,.
and consequently the limit distribution is the
distribution of which can be arbitrary.
In order to have a more meaningful problem, we shall suppose in the
present section that the variables
1
are, for each n ~ 1, not only
independent, but also identically distributed.
Recall that this was the situation in Poisson's theorem (Theorem 4 of 3).
The same framework also includes the central limit theorem (Theorem 3
of 3) for sums sk =
+ ... + ~n n ~ 1, of independent identically distributed random variables 1 , 2 , In fact, ifwe put
e.
=e
en,
e1
en,
'
e, en,
en,n
ee
then
T.
~~
k~1 n,k
Sn  ES"
V,
Consequently both the normal and the Poisson distributions can be
presented as limits in a triangular array. If T,.+ T, it is intuitively clear that
since T, is a sum of independent identically distributed random variables, the
limit variable T must also be a sum of independent identically distributed
random variables. With this in mind, we introduce the following definition.
Definition 1. A random variable T, its distribution F T and its characteristic function q>T are said to be infinitely divisible if, for each n ~ 1, there are
independent identically distributed random variables r, 1, , 'In such thatt
T 4 r, 1 + + 'In (or, equivalently, F T = F~, * * F~"' or q>T = (q>,11 )n).
Theorem 1. A random variable T can be a limit of sums T, =
only if T is infinitely divisible.
t The
Lk= 1 en,
if and
notation ~ 4. '1 means that the random variables ~ and '1 agree in distribution, i.e.
F,(x), x e R.
F~(x) =
336
Ill. Convergence of Probability Measures. Central Limit Theorem
PROOF. If T is infinitely divisible, for each n 2:: 1 there are independent
identically distributed random variables ~n.t .. , ~n.k such that T 4.
~n. 1 + + ~n.k and this means that T 4. T,, n 2:: 1.
Conversely, let T, ~ T. Let us show that T is infinitely divisible, i.e. for
each k there are independent identically distributed random variables
1/t, , 1'/k such that T 4. 17 1 + + 1'/k
Choose a k 2:: 1 and represent T,k in the form C~l) + + C~k>, where
Y(l) _
'>n
J!
':.nk,1
+ ... + '>nk,n"'>n
J!
y(k)
J!
'>nk,n(k1)+1
+ ... + '>nk,nk
;;
Since T,k ~ T, n ~ oo, the sequence of distribution functions corresponding
to the random variables T,k> n 2:: 1, is relatively compact and therefore, by
Prohorov's theorem, is tight. Moreover,
[P(C~1 >
> z)]k
= P(C~1 >
> z, ... , C~k> > z) ::; P(T,k > kz)
and
[P(C~1 >
< z)]k
= P(C~1 >
< z, ... , C~k> < z)::; P(T,k < kz).
The family of distributions for C~1 >, n 2:: 1, is tight because of the preceding
two inequalities and because the family of distributions for T,k> n 2:: 1, is
tight. Therefore there are a subsequence {ni} !:; {n} and a random variable
17 1 such thatC~!> ~ 17 1 asni ~ oo. SincethevariablesC~l)' ... , C~k>areidentically
distributed, we have c~:) ~ 1'/z, .. ''~~) .!41'/k, where 1'11 4. 172 4.
= 1'/k
Since C~l), ... , C~k> are independent, it follows from the corollary to Theorem 1
of 3 that 1'/ 1 , , 11k are independent and
0
T,,k = C~!>
+ .. + C~~> 4 1'11 + .. + 11k
But T,,k.!!... T; therefore (Problem 1)
T 4. 1'11 + .. + 1'/k
This completes the proof of the theorem.
Remark. The conclusion of the theorem remains valid if we replace the
hypothesis that ~n. 1, , ~n." are identically distributed for each n 2:: 1 by the
hypothesis that they are uniformly asymptotically infinitesimal (4.2).
2. To test whether a given random variable T is infinitely divisible, it is
simplest to begin with its characteristic function qJ(t). If we can find characteristic functions qJnCt) such that qJ(t) = [({J"(t)]" for every n 2:: 1, then T is
infinitely divisible.
In the Gaussian case,
and if we put
we see at once that qJ(t) = [({Jn(t)]".
337
5. Infinitely Divisible and Stable Distributions
In the Poisson case,
<p(t) = exp{A.(ei'  1)},
and if we put <pit)= exp{(A./n)(ei' 1)} then <p(t) = [<pit)]n.
If a random variable T has a r distribution with density
Xa1exfP
{
f(x) =
r(a)pa '
X> O
 '
0,
X< 0,
it is easy to show that its characteristic function is
1
<p(t) = (1  i{3t)a
Consequently <p(t)
= [<pn(t)]n where
<fJn(t)
1
(1  i{3t)afn'
and therefore T is infinitely divisible.
We quote without proof the following result on the general form of the
characteristic functions of infinitely divisible distributions.
Theorem 2 (LevyKhinchin Theorem). A random variable T is irifinitely
divisible if and only if <p(t) = exp ljl(t) and
1/J(t) =
where
f3 E R, cr 2
2 2
. it/3  t cr + fro (e"x
~ 0 and
1 +x 2 dA.(x),
1  itx) 2
2
1+x
x
oo
A. is a .finite measure on (R, PA(R)) with A.{O}
(2)
=
0.
3. Let ~ 1 , ~ 2 , .. be a sequence of independent identically distributed
random variables and Sn = ~ 1 + + ~n. Suppose that there are constants
bn and an > 0, and a random variable T, such that
Sn bn
__::__....:.: "+
(3)
We ask for a description of the distributions (random variables T) that can be
obtained as limit distributions in (3).
If the independent identically distributed random variables ~ 1 , ~ 2 , ..
satisfy 0 < cr 2 = V~ 1 < oo, then if we put bn = nE~ 1 and an= crJn, we
find by 4 that Tis the normal distribution %(0, 1).
If f(x) = ()jn(x 2 + () 2 ) is the Cauchy density (with parameter () > 0)
and ~ 1 , ~ 2 , ... are independent random variables with density f(x), the
characteristic functions <p~ 1 (t) are equal to eOitl and therefore <fJsn;n(t) =
(eOitlfn)n = e 6 111, i.e. Sn/n also has a Cauchy distribution (with the same
parameter ()).
338
III. Convergence of Probability Measures. Central Limit Theorem
Consequently there are other limit distributions besides the normal: the
Cauchy distribution, for example.
If we put ~nk = (~Jan)  (bJnan), 1 ~ k ~ n, we find that
i ~n. k
sn  bn =
an
( = T,.).
k= 1
Therefore all conceivable distributions for T that can conceivably appear as
limits in (3) are necessarily (in agreement with Theorem 1) infinitely divisible.
However, the specific characteristics of the variable T,. = (Sn  bn)/an may
make it possible to obtain further information on the structure of the limit
distributions that arise.
For this reason we introduce the following definition.
Definition 2. A random variable T, its distribution function F(x), and its
characteristic function cp(t) are stable if, for every n ~ 1, there are constants
an > 0, bn, and independent random variables ~ 1 , , ~n distributed like T,
such that
(4)
anT + bn 4 ~ 1 + + ~n
or, equivalently, F[(x  bn)/an] = F * * F(x), or
'..'
ntimes
(5)
Theorem 3. A necessary and sufficient condition for the random variable T
to be a limit in distribution of random variables (Sn  bn)/an, an > 0, is that Tis
stable.
PRooF. If T is stable, then by (4)
T 4 Sn bn ,
an
where Sn = ~ 1 + + ~n and consequently (Sn  bn)/an .4 T.
Conversely, let ~ 1, ~ 2 , be a sequence of independent identically distributed random variables, Sn = ~ 1 + + ~n and (Sn bn)/an ~ T, an> 0.
Let us show that T is a stable random variable.
If T is degenerate, it is evidently stable. Let us suppose that T is nondegenerate.
Choose k ~ 1 and write
(1) _ )!
Sn
'o1
T(l)
+ ... + 'on
)!
Sn<1)  bn
T(k) =
' '
n
an
s<k) _ )!
n'o(k1)n+1
s<k)
n an
+ + 'okn
)!
It is clear that all the variables T~ll, ... , T~kl have the same distribution and
n ~ oo, i = 1, ... , k.
339
5. Infinitely Divisible and Stable Distributions
Write
Then
u~k> ~ y(l>
+ ... +
y<k>,
where y<l> 4 . 4 y<k> 4:. T.
On the other hand,
u<k) = ~ 1
n
+ ... + ~kn
 kb.
a.
v. + p<k)
n '
ll(k)
n kn
(6)
where
and
It is clear from (6) that
V.
kn =
u<k> _ p<k>
n
(k)
lln
'
where J.k. ~ T, U~k> ~ T(l> + + r<k>, n+ oo.
It follows from the lemma established below that there are constants
ll<k> > 0 and p<k> such that ll~k) + ll<k> and f3~k> + p<k> as n + oo. Therefore
d y<l)
T
+ ... +
y<k)  p<k)
<k>
ll
'
which shows that T is a stable random variable.
This completes the proof of the theorem.
We now state and prove the lemma that we used above.
Lemma. Let ~. ~ ~ and let there be constants a. > 0 and b. such that
d
a.~.+ b.+~'
where the random variables ~ and ~ are not degenerate. Then there are constants a > 0 and b such that lim a. = a, lim b. = b, and
~=a~+ b.
340
III. Convergence of Probability Measures. Central Limit Theorem
PRooF. Let cpn, cp and iP be the characteristic functions of ~n ~ and ~' respectively. Then cp""~"+b"(t), the characteristic function of an~n + bn, is equal
to eirb"cpn(an t) and, by Theorem 1 and Problem 3 of 3,
eitb"cpn(an t)
+
ip(t),
(7)
cpn(t)
+
cp(t)
(8)
uniformly on every finite interval of length t.
Let {n;} be a subsequence of {n} such that an1 + a. Let us first show that
a < oo. Suppose that a = oo. By (7),
sup llcpn(ant)l
ltiSc
lip(t)ll+ 0,
n+ oo
for every c > 0. We replace t by t 0 /an; Then, since an;+ oo, we have
and therefore
Icpn.(to) I + Iip(O) I = 1.
But Icpn1(t 0) I + Icp(t 0 ) 1. Therefore Icp(t 0 ) I = 1 for every t 0 E R, and consequently, by Theorem 5, 12, Chapter II, the random variable ~ must be
degenerate, which contradicts the hypotheses of the lemma.
Thus a < oo. Now suppose that there are two subsequences {n;} and {n;}
such that an, + a, ani + a', where a =f a'; suppose for definiteness that
0 :::; a' < a. Then by (7) and (8),
lcpn 1(anJ)I+ lcp(at)l,
lcpn1(an,t)l+ liP(t)l
and
Consequently
lcp(at)l
lcp(a't)l,
and therefore, for all t E R,
lcp(t)l =
lcp(~t)l = ... = lcp((~)"t)l+ 1,
n+ oo.
Therefore Icp(t) I = 1 and, by Theorem 5 of 12, Chapter II, it follows that ~
is a degenerate random variable. This contradiction shows that a = a'
and therefore that there is a finite limit lim an = a, with a ~ 0.
Let us now show that there is a limit lim b" = b, and that a > 0. Since (8)
is satisfied uniformly on each finite interval, we have
cpn(ant)+ cp(at),
341
5. Infinitely Divisible and Stable Distributions
and therefore, by (7), the limit limn+oo eitbn exists for all t such that qJ(at) =F 0.
Let [J > 0 be such that qJ(at) =F 0 for all It I < b. For such t, lim eitbn exists.
Hence we can deduce (Problem 9) that lim Ibn I < oo.
Let there be two sequences {n;} and {ni} such that lim bn, = b and
lim bn; = b'. Then
for It I < [J, and consequently b = b'. Thus there is a finite limit b = lim bn
and, by (7),
ijJ(t) = eitbqJ(at),
which means that ~ 4 a~ + b. Since ~ is not degenerate, we have a > 0.
This completes the proof of the lemma.
4. We quote without proof a theorem on the general form of the characteristic functions of stable distributions.
Theorem 4 (LevyKhinchin Representation). A random variable Tis stable
if and only if its characteristic function qJ(t) has the form qJ(t) = exp 1/J(t),
t/l(t) = itf3  d It 1"'(1
whereO <a< 2,f3ER,d
G(t (X) =
'
+ iO I~ I G(t, a)).
1,t/ltl = Ofort
tan !na
if a =F 1,
{
(2/n) log It I if a= 1.
0,101:::;;
(9)
O,and
(10)
Observe that it is easy to exhibit characteristic functions of symmetric
stable distributions:
qJ(t) = edltl"',
where 0 < a :::;; 2, d
5.
(11)
0.
PROBLEMS
1. Show that ~ ~ '1 if ~n .!!... ~ and ~n .!!... '1
2. Show that if qJ 1 and qJ 2 are infinitely divisible characteristic functions, so is qJ 1
qJ 2
3. Let (/Jn be infinitely divisible characteristic functions and let ({Jn(t)+ ({J(t) for every
t e R, where ({J(t) is a characteristic function. Show that ({J(t) is infinitely divisible.
4. Show that the characteristic function of an infinitely divisible distribution cannot take
the value 0.
5. Give an example of a random variable that is infinitely divisible but not stable.
6. Show that a stable random variable ~always satisfies the inequality E I~ I' < oo for all
r E (0, IX).
342
III. Convergence of Probability Measures. Central Limit Theorem
7. Show that if~ is a stable random variable with parameter 0 <
differentiable at t = 0.
IX
:s; 1, then rp(t) is not
8. Prove that edlrla is a characteristic function provided that d ~ 0, 0 <
9. Let (bn)n;d be a sequence of numbers such that limn eirb. exists for all
Show that lim lbnl < oo.
IX
:s; 2.
It I < c5, c5 > 0.
6. Rapidity of Convergence in the
Central Limit Theorem
1. Let ~nl ... , ~"" be a sequence of independent random variables, S" =
~nl + + ~""' Fn(x) = P(Sn ~ x). If Sn+ %(0, 1), then FnCx)+ <l>(x) for
every x e R. Since <l>(x) is continuous, the convergence here is actually uniform (Problem 5 in 1):
n + oo.
supJFn(x) <l>(x)l+ 0,
(1)
In particular, it follows that
P(Sn ~ x) <I> (
xES)
JVS,"
n+ oo
+ 0,
(under the assumption that ESn and VSn exist and are finite).
It is natural to ask how rapid the convergence in (1) is. We shall establish
a result for the case when
n;;::: 1,
where ~ 1 , ~ 2 , .. is a sequence of independent identically distributed random
variables with E~k = 0, V~k = a 2 and El~ 1 1 3 < oo.
Theorem (Berry and Esseen). We have the bound
sup IFn(x) <D(x)l
x
~ CEIJt,
~
(2)
where Cis an absolute constant ((2n) 112 ~ C < 0.8).
PRooF. For simplicity, let a 2 = 1 and
(Subsection 10, 12, Chapter II)
sup IFnCx) <D(x)l
x
~~
(I
Jo
P3
= El~ 1 1 3 . By Esseen's inequality
fn(t) qJ(t) dt
t
+ 2T4 ~
n v 2n
(3)
343
6. Rapidity of Convergence in the Central Limit Theorem
where qJ(t) = e 1212 and
fit) = [f(tjjn)]",
with f(t) = Ee; 1 ~'.
In (3) we may take T arbitrarily. Let us choose
jn/(5/33).
T =
We are going to show that for this T,
I fit) ({J(t)l::; ~
}n ltl
3 e 1214 ,
ltl ::; T.
(4)
The required estimate (2), with C an absolute constant, will follow immediately from (3) by means of (4). (A more detailed analysis shows that
c < 0.8.)
We now turn to the proof of (4).
By Taylor's formula with integral remainder,
t
f(t) = 1 + 1 ! f'(O)
t3 Jl (1 
t2
+ 2 ! f"(O) + 2
v) 2f"'(vt) dv.
(5)
Hence
where 181 ::; 1.
If It I ::; Jn/(5/3 3 ), we have, since
t2
2n
/3 3 z
a 3 = 1.
lt 3 1/33
6n 312
:$
1
25
Consequently
f(t/Jn)
for
It I ::;
z ~~
= Jn/(5/3 3 ), and hence we may write
[f(t/jn)]" =
(6)
enlnnf(t/,fti).
By (5) (with f(t) replaced by In f(tjjll)) we find that
t2
In f(t/Jn) =  2n
here
8t 3
+ 6n 312
18 1 1::; 1 and
IOn f)/Ill = 1[f'" ! 2
3f"
r J + 2U'n f
::; (/33 + 3PzPt + 2PD<~~) 3
where
Pk =
E I~ 1 1\ k = 1, 2, 3.
(7)
(In f) 111(8 1 t/jn);
::;
7/3 3 ,
1
(8)
344
III. Convergence of Probability Measures. Central Limit Theorem
Using the inequality le 11 :s; lzlell, we can now show, for I tl :s; T =
Jn/(5P 3 ), that
l[f(t/Jn)J" _ et2 /21 = le"In/(t/Jiil _
<
et2f2l
2P31tl3 exp{ t2 + 21tl3 ~}
6
Jn
Jn
< 2P31 t 13et2f4.
6
Jn
This completes the proof of the theorem.
Remark. We observe that unless we make some supplementary hypothesis
about the behavior of the random variables that are added, (2) cannot be
improved. In fact, let ~ 1 , ~ 2 , be independent identically distributed random
variables with
It is evident by symmetry that
and hence by Stirling's formula
= l.Cn . 22n""' _1_ =
2
2fo
2"
j(2n) (2n).
It follows, in particular, that the constant C in (2) cannot be less than (2n)  1 12
2.
PROBLEMS
1. Prove (8).
2.
be independent identically distributed random variables with E~k = 0,
Ele 1 l3 < oo.
It is known [53] that the following nonuniform inequality holds: for all x E R,
Let~~> ~ 2 ,
vek
= u 2and
IFn(x) ~x)l :::;;
CEI~ 1 1 3
u\fo . (1 + lxl)3.
Prove this, at least for Bernoulli random variables.
345
7. Rapidity of Convergence in Poisson's Theorem
7. Rapidity of Convergence in Poisson's Theorem
1. Let 1] 1, 1] 2 , . , 1Jn be independent Bernoulli random variables, taking the
values 1 and 0 with probabilities
P(1Jk = 1) = Pk>
P(1Jk
= 0) =
1~ k
1  Pk>
n.
Write
sn
= 111 + ... + 1Jn,
Pk = P(S = k),
= 0, 1, ... ; A. > 0.
In 6 of Chapter I we observe that when p 1 = = P. = p and A. = np
we have Prohorov's inequality,
L IPk nkl ~ C1(A.)p,
00
k=O
where
C 1 (A.) = 2 min(2, A.).
When the Pk are not necessarily equal, but
that
Lk=
Pk
= A., LeCam showed
L IPk nkl ~ Cz(A.) max Pk>
00
k=O
1:5k:5n
where Cz(A.) = 2 min(9, A.).
The object of the present section is to prove the following theorem, which
provides an estimate of the approximation of the P k by the nk, specifically
not including the assumption that
1 Pk = A.. Although the proof does not
produce very good constants (see C(A.) below), it may be of interest because
of its simplicity.
Li:=
Theorem.
(1) We have the following inequality:
(1)
where min is taken over all permutations i = (i 1, i2 , , in) of(l, 2, ... , n),
P;0 = 0, and [ns] is the integral part of ns.
(2) IfLi:= 1 Pk = A., then
k~0 1Pk nkl ~ C(A.)~in 0~~~~~k~/;k A.sl
~
C(A.) max pk,
where
C(A.) = (2
+ 4A.)eu.
(2)
346
III. Convergence of Probability Measures. Central Limit Theorem
2. The key to the proof is the following lemma.
Lemma 1. Let S(t)
= Ll"2o 1'fk, where 1'fo = 0, 0 ~ t
Pk(t) = P(S(t) = k),
Then for every tin 0
(A.t)ke.<t
nk(t) =
k!
~ 1,
k = 0, 1, ....
1,
k~o IPk(t) nk(t)l ~ eur(2 + 4~~!k) ~~~srI ~~k A.s
(3)
PRooF. Let us introduce the variables
Xk(t) = J(S(t) = k),
where I(A) is the indicator of A. For each sample point, the function S(t),
0 ~ t ~ 1, is a generalized distribution function, for which, consequently, the
LebesgueStieltjes integral
J~xk(s) dS(s)
is defined; it is in fact just the sum
(j _ 1)
[nt]
LXk   'lj
i= 1
A simple argument shows that Xk(t), k ;;::: 0, satisfy
X 0 (t)
1
Xk(t) = 
{x
dS(s),
f~[Xk(s) Xk_ 1(s)] dS(s),
where X 0 (0) = 1 and Xk(O)
Now EXk(t) = Pk(t) and
0 (s)
1,
= 0 fork~ 1.
[nt]
Xk(s) dS(s) = E Lxk
o
i= 1
=
=
[nt]
1)
n
(j _1)E17j =
~ 17;
('
LEXk  n
i= 1
{Pk(s) dA(s),
where
[nt]
A(t) =
L Pk
k=O
( =ES(t)).
[nt]
(j _ 1)
Lpk Pi
i= 1
(4)
347
7. Rapidity of Convergence in Poisson's Theorem
Hence if we average the lefthand and righthand sides of(4) we find that
P 0 (t)
= 1  J:P 0 (s) dA(s),
(5)
Pk(t) =  s:[Pk(s) Pk_ 1(s)] dA(s),
In turn, it is easily verified that nk(t), k
system
n 0 (t)
1.
0, 0 :::;:; t :::;:; 1, satisfy the similar
= 1  J:n 0 (s) d(A.s),
nk(t) = 
[nis)  nkl (s )] d(A.s),
~ 1.
Therefore
n 0 (t) P 0 (t)
=  {[n0 (s) P 0 (s)] d(A.s)
+ {P0 (s) d(A(s)
(6)
A.s)
and
nk(t) Pk(t)
=  {[nk(s)
Pk(s )] d(A.s)
+ {[nk 1(s) Pk_ 1(s)] d(A.s)
+ {[Pk(s)
Pk_ 1 (s )] d[A(s) A.s].
By the formula for integration by parts (namely "dUV = U dV
see Theorem 11, 6, Chapter II) and by (5),
J~P 0 (s) d(A(s) 
A.s) = (A(t) A.t)P 0 (t)
+ J~(A(s)
(7)
+ VdU";
A.s)P 0 (s) dA(s),
(8)
{ [Pk(s)  Pkt (s)] d(A(s)  A.s)
(A(t)  A.t)(Pk(t)  Pk 1(t))
+ {[Pk(s)
2Pk_ 1(s)
+ Pk_ 2 (s)](A(s)
where it is convenient to suppose that P _ 1 (s)
=0.
A.s)dA(s), (9)
348
III. Convergence of Probability Measures. Central Limit Theorem
From (6)(9) we obtain
k~0 lnk(t) Pk(t)l ~ 2 k~o lnk(s) Pk(s )ld(A.s)
+ 21A(t) A.tl + 4A(t)max IA(s) A.sl
O:;;;s:;;;t
~ 2 k~o lnk(s) Pk(s )ld(A.s)
+(2 + 4A(t)) max IA(s) A.sl.
O:;;;s:;;;t
Hence, by Lemma 2, which will be proved in Subsection 3,
max IA(s) ..lsi,
L IPk(t) nk(t)l ~ e M(2 + 4A(t)) o:;;;s:;;;t
k=O
where we recall that A(t) = 21"2 Pk
00
(10)
This completes the proof of Lemma 1.
3. PROOF OF THE THEOREM. Inequality (1) follows immediately from Lemma
1 if we observe that Pk = Pk(l), nk = nk(1), and that the probability Pk =
P{1] 1 + + 1Jn = k} is the same as P{1]; 1 + + 1];" = k}, where
(i 1 , i2 , , in) is an arbitrary permutation of(1, 2, ... , n).
Moreover, the first inequality in (2) and the estimate (3) follow from ( 1).
We have only to show that
I L Pik O:;;;s:;;; I k=O
min sup
i
[ns]
(11)
A.s ~ max Pk>
1 :;;;k:;;;n
where we may evidently suppose that A. = 1.
We write F;(s) = Lk~o P;k, G(s) = s, 0 ~ s ~ 1. These are distribution
functions: Fi(s) is a discrete distribution function with masses Pi 1 , Pi 2 , , Pi"
at the points 1/n, 2/n, ... , 1; and G(s) is a uniform distribution on [0, 1].
Our object is to find a permutation i* = (i!, ... , i:) such that
sup IFi(s) G(s)l ~ max Pk
Since
sup IF;"(s) G(s)l = max
!:;;;k:;;;n
O:;;;s:;;;l
I (~
Fi
)
~I
n
it is sufficient to study the deviations F;(s)  G(s) only at the points
s = kjn, k = 1, ... , n.
We observe that if all the Pk are equal (p 1 = = Pn = 1/n), then
1
n
sup IF;(s) G(s)l =  = max Pk
O:;;;s:;;;l
!:;;;k:;;;n
349
7. Rapidity of Convergence in Poisson's Theorem
We shall accordingly suppose that at least one of p 1, ... , Pn is not equal to
1/n. With this assumption, we separate the set of numbers P1> . .. , Pn into the
union of two (nonempty) sets
A = {pi: Pi> 1/n}
B = {pi:
and
Pi~
To simplify the notation, we write F*(s) = Fi.(s), Pit
It is clear that F*(1) = 1, F*(1  (1/n)) = 1  p:,
F*(1
~) =
1
(p: + P: 1), ,F*(~)
1
1/n}.
= Pt.
(p~ + + P2*).
Hence we see that the distribution F*(s), 0 ~ s ~ 1, can be generated by
successively choosing p~, then p~_ 1, etc.
The following picture indicates the inductive construction of p:,
P!1 , pf:
(0, 1 )  
~ ~
I
I
ii : ](1,
I
  :    J    L   ~  I
I
1
1
  _j    ...!   t   :
I
I
,I
~~~
I/
/I
r7{_ ~ _
p,
/1
/ I
2
n
/I
~:~ I
}fL/~~
I
/1
~~~~~~
P:
~J
 ~~~ 
I)
I
I
I
I
I
I
_j_
I
I
1 __ ,
2
n
1
1n
On the righthand side of the square [0, 1] x [0, 1] we start from (1, 1)
and descend an amount p~, where p~ is the largest number (or one of the
largest numbers) in A, and from the point (1, 1  p~) we draw a (dotted) line
parallel to the diagonal of the square that extends from (0, 0) to {1, 1).
350
III. Convergence of Probability Measures. Central Limit Theorem
Now we draw a horizontal line segment of length 1/n from the point
(1, 1  p:>. Its lefthand end will be at the point (1  (1/n), 1  p:), from
which we descend an amount
1 is the largest number (or one
1 , where
of the largest numbers) in B. Since
1 :::;; 1/n, it is easily seen that this line
segment does not intersect the dotted line and therefore G(1  (1/n)) F*((1  (1/n))) :::;;
From the point (1  (1/n), 1 1 ), we again draw a horizontal
either the lefthand end
possibilities:
two
are
There
line segment oflength 1/n.
the diagonal or on it;
below
falls
)
1
of this interval (1  (2/n),
1
diagonal. In the first
the
above
is
)
1
or the point (1  (2/n),
1
length
of
segment
2 , where
case, we descend from this point by a line
{p: _ 1 }.
B\
set
the
in
numbers)
largest
the
of
p:_ 2 is the largest number (or one
is clear
it
Again,
2)/n.)
2 > (n(This set is not empty, since Pi++
descend
we
case
second
the
In
:::;;
that G(1  (2/n))  F*((1  (2/n))
2 is the largest number (or one of
by a line segment oflength
2 , where
the largest numbers) in the set A\{p:}. (This set is not empty since
Pi++ P: 2 > (n 2)/2).) Since P: 2 :::;; p:, it is clear that in this case
p:_
p:.
p:_
p:_
p:  p:_
p:  p:_
p:  p:_
P:
p:_
p:.
P:
p:_
Continuing in this way, we construct the sequence P: 3 , . , Pi
It is clear from the construction that
for 1 :::;; k:::;; n.
Since
min sup
i
05s51
II
k=O
p;k s
I : :;
sup IF*(s) G(s)l :::;; p:,
05s51
we have also established the second inequality in (2).
Remark. Let us observe that the lower bound
min sup
i
05s51
I L P;k  s I~ !P:
[ns]
k=O
is evident.
4. Let A = A(t), t ~ 0, be a nondecreasing function, continuous on the right,
with limits on the left, and with A(O) = 0. In accordance with Theorem 12
of 6, Chapter II, the equation
Zr = K +
J~z._ dA(s)
(12)
351
7. Rapidity of Convergence in Poisson's Theorem
has (in the class of locally bounded functions that are continuous on the right
and have limits on the left) a unique solution, given by the formula
Z, = KtS',(A),
where
tS',(A) = eA<tJ
(13)
n (1 + L\A(s))e&<>.
(14)
oss:!>f
Let us now suppose that the function V(t), t ~ 0, which is locally bounded,
continuous on the right, and with limits on the left, satisfies
Jt;
~ K + LV. dA(s),
where K is a constant, for every t
Lemma 2. For every t
(15)
0.
0,
(16)
PRooF. Let T = inf{t ~ 0: Jt; > KtS',(A)}, where inf{0} = oo. If T = oo,
then (16) holds. Suppose T < oo; we show that this leads to a contradiction.
By the definition of T,
From this and (12),
Vr
~ KtS'r(A) =
~K
+K
LTs._(A) dA.
V.
s:
dA(s)
~ Vr.
(17)
If Vr > KtS' r(A), inequality (17) yields Vr > Vr, which is impossible, since
IVrl <
00.
Now let Vr = Kef r(A). Then, by (17),
Vr
=K +
s:
V.
dA(s).
From the definition ofT, and the continuity on the right of Jt;, Kcf,(A) and
A(t), it follows that there is an h > 0 such that when T < t ~ T + h we have
Jt; > KtS',(A) and Ar+h  Ar ~
t.
Let us write 1/11 = Jt;  Kcf,(A). Then
0 < 1/1,
~ J:t/1. dA.,
T < t
~ T + h,
and therefore
0:::;;; sup 1/11
:::;;;
sup
1/J,.
T:s;r:s;T+.!
352
III. Convergence of Probability Measures. Central Limit Theorem
Hence it follows that ljl, = 0 for T
assumption that T < oo.
+ h,
which contradicts the
Corollary (GronwallBellman Lemma). In (15) let A(t) = J~ a(s) ds and
K;;;::: 0. Then
(18)
5.
PROBLEMS
1. Prove formula (4).
~ 0, be functions of locally bounded variation, continuous
on the right and with limits on the left. Let A(O) = B(O) = 0 and L\A(t) > 1,
t ~ 0. Show that the equation
2. Let A = A(t), B = B(t), t
z,
J~z._ dA(s) + B(t)
has a unique solution $.(A, B) of locally bounded variation, given by
3. Let ~ and 11 be random variables taking the values 0, 1, ....
Let
p(~,
1'/) = supj P(~
A) P(l'/
A)j,
where sup is taken over all subsets of {0, 1, ... }.
Prove that
(1) p(~, 1'/) = t Ik'=ol P(~ = k) P('l = k)l;
(2) p(~. 1'/) s p(~. ~) + p(~. 1'/);
(3) if Cis independent of(~, 1'/), then
p(~
(4) If the vectors (~ 1 , ,
~.)and
+ ,, '1 + 0 s
it 1'/i)
J/(~;. 1'/;).
Let~= ~(p) be a Bernoulli random variable with P(~ = 1) = p, P(~ = 0) = 1  p,
0 < p < 1; and n = n(p) a Poisson random variable with En= p. Show that
p(~(p),
5.
1'/);
('7 1 , , 11.) are independent, then
p(t1 ~i
4.
p(~.
n(p)) = p(1  eP)
p2
Let~ = ~(p) be a Bernoulli random variable with P(~ = 1) = p, P(~ = 0) = 1  p,
0 < p < 1, and let n = n(A.) be a Poisson random variable such that P(~ = 0) =
P(n = 0). Show that A.= ln(1  p) and
p(~(p),
n(A.)) = 1  e< A_e.t
s tA. 2
7. Rapidity of Convergence in Poisson's Theorem
353
6. Show, using Property (4) of Problem 3, and the conclusions of Problems 4 and 5,
that if ~ 1 = ~ 1(pd, ... , ~. = ~.(p.) are independent Bernoulli random variables,
0 < P; < 1, and A; = ln(1  p;), 1 :::;; i :::;; n, then
and
CHAPTER IV
Sequences and Sums of Independent
Random Variables
1. ZeroorOne Laws
1. The series 2::': 1 (1/n) diverges and the series 2::': 1 ( 1)"(1/n) converges.
We ask the following question. What can we say about the convergence or
divergence of a series L.."'= 1 (~n/n), where ~ 1 , ~ 2 , isasequenceofindependent
identically distributed Bernoulli random variables with P(~ 1 = + 1) =
P(~ 1 = '1) =
In other words, what can be said about the convergence
of a series whose general term is 1/n, where the signs are chosen in a random
manner, according to the sequence ~ 1 , ~ 2 , . ?
Let
t?
A1 =
{w: I
n= 1
~"converges}
n
2::':
be the set of sample points for which
1 (~n/n) converges (to a finite
number) and consider the probability P(A 1 ) of this set. It is far from clear,
to begin with, what values this probability might have. However, it is a
remarkable fact that we are able to say that the probability can have only two
values, 0 or 1. This is a corollary of Kolmogorov's "zeroone law," whose
statement and proof form the main content of the present section.
2. Let (Q, , P) be a probability space, and let ~I> ~ 2 , be a sequence of
random variables. Let !F': = a(~"' ~n+ 1 , ) be the aalgebra generated by
~"' ~n+ 1 , , and write
!!'
n ~F:.
00
n=1
355
I. ZeroorOne Laws
Since an intersection of aalgebras is again a aalgebra, X is a aalgebra. It is
called a tail algebra (or terminal or asymptotic algebra), because every
event A EX is independent of the values of ~ 1 , .. , ~n for every finite number n,
and is determined, so to speak, only by the behavior of the infinitely remote
values of~ 1 , ~ 2 ,
Since, for every k 2::: 1,
A1
we have A1
={I ~nn converges} = {I ~nn converges}
n=k
n=1
nk ff'k =X. In the same way, if
A2 = {
ff'k,
~1 ~2 ... is any sequence,
~ ~n converges} EX.
The following events are also tail events:
A3 =
where In
{~n E
In for infinitely many n},
fJ6(R), n 2::: 1;
A4
{Ji~ ~n < 00 };
A5 =
+ " ' + ~n
{I~
1m ~ 1
A6 =
{I
A7
{~n converges};
As =
1m ~ 1
+ ' ' ' + ~n
n
{rrm:
n
Sn
J2n log n
}
<oo;
}
<c;
1}.
On the other hand,
= 0 for all n
B1 =
{~n
B2 =
{li~(~ 1 + + ~n) exists and is less than c}
2::: 1},
are examples of events that do not belong to X.
Let us now suppose that our random variables are independent. Then by
the BorelCantelli lemma it follows that
P(A3)
P(A3) =
L P(~n E Jn) <
1 ~ L P(~n E Jn) =
0~
00,
00.
Therefore the probability of A 3 can take only the values 0 or 1 according to
the convergence or divergence of L P(~n E Jn). This is Borel's zeroone law.
356
IV. Sequences and Sums of Independent Random Variables
Theorem 1 (Kolmogorov's ZeroOne Law). Let ~ 1 , ~ 2 , . be a sequence of
independent random variables and let A E !!f. The P(A) can only have one of the
values zero or one.
PROOF. The idea of the proof is to show that every tail event A is independent
of itself and therefore P(A n A) = P(A) P(A), i.e. P(A) = P 2 (A), so that
P(A) = 0 or 1.
If A EX then A E ffr" = crg 1 , ~ 2 , ... } = cr(Un ffD, where ff~ =
a{~ 1, ... , ~n}, and we can find (Problem 8, 3, Chapter II) sets An E ff~,
m :2: 1, such that P(A !::,. An) + 0, n + oo. Hence
P(An)+ P(A),
P(An n A)+ P(A).
(1)
But if A E X, the events An and A are independent for every n :2: 1. Hence it
follows from (1) that P(A) = P 2 (A) and therefore P(A) = 0 or 1.
This completes the proof of the theorem.
Corollary. Let 1J be a random variable that is measurable with respect to the tail
aalgebra X, i.e. {IJ E B} E !!f, BE ?4(R). Then 1J is degenerate, i.e. there is a
constant c such that P(IJ = c) = 1.
3. Theorem 2 below provides an example of a nontrivial application of
Kolmogorov's zeroone law.
Let ~ 1 , ~ 2 , . be a sequence of independent Bernoulli random variables
with P(~n = 1) = p, P(~n = 1) = q, p + q = 1, n :2: 1, and let Sn =
~ 1 + + ~n It seems intuitively clear that in the symmetric case (p = !)
a "typical" path of the random walk Sn, n :2: 1, will cross zero infinitely often,
whereas when p # 1it will go off to infinity. Let us give a precise formulation.
Theorem 2. (a)
(b) lfp #
~fp =
1 then P(Sn =
1, then P(Sn =
0 i.o.) = 1.
0 i.o.) = 0.
PRooF. We first observe that the event B = (Sn = 0 i.o.) is not a tail event, i.e.
B!!f=nff;:o, ff:O=a{~n,~n+t .. } Consequently it is, in principle,
not clear that B should have only the values 0 or 1.
Statement (b) is easily proved by applying (the first part of) the BorelCantelli lemma. In fact, if B 2 n = {S 2 n = 0}, then by Stirling's formula
P(B ) =
2n
en
n n
2nP q
"'(4pqt
!::
vnn
and therefore L P(B 2 n) < oo. Consequently P(Sn = 0 i.o.) = 0.
To prove (a), it is enough to prove that the event
A =
~ Sn
{hm
Jn =
has probability 1 (since A
5;:;
B).
Jn
. Sn
}
oo, hm
=  oo
357
1. ZeroorOne Laws
Let
Then Ac! A, c+ oo, and all the events A, A 0 A~, A~ are tail events. Let us
show that P(A~) = P(A~) = 1 for each c > 0. Since A~ E ?and A~ E ?, it is
sufficient to show only that P(A~) > 0, P(A~) > 0. But by Problem 5
P(lim }n <c)= P(rrm }n >c)~ umP(}n >c)> O,
where the last inequality follows from the De MoivreLaplace theorem.
Thus P(Ac) = 1 for all c > 0 and therefore P(A) = limc .... oo P(Ac) = 1.
This completes the proof of the theorem.
4. Let us observe again that B = { Sn = 0 i.o.} is not a tail event. Nevertheless,
it follows from Theorem 2 that, for a Bernoulli scheme, the probability of this
event, just as for tail events, takes only the values 0 and 1. This phenomenon
is not accidental: it is a corollary of the HewittSavage zeroone law, which
for independent identically distributed random variables extends the result
of Theorem 1 to the class of" symmetric" events (which includes the class of
tail events).
Let us give the essential definitions. A onetoone mapping n =
(n 1, n 2 , ) of the set (1, 2, ... ) on itself is said to be a finite permutation if
rt 11 = n for every n with a finite number of exceptions.
If~= ~ 1 , ~ 2 , is a sequence of random variables, n(~) denotes the
sequence (~,11 , ~ 112 , ). If A is the event {~ E B}, BE PJ(R 00 ), then n(A)
denotes the event {n(~) E B}, BE .94(R
We call an event A = {~ E B}, BE PJ(R 00 ), symmetric if n(A) coincides with
A for every finite permutation n.
An example of a symmetric event is A = {Sn = 0 i.o.}, where Sn =
~ 1 + + ~n Moreover, we may suppose (Problem 4) that every event in
ff;:'(S) = a{w: Sn, Sn+ 1 , .. } generated by
the tail aalgebra ?I(S) =
s1 = ~1 s2 = ~1 + ~2 ... is symmetric.
00
).
Theorem 3 (HewittSavage ZeroOne Law). Let ~ 1 , ~ 2 , be a sequence of
independent identically distributed random variables, and
A=
a symmetric event. Then P(A)
{w:(~ 1 ,~ 2 , ... )EB}
= 0
or 1.
PRooF. Let A = { ~ E B} be a symmetric event. Choose sets Bn E PJ(R") such
that, for An = {w: (~ 1 ... , ~n) E Bn},
n+ oo.
(2)
358
IV. Sequences and Sums of Independent Random Variables
Since the random variables ~ 1 , ~ 2 , . are independent and identically
distributed, the probability distributions P~(B) = P(~ E B) and P,..w(B) =
P(n.(O E B) coincide. Therefore
P(A lJ. A.)
= P~(B lJ. B.) = P,..w(B lJ. B.).
Since A is symmetric, we have
A
Therefore
P,.<~>(B
=g
B} = n.(A)
{nnC~) E
(3)
B}.
lJ. B.)= P{n.(~) E B) lJ. (n.(~) E B.)}
= P{(~ E B) lJ. (n.(~) E B.)} = P{A i:J. n.(A.)}.
(4)
Hence, by (3) and (4),
P(A lJ. A.) = P(A lJ. n.(A.)).
(5)
It then follows from (2) that
P(A lJ. (A. n n.(A.)))
+
n+ oo.
0,
(6)
Hence, by (2), (5) and (6), we obtain
P(A.)
+
P(A),
P(n.(A))
P(A. n n.(A.))
+
+
P(A),
P(A).
(7)
Moreover, since ~ 1 and ~ 2 are independent,
P(A.
f1
n.(A.)) = P{(~ 1 , ... , ~.) E B., (~n+ 1 ... , ~2n) E B;,}
= P{(~I ... , ~.) E B.} P{(~n+I ... , ~2n) E B.}
= P(A.)P(n:.(A.)),
whence by (7)
P(A)
= P 2 (A)
and therefore P(A) = 0 or 1.
This completes the proof of the theorem.
5.
PROBLEMS
1. Prove the corollary to Theorem 1.
2. Show that if (e.) is a sequence of independent random variables, the random variables
Urn and lim are degenerate.
e.
e.
e.,
3. Let (e.) be a sequence of independent random variables, s. = 1 + +
and let
the constants b. satisfy 0 < b. j oo. Show that the random variables Iiiii(S./b.) and
lim(S.Jb.) are degenerate.
e.,
4. Lets.= el + ... +
n ~ 1, and ~(S) =
Show that every event in ~(S) is symmetric.
n$';:'(S), $';:'(S) = a{ro: s., s.+ I ...}.
S. Let (e.) be a sequence of random variables. Show that {Iiiii
for each c > 0.
e.> c} 2ilm{e. > c}
359
2. Convergence of Series
2. Convergence of Series
1. Let us suppose that ~ 1 , ~ 2 , . is a sequence of independent random
variables, Sn = ~ 1 + + ~"' and let A be the set of sample points w for
which
~nCw) converges to a finite limit. It follows from Kolmogorov's
zeroone law that P(A) = 0 or 1, i.e. the series ~n converges or diverges with
probability 1. The object of the present section is to give criteria that will
determine whether a sum of independent random variables converges or
diverges.
L:
L:
Theorem 1 (Kolmogorov and Khinchin).
(a) Let E~n = 0, n ~ 1. Then if
(1)
the series ~n converges with probability 1.
(b) !{the random variables~"' n ~ 1, are unfformly bounded (i.e., P(l~nl s c)
= 1, c < oo ), the converse is true: the convergence of
~n with probability
1 implies (1).
The proof depends on
Kolmogorov's Inequality
(a) Let~ 1, ~ 2 , .. , ~n be independent random variables with E~i = 0, E~t < oo,
i S n. Then for every c; > 0
(2)
(b) If also
P(\~d
S c)= 1, iS n, then
(3)
PRooF. (a) Put
A= {max\Ski ~ e},
Ak = {\Sd <
Then A= LAk and
t:,
i = 1, ... ,k 1, \Ski~ e},
1 S k S n.
360
IV. Sequences and Sums of Independent Random Variables
But
ES;IAk = E(Sk + (ek+t + + en)) 2 1Ak
= ESf/Ak + 2ESiek+l + ... + en)/Ak
~ ESf/Ak'
+ E(ek+l + ... + en) 2 1Ak
since
+ .. + en)/Ak = ESk/Ak E(ek+t + .. +en)= 0
because of independence and the conditions Eei = 0, i :::;; n. Hence
ESk(ek+t
Es; ~
L ESf/ Ak ~ 8 2 L P(Ak) =
8 2 P(A),
which proves the first inequality.
(b) To prove (3), we observe that
ES;IA
= Es;
= ES;
 ES;Ix ~ Es;  82 P(A)
On the other hand, on the set Ak
1sktl:::;;
and therefore
Es;IA
1sk1:::;; 1sktl
8,
:::;; (8
82
+ 82 P(A).
(4)
+ 1ek1:::;; 8 + c
L ESf/Ak + L E(/Ak(Sn Sk)
k
2)
k=l
j=k+l
+ c) 2 L P(Ak) + L P(Ak) L Ee]
k
:::;; P(A{(8
+ c) 2 +
J
1
+ c) 2 + ES;),
Ee] = P(A)[(8
(5)
From (4) and (5) we obtain
P(A) >

Es; 
(8
= 1
82
+ c) 2 + Es;
82
(8
(8
> 1
c)2
+ c) 2 + Es;
(8
c)2
Es;
82 
This completes the proof of (3).
PROOF OF THEOREM 1. (a) By Theorem 4 of 10, Chapter II, the sequence
(Sn), n ~ 1, converges with probability 1, if and only if it is fundamental with
probability 1. By Theorem 1 of 10, Chapter II, the sequence (Sn), n ~ 1, is
fundamental (Pa.s.) if and only if
n... oo.
(6)
By (2),
P{sup ISn+k Snl
k~l
~ 8} =
limP{ max ISn+k Snl
N+oo
1Sk:SN
Therefore (6) is satisfied if:Lk"'=. 1 Ee~ < oo, and consequently
with probability 1.
~ 8}
L ek converges
361
2. Convergence of Series
(b) Let
L ~k converge. Then, by (6), for sufficiently large n,
P{sup ISn+k Snl ~ e} < 1.
(7)
k~l
By (3),
P{ supiSn+k Snl ~ e} ~ 1k~
Therefore if we suppose that
Lf=
(c + e)
E1'2"
"oo
2
L...k=n
'ok
E~f = oo, we obtain
P{sup ISn+k S.l
k~1
~ e} = 1,
which contradicts (7).
This completes the proof of the theorem.
If ~ 1 , ~ 2 , ... is a sequence of independent Bernoulli random
variables with P(~n = + 1) = P(~n = 1) = !, then the series ~nan, with
Ian I ~ c, converges with probability 1, if and only if La; < oo.
ExAMPLE.
2. Theorem 2 (TwoSeries Theorem). A sufficient condition for the convergence
of the series L ~" of independent random variables, with probability 1, is that
both series L E~n and L V~"converge. If P( I~n I ~ c) = 1, the condition is also
necessary.
PROOF. Ifl: V~" < oo, then by Theorem 1 the series L (~n E~n) converges
(Pa.s.). But by hypothesis the series E~n converges; hence ~"converges
(Pa.s.)
To prove the necessity we use the following symmetrization method. In
addition to the sequence ~ 1 , ~ 2 , ... we consider a different sequence ~ 1 ,
~2 ,
of independent random variables such that ~" has the same distribu
tion as ~"' n ~ 1. (When the original sample space is sufficiently rich, the
existence of such a sequence follows from Theorem 1 of 9, Chapter II. We
can also show that this assumption involves no loss of generality.)
Then if ~"converges (Pa.s.), the series ~"also converges, and hence
so does L (~.  ~.). But E(~.  ~.) = 0 and P( I~"  ~"I ~ 2c) = 1. Therefore LV(~. ~.) < oo by Theorem 1. In addition,
L

L: v~. = 1 L: V(~.  ~.) < oo.
Consequently, by Theorem 1, L (~.  E~.) converges with probability 1,
and therefore L E~" converges.
Thus if L ~"converges (Pa.s.)(and P( I~" I ~ c) = 1, n ~ 1) it follows that
both L E~. and LV~. converge.
This completes the proof of the theorem.
3. The following theorem provides a necessary and sufficient condition for
the convergence of L ~" without any boundedness condition on the random
variables.
362
IV. Sequences and Sums of Independent Random Variables
Let
c be a constant and
~c = {~' ~~~ ~ C,
0,
>c.
1~1
Theorem 3 (Kolmogorov's ThreeSeries Theorem). Let ~ 1 , ~ 2 , be a
sequence of independent random variables. A necessary condition for the convergence of"[. ~ .. with probability 1 is that the series
"[. E~~.
v~~.
"[. P(l~.. l;?: c)
converge for every c > 0; a sufficient condition is that these series converge
for some c > 0.
PRooF. Sufficiency. By the twoseries theorem,"[. ~~converges with probability
1. But if "[. P( I~.. I ;?: c) < oo, then by the BorelCantelli lemma we have
I(l ~ .. I ;?: c) < oo with probability 1. Therefore ~ .. = ~~ for all n with at
L.
most finitely many exceptions. Therefore
~ also converges (Pa.s.).
Necessity. If ~ converges (Pa.s.) then ~ .. + 0 (Pa.s.), and therefore,
for every c > 0, at most a finite number of the events {I~ .. I ;?: c} can occur
(Pa.s.). Therefore I( I~ .. I ;?: c) < oo (Pa.s.), and, by the second part of the
BorelCantelli lemma,
P( I~ .. I > c) < oo. Moreover, the convergence of
~ implies the convergence of
~~. Therefore, by the twoseries theorem,
both of the series E~~ and LV~~ converge.
This completes the proof of the theorem.
L.
L
L .
Corollary. Let ~ 1 , ~ 2 , be independent variables withE~, = 0. Then if
~2
L E 1 + ~~~~~~~ <
00,
the series "[. ~ .. converges with probability 1.
For the proof we observe that
~2
L E 1 + ~~~~~~~ < 00 <=> L E[~;I(I~nl ~ 1) + l~nii(I~nl > 1)] < 00.
Therefore if~! = ~ .. I( I~ .. I ~ 1), we have
L E(~!)2
<
00.
Since E~, = 0, we have
LIE~! I=
L IE~,I(I~.. I ~
1)1 =
L IE~,I(I~.. I > 1)1
~ l:EI~ .. II(I~ .. I > 1) < oo.
Therefore both
inequality,
P{l~ .. l
L E~!
> 1} =
and LV~! converge. Moreover, by Chebyshev's
P{I~.. II(I~ .. I > 1)
Therefore
P( I~ .. I > 1)
the threeseries theorem.
> 1}
~ E(I~ .. II(I~ .. I > 1).
< oo. Hence the convergence of L ~.. follows from
363
3. Strong Law of Large Numbers
4.
PROBLEMS
1. Let ~ 1 , ~ 2 , .. be a sequence of independent random variables, S" = ~ 1 ,
. , ~n
Show, using the threeseries theorem, that
(a) if I ~;; < oo (Pa.s.) then I ~" converges with probability 1, if and only if
IE
U(l~d ~!)converges;
(b) if I~" converges (Pa.s.) then I~;; < oo (Pa.s.) if and only if
(E l~nll(lenl ~ 1))2 < oo.
2. Let eI ~2' ... be a sequence of independent random variables. Show that I
(Pa.s.) if and only if
~;;
IE   2 <
1 + ~"
e;;
<
00
00.
3. Let ~ 1 , ~ 2 , be a sequence of independent random variables. Show that I~"
converges (Pa.s.) if and only if it converges in probability.
3. Strong Law of Large Numbers
1. Let ~ 1 , ~ 2 , . be a sequence of independent random variables with finite
second moments; Sn = ~ 1 + + ~n By Problem 2, 3, Chapter III, if the
numbers V~i are uniformly bounded, we have the law of large numbers:
Sn ESn
"'+
n+ oo.
'
(1)
A strong law of large numbers is a proposition in which convergence in
probability is replaced by convergence with probability 1.
One of the earliest results in this direction is the following theorem.
Theorem 1 (Cantelli). Let ~ 1 ,
finite fourth moments and let
~ 2 , .
be independent random variables with
El~nE~ni 4 ~C,
n~l,
for some constant C. Then as n+ oo
Sn ES"
   +
(Pa.s.).
(2)
PRooF. Without loss of generality, we may assume that E~" = 0 for n ~ 1.
By the corollary to Theorem 1, 10, Chapter II, we will have Sn/n+ 0 (Pa.s.)
provided that
for every e > 0. In turn, by Chebyshev's inequality, this will follow from
364
IV. Sequences and Sums oflndependent Random Variables
Let us show that this condition is actually satisfied under our hypotheses.
We have
4
s.
= (~1 + ... + ~.) = i~1~i
4! 2 2
2!2! ~i~j
i<j
"
4!
+ loFJ
.L.. 3'1'
~i~j
Remembering that
E~k
ES! =
= 0, k ::::;; n, we then obtain
n
L E~i + 6 L:
i=1
i,j=1
nC
E~?E~J s nC + 6
6n(n  1)
C = (3n 2
2
i,j=1
i<j
JE~i E~J
2n)C < 3n 2 C.
Consequently
L E(s: )
::::;; 3C
L n1
< oo.
This completes the proof of the theorem.
2. The hypotheses of Theorem 1 can be considerably weakened by the use of
more precise methods. In this way we obtain a stronger law of large numbers.
Theorem 2 (Kolmogorov). Let~ 1 , ~ 2 , .. be a sequence of independent random
variables with finite second moments, and let there be positive numbers b. such
that b. i oo and
"v~. < oo.
(3)
L.f?
n
Then
s. b ES. __,. O
n
In particular,
Pa.s..
(4)
if
(5)
then
S"  ES. .. 0 (P a.s. )
'n
For the proof ofthis, and of Theorem 2 below, we need two lemmas.
(6)
365
3. Strong Law of Large Numbers
Lemma 1 (Toeplitz). Let {a.} be a sequence of nonnegative numbers, b. =
> 0 for n ~ 1, and b. i oo, n+ oo. Let {x.} be a sequence of
L~= 1 a;, b.
numbers converging to x. Then
(7)
In particular,
if a.
1 then
x1
+ + x.
(8)
+X.
PROOF. Let e > 0 and let n0
Choose n1 > n0 so that
n0 (e) be such that Ix. 
no
.L
" ' j=
Ixi
xI ::;
~:/2
for all n
n0 .
xI < e/2.
Then, for n > n 1,
This completes the proof of the lemma.
Lemma 2 (Kronecker). Let {b.} be a sequence of positive increasing numbers,
b. j oo, n+ oo, and let {x.} be a sequence of numbers such that L x. converges.
Then
1
b.
j= 1
 L bixi+ 0,
In particular, if b. = n, x. = Y./n and
Y1
Let b 0
n + oo.
(10)
= 0, S 0 = 0, s. = L}= 1 xi. Then (by summation by parts)
n
L bjxj = L b/Sj sj1) =
j=1
(9)
L (y./n) converg_es, then
+ + Yn +,
0
n
PROOF.
n + oo.
j=1
bnSn  boSo 
L sj1(bj b;1)
j=l
366
IV. Sequences and Sums of Independent Random Variables
and therefore
since, if sn
~X,
then by Toeplitz' lemma,
1
b
L Si _ 1ai ~ x.
n j= 1
This establishes the lemma.
PROOF OF THEOREM 1. Since
Sn  ES"
bn
2_
I bk(~k bkE~k),
bn k=1
a sufficient condition for (4) is, by Kronecker's lemma, that the series
[(~k  E~k)/bk] converges (Pa.s.). But this series does converge by (3) of
Theorem 1, 2.
This completes the proof of the theorem.
EXAMPLE 1. Let ~ 1 , ~ 2 , . be a sequence of independent Bernoulli random
variables with P(~n= l)=P(~n= 1)=1. Then, since L [l/(n log 2 n)] < oo,
we have
r::. sn
v n log n
)
~
~ 0 (Pa.s ..
(11)
3. In the case when the variables ~ 1 , ~ 2 , . are not only independent but
also identically distributed, we can obtain a strong Jaw of large numbers
without requiring (as in Theorem 2) the existence of the second moment,
provided that the first absolute moment exists.
Theorem 3 (Kolmogorov). Let ~ 1 , ~ 2 , . be a sequence of independent
identically distributed random variables withE I~ 1 1< oo. Then
sn ~ m
n
where m
(Pa.s.)
(12)
= E~ 1 .
We need the following lemma.
Lemma 3. Let~ be a nonnegative random variable. Then
00
00
n=l
n=l
L P( ~ :2: n) ~ E~ ~ 1 + L P( ~ ;;:: n).
(13)
367
3. Strong Law of Large Numbers
The proof consists of the following chain of inequalities:
L P(e ~ n) = L L P(k ~ e<
00
00
n= 1
k + 1)
n= 1 k;;,n
00
= I
kP(k ~ e < k
k=1
+ 1) =
00
k=O
E[ei<k ~ e < k
r E[kl(k ~ e < k + 1)J
00
k=O
+ 1)J
r E[(k + 1)I(k ~ e< k + 1n
00
= Ee ~
00
k=O
(k
k=O
+ 1)P(k ~ e< k + 1)
00
00
L P(e ~ n) + L P(k ~ e<
n=1
k=O
00
k + 1) =
L P(e ~ n) +
n=1
1.
PRooF OF THEOREM 3. By Lemma 3 and the BorelCantelli lemma,
E le11 <
00
>
>
L P{le11 ~ n} < 00
L P{lenl ~ n} < oo
>
P{lenl ~ n i.o.} = 0.
Hence Ien I < n, except for a finite number of n, with probability 1.
Let us put
and suppose that Een
and only if ( 1 + +
Een
= 0, n
1. Then (e 1 + + en)/n. 0 (Pa.s.), if
en )/n . 0 (Pa.s. ). Note that in general Een # 0 but
= Een !(len I<
n)
= Eeti(Ietl
< n) ___. Eet
= 0.
Hence by Toeplitz' lemma
n. oo,
and consequently (e 1 + + en)!n+ 0 (Pa.s.), if and only if
<e~  E~~) ++<en Ee") .,
0
n
Write ~n = ~n
n .
oo
(Pa.s.),
n .
oo. (14)
E~n By Kronecker's lemma, (14) will be established if
I (~jn) converges (Pa.s.). In turn, by Theorem 1 of 2, this will follow if we
show that, when E Ie1 I < 00, the series L (V ~n/n 2 ) converges.
368
IV. Sequences and Sums of Independent Random Variables
We have
"V~n < ~ E~;
L.,
2L.
n=l
00
L E[~iJ(k
1 ~ l~tl < k)].
k=1
00
00
n=k
~ 2 k~1 kE[~if(k 1 ~ 1~ 1 1 < k)]
L E[1~11J(k 1 ~ l~tl < k)] =
00
~ 2
2E1~11
<
00.
k= 1
This completes the proof of the theorem.
Remark 1. The theorem admits a converse in the following sense. Let
be a sequence of independent identically distributed random
variables such that
~ 1 , ~ 2 , .
~1
+ "' + ~n
="+
C,
with probability 1, where Cis a (finite) constant. Then E I~ 1 I < oo and C
In fact, if Sn/n + C (Pa.s.) then
~n =
n
Sn _ (n n
n
Sn1) n1
1 +
E~ 1
O (Pa.s.)
and therefore P( I~n I > n i.o.) = 0. By the BorelCantelli lemma,
L P(l~1l > n) < oo,
and by Lemma 3 we have E I~ 1 I < oo. Then it follows from the theorem that
C = E~ 1 .
Consequently for independent identically distributed random variables
the conditionE I ~ 1 1 < oo is necessary and sufficient for the convergence (with
probability 1) of the ratio Sn/n to a finite limit.
Remark 2. If the expectation m = E~ 1 exists but is not necessarily finite, the
conclusion (10) of the theorem remains valid.
In fact, let, for example, E~1 < oo and E~i = oo. With C > 0, put
n
s; = L ~J(~; ~ C).
i= 1
369
3. Strong Law of Large Numbers
Then (Pa.s.).
But as C + oo,
Ee 1 I(e 1 ~C)+ Ee 1 = oo;
therefore Sn/n +
+ oo (Pa.s.)
4. Let us give some applications of the strong law oflarge numbers.
1 (Application to number theory). Let Q = [0, 1), let 11 be the
algebra of Borel subsets of Q and let P be Lebesgue measure on [0, 1). Consider the binary expansions OJ=O. OJ 1 0J 2 . of numbers OJ e Q (with infinitely
many O's) and define random variables 1(OJ), 2(OJ), ... by putting en(OJ) = OJn.
Since, for all no~ 1 and all x 1 , .. , xn taking the values 0 or 1,
EXAMPLE
e e
the Pmeasure of this set is 1/2n. It follows that e 1 , en, ... is a sequence of
independent identically distributed random variables with
P<el =
o> =
P<el = 1) = !.
Hence, by the strong law of large numbers, we have the following result of
Borel: almost every number in [0, 1) is normal, in the sense that with probability
1 the proportion of zeros and ones in its binary expansion tends to ! , i.e.
 L I(ek =
n k=t
1)+! (Pa.s.).
EXAMPLE 2 (The Monte Carlo method). Let f(x) be a continuous function
defined on [0, 1], with values on [0, 1]. The following idea is the foundation
of the statistical method of calculating JAf(x) dx (the "Monte Carlo
method").
Let e 1 , '7 1 , e 2 , '7 2 , . be a sequence of independent random variables,
uniformly distributed on [0, 1]. Put
P =
It is clear that
{1
if f(O > '7;.
if
< 'li
f(ei>
370
IV. Sequences and Sums of Independent Random Variables
By the strong law of large numbers (Theorem 3)
1 n
(1
n;~/; + Jo f(x) dx
(Pa.s.).
Consequently we can approximate an integral f~ f(x) dx by taking a
simulation consisting of a pair of random variables(~;, '7;), i ;;:: 1, and then
1 P;
calculating P; and (1/n)
L7=
5. PROBLEMS
1. Show that
Ee
<
00
eI > n) <
if and only if L:'=t nP( I
00.
ee
2. Supposing that 1, 2 , are independent and identically distributed, show that if
E Ietl" < 00 for some IX, 0 < IX < 1, then S./n 11"> 0 (Pa.s.), and if EI~tiP < 00 for
some {J, 1 ~ fJ < 2, then (S.  nE~ 1 )/n 11 P+ 0 (Pa.s.).
ee
3. Let 1, 2 , be a sequence of independent identically distributed random variables
and let E I ~ 1 1 = oo. Show that
~I~
I
a.
oo
(Pa.s.)
for every sequence of constants {a.}.
4. Show that a rational number on [0, 1) is never normal (in the sense of Example 1,
Subsection 4).
4. Law of the Iterated Logarithm
1. Let ~ 1 , ~ 2 , be a sequence of independent Bernoulli random variables
with P(~n = 1) = P(~n = 1) = t; let Sn = ~ 1 + + ~n It follows from
the proof of Theorem 2, 1, that
(1)
with probability 1. On the other hand, by (3.11),
JnnSnlog n + 0
(Pa.s.).
(2)
Let us compare these results.
It follows from (1) that with probability 1 the paths of (Sn)n~ 1 intersect
the "curves"
infinitely often for any given but at the same time (2)
eJn
e;
~4.
371
Law of the Iterated Logarithm
shows that they only finitely often leave the region bounded by the curves
eJn log n. These two results yield useful information on the amplitude
of the oscillations of the symmetric random walk (Sn)n ~ 1 The law of the
iterated logarithm, which we present below, improves this picture of the
amplitude of the oscillations of (Sn)n> 1 .
Let us introduce the following definition. We call a function cp* = cp*(n),
n ;;::: 1, upper (for (Sn)n~ 1) if, with probability 1, Sn ~ cp*(n) for all n from
n = n 0 (w) on.
Wecallafunctioncp* = cp*(n),n;;::: 1,lower(for(Sn)n~ 1 )if,withprobability
1,
> cp*(n) for infinitely many n.
Using these definitions, and appealing to (1) and (2), we can say that every
function cp* = eJn log n, e > 0, is upper, whereas cp* = eJn is lower, e > 0.
Let cp = cp(n) be a function and cp: = (1 + e)cp, cp*" = (1  e)cp, where
e > 0. Then it is easily seen that
sn
{rrm (/)~~> ~ 1} = {1i~ [~~~ (/)~:>] ~ t}
:> { sup
m~n,(<)
:.
S(m)
cp m
{Sm ~ (1
~ 1 + e for every e > 0, from some n (e) on}
1
+ e)cp(m) for every e >
0, from some n 1(e) on}.
(3)
In the same way,
{11m(/)~~) ; : : 1} = {li~ [~~~ (/)~:)] ;;::: 1}
:>{sup S(m)
m2:n2(<) cp m
~ 1 + e.foreverye > O,.fromsomen (e)on}
1
:> {Sm ;;::: (1  e)cp(m) for every e > 0 and for infinitely
many m larger than some n 3 (e) ;;::: n 2 (e)}.
(1
(4)
It follows from (3) and (4) that in order to verify that each function cp: =
+ e)cp, e > 0, is upper, we have to show that
(5)
But to show that cp*. = (1  e)cp, e > 0, is lower, we have to show that
(6)
372
IV. Sequences and Sums of Independent Random Variables
2. Theorem 1 (Law of the Iterated Logarithm). Let~ 1 , ~ 2 , . be a sequence of
independent identically distributed random variables with E~; = 0 and E~t =
CJ 2 > 0. Then
(7)
where
tf;(n) = J2CJ 2 n log log n.
(8)
For uniformly bounded random variables, the law of the iterated logarithm
was established by Khinchin (1924). In 1929 Kolmogorov generalized this
result to a wide class of independent variables. Under the conditions of
Theorem 1, the law of the iterated logarithm was established by Hartman
and Wintner (1941).
Since the proof of Theorem 1 is rather complicated, we shall confine
ourselves to the special case when the random variables ~n are normal,
~n " ' %(0, 1), n ~ 1.
We begin by proving two auxiliary results.
Lemma 1. Let~ 1, . , ~n be independent random variables that are symmetrically
distributed (P( ~k E B) = P( ~k E B) for every B E 84 (R), k ::;; n). Then for
every real number a
P( max
Sk >
1'5k:5,n
a) : ; 2P(S" > a).
(9)
PROOF. Let A= {max1:5k:5n sk >a}, Ak = {S;::;; a, i::;; k 1; sk >a} and
B = {Sn > a}. Since S" > a on Ak (because Sk ::;; S"), we have
P(B n Ak) ~ P(Ak n {S" ~ Sd) = P(Ak)P(S" ~ Sk)
= P(Ak)P(~k+ 1 + + ~n ~ 0).
By the symmetry of the distributions of the random variables ~ 1, ... , ~n,
we have
P(~k+ 1 +
Hence P(~k+ 1 +
+ ~n >
~n > 0) = P(~k+ 1
+ + ~n <
0).
0) ~ !,and therefore
P(B) ~ k~1 P(Ak n B) ~
2k~1 P(Ak) =
2 P(A),
which establishes (9).
Lemma 2. Let Sn ,...., JV (0, CJ 2 (n)), CJ 2 (n) j oo, and let a(n), n ~ 1, satisfy
a(n)/CJ(n).. oo, n.. oo. Then
P(Sn > a(n)) ,....,
CJ(n)
11::..
y
2na(n)
exp{ ta 2 (n)/CJ 2 (n)}.
(10)
373
4. Law of the Iterated Logarithm
The proof follows from the asymptotic formula
1
faa e  y1/2 dy  1 e x2/2
fox
fox
X+ 00,
since S,Ju(n) ,.... .Af(O, 1).
PROOF OF THEOREM 1 (for~; "' .Af(O, 1)).
Let us first establish (5). Let e > 0, A.= l + e, nk =A.\ where k ~ k0 , and
k0 is chosen so that In In k0 is defined. We also define
Ak = {Sn > A.t/l(n) for some n E (nk, nk+ tJ},
(11)
and put
A = {Ak i.o.}
= {Sn > A.t/J(n) for infinitely many n}.
In accordance with (3), we can establish (5) by showing that P(A) = 0.
Let us show that
P(Ak) < oo. Then P(A) = 0 by the BorelCantelli
lemma.
From (11), (9) and (10) we find that
P(Ak)
P{Sn > A.t/J(nk) for some n E (nk, nk+ 1)}
~ P{S"
> A.t/J(nk) for some n
nk+ tl
~ 2P{Snk+1 > A.t/J(nk)} "'fo~nk) exp{ tA.2[t/J(nk)/AJl}
~
C 1 exp( A. In In A.k)
CeJ.Ink = C2 k;.,
where C 1 and C2 are constants. But L::;. 1 k;. < oo, and therefore
L P(Ak) <
oo.
Consequently (5) is established.
We turn now to the proof of(6). In accordance with (4) we must show that,
with A. = 1  e, e > 0, we have with probability 1 that sn ~ A.t/J(n) for infinitely many n.
Let us apply (5), which we just proved, to the sequence ( Sn)n<::.t Then we
find that for all n, with finitely many exceptions, Sn ~ 21/J(n) (Pa.s.).
Consequently if nk = N\ N > 1, then for sufficiently large k, either
snk1 ~ 21/J(nkl)
or
(12)
where Yk = snk  snk1.
Hence if we show that for infinitely many k
lk >
A.t/J(nk)
+ 21/J(nkl),
(13)
374
IV. Sequences and Sums of Independent Random Variables
this and (12) show that (Pa.s.) s.k > A.t/J(nk) for infinitely many k. Take some
).' E (A., 1). Then there is anN > 1 such that for all k
A.'[2(Nk Nk 1) In In Nk] 112 > A.(2Nk In In Nk) 112
+ 2(2Nk 1 In In Nk 1) 112
=A.t/J(Nk) + 21/J(Nk 1).
It is now enough to show that
~
> A.'[2(Nk  Nk 1) In In Nk] 112
(14)
for infinitely many k. Evidently ~ ""' %(0, Nk  Nk 1). Therefore, by
Lemma2,
P{Y. > A.'[2(Nk Nk1) In In Nkr/2}"'
>
1
foX(2 In In Nk) 112
c1
k(A')2
 (In k) 112
e0.'>2JntnNk
c2>
k In k
Since (1/k Ink) = oo, it follows from the second part of the BorelCantelli
lemma that, with probability 1, inequality (14) is satisfied for infinitely many
k, so that (6) is established.
This completes the proof of the theorem.
Remark 1. Applying (7) to the random variables ( s.).;;,; 1, we find that
s.
(15)
tm cp(n) = 1.
It follows from (7) and (15) that the law of the iterated logarithm can be put
in the form
(16)
Remark 2. The law of the iterated logarithm says that for every 8 > 0 each
function t/Ji = (1 + 8)1/1 is upper, and t/1*, = (1  8)1/1 is lower.
The conclusion (7) is also equivalent to the statement that, for each 8 > 0,
3.
P{IS.I
(1  8)1/J(n) i.o.}
1,
P{IS.I
(1
+ 8)1/J(n) i.o.}
o.
PROBLEMS
1. Let ~ 1 , ~ 2 , ... be a sequence of independent random variables
Show that
= 1} = 1.
P{IIm v~
2ln n
with~.~
JV(O, 1).
375
4. Law of the Iterated Logarithm
2. Let ~ 1 , ~ 2 , .. be a sequence of independent random variables, distributed according
to Poisson's law with parameter A. > 0. Show that (independently of A.)
{
Pnm
=~.In Inn =1 } =1.
Inn
3. Let ~ 1 , ~ 2 , .. be a sequence of independent identically distributed random variables
with
0 <IX< 2.
Show that
{ I
}
S lll(lnlnn)
= ell = 1.
p Ilrri "
nil
4. Establish the following generalization of (9). Let ~ 1 , . , ~.be independent random
variables. Levy's inequality
PL~::. [Sk + Jl(S. 
Sk)] >
a}
:$
2P(S. a),
>
S0
0,
holds for every real a, where /1(~) is the median of~, i.e. a constant such that
CHAPTER V
Stationary (Strict Sense) Random
Sequences and Ergodic Theory
1. Stationary (Strict Sense) Random Sequences.
MeasurePreserving Transformations
1. Let (Q, :F, P) be a probability space and~= (~t. ~ 2 , .. ) a sequence of
random variables or, as we say, a random sequence. Let (}k ~denote the sequence
(~k+1 ~k+2 ... ).
Definition l. A random sequence ~ is stationary (in the strict sense) if the
probability distributions of (}k~ and~ are the same for every k 2: 1:
BE PA(R 00 ).
The simplest example is a sequence ~ = (~ 1 , ~ 2 , ..) of independent
identically distributed random variables. Starting from such a sequence, we
can construct a broad class of stationary sequences 11 = (17 t. 17 2 , ) by choosing any Borel function g(x 1, ... , x.) and setting '1k = g(~k ~k+ 1, ... , ~k+ 1).
If ~ = (~ 1 , ~ 2 , .. ) is a sequence of independent identically distributed
random variables with E I ~ 1 1 < oo and E~ 1 = m, the law of large numbers
tells us that, with probability 1,
~1 + ... + ~. __.m,
_____
n ~ oo.
In 1931 Birkhoff obtained a remarkable generalization of this fact for the case
of stationary sequences. The present chapter consists mainly of a proof of
Birkhoff's theorem.
The following presentation is based on the idea of measurepreserving
transformations, something that brings us in contact with an interesting
I. Stationary (Strict Sense) Random Sequences. MeasurePreserving Transformations
377
branch of analysis (ergodic theory), and at the same time shows the connection between this theory and stationary random proceesses.
Let (Q, fF, P) be a probability space.
Definition 2. A transformation T ofQ into n is measurable if, for every A E fF,
T 1 A = {w: TwEA}EfF.
Definition 3. A measurable transformation T is a measurepreserving transformation (or morphism) if, for every A E F,
P(T 1A) = P(A).
LetT be a measurepreserving transformation, T" its nth iterate, and ~ 1 =
n ~ 2, and consider the
sequence~= (~ 1 , ~ 2 , ... ). We claim that this sequence is stationary.
In fact, let A= {w: ~EB} and A 1 = {w: 0 1 ~EB}, where BE~(R 00 ).
Since A = {w: (~ 1 (w), ~ 1 (Tw), .. .) E B}, and A 1 = {w: (~ 1 (Tw), ~ 1 (T 2 w), ...)
E B}, we have wE A 1 if and only if either Tw E A or A 1 = r 1A. But P(T 1A)
= P(A), and therefore P(Ad = P(A). Similarly P(Ak) = P(A) for every
Ak = {w: Ok~ EB}, k ~ 2.
Thus we can use measurepreserving transformations to construct
stationary (strict sense) random variables.
In a certain sense, there is a converse result: for every stationary sequence
~considered on (Q, fF, P) we can construct a new probability space (0, #, P),
a random variable ~ 1 (w) and a measurepreserving transformation f,
such that the distribution of~= {~ 1 (&), ~ 1 (fw), ... } coincides with the
distribution of~.
In fact, take 0 to be the coordinate space Roo and put :fft = BI(R 00 ),
P = P~, where P~(B) = P{w: ~ E B}, BE BH(R 00 ). The action of f on Q is
given by
~ 1 (w) a random variable. Put ~k(w) = ~ 1 (Tn 1 w),
If w = (xt. x 2 ,
),
put
~1(&)
Now let A = {w: (x 1,
x1,
... ,
t
Uw) = ~ 1 (fn 1 w),
n ~ 2.
xk) E B}, BE ~(Rk), and
A = {w:(x 2 ,
...
,xk+ 1 )EB}.
Then the property of being stationary means that
P(A) = P{w:(~ 1 ,
...
,~dEB} = P{w:(~ 2 ,
...
,~k+ 1 )EB} =
P(f 1A),
i.e. Tis a measurepreserving transformation. Since P{w: (~ 1 , , ~k) E B} =
P {w: (~ 1 , ... , ~k)EB} for every k, it follows that~ and~ have the same
distribution.
Here are some examples of measurepreserving transformations.
378
V. Stationary (Strict Sense) Random Sequences and Ergodic Theory
EXAMPLE 1.
Let Q = {w 1 , ... , ron} consist ofn points (a finite number), n 2:: 2,
letffbethecollectionofitssubsets,andlet Tw; = W;+ 1, 1::;; i::;; n 1,and
Tw" = w 1 If P (w;) = 1/n, the transformation T is measurepreserving.
2.1f n = [0, 1), ' = ~([0, 1)), Pis Lebesgue measure, A. E [0, 1),
then Tx = (x + A.) mod 1 and T = 2x mod 1 are both measurepreserving
transformations.
ExAMPLE
2. Let us pause to consider the physical hypotheses that led to the consideration of measurepreserving transformations.
Let us suppose that Q is the phase space of a system that evolves (in discrete
time) according to a given law of motion. If w is the state at instant n = 1, then
Tnw, where Tis the translation operator induced by the given law of motion,
is the state attained by the system after n steps. Moreover, if A is some set of
states w then T 1A = { w: Tw E A} is, by definition, the set of states w that
lead to A in one step. Therefore if we interpret Q as an incompressible fluid, the
condition P(T 1 A) = P(A) can be thought of as the rather natural condition
of conservation of volume. (For the classical conservative Hamiltonian systems, Liouville's theorem asserts that the corresponding transformation T
preserves Lebesgue measure.)
3. One of the earliest results on measurepreserving transformations was
Poincare's recurrence theorem (1912).
Theorem 1. Let (Q, ff, P) be a probability space, let T be a measurepreserving
transformation, and let A E F. Then, for almost every point wE A, we have
Tnw E A for infinitely many n 2:: 1.
PROOF. Let C = {wEA: Tnw~A, for all n 2:: 1}. Since C n Tnc = 0 for
all n 2:: 1, we have Tmc n T<m+nlc = Tm(C n TnC) = 0. Therefore
the sequence {TnC} consists of disjoint sets of equal measure. Therefore
L:':o P(C) = L:':o P(TnC)::;; P(O) = 1 and consequently P(C) = 0.
Therefore, for almost every point wE A, for at least one n 2:: 1, we have
Tnw EA. It follows that Tnw E A for infinitely many n.
Let us apply the preceding result to T\ k 2:: 1. Then for every wE A\N,
where N is a set of probability zero, the union of the corresponding sets corresponding to the various values of k, there is an nk such that (Tkrw EA.
It is then clear that rnw E A for infinitely many n. This completes the proof of
the theorem.
Corollary. Let ~(w) 2:: 0. Then
L ~(Tkw) =
00
k=O
on the set {w: ~(w)
> 0}.
oo
(Pa.s.)
379
2. Ergodicity and Mixing
In fact, let An= {w: ~(w) ~ 1/n}. Then, according to the theorem,
= oo (Pa.s.) on An, and the required result follows by letting
n+ oo.
Lk'=o ~(Tkw)
Remark. The theorem remains valid if we replace the probability measure P
by any finite measure Jl with Jl(Q) < oo.
4. PROBLEMS
1. Let Tbeameasurepreservingtransformationand~ = ~(w)arandom variable whose
expectation E~(w) exists. Show that E~(w) = E~(Tw).
2. Show that the transformations in Examples 1 and 2 are measurepreserving.
3. Let n = [0, 1), F = af([O, 1)) and let P be a measure whose distribution function is
continuous. Show that the transformations Tx = ..l.x, 0 < A. < 1, and Tx = x 2 are not
measurepreserving.
2. Ergodicity and Mixing
1. In the present section T denotes a measurepreserving transformation on
the probability space (Q, fl', P).
Definition 1. A set A E fJ' is invariant if T 1A = A. A set A E fJ' is almost
invariant if A and r 1A differ only by a set of measure zero, i.e. P(A 6. r 1 A)
=Q
.
It is easily verified that the classes J and J* of invariant or almost invariant sets, respectively, are ualgebras.
Definition 2. A measurepreserving transformation Tis ergodic (or metrically
transitive) if every invariant set A has measure either zero or one.
Definition 3. A random variable~ = ~(w) is invariant (or almost invariant) if
~(w) = ~(Tw) for all wEn (or for almost all wE Q).
The following lemma establishes a connection between invariant and
almost invariant sets.
Lemma 1. If A is almost invariant, there is an invariant set B such that
P(A!::. B) = 0.
PRooF. Let B = Ilm rnA. Then T 1B = Ilm r<n+ 1 >A = B, i.e. B EJ.It is
easily seen that A!::. B s;;; Uk'=o (TkA!::. r<k+ 1>A). But
P(TkA!::. r<k+ 1 >A)
Hence P(A !'::!,. B) = 0.
= P(A 6. T 1A) = 0.
380
V. Stationary (Strict Sense) Random Sequences and Ergodic Theory
Lemma 2. A transformation T is ergodic if and only if every almost invariant set
has measure zero or one.
PROOF. Let A E *; then according to Lemma 1 there is an invariant set B
such that P(A 6. B) = 0. But T is ergodic and therefore P(B) = 0 or 1.
Therefore P(A) = 0 or 1. The converse is evident, since J J*. This completes the proof of the lemma.
Theorem 1. Let T be a measurepreserving transformation. Then the following
conditions are equivalent:
(1) Tis ergodic;
(2) every almost invariant random variable is (Pa.s.) constant;
(3) every invariant random variable is (Pa.s.) constant.
PROOF. (1) <o> (2). Let T be ergodic and ~ almost invariant, i.e. (Pa.s.) ~( w) =
~(Tw). Then for every c E R we have Ac = {w: ~(w) ~ c} E J*, and then
P(AJ = 0 or 1 by Lemma 2. Let C = sup{c: P(Ac) = 0}. Since Ac j Q as
c j oo and Ac! 0 as c! w, we have ICI < oo. Then
P{w:
~(w) < C} =
P{.91
{~(w) ~ C ~}} =
and similarly P{w: ~(w) > C} = 0. Consequently P{w:
(2) => (3). Evident.
(3)
=>
(1 ). Let A
~(w)
= C} = 1.
J; then I A is an invariant random variable and therefore,
(Pa.s.), IA = 0 or IA = 1, whence P(A) = 0 or 1,
Remark. The conclusion of the theorem remains valid in the case when
"random variable" is replaced by "boundedrandom variable".
We illustrate the theorem with the following example.
ExAMPLE. Let Q = [0, 1), ' = .?4([0, 1)), let P be Lebesgue measure and let
Tw = (w + A.) mod 1. Let us show that T is ergodic if and only if A. is irrational.
Let~ = ~(w) be a random variable with Ee(w) < oo. Then we know that
the Fourier series L~ _"' c.e 2 "inw of ~(w) converges in the mean square sense,
L Ic. 12 < oo, and, because T is a measurepreserving transformation
(Example 2, 1), we have (Problem 1, 1) that for the random variable~
c.E~(w)e2xin~<w>
=
=
E~(Tw)e2xinTw
ez";";.E~(w)e2xinw
e2xin.i.E~(Tw)ez"inw
= c.ez";";..
So c.(1  e 2 ";";.) = 0. By hypothesis, A. is irrational and therefore e2 "in.l. # 1
for all n # 0. Therefore c. = 0, n # 0, ~(w) = c0 (Pa.s.), and Tis ergodic by
Theorem 1.
381
3. Ergodic Theorems
On the other hand, let A. be rational, i.e. A.
integers. Consider the set
A =
U
k=O
2m2 {
= k/m,
where k and m are
1}
w: 2 ~ w < 2 .
m
m
ltisclearthatthissetisinvariant;butP(A) =
t. Consequently Tis not ergodic.
2. Definition 4. A measurepreserving transformation is mixing (or has the
mixing property) if, for all A and BE F,
lim P(A n ynB)
= P(A)P(B).
(1)
n+ oo
The following theorem establishes a connection between ergodicity and
mixing.
Theorem 2. Every mixing transformation T is ergodic.
PRooF.
Let A E ~. BE ..F. Then B = ynB, n ;::: 1, and therefore
P(A n ynB) = P(A n B)
for all n ;::: 1. Because of (1), P(A n B) = P(A)P(B). Hence we find, when
A = B, that P(B) = P 2 (B), and consequently P(B) = 0 or 1. This completes
the proof.
3.
PROBLEMS
1. Show that a random variable is invariant if and only if it is Jmeasurable.
2. Show that a set A is almost invariant if and only if either
P(T 1A\A) = 0
or
3. Show that the transformation considered in the example of Subsection 1 of the
present section is not mixing.
4. Show that a transformation is mixing if and only if, for all random variables and 11
with Ee 2 < oo and E11 2 < oo,
3. Ergodic Theorems
1. Theorem 1 (Birkhoff and Khinchin). Let T be a measurepreserving trans~(w) a random variable withE I~ I < oo. Then (Pa.s.)
formation and~ =
(1)
382
V. Stationary (Strict Sense) Random Sequences and Ergodic Theory
lf also Tis ergodic then (Pa.s.)
1 n1
limn
L ~(T"w) =
n k=O
(2)
E~.
The proof given below is based on the following proposition, whose simple
proof was given by A. Garsia (1965).
Lemma (Maximal Ergodic Theorem). Let T be a measurepreserving transformation, let~ be a random variable with E I~ I < oo, and let
Sk(w)
~(w)
Mk(w)
+ ~(Tw) + + ~(T" 1 w),
= max{O, S 1(ro), ... , S"(ro)}.
Then
for every n
PRooF. Ifn ~
~(w)
1.
k, we have Mn(Tw) ~ Sk(Tw) and therefore ~(ro) + Mn(Tw) ~
Sk+ 1(w). Since it is evident that ~(ro) ~ S 1 (ro) Mn(Tw),
+ Sk(Tw) =
we have
Therefore
E[c;(ro)/1M">olro)] ~ E(max(S 1 (ro), ... , Sn(w))  Mn(Tw)),
But max(Sl> ... , Sn) = Mn on the set {Mn > 0}. Consequently,
E[~(ro)/{Mn>O}(ro)] ~ E{(Mn(ro) Mn(Tro))/{Mn(w)>OJ}
~
E{Mn(ro) Mn(Tw)}
= 0,
since if T is a measurepreserving transformation we have EMn(ro)
EM,.(Tw) (Problem 1, 1).
This completes the proof of the lemma.
PRooF OF THE
THEOREM. Let us suppose that E(~l.f) = O(otherwise replace
~by~ E(~l.f)).
Let tf
(Pa.s.)
= Ilm(SJn)
and!!.
= lim(SJn). It
0
$;
!!.
$;
will be enough to establish that
'1 $; 0.
Consider the random variable i7 = q(w). Since q(ro) = q(Tw), the variable i7 is
invariant and consequently, for every e > 0, the set A. = {q(ro) > e} is also
invariant. Let us introduce the new random variable
and put
St(w) = ~*(ro)
+ .. + ~cr" 1 ro),
Mt(w)
max(O, Sf, ... , St).
383
3. Ergodic Theorems
Then, by the lemma,
E[~*J 1 M~>oa ~
for every n
1. But as n+ oo,
{M~ > 0} = { maxSt > o} j {sup St > o} ={sup Skt > o}
= {sup
k;;>: 1
k<: 1
k<': 1
1 ,;;k,;;n
~k > e} nA, = A.,
where the last equation follows because supk,., 1(S:/k) ~ ij, and A,=
{w: ij > e}.
Moreover, El~*l ~ El~l +e. Hence, by the dominated convergence
theorem,
Thus
~ E[~*IAJ
= E[(~  e)IAJ = E[UAJ  eP(A,)
= E[E<ei.Y)IAJ  eP(A,) = eP(A,),
so that P(A,) = 0 and therefore P(ij ~ 0) = 1.
Similarly, if we consider ~(w) instead of ~(w), we find that
=(
11111
Sn) =  1l. mSn = 1]

and P( 17 ~ 0) = 1, i.e. P(17 ~ 0) = 1. Therefore 0 ~ 11 ~ ij ~ 0 (Pa.s.) and
the first part of the theorem is established.
To prove the second part, we observe that ~ince E( ~ I.Y) is an invariant
random variable, we have E(~IJ) = E~ (Pa.s.) in the ergodic case.
This completes the proof of the theorem.
Corollary. A measurepreserving transformation T is ergodic
all A and BE f7,
1 n1
lim L P(A n rkB) = P(A)P(B).
n
n k=O
if and only if, for
(3)
To prove the ergodicity of Twe use A = BE J in (3). Then A n T kB = B
and therefore P(B) = P 2 (B), i.e. P(B) = 0 or 1. Conversely, let T be ergodic.
Then if we apply (2) to the random variable~ = Ja(w), where BE f7, we find
that (Pa.s.)
1 n1
lim L ITkiw) = P(B).
n
n k=O
384
V. Stationary (Strict Sense) Random Sequences and Ergodic Theory
If we now integrate both sides over A
theorem, we obtain (3) as required.
fF and use the dominated convergence
2. We now show that, under the hypotheses of Theorem 1, there is not only
almost sure convergence in (1) and (2), but also convergence in mean. (This
result will be used below in the proof of Theorem 3.)
Theoreml. LetT be a measurepreserving transformation and let
random variable with EI I < 00. Then
~~ :t>(Tkw) E(el~ I. 0,
e=
e(w) be a
n. oo.
(4)
If also T is ergodic, then
n. oo.
(5)
PRooF. For every e > 0 there is a bounded random variable
such that E1e '71:::;; e. Then
'7( I'7(w) I :::;; M)
(6)
Since I'71 :::;; M, then by the dominated convergence theorem and by using (1)
we find that the second term on the right of (6) tends to zero as n. oo. The
first and third terms are each at most e. Hence for sufficiently large n the lefthand side of(6) is less than 2e, so that (4) is proved. Finally, if Tis ergodic, then
(5) follows from (4) and the remark that E(eii) = Ee (Pa.s.).
This completes the proof of the theorem.
3. We now turn to the question of the validity of the ergodic theorem for
)defined on a probstationary (strict sense) random sequences = <el>
ability space (Q, !F, P). In general, (Q, !F, P) need not carry any measurepreserving transformations, so that it is not possible to apply Theorem 1
directly. However, as we observed in 1, we can construct a coordinate probability space (fi, ;, P), random variables ~ = (~ 1 , ~ 2 , . ), and a measurepreserving transformation T such that ~n(w) = ~ 1 (T" 1 w) and the distributions of and ~ are the same. Since such properties as almost sure
convergence and convergence in the mean are defined only for proba1w) (Pa.s.
bility distributions, from the convergence of (1/n)
1 ~1
and in mean) to a random variable ij it follows that (1/n)
also
1
converges (Pa.s. and in mean) to a random variable '1 such that '1 :4: ij. It
ez, .. .
Li=
(Tk
:D= ek<w)
385
3. Ergodic Theorems
follows from Theorem 1 that if EI~ 1 I < oo then = E( ~ 1 If), where .J is a
collection of invariant sets (E is the average with respect to the measure P). We
now describe the structure of '1
Definition 1. A set A E i' is invariant with respect to the sequence~ if there is a
set BE 86(R oo) such that for n ~ 1
A = {w:
(~., ~n+ 1, ...) E
B}.
The collection of all such invariant sets is a aalgebra, denoted by
J~.
Definition 2. A stationary sequence ~ is ergodic if the measure of every invariant set is either 0 or 1.
Let us now show that the random variable 1J can be taken equal toE(~ 1 IJ ~).
In fact, let A E f ~. Then since
1 n1
E1~k
n k~ 1
we have
f ~k ~ f
!n I
k~ 1
{w;
dP
(7)
'1 dP.
= {w: (~k ~k+ 1, ... ) E B}
Let B Eei(R 00 ) be such that A
since ~ is stationary,
f ~k dP = Jl
'1 ~ 0,
(~k. ~k+ 1 ....)eB}
~k dP =
Hence it follows from (7) that for all A
that '1 = E( 1 1...';). Here E( 1 1...';) =
f{ro;(~t. ~2
J
~,
... )eB}
for all k ~ 1. Then
~1 dP =
f ~1 dP.
A
which implies (see 7, Chapter II)
E~ 1 if~
is ergodic.
Therefore we have proved the following theorem.
Theorem 3 (Ergodic Theorem). Let ~ = (~ 1 , ~ 2 , .. ) be a stationary (strict
sense) random sequence withE I~ 1 1< oo. Then (Pa.s., and in the mean)
If~
is also an ergqdic sequence, then (Pa.s., and in the mean)
1
lim
L ~k(w) = E~ 1 .
n k~ 1
4.
PROBLEMS
1. Let ~ = (~ 1 , ~ 2 ,
function R(n)
ergodic.
. )
be a Gaussian stationary sequence withE~" = 0 and covariance
Show that R(n)+ 0 is a sufficient condition for ~ to be
= E~k+n~k
386
V. Stationary (Strict Sense) Random Sequences and Ergodic Theory
2. Show that every sequence ~ =
random variables is ergodic.
(~ 1 , ~ 2 , . )
3. Show that a stationary sequence
for every BE [J(R), k = 1,2, ....
of independent identically distributed
is ergodic if and only if
CHAPTER VI
Stationary (Wide Sense) Random
Sequences. L 2  Theory
1. Spectral Representation of the
Covariance Function
1. According to the definition given in the preceding chapter, a random
sequence ~ = (~ 1 , ~ 2 , ..) is stationary in the strict sense if, for every set
BE&l(R 00 )andeveryn ~ 1,
P{(~l' ~2 ... )EB} = P{(~n+l ~n+2 ... )EB}.
(1)
It follows, in particular, that if E~i < oo then E~n is independent of n:
E~n
and the covariance cov(~n+m~n) =
onm:
E~1'
E(~n+m E~n+m)(~n  E~n)
(2)
depends only
(3)
In the present chapter we study sequences that are stationary in the wide
sense (and have finite second moments), namely those for which (1) is
replaced by the (weaker) conditions (2) and (3).
The random variables ~n are understood to be defined for n E 7L. =
{0, 1, ... } and to be complexvalued. The latter assumption not only does
not complicate the theory, but makes it more elegant. It is also clear that
results for real random variables can easily be obtained as special cases of the
corresponding results for complex random variables.
Let H 2 = H 2 (Q, Jl', P) be the space of (complex) random variables
2 < oo, where I~ 1
2 = r~. 2 + {3 2 If ~ and
~ = r:1. + i{3, r:J., {3 E R, with E I~ 1
17 E H 2 , we put
(4)
388
where
VI. Stationary (Wide Sense) Random Sequences. L 2 Theory
r; =
(J( 
if3 is the complex conjugate of 11 =
(J(
+ if3 and
WI = (~, 0 112
(5)
As for real random variables, the space H 2 (more precisely, the space of
equivalence classes of random variables; compare 10 and 11 of Chapter II)
is complete under the scalar product ( ~, 11) and norm II~ 11. In accordance with
the terminology of functional analysis, H 2 is called the complex (or unitary)
Hilbert space (of random variables considered on the probability space
(Q, !F, P)).
If~' 11 E H 2 their covariance is
11) =
cov(~,
E(~  E~)(11
 E11).
(6)
It follows from (4) and (6) that if E~ = E11 = 0 then
cov(~,
11) = (~, 11).
(7)
Definition. A sequence of complex random variables ~ = (~")" E 2 with
E I~" 12 < oo, n E 7l., is stationary (in the wide sense) if, for all n E 7l.,
E~" = E~ 0 ,
cov(~k+n ~k)
k E 7i..
cov(~"' ~ 0 ),
(8)
As a matter of convenience, we shall always suppose that E~ 0 = 0. This
involves no loss of generality, but does make it possible (by (7)) to identify the
covariance with the scalar product and hence to apply the methods and results
of the theory of Hilbert spaces.
Let us write
R(n) =
cov(~"' ~ 0 ),
n E 7l.,
(9)
and (assuming R(O) = E I ~ 0 12 # 0)
R(n)
p(n) = R(O)'
n E 7l..
(10)
We call R(n) the covariance function, and p(n), the correlationfunction, of the
sequence~ (assumed stationary in the wide sense).
It follows immediately from (9) that R(n) is nonnegativedefinite, i.e. for all
complex numbers a 1, ... , am and t 1 , .. , tm E 7l., m ;::::: 1, we have
m
L aJijR(t; 
t) ;: : : 0.
(11)
,i,j= 1
It is then easy to deduce (either from (11) or directly from (9)) the following
properties of the covariance function (see Problem 1):
R(O);::::: 0,
R( n) = R(n),
IR(n) I :s; R(O),
IR(n) R(mW :s; 2R(O)[R(O) Re R(n m)].
(12)
389
1. Spectral Representation of the Covariance Function
2. Let us give some examples of stationary sequences ~ = (~n)nEZ (From
now on, the words "in the wide sense" and the statement n E Z will both be
omitted.)
ExAMPLE 1. Let ~n = ~ 0 g(n), where E~ 0 = 0, E~6 = 1 and g = g(n) is a
function. The sequence~ = (~n) will be stationary if and only if g(k + n)g(k)
depends only on n. Hence it is easy to see that there is a Asuch that
g(n) = g(O)ei;.n.
Consequently the sequence of random variables
~n = ~0
g(O)eiln
is stationary with
In particular, the random
EXAMPLE
"constant"~
= 0 is a stationary sequence.
~
2. An almost periodic sequence. Let
(13)
where z 1, . , zN are orthogonal (Ez;zi = 0, i =F j) random variables with zero
means and E lzkl 2 =a~ > 0; n ~ Ak < n, k = 1, ... , N; A; =F Ai, i =F j.
The sequence~ = (~n) is stationary with
N
L a~eiJ.n.
R(n) =
(14)
k=l
As a generalization of (13) we now suppose that
~n
L
00
k=
zkei).n,
(15)
00
where zk, k E Z, have the same properties as in (13). If we suppose that
oo a~ < oo, the series on the right of (15) converges in meansquare and
Lr'= _
R(n)
L
00
k=
a~e;;.n.
(16)
00
Let us introduce the function
F(A) =
af.
(17)
{k: Ak:5 ).)
Then the covariance function (16) can be written as a LebesgueStieltjes
integral,
R(n)
r"
e;;.n dF(A.).
(18)
390
VI. Stationary (Wide Sense) Random Sequences. L 2 Theory
The stationary sequence (15) is represented as a sum of "harmonics" ei).kn
with "frequencies" Ak and random "amplitudes" zk of "intensities" (Jl =
E jzkj 2. Consequently the values of F(A) provide complete information on the
"spectrum" of the sequence ~' i.e. on the intensity with which each frequency
appears in (15). By (18), the values of F(A) also completely determine the
structure of the covariance function R(n).
Up to a constant multiple, a (nondegenerate) F(A) is evidently a distribution function, which in the examples considered so far has been piecewise
constant. It is quite remarkable that the covariance function of every stationary
(wide sense) random sequence can be represented (see the theorem in Subsection 3) in the form (18), where F(A) is a distribution function (up to normalization), whose support is concentrated on [ n, n), i.e. F(A) = 0 for
A < nand F(A) = F(n) for A > n.
The result on the integral representation of the covariance function, if
compared with (15) and (16), suggests that every stationary sequence also
admits an "integral" representation. This is in fact the case, as will be shown in
3 by using what we shall learn to call stochastic integrals with respect to
orthogonal stochastic measures (2).
EXAMPLE 3 (White noise). Let e = (en) be an orthonormal sequence of random
variables, Een = 0, Eeiei = bii' where bii is the Kronecker delta. Such a
sequence is evidently stationary, and
R(n) =
{10,
0,
n=
n # 0.
Observe that R(n) can be represented in the form
R(n) =
f/""
(19)
dF(A),
where
F(A)
f/(v) dv;
j(A) = 2n'
n::;; A< n.
(20)
Comparison of the spectral functions (17) and (20) shows that whereas the
spectrum in Example 2 is discrete, in the present example it is absolutely
continuous with constant "spectral density" j(A) tn. In this sense we can
say that the sequence e = (en) "consists of harmonics of equal intensities." It
is just this property that has led to calling such a sequence e = (en) "white
noise" by analogy with white light, which consists of different frequencies with
the same intensities.
EXAMPLE 4 (Moving averages) Starting from the white noise e = (en) introduced in Example 3, let us form the new sequence
C()
~" =
2: akenk,
k=oo
(21)
391
I. Spectral Representation of the Covariance Function
where ak are complex numbers such that
equation,
Lk'= _
oo I ak 12
<
rx:;.
By Parseval's
00
cov(~n+m ~m) = cov(~"' ~ 0 ) =
k=
an+kak,
00
so that~ = (~k) is a stationary sequence, which we call the sequence obtained
from e = (ek) by a (twosided) moving average.
In the special case when the ak of negative index are zero, i.e.
00
~"
the sequence~ =
k > p, i.e. if
(~n)
L akenk
k=O
is a onesided moving average. If, in addition, ak = 0 for
(22)
then ~ = (~") is a moving average of order p.
We can show (Problem 5) that (22) has a covariance function of the form
R(n) =
e;;."f(A.) dA., where the spectral density is
J':.,
(23)
with
P(z) = a0
+ a1 z + + aPzP.
EXAMPLE 5 (Autoregression). Again let e = (en) be white noise. We say that a
random sequence~ = ( ~n) is described by an autoregressive model of order q if
~n
+ bl~n1 + ''' + bq~nq =
Sn.
(24)
Under what conditions on b 1 , , bn can we say that (24) has a stationary
solution? To find an answer, let us begin with the case q = 1:
(25)
where a.= b 1 . If la.l < 1, it is easy to verify that the stationary sequence
~ = (~n) with
L a.ien00
~n =
(26)
j=O
is a solution of (25). (The series on the right of (26) converges in meansquare.)
Let us now show that, in the class of stationary sequences ~ = (~") (with
finite second moments) this is the only solution. In fact, we find from (25),
by successive iteration, that
392
VI. Stationary (Wide Sense) Random Sequences. L 2 Theory
Hence it follows that
k+
00.
Therefore when Ia I < 1 a stationary solution of (25) exists and is representable as the onesided moving average (26).
There is a similar result for every q > 1 : if all the zeros of the polynomial
(27)
lie outside the unit disk, then the autoregression equation (24) has a unique
stationary solution, which is representable as a onesided moving average
(Problem 2). Here the covariance function R(n) can be represented (Problem
5) in the form
R(n) =
f/An dF().),
F().) =
f/(v)
dv,
(28)
where
(29)
In the special case q = 1, we find easily from (25) that Eeo = 0,
2
Eeo
= t 
lo::l2'
and
R(n)
a"
= 1lo::l2'
n~O
(when n < 0 we have R(n) = R( n)), Here
1
1
J().) = 2n '11  ae;;.1 2 '
EXAMPLE 6.
This example illustrates how autoregression arises in the construction of probabilistic models in hydrology. Consider a body of water;
we try to construct a probabilistic model of the deviations of the level ofthe
water from its average value because of variations in the inflow and evaporation from the surface.
If we take a year as the unit of time and let H n denote the water level in
yearn, we obtain the following balance equation:
Hn+ 1 = Hn  KS(H")
+ ~n+ 1
(30)
where ~n+ 1 is the inflow in year (n + 1), S(H) is the area of the surface of the
water at level H, and K is the coefficient of evaporation.
393
I. Spectral Representation of the Covariance Function
Let ~n = Hn  H be the deviation from the mean level (which is obtained
from observations over many years) and suppose that S(H) = S(H) +
c(H  H). Then it follows from the balance equation that ~n satisfies
(31)
with a = 1  cK, en = :En  KS(H). It is natural to assume that the random
variables ~>n have zero means and are identically distributed. Then, as we
showed in Example 5, equation (31) has (for Ia I < 1) a unique stationary
solution, which we think of as the steadystate solution (with respect to time
in years) of the oscillations of the level in the body of water.
As an example of practical conclusions that can be drawn from a
(theoretical) model (31 ), we call attention to the possibility of predicting the
level for the following year from the results of the observations of the present
and preceding years. It turns out (see also Example 2 in 6) that (in the meansquare sense) the optimal linear estimator of ~n+ 1 in terms of the values of
. , ~nl ~n is simply O!~n
7 (Autoregression and moving average (mixed model)). If we
suppose that the righthand side of (24) contains a0 Bn + a 1Bn 1 + +
apsn P instead of en, we obtain a mixed model with autoregression and
moving average of order (p, q):
ExAMPLE
~n
+ b1~n1 + + bq~nq = aoBn + a1Bn1 + + apBnp
(32)
Under the same hypotheses as in Example 5 on the zeros it will be shown later
(Corollary 2 to Theorem 3 of3) that (32) has the stationary solution~ = (~")
e;;.n dF(A.) with F(A.) =
for which the covariance function is R(n) =
J~,f(v) dv, where
J':,
3. Theorem (Herglotz). Let R(n) be the covariance function of a stationary
(wide sense) random sequence with zero mean. Then there is, on
([ n, n), g6([ n, n))),
a finite measure F = F(B),
BE g6([ n,
R(n) =
PRooF. For N
n)), such that for every n E 7L
f/;." F(dA.).
(33)
1 and A. E [ n, n], put
fN(A.) = 2 N
1t
I I
k=1 1=1
R(k  T)e ik;.ew'.
(34)
394
VI. Stationary (Wide Sense) Random Sequences. L 2 Theory
Since R(n) is nonnegative definite, fN(A.) is nonnegative. Since there are
N  Im I pairs (k, I) for which k  l = m, we have
fN(A.) = __!_
2n lmi<N
(1  ~)R(m)eim'.
N
(35)
Let
BE~([ n,
n)).
Then
lnl <
lnl
N,} (36)
~N.
The measures F N N ~ 1, are supported on the interval [ n, n] and
F N([ n, n]) = R(O) < oo for all N ~ 1. Consequently the family of measures
{FN}, N ~ 1, is tight, and by Prohorov's theorem {Theorem 1 of 2 of
Chapter III) there are a sequence {Nk}
{N} and a measure F such that
F Nk ~ F. {The concepts of tightness, relative compactness, and weak
convergence, together with Prohorov's theorem, can be extended in an
obvious way from probability measures to any finite measures.)
It then follows from (36) that
f/'"nF(dA.)
N~~oo f~/'npNk(dA.) =
R(n).
The measure F so constructed is supported on [ n, n]. Without changing
the integral J~oo ei'nF(dA.), we can redefine F by transferring the "mass"
F( { n} ), which is concentrated at n, to n. The resulting new measure (which
we again denote by F) will be supported on [ n, n).
This completes the proof of the theorem.
Remark 1. The measure F = F(B) involved in (33) is known as the spectral
measure, and F(A.) = F([ n, A.]) as the spectral function, of the stationary
sequence with covariance function R(n).
In Example 2 above the spectral measure was discrete (concentrated at
A.k, k = 0, 1, ...). In Examples 36 the spectral measures were absolutely
continuous.
Remark 2. The spectral measure F is uniquely defined by the covariance
function. In fact, let F 1 and F 2 be two spectral measures and let
f/'nFl(dA.)
= f/'"nFz(dA.),
n E Z.
2. Orthogonal Stochastic Measures and Stochastic Integrals
395
Since every bounded continuous function g(A.) can be uniformly approximated on [ n, n) by trigonometric polynomials, we have
{,.g(A.)F 1(dA.) = {,.g(A.)FidA.).
It follows (compare the proof in Theorem 2, 12, Chapter II) that F 1(B)
=FiB) for all Be&l([ n, n)).
Remark 3. If e=(en) is a stationary sequence of real random variables en,
then
R(n) = {,.cos A.n F(dA.).
4.
PROBLEMS
1. Derive (12) from (11).
2. Show that the autoregression equation (24) has a stationary solution if all the zeros
ofthe polynomial Q(z) defined by (27) lie outside the unit disk.
3. Prove that the covariance function (28) admits the representation (29) with spectral
density given by (30).
4. Show that the sequence
e= (e.) of random variables, where
e. = I <cxk sin A.kn + pk cos A.kn>
<Xl
k=l
and cxk and f3k are real random variables, can be represented in the form
e. = I
<Xl
zkeu
k=oo
with zk = !(f3k icxk) fork~ 0 and zk = z_k, A.k = A.k fork < 0.
5. Show that the spectral functions of the sequences (22) and (24) have densities given
respectively by (23) and (29).
6. Show that if I IR(n) I < oo, the spectral function F(A.) has density f(A.) given by
1
f(A.} =  I ei.<nR(n).
2nn=oo
<Xl
2. Orthogonal Stochastic Measures and Stochastic
Integrals
1. As we observed in 1, the integral representation of the covariance
function and the example of a stationary sequence
en=
00
k=oo
zke;;.
(1)
396
VI. Stationary (Wide Sense) Random Sequences. 2 Theory
with pairwise orthogonal random variables zk, k e 7L, suggest the possibility
of representing an arbitrary stationary sequence as a corresponding integral
generalization of (1).
If we put
(2)
Z(A.) =
zk,
{k:.l.kS.l.}
we can rewrite (1) in the form
co
e. = L
k= co
eilk".1Z(A.k),
(3)
where .1Z(A.k) Z(A.k)  Z(A.k ) = zk.
The righthand side of (3) reminds us of an approximating sum for an
integral j'':., ei.n dZ(A.) of RiemannStieltjes type. However, in the present
case Z(A.) is a random function (it also depends on w). Hence it is clear that
for an integral representation of a general stationary sequence we need to use
functions Z(A.) that do not have bounded variation for each ro. Consequently
the simple interpretation of
eiAn dZ(A.) as a RiemannStieltjes integral
for each w is inapplicable.
J':.,
2. By analogy with the general ideas of the Lebesgue, LebesgueStieltjes
and RiemannStieltjes integrals (6, Chapter II), we begin by defining
stochastic measure.
Let (Q, JF, P) be a probability space, and let E be a subset, with an algebra
S 0 of subsets and the ualgebra S generated by S 0
Definition 1. A complexvalued function Z(.1) = Z(w; .1), defined for wen
and .1 e S 0 , is a finitely additive stochastic measure if
(1) E I Z(.1) 12 < oo for every .1 E S 0 ;
(2) for every pair A 1 and .12 of disjoint sets in tff 0 ,
Z(.11
+ .12) =
Z(.1 1)
+ Z(.12)
(Pa.s.)
(4)
Definition 2. A finitely additive stochastic me
Гораздо больше, чем просто документы.
Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.
Отменить можно в любой момент.