Probability Theory and Stochastic Processes With Applications

Probability Theory
and Stochastic
Processes
with Applications
Oliver Knill
Overseas Press
Probability and Stochastic
Processes with Applications
Probability and Stochastic
Processes with Applications
Oliver Knill
OVERSEAS PRESS (INDIA) PVT. LTD.

Copyright © Oliver Knill
Read. Office:
Overseas Press India Private Limited
7/28, Ansari Road, Daryaganj
NewDelhi-110 002
Email: [email protected]
Website: www.overseaspub.com
Sales Office:
Overseas Press India Private Limited
2/15, Ansari Road, Daryaganj
NewDelhi-110 002
Email: [email protected]
Website: www.overseaspub.com
All rights reserved. No part of this publication may be reproduced or transmitted

in any form or by any means, electronic or mechanical, including photocopying,
recording, or any information storage and retrieval system, without permission in
writing from the publisher/Authors.
Edition : 2009
10 digit ISBN : 81 - 89938 - 40 -1

13 digit ISBN : 978 - 81 - 89938 - 40 - 6
Published by Narinder Kumar Lijhara for Overseas Press India Private Limited,
7/28, Ansari Road, Daryaganj, New Delhi-110002 and Printed in India.
Contents
Preface 3
1 Introduction 5
1.1 What is probability theory? Vt 5
1.2 Some paradoxes in probability theory 12
1.3 Some applications of probability theory 16
2 Limit theorems 23
2.1 Probability spaces, random variables, independence 23
2.2 Kolmogorov's 0 — 1 law, Borel-Cantelli lemma 34
2 . 3 I n t e g r a t i o n , E x p e c t a t i o n , Va r i a n c e 3 9
2.4 Results from real analysis 42
2.5 Some inequalities 44
2.6 The weak law of large numbers 50
2.7 The probability distribution function 56
2.8 Convergence of random variables 59
2.9 The strong law of large numbers 64
2.10 Birkhoff's ergodic theorem 68
2 . 11 More convergence results 72
2.12 Classes of random variables 78
2.13 Weak convergence 90
2.14 The central limit theorem 92
2.15 Entropy of distributions 98
2.16 Markov operators 107
2.17 Characteristic functions 11 0
2 . 1 8 T h e l a w o f t h e i t e r a t e d l o g a r i t h m 11 7
3 Discrete Stochastic Processes 123

3.1 Conditional Expectation 123
3.2 Martingales 131
3.3 Doob's convergence theorem 143
3.4 Levy's upward and downward theorems 150
3.5 Doob's decomposition of a stochastic process 152
3.6 Doob's submartingale inequality 157
3.7 Doob's Cp inequality 159
3.8 Random walks 162
1
Contents
3.9 The arc-sin law for the ID random walk 167
3.10 The random walk on the free group 171
3.11 The free Laplacian on a discrete group . . 175
3.12 A discrete Feynman-Kac formula 179
3.13 Discrete Dirichlet problem 181
3.14 Markov processes 186
Continuous Stochastic Processes 191

4.1 Brownian motion 191
4.2 Some properties of Brownian motion 198
4.3 The Wiener measure 205
4.4 Levy's modulus of continuity 207
4.5 Stopping times 209
4.6 Continuous time martingales 215
4.7 Doob inequalities 217
4.8 Khintchine's law of the iterated logarithm 219
4.9 The theorem of Dynkin-Hunt 222
4.10 Self-intersection of Brownian motion 223
4 . 11 R e c u r r e n c e o f B r o w n i a n m o t i o n 228
4.12 Feynman-Kac formula 230
4.13 The quantum mechanical oscillator 235
4.14 Feynman-Kac for the oscillator 238
4.15 Neighborhood of Brownian motion 241
4.16 The Ito integral for Brownian motion 245
4.17 Processes of bounded quadratic variation 255
4.18 The Ito integral for martingales 260
4.19 Stochastic differential equations 264
Selected To p i c s 275
5.1 Percolation 275
5.2 Random Jacobi matrices 286
5.3 Estimation theory 292
5.4 Vlasov dynamics 298
5.5 Multidimensional distributions 306
5.6 Poisson processes 3 11
5.7 Random maps 316
5.8 Circular random variables 319
5.9 Lattice points near Brownian paths 327
5.10 Arithmetic random variables 333
5 . 11 S y m m e t r i c D i o p h a n t i n e E q u a t i o n s 343
5.12 Continuity of random variables 349
Preface
These notes grew from an introduction to probability theory taught during

the first and second term of 1994 at Caltech. There was a mixed audience of
undergraduates and graduate students in the first half of the course which
covered Chapters 2 and 3, and mostly graduate students in the second part
which covered Chapter 4 and two sections of Chapter 5.
Having been online for many years on my personal web sites, the text got
reviewed, corrected and indexed in the summer of 2006. It obtained some
enhancements which benefited from some other teaching notes and research,
I wrote while teaching probability theory at the University of Arizona in
Tucson or when incorporating probability in calculus courses at Caltech
and Harvard University.
Most of Chapter 2 is standard material and subject of virtually any course

on probability theory. Also Chapters 3 and 4 is well covered by the litera
ture but not in this combination.
The last chapter "selected topics" got considerably extended in the summer
of 2006. While in the original course, only localization and percolation prob
lems were included, I added other topics like estimation theory, Vlasov dy
namics, multi-dimensional moment problems, random maps, circle-valued
random variables, the geometry of numbers, Diophantine equations and
harmonic analysis. Some of this material is related to research I got inter
ested in over time.
While the text assumes no prerequisites in probability, a basic exposure to

calculus and linear algebra is necessary. Some real analysis as well as some
background in topology and functional analysis can be helpful.
I would like to get feedback from readers. I plan to keep this text alive and
update it in the future. You can email this to [email protected] and
also indicate on the email if you don't want your feedback to be acknowl
edged in an eventual future edition of these notes.
4 Contents
To get a more detailed and analytic exposure to probability, the students
of the original course have consulted the book [105] which contains much
more material than covered in class. Since my course had been taught,
many other books have appeared. Examples are [21, 34].
For a less analytic approach, see [40, 91, 97] or the still excellent classic
[26]. For an introduction to martingales, we recommend [108] and [47] from
both of which these notes have benefited a lot and to which the students
of the original course had access too.
For Brownian motion, we refer to [73, 66], for stochastic processes to [17],
for stochastic differential equation to [2, 55, 76, 66, 46], for random walks
to [100], for Markov chains to [27, 87], for entropy and Markov operators
[61]. For applications in physics and chemistry, see [106].
For the selected topics, we followed [32] in the percolation section. The
books [101, 30] contain introductions to Vlasov dynamics. The book of [1]
gives an introduction for the moment problem, [75, 64] for circle-valued
random variables, for Poisson processes, see [49, 9]. For the geometry of
numbers for Fourier series on fractals [45].
The book [109] contains examples which challenge the theory with counter
examples. [33, 92, 70] are sources for problems with solutions.
Probability theory can be developed using nonstandard analysis on finite

probability spaces [74]. The book [42] breaks some of the material of the
first chapter into attractive stories. Also texts like [89, 78] are not only for
mathematical tourists.
We live in a time, in which more and more content is available online.

Knowledge diffuses from papers and books to online websites and databases
which also ease the digging for knowledge in the fascinating field of proba
bility theory.
Oliver Knill
Chapter 1
Introduction
1.1 What is probability theory?

Probability theory is a fundamental pillar of modern mathematics with
relations to other mathematical areas like algebra, topology, analysis, ge
ometry or dynamical systems. As with any fundamental mathematical con
struction, the theory starts by adding more structure to a set ft. In a similar
way as introducing algebraic operations, a topology, or a time evolution on
a set, probability theory adds a measure theoretical structure to ft which
generalizes "counting" on finite sets: in order to measure the probability
of a subset A C ft, one singles out a class of subsets A, on which one can
hope to do so. This leads to the notion of a cr-algebra A. It is a set of sub
sets of ft in which on can perform finitely or countably many operations
like taking unions, complements or intersections. The elements in A are
called events. If a point u in the "laboratory" ft denotes an "experiment",
an "event" A £ A is a subset of ft, for which one can assign a proba
bility P[A] e [0,1]. For example, if P[A] = 1/3, the event happens with
probability 1/3. If P[A] = 1, the event takes place almost certainly. The
probability measure P has to satisfy obvious properties like that the union
AUB of two disjoint events A, B satisfies P[iUJB]=P[A]+ P[J5] or that
the complement Ac of an event A has the probability P[AC] = 1 - P[A].
With a probability space (ft,.4,P) alone, there is already some interesting
mathematics: one has for example the combinatorial problem to find the
probabilities of events like the event to get a "royal flush" in poker. If ft
is a subset of an Euclidean space like the plane, P[A] = JAf(x,y) dxdy
for a suitable nonnegative function /, we are led to integration problems
in calculus. Actually, in many applications, the probability space is part of
Euclidean space and the cr-algebra is the smallest which contains all open
sets. It is called the Borel cr-algebra. An important example is the Borel
cr-algebra on the real line.
Given a probability space (ft, A, P), one can define random variables X. A
random variable is a function X from ft to the real line R which is mea
surable in the sense that the inverse of a measurable Borel set B in R is
" Chapter 1. Introduction
in A. The interpretation is that if uj is an experiment, then X(u>) mea
sures an observable quantity of the experiment. The technical condition of
measurability resembles the notion of a continuity for a function / from a
topological space (fi, O) to the topological space (R,U). A function is con
tinuous if f~l{U) G O for all open sets U € U. In probability theory, where
functions are often denoted with capital letters, like X, Y,..., a random
variable X is measurable if X~l(B) € A for all Borel sets B £ B. Any
continuous function is measurable for the Borel a-algebra. As in calculus,
where one does not have to worry about continuity most of the time, also in
probability theory, one often does not have to sweat about measurability is
sues. Indeed, one could suspect that notions like a-algebras or measurability
were introduced by mathematicians to scare normal folks away from their
realms. This is not the case. Serious issues are avoided with those construc
tions. Mathematics is eternal: a once established result will be true also in
thousands of years. A theory in which one could prove a theorem as well as
its negation would be worthless: it would formally allow to prove any other
result, whether true or false. So, these notions are not only introduced to
keep the theory "clean", they are essential for the "survival" of the theory.
We give some examples of "paradoxes" to illustrate the need for building
a careful theory. Back to the fundamental notion of random variables: be
cause they are just functions, one can add and multiply them by defining
(X + y)(w) = X(lj) + Y(w) or (XY)(u>) = X(lo)Y(uj). Random variables
form so an algebra C. The expectation of a random variable X is denoted
by E[X] if it exists. It is a real number which indicates the "mean" or "av
erage" of the observation X. It is the value, one would expect to measure in
the experiment. If X = 1B is the random variable which has the value 1 if
u> is in the event B and 0 if lj is not in the event B, then the expectation of
X is just the probability of B. The constant random variable X(w) = a has
the expectation E[X] = a. These two basic examples as well as the linearity
requirement E[aX + bY] = aE[X] +bE[Y] determine the expectation for all
random variables in the algebra C: first one defines expectation for finite
sums YJi=\ adBt called elementary random variables, which approximate
general measurable functions. Extending the expectation to a subset C1 of
the entire algebra is part of integration theory. While in calculus, one can
live with the Riemann integral on the real line, which defines the integral
by Riemann sums f* f(x) dx ~ \ J2i/ne[a,b] /(*/«)> the integral defined in
measure theory is the Lebesgue integral. The later is more fundamental
and probability theory is a major motivator for using it. It allows to make
statements like that the probability of the set of real numbers with periodic
decimal expansion has probability 0. In general, the probability of A is the
expectation of the random variable X(x) = f(x) = lA{x). In calculus, the
integral f0 f(x) dx would not be defined because a Riemann integral can
give 1 or 0 depending on how the Riemann approximation is done. Probabil
ity theory allows to introduce the Lebesgue integral by defining f* f(x) dx
as the limit of £ Yn=i f(xi) for n -> oo, where n are random uniformly
distributed points in the interval [a, b]. This Mcnte Carlo definition of the
Lebesgue integral is based on the law of large numbers and is as intuitive
1.1. What is probability theory? '
to state as the Riemann integral which is the limit of £ Y.Xj=j/ne[a,b\ f(xo)
for n —▶ oo.
With the fundamental notion of expectation one can define the variance,
Var[X] = E[X2] - E[X]2 and the standard deviation a[X] = y/Var[X] of a
random variable X for which X2 e C1. One can also look at the covariance
Cov[XY] = E[XY] - E[X]E[Y] of two random variables X,Y for which
X2,Y2 € C1. The'correlation Corr[X,Y] = Cov[XY]/(a[X]a[Y)) of two
random variables with positive variance is a number which tells how much
the random variable X is related to the random variable Y. If E[XY] is
interpreted as an inner product, then the standard deviation is the length
of X - E[X] and the correlation has the geometric interpretation as cos(a),
where a is the angle between the centered random variables X - E[X] and
Y - E[Y}. For example, if Cov[X, Y] = 1, then Y = XX for some A > 0, if
Cov[X, Y] = -1, they are anti-parallel. If the correlation is zero, the geo
metric interpretation is that the two random variables are perpendicular.
Decorrelated random variables still can have relations to each other but if
for any measurable real functions / and g, the random variables f(X) and
g(X) are uncorrected, then the random variables X,Y are independent.
A random variable X can be described well by its distribution function

Fx> This is a real-valued function defined as Fx(s) = P[X < s] on R,
where {X < s } is the event of all experiments uj satisfying X(u) < s. The
distribution function does not encode the internal structure of the random
variable X; it does not reveal the structure of the probability space for ex
ample. But the function Fx allows the construction of a probability space
with exactly this distribution function. There are two important types of
distributions, continuous distributions with a probability density function
fx = Ffx and discrete distributions for which F is piecewise constant. An
example of a continuous distribution is the standard normal distribution,
where fx(x) = e~*2/2/\/27r. One can characterize it as the distribution
with maximal entropy 1(f) = - Jlog{f{x))f{x) dx among all distributions
which have zero mean and variance 1. An example of a discrete distribu
tion is the Poisson distribution P[X = k] = e_A^ on N = {0,1,2,... }.
One can describe random variables by their moment generating functions
Mx(t) = E[ext] or by their characteristic function </>x(t) = E[eiXt]. The
later is the Fourier transform of the law fix = F'x which is a measure on
the real line R.
The law /j,x of the random variable is a probability measure on the real
line satisfying /ix((a, b\) = Fx(b) - Fx(a). By the Lebesgue decomposition
theorem, one can decompose any measure // into a discrete part /xpp, an
absolutely continuous part jiac and a singular continuous part /jlsc. Random
variables X for which fix is a discrete measure are called discrete random
variables, random variables with a continuous law are called continuous
random variables. Traditionally, these two type of random variables are
the most important ones. But singular continuous random variables appear
too: in spectral theory, dynamical systems or fractal geometry. Of course,
the law of a random variable X does not need to be pure. It can mix the
8 - Chapter 1. Introduction
three types. A random variable can be mixed discrete and continuous for
example^
Inequalities play an important role in probability theory. The Chebychev

inequality P[\X - E[X]\ > c] < ^2£j*l is used very often. It is a spe
cial case of the Chebychev-Markov mejquality h(c) • P[X > c] < E[h(X)]
for monotone nonnegative functions ft. Other inequalities are the Jensen
inequality E[h(X)] > h(E[X]) for convex functions ft, the Minkowski in
equality \\X + Y\\p < \\X\\p + ||F||P or the Holder inequality ||AT||i <
ll*llpll*1lg,l/P + VQ = 1 for random variables, X, Y, for which \\X\\P =
E[|^lp]> ll*1lg = E[|Y|9] are finite. Any inequality which appears in analy
sis can be useful in the toolbox of probability theory.
Independence is an central notion in probability theory. Two events A, B

are called independent, if P[A n B] = P[A] • P[B]. An arbitrary set of
events A{ is called independent, if for any finite subset of them, the prob
ability of their intersection is the product of their probabilities. Two o-
algebras A, B are called independent, if for any pair A e A, B e B, the
events A, B are independent. Two random variables X, Y are independent,
if they generate independent a-algebras. It is enough to check that the
events A = {X e (a, 6)} and B = {Y e (c,d)} are independent for
all intervals (a, b) and (c,d). One should think of independent random
variables as two aspects of the laboratory Q which do not influence each
other. Each event A = {a < X(u) < b } is independent of the event
B = {c< Y(uj) <d}. While the distribution function Fx+Y of the sum of
two independent random variables is a convolution /R Fx(t-s) dFy(s), the
moment generating functions and characteristic functions satisfy the for
mulas Mx+Y{t) = Mx(t)MY{t) and </>x+y(£) = fo(*)0y(*)- These identi
ties make Mx, (j)X valuable tools to compute the distribution of an arbitrary
finite sum of independent random variables.
Independence can also be explained using conditional probability with re

spect to an event B of positive probability: the conditional probability
P[A\B] = P[A ft B]/P[B] of A is the probability that A happens when we
know that B takes place. If B is independent of A, then P[A\B] = P[A] but
in general, the conditional probability is larger. The notion of conditional
probability leads to the important notion of conditional expectation E[X|B]
of a random variable X with respect to some sub-a-algebra B of the a al
gebra A; it is a new random variable which is S-measurable. For B = A, it
is the random variable itself, for the trivial algebra B = {0, ft }, we obtain
the usual expectation E[X] = E[X|{0,fi }]. If i? is generated by a finite
partition Bi,..., Bn of Q of pairwise disjoint sets covering Q, then E[X|S]
is piecewise constant on the sets Bi and the value on Bi is the average
value of X on Bi. If B is the cr-algebra of an independent random variable
Y, then E[X|Y] = E[X\B] = E[X]. In general, the conditional expectation
with respect to B is a new random variable obtained by averaging on the
elements of B. One has E[X|Y] = ft(Y) for some function ft, extreme cases
being E[X|1] = E[X], E[X\X] = X. An illustrative example is the situation
1.1. What is probability theory? 9
where X(x,y) is a continuous function on the unit square with P = dxdy

as a probability measure and where Y(x,y) = x. In that case, E[X|Y] is
a function of x alone, given by E[X|Y](x) = J* f(x,y) dy. This is called a
conditional integral.
A set {Xt}teT of random variables defines a stochastic process. The vari

able t e T is a parameter called "time". Stochastic processes are to prob
ability theory what differential equations are to calculus. An example is a
family Xn of random variables which evolve with discrete time n e N. De
terministic dynamical system theory branches into discrete time systems,
the iteration of maps and continuous time systems, the theory of ordinary
and partial differential equations. Similarly, in probability theory, one dis
tinguishes between discrete time stochastic processes and continuous time
stochastic processes. A discrete time stochastic process is a sequence of ran
dom variables Xn with certain properties. An important example is when
Xn are independent, identically distributed random variables. A continuous
time stochastic process is given by a family of random variables X*, where
t is real time. An example is a solution of a stochastic differential equation.
With more general time like Zd or Rd random variables are called random
fields which play a role in statistical physics. Examples of such processes
are percolation processes.
While one can realize every discrete time stochastic process Xn by a measure-
preserving transformation T : ft —> ft and Xn(u) = X(Tn(uo)), probabil
ity theory often focuses a special subclass of systems called martingales,
where one has a filtration An C An+\ of a-algebras such that Xn is An-
measurable and E[Xn|Ai-i] = ^n-i, where E[Xn|Ai-i] is the conditional
expectation with respect to the sub-algebra An-i- Martingales are a pow
erful generalization of the random walk, the process of summing up IID
random variables with zero mean. Similar as ergodic theory, martingale
theory is a natural extension of probability theory and has many applica
tions.
The language of probability fits well into the classical theory of dynam
ical systems. For example, the ergodic theorem of Birkhoff for measure-
preserving transformations has as a special case the law of large numbers
which describes the average of partial sums of random variables ^ XX=i ^k-
There are different versions of the law of large numbers. "Weak laws"
make statements about convergence in probability, "strong laws" make
statements about almost everywhere convergence. There are versions of
the law of large numbers for which the random variables do not need to
have a common distribution and which go beyond Birkhoff's theorem. An
other important theorem is the central limit theorem which shows that
Sn = Xi 4- X2 + • • • + Xn normalized to have zero mean and variance 1
converges in law to the normal distribution or the law of the iterated loga
rithm which says that for centered independent and identically distributed
Xfc, the scaled sum Sn/An has accumulation points in the interval [—cr, a]
if An = y/2n log log n and a is the standard deviation of X&. While stating
10 Chapter 1. Introduction
the weak and strong law of large numbers and the central limit theorem,
different convergence notions for random variables appear: almost sure con
vergence is the strongest, it implies convergence in probability and the later
implies convergence convergence in law. There is also /^-convergence which
is stronger than convergence in probability.
As in the deterministic case, where the theory of differential equations is

more technical than the theory of maps, building up the formalism for
continuous time stochastic processes Xt is more elaborate. Similarly as
for differential equations, one has first to prove the existence of the ob
jects. The most important continuous time stochastic process definitely is
Brownian motion Bt. Standard Brownian motion is a stochastic process
which satisfies B0 = 0, E[Bt] = 0, Cov[Bs,Bt] = s for s < t and for
any sequence of times, 0 = t0 < tx < • • • where £(t) is called white noise. Because white noise is only
defined as a generalized function and is not a stochastic process by itself,
this stochastic differential equation has to be understood in its integrated
form St = /o dBs = f*((s)ds.
More generally, a solution to a stochastic differential equation j-tXt =
f(Xt)Ct(t) + g(Xt) is defined as the solution to the integral equation Xt =
ô + J0 f(Xs) dBt + /0 g(Xs) ds. Stochastic differential equations can
be defined in different ways. The expression f£ f(X8) dBt can either be
defined as an Ito integral, which leads to martingale solutions, or the
Stratonovich integral, which has similar integration rules than classical
differentiation equations. Examples of stochastic differential equations are
ftXt = Xt£(t) which has the solution Xt = eBt~^2. Or ftXt = B?((t)
which has as the solution the process Xt = B% - 10B? + 15Bt. The key tool
to solve stochastic differential equations is Ito's formula f(Bt) - f{B0) —
Jo f'{Bs)dBs + \ f0 f"(Ba) ds, which is the stochastic analog of the fun
damental theorem of calculus. Solutions to stochastic differential equations
are examples of Markov processes which show diffusion. Especially, the so
lutions can be used to solve classical partial differential equations like the
Dirichlet problem Au = 0 in a bounded domain D with u = f on the
boundary SD. One can get the solution by computing the expectation of
/ at the end points of Brownian motion starting at x and ending at the
boundary u = EX[/(£T)]. On a discrete graph, if Brownian motion is re
placed by random walk, the same formula holds too. Stochastic calculus is
also useful to interpret quantum mechanics as a diffusion processes [73, 71]
or as a tool to compute solutions to quantum mechanical problems using
Feynman-Kac formulas.
Some features of stochastic process can be described using the language of

Markov operators P, which are positive and expectation-preserving trans
formations on C1. Examples of such operators are Perron-Frobenius op
erators X —▶ X(T) for a measure preserving transformation T defining a
1.1. What is probability theory? 11
discrete time evolution or stochastic matrices describing a random walk
on a finite graph. Markov operators can be defined by transition proba
bility functions which are measure-valued random variables. The interpre
tation is that from a given point u, there are different possibilities to go
to. A transition probability measure V(u,-) gives the distribution of the
target. The relation with Markov operators is assured by the Chapman-
Kolmogorov equation Pn+m = pn o Pm. Markov processes can be obtained
from random transformations, random walks or by stochastic differential
equations. In the case of a finite or countable target space 5, one obtains
Markov chains which can be described by probability matrices P, which
are the simplest Markov operators. For Markov operators, there is an ar
row of time: the relative entropy with respect to a background measure
is non-increasing. Markov processes often are attracted by fixed points of
the Markov operator. Such fixed points are called stationary states. They
describe equilibria and often they are measures with maximal entropy. An
example is the Markov operator P, which assigns to a probability density
fy the probability density of /y+x where Y + X is the random variable
Y + X normalized so that it has mean 0 and variance 1. For the initial
function / = 1, the function Pn(fx) is the distribution of 5* the nor
malized sum of n IID random variables X*. This Markov operator has a
unique equilibrium point, the standard normal distribution. It has maxi
mal entropy among all distributions on the real line with variance 1 and
mean 0. The central limit theorem tells that the Markov operator P has
the normal distribution as a unique attracting fixed point if one takes the
weaker topology of convergence in distribution on Cl. This works in other
situations too. For circle-valued random variables for example, the uniform
distribution maximizes entropy. It is not surprising therefore, that there is
a central limit theorem for circle-valued random variables with the uniform
distribution as the limiting distribution.
In the same way as mathematics reaches out into other scientific areas,
probability theory has connections with many other branches of mathe
matics. The last chapter of these notes give some examples. The section
on percolation shows how probability theory can help to understand criti
cal phenomena. In solid state physics, one considers operator-valued ran
dom variables. The spectrum of random operators are random objects too.
One is interested what happens with probability one. Localization is the
phenomenon in solid state physics that sufficiently random operators of
ten have pure point spectrum. The section on estimation theory gives a
glimpse of what mathematical statistics is about. In statistics one often
does not know the probability space itself so that one has to make a statis
tical model and look at a parameterization of probability spaces. The goal
is to give maximum likelihood estimates for the parameters from data and
to understand how small the quadratic estimation error can be made. A
section on Vlasov dynamics shows how probability theory appears in prob
lems of geometric evolution. Vlasov dynamics is a generalization of the
n-body problem to the evolution of of probability measures. One can look
at the evolution of smooth measures or measures located on surfaces. This
deterministic stochastic system produces an evolution of densities which
can form singularities without doing harm to the formalism. It also defines
the evolution of surfaces. The section on moment problems is part of multi
variate statistics. As for random variables, random vectors can be described
by their moments. Since moments define the law of the random variable,
the question arises how one can see from the moments, whether we have a
continuous random variable. The section of random maps is an other part
of dynamical systems theory. Randomized versions of diffeomorphisms can
be considered idealization of their undisturbed versions. They often can
be understood better than their deterministic versions. For example, many
random diffeomorphisms have only finitely many ergodic components. In
the section in circular random variables, we see that the Mises distribu
tion has extremal entropy among all circle-valued random variables with
given circular mean and variance. There is also a central limit theorem
on the circle: the sum of IID circular random variables converges in law
to the uniform distribution. We then look at a problem in the geometry
of numbers: how many lattice points are there in a neighborhood of the
graph of one-dimensional Brownian motion? The analysis of this problem
needs a law of large numbers for independent random variables Xk with
uniform distribution on [0,1]: for 0 < 5 < 1, and An = [0, l/n6] one has
lininôo ^ Ylk=i n* = 1- Probability theory also matters in complex
ity theory as a section on arithmetic random variables shows. It turns out
that random variables like Xn(k) = fc, Yn(k) = k2 + 3 mod n defined on
finite probability spaces become independent in the limit n —▶ oc. Such
considerations matter in complexity theory: arithmetic functions defined
on large but finite sets behave very much like random functions. This is
reflected by the fact that the inverse of arithmetic functions is in general
difficult to compute and belong to the complexity class of NP. Indeed, if
one could invert arithmetic functions easily, one could solve problems like
factoring integers fast. A short section on Diophantine equations indicates
how the distribution of random variables can shed light on the solution
of Diophantine equations. Finally, we look at a topic in harmonic analy
sis which was initiated by Norbert Wiener. It deals with the relation of
the characteristic function cj)X and the continuity properties of the random
variable X.
1.2 Some paradoxes in probability theory

Colloquial language is not always precise enough to tackle problems in
probability theory. Paradoxes appear, when definitions allow different in
terpretations. Ambiguous language can lead to wrong conclusions or con
tradicting solutions. To illustrate this, we mention a few problems. The
following four examples should serve as a motivation to introduce proba
bility theory on a rigorous mathematical footing.
1) Bertrand's paradox (Bertrand 1889)

We throw at random lines onto the unit disc. What is the probability that
1.2. Some paradoxes in probability theory 13
the line intersects the disc with a length > y/3, the length of the inscribed
equilateral triangle?
First answer: take an arbitrary point P on the boundary of the disc. The
set of all lines through that point are parameterized by an angle 0. In order
that the chord is longer than -y/3, the line has to lie within a sector of 60°
within a range of 180°. The probability is 1/3.
Second answer: take all lines perpendicular to a fixed diameter. The chord
is longer than \/3 if the point of intersection lies on the middle half of the
diameter. The probability is 1/2.
Third answer: if the midpoints of the chords lie in a disc of radius 1/2, the
chord is longer than \/3- Because the disc has a radius which is half the
radius of the unit disc, the probability is 1/4.
''V
Figure. Random an- Figure. Random Figure. Random area.

gle- translation.
Like most paradoxes in mathematics, a part of the question in Bertrand's

problem is not well defined. Here it is the term " random line". The solu
tion of the paradox lies in the fact that the three answers depend on the
chosen probability distribution. There are several "natural" distributions.
The actual answer depends on how the experiment is performed.
2) Petersburg paradox (D.Bernoulli, 1738)

In the Petersburg casino, you pay an entrance fee c and you get the prize
2T, where T is the number of times, the casino flips a coin until "head"
appears. For example, if the sequence of coin experiments would give "tail,
tail, tail, head", you would win 23 - c = 8 - c, the win minus the entrance
fee. Fair would be an entrance fee which is equal to the expectation of the
win, which is
£2fcP[T = fc] = ]Tl = oo.

fc=l fc=l
The paradox is that nobody would agree to pay even an entrance fee c = 10.
The problem with this casino is that it is not quite clear, what is "fair".
For example, the situation T = 20 is so improbable that it never occurs
in the life-time of a person. Therefore, for any practical reason, one has
not to worry about large values of T. This, as well as the finiteness of
money resources is the reason, why casinos do not have to worry about the
following bullet proof martingale strategy in roulette: bet c dollars on red.
If you win, stop, if you lose, bet 2c dollars on red. If you win, stop. If you
lose, bet Ac dollars on red. Keep doubling the bet. Eventually after n steps,
red will occur and you will win 2nc - (c + 2c H h 2n_1c) = c dollars.
This example motivates the concept of martingales. Theorem (3.2.7) or
proposition (3.2.9) will shed some light on this. Back to the Petersburg
paradox. How does one resolve it? What would be a reasonable entrance
fee in "real life"? Bernoulli proposed to replace the expectation E[G] of the
profit G = 2T with the expectation (E[\/G])2, where u(x) = y/x is called a
utility function. This would lead to a fair entrance
oo -
(E[VG})2 = £>fc/22-fc)2 = ^ ~ 5.828... .
It is not so clear if that is a way out of the paradox because for any proposed
utility function u(fc), one can modify the casino rule so that the paradox
reappears: pay (2fc)2 if the utility function u(k) = \fk or pay e2 dollars,
if the utility function is u(k) = log(fe). Such reasoning plays a role in
economics and social sciences.
Figure. The picture to the right

shows the average profit devel
opment during a typical tourna
ment of 4000 Petersburg games.
After these 4000 games, the
player would have lost about 10
thousand dollars, when paying a
10 dollar entrance fee each game.
The player would have to play a
very, very long time to catch up.
V
Mathematically, the player will | ^
do so and have a profit in the ? %
lonq run, but it is• unlikely thatI jT1 0 0„^° ^_.__..______, _l.
. .„ , . 1 1 I-/- 200° 30Q0 4000
it will happen in his or her life '
time.
3) The three door problem (1991) Suppose you're on a game show and
you are given a choice of three doors. Behind one door is a car and behind
the others are goats. You pick a door-say No. 1 - and the host, who knows
what's behind the doors, opens another door-say, No. 3-which has a goat.
(In all games, he opens a door to reveal a goat). He then says to you, "Do
1.2. Some paradoxes in probability theory 15
you want to pick door No. 2?" (In all games he always offers an option to
switch). Is it to your advantage to switch your choice?
The problem is also called "Monty Hall problem" and was discussed by
Marilyn vos Savant in a "Parade" column in 1991 and provoked a big
controversy. (See [98] for pointers and similar examples.) The problem is
that intuitive argumentation can easily lead to the conclusion that it does
not matter whether to change the door or not. Switching the door doubles
the chances to win:
No switching: you choose a door and win with probability 1/3. The opening
of the host does not affect any more your choice.
Switching: when choosing the door with the car, you loose since you switch.
If you choose a door with a goat. The host opens the other door with the
goat and you win. There are two such cases, where you win. The probability
to win is 2/3.
4) The Banach-Tarski paradox (1924)

It is possible to cut the standard unit ball ft = {:rER3||:r|<l} into 5
disjoint pieces ft = Yx U Y2 U Y3 U Y4 U Y5 and rotate and translate the pieces
with transformations T{ so that T1(Y1)UT2{Y2) = ft and T3(Y3)UT4(Y4) U
T*>{Y5) = ft' is a second unit ball ft' = {x e M3 | \x - (3,0,0)| < 1} and all
the transformed sets again don't intersect.
While this example of Banach-Tarski is spectacular, the existence of bounded
subsets A of the circle for which one can not assign a translational invari
ant probability P[A] can already be achieved in one dimension. The Italian
mathematician Giuseppe Vitali gave in 1905 the following example: define
an equivalence relation on the circle T = [0, 2tt) by saying that two angles
are equivalent x ~ y if {x-y)/n is a rational angle. Let A be a subset in the
circle which contains exactly one number from each equivalence class. The
axiom of choice assures the existence of A If xi, x2, • • • is a enumeration
of the set of rational angles in the circle, then the sets A = A + xt are
pairwise disjoint and satisfy |J~i Ai = T. If we could assign a translational
invariant probability P[A] to A, then the basic rules of probability would
give
oo oo oo
l=P[T]=P[(J^] = X>[^] = £p.

i=l
But there is no real number p = P[A] = P[A] which makes this possible.
Both the Banach-Tarski as well as Vitalis result shows that one can not
hope to define a probability space on the algebra A of all subsets of the unit
ball or the unit circle such that the probability measure is translational
and rotational invariant. The natural concepts of "length" or "volume",
which are rotational and translational invariant only makes sense for a
smaller algebra. This will lead to the notion of a-algebra. In the context
of topological spaces like Euclidean spaces, it leads to Borel cr-algebras,
algebras of sets generated by the compact sets of the topological space.
This language will be developed in the next chapter.
1.3 Some applications of probability theory
Probability theory is a central topic in mathematics. There are close re
lations and intersections with other fields like computer science, ergodic
theory and dynamical systems, cryptology, game theory, analysis, partial
differential equation, mathematical physics, economical sciences, statistical
mechanics and even number theory. As a motivation, we give some prob
lems and topics which can be treated with probabilistic methods.
1) Random walks: (statistical mechanics, gambling, stock markets, quan

tum field theory).
Assume you walk through a lattice. At each vertex, you choose a direction
at random. What is the probability that you return back to your start
ing point? Polya's theorem (3.8.1) says that in two dimensions, a random
walker almost certainly returns to the origin arbitrarily often, while in three
dimensions, the walker with probability 1 only returns a finite number of
times and then escapes for ever.
iI
^
\
/iV
wy
Figure. A random Figure. A piece of a Figure. A piece of a
walk in one dimen random walk in two random walk in three
sions displayed as a dimensions. dimensions.
graph (t,Bt).
2) Percolation problems (model of a porous medium, statistical mechanics,

critical phenomena).
Each bond of a rectangular lattice in the plane is connected with probability
p and disconnected with probability 1 - p. Two lattice points x, y in the
lattice are in the same cluster, if there is a path from x to y. One says that
"percolation occurs" if there is a positive probability that an infinite cluster
appears. One problem is to find the critical probability pc, the infimum of all
p, for which percolation occurs. The problem can be extended to situations,
where the switch probabilities are not independent to each other. Some
random variables like the size of the largest cluster are of interest near the
critical probability pc.
1.3. Some applications of probability theory 17
..'■.-.::
St
_u -iu
' ' ^
j hJ, i J_
r-i m&m
Z " J ' j" 'i i— : .i — . _„ iJ. . .! _r
:.:—
i H jj j z •'_- -c :U 3 n z i_
Figure. Sond percola Figure. Bond percola Figure. Bond percola

tion with p=0.2. tion with p=0.4- tion with p=0.6.
A variant of bond percolation is site percolation where the nodes of the

lattice are switched on with probability p.
V ■■ J S ' ?
\ ■
Figure. Site percola Figure. Site percola Figure. Site percola

tion with p=0.2. tion with p=0.4- tion with p=0.6.
Generalized percolation problems are obtained, when the independence

of the individual nodes is relaxed. A class of such dependent percola
tion problems can be obtained by choosing two irrational numbers a,/?
like a = y/2 — 1 and (3 = y/3 — 1 and switching the node (n, m) on if
(na + ra/3) mod 1 G [0,p). The probability of switching a node on is again
p, but the random variables
Xn m — l(na+m/3) mod l€[0,p)
are no more independent.

^ ■■ ' . ■ ! ■■ V
■ ■■ ^ ■■■'■ v ■
■^ ■ ■ ■ _■ ■ ■
■ ■ r_■ ■ ■
.v. %:■-/'- ■■*■
Figure. Dependent Figure. Dependent Figure. Dependent

site percolation with site percolation with site percolation with
p=0.2. p=0.4. p=0.6.
Even more general percolation problems are obtained, if also the distribu
tion of the random variables Xn,m can depend on the position (n, m).
3) Random Schrodinger operators, (quantum mechanics, functional analy

sis, disordered systems, solid state physics)
Consider the linear map Lu(n) = ]C|m-n|=i u(n) + V(n)u(n) on the space
of sequences u = (..., u-2, "U-i, uo, ui, u2,...). We assume that V(n) takes
random values in {0,1}. The function V is called the potential. The problem
is to determine the spectrum or spectral type of the infinite matrix L on
the Hilbert space I2 of all sequences u with finite ||w||2 = S^L-oo^n-
The operator L is the Hamiltonian of an electron in a one-dimensional
disordered crystal. The spectral properties of L have a relation with the
conductivity properties of the crystal. Of special interest is the situation,
where the values V(n) are all independent random variables. It turns out
that if V(n) are IID random variables with a continuous distribution, there
are many eigenvalues for the infinite dimensional matrix L - at least with
probability 1. This phenomenon is called localization.
1.3. Some applications of probability theory l'J
ll ! I ft I
if
...,l, ll, iii.n...ii.tlill lillt.ttld.ri.il Hi.ti.,.iltil iiiiiitiii.i..i
Figure. A wave Figure. wave

Figure. A wave
$ { t ) = e i LV ( 0 ) iP(t) = eiLtip(0) m = ALt
V(o)
evolving in a random evolving in a random evolving in a random
potential at t — 0. potential at t = 1. potential at t = 2.
Shown are both the Shown are both the Shown are both the
potential Vn and the potential Vn and the potential Vn and the
wave i/>(0). wave ?/>(l). wave i/j(2).
More general operators are obtained by allowing V(n) to be random vari

ables with the same distribution but where one does not persist on indepen
dence any more. A well studied example is the almost Mathieu operator,
where V(n) = Acos(0 + net) and for which a/(2ir) is irrational.
4) Classical dynamical systems (celestial mechanics, fluid dynamics, me

chanics, population models)
The study of deterministic dynamical systems like the logistic map x h^

4x(l - x) on the interval [0,1] or the three body problem in celestial me
chanics has shown that such systems or subsets of it can behave like random
systems. Many effects can be described by ergodic theory, which can be
seen as a brother of probability theory. Many results in probability the
ory generalize to the more general setup of ergodic theory. An example is
Birkhoff's ergodic theorem which generalizes the law of large numbers.
Figure. Iterating the Figure. The simple Figure. A short time

logistic map evolution of the New
mechanical system of
tonian three body
a double pendulum
T{x) = 4x(l - x) exhibits complicated problem. There are
dynamics. The dif energies and subsets
on [0,1] produces
ferential equation of the energy surface
independent random which are invari
variables. The in defines a measure ant and on which
variant measure P is preserving flow Tt on there is an invariant
continuous. a probability space.
probability measure.
Given a dynamical system given by a map T or a flow Tt on a subset Q of

some Euclidean space, one obtains for every invariant probability measure
P a probability space (ft?tA,P). An observed quantity like a coordinate of
an individual particle is a random variable X and defines a stochastic pro
cess Xn(u) = X(Tnuj). For many dynamical systems including also some 3
body problems, there are invariant measures and observables X for which
Xn are IID random variables. Probability theory is therefore intrinsically
relevant also in classical dynamical systems.
5) Cryptology. (computer science, coding theory, data encryption)

Coding theory deals with the mathematics of encrypting codes or deals
with the design of error correcting codes. Both aspects of coding theory
have important applications. A good code can repair loss of information
due to bad channels and hide the information in an encrypted way. While
many aspects of coding theory are based in discrete mathematics, number
theory, algebra and algebraic geometry, there are probabilistic and combi
natorial aspects to the problem. We illustrate this with the example of a
public key encryption algorithm whose security is based on the fact that
it is hard to factor a large integer N = pq into its prime factors p, q but
easy to verify that p, q are factors, if one knows them. The number N can
be public but only the person, who knows the factors p, q can read the
message. Assume, we want to crack the code and find the factors p and q.
The simplest method is to try to find the factors by trial and error but this is
impractical already if N has 50 digits. We would have to search through 1025
numbers to find the factor p. This corresponds to probe 100 million times
1.3. Some applications of probability theory 21
every second over a time span of 15 billion years. There are better methods
known and we want to illustrate one of them now: assume we want to find
the factors of N = 11111111111111111111111111111111111111111111111.
The method goes as follows: start with an integer a and iterate the quadratic
map T(x) =x2 + c mod N on {0,1.,,, .N - 1 }. If we assume the numbers
x0 = a, xi = T(a),x2 = T(T(a))... to be random, how many such numbers
do we have to generate, until two of them are the same modulo one of the
prime factors p? The answer is surprisingly small and based on the birthday
paradox: the probability that in a group of 23 students, two of them have the
same birthday is larger than 1/2: the probability of the event that we have
no birthday match is 1(364/365) (363/365) • • • (343/365) = 0.492703..., so
that the probability of a birthday match is 1 - 0.492703 = 0.507292. This
is larger than 1/2. If we apply this thinking to the sequence of numbers
Xi generated by the pseudo random number generator T, then we expect
to have a chance of 1/2 for finding a match modulo p in yfp iterations.
Because p < y/n, we have to try iV1/4 numbers, to get a factor: if xn and
xm are the same modulo p, then gcd(xn -xm,N) produces the factor p of
N. In the above example of the 46 digit number AT, there is a prime factor
p = 35121409. The Pollard algorithm finds this factor with probability 1/2
in y/p == 5926 steps. This is an estimate only which gives the order of mag
nitude. With the above N, if we start with a = 11 and take a = 3, then we
have a match #27720 = xi3860- It can be found very fast.
This probabilistic argument would give a rigorous probabilistic estimate

if we would pick truly random numbers. The algorithm of course gener
ates such numbers in a deterministic way and they are not truly random.
The generator is called a pseudo random number generator. It produces
numbers which are random in the sense that many statistical tests can
not distinguish them from true random numbers. Actually, many random
number generators built into computer operating systems and program
ming languages are pseudo random number generators.
Probabilistic thinking is often involved in designing, investigating and at

tacking data encryption codes or random number generators.
6) Numerical methods, (integration, Monte Carlo experiments, algorithms)

In applied situations, it is often very difficult to find integrals directly. This
happens for example in statistical mechanics or quantum electrodynamics,
where one wants to find integrals in spaces with a large number of dimen
sions. One can nevertheless compute numerical values using Monte Carlo
Methods with a manageable amount of effort. Limit theorems assure that
these numerical values are reasonable. Let us illustrate this with a very
simple but famous example, the Buffon needle problem.
A stick of length 2 is thrown onto the plane filled with parallel lines, all
of which are distance d = 2 apart. If the center of the stick falls within
distance y of a line, then the interval of angles leading to an intersection
with a grid line has length 2 arccos(y) among a possible range of angles
[0,7r]. The probability of hitting a line is therefore J* 2 arccos(2/)/7r = 2/ir.

This leads to a Monte Carlo method to compute n. Just throw randomly
n sticks onto the plane and count the number k of times, it hits a line. The
number 2n/k is an approximation of n. This is of course not an effective
way to compute tt but it illustrates the principle.
Figure. The Buff on needle prob

lem is a Monte Carlo method
to compute n. By counting the
number of hits in a sequence of
experiments, one can get ran
dom approximations of n. The
law of large numbers assures that
the approximations will converge
to the expected limit. All Monte
Carlo computations are theoreti
cally based on limit theorems.
Chapter 2
Limit theorems
2.1 Probability spaces, random variables, indepen

dence
In this section we define the basic notions of a "probability space" and
"random variables" on an arbitrary set ft.
Definition. A set A of subsets of ft is called a cr-algebra if the following

three properties are satisfied:
(i) ft € A,
(ii) A £ A =▶ Ac =■-Q\A €.4,
(iii) An G A => Ur eA
A pair (ft, A) for which A is a cr-algebra in ft is called a measurable space.
Properties. If A is a cr-algebra, and An is a sequence in A, then the fol

lowing properties follow immediately by checking the axioms:
2) limsupn An := fl~i Um=„ A, € A

3) liminfn An := U~ i fC=„ ^ e A
4) .4, i? are algebras, then A fl 6 is an algebra.
5) If {A\}ie/ is a family of a- sub-algebras of A then f]ieI A% is a cr-algebra.
Example. For an arbitrary set fi, >t = {0,0}) is a cr-algebra. It is called

the trivial cr-algebra.
Example. If fi is an arbitrary set, then A = {A C Cl}) is a cr-algebra. The

set of all subsets of Q is the largest cr-algebra one can define on a set.
23
24 Chapter 2. Limit theorems
Example. A finite set of subsets Au A2,..., An of ft which are pairwise

disjoint and whose union is ft, it is called a partition of ft. It generates the
cr-algebra: A = {A = \JjeJ Aj } where J runs over all subsets of {1,.., n}.
This a-algebra has 2n elements. Every finite a-algebra is of this form. The
smallest nonempty elements {Au ..., An} of this algebra are called atoms.
Definition. For any set C of subsets of ft, we can define <r(C), the smallest
a-algebra A which contains C. The cr-algebra A is the intersection of all
cr-algebras which contain C. It is again a cr-algebra.
Example. For ft = {1,2,3}, the set C = {{1,2}, {2,3 }} generates the

cr-algebra A which consists of all 8 subsets of ft.
Definition. If (E, O) is a topological space, where O is the set of open sets

in E. then a(0) is called the Borel cr-algebra of the topological space. If
A C B, then A is called a subalgebra of B. A set B in B is also called a
Borel set.
Remark. One sometimes defines the Borel cr-algebra as the cr-algebra gen
erated by the set of compact sets C of a topological space. Compact sets
in a topological space are sets for which every open cover has a finite sub-
cover. In Euclidean spaces Rn, where compact sets coincide with the sets
which are both bounded and closed, the Borel cr-algebra generated by the
compact sets is the same as the one generated by open sets. The two def
initions agree for a large class of topological spaces like "locally compact
separable metric spaces".
Remark. Often, the Borel a-algebra is enlarged to the a-algebra of all

Lebesgue measurable sets, which includes all sets B which are a subset
of a Borel set A of measure 0. The smallest a-algebra B which contains
all these sets is called the completion of B. The completion of the Borel
a-algebra is the a-algebra of all Lebesgue measurable sets. It is in general
strictly larger than the Borel a-algebra. But it can also have pathological
features like that the composition of a Lebesgue measurable function with
a continuous functions does not need to be Lebesgue measurable any more.
(See [109], Example 2.4).
Example. The a-algebra generated by the open balls C = {A = Br(x) } of

a metric space (X, d) need not to agree with the family of Borel subsets,
which are generated by 0, the set of open sets in (X, d).
Proof. Take the metric space (R,d) where d(x,y) = l{x=y} is the discrete
metric. Because any subset of R is open, the Borel a-algebra is the set of
all subsets of R. The open balls in R are either single points or the whole
space. The a-algebra generated by the open balls is the set of countable
subset of R together with their complements.
2.1. Probability spaces, random variables, independence 25
Example. If ft = [0,1] x [0,1] is the unit square and C is the set of all sets
of the form [0,1] x [a, b] with 0 < a < b < 1, then a(C) is the a-algebra of
all sets of the form [0,1] x A, where A is in the Borel a-algebra of [0,1].
Definition. Given a measurable space (ft, .4). A function P : A -» R is

called a probability measure and (ft, A, P) is called a probability space if
the following three properties called Kolmogorov axioms are satisfied:
(i) P[A] >0 for all Ae A,

(h) P[ft] = 1,
(hi) An G A disjoint =* P[Un An] = En P[^n]
The last property is called a-additivity.
Properties. Here are some basic properties of the probability measure

which immediately follow from the definition:
1)P[0]=O.
2)AcB=>P[A] <P[B).
3)P[Un^n]<E„P[AJ-
A)P[AC] = 1-P[A}.
5) 0 < P[A] < 1.
6) Ai C A2, C • • • with An G A then P[Un°=i An] = limn->oo P[An}-
Remark. There are different ways to build the axioms for a probability
space. One could for example replace (i) and (ii) with properties 4),5) in
the above list. Statement 6) is equivalent to a-additivity if P is only assumed
to be additive.
Remark. The name "Kolmogorov axioms" honors a monograph of Kol

mogorov from 1933 [53] in which an axiomatization appeared. Other math
ematicians have formulated similar axiomatizations at the same time, like
Hans Reichenbach in 1932. According to Doob, axioms (i)-(iii) were first
proposed by G. Bohlmann in 1908 [22].
Definition. A map X from a measure space (ft, A) to an other measure
space (A,B) is called measurable, if X~\B) G A for all B e B. The set
X~1(B) consists of all points x G ft for which X(x) G B. This pull back set
X~1(B) is defined even if X is non-invertible. For example, for X(x) = x2
on (R,B) one has X^flM]) - [1,2] U [-2,-1].
Definition. A function X : ft —> R is called a random variable, if it is a

measurable map from (ft, .4) to (R, #), where B is the Borel a-algebra of
R. Denote by C the set of all real random variables. The set C is an alge
bra under addition and multiplication: one can add and multiply random
variables and gets new random variables. More generally, one can consider
random variables taking values in a second measurable space (E,B). If
E = Rd, then the random variable X is called a random vector. For a ran
dom vector X = (Xi,..., Xd), each component Xi is a random variable.
Example. Let ft = R2 with Borel a-algebra A and let
P[A] = ^J J e-^-y^dxdy.
Any continuous function X of two variables is a random variable on ft. For
example, X(x,y) = xy(x + y) is a random variable. But also X(x,y) =
l/(x + y) is a random variable, even so it is not continuous. The vector-
valued function X(x, y) = (x, y, x3) is an example of a random vector.
Definition. Every random variable X defines a a-algebra
X-\B) = {X-\B)\BeB}.
We denote this algebra by cr(X) and call it the a-algebra generated by X.
Example. A constant map X(x) = c defines the trivial algebra A = {0, ft }.
Example. The map X(x,y) = x from the square ft = [0,1] x [0,1] to the
real line R defines the algebra B={4x[0,l]}, where A is in the Borel
a-algebra of the interval [0,1].
Example. The map X from Z6 = {0,1,2,3,4,5} to {0,1} c R defined by

X(x) =x mod 2 has the value X(x) = 0 if x is even and X(x) = 1 if x is
odd. The a-algebra generated by X is A = {0, {1,3,5}, {0,2,4}, ft }.
Definition. Given a set B G A with P[B] > 0, we define
F[AlB] ~ ~P[BT '

the conditional probability of A with respect to B. It is the probability of
the event A, under the condition that the event B happens.
Example. We throw two fair dice. Let A be the event that the first dice is
6 and let B be the event that the sum of two dices is 11. Because P[B] =
2/36 = 1/18 and P[A n B] = 1/36 (we need to throw a 6 and then a 5),
we haveP[A\B] = (1/16)/(1/18) = 1/2. The interpretation is that since
we know that the event B happens, we have only two possibilities: (5,6)
or (6,5). On this space of possibilities, only the second is compatible with
the event B.
Exercice. a) Verify that the Sicherman dices with faces (1,3,4,5,6,8) and
(1,2,2,3,3,4) have the property that the probability of getting the value
k is the same as with a pair of standard dice. For example, the proba
bility to get 5 with the Sicherman dices is 3/36 because the three cases
(1,4), (3,2), (3,2) lead to a sum 5. Also for the standard dice, we have
three cases (1,4), (2,3), (3,2).
b) Three dices A, B, C are called non-transitive, if the probability that A >
B is larger than 1/2, the probability that B > C is larger than 1/2 and the
probability that C > A is larger than 1/2. Verify the nontransitivity prop
erty for A = (1,4,4,4,4,4), B = (3,3,3,3,3,6) and C = (2,2,2,5,5,5).
Properties. The following properties of conditional probability are called

Keynes postulates. While they follow immediately from the definition
of conditional probability, they are historically interesting because they
appeared already in 1921 as part of an axiomatization of probability theory:
1) P[A\B] > 0.
2) P[A\A] = 1.
3) P[A\B] + P[AC\B] = 1
4) P[AnB\C] =- P[A\C] ■ p[B\An C}.
Definition. A finite set {Ai,..., An } c A is called a finite partition of fi if

IJn=i Aj = Q and AjClAi =0 for i / j. A finite partition covers the entire
space with finitely many, pairwise disjoint sets.
If all possible experiments are partitioned into different events Aj and the
probabilities that B occurs under the condition Aj, then one can compute
the probability that Ai occurs knowing that B happens:
Theorem 2.1.1 (Bayes rule). Given a finite partition {î,.., An} in A and
B e A with P[B] > 0, one has
P[B\Ai]P[Ai]
P[Ai\B]
Proof. Because the denominator is P[B] = E"=1 P[B|Aj]P[t4j], the Bayes

rule just says P[Ai|S]P[B] = P[B|Ai]P[A,]. But these are by definition
bothP^nB]. □
Example. A fair dice is rolled first. It gives a random number k from

{1,2,3,4,5,6}. Next, a fair coin is tossed k times. Assume, we know that
all coins show heads, what is the probability that the score of the dice was
equal to 5?
Solution. Let B be the event that all coins are heads and let Aj be the
event that the dice showed the number j. The problem is to find P[i45|B].
We know P[B\Aj] = 2~K Because the events AjJ = l,...,6 form a par
tition of ft, we have P[B] = EUFlB n aj] = E-=i P[^|^]P[^] =
£5=i 2"V6 = (1 + 1/2 + 1/3 + 1/4 + 1/5 + l/6)/6 = 49/120. By Bayes
rule,
P[A5\B] = nB\A5]P[A5] = (l/32)(l/6) = J_

(E5=ip[BlAi]p[^']) 49/120 392'
which is about 1 percent.
Example. The Girl-Boy problem: "Dave has two child. One child is a boy.
What is the probability that the other child is a girl"?
Most people would intuitively say 1/2 because the second event looks inde
pendent of the first. However, it is not and the initial intuition is mislead
ing. Here is the solution: first introduce the probability space of all possible
events ft = {BG,GB,BB,GG} with P[{BG}] = P[{GB}] = P[{BB}] =
P[{GG}] = 1/4. Let B = {BG, GB, BB} be the event that there is at least
one boy and A = {GB, BG, GG} be the event that there is at least one
girl. We have
•p[A\B]
L ' J = PÂn^
[ B ] (=
3 /m
4 ) =3 * *
Definition. Two events A, B in s probability space (ft, A, P) are called in

dependent, if
P[AnB] = P[A].P[B].
Example. The probability space ft = {1,2,3,4,5,6 } and Pi = P[{i}} = 1/6

describes a fair dice which is thrown once. The set A = {1,3,5 } is the
event that "the dice produces an odd number". It has the probability 1/2.
The event B = {1,2 } is the event that the dice shows a number smaller
than 3. It has probability 1/3. The two events are independent because
P[A HB}= P[{1}] = 1/6 = P[A] . P[B}.
Definition. Write J C/ I if J is a finite subset of J. A family {Ai}ieI of cr-
sub-algebras of A is called independent, if for every J C/ I and every choice
Aj e Aj P[f]jeJ Ao\ = Ylpej p[^]- A family {Xj}jeJ of random variables
is called independent, if {a(Xj)}jej are independent a-algebras. A family
of sets {Aj}jeI is called independent, if the a-algebras Aj = {0, Aj, ACj, ft }
are independent.
Example. On ft = {1,2,3,4 } the two cr-algebras A = {0, {1,3 }, {2,4 }, ft }

and B = {0, {1,2 }, {3,4 }, ft } are independent.
Properties. (1) If a cr-algebra T C A is independent to itself, then P[A fl

A] = P[A] = P[A]2 so that for every A e T, P[A] e {0,1}. Such a a-algebra
is called P-trivial.
(2) Two sets A,B eA are independent if and only if P[AnB] = P[A] P[B].
(3) If A, B are independent, then A, Bc are independent too.
(4) If P[B] > 0, and A, B are independent, then P[j4|B] = P[A] because
P[A\B]-=(P[A].P[B})/P[B} = P[A}.
(5) For independent sets A, B, the cr-algebras A = {0, A, Ac, ft} and B =
{0, B, Bc, ft} are independent.
Definition. A family J of subsets of ft is called a 7r-system, if T is closed

under intersections: if A, B are in J, then A fl B is in T. A cr-additive map
from a 7r-system I to [0, oo) is called a measure.
Example. 1) The family X = {0, {1}, {2}, {3}, {1,2}, {2,3}, ft} is a 7r-system
on ft = {1,2,3}.
2) The set J = {{a, b) |0 < a < b < 1} U {0} of all half closed intervals is a
7r-system on ft = [0,1] because the intersection of two such intervals [a, b)
and [c, d) is either empty or again such an interval [c, b).
Definition. We use the notation An /* A if An c ^4n+i and |Jn An = A.

Let ft be a set. (ft, V) is called a Dynkin system if V is a set of subsets of
ft satisfying
(i) tie A,
(ii) A, B 6 V, A C B = > B \ A e V.
(hi) An € V, An / A = > i e P
Lemma 2.1.2. (ft, A) is a cr-algebra if and only if it is a 7r-system and a

Dynkin system.
Proof If A is a cr-algebra, then it certainly is both a 7r-system and a Dynkin

system. Assume now, A is both a 7r-system and a Dynkin system. Given
A, B e A. The Dynkin property implies that Ac = ft \ A, Bc = ft \ B are
in A and by the 7r-system property also A U B = ft \ (Ac C\BC) E A. Given
a sequence An G A. Define Bn = U/c=i ^ ^ ^ an<^ ^ = Un^- Then
An / A and by the Dynkin property A e A. Also f|n An = ft \ Un An G A

so that A is a a-algebra. □
Definition. If X is any set of subsets of ft, we denote by d(T) the smallest

Dynkin system, which contains X and call it the Dynkin system generated
by J.
Lemma 2.1.3. If I is a n- system, then d(l) = a(X).
Proof By the previous lemma, we need only to show that d(l) is a 7r—
system.
(i) Define Vx = {B e d(l) | B n C G d(I),VC El}. Because J is a
7r-system, we have X C V\.
Claim. V\ is a Dynkin system.
Proof. Clearly ft G Vx. Given A,B G V - 1 with A C B. For C G J
we compute (B \A) C\C = (B nC)\(Af]C) which is in d(I). Therefore
A\B G Vi. Given An / A with An G Pi and C G I. Then AnnC / AnC
so that A D C G d(X) and A G Pi.
(ii) Define D2 = {Ae d(X) | 5 n A G d(X), VB G d(I) }. From (i) we know
that 1 C T>2. Like in (i), we show that V2 is a Dynkin-system. Therefore
V2 = d(l), which means that d(J) is a 7r-system. D
Lemma 2.1.4. (Extension lemma) Given a 7r-system T. If two measures //, v

on a(J) satisfy /i(ft), i/(ft) < oo and //(A) = i/(A) for A G T, then H = v.
Proof. Proof of lemma (2.1.5). The set V = {A G cr(J) | /u(;4) = i/(A) }

is Dynkin system: first of all ft G V. Given A,B e V,A C B. Then
A*(B\A) = /i(5)-/i(A) = v(B)-v(A) = v(B\A) so that B\AeV. Given
An ET> with An /* A, then the a additivity gives fi(A) = limsupn ii{An) =
limsupnz/(An) = v(A), so that A G V. Since V is a Dynkin system con
taining the 7r-system X, we know that a (J) = d(2") C £> which means that
u. = z/ on cr(J). D
Definition. Given a probability space (ft,^4,P). Two 7r-systems X,JcA

are called P-independent, if for all A G X and 5eJ, P[AnJ3] = P[A] >P[B].
Lemma 2.1.5. Given a probability space (ft,*4, P). Let Q,H be two o-
subalgebras of A and X and J be two 7r-systems satisfying a (I) = Q,
a(J) = H. Then Q and H are independent if and only if X and J are
independent.
Proof, (i) Fix J G X and define on (ft,W) the measures //(if) = P[J H
H],u{H) = P[I]P[H] of total probability P[J]. By the independence of X
and J", they coincide on J and by the extension lemma (2.1.4), they agree
on H and we have P[I n H] = P[I]P[H] for all I G X and if G H.
(ii) Define for fixed H G W the measures /j,(G) = P[G D H] and i/(G) =
P[G]P[fl] of total probability P[H] on (ft, <?). They agree on X and so on Q.
We have shown that P[Gf)H} =P[G]P[H] for all G G 0 and all H eH. D
Definition. *4 is an algebra if ^4 is a set of subsets of ft satisfying
(i) ft G A,
( i i ) i G ^ ^ 4 c G .4,
(hi) A,B eA=> A u £ g * 4
A a-additive map from A to [0, oo) is called a measure.
Theorem 2.1.6 (Caratheodory continuation theorem). Any measure on an

algebra 1Z has a unique continuation to a measure on a(1Z).
Before we launch into the proof of this theorem, we need two lemmas:
Definition. Let A be an algebra and A : A •-▶ [9, oo] with A(0) = 0. A set
A G A is called a A-set, if X(A n G) + X(AC n G) = A(G) for all G G A.
Lemma 2.1.7. The set A\ of A-sets of an algebra A is again an algebra and

satisfies Y2=\ X(Ak C\G) = A((|JLi Ak) nG) for all finite disjoint families
{Ak}l=i and al1 G eA.
Proof From the definition is clear that ft G A\ and that \f B e A\, then
Bc e A\. Given B, C G A\. Then A = 5nCGiA. Proof. Since C G *4A,
we get
A(Cn AcnG) + A(CCnicnG) = A(ACnG).
This can be rewritten with C fl Ac = C n Bc and Gc n Ac = Cc as
\(AC n G) = \{C n £c n G) + x(Cc n G). (2.1)

Because B is a A-set, we get using B DC = A.
\{A n G) + A(BC n G n G) = x(c n G). (2.2)

Since C is a A-set, we have
A(G n G) + A(GC n G) = A(G) . (2.3)
Adding up these three equations shows that B D C is a A-set. We have so

verified that ,4a is an algebra. If B and G are disjoint in ,4a we deduce
from the fact that B is a A-set
\(B n (B u C) n G) + A(BC n (B u c) n G) = A((B u G) n G).

This can be rewritten as A(BflG) + A(GnG) = A((BuG)nG). The analog
statement for finitely many sets is obtained by induction. □
Definition. Let A be a cr-algebra. A map A : A —» [0, oo] is called an outer

measure, if
A(0) = O,
A, B e A with AcB=> X(A) < X(B).
An e A => X(\JnAn) < EnPW (<t subadditivity)
Lemma 2.1.8. (Caratheodory's lemma) If A is an outer measure on a mea

surable space (ft, A), then ,4a C A defines a a-algebra on which A is count-
ably additive.
Proof. Given a disjoint sequence An G A\. We have to show that A =

\JnAn e A\ and X(A) = 52n\(An). By the above lemma (2.1.7), Bn =
Ufc=i Ak is in *4a- By the monotonicity, additivity and the a -subadditivity,
we have
A(G) = A(Bn n G) + X{Bn n G) > X{Bn n G) + X{AC n G)

n
= y, x(Ak n G)+A(^c n G) ^ X(An G)+A(^c n G) •
fc=l
Subadditivity for A gives A(G) < X(A C\G) + X(AC n G). All the inequalities
in this proof are therefore equalities. We conclude that A G C and that A
is a additive on A\. □
We now prove the Caratheodory's continuation theorem (2.1.6):

Proof. Given an algebra 71 with a measure fi. Define A = o~(1Z) and the
a-algebra V consisting of all subsets of ft. Define on V the function
X(A) = inf{yZ ^(An) I {Ai}n€N sequence in 1Z satisfying A C {jAn } •

neN n
(i) A is an outer measure on V.

A(0) = 0 and X(A) D X(B) forB>A are obvious. To see the a subad
ditivity, take a sequence An e V with X(An) < oo and fix e > 0. For all
neN, one can (by the definition of A) find a sequence {Bn,fc}fcGN in TZ
such that An C \JkeN Bn,k and YlkeN M^n,*) < X(An) + c2~n. Define A =
UneNAn C Un,fc€N5n,fe, SO that X{A) < £n?fc/i(Bn,fc) < £n X(An) + 6.
Since e was arbitrary, the a-subadditivity is proven.
(ii) A = /i on TZ.
Given A ell. Clearly X(A) < u.(A). Suppose that Ac\JnAn, with An G
TZ. Define a sequence {Bn}neN of disjoint sets in TZ inductively Bi = A\,
Bn = Ann({Jk<n Ak)c such that Bn c An and |Jn Bn = \Jn An D A. From
the a-additivity of /x on TZ follows
/x(A)<|J/i(Bn)<|J/x(^n),
so that fi(A) > X(A).
(hi) Let V\ be the set of A-sets in V. Then TZcV\.

Given A G TZ and G eV. There exists a sequence {Bn}nGN in 1Z such that
G C Un ^n and En /*(Bn) < A(G) + 6. By the definition of A
Y KBn) = YÂn Bn) + E ^ n B-) ^ A(^ ° G) +• X(A° n G)

because AflG C \JnAn Bn and Ac fl G C \JnAC n Bn- Since e is ar
bitrary, we get X(A) > X(A fl G) + X(AC n G). On the other hand, since
A is sub-additive, we have also X(A) < X(AnG)-T-X(Acf)G) and A is a A-set.
(iv) By (i) A is an outer measure on (ft, Pa)- Since by step (hi), 1Z C V\,
we know by Caratheodory's lemma that A C V\, so that we can define \i
on A as the restriction of A to A. By step (ii), this is an extension of the
measure /x on 1Z. □
Here is an overview over the possible set of subsets of ft we have considered.

We also include the notion of ring and a-ring, which is often used in measure
theory and which differ from the notions of algebra or a-algebra in that
ft does not have to be in it. In probability theory, those notions are not
needed at first. For an introduction into measure theory, see [3, 37, 18].
Set of ft subsets contains closed under
topology 0,ft arbitrary unions, finite intersections
7r-system finite intersections
Dynkin system ft increasing countable union, difference
ring 0 complement and finite unions
a-ring 0 countably many unions and complement
algebra ft complement and finite unions
a-algebra 0,12 countably many unions and complement
Borel a-algebra 0,ft a-algebra generated by the topology
Remark. The name "ring" has its origin to the fact that with the "addition"
A + B = AAB = (A U B) \ (A n B) and "multiplication" A • B = A n B,
a ring of sets becomes an algebraic ring like the set of integers, in which
rules like A-(B + C) = A-B + A-C hold. The empty set 0 is the zero
element because AA0 = A for every set A. If the set ft is also in the ring,
one has a ring with 1 because the identity A 0 ft = A shows that ft is the
1-element in the ring.
Lets add some definitions, which will occur later:
Definition. A nonzero measure // on a measurable space (CI, A) is called

positive, if ji(A) > 0 for all A G A. If n+ ,u~ are two positive measures
so that ji(A) = u.+ — /j,~ then this is called the Hahn decomposition of /i.
A measure is called finite if it has a Hahn decomposition and the positive
measure |/z| defined by \/jl\(A) = n+(A) +ir(A) satisfies |/x|(ft) < oo.
Definition. Let v be a measure on the measurable space (CI, A). We write

v « fi if for every A in the a-algebra A, the condition ^(A) = 0 implies
u(A) = 0. One says that v is absolutely continuous with respect to /i.
2.2 Kolmogorov's 0-1 law, Borel-Cantelli lemma

Definition. Given a family {Ai}iei of a-subalgebras of A. For any nonempty
set J C I, let Aj := \fjeJAj be the a-algebra generated by \JjeJAj.
Define also A® = {0,ft}. The tail a-algebra T of {A}iei is defined as
T = f|JC// Ajc, where JC=X\X.
Theorem 2.2.1 (Kolmogorov's 0-1 law). If {Ai}iei are independent cr-

algebras, then the tail a-algebra T is P-trivial: P[A] = 0 or P[A] = 1 for
every AeT.
Proof, (i) The algebras Af and Ag are independent, whenever F,G C I

are disjoint.
2 . 2 . K o l m o g o r o v ' s 0 - 1 J a w, B o r e l - C a n t e l l i l e m m a 3 5
Proof. Define for H C I the 7r-system
XH = {AeA\A= f)Ai,KcfH,Aie Ai} .

ieK
The 7r-systems XF and JG are independent and generate the a-algebras Af

and Ag- Use lemma (2.1.5).
(ii) Especially: Aj is independent of Ajc for every J C I.
(hi) T is independent of Ai-
Proof. T = f|jc iAjc is independent of any Ak for K C/ /. It is
therefore independent to the 7r-system J/ which generates Ai. Use again
lemma (2.1.5).
(iv) T is a sub-a-algebra of Ai. Therefore T is independent of itself which
implies that it is P-trivial. ^
Example. Let Xn be a sequence of independent random variables and let

oo
A = {uo e ft | YXn converges} •

71=1
Then P[A] = 0 or P[A] = 1. Proof. Because Yln=i X™ converges if and

only if Y,n=NXn converges, A G a(AN, AN+!...) and so A G T, the
tail a- algebra defined by the independent a-algebras An = a(Xn). If for
example, Xn takes values ±l/ra, each with probability 1/2, then P[i4] = 0.
If Xn takes values dbl/n2 each with probability 1/2, then P[A] = 1. As you
might guess, the decision whether P[A] =0 or P[A] = 0 is related to the
convergence or divergence of a series. We will come back to that later.
Example. Let {An}neN be a sequence of of subsets of ft. The set

oo
Aoo := limsup An = Q (J An
n_>°° m=ln>m
consists of the set {u G ft} such that uj e An for infinitely many n G N. The
set ôo is contained in the tail a-algebra of An = {0,A^c,ft}. It follows
from Kolmogorov's 0 - 1 law that P[Aoo] G {0,1} if An e A and {An} are
P-independent.
Remark. In the theory of dynamical systems, a measurable map T : ft —▶ ft

of a probability space (ft, A, P) onto itself is called a if-system, if there
exists a a-subalgebra T C A which satisfies T C o^T(T)) for which the
sequence Tn = v(Tn(F)) satisfies fN = A and which has a trivial tail
a-algebra T = {0, ft}. An example of such a system is a shift map T(x)n =
xn+i on ft = AN, where A is a compact topological space. The K-system
property follows from Kolmogorov's 0 -1 law: take T = Vfcli ^(ô), with
To = {x e ft = Az | x0 = r e A }.
Theorem 2.2.2 (First Borel-Cantelli lemma). Given a sequence of events

An G A. Then
^P[^n]<OO^P[Aoo]=0.
neN
Proof. P[Aoo] = lining P[|Jfc>n Ak] < lim™ ^k>n p[^l = 0.

D
Theorem 2.2.3 (Second Borel-Cantelli lemma). For a sequence An G A of

independent events,
5^P[i4n]=00=»P[i4oo] = l.
n€N
Proof For every integer n G N,
k>n kfc>n
>n
= n(i-p[^])^nexp(-p[^])
k>n k>n
= exp(-5>[4t]).
k>n
The right hand side converges to 0 for n —» oo. From
p[^,]=p[U nAaÊpn^]=o
nGNfe>n n£N
follows Pfylgo] = 0. G
Example. The following example illustrates that independence is necessary

in the second Borel-Cantelli lemma: take the probability space ([0,1], B, P),
where P = dx is the Lebesgue measure on the Borel a-algebra B of [0,1],
For An = [0,1/n] we get A00 = ® and so P^] = 0. But because P[An] =
l/n we have Y!Z=i ^\An] = Yln=i h = °° because tne harmonic series
5Z£Li Vn diverges:
E^f^-iog^).
2 . 2 . K o l m o g o r o v ' s 0 — 1 l a w, B o r e l - C a n t e l l i l e m m a 3 7
Example. ("Monkey typing Shakespeare") Writing a novel amounts to en

ter a sequence of N symbols into a computer. For example, to write "Ham
let" , Shakespeare had to enter N = 180'000 characters. A monkey is placed
in front of a terminal and types symbols at random, one per unit time, pro
ducing a random sequence Xn of identically distributed sequence of random
variables in the set of all possible symbols. If each letter occurs with prob
ability at least e, then the probability that Hamlet appears when typing
the first N letters is eN. Call A\ this event and call Ak the event that
this happens when typing the (k - l)N + 1 until the fcAT'th letter. These
sets Ak are all independent and have all equal probability. By the second
Borel-Cantelli lemma, the events occur infinitely often. This means that
Shakespeare's work is not only written once, but infinitely many times. Be
fore we model this precisely, lets look at the odds for random typing. There
are 30^ possibilities to write a word of length N with 26 letters together
with a minimal set of punctuation: a space, a comma, a dash and a period
sign. The chance to write "To be, or not to be - that is the question."
with 43 random hits onto the keyboard is 1/10635. Note that the life time
of a monkey is bounded above by 131400000 ~ 108 seconds so that it is
even unlikely that this single sentence will ever be typed. To compare the
probability probability, it is helpful to put the result into a list of known
large numbers [10, 38].
104 One "myriad". The largest numbers, the Greeks were considering.
105 The largest number considered by the Romans.
1010 The age of the universe in years.
1017 The age of the universe in seconds.
1022 Distance to our neighbor galaxy Andromeda in meters.
1023 Number of atoms in two gram Carbon which is 1 Avogadro.
1027 Estimated size of universe in meters.
1030 Mass of the sun in kilograms.
1041 Mass of our home galaxy "milky way" in kilograms.
1051 Archimedes's estimate of number of sand grains in universe.
1080 The number of protons in the universe.
10100 One "googol". (Name coined by 9 year old nephew of E. Kasner).

10153 Number mentioned in a myth about Buddha.
10155 Size of ninth Fermat number (factored in 1990).
1010 Size of large prime number (Mersenne number, Nov 1996).
1010 Years, ape needs to write "hound of Baskerville" (random typing).
1010 Inverse is chance that a can of beer tips by quantum fluctuation.
1010 Inverse is probability that a mouse survives on the sun for a week.
-ir»50
10 Estimated number of possible games of chess.
1010 Inverse is chance to find yourself on Mars by quantum fluctuations
1010l°° One"Gogoolplex"
Lemma 2.2.4. Given a random variable X on a finite probability space A,

there exists a sequence X\,X2,... of independent random variables for
which all random variables Xi have the same distribution as X.
Proof. The product space ft = AN is compact by Tychonov's theorem. Let

A be the Borel-a-algebra on ft and let Qdenote the probability measure on
A. The probability measure P = Qz is defined on (ft, A) has the property
that for any cylinder set
Z(w) = {u e ft | ujk = rkluk+i = r-fc+i,... ,un = rn }
defined by a "word" w = [r^,,... rn],
P[Z(w)) = f[P[wi = ri] = f[Q({ri}).

i=k i=k
Finite unions of cylinder sets form an algebra TZ which generates a (TZ) = A.

The measure P is a-additive on this algebra. By Caratheodory's continu
ation theorem (2.1.6), there exists a measure P on (CI, A). For this proba
bility space (Ct,A,P), the random variables Xi(uS) — Ui) are independent
and have the same distribution as X. □
Example. In the example of the monkey writing a novel, the process of

authoring is given by a sequence of independent random variables Xn(u) =
un. The event that Hamlet is written during the time [Nk + 1, N(k + 1)]
is given by a cylinder set Ak> They have all the same probability. By the
second Borel-Cantelli lemma, P[Aoo] = 1. The set A^, the event that the
Monkey types this novel arbitrarily often, has probability 1.
Remark. Lemma (2.2.4) can be generalized: given any sequence of prob

ability spaces (R,5,Pi) one can form the product space (CI, A, P). The
random variables Xi(u) = Ui are independent and have the law Pi. For an
other construction of independent random variables is given in [105].
Exercice. In this exercise, we experiment with some measures on ft = N

[108].
a) The distance d(n,m) = \n — m\ defines a topology O on ft = N. What
is the Borel a-algebra A generated by this topology?
b) Show that for every A > 0
P[A] = ê-A"
neA
2.3. Integration, Expectation, Va r i a n c e 39
is a probability measure on the measurable space (CI, A) considered in a).
c) Show that for every s > 1
n£A
is a probability measure on the measurable space (ft,»4). The function
nefi
is called the Riemann zeta function.
d) Show that the sets Ap = {n G ft| p divides n} with prime p are indepen
dent. What happens if p is not prime.
e) Give a probabilistic proof of Euler's formula
p prime
f) Let A be the set of natural numbers which are not divisible by a square
different from 1. Prove
2.3 Integration, Expectation, Variance

In this entire section, (ft, A, P) will denote a fixed probability space.
Definition. A statement S about points u G ft is a map from ft to {true, false}.

A statement is said to hold almost everywhere, if the set P[{v \ S(u) =
false }] = 0. For example, the statement "let Xn —▶ X almost everywhere",
is a short hand notation for the statement that the set {x G ft | Xn(x) —>
X(x) } is measurable and has measure 1.
Definition. The algebra of all random variables is denoted by C. It is a

vector space over the field R of the real numbers in which one can multiply.
A elementary function or step function is an element of C which is of the
form
x = 5>,-u
2=1
with ai e R and where Ai G A are disjoint sets. Denote by S the algebra

of step functions. For X G S we can define the integral
E[X}:= f XdP = y2îP[Ai}.

Definition. Define C1 C C as the set of random variables X, for which
sup / Y dP
YeS,Y<\x\*
\X\J
is finite. For X e C1, we can define the integral or expectation
E[X] := f XdP= sup [ Y d P - s u p f Y d P,

J yeS,y<x+ J Ye S , Y < x - J
where I+ = IV0 = max(X,0) and X" = -X V 0 = max(-X,0). The

vector space C1 is called the space of integrable random variables. Similarly,
for p > 1 write Cp for the set of random variables X for which E[|X|P] < oo.
Definition. It is custom to write L1 for the space C1, where random vari
ables X,Y for which E[|X - Y\] = 0 are identified. Unlike Cp, the spaces
Lp are Banach spaces. We will come back to this later.
Definition. For X e C2,we can define the variance
Var[X] := E[(X - E[X})2} = E[X2) - E[X}2 .

The nonnegative number
a[X]=Var[X]1/2
is called the standard deviation of X.
The names expectation and standard deviation pretty much describe al

ready the meaning of these numbers. The expectation is the "average",
"mean" or "expected" value of the variable and the standard deviation
measures how much we can expect the variable to deviate from the mean.
Example. The m'th power random variable X(x) = xm on ([0,1], B, P) has

the expectation
1
E[X}= I':xmdx =
Jo
/o ra + 1 '
the variance
Va r | X ] = s p fl - E [ X ] > = ^ - ^ = { 1 + m f ( 1 + 2 m )
and the standard deviation L cr\X]

J ( l=+ nm.O vBoth
' U + ^the
m ) expectation
y
as well as the standard deviation converge to 0 if m -^ oo.
Definition. If X is a random variable, then E[Xm] is called the m'th mo

ment of X. The m'th central moment of X is defined as E[(X - E[X])rn}.
2.3. Integration, Expectation, Va r i a n c e 41
Definition. The moment generating function of X is defined as Mx(t) =
E[etx]. The moment generating function often allows a fast simultaneous
computation of all the moments. The function
KX(t)=\og(Mx(t))
is called the cumulant generating function.
Example. For X(x) = x on [0,1] we have both
° U
7 n =1l
m—1 . . r.a. =™0 —nC( mv + l ) '!
and
00 fmym 00 rfyml
M*W = E[e"] r=t E[£L* ]=£^_1.

m! ^—' m!
m=0 ra=0
Comparing coefficients shows E[Xm] = l/(m -f 1).
Example. Let ft = R. For given m G R, a > 0, define the probability

measure P[[a,6]] = / f(x) dx with
1 (x-m)2
/(X) = e 2*2 #
This is a probability measure because after a change of variables y =

(x—m)/(y/2a), the integral J^ f(x) dx becomes -4= f^ e~y dy = 1. The
random variable X(x) — x on (ft, A, P) is a random variable with Gaussian
distribution mean m and standard deviation a. One simply calls it a Gaus
sian random variable or random variable with normal distribution. Lets
justify the constants m and a: the expectation of X is E[X] = J X dP =
J^°00xf{x) dx = m. The variance is E[(X - m)2] = f™00x2f(x) dx = a2
so that the constant a is indeed the standard deviation. The moment gen
erating function of X is Mx(t) = emt+<T * /2. The cumulant generating
function is therefore nx(t) = mt + a2t2/2.
Example. If X is a Gaussian random variable with mean m = 0 and

standard deviation a, then the random variable Y = ex has the mean
E[ey] = e^l2. Proof:
1 A*00 2 1 yOO , 2n2
-L= I ey~& dy = e-2/2 1 / e^^~ dy = e*>/2
The random variable y has the log normal distribution.
Example. A random variable X G C2 with standard deviation a = 0 is a

constant random variable. It satisfies X(uj) = m for all u G ft.
Definition. If X G C2 is a random variable with mean m and standard
deviation a, then the random variable Y = (X — m)/o~ has the mean ■m = 0
and standard deviation a = 1. Such a random variable is called normalized.
One often only adjusts the mean and calls X — E[X] the centered random
variable.
Exercice. The Rademacher functions rn(x) are real-valued functions on

[0,1] defined by
f 1 2&=i < X < ^

rn(x) 1^ - 1n 2 *—
< X < 2fc+l
They are random variables on the Lebesgue space ([0,1], A, P = dx).

a) Show that 1—2x = X^Li ^r^. This means that for fixed x, the sequence
rn(x) is the binary expansion of 1 — 2x.
b) Verify that rn(x) = sign(sin(27r2n_1x)) for almost all x.
c) Show that the random variables rn(x) on [0,1] are IID random variables
with uniform distribution on {—1,1 }.
d) Each rn(x) has the mean E[rn] = 0 and the variance Var[rn] = 1.
11 h————i (I
1 6 I I
Figure. The Figure. The Figure. The

Rademacher Function Rademacher Function Rademacher Function
7*1 (x) r2(x) r3(x)
2.4 Results from real analysis

In this section we recall some results of real analysis with their proofs.
In the measure theory or real analysis literature, it is custom to write
/ f(x) d/j,(x) instead of E[X] or f,g,h,... instead of X, Y, Z,..., but this
is just a change of vocabulary. What is special about probability theory is
that the measures /i are probability measures and so finite.
2.4. Results from real analysis 43
Theorem 2.4.1 (Monotone convergence theorem, Beppo Levi 1906). Let Xn

be a sequence of random variables in C1 with 0 < X1 < X2,... and assume
X = \imn-+ocXn converges point wise. If supnE[Xn] < oo, then X G C1
and
E[X] = lim E[Xn] .
Proof. Because we can replace Xn by Xn - Xi, we can assume Xn > 0.

Find for each n a monotone sequence of step functions Xn,m £ S with
Xn = supm Xnim. Consider the sequence of step functions
Yn := SUp Xk,n < SUp Xk,n+1 < SUp Xk,n+1 = Yn+\ .

l<fc<n l<fc<n l<fc<n+l
Since Yn < sup™=1 Xm = Xn also E[yn] < E[Xn]. One checks that
supn Yn = X implies supn E[yn] = supyG£ y<x E[Y] and concludes
E[X] = sup E[y] = supE[yn] < supE[Xn] < E[supXn] = E[X) .

Ye S , Y < x n n n
We have used the monotonicity E[Xn] < E[Xn+i] in supnE[Xn] = E[X\.

□
Theorem 2.4.2 (Fatou lemma, 1906). Let Xn be a sequence of random

variables in C1 with \Xn\ < X for some X G C1. Then
E[liminf Xn] < liminf E[Xn] < limsupE[Xn] < E[limsupXn] .
Proof. For p>n,we have
inf Xm < Xp n m>n
Therefore
E[minf
> nXm] < E[XP] < E[sup
m > n Xm] .
Because p> n was arbitrary, we have also
E[ inf Xm] < inf E[XP] < supE[XJ < E[sup Xm] .
m^n P>n p>n m>n
Since Yn = infm>nXm is increasing with supnE[yn] < oo and Zn =

suPm>n^m is decreasing with infnE[Zn] > -co we get from Beppo-Levi
theorem (2.4.1) that Y = supnyn = limsupnXn and Z = infn Zn =
liminfnXn are in C1 and
E[\immfXn]
n =n supE[ minf
> nXm] < nsup inf " E[Xm]
r n > n = liminf
n E[Xn]
< limsupE[Xn] = inf sup E[Xm]
n n m>n
< inf E[ sup Xm] = E[limsupXn] .
n m>n n
Theorem 2.4.3 (Lebesgue's dominated convergence theorem, 1902). Let Xn

be a sequence in C1 with \Xn\ < Y for some Y G C . If Xn —> X almost
everywhere, then E[Xn] —> E[X].
Proof. Since X = liminfnXn = limsupnXn we know that X G C1 and

from Fatou lemma (2.4.2)
E [ X ) = E [ l i mni n f X n ] < l i m ni n f E [ X n ]
< limsup E[Xn] < E[limsup Xn] = E[X].
n n
A special case of Lebesgue's dominated convergence theorem is when Y =

K is constant. The theorem is then called the bounded dominated conver
gence theorem. It says that E[Xn] —> E[X] if Xn < K and Xn —> X almost
everywhere.
Definition. Define also for p G [1, oo) the vector spaces Cp = {X G C \ \X\P G
C1 } and C°° = {X G C \ 3K G R X < K, almost everywhere }.
Example. For ft = [0,1] with the Lebesgue measure P = dx and Borel

a-algebra A, look at the random variable X(x) = xa, where a is a real
number. Because X is bounded for a > 0, we have then X G £°°. For
a < 0, the integral E[|X|P] = /0 xap dx is finite if and only if ap < 1 so
that X is in Cp whenever p > 1/a.
2.5 Some inequalities

Definition. A function h : R —*• R is called convex, if there exists for all
x0 e R a linear map l(x) = ax + b such that l(x0) = h(x0) and for all x G R
the inequality l(x) < h(x) holds.
2.5. Some inequalities 45
Example. h(x) = x2 is convex, h(x) = ex is convex, h(x) = x is convex.

h(x) = -x2 is not convex, h(x) = x3 is not convex on R but convex on
R+ = [0,00).
Figure. The Jensen inequality in

the case ft = {u,v}, P[{^}] =
P[{v}} = 1/2 and with X(u) =
a,X(v) = b. The function h in
this picture is a quadratic func
tion of the form h(x) = (x — s)2 +
t.
E[X]=(a+b)/2
Theorem 2.5.1 (Jensen inequality). Given X e C1. For any convex function
h : R -> R, we have
E[h(X)} > h(E[X}) ,
where the left hand side can also be infinite.
Proof Let I be the linear map defined at x0 = E[X}. By the linearity and
monotonicity of the expectation, we get
h(E[X}) = l(E{X]) = E[l(X)} < E[h(X)] .

D
Example. Given p < q. Define h(x) = \x\q/p. Jensen's inequality gives

E[|X|«] = E[h{\X\P)} < h(E[\X\P] = E[\X\P}i/r. This implies that ||A-||, :=
EOXJ"]1^ < EflXlP]1/" = ||X||P for p < q and so
£°° c CP c Cq C C1
for p > q. The smallest space is £°° which is the space of all bounded
random variables.
Exercice. Assume X is a nonnegative random variable for which X and

l/X are both in C1. Show that E[X + 1/X] > 2.
We have defined Cp as the set of random variables which satisfy E[\X\P] <
oo for p e [1, oo) and \X\ < K almost everywhere for p = oo. The vector
space Cp has the semi-norm \\X\\P = E[|X|p]Vp rsp. H^H^ = inf{K G
^ I |-^| < K almost everywhere }.
Definition. One can construct from Cp a real Banach space Lp = Cp/Af
which is the quotient of Cp with M = {X G Cp \ \\X\\P = 0 }. Without this
identification, one only has a pre-Banach space in which the property that
only the zero element has norm zero is not necessarily true. Especially, for
p = 2, the space L2 is a real Hilbert space with inner product < X, Y >=
E[XY}.
Example. The function f(x) = lq(x) which assigns values 1 to rational

numbers x on [0,1] and the value 0 to irrational numbers is different from
the constant function g(x) = 0 in Cp. But in Lp, we have / = g.
The finiteness of the inner product follows from the following inequality:
Theorem 2.5.2 (Holder inequality, Holder 1889). Given p,q e [l,oo] with
P'1 + q-1 = 1 and X G C? and Y G Cq. Then XY e C1 and
\ \ X Y \ U K l \ X U Y \ \q
Proof. Without loss of generality, we can restrict the situation to X, Y > 0

and ||X||p > 0. Define the probability measure
Q=XPP
E[Xp]
and define u = 1{a>o}57*p_1- Jensen's inequality
sen's in € gives Q{u)q < Q(vfl) so
that
E[|^|]<||A:||p||i{z>0}y||,<||x||p||y||,.
□
A special case of Holder's inequality is the Cauchy-Schwarz inequality
ll*l%<||*||2-|M|2-
The semi-norm property of Cp follows from the following fact:
Theorem 2.5.3 (Minkowski inequality (1896)). Given p e [1, oo] and X, Y e

CP. Then
II*+ *%< 11*11,+ IMI„.
2.5. Some inequalities
Proof. We use Holder's inequality from below to get
e[|x + Yf] = e[\x\\x + rr1] + E[|y||* + rr1] < ll*llPc + ||r||pc,

where C = \\\X + ^|p_1||9 = E[\X + Y\p]^q which leads to the claim. D
Definition. We use the short-hand notation P[X > c] for P[{w £ fi | I(w) >
Mi-
Theorem 2.5.4 (Chebychev-Markov inequality). Let h be a monotone func

tion on K with h > 0. For every c> 0, and h(X) G £l we have
h{c)-P[X>c] <E[h{X)].
Proof. Integrate the inequality h <h(X) and use the monotonicity

and linearity of the expectation. □
Figure. The proof of the

Chebychev-Markov inequality in
I I
the case h(x) = x. The left hand
side h(c) ■ P[X > c) is the area of
the rectangles {X > c} x [0, h(x)}
and E[h(X)] = E[X] is the area
under the graph of X.
Example. h(x) = |x| > c] < ||X||i/c which implies for

example the statement
= 0 => P[X = 0] = 1
Exercice. Prove the Chernoff bound
t>„ e~tcMx{t)
where Mx(t) = E[ext] is the moment generating function of X.

An important special case of the Chebychev-Markov inequality is the Cheby-
chev inequality:
Theorem 2.5.5 (Chebychev inequality). If X G C2, then
P[|X-E[X]|>c]<^pl.
Proof Take h(x) = x2 and apply the Chebychev-Markov inequality to the

random variable Y = X - E[X] e C2 satisfying h(Y) eC1. □
Definition. For X,Y e C2 define the covariance
Cov[X, Y] := E[(X - E[X])(Y - E[Y])} = E[XY] - E[X]E[Y] .
Two random variables in C2 are called uncorrected if Cov[X, Y] = 0.
Example. We have Cov[X,X] = Var[X] = E[(X - E[X})2} for a random

variable X e C2.
Remark. The Cauchy-Schwarz-inequality can be restated in the form
\Cov[X,Y]\<a[X]a[Y]
Definition. The regression line of two random variables X, Y is defined as

y = ax + b, where
Iff2 = {l,...,n}isa finite set, then the random variables X, Y define the
vectors
X = (X(l),...,X(n)),Y = (Y(l),...,Y(n))
or n data points (X(i),Y(i)) in the plane. As will follow from the proposi
tion below, the regression line has the property that it minimizes the sum
of the squares of the distances from these points to the line.
2.5. Some inequalities 49
Figure. Regression line com

puted from a finite set of data
points (X(i),Y(i)).
Example. If X, Y are independent, then a = 0. It follows that b = E[Y].

Example. If X = Y, then a = 1 and 6 = 0. The best guess for Y is X.
Proposition 2.5.6. If y = ax -f b is the regression line of of X, Y, then the

random variable Y = aX + b minimizes Var[Y - Y] under the constraint
E[Y] = E[Y] and is the best guess for Y, when knowing only E[Y] and
Cov[X, Y]. We check Cov[X, Y] = Cov[X, y ].
Proo/. To minimize Yar[aX-\-b—Y] under the constraint E[aX-\-b—Y] = 0 is

equivalent to find (a, b) which minimizes /(a, 6) = E[(aX + b — Y)2] under
the constraint g(a,b) = E[aX + b - Y] = 0. This least square solution
can be obtained with the Lagrange multiplier method or by solving b =
E[Y]-aE[X] and minimizing h(a) = E[(aX-Y-E[aX-Y])2} = a2(E[X2]-
E[X]2)-2a(E[XY}-E[X]E[Y}) = a2Var[X]-2aCov[X,Y]. Setting h'(a) =
0 gives a = Cov[X, y ] / Va r [ X ] . □
Definition. If the standard deviations cr[X],cr[Y] are both different from

zero, then one can define the correlation coefficient
Cov[X,Y]
Corr[X,Y] =
a[X]a[Y]
which is a number in [—1,1]. Two random variables in C2 are called un-
correlated if Corr[X,F] = 0. The other extreme is |Corr[X,y]| = 1, then
Y = aX + b by the Cauchy-Schwarz inequality.
Theorem 2.5.7 (Pythagoras). If two random variables X,Y G C2 are

independent, then Cov[X, Y] = 0. If X and Y are uncorrelated, then
Var[X + Y]= Var[X] + Var[y].
Proof. We can find monotone sequences of step functions

n n
Xn = 5>ilAi - X ,Yn = J2& ' 1* "> Y '

i=l j=l
We can choose these functions in such a way that Ai e A = cr(X) and

Bj e B = o~(Y). By the Lebesgue dominated convergence theorem (2.4.3),
E[Xn] —▶ E[X] and E[yn] —▶ E[y] almost everywhere. Compute Xn •
Fn = ^^^xCKi^lAiDBj- By the Lebesgue dominated convergence theo
rem (2.4.3) again, E[Xnyn] —> E[Xy]. By the independence of X,Y we
have E[Xnyn] = E[Xn) • E[yn] and so E[XY] = E[X]E[F] which implies
Cov[X, Y] = E[XY] - E[X] • E[Y] = 0.
The second statement follows from
Var[X + y] = Var[X] + Var[y] + 2 Cov[X, Y] .
Remark. If fi is a finite set, then the covariance Cov[X, Y] is the dot prod
uct between the centered random variables X — E[X] and Y — E[Y\, and
cr[X] is the length of the vector X — E[X] and the correlation coefficient
Corr[X,y] is the cosine of the angle a between X - E[X] and Y - E[Y]
because the dot product satisfies v- w = \v\\w\ cos(a). So, uncorrelated
random variables X, Y have the property that X — E[X] is perpendicular
to y — E[y]. This geometric interpretation explains, why lemma (2.5.7) is
called Pythagoras theorem.
For more inequalities in analysis, see the classic [29, 58]. We end this sec
tion with a list of properties of variance and covariance:
Var[X] > 0.
Vai[X] = E[X2}-E[X}2.
Var[AX] = A2Var[X].
Var[X + y] = Var[X] + Var[y] + 2Cov[X, Y}. Corr[X, Y] G [0,1].
Cov[X, y] = E[XY] - E[X]E[y].
Cov[X,y] <a[X]a[Y}.
Corr[X, y] = 1 if X - E[X] = Y - E[Y]
2.6 The weak law of large numbers

Consider a sequence X\, X2,... of random variables on a probability space
(Cl,A,P). We are interested in the asymptotic behavior of the sums Sn =
Xi + X2 + • • • + Xn for n —▶ oo and especially in the convergence of the
averages Sn/n. The limiting behavior is described by "laws of large num
bers". Depending on the definition of convergence, one speaks of "weak"
2.6. The weak law of large numbers 51
and "strong" laws of large numbers.
We first prove the weak law of large numbers. There exist different ver
sions of this theorem since more assumptions on Xn can allow stronger
statements.
Definition. A sequence of random variables Yn converges in probability to

a random variable Y, if for all e > 0,
lim p[|y„-y|>c] = o.
71—>-00
One calls convergence in probability also stochastic convergence.
Remark. If for some p G [l,oo), \\Xn - X\\p -> 0, then Xn -» X in

probability since by the Chebychev-Markov inequality (2.5.4), P[|Xn-X| >
e]<\\X-Xn\\p/ep.
Exercice. Show that if two random variables X,Y G C2 have non-zero

variance and satisfy |Corr(X,y)| = 1, then Y = aX + b for some real
numbers a,b.
Theorem 2.6.1 (Weak law of large numbers for uncorrelated random vari
ables). Assume Xi G C2 have common expectation E[Xi] = m and satisfy
suPn n Ya=i Vaxpfi] < °°. If xn are pairwise uncorrelated, then ^ —▶ m
in probability.
Proof Since Var[X + Y] = Var[X] + Var[y] + 2 • Cov[X,y] and Xn are

pairwise uncorrelated, we get Var[Xn + Xm] = Var[Xn] + Var[Xm] and by
induction Var[5n] = £?=1 Vax[Xn]. Using linearity, we obtain E[Sn/n] = m
and
Vflr(^]
L n_ n%
n 2 i- S%1!
n2 .n ^&1
2 n_
2 -I ±
f - 'vM[x„l.
The right hand side converges to zero for n —▶ oo. With Chebychev's in
equality (2.5.5), we obtain
P [ni ^ _ T O i > ee] 2< ^ .

D
As an application in analysis, this leads to a constructive proof of a theorem
of Weierstrass which states that polynomials are dense in the space C[0,1]
of all continuous functions on the interval [0,1]. Unlike the abstract Weier
strass theorem, the construction with specific polynomials is constructive
and gives explicit formulas.
Figure. Approximation of a
function f(x) by Bernstein poly
nomials B2,B5, B10, B2Q, B30.
Theorem 2.6.2 (Weierstrass theorem). For every / G C[0,1], the Bernstein

polynomials
B^) = E/(^)(r)^(i-^r-fc
converge uniformly to /. If f(x) > 0, then also Bn(x) > 0.
Proof. For x £ [0,1], let Xn be a sequence of independent {0,1}- valued

random variables with mean value x. In other words, we take the proba
bility space ({0,1}N,.4,P) denned by P[wn = 1] = x. Since P[Sn = k] =
we can write Bn(x) = E[/(^)]. We estimate
^pk{l-p)n-k,
\Bn(x)-f(x)\ = \E[fn
( ^ ) } - f ( x ) } \ < E [ \ f n( ^ ) - f ( x )
< 2||/||.P[|^-x|><5]
n
+ s u p \ f ( x ) - f ( y ) \ . - p [I\*->n
ZL-x\<6\
\x-y\<6 n
< 2||/||-P[|^-x|><5]
+ sup \f(x)-f(y)\.
\x-y\<8
The second term in the last line is called the continuity module of /. It
converges to zero for S —▶ 0. By the Chebychev inequality (2.5.5) and the
proof of the weak law of large numbers, the first term can be estimated
from above by
,Var[X,]
n62 '
a bound which goes to zero for n —▶ oo because the variance satisfies
\ar[Xi] = x(l -x)< 1/4. □
In the first version of the weak law of large numbers theorem (2.6.1), we
only assumed the random variables to be uncorrelated. Under the stronger
condition of independence and a stronger conditions on the moments (X4 G
C1), the convergence can be accelerated:
Theorem 2.6.3 (Weak law of large numbers for independent L4 random

variables). Assume Xi G C4 have common expectation E[Xi] = m and
satisfy M = supn ||X||4 < oo. If Xi are independent, then Sn/n —▶ m in
probability. Even Y^>=1 P[|^f ~m\>e] converges for all e > 0.
Proof. We can assume without loss of generality that m = 0. Because the

Xi are independent, we get
n
E[£n] = X/ ElX^X^X^Xi^ .
ii ,^2^3^4 = 1
Again by independence, a summand ElX^X^X^X^] is zero if an index

i = ik occurs alone, is E[X4] if all indices are the same and E[X?]E[X?], if
there are two pairwise equal indices. Since by Jensen's inequality E[X2]2 <
E[Xf] < M we get
E[Sn] < nM + n(n + l)Af .
Use now the Chebychev-Markov inequality (2.5.4) with h(x) — x4 to get
PA > e l < n ( S n / n ) * \
n
^ ^ n + n <2 2M—
,, 1
< M—j—r-
We can weaken the moment assumption in order to deal with C1 random

variables. Of course, the assumptions have to be made stronger at some
other place.
Definition. A family {Xi}iei of random variables is called uniformly inte-
grable, if sup^/E^jî^r] —▶ 0 for R —▶ oo. A convenient notation which
we will quite often use in the future is E[1^X] = E[X; A] for X e C1 and
AeA.
Theorem 2.6.4 (Weak law for uniformly integrable, independent L1 random

variables). Assume Xi G C1 are uniformly integrable. If Xi are indepen
dent, then ^ ^"=1(Im — E[Xm]) —> 0 in C1 and therefore in probability.
Proof Without loss of generality, we can assume that E[Xn] = 0 for all
n e N, because otherwise Xn can be replaced by Yn = Xn — E[Xn], Define
/#(£) = tl[-R,R], the random variables
XW = fR{Xn) - E[fR(Xn)}, Y™ =Xn- XW
as well as the random variables
n . n .i = \
i=i
We estimate, using the Minkowski and Cauchy-Schwarz inequalities
||Sn||l < 115^11! + HT^Hx

< ||5ifl)||2 + 2 sup E[|X,|; |X,| > R]
KKn
< -R=+2sapE[\Xl\;\Xi\>R].
In the last step we have used the independence of the random variables and
E[xiR)] = 0 to get
The claim follows from the uniform integrability assumption

supi€NE[|Xz|; \Xl\>R]-+0foTR-+oo- □
A special case of the weak law of large numbers is the situation, where all
the random variables are IID:
Theorem 2.6.5 (Weak law of large numbers for IID L1 random variables).
Assume Xi G C1 are IID random variables with mean m. Then Sn/n —> m
in C1 and so in probability.
Proof We show that a set of IID C1 random variables is uniformly inte

grable: given X G C1, we have K • P[|X| > K] < \\X\\i so that P[|X| >
K] -> 0 for K -> oo.
Because the random variables Xi are identically distributed, P[|X»|; \Xi\ >
R] is independent of i. Consequently any set of IID random variables is also
uniformly integrable. We can now use theorem (2.6.4). □
Example. The random variable X(x) = x2 on [0,1] has the expectation

m = E[X] = Jo x2 dx = 1/2. For every n, we can form the sum Sn/n =
(x\ + x2 H h xn)/n. The weak law of large numbers tells us that P[|Sn -
1/2| > e] —▶ 0 for n —> oo. Geometrically, this means that for every e > 0,
the volume of the set of points in the n-dimensional cube for which the
distance r(xi,..,xn) = y/x\ + • • • + x2n to the origin satisfies ^Jn~J2 - e <
r < yjnj2 + e converges to 1 for n -* oo. In colloquial language, one
could rephrase this that asymptotically, as the number of dimensions to go
infinity, most of the weight of a n-dimensional cube is concentrated near a
shell of radius l/>/2 ~ 0.7 times the length y/n of the longest diagonal in
the cube.
Exercice. Show that if X, Y G C1 are independent random variables, then

XY e C1. Find an example of two random variables X,Y G C1 for which
XY {C1.
Exercice. a) Given a sequence pn e [0,1] and a sequence Xn of IID random

variables taking values in {-1,1} such that P[Xn = 1} = pn and P[Xn =
— 1] = 1 — pn. Show that
-T(Xk-mk)^0
nfe=i
in probability, where mk = 2pk — 1.

b) We assume the same set up like in a) but this time, the sequence pn is
dependent on a parameter. Given a sequence Xn of independent random
variables taking values in {-1,1} such that P[Xn = 1] = Pn and P[Xn =
-1] = 1 - pn with pn = (1 + cos[0 + na])/2, where 6 is a parameter. Prove
that - Y,n Xn -> ° in £l for almost all 6. You can take for granted the fact
that - J2k=i Pk -> V2 f°r almost all real parameters 6 G [0,27r]
Exercice. Prove that Xn —> X in C1, then there exists of a subsequence

Yn = Xnk satisfying Yn —> X almost everywhere.
Exercice. Given a sequence of random variables Xn. Show that Xn con
verges to X in probability if and only if
E[ I*""*' 1 _+ o
[l-r\Xn-X\l
for n • oo.
Exercice. Give an example of a sequence of random variables Xn which

converges almost everywhere, but not completely.
Exercice. Use the weak law of large numbers to verify that the volume of
an n-dimensional ball of radius 1 satisfies Vn -▶ 0 for n -> oo. Estimate,
how fast the volume goes to 0. (See example (2.6))
2.7 The probability distribution function

Definition. The law of a random variable X is the probability measure p on
R defined by p(B) = Ppf"1^)] for all B in the Borel a-algebra of R. The
measure p is also called the push-forward measure under the measurable
map X : Ct -> R.
Definition. The distribution function of a random variable X is defined as
Fx(s) = p((-oc,s]) = P[X < s] .

The distribution function is sometimes also called cumulative density func
tion (CDF) but we do not use this name here in order not to confuse it
with the probability density function (PDF) fx(s) = F^(s) for continuous
random variables.
Remark. The distribution function F is very useful. For example, if X is a

continuous random variable with distribution function F, then Y = F(X)
has the uniform distribution on [0,1]. We can reverse this. If we want to pro
duce random variables with a distribution function F, just take a random
variable Y with uniform distribution on [0,1] and define X = F~1(Y). This
random variable has the distribution function F because {X e [a, b] } =
{F~\Y) G [a,b] } = {Ye F([a,b}) } = {Y e [F(a),F(b)}} = F(b) - F(a).
We see that we need only to have a random number generator which pro
duce^ uniformly distributed random variables in [0,1] to produce random
variables with a given continuous distribution.
2.7. The probability distribution function 57
Definition. A set of random variables is called identically distributed, if
each random variable in the set has the same distribution function. It is
called independent and identically distributed if the random variables are
independent and identically distributed. A common abbreviation for inde
pendent identically distributed random variables is IID.
Example. Let Cl = [0,1] be the unit interval with the Lebesgue measure p
and let m be an integer. Define the random variable X(x) = x™. One calls
its distribution a power distribution. It is in C1 and has the expectation
E[X] = l/(ra + 1). The distribution function of X is Fx(s) = s^/™"* on
[0,1] and Fx(s) = 0 for s < 0 and Fx(s) = 1 for s > 1. The random
variable is continuous in the sense that it has a probability density function
fx(s) = F'x(s) = sVn-i/m so that Fx(s) = /^ fx(t) dt.
0.2 0.4 0.6 0. 0.2 0.4 0.6 0.8 1 1.2
Figure. The distribution function Figure. The density function

Fx(s) of X(x) = xm in the case fx(s) of X(x) = xm in the case
ra = 2. m = 2.
Given two IID random variables X, Y with the m'th power distribution as
above, we can look at the random variables V = X+Y, W = X-Y. One can
realize V and W on the unit square Cl = [0,1] x [0,1] by V(x, y) = xm + ym
and W(x,y) = xm - ym. The distribution functions Fv(s) = P[V < s] and
Fw(s) = P[V < s] are the areas of the set A(s) = {(x,y) | xm + ym < s }
and B(s) = {(x, y) \ xm - ym < s }.
Chapter 2. Limit theorems
Figure. Fy(s) is the area of the Figure. Fw(s) is the area of the
set A(s), shown here in the case set B(s), shown here in the case
m = 4. m — 4.
We will later see how to compute the distribution function of a sum of in

dependent random variables algebraically from the probability distribution
function Fx ■ From the area interpretation, we see in this case
/„* {s-xm)^mdx, «e[0,l]

jLlv/ml-(s-xm)l^dx, s€[l,2]
JO
Fw(s) =
f}Um l-(xm- sf/m dx, s G [0,1]
Figure. The function Fv(s) with Figure. The function Fw(s) with
density (dashed) fv{s) of the sum density (dashed) fw{s) of the dif
of two power distributed random ference of two power distributed
variables with m = 2. random variables with m = 2.
2.8. Convergence of random variables 59
Exercice. a) Verify that for 9 > 0 the Maxwell distribution
f(x) = A03/2x2e-**2
is a probability distribution on R+ = [0,oo). This distribution can model

the speed distribution of molecules in thermal equilibrium.
a) Verify that for 6 > 0 the Rayleigh distribution
f(x) = 26xe-9x2
is a probability distribution on R+ = [0,oo). This distribution can model

the speed distribution y/X2 + Y2 of a two dimensional wind velocity (X, Y),
where both X, Y are normal random variables.
2.8 Convergence of random variables

In order to formulate the strong law of large numbers, we need some other
notions of convergence.
Definition. A sequence of random variables Xn converges in probability to

a random variable X, if
F[\Xn-X\>e]^0
for all e > 0.
Definition. A sequence of random variables Xn converges almost every

where or almost surely to a random variable X, if P[Xn -> X] = 1.
Definition. A sequence of CP random variables Xn converges in Cp to a

random variable X, if
||Xn.-X||p-0
for n —▶ oo..
Definition. A sequence of random variables Xn converges fast in probabil

ity, or completely if
£>[|Xn-X|>e]<co
n
for all e > 0.
We have so four notions of convergence of random variables Xn —▶ X, if the
random variables are defined on the same probability space (Cl, A, P). Later
we will the two equivalent but weaker notions convergence in distribution
and weak convergence, which not necessarily assume Xn and X to be de
fined on the same probability space. Lets add these two definitions also
here. We will see later, in theorem (2.13.2) that the following definitions
are equivalent:
Definition. A sequence of random variables Xn converges in distribution,

if FXri(x) —> Fx(x) for all points s, where Fx is continuous.
Example. Let Cln = {1,2,..., n) with the uniform distribution P[{fc}] = 1/n
and Xn the random variable Xn(x) = x/n. Let X(x) = x on the probability
space [0,1] with probability P[[a, b)] = b — a. The random variables Xn and
X are defined on a different probability spaces but Xn converges to X in
distribution for n —> oo.
Definition. A sequence of random variables Xn converges in law to a ran

dom variable X, if the laws pn of Xn converge weakly to the law p of
X.
Remark. In other words, Xn converges weakly to X if for every continuous

function / on R of compact support, one has
/ f(x) dpn(x) -+ / f(x) dp(x) .

2.8. Convergence of random variables
61
Proposition 2.8.1. The next figure shows the relations between the different
convergence types.
0) In distribution = in law
Fxn(s) -> Fx(s), Fx cont. at s
1) In probability
P [ | X „ - X | > e ] ^ 0 , Ve > 0 .
2) Almost everywhere 3) In £p
P[Xn - X] = 1 ||*n-A-||p-0
4) Complete
E „ P [ | X n - * | > e ] < o o , Ve > 0
Proof. 2) =▶ 1): Since
{Xn->X} = f){Jf){\Xn-X\<l/k}
k m n>m
"almost everywhere convergence" is equivalent to
1 = PflJ fl J{l*»
A n -~X\<\}]=
A| < Jim P[ fl {\Xn - X\ m n>r
for all k. Therefore
P[|*m - *| > e] < P[ fl {|X„ - X\ > e }} -+ 0

n>m
for all e > 0.

4) => 2): The first Borel-Cantelli lemma implies that for all e > 0
P[\Xn - X\> e, infinitely often] = 0 .

We get so for en —> 0
P[(J |xn-X| > ek, infinitely often] < ^P[|Xn-X| > ek, infinitely often] =
n n
from which we obtain P[Xn —▶ X] = 1.

3) => 1): Use the Chebychev-Markov inequality (2.5.4), to get
, , E[\Xn-X\P]
P[[Xn-Xl>6]< u ngp M.
D
Example. Here is an example of convergence in probability but not almost

everywhere convergence. Let ([0,1], A,P) be the Lebesgue measure space,
where A is the Borel a-algebra on [0,1]. Define the random variables
Xn,k = l[fc2-",(fc+i)2-"]> n = 1,2,..., fc = 0, ...,2 - 1 .
By lexicographical ordering X\ — X\,\,Xoo n—>oo
but P[|Xn,fc > c] < 2"n.
Example. And here is an example of almost everywhere but not Cp con

vergence: the random variables
Xn = 2nl[o,2-n]
on the probability space ([0,1],A,P) converge almost everywhere to the

constant random variable X = 0 but not in Cp because ||Xn||p = 2n(p 1)/p.
With more assumptions other implications can hold. We give two examples.
Proposition 2.8.2. Given a sequence Xn £ C°° with \\Xn\\oo < K for all n,
then Xn -> X in probability if and only if Xn -> X in C .
Proof, (i) P[|X| < K) = 1. Proof. For k G N,
P[|X| > tf + i] < P[|X - Xn\ > ^] - 0, n - oo
so that P[|X| > K + £] = 0. Therefore
P[|X|>K] = P[U{|X|>K + i}] = 0.

2.8. Convergence of random variables 63
(ii) Given e > 0. Choose m such that for all n > m
Then, using (i)
E[|Xn-X|) = E[(|Xn-X|;|Xn-X|>e]+E[(|Xn-X|;iXn-X|<6]
< 2^P[|Xn-X|>^] + i<e.
Definition. A family C C C1 of random variables is called uniformly inte

grable, if
lim sup E[l|X|>jR] -^0
R^°°xeC
for all X e C. The next lemma was already been used in the proof of the
weak law of large numbers for IID random variables.
Lemma 2.8.3. Given X £ C1 and e > 0. Then, there exists K > 0 with
E[|X|;|X|>^]<e.
Proof. Given e > 0. If X e C1, we can find 6 > 0 such that if P[A] < S,
then E[\X\,A] < e. Since KP[\X\ > K] < E[|X|], we can choose K such
that P[|X| >K)<6. Therefore E[|X|; |X| > K] < e. □
The next proposition gives a necessary and sufficient condition for C1 con
vergence.
Proposition 2.8.4. Given a sequence random variables Xn e C1. The fol

lowing is equivalent:
a) Xn converges in probability to X and {Xn}n€N is uniformly integrable.
b) Xn converges in C1 to X.
Proof, a) => b). Define for K > 0 and a random variable X the bounded
variable
X(K) = X ' 1{-k<x<k} + K • 1{X>K} - K • 1{X<-K} .
By the uniform integrability condition and the above lemma (2.8.3), we

can choose K such that for all n,
E[|JfW-Xn|]<|,E[|A-W-X|]<|.
Since \XnK) - XW| < |Xn - X|, we have XnK) - X<*> in probability.
We have so by the last proposition (2.8.2) XnK) -> X(i° in C1 so that for
n>m E[\XnK) - X(/°|] < e/3. Therefore, for n > m also
E[|Xn-X|]<E[|Xn-XW|]+E[|xW-XW|] + E[|XW-X|]<6.
a) => b). We have seen already that Xn -> X in probability if ||Xn-X||i ->
0. We have to show that Xn -> X in C1 implies that Xn is uniformly
integrable.
Given e > 0. There exists m such that E[|Xn - X|] < e/2 for n > m. By
the absolutely continuity property, we can choose S > 0 such that P[A] <e
implies
E[|Xn|; A] < e, 1 < n < m,E[|X|; A] < e/2 .
Because Xn is bounded in C1, we can choose K such that K"1 supn E[|Xn|] <
5 which implies E[|Xn| > K) < S. For n > m, we have therefore
E[|Xn|; |Xn| >K}< E[|X|; |Xn| > K] + E[|X - Xn|] < e .

□
Exercice. a) P[supfc>n \Xk - X\ > c] -> 0 forn -^ oo and all e > 0 if and
only if Xn —▶ X almost everywhere.
b) A sequence Xn converges almost surely if and only if
lim P[sup \Xn+k - Xn| > e] = 0
^-^°° fc>i
for all e > 0.
2.9 The strong law of large numbers

The weak law of large numbers makes a statement about the stochastic
convergence of sums
Sn X\ + • • • + Xn
n n
of random variables Xn. The strong laws of large numbers make analog
statements about almost everywhere convergence.
The first version of the strong law does not assume the random variables to
have the same distribution. They are assumed to have the same expectation
and have to be bounded in C .
Theorem 2.9.1 (Strong law for independent Z^-random variables). Assume

Xn are independent random variables in £4 with common expectation
E[Xn] = m and for which M = supn ||Xn||| < oo. Then Sn/n -> m almost
everywhere.
2.9. The strong law of large numbers 65
Proof. In the proof of theorem (2.6.3), we derived
P[|^-m|>6]<2MJ_
This means that Sn/n —▶ m converges completely. By proposition (2.8) we

have almost everywhere convergence. □
Here is an application of the strong law:
Definition. A real number x G [0,1] is called normal to the base 10, if its
decimal expansion x = X\X<i... has the property that each digit appears
with the same frequency 1/10.
Corollary 2.9.2. (Normality of numbers) On the probability space

([0,1], B, Q = dx), Lebesgue almost all numbers x are normal.
Proof. Define the random variables Xn(x) = xn, where xn is the n'th
decimal digit. We have only to verify that Xn are IID random variables. The
strong law of large numbers will assure that almost all x are normal. Let Cl =
{0,1,..., 9 }N be the space of all infinite sequences u = (u)\,u>2, ^3,...).
Define on Cl the product cr-algebra A and the product probability measure
P. Define the measurable map S(u>) = ]C^Li ^fc/lO^ = x from Cl to [0,1].
It produces for every sequence in Cl a real number x G [0,1]. The integers
LJk are just the decimal digits of x. The map S is measure preserving and
can be inverted on a set of measure 1 because almost all real numbers have
a unique decimal expansion.
Because Xn(x) = Xn(S(u)) = Yu(uj) = wn, if S(u) = x. We see that Xn
are the same random variables than Yn. The later are by construction IID
with uniform distribution on{0,l,...,9}. □
Remark. While almost all numbers are normal, it is difficult to decide

normality for specific real numbers. One does not know for example whether
tt - 3 = 0.1415926... of y/2 - 1 = 0.41421... is normal.
The strong law for IID random variables was first proven by Kolmogorov
in 1930. Only much later in 1981, it has been observed that the weaker
notion of pairwise independence is sufficient [25].
Theorem 2.9.3 (Strong law for pairwise independent L1 random variables).

Assume Xn G C1 are pairwise independent and identically distributed ran
dom variables. Then Sn/n —▶ E[Xi] almost everywhere.
Proof. We can assume without loss of generality that Xn > 0 (because we

can split Xn = X+ + X~ into its positive X+ = Xn V 0 = max(Xn, 0) and
negative part X~ = —X V 0 = max(—X, 0). Knowing the result for X^
implies the result for Xn.).
Define /#(£) = t • 1[-r,r\, the random variables Xn = /r(Xu) and Yn =
Xn as well as
1 n i n
bn = ~~ / J X{, ln = — / Y% •
(i) It is enough to show that Tn — E[Tn] —▶ 0.

Proof. Since E[Yn] —> E[Xi] = m, we get E[Tn] —▶ m. Because
^P[Yn^Xn] < £P[X»>n] = £P[Xi>n]

n>l n >nl > l n>l
n>lk>n
= J3fc.P[XiG[Jfe,*+l]]<E[Jfi]<oo,
fe>l
we get by the first Borel-Cantelli lemma that P[Yn ^ Xn, infinitely often] =
0. This means Tn — Sn —▶ 0 almost everywhere, proving E[Sn] —> m.
(ii) Fix a real number a > 1 and define an exponentially growing subse
quence kn = [an] which is the integer part of an. Denote by p the law of
the random variables Xn. For every e > 0, we get using Chebychev inequal
ity (2.5.5), pairwise independence for kn = [an] and constants C which can
vary from line to line:
fn > [ | 7 k - En =[ lr f e j | > Ce ] < f K; n^

=l
- £ Var[Ym]
n m=l
1 °° 1
m=l n:kn>m n
<W e if;Var[Fm]4
2 *—' m2
7Tl=l
oo 1
m=l
In (1) we used that with kn = [ctn] one has Yln:kn>m- 1 ^n

<r r» .-TK.-2
<C-m
2.9. The strong law of large numbers 67
Lets take some breath and continue, where we have just left off:
oo oo ^
£P[|rfcB-E[7U>e] < CY,-^V{Yl]

Ti l
n=l r am== l
o oo -1 mm—- 1 »
„/+i
oo oo 1 „/+!
i=0m=(+l -71
/=0 m=Z+l m 7<

oo /./+1
<(2) C$^/ Xd/i(x)
Z=0^
< C • E[Xi] < oo .
In (2) we used that £m=z+i m~2 < C • (I + l)"1.
We have now proved complete (=fast stochastic) convergence. This implies

the almost everywhere convergence of Tkn — E[Tkn] —▶ 0.
(iii) So far, the convergence has only be verified along a subsequence kn-
Because we assumed Xn > 0, the sequence Un = Y^=i Yn = nTn is mono-
tonically increasing. For k G [km, fcm+i]> we get therefore
km Ukm _ Ukrn <Un < Uhrn+i_ _ fcm+l ^fcm+1
fcm-\-l Km Km+1 ^ ^m ^m ^ra+1
and from limn-^ Tn = E[Xi] almost everywhere,
-E[Xi] < lim n

inf Tn < lim nsup Tn < aE[Xi]
a
follows. □
Remark. The strong law of large numbers can be interpreted as a statement
about the growth of the sequence X]fc=i Xn. For E[Xi] = 0, the convergence
n Sfc=i Xn ~* 0 means that for all e > 0 there exists m such that for n > m
\J2x"\ên-
k=l
This means that the trajectory Ylk=i Xn ls nnahy contained in any arbi
trary small cone. In other words, it grows slower than linear. The exact
description for the growth of XX=i Xn *s given by the law of the iterated
logarithm of Khinchin which says that a sequence of IID random variables
Xn with E[Xn] = m and cr(Xn) = o i1 0 satisfies
lim sup —- = +1, lim inf —— = — 1 ,

n->oo An n-*oo An
with An = ^2cr2nloglogn.
Remark. The IID assumption on the random variables can not be weakened
without further restrictions. Take for example a sequence Xn of random
variables satisfying P[Xn = ±2n] = 1/2. Then E[Xn] = 0 but even Sn/n
does not converge.
Exercice. Let Xi be IID random variables in C2. Define Yk = \ Ylt=i xi-

What can you say about Sn = £ Ylk=i ^fc?
2.10 Birkhoff's ergodic theorem

In this section we fix a probability space (Cl,A,P) and consider sequences
of random variables Xn which are defined dynamically by a map T on Cl
by
XnM=X(T»),
where Tn(u) = T(T(.. .T(u))) is the n'th iterate of uj. This can include
as a special case the situation that the random variables are independent,
but it can be much more general. Similarly as martingale theory covered
later in these notes, ergodic theory is not only a generalization of classical
probability theory, it is a considerable extension of it, both by language as
by scope.
Definition. A measurable map T : Cl —> Cl from the probability space onto

itself is called measure preserving, if P[T_1(A)] = P[A] for all A £ A. The
map T is called ergodic if T(A) = A implies P[A] = 0 or P[A] = 1. A
measure preserving map T is called invertible, if there exists a measurable,
measure preserving inverse T_1 of T. An invertible an measure preserving
map T is also called an automorphism of the probability space.
Example. Let fi = {|z| = l}cCbe the unit circle in the complex plane
with the measure P[Arg(z) G [a, b]] = (b - a)/(2n) for 0 < a < b < 2n
and the Borel a-algebra A. If w = e2nta is a complex number of length 1,
then the rotation T(z) = wz defines a measure preserving transformation
on (Cl,B, P). It is invertible with inverse T~1(z) = z/w.
Example. The transformation T(z) = z2 on the same probability space as

in the previous example is also measure preserving. Note that P[T(A)] =
2P[A| but P[T_1(A)] = P[A] for all A G B. The map is measure preserving
but it is not invertible.
Remark. T is ergodic if and only if for any X G C2 the condition X(T) = X

implies that X is constant almost everywhere.
2.10. Birkhoff's ergodic theorem 69
Example. The rotation on the circle is ergodic if a is irrational Proof:

with z = e2nix one can write a random variable X on Cl as a Fourier series
/(*) = £~=-oo a^n which is the sum fo+U+f-, where /+ = £~=1 anZn
is analytic in \z\ < 1 and /_ = £^°=1 anz~n is analytic in \z\ > 1 and f0 is
constant. By doing the same decomposition for f(T(z)) = ]C^L-oo anWnzn,
we see that /+ = £~=1 anzn = £~=1 «n^^. But these are the Taylor
expansions of /+ = /+(T) and so an = anwn. Because wn ^ 1 for irrational
a, we deduce an = 0 for n > 1. Similarly, one derives an = 0 for n < -1.
Therefore /(s) = a0 is constant.
Example. Also the non-invertible squaring transformation T(x) = x2 on

the circle is ergodic as a Fourier argument shows again: T preserves again
the decomposition of / into three analytic functions f = f- + fo + f+
SO that f(T(z)) = EZ-oc *nZ2n = Y.n=-oo*nZn implies Y,n=l"nZ2n =
Yln=i anzn- Comparing Taylor coefficients of this identity for analytic func
tions shows an = 0 for odd n because the left hand side has zero Taylor
coefficients for odd powers of z. But because for even n = 2lk with odd
k, we have an = a2ik = a2i-ik = • • • = ak = 0, all coefficients ak = 0 for
k > 1. Similarly, one sees ak = 0 for k < — 1.
Definition. Given a random variable X G C, one obtains a sequence of

random variables Xn = X(Tn) G C by X(Tn)M = X(Tno;). Define S0 = 0
andSn = £Lo*(T/c)-
Theorem 2.10.1 (Maximal ergodic theorem of Hopf). Given X G £ , the

event A = {supn Sn > 0 } satisfies
E[X;A] = E[1AX]>0.
Proo/. Define Zn = max0<k<nSk and the sets An = {Zn > 0} C An+i.

Then A = |Jn ^n- Clearly Zn G £x. For 0 < k < n, we have Zn > Sk and
so Zn(T) > Sk(T) and hence
Zn(T) + X > Sk+i .
By taking the maxima on both sides over 0 < k < n, we get
Zn(T) + X> max Sk .

~ l<fc<n+l
On An = {Zn > 0}, we can extend this to Zn(T) + X > max:i<fc<n+i Sk >
max0<jk<n-i-i Sk = Zn+\ > Zn so that on An
X>Zn- Zn(T) .
Integration over the set An gives
E[X; An] > E\Zn\ An] - E[Zn(T); An] .
Using (1) this inequality, the fact (2) that Zn = 0 on X \ An, the (3) in
equality Zn(T) > Sn(T) > 0 on An and finally that T is measure preserving
(4), leads to
E[X;An] >(1) E[Zn,An]-E[Zn(T)',An]

=(2) E[Zn]-E[Zn(T);An]
>(3) E[Zn - Zn(T)] =(4) 0
for every n and so to E[X; A] > 0. D
Theorem 2.10.2 (Ergodic theorem of Birkhoff, 1931). For any X G £r the

time average
<? i n_1
n n ^^
2=0
converges almost everywhere to a T-invariant random variable X satisfying

E[X] = E[X]. Especially, if T is ergodic, then Sn/n converges to E[X].
Proof. Define X = limsup^^ Sn, X = lim inf n_+oo Sn . We get X =

X(T) and X = X(T) because
on — Sn(T) = —
n n .
(i)X = X.
Define for a < f3 G R the sets Aa>/3 = {X < (3, a < X}. Because {X <
X} = Ua</3,a,/36Qâ^' lt is enougn to show that P[i4a,/3] = 0 for rational
a < p. Define
A = {sup(5n - na) > 0 } = {sup(5n - a) > 0 } .

n n
Because j4q>0 C .A and Aa^ is T-invariant, we get from the maximal ergodic
theorem E[X — a, Aa^] > 0 and so
E[X,A^]>a^P[Aa^].
Replacing X, a, (3 with -X, -/?, -a and using -X = -X, -X = -X gives

E[X; Aa^] < (3 • P[Aa,/3] and because (3 < a, the claim PfA^] = 0 follows.
(u)XG/:1. _ _
\Sn\ < \X\, and Sn converges point wise to X = X and X G Cl. Lebesgue's
2.10. BirkhofFs ergodic theorem 71
dominated convergence theorem gives X G C.
(iii)E[X] = E[X]. _
Define the sets Bk,n = {X G [£, *±I)} for fc G Z,n > 1. Define for e > 0,
y = X - - + e. Using the maximal ergodic theorem, we get E[Y; Bk,n] > 0.
Because e > 0 was arbitrary,
E [ X ; £ f c , n ] >n- .
With this inequality
E[X,Bk,n] < —P[Bfc,„]

n < ip[Bfc,„]
n +E[X;Bk,n] .
Summing over k gives
E[X] < - + E[X]
and because n was arbitrary, E[X] < E[X]. Doing the same with -X and
using (i), we end with
E[-X] = EbX] < E[^X] < E[-X] .
Corollary 2.10.3. The strong law of large numbers holds for IID random
variables Xn G C1.
Proof. Given a sequence of IID random variables Xn G C1. Let p be the

law of Xn. Define the probability space Cl = (RZ,.4,P), where P = pz is
the product measure. If T : Cl -» Cl, T(u))n = ^n+i denotes the shift on Cl,
then Xn = X(Tn) with with X(u>) = u0. Since every T-invariant function
is constant almost everywhere, we must have X = E[X] almost everywhere,
s o t h a t S n / n —▶ E [ X ] a l m o s t e v e r y w h e r e . □
Remark. While ergodic theory is closely related to probability theory, the

notation in the two fields is different. The reason is that the origin of
the theories are different. One usually writes (X,A,m) for a probability
space. An example of different language is also that ergodic theorists do
not use the word "random variables" X but speak of "functions" /. Good
introductions to ergodic theory are [36, 13, 8, 77, 54, 107].
2.11 More convergence results
We mention now some results about the almost everywhere convergence of
sums of random variables in contrast to the weak and strong laws which
were dealing with averaged sums.
Theorem 2.11.1 (Kolmogorov's inequalities), a) Assume Xk G C2 are inde

pendent random variables. Then
P[ sup \Sk - E[Sk]\ >e]< \var[Sn]

Kk<n
b) Assume Xk G C°° are independent random variables and HXJU < R.

Then
P[sup \Sk-E[Sk]\>e]>l- J;Rle)*
i<k<n J2k=1 Va r [ X ; f c ]
Proof. We can assume E[Xk] = 0 without loss of generality,

a) For 1 < k < n we have
Si ~ S2 = (Sn - Sk)2 + 2(Sn - Sk)Sk = 2(Sn - Sk)Sk
and therefore E[S2;Ak] > E[Sl,Ak] for all Ak G a(Xu...,Xk) by the

independence of Sn - Sk and Sk. The sets Ax = {|5i| > e}, Ak+X =
{\Sk+i\ > e,maxi<i<k \Si\ < e} are mutually disjoint. We have to estimate
the probability of the events
Bn = {max \Sk\>e}= M Ak .
Kk<n v^
fc=l
We get
E[S>] > E[Sl;Bn] = Y,E[Sl; Ak] > Ê[5fc2; A*] > €2^P[^fe] = 62P[Bn]
fe=i k=i fc=i
b)
E[S2k; Bn] = E[Sl] - E[Sl Bcn] > E[5fc2] - 62(1 - P[Bn}) .
On Ak, \Sk-i\ < e and \Sk\ < |5fe_i| + \Xk\ <e + R holds. We use that in
2 . 11 . More convergence results 73
the estimate
E[52;Bn] = f>[S! + (Sn-Sfc)2;Afc]

fc=l
= Ê[S2;Afc] + Ê[(5n-5fe)2;Afc]
fc=l fe=l
< (i? + 6)2^P[A,]+^P[Afc] £ Var[X;]

jb=i fc=i j=fc+i
< P[5n]((6 + i?)2 + E[52])
so that
E[52] < P[Bn]((e + i?)2 + E[52]) + e2 - 62P[£n] .
and so
rPlBnJ
[ c ] -, (c +e i?)2
[ s 2+ ]E[Sn]
- 6 2- 6^ >" (6
x + R)2
( ^ ^+ )£[S2]
2 - >62i _- iE[52]
i±^
Remark. The inequalities remain true in the limit n -> oo. The first in
equality is then
1 °°
P[sup \Sk - E[Sk]\ > e] < -j £ Var[Xfc] .
fc e fe=i
Of course, the statement in a) is void, if the right hand side is infinite. In

this case, however, the inequality in b) states that supfc \Sk — E[Sk]\ > e
almost surely for every e > 0.
Remark. For n = 1, Kolmogorov's inequality reduces to Chebychev's in

equality (2.5.5)
Lemma 2.11.2. A sequence Xn of random variables converges almost ev

erywhere, if and only if
lim P[sup |Xn+fc - Xn\ > e] = 0
n^°° fc>i
for all e > 0.
Proof This is an exercise. D

74
Theorem 2.11.3 (Kolmogorov). Assume Xn e C2 are independent and

Er=i Var[Xn] < oo. Then
f^(Xn-E[Xn})
n=l
converges almost everywhere.
Proof. Define Yn = Xn - E[Xn] and Sn = £Li Yk. Given m G N. Apply

Kolmogorov's inequality to the sequence Ym+k to get
i °°
P[sup|5n-5m|>6]<4
n>m €
Y E[X2]^0
*—**
fc=77l+l
for m -> oo. The above lemma implies that Sn(u;) converges. D
Figure. We sum up indepen

dent random variables Xk
which take values ^ with
equal probability. According to
theorem (2.11.3), the process
Sn = X>fc-E[Xfe]) = £x,
k=l k=i
converges if
oo oo 1
k=l k=l
converges. This is the case ifa>

1/2. The picture shows some ex
periments in the case a = 0.6.
The following theorem gives a necessary and sufficient condition that a

sum Sn = $X=i Xk converges for a sequence Xn of independent random
variables.
Definition. Given R e R and a random variable X, we define the bounded
random variable
X{R) = 1\x\<rX .
Theorem 2.11.4 (Three series theorem). Assume Xn G C be independent.

Then £^°=1 Xn converges almost everywhere if and only if for some R > 0
all of the following three series converge:
oo
< 00 , (2.4)
Y,n\Xk\>R}
fc=i
oo
£|E[4R)]l < 00 , (2.5)

fc=i
oo
< oo . (2.6)
EVar[4ft)]
fc=l
Proof. "=>" Assume first that the three series all converge. By (3) and
Kolmogorov's theorem, we know that YlkLi xiR) ~ ElxiR)] converges al
most surely. Therefore, by (2), YlkLi XkR) converges almost surely. By
(1) and Borel-Cantelli, P[Xfc ^ X(kR) infinitely often) = 0. Since for al
most all u, xlR\v) = Xk(uo) for sufficiently large k and for almost all
uj, YlkLi Xk^fa) converges, we get a set of measure one, where Ylh=i X^
converges.
"<^=" Assume now that X^Li xn converges almost everywhere. Then Xk —▶
0 almost everywhere and P[|Xfc| > R, infinitely often) = 0 for every R > 0.
By the second Borel-Cantelli lemma, the sum (1) converges.
The almost sure convergence of Y^=i xn implies the almost sure conver
gence of £~=1 XnR) since P[|Xfc| > R, infinitely often) = 0.
Let R > 0 be fixed. Let Yk be a sequence of independent random vari
ables such that Yk and X^ have the same distribution and that all the
random variables X^,Yk are independent. The almost sure convergence
of £~ i XnR) implies that of ZZi X™ " ^- Since Et4*} " y*l = °
and P[|X^} - Yk\ < 2R) = 1, by Kolmogorov inequality b), the series
Tn = ELi XiR) ~ Y* satisfies for all e > 0
P[sup|Tn+fc -Tn\ > 6] > 1 - [R^\r) v1 '

*>* E f c = n Va r K ~Ykl
Claim: £~ x Varfxf} - Yk] < oo.

Assume, the sum is infinite. Then the above inequality gives P[supfc> \Tn+k-
Tn | > e] = 1. But this contradicts the almost sure convergence of YlkLi Xk "
Yk because the latter frnplies by Kolmogorov inequality that Pûp^ |Sn+fc-
5n| > e] < 1/2 for large enough n. Having shown that £Jî(Var[x£ -
Yk)] < oo, we are done because then by Kolmogorov's theorem (2.11.3),
the sum £21i XkR) ~ ElXkR)] converges, so that (2) holds.
□
Figure. A special case of the

three series theorem is when Xk
are uniformly bounded Xk <
R and have zero expectation
E[Xfc] = 0. In that case, almost
everywhere convergence of Sn =
J2k=i xk is equivalent to the
convergence of J2kLi Var[Xfc].
For example, in the case
X i = {A >
l ka
and a = 1/2, we do not have

almost everywhere convergence
of Sn, because J2T=i Var[Xfc] =
E o fc=l
o ki = oo.
Definition. A real number a G R is called a median of X G C if P[X <

a] > 1/2 and P[X > a] > 1/2. We denote by med(X) the set of medians
ofX.
Remark. The median is not unique and in general different from the mean.
It is also defined for random variables for which the mean does not exist.
Proposition 2.11.5. (Comparing median and mean) For Y G C2. Then every
a G med(y) satisfies
\a-E[Y]\<V2a[Y].
Proof. For every (3 eR, one has

\a- 8\2
2 < \<* ~ /?|2min(P[y > a],P[Y < a]) < E[(Y - (3)2] .
Nowput/3 = E[y]. D
Theorem 2.11.6 (Levy). Given a sequence Xn G C which is independent.

Choose aiik G med(5/ - Sk). Then, for all n G N and all e > 0
P[ Kmax
k < n\Sn -r an,k\ > e] < 2P[5n > e]
J
Proof. Fix n G N and e > 0. The sets
Ax = {Si + an,i > e}, Afc+i = { max (Sn + an,i) < 6, Sfc+i + an,fc+i > e}
for 1 < k < n are disjoint and |Jfc=i Ak = {maxi<k<n(Sk + an,k) > e}.
Because {Sn > e } contains all the sets Ak as well as {Sn - Sk > an,k} for
1 < k < n, we have using the independence of o(Ak) and a(Sn - Sk)
P[Sn>e] > J2P[{Sn-Sk>an,k}nAk]

n
= ^P[{5n-5fe>an,fe}]P[Afc]
fe=i
zfc=i
= 5plU*
Z fc=l
= -P[ max (5n + an>fc) > e] .

2 l<k<n
Applying this inequality to — Xn, we get also P[—Sn — ctn,m > —6] >
2P[-Sn > -e] and so
P[ l</c<n
max |5„ + an,fc| > e] < 2P[Sn > 6] .
Corollary 2.11.7. (Levy) Given a sequence Xn G C of independent random

variables. If the partial sums Sn converge in probability to 5, then Sn
converges almost everywhere to S.
Proof Take a^k G med(Sj - Sk). Since Sn converges in probability, there

exists mi G N such that \cxi,k\ < e/2 for all m\ < k < I. In addition,
there exists m2 G N such that supn>1 P[|Sn+m - 5m| > e/2] < e/2 for all
m > 777,2. For m — max{mi,7722}, we have for n > 1
P[ l max
< l < n|S/+m - Sm\ > e] < P[l <
max
l < n\Si+m - Sm + an+m,z+m| > e/2] .
The right hand side can be estimated by theorem (2.11.6) applied to Xn+m
with
<2P[|Sn+m-Sm|>^]<e.
N o w a p p l y t h e c o n v e r g e n c e l e m m a ( 2 . 11 . 2 ) . □
Exercice. Prove the strong law of large numbers of independent but not
necessarily identically distributed random variables: Given a sequence of
independent random variables Xn G C2 satisfying E[Xn] = m. If
f>ar[Xfc]/fc2 < co ,
fc=i
then Sn/n —▶ m almost everywhere.

Hint: Use Kolmogorov's theorem for Yk = Xk/k.
Exercice. Let Xn be an IID sequence of random variables with uniform

distribution on [0,1]. Prove that almost surely
En^<°°
71=1 2=1
Hint: Use Var^ Xi] = U E[X2] - U E[X{]2 and use the three series theo
rem.
2.12 Classes of random variables

The probability distribution function Fx : R —> [0,1] of a random variable
X was defined as
Fx(x)=P[X<x],
where P[X < x] is a short hand notation for P[{uj G Cl | X(u) < x }. With
the law p,x = X*P of X on R has Fx(x) = f*^ dp(x) so that F is the
anti-derivative of \x. One reason to introduce distribution functions is that
one can replace integrals on the probability space Cl by integrals on the real
line R which is more convenient.
Remark. The distribution function Fx determines the law px because the

measure u((—oo,a]) = Fx(a) on the 7r-system 1 given by the intervals
{(—oo, a]} determines a unique measure on R. Of course, the distribution
function does not determine the random variable itself. There are many
different random variables defined on different probability spaces, which
have the same distribution.
2.12. Classes of random variables 79
Proposition 2.12.1. The distribution function Fx of a random variable is
a) non-decreasing,
b)Fx(-oo) = 0,Fx(oo) = l
c) continuous from the right: Fx(x + h) = Fx-
Furthermore, given a function F with the properties a),b),c), there exists

a random variable X on the probability space (Cl,A,P) which satisfies
F X = F.
Proof, a) follows from {X < x } C {X < y } for x < y. b) P[{X < -n}] ->
0 and P[{X < n}] -> 1. c) Fx(x + h) - Fx = P[x < X < x + h] -> 0 for
h-*0.
Given F, define Cl = R and .4 as the Borel cr-algebra on R. The measure
P[(—oo, a]] = F[a] on the 7r-system 1 defines a unique measure on (Cl,A).
□
Remark. Every Borel probability measure \x on R determines a distribution

function Fx of some random variable X by
dp(x) = F(x) .
J —c
The proposition tells also that one can define a class of distribution func
tions, the set of real functions F which satisfy properties a), b), c).
Example. Bertrands paradox mentioned in the introduction shows that the

choice of the distribution functions is important. In any of the three cases,
there is a distribution function f(x,y) which is radially symmetric. The
constant distribution f(x, y) = 1/n is obtained when we throw the center of
the line into the disc. The disc Ar of radius r has probability P[Ar] = r2/tt.
The density in the r direction is 2r/n. The distribution f(x, y) = \/r =
1/^/x2 + y2 is obtained when throwing parallel lines. This will put more
weight to center. The probability P[Ar] = r/ir is bigger than the area of
the disc. The radial density is 1/tt. f(x,y) is the distribution when we
rotate the line around a point on the boundary. The disc Ar of radius r
has probability arcsin(r). The density in the r direction is l/\/l — r2.
Figure. A plot of the radial Figure. A plot of the radial dis

density function f(r) for the tribution function F(r) — P[Ar]
three different interpretation of There are different values at
the Bertrand paradox. F(l/2).
So, what happens, if we really do an experiment and throw randomly lines

onto a disc? The punch line of the story is that the outcome of the ex
periment very much depends on how the experiment will be performed. If
we would do the experiment by hand, we would probably try to throw the
center of the stick into the middle of the disc. Since we would aim to the
center, the distribution would be different from any of the three solutions
given in Bertrand's paradox.
Definition. A distribution function F is called absolutely continuous (ac), if

there exists a Borel measurable function / satisfying F(x) = f*^ f(x) dx.
One calls a random variable with an absolutely continuous distribution
function a continuous random variable.
Definition. A distribution function is called pure point (pp) or atomic if

there exists a countable sequence of real numbers xn and a sequence of
positive numbers pn,SnPn = 1 such that F(x) = J2n,xn<xPn' ®ne ca^s
a random variable with a discrete distribution function a discrete random
variable.
Definition. A distribution function F is called singular continuous (sc) if F

is continuous and if there exists a Borel set S of zero Lebesgue measure such
that Pf(S) = 1. One calls a random variable with a singular continuous
distribution function a singular continuous random variable.
Remark. The definition of (ac),(pp) and (sc) distribution functions is com

patible for the definition of (ac),(pp) and (sc) Borel measures on R. A Borel
measure is (pp), if p(A) = J2xeA MW)- Jt is continuous, if it contains no
atoms, points with positive measure. It is (ac), if there exists a measurable
function / such that p = / dx. It is (sc), if it is continuous and if p(S) = 1
for some Borel set S of zero Lebesgue measure.
The following decomposition theorem shows that these three classes are
natural:
Theorem 2.12.2 (Lebesgue decomposition theorem). Every Borel measure

p on (R, B) can be decomposed in a unique way as p = ppp + pac + Use,
where ppp is pure point, psc is singular continuous and pac is absolutely
continuous with respect to the Lebesgue measure A.
Proof. Denote by A the Lebesgue measure on (R,B) for which X([a, b]) =
b — a. We first show that any measure p can be decomposed as p = pac+V<s,
where pac is absolutely continuous with respect to A and ps is singular. The
decomposition is unique: p — p>a} + Ps — pac + ps implies that p^c -
p(2J = p)?> —ps2' is both absolutely continuous and singular continuous with
respect to p which is only possible, if they are zero. To get the existence
of the decomposition, define c = supAG^ \(A)=oM^)- If c = 0, then p is
absolutely continuous and we are done. If c > 0, take an increasing sequence
An G B with p(An) -» c. Define A = Un>i ^n and pac as pac(B) =
p(AilB). To split the singular part ps into a singular continuous and pure
(1) (2) (2) (2)
point part, we again have uniqueness because ps = psc + psc = Ppp + pPP
implies that v = plV — Psc = ppp — ppp are both singular continuous and
pure point which implies that v = 0. To get existence, define the finite or
countable set A = {u | p(u) > 0 } and define ppp(B) = p(A fl B). □
Definition. The Gamma function is defined for x > 0 as

/•OO
T(x) = / tx-le-1 dt .
Jo
It satisfies T(n) = (n - 1)! for n G N. Define also
B(p,q)= [ xp~l(l-x)q-1 dx ,
Jo
the Beta function.
Here are some examples of absolutely continuous distributions:
acl) The normal distribution N(m, a2) on Cl = R has the probability den
sity function
1 (x-m)2
ac2) The Cauchy distribution on Cl = R has the probability density function
t( \ l b
7r b2 + (x — m)2
ac3) The uniform distribution on Cl = [a, b] has the probability density

function
b—a
ac4) The exponential distribution A > 0 on Cl = [0, oo) has the probability
density function
f(x) = \e~Xx .
ac5) The log normal distribution on Cl = [0, oo) has the density function
f(\— 1 (log(x)-m)2
y/2irx2a2
ac6) The beta distribution on Cl = [0,1] with p > 1, q > 1 has the density
fix)
H )~ - Bgâ-*)'-1
(p,q) •
ac7) The Gamma distribution on Cl = [0, oo) with parameters a > 0, (3 > 0
m x*-lp*e-x/(3
T(a)
Figure. The probability density Figure. The probability density

and the CDF of the normal dis- and the CDF of the Cauchy dis
tribution, tribution.
83
2.12. Classes of random variables
Figure. The probability density Figure. The probability density

and the CDF of the uniform dis- and the CDF of the exponential
tribution. distribution.
Definition. We use the notation

n n!
k J (n-k)\k\
for the Binomial coefficient, where fc! = fc(fc-l)(fc-2) • • • 2-1 is the factorial
of k with the convention 0! = 1. For example,
10 10! = 10 * 9 * 8/6 = 120 .
7!3!
Examples of discrete distributions:
ppl) The binomial distribution on Cl = {1,..., n }
p[* = fc] = (fc)pfc(i-p)n~fc
pp2) The Poisson distribution on ft = N
p[X = k] = e~ k\
pp3) The Discrete uniform distribution on ft = {1,.., n }
P[X = k] = n-
pp4) The geometric distribution on ft = N = {0,1,2,3,... }
P[X = k}=p(l-p)k
pp5) The distribution of first success on ft = N \ {0} = {1,2,3,... }
F[X = k] = p(l - p)*-1

Figure. The probabilities and the Figure. The probabilities and the
CDF of the binomial distribution. CDF of the Poisson distribution.
Figure. The probabilities and the Figure. The probabilities and the
CDF of the uniform distribution. CDF of the geometric distribution.
An example of a singular continuous distribution:
scl) The Cantor distribution. Let C = f)™=0 En be the Cantor set,

where E0 = [0,1], E1 = [0,1/3] U [2/3,1] and En is inductively
obtained by cutting away the middle third of each interval in
En-i. Define
F[x) = n—too
lim Fn{x)
where Fn (x) has the density (3/2)n • lEn. One can realize a random
variable with the Cantor distribution as a sum of IID random
variables as follows:
oo v
x = y^,
Figure. The CDF of the Cantor

distribution is continuous but not
absolutely continuous. The func
tion Fx(x) is in this case called
the Cantor function. Its graph is
also called a Devils staircase
Lemma 2.12.3. Given X e £ with law p. For any measurable map h : R1 —>
[0,oo) for which h(X) e C1, one has E[/i(X)] = fRh(x) dp(x). Especially,
if p = pac = f dx then
E[h(X)] = f h(x)f(x) dx
Jr
If p — ppp, then
E[h(X)] = £ h(x)p({x}) .
x^({x})^0
Proof. If the function h is nonnegative, prove it first for X = clxeA, then

for step functions X e S and then by the monotone convergence theorem
for any X e C for which h(x) eCl.lf h(X) is integrable, then E[h(X)] =
E[h+(X)]-E[h-(X)]. □
Proposition 2.12.4.
Distribution Parameters Mean Variance
acl) Normal m e R, a2 > 0 m a*
ac2) Cauchy m e R, b > 0 "m" oo
ac3) Uniform a0 1/A 1/A2
ac5) Log-Normal m e R, a2 > 0 e/i+cra/2 (e^ - l)e2m+<T"
PQ
ac6) Beta p,q>0 p/(p + q) (p+q)Hp+<,+D
ac7) Gamma a,/3>0 a(3 a/F
Proposition 2.12.5.
ppl) Bernoulli n € N, p e [0,1] np np{\ - p)
pp2) Poisson A>0 A A
pp3) Uniform neN (l+n)/2 (n2 - 1)/12
pp4) Geometric pG (0,1) (l-p)/p (l-pj/p*
pp5) First Success p e (o, l) 1/p (i-pVp*
scl) Cantor - 1/2 1/8
Proof These are direct computations, which we do in some of the examples:

Exponential distribution:
E[Xp] = r xp\e~Xx dx = Ê^"1] = ^ .

Poisson distribution:
oo >^ oo \k—1
fc=0 k=\ v y
For calculating higher moments, one can also use the probability generating
function
E[**] = E _-*(**)'
k\
-A(l-z)
k=0
and then differentiate this identity with respect to z at the place z = 0. We

get then
E[X] = X,E[X(X - 1)] = A2,E[X3] = E[X(X - 1)(X - 2)],...
so that E[X2] = A + A2 and Var[X] - A.

Geometric distribution. Differentiating the identity for the geometric series
E**
fe=0
1 - 3
gives W 1
k=0 {1-xf
Therefore
oo oo oo 1
E[xp] = Efc(i-P)fcp=Efc(1-p)fep=pEfc(1-p)fc = S = p-
k=0 k=0 fc=l
For calculating the higher moments one can proceed as in the Poisson case
or use the moment generating function.
Cantor distribution: because one can realize a random variable with the
Cantor distribution asX = X^Li -W3n, where the IID random variables
Xn take the values 0 and 2 with probability p = 1/2 each, we have
M * J -n2=^l - 3 ^ - - 2n ^= l3 n - -i _1i //' 33 x - 2

and
Va r [ X ] = ^ — ^ — - ^ — g s — - 2 ^ 9 ^ ~ i _ 1 / 9 ^ 8 l ~ 8 '
n=l n=l n=l
See also corollary (3.1.6) for an other computation. □
Computations can sometimes be done in an elegant way using character

istic functions 4>x(t) = E[eitX] or moment generating functions Mx(t) =
E[etx]. With the moment generating function one can get the moments
with the moment formula
E[Xn}= j xndix = ^f{t)\t=a.
For the characteristic function one obtains
E[Xn}= [xndp = (-i)n^(t)\t=o.
Example. The random variable X(x) = x has the uniform distribution

on [0,1]. Its moment generating function is Mx(t) = JQ etx dx = (e* -
l)/t = 1 + tj2\ + £2/3! + A comparison of coefficients gives the moments
E[Xm] = l/(ra + 1), which agrees with the moment formula.
Example. A random variable X which has the Normal distribution N(m, a)

has the moment generating function Mx(t) = etrn+°2t*"I2. All the moments
can be obtained with the moment formula. For example, E[X] = M'x(0) =
m,E[X2]=M,J[(0) = m2 + a2.
Example. For a Poisson distributed random variable X on fi = N =

{0,1,2,3,... } with P[X = k] = e_A^r, the moment generating function is
oo
Mx(t) = J2F^X = k^k = eMl~et) '

k=0
Example. A random variable X on Q = N = {0,1,2,3,... } with the

geometric distribution P[X = k] = p(l - p)k has the moment generating
function
oo oo
Mx(t) = ê^(l-p)l=p^((l-P)e? = _ P_ .
k=0 k=0 V F)
A random variable X on f2 = {1,2,3,...} with the distribution of first
success P[X = k]= p(l - p)k~x, has the moment generating function
oo OO £
Mx(t) =£ r Y,ektp(l-p)k~1
i r-i l
=- Sp^-py)'
(1 -
= , ,fp )we t •
k=l k=0
Exercice. Compute the mean and variance of the Erlang distribution

\k+k—1
on the positive real line ft = [0, oo) with the help of the moment generating
function. If k is allowed to be an arbitrary positive real number, then the
Erlang distribution is called the Gamma distribution.
Lemma 2.12.6. If X, Y are independent random variables, then their mo

ment generating functions satisfy
Mx+Y(t) = Mx(t) - MY(t) .
Proof If X and Y are independent, then also etx and etY are independent.
Therefore,
E[e*(*+y>] = E[etxetY] = E[etx]E[ety] = Mx(t) • MY(t) .

D
Example. The lemma can be used to compute the moment generating

function of the binomial distribution. A random variable X with bino
mial distribution can be written as a sum of IID random variables Xi
taking values 0 and 1 with probability 1 - p and p. Because for n = 1,
we have MXi(t) = (1 - p) + pef, the moment generating function of X
is Mx(t) = [(1 - p) +pet]n. The moment formula allows us to compute
moments E[Xn] and central moments E[(X - E[X})n] of X. Examples:
E[X] =
np
E[X2} =
np(l-p + np)
Var[X] =
E[(X-E[X])2} = E[X2}-E[X}2 = np(l-p)
E[X3] =
np{l + 3(n-l)p+(2-3n + n2)p2)
E[X4] =np(l + 7(n-l)p+6(2-3n
+n2)p2 + (-6 + lln - 6n2 + n3)p3)
E[(X-E[X})4} = E[X4]-8E[X]E[X3] + 6E[X2}2 + E[X}4
= np(l - p)(l + (5n - 6)p - (-6 + n + 6n2)p2)
Example. The sum X + Y of a Poisson distributed random variable X with

parameter A and a Poisson distributed random variable Y with parameter
p is Poisson distributed with parameter \ + p as can be seen by multiplying
their moment generating functions.
Definition. An interesting quantity for a random variable with a continuous

distribution with probability density fx is the Shannon entropy or simply
entropy
H(X) = - f f(x)\og(f(x))dx.
JR
Without restricting the class of functions, H(X) is allowed to be -oo or

oo. The entropy allows to distinguish several distributions from others by
asking for the distribution with the largest entropy. For example, among all
distribution functions on the positive real line [0, oo) with fixed expectation
m = 1/A, the exponential distribution Ae_A is the one with maximal en
tropy. We will return to these interesting entropy extremization questions
later.
Example. Let us compute the entropy of the random variable X(x) = xm

on ([0, l],B,dx). We have seen earlier that the density of X is fx(x) =
xl/rn-l/m go that
H(X) = - [ (x1/m-1/m)log(x^m-1/m) dx .
Jo
To compute this integral, note first that f(x) = xa log(a:a) = axa log(x) has
the antiderivative ax1+a((l+a) log(x)-l)/(l+a)2 so that J* xa log(xa) dx =
-a/(l + a2) andH(X) = (l-ra + log(ra)). Because £lH(Xrn) = (l/m)-l
and -^2H(Xm) = -1/ra2, the entropy has its maximum at m = 1, where
the density is uniform. The entropy decreases for m —▶ oo. Among all ran
dom variables X(x) = xm, the random variable X(x) = x has maximal
entropy.
Figure. The entropy of the ran

dom variables X(x) = xm on
[0,1] as a function of m. The
maximum is attained for m = 1,
which is the uniform distribution
2.13 Weak convergence
Definition. Denote by Cb(R) the vector space of bounded continuous func
tions on R. This means that H/H^ = £X(_R \f(x)\ < oo for every / e Cb(R).
A sequence of Borel probability measures pn on R converges weakly to a
probability measure p on R if for every / e Cb(R) one has
/ / dpn -» / / dp
JR Jx
in the limit n —▶ oo.
Remark. For weak convergence, it is enough to test /R / dpn -> jx f dp

for a dense set in Cb(R). This dense set can consist of the space P(R) of
polynomials or the space C6°°(R) of bounded, smooth functions.
An important fact is that a sequence of random variables Xn converges
in distribution to X if and only if E[h(Xn)] -> E[h(X)] for all smooth
functions h on the real line. This will be used the proof of the central limit
theorem.
Weak convergence defines a topology on the set MX(R) of all Borel proba
bility measures on R. Similarly, one has a topology for Mi ([a, &]).
Lemma 2.13.1. The set MX(I) of all probability measures on an interval

/ = [a, b] is a compact topological space.
Proof We need to show that any sequence pn of probability measures on

J has an accumulation point. The set of functions fk(x) = xk on [a, b] span
all polynomials and so a dense set in Cb([a, b]). The sequence pn converges
to p if and only if all the moments Ja xk dpn converge for n —▶ oo and for
all k e N. In other words, the compactness of Mi ([a, b]) is equivalent to the
compactness of the product space IN with the product topology, which is
Tychonovs theorem. □
Remark. In functional analysis, a more general theorem called Banach-

Alaoglu theorem is known: a closed and bounded set in the dual space X*
of a Banach space X is compact with respect to the weak-* topology, where
the functionals pn converge to p if and only if pn(f) converges to p(f) for
all / e X. In the present case, X = Cb[a,b] and the dual space X* is the
space of all signed measures on [a, b] (see [7]).
Remark. The compactness of probability measures can also be seen by
looking at the distribution functions F^(s) = p((-oo, s]). Given a sequence
Fn of monotonically increasing functions, there is a subsequence Fnk which
converges to an other monotonically increasing function F, which is again
2.13. Weak convergence 91
a distribution function. This fact generalizes to distribution functions on
the line where the limiting function F is still a right-continuous and non-
decreasing function Helly's selection theorem but the function F does not
need to be a distribution function any more, if the interval [a, b] is replaced
by the real line R.
Definition. A sequence of random variables Xn converges weakly or in law
to a random variable X, if the laws pxn of Xn converge weakly to the law
px of X.
Definition. Given a distribution function F, we denote by Cont(F) the set

of continuity points of F.
Remark. Because F is nondecreasing and takes values in [0,1], the only

possible discontinuity is a jump discontinuity. They happen at points U,
where a^ = p({U}) > 0. There can be only countably many such disconti
nuities, because for every rational number p/q > 0, there are only finitely
many a* with a* > p/q because £V a* < 1.
Definition. We say that a sequence of random variables Xn converges in
distribution to a random variable X, if Fxn(x) —▶ Fx(x) point wise for all
x e Cont(F).
Theorem 2.13.2 (Weak convergence = convergence in distribution). A se

quence Xn of random variables converges in law to a random variable X if
and only if Xn converges in distribution to X.
Proof, (i) Assume we have convergence in law. We want to show that we

have convergence in distribution. Given s G Cont(/) and 6 > 0. Define a
continuous function l(_oo,s] < / < l(_oo)S+(5]. Then
Fn(s) = / l(_oo,s] dpn< f dpn < / l(_oojfl+a] dpn = Fn(s + S) .

Jk Jr Jr
This gives
limsupFn(s) < lim / / dpn — If dp < F(x -f S) .

TWOO n-*°° J J
Similarly, we obtain with a function I^qo^.^j < / < l(_oo,s]
lim inf Fn (s)n —

> >lim / / dpn = IfJ dp > F(s — S) .
n—>oo oo J
Since F is continuous at x we have for 5 —> 0:
F(s) = lim F(s - 6) < lim inf Fn (s) <n limsupFn(s) < F(s) .
<5->0 n-+oo -+oo
That is we have established convergence in distribution,
(ii) Assume now we have no convergence in law. There exists then a con
tinuous function / so that // dpn to J f dp fails. That is, there is a
subsequence and e > 0 such that | / / dpnk - J f dp\ > e > 0. There exists
a compact interval J such that | /7 / dpnk - JT f dp\ > e/2 > 0 and we
can assume that pnk and p have support on /. The set of all probability
measures on J is compact in the weak topology. Therefore, a subsequence
of dpnk converges weakly to a measure v and \v(f) — p(f)\ > e/2. De
fine the 7r-system 1 of all intervals {(—oo,s] | s continuity point of F }.
We have pn((-oc,s]) = FXn(s) -> Fx(s) = p(-oo,s]). Using (i) we see
Pnk ((—oo, s]) —> v(—oo, s] also, so that p and v agree on the n system X. If
p and v agree on T, they agree on the 7r-system of all intervals {(—oo, s]}.
By lemma (2.1.4), we know that p — v on the Borel cr-algebra and so p = v.
This contradicts \v(f) - p(f)\ > e/2. So, the initial assumption of having
no convergence in law was wrong. □
2.14 The central limit theorem
Definition. For any random variable X with non-zero variance, we denote

by
X ~ ( X* -{EX[ )X } )
the normalized random variable, which has mean E[X*] = 0 and variance
a(X*) = i/Var[X*] = 1. Given a sequence of random variables Xk, we
again use the notation Sn = Ylk=i ^fc-
Theorem 2.14.1 (Central limit theorem for independent L3 random vari

ables). Assume Xi G C3 are independent and satisfy
1 n
M = sup \\Xi\\s < oo, 6 = liminf - V Var[Xi] > 0 .
i
1 n-->oo n ^z—
= l'
Then 5* converges in distribution to a random variable with standard

normal distribution N(0,1):
lim P[S; <x] = --L / e~y2/2 dy, \/x G R .

n^°° V27T J-oo
2.14. The central limit theorem 93
Figure. The probabil Figure. The probabil Figure. The probabil

ity density function ity density function ity density function
fs* of the random fs* of the random Fs* of the random
variable X(x) — x on variable X(x) = x on variable X(x) = x on
[-1,1]. [-i,i]. [-1,1].
Lemma 2.14.2. A JV(0,<72) distributed random variable X satisfies
E [ | X H = ^ = 2 " / Vr ( i ( p + 1 ) ) .
V71" z
Especially E[\X\3} = ^fa3.
Proof. With the density function f(x) = (2iro-2)-1/2e~^, we have E[\X\P]

2 J0°° xpf(x) dx which is after a substitution z = x2/(2a2) equal to
1 r°°
-±=2p/2ap / x^p^-le~xdx.
V* Jo
The integral to the right is by definition equal to r(|(p + 1)). □
After this preliminary computation, we turn to the proof of the central

limit theorem.
Proof. Define for fixed n > 1 the random variables
v..<*z|M,1<i<„
so that S* = Y%=i Yi- Define iV(0,a)-distributed random variables Yt hav
ing the property that the set of random variables
{Yl,...,Yn,Yl,...Yn}
are independent. The distribution of Sn = Yli=i YJ is just the normal distri
bution JV(0,1). In order to show the theorem, we have to prove E[/(5*)] -
'(l '0)aT uoi^nqu^stp reuiiou prapire^s uîav ajqBUBA uioprrei
b oq. Appra/A s9Sj8auoo "<? U9qq '[?x]J,BA > 0 ^jsiq.'Bs pros qjj qjb j 3 ?_y
JI •(saiqBiJBA.nioptrej ^7 qh joj raaaoam %\m\\ \bi%u3q) Z'VVZ majoaqx
:87 B ?X <* e7 9 -X
uoiqipuoo gqj x^jaj ubo 9m 'pgqnquqsip A*[reoi:rii9pi aq oq >x 3TW aturissB 9m ji
□ -Â/U)D > \[(uS)m - i(u*S)f}3\ »«W ipns
(/)£) ÎIB^SUOO B SJSIX9 9191^ ($j[)«3 9 / U^C-OIIIS AJ9A9 JOJ %VT\% U99S 9A«I{ 9^.
HA g/e("/[ug]"A) lsuo. _
T £||?x!|?dns '*SU°D ~
Z/S[US]Â/S\\-X\\^s.u.^saoo >
i=<y i==j
U^|]a3-*suo°
u
> [l(*4'Stella+ [|(U'*zM]a u3
yw\% os e|^| • ;suoo > KÂ^Z^l S9A!§ uraioaift scjot^x
(l"Q'2) Aipnb9ui uosuaf
9q^ pu-e (g'f^X'Z) raraiai Xq âS 9M 'pa^nqTj^sip-(si?'o)A[ 9J13 *A 9s™so9g
u-z) • [\(^z)v\u+[m^z)u\u^>u \[(us)m-[cs)m\
ojojojoq; ^9§ 9^/V 7 no pu9d9p ui3D uoium '(^ <£)# mj9^ ^S9JC J°l^L « q^TM
1=3/
[(*A'*z)a + (U'*z)tf]2 +
it
*■ u it
1=2/
[(*x + *z)/-(% + *z)/]3 = (us)f-(*s)f
9^TJM UT30 9M 'U!9J09q; StIOpft3X U9U^ PU13
rans oidoosap^ * ^sju Smsfi -^ = ^ + uz pire *£ = XA + TZ ^^ 9*°N-
9UU9Q
7 irôouis JOj in; Aju9a o^ uSnou9 si %[ '(^Q 3 f jftre joj o <- [("£)/] a'
SUI9109lfl IJUIiq £ J9^dVXJQ ^g
2.14. The central limit theorem 95
Proof. The same proof gives equation (5.4). We change the estimation of
Taylor \R(z,y)\ < 6(y) • y2 with 8(y) — 0 for \y\ -+ 0. Using the IID
property and using dominated convergence we can estimate the rest term
n
R = Ê[|i?(Zfc,n)|] +E[\R(Zk,Yk)\]
k=i
as follows:
R < J2EiS(Yk)Y2]+E[S(Yk)Y2}
fc=l
= n.E[S(^)^}+n-E[S(^)^-}
s »-El0IEI^1+nEI^)1Els'
= E0|E$1 + E01E|&
s El0c^0
D
The central limit theorem can be interpreted as a solution to a fixed point

problem:
Definition. Let Po,i be the space of probability measure p on (R, Br) which
have the properties that JR x2 dp(x) = 1, JR x dp(x) = 0. Define the map
Tn(A) = [ [ lA(^J)p(dx) nidy)

JRJR V*
on P0,i-
Corollary 2.14.4. The only attracting fixed point of T on Poa is the law of
the standard normal distribution.
Proof. If p is the law of a random variable X with Var[X] = 1 and E[X] —

0. Then T(p) is the law of the normalized random variable (X + X)/y/2 be
cause the independent random variables X, Y can be realized on the proba
bility space (R2, B,pxp) as coordinate functions X((x, y)) = x, Y((x, y)) =
y. Then T(p) is obviously the law of (X + Y)/y/2. Now use that Tn(X) =
( S 2 ™ ) * c o n v e r g e s i n d i s t r i b u t i o n t o AT ( 0 , 1 ) . □
For independent 0-1 experiments with win probability p € (0,1), the
central limit theorem is quite old. In this case
limP[f-^
»-«> <x] ~ =J ±-[X
y/np{l-p) 2^7.^ e^dy
as had been shown by de Moivre in 1730 in the case p=l/2 and for general
p e (0,1) by Laplace in 1812. It is a direct consequence of the central limit
theorem:
Corollary 2.14.5. (DeMoivre-Laplace limit theorem) The distribution of X*

converges to the normal distribution if Xn has the binomial distribution
B(n,p).
For more general versions of the central limit theorem, see [105]. The next
limit theorem for discrete random variables illustrates, why Poisson dis
tribution on N is natural. Denote by B(n,p) the binomial distribution on
{1,..., n } and with Pa the Poisson distribution on N \ {0 }.
Theorem 2.14.6 (Poisson limit theorem). Let Xn be a J3(n,pn)-distributed

and suppose npn —▶ a. Then Xn converges in distribution to a random
variable X with Poisson distribution with parameter a.
Proof We have to show that P[Xn = jfe] -> P[X = k] for each fixed keN.
p[xn = k] = (j)p£(i-p»: n—k

= "(n-1)(-2t>,-("-t+1)pi(i-p,)->
□
2.14. The central limit theorem
Figure. The binomial Figure. The binomial Figure. The Pois-

distribution 5(2,1/2) distribution 5(5,1/5) son distribution
has its support on has its support on with a — 1 on
{0,1,2}. {0,1,2,3,4,5}. N= {0,1,2,3,...}.
Exercice. It is custom to use the notation
*(«) = Fx(s) = -L f e^2/2dy

V ^7T J-oo
for the distribution function of a random variable X which has the standard
normal distribution N(0,1). Given a sequence of IID random variables Xn
with this distribution.
a) Justify that one can estimate for large n probabilities
P[a < S; < b] ~ *(b) - *(a) .
b) Assume X, are all uniformly distributed random variables in [0,1].

Estimate for large n
P[|S„/n-0.5|>e]
in terms of <E>, e and n.
large numbers.
Exercice. Define for A > 0 the transformation
Tx(jjl){A) =11 U(^) dn(x) dn(y)

JrJs. a
in V — Mi(R), the set of all Borel probability measures on R. For which A
can you describe the limit?
2.15 Entropy of distributions
Denote by v a (not necessarily finite) measure on a measure space (SI, A).
An example is the Lebesgue measure on R or the counting measure on N.
Note that the measure is defined only on a ^-subring of A since we did not
assume that v is finite.
Definition. A probability measure p on R is called v absolutely continuous,

if there exists a density / e Cl(v) such that p = fv. If p is ^-absolutely
continuous, one writes p < v. Call V(v) the set of all v absolutely contin
uous measures. The set V(v) is the set of functions / G C}(v) satisfying
f>0andjf(x)du(x) = l.
Remark. The fact that p < v defined earlier is equivalent to this is called
the Radon-Nykodym theorem ([?]). The function / is therefore called the
Radon-Nykodym derivative of p with respect to v.
Example. If v is the counting measure N = {0,1,2,... } and v is the
law of the geometric distribution with parameter p, then the density is
f(k)=p(l-p)k.
Example. If v is the Lebesgue measure on (-co, oo) and p is the law of
the standard normal distribution, then the density is f(x) = e~x2/2/y/2n.
There is a multi-variable calculus trick using polar coordinates, which im
mediately shows that / is a density:
J[ I e-^2+y^2
J r 2 dxdy
J o= P f Je-*"2/2
o rdOdr = 2n .
Definition. For any probability measure p e V(v) define the entropy
H ( p ) = f - f( u j ) \o g ( f( u ) ) d u ( u ) .
Jn
It generalizes the earlier defined Shannon entropy, where the assumption
had been dv — dx.
Example. Let v be the counting measure on a countable set SI, where A

is the cr-algebra of all subsets of Q and let the measure v is defined on the
5-ring of all finite subsets of Q,. In this case,
#(M) = £-/Miog(/(uO).
For example, for fl = N = {0,1,2,3,... } with counting measure v, the

geometric distribution P[{&}] = p(l - p)k has the entropy
±
is
-a - P) Wd - P)kP)
p
= bg(l^) -p«^ •
2.15. Entropy of distributions "
Example. Let v be the Lebesgue measure on R. If p = fdx has a density

function /, we have
H(pi)= f -f(x)\og(f(x))dx
JfL
For example, for the standard normal distribution fi with probability den
sity function f(x) = -jê-*2/2, the entropy is H(f) = (1 + log(27r))/2.
Example. If v is the Lebesgue measure dx on fi = E+ = [0, oo). A random

variable on 9. with probability density function f{x) = \e~Xx is called the
exponential distribution. It has the mean 1/A. The entropy of this distri
bution is (log(A) - 1)/A.
Example. If v is a probability measure on R, / a density and
A = {A1,...,.An}
is a partition on R. For the step function
f = Yi f f d u ) l A i e S ( u ) ,
the entropy H(fu) is equal to
H({Ai}) = Yi-v(Ai)log(v{Ai))
i
which is called the entropy of the partition {Ai}. The approximation of the
density / by a step functions / is called coarse graining and the entropy
of / is called the coarse grained entropy. It has first been considered by
Gibbs in 1902.
Remark. In ergodic theory, where one studies measure preserving trans

formations T of probability spaces, one is interested in the growth rate of
the entropy of a partition generated by A,T(A),..,Tn(A). This leads to
the notion of an entropy of a measure preserving transformation called
Kolmogorov-Sinai entropy.
Interpretation. Assume that ft is finite and that v the counting measure

and p({v}) = f(u) the probability distribution of random variable de
scribing the measurement of an experiment. If the event {u;} happens, then
-log(/(u>)) is a measure for the information or "surprise" that the event
happens. The averaged information or surprise is
H(ji) = ^-f{u>)log(f{u)).
If / takes only the values 0 or 1, which means that p is deterministic,
then H(p) = 0. There is no surprise and what the measurements show, is
the reality. On the other hand, if / is the uniform distribution on ft, then
H(p) = log(|fi|). We will see in a moment that this is the maximal entropy.
Definition. Given two probability measures p = fv and p = fu which are

both absolutely continuous with respect to v. Define the relative entropy
H(p\p) = f f(u)\og(f^)dv(x) €[0,oo].
It is the expectation Ep,[l] of the Likelihood coefficient / = log(|£4). The

negative relative entropy -H(p\p) is also called the conditional entropy.
One writes also H(f\f) instead of H(p\p).
Theorem 2.15.1 (Gibbs inequality). 0 < H(p\p) < +oo and H(p\p) = 0 if
and only if p — p.
Proof. We can assume H(p\p) < oo. The function u(x) = x log(x) is convex
on R+ = [0, oo) and satisfies u(x) > x - 1.
H(p\p) = f /»u(M) dv>f f(u;)(M - 1) dv = 0 .

Jn f(x) Jn f(x)
If p = p, then / = / almost everywhere and H(p\p) = 0.

On the other hand, if H(p\p) = 0, then by the Jensen inequality (2.5.1)
0 = EM[u(^)]>ii(EM[^]) = ti(l)=0.
Therefore, E^[u(j)] = ^(E^]). The strict convexity of u implies that j

must be a constant. Since both / and / are densities, we have / = /. □
Remark. The relative entropy can be used to measure the distance between
two distributions. It is not a metric although. The relative entropy is also
known under the name Kullback-Leibler divergence or Kullback-Leibler
metric, if v = dx [85].
2.15. Entropy of distributions 101
Theorem 2.15.2 (Distributions with maximal entropy). The following dis

tributions have maximal entropy.
a) If ft is finite with counting measure v. The uniform distribution on ft
has maximal entropy among all distributions on ft. It is unique with this
property.
b) ft = N with counting measure v. The geometric distribution with
parameter p = c-1 has maximal entropy among all distributions on
N = {0,1,2,3,... } with fixed mean c. It is unique with this property.
c) ft = {0,1}^ with counting measure v. The product distribution rjN,
where 77(1) = p,77(0) = l-p with p = c/N has maximal entropy among all
distributions satisfying E[SN] = c, where SN(u) = 5Zi=1u>*. It is unique
with this property.
d) ft = [0,00) with Lebesgue measure v. The exponential distribution with
density f(x) = Xe~Xx with parameter A on ft has the maximal entropy
among all distributions with fixed mean c = 1/A. It is unique with this
property.
e) ft = R with Lebesgue measure v. The normal distribution N(m,cr2)
has maximal entropy among all distributions with fixed mean m and fixed
variance a2. It is unique with this property.
f) Finite measures. Let (ft, A) be an arbitrary measure space for which
0 < v(ft) < 00. Then the measure v with uniform distribution / = l/u(ft)
has maximal entropy among all other measures on ft. It is unique with this
property.
Proof. Let p = fv be the measure of the distribution from which we want

to prove maximal entropy and let p = fv be any other measure. The aim
is to show H(p\p) = H(p) - H(p) which implies the maximality since by
the Gibbs inequality lemma (2.15.1) H(p\p) > 0.
In general,
H(p\p) = -H(p) - [ f(cu)\og(f(u)) dv
so that in each case, we have to show
ff(/i) = -/7(w)log(/M)di/. (2.8)
With
H<fi\v) = H(ji) - H(Ji)
we also have uniqueness: if two measures p, p have maximal entropy, then
H(p\p) = 0 so that by the Gibbs inequality lemma (2.15.1) p = p.
a) The density / = l/\ft\ is constant. Therefore H(p) = log(|ft|) and equa

tion (2.8) holds.
b) The geometric distribution on N = {0,1,2,... } satisfies P[{&}] = f(k) =

p(l —p)k, so that
J f(u)\og(f(u;))du = \og(p) + Jf(u;)k\og(l-p)du

= log(p) - log(l - p) / f(u) dv(u)
which is also the entropy of p.
c) The discrete density is f(u) = ps"(l - p)N~s» so that
log(/(fc)) = SN \og(p) + (N- SN) log(l - p)

and
£ f(k) log(/(fc)) = E[SN] log(p) + (N - E[SN]) log(l - p) .

k
The claim follows since we fixed E[Sjv]-
d) The density is f(x) = ote'^, so that log(/(z)) = log(a) - ax. The

claim follows since we fixed E[X] = J x dp(x) was assumed to be fixed for
all distributions.
e) For the normal distribution log(/(x)) = a + b(x - m)2 with two real
number a, b depending only on m and a. The claim follows since we fixed
Var[X] = E[(x - m)2] for all distributions.
f) The density / = 1 is constant. Therefore H(p) = 0 which is also on the

right hand side of equation (2.8). □
Remark. This result has relations to the foundations of thermodynamics,
where one considers the phase space of N particles moving in a finite region
in Euclidean space. The energy surface is then a compact surface ft and the
motion on this surface leaves a measure v invariant which is induced from
the flow invariant Lebesgue measure. The measure v is called the micro-
canonical ensemble. According to /) in the above, it is the measure which
maximizes entropy.
Remark. Let us try to get the maximal distribution using calculus of vari
ations. In order to find the maximum of the functional
H{f) = - j f\og{f)du
on Cx(v) under the constraints
F(f) = ! fdu = l, G(f) = f Xfdv = c,

Jet Jn
we have to find the critical points of H = H - XF - pG In infinite dimen
sions, constrained critical points are points, where the Lagrange equations
^jH(f) = \±F(f) + VL-JljG(f)

F(f) = 1
G(f) = c
are satisfied. The derivative d/df is the functional derivative and A, p are
the Lagrange multipliers. We find (/, A, v) as a solution of the system of
equations
-l-log(/(x)) = A + //X,
\ f(x)dv(x) = 1,
Jn
/ xf(x) dv(x) = c
Jq
by solving the first equation for /:
-A-Mx+1 dv^ = x
/ ■
x e --A—fix+1
x-»x+i djy^
/ ■
dividing the third equation by the second, so that we can get p from the
equation f xe~^xx du(x) = cfe~^x^ dv(x) and A from the third equation
ei+A _ j e-fix djy^xy This variational approach produces critical points of
the entropy. Because the Hessian D2(H) = —l/f is negative definite, it is
also negative definite when restricted to the surface in C1 determined by
the restrictions F = 1, G = c. This indicates that we have found a global
maximum.
Example. For ft = R, X(x) = x2, we get the normal distribution N(0,1).
Example. For ft = N, X(n) = en, we get f(n) = e~enXl/Z(f) with Z(f) =
Sn e~6nXl and where Ai is determined by ^n ene~CnAl = c. This is called
the discrete Maxwell-Boltzmann distribution. In physics, one writes A-1 =
kT with the Boltzmann constant k, determining T, the temperature.
Here is a dictionary matching some notions in probability theory with cor
responding terms in statistical physics. The statistical physics jargon is
often more intuitive.
Probability theory Statistical mechanics
Set ft Phase space
Measure space Thermodynamic system
Random variable Observable (for example energy)
Probability density Thermodynamic state
Entropy Boltzmann-Gibbs entropy
Densities of maximal entropy Thermodynamic equilibria
Central limit theorem Maximal entropy principle
Distributions, which maximize the entropy possibly under some constraint

are mathematically natural because they are critical points of a variational
principle. Physically, they are natural, because nature prefers them. From
the statistical mechanical point of view, the extremal properties of entropy
offer insight into thermodynamics, where large systems are modeled with
statistical methods. Thermodyanamic equilibria often extremize variational
problems in a given set of measures.
Definition. Given a measure space (ft, A) with a not necessarily finite

measure v and a random variable X G C. Given / G C1 leading to the
probability measure p = fv. Consider the moment generating function
Z(\) = E^[exx] and define the interval A = {A G R \ Z(\) < oo } in R.
For every A G A we can define a new probability measure
exx
on Q. The set
{fix | A € A }
of measures on (ft, A) is called the exponential family defined by v and X.
Theorem 2.15.3 (Minimizing relative entropy). For all probability measures

p which are absolutely continuous with respect to v, we have for all A G A
H(p\p)-XEil[X]>^\ogZ(X).
The minimum — logZ(A) is obtained for p\.
Proof. For every p = fv, we have
H(H\n) = J Q[ fl0g(l).l±dv
JX J
= -H(fi\u.x) + (- log(Z(A)) + AEA[X]) .
For p = p\, we have
H(»x\») =-log(Z(\)) + \E^[X] .
Therefore
H(p\p) - \Efi[X] = H(p\px) - \og(Z(X)) > -logZ(A) .
The minimum is obtained for p = p\. D

Corollary 2.15.4. (Minimizers for relative entropy)

a) p\ minimizes the relative entropy p »-» H(p\p) among all ^-absolutely
continuous measures p with fixed E/x[X].
b) If we fix A by requiring EMA [X] = c, then p\ maximizes the entropy
H(p) among all measures p satisfying Ep,[X] = c.
Proof, a) Minimizing p h-> H(p\p) under the constraint Ep,[X] = c is equiv

alent to minimize
H{ii\v) - AEA[X],
and to determine the Lagrange multiplier A by EMJX] = c. The above
theorem shows that nx is minimizing that,
b) If n = fu,nx = e~xxf/Z, then
0 < #(m,Ma) = -H(fl) + (- log(Z)) - AE„[X] = -H(fi) + H(jtx) .
a
Corollary 2.15.5. If v = \i is a probability measure, then fxx maximizes
F{u) = H(n) + XE^[X]
among all measures p which are absolutely continuous with respect to p.
Proof. Take p = v. Since then f = 1, H(p\p) = -H(p). The claim follows

from the theorem since a minimum of H(p\p) — XEp,[X] corresponds to a
maximum of F(p). □
This corollary can also be proved by calculus of variations, namely by

finding the minimum of F(f) = J/log(/) + Xf dv under the constraint
jfdv=l.
Remark. In statistical mechanics, the measure p\ is called the Gibbs distri

bution or Gibbs canonical ensemble for the observable X and Z(X) is called
the partition function. In physics, one uses the notation A = — (kT)~l,
where T is the temperature. Maximizing H(p) - (fcT)_1EM[X] is the same
as minimizing E^X] - kTH(p) which is called the tree energy if X is
the Hamiltonian and E^JX] is the energy. The measure p is the a priori
model, the micro canonical ensemble. Adding the restriction that X has
a specific expectation value c = EM[X] leads to the probability measure
px, the canonical ensemble. We illustrated two physical principles: nature
maximizes entropy when the energy is fixed and minimizes the free energy,
when energy is not fixed.
Example. Take on the real line the Hamiltonian X(x) = x2 and a measure
p = fdx, we get the energy J x2 dp. Among all symmetric distributions
fixing the energy, the Gaussian distribution maximizes the entropy.
Example. Let ft = N = {0,1,2,... } and X(k) = k and let v be the

counting measure on ft and p the Poisson measure with parameter 1. The
partition function is
Z(A) = £e^=exp(e*-l)
k
so that A = R and p\ is given by the weights
MA(fc) = exp(e-A + l)eAfc^ = e-Q^,
where a = ex — 0. The exponential family of the Poisson measure is the

family of all Poisson measures.
Example. The geometric distribution on N = {0,1,2,3,... } is an expo

nential family.
Example. The product measure on ft = {0,1 }N with win probability p is

an exponential family with respect to X(k) = k.
Example, ft = {1,... ,N}, v the counting measure and let pp be the bino
mial distribution with p. Take p = pi/2 and X(k) = k. Since
0 < H(p\p) = H(p\pp) + log(p)E[X] + log(l - p)E[(N - E[X])]

= -H(p\pp) + H(pp) ,
pp is an exponential family.
Remark. There is an obvious generalization of the maximum entropy prin

ciple to the case, when we have finitely many random variables {Xi}'2=1.
Given p = fv we define the (n-dimensional) exponential family
where
Z ( X ) = E fl [ e ^ ^ X i X i ]
is the partition function defined on a subset A of Rn.

2.16. Markov operators 107
Theorem 2.15.6. For all probability measures p which are absolutely con
tinuous with respect to v, we have for all A G A
H(p\p) - £ AiEA[Xi] > -logZ(A) .

i
The minimum -logZ(A) is obtained for p\. If we fix A* by requiring

EMA[Xi] = Ci, then p\ maximizes the entropy H(p) among all measures
p satisfying E^X*] = c*.
Assume v = p is a probability measure. The measure p\ maximizes
F(£)=JJ(£) + AEA[X].
Proof. Take the same proofs as before by replacing AX with A • X =
2.16 Markov operators

Definition. Given a not necessarily finite probability space (ft, A, v). A
linear operator P : ^(ft) —▶ C*(ft) is called a Markov operator, if
/ > 0 => Pf > 0,

/>o=H|p/||i =
Remark. In other words, a Markov operator P has to leave the closed
positive cone invariant C\ = {f € C1 | / > 0 } and preserve the norm on
that cone.
Remark. A Markov operator on (ft, A, v)'leaves invariant the set V(v) =

{/ £ C1 |/>0,||/||i = l} of probability densities. They correspond
bijectively to the set V(v) of probability measures which are absolutely
continuous with respect to v. A Markov operator is therefore also called a
stochastic operator.
Example. Let T be a measure preserving transformation on (ft, A, v). It is

called nonsingular if T*v is absolutely continuous with respect to v. The
unique operator P : C1 —> C1 satisfying
[ PAf d vJ=T ~ 1f A f dv
J
is called the Perron-Frobenius operator associated to T. It is a Markov
operator. Closely related is the operator Pf(x) = f(Tx) for measure pre
serving invertible transformations. This Koopman operator is often studied
on C2, but it becomes a Markov operator when considered as a transfor
mation on C1.
Exercice. Assume ft = [0,1] with Lebesgue measure p. Verify that the
Perron-Frobenius operator for the tent map
' T(x) = | 2x ,are [0,1/2]
2(1 -x) ,xe [1/2,1]
isPf(x) = \(f(\x) + f(l-\x))
Here is an abstract version of the Jensen inequality (2.5.1). It is due to M.

Kuczma. See [61].
Theorem 2.16.1 (Jensen inequality for positive operators). Given a convex

function u and an operator P : Cl -> Cl mapping positive functions into
positive functions satisfying PI = 1, then
u(Pf) < Pu(f)
for all / e C\ for which Pu(f) exists.
Proof. We have to show u(Pf)(u) < Pu(f)(u) for almost all uj e ft. Given
x = (Pf)(uj), there exists by definition of convexity a linear function y h+
ay + b such that u(x) =ax + b and u(y) >ay + b for all y eR. Therefore,
since af + b < u(f) and P is positive
u(Pf)(u) = a(Pf)(u) + b = P(af + b){u>) < P{u(f))(u>) .

□
The following theorem states that relative entropy does not increase along
orbits of Markov operators. The assumption that {/ > 0 } is mapped into
itself is actually not necessary, but simplifies the proof.
Theorem 2.16.2 (Voigt, 1981). Given a Markov operator V which maps

{/ > 0} into itself. For all f,geC\,
H(Vf\Vg) < H(f\g) .
Proof. We can assume that {g(uo) = 0} C A = {f(u) = 0} because nothing

is to show in the case H(f\g) = oo. By restriction to the measure space
space (Ac,AnAc,v(- n A)), we can assume / > 0,0 > 0 so that by our
assumption also Pf > 0 and Pg > 0.
2.16. Markov operators 109
(i) Assume first (f/g)(u) < c for some constant ceR.
For fixed g, the linear operator Rh = P(hg)/P(g) maps positive functions
into positive functions. Take the convex function u(x) = xlog(x) and put
h = f/g. Using Jensen's inequality, we get
which is equivalent to Pf log ^ < P(flog(f/g)). Integration gives
H(Pf\Pg) = J Pf iog ?ld„
< j P(f log(//fl)) dv = J f log(f/g) dv = H(f\g) .
(ii) Define fk = inf(/, kg) so that fk/g < k. We have fk c /fc+i and
fk -▶ / in C1. From (i) we know that H(Pfk\Pg) < H(fk\g). We can
assume H(f\g) < oo because the result is trivially true in the other case.
Define B = {/ < g}. On B, we have fklog(fk/g) = flog(f/g) and on ft\B
we have
fklog(fk/g) < fk+ilog(fk+1/g)u-+ flog(f/g)
so that by Lebesgue dominated convergence theorem,
H(f\g) = k—+oc
lim H(fk\g) .
As an increasing sequence, Pfk converges to Pf almost everywhere. The

elementary inequality x log(x) - x > x log(y) - y for all x > y > 0 gives
(Pfk) log(Pfk) - (Pfk) log(Pg) - (Pfk) + (Pg) > 0 .

Integration gives with Fatou's lemma (2.4.2)
H(Pf\Pg) - \\Pf\\ + \\Pg\\ < lim inf H(Pfk\Pg) - \\Pfk\\ + ||Pfl||

k—>oc
and so H(Pf\Pg) < liminf^^ H(Pfk\Pg). □
Corollary 2.16.3. For an invertible Markov operator V, the relative entropy

is constant: H{Vf\Vg) = H(f\g).
Proof. Because V and V : are both Markov operators,
H(f\g) = H{VV-lf\VV-xg) < HiJ>-x!\V~xg) < H(f\g) .
□
11 0 Chapter 2. Limit theorems
Example. If a measure preserving transformation T is invertible, then the

corresponding Koopman operator and Perron-Frobenius operators preserve
relative entropy.
Corollary 2.16.4. The operator T(p)(A) = JR2 U(^) dp(x) dp(y) does
not decrease entropy.
Proof. Denote by X^ a random variable having the law p and with p(X)
the law of a random variable. For a fixed random variable Y, we define the
Markov operator
Because the entropy is nondecreasing for each Py, we have this property
also for the nonlinear map T(p) = Px^(p). D
We have shown as a corollary of the central limit theorem that T has a

unique fixed point attracting all of -Po,i. The entropy is also strictly in
creasing at infinitely many points of the orbit Tn(p) since it converges to
the fixed point with maximal entropy. It follows that T is not invertible.
More generally: given a sequence Xn of IID random variables. For every n,

the map Pn which maps the law of 5* into the law of 5*+1 is a Markov
operator which does not increase entropy. We can summarize: summing up
IID random variables tends to increase the entropy of the distributions.
A fixed point of a Markov operator is called a stationary state or in more
physical language a thermodynamic equilibrium. Important questions are:
is there a thermodynamic equilibrium for a given Markov operator V and
if yes, how many are there?
2.17 Characteristic functions

Distribution functions are in general not so easy to deal with, as for ex
ample, when summing up independent random variables. It is therefore
convenient to deal with its Fourier transforms, the characteristic functions.
It is an important topic by itself [60].
Definition. Given a random variable X, its characteristic function is a real-
valued function on R defined as
4>x(u) = E[eiuX].
If Fx is the distribution function of X and px its law, the characteristic
function of X is the Fourier-Stieltjes transform
^(t) = f eitx dFx(x) = [ eitx px(dx) .

JR JR
2.17. Characteristic functions 111
Remark. If Fx is a continuous distribution function dFx(x) = fx(x) dx,
then 0x is the Fourier transform of the density function fx:
[ eitx fx(x) dx .
Jr
Remark. By definition, characteristic functions are Fourier transforms of
probability measures: if p is the law of X, then (j>x = P-
Example. For a random variable with density fx(%) = xm'/(m + 1) on
ft = [0,1] the characteristic function is
-«x„mdx/tm
^Hm-f MU_
0X(*) = / e2t*xm 1) = m!(l-e"em(-it))
Jo (-2t)1+m(ra+l)
where en(x) = Ylô xk/(k\) is the n'th partial exponential function.
Theorem 2.17.1 (Levy formula). The characteristic function 4>x determines

the distribution of X. If a, b are points of continuity of F, then
i /"oo —ita _ p—itb
Fx(b) -Fx(a) = —J <f>x(t) dt (2.9)
In general, one has

oo —ita p—itb i i
- / r^<f>x(t) dt = ri(a,b)] + -MW] + ^[{b}} .
Proof Because a distribution function F has only countably many points of

discontinuities, it is enough to determine F(b) — F(a) in terms of <j> if a and
b are continuity points of F. The verification of the Levy formula is then
a computation. For continuous distributions with density Ffx — fx is the
inverse formula for the Fourier transform: fx(o) = ^ J^° c~lta(j)x(t) dt
so that Fx(a) = ^ f^ ^^(t>x(t) dt. This proves the inversion formula
if a and b are points of continuity.
The general formula needs only to be verified when p is a point measure
at the boundary of the interval. By linearity, one can assume p is located
on a single point b with p = P[X = b] > 0. The Fourier transform of the
Dirac measure p5b is 4>x(t) = peltb. The claim reduces to
i r
it^ P e « » d t = P -
which is equivalent to the claim lim^ôo J_R e it 1 dt = tt for c > 0.

Because the imaginary part is zero for every R by symmetry, only
. sin(tc) ,
l i m / / —K—t dt = ir
hm
R-+ooJ_p> t
remains. The verification of this integral is a prototype computation in
residue calculus. D
Theorem 2.17.2 (Characterization of weak convergence). A sequence Xn

of random variables converges weakly to X if and only if its characteristic
functions converge point wise:
<t>xn (x) -▶ (j)X
Proof. Because the exponential function eitx is continuous for each t, it

follows from the definition that weak convergence implies the point wise
convergence of the characteristic functions. From formula (2.9) follows that
if the characteristic functions converge point wise, then convergence in dis
tribution takes place. We have learned in lemma (2.13.2) that weak con
vergence is equivalent to convergence in distribution. □
Example. Here is a table of characteristic functions (CF) <f>x(t) = E[eux]

and moment generating functions (MGF) Mx(t) = E[etx] for some familiar
random variables:
Distribution Parameter CF MGF
Normal m e R, a2 > 0 emit-o*tz /2 emt+<r'!t'72
N(0,l) e-e/2 e*72

Uniform [-a, a] sm(at)/(at) smh(at) / (at)
Exponential A>0 A/(A - it) X/(X-t)
binomial n>l,pe [0,1] (l-p + pelt)n (l-p + pe1)71
Poisson A>0, A eA(e"-l) eHel-l)
p p
Geometric pe(0,i) (l-(l-p)e>< (l-(l-p)e*
pe" pe1
first success pe(0,i) (1-(1-P)e» (l-(l-p)e*
Cauchy m e R, b > 0 pimt—\t\ emt-\t\
Definition. Let F and G be two probability distribution functions. Their

convolution F • G is defined as
F*G(x)= [ F(x-y)dG(y)
Jr
Lemma 2.17.3. If F and G are distribution functions, then F • G is again

a distribution function.
2.17. Characteristic functions 11 3
Proof. We have to verify the three properties which characterize distribu

tion functions among real-valued functions as in proposition (2.12.1).
a) Since F is nondecreasing, also F • G is nondecreasing.
b) Because F(-oo) = 0 we have also F*G(-oo) = 0. Since F(oo) = 1 and
dG is a probability measure, also F • G(oo) = 1.
c) Given a sequence hn —> 0. Define Fn(x) = F(x + hn). Because F is con
tinuous from the right, Fn(x) converges point wise to F(x). The Lebesgue
dominated convergence theorem implies that Fn • G(x) = F • G(x -f hn)
converges to F • G(x). □
Example. Given two discrete distributions
F(x) = Yl^ G(x)= Ylqn •

n<x n<x
Then F*G(x) = *52n<x(p*q)n, where p*q is the convolution of the sequences

p,q defined by (p • q)n = Ylk=oPkQn-k- We see that the convolution of
discrete distributions gives again a discrete distribution.
Example. Given two continuous distributions F, G with densities h and k.

Then the distribution of F • G is given by the convolution
h*k(x)= I h(x — y)k(y)dy

Jr
because
(F * G)'(x) = ^-JF(x- y)k(y) dy = J h(x - y)k(y) dy .
Lemma 2.17.4. If F and G are distribution functions with characteristic

functions </> and i/j, then F • G has the characteristic function 0 • ip.
Proof. While one can deduce this fact directly from Fourier theory, we
prove it by hand: use an approximation of the integral by step functions:
J[ r eiuxd(F*G)(x)
iV2n
N2n
= lim
N.n—+oo E eiuk2~n
fc=-N2n
[lFd-y)-F(t-±-y))dG(y)
- N 2 +n +
l 1
JR
^2"
= lim
N,n—▶oo £ / c-*-"[F(£ - y) - F(^— - y)] • e^ dGfo)
^=-^2^ + 1
r i V- y
= / [Urn / V eiux dF(x)]eiuy dG(y) = [ 4>(u)eiuy dG(y)
JRN->ooJ_N_y JR
= <p(u)ip(u) .
D
It follows that the set of distribution functions forms an associative com

mutative group with respect to the convolution multiplication. The reason
is that the characteristic functions have this property with point wise mul
tiplication.
Characteristic functions become especially useful, if one deals with inde

pendent random variables. Their characteristic functions multiply:
Proposition 2.17.5. Given a finite set of independent random variables

Xj,j = 1,... ,n with characteristic functions <j>j. The characteristic func
tion of y%=i Xj is (/) = n£=i <t>j.
Proof. Since Xj are independent, we get for any set of complex valued
measurable functions gj, for which E[gj(Xj)] exists:
nf[9J(Xj)] = f[E[gj(Xj)].
3=1 j=l
Proof: This follows almost immediately from the definition of independence

since one can check it first for functions gj = 1^, where Aj are a(Xj
measurable functions for which gj(Xj)gk(Xk) = IajHA* and
E[9j(Xj)gk(Xk)] = m(Aj)m(Ak) = E[gj(Xj)]E[gk(Xk)] ,
then for step functions by linearity and then for arbitrary measurable func
tions.
If we put gj(x) = exp(ix), the proposition is proved. □
Example. If Xn are IID random variables which take the values 0 and 2 with
probability 1/2 each, the random variable X = Y^=i Xn/^n is a random
variable with the Cantor distribution. Because the characteristic function
of Xn is (j)xn/3n(t) = E[e*tXn/3n] = e* 2 -1, we see that the characteristic
function of X is
™ eiV3n _ 1
M*)=nf—2—^-
The centered random variable Y = X - 1/2 can be written as Y —
Z^Li yn/3n, where Yn takes values -1,1 with probability 1/2. So
4>Y{t) = YlneitY:/3n]=n f-Jr— = II cos(f} •

n .«» n n=l
2.17. Characteristic functions 115
This formula for the Fourier transform of a singular continuous measure p
has already been derived by Wiener. The Fourier theory of fractal measures
has been developed much more since then.
Figure. The characteristic func

tion 4>y(t) of a random variable
Y with a centered Cantor distri
bution supported on [—1/2,1/2]
has an explicit formula 4>y(t) =
n^=icos(^r) and already been
derived by Wiener in the early
20'th century. The formula can
also be used to compute moments
of Y with the moment formula
E[X™] = (-ir£rcf>x(t)\t=o.
Corollary 2.17.6. The probability density of the sum of independent ran

dom variables YTj=i x3; is fx • f2 * • • • • fn, if Xj has the density fj.
Proof. This follows immediately from proposition (2.17.5) and the alge
braic isomorphisms between the algebra of characteristic functions with
convolution product and the algebra of distribution functions with point
wise multiplication. □
Example. Let Yk be IID random variables and let Xk = XkYk with 0 < A <
1. The process Sn = XX=i Xk is called the random walk with variable step
size or the branching random walk with exponentially decreasing steps. Let
p be the law of the random sum X = J2k=i Xk. If (fry (t) is the characteristic
function of Y, then the characteristic function of X is
<t>x(t) = \[<t>x{t\n)-
71=1
For example, if the random Yn take values -1,1 with probability 1/2, where
(j)Y(t) = cos(t), then
<Px(t) = l[cos(tXn)
n=l
The measure p is then called a Bernoulli convolution. For example, for
A = 1/3, the measure is supported on the Cantor set as we have seen
above. For more information on this stochastic process and the properties
of the measure p which in a subtle way depends on A, see [41].
Exercice. Show that Xn —> X in distribution if and only if the distribution
functions satisfy 4>xn{t) —▶ <t>x{t) for all <6M.
Exercice. The characteristic function of a vector valued random variable

X = (Xi,..., Xk) is the real-valued function
tf>x(t) = E[ert-x]
on Rfc, where we wrote t = (t\,... ,tfc). Two such random variables X, F
are independent, if the cr-algebras X"1(B) and y_1(S) are independent,
where B is the Borel cr-algebra on Rk.
a) Show that if X and Y are independent then 4>x+y = <t>x • 4>y-
b) Given a real nonsingular k x k matrix A called the covariance matrix
and a vector m = (mi,... , m^) called the mean of X. We say, a vector
valued random variable X has a Gaussian distribution with covariance A
and mean m, if
<M*) eimt-±(t-At)
Show that the sum X + Y of two Gaussian distributed random variables is

again Gaussian distributed.
c) Find the probability density of a Gaussian distributed random variable
X with covariance matrix A and mean m.
Exercice. The Laplace transform of a positive random variable X > 0 is

defined as lx(t) = E[e~tx]. The moment generating function is defined as
M(t) — E[etx] provided that the expectation exists in a neighborhood of
0. The generating function of an integer-valued random variable is defined
as C(X) = E[ux] for u € (0,1). What does independence of two random
variables X, Y mean in terms of (i) the Laplace transform, (ii) the moment
generating function or (iii) the generating function?
Exercice. Let (0, A, /x) be a probability space and let U,V E. X be ran
dom variables (describing the energy density and the mass density of a
thermodynamical system). We have seen that the Helmholtz free energy
EA[t/] - kTH[p\
(k is a physical constant), T is the temperature, is taking its minimum for

the exponential family. Find the measure minimizing the free enthalpy or
Gibbs potential
Efi[U}-kTH[fi\-PE^V},
where p is the pressure.
2.18. The law of the iterated logarithm H7
Exercice. Let (fi, A, p) be a probability space and Xi e C random variables.
Compute E^Xi] and the entropy of p\ in terms of the partition function
Z(X).
Exercice. a) Given the discrete measure space (ft = {e0 + nS},u), with
e0 G M and 6 > 0 and where v is the counting measure and let X(k) = k.
Find the distribution / maximizing the entropy H(f) among all measures
p = fv fixing Ep,[X] = e.
b) The physical interpretation is as follows: ft is the discrete set of ener
gies of a harmonic oscillator, e0 is the ground state energy, 8 = hw is the
incremental energy, where u is the frequency of the oscillation and h is
Planck's constant. X(k) = k is the Hamiltonian and E[X] is the energy.
Put A = 1/fcT, where T is the temperature (in the answer of a), there ap
pears a parameter A, the Lagrange multiplier of the variational problem).
Since can fix also the temperature T instead of the energy e, the distribu
tion in a) maximizing the entropy is determined by u and T. Compute the
spectrum e(uo,T) of the blackbody radiation defined by
J1
e(uj,T) = (E[X]-e0)
where c is the velocity of light. You have deduced then Planck's blackbody
radiation formula.
2.18 The law of the iterated logarithm

We will give only a proof of the law of iterated logarithm in the special
case, when the random variables Xn are independent and have all the
standard normal distribution. The proof of the theorem for general IID
random variables Xn can be found for example in [105]. The central limit
theorem makes the general result plausible from the special case.
Definition. A random variable X G C is called symmetric if its law px
satisfies:
p([-b,-a)) =p([a,b))
for all a < b. A symmetric random variable X G C1 has zero mean. We
again use the notation Sn — Y^k=\ Xk in this section.
Lemma 2.18.1. Let Xn by symmetric and independent. For every e > 0
P[ Kmax
k < nSk > e] < 2P[(Sn > e] .
Proof This is a direct consequence of Levy's theorem (2.11.6) because we

can take m = 0 as the median of a symmetric distribution. □
Definition. Define for n > 2 the constants An = V2nloglogn. It grows only

slightly faster than \/2n. For example, in order that the factor Vlog log n
is 3, we already have n = exp(exp(9)) > 1.33 • 103519.
Theorem 2.18.2 (Law of iterated logarithm for JV(0,1)). Let Xn be a se

quence of IID JV(0, l)-distributed random variables. Then
limsup —- = 1, liminf -^ = -1 .
n-KX) An n-voo An
Proof We follow [47]. Because the second statement follows obviously from
the first one by replacing Xn by -Xn, we have only to prove
limsup£n/An = 1 .
n—>oo
(i) P[5n > (1 + e)An, infinitely often] = 0 for all e > 0.
Define nk = [(1 + e)k] G N, where [x] is the integer part of x and the events
Ak = {Sn > (1 + e)An, for some n G (nk, rik+i] }•
Clearly limsupfc Ak = {Sn > (l-he)An, infinitely often}. By the first Borel-
Cantelli lemma (2.2.2), it is enough to show that Y,k P[^fc] < oo. For each
k, we get with the above lemma
P[Ak] < P[ max Sn>(l + e)Afc]

nk<n<nk+i
< P[ max Sn > (1 + c)Afc]
l<n<nk+1
< 2P[Snk+1 > (1 + c)Afc] .
The right-hand side can be estimated further using that Snk+l/y/rik+i

is N(0, l)-distributed and that for a iV(0, l)-distributed random variable
P[X >t)< const • e-'2/2
2P[5„fc+1>A,] = 2P[(^>(l + e)^nk^nk)}
< Cexp(-l(l+e)2)2nfclQgl0gnfc)
2 rik+i
< Ci exp(-(l + e) loglog(nfc))
= Cilog(nfc)-(1+£><C2fc-(1+€).
2.18. The law of the iterated logarithm 11 9
Having shown that P[Ak] < const ./H1+e) proves the claim £fc P[Ak] < oo.
(ii) P[Sn > (1 - e)An, infinitely often] = 1 for all e > 0.
It suffices to show, that for all e > 0, there exists a subsequence nk
P[^nfc > (1 - <0Anfc, infinitely often] = 1 .
Given e > 0. Choose N > 1 large enough and c < 1 near enough to 1 such
that
c^l - l/N - 2/y/N > 1 - e . (2.10)
Define nk = Nk and Ank = nk - nk-i. The sets
Ak = {Snk - Sn^ > c^/2AnkloglogAnk}
are independent. In the following estimate, we use the fact that Jt°° e~x /2 dx >
C • e~l I2 for some constant C.
P[Ak] = P[{Snk - 5nfc_x > <V2AnfcloglogAnfc}]

" ^rrgnfc-gnfc-!
FK y / A h - k^ y2AnfcloglogAnfcl1
>C y/A^ il
> C • exp(-c2 log log Ank) = C • exp(-c2 log(fc log N))

= d • exp(-c2 logfc) = Cik~c2
so that J2k p[^fc] = °°- We nave tnerefore °y Borel-Cantelli a set A of full

measure so that for lj G A
Snk - Sn^ > c^2AnkloglogAnk
for infinitely many fc. From (i), we know that
Snk > -2\]2nk log log nk
for sufficiently large fc. Both inequalities hold therefore for infinitely many
values of fc. For such fc,
Snk(u) > Sn^ (u) + cy/2Ank log log Anfc

> -2y/2nk-i loglognfc_i + cy/2Ank log log Ank
> (-2/ v^ + cy/1 - l/N) ^2nk log log nk
> (1 -e)y/2nk log log nk ,
where we have used assumption (2.10) in the last inequality. □
We know that JV(0,1) is the unique fixed point of the map T by the central
limit theorem. The law of iterated logarithm is true for T(X) implies that
it is true for X. This shows that it would be enough to prove the theorem
in the case when X has distribution in an arbitrary small neighborhood of
N(0,1). We would need however sharper estimates.
We present a second proof of the central limit theorem in the IID case, to
illustrate the use of characteristic functions.
Theorem 2.18.3 (Central limit theorem for IID random variables). Given
Xn G C which are IID with mean 0 and finite variance a2. Then
Sn/(ay/n) -> N(0,1) in distribution.
Proof. The characteristic function of N(0,1) is (j)(t) = e-*2/2. We have to

show that for all * G R
E[eft3k] - e"<2/2 .
Denote by </>Xn the characteristic function of Xn. Since by assumption
E[Xn] = 0 and E[X2] = a2, we have
4>xn(t) = l-Yt2 + o(t2)
Therefore
E[ea^] = <j>xA—r)n
CFy/n
= (i-J-I t 2+ *R)n
1
2 n n
= e-<2/2 + o(l).
D
This method can be adapted to other situations as the following example
shows.
Proposition 2.18.4. Given a sequence of independent events An c ft with

P[An] = 1/n. Define the random variables Xn = lAn and Sn = X)£=i Xk.
Then
T = Sn - log(n)
\/lo8(n)
converges to JV(0,1) in distribution.
Proof
E[^]=El=1°g(n) + 7 + o(l),
fc=l'c
2.18. The law of the iterated logarithm 121
where 7 = limn-oo ££=1 \ ~ los(n) is the Euler constant
1,- lx , / n *"*
Var[5n] = £ -(1 - -) = log(n) + 7 - y + o(l)
fc=i
satisfy E[Tn] -> 0 and Var[Tn] -▶ 1. Compute 0Xn = 1 - £ + ^ so that

0sn(i) = IlLi(l " i + x) and ^nW = ^(*(*))^log(n), where * =
t/y/\og(n). For n —▶ 00, we compute
n x
log^rnW = -ity/log(n) + £log(l + -(e*s - 1))
n 1 1
= -itv/log(n) + J^log(l + ^(is - -s2 + o(s2)))
fc=i
n 1 1 n 52
= -itVic^H + ^-(i5 + -s2 + o(52)) + o(^^)
fc=i k=i
= -ity/l^fn) + (is - l-s2 + o(s2))(log(n) + 0(1)) + t20(l)
We see that Tn converges in law to the standard normal distribution. □

Chapter 3
Discrete Stochastic Processes
3.1 Conditional Expectation

Definition. Given a probability space (Q,A,P). A second measure P' on
(ft, A) is called absolutely continuous with respect to P, if P[A] = 0 implies
Pf[A] = 0 for all A G A. One writes P' < P.
Example. If P [a, b] = b — a is the uniform distribution on Q = [0,1] and A

is the Borel cr-algebra, and Y G C1 satisfies Y(x) > 0 for all x G fl, then
P'[a,6] = / Y(x) dx is absolutely continuous with respect to P.
Example. Assume P is again the Lebesgue measure on [0,1] as in the last

example. If Y(x) = lB(x), then P'[A] = P[AnB] for all A G A. IfP[B] < 1,
then P is not absolutely continuous with respect to P'. We have P'[£?c] = 0
butP[J3c] = l-P[£] >0.
f 1 1/2 G A
Example. If P'[A\ = < n J icy a a > then P' ls not absolutely continuous
with respect to P. For B = {1/2}, we have P[B] = 0 but P'[B] = 1^0.
The next theorem is a reformulation of a classical theorem of Radon-

Nykodym of 1913 and 1930.
Theorem 3.1.1 (Radon-Nykodym equivalent). Given a measure P' which

is absolutely continuous with respect to P, then there exists a unique
Y G £*(P) with P' = YP. The function Y is called the Radon-Nykodym
derivative of P' with respect to P. It is unique in L1.
Proof. We can assume without loss of generality that P' is a positive mea
sure (do else the Hahn decomposition P = P+ — P~), where P+ and P~
123
124 Chapter 3. Discrete Stochastic Processes
are positive measures).
(i) Construction: We recall the notation E[Y; A] = E[U Y] = JA Y dP.

The set T = {Y > 0 | E[Y; A] < P'[A], VA G A } is closed under formation
of suprema
E [ Yi v y 2 ; 4 = E [ y i ; A n { y i > y 2 } ] + E [ y 2 ; A n { y 2 > Yi } ]
< p'[a n {Fi > y2}] + p'[a n {y2 > yx}] = p'[A]
and contains a function Y different from 0 since else, P' would be singular
with respect to P according to the definition (2.15) of absolute continuity.
We claim that the supremum Y of all functions T satisfies yP = Pf: an
application of Beppo-Levi's theorem (2.4.1) shows that the supremum of T
is in T. The measure P" — P' — YP is the zero measure since we could do
the same argument with a new set V for the absolutely continuous part of
Pn.
(ii) Uniqueness: assume there exist two derivatives Y,Y'. One has then
E[y - Yf; [Y > Y'}] = 0 and so Y > Y' almost everywhere. A similar
argument gives Y' <Y almost everywhere, so that Y — Y' almost every
where. In other words, Y — Y' in L1. □
Theorem 3.1.2 (Existence of conditional expectation, Kolmogorov 1933).

Given X G £X(A) and a sub a-algebra B C A. There exists a random
variable Y G Cl(B) with fAY dP = JAX dP for all A G B.
Proof Define the measures P[A] = P[A] and P'[A] = JAX dP = E[X;A]
on the probability space (tt,B). Given a set B G B with P[B] = 0, then
Pf[B] = 0 so that P' is absolutely continuous with respect to P. Radon-
Nykodym's theorem (3.1.1) provides us with a random variable Y G Cl(B)
with P'[A] = JAXdP = JA Y d P. D
Definition. The random variable Y in this theorem is denoted with E[X|/3]

and called the conditional expectation of X with respect to B. The random
variable Y G CX(B) is unique in LX(B). If Z is a random variable, then
E[-X"|Z] is defined as E[X|<r(Z)]. If {Z}j is a family of random variables,
then E[X|{Z}j] is defined as E[X\a({Z}x)].
Example. If B is the trivial a-algebra B = {0, ft}, then E[X\B] = X.
Example. If B = A, then E[X\B] = E[X].

Example. If B = {0, Y, Yc, Q} then
Frvimr ^ J ^y)^XdP for "Gy'

3.1. Conditional Expectation 125
Example. Let (Sl,A,V) = ([0,1} x [0,l],A,dxdy), where A is the Borel

a-algebra defined by the Euclidean distance metric on the square ft. Let B
be the a-algebra of sets A x [0,1], where A is in the Borel a-algebra of the
interval [0,1]. If X(x, y) is a random variable on ft, then Y = E[X\B] is the
random variable
Y(x,y) = / X(x,y)dy .
Jo
This conditional integral only depends on x.
Remark. This notion of conditional entropy will be important later. Here

is a possible interpretation of conditional expectation: for an experiment,
the possible outcomes are modeled by a probability space (ft, A) which is
our "laboratory". Assume that the only information about the experiment
are the events in a subalgebra B of A. It models the "knowledge" obtained
from some measurements we can do in the laboratory and B is generated by
a set of random variables {Zi}iej obtained from some measuring devices.
With respect to these measurements, our best knowledge of the random
variable X is the conditional expectation E[X|B]. It is a random variable
which is a function of the measurements Z*. For a specific "experiment
uj, the conditional expectation E[X|S](u;) is the expected value of X(uj),
conditioned to the a-algebra B which contains the events singled out by
data from Xi.
Proposition 3.1.3. The conditional expectation X i-> E[X|S] is the projec

tion from C2(A) onto C2(B).
Proof. The space C2(B) of square integrable B-measurable functions is a

linear subspace of £2(^4). When identifying functions which agree almost
everywhere, then L2(B) is a Hilbert space which is a linear subspace of the
Hilbert space L2(A). For any X G C2(A), there exists a unique projection
p(X) G C2(B). The orthogonal complement £2(S)-L is defined as
C2(B)L = {Ze C2(A) | (Z, Y) := E[Z -Y]=0 for all Y G C2(B) } .
By the definition of the conditional expectation, we have for A G B
(X - E[X\B], 1A) = E[X - E[X\B\; A]=0.
Therefore X-E[X|B] G C2(B)^. Because the map q(X) = E[X\B] satisfies

q2 = q, it is linear and has the property that (1 — q)(X) is perpendicular
to C2(B), the map q is a projection which must agree with p. □
Example. Let ft = {1,2,3,4} and A the a-algebra of all subsets of ft. Let
B = {0, {1,3}, {2,4}, ft}. What is the conditional expectation Y = E[X\B]
of the random variable X(k) = fc2? The Hilbert space C2(A) is the four-
dimensional space R4 because a random variable X is now just a vector
X = (X(1),X(2),X(3),X(4)) = (1,4,9,16). The Hilbert space C2(B) is
the set of all vectors v = (v\,V2,vz,v±) for which v\ = v2 and V3 = V4
because functions which would not be constant in (vi,v2) would gener
ate a finer algebra. It is the two-dimensional subspace of all vectors {v =
(a,a,b,b) \ a,b G R}. The conditional expectation projects onto that plane.
The first two components (X(1),X(2)) project to (X(1HX(2) ? xi1)î2))^
the second two components project to (*(3)+_x(4) ? x&)+x(*)y Therefore,
X(l) + X(2) X(l)+X(2) X(3) + X(4) X(3) + X(4),
Remark. This proposition 3.1.3 means that Y is the least-squares best Im

measurable square integrable predictor. This makes conditional expectation
important for controlling processes. If B is the a-algebra describing the
knowledge about a process (like for example the data which a pilot knows
about an plane) and X is the random variable (which could be the actual
data of the flying plane), we want to know, then E[X|S] is the best guess
about this random variable, we can make with our knowledge.
Exercice. Given two independent random variables X,Y G C2 such that

X has the Poisson distribution Pa and Y has the Poisson distribution PM.
The random variable Z = X -\-Y has Poisson distribution Pa+// as can
be seen with the help of characteristic functions. Let B be the a-algebra
generated by Z. Show that
E[X\B] =X +-±-Z
p
Hint: It is enough to show
E[X;{Z = k}] = A -±-P[Z = k]

-f- p
Even if random variables are only in C1, the next list of properties of
conditional expectation can be remembered better with proposition 3.1.3
in mind which identifies conditional expectation as a projection, if they are
in£2.
Theorem 3.1.4 (Properties of conditional expectation). For given random

variables X, Xn, Y G C, the following properties hold:
(1) Linearity: The map X i—▶ E[X|S] is linear.
(2) Positivity: X > 0 => E[X\B] > 0.
(3) Tower property: CcBcA^ E[E[X\B]\C] = E[X\C].
(4) Conditional Fatou: \Xn\ < X, Epiminfn-.oo Xn\B] <
liminfôoEÎS].
(5) Conditional dominated convergence: \Xn\ < X,Xn —> X a.e.
Ê[Xn\B]Ê[X\B] a.e.
(6) Conditional Jensen: if h is convex, then E[h(X)\B] > h(E[X\B]).
Especially ||E[X|S]||P < |\X\\p.
(7) Extracting knowledge: For Z G C°°(B), one has E[ZX\B] = ZE[X\B].
(8) Independence: if X is independent of C, then E[X\C] = E[X].
Proof. (1) For positivity, note that if Y = E[X\B] would be negative on a

set of positive measure, then A = Y~l([-l/n, 0]) G B would have positive
probability for some n. This would lead to the contradiction 0 < E[1^X] =
E[Uy] < -n-1m(A) < 0.
(2) Use that P" < P' < P implies P" = Y'P' = YfYP and P" < P gives
P" = ZP so that Z = Y'Y almost everywhere.
(3) This is especially useful when applied to the algebra Cy = {0, Y, Yc, ft}.
Because X < Y almost everywhere if and only if E[X|Cy] < E[y|Cy] for
all Y eB.
(4)-(5) The conditional versions of the Fatou lemma or the dominated
convergence theorem are true, if they are true conditioned with Cy for
each Y eB. The tower property reduces these statements to versions with
B = CY which are then on each of the sets Y, Yc the usual theorems.
(6) Chose a sequence (an, bn) G R2 such that h(x) = supn anx + bn for all
x G R. We get from h(X) > anX + bn that almost surely E[h(X)\G] >
anE[X\Q] + bn. These inequalities hold therefore simultaneously for all n
and we obtain almost surely
E[h(X)\Q] > sup(anE[X|£] + bn) = h(E[X\Q]) .

n
The corollary is obtained with h(x) = \x\p.
(7) It is enough to condition it to each algebra Cy for Y € B. The tower

property reduces these statements to linearity.
(8) By linearity, we can assume X > 0. For B G B and C G C, the random

variables XIB and lc are independent so that
E[XlBnC] = E[XlBlc] = E[X1B]P[C] .

The random variable Y = E[X\B] is B measurable and because Y1B is
independent of C we get
E[(yifl)ic] = E[yiB]P[ci
so that E[lBnc^] = E[lBnCY]. The measures on o(B,C)
p\A^ E\\AX], v\A*-+ E[1AY]
agree therefore on the 7r-system of the form Bf\C with B G B and C eC

and consequently everywhere on a(B, C). □
Remark. From the conditional Jensen property in theorem (3.1.4), it fol

lows that the operation of conditional expectation is a positive and contin
uous operation on CP for any p > 1.
Remark. The properties of Conditional Fatou, Lebesgue and Jensen are
statements about functions in CX(B) and not about numbers as the usual
theorems of Fatou, Lebesgue or Jensen.
Remark. Is there for almost all u G ft a probability measure P^ such that
E[X\B](u)= [ XdP^l
Jn
If such a map from ft to Mi (ft) exists and if it is S-measurable, it is called
a regular conditional probability given B. In general such a map u h-> P^
does not exist. However, it is known that for a probability space (ft, A, P)
for which ft is a complete separable metric space with Borel a-algebra A,
there exists a regular probability space for any sub a-algebra B of A.
Exercice. This exercise deals with conditional expectation.

a) What is E[y|y]?
b) Show that if E[X\A] = 0 and E[X\B] = 0, then E[X\a(A,B)] = 0.
c) Given X,Y G C1 satisfying E[X\Y] = Y and E[Y\X] = X. Verify that
X = Y almost everywhere.
We add a notation which is commonly used.
Definition. The conditional probability space (ft, ^4, P[-|S]) is defined by
P[B\B] = E[1B\B].
For X G Cp, one has the conditional moment E[X?|S] = E[XP|B] if B be a
a-subalgebra of A. They are S-measurable random variables and generalize
the usual moments. Of special interest is the conditional variance:
Definition. For X G C2, the conditional variance Var[X|S] is the random
variable E[X2|S] - E[X|Z3]2. Especially, if B is generated by a random vari
able y, one writes Var[X|y] = E[X2|y] - E[X|y]2.
Remark. Because conditional expectation is a projection, all properties
known for the usual variance hold the more general notion of conditional
variance. For example, if X, Z are independent random variables in C ,
then Var[X + Z\Y] = Var[X|y] + Var[Z|y]. One also has the identity
Var[X|y] = E[(X - E[x|y])2|y].
Lemma 3.1.5. (Law of total variance) For X G C2 and an arbitrary random

variable Y, one has
Var[X] = E[Var[X|y]] + Var[E[X|y]] .
Proof. By the definition of the conditional variance as well as the properties

of conditional expectation:
Va r [ X ] = E[X2]-E[X]2
= E[E[X2|y]]-E[E[X|y]]2
= E[Var[X|y]] + E[E[X|y]2] - E[E[X|y]]2
= E[Var[X|y]] + Var[E[X|y]] .
Here is an application which illustrates how one can use of the conditional
variance in applications: the Cantor distribution is the singular continuous
distribution with the law p has its support on the standard Cantor set.
Corollary 3.1.6. (Variance of the Cantor distribution) The standard Cantor

distribution for the Cantor set on [0,1] has the expectation 1/2 and the
variance 1/8.
Proof. Let X be a random variable with the Cantor distribution. By sym

metry, E[X] = JQ x dp(x) = 1/2. Define the a-algebra
{0, [0,1/3), [1/3,1], [0,1]}
on ft = [0,1]. It is generated by the random variable Y = l[o,i/3)- Define

Z = E[X|y]. It is a random variable which is constant 1/6 on [0,1/3)
and equal to 5/6 on [1/3,1]. It has the expectation E[Z] = (l/6)P[y =
1] + (5/6)P[y = 0] = 1/12 + 5/12 = 1/2 and the variance
Var[Z] = E[Z2} - E[Z}2 = -P[Y = 1] + ^P[y = 0] - 1/4 = 1/9 .

Define the random variable W = Var[X|y] = E[X2|y] - E[X|y]2 =
E[X2|y] - Z. It is equal to /01/3(a; - 1/6)2 dx on [0,1/3] and equal to
J2,3(x — 5/6)2 dx on [2/3,3/3]. By the self-similarity of the Cantor set, we
see that W = Var[X|y] is actually constant and equal to Var[X]/9. The
identity E[Var[X|y]] - Var[X]/9 implies
Var[X] = E[Var[X|y]] + Var[E[X|y]] = E[W] + Var[Z] = Y^l + ]- .
Solving for Va r [ X ] gives Va r [ X ] = 1/8. □
Exercice. Given a probability space (ft, .4, P) and a a-algebra B C A.

a) Show that the map P : X G C1 i—▶ E[X|S] is a Markov operator from
Cl(A, P) to CX(B, Q), where Q is the conditional probability measure on
(ft, B) defined by Q[A] = P[A] for A G B.
b) The map T can also be viewed as a map on the new probability space
(ft,B, Q), where Q is the conditional probability. Denote this new map by
S. Show that S is again measure preserving and invertible.
Exercice. a) Given a measure preserving invertible map T : ft —> ft we call

(ft, T, A, P) a dynamical system. A complex number A is called an eigen
value of T, if there exists X G C2 such that X(T) = XX. The map T is said
to have pure point spectrum, if there exists a countable set of eigenvalues
Xi such that their eigenfuctions Xi span C2. Show that if T has pure point
spectrum, then also S has pure point spectrum.
b) A measure preserving dynamical system (A, S, B, v) is called a factor of a
measure preserving dynamical system (ft, T, A, p) if there exists a measure
preserving map U : ft —▶ A such that SoJJ(x) = UoT(x) for all x G ft. Ex
amples of factors are the system itself or the trivial system (ft, S(x) = x,p).
If S is a factor of T and T is a factor of S, then the two systems are called
isomorphic. Verify that every factor of a dynamical system (ft, T, A, p) can
be realized as (ft, T, B, p) where B is a a-subalgebra of A.
c) It is known that if a measure preserving transformation T on a proba
bility space has pure point spectrum, then the system is isomorphic to a
translation on the compact Abelian group G which is the dual group of the
discrete group G formed by the spectrum cr(T) C T. Describe the possible
factors of T and their spectra.
Exercice. Let ft = T1 be the one-dimensional circle. Let A be the Borel a-

algebra on T1 = R/(27rZ) and P = dx the Lebesgue measure. Given k G N,
denote by Bk the a-algebra consisting of all A G A such that A + ^ =
A (mod 2ir) for all 1 < n < k. What is the conditional expectation E[X|Sfc]
for a random variable X G C1!
3.2. Martingales _ 131
3.2 Martingales
It is typical in probability theory is that one considers several cr-algebras on
a probability space (Cl,A,P). These algebras are often defined by a set of
random variables, especially in the case of stochastic processes. Martingales
are discrete stochastic processes which generalize the process of summing
up IID random variables. It is a powerful tool with many applications.
Definition. A sequence {An}nen of sub a-algebras of A is called a fil

tration, if Ao C Ai C • • ■ C A. Given a filtration {AJneN, one caus
(fl,A, {w4„}„€n,P) a filtered space.
Example. If Cl — {0,1}N is the space of all 0 - 1 sequences with the Borel

cr-algebra generated by the product topology and An is the finite set of
cylinder sets A — {x\ = ai,... ,xn = an} with a, e {0,1}, which contains
2n elements, then {An}neN is a filtered space.
Definition. A sequence X = {Xn}ne^ of random variables is called a dis

crete stochastic process or simply process. It is a £p-process, if each Xn
is in £p. A process is called adapted to the filtration {An} if Xn is În-
measurable for all n G N.
Example. For fl = {0,1}N as above, the process Xn(x) = FliLi x« *s

a stochastic process adapted to the filtration. Also Sn(x) = J2i=i Xi 's
adapted to the filtration.
Definition. A C1 -process which is adapted to a filtration {An} is called a

martingale if
E[X„|«4n_i] = Xn-\
for all n > 1. It is called a supermartingale if E[Xn|.4n_i] < Xn-\ and a
submartingale if E[Xn|.4n_i] > Xn-.\. If we mean either submartingale or
supermartingale (or martingale) we speak of a semimartingale.
Remark. It immediately follows that for a martingale
E \Xn \Am\ — Xm
if m < n and that E[X„] is constant. Allan Gut mentions in [34] that a
martingale is an allegory for "life" itself: the expected state of the future
given the past history is equal the present state and on average, nothing
happens.
1 IR> I m 1 ■ ! m i
!
1*J t mm I
Figure. A random variable X on the unit square defines a gray scale picture
if we interpret X(x, y) is the gray value at the point (x, y). It shows Joseph
Leo Doob (1910-2004), who developed basic martingale theory and many
applications. The partitions An = {[k/2n(k + l)/2n) x [j/2n(j + l)/2n)}
define a filtration of Q, = [0,1] x [0,1]. The sequence of pictures shows the
conditional expectations E[Xn|.4n]. It is a martingale.
Exercice. Determine from the following sequence of pictures, whether it is

a supermartingale or a submartingale. The images get brighter and brighter
in average as the resolution becomes better.
1 ",s » H H 1 'fc 1
Definition. If a martingale Xn is given with respect to a filtered space

An = <t(Yo, ... ,Yn), where Yn is a given process, X is is called a martingale
with respect Y.
Remark. The word " martingale" means a gambling system in which losing
bets are doubled. It is also the name of a part of a horse's harness or a belt
on the back of a man's coat.
Remark. If A" is a supermartingale, then -X is a submartingale and vice

versa. A supermartingale, which is also a submartingale is a martingale.
Since we can change X to X - X0 without destroying any of the martingale
properties, we could assume the process is null at 0 which means Xq = 0.
Exercice. a) Verify that if Xn,Yn are two submartingales, then sup(A, Y)

is a submartingale.
b) If Xn is a submartingale, then E[Xn] < E[An_i].
c) If Xn is a martingale, then E[Xn] = E[Xn_i].
Remark. Given a martingale. From the tower property of conditional ex-
E[Xn|A„] = E[E{Xn\An-i}\Am} = EiXn-Mm] = • • • = E[Xm}

3.2. Martingales !33
Example. Sum of independent random variables

Let Xi G C1 be a sequence of independent random variables with mean
E[Xi] = 0. Define S0 = 0,Sn = £Li Xk and An = <r{Xu • • •, Xn) with
Aq = {0,ft}. Then Sn is a martingale since Sn is an {AJ-adapted C -
process and
E[Sn\An-l] = E[Sn-l\An-l] + E[Xn| A>-i] = Sn-1 + E[Xn] = Sn_i .
We have used linearity, the independence property of the conditional ex

pectation.
Example. Conditional expectation

Given a random variable X G C1 on a filtered space (ft, A, {AJneN,P).
Then Xn = E[X\An] is a martingale.
Especially: given a sequence Yn of random variables. Then An — cr(Y0,..., Yn)
is a filtered space and Xn = E[X\Y0,..., Yn] is a martingale. Proof: by the
tower property
E[X„|Y0,...,y„-i] = E[[Xn\Y0,...,Yn}\Y0,...,Yn-1]
= E[Xn\Yo,..., yn-i] = Xn-i
We say X is a martingale with respect to Y. Note that because Xn is by

definition a(Y0,..., yn)-measurable, there exist Borel measurable functions
hn : Rn+1 -^ R such that Xn = hn(Y0,..., yn-i).
Example. Product of positive variables

Given a sequence Yn of independent random variables Yn > 0 satisfying
with E[yn] = 1. Define X0 = 1 and Xn = n?=0 y< and ^ = a^ '''' Yn">'
Then Xn is a martingale. This is an exercise. Note that the martingale
property does not follow directly by taking logarithms.
Example. Product of matrix-valued random variables

Given a sequence of independent random variables Zn with values in the
group GL(N, R) of invertible N x N matrices and let An = &(Zi, • • •, Zn)-
Assume E[log||Zn||] < 0, if \\Zn\\ denotes the norm of the matrix (the
square root of the maximal eigenvalue of Zn • Zn, where Zn is the adjoint).
Define the real-valued random variables Xn = log \\Z\ • Z2 • • • Zn\\, where •
denotes matrix multiplication. Because Xn < log \\Zn\\ + Xn-i, we get
E[Xn|^n_!] < E[l0g||Zn|| lAn-^+EiXn-ilAn-l]

= E[log||Zn||]+Xn_i<Xn_i
so that Xn is a supermartingale. In ergodic theory, such a matrix-valued

process Xn is called sub-additive.
Example. If Zn is a sequence of matrix valued random variables, we can

also look at the sequence of random variables Yn = \\Z\ • Z2 • • • Zn\\. If
E[||Zn||] = 1, then Yn is a supermartingale.
Example. Polya's urn scheme

An urn contains initially a red and a black ball. At each time n > 1, a
ball is taken randomly, its color noted, and both this ball and another
ball of the same color are placed back into the urn. Like this, after n
draws, the urn contains n + 2 balls. Define Yn as the number of black balls
after n moves and Xn = Yn/(n + 2), the fraction of black balls. We claim
that X is a martingale with respect to Y: the random variables Yn take
values in {1,..., n + 1}. Clearly P[yn+i = k + l\Yn = k] = k/(n + 2) and
P[yn+1 = k\Yn = fc] = 1 - k/(n + 2). Therefore
E[Xn+i|yi,...,yn] = -J—E\Yn+1\Y1,...,Yn]
n ] o
= —|^P[r„+i = k + i\Yn = k]- p[Yn+1]
+p[yn+1 = fc | Yn = fc] • p[yn]

1 [(yn + i)-^4-yn(l-^f-)]
n + 3 LV ' n + 2 " v n + 2
^ n
n+2
Note that Xn is not independent of Xn-\. The process "learns" in the sense
that if there are more black balls, then the winning chances are better.
oo
ooo
oooo
ooooo
oooooo
ooooooo
oooooooo
ooooooooo
oooooooooo
ooooooooooo
oooooooooooo
ooooooooooooo
F °i g u r e . A t &yrp i c a l run J of 30 o o oo ooooooooo ooooooooooo oooooo
oooo*
oooooooooooo*
experiments with Polya's urn oooooooooooo**
oooooooooooo*••
GPnPTYiP oo ooo oo oo ooo •• ••
oooooooooooo*
oooooooooooo**
ooooooooo******
oooooo**********
oooooo***********
oooooo************
oooooo*••••••••••••
ooooo***************
oooo**•••••••••••••••
oooo******************
ooo********************
Example. Branching processes

Let Zni be IID, integer-valued random variables with positive finite mean
m. Define Yq = 1 and
Yn
Yn+\ = 2_j Znk

k=l
with the convention that for Yn = 0, the sum is zero. We claim that Xn =
Yn/mn is a martingale with respect to Y. By the independence of Yn and
3.2. Martingales 135
Zni,i > 1, we have for every n
Yn Yn
E[Yn+1\Y0,...,Yn]=E[Y,Znk\Y0,...,Yn]=E[Y,Znk]=mYn
fc=i fc=i
so that
E[Xn+1|y0,..., Yn] = E[yn+1|y0,... Yn]/mn+1 = mYn/mn+l = xn.

The branching process can be used to model population growth, disease
epidemic or nuclear reactions. In the first case, think of Yn as the size of a
population at time n and with Zni the number of progenies of the i - th
member of the population, in the n'th generation.
Figure. A typical growth of Yn

of a branch process. In this ex
ample, the random variables Zni
had a Poisson distribution with
mean m = 1.1. It is possible that
the process dies out, but often, it
grows exponentially.
Proposition 3.2.1. Let An be a fixed filtered sequence of a-algebras. Lin

ear combinations of martingales over An are again martingales over An-
Submartingales and supermartingales form cones: if for example X, Y are
submartingales and a, b > 0, then aX + bY is a submartingale.
Proof. Use the linearity and positivity of the conditional expectation. □
Proposition 3.2.2. a) If X is a martingale and u is convex such that u(Xn) G

C1, then y = u(X) is a submartingale. Especially, if X is a martingale,
then \X\ is a submartingale.
b) If u is monotone and convex and X is a submartingale such that u(Xn) G
C1, then u(X) is a submartingale.
Proof, a) We have by the conditional Jensen property (3.1.4)
Yn = u(Xn) = u(E[Xn+l\An]) < E[u(Xn^)\An] = E[yn+1| \An] .
b) Use the conditional Jensen property again and the monotonicity of u to

get
Yn = u(Xn) < u(E[Xn+l\An]) < E[u(Xn+l)\An] = E[yn+1| \An] .
D
Definition. A stochastic process C = {Cn}n>i is called previsible if Cn is
*4n-i-measurable. A process X is called bounded, if Xn G £°° and if there
exists K G R such that ||Xn||oo < # for all neN.
Definition. Given a semimartingale X and a previsible process C, the pro

cess
/
CdX)n = nY,Ck(Xk-Xk.l).
k=l
It is called a discrete stochastic integral or a martingale transform.
Theorem 3.2.3 (The system can't be beaten). If C is a bounded nonnega-

tive previsible process and X is a supermartingale then / C dX is a super
martingale. The same statement is true for submartingales and martingales.
Proof. Let Y = fC dX. Prom the property of "extracting knowledge" in

theorem (3.1.4), we get
E^-y^^-l] = E[Cn(Xn-Xn_!)|A-l] = Cn-ElXn-XnÂn-l] < 0

because Cn is nonnegative and Xn is a supermartingale. □
Remark. If one wants to relax the boundedness rof C, then one has to
strengthen the condition for X. The proposition stays true, if both C and
X are £2-processes.
Remark. Here is an interpretation: if Xn represents your capital in a game,
then Xn - Xn_x are the net winnings per unit stake. If Cn is the stake on
game n, then
CdX = Y,Ck(Xk-Xk^)
/ • k=l
are the total winnings up to time n. A martingale represents a fair game
since E[Xn — Xn-i\An-i] = 0, whereas a supermartingale is a game which
is unfavorable to you. The above proposition tells that you can not find a
strategy for putting your stake to make the game fair.
Figure. In this example, Xn =

±1 with probability 1/2 and
Cn — \ if Xn-\ is even and
Cn = 0 if Xn-\ is odd. The orig
inal process Xn is a symmetric
random walk and so a martin
gale. The new process J C dX is
again a martingale.
Exercice. a) Let Y\, Y2,... be a sequence of independent non-negative ran

dom variables satisfying E[Yk] = 1 for all fc G N. Define Xq = l,Xn —
Yi"'Yn and An = o~(Yi, Y2, ...,Yn). Show that Xn is a martingale,
b) Let Zn be a sequence of independent random variables taking values in
the set of n x n matrices satisfying E[||Zn||] = 1. Define X$ = l,Xn =
\\Z\ • • • Zn\\. Show that Xn is a supermartingale.
Definition. A random variable T with values inN = NU{oo}is called

a random time. Define Aoo = cr(Un>o^)- ^ random time T is called a
stopping time with respect to a filtration An, if {T < n} G An for all
neN.
Remark. A random time T is a stopping time if and only if {T = n } G An

for all n G N since {T < n} = Uo<fc<n{r = k} e An.
Remark. Here is an interpretation: stopping times are random times, whose

occurrence can be determined without pre-knowledge of the future. The
term comes from gambling. A gambler is forced to stop to play if his capital
is zero. Whether or not you stop after the n—th game depends only on the
history up to and including the time n.
Example. First entry time.

Let Xn be a *4n-adapted process and given a Borel set B G B in Rd. Define
T(lj) = mf{n > 0 | Xn(u) G B}
which is the time of first entry of Xn into B. The set {T = oo} is the set
which never enters into B. Obviously
{T<n}= [J{XkeB}eAn
k=0
so that T is a stopping time.

Example. "Continuous Black-Jack": let Xi be IID random variables with

uniform distribution in [0,1]. Define Sn = YJk=\Xi and let T(u) be the
smallest integer so that Sn(uj) > 1. This is a stopping time. A popular
problem asks for the expectation of this random variable T: How many
"cards" Xi do we have to draw until we get busted and the sum is larger
than 1? We obviously have P[T = 1] = 0. Now, P[T = 2] = P[X2 >1-X1]
is the area of region {(x,y) G [0,1] x [0,1] | y > 1 - x } which is 1/2.
Similarly P[T = 3] = P[X3 > 1 - Xl - X2] is the volume of the solid
{(x,y,z) e [0, l]3 | z > 1 - x - y } which is 1/6 = 1/3!. Inductively we
see P[T = fc] = 1/fc! and the expectation of T is E[T] = J^=1 fc/fc! =
J2kLo l/fc! = e- This means that if we play Black-Jack with uniformly
distributed random variables and threshold 1, we expect to get busted in
more than 2, but less than 3 "cards".
Example. Last exit time. °

Assume the same setup as in 1). But this time
T(u) = sup{n > 0 | Xn(uj) e B}

is not a stopping time since it is impossible to know that X will return to
B after some time fc without knowing the whole future.
Proposition 3.2.4. Let T\,T2 be two stopping times. The infimum T\ l\T2,
the maximum T\ V T2 as well as the sum T\ + T2 are stopping times.
Proof. This is obvious from the definition because w4n-measurable functions

are closed by taking minima, maxima and sums. □
Definition. Given a stochastic process Xn which is adapted to a filtration

An and let T be a stopping time with respect to An, define the random
variable
*tm={V :^)<cc
or equivalently Xt — Y^=QXn^{T=n}- The process X% = Xr/\n is called
the stopped process. It is equal to Xt for times T <n and equal to Xn if
T>n.
Proposition 3.2.5. If X is a supermartingale and T is a stopping time, then

the stopped process XT is a supermartingale. In particular E[XT] < E[Xo].
The same statement is true if supermartingale is replaced by martingale in
which case E[XT]=E[X0].
Proof. Define the "stake process" C^ by Cn = ln<T- You can think of

it as betting 1 unit and quit playing immediately after time T. Define then
the "winning process"
( / C<T> dX)n = JT Ck(Xk - Xfc_i) = XTAn ~ X0 .

J fc=l
or shortly J C^ dX = Xt — X0. The process C is previsible, since it can

only take values 0 and 1 and {CnT) = 0} = {T < n- 1} G Ai-i- The claim
follows from the "system can't be beaten" theorem. □
Remark. It is important that we take the stopped process XT and not the
random variable Xt-
for the random walk X on Z starting at 0, let T be the stopping time
T = inf{n | Xn = 1 }• This is the martingale strategy in casino which gave
the name of these processes. As we will see later on, the random walk is
recurrent P[T < oo] = 1 in one dimensions. However
1 = E[XT] ^ E[X0] = 0 .
The above theorem gives E[XT] = E[X0].

When can we say E[Xr] = E[Xo]? The answer gives Doob's optimal stop
ping time theorem:
Theorem 3.2.6 (Doob's optimal stopping time theorem). Let X be a

supermartingale and T be a stopping time. If one of the five following
conditions are true
(i) T is bounded.
(ii) X is bounded and T is almost everywhere finite.
(hi) T e C1 and \Xn — Xn-i\ is bounded.
(iv) XT e C1 and limôo E[Xk; {T > fc}] = 0.
(v) X is uniformly integrable and T is almost everywhere finite.
thenE[XT] <E[X0].
If X is a martingale and any of the five conditions is true, then E[Xr] =
E[X0].
Proof. We know that E[Xtati — ô] < 0 because X is a supermartingale.

(i) Because T is bounded, we can take n = supT(cj) < oo and get
E[XT - X0] = E[XTAn - X0] < 0 .
(ii) Use the dominated convergence theorem to get

lim E[XTAn - X0] < 0 .
(iii) We have a bound \Xn - Xn_i| < K and so
TA n
\XTAn -X0\ = \^Xk- Xfc-i| < KT .

k=l
Because KT G C1, the result follows from the dominated convergence the
orem.
(iv) By (i), we get E[X0] = E[XTAk] = E[XT; {T < fc}] + E[Xk; {T > fc}]
and taking the limit gives E[X0] = limn-+oo*E>[Xk;{T < fc}] -▶ E[XT] by
the dominated convergence theorem and the assumption,
(v) The uniformly integrability E[\Xn\; \Xn\ > R] -> 0 for R -> oo assures
that XT e C1 since E[\XT\] < k • maxi<;<nE[|Xfc|] + supnE[|Xn|; {T >
fc}] < oo. Since \E[Xk;{T > k}]\ < supnE[|Xn|;{T > fc}] -▶ 0, we can
apply (iv).
If X is a martingale, we use the supermartingale case for both X and

-X. □
Remark. The interpretation of this result is that a fair game cannot be

made unfair by sampling it with bounded stopping times.
Theorem 3.2.7 (No winning strategy). Assume X is a martingale and sup

pose \Xn—Xn-\\ is bounded. Given a previsible process C which is bounded
and let T G Cl be a stopping time, then E[(f CdX)T] = 0.
Proof. We know that / C dX is a martingale and since (/ C dX)0 = 0, the

claim follows from the optimal stopping time theorem part (iii). □
Remark. The martingale strategy mentioned in the introduction shows

that for unbounded stopping times, there is a winning strategy. With the
martingale strategy one has T == n with probability l/2n. The player always
wins, she just has to double the bet until the coin changes sign. But it
assumes an "infinitely thick wallet". With a finite but large initial capital,
there is a very small risk to lose, but then the loss is large. You see that in
the real world: players with large capital in the stock market mostly win,
but if they lose, their loss can be huge.
Martingales can be characterized involving stopping times:

Theorem 3.2.8 (Komatsu's lemma). Let X be an ^4n-adapted sequence of

random variables in C1 such that for every bounded stopping time T
E[XT] = E[X0] ,
then X is a martingale with respect to An.
Proof. Fix neN and A e An. The map

n u e A
r = n + i-:u = { n+ 1 lj £ A
is a stopping time because cr(T) = {0,^4, Ac, ft } C An. Apply E[XT] =

E[X0] and E[Xr'] = E[X0] for the bounded constant stopping time X" =
n -f 1 to get
E[Xn;^] + E[Xn+1;^c] = E[XT] = E[X0] = E[XT,] - E[Xn+1]

= E[Xn+1;A]+E[Xn+l;Ac]
so that E[Xn+i; A] = E[Xn; A]. Since this is true, for any A G An, we know
that E[Xn+i|.4n] = E[Xn|^4n] = Xn and X is a martingale. □
Example. The gambler's ruin problem is the following question: Let Yi be

IID with P[Yi = ±1] = 1/2 and let Xn = ££=1 Y{ be the random walk
with Xo = 0. We know that X is a martingale with respect to Y. Given
a, b > 0, we define the stopping time
T = min{n > 0 | Xn = b, or Xn = — a } .
We want to compute P[Xt = -a] and P[Xt = 6] in dependence of a, b.
Figure. Three samples of a pro

cess Xn starting at Xo = 0.
The process is stopped with the
stopping time T, when Xn hits
the lower bound —a or the upper
bound b. If Xn is the winning of a
first gambler, which is the loss of
a second gambler, then T is the
time, for which one of the gam
blers is broke. The initial capital
of the first gambler is a, the ini
tial capital of the second gambler
is b.
Remark. If YJ are the outcomes of a series of fair gambles between two
players A and B and the random variables Xn are the net change in the
fortune of the gamblers after n independent games. If at the beginning, A
has fortune a and B has fortune b, then P[XT = -a] is the ruin probability
of A and P[XT = b] is the ruin probability of B.
Proposition 3.2.9.
P[XT = -a] = l-P[XT = b]

(a + b)'
Proof. T is finite almost everywhere. One can see this by the law of the
iterated logarithm,
limsup —— = 1, liminf —- = -1 .
n An n An
(We will give later a direct proof the finiteness of T, when we treat the
random walk in more detail.) It follows that P[XT = -a] = 1- P[XT = b].
We check that Xk satisfies condition (iv) in Doob's stopping time theorem:
since XT takes values in {a, b }, it is in C1 and because on the set {T > k },
the value of Xk is in (-a, b), we have \E[Xk; {T > k }]| < max{a, b}P[T >
fc] - o. a
Remark. The boundedness of T is necessary in Doob's stopping time the

orem. Let T = inf{n | Xn = 1 }. Then E[XT] = 1 but E[X0] = 0] which
shows that some condition on T or X has to be imposed. This fact leads
to the "martingale" gambling strategy defined by doubling the bet when
loosing. If the casinos would not impose a bound on the possible inputs,
this gambling strategy would lead to wins. But you have to go there with
enough money. One can see it also like this, If you are A and the casino is
B and b = 1, a = oo then P[XT — b] = 1, which means that the casino is
ruined with probability 1.
Theorem 3.2.10 (Wald's identity). Assume T is a stopping time of a C1-

process Y for which Yi are IID random variables with expectation E[Yi] = m
and T eC1. The process Sn = £Li Yk satisfies
E[ST] = mE[T] .
3.3. Doob's convergence theorem 143
Proof. The process Xn = Sn - nE[yi] is a martingale satisfying condition

(iii) in Doob's stopping time theorem. Therefore
0 = E[X0] = E[XT] = E[5T - TE[YX]] .
Now solve for E[Sr]. D
3.3 Doob's convergence theorem

Definition. Given a stochastic process X and two real numbers a < b, we
define the random variable
Un[a, b](u) = max{fc G N | 3

0 < si < t\ < • • - < sk < tk < n,
XSi(uj) <a,Xti(uo) >b,l<i<k}
called the number of up-crossings of [a, b]. Denote with [Too [a, b] the limit
Uoo[a,b] = n—»>oo
lim Un[a,b] .
Because n \—> Un[a, b] is monotone, this limit exists in N U {oo}.
Figure. A random walk crossing

two values a < b. An up-crossing
is a time s, where Xs < a un
til the time, when the first time
Xt > b. The random variable
Un[a,b] with values in N mea
sures the number of up-crossings
in the time interval [0,n].
Theorem 3.3.1 (Doob's up-crossing inequality). If X is a supermartingale.

Then
(b-a)E[Un[a,b]]<E[(Xn-a)-].
Proof Define C\ — l{x0<a} and inductively for n > 2 the process
Cn := l{Cn_i = l}l{Xn_!<6 } + l{Cn_i=0}l{Xn_i<a } •

It is previsible. Define the winning process Y = J C dX which satisfies by
definition Y0 = 0. We have the winning inequality
Yn(uj) >(b- a)Un[a, b]{u) - (Xn{u) - a)~ .
Every up-crossing of [a, b] increases the Y-value (the winning) by at least

(b - a), while (Xn - a)~ is essentially the loss during the last interval of
play-
Since C is previsible, bounded and nonnegative, we know that Yn is also a
supermartingale (see "the system can't be beaten") and we have therefore
E[yJ < 0. Taking expectation of the winning inequality, gives the claim.
□
Remark. The proof uses the following strategy for putting your stakes C:
wait until X gets below a. Play then unit stakes until X gets above b and
stop playing. Wait again until X gets below a, etc.
Definition. We say, a stochastic process Xn is bounded in Cp, if there exists
Mel such that ||Xn||p < M for all neN.
Corollary 3.3.2. If X is a supermartingale which is bounded in C . Then
P[tfoo[M] = Oo]=0.
Proof. By the up-crossing lemma, we have for each n G N
(b - a)E[Un[a, b]] < \a\ + E[|X„|] < \a\ + sup

n ||Xn||i < oo .
By the dominated convergence theorem
(b-a)E[Uoo[a,b]]<oo,
which gives the claim. ^
Remark. If Sn = Y2=i Xk is tne one dimensional random walk, then it is

a martingale which is unbounded in C1. In this case, E[Uoo[a, b]] = oo.
Theorem 3.3.3 (Doob's convergence theorem). Let Xn be a supermartingale

which is bounded in C1. Then
Xoo = n—*oo
lim Xn
exists almost everywhere.

Proof.
A = {uj e ft I Xn has no limit in [—00,00] }
= {uj e ft I lim inf Xn < lim sup Xn }
= I) {uj G ft I lim inf Xn < a < b < lim sup Xn }
a<b,a,beQ
U A°<»-
a<b,a,b£Q
Since Aa,b C {Uoo[a, b] = 00 } we have P[Aa?b] = 0 and therefore also

P[A] = 0. Therefore Xqo = limnôoXn exists almost surely. By Fatou's
lemma
E[|Xoo|] = Epiminf
n — ▶\Xn\)
0 0 < liminf
n — E[|Xn|]
> o o < supE[|Xn|]
n < 00
so that P[Xoo < 00] = 1. □
Example. Let X be a random variable on ([0,1), A, P), where P is the

Lebesgue measure. The finite a-algebra An generated by the intervals
Ak -- r l2n,
— k 2n
+ 1^)
defines a filtration and Xn = E[X|^4n] is a martingale which converges. We

will see below with Levys upward theorem (3.4.2 that the limit actually is
the random variable X.
Example. Let Xk be IID random variables in C1. For 0 < A < 1, the
branching random walk Sn = ]C/c=o ^k^k is a martingale which is bounded
in C1 because
||S„||l < j-^HXolk .
The martingale converges by Doob's convergence theorem almost surely.
One can also deduce this from Kolmogorov's theorem (2.11.3) if Xk G C2.
Doobs convergence theorem (3.3.3) assures convergence for Xk G C1.
Remark. Of course, we can replace supermartingale by submartingale or

martingale in the theorem.
Example. We look again at Polya's urn scheme, which was defined earlier.
Since the process Y giving the fraction of black balls is a martingale and
bounded 0 < Y < 1, we can apply the convergence theorem: there exists
y'oo with Yn^Y^.
Corollary 3.3.4. If X is a non-negative supermartingale, then X^ =

limn_>00 Xn exists almost everywhere and is finite.
Proof Since the supermartingale property gives E[\Xn\] = E[Xn] < E[X0],
the process Xn is bounded in C1. Apply Doob's convergence theorem. □
Remark. This corollary is also true for non-positive submartingales or mar

tingales, which are either nonnegative or non-positive.
Example. For the Branching process, we had IID random variables Zni
with positive finite mean m and defined Y0 = 0, Yn+1 = Ylk=i Znk- We
saw that the process Xn = Yn/mn is non-negative and a martingale. Ac
cording to the above corollary, the limit Xoo exists almost everywhere. It
is an interesting problem to find the distribution of X^: Assume ZUi have
the generating function f(6) = E[9Zni].
(i) Yn has the generating function fn(6) = f(fn~1)(9).

We prove this by induction. For n = 1 this is trivial. Using the independence
of Znk we have
E[0Y^\Yn = k] = f(6)k
and so
E[9Y^\Yn] = f(e)z«.
By the tower property, this leads to
E[0y"+'] = E[/(0)z"] .
Write a — f{6) and use induction to simplify the right hand side to
E[f{6)Y"\ = E[ay"] = /» = p{f{9)) = fn+\6) .

(ii) In order to find the distribution of .X^ we calculate instead the char
acteristic function
L(X) = L(Xoo)(A) = EfexpÂXoo)] .

Since X„ —▶ X^ almost everywhere, we have L(Xn)(\) —▶ L(Xoo)(X).
Since Xn = Yn/mn and E[0Y"] = fn(6), we have
L{Xn){\) = fn{eiXlmn)
so that L satisfies the functional equation
L(Xm) = f(L(\)) .
Theorem 3.3.5 (Limit distribution of the branching process). For the

branching process defined by IID random variables Zn% having the gen
erating function /, the Fourier transform L(X) = E[eiXX°°] of the distribu
tion of the limit martingale Xoo can be computed by solving the functional
equation
L(X-m) = f(L(X)) .
Remark. If / has no analytic extension to the complex plane, we have to
replace the Fourier transform with the Laplace transform
L(X) = E[e-AX~] .
Remark. Related to Doob's convergence theorem for supermartingales is
Kingman's subadditive ergodic theorem, which generalizes Birkhoff's er
godic theorem and which we state without proof. Neither of the two theo
rems are however corollaries of each other.
Definition. A sequence of random variables Xn is called subadditive with
respect to a measure preserving transformation T, if Xm+n < Xrn-\-Xn(Trn)
almost everywhere.
Theorem 3.3.6 (The subadditive ergodic theorem of Kingmann). Given a

sequence of random variables, which Xn : X —▶ RU{-oo} with X+ :=
max(0,Xn) G Ll(X) and which is subadditive with respect to a measure
preserving transformation T. Then there exists a T-invariant integrable
measurable function X : ft -> R U {-oo} such that \Xn(x) -> X(x) for
almost all x G X. Furthermore Ê[Xn] -> E[X].
If the condition of boundedness of the process in Doob's convergence the

orem is strengthened a bit by assuming that Xn is uniformly integrable,
then one can reverse in some sense the convergence theorem:
Theorem 3.3.7 (Doob's convergence theorem for uniformly integrable su

permartingales). A supermartingale Xn is uniformly integrable if and only
if there exists X such that Xn —> X in C1.
Proof. If Xn is uniformly integrable, then Xn is bounded in Cl and Doob's

convergence theorem gives Xn —» X almost everywhere. But a uniformly
integrable family Xn which converges almost everywhere converges in C .
On the other hand, a sequence Xn G C1 converging to X G C1 is uniformly
integrable. D
Theorem 3.3.8 (Characterization of uniformly integrable martingales). An

*4n-adapted process is an uniformly integrable martingale if and only if
Xn -> X in C1 and Xn = E[X\An]>
Proof. By Doob's convergence theorem for uniformly integrable supermartin-

gales (3.3.7), we know the "if'-part. We already know that Xn = E[X\An]
is a martingale. What we have to show is that it is uniformly integrable.
Given e > 0. Choose S > 0 such that for all A G A, the condition P[A] < S
implies E[|X|; A] < e. Choose further K G R such that K~l • E[\X\] < S.
By Jensen's inequality
\Xn\ = \E[X\An]\ < E[\X\\An] < E[\X\] .

Therefore
K • P[|Xn| > K] < E[\Xn\] < E[\X\] <6K
so that P[|Xn| > K] < S. By definition of conditional expectation , \Xn\ <
E[\X\\An]tmd{\Xn\>K}eAn
E[\Xn\; \Xn\ >K]< E[|X|; |Xn| > K] < e .

D
Remark. As a summary we can say that supermartingale Xn which is either

bounded in C1 or nonnegative or uniformly integrable converges almost
everywhere.
Exercice. Let S and T be stopping times satisfying S <T.

a) Show that the process
Cn(uj) = l{S(u/)<n<T(u/)}
is previsible.
b) Show that for every supermartingale X and stopping times S < T the
inequality
E[XT] < E[XS]
holds.
Exercice. In Polya's urn process, let Yn be the number of black balls after
n steps. Let Xn = Yn/(n + 2) be the fraction of black balls. We have seen
that X is a martingale.
a) Prove that P[yn = fc] = l/(n + 1) for every 1 < fc < n + 1.
b) Compute the distribution of the limit Xqo .
Exercice. a) Which polynomials / can you realize as generating functions

of a probability distribution? Denote this class of polynomials with V.
b) Design a martingale Xn, where the iteration of polynomials P eV plays
a role.
c) Use one of the consequences of Doob's convergence theorem to show

that the dynamics of every polynomial P e V on the positive axis can be
conjugated to a linear map T : z •-> mz: there exists a map L such that
LoT(z) = PoL(z)
for every z G
Example. The branching process yn+i = YX=\ Znk defined by random

variables Znk having generating function / and mean m defines a mar
tingale Xn = Yn/mn. We have seen that the Laplace transform L(X) =
E[e~xx°°] of the limit Xoo satisfies the functional equation
L(m\) = f(L(X)) .
We assume that the IID randoril variables Znk have the geometric distribu
tion P[Z = fc] = p(l -p)k = pqk with parameter 0 )=E[0Z]=XX0* 1-qO
fc=l
As we have seen in proposition (2.12.5),

oo
E[Z] = J>fefc = V
k=\
The function fn(0) can be computed as
J [ }= qpmn(l-0)
mn(l-6)++ q qO-p
e-p '
This is because / is a Mobius transformation and iterating / corresponds

0 P
to look at the power An . This power can be computed by
1
diagonalisating A:
An = (q-p) - 1 Q
0 qn -1
We get therefore
pX-\-q-p
L(X) = E[e-xx~] = lim E[e"Ay"/mn] - lim fn(ex/mU) =
qX + q-p
If m < 1, then the law of X(oo is a Dirac mass at 0. This means that the
process dies out. We see that in this case directly that limn_*oc fn(Q) = 1. In
the case m > 1, the law of Xoo has a point mass at 0 of weight p/q = 1/ra
and an absolutely continuous part (1/ra - l)2e(1/m_1)x dx. This can be
seen by performing a "look up" in a table of Laplace transforms
/•oo
L(X) = V*° + / {l-p/q)2eWq-Vx-e-Xxdx

Jo
Definition. Define pn = P[Yn = 0], the probability that the process dies
out until time n. Since pn = fn(0) we have pn+\ = f(pn)- If f(p) = p, p is
called the extinction probability.
Proposition 3.3.9. For a branching process with E[Z] > 1, the extinction
probability is the unique solution of f(x) = x in (0,1). For E[Z] < 1, the
extinction probability is 1.
Proof The generating function f(9) = E[6Z] = J2n=oFiZ = n\6n =

^2nPn@n is analytic in [0,1]. It is nondecreasing and satisfies /(l) = 1.
If we assume that P[Z = 0] > 0, then /(0) > 0 and there exists a unique
solution of f(x) = x satisfying f'(x) < 1. The orbit fn(u) converges to
this fixed point for every u G (0,1) and this fixed point is the extinction
probability of the process. The value of /'(0) = E[Z] decides whether there
exists an attracting fixed point in the interval (0,1) or not. □
3.4 Levy's upward and downward theorems
Lemma 3.4.1. Given X e C1. Then the class of random variables
{y = E[X\B] | B C A, B is a - algebra }
is uniformly integrable.
Proof. Given e > 0. Choose 6 > 0 such that for all A G A, P[A] < S
implies E[\X\; A] < e. Choose further K G R such that K~l • E[|X|] < S.
By Jensen's inequality, Y = E[X\B] satisfies
\Y\ = \.E\X\B]\ <E[\X\\B] <E[\X\] .
Therefore
K • P[|Xn| > K] < E[\Y\] < E[\X\] <6-K
so that P[|y| > K] < 6. Now, by definition of conditional expectation ,
|y|<E[|X||£]and{|y|>X}GS
E[|XB|; \XB\ >K]< E[|X|; \XB\ > K] < e .

□
Definition. Denote by Aoo the cr-algebra generated by (Jn An>

3.4. Levy's upward and downward theorems 151
Theorem 3.4.2 (Levy's upward theorem). Given X G C . Then Xn =

E[X\An] is a uniformly integrable martingale and Xn converges in C1 to
XQO=E[X\Aoo\.
Proof. The process X is a martingale. The sequence Xn is uniformly in

tegrable by the above lemma. Therefore Xoo exists almost everywhere by
Doob's convergence theorem for uniformly integrable martingales, and since
the family Xn is uniformly integrable, the convergence is in C1. We have
to show that Xoo = Y := E[X|A»].
By proving the claim for the positive and negative part, we can assume
that X > 0 (and so Y > 0). Consider the two measures
Q1(A)=E[X-,A], Q2(A) = E[X00;A] .

Since E^^lAi] = E[X|*4n], we know that Qi and Q2 agree on the n-
system \JnAn- They agree therefore everywhere on Aoo- Define the event
A = {E[X\Aoo] > Xoo } G Aoo- Since QX(A) - Q2(A) = E[E[X\Aco] ~
Aoo]; A] = 0 we have E[X|^4oo] < Xoo almost everywhere. Similarly also
-ôo < A|>too] almost everywhere. □
As an application, we see a martingale proof of Kolmogorov's 0 — 1 law:
Corollary 3.4.3. For any sequence An of independent cr-algebras, the tail
cr-algebra T = f]nBn with Bn = \Jm>n Am is trivial.
Proof. Given A G T, define X = 1A G C°°(T) and the a-algebras Cn =
cr(A\,... ,An)- By Levy's upward theorem (3.4.2),
X = E[X|Coo]= lim E[X\Cn] .
—▶o
n
But since Cn is independent of An and X is Cn measurable, we have
P[A]=E[X] = E[X\Cn]-+X
and because X takes only the value 0 or 1 and X = P[A] shows that it
must be constant, we get P[A] — 1 or P[A] = 0. □
Definition. A sequence A-n of cr-algebras A-n satisfying
• • • c A-n c A_(n_i) c • • • c A-i
is called a downward filtration. Define *4-oo = Dn^-™-
Theorem 3.4.4 (Levy's downward theorem). Given a downward filtration

A-n and X G C1. Define X-n = E[X\A-n]. Then X-^ = limnôoX_n
converges in C1 and X-oo = E[X\A-oo\.
Proof. Apply Doob's up-crossing lemma to the uniformly integrable mar

tingale
Xk,-n<k< -1 :
for all a < b, the number of up-crossings is bounded
£/fc[a,6]<(|a| + ||X||i)/(6-a).
This implies in the same way as in the proof of Doob's convergence theorem
that limn-ôo Y-n converges almost everywhere.
We show now that X_oo = E[X|,4_oo]: given A G *4-oo- We have E[X; A] =
E[X-n] A] —> E[X_oo; A]. The same argument as before shows that X_oo =
E[X;A-oo]. □
Lets also look at a martingale proof of the strong law of large numbers.
Corollary 3.4.5. Given Xn G C1 which are IID and have mean ra. Then
Sn/n —> Tn in C1.
Proof. Define the downward filtration A~n = o~(Sn, Sn+i, • • • )•

SmceE[X1\A-n]=E[Xi\A-n]=E[Xi\Sn,Sn+l,...] = Xi,andE[X1\An] =
Sn/n. We can apply Levy's downward theorem to see that Sn/n converges
in C1. Since the limit X is in T, it is by Kolmogorov's 0-1 law a constant
c and c = E[X] = limn_>oo E[Sn/n] = ra. □
3.5 Doob's decomposition of a stochastic process

Definition. A process Xn is increasing, if P[Xn < Xn+i] = 1.
Theorem 3.5.1 (Doob's decomposition). Let Xn be an >tn-adapted C1-

process. Then
X = X0 -r N + A
where AT is a martingale null at 0 and A is a previsible process null at 0.
This decomposition is unique in L1. X is a submartingale if and only if A
is increasing.
Proof. If X has a Doob decomposition X = X0 + N + A, then
E[Xn-Xn_i|.An_i] = E[N„-JVn-l|Ai]4-E[i4n-An-l|Ai-l] = An-An-i

3.5. Doob's decomposition of a stochastic process 153
which means that
n
An = Y,E[Xk-Xk-1\An-i]-
fc=i
If we define A like this, we get the required decomposition and the sub
martingale characterization is also obvious. D
Remark. The corresponding result for continuous time processes is deeper

and called Doob-Meyer decomposition theorem. See theorem (4.17.2).
Lemma 3.5.2. Given s,t,u,v G N with s < t < u < v. If Xn is a C -

martingale, then
E[(Xt-Xs)(Xv-Xu)]=0
and
E[Xl\ = E[Xl] + J2 E[(Xfc - **-i)2] •
Proof. Because E[XV - XU\AU] = Xu - Xu = 0, we know that Xv - Xu

is orthogonal to C2(AU)- The first claim follows since Xt - Xx G C2(AU)-
The formula
Xn = Xq -h 2^(Xk - Xk-l)
fc=l
expresses Xn as a sum of orthogonal terms and Pythagoras theorem gives

the second claim. □
Corollary 3.5.3. A £2-martingale X is bounded in £2 if and only if

EîE[(Xfc-Xfc_1)2]<oo.
Proof.
n oo
E[Xl\ = ElXft+ÊKXk-Xk^)2} < EiXft+Y^rnXk-X^)2} < oo .

fc=l k=l
If on the other hand, Xn is bounded in £2, then ||Xn||2 < K < oo and
^[(Xk-Xk-^^K + EiX*]. D
Theorem 3.5.4 (Doob's convergence theorem for L2-martingales). Let Xn

be a £2-martingale which is bounded in C2, then there exists X e C2 such
that Xn -» X in C2.
Proof. If X is bounded in C2, then, by monotonicity of the norm ||X||i <

||X||2, it is bounded in C1 so that by Doob's convergence theorem, Xn —▶ X
almost everywhere for some X. By Pythagoras and the previous corol
lary (3.5.3), we have
E[(X-Xn)2}< £ E[(Xk - Xfc_i)2] - 0

k>n+l
so that Xn -> X in £2. □
Definition. Let Xn be a martingale in C2 which is null at 0. The conditional

Jensen's inequality (3.1.4) shows that X2 is a submartingale. Doob's de
composition theorem allows to write X2 = N + A, where N is a martingale
and A is a previsible increasing process. Define A^ = limn_ôo An point
wise, where the limit can take the value oo also. One writes also (X) for A
so that
X2 = N + (X) .
Lemma 3.5.5. Assume X is a C2-martingale. X is bounded in C2 if and

onlyifE[(X)00]<oo.
Proof. From X2 = N + A, we get E[X2] = E[An] since for a martingale N,

the equality E[Nn] = E[iVo] holds and AT is null at 0. Therefore, X is in C2
if and only if E[Aoo] < oo since = E[X2] = E[An] and An is increasing. □
We can now relate the convergence of Xn to the finiteness of Aoo = (X)^.
Proposition 3.5.6. a) If limn_>00 Xn(uj) converges, then Aoo(uj) < oo.

b) On the other hand, if \\Xn — Xn_i||oo < K, then Aûj) < oo implies
the convergence of limn_oo Xn(uj).
3.5. Doob's decomposition of a stochastic process 155
Proof, a) Because A is previsible, we can define for every fc a stopping time

S(k) = inf{n G N | An+i > k }. The stopped process As^ is previsible
because for B e B^ and neN,
{AnAS(k) eB} = A1uA2

with
n—i
M = \J{S(k) = i',AieB}eAn-i
A2 = {An eB}n {S(k) < n - 1}C G An-i •
Since
(Xs(kh2 - ASk = (X2 - A)s{-k)
is a martingale, we see that (Xs(k) — As^. The later process As^ is
bounded by fc so that by the above lemma Xs^ is bounded in C2. There
fore limn_>oo XnAs(k) exists almost surely. Combining this with
{Aoo < oo} = [j{Sk = oo}

k
proves the claim.

b) Suppose the claim is wrong and that
P[Aoo = 00,SUp|Xn| < OO] > 0 .

n
Then,
P[T(c) = oo;A» = oo] >0
where T(c) is the stopping time
T(c) = inf{n \ \Xn\ > c] .

Now
ElXT(c)An ~ AT(c)An] = 0
and XT^ is bounded by c + K. Thus
E[AT{c)An]<(c + K)2
for all n. This is a contradiction to P[Ax> = oo, supn \Xn\ < oo] > 0. D
Theorem 3.5.7 (A strong law for martingales). Let X be a £2-martingale

zero at 0 and let A = (X). Then
An
almost surely on {A^ = oo }.

Proof, (i) Cesaro's lemma: Given 0 = bo < h < ... ,bn < bn+i —> oo and a
sequence vn G R which converges vn —▶ ôo, then ^- $^fc=i(&fc — bk-i)vk —>
Proof. Let e > 0. Choose ra such that ^ > i>oo - e if fc > m. Then
^ n ^ rn
lim inf — Y^(6fc - bk-i)vk > lim inf — Y\&fc - bk-i)vk
+—r (ôo -e)

On
> 0 + Voo - C
Since this is true for every e > 0, we have lim inf > ôo- By a similar
argument lim sup > Vqo- '-'
(ii) Kronecker's lemma: Given 0 = b0 < h < ..., bn < 6n+i -> oo and a se
quence xn of real numbers. Define sn = xx + • • • + xn. Then the convergence
of un = Ylk=i xk/bk implies that sn/bn -> 0.
Proof. We have un - wn_i = xn/6n and

n w
Sn=Ylbk(Uk~Uk-1^=bnUn~X^fc"&fc-1)Wfc-1*
Cesaro's lemma (i) implies that sn/bn converges to ôo - ôo = 0. □
(iii) Proof of the claim: since A is increasing and null at 0, we have An > 0
and 1/(1+An) is bounded. Since A is previsible, also 1/(1+An) is previsible,
we can define the martingale
Moreover, since (1 4- An) is An-\-measurable, we have
E[(Wn-Wn-l)2\An-l] = (1+An)'2(An~An-l) < (1+An-l)"1 "(l+^n)"1
almost surely. This implies that (W)^ < 1 so that limn_oo Wn exists
almost surely. Kronecker's lemma (ii) applied point wise implies that on
{Aoo = 00}
lim Xn/(l + An) = lim Xn/An -+ 0 .
n—>oo n—>oo
D
3.6. Doob's submartingale inequality 157
3.6 Doob's submartingale inequality
Theorem 3.6.1 (Doob's submartingale inequality). Let X be a non-negative

submartingale. For any e > 0
e • P[ sup Xk > e] < E[Xn', { sup Xk > e}] < E[Xn] .

l<fc<n Kk<n
Proof. The set A = {sup1<fc<n Xk > e} is a disjoint union of the sets
A0 = {X0 >e}eA0
n
Ak = {Xk>e}n({jA1)€Ak.
Since X is a submartingale, and Xk > e on Ak we have for fc < n
E[Xn;Ak]>E[Xk',Ak]>eP[Ak].
Summing up from fc = 0 to n gives the result. □

We have seen the following result already as part of theorem (2.11.1). Here
it appears as a special case of the submartingale inequality:
Theorem 3.6.2 (Kolmogorov's inequality). Given Xn G C2 IID with

E[Xi] = 0 and Sn = £Li Xk. Then for e > 0,
P [K.ks<un p€ * \ S k \ > e ] < ? ^ .
Proof. Sn is a martingale with respect to An = cf(X1 ,X2,..., Xn). Because

u(x) = x2 is convex, S2 is a submartingale. Now apply the submartingale
inequality (3.6.1). □
Here is an other proof of the law of iterated logarithm for independent
iV(0,1) random variables.
Theorem 3.6.3 (Special case of law of iterated logarithm). Given Xn IID

with standard normal distribution JV(0,1). Then limsup^^ Sn/A(n) = 1.
Proof. We will use for

/•oo poo
1 - *(*) = / 4>(y) dy = / (27T)-1/2 exp(-j/2/2) dy
Jx Jx
the elementary estimates
(x + x"1)-1^) < 1 - *(s) < x"V(a?) •
(i) 5n is a martingale relative to An = <r(Xi,..., Xn). The function x h->

e61* is convex on R so that e^5n is a submartingale. The submartingale
inequality (3.6.1) gives
P[ sup Sk > e] = P[ sup eeSk > e6e] < e-0eE[e0S»] = eê**"1'2 .

l<fc<n l<fc<n
For given e > 0, we get the best estimate for 6 = e/n and obtain
P[ sup Sk > e] < e"e2/(2n) .

l<k<n
(ii) Given K>\ (close to 1). Choose en = KK(Kn-1). The last inequality
in (i) gives
P[ sup Sk>en]<exp(-e2J(2Kn)) = (n-l)-K(logK)-K .

l<k<Kn
The Borel-Cantelli lemma assures that for large enough n and Kn~l < k <
Kn
Sk 1 almost surely
limsuP777X -K '
fcôo A(fc)
By taking a sequence of K's converging down to 1, we obtain almost surely
limsup —- < 1 .
fcôo A(fc)
(iii) Given N > 1 (large) and S > 0 (small). Define the independent sets
An = {S(Nn+1) - S(Nn) > (1 - 8)A(Nn+l - Nn)} .
Then
P[An] = 1 - 9(y) = (27r)-1/2(y + y-i)-ic-va/2
with y = (1 -5)(21oglog(iVn-1 - N"))1'2. Since P[An] is up to logarithmic
terms equal to (nlogiV)-^1-^2, we have ]Cnp[î] = °°- Borel-Cantelli
shows that P[limsupn An] = 1 so that
S(Nn^) > (1- <J)A(JVn+1 - Nn) + S(Nn) .

3.7. Doob's Cp inequality 159
By (ii), S(Nn) > -2A(Nn) for large n so that for infinitely many n, we
have
S(Nn+1) > (1 - S)A(Nn+1 - Nn) - 2A(Nn) .
It follows that
Sn . r S(Nn^) 1 1/2 1/2

limsup j- > limsup ^^^y > (1 - *)(1 - ^) - 2iV .
3.7 Doob's Cp inequality
Lemma 3.7.1. (Corollary of Holder inequality) Fix p > 1 and q satisfying

p-1 + q-1 = 1. Given X,Y e Cp satisfying
eP[|X|>e]<E[|y|;|X|>e]
Ve>0, then \\X\\P < q • \\Y\\p.
Proof. Integrating the assumption multiplied with pep 2 gives

/•OO /"OO
L = / pcP-ÔXl > e] de < / pep-2E[\Y\; \X\ > e] de =: R .

Jo Jo
By Fubini's theorem, the the left hand side is
/•oo /»oo
L= E[pC"-1l{|X|>e}](fc = E[/ pC"-1l{|X|>e}dc]=E[|X|"].

7o Jo
Similarly, the right hand side is R = E[q • IX^"1^!]. With Holder's in

equality, we get
E[\xn < EteixMYi] < q\\Y\\p • \\\xr%.

Since (p - l)q = p, we can substitute IHXI"-1!!, = EOXIp]1/* on the right
hand side, which gives the claim. □
Theorem 3.7.2 (Doob's Lp inequality). Given a non-negative submartingale

X which is bounded in Cp. Then X* = supn Xn is in Cp and satisfies
||X*||<g-sup||Xn||p.
Proof. Define X* = sup^^X* for neN. From Doob's submartingale

inequality (3.6.1) and the above lemma (3.7.1), we see that
\\X*\\P < q\\Xn\\p < qs\ip\\Xn\\P .
Corollary 3.7.3. Given a non-negative submartingale X which is bounded

in Cp. Then X^ = limôo Xn exists in Cp and ||Xoo||P = supn ||Xn||p .
Proof. The submartingale X is dominated by the element X* in the Cp-

inequality. The supermartingale -X is bounded in Cp and so bounded in
C1. We know therefore that X^ = lininôoXn exists almost everywhere.
From \Xn - X^ < (2X*)P e Cp and the dominated convergence theorem
we deduce Xn —▶ Xoc in Cp. D
Corollary 3.7.4. Given a martingale Y bounded in Cp and X = \Y\. Then
Xqo = nlim
— ▶o Xn
exists in Cp and HX^H^ = supn \\Xn\
Proof. Use the above corollary for the submartingale X = |F|. □
Theorem 3.7.5 (Kakutani's theorem). Let Xn be a non-negative indepen

dent C1 process with E[Xn] = 1 for all n. Define 50 = 1 and Sn = JlLi x*-
Then Sqq = limn 5n exists, because Sn is a nonnegative £\ martingale.
Then 5n is uniformly integrable if and only if H™=1 E[X*/2] > 0.
Proof. Define an = E[xl/2]. The process
Yl/2 Yl/2 „l/2

T
J-n —— *1 * 2 ' n
ai a2 an
3.7. Doob's Cp inequality 161
is a martingale. We have E[T2] = (cl\cl2 • • • an)~2 < (Y\n an)_1 < oo so that
T is bounded in C2, By Doob's £2-inequality
E[sup|5n|] < E[sup|Tn|2] < 4supE[|Tn|2] < oo

n n n
so that S is dominated by S* = supn |5n| G C1. This implies that S is

uniformly integrable.
If Sn is uniformly integrable, then Sn —▶ Sex, in C1. We have to show that

n^Li an > 0. Aiming to a contradiction, we assume that Y[n &n = 0. The
martingale T defined above is a nonnegative martingale which has a limit
T^. But since Yln fln = 0 we must then have that S^ = 0 and so Sn —> 0
in C1. This is not possible because E[5n] = 1 by the independence of the
Xn. D
Here are examples, where martingales occur in applications:
Example. This example is a primitive model for the Stock and Bond mar
ket. Given a < r < b < oo real numbers. Define p = (r - a)/(b - a). Let en
be IID random variables taking values 1,-1 with probability p respectively
1-p. Define a process Bn (bonds with fixed interest rate /) and Sn (stocks
with fluctuating interest rates) by
Bn = (l+r)nBn-i,B0 = l •
Sn = (1 + Rn)Sn-i, So = 1
with Rn = (a + b)/2 + en(a - b)/2. Given a sequence An (the portfolio),
your fortune is Xn and satisfies
Xn = (1 + r)Xn_! + AnSn-l(Rn - r) .
We can write Rn - r = \(b - a)(Zn - Zn-i) with the martingale
n
Zn = ^(ek-2p+l).
fc=i
The process Yn = (1 + r)~nXn satisfies then
yn-yn_! = (l + r)-nA,5„_i(iin-r)
= Itb-aXl+ryAnSn-iiZn-Zn-!)
= Cn(Zn — Zn-i)
showing us that Y is the stochastic integral / C dZ. So, if your portfolio

An is previsible (An-i measurable), then Y is a martingale.
Example. Let X, Xi,X2... be independent random variables satisfying

that the law of X is JV(0, (J2) and the law of Xk is N(0,0%). We define the
random variables
Yk=X-rXk
which we consider as a noisy observation of the random variable X. Define
An = <j(Xi, ...,Xn) and the martingale
Mn = E[X\An] •
By Doob's martingale convergence theorem (3.5.4), we know that Mn con

verges in C2 to a random variable M^. One can show that
E[(X-M„)2] = ((r-2 + ^(Tfe-2)-1.

Jfe=l
This implies that X = M^ if and only if ^n o~2 = oo. If the noise grows
too much, for example for o~n = n, then we can not recover X from the
observations Yk.
3.8 Random walks

Consider the d-dimensional lattice Zd, where each point has 2d neighbors.
A particle starts at the origin 0 G Zd and makes in each time step a random
step into one of the 2d directions. What is the probability that the particle
returns back to the origin?
Definition. Define a sequence of IID random variables Xn which take values

in
d
I = {eeZd\\e\ = '£\ei\ = l}
1=1
and which have the uniform distribution defined by P[Xn = e] = (2d)"1

for all eel. The random variable Sn = Y17=i î with So = 0 describes
the position of the particle at time n. The discrete stochastic process Sn is
called the random walk on the lattice Zd.
Figure. A random walk sample

path Si(uj), ..., Sn(uj) in the lat
tice 1? after 2000 steps. Bn(uj)
is the number of revisits of the
starting points 0.
As a probability space, we can take fi, = IN with product measure z/N,

where v is the measure on E, which assigns to each point e the probability
3.8. Random walks 163
v({e}) = (2d)"1. The random variables Xn are then defined by Xn(oj) =

ujn. Define the sets An = {Sn = 0 } and the random variables
Yn = lAn •
if the particle has returned to position 0 G Zd at time n, then Yn = 1,

otherwise Fn = 0. The sum £n = ELo1* counts the number of visits °f
the origin 0 of the particle up to time n and B = £felo Yk counts the total
number of visits at the origin. The expectation
oo
E[B] = £P[Sn = 0]
71=0
tells us how many times a particle is expected to return to the origin. We

write E[JB] = oo, if the sum diverges. In this case, the particle returns back
to the origin infinitely many times.
Theorem 3.8.1 (Polya). E[B] = oo for d = 1,2 and E[B] < oo for d > 2.
Proof. Fix n G N and define a^n\k) = P[Sn = fc] for fc G Zd. Because
the particle can reach in time n only a bounded region, the function a(n) :
Zd -> R is zero outside a bounded set. We can therefore define its Fourier
transform
<^(*) = Ea(n)(fe)e27rifcx'
kezd
which is smooth function on Td = Md/Zd. It is the characteristic function
of Sn because
E[eteS»] = Y, P(5» = W^ ■
kezd
The characteristic function <\>x of Xk is
iji=i f=i
Because the Sn is a sum of n independent random variables Xj
1 d
<t>sn = </>*, (*)</>x2 (a:)... 0xn (*) = — (£ cos(27rx0)n .
2=1
Note that 05n(O) = P[Sn = 0].
We now show that E[B] = En>o <^n(0) is finite if and only if d < 3- The
Fourier inversion formula gives
£p[s„ = o] - / £«<*> * - jC rr^MdI •

n u n=U
2
A Taylor expansion (j)X(x) = 1 - £. ^(2tt)2 + ... shows
The claim of the theorem follows because the integral
/ T~~i2 dx
J{\x\<e} m2
over the ball of radius e in Rd is finite if and only if d > 3. D
Corollary 3.8.2. The particle returns to the origin infinitely often almost
surely if d < 2. For d > 3, almost surely, the particle returns only finitely
many times to zero and P[limn_>oo \Sn\ = oo] = 1.
Proof. If d > 2, then Aoo = limsupn An is the subset of ft, for which the
particles returns to 0 infinitely many times. Since E[B] = Y^Lo^lAn],
the Borel-Cantelli lemma gives P[j4oo] = 0 for d > 2. The particle returns
therefore back to 0 only finitely many times and in the same way it visits
each lattice point only finitely many times. This means that the particle
eventually leaves every bounded set and converges to infinity.
If d < 2, let p be the probability that the random walk returns to 0:
71
Then pm~1 is the probability that there are at least m visits in 0 and the
probability is p™"1 -pm = pm~l(l -p) that there are exactly m visits. We
can write
E[JB]=^mp—1(l-p) = -^.
m>l P
Because E[B] = oo, we know that p = 1. □
The use of characteristic functions allows also to solve combinatorial prob
lems like to count the number of closed paths starting at zero in the graph:
Proposition 3.8.3. There are

d
Jr*d tl^ eos(2mxk))n dx\ • • • dxd

L
closed paths of length n in the lattice Zd.

3.8. Random walks 165
Proof If we know the probability P[Sn = 0] that a path returns to 0 in n

step, then (2d)nP[5n = 0] is the number of closed paths in Zd of length n.
But P[Sn = 0] is the zeroth Fourier coefficient
J/ r (t>Sn(x)dx=
d J j d a f £l-(y2cos(27rixk))n
—-J dx
Example. In the case d = 1, we have
J122ncos2n(27rx)dx=f 2M
closed paths of length 2n starting at 0. We know that also because
P L[ S 2J n \ = n0 ] J= ( 22 nn)2± n± .
The lattice Zd can be generalized to an arbitrary graph G which is a regular

graph that is a graph, where each vertex has the same number of neighbors.
A convenient way is to take as the graph the Cayley graph of a discrete
group G with generators a\,..., ad.
Corollary 3.8.4. If G is the Cayley graph of an Abelian group Q then the

random walk on G is recurrent if and only at most two of the generators
have infinite order.
Proof. An Abelian group Q is isomorphic to Zk x Zni x ... ZUd. The char

acteristic function of Xn is a function on the dual group Q
EPiSn =n =0]0 = £JUSn(x) dx = Y: JL*xb) d*<=! >LXr-4-TT dx

n=0 G tZoJG Ql- (x)
is finite if and only if Q contains a three dimensional torus which means

fc > 2. |-j
The recurrence properties on non-Abelian groups is more subtle, because

characteristic functions loose then some of their good properties.
Example. An other generalization is to add a drift by changing the prob

ability distribution v on /. Given Pj e (0,1) with Em=iPj = 1- In this
case
<t>x(x) = J2Pje27riXj •
1-71=1
We have recurrence if and only if
1
dx = oo .
/Jjdt 1 - </>x(x)
Take for example the case d = 1 with drift parameterized by p G (0,1).

Then
<t>x(x) = pe2™ + (1 - p)e~2nix = cos(2ttx) + i(2p - 1) sin(27rz) .
which shows that

dx < oo
/Jfad
j d 1 -(/>x(x)
if p ^ 1/2. A random walk with drift will almost certainly not return to 0
infinitely often.
Example. An other generalization of the random walk is to take identically

distributed random variables Xn with values in /, which need not to be
independent. An example which appears in number theory in the case d = 1
is to take the probability space Q, = T1 = E/Z, an irrational number a and
a function / which takes each value in / on an interval [^, ^jr). The
random variables Xn(uj) = f(uJ + na) define an ergodic discrete stochastic
process but the random variables are not independent. A random walk
Sn = Efc=i Xk with random variables Xk which are dependent is called a
dependent random walk.
Figure. // Yk are IID random

variables with uniform distri
bution in [0,a], then Zn =
Efc=i Yk mod 1 are dependent.
Define Xk = (1,0) if Zk G
[0,1/4), Xk = (-1,0) if Zk e
[1/4,1/2), Xk = (0,1) if zk e
[1/2,3/4) and Xk = (0,-1) if
Zk e [3/4,1). Also Xk are no
more independent. For small a,
there can belong intervals, where
Xk is the same because Zk stays
in the same quarter interval. The
picture shows a typical path of
the process Sn = YJk=i Xk.
Example. An example of a one-dimensional dependent random walk is the

problem of "almost alternating sums" [52]. Define on the probability space
Q, = ([0,1], A,dx) the random variables Xn(x) = 21[o,i/2](^ + na) — 1,
where a is an irrational number. This produces a symmetric random walk,
3.9. The arc-sin law for the ID random walk 167
but unlike for the usual random walk, where Sn(x) grows like y/n, one sees
a much slower growth 5„(0) < log(n)2 for almost all a and for special
numbers like the golden ratio (\/5 + l)/2 or the silver ratio y/2 + 1 one has
for infinitely many n the relation
a ■ log(n) + 0.78 < S„(0) < a • log(n) + 1
with a = 1/(2 log(l + \/2)). It is not known whether Sn{0) grows like log(n)
for almost all a.
Figure. An almost periodic ran

dom walk in one dimensions. In
stead of flipping coins to decide
whether to go up or down, one
turns a wheel by an angle a after
each step and goes up if the wheel
position is in the right half and
goes down if the wheel position is
in the left half. While for periodic
a the growth of Sn is either lin
ear (like for a — 0), or zero (like
for a = 1/2), the growth for most
irrational a seems to be logarith
mic.
3.9 The arc-sin law for the ID random walk

Definition. Let Xn denote the { — 1, l}-valued random variable with P[Sn =
±1] = 1/2 and let Sn = Ylk=i Xk be the random walk. We have seen that
it is a martingale with respect to Xn. Given a € Z, we define the stopping
time
Ta = min{n £ N | Sn = a} .
Theorem 3.9.1 (Reflection principle). For integers a, b > 0, one has
P[a + Sn = b , T_Q < n] = P[Sn = a + b}.
Proof. The number of paths from a to b passing zero is equal to the number
of paths from —a to b which in turn is the number of paths from zero to
a + b. □
Figure. The proof of the reflec

tion principle: reflect the part of
the path above 0 at the line 0. To
every path which goes from a to
b and touches 0 there corresponds
a path from —a to b.
The reflection principle allows to compute the distribution of the random

variable T_a:
Theorem 3.9.2 (Ruin time). We have the following distribution of the stop
ping time:
a) P[T_a < n] = P[Sn < -a] + P[Sn > a].
b)P[T-a = n] = £P[Sn = a].
Proof, a) Use the reflection principle in the third equality:
P[T_a<n] = ^2P[T-a<n,a-rSn = b]
bez
= ^P[a + Sn = &] + ^P[T_a<n,a + Sn = 6]
= ]Tp[a + Sn = 6] + ^P[Sn = a + &]

b<0 fe>o
P[Sn<-a]+P[Sn>a]
b) From
P[S» = a]=( Jjk )
we get
^P[5n
n = a] = \(P[Sn-i = a - 1] - P[5n_! = a + 1]) .
Also
P[Sn > a] - P[5„_i > a] = P[5„ > a , Sn-i < a]

+P[Sn > a , Sn-! > a] - P[5„_i > a]
= i(P[5„_i = a]-P[5„_1 = a + l])
3.9. The arc-sin law for the ID random walk 169
and analogously
P[Sn < -a] - P[5n_! < -a] = |(P[5„_i = a - 1] - P[5„_! = a]) .
Therefore, using a)
P[T_a = n] = P[T_Q < n] - P[T_a < n - 1]
= P[5„ < -a] - P[5„_i < -a]
P[5n > a] - P[5„_i > a]
= ^(P^n-âl-P^n-Ô+l])
^(PlSn-â-lJ-P^n-ô])
= 5(P[5„_i = a - 1] - P[5„_i = a + 1]) = -P[5n = a]
* n
D
Theorem 3.9.3 (Ballot theorem).
P[5„ = a, 5i>0,...,5B_1>0] = --P[5n = o]

n
Proof. When reversing time, the number of paths from 0 to a of length n

which do no more hit 0 is the number of paths of length n which start in
a and for which T_a = n. Now use the previous theorem
P[T.a = n] =n -P[5n=a].
Corollary 3.9.4. The distribution of the first return time is
P[r0 > 2n] = P[S2n = 0] .
Proof
P[T0>2n] = ^P[T_1>2n-l] + ip[r1>2n-l]

= P[T_i > 2n - 1] (by symmetry)
= P[52n-i > -1 and S2n-i < 1]
= P[52n-l€{0,l}]
= P[52n_! = 1] = P[52n = 0] .
D
Remark. We see that limn_+ooP[To > 2n] = 0. This restates that the
random walk is recurrent. However, the expected return time is very long:
oo oo oo
E[T0] = £ nP[T0 = n] = ]T P[T0 > n] = £ P[Sn = 0] = oo

n=0 n=0 n=0
because by the Stirling formula n! ~ nne~ny/2wn, one has f j ~
22n/v/7rn and so
P[^ = 0]=(^)^~(-nr1/a
Definition. We are interested now in the random variable
L(uj) = max{0 <n<2N\ Sn(uj) = 0}
which describes the last visit of the random walk in 0 before time 2N. If
the random walk describes a game between two players, who play over a
time 2iV, then L is the time when one of the two players does no more give
up his leadership.
Theorem 3.9.5 (Arc Sin law). L has the discrete arc-sin distribution:
2N -2n
p[^»]=^(2;)( N - n
and for N —▶ oo, we have

L 2
P^2iV ~ ^ ~* ~arcsin(^
Proof.
P[L = 2n] = P[52n = 0] • P[T0 > 2N - 2n] = P[52n = 0] • P[S2N-2n = 0]
which gives the first formula. The Stirling formula gives P[S2k = 0] ~ ^
so that ,
with 1
fix) = , n , ■
7Ty/x(l - X)
3.10. The random walk on the free group 171
It follows that
L fz 2
y[wz?
2 N <z]-+ f(x)
J 0 dx = - n arcsin(x/i) .
0.2 0.4 0.6 0
Figure. The distribution function Figure. The density function of

P[L/2N < z] converges in the this distribution in the limit N —>
limit N —> oo to the function oo is called the arc-sin distribu
2arcsm(y/z)/7r. tion.
Remark. From the shape of the arc-sin distribution, one has to expect that
the winner takes the final leading position either early or late.
Remark. The arc-sin distribution is a natural distribution on the interval

[0,1] from the different points of view. It belongs to a measure which is
the Gibbs measure of the quadratic map x i—> 4 • x(l — x) on the unit
interval maximizing the Boltzmann-Gibbs entropy. It is a thermodynamic
equilibrium measure for this quadratic map. It is the measure p on the
interval [0,1] which minimizes the energy
I(fi) = - j [ log \E - E'\ dp(E) d»(E') .

Jo Jo
One calls such measures also potential theoretical equilibrium measures.
3.10 The random walk on the free group

Definition. The free group Fd with d generators is the set of finite words
w written in the 2d letters
A = {a1,a2,...,ad,a11,a21,...,ad1 }
modulo the identifications aia~ = a~ ai = 1. The group operation is
concatenating words vow = vw. The inverse of w = W\W2 • • • wn is w~1 =
wn1 "'w2lwi1' Elements w in the group Fd can be uniquely represented
by reduced words obtained by deleting all words vv~l in w. The identity
e in the group Fd is the empty word. We denote by l(w) the length of the
reduced word of w.
Definition. Given a free group G with generators A and let Xk be uniformly

distributed random variables with values in A. The stochastic process Sn =
Xi • • • Xn is called the random walk on the group G. Note that the group
operation Xk needs not to be commutative. The random walk on the free
group can be interpreted as a walk on a tree, because the Cayley graph of
the group Fd with generators A contains no non-contractible closed circles.
+■■+
Figure. Part of the Cayley graph *

of the free group F2 with two gen
erators a, b. It is a tree. At ev 4--+ 4--f
ery point, one can go into 4 dif

ferent directions. Going into one
of these directions corresponds to
multiplying with a,a~l,b orb-1.
* *
Definition. Define for n G N
rn = P[5n = e , Si ^ e,52 + e,.. .Sn-i ± e]

which is the probability of returning for the first time to e if one starts at
e. Define also for n G N
mn = P[Sn = e]
with the convention m(0) = 1. Let r and m be the probability generating
functions of the sequences rn and ran:
m(x) = y^mnxn, r(x) = ^Vnxn .

n=0 n=0
These sums converge for \x\ < 1.
Lemma 3.10.1. (Feller)

1
m(x) = 1 - r(x)
3.10. The random walk on the free group 173
Proof. Let T be the stopping time

T = min{n G N | Sn = e} .
With P[T = n] = rn, the function r(x) = 2nrnxn is tne probability
generating function of T. The probability generating function of a sum in
dependent random variables is the product of the probability generating
functions. Therefore, if T» are independent random variables with distribu
tion T, then Y17=i ^ nas tne probability generating function x (-▶ rn(x).
We have
oo oo
n=0 n=0
oo
= z2 z2 'p[Sn1=e,Sn2=e,...,Snk=e,
n=0 0<ni<n2<---<nk
Sn 7^ e for n ^ {m,..., nk }]xn

oo n oo
= £ pEr* = »]*" = £'"(*) = iz7£T-
n=0 fc=l fc=0 'W
D
Remark. This lemma is true for the random walk on a Cayley graph of any
finitely presented group.
The numbers r2n+i are zero for odd 2n+1 because an even number of steps
are needed to come back. The values of r2n can be computed by using basic
combinatorics:
Lemma 3.10.2. (Kesten)
^=(2^^ »-"l2)M(2d-ir
Proof We have
To count the number of such words, map every word with 2n letters into
a path in Z2 going from (0,0) to (n, n) which is away from the diagonal
except at the beginning or the end. The map is constructed in the following
way: for every letter, we record a horizontal or vertical step of length 1.
If l(wk) = l(wk~x) + 1, we record a horizontal step. In the other case, if
l(wk) = l(wk~l) - 1, we record a vertical step. The first step is horizontal
independent of the word. There are
If 2n-2
n\ ra-1
such paths since by the distribution of the stopping time in the one dimen
sional random walk
P[T.1 = 2n-l] = ^J—.p[S2n,1 = i]

1 / 2n - 1
2n-l\ n
l / 2 n - 2
n \ n - l
Counting the number of words mapped into the same path, we see that we
have in the first step 2d possibilities and later (2d- 1) possibilities in each
of the n - 1 horizontal step and only 1 possibility in a vertical step. We
have therefore to multiply the number of paths by 2d(2d - l)n_1. □
Theorem 3.10.3 (Kesten). For the free group Fd, we have
2d-l
m(x)
(d - 1) + yjd? - (2d - l)x2
Proof. Since we know r2n we can compute
r ,W
( x, ) d-^/d2- (2d - l)x2
2d -= 1 —
and get the claim with Feller's lemma m(x) = 1/(1 - r(x)). □
Remark. The Cayley graph of the free group is also called the Bethe lattice.
One can read of from this formula that the spectrum of the free Laplacian
L : l2(Fd) -> l2(Fd) on the Bethe lattice given by
Lu(d) = J2u(q+a)
aeA
is the whole interval [-a, a] with a = 2y/2d- 1.
Corollary 3.10.4. The random walk on the free group Fd with d generators
is recurrent if and only if d = 1.
Proof. Denote as in the case of the random walk on Zd with B the random
variable counting the total number of visits of the origin. We have then
again E[B] = £n P[5n = e] = £n mn = m(l). We see that for d = 1 we
3 . 11 . T h e f r e e L a p l a c i a n o n a d i s c r e t e g r o u p 1 7 5
have m(l) = oo and that m(d) < oo for d > 1. This establishes the analog
of Polya's result on Zd and leads in the same way to the recurrence:
(i) d = 1: We know that Zi = Fu and that the walk in Z1 is recurrent.
(ii) d > 2: define the event An = {Sn = e). Then A^ = limsupn An is the
subset of fi, for which the walk returns to e infinitely many times. Since
for d > 2,
oo
E[£] = IZP^n]m(d)<oc,
The Borel-Cantelli lemma gives P[Ax>] = 0 for d > 2. The particle returns
therefore to 0 only finitely many times and similarly it visits each vertex in
Fd only finitely many times. This means that the particle eventually leaves
e v e r y b o u n d e d s e t a n d e s c a p e s t o i n f i n i t y. □
Remark. We could say that the problem of the random walk on a discrete
group G is solvable if one can give an algebraic formula for the function
m(x). We have seen that the classes of Abelian finitely generated and free
groups are solvable. Trying to extend the class of solvable random walks
seems to be an interesting problem. It would also be interesting to know,
whether there exists a group such that the function m(x) is transcendental.
3.11 The free Laplacian on a discrete group

Definition. Let G be a countable discrete group and A C G a finite set
which generates G. The Cayley graph V of (G, A) is the graph with edges
G and sites (ij) satisfying i - j G A or j - i G A.
Remark. We write the composition in G additively even so we do not

assume that G is Abelian. We allow A to contain also the identity e G G.
In this case, the Cayley graph contains two closed loops of length 1 at each
site.
Definition. The symmetric random walk on T(G, A) is the process obtained
by summing up independent uniformly distributed (A U A~ ^-valued ran
dom variables Xn. More generally, we can allow the random variables Xn
to be independent but have any distribution on AU A'1. This distribution
is given by numbers pa = p~l e [0,1] satisfying EaeAuA-i Pa = 1.
Definition. The free Laplacian for the random walk given by (G,A,p) is
the linear operator on l2(G) defined by
Lgh = Pg-h •
Since we assumed pa =pa-i, the matrix L is symmetric: Lgh = Lhg and
the spectrum
a(L) = {E eC \(L-E) is invertible }

is a compact subset of the real line.
Remark. One can interpret L as the transition probability matrix of the
random walk which is a "Markov chain". We will come back to this inter
pretation later.
Example. G = Z, A = {1}. p = pa = 1/2 for a = 1, -1 and pa = 0 for

a £ {1, —1}. The matrix
0 p
p 0 p
p 0 p
L =
p 0 p
p 0 p
p 0
is also called a Jacobi matrix. It acts on the Hilbert space l2(Z) by (Lu)n =
p(un+i +un-i).
Example. Let G = £>3 be the dihedral group which has the presentation
G = (a,b\a3 = b2 = (ab)2 = 1). The group is the symmetry group of the
equilateral triangle. It has 6 elements and it is the smallest non-Abelian
group. Let us number the group elements with integers {1,2 = a, 3 =
a2,4 = 6,5 = a&,6 = o?b }. We have for example 3 • 4 = o?b = 6 or
3*5 = a2ab = o?b = b = 4. In this case A = {a, &}, A'1 = {a-1, b} so that
A U A'1 = {a, a"1, b}. The Cayley graph of the group is a graph with 6
vertices. We could take the uniform distribution pa = Pb = Pa-1 = I/3 on
A\JA~1, but lets instead chose the distribution pa = pa~l = ^-I^Pb = V2>
which is natural if we consider multiplication by b and multiplication by
6-1 as different.
Example. The free Laplacian on D3 with the random walk transition prob
abilities pa = Pa-1 = ^/^Pb = 1/2 is the matrix
0 1/4 1/4 1/2 0 0

1/4 0 1/4 0 1/2 0
L = 1/40000 1/2
1/2 0 0 0 1/4 1/4
0 1/2 0 1/4 0 1/4
0 0 1/2 1/4 0 0
which has the eigenvalues (-3 ± >/5)/8, (5 ± \/5)/8,1/4, -3/4.

3 . 11 . T h e f r e e L a p l a c i a n o n a d i s c r e t e g r o u p 1 7 7
Figure. The Cayley graph of the

dihedral group G = D$ is a reg
ular graph with 6 vertices and 9
edges.
A basic question is: what is the relation between the spectrum of L, the
structure of the group G and the properties of the random walk on GL
Definition. As before, let mn be the probability that the random walk

starting in e returns in n steps to e and let
m(x) = 2_] mnxn

neG
be the generating function of the sequence ran.
Proposition 3.11.1. The norm of L is equal to limsupn_+oc(mn)1/n, the

inverse of the radius of convergence of m(x).
Proof. Because L is symmetric and real, it is self-adjoint and the spectrum

of L is a subset of the real line R and the spectral radius of L is equal to
its norm ||L||.
We have [Ln]ee = mn since [Ln]ee is the sum of products YlîPaj eacn
of which is the probability that a specific path of length n starting and
landing at e occurs.
It remains therefore to verify that
limsup ||Ln||1/n = limsup[Ln]^n

n—+00 n—>oo
and since the > direction is trivial we have only to show that < direction.
Denote by E(A) the spectral projection matrix of L, so that dE(X) is a
projection-valued measure on the spectrum and the spectral theorem says
that L can be written as L = J A dE(X). The measure pe = dEee is called
a spectral measure of L. The real number E(X) — E(p) is nonzero if and
only if there exists some spectrum of L in the interval [A, p). Since
(-i)
£l^ = f'{E - X)-Uk(E)
can't be analytic in A in a point Ao of the support of dk which is the
spectrum of Z,, the claim follows. □
Remark. We have seen that the matrix L defines a spectral measure pe on

the real line. It can be defined for any group element g, not only g — e and
is the same measure. It is therefore also the so called density of states of L.
If we think of p as playing the role of the law for random variables, then
the integrated density of states E(X) = Fl(X) = J^ dp(X) plays the role
of the distribution function for real-valued random variables.
Example. The Fourier transform U : l2(Zl) -> L2(TX):
u(x) = (Uu)(x) = J2uneinx

nez
diagonalises the matrix L for the random walk on Z1
(ULU*)u(x) = ((UL)(un)(x)=pU(un+i+un-1)(x)
= p]P(un+i +un-i)eVl
= p]Tun(e^n-1)x + e'(n+1)x)
= p Y, u n ( e i x + e - i x ) e i n x
nez
= pûn2cos(x)einx
nez
= 2pcos(x) • u(x) .
This shows that the spectrum of ULU* is [—1,1] and because U is an

unitary transformation, also the spectrum of L is in [—1,1].
Example. Let G = Zd and A = {ei}f=1, where {e^} is the standard bases.
Assume p = pa = l/(2d). The analogous Fourier transform F : l2(Zd) —>
L2(Td) shows that FLF* is the multiplication with \ Y?j=i cos(xj). The
spectrum is again the interval [—1,1].
Example. The Fourier diagonalisation works for any discrete Abelian group
with finitely many generators.
Example. G = Fd the free group with the natural d generators. The spec
trum of L is
[ d~~' d J
which is strictly contained in [—1,1] if d > 1.
Remark. Kesten has shown that the spectral radius of L is equal to 1 if and
only if the group G has an invariant mean. For example, for a finite graph,
where L is a stochastic matrix, for which each column is a probability
vector, the spectral radius is 1 because LT has the eigenvector (1,..., 1)
with eigenvalue 1.
3.12. A discrete Feynman-Kac formula 179
Random walks and Laplacian can be defined on any graph. The spectrum
of the Laplacian on a finite graph is an invariant of the graph but there are
non-isomorphic graphs with the same spectrum. There are known infinite
self-similar graphs, for which the Laplacian has pure point spectrum [63].
There are also known infinite graphs, such that the Laplacian has purely
singular continuous spectrum [95]. For more on spectral theory on graphs,
start with [6].
3.12 A discrete Feynman-Kac formula

Definition. A discrete Schrodinger operator is a bounded linear operator
L on the Hilbert space l2(Zd) of the form
d
(Lu)(n) = ^2 u(n + e*) ~ Mn) + n(n ~ e*) + v(n)u(n) *

i=i
where V is a bounded function on Zd. They are discrete versions of op

erators L = -A + V(x) on L2(Rd), where A is the free Laplacian. Such
operators are also called Jacobi matrices.
Definition. The Schrodinger equation
ihu = Lu, u(0) = uo
is a differential equation in l2(Zd, C) which describes the motion of a com

plex valued wave function u of a classical quantum mechanical system. The
constant ft is called the Planck constant and i — \f—\ is the imaginary
unit. Lets assume to have units where h = 1 for simplicity.
Remark. The solution of the Schrodinger equation is
ut = e~iLu0 .
The solution exists for all times because the von Neumann series
JL t2L2 t3L3
is in the space of bounded operators.
Remark. It is an achievement of the physicist Richard Feynman to see

that the evolution as a path integral. In the case of differential operators
L, where this idea can be made rigorous by going to imaginary time and
one can write for L = — A + V
e~uu(x) = Ex[eX v™*» dsu0(7(t))] ,
where Ex is the expectation value with respect to the measure Px on the

Wiener space of Brownian motion starting at x.
Here is a discrete version of the Feynman-Kac formula:
Definition. The Schrodinger equation with discrete time is defined as
i(ut+e -ut) = eLut ,
where e > 0 is fixed. We get the evolution
ut+ne = (1 - ieL)nut
and we denote the right hand side with LnUt.
Definition. Denote by Tn(i,j) the set of paths of length n in the graph

G having as edges Zd and sites pairs [i,j] with \i — j\ < 1. The graph G
is the Cayley graph of the group Zd with the generators A U A'1 U {e},
where A = {ei,..., e^, } is the set of natural generators and where e is the
identity.
Definition. Given a path 7 of finite length n, we use the notation
f n
exp( / L) = f[L7Wj7(i+i) .
Ji t=i
Let Q is the set of all paths on G and E denotes the expectation with
respect to a measure P of the random walk on G starting at 0.
Theorem 3.12.1 (Discrete Feynman-Kac formula). Given a discrete

Schrodinger operator L. Then
(Lnu)(0) = Eo[exp(rL)ti(7(n))].
Jo
Proof.
(L"u)(0) = $>")(,;"(;)
= E E eM[nL)u(j)
nn
= Yl exP( / LM7(™)) .
iern
□
Remark. This discrete random walk expansion corresponds to the Feynman-

Kac formula in the continuum. If we extend the potential to all the sites of
3.13. Discrete Dirichlet problem 181
the Cayley graph by putting V([k,k]) = V(k) and V([kJ]) = 0 for k j- I,
we can define exp(/7 V) as the product Uti V([y{i)Mi + !)])• Then
(Lnu)(0) = E[exp(/ V)u(7(n))]

Jo
which is formally the Feynman-Kac formula.
In order to compute (Lnu)(k) with L = (1 - fceL), we have to take the

potential v defined by
v([k,k]) = l-iev(j(k)) .
Remark. The Schrodinger equation with discrete time has the disadvantage
that the time evolution of the quantum mechanical system is no more
unitary. This draw-back could be overcome by considering also ih(ut -
ut-e) = eLut so that the propagator from ut-e to ut+€ is given by the
unitary operator
which is a Cayley transform of L. See also [50], where the idea is disussed
to use L = arccos(aL'), where L has been rescaled such that ah has norm
smaller or equal to 1. The time evolution can then be computed by iterating
the map A : (ip, (j)) h-> (2aL^ - 0, ip) on H 0 H.
3.13 Discrete Dirichlet problem

Also for other partial differential equations, solutions can be described prob
abilistically. We look here at the Dirichlet problem in a bounded discrete
region. The formula which we derive in this situation holds also in the
continuum limit, where the random walk is replaced by Brownian motion.
Definition. The discrete Laplacian on Z2 is defined as
A/(n,m) = /(n-hl,m) + /(n-l,m) + /(n,m+l) + /(n,m-l)-4/(n,m) .

With the discrete partial derivatives
<5+/(n,m) = i(/(n+l,m)-/(n,m)), <5"/(n,m) = -(/(n,m)-/(n-l,m)) ,
8+f(n,m) = ^(/(n,m+l)-/(n,m)), S~f(n,m) = -(/(n,m)-/(n,m-l)) ,

the Laplacian is the sum of the second derivatives as in the continuous case,
where A = fxx + fyy:
A = #*"+*+*;.
The discrete Laplacian in Z3 is defined in the same way as a discretisation
of A = fxx + fyy + fzz. The setup is analogue in higher dimensions
1 d
(Au)(n) = — ^2(u(n + a) + u(n - a) - 2u(n)) ,
i=l
where ei,..., e<* is the standard basis in Zd.
Definition. A bounded region D in Zd is a finite subset of Zd. Two points

are connected in D if they are connected in Z3. The boundary 5D of D
consists of all lattice points in D which have a neighboring lattice point
which is outside D. Given a function / on the boundary 5D, the discrete
Dirichlet problem asks for a function u on D which satisfies the discrete
Laplace equation An = 0 in the interior mi(D) and for which u = f on the
boundary 5D.
Figure. The discrete Dirichlet

problem is a problem in lin
ear algebra. One algorithm to
solve the problem can be restated
as a probabilistic "path integral
method". To find the value of u
at a point x, look at the "dis
crete Wiener space" of all paths
7 starting at x and ending at
some boundary point St(u) G
5D of D. The solution is u(x) =
Ex[/(St)].
Definition. Let ilXin denote the set of all paths of length n in D which start
at a point x G D and end up at a point in the boundary SD. It is a subset
of T^n, the set of all paths of length n in Zd starting at x. Lets call it the
discrete Wiener space of order n defined by x and D. It is a subset of the
set TXin which has 2dn elements. We take the uniform distribution on this
finite set so that P*,n[{7}] = l/2dn.
Definition. Let L be the matrix for which LXiV = If (2d) if x,y G Zd are
connected by a path and x is in the interior of D. The matrix L is a bounded
linear operator on l2(D) and satisfies Lx,z = Lz,x for x,zG int(D) = D\5D.
Given / : SD -> R, we extend / to a function F(x) = 0 on / D = D\SD
and F(x) = f(x) for x G SD. The discrete Dirichlet problem can be restated
as the problem to find the solution u to the system of linear equations
(1-L)u = f.
Lemma 3.13.1. The number of paths in Qx,n starting at x G D and ending

at a different point y G D is equal to (2d)nLny.
Proof. Use induction. By definition, Lxz is l/(2d) if there is a path from x

to z. The integer L™y is the number of paths of length n from x to y. □
Figure, //"ere is an example of a

problem where D C Z2 /ms i0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
000 0 000000
10 10 10 0 10 0
0 10 10 10 0 10
4L =
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 1 0 0 11
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
Only the rows corresponding to

interior points are nonzero.
Definition. For a function / on the boundary SD, define
E*,n[/] = E f(y)LZv
yeSD
and
oo
Ex[/] = £>*.»[/]•
n=0
This functional defines for every point x G D a probability measure /ix on
the boundary SD. It is the discrete analog of the harmonic measure in the
continuum. The measure Px on the set of paths satisfies Ex[l] = 1 as we
will just see.
Proposition 3.13.2. Let Sn be the random walk on Zd and let T be the

stopping time which is the first exit time of S from D. The solution to the
discrete Dirichlet problem is
u(x)=Ex[f(ST)].
Proof. Because (1 - L)u = f and
E*,n[/] = (Lnf)x ,
we have from the geometric series formula
( l - ^ r 1k=0^ ^
the result
oo oo
u(x) = (1 - L)-Xf[x) = V>"/]s = 5>x,n[/] = EX[ST] .

n=0 n=0
Define the matrix K by Kjj = 1 for j G SD and ^ = L^/4 else. The

matrix K is a stochastic matrix: its column vectors are probability vectors.
The matrix K has a maximal eigenvalue 1 and so norm 1 (KT has the
maximal eigenvector (1,1,..., 1) with eigenvalue 1 and since eigenvalues of
K agree with eigenvalues of KT). Because ||L|| < 1, the spectral radius of
L is smaller than 1 and the series converges. If / = 1 on the boundary,
then u = 1 everywhere. From E^l] = 1 follows that the discrete Wiener
measure is a probability measure on the set of all paths starting at x. D
• *■■■• ■-• . . ♦ i ■■# -i -•'
Figure. The random Figure. The diffusion Figure. The diffusion

walk defines a diffu- process after time t = process after time t =
sion process. 2. 3.
The path integral result can be generalized and the increased generality
makes it even simpler to describe:
Definition. Let (D, E) be an arbitrary finite directed graph, where D is

a finite set of n vertices and E C D x D is the set of edges. Denote an
edge connecting i with j with e^. Let K be a stochastic matrix on l2(D):
the entries satisfy K^ > 0 and its column vectors are probability vectors
îeD Kij = 1 f°r a^ 3 ^ D- The stochastic matrix encodes the graph and
additionally defines a random walk on D if K^ is interpreted as the tran
sition probability to hop from j to i. Lets call a point j G SD a boundary
point, if Kjj = 1. The complement mtD = D\SD consists of interior points.
Define the matrix L as Ljj = 0 if j is a boundary point and Lij = Kji
otherwise.
The discrete Wiener space Qx c D on D is the set of all finite paths 7 =
(x = xo,xi,X2,... ,xn) starting at a point x G D for which KXiXi+1 > 0.
The discrete Wiener measure on this countable set is defined as Px[{7}] =
YYjZo Kjj+i- A function u on D is called harmonic if (Lu)x = 0 for all
x € D. The discrete Dirichlet problem on the graph is to find a function u
on D which is harmonic and which satisfies u = f on the boundary 5D of
D.
Theorem 3.13.3 (The Dirichlet problem on graphs). Assume D is a directed

graph. If Sn is the random walk starting at x and T is the stopping time
to reach the boundary of Z), then the solution
u = Ex[f(ST)]
is the expected value of St on the discrete Wiener space of all paths starting
at x and ending at the boundary of D.
Proof. Let F be the function on D which agrees with / on the boundary of

D and which is 0 in the interior of D. The Dirichlet problem on the graph
is the system of linear equations (1 - L)u = f. Because the matrix L has
spectral radius smaller than 1, the problem is given by the geometric series
ELnf
n=0
But this is the sum Ex[f(Sr)] over all paths 7 starting at x and ending at
the boundary of /. □
Example. Lets look at a directed graph (D, E) with 5 vertices and 2 bound
ary points. The Laplacian on D is defined by the stochastic matrix
0 1/3 0 0 0"
1/2 0 10 0
K = 1/4 1/2 0 0 0
1/8 1/6 0 10
1/8 0 0 0 1
or the Laplacian
0 1/2 1/4 1/8 1/8

1/3 0 1/2 1/6 0
L = 0 10 0 0
0 0 0 10
0 0 0 0 1
Given a function / on the boundary of D, the solution u of the discrete
Dirichlet problem (1 - L)u = f on this graph can be written as a path
integral En=oLnf = Ex[f(ST)] for the random walk Sn on D stopped at

the boundary SD.
Figure. The directed graph

(D,E) with 5 vertices and 2
boundary points.
Remark. The interplay of random walks on graphs and discrete partial

differential equations is relevant in electric networks. For mathematical
treatments, see [19, 99].
3.14 Markov processes

Definition. Given a measurable space (S, B) called state space, where S is
a set and B is a a-algebra on S. A function P : S x B -> R is called a
transition probability function if P(x, •) is a probability measure on (5, B)
for all x G S and if for every B G S, the map 5 -> P(s, B) is B-measurable.
Define P1(x,B) = P(x,B) and inductively the measures Pn+1(x,B) =
fsPn(y,B)P(x,dy), where we write f P(x,dy) for the integration on S
with respect to the measure P(x, •).
Example. If 5 is a finite set and B is the set of all subsets of S. Given

a stochastic matrix K and a point s G S, the measures P(s, •) are the
probability vectors, which are the columns of K.
Remark. The transition probability functions are elements in £(5, Mi(S)),
where Mi(S) is the set of Borel probability measures on S. With the mul
tiplication
(PoQ)(x,B)= [ P(y,B)dQ(x)
Js
we get a commutative semi-group. The relation Pn+m = pn o Pm is also
called the Chapmann-Kohnogorov equation.
Definition. Given a probability space (fi, A, P) with a filtration An of a-

algebras. An *4n-adapted process Xn with values in S is called a discrete
3.14. Markov processes 187
time Markov process if there exists a transition probability function P such
that
V[Xn € B | Ak]{u>) = Pn-k{Xk{w),B) .
Definition. If the state space 5 is a discrete space, a finite or countable

set, then the Markov process is called a Markov chain, A Markov chain is
called a denumerable Markov chain, if the state space S is countable, a
finite Markov chain, if the state space is finite.
Remark. It follows from the definition of a Markov process that Xn satisfies

the elementary Markov property: for n > A:,
P [Xn G B | X1,..., Xh] = P [Xn G B I Xk] .
This means that the probability distribution of Xn is determined by know

ing the probability distribution of Xn-i. The future depends only on the
present and not on the past.
Theorem 3.14.1 (Markov processes exist). For any state space (S,B) and
any transition probability function P, there exists a corresponding Markov
process X.
Proof. Choose a probability measure p on (£, B) and define on the prod

uct space (Q,A) = (SN,BN) the 7r-system C consisting of of cylinder-sets
IlneN Bn given by a sequence Bn G B such that Bn = S except for finitely
many n. Define a measure P = PM on (fi,C) by requiring
P[cjfe G Bk,k= l,...n] = / /i(dx0) / P(x0,dxi)... P(xn-i,dxn) .

JBQ JBi JBn
This measure has a unique extension to the cr-algebra A.

Define the increasing sequence of cr-algebras An = Bn x nr=i(^'^} con"
taining cylinder sets. The random variables Xn((jj) = xn are ^4n-adapted.
In order to see that it is a Markov process, we have to check that
P[Xn G Bn | An-l](0j) = P(Xn-l (w), Bn)

which is a special case of the above requirement by taking Bk = S for
k^n. □
Example. Independent 5-valued random variables

Assume the measures P(x, •) are independent of x. Call this measure P. In
this case
P[XneBn\An-l](0j)=P[Bn]
which means that P[Xn G Bn | An-i] = P[Xn G Bn]. The S- valued
random variables Xn are independent and have identical distribution and
P is the law of Xn. Every sequence of IID random variables is a Markov
process.
Example. Countable and finite state Markov chains.
Given a Markov process with finite or countable state space S. We define
the transition matrix P^ on the Hilbert space l2(S) by
Pij = p(i,{j})-
The matrix P transports the law of Xn into the law of Xn+\.
The transition matrix Pij is a stochastic matrix: each column is a proba
bility vector: ^. P^ = 1 with P^- > 0. Every measure on S can be given
by a vector n G l2(S) and Pn is again a measure. If Xo is constant and
equal to i and Xn is a Markov process with transition probability P, then
pn=P[Xn = j}.
Example. Sum of independent 5-valued random variables Let S be a count

able Abelian group and let tt be a probability distribution on S assigning
to each j G S the weight ttj. Define P{j — ttj-i. Now Xn is the sum of n
independent random variables with law tt. The sum changes from i to j
with probability Pij = pi-j.
Example. Branching processes Given S = {0,1,2... } = N with fixed

probability distribution tt. If X is a 5-valued random variable with distri
bution tt then Ylk=i Xk has a distribution which we denote by ir(n\ Define
the matrix P^ = iry. The Markov chain with this transition probability
matrix on S is called a branching process.
Definition. The transition probability function P acts also on measures tt

of 5 by
V(tt)(B)= [ P(x,B)dTr(x) .
Js
is
A probability measure tt is called invariant if Vtt = tt. An invariant measure
tt on S is called stationary measure of the Markov process.
This operator on measures leaves a subclass of measures with densities with
respect to some measure v invariant. We can so assign a Markov operator
to a transition probability function:
Lemma 3.14.2. For any x e S define the measure

oo 1
n=0
on (S,B), has the property that if ji is absolutely continuous with respect

to z^, then also Vji is absolutely continuous with respect to v.
3.14. Markov processes 189
Proof. Given p = f -v with / G LX(S). Then
Vn= j P(x,B)f(x)dv(x
is absolutely continuous with respect to v because Vfi(B) = 0 implies

P(x, J5) = 0 for almost all x with f(x) > 0 and so fu(B) =0. D
Corollary 3.14.3. To each transition probability function can be assigned a

Markov operator V : LX(S, v) -> Lx(5, v).
Proof. Choose v as above and define
Vh = h
if Ppx = ^2 with /Xi = fiVi. To check that V is a Markov operator, we have

to check Vf > 0 if / > 0, which follows from
Vfv(B) - / P(x,B)f(x) dv(x) > 0 .

Js
We also have to show that \\Vf\\i = 1 if ||/||i. It is enough to show this
for elementary functions / = ^ UjIbj with aj > 0 with Bj G B satisfying
^2jaju(Bj) = 1 satisfies ||P1bi/|| = v(B). But this is obvious ||PlBz/|| =
fBP(x,-)dv(x) = v(B). □
We see that the abstract approach to study Markov operators on LX(S) is
more general, than looking at transition probability measures. This point
of view can reduce some of the complexity, when dealing with discrete time
Markov processes.
Chapter 4
Continuous Stochastic
Processes
4.1 Brownian motion

Definition. Let (Q,A,P) be a probability space and let T C R be time.
A collection of random variables Xt, t G T with values in R is called a
stochastic process. If Xt takes values in S = Rd, it is called a vector-valued
stochastic process but one often abbreviates this by the name stochastic
process too. If the time T can be a discrete subset of R, then Xt is called
a discrete time stochastic process. If time is an interval, R+ or R, it is
called a stochastic process with continuous time. For any fixed uj G fi, one
can regard Xt(uj) as a function of t. It is called a sample function of the
stochastic process. In the case of a vector-valued process, it is a sample
path, a curve in Rd.
Definition. A stochastic process is called measurable, if X : T x f2 —▶ S is

measurable with respect to the product a-algebra B(T) x A In the case of
a real-valued process (5 = R), one says X is continuous in probability if
for any t G R the limit Xt+h -> Xt takes place in probability for h —▶ 0.
If the sample function Xt(uj) is a continuous function of t for almost all uj,
then Xt is called a continuous stochastic process. If the sample function is
a right continuous function in t for almost all uj G fi, Xt is called a right
continuous stochastic process. Two stochastic process Xt and Yt satisfying
p[Xt - Yt = 0] = 1 for all t G T are called modifications of each other
or indistinguishable. This means that for almost all uj G f£, the sample
functions coincide Xt(uj) — Yt(uj).
Definition. A Revalued random vector X is called Gaussian, if it has the

multidimensional characteristic function
ct>X(s) = E[eiS'X] = e-^Vs)/2+i(m,s)
191
192 Chapter 4. Continuous Stochastic Processes
for some nonsingular symmetric nxn matrix V and vector m = E[X]. The
matrix V is called covariance matrix and the vector m is called the mean
vector.
Example. A normal distributed random variable X is a Gaussian random

variable. The covariance matrix is in this case the scalar Var[X].
Example. If V is a symmetric matrix with determinant det(F) ^ 0, then

the random variable
X(X) = 1t p-{x-m,V-\x-m))l2
(27r)"/2^t(l0
on Q = Rn is a Gaussian random variable with covariance matrix V. To
see that it has the required multidimensional characteristic function (j>x (u).
Note that because V is symmetric, one can diagonalize it. Therefore, the
computation can be done in a bases, where V is diagonal. This reduces the
situation to characteristic functions for normal random variables.
Example. A set of random variables Xx,..., Xn are called jointly Gaussian

if any linear combination ^™=1 aiXi is a Gaussian random variable too.
For a jointly Gaussian set of of random variables Xj, the vector X =
(Xi,..., Xn) is a Gaussian random vector.
Example. A Gaussian process is a Revalued stochastic process with con

tinuous time such that (Xto, Xtl,..., Xtn) is jointly Gaussian for any to <
t\ < • • • < tn- It is called centered if mt = E[Xt] = 0 for all t.
Definition. An Revalued continuous Gaussian process Xt with mean vector

mt = E[Xt] and the covariance matrix V(s,t) = Cov[Xs,Xt] = E[(XS —
ms) - (Xt — mt)*] is called Brownian motion if for any 0 < to < t\ < • • • < £n,
the random vectors Xto,Xti+1 — Xti are independent and the covariance
matrix V satisfies V(s,t) = V(r,r), where r = min(s,t) and s h-+ V(s,s).
It is called the standard Brownian motion if mt = 0 for all t and V(s, t) =
min{s,£}.
Figure. A path Xt(uJi) of Brow

nian motion in the plane S = R2
with a drift mt = E[Xt] = (t,0).
This is not standard Brownian
motion. The process Yt = Xt —
(t,0) is standard Brownian mo
tion.
4.1. Brownian motion 193
Recall that for two random vectors X, Y with mean vectors ra, n, the covari
ance matrix is Cov[X, Y]„ = E[(X{ - m{)(Yj - nj)]. We say Cov[X, Y]=0
if this matrix is the zero matrix.
Lemma 4.1.1. A Gaussian random vector (X, Y) with random vectors X, Y

satisfying Cov[X, Y] = 0 has the property that X and Y are independent.
Proof. We can assume without loss of generality that the random variables
X, Y are centered. Two Rn-valued Gaussian random vectors X and Y are
independent if and only if
<t>(x,Y){s,t) = (f>x(s) • (j>y(t)ys,t G Rn
Indeed, if V is the covariance matrix of the random vector X and W is the

covariance matrix of the random vector Y, then
U Cov[X,y] U 0
U=
C o v [ y, x ] v 0 V
is the covariance matrix of the random vector (X, Y). With r = (£, s), we
have therefore
*(*,y)(r) = E[e^^]=e-^^
= e-Hs-V8)-±(t-Wt)
= e - i ( s - Va ) e - i ( t - W t )
= <l>x(s)<l>Y(t) •
Example. In the context of this lemma, one should mention that there
exist uncorrelated normal distributed random variables X, Y which are not
independent [109]: Proof. Let X be Gaussian on R and define for a > 0 the
variable Y(uj) — —X(uj), if uj > a and Y = X else. Also Y is Gaussian and
there exists a such that E[X7] =0. But X and Y are not independent and
X+Y = 0 on [—a, a] shows that X+Y is not Gaussian. This example shows
why Gaussian vectors (X,Y) are defined directly as R2 valued random
variables with some properties and not as a vector (X, Y) where each of
the two component is a one-dimensional random Gaussian variable.
Proposition 4.1.2. If Xt is a Gaussian process with covariance V(s, t) =

V(r,r) with r = min(s,£), then it is Brownian motion.
Proof By the above lemma (4.1.1), we only have to check that for all i < j
Cov[Xt0,X,.+1 - Xtj] = 0, Cov[Xt2+1 - Xti,Xtj+1 -Xtj]=0.

But by assumption
Cov[Xt0,Xt.+1 - Xtj] = V(to,tj^)-V(to,tj) = V(t0,t0) - V(t0,t0) = 0

and
Cov[Xti+1 - Xu,Xtj+1 - Xtj] = V(ti+Utj+1) - ViU+utj)

-V(U,tj+1) + V(tutj)
= V(U+i,U+i) — V(tf+i,ti+i)
-V(tuU) + V(tuti) = 0.
D
Remark. Botanist Robert Brown was studying the fertilization process in a

species of flowers in 1828. While watching pollen particles in water through
a microscope, he observed small particles in "rapid oscillatory motion".
While previous studies concluded that these particles were alive, Brown's
explanation was that matter is composed of small" active molecules", which
exhibit a rapid, irregular motion having its origin in the particles themselves
and not in the surrounding fluid. Brown's contribution was to establish
Brownian motion as an important phenomenon, to demonstrate its presence
in inorganic as well as organic matter and to refute by experiment incorrect
mechanical or biological explanations of the phenomenon. The book [73]
includes more on the history of Brownian motion.
The construction of Brownian motion happens in two steps: one first con
structs a Gaussian process which has the desired properties and then shows
that it has a modification which is continuous.
Proposition 4.1.3. Given a separable real Hilbert space (H, || • ||). There
exists a probability space (0, A P) and a family X(h), h G H of real-valued
random variables on Q such that h*-> X(h) is linear, and X(h) is Gaussian,
centered and E[X(/i)2] =
Proof. Pick an orthonormal basis {en} in H and attach to each en a cen

tered Gaussian IID random variable Xn G C2 satisfying ||Xn||2 = 1. Given
a general h = ^ hnen G H\ define
X(h) = }hnXn
n
which converges in C2. Because Xn are independent, they are orthonormal

in C2 so that
l|x(fc)ll2 = £ft»llx»H2 = E',n = IWI

4.1. Brownian motion 1^5
□
2
Definition. If we choose H = L2(R+,dx), the map X : H ■--> C IS
also called a Gaussian measure. For a Borel set A C R+ we define then
X(A) = X(1a)> The term "measure" is warranted by the fact that X(A) =
J2n x(An) if A is a countable disjoint union of Borel sets A„. One also has
X(0) = 0.
Remark. The space X(H) C C2 is a Hilbert space isomorphic to H and in
particular
E[X(h)X(ti)] = (h,ti).
We know from the above lemma that h and h! are orthogonal if and only
if X(h) and X(/i') are independent and that
E[X(A)X(B)] = Cov[X(A), X(B)] = (1A, lB) = \AnB\.
Especially X(A) and X(B) are independent if and only if A and £ are
disjoint.
Definition. Define the process Bt = X([0, £]). For any sequence ti,t2, • • • G
T, this process has independent increments Bti - Bu_x and is a Gaussian
process. For each t, we have E[B2] = t and for s <t, the increment Bt - Bs
has variance t — s so that
E[B,flt] = E[PS2] + E[Ps(Pt - Ba)] = E[P2] = s .
This model of Brownian motion has everything except continuity.
Theorem 4.1.4 (Kolmogorov's lemma). Given a stochastic process Xt with

t e[a,b] for which there exist three constants p> r,K such that
E[\Xt+h-Xt\p]<K-h^r
for every t,t + h G [a,6], then Xt has a modification Yt which is almost

everywhere continuous: for all s,t G [a, 6]
|yt(a;)-yfl(a;)| < C(u) \t-s\a,0<a< - .
Proof. We can assume without loss of generality that a = 0, b = 1 because

we can translate and rescale the time variable to be in this situation. Define
e — r — cap. By the Chebychev-Markov inequality (2.5.4)
P[|Xt+h - Xt\] > \h\a] < \h\-a>>E[\Xt+h - Xt\*] < K\h\1+<
so that
P[|X(fc+1)/2- - Xfc/2»| > 2-™} < K2-"(1+£) .
Therefore
oo 2n-l
£ £ P[|X(fc+1)/2» - Xk/2n\ > 2-™} < oo .

71=1 k=0
By the first Borel-Cantelli's lemma (2.2.2), there exists n(uj) < oo almost
everywhere such that for all n > n(uj) and k = 0,..., 2n — 1
\X(k+i)/2n(v) - Xk/2n(uj)\ < 2~not .
Let n > n(uj) and t G [fc/2n, (fc+l)/2n] of the form * = k/2n+Y7=i 7z/2n_N
with 7* G {0,1}. Then
m
i=l
with d = (1 - 2~a)~1. Similarly
\Xt - X(k+i)2-n\ < d 2~na .
Given t,t + h e D = {k2~n \ n G N,k = 0,...n - 1}. Take n so that

2-71-1 < ft < 2~n and fc so that fc/2n+1 < t < (fc + l)/2n+1. Then (fc +
l)/2n+1 < t + h < (fc + 3)/2n+1 and
l-Xt+h ~ X*l ^ 2d2"(n+1)a < 2dha .
For almost all a;, this holds for sufficiently small /i.
We know now that for almost all uj, the path Xt(uj) is uniformly continuous
on the dense set of dyadic numbers D = {fc/2n}. Such a function can be
extended to a continuous function on [0,1] by defining
Yt(uj)= seD^t
lim Xs(uj).
Because the inequality in the assumption of the theorem implies E[Xt(u;) —

limseD-^t Xs(uj)] = 0 and by Fatou's lemma E[Yt(uj)-limseD^t Xs(uj)] = 0
we know that Xt = Yt almost everywhere. The process Y is therefore a
modification of X. Moreover, Y satisfies
\Yt(uj)-Ys(uj)\<C(uj)\t-s\a
for all s,te [a, 6]. □
Corollary 4.1.5. Brownian motion exists.

4.1. Brownian motion 197
Proof. In one dimension, take the process Bt from above. Since Xh =

Bt+h - Bt is centered with variance ft, the fourth moment is E[X£] =
-£s exp(-x2h/2){x=0 = 3ft2, so that
E[(Bt+h - Bt)4] = 3ft2 .
Kolmogorov's lemma (4.1.4) assures the existence of a continuous modifi
cation of B.
To define standard Brownian motion in n dimension, we take the joint

motion Bt = (B[l\ ..., B[n)) of n independent one- dimensional Brownian
motions. LJ
Definition. Let Bt be the standard Brownian motion. For any x G Rn, the
process Xf = x + Bt is called Brownian motion started at x.
The first rigorous construction of Brownian motion was given by Norbert
Wiener in 1923. By construction of a Wiener measure on C[0,1], one has
a construction of Brownian motion, where the probability space is directly
given by the set of paths. One has then the process Xt(uj) = oj(t). We will
come to this later. A general construction of such measures is possible given
a Markov transition probability function [104]. The construction given here
is due to Neveu and goes back to Kakutani. It can be found in Simon's book
on functional integration [93] or in the book of Revuz and Yor [83] about
continuous martingales and Brownian motion. This construction has the
advantage that it can be applied to more general situations.
In McKean's book "Stochastic integrals" [66] one can find Levy's direct
proof of the existence of Brownian motion. Because that proof gives an ex
plicit formula for the Brownian motion process Bt and is so constructive,
we outline it shortly:
1) Take as a basis in L2([0,1] the Haar functions

fk,n '•= 2(n~1)/2(lp-l)2-*\fc2-"] ~ l[fc2-n,(fc+l)2-n])
for {(fc,n)|n > l,fc < 2n } and /0j0 = 1.
2) Take a family Xfc?n for (fc,n) G I = {(fc,n) | n > l,fc < 2n,fc odd } U
{(0,0) } of independent Gaussian random variables.
3) Define
Bt = 22 ^k>n / ^fc'n '
(k,n)ei Jo
4) Prove convergence of the above series.
5) Check
E[BsBt] = V / / f{k,n)fc = I l[o,*]l[o,t] = inf{M } •

(k,n)eiJo Jo Jo
6) Extend the definition from t G [0,1] to t G [0, oo) by taking independent

Brownian motions b[i) and defining Bt = £n<m £*-n, where [t] is the
largest integer smaller or equal to t.
4.2 Some properties of Brownian motion

We first want to establish that Brownian motion is unique. To do so, we
first have to say, when two processes are the same:
Definition. Two processes Xt on (ft, A P) and X[ on (ft', A, P') are called

indistinguishable, if there exists an isomorphism U : ft —▶ ft' of probability
spaces, such that X't(Uuj) = Xt(u). Indistinguishable processes are consid
ered the same. A special case is if the two processes are defined on the same
probability space (ft, AP) and Xt(uj) = Yt(u) for almost all uj.
Proposition 4.2.1. Brownian motion is unique in the sense that two stan
dard Brownian motions are indistinguishable.
Proof. The construction of the map H -▶ C2 was unique in the sense that
if we construct two different processes X(ft) and Y(h), then there exists an
isomorphism U of the probability space such that X(ft) = Y(U(h)). The
continuity of Xt and Yt implies then that for almost all uj, Xt(uj) = Yt(Uuj).
In other words, they are indistinguishable. D
We are now ready to list some symmetries of Brownian motion.
Theorem 4.2.2 (Properties of Brownian motion). The following symmetries

exist:
(i) Time-homogeneity: For any s > 0, the process Bt = Bt+S - Bs is a
Brownian motion independent of o(Bu,u < s).
(ii) Reflection symmetry: The process Bt — -Bt is a Brownian motion,
(iii) Brownian scaling: For every c> 0, the process Bt = cBt/c2 is a Brow
nian motion.
(iv) Time inversion: The process B0 = 0,Bt = tBXjUt > 0 is a Brownian
motion.
Proof. (i),(ii),(iii) In each case, Bt is a continuous centered Gaussian pro

cess with continuous paths, independent increments and variance t.
4.2. Some properties of Brownian motion 199
(iv) B is a centered Gaussian process with covariance
Cov[Bs,Bt] = E[Ba,Bt] = st • E[B1/a,B1/t] = at • inf(-, -) = inf(M) .
Continuity of Bt is obvious for t > 0. We have to check continuity only for

t = 0, but since E[B2] = s -> 0 for s -> 0, we know that Bs —▶ 0 almost
everywhere. LJ
It follows the strong law of large numbers for Brownian motion:
Theorem 4.2.3 (SLLN for Brownian motion). If Bt is Brownian motion,

then
lim -Bt = 0
t—>oo t
almost surely.
Proof. From the time inversion property (iv), we see that t 1Bt, = Pi/t
which converges for t -> oo to 0 almost everywhere, because of the almost
everywhere continuity of Bt. ^J
Definition. A parameterized curve t G [0, oo) i-> Xt G Rn is called Holder

continuous of order a if there exists a constant C such that
\\Xt+h-Xt\\<C.ha
for all ft > 0 and all t. A curve which is Holder continuous of order a = 1
is called Lipshitz continuous.
The curve is called locally Holder continuous of order a if there exists for
each t a constant C = C(t) such that
\\Xt+h-Xt\\<C.ha
for all small enough ft. For a Revalued stochastic process, (local) Holder
continuity holds if for almost all uj G ft the sample path Xt(uj) is (local)
Holder continuous for almost all uj G ft.
Proposition 4.2.4. For every a < 1/2, Brownian motion has a modification
which is locally Holder continuous of order a.
Proof. It is enough to show it in one dimension because a vector func

tion with locally Holder continuous component functions is locally Holder
continuous. Since increments of Brownian motion are Gaussian, we have
E[(Bt - Bs)2p] = Cp.\t- s\p
for some constant Cp. Kolmogorov's lemma assures the existence of a mod
ification satisfying locally
\Bt-Bs\ <C\t-s\a,0<a<^^ .
2p
Because p can be chosen arbitrary large, the result follows. □
Because of this proposition, we can assume from now on that all the paths
of Brownian motion are locally Holder continuous of order a < 1/2.
Definition. A continuous path Xt = (X| ,..., X^ ) is called nowhere

differentiate, if for all t, each coordinate function X^ is not differentiable
at t.
Theorem 4.2.5 (Wiener). Brownian motion is nowhere differentiable: for

almost all uj, the path 11—▶ Xt(uj) is nowhere differentiable.
Proof. We follow [66]. It is enough to show it in one dimensions. Suppose

Bt is differentiable at some point 0 < s < 1. There exists then an integer /
such that \Bt — Bs\ < l(t — s) for t — s > 0 small enough. But this means
that
\Bj/n-B{j_1)/n\ <7-
for all j satisfying
i = [ns] + 1 < j < [ns] + 4 = i + 3
and sufficiently large n so that the set of differentiable paths is included in

the set
b=u/>1 um>ln n>m

u n0<i<n+l
{i**/» - *u-D/»i < 7l > •
i<j<i+3
Using Brownian scaling, we show that P[JB] =0 as follows
pm u n {i*j/»-*o-D/»i<7i}]
n>m 0<2<n+l i<j<i+3
< liminfnP[|Bi/n| < 7-]

n—>oo ' n
I ,3
liminfnPNPil <7-=]
n—oc y/n
C
< lim —= = 0 .
n—kx) yjn
□
Remark. This proposition shows especially that we have no Lipshitz con
tinuity of Brownian paths. A slight generalization shows that Brownian
motion is not Holder continuous for any a > 1/2. One has just to do the
same trick with fc instead of 3 steps, where k(a — 1/2) > 1. The actual
modulus of continuity is very near to a = 1/2: \Bt — Bt+e\ is of the order
h(e) = ^2elog(-e).
More precisely, Pflimsup^QSupig,^^ ' hU) — 1] = 1' as we wm* see

later in theorem (4.4.2).
The covariance of standard Brownian motion was given by E[BsBt] =
min{s, t}. We constructed it by implementing the Hilbert space L2([0, oo))
as a Gaussian subspace of £2(ft, A, P). We look now at a more general class
of Gaussian processes.
Definition. A function V : T x T —> R is called positive semidefinite,
if for all finite sets {t\,...,td) c T, the matrix Vtj = V(U,tj) satisfies
(u, Vu) > 0 for all vectors u— (u\,...,un).
Proposition 4.2.6. The covariance of a centered Gaussian process is positive

semidefinite. Any positive semidefinite function V on TxT is the covariance
of a centered Gaussian process Xt.
Proof. The first statement follows from the fact that for all u = (u\,..., un)
n
^ViU^muj = E[(%2uiXti)2} > 0 .
We introduce for t G T a formal symbol 5t- Consider the vector space of

finite sums YH=i aiû with inner product
d d
QâA^XÂ) = ^fli6/(tt,tj) •
i=l j=l ij
This is a positive semidefinite inner product. Multiplying out the null vec
tors {||v|| = 0 } and doing a completion gives a separable Hilbert space
H. Define now as in the construction of Brownian motion the process
Xt = X(5t). Because the map X : H -» C2 preserves the inner product, we
have
E[XuXs] = (5s,5t) = V(s,t).
□
Lets look at some examples of Gaussian processes:
Example. The Ornstein-Uhlenbeck oscillator process Xt is a one-dimensional

process which is used to describe the quantum mechanical oscillator as we
will see later. Let T = R+ and take the function V(s,t) = ê-|t_s| on
T x T. We first show that V is positive semidefinite: The Fourier transform
of /(*) = e-l*' is
eikte-^dt = l
/, R 2 7 r ( f c 2 + 1 ) •
By Fourier inversion, we get
J_ f(k2 + l)-leik(t-sUk=le-\t-s\,
27T JR *
and so
0 < (2tt)~1 [(k2 + l)-lJ2\ujeiktj\2dk

Jr j
n ^
j,k=l
This process has a continuous modification because
E[(Xt - Xs)2] = (e-l'-*' + e-ls^l - 2e"lt-sl)/2 = (1 - c-l*—I) <\t-s\
and Kolmogorov's criterion. The Ornstein-Uhlenbeck is also called the os

cillatory process.
Proposition 4.2.7. Brownian motion Bt and the Ornstein-Uhlenbeck pro

cess Ot are for t > 0 related by
Ot = -7f~'^ •
Proof. Denote by O the Ornstein-Uhlenbeck process and let
Xt = 2~1/2e-tBe2t .
We want to show that X = Y. Both X and O are centered Gaussian,

continuous processes with independent increments. To verify that they are
the same, we have to show that they have the same covariance. This is a
computation:
E[OtOs] = ie-Vsmin{e2',e2s } = e'—*'/2 .
It follows from this relation that also the Ornstein-Uhlenbeck process is

not differentiable almost everywhere. There are also generalized Ornstein-
Uhlenbeck processes. The case V(sA) = J^-^'-^ dp(k) = p(t - s)
with the Cauchy measure p = 2n(k2+i)dx on ^ can ê generalized to take
any symmetric measure /ionR and let fi denote its Fourier transform
fRe~ikt d/i(k). The same calculation as above shows that the function
V(a,t) = jl(t — s) is positive semidefinite.
Figure. Three paths of the

Ornstein- Uhlenbeck process.
Example. Brownian bridge is a one-dimensional process with time T —

[0,1] and V(s, t) = s(l - t) for 1 < a < t < 1 and V(s,t) = V(t, s) else. It
is also called tied down process.
In order to show that V is positive semidefinite, one observes that Xt =
Bs — sBi is a Gaussian process, which has the covariance
E[XSX*] = E[(BS - sB1)(Bt - tBi)] = s + st-2st = s(l - t) .
Since E[X2] = 0, we have X\ = 0 which means that all paths start from 0
at time 0 and end at 1 at time 1.
The realization Xt = B8 — sB\ shows also that Xt has a continuous real
ization.
Figure. Three paths of Brownian

bridge.
Let Xt be the Brownian bridge and let y be a point in Rd. We can consider
the Gaussian process Yt = ty + Xt which describes paths going from 0 at
time 0 to y at time 1. The process Y has however no more zero mean.
Brownian motion B and Brownian bridge X are related to each other by
the formulas:
Bt = Bt := (t + l)Xt/(£+1), Xt = Xt := (1 - t)Bt/{1.t) .
These identities follow from the fact that both are continuous centered
Gaussian processes with the right covariance:
E[B.Bt] = (t + l)(s + 1) min{^y, ^y } = min{s,t} = E[BsBt] ,
E[XsXt] = (l-t)(l-s)min{-^—y-^T)} = s(l-t)=E[XsXt]
and uniqueness of Brownian motion.
Example. If V(s,t) = l{s==t}> we get a Gaussian process which has the

property that Xs and Xt are independent, if s ^ t. Especially, there is no
autocorrelation between different Xs and Xt. This process is called white
noise or "great disorder". It can not be modified so that (t,uj) i-> Xt(uj) is
measurable: if (t,uj) \-> Xt(uj) were measurable, then Yt = J0 Xs ds would
be measurable too. But then
E[Yt2] = E[( / Xs)2] = / / E[Xa,Xa>] ds' ds = 0

Jo Jo Jo
which implies Yt = 0 almost everywhere so that the measure dp(uj) =
Xs(uj) ds is zero for almost all uj.
t = E[[ X2] = E[[ XSXS ds] =E[[ Xs dp(s)] = 0 .

./o Jo Jo
In a distributional sense, one can see Brownian motion as a solution of
the stochastic differential equation and white noise as a generalized mean-
square derivative of Brownian motion. We will look at stochastic differential
equations later.
4.3. The Wiener measure 205
Example. Brownian sheet is not a stochastic process with one dimensional

time but a random field: time T = R+ is two dimensional. Actually, as
long as we deal only with Gaussian random variables and do not want to
tackle regularity questions, the time T can be quite arbitrary and proposi
tion (4.2.6) stated at the beginning of this section holds true. The Gaussian
process with
V({s1,s2),(ti,t2)) = min(si,ii)-min(s2,t2)
is called Brownian sheet. It has similar scaling properties as Brownian mo
tion.
Figure. Illustrating a sample of a

Brownian sheet Bt,s- Time is two
dimensional. Every trace Bt =
Bt,So or Bt = Bt,so is standard
Brownian motion.
4.3 The Wiener measure

Let (E,£) be a measurable space and let T be a set called "time". A
stochastic process on a probability space (fi, A, P) indexed by T and with
values in E defines a map
<p: fi-> Et,uj^ Xt(u) .
The product space ET is equipped with the product a-algebra £ , which
is the smallest algebra for which all the functions Xt are measurable which
is the cr-algebra generated by the 7r-system
n
ti,...,t„
consisting of cylinder sets. Denote by Yt{w) = w(t) the coordinate maps on

ET. Because Yt o is measurable. Denote by
Px the push-forward measure of </> from {fl,A,P) to (ET,£T) defined by
PX[A] = P[X-1{A)}. For any finite set (*i,... ,t„) C T and all sets At G £,
we have
P[XU eAi,i = l,...,n]=Px[Yti eAi,l = l,...n].

One says, the two processes X and Y are versions of each other.
Definition. Y is called the coordinate process of X and the probability
measure Px is called the law of X.
Definition. Two processes X, X' possibly defined on different probability

spaces are called versions of each other if they have the same law Px = Px> •
One usually does not work with the coordinate process but prefers to work
with processes which have some continuity properties. Many processes have
versions which are right continuous and have left hand limits at every point.
Definition. Let Dbea measurable subset of ET and assume the process has
a version X such that almost all paths X(uj) are in D. Define the probability
space (D,£T n D,Q), where Q is the measure Q = (j)*P. Obviously, the
process Y defined on (D, ST n £>, Q) is another version of X. If D is right
continuous with left hand limits, the process is called the canonical version
of X.
Corollary 4.3.1. Let E = Rd and T = R+. There exists a unique probability

measure W on C(T, E) for which the coordinate process Y is the Brownian
motion B.
Proof. Let D = C(T,E) C ET. Define the measure W = (j)*Px and let
Y be the coordinate process of B. Uniqueness: assume we have two such
measures W, W and let Y, Y' be the coordinate processes of B on D with
respect to W and W. Since both Y and Yf are versions of X and "being
a version" is an equivalence relation, they are also versions of each other.
This means that W and W coincide on a n- system and are therefore the
same. □
Definition. If E = Rd and T = [0, oo), the measure W on C(T, E) is called

the Wiener measure. The probability space (C(T, E), £T n C(T, E), W) is
called the Wiener space.
Let B' be the cr-algebra £T n C(T,E), which is the Borel a-algebra re

stricted to C(T,E). The space C(T,E) carries an other cr-algebra, namely
the Borel a-algebra B generated by its own topology. We have B C B',
since all closed balls {/ e C(T, E)\\f - /0| < r} <E B are in B'. The other
relation B' C B is clear so that B = B'. The Wiener measure is therefore a
Borel measure.
Remark. The Wiener measure can also be constructed without Brownian

motion and can be used to define Brownian motion. We sketch the idea.
Let S = Rn denote the one point compactification of Rn. Define Q = S^
4.4. Levy's modulus of continuity 207
be the set of functions from [0, t] to 5 which is also the set of paths in R .
It is by Tychonov a compact space with the product topology. Define
Cfiniil) = {<t> € C(n,R) | 3F : Rn - R,0M = F(u;(ti),... ,o;(tn))} •
Define also the Gauss kernelp(x, y, t) = (47rt)-n/2 exp(-|x-2/|2/4t). Define

on Cfin(ft) the functional
(L0)(si,..., am) = / F(xux2,.. •, xm)p(0,xua1)p(xux2, s2)

• • mp(xm—i, xm, smj axi • • • dxm
with si = ti and sk = tk - tk-\ for fc > 2. Since L((j>) < |0(<*>)|oo, it

is a bounded linear functional on the dense linear subspace Cfin(Q) C
C(Q). It is nonnegative and L(l) = 1. By the Hahn Banach theorem, it
extends uniquely to a bounded linear functional on C(Cl). By the Riesz
representation theorem, there exists a unique measure p on C(Q) such that
L(<t>) — J (j)(uj) dp(uj). This is the Wiener measure on fJ.
4.4 Levy's modulus of continuity

We start with an elementary estimate
Lemma 4.4.1.
ie-2/2> / e-xV2dx> « e, - a 2 / 2
Proof.
/ e-2/2 dx < / e"* /2(x/a) dx = Va /2 .
For the right inequality consider
Integrating by parts of the left hand side of this gives
Ie-aV2
a J
_ f°
a
e-xV2
a
dx<\r
J
e-**l2 dX .
a
D
Theorem 4.4.2 (Levy's modulus of continuity). If B is standard Brownian

motion, then
r >[hm
P r r sup
\ B sup
S —j-T-r—
~ B t \= 1] , = 1 ,
e-^0 |s_t|<e ™(e)
where h(e) = ^2elog(l/e).
Proof. We follow [83]:
(i) Proof of the inequality " > 1".

Take 0 < S < 1. Define an = (1 - 5)/i(2~n) = (1 - S)y/n2\og2. Consider
P[An] = P[ l<fc<2n
max |J3fc2-n - B(fc_i)2-»| < an] .
Because Bfc/2" -#(/c-i)/2" are independent Gaussian random variables, we

compute, using the above lemma (4.4.1) and 1 - s < e~~s
/•oo 1
P[4] < (1-2/ -=e-x'2dx)

Jan V27T
< (i_2-^-re-a2/2)2"
an + 1
< e x p ( -r 2
> nn - 2#a=n- ca - f l » / 2 ) < e - ^ W - W ^ ) ,
where C is a constant independent of n. Since £nP[^n] < oo, we get by

the first Borel-Cantelli that P[limsupn An] = 0 so that
P[lim
n - +max
o o \Bk2-n
l < / c <-2 n
J5(fc_1)2n|
v > h(2~n)] = 1 .
(ii) Proof of the inequality "< 1".

Take again 0 < S < 1 and pick e > 0 such that (1 + e)(l - 5) > (1 + S).
Define
P [ A n ] = P [ k=j—i£K
max \Bj2-n-Bi2-n\/h(k2-n)>(l + e)}
= P( U {!%-" _ B«-»|] > «n,fc} ,
where
K = {0 < fc < 2nS }
&ndan,k = h{k2-n){l + e).

Using the above lemma, we get with some constants C which may vary
209
4.5. Stopping times
from line to line:
. keK
keK
< c . 2-n(i-«5)(i+e)2 ^P(log(Ar12n)r1/2 ( since fc-1 > 2~nS)
k€K
< C ■ n-1/22Tl(5_(1_,s)(1+e)2) •
In the last step was used that there are at most 2nS points in K and for
each of them logCAr^) > log(2"(l - 5)).
We see that £n P[An} converges. By Borel-Cantelli we get for almost every
uj an integer ti(oj) such that for n > n(ui)
\Bj2-«-Bi2-n\<(l + e)-h(k2-n),
where k = j-ieK. Increase possibly n{uS) so that for n > n(w)
Y,h(2-m)<e-h(2-(n+1){l-S)).
m>n
Pick 0 < h < h < 1 such that t = t2 - h < 2-"(^1-*>. Take next
n > n(w) such that 2^n+1^-s) < t < 2~n^ and write the dyadic
development of ti, t2\
h = i2-n _ 2-Pl _ 2-P2 _^M= j2-» + 2~^ + 2"* ...
with h < i2~n < j2~n < t2 and 0 < fc = j - i < t2n < 2n5. We get
|Btl(w)-B*a(w)| < \Btl - Bi2-n(u)\

+ \Bi2-n{u) - Bj2-n{0j)\
+ \Bj2-n(0j) - Bt2\
< 2^(l + e)/i(2-p) + (l + €)Mfc2-n)
p>n
< (l + 3e + 2e2)h(t) .
B e c a u s e e > 0 w a s a r b i t r a r y, t h e p r o o f i s c o m p l e t e . □
4.5 Stopping times

Stopping times are useful for the construction of new processes, in proofs
of inequalities and convergence theorems as well as in the study of return
time results. A good source for stopping time results and stochastic process
in general is [83].
Definition. A filtration of a measurable space (ft, A) is an increasing family
(At)t>o of sub-cr-algebras of A. A measurable space endowed with a filtra
tion (At)t>o is called a filtered space. A process X is called adapted to the
filtration At, if xt is A-measurable for all t.
Definition. A process X on (tl,A,P) defines a natural filtration At =
<t{Xs I s <t), the minimal filtration of X for which X is adapted. Heuris-
tically, At is the set of events, which may occur up to time t.
Definition. With a filtration we can associate two other filtration by setting
for t > 0 8
At- =<T(Aa,s<t),At+ = f)As.
s>t
For * = 0 we can still define A0+ = f]s>0 A8 and define A0- = A0. Define
alsoA^ = v(As,s>0).
Remark. We always have At- C At C At+ and both inclusions can be
strict.
Definition. If At = At+ then the filtration At is called right continuous. If
At = At-, then At is left continuous. As an example, the filtration At+ of
any filtration is right continuous.
Definition. A stopping time relative to a filtration At is a map T • ft -+

[0, oo] such that {T <t} eAt.
Remark. If At is right continuous, then T is a stopping time if and only

if {T <t} eAt. Also T is a stopping time if and only if Xt = 1(0 T](*) is
adapted. X is then a left continuous adapted process.
Definition. If T is a stopping time, define
AT = {A e Aoo | An {T < t} e Au\ft}.

It is a a-algebra. As an example, if T = s is constant, then AT = A3. Note
also that
AT+ = {A e Aoo I A n {T < t} e Au Mt} .
We give examples of stopping times.
Proposition 4.5.1. Let X be the coordinate process on C(R+,E), where E

is a metric space. Let A be a closed set in E. Then the so called entry time
TA(uj) = mf{t>0\Xt(uj)eA}
is a stopping time relative to the filtration At = a({Xs }5<t).
Proof. Let d be the metric on E. We have
{{TA<t}
TA < t } == {{ s inf
m ^ t dd(X.
(Xs(uj)1A)=0}
which is in At = a(Xs, x<t). rj

4.5. Stopping times
Proposition 4.5.2. Let X be the coordinate process on D{R+,E), the space

of right continuous functions, where E is a metric space. Let A be an open
subset of E. Then the hitting time
Ta(uj) = ini{t > 0 | Xt(w) G A }
is a stopping time with respect to the filtration At+ ■
Proof. TA is a At+ stopping time if and only if {TA < t} € At for all t.
If A is open and X,(w) G A, we know by the right-continuity of the paths
that Xt(w) G A for every t € [s, s + e) for some e > 0. Therefore
{TA
1 <t}
J = s {6 Q
inf, s Xs
<t G A } G A •
Definition. Let At be a filtration on {Q,A) and let T be a stopping time.

For a process X, we define a new map XT on the set {T < oc} by
Xt(uj) = XT{u>){u>) ■
Remark. We have met this definition already in the case of discrete time
but in the present situation, it is not clear whether XT is measurable. It
turns out that this is true for many processes.
Definition. A process X is called progressively measurable with respect to a

filtration At if for all t, the map (s, u) f* Xs(u) from ([0, t] x fl, B([0, t] x At)
into (E, £) is measurable.
A progressively measurable process is adapted. For some processes, the

inverse holds:
Lemma 4.5.3. An adapted process with right or left continuous paths is

progressively measurable.
Proof. Assume right continuity (the argument is similar in the case of left
continuity). Write X as the coordinate process D([0, t],E). Denote the map
(s to) ^ Xs(u>) with Y = Y(s,u). Given a closed ball U G £. We have to
show that Y-X(U) = {(s,w) | Y(s,oj) G U} G B([0,t}) x At- Given k = N,
we define E0,i/ = 0 and inductively for fc > 1 the fc'th hitting time (a
stopping time)
Hk,u(w) = inf{s G Q | Efc_i,uM < s < t, Xs G U }

as well as the fc'th exit time (not necessarily a stopping time)
EktU(v) = mf{s e Q\Hku(uj) <s<t, XS<£U} .

These are countably many measurable maps from D([0, t],E) to [0, t]. Then
by the right-continuity
oo
Y~\U) = \J{(s,uj) I Hk,v(uj) < s < Ek,u(uj)}

k=i
which is in S([0, t]) x At. rj
Proposition 4.5.4. If X is progressively measurable and T is a stopping

time, then XT is ^T-measurable on the set {T < oo}.
Proof. The set {T < oo} is itself in AT. To say that XT is AT- measurable
on this set is equivalent with XT • l{T<t} G A for every t. But the map
S:({T< t},At D {T < *}) - ([0,*],B[0,*])

is measurable because T is a stopping time. This means that the map
uj h- (r(o;),o;) from (ft, At) to ([0,t] x ft,B([0,*]) x At) is measurable and
XT is the composition of this map with X which is B[0, *] x At measurable
by hypothesis. q
Definition. Given a stopping time T and a process X, we define the stopped

process (XT)t(w) = XTAt(v).
Remark. If At is a filtration then AtAT is a filtration since if 7\ and T2 are

stopping times, then Ti A T2 is a stopping time.
Corollary 4.5.5. If X is progressively measurable with respect to At and

T is a stopping time, then (XT)t = XtAT is progressively measurable with
respect to AtAT-
Proof. Because t A T is a stopping time, we have from the previous propo

sition that XT is AtAT measurable.
We know by assumption that 0 : (s,uj) •-» Xs(o;) is measurable. Since also
ip : (s,uj) h(sa T)(uj) is measurable, we know also that the composition
(s,uj) (-▶ Xt(uj) = X^8^(uj) = cj)(ip(s,uj),uj) is measurable. □
4.5. Stopping times 213
Proposition 4.5.6. Every stopping time is the decreasing limit of a sequence

of stopping times taking only finitely many values.
Proof. Given a stopping time T, define the discretisation Tk = +oo if T > k

and Xfc = q2~k if (q - l)2"fc <T <q 2~k, q < 2kk. Each Tk is a stopping
time and Tk decreases to T. ^
Many concepts of classical potential theory can be expressed in an elegant

form in a probabilistic language. We give very briefly some examples with
out proofs, but some hints to the literature.
Let Bt be Brownian motion in Rd and TA the hitting time of a set A C Rd.

Let Dbea domain in Rd with boundary 5(D) such that the Green function
G(x, y) exists in D. Such a domain is then called a Green domain.
Definition. The Green function of a domain D is defined as the fundamental

solution satisfying AG(x, y) = S(x-y), where 5(x-y) is the Dirac measure
at y G D. Having the fundamental solution G, we can solve the Poisson
equation Au = v for a given function v by
u= / G(x,y)-v(y) dy .
Jd
The Green function can be computed using Brownian motion as follows:
/»oo
G(x,y) = / g(t,x,y)dt,
Jo
where for x G D,
[ g(t, x, y) dy = Px[Bt € C, TSD > t]

Jc
and Px is the Wiener measure of Bt starting at the point x.
We can interpret that as follows. To determine G(x,y), consider the killed

Brownian motion Bt starting at x, where T is the hitting time of the bound
ary. G(x, y) is then the probability density, of the particles described by the
Brownian motion.
Definition. The classical Dirichlet problem for a bounded Green domain

D eRd with boundary SD is to find for a given function / G C(S(D)), a
solution u G C(D) such that Au = 0 inside D and
lim u(x) = f(y)
for every y G SD.

This problem can not be solved in general even for domains with piecewise
smooth boundaries if d > 3.
Definition. The following example is called Lebesgue thorn or Lebesgue

spine has been suggested by Lebesgue in 1913. Let D be the inside of a
spherical chamber in which a thorn is punched in. The boundary SD is
held on constant temperature /, where / = 1 at the tip of the thorn y
and zero except in a small neighborhood of y. The temperature u inside
D is a solution of the Dirichlet problem ADu = 0 satisfying the boundary
condition u = f on the boundary SD. But the heat radiated from the thorn
is proportional to its surface area. If the tip is sharp enough, a person sitting
in the chamber will be cold, no matter how close to the heater. This means
limmfx_>yiXeDu(x) < 1 = f(y). (For more details, see [43, 46]).
Because of this problem, one has to modify the question and one says, u is
a solution of a modified Dirichlet problem, if u satisfies ADu = 0 inside D
and limx^yjXeD u(x) = f(y) for all nonsingular points y in the boundary
SD. Irregularity of a point y can be defined analytically but it is equivalent
with Py[TDc > 0] = 1, which means that almost every Brownian particle
starting at y G SD will return to SD after positive time.
Theorem 4.5.7 (Kakutani 1944). The solution of the regularized Dirichlet

problem can be expressed with Brownian motion Bt and the hitting time
T of the boundary:
u(x)=Ex[f(BT)].
In words, the solution u(x) of the Dirichlet problem is the expected value
of the boundary function / at the exit point BT of Brownian motion Bt
starting at x. We have seen in the previous chapter that the discretized
version of this result on a graph is quite easy to prove.
Figure. To solve the Dirichlet

problem in a bounded domain
with Brownian motion, start the
process at the point x and run it
until it reaches the boundary Bt,
then compute /(J9r) and aver
age this random variable over all
paths uj.
4.6. Continuous time martingales 215
Remark. Ikeda has discovered that there exists also a probabilistic method
for solving the classical von Neumann problem in the case d = 2. For more
information about this, one can consult [43, 79]. The process for the von
Neumann problem is not the process of killed Brownian motion, but the
process of reflected Brownian motion.
Remark. Given the Dirichlet Laplacian A of a bounded domain D. One

can compute the heat flow e~tAu by the following formula
(e~tAu)(x) = Ex[u(Bt);t <T] ,

where T is the hitting time of SD for Brownian motion Bt starting at x.
Remark. Let K be a compact subset of a Green domain D. The hitting

probability
p(x)=Px[TK<TSD]
is the equilibrium potential of K relative to D. We give a definition of the
equilibrium potential later. Physically, the equilibrium potential is obtained
by measuring the electrostatic potential, if one is grounding the conducting
boundary and charging the conducting set B with a unit amount of charge.
4.6 Continuous time martingales

Definition. Given a filtration At of the probability space (ft, A, P). A real-
valued process Xt G C1 which is At adapted is called a submartingale, if
E[Xt|A] > Xs, it is called a supermartingale if -X is a submartingale
and a martingale, if it is both a super and sub-martingale. If additionally
Xt G CP for all t, we speak of Cp super or sub-martingales.
We have seen martingales for discrete time already in the last chapter.
Brownian motion gives examples with continuous time.
Proposition 4.6.1. Let Bt be standard Brownian motion. Then BUB\ —t

and eaBt~a tl2 are martingales.
Proof. Bt — Bs is independent of B3. Therefore
E[Bt | As] -Bs = E[Bt - BS\AS] = E[Bt -Bs]=0.

Since by the "extracting knowledge" property
E[BtBs | As] = Bs • E[Bt \ As] = 0 ,

we get
E[B?-t\As}-(B2s-s) = E[B?-B2a\As}-(t-s)
= E[(Bt-Ba)2\As}-(t-s)=0
Since Brownian motion begins at any time s new, we have
E[ea^-B^\AS] = E[eaBt~°] = e^*"5)/2

from which
E[eaB*\As]e-aH!2 = E[eaBs]e~a2s/2
follows. □
As in the discrete case, we remark:
Proposition 4.6.2. If Xt is a £p-martingale, then \Xt\p is a submartingale

for p > 1.
Proof. The conditional Jensen inequality gives
E[\Xt\p\As]>\E[Xt\As]\p = \Xs\p.
□
Example. Let Xn be a sequence of IID exponential distributed random

variables with probability density fx(x) = e~cxc. Let Sn = ££=1 Xk. The
Poisson process Nt with time T = R+ = [0, oo) is defined as
oo
*i = $>»<«.
k=i
It is an example of a martingale which is not continuous, This process
takes values in N and measures, how many jumps are necessary to reach
t. Since E[Nt] = ct, it follows that Nt — ct is a martingale with respect to
the filtration At = <r(N8, s < t). It is a right continuous process. We know
therefore that it is progressively measurable and that for each stopping
time T, also NT is progressively measurable. See [49] or the last chapter
for more information about Poisson processes.
Figure. The Poisson point pro

cess on the line. Nt is the num
ber of events which happen up to s, % % % % % % % % #o §1
time t. It could model for exam- ■ v ■ v ■ „ ■ v ■ v «..■■■■ v ■ „ w .. ■
pie the number Nt of hits onto a
website.
4.7. Doob inequalities 217
Proposition 4.6.3. (Interval theorem) The Poisson process has independent

increments oo
Nt-Ns = ^2 1s<sn<t •
n=l
Moreover, Nt is Poisson distributed with parameter tc:
m - fc] - fc!{^-tc
Proof The proof is done by starting with a Poisson distributed process Nt.
Define then
Sn(uj) = {t\Nt = n,Nt-o = n - 1 }
and show that Xn = Sn - Sn-i are independent random variables with
exponential distribution. ^
Remark. Poisson processes on the lattice Zd are also called Brownian mo
tion on the lattice and can be used to describe Feynman-Kac formulas for
discrete Schrodinger operators. The process is defined as follows: take Xt
as above and define oo
Yt = 22Zklsk<t ,
k=l
where Zn are IID random variables taking values in {ra G Zd\\m\ = 1}.
This means that a particle stays at a lattice site for an exponential time
and jumps then to one of the neighbors of n with equal probability. Let
Pn be the analog of the Wiener measure on right continuous paths on the
lattice and denote with En the expectation. The Feynman-Kac formula for
discrete Schrodinger operators H = H0 + V is
(e-itHu)(n) = e2dtEn[u(Xt)iNie-ô v(*.) *] .
4.7 Doob inequalities

We have already established inequalities of Doob for discrete times T = N.
By a limiting argument, they hold also for right-continuous submartingales.
Theorem 4.7.1 (Doob's submartingale inequality). Let X be a non-negative

right continuous submartingale with time T = [a, b]. For any e > 0
e • P[ sup Xt > e] < E[Xb; { sup Xt > e}] < E[Xb] .

a<t e] < E[X6;{sup Xt > c}] < E[Xb] .

teDn teDn
since E[Xt] is nondecreasing in t. Going to the limit n -▶ oo gives the claim
with T = D. Since X is right continuous, we get the claim for T = [a, 6]. □
One often applies this inequality to the non-negative submartingale |X| if

X is a martingale.
Theorem 4.7.2 (Doob's Lp inequality). Fix p > 1 and q satisfying p~l +

<T =1- Given a non-negative right-continuous submartingale X with
time T = [a, b] which is bounded in Cp. Then X* = suptGT Xt is in Cp and
satisfies
iMniP<g.sup||xtnP.
Proof Take a countable subset D of T and choose an increasing sequence

Dn of finite sets such that |Jn Dn = D.
We had
I sup Xt\\ <q- sup \\Xt\\p
teDn teDn
Going to the limit gives
|supX*|| <q-sup\\Xt\\p .
t€D teD
Since D is dense and X is right continuous we can replace D by T. D
The following inequality measures, how big is the probability that one-
dimensional Brownian motion will leave the cone {(t, x), \x\ < a • t}.
Theorem 4.7.3 (Exponential inequality). St = sup0<s<t Bs satisfies for any

a>0
P[St > a • t] < e-aH/2 .
Proof. We have seen in proposition (4.6.1) that Mt = eaBt~ V is a mar

tingale. It is nonnegative. Since
exp(aSt --tt)< exp(supJBs -* ^—) <8 < tsupexp(£s -~)=supMs,

Z S<t * S<t
4.8. Khintchine's law ot the iterated logarithm 219
we get with Doob's submartingale inequality (4.7.1)
P[5t > at] < P[supMs > eaat"^]

s<t
a2t
< exp(-aat + — )E[Mt] .
The result follows from E[Bt] = E[J30] = 1 and infa>0 exp(-aat 4- ^) =

exp(-^). D
An other corollary of Doob's maximal inequality will also be useful.
Corollary 4.7.4. For a, b > 0,
P[ sup (Bs -^-)>(3]< e-"P .

«€[0,1] l
Proof.
P[sup(£s-^)>/?] < P[ sup (Bs - %) > 0\

«€[0,1] Z «€[0,1] Z
= P[ sup (eQBs"V)>e^]
*€[0,1]
= P[ sup Ms > e0a]
s6[0,l]
< eâ sup E[MS] = e_/3a
«€[0,1]
since E[MS] = 1 for all 5. □
4.8 Khintchine's law of the iterated logarithm

Khinchine's law of the iterated logarithm for Brownian motion gives a pre
cise statement about how one-dimensional Brownian motion oscillates in a
neighborhood of the origin. As in the law of the iterated logarithm, define
A(t) = V^logllog*!
Theorem 4.8.1 (Law of iterated logarithm for Brownian motion).
P[limsup-^-
L t - > o =A (1]t ) = J 1,' P[liminf-^-
L t - o A =( t -1]
) J= 1
Proof. The second statement follows from the first by changing Bt to -Bt.
(i) limsups^0 -j&hj < 1 almost everywhere:

Take 6,5 e (0,1) and define
an = (i + 8)6-nA(en), pn = ^p-.
We have an(3n = loglog(0n)(l + 5) = log(n) log(0). From corollary (4.7.4),

we get
P[sup(Bs - » > pn] < e~a^ = W"1+i» .
s<l ^
The Borel-Cantelli lemma assures

Pfliminf sup(Ss - ^) < &»] = 1
which means that for almost every u, there is n0(w) such that for n > n0(ui)
andse [O,^""1),
B8(u>) < aj- +{3n< aj-^- + 0n = {^- + \)A(9n) .
Since A is increasing on a sufficiently small interval [0,a), we have for

sufficiently large n and s G (0n,0n_1]
B.(«) < (fi^ + j)A(«) •

In the limit 0 -> 1 and 6 -> 0, we get the claim.
(ii) limsups^0 ^ > 1 almost everywhere.

For 0G (0,1), the sets
An = {Ben - Ben+i > (1 - y/9)A(6n)}
are independent and since B^ - B„„+i is Gaussian we have
/ 2 _ ^ _ . fl r~a2/2
'[An] = / e-' ^ a' + l*
with a = (1 - v/0)A(0n) < Kn"a with some constants K and a < 1.
Therefore En1^"] = °° and by the second Borel-Cantelli lemma,
Ben > (1 - V^)A(0n) + Ben+i (4.1)
for infinitely many n. Since -B is also Brownian motion, we know from (i)
that
-B*.+i < 2A(0n+1) (4-2)
for sufficiently large n. Using these two inequalities (4.1) and (4.2) and
A(0"+1) < 2v/0A(0n) for large enough n, we get
B6n > (1 - v/0)A(0n) - 4A(0n+1) > A(0n)(l - V5 - 4>/0)

4.8. Khintchine's law of the iterated logarithm 221
for infinitely many n and therefore
lim inf —7-7 > limsup ttttt > 1 — 5v#.

t-o A(*) " „-<xf A(0»)
The claim follows for 0 -> 0. □
Remark. This statement shows also that Bt changes sign infinitely often
for t —* 0 and that Brownian motion is recurrent in one dimension. One
could show more, namely that the set {Bt = 0 } is a nonempty perfect set
with Hausdorff dimension 1/2 which is in particularly uncountable.
By time inversion, one gets the law of iterated logarithm near infinity:
Corollary 4.8.2.
Pflimsup-ryL = 1] = 1, Pfliminf —±r = -1] = 1
Proof. Since Bt = £.Bi/t (with Bq = 0) is a Brownian motion, we have with

s = 1/t
Bs B\is
1 = lim sup . , . = lim sups . , ■
s_+o A(s) sô A(s)
= lim sup = lim sup ■
T£*t\(i/t) ;rÂ(i)"
The other statement follows again by reflection. □
Corollary 4.8.3. For d-dimensional Brownian motion, one has
Proof. Let e be a unit vector in Rd. Then Bt • e is a 1-dimensional Brown

ian motion since Bt was defined as the product of d orthogonal Brownian
motions. From the previous theorem, we have
Ppimsup-^-f
t — 0 A ( t ) = 1] = 1.
Since Bt • e < \Bt\, we know that the limsup is > 1. This is true for all
unit vectors and we can even get it simultaneously for a dense set {en}n€N
of unit vectors in the unit sphere. Assume the limsup is 1 + e > 1. Then,
there exists en such that
in contradiction to the law of iterated logarithm for Brownian motion.

Therefore, we have limsup = 1. By reflection symmetry, lim inf = — 1. □
Remark. It follows that in d dimensions, the set of limit points of Bt/A(t)

for t —▶ 0 is the entire unit ball {\v\ < 1}.
4.9 The theorem of Dynkin-Hunt

Definition. Denote by I(k, n) the interval [^=r, £). If T is a stopping time,
then T^ denotes its discretisation
r(n)M = Ei/(M)(rM)^
which is again a stopping time. Define also:
AT + = { A e A o o \ A n { T < t } e A u V t } .
The next theorem tells that Brownian motion starts afresh at stopping
times.
Theorem 4.9.1 (Dynkin-Hunt). Let T be a stopping time for Brownian

motion, then Bt = Bt4-T — Bt is Brownian motion when conditioned to
{T < oo} and Bt is independent of At+ when conditioned to {T < oo}.
Proof. Let A be the set {T < oo}. The theorem says that for every function
f(Bt) = g(Bt+u, Bt+t2, •.., Bt+tn)
with g e C(Rn)
E[f(Bt)lA]=E[f(Bt)]'P[A]
and that for every set C G At+
E[f(Bt)lAnc] ■ P[A] = E[f(Bt)lA} ■ P[A n C] .
This two statements are equivalent to the statement that for every C G At+
E[f(Bt) ■ lAnc] = E[f(Bt)} ■ P[A n C] .

4.10. Self-intersection of Brownian motion 223
Let T(n) be the discretisation of the stopping time T and An = {T(n) < 00}
as well as An,k = {T^ = fc/2n}. Using A = {T < oo},P[|jr=i An,knC] ->
P[AflC] for n —> 00, we compute
E[/(Bt)Unc] = lim
n—>ooE[f(BTin))lAnnc]
00
= lim y>[/(£fe/2")Un,fcnd
n—»oo ^—'
fc=0
00
= lim VE[/(B0)]-PK,fenC]
n.—▶n'o'^
n—»oo
fc=0
= E[/(B0)] lim P[f)4,fcnC]

—▶o
n fc=l
= E[/(B0)Unc]
= E\f(B0)]-P[AnC]
= E[/(J5t)]-P[AHC]
Remark. If T < 00 almost everywhere, no conditioning is necessary and

Bt-\-T — Bt is again Brownian motion.
Theorem 4.9.2 (Blumental's zero-one law). For every set A G Ao+ we have
P[A] = 0 or P[A] = 1.
Proof. Take the stopping time T which is identically 0. Now B = BtA-T -

Bt = B. By Dynkin-Hunt's result, we know that B = B is independent of
BT+ = A0+. Since every C G Aq+ is {#s> 5 > 0} measurable, we know that
Aq+ is independent to itself. D
Remark. This zero-one law can be used to define regular points on the
boundary of a domain D eRd. Given a point y G SD. We say it is regular,
if Py[T5D > 0] = 0 and irregular Py[TSD > 0] = 1. This definition turns
out to be equivalent to the classical definition in potential theory: a point
y G SD is irregular if and only if there exists a barrier function / : N —▶ R
in a neighborhood Nofy.A barrier function is defined as a negative sub-
harmonic function on int(JV n D) satisfying f(x) -> 0 for x -> y within
D.
4.10 Self-intersection of Brownian motion

Our aim is to prove the following theorem:
Theorem 4.10.1 (Self intersections of random walk). For d < 3, Brownian

motion has infinitely many self intersections with probability 1.
Remark. Kakutani, Dvoretsky and Erdos have shown that for d > 3, there
are no self-intersections with probability 1. It is known that for d < 2, there
are infinitely many n-fold points and for d > 3, there are no triple points.
Proposition 4.10.2. Let K be a compact subset of Rd and T the hitting time

of K with respect to Brownian motion starting at y. The hitting probability
h(y) = P[y + Bs G K, T < s < oo] is a harmonic function on Rd \ K.
Proof. Let Ts be the hitting time of Ss = {\x - y\ = S}. By the law of

iterated logarithm, we have Ts < oo almost everywhere. By Dynkin-Hunt,
we know that Bt = Bt+Ts - Bt is again Brownian motion.
If S is small enough, then y + Bs £ K for t < T6. The random variable

BTs G Ss has a uniform distribution on Ss because Brownian motion is
rotational symmetric. We have therefore
h(y) = F[y + BseK,s>Ts]

= P[y + BTs+BeK]
= h(y + x) dp(x) ,
where p is the normalized Lebesgue measure on 5,5. This equality for small
e n o u g h S i s t h e d e f i n i t i o n o f h a r m o n i c i t y. □
Proposition 4.10.3. Let K be a countable union of closed balls. Then

h(K,y) -+1 for y-^K.
Proof, (i) We show the claim first for one ball K = Br(z) and let R = \z-y\.
By Brownian scaling Bt~c- Bt/C2. The hitting probability of K can only
be a function f(r/R) of r/R:
h(y,K) = P[y + BseK,T<s] = P[cy + Bs/c2 G cK,TK < s]

= P[cy-r Bs/C2 G cK, TcK < s/c2]
= P[cy + B§,TcK<s]
= h(cy, cK) .
We have to show therefore that f(x) —▶ 1 as x —> 1. By translation invari-
ance, we can fix y = y0 = (1,0,..., 0) and change Ka, which is a ball of
radius a around (—a, 0,...). We have
h(y0,Ka) = f(a/(l + a))

and take therefore the limit a —> oo
lim f(x) = lim %0, Ka) = h(y0,\\Ka)

x—>1 a—+oo >^
= E[inf (Bs)x < -1] = 1
s>0
because of the law of iterated logarithm.

(ii) Given yn -> y0 G K. Then y0 € K0 for some ball if0-
lim inf h(yn,K) > lim h(yn,K0) = 1
n—>oo n—ôo
by (i). □
Definition. Let p be a probability measure on R3. Define the potential
theoretical energy of p as
J(M) = / \z- y\~X dp(x) dp(y) .

7]R3 JR3
Given a compact set K cR3, the capacity of K is defined as
/x€M(K)
where M(X) is the set of probability measures on K. A measure on K

minimizing the energy is called an equilibrium measure.
Remark. This definitions can be done in any dimension. In the case d =

2, one replaces \x - y\~l by log|z - y\~l. In the case d > 3, one takes
k - y\~{d~2)- The capacity is for d = 2 defined as exp(- infM I(p)) and for
d>3as(infM/(^))-(d-2).
Definition. We say a measure pn on Rd converges weakly to //, if for all con
tinuous functions /, / / dpn -* / / dp. The set of all probability measures
on a compact subset E of Rd is known to be compact.
The next proposition is part of Frostman's fundamental theorem of poten
tial theory. For detailed proofs, we refer to [39, 80].
Proposition 4.10.4. For every compact set K C Rd, there exists an equilib
rium measure p on K and the equilibrium potential / \x - y|~(d~2) dp(y)
rsp. /logflx - y\~l) dp(y) takes the value C(K)'1 on the support K* of
p.
Proof, (i) (Lower semicontinuity of energy) If pn converges to /x, then
lim inf I(pn) > I(p) .

n—▶o
(ii) (Existence of equilibrium measure) The existence of an equilibrium mea

sure p follows from the compactness of the set of probability measures on
K and the lower semicontinuity of the energy since a lower semi-continuous
function takes a minimum on a compact space. Take a sequence pn such
that
H^n)-* ini I(p).
Then pn has an accumulation point /i and I(p) < infAt€M(x) /(/i).
(iii) (Value of capacity) If the potential <f>(x) belonging to p is constant on

K, then it must take the value C(K)~l since
/ <f>(x) dp(x) = I(u) .
(iv) (Constancy of capacity) Assume the potential is not constant C(K)~X

on K*. By constructing a new measure on K* one shows then that one can
strictly decrease the energy. This is physically evident if we think of </> as
the potential of a charge distribution fi on the set K. □
Corollary 4.10.5. Let \x be the equilibrium distribution on K. Then
h(y,K) = <t>^C(K)
and therefore h(y, K) > C(K) • inf^^ \x - y\~l.
Proof. Assume first K is a countable union of balls. According to proposi

tion (4.10.2) and proposition (4.10.3), both functions h and <^ • C(K) are
harmonic, zero at oo and equal to 1 on S(K). They must therefore be equal.
For a general compact set K, let {yn} be a dense set in K and let Ke =
\Jn Be(yn). One can pass to the limit e —> 0. Both h(y, Kt) —▶ h(y, K) and
infx€K€ \x — y\~l —▶ infx€K \x - y\~l are clear. The statement C(K€) —>
C(K) follows from the upper semicontinuity of the capacity: if Gn is a se
quence of open sets with DGn = 22, then C(Gn) —▶ C(E).
The upper semicontinuity of the capacity follows from the lower semicon
tinuity of the e n e r g y. □
Proposition 4.10.6. Assume, the dimension d = 3. For any interval J =

[a, b], the set
Bj(uj) = {Bt(uj)\te[a,b]}
has positive capacity for almost all uj.
Proof. We have to find a probability measure p(u) on Bûj) such that its
energy I(p(w)) is finite almost everywhere. Define such a measure by
{gG[a,b]15,GA}
<W) = l (6^) 1-
Then
I(u) = I [\x- Vl'1 dp(x)du.(y) = J J (b- ay^Bs - Bt\ X dsdt.
To see the claim we have to show that this is finite almost everywhere, we
integrate over fl which is by Fubini
E[I(p)] = f I (b - a)~lE[\Bs - Btl'1] dsdt

JaJa
which is finite since Bs - Bt has the same distribution as y/s - tB\ by

Brownian scaling and since Eîl"1] = / \x\-xe~^l2 dx < oo in dimen
sion d > 2 and £ f* Vs~=t ds dt < oo. □
Now we prove the theorem
Proof. We have only to show that in the case d = 3. Because Brownian

motion projected to the plane is two dimensional Brownian and to the line
is one dimensional Brownian motion, the result in smaller dimensions fol
low.
(i)a = PBJt€[o,i],.>2B* = BJ>0-

Proof. Let K be the set Ug[o,i] Bt> We know that it has positive capacity
almost everywhere and that therefore h(Bs,K) > 0 almost everywhere.
But h(Bs,K) = a since Bs+2 - Bs is Brownian motion independent of
B8,0<s<l.
(ii) aT = P[Ut€[o,il,2<TB* = BS]>0 for some T > 0. Proof. Clear since

ost —> ol for T —> oo.
(iii) Proof of the claim. Define the random variables Xn = lcn with
Cn = {uj\Bt = Bs, for some te [nT,nT+l],se [nT + 2, (n + 1)T] }.
They are independent and by the strong law of large numbers YLn X™ = °°
almost everywhere. '-'
22** Chapter 4. Continuous Stochastic Processes
Corollary 4.10.7. Any point Bs(uj) is an accumulation point of self-crossings

of {Bt(uj)}t>0.
Proof. Again, we have only to treat the three dimensional case. Let T > 0
be such that
aT = P[ |J Bt = B8]>0
t€[0,l],2<T
in the proof of the theorem. By scaling,
P[^ = Bs|tG[0,^],5G[2/3,T/?]]
is independent of /3. We have thus self-intersections of the random walk in
any interval [0, b] and by translation in any interval [a, b]. □
4.11 Recurrence of Brownian motion

We show in this section that like its discrete brother, the random walk,
Brownian motion is transient in dimensions d > 3 and recurrent in dimen
sions d < 2.
Lemma 4.11.1. Let T be a finite stopping time and Rt(uj) be a rotation in

Rd which turns BT(u) onto the first coordinate axis
Rt(u)Bt(uj) = (\Bt(uj)\,0,...Q).
Then Bt = RT(Bt-\-T — Bt) is again Brownian motion.
Proof. By the Dynkin-Hunt theorem, Bt = Bt+T - BT is Brownian motion

and independent of At- By checking the definitions of Brownian motion,
it follows that if B is Brownian motion, also R(x)Bt is Brownian motion,
if R(x) is a random rotation on Rd independent of Bt. Since RT is AT
measurable and Bt is independent of At, the claim follows. □
Lemma 4.11.2. Let Kr be the ball of radius r centered at 0 G Rd with

d>3. We have for y (£ Kr
d-2
h(y,Kr) = (r/\y\)
4 . 11 . Recurrence of Brownian motion 229
Proof. Both h(y, Kr) and (r/\y\)d~2 are harmonic functions which are 1 at
S K r a n d z e r o a t i n f i n i t y. T h e y a r e t h e s a m e . □
Theorem 4.11.3 (Escape of Brownian motion in three dimensions). For

d > 3, we have limôo \Bt\ = oo almost surely.
Proof. Define a sequence of stopping times Tn by

Tn = inf{s>0||£s| = 2n},
which is finite almost everywhere because of the law of iterated logarithm.
We know from the lemma (4.11.1) that
Bt = RTn(Bt+Tn - BTn)
is a copy of Brownian motion. Clearly also |i?Tn | = 2n.
We have Bs G Kr(0) = {\x\ < r} for some s > Tn if and only if Bt G
(2n, 0..., 0) 4- Kr(0) for some t > 0.
Therefore using the previous lemma
P[BS eKr(0);s> Tn] = P[Bt G (2n, 0..., 0) + ffr(0); t > 0] = (^)d~2
which implies in the case r2_n < 1 by the Borel-Cantelli lemma that for
almost all uj, Bs(uj) > r for s > Tn. Since Tn is finite almost everywhere,
we get lim infs \Ba\ > r. Since r is arbitrary, the claim follows. □
Brownian motion is recurrent in dimensions d < 2. In the case d = 1, this

follows readily from the law of iterated logarithm. First a lemma
Lemma 4.11.4. In dimensions d = 2, almost every path of Brownian motion

hits a ball Kr if r > 0: one has h(y, K) = l.
Proof. We know that h(y) = h(y, K) is harmonic and equal to 1 on SK. It

is also rotational invariant and therefore h(y) = a + 61og \y\. Since h G [0,1]
we have h(y) = a and so a = 1. d
Theorem 4.11.5 (Recurrence of Brownian motion in 1 or 2 dimensions). Let

d < 2 and S be an open nonempty set in Rd. Then the Lebesgue measure
of {t\Bte S} is infinite.
Proof. It suffices to take S = Kr(x0), a ball of radius r around x0. Since

by the previous lemma, Brownian motion hits every ball almost surely, we
can assume that x0 = 0 and by scaling that r = 1.
Define inductively a sequence of hitting or leaving times Tn, Sn of the
annulus {1/2 < \x\ < 2}, where Ti = inf{£ | \Bt\ = 2} and
S n = i n f { * > r n | | fl t | = l / 2 }
Tn = inf{*>Sn_i||Bt| = 2}.
These are finite stopping times. The Dynkin-Hunt theorem shows that Sn -
Tn and Tn - Sn-i are two mutually independent families of IID random
variables. The Lebesgue measures Yn = \In\ of the time intervals
In = {t\\Bt\<l, Tn<t<Tn+1},
are independent random variables. Therefore, also Xn = min(l,yn) are
independent bounded IID random variables. By the law of large numbers,
J2n Xn = oo which implies J2n Yn = oo and the claim follows from
|{*e[o,oo)||ft|<i}|>2rn.
n
D
Remark. Brownian motion in Rd can be defined as a diffusion on Rd with
generator A/2, where A is the Laplacian on Rd. A generalization of Brow
nian motion to manifolds can be done using the diffusion processes with
respect to the Laplace-Beltrami operator. Like this, one can define Brown
ian motion on the torus or on the sphere for example. See [57].
4.12 Feynman-Kac formula

In quantum mechanics, the Schrodinger equation ihii = Hu defines the
evolution of the wave function u(t) = e~itH/hu(0) in a Hilbert space H. The
operator H is the Hamiltonian of the system. We assume, it is a Schrodinger
operator H = H0 + V, where H0 = -A/2 is the Hamiltonian of a free
particle and V : Rd -> R is the potential. The free operator H0 already is
not defined on the whole Hilbert space Ti = L2(Rd) and one restricts H to
a vector space D(H) called domain containing the in H dense set C%°(Rd)
of all smooth functions which are zero at infinity. Define
D(A*) = {u G H | v h-> (Av, u) is a bounded linear functional on D(A)}.

Hue D(A*), then there exists a unique function w = A*u GH such that
(Av,u) = (v,w) for all u G D(A). This defines the adjoint A* of A with
domain D(A*).
Definition. A linear operator A : D(A) C H —▶ H is called symmetric if

(Au, v) = (u, Av) for all u,v e D(A) and self-adjoint, if it is symmetric and
D(A) = D(A*).
4.12. Feynman-Kac formula 231
Definition. A sequence of bounded linear operators An converges strongly
to A, if Anu -> Au for all u G Ti. One writes A = s- limnôo An.
Define eA = 1 4- A + A2/2! + A3/3! + • • •. We will use the fact that a

self-adjoint operator defines a one parameter family of unitary operators
t h-> eitA which is strongly continuous. Moreover, eltA leaves the domain
D(A) of A invariant. For more details, see [81, 7].
Theorem 4.12.1 (Trotter product formula). Given self-adjoint operators

A,B defined on D(A),D(B) C H. Assume A + B is self-adjoint on D =
D(A)C)D(B), then
eit(A+B) =s_ Um (eitA/neitB/ny #
n—▶oo
If A, B are bounded from below, then
e-t(A+B) = s _ lim {e-tA/ne-tB/n)n
Proof. Define
5t = ^(A+B)^ = eitA^Wf = eitBÛt = ytWt
and vt = Stv for v G D. Because A + B is self-adjoint on £>, one has vt G £>.

Use a telescopic sum to estimate
\\{St-U?/nM = \\T,Uhn(St/n-Ut/n)S^^
3=0
< n SUp \\(St/n - Ut/n)Vs\\ •
0<s<t
We have to show that this goes to zero for n -^ oo. Given u G D =

D(A)HD(B),
lim s~ u = i(A + B)u = lim — u

s—o s s^° $
so that for each u £ D
n
—l▶
im
o n-\\(St/n-Ut/n)u\\ = 0. (4.3)
The linear space D with norm |||u||| = \\{A + B)u\\ + \\u\\ is a Banach
space since A + B is self-adjoint on D and therefore closed. We have a
bounded family {n(St/n - Ut/n)}neN of bounded operators from D to H.
The principle of uniform boundedness states that
\\n{St/n-Ut/n)u\\<C-\\\u\\\.
An e/3 argument shows that the limit (4.3) exists uniformly on compact
subsets of£> and especially on {vs}ae[0>t] c D and so nsup0<s<t \\(St/n-
Ut/n)va11 - U. 1 he second statement is proved in exactly the same way. □
Remark. Trotter's product formula generalizes the Lie product formula
Ji™^-) exP(f))" = exp(,4 + B)

for finite dimensional matrices A,B, which is a special case.
Corollary 4.12.2. (Feynman 1948) Assume H = H0 + V is self-adjoint on

D(H). Then
e - i t H U{X0) = n1™^^)-"72 / n ei5"(x°'Xl'X2 *»•«>«(*») dXl... dxn
where
t ^ 1 \xi -Xj-x
Sn(xo,Xlt... tXn,t) = ± £ -2{^-^-? ~ V(Xi)
i=l '
Proof. (Nelson) From u = -iH0u, we get by Fourier transform fi = i^-u

which gives ut(k) = eXp(iUpt)u0(k) and by inverse Fourier transform
e-itH°u(x) = ut{x) = {2mt)-d'2 f eikû(y) dy .

JRd
The Trotter product formula
e-it{HQ+V) z=s_ jim /eittf0/neitV/nxn .

n—>oo '
gives now the claim. rj

Remark. We did not specify the set of potentials, for which H0 + V can be
made self-adjoint. For example, V G Cg°(R^) is enough or V G L2(R3) n
L°°(R3) in three dimensions.
We have seen in the above proof that e~itH° has the integral kernel Pt(x, y) =
(2mt)-dl2e% V . The same Fourier calculation shows that e~tH° has the
integral kernel
Pt(x,y) = M-ê-1^ ,
where gt is the density of a Gaussian random variable with variance t.
Note that even tf u € L2(Rd) is only denned almost everywhere, the func
tion ut{x) = e tH°u(x) = fPt(x - y)u{y)dy is continuous and denned
4.12. Feynman-Kac formula 233
everywhere.
Lemma 4.12.3. Given /i,..., fn G L°°(Rd)nL2(Rd) and 0 < si < • • • < sn.
Then
(e-ô/i---e-t"i/o/n)(0) = |/i(BSl).-./n(^Jd5,
where t\ = S\,U = Si - Si-\,i > 2 and the fa on the left hand side are
understood as multiplication operators on L2(Rd).
Proof. Since BSl, B82 - B8l,... BSn - B8n_x are mutually independent
Gaussian random variables of variance ti, t2, •.. ,tn, their joint distribu
tion is
Pt1(0,»i)Pta(0,^)...Ptn(0,yn)d»
which is after a change of variables y\ = #i, yi = Xi — Xi-\
Ptl(0,xi)Pt2(xi,x2)...Ptn(xn_i,xn) dx .
Therefore,
Jfi(BSl)-'fn(Bsn)dB
[ Ptl(0,yi)Pta(0,|fc) • • • Ptn(0,yn)/i(yi) • • • fn(Vn) dy
J(Rd)n
= / Ptl(01Xi)Pt2(x1,X2) . ..Ptn(xn-i,Xn)fi(xi) . . . /n(Xn) dx
J(Rd)n
= (e"tlHo/i---e"tnHo/n)(0).
Denote by df? the Wiener measure on C([0, oo),Rd) and with dx the
Lebesgue measure on Rd. We define also an extended Wiener measure
dW = dxx dB on C([0, oo),Rd) on all paths s^Ws=x + Bs starting at
x eRd.
Corollary 4.12.4. Given f0,fi,...Jn e L°°(Rd) n L2(Rd) and 0 < sx <

• • • < sn. Then
j /oW • • • fn(WSn) dW = (To, e-tl/fo/i • • • e-*-Ho/n) •

Proof, (i) Case s0 = 0. From the above lemma, we have after the dB
integration that
[fo(WSo)---fn(WSn)dW = f fo{x)e-t^h{x)...e-t^fn{x)dx
= (7o,e-tlHofi---e-t"H"fn).
(ii) In the case s0 > 0 we have from (i) and the dominated convergence
theorem
/ fo(WSo) ■ ■ ■ fn(WSn) dW
= lim / l{\x\<R}{W0)
fo(WSoJ---fn(WSn)dW
= l™Jfoe-SoH°l{\x\<R}, e-tlHoh ■ ■ ■ e-f""°/„(x))
= (fo,e-tlHoh---e-tnHofn).
D
We prove now the Feynman-Kac formula for Schrodinger operators of the
form H = H0 + V with V € C03O(Rd). Because V is continuous, the integral
/0 V(Ws(w)) ds can be taken for each w as a limit of Riemann sums and
/0 V(Wa) ds certainly is a random variable.
Theorem 4.12.5 (Feynman-Kac formula). Given H = H0 + V with V e

Cg>(Rd), then
(/, e-tHg) = J7(Wo)9(Wt)e- SS v(w°» d° dW .
Proof. (Nelson) By the Trotter product formula
(f,e-tMg) = n—*oo
lim (/, {e-tH"'ne-tv'n)ng)
so that by corollary (4.12.4)
(/, e~tHg) n=— lim

^00
[j(Wo)g(Wt)
J
exp(--
fi
T V(Wtj/n)) dW (4.4)
and since s i—▶ Ws is continuous, we have almost everywhere
tn-l rt
n£
■n„ ^/n)- / V(Ws)ds.
3=0
4.13. The quantum mechanical oscillator 235
The integrand on the right hand side of (4.4) is dominated by
l/(Wb)|-|ff(Wi)|-^l|vr|l~
which is in Ll(dW) because again by corollary (4.12.4),
|/(Wb)| • \9(Wt)\ dW = (|/|,e-tH«\g\) < oo .

/ '
The dominated convergence theorem leads us now to the claim. □
Remark. The formula can be extended to larger classes of potentials like
potentials V which are locally in L1. The selfadjointness, which needed in
Trotter's product formula, is assured if V G L2 n Lp with p > d/2. Also
Trotter's product formula allows further generalizations [93, 31].
Why is the Feynman-Kac formula useful?

• One can use Brownian motion to study Schrodinger semigroups. It al
lows for example to give an easy proof of the ArcSin-law for Brownian
motion.
• One can treat operators with magnetic fields in a unified way.
• Functional integration is a way of quantization which generalizes to
more situations.
• It is useful to study ground states and ground state energies under
perturbations.
• One can study the classical limit ft —▶ 0.
4.13 The quantum mechanical oscillator

The one-dimensional Schrodinger operator
Id2 1 o 1
H = H° + U = -2d^ + 2X-2
is the Hamiltonian of the quantum mechanical oscillator. It is a quantum

mechanical system which can be solved explicitly like its classical analog,
which has the Hamiltonian H(x,p) — \p2 4- \x2 - \.
One can write

H = AA* - 1 = A* A ,
with
A* = 4=(s-T-)> A=4=(*-+T-)-
y/2 dxh y/2 dx'
The first order operator A* is also called particle creation operator and A,
the particle annihilation operator. The space Cq° of smooth functions of
compact support is dense in L2(R). Because for all u,v G Cq°(R)
(Au,v) = (u,A*v)
the two operators are adjoint to each other. The vector
fi0 = -x2/2
rV4
is a unit vector because fig is the density of a JV(0, l/\/2) distributed ran
dom variable. Because ^fi0 = 0, it is an eigenvector of H = A* A with
eigenvalue 1/2. It is called the ground state or vacuum state describing the
system with no particle. Define inductively the n-particle states
fi n = - ^ * fi n _ !
by creating an additional particle from the (n - l)-particle state fin_! .
Figure. The first Hermit func

tions fin. They are unit vectors
in L2(R) defined by
Hn(x)vo(x)
fin(x) =
V2nnl
where Hn(x) are Hermite poly
nomials, H0(x) = l,Hi(x) =
2x,H2(x) = 4x2 - 2,H3(x) =
Sx3 - \2x,....
Theorem 4.13.1 (Quantum mechanical oscillator). The following properties

hold:
a) The functions are orthonormal (fin,fim) = 5n,m.
b) ,4fin = y/n£ln_uA*nn = v/n+Tfin+1.
c) (n ~~ 2) are tne eigenvalues of H
H = (A*A-\)Sln = (n-l-)Sln
d) The functions fin form a basis in L2(R).
Proof. Denote by [A, B] = AB - BA the commutator of two operators A

and B. We check first by induction the formula
*\n—1
[A,(AT} = n-(A*)
4.13. The quantum mechanical oscillator 237
For n = 1, this means [A, A*] = 1. The induction step is
[A,{A*)n\ = [MAT~1]A* + {A*)n-1[A,A*]

= (n - 1)(A*)B_1 + {AT"1 = "(A*)"-1 .
a) Also
((A*)nfi0,(A*)mfi0) = n!5mn.
can be proven by induction. For n = 0 it follows from the fact that fio is
normalized. The induction step uses [A, (A*)71} = n- (A*)n~l and Afi0 = 0:
((A*)nfio,04*rfio) = (A(A*)nno(A*)m-lno)
= n ^ r - ^ c ^ r - ' fi o ) .
If n < m, then we get from this 0 after n steps, while in the case n = ra,
we obtain ((A*)nfi0, (A*)nfi0) = n • ((A*)n_1"o, (A*)n_1"o), which is by
induction n(n — l)!(Sn-i,n-i = n!.
b) A* fin = \/n + l • fin+i is the definition of fin.
AQn = -^=A(A*)nn0 = 4=nfi0 = VnVn-i •

Vn! Vn!
c) This follows from b) and the definition fin = Â*fin-i-
d) Part a) shows that {fin}n°=o 'lt is an orthonormal set in L2(R). In order

to show that they span L2(E), we have to verify that they span the dense
set
S = {/ g C^(R) | xmfM(x) -+ 0, |x| -+ oo,Vra,n G N }
called the Schwarz space. The reason is that by the Hahn-Banach theorem,
a function / must be zero in L2(R) if it is orthogonal to a dense set. So,
lets assume (/, fin) = 0 for all n. Because A* + A = y/2x
0 = VnlF (/, fin) = (/, (ATfio) = (/, (** + ^)n"o) = 2n/2 (/, a?nfi0)
we have
f(x)ilQ(x)eikx dx
/ o-oo
o
= (/,fioeite) = (/,£^«o)
n>0
n!
n>0
and so /fio = 0. Since fio0*0 is positive for all x, we must have / = 0. This
finishes the proof that we have a complete basis. □
Remark. This gives a complete solution to the quantum mechanical har
monic oscillator. With the eigenvalues {An = n-1/2}£L0 and the complete
set of eigenvectors fin one can solve the Schrodinger equation
d
ih—u = Hu
dt
by writing the function u(x) = En=oun^n(x) as a sum of eigenfuctions,
where un = (u, fin). The solution of the Schrodinger equation is
oo
«(t,a;) = yjune<ft<n-1/2)*fin(x).
71=0
Remark. The formalism of particle creation and annihilation operators

can be extended to some potentials of the form U(x) = q2(x) - q'(x) the
operator H = -D2/2 + 17/2 can then be written as H = A*A, where
The oscillator is the special case q(x) = x. See [12]. The Backlund transfor
mation H = A" A h-> H = A A* is in the case of the harmonic oscillator the
map H >-▶ H + 1 has the effect that it replaces U with U = U - d2 logfi0,
where fi0 is the lowest eigenvalue. The new operator H has the same spec
trum as H except that the lowest eigenvalue is removed. This procedure
can be reversed and to create "soliton potentials" out of the vacuum. It
is also natural to use the language of super-symmetry as introduced by
Witten: take two copies Hf®Hh of the Hilbert space where "/" stands for
Fermion and "6" for Boson. With
0 -A* 1 0
Q= A 0 ,P = 0 -1
one can write H 0 H = Q2, P2 = 1, QP + PQ = 0 and one says (#, P, Q)

has super-symmetry. The operator Q is also called a Dirac operator. A
super-symmetric system has the property that nonzero eigenvalues have
the same number of bosonic and fermionic eigenstates. This implies that H
has the same spectrum as H except that lowest eigenvalue can disappear.
Remark. In quantum field theory, there exists a process called canonical

quantization, where a quantum mechanical system is extended to a quan
tum field. Particle annihilation and creation operators play an important
role.
4.14 Feynman-Kac for the oscillator

We want to treat perturbations L = L0 + V of the harmonic oscillator
L0 with an similar Feynman-Kac formula. The calculation of the integral
4.14. Feynman-Kac for the oscillator 239
kernel pt(x, y) of e~tL° satisfying
{e-tL°f){x)= fPt(x,y)f(y) dy
JR
is slightly more involved than in the case of the free Laplacian. Let fio be
the ground state of L0 as in the last section.
Lemma 4.14.1. Given /0, /i,..., fn G L°°(R) and -oo < s0 < si < • • • <
sn < oo. Then
(fio, /oe-tlLo/i • • • e-'-^/nOo) = JMQ,0) • • • fn(QsJ dQ ,
where to = so,U — Si — Sj_i,i > 1.
Proof. The Trotter product formula for L0 = Ho + U gives
(fi0,/oe-tlLo/i---e-t"Lo/nfio)
lim (fio,/o(e-'lifo/mie-'lt//miri/1---e-t"Ko/nn0)
m=(mi,...,mn),mi—▶oo
/o(ô) * * ' fn(xn) dGm(x, y)

/
and Gm is a measure. Since e~tH° has a Gaussian kernel and e~tu is a

multiple of a Gaussian density and integrals are Gaussian, the measure dGm
is Gaussian converging to a Gaussian measure dG. Since Lo(xQo) = #fio
and (xfi0,xfi0) = 1/2 we have
jXiXj dG = (xn0,e-^-s^L0xQo) = ]-e~^-Si)
which shows that dG is the joint probability distribution of QSo,... QSn.

The claim follows. □
Theorem 4.14.2 (Mehler formula). The kernel pt(x, y) of L0 is given by the

Mehler formula
, x 1 / (x2+y2)(l + e-2t)-4xye-t
p t ( x , y ) = — = = e x p - ^ -2(J2
/v '
with<j2 = (l-e-2*).
Proof. We have
(f,e-tL°g) = J f(y)^1(y)g(x)^\X) dG(x,y) = jf(y)pt(x,y) dy

with the Gaussian measure dG having covariance
H 1 e"'
r* l
We get Mehler's formula by inverting this matrix and using that the density
is
(2w)det(A)-1/2e-^x'y'''A^y^ .
D
Definition. Let dQ be the Wiener measure on C(R) belonging to the os
cillator process Qt-
Theorem 4.14.3 (Feynman-Kac for oscillator process). Given L = L0 + V

with V £ Cg°(R), then
(fn0,e-iL9n0) = jj{Qo)9{Qt)e-S°v^-»ds dQ
forall/,5e.L2(R,figda:).
Proof. By the Trotter product formula

(/fio,e-iL<A))= lim {fno,{e-tLo'ne-tv'n)ngno)
so that
(/fi0, e-iL9n0) = Jim fj(Qo)g(Qt) exp(-- ]T V(Qtj/n)) dQ . (4.5)

J n ^
3=0
and since Q is continuous, we have almost everywhere
t^ *
n I>«t;/n)- I V(Q8)ds.
37 = 0 J0
The integrand on the right hand side of (4.5) is dominated by
. l/(Qo)||p(Q*)|e*"^H-
which is in Ll(dQ) since
/ \f(Qo)\\9(Qt)\ dQ = (fio|/|,e-<L°fi0|<7|) < oo .

The dominated convergence theorem gives the claim. □
4.15. Neighborhood of Brownian motion 241
4.15 Neighborhood of Brownian motion
The Feynman-Kac formula can be used to understand the Dirichlet Lapla
cian of a domain D C Rd. For more details, see [93].
Example. Let D be an open set in Rd such that the Lebesgue measure \D\ is
finite and the Lebesgue measure of the boundary \6D\ is zero. Denote by Hd
the Dirichlet Laplacian -A/2. Denote by kD(E) the number of eigenvalues
of Hd below E. This function is also called the integrated density of states.
Denote with Kd the unit ball in Rd and with \Kd\ = Vo\(Kd) = 7rd/2r(f +
1)_1 its volume. Weyl's formula describes the asymptotic behavior of ko(E)
for large E:
kD(E) \Kd\ ■ |g|
E™oo Edl2 2d/27Td '
It shows that one can read off the volume of D from the spectrum of the
Laplacian.
Example. Put n ice balls KjtTl, 1 < j < n of radius rn into a glass of water
so that n • rn = a. In order to know, how good this ice cools the water it is
good to know the lowest eigenvalue Ei of the Dirichlet Laplacian Hd since
the motion of the temperature distribution u by the heat equation u = Hdu
is dominated by e~tEl. This motivates to compute the lowest eigenvalue of
the domain D \ (J?=i ^0,n- This can be done exactly in the limit n —▶ oo
and when ice Kj,n is randomly distributed in the glass. Mathematically,
this is described as follows:
Let D be an open bounded domain in Rd. Given a sequence x = (x\, #2, • • •)
which is an element in DN and a sequence of radii n, r2,..., define
n
Dn = D\\J{\x-Xi\<rn} .
i=l
This is the domain D with n points balls K2,n with center xi,...xn and ra
dius rn removed. Let H(x, n) be the Dirichlet Laplacian on Dn and Ek(x, n)
the fc-th eigenvalue of H(x, n) which are random variable Efc(n) in x, if DN
is equipped with the product Lebesgue measure. One can show that in the
case nrn —▶ a
Efc(n)Êfc(0) + 27ra|Z)|-1
in probability. Random impurities produce a constant shift in the spectrum.
For the physical system with the crushed ice, where the crushing makes
nrn —> oo, there is much better cooling as one might expect.
Definition. Let Ws(t) be the set
{x e Rd | \x - Bt(u)\ < S, for some s <E [0, t]} .

It is of course dependent on u and just a ^-neighborhood of the Brownian
path B[0it](uj). This set is called Wiener sausage and one is interested in the
expected volume |W$(£)| of this set as S —> 0. We will look at this problem
a bit more closely in the rest of this section.
Chapter 4. Continuous Stochastic Processes
Figure. A sample of Wiener

sausage in the plane d = 2. A
finite path of Brownian motion
with its neighborhood W$.
Lets first prove a lemma, which relates the Dirichlet Laplacian HD — - A/2
on D with Brownian motion.
Lemma 4.15.1. Let D be a bounded domain in Rd containing 0 and

PD{x,V,t), the integral kernel of e~tH, where H is the Dirichlet Laplacian
on D. Then
E[BS eD;0<s<(] = l- pD{0,x,t) dx .
Proof, (i) It is known that the Dirichlet Laplacian can be approximated in

the strong resolvent sense by operators H0 + AV, where V = 1dc is the
characteristic function of the exterior Dc of D. This means that
{Ho + A • V)~lu -+ {HD -z) lu,\->oc
for 2 outside [0, oo) and all u € C™(Rd).
(ii) Since Brownian paths are continuous, we have /„ V(BS) ds > 0 if and
only if Bs £ Cc for some s G [0, i\. We get therefore
e-A/<V(Bs)ds ,
point wise almost everywhere.
Let un be a sequence in C~ converging point wise to 1. We get with the

4.15. Neighborhood of Brownian motion 243
dominated convergence theorem, using (i) and (ii) and Feynman-Kac
E[BS G Dc; 0 < s < t] = lim E[un(Bs) eDc;0<s< t]

= lim limE[e-^v<B''\(Bt)]
n—▶ooA—▶oo
= lim lim e-t(ifo+AV)ix„(0)

n—ôo A—^00
= lim e-tHDun{0)
n—»oo
= lim [pD(0,x,t)un(0)dx= pD{0,x,t)dx
Theorem 4.15.2 (Spitzer). In three dimensions d = 3

471
E{\W5(t)\} = 27T(Jt + 4(52v/27rt + —~3<53 .
Proo/. Using Brownian scaling,

E[\Wxs(X2t)\\ = E[\{\x - Ba\ < XS,0 < s < X2t}\]
= E[\{\j-^\<S,0<s = s/\2<t}\]
= E[|{|^-Bg|<*,0<5<*}|]
= A3-E[|W*(t)|],
so that one assume without loss of generality that S = I: knowing E[| Wi (t) \],
we get the general case with the formula E[|W«(i)|] = S3 ■ E[|Wi(«-2t)|].
Let K be the closed unit ball in Rd. Define the hitting probability
f(x,t)=P[x + BseK;0<s<t].
We have
E[|Wi(t)|]= / f(x,t)dx.
Proof.
E[|Wi(t)|] = j [p[xeWi(t)]dxdB
= f fp[Bs-x€K;0<s<t}dxdB
= f f P[BS -x€K;0<s<t]dBdx
= / f(x, t) dx .
The hitting probability is radially symmetric and can be computed explic

itly in terms of r = \x\: for \x\ > 1, one has
/(M) = —7= e 2t dz-

r\J2-Kt Jo
Proof. The kernel of e~tH satisfies the heat equation
dtP(x,0,t) = (A/2)p(x,0,t)
inside D. From the previous lemma follows that / = (A/2)/, so that the
function g(r, t) = r/(x, t) satisfies g = 2(dr)29(r' *) wu^ boundary condition
g(r,Q) = 0, <?(1, t) = 1. We compute
dx = 2irt + 4v27rt
/ /(a,*)-
J\x\>l
and Jixi<1 f(x,t) dx = 4ir/3 so that
E[|Wi(*)| = 2tt* + 4>/2tS + 4tt/3 .
Corollary 4.15.3. In three dimensions, one has:
limÊ[|W«(t)|] = 27rt
and
lim \>E[\W6(t)\]=2ir6.
t—+oo t
Proof. The proof follows immediately from Spitzer's theorem (4.15.2). □
Remark. If Brownian motion were one-dimensional, then <$~2E[|W<$(£)|]

would stay bounded as S —▶ 0. The corollary shows that the Wiener sausage
is quite "fat". Brownian motion is rather "two-dimensional".
Remark. Kesten, Spitzer and Wightman have got stronger results. It is

even true that ]ims->o\Ws(t)\/t = 2n5 and ]imt->oo \Ws(t)\/t = 2n8 for
almost all paths.
4.16. The Ito integral for Brownian motion 245
4.16 The Ito integral for Brownian motion
We start now to develop stochastic integration first for Brownian motion
and then more generally for continuous martingales. Lets start with a mo
tivation. We know by theorem (4.2.5) that almost all paths of Brownian
motion are not differentiable. The usual Lebesgue-Stieltjes integral
rt
f(Bs)Bs ds
//o*
Jo
can therefore not be defined. We are first going to see, how a stochas
tic integral can still be constructed. Actually, we were already dealing
with a special case of stochastic integrals, namely with Wiener integrals
/0* f(Bs) dBs, where / is a function on C([0, oo],Rd) which can contain for
example /0* V(BS) ds as in the Feynman-Kac formula. But the result of this
integral was a number while the stochastic integral, we are going to define,
will be a random variable.
Definition. Let Bt be the one-dimensional Brownian motion process and
let / be a function / : R -▶ R. Define for n e N the random variable
2n 2n
Jn(f) = 51 f(B(m-l)2-*){Bm2-n ~ ^(m-l)2—) =: ^ Jn,rn(f) •

m=l m=1
We will use later for Jn,m(f) also the notation f(Btrn_1)5nBtrn, where
SnBt = Bt — Bt_2-n.
Remark. We have earlier defined the discrete stochastic integral for a pre
visible process C and a martingale X
f n
( / C dX)n=
(fcdx)n = Y/
y. Cm(Xm — Xm-i) .
J m=l
If we want to take for C a function of X, then we have to take Cm =

f(Xm-i)- This is the reason, why we have to take the differentials SnBtm
to "stick out into future".
The stochastic integral is a limit of discrete stochastic integrals:
Lemma 4.16.1. If / G CX(R) such that /, /' are bounded on R, then Jn(f)
converges in C2 to a random variable
/ f(Bs) dB = lim Jn
Jo n-*°°
satisfying
|| ff(B8)dB\\2=E[f f(Bs)2ds]
Jo Jo
Proof, (i) For i ? j we have E[Jnii(f)JnJ{f)] = 0.

Proof. For j > i, there is a factor Bj2-n - B{j_1)2-n of Jn,i(f)Jnj(f) inde
pendent of the rest of Jn,i(f)Jnj(f) and the claim follows from E[Bj2-n -
B(j-l)2-"} =0.
(ii) E[Jn,m(/)2] = E[f(B{m_l)2-n)2}2-\

Proof. f(B(m_1)/2n) is independent of (Bm2-n - £(m_1)2-n)2 which has
expectation 2~n.
(iii) From (ii) follows
2n
\\Jn(f)\\2='£E[f(B{m_1)2-n)2}2-n .
m=l
(iv) The claim: Jn converges in C2.

Since / G C\ there exists C = ||/'||^ and this gives \f(x) - f(y)\2 <
C >\x- y\2. We get
\\Jn+l(f)-Jn(f)\\l
2n-l
= J2 E[(/(B(2ro+1)2-(»+i,) - /(5(2m)2_(.+1)))2]2-("+1)
m=l
2n-l
< C £ E[(5(2m+1)2_(.+1) - S(2m)2-(n+l))2]2-("+1)
m=l
= C.2"n-2,
where the last equality followed from the fact that E[(jB(2m+i)2-(n+i) -
B{2m)2-(n+1))2] = 2~n since J3 is Gaussian. We see that Jn is a Cauchy
sequence in C2 and has therefore a limit.
(v) The claim || jj f(B8) dB\\2 = E[jJ f(Bs)2 ds}.

Proof. Since Em/(s(m-i)2—)22"n converges point wise to /J" f(Bs)2 ds,
(which exists because / and Bs are continuous), and is dominated by H/H^,
the claim follows since Jn converges in C2. D
We can extend the integral to functions /, which are locally L1 and bounded
near 0. We write £foc(R) for functions / which are in LP(I) when restricted
to any finite interval / on the real line.
Corollary 4.16.2. /0 f(Bs) dB exists as a C2 random variable for / G

Ljoc(R) fl L°°(-e, e) and any e > 0.
Proof, (i) If / € Ljoc{R) n L°°{-e, e) for some e > 0, then

ft*? e-X*/2s dxds <OQ.
E [Jfo1 f ( B s )J2od s } = f 1J [r /27TS
(ii) If / G L]oc(R) fl L°°(-e, e), then for almost every B(u), the limit
lim / l{.aM(Bs)f(Bs)2 ds
<*-oo J0
exists point wise and is finite.

Proof. Bs is continuous for almost all lj so that l[_0>a](Bs)/(B) is indepen
dent of a for large a. The integral E^1 l[_a,a](Bs)f(Bs)2 ds] is bounded
by E[f(Bs)2 ds] < oo by (i).
(iii) The claim.

Proof. Assume / G L[oc(R)nL°°(-6, e). Given fn G C2(R) with l[_a,a]/n -▶
/ in L2(R).
By the dominated convergence theorem, we have
/" 1[ - a, a]/n(Ba) dB - / 1( - a, a)f(Bs) dB
in £2. Since by (ii), the £2 bound is independent of a, we can also pass to

the limit a —▶ oo. □
Definition. This integral is called an Ito integral. Having the one-dimensional

integral allows also to set up the integral in higher dimensions: with Brow
nian motion in Rd and / G Lfoc(Rd) define the integral /J" f(Bs) dBs
component wise.
Lemma 4.16.3. For n —^ oo,
2n 2n
Yl Jnj(1)2 = X^72n ~ B(j"l)/2n)2 "* X '

j=l j=l
Proof. By definition of Brownian motion, we know that for fixed n, Jn,j

are N(0,2~n)-distributed random variables and so
2n
EE J^"(1)2] = ^ • VarI%2" - B0-i)/2»] = 2n2"n = 1 •

3=1
Now, Xj = 2nJnj are IID iV(0,1)-distributed random variables so that by

the law of large numbers
2
iE*^1
2~
for n —> oo. □

The formal rules of integration do not hold for this integral. We have for
example in one dimension:
I 1,-, .. .1
BsdB = -{Bl-l)^l-{B\-Bl).
Proof. Define
Jn = 2^ f(B(m-l)2-n)(Bm2-n - J5(m_1)2-») ,
771=1
2 n
^™ = Z^ f(Bm2-n)(Bm2-n - B(m_i)2-n) .
m=l
The above lemma implies that J+ - J~ -> 1 almost everywhere for n -> oo
and we check also J+ + J~ = B\. Both of these identities come from
cancellations in the sum and imply together the claim. □
We mention now some trivial properties of the stochastic integral.
Theorem 4.16.4 (Properties of the Ito integral). Here are some basic prop
erties of the Ito integral:
(1) SS f(B.) + g(B.) dBa = /„< f(B.) dBs + jj g(Bs) dBs.
(2) So * ■ /(*.) dB. = \. Ji f(Bs) dBa.
(3) t>-+ J0 f(Bs) dBs is a continuous map from K+ to C2.
(4)E[f*f(Bs)dB3) = 0.
(5) f0 f(Bs) dBa is At measurable.
Proof. (1) and (2) follow from the definition of the integral.
For (3) define Xt = /„* f(Bs) dB. Since
/t+e
f(B,)2ds]
Jt Jr v 2ns
for e —> 0, the claim follows.
(4) and (5) can be seen by verifying it first for elementary functions /. □
It will be useful to consider an other generalizations of the integral.
Definition. If dW = dxdB is the Wiener measure on Rd x C([0, oo), define
f f(Ws)dWs=
JO jRd
[ J fo f(x-rB3)dBsdx.
Definition. Assume / is also time dependent so that it is a function on
Rd x R. As long as E[/0 \f(Bs, s)\2 ds] < oo, we can also define the integral
f(Ba,s) ds .
Jo
The following formula is useful for understanding and calculating stochas

tic integrals. It is the "fundamental theorem for stochastic integrals" and
allows to do "change of variables" in stochastic calculus similarly as the
fundamental theorem of calculus does for usual calculus.
Theorem 4.16.5 (Ito's formula). For a C2 function f(x) on Rd
f(Bt) - f(Bo) = J Vf(Bs) • dBs + \J Af(Bs) ds .
If Bs would be an ordinary path in Rd with velocity vector dBs = Bs ds,

then we had
t
f(Bt) - /(Bo) = / V/(£s) • Ba ds
Jo
by the fundamental theorem of line integrals in calculus. It is a bit surprising
that in the stochastic setup, a second derivative A/ appears in a first order
differential. One writes sometimes the formula also in the differential form
df = VfdB + ±Afdt.
Remark. We cite [11]: "Ito's formula is now the bread and butter of the
"quant" department of several major financial institutions. Models like that
of Black-Scholes constitute the basis on which a modern business makes de
cisions about how everything from stocks and bonds to pork belly futures
should be priced. Ito's formula provides the link between various stochastic
quantities and differential equations of which those quantities are the so
lution." For more information on the Black-Scholes model and the famous
Black-Scholes formula, see [16].
It is not much more work to prove a more general formula for functions
/(x,t), which can be time-dependent too:
Theorem 4.16.6 (Generalized Ito formula). Given a function /(#, t) on Rd x

[0,t] which is twice differentiable in x and differentiable in t. Then
f{Bt,t)-f(B0,0)= Jj Vf(Bs,s)-dBs-rl [ Af(Bs,s)ds+[ f(Bs,s)ds.

o * Jo Jo
In differential notation, this means
df = VfdB + (±Af + f)dt.
Proof. By a change of variables, we can assume t = 1. For each n, we

discretized time
{0<2-n<...,tfc = fc.2"n,...,l}
and define SnBtk = Btk — Btk_1. We write
2n
f(Bul) - f(B0,0) = J2(Vf)(Btk_1,tk-1)6nBtk

fc=i
2n
+ E/(5tfc,tfc-i) - f(Btk^tk-i) ~ (Vf)(Btk_,M-^nBtk

k=l
2n
+ J2f(Btk,tk)-f(Btk,tk^)
= In + IIn + IIIn.
(i) By definition of the Ito integral, the first sum In converges in C2 to

f0\Vf)(Bs,s)dBs.
(ii) If p > 2, we have ^fe=1 \6nBtk\p —▶ 0 for n —▶ oo.

Proof. 0n-Btfc is a iV(0,2_n)-distributed random variable so that
|z|pe-x2/2 dx = C2-(np)/2 .
/ o-oo
o
This means
2n
EE |JnStJp] = C2n2-^/2
fc=i
which goes to zero for n —» oo and p > 2.
(iii) J2kLi Ei(Btk ~ Btk^)4} -* 0 follows from (ii). We have therefore

2n 2n
YÊ[g(Btk,tk)2((Btk-Btk_l)2-2-n)2} < c£ Var[(i?u - i^J2]

fc=i f cf c== il
2n
fc=l
(iv) Using a Taylor expansion
f(x) = f(y) - Vf(y)(x-y) - \2 £dXiXjf(y)(x- y)i(x-y)j + 0(\x - y\3 ,

we get for n —▶ oo
2"
n- - Y.2^dxix'f{Btk-^tk-l)(-5nBtk)i(-5nBtk)i "* °
k—l i,j
in £2. Since
2"
k=l
goes to zero in C2 (applying (ii) for g = dXiXjf and note that (SnBtk)i and
(6nBtk)j are independent for i^ j), we have therefore
i r*
in£2.
(v) A Taylor expansion with respect to t
f(x, t) - f(x, s) - f{x, s)it -s) + OUt - s)2)
gives
IIIn^ [ [ HBS,S)
f(Bs,s)ds
Jo
in C1 because s —> f(Bs,s) is continuous and IIIn is a Riemann sum

approximation. □
Example. Consider the function
f(x,t)=eax~a2t/2 .
Because this function satisfies the heat equation / + f" /2 = 0, we get from
Ito's formula
f(But) - f(Bo,t) = a j f(Bs,s) • dBs .

Jo
We see that for functions satisfying the heat equation / + f"/2 = 0 Ito's
formula reduces to the usual rule of calculus. If we make a power expansion
in a of
j' eaB8-**s/2 dB = \ecBs-c?sl2 _ I
Jo a a
we get other formulas like
J BsdB=l-iB2-t).
Wick ordering.
There is a notation used in quantum field theory developed by Gian-Carlo
Wick at about the same time as Ito's invented the integral. This Wick
ordering is a map on polynomials Yl7=i a^x% which leave monomials (poly
nomials of the form xn + an-\xn~x • • •) invariant.
Definition. Let
Hn(x)Qo(x)
be the n'-th eigenfunction of the quantum mechanical oscillator. Define
,_2-/2 nKy/2J
and extend the definition to all polynomials by linearity. The Polynomials

: xn : are orthogonal with respect to the measure tiody = 7r-1/2e-2/ dy
because we have seen that the eigenfunctions fin are orthonormal.
Example. Here are the first Wick powers:
:x = X
:x2 = x2-l
:xz = x3 -3x
:z4 = x4 - 6x2 + 3
:x5 = x5 - 10a;3 + 15x .
Definition. The multiplication operator Q : / i—▶ xf is called the position

operator. By definition of the creation and annihilation operators one has
Q = -^(A + A*).
The following formula indicates, why Wick ordering has its name and why
it is useful in quantum mechanics:
Proposition 4.16.7. As operators, we have the identity
■Qn ■= ^ ■ V + *T •■- ^ E ( " ) (*?*"-'
Definition. Define L = £"=0 ( " ) iA*)jAn~K

Proof. Since we know that ftn forms a basis in L2, we have only to verify
that : Qn : ttk = 2"n/2LQfc for all fc. From
2"1/2[Q,L] = [A + A*E(^)(A*)M^]
= E ( n ) Ji^y-'A^ - (n - 3)(A*yA^~l
3=0 ^ J '
= 0
we obtain by linearity [Hk(y/2Q), L\. Because : Qn : Q0 = 2"n/2(n!)1/2nn =

2-n/2(A*)nft0 = 2"n/2LQ0, we get
0 = (: Qn : -2-n/2L)ft0
= (A:!)"1/2^(^Q)(: Qn : -2"^2L)fi0
= (: Qn : -2"n/2L)(/c!)-1/2i/fc(v^Q)fio
= (: Qn : -2"n/2L)ftfc .
Remark. The new ordering made the operators A, A* behaves as if A, B

would commutate. even so they don't: they satisfy the commutation rela
tions [A, A*] = 1:
The fact that stochastic integration is relevant to quantum mechanics can

be seen from the following formula for the Ito integral:
Theorem 4.16.8 (Ito Integral of Bn). Wick ordering makes the Ito integral
behave like an ordinary integral.
• : K :dB. = ^.: B,"« :
I
Remark. Notation can be important to make a concept appear natural. An
other example, where an adaption of notation helps is quantum calculus,
"calculus without taking limits" [44], where the derivative is defined as
Dqf(x) = dqf(x)/dq(x) with dqf(x) = f(qx) - f(x). One can see that
Dqxn = [n]xn_1, where [n] = ^-fr- The limit q —▶ 1 corresponds to the
classical limit case h —> 0 of quantum mechanics.
Proof. By rescaling, we can assume that t = 1.

We prove all these equalities simultaneously by showing
l
I
o
: eaBs : dB = a'1 : eaBl : -a"1
The generating function for the Hermite polynomials is known to be
J2Hnix)^j
n—0
= ea^-£
(We can check this formula by multiplying it with ft0, replacing x with
x/y/2 so that we have
E
2 2
^njx)an
71=0 (n!)V2
If we apply A* on both sides, the equation goes onto itself and we get after
k such applications of A* that that the inner product with flk is the same
on both sides. Therefore the functions must be the same.)
This means
<* : X ' = e«x-W
-E 3=0
Since the right hand side satisfies / + f" 12 = 2, the claim follows from the
Ito formula for such functions. □
We can now determine all the integrals J B™ dB:
I t
o
1 dB = Bt
I0
BsdB = \iB2-l)
*
Jf oB2dB = f : B2J :+ldB

o = Bt + \i: 6Bt :3) = Bt + 6±(B? - Wt)
and so on.
Stochastic integrals for the oscillator and the Brownian bridge process.
Let Qt = e~tBeitj\/2 the oscillator process and At = (1 - t)Bt/(i-t) the
Brownian bridge. If we define new discrete differentials
SnQtk = Qtk+1-e-^-^Qtk
tk+i —tk .
SnAtk = Atk+1 - Atk +
il-t)
the stochastic integrals can be defined as in the case of Brownian motion
as a limit of discrete integrals.
Feynman-Kac formula for Schrodinger operators with magnetic fields.

Stochastic integrals appear in the Feynman-Kac formula for particles mov
ing in a magnetic field. Let A(x) be a vector potential in R3 which gives
4.17. Processes of bounded quadratic variation 255
the magnetic field B(x) = curl(A). Quantum mechanically, a particle mov
ing in an magnetic field together with an external field is described by the
Hamiltonian
H = (iV + A)2 + V .
In the case A = 0, we get the usual Schrodinger operator. The Feynman-

Kac formula is the Wiener integral
e~tHu(0)= j e-F^Bû(Bt)dB,
where F(B, t) is a stochastic integral.
F(B, t) = i f a(Bs) dB+%- j div(A) ds + I V(BS) ds .
4.17 Processes of bounded quadratic variation

We develop now the stochastic Ito integral with respect to general martin
gales. Brownian motion B will be replaced by a martingale M which are
assumed to be in £2. The aim will be to define an integral
/ Ks dMs ,
Jo
where K is a progressively measurable process which satisfies some bound-

edness condition.
Definition. Given a right-continuous function / : [0, oo) -> R. For each

finite subdivision
A = {0 = to,*i,..-,* = *n}
of the interval [0, t] we define | A| = sup[=1 |t»+i - U\ called the modulus of
A. Define
n-l
||/||A = £|/ti+1-/tJ-
i=0
A function with finite total variation ||/||t = supA ||/||A < oo is called a
function of finite variation. If supt |/|t < oo, then / is called of bounded
variation. One abbreviates, bounded variation with BV.
Example. Differentiable C1 functions are of finite variation. Note that for

functions of finite variations, Vt can go to oo for t -> oo but if Vt stays
bounded, we have a function of bounded variation. Monotone and bounded
functions are of finite variation. Sums of functions of bounded variation are
of bounded variation.
Remark. Every function of finite variation can be written as / = /+ - /",

where ^ are both positive and increasing. Proof: define /± = (±/t +
ll/IW/2.
Remark. Functions of bounded variation are in one to one correspondence
to Borel measures on [0, oo) by the Stieltjes integral /0* \df\ = /+ + /-.
Definition. A process Xt is called increasing if the paths Xt(u) are finite,

right-continuous and increasing for almost all u e ft. A process Xt is called
of finite variation, if the paths Xt(u) are finite, right-continuous and of
finite variation for almost all u e ft.
Remark. Every bounded variation process A can be written as At = A? -

At , where Af are increasing. The process Vt = f* \dA\s = A+ + A< is
increasing and we get for almost all u G ft a measure called the variation
of A.
If Xt is a bounded A-adapted process and A is a process of bounded

variation, we can form the Lebesgue-Stieltjes integral
(X-A)t(u) = [ Xs(u) dAs(u) .

Jo
We would like to define such an integral for martingales. The problem is:
Proposition 4.17.1. A continuous martingale M is never of finite variation,

unless it is constant.
Proof. Assume M is of finite variation. We show that it is constant.
(i) We can assume without loss of generality that M is of bounded varia

tion.
Proof. Otherwise, we can look at the martingale M5-, where Sn is the
stopping time Sn = inf{s \Vs>n} and Vt is the variation of M on [0, t\.
(ii) We can also assume also without loss of generality that M0 = 0.
(iii) Let A - {£0 = 0, tu ..., tn = i) be a subdivision of [0, t\. Since M is a

martingale, we have by Pythagoras
fc-i
E[M2} = E[£(M2+1-M2)]
i=0
k-1
= E($2(Mti+l - Mti)(Mti+1 + Mti)}
i=l
k-1
= E[V>4+1-Mti)2]
2=1
and so
E[M2] < E[Vt(sup \Mti+1 -Mti\]<K- E[supi |Mti+1 - Mu\] .

i
If the modulus | A| goes to zero, then the right hand side goes to zero since
M is continuous. Therefore M — 0. □
Remark. This proposition applies especially for Brownian motion and un

derlines the fact that the stochastic integral could not be defined point wise
by a Lebesgue-Stieltjes integral.
Definition. If A = {to = 0 < t\ < ... } is a subdivision of R+ = [0, oo) with

only finitely many points {to, h,... ,tk } in each interval [0,t], we define
for a process X
TtA = Tt\X) = C£iXti+1 - Xu)2) + iXt - Xtkf .

i=0
The process X is called of finite quadratic variation, if there exists a process

< X, X > such that for each t, the random variable TtA converges in
probability to < X,X >t as |A| —> 0.
Theorem 4.17.2 (Doob-Meyer decomposition). Given a continuous and

bounded martingale M of finite quadratic variation. Then < M, M > is
the unique continuous increasing adapted process vanishing at zero such
that M2— < M, M > is a martingale.
Remark. Before we enter the not so easy proof given in [83], let us mention
the corresponding result in the discrete case (see theorem (3.5.1), where
M2 was a submartingale so that M2 could be written uniquely as a sum
of a martingale and an increasing previsible process.
Proof. Uniqueness follows from the previous proposition: if there would be

two such continuous and increasing processes A, £?, then A — B would be
a continuous martingale with bounded variation (if A and B are increas
ing they are of bounded variation) which vanishes at zero. Therefore A — B.
(i) M2 — T^(M) is a continuous martingale.

Proof. For U < s < £i+i, we have from the martingale property using that
(Mti+1 — Ms)2 and (M3 — Mti)2 are independent,
E[(Mti+1 - Mti)2 | As] = E[(MU+1 - Ms)2\As] + (Ms - Mti)2 .

This implies with 0 = t0 < h < • • • < U < s < *z+1 < .. • < tk < t and
using orthogonality
E [ TA ( M ) - TA ( M ) \ A S } = E [ ^ ( M , , + 1 - M £ j . ) 2 | A ]
3=1
+ E[(Mt-Mtk)2\As}+E[(Ms-Mtl)2\As}
= E[(Mt-Ms)2\As} = E[M2-M2\As}.
This implies that Mt2 - TtA(M) is a continuous martingale.
(ii) Let C be a constant such that \M\ < C in [0,a]. Then E[TA] < 4C2,
independent of the subdivision A = {to,..., tn) of [0, a].
Proof. The previous computation in (i) gives for 5 = 0, using T^(M) = 0
E[TA(M)\A0] = E[M2 - M2\Ao) < E[(Mt - M0)(Mt + M0)\ < 4C2 .

(iii) For any subdivision A, one has E[(TA)2] < 48C4.
Proof. We can assume tn = a. Then
iTtiM))2 = (E(M(fe-Mtfc_J2)2
= 2f>* - TftiT* - Ttj + f>(fc - Mtk_xf .

fe=l k=l
From (i), we have
E[TaA - T?k\Atk\ = E[(M„ - Mtkf \ Atk]

and consequently, using (ii)
n(TaA)2} = 2J2n(Ma-Mtk)2iT*-T*+J} + J2n(Mtk-Mtk_1)*]

k=i k=i
< E[(2 sup \Ma - MtJ2 + sup |Mtfc - Mtk_x |2)TQA]
< 12C2E[TaA] < 48C4 .
(iii) For fixed a > 0 and subdivisions An of [0,a] satisfying |An| -+ 0, the
sequence TAn has a limit in C2.
Proof. Given two subdivisions A', A" of [0,a], let A be the subdivision
obtained by taking the union of the points of A' and A". By (i), the process
X = TA -TA is a martingale and by (i) again, applied to the martingale
X instead of M we have, using (x + y)2 < 2(x2 + y2)
E[X2a] = E[(TA' - TA")2) = E[TA(X)\ < 2(E[TA(TA')\ + E[TA(TA")}) .

We have therefore only to show that E[TA(TA')\ -▶ 0 for |A'| + |A,;| -> 0.
Let sk be in A and tm the rightmost point in A7 such that £m < sk <
Sfc+i < tm+i. We have
r ^ - T. f = ( M S f e + 1 - M t m ) 2 - ( M „ - M t m ) 2
= (MSfc+1 - MSk)iMSk+1 + MSk - 2Mtm)
and so A
TA(TA') < (supk |MSte+1 + MSk - 2Mtm\2)TA .
By the Cauchy Schwarz-inequality
Ep^(I*')] <E[supk \MSk+1+MSk-2MtJi]1'2E{iT*)2}V2
and the first factor goes to 0 as |A| -▶ 0 and the second factor is bounded
because of (iii).
(iv) There exists a sequence of An C An+i such that TtAn(M) converges

uniformly to a limit (M, M) on [0, a]. a a
Proof. Doob's inequality applied to the discrete time martingale T n-T m
gives
E[sup |7f " - TA™\2} < 4E[(TaA- - TA™)2} .
t<a
Choose the sequence An such that An+i is a refinement of An and such

that |JnAn is dense in [0,a], we can achieve that the convergence is uni
form. The limit (M,M) is therefore continuous.
(v) (M, M) is increasing.

Proof. Take An C An+i. For any pair s < t in |Jn An, we have Ts »(M) <
TAn(M) if n is so large that An contains both s and t. Therefore (M, M)
is increasing on \Jn An, which can be chosen to be dense. The continuity
of M implies that (M, M) is increasing everywhere. □
Remark. The assumption of boundedness for the martingales is not essen

tial. It holds for general martingales and even more generally for so called
local martingales, stochastic processes X for which there exists a sequence
of bounded stopping times Tn increasing to oo for which XTn are martin
gales.
Corollary 4.17.3. Let M, N be two continuous martingales with the same

filtration. There exists a unique continuous adapted process (M, N) of finite
variation which is vanishing at zero and such that
MN-(M,N)
is a martingale.
Proof. Uniqueness follows again from the fact that a finite variation mar
tingale must be zero. To get existence, use the parallelogram law
(M, N) = J«M + AT, M + N) - (M - N, M - N)) .

This is vanishing at zero and of finite variation since it is a sum of two
processes with this property.
We know that M2 - (M, M),N2- (N, N) and so that (M ± N)2 - (M ±
N,M±N) are martingales. Therefore
(M + N)2 - (M + N, M + N) - (M - N)2 - (M - N, M - N)
= 4MN-(M + N,M + N)-(M-N,M-N) .
and MN - (M, N) is a martingale. □
Definition. The process (M, N) is called the bracket of M and N and

(M, M) the increasing process of M.
Example. If B = (B™,.. .,£(*) is Brownian motion, then (< B^\B^) =

Sijt as we have computed in the proof of the Ito formula in the case t = 1.
It can be shown that every martingale M which has the property that
(M^,M^) = 5irt
must be Brownian motion. This is Levy's characterization of Brownian
motion.
Remark. If M is a martingale vanishing at zero and (M, M) = 0, then

M = 0. Since M2 - (M,M)t is a martingale vanishing at zero, we have
E[M2]=E[(M,M)*].
Remark. Since we have got (M,M) as a limit of processes TtA, we could
also write (M, N) as such a limit.
4.18 The Ito integral for martingales

In the last section, we have defined for two continuous martingales Af ,JV,
the bracket process (M,N). Because (M,M) was increasing, it was of fi
nite variation and therefore also (M, N) is of finite variation. It defines a
random measure d(M, TV).
Theorem 4.18.1 (Kunita-Watanabe inequality). Let M, N be two continu

ous martingales and H, K two measurable processes. Then for all p, q > 1
satisfying 1/p + 1/q = 1, we have for alH < oo
E[fJ o\Hs\\Ks\\d(M,N)8\] < \\(fJ oH2sd(M,M)Y?2\\p

■ Wif K2td(N,N)m\q.
Jo
4.18. The Ito integral for martingales 261
Proof, (i) Define (M, AT)* = (M,N)t - (M, A/%. Claim: almost surely
|(M, AT)*| < ((MiMW'îNtN),)1'2 .
Proof. For fixed r, the random variable
(M, M)l + 2r(M, N)*a + r2(A, iV)* = (M + rTV, M + rA^
is positive almost everywhere and this stays true simultaneously for a dense
set of r G R. Since M, AT are continuous, it holds for all r. The claim follows,
since a + 2rb+cr2 > 0 for al] r > 0 with nonnegative a, c implies 6 < yja^fc.
(ii) To prove the claim, it is, using Holder's inequality, enough to show
almost everywhere, the inequality
J/ o\H,\ \K,\ d\(M,N)\s <J oi f H2sd(M,M))V2 ■J o( f K2sdiN,N)f'2

holds. By taking limits, it is enough to prove this for t < oo and bounded
K,H. By a density argument, we can also assume the both K and H are
step functions H = YlZ=i H^Ji and K = XXa Kdjn where J* = [ti,U+i).
(iii) We get from (i) for step functions H, K as in (ii)
\ J[ oH a K a d ( M , N ) s \ < Ti l H i K i W i M ^ N f t ^ l
< ^|i/^|((M,M)^1)1/2((M,M)^1)1/2
i
< C£H2(M,M)l^y/2iJ2K2(N,N)^y/2
= i f H2diM, M))"2 ■ i f K2d(N, N))1'2 ,

Jo Jo
where we have used Cauchy-Schwarz inequality for the summation over
i. □
Definition. Denote by H2 the set of £2-martingales which are ^-adapted

and satisfy
llA/ll^^supEIM2])1/^^.
Call H2 the subset of continuous martingales in H2 and with Hq the subset

of continuous martingales which are vanishing at zero.
Given a martingale M G H2, we define C2(M) the space of progressively
measurable processes K such that
/•OO
II^HrV)^/, K?d(M,M).] <oo..

Both H2 and C2(M) are Hilbert spaces.
Lemma 4.18.2. The space H2 of continuous C2 martingales is closed in H2

and so a Hilbert space. Also H$ is closed in H2 and is therefore a Hilbert
space.
Proof. Take a sequence M(n) in H2 converging to M G H2. By Doob's

inequality
E[(sup |M,n - Mt\)2} < 4\\M^ - M\\2n2 .
We can extract a subsequence, for which supt \M^ — Mt\ converges point
wise to zero almost everywhere. Therefore M G H2. The same argument
shows also that Hq is closed. □
Proposition 4.18.3. Given M G H2 and K G C2(M). There exists a unique

element /0 KdM G Hq such that
< / KdM,N>= [ Kd(M,N)

Jo Jo
for every N G H2. The map K h+ /q* KdM is an isometry form C2(M) to
H2.
Proof. We can assume M e Ho since in general, we define /0 K dM =

fi K d i M - M o ) .
(i) By the Kunita-Watanabe inequality, we have for every N G Hq
lElfJ 0 Ksd(M,N)s}\<\\N\\n2-\\K\\C2{M).
The map
N ^ E [ ( [ JoK 8 ) d ( M , N ) s ]
is therefore a linear continuous functional on the Hilbert space Hq. By

Riesz representation theorem, there is an element / K dM G Hq such that
E[( / Ks dMs)Nt] =E[[ Ksd(M,N)s

Jo Jo
for every N G Hq.
4.18. The Ito integral for martingales 263
(ii) Uniqueness. Assume there exist two martingales L, V G Hi such that

(L, AT) = (I/, AT) for all N e H$. Then, in particular, (L -L',L- V) = 0,
from which L = L' follows.
(iii) The integral K ^ Jq K dM is an isometry because
|| / KdM\\2Ho = E[( KsdMs)2}

/•OO
= E[ K2d(M,M)]
Jo
= HKll£2(M)
Definition. The martingale /0* Ks dMs is called the Ito integral of the
progressively measurable process K with respect to the martingale M. We
can take especially, K = /(M), since continuous processes are progressively
measurable. If we take M = B, Brownian motion, we get the already
familiar Ito integral.
Definition. An At adapted right-continuous process is called a local martin

gale if there exists a sequence Tn of increasing stopping times with Tn —▶ oo
almost everywhere, such that for every n, the process XTnl{Tn>o} is a uni
formly integrable ^-martingale. Local martingales are more general than
martingales. Stochastic integration can be defined more generally for local
martingales.
We show now that Ito's formula holds also for general martingales. First,
a special case, the integration by parts formula.
Theorem 4.18.4 (Integration by parts). Let X,Y be two continuous mar

tingales. Then
XtYt - X0Yo = [ XsdYs+ [ YsdXs + (X, Y)t

Jo Jo
and especially
X2 - X2 = 2 / Xs dXs + (X,X)t •
Proof. The general case follows from the special case by polarization: use
the special case for X ± Y as well as X and Y.
The special case is proved by discretisation: let A = {to,h,..., tn} be a
finite discretisation of [0,£]. Then
J2ixti+1 - xti)2 = X2 - X2 - 2]Txti(Xti+1 - xti).

i=l i=l
Letting |A| going to zero, we get the claim. □
Theorem 4.18.5 (Ito formula for martingales). Given vector martingales

M = (M^\ ..., Af<d>) and X and a function / G C2(Rd, R). Then
f(Xt)-f(Xo) = f\f(X) dMt+\Y,jl5Xi5XjfXiXj(Xs) d(M^\Mtt(i)v/
Proof. It is enough to prove the formula for polynomials. By the integration

by parts formula, we get the result for functions f(x) = xig(x), if it is
established for a function g. Since it is true for constant functions, we are
done by induction. rj
Remark. The usual Ito formula in one dimensions is a special case
f(Xt) - f(X0) = f f'(Xs) dBs + \f fn(Xs) ds .

Jo 2 J0
In one dimension and if Mt = Bt is Brownian motion and Xt is a martin
gale, we have We will use it later, when dealing with stochastic differential
equations. It is a special case, because (Bt, Bt) = *, so that d(Bu Bt) = dt.
Example. If f(x) = x2, this formula gives for processes satisfying X0 = 0
X2/2 = f Xs dBs + It.

Jo *
This formula integrates the stochastic integral /0* Xs dBs = A^2/2 - t/2.
Example. If f(x) = log(x), the formula gives
\og(Xt/Xo) = f dBs/X8 - 1 f ds/X2s .

Jo * Jo
4.19 Stochastic differential equations

We have seen earlier that if Bt is Brownian motion, then X = /(J5, t) =
eaBt-a t/2 is a martingaie. In the last section we learned using Ito's formula
and and \ A/ + / = 0 that
t
Io
aXs dMs = Xt - 1 .
We can write this in differential form as
dXt = aXtdMuXo = l.
4.19. Stochastic differential equations 265
This is an example of a stochastic differential equation (SDE) and one
would use the notation
——- = olX
dM
if it would not lead to confusion with the corresponding ordinary differential
equation, where M is not a stochastic process but a variable and where the
solution would be X — eaB. Here, the solution is the stochastic process
Xt = e^t-o^tii
Definition. Let Bt be Brownian motion in Rd. A solution of a stochastic
differential equation
dXt = f(XuBt) .dBt + g(Xt)dt,
is a Rd-valued process Xt satisfying
Xt = / f(Xa,B9) dBs+ f g(Xs) ds ,

Jo Jo
where / : and g :
As for ordinary differential equations, where one can easily solve separable
differential equations dx/dt = f(x)+ g(t) by integration, this works for
stochastic differential equations. However, to integrate, one has to use an
adapted substitution. The key is Ito's formula (4.18.5) which holds for
martingales and so for solutions of stochastic differential equations which
is in one dimensions
fiXt) - /(Xo) = f f'iXa) dXs + \f f'iX.) d(Xs, Xa) .

Jo * Jo
The following "multiplication table" for the product (•, •) and the differen
tials dt, dBt can be found in many books of stochastic differential equations
[2, 46, 66] and is useful to have in mind when solving actual stochastic dif
ferential equations:
dt dBt
dt 0 0
dBt 0 t
Example. The linear ordinary differential equation dX/dt = rX with solu
tion Xt = ertX0 has a stochastic analog. It is called the stochastic popula
tion model. We look for a stochastic process Xt which solves the SDE
dXt = rXt + aXtCt •

dt
Separation of variables gives
dX = rtdt + a(dt
—
and integration with respect to t
t dXt
I
Joo
—A=rt + aBt.
^t
In order to compute the stochastic integral on the left hand side, we have to
do a change of variables with f(X) = log(#). Looking up the multiplication
table:
(dXt, dXt) = (rXtdt + aXtdBu rXtdt + a2XtdBt) = a2X2dt.

Ito's formula in one dimensions
fiXt) - /(Xo) = J* f'iX.) dXs + ^j f"iXs)(Xs,X6

gives therefore
\og(Xt/Xo) = / dXs/Xs -\j a2ds
so that Jq dXs/Xs = a2t/2 + log(Xt/-Xo). Therefore,
a2t/2 + \og(Xt/X0) = rt + aBt
and so Xt = Xoert~a2t/2+aBt. This process is called geometric Brownian

motion. We see especially that X = X/2 + X£ has the solution Xt = eBt.
Figure. Solutions to the stochastic Figure. Solutions to the stochastic

population model for r > 0. population model for r < 0.
Remark. The stochastic population model is also important when modeling

financial markets. In that area the constant r is called the percentage drift
or expected gain and a is called the percentage volatility. The Black-Scholes
model makes the assumption that the stock prices evolves according to
geometric Brownian motion.
Example. In principle, one can study stochastic versions of any differential

equation. An example from physics is when a particle move in a possibly
time-dependent force field F(x,t) with friction b for which the equation
without noise is
x = —bx + F(x, t) .
If we add white noise, we get a stochastic differential equation
x = -bx + F(x,t) + a((t) .
For example, with X = x and F = 0, the function v(t) satisfies the stochas
tic differential equation
dXt = -bXt + aCt ,
dt
which has the solution
Xt e~bt + aBt
With a time dependent force F(x, t), already the differential equation with
out noise can not be given closed solutions in general. If the friction constant
b is noisy, we obtain
*?± = i-b + aCt)Xt
which is the stochastic population model treated in the previous example.
Example. Here is a list of stochastic differential equations with solutions.

We again use the notation of white noise £(£) = ^ which is a generalized
function in the following table. The notational replacement dBt = Ctdt is
quite popular for more applied sciences like engineering or finance.
Stochastic differential equation Solution

f xt = Kit) Xt = Bt
&Xt = BtCit) xt = B2t /2 = iB2-l)/2
iXt = B2«t) xt = Bf /3 = iBf - 3Bt)/3
ixt = Bfat) xt = Bf /4 = (JBt4-6B2 + 3)/4
&xt = Btat) xt = Bf /5 = (Bf5 - 10B? + 15Bj)/5
£xt = aXtdt) Xt = eaBt-aH'2
iXt=rXt+aXtat) Xt = ert+aBt-a''t/2
Remark. Because the Ito integral can be defined for any continuous martin
gale, Brownian motion could be replaced by an other continuous martingale
M leading to other classes of stochastic differential equations. A solution
must then satisfy
Xt= f f(Xs,Ms,s) -d,Ms+ [ g(Xs,s)ds.

Jo Jo
Example.
Xt = eaMt-c*2(X,X)t/2
is a solution of dXt = aMtdMt, M0 = 1. ■
Remark. Stochastic differential equations were introduced by Ito in 1951.
Differential equations with a different integral came from Stratonovich but
there are formulas which relating them with each other. So, it is enough
to consider the Ito integral. Both versions of stochastic integration have
advantages and disadvantages. Kunita shows in his book [55] that one can
view solutions as stochastic flows of diffeomorphisms. This brings the topic
into the framework of ergodic theory.
For ordinary differential equations x = f(x, £), one knows that unique solu
tions exist locally if / is Lipshitz continuous, in x and continuous in t. The
proof given for 1-dimensional systems generalizes to differential equations
in arbitrary Banach spaces. The idea of the proof is a Picard iteration of
an operator which is a contraction. Below, we give a detailed proof of this
existence theorem for ordinary differential equations. For stochastic differ
ential equations, one can do the same. We will do such an iteration on the
Hilbert space H?0 t, of C2 martingales X having finite norm
||X||T = E[supX2].
t<T
We will need the following version of Doob's inequality:
Lemma 4.19.1. Let X be a Cp martingale with p > 1. Then
E[sup|Xsn<(^-)Ê[|X,H.
s<t P-1
Proof. We can assume without loss of generality that X is bounded. The

general result follows by approximating X by X A k with k —> oo.
Define X* = sups<t \XS\P. From Doob's inequality
P[X>\]<E[\Xt\'lx*>\]
we get
rx*
E[\X*\P] = E[/ p\p-1 d\]
Jo
/•OO
= E[/ pAp_1l{x*>A} dX]

Jo
/»oo
= E[ / pA^PpT > A] dX]
Jo
/•OO
< E[/ pA'ÊpTtl-lx.^MA]

Jo
X*
pE[|Xt| / Xp-2 dX
Jo
-£-E[\Xt\ ■ (XT1} ■
p-1
Holder's inequality gives
E[\X*\P] < -JÊ[(X*)p](p~1VpE[\Xt\p]1/p
and the claim follows. □
Theorem 4.19.2 (Local existence and uniqueness of solutions). Let M be a

continuous martingale. Assume f(x, t) and g(x, t) are continuous in t and
Lipshitz continuous in x. Then there exists T > 0 and a unique solution
Xt of the SDE
dX = f(x,t)dM + g(x,t)ds
with initial condition X0 = X0.
Proof. Define the operator
SiX) = I fis, X.) dM. + f gis, X.) ds

Jo Jo
on £2-processes. Write 5(X) = SiiX) +<S2(X). We will show that on some

time interval (0,T], the map S is a contraction and that 5n(X) converges
in the metric \\\X - Y\\\T = E[sup,<T(X, - Y.)2], if T is small enough to
a unique fixed point. It is enough that for i = 1,2
ll|5«(A-)-54(y)|||T<(i/4)-||A:-y||T
then S is a contraction
| | | 5 ( x ) - 5 ( y ) | | | r < ( i / 2 ) . | | x - y | | T.
By assumption, there exists a constant K, such that
\fit, w) - fit, w')\<K- sup \w - w'\ .
(i) HISipO-SiiY)\\\T = \\\f*fis,X.)-fia,Y.) dM.\\\T< (1/4)- \\\X-

Y\\\t for T small enough.
Proof. By the above lemma for p = 2, we have
\\\S1(X)-S1(Y)\\\T = nsM^f(s1X)-f(s,Y)dMs)2}
rp
< 4E[(/ fit,X)-fit,Y)dMt)2)

Jo
= 4E[(/"T f it,X) - fit,Y))2 d(M,M)t]

Jo
rp
< 4K2E[[ sup|Xs-ys|2dt]
= 4K2/ |||X-y|||.ds
< ( l / 4 ) . | | | X - y | | | r,
where the last inequality holds for T small enough.
(ii) \\\S2iX) - S2iY)\\\T = \\\^9is,Xs)-9is,Ys)ds\\\T ^ (1/4) ■\\\X-

Y\\\t for T small enough. This is proved for differential equations in Banach
spaces. , • 4.i,«
The two estimates (i) and (it) prove the claim in the same way as in the
classical Cauchy-Picard existence theorem.
Appendix. In this Appendix, we add the existence of solutions of ordinary
differential equations in Banach spaces. Let X be a Banach space and/ an
interval in R. The following lemma is useful for proving existence of faxed
points of maps.
Lemma 4.19.3. Let X = Brix0) C X and assume <t> is a differentiable map

X -» X. If for all xeX, \\D<l>(x)\\ < |A| < 1 and
then 4> has exactly one fixed point in X.
Proof. The condition ||z - x0|| < r implies that

||*(s) - *o|| < IW*) - *(*o)ll + I W*o) - *o|| < Ar + (1 - A)r = r .
The map <f> maps therefore the ball X into itself. Banach's fixed point
theorem applied to the complete metric space X and the contraction^
implies the result.
Let / be a map from I x X to X . A differentiable map u : J -> X of an
open ball J C /in * is called a solution of the differential equation
x = /(*, x)
if we have for allt £ J the relation
ii(t) = f(t,u(t)) .
Theorem 4.19.4 (Cauchy-Picard Existence theorem). Let / : I x X —▶ X

be continuous in the first coordinate and locally Lipshitz continuous in the
second. Then, for every (to, xq) E/xA', there exists an open interval J C I
with midpoint to, such that on J, there exists exactly one solution of the
differential equation x = f(t, x).
Proof. There exists an interval J (to, a) = (to - a,t0 + a) C / and a ball

B(xo,b)1 such that
M = sup{||/(t,x)|| | (t,x) e J(t0,a) x B(x0,b)}
as well as
fc = sup{ll/(t'f)~/(^2)l1 I it,Xl),it,x2) € Jit0,a)xBix0,b),x1^x2}

| \X\ X%11
are finite. Define for r < a the Banach space
Xr = C(J(to,r),X) = {y : J(to,r) —▶ X, y continuous}

with norm
||y||= sup ||y(t)||
teJ(to,r)
Let Vrfi be the open ball in Xr with radius b around the constant map
t •-▶ xo. For every y e Vr,b we define
(j)(y) : *H+x0+ / f(syy(s))ds

Jt0
which is again an element in Xr. We prove now, that for r small enough,
^ is a contraction. A fixed point of (j> is then a solution of the differential
equation x = f(t, x), which exists on J = Jr(to). For two points t/i, y2 G Vr,
we have by assumption
\ \ fi a , y i i s ) ) - fi s 1 y 2 i ' > ) ) \ \ < k - \ \ y i i a ) - y 2 i 8 ) \ \ < k - \ \ y 1 - y 2 \ \

for every s € Jr. Thus, we have
IWvi)-tf(lfc)|| = ll//(«,Vi(*))-/(«,w(«))d*||
Jt0
< f Wfi»,Vlis))-fi»,V2i8))\\d8
Jt0
< kr - \\yi - y2\\ .
On the other hand, we have for every s G Jr
\ \ fi s , y i s ) ) \ \ < M
and so
IMxo)-so|| = || / /(*,*<>(*)) <k||< / \\fis,xois))\\ds<M-r.

Jt0 Jt0
We can apply the above lemma, if kr < 1 and Mr < 6(1 - kr). This is
the case, if r < b/(M + kb). By choosing r small enough, we can get the
contraction rate as small as we wish. Q
Definition. A set X with a distance function d(x, y) for which the following
properties
(i) d(y, x) = d(x, y)>0 for all x, t/ G X.
(ii) d(x, x) =0 and d(x, y) > 0 for x ^ y.
(iii) d(x, 2) < d(x, y) + d(j/, z) for all x, 2/, 2;. hold is called a metric space.
Example. The plane R2 with the usual distance d(x, y) = \x-y\. An other
metric is the Manhattan or taxi metric d(x,y) = \xi-yi\ + |x2 - 2/21-
Example. The set C([0,1]) of all continuous functions x(t) on the interval
[0,1] with the distance d(x,y) = maxt \x(t) - y(t)\ is a metric space.
Definition. A map 0 : X -> X is called a contraction, if there exists A < 1
such that d(<t>(x), <f)(y)) < A • d(x, y) for all x, y G X. The map 0 shrinks the
distance of any two points by the contraction factor A.
Example. The map </>(x) = \x + (1,0) is a contraction on R2.
Example. The map </>(x)(t) = sin(t)x(t) + t is a contraction on C([0,1])

because \</>{x)(t) - <Ky){t)\ = \ sin(t)| • |x(t) - y(t)\ < sin(l) • |x(t) - y(t)\.
Definition. A Cauchy sequence in a metric space (X, d) is defined to be a

sequence which has the property that for any e > 0, there exists n0 such
that |xn - xm| < e for n>no,m>n0.
A metric space in which every Cauchy sequence converges to a limit is
called complete.
Example. The n-dimensional Euclidean space
(Rn, d(x, y) = \x-y\ = ^x2 + -- + x2n)
is complete. The set of rational numbers with the usual distance
(Q,d(x,y) = \x-y\)
is not complete.
Example. The space C[0,1] is complete: given a Cauchy sequence xn, then
xn(t) is a Cauchy sequence in R for all t. Therefore xn(t) converges point
wise to a function x(t). This function is continuous: take e > 0, then \x(t) -
x(s)\ < \x(t) - xn(t)\ + \xn(t) - yn(s)\ -r \yn(s) - y(s)\ by the triangle
inequality. If s is close to £, the second term is smaller than e/3. For large
n, \x(t) - xn(t)\ < 6/3 and \yn(s) - y(s)\ < e/3. So, \x(t) - x(s)\ < e if
\t — s\ is small.
Theorem 4.19.5 (Banachs fixed point theorem). A contraction 0 in a com

plete metric space (X, d) has exactly one fixed point in X.
Proof, (i) We first show by induction that
d($n(x),<t>n(y))<\n-d(x,y)
for all n.
(ii) Using the triangle inequality and Ylk^k = i1 ~ *)_1» we Set for a11
x GX,
n—1 n—1 ^
dix,^x) < 5>(0fcx,0fe+1x) < J] Afcd(a;,<A(x)) < ^—^ • d(a;,^(x)) .

fc=0 fe=0
(iii) For all x G X the sequence xn = (f>n(x) is a Cauchy sequence because

by (i),(ii),
d(xn, xn+fc) < An • d(x, xk) < Xn • ;j—-^ ' ^, î) •
By completeness of X it has a limit x which is a fixed point of (j).
(iv) There is only one fixed point. Assume, there were two fixed points x, y
of (/>. Then
d(x, y) = d(0(x), 0(£)) < A • d(x, y) .
This is impossible unless x = y. D
Chapter 5
Selected Topics
5.1 Percolation
Definition. Let a be the standard basis in the lattice Zd. Denote with hd
the Cayley graph of Zd with the generators A = {eu ..., ed }. This graph
hd = (V, E) has the lattice Zd as vertices. The edges or bonds in that
graph are straight line segments connecting neighboring points x, y. Points
satisfying |x - y\ = J2i=i \x* ~ 2/<l = 1*
Definition. We declare each bond of hd to be open with probability p G

[0,1] and closed otherwise. Bonds are open ore closed independently of all
other bonds. The product measure Pp is defined on the probability space
fi = lie e{°> 1) of a11 configurations. We denote expectation with respect
to Pp with Ep[-].
Definition. A path in hd is a sequence of vertices (x0, xi,..., xn) such that

(x^ xi+i) = ei are bonds of hd. Such a path has length n and connects x0
with xn. A path is called open if all its edges are open and closed if all its
edges are closed. Two subgraphs of Ld are disjoint if they have no edges
and no vertices in common.
Definition. Consider the random subgraph of hd containing the vertex set

Zd and only open edges. The connected components of this graph are called
open clusters. If it is finite, an open cluster is also called a lattice animal.
Call C(x) the open cluster containing the vertex x. By translation invari-
ance, the distribution of C(x) is independent of x and we can take x = 0
for which we write C(0) = C.
275
97fi
Chapter 5. Selected Topics
Figure. A lattice animal.
Definition. Define the percolation probability 6(p) being the probability

that a given vertex belongs to an infinite open cluster.
*(p) = P[|C| = oo] = l-£p[|C|=n].

n=l
One of the goals of bond percolation theory is to study the function 0(p).
Lemma 5.1.1. There exists a critical value pc = pc(d) such that 0(p) = 0
for p < pc and 0(p) > 0 for p > pc. The value d ^ Pc(d) is non-increasing
with respect to the dimension pc(d + 1) < pc(d).
Proof. The function p h-> 6(p) is non-decreasing and 0(0) = 0,0(1) = 1. We

can therefore define
Pc = inf{p G [0,1] | 9(p) > 0 }.

The graph Zd can be embedded into the graph Zd' for d < d' by realizing Zd
as a linear subspace of Zd' parallel to a coordinate plane. Any configuration
in Ld projects then to a configuration in Ld. If the origin is in an infinite
cluster of Zd, then it is also in an infinite cluster of Zd'. □
Remark. The one-dimensional case d = 1 is not interesting because pc = 1

there. Interesting phenomena are only possible in dimensions d > 1. The
planar case d = 2 is already very interesting.
Definition. A self-avoiding random walk in Ld is the process ST obtained
by stopping the ordinary random walk Sn with stopping time
T(u) = inf{n G N | u(n) = v(m), m<n] .

Let a(n) be the number of self-avoiding paths in Ld which have length n.
The connective constant of Ld is defined as
A(d) = n
—lim
▶o a(n)^n .
5.1. Percolation 277
Remark. The exact value of A(d) is not known. But one has the elementary
estimate d < A(d) < 2d - 1 because a self-avoiding walk can not reverse
direction and so a(n) < 2d(2d - l)n_1 and a walk going only forward
in each direction is self-avoiding. For example, it is known that A(2) G
[2.62002,2.69576] and numerical estimates makes one believe that the real
value is 2.6381585. The number cn of self-avoiding walks of length n in L2 is
for small values cx = 4, c2 = 12, c3 = 36, c4 = 100, c5 = 284, c6 = 780, c7 =
2172, Consult [62] for more information on the self-avoiding random
walk.
Theorem 5.1.2 (Broadbent-Hammersley theorem). If d > 1, then
0<A(d)-x < pc(d) < pc(2) < 1 .
Proof. (i)pc(d)>X(d)~1.
Let N(n) < a(n) be the number of open self-avoiding paths of length n in
Ln. Since any such path is open with probability pn, we have
Ep[N(n)]=pna(n).
If the origin is in an infinite open cluster, there must exist open paths of
all lengths beginning at the origin so that
0(p) < Pp[N(n) > 1] < Ep[N(n)} = pna(n) = (pX(d) + o(l))n

which goes to zero for p < X(p)~l. This shows that pc(d) > A(d)-1.
(ii) pc(2) < 1.

Denote by L2 the dual graph of L2 which has as vertices the faces of L2 and
as vertices pairs of faces which are adjacent. We can realize the vertices as
Z2 + (1/2,1/2). Since there is a bijective relation between the edges of L2
and L2 and we declare an edge of L2 to be open if it crosses an open edge
in L2 and closed, if it crosses a closed edge. This defines bond percolation
onL2.
The fact that the origin is in the interior of a closed circuit of the dual
lattice if and only if the open cluster at the origin is finite follows from the
Jordan curve theorem which assures that a closed path in the plane divides
the plane into two disjoint subsets.
Let p(n) denote the number of closed circuits in the dual which have length
n and which contain in their interiors the origin of L2. Each such circuit
contains a self-avoiding walk of length n — 1 starting from a vertex of the
form (fc H-1/2,1/2), where 0 < fc < n. Since the number of such paths 7 is
at most na(n— 1), we have
p(n) < na(n — 1)

278 Chapter 5. Selected To p i c s
and with q = 1 - p
oo oo
J2 p[7 is closed] < £ qnn(j(n - 1) = £ «nfaA(2) + o(l))n-x

7 n=l n=l
which is finite if gA(2) < 1. Furthermore, this sum goes to zero if q -+ 0 so

that we can find 0 < 5 < 1 such that for p > S
^P[7 is closed] < 1/2.

7
We have therefore
P[|C| = oo] = P[no 7 is closed] > 1 - ]T P[7 is closed] > 1/2

7
so thatpc(2) < S < 1. □
Remark. We will see below that even pc(2) < 1 - A(2)"1. It is however
known that pc(2) = 1/2.
Definition. The parameter set p < pc is called the sub-critical phase, the
set p > pc is the supercritical phase.
Definition. For p < pc, one is also interested in the mean size of the open
cluster
x(p) = Ep[|C|].
For p > pc, one would like to know the mean size of the finite clusters
X/(p) = Ep[|C|||C|<oo].
It is known that \(p) < oo for p < pc but only conjectured that xf(p) < oo
for p > pc.
An interesting question is whether there exists an open cluster at the critical
point p = pc. The answer is known to be no in the case d = 2 and generally
believed to be no for d > 3. For p near pc it is believed that the percolation
probability 0(p) and the mean size x(p) behave as powers of \p - pc\. It is
conjectured that the following critical exponents
7 = -pSpc
iimlog ^xiiL
\p- pc\
0 = lim ***«
P\Pc lOg \p~ pc\
*-i = - lim lQgP^Cl ^]

n->oo log n
exist.
Percolation deals with a family of probability spaces (Q,A,PP), where

^ = {0? 1}L is the set of configurations with product <j-algebra A and
product measure Pp = (p, 1 — p)h<i.
Definition. There exists a natural partial ordering in ft coming from the
ordering on {0,1}: we say u> < u/, if u(e) < a/(e) for all bonds e G L .
We call a random variable X on (ft, AP) increasing if w < J implies
X(u) < X(u/)- It is called decreasing if -X is increasing. As usual, this
notion can also be defined for measurable sets A G A: a set A is increasing
if \a is increasing.
Lemma 5.1.3. If X is a increasing random variable in C1 (ft, Pq)nC (ft, Pp),

then '•'
EP[X] < Eq[X]
if p < q.
Proof. If X depends only on a single bond e, we can write EP[X] = pX(l) +

(1 - p)X(O). Because X is assumed to be increasing, we have ÊP[X] =
X(l) - X(0) > 0 which gives EP[X] < Eq[X] for p < q. If X depends only
on finitely many bonds, we can write it as a sum X = Yli=ix* of variables
Xi which depend only on one bond and get again
±Ep[X] = J2(Xi(l)-Xi(0))>0.
In general we approximate every random variable in C (ft, Pp) fl C (ft, Pq)

by step functions which depend only on finitely many coordinates X*. Since
Ep[Xi] -+ EP[X] and Eq[Xi] -+ Eg[X], the claim follows. □
The following correlation inequality is named after Fortuin, Kasterleyn and
Ginibre (1971).
Theorem 5.1.4 (FKG inequality). For increasing random variables X,F G

£2(ft,Pp), we have
Ep[XY]>Ep[X].Ep[y].
Proof. As in the proof of the above lemma, we prove the claim first for ran
dom variables X which depend only on n edges ci, e2,..., en and proceed
by induction.
(i) The claim, if X and Y only depend on one edge e.

We have
(XM - X(J)(Y(uj) - Y(J)) > 0
since the left hand side is 0 if w(e) = u/(e) and if 1 = w(e) = u/(e) = 0, both
factors are nonnegative since X, Y are increasing, if 0 = w(e) = a/(e) = 1
both factors are non-positive since X, Y are increasing. Therefore
o < VJ (i(w)-i(w'))(r(w)-y(w'))pp[w(e)=(7]pp[w(e) = (Tl

<t,<t'€{o,i}
= 2(EP[XF]-EP[X]EP[F]).
(ii) Assume the claim is known for all functions which depend on A; edges
with k < n. We claim that it holds also for X,Y depending on n edges
ei,e2,...,en.
Let Ak = Aiei, ...ek) be the a-algebra generated by functions depending
only on the edges ek. The random variables
Xk = Ep[X\Ak],Yk = Ep\Y\Ak]
depend only on the eu ■ ■ ■, ek and are increasing. By induction,
EplXn-iYn-i] > EplXnÊ^Yn^} .

By the tower property of conditional expectation, the right hand side is
EP[X}EP[Y]. For fixed d,...,en^, we have iXY)n^ > Xn^Y^ and so
EP[XF] = Ep[(XF)n_!] > EriXn-iYn-t] .

(iii) Let X, Y be arbitrary and define Xn = Ep[X\An), Yn = Ep[F|.4n]. We
know from iii) that Ep[XnYn] > Ep[Xn]Ep[Fn]. Since Xn = E[X\An] and
Yn - E[X\An] are martingales which are bounded in £2(ft,Pp), Doob's
convergence theorem (3.5.4) implies that Xn -» X and Yn -+ Y in £2 and
therefore E[Xn) -> E[X] and E[Fn] -▶ E[Y]. By the Schwarz inequality, we
get also in C1 or the C2 norm in (ft, .4, Pp)
||*»y„-xy||i < IIPfn-JOrnlli + HW-y)!!!

< ll^n-x||2||yn||2 + ||x||2||yn-y||2
< c(||xn-A-||2 + ||yB-r||2)->o
where C = max(||JC||a> ||F||2) is a constant. This means EJXnYn] -»
Remark. It follows immediately that if A, B are increasing events in ft

then PP[A DB}> PP[A] ■ PP[B].
Example. Let Tt be families of paths in L? and let At be the event that

some path in r» is open. Then Ai are increasing events and so after applying
the inequality fc times, we get
k k
pP[fv«]^npp[^]-
i=l i=l
We show now, how this inequality can be used to give an explicit bound for
the critical percolation probability pc in L2. The following corollary belongs
still to the theorem of Broadbent-Hammersley.
Corollary 5.1.5.
Pc(2)<(l-A(2)-1)
Proof. Given any integer N G N, define the events

FN = {3 no closed path of length < N in L?}
GN = {3 no closed path of length > N in L?} .
We know that FNnGNC {\C\ = oo}. Since FN and GN are both increas
ing, the correlation inequality says Pp[Fjv H GN] > ?p[Fn} • Pp[Gn]- We
deduce
0(p) = PP[|C| = oo] = PP[FN H GN] > PP[FN] - PP[GN] .

If (1 - p)A(2) < 1, then we know that
oo
PP[GCN)< £(l-pyW(n-l)
n=N
which goes to zero for N -▶ oo. For N large enough, we have therefore
PP[GN] > 1/2. Since also PP[FN] > 0, it follows that 9P > 0, if (1 -p) A(2) <
1 or p < (1 - A(2)_1) which proves the claim. n
Definition. Given A € A and w € ft. We say that an edge e € Ld is pivotal

for the pair (A,w) if UM £ Ufa), where we is the unique configuration
which agrees with u except at the edge e.
Theorem 5.1.6 (Russo's formula). Let A be an increasing event depending

only on finitely many edges of Ld. Then
±Pp[A]=Ep{NiA)},
dp
where N(A) is the number of edges which are pivotal for A.
Proof, (i) We define a new probability space.

The family of probability spaces (ft, A, Pp), can be embedded in one prob
ability space
([0,l]Ld,B([0,l]L<i),P),
where P is the product measure dxL . Given a configuration 77 G [0,1]L and
p G [0,1], we get a configuration in £2 by defining rjp(e) = 1 if 77(e) < p and
rjp = 0 else. More generally, given p G [0,1]L , we get configurations np(e) =
1 if 77(e) < p(e) and np = 0 else. Like this, we can define configurations
with a large class of probability measures Pp = IIe€Ld(p(e)> 1 — P(e)) wrtn
one probability space and we have
PP[A] = P{rjpeA}.
(ii) Derivative with respect to one p(f).

Assume p and p' differ only at an edge / such that p(f) < p'(f). Then
{t7p G A} C {t7p/ G A} so that
PP,[A]-?P[A] = P[rjp,eA]-P[npeA}
= P[v6il;?j,M]
= (p,(/)"P(/))Pp[/ Pivotal for A].
Divide both sides by (p'(f) -p(f)) and let p'(f) —▶ p(f). This gives
a
PP[A] = Pp[f pivotal for A]
dp(f)
(iii) The claim, if A depends on finitely many edges. If A depends on finitely
many edges, then PP[A] is a function of a finite set {p(fi)}iLi of edge
probabilities. The chain rule gives then
d m ft
-^pp\A\ = Ea^)Pp[A1|p=(p>p'p--p)
m
= ^2 Ppifi Pivotal for A]
i=l
= EP[AT(A)].
(iv) The general claim.

In general, define for every finite set F C E
PF(e) =P+l{eeF}$
where 0 PPF[A]
and therefore
i(Pp+5[A] - PP[A}) > i(PPF[A] - PP[A}) -+ J2 pv\* Pivotal for A\

e£F
as 6 —> 0. The claim is obtained by making F larger and larger filling out
E. □
Example. Let F = {eu e2,..., em} C E be a finite set in of edges.

A = {the number of open edges in F is > fc} .
An edge e G F is pivotal for A if and only if A \ {e} has exactly fc - 1 open
edges. We have
Pp[eispivotal] = (^:i1)Pfc-1(l-Pr-fc
so that by Russo's formula
±PP[A) = £ PP[e is pivotal] = m ( ^ ) 1^(1 - P)"1^ ■
Since we know P0[A] = 0, we obtain by integration
Remark. If A does no more depend on finitely many edges, then PP[A]

need no more be differentiable for all values of p.
Definition. The mean size of the open cluster is x(p) = EP[|C|].
Theorem 5.1.7 (Uniqueness). For p < pci the mean size of the open cluster
is finite x(p) < °°-
The proof of this theorem is quite involved and we will not give the full
argument. Let S(n,x) = {y G Zd \ \x - y\ = £ti \xi\ < n] be the ball of
radius n around x in Zd and let An be the event that there exists an open
path joining the origin with some vertex in 5S(n, 0).
Lemma 5.1.8. (Exponential decay of radius of the open cluster) If p < pc,
there exists ^p such that Pp[An] < e'n^.
Proof. Clearly, |S(n,0)| <Cd-(n + l)d with some constant Cd. Let M =
max{n | An occurs }. By definition of pc, if p < pc, then PP[M < oo] = 1.
We get
EP[|C|] < 2Ep[|C||Af = n]-Pp[M = n]
< ^|5(n,0)|Pp[An]
n
< ^Cd(n + l)de-n^<oo
284
Chapter 5. Selected Topics
D
Proof. We are concerned with the probabilities <?p(n) = Pp[An}. Sine An

are increasing events, Russo's formula gives
g'pin) = Ep[NiAn)} ,
where NiAn) is the number of pivotal edges in An. We have
9pin) = VJ pp[e pivotal for A]

e
= 5Z ZPple °Pen and Pivotal for A]

e ^
= ^2 "Pp iA n ie Pôtal for A}]

e ^
= £ pPp[^ n (e Pivotal for 4}I4 • PP[4]

e ^
e ^
= J2l^mA)\A].gPin)
so that
9v(n) 1
^y-EP[iV(^)Mn]
By integrating up from a to /?, we get
r 1
9a(n) = 9(3(n) exp(- / -Ep[AT(An) | ^J dp)
«/a P
f0
< g0in)expi- / Ep[AT(,4n) | An] dp)
J ex.
< exp(- / Ep[AT(^ln) | An] dp) .

J (X
One needs to show then that Ep[N(An) \An] grows roughly linearly when
p < pc. This is quite technical and we skip it. □
Definition. The number of open clusters per vertex is defined as
oo -
«(p) = Ep[|C|-1] = Vjipp[|C7| = n].

n=l
Let Bn the box with side length 2ra and center at the origin and let Kn be
the number of open clusters in Bn. The following proposition explains the
name of k.
Proposition 5.1.9. In £1(n,^,Pp) we have
Kn/\Bn\ -* nip) .
Proof. Let Cn(x) be the connected component of the open cluster in Bn

which contains x 6 Zd. Define T(x) = \Cix)\~l.
(i)E*6B„rn(x) = #n.
Proof. If S is an open cluster of Bn, then each vertex x G E contributes
|E|_1 to the left hand side. Thus, each open cluster contributes 1 to the
left hand side.
(ii) jt\>\k\ £*6B„ rW where r(*) = \Cix)\~l.

Proof. Follows from (i) and the trivial fact F(x) < rn(x).
(i") ^ £*6B„ *» - E^r(0)] = «&>)•

Proof. r(x) are bounded random variables which have a distribution which
is invariant under the ergodic group of translations in Zd. The claim follows
from the ergodic theorem.
(iv) lim inf nôo p^ > nip) almost everywhere.

Proof. Follows from (m) and [iii).
(v) 22xeB(n) rn(x) < X)X€B(n) r(x) + Ex~5b„ rn(x), where x ~ Y means
that x is in the same cluster as one of the elements y e Y c Zd.
(vi) r^ £x6Bn rn(x) < ^ £x6Bn m + Jî. n
Remark. It is known that function «(p) is continuously differentiable on

[0,1], It is even known that k and the mean size of the open cluster xip) are
real analytic functions on the interval [0,pc). There would be much more
to say in percolation theory. We mention:
The uniqueness of the infinite open cluster:
For p> pc and if %c) > 0 also for p = pc, there exists a unique infinite
open cluster.
Regularity of some functions 9{p)
For p > Pc, the functions 0(p),xfip),nip) are differentiable. In general,
Dip) is continuous from the right.
The critical probability in two dimensions is 1/2.
5.2 Random Jacobi matrices
Definition. A Jacobi matrix with IID potential V^n) is a bounded self-
adjoint operator on the Hilbert space
oo
Z2(Z) = {(...,x_1,x0,x1,x2...)| Y, 4 = 1}
k=—oo
of the form
Luu(n) = Y, <m) + V„(n)u(n) = (A + Vu)(u)(n) ,

|m—n| = l
where K,(n) are IID random variables in £°°. These operators are called
discrete random Schrodinger operators. We are interested in properties of
L which hold for almost all u e ft. In this section, we mostly write the
elements u of the probability space (ft, A, P) as a lower index.
Definition. A bounded linear operator L has pure point spectrum, if there

exists a countable set of eigenvalues A* with eigenfunctions ^ such that
L(f)i = X^ and ^ span the Hilbert space l2(Z). A random operator has
pure point spectrum if L^ has pure point spectrum for almost all u € ft.
Our goal is to prove the following theorem:
Theorem 5.2.1 (Prohlich-Spencer). Let V(n) are IID random variables with
uniform distribution on [0,1]. There exists A0 such that for A > A0, the
operator L^ = A + A • K, has pure point spectrum for almost all u.
We will give a recent elegant proof of Aizenman-Molchanov following [94].
Definition. Given E € C \ R, define the Green function
Gu,(m,n,E) = [(L0J-E)-1]mn.
Let p = pu be the spectral measure of the vector e0. This measure is
defined as the functional C(R) -♦ R, / .-* /(Lw)oo by f(L„)0o = E[/(L)00].
Define the function
f(Z) =fdm.
J r V- z
It is a function on the complex plane and called the Borel transform of the
measure p. An important role will play its derivative
= f M\)
h iv- z f
5.2. Random Jacobi matrices 287
Definition. Given any Jacobi matrix L, let La be the operator L + aP0,
where Po is the projection onto the one-dimensional space spanned by S0.
One calls La a rank-one perturbation of L.
Theorem 5.2.2 (Integral formula of Javrjan-Kotani). The average over all

specral measures dpa is the Lebesgue measure:
dpot da = dE .
/JR
r
Proof. The second resolvent formula gives
(LQ - z)-1 - (L - z)-1 = -a(La - z^PoiL - z)~l .
Looking at 00 entry of this matrix identity, we obtain
Fa(z) - F(z) = -aFa(z)F(z)
which gives, when solved for Fa, the Aronzajn-Krein formula
taKZ)
F { z ) = 1 *+" (aFiz)-
*)
We have to show that for any continuous function / : C —> C
Jf r f fix) dpaix)
Jr da= Jf fix) dEix)
and it is enough to verify this for the dense set of functions
{/,(*) = (x - z)-1 -(x + i)-1 \z e C \ R} .

Contour integration in the upper half plane gives JRfz(x) dx = 0 for
Im(z) < 0 and 2ni for lm(z) > 0. On the other hand
J fz(x)dpa(x) = Fa(z) - Fa(-i)

which is by the Aronzajn-Krain formula equal to
1 1
hz(a) :=
a-^F(z)-1 a + F(-2)"1
Now, if ±Im(z) > 0, then ±lmF(z) > 0 so that ±lmF(z)~1 < 0. This
means that hz(a) has either two poles in the lower half plane if Im(z) < 0
or one in each half plane if Im(z) > 0. Contour integration in the upper
half plane (now with a) implies that JRhz(a) da = 0 for lm(z) < 0 and
2-ki for Im(z) > 0. □
In theorem (2.12.2), we have seen that any Borel measure p on the real line
has a unique Lebesgue decomposition dp = dpac + dpsing = dpac + dpsc +
dppp. The function F is related to this decomposition in the following way:
Proposition 5.2.3. (Facts about Borel transform) For e -+ 0, the measures

7r~1ImF(£' + ie) dE converges weakly to p.
dVsin9({E I lmF(E + i0) = oo }) = 1,
dn({E0}) = limeô lmF(E0 + ze)e,
dvac(E) = 7r-1ImF(E + iO) dE.
Definition. Define for a ^ 0 the sets
Sa = {xeR\F(x + iO) = -a~\ F'(x) = oc}

Pa = {xeR\F(x-riO) = -a-\ F'(x)<oo}
L = {x E R | lmF(x +10) ^ 0 }
Lemma 5.2.4. (Aronzajn-Donoghue) The set Pa is the set of eigenvalues of

La. One has (dfia)sc(Sa) = 1 and (dfia)ac(L) = 1. The sets Pa,Sa,L are
mutually disjoint.
Proof. If F(E + iO) = -1/a, then
lim c ImFa(£7 + ie) = (a2F'(E))~2
since F(E-rie) = -l/a+ieF'(x)+o(e) HF'(E) < oo and e-1Im(l+aF) ->

oo if F'(E) = oo which means e|l + aF|_1 -^ 0 and since F -> -1/a, one
gets c|F/(l + aF)|-* 0.
The theorem of de la Vallee Poussin (see [88]) states that the set
{E\ \Fa(E + iO)| = oo }
has full (dfj,a)sing measure. Because Fa = F/(l -f aF), we know that

\Fa(E + iO)| = oo is equivalent to F(E + iO) = -1/a. □
The following criterion of Simon-Wolff [96] will be important. In the case of

IID potentials with absolutely continuous distribution, a spectral averaging
argument will then lead to pure point spectrum also for a = 0.
Theorem 5.2.5 (Simon-Wolff criterion). For any interval [a, b] C R, the

random operator L has pure point spectrum if
F'(E) < oo
for almost almost all E G [a, 6].
Proof. By hypothesis, the Lebesgue measure of S = {E \ F'(E) = oo } is

zero. This means by the integral formula that d/xa(5) = 0 for almost all a.
The Aronzajn-Donoghue lemma (5.2.4) implies
Va(Sa n [a,b}) = /ia(Ln[a,6]) = 0
so that pa has only point spectrum. □
Lemma 5.2.6. (Formula of Simon-Wolff) For each E G R, the sum

Y^nez \(L — E — ie)on |2 increases monotonically as e \ 0 and converges
point wise to F'(E).
Proof. For e > 0, we have
J2\iL-E-ie)^\2 = \\iL-E-ie)-H0\\2
nez
= |[(Z,-E-ie)-1(L-.E + *e)-1]oo|
f Mx)
iR(x-S)2 + e2
f r o m w h i c h t h e m o n o t o n i c i t y a n d t h e l i m i t f o l l o w. □
Lemma 5.2.7. There exists a constant C, such that for all a, /3 G C
J/ o \ x - a \ ^ 2 \ x - fi \ - ^ 2 d x > C Jj o \ x - p \ ~ 1 / 2 d x .
Proof. We can assume without loss of generality that a G [0,1], because

replacing a general a G C with the nearest point in [0,1] only decreases the
290 Chapter 5. Selected Topics
left hand side. Because the symmetry anl-a leaves the claim invariant,
we can also assume that a G [0,1/2]. But then
f1 \x _ a|l/2|x _ fl-1/2 dx > ( 1)1/2 j' {x _/5|-l/2 dx

Jo 4 y3/4
The function
/3/4l*-/3|-1/2cte
K0) = -r
/olx-aMx-zSI-^dx
is non-zero, continuous and satisfies ft(oo) = 1/4. Therefore
C := inf h(J3) > 0 .
The next lemma is an estimate for the free Laplacian.
Lemma 5.2.8. Let f,g e i°°(Z) be nonnegative and let 0 < o < (2d)-1.
(1 - oA)/ <g^f< (1 - aA)-x5 .
[(1 - aA)-1]^- < (2do)lJ'-"(l - 2da)~1 .
Proo/. Since ||A|| < 2d, we can write (1 - aA)"1 = Em=o(aA)m which is
preserving positivity. Since [(aA)"% = 0 for m < \i - j\ we have
oo oo
[ ( o AH y= j ; [(a AH y< ^ (2 d a )m.

m=|i-j| m=|i-j|
We come now to the proof of theorem (5.2.1):
Proof. In order to prove theorem (5.2.1), we have by Simon-Wolff only to

show that F'(E) < oo for almost all E. This will be achieved by proving
EfF'^)1/4] < oo. By the formula of Simon-Wolff, we have therefore to
show that
suPE[(^|G(n,0,2)|2)1/4]<oo.
^c „
Since
EiG(».o^)ia)1/4ÊiG(n'°^)i1/a.
we have only to control the later the term. Define gz(n) = G(n,0,z) and
kz(n) = E[\gz(n)\1/2}. The aim is now to give an estimate for
5>(n)
which holds uniformly for Im(z) ^ 0.
(i)
E[\XVin) - z\^2\9zin)\1/2} < 5n,0 + £ hin + j) .
Proof. (L - z)gzin) = Sn0 means
iXVin) - z)gzin) = 8n0 - ^ 0*(™ + i) •

bl=i
Jensen's inequality gives
E[\XVin) - z\1/2\gzin)\1/2} < Sn0 + £ M« + J') .
(U) E[|AV(n) - z|1/2|<7*(n)|1/2] > CA1/2*^) •
Proof. We can write gzin) = ;4/(AV(n) + B), where A, B are functions of

{^(/)}i#n- The independent random variables V(fc) can be realized over
the probability space Q = [0, l]z = flfcez fi(fc)- We avera6e now \xv(n) ~
z\1/2\gzin)\1/2 over fi(n) and use an elementary integral estimate:
/ l^-^l1/2l^l1/2dt; = \A\w[1\v-z\-*\\v + B\-*\-l/2*>

jQ(n) |A« + BI1'2 Jo
> C\A\V2 f \v + B\-l\-1'2dv
Jo
= CX1'2 f'lA/iXv + B)^2
Jo
= E[gzin)l'2\ = kzin) .
(iii)
fcz(n) < (CA1/2)-1 I £ k*(n + 3) + *»o
Proof. Follows directly from (i) and (ii).
(iv)
(1 - CA1/2A)fc < 6n0 .
Proof. Rewriting (iii).
(v) Define a = CA1/2.
kz(n) < a~1(2d/a)^(l - 2d/a)~l .

Proof. For lm(z) ^ 0, we have kz G l°°(Z). From lemma (5.2.8) and (iv),
we have
k(n) < a-'[(l - A/a)-1^ < a~l(-)^(l - l)-i .

a a
(vi) For A > 4C~2, we get a uniform bound for £n kz(n).
Proof. Since CA1/2 < 1/2, we get the estimate from (v).
(vii) Pure point spectrum.
Proof. By Simon-Wolff, we have pure point spectrum for La for almost all
a. Because the set of random operators of La and L0 coincide on a set of
measure > 1 - 2a, we get also pure point spectrum of L^ for almost all
5.3 Estimation theory

Estimation theory is a branch of mathematical statistics. The aim is to
estimate continuous or discrete parameters for models in an optimal way.
This leads to extremization problems. We start with some terminology.
Definition. A collection (Q,A,Pe) of probability spaces is called a statis

tical model. If X is a random variable, its expectation with respect to the
measure Pe is denoted by Ee[X], its variance is Var#[X] = Ee[(X-Ee[X])2].
If X is continuous, then its probability density function is denoted by fe.
In that case one has of course Ee[X] = JQ fe(x) dx. The parameters 9 are
taken from a parameter space 9, which is assumed to be a subset of R or
Rk.
Definition. A probability distribution p = p(6) d6 on (6,B) is called an

a priori distribution on 6 C R. It allows to define the global expectation
E[X] = fQEe[X]dp(6).
Definition. Given n independent and identically distributed random vari

ables Xi,..., Xn on the probability space (fi, A, Pa), we want to estimate
a quantity g(6) using an estimator T(uo) = t(Xx(uj),..., Xn(u)).
Example. If the quantity g(0) = Ee[Xi] is the expectation of the ran

dom variables, we can look at the estimator T(uj) = £ £n=1 ^(cj), the
arithmetic mean. The arithmetic mean is natural because for any data
xi,... ,xn, the function f(x) = £ILi(^ - x)2 is minimized by the arith
metic mean of the data.
Example. We can also take the estimator T(uj) which is the median of
Xi(lj), ...,Xn(uj). The median is a natural quantity because the function
f(x) = Yl7=i \xi ~ x\ is minimized by the median. Proof. \a - x\ + |6 - x\ =
5.3. Estimation theory 29^
|6 _ a\ + C(x), where C(x) is zero if a < x 6 and D(x) = a - x if x < a. If n = 2m + 1 is odd, we have f(x) =
E£i ki-^n+i-il+E^^ C(xj)+ZXj<Xm D(Xj) which is minimized for
x = Xm- If n = 2m, we have /(x) = ^^=i l^-^n+i-i|+ZXj>xm+1 C'fe'H
E £>(#,) which is minimized for x G [xm,xm+i].
Example. Define the bias of an estimator T as
B(0) = Be[T\ = Ee[T\-g(0).
The bias is also called the systematic error. If the bias is zero, the estimator
is called unbiased. With an a priori distribution on ©, one can define the
global error B(T) = JQ B(6) dp,(0).
Proposition 5.3.1. A linear estimator T(u) = £"=1 atXû) with £» a» =

1 is unbiased for the estimator g(0) = Ee[Xi\.
Proof. Ee[T] = £" a,E,[X,] = E,[X,]. □
Proposition 5.3.2. For g(6) = Vaie[Xi\ and fixed mean m, the estimator
T = - Yln=i(xi ~ m)2 is unbiased- If tne mean is unknown, the estimator
T = î Zti(Xi ~ X)2 with X = £ ElLi *i ^ unbiased.
Proof, a) E,[r] = 1 £i=i(*« - m)2 = Var,[T] = ff(0).
b) For T = 1 £•(*; " **)2, we get
E*[T] = E,[X2]-E,[^^X^]
= Ee[X2] - l-Ee[Xl\ - ^^W,]2

n n
= ii-l)Ee[X?]-Êe[Xi?
= — Va i 9 [ X i ] .
n
Therefore n/(n - l)T is the correct unbiased estimate. □
Remark. Part b) is the reason, why statisticians often take the average of
r^br\(xi -x)2 as an estimate for the variance of n data points Xi with mean
m if the actual mean value m is not known.
Definition. The expectation of the quadratic estimation error
E r r, [ T ] = E , [ ( T- ^ ) ) 2 ]
is called the risk function or the mean square error of the estimator T. It
measures the estimator performance. We have
Err$[T\=Vax9[T\ + B$[T],
where B$[T] is the bias.
Example. If T is unbiased, then Err^T] = Vaxe[T\.
Example. The arithmetic mean is the "best linear unbiased estimator".

Proof. With T = Y,i (x%Xu where £• a* = 1, the risk function is
Err*[T] = Var^T] = J^Var^] .

i
It is by Lagrange minimal for ai = 1/n.
Definition. For continuous random variables, the maximum likelihood func

tion t(x\,..., xn) is defined as the maximum of 6 i-* Lq(x\, ..., xn) :=
fe(xi) fe(xn)- The maximum likelihood estimator is the random vari
able
T(u) = t(X1(u),...,Xn(u)).
For discrete random variables, L$(xi,..., xn) would be replaced by Po[Xi =
Xi,..., An = xnJ.
One also looks at the maximum a posteriori estimator, which is the maxi
mum of
0 ■-▶ Le(xu..., xn) = fg(xi) fe(xn)p(0) ,
where p(6) d6 was the a priori distribution on 6.
Definition. The minimax principle is the aim to find
minmax.R(0,T) .
t e v
The Bayes principle is the aim to find
min [ (R(6,T)dp(6) .
t jo
Example. Assume fe(x) = Ie-'*"*1. The maximum likelihood function
L,(x1,...,xn) = ê-^"^-^
is maximal when ]T\ |x* - 0\ is minimal which means that t(x\,...,xn) is

the median of the data x\,..., xn.
5.3. Estimation theory 295
Example. Assume f0(x) = 0xe~e/x\ is the probability density of the Pois
son distribution. The maximal likelihood function
e£ilog(0)*i-n0
^ ( x i , . . . , x n ) = X i\ .\ . .d Tj - r iI-
is maximal for 0 = Yh=i xi/n-

Example. The maximum likelihood estimator for 0 = (m, a2) for Gaussian
-jx-rn)2
distributed random variables /«(*) = 7^=2 e"^5"^ has the maximum
likelihood function maximized for
t(X!, . . . , X„) = (-5>i, - ^(Xi - X)2) .

i i
Definition. Define the Fisher information of a random variable X with

density /# as
If 0 is a vector, one defines the Fisher information matrix
IiJ(0) = Jf-^-fedx.
£1
Lemma 5.3.3. 7(0) = Var«[^].
fZa.1 =
Proof. E[£] -= /n /*dx = 0 so that
Vare[f]
J e = Ee[(f)2]
je
a
Lemma 5.3.4. 1(61) = -Ee[(log(/<,)"]
Proof. Integration by parts gives:
Epog(/,)"] = [logifeY'fe dx = - j \ogife)'f'e dx = - jif'e/feffe dx

□
Definition. The score function for a continuous random variable is defined
as the logarithmic derivative pe = f'e/fe. One has 1(0) = Ee[p2} = Vare[pe).
Example. If X is a Gaussian random variable, the score function pe =

/'(0)//(0) = ~(x - m)/(cr2) is linear and has variance 1. The Fisher in
formation / is 1/a2. We see that Var[X] = 1/J. This is a special case
n = 1> T = X, 0 = m of the following bound:
Theorem 5.3.5 (Rao-Cramer inequality).
****W--
In the unbiased case, one has
1
Erv0[T] >
nli$)
Proof l)0 + Bid) = Ee[T} = Jtix1,...,xn)Leix1,...,xn)dx1...dxn.

2)
1 + B'ie) = Jtixu...,xn)L'eix1,...,xn)dx1---dxn
= hix1,...,xn)L'^--^Xn\dx1...dxn
J Leixi,...,xn)
3) ! = / Leixi,..., xn) dxi • • • dxn implies
0 - J L'eixlt... ,xn)/Leixlt... ,xn) = E[L'0/Le] .
4) Using 3) and 2)
Cov[T,L'e/L0] = Ee[TL'e/Le}-0
= i + B'ie).
5)
il + B'ie))2 = Cov2[T,^]
< Var*[T]Vare[^]
= Va r e [ T ] £ E , [(feixi)\
(M^)2]
= Var„[T] n/(0) ,
5.3. Estimation theory 297
where we used 4), the lemma and
n
L'eILe = YJf'e{xi)lfeixi)-
i=l
□
Definition. Closely related to the Fisher information is the already defined

Shannon entropy of a random variable X:
Si6) = - J felogife) dx ,
as well as the power entropy
Nie) =27re
J-e*m.
Theorem 5.3.6 (Information Inequalities). If X, Y are independent random

variables then the following inequalities hold:
a) Fisher information inequality: Ix\y — ^x* + ^y1-
b) Power entropy inequality: Nx+y > Nx + Ny.
c) Uncertainty property: IxNx > 1-
In all cases, equality holds if and only if the random variables are Gaussian.
Proof, a) Ix+y < c2Ix + (1 — c)2Iy is proven using the Jensen inequal
ity (2.5.1). Take then c = Iy/(IX + W)-
b) and c) are exercises. □
Theorem 5.3.7 (Rao-Cramer bound). A random variable X with mean m

and variance a2 satisfies: Ix > 1/cr2. Equality holds if and only if X is the
Normal distribution.
Proof. This is a special case of Rao-Cramer inequality, where 0 is fixed,

n = 1. The bias is automatically zero. A direct computation giving also
uniqueness: E[(aX + b)p(X)] = J (ax + b)f'(x) dx = —aff(x) dx = —a
implies
0 < E[ipiX) + iX-m)/a2)2}
= E\ipiX)2] + 2E[(X - m)piX)]/a2 + E[(X - m)2/a4]
< Ix- 2/a2 + 1(72 •
Equality holds if and only if px is linear, that is if X is normal. □

We see that the normal distribution has the smallest Fisher information
among all distributions with the same variance a2.
5.4 Vlasov dynamics

Vlasov dynamics generalizes Hamiltonian n-body particle dynamics. It deals
with the evolution of the law P* of a discrete random vector X*. If P* is
a discrete measure located on finitely many points, then it is the usual
dynamics of n bodies which attract or repel each other. In general, the
stochastic process Xt describes the evolution of densities or the evolution
of surfaces. It is an important feature of Vlasov theory that while the ran
dom variables Xt stay smooth, their laws P* can develop singularities. This
can be useful to model shocks. Due to the overlap of this section with geom
etry and dynamics, the notation slightly changes in this section. We write
X* for the stochastic process for example and not Xt as before.
Definition. Let ft = M be a 2p-dimensional Euclidean space or torus with
a probability measure m and let N be an Euclidean space of dimension 2q.
Given a potential V : Rq -+ R, the Vlasov flow X* = (/*,#*) : M -▶ TV is
defined by the differential equation
f = 9,9'- - [ VF(/(u;)-/(n))dm(n).
Jm
These equations are called the Hamiltonian equations of the Vlasov flow.
We can interpret X1 as a vector-valued stochastic process on the probability
space (M, A, m). The probability space (M, A, m) labels the particles which
move on the target space N.
Example. If p = 0 and M is a finite set ft = {uu..., wn}, then X* describes

the evolution of n particles (/»,#) = X(ui). Vlasov dynamics is therefore
a generalization of n-body dynamics. For example, if
y(xi,....xn) = X^ >
i
then W(x) = x and the Vlasov Hamiltonian system
/ = 9,9(u) = ~ /M - f(v) dm(n)

Jm
is equivalent to the n-body evolution
/t = 9%
3=1
In a center of mass coordinate system where Yl7=i fi(x) = ®>tms simplifies

to a system of coupled harmonic oscillators
d2
5.4. Viasov dynamics 299
Example. If N = M = R2 and m is a measure, then the process Xt

describes a volume-preserving deformation of the plane M. In other words,
X1 is a one-parameter family of volume-preserving diffeomorphisms in the
plane.
H — <▶— < •——( > — — (▶— ( h——<>—I ▶— < > — — <▶— - <▶
I I > I I ( > I > O I > <> (> —<»——< h—( h—« H—< h—I H-—H h—I h—I H—I
h — (▶— H I—(»—« »——<>——< >——« l-HKHI—H
▶— <▶— — I > — (▶— — ( >——<>-—< H—H »—H I——( >—H
h—Ih-—I >—( ▶— — < I—-H>—I >—H>—O—( H—(
> II ( » <> I I l> ( I I > I
h—( •——I H--H»—« I—H I—I >—<>——< ▶— — I H——(
I——( h——< •——( I—H •—H »——(I——H ▶— <▶— — 4 1 1
I—H I—I »—<>—I ▶— — « » — — I H—(h—I K—( f—(
Figure. The situation at time t =

Figure. i4n example with M = 0.1. The two particles have evolved
AT = R2, where the measure m in the phase space N. Each point
is located on 2 points. The Vlasov moves as "test particle" in the
evolution describes a deformation
force field of the 2 particles. Even
of the plane. The situation is so the 2 body problem is inte
shown at time t = 0. The coor
grable, its periodic motion acts like
dinates (x, y) describe the position
a "mixer" for the complicated evo
and the speed of the particles.
lution of the test particles.
Example. Let M = N = R2 and assume that the measure m has its support
on a smooth closed curve C. The process Xf is again a volume-preserving
deformation of the plane. It describes the evolution of a continuum of par
ticles on the curve. Dynamically, it can for example describe the evolution
of a curve where each part of the curve interacts with each other part. The
picture sequence below shows the evolution of a particle gas with support
on a closed curve in phase space. The interaction potential is V(x) = e~x.
Because the curve at time t is the image of the diffeomorphism X*, it
will never have self intersections. The curvature of the curve is expected
to grow exponentially at many points. The deformation transformation
X1 = (fl,gl) satisfies the differential equation
d t
dtf = 9
o-(f(")-f{r)))
dt9 dm(n)
Jm
If r(s), 5 G [0,1] is the parameterization of the curve C so that m(r[a, b\) =

(b — a), then the equations are
ifi*) = fix)
±g*ix) = J* e-W-fW))) & .
The evolved curve C at time t is parameterized by s -» (/'(r(s)),5*(r(s))).
O 0
Figure. Tfte support Figure. The support Figure. The support
of the measure P° on of the measure P0A on of the measure P12 on
N = R2. N = R2. 7V = R2.
Example. If X1 is a stochastic process on (ft = M, A, m) with takes values

in JV, then P* is a probability measure on N defined by P*[A] = m(X~1A).
It is called the push-forward measure or law of the random vector X. The
measure P* is a measure in the phase space N. The Vlasov evolution defines
a family of probability spaces (N,B,Pf). The spatial particle density p is
the law of the random variable x(x, y) = x.
Example. Assume the measure P° is located on a curve f(s) = (s,sin(s))

and assume that there is no particle interaction at all: V = 0. Then P* is
supported on a curve (s + sin(s),sin(s)). While the spatial particle density
has initially a smooth density y/1 + cos(s)2, it becomes discontinuous after
some time.
5.4. Vlasov dynamics 301
t---0 ■>:-■■'. t=2 t=3
Figure. Already for the free evo

lution of particles on a curve in
phase space, the spatial particle
density can become non-smooth
after some time.
Example. In the case of the quadratic potential V(x) = x2/2 assume ra has
a density p(x, y) = e~x2-2y2, then P* has the density p\x, y) = f(x cos(t) +
y sin(t), -xsin(t) + ycos(t)). To get from this density in the phase space,
the spatial density of particles, we have to do integrate y out and do a
conditional expectation.
Lemma 5.4.1. (Maxwell) If X1 = (/*, fl*) is a solution of the Vlasov Hamil

tonian flow, then the law P* = (X*)*ra satisfies the Vlasov equation
P*(s, V) + V • V.P'foy) - W(x) • VyP*(x,») = 0
with W(x) = fM VxV(x - x') • P'(x', y')) dy'dx'.
Proof. We have / W(/(u>) - f(n)) dm(n) = W(/M). Given a smooth

function h on N oi compact support, we calculate
L= h(x,y)jtP\x,y)dxdy
as follows:
302 Chapter & * Selected Tbpics
L = dt] h(x>y)pt(x>y) dxdy
= Jtd J h(f(v,t),giu>,t))dmiw)
f ' .'.':...
= / Vxhif(v,t),giu,t))giu,t)dmiu)
JM
- /-■.Vy7»(/.(w,t),0(w,t)) / VVifiu) - fir,)) dmirj) 4n(w)
= / ^xhix,y)yPtix,y)dxdy- f P*(x,»)VyM».y)
J/NW(x - x')P'(x', j/') dx'dy'dxdy

= - J hi^,y)^xPtix,y)ydxdy
Jn
+ I h i x , y ) W i x ) - Vy P t i x , y ) d x d y. - , .
./AT
n
Remark. The Vlasov equation is an example of an integro-differential equa-
tion The right hand side is an integral. In a short hand notation, the Vlasov
equation is
■' ■ '■ - ■■'■•--..■■■' ■' ::''«■■ P+^-^^^(x)-P^0; , i .]rS.4C 5;j,^*fc]
whereW ^VXV'*P is'the cdhvblution of the force1 V^ with P: ;
Example. V(x) = 0. Particles move freely. The Vlasov equation becomes

the transport equation P(^rf^)^.y-9:plPf\^y) =*01vhJbh is inItoedi
mensions a partial differential equation ut + yux = 0. It has solutions
u(t, x, y)- = u(% x + fy).Restricting this function to y = x gives the Burg
ers equation ut + x^ = 0.
Example. For a quadratic potential V(x) = x2, the Hamilton equations are
/M = ~(/M - / f(v) dm(n)) .

Jm
In center-of-mass-coordinates / = /-£[/], the system is a decoupled
system pf a^qntinuum o£oscill#tprs / ==;g, g ==~f with solutions; «; \
/(*) = /(o)cos(o ^^(oykmfij; -^(*yi±dy '^/(ayafax^'-H-^co^cosC^y -'f/: ?';
The evolution for the density P is the partial differential equation
written in short hand as ut+y-ux-X'Uy = 0, which has the explicit solution

P*(x, y) = P°(cos(*)x + sin(t)y, - sin(*)x + cos(t)y). It is an example of a
Hamilton-Jacobi equation.
Example. .On any Riemanniap manifold w^

A, there êâteal potentials; ^ fsplyed, t^y
(j) = V * /*, where • is the convolution^ This; defines Newton potentials on
the manifold. Here are some examples: .
• jv == R: Vix) = M
2-
• N--= T: Vix) = |x(27T-x)|
4x
• N--= S2 ■Yix) == 10g(l,-X-x). ; !;:|
• N-.= E2 Vfal*
• N-.=M3Jix) = 47Tx|| ■'
• N = R4 V(x) =_ 1 1
For example, for ^ - R, the Laplaciati A/ =±> /'Ms the second deriva
tive. It is diagonal in Fourier space: A/(fc) = -fc2/, wheifefc €fe FVom
Deltaf{k) = -fc2/ = p(k) we get /(fc) = -(l/^îfc^.sp^that / = V*p,
where V" is the function which has the Fourier transform V'(fc) = -1/fc2.
B^tV^x) =; |xj/2 has this Fourier transform; I .:.■.■■);'•• ^ ; ;^ !
00 ItI 1
/ 2C , ^: ... fc2
Also for JV = T, the Laplacian A/ '== /" is diagonal in Fourier space. It

is. the gTr-periodic function y(x)^x(27r - x)/(47r)j, which has the Fourier
series,V(fc) == - 1 / f c P. ' >
For generalM;=? Rn, see for example [58] ; ; ;: ^
Remark. The function Gy(x) =V(x - y) i$ also called; the Gr^:f^nc^<m

of the Laplacian. Because Newton potentials V are not smooth, establishing
global existence for the Vlasov dynamics is npt easy but it has bqen done
in many cases [30]. The potential |x| models galaxy motion and appears in
plasma dynamics [90, 65, 82],
Lemma 5.4.2. (Gronwall) If a function u satisfies uf(t) < \g(t)\u(t) for all
0 < t < T, then u(t) < tz(0) exp(/0*'\g(s)\ ..factor 0 <t<T.
Proof. Integrating the assumption gives u(t) < u(0) + JQ g(s)u(s) ds. The
function h(t) satisfying the differential equ&tirin ''&($) == \ff(t)\u(t) sîsffes
Jhr(i) < \g(tyh(t). This leads to A(t>."</?i<-0>^^^jp^^)!* 0'^'^ Hty -
u(0) exp(.J* \g(s)\ ds). This proof for real value<l functions [20] generalizes
to the cafe, where ul (x)( evolves in a function spacevQne just can apply the
same proof for any fi x e d x. D
Theorem 5.4.3 (Batt-Neunzert-Brown-Hepp-Dobrushin). If VXV is

bounded and globally Lipshitz continuous, then the Hamiltonian Vlasov
flow has a unique global solution Xt and consequently, the Vlasov equa
tion has a unique and global solution P* in the space of measures. If V and
P° are smooth, then P* is piecewise smooth.
Proof. The Hamiltonian differential equation for X = (/, g) evolves on on

the complete metric space of all continuous maps from M to N. The dis
tance is d(X, Y) = supw€Af d(X(u), Y(u>)), where d is the distance in N.
We have to show that the differential equation / = g and g = G(f) =

- fM VxV(f(u) - f(n)) dm(n) in C(M, N) has a unique solution: because
of Lipshitz continuity
||G(/) ~ G(/;)||oo < 2||D(V,V)||oo • 11/ - /'Hoc

the standard Piccard existence theorem for differential equations assures
local existence of solutions.
The Gronwall's lemma assures that ||X(fj)|| can not grow faster than ex
p o n e n t i a l l y. T h i s g i v e s t h e g l o b a l e x i s t e n c e . Q
Remark. If m is a point measure supported on finitely many points, then
one could also invokl the global existence theorem for differential equations.
For smooth potentials, the dynamics depends continuously on the measure
m. One could approximate a smooth measure m by point measures.
Definition. The evolution of DX1 at a point uj € M is called the linearized
Vlasov flow. It is the differential equation
Df{«>) = - [ V2V(/M - f(n)) dm(rj)Df(uj) =: B(f)Df(u

Jm
and we can write it as a first order differential equation
—DX =
dt dt
0 1
./m-vM/M-/(^))^(^) o
A(n rf
Remark. The rank of the matrix DX\u) stays constant. Df\u) is a lin
ear combination of Df°(u) and Dg°(u). Critical points of /* can only
appear for u;, where Df°(uj), Df°(uj) are linearly dependent. More gen
erally Yk(t) = {uj e M | DX\uj) has rank 2q - fc = dim(N) - fc} is time
independent. The set Yq contains {uj \ D(f)(u) = \D(g)(u), A G RU{oc}}.
Definition. The random variable
\(uj) = limsup 7 logdl^Xô;))!!) € [0,oo]
is called the maximal Lyapunov exponent of the SL (2g,R)-cocycle At =

A(fl) along an orbit X1 = (/*, gl) of the Vlasov flow. The Lyapunov expo
nent could be infinite. Differentiation of Df = B(ft)ft at a critical point
u* gives D2}*(&*) = B(ft)D2ft(ujt). The eigenvalues Xj of the Hessian
D2f satisfy Xj = B(ft)Xj.
Definition. Time independent solutions of the Vlasov equation are called
equilibrium measures or stationary solutions.
Definition. One can construct some of them with a Maxwellian ansatz

2 r
P(x,y) = Cexp(-/3(|- + JV(x- x')Q(x') dx)) = S(y)Q(x) ,
The constant C is chosen such that /Rd S(y) dy = 1. These measures are
called Bernstein-Green-Kruskal (BGK) modes.
Proposition 5.4.4. If Q : N •-> R satisfies the integral equation
Q(x) = exp(- / (3V(x - x')Q(x')) dx1 = exp(-(3V * Q(x))

Jnd
then the Maxwellian distribution P(x,y) = S(y)Q(x) is an equilibrium

solution of the Vlasov equation to the potential V.
Proof.
yVxP = yS(y)Qx(x)
= yS(y)(-(3Q(x) [ VxV(x-x')Q(x')dx')
Jnd
and
/ VXV(x - x')VyP(x, y)P(x',y') dx'dy'

Jn
= Q(x)(-(3S(y)y) J VxV(x - x')Q(x') dx'
gives yVxP(x, y) = JN VxV(x - x')VyP(x, y)P(x',y') dx1 dy'. □

3P& Chapter 5. Selected Tropics
5.5 Multidimensional distributions
Random variables which are vector-valued can be treated in an analogous
way as random variables. One often adds the term to indi
cate that one has multiple dimensions.
P^flnitipn? A rwdpm vector jjs, a v^ctpr^valued' randojn y^ri^ble. It is in

£p tf each copr^lmate -is, in, JC?.TH expectati9ii1E[Z], of a random vector
%.=.-$f\i\' <\\x*) .« ^he^vector .(E^],, ,E[Zd]), the variance is the
v e c t o r ( y a r [ X i J > : : . , Va r, J f d ] ) r ' M \ ; ' ' ! : . ^
Example. The random vector ^ = (x3,j/4,25) on the unit cube ft =;JOrlE

with Lebesgue measure £ nas'the^^
Definition. Assume X = (Xi,...,Xd) is a random vector in C°°. The law

of %^rM^W?^#^^ 4s ft ineûrp^ pn,]Rd?witli^pnip^ct supppr^^ft^rj
some scaling and translation we can assume that p be a bounded Borel
measure on the unit cube Id = [0, l]d.
Definition. The multi-ênîqn^l distribution fimction pf a randomy^ctorj

X = (Xi,- • • ,Xd) is defined as^_. *ViJY'-; !>-,{■■-> -.•■»! ^--» -'■» -:..i-.l-(-.'u! ' *
Fx(t) = F{Xl^Xd)(tu..,td) = P[X! < tu...,Xd < td) .
For a continuous randoinvariable, there is ^ density falty satisjF^ng 0
-OO J— OO «;;
The multidimensional distribution function is also called multivariate dis-

t r i b u t i o n f u n c t i o n . . " , > , - < v . : - j " . / ■ ) , ] ■■ u ^ n u i . . > .
Definition. We use in this section the multi-index notation xn = H?=1 x"\

Denote by pn = fJd xn dp the n'th moment of p. If X is a random
vector, with law p, call pn(X) the n'th moment of X. It is equal to^
E[X»] = E[X^* • • • *»-]. We call the map n e Nd ~ pn the moment
configuration or, if d = 1, the moment sequence. We will tacitly assume
Pn = 0, if at least one coordinate nîn n = (m,..., nrf) is negative.
If X is a continuous random Vector, the moments satisfy
Mn(*)= / xUf(x)dx
Jud
which is a short hand notation for \
• • f ^ ■■■ ^ f \ x i , ^ x d ) d x 1 ^ - d x d .
-oo « / — ooo
o
5.5. Multidimensional distributions 307
Example. Then = (7,3,4) 'th moment of the random vector X = (x3,j/4f^)
is :\ ■■ - . ■ 1 - - 1 - - : 1 - - ■■ ' i; ' ■■
E[X^X^X^\ = E[x21yl2z*0} = -Y32Q •
The random vector X is continuous and has the pfobability density
T- 2 / 3 J . - 3 / 4 ^ - 4 / 5
f(x^z) = (^)(V-ir)(-r)-
Remark. As in one dimension, one can define a multidunensional moment
generating function li:l ;.-,; ;..-.:.•.,.;-i [^^^y^-i^nûl/.) .LO-.C xwiH'.tA
Mx(t) = E[el'x] = Effe'1*1^- • -e^f "'/{-l! : ' '
which contains^,!! the information about the moments because of the multi
dimensional moment formula
1 Jud » "^
where the n*th derivative is defined $s

d dni dn2 dnd
Example. The random variable X = (x, ^y, z1/3) has the moment gener
ating function
Af(M,«) = t f f esx+t^+uzU3 dxdydz

Jo Jo Jo j m i. *?
Because^componentsiXi?,X2,X$inthisExample;wereindependent ran
dom variables, the moment generating function is of the form
M(s)M(t)M(u) ,
where the factors are the one-dimensional moments of the one-dimensional
rahclohivariaibliesilX\, X$ aiid iX^l" *
Definition. Let a be the standard Ipaîs in £f. 1 JDfBê tê #artM #B^fwe
(Aia)n = an~ei - an on configurations and write Afe = ]li V- Unlike the
uûal convention, we take a particular sign convîtiop for A. This allows
ui to avoid many negative signs in tnis section. By induction in ]£i=in*>
one proves the relation
' ' (Afc/i)n = I ,xT^(l-^fcdM, / ^ : (5.1)
using xn-^-fe(l-x)fc-xn~fe(l-x)fe = xn-ei-k^^k+ei lb improve read

ability, we also use notation like £ = fllLi nt or ( fc ) = ^=1 ( fc- J
of rfViTf1 n- • -xr a .,we mdan n-> po m thfe sense that m -voo
F>(fcrtff?9> wt4f%?* «-)^**rfldo-i(i .mnwotu our ot no:um»* j/rpmii u «.»diy)
^ 3 a? 'fcxi oViiVm ,.v\ ;-'/rnrr;r/)iii >s -ioH .(&.£'} ifoi.Hbijo-> T»b;iP v\ '>iiJHi.v>isi j; 1o
Definition. Given a continuous function / : Id -+ R. For n € Nd, n* > 0 we
define the higher dimensional Bernstein polynomials
«.(flw-f/(| £>(; )*•<'-**-.

fc=0 u \ /
Lemma 5.5.1. (Multidimensional Bernstein) In the uniform topology in

C(Id), we have Bn(f) -+ / if n -> oo.
Proof. By the Weierstrass theorem, multi-dimensional polynomials are dense

in C(Id) as they separate points in C(Id). It is therefore enough to prove
the claim for f(x) = xm = ]lt=i XT- Because Bn(ym)(x) is the product of
one dimensional Bernstein polynomials
5„(j/m)(x) = nB„j(yr)(^),
i=l
the claim follows from the result corollary (2.6.2) in one dimensions. □
Remark. Hildebrandt and Schoenberg refer for the proof of lemma (5.5.1)
to Bernstein's proof in one dimension. While a higher dimensional adapta
tion of the probabilistic proof could be done involving a stochastic process
in Zd with drift Xi in the z'th direction, the factorization argument is more
elegant.
Theorem 5.5.2 (Hausdorff,Hildebrandt-Schoenberg). There is a bijection

between signed bounded Borel measures u on [0, l]d and configurations /xn
for which there exists a constant C such that
(Afc/x)n| < C, Vn e Nd . (5.2)

k=o v '
A configuration fin belongs to a positive measure if and only if additionally

to (5.2) one has (Afc/x)n > 0 for all fc,n E Nd.
Proof, (i) Because by lemma (5.5.1), polynomials are dense in C(Id), there
exists a unique solution to the moment problem. We show now existence
of a measure /z under condition (5.2). For a measures /x, define for n € Nd
5.5. Multidimensional distributions 309
the atomic measures ;u(n) on Id which have weights [\\ (&kt*)n on the
IllLifa + !) P°ints (^^T1' • • •' rbÎA>> € /d with ° - fci " "*' Because
J^^Hx) = £ (*) (^nAV)n

= / E (^) (^)m^n_fe(i - *)fc <*mo«o
^/d fc=0 ^ '
= j Bn(ym)(x) dfi(x) - / xm d/x(x) ,
we know that any signed measure \i which is an accumulation point of /x(n),

where n* -▶ oo solves the moment problem. The condition (5.2) implies that
the variation of the measures /i(n) is bounded. By Alaoglu's theorem, there
exists an accumulation point /i.
(ii) The left hand side of (5.2) is the variation ||/x(n)|| of the measure /i(n).
Because by (i) /x^ —▶ /x, and /j, has finite variation, there exists a constant
C such that ||/i(n)|| < C for all n. This establishes (5.2).
(iii) We see that if (Afc/x)n > 0 for all fc, then the measures /x(n) are all
positive and therefore also the measure /i.
(iv) If \x is a positive measure, then by (5.1)
Remark. Hildebrandt and Schoenberg noted in 1933, that this result gives
a constructive proof of the Riesz representation theorem stating that the
dual of C(Id) is the space of Borel measures M(Id).
Definition. Let S(x) denote the Dirac point measure located on x G Id. It
satisfies JId S(x) dy = x.
We extract from the proof of theorem (5.5.2) the construction:
Corollary 5.5.3. An explicit finite constructive approximations of a given

measure /i on Id is given for n G Nd by the atomic measures
**- £ ( « ) < AV W « ^ =^»-

0<fci<Tli V '
31° Chapter5. Selected To p i c s
Hausdorff established a criterion for absolutely continuity of a measure /z
with respect to the Lebesgue measure on [0,1] [72]. This can be generalized
to find a criterion for comparing two arbitrary measures and works in d
dimensions. ■
Definition. As usual, we call a measure p, on "if uniformly absolutely con

tinuous with respect to v, if it satisfies/i ^ f dv with / e L^(Id). \
Corollary 5.5.4. A positive probability measure p is uniformly absolutely

continuous with respect to a second probability measure v if and only if
there exists a constant C such that (Afc/i)n < C • (Akv)n for all fc, n € Nd.
Proof Up ^ fv wfth / € L°°(/d), we get using <5.1)
(Afc/i)n = [ x^-k(l-x)kdp(x) \if

J j d ■ • * : ■ « ■ ; ■ : ■■ ' ■■ >
- I Xn-k(l^x)kfdu(x)
Jld
< H/IU f ^^(l ^x)* du{x)
Jld
On the other hand, if (Akp)n <C(A*i/)n then pn = C(A*i/)n - (A*/i)n

defines by theorem (5.5.2) a positive measure p on Id. Since p = Qv- py
we have for any Borel set A C Id p(A) > 0. This gives p(A) < Cu(A) and
implies that p is absolutely continuous with respect to v with a function /
satisfying f(x) < C almost everywhere \ □
This leads to a higher dimensional generalization of Hausdorff's result
which allows to characterize the continuity of a multidimensional randona
vector from its moments: '' ;' l' —''-"■ <':- -' >:^^.ui'>hî\ >A~uaif.r4
Corollary 5-5.5. A BorsI probability meapure p on Id is unifbrn^;||jbp^

lutely continuous with respect to Lebesgue measure oii\Jd if and only if
Proof. Use corollary ^
There is albo a âracterizat^ mea&ureson f1 = [0,1]

for p > 2. This has an obvious generalization to d dim^M&ns:
5.6. Poisson processes 3 11
Proposition 5.5.6. Given a bounded positive probability measure p €

M(Id) and assume 1 < p < oo. Then p € Lp(Id) if and only if there
exists a constant C such that for all fc,n
.'.". - (n + ^-'fî^n^Jr^C. , L: .-(5,3).

fc=0
Proof: (i) Vet ^ Wthe measures of corollary (5:5.3). We' construct first
from the atomic measures pM1^ absolutely cbntihuous measures pW =
g^&x dii Id given by afuikition g whidh takes the Constant v&lue r
d
i\Akin)n\(nkyf[(ni + iy
on a cube of side lengths l/(n* + 1) centered at the point (n - k)/ne Id.
Because the cube has Lebesgue volume (n + 1)_1 = rL=i(n* + I)""1' ^ nas
the same measure with respect to both /x(n) and g^dx. We have therefore
also g^dx —»/x weakly.
(ii) Assume |u f /dx with / G Lp. Because g^dx■-^ /dx in the we&M
topology for measures, we have (^ —▶ / weakly in Lp/ But then, there
exists a constant C such that ||p^||P < C and this is equivalent to (5.3).
(iii) On the other hand, assumption (5.3) means that ||^(n)[|p < C, where
g(n) was constructed in (i). Since the unit-ball in the reflexive Banach space
Lp(Id) is weakly compact for p € (0,1), a subsequence of #(n) converges to
a function g E Lp. This implies that a subsequence of g^dx converges as
a measure to gdx which is in Lp and which is equal to p by the uniqueness
of the moment problem (Weierstrass). □
5.6 Poisson processes i

Definition. A Poisson process (5, P, II, N) over a probability space (ft, T, Q)
is given by a complete metric space 5, a non-atomic finite Borel measure
P on S and a function uj »-> II(uj) C S from Ct to the set of finite subsets of
S such that for evky measurable set B C 5, the map
v. " ; ; ; . : i . A \ ' ' ' . : ' " - ^ r ' ^ ^ ^ i ^ ^ p ^ ^ r - v : ! i v . : " " : ' .■ -
is,a IjOissojidistributed random variable with parameter P[jB],: For any

finite partition {^}^=1 of 5, the set of random variab^^^^^ havejtpi
be independent. The measure P is called the mean measure of the process.
H^rJeu|iAf-denotes the cardinality pf a finite set A. It is understood that
- ( " i ■ - ■ ! - -■ \
Example. We have encou|tered the one-dimensional Poisson process in

the last chapter as a martingale. We started with IID Poisson distributed
random variables Xk which are "waiting times" and defined Nt(uj) =
YlkLi lsk(u>)<t- Lets translate this into the current framework. The set S
is [0, t] with~Lebesgue measure P as mean measure. The set II(u;) is the
discrete point set II(u;) = {Sn(uj) \ n = 1,2,3,... } n S. For every Borel set
B in 5, we have
JV3(«) = t |n(w)| •
Remark. The Poisson process is an example of a point process, because

we can see it as assigning a random point set II(u;) on 5 which has density
P on S. If S is part of the Euclidean space and the mean measure P is
continuous P = /dx, then the interpretation is that f(x) is the average
density of points at x.
Figure. A Poisson process in R2 **r

with mean density } * jfer^
P = dxdy . . .' •; ^ —V»*

2tt • %>• •• • v. , "
Theorem 5.6.1 (Existence of Poisson processes). For every non-atomic mea

sure P on 5, there exists a Poisson process.
Proof. Define ft = US=o 5d' where Sd = Sx--xS isthe Cartesian product

and 5° = {0}. Let T be the Borel a-algebra on ft. The probability measure
Q restricted to Sd is the product measure (P x P x • • • x P) -Q[NS = d], where
Q[NS = d} = Q[Sd] = c-p[sl(d!)-1P[5]d. Define U(uj) = {u>u... ,ujd} if
u e Sd and NB as above. One readily checks that (5, P, II, N) is a Poisson
process on the probability space (ft, T,Q): For any measurable partition
{Bj}f=0 of 5, we have
™ d! fVP[5i]d
0[NBl = di,. • •, NBrn = dm I iVs = d0+2^ di = d\ = do\"'dm\ AI TWj
j=i j—^
5.6. Poisson processes 313
so that the independence of {NBj}^lx follows:
oo
Q[NBl=du...,NBm=dm} = £ QWs = d]Q[NBl

= di,...,NBm=dm\Ns = d]
" 12- dj. J|l d,!

d0=0 J=l
J=l 3
m
= ^Q^B^dj).
5=1
This calculation in the case m = 1, leaving away the last step shows that NB
is Poisson distributed with parameter P[B]. The last step in the calculation
is then j u s t i fi e d . ^
Remark. The random discrete measure P(uj)[B] = Nb(uj) is a normal

ized counting measure on S with support on U(uj). The expectation of
the random measure P(uj) is the measure P on S defined by P[B] =
fQ P(uj)[B] dQ(u). But this measure is just P:
Lemma 5.6.2. P = /n P(^) dQ(uj) = P.
Proof. Because the Poisson distributed random variable NB(u) = P(uj)[B]

has by assumption the Q-expectation P[B] = YlkLok Q[nb = fc] =
/ n P ( u j ) [ B ] d Q ( u ) o n e g e t s P = / n P ( u j ) d Q ( u ) = P. □
Remark. The existence of Poisson processes can also be established by
assigning to a basis {a } of the Hilbert space L2(S,P) some independent
Poisson-distributed random variables Zi = (j)(ei) and define then a map
(j)(f) = J2iai<f>(ei) if / = H2iaiei' The imaSe °f this m8LP iS a Hilbert
space of random variables with dot product Cov[^(/), </>(g)] = (/,#)• Define
NB = 4>(lB). These random variables have the correct distribution arid are
uncorrelated for disjoint sets Bj.
Definition. A point process is a map II a probability space (ft, F,Q) to

the set df finite subsets of a probability space (5, B, P) such that Nb(uj) :=
la; fl BI is a random variable for all measurable sets B € B.
' 3i4 Chapter 5. Selected To p i c s
Definition. Assume II is a point process oh (5,8,P). Fora function / :

S —▶ R+ in LX(S, P), define the random variable
z€U(u)
Example. For a Poisson process and / = 1b, one gets E/(u;) = Nb(uj).
Definition. The moment generating function of E/ is defined as for any
random variable as
■■ I T ;MEf(i)AE\e^f] .
It is called the characteristic functional of the point process.
Example. For a Poisson process and / = al5, the moment generating

function of E/(cj) = Nb(uj) is E^ff] = ep[BKi-ât). We have computed
the moment generating function of a Poisson distributed random variable
in t h e f fi r s t chapter. ; , i
Example. For a Poisson process aiid / = Efc=i ao *Bk, where B^ aire disjoint
sets, we have the characteristic functional
Example. For a Poisson process, and / G LX(S, P), the moment generating
function of E/ is
ME/(t) = exp(- /(l-exp(t/(z)))dP(2;)), tl ,, ly^,Tj

Js
This is called CampbellV theorem. The proof is done by writing / =
/+-/-, where both /+ and /" are nonnegative, then approximating
both functions with step functionsf£ ^= â^lB+, ;=^ ^/^ arid/^ =
"52 a'ljg- 53; fkj- Because for Poisson process,, the:random variables E^
are independent for different j or different sign, the-moment generating
function of E/ is the product of the moment generating functions E^± =
N f r ^ ■ ^ : . , : • ' ■■■" — - i ; v : - - ' » ■ ■•■:■■■•■■■' : •■■' [ / : ^ ? : d
Thfe next the6rerii of Alfred Renyr(19^1-1OT0) ghres ia haM^ tool to check
whether a point process, a randdm variable II with values in the set of
firiitW subsets of S, defined a PbiSsori process.
Definition. A fc-cube in an open subset 5 of Rd is is a set
nrn ( n j - fl ) v
2=1
5.6. Poisson processes 3l5
Theorem 5.6.3 (Renyi's theorem, 1967). Let P be a non-atomic probability

measure on (5, B) and let II be a point process on (ft, T, Q). Assume for any
finite union of fc-cubes B C 5, Q[NB = 0] = exp(-P[B]). Then (5, P; II, N)
is a Poisson process with mean measure P.
Proof. (i) Define 0(B) = {uj £ ft | Nb(uj) = 0 } C ft for any measurable

setBmS. By assumption, Q[0(B)} = exp(-P[J5]).
(ii) For m disjoint fc-cubes {Bj}f=lr the sets 0(Bj) C ft are independent.
Proof:
■ .. ■ . rn . . ■ ' . ' ■■■
QlC\0(Bj)) = Q[{NUT=lBj=o}}
■»; ■ ■'•• .-' •• :....■ i=i ■■ ■ -.,;..:■ . ■ ; .:i!
rn
= expc-ny^])
■ . i=l .■.■■■■■■:
m
= ]jeM-P[Bj\)
m * ■■ ' { - *■ : .■
(iii) We, count the number pf points in ah open open subset U of S using
^-cubes: define\ for k > 0 the random variable tiy{uj) as the number fc-
cubes B for which uj £ 0(Bhi/).' These random variable NJj(u) converge
to Nu(uj) for fc —> oo, for almost all uj.
(iv) For an open set [/, the random variable Nu is Poisson distributed
with parameter P[U]. Proof: we compute its moment generating function.
Because for different fc-cubes, the sets O(Bj) C 0(U) are independent,
the ttibmerit '"generating function of Nfr = ^fc l'o(£)j) is the product bf trie
moment generating functions of lo(B)j):
E[e^] = n (0|O(B)].W(l-i?[O(BJ]i :
fc—cube B
= J] (exp(-P[B]) + e'(l-exp(-P[S])))
fc—cube B
Each factor of this product is positive and the monotone convergence the
orem shows that the moment generating function of Nu is
E[ew"] =cf—>olim J]
o ■*■■»• (exp(-P[B])+e*(l-exp(-P[B]))).
fc—cube B
which converges to exp(P[C/](l - e*)) for fc -> oo if the measure P is non-
atomic.
Because the generating function determines the distribution of Nu, this

assures that the random variables Nu are Poisson distributed with param
eter P[U).
(v) For any disjoint open sets U\,..., C/m, the random variables {JV^ )}f=i
are independent. Proof: the random variables {N^.)}^ are independent
for large enough fc, because no fc-cube can be in more than one of the sets
Uj, The random variables {Nj}.)}^ are then independent for fixed fc. Let
ting fc —▶ oo shows that the variables Nuj are independent.
(vi) To extend (iv) and (v) from open sets to arbitrary Borel sets, one
can use the characterization of a Poisson process by its moment generating
function of / G Ll(S,P). If / = âj^ for disjoint open sets Uj and
real numbers aj, we have seen that the characteristic functional is the
characteristic functional of a Poisson process. For general / 6 L^S, P) the
characteristic functional is the one of a Poisson process by approximation
and the Lebesgue dominated convergence theorem. Use / = 1B to verify
that NB is Poisson distributed and / = Yl^Bj with disjoint Borel sets
Bj to see that {NBj)}fz=il are independent. □
5.7 Random maps

Definition. Let (ft, A, P) be a probability space and M be a manifold with
Borel cr-algebra B. A random diffeomorphism on M is a measurable map
from M x ft —▶ M so that x »-* f(x,u>) is a diffeomorphism for all uj e ft.
Given a V measure preserving transformation T on ft, it defines a cocycle
S(x,u) = (f(x,uj),T(uj))
which is a map on M x ft.
Example. If M is the circle and /(#, c) = x + csin(x) is a circle diffeomor

phism, we can iterate this map and assume, the parameter c is given by
IID random variables which change in each iteration. We can model this
by taking (ft,.4,P) = ([0, l]N,5N,i/N) where v is a measure on [0,1] and
take the shift T(xn) = xn+i and to define
S(x,uj) = (f(x,uJo),T(u>)).
Iterating this random logistic map is done by taking IID random variables
cn with law v and then iterate
xo,xx =/(x0,co),a?2 = /(xi,ci)... .

5.7. Random maps 317
Example. If (Sl,A,P,T) is an ergodic dynamical system, and A : ft -▶

5L(d, R) is measurable map with values in the special linear group SL(d, R)
of all d x d matrices with determinant 1. With M = Rd, the random
diffeomorphism f(x,v) = A(x)v is called a matrix cocycle. One often uses
the notation
An(x) = A(Tn~\x)) • A(Tn~2(x)) • • • A(T(x)) • A(x)
for the n'th iterate of this random map.
Example. If M is a finite set {1,.., n} and P = P{j is a Markov transition

matrix, a matrix with entries Pij > 0 and for which the sum of the column
elements is 1 in each column. A random map for which f(xi,uj) = Xj with
probability Pij is called a finite Markov chain.
Random diffeomorphisms are examples of Markov chains as covered in Sec
tion (3.14) of the chapter on discrete stochastic processes:
Lemma 5.7.1. a) Any random map defines transition probability functions

P:MxS-+[0,l]:
V(x,B)=F[f(x,uj)eB}.
b) If An is a filtration of a-algebras and Xn(uj) = Tn(uj) is An adapted,

then V is a discrete Markov process.
Proof, a) We have to check that for all x, the measure V(x, •) is a prob
ability measure on M. This is easily be done by checking all the axioms.
We further have to verify that for all B E S, the map x -> V(x,B) is
B-measurable. This is the case because / is a diffeomorphism and so con
tinuous and especially measurable,
b) is the definition of a discrete Markov process. □
Example. If ft = (AN, ^N, vN) and T(x) is the shift, then the random map
defines a discrete Markov process.
Definition. In case, we get IID A-valued random variables Xn = Tn(x)0.

A random map f(x,u) defines so a IID diffeomorphism-valued random
variables h(x)(uj) = f(x,X1(u)),f2(x) = f(x,X2(uj)). We will call a ran
dom diffeomorphism in this case an IID random diffeomorphism. If the
transition probability measures are continuous, then the random diffeomor
phism is called a continuous IID random diffeomorphism. If f(x, uj) depends
smoothly on uj and the transition probability measures are smooth, then
the random diffeomorphism is called a smooth IID random diffeomorphism.
It is important to note that "continuous" and "smooth" in this definition is
only with respect to the transition probabilities that A must have at least
dimension d > 1. With respect to M, we have already assumed smoothness
from the beginning.
Definition. A measure p on M is called a stationary measure for the random

diffeomorphism if the measure p x P is invariant under the map S.
Remark. If the random diffeomorphism defines a Markov process, the sta

tionary measure p is a stationary measure of the Markov process.
Example. If every diffeomorphism x -> f(x,uj) from v e ft preserves a

measure /x, then p is a automatically a stationary measure.
Example. Let M = T2 = R2/Z2 denote the two-dimensional torus. It is a

group with addition modulo 1 in each coordinate. Given an IID random
map:
f (x\ '= \{ x-r/3
JnK x + a with
with probability
probability 1/2
1/2 "
Each map either rotates the point by the vector a = (ai,a2) or by the
vector /? = (/?i,/?2). The Lebesgue measure on T2 is invariant because
it is invariant for each of the two transformations. If a and /? are both
rational vectors, then there are infinitely many ergodic invariant measures.
For example, if a = (3/7,2/7),/? = (1/11,5/11) then the 77 rectangles
[ill, (i + l)/7] x jj/ll, (j + i)/n] are permuted by both transformations.
Definition. A stationary measure p of a random diffeomorphism is called
ergodic, if p x P is an ergodic invariant measure for the map S on (M x
ft,/xxP). v
Remark. If p is a stationary invariant measure, one has
p(A)= f P(x,A)dp
Jm
for every Borel set A e A. We have earlier written this as a fixed point
equation for the Markov operator V acting on measures: Vp = p. In the
context of random maps, the Markov operator is also called a transfer
operator.
Remark. Ergodicity especially means that the transformation T on the

"base probability space" (ft,^,P) is ergodic.
Definition. The support of a measure p is the complement of the open set

of points x for which there is a neighborhood U with p(U) = 0. It is by
definition a closed set.
The previous example 2) shows that there can be infinitely many ergodic in
variant measures of a random diffeomorphism. But for smooth IID random
diffeomorphisms, one has only finitely many, if the manifold is compact:
5.8. Circular random variables 319
Theorem 5.7.2 (Finitely many ergodic stationary measures (Doob)). If M

is compact, a smooth IID random diffeomorphism has finitely many ergodic
stationary measures pi. Their supports are mutually disjoint and separated
by open sets.
Proof, (i) Let p\ and p2 be two ergodic invariant measures. Denote by Si

and E2 their support. Assume Ei and E2 are not disjoint. Then there ex
ist points Xi G Ei and open sets Ui of Xi so that the transition probability
P(x\,U2) is positive. This uses the assumption that the transition probabil
ities have smooth densities. But then p2(U x ft) = 0 and p2(S(U x ft)) > 0
violating the measure preserving property of S.
(ii) Assume there are infinitely many ergodic invariant measures, there
exist at least countably many. We can enumerate them as p\, p2,... Denote
by Ei their supports. Choose a point yi in E*. The sequence of points
has an accumulation point y e M by compactness of M. This implies
that an arbitrary e-neighborhood U of y intersects with infinitely many E*.
Again, the smoothness assumption of the transition probabilities P(y, •)
contradicts with the S invariance of the measures pi having supports E*.
□
Remark. If p\, p2 are stationary probability measures, then \p\ + (1 — \)p2

is an other stationary probability measure. This theorem implies that the
set of stationary probability measures forms a closed convex simplex with
finitely many corners. It is an example of a Choquet simplex.
5.8 Circular random variables

Definition. A measurable function from a probability space (Sl,A,P) to
the circle (T, B) with Borel a-algebra B is is called a circle-valued random
variable. It is an example of a directional random variable. We can realize
the circle as T = [-tt, tt) or T = [0,2n) = R/(2ttZ).
Example. If (Sl,A,P) = (R,A,e-x^2/V27rdx, then X(x) = x mod 2n is a

circle-valued random variable. In general, for any real-valued random vari
able Y, the random variable X(x) = X mod 27r is a circle-valued random
variable.
Example. For a positive integer fc, the first significant digit is X(k) =
27rlog10(fc) mod 1. It is a circle-valued random variable on every finite
probability space (ft = {1,..., n }, A, P[{fc}] = 1/n).
Example. A dice takes values in 0,1,2,3,4,5 (count 6 = 0). We roll it two

times, but instead of adding up the results X and Y, we add them up
modulo 6. For example, if X = 4 and Y = 3, then X + Y = 1. Note that
E[X + Y] = E[X] ^ E[X] + E[F]. Even if X is an unfair dice and if Y is
fair, then X + y is a fair dice.
Definition. The law of a circular random variable X is the push-forward

measure p = X*P on the circle T. If the law is absolutely continuous, it
has a probability density function fx on the circle and p = fx(x)dx. As
on the real line the Lebesgue decomposition theorem (2.12.2) assures that
every measure on the circle can be decomposed p = ppp + pac + Use, where
ppp is (pp), psc is (sc) and pac is (ac).
Example. The law of the wrapped normal distribution in the first example
is a measure on the circle with a smooth density
oo
fxix)= £ e-(*+2*fe>2/2A/2^-
k=—oo
It is an example of a wrapped normal distribution.
Example. The law of the first significant digit random variable Xn(k) =
27rlog10(fc) mod 1 defined on {1,... ,n } is a discrete measure, supported
on {fc27r/10|0 < fc < 10 }. It is an example of a lattice distribution.
Definition. The entropy of a circle-valued random variable X with prob

ability density function fx is defined as H(f) = - J^ f(x)\og(f(x)) dx.
The relative entropy for two densities is defined as
H(f\g)= / f(x)log(f(x)/g(x))dx.
Jo
The Gibbs inequality lemma (2.15.1) assures that H(f\g) > 0 and that
H(f\g) = 0, if / = g almost everywhere.
Definition. The mean direction m and resultant length p of a circular

random variable taking values in {\z\ = 1} C C are defined as
peim = E[eiX] .
One can write p = E[cos(X - ra)]. The circular variance is defined as

V = 1 - p = E[l - cos(X - ra)] = E[(X - ra)2/2 - (X - ra)4/4!...].
The later expansion shows the relation with the variance in the case of
real-valued random variables. The circular variance is a number in [0,1]. If
p = 0, there is no distinguished mean direction. We define ra = 0 just to
have one in that case.
Example. If the distribution of X is located a single point x0, then p =

l,ra = rr0 and V = 0. If the distribution of X is the uniform distribution
on the circle, then p = 0, V = 1. There is no particular mean direction in
this case. For the wrapped normal distribution ra = 0,p = e~G '2,V =
l-e"'2/2.
The following lemma is analogous to theorem (2.5.5):
Theorem 5.8.1 (Chebychev inequality on the circle). If X is a circular

random variable with circular mean ra and variance V, then
P[|sm((X-ro)/2)|>e]<^.
Proof. We can assume without loss of generality that ra = 0, otherwise

replace X with X - ra which does not change the variance. We take T =
[-7r,7r). We use the trigonometric identity 1 - cos(z) = 2sin2(x/2), to get
= E[l-cos(X)] = 2E[sin2(|)]
> 2E[l|sin(f)|>£sin(y)]
> 2e2P[|sin(f)|>e].
Example. Let X be the random variable which has a discrete distribution

with a law supported on the two points x = x0 = 0 and x = x± =
±2arcsin(€) and P[X = x0] = 1 - V/(2e2) and P[X = x±] = V/(4e2). This
distribution has the circular mean ra and the variance V. The equality
P[| sin(X/2)| > c] = 2V/(4e2) = V/(2e2) .

shows that the Chebychev inequality on the circle is "sharp": one can not ,
improve it without further assumptions on the distribution.
Definition. A sequence of circle-valued random variables Xn converges
weakly to a circle-valued random variable X if the law of Xn converges
weakly to the law of X. As with real valued random variables weak con
vergence is also called convergence by law.
Example. The sequence Xn of significant digit random variables Xn con
verges weakly to a random variable with lattice distribution P[X = fc] =
logio(fc + 1) - logio(fc) supported on {fc27r/10 | 0 < fc < 10 }. It is called
the distribution of the first significant digit. The interpretation is that if
you take a large random number, then the probability that the first digit
is 1 is log(2), the probability that the first digit is 6 is log(7/6). The law is
also called Benford's law.
Definition. The characteristic function of a circle-valued random variable
X is the Fourier transform <$>x = 0 of the law of X. It is a sequence (that
is a function on Z) given by
<Px(n) = E[einX] = j einx dvx(x)

Jj
Definition. More generally, the characteristic function of a Trf-valued ran
dom variable (circle-valued random vector) is the Fourier transform of the
law of X. It is a function on Zd given by
<t>x(n) = E[ein-X] = [ ein'x dvx(x) .

Jrd
The following lemma is analog to corollary (2.17).
Lemma 5.8.2. A sequence Xn of circle-valued random variables converges

in law to a circle-valued random variable X if and only if for every integer
fc, one has <t>xn(k) -> <£x(fc) for n -» oo.
Example. A circle valued random variable with probability density function

f(x) = CeKCO<x~^ is called the Mises distribution. It is also called the
circular normal distribution. The constant C is l/(27r/o (*;)), where I0(k) =
E^=o(K/2)2n/(n!2) a modified Bessel function. The parameter k is called
the concentration parameter, the parameter a is called the mean direction.
For k -» 0, the Mises distribution approaches the uniform distribution on
the circle.
0 12 3 4 5 6
Figure. The density function of Figure. The density function of

the Mises distribution on [—7r, n]. the Mises distribution plotted as a
polar graph.
Proposition 5.8.3. The Mises distribution maximizes the entropy among all
circular distributions with fixed mean a and circular variance V.
Proof. If g is the density of the Mises distribution, then log(g) = /ccos(x -

a) + log(C) and H(g) = Kp + 2nlog(C).
Now compute the relative entropy
0 > H(f\g) = j f(x) log(f(x))dx - J f{x) \og(g(x))dx .
This means with the resultant length p of f and g:
H(f) > -E[kcos(x - a) + log(C)] = -«p + 2?r log(C> - H(g) .
Definition. A circle-valued random variable with probability density func

tion
V27r<T2 fc= — OO
is the wrapped normal distribution. It is obtained by taking the normal

distribution and wrapping it around the circle: if X is a normal distribu
tion with mean a and variance cr2, then X mod 1 is the wrapped normal
distribution with those parameters.
Example. A circle-valued random variable with constant density is called

a random variable with the uniform distribution.
Example. A circle-valued random variable with values in a closed finite

subgroup H of the circle is called a lattice distribution. For example, the
random variable which takes the value 0 with probability 1/2, the value
27r/3 with probability 1/4 and the value 47r/3 with probability 1/4 is an
example of a lattice distribution. The group H is the finite cyclic group Z3.
Remark. Why do we bother with new terminology and not just look at real-
valued random variables taking values in [0,27r)? The reason to change the
language is that there is a natural addition of angles given by rotations.
Also, any modeling by vector-valued random variables is kind of arbitrary.
An advantage is also that the characteristic function is now a sequence and
no more a function.
Distribution Parameter characteristic function

point Xo <t>x(k) = e%**°
uniform 4>x(k) = 0 for fc ^ 0 and <f>x(0) = 1
Mises k, a = 0 4W//oW
wrapped normal cr, a = 0 e-fcV/2 = ^
The functions Ik(n) are modified Bessel functions of the first kind of fc'th
order.
Definition. If X\,X2,... is a sequence of circle-valued random variables,
define 5n = Xi + ..- + Xn.
Theorem 5.8.4 (Central limit theorem for circle-valued random variable).

The sum Sn of IID-valued circle-valued random variables Xi which do
not have a lattice distribution converges in distribution to the uniform
distribution.
Proof. We have \<j)x(k)\ < 1 for all fc ^ 0 because if </>x(k) = 1 for some
fc ^ 0, then X has a lattice distribution. Because </>sn(k) = n*=i fe(fc),
all Fourier coefficients (/)Sn (fc) converge to 0 for n -> oo for fc ^ 0. * □
Remark. The IID property can be weakened. The Fourier coefficients
<t>xn (fc) = 1 - ank

should have the property that Y£=\ ank diverges, for all fc, because then,
n^Li(! ~ ank) -▶ 0. If Xi converges in law to a lattice distribution, then
there is a subsequence, for which the central limit theorem does not hold.
Remark. Every Fourier mode goes to zero exponentially. If <f>x(k) < 1 - S
for S > 0 and all fc ^ 0, then the convergence in the central limit theorem
is exponentially fast.
Remark. Naturally, the usual central limit theorem still applies if one con
siders a circle-valued random variable as a random variable taking values in
[-7T, 7r] Because the classical central limit theorem shows that £?=1 Xn/y/n
converges weakly to a normal distribution,.^=1 Xn/y/n mod 1 converges
to the wrapped normal distribution. Note that such a restatement of the
central limit theorem is not natural in the gontext of circular random vari
ables because it assumes the circle to be embedded in a particular way in
the real line and also because the operation of dividing by n is not natural
on the circle. It uses the field structure of the cover R.
Example. Circle-valued random variables appear as magnetic fields in math

ematical physics. Assume the plane is partitioned into squares \j,j + 1) x
[fc, fc+1) called plaquettes. We can attach IID random variables Bjk = eiX*k
on each plaquette. The total magnetic field in a region G is the product of
all the magnetic fields Bjk in the region:
TT Bjk = e^0,k€GX3k e
(J,*)€G
The central limit theorem assures that the total magnetic field distribution
in a large region is close to a uniform distribution.
Example. Consider standard Brownian motion Bt on the real line and its
graph of {(t, Bt) \ t e R } in the plane. The circle-valued random variables
Xn = Bn mod 1 gives the distance of the graph at time t = n to the
next lattice point below the graph. The distribution of Xn is the wrapped
normal distribution with parameter ra = 0 and a = n.
w^
Figure. The graph of one-
dimensional Brownian motion
with a grid. The stochastic pro
cess produces a circle-valued ran
dom variable Xn = Bn mod 1. I ^^Pl—.—,—h -
-—j j
1 1 L L....I. i -
If X, y are real-valued IID random variables, then X+Y is not independent
of X. Indeed X + Y and Y are positively correlated because
Cov[x + y,y) = Cov[x, y] + Cov[y, y] = Cov[y, y] = var[y] > o.
The situation changes for circle-valued random variables. The sum of two
independent random variables can be independent to the first random vari
able. Adding a random variable with uniform distribution immediately ren
ders the sum uniform:
Theorem 5.8.5 (Stability of the uniform distribution). If X, Y are circle-

valued random variables. Assume that Y has the uniform distribution and
that X, y are independent, then X + Y is independent of X and has the
uniform distribution.
Proof. We have to show that the event A = {X + Y 6 [c, d] } is indepen

dent of the event B = {X e [a, 6] }. To do so we calculate P[A D B] =
jb jd-x yx(x)/y(y) dydx. Because Y has the uniform distribution, we get
after a substitution u — y - x,
b rd
f I * Jfx(x)fv(y)dydx=
Ja c—x J Q>
f f fx(x)fy(u)
Jc
dudx = P[A]P[B] .
By looking at the characteristic function <£x+y = <J>x<t>Y = <\>x, we see that

X + Y has the uniform distribution. d
The interpretation of this lemma is that adding a uniform random noise to
a given uniform distribution makes it uniform.
On the n-dimensional torus Td, the uniform distribution plays the role of
the normal distribution as the following central limit theorem shows:
Theorem 5.8.6 (Central limit theorem for circular random vectors). The
sum Sn of IID-valued circle-valued random vectors X converges in distri
bution to the uniform distribution on a closed subgroup H of G.
Proof. Again \</>x(k)\ < 1. Let A denote the set of fc such that (j>x(k) = 1.
(i) A is a lattice. If / eikx& dx = 1 then X(x)k = 1 for all x. If A, A2 are

in A, then Ai + A2 e A.
(ii) The random variable takes values in a group H which is the dual group
ofZd/H. B y
(iii) Because </>Sn(k) = IE=i **(*), all Fourier coefficients </>sn(k) which
are not 1 converge to 0.
(iv) <As„(fc) -» U, which is the characteristic function of the uniform dis

tribution on H. □
Example. If G = T2 and A = {..., (-1,0), (1,0), (2,0),... }, then the ran

dom variable X takes values in H = {(0,2/) | y e T1 }, a one dimensional
circle and there is no smaller subgroup. The limiting distribution is the
uniform distribution on that circle.
Remark. If X is a random variable with an absolutely continuous distribu

tion on Td, then the distribution of Sn converges to the uniform distribution
onTd.
Exercice. Let Y be a real-valued random variable which has standard

normal distribution. Then X(x) = Y(x) mod 1 is a circle-valued ran
dom variable. If Yi are IID normal distributed random variables, then
Sn = Yi + •-• + Yn mod 1 are circle-valued random variable. What is
Cov[Sn,Sm]?
The central limit theorem applies to all compact Abelian groups. Here is
the setup:
5.9. Lattice points near Brownian paths 327
Definition. A topological group G is a group with a topology so that addi
tion on this group is a continuous map from GxG —▶ G and such that the
inverse x -▶ rr-1 from G to G is continuous. If the group acts transitively
as transformations on a space B, the space H is called a homogeneous
space. In this case, H can be identified with G/Gx, where Gx is the isotopy
subgroup of G consisting of all elements which fix a point x.
Example. Any finite group G with the discrete topology d(x, y) = liix^y
and d(x, y) = 0 if x = y is a topological group.
Example. The real line R with addition or more generally, the Euclidean
space Rd with addition are topological groups when the usual Euclidean
distance is the topology.
Example. The circle T with addition or more generally, the torus Td with
addition is a topological group with addition. It is an example of a compact
Abelian topological group.
Example. The general linear group G = GZ(n,R) with matrix multiplica

tion is a topological group if the topology is the topology inherited as a sub
set of the Euclidean space Rn2 of n x n matrices. Also subgroup of Gl(n, R),
like the special linear group 5L(n,R) of matrices with determinant 1 or
the rotation group SO(n, R) of orthogonal matrices are topological groups.
The rotation group has the sphere Sn as a homogeneous space.
Definition. A measurable function from a probability space (ft,^4,P) to

a topological group (G, B) with Borel cr-algebra B is is called a G-valued
random variable.
Definition. The law of a spherical random variable X is the push-forward

measure p = X*P on G.
Example. If (G, A, P) is a the probability space by taking a compact topo

logical group G with a group invariant distance d, a Borel cr-algebra A and
the Haar measure P, then X(x) = x is a group valued random variable.
The law of X is called the uniform distribution on G.
Definition. A measurable function from a probability space (ft, A, P) to the

group (G, B) is called a G-valued random variable. A measurable function
to a homogeneous space is called B-valued random variable. Especially,
if H is the d-dimensional sphere (Sd, B) with Borel probability measure,
then X is called a spherical random variable. It is used to describe spherical
data.
5.9 Lattice points near Brownian paths

The following law of large numbers deals with sums Sn of n random vari
ables, where the law of random variables depends on n.
Theorem 5.9.1 (Law of large numbers for random variables with shrinking
support). IfXi are IID random variables with uniform distribution on [0,1].
Then for any 0 < S < 1, and An = [0,1/n6], we have
fc=i
in probability. For S < 1/2, we have almost everywhere convergence.
Proof. For fixed n, the random variables Zk(x) = l[0}1/nS](Xk) are indepen
dent, identically distributed random variables with mean E[Zk] =p = l/ns
and variance p(l -p). The sum Sn = ££=1 Xk has a binomial distribution
with mean np = nl~8 and variance Var[5n] = np(l - p) = n1_<5(l - p).
Note that if n changes, then the random variables in the sum Sn change
too, so that we can not invoke the law of large numbers directly. But the
tools for the proof of the law of large numbers still work.
For fixed e > 0 and n, the set
Bn = {x€[0,l]||^-l|>e}
n
has by the Chebychev inequality (2.5.5), the measure
PfB 1 < Varf-^-1/f2 - Var^n] - l~p < l
This proves convergence in probability and the weak law version for all
8 < 1 follows.
In order to apply the Borel-Cantelli lemma (2.2.2), we need to take a sub

sequence so that YlkLi P[Bnk] converges. Like this, we establish complete
convergence which implies almost everywhere convergence.
Take k = 2 with k(1 - S) > 1 and define nk = fc* = fc2. The event B =
limsupfcBnfc has measure zero. This is the event that we are in infinitely
many of the sets Bnk. Consequently, for large enough fc, we are in none of
the sets Bnk: if x e B, then
\Snk(x) ,,
i x ■*■ ^
for large enough fc. Therefore,
■ &»+!(») ,.<■&*(») ,, . SijT^jx))

Because for nk = fc2 we have nk+\ -nk = 2k + l and
Si(T£(x)) < 2fc + l

n1 - 5 - f c 2 ( l - < 5 )
For 8 < 1/2, this goes to zero assuring that we have not only convergence
of the sum along a subsequence Snk but for Sn (compare lemma (2.11.2)).
We know now | ^^ — 1| —▶ 0 almost everywhere for n -» oo. □
Remark. If we sum up independent random variables Zk = n*l[O,i/n*]P0c)

where Xk are IID random variables, the moments E[Z£*] = n^™"1^ be
come infinite for ra > 2. The laws of large numbers do not apply be
cause E[Z%] depends on n and diverges for n -+ oo. We also change the
random variables, when taking larger sums. For example, the assumption
supn £ Y!i=i Var[Xi] < oo does not apply.
Remark. We could not conclude the proof in the same way as in theo
rem (2.9.3) because Un = £Li Zk is not monotonically increasing. For
8 e [1/2,1) we have only proven a weak law of large numbers. It seems
however that a strong law should work for all 8 < 1.
Here is an application of this theorem in random geometry.
Corollary 5.9.2. Assume we place randomly n discs of radius r = 1/n1/2-5/2

onto the plane. Their total area without overlap is 7rnr2 = im6. If Sn is the
number of lattice points hit by the discs, then for 8 < 1/2
7T ,
n°
almost surely.
—
1
— m$ 1 —P—r~*
-j ■* ii
Figure. Throwing randomly
discs onto the plane and count
~1—i
4
^j *# »
ing the number of lattice points m <p^
T .
— i
which are hit. The size of the
discs depends on the number of
ti? ■P i
^ r
i r ——
discs on the plane. If 8 = 1/3 m
1»
and if n = l'OOO'OOO, then we 1! j ., m* •
have discs of radius 1/10000 f
and we expect Snj the number of fl 1 i
0
lattice point hits, to be 1007T.
P i%
§ j.
!
j i
Remark. Similarly as with the Buffon needle problem mentioned in the in
troduction, we can get a limit. But unlike the Buffon needle problem, where
we keep the setup the same, independent of the number of experiments. We
adapt the experiment depending on the number of tries. If we make a large
number of experiments, we take a small radius of the disk. The case 8 = 0
is the trivial case, where the radius of the disc stays the same.
The proof of theorem (5.9.1) shows that the assumption of independence

can be weakened. It is enough to have asymptotically exponentially decor-
related random variables.
Definition. A measure preserving transformation T of [0,1] has decay of

correlations for a random variable X satisfying E[X] = 0, if
Cov[X,X(Tn)]-+0
for n —> oo. If
Cov[X,X(Tn)]<e-Cn
for some constant C > 0, then X has exponential decay of correlations.
Lemma 5.9.3. If Bt is standard Brownian motion. Then the random vari

ables Xn = Bn mod 1 have exponential decay of correlations.
Proof. Bn has the standard normal distribution with mean 0 and standard
deviation o = n. The random variable Xn is a circle-valued random variable
with wrapped normal distribution with parameter a = n. Its characteris
tic function is <j>x(k) = e^*2!2. We have Xn+m = Xn + Ym mod 1,
where Xn and Yp are independent circle-valued random variables. Let
9n = ££L0 e"fc2n /2 cos(fcx) = 1 - e(x) > 1 - e~Cn2 be the density of Xn
which is also the density of Yn. We want to know the correlation between
Xn+m and Xn:
ri rl
/ /Jof(x)f(x + y)g(x)g(y) dy dx.
/o
Jo Jo
With u = x + y, this is equal to
/ /Jof(x)f{u>)9(x)9(u - x) dudx
Jo
= [ [ f(x)f(u)(l-e(x))(l-e(u-x))dudx
Jo Jo
< Cr\f\le-Cn2.
a
Proposition 5.9.4. If T : [0,1] —▶ [0,1] is a measure-preserving transfor

mation which has exponential decay of correlations for Xj. Then for any
S € [0,1/2), and An = [0, l/ns], we have
llm ^l>n(^(*))-l.
n—>oo n ,i-—'
fc=l
Proof. The same proof works. The decorrelation assumption implies that
there exists a constant C such that
J2 Cov[Xi,Xj] <C .
Therefore,
Var[5n] = nVar[X„] + £ Cov^.X,] < Ci|/£, £ B-C(*-i)a .
T h e s u m c o n v e r g e s a n d s o Va r [ 5 „ ] — n Va r [ X i ] + C . D
Remark. The assumption that the probability space fi is the interval [0,1] is
not crucial. Many probability spaces (fi, A, P) where Q is a compact metric
space with Borel cr-algebra A and P[{z}] = 0 for all x £ Q is measure
theoretically isomorphic to ([0,1],B, dx), where B is the Borel cr-algebra
on [0,1] (see [13] proposition (2.17). The same remark also shows that
the assumption An = [0, l/ns] is not essential. One can take any nested
sequence of sets An € A with P[An] = \/ns, and An+\ c An.
Figure. We can apply this propo

sition to a lattice point prob
lem near the graphs of one-
J / V ^
dimensional Brownian motion,
where we have a probability space
of paths and where we can make
a statement about almost every
path in that space. This is a re
sult in the geometry of numbers
for connected sets with fractal
boundary.
Corollary 5.9.5. Assume Bt is standard Brownian motion. For any 0 < 8 <
1/2, there exists a constant C, such that any l/n1+(5 neighborhood of the
graph of B over [0,1] contains at least C/n1_<5 lattice points, if the lattice
has a minimal spacing distance of 1/n.
Proof. Bt+i/n mod 1/n is not independent of Bt but the Poincare return
map T from time t = k/n to time (fc + l)/n is a Markov process from.
[0,1/n] to [0,1/n] with transition probabilities. The random variables Xi
have exponential decay of correlations as we have seen in lemma (5.9.3). □
Remark. A similar result can be shown for other dynamical systems with
strong recurrence properties. It holds for example for irrational rotations
with T(x) = x + a mod 1 with Diophantine a, while it does not hold for
Liouville a. For any irrational a, we have fn = ^=7 Xwb=i Ân(Tk(x)) near
1 for arbitrary large n = qi, where pi/qi is the periodic approximation of
8. However, if the qi are sufficiently far apart, there are arbitrary large n,
where fn is bounded away from 1 and where fn do not converge to 1.
The theorem we have proved above belongs to the research area of geome
try of numbers. Mixed with probability theory it is a result in the random
geometry of numbers.
A prototype of many results in the geometry of numbers is Minkowski's

theorem:
Theorem 5.9.6 (Minkowski theorem). A convex set M which is invariant

under the map T(x) = —x and with area > 4 contains a lattice point
different from the origin.
Proof. One can translate all points of the set M back to the square ft =
[-1,1] x [-1,1]. Because the area is > 4, there are two different points
(x, j/), (a, 6) which have the same identification in the square f2. But if
(x, y) = (u+2k, v+2l) then (x-u, y-v) = (2fc, 21). By point symmetry also
(a, b) = (-ix, -v) is in the set M. By convexity ((x+a)/2, (y+b)/2) = (fc, I)
is in M. This is the lattice point we were looking for. D
5.10. Arithmetic random variables
Figure. A convex, symmetric set

M. For illustration purposes, the
area has been chosen smaller
than 4 in this picture. The theo
rem of Minkowski assumes, it is
larger than 4.
Figure. Translate all points back

to the square [—1,1] x [—1,1] of
5SH ]
area 4. One obtains overlapping
points. The symmetry and con
vexity allows to conclude the ex ^MtlDl^n,
istence of a lattice point in M.
There are also open questions:
• The Gauss circle problem asks to estimate the number of 1/n-lattice

points gin) = im2 + E{n) enclosed in the unit disk. One believes that
an estimate £(n) < Cne holds for every 0 > 1/2. The smallest 6 for
which one knows the is 9 = 46/73.
• For a smooth curve of length 1 which is not a line, we have a similar

result as for the random walk but we need 8 < 1/3. Is there a result
for 6 < 1?
• If we look at Brownian motion in W1. How many 1/n lattice points

are there in a Wiener sausage, in a l/n1+s neighborhood of the path?
5.10 Arithmetic random variables

Because large numbers are virtually infinite - we have no possibility to in
spect all of of the numbers from Qn = {1,... n = 10100} for example -
functions like Xn = k2 + 5 mod n are accessible on a small subset only. The
function Xn behaves as random variable on an infinite probability space. If
we could find the events Un = {Xn = 0} easily, then factorization would
be easy as its factors can be determined from in Un. A finite but large
probability space Qn can be explored statistically and the question is how
much information we can draw from a small number of data. It is unknown
how much information can we get from a large integer n with finitely many
computations. Can, we statistically recover the factors of n from 0(log(n))
data points (kj,Xj), where Xj = n mod kj for example?
As an illustration of how arithmetic complexity meets randomness, we con

sider in this section examples of number theoretical random variables, which
can be computed with a fixed number of arithmetic operations. Both have
the property that they appear to be "random" for large n. These functions
belong to a class of random variables
X(k) = p(fc, n) mod g(fc, n) ,
where p and q are polynomials in two variables. For these functions, the
sets X~1(o) = {X(k) = a } are in general difficult to compute and
Y0(k) =X(k),Yx(k) =X(k + l),...,Yl(k) = X(k + l) behave very much
as independent random variables.
To deal with " number theoretical randomness", we use the notion of asymp
totically independence. Asymptotically independent random variables ap
proximate independent random variables in the limit n —▶ oo. With this
notion, we can study fixed sequences or deterministic arithmetic functions
on finite probability spaces with the language of probability, even so there is
no fixed probability space on which the sequences form a stochastic process.
Definition. A sequence of number theoretical random variables is a col
lection of integer valued random variables Xn defined on finite probability
spaces (ftn,.An,Pn) for which fin c f*n+i and An is the set of all subsets
of ftn. An example is a sequence Xn of integer valued functions defined
•on fin = {0,...,n — 1 }. If there exists a constant C such that Xn on
{0,..., n } is computable with a total of less than C additions, multiplica
tions, comparisons, greatest common divisor and modular operations, we
call X a sequence of arithmetic random variables.
Example. For example
Xn(x) = (((x5 - 7) mod 9)3x - x2) mod n
defines a sequence of arithmetic random variables on fin = {0,..., n - 1 }.
Example. If xn is a fixed integer sequence, then Xn(k) = Xk on Qn =
{0,..., n — 1 } is a sequence of number theoretical random variables. For
example, the digits xn of the decimal sequence of tt defines a sequence
of number theoretical random variables Xn(k) = xn for fc < n. However,
in the case of 7r, it is not known, whether this sequence is an arithmetic
sequence. It would be a surprise, if one could compute xn with a finite n-
independent number of basic operations. Also other deterministic sequences
like the decimal expansions of 7r, \[2 or the Mobius function p(n) appear
"random".
5.10. Arithmetic random variables 335
Remark. Unlike for discrete time stochastic processes Xn, where all ran
dom variables Xn are defined on a fixed probability space (fi,-4,P), an
arithmetic sequence of random variables Xn uses different finite probabil
ity spaces (ftn> Ai?Pn)-
Remark. Arithmetic functions are a subset of the complexity class P of

functions computable in polynomial time. The class of arithmetic sequences
of random variables is expected to be much smaller than the class of se
quences of all number theoretical random variables. Because computing
gcd(x,y) needs less than C(x + y) basic operations, we have included it
too in the definition of arithmetic random variable.
Definition. If limnôc E[Xn] exists, then it is called the asymptotic expec

tation of a sequence of arithmetic random variables. If limnôo Var[Xn]
exists, it is called the asymptotic variance. If the law of Xn converges, the
limiting law is called the asymptotic law.
Example. On the probability space ft„ = [1,..., n] x [1,..., n], consider the
arithmetic random variables Xd = lSd, where Sd = {(n,m),gcd(n,m) =
d}.
.2
Proposition 5.10.1. The asymptotic expectation Pn[Si] = En[Xi] is 6/n
In other words, the probability that two random integers are relatively
prime is 6/7T2.
Proof. Because there is a bijection (j) between Si on [1,..., n]2 and Sd on

[l,...,dn]2 realized by 0(j,fc) -▶ (dj.dk)', we have |Si|/n2 = \Sd\/(d2n2).
This shows that En[Xi]/En[Xd] -> d2 has a limit 1/d2 for n ^ oo. To
know P[5i], we note that the sets Sd form a partition'of N2 and also when
restricted to fin. Because P[Sd] = P[Si]/d2, one has
PN-(^4 + P + -) = P[Sl]I = 1,
so that P[Si] = 6/tt2. D
Figure. The probability that two

random integers are relatively
prime is 6/tt2. A cell (j, fc)
in the finite probability space
[1,..., n] x [1,..., n] is painted
black if gcd(j, fc) = 1. The proba
bility that gcd(j, fc) = 1 is 6/7T2 =
0.607927... in the limit n —> oo. s1
So, if you pick two large num
bers (j, fc) at random, the change
BS!:!!!!!:R Ifl-H-t
to have no common divisor is
slightly larger than to have a
common divisor.
Exercice. Show that the asymptotic expectation of the arithmetic random

variable Xn(x, y) = gcd(x, y) on [1,..., n]2 is infinite.
Example. A large class of arithmetic random variables is defined by
Xn(k) = p(n, fc) mod q(n, fc)
on f2n = {0,..., n — 1} where p and q are not simultaneously linear poly

nomials. We will look more closely at the following two examples:
1) Xn(k) — n2 + c mod fc
2) Xn(k) = fc2 +cmodn
Definition. Two sequences Xn,Yn of arithmetic random variables, (where

Xn,Yn are defined on the same probability spaces fin), are called uncor
related if Cov[Xn,Fn] — 0. The are called asymptotically uncorrelated, if
their asymptotic correlation is zero:
Cov[xn,yn]-+o
for n oo.
Definition. Two sequences X,Y of arithmetic random variables are called

independent if for every n, the random variables Xn,Yn are independent.
Two sequences X, Y of arithmetic random variables with values in [0, n]
are called asymptotically independent, if for all /, J, we have
P[*H
n e j ^ G
n j] _ P[^n e /] pfii ne j] _> o
for n oo.
Remark. If there exist two uncorrelated sequences of arithmetic random
variables U,V such that \\Un - Xn\\L2^n) —▶ 0 and \\Vn - l^||i,2(n„) -+ °>
then X, Y are asymptotically uncorrelated. If the same is true for indepen
dent sequences U, V of arithmetic random variables, then X, Y are asymp
totically independent.
Remark. If two random variables are asymptotically independent, they are

asymptotically uncorrelated.
Example. Two arithmetic random variables Xn(k) = fc mod n and Yn(k) —

ak + b mod n are not asymptotic independent. Lets look at the distribution
of the random vector (Xn, Yn) in an example:
Figure. The figure shows

the points (Xn(k),Yn(k)) for
Xn(k) = k,Yn(k) = 5fc + 3
modulo n in the case n = 2000.
There is a clear correlation be
tween the two random variables.
Exercice. Find the correlation of Xn(k) = fc mod n and Yn(k) = 5fc +

3 mod n.
Having asymptotic correlations between sequences of arithmetic random

variables is rather exceptional. Most of the time, we observe asymptotic
independence. Here are some examples:
Example. Consider the two arithmetic variables Xn(k) = k and
Yn(k) = cfc-1 mod p(n) ,
where c is a constant and p(n) is the n'th prime number. The random
variables Xn and Yn are asymptotically independent. Proof: by a lemma of
Merel [67, 23], the number of solutions of (x, y) G I x J of xy = c mod p is
This means that the probability that Xn/n G In, Yn/n G Jn is |JTn| • \Jn\.
Figure. Illustration of the lemma

of Merel. The picture shows the
points {(fc, 1/fc) mod p }, where
p is the 200'th prime number
p(200) = 1223.
Nonlinear polynomial arithmetic random variables lead in general to asymp

totic independence. Lets start with an experiment:
Figure. We see the points

(Xn(k),Yn(k)) for Xn(k) =
fc,l^(fc) = fc2 + 3 in the case -..•",- » --'•..
n = 2001. Even so there are
narrow regions in which some
correlations are visible, these
regions become smaller and . ' ' ' /
smaller for n —> oo. Indeed, we - ' '
will show that Xn,Yn are asymp
totically independent random
variables.
The random variable Xn(k) = (n2 + c) mod fc on {1,..., n} is equivalent

to Xn(k) = n mod fc on {0,..., [y/n - c] }, where [x] is the integer part of
x. After the rescaling the sequence of random variables is easier to analyze.
To study the distribution of the arithmetic random variable Xni we can

also rescale the image, so that the range in the interval [0,1]. The random
variable Yn = Xn(x • |fin|) can be extended from the discrete set {fc/|fin|)}
to the interval [0,1]. Therefore, instead of n2 + c mod fc, we look at
n mod fc n rn,
xnik) = ^r- = --[-}
on fim(n) = {1,..., m(n) }, where m(n) = y/n — c.
Elements in the set X~x(0) are the integer factors of n. Because factoring
is a well studied NP problem, the multi-valued function X~x is probably
hard to compute in general because if we could compute it fast, we could
factor integers fast.
Proposition 5.10.2. The rescaled arithmetic random variables

n mod fc n
Xn(k) = ["]
fc LfcJ
converge in law to the uniform distribution on [0,1].
Proof. The functions f„(k) = n/(k+r)-[n/(k+r)] are piecewise continuous

circle maps on [0,1]. When rescaling the argument [0,... ,n], the slope of
the graph becomes larger and larger for n —▶ oo. We can use lemma (5.10.3)
below. □
Figure. Data points
,, n mod k N
for n = lO'OOO and 1 < fc <

n. For smaller values of k, the
data points appear random. The
points are located on the graph of
the circle map
n
Ut) 1 ?>
'
To show the asymptotic independence of Xn with any of its translations,
we restrict the random vectors to [1, l/na] with a < 1.
Lemma 5.10.3. Let fn be a sequence of smooth maps from [0,1] to the circle
T1 = R/Z for which (f-l)"(x) -> 0 uniformly on [0,1], then the law \xn of
the random variables Xn(x) = (x, fn(x)) converges weakly to the Lebesgue
measure /i = dxdy on [0,1] x T1.
Proof. Fix an interval [a, b] in [0,1]. Because ^n([a, b] x T1) is the Lebesgue
measure of {(#, y) \Xn(x, y) G [a, b]} which is equal to b — a, we only need
to compare
pn([a,b] x[c,c + dy])
and
pn([a,b] x [d,d + dy])
in the limit n -^ oo. But p,n([a,b] x [c,c + dy]) - p,n([a,b] x [c,c + dy}) is
bounded above by
K/n'/w-^mi^K/n-1)"^)!
which goes to zero by assumption.
Figure. Proof of the lemma. The

measure \in with support on the
graph of fn(x) converges to the
Lebesgue measure on the prod
uct space [0,1] x T1. The con
dition f'jf2 —> 0 assures that
the distribution in the y direction
smooths it out. .
Theorem 5.10.4. Let c be a fixed integer and Xn(k) = (n2 + c) mod fc

on {1,... ,n} For every integer r > 0,0 < a < 1, the random variables
X(k),Y(k) = X(k + r) are asymptotically independent and uncorrelated
on [0,na].
Proof. We have to show that the discrete measures Y%=i $(x(k),Y(k))

converge weakly to the Lebesgue measure on the torus. To do so, we first
look at the measure pn = /0 Y%=i&(X(k))Y(k)) which is supported on
the curve 11-> (X(t),Y(t)), where fc G [0,na] with a < 1 converges weakly
to the Lebesgue measure. When rescaled, this curve is the graph of the
circle map fn(x) = l/x mod 1 The result follows from lemma (5.10.3). □
Remark. Similarly, we could show that the random vectors (X(k),X(k +

ri), X(k + r2),..., X(k + ri)) are asymptotically independent.
Remark. Polynomial maps like T(x) = x2 + c are used as pseudo random

number generators for example in the Pollard p method for factorization
[84]. In that case, one considers the random variables {0,... ,n — 1} de
fined by -Xo(fc) = fc, Xn+\(k) = T(Xn(k)). Already one polynomial map
produces randomness asymptotically as n —▶ oo.
Theorem 5.10.5. If p is a polynomial of degree d > 2, then the distribution

of Y(k) = p(k) mod n is asymptotically uniform. The random variables
X(k) = fc and Y(k) = p(k) mod n are asymptotically independent and
uncorrelated.
Proof. The map can be extended to a map on the interval [0, n]. The graph
(x, T(x)) in {1,..., n} x {1,..., n} has a large slope on most of the square.
Again use lemma (5.10.3) for the circle maps fn(x) = p(nx) mod n on
[0,1]. D
Figure. The slope of the graph

of p(x) mod n becomes larger
and larger as n —▶ oo. Choos
ing an integer fc G [0,n] pro
duces essentially a random value
p(k) mod n. To prove the asymp
totic independence, one has to
verify that in the limit, the push
forward of the Lebesgue measure
on [0, n] under the map f(x) =
(x,p(x)) mod n converges in
law to the Lebesgue measure on
[0,n]2.
Remark. Also here, we deal with random variables which are difficult to
invert: if one could find Y~l(c) in 0(P(log(n)) times steps, then factoriza
tion would be in the complexity class P of tasks which can be computed
in polynomial time. The reason is that taking square roots modulo n is at
least as hard as factoring is the following: if we could find two square roots
x, y of a number modulo n, then x2 = y2 mod n. This would lead to factor
gcd(x - y, n) of n. This fact which had already been known by Fermat. If
factorization was a NP complete problem, then inverting those maps would
be hard.
Remark. The Mobius function is a function on the positive integers defined

as follows: the value of //(n) is defined as 0, if n has a factor p2 with a prime p
and is (-l)fc, if it contains fc distinct prime factors. For example, /i(14) = 1
and //(18) = 0 and //(30) = -1. The Mertens conjecture claimed hat
M(n) = |/x(l) + • • • + p(n)\ < Cyfr

for some constant C. It is now believed that M(n)/y/n is unbounded but it
is hard to explore this numerically, because the x/loglog(n) bound in the
law of iterated logarithm is small for the integers n we are able to compute
- for example for n = IO100, one has ^/loglog(n) is less then 8/3. The fact
M(n)
n =£i>(*)-o
k=i
is known to be equivalent to the prime number theorem. It is also known

that lim sup M(n)/y/n > 1.06 and lim inf M(n)/y/n < -1.009.
If one restricts the function fi to the finite probability spaces Qn of all
numbers < n which have no repeated prime factors, one obtains a sequence
of number theoretical random variables Xn, which take values in {-1,1}.
Is this sequence asymptotically independent? Is the sequence /x(n) random
enough so that the law of the iterated logarithm
lim sup J2 , = < 1

n-ôo £-J y^nloglogtn)
holds? Nobody knows. The question is probably very hard, because if it

were true, one would have
M(n) < n1/2+e, for all e > 0

which is called the modified Mertens conjecture . This conjecture is known
to be equivalent to the Riemann hypothesis, the probably most notori
ous unsolved problem in mathematics. In any case, the connection with
the Mobius functions produces a convenient way to formulate the Rie
mann hypothesis to non-mathematicians (see for example [14]). Actually,
the question about the randomness of p,(n) appeared in classic probability
text books like Fellers. Why would the law of the iterated logarithm for
the Mobius function imply the Riemann hypothesis? Here is a sketch of
the argument: the Euler product formula - sometimes referred to as "the
Golden key" - says
« • ) - £n ?= l- n » - ? ) - '■
p prime
The function £(s) in the above formula is called the Riemann zeta function.
With M(n) < n1/2*6, one can conclude from the formula
J_ = y> Mn)
cm h n*
that £(s) could be extended analytically from Re(s) > 1 to any of the
half planes Re(s) > 1/2 + e. This would prevent roots of ((s) to be to the
right of the axis Re(s) = 1/2. By a result of Riemann, the function A(s) =
-ir~s/2r(s/2)C(s) is a meromorphic function with a simple pole at s = 1 and
satisfies the functional equation A(s) = A(l — s). This would imply that
C(s) has also no nontrivial zeros to the left of the axis Re(s) = 1/2 and
343
5.11. Symmetric Diophantine Equations
that the Riemann hypothesis were proven. The upshot is that the Riemann
hypothesis could have aspects which are rooted in probability theory.
Figure. The sequence Xk =

p(l(k)), where /(fc) is the k
nonzero entry in the sequence
{/i(l), /x(2), /x(3),... } produces a
"random walk" Sn = ^2k=i^k'
While Xk is a deterministic se
quence, the behavior of Sn re
sembles a typical random walk.
If that were true and the law of
the iterated logarithm would hold,
this would imply the Riemann
hypothesis.
5.11 Symmetric Diophantine Equations

Definition. A Diophantine equation is an equation /(xi,..., xk) = 0, where
p is a polynomial in fc integer variables xi, ..., xfc and where the polynomial
/ has integer coefficients. The Diophantine equation has degree m if the
polynomial has degree ra. The Diophantine equation is homogeneous, if
every summand in the polynomial has the same degree. A homogeneous
Diophantine equation is also called a form.
Example. The quadratic equation x2 + y2 - z2 = 0 is a homogeneous

Diophantine equation of degree 2. It has many solutions. They are called
Pythagorean triples. One can parameterize them all with two parameters
5, t with x = 2s£, y = s2 -12, z = s2+t2, as has been known since antiquity
already [15].
Definition. A Diophantine equation of the form
p(xi,...,xfc) =p(yi,...,2/fc)
is called a symmetric Diophantine equation. More generally, a Diophantine

equation i
£*r = £
i=l j=l
m
is called an Euler Diophantine equation of type (fc, I) and degree ra. It is a

symmetric Diophantine equation if fc = I. [28, 35, 15, 4, 5]
Remark. An Euler Diophantine equation is equivalent to a symmetric Dio

phantine equation if m is odd and fc + / is even.
Definition. A solution (xl5.., xk), (yu ..., yk) to a symmetric Diophantine
equation p(x) = p(y) is called nontrivial, if {Xl, ... ,xk } and {yu... ,yk }
are different sets. For example, 53 + 73 + 33 = 33 + 73 + 53 is a trivial
solution of p(x) = p(y) with p(x, y, z) = x3 + y3 + z3.
The following theorem was proved in [68]:
Theorem 5.11.1 (Jaroslaw Wroblewski 2002). For fc > ra, the Diophantine
equation xf + • • • + x£* = y™ + ... + j,™ has infinitely many nontrivial
solutions.
Proof. Let R be a collection of different integer multi-sets in the finite

set [0,...,n]fc. It contains at least nk/k\ elements. The set S = {p(x) =
XT + ' • • + xk € [0, Vknml2} | x G R } contains at least nk/k\ numbers.
By the pigeon hole principle, there are different multi-sets x,y for which
p(x) = p(y). This is the case if nk/k\ > Vknm or nfc"m > klVk. □
The proof generalizes to the case, where p is an arbitrary polynomial of
degree ra with integer coefficients in the variables xx,..., xk.
Theorem 5.11.2. For an arbitrary polynomial p in fc variables of degree

ra, the Diophantine equation p(x) = p(y) has infinitely many nontrivial
solutions.
Remark. Already small deviations from the symmetric case leads to local
constraints: for example, 2p(x) = 2p(y) +1 has no solution for any nonzero
polynomial p in fc variables because there are no solutions modulo 2.
Remark. It has been realized by Jean-Charles Meyrignac, that the proof
also gives nontrivial solutions to simultaneous equations like p(x) = p(y) =
p(z) etc. again by the pigeon hole principle: there are some slots, where more
than 2 values hit. Hardy and Wright [28] (theorem 412) prove that in the
case fc = 2, ra = 3: for every r, there are numbers which are representable
as sums of two positive cubes in at least r different ways. No solutions
of xf + yf=x% + y$ = x$ + j/| were known to those authors [28], nor
whether there are infinitely many solutions for general (fc,ra) = (2,ra).
Mahler proved that x3 + y3 + z3 = 1 has infinitely many solutions. It is
believed that x3+y3+z3+w3 = n has solutions for all n. For (fc, ra) = (2,3),
multiple solutions lead to so called taxi-cab or Hardy-Ramanujan numbers.
Remark. For general polynomials, the degree and number of variables alone
does not decide about the existence of nontrivial solutions of p(xi, ...,xk) =
P(yi, • • •, 2/fc). There are symmetric irreducible homogeneous equations with
5 . 11 . Symmetric Diophantine Equations 345
fc < ra/2 for which one has a nontrivial solution. An example is p(x, y) =
x5 - Ay5 which has the nontrivial solution p(l, 3) = p(4, 5).
Definition. The law of a symmetric Diophantine equation p(xi,..., xk) =

p(xi,..., Xk) with domain Q = [0,..., n]h is the law of the random variable
defined on the finite probability space fi.
Remark. Wroblewski's theorem holds because the random variable has an

average density which is larger than the lattice spacing of the integers. So,
there have to be different integers, which match. The continuum analog is
that if a random variable X on a domain Q takes values in [a, b] and b — a
is smaller than the area of Q, then the density fx is larger than 1 at some
point.
Remark. Wroblewski's theorem covers cases like x2 +y2 + z2 = u2 + v2 +w2

or x3 + y3 + z3 + w3 = a3 + b3 + c3 + d3. It is believed that for fc > ra/2,
there are infinitely many solutions and no solution for fc < ra/2. [59].
Remark. For homogeneous Diophantine equations, it is enough to find a

single nontrivial solution (xi,...,xk) to obtain infinitely many. The reason
is that (raxi,..., mxk) is a solution too, for any ra ^ 0.
Here are examples of solutions. Sources are [69, 35, 15]:
k=2,m = 4 (59, 158)4 = (133, 134)4 (Euler, gave algebraic solutions in 1772 and 1778)
k = 2,m = 5 (open problem ([35]) all sums < 1.02 • IO26 have been tested)
k=3,m = 5 (3, 54, 62)5 = (24, 28, 67)5 ([59], two parametric solutions by Moessner 1939, Swinnerton-Dyer)
k=3,m=6 (3, 19, 22)6 = (10, 15, 23)6 ([28],Subba Rao, Bremner and Brudno parametric solutions)
k=3,m=7 open problem?
k=4,m=7 (10, 14, 123, 149)7 = (15, 90, 129, 146)7 (Ekl)
k=4,m = 8 open problem?
k=5,m=7 (8, 13, 16, 19)7 = (2, 12, 15, 17, 18)7 ([59])
k = 5,m = 8 (1, 10, 11, 20, 43)8 = (5, 28, 32, 35, 41)8.
k=5,m = 9 (192, 101, 91, 30, 26)9 = (180, 175, 116, 17, 12)9 (Randy Ekl, 1997)
k = 5,m=10 open problem
k=6,m=3 (3, 19, 22)6 = (10, 15, 23)6 (Subba Rao [59])
k=6,m=10 (95, 71, 32, 28, 25, 16)10 = (92, 85, 34, 34, 23, 5)10 (Randy Ekl,1997)
k=6,m = ll open problem?
k=7,m=10 (1, 8, 31, 32, 55, 61, 68)10 = (17, 20, 23, 44, 49, 64, 67)10 ([59])
k=7,m=12 (99, 77, 74, 73, 73, 54, 30)12 = (95, 89, 88, 48, 42, 37, 3)12 (Greg Childers, 2000)
k=7,m=13 open problem?
k=8,m=ll (67, 52, 51, 51, 39, 38, 35, 27)11 = (66, 60, 47, 36, 32, 30, 16, 7)11 (Nuutti Kuosa, 1999)
k=20,m = 21 (76, 74, 74, 64, 58, 50, 50, 48, 48, 45, 41, 32, 21, 20, 10, 9, 8, 6, 4, 4)21
= (77, 73, 70, 70, 67, 56, 47, 46, 38, 35, 29, 28, 25, 23, 16, 14, 11, 11, 3, 3)21 (Greg Childers, 2000)
k=22,m=22 (85, 79, 78, 72, 68, 63, 61, 61, 60, 55, 43, 42, 41, 38, 36, 34, 30, 28, 24, 12, 11, ll)22
= (83, 82, 77, 77, 76, 71, 66, 65, 65, 58, 58, 54, 54, 51, 49, 48, 47, 26, 17, 14, 8, 6)22 (Greg Childers, 2000)
Figure. Known cases of (fc,ra)

with nontrivial solutions x,y
of symmetric Diophantine equa
tions g(x) = g(y) with g(x) =
Xj1 H hxf. Wroblewski 's theo
rem assures that for fc > ra, there
are solutions. The points above
the diagonal beat Wroblewski's
theorem. The steep line m =
2fc is believed to be the thresh
old for the existence of nontrivial
solutions. Above this line, there
should be no solutions, below,
there should be nontrivial solu
tions.
What happens in the case fc = ra? There is no general result known. The
problem has a probabilistic flavor because one can look at the distribution
of random variables in the limit n —▶ oo:
Lemma 5.11.3. Given a polynomial p(x\,... , x^) with integer coefficients

of degree fc. The random variables
Xn(xi,...,xfc) =p(xi,..,xk)/nk
on the finite probability spaces Vtn = [0, ...,n]fc converge in law to the
random variable X(x\,... ,xn) = p(xi,..,Xfc) on the probability space
([0, l]fc,Z3,P), where B is the Borel cr-algebra and P is the Lebesgue mea
sure.
Proof. Let Sa^(n) be the number of points (xi,... ,x^) satisfying
p(xi,...,xfc) G [nka,nkb] .
This means
^W=Fn(&)-Fn(a),
where Fn is the distribution function of Xn. The result follows from the fact
that Fn(b) — Fn(a) = Sa^(n)/nk is a Riemann sum approximation of the in
tegral F(b) - F(a) = JAa b 1 dx, where Aa,b = {x G [0, l]k \ X(xi,..., xk) G
(a, b)}. ^ □
Definition. Lets call the limiting distribution the distribution of the sym
metric Diophantine equation. By the lemma, it is clearly a piecewise smooth
function.
5 . 11 . Symmetric Diophantine Equations 347
Example. For fc = 1, we have F(s) = P[X(x) < s] = P[xm < s] = a^/n.

The distribution for fc = 2 for p(x,y) = x2 + y2 and p(x,y) = x2 - y
were plotted in the first part of these notes. The distribution function of
p(xi,x2,.. • ,xfc) is a k'th convolution product Fk = F * • • • • F, where
p(sj = o(sl/m) near 5 = 0. The asymptotic distribution of p(x, y) = x2+y2
is bounded for all ra. The asymptotic distribution of p(x, y) = x2 - y2
is unbounded near s = 0 Proof. We have to understand the laws of the
random variables X(x, y) = x2+y2 on [0, l]2. We can see geometrically that
(?r/4)s2 < Fx(s) < s2. The density is bounded. For Y(x,y) = x2 - y , we
use polar coordinates F(s) = {(r, 6) \ r2 cos(26)/2 < s }. Integration shows
that F(s) = Cs2 + f(s), where f(s) grows logarithmically as -log(s). For
ra > 2, the area xm - ym < s is piecewise differentiable and the derivative
stays bounded.
Remark. If p is a polynomial of fc variables of degree fc. If the density
/ = F' of the asymptotic distribution is unbounded, then then there are
solutions to the symmetric Diophantine equation p(x) = p(y).
Corollary 5.11.4. (Generalized Wroblewski) Wroblewski's result extends to

polynomials p of degree fc for which at least one variable appears in a term
of degree smaller than fc.
Proof. We can assume without loss of generality that the first variable
is the one with a smaller degree ra. If the variable xi appears only in
terms of degree fc - 1 or smaller, then the polynomial p maps the finite
space [0, n]fc/m x [0, n}k~l with n*^/™"1 = nfc+e elements into the interval
[min(p), max(p)] C [-Cnk, Cnk). Apply the pigeon hole principle. □
Example. Let us illustrate this in the case p(x, y, z, w) = x4 + x3 + zA + w4.

Consider the finite probability space ftn = [0,n] x [0,n] x [0,n4/3] x [0,n]
with n4+1/3. The polynomial maps f>n to the interval [0,4n4]. The pigeon
hole principle shows that there are matches.
Theorem 5.11.5. If the density fp of the random variable p on a surface

ft C [0,n]fc is larger than fc!, then there are nontrivial solutions to p(x) =
p(y)-
In general, we try to find a subsets ft C [0,n]fc C Rk which contains nk~^

points which is mapped by X into [0,nm"a]. This includes surfaces, sub
sets or points, where the density of X is large. To decide about this, we
definitely have to know the density of X on subsets. This works often be
cause the polynomials p modulo some integer number L do not cover all
'the conjugacy classes. Much of the research in this part of Diophantine
equations is devoted to find such subsets and hopefully parameterize all of

the solutions.
Figure. X(x, y, z) = x3 + y3 + z3. Figure. X(x, y\ z) = x3 + y3 - z3
Exercice. Show that there are infinitely many integers which can be written
in non trivially different ways as x4 + y4 + z4 - w2.
Remark. Here is a heuristic argument for the "rule of thumb" that the Euler
Diophantine equation xf + • + xf = xjf1 has infinitely many solutions for
fc > ra and no solutions if fc < ra.
For given n, the finite probability space ft = {(xi,..., xk) | 0 < x{ < n1/™ }
contains nk/m different vectors x = (xu ..., xk). Define the random variable
X(x) = (x? + -.. + xf)1/™.

We expect that X takes values l/nfc/m = nmlk close to an integer for large
n because Y(x) = X(x) mod 1 is expected to be uniformly distributed on
the interval [0,1) as n —▶ oo.
How close do two values Y(x),Y(y) have to be, so that Y(x) = Y(y)l
Assume Y(x) = Y(y) + e. Then
X(xr = X(yr + eX(yr-1 + 0(e2)

with integers X(x)™, X(y)™. lfX(y)rn~1e < 1, then it must be zero so that
Y(x) = Y(y). With the expected e = nmlk and X(y)m~1 < Cn^-V/™ we
see we should have solutions if fc > ra - 1 and none for fc < ra - 1. Cases
like ra = 3, fc = 2, the Fermat Diophantine equation
x3 + y3 = z3
5.12. Continuity of random variables 349
are tagged as threshold cases by this reasoning.
This argument has still to be made rigorous by showing that the distri
bution of the points f(x) mod 1 is uniform enough which amounts to
understand a dynamical system with multidimensional time. We see nev
ertheless that probabilistic thinking can help to bring order into the zoo
of Diophantine equations. Here are some known solutions, some written in
the Lander notation
xm = ix1,...,xk)m = x? + --- + xZ.
m = 2fc = 2: x2 + y2 = z2 Pythagorean triples like 32 + 42 = 52 (1900 BC).
m = 3fc = 2: xm + ym = zm impossible, by Fermat's theorem.
m = 3, k = 3: x3 + y3 + u3 = v3 derived from taxicab numbers, like IO3 + 93 = l3 + 123 (Viete 1591).
m = 4, fc = 3: 26824404 + 153656394 + 187967604 = 206156734 (Elkies 1988 [24]) m = 5, fc = 3: like
x5 + y5 + z5 = w5 is open
m = 4, fc = 4: 304 + 1204 + 2724 + 3154 = 3534. (R. Norrie 1911 [35])
m = 5, fc = 4 275 + 845 + HO5 + 1335 = 1445 (Lander Parkin 1967).
m = 6, k = 5: x6 + y6 + z6 + %i6 + v6 = w6 is open.
m = 6, fc = 6: (74, 234, 402, 474, 702, 894, 1077)6 = 11416.
m = 7, fc = 7: (525, 439, 430, 413, 266, 258, 127)7 = 5687 (Mark Dodrill, 1999)
m = 8, fc = 8: (1324, 1190, 1088, 748, 524, 478, 223, 90)8 = 14098 (Scott Chase)
m = 9, fc = 12, (91, 91, 89, 71, 68, 65, 43, 42, 19, 16, 13, 5)9 = 1039 (Jean-Charles Meyrignac,1997)
5.12 Continuity of random variables

Let X be a random variable on a probability space (ft,.A,P). How can
we see from the characteristic function </>x whether X is continuous or
not? If it is continuous, how can we deduce from the characteristic function
whether X is absolutely continuous or not? The first question is completely
answered by Wieners theorem given below. The decision about singular
or absolute continuity is more subtle. There is a necessary condition for
absolute continuity:
Theorem 5.12.1 (Riemann Lebesgue-lemma). If X e C1, then <j>x(n)

for \n\ —▶ oo.
Proof. Given e > 0, choose n so large that the n'th Fourier approximation
Xn(x) = EL-n <t>x(n)einx satisfies \\X - Xn\\i < e. For ra > n, we have
(MXn) = E[eimXri] = 0 so that
\<t>x(m)\ = \4>x-xn(m)\ < \\X - Xn||i < e .
D
Remark. The Riemann-Lebesgue lemma can not be reversed. There are
random variables X for which <£x(n) -> 0, but which X is not in Cl.
Here is an example of a criterion for the characteristic function which as
sures that X is absolutely continuous:
Theorem 5.12.2 (Convexity). If an = a_n satisfies an -▶ 0 for n -+ oo and

an+i _ 2an + an_i > 0, then there exists a random variable X e C1 for
which (f)x(n) = an.
Proof. We follow [48].

(i) bn = an — an+i decreases monotonically.
Proof: the convexity condition is equivalent to an - an+1 < an_i - an.
(ii) bn — an — an+i is non-negative for all n.
Proof: bn decreases monotonically. If some bn = c < 0, then by (i), also
bm < c for all ra contradicting the assumption that bn —▶ 0.
(iii) Also nbn goes to zero.
Proof: Because Y!k=i(ak-a>k+i) = ai -an+i is bounded and the summands
are positive, we must have k(ak - ak+{) —▶ 0.
(iv) ELi k(ak-i - 2ak + ak+i) -♦ 0 for n -» oo.
Proof. This sum simplifies to a0 - an+i - n(an - an+i. By (iiii), it goes to
0 for n -^ oo.
(v) The random variable Y(x) = ^Zi K<*k-i - 2ak + ak+i)Kk(x) is in
C , if Kk(x) is the Fejer kernel with Fourier coefficients 1 - \j\/(k + 1).
Proof. The Fejer kernel is a positive summability kernel and satisfies
1 f27r
ll^lli = ^y Kk(x)dx = l.
for all fc. The sum converges by (iv).

(vi) The random variables X and Y have the same characteristic functions.
Proof.
cj)Y(n) = ]T k(ak-i - 2ak + ak+i)Kk(n)

fc=i
= ^k(ak-i-2ak + aM)(l--^-)
k=i K + L
oo I ..
= YI kiak-i - 2ak -r ak+i)(l - £^y) = an
n+l
□
For bounded random variables, the existence of a discrete component of
the random variable X is decided by the following theorem. It will follow
from corollary (5.12.5) given later on.
Theorem 5.12.3 (Wiener theorem). Given X e C°° with law p supported

in [—7r,7r] and characteristic function (j) = <t>x- Then
lim if>x(A0|2
n — >' n
oon k = î - ^
= ^x$>[X
~e -R'
= z]2
Therefore, X is continuous if and only if the Wiener averages

££Lil<Mfc)l2 converge to 0.
Lemma 5.12.4. If p is a measure on the circle T with Fourier coefficients

/ifc, then for every x G T, one has
M'V-^dritte"'-
Proof. We follow [48]. The Dirichlet kernel
t_ „ sm(t/2)
satisfies n
A,*/(ao = s„(/)o»o= E /(*)cite
k=—n
The functions
_—inx jint
^) = 2^riD"(i-x) = 2^n,^
k=—n
are bounded by 1 and go to zero uniformly outside any neighborhood of

t = x. From
lim / \d(p - p({x})Sx)\ = 0
follows
lim (fn,p-p({x})) =0
so that
(/„,/*- MW)) = 2^Vi E ^n)einx - MW) - o

Theorem 5.12.7 (Strichartz). Let p be a uniformly /i-continuous measure

on the circle. There exists a constant C such that for all n
it^nc-hi1-).
Proof. The computation ([102, 103] for the Fourier transform was adapted
to Fourier series in [51]). In the following computation, we abbreviate dp(x)
with dx:
n-l i n-1 ( * + » )2
1 *-^T r1 n ~ l e n =T-
I E we <. ./ E
k=-n '» k=-n U
d0 \pk |2
-1 n-1 -i*±Jli -
=2 e Y ~ / c"*^-*)* dxd</d0
i m eP - ^n - i { * - y ) k
dOdxdy
u U
=4 e / / ( ^-"l2"2+i(x-v)e
n-1 e_{h±±+i{x_y)n?
/ J dOdxdy
k——n
and continue
.. n—1
-EM *
k=—n
e / e-(*-»)9l5r| /'
A2 Jo
^1 ,,_«•£•+(._,,)$)»
/ „ dd\ dxdy
«——n
=6 e f / . [ f°°/ =
e-(*+<(*-v)*)a
e f t e - ( s - ,t f ), ,2„2
* dxdy
=7 e V^/
" " (e-(x~y)2l£)dxdy
Jt2
<8 e yft{( e-^-y^^dxdy)1/2
Jt2
oo «
=9 eV 5FE/ e-^-^^dxd^)1^
fc=0 ^/^<l^-2/l<(fc+l)/n
oo
<io e^Ci/i(n-1)(Ee_fc2/2)1/2
fc=o
are tagged as threshold cases by this reasoning.
This argument has still to be made rigorous by showing that the distri
bution of the points f(x) mod 1 is uniform enough which amounts to
understand a dynamical system with multidimensional time. We see nev
ertheless that probabilistic thinking can help to bring order into the zoo
of Diophantine equations. Here are some known solutions, some written in
the Lander notation
xm = (x1,...,Xk)rn=x? + --- + xf.
m = 2fc = 2: x2 + y2 = z2 Pythagorean triples like 32 + 42 = 52 (1900 BC).
m = 3fc = 2: xm + yrn = zm impossible, by Fermat's theorem.
m = 3, fc = 3: x + y + u3 = v3 derived from taxicab numbers, like IO3 + 93 = l3 -f 123 (Viete 1591).
m = 4, fc = 3: 26824404 + 153656394 + 187967604 = 206156734 (Elkies 1988 [24]) m = 5, fc = 3: like
x5 + y5 + z5 = w5 is open
m = 4, fc = 4: 304 + 1204 + 2724 + 3154 = 3534. (R. Norrie 1911 [35])
m = 5, fc = 4 275 + 845 + HO5 + 1335 = 1445 (Lander Parkin 1967).
m = 6, fc = 5 xu + y° + zD + u° + v°
m = 6, fc = 6 (74, 234, 402,474,702,894,1077)6 = 11416.
m = 7, fc = 7: (525, 439, 430, 413, 266, 258, 127)7 = 5687 (Mark Dodrill, 1999)
m = 8, fc = 8: (1324, 1190, 1088, 748, 524, 478, 223, 90)8 = 14098 (Scott Chase)
m = 9, fc = 12, (91, 91, 89, 71, 68, 65, 43, 42, 19, 16, 13, 5)9 = 1039(Jean-Charles Meyrignac,1997)
5.12 Continuity of random variables

Let X be a random variable on a probability space (Q,A,P). How can
we see from the characteristic function (j)X whether X is continuous or
not? If it is continuous, how can we deduce from the characteristic function
whether X is absolutely continuous or not? The first question is completely
answered by Wieners theorem given below. The decision about singular
or absolute continuity is more subtle. There is a necessary condition for
absolute continuity:
Theorem 5.12.1 (Riemann Lebesgue-lemma). If X <E C1, then <j)X(n)

for \n\ —> oo.
Proof. Given e > 0, choose n so large that the n'th Fourier approximation
xn(x) = EL-n 4>x(n)einx satisfies \\X - Xn\\x < e. For m > n, we have
<t>rn(Xn) = E[eimX"} = 0 so that
\<f>x(m)\ = \cf>x-xn(™)\ < \\X - X^^ < e .
D
Remark. The Riemann-Lebesgue lemma can not be reversed. There are
random variables X for which (j)X(n) -+ 0, but which X is not in C1.
Here is an example of a criterion for the characteristic function which as
sures that X is absolutely continuous:
Theorem 5.12.2 (Convexity). If an = a_n satisfies an -▶ 0 for n -▶ oo and

an+i - 2an + an_i > 0, then there exists a random variable X e C1 for
which (f)X(n) = an.
Proof. We follow [48].

(i) bn = an — an+i decreases monotonically.
Proof: the convexity condition is equivalent to an - an+i < an-i - an.
(ii) bn = an — an+i is non-negative for all n.
Proof: bn decreases monotonically. If some bn = c < 0, then by (i), also
bm <c for all m contradicting the assumption that bn —▶ 0.
(iii) Also nbn goes to zero.
Proof: Because Ylk=i(ak-ak+i) = ai-an+i is bounded and the summands
are positive, we must have k(dk - aî) —▶ 0.
(iv) ELi k(ak-i - 2ak + afc+i) -▶ 0 for n -> oo.
Proof. This sum simplifies to a0 - an+i - n(an - an+i. By (iiii), it goes to
0 for n —> oo.
(v) The random variable Y(x) = Yî k^k-i - 2ak + ak+1)Kk(x) is in
C , if Kk(x) is the Fejer kernel with Fourier coefficients 1 - \j\/(k + 1).
Proof. The Fejer kernel is a positive summability kernel and satisfies
1 f2n
for all k. The sum converges by (iv).

(vi) The random variables X and Y have the same characteristic functions.
Proof.
(j)Y(n) = Y^ k(ak-i - 2ak + ak+1)Kk(n)

fc=i
= ^%fc_i-2afc + aife+1)(l--^T)
k=i K + i
= Y^^-1 -2afcH-afc+i)(l- ^y) = On •

n+l
D
For bounded random variables, the existence of a discrete component of
the random variable X is decided by the following theorem. It will follow
from corollary (5.12.5) given later on.
Theorem 5.12.3 (Wiener theorem). Given X e C°° with law p supported

in [—7r,7r] and characteristic function </> = (frx- Then
lim -Y>x(fc)|2 = £P[X = a;]2-
Therefore, X is continuous if and only if the Wiener averages

ÊLil^WI2 converge toO.
Lemma 5.12.4. If n is a measure on the circle T with Fourier coefficients

/tk, then for every xeT, one has
ikx
^ ^ i t ^ T ik = —En J i^k t
Proof. We follow [48]. The Dirichlet kernel
n .n_W
V V J- k t2 ^_ e s i'n (s( fi cn (+t / 2l / 2
) )t)
k=-n V ' '
satisfies n
Akx
Dn*fix) = Snif)ix)= J2 f(k)el
k——n
The functions
—inxint
>»<«> = ^»<«-x) = 2^1 £
k=—n
are bounded by 1 and go to zero uniformly outside any neighborhood of

t = x. From
lim / \d(p - p({x})5x)\ = 0
e^°Jx-e
follows
lim (/n,/i-/i(W)) =0
so that
ifn,H - M({*})> = ^ J2 <f>iny™ - »i{x)) - 0

k=—n
D
Definition. If p and v are two measures on (fi = T, A), then its convolution
is defined as
p • v(A) = / p(A - x) dv(x)
Jt
for any A e A. Define for a measure on [-7r,7r] also p*(A) = p(-A).
Remark. We have p*(n) = p(n) and p*v(n) = p(n)i>(n). If p = J2ajSxj
is a discrete measure, then p* = J2^js-xj- Because p*p* = ^ \o,j\2, we
have in general
(^^)({o}) = EWW)2l'
Corollary 5.12.5. (Wiener) £*gt \p({x})\2 = limnôo ^pi EL-n IAn
Remark. For bounded random variables, we can rescale the random vari
able so that their values is in [~7r, it] and so that we can use Fourier series
instead of Fourier integrals. We have also
x£R J n
We turn our attention now to random variables with singular continuous

distribution. For these random variables, one does have P[X = c) = 0 for
all c. Furthermore, the distribution function Fx of such a random variable
X does not have a density. The graph of Fx looks like a Devil staircase.
Here is a refinement of the notion of continuity for measures.
Definition. Given a function h : R -» [0, oo) satisfying limx^0 h(x) = 0. A
measure p on the real line or on the circle is called uniformly /i-continuous,
if there exists a constant C such that for all intervals J = [a, b] on T the
inequality
p(I) < Ch(\I\)
holds, where |/| = b - a is the length of L For h(x) = xa with 0 < a < 1,
the measure is called uniformly a-continuous. It is then the derivative of a
a-H61der continuous function.
Remark. If p is the law of a singular continuous random variable X with
distribution function Fx, then Fx is a-Holder continuous if and only if p is
a-continuous. For general h, one calls F uniformly lip - h continuous [86].
Theorem 5.12.6 (Y. Last). If there exists C, such that ^Ylk=i lAfc|2 <
C • h(±) for all n > 0, then p is uniformly v^-continuous.
Proof. We follow [56]. The Dirichlet kernel satisfies
YI l^l2 = / / 2 Dn^y ~~x) M*W(y)

k=—n
and the Fejer kernel Kn(t) satisfies
1 /sin(^tr2
K n ( t ) ~n +<1 VKsin(t/2)
2 '
Ifcl >
= y (i _ JJ^Lykt
fc=—n
n + lJ
1^1 îfct
k=—n
Therefore
0 < ^ "Y \k\\pk\2 = f [(Dn(y - x) - Kn(y - x))dp(x)dp(y)

n+l /—' Jt7t
T \M2 -It Kniy - x)d»ix)dniy) . (5.4)

k——n
Because pn = /Ln, we can also sum from -n to n, changing only the

constant C. If /x is not uniformly y/h continuous, there exists a sequence
of intervals \Ik\ -> 0 with p(It) > ly/h(\Ii\). A property of the Fejer kernel
Kn(t) is that for large enough n, there exists S > 0 such that ^Kn(t) >
S > 0 if 1 < n\t\ < tt/2. Choose nu so that 1 < nt • \It\ < tt/2. Using
estimate (5.4), one gets
n* I - |2
E
> W)2><tt2MW)
> c-hi—).
This contradicts the existence of C such that
l-±\M*<Chil).
k=—n
□
Theorem 5.12.7 (Strichartz). Let p be a uniformly h-continuous measure

on the circle. There exists a constant C such that for all n
fc=l
Proof. The computation ([102, 103] for the Fourier transform was adapted
to Fourier series in [51]). In the following computation, we abbreviate dp(x)
with dx:
n-1 ,1 n~1 -{k+6/
a
' fc|2
fc=—n fc=-n
j n-1 iS±|li .
./o fc=—n
,._ . n Jv>
==3 e / / V] 1 dOdxdy
=4 e / / e ^ 2"2+i(x-y)e
Jt2 Jo
n-1 c_(-E±ft+<(x_y)t)a
N dOdxdy
k=—n
and continue
n-1
k=—n
n
=6 e / [ / — dt]e"^-^ ^ dxdy
Jt2 J-oo n
=7 e 0F / (e^x-y)2^)dxdy
Jt2
<8 e VtF(/ e-^-tf'tdxdy)1'2

Jt2
OO - 2
= 9 e^
> (E/ e-^-^^dxdy)1/2
^ A/n<|a;-»|<(fc+l)/n
oo
<io .cVJFCiMn-Êc-*2/2)1/2
fc=0
<n Ch(n~l) .
Here are some remarks about the steps done in this computation:
(1) is the trivial estimate
, n-1 -Lh±<pl
> 1
n
fc=—n
(2)
[ e-Hv-x)k dp(x)dp(y) = [ e~iyhdp(x) f eixkdp(x) = pkK = \M2

Jj2 Jt Jt
(3) uses Fubini's theorem.

(4) is a completion of the square.
(5) is the Cauchy-Schwartz inequality,
(6) replaces a sum and the integral /0 by j_oo,
(7) uses Ho e"(-+iry)t)2 dt = V*F because
'<*> e-(t/n+b)2
dt = \TK
/
for all n and complex 6,

(8) is Jensen's inequality.
(9) splits the integral over a sum of small intervals of strips of width 1/n.
(10) uses the assumption that p is h-continuous.
(11) This step uses that oo
(£e-f
k=0
c2/2)1/2
is a constant.
356 Chapter 5. Selected Tbpics
Bibliography
[I] N.I. Akhiezer. The classical moment problem and some related ques
tions in analysis. University Mathematical Monographs. Hafner pub
lishing company, New York, 1965.
[2] L. Arnold. Stochastische Differentialgleichungen. Oldenbourg Verlag,

Miinchen, Wien, 1973.
[3] S.K. Berberian. Measure and Integration. MacMillan Company, New

York, 1965.
[4] A. Choudhry. Symmetric Diophantine systems. Acta Arithmetica,

59:291-307, 1991.
[5] A. Choudhry. Symmetric Diophantine systems revisited. Acta Arith

metical 119:329-347, 2005.
[6] F.R.K. Chung. Spectral graph theory, volume 92 of CBMS Regional

Conference Series in Mathematics. AMS.
[7] J.B. Conway. A course in functional analysis, volume 96 of Graduate

texts in Mathematics. Springer-Verlag, Berlin, 1985.
[8] LP. Cornfeld, S.V.Fomin, and Ya.G.Sinai. Ergodic Theory, volume

115 of Grundlehren der mathematischen Wissenschaften in Einzel-
darstellungen. Springer Verlag, 1982.
[9] D.R. Cox and V. Isham. Point processes. Chapman & Hall, Lon
don and New York, 1980. Monographs on Applied Probability and
Statistics.
[10] R.E. Crandall. The challenge of large numbers. Scientific American,

(Feb), 1997.
[II] Daniel D. Stroock.
[12] P. Deift. Applications of a commutation formula. Duke Math. J.,

45(2):267-310, 1978.
[13] M. Denker, C. Grillenberger, and K. Sigmund. Ergodic Theory on

Compact Spaces. Lecture Notes in Mathematics 527. Springer, 1976.
358 Bibliography
[14] John Derbyshire. Prime obsession. Plume, New York, 2004. Bernhard
Riemann and the greatest unsolved problem in mathematics, Reprint
of the 2003 original [J. Henry Press, Washington, DC; MR1968857].
[15] L.E. Dickson. History of the theory of numbers. Vol.ILDiophantine

analysis. Chelsea Publishing Co., New York, 1966.
[16] S. Dineen. Probability Theory in Finance, A mathematical Guide to

the Black-Scholes Formula, volume 70 of Graduate Studies in Math
ematics. American Mathematical Society, 2005.
[17] J. Doob. Stochastic processes. Wiley series in probability and math

ematical statistics. Wiley, New York, 1953.
[18] J. Doob. Measure Theory. Graduate Texts in Mathematics. Springer

Verlag, 1994.
[19] P.G. Doyle and J.L. Snell. Random walks and electric networks, vol
ume 22 of Carus Mathematical Monographs. AMS, Washington, D.C.,
1984.
[20] T.P. Dreyer. Modelling with Ordinary Differential equations. CRC

Press, Boca Raton, 1993.
[21] R. Durrett. Probability: Theory and Examples. Duxburry Press, sec

ond edition edition, 1996.
[22] M. Eisen. Introduction to mathematical probability theory. Prentice-

Hall, Inc, 1969.
[23] N. Elkies. An application of Kloosterman sums.

http://www.math.harvard.edu/ elkies/M259.02/kloos.pdf, 2003.
[24] Noam Elkies. On a4 + b4 + c4 = d4. Math. Comput., 51:828-838,
1988.
[25] N. Etemadi. An elementary proof of the strong law of large numbers.

Z. Wahrsch. Verw. Gebiete, 55(1):119-122, 1981.
[26] W. Feller. An introduction to probability theory and its applications.

John Wiley and Sons, 1968.
[27] D. Freedman. Markov Chains. Springer Verlag, New York Heidelberg,

Berlin, 1983.
[28] E.M. Wright G.H. Hardy. An Introduction to the Theory of Numbers.

Oxford University Press, Oxford, fourth edition edition, 1959.
[29] J.E. Littlewood G.H. Hardy and G. Polya. Inequalities. Cambridge

at the University Press, 1959.
[30] R.T. Glassey. The Cauchy Problem in Kinetic Theory. SIAM,

Philadelphia, 1996.
Bibliography 359
[31] J. Glimm and A. Jaffe. Quantum physics, a functional point of view.

Springer Verlag, New York, second edition, 1987.
[32] G. Grimmet. Percolation. Springer Verlag, 1989.
[33] G. Grimmet and D.R. Stirzaker. Probability and Random Processes,

Problems and Solutions. Clarendon PRess, Oxford.
[34] A. Gut. Probability: A graduate Course. Springer texts in statistics.

Springer, 2005.
[35] Richard K. Guy. Unsolved Problems in Number Theory. Springer,

Berlin, 3 edition, 2004.
[36] P. Halmos. Lectures on ergodic theory. The mathematical society of

Japan, 1956.
[37] Paul R. Halmos. Measure Theory. Springer Verlag, New York, 1974.
[38] G.H. Hardy. Ramanujan. Cambridge at the University Press, 1940.

Twelve Lectures on Subjects by his Life and Work.
[39] W.K. Hayman. Subharmonic functions I,II, volume 20 of London

Mathematical Society Monographs. Academic Press, Inc. Harcourt
Brace Jovanovich, Publishers, London, 1989.
[40] H.G.Tucker. A graduate course in probability. Probability and Math

ematical Statistics. Academic Press, 1967.
[41] O. Gurel-Gurevich I. Benjamini and B. Solomyak. Branching ran

dom walk with exponentially decreasing steps and stochastically self-
similar measures. arXiv PR/0608271, 2006.
[42] R. Isaac. The Pleasures of Probability. Graduate Texts in Mathemat

ics. Springer Verlag, 1995.
[43] K. Ito and H.P. McKean. Diffusion processes and their sample paths,
volume 125 of Die Grundlehren der mathematischen Wissenschaften.
Springer-Verlag, Berlin, second printing edition, 1974.
[44] V. Kac and P. Cheung. Quantum calculus. Universitext. Springer-

Verlag, New York, 2002.
[45] J-P. Kahane and R. Salem. Ensembles parfaits et series

trigonometriques. Hermann, 1963.
[46] I. Karatzas and S. Shreve. Brownian motion and stochastic calculus,

volume 113 of Graduate Texts in Mathematics. Springer-Verlag, New
York, second edition, 1991.
[47] A.F. Karr. Probability. Springer texts in statistics. Springer-Verlag,

1993.
360 Bibliography
[48] Y. Katznelson. An introduction to harmonic analysis. Dover publi

cations, Inc, New York, second corrected edition edition, 1968.
[49] J.F.C. Kingman. Poisson processes, volume 3 of Oxford studies in

probability. Clarendon Press, New York: Oxford University Press,
1993.
[50] O. Knill. A remark on quantum dynamics. Helvetica Physica Acta,

71:233-241, 1998.
[51] O. Knill. Singular continuous spectrum and quantitative rates of

weakly mixing. Discrete and continuous dynamical systems, 4:33-42,
1998.
[52] B. Reznick K.O. Bryant and M. Serbinowska. Almost alternating

sums. Amer. Math. Monthly.
[53] N. Kolmogorov. Grundbegriffe der Wahrscheinlichkeitsrechnung.

Berlin, 1933. English: Foundations of Probability Theory, Chelsea,
New York, 1950., 1933.
[54] U. Krengel. Ergodic Theorems, volume 6 of De Gruyter Studies in

Mathematics. Walter de Gruyter, Berlin, 1985.
[55] H. Kunita. Stochastic flows and stochastic differential equations.

Cambridge University Press, 1990.
[56] Y. Last. Quantum dynamics and decompositions of singular contin

uous spectra. J. Func. Anal, 142:406-445, 1996.
[57] J. Lewis. An elementary approach to Brownian motion on manifolds.

In Stochastic processes—mathematics and physics (Bielefeld, 1984),
volume 1158 of Lecture Notes in Math., pages 158-167. Springer,
Berlin, 1986.
[58] E.H. Lieb and M. Loss. Analysis, volume 14 of Graduate Studies in

Mathematics. American Mathematical Society, 1996.
[59] J.L. Selfridge L.J. Lander, T.R. Parken. A survey of equal sums of
like powers. Mathematics of Computation, (99):446-459, 1967.
[60] E. Lukacs and R.G.Laha. Applications of Characteristic Functions.

Griffin's Statistical Monogrpaphs and Courses.
[61] M.C. Mackey. Time's arrow: the origins of thermodynamics behavior.

Springer-Verlag, New York, 1992.
[62] N. Madras and G. Slade. The self-avoiding random walk. Probability

and its applications. Birkh 1993.
[63] L. Malozemov and A. Teplyaev. Pure point spectrum of the Lapla-

cians on fractal graphs. J. Fund. Anal, 129(2):390-405, 1995.
Bibliography 361
[64] K.V. Mardia. Statistics of directional data. Academic press, London
and New York, 1972.
[65] A. Matulich and B.N. Miller. Gravity in one dimension: stability of

a three particle system. Colloq. Math., 39:191-198, 1986.
[66] H.P. McKean. Stochastic integrals. Academic Press, 1969.
[67] L. Merel. Bornes pour la torson des courbes elliptiques sur les corps
de nombres. Inv. Math., 124:437-449, 1996.
[68] Jean-Charles Meyrignac. Existence of solutions of (n,n+ l,n+l).

http://euler.free.fr/theorem.htm.
[69] Jean-Charles Meyrignac. Records of equal sums of like powers.

http://euler.free.fr/records.htm.
[70] F. Mosteller. Fifty Challenging Problems in Probability with solutions.

Dover Publications, inc, New York, 1965.
[71] M. Nagasawa. Schrodinger equations and diffusion theory, volume 86

of Monographs in Mathematics. Birkhauser Verlag, Basel, 1993.
[72] LP. Natanson. Constructive theory of functions. Translation se

ries. United states atomic energy commission, 1949. State Publishing
House of Technical-Theoretical Literature.
[73] E. Nelson. Dynamical theories of Brownian motion. Princeton uni

versity press, 1967.
[74] E. Nelson. Radically elementary probability theory. Princeton univer

sity text, 1987.
[75] T.Lewis N.I. Fisher and B.J. Embleton. Statistical analysis of spher
ical data. Cambridge University Press, 1987.
[76] B. Oksendal. Stochastic Differential Equations. Universitext.

Springer-Verlag, New York, fourth edition edition, 1995.
[77] K. Petersen. Ergodic theory. Cambridge University Press, Cambridge,

1983.
[78] I. Peterson. The Jungles of Randomness, A mathematical Safari.

John Wiley and Sons, Inc.
[79] S. C. Port and C.J. Stone. Brownian motion and classical poten
tial theory. Probability and Mathematical Statistics. Academic Press
(Harcourt Brace Jovanovich Publishers),New York, 1978.
[80] T. Ransford. Potential theory in the complex plane, volume 28 of

London Mathematical Society Student Texts. Cambridge University
Press, Cambridge, 1995.
362 Bibliography
[81] M. Reed and B. Simon. Methods of modern mathematical physics ,

Volume I. Academic Press, Orlando, 1980.
[82] C.J. Reidl and B.N. Miller. Gravity in one dimension: The critical
population. Phys. Rev. E, 48:4250-4256, 1993.
[83] D. Revuz and M.Yor. Continuous Martingales and Brownian Mo

tion. Springer Verlag, 1991. Grundlehren der mathmatischen Wis-
senschaften, 293.
[84] H. Riesel. Prime numbers and computer methods for factorization,

volume 57 of Progress in Mathematics. Birkhauser Boston Inc., 1985.
[85] RE. Hart R.O. Duda and D.G. Stork. Pattern Classification. John
Wiley and Sons, Inc, New York, second edition edition.
[86] C.A. Rogers. Hausdorff measures. Cambridge University Press, 1970.
[87] S.M. Ross. Applied Probability Models with optimization Applications.

Dover Publications, inc, New York, 1970.
[88] W. Rudin. Real and Complex Anslysis. McGraw-Hill Series in Higher

Mathematics, 1987.
[89] D. Ruelle. Chance and Chaos. Princeton Science Library. Princeton

University Press, 1991.
[90] G.B. Rybicki. Exact statistical mechanics of a one-dimensional

self-gravitating system. In M. Lecar, editor, Gravitational N-Body
Problem, pages 194-210. D. Reidel Publishing Company, Dordrecht-
Holland, 1972.
[91] A. Shiryayev. Probability, volume 95 of Graduate Texts in Mathemat

ics. Springer-Verlag, New York, 1984.
[92] Hwei P. Shu. Probability, Random variables and Random Processes.

Schaum's Outlines. McGraw-Hill, 1997.
[93] B. Simon. Functional Integration and Quantum Physics. Academic

Press, 1979. Pure and applied mathematics.
[94] B. Simon. Spectral analysis of rank one perturbations and appli

cations. In Mathematical quantum theory. II. Schrodinger operators
(Vancouver, BC, 1993), volume 8 of CRM Proc. Lecture Notes, pages
109-149. AMS, Providence, Rl, 1995.
[95] B. Simon. Operators with singular continuous spectrum. VI. Graph

Laplacians and Laplace-Beltrami operators. Proc. Amer. Math. Soc,
124(4):1177-1182, 1996.
[96] B. Simon and T. Wolff. Singular continuous spectrum under rank one
perturbations and localization for random Hamiltonians. Commun.
Pure Appl. Math., 39:75.
Bibliography 363
[97] Ya. G. Sinai. Probability Theory, An Introductory Course. Springer
Textbook. Springer Verlag, Berlin, 1992.
[98] J.L. Snell and R. Vanderbei. Three bewitching paradoxes. Probability

and Stochastics Series. CRC Press, Boca Raton, 1995.
[99] P.M. Soardi. Potential theory on infinite networks, volume 1590 of

Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1994.
[100] F. Spitzer. Principles of Random walk. Graduate texts in mathemat

ics. Springer-Verlag, New York Heidelberg Berlin, 1976.
[101] H. Spohn. Large scale dynamics of interacting particles. Texts and

monographs in physics. Springer-Verlag, New York, 1991.
[102] R.S. Strichartz. Fourier asymptotics of fractal measures. J. Func.

Anal, 89:154-187, 1990.
[103] R.S. Strichartz. Self-similarity in harmonic analysis. The journal of

Fourier analysis and Applications, 1:1-37, 1994.
[104] D.W. Stroock. Lectures on Stochastic Analysis and Diffusion Theory.

London Mathematical Society Students Texts. Cambridge University
Press, 1987.
[105] D.W. Stroock. Probability theory, an analytic view. Cambridge Uni

versity Press, 1993.
[106] N. van Kampen. Stochastic processes in physics and chemistry.

North-Holland Personal Library, 1992.
[107] P. Walters. An introduction to ergodic theory. Graduate texts in

mathematics 79. Springer-Verlag, New York, 1982.
[108] D. Williams. Probability with Martingales. Cambridge mathematical

Texbooks, 1991.
[109] G.L. Wise and E.B. Hall. Counterexamples in probability and real
analysis. Oxford University Press, 1993.
Index
IP, 40 atomic random variable, 80

P-independent, 30 automorphism probability space,
P-trivial, 34 68
A-set, 31 axiom of choice, 15
7r-system, 29
a ring, 34 Ballot theorem, 169
cr-additivity, 25 Banach space, 46
a-algebra, 23 Banach-Tarski paradox, 15
a-algebra Bayes rule, 27
P-trivial, 29 Benford's law, 322
Borel, 24 Beppo-Levi theorem, 43
cr-algebra generated subsets, 24 Bernoulli convolution, 115
cr-ring, 33 Bernstein polynomial, 308
Bernstein polynomials, 52
absolutely continuous, 123 Bernstein-Green-Kruskal modes, 305
absolutely continuous distribution Bertrand, 13
function, 80 beta distribution, 82
absolutely continuous measure, 34 Beta function, 81
algebra, 31 Bethe lattice, 174
algebra BGK modes, 305
a,23
bias, 293
Borel, 24 Binomial coefficient, 83
generated by a map, 26 binomial distribution, 83
tail, 34 Birkhoff's ergodic theorem, 19
algebra Birkhoffs ergodic theorem, 70
trivial, 23
birthday paradox, 21
algebra of random variables, 40 Black and Scholes, 249, 266
algebraic ring, 34
almost alternating sums, 167 Black-Jack, 138
almost everywhere statement, 39 blackbody radiation, 117
almost Mathieu operator, 19 Blumental's zero-one law, 223
Boltzmann distribution, 103
angle, 50
arc-sin law, 170 Boltzmann-Gibbs entropy, 98
arithmetic random variable, 334 bond of a graph, 275
Aronzajn-Krein formula, 287 bonds, 161
Borel cr-algebra, 24
asymptotic expectation, 335
Borel set, 24
asymptotic variance, 335
Borel transform, 286
atom, 24, 81
atomic distribution function, 80 Borel-Cantelli lemma, 34, 36
364
Index 365
bounded dominated convergence, Chapmann-Kolmogorov equation,

44 186
bounded process, 136 characteristic function, 111
bounded stochastic process, 144 characteristic function
bounded variation, 255 Cantor distribution, 115
boy-girl problem, 28 characteristic functional
bracket process, 260 point process, 314
characteristic functions
branching process, 135, 188
branching random walk, 115, 145 examples, 112
Brownian bridge, 204 Chebychev inequality, 48
Brownian motion, 192 Chebychev-Markov inequality, 47
Brownian motion Chernoff bound, 48
existence, 196 Choquet simplex, 319
circle-valued random variable, 319
geometry, 266
on a lattice, 217 circular variance, 320
classical mechanics, 19
strong law , 199
Brownian sheet, 205 coarse grained entropy, 99
Buffon needle problem, 21 compact set, 24
complete, 272
complete convergence, 60
calculus of variations, 103
completion of cr-algebra, 24
Campbell's theorem, 314 concentration parameter, 322
canonical ensemble, 106 conditional entropy, 100
canonical version of a process, 206 conditional expectation, 124
Cantor distribution, 84 conditional integral, 9, 125
Cantor distribution conditional probability, 26
characteristic function, 115 conditional probability space, 128
Cantor function, 85 conditional variance, 128
Cantor set, 115 cone, 107
capacity, 225 cone of martingales, 135
Caratheodory lemma, 32 continuation theorem, 31
casino, 14 continuity module, 53
Cauchy distribution, 82 continuity points, 91
Cauchy sequence, 272 continuous random variable, 80
Cauchy-Bunyakowsky-Schwarz in contraction, 272
equality, 46 convergence
Cauchy-Picard existence, 271 in distribution, 60, 91
Cauchy-Schwarz inequality^ 46 in law, 60, 91
Cayley graph, 165, 175 in probability, 51
Cayley transform, 181 stochastic, 51
CDF, 56 weak, 91
celestial mechanics, 19 convergence almost everywhere, 59
centered Gaussian process, 192 convergence almost sure, 59
centered random variable, 42 convergence complete, 60
central limit theorem, 120 convergence fast in probability, 60
central limit theorem convergence in Cp, 59
circular random vectors, 326 convergence in probability, 59
central moment, 40 convex function, 44
366 Index
convolution, 352 discrete stochastic process, 131
convolution discrete Wiener space, 185
random variable, 112 discretized stopping time, 222
coordinate process, 206 distribution
correlation coefficient, 49 beta, 82
covariance, 48 binomial, 83
Covariance matrix, 192 Cantor, 84
crushed ice, 241 Cauchy, 82
cumulant generating function, 41 Diophantine equation, 346
cumulative density function, 56 Erlang, 88
cylinder set, 38 exponential, 82
first success, 83
de Moivre-Laplace, 96 Gamma, 82, 88
decimal expansion, 65 geometric, 83
decomposition of Doob, 152 log normal, 41, 82
decomposition of Doob-Meyer, 153 normal, 81
density function, 57 Poisson , 83
density of states, 178 uniform, 82, 83
dependent percolation, 17 distribution function, 57
dependent random walk, 166 distribution function
derivative of Radon-Nykodym, 123 absolutely continuous, 80
Devil staircase, 352 discrete, 80
devils staircase, 85 singular continuous, 80
dice distribution of the first significant
fair, 26 digit, 322
non-transitive, 27 dominated convergence theorem,
Sicherman, 27 44
differential equation Doob convergence theorem, 154
solution, 271 Doob submartingale inequality, 157
stochastic, 265 Doob's convergence theorem, 144
dihedral group, 176 Doob's decomposition, 152
Diophantine equation, 343 Doob's up-crossing inequality, 143
Diophantine equation Doob-Meyer decomposition, 153,
Euler, 343 257
symmetric, 343 dot product, 50
Dirac point measure, 309 downward filtration, 151
directed graph, 185 dyadic numbers, 196
Dirichlet kernel, 352 Dynkin system, 29
Dirichlet problem, 214 Dynkin-Hunt theorem, 222
Dirichlet problem
discrete, 182 economics, 14
Kakutani solution, 214 edge of a graph, 275
discrete Dirichlet problem, 182 edge which is pivotal, 281
discrete distribution function, 80 Einstein, 194
discrete Laplacian, 182 electron in crystal, 18
discrete random variable, 80 elementary function, 40
discrete Schrodinger operator, 286 elementary Markov property, 187
discrete stochastic integral, 136 Elkies example, 349
Index 367
ensemble finite measure, 34

micro-canonical, 102 finite quadratic variation, 257
entropy finite total variation, 255
Boltzmann-Gibbs, 98 first entry time, 137
circle valued random variable, first significant digit, 322
320 first success distribution, 83
coarse grained, 99 Fisher information, 295
distribution, 98 Fisher information matrix, 295
FKG inequality, 279, 280
geometric distribution, 99
Kolmogorov-Sinai, 99 form, 343
measure preserving transfor formula
mation, 99 Aronzajn-Krein, 287
normal distribution, 99 Feynman-Kac, 180
partition, 99 Levy, 111
random variable, 89 formula of Russo, 281
equilibrium measure, 171, 225 Formula of Simon-Wolff, 289
equilibrium measure Fortuin Kasteleyn and Ginibre, 279
existence, 226 Fourier series, 69, 303, 351
Vlasov dynamics, 305 Fourier transform, 111
ergodic theorem, 70 Frohlich-Spencer theorem, 286
ergodic theorem of Hopf, 69 free energy, 117
ergodic theory, 35 free group, 172
ergodic transformation, 68 free Laplacian, 176
Erlang distribution, 88 function
estimator, 292 characteristic, 111
Euler Diophantine equation, 343, convex, 44
348 distribution, 57
Euler's golden key, 343 Rademacher, 42
Eulers golden key, 39 functional derivative, 103
expectation, 40
expectation gamblers ruin probability, 142
E[X;A], 54 gamblers ruin problem, 141
conditional, 124 game which is fair, 137
exponential distribution, 82 Gamma distribution, 82, 88
extended Wiener measure, 233 Gamma function, 81
extinction probability, 150 Gaussian
distribution, 41
Fejer kernel, 350, 353 process, 192
factorial, 83 random vector, 192
factorization of numbers, 21 vector valued random variable,
fair game, 137 116
Fatou lemma, 43 generalized function, 10
Fermat theorem, 349 generalized Ornstein-Uhlenbeck pro
Feynman-Kac formula, 180, 217 cess, 203
Feynman-Kac in discrete case, 180 generating function, 116
filtered space, 131, 209 generating function
filtration, 131, 209 moment, 87
finite Markov chain, 317 geometric Brownian motion, 266
368 Index
geometric distribution, 83 Kolmogorov, 72
geometric series, 87 inequality
Gibbs potential, 117
Cauchy-Schwarz, 46
global error, 293 Chebychev, 48
global expectation, 292 Chebychev-Markov, 47
Golden key, 343 Doob's up-crossing, 143
golden ratio, 167 Fisher , 297
great disorder, 205 FKG, 279
Green domain, 213 Holder, 46
Green function, 213, 303 Jensen, 45
Green function of Laplacian, 286 Jensen for operators, 108
group-valued random variable, 327 Minkowski, 46
power entropy, 297
Holder continuous, 199, 352
submartingale, 218
Holder inequality, 46
inequality Kolmogorov, 157
Haar function, 197
inequality of Kunita-Watanabe, 260
Hahn decomposition, 34, 124 information inequalities, 297
Hamilton-Jacobi equation, 303 inner product, 46
Hamlet, 37 integrable, 40
Hardy-Ramanujan number, 344 integrable
harmonic function on finite graph,
uniformly, 54, 63
185 integral, 40
harmonic series, 36 integral of Ito , 263
heat equation, 252 integrated density of states, 178
heat flow, 215 invertible transformation, 68
Helly's selection theorem, 91 iterated logarithm
Helmholtz free energy, 117 law, 118
Hermite polynomial, 252 Ito integral, 247, 263
Hilbert space, 46 Ito's formula, 249
homogeneous space, 327
Jacobi matrix, 179, 286
identically distributed, 57 Javrjan-Kotani formula, 287 ,
identity of Wald, 143 Jensen inequality, 45
IID, 57 Jensen inequality for operators, 108
IID random diffeomorphism, 318
jointly Gaussian, 192
IID random diffeomorphism
continuous, 318 K-system, 35
smooth, 318 Keynes postulates, 27
increasing process, 152, 256 Kingmann subadditive ergodic the
independent orem, 147
7r-system, 30 Kintchine's law of the iterated log
independent events, 28 arithm, 219
independent identically distributed, Kolmogorov 0-1 law, 34
57 Kolmogorov axioms, 25
independent random variable, 28 Kolmogorov inequalities, 72
independent subalgebra, 28 Kolmogorov inequality, 157
indistinguishable process, 198 Kolmogorov theorem, 74
inequalities Kolmogorov theorem
Index 369
conditional expectation, 124 Fatou, 43

Kolmogorov zero-one law, 34 Riemann-Lebesgue, 349
Komatsu lemma, 141 Komatsu, 141
Koopman operator, 108 length, 50
Kronecker lemma, 156 lexicographical ordering, 62
Kullback-Leibler divergence, 100 likelihood coefficient, 100
Kunita-Watanabe inequality, 260 limit theorem
de Moive-Laplace, 96
Levy theorem, 76 Poisson, 96
Levy formula, 111 linear estimator, 293
Lander notation, 349 linearized Vlasov flow, 304
Langevin equation, 267 lip-h continuous, 352
Laplace transform, 116 Lipshitz continuous, 199
Laplace-Beltrami operator, 303 locally Holder continuous, 199
Laplacian, 176 log normal distribution, 41, 82
Laplacian on Bethe lattice, 174 logistic map, 19
last exit time, 138
Last's theorem, 353 Mobius function, 343
lattice animal, 275 Marilyn vos Savant, 15
lattice distribution, 320 Markov chain, 187
law Markov operator, 107, 318
group valued random variable, Markov process, 187
327 Markov process existence, 187
iterated logarithm, 118 Markov property, 187
random vector, 300, 306 martingale, 131, 215
symmetric Diophantine equa martingale inequality, 159
tion, 345 martingale strategy, 14
uniformly h-continuous, 352 martingale transform, 136
law of a random variable, 56 martingale, etymology, 132
law of arc-sin, 170 matrix cocycle, 317
law of group valued random vari maximal ergodic theorem of Hopf,
able, 320 69
law of iterated logarithm, 157 maximum likelihood estimator, 294
law of large numbers, 51 Maxwell distribution, 59
law of large numbers Maxwell-Boltzmann distribution,
strong, 65, 71 103
weak, 53 mean direction, 320, 322
law of total variance, 129 mean measure
Lebesgue decomposition theorem, Poisson process, 312
81 mean size, 278
Lebesgue dominated convergence, mean size
44 open cluster, 283
Lebesgue integral, 7 mean square error, 294
Lebesgue measurable, 24 mean vector, 192
Lebesgue thorn, 214 measurable
lemma progressively, 211
Borel-Cantelli, 34, 36 measurable map, 25, 26
Caratheodory, 32 measurable space, 23
370 Index
measure, 29 number of open clusters, 284
measure , finite34
absolutely continuous, 98 operator
algebra, 31 Koopman, 108
equilibrium, 225 Markov, 107
outer, 32 Perron-Frobenius, 108
positive, 34 Schrodinger, 18
push-forward, 56, 205 symmetric, 231
uniformly h-continuous, 352 Ornstein-Uhlenbeck process, 202
Wiener , 206 oscillator, 235
measure preserving transformation, outer measure, 32
68
median, 76 paradox
Mehler formula, 239 Bertrand, 13
Mertens conjecture, 343 Petersburg, 14
metric space, 272 three door , 15
micro-canonical ensemble, 102,106 partial exponential function, 111
minimal filtration, 210 partition, 27
Minkowski inequality, 46 path integral, 180
Minkowski theorem, 332 percentage drift, 266
Mises distribution, 322 percentage volatility, 266
moment, 40 percolation, 16
moment percolation
formula, 87 bond, 16
generating function, 41, 87, cluster, 16
116 dependent, 17
measure, 306 percolation probability, 276
random vector, 306 perpendicular, 50
moments, 87 Perron-Frobenius operator, 108
monkey, 37 perturbation of rank one, 287
monkey typing Shakespeare, 37 Petersburg paradox, 14
Monte Carlo Picard iteration, 268
integral, 7 pigeon hole principle, 344
Monte Carlo method, 21 pivotal edge, 281
Multidimensional Bernstein theo Planck constant, 117, 179
rem, 308 point process, 312
multivariate distribution function, Poisson distribution, 83
306 Poisson equation, 213, 303
Poisson limit theorem, 96
neighboring points, 275 Poisson process, 216, 312
net winning, 137 Poisson process
normal distribution, 41, 81, 98 existence, 312
normal number, 65 Pollard p method, 21
normality of numbers, 65 Pollare p method, 340
normalized random variable, 42 Polya theorem
nowhere differentiable, 200 random walks, 163
NP complete, 341 Polya urn scheme, 134
null at 0, 132 portfolio, 161
Index 371
position operator, 252 integrable, 40

positive cone, 107 normalized, 42
positive measure, 34 singular continuous, 80
positive semidefinite, 201 spherical, 327
postulates of Keynes, 27 symmetric, 117
power distribution, 57 uniformly integrable, 63
previsible process, 136 random variable independent, 28
prime number theorem, 343 random vector, 26
probability random walk, 16, 162
P[A > c] , 47 rank one perturbation, 287
conditional, 26 Rao-Cramer bound, 297
probability density function, 57 Rao-Cramer inequality, 296
probability generating function, 87 Rayleigh distribution, 59
probability space, 25 reflected Brownian motion, 215
process reflection principle, 167
bounded, 136 regression line, 48
finite variation, 256 regular conditional probability, 128
increasing, 256 relative entropy, 100
previsible, 136 relative entropy
process indistinguishable, 198 circle valued random variable,
progressively measurable, 211 320
pseudo random number generator, resultant length, 320
21, 340 Riemann hypothesis, 343
pull back set, 25 / Riemann integral, 7
pure point spectrum, 286 / Riemann zeta function, 39, 343
push-forward measure, 56, 205, 300 Riemann-Lebesgue lemma, 349
Pythagoras theorem, 50 / / right continuous filtration^ 210
Pythagorean triples, 343, 349 ring, 33, 34
risk function, 294
quantum mechanical oscillator, 235 ruin probability, 142
ruin problem, 141
Renyi's theorem, 315 Russo's formula, 281
Rademacher function, 42
Radon-Nykodym theorem, 123 Schrodinger equation, 179
random circle map, 316 Schrodinger operator, 18
random diffeomorphism, 316 score function, 296
random field, 205 SDE, 265
random number generator, 57 semimartingale, 131
random variable, 26 set of continuity points, 91
random variable Shakespeare, 37
LP, 44 Shannon entropy , 89
absolutely continuous, 80 significant digit, 319
arithmetic, $34 silver ratio, 167
centered, 42 Simon-Wolff criterion, 289
circle valued, 319 singular continuous distribution,
continuous, 57, 80 80
discrete, 80 singular continuous random vari
group valued, 327 able, 80
372 Index
solution subgraph, 275
differential equation, 271 submartingale, 131, 215
spectral measure, 178 submartingale inequality, 157
spectrum, 18 submartingale inequality
spherical random variable, 327 continuous martingales, 218
Spitzer theorem, 244 sum circular random variables, 324
stake of game, 137 sum of random variables, 69
Standard Brownian motion, 192 super-symmetry, 238
standard deviation, 40 supercritical phase, 278
standard normal distribution, 92, supermartingale, 131, 215
98 support of a measure, 318
state space, 186 symmetric Diophantine equation,
statement 343
almost everywhere, 39 symmetric operator, 231
stationary measure, 188 symmetric random variable, 117
stationary measure symmetric random walk, 175
discrete Markov process, 318 systematic error, 293
ergodic, 318
random map, 318 tail cr-algebra, 34
stationary state, 110 taxi-cab number, 344
statistical model, 292 taxicab numbers, 349
step function, 40 theorem
Stirling formula, 170 Ballot, 169
stochastic convergence, 51 Banach-Alaoglu, 90
stochastic differential equation, 265 Beppo-Levi, 43
stochastic differential equation Birkhoff, 19
existence, 269 Birkhoff ergodic, 70
stochastic matrix, 185, 188 bounded dominated conver
stochastic operator, 107 gence, 44
stochastic population model, 266 Caratheodory continuation, 31
stochastic process, 191 central limit, 120
stochastic process dominated convergence, 44
discrete, 131 Doob convergence, 154
stocks, 161 Doob's convergence, 144
stopping time, 137, 210 Dynkin-Hunt, 222
stopping time for random walk, Helly, 91
167 Kolmogorov, 74
Strichartz theorem, 354 Kolmogorov's 0 — 1 law, 34
strong convergence Levy, 76
operators, 231 Last, 353
strong law Lebesgue decomposition, 81
Brownian motion, 199 martingale convergence, 154
strong law of large numbers for maximal ergodic theorem, 69
Brownian motion, 199 Minkowski, 332
sub-critical phase, 278 monotone convergence, 43
subadditive, 147 Polya, 163
subalgebra, 24 Pythagoras, 50
subalgebra independent, 28 Radon-Nykodym, 123
Index 373
Strichartz, 354 Vitali, Giuseppe, 15

three series , 75 Vlasov flow, 298
Tychonov, 90 Vlasov flow Hamiltonian, 298
Voigt, 108 von Mises distribution, 322
Weierstrass, 52
Wiener, 351, 352 Wald identity, 143
Wroblewski, 344 weak convergence
thermodynamic equilibrium, 110 by characteristic functions, 112
for measures, 225
thermodynamic equilibrium mea
measure , 90
sure, 171
three door problem, 15 random variable, 91
three series theorem, 75 weak law of large numbers, 51, 53
tied down process, 203 weak law of large numbers for L1,
54
topological group, 327
total variance Weierstrass theorem, 52
law, 129 Weyl formula, 241
transfer operator, 318 white noise, 10, 205
transform Wick ordering, 252
Wick power, 252
Fourier, 111
Laplace, 116 Wick, Gian-Carlo, 252
Wiener measure, 206
martingale, 136
transition probability function, 186 Wiener sausage, 242
Wiener space, 206
tree, 172
trivial cr-algebra, 34 Wiener theorem, 351
trivial algebra, 23, 26 Wiener, Norbert, 197
Wieners theorem, 352
Tychonovs theorem, 90
wrapped normal distribution, 320,
323
uncorrelated, 48, 49
uniform distribution, 82, 83 Wroblewski theorem, 344
uniform distribution
zero-one law of Blumental, 223
circle valued random variable,
zero-one law of Kolmogorov, 34
323
Zeta function, 343
uniformly h-continuous measure,
352
uniformly integrable, 54, 63
up-crossing, 143
up-crossing inequality, 143
urn scheme, 134
utility function, 14
variance, 40
variance
Cantor distribution, 129
conditional, 128
variation
stochastic process, 256
vector valued random variable, 26
vertex of a graph, 275
Probability Theory and Stochastic Processes with Applications
About the Book-
Chapter 1-2 of this text covers material of a basic probability course. Chapter 3 deals with discrete
stochastic processes including Martingale theory. Chapter 4 covers continuous time stochastic
processes like Brownian motion and stochastic differential equations. The last chapter "selected
topics" got considerably extended in the summer of 2006. In the original course, only localization
and percolation problems were included. Now, more topics like estimation theory, Vlasov
dynamics, multi-dimensional moment problems, random maps, circle-valued random variables,
the geometry of numbers, Diophantine equations and harmonic analysis have been added. No
previous knowledge in probability are necessary, but a basic exposure to calculus and linear algebra
is required. Some real analysis as well as some background in topology, functional analysis and
harmonic analysis can be helpful for the later chapters.
About the Author-

Oliver Knill grew up near Schaffhausen in Switzerland. He got
his BA in Mathematics at ETH Zurich with a senior theses in
dynamical systems advised by Jiirgen Moser. After a
postgraduate visit at Technion in Haifa he wrote his PhD theses
at the ETH Zurich advised by Oscar Lanford III in the area of
dynamical systems and random operator theory. After three
years in Pasadena at Caltech as a Taussky-Todd instructor in the
group of Barry Simon, and one year at the University of Arizona
in Tucson as a Hanno-Rund visiting assistant professor, he was
a Swiss National Science foundation fellow at the University of
Texas in Austin and worked a year as a research assistant at the
Institute for Fusion Studies at UT. Since Fall 2000, he teaches at
the Mathematics department at Harvard University.
Oliver Knill started his research in the field of dynamical systems, tackling ergodic and spectral
theoretical questions. He worked out and published Moser's lecture notes "Selected chapters in the
calculus of variations" which appeared in 2003 in the series Lectures in Mathematics ETH Zurich.
More recently he works on applications of dynamical systems to probability theory and elementary
number theory as well as inverse problem in analysis, geometry, and computer vision.
Oliver Knill, Department of Mathematics,

One Oxford Street, Harvard University,
Cambridge, MA 02138, USA
Visit Overseas Press on the World Wide Web at: www.overseaspub.com.
ISBN 81-89938-40-I
OVERSEAS PRESS (INDIA) PRIVATE LIMITEti

7/28, Ansari Road, Daryaganj, New Delhi-110002 78 8 189 93 8406

Probability Theory and Stochastic Processes With Applications

Uploaded by

Copyright:

Available Formats

Probability Theory and Stochastic Processes With Applications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Probability Theory and Stochastic Processes With Applications

Uploaded by

Copyright:

Available Formats

Probability Theory

OVERSEAS PRESS (INDIA) PVT. LTD.

All rights reserved. No part of this publication may be reproduced or transmitted

10 digit ISBN : 81 - 89938 - 40 -1

3 Discrete Stochastic Processes 123

Continuous Stochastic Processes 191

These notes grew from an introduction to probability theory taught during

Most of Chapter 2 is standard material and subject of virtually any course

While the text assumes no prerequisites in probability, a basic exposure to

Probability theory can be developed using nonstandard analysis on finite

We live in a time, in which more and more content is available online.

1.1 What is probability theory?

A random variable X can be described well by its distribution function

Inequalities play an important role in probability theory. The Chebychev

Independence is an central notion in probability theory. Two events A, B

Independence can also be explained using conditional probability with re

where X(x,y) is a continuous function on the unit square with P = dxdy

A set {Xt}teT of random variables defines a stochastic process. The vari

As in the deterministic case, where the theory of differential equations is

Some features of stochastic process can be described using the language of

1.2 Some paradoxes in probability theory

1) Bertrand's paradox (Bertrand 1889)

Figure. Random an- Figure. Random Figure. Random area.

Like most paradoxes in mathematics, a part of the question in Bertrand's

2) Petersburg paradox (D.Bernoulli, 1738)

£2fcP[T = fc] = ]Tl = oo.

(E[VG})2 = £>fc/22-fc)2 = ^ ~ 5.828... .

Figure. The picture to the right

4) The Banach-Tarski paradox (1924)

l=P[T]=P[(J^] = X>[^] = £p.

1) Random walks: (statistical mechanics, gambling, stock markets, quan

2) Percolation problems (model of a porous medium, statistical mechanics,

Figure. Sond percola Figure. Bond percola Figure. Bond percola

A variant of bond percolation is site percolation where the nodes of the

Figure. Site percola Figure. Site percola Figure. Site percola

Generalized percolation problems are obtained, when the independence

Xn m — l(na+m/3) mod l€[0,p)

are no more independent.

Figure. Dependent Figure. Dependent Figure. Dependent

3) Random Schrodinger operators, (quantum mechanics, functional analy

Figure. A wave Figure. wave

More general operators are obtained by allowing V(n) to be random vari

4) Classical dynamical systems (celestial mechanics, fluid dynamics, me

The study of deterministic dynamical systems like the logistic map x h^

Figure. Iterating the Figure. The simple Figure. A short time

Given a dynamical system given by a map T or a flow Tt on a subset Q of

5) Cryptology. (computer science, coding theory, data encryption)

This probabilistic argument would give a rigorous probabilistic estimate

Probabilistic thinking is often involved in designing, investigating and at

6) Numerical methods, (integration, Monte Carlo experiments, algorithms)

[0,7r]. The probability of hitting a line is therefore J* 2 arccos(2/)/7r = 2/ir.

Figure. The Buff on needle prob

2.1 Probability spaces, random variables, indepen

Definition. A set A of subsets of ft is called a cr-algebra if the following

Properties. If A is a cr-algebra, and An is a sequence in A, then the fol

2) limsupn An := fl~i Um=„ A, € A

Example. For an arbitrary set fi, >t = {0,0}) is a cr-algebra. It is called

Example. If fi is an arbitrary set, then A = {A C Cl}) is a cr-algebra. The

Example. A finite set of subsets Au A2,..., An of ft which are pairwise