This action might not be possible to undo. Are you sure you want to continue?
Markov Chains
J. R. Norris
University of Cambridge
~ u ~ ~ u CAMBRIDGE
::: UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE
The Pitt Building, Trumpington Street, Cambridge CB2 lRP, United Kingdom
CAMBRIDGE UNIVERSITY PRESS
The Edinburgh Building, Cambridge CB2 2RU, United Kingdom
40 West 20th Street, New York, NY 100114211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia
© Cambridge University Press 1997
This book is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.
First published 1997
Reprinted 1998
First paperback edition 1998
Printed in the United States of America
TYPeset in Computer Modem
A catalogue recordfor this book is available from the British Library
Library of Congress CataloguinginPublication Data is available
ISBN 0521481813 hardback
ISBN 0521633966 paperback
For my parents
Contents
Preface
Introduction
1. Discretetime Markov chains
1.1 Definition and basic properties
1.2 Class structure
1.3 Hitting times and absorption probabilities
1.4 Strong Markov property
1.5 Recurrence and transience
1.6 Recurrence and transience of random walks
1.7 Invariant distributions
1.8 Convergence to equilibrium
1.9 Time reversal
1.10 Ergodic theorem
1.11 Appendix: recurrence relations
1.12 Appendix: asymptotics for n!
2. Continuoustime Markov chains I
2.1 Qmatrices and their exponentials
2.2 Continuoustime random processes
2.3 Some properties of the exponential distribution
ix
xiii
1
1
10
12
19
24
29
33
40
47
52
57
58
60
60
67
70
viii Contents
2.4 Poisson processes 73
2.5 Birth processes 81
2.6 Jump chain and holding times 87
2.7 Explosion 90
2.8 Forward and backward equations 93
2.9 Nonminimal chains 103
2.10 Appendix: matrix exponentials 105
3. Continuoustime Markov chains II 108
3.1 Basic properties 108
3.2 Class structure 111
3.3 Hitting times and absorption probabilities 112
3.4 Recurrence and transience 114
3.5 Invariant distributions 117
3.6 Convergence to equilibrium 121
3.7 Time reversal 123
3.8 Ergodic theorem 125
4. Further theory 128
4.1 Martingales 128
4.2 Potential theory 134
4.3 Electrical networks 151
4.4 Brownian motion 159
5. Applications 170
5.1 Markov chains in biology 170
5.2 Queues and queueing networks 179
5.3 Markov chains in resource management 192
5.4 Markov decision processes 197
5.5 Markov chain Monte Carlo 206
6. Appendix: probability and measure 217
6.1 Countable sets and countable sums 217
6.2 Basic facts of measure theory 220
6.3 Probability spaces and expectation 222
6.4 Monotone convergence and Fubini's theorem 223
6.5 Stopping times and the strong Markov property 224
6.6 Uniqueness of probabilities and independence of aalgebras 228
Further reading 232
Index 234
Preface
Markov chains are the simplest mathematical models for random phenom
ena evolving in time. Their simple structure makes it possible to say a great
deal about their behaviour. At the same time, the class of Markov chains
is rich enough to serve in many applications. This makes Markov chains
the first and most important examples of random processes. Indeed, the
whole of the mathematical study of random processes can be regarded as a
generalization in one way or another of the theory of Markov chains.
This book is an account of the elementary theory of Markov chains,
with applications. It was conceived as a text for advanced undergraduates
or master's level students, and is developed from a course taught to un
dergraduates for several years. There are no strict prerequisites but it is
envisaged that the reader will have taken a course in elementary probability.
In particular, measure theory is not a prerequisite.
The first half of the book is based on lecture notes for the undergradu
ate course. Illustrative examples introduce many of the key ideas. Careful
proofs are given throughout. There is a selection of exercises, which forms
the basis of classwork done by the students, and which has been tested
over several years. Chapter 1 deals with the theory of discretetime Markov
chains, and is the basis of all that follows. You must begin here. The
material is quite straightforward and the ideas introduced permeate the
whole book. The basic pattern of Chapter 1 is repeated in Chapter 3 for
continuoustime chains, making it easy to follow the development byanal
ogy. In between, Chapter 2 explains how to set up the theory of continuous
x Preface
time chains, beginning with simple examples such as the Poisson process
and chains with finite state space.
The second half of the book comprises three independent chapters in
tended to complement the first half. In some sections the style is a lit
tle more demanding. Chapter 4 introduces, in the context of elementary
Markov chains, some of the ideas crucial to the advanced study of Markov
processes, such as martingales, potentials, electrical networks and Brownian
motion. Chapter 5 is devoted to applications, for example to population
growth, mathematical genetics, queues and networks of queues, Markov de
cision processes and Monte Carlo simulation. Chapter 6 is an appendix to
the main text, where we explain some of the basic notions of probability
and measure used in the rest of the book and give careful proofs of the few
points where measure theory is really needed.
The following paragraph is directed primarily at an instructor and as
sumes some familiarity with the subject. Overall, the book is more focused
on the Markovian context than most other books dealing with the elemen
tary theory of stochastic processes. I believe that this restriction in scope
is desirable for the greater coherence and depth it allows. The treatment
of discretetime chains in Chapter 1 includes the calculation of transition
probabilities, hitting probabilities, expected hitting times and invariant dis
tributions. Also treated are recurrence and transience, convergence to equi
librium, reversibility, and the ergodic theorem for longrun averages. All
the results are proved, exploiting to the full the probabilistic viewpoint.
For example, we use excursions and the strong Markov property to obtain
conditions for recurrence and transience, and convergence to equilibrium is
proved by the coupling method. In Chapters 2 and 3 we proceed via the
jump chain/holding time construction to treat all rightcontinuous, mini
mal continuoustime chains, and establish analogues of all the main results
obtained for discrete time. No conditions of uniformly bounded rates are
needed. The student has the option to take Chapter 3 first, to study the
properties of continuoustime chains before the technically more demand
ing construction. We have left measure theory in the background, but
the proofs are intended to be rigorous, or very easily made rigorous, when
considered in measuretheoretic terms. Some further details are given in
Chapter 6.
It is a pleasure to acknowledge the work of colleagues from which I have
benefitted in preparing this book. The course on which it is based has
evolved over many years and under many hands  I inherited parts of it
from Martin Barlow and Chris Rogers. In recent years it has been given
by Doug Kennedy and Colin Sparrow. Richard Gibbens, Geoffrey Grim
Preface xi
mett, Frank Kelly and Gareth Roberts gave expert advice at various stages.
Meena Lakshmanan, Violet Lo and David Rose pointed out many typos ,and
ambiguities. Brian Ripley and David Williams made constructive sugges
tions for improvement of an early version.
I am especially grateful to David Thanah at Cambridge University Press
for his suggestion to write the book and for his continuing support, and to
Sarah SheaSimonds who typeset the whole book with efficiency, precision
and good humour.
Cambridge, 1996 James Norris
Introduction
This book is about a certain sort of random process. The characteristic
property of this sort of process is that it retains no memory of where it has
been in the past. This means that only the current state of the process can
influence where it goes next. Such a process is called a Markov process. We
shall be concerned exclusively with the case where the process can assume
only a finite or countable set of states, when it is usual to refer it as a
Markov chain.
Examples of Markov chains abound, as you will see throughout the book.
What makes them important is that not only do Markov chains model
many phenomena of interest, but also the lack of memory property makes
it possible to predict how a Markov chain may behave, and to compute
probabilities and expected values which quantify that behaviour. In this
book we shall present general techniques for the analysis of Markov chains,
together with many examples and applications. In this introduction we
shall discuss a few very simple examples and preview some of the questions
which the general theory will answer.
We shall consider chains both in discrete time
n E Z+ = {O, 1, 2, ... }
and continuous time
t E jR+ = [0, (0).
The letters n, m, k will always denote integers, whereas t and s will refer
to real numbers. Thus we write ( X n ) n ~ O for a discretetime process and
( X t ) t ~ O for a continuoustime process.
XIV Introduction
Markov chains are often best described by diagrams, of which we now
give some simple examples:
(i) (Discrete time)
1
3
1
3
2
You move from state 1 to state 2 with probability 1. From state 3 you
move either to 1 or to 2 with equal probability 1/2, and from 2 you jump
to 3 with probability 1/3, otherwise stay at 2. We might have drawn a loop
from 2 to itself with label 2/3. But since the total probability on jumping
from 2 must equal 1, this does not convey any more information and we
prefer to leave the loops out.
(ii) (Continuous time)
A
o••   ~ . ~   ...... 1
When in state 0 you wait for a random time with exponential distribution
of parameter A E (0, 00), then jump to 1. Thus the density function of the
waiting time T is given by
for t ~ o.
We write T rv E(A) for short.
(iii) (Continuous time)
A
• •
o 1
A
•
2 3 4
Here, when you get to 1 you do not stop but after another independent
exponential time of parameter Ajump to 2, and so on. The resulting process
is called the Poisson process of rate A.
3
Introduction
1
4
2
xv
(iv) (Continuous time)
In state 3 you take two independent exponential times T
1
rv E(2) and
T2 rv E (4); if T
1
is the smaller you go to 1 after time T
1
, and if T2 is the
smaller you go to 2 after time T
2
. The rules for states 1 and 2 are as given
in examples (ii) and (iii). It is a simple matter to show that the time spent
in 3 is exponential of parameter 2 + 4 = 6, and that the probability of
jumping from 3 to 1 is 2/(2 +4) = 1/3. The details are given later.
(v) (Discrete time)
3 6
2
o
5
We use this example to anticipate some of the ideas discussed in detail
in Chapter 1. The states may be partitioned into communicating classes,
namely {O}, {I, 2, 3} and {4, 5, 6}. Two of these classes are closed, meaning
that you cannot escape. The closed classes here are recurrent, meaning
that you return again and again to every state. The class {O} is transient.
The class {4, 5, 6} is periodic, but {I, 2, 3} is not. We shall show how to
establish the following facts by solving some simple linear/ equations. You
might like to try from first principles.
(a) Starting from 0, the probability of hitting 6 is 1/4.
(b) Starting from 1, the probability of hitting 3 is 1.
(c) Starting from 1, it takes on average three steps to hit 3.
(d) Starting from 1, the longrun proportion of time spent in 2 is 3/8.
xvi Introduction
Let us write pij for the probability starting from i of being in state j after
n steps. Then we have:
(e) lim POI = 9/32;
n+oo
(f) P04 does not converge as n ~ 00;
(g) lim pg4 = 1/124.
n+oo
1
Discretetime Markov chains
This chapter is the foundation for all that follows. Discretetime Markov
chains are defined and their behaviour is investigated. For better orien
tation we now list the key theorems: these are Theorems 1.3.2 and 1.3.5
on hitting times, Theorem 1.4.2 on the strong Markov property, Theorem
1.5.3 characterizing recurrence and transience, Theorem 1.7.7 on invariant
distributions and positive recurrence. Theorem 1.8.3 on convergence to
equilibrium, Theorem 1.9.3 on reversibility, and Theorem 1.10.2 on long
run averages. Once you understand these you will understand the basic
theory. Part of that understanding will come from familiarity with exam
ples, so a large number are worked out in the text. Exercises at the end of
each section are an important part of the exposition.
1.1 Definition and basic properties
Let I be a countable set. Each i E I is called a state and I is called the
statespace. We say that A = (Ai: i E I) is a measure on I if 0 ~ Ai < 00
for all i E I. If in addition the total mass EiEI Ai equals 1, then we call
A a distribution. We work'throughout with a probability space (0, F, lP).
Recall that a random variable X with values in I is a function X : 0 + I.
Suppose we set
Ai = IF(X = i) = IF( {w : X(w) = i}).
2 1. Discretetime Markov chains
Then A defines a distribution, the distribution of X. We think of X as
modelling a random state which takes the value i with probability Ai. There
is a brief review of some basic facts about countable sets and probability
spaces in Chapter 6.
We say that a matrix P == (Pij : i, j E I) is stochastic if every row
(Pij : j E I) is a distribution. There is a onetoone correspondence between
stochastic matrices P and the sort of diagrams described in the Introduc
tion. Here are two examples:
(Ia
1 /3)
P
 (3
1
p= (
1
)
1/2
1/2 0 1/2
3 1 2
2
We shall now formalize the rules for a Markov chain by a definition in
terms of the corresponding matrix P. We say that is a Markov
chain wit'h initial distribution A and transition matrix P if
(i) X
o
has distribution A;
(ii) for n 0, conditional on X
n
== i, X
n
+
1
has distribution (Pij : j E I)
and is independent of X
o
,.·· ,X
n

1
.
More explicitly, these conditions state that, for n 2: 0 and io, ... ,in+l E I,
(i) P(X0 == io) == Aio;
(ii) P(X
n
+1 == i
n
+
1
I X
o
== io, ... ,X
n
== in) == Pi
n
i
n
+
1
•
We say that is Markov (A, P) for short. If is a finite
sequence of random variables satisfying (i) and (ii) for n == 0, ... ,N  1,
then we again say is Markov (A, P).
It is in terms of properties (i) and (ii) that most realworld examples are
seen to be Markov chains. But mathematically the. following result appears
to give a more comprehensive description, and it is the key to some later
calculations.
Theorem 1.1.1. A discretetime random process (Xn)O<n<N is
Markov(A, P) if and only if for all io, ... ,iN E I
1.1 Definition and basic properties
Proof. Suppose is Markov(A, P), then
P(X
o
= iO,X
I
= i
l
, ... ,X
N
= iN)
= P(X
o
= iO)P(X
I
= il I X
o
= io)
... lP(X
N
= iN I X
o
= io,· .. ,X
N

I
= iNI)
3
On the other hand, if (1.1) holds for N, then by summing both sides over
iN E I and using EjEI Pij = 1 we see that (1.1) holds for N  1 and, by
induction
for all n = 0,1, ... ,N. In particular, P(X
o
= io) = Aio and, for n
0,1, ... ,N  1,
P(X
n
+
1
= i
n
+
1
I X
o
= io, ... ,X
n
= in)
= P(X
o
= i
o
, ... ,X
n
= in, X
n
+
1
= in+I)/P(X
O
= i
o
,· .. ,X
n
= in)
So is Markov(A, P). D
The next result reinforces the idea that Markov chains have no memory.
We write 8
i
= (8
ij
: j E I) for the unit mass at i, where
{
I ifi=j
8ij =
°otherwise.
Theorem 1.1.2 (Markov property). Let be Markov(A, P).
Then, conditional on X
m
= i, is Markov(8
i
, P) and is indepen
dent of the random variables X
o
, . .. ,X
m
.
Proof. We have to show that for any event A determined by ... ,X
m
we have
lP({Xm = im,··· ,Xm+n = im+n} n A I Xm = i)
= 8iirnPirnirn+l .. ·Pirn+nlirn+nlP(A I Xm = i) (1.2)
then the result follows by Theorem 1.1.1. First consider the case of elemen
taryevents
A = {X
o
= i
o
, ... ,X
m
= i
m
}.
4 1. Discretetime Markov chains
In that case we have to show
P(X
o
= i
o
,· .. ,X
m
+
n
= i
m
+
n
and i = im)/P(X
m
= i)
= biirnPi
rn
irn+l Pirn+nl i
rn
+
n
X lP(X
o
= io, ,X
m
= i
m
and i = im)/lP(X
m
= i)
which is true by Theorem 1.1.1. In general, any event A determined by
X
o
, ... ,X
m
may be written as a countable disjoint union of elementary
events
00
A= UAk.
k=l
Then the desired identity (1.2) for A follows by summing up the corre
sponding identities for A
k
. D
The remainder of this section addresses the following problem: what is
the probability that after n steps our Markov chain is in a given First
we shall see how the problem reduces to calculating entries in the nth power
of the transition matrix. Then we shall look at some examples where this
may be done explicitly.
We regard distributions and measures A as row vectors whose compo
nents are indexed by I, just as P is a matrix whose entries are indexed by
I x I. When I is finite we will often label the states 1,2, ... ,N; then A
will be an Nvector and P an N x Nmatrix. For these objects, matrix
multiplication is a familiar operation. We extend matrix multiplication to
the general case in the obvious way, defining a new measure AP and a new
matrix p
2
by
(AP)j = L AiPij,
iEI
(p2)ik = LPijPjk.
jEI
We define pn similarly for any n. We agree that pO is the identity matrix
I, where (I)ij = 8
ij
. The context will make it clear when I refers to the
statespace and when to the identity matrix. We write = (pn)ij for
the (i, j) entry in pn.
In the case where Ai > °we shall write Pi(A) for the conditional prob
ability P(A I X
o
= i). By the Markov property at time m = 0, under lPi,
is Markov(8
i
, P). So the behaviour of under lPi does not
depend on A.
Theorem 1.1.3. Let be Markov(A, P). Then, for all n, m 0,
(i) P(X
n
= j) = (Apn)j;
(ii) lP\(X
n
= j) = JP(X
n
+
m
= j I X
m
= i) = ·
1.1 Definition and basic properties
Proof. (i) By Theorem 1.1.1
P(Xn = j) = L ·.. L P(X
o
= i
o
,··· ,Xn I = inI, X n = j)
ioEI inlEI
= L ... L AioPioil" · PinIi = (Apnk
ioEI inlEI
5
(ii) By the Markov property, conditional on X
m
= i, is Markov
(8
i
, P), so we just take A = 8
i
in (i). D
In light of this theorem we call the nstep transition probability from i
to j. The following examples give some methods for calculating
Example 1.1.4
The most general twostate chain has transition matrix of the form
(
Ia
p = {3
and is represented by the following diagram:
{3
We exploit the relation p
n
+
I
= pnP to write
(n+I) _ (n){3 + (n) (1 _ )
PII  PI2 PII a .
We also know that + = IP\ (X
n
= 1 or 2) = 1, so by eliminating
we get a recurrence relation for
(n+I)  (1   (3) (n) + {3
PII  a PII ,
This has a unique solution (see Section 1.11):
(n) {_(3_ + _a_(l_ a  (3)n for a +{3 > 0
PII = a + {3 a + {3
1 for a +{3 = O.
6 1. Discretetime Markov chains
Example 1.1.5 (Virus mutation)
Suppose a virus can exist in N different strains and in each generation
either stays the same, or with probability a mutates to another strain,
which is chosen at random. What is the probability that the strain in the
nth generation is the same as that in the Oth?
We could model this process as an Nstate chain, with N x N transition
matrix P given by
Pii = 1  a, Pij = a / (N  1) for i =I j.
Then the answer we want would be found by computing p i ~ ) . In fact, in
this example there is a much simpler approach, which relies on exploiting
the symmetry present in the mutation rules.
At any time a transition is made from the initial state to another with
probability a, and a transition from another state to the initial state with
probability a/(N  1). Thus we have a twostate chain with diagram
a/(N  1)
and by putting (3 = a/(N  1) in Example 1.1.4 we find that the desired
probability is
!. + (1 !.) (1 ~ ) n
N N N1·
Beware that in examples having less symmetry, this sort of lumping together
of states may not produce a Markov chain.
Example 1.1.6
Consider the threestate chain with diagram
1
3
1
2
2
1.1 Definition and basic properties
and transition matrix
7
p= I)·
The problem is to find a general formula for
First we compute the eigenvalues of P by writing down its characteristic
equation
o= det (x  P) = x(x   =  1)(4x
2
+1).
The eigenvalues are 1, i/2, i/2 and from this we deduce that has the
form
(
.)n (")n
(n)
Pu = a + b 2 + c  2
for some constants a, band c. (The justification comes from linear algebra:
having distinct eigenvalues, P is diagonalizable, that is, for some invertible
matrix U we have
) U
1
i/2
and hence
p
n
= U (00
1
U
1
o (i/2)n
which forces to have the form claimed.) The answer we want is real
and
= (cosn
2
7r
±isin
n
2
7r
)
so it makes sense to rewrite in the form
for constants Q, (3 The first few values of are easy to write
down, so we get equations to solve for Q, (3 and
1 = = Q +(3
o= pii) = Q +
O
 p(2)  r\J  1(3
 11  L.(, 4
8 1. Discretetime Markov chains
SO Q: = 1/5, (3 = 4/5, = 2/5 and
More generally, the following method may in principle be used to find a
formula for for any Mstate chain and any states i and j.
(i) Compute the eigenvalues AI, ... ,AM of P by solving the character
istic equation.
(ii) If the eigenvalues are distinct then has the form
(n) _ \ n \ n
Pij  alAI + ... +aMAM
for some constants al, . .. ,aM (depending on i and j). If an eigen
value A is repeated (once, say) then the general form includes the
term (an+b)A
n
.
(iii) As roots of a polynomial with real coefficients, complex eigenvalues
will come in conjugate pairs and these are best written using sine
and cosine, as in the example.
Exercises
1.1.1 Let B
I
, B
2
, • •• be disjoint events with B
n
= O. Show that if A
is another event and P(AIB
n
) = P for all n then P(A) = p.
Deduce that if X and Yare discrete random variables then the following
are equivalent:
(a) X and Yare independent;
(b) the conditional distribution of X given Y = y is independent of y.
1.1.2 Suppose that is Markov (A, P). If Y
n
= Xkn, show that
is Markov (A, p
k
).
1.1.3 Let X
o
be a random variable with values in a countable set I. Let
Y
I
, Y
2
, . .. be a sequence of independent random variables, uniformly dis
tributed on [0, 1]. Suppose we are given a function
G : I x [0, 1] + I
and define inductively
Show that is a Markov chain and express its transition matrix P
in terms of G. Can all Markov chains be realized in this way? How would
you simulate a Markov chain using a computer?
1.1 Definition and basic properties 9
Suppose now that Zo, Z1, . .. are independent, identically distributed
random variables such that Zi = 1 with probability p and Zi = 0 with
probability 1  p. Set So = 0, Sn = Z1 + ... + Zn. In each of the following
cases determine whether ( X n ) n ~ O is a Markov chain:
(a) X
n
= Zn, (b) X
n
= Sn,
(c) X
n
= So + ... + Sn, (d)X
n
= (Sn, So + ... + Sn).
In the cases where ( X n ) n ~ O is a Markov chain find its statespace and
transition matrix, and in the cases where it is not a Markov chain give an
example where P(X
n
+
1
= ilX
n
= j, X
n

1
= k) is not independent of k.
1.1.4 A flea hops about at random on the vertices of a triangle, with all
jumps equally likely. Find the probability that after n hops the flea is back
where it started.
A second flea also hops about on the vertices of a triangle, but this flea is
twice as likely to jump clockwise as anticlockwise. What is the probability
that after n hops this second flea is back where it started? [Recall that
e±i7r/6 = V3/2 ± i/2.]
1.1.5 A die is 'fixed' so that each time it is rolled the score cannot be the
same as the preceding score, all other scores having probability 1/5. If the
first score fS 6, what is the probability p that the nth score is 6? What is
the probability that the nth score is I?
Suppose now that a new die is produced which cannot score one greater
(mod 6) than the preceding score, all other scores having equal probability.
By considering the relationship between the two dice find the value of p for
the new die.
1.1.6 An octopus is trained to choose object A from a pair of objects A, B
by being given repeated trials in which it is shown both and is rewarded
with food if it chooses A. The octopus may be in one of three states of mind:
in state 1 it cannot remember which object is rewarded and is equally likely
to choose either; in state 2 it remembers and chooses A but may forget
again; in state 3 it remembers and chooses A and never forgets. After each
tr·ial it may change its state of mind according to the transition matrix
State 1 ~ ~ 0
State 2 ~ l2 1
5
2
State 3 0 0 1.
It is in state 1 before the first trial. What is the probablity that it is
in state 1 just before the (n+l)th trial? What is the probability Pn+1 (A)
that it chooses A on the (n + 1)th trial ?
10 1. Discretetime Markov chains
Someone suggests that the record of successive choices (a sequence of As
and Bs) might arise from a twostate Markov chain with constant transition
probabilities. Discuss, with reference to the value of P
n
+
1
(A) that you have
found, whether this is possible.
1.1.7 Let ( X n ) n ~ O be a Markov chain on {1,2,3} with transition matrix
1 0 )
2/3 1/3 .
1 P 0
Calculate P(X
n
= 11X
o
= 1) in each of the following cases: (a) p = 1/16,
(b) p = 1/6, (c) p = 1/12.
1.2 Class structure
It is sometimes possible to break a Markov chain into smaller pieces, each
of which is relatively easy to understand, and which together give an un
derstanding of the whole. This is done by identifying the communicating
classes of the chain.
We say that i leads to j and write i ~ j if
Pi(X
n
= j for some n ~ 0) > o.
We say i communicates with j and write i ~ j if both i ~ j and j ~ i.
Theorem 1.2.1. For distinct states i and j the following are equivalent:
(i) i ~ j;
(ii) P
i
oi
l
P
i
li2 ... Pinli
n
> 0 for some states io,il, ... ,in with io = i and
in = j;
(iii) p ~ j ) > 0 for some n ~ O.
Proof. Observe that
00
p ~ j ) ::; lPi(X
n
= j for some n ~ 0) ::; L p ~ j )
n=O
which proves the equivalence of (i) and (iii). Also
p ~ j ) = L Pii
1
P i li2 .. ·Pind
il , ... ,inl
so that (ii) and (iii) are equivalent. 0
1.3 Hitting times and absorption probabilities 11
It is clear from (ii) that i ~ j and j ~ k imply i ~ k. Also i ~ i for
any state i. So ~ satisfies the conditions for an equivalence relation on I,
and thus partitions I into communicating classes. We say that a class C is
closed if
i E C, i ~ j imply j E C.
Thus a closed class is one from which there is no escape. A state i is
absorbing if {i} is a closed class. The smaller pieces referred to above are
these communicating classes. A chain or transition matrix P where I is a
single class is called irreducible.
As the following example makes clear, when one can draw the diagram,
the class structure of a chain is very easy to find.
Example 1.2.2
Find the communicating classes associated to the stochastic matrix
1 1
0 0 0 0
2 2
0 0 1 0 0 0
1
0 0
1 1
0
P=
3 3 3
0 0 0
1 1
0
2 2
0 0 0 0 0 1
0 0 0 0 1 0
The solution is obvious from the diagram
1 4
2 6
the classes being {1,2,3}, {4} and {5,6}, with only {5,6} being closed.
Exercises
1.2.1 Identify the communicating classes of the following transition matrix:
1
0 0 0
1
2 2
0
1
0
1
0
2 2
P= 0 0 1 0 0
0
1 1 1 1
:4 :4 :4 :4
1
0 0 0
1
2 2
Which classes are closed?
12 1. Discretetime Markov chains
1.2.2 Show that every transition matrix on a finite statespace has at least
one closed communicating class. Find an example of a transition matrix
with no closed communicating class.
1.3 Hitting times and absorption probabilities
Let (Xn)n>O be a Markov chain with transition matrix P. The hitting time
of a subset A of I is the random variable H
A
: n ~ {O,1,2, ... } U {oo}
given by
HA(w) = inf{n ~ 0 : Xn(w) E A}
where we agree that the infimum of the empty set 0 is 00. The probability
starting from i that ( X n ) n ~ O ever hits A is then
When A is a closed class, hf is called the absorption probability. The mean
time taken for ( X n ) n ~ O to reach A is given by
kt = lEi(H
A
) = 2: nJP>(H
A
= n) + ooJP>(H
A
= (0).
n<oo
We shall often write less formally
Remarkably, these quantities can be calculated explicitly by means of cer
tain linear equations associated with the transition matrix P. Before we
give the general theory, here is a simple example.
Example 1.3.1
Consider the chain with the following diagram:
1
1
1
2
2
2
•
•
E
• 41( •
~
•
1 2
1
3 4
2
Starting from 2, what is the probability of absorption in 4? How long does
it take until the chain is absorbed in 1 or 4?
Introduce
1.3 Hitting times and absorption probabilities 13
Clearly, hI = 0, h
4
= 1 and k
l
= k
4
= o. Suppose now that we start at 2,
and consider the situation after making one step. With probability 1/2 we
jump to 1 and with probability 1/2 we jump to 3. So
The 1 appears in the second formula because we count the time for the first
step. Similarly,
Hence
h2 = h3 = ( h2 +
k2 = 1 + = 1 + +
So, starting from 2, the probability of hitting 4 is 1/3 and the mean time to
absorption is 2. Note that in writing down the first equations for h
2
and k
2
we made implicit use of the Markov property, in assuming that the chain
begins afresh from its new position after the first jump. Here is a general
result for hitting probabilities.
Theorem 1.3.2. The vector of hitting probabilities h
A
= (hf : i E I) is
the minimal nonnegative solution to the system of linear equations
{
hf = 1 for i E A
hf = EjEI Pij
h
1 for i fj. A.
(1.3)
(Minimality means that if x = (Xi: i E I) is another solution with Xi 0
for all i, then Xi hi for all i.)
Proof. First we show that h
A
satisfies (1.3). If X
o
= i E A, then H
A
= 0,
so hf = 1. If X
o
= i fj. A, then H
A
1, so by the Markov property
and
hf = lP\(H
A
< 00) = LJPi(H
A
< oo,X
I
= j)
jEI
=LJPi(H
A
< 00 I Xl = j)JPi(XI = j) = LPijhf.
jEI jEI
14 1. Discretetime Markov chains
Suppose now that X = (Xi: i E I) is any solution to (1.3). Then hf = Xi = 1
for i E A. Suppose i ¢ A, then
Xi = '2:PijXj = '2:Pij + '2:PijXj.
jEI jEA
Substitute for Xj to obtain
Xi = LPij + LPij(LPjk + LPjkXk)
jEA kEA
= JPli(X
1
E A) + JPli(X
1
¢ A, X2 E A) + '2: '2: PijPjkXk·
By repeated substitution for X in the final term we obtain after n steps
Xi = JP>i(X
I
E A) + ... + JP>i(X
I
fj. A, ... ,X
n

I
fj. A, X
n
E A)
+ L ... L PiiIPjli2 · · · Pjnlin Xjn'
Now if X is nonnegative, so is the last term on the right, and the remaining
terms sum to JP>i(H
A
n). So Xi JP>i(H
A
n) for all n and then
Xi lim JP>i(H
A
n) = JP>i(H
A
< 00) = hi.
n+oo
Example 1.3.1 (continued)
The system of linear equations (1.3) for h = h{4} are given here by
h
4
= 1,
h2 = + h3 = +
so that
and
D
The value of hI is not determined by the system (1.3), but the minimality
condition now makes us take hI = 0, so we recover h
2
= 1/3 as before. Of
course, the extra boundary condition hI = 0 was obvious from the beginning
1.3 Hitting times and absorption probabilities 15
so we built it into our system of equations and did not have to worry about
minimal nonnegative solutions.
In cases where the statespace is infinite it may not be possible to write
down a corresponding extra boundary condition. Then, as we shall see in
the next examples, the minimality condition is essential.
Example 1.3.3 (Gamblers' ruin)
Consider the Markov chain with diagram
•
o
q p
I( • ..
1
q p q p
I( ... 1( ••
i i + 1
where 0 < P = 1  q < 1. The transition probabilities are
Poo = 1,
Pi,il = q, Pi,i+l = P for i = 1,2, ....
Imagine that you enter a casino with a fortune of £i and gamble, £1 at a
time, with probability P of doubling your stake and probability q of losing
it. The resources of the casino are regarded as infinite, so there is no upper
limit to your fortune. But what is the probability that you leave broke?
Set hi = IPi(hit 0), then h is the minimal nonnegative solution to
h
o
= 1,
hi = ph
i
+
1
+ qh
i

1
, for i = 1,2, ....
If p =I q this recurrence relation has a general solution
hi = A + B ( ~ ) i .
(See Section 1.11.) If P < q, which is the case in most successful casinos,
then the restriction 0 ~ hi ~ 1 forces B = 0, so hi = 1 for all i. If p > q,
then since h
o
= 1 we get a family of solutions
for a nonnegative solution we must have A ~ 0, so the minimal non
negative solution is hi = (q/ p)i. Finally, if p = q the recurrence relation
has a general solution
hi = A+ Bi
16 1. Discretetime Markov chains
and again the restriction 0 hi 1 forces B = 0, so hi = 1 for all i.
Thus, even if you find a fair casino, you are certain to end up broke. This
apparent paradox is called gamblers' ruin.
Example 1.3.4 (Birthanddeath chain)
Consider the Markov chain with diagram
ql PI qi Pi qi+1 Pi+1
....t(...... ......     • ....... t(I..••    
o 1 i i+l
where, for i = 1,2, ... , we have 0 < Pi = 1  qi < 1. As in the preceding
example, 0 is an absorbing state and we wish to calculate the absorption
probability starting from i. But here we allow Pi and qi to depend on i.
Such a chain may serve as a model for the size of a population, recorded
each time it changes, Pi being the probability that we get a birth before
a death in a population of size i. Then hi = IPi(hit 0) is the extinction
probability starting from i.
We write down the usual system of equations
h
o
= 1,
hi = Pihi+1 + qihiI, for i = 1,2, ....
This recurrence relation has variable coefficients so the usual technique fails.
But consider Ui = h
i

I
 hi, then PiUi+1 = qiUi, so
where the final equality defines Then
UI + ... +Ui = h
o
 hi
so
where A = UI and = 1. At this point A remains to be determined. In
the case L::o = 00, the restriction 0 hi 1 forces A = 0 and hi = 1
for all i. But if L::o < 00 then we can take A > 0 so long as
... for all i.
1.3 Hitting times and absorption probabilities 17
(1.4)
Thus the minimal nonnegative solution occurs when A = (E:o Ii) 1 and
then
In this case, for i = 1,2, ... , we have hi < 1, so the population survives
with positive probability.
Here is the general result on mean hitting times. Recall that kf =
Ei(H
A
), where H
A
is the first time hits A. We use the notation
1
B
for the indicator function of B, so, for example, 1
x1
==j is the random
variable equal to 1 if Xl = j and equal to 0 otherwise.
Theorem 1.3.5. The vector of mean hitting times k
A
= (k
A
: i E I) is
the minimal nonnegative solution to the system of linear equations
{
kf = 0 for i E A
kf = 1 + pijkf for i f/: A.
Proof. First we show that k
A
satisfies (1.4). If X
o
= i E A, then HA = 0,
so kf = o. If X
o
= i f/: A, then H
A
1, so, by the Markov property,
and
kt = Ei(H
A
) = L
E
i(H
A
1
X1
=j)
JEI
= LEi(H
A
I Xl = j)JP\(XI = j) = 1 + LPijkf·
JEI
Suppose now that Y = (Yi : i E I) is any solution to (1.4). Then kf = Yi = 0
for i E A. If i f/: A, then
Yi = 1 + LPijYj
= 1 + LPi
j
(l + LPjkYk)
= lP'i(H
A
1) + lP'i(H
A
2) + L L PijPjkYk·
By repeated substitution for Y in the final term we obtain after n steps
Yi = lP'i(H
A
1) + ... + lP'i(H
A
n) + L ... L PiilPilh .. 'PjndnYjn'
18 1. Discretetime Markov chains
So, if y is nonnegative,
Yi 2: Wi(H
A
2: 1) +... +Wi(H
A
2: n)
and, letting n * 00,
Yi 2': L IP\(H
A
2': n) = JEi(H
A
) = Xi
n==l
D
Exercises
1.3.1 Prove the claims (a), (b) and (c) made in example (v) of the Intro
duction.
1.3.2 A gambler has £2 and needs to increase it to £10 in a hurry. He
can play a game with the following rules: a fair coin is tossed; if a player
bets on the right side, he wins a sum equal to his stake, and his stake is
returned; otherwise he loses his stake. The gambler decides to use a bold
strategy in which he stakes all his money if he has £5 or less, and otllerwise
stakes just enough to increase his capital, if he wins, to £10.
Let X
o
== 2 and let X
n
be his capital after n throws. P r o ~ e that the
gambler will achieve his aim with probability 1/5.
What is the expected number of tosses until the gambler either achieves
his aim or loses his capital?
1.3.3 A simple game of 'snakes and ladders' is played on a board of nine
squares.
1.4 Strong Markov property 19
At each turn a player tosses a fair coin and advances one or two places
according to whether the coin lands heads or tails. If you land at the foot
of a ladder you climb to the top, but if you land at the head of a snake you
slide down to the tail. How many turns on average does it take to complete
the game?
What is the probability that a player who has reached the middle square
will complete the game without slipping back to square I?
1.3.4 Let ( X n ) n ~ O be a Markov chain on {O, 1, ... } with transition proba
bilities given by
(
i +1) 2
POI = 1, Pi,i+l +Pi,il = 1, Pi,i+l = i Pi,il,
i ~ 1.
Show that if X
o
= 0 then the probability that X
n
~ 1 for all n ~ 1 is 6/1T
2
•
1.4 Strong Markov property
In Section 1.1 we proved the Markov property. This says that for each time
m, conditional on X
m
= i, the process after time m begins afresh from
i. Suppose, instead of conditioning on X
m
= i, we simply waited for the
process to hit state i, at some random time H. What can one say about the
process after time H? What if we replaced H by a more general random
time, for example H  I? In this section we shall identify a class of random
times at which a version of the Markov property does hold. This class will
include H but not H  1; after all, the process after time H  1 jumps
straight to i, so it does not simply begin afresh.
A random variable T : n ~ {O, 1,2, ... } U {(X)} is called a stopping time
if the event {T = n} depends only on X
o
, Xl, ... ,X
n
for n = 0,1,2, ....
Intuitively, by watching the process, you know at the time when T occurs.
If asked to stop at T, you know when to stop.
Examples 1.4.1
(a) The first passage time
T
j
= inf{n ~ 1 : X
n
= j}
is a stopping time because
(b) The first hitting time H
A
of Section 1.3 is a stopping time because
{H
A
= n} = {X
o
~ A, ... ,X
n

l
~ A,X
n
E A}.
20
(c) The last exit time
1. Discretetime Markov chains
LA = sup{n 0 : X
n
E A}
is not in general a stopping time because the event {LA == n} depends on
whether visits A or not.
We sllall show that the Markov property holds at stopping times. The
crucial point is that, if T is a stopping time and B n is determined by
X
o
, Xl, ... ,X
T
, then B n {T == m} is determined by X
o
, Xl, ... ,X
m
, for
all m = 0, 1, 2, ....
Theorem 1.4.2 (Strong Markov property). Let be
Markov(A, P) and let T be a stopping time of Then, conditional
on T < 00 and X
T
= i, is Markov(8
i
, P) and independent of
XO,X
I
,.·. ,XT.
Proof. If B is an event determined by X
o
, Xl, ... ,X
T
, then B n {T = m}
is determined by X
o
, Xl, ... ,X
m
, so, by the Markov property at time m
lP( {X
T
= io, XT+I = il, ,XT+n = in} n B n {T = m} n {X
T
= i})
= JP>i(X
O
= io,X
I
= jl, ,X
n
= jn)JP>(B n {T = m} n{X
T
= i})
where we have used the condition T = m to replace m by T. Now sum over
m == 0, 1, 2, ... and divide by lP(T < 00, X
T
= i) to obtain
JP>( {X
T
= jo, X
T
+
I
= jl, ,X
T
+
n
= jn} n BIT < 00, X
T
= i)
= JP>i(X
O
= jo, Xl = jl, ,X
n
= jn)JP>(B I T < 00, X
T
= i). D
The following example uses the strong Markov property to get more
information on the hitting times of the chain considered in Example 1.3.3.
Example 1.4.3
Consider the Markov chain with diagram
•
o
q p
I( •
1
q p q p
I( •
i i + 1
1.4 Strong Markov property 21
where 0 < p = 1  q < 1. We know from Example 1.3.3 the probability of
hitting 0 starting from 1. Here we obtain the complete distribution of the
time to hit 0 starting from 1 in terms of its probability generating function.
Set
H
j
= inf{n ~ 0 : X
n
= j}
and, for 0 ~ s < 1
</>(s) = lEI (sHo) = 2: snIP
1
(H
o
= n).
n<oo
Suppose we start at 2. Apply the strong Markov property at !II to see
that under lP
2
, conditional on HI < 00, we have H
o
= HI + H
o
, where
ii
o
, the time taken after HI to get to 0, is independent of HI and has the
(unconditioned) distribution of HI. So
E
2
(SHO) = E
2
(sHl I HI < 00)E
2
(sHO I HI < 00)JP>2(H
l
< 00)
=E2(sHI1Hl<OO)E2(sHO I HI < 00)
= E
2
(sHl)2 = ¢(s)2.
Then, by the Markov property at time 1, conditional on Xl = 2, we have
H
o
= 1 +H
o
, where H
o
, the time taken after time 1 to get to 0, has the
same distribution as H
o
does under lP
2
. So
¢(s) = El(sH
o
) = pEl (sHo I Xl = 2) +qEl(sH
o
I Xl = 0)
= pEl (sl+Ho I Xl = 2) +qEl(s I Xl = 0)
= psE
2
(sH
o
) +qs
=ps¢(s)2 +qs.
Thus ¢ = ¢(s) satisfies
pS¢2  ¢ +qs = 0
and
(1.5)
¢ = (1 ± VI  4pqs2)/2ps.
Since ¢(O) ~ 1 and ¢ is continuous we are forced to take the negative root
at s = 0 and stick with it for all 0 ~ s < 1.
To recover the distribution of H
o
we expand the squareroot as a power
series:
</>(s) = 2 ~ S { 1 (1 + !(4
pqs
2) + !(lH_4
pqs
2)2/2! + ... ) }
2 3
= qs +pq s +...
= slPl(H
o
= 1) +s2lP
l
(H
o
= 2) +s3lP
l
(H
o
= 3) +....
22 1. Discretetime Markov chains
The first few probabilities P1(H
o
= 1),P
1
(H
o
= 2), ... are readily checked
from first principles.
On letting s i 1 we have ¢(s) ~ P1(H
o
< 00), so
lI]) (Ii ) _ 1  VI  4pq _ { 1 if p ~ q
.1rl 0 < 00  
2p q/p if p > q.
(Remember that q = 1  p, so
VI  4pq = VI  4p +4
p
2 = 11  2pl = 12q  11.)
We can also find the mean hitting time using
It is only worth considering the case p ~ q, where the mean hitting time
has a chance of being finite. Differentiate (1.5) to obtain
2ps¢¢' +p¢2  ¢' +q = 0
so
¢'(s) = (p¢(S)2 +q)/(l  2ps¢(s)) ~ 1/(1  2p) = l/(q  p) as s i 1.
See Example 5.1.1 for a connection with branching processes.
Example 1.4.4
We now consider an application of the strong Markov property to a Markov
chain ( X n ) n ~ O observed only at certain times. In the first instance suppose
that J is some subset of the statespace I and that we observe the chain
only when it takes values in J. The resulting process ( Y m ) m ~ O may be
obtained formally by setting Y
m
= X
Trn
, where
To = inf{n ~ 0 : X
n
E J}
and, for m = 0, 1, 2, ...
T
m
+
1
= inf{n > T
m
: X
n
E J}.
Let us assume that lP(Tm < 00) = 1 for all m. For each m we can check
easily that Tm, the time of the mth visit to J, is a stopping time. So the
strong Markov property applies to show, for i
o
, ... ,im+l E J, that
P(Ym +1 = i m +1 I Yo = i o,··· , Ym = im )
= P(X
Trn
+
1
= i m +1 I X
To
= i o,··· ,X
Trn
= im)
= Pi (X
T
= i
m
+
1
) = p. .
rn 1 't rn't rn+l
where, for i, j E J
1.4 Strong Markov property 23
(1.6)
 hi
Pii = i
and where, for j E J, the vector (h{ : i E I) is the minimal nonnegative
solution to
h{ = Pij + LPik
h
1·
Thus is a Markov chain on J with transition matrix P.
A second example of a similar type arises if we observe the original chain
only when it moves. The resulting process is given by
Zm = XS
rn
where 8
0
= 0 and for m = 0,1,2, ...
Let us assume there are no absorbing states. Again the random times 8
m
for m 0 are stopping times and, by the strong Markov property
lP(Zm+l = im+l I Zo = io,··· ,Zm = im)
= lP(X
Srn
+
1
= im+l I XS
o
= io, ... ,Xs
rn
= i
m
)
= lPirn (XS1 = i m +1 ) = Pirnirn+l
where Pii = 0 and, for i =I j
Pij = Pij/ LPik.
k=j:i
Thus is a Markov chain on I with transition matrix P.
Exercises
1.4.1 Let Y
1
, Y
2
, ••• be independent identically distributed random vari
ables with
lP(Y
1
= 1) = lP(Y
1
= 1) = 1/2 and set X
o
= 1, X
n
= X
o
+Y
1
+... +Y
n
for n 1. Define
H
o
= inf{n 0 : X
n
= O}.
Find the probability generating function ¢(s) = E(sHO).
Suppose the distribution of Y
1
, "Y2, ... is changed to lP(Y
1
= 2) = 1/2,
lP(Y
1
= 1) = 1/2. Show that ¢ now satisfies
s¢3  2¢ +s = 0.
1.4.2 Deduce carefully from Theorem 1.3.2 the claim made at (1.6).
24 1. Discretetime Markov chains
1.5 Recurrence and transience
Let ( X n ) n ~ O be a Markov chain with transition matrix P. We say that a
state i is recurrent if
lPi(X
n
= i for infinitely many n) = 1.
We say that i is transient if
IPi(X
n
= i for infinitely many n) = O.
Thus a recurrent state is one to which you keep coming back and a transient
state is one which you eventually leave for ever. We shall show that every
state is either recurrent or transient.
Recall that the first passage time to state i is the random variable T
i
defined by
Ti(W) = inf{n ~ 1 : Xn(w) = i}
where inf 0 = 00. We IIOW define inductively the rth passage time Ti(r) to
state i by
and, for r = 0,1,2, ... ,
The length of the rth excursion to i is then
{
T(r) _ T(rI)
S ~ r ) = i i
~ 0
·f T(rI)
1 i < 00
otherwise.
The following diagram illustrates these definitions:
X
n
i
T.(O)
~
n
1.5 Recurrence and transience 25
Our analysis of recurrence and transience will rest on finding the joint
distribution of these excursion lengths.
Lemma 1.5.1. For r = 2,3, ... , conditional on Ti(rI) < 00, sir) is inde
pendent of {X
m
: m Ti(rI)} and
Proof. Apply the strong Markov property at the stopping time T = Ti(rI).
It is automatic that X
T
= i on T < 00. So, conditional on T < 00,
is Markov(8
i
, P) and independent of X
o
, Xl, ... ,X
T
. But
sir) = inf{n 1 : X
T
+
n
= i},
so sir) is the first passage time of to state i. D
Recall that the indicator function l{xl==j} is the random variable equal
to 1 if Xl = j and 0 otherwise. Let us introduce the number of visits Vi to
i, which may be written in terms of indicator functions as
00
Vi = L l{x
n
=i}
n==O
and note that
00 00 00 00
lEi(Vi) = lEi L l{X
n
=i} = L
lE
i(l{X
n
=i}) = LlPi(Xn = i) =
n==O n==O n==O n==O
Also, we can compute the distribution of Vi under Pi in terms of the return
probability
Lemma 1.5.2. For r = 0,1,2, ... , we have Pi (Vi > r) = fi.
Proof. Observe that if X
o
= i then {Vi > r} = {Ti(r) < oo}. When r = 0
the result is true. Suppose inductively that it is true for r, then
( ) (
(r+I) )
Pi Vi > r +1 = Pi T
i
< 00
= Pi(Ti(r) < 00 and si
r
+
I
) < 00)
= Pi (Sfr+I) < 00 I Ti(r) < oo)Pi(Ti(r) < 00)
= fif; = f[+1
by Lemma 1.5.1, so by induction the result is true for all r. D
26 1. Discretetime Markov chains
Recall that one can compute the expectation of a nonnegative integer
valued random variable as follows:
00 00 00
LP(V > r) = L L P(V = v)
r=O r=Ov=r+1
00 vI 00
= L LP(V ="' v) = L vP(V = v) = E(V).
v=lr=O v=1
The next theorem is the means by which we establish recurrence or
transience for a given state. Note that it provides two criteria for this, one
in terms of the return probability, ttle other in terms of the nstep transition
probabilities. Both are useful.
Theorem 1.5.3. The following dichotomy holds:
(i) ifIPi(T
i
< 00) = 1, then i is recurrent and = 00;
(ii) ifIPi(T
i
< 00) < 1, then i is transient and < 00.
In particular, every state is either transient or recurrent.
Proof. If IPi(T
i
< 00) = 1, then, by Lemma 1.5.2,
IPi(Vi = 00) = lim IPi(Vi > r) = 1
r+oo
so i is recurrent and
00
= Ei(Vi) = 00.
n=O
On the other hand, if Ii = IPi(Ti < 00) < 1, then by Lemma 1.5.2
= Ei(Vi) = fpi(Vi > r) = fl[ = 1 . < 00
n=O r=O r=O f'l,
so IPi (Vi = 00) = 0 and i is transient. D
From this theorem we can go on to solve completely the problem of
recurrence or transience for Markov chains with finite statespace. Some
cases of infinite statespace are dealt with in the following chapter. First
we show that recurrence and transience are class properties.
Theorem 1.5.4. Let C be a communicating class. Then either all states
in C are transient or all are recurrent.
Proof. Take any pair of states i, j E C and suppose that i is transient.
There exist n, m 0 with > 0 and PJ":) > 0, and, for all r 0
(n+r+m) > (n) (r) (m)
Pii  Pij Pjj Pji
1.5 Recurrence and transience
so
~ (r) < 1 ~ (n+r+m)
LJPjj  (n) (m) LJPii < 00
r==O Pij Pji r==O
by Theorem 1.5.3. Hence j is also transient by Theorem 1.5.3. D
27
In the light of this theorem it is natural to speak of a recurrent or transient
class.
Theorem 1.5.5. Every recurrent class is closed.
Proof. Let C be a class which is not closed. Then there exist i E C, j fj. C
and m ~ 1 with
Since we have
lP
i
( {X
m
= j} n {X
n
= i for infinitely many n}) = 0
this implies that
lPi(X
n
= i for infinitely many n) < 1
so i is not recurrent, and so neither is C. D
Theorem 1.5.6. Every finite closed class is recurrent.
Proof. Suppose C is closed and finite and that ( X n ) n ~ O starts in C. Then
for some i E C we have
o< P(X
n
= i for infinitely many n)
= lP(X
n
= i for some n)Pi(X
n
= i for infinitely many n)
by the strong Markov property. This shows that i is not transient, so C is
recurrent by Theorems 1.5.3 and 1.5.4. D
It is easy to spot closed classes, so the transience or recurrence of finite
classes is easy to determine. For example, the only recurrent class in Ex
ample 1.2.2 is {5, 6}, the others being transient. On the other hand, infinite
closed classes may be transient: see Examples 1.3.3 and 1.6.3.
We shall need the following result in Section 1.8. Remember that irre
ducibility means that the chain can get from any state to any other, with
positive probability.
28 1. Discretetime Markov chains
Theorem 1.5.7. Suppose P is irreducible and recurrent. Then for all
j E I we have lP(T
j
< 00) = 1.
Proof. By the Markov property we have
lP(T
j
< (0) = LlP(Xo = i)lPi(T
j
< (0)
iEI
so it suffices to show lPi (T
j
< (0) = 1 for all i E I. Choose m with p;:n) > o.
By Theorem 1.5.3, we have
1 = Pj(X
n
= j for infinitely many n)
= lPj(X
n
= j for some n m +1)
=L lPj(X
n
= j for some n m + 1 I X
m
= k)lPj(X
m
= k)
kEI
= L lPk(T
j
< oo)p;7:)
kEI
where the final equality uses the Markov property. But EkE! p;7:) = 1 so
we must have Pi(T
j
< 00) = 1. D
Exercises
1.5.1 In Exercise 1.2.1, which states are recurrent and which are transient?
1.5.2 Show that, for the Markov chain in Exercise 1.3.4 we have
lP(X
n
00 as n 00) = 1.
Suppose, instead, the transition probabilities satisfy
(
i+l)O
Pi,i+l = i Pi,il·
For each Q E (0,00) find the value of P(X
n
00 as n 00).
1.5.3 (First passage decomposition). Denote by T
j
the first passage
time to state j and set
Justify the identity
n
= '"'
I) I) ))
k==l
for n 1
1.6 Recurrence and transience of random walks
and deduce that
where
29
CX)
Pij(s) =
n==O
00
F
ij
(s) = L sn·
n==O
Hence show that Pi (T
i
< 00) = 1 if and only if
00
= 00
n==O
without using Theorem 1.5.3.
1.5.4 A random sequence of nonnegative integers is obtained by
setting F
o
= 0 and F
1
= 1 and, once F
o
, ... ,F
n
are known, taking F
n
+
1
to
be either the sum or the difference of F
n

1
and F
n
, each with probability
1/2. Is a Markov chain?
By considering the Markov chain X
n
= (F
n

1
, F
n
), find the probability
that reaches 3 before first returning to O.
Draw enough of the flow diagram for to establish a general
pattern. Hence, using the strong Markov property, show that the hitting
probability for (1,1), starting from (1,2), is (3  V5)/2.
Deduce that is transient. Show that, moreover, with probability
1, F
n
00 as n 00.
1.6 Recurrence and transience of random walks
In the last section we showed that recurrence was a class property, that all
recurrent classes were closed and that all finite closed classes were recurrent.
So the only chains for which the question of recurrence remains interesting
are irreducible with infinite statespace. Here we shall study some simple
and fundamental examples of this type, making use of the following criterion
for recurrence from Theorem 1.5.3: a state i is recurrent if and only if
(n)_
L..Jn==O Pii  00.
Example 1.6.1 (Simple random walk on Z)
The simple random walk on Z has diagram
iI
q P
... ...
i i+l
30 1. Discretetime Markov chains
where 0 < P = 1  q < 1. Suppose we start at O. It is clear that we cannot
return to 0 after an odd number of steps, so = 0 for all n. Any
given sequence of steps of length 2n from 0 to 0 occurs with probability
pnqn, there being n steps up and n steps down, and the number of such
sequences is the number of ways of choosing the n steps up from 2n. Thus
(2n) (2n) n n
Poo = p q .
n
Stirling's formula provides a good approximation to n! for large n: it is
known that
where an rv b
n
means an/b
n
1. For a proof see W. Feller, An Introduction
to Probability Theory and its Applications, Vol I (Wiley, New York, 3rd
edition, 1968). At the end of this chapter we reproduce the argument used
by Feller to show that
for some A E [1,00). The additional work needed to show A = y'2;IT is
omitted, as this fact is unnecessary to our applications.
For the nstep transition probabilities we obtain
(2n) _ (2n)! ( )n rv (4pq)n
Poo  (n!)2 pq AJn/2
as n 00.
In the symmetric case p = q = 1/2, so 4pq = 1; then for some N and all
n N we have
(2n) > 1
Poo  2AVii
so
(2n) > .!... _1 _
Poo  2A Vii  00
which shows that the random walk is recurrent. On the other hand, if p =I q
then 4pq = r < 1, so by a similar argument, for some N
00 1 00
'"" p(n) < _ '"" r
n
< 00
00 A
n==N n=N
showing that the random walk is transient.
1.6 Recurrence and transience of random walks
Example 1.6.2 (Simple symmetric random walk on Z2)
The simple symmetric random walk on Z2 has diagram
1
1
:4
:4
......
,
1
1 , ~
:4
:4
31
and transition probabilities
{
1/4 iflijl==l
Pij == 0 otherwise.
Suppose we start at o. Let us call the walk X
n
and write X ~ and X; for
the orthogonal projections of X
n
on the diagonal lines y == ±x:
X+
n
Then X ~ and X; are independent simple symmetric random walks on
2
1
/
2
Z and X
n
== 0 if and only if X ~ == 0 == X;. This makes it clear that
for X
n
we have
(2n)
Poo ==
a s n ~ o o
32 1. Discretetime Markov chains
by Stirling's formula. Then E::o = 00 by comparison with E::o l/n
and the walk is recurrent.
Example 1.6.3 (Simple symmetric random walk on Z3)
The transition probabilities of the simple symmetric random walk on Z3
are given by
P
.. _ {1/6 if Ii  jl = 1
'1,) 
o otherwise.
Thus the chain jumps to each of its nearest neighbours with equal probabil
ity. Suppose we start at O. We can only return to 0 after an even number
2n of steps. Of these 2n steps there must be i up, i down, j north, j south,
k east and k west for some i,j, k 0, with i +j +k = n. By counting the
ways in which this can be done, we obtain
(2n) _
Poo 
Now
i+j+k=n
(2n)!
(
., ·'k')2
..
i+j+k=n
i+j+k=n
the lefthand side being the total probability of all the ways of placing n
balls randomly into three boxes. For the case where n = 3m, we have
(
n ) n! (n)
ij k = i!j!k! mmm
for all i, j, k, so
3/2
by Stirling's formula. Hence, E:=o < 00 by comparison with
,",00 3/2 B t (6m) > (1/6)2 (6m2) d (6m) > (1/6)4 (6m4) £
L...Jn=O n . u Poo Poo an Poo _ Poo or
all m so we must have
and the walk is transient.
Exercises
1. 7 Invariant distributions 33
1.6.1 The rooted binary tree is an infinite graph T with one distinguished
vertex R from which comes a single edge; at every other vertex there are
three edges and there are no closed loops. The random walk on T jumps
from a vertex along each available edge with equal probability. Show that
the random walk is transient. \
1.6.2 Show that the simple symmetric random walk in Z4 is transient.
1.7 Invariant distributions
Many of the longtime properties of Markov chains are connected with the
notion of an invariant distribution or measure. Remember that a measure
A is any row vector (Ai : i E I) with nonnegative entries. We say A is
invariant if
AP == A.
The terms equilibrium and stationary are also used to mean the same. The
first result explains the term stationary.
Theorem 1.7.1. Let be Markov(A,P) and suppose that A is in
variant for P. Then is also Markov(A, P).
Proof. By Theorem 1.1.3, P(X
m
== i) == (Apm)i == Ai for all i and, clearly,
conditional on X
m
+
n
== i, X
m
+
n
+
1
is independent of X
m
, X
m
+
1
, ... ,X
m
+
n
and has distribution (Pij : j E I). D
The next result explains the term equilibrium.
Theorem 1.7.2. Let I be finite. Suppose for some i E I that
t 7rj as n t 00 for all j E I.
Then 7r == (7rj : j E I) is an invariant distribution.
Proof. We have
L L
1
· (n) 1· L (n) 1
7rj == 1m P  == 1m P  ==
jEI jEI jEI
and
1
· (n) 1· L (n) L 1· (n) L
7rj == 1m P  == 1m Pk Pkj == 1m Pk Pkj == 7rkPkj
kEI kEI kEI
34 1. Discretetime Markov chains
where we have used finiteness of I to justify interchange of summation and
limit operations. Hence 7r is an invariant distribution. D
Notice that for any of the random walks discussed in Section 1.6 we have
t 0 as n t 00 for all i, j E I. The limit is certainly invariant, but it
is not a distribution!
Theorem 1.7.2 is not a very useful result but it serves to indicate a rela
tionship between invariant distributions and nstep transition probabilities.
In Theorem 1.8.3 we shall prove a sort of converse, which is much more
useful.
Example 1.7.3
Consider the twostate Markov chain with transition matrix
(
la
p
 (3
Ignore the trivial cases 0: = (3 = 0 and 0: = (3 = 1. Then, by Example 1.1.4
pn ((3/(a +(3)
(3/(a +(3)
a/(a +(3))
a/(a +(3)
as n 00,
so, by Theorem 1.7.2, the distribution ((3/(0: + (3),0:/(0: + (3)) must be
invariant. There are of course easier ways to discover this.
Example 1.7.4
Consider the Markov chain with diagram
1
3
1
2
2
To find an invariant distribution we write down the components of the
vector equation 7rP = 7r
7rl =
1 1
7r2 = 27r1 + 2
7r3
_ 1 1
7r3  27r2 + 2
7r3
·
1. 7 Invariant distributions 35
In terms of the chain, the righthand sides give the probabilities for Xl,
when X
o
has distribution 7r, and the equations require Xl also to have
distribution 7r. The equations are homogeneous so one of them is redundant,
and another equation is required to fix 7r uniquely. That equation is
and we find that 7r = (1/5,2/5,2/5).
According to Example 1.1.6
p i ~ ) ~ 1/5 as n ~ 00
so this confirms Theorem 1.7.2. Alternatively, knowing that p ~ ~ ) had the
form
(n) (l)n ( n7r . n7r)
PII = a + 2 bcos 2 + CSln 2
we could have used Theorem 1.7.2 and knowledge of 7r1 to identify a = 1/5,
instead of working out p ~ ; ) in Example 1.1.6.
In the next two results we shall show that every irreducible and recurrent
stochastic matrix P has an essentially unique positive invariant measure.
The proofs rely heavily on the probabilistic interpretation so it is worth
noting at the outset that, for a finite statespace I, the existence of an
invariant row vector is a simple piece of linear algebra: the row sums of P
are alII, so the column vector of ones is an eigenvector with eigenvalue 1,
so P must have a row eigenvector with eigenvalue 1.
For a fixed state k, consider for each i the expected time spent in i between
visits to k:
Tk l
')'f = lEk L l{x
n
=i}'
n=O
Here the sum of indicator functions serves to count the number of times n
at which X
n
= i before the first passage time T
k
.
Theorem 1.7.5. Let P be irreducible and recurrent. Then
(i) l ' ~ = 1;
(ii) l'k = (l'f : i E I) satisfies l'k P = l'k ;
(iii) 0 < Tf < 00 for all i E I.
Proof. (i) This is obvious. (ii) For n = 1,2, ... the event {n ::; Tk} depends
only on X
o
, Xl, . .. ,X
n

l
, so, by the Markov property at n  1
36 1. Discretetime Markov chains
Since P is recurrent, under IPk we have Tk < 00 and X
o
= X
Tk
= k with
probability one. Therefore
Tk 00
,j = lEk L l{x
n
=j} = lEk L l{X
n
=j and nSTd
n=l n=l
00
= L JP>k(X
n
= j and n ~ Tk)
n=l
00
= L L JP>k(X
n

1
= i, X
n
= j and n ~ Tk)
iEI n=l
00
= LPij L JP>k(X
n

1
= i and n ~ Tk)
iEI n=l
00
=LpijlEk L l{X
tn
=i and msTk 1}
iEI m=O
Tk 1
= LPijlEk L l{X
tn
=i} = L ,fpij.
iEI m=O iEI
(iii) Since P is irreducible, for each state i there exist n, m ~ 0 with
(n) (m) 0 Th k k (m) 0 d k (n) k b () d
Pik ,Pki >. en ~ i ~ ~ k P k i > an ~ i Pik ~ ~ k = 1 Y i an
(ii). D
Theorem 1.7.6. Let P be irreducible and let A be an invariant measure
for P with Ak = 1. Then A ~ ~ k . If in addition P is recurrent, then A = ~ k .
Proof. For each j E I we have
Aj = L AioPioj = L AioPioj + Pkj
ioEI io#k
= L Ai1Pi
1
ioPioj + (Pk
j
+ L PkioPiOj )
io,il#k io#k
L AinPininl · · · Pioj
i
o
,... ,in=j:k
+ (Pk
j
+ L PkioPioj + · · · + L Pkinl · · · PilioPioj )
io#k io, ... ,inl#k
~ IPk(X
1
= j and Tk ~ 1) +IPk(X2 = j and Tk ~ 2)
+... +IPk(X
n
=j and Tk ~ n)
~ ~ j as n ~ 00.
1. 7 Invariant distributions 37
So A If P is recurrent, then is invariant by Theorem 1.7.5, so
Jl = A  is also invariant and /l O. Since P is irreducible, given i E I,
h
(n) 0 £ dO'" (n) (n)
we ave Pik > or some n, an = /lk = LJjEI /ljPjk /liPik , so
Jli = O. 0
Recall that a state i is recurrent if
IPi(X
n
= i for infinitely many n) = 1
and we showed in Theorem 1.5.3 that this is equivalent to
If in addition the expected return time
is finite, then we say i is positive recurrent. A recurrent state which fails to
have this stronger property is called null recurrent.
Theorem 1.7.7. Let P be irreducible. Then the following are equivalent:
(i) every state is positive recurrent;
(ii) some state i is positive recurrent;
(iii) P has an invariant distribution, 7r say.
Moreover, when (iii) holds we have mi = 1/7ri for all i.
Proof. (i) => (ii) This is obvious.
(ii) => (iii) If i is positive recurrent, it is certainly recurrent, so P is recur
rent. By Theorem 1.7.5, is then invariant. But
= mi < 00
jEI
so 7rj = / mi defines an invariant distribution.
(iii) => (i) Take any state k. Since P is irreducible and EiEI 7ri = 1 we have
7rk = EiEI > 0 for some n. Set Ai = 7ri/7rk. Then A is an invariant
measure with Ak = 1. So by Theorem 1.7.6, A Hence
L
k L 7ri 1
mk =  =  < 00
7r 7r
iEI iEI k k
and k is positive recurrent.
(1.7)
38 1. Discretetime Markov chains
To complete the proof we return to the argument for (iii) => (i) armed
with the knowledge that P is recurrent, so A = ~ k and the inequality (1.7)
is in fact an equality. 0
Example 1.7.8 (Simple symmetric random walk on Z)
The simple symmetric random walk on Z is clearly irreducible and, by
Example 1.6.1, it is also recurrent. Consider the measure
7ri = 1 for all i.
Then
so 7r is invariant. Now Theorem 1.7.6 forces any invariant measure to be
a scalar multiple of 7r. Since LiEZ 7ri = 00, there can be no invariant
distribution and the walk is therefore null recurrent, by Theorem 1.7.7.
Example 1.7.9
The existence of an invariant measure does not guarantee recurrence: con
sider, for example, the simple symmetric random walk on Z3, which is
transient by Example 1.6.3, but has invariant measure 7r given by 7ri = 1
for all i.
Example 1.7.10
Consider the asymmetric random walk on Z with transition probabilities
Pi,il = q < P = Pi,i+l· In components the invariant measure equation
7rP = 7r reads
This is a recurrence relation for 7r with general solution
So, in this case, there is a twoparameter family of invariant measures 
uniqueness up to scalar multiples does not hold.
Example 1.7.11
Consider a successrun chain on Z+ , whose transition probabilities are given
by
Pi,i+l = Pi, PiO = qi = 1  Pi·
1. 7 Invariant distributions
Then the components of the invariant measure equation 7rP = 7r read
00
1ro = L qi
1r
i,
i=O
7ri = Pil7ril, for i ~ 1.
Suppose we choose Pi converging sufficiently rapidly to 1 so that
00
P = IIpi > 0
i=O
which is equivalent to
00
Lqi = 00.
i=O
Then for any solution of 7rP = 7r we have
and so
00
1ro ~ JJ7ro L qi·
i=O
39
This last equation forces either 7ro = 0 or 7ro = 00, so there is no invariant
measure.
Exercises
1.7.1 Find all invariant distributions of the transition matrix in Exercise
1.2.1.
1.7.2 Gas molecules move about randomly in a box which is divided into two
halves symmetrically by a partition. A hole is made in the partition. Sup
pose there are N molecules in the box. Show that the number of molecules
on one side of the partition just after a molecule has passed through the hole
evolves as a Markov chain. What are the transition probabilities? What is
the invariant distribution of this chain?
1.7.3 A particle moves on the eight vertices of a cube in the following
way: at each step the particle is equally likely to move to each of the three
adjacent vertices, independently of its past motion. Let i be the initial
vertex occupied by the particle, 0 the vertex opposite i. Calculate each of
the following quantities:
40 1. Discretetime Markov chains
(i) the expected number of steps until the particle returns to i;
(ii) the expected number of visits to 0 until the first return to i;
(iii) the expected number of steps until the first visit to o.
1.7.4 Let be a simple random walk on Z with Pi,il = q < P =
Pi,i+l. Find
(
TOl )
,f = lEo L l{x
n
=i}
n=O
and verify that
= inf Ai for all i
A
where the infimum is taken over all invariant measures A with Ao 1.
(Compare with Theorem 1.7.6 and Example 1.7.10.)
1.7.5 Let P be a stochastic matrix on a finite set I. Show that a distribution
7r is invariant for P if and only if 7r(I P+A) = a, where A = (aij : i,j E I)
with aij = 1 for all i and j, and a = (ai : i E I) with ai = 1 for all i. Deduce
that if P is irreducible then I P+A is invertible. Note that this enables one
to compute the invariant distribution by any standard method of inverting
a matrix.
1.8 Convergence to equilibrium
We shall investigate the limiting behaviour of the nstep transition proba
bilities as n 00. As we saw in Theorem 1.7.2, if the statespace is
finite and if for some i the limit exists for all j, then it must be an invariant
distribution. But, as the following example shows, the limit does not always
exist.
Example 1.8.1
Consider the twostate chain with transition matrix
Then p2 = I, so p2n = I and p
2n
+l = P for all n. Thus fails to
converge for all i, j.
Let us call a state i aperiodic if > 0 for all sufficiently large n. We
leave it as an exercise to show that i is aperiodic if and only if the set
{n 2:: 0 : > O} has no common divisor other than 1. This is also
a consequence of Theorem 1.8.4. The behaviour of the chain in Example
1.8.1 is connected with its periodicity.
1.8 Convergence to equilibrium 41
Lemma 1.8.2. Suppose P is irreducible and has an aperiodic state i.
Then, for all states j and k, > 0 for all sufflciently large n. In particular,
all states are aperiodic.
P
,I Th · > 0 · h (r) (s) 0 Th
rooJ. ere eXIst r, S _ WIt Pji' Pik >. en
(r+n+s) > (r) (n) (s) > 0
Pjk  Pji Pii Pik
for all sufficiently large n. 0
Here is the main result of this section. The method of proof, by coupling
two Markov chains, is ingenious.
Theorem 1.8.3 (Convergence to equilibrium). Let P be irreducible
and aperiodic, and suppose that P has an invariant distribution 7r. Let A
be any distribution. Suppose that is Markov(A, P). Then
P(X
n
= j) 7rj as n 00 for all j.
In particular,
t 'lrj as n t 00 for all i, j.
Proof. We use a coupling argument. Let be Markov(7r, P) and
independent of Fix a reference state b and set
T = inf{n 2:: 1 : X
n
= Y
n
= b}.
Step 1. We show P(T < 00) = 1. The process W
n
= (X
n
, Y
n
) is a Markov
chain on I x I with transition probabilities
P(i,k)(j,l) = PijPkl
and initial distribution
jj(i,k) = Ai
7r
k.
Since P is aperiodic, for all states i, j, k, l we have
(n) (n) 0
P(i,k)(j,l) = Pij Pkl >
for all sufficiently large n; so P is irreducible. Also, P has an invariant
distribution given by
7r(i,k) = 7ri7rk
so, by Theorem 1.7.7, P is positive recurrent. But T is the first passage
time of W
n
to (b, b) so P(T < 00) = 1, by Theorem 1.5.7.
42 1. Discretetime Markov chains
Step 2. Set
ifn < T
if n T.
The diagram below illustrates the idea. We show that is
Markov(;\, P).
I
n
The strong Markov property applies to at time T, so
(XT+n, is Markov(8(b,b), P) and independent of (X
o
, Yo),
(XI, YI), . .. , (XT, YT ) . By symmetry, we can replace the process
(XT+n, by (Y
T
+
n
, which is also Markov(8(b,b), P) and
remains independent of (X
o
, Yo), (Xl, YI), ... ,(X
T
, YT). Hence =
(Zn, is Markov(j.t, P) where
, { Y
n
Z =
n X
n
ifn < T
if n 2 T.
In particular, is Markov(;\, P).
Step 3. We have
lP(Zn = j) = lP(X
n
= j and n < T) +lP(Y
n
= j and n 2 T)
so
IlP(X
n
= j)  7rjl = IlP(Zn = j) lP(Y
n
= j)1
= IlP(X
n
= j and n < T) lP(Y
n
= j and n < T)I
lP(n < T)
and P(n < T) 0 as n 00. D
1.8 Convergence to equilibrium 43
To understand this proof one should see what goes wrong when P is
not aperiodic. Consider the twostate chain of Example 1.8.1 which has
(1/2, 1/2) as its unique invariant distribution. We start from 0
and with equal probability from 0 or 1. However, if Yo = 1, then,
because of periodicity, and will never meet, and the proof
fails. We move on now to the cases that were excluded in the last theorem,
where is periodic or transient or null recurrent. The remainder of
this section might be omitted on a first reading.
Theorem 1.8.4. Let P be irreducible. There is an integer d 1 and a
partition
I = Co U C
1
U ... U Cd1
such that (setting Cnd+r = Cr)
(i) > 0 only if i E C
r
and j E C
r
+
n
for some r;
(ii) > 0 for all sufficiently large n, for all i,j E C
r
, for all r.
Proof. Fix a state k and consider S = {n 0 : > O}. Choose n1, n2 E S
with n1 < n2 and such that d := n2  n1 is as small as possible. (Here and
throughout we use the symbol := to mean 'defined to equal'.) Define for
r = 0, ... ,d1
C {
. I (nd+r) £ O}
r = E : Pki > 0 or some n .
Then Co U ... U Cd1 = I, by irreducibility. Moreover, if > 0 and
> 0 for some r, S E {O, 1, ... ,d  I}, then, choosing m 0 so that
> 0, we have > 0 and > 0 so r = s by minimality
of d. Hence we have a partition.
To prove (i) suppose > 0 and i E Cr. Choose m so that > 0,
then > 0 so j E C
r
+
n
as required. By taking i = j = k we now
see that d must divide every element of S, in particular n1.
Now for nd we can write nd = qn1 + r for integers q n1 and
o r n1  1. Since d divides n1 we then have r = md for some integer
m and then nd = (q  m)n1 +mn2. Hence
and hence nd E S. To prove (ii) for i, j E C
r
choose m1 and m2 so that
> 0 and > 0, then
44 1. Discretetime Markov chains
whenever nd Since ml +m2 is then necessarily a multiple of d, we
are done. D
We call d the period of P. The theorem just proved shows in particular for
all i E I that d is the greatest common divisor of the set {n 0 : > O}.
This is sometimes useful in identifying d.
Finally, here is a complete description of limiting behaviour for irre
ducible chains. This generalizes Theorem 1.8.3 in two respects since we
require neither aperiodicity nor the existence of an invariant distribution.
The argument we use for the null recurrent case was discovered recently by
B. Fristedt and L. Gray.
Theorem 1.8.5. Let P be irreducible of period d and let Co, C
1
, ... ,Cdl
be the partition obtained in Theorem 1.8.4. Let A be a distribution with
LiEC
o
Ai = 1. Suppose that is Markov(A, P). Then for r =
0,1, ... ,d  1 and j E C
r
we have
P(X
nd
+
r
= j) dlmj as n 00
where mj is the expected return time to j. In particular, for i E Co and
j E C
r
we have
Proof
(nd+r) dl
Pij mj
as n 00.
Step 1. We reduce to the aperiodic case. Set v = Apr, then by Theorem
1.8.4 we have
Set Y
n
= Xnd+r, then is Markov(v, pd) and, by Theorem 1.8.4, pd
is irreducible and aperiodic on Cr. For j E C
r
the expected return time of
to j is mjld. So if the theorem holds in the aperiodic case, then
P(Xnd+r = j) = P(Y
n
= j) dlmj as n 00
so the theorem holds in general.
Step 2. Assume that P is aperiodic. If P is positive recurrent then 1/mj =
1rj, where 1r is the unique invariant distribution, so the result follows from
Theorem 1.8.3. Otherwise mj = 00 and we have to show that
P(X
n
= j) 0 as n 00.
1.8 Convergence to equilibrium 45
If P is transient this is easy and we are left with the null recurrent
case.
Step 3. Assume that P is aperiodic and null recurrent. Then
00
:EPj(Tj > k) = lEj(Tj) = 00.
k=O
Given € > 0 choose K so that
Then, for n K  1
n
1 :E P(Xk = j and X
m
1= j for m = k + 1, ... ,n)
k=nK+l
n
L P(X
k
= j)Pj(Tj > n  k)
k=nK+l
Kl
= :E P(Xnk = j)Pj(Tj > k)
k=O
so we must have P(X
n

k
= j) €/2 for some k E {O, 1, ... ,K  I}.
Return now to the coupling argument used in Theorem 1.8.3, only now let
be Markov(j.t, P), where j.t is to be chosen later. Set W
n
= (X
n
, Y
n
).
As before, aperiodicity of ensures irreducibility of If
is transient then, on taking j.t = A, we obtain
P(X
n
= j)2 = P(W
n
= (j, j)) 0
as required. Assume then that is recurrent. Then, in the notation
of Theorem 1.8.3, we have P(T < 00) = 1 and the coupling argument shows
that
IP(X
n
= j)  P(Y
n
= j)1 0 as n 00.
We exploit this convergence by taking j.t = Ap
k
for k = 1, ... ,K 1, so
that P(Y
n
= j) = P(X
n
+
k
= j). We can find N such that for n 2:: Nand
k = 1, ... ,K 1,
IP(Xn = j)  P(Xn+k = j)1 ::; ·
46 1. Discretetime Markov chains
But for any n we can find k E {O, 1, ... ,K  I} such that P(X
n
+
k
= j) ~
c/2. Hence, for n ~ N
P(X
n
= j) ~ c.
Since c > 0 was arbitrary, this shows that lP(X
n
= j) ~ 0 as n ~ 00, as
required. D
Exercises
1.8.1 Prove the claims (e), (f) and (g) made in example (v) of the Intro
duction.
1.8.2 Find the invariant distributions of the transition matrices in Exercise
1.1.7, parts (a), (b) and (c), and compare them with your answers there.
1.8.3 A fair die is thrown repeatedly. Let X
n
denote the sum of the first n
throws. Find
lim P(X
n
is a multiple of 13)
n+oo
quoting carefully any general theorems that you use.
1.8.4 Each morning a student takes one of the three books he owns from
his shelf. The probability that he chooses book i is Qi, where 0 < Qi < 1 for
i = 1,2,3, and choices on successive days are independent. In the evening
he replaces the book at the lefthand end of the shelf. If Pn denotes the
probability that on day n the student finds the books in the order 1,2,3,
from left to right, show that, irrespective of the initial arrangement of the
books, Pn converges as n ~ 00, and determine the limit.
1.8.5 (Renewal theorem). Let Y
1
, Y
2
, ••• be independent, identically
distributed random variables with values in {I, 2, ... }. Suppose that the
set of integers
{n : P(Y
1
= n) > I}
has greatest common divisor 1. Set Jl = E(Y
1
). Show that the following
process is a Markov chain:
X
n
= inf{m ~ n: m = Y
1
+ ... +Y
k
for some k ~ O}  n.
Determine
lim lP(X
n
= 0)
n+oo
and hence show that as n ~ 00
P(n = Y
1
+ ... +Y
k
for some k ~ 0) ~ 1/Jl.
1.9 Time reversal 47
(Think of Y
I
, Y
2
, • •• as lightbulb lifetimes. A bulb is replaced when it fails.
Thus the limiting probability that a bulb is replaced at time n is 1/Jl. Al
though this appears to be a very special case of convergence to equilibrium,
one can actually recover the full result by applying the renewal theorem to
the excursion lengths ... from state i.)
1.9 Time reversal
For Markov chains, the past and future are independent given the present.
This property is symmetrical in time and suggests looking at Markov chains
with time running backwards. On the other hand, convergence to equilib
rium shows behaviour which is asymmetrical in time: a highly organised
state such as a point mass decays to a disorganised one, the invariant dis
tribution. This is an example of entropy increasing. It suggests that if
we want complete timesymmetry we must begin in equilibrium. The next
result shows that a Markov chain in equilibrium, run backwards, is again a
Markov chain. The transition matrix may however be different.
Theorem 1.9.1. Let P be irreducible and have an invariant distribution
1r. Suppose that is Markov(1T', P) and set Y
n
= X
N

n
. Then
is MarkovCrr, P), where P= (Pij) is given by
1rjPji = 1riPij for all i, j
and P is also irreducible with invariant distribution 1r.
Proof. First we check that Pis a stochastic matrix:
since 1r is invariant for P. Next we check that 1r is invariant for P:
L 1rjPji = L 1riPij = 1ri
JEI JEI
since P is a stochastic matrix.
We have
P(YO= io, Y
I
= i
l
,··· ,YN = iN)
= P(Xo = iN, Xl = iNI,·.· ,XN = io)
= 1riNPiNiN1 ... P i 1io = 1rioPioi1 ... PiN1iN
48 1. Discretetime Markov chains
so, by Theorem 1.1.1, is Markov(7r, P). Finally, since P is
irreducible, for each pair of states i, j there is a chain of states io =
i,il, ... ,in1,i
n
=j withpioil ... Pinli
n
> O. Then
Pininl ... Pil io = 7rioPioil ... Pinl in /
7r
i
n
> 0
so Pis also irreducible. D
The chain is called the timereversal of
A stochastic matrix P and a measure A are said to be in detailed balance
if
AiPij = AjPji for all i, j.
Though obvious, the following result is worth remembering because, when
a solution A to the detailed balance equations exists, it is often easier to
find by the detailed balance equations than by the equation A = AP.
Lemma 1.9.2. If P and A are in detailed balance, then A is invariant for
P.
Proof· We have (AP)i = LjE! AjPji = LjE! AiPij = Ai· D
Let be Markov(A, P), with P irreducible. We say that
is reversible if, for all N 1, is also Markov(A, P).
Theorem 1.9.3. Let P be an irreducible stochastic matrix and let A be
a distribution. Suppose that is Markov(A, P). Then the following
are equivalent:
(a) is reversible;
(b) P and A are in detailed balance.
Proof. Both (a) and (b) imply that A is invariant for P. Then both (a) and
(b) are equivalent to the statement that P= P in Theorem 1.9.1. D
We begin a collection of examples with a chain which is not reversible.
Example 1.9.4
Consider the Markov chain with diagram:
1
3
2
3
2
The transition matrix is
1.9 Time reversal
(
0 2/3 1/3)
p= 1/3 0 2/3
2/3 1/3 0
49
and 1r = (1/3, 1/3, 1/3) is invariant. Hence P = pT, the transpose of P.
But p is not symmetric, so P =I P and this chain is not reversible. A
patient observer would see the chain move clockwise in the long run: under
timereversal the clock would run backwards!
Example 1.9.5
Consider the Markov chain with diagram:
P
. ~
o 1
q P
I I ( . ~
iI i i+l
q
• II( •
Ml M
where 0 < P = 1  q < 1. The nonzero detailed balance equations read
AiPi,i+l = Ai+lPi+l,i for i = 0, 1, ... ,M  1.
So a solution is given by
A= ((p/q)i : i = 0, 1, ... ,M)
and this may be normalised to give a distribution in detailed balance with
P. Hence this chain is reversible.
If P were much larger than q, one might argue that the chain would tend
to move to the right and its timereversal to the left. However, this ignores
the fact that we reverse the chain in equilibrium, which in this case would
be heavily concentrated near M. An observer would see the chain spending
most of its time near M and making occasional brief forays to the left,
which behaviour is symmetrical in time.
Example 1.9.6 (Random walk on a graph)
A gr.aph G is a countable collection of states, usually called vertices, some
of which are joined by edges, for example:
50 1. Discretetime Markov chains
1
4
2
3
Thus a graph is a partially drawn Markov chain diagram. There is a natural
way to complete the diagram which gives rise to the random walk on G.
The valency Vi of vertex i is the number of edges at i. We have to assume
that every vertex has finite valency. The random walk on G picks edges
with equal probability:
1
1
2
1
3
4
1
2
1
3
1
3
1
2
2
1
3
1
2
3
Thus the transition probabilities are given by
if (i, j) is an edge
otherwise.
We assume G is connected, so that P is irreducible. It is easy to see that
P is in detailed balance with V == (Vi : i E G). So, if the total valency
a == LiEG Vi is finite, then 1r == V / a is invariant and P is reversible.
Example 1.9.7 (Random chessboard knight)
A random knight makes each permissible move with equal probability. If it
starts in a corner, how long on average will it take to return?
This is an example of a random walk on a graph: the vertices are the
squares of the chessboard and the edges are the moves that the knight can
take:
1.9 Time reversal 51
The diagram shows a part of the graph. We know by Theorem 1.7.7 and
the preceding example that
so all we have to do is identify valencies. The four corner squares have
valency 2, and the eight squares adjacent to the corners have valency 3.
There are 20 squares of valency 4, 16 of valency 6, and the 16 central
squares have valency 8. Hence
lEc(T
c
) = 8 + 24 + 80 + 96 + 128 = 168.
2
Alternatively, if you enjoy solving sets of 64 simultaneous linear equations,
you might try finding 1r from 1rP == 1r, or calculating lEe (T
e
) using Theorem
1.3.5!
Exercises
1.9.1 In each of the following cases determine whether the stochastic matrix
P, which you may assume is irreducible, is reversible:
(a)
p ) .
lq ,
(b)
(c) 1== {O,l, ... ,N} a n d p ~ J == 0 if Ij il2:: 2;
52 1. Discretetime Markov chains
(d) I = {a, 1,2, ... } and POI = 1, Pi,i+l = P, Pi,il = 1  P for i 2:: 1;
(e) Pij = Pji for all i,j E S.
1.9.2 Two particles X and Y perform independent random walks on the
graph shown in the diagram. So, for example, a particle at A jumps to B,
C or D with equal probability 1/3.
D p_.......
B
C ~    " I I
E
Find the probability that X and Y ever meet at a vertex in the following
cases:
(a) X starts at A and Y starts at B;
(b) X starts at A and Y starts at E. For I = B, DIet MI denote the
expected time, when both X and Y start at I, until they are once
again both at I. Show that 9MD = 16MB.
1.10 Ergodic theorem
Ergodic theorems concern the limiting behaviour of averages over time.
We shall prove a theorem which identifies for Markov chains the longrun
proportion of time spent in each state. An essential tool is the following
ergodic theorem for independent random variables which is a version of the
strong law of large numbers.
Theorem 1.10.1 (Strong law of large numbers). Let Y
I
, Y
2
, ... be
a sequence of independent, identically distributed, nonnegative random
1.10 Ergodic theorem
variables with JE(Y
1
) = jj. Then
11']) (Yl + ... +Y
n
) 1
c = .
n
53
Proof. A proof for the case J_l < 00 may be found, for example, in Probability
with Martingales by David Williams (Cambridge University Press, 1991).
The case where J_l = 00 is a simple deduction. Fix N < 00 and set yJN) =
YnAN. Then
(N) (N)
Y
1
+ ···+Y
n
> Y
1
+ ···+Y
n
_ t E(Y
1
/\ N)
n n
with probability one. As N i 00 we have E(YI A N) i J_l by monotone
convergence (see Section 6.4). So we must have, with probability 1
Y
1
+ ... + Y
n
n
as n 00.
D
We denote by Vi (n) the number of visits to i before n:
nl
Vi(n) = L l{xk=i}'
k=O
Then Vi (n )/ n is the proportion of time before n spent in state i. The
following result gives the longrun proportion of time spent by a Markov
chain in each state.
Theorem 1.10.2 (Ergodic theorem). Let P be irreducible and let .x
be any distribution. If (Xn)n?O is Markov(.x, P) then
]p> (Vi (n) t ...! as n t 00) = 1
n mi
where mi = Ei(T
i
) is the expected return time to state i. Moreover, in the
positive recurrent case, for any bounded function f : I lR we have
where
and where (7ri : i E I) is the unique invariant distribution.
54 1. Discretetime Markov chains
Proof. If P is transient, then, with probability 1, the total number Vi of
visits to i is finite, so
Vi(n) < Vi t 0 = ...!.
n  n mi
Suppose then that P is recurrent and fix a state i. For T = T
i
we have
P(T < 00) = 1 by Theorem 1.5.7 and is Markov(8
i
, P) and
independent of X
o
, Xl, . .. ,X
T
by the strong Markov property. The long
run proportion of time spent in i is the same for (XT+n)n>O and (Xn)n>O,
 
so it suffices to consider the case A = bi.
Write sir) for the length of the rth excursion to i, as in Section 1.5. By
Lemma 1.5.1, the nonnegative random variables S;l), 8;2), ... are indepen
dent and identically distributed with Ei(S;r)) = mi. Now
+ + <  1
't ••• 't _n,
the lefthand side being the time of the last visit to i before n. Also
the lefthand side being the time of the first visit to i after n  1. Hence
S;l) + ... + n S;l) + ... +
Vi(n) :::; Vi(n) < Vi(n) (1.8)
By the strong law of large numbers
+ + )
JP> • t mi as n t 00 = 1
and, since P is recurrent
P(Vi(n) 00 as n 00) = 1.
So, letting n 00 in (1.8), we get
JP> (Vi7n) t mi as n t 00) = 1,
which implies
P (Vi(n) t ...! as n t (0) = 1.
n mi
1.10 Ergodic theorem 55
Assume now that has an invariant distribution (1ri : i E I). Let
I : I lR be a bounded function and assume without loss of generality that
III 1. For any J I we have
I
nl I ( )
1  Vi(n)
L !(Xk) ! = n  1ri Ii
k=O 'tEl
L I  7ril + L I  7ril
iEJ
L I  7ril +L +7ri)
iEJ
" IVi(n) I " n 
1r
i
iEJ
We proved above that
JP> (Vi t 7ri as n t 00 for all i) = 1.
Given c > 0, choose J finite so that
and then N = N(w) so that, for n N(w)
"I Vi(n) I n  1ri < c/4.
iEJ
Then, for n N(w), we have
< c,
which establishes the desired convergence. D
We consider now the statistical problem of estimating an unknown tran
sition matrix P on the basis of observations of the corresponding Markov
chain. Consider, to begin, the case where we have N + 1 observations
The loglikelihood function is given by
I(P) = log(..\xoPxox
l
•• ,PXN1XN) = L Nij logpij
i,jEl
56 1. Discretetime Markov chains
up to a constant independent of P, where N
ij
is the number of transitions
from i to j. A standard statistical procedure is to find the maximum likeli
hood estimate P, which is the choice of P maximizing l(P). Since P must
satisfy the linear constraint E
j
Pij = 1 for each i, we first try to maximize
l (P) + 2: /LiPij
i,jEI
and then choose (/li : i E I) to fit the constraints. This is the method of
Lagrange multipliers. Thus we find
Nl Nl
Pij = L 1{Xn =i,X
n
+
1
=j}/ L l{x
n
=i}
n=O n=O
which is the proportion of jumps from i which go to j.
We now turn to consider the consistency of this sort of estimate, that is
to say whether Pij ~ Pij with probability 1 as N ~ 00. Since this is clearly
false when i is transient, we shall slightly modify our approach. Note that
to find Pij we simply have to maximize
2: Nij logPij
jEI
subject to E
j
Pij = 1: the other terms and constraints are irrelevant. Sup
pose then that instead of N + 1 observations we make enough observations
to ensure the chain leaves state i a total of N times. In the transient case
this may involve restarting the chain several times. Denote again by N
ij
the number of transitions from i to j.
To maximize the likelihood for (Pij : j E I) we still maximize
2: Nij logPij
jEI
subject to E
j
Pij = 1, which leads to the maximum likelihood estimate
Pij = Nij/N.
But Nij = Y
1
+ ... +Y
N
, where Y
n
= 1 if the nth transition from i is to
j, and Y
n
= 0 otherwise. By the strong Markov property Yl, . .. ,YN are
independent and identically distributed random variables with mean Pij.
So, by the strong law of large numbers
P(Pij ~ Pij as N ~ 00) = 1,
which shows that Pij is consistent.
Exercises
1.11 Appendix: recurrence relations 57
1.10.1 Prove the claim (d) made in example (v) of the Introduction.
1.10.2 A professor has N umbrellas. He walks to the office in the morning
and walks home in the evening. If it is raining he likes to carry an um
brella and if it is fine he does not. Suppose that it rains on each journey
with probability p, independently of past weather. What is the longrun
proportion of journeys on which the professor gets wet?
1.10.3 Let ( X n ) n ~ O be an irreducible Markov chain on I having an invariant
distribution Jr. For J ~ I let ( Y m ) m ~ O be the Markov chain on J obtained
by observing ( X n ) n ~ O whilst in J. (See Example 1.4.4.) Show that ( Y m ) m ~ O
is positive recurrent and find its invariant distribution.
1.10.4 An opera singer is due to perform a long series of concerts. Hav
ing a fine artistic temperament, she is liable to pullout each night with
probability 1/2. Once this has happened she will not sing again until the
promoter convinces her of his high regard. This he does by sending flowers
every day until she returns. Flowers costing x thousand pounds, 0 ~ x ~ 1,
bring about a reconciliation with probability yIX. The promoter stands to
make £750 from each successful concert. How much should he spend on
flowers?
1.11 Appendix: recurrence relations
Recurrence relations often arise in the linear equations associated to Markov
chains. Here is an account of the simplest cases. A more specialized case
was dealt with in Example 1.3.4. In Example 1.1.4 we found a recurrence
relation of the form
Xn+l = aXn +b.
We look first for a constant solution X
n
= x; then x = ax +b, so provided
a =I 1 we must have x = b/(l  a). Now Yn = X
n
 b/(l  a) satisfies
Yn+l = aYn, so Yn = anyo. Thus the general solution when a =I 1 is given
by
X
n
= Aan +b/ (1  a)
where A is a constant. When a = 1 the general solution is obviously
X
n
= Xo +nb.
In Example 1.3.3 we found a recurrence relation of the form
aXn+l +bXn +CXnl = 0
58 1. Discretetime Markov chains
where a and c were both nonzero. Let us try a solution of the form X
n
= Xn;
then aX
2
+ bX + c = O. Denote by 0: and {3 the roots of this quadratic. Then
Yn = Ao:
n
+B{3n
is a solution. If 0: =1= (3 then we can solve the equations
xo=A+B, xl=Ao:+B{3
so that Yo = Xo and Yl = Xl; but
for all n, so by induction Yn = X
n
for all n. If 0: = {3 # 0, then
Yn = (A +nB)o:n
is a solution and we can solve
so that Yo = Xo and Yl = Xl; then, by the same argument, Yn = X
n
for all
n. The case 0: = (3 = 0 does not arise. Hence the general solution is given
by
{
Ao:n + B{3n
X
n
= (A + nB)a
n
if 0: # {3
if 0: = {3.
1.12 Appendix: asymptotics for n!
Our analysis of recurrence and transience for random walks in Section 1.6
rested heavily on the use of the asymptotic relation
n! rv Ayri(nje)n as n ~ 00
for some A E [1, 00). Here is a derivation.
We make use of the power series expansions for It I < 1
10g(1 +t) = t  ~ t 2 + ~ t 3  .
10g(1  t) = t  ~ t 2  ~ t 3  .
By subtraction we obtain
1 (l+t) 13 15
2 log 1 _ t = t + 3t +:5t +....
1.12 Appendix: asymptotics for n! 59
Set An = n!/(nn+l/2en) and an = logAn. Then, by a straightforward
calculation
1 (1+(2n+l)1)
an  an+! = (2n + 1)2" log 1 _ (2n + 1)1  1.
By the series expansion written above we have
{
Ill II}
anan+! = (2n+1) (2n+1) +3"(2n+1)3 +5 (2n+1)5 + ... 1
1 1 1 1
= 3 (2n + 1)2 + 5 (2n + 1)4 + · ..
1{II}
::; 3" (2n + 1)2 + (2n + 1)4 + · · ·
1 1 1 1
3 (2n + 1)2  1 12n  12(n + 1) .
It follows that an decreases and an  1/(12n) increases as n ~ 00. Hence
an ~ a for some a E [0, 00) and hence An ~ A, as n ~ 00, where A = ea.
2
Continuoustime Markov chains I
The material on continuoustime Markov chains is divided between this
chapter and the next. The theory takes some time to set up, but once up
and running it follows a very similar pattern to the discretetime case. To
emphasise this we have put the settingup in this chapter and the rest in the
next. If you wish, you can begin with Chapter 3, provided you take certain
basic properties on trust, which are reviewed in Section 3.1. The first three
sections of Chapter 2 fill in some necessary background information and are
independent of each other. Section 2.4 on the Poisson process and Section
2.5 on birth processes provide a gentle warmup for general continuous
time Markov chains. These processes are simple and particularly important
examples of continuoustime chains. Sections 2.62.8, especially 2.8, deal
with the heart of the continuoustime theory. There is an irreducible level
of difficulty at this point, so we advise that Sections 2.7 and 2.8 are read
selectively at first. Some examples of more general processes are given in
Section 2.9. As in Chapter 1 the exercises form an important part of the
text.
2.1 Qmatrices and their exponentials
In this section we shall discuss some of the basic properties of Qmatrices
and explain their connection with continuoustime Markov chains.
Let I be a countable set. A Qmatrix on I is a matrix Q = (qij : i,j E I)
satisfying the following conditions:
2.1 Qmatrices and their exponentials 61
(i) 0 ~ qii < 00 for all i;
(ii) qij ~ 0 for all i =I j;
(iii) L qij = 0 for all i.
jEI
Thus in each row of Q we can choose the offdiagonal entries to be any non
negative real numbers, subject only to the constraint that the offdiagonal
row sum is finite:
qi = Lqij < 00.
j#i
The diagonal entry qii is then qi, making the total row sum zero.
A convenient way to present the data for a continuoustime Markov chain
is by means of a diagram, for example:
1
3
1
2
Each diagram then corresponds to a unique Qmatrix, in this case
Q = ( ~ 2 !1 ~ )
2 1 3
Thus each offdiagonal entry qij gives the value we attach to the (i, j) arrow
on the diagram, which we shall interpret later as the rate of going from i to
j. The numbers qi are not shown on the diagram, but you can work them
out from the other information given. We shall later interpret qi as the rate
of leaving i.
We may think of the discrete parameter space {O, 1,2, ... } as embedded
in the continuous parameter space [0,00). For P E (0,00) a natural way to
interpolate the discrete sequence (pn : n = 0,1,2, ... ) is by the function
(e
tq
: t ~ 0), where q = logp. Consider now a finite set I and a matrix
62 2. Continuoustime Markov chains I
P = (Pij : i,j E I). Is there a natural way to fill in the gaps in the discrete
sequence (p
n
: n = 0,1,2, ... )?
For any matrix Q = (qij : i, j E I), the series
converges componentwise and we denote its limit by e
Q
. Moreover, if two
matrices Q1 and Q2 commute, then
The proofs of these assertions follow the scalar case closely and are given
in Section 2.10. Suppose then that we can find a matrix Q with e
Q
= P.
Then
e
nQ
= (eQ)n = p
n
so (e
tQ
: t ~ 0) fills in the gaps in the discrete sequence.
Theorem 2.1.1. Let Q be matrix on a finite set I. Set P(t) = e
tQ
. Then
(P(t) : t ~ 0) has the following properties:
(i) P(s +t) = P(s)P(t) for all s, t (semigroup property);
(ii) (P(t) : t ~ 0) is the unique solution to the forward equation
~ P(t) = P(t)Q,
P(O) = I;
(iii) (P(t) : t ~ 0) is the unique solution to the backward equation
P(O) = I;
(iv) for k = 0,1,2, ... , we have
Proof. For any s, t E lR, sQ and tQ commute, so
e
sQ
e
tQ
= e(s+t)Q
proving the semigroup property. The matrixvalued power series
P(t) = f: ( t ~ ) k
k=O
2.1 Qmatrices and their exponentials 63
has infinite radius of convergence (see Section 2.10). So each component is
differentiable with derivative given by termbyterm differentiation:
, 00 t
k

1
Qk
P (t) = L (k _ 1)! = P(t)Q = QP(t).
k=1
Hence P(t) satisfies the forward and backward equations. Moreover by
repeated termbyterm differentiation we obtain (iv). It remains to show
that P(t) is the only solution of the forward and backward equations. But
if M(t) satisfies the forward equation, then
.!!.(M(t)e
tQ
) = (.!!.M(t)) e
tQ
+ M(t) (.!!.e
tQ
)
dt dt dt
= M(t)Qe
tQ
+ M(t)( Q)e
tQ
= 0
so M(t)e
tQ
is constant, and so M(t) = P(t). A similar argument proves
uniqueness for the backward equation. D
The last result was about matrix exponentials in general. Now let us see
what happens to Qmatrices. Recall that a matrix P = (Pij : i, j E I) is
stochastic if it satisfies
(i) 0 ~ Pij < 00 for all i,j;
(ii) LPij = 1 for all i.
jEI
We recall the convention that in the limit t ~ 0 the statement f(t) = O(t)
means that f(t)/t ~ C for all sufficiently small t, for some C < 00. Later
we shall also use the convention that f(t) = o(t) means f(t)/t ~ 0 as t ~ o.
Theorem 2.1.2. A matrix Q on a finite set I is a Qmatrix if and only if
P(t) = e
tQ
is a stochastic matrix for all t ~ o.
Proof. As t ! 0 we have
so qij ~ 0 for i =1= j if and only if Pij (t) ~ 0 for all i, j and t ~ 0 sufficiently
small. Since P(t) = P(t/n)n for all n, it follows that qij ~ 0 for i =1= j if
and only if Pij(t) ~ 0 for all i,j and all t ~ o.
If Q has zero row sums then so does Qn for every n:
""" (n) """ """ (n1) """ (n1) """ 0
LJ qik = LJ LJ qij qjk = LJ qij LJ qjk = .
kEI kEI jEI jEI kEI
64
So
2. Continuoustime Markov chains I
00 t
n
(n)
LPij(t) = 1 + L I" Lqij = 1.
n.
jEI n=l jEI
On the other hand, if LjE! Pij(t) = 1 for all t 0, then
L qij = I LPij(t) = O.
jEI t=O jEI
D
Now, if P is a stochastic matrix of the form e
Q
for some Qmatrix, we
can do some sort of fillingin of gaps at the level of processes. Fix some
large integer m and let be discretetime Markov(.x, e
Q
/
m
). We
define a process indexed by {n/m : n = 0,1,2, ... } by
Then (X
n
: n = 0,1,2, ... ) is discretetime Markov(.x, (eQ/m)m) (see Exer
cise 1.1.2) and
Thus we can find discretetime Markov chains with arbitrarily fine
grids {n/m : n = 0,1,2, ... } as timeparameter sets which give rise to
Markov(.x, P) when sampled at integer times. It should not then be too
surprising that there is, as we shall see in Section 2.8, a continuoustime
process which also has this property.
To anticipate a little, we shall see in Section 2.8 that a continuoustime
Markov chain with Qmatrix Q satisfies
for all n = 0,1,2, ... , all times ° to ... t
n
+l and all states
io, ... ,i
n
+
1
, where Pij(t) is the (i,j) entry in e
tQ
. In particular, the tran
sition probability from i to j in time t is given by
(Recall that := means 'defined to equal'.) You should compare this with
the defining property of a discretetime Markov chain given in Section 1.1.
We shall now give some examples where the transition probabilities Pij(t)
may be calculated explicitly.
2.1 Qmatrices and their exponentials 65
Example 2.1.3
We calculate PII(t) for the continuoustime Markov chain with Qmatrix
Q = ( ~ 2 !1 ~ )
2 1 3
The method is similar to that of Example 1.1.6. We begin by writing down
the characteristic equation for Q:
o= det (x  Q) = x(x +2) (x +4).
This shows that Q has distinct eigenvalues 0, 2, 4. Then PII(t) has the
form
PII(t) = a +be
2t
+ce
4t
for some constants a, band c. (This is because we could diagonalize Q by
an invertible matrix U:
Then
00)
(2t)k 0
o (4t)k
~ ) U
I
,
e
4t
U
I
so PII(t) must have the form claimed.) To determine the constants we use
1 = PII (0) = a +b+c,
2 = qII = P ~ I (0) = 2b  4c,
7 = qii) = P ~ I (0) = 4b +16c,
so
66 2. Continuoustime Markov chains I
A A A
... ..... .. ... ..
o 1 2 Nl N
Q=
Example 2.1.4
We calculate Pij(t) for the continuoustime Markov chain with diagram
given above. The Qmatrix is
A A
A A
A
A A
o
where entries off the diagonal and superdiagonal are all zero. The expo
nential of an uppertriangular matrix is uppertriangular, so Pij(t) = 0 for
i > j. In components the forward equation P'(t) = P(t)Q reads
= APii(t),
= APij(t) + APi,jl (t),
= APiNl(t),
Pii (0) == 1,
Pij (0) == 0,
PiN(O) = 0,
for i < N,
for i < j < N,
for i < N.
We can solve these equations. First, pii(t) = e
At
for i < N. Then, for
i <j < N
so, by induction
(At)ji
Pij(t) = e>.t (j _ i)!'
If i = 0, these are the Poisson probabilities of parameter At. So, start
ing from 0, the distribution of the Markov chain at time t is the same as
the distribution of min{yt, N}, where yt is a Poisson random variable of
parameter At.
Exercises
2.1.1 Compute Pll(t) for P(t) = e
tQ
, where
Q = !4
2 1 3
2.2 Continuoustime random processes
2.1.2 Which of the following matrices is the exponential of a Qmatrix?
67
(a) ( ~ ~ ) (b) ( ~ ~ ) (c) ( ~ ~ ) .
What consequences do your answers have for the discretetime Markov
chains with these transition matrices?
2.2 Continuoustime random processes
Let I be a countable set. A continuoustime random process
with values in I is a family of random variables X
t
: n ~ I. We are going
to consider ways in which we might specify the probabilistic behaviour (or
law) of ( X t ) t ~ o . These should enable us to find, at least in principle,
any probability connected with the process, such as lP(X
t
= i) or
lP(X
to
= io, ... ,X
tn
= in), or P(X
t
= i for some t). There are subtleties in
this problem not present in the discretetime case. They arise because, for
a countable disjoint union
whereas for an uncountable union Ut>o At there is no such rule. To avoid
these subtleties as far as possible we shall restrict our attention to processes
( X t ) t ~ O which are rightcontinuous. This means in this context that for all
wEn and t ~ 0 there exists € > 0 such that
for t ~ s ~ t + €.
By a standard result of measure theory, which is proved in Section 6.6,
the probability of any event depending on a rightcontinuous process can
be determined from its finitedimensional distributions, that is, from the
probabilities
lP(X
to
= io, X
t1
= il, ,X
tn
= in)
for n ~ 0, 0 ~ to ~ tl ~ ... ~ t
n
and i
o
, ,in E I. For example
P(X
t
= i for some t E [0,00)) = 1 lim '""" P(X
q1
= ji, ... ,X
qn
= jn)
n+oo L.J
jl , ... ,in=l=i
where ql, q2, . .. is an enumeration of the rationals.
68 2. Continuoustime Markov chains I
Every path t ~ Xt(w) of a rightcontinuous process must remain con
stant for a while in each new state, so there are three possibilities for the
sorts of path we get. In the first case the path makes infinitely many jumps,
but only finitely many in any interval [0, t]:
•
o
•
•
o
~ 
J
o
= 0
1.1
t
The second case is where the path makes finitely many jumps and then
becomes stuck in some state forever:
•
J
o
= 0
J.2
t
In the third case the process makes infinitely many jumps in a finite interval;
this is illustrated below. In this case, after the explosion time ( the process
starts up again; it may explode again, maybe infinitely often, or it may
not.
2.2 Continuoustime random processes
...
:.
69
•
"
. ...
.......
• •
....
J
o
= 0
..........e.
t
..
14....._....
We call J
o
, J
1
, ... the jump times of and 8
1
,8
2
, ... the holding
times. They are obtained from by
for n = 0,1, ... , where inf 0= 00, and, for n = 1,2, ... ,
if I
n
1 < 00
otherwise.
Note that rightcontinuity forces 8
n
> 0 for all n. If I
n
+
1
= 00 for some
n, we define X
oo
= XJ
n
, the final value, otherwise X
oo
is undefined. The
(first) explosion time ( is defined by
00
(= supJ
n
= I:Sn.
n n==1
The discretetime process given by Y
n
= X
Jn
is called the jump
process of or the jump chain if it is a discretetime Markov chain.
This is simply the sequence of values taken by up to explosion.
We shall not consider what happens to a process after explosion. So it
is convenient to adjoin to I a new state, 00 say, and require that X
t
= 00
if t (. Any process satisfying this requirement is called minimal. The
terminology 'minimal' does not refer to the state of the process but to the
70 2. Continuoustime Markov chains I
interval of time over which the process is active. Note that a minimal
process may be reconstructed from its holding times and jump process.
Thus by specifying the joint distribution of 8
1
,82, ... and we have
another 'countable' specification of the probabilistic behaviour of
For example, the probability that X
t
= i is given by
00
IP(X
t
= i) = L IP(Y
n
= i and I
n
:::; t < I
n
+1)
n=O
and
lP(X
t
= i for some t E [0,00)) = lP(Y
n
= i for some n 0).
2.3 Some properties of the exponential distribution
A random variable T : n [0,00] has exponential distribution of parameter
X (0 X < 00) if
lP(T > t) = e
At
for all t O.
We write T rv E(X) for short. If X > 0, then T has density function
The mean of T is given by
E(T) = 1
00
IP(T > t)dt = AI.
The exponential distribution plays a fundamental role in continuoustime
Markov chains because of the following results.
Theorem 2.3.1 (Memoryless property). A random variable T: n
(0,00] has an exponential distribution if and only if it has the following
memoryless property:
lP(T> s +tiT> s) = lP(T > t) for all s, t O.
Proof. Suppose T rv E(X), then
P(T> s +t) eA(s+t) At
IP(T> s + tiT> s) = IP(T> s) = e>'s = e = IP(T > t).
2.3 Some properties of the exponential distribution 71
On the other hand, suppose T has the memoryless property whenever
JP>(T> s) > O. Then g(t) = JP>(T > t) satisfies
g(s +t) = g(s)g(t) for all s, t ~ O.
We assumed T > 0 so that g(l/n) > 0 for some n. Then, by induction
so g(l) = e
A
for some 0 ::; A < 00. By the same argument, for integers
p,q ~ 1
g(p/q) = g(l/q)P = g(l)pjq
so g(r) = e
AT
for all rationals r > O. For real t > 0, choose rationals
r, s > 0 with r ::; t ::; s. Since 9 is decreasing,
e
AT
= g(r) ~ g(t) ~ g(s) = e
AS
and, since we can choose rand s arbitrarily close to t, this forces g(t) = e
At
,
so T rv E(A). 0
The next result shows that a sum of independent exponential random
variables is either certain to be finite or certain to be infinite, and gives a cri
terion for deciding which is true. This will be used to determine whether or
not certain continuoustime Markov chains can take infinitely many jumps
in a finite time.
Theorem 2.3.2. Let 8
1
, 8
2
, . .. be a sequence ofindependent random vari
ables with 8
n
rv E(A
n
) and 0 < An < 00 for all n.
00 1 (00)
(i) I f ~ An < 00, then JP> ~ Sn < 00 = l.
00 1 (00)
(ii) If ~ An = 00, then JP> ~ Sn = 00 = l.
Proof. (i) Suppose E ~ 1 1/An < 00. Then, by monotone convergence
so
72 2. Continuoustime Markov chains I
(ii) Suppose instead that E ~ 1 1 / An = 00. Then r r ~ o ( 1 + 1/An) = 00.
By monotone convergence and independence
so
JP> (fSn = 00) = 1.
n==l
D
The following result is fundamental to continuoustime Markov chains.
Theorem 2.3.3. Let I be a countable set and let Tk' k E I, be independent
random variables with Tk rv E(qk) and 0 < q := EkE! qk < 00. Set
T = infk T
k
. Then this infimum is attained at a unique random value K of
k, with probability 1. Moreover, T and K are independent, with T rv E(q)
and lP(K = k) = qk/q.
Proof. Set K = k if T
k
< T
j
for all j =I k, otherwise let K be undefined.
Then
lP(K = k and T ~ t)
=lP(Tk ~ t and Tj > Tk for all j =I k)
= 1
00
qkeqkBP(Tj > s for all j 1= k)ds
= (OO qk
e

qkB
IT eqjBds
it j#
1
00
 8 qk  t
= qke q ds = e q .
t q
Hence lP(K = k for some k) = 1 and T and K have the claimed joint
distribution. D
The following identity is the simplest case of an identity used in Section
2.8 in proving the forward equations for a continuoustime Markov chain.
Theorem 2.3.4. For independent random variables S rv E(A) and R rv
E(jj) and for t ~ 0, we have
jjJP>(S ~ t < S +R) = AJP>(R ~ t < R+ S).
2.4 Poisson processes
Proof. We have
73
from which the identity follows by symmetry. D
Exercises
2.3.1 Suppose Sand T are independent exponential random variables of
parameters Q and (3 respectively. What is the distribution of min{S, T}?
What is the probability that S :::; T? Show that the two events {S < T}
and {min{S, T} ~ t} are independent.
2.3.2 Let Tl, T
2
, . .. be independent exponential random variables of pa
rameter A and let N be an independent geometric random variable with
JP>(N = n) = (3(1  (3)nl,
n = 1,2, ....
Show that T = 2::1 Ti has exponential distribution of parameter A(3.
2.3.3 Let 81, 8
2
, . .. be independent exponential random variables with
parameters AI, A2, . .. respectively. Show that AlSl is exponential of pa
rameter 1.
Use the strong law of large numbers to show, first in the special case
An = 1 for all n, and then subject only to the condition sUPn An < 00, that
JP> (fSn = 00) = 1 .
n=l
Is the condition sUPn An < 00 absolutely necessary?
2.4 Poisson processes
Poisson processes are some of the simplest examples of continuoustime
Markov chains. We shall also see that they may serve as building blocks
for the most general continuoustime Markov chain. Moreover, a Poisson
process is the natural probabilistic model for any uncoordinated stream of
discrete events in continuous time. So we shall study Poisson processes
first, both as a gentle warmup for the general theory and because they
are useful in themselves. The key result is Theorem 2.4.3, which provides
three different descriptions of a Poisson process. The reader might well
begin with the statement of this result and then see how it is used in the
74 2. Continuoustime Markov chains I
theorems and examples that follow. We shall begin with a definition in
terms of jump chain and holding times (see Section 2.2). A rightcontinuous
process with values in {O, 1,2, ... } is a Poisson process of rate A
(0 < A < 00) if its holding times 8
1
, 8
2
, . .. are independent exponential
random variables of parameter A and its jump chain is given by Y
n
= n.
Here is the diagram:
A A A A
.. . .. . . .. .
o 1 234
The associated Qmatrix is given by
A A
A A
Q=
By Theorem 2.3.2 (or the strong law of large numbers) we have
P(J
n
00) = 1 so there is no explosion and the law of is uniquely
determined. A simple way to construct a Poisson process of rate A is to
take a sequence 8
1
, 8
2
, . .. of independent exponential random variables of
parameter A, to set J
o
= 0, I
n
= 81 + ... + 8
n
and then set
. .
. . . ..
5 : : : : .. 
4 : : :....................... •'00········
3 ; :.... • 0,······ .: .
. . . ..
.. .
2 : : : .
. . ..
1 ........... .••ec;> : : : .
t
J.4
J
o
= (}
2.4 Poisson processes 75
The diagram illustrates a typical path. We now show how the memory
less property of the exponential holding times, Theorem 2.3.1, leads to a
memoryless property of the Poisson process.
Theorem 2.4.1 (Markov property). Let be a Poisson process
of rate X. Then, for any s 0, (X
s
+
t
 is also a Poisson process of
rate X, independent of(X
r
: r:::; s).
Proof. It to prove the claim conditional on the event X
s
= i, for
each i O. Set X
t
= X
s
+
t
 X
s
. We have
On this event
i
X r = L l{srSt} for r ::; s
j==1
and the holding times 8
1
,8
2
, ... of are given by
8
1
= 8
i
+
1
 (s  J
i
), 8
n
= 8
i
+
n
for n
as shown in the diagram.
o
ill(
s
Recall that the holding times 8
1
,8
2
, ... are independent E(A). Condition
on 8
1
, ... ,8
i
and {X
s
= i}, then by the memoryless property of 8
i
+
1
and independence, 81,8
2
, ... are themselves independent E(X). Hence,
conditional on {X
s
= i}, 8
1
,8
2
, ... are independent E(X), and independent
of 8
1
, ... ,8
i
. Hence, conditional on {X
s
= i}, is a Poisson process
of rate A and independent of (X
r
: r s). D
In fact, we shall see in Section 6.5, by an argument in essentially the
same spirit that the result also holds with s replaced by any stopping time
T of
76 2. Continuoustime Markov chains I
Theorem 2.4.2 (Strong Markov property). Let be a Poisson
process oErate A and let T be a stopping time Then, conditional
on T < 00, (X
T
+
t
 XT is also a Poisson process of rate A, independent
Here is some standard terminology. If is a realvalued process,
we can consider its increment X
t
 X
s
over any interval (s, t]. We say that
has stationary increments if the distribution of X
s
+
t
 X
s
depends
only on t O. We say that has independent increments if its
increments over any finite collection of disjoint intervals are independent.
We come to the key result for the Poisson process, which gives two condi
tions equivalent to the jump chain/holding time characterization which we
took as our original definition. Thus we have three alternative definitions
of the same process.
Theorem 2.4.3. Let be an increasing, rightcontinuous integer
valued process starting from O. Let 0 < A < 00. Then the following three
conditions are equivalent:
(a) (jump chain/holding time definition) the holding times 8
1
,8
2
, ... of
are independent exponential random variables of parameter
A and the jump chain is given by Y
n
= n for all n;
(b) (infinitesimal definition) has independent increments and, as
h ! 0, uniformly in t,
IF(X
t
+
h
 X
t
= 0) = 1  Ah +o(h), IF(X
t
+
h
 X
t
= 1) = Ah +o(h);
(c) (transition probability definition) has stationary independent
increments and, for each t, X
t
has Poisson distribution of parameter
At.
If satisfies any of these conditions then it is called a Poisson process
of rate A.
Proof. (a) => (b) If (a) holds, then, by the Markov property, for any t, h 0,
the increment X
t
+
h
 X
t
has the same distribution as X
h
and is independent
of (X
s
: s t). So has independent increments and as h ! 0
lP(Xt+h  X
t
1) = lP(X
h
1) = lP(J1 h) = 1  e)..h = Ah +o(h),
lP(Xt+h  X
t
2) = lP(X
h
2) = lP(J2 h)
IF(8
1
h and 8
2
h) = (1  e
Ah
)2 = o(h),
which implies (b).
2.4 Poisson processes 77
(b) ::::} (c) If (b) holds, then, for i = 2,3, ... , we have P(X
t
+
h
 X
t
= i) =
o(h) as h ! 0, uniformly in t. Set Pj(t) = P(X
t
= j). Ther:, for j = 1,2, ... ,
j
Pj(t + h) = P(X
Hh
= j) = L P(X
Hh
 X
t
= i) P(X
t
= j  i)
i==O
= (1  Ah +o(h) )Pj(t) + (Ah +o(h))Pjl (t) +o(h)
so
Pj(t +  Pj(t) = >'Pj(t) + >'Pjl(t) + O(h).
Since this estimate is uniform in t we can put t = s  h to obtain for all
s,?h
Now let h ! 0 to see that Pj(t) is first continuous and then differentiable
and satisfies the differential equation
By a simpler argument we also find
= APO(t).
Since X
o
= 0 we have initial conditions
PO(O) = 1, Pj(O) = 0 for j = 1,2, ....
As we saw in Example 2.1.4, this system of equations has a unique solution
given by
(At)j
Pj(t) = e
At
_.,_, j = 0,1,2, ....
J.
Hence X
t
rv P(At). If satisfies (b), then certainly has
independent increments, but also (X
s
+
t
 satisfies (b), so the above
argument shows X
s
+
t
 X
s
rv P(At), for any s, which implies (c).
(c) ::::} (a) There is a process satisfying (a) and we have shown that it must
then satisfy (c). But condition (c) determines the finitedimensional distri
butions of and hence the distribution of jump chain and holding
times. So if one process satisfying (c) also satisfies (a), so must every process
satisfying (c). D
The differential equations which appeared in the proof are really the
forward equations for the Poisson process. To make this clear, consider the
78 2. Continuoustime Markov chains I
possibility of starting the process from i at time 0, writing Pi as a reminder,
and set
Pij(t) = Pi(X
t
= j).
Then, by spatial homogeneity Pij(t) = Pji(t), and we could rewrite the
differential equations as
= APiO(t),
= APi,jl(t)  APij(t),
or, in matrix form, for Q as above,
PiO(O) = 8
iO
,
Pij(O) = 8ij
P'(t) = P(t)Q, P(O) = I.
Theorem 2.4.3 contains a great deal of information about the Poisson
process of rate A. It can be useful when trying to decide whether a given
process is a Poisson process as it gives you three alternative conditions to
check, and it is likely that one will be easier to check than another. On the
other hand it can also be useful when answering a question about a given
Poisson process as this question may be more closely connected to one defi
nition than another. For example, you might like to consider the difficulties
in approaching the next result using the jump chain/holding time definition.
Theorem 2.4.4. If and are independent Poisson processes
of rates A and Jl, respectively, then (X
t
+ is a Poisson process of rate
A+ Jl.
Proof. We shall use the infinitesimal definition, according to which
and have independent increments and, as h ! 0, uniformly in t,
P(X
t
+
h
 X
t
= 0) == 1  Ah + o(h), P(X
t
+
h
 X
t
== 1) == Ah + o(h),
P(¥t+h  ¥t = 0) = 1  Jlh + o(h), P(¥t+h  yt = 1) = Jlh + o(h).
Set Zt = X
t
+yt. Then, since and are independent,
has independent increments and, as h ! 0, uniformly in t,
P(Zt+h  Zt = 0) = P(X
t
+
h
 X
t
= O)P(yt+h  yt = 0)
= (1  Ah + o(h))(l  Jlh + o(h)) = 1  (A + Jl)h +o(h),
lP(Zt+h  Zt = 1) = P(X
t
+
h
 X
t
= l)lP(yt+h  yt == 0)
+lP(X
t
+
h
 X
t
== O)lP(yt+h  yt == 1)
== (Ah +o(h) )(1  Jlh +o(h)) + (1  Ah +o(h) )(Jlh +o(h))
= (A + Jl)h + o(h).
Hence is a Poisson process of rate A+ Jl. D
2.4 Poisson processes 79
Next we establish some relations between Poisson processes and the uni
form distribution. Notice that the conclusions are independent of the rate
of the process considered. The results say in effect that the jumps of a
Poisson process are as randomly distributed as possible.
Theorem 2.4.5. Let be a Poisson process. Then, conditional on
having exactly one jump in the interval [s, s +t], the time at which
that jump occurs is uniformly distributed on [s, s +t].
Proof. We shall use the finitedimensional distribution definition. By sta
tionarity of increments, it suffices to consider the case s = O. Then, for
o u t,
IF(J1 u I X
t
= 1) = IF(J
1
u and X
t
= l)/lF(X
t
= 1)
= IF(X
u
= 1 and X
t
 Xu = O)/lF(X
t
= 1)
= Aue'xue,X(tu) /(Ate'xt) = u/t. D
Theorem 2.4.6. Let be a Poisson process. Then, conditional on
the event {X
t
= n}, the jump times J
1
, ... ,I
n
have joint density function
f(t1, . .. ,tn) = n! ...
Thus, conditional on {X
t
= n}, the jump times J
1
, ... ,I
n
have the same
distribution as an ordered sample of size n from the uniform distribution
on [O,t].
Proof. The holding times S1, . .. ,Sn+1 have joint density function
An+1 e'x(Sl +...+Sn+l) 1
{Sl ,... ,Sn+l
so the jump times J1, ... ,I
n
+
1
have joint density function
So for A jRn we have
IF((J
1
, ... ,I
n
) E A and X
t
= n) = IF((J
1
, ... ,I
n
) E A and I
n
t < I
n
+1)
= e,Xt An 1 ... ... dtn
(tl ,... ,tn)EA
and since IF(X
t
= n) = e,Xt AnIn! we obtain
as required. 0
80 2. Continuoustime Markov chains I
We finish with a simple example typical of many problems making use
of a range of properties of the Poisson process.
Example 2.4.7
Robins and blackbirds make brief visits to my birdtable. The probability
that in any small interval of duration h a robin will arrive is found to be
ph+o(h), whereas the corresponding probability for blackbirds is (3h+o(h).
What is the probability that the first two birds I see are both robins? What
is the distribution of the total number of birds seen in time t? Given that
this number is n, what is the distribution of the number of blackbirds seen
in time t?
By the infinitesimal characterization, the number of robins seen by time
t is a Poisson process of rate p, and the number of blackbirds is
a Poisson process of rate (3. The times spent waiting for the first
robin or blackbird are independent exponential random variables 8
1
and T
1
of parameters p and (3 respectively. So a robin arrives first with probability
p/(p + (3) and, by the memoryless property of T
1
, the probability that
the first two birds are robins is p2 / (p +(3)2. By Theorem 2.4.4 the total
number of birds seen in an interval of duration t has Poisson distribution
of parameter (p +(3)t. Finally
JP>(B
t
= k I R
t
+B
t
= n) = JP>(B
t
= k and R
t
= n  k)/JP>(R
t
+B
t
= n)
= (e
f3
(3k) (e
ppn

k
)/ (e
CP
+
f3
)(p+(3)n)
k! (n  k)! n!
=
so if n birds are seen in time t, then the distribution of the number of
blackbirds is binomial of parameters nand {3/ (p +(3).
Exercises
2.4.1 State the transition probability definition of a Poisson process. Show
directly from this definition that the first jump time J1 of a Poisson process
of rate A is exponential of parameter A.
Show also (from the same definition and without assuming the strong
Markov property) that
and hence that J
2
 J1 is also exponential of parameter A and independent
of J1.
2.5 Birth processes 81
2.4.2 Show directly from the infinitesimal definition that the first jump time
J
1
of a Poisson process of rate A has exponential distribution of parameter
A.
2.4.3 Arrivals of the Number 1 bus form a Poisson process of rate one bus
per hour, and arrivals of the Number 7 bus form an independent Poisson
process of rate seven buses per hour.
(a) What is the probability that exactly three buses pass by in one hour?
(b) What is the probability that exactly three Number 7 buses pass by
while I am waiting for a Number I?
(c) When the maintenance depot goes on strike half the buses break down
before they reach my stop. What, then, is the probability that I wait
for 30 minutes without seeing a single bus?
2.4.4 A radioactive source emits particles in a Poisson process of rate A.
The particles are each emitted in an independent random direction. A
Geiger counter placed near the source records a fraction p of the particles
emitted. What is the distribution of the number of particles recorded in
time t?
2.4.5 A pedestrian wishes to cross a single lane of fastmoving traffic. Sup
pose the number of vehicles that have passed by time t is a Poisson process
of rate A, and suppose it takes time a to walk across the lane. Assuming
that the pedestrian can foresee correctly the times at which vehicles will
pass by, how long on average does it take to cross over safely? [Consider
the time at which the first car passes.]
How long on average does it take to cross two similar lanes (a) when one
must walk straight across (assuming that the pedestrian will not cross if,
at any time whilst crossing, a car would pass in either direction), (b) when
an island in the middle of the road makes it safe to stop halfway?
2.5 Birth processes
A birth process is a generalization of a Poisson process in which the param
eter A is allowed to depend on the current state of the process. The data
for a birth process consist of birth rates 0 ~ qj < 00, where j == 0,1,2, ....
We begin with a definition in terms of jump chain and holding times. A
minimal rightcontinuous process ( X t ) ( ~ O with values in {O, 1,2, ... } U{oo}
is a birth process of rates (qj : j ~ 0) if, conditional on X
o
== i, its holding
times 8
1
,8
2
, are independent exponential random variables of param
eters qi, qi+1, , respectively, and its jump chain is given by Y
n
== i +n.
82 2. Continuoustime Markov chains I
ql q2 q3
~ . ~ . .. .
o 1 2 3 4
The flow diagram is shown above and the Qmatrix is given by:
Q=
Example 2.5.1 (Simple birth process)
Consider a population in which each individual gives birth after an expo
nential time of parameter X, all independently. If i individuals are present
then the first birth will occur after an exponential time of parameter iA.
Then we have i + 1 individuals and, by the memoryless property, the pro
cess begins afresh. Thus the size of the population performs a birth process
with rates qi = iX. Let X
t
denote the number of individuals at time t and
suppose X
o
= 1. Write T for the time of the first birth. Then
E(X
t
) = E ( X t l T ~ t ) +E(X
t
lT>t)
= it ,xeASlE(X
t
I T = s )ds + e
At
•
Put Jl(t) = E(X
t
), then E(X
t
IT = s) = 2Jl(t  s), so
p,(t) = it 2,xe
AS
p,(t  s)ds + e
At
and setting r = t  s
By differentiating we obtain
so the mean population size grows exponentially:
2.5 Birth processes 83
t
Much of the theory associated with the Poisson process goes through
for birth processes with little change, except that some calculations can no
longer be made so explicitly. The most interesting new phenomenon present
in birth processes is the possibility of explosion. For certain choices of birth
rates, a typical path will make infinitely many jumps in a finite time, as
shown in the diagram. The convention of setting the process to equal 00
after explosion is particularly appropriate for birth processes!
· ..
8 : : :.. .:•..................
7 ; : : :.. : .•..................
· .
6 ; : : :.. •...................
· ..
5 : : : :.. .
· "
4 : : .
· ..
3 : : ......0. : .:.: .
2 ..............• .: .. : .:.:: .
1 ....................•• : :.. :.:: .
J
o
== 0
In fact, Theorem 2.3.2 tells us exactly when explosion will occur.
Theorem 2.5.2. Let be a birth process of rates (qj : j > 0),
starting from o.
00 1
(i) IfI:  < 00, then P(( < 00) = 1.
j==O qj
00 1
(ii) If:E  = 00, then P(( = 00) = 1.
j==O qj
Proof. Apply Theorem 2.3.2 to the sequence of holding times 8
1
,8
2
,. . .. D
The proof of the Markov property for the Poisson process is easily
adapted to give the following generalization.
84 2. Continuoustime Markov chains I
Theorem 2.5.3 (Markov property). Let be a birth process of
rates (qj : j 2:: 0). Then, conditional on X
s
= i, is a birth process
of rates (qj : j 0) starting from i and independent of (X
r
: r s).
We shall shortly prove a theorem on birth processes which generalizes the
key theorem on Poisson processes. First we must see what will replace the
Poisson probabilities. In Theorem 2.4.3 these arose as the unique solution
of a system of differential equations, which we showed were essentially the
forward equations. Now we can still write down the forward equation
P'(t) = P(t)Q, P(O) = I.
or, in components
and, for j = 1, 2, ...
Moreover, these equations still have a unique solution; it is just not as
explicit as before. For we must have
which can be substituted in the equation
1( t) = PiO(t )qo  Pi 1( t )q1 , Pi 1( 0) = 8
i
1
and this equation solved to give
Now we can substitute for Pi1(t) in the next equation up the hierarchy and
find an explicit expression for Pi2(t), and so on.
Theorem 2.5.4. Let be an increasing, rightcontinuous process
with values in {O, 1,2, ... } U {oo}. Let 0 qj < 00 for all j o. Then the
following three conditions are equivalent:
(a) (jump chain/holding time definition) conditional on X
o
= i, the hold
ing times 8
1
,8
2
, ... are independent exponential random variables of
parameters qi, qi+1, . .. respectively and the jump chain is given by
Y
n
= i +n for all n;
2.5 Birth processes 85
(b) (infinitesimal definition) for all t, h ~ 0, conditional on X
t
= i, X
t
+
h
is independent of (X
s
: s ~ t) and, as h ! 0, uniformly in t,
IF(X
t
+
h
= i I X
t
= i) = 1  qih +o(h),
lP(X
t
+
h
= i +1 I X
t
= i) = qih +o(h);
(c) (transition probability definition) for all n = 0,1,2, ... , all times °~
to ~ ... ~ tn+l and all states io, ... ,in+l
where (pij(t) : i,j = 0,1,2, ... ) is the unique solution of the forward
equations.
If ( X t ) t ~ O satisfies any of these conditions then it is called a birth process
of rates (qj : j ~ 0).
Proof. (a) => (b) If (a) holds, then, by the Markov property for any t, h ~ 0,
conditional on X
t
= i, X
t
+
h
is independent of (X
s
: s ~ t) and, as h ! 0,
uniformly in t,
lP(X
t
+
h
~ i +1 I X
t
= i) = lP(X
h
~ i +1 I X
o
= i)
= IF(Jl ~ h I X
o
= i) = 1  e
qih
= qih +o(h),
and
IF(X
t
+
h
~ i +2 I Xt = i) = IF(X
h
~ i +2 I X
o
= i)
= IF( J
2
~ h I X
o
= i) ~ IF(Sl ~ hand S2 ~ h I X
o
= i)
= (1  e
qih
)(l  e
qi
+
1h
) = o(h),
which implies (b).
(b) => (c) If (b) holds, then certainly for k = i +2, i +3, ...
IF(X
t
+
h
= k I X
t
= i) = o(h) as h ! 0, uniformly in t.
Set Pij(t) = IF(X
t
= j I X
o
= i). Then, for j = 1,2, ...
Pij(t +h) = IF(X
t
+
h
= j I X o = i)
j
= LJP>(X
t
= k I X
o
= i)JP>(X
Hh
= j I X
t
= k)
k==i
=Pij(t)(l  qjh +o(h)) +Pi,jl(t)(qjlh +o(h)) +o(h)
86
so
2. Continuoustime Markov chains I
As in the proof of Theorem 2.4.3, we can deduce that Pij(t) is differentiable
and satisfies the differential equation
By a simpler argument we also find
Thus (Pij(t) : i,j = 0,1,2, ... ) must be the unique solution to the forward
equations. If satisfies (b), then certainly
but also satisfies (b), so
by uniqueness for the forward equations. Hence satisfies (c).
(c) =* (a) See the proof of Theorem 2.4.3. D
Exercise
2.5.1 Each bacterium in a colony splits into two identical bacteria after
an exponential time of parameter X, which then split in the same way but
independently. Let X
t
denote the size of the colony at time t, and suppose
X
o
= 1. Show that the probability generating function ¢(t) = E(zX
t
)
satisfies
Make a change of variables u = t  s in the integral and deduce that
d¢/dt = X¢(¢  1). Hence deduce that, for q = 1  e
At
and n = 1,2, ...
2.6 Jump chain and holding times
2.6 Jump chain and holding times
87
This section begins the theory of continuoustime Markov chains proper,
which will occupy the remainder of this chapter and the whole of the next.
The approach we have chosen is to introduce continuoustime chains in
terms of the joint distribution of their jump chain and holding times. This
provides the most direct mathematical description. It also makes possible
a number of constructive realizations of a given Markov chain, which we
shall describe, and which underlie many applications.
Let I be a countable set. The basic data for a continuoustime Markov
chain on I are given in the form of a Qmatrix. Recall that a Qmatrix on
I is any matrix Q = (qij : i, j E I) which satisfies the following conditions:
(i) 0 ~ qii < 00 for all i;
(ii) qij ~ 0 for all i =I j;
(iii) L qij = 0 for all i.
JEI
We will sometimes find it convenient to write qi or q(i) as an alternative
notation for qii.
We are going to describe a simple procedure for obtaining from a Q
matrix Q a stochastic matrix IT. The jump matrix IT = (7rij : i, j E I) of Q
is defined by
{
qij / qi if j =I i and qi =I 0
7rij = 0 if j =I i and qi = 0,
7rii = {O ~ f qi 1= 0
1 If qi = O.
This procedure is best thought of row by row. For each i E I we take,
where possible, the offdiagonal entries in the ith rovJ of Q and scale them
so they add up to 1, putting a 0 on the diagonal. This is only impossible
when the offdiagonal entries are all 0, then we leave them alone and put a
1 on the diagonal. As you will see in the following example, the associated
diagram transforms into a discretetime Markov chain diagram simply by
rescaling all the numbers on any arrows leaving a state so they add up to
1.
Example 2.6.1
The Qmatrix
Q = ( ~ 2 ~ 1 ~ )
2 1 3
88
has diagram:
2. Continuoustime Markov chains I
1
3
1
2
The jump matrix IT of Q is given by
II = ( ~
2/3
and has diagram:
1/2
o
1/3
1
3
1
3
2
Here is the definition of a continuoustime Markov chain in terms of its
jump chain and holding times. Recall that a minimal process is one which
is set equal to 00 after any explosion  see Section 2.2. A minimal right
continuous process ( X t ) t ~ O on I is a Markov chain with initial distribution
.x and generator matrix Q if its jump chain ( Y n ) n ~ O is discretetime Mar
kov(.x, IT) and if for each n ~ 1, conditional on yo, ... ,Y
n

1
, its holding
times Sl, ,Sn are independent exponential random variables of param
eters q(Y
o
), ,q(Yn1) respectively. We say ( X t ) t ~ O is Markov(.x, Q) for
short. We can construct such a process as follows: let ( Y n ) n ~ O be discrete
time Markov(.x, IT) and let T
1
, T
2
, ••• be independent exponential random
2.6 Jump chain and holding times 89
variables of parameter 1, independent of Set 8
n
= T
n
/q(Y
n

1
),
I
n
= 8
1
+... +8
n
and
{
Yn if I
n
t < I
n
+
1
for some n
X
t
= 00
otherwise.
Then has the required properties.
We shall now describe two further constructions. You will need to un
derstand these constructions in order to identify processes in applications
which can be modelled as Markov chains. Both constructions make direct
use of the entries in the Qmatrix, rather than proceeding first via the jump
matrix. Here is the second construction.
We begin with an initial state X
o
= Yo with distribution .x, and with an
array : n 1, j E I) of independent exponential random variables of
parameter 1. Then, inductively for n = 0,1,2, ... , if Y
n
= i we set
= for j =I i,
Sn+l =
Jr'I,
y: _ {j if = 8n +1 < 00
n+1  . ·f S
'l 1 n+1 = 00.
Then, conditional on Y
n
= i, the random variables are independent
exponentials of parameter qij for all j =I i. So, conditional on Y
n
= i,
by Theorem 2.3.3, 8
n
+
1
is exponential of parameter qi = Ej:j=i qij, Y
n
+
1
has distribution (7rij : j E I), and 8
n
+
1
and Y
n
+
1
are independent, and
independent of yo, . .. ,Y
n
and 8
1
, ... ,8
n
, as required. This construction
shows why we call qi the rate of leaving i and qij the rate of going from i
to j.
Our third and final construction of a Markov chain with generator matrix
Q and initial distribution .x is based on the Poisson process. Imagine the
statespace I as a labyrinth of chambers and passages, each passage shut
off by a single door which opens briefly from time to time to allow you
through in one direction only. Suppose the door giving access to chamber
j from chamber i opens at the jump times of a Poisson process of rate qij
and you take every chance to move that you can, then you will perform
a Markov chain with Qmatrix Q. In more mathematical terms, we begin
with an initial state X
o
= Yo with distribution .x, and with a family of
independent Poisson processes : i,j E I,i =I j}, having
rate qij. Then set J
o
= °and define inductively for n = 0,1,2, ...
I
n
+
1
= inf{t > I
n
: Nr
nj
=I Nj:
j
for some j #ly
n
}
{
j if I
n
+
1
< 00 and Njnj =I Njnj
Y
n
+
1
= . . n+l n
'l If I
n
+
1
= 00.
90 2. Continuoustime Markov chains I
The first jump time of is exponential of parameter qij. So, by
Theorem 2.3.3, conditional on Yo = i, J
1
is exponential of parameter qi =
Lj#i qij, Y
1
has distribution (7rij : j E I), and J
1
and Y
1
are independent.
Now suppose T is a stopping time of If we condition on X
o
and
on the processes for (k,l) =1= (i,j), which are independent of N;j,
then {T t} depends only on (N;j : s t). So, by the strong Markov
property of the Poisson process N;j := N!}+t  N!} is a Poisson process of
rate qij independent of (N;j : s T), and independent of X
o
and
for (k,l) =1= (i,j). Hence, conditional on T < 00 and XT = i,
has the same distribution as and is independent of (X
s
: s T).
In particular, we can take T = I
n
to see that, conditional on I
n
< 00
and Y
n
= i, 8
n
+
1
is exponential of parameter qi, Y
n
+
1
has distribution
(7rij : j E I), and 8
n
+
1
and Y
n
+
1
are independent, and independent of
yo, ... ,Y
n
and 8
1
, ... ,8
n
· Hence is Markov(X, Q) and, more
over, has the strong Markov property. The conditioning on which
this argument relies requires some further justification, especially when the
statespace is infinite, so we shall not rely on this third construction in the
development of the theory.
2.7 Explosion
We saw in the special case of birth processes that, although each holding
time is strictly positive, one can run through a sequence of states with
shorter and shorter holding times and end up taking infinitely many jumps
in a finite time. This phenomenon is called explosion. Recall the notation
of Section 2.2: for a process with jump times Jo, J
1
, J2, ... and holding
times 8
1
, 8
2
, ... , the explosion time ( is given by
00
(= supJ
n
= 2: Sn o
n n==l
Theorem 2.7.1. Let be Markov(X, Q). Then does not
explode if anyone of the following conditions holds:
(i) I is finite;
(ii) sup qi < 00;
iEI
(iii) X
o
= i, and i is recurrent for the jump chain.
Proof. Set Tn = q(Y
n

1
)Sn, then T
1
, T
2
, . .. are independent E(l) and in
dependent of In cases (i) and (ii), q = SUPi qi < 00 and
00
2. 7 Explosion 91
with probability 1. In case (iii), we know that visits i infinitely
often, at times N
1
, N
2
, • •• , say. Then
00
qi( L TN",+! = 00
m==l
with probability 1. D
We say that a Qmatrix Q is explosive if, for the associated Markov chain
Pi (( < 00) > 0 for some i E I.
Otherwise Q is nonexplosive. Here as in Chapter 1 we denote by Pi the
conditional probability Pi(A) = P(AIX
o
= i). It is a simple consequence
of the Markov property for (Yn)n>O that under Pi the process (Xt)t>o is
 
Markov(8
i
, Q). The result just proved gives simple conditions for non
explosion and covers many cases of interest. As a corollary to the next
result we shall obtain necessary and sufficient conditions for Q to be explo
sive, but these are not as easy to apply as Theorem 2.7.1.
Theorem 2.7.2. Let be a continuoustime Markov chain with
generator matrix Q and write ( for the explosion time of (Xt)t>o. Fix
() > 0 and set Zi = Ei(e(}(). Then Z = (Zi : i E I) satisfies: 
(i) IZil 1 for all i;
(ii) Qz = Oz.
Moreover, ifz also satisfies (i) and (ii), then Zi Zi for all i.
Proof. Condition on X
o
= i. The time and place of the first jump are
independent, J
1
is E(qi) and
Moreover, by the Markov property of the jump chain at time n = 1, con
ditional on XJ
1
= k, is Markov(8k,Q) and independent of J
1
•
So
and
92 2. Continuoustime Markov chains I
Recall that qi = qii and q(Jrik = qik. Then
(0  qii)Zi = L qikZk
k#i
so
OZi = Lqikzk
kEI
and so z satisfies (i) and (ii). Note that the same argument also shows that
Suppose that zalso satisfies (i) and (ii), then, in particular
for all i. Suppose inductively that
z. < E·(e
9Jn
)
~  ~
then, since zsatisfies (ii)
Hence Zi ~ E
i
(e
9Jn
) for all n. By monotone convergence
as n ~ 00, so Zi ~ Zi for all i. D
Corollary 2.7.3. For each (J > 0 the following are equivalent:
(a) Q is nonexplosive;
(b) Qz = (Jz and IZil ~ 1 for all i imply z = o.
Proof. If (a) holds then JP>i(( = 00) = 1 so E
i
(e
9
() = o. By the theorem,
Qz = (Jz and Izl ~ 1 imply Zi ~ E
i
(e
9
(), hence z ~ 0, by symmetry z ~ 0,
and hence (b) holds. On the other hand, if (b) holds, then by the theorem
E
i
(e
9
() = 0 for all i, so JP>i(( = 00) = 1 and (a) holds. D
2.8 Forward and backward equations 93
Exercise
2.7.1 Let (X
t
be a Markov chain on the integers with transition rates
qi,i+l = Aqi, qi,il = Jlqi
and qij = °if Ij  il 2, where A + Jl = 1 and qi > °for all i. Find for all
integers i:
(a) the probability, starting from 0, that X
t
hits i;
(b) the expected total time spent in state i, starting from 0.
In the case where Jl = 0, write down a necessary and sufficient condition
for to be explosive. Why is this condition necessary for to
be explosive for all Jl E [0,1/2)?
Show that, in general, is nonexplosive if and only if one of the
following conditions holds:
(i) A = Jl;
(ii) A > Jl and 1/qi = 00;
(iii) A < Jl and = 00.
2.8 Forward and backward equations
Although the definition of a continuoustime Markov chain in terms of its
jump chain and holding times provides a clear picture of the process, it does
not answer some basic questions. For example, we might wish to calculate
IPi(X
t
= j). In this section we shall obtain two more ways of characterizing
a continuoustime Markov chain, which will in particular give us a means
to find IPi(X
t
= j). As for Poisson processes and birth processes, the
first step is to deduce the Markov property from the jump chain/holding
time definition. In fact, we shall give the strong Markov property as this
is a fundamental result and the proof is not much harder. However, the
proof of both results really requires the precision of measure theory, so we
have deferred it to Section 6.5. If you want to understand what happens,
Theorem 2.4.1 on the Poisson process gives the main idea in a simpler
context.
Recall that a random variable T with values in [0,00] is a stopping time of
if for each t E [0,00) the event {T t} depends only on (X
s
: s t).
Theorem 2.8.1 (Strong Markov property). Let be
Markov(A, Q) and let T be a stopping time of Then, conditional
on T < 00 and XT = i, is Markov(8
i
, Q) and independent of
(X
s
: s T).
We come to the key result for continuoustime Markov chains. We shall
present first a version for the case of finite statespace, where there is a
94 2. Continuoustime Markov chains I
simpler proof. In this case there are three alternative definitions, just as for
the Poisson process.
Theorem 2.8.2. Let be a rightcontinuous process with values in
a finite set I. Let Q be a Qmatrix on I with jump matrix II. Then the
following three conditions are equivalent:
(a) (jump chain/holding time definition) conditional on X
o
= i, the
jump chain of is discretetime Markov(8
i
, IT) and for
each n 1, conditional on yo, . .. ,Y
n

1
, the holding times 8
1
, ... ,8
n
are independent exponential random variables of parameters
q(Y
o
), ... ,q(Y
n

1
) respectively;
(b) (infinitesimal definition) for all t, h 0, conditional on X
t
= i, Xt+h
is independent of (X
s
: s t) and, as h ! 0, uniformly in t, for all j
(c) (transition probability definition) for all n = 0, 1, 2, ... , all times °::;
to t
1
... t
n
+
1
and all states io, ... ,i
n
+
1
where (Pij (t) : i, j E I, t 0) is the solution of the forward equation
P'(t) = P(t)Q, P(o) = I.
If satisfies any of these conditions then it is called a Markov chain
with generator matrix Q. We say that is Markov(>.., Q) for short,
where>.. is the distribution of X
o
.
Proof. (a) =* (b) Suppose (a) holds, then, as h ! 0,
and for j =I i we have
JP>i(Xh = j) JP>(J1 h, Y1 = j, 8
2
> h)
= (1  eqih)7rijeqjh = qijh + o(h).
Thus for every state j there is an inequality
2.8 Forward and backward equations 95
and by taking the finite sum over j we see that these must in fact be
equalities. Then by the Markov property, for any t, h 0, conditional on
X
t
= i, X
t
+
h
is independent of (X
s
: s t) and, as h ! 0, uniformly in t
(b) :::} (c) Set Pij(t) = lFi(X
t
= j) = IF(X
t
= j I X
o
= i). If (b) holds, then
for all t, h 0, as h ! 0, uniformly in t
Pij(t + h) = LlPi(X
t
= k)IP(XHh = j I X
t
= k)
kEI
= LPik(t)(8kj + qkjh + o(h)).
kEI
Since I is finite we have
Pij(t +  Pij(t) = LPik(t)qkj + O(h)
kEI
so, letting h ! 0, we see that Pij(t) is differentiable on the right. Then by
uniformity we can replace t by t  h in the above and let h ! °to see first
that Pij(t) is continuous on the left, then differentiable on the left, hence
differentiable, and satisfies the forward equations
= LPik(t)qkj, Pij(O) = 8ij .
kEI
Since I is finite, Pij(t) is then the unique solution by Theorem 2.1.1. Also,
if (b) holds, then
and, moreover, (b) holds for so, by the above argument,
proving (c).
(c) :::} (a) See the proof of Theorem 2.4.3. D
We know from Theorem 2.1.1 that for I finite the forward and backward
equations have the same solution. So in condition (c) of the result just
proved we could replace the forward equation with the backward equation.
Indeed, there is a slight variation of the argument from (b) to (c) which
leads directly to the backward equation.
96 2. Continuoustime Markov chains I
The deduction of (c) from (b) above can be seen as the matrix version
of the following result: for q E lR we have
(1 + + o( )n 4 e
q
as n 4 00.
Suppose (b) holds and set
Pij(t, t +h) = P(X
t
+
h
= j I X
t
= i);
then P(t, t +h) = (Pij(t, t +h) : i,j E I) satisfies
P(t, t +h) = I +Qh +o(h)
and
... =
Some care is needed in making this precise, since the o(h) terms, though
uniform in t, are not a priori identical. On the other hand, in (c) we see
that
P(O, t) = e
tQ
.
We turn now to the case of infinite statespace. The backward equation
may still be written in the form
P'(t) = QP(t), P(O) = I
only now we have an infinite system of differential equations
= L qikPkj(t), Pij(O) = 15ij
kEI
and the results on matrix exponentials given in Section 2.1 no longer apply.
A solution to the backward equation is any matrix (pij(t) : i,j E I) of
differentiable functions satisfying this system of differential equations.
Theorem 2.8.3. Let Q be a Qmatrix. Then the backward equation
P'(t) = QP(t), P(O) = I
has a minimal nonnegative solution (P(t) : t 2: 0). This solution forms a
matrix semigroup
P(s)P(t) = P(s +t) for all s, t O.
We shall prove this result by a probabilistic method in combination with
Theorem 2.8.4. Note that if I is finite we must have P(t) = e
tQ
by Theorem
2.1.1. We call (P(t) : t 0) the minimal nonnegative semigroup associated
to Q, or simply the semigroup of Q, the qualifications minimal and non
negative being understood.
Here is the key result for Markov chains with infinite statespace. There
are just two alternative definitions now as the infinitesimal characterization
become problematic for infinite statespace.
2.8 Forward and backward equations 97
(2.2)
Theorem 2.8.4. Let be a minimal rightcontinuous process with
values in I. Let Q be a Qmatrix on I with jump matrix IT and semigroup
(P(t) : t 0). Then the following conditions are equivalent:
(a) (jump chain/holding time definition) conditional on X
o
= i, the
jump chain of is discretetime Markov(8i, II) and for
each n 1, conditional on yo, . .. ,Y
n

1
, the holding times 8
1
, ... ,8
n
are independent exponential random variables of parameters
q(Y
o
), ... ,q(Y
n

1
) respectively;
(b) (transition probability definition) for all n = 0,1,2, ... , all times 0
to tl ... t
n
+l and all states io, il, ... ,i
n
+l
If satisfies any of these conditions then it is called a Markov chain
with generator matrix Q. We say that is Markov (A, Q) for short,
where A is the distribution of X
o
.
Proof of Theorems 2.8.3 and 2.8.4. We know that there exists a process
satisfying (a). So let us define P(t) by
Pij(t) =JP>i(X
t
=j).
Step 1. We show that P(t) satisfies the backward equation.
Conditional on X
o
= i we have J
1
rv E(ql) and X
J1
rv (7rik : k E I).
Then conditional on J
1
= sand X
J1
= k we have rv
Markov(8k, Q). So
and
JP>i(J1 ::; t, XJI = k, X
t
= j) = I
t
qieQiS7rikPkj(t  s)ds.
Therefore
Pij(t) = JP>i(X
t
= j, t < J1) + LJP>i(J1 ::; t, XJl = k, X
t
= j)
k#i
= eQit8ij + 2: t qieQiS7rikPkj(t  s)ds. (2.1)
k#i Jo
Make a change of variable u = t  s in each of the integrals, interchange
sum and integral by monotone convergence and multiply by e
qit
to obtain
eQitpij(t) = 8
ij
+ t L qieQiU7rikPkj(u)du.
Jo k#i
98 2. Continuoustime Markov chains I
This equation shows, firstly, that Pij (t) is continuous in t for all i, j.
Secondly, the integrand is then a uniformly converging sum of continuous
functions, hence continuous, and hence Pij (t) is differentiable in t and sat
isfies
eqit(qiPij(t) + P ~ j ( t ) ) = L qieqit7rikPkj(t).
k#i
Recall that qi = qii and qik = qi'lrik for k =I i. Then, on rearranging, we
obtain
P ~ j ( t ) = L qikPkj(t)
kEI
(2.3)
so P(t) satisfies the backward equation.
The integral equation (2.1) is called the integral form of the backward
equation.
Step 2. We show that if P(t) is another nonnegative solution of the back
ward equation, then P(t) ~ P(t), hence P(t) is the minimal nonnegative
solution.
The argument used to prove (2.1) also shows that
JP>i(X
t
=j,t < I
n
+l)
= eqit15ij + L rt qieQiS7rikJP>k(Xts = j, t  s < In)ds.
k=li Jo (2.4)
On the other hand, if P(t) satisfies the backward equation, then, by revers
ing the steps from (2.1) to (2.3), it also satisfies the integral form:
If P(t) 2 0, then
for all i, j and t.
Let us suppose inductively that
for all i, j and t,
then by comparing (2.4) and (2.5) we have
for all i, j and t,
2.8 Forward and backward equations
and the induction proceeds. Hence
99
for all i, j and t.
Step 3. Since does not return from 00 we have
Pij(S + t) = lPi(Xs+t = j) = LlPi(XS+
t
= j I X s = k)lPi(Xs = k)
kEI
= :ElPi(X
s
= k)lPk(X
t
= j) = :EPik(S)Pkj(t)
kEI kEI
by the Markov property. Hence (P(t) : t 0) is a matrix semigroup. This
completes the proof of Theorem 2.8.3.
Step 4. Suppose, as we have throughout, that satisfies (a). Then,
by the Markov property
P(X
tn
+
1
= i
n
+
1
IX
to
= io, ... ,X
tn
= in)
=Pin (Xtn+ltn = i n+1) = Pi
n
i
n
+l (t n+l  tn)
so satisfies (b). We complete the proof of Theorem 2.8.4 by the
usual argument that (b) must now imply (a) (see the proof of Theorem
2.4.3, (c) :::} (a)). D
So far we have said nothing about the forward equation in the case of
infinite statespace. Remember that the finite statespace results of Section
2.1 are no longer valid. The forward equation may still be written
P'(t) = P(t)Q, P(O) = I,
now understood as an infinite system of differential equations
= LPik(t)qkj, Pij(O) = 8ij .
kEI
A solution is then any matrix (pij(t) : i,j E I) of differentiable functions
satisfying this system of equations. We shall show that the
(P(t) : t 0) of Q does satisfy the forward equations, by a probabilistic
argument resembling Step 1 of the proof of Theorems 2.8.3 and 2.8.4. This
time, instead of conditioning on the first event, we condition on the last
event before time t. The argument is a little longer because there is no
reversetime Markov property to give the conditional distribution. We need
the following timereversal identity, a simple version of which was given in
Theorem 2.3.4.
100 2. Continuoustime Markov chains I
Lemma 2.8.5. We have
qinJP(J
n
~ t < I
n
+
1
I Yo = io, Y
1
= i
1
, ... ,Y
n
= in)
= qioJP(Jn ~ t < I n+
1
I Yo = in,· .. , Y
n

1
= iI, Y
n
= io).
Proof· Conditional on Yo = i
o
, ... ,Y
n
= in, the holding times 8
1
, ... ,8
n
+
1
are independent with 8k rv E(qik_l). So the lefthand side is given by
{ qi
n
exp{qi
n
(t  Sl  · · ·  sn)} IT qikl exp{qikl Sk}dsk
Jt:!..(t) k=l
where Ll(t) = {(SI, ... ,sn) : SI + ... + Sn ~ t and SI, ... ,8
n
~ O}. On
making the substitutions Ul = t  81  ...  Sn and Uk = 8
n
k+2, for
k = 2, ... ,n, we obtain
qinJP(Jn ~ t < I n+
1
I Yo = i
o
,· .. ,Y
n
= in)
= ( qio exp{qio(t  U1  · ..  Un)} IT qink+l exp{qink+l Uk}duk
Jt:!..(t) k=l
=qioJP(J
n
~ t < I
n
+
1
I Yo = in, . .. ,Y
n

1
= iI, Y
n
= io). D
Theorem 2.8.6. The minimal nonnegative solution (P(t) : t 2: 0) of the
backward equation is also the minimal nonnegative solution of the forward
equation
P'(t) = P(t)Q, P(O) = I.
Proof. Let ( X t ) t ~ O denote the minimal Markov chain with generator matrix
Q. By Theorem 2.8.4
00
= L LJP>i(Jn ::; t < In+i, Y
n

1
= k, Y
n
= j).
n==O k=j:j
Now by Lemma 2.8.5, for n ~ 1, we have
JPi(J
n
~ t < I
n
+
1
I Y
n

1
= k, Y
n
= j)
= (qi/qj)JPj(J
n
~ t < I
n
+
1
I Y
1
= k, Y
n
= i)
= (qi/Qj) it qjeQjBJP>k(Jn_1 ::; t  S < I
n
I Yn1 = i)ds
= qi I
t
e
QjB
(qk/qi)JP>i (In1 ::; t  s < I
n
I Y
n

1
= k)ds
2.8 Forward and backward equations 101
where we have used the Markov property of for the second equality.
Hence
Pij(t) = 8ijeqit + f '2: it JP>i(Jnl ::; t  s < I
n
I Y
n

1
= k)
n=lki=i 0
X IPi(Y
n

1
= k, Y
n
= j)Qkeqj8ds
00 ft
=8ijeqit + L L io JP>i(Jnl ::; t  s< I
n
, Y
n

1
= k)qk'lrkjeqjSds
n=lki=i 0
= 8ijeqit + t '2:Pik(t  s)qkjeqjSds (2.6)
io k#j
where we have used monotone convergence to interchange the sum and
integral at the last step. This is the integral form of the forward equation.
Now make a change of variable u = t  s in the integral and multiply by
e
qjt
to obtain
(2.7)
We know by equation (2.2) that eqitpik(t) is increasing for all i, k. Hence
either
'2:Pik(U)qkj converges uniformly for u E [0, t]
ki=i
or
LPik(U)qkj = 00 for all u t.
ki=j
The latter would contradict (2.7) since the lefthand side is finite for all t,
so it is the former which holds. We know from the backward equation that
Pii (t) is continuous for all i, j; hence by uniform convergence the integrand
in (2.7) is continuous and we may differentiate to obtain
+ Pij(t)qj = '2:Pik(t)qkj.
ki=i
Hence P(t) solves the forward equation.
To establish minimality let us suppose that Pij(t) is another solution of
the forward equation; then we also have
102 2. Continuoustime Markov chains I
A small variation of the argument leading to (2.6) shows that, for n ~ 0
Pi(X
t
= j, t < I
n
+
1
)
= 8ijeqit + L (t IP\(X
t
= j, t < In)qkjeQjBds. (2.8)
k#jJo
If P(t) ~ 0, then
P(X
t
= j, t < J
o
) = 0 ~ Pij (t) for all i, j and t.
Let us suppose inductively that
then by comparing (2.7) and (2.8) we obtain
and the induction proceeds. Hence
Exercises
2.8.1 Two fleas are bound together to take part in a ninelegged race on the
vertices A, B, C of a triangle. Flea 1 hops at random times in the clockwise
direction; each hop takes the pair from one vertex to the next and the times
between successive hops of Flea 1 are independent random variables, each
with with exponential distribution, mean 1/A. Flea 2 behaves similarly,
but hops in the anticlockwise direction, the times between his hops having
mean 1/Il. Show that the probability that they are at A at a given time
t > 0 (starting from A at time t = 0) is
2.8.2 Let ( X t ) t ; ~ o be a birthanddeath process with rates An = nA and
Iln = nil, and assume that X
o
= 1. Show that h(t) = P(X
t
= 0) satisfies
2.9 Nonminimal chains
and deduce that if A =I Jl then
2.9 Nonminimal chains
103
This book concentrates entirely on processes which are rightcontinuous
and minimal. These are the simplest sorts of process and, overwhelmingly,
the ones of greatest practical application. We have seen in this chapter
that we can associate to each distribution A and Qmatrix Q a unique
such process, the Markov chain with initial distribution A and generator
matrix Q. Indeed we have taken the liberty of defining Markov chains to be
those processes which arise in this way. However, these processes do not by
any means exhaust the class of memoryless continuoustime processes with
values in a countable set I. There are many more exotic possibilities, the
general theory of which goes very much deeper than the account given in
this book. It is in the nature of things that these exotic cases have received
the greater attention among mathematicians. Here are some examples to
help you imagine the possibilities.
Example 2.9.1
Consider a birth process ( X t ) t ~ O starting from 0 with rates qi = 2
i
for i ~ o.
We have chosen these rates so that
00 00
Lq:;l = L2i < 00
i=O i=O
which shows that the process explodes (see Theorems 2.3.2 and 2.5.2). We
have until now insisted that X
t
= 00 for all t 2 (, where ( is the explosion
time. But another obvious possibility is to start the process off again from
oat time (, and do the same for all subsequent explosions. An argument
based on the memoryless property of the exponential distribution shows
that for 0 ~ to ~ . . . ~ t
n
+1 this process satisfies
for a semigroup of stochastic matrices (P(t) : t ~ 0) on I. This is the
defining property for a more general class of Markov chains. Note that
the chain is no longer determined by A and Q alone; the rule for bringing
( X t ) t ~ O back into I after explosion also has to be given.
104 2. Continuoustime Markov chains I
Example 2.9.2
We make a variation on the preceding example. Suppose now that the jump
chain of is the Markov chain on Z which moves one step away from
owith probability 2/3 and one step towards 0 with probability 1/3, and
that Yo = O. Let the transition rates for be qi = 2
1il
for i E Z. Then
is again explosive. (A simple way to see this using some results of
Chapter 3 is to check that is transient but has an invariant
distribution  by solution of the detailed balance equations. Then Theorem
3.5.3 makes explosion inevitable.) Now there are two ways in which
can explode, either X
t
00 or X
t
00.
The process may again be restarted at 0 after explosion. Alternatively,
we may choose the restart randomly, and according to the way that explo
sion occurred. For example
if X
t
00 as t i (
if X
t
00 as t i (
where Z takes values ±1 with probability 1/2.
Example 2.9.3
The processes in the preceding two examples, though no longer minimal,
were at least rightcontinuous. Here is an altogether more exotic example,
due to P. Levy, which is not even rightcontinuous. Consider
for n 0
and set I = UnD
n
. With each i E D
n
\D
n

1
we associate an independent
exponential random variable Si of parameter (2
n
)2. There are 2
n

1
states
in (Dn\Dnl) n [0,1), so, for all i E I
and
Now define
if L Sj :::; t < L Sj for some i E I
j<i
otherwise.
2.10 Appendix: matrix exponentials 105
This process runs through all the dyadic rationals i E I in the usual order.
It remains in i E D
n
\D
n

1
for an exponential time of parameter 1. Between
any two distinct states i and j it makes infinitely many visits to 00. The
Lebesgue measure of the set of times t when X
t
= 00 is zero. There is a
semigroup of stochastic matrices (P(t) : t ~ 0) on I such that, for 0 ~ to ~
... ~ tn+l
In particular, P(X
t
= 00) = 0 for all t ~ O. The details may be found in
Markov Chains by D. Freedman (HoldenDay, San Francisco, 1971).
We hope these three examples will serve to suggest some of the possibil
ities for more general continuoustime Markov chains. For further reading,
see Freedman's book, or else Markov Chains with Stationary Transition
Probabilities by K.L. Chung (Springer, Berlin, 2nd edition, 1967), or Dif
fusions, Markov Processes and Martingales, Vol 1: Foundations by L. C. G.
Rogers and D. Williams (Wiley, Chichester, 2nd edition, 1994).
2.10 Appendix: matrix exponentials
Define two norms on the space of realvalued N x N matrices
IQI = sup IQvl/lvl, IIQlloo = sup Iqijl·
v#O i,j
Obviously, IIQlloo is finite for all Q and controls the size of the entries in
Q. We shall show that the two norms are equivalent and that IQI is well
adapted to sums and products of matrices, which IIQlloo is not.
Lemma 2.10.1. We have
(a) IIQlloo ~ IQI ~ NIIQlloo;
(b) IQl +Q21 ~ IQll + IQ21 and IQIQ21 ~ IQIIIQ21·
Proof. (a) For any vector v we have IQvl ~ IQllvl. In particular, for the
vector Cj = (0, ... ,1, ... ,0), with 1 in the jth place, we have IQcjl ~ IQI.
The supremum defining IIQlloo is achieved, at j say, so
I I Q I I ~ ~ L(qij)2 = IQcjl2 ~ IQI
2
.
i
106
On the other hand
2. Continuoustime Markov chains I
IQvI
2
= 2
( IIQllooIVjl) 2
= ( IVjl) 2
and, by the CauchySchwarz inequality
(b) For any vector v we have
I(QI +Q2)vl IQI
V
I+ IQ2
V
l (IQll + IQ21)1vl,
IQIQ2
V
l IQI11Q2
V
l IQIIIQ21Ivl· D
Now for n = 0,1,2, ... , consider the finite sum
n Qk
E(n) = L kf'
k=O
For each i and j, and m n, we have
I(E(n)  E(m))iil IIE(n)  E(m)lloo IE(n)  E(m)1
n Qk
Lkf
k=m+l
n IQl
k
< '" 
 LJ k!·
k=m+l
Since IQI NllQlloo < 00, IQl
k
/k! converges by the ratio test, so
n IQl
k
'" 
LJ k!
k=m+l
as m,n 00.
2.10 Appendix: matrix exponentials 107
Hence each component of E(n) forms a Cauchy sequence, which therefore
converges, proving that
CX) Qk
e
Q
= L kf
k=O
is well defined and, indeed, that the power series
(
t
Q
) .. _ ~ (tQ)fj
e '1,)  L.J k!
k=O
has infinite radius of convergence for all i,j. Finally, for two commuting
matrices Q1 and Q2 we have
3
Continuoustime Markov chains II
This chapter brings together the discretetime and continuoustime theo
ries, allowing us to deduce analogues, for continuoustime chains, of all the
results given in Chapter 1. All the facts from Chapter 2 that are necessary
to understand this synthesis are reviewed in Section 3.1. You will require
a reasonable understanding of Chapter 1 here, but, given such an under
standing, this chapter should look reassuringly familiar. Exercises remain
an important part of the text.
3.1 Basic properties
Let I be a countable set. Recall that a Qmatrix on I is a matrix Q =
(qij : i,j E I) satisfying the following conditions:
(i) 0 ~ qii < 00 for all i;
(ii) qij 2 0 for all i =I j;
(iii) L qij = 0 for all i.
jEI
We set qi = q(i) = qii. Associated to any Qmatrix is a jump matrix
IT = (7rij : i,j E I) given by
1r.. _ { qij / qi if j =I i and qi =I 0
~ J  0 if j =I i and qi = 0,
7r.. _ {o if qi =I 0
~ ~  1 if qi = o.
Note that II is a stochastic matrix.
3.1 Basic properties 109
A substochastic matrix on I is a matrix P = (Pij : i, j E I) with non
negative entries and such that
LPij ::; 1 for all i.
jEI
Associated to any Qmatrix is a semigroup (P(t) : t 2 0) of substochastic
matrices P(t) = (Pij(t) : i,j E I). As the name implies we have
P(s)P(t) = P(s +t) for all s, t 2 o.
You will need to be familiar with the following terms introduced in Sec
tion 2.2: minimal rightcontinuous random process, jump times, holding
times, jump chain and explosion. Briefly, a rightcontinuous process ( X t ) t ~ O
runs through a sequence of states Yo, Y
1
, Y
2
, . .. , being held in these states
for times 8
1
,8
2
,8
3
, ... respectively and jumping to the next state at times
J
1
, J
2
, J
3
,· ... Thus I
n
= 8
1
+... +8
n
. The discretetime process ( Y n ) n ~ O
is the jump chain, (8
n
)n>1 are the holding times and (I
n
)n>1 are the jump
 
times. The explosion time ( is given by
00
( = '""" 8n = lim I n .
L.J n+oo
n=l
For a minimal process we take a new state 00 and insist that X
t
= 00 for
all t 2 (. An important point is that a minimal rightcontinuous process is
determined by its jump chain and holding times.
The data for a continuoustime Markov chain ( X t ) t ~ O are a distribution
X and a Qmatrix Q. The distribution X gives the initial distribution, the
distribution of X
o
. The Qmatrix is known as the generator matrix of
( X t ) t ~ O and determines how the process evolves from its initial state. We
established in Section 2.8 that there are two different, but equivalent, ways
to describe how the process evolves.
The first, in terms of jump chain and holding times, states that
(a) ( Y n ) n ~ O is Markov(X, IT);
(b) conditional on Yo = i
o
, ... ,Y
n

1
= inI, the holding times 8
1
, ... ,Sn
are independent exponential random variables of parameters
qi
o
' . .. ,qi
n
1 .
Put more simply, given that the chain starts at i, it waits there for an
exponential time of parameter qi and then jumps to a new state, choosing
state j with probability 1rij. It then starts afresh, forgetting what has gone
before.
110 3. Continuoustime Markov chains II
The second description, in terms of the semigroup, states that the finite
dimensional distributions of the process are given by
(c) for all n = 0,1,2, ... , all times 0 to tl ... tn+l and all states
i
o
, iI, . .. ,i
n
+1
Again, put more simply, given that the chain starts at i, by time t it is
found in state j with probability Pij (t). It then starts afresh, forgetting
what has gone before. In the case where
Pioo(t) := 1  :EPij(t) > 0
jEI
the chain is found at 00 with probability Pi 00 (t). The semigroup P(t) is re
ferred to as the transition matrix of the chain and its entries Pij (t) are the
transition probabilities. This description implies that for all h > 0 the dis
crete skeleton is Markov(.x, P(h)). Strictly, in the explosive case,
that is, when P(t) is strictly substochastic, we should say P(h)),
where and P(h) are defined on IU{oo}, extending.x and P(h) by = 0
and Pooj(h) = O. But there is no danger of confusion in using the simpler
notation.
The information coming from these two descriptions is sufficient for most
of the analysis of continuoustime chains done in this chapter. Note that
we have not yet said how the semigroup P(t) is associated to the Qmatrix
Q, except via the process! This extra information will be required when
we discuss reversibility in Section 3.7. So we recall from Section 2.8 that
the semigroup is characterized as the minimal nonnegative solution of the
backward equation
P'(t) = QP(t), P(O) = I
which reads in components
= L qikPkj(t), Pij(O) = 8ij .
kEI
The semigroup is also the minimal nonnegative solution of the forward
equation
P'(t) = P(t)Q, P(O) = I.
In the case where I is finite, P(t) is simply the matrix exponential e
tQ
, and
is the unique solution of the backward and forward equations.
3.2 Class structure
3.2 Class structure
111
A first step in the analysis of a continuoustime Markov chain is to
identify its class structure. We emphasise that we deal only with minimal
chains, those that die after explosion. Then the class structure is simply
the discretetime class structure of the jump chain as discussed in
Section 1.2.
We say that i leads to j and write i j if
lPi(X
t
= j for some t 0) > o.
We say i communicates with j and write i j if both i j and j
i. The notions of communicating class, closed class, absorbing state and
irreducibility are inherited from the jump chain.
Theorem 3.2.1. For distinct states i and j the following are equivalent:
(i) i
(ii) i j for the jump chain;
(iii) Q
i
oilqili2 ... Qinli
n
> 0 for some states io,iI, ... ,in with i
o
= i,
in = j;
(iv) Pij(t) > 0 for all t > 0;
(v) Pij(t) > 0 for some t > o.
Proof. Implications (iv) => (v) => (i) => (ii) are clear. If (ii) holds, then
by Theorem 1.2.1, there are states io, i
I
, ... ,in with io = i, in = j and
'1rioi
l
'1rili2 ... '1ri
n
li
n
> 0, which implies (iii). If qij > 0, then
for all t > 0, so if (iii) holds, then
Pij (t) Pioil (tin) ... Pinlin (tin) > 0
for all t > 0, and (iv) holds. D
Condition (iv) of Theorem 3.2.1 shows that the situation is simpler than
in discretetime, where it may be possible to reach a state, but only after a
certain length of time, and then only periodically.
3.3 Hitting times and absorption probabilities
Let (Xt)t>o be a Markov chain with generator matrix Q. The hitting time
of a subset A of I is the random variable D
A
defined by
112 3. Continuoustime Markov chains II
with the usual convention that inf 0 = 00. We emphasise that (Xt)t>o is
minimal. So if H A is the hitting time of A for the jump chain, then 
{H
A
< oo} = {D
A
< oo}
and on this set we have
D
A
= JHA.
The probability, starting from i, that ( X t ) t ~ O ever hits A is then
When A is a closed class, hf is called the absorption probability. Since the
hitting probabilities are those of the jump chain we can calculate them as
in Section 1.3.
Theorem 3.3.1. The vector of hitting probabilities h
A
= (hf : i E I) is
the minimal nonnegative solution to the system of linear equations
Proof. Apply Theorem 1.3.2 to the jump chain and rewrite (1.3) in terms
ofQ. D
The average time taken, starting from i, for ( X t ) t ~ O to reach A is given
by
In calculating kf we have to take account of the holding times so the rela
tionship to the discretetime case is not quite as simple.
1
1 2
2
1
3
3
3
2
4
3.3 Hitting times and absorption probabilities 113
(3.1)
Example 3.3.2
Consider the Markov chain with the diagram given on the preceding
page. How long on average does it take to get from 1 to 4?
Set k
i
= E
i
(time to get to 4). On starting in 1 we spend an average time
ql
1
= 1/2 in 1, then jump with equal probability to 2 or 3. Thus
k 1 = + k2 + k3
and similarly
k 2 = + k 1 + k 3 , k 3 = + k 1 + k 2 .
On solving these linear equations we find k
1
= 17/12.
Here is the general result. The proof follows the same lines as Theorem
1.3.5.
Theorem 3.3.3. Assume that qi > 0 for all i fj. A. The vector of expected
hitting times k
A
= (kt : i E I) is the minimal nonnegative solution to the
system of linear equations
{
kt = 0 for i E A
 LjEI qijkf = 1 for i fj. A.
Proof. First we show that k
A
satisfies (3.1). If X
o
= i E A, then D
A
= 0,
so kt = o. If X
o
= i fj. A, then D
A
J
1
, so by the Markov property of the
jump chain
so
kt = Ei(D
A
) = Ei(J
1
)+LE(DAJ
1
I Y
1
= j)IP\(Y
l
= j) = q;l+L 'lrij
k
1
j#i j#i
and so
 Lqijk1 = 1.
JEI
Suppose now that Y = (Yi : i E I) is another solution to (3.1). Then
kf = Yi = 0 for i E A. Suppose i fj. A, then
1 '"' 1 '"' (1 '"' ) Yi = qi + 'lrijYj = qi + 'lrij qj + 'lrjkYk
= Ei(Sl) + Ei (S2
1
{HA:2:2}) + L L'lrij'lrjkYk.
114 3. Continuoustime Markov chains II
By repeated substitution for y in the final term we obtain after n steps
Yi = lEi(Sd + ... + lEi (Sn
1
{HA:2:n}) + L ... L 1riil .. ·
1r
indnYin·
So, if y is nonnegative
where we use the notation HA /\ n for the minimum of H
A
and n. Now
so, by monotone convergence, Yi E
i
(DA) = kf, as required. D
Exercise
3.3.1 Consider the Markov chain on {I, 2, 3, 4} with generator matrix
Q=
1/6
o
1/2
1/2
o
o
1/2
o
1/3
o
1/6
o.
Calculate (a) the probability of hitting 3 starting from 1, (b) the expected
time to hit 4 starting from 1.
3.4 Recurrence and transience
Let be Markov chain with generator matrix Q. Recall that we insist
be minimal. We say a state i is recurrent if
JP>i({t 0: X
t
= i} is unbounded) = 1.
We say that i is transient if
JP>i( {t 0 : X
t
= i} is unbounded) = O.
Note that if can explode starting from i then i is certainly not
recurrent. The next result shows that, like class structure, recurrence and
transience are determined by the jump chain.
3.4 Recurrence and transience 115
Theorem 3.4.1. We have:
(i) if i is recurrent for the jump chain then i is recurrent for
(ii) if i is transient for the jump chain, then i is transient for
(iii) every state is either recurrent or transient;
(iv) recurrence and transience are class properties.
Proof. (i) Suppose i is recurrent for If X
o
= i then does
not explode and I
n
00 by Theorem 2.7.1. Also X(J
n
) = Y
n
= i infinitely
often, so {t 0 : X
t
= i} is unbounded, with probability 1.
(ii) Suppose i is transient for If X
o
= i then
N = sup{n 0 : Y
n
= i} < 00,
so {t 0: X
t
= i} is bounded by J(N +1), which is finite, with probability
1, because (Y
n
: n N) cannot include an absorbing state.
(iii) Apply Theorem 1.5.3 to the jump chain.
(iv) Apply Theorem 1.5.4 to the jump chain. D
The next result gives continuoustime analogues of the conditions for
recurrence and transience found in Theorem 1.5.3. We denote by T
i
the
first passage time of to state i, defined by
Ti(w) = inf{t J
1
(w) : Xt(w) = i}.
Theorem 3.4.2. The following dichotomy holds:
(i) if qi = 0 or IPi(T
i
< 00) = 1, then i is recurrent and Jo
oo
Pii(t)dt = 00;
(ii) ifqi > 0 and IPi(T
i
< 00) < 1, then i is transient and Jo
oo
Pii(t)dt < 00.
Proof. If qi = 0, then cannot leave i, so i is recurrent, Pii(t) = 1
for all t, and Jo
oo
Pii(t)dt = 00. Suppose then that qi > o. Let N
i
denote
the first passage time of the jump chain to state i. Then
IPi(N
i
< 00) = IPi(T
i
< 00)
so i is recurrent if and only if IPi(T
i
< 00) = 1, by Theorem 3.4.1 and the
corresponding result for the jump chain.
Write 1r}j) for the (i,j) entry in rrn. We shall show that
1
00 1 00
_ (n)
pii(t)dt  : L 1r
ii
o q1, n=O
(3.2)
so that i is recurrent if and only if Jo
oo
Pii(t)dt = 00, by Theorem 3.4.1 and
the corresponding result for the jump chain.
116 3. Continuoustime Markov chains II
To establish (3.2) we use Fubini's theorem (see Section 6.4):
00
= lEi L Sn+l
1
{Yn=i}
n=O
00 1 00
= L lEi (Sn+l I Yn = i)JP>i(Yn = i) = : L 0
n=O n=O
Finally, we show that recurrence and transience are determined by any
discretetime sampling of
Theorem 3.4.3. Let h > 0 be given and set Zn = X
nh
.
(i) If i is recurrent for then i is recurrent for
(ii) If i is transient for then i is transient for
Proof. Claim (ii) is obvious. To prove (i) we use for nh t < (n + l)h the
estimate
which follows from the Markov property. Then, by monotone convergence
and the result follows by Theorems 1.5.3 and 3.4.2. D
Exercise
3.4.1 Customers arrive at a certain queue in a Poisson process of rate A.
The single 'server' has two states A and B, state A signifying that he is 'in
attendance' and state B that he is having a teabreak. Independently of
how many customers are in the queue, he fluctuates between these states
as a Markov chain Y on {A, B} with Qmatrix
(
a a)
(3 (3 .
The total service time for any customer is exponentially distributed with
parameter Jl and is independent of the chain Y and of the service times of
other customers.
3.5 Invariant distributions
Describe the system as a Markov chain X with statespace
117
An signifying that the server is in state A and there are n people in the
queue (including anyone being served) and B
n
signifying that the server is
in state B and there are n people in the queue.
Explain why, for some fJ in (0,1], and k = 0,1,2, ... ,
Show that (fJ  1) f (fJ) = 0, where
By considering f(l) or otherwise, prove that X is transient if 1l(3 < A(a+(3),
and explain why this is intuitively obvious.
3.5 Invariant distributions
Just as in the discretetime theory, the notions of invariant distribution
and measure play an important role in the study of continuoustime Markov
chains. We say that A is invariant if
AQ=O.
Theorem 3.5.1. Let Q be a Qmatrix with jump matrix IT and let A be
a measure. The following are equivalent:
(i) A is invariant;
(ii) Il IT = Il where Ili = Aiqi.
Proof· We have qi ('lrij  bij) = qij for all i, j, so
(Jl(II  I))j = :EJli(1rij  8ij ) = :E,xiqij = (,xQk D
iEI iEI
This tieup with measures invariant for the jump matrix means that we
can use the existence and uniqueness results of Section 1.7 to obtain the
following result.
118 3. Continuoustime Markov chains II
Theorem 3.5.2. Suppose that Q is irreducible and recurrent. Then Q has
an invariant measure A which is unique up to scalar multiples.
Proof. Let us exclude the trivial case I = {i}; then irreducibility forces
qi > 0 for all i. By Theorems 3.2.1 and 3.4.1, II is irreducible and recurrent.
Then, by Theorems 1.7.5 and 1.7.6, II has an invariant measure jj, which is
unique up to scalar multiples. So, by Theorem 3.5.1, we can take Ai = jji/qi
to obtain an invariant measure unique up to scalar multiples. D
Recall that a state i is recurrent if qi = 0 or lPi (Ti < 00) = 1. If qi = 0
or the expected return time mi = Ei(Ti) is finite then we say i is positive
recurrent. Otherwise a recurrent state i is called null recurrent. As in the
discretetime case positive recurrence is tied up with the existence of an
invariant distribution.
Theorem 3.5.3. Let Q be an irreducible Qmatrix. Then the following
are equivalent:
(i) every state is positive recurrent;
(ii) some state i is positive recurrent;
(iii) Q is nonexplosive and has an invariant distribution A.
Moreover, when (iii) holds we have mi = l/(Aiqi) for all i.
Proof. Let us exclude the trivial case I = {i}; then irreducibility forces
qi > 0 for all i. It is obvious that (i) implies (ii). Define Jli = (Jl; : j E I)
by
fTil\(
f . L ~ = Ei 10 1{xs=j}ds,
where T
i
/\ ( denotes the minimum of T
i
and (. By monotone convergence,
L f.L; = Ei(Ti 1\ ().
jEI
Denote by N
i
the first passage time of the jump chain to state i. By Fubini's
theorem
00
f.L; = Ei L Sn+l
1
{Yn=j,n<Nd
n=O
00
= LEi(Sn+l IYn =j)Ei (l{Y
n
=j,n<N
i
})
n=O
00
= ~ l E · '""" 1 .
qJ ~ L.J {Yn=J,n<Ni }
n=O
Nil
= q;l Ei L l{Y
n
=j} = ,j/qj
n=O
3.5 Invariant distributions 119
where, in the notation of Section 1.7, ,,; is the expected time in j between
visits to i for the jump chain.
Suppose (ii) holds, then i is certainly recurrent, so the jump chain is
recurrent, and Q is nonexplosive, by Theorem 2.7.1. We know that "ill =
"i by Theorem 1.7.5, so JliQ = 0 by Theorem 3.5.1. But Jli has finite total
mass
L Jl; = Ei(Ti) = mi
jEI
so we obtain an invariant distribution A by setting Aj = Jl;/mi.
On the other hand, suppose (iii) holds. Fix i E I and set Vj =
Ajqj/(Aiqi); then Vi = 1 and vll = v by Theorem 3.5.1, so Vj ~ ,,; for
all j by Theorem 1.7.6. So
mi = LJl; = L ,;/qj ~ L Vj/Qj
jEI jEI jEI
= 2: Aj/(Aiqi) = l/(Aiqi) < 00
jEI
showing that i is positive recurrent.
To complete the proof we return to the preceding calculation armed
with the knowledge that Q is recurrent, hence 11 is recurrent, Vj = ,,; and
mi = l/(Aiqi) for all i. D
The following example is a caution that the existence of an invariant
distribution for a continuoustime Markov chain is not enough to guarantee
positive recurrence, or even recurrence.
Example 3.5.4
Consider the Markov chain ( X t ) ( ~ O on Z+ with the following diagram,
where qi > 0 for all i and where 0 < A = 1  Jl < 1:
1
AqO
...   ~   .......     
o iI i i+1
The jump chain behaves as a simple random walk away from 0, so ( X t ) ( ~ O
is recurrent if A ~ Jl and transient if A > Jl. To compute an invariant
measure v it is convenient to use the detailed balance equations
for all i, j.
120 3. Continuoustime Markov chains II
Look ahead to Lemma 3.7.2 to see that any solution is invariant. In this
case the nonzero equations read
for all i.
So a solution is given by Vi = q:;l(A/Il)i. If the jump rates qi are constant
then v can be normalized to produce an invariant distribution precisely
when A < Il.
Consider, on the other hand, the case where qi = 2
i
for all i and
1 < AIIl < 2. Then v has finite total mass so ( X t ) t ~ O has an invariant
distribution, but ( X t ) t ~ O is also transient. Given Theorem 3.5.3, the only
possibility is that ( X t ) t ~ O is explosive.
The next result justifies calling measures A with AQ = 0 invariant.
Theorem 3.5.5. Let Q be irreducible and recurrent, and let A be a mea
sure. Let s > 0 be given. The following are equivalent:
(i) AQ = 0;
(ii) AP(S) = A.
Proof. There is a very simple proof in the case of finite statespace: by the
backward equation
d
ds>'P(s) = >.P'(s) = >.QP(s)
so AQ = 0 implies AP(S) = AP(O) = A for all s; P(s) is also recurrent, so
Il P (s) = Il implies that Il is proportional to A, so IlQ = O.
For infinite statespace, the interchange of differentiation with the sum
mation involved in multiplication by A is not justified and an entirely dif
ferent proof is needed.
Since Q is recurrent, it is nonexplosive by Theorem 2.7.1, and P(s) is
recurrent by Theorem 3.4.3. Hence any A satisfying (i) or (ii) is unique up
to scalar multiples; and from the proof of Theorem 3.5.3, if we fix i and set
then IlQ = o. Thus it suffices to show IlP( s) = Il. By the strong Markov
property at T
i
(which is a simple consequence of the strong Markov property
of the jump chain)
3.6 Convergence to equilibrium
Hence, using Fubini's theorem,
l
s
+
Ti
/£j = lEi s l{X
t
=j}dt
= 1
00
JP\(X
s
+t = j, t < Ti)dt
= roo '2:)J.»i(X
t
= k, t < Ti)Pkj(s)dt
Jo kEI
= L(lE
i
(Ti l{Xt=k}dt)Pk
j
(s)
kEI Jo
= :E/£kPkj(S)
kEI
121
as required. D
Theorem 3.5.6. Let Q be an irreducible nonexplosive Qmatrix having
an invariant distribution A. If ( X t ) ( ~ O is Markov(A, Q) then so is ( X s + t ) ( ~ O
for any S 2: o.
Proof. By Theorem 3.5.5, for all i,
IF(X
s
= i) = (AP(S))i = Ai
so, by the Markov property, conditional on X
s
Markov(8
i
, Q). D
3.6 Convergence to equilibrium
We now investigate the limiting behaviour of Pij (t) as t ~ 00 and its relation
to invariant distributions. You will see that the situation is analogous to the
case of discretetime, only there is no longer any possibility of periodicity.
We shall need the following estimate of uniform continuity for the tran
sition probabilities.
Lemma 3.6.1. Let Q be a Qmatrix with semigroup P(t). Then, for all
t, h 2: 0
Proof. We have
Ipij(t + h)Pij(t)1 = I:EPik(h)Pkj(t)  Pij(t)1
kEI
= !:EPik(h)Pkj(t)  (1  pii(h))pij(t) I
k#i
:s; 1  pii(h) :s; lF
i
(J
1
:s; h) = 1  e
qih
. D
122 3. Continuoustime Markov chains II
Theorem 3.6.2 (Convergence to equilibrium). Let Q be an irre
ducible nonexplosive Qmatrix with semigroup P(t), and having an in
variant distribution A. Then for all states i, j we have
Proof Let be Markov(8
i
, Q). Fix h > 0 and consider the hskeleton
Zn = Xnh. By Theorem 2.8.4
so is discretetime Markov(8
i
, P(h)). By Theorem 3.2.1 irreducibil
ity implies pij(h) > 0 for all i,j so P(h) is irreducible and aperiodic. By
Theorem 3.5.5, A is invariant for P(h). So, by discretetime convergence to
equilibrium, for all i, j
Thus we have a lattice of points along which the desired limit holds; we fill
in the gaps using uniform continuity. Fix a state i. Given € > 0 we can
find h > 0 so that
for 0 s h
and then find N, so that
for n 2: N.
For t 2: Nh we have nh t < (n + l)h for some n 2: Nand
by Lemma 3.6.1. Hence
as
D
The complete description of limiting behaviour for irreducible chains in
continuous time is provided by the following result. It follows from Theorem
1.8.5 by the same argument we used in the preceding result. We do not
give the details.
3. 7 Time reversal 123
Theorem 3.6.3. Let Q be an irreducible Qmatrix and let v be any dis
tribution. Suppose that is Markov(v, Q). Then
as t 00 for all j E I
where mj is the expected return time to state j.
Exercises
3.6.1 Find an invariant distribution A for the Qmatrix
Q = !4
2 1 3
and verify that = Al using your answer to Exercise 2.1.1.
3.6.2 In each of the following cases, compute P(X
t
= 21X
o
= 1) for
the Markov chain with the given Qmatrix on {1,2,3,4}:
(a)
(c)
(1
2 1 1
11)
1 1
0 1
0 0
(1
1 1 0
]2)
1 0
0 2
0 2
(b)
(d)
(1
2 1 1
1 1
0 1
0 0
(1
2 1 0
2 2
1 1
0 0
3.6.3 Customers arrive at a singleserver queue in a Poisson stream of rate
A. Each customer has a service requirement distributed as the sum of two
independent exponential random variables of parameter Jl. Service require
ments are independent of one another and of the arrival process. Write
down the generator matrix Q of a continuoustime Markov chain which
models this, explaining what the states of the chain represent. Calculate
the essentially unique invariant measure for Q, and deduce that the chain
is positive recurrent if and only if AIIl < 1/2.
3. 7 Time reversal
Time reversal of continuoustime chains has the same features found in the
discretetime case. Reversibility provides a powerful tool in the analysis
of Markov chains, as we shall see in Section 5.2. Note in the following
124 3. Continuoustime Markov chains II
result how time reversal interchanges the roles of backward and forward
equations. This echoes our proof of the forward equation, which rested on
the time reversal identity of Lemma 2.8.5.
A small technical point arises in time reversal: rightcontinuous processes
become leftcontinuous processes. For the processes we consider, this is
unimportant. We could if we wished redefine the timereversed process
to equal its right limit at the jump times, thus obtaining again a right
continuous process. We shall suppose implicitly that this is done, and
forget about the problem.
Theorem 3.7.1. Let Q be irreducible and nonexplosive and suppose
that Q has an invariant distribution A. Let T E (0,00) be given and let
(Xt)O<t<T be Markov(A, Q). Set X
t
= X
T

t
. Then the process (Xt)O<t<T
is Q), where Q= : i,j E 1) is given by = Aiqijo Mo;e
over, Qis also irreducible and nonexplosive with invariant distribution A.
Proof. By Theorem 2.8.6, the semigroup (P(t) : t 2: 0) of Q is the minimal
nonnegative solution of the forward equation
P'(t) = P(t)Q, P(O) = I.
Also, for all t > 0, P(t) is an irreducible stochastic matrix with invariant
distribution A. Define P(t) by
AjPJi(t) = AiPij(t),
then P(t) is an irreducible stochastic matrix with invariant distribution A,
and we can rewrite the forward equation transposed as
P'(t) = QP(t) .
.
But this is the backward equation for Q, which is itself a Qmatrix, and
P(t) is then its minimal nonnegative solution. Hence Qis irreducible and
nonexplosive and has invariant distribution A.
Finally, for 0 = to < ... < t
n
= T and Sk = tk  tk1, by Theorem 2.8.4
we have
P(X
to
= io, ... ,X
tn
= in) = P(XTto = io, ... ,X
T

tn
= in)
= AinPininl (sn) ... Pi
l
io (Sl)
= AioPioi
l
(Sl) ... Pinlin (Sn)
SO, by Theorem 2.8.4 again, is Markov(A, Q). D
The chain is called the timereversal of
A Qmatrix Q and a measure A are said to be in detailed balance if
for all i, j.
3.8 Ergodic theorem 125
Lemma 3.7.2. IfQ and A are in detailed balance then A is invariant for
Q.
Proof· We have (AQ)i = EjEI Ajqji = EjEI Aiqij = 0. D
Let ( X t ) ( ~ O be Markov(A, Q), with Q irreducible and. nonexplosive.
We say that ( X t ) ( ~ O is reversible if, for all T > 0, (XTt)OstsT is also
Markov(A, Q).
Theorem 3.7.3. Let Q be an irreducible and nonexplosive Qmatrix and
let A be a distribution. Suppose that ( X t ) ( ~ O is Markov(A, Q). Then the
following are equivalent:
(a) ( X t ) ( ~ O is reversible;
(b) Q and A are in detailed balance.
Proof. Both (a) and (b) imply that A is invariant for Q. Then both (a) and
(b) are equivalent to the statement that Q= Q in Theorem 3.7.1. D
Exercise
3.7.1 Consider a fleet of N buses. Each bus breaks down independently
at rate Jl, when it is sent to the depot for repair. The repair shop can
only repair one bus at a time and each bus takes an exponential time of
parameter A to repair. Find the equilibrium distribution of the number of
buses in service.
3.7.2 Calls arrive at a telephone exchange as a Poisson process of rate A,
and the lengths of calls are independent exponential random variables of
parameter Jl. Assuming that infinitely many telephone lines are available,
set up a Markov chain model for this process.
Show that for large t the distribution of the number of lines in use at
time t is approximately Poisson with mean AIJl.
Find the mean length of the busy periods during which at least one line
is in use.
Show that the expected number of lines in use at time t, given that n
are in use at time 0, is neJLt + A(l  eJLt)1Jl.
Show that, in equilibrium, the number Nt of calls finishing in the time
interval [0, t] has Poisson distribution of mean At.
Is ( N t ) t ~ o a Poisson process?
3.8 Ergodic theorem
Longrun averages for continuoustime chains display the same sort of be
haviour as in the discretetime case, and for similar reasons. Here is the
result.
126 3. Continuoustime Markov chains II
Theorem 3.8.1 (Ergodic theorem). Let Q be irreducible and let v be
any distribution. If is Markov(v, Q), then
]p> t l{x
s
=i}ds _1_ as t 00) = 1
t io miqi
where mi = IEi(T
i
) is the expected return time to state i. Moreover, in the
positive recurrent case, for any bounded function f : I lR we have
where
and where (Ai: i E I) is the unique invariant distribution.
Proof. If Q is transient then the total time spent in any state i is finite, so
1 t 1 1
t Jo l{xs =i}ds::; t Jo l{xs =i}ds 0 = mi ·
Suppose then that Q is recurrent and fix a state i. Then hits i
with probability 1 and the longrun proportion of time in i equals the long
run proportion of time in i after first hitting i. So, by the strong Markov
property (of the jump chain), it suffices to consider the case v = bi.
Denote by Mi the length of the nth visit to i, by Tin the time of the
nth return to i and by Li the length of the nth excursion to i. Thus for
n = 0,1,2, ... , setting Tp = 0, we have
M;+l = inf{t > Tin: X
t
=1= i}  Tin
T
i
n
+
1
= inf{t > Tin + M;+l : X
t
= i}
= T?1'+l  T!"
.
By the strong Markov property (of the jump chain) at the stopping times
Tin for n 0 we find that L}, , . .. are independent and identically dis
tributed with mean mi, and that Ml, M?, ... are independent and identi
cally distributed with mean l/qi. Hence, by the strong law of large numbers
(see Theorem 1.10.1)
+ ... + Lr:t
mi
n
+ ... +M!L 1
n qi
and hence
3.8 Ergodic theorem
+... +M!L 1
L} +··· +Li miqi
127
with probability 1. In particular, we note that Tin /T
i
n
+
1
1 as n 00
with probability 1. Now, for Tin t < T
i
n
+
1
we have
T!L +··· +M!L 1 it T!"+l +··· +M!"+l
< 1
T!"+l +···+Lr:t  t 0  T!L +... +L,,:+l
so on letting t 00 we have, with probability 1
In the positive recurrent case we can write
where Ai = 1/(miqi). We conclude that
1 it
 f(Xs)ds 1
t 0
with probability 1, by the same argument as was used in the proof of The
orem 1.10.2. D
4
Further theory
In the first three chapters we have given an account of the elementary theory
of Markov chains. This already covers a great many applications, but is
just the beginning of the theory of Markov processes. The further theory
inevitably involves more sophisticated techniques which, although having
their own interest, can obscure the overall structure. On the other hand,
the overall structure is, to a large extent, already present in the elementary
theory. We therefore thought it worth while to discuss some features of the
further theory in the context of simple Markov chains, namely, martingales,
potential theory, electrical networks and Brownian motion. The idea is that
the Markov chain case serves as a guiding metaphor for more complicated
processes. So the reader familiar with Markov chains may find this chapter
helpful alongside more general higherlevel texts. At the same time, further
insight is gained into Markov chains themselves.
4.1 Martingales
A martingale is a process whose average value remains constant in a par
ticular strong sense, which we shall make precise shortly. This is a sort of
balancing property. Often, the identification of martingales is a crucial step
in understanding the evolution of a stochastic process.
We begin with a simple example. Consider the simple symmetric random
walk ( X n ) n ~ O on Z, which is a Markov chain with the following diagram
4.1 Martingales 129
iI
1 1
2 2
I( • •
i i+l
The average value of the walk is constant; indeed it has the stronger prop
erty that the average value of the walk at some future time is always simply
the current value. In precise terms we have
and the stronger property says that, for n m,
E(X
n
 X
m
I X
o
= io, ... ,X
m
= i
m
) = o.
This stronger property says that is in fact a martingale.
Here is the general definition. Let us fix for definiteness a Markov chain
and write F
n
for the collection of all sets depending only on
X
o
, ... ,X
n
. The sequence is called the filtration of and
we think of F
n
as representing the state of knowledge, or history, of the
chain up to time n. A process is called adapted if M
n
depends
only on X
o
, . .. ,X
n
. A process is called integmble if EIMnl < 00
for all n. An adapted integrable process is called a martingale if
for all A E F
n
and all n. Since the collection F
n
consists of countable
unions of elementary events such as
this martingale property is equivalent to saying that
for all i
o
, . .. ,in and all n.
A third formulation of the martingale property involves another notion of
conditional expectation. Given an integrable random variable Y, we define
lE(Y I F
n
) = L lE(Y I Xo = i o,· .. ,Xn = i n )l{xo=io, ... ,Xn=i
n
}'
io,··· ,in
130 4. Further theory
The random variable E(Y I F
n
) is called the conditional expectation of Y
given F
n
. In passing from Y to E(Y I F
n
), what we do is to replace on
each elementary event A E F
n
, the random variable Y by its average value
E(Y I A). It is easy to check that an adapted integrable process ( M n ) n ~ o
is a martingale if and only if
Conditional expectation is a partial averaging, so if we complete the process
and average the conditional expectation we should get the full expectation
E(E(Y I F
n
)) = E(Y).
It is easy to check that this formula holds.
In particular, for a martingale
so, by induction
This was already clear on taking A = n in our original definition of a
martingale.
We shall prove one general result about martingales, then see how it
explains some things we know about the simple symmetric random walk.
Recall that a random variable
T : n ~ {O, 1, 2, ... } u {(X)}
is a stopping time if {T = n} E F
n
for all n < 00. An equivalent condition
is that {T ::; n} E F
n
for all n < 00. Recall from Section 1.4 that all sorts
of hitting times are stopping times.
Theorem 4.1.1 (Optional stopping theorem). Let ( M n ) n ~ o be a
martingale and let T be a stopping time. Suppose that at least one of the
following conditions holds:
(i) T::; n for some n;
(ii) T < 00 and IMnl ~ C whenever n ~ T.
Then lEMr = lEMo.
Proof. Assume that (i) holds. Then
M
r
 M
o
= (M
r
 M
r
 1 ) + ···+ (M1  Mo)
nl
= :2)Mk+l  Mk)lk<T.
k=O
4.1 Martingales
Now {k.< T} = {T k}C E Fk since T is a stopping time, and so
since is a martingale. Hence
nl
EM
T
 EM
o
= L E[(Mk+l  Mk)lk<T] = O.
k=O
131
If we do not assume (i) but (ii), then the preceding argument applies to the
stopping time T 1\ n, so that IEMTAn = IEM
o
. Then
IIEM
T
 IEMol = IIEM
T
 IEMTAnl IEIM
T
 MTAnl 2CJP>(T > n)
for all n. But JP>(T > n) 0 as n 00, so EM
T
= EM
o
. D
Returning to the simple symmetric random walk suppose that
X
o
= 0 and we take
T = inf{n 0 : X
n
= a or X
n
= b}
where a, bEN are given. Then T is a stopping time and T < 00 by
recurrence of finite closed classes. Thus condition (ii) of the optional stop
ping theorem applies with M
n
= X
n
and C = a V b. We deduce that
IEXT = EX
o
= O. So what? Well, now we can compute
p = JP>(X
n
hits a before b).
We have X
T
= a with probability p and X
T
= b with probability 1  p,
so
o= EXT = p( a) + (1  p)b
giving
p = b/(a +b).
There is an entirely different, Markovian, way to compute p, using the
methods of Section 1.4. But the intuition behind the result EXT = 0 is
very clear: a gambler, playing a fair game, leaves the casino once losses
reach a or winnings reach b, whichever is sooner; since the game is fair, the
average gain should be zero.
We discussed in Section 1.3 the counterintuitive case of a gambler who
keeps on playing a fair game against an infinitely rich casino, with the
certain outcome of ruin. This game ends at the finite stopping time
T = inf{n 0 : X
n
= a}
132 4. Further theory
where a is the gambler's initial fortune. Since XT = a we have
EXT = a =I 0 = EX
o
but this does not contradict the optional stopping theorem because neither
condition (i) nor condition (ii) is satisfied. Thus, while intuition might
suggest that EXT = EX
o
is rather obvious, some care is needed as it is not
always true.
The example just discussed was rather special in that the chain
itself was a martingale. Obviously, this is not true in general; indeed a
martingale is necessarily realvalued and we do not in general insist that
the statespace I is contained in JR. Nevertheless, to every Markov chain is
associated a whole collection of martingales, and these martingales charac
terize the chain. This is the basis of a deep connection between martingales
and Markov chains.
We recall that, given a function f : I JR and a Markov chain
with transition matrix P, we have
(pn J)(i) = Ii = lEi (J(X
n
)).
jEI
Theorem 4.1.2. Let be a random process with values in I and
let P be a stochastic matrix. Write for the filtration of
Then the following are equivalent:
(i) is a Markov chain with transition matrix P;
(ii) for all bounded functions f : I JR, the following process is a mar
tingale:
nl
= f(X
n
)  f(X
o
)  L (P  I)f(X
m
).
m=O
Proof. Suppose (i) holds. Let f be a bounded function. Then
I(PJ)(i)1 = ILPij1i1 ::; llil
jEI J
so
2(n+1)suplfjl < 00
j
showing that is integrable for all n.
Let A = {X
o
= i
o
, ... ,X
n
= in}. By the Markov property
4.1 Martingales
so
 M! I A) = E[f(X
n
+
l
)  (Pf)(X
n
) I A] = 0
and so is a martingale.
On the other hand, if (ii) holds, then
for all bounded functions f. On taking f = l{i
n
+l} we obtain
133
so is Markov with transition matrix P. D
Some more martingales associated to a Markov chain are described in
the next result. Notice that we drop the requirement that f be bounded.
Theorem 4.1.3. Let be a Markov chain with transition matrix
P. Suppose that a function f : N x I lR satisfies, for all n 0, both
and
(Pf)(n+ 1,i) = LPijf(n+ 1,j) = f(n,i).
JEI
Then M
n
= f(n, X
n
) is a martingale.
Proof. We have assumed that M
n
is integrable for all n. Then, by the
Markov property
E(M
n
+
1
 M
n
I X
o
= i
o
,· .. ,X
n
= in) = E
in
[f(n +1, Xl)  f(n, X
o
)]
= (Pf)(n +1, in)  f(n, in) = O.
So is a martingale. D
Let us see how this theorem works in the case where is a simple
random walk on Z, starting from O. We consider f (i) = i and g(n, i) =
i
2
 n. Since IXnl n for all n, we have
Also
(Pf)(i) = (i  1)/2 + (i +1)/2 = i = f(i),
(Pg)(n +1, i) = (i  1)2/2 + (i  1)2/2  (n +1) = i
2
 n = g(n, i).
Hence both X
n
= f(X
n
) and Y
n
= g(n, X
n
) are martingales.
134 4. Further theory
In order to put this to some use, consider again the stopping time
T = inf{n ~ 0 : X
n
= a or X
n
= b}
where a, bEN. By the optional stopping theorem
Hence
On letting n + 00, the left side converges to E(T), by monotone conver
gence, and the right side to E(Xf) by bounded convergence. So we obtain
We have given only the simplest examples of the use of martingales in
studying Markov chains. Some more will appear in later sections. For
an excellent introduction to martingales and their applications we recom
mend Probability with Martingales by David Williams (Cambridge Univer
sity Press, 1991).
Exercise
4.1.1 Let ( X n ) n ~ O be a Markov chain on I and let A be an absorbing set
in I. Set
T = inf{n ~ 0: X
n
E A}
and
hi = IPi(X
n
E A for some n ~ 0) = IPi(T < 00).
Show that M
n
= h(X
n
) is a martingale.
4.2 Potential theory
Several physical theories share a common mathematical framework, which is
known as potential theory. One example is Newton's theory of gravity, but
potential theory is also relevant to electrostatics, fluid flow and the diffusion
of heat. In gravity, a distribution of mass, of density p say, gives rise to a
gravitational potential ¢, which in suitable units satisfies the equation
Da¢ = p,
where ~ = 8
2
/8x
2
+8
2
/8y2 +8
2
/ 8z
2
. The potential ¢ is felt physically
through its gradient
'\l¢ = (8¢ 8¢ 8¢)
8x' 8y' 8z
4.2 Potential theory 135
which gives the force of gravity acting on a particle of unit mass. Markov
chains, where space is discrete, obviously have no direct link with this the
ory, in which space is a continuum. An indirect link is provided by Brownian
motion, which we shall discuss in Section 4.4.
In this section we are going to consider potential theory for a count
able statespace, which has much of the structure of the continuum version.
This discrete theory amounts to doing Markov chains without the proba
bility, which has the disadvantage that one loses the intuitive picture of the
process, but the advantage of wider applicability. We shall begin by intro
ducing the idea of potentials associated to a Markov chain, and by showing
how to calculate these potentials. This is a unifying idea, containing within
it other notions previously considered such as hitting probabilities and ex
pected hitting times. It also finds application when one associates costs to
Markov chains in modelling economic activity: see Section 5.4.
Once we have established the basic link between a Markov chain and its
associated potentials, we shall briefly run through some of the main features
of potential theory, explaining their significance in terms of Markov chains.
This is the easiest way to appreciate the general structure of potential
theory, unobscured by technical difficulties. The basic ideas of boundary
theory for Markov chains will also be introduced.
Before we embark on a general discussion of potentials associated to
a Markov chain, here are two simple examples. In these examples the
potential cP has the interpretation of expected total cost.
1
2
5
4
Example 4.2.1
Consider the discretetime random walk on the directed graph shown above,
which at each step choses among the allowable transitions with equal proba
bility. Suppose that on each visit to states i = 1,2,3,4 a cost Ci is incurred,
where Ci = i. What is the fair price to move from state 3 to state 4?
The fair price is always the difference in the expected total cost. We
denote by cPi the expected total cost starting from i. Obviously, cPs = 0 and
136 4. Further theory
by considering the effect of a single step we see that
¢l = 1 + ¢2,
¢2 = 2 + ¢3,
¢3 = 3 + +
¢4 = 4.
Hence ¢3 = 8 and the fair price to move from 3 to 4 is 4.
We shall now consider two variations on this problem. First suppose
our process is, instead, the continuoustime random walk on the
same directed graph which makes each allowable transition at rate 1, and
suppose cost is incurred at rate Ci = i in state i for i = 1,2,3,4. Thus the
total cost is now
1
00
c(Xs)ds.
What now is the fair price to move from 3 to 4? The expected cost incurred
on each visit to i is given by Ci/ qi and ql = 1, q2 = 1, q3 = 3, q4 = 1. So we
see, as before
¢l = 1 + ¢2,
¢2 = 2 + ¢3,
¢3 = i + +
¢4 = 4.
Hence ¢3 = 5 and the fair price to move from 3 to 4 is 1.
1
2
5
4
In the second variation we consider the discretetime random walk
on the modified graph shown above. Where there is no arrow,
transitions are allowed in both directions. Obviously, states 1 and 5 are
absorbing. We impose a cost Ci = i on each visit to i for i = 2, 3, 4, and a
final cost Ii on arrival at i = 1 or 5, where Ii = i. Thus the total cost is
now
Tl
L c(X
n
) + f(X
T
)
n=O
4.2 Potential theory 137
where T is the hitting time of {I, 5}. Write, as before, cPi for the expected
total cost starting from i. Then cPl = 1, cPs = 5 and
cP2 = 2 + ~ ( c P l + cP3),
cP3 = 3 + ~ ( c P l +cP2 + cP4 + cPs),
cP4 = 4 + ~ ( c P 3 + cPs).
On solving these equations we obtain cP2 = 7, cP3 = 9 and cP4 = 11. So in
this case the fair price to move from 3 to 4 is 2.
Example 4.2.2
Consider the simple discretetime random walk on Z with transition prob
abilities Pi,il = q < P = Pi,i+l. Let c > 0 and suppose that a cost c
i
is
incurred every time the walk visits state i. What is the expected total cost
cPo incurred by the walk starting from O?
We must be prepared to find that cPo = 00 for some values of c, as the
total cost is a sum over infinitely many times. Indeed, we know that the
walk X
n
~ 00 with probability 1, so for c ~ 1 we shall certainly have
cPo = 00.
Let cPi denote the expected total cost starting from i. On moving one
step to the right, all costs are multiplied by c, so we must have
By considering what happens on the first step, we see
cPo = 1 + PcPl + qcPl = 1 + (cp + q/c)cPo.
Note that cPo = 00 always satisfies this equation. We shall see in the general
theory that cPo is the minimal nonnegative solution. Let us look for a finite
solution: then
so
c
cPo = .
c  c
2
p  q
The quadratic c
2
p  c + q has roots at q/p and 1, and takes negative values
in between. Hence the expected total cost is given by
{
c/(c c
2
pq)
cPo =
00
if c E (q/p, l)
otherwise.
138 4. Further theory
It was clear at the outset that cPo = 00 when c ~ 1. It is interesting that
cPo = 00 also when c is too small: in this case the costs rapidly become large
to the left of 0, and although the walk eventually drifts away to the right,
the expected cost incurred to the left of 0 is infinite.
In the examples just discussed we were able to calculate potentials by
writing down and solving a system of linear equations. This situation is
familiar from hitting probabilities and expected hitting times. Indeed, these
are simple examples of potentials for Markov chains. As the examples show,
one does not really need a general theory to write down the linear equations.
Nevertheless, we are now going to give some general results on potentials.
These will help to reveal the scope of the ideas used in the examples, and
will reveal also what happens when the linear equations do not have a
unique solution. We shall discuss the cases of discrete and continuous time
sidebyside. Throughout, we shall write ( X n ) n ~ O for a discretetime chain
with transition matrix P, and ( X t ) t ~ O for a continuoustime chain with
generator matrix Q. As usual, we insist that ( X t ) ( ~ O be minimal.
Let us partition the statespace I into two disjoint sets D and aD; we call
aD the boundary. We suppose that functions (Ci : i E D) and (Ii: i E aD)
are given. We shall consider the associated potential, defined by
in discrete time, and in continuous time
where T denotes the hitting time of aD. To be sure that the sums and
integrals here are well defined, we shall assume for the most part that c
and I are nonnegative, that is, Ci ~ 0 for all i E D and Ii ~ 0 for all
i E aD. More generally, ¢ is the difference of the potentials associated with
the positive and negative parts of c and I, so this assumption is not too
restrictive. In the explosive case we always set c(00) = 0, so no further
costs are incurred after explosion.
The most obvious interpretation of these potentials is in terms of cost:
the chain wanders around in D until it hits the boundary: whilst in D, at
state i say, it incurs a cost Ci per unit time; when and if it hits the boundary,
at j say, a final cost Ij is incurred. Note that we do not assume the chain
will hit the boundary, or even that the boundary is nonempty.
4.2 Potential theory 139
Theorem 4.2.3. Suppose that (Ci : i E D) and (Ii: i E aD) are non
negative. Set
where T denotes the hitting time of aD. Then
(i) the potential ¢ = (¢i : i E I) satisfies
{
¢ = P¢+ C
¢=I
(ii) if'l/J = ('l/Ji : i E I) satisfies
in D
in aD;
(4.1)
{
'l/J ~ P'l/J+c in D
'l/J ~ I in aD
(4.2)
and 'l/Ji ~ 0 for all i, then 'l/Ji ~ ¢i for all i;
(iii) if IPi(T < 00) = 1 for all i, then (4.1) has at most one bounded
solution.
Proof. (i) Obviously, ¢ = f on aD. For i E D by the Markov property
IEi ( L c(Xn ) + f(XT )IT<oo Xl = j)
l ~ n < T
= IEj (L c(Xn ) + f(XT )lT <OO) = ¢j
n<T
so we have
<Pi = Ci +LpijlE ( L c(Xn ) + f(XT )lT <oo Xl = j)
jEI l ~ n < T
=Ci+ LPWPj
jEI
as required.
(ii) Consider the expected cost up to time n:
140 4. Further theory
By monotone convergence, ¢i(n) i ¢i as n 00. Also, by the argument
used in part (i), we find
{
¢(n + 1) = c + P¢(n) in D
¢(n + 1) = f in aD.
Suppose that 1/J satisfies (4.2) and 1/J 0 = ¢(O). Then 1/J P1/J + c
P¢(O) + c = ¢(1) in D and 1/J f = ¢(1) in aD, so 1/J ¢(1). Similarly
and by induction, 1/J ¢(n) for all n, and hence 1/J ¢.
(iii) We shall show that if 1/J satisfies (4.2) then
with equality if equality holds in (4.2). This is another proof of (ii). But
also, in the case of equality, if l1/Ji I M and Pi (T < 00) = 1 for all i, then
so 1/J = ¢(n) = ¢, proving (iii).
For i E D we have
'ljJi Ci + L Pij!j + LPij'ljJj
jEaD JED
and, by repeated substitution for 1/J on the right
'ljJi Ci + L Pijli + LPijCj
jEaD JED
+ ···+ 2: ... 2: Piil'" Pjn2jnl Cjn_l
jlED jnIED
+ L '" L L PijI'" Pjnlin !jn
jlED jnIEDjnEaD
+ 2: ... L Pijl · · · Pjnlin'ljJjn
jlED jnED
= E
i
(c(XO)lT>O + !(XdlT=l + C(Xdlr>l
+ ···+C(Xn1)lT>nl +!(Xn)lT=n +'ljJ(Xn)lT>n)
= ¢i(n  1) + IE
i
as required, with equality when equality holds in (4.2). D
4.2 Potential theory 141
It is illuminating to think of the calculation we have just done in terms
of martingales. Consider
nl
Mn = L c(Xk)lk<T + f(XT )IT<n +'ljJ(Xn)ln::;T'
k==O
Then
nl
lE(Mn+l I Fn) = L c(Xk)lk<T + f(X
T
)lT<n
k==O
+(P1jJ + c)(X
n
)lT>n + I(X
n
)lT==n
:::;M
n
with equality if equality holds in (4.2). We note that M
n
is not necessarily
integrable. Nevertheless, it still follows that
with equality if equality holds in (4.2).
For continuoustime chains there is a result analogous to Theorem 4.2.3.
We have to state it slightly differently because when cP takes infinite val
ues the equations (4.3) may involve subtraction of infinities, and therefore
not make sense. Although the conclusion then appears to depend on the
finiteness of cP, which is a priori unknown, we can still use the result to
determine cPi in all cases. To do this we restrict our attention to the set of
states J accessible from i. If the linear equations have a finite nonnegative
solution on J, then (cPj : j E J) is the minimal such solution. If not, then
cPj = 00 for some j E J, which forces cPi = 00, since i leads to j.
Theorem 4.2.4. Assume that ( X t ) t ~ O is minimal, and that (Ci : i E D)
and (Ii: i E aD) are nonnegative. Set
where T is the hitting time of aD. Then cP = (cPi : i E I), if finite, is the
minimal nonnegative solution to
{
QcP = C
cP=1
in D
in aD.
(4.3)
142 4. Further theory
If ¢i = 00 for some i, then (4.3) has no finite nonnegative solution. More
over, if IPi(T < 00) == 1 for all i, then (4.3) has at most one bounded
solution.
Proof. Denote by (Yn)n>O and 8
1
,8
2
, ... the jump chain and holding times
of and by IT the jump matrix. Then
iTc(Xt)dt +!(XT)lT<oo = L c(Yn)Sn+l +!(YN)lN<oo
o n<N
where N is the first time hits aD, and where we use the convention
ox 00 = 0 on the right. We have
()
") __ { Cj / qj
lE c(Yn Sn+l I Yn = J = Cj = 0
so, by Fubini's theorem
if Cj > 0
if Cj = 0,
¢i = lEi(L c(Yn) +!(YN)lN<oo).
n<N
By Theorem 4.2.3, ¢ is therefore the minimal nonnegative solution to
{
¢ = IT¢ + c in D
¢ == f in aD,
(4.4)
which equations have at most one bounded solution if IPi(N < 00) = 1 for
all i. Since the finite solutions of (4.4) are exactly the finite solutions of
(4.3), and since N is finite whenever T is finite, this proves the result. D
It is natural in some economic applications to apply to future costs a
discount factor a E (0, 1) or rate A E (0, 00), corresponding to an interest
rate. Potentials with discounted costs may also be calculated by linear
equations; indeed the discounting actually makes the analysis easier.
Theorem 4.2.5. Suppose that (Ci : i E I) is bounded. Set
00
¢i = lEi L anc(X
n
)
n==O
then ¢ = (¢i : i E I) is the unique bounded solution to
¢ = aP¢+c.
4.2 Potential theory
Proof. Suppose that ICil ::; C for all i, then
00
I<Pi I :::; CLan = C/(l a)
n==O
so ¢ is bounded. By the Markov property
Then
00
<Pi = lEi L anc(X
n
)
n==O
= ci +a LPijlE (fanIc(X
n
) Xl = j)
jEI n==l
= Ci +a LPij<Pj,
jEI
so
143
¢ = c+aP¢.
On the other hand, suppose that 1/J is bounded and also that 1/J = c + aP1/J.
Set M = sUPi l1/Ji  ¢il, then M < 00. But
1/J  ¢ = aP(1/J  ¢)
so
I'l/Ji  <Pi I :::; a LPijl'l/Jj  <pjl :::; aM.
jEI
Hence M ::; aM, which forces M = 0 and 1/J = ¢. D
We have a similar looking result for continuous time, which however lies
a little deeper, because it really corresponds to a version of the discretetime
result where the discount factor may depend on the current state.
Theorem 4.2.6. Assume that ( X t ) t ~ O is nonexplosive. Suppose that
(Ci : i E I) is bounded. Set
<Pi = lEi 1
00
e>..tc(Xt)dt,
then ¢ = (¢i : i E I) is the unique bounded solution to
(A  Q)¢ = c. (4.5)
144 4. Further theory
Proof. Assume for now that c is nonnegative. Introduce a new state awith
Ca = O. Let T be an independent E(X) random variable and define
 {Xt for t < T
X
t
=
a for t ~ T.
Then ( X t ) t ~ O is a Markov chain on I U {a} with modified transition rates
Qi = qi + X, Qia = X, qa = O.
Also T is the hitting time of a, and is finite with probability 1. By Fubini's
theorem
<Pi = Ei l
T
c(Xt)dt.
Suppose Ci :::; C for all i, then
so ¢ is bounded. Hence, by Theorem 4.2.4, ¢ is the unique bounded solution
to
Q¢ = c,
which is the same as (4.5).
When C takes negative values we can apply the preceding argument to
the potentials
<P; = E
i
1
00
e>.tc±(Xt)dt
where ct = (±c) V o. Then ¢ = ¢+  ¢ so ¢ is bounded. We have
so, subtracting
(X  Q)¢ = c.
Finally, if 1jJ is bounded and (XQ)1jJ = c, then (XQ)(1jJ¢) = 0, so 1jJ¢
is the unique bounded solution for the case when c = 0, which is O. D
The point of view underlying the last four theorems was that we were
interested in a given potential associated to a Markov chain, and wished to
calculate it. We shall now take a brief look at some structural aspects of
the set of all potentials of a given Markov chain. What we describe is just
the simplest case of a structure of great generality. First we shall look at
the Green matrix, and then at the role of the boundary.
4.2 Potential theory 145
Let us consider potentials with nonnegative costs c, and without bound
ary. The potential is defined by
00
¢i = lEi L c(X
n
)
n==O
in discrete time, and in continuous time
By Fubini's theorem we have
00 00
¢i = L lEic(X
n
) = L(pnC)i = (GC)i
n==O n==O
where G = (9ij : i, j E I) is the Green matrix
00
Similarly, in continuous time ¢ = Gc, with
G = 1
00
P(t)dt.
Thus, once we know the Green matrix, we have explicit expressions for
all potentials of the Markov chain. The Green matrix is also called the
fundamental solution of the linear equations (4.1) and (4.3). The jth column
(9ij : i E I) is itself a potential. We have
00
9ij = lEi L lXn=j
n==O
in discrete time, and in continuous time
Thus 9ij is the expected total time in j starting from i. These quantities
have already appeared in our discussions of transience and recurrence in
Sections 1.5 and 2.11: we know that 9ij = 00 if and only if i leads to j and
j is recurrent. Indeed, in discrete time
146 4. Further theory
where hi is the probability of hitting j from i, and /j is the return proba
bility for j. The formula for continuous time is
For potentials with discounted costs the situation is similar: in discrete
time
00 00
<Pi = lEi L anc(X
n
) = L anlEic(X
n
) = (RaC)i
n=O n=O
where
and in continuous time
where
R>.. = 1
00
e>..t P(t)dt.
We call (R
Q
: Q: E (0,1)) and (R>.. : ,\ E (0,00)) the resolvent of the Markov
chain. Unlike the Green matrix the resolvent is always finite. Indeed, for
finite statespace we have
and
We return to the general case, with boundary aD. Any bounded function
(¢i : i E I) for which
¢ = P¢ in D
is called harmonic in D. Our object now is to examine the relation between
nonnegative functions, harmonic in D, and the boundary aD. Here are
two examples.
a
4.2 Potential theory
, ~ , ~

.....
......

b
147
Example 4.2.7
Consider the random walk ( X n ) n ~ O on the above graph, where each allow
able transition is made with equal probability. States a and bare absorbing.
We set aD = {a, b}. Let hi denote the absorption probability for a, starting
from i. By the method of Section 1.3 we find
1/2 5/12)
1/2 1/3
1/2 0
in D
where we have written the vector h
a
as a matrix, corresponding in an ob
vious way to the statespace. The linear equations for the vector h
a
read
{
ha = Ph
a
h ~ = 1, hg = o.
Thus we can find two nonnegative functions h
a
and h
b
, harmonic in D,
but with different boundary values. In fact, the most general nonnegative
harmonic function ¢ in D satisfies
{
¢=P¢
¢=f
where fa' fb 2:: 0, and this implies
in D
in aD
Thus the boundary points a and b give us extremal generators h
a
and h
b
of the set of all nonnegative harmonic functions.
148 4. Further theory
Example 4.2.8
Consider the random walk on Z which jumps towards 0 with prob
ability q and jumps away from 0 with probability p = 1  q, except that at
oit jumps to lor 1 with probability 1/2. We choose p > q so that the
walk is transient. In fact, starting from 0, we can show that is
equally likely to end up drifting to the left or to the right, at speed p  q.
Let us consider the problem of determining for the set C of all
nonnegative harmonic functions cP. We must have
cPi = PcPi+l + qcPil
cPo = +
cPi = qcPi+l + PcPil
The first equation has general solution
for i = 1, 2, . .. ,
for i = 1,  2, . .. .
cPi = A + B(l  (q/p)i) for i = 0,1,2, ... ,
which is nonnegative provided A + B O. Similarly, the third equation
has general solution
cPi = A' + B' (1  (q/p)i) for i = 0, 1, 2, ... ,
nonnegative provided A' + B' O. To obtain a general harmonic function
we must match the values cPo and satisfy cPo = (cPl + cPl)/2. This forces
A = A' and B + B' = O. It follows that all nonnegative harmonic functions
have the form
where f, f+ 0 and where hi = and
+ _ { + (1  (q/p)i)
hi  1 1 ( ")
2  2 1  (q/p)t
for i = 0, 1, 2, ... ,
for i = 1,  2, . .. .
In the preceding example the generators of C were in onetoone corre
spondence with the points of the boundary  the possible places for the
chain to end up. In this example there is no boundary, but the generators
of C still correspond to the two possibilities for the longtime behaviour of
the chain. For we have
The suggestion of this example, which is fully developed other works, is
that the set of nonnegative harmonic functions may be used to identify a
4.2 Potential theory 149
generalized notion of boundary for Markov chains, which sometimes just
consists of points in the statespace, but more generally corresponds to the
varieties of possible limiting behaviour for X
n
as n ~ 00. See, for example,
Markov Chains by D. Revuz (NorthHolland, Amsterdam, 1984).
We cannot begin to give the general theory corresponding to Example
4.2.8, but we can draw some general conclusions from Theorem 4.2.3 when
the situation is more like Example 4.2.7. Suppose we have a Markov chain
( X n ) n ~ O with absorbing boundary aD. Set
h? = JP>i(T < 00)
(4.6)
{
h8 = Ph
8
in D
h
8
= 1 in aD.
Note that hf = 1 for all i always gives a possible solution. Hence if (4.6) has
a unique bounded solution then hf = JP>i(T < 00) = 1 for all i. Conversely,
if Pi (T < 00) = 1 for all i, then, as we showed in Theorem 4.2.3, (4.6) has
a unique bounded solution. Indeed, we showed more generally that this
condition implies that
where T is the hitting time of aD. Then by the methods of Section 1.3 we
have
{
¢ = P¢ + c in D
¢ = f in aD
has at most one bounded solution, and since
(4.7)
is the minimal solution, any bounded solution is given by (4.7). Suppose
from now on that JP>i(T < 00) = 1 for all i. Let ¢ be a bounded nonnegative
function, harmonic in D, with boundary values ¢i = Ii for i E aD. Then,
by monotone convergence
<Pi = lEi (J(X
T
)) = L hJI»i(XT=j).
jE8D
Hence every bounded harmonic function is determined by its boundary
values and, indeed
<P = L !i
hi
,
jE8D
where
150 4. Further theory
Just as in Example 4.2.7, the hitting probabilities for boundary states form
a set of extremal generators for the set of all bounded nonnegative harmonic
functions.
Exercises
4.2.1 Consider a discretetime Markov chain ( X n ) n ~ O and the potential ¢
with costs (Ci : i E D) and boundary values (Ii: i E aD). Set
ifn ~ T
if n > T,
n<T
where T is the hitting time of aD and a is a new state. Show that ( X n ) n ~ o
is a Markov chain and determine its transition matrix.
Check that
00
<Pi = E
i
L c(X
n
) = E
i
L c(X
n
)
n=O
where T = T + 1 and where we set Ci = Ii on aD and Ca = O. This
shows that a general potential may always be considered as a potential
with boundary value zero or, indeed, without boundary at all.
Can you find a similar reduction for continuoustime chains?
4.2.2 Prove the fact claimed in Example 4.2.8 that
4.2.3 Let (Ci : i E I) be a nonnegative function. Partition I as D U aD and
suppose that the linear equations
{
¢ = P¢ +c in D
¢ = 0 in aD
have a unique bounded solution. Show that the Markov chain ( X n ) n ~ O
with transition matrix P is certain to hit aD.
Consider now a new partition Du aD, where D ~ D. Show that the
linear equations
{
'l/J = P'l/J + c in 15_
1f;=0 in aD
also have a unique bounded solution, and that
11/" < ,h.
0/'1,  0/'1,
for all i E I.
4.3 Electrical networks
4.3 Electrical networks
151
An electrical network has a countable set I of nodes, each node i having a
capacity 1ri > O. Some nodes are joined by wires, the wire between i and
j having conductivity aij = aji 2:: o. Where no wire joins i to j we take
aij = O. In practice, each 'wire' contains a resistor, which determines the
conductivity as the reciprocal of its resistance. Each node i holds a certain
charge Xi, which determines its potential ¢i by
A current or flow of charge is any matrix (T'ij : i, j E I) with T'ij = T'ji.
Physically it is found that the current T'ij from i to j obeys Ohm's law:
Thus charge flows from nodes of high potential to nodes of low potential.
The first problem in electrical networks is to determine equilibrium flows
and potentials, subject to given external conditions. The nodes are parti
tioned into two sets D and aD. External connections are made at the nodes
in aD and possibly at some of the nodes in D. These have the effect that
each node i E aD is held at a given potential Ii, and that a given current 9i
enters the network at each node i ED. The case where gi = 0 corresponds
to a node with no external connection. In equilibrium, current may also
enter or leave the network through aD , but here it is not the current but
the potential which is determined externally.
Given a flow (T'ij : i, j E I) we shall write T'i for the total flow from i to
the network:
'Yi = 2: 'Yij'
JEI
In equilibrium the charge at each node is constant, so
T'i = 9i
for i E D.
Therefore, by Ohm's law, any equilibrium potential ¢ = (¢i : i E I) must
satisfy
{
LjEIaij(¢i  ¢j) = 9i,
¢i = fi'
for i E D
for i E aD.
(4.8)
There is a simple correspondence between electrical networks and reversible
Markov chains in continuoustime, given by
for i =I j.
152 4. Further theory
We shall assume that the total conductivity at each node is finite:
ai = Laij < 00.
j#i
Then ai = 7riqi = 7riqii. The capacities 7ri are the components of an
invariant measure, and the symmetry of aij corresponds to the detailed
balance equations. The equations for an equilibrium potential may now be
written in a form familiar from the preceding section:
{
Q¢ = c in D
¢ = f in aD,
(4.9)
(4.10)
where Ci = 9i/7ri. It is natural that c appears here and not 9, because ct
and f have the same physical dimensions. We know that these equations
may fail to have a unique solution, indicating the interesting possibility
that there may be more than one equilibrium potential. However, to keep
matters simple here, we shall assume that I is finite, that the network is
connected, and that aD is nonempty. This is enough to ensure uniqueness
of potentials. Then, by Theorem 4.2.4, the equilibrium potential is given
by
¢i = E
i
(iT c(Xt)dt + f(X
T
))
where T is the hitting time of aD.
In fact, the case where aD is empty may be dealt with as follows: we
must have
L9i=O
iEI
or there is no possibility of equilibrium; pick one node k, set aD = {k}, and
replace the condition ~ k = 9k by ¢i = O. The new problem is equivalent to
the old, but now aD is nonempty.
A
B
c
2
1
2
1
2
1
2
D
E
F
4.3 Electrical networks 153
Example 4.3.1
Determine the equilibrium current in the network shown on the preceding
page when unit current enters at A and leaves at F. The conductivities are
shown on the diagram. Let us set cPA = 1 and cPF = O. This will result in
some flow from A to F, which we can scale to get a unit flow. By symmetry,
cPE = 1 cPB and cPD = 1 cPe. Then, by Ohm's law, since the total current
leaving Band C must vanish
(cPB  cPA) + (cPB  cPE) + 2(cPB  cPe) = 0,
2(cPe  cPF) + 2(cPe  cPB) = O.
Hence, cPB = 1/2 and cPe = 1/4, and the associated flow is given by =
1/2, = 1/2, = 1/2, = O. In fact, we were lucky  no scaling
was necessary.
Note that the node capacities do not affect the problem we considered.
Let us arbitrarily assign to each node a capacity 1. Then there is an asso
ciated Markov chain and, according to (4.10), the equilibrium potential is
given by
cPi = Ei (lxT=A) = JP>i(XT = A)
where T is the hitting time of {A, F}. Different node capacities result
in different Markov chains, but the same jump chain and hence the same
hitting probabilities.
Here is a general result expressing equilibrium potentials, flows and
charges in terms of the associated Markov chain.
Theorem 4.3.2. Consider a finite network with external connections at
two nodes A and B, and the associated Markov chain
(a) The unique equilibrium potential cP with cPA = 1 and cPB = 0 is given
by
where TA and TB are the hitting times of A and B.
(b) The unique equilibrium flow with = 1 and = 1 is given by
where r
ij
is the number of times that jumps from i to j before
hitting B.
(c) The charge X associated with subject to XB = 0, is given by
154 4. Further theory
Proof. The formula for ¢ is a special case of (4.10), where c = 0 and f =
l{A}. We shall prove (b) and (c) together. Observe that if X
o
= A then
if i = A
if i {A,B}
if i = B
so if = EA(r
ij
 r
ji
) then is a unit flow from A to B. We have
00
roo = '""'1{,," _0,," 0 N}
'1,) LJ B
n=O
where NB is the hitting time of B for the jump chain So, by the
Markov property of the jump chain
00
lEA(rij) = LIPA(Y
n
= i, Y
n
+! = j, n < NB)
n=O
00
= LIPA(Y
n
= i,n < NB)1rij.
n=O
Set
fTB
Xi = lEA 10 l{Xt=i}dt
and consider the associated potential'l/Ji = Xi/1ri. Then
00
Xiqij = XiQi
1r
ij = LIPA(Y
n
= i,n < NB)
1r
ij = lEA(rij)
n=O
so
('l/Ji 'l/Jj)aij = Xiqij  Xjqij =
Hence'l/J = ¢, is the equilibrium unit flow and X the associated charge, as
required. D
The interpretation of potential theory in terms of electrical networks
makes it natural to consider notions of energy. We define for a potential
¢ = (¢i : i E I) and a flow = : i, j E I)
E(<p) = l L (<Pi  <pj)2aij ,
i,jEI
I(r) = l L ,f
j
aijl.
i,jEI
4.3 Electrical networks 155
The 1/2 means that each wire is counted once. When ¢ and ~ are related
by Ohm's law we have
E(¢) = l L (¢i  ¢j)rij = I(r)
i,jEI
and E(¢) is found physically to give the rate of dissipation of energy, as heat,
by the network. Moreover, we shall see that certain equilibrium potentials
and flows determined by Ohm's law minimize these energy functions. This
characteristic of energy minimization can indeed replace Ohm's law as the
fundamental physical principle.
Theorem 4.3.3. The equilibrium potential and flow may be determined
as follows.
(a) The equilibrium potential ¢ = (¢i : i E I) with boundary values
¢i = Ii for i E 8D and no current sources in D is the unique solution
to
minimize E (¢)
subject to ¢i = fi for i E aD.
(b) The equilibrium flow r = (rij : i, j E I) with current sources ri = 9i
for i E D and boundary potential zero is the unique solution to
minimize I (~ )
subject to ~ i = 9i for i E D.
Proof. For any potential ¢ = (¢i : i E I) and any flow ~ = ( ~ i j : i,j E I)
we have
L (¢i  ¢j)rij = 2 L ¢i'Yi.
i,jEI iEI
(a) Denote by ¢ = (¢i : i E I) and by ~ = ( ~ i j : i,j E I) the equilibrium
potential and flow. We have ~ i = 0 for i E D. We can write any potential
in the minimization problem in the form ¢ +c, where c = (ci : i E I) with
Ci = 0 for i E aD. Then
L (Ei  Ej)(¢i  ¢j)aij = L (Ei  Ej)rij = 2 LEi'Yi = 0
i,j EI i,j EI iEI
so
E(¢ +c) = E(¢) +E(c) ~ E(¢)
with equality only if c = o.
156 4. Further theory
(b) Denote by cP = (cPi : i E I) and by ~ = ( ~ i j : i, j E I) the equilibrium
potential and flow. We have cPi = 0 for i E aD. We can write any flow in
the minimization problem in the form ~ +8, where 8 = (8
ij
: i,j E I) is a
flow with 8
i
= 0 for i E D. Then
L 'Yij8ijai/ = L (¢i  ¢j )8ij = 2 L ¢i
8
i = 0,
i,jEI i,jEI iEI
so
I ( ~ +8) = I ( ~ ) +1(8) ~ 1(8)
with equality only if 8 = o. D
The following reformulation of part (a) of the preceding result states that
harmonic functions minimize energy.
Corollary 4.3.4. Suppose that ¢ = (¢i : i E I) satisfies
{
Q¢ = 0 in D
cP = I in aD.
Then cP is the unique solution to
minimize E (¢)
subject to ¢ = f in aD.
An important feature of electrical networks is that networks with a small
number of external connections look like networks with a small number
of nodes altogether. In fact, given any network, there is always another
network of wires joining the externally connected nodes alone, equivalent
in its response to external flows and potentials.
Let J ~ I. We say that a = (aij : i,j E J) is an effective conductivity on
J if, for all potentials I = (Ii : i E J), the external currents into J when J
is held at potential I are the same for (J, a) as for (I, a). We know that I
determines an equilibrium potential ¢ = (cPi : i E I) by
{
L.jEI(¢i  ¢j)aij : 0 for ~ ~ J
cPi  Ii for 'I, E J.
Then a is an effective conductivity if, for all I, for i E J we have
L(¢i  ¢j)aij = :E(Ji  Ji) aij'
JEI jEJ
For a conductivity matrix a on J, for a potential I = (Ii : i E J) and a flow
8 = (8
ij
: i, j E J) we set
 1" 2
E(/) = 2 L.J (Ii  Ij) aij,
i,jEJ
1(8) = ~ L 8fj
u
i?·
i,jEJ
4.3 Electrical networks
Theorem 4.3.5. There is a unique effective conductivity a given by
aij = aij + L aik<P{
where for each j E J, q) = : i E I) is the potential defined by
157
{
 = 0
= 8
ij
for i tf J
for i E J.
(4.11)
Moreover, a is characterized by the Dirichlet variational principle
E(/) = inf E(¢),
<Pi==fi on J
and also by the Thompson variational principle
inf 1(8) = inf
Oi==gi on J {J
gi on
"Yi==
o off J
Proof. Given I = (Ii: i E J), define ¢ = (¢i : i E I) by
<Pi = L 1i<p1
jEJ
then ¢ is the equilibrium potential given by
{
aij(¢i  ¢j) = 0
¢i = Ii
and, by Corollary 4.3.4, ¢ solves
for i fj. J
for i E J,
minimize E(¢)
subject to ¢i = Ii for i E J.
We have, for i E J
Z:=aij<Pj = z:=aijfj + z:=z:=aik<p{fj = z:=aijfjo
JEI jEJ jEJ
In particular, taking I == 1 we obtain
Laij = Laijo
JEI jEJ
158 4. Further theory
Hence we have equality of external currents:
2)cPi  cPj )aij = 'r)1i  Ii )Uij ·
JEI jEJ
Moreover, we also have equality of energies:
L (cPi  cPj)2aij = 2 L cPi L(cPi  cPj)aij
i,jEI iEI JEI
= 2 L Ii LUi !i)Uij = L Ui _!i)2
Uij
.
iEJ jEJ i,jEJ
Finally, if gij = (Ii  Ij )aij and = (¢i  ¢j )aij, then
L "Yljai/ = L (cPi  cPj)2
aij
i,jEI i,jEI
= L (Ii  Ii )2Uij = L gfjUi/,
i,jEJ i,jEJ
so, by Theorem 4.3.3, for any flow 8 = (8ij : i,j E I) with 8i = gi for i E J
and bi = 0 for i ¢. J, we have
L
1 > L 2 1
v··a·· g··a··
 •
i,jEI i,jEJ
o
Effective conductivity is also related to the associated Markov chain
in an interesting way. Define the time spent in J
At = it l{X
s
EJ}ds
and a timechanged process (X by
X
t
= Xr(t)
where
r(t) = inf{s 0 : As > t}.
We obtain by observing whilst in J, and stopping the clock
whilst makes excursions outside J. This is really a transformation
of the jump chain. By applying the strong Markov property to the jump
chain we find that (X is itself a Markov chain, with jump matrix IT
given by
1'fij = 'lrij + L'lrik¢k
for i,j E J,
where
4.4 Brownian motion 159
¢{ = JP>k(X
T
= j)
and T denotes the hitting time of J. See Example 1.4.4. Hence
has Qmatrix given by
Qij = qij + L qik¢{
Since ¢i = (¢{ : k E I) is the unique solution to (4.11), this shows that
so is the Markov chain on J associated with the effective conduc
tivitya.
There is much more that one can say, for example in tying up the non
equilibrium behaviour of Markov chains and electrical networks. More
over, methods coming from one theory one provide insights into the other.
For an entertaining and illuminating account of the subject, you should
see Random Walks and Electrical Networks by P. G. Doyle and J. L. Snell
(earus Mathematical Monographs 22, Mathematical Association of Amer
ica, 1984).
4.4 Brownian motion
Imagine a symmetric ran'dom walk in Euclidean space which takes infinitesi
mal jumps with infinite frequency and you will have some idea of Brownian
motion. It is named after a botanist who observed such a motion when
looking at pollen grains under a microscope. The mathematical object now
called Brownian motion was actually discovered by Wiener, and is also
called the Wiener process.
A discrete approximation to Euclidean space ]Rd is provided by
where c is a large positive number. The simple symmetric random walk
on Zd is a Markov chain which is by now quite familiar. We shall
show that the scaleddown and speededup process
X
(c)  1/2X
t  c ct
is a good approximation to Brownian motion. This provides an elementary
way of thinking about Brownian motion. Also, it makes it reasonable to
160 4. Further theory
suppose that some properties of the random walk carryover to Brownian
motion. At the end of this section we state some results which confirm that
this is true to a remarkable extent.
Why is space rescaled by the squareroot of the timescaling? Well, if
we hope that xi
c
) converges in some sense as c ~ 00 to a nondegenerate
limit, we will at least want IE[lxi
c
) 1
2
] to converge to a nondegenerate limit.
For ct E Z+ we have
so the squareroot scaling gives
which is independent of c.
We begin by defining Brownian motion, and then show that this is not
an empty definition; that is to say, Brownian motions exist.
A realvalued random variable is said to have Gaussian distribution with
mean 0 and variance t if it has density function
cPt(X) = (27T"t)1/2 exp{ _x
2
/2t}.
The fundamental role of Gaussian distributions in probability derives from
the following result.
Theorem 4.4.1 (Central limit theorem). Let Xl, X
2
, ••• be a se
quence of independent and identically distributed realvalued random vari
ables with mean 0 and variance t E (0,00). Then, for all bounded continu
ous functions f, as n ~ 00 we have
We shall take this result and a few other standard properties of the
Gaussian distribution for granted in this section. There are many introduc
tory texts on probability which give the full details.
A realvalued process (Xt)t;:::o is said to be continuous if
lP({w : t..+ Xt(w) is continuous}) = 1.
A continuous realvalued process (Bt)t;:::o is called a Brownian motion if
B
o
= 0 and for all 0 = to < tl < ... < t
n
the increments
4.4 Brownian motion
are independent Gaussian random variables of mean 0 and variance
161
The conditions made on ( B t ) t ~ o are enough to determine all the probabil
ities associated with the process. To put it properly, the law of Brownian
motion, which is a measure on the set of continuous paths, is uniquely de
termined. However, it is not obvious that there is any such process. We
need the following result.
Theorem 4.4.2 (Wiener's theorem). Brownian motion exists.
Proof. For N = 0,1,2, ... , denote by D
N
the set of integer multiples of
2
N
in [0,00), and denote by D the union of these sets. Let us say that
(B
t
: t E D
N
) is a Brownian motion indexed by D
N
if B
o
= 0 and for all
o= to < tl < ... < t
n
in D
N
the increments
are independent Gaussian random variables of mean 0 and variance
We suppose given, for each tED, an independent Gaussian random variable
yt of mean 0 and variance 1. For t E Do = Z+ set
then (B
t
: t E Do) is a Brownian motion indexed by Do. We shall show
how to extend this process successively to Brownian motions (B
t
: tEDN )
indexed by D
N
. Then (B
t
: tED) is a Brownian motion indexed by D.
Next we shall show that (B
t
: tED) extends continuously to t E [0,00),
and finally check that the extension is a Brownian motion.
Suppose we have constructed (B
t
: t E D
N

1
), a Brownian motion in
dexed by D
N

1
• For t E DN\D
N

1
set r = t  2
N
S = t +2
N
so that
s, t E D
N

1
and define
Zt = 2(N+l)/2yt,
Bt = ~ ( B r +Bs) +Zt·
We obtain two new increments:
B
t
 B
r
= ~ (Bs  B
r
) +Zt,
B
s
 B
t
= ~ ( B s  B
r
)  Zt.
162
We compute
4. Further theory
E[(B
t
 B
r
)2] = E[(B
s
 B
t
)2] = +2(N+I) = 2
N
,
E[(B
t
 Br)(B
s
 B
t
)] =  2(N+I) = o.
The two new increments, being Gaussian, are therefore independent and of
the required variance. Moreover, being constructed from B
s
 B
r
and yt,
they are certainly independent of increments over intervals disjoint from
(r, s). Hence (B
t
: tEDN) is a Brownian motion indexed by D N, as
required. Hence, by induction, we obtain a Brownian motion (B
t
: tED).
For each N denote by the continuous process obtained by
linear interpolation from (B
t
: t E D
N
). Also, set zi
N
) = BiN)  BiNI).
For t E D
N

I
we have zi
N
) = O. For t E DN\D
N

I
, by our construction
we have
with yt Gaussian of mean 0 and variance 1. Set
M
N
= sup IziN)I.
tE[O,I]
Then, since interpolates linearly between its values on D
N
, we
obtain
M
N
= sup 2(N+I)/2Iytl.
tE(DN\DNl)n[o,l]
There are 2
N

I
points in (DN\D
N

I
) n [0,1]. So for A > 0 we have
For a random variable X 0 and p > 0 we have the formula
Hence
= 1
00
p,Xp
1
IP(2(N+l)/2M
N
> 'x)d'x
2
N

1
1
00
p,Xp1IP(IY11 > 'x)d'x = 2
N

1
E(IY
1
IP)
4.4 Brownian motion
and hence, for any p > 2
00 00
ELMn= LE(MN )
N=O N=O
163
00 00
:::; L E(MKr)l/P :::; E(IY1IP)1/p L (2(p2)/2
p
)N < 00.
N=O N=O
It follows that, with probability 1, as N 00
BiN) = Bi
O
) +zi
1
) +... + zi
N
)
converges uniformly in t E [0,1], and by a similar argument uniformly for
t in any bounded interval. Now BiN) eventually equals B
t
for any tED
and the uniform limit of continuous functions is continuous. Therefore,
(B
t
: tED) has a continuous extension (Bt)t,?:o, as claimed.
It remains to show that the increments of (Bt)t,?:o have the required joint
distribution. But given 0 < t
1
< ... < t
n
we can find sequences (t''k)mEN
in D such that 0 < t
1
< ... < for all m and tf: tk for all k. Set
to = to = o. We know that the increments
are Gaussian of mean 0 and variance
tr  t(f, · · · 
Hence, using continuity of (Bt)t,?:o we can let m 00 to obtain the desired
distribution for the increments
D
Having shown that Brownian motion exists, we now want to show how
it appears as a universal scaling limit of random walks, very/much as the
Gaussian distribution does for sums of independent random variables.

Theorem 4.4.3. Let (Xn)n,?:O be a discretetime realvalued random walk
with steps of mean 0 and variance 0'2 E (0,00). For c > 0 consider the
rescaled process
X
(c)  1/2X
t  c ct
where the value of X
ct
when ct is not an integer is found by linear interpo
lation. Then, for all m, for all bounded continuous functions f : jRm jR
and all 0 tl < ... < t
m
, we have
tE[f(aB
h
,· .. ,aB
tm
)]
as c 00, where (Bt).t,?:o is a Brownian motion.
164 4. Further theory
Proof The claim is that ... , converges weakly to
(aBtl, ... ,aB
trn
) as c 00. In the proof we shall take for granted some
basic properties of weak convergence. First define
X
(e)  I/2X
t  e [et]
where (et] denotes the integer part of ct. Then
\(x
(e) X(e)) (X(e) X(e))\ I/2\('V: 'V:)\
t 1 ,..., t
rn
 t 1 ,..., t
rn
:::; e .L [et 1 ] +1, · .. ,.L [et n] +1
where Y
n
denotes the nth step of The right side converges weakly
to 0, so it suffices to prove the claim with replacing xi
e
).
Consider now the increments
U
(e)  X(e)  X(e) Z (B B )
k  tk tkl' k = a tk  tkl
for k = 1, ... , m. Since .ide) = B
o
= 0 it suffices to show that
(Ui
e
), . .. converges weakly to (ZI, ... ,Zm). Then since both sets of
increments are independent, it suffices to show that converges weakly
to Zk for each k. But
[etk]
= C
1
/
2
L Y
n
rv (C
1
/
2
Nk(C)1/2)Nk(C)1/2(Yi + ... +YN(c»)
n=[etk_l]+1
where rv denotes identity of distribution and Nk(e) = [etk]  [etkI]. By
the central limit theorem Nk(e)I/2(Y
I
+ ... + YN(e)) converges weakly
to (tk  tk_l)1/2Zk, and (C
1
/
2
Nk(C)1/2) t (tk  tk_l)1/2. Hence
converges weakly to Zk, as required. D
To summarize the last two results, we have shown, using special proper
ties of the Gaussian distribution, that there is a continuous process
with stationary independent increments and such that B
t
is Gaussian of
mean 0 and variance t, for each t o. That was Wiener's theorem. Then,
using the central limit theorem applied to the increments of a rescaled ran
dom walk, we established a sort of convergence to Brownian motion. There
now follows a series of related remarks.
Note the similarity to the definition of a Poisson process as a right
continuous integervalued process starting from 0, having station
ary independent increments and such that X
t
is Poisson of parameter At
for each t O.
Given d independent Brownian motions (Bi . .. let us con
sider the JRdvalued process B
t
= (Bi, ... ,Bt). We call a Brownian
4.4 Brownian motion 165
motion in ]Rd. There is a multidimensional version of the central limit the
orem which leads to a multidimensional version of Theorem 4.4.3, with no
essential change in the proof. Thus if is a random walk in ]Rd with
steps of mean 0 and covariance matrix
and if V is finite, then for all bounded continuous functions f : (]Rd)m ]R,
as c 00 we have
Here are two examples. We might take to be the simple symmetric
random walk in Z3, then V = Alternatively, we might take the compo
nents of to be three independent simple symmetric random walks
in Z, in which case V = I. Although these are different random walks, once
the difference in variance is taken out, the result shows that in the scaling
limit they behave asymptotically the same. More generally, given a random
walk with a complicated step distribution, it is useful to know that on large
scales all one needs to calculate is the variance (or covariance matrix). All
other aspects of the step distribution become irrelevant as c 00.
The scaling used in Theorem 4.4.3 suggests the following scaling invari
ance property of Brownian motion (Bt)t>o, which is also easy to check from
the definition. For any c > 0 the proces; defined by
B
(e)  1/2B
t  C et
is a Brownian motion. Thus Brownian motion appears as a fixed point of the
scaling transformation, which attracts all other finite variance symmetric
random walks as c 00.
The sense in which we have shown that converges to Brownian
motion is very weak, and one can with effort prove stronger forms of con
vergence. However, what we have proved is strong enough to ensure that
does not converge, in the same sense, to anything else.
The discussion to this point has not really been about the Markov prop
erty, but rather about processes with independent increments. To remedy
this we must first define Brownian motion starting from x: this is simply
any process such that B
o
= x and (B
t
 is a Brownian mo
tion (starting from 0). As a limit of Markov chains it is natural to look in
Brownian motion for the structure of a Markov process. By analogy with
continuoustime Markov chains we look for a transition semigroup
166 4. Further theory
and a generator G. For any bounded measurable function f : ]Rd ]R we
have
lEx[J(B
t
)] = lEo[J(x + B
t
)] = { f(x + Y)¢t(Yd··· ¢t(Yd)dYl .. , dYd
JRd
= ( p(t, x, y)f(y)dy
JRd
where
p(t, x, y) = (27rt)d/2 exp{Iy  xI
2
/2t}.
This is the transition density for Brownian motion and the transition semi
group is given by
(Ptf)(x) = ( p(t, x, y)f(y)dy = lEx [f(B
t
)].
JRd
To check the semigroup property PsP
t
= P
s
+
t
we note that
Ex[f(B
s
+
t
)] = Ex [f(B
s
+ (B
s
+
t
 B
s
))]
= Ex [Ptf(B
s
)] = (PsPtf)(x)
where we first took the expectation over the independent increment
B
s
+
t
 B
s
. For t > 0 it is easy to check that
:t p(t, x, y) = x, y)
where
a
2
a
2
= a 2 + · · · + a 2·
Xl x
d
Hence, if f has two bounded derivatives, we have
a
8
(Ptf)(x) = { x, y)f(y)dy
t JRd
= { x, y)f(y)dy
JRd
= ( p(t, x, y)(
JRd
=
as t ! O. This suggests, by analogy with continuoustime chains, that the
generator, a term we have not defined precisely, should be given by
4.4 Brownian motion 167
Where formerly we considered vectors (fi : i E I), now there are functions
f : jRd jR, required to have various degrees of local regularity, such as
measurability and differentiability. Where formerly we considered matrices
P
t
and Q, now we have linear operators on functions: P
t
is an integral
operator, G is a differential operator.
We would like to explain the appearance of the Laplacian by refer
ence to the random walk approximation. Denote by the simple
symmetric random walk in tl
d
and consider for N == 1,2, ... the rescaled
process
xi
N
) == N
I
/
2
X
Nt
, t == 0, liN, 2/N, . ...
For a bounded continuous function f : jRd jR, set
(pt(N) f)(x) == IEx[f(Xi
N
))], X E N
I
/
2
71
d
•
The closest thing we have to a derivative in t at 0 for (pt(N))t=O,I/N,2/N, ...
is
(
(N) )) _ [ ( (N) ) ((N))]
N PI/Nf  f (x  NIE
x
f X
I
/
N
 f X
o
== NIE
N
l/2
x
[f(N
I
/
2
Xl)  f(N
I
/
2
X
O
)]
= (N/2){f(x  N
I
/
2
)  2f(x) + f(x +N
I
/
2
)}.
If we assume that f has two bounded derivatives then, by Taylor's theorem,
as N 00,
f(x  N
I
/
2
)  2f(x) + f(x +N
I
/
2
) == +o(N)),
so
N(pi;Jf  f)(x) t
We finish by stating some results about Brownian motion which empha
sise how much of the structure of Markov chains carries over. You will notice
some weasel words creeping in, such as measurable, continuous and differ
entiable. These are various sorts of local regularity for functions defined on
the statespace jRd. They did not appear for Markov chains because a dis
crete statespace has no local structure. You might correctly guess that the
proofs would require additional real analysis, relative to the corresponding
results for chains, and a proper measuretheoretic basis for the probabil
ity. But, this aside, the main ideas are very similar. For further details
see, for example, Probability Theory  an analytic view by D. W. Stroock
(Cambridge University Press, 1993), or Diffusions, Markov Processes and
Martingales, Volume 1: Foundations by L. C. G. Roger.s and David Williams
(Wiley, Chichester, 2nd edition 1994).
First, here is a result on recurrence and transience.
168 4. Further theory
Theorem 4.4.4. Let ( B t ) t ~ o be a Brownian motion in jRd.
(i) If d = 1, then
JP>({t ~ 0: B
t
= O} is unbounded) = 1.
(ii) If d = 2, then
JP>(B
t
= 0 for some t > 0) = 0
but, for any € > 0
JP>( {t ~ 0 : IBt I < €} is unbounded) = 1.
(iii) If d = 3, then for any N < 00
JP>(I B
t
I ~ 00 as t ~ 00) = 1.
It is natural to compare this result with the facts proved in Section 1.6,
that in Z and Z2 the simple symmetric random walk is recurrent, whereas
in Z3 it is transient. The results correspond exactly in dimensions one and
three. In dimension two we see the fact that for continuous statespace
it makes a difference to demand returns to a point or to arbitrarily small
neighbourhoods of a point. If we accept this latter notion of recurrence the
correspondence extends to dimension two.
The invariant measure for Brownian motion is Lebesgue measure dx.
This has infinite total mass so in dimensions one and two Brownian motion
is only null recurrent. So that we can state some results for the positive
recurrent case, we shall consider Brownian motion in jRd projected onto the
torus T
d
= jRd/Zd. In dimension one this just means wrapping the line
round a circle of circumference 1. The invariant measure remains Lebesgue
measure but this now has total mass 1. So the projected process is positive
recurrent and we can expect convergence to equilibrium and ergodic results
corresponding to Theorems 1.8.3 and 1.10.2.
Theorem 4.4.5. Let (Bt)t>o be a Brownian motion in jRd and let
f : jRd ~ jR be a continuous periodic function, so that
f(x +z) = f(x) for all z E Zd.
Then for all x E jRd, as t ~ 00, we have
lEx[f(B
t
)] t 1= [ f(z)dz
J[O,l]d
and, moreover
JI»x ( ~ it f(Bs)ds t 1ast t 00) = 1.
The generator ! ~ of Brownian motion in jRd reappears as it should in
the following martingale characterization of Brownian motion.
4.4 Brownian motion 169
Theorem 4.4.6. Let be a continuous ]Rdvalued random pro
cess. Write (F
t
)t2::o for the filtration of Then the following are
equivalent:
(i) (X
t
)t2::o is a Brownian motion;
(ii) for all bounded functions f which are twice differentiable with
bounded second derivative, the following process is a martingale:
This result obviously corresponds to Theorem 4.1.2. In case you are unsure,
a continuous time process (M
t
)t2::o is a martingale if it is adapted to the
given filtration (F
t
)t2::o, if JEIMtl < 00 for all t, and
whenever s :::; t and A E F
8
•
We end with a result on the potentials associated with Brownian motion,
corresponding very closely to Theorem 4.2.3 for Markov chains. These
potentials are identical to those appearing in Newton's theory of gravity,
as we remarked in Section 4.2.
Theorem 4.4.7. Let D be an open set in ]Rd with smooth boundary aD.
Let c : D [0, 00) be measurable and let f : aD [0,00) be continuous.
Set
<jJ(x) = JEx [I
T
c(Bt)dt + f(XT )IT<oo]
where T is the hitting time of aD. Then
(i) ¢ if finite belongs to C
2
(D) n C(D) and satisfies
in D
in aD;
(4.12)
(ii) if 1/J E C
2
(D) n C(D) and satisfies
{
c in D
1/J f in aD
and 'l/J 0, then 'l/J ¢;
(iii) if ¢(x) = 00 for some x, then (4.12) has no finite solution;
(iv) if JPx(T < 00) = 1 for all x, then (4.12) has at most one bounded
solution in C
2
(D) n C(D).
5
Applications
Applications of Markov chains arise in many different areas. Some have
already appeared to illustrate the theory, from games of chance to the
evolution of populations, from calculating the fair price for a random reward
to calculating the probability that an absentminded professor is caught
without an umbrella. In a realworld problem involving random processes
you should always look for Markov chains. They are often easy to spot.
Once a Markov chain is identified, there is a qualitative theory which limits
the sorts of behaviour that can occur  we know, for example, that every
state is either recurrent or transient. There are also good computational
methods  for hitting probabilities and expected rewards, and for longrun
behaviour via invariant distributions.
In this chapter we shall look at five areas of application in detail: bi
ological models, queueing models, resource management models, Markov
decision processes and Markov chain Monte Carlo. In each case our aim is
to provide an introduction rather than a systematic account or survey of
the field. References to books for further reading are given in each section.
5.1 Markov chains in biology
Randomness is often an appropriate model for systems of high complex
ity, such as are often found in biology. We have already illustrated some
aspects of the theory by simple models with a biological interpretation.
See Example 1.1.5 (virus), Exercise 1.1.6 (octopus), Example 1.3.4 (birth
anddeath chain) and Exercise 2.5.1 (bacteria). We are now going to give
5.1 Markov chains in biology 171
some more examples where Markov chains have been used to model bio
logical processes, in the study of population growth, epidemics and genetic
inheritance. It should be recognised from the start that these models are
simplified and somewhat stylized in order to make them mathematically
tractable. Nevertheless, by providing quantitative understanding of various
phenomena they can provide a useful contribution to science.
Example 5.1.1 (Branching processes)
The original branching process was considered by Galton and Watson in the
1870s while seeking a quantitative explanation for the phenomenon of the
disappearance of family names, even in a growing population. Under the
assumption that each male in a given family had a probability Pk of having
k sons, they wished to determine the probability that after n generations
an individual had no male descendents. The solution to this problem is
explained below.
The basic branching process model has many applications to problems
of population growth, and also to the study of chain reactions in chemistry
and nuclear fission. Suppose at time n = 0 there is one individual, who dies
and is replaced at time n = 1 by a random number of offspring N. Suppose,
next, that these offspring also die and are themselves replaced at time n = 2,
each independently, by a random number of further offspring, having the
same distribution as N, and so on. We can construct the process by taking
for each n E N a sequence of independent random variables (N;:)kEN, each
with the same distribution as N, by setting X
o
= 1 and defining inductively,
for n 2: 1
X
n
= Nf + ... + N
Xn
_
1
•
Then X
n
gives the size of the population in the nth generation. The process
( X n ) n ~ O is a Markov chain on I = {O, 1,2, ... } with absorbing state o. The
case where P(N = 1) = 1 is trivial so we exclude it. We have
P(X
n
= 0 I X
n

1
= i) = P(N = O)i
so if P(N = 0) > 0 then i leads to 0, and every state i 2: 1 is transient. If
P(N = 0) = 0 then P(N 2: 2) > 0, so for i 2: 1, i leads to j for some j > i,
and j does not lead to i, hence i is transient in any case. We deduce that
with probability 1 either X
n
= 0 for some n or X
n
~ 00 as n ~ 00.
Further information on ( X n ) n ~ O is obtained by exploiting "the branching
structure. Consider the probability generating function
00
</J(t) = E(t
N
) = :EtkJP(N = k),
k==O
172 5. Applications
defined for 0 ~ t ~ 1. Conditional on X
n

l
= k we have
Xn =Nf+···+Nr
so
and so
00
lE(t
Xn
) = 'LlE(t
Xn
I X
n

1
= k)JP>(X
n

1
= k) = lE(</>(t)x
n

1
).
k==O
Hence, by induction, we find that E(t
Xn
) = ¢(n)(t), where ¢(n) is the nfold
composition ¢ 0 ... 0 ¢. In principle, this gives the entire distribution of
X
n
, though ¢(n) may be a rather complicated function. Some quantities
are easily deduced: we have
E(X
n
) = lim dd lE(t
Xn
) = lim dd </>(n)(t) = (lim</>'(t)r = p,n,
til t til t til
where Jl = E(N); also
so, since 0 is absorbing, we have
q = P(X
n
= 0 for some n) = lim ¢(n)(o).
n+oo
Now ¢(t) is a convex function with ¢(1) = 1. Let us set r = inf{t E [0,1] :
¢(t) = t}, then ¢(r) = r by continuity. Since ¢ is increasing and 0 ~ r, we
have ¢(O) ~ r and, by induction, ¢(n)(o) ~ r for all n, hence q ~ r. On the
other hand
q = lim ¢(n+l) (0) = lim ¢(¢(n) (0)) = ¢(q)
n + 00 n + 00
so also q 2:: r. Hence q = r. If ¢'(1) > 1 then we must have q < 1, and if
¢'(1) ~ 1 then since either ¢" = 0 or ¢" > 0 everywhere in [0,1) we must
have q = 1. We have shown that the population survives with positive
probability if and only if Jl > 1, where Jl is the mean of the offspring
distribution.
There is a nice connection between branching processes and random
walks. Suppose that in each generation we replace individuals by their off
spring one at a time, so if X
n
= k then it takes k steps to obtain X
n
+
l
.
5.1 Markov chains in biology 173
The population size then performs a random walk ( Y m ) m ~ O with step dis
tribution N  1. Define stopping times To = 0 and, for n 2: 0
Observe that X
n
= YT
n
for all n, and since ( Y m ) m ~ O jumps down by at
most 1 each time, (Xn)n>O hits 0 if and only if (Ym)m>O hits O. Moreover
 
we can use the strong Markov property and a variation of the argument of
Example 1.4.3 to see that, if
qi = P(Y
m
= 0 for some m I Yo = i)
then qi = qf for all i and so
00
ql = P(N = 0) + L qfP(N = i) = ¢(ql)'
k==l
Now each nonnegative solution of this equation provides a nonnegative
solution of the hitting probability equations, so we deduce that ql is the
smallest nonnegative root of the equation q = ¢(q), in agreement with the
generating function approach.
The classic work in this area is The Theory of Branching Processes by
T. E. Harris (Dover, New York, 1989).
Example 5.1.2 (Epidemics)
Many infectious diseases persist at a low intensity in a population for long
periods. Occasionally a large number of cases arise together and form an
epidemic. This behaviour is to some extent explained by the observation
that the presence of a large number of infected individuals increases the
risk to the rest of the population. The decline of an epidemic can also be
explained by the eventual decline in the number of individuals susceptible
to infection, as infectives either die or recover and are then resistant to fur
ther infection. However, these naive explanations leave unanswered many
quantitative questions that are important in predicting the behaviour of
epidemics.
In an idealized population we might suppose that all pairs of individu
als make contact randomly and independently at a common rate, whether
infected or not. For an idealized disease we might suppose t h ~ t on contact
with an infective, individuals themselves become infective and remain so for
an exponential random time, after which they either die or recover. These
two possibilities have identical consequences for the progress of the epi
demic. This idealized model is obviously unrealistic, but it is the simplest
mathematical model to incorporate the basic features of an epidemic.
174 5. Applications
We denote the number of susceptibles by St and the number of infectives
by It. In the idealized model, X
t
= (St, It) performs a Markov chain on
(Z+)2 with transition rates
q(s,i)(si,i+l) = Asi, q(s,i)(s,il) = /li
for some .x, /l E (0, 00). Since St + It does not increase, we effectively
have a finite statespace. The states (s,O) for s E Z+ are all absorbing
and all the other states are transient; indeed all the communicating classes
are singletons. The epidemic must therefore eventually die out, and the
absorption probabilities give the distribution of the number of susceptibles
who escape infection. We can calculate these probabilities explicitly when
So +1
0
is small.
Of greater concern is the behaviour of an epidemic in a large population,
of size N, say. Let us consider the proportions sf = St/N and if' = It/N
and suppose that .x = v/N, where v is independent of N. Consider now a
sequence of models as N ~ 00 and choose s ~ ~ So and i ~ ~ i
o
. It can
be shown that as N ~ 00 the process (sf', if') converges to the solution
(St, it) of the differential equations
(d/dt)st = vstit
(d/dt)i
t
= vStit  /lit
starting from (so, i
o
). Here convergence means that E[I( sf', if')  (St, it) I] ~
ofor all t 2:: o. We will not prove this result, but will give an example of
another easier asymptotic calculation.
Consider the case where So = N 1,1
0
= 1, .x = l/N and /l = O. This has
the following interpretation: a rumour is begun by a single individual who
tells it to everyone she meets; they in turn pass the rumour on to everyone
they meet. We assume that each individual meets another randomly at
the jump times of a Poisson process of rate 1. How long does it take until
everyone knows the rumour? If i people know the rumour, then N  i do
not, and the rate at which the rumour is passed on is
qi = i(N  i)/N.
The expected time until everyone knows the rumour is then
Nl 1 Nl N Nl(l 1) Nl
1
'" qi = '" . . = '" = + . = 2 '" = rv 210gN
~ ~ ~ ( N  ~ ) ~ ~ N  ~ ~ ~
i=l i=l i=l i=l
as N ~ 00. This is not a limit as above but, rather, an asymptotic equiva
lence. The fact that the expected time grows with N is related to the fact
5.1 Markov chains in biology 175
that we do not scale /0 with N: when the rumour is known by very few or
by almost all, the proportion of 'infectives' changes very slowly.
The final two examples come from population genetics. They represent
an attempt to understand quantitatively the consequences of randomness
in genetic inheritance. The randomness here might derive from the choice
of reproducing individual, in sexual reproduction the choice of partner, or
the choice of parents' alleles retained by their offspring. (The word gene
refers to a particular chromosomal locus; the varieties of genetic material
that can be present at such a locus are known as alleles.) This sort of
study was motivated in the first place by a desire to find mathematical
models of natural selection, and thereby to discriminate between various
competing accounts of the process of evolution. More recently, as scientists
have gained access to the genetic material itself, many more questions of a
statistical nature have arisen. We emphasise that we present only the very
simplest examples in a rich theory, for which we refer the interested reader
to Mathematical Population Genetics by W. J. Ewens (Springer, Berlin,
1979).
Example 5.1.3 (WrightFisher model)
This is the discretetime Markov chain on {O, 1, ... ,m} with transition
probabilities
In each generation there are m alleles, some of type A and some of type a.
The types of alleles in generation n+1 are found by choosing randomly (with
replacement) from the types in generation n. If X
n
denotes the number of
alleles of type A in generation n, then ( X n ) n ~ O is a Markov chain with the
above transition probabilities.
This can be viewed as a model of inheritance for a particular gene with
two alleles A and a. We suppose that each individual has two genes, so the
possibilities are AA, Aa and aa. Let us take m to be even with m = 2k.
Suppose that individuals in the next generation are obtained by mating
randomly chosen individuals from the current generation and that offspring
inherit one allele from each parent. We have to allow that both parents
may be the same, and in particular make no requirement that parents be
of opposite sexes. Then if the generation n is, for example.
AA aA AA AA aa,
then each gene in generation n + 1 is A with probability 7/10 and a with
probability 3/10, all independent. We might, for example, get
aa aA Aa AA AA.
176 5. Applications
The structure of pairs of genes is irrelevant to the Markov chain ( X n ) n ~ O ,
which simply counts the number of alleles of type A.
The communicating classes of ( X n ) n ~ O are {O}, {I, ... ,m  I}, {m}.
States 0 and m are absorbing and {I, ... ,m  1} is transient. The hit
ting probabilities for state m (pure AA) are given by
This is obvious when one notes that ( X n ) n ~ O is a martingale; alternatively
one can check that
m
hi = LPijh
j
.
j==o
According to this model, genetic diversity eventually disappears. It is
known, however, that, for p E (0,1), as m ~ 00
lEpm(T) ~ 2m{(1  p) log(l  p) +plogp}
where T is the hitting time of {O, m}, so in a large population diversity does
not disappear quickly.
Some modifications are possible which model other aspects of genetic
theory. Firstly, it may be that the three genetic types AA, Aa, aa have a
relative selective advantage given by lX, (3, ~ > 0 respectively. This means
that the probability of choosing allele A when X
n
= i is given by
'l/J.  a(ijm)2 + (lj2)(3i(m  i)jm
2
~  a(ijm)2 + (3i(m  i)jm
2
+ ,((m  i)jm)2
and the transition probabilities are
Secondly, we may allow genes to mutate. Suppose A mutates to a with
probability u and a mutates to A with probability v. Then the probability
of choosing A when X
n
= i is given by
¢i = {i(l  u) + (m  i)v}/m
and
5.1 Markov chains in biology 177
With u, v > 0, the states 0 and m are no longer absorbing, in fact the chain
is irreducible, so attention shifts from hitting probabilities to the invariant
distribution 7r. There is an exact calculation for the mean of 7r: we have
m m
p, = L i7ri = E
7r
(XI) = L 7riEi(X1)
i=O i=O
m m
= 2: m7ri<Pi = 2:{i(l  u) + (m  i)V}7ri = (1  u)p, + mv  vp,
i=O i=O
so that
JL = mv/(u +v).
Example 5.1.4 (Moran model)
The Moran model is the birthanddeath chain on {O, 1, ... ,m} with tran
sition probabilities
Pi,il = i(m  i)/m
2
, Pii = (i
2
+ (m  i)2)/m
2
, Pi,i+l = i(m  i)/m
2
.
Here is the genetic interpretation: a population consists of individuals of
two types, a and A; we choose randomly one individual from the population
at time n, and add a new individual of the same type; then we choose, again
randomly, one individual from the population at time n and remove it; so
we obtain the population at time n +1. The same individual may be chosen
each time, both to give birth and to die, in which case there is no change
in the makeup of the population. Now, if X
n
denotes the number of type
A individuals in the population at time n, then ( X n ) n ~ O is a Markov chain
with transition matrix P.
There are some obvious differences from the WrightFisher model: firstly,
the Moran model cannot be interpreted in terms of a species where genes
come in pairs, or where individuals have more than one parent; secondly in
the Moran model we only change one individual at a time, not the whole
population. However, the basic Markov chain structure is the same, with
communicating classes {O}, {I, . . . ,m  I}, {m}, absorbing states 0 and m
and transient class {I, ... ,m  I}. The Moran model is reversible, and,
like the WrightFisher model, is a martingale. The hitting probabilities are
given by
IPi(X
n
= m for some n) = i/m.
We can also calculate explicitly the mean time to absorption
178 5. Applications
where T is the hitting time of {O, m}. The simplest method is first to fix j
and write down equations for the mean time kf spent in j, starting from i,
before absorption:
kf = bij + (Pi,i1
k
t1 +Piikt +Pi,i+1
k
t+1) for i = 1, ... ,m  1
kg = kin = 0.
Then, for i = 1, ... ,m  1
so that
. {(i/j)k;
kf = ((mi)/(mj))k;
where k; is determined by
for i ~ j
for i ~ j
(
m
j
l jl). m
2
 2+ k ~ =
mj j j j(mj)
which gives k; = m. Hence
m1 { i ( . ) m1. }
ki=Lkf=m L m=z. + L ~ ·
j=1 j=1 m J j=i+1 J
As in the WrightFisher model, one is really interested in the case where
m is large, and i = pm for some P E (0,1). Then
mp 1 m1 1
m
2
k
pm
= (1 p) L m _ . + P L ~ t (1 p) log(1 p)  plogp
j==1 J j==mp+1 J
as m ~ 00. So, as m ~ 00
Epm(T) ~ m
2
{(1  p) 10g(1  p) +plogp}.
For the WrightFisher model we claimed that
Epm(T) ~ 2m{(1  p) 10g(1  p) +plogp}
which has the same functional form in p and differs by a factor of m/2.
This factor is partially explained by the fact that the Moran model deals
5.2 Queues and queueing networks 179
with one individual at a time, whereas the WrightFisher model changes
all m at once.
Exercises
5.1.1 Consider a branching process with immigration. This is defined, in
the notation of Example 5.1.1, by
X
n
= N
1
n
+ ... + N
x
n
+ In
nl
where (In)n'?:o is a sequence of independent Z+valued random variables
with common generating function 'ljJ(t) = E(t
1n
). Show that, if X
o
= 1,
then
nl
lE(t
X
n) = ¢(n) ( t) II 1/J (¢(k) ( t) ) .
k==O
In the case where the number of immigrants in each generation is Poisson
of parameter A, and where P(N = 0) = 1  P and P(N = 1) = p, find the
longrun proportion of time during which the population is zero.
5.1.2 A species of plant comes in three genotypes AA, Aa and aa. A
single plant of genotype Aa is crossed with itself, so that the offspring has
genotype AA, Aa or aa with probabilities 1/4, 1/2 and 1/4. How long on
average does it take to achieve a pure strain, that is, AA or aa? Suppose it
is desired to breed an AA plant. What should you do? How many crosses
would your procedure require, on average?
5.1.3 In the Moran model we may introduce a selective bias by making
it twice as likely that a type a individual is chosen to die, as compared
to a type A individual. Thus in a population of size m containing i type
A individuals, the probability that some type A is chosen to die is now
i/(i + 2(m  i)). Suppose we begin with just one type A. What is the
probability that eventually the whole population is of type A?
5.2 Queues and queueing networks
Queues form in many circumstances and it is important to be able to pre
dict their behaviour. The basic mathematical model for queues runs as
follows: there is a succession of customers wanting service; on arrival each
customer must wait until a server is free, giving priority to earlier arrivals;
it is assumed that the times between arrivals are independent random vari
ables of the same distribution, and the times taken to serve customers are
also independent random variables, of some other distribution. The main
180 5. Applications
quantity of interest is the random process ( X t ) t ~ O recording the number
of customers in the queue at time t. This is always taken to include both
those being served and those waiting to be served.
In cases where interarrival times and service times have exponential
distributions, ( X t ) t ~ O turns out to be a continuoustime Markov chain, so
we can answer many questions about the queue. This is the context of
our first six examples. Some further variations on queues of this type have
already appeared in Exercises 3.4.1, 3.6.3, 3.7.1 and 3.7.2.
If the interarrival times only are exponential, an analysis is still pos
sible, by exploiting the memorylessness of the Poisson process of arrivals,
and a certain discretetime Markov chain embedded in the queue. This is
explained in the final two examples.
In each example we shall aim to describe some salient features of the
queue in terms of the given data of arrivaltime and servicetime distribu
tions. We shall find conditions for the stability of the queue, and in the
stable case find means to compute the equilibrium distribution of queue
length. We shall also look at the random times that customers spend wait
ing and the length of time that servers are continuously busy.
Example 5.2.1 (M/M/I queue)
This is the simplest queue of all. The code means: memoryless interarrival
times/memoryless service times/one server. Let us suppose that the inter
arrival times are exponential of parameter A, and the service times are
exponential of parameter Jl. Then the number of customers in the queue
( X t ) t ~ O evolves as a Markov chain with the following diagram:
Jl A Jl A
III( •• 111( ••
i i + 1
To see this, suppose at time 0 there are i customers in the queue, where
i > O. Denote by T the time taken to serve the first customer and by A
the time of the next arrival. Then the first jump time J
1
is A 1\ T, which is
exponential of parameter A + JL, and XJl = iI if T < A, XJl = i + 1 if
T > A, which events are independent of J
1
, with probabilities Jl/(A+Jl) and
A/(A+Jl) respectively. If we condition on J
1
= T, then AJ
1
is exponential
of parameter Aand independent of J1: the time already spent waiting for an
arrival is forgotten. Similarly, conditional on J
1
= A, T  J
1
is exponential
of parameter JL and independent of J
1
. The case where i = 0 is simpler
as there is no serving going on. Hence, conditional on XJl = j, ( X t ) t ~ O
5.2 Queues and queueing networks 181
begins afresh from j at time Jl. It follows that ( X t ) t ; ~ o is the claimed
Markov chain. This sort of argument should by now be very familiar and
we shall not spell out the details like this in later examples.
The MIMll queue thus evolves like a random walk, except that it does
not take jumps below o. We deduce that if ,\ > Jl then ( X t ) t ~ O is transient,
that is X
t
~ 00 as t ~ 00. Thus if A > Jl the queue grows without limit
in the long term. When A < Jl, ( X t ) t ~ O is positive recurrent with invariant
distribution
So when A < Jl the average number of customers in the queue in equilibrium
is given by
00 00
lE
7r
{X
t
) = LJP>7r{X
t
;::: i) = L{,\/JL)i = ,\/{JL  ,\).
i=l i=l
Also, the mean time to return to 0 is given by
so the mean length of time that the server is continuously busy is given by
rnO  (llqo) = 1/(Jl  A).
Another quantity of interest is the mean waiting time for a typical customer,
when A < Jl and the queue is in equilibrium. Conditional on finding a queue
of length i on arrival, this is (i + 1)I Jl, so the overall mean waiting time is
A rough check is available here as we can calculate in two ways the expected
total time spent in the queue over an interval of length t: either we multiply
the average queue length by t, or we multiply the mean waiting time by the
expected number of customers At. Either way we get AtlJl  A. The first
calculation is exact but we have not fully justified the sec9nd.
Thus, once the queue size is identified as a Markov chain, its behaviour
is largely understood. Even in more complicated examples where exact
calculation is limited, once the Markovian character of the queue is noted
we know what sort of features to look for  transience and recurrence,
convergence to equilibrium, longrun averages, and so on.
182 5. Applications
Example 5.2.2 (M/M/s queue)
This is a variation on the last example where there is one queue but there
are S servers. Let us assume that the arrival rate is A and the service
rate by each server is Jl. Then if i servers are occupied, the first service is
completed at the minimum of i independent exponential times of parameter
Jl. The first service time is therefore exponential of parameter iJl. The total
service rate increases to a maximum SJl when all servers are working. We
emphasise that the queue size includes those customers who are currently
being served. By an argument similar to the preceding example, the queue
size performs a Markov chain with the following diagram:
A Jl A 2Jl A
•• ••• •••
012
SJl A SJl A
• •• •••
s s+l
So this time we obtain a birthanddeath chain. It is transient in the
case A > SJl and otherwise recurrent. To find an invariant measure we look
at the detailed balance equations
Hence
for i = 0,1, ... ,S
for i = s + 1, S + 2, ....
The queue is therefore positive recurrent when A < SJl. There are two cases
when the invariant distribution has a particularly nice form: when S = 1
we are back to Example 5.2.1 and the invariant distribution is geometric of
parameter AIJl:
When S = 00 we normalize 1r by taking 1ro = e)../JL so that
and the invariant distribution is Poisson of parameter AIJl.
The number of arrivals by time t is a Poisson process of rate A. Each
arrival corresponds to an increase in X
t
, and each departure to a decrease.
Let us suppose that A < SJl, so there is an invariant distribution, and
consider the queue in equilibrium. The detailed balance equations hold and
is nonexplosive, so by Theorem 3.7.3 for any T > 0,
5.2 Queues and queueing networks 183
and ( X T  t ) O ~ t ~ T have the same law. It follows that, in equilibrium, the
number of departures by time t is also a Poisson process of rate A. This is
slightly counterintuitive, as one might imagine that the departure process
runs in fits and starts depending on the number of servers working. Instead,
it turns out that the process of departures, in equilibrium, is just as regular
as the process of arrivals.
Example 5.2.3 (Telephone exchange)
A variation on the M/M/s queue is to turn away customers who cannot
be served immediately. This might serve as a simple model for a telephone
exchange, where the maximum number of calls that can be connected at
once is s: when the exchange is full, additional calls are lost. The maximum
queue size or buffer size is s and we get the following modified Markov chain
diagram:
A JL A 2JL A
• I( • I( •
012
sl s
We can find the invariant distribution of this finite Markov chain by solving
the detailed balance equations, as in the last example. This time we get a
truncated Poisson distribution
By the ergodic theorem, the longrun proportion of time that the exchange
is full, and hence the longrun proportion of calls that are lost, is given by
This is known as Erlang's formula. Compare this example with the bus
maintenance problem in Exercise 3.7.1.
Example 5.2.4 (Queues in series)
Suppose that customers have two service requirements: they arrive as a
Poisson process of rate A to be seen first by server A, and then by server
184 5. Applications
B. For simplicity we shall assume that the service times are independent
exponentials of parameters Q and j3 respectively. What is the average queue
length at B?
Let us denote the queue length at A by (Xt)t>o and that by B by (¥t)t>o.
 
Then is simply an MIMll queue. If A > Q, then is transient
so there is eventually always a queue at A and departures form a Poisson
process of rate a. If A < a, then, by the reversibility argument of Example
5.2.2, the process of departures from A is Poisson of rate A, provided queue
A is in equilibrium. The question about queue length at B is not precisely
formulated: it does not specify that the queues should be in equilibrium;
indeed if A Q there is no equilibrium. Nevertheless, we hope you will agree
to treat arrivals at B as a Poisson process of rate Q /\ A. Then, by Example
5.2.1, the average queue length at B when Q /\ A < j3, in equilibrium, is
given by (Q /\ A) 1((3  (Q /\ A)). If, on the other hand, Q /\ A > (3, then
is transient so the queue at B grows without limit.
There is an equilibrium for both queues if A < Q and A < j3. The
fact that in equilibrium the output from A is Poisson greatly simplifies the
analysis of the two queues in series. For example, the average time taken
by one customer to obtain both services is given by
1/(a  A) + 1/(j3  A).
Example 5.2.5 (Closed migration process)
Consider, first, a single particle in a finite statespace I which performs
a Markov chain with irreducible Qmatrix Q. We know there is a unique
invariant distribution Jr. We may think of the holding times of this chain
as service times, by a single server at each node i E I.
Let us suppose now that there are N particles in the statespace, which
move as before except that they must queue for service at every node. If
we do not care to distinguish between the particles, we can regard this as
a new process with statespace Y= N
1
, where X
t
= (ni : i E I) if
at time t there are ni particles at state i. !n fact, this new process is
Markov chain. To describe its Qmatrix Q we define a function bi : I + I
by
Thus bi adds a particle at i. Then for i =I j the nonzero transition rates
are given by
n E I, i,j E I.
5.2 Queues and queueing networks 185
Observe that we can write the invariant measure equation 1rQ = 0 in the
form
1ri L qij = L 1rjqji·
j#i j#i
For n = (ni : i E I) we set
1f(n) = II1rfi.
iEI
Then
(5.1)
1f(8
i
n) L q(8
i
n, 8
j
n) = II 1 r ~ k
j#i kEI
(
1r
i
Lqji)
j#i
(
L:: 1rjqji)
j#i
= L 1f(8
j
n)q(8
j
n, 8m).
j#i
Given mEl we can put m = 8
i
n in the last identity whenever mi ~ 1. On
summing the resulting equations we obtain
1f(m) L q(m, n) = L 1f(n)q(n, m)
n#m n#m
so 7r is an invariant measure for Q. The total number of particles is con
served so Qhas communicating classes
and the unique invariant distribution for the Nparticle system is given by
normalizing 7f restricted to eN.
Example 5.2.6 (Open migration process)
We consider a modification of the last example where new customers, or
particles, arrive at each node i E I at rate Ai. We suppose also that
customers receiving service at node i leave the network at rate /li. Thus
customers enter the network, move from queue to queue according to a
Markov chain and eventually leave, rather like a shopping centre. This
model includes the closed system of the last example and also the queues
186 5. Applications
in series of Example 5.2.4. Let X
t
= (X; : i E I), where X; denotes the
number of customers at node i at time t. Then (Xt)t>o is a Markov chain
in Y= N
I
and the nonzero transition rates are given by
q(n, bin) = Ai, q(bi n , bjn) = qij, q(bjn, n) = /lj
for n E I and distinct states i, LEI. We shall ass':,me that Ai > 0 for some
i and /lj > 0 for some j; then Q is irreducible on I.
The system of equations (5.1) for an invariant measure is replaced here
by
1ri (f.1,i + L qi
j
) = Ai + L 1rjqji·
j#i j#i
This system has a unique solution, with 1ri > 0 for all i. This may be seen
by considering the invariant distribution for the extended Qmatrix Q on
I U {8} with offdiagonal entries
q8j = Aj, qij = qij, qi8 = /li·
On summing the system over i E I we find
L 1rif.1,i = L Ai.
iEI iEI
As in the last example, for n = (ni : i E I) we set
1f(n) = II1 r ~ i .
iEI
Transitions from m E I may be divided into those where a new particle is
added and, for each i E I with mi 2:: 1, those where a particle is moved
from i to somewhere else. We have, for the first sort of transition
1f(m) Lq(m,8j m) = 1f(m) LAj
JEI JEI
= 1f(m) L 1rjf.1,j = L 1f(8j m)q(8j m, m)
JEI JEI
and for the second sort
1f(8
i
n) (q(8in, n) + L q(8
i
n, 8j n))
j#i
= II 1 r ~ k (1ri (f.1,i + L qij))
kEI j#i
= II 1 r ~ k ( Ai + L 1rjq
j
i)
kEI j#i
= 1f(n)q(n, 8
i
n) + L 1f(8j n)q(8j n, 8
i
n).
j#i
5.2 Queues and queueing networks
On summing these equations we obtain
1f(m) L: q(m, n) = L: 1f(n)q(n, m)
n#m n#m
187
so 7r is an invariant measure for Q. If 1ri < 1 for all i then 7r has finite total
mass niEI(l1ri), otherwise the total mass if infinite. Hence, Qis positive
recurrent if and only if 1ri < 1 for all i, and in that case, in equilibrium, the
individual queue lengths (Xi: i E I) are independent geometric random
variables with
Example 5.2.7 (M/G/! queue)
As we argued in Section 2.4, the Poisson process is the natural probabilistic
model for any uncoordinated stream of discrete events. So we are often justi
fied in assuming that arrivals to a queue form a Poisson process. In the pre
ceding examples we also assumed an exponential servicetime distribution.
This is desirable because it makes the queue size into a continuoustime
Markov chain, but it is obviously inappropriate in many realworld exam
ples. The service requirements of customers and the duration of telephone
calls have observable distributions which are generally not exponential. A
better model in this case is the MIGll queue, where G indicates that the
servicetime distribution is general.
We can characterize the distribution of a service time T by its distribu
tion function
F(t) = JP>(T ~ t),
or by its Laplace transform
(The integral written here is the LebesgueStieltjes integral: when T has
a density function f(t) we can replace dF(t) by f(t)dt.) Then the mean
service time Jl is given by
Jl = E(T) = £'(0+).
To analyse the MIGll queue, we consider the queue size X
n
immediately
following the nth departure. Then
(5.2)
188 5. Applications
where Y
n
denotes the number of arrivals during the nth service time. The
case where X
n
= 0 is different because then we get an extra arrival before
the (n + 1)th service time begins. By the Markov property of the Poisson
process, Y
1
, Y
2
, • •• are independent and identically distributed, so ( X n ) n ~ O
is a discretetime Markov chain. Indeed, except for visits to 0, ( X n ) n ~ O
behaves as a random walk with jumps Y
n
 1.
Let Tn denote the nth service time. Then, conditional on Tn = t, Y
n
is
Poisson of parameter At. So
and, indeed, we can compute the probability generating function
A(z) = E(zY
n
) = 1
00
E(zY
n
I Tn = t)dF(t)
= 1
00
eAt(lZ)dF(t) = £('\(1  z)).
Set p = E(Y
n
) = AJl. We call p the service intensity. Let us suppose
that p < 1. We have
Xn = Xo + (Y1 + · · · + Yn)  n + Zn
where Zn denotes the number of visits of X
n
to 0 before time n. So
E(X
n
) = E(Xo)  n(l  p) +E(Zn).
Take X
o
= 0, then, since X
n
~ 0, we have for all n
o< 1  p ~ E(Znln).
By the ergodic theorem we know that, as n + 00
E(Znln) + limo
where mo is the mean return time to o. Hence
mo ~ 1/(1  p) < 00
showing that ( X n ) n ~ O is positive recurrent.
Suppose now that we start ( X n ) n ~ O with its equilibrium distribution Jr.
Set
00
G(z) = E(zX
n
) = I: 1riZi
i==O
5.2 Queues and queueing networks
then
00
= lE(ZYn+l) (1r
O
Z+ L 1r
i
Z
i
)
i==l
= A(z) (1t"OZ + G(z) 1t"o)
so
(A(z)  z)G(z) = 1t"
o
A(z)(l  z).
By I'Hopital's rule, as z i 1
(A(z)  z)/(l z) + 1 A'(l) = 1 p.
189
(5.3)
Since G(l) = 1 = A(l), we must therefore have 1t"o = 1  p, rno = 1/(1  p)
and
G(z) = (1  p)(l  z)A(z)/(A(z)  z).
Since A is given explicitly in terms of the servicetime distribution, we
can now obtain, in principle, the full equilibrium distribution. The fact
that generating functions work well here is due to the additive structure of
(5.2).
To obtain the mean queue length we differentiate (5.3)
(A(z)  z)G'(z) = (A'(z)  l)G(z) = (1  p){ A'(z)(l  z)  A(z)},
then substitute for G(z) to obtain
G' (z) = (lp)A' (z) (1 z) (lp)A(z) {(A' (z) 1) (1 z) + A(z)  z} .
(A(z)  z) (A(z) _ Z)2
By l'Hopital's rule:
lim (A'(z)l)(lz) + A(z)z lim A"(z)(lz) _ __A_"_(l__)
zjl (A(Z)Z)2  zjl 2(A'(z)1) (A(z)z)  2(1p)2 ·
Hence
E(X
n
) = G'(l) = p + A"(l)/2(1  p)
=p + A
2
£"(0+)/2(1  p) = p + A
2
E(T
2
)/2(1  p).
In the case of the M/M/1 queue p = AIJl, E(T
2
) = 2/Jl2 and E(X
n
) =
p/(l  p) = A/(Jl  A), as we found in Example 5.2.1.
190 5. Applications
We shall use generating functions to study two more quantities of interest
 the queueing time of a typical customer and the busy periods of the server.
Consider the queue (Xn)nEZ in equilibrium. Suppose that the customer
who leaves at time 0 has spent time Q queueing to be served, and time T
being served. Then, conditional on Q + T = t, X
o
is Poisson of parameter
>..t, since the customers in the queue at time 0 are precisely those who
arrived during the queueing and service times of the departing customer.
Hence
G(z) = E(e
A
(Q+T)(lZ)) = M(A(l z))L(A(l z))
where M is the Laplace transform
On substituting for G(z) we obtain the formula
M(w) = (1  p)w/(w  .x(1  L(w))).
Differentiation and l'Hopital's rule, as above, lead to a formula for the mean
queueing time
E(Q) = M'(O+) = .xL"(0+)
2 (1 + AL' (0+)) 2
We now turn to the busy period 8. Consider the Laplace transform
Let T denote the service time of the first customer in the busy period. Then
conditional on T = t, we have
8=t+8
1
+ ... +8N,
where N is the number of customers arriving while the first customer is
served, which is Poisson of parameter At, and where 8
1
,8
2
, ... are inde
pendent, with the same distribution as 8. Hence
B(w) = 1
00
E(e
WS
IT = t)dF(t)
= 1
00
ewte>.t(lB(w»dF(t) = L(w + .x(1 B(w))).
5.2 Queues and queueing networks 191
Although this is an implicit relation for B(w), we can obtain moments by
differentiation:
E(S) = B'(O+) = £'(0+)(1  AB'(O+)) = Jl(l + AE(S))
so the mean length of the busy period is given by
E(S) = Jl/(l  p).
Example 5.2.8 (M/G/oo queue)
Arrivals at this queue form a Poisson process, of rate A, say. Service times
are independent, with a common distribution function F(t) = lP(T ~ t).
There are infinitely many servers, so all customers in fact receive service
at once. The analysis here is simpler than in the last example because
customers do not interact. Suppose there are no customers at time O.
What, then, is the distribution of the number X
t
being served at time t?
The number Nt of arrivals by time t is a Poisson random variable of
parameter At. We condition on Nt = n and label the times of the n arrivals
randomly by AI, ... ,An. Then, by Theorem 2.4.6, AI, ... ,An are inde
pendent and uniformly distributed on the interval [0, t]. For each of these
customers, service is incomplete at time t with probability
1 it 1 it
p= JP>(T>s)ds= (lF(s))ds.
tot 0
Hence, conditional on Nt = n, X
t
is binomial of parameters nand p. Then
00
P(X
t
= k) = L P(X
t
= k I Nt = n)P(N
t
= n)
n==O
00
= e>.t(>.pt)k /k! L('\(1  p)tt
k
/(n  k)!
n==k
= eAPt(Apt)k /k!
So we have shown that X
t
is Poisson of parameter
,\ it(1  F(s))ds.
192
Recall that
5. Applications
Hence if E(T) < 00, the queue size has a limiting distribution, which is
Poisson of parameter ,xE(T).
For further reading see Reversibility and Stochastic Networks by F. P.
Kelly (Wiley, Chichester, 1978).
5.3 Markov chains in resource management
Management decisions are always subject to risk because of the uncertainty
of future events. If one can quantify that risk, perhaps on the basis of past
experience, then the determination of the best action will rest on the cal
culation of probabilities, often involving a Markov chain. Here we present
some examples involving the management of a resource: either the stock
in a warehouse, or the water in a reservoir, or the reserves of an insurance
company. See also Exercise 3.7.1 on the maintenance of unreliable equip
ment. The statistical problem of estimating transition rates for Markov
chains has already been discussed in Section 1.10.
Example 5.3.1 (Restocking a warehouse)
A warehouse has a capacity of c units of stock. In each time period n, there
is a demand for D
n
units of stock, which is met if possible. We denote
the residual stock at the end of period n by X
n
. The warehouse manager
restocks to capacity for the beginning of period n + 1 whenever X
n
~ m,
for some threshold m. Thus ( X n ) n ~ O satisfies
if X
n
~ m
ifm < X
n
~ c.
Let us assume that D
1
, D
2
, ••• are independent and identically distributed;
then ( X n ) n ~ O is a Markov chain, and, excepting some peculiar demand
structures, is irreducible on {O, 1, ... ,c}. Hence ( X n ) n ~ O has a unique
invariant distribution 7r which determines the longrun proportion of time
in each state. Given that X
n
= i, the expected unmet demand in period
n +1 is given by
if i ~ m
if m < i :s; c.
5.3 Markov chains in resource management
Hence the longrun proportion of demand that is unmet is
c
u(m) = L 7l"i
U
i·
i==O
The longrun frequency of restocking is given by
m
r(m) = L7l"i.
i==O
193
Now as m increases, u(m) decreases and r(m) increases. The warehouse
manager may want to compute these quantities in order to optimize the
longrun cost
ar(m) +bu(m)
where a is the cost of restocking and b is the profit per unit.
There is no general formula for 1r, but once the distribution of the demand
is known, it is a relatively simple matter to write down the (c +1) x (c +1)
transition matrix P for and solve 1rP = 1r subject to 1ri = 1.
We shall discuss in detail a special case where the calculations work out
nicely.
Suppose that the capacity c = 3, so possible threshold values are m =
0,1,2. Suppose that the profit per unit b = 1, and that the demand satisfies
P(D i) = 2
i
for i = 0,1,2, ....
Then
00 00
lE((D  i)+) = LIP((D  i)+ 2: k) = LIP(D 2: i + k) = 2
i
.
k=l k=l
The transition matrices for m = 0,1,2 are given, respectively, by
(
1/8 1/8 1/4 1/2) (1/8 1/8 1/4 1/2) (1/8 1/8 1/4 1/2)
1/2 1/2 0 0 1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/2
1/4 1/4 1/2 0 1/4 1/4 1/2 0 1/8 1/8 1/4 1/2
1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/2
with invariant distributions
(1/4,1/4,1/4,1/4), (1/6,1/6,1/3,1/3), (1/8,1/8,1/4,1/2).
Hence
u(O) = 1/4, u(l) = 1/6, u(2) = 1/8
194
and
5. Applications
r(O) = 1/4, r(l) = 1/3, r(2) = 1/2.
Therefore, to minimize the longrun cost ar(m) +u(m) we should take
{
2 if a ~ 1/4
m = 1 if 1/4 < a ~ 1
o if 1 < a.
Example 5.3.2 (Reservoir model  discrete time)
We are concerned here with a storage facility, for example a reservoir, of
finite capacity c. In each time period n, An units of resource are available
to enter the facility and B
n
units are drawn off. When the reservoir is
full, surplus water is lost. When the reservoir is empty, no water can be
supplied. We assume that newly available resources cannot be used in the
current time period. Then the quantity of water X
n
in the reservoir at the
end of period n satisfies
X
n
+
1
= ((X
n
Bn+I)+ +A
n
+
l
) Ac.
If we assume that An, B
n
and c are integervalued and that AI, A
2
, ••• are
independent and identically distributed, likewise B
I
, B
2
, ••• , then ( X n ) n ~ O
is a Markov chain on {O, 1, ... ,c}, whose transition probabilities may be
deduced from the distributions of An and B
n
. Hence we know that the long
run behaviour of ( X n ) n ~ O is controlled by its unique invariant distribution
7r, assuming irreducibility. For example, the longrun proportion of time
that the reservoir is empty is simply 7ro. So we would like to calculate 7r.
A simplifying assumption which makes some calculations possible is to
assume that consumption in each period is constant, and that our units are
chosen to make this constant 1. Then the infinite capacity model satisfies
a recursion similar to the M/G/l queue:
X
n
+
1
= (X
n
 1)+ + A
n
+
l
.
Hence, by the argument used in Example 5.2.7, if E(A
n
) < 1, then ( X n ) n ~ O
is positive recurrent and the invariant distribution 7r satisfies
00
L 7riZi = (1  EA
n
)(1  z)A(z)j(A(z)  z)
i=O
where A(z) = E(zA
n
). In fact, whether or not E(A
n
) < 1, the equation
00
L lIiZi = (1  z)A(z)j(A(z)  z)
i=O
5.3 Markov chains in resource management 195
serves to define a positive invariant measure v for To see this,
multiply by A(z)  z and equate powers of z: the resulting equations are
the equilibrium equations for
Vo = lIo(ao +a1) +V1
a
O
i
l/i = l/i+l
a
O + L l/j
a
ij+1, for i 1
j=O
where ai = P(A
n
= i).
Note that can only enter {O, 1, ... ,c} through c. Hence, by the
strong Markov property, observed whilst in {O, 1, ... ,c} is simply
the finitecapacity model. In the case where IE(A
n
) < 1, we can deduce for
the finitecapacity model that the longrun proportion of time in state i is
given by vi/(vo +... +v
e
). In fact, this is true in general as the equilibrium
equations for the finitecapacity model coincide with those for v up to level
c  1, and the level c equation is redundant.
In reality, it is to be hoped that, in the long run, supply will exceed
demand, which is true if E(A
n
) > 1. Then is transient, so II must
have infinite total mass. The problem faced by the water company is to
keep the longrun proportion of time 1ro(c) that the reservoir is empty below
a certain acceptable fraction, € > °say. Hence c should be chosen large
enough to make
lIO/(lIo +... +lie) < c
which is always possible in the transient case.
Example 5.3.3 (Reservoir model  continuous time)
Consider a reservoir model where fresh water arrives at the times of
a Poisson process of rate A. The quantities of water 8
1
,8
2
, ... arriving
each time are assumed independent and identically distributed. We assume
that there is a continuous demand for water of rate 1. For a reservoir of
infinite capacity, the quantity of water held is just the stored
work in an M/G/I queue with the same arrival times and service times
8
1
,8
2
,. ... The periods when the reservoir is empty correspond to idle
periods of the queue. Hence in the positive recurrent case where AE(8
n
) <
1, the longrun proportion of time that the reservoir is empty is given by
IE(Sn)/(l  ,xIE(Sn)). Note that can enter [0, c] only through c.
As in the preceding example we can obtain the finite capacity model by
observing whilst in [0, c], but we shall not pursue this here.
The next example is included, in part, because it illustrates a surprising
and powerful connection between reflected random walks and the maxima
196 5. Applications
of random walks, which we now explain. Let Xl, X
2
, • •• denote a sequence
of independent, identically distributed random variables. Set Sn = Xl +
... + X
n
and define ( Z n ) n ~ O by Zo = 0 and
Zn+l = (Zn + X
n
+
l
)+.
Then, by induction, we have
so Zn has the same distribution as M
n
where
Example 5.3.4 (Ruin of an insurance company)
An insurance company receives premiums continuously at a constant rate.
We choose units making this rate 1. The company pays claims at the times
of a Poisson process of rate X, the claims YI, Y
2
, ••• being independent and
identically distributed. Set p = XE(Yi) and assume that p < 1. Then in
the long run the company can expect to make a profit of 1  P per unit
time. However, there is a danger that large claims early on will ruin the
company even though the longterm trend is good.
Denote by Sn the cumulative net loss following the nth claim. Thus
Sn = Xl + .. · + X
n
, where X
n
= Y
n
 Tn and Tn is the nth interarrival
time. By the strong law of large numbers
as n ~ 00. The maximum loss that the company will have to sustain is
M= lim M
n
n+oo
where
By the argument given above, M
n
has the same distribution as Zn, where
Zo = 0 and
Zn+l = (Zn +Y
n
 T
n
)+.
But Zn is the queueing time of the nth customer in the M/G/1 queue with
interarrival times Tn and service times Y
n
. We know by Example 5.2.7 that
the queuelength distribution converges to equilibrium. Hence, so does the
5.4 Markov decision processes 197
queueingtime distribution. Also by Example 5.2.7, we know the Laplace
transform of the equilibrium queueingtime distribution. Hence
The probability of eventual bankruptcy is P(M > a), where a denotes the
initial value of the company's assets. In principle, this may now be obtained
by inverting the Laplace transform.
5.4 Markov decision processes
In many contexts costs are incurred at a rate determined by some process
which may best be modelled as a Markov chain. We have seen in Section
1.10 and Section 4.2 how to calculate in these circumstances the longrun
average cost or the expected total cost. Suppose now that we are able to
choose the transition probabilities for each state from a given class and that
our choice determines the cost incurred. The question arises as to how best
to do this to minimize our expected costs.
Example 5.4.1
A random walker on {O, 1, 2, ... } jumps one step to the right with proba
bility p and one step to the left with probability q = 1  p. Any value of
p E (0,1] may be chosen, but incurs a cost
c(p) = l/p.
The walker on reaching 0 stays there, incurring no further costs.
If we are only concerned with minimizing costs over the first few time
steps, then the choice p = 1 may be best. However, in the long run the only
way to avoid an infinite total cost is to get to O. Starting from i we must
first hit iI, then i  2, and so on. Given the lack of memory in the model,
this makes it reasonable to pick the same value of p throughout, and seek
to minimize ¢(p), the expected total cost starting from 1. The expected
total cost starting from 2 is 2¢(p) since we must first hit 1. Hence
¢(p) = c(p) + 2p¢(p)
so that
¢(p) = { c(p)/(l  2p) for p < 1/2
00 for p ~ 1/2.
Thus for c(p) = lip the choice p = 1/4 is optimal, with expected cost 8.
The general discussion which follows will make rigorous what we claimed
was reasonable.
198 5. Applications
Generally, let us suppose given some distribution A = (Ai: i E I) and,
for each action a E A, a transition matrix P(a) = (Pij(a) : i,j E I) and
a cost function c(a) = (Ci (a) : i E I). These are the data for a Markov
decision process, though so far we have no process and when we do it will
not in general be Markov. To get a process we must choose a policy, that
is, a way of determining actions by our current knowledge of the process.
Formally, a policy U is a sequence of functions
n=O,1,2, ....
Each policy u determines a probability law pu for a process ( X n ) n ~ O with
values in I by
(i) PU(X
o
= io) = Ai
o
;
(ii) PU(X
n
+
1
=in+11 X
o
=io, ... ,X
n
=i
n
) =Pi
n
i
n
+l(u
n
(i
o
, ... ,in)).
A stationary policy u is a function u : I ~ A. We abuse notation and write
u also for the associated policy given by
Under a stationary policy u, the probability law pu makes (Xn)n>O Markov,
with transition probabilities P"tj = Pij (u(i))· 
We suppose that a cost c(i, a) = ci(a) is incurred when action a is chosen
in state i. Then we associate to a policy u an expected total cost starting
from i, given by
00
VU(i) =E
U
LC(Xn,Un(X
o
, ... ,X
n
)).
n=O
So that this sum is well defined, we assume that c(i, a) ~ 0 for all i and a.
Define also the value function
V*(i) = infVU(i)
U
which is the minimal expected total cost starting from i.
The basic problem of Markov decision theory is how to minimize expected
costs by our choice of policy. The minimum expected cost incurred before
time n = 1 is given by
VI (i) = inf c(i, a).
a
Then the minimum expected cost incurred before time n = 2 is
V2 (i) = i ~ { c(i, a) + LPij(a)V
1
(j)}.
jE!
5.4 Markov decision processes
Define inductively
Vn+l (i) = i ~ f { c(i, a) +LPij(a)VnU)}.
jEI
199
(5.4)
It is easy to see by induction that V
n
(i) ~ V
n
+l (i) for all i, so V
n
(i) increases
to a limit Voo(i), possibly infinite. We have
Vn+l(i) :::; c(i,a) +LPij(a)VnU)
jEI
so, letting n ~ 00 and then minimizing over a,
for all a
Voo(i) :::; i ~ { c(i, a) +LPij(a)VOOU)}.
jEI
(5.5)
It is a reasonable guess that Voo(i), being the limit of minimal expected
costs over finite time intervals, is in fact the value function V*(i). This is
not always true, unless we can show that the inequality (5.5) is actually an
equality. We make three technical assumptions to ensure this. We assume
that
(i) for all i, j the functions Ci : A ~ [0,00) and Pij : A ~ [0,00) are
continuous;
(ii) for all i and all B < 00 the set {a : ci(a) ~ B} is compact;
(iii) for each i, for all but finitely many j, for all a E A we have Pij (a) = 0.
A simple case where (i) and (ii) hold is when A is a finite set. It is easy
to check that the assumptions are valid in Example 5.4.1, with A = (0,1],
ci(a) = l/a and
{
a ifj=i+1
Pij (a) = 0
1
 a if j = iI
otherwise,
with obvious exceptions at i = O.
Lemma 5.4.2. There is a stationary policy u such that
Voo(i) = c(i, u(i)) + LPij (u(i))VooU)·
jEI
(5.6)
Proof. If Voo(i) = 00 there is nothing to prove, so let us assume that
Voo(i) ~ B < 00. Then
Vn+l(i) = ! ~ {C(i,a) +LPij(a)Vn(j)}
jEJ
200 5. Applications
where K is the compact set {a : c(i, a) ~ B} and where J is the finite set
{j : Pij ¢. O}. Hence, by continuity, the infimum is attained and
V
n
+l(i) = c(i,un(i)) + :EPij(un(i))Vn(j)
jEJ
(5.7)
for some un(i) E K. By compactness there is a convergent subsequence
u
nk
(i) ~ u(i), say, and, on passing to the limit nk ~ 00 in (5.7), we obtain
(5.6). D
Theorem 5.4.3. We have
(i) Vn(i) i V*(i) as n ~ 00 for all i;
(ii) if u * is any stationary policy such that a = u*(i) minimizes
c(i,a) + :EPij (a)V* (j)
jEI
for all i, then u* is optimal, in the sense that
v
u
* (i) = V*(i)
Proof. For any policy u we have
00
for all i.
VU(i) = Ei L c(X
n
, Un (X
o
, · " ,X
n
))
n=O
= c(i,uo(i)) + :EPij(UO(i))vu[i](j)
jEI
where u[i] is the policy given by
Hence we obtain
VU(i) ~ i ~ { c(i, a) + LPij (a)V* (j) }
jEI
and, on taking the infimum over u
V*(i) ~ i ~ { c(i, a) +LPij (a)V* (j) }.
jEI
(5.8)
Certainly, Vo(i) = 0 ~ V*(i). Let us suppose inductively that Vn(i) ~ V*(i)
for all i. Then by substitution in the right sides of (5.4) and (5.8) we find
V
n
+
1
(i) ~ V*(i) and the induction proceeds. Hence Voo(i) ~ V*(i) for all i.
5.4 Markov decision processes
Let u* be any stationary policy for which
Voo(i) ~ c(i,u*(i)) +LPij(U*(i))Voo(j).
jEI
201
We know such a policy exists by Lemma 5.4.2. Then by Theorem 4.2.3 we
have vu* (i) ~ Voo(i) for all i. But V*(i) ~ vu* (i) for all i, so
Voo(i) = V*(i) = v
u
* (i)
and we are done. D
for all i
The theorem just proved shows that the problem of finding a good policy
is much simpler than we might have supposed. For it was not clear at the
outset that there would be a single policy which was optimal for all i, even
less that this policy would be stationary. Moreover, part (i) gives an explicit
way of obtaining the value function V* and, once this is known, part (ii)
identifies an optimal stationary policy.
In practice we may know only an approximation to V*, for example V
n
for n large. We may then hope that, by choosing a = u(i) to minimize
c(i, a) + :EPij(a)Vn(j)
jEI
we get a nearly optimal policy. An alternative means of constructing nearly
optimal policies is sometimes provided by the method of policy improve
ment. Given one stationary policy u we may define another Ou by the
requirement that a = (Ou)(i) minimizes
c(i,a) +:Epij(a)Vu(j).
jEI
Theorem 5.4.4 (Policy improvement). We have
(i) VeU(i) ~ VU(i) for all i;
(ii) VenU(i) ! V*(i) as n ~ 00 for all i, provided that
for all i. (5.9)
Proof. (i) We have, by Theorem 4.2.3
VU(i) = c(i,u(i)) + LPij(u(i))VU(j)
jEI
~ c(i,Ou(i)) + :EPij(Ou(i))vu(j)
jEI
so VU(i) ~ VeU(i) for all i, by Theorem 4.2.3.
202 5. Applications
(ii) We note from part (i) that
VOU(i) :::; c(i,a) +LPii(a)Vu(j)
jEI
for all i and a. (5.10)
Fix N 0 and consider for n = 0, 1, . .. ,N the process
nl
M
n
= VONnU(Xn) +LC(Xk,U*(Xk))'
k=O
Recall the notation for conditional expectation introduced in Section 4.1.
We have
E
U
* (M
n
+
1
I:F
n
) = LPXni (U*(Xn))VoNnlu(j) + c(X
n
, u*(X
n
))
jEI
nl
+ LC(Xk,U*(Xk))
k=O
where we used (5.10) with u replaced by (INnl u, i = X
n
and a = u*(X
n
).
It follows that EU· (M
n
+
1
) EU· (M
n
) for all n. Hence if we assume (5.9),
then
V
9N
U(i) = IEy· (M
o
) IEY· (M
N
)
=Ef(VU(XN)) +E
u
*
D
We have been discussing the minimization of expected total cost, which is
only relevant to the transient case. This is because we will have V* (i) = 00
unless for some stationary policy u, the only states j with positive cost
c(j, u(j)) > 0, accessible from i, are transient. The recurrent case is also
of practical importance and one way to deal with this is to discount costs
at future times by a fixed factor Q E (0,1). We now seek to minimize the
expected total discounted cost
00
V:(i) = EiLOnC(Xn,Un(X
o
, ... ,X
n
)).
n=O
Define the discounted value function
V;(i) = infV:(i).
U
5.4 Markov decision processes 203
In fact, the discounted case reduces to the undiscounted case by intro
ducing a new absorbing state 8 and defining a new Markov decision process
by
Pij (a) = apij (a),
ci(a) = ci(a),
Pia(a) = 1  a,
ca(a) = O.
Thus the new process follows the old until, at some geometric time of pa
rameter a, it jumps to 8 and stays there, incurring no further costs.
Introduce VO,a (i) = 0 and, inductively
Vn+l,o:(i) = i ~ f {c(i, a) + a 2:Pij (a)Vn,o: (in
jEJ
and, given a stationary policy u, define another Oau by the requirement
that a = (Oau)(i) minimizes
c(i, a) + a 2:Pij(a)V
U
U).
jEJ
Theorem 5.4.5. Suppose that the cost function c(i, a) is uniformly
bounded.
(i) We have Vn,a(i) i VC:(i) as n ~ 00 for all i.
(ii) The value function VC: is the unique bounded solution to
V;(i) = i ~ f { c(i, a) + a 2: Pij(a)V;U) }.
jEI
(iii) Let u* be a stationary policy such that a = u*(i) minimizes
c(i,a) +a 2:Pij(a)V;U)
jEI
for all i. Then u* is optimal in the sense that
(5.11)
V:* (i) = VC:(i)
(iv) For all stationary policies u we have
for all i.
as n ~ 00 for all i.
Proof. With obvious notation we have
204 5. Applications
so parts (i), (ii) and (iii) follow directly from Theorems 5.4.3 and 5.4.4,
except for the uniqueness claim in (ii). But given any bounded solution V
to (5.11), there is a stationary policy u such that
V(i) = c(i,u(i)) +0: :Epij(u(i))V(j).
JEI
Then V = V ~ , by Theorem 4.2.5. Then (}au = U so (iv) will show that u is
optimal and V = VC:.
We have c(i, a) ~ B for some B < 00. So for any stationary policy u we
have
00
V:(i) = Ei:E o:nc(x
n
, u(X
n
)) :::; B/(l  0:)
n=O
and so
iEf* (VU(X
n
)) = anEf* ( V ~ ( X n ) ) ~ Ba
n
/(l a) ~ 0
as n ~ 00. Hence (iv) also follows from Theorem 5.4.4. D
We finish with a discussion of longrun average costs. Here we are con
cerned with the limiting behaviour, as n ~ 00, of
We assume that
Ic(i, a)1 ~ B < 00
for all i and a.
This forces IV:(i) I ~ B for all n, but in general the sequence ~ (i) may
fail to converge as n ~ 00. In the case of a stationary strategy u for which
(Xn)n>O has a unique invariant distribution 1r
u
, we know by the ergodic
theorem that
1 nl
;;: LC(Xk,U(Xk)) ~ L
7r
j
c
U,u(j))
k=O JEI
as n ~ 00, Pialmost surely, for all i. So ~ (i) does converge in this
case by bounded convergence, with the same limit. This suggests that one
approach to minimizing longrun costs might be to minimize
L 7rjcU, u(j)).
JEI
But, although this is sometimes valid, we do not know in general that the
optimal policy is positive recurrent, or even stationary. Instead, we use a
martingale approach, which is more general.
5.4 Markov decision processes 205
Theorem 5.4.6. Suppose we can find a constant V* and a bounded func
tion W(i) such that
V* +W(i) = +LPii(a)W(j)}
jEI
for all i. (5.12)
Let u* be any stationary strategy such that a = u*(i) achieves the infimum
in (5.12) for each i. Then
(i) v::* (i) t V* as n t 00 for all i;
(ii) v:: (i) V* for all i, for all u.
Proof. Fix a strategy u and set Un = un(Xo, ... ,X
n
). Consider
nl
M
n
= W(X
n
)  nV* + L C(Xk' Uk).
k=O
Then
EU(Mn+l I F
n
)
= M
n
+ {C(X
n
, Un) +LPXni(Un)W(j)}  (V* +W(X
n
))
jEI
with equality if u = u*. Therefore
So we obtain
v* +2 sup IW(i)l/n.
i
This implies (ii) on letting n 00. When u = u* we also have
v::* (i) V* + 2 sup IW(i)l/n
i
and hence (i). D
The most obvious point of this theorem is that it identifies an optimal
stationary policy when the hypothesis is met. Two further aspects also
deserve comment. Firstly, if u is a stationary policy for which has
an invariant distribution 1r
u
, then
2: 1I"f (V* + W(i)) 2:1I"f (C(i'U(i)) + 2:Pii(U(i))W(j))
iEI iEI jEI
= 2:1I"fc(i,u(i)) + L1I"jW(j)
iEI jEI
206
so
5. Applications
v* ~ L 1rfc(i, u(i))
iEI
with equality if we can take u = u* .
Secondly, there is a connection with the case of discounted costs. Assume
that I is finite and that P(a) is irreducible for all a. Then we can show
that as a i 1 we have
V;(i) = V* /(1  a) +W(i) +0(1  a).
On substituting this into (5.11) we find
v* /(1  a) +W(i) + 0(1  a)
= i ~ f {C(i' a) + a LPij(a)(V*/(1 a) + W(j) + 0(1  a))}
jEI
so
V* + W(i) = i ~ f {C(i' a) + a LPij(a)W(j)} + 0(1  a)
jEI
which brings us back to (5.12) on letting a i 1.
The interested reader is referred to S.11. Ross, Applied Probability Mod
els with Optimization Applications (HoldenDay, San Francisco, 1970) and
to H. C. Tijms, Stochastic Models  an algorithmic approach (Wiley, Chich
ester, 1994) for more examples, results and references.
5.5 Markov chain Monte Carlo
Most computers may be instructed to provide a sequence of numbers
Ul = O.Ull U12
U
13 Ulm
U2 = O. U21 U22
U
23 U2m
U3 = 0.U31U32
U
33 U3m
written as decimal expansions of a certain length, which for many purposes
may be regarded as sample values of a sequence of independent random
variables, uniformly distributed on [0,1]:
5.5 Markov chain Monte Carlo 207
We are cautious in our language because, of course, Ul, U2, U3, ... are actu
ally all integer multiples of 10
m
and, more seriously, they are usually de
rived sequentially by some entirely deterministic algorithm in the computer.
Nevertheless, the generators of such pseudorandom numbers are in general
as reliable an imitation as one could wish of U1(w), U
2
(w), U
3
(w), .... This
makes it worth while considering how one might construct Markov chains
from a given sequence of independent uniform random variables, and then
might exploit the observed properties of such processes.
We shall now describe one procedure to simulate a Markov chain ( X n ) n ~ O
with initial distribution A and transition matrix P. Since EiEI Ai = 1 we
can partition [0,1] into disjoint subintervals (Ai: i E I) with lengths
Similarly for each i E I, we can partition [0,1] into disjoint subintervals
(A
ij
: j E I) such that
Now define functions
Go : [0, 1] ~ I,
G : I x [0, 1] ~ I
by
Go(U) = i
G(i,u)=j
if u E Ai,
if u E A
ij
.
Suppose that U
o
, U
1
, U
2
, • •• is a sequence of independent random variables,
uniformly distributed on [0,1], and set
X
o
= Go(U
o
),
X
n
+
1
= G(X
n
, U
n
+
1
) for n ~ 0.
Then
lP(X
o
= i) = lP(U
o
E Ai) = Ai,
lP(Xn +1 = i n +1 I Xo = io, ... ,Xn = in) = JP>(Un +1 E Ainin+l) = Pi
n
i
n
+l
so ( X n ) n ~ O is Markov(A, P).
This simple procedure may be used to investigate empirically those as
pects of the behaviour of a Markov chain where theoretical calculations
become infeasible.
208 5. Applications
The remainder of this section is devoted to one application of the simu
lation of Markov chains. It is the application which finds greatest practical
use, especially in statistics, statistical physics and computer science, known
as Markov chain Monte Carlo. Monte Carlo is another name for computer
simulation so this sounds no different from the procedure just discussed.
But what is really meant is simulation by means of Markov chains, the
object of primary interest being the invariant distribution of the Markov
chain and not the chain itself. After a general discussion we shall give two
examples.
The context for Markov chain Monte Carlo is a statespace in product
form
I = II 8m
mEA
where A is a finite set. For the purposes of this discussion we shall also
assume that each component 8
m
is a finite set. A random variable X with
values in I is then a family of component random variables (X(m) : mEA),
where, for each site mEA, X(m) takes values in 8
m
.
We are given a distribution 7r = (7ri : i E I), perhaps up to an unknown
constant multiple, and it is desired to compute the number
(5.13)
for some given function I = (Ii : i E I). The essential point to understand is
that A is typically a large set, making the statespace I very large indeed.
Then certain operations are computationally infeasible  performing the
sum (5.13) state by state for a start.
An alternative approach would be to simulate a large number of inde
pendent random variables Xl, ... ,X
n
in I, each with distribution 7r, and
to approximate (5.13) by
1 n
;;, L!(Xk).
k=l
The strong law of large numbers guarantees that this is a good approxi
mation as n ~ 00 and, moreover, one can obtain error estimates which
indicate how large to make n in practice. However, simulation from the
distribution 7r is also difficult, unless 7r has product form
1r(X) = II 1r
m
{x(m)).
mEA
For recall that a computer just simulates sequences of independent U[O, 1]
random variables. When 7r does not have product form, Markov chain
Monte Carlo is sometimes the only way to simulate samples from 7r.
5.5 Markov chain Monte Carlo 209
The basic idea is to simulate a Markov chain ( X n ) n ~ O , which is con
structed to have invariant distribution Jr. Then, assuming aperiodicity and
irreducibility, we know, by Theorem 1.8.3, that as n ~ 00 the distribution
of X
n
converges to Jr. Indeed, assuming only irreducibility, Theorem 1.10.2
shows that
with probability 1. But why should simulating an entire Markov chain be
easier than simulating a simple distribution Jr? The answer lies in the fact
that the statespace is a product.
Each component Xo(m) of the initial state X
o
is a random variable in 8
m
.
It does not matter crucially what distribution X
o
is given, but we might,
for example, make all components independent. The process ( X n ) n ~ O is
made to evolve by changing components one site at a time. When the
chosen site is m, we simulate a new random variable X
n
+
1
(m) with values
in 8
m
according to a distribution determined by X
n
, and for k =I m we set
Xn+1 ( k) = Xn ( k) . Thus at each step we have only to simulate a random
variable in 8
m
, not one in the much larger space I.
Let us write i ~ j if i and j agree, except possibly at site m. The law for
simulating a new value at site m is described by a transition matrix P(m),
where
pij(m) = 0 unless i ~ j.
We would like Jr to be invariant for P(m). A sufficient condition is that the
detailed balance equations hold: thus for all i, j we want
There are many possible choices for P(m) satisfying these equations. In
deed, given any stochastic matrix R(m) with
rij(m) = 0 unless i ~ j
we can determine such a P(m) by
for i =I j, and then
pii(m) = 1 2:Pij(m) ~ O.
j#i
210 5. Applications
This has the following interpretation: if X
n
= i we simulate a new random
variable Y
n
so that Y
n
= j with probability rij(m), then if Y
n
= j we set
with probability (7ririj(m)/7rjrji(m)) 1\ 1
otherwise.
This is called a Hastings algorithm.
There are two commonly used special cases. On taking
for i ~ j
we also find
for i ~ j.
So we simply resample X
n
(m) according to the conditional distribution
under 7r, given the other components. This is called the Gibbs sampler. It
is particularly useful in Bayesian statistics.
On taking rij(m) = rji(m) for all i and j we find
for i ~ j, i =1= j.
This is called a Metropolis algorithm. A particularly simple case would be
to take
rij(m) == l/(N
m
 1) for i ~ j, i =1= j
where N
m
== ISml. This amounts to choosing another value jm at site m
uniformly at random; if 7rj > 7ri, then we adopt the new value, whereas if
7rj ~ 7ri we adopt the new value with probability 7rj/7ri.
We have not yet specified a rule for deciding which site to visit when.
In practice this may not matter much, provided we keep returning to every
site. For definiteness we mention two possibilities. We might choose to visit
every site once and then repeat, generating a sequence of sites ( m n ) n ~ O .
Then (m
n
, X n ) n ~ O is a Markov chain in A x I. Alternatively, we might
choose a site randomly at each step. Then ( X n ) n ~ O is itself a Markov chain
with transition matrix
P = IAI
1
:E P(m).
mEA
We shall stick with this second choice, where the analysis is simpler to
present. Let us assume that P is irreducible, which is easy to ensure in the
examples. We know that
5.5 Markov chain Monte Carlo
for all m and all i, j, so also
211
7riPij = 7rjPji
and so 7r is the unique invariant measure for P. Hence, by Theorem 1.10.2,
we have
1 n1
 L f(Xk) t L 7rdi
n k=O iEI
as n ~ 00 with probability 1. Thus the algorithm works eventually. In
practice one is concerned with how fast it works, but useful information
of this type cannot be gained in the present general context. Given more
information on the structure of 8
m
and the distribution 7r to be simulated,
much more can be said. We shall not pursue the matter here. It should
also be emphasised that there is an empirical side to simulation: with due
caution informed by the theory, the computer output gives a good idea
of how well we are doing. For further reading we recommend Stochastic
Simulation by B. D. Ripley (Wiley, Chichester, 1987), and Markov Chain
Monte Carlo in practice by W. R. Gilks, S. Richardson and D. J. Spiegelhal
ter (Chapman and Hall, London, 1996). The recent survey article Bayesian
computation and stochastic systems by J. Besag, P. Green, D. Higdon and
K. Mengersen (Statistical Science, 10 (1), pp. 340, 1995) contains many
interesting references. We finish with two examples.
Example 5.5.1 (Bayesian statistics)
In a statistical problem one may be presented with a set of independent
observations Y
1
, . .. ,Y
n
, which it is reasonable to assume are normally dis
tributed, but with unknown mean /l and variance 7
1
. One then seeks
to draw conclusions about /l and 7 on the basis of the observations. The
Bayesian approach to this problem is to assume that /l and 7 are themselves
random variables, with a given prior distribution. For example, we might
assume that
/l rv N(0
0
, cPo
1
), 7 rv r(0:0, ,80),
that is to say, /l is normal of mean 0
0
and variance cPo
1
, and 7 has gamma
distribution of parameters 0:0 and ,80. The parameters 0
0
, cPo, 0:0 and ,80
are known. Then the prior density for (/l, 7) is given by
7r(/l, 7) ex exp{cPO(/l O
o
)2/2}7
ao

1
exp{,807}.
The posterior density for (/l, 7), which is the conditional density given
the observations, is then given by Bayes' formula
7r(/l, 7 I y) ex 7r(/l, 7) f (y I /l, 7)
ex exp{¢o(p (
0
)2/2} exp { Tt.(Yi p)2/2} T
ao
l+
n
/
2
exp{{jOT}.
212 5. Applications
Note that the posterior density is no longer in product form: the condition
ing has introduced a dependence between Jl and T. Nevertheless, the full
conditional distributions still have a simple form
1r(pl Y, r) ex exp{¢o(p (
0
)2/2} exp { r t.(Yi  p)2/2} I'.J N(On, ¢;;1),
1r(r I Y, p) ex r
CYo
l+
n
/
2
exp { r (,80 +t,(Yi  p)2/2) }I'.J r(a
n
, ,8n)
where
n
an = ao + n/2, ,8n = ,80 + 2)Yi  p)2/2.
i==l
Our final belief about Jl and T is regarded as measured by the posterior
density. We may wish to compute probabilities and expectations. Here the
Gibbs sampler provides a particularly simple approach. Of course, numeri
cal integration would also be feasible as the dimension is only two. To make
the connection with our general discussion we set
1=8
1
X 8
2
= JR x [0,00).
We wish to simulate X = (Jl, T) with density 7r(Jl, T I y). The fact that JR
and [0, 00) are not finite sets does not affect the basic idea. In any case the
computer will work with finite approximations to JR and [0,00). First we
simulate X
o
, say from the product form density 7r(Jl, T). At the kth stage,
given X
k
= (Jlk, Tk), we first simulate Jlk+1 from 7r(Jl I y, Tk) and then Tk+1
from 7r(T I y,Jlk+1), then set Xk+1 = (Jlk+1,Tk+1). Then ( X k ) k ~ O is a
Markov chain in I with invariant measure 7r(Jl, T I y), and one can show
that
k1
~ ~ f(Xj ) t 1f(x)1r(x I y)dx as k t 00
with probability 1, for all bounded continuous functions f : I + JR. This
is not an immediate consequence of the ergodic theorem for discrete state
space, but you may find it reasonable at an intuitive level, with a rate of
convergence depending on the smoothness of 7r and f.
We now turn to an elaboration of this example where the Gibbs sampler
is indispensible. The model consists of m copies of the preceding one, with
5.5 Markov chain Monte Carlo 213
different means but a common variance. Thus there are mn independent
observations }!ij, where i = 1, ... n, and j = 1, ... ,m, normally distributed,
with means jjj and common variance 7
1
. We take these parameters to be
independent random variables as before, with
Let us write jj = (jj1, . .. ,jjn). The prior density is given by
and the posterior density is given by
1r(Jl, T I y) ex exp { ¢o ~ ( J  l j  ( 0 )2/2}
x exp { Tt.~ ( Y i j  Jlj)2/2} TQoHmn/2 exp{,8oT}.
Hence the full conditional distributions are
where
n m
On = 00 + mn/2, ,8n = f30 + L L(Yij  Jlj)2/2.
i=1 j=1
We can construct approximate samples from 7r(jj, 7 I y), just as in the case
m = 1 discussed above, by a Gibbs sampler method. Note that, conditional
on 7, the means jjj, for j = 1, ... ,m, remain independent. Thus one can
update all the means simultaneously in the Gibbs sampler. This has the.
effect of speeding convergence to the equilibrium distribution. In cases
where m is large, numerical integration of 7r(jj,7 I y) is infeasible, as is
direct simulation from the distribution, so the Markov chain approach is
the only one available.
214 5. Applications
Example 5.5.2 (Ising model and image analysis)
Consider a large box A = AN in 71
2
A = {N, ... ,1,0,1, ... ,N}2
with boundary 8A = AN\A
N

1
, and the configuration space
For x E A define
H(x) = ! I::(x(m)  x(m,))2
where the sum is taken over all pairs {m, m
/
} ~ A with 1m  m'l = 1. Note
that H(x) is small when the values taken by x at neighbouring sites are
predominantly the same. We write
I+ = {x E I : x(m) = 1 for all m E 8A}
and for each (3 > °define a probability distribution (rr(x) : x E I+) by
1t"(x) ex e(3H(x).
As (3 ! °the weighting becomes uniform, whereas, as (3 i 00 the mass
concentrates on configurations x where H(x) is small. This is one of the
fundamental models of statistical physics, called the Ising model. A famous
and deep result of Onsager says that if X has distribution 1t", then
In particular, if sinh 2(3 ~ 1, the fact that X is forced to take boundary
values 1 does not significantly affect the distribution of X(O) when N is
large, whereas if sinh 2{3 > 1 there is a residual effect of the boundary
values on X(O), uniformly in N.
Here we consider the problem of simulating the Ising model. Simulations
may sometimes be used to guide further developments in the theory, or even
to detect phenomena quite out of reach of the current theory. In fact, the
Ising model is rather well understood theoretically; but there are many
related models which are not, where simulation is still possible by simple
modifications of the methods presented here.
First we describe a Gibbs sampler. Consider the sets of even and odd
sites
A+ = {(ml' m2) E A : ml +m2 is even},
A = {(ml' m2) E A : ml +m2 is odd}
and for x E I set
5.5 Markov chain Monte Carlo 215
x± = (x(m) : m E A±).
We can exploit the fact that the conditional distribution 1r(x+ I x) has
product form
1r(X+ I x) ex II e/3x(m)s(m)
mEA+\8A
where, for mEA+\8A
s(m) = L x(m').
Im'ml=l
Therefore, it is easy to simulate from 1r(x+ I x) and likewise from
1r(x I x+). Choose now some simple initial configuration X
o
in 1+. Then
inductively, given X;; = x, simulate firstly X ~ + l with distribution
1r(. I x) and then given X ~ + l = x+, simulate X;+l with distribution
1r(. I x+). Then according to our general discussion, for large n, the distri
bution of X
n
is approximately 1r. Note that we did not use the value of the
normalizing constant
Z = L e/3H(x)
xEI+
wllich is hard to compute by elementary means when N is large.
An alternative approach is to use a Metropolis algorithm. We can again
exploit the even/odd partition. Given that X
n
= x, independently for each
m E A+\8A, we change the sign of Xt(m) with probability
p(m,x) = (1r(x)/1r(x)) 1\ 1 = e
2
,Bx(m)s(m) 1\ 1
where x ~ x with x(m) = x(m). Let us call the resulting configuration
Y
n
. Next we apply the corresponding transformation to Y
n
 ( m) for the odd
sites m E A\8A, to obtain X
n
+
1
. The process ( X n ) n ~ O is then a Markov
chain in 1+ with invariant distribution 1r.
Both methods we have described serve to simulate samples from 1r; there
is little to choose between them. Convergence is fast in the subcritical case
sinh 2,8 < 1, where 1r has an approximate product structure on large scales.
In a Bayesian analysis of twodimensional images, the Ising model is
sometimes used as a prior. We may encode a digitized image on a two
dimensional grid as a particular configuration (x(m) : mEA) E I, where
x(m) = 1 for a white pixel and x(m) = 1 for a black pixel. By varying
the parameter ,8 in the Ising model, we vary the tendency of black pixels
216 5. Applications
to clump together; the same for white pixels. Thus (3 is a sort of texture
parameter, which we choose according to the sort of image we expect, thus
obtaining a prior 7r(x). Observations are now made at each site which record
the true pixel, black or white, with probability p E (0,1). The posterior
distribution for X given observations Y is then given by
7r(x I y) ex 7r(x)f(y I x) ex e
f3H
(x)pa(x,y) (1  p)d(x,y)
where a(x, y) and d(x, y) are the numbers of sites at which x and y agree and
disagree respectively. 'Cleanedup' versions of the observed image Y may
now be obtained by simulating from the posterior distribution. Although
this is not exactly the Ising model, the same methods work. We describe
the appropriate Metropolis algorithm: given that X
n
= x, independently
for each m E A+\8A, change the sign of X:(m) with probability
p(m, x, y) = (7r(x I Y)/7r(x I y)) 1\ 1
= e
2
,6x(m)s(m)((1_ p)/pt(m)y(m)
where x x with x(m) = x(m). Call the resulting configuration X
n
+
1
/
2
.
Next apply the corresponding transformation to for the odd sites
to obtain X
n
+
1
. Then is a Markov chain in /+ with invariant
distribution 7r(. I y).
6
Appendix: probability and measure
Section 6.1 contains some reminders about countable sets and the discrete
version of measure theory. For much of the book we can do without explicit
mention of more general aspects of measure theory, except an elementary
understanding of Riemann integration or Lebesgue measure. This is because
the statespace is at worst countable. The proofs we have given may be read
on two levels, with or without a measuretheoretic background. When in
terpreted in terms of measure theory, the proofs are intended to be rigorous.
The basic framework of measure and probability is reviewed in Sections 6.2
and 6.3. Two important results of measure theory, the monotone conver
gence theorem and Fubini's theorem, are needed a number of times: these
are discussed in Section 6.4. One crucial result which we found impossi
ble to discuss convincingly without measure theory is the strong Markov
property for continuoustime chains. This is proved in Section 6.5. Finally,
in Section 6.6, we discuss a general technique for determining probability
measures and independence in terms of 1rsystems, which are often more
convenient than aalgebras.
6.1 Countable sets and countable sums
A set I is countable if there is a bijection f : {I, . .. ,n} ~ I for some n E N,
or a bijection f : N ~ I. In either case we can enumerate all the elements
of I
218 6. Appendix: probability and measure
where in one case the sequence terminates and in the other it does not.
There would have been no loss in generality had we insisted that all our
Markov chains had statespace N or {I, ... ,n} for some n E N: this just
corresponds to a particular choice of the bijection f.
Any subset of a countable set is countable. Any finite cartesian product
of countable sets is countable, for example tl
n
for any n. Any countable
union of countable sets is countable. The set of all subsets of N is uncount
able and so is the set of real numbers JR.
We need the following basic fact.
Lemma 6.1.1. Let I be a countably infinite set and let Ai ~ 0 for all i E I.
Then, for any two enumerations of I
~ 1 , ~ 2 , ~ 3 , ••• ,
we have
00 00
LAin = LAin'
n=l n=l
Proof. Given any N E N we can find M ~ Nand N' ~ M such that
Then
N M N'
""" A· < """ A· < """ A·
L.J 't n  L.J In  L.J 't
n
n=l n=l n=l
and the result follows on letting N ~ 00. D
Since the value of the sum does not depend on the enumeration we are
justified in using a notation which does not specify an enumeration and
write simply
More generally, if we allow Ai to take negative values, then we can set
where
6.1 Countable sets and countable sums 219
allowing that the sum over I is undefined when the sums over I+ and I
are both infinite. There is no difficulty in showing for Ai, jji 2 0 that
I)Ai + Pi) = LAi + LPi.
iEI iEI iEI
By induction, for any finite set J and for Aij 2 0, we have
L (LAi
j
) = L (LAij).
iEI jEJ jEJ iEI
The following two results on sums are simple versions of fundamental
results for integrals. We take the opportunity to prove these simple versions
in order to convey some intuition relevant to the general case.
Lemma 6.1.2 (Fubini's theorem  discrete case). Let I and J be
countable sets and let Aij 2 0 for all i E I and j E J. Then
L (LAi
j
) = L (LAij).
iEI jEJ jEJ iEI
Proof. Let jl,j2,j3, ... be an enumeration of J. Then
as n ~ 00. Hence
and the result follows by symmetry. D
Lemma 6.1.3 (Monotone convergence  discrete case). Suppose for
each i E I we are given an increasing sequence ( A i ( n ) ) n ~ O with limit Ai, .
and that Ai ( n) 2 0 for all i and n. Then
LAi(n) i LAi as nt 00.
iEI iEI
220 6. Appendix: probability and measure
Proof. Set 8
i
(1) = Ai(l) and for n ~ 2 set
Then 8
i
(n) ~ 0 for all i and n, so as n ~ 00, by Fubini's theorem
~ A i ( n ) = ~ (t,8
i
(k))
= t. ( ~ 8 i ( k ) ) i t. ( ~ 8 i ( k ) )
= L(f
8i
(k)) = LAi
o
iEI k=l iEI
6.2 Basic facts of measure theory
D
We state here for easy reference the basic definitions and results of measure
theory. Let E be a set. A aalgebra £ on E is a set of subsets of E satisfying
(i) 0 E £;
(ii) A E £ =* AC E £;
(iii) (An E £,n E N) =* Un An E £.
Here AC denotes the complement E\A of A in E. Thus £ is closed under
countable set operations. The pair (E, £) is called a measurable space. A
measure Jl on (E, £) is a function Jl : £ ~ [0,00] which has the following
countable additivity property:
The triple (E, £, Jl) is called a measure space. If there exist sets En E £,
n E N with Un En = E and Jl(E
n
) < 00 for all n, then we say Jl is afinite.
Example 6.2.1
Let I be a countable set and denote by I the set of all subsets of I. Recall
that A = (Ai: i E I) is a measure in the sense of Section 1.1 if Ai E [0,00)
for all i. For such A we obtain a measure on the measurable space (I,I) by
setting
In fact, we obtain in this way all afinite measures Jl on (I,I).
6.2 Basic facts of measure theory 221
Example 6.2.2
Let A be any set of subsets of E. The set of all subsets of E is a a
algebra containing A. The intersection of any collection of aalgebras is
again a aalgebra. The collection of aalgebras containing A is therefore
nonempty and its intersection is a aalgebra a(A), which is called the
aalgebra generated by A.
Example 6.2.3
In the preceding example take E = JR and
A = {(a,b): a,b E JR,a < b}.
The aalgebra B generated by A is called the Borel aalgebra of JR. It can
be shown that there is a unique measure J.t on (JR, B) such that
J.t(a, b) = b  a for all a, b.
This measure J.t is called Lebesgue measure.
Let (E
1
, £1) and (E
2
, £2) be measurable spaces. A function f : E
1
~ E
2
is measurable if f1(A) E £1 whenever A E £2. When the range E
2
= JR we
take £2 = B by default. When the range E
2
is a countable set I we take £2
to be the set of all subsets I by default.
Let (E, £) be a measllrable space. We denote by m£ the set of measurable
functions f : E ~ JR. Then m£ is a vector space. We denote by m£+ the
set of measurable functions f : E ~ [0, 00], where we take on [0, 00] the
aalgebra generated by the open intervals (a, b). Then m£+ is a cone
(f, 9 E m£+ ,0:, (3 ~ 0) ~ o:f +{3g E m£+.
Also, m£+ is closed under countable suprema:
(fi E m£+,i E I) ~ SUpfi E m£+.
i
It follows that, for a sequence of functions f n E m£+, both limsUPn f nand
liminf
n
fn are in m£+, and so is limn fn when this exists. It can be shown
that there is a unique map ji : m£+ ~ [0,00] such that
(i) ji(lA) = J.t(A) for all A E £;
(ii) ji(o:f +(3g) = o:ji(f) +(3ji(f) for all f,g E m£+, 0:, {3 ~ 0;
(iii) (fn E m£+, n E N) ~ ji(En fn) = En ji(fn).
222 6. Appendix: probability and measure
For f E mE, set f± = (±f) V 0, then f+,f E m£+, f = f+  f and
IfI = f+ + f· If jt(lfl) < 00 then f is said to be integrable and we set
We call ji(f) the integral of f. It is conventional to drop the tilde and
denote the integral by one of the following alternative notations:
p(J) = r fdp = r f(x)p(dx).
lE lXEE
In the case of Lebesgue measure jj, one usually writes simply
r f(x)dx.
lXEJR
6.3 Probability spaces and expectation
The basic apparatus for modelling randomness is a probability space
(0, F, P). This is simply a measure space with total mass P(O) = 1. Thus
F is a aalgebra of subsets of 0 and P : F ~ [0,1] satisfies
(i) P(O) = 1;
(ii) P(AI n A
2
) = P(A
I
) +P(A
2
) for AI, A
2
disjoint;
(iii) P(A
n
) i P(A) whenever An i A.
In (iii) we write An i A to mean Al ~ An ~ ... with Un An = A. A
measurable function X defined on (0, F) is called a random variable. We
use random variables Y : 0 ~ lR to model random quantities, where for a
Borel set B ~ lR the probability that Y E B is given by
P(Y E B) = P({w: Y(w) E B}).
Similarly, given a countable statespace I, a random variable X : 0 ~ I
models a random state, with distribution
Ai = P(X = i) = p({w : X(w) = i}).
To every nonnegative or integrable realvalued random variable Y is asso
ciated an average value or expectation E(Y), which is the integral of Y with
respect to P. Thus we have
(i) E(IA) = P(A) for A E F;
(ii) E(oX +(3Y) = oE(X) +(3E(Y) for X, Y E mF+, o,{3 ~ 0;
6.4 Monotone convergence and Fubini's theorem 223
(iii) (Y
n
E mF+, n E N, Y
n
i Y) :::} IE(Y
n
) i IE(Y).
When X is a random variable with values in I and f : I [0,00] the
expectation of Y = f(X) = foX is given explicitly by
E(J(X)) = LAdi
iEI
where A is the distribution of X. For a realvalued random variable Y the
probabilities are sometimes given by a measurable density function p in
terms of Lebesgue measure:
P(Y E B) = Lp(y)dy.
Then for any measurable function f : lR [0,00] there is an explicit formula
E(J(Y)) = Lf(y)p(y)dy.
6.4 Monotone convergence and Fubini's theorem
Here are the two theorems from measure theory that come into play in
the main text. First we shall state the theorems, then we shall discuss
some places where they are used. Proofs may be found, for example, in
Probability with Martingales by D. Williams (Cambridge University Press,
1991).
Theorem 6.4.1 (Monotone convergence). Let (E, £, Jt) be a measure
space and let be a sequence of nonnegative measurable functions.
Then, as n 00
(fn(x) i f(x) for all x E E) :::} Jt(fn) i Jt(f)·
Theorem 6.4.2 (Fubini's theorem). Let (E
1
, £1, Jl1) and (E
2
, £2, Jl2) be
two afinite measure spaces. Suppose that f : E
1
x E
2
[0, 00] satisfies
(i) x f(x, y) : E
1
[0,00] is £1 measurable for all Y E E
2
;
(ii) Y IXEE
1
f(x, y)Jt1(dx) : E
2
[0,00] is £2 measurable.
Then
(a) y f(x, y) : E
2
[0,00] is £2 measurable for all x E E
1
;
(b) x f
yE
E
2
f(x, Y)Jl2(dy) : E
1
[0,00] is £1 measurable;
(c) r (1 (r
JxEE
l
yEE2 ') yEE2 JxEE
l
')
224 6. Appendix: probability and measure
The measurability conditions in the above theorems rarely need much
consideration. They are powerful results and very easy to use. There is
an equivalent formulation of monotone convergence in terms of sums: for
nonnegative measurable functions 9n we have
To see this just take .fn = 91 +... +9n. This form of monotone convergence
has already appeared in Section 6.2 as a defining property of the integral.
This is also a special case of Fubini's theorem, provided that (E, £, Jt) is
afinite: just take E
2
= {I, 2, 3, ... } and Jt2( {n}) = 1 for all n.
We used monotone convergence in Theorem 1.10.1 to see that for a non
negative random variable Y we have
IE(Y) = lim IE(Y /\ N).
N+oo
We used monotone convergence in Theorem 2.3.2 to see that for random
variables Sn ~ 0 we have
E(LSn) = LE(Sn)
n n
and
E(ex
p
{ LSn}) = E ( J ~ = ex
p
{ L Sn})
n n ~ N
=J ~ = E(ex
p
{  L Sn}).
n ~ N
In the last application convergence is not monotone increasing but mono
tone decreasing. But if 0 ~ X
n
~ Y and X
n
! X then Y  X
n
i Y  X.
So IE(Y  X
n
) i IE(Y  X) and if IE(Y) < 00 we can deduce IE(X
n
) ! IE(X).
Fubini's theorem is used in Theorem 3.4.2 to see that
Thus we have taken (E
1
, £1, Jt1) to be [0,00) with Lebesgue measure and
(E
2
, £2, Jt2) to be the probability space with the measure Pi.
6.5 Stopping times and the strong Markov property
The strong Markov property for continuoustime Markov chains cannot
properly be understood without measure theory. The problem lies with the
6.5 Stopping times and the strong Markov property 225
notion of 'depending only on', which in measure theory is made precise as
measurability with respect to some aalgebra. Without measure theory the
statement that a set A depends only on (X
s
: s t) does not have a precise
meaning. Of course, if the dependence is reasonably explicit we can exhibit
it, but then, in general, in what terms would you require the dependence
to be exhibited? So in this section we shall give a precise measuretheoretic
account of the strong Markov property.
Let be a rightcontinuous process with values in a countable set
I. Denote by F
t
the aalgebra generated by {X
s
: s t}, that is to say,
by all sets {X
s
= i} for s t and i E I. We say that a random variable T
with values in [0,00] is a stopping time of if {T t} E F
t
for all
t O. Note that this certainly implies
{T<t}=U{T::;tl/n}EF
t
forall
n
We define for stopping times T
F
T
= {A E F : A n {T t} E F
t
for all t O}.
This turns out to be the correct way to make precise the notion of sets
which 'depend only on {X
t
: t T}'.
Lemma 6.5.1. Let Sand T be stopping times of Then both X
T
and {S T} are FTmeasurable.
Proof. Since is rightcontinuous, on {T < t} there exists an n 0
such that for all m n, for some k 1, (k  1)2
m
T < k2
m
t and
X
k2
rn = X
T
. Hence
so X
T
is FTmeasurable.
We have
U
so {S > T} E Fr, and so {S T} E Fr· D
226 6. Appendix: probability and measure
Lemma 6.5.2. For all m 0, the jump time J
m
is a stopping time of
Proof. Obviously, Jo = 0 is a stopping time. Assume inductively that J
m
is a stopping time. Then
{J
m
+! t} = U {J
m
s} n {X
s
=1= X
J71
J E F
t
for all t 0, so J
m
+
1
is a stopping time and the induction proceeds. D
We denote by Qm the aalgebra generated by yo,··· ,Y
m
and 8
1
, ... ,8
m
,
that is, by events of the form {Y
k
= i} for k m and i E I or of the form
{8k > s} for k m and s > o.
Lemma 6.5.3. Let T be a stopping time of and let A EFT.
Then for all m 0 there exist a random variable Tm and a set Am, both
measurable with respect to Qm, such that T = T
m
and lA = lA
Tn
on
{T < J
m
+
1
}.
Proof. Fix t 0 and consider
Since Qm is a aalgebra, so is At. For s t we have
{X
s
= i} n {t < J
m
+1}
= (D\Yk =i,Jk s < Jk+l}U{Ym =i,Jm s}) n{t< Jm+d
k=O
so {X
s
= i} E At. Since these sets generate F
t
, this implies that At = Ft·
For T a stopping time and A E FT we have B(t) := {T :s; t} E F
t
and
A(t) := An {T t} E F
t
for all t O. So we can find Bm(t), Am(t) E Qm
such that
B(t) n {T < J
m
+
1
} = Bm(t) n {T < J
m
+1},
A(t) n {T < J
m
+
1
} = Am(t) n {T < J
m
+
1
}.
Set
Am = UAm(t)
tEQ
then T
m
and Am are Qmmeasurable and
Tml{T<J
Tn
+l} = suptlB
Tn
(t)n{T<J
Tn
+l}
tEQ
= (sup l{T<J
Tn
+l} = Tl{T<J
Tn
+l}
tEQ
6.5 Stopping times and the strong Markov property 227
and
Am n {T < Jm+d = UAm(t) n {T < Jm+d
tEQ
= U(A n {T t}) n {T < Jm+d = An {T < Jm+d
tEQ
as required. D
Theorem 6.5.4 (Strong Markov property). Let be
Markov(A, Q) and let T be a stopping time of Then, conditional
on T < ( and X
T
= i, is Markov(8
i
, Q) and independent of FT.
Proof On {T < (} set X
t
= X
T
+
t
and denote by the jump chain
and by the holding times of We have to show that, for all
A EFT, all io, ... ,in E I and all S1, ... ,Sn 0
IF({Yo = io, ... , Y
n
= in, 8
1
> S1, ... ,8
n
> sn} nAn {T < (} n{X
T
= i})
= lFi(Y
O
= io, ... , Y
n
= in, 8
1
> S1, ... ,8
n
> sn)
X IF(A n {T < (} n {X
T
= i}).
It suffices to prove this with {T < (} replaced by {J
m
T < J
m
+
1
}
for all m 0 and then sum over m. By Lemmas 6.5.1 and 6.5.2,
{J
m
T} n {X
T
= i} E FT so we may assume without loss of generality
that A {J
m
T} n {X
T
= i}. By Lemma 6.5.3 we can write T = T
m
and 1A = 1A
Tn
on {T < J
m
+
1
}, where T
m
and Am are Qmmeasurable.
. 8
1
: 82
:c: .:c: .:
o
8m +1 :
i
T
On {J
m
T < J
m
+
1
} we have, as shown in the diagram
228 6. Appendix: probability and measure
Now, conditional on Y
m
= i, 8
m
+
1
is independent of gm and hence of
T
m
 J
m
and Am and, by the memoryless property of the exponential
Hence, by the Markov property of the jump chain
P({Yo = io, ,Y
n
= in,
8
1
> 81, ,8
n
> 8
n
} nAn {J
m
~ T < J
m
+
1
} n {X
T
= i})
= p({Y
m
= io,· .. , Y
m
+
n
= in, 8
m
+
1
> 81 +(T
m
 J
m
),
8
m
+
2
> 82, ,8
m
+
n
> 8 n} n Am n {8m+
1
> T
m
 J
m
})
= Pi(Y
o
= io, ,Y
n
= in,
8
1
> 81, ,8
n
> 8
n
)P(A n {J
m
~ T < J
m
+
1
} n {X
T
= i})
as required. D
6.6 Uniqueness of probabilities and independence of aalgebras
For both discretetime and continuoustime Markov chains we have given
definitions which specify the probabilities of certain events determined by
the process. From these specified probabilities we have often deduced ex
plicitly the values of other probabilities, for example hitting probabilities.
In this section we shall show, in measuretheoretic terms, that our defini
tions determine the probabilities of all events depending on the process.
The constructive approach we have taken should make this seem obvious,
but it is illuminating to see what has to be done.
Let 0 be a set. A 7rsystem A on 0 is a collection of subsets of 0 which
is closed under finite intersections; thus
We denote as usual by a(A) the aalgebra generated by A. If a(A) = :F
we say that A generates :F.
Theorem 6.6.1. Let (O,:F) be a measurable space. Let PI and P
2
be
probability measures on (0, F) which agree on a 7rsystem A generating :F.
Then PI = P
2
·
Proof. Consider
6.6 Uniqueness of probabilities and independence of aalgebras 229
We have assumed that A ~ V. Moreover, since PI and P
2
are probability
measures, V has the following properties:
(i) 0 E V;
(ii) (A, B E V and A ~ B) :::} B\A E V;
(iii) (An E V, An i A) :::} A E V.
Any collection of subsets having these properties is called a dsystem. Since
A generates F, the result now follows from the following lemma. D
Lemma 6.6.2 (Dynkin's 1rsystem lemma). Let A be a 1rsystem and
let V be a dsystem. Suppose A ~ V. Then a(A) ~ V.
Proof. Any intersection of dsystems is again a dsystem, so we may without
loss assume that V is the smallest dsystem containing A. You may easily
check that any dsystem which is also a 1rsystem is necessarily a aalgebra,
so it suffices to show V is a 1rsystem. This we do in two stages.
Consider first
VI = {A E V : A n B E V for all B E A}.
Since A is a 1rsystem, A ~ VI. You may easily check that VI is adsystem
 because V is a dsystem. Since V is the smallest dsystem containing A,
this shows VI = V.
Next consider
V
2
= {A E V : A n B E V for all B E V}.
Since VI = V, A ~ V
2
. You can easily check that V
2
is also adsystem.
Hence also V
2
= V. But this shows V is a 1rsystem. D
The notion of independence used in advanced probability is the indepen
dence of aalgebras. Suppose that (0, F, P) is a probability space and F
1
and F
2
are subaalgebras of F. We say that Fl and F
2
are independent if
The usual means of establishing such independence is the following corollary
of Theorem 6.6.1.
Theorem 6.6.3. Let Al be a 1rsystem generating F
1
and let A2 be a
1rsystem generating F2. Suppose that
Then Fl and F
2
are independent.
230 6. Appendix: probability and measure
Proof. There are two steps. First fix A
2
E A
2
with P(A
2
) > 0 and consider
the probability measure
We have assumed that P(A) = P(A) for all A E AI, so, by Theorem 6.6.1,
jp> = P on :Fl. Next fix Al E :F1 with P(A
1
) > 0 and consider the probability
measure
We showed in first step that P(A) = P(A) for all A E A2, so, by
Theorem 6.6.1, P = P on :F
2
. Hence:F1 and :F
2
are independent. D
We now review some points in the main text where Theorems 6.6.1 and
6.6.3 are relevant.
In Theorem 1.1.1 we showed that our definition of a discretetime Markov
chain with initial distribution ,\ and transition matrix P deter
mines the probabilities of all events of the form
But subsequently we made explicit calculations for probabilities of events
which were not of this form  such as the event that visits a set
of states A. We note now that the events {X
o
= io, ... ,X
n
= in} form a
1rsystem which generates the aalgebra a(X
n
: n 0). Hence, by Theorem
6.6.1, our definition determines (in principle) the probabilities of all events
in this aalgebra.
In our general discussion of continuoustime random processes in Section
2.2 we claimed that for a rightcontinuous process the probabilities
of events of the form
for all n 0 determined the probabilities of all events depending on
Now events of the form {X
to
= io, ... ,X
tn
= in} form a 1rsystem which
generates the aalgebra a(X
t
: t 0). So Theorem 6.6.1 justifies (a precise
version) of this claim. The point about rightcontinuity is that without
such an assumption an event such as
{X
t
= i for some t > O}
which might reasonably be considered to depend on is not nec
essarily measurable with respect to a(X
t
: t 0). An argument given in
6.6 Uniqueness of probabilities and independence of aalgebras 231
Section 2.2 shows that this event is measurable in the rightcontinuous case.
We conclude that, without some assumption like rightcontinuity, general
continuoustime processes are unreasonable.
Consider now the method of describing a minimal rightcontinuous pro
cess via its jump process and holding times Let
us take F = a(X
t
: t 0). Then Lemmas 6.5.1 and 6.5.2 show that
and are Fmeasurable. Thus 9 F where
9 =
On the other hand, for all i E I
{X
t
= i} = U{I
n
t < In+d n {Y
n
= i} E g,
so also F C g.
A useful1rsystem generating 9 is given by sets of the form
B = {Yo = i
o
, .. · ,Y
n
= i
n
,S1 > S1,·.· ,Sn > sn}.
Our jump chain/holding time definition of the continuoustime chain
with initial distribution .x and generator matrix Q may be read
as stating that, for such events
l1J)(B)  \ . I'fr.. I'fr. • eqio81 eqin18n
C  A'tO"'tO'tl ••• "'tnl't
n
••• •
Then, by Theorem 6.6.1, this definition determines JP> on 9 and hence on F.
Finally, we consider the strong Markov property, Theorem 6.5.4. Assume
that is Markov(.x, Q) and that T is a stopping time of On
the set n= {T < (} define X
t
= X
T
+
t
and let j = a(X
t
: t 0); write
and for the jump chain and holding times of and
set
9=
Thus F and 9are aalgebras on n, and coincide by the same argument as
for F = g. Set
B= {Yo = i
o
, ... ,Y
n
= i
n
,S1 > S1, .. · ,Sn > sn}.
Then the conclusion of the strong Markov property states that
JP>(B I T < (, X
T
= i) = lPi(B)
with B as above, and that
JP>(C n A IT < (,XT = i) = JP>(C IT < (,X
t
= i)JP>(A IT < (,XT = i)
for all C E F and A E FT. By 6.6.3 it suffices to prove the
independence assertion for the case C = B, which is what we did in the
proof of Theorem 6.5.4.
Further reading
We gather here the references for further reading which have appeared in
the text. This may provide a number of starting points for your exploration
of the vast literature on Markov processes and their applications.
J. Besag, P. Green, D. Higdon and K. Mengersen, Bayesian computation
and stochastic systems, Statistical Science 10 (1) (1995), 340.
K.L. Chung, Markov Chains with Stationary Transition Probabilities,
Springer, Berlin, 2nd edition, 1967.
P.G. Doyle and J.L. Snell, Random Walks and Electrical Networks, Carus
Mathematical Monographs 22, Mathematical Association of America,
1984.
W.J. Ewens, Mathematical Population Genetics, Springer, Berlin, 1979.
D. Freedman, Markov Chains, HoldenDay, San Francisco, 1971.
W.R. Gilks, S. Richardson and D.J. Spiegelhalter, Markov Chain Monte
Carlo in Practice, Chapman and Hall, London, 1996.
T.E. Harris, The Theory of Branching Processes, Dover, New York, 1989.
F.P. Kelly, Reversibility and Stochastic Networks, Wiley, Chichester, 1978.
D. Revuz, Markov Chains, NorthHolland, Amsterdam, 1984.
B.D. Ripley, Stochastic Simulation, Wiley, Chichester, 1987.
L.C.G. Rogers and D. Williams, Diffusions, Markov Processes and Martin
gales, Vol 1: Foundations, Wiley, Chichester, 2nd edition, 1994.
S.M. Ross, Applied Probability Models with Optimization Applications,
HoldenDay, San Francisco, 1970.
Further reading 233
D.W. Stroock, Probability Theory  An Analytic View, Cambridge Univer
sity Press, 1993.
H.C. Tijms, Stochastic Models  an Algorithmic Approach, Wiley, Chich
ester, 1994.
D. Williams, Probability with Martingales, Cambridge University Press,
1991.
Index
absorbing state 11, 111
absorption probability 12, 112
action 198
adapted 129
alleles 175
aperiodicity 40
average number of customers 181
backward equation 62, 96
Bayesian statistics 211
biological models 6, 9, 16, 82, 170
birth process 81
infinitesimal definition 85
jump chain/holding time definition
85
transition probability definition 85
birthanddeath chain 16
boundary 138
boundary theory for Markov chains
147
branching process 171
with immigration 179
Brownian motion 159
as limit of random walks 164
existence 161
in jRd 165
scaling invariance 165
starting from x 165
transition density 166
busy period 181
capacity 151
central limit theorem 160
charge 151
closed class 11, 111
closed migration process 184
communicating class 11, 111
conditional expectation 129
conductivity 151
continuoustime Markov chains 87
construction 89
infinitesimal definition 94
jump chain/holding time definition
94, 97
transition probability definition
94, 97
continuoustime random process 67
convergence to equilibrium 41, 121,
168
countable set 217
coupling method 41
current 151
Index
detailed balance 48, 124
discounted value function 202
distribution 1
lE, expectation 222
E(:A), exponential distribution of
parameter :A 70
e
Q
, exponential of Q 62
effective conductivity 156
electrical network 151
energy 154·
epidemic model 173
equilibrium distribution 33, 117
ergodic theorem 53, 126, 168
Erlang's formula 183
excursions 24
expectation 222
expected hitting time 12, 113
expected return time 37, 118
explosion
for birth processes 83
for continuoustime chains 90
explosion time 69
explosive Qmatrix 91
exponential distribution 70
F
n
, filtration 129
fair price 135
filtration 129
finitedimensional distributions 67
first passage decomposition 28
first passage time 19, 115
flow 151
forward equation 62, 100
for birth processes 84
for Poisson process 78
Fubini's theorem 223
discrete case 219
full conditional distributions 212
fundamental solution 145
1'1, expected time in i between visits
to j 35
GaltonWatson process 171
gambling 131
Gaussian distribution 160
generator 166
generator matrix 94, 97
235
Gibbs sampler 210
gravity 134, 169
Green matrix 144, 145
harmonic function 146
Hastings algorithm 210
hitting probability 12
hitting time 12, 111
holding times 69
I, statespace 2
infective 173
integrable 129, 222
integral form of the backward
equation 98
integral form of the forward equation
101
interarrival times 180
invariant distribution 33, 117
computation of 40
irreducibility 11, 111
Ising model 214
jump chain 69
jump matrix 87
jump times 69
last exit time 20
longrun proportion of time 53, 126
mi, expected return time 37, 118
J.t{, expected time in i between visits
to j 118
Markov(:A, P) 2
Markov(:A, Q) 94,97
Markov chain
continuoustime 88
discretetime 2
Markov chain Monte Carlo 206, 208
Markov decision process 197
expected total cost 198
expected total discounted cost 202
longrun average costs 204
Markov property 3
for birth processes 84
for continuoustime chains 93
for Poisson process 75
martingale 129, 141, 176, 204
236
associated to a Markov chain 132
associated to Brownian motion
169
matrix exponentials 105
maximum likelihood estimate 56
measure 1
memoryless property 70
Metropolis algorithm 210
minimal nonnegative solution 13
minimal process 69
monotone convergence 223
discrete case 219
Moran model 177
mutation 6, 176
nonminimal chains 103
null recurrence 37, 118
o(t),O(t), order notation 63
Ohm's law 151
open migration process 185
optional stopping theorem 130
IP, probability 222
P, transition matrix 2
[>, transition matrix of reversed
chain 47
P(t), transition semigroup 96
P(t), semigroup of reversed chain
124
p ~ j ) , nstep transition probability 5
II, jump matrix 87
1rsystem 228
Poisson process 74
infinitesimal definition 76
jump chain/holding time definition
76
transition probability definition 76
policy 198
policy improvement 201
population genetics 175
population growth 171
positive recurrence 37, 118
potential 138
associated to a Markov chain 138
associated to Brownian motion
169
Index
gravitational 134
in electrical networks 151
with discounted costs 142
potential theory 134
probability generating function 171
probability measure 222
probability space 222
Qmatrix 60
Q, generator matrix of reversed chain
124
qi, rate of leaving i 61
qij, rate of going from i to j 61
queue 179
MIGll 187
M/G/oo 191
MIMll 180
M/M/s 182
queueing network 183185
queues in series 183
random chessboard knight 50
random walk
on tl
d
29
on a graph 49
recurrence 24, 114, 167
recurrence relations 57
reflected random walks 195
reservoir model 194, 195
resolvent 146
resource management 192
restocking a warehouse 192
return probability 25
reversibility 48, 125
rightcontinuous process 67
ruin
gambler 15
insurance company 196
selective advantage 176
semigroup 96
semigroup property 62
service times 180
shopping centre 185
simple birth process 82
simulation 206
skeleton 122
statespace 1
stationary distribution 33, 117
stationary increments 76
stationary.policy 198
statistics 55, 211, 215
stochastic matrix 2
stopping time 19
strong law of large numbers 52
strong Markov property 19, 93, 227
successrun chain 38
susceptible 173
telephone exchange 183
texture parameter 216
time reversal 47, 123
transience 24, 114, 167
transition matrix 2
irreducible 11
Index 237
maximum likelihood estimate 56
transition semigroup 165
truncated Poisson distribution 183
unit mass 3
Vi(n), number of visits to i before n
53
valency 50
value function 198
weak convergence 164
Wiener process 159
Wiener's theorem 161
WrightFisher model 175
(, explosion time 69
Markov Chains
J. R. Norris University of Cambridge
~u~~u CAMBRIDGE
::: UNIVERSITY PRESS
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE
The Pitt Building, Trumpington Street, Cambridge CB2 lRP, United Kingdom
CAMBRIDGE UNIVERSITY PRESS
The Edinburgh Building, Cambridge CB2 2RU, United Kingdom 40 West 20th Street, New York, NY 100114211, USA 10 Stamford Road, Oakleigh, Melbourne 3166, Australia
© Cambridge University Press 1997
This book is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 1997 Reprinted 1998 First paperback edition 1998 Printed in the United States of America TYPeset in Computer Modem
A catalogue record for this book is available from the British Library Library of Congress CataloguinginPublication Data is available
ISBN 0521481813 hardback ISBN 0521633966 paperback
For my parents
Contents
Preface Introduction 1. Discretetime Markov chains 1.1 Definition and basic properties 1.2 Class structure 1.3 Hitting times and absorption probabilities 1.4 Strong Markov property 1.5 Recurrence and transience 1.6 Recurrence and transience of random walks 1.7 Invariant distributions 1.8 Convergence to equilibrium 1.9 Time reversal 1.10 Ergodic theorem 1.11 Appendix: recurrence relations 1.12 Appendix: asymptotics for n! 2. Continuoustime Markov chains I 2.1 Qmatrices and their exponentials 2.2 Continuoustime random processes 2.3 Some properties of the exponential distribution
ix xiii 1
1
10 12 19 24
29
33
40 47
52
57 58 60 60
67
70
5 Birth processes 2. Applications 5.3 Electrical networks 4.7 Time reversal 3.6 Jump chain and holding times 2.2 Queues and queueing networks 5.2 Potential theory 4.4 Brownian motion 5. Continuoustime Markov chains II 3.3 Hitting times and absorption probabilities 3.5 Invariant distributions 3.4 Poisson processes 2.6 Uniqueness of probabilities and independence of aalgebras Further reading Index .2 Class structure 3.8 Ergodic theorem 4.1 Markov chains in biology 5.viii Contents 2.4 Markov decision processes 5.2 Basic facts of measure theory 6.4 Recurrence and transience 3.10 Appendix: matrix exponentials 73 81 87 90 93 103 105 108 108 111 112 114 117 121 123 125 128 128 134 151 159 170 170 179 192 197 206 217 217 220 222 223 224 228 232 234 3. Appendix: probability and measure 6.3 Probability spaces and expectation 6.4 Monotone convergence and Fubini's theorem 6.1 Countable sets and countable sums 6.5 Stopping times and the strong Markov property 6.5 Markov chain Monte Carlo 6.7 Explosion 2.6 Convergence to equilibrium 3.8 Forward and backward equations 2.1 Martingales 4.3 Markov chains in resource management 5. Further theory 4.9 Nonminimal chains 2.1 Basic properties 3.
Illustrative examples introduce many of the key ideas. and is the basis of all that follows. In between. The basic pattern of Chapter 1 is repeated in Chapter 3 for continuoustime chains. making it easy to follow the development byanalogy. This makes Markov chains the first and most important examples of random processes. measure theory is not a prerequisite.Preface Markov chains are the simplest mathematical models for random phenomena evolving in time. There is a selection of exercises. Indeed. Careful proofs are given throughout. At the same time. This book is an account of the elementary theory of Markov chains. There are no strict prerequisites but it is envisaged that the reader will have taken a course in elementary probability. the class of Markov chains is rich enough to serve in many applications. with applications. and which has been tested over several years. In particular. The first half of the book is based on lecture notes for the undergraduate course. and is developed from a course taught to undergraduates for several years. which forms the basis of classwork done by the students. The material is quite straightforward and the ideas introduced permeate the whole book. the whole of the mathematical study of random processes can be regarded as a generalization in one way or another of the theory of Markov chains. You must begin here. Chapter 1 deals with the theory of discretetime Markov chains. Chapter 2 explains how to set up the theory of continuous . It was conceived as a text for advanced undergraduates or master's level students. Their simple structure makes it possible to say a great deal about their behaviour.
beginning with simple examples such as the Poisson process and chains with finite state space. Richard Gibbens. and the ergodic theorem for longrun averages. convergence to equilibrium. I believe that this restriction in scope is desirable for the greater coherence and depth it allows. mathematical genetics. We have left measure theory in the background. All the results are proved. reversibility. where we explain some of the basic notions of probability and measure used in the rest of the book and give careful proofs of the few points where measure theory is really needed. The student has the option to take Chapter 3 first. Geoffrey Grim . electrical networks and Brownian motion. exploiting to the full the probabilistic viewpoint. Chapter 5 is devoted to applications. in the context of elementary Markov chains.I inherited parts of it from Martin Barlow and Chris Rogers. Markov decision processes and Monte Carlo simulation. we use excursions and the strong Markov property to obtain conditions for recurrence and transience. to study the properties of continuoustime chains before the technically more demanding construction. for example to population growth. The second half of the book comprises three independent chapters intended to complement the first half. some of the ideas crucial to the advanced study of Markov processes. Some further details are given in Chapter 6. but the proofs are intended to be rigorous. Chapter 6 is an appendix to the main text. It is a pleasure to acknowledge the work of colleagues from which I have benefitted in preparing this book. hitting probabilities. and convergence to equilibrium is proved by the coupling method. The treatment of discretetime chains in Chapter 1 includes the calculation of transition probabilities. minimal continuoustime chains. potentials. In recent years it has been given by Doug Kennedy and Colin Sparrow. when considered in measuretheoretic terms. and establish analogues of all the main results obtained for discrete time.x Preface time chains. The following paragraph is directed primarily at an instructor and assumes some familiarity with the subject. expected hitting times and invariant distributions. or very easily made rigorous. In Chapters 2 and 3 we proceed via the jump chain/holding time construction to treat all rightcontinuous. queues and networks of queues. Overall. such as martingales. In some sections the style is a little more demanding. For example. Also treated are recurrence and transience. Chapter 4 introduces. the book is more focused on the Markovian context than most other books dealing with the elementary theory of stochastic processes. No conditions of uniformly bounded rates are needed. The course on which it is based has evolved over many years and under many hands .
precision and good humour. Brian Ripley and David Williams made constructive suggestions for improvement of an early version. I am especially grateful to David Thanah at Cambridge University Press for his suggestion to write the book and for his continuing support. Cambridge.and ambiguities. Violet Lo and David Rose pointed out many typos . and to Sarah SheaSimonds who typeset the whole book with efficiency. Meena Lakshmanan.Preface xi mett. 1996 James Norris . Frank Kelly and Gareth Roberts gave expert advice at various stages.
and to compute probabilities and expected values which quantify that behaviour. .Introduction This book is about a certain sort of random process. 1. when it is usual to refer it as a Markov chain. (0). but also the lack of memory property makes it possible to predict how a Markov chain may behave. The characteristic property of this sort of process is that it retains no memory of where it has been in the past.. . We shall be concerned exclusively with the case where the process can assume only a finite or countable set of states. Examples of Markov chains abound. whereas t and s will refer to real numbers. together with many examples and applications. Such a process is called a Markov process. } t E jR+ = [0. We shall consider chains both in discrete time n E Z+ and continuous time = {O. as you will see throughout the book. The letters n. m. What makes them important is that not only do Markov chains model many phenomena of interest. k will always denote integers. In this book we shall present general techniques for the analysis of Markov chains. Thus we write (Xn)n~O for a discretetime process and (Xt)t~O for a continuoustime process. 2.. In this introduction we shall discuss a few very simple examples and preview some of the questions which the general theory will answer. This means that only the current state of the process can influence where it goes next.
Thus the density function of the waiting time T is given by for t We write T rv A ~ o.. 1 When in state 0 you wait for a random time with exponential distribution of parameter A E (0. when you get to 1 you do not stop but after another independent exponential time of parameter A jump to 2. From state 3 you move either to 1 or to 2 with equal probability 1/2. The resulting process is called the Poisson process of rate A. this does not convey any more information and we prefer to leave the loops out. otherwise stay at 2. (ii) (Continuous time) o••~.XIV Introduction Markov chains are often best described by diagrams.. and so on.. then jump to 1. and from 2 you jump to 3 with probability 1/3.. E(A) for short. We might have drawn a loop from 2 to itself with label 2/3. But since the total probability on jumping from 2 must equal 1. of which we now give some simple examples: (i) (Discrete time) 1 3 3 1 2 You move from state 1 to state 2 with probability 1.. (iii) (Continuous time) o • • A 1 • A 2 3 4 Here.. 00 ).~. .
the probability of hitting 3 is 1. {I. 2. The rules for states 1 and 2 are as given in examples (ii) and (iii). You might like to try from first principles. meaning that you cannot escape. 6} is periodic. meaning that you return again and again to every state. 6}. 5. if T 1 is the smaller you go to 1 after time T 1 . but {I. it takes on average three steps to hit 3. The states may be partitioned into communicating classes.Introduction 1 xv 3 4 2 (iv) (Continuous time) In state 3 you take two independent exponential times T1 rv E(2) and T2 rv E (4). It is a simple matter to show that the time spent in 3 is exponential of parameter 2 + 4 = 6. the probability of hitting 6 is 1/4. The class {O} is transient. (c) Starting from 1. and if T2 is the smaller you go to 2 after time T 2 . and that the probability of jumping from 3 to 1 is 2/(2 + 4) = 1/3. The class {4. The closed classes here are recurrent. namely {O}. . (d) Starting from 1. (b) Starting from 1. 3} and {4. 5. 2. Two of these classes are closed. We shall show how to establish the following facts by solving some simple linear/ equations. (a) Starting from 0. (v) (Discrete time) 3 6 2 5 o We use this example to anticipate some of the ideas discussed in detail in Chapter 1. 3} is not. The details are given later. the longrun proportion of time spent in 2 is 3/8.
Then we have: (e) lim POI n+oo = 9/32. (g) lim pg4 = 1/124.xvi Introduction Let us write pij for the probability starting from i of being in state j after n steps. n+oo . (f) P04 does not converge as n ~ 00.
so a large number are worked out in the text. Exercises at the end of each section are an important part of the exposition. Discretetime Markov chains are defined and their behaviour is investigated. For better orientation we now list the key theorems: these are Theorems 1. F.2 on longrun averages.3 characterizing recurrence and transience.3. Theorem 1. Suppose we set Ai = IF(X = i) = IF( {w : X(w) = i}).4. We work'throughout with a probability space (0. Recall that a random variable X with values in I is a function X : 0 + I.7 on invariant distributions and positive recurrence.3 on convergence to equilibrium.7.10.1 Discretetime Markov chains This chapter is the foundation for all that follows. Once you understand these you will understand the basic theory.5 on hitting times. Theorem 1.5. Part of that understanding will come from familiarity with examples. lP).3. We say that A = (Ai: i E I) is a measure on I if 0 ~ Ai < 00 for all i E I.9. . then we call A a distribution. Theorem 1. Theorem 1.1 Definition and basic properties Let I be a countable set. Each i E I is called a state and I is called the statespace. 1. Theorem 1. and Theorem 1. If in addition the total mass EiEI Ai equals 1.8.3 on reversibility.2 on the strong Markov property.2 and 1.
(ii) for n ~ 0.2 1. . It is in terms of properties (i) and (ii) that most realworld examples are seen to be Markov chains. P) for short.. (i) P (X 0 == io) == Aio. the distribution of X. . and it is the key to some later calculations.1. . then we again say (Xn)O~n~N is Markov (A. But mathematically the... . X n + 1 has distribution (Pij : j E I) and is independent of X o. P). .in+l E I. conditional on X n == i.iN E I (Xn)O<n<N is .Xn == in) == Pi n i n +1 • We say that (Xn)n~O is Markov (A.1.1 .·· . A discretetime random process Markov(A... We say that (Xn)n~O is a Markov chain wit'h initial distribution A and transition matrix P if (i) X o has distribution A. . We think of X as modelling a random state which takes the value i with probability Ai. We say that a matrix P == (Pij : i.Xn . (ii) P(Xn + 1 == i n + 1 I X o == io. .1. There is a onetoone correspondence between stochastic matrices P and the sort of diagrams described in the Introduction. More explicitly. Here are two examples: P (Ia (3 1 ~ /3) 1 p= ( ~ 1/2 1 1/2 0 1/2 1~2 ) 3 1 2 2 We shall now formalize the rules for a Markov chain by a definition in terms of the corresponding matrix P. for n 2: 0 and io.. these conditions state that.N . . Theorem 1. P) if and only if for all io. following result appears to give a more comprehensive description. Discretetime Markov chains Then A defines a distribution... j E I) is stochastic if every row (Pij : j E I) is a distribution. If (Xn)O~n~N is a finite sequence of random variables satisfying (i) and (ii) for n == 0. There is a brief review of some basic facts about countable sets and probability spaces in Chapter 6..
. .. for n 0.2) then the result follows by Theorem 1.1... . by induction for all n = 0...I = iNI) On the other hand.Xn = in) So (Xn)O~n~N is Markov(A.Xm = 8iirnPirnirn+l . .2 (Markov property). P(Xn+ 1 = i n+ 1 I X o = io. if (1. ·Pirn+nlirn+nlP(A I Xm = i) (1. .N . . P).. First consider the case of elementaryevents A = {X o = i o. . In particular.. . We have to show that for any event A determined by we have lP({Xm = im.. then P(Xo = iO. (Xm+n)n~O is Markov(8 i .Xm = i m }. ..1. . P) and is independent of the random variables X o. .Xn = in) = P(Xo = i o.XN. where 8ij ={ ° I ifi=j otherwise. P(Xo = io) = Aio and. X n+1 = in+I)/P(XO = i o. .1. D The next result reinforces the idea that Markov chains have no memory.Xn = in. Theorem 1. . lP(XN = iN I X o = io.1) holds for N.1 Definition and basic properties 3 Proof. . .1.1 and.1. Then.. Let (Xn)n~O be Markov(A. . . We write 8i = (8 ij : j E I) for the unit mass at i.1. P).1..Xm+n = im+n } n A I X m = i) Xo~.N. then by summing both sides over iN E I and using EjEI Pij = 1 we see that (1. Proof.. .XI = i l .· . . . P).. Suppose (Xn)O~n~N is Markov(A.. conditional on X m = i.Xm ..1) holds for N ..XN = iN) = P(Xo = iO)P(XI = il I X o = io) ....··· .· ..
.Xm+n = i m+n and i = im)/P(Xm = i) = biirnPi rn irn+l X Pirn+nl i rn + n lP(Xo = io.. In the case where Ai > we shall write Pi(A) for the conditional probability P(A I X o = i).2. . Then. (ii) lP\(Xn = j) = JP(Xn + m = j I X m = i) = p~j) · . . defining a new measure AP and a new matrix p 2 by (AP)j = L iEI AiPij. ... D The remainder of this section addresses the following problem: what is the probability that after n steps our Markov chain is in a given state~ First we shall see how the problem reduces to calculating entries in the nth power of the transition matrix. (p2)ik = LPijPjk. matrix multiplication is a familiar operation.1. . So the behaviour of (Xn)n~O under lPi does not depend on A. Discretetime Markov chains In that case we have to show P(Xo = i o.N. P). P). Then we shall look at some examples where this may be done explicitly.. In general. We regard distributions and measures A as row vectors whose components are indexed by I. k=l 00 Then the desired identity (1.4 1. For these objects. .Xm may be written as a countable disjoint union of elementary events A= U Ak. m ~ 0. under lPi. We agree that pO is the identity matrix I. We write p~j) = (pn)ij for the (i.2) for A follows by summing up the corresponding identities for A k . We extend matrix multiplication to the general case in the obvious way. Theorem 1.Xm = i m and i = im)/lP(Xm = i) which is true by Theorem 1. j) entry in pn. The context will make it clear when I refers to the statespace and when to the identity matrix. for all n. then A will be an Nvector and P an N x Nmatrix.3. just as P is a matrix whose entries are indexed by I x I. . Let (Xn)n~O ° be Markov(A.1. jEI We define pn similarly for any n. where (I)ij = 8ij . (Xn)n~O is Markov(8i . When I is finite we will often label the states 1.· . any event A determined by X o. By the Markov property at time m = 0.1. (i) P(Xn = j) = (Apn)j.
1 Definition and basic properties Proof. L P(Xo = i o.a . L L . P).(3) PII + {3 . The following examples give some methods for calculating p~j).I = inI. .1. conditional on X m (8i . so by eliminating pi~) we get a recurrence relation for pi~): PII (n+I)  (1 .1 5 P(Xn = j) = = ioEI L ·.. (i) By Theorem 1. (n) We also know that pi~) + pi~) = IP\ (Xn = 1 or 2) = 1. Example 1.1.4 The most general twostate chain has transition matrix of the form p = ( Ia {3 and is represented by the following diagram: {3 We exploit the relation p n+ I = pn P to write PII (n+I) _  PI2 (n){3 + PII (1 _ a ) .11): (n) PII = {_(3_ + a + {3 1 _a_(l_ a a + {3 (3)n for a for a + {3 > 0 + {3 = O. X n = j) AioPioil" · PinIi = (Apnk inlEI ioEI inlEI (ii) By the Markov property.1.. (n) This has a unique solution (see Section 1. D = i.. (Xm+n)n~O is Markov In light of this theorem we call p~j) the nstep transition probability from i to j.Xn. so we just take A = 8i in (i).··· .
Discretetime Markov chains Example 1. + (1. Example 1. and a transition from another state to the initial state with probability a/(N . in this example there is a much simpler approach. In fact. which relies on exploiting the symmetry present in the mutation rules. Pij = a / (N  1) for i =I j. with N x N transition matrix P given by Pii = 1 .1) and by putting (3 = a/(N . Then the answer we want would be found by computing pi~).1). or with probability a mutates to another strain.!. At any time a transition is made from the initial state to another with probability a. this sort of lumping together of states may not produce a Markov chain.1) in Example 1.a.6 1. Thus we have a twostate chain with diagram a/(N .~)n N N N1· Beware that in examples having less symmetry. What is the probability that the strain in the nth generation is the same as that in the Oth? We could model this process as an Nstate chain.) (1. which is chosen at random.5 (Virus mutation) Suppose a virus can exist in N different strains and in each generation either stays the same.6 Consider the threestate chain with diagram 1 3 2 1 2 .4 we find that the desired probability is !.1.1.1.
The eigenvalues are 1. (3 aDd~. band c. First we compute the eigenvalues of P by writing down its characteristic equation o = det (x  P) = x(x  ~)2 .~ + 1).  1(3 4 . for some invertible matrix U we have i/2 and hence ~ ) U 1 p n = U 1 (00 (i/~t o ~) U(i/2)n 2 1 which forces pi~) to have the form claimed. i/2 and from this we deduce that pi~) has the form Pu (n) =a+b 2 .)n + c (")n ~ ~ for some constants a. P is diagonalizable.2 ( . i/2. (3 and ~: 1= pi~) = Q p(2) 11  o = pii) O + (3 = + ~l' Q r\J L. (The justification comes from linear algebra: having distinct eigenvalues.1 Definition and basic properties and transition matrix 7 p= (~ ~ I)· = ~(x . The first few values of pi~) are easy to write down. so we get equations to solve for Q.) The answer we want is real and (±~)n = (~)ne±in~/2= (~)n (cosn 7r ±isin n 7r ) 2 so it makes sense to rewrite pi~) in the form for constants Q. that is.1)(4x 2 The problem is to find a general formula for pi~).(.1.
+ aMAM for some constants al. Discretetime Markov chains = 1/5. If an eigenvalue A is repeated (once. Can all Markov chains be realized in this way? How would you simulate a Markov chain using a computer? . Exercises 1. . P). 1.) for any Mstate chain and any states i and j. • •• be disjoint events with U~l B n = O. be a sequence of independent random variables. 1. (b) the conditional distribution of X given Y = y is independent of y. .8 SO Q: 1. Let YI .1. ~ = 2/5 and More generally..1. Suppose we are given a function G : I x [0.3 Let X o be a random variable with values in a countable set I. (iii) As roots of a polynomial with real coefficients.. p k ). Deduce that if X and Yare discrete random variables then the following are equivalent: (a) X and Yare independent. . If Y n = Xkn. . the following method may in principle be used to find a formula for p~.aM (depending on i and j). . say) then the general form includes the term (an+b)A n . (i) Compute the eigenvalues AI.2 Suppose that (Xn)n~O is Markov (A..) has the form Pij (n) _  alAI \n \n + . 1]. as in the example.. Show that if A is another event and P(AIBn ) = P for all n then P(A) = p. uniformly distributed on [0.. complex eigenvalues will come in conjugate pairs and these are best written using sine and cosine. 1] and define inductively + I Show that (Xn)n~O is a Markov chain and express its transition matrix P in terms of G. . (3 = 4/5.1 Let B I . show that (Yn)n~O is Markov (A. .AM of P by solving the characteristic equation. Y2 .. (ii) If the eigenvalues are distinct then p~.1. B 2 .
in state 2 it remembers and chooses A but may forget again. with all jumps equally likely. are independent. Sn = Z1 + . A second flea also hops about on the vertices of a triangle. What is the probability that after n hops this second flea is back where it started? [Recall that e±i7r/6 = V3/2 ± i/2. Find the probability that after n hops the flea is back where it started.1. In each of the following cases determine whether (Xn)n~O is a Markov chain: (a) X n = Zn. In the cases where (Xn)n~O is a Markov chain find its statespace and transition matrix. B by being given repeated trials in which it is shown both and is rewarded with food if it chooses A. After each tr·ial it may change its state of mind according to the transition matrix State 1 State 2 State 3 ~ ~ 0 5 12 ~ 0 l2 0 1. identically distributed random variables such that Zi = 1 with probability p and Zi = 0 with probability 1 .1 = k) is not independent of k. . + Zn. Set So = 0..... By considering the relationship between the two dice find the value of p for the new die. (b) X n = Sn. in state 3 it remembers and chooses A and never forgets. and in the cases where it is not a Markov chain give an example where P(Xn+1 = ilXn = j. all other scores having equal probability. So + . The octopus may be in one of three states of mind: in state 1 it cannot remember which object is rewarded and is equally likely to choose either. 1.4 A flea hops about at random on the vertices of a triangle. It is in state 1 before the first trial. + Sn. What is the probablity that it is in state 1 just before the (n+l)th trial? What is the probability Pn+1 (A) that it chooses A on the (n + 1)th trial ? .6 An octopus is trained to choose object A from a pair of objects A. .. but this flea is twice as likely to jump clockwise as anticlockwise.1. X n. all other scores having probability 1/5. If the first score fS 6. (c) X n = So + . + Sn).. 1.1.] 1.p. (d)X n = (Sn.1.1 Definition and basic properties 9 Suppose now that Zo.. what is the probability p that the nth score is 6? What is the probability that the nth score is I? Suppose now that a new die is produced which cannot score one greater (mod 6) than the preceding score. Z1.5 A die is 'fixed' so that each time it is rolled the score cannot be the same as the preceding score.
and which together give an understanding of the whole..il. Discuss. lPi(Xn =j for some n ~ 0) ::.10 1.2. For distinct states i and j the following are equivalent: (i) i (ii) ~ j..2 Class structure = 1/16. P i oi l P i li2 . L p~j) n=O 00 which proves the equivalence of (i) and (iii). .inl Pii 1 P i li2 . It is sometimes possible to break a Markov chain into smaller pieces. (iii) p~j) > 0 for some n ~ O. .. Observe that p~j) ::. Discretetime Markov chains Someone suggests that the record of successive choices (a sequence of As and Bs) might arise from a twostate Markov chain with constant transition probabilities.3} with transition matrix 1 2/3 1. Proof. Calculate P(Xn = 11Xo = 1) in each of the following cases: (a) p (b) p = 1/6. Also p~j) = L il . This is done by identifying the communicating classes of the chain. 1. 1.. We say that i leads to j and write i ~ j if ~ Pi(Xn = j for some n We say i communicates with j and write i 0) > o.. with reference to the value of Pn+ 1 (A) that you have found.in with io = i and in = j..7 Let (Xn)n~O be a Markov chain on {1. 0 . Pinli n > 0 for some states io. (c) p = 1/12.1.1. ~ ~ j if both i j and j ~ i. Theorem 1.P 0 ) 1/3 0 .2. . ·Pind so that (ii) and (iii) are equivalent... each of which is relatively easy to understand. whether this is possible.
Example 1. So ~ satisfies the conditions for an equivalence relation on I. and thus partitions I into communicating classes.6} being closed. when one can draw the diagram.3 Hitting times and absorption probabilities 11 It is clear from (ii) that i ~ j and j ~ k imply i ~ k.2. with only {5.6}. Exercises 1.1 Identify the communicating classes of the following transition matrix: 2 0 0 0 2 1 1 0 2 0 :4 0 1 1 P= 0 0 1 :4 0 1 0 2 0 :4 0 1 1 2 0 0 :4 1 2 1 1 Which classes are closed? .2.3}.2 Find the communicating classes associated to the stochastic matrix 2 0 1 P= 3 1 0 0 0 2 0 0 0 0 0 1 0 1 0 0 0 0 0 0 3 1 1 0 0 3 1 1 2 0 0 2 0 1 0 0 0 0 1 0 The solution is obvious from the diagram 1 4 2 6 the classes being {1. {4} and {5. As the following example makes clear. A chain or transition matrix P where I is a single class is called irreducible. We say that a class C is closed if i E C. Also i ~ i for any state i. i ~ j imply j E C. A state i is absorbing if {i} is a closed class.1. Thus a closed class is one from which there is no escape.2. the class structure of a chain is very easy to find. The smaller pieces referred to above are these communicating classes.
Find an example of a transition matrix with no closed communicating class. Before we give the general theory.1. these quantities can be calculated explicitly by means of certain linear equations associated with the transition matrix P. The mean time taken for (Xn)n~O to reach A is given by kt = lEi(H A ) = 2: nJP>(H n<oo A = n) + ooJP>(H A = (0). .3 Hitting times and absorption probabilities Let (Xn)n>O be a Markov chain with transition matrix P.3. hf is called the absorption probability.. The hitting time of a subset A of I is the random variable H A : n ~ {O. what is the probability of absorption in 4? How long does it take until the chain is absorbed in 1 or 4? Introduce .12 1. Discretetime Markov chains 1.2. Example 1.1 Consider the chain with the following diagram: 1 1 1 • 1 2 E 2 • • 2 41( 2 1 3 • 2 ~ 4 • Starting from 2..2 Show that every transition matrix on a finite statespace has at least one closed communicating class.2. The probability When A is a closed class. here is a simple example. 1. We shall often write less formally Remarkably. } U {oo} given by HA(w) = inf{n ~ 0 : Xn(w) E A} where we agree that the infimum of the empty set 0 is starting from i that (Xn)n~O ever hits A is then 00.
then H A ~ 1. and consider the situation after making one step.2. the probability of hitting 4 is 1/3 and the mean time to absorption is 2. A. Here is a general result for hitting probabilities.3 Hitting times and absorption probabilities 13 Clearly. (1. The vector of hitting probabilities h A = (hf : i E I) is the minimal nonnegative solution to the system of linear equations hf = 1 { hf = EjEI Pij h (Xi: for i E A 1 for i fj. A. If X o = i fj.3). in assuming that the chain begins afresh from its new position after the first jump. With probability 1/2 we jump to 1 and with probability 1/2 we jump to 3. First we show that h A satisfies (1. Hence h2 = ~ h 3 = ~ ( ~ h 2 + ~). So. Note that in writing down the first equations for h2 and k2 we made implicit use of the Markov property. If X o = i E A. k2 = 1 + ~k3 = 1 + ~(1 + ~k2).3) Xi ~ (Minimality means that if x = for all i.1. starting from 2. Theorem 1. h4 = 1 and k l = k4 = o. and hf = lP\(H A < 00) = LJPi(H A < oo. then Xi ~ hi for all i. jEI jEI A .) i E I) is another solution with 0 Proof.XI = j) jEI = LJPi(H < 00 I Xl = j)JPi(XI = j) = LPijhf. So The 1 appears in the second formula because we count the time for the first step. hI = 0.3. then H A so hf = 1. Similarly. Suppose now that we start at 2. so by the Markov property = 0.
so is the last term on the right.3)...3) for h = h{4} are given here by = 1. Of course. then Xi Substitute for Xi Xj = Xi = 1 = '2:PijXj jEI = '2:Pij jEA + '2:PijXj.. X n E A) + L jl~A .. So Xi ~ JP>i(H A ~ n) for all n and then Xi ~ lim JP>i(H A ~ n) n+oo = JP>i(H A < 00) = hi. so we recover h 2 = 1/3 as before. Then hf for i E A. + JP>i(XI jn~A fj.. L PiiIPjli2 · · · Pjnlin Xjn' Now if X is nonnegative. Discretetime Markov chains Suppose now that X = (Xi: i E I) is any solution to (1.14 1.3). A. j~A to obtain = LPij jEA + LPij(LPjk + j~A LPjkXk) k~A kEA = JPli(X1 E A) + JPli(X1 ¢ A. . A.Xn I + . Suppose i ¢ A. h2 = ~hl + ~h3.3..1 (continued) The system of linear equations (1. h4 so that h3 = ~h2 + ~h4 and The value of hI is not determined by the system (1. X 2 X E A) + '2: '2: PijPjkXk· j~A k~A By repeated substitution for Xi = JP>i(XI E A) in the final term we obtain after n steps fj. but the minimality condition now makes us take hI = 0. the extra boundary condition hI = 0 was obvious from the beginning . and the remaining terms sum to JP>i(H A ~ n). D Example 1. .
then h is the minimal nonnegative solution to h o = 1.3 Hitting times and absorption probabilities 15 so we built it into our system of equations and did not have to worry about minimal nonnegative solutions. hi = ph i+1 + qhi . But what is the probability that you leave broke? Set hi = IPi(hit 0). so the minimal nonq the recurrence relation .3. then since h o = 1 we get a family of solutions for a nonnegative solution we must have A negative solution is hi = (q / p) i..) If P < q. £1 at a time. Example 1.1.. Pi. If p > q.2.2..11.il = q. so hi = 1 for all i. q 1( p o • I( 1 i i +1 •• where 0 < P = 1 . (See Section 1.. .. if p has a general solution hi = A+ Bi ~ = 0.q < 1.3 (Gamblers' ruin) Consider the Markov chain with diagram q p • . then the restriction 0 ~ hi ~ 1 forces B = 0. q I( p .1 .. the minimality condition is essential. . Finally.. The resources of the casino are regarded as infinite. which is the case in most successful casinos.. If p =I q this recurrence relation has a general solution hi = A +B (~) i . Pi.i+l = P Imagine that you enter a casino with a fortune of £i and gamble. In cases where the statespace is infinite it may not be possible to write down a corresponding extra boundary condition. for i = 1. The transition probabilities are Poo = 1. as we shall see in the next examples. so there is no upper limit to your fortune.. Then. for i = 1. with probability P of doubling your stake and probability q of losing it.
. Then hi = IPi(hit 0) is the extinction probability starting from i. Thus..hi. But here we allow Pi and qi to depend on i....... 0 is an absorbing state and we wish to calculate the absorption probability starting from i....qi < 1. the restriction 0 ~ hi ~ 1 forces A = 0 and hi = 1 for all i...2.. ...••. But consider Ui = h i . .t(. This recurrence relation has variable coefficients so the usual technique fails. Discretetime Markov chains and again the restriction 0 ~ hi ~ 1 forces B = 0. In the case L::o ~i = 00. Such a chain may serve as a model for the size of a population..I . you are certain to end up broke...2. for i = 1.~ ql PI qi Pi . As in the preceding example. recorded each time it changes. then PiUi+1 = qiUi. But if L::o ~i < 00 then we can take A > 0 so long as lA(~o+ .. At this point A remains to be determined.3....• t i+l qi+1 Pi+1 where.16 1. . .. so hi = 1 for all i. even if you find a fair casino... so where the final equality defines UI ~i. (I. Example 1. + Ui so where A = UI and ~o = 1.. This apparent paradox is called gamblers' ruin.... Then = h o . we have 0 < Pi = 1 .4 (Birthanddeath chain) Consider the Markov chain with diagram o ... hi = Pihi+1 + qihiI....t(~ 1 i . .... Pi being the probability that we get a birth before a death in a population of size i. We write down the usual system of equations ho = 1. . +~iI)~O for all i.hi + . for i = 1.
'PjndnYjn' . L jn~A PiilPilh . . Theorem 1. then HA so kf = o. If X o = i f/: A.2. (1. First we show that k A satisfies (1. Then for i E A. 1x1 ==j is the random variable equal to 1 if Xl = j and equal to 0 otherwise. If X o = i E A.3.1. Recall that kf = Ei(H A ). + lP'i(H A ~ n) + L jl~A .4) Proof. Here is the general result on mean hitting times. where H A is the first time (Xn)n~O hits A. so..4).. . = 0.5. If i f/: A.3 Hitting times and absorption probabilities 17 Thus the minimal nonnegative solution occurs when A = then (E:o Ii) 1 and In this case. so the population survives with positive probability. we have hi < 1. and kt = Ei(H A ) = = LEi(H JEI A L JEI E i(H A 1X1 =j) I Xl = j)JP\(XI = j) = 1 + LPijkf· j~A Suppose now that Y = (Yi : i E I) is any solution to (1.. then Yi kf = Yi = 0 = 1 + LPijYj j~A = 1 + LPij j~A (l + LPjkYk) k~A = lP'i(H A ~ 1) + lP'i(H A ~ 2) + L L PijPjkYk· j~A k~A By repeated substitution for Y in the final term we obtain after n steps Yi = lP'i(H A ~ 1) + .4). The vector of mean hitting times k A = (k A : i E I) is the minimal nonnegative solution to the system of linear equations kf { kf = 0 = 1 for i E A + Ej~A pijkf for i f/: A. by the Markov property..... for i = 1. so. then H A ~ 1. for example. We use the notation 1B for the indicator function of B.
letting n * 00.1 Prove the claims (a). if he wins. 1.3.. He can play a game with the following rules: a fair coin is tossed.3.3 A simple game of 'snakes and ladders' is played on a board of nine squares. and his stake is returned. and otllerwise stakes just enough to increase his capital. if y is nonnegative.3. The gambler decides to use a bold strategy in which he stakes all his money if he has £5 or less. . he wins a sum equal to his stake. + Wi(H A = 2: n) Yi 2': Exercises L IP\(H n==l A 2': n) JEi(H A ) = Xi D 1. otherwise he loses his stake. to £10. Yi 2: Wi(H A 2: 1) and. Discretetime Markov chains So. + .18 1. if a player bets on the right side. Pro~e that the What is the expected number of tosses until the gambler either achieves his aim or loses his capital? 1.2 A gambler has £2 and needs to increase it to £10 in a hurry. gambler will achieve his aim with probability 1/5. Let X o == 2 and let X n be his capital after n throws.. (b) and (c) made in example (v) of the Introduction.
Xn E A}.il = 1.Xn for n = 0. . after all.. } with transition probabilities given by POI = 1. so it does not simply begin afresh.3 is a stopping time because {H A = n} = {X o ~ A.4 Strong Markov property 19 At each turn a player tosses a fair coin and advances one or two places according to whether the coin lands heads or tails. .Xn l ~ A.. . you know at the time when T occurs.I? In this section we shall identify a class of random times at which a version of the Markov property does hold. . Intuitively.. Xl.. . 1. instead of conditioning on X m = i.2.. ...4 Strong Markov property In Section 1.i+l + Pi. Pi. i ~ 1.4 Let (Xn)n~O be a Markov chain on {O. conditional on X m = i.1 jumps straight to i. If asked to stop at T...3. A random variable T : n ~ {O. at some random time H. . the process after time H .2. by watching the process.. you know when to stop. for example H . 1. but if you land at the head of a snake you slide down to the tail.1. This says that for each time m.1. Show that if X o = 0 then the probability that X n ~ 1 for all n ~ 1 is 6/1T 2 • 1.1 we proved the Markov property. If you land at the foot of a ladder you climb to the top.1. This class will include H but not H .1 (a) The first passage time Tj = inf{n ~ 1 : Xn = j} is a stopping time because (b) The first hitting time H A of Section 1.il. What can one say about the process after time H? What if we replaced H by a more general random time. How many turns on average does it take to complete the game? What is the probability that a player who has reached the middle square will complete the game without slipping back to square I? 1.i+l = i + ( i 1) 2 Pi. Pi. } U {(X)} is called a stopping time if the event {T = n} depends only on X o . the process after time m begins afresh from i.4.. . Examples 1. Suppose. we simply waited for the process to hit state i.
then B n {T == m} is determined by X o. P) and let T be a stopping time of (Xn)n~O.. . ..3.Xm . . Xl.Xn = jn)JP>(B I T < 00.XT+n = jn} n BIT < 00. X T = i) = JP>i(XO = jo.3 Consider the Markov chain (Xn)n~O with diagram q p • ~ q I( • p ~ q i p o • I( I(.. conditional on T < 00 and X T = i. and divide by lP(T < 00. Xl. .4. Let (Xn)n~O be Markov(A. Example 1. X T = i). Xl. D The following example uses the strong Markov property to get more information on the hitting times of the chain considered in Example 1. Theorem 1.2 (Strong Markov property).XI .. X T+I = jl..·. . for all m = 0..Xn = jn)JP>(B n {T = m} n {XT = i}) where we have used the condition T = m to replace m by T.. then B n {T = m} is determined by X o. Proof. . Xl = jl. Discretetime Markov chains (c) The last exit time LA = sup{n ~ 0 : X n E A} is not in general a stopping time because the event {LA whether (Xn+m)m~l visits A or not. If B is an event determined by X o. . . .XT.3.. 2.. X T = i) to obtain JP>( {XT = jo. Now sum over m == 0.~ 1 i +1 . P) and independent of XO. 1. (XT+n)n~O is Markov(8 i . .. . Xl. == n} depends on We sllall show that the Markov property holds at stopping times..XT .. . . Then.4. by the Markov property at time m lP( {XT = io. The crucial point is that.X I = jl. . XT+I = il.Xm . . if T is a stopping time and B ~ n is determined by X o. so...20 1.XT+n = in} n B n {T = m} n {XT = i}) = JP>i(XO = io. 2. 1.XT .
Here we obtain the complete distribution of the time to hit 0 starting from 1 in terms of its probability generating function. the time taken after HI to get to 0.4 Strong Markov property 21 where 0 < p = 1 .4pqs2)/2ps.5) ¢ = (1 ± VI .1. where iio. To recover the distribution of H o we expand the squareroot as a power series: </>(s) = 2~S { 1  (1 + !( 4pqs 2) + !(lH _4pqs 2)2 /2! + . for 0 ~ s < 1 </>(s) = lEI (sHo) = 2: snIP (Ho = n). 2 3 . we have H o = HI + H o. ) } = qs +pq s = + .. we have H o = 1 + H o.3. where H o. Thus ¢ = ¢( s) satisfies and pS¢2 . Set H j = inf{n ~ 0 : X n = j} and. So E 2(SHO) = E 2(sHl I HI < 00)E 2(sHO I HI < 00)JP>2(Hl < 00) = E2(sHI1Hl<OO)E2(sHO I HI < 00) = E 2(sHl)2 = ¢(s)2.q < 1.. slPl(Ho = 1) + s2lP l (Ho = 2) + s3lP l (Ho = 3) + . We know from Example 1.. 1 n<oo Suppose we start at 2. Then.. Since ¢(O) ~ 1 and ¢ is continuous we are forced to take the negative root at s = 0 and stick with it for all 0 ~ s < 1. the time taken after time 1 to get to 0. conditional on Xl = 2.. So ¢(s) = El(sHo) = pEl (sHo I Xl = = pEl (sl+Ho I Xl = psE 2(sHo ) + qEl(sHo I Xl = 0) = 2) + qEl(s I Xl = 0) 2) + qs = ps¢(s)2 + qs. conditional on HI < 00.. by the Markov property at time 1. has the same distribution as H o does under lP 2 .3 the probability of hitting 0 starting from 1. is independent of HI and has the (unconditioned) distribution of HI. Apply the strong Markov property at !II to see that under lP 2..¢ + qs =0 (1.
(Remember that q = 1 p. is a stopping time. Discretetime Markov chains The first few probabilities P1(Ho = 1).4 We now consider an application of the strong Markov property to a Markov chain (Xn)n~O observed only at certain times.1rl (Ii < 00 ) _ 1 VI  4pq _ { 1 if p ~ q 2p q/p if p > q. so VI . the time of the mth visit to strong Markov property applies to show.. On letting s lI]) i 1 we have ¢(s) ~ P1(Ho < 0 00).··· . are readily checked from first principles.XTrn = i m) = Pi rn (XT 1 = i m +1 ) = p. that P(Ym +1 = i m +1 I Yo = i o. For each m we can check J.. Ym = i m ) = P(XTrn + 1 = i m + 1 I X To = i o.4p + 4p2 = 11 We can also find the mean hitting time using 2pl = 12q . where To = inf{n ~ 0 : X n E J} and. rn .1 for a connection with branching processes.22 1.) It is only worth considering the case p ~ q. 2.5) to obtain 2ps¢¢' so + p¢2  ¢' + q = 0 ¢'(s) = (p¢(S)2 + q)/(l  2ps¢(s)) ~ 1/(1 . where the mean hitting time has a chance of being finite.··· . 1.4pq = VI ..2p) = l/(q . So the i o. for m... .11. .p) as s i 1. for m = 0.im+l E J.P 1(Ho = 2). The resulting process (Ym)m~O may be obtained formally by setting Ym = X Trn . T m + 1 = inf{n > Tm : X n E J}. . In the first instance suppose that J is some subset of the statespace I and that we observe the chain only when it takes values in J. . Example 1. Differentiate (1.1. Let us assume that lP(Tm < 00) = 1 for all easily that T m.. See Example 5.4.rn +l 't 't . so .
1.4 Strong Markov property 23 where. X n = X o + Y1 + ...6) h{ = Pij + LPik h 1· k~J Thus (Ym)m~O is a Markov chain on J with transition matrix P.3.2 Deduce carefully from Theorem 1.1 Let Y1 . k=j:i Thus (Zm)m~O is a Markov chain on I with transition matrix Exercises P. Y2 .6).Xsrn = im ) = lPirn (XS1 = i m +1 ) = Pirnirn+l where Pii = 0 and. "Y2.2¢ + s = 0 .4. . Show that ¢ now satisfies s¢3 . for i.. The resulting process (Zm)m~O is given by Zm = XS rn where 8 0 = 0 and for m = 0. by the strong Markov property lP(Zm+l = im+l I Zo = io.. for j E J. = 2) = 1/2. A second example of a similar type arises if we observe the original chain (Xn)n~O only when it moves. Find the probability generating function ¢(s) = E(sHO).2 the claim made at (1.Zm = i m) = lP(XSrn + 1 = im+l I XS o = io.2. .. .. is changed to lP(Y1 lP(Y1 = 1) = 1/2. the vector (h{ : i E I) is the minimal nonnegative solution to (1. Again the random times 8 m for m ~ 0 are stopping times and. Let us assume there are no absorbing states. ••• be independent identically distributed random variables with lP(Y1 = 1) = lP(Y1 = 1) = 1/2 and set X o = 1. j E J Pii  = hi i and where. Define H o = inf{n ~ 0 : X n = O}.··· . for i =I j Pij = Pij/ LPik. + Yn for n ~ 1. 1.1. Suppose the distribution of Y1 . . 1. ...4.
We shall show that every state is either recurrent or transient. We IIOW define inductively the rth passage time Ti(r) to and. 00 The following diagram illustrates these definitions: Xn i T.24 1.(O) ~ n . for r = 0.1. .5 Recurrence and transience Let (Xn)n~O be a Markov chain with transition matrix P.2. .. We say that a state i is recurrent if lPi(Xn =i for infinitely many n) = 1. The length of the rth excursion to i is then S~r) ~ = { T(r) _ T(rI) 0 i i 1 ·f T(rI) < i otherwise. Recall that the first passage time to state i is the random variable T i defined by Ti(W) = inf{n ~ 1 : Xn(w) = i} where inf 0 = state i by 00. Thus a recurrent state is one to which you keep coming back and a transient state is one which you eventually leave for ever. We say that i is transient if IPi(Xn =i for infinitely many n) = O.. Discretetime Markov chains 1.
For r = 2. (XT+n)n~O is Markov(8i .5. Recall that the indicator function l{xl==j} is the random variable equal to 1 if Xl = j and 0 otherwise. . . So. then Pi ( Vi > r + 1) = Pi ( T i(r+I) < 00 00 ) 00) = Pi(Ti(r) < = Pi (Sfr+I) = fif. we have Pi (Vi > r) = fi. P) and independent of X o. conditional on T < 00. conditional on Ti(rI) < pendent of {Xm : m ~ Ti(rI)} and 00. Lemma 1.3.. For r = 0. But sir) so = inf{n ~ 1 : X T +n = i}.1. .2. D . Observe that if X o = i then {Vi > r} = {Ti(r) < oo}. n==O n==O Also. we can compute the distribution of Vi under Pi in terms of the return probability Lemma 1. When r the result is true. .5..2. sir) is inde Proof.. . =0 Proof. which may be written in terms of indicator functions as Vi and note that 00 = L n==O 00 l{x n =i} lEi(Vi) = lEi L n==O l{X n =i} =L n==O 00 00 00 lEi(l{Xn =i}) = LlPi(Xn = i) = LP~7). Suppose inductively that it is true for r.1. Apply the strong Markov property at the stopping time T = Ti(rI)....1.1.5. Xl.5 Recurrence and transience 25 Our analysis of recurrence and transience will rest on finding the joint distribution of these excursion lengths. D sir) is the first passage time of (XT+n)n~O to state i. and si r+ I) < < 00 I Ti(r) < oo)Pi(Ti(r) < 00) = f[+1 by Lemma 1. so by induction the result is true for all r.XT . Let us introduce the number of visits Vi to i. It is automatic that X T = i on T < 00. .
26
1. Discretetime Markov chains
Recall that one can compute the expectation of a nonnegative integervalued random variable as follows:
00 00 00
LP(V > r)
r=O
=
L
L
P(V
=
v)
00
r=Ov=r+1 00 vI
=
L LP(V
v=lr=O
="'
v) = L vP(V = v) = E(V).
v=1
The next theorem is the means by which we establish recurrence or transience for a given state. Note that it provides two criteria for this, one in terms of the return probability, ttle other in terms of the nstep transition probabilities. Both are useful.
Theorem 1.5.3. The following dichotomy holds:
(i) ifIPi(Ti < 00) = 1, then i is recurrent and E~op~~) = 00; (ii) ifIPi(Ti < 00) < 1, then i is transient and E~op~~) < 00. In particular, every state is either transient or recurrent.
Proof. If IPi(Ti < 00) = 1, then, by Lemma 1.5.2, IPi(Vi
so i is recurrent and
= 00) = r+oo IPi(Vi > r) = 1 lim
00
LP~~) = Ei(Vi)
n=O
=
00.
On the other hand, if Ii = IPi(Ti < 00) < 1, then by Lemma 1.5.2
n=O
fp~~) = Ei(Vi) = fpi(Vi > r) = fl[
r=O r=O
=
1
~
. < f'l,
00
so IPi (Vi
= 00) = 0 and i
is transient.
D
From this theorem we can go on to solve completely the problem of recurrence or transience for Markov chains with finite statespace. Some cases of infinite statespace are dealt with in the following chapter. First we show that recurrence and transience are class properties.
Theorem 1.5.4. Let C be a communicating class. Then either all states in C are transient or all are recurrent.
Proof. Take any pair of states i, j E C and suppose that i is transient. There exist n, m ~ 0 with p~j) > 0 and PJ":) > 0, and, for all r ~ 0 Pii
(n+r+m)
 Pij Pjj Pji
>
(n)
(r) (m)
1.5 Recurrence and transience
so
27
~
r==O
(r) LJPjj 
<
1
Pij Pji
(n) (m) LJPii
r==O
~
(n+r+m)
<
00
by Theorem 1.5.3. Hence j is also transient by Theorem 1.5.3.
D
In the light of this theorem it is natural to speak of a recurrent or transient class. Theorem 1.5.5. Every recurrent class is closed.
Proof. Let C be a class which is not closed. Then there exist i E C, j fj. C and m ~ 1 with
Since we have lPi ( {Xm = j} n {X n = i for infinitely many n}) = 0 this implies that
lPi(Xn = i for infinitely many n) < 1
so i is not recurrent, and so neither is C. D
Theorem 1.5.6. Every finite closed class is recurrent.
Proof. Suppose C is closed and finite and that for some i E C we have
(Xn)n~O
starts in C. Then
o < P(Xn = i
= lP(Xn
for infinitely many n) for some n)Pi(Xn
=i
=i
for infinitely many n)
by the strong Markov property. This shows that i is not transient, so C is recurrent by Theorems 1.5.3 and 1.5.4. D It is easy to spot closed classes, so the transience or recurrence of finite classes is easy to determine. For example, the only recurrent class in Example 1.2.2 is {5, 6}, the others being transient. On the other hand, infinite closed classes may be transient: see Examples 1.3.3 and 1.6.3. We shall need the following result in Section 1.8. Remember that irreducibility means that the chain can get from any state to any other, with positive probability.
28
1. Discretetime Markov chains Then for all
j E I we have lP(Tj
Theorem 1.5.7. Suppose P is irreducible and recurrent. < 00) = 1.
Proof. By the Markov property we have
lP(Tj < (0) = LlP(Xo = i)lPi(Tj < (0)
iEI
so it suffices to show lPi (Tj < (0) By Theorem 1.5.3, we have
1 = Pj(Xn
=
1 for all i E I. Choose m with
p;:n) > o.
= j for infinitely many
= j
n)
= lPj(Xn
kEI
=
for some n ~ m
= j for some n
+ 1)
+ 1 IXm
= k)lPj(Xm = k)
= L lPj(Xn
~m
L lPk(T
kEI
j
< oo)p;7:)
where the final equality uses the Markov property. But we must have Pi(Tj < 00) = 1. D
Exercises
EkE!
p;7:) =
1 so
1.5.1 In Exercise 1.2.1, which states are recurrent and which are transient? 1.5.2 Show that, for the Markov chain
lP(Xn
~
(Xn)n~O
~
in Exercise 1.3.4 we have
00 as n
00) = 1.
Suppose, instead, the transition probabilities satisfy
Pi,i+l =
(i+l)O
i
Pi,il·
For each
Q
E (0,00) find the value of P(Xn ~ 00 as n ~ 00).
1.5.3 (First passage decomposition). Denote by T j the first passage time to state j and set
Justify the identity
n
P ~~) = '"' f~~)p~~k)
I)
~
I)
))
for n
~
1
k==l
1.6 Recurrence and transience of random walks and deduce that where
29
CX)
00
Pij(s) = LP~;)sn,
F ij (s) = L
fi~n) sn ·
n==O
Hence show that Pi (Ti <
00)
n==O
= 1 if and only if
00
n==O
without using Theorem 1.5.3.
LP~~) =
00
1.5.4 A random sequence of nonnegative integers (Fn)n~o is obtained by setting Fo = 0 and F 1 = 1 and, once Fo, ... ,Fn are known, taking Fn+ 1 to be either the sum or the difference of F n  1 and F n , each with probability 1/2. Is (Fn)n~o a Markov chain? By considering the Markov chain X n = (Fn  1 , Fn ), find the probability that (Fn)n~o reaches 3 before first returning to O. Draw enough of the flow diagram for (Xn)n~O to establish a general pattern. Hence, using the strong Markov property, show that the hitting probability for (1,1), starting from (1,2), is (3  V5)/2. Deduce that (Xn)n~O is transient. Show that, moreover, with probability 1, Fn ~ 00 as n ~ 00.
1.6 Recurrence and transience of random walks
In the last section we showed that recurrence was a class property, that all recurrent classes were closed and that all finite closed classes were recurrent. So the only chains for which the question of recurrence remains interesting are irreducible with infinite statespace. Here we shall study some simple and fundamental examples of this type, making use of the following criterion for recurrence from Theorem 1.5.3: a state i is recurrent if and only if
~oo
L..Jn==O Pii 
(n)_
00.
Example 1.6.1 (Simple random walk on Z)
The simple random walk on Z has diagram
q
iI
...
..
i
P
i+l
30
1. Discretetime Markov chains
where 0 < P = 1  q < 1. Suppose we start at O. It is clear that we cannot return to 0 after an odd number of steps, so p~~n+l) = 0 for all n. Any given sequence of steps of length 2n from 0 to 0 occurs with probability pnqn, there being n steps up and n steps down, and the number of such sequences is the number of ways of choosing the n steps up from 2n. Thus
(2n) Poo = (2n) p n qn . n
Stirling's formula provides a good approximation to n! for large n: it is known that
asn~oo
where an rv bn means an/bn ~ 1. For a proof see W. Feller, An Introduction to Probability Theory and its Applications, Vol I (Wiley, New York, 3rd edition, 1968). At the end of this chapter we reproduce the argument used by Feller to show that
asn~oo
for some A E [1,00). The additional work needed to show A omitted, as this fact is unnecessary to our applications. For the nstep transition probabilities we obtain
= y'2;IT
is
Poo
(2n) _ (2n)! ( )n  (n!)2 pq
rv
(4pq)n
AJn/2
as n
~
00.
In the symmetric case p n ~ N we have
= q = 1/2, so 4pq = 1; then for some N and all
(2n) > Poo 
2AVii
_1 _
1
so ~ Poo
~
(2n)
> .!... ~
 2A ~
Vii 
00
which shows that the random walk is recurrent. On the other hand, if p then 4pq = r < 1, so by a similar argument, for some N
=I q
'"" p(n) < _ '"" r n < 00 ~ 00 A ~
n==N n=N
00
1
00
showing that the random walk is transient.
1 1 .6. Suppose we start at o.~ :4 1 :4 and transition probabilities Pij 1/4 { 0 == iflijl==l otherwise. 1 :4 ..2 (Simple symmetric random walk on Z2) The simple symmetric random walk on Z2 has diagram 31 :4 .. for the orthogonal projections of X n on the diagonal lines y == ±x: X+ n Then X~ and X.1. This makes it clear that for X n we have Poo (2n) == asn~oo ..... Let us call the walk X n and write X~ and X. are independent simple symmetric random walks on 2.6 Recurrence and transience of random walks Example 1.1 / 2 Z and X n == 0 if and only if X~ == 0 == X.
By counting the ways in which this can be done. Then and the walk is recurrent.J . ·'k')2 (~. we have n ( ij k) for all i..6. For the case where n = 3m.Jn=O n and Poo _ Poo or all m so we must have and the walk is transient. k. j.jl = 1 P'1. _ {1/6 if Ii .j.j..j.k~O i+j+k=n i+j+k=n Now i. Thus the chain jumps to each of its nearest neighbours with equal probability.k~O (2n)! .k~O i+j+k=n the lefthand side being the total probability of all the ways of placing n balls randomly into three boxes.) o otherwise. k ~ 0. so = i!j!k! ~ mmm n! (n) (~) 3/2 asn~oo by Stirling's formula.. j south. . we obtain (2n) _ Poo  i. Discretetime Markov chains by Stirling's formula. B u t Poo > (1/6)2 Poo (6m) (6m2) (6m) > (1/6)4 (6m4) £ L.j. Suppose we start at O.. . i down. Of these 2n steps there must be i up. i.00 3/2 . E::o p~~) = 00 by comparison with E::o l/n Example 1. k east and k west for some i. E:=o p~~m) < 00 by comparison with .32 1.3 (Simple symmetric random walk on Z3) The transition probabilities of the simple symmetric random walk on Z3 are given by . with i + j + k = n.". j north. We can only return to 0 after an even number 2n of steps. Hence.
7 Invariant distributions 33 Exercises 1.== n~oo 1m (n) 1m L kEI Pk Pkj ~ (n) == L kEI n~oo 1· 1m Pk Pkj ~ (n) == L kEI 7rkPkj . X m + 1 . Theorem 1..7 Invariant distributions Many of the longtime properties of Markov chains are connected with the notion of an invariant distribution or measure. Proof. Then (Xm+n)n~O is also Markov(A. 1. == (7rj : j E I) is an invariant distribution. P). Suppose for some i E I that p~j) Then 7r t 7rj as n t 00 for all j E I.P) and suppose that A is invariant for P. Show that the random walk is transient. Theorem 1. P(Xm == i) == (Apm)i == Ai for all i and. The terms equilibrium and stationary are also used to mean the same.7. Proof. The first result explains the term stationary.1 The rooted binary tree is an infinite graph T with one distinguished vertex R from which comes a single edge. Let (Xn)n~O be Markov(A. D The next result explains the term equilibrium.3. . X m+n+ 1 is independent of X m . at every other vertex there are three edges and there are no closed loops. The random walk on T jumps from a vertex along each available edge with equal probability. clearly. Let I be finite.Xm+n and has distribution (Pij : j E I).6.7. \ 1.2 Show that the simple symmetric random walk in Z4 is transient. . Remember that a measure A is any row vector (Ai : i E I) with nonnegative entries. By Theorem 1.1.. We say A is invariant if AP == A.2. conditional on X m+n == i.1.6.1. We have L and 7rj 7rj == jEI L n~oo 1· P~J1m (n)  1· == n~oo 1m L P ~J (n) == 1 jEI jEI 1· 1· == n~oo P~J.
4 Consider the Markov chain (Xn)n~O with diagram 1 3 1 2 2 To find an invariant distribution we write down the components of the vector equation 7r P = 7r 7rl 7r2 = ~7r3 1 1 = 27r1 + 2 7r3 1 27r2 1 + 2 7r3 · _ 7r3  . + (3)) must be Example 1. but it is not a distribution! Theorem 1. so. by Example 1. Example 1.6 we have p~.2 is not a very useful result but it serves to indicate a relationship between invariant distributions and nstep transition probabilities. j E I.( la (3 Ignore the trivial cases 0: = (3 = 0 and 0: = (3 = 1. by Theorem 1.0:/(0: invariant. Then. D Notice that for any of the random walks discussed in Section 1.1.3 Consider the twostate Markov chain with transition matrix p.7.34 1.7. The limit is certainly invariant. Hence 7r is an invariant distribution. In Theorem 1. the distribution ((3/(0: + (3).7. Discretetime Markov chains where we have used finiteness of I to justify interchange of summation and limit operations.2.3 we shall prove a sort of converse.7. which is much more useful.) t 0 as n t 00 for all i. There are of course easier ways to discover this.8.4 pn ~ ((3/(a (3/(a + (3) + (3) a/(a + (3)) a/(a + (3) as n ~ 00.
6 pi~) ~ 1/5 as n ~ 00 so this confirms Theorem 1.2. Tk} depends only on X o.2/5).l .7. instead of working out p~. Then (i) l'~ = 1. (i) This is obvious.2/5. n7r) + CSln 2 we could have used Theorem 1.6. knowing that p~~) had the form PII (n) = a + (l)n (bcos 2 2 n7r . so P must have a row eigenvector with eigenvalue 1. consider for each i the expected time spent in i between visits to k: Tk. when X o has distribution 7r.l ')'f = lEk L l{x n =i}' n=O Here the sum of indicator functions serves to count the number of times n at which X n = i before the first passage time T k . . the righthand sides give the probabilities for Xl. Proof. by the Markov property at n .2. and another equation is required to fix 7r uniquely. (ii) l'k = (l'f : i (iii) 0 < E Tf < 00 I) satisfies l'k P = l'k . . In the next two results we shall show that every irreducible and recurrent stochastic matrix P has an essentially unique positive invariant measure.7. The equations are homogeneous so one of them is redundant..1. Let P be irreducible and recurrent. the existence of an invariant row vector is a simple piece of linear algebra: the row sums of P are alII. so. That equation is and we find that 7r = (1/5. Theorem 1. and the equations require Xl also to have distribution 7r. . Alternatively..1. .1 . for all i E I. the event {n ::. (ii) For n = 1. The proofs rely heavily on the probabilistic interpretation so it is worth noting at the outset that. for a finite statespace I.) in Example 1. For a fixed state k.1.. so the column vector of ones is an eigenvector with eigenvalue 1. According to Example 1. Xl.5.2 and knowledge of 7r1 to identify a = 1/5.7.Xn . 7 Invariant distributions 35 In terms of the chain.
Discretetime Markov chains Tk Since P is recurrent.i n l#k Pkinl · · · PilioPioj ) ~ IPk(X1 = j and Tk ~ 1) + IPk(X2 n) = j and Tk ~ 2) + . D Theorem 1.1 } Tk.. Then ~ik ~ ~kPki > 0 and ~ik Pik ~ ~k = 1 bY () and i (ii).. (iii) Since P is irreducible. Then A ~ ~k. If in addition P is recurrent. +IPk(Xn ~ ~j as =j and n ~ 00.. Proof.. for each state i there exist n. Let P be irreducible and let A be an invariant measure for P with Ak = 1.. then A = ~k .1 = iEI n=l 00 = LpijlEk L iEI m=O l{Xtn =i and msTk.Pki >.j = lEk L n=l 00 Tk l{x n =j} = lEk L 00 l{Xn =j and nSTd n=l = L n=l JP>k(Xn 00 =j and n ~ Tk) i.il#k Ai1Pi 1 ioPioj + (Pk j PkioPiOj ) L i o . m ~ 0 with (n) (m) 0 k (m) (n) k Pik . . X n = j and n ~ Tk) i and n ~ Tk) = L L JP>k(X iEI n=l 00 n.1 = LPijlEk iEI L m=O l{Xtn =i} = L iEI .in=j:k AinPininl · · · Pioj L io#k + (Pk j + PkioPioj + ···+ Tk ~ L io. For each j E I we have Aj = = L ioEI AioPioj = L io#k AioPioj + Pkj + L io#k L io.fpij. .36 1. Therefore < 00 and X o = X Tk = k with . .7..1 = = LPij L JP>k(Xn .6. under IPk we have probability one.. .
it is certainly recurrent. (n) (n) (n) we have Pik > 0 £ some n. / mi defines an invariant distribution. given i E I. Since P is irreducible. (ii) some state i is positive recurrent. Hence mk =L k ~.1. If P is recurrent. then ~k is invariant by Theorem 1. (iii) P has an invariant distribution. so Jl = A .7. 0 Recall that a state i is recurrent if IPi(Xn = i for infinitely many n) = 1 and we showed in Theorem 1. Let P be irreducible. A ~ ~k. Then A is an invariant measure with Ak = 1. By Theorem 1. A recurrent state which fails to have this stronger property is called null recurrent.7. Proof. Set Ai = 7ri/7rk.5. 7r say. ~i is then invariant.7.3 that this is equivalent to If in addition the expected return time is finite.5. and O ' " or = /lk = LJjEI /ljPjk ~ /liPik . (iii) => (i) Take any state k. Moreover.7) and k is positive recurrent. (ii) => (iii) If i is positive recurrent.~k is also invariant and /l ~ O.7. So by Theorem 1. so P is recurrent. (i) => (ii) This is obvious.= 1 < 7r 7r k k 00 (1. ~ ~ L iEI iEI 7ri .6. when (iii) holds we have mi = 1/7ri for all i.5.7. so Jli = O. . But L~j jEI = mi < 00 so 7rj = ~. Then the following are equivalent: (i) every state is positive recurrent. then we say i is positive recurrent. 7 Invariant distributions 37 So A ~ ~k. Theorem 1. Since P is irreducible and EiEI 7ri = 1 we have 7rk = EiEI 7riP~~) > 0 for some n.
so A = ~k and the inequality (1.11 Consider a successrun chain on Z+ . Example 1. Consider the measure 7ri Then =1 for all i.i+l· In components the invariant measure equation 7r P = 7r reads This is a recurrence relation for 7r with general solution So. Example 1.7. Example 1. the simple symmetric random walk on Z3. which is transient by Example 1. so 7r is invariant.7.7. in this case. PiO = qi = 1 .Pi· .7. Since LiEZ 7ri = 00. by Example 1. 0 Example 1.7) is in fact an equality. for example.9 The existence of an invariant measure does not guarantee recurrence: consider.7.il = q < P = Pi.38 1.7. it is also recurrent.6. Now Theorem 1.6 forces any invariant measure to be a scalar multiple of 7r. there is a twoparameter family of invariant measures uniqueness up to scalar multiples does not hold.3.i+l = Pi. Discretetime Markov chains To complete the proof we return to the argument for (iii) => (i) armed with the knowledge that P is recurrent. whose transition probabilities are given by Pi. but has invariant measure 7r given by 7ri = 1 for all i.8 (Simple symmetric random walk on Z) The simple symmetric random walk on Z is clearly irreducible and. there can be no invariant distribution and the walk is therefore null recurrent.10 Consider the asymmetric random walk on Z with transition probabilities Pi. by Theorem 1.1.7.6.
1. A hole is made in the partition.2 Gas molecules move about randomly in a box which is divided into two halves symmetrically by a partition. Suppose we choose Pi converging sufficiently rapidly to 1 so that P = IIpi > 0 i=O 00 which is equivalent to 00 Lqi = i=O 00. What are the transition probabilities? What is the invariant distribution of this chain? 1. Let i be the initial vertex occupied by the particle. Then for any solution of 7r P = 7r we have and so 1ro ~ JJ7ro 7ro L qi· i=O 00 This last equation forces either measure. independently of its past motion. Suppose there are N molecules in the box. 1. for i ~ 7ri = Pil7ril. 0 the vertex opposite i.7.2. Calculate each of the following quantities: . so there is no invariant Exercises 1.7.3 A particle moves on the eight vertices of a cube in the following way: at each step the particle is equally likely to move to each of the three adjacent vertices. 7 Invariant distributions 39 7r P Then the components of the invariant measure equation 00 = 7r read 1ro = L i=O qi 1ri. Show that the number of molecules on one side of the partition just after a molecule has passed through the hole evolves as a Markov chain.7.1 Find all invariant distributions of the transition matrix in Exercise 1.1.1. = 0 or 7ro = 00.
Example 1.8 Convergence to equilibrium We shall investigate the limiting behaviour of the nstep transition probabilities p~.j E I) with aij = 1 for all i and j. Deduce that if P is irreducible then I P+A is invertible. Show that a distribution 7r is invariant for P if and only if 7r(I P+A) = a.2. so p2n converge for all i.40 1.il = q < P = Pi. 1. Then p2 = I.7.8. Note that this enables one to compute the invariant distribution by any standard method of inverting a matrix. We leave it as an exercise to show that i is aperiodic if and only if the set {n 2:: 0 : p~~) > O} has no common divisor other than 1.8. Thus p~j) fails to Let us call a state i aperiodic if p~~) > 0 for all sufficiently large n.1 Consider the twostate chain with transition matrix p=(~ ~). The behaviour of the chain in Example 1. j.7.7. 1. .10.4.6 and Example 1.) as n ~ 00. then it must be an invariant distribution. and a = (ai : i E I) with ai = 1 for all i.7. (iii) the expected number of steps until the first visit to o.i+l. (ii) the expected number of visits to 0 until the first return to i. But.4 Let (Xn)n~O be a simple random walk on Z with Pi.8. Find . where A = (aij : i. This is also a consequence of Theorem 1. the limit does not always exist. As we saw in Theorem 1.f = and verify that T'~ ~ TOl lEo ( n=O L ) l{x n =i} = inf Ai for all i A 1.1 is connected with its periodicity.) 1. as the following example shows.5 Let P be a stochastic matrix on a finite set I. =I and p 2n+l = P for all n.7. Discretetime Markov chains (i) the expected number of steps until the particle returns to i. where the infimum is taken over all invariant measures A with Ao (Compare with Theorem 1. if the statespace is finite and if for some i the limit exists for all j.
and suppose that P has an invariant distribution 7r.k)(j. Since P is aperiodic.k) = 7ri7rk so.l) is a Markov = PijPkl and initial distribution jj(i. b) so P(T < 00) = 1. P is positive recurrent. Then · (r) (s) 0 (r+n+s) (r) (n) (s) Pjk . In particular. for all states j and k. P) and independent of (Xn)n~O.8. k. P has an invariant distribution given by 7r(i.8 Convergence to equilibrium 41 Lemma 1.Pji Pii Pik > >0 for all sufficiently large n.7. Y n ) Step 1.l) = Pij Pkl > 0 for all sufficiently large n.) t 'lrj as n t 00 for all i. Let A be any distribution. by Theorem 1. Proof. . l we have ~n) (n) (n) P(i. The process W n = chain on I x I with transition probabilities P(i. . P).I · P rooJ. for all states i. Fix a reference state b and set T = inf{ n 2:: 1 : X n = Yn = b}. PJ~) > 0 for all sufflciently large n. We show P(T < 00) = 1.2.5. Then P(Xn = j) ~ 7rj as n ~ 00 for all j.7.8. The method of proof. There eXIst r. 0 Here is the main result of this section. Also. Suppose that (Xn)n~O is Markov(A. Suppose P is irreducible and has an aperiodic state i. Let P be irreducible and aperiodic. so P is irreducible.k) = Ai 7rk. (X n . by coupling two Markov chains. But T is the first passage time of W n to (b. all states are aperiodic. In particular. S _ > 0 WIt h Pji' Pik >. We use a coupling argument. p~.3 (Convergence to equilibrium). Theorem 1. Let (Yn)n~O be Markov(7r.1. is ingenious. by Theorem 1.7. j. Then. j.k)(j.
P).. P). YT+n)n~O is Markov(8(b. (XT. Yo). YI). (X I. (Zn)n~O is Markov(. we can replace the process (XT+n. so (XT+n. Discretetime Markov chains Step 2.(XT . P) where n . We show that (Zn)n~O is n The strong Markov property applies to (Wn)n~O at time T. YT+n)n~O by (YT+n . In particular.7rjl = IlP(Zn = j) lP(Yn = j)1 = IlP(Xn = j and n < T) lP(Yn = j and n < T)I lP(n < T) 0 as n ~ 00. ~ and P(n < T) D .42 1. .b). .. Zn = { Y Xn ifn < T if n 2 T. I ~ T. Z~) is Markov(j. P) and remains independent of (X o. Markov(. P) and independent of (X o. .b). Y T ) . YI).Yo). We have lP(Zn = j) = lP(Xn = j and n < T) +lP(Yn = j and n 2 T) so IlP(Xn ~ = j) .t..\. Step 3. XT+n)n~O which is also Markov(8(b. Hence W~ = (Zn.YT). . By symmetry. (Xl. .\. Set ifn < T if n The diagram below illustrates the idea.
Hence we have a partition. Choose n1..j E C r . (ii) p~jd) > 0 for all sufficiently large n. choosing m ~ 0 so that p~". because of periodicity. To prove (i) suppose p~j) > 0 and i E Cr. o~ r Now for nd ~ n~. Then Co U . then p~7d+r+k) > 0 so j E C r+n as required. and the proof fails. Since d divides n1 we then have r = md for some integer m and then nd = (q .8.. where (Xt)~o is periodic or transient or null recurrent. . U Cd1 ~ 1 and a such that (setting Cnd+r = Cr) (i) p~j) > 0 only if i E Cr and j E C r+ n for some r. 1. 1/2) as its unique invariant distribution... Hence and hence nd E S. (Here and throughout we use the symbol := to mean 'defined to equal'. then. Let P be irreducible. By taking i = j = k we now see that d must divide every element of S.1. However. E I (nd+r) : Pki > 0 £or some n ~ O} . To prove (ii) for i. Proof.1 which has (1/2. . We move on now to the cases that were excluded in the last theorem.I}.) Define for r = 0. if Yo = 1.. . . by irreducibility.) > 0. Fix a state k and consider S = {n ~ 0 : P~~ > O}. S E {O. for all i. in particular n1. n2 E S with n1 < n2 and such that d := n2 . if p~~d+r) > 0 and p~~d+8) > 0 for some r. j E Cr choose m1 and m2 so that p~:l) > 0 and p~72) > 0.8..m)n1 + mn2.. Moreover.n1 is as small as possible. for all r. U Cd1 = I. We start (Xn)n~O from 0 and (Yn)n~O with equal probability from 0 or 1. (Xn)n~O and (Yn)n~O will never meet. we have p~~d+r+m) > 0 and p~~d+8+m) > 0 so r = s by minimality of d. we can write nd = qn1 + r for integers q ~ n1 and ~ n1 . Theorem 1.8 Convergence to equilibrium 43 To understand this proof one should see what goes wrong when P is not aperiodic.. Consider the twostate chain of Example 1. The remainder of this section might be omitted on a first reading. then .d .d1 Cr = {~.4. then. Choose m so that p~7d+r) > 0. There is an integer d partition I = Co U C 1 U .1.
Suppose that (Xn)n~O is Markov(A. here is a complete description of limiting behaviour for irreducible chains.4. .1 and j E C r we have P(Xnd +r = j) ~ dlmj as n ~ 00 where mj is the expected return time to j.8. Pij Proof Step 1.. This is sometimes useful in identifying d. This generalizes Theorem 1. Set v 1.3 in two respects since we require neither aperiodicity nor the existence of an invariant distribution.4 we have = Apr. where 1r is the unique invariant distribution.1.Cdl be the partition obtained in Theorem 1. Since are done.5. D + m2 is then necessarily a multiple of d. then P(Xnd+r = j) = P(Yn = j) so the theorem holds in general. For j E Cr the expected return time of (Yn)n~O to j is mjld. Assume that P is aperiodic. Let P be irreducible of period d and let Co.4. The theorem just proved shows in particular for all i E I that d is the greatest common divisor of the set {n ~ 0 : p~~) > O}. . . Otherwise mj = 00 and we have to show that P(Xn = j) ~ 0 as n ~ 00. The argument we use for the null recurrent case was discovered recently by B.. pd is irreducible and aperiodic on Cr. . so the result follows from Theorem 1. pd) and.8. Then for r = 0. ~ dlmj as n ~ 00 Step 2. Discretetime Markov chains ml whenever nd ~ n~. for i E Co and j E C r we have (nd+r) ~ dl mj as n ~ 00. In particular.44 1. So if the theorem holds in the aperiodic case.8. we We call d the period of P. Finally. C 1 . by Theorem 1. Theorem 1.d .. If P is positive recurrent then 1/mj = 1rj. We reduce to the aperiodic case. Gray. P). then by Theorem Set Y n = Xnd+r. Fristedt and L. Let A be a distribution with LiECo Ai = 1. then (Yn)n~O is Markov(v..3. .8.8.8.
. .8 Convergence to equilibrium 45 If P is transient this is easy and we are left with the null recurrent case. j)) ~0 as required..3.K 1. on taking j.K .1 :E k=nK+l n P(Xk = j and X m 1= j for m = k + 1. . where j.t. Given € > 0 choose K so that Then. Then 00 :EPj(Tj > k) = lEj(Tj) = k=O 00. . We can find N such that for n 2:: Nand k = 1.8. ~· . . Then. . If (Wn)n~O is transient then. Assume then that (Wn)n~O is recurrent.t = A. we have P(T < 00) = 1 and the coupling argument shows that as n ~ 00. . IP(Xn = j) . we obtain P(Xn = j)2 = P(Wn = (j.t is to be chosen later.k) = :E P(Xnk = j)Pj(Tj > k) k=O so we must have P(Xn . Step 3.P(Xn+k = j)1 ::. aperiodicity of (Xn)n~O ensures irreducibility of (Wn)n~O..3. 1. As before.n) L k=nK+l Kl n P(Xk = j)Pj(Tj > n . Set W n = (Xn .P(Yn = j)1 ~ 0 We exploit this convergence by taking j. Return now to the coupling argument used in Theorem 1.K . IP(Xn = j) ..1. so that P(Yn = j) = P(Xn +k = j). . in the notation of Theorem 1....I}.Yn ).8.t = Ap k for k = 1.1..k = j) ~ €/2 for some k E {O. Assume that P is aperiodic and null recurrent. for n 1~ ~ K . . only now let (Yn)n~O be Markov(j. P).
4 Each morning a student takes one of the three books he owns from his shelf.7. The probability that he chooses book i is Qi. and compare them with your answers there.8. .. and determine the limit. this shows that lP(Xn = j) required. show that. Discretetime Markov chains ~ But for any n we can find k E {O. 1. (f) and (g) made in example (v) of the Introduction..3 A fair die is thrown repeatedly. from left to right.2 Find the invariant distributions of the transition matrices in Exercise 1.1.8. (b) and (c). Let X n denote the sum of the first n throws.2. Show that the following = inf{m ~ n: m = Y1 + .. Let Y1 . . for n ~ N P(Xn = j) ~ c. Pn converges as n ~ 00. parts (a).3. Determine n+oo and hence show that as n P(n ~ 00 = Y1 + . where 0 < Qi < 1 for i = 1. and choices on successive days are independent.5 (Renewal theorem).I} such that P(Xn + k = j) c/2. Find lim P(Xn is a multiple of 13) n+oo quoting carefully any general theorems that you use.3. If Pn denotes the probability that on day n the student finds the books in the order 1. Since c > 0 was arbitrary.. }. identically distributed random variables with values in {I.8. 1.n. irrespective of the initial arrangement of the books. Suppose that the set of integers {n : P(Y1 = n) > I} has greatest common divisor 1. 2.K . Set Jl process is a Markov chain: Xn = E(Y1 ).1 Prove the claims (e)..2.46 1. 1.8. Hence... D Exercises ~ 0 as n ~ 00. ••• be independent. . 1. + Yk for some k ~ 0) ~ 1/ Jl. . 1. as 1. Y 2 .8. + Yk lim lP(Xn = 0) for some k ~ O} .. In the evening he replaces the book at the lefthand end of the shelf.
Suppose that (Xn)O~n~N is Markov(1T'.1.. . Xl = iNI. This is an example of entropy increasing. P).n .9. Let P be irreducible and have an invariant distribution 1r. is again a Markov chain. A bulb is replaced when it fails. This property is symmetrical in time and suggests looking at Markov chains with time running backwards. Thus the limiting probability that a bulb is replaced at time n is 1/ Jl. P i 1io = 1rioPioi1 .XN = io) = 1riN PiNiN 1 . The next result shows that a Markov chain in equilibrium. • •• as lightbulb lifetimes.. Although this appears to be a very special case of convergence to equilibrium. where P = (Pij) is given by 1rjPji = 1riPij for all i. We have P(YO= io.. j and P is also irreducible with invariant distribution 1r.. On the other hand. First we check that P is a stochastic matrix: since 1r is invariant for P. one can actually recover the full result by applying the renewal theorem to the excursion lengths S~l).9 Time reversal For Markov chains. Then (Yn)O~n~N is MarkovCrr. Y 2 .. from state i.) 1.9 Time reversal 47 (Think of YI . the past and future are independent given the present..··· . Theorem 1. run backwards. PiN 1 iN .YN = iN) = P(Xo = iN. S~2). Next we check that 1r is invariant for P: L JEI 1rjPji = L JEI 1riPij = 1ri since P is a stochastic matrix. the invariant distribution. The transition matrix may however be different.·. convergence to equilibrium shows behaviour which is asymmetrical in time: a highly organised state such as a point mass decays to a disorganised one. YI = i l . P) and set Y n = X N . Proof. It suggests that if we want complete timesymmetry we must begin in equilibrium.· .1.
The chain (Yn)O~n~N is called the timereversal of (Xn)O~n~N.48 1. Then Pin inl . Let P be an irreducible stochastic matrix and let A be a distribution.. the following result is worth remembering because. Pinl in / 7ri n >0 so P is also irreducible.9. If P and A are in detailed balance. Suppose that (Xn)n~O is Markov(A.1. P). Proof. since P is irreducible. Then both (a) and (b) are equivalent to the statement that P = P in Theorem 1.2. when a solution A to the detailed balance equations exists. Both (a) and (b) imply that A is invariant for P. D We begin a collection of examples with a chain which is not reversible. with P irreducible. Discretetime Markov chains so..il. Theorem 1. Pil io D = 7rioPioil . Example 1.9. then A is invariant for P. (XNn)O~n~N is also Markov(A.. Though obvious. A stochastic matrix P and a measure A are said to be in detailed balance if AiPij = AjPji for all i.. for each pair of states i.4 Consider the Markov chain with diagram: 1 3 3 2 2 .. j. Then the following are equivalent: (a) (Xn)n~O is reversible.. (b) P and A are in detailed balance. P). Lemma 1.9. (Yn)O~n~N is Markov(7r.1. Pinli n > O. Finally. .in1. P). j there is a chain of states io = i..3.in =j withpioil . We say that (Xn)n~O is reversible if. by Theorem 1..1. Proof· We have (AP)i = LjE! AjPji = LjE! AiPij = Ai· D Let (Xn)n~O be Markov(A. . for all N ~ 1. P).9. it is often easier to find by the detailed balance equations than by the equation A = AP.
.~ iI i i+l Ml M <P= 1 q < 1.9 Time reversal The transition matrix is 0 1/3 ( 2/3 49 p= 2/3 0 1/3 1/3) 2/3 0 and 1r = (1/3. Example 1. 1/3. so P =I P and this chain is not reversible.5 Consider the Markov chain with diagram: P q 1 P • q II( • o where 0 . But p is not symmetric. The nonzero detailed balance equations read AiPi.. If P were much larger than q. . .aph G is a countable collection of states. A patient observer would see the chain move clockwise in the long run: under timereversal the clock would run backwards! Example 1. 1/3) is invariant. which in this case would be heavily concentrated near M. some of which are joined by edges. 1. one might argue that the chain would tend to move to the right and its timereversal to the left.6 (Random walk on a graph) A gr.i for i = 0.M) and this may be normalised to give a distribution in detailed balance with P. Hence this chain is reversible. which behaviour is symmetrical in time. the transpose of P. However. for example: .1.M . So a solution is given by A = ((p/q)i : i = 0..9. 1.. An observer would see the chain spending most of its time near M and making occasional brief forays to the left. this ignores the fact that we reverse the chain in equilibrium.i+l = Ai+lPi+l..1.9. Hence P = pT. . usually called vertices.~ II(.
It is easy to see that P is in detailed balance with V == (Vi : i E G). There is a natural way to complete the diagram which gives rise to the random walk on G.9.50 1. The valency Vi of vertex i is the number of edges at i. then 1r == V / a is invariant and P is reversible. We assume G is connected. j) is an edge otherwise. how long on average will it take to return? This is an example of a random walk on a graph: the vertices are the squares of the chessboard and the edges are the moves that the knight can take: . if the total valency a == LiEG Vi is finite.7 (Random chessboard knight) A random knight makes each permissible move with equal probability. The random walk on G picks edges with equal probability: 1 1 2 1 3 1 2 1 2 1 3 1 3 4 1 1 2 3 2 3 Thus the transition probabilities are given by if (i. We have to assume that every vertex has finite valency. Example 1. so that P is irreducible. So. Discretetime Markov chains 1 2 4 3 Thus a graph is a partially drawn Markov chain diagram. If it starts in a corner.
if you enjoy solving sets of 64 simultaneous linear equations. (b) (c) 1== {O. and the 16 central squares have valency 8.1 In each of the following cases determine whether the stochastic matrix P. Hence lEc(Tc ) = 8 + 24 + 80 + 96 + 128 = 168. We know by Theorem 1. .7 and the preceding example that so all we have to do is identify valencies.5! Exercises 1. There are 20 squares of valency 4. you might try finding 1r from 1r P == 1r. 16 of valency 6.9..1. .N} andp~J == 0 if Ij il2:: 2. . which you may assume is irreducible.3. .7. and the eight squares adjacent to the corners have valency 3.l. The four corner squares have valency 2.. 2 Alternatively. or calculating lEe (Te ) using Theorem 1. is reversible: (a) lq p ) .9 Time reversal 51 The diagram shows a part of the graph.
1.9.. when both X and Y start at I.. We shall prove a theorem which identifies for Markov chains the longrun proportion of time spent in each state. Show that 9MD = 16MB. I P for i 2:: 1. identically distributed. For I = B. Pi. a particle at A jumps to B. C or D with equal probability 1/3.il = 1 Pij = Pji for all i. until they are once again both at I... Pi.52 1. 1. . } and POI = 1...j E S. Discretetime Markov chains (d) (e) = {a.10 Ergodic theorem Ergodic theorems concern the limiting behaviour of averages over time.. nonnegative random . 1._..10.1 (Strong law of large numbers).. So. .. Y 2 . Let YI . Theorem 1.2 Two particles X and Y perform independent random walks on the graph shown in the diagram. An essential tool is the following ergodic theorem for independent random variables which is a version of the strong law of large numbers. DIet MI denote the expected time. for example. (b) X starts at A and Y starts at E. B C ~"II E Find the probability that X and Y ever meet at a vertex in the following cases: (a) X starts at A and Y starts at B.. D p . be a sequence of independent.i+l = P.2.
with probability 1 ~oo Y1 + . If (Xn)n?O is Markov(.x (Vi (n) n t . So we must have.4).. P) then ]p> .x. in Probability with Martingales by David Williams (Cambridge University Press.as n mi t 00) = 1 where mi = Ei(Ti ) is the expected return time to state i.!.. A proof for the case J_l < 00 may be found. Moreover. .. + Yn n ~J_lasn~oo ) = 1. in the positive recurrent case. Let P be irreducible and let be any distribution. The following result gives the longrun proportion of time spent by a Markov chain in each state. for example. 1991). Then Y1 + ···+ Yn > Y1 n (N) + ···+ Yn n (N) _ t E(Y1 /\ N) asn~oo with probability one. Proof.10. Fix N < 00 and set yJN) = YnAN. Theorem 1. As N i 00 we have E(YI A N) i J_l by monotone convergence (see Section 6.1 0 Ergodic theorem 53 variables with JE(Y1 ) = 11']) jj. D We denote by Vi (n) the number of visits to i before n: nl Vi(n) = L l{xk=i}' k=O Then Vi (n )/ n is the proportion of time before n spent in state i..1. + Yn n as n ~ 00. The case where J_l = 00 is a simple deduction...2 (Ergodic theorem). for any bounded function f : I ~ lR we have where and where (7ri : i E I) is the unique invariant distribution. Then c (Yl + .
. the total number Vi of visits to i is finite.l) :::...5. .V~(n)) Vi(n) (1.8).5. we get JP> (Vi7n) t mi as n . . letting n ~ ~ 00 as n ~ 00) = 1. since P is recurrent P(Vi(n) So..1.!.XT by the strong Markov property.2).l) + . mi Suppose then that P is recurrent and fix a state i..1.. 8.l). By Lemma 1.r)) = mi. + S.. Also the lefthand side being the time of the first visit to i after n . Vi(n) < + .7 and (XT+n)n~O is Markov(8i .54 1. .as n mi t (0) = 1.!. Now S't~l) + ••• + S~V~(n)l) < n 1 _ . then. + S. . For T = T i we have P(T < 00) = 1 by Theorem 1. If P is transient...8) By the strong law of large numbers JP> (S~l) + • ~ ~. so it suffices to consider the case A = bi. . The longrun proportion of time spent in i is the same for (XT+n)n>O and (Xn)n>O. Hence S. + S~n) ~ t mi as n ) . with probability 1. the nonnegative random variables S..V~(n)l) Vi(n) n S. 't the lefthand side being the time of the last visit to i before n. Write sir) for the length of the rth excursion to i. as in Section 1. Xl. which implies P (Vi(n) n t .n t 0= .t 00) = 1.5.t 00 = 1 and. are independent and identically distributed with Ei(S. P) and independent of X o. 00 in (1.. so Vi(n) < Vi n . Discretetime Markov chains Proof..
.7ril i~J ~ L I Vi~n) ~ 2~ " iEJ 7ril + L i~J (Vi~n) + 7ri) i~J InVi(n)  1ri I +2~1ri. we have < c. 1 L  Vi(n) 1ri ) Ii k=O 'tEl ~ L IVi~n) iEJ iEJ . D We consider now the statistical problem of estimating an unknown transition matrix P on the basis of observations of the corresponding Markov chain. Then. For any J ~ I we have I ~ nl !(Xk) ! I = ~ ( n. I < c/4. The loglikelihood function is given by I(P) = log(. which establishes the desired convergence.PXN1XN) = L i. " We proved above that JP> (Vi ~n) t 7ri as n t 00 for all i) = 1.1 0 Ergodic theorem 55 Assume now that (Xn)n~O has an invariant distribution (1ri : i E I). for n ~ N(w) 1ri "I iEJ Vi(n) ~ n. for n ~ N(w). Consider. to begin.\xoPxoxl •• .7ril + L IVi~n) .jEl N ij logpij . the case where we have N + 1 observations (Xn)O~n~N. Given c > 0. Let I : I ~ lR be a bounded function and assume without loss of generality that III ~ 1.1. choose J finite so that and then N = N(w) so that.
We now turn to consider the consistency of this sort of estimate. Note that to find Pij we simply have to maximize 2: N jEI ij log Pij subject to E j Pij = 1: the other terms and constraints are irrelevant. where Y n = 1 if the nth transition from i is to j.56 1. which shows that Pij is consistent. Denote again by N ij the number of transitions from i to j. we first try to maximize l (P) + 2: i. + Y N . Thus we find Nl Pij = L Nl 1{Xn =i. we shall slightly modify our approach.X n +1 =j}/ L l{x n =i} n=O n=O which is the proportion of jumps from i which go to j. By the strong Markov property Yl.. In the transient case this may involve restarting the chain several times. To maximize the likelihood for (Pij : ij j E I) we still maximize 2: N jEI log Pij subject to Ej Pij = 1. where N ij is the number of transitions from i to j. . .. But Nij = Y1 + . Suppose then that instead of N + 1 observations we make enough observations to ensure the chain leaves state i a total of N times.. that is to say whether Pij ~ Pij with probability 1 as N ~ 00. .Y are N independent and identically distributed random variables with mean Pij. Discretetime Markov chains up to a constant independent of P. which is the choice of P maximizing l(P). . and Yn = 0 otherwise. which leads to the maximum likelihood estimate Pij = Nij/N. by the strong law of large numbers P(Pij ~ Pij as N ~ 00) = 1.jEI /LiPij and then choose (/li : i E I) to fit the constraints. A standard statistical procedure is to find the maximum likelihood estimate P. So. Since this is clearly false when i is transient. Since P must satisfy the linear constraint E j Pij = 1 for each i. This is the method of Lagrange multipliers.
4. 1. (See Example 1. A more specialized case was dealt with in Example 1. Flowers costing x thousand pounds.4.4 An opera singer is due to perform a long series of concerts.a).1. so provided a =I 1 we must have x = b/(l . then x = ax + b.b/(l . independently of past weather. He walks to the office in the morning and walks home in the evening. This he does by sending flowers every day until she returns.10. For J ~ I let (Ym)m~O be the Markov chain on J obtained by observing (Xn)n~O whilst in J.10. Suppose that it rains on each journey with probability p.4 we found a recurrence relation of the form Xn+l = aX n + b.3 we found a recurrence relation of the form aXn+l + bXn + CXnl = 0 . The promoter stands to make £750 from each successful concert. 1. Having a fine artistic temperament. Here is an account of the simplest cases.3. We look first for a constant solution X n = x. When a = 1 the general solution is obviously Xn = Xo +nb. What is the longrun proportion of journeys on which the professor gets wet? 1.a) satisfies Yn+l = aYn. In Example 1. so Yn = anyo. Now Yn = Xn .11 Appendix: recurrence relations Recurrence relations often arise in the linear equations associated to Markov chains.2 A professor has N umbrellas. Once this has happened she will not sing again until the promoter convinces her of his high regard. she is liable to pullout each night with probability 1/2.10. If it is raining he likes to carry an umbrella and if it is fine he does not.1 Prove the claim (d) made in example (v) of the Introduction.3.1. Thus the general solution when a =I 1 is given by Xn = Aa n + b/ (1  a) where A is a constant.4.10. How much should he spend on flowers? 1. In Example 1.11 Appendix: recurrence relations 57 Exercises 1. bring about a reconciliation with probability yIX.3 Let (Xn)n~O be an irreducible Markov chain on I having an invariant distribution Jr.) Show that (Ym)m~O is positive recurrent and find its invariant distribution. 0 ~ x ~ 1.
Discretetime Markov chains where a and c were both nonzero. .. Then Yn = Ao: n + B{3n is a solution. If 0: = {3 # 0. then aX 2 + bX + c = O. Denote by 0: and {3 the roots of this quadratic. Yn = Xn for all n. but for all n. 1 + t) = t  ~t2 + ~t3  10g(1 . then Yn = (A + nB)o:n is a solution and we can solve so that Yo = Xo and Yl = Xl. . Here is a derivation.~t2 . Hence the general solution is given by Ao:n + B{3n if 0: # {3 n { (A + nB)an if 0: = {3. by the same argument.. If 0: =1= (3 then we can solve the equations xo=A+B. Let us try a solution of the form X n = X n. then. so that Yo = Xo and Yl xl=Ao:+B{3 = Xl.6 rested heavily on the use of the asymptotic relation n! rv Ayri(nje)n as n ~ 00 for some A E [1.~t3  By subtraction we obtain 2 log 1 (l+t) t 1_ = 13 15 t + 3 t +:5 t +. so by induction Yn = Xn for all n. 00 ).. X = 1.12 Appendix: asymptotics for n! Our analysis of recurrence and transience for random walks in Section 1.58 1. We make use of the power series expansions for 10g(1 It I < . The case 0: = (3 = 0 does not arise.t) = t .
1 3 (2n + 1)2 . where A = ea.. as n ~ 00. 00) and hence An ~ A.. . Hence an ~ a for some a E [0.12 Appendix: asymptotics for n! 59 Set An = n!/(nn+l/2 en) and an calculation an .1/(12n) increases as n ~ 00. (2n+1)5 1 1 1 = 3 (2n + 1)2 + 5 (2n + 1)4 + · . Then.1.12(n + 1) ..1 It follows that an decreases and an .1.an+! = (2n = logAn. ::. by a straightforward + 1)2" log 1 (1+(2n+l)1) 1 _ (2n + 1)1 . 3" (2n (2n 1)4 1 + II} { + + + ··· 1 1 1 12n . By the series expansion written above we have anan+! = (2n+1) { Ill 1 1 1)2 (2n+1) +3"(2n+1)3 +5 II} + .
Section 2. 2. Sections 2. The first three sections of Chapter 2 fill in some necessary background information and are independent of each other. If you wish. Some examples of more general processes are given in Section 2. but once up and running it follows a very similar pattern to the discretetime case. As in Chapter 1 the exercises form an important part of the text.8 are read selectively at first.9. which are reviewed in Section 3. Let I be a countable set. There is an irreducible level of difficulty at this point.1 Qmatrices and their exponentials In this section we shall discuss some of the basic properties of Qmatrices and explain their connection with continuoustime Markov chains.8. especially 2.j E I) . provided you take certain basic properties on trust. These processes are simple and particularly important examples of continuoustime chains.5 on birth processes provide a gentle warmup for general continuoustime Markov chains.4 on the Poisson process and Section 2. you can begin with Chapter 3.2 Continuoustime Markov chains I The material on continuoustime Markov chains is divided between this chapter and the next. A Qmatrix on I is a matrix Q satisfying the following conditions: = (qij : i. so we advise that Sections 2.1. The theory takes some time to set up.7 and 2. To emphasise this we have put the settingup in this chapter and the rest in the next.8. deal with the heart of the continuoustime theory.62.
We may think of the discrete parameter space {O. where q = logp.00) a natural way to interpolate the discrete sequence (pn : n = 0. We shall later interpret qi as the rate of leaving i. For P E (0. (iii) L jEI qij = 0 for all Thus in each row of Q we can choose the offdiagonal entries to be any nonnegative real numbers. The diagonal entry qii is then qi. } as embedded in the continuous parameter space [0. i. 1. in this case Q= (~2 2 !1 1 3 ~) Thus each offdiagonal entry qij gives the value we attach to the (i. Consider now a finite set I and a matrix . making the total row sum zero.00). ) is by the function (e tq : t ~ 0). for all i qij ~ 0 =I j. j) arrow on the diagram.1 Qmatrices and their exponentials (i) 0 ~ (ii) qii 61 < 00 for all i. for example: 1 3 1 2 Each diagram then corresponds to a unique Qmatrix.... but you can work them out from the other information given. subject only to the constraint that the offdiagonal row sum is finite: qi = Lqij j#i < 00.2. .2. A convenient way to present the data for a continuoustime Markov chain is by means of a diagram.. The numbers qi are not shown on the diagram.2.1. which we shall interpret later as the rate of going from i to j. .
. 0) is the unique solution to the backward equation P(O) = I.1. t E lR. .. Then (P(t) : t ~ 0) has the following properties: (i) P(s + t) = P(s)P(t) for all s. The matrixvalued power series P(t) = f: (t~)k k=O . For any s. . )? For any matrix Q = (qij : i.1. we have Proof. Set P(t) = e tQ . if two matrices Q1 and Q2 commute. Moreover. (iv) for k = 0. .. Suppose then that we can find a matrix Q with eQ = P. j E I). so e sQ e tQ = e(s+t)Q proving the semigroup property. Is there a natural way to fill in the gaps in the discrete sequence (p n : n = 0. (iii) (P(t) : t ~ P(O) = I.2.62 2.2.j E I). Theorem 2. (ii) (P(t) : t ~ 0) is the unique solution to the forward equation ~ P(t) = P(t)Q.1. t (semigroup property).. then The proofs of these assertions follow the scalar case closely and are given in Section 2.1. Let Q be matrix on a finite set I. Continuoustime Markov chains I P = (Pij : i. sQ and tQ commute. Then e nQ = (eQ)n = p n so (e tQ : t ~ 0) fills in the gaps in the discrete sequence. the series converges componentwise and we denote its limit by eQ .10.
(ii) LPij = 1 for all i.e.j and all t ~ o. It remains to show that P(t) is the only solution of the forward and backward equations.!!. and so M(t) = P(t). Since P(t) = P(t/n)n for all n.!!. Moreover by repeated termbyterm differentiation we obtain (iv).1 Qk 1)! = P(t)Q = QP(t). But if M(t) satisfies the forward equation.!!. P (t) = L (k _ k=1 00 t k. then .tQ + M(t) (. jEI We recall the convention that in the limit t ~ 0 the statement f(t) = O(t) means that f(t)/t ~ C for all sufficiently small t.2. Later we shall also use the convention that f(t) = o(t) means f(t)/t ~ 0 as t ~ o.tQ is constant. LJ (n) LJ LJ (n LJ ( n LJ kEI kEI jEI jEI kEI . If Q has zero row sums then so does Qn for every n: """ qik = """ """ qij 1) qjk = """ qij 1) """ qjk = 0 . A similar argument proves uniqueness for the backward equation.tQ ) dt dt dt = M(t)Qe. Recall that a matrix P = (Pij : i.2. it follows that qij ~ 0 for i =1= j if and only if Pij(t) ~ 0 for all i. D The last result was about matrix exponentials in general. A matrix Q on a finite set I is a Qmatrix if and only if P(t) = etQ is a stochastic matrix for all t ~ o. So each component is differentiable with derivative given by termbyterm differentiation: .j.1 Qmatrices and their exponentials 63 has infinite radius of convergence (see Section 2.tQ + M(t)( Q)e. for some C < 00. j E I) is stochastic if it satisfies (i) 0 ~ Pij < 00 for all i.10). Proof. Theorem 2. Hence P(t) satisfies the forward and backward equations. As t ! 0 we have so qij ~ 0 for i =1= j if and only if Pij (t) ~ 0 for all i.tQ = 0 so M(t)e.1. j and t ~ 0 sufficiently small. Now let us see what happens to Qmatrices.(M(t)etQ ) = (.M(t)) e.
.2. we can do some sort of fillingin of gaps at the level of processes. (eQ/m)m) (see Exercise 1.1. ) is discretetime Markov(. } as timeparameter sets which give rise to Markov(. the transition probability from i to j in time t is given by ° (Recall that := means 'defined to equal'..8 that a continuoustime Markov chain (Xt)t~O with Qmatrix Q satisfies ~ to ~ . .2. then L qij jEI ~I LPij(t) jEI = O. we shall see in Section 2. .) You should compare this with the defining property of a discretetime Markov chain given in Section 1. .. Fix some large integer m and let (X~)n~O be discretetime Markov(. Continuoustime Markov chains I tn LPij(t) jEI = 1 + L I" Lqij n.1.2.j) entry in etQ .2..1. .1. } by Then (Xn : n = 0. where Pij(t) is the (i. all times io. . .2) and Thus we can find discretetime Markov chains with arbitrarily fine grids {n/m : n = 0.~o which also has this property.64 So 2. On the other hand.. P) when sampled at integer times. D t=O Now. .. if P is a stochastic matrix of the form eQ for some Qmatrix. We define a process indexed by {n/m : n = 0. ~ t n +l and all states for all n = 0. as we shall see in Section 2.i n + 1 .. eQ / m ). if LjE! Pij(t) = 1 for all t = ~ 0. It should not then be too surprising that there is. We shall now give some examples where the transition probabilities Pij(t) may be calculated explicitly. ..1.x.x....x..8. In particular.1. a continuoustime process (Xt)t. n=l jEI 00 (n) = 1. To anticipate a little.
2. 4.3 65 We calculate PII(t) for the continuoustime Markov chain with Qmatrix Q= (~2 2 !1 1 3 ~) The method is similar to that of Example 1.1.1 Qmatrices and their exponentials Example 2.) To determine the constants we use 1 = PII (0) = a + b + c. We begin by writing down the characteristic equation for Q: o = det (x  Q) = x (x + 2) (x + 4).4t for some constants a.I 0 o (4t)k ~ e ) U I . This shows that Q has distinct eigenvalues 0. 2 = qII = P~I (0) = 2b . Then PII(t) has the form PII(t) = a + be. 7= so qii) = P~I (0) = 4b + 16c. 2.4c. 4t so PII(t) must have the form claimed. (This is because we could diagonalize Q by an invertible matrix U: Then 0 (2t)k 0 ) U.2t + ce. band c.1.6. .
. by induction = e.1 Compute Pll(t) for P(t) = etQ .4 We calculate Pij(t) for the continuoustime Markov chain with diagram given above. P~j(t) = APij(t) + APi... We can solve these equations. for i < N.. P~N(t) = APiNl(t).. PiN(O) = 0. N}.A. these are the Poisson probabilities of parameter At.jl (t). 41. So. so Pij(t) = 0 for i > j. Then. pii(t) i <j < N so.~~.At for i < N. The exponential of an uppertriangular matrix is uppertriangular. where yt is a Poisson random variable of parameter At. The Qmatrix is A A A A A A Q= o A where entries off the diagonal and superdiagonal are all zero. Exercises 2. for Pij(t) = e>. Pii (0) == 1.o. Pij (0) == 0. First. A ~.. In components the forward equation P'(t) = P(t)Q reads P~i(t) = APii(t).t (At)ji (j _ i)!' If i = 0.... the distribution of the Markov chain at time t is the same as the distribution of min{yt. Continuoustime Markov chains I A . where Q= (~2 2 !4 1 3 ~). starting from 0.1.66 2. for i < j < N... 1 . .. for i < N. 2 Nl N Example 2.1.
0 ~ to ~ tl ~ . which is proved in Section 6.in E I.. at least in principle. .. . X t1 = il. .2 Continuoustime random processes 2.6. What consequences do your answers have for the discretetime Markov chains with these transition matrices? 2. for a countable disjoint union whereas for an uncountable union Ut>o At there is no such rule. A continuoustime random process with values in I is a family of random variables X t : n ~ I. or P(Xt = i for some t). . These should enable us to find. such as lP(Xt = i) or lP(Xto = io.2. We are going to consider ways in which we might specify the probabilistic behaviour (or law) of (Xt)t~o. To avoid these subtleties as far as possible we shall restrict our attention to processes (Xt)t~O which are rightcontinuous. . the probability of any event depending on a rightcontinuous process can be determined from its finitedimensional distributions. .....J P(Xq1 P(Xt = i for some t E [0. any probability connected with the process.. from the probabilities lP(Xto = io. is an enumeration of the rationals. . . For example '""" L.in =l=i where ql.. There are subtleties in this problem not present in the discretetime case. This means in this context that for all wEn and t ~ 0 there exists € > 0 such that for t ~ s ~t + €.2 Which of the following matrices is the exponential of a Qmatrix? 67 (a) (~ ~) (b) (~ ~) (c) (~ ~). They arise because.2 Continuoustime random processes Let I be a countable set.Xqn = jn) jl .. .1.Xtn = in). that is. q2.Xtn = in) for n ~ 0. By a standard result of measure theory. ~ t n and i o. ..00)) = 1n+oo lim = ji.
In this case. after the explosion time ( the process starts up again. In the first case the path makes infinitely many jumps. so there are three possibilities for the sorts of path we get. this is illustrated below.1 t The second case is where the path makes finitely many jumps and then becomes stuck in some state forever: • Jo = 0 J. or it may not. . Continuoustime Markov chains I Every path t ~ Xt(w) of a rightcontinuous process must remain constant for a while in each new state. maybe infinitely often. t]: • o • • o ~ Jo = 0 1. but only finitely many in any interval [0.68 2.2 t In the third case the process makes infinitely many jumps in a finite interval. it may explode again.
The (first) explosion time ( is defined by (= supJn = n I:Sn.. we define X oo = XJn . or the jump chain if it is a discretetime Markov chain. Note that rightcontinuity forces 8 n > 0 for all n.. the jump times of (Xt)t~O and 8 1 . . This is simply the sequence of values taken by (Xt)t~O up to explosion. Any process satisfying this requirement is called minimal.. The terminology 'minimal' does not refer to the state of the process but to the ... for n = 1. Jo = 0 t 14...1. where inf 0 = 00.. So it is convenient to adjoin to I a new state.2 Continuoustime random processes 69 :.... . . 00 say.e... and require that X t = 00 if t ~ (. . the holding times. ..  We call Jo..82 . We shall not consider what happens to a process after explosion._. • " . J 1 ...... otherwise X oo is undefined.. • . .... .. and... They are obtained from (Xt)t~O by for n = 0.... • . the final value.. n==1 00 The discretetime process (Yn)n~O given by Y n = X Jn is called the jump process of (Xt)t~O.... if I n 1 < 00 otherwise.... If I n + 1 = 00 for some n. ~~~~~ ..2.2. . .
70 2.1 (Memoryless property).00)) = lP(Yn = i for some n ~ 0). t < I n +1) = i for some t E [0.00] has exponential distribution of parameter X (0 ~ X < 00) if lP(T > t) = e. and (Yn)n~O we have another 'countable' specification of the probabilistic behaviour of (Xt)t~o.. Thus by specifying the joint distribution of 8 1 . If X > 0. Note that a minimal process may be reconstructed from its holding times and jump process.00] has an exponential distribution if and only if it has the following memoryless property: lP(T> s Proof. The exponential distribution plays a fundamental role in continuoustime Markov chains because of the following results. then T has density function The mean of T is given by E(T) = 1 00 IP(T > t)dt = AI. Continuoustime Markov chains I interval of time over which the process is active. We write T rv E(X) for short.82. . then lP(T > t) for all s. A random variable T: n ~ (0. t ~ O.At for all t ~ O. . rv IP(T> s + tiT> s) = P(T> s + t) IP(T> s) = eA(s+t) e>'s = At e = IP(T > t). the probability that X t = i is given by IP(Xt = i) = and lP(Xt n=O L IP(Y 00 n = i and I n :::.3. Suppose T + tiT> s) = E(X). For example. 2.3 Some properties of the exponential distribution A random variable T : n ~ [0. Theorem 2..
2. 0 The next result shows that a sum of independent exponential random variables is either certain to be finite or certain to be infinite. Proof. (i) Suppose E~1 1/ An < Then. Then. by induction so g(l) = e.AT = g(r) ~ g(t) ~ g(s) = e. t ~ O. this forces g(t) so T rv E(A).. Then g(t) = JP>(T > t) satisfies g(s + t) = g(s)g(t) for all s. suppose T has the memoryless property whenever JP>(T> s) > O. (ii) If ~ An 00 1 = 00. by monotone convergence so . A < 00.2. We assumed T > 0 so that g(l/n) > 0 for some n. > 0. . then then JP> (00) = ~ Sn < 00 l.At .AS = e. t ::.AT for all rationals r > O. since we can choose rand s arbitrarily close to t. l. By the same argument.q ~ 1 g(p/q) = g(l/q)P = g(l)pjq so g(r) = e. JP> (00) = ~ Sn = 00 00. . 8 2 . Let 8 1 . choose rationals e. Theorem 2.3. For real t r. be a sequence of independent random variables with 8 n rv E(A n ) and 0 < An < 00 for all n.3 Some properties of the exponential distribution 71 On the other hand. and gives a criterion for deciding which is true.A for some 0 ::. Since 9 is decreasing. for integers p. and. s. 1 (i) If~ An < 00 00. s > 0 with r ::. This will be used to determine whether or not certain continuoustime Markov chains can take infinitely many jumps in a finite time.
we have rv E(A) and R rv jjJP>(S ~ t < S + R) = AJP>(R ~ t < R+ S).3.qkB 00 IT eqjBds j# = 1 qke  q 8 qk ds = e .4.72 2. . with T rv E(q) and lP(K = k) = qk/q.3. Proof. (ii) Suppose instead that E~11/An = 00. For independent random variables S E(jj) and for t ~ 0. Set T = inf k T k . D = 1 and T and K have the claimed joint The following identity is the simplest case of an identity used in Section 2. Continuoustime Markov chains I 00. Set K Then = k if T k < T j for all j =I k. be independent random variables with Tk rv E(qk) and 0 < q := EkE! qk < 00. q Hence lP(K = k for some k) distribution. lP(K = k and T ~ t) ~ = lP(Tk = 1 it t t and Tj > Tk for all j =I k) 00 qkeqkBP(Tj > s for all j 1= k)ds = (OO qk e . T and K are independent. Then rr~o(1 + 1/ An) = By monotone convergence and independence so JP> (f n==l Sn = 00) = 1. Theorem 2.q t . Theorem 2. Let I be a countable set and let Tk' k E I. D The following result is fundamental to continuoustime Markov chains.3.8 in proving the forward equations for a continuoustime Markov chain. Moreover. with probability 1. Then this infimum is attained at a unique random value K of k. otherwise let K be undefined.
. 8 2 . .3.. What is the distribution of min{S. 2. The reader might well begin with the statement of this result and then see how it is used in the . Show that AlSl is exponential of parameter 1. both as a gentle warmup for the general theory and because they are useful in themselves.3 Let 81. be independent exponential random variables with parameters AI. T} ~ t} are independent. . . T} ? What is the probability that S :::. n = 1. An Use the strong law of large numbers to show.4 Poisson processes Poisson processes are some of the simplest examples of continuoustime Markov chains. respectively.2 Let Tl.3. . a Poisson process is the natural probabilistic model for any uncoordinated stream of discrete events in continuous time. . We shall also see that they may serve as building blocks for the most general continuoustime Markov chain.. T 2 . which provides three different descriptions of a Poisson process. . and then subject only to the condition sUPn An < 00. A2.4 Poisson processes Proof.. We have 73 from which the identity follows by symmetry. Is the condition sUPn An < 00 absolutely necessary? 2. Exercises D 2.1 Suppose Sand T are independent exponential random variables of parameters Q and (3 respectively.4.3.2. T? Show that the two events {S < T} and {min{S. be independent exponential random variables of parameter A and let N be an independent geometric random variable with JP>(N Show that T = n) = (3(1  (3)nl. = 2::1 Ti has exponential distribution of parameter A(3.2.3. 2. that JP> (f Sn = n=l 00) = 1. The key result is Theorem 2... So we shall study Poisson processes first.. first in the special case = 1 for all n. Moreover.
to set Jo = 0.o4l~""''. We shall begin with a definition in terms of jump chain and holding times (see Section 2. : O. 1. .: .... . : ..2).. ~ . A A A A Q= By Theorem 2.... : . I n = 81 + . .. : ~ .. . : .... Here is the diagram: A A A A 234 The associated Qmatrix is given by 1 o .2 (or the strong law of large numbers) we have P(Jn ~ 00) = 1 so there is no explosion and the law of (Xt)t.. + 8 n and then set 5 .. 8 2 . .. : : ..~o is uniquely determined. • . •00········ ' .... .. .~o with values in {O... . Continuoustime Markov chains I theorems and examples that follow.. 2 ...4 t . A simple way to construct a Poisson process of rate A is to take a sequence 8 1 . .74 2...... : : 0. ... } is a Poisson process of rate A (0 < A < 00) if its holding times 8 1.. :.2.••ec.······ .. .. .3.... : .. .. ... ... .'_""""_" Jo = (} J.. of independent exponential random variables of parameter A... .> . . 1 . 4 3 : : :.. .. are independent exponential random variables of parameter A and its jump chain is given by Yn = n. 8 2. A rightcontinuous process (Xt)t. ..
Theorem 2. we shall see in Section 6. for any s ~ 0. . independent of(Xr : r:::. .2. Condition on 8 1 .4. Hence. Proof. Set X t = X s +t .1 (Markov property). leads to a memoryless property of the Poisson process. . ..82 .. s 81 . 8 n = 8 i +n for n ~ o ill( s Recall that the holding times 8 1 . . Theorem 2. and independent of 8 1 . 81 . are themselves independent E(X). conditional on {Xs = i}... Let (Xt)t~O be a Poisson process of rate X.82 . conditional on {X s = i}. It suffice~to prove the claim conditional on the event X s each i ~ O.82 .. D In fact.. (X s+t . (s . are independent E(A)..Xs)t~O is also a Poisson process of rate X.1.. Hence. . . 81.. then by the memoryless property of 8 i +1 and independence.5. of (Xt)t~O are given by 81 = 8 i +1 as shown in the diagram.4 Poisson processes 75 The diagram illustrates a typical path.8i and {X s = i}.3. (Xt)t~O is a Poisson process of rate A and independent of (X r : r ~ s). are independent E(X). . We have = i. by an argument in essentially the same spirit that the result also holds with s replaced by any stopping time T of (Xt)t~o.8i .J i ). We now show how the memoryless property of the exponential holding times.X s .. s).. .82 . Then.. for On this event Xr = and the holding times L j==1 i l{srSt} for r ::.
.X t ~ 1) lP(Xt+h .4. (a) => (b) If (a) holds. (c) (transition probability definition) (Xt)t~O has stationary independent increments and.e).h = Ah + o(h). (XT +t .3. We say that (Xt)t~O has independent increments if its increments over any finite collection of disjoint intervals are independent. rightcontinuous integervalued process starting from O. Let 0 < A < 00.X t = 1) = Ah + o(h).4. Theorem 2.XT )t~O is also a Poisson process of rate A. which gives two conditions equivalent to the jump chain/holding time characterization which we took as our original definition. Let (Xt)t~O be an increasing. then. independent of(Xs:s~T). h ~ 0. we can consider its increment X t .82 . If (Xt)t~O satisfies any of these conditions then it is called a Poisson process of rate A. by the Markov property.76 2. ..2 (Strong Markov property). the increment X t +h . as h ! 0. IF(Xt +h . (b) (infinitesimal definition) (Xt)t~O has independent increments and. of (Xt)t~O are independent exponential random variables of parameter A and the jump chain is given by Yn = n for all n. IF(Xt +h .X t ~ 2) = lP(Xh ~ = lP(Xh ~ 1) 2) = lP(J1 = lP(J2 ~ h) ~ h) = 1 .X t = 0) = 1 . ~ IF(81 ~ h and 8 2 ~ h) = (1  e. conditional on T < 00.X s over any interval (s. Continuoustime Markov chains I Theorem 2. We come to the key result for the Poisson process. If (Xt)t~O is a realvalued process. Proof.Ah )2 = o(h). uniformly in t. X t has Poisson distribution of parameter At. Thus we have three alternative definitions of the same process. Let (Xt)t. So (Xt)t~O has independent increments and as h ! 0 lP(Xt+h . t].~o be a Poisson process oErate A and let T be a stopping time oE(Xt)t~o. We say that (Xt)t~O has stationary increments if the distribution of X s+t .X t has the same distribution as X h and is independent of (X s : s ~ t). Here is some standard terminology.. Then the following three conditions are equivalent: (a) (jump chain/holding time definition) the holding times 8 1 ..X s depends only on t ~ O. for any t. which implies (b). for each t. Then.Ah + o(h).
which implies (c).4.i) (1 . Set Pj(t) = P(Xt = j). for any s. D The differential equations which appeared in the proof are really the forward equations for the Poisson process. Since X o = 0 we have initial conditions PO(O) = 1. If (Xt)(~O satisfies (b). Hence X t rv P(At). Since this estimate is uniform in t we can put t = s . . we have P(Xt + h . So if one process satisfying (c) also satisfies (a). this system of equations has a unique solution given by (At)j Pj(t) = e.2.. for i = 2..1. then certainly (Xt)(~O has independent increments. J.. Ther:.?h Now let h ! 0 to see that Pj(t) is first continuous and then differentiable and satisfies the differential equation By a simpler argument we also find p~(t) = APO(t). .. .. Pj(O) = 0 for j = 1.Ah + o(h) )Pj(t) + (Ah + o(h))Pjl (t) + o(h) so Pj(t + h~ . . uniformly in t.X s rv P(At).X t = i) = o(h) as h ! 0. so must every process satisfying (c).. But condition (c) determines the finitedimensional distributions of (Xt)t~O and hence the distribution of jump chain and holding times. To make this clear.At _.3.2. so the above argument shows X s +t .2.1..4 Poisson processes 77 (b) ::::} (c) If (b) holds.. for j = 1. . Pj(t + h) = P(XHh = j) = = L P(X i==O j Hh  X t = i) P(Xt = j ._.Xs)t~O satisfies (b). (c) ::::} (a) There is a process satisfying (a) and we have shown that it must then satisfy (c). but also (X s+t . j = 0.h to obtain for all s. . consider the .Pj(t) = >'Pj(t) + >'Pjl(t) + O(h).2.. then.. As we saw in Example 2..
(A + Jl)h + o(h). uniformly in t.X t = O)P(yt+h .4.Jlh + o(h). as h ! 0. writing Pi as a reminder.Jlh + o( h)) + (1 . P~j(t) = Pji(t).X t = l)lP(yt+h . Pij(O) = 8ij or. Theorem 2. Then.3 contains a great deal of information about the Poisson process of rate A. according to which (Xt)t~O and (yt)t~O have independent increments and. We shall use the infinitesimal definition. P(O) = I.jl(t) .X t == 1) == Ah + o(h). respectively. P(Xt+h .78 2. Proof. (Zt)t~O has independent increments and. since (Xt)t~O and (yt)t~O are independent. P'(t) = P(t)Q.yt = 0) = (1 . It can be useful when trying to decide whether a given process is a Poisson process as it gives you three alternative conditions to check. Then. P(¥t+h .yt = 1) = Jlh + o(h). = APi.Jlh + o(h)) = 1 . If (Xt)t~O and (yt)t~O are independent Poisson processes of rates A and Jl. P(Zt+h . Hence (Zt)t~O is a Poisson process of rate A + Jl.Ah + o(h))(l . in matrix form.X t == O)lP(yt+h . uniformly in t. D .¥t = 0) = 1 . Set Zt = X t + yt. P(Xt+h .Ah + o( h) )(Jlh + o( h)) = (A + Jl)h + o(h). On the other hand it can also be useful when answering a question about a given Poisson process as this question may be more closely connected to one definition than another. and it is likely that one will be easier to check than another. and we could rewrite the PiO(O) = 8iO . Theorem 2. P(¥t+h . then (Xt + yt)t~O is a Poisson process of rate A + Jl.APij(t). for Q as above.yt == 0) + lP(Xt+h .4. Continuoustime Markov chains I possibility of starting the process from i at time 0. by spatial homogeneity Pij(t) differential equations as p~O(t) = APiO(t). For example.yt == 1) == (Ah + o( h) ) (1 .Ah + o(h). and set Pij(t) = Pi(Xt = j). you might like to consider the difficulties in approaching the next result using the jump chain/holding time definition.Zt = 0) = P(Xt+h .X t = 0) == 1 .4.Zt = 1) = P(Xt+h . lP(Zt+h . as h ! 0.
.Sn+l ~O} so the jump times J1. ...In+1 have joint density function So for A ~ jRn we have IF((J1. for o ~ u ~ t. The results say in effect that the jumps of a Poisson process are as randomly distributed as possible. the jump times J 1 . IF(J1 ~ u I X t = IF(J1 ~ u and X t = l)/lF(Xt = = IF(Xu = 1 and X t . .4 Poisson processes 79 Next we establish some relations between Poisson processes and the uniform distribution. Let (Xt)t~O be a Poisson process.. . conditional on {Xt = n}. . .4.. 1) = 1) D Theorem 2.. conditional on (Xt)t~O having exactly one jump in the interval [s. s + t].. Notice that the conclusions are independent of the rate of the process considered. 0 . The holding times S1.tn)EA 1{O~tl~ .6.. Let (Xt)(~O be a Poisson process. We shall use the finitedimensional distribution definition.In have joint density function f(t1....Xt as required.+Sn+l) 1 {Sl ....Xt An 1 = n) = IF((J1. the jump times J 1 ... Then.Xu = O)/lF(Xt = 1) = Aue'xue. ~tn~t}.. Then.X(tu) /(Ate'xt) = u/t. dtn and since IF(Xt = n) = e. conditional on the event {Xt = n}. Theorem 2. . it suffices to consider the case s = O. . . .. . s + t]. By stationarity of increments.Sn+1 have joint density function An+1 e'x(Sl +. . Proof... Then. Thus.. .2. .. . the time at which that jump occurs is uniformly distributed on [s...4. ~tn~t}dt1 An In! we obtain .. .In have the same distribution as an ordered sample of size n from the uniform distribution on [O. Proof.t].In ) E A and X t = e.5. .. .tn ) = n! l{o~tl~ .In ) E A and I n ~ t < I n+1) (tl ..
k)/JP>(Rt + B t = n) ppn k f3 CP f3 = (3k) . The times spent waiting for the first robin or blackbird are independent exponential random variables 8 1 and T 1 of parameters p and (3 respectively.4. By Theorem 2. Exercises 2.4. Show also (from the same definition and without assuming the strong Markov property) that and hence that J 2 .)/ + )(p+(3)n) k! (n .4. what is the distribution of the number of blackbirds seen in time t? t is a Poisson process By the infinitesimal characterization. then the distribution of the number of blackbirds is binomial of parameters nand {3/ (p + (3).k)! n! (e (e (e = (~) (p~~)k (p:~)nk so if n birds are seen in time t. Continuoustime Markov chains I We finish with a simple example typical of many problems making use of a range of properties of the Poisson process.4 the total number of birds seen in an interval of duration t has Poisson distribution of parameter (p + (3)t.7 Robins and blackbirds make brief visits to my birdtable. What is the probability that the first two birds I see are both robins? What is the distribution of the total number of birds seen in time t? Given that this number is n. So a robin arrives first with probability p/(p + (3) and. Finally JP>(Bt = k I R t + Bt = n) = JP>(Bt = k and R t = n . the probability that the first two birds are robins is p2 / (p + (3)2. whereas the corresponding probability for blackbirds is (3h+o(h).80 2. and the number of blackbirds is a Poisson process (Bt)t~O of rate (3. . by the memoryless property of T 1 .1 State the transition probability definition of a Poisson process. Show directly from this definition that the first jump time J1 of a Poisson process of rate A is exponential of parameter A. Example 2. the number of robins seen by time (Rt)t~O of rate p.J1 is also exponential of parameter A and independent of J1. The probability that in any small interval of duration h a robin will arrive is found to be ph+o(h).
.4 A radioactive source emits particles in a Poisson process of rate A.5 Birth processes A birth process is a generalization of a Poisson process in which the parameter A is allowed to depend on the current state of the process. and its jump chain is given by Y n == i + n. . We begin with a definition in terms of jump chain and holding times. What is the distribution of the number of particles recorded in time t? 2. The particles are each emitted in an independent random direction. qi+1. A minimal rightcontinuous process (Xt)(~O with values in {O. .2.. 1.. (b) when an island in the middle of the road makes it safe to stop halfway? 2.5 Birth processes 81 2. then.4.5 A pedestrian wishes to cross a single lane of fastmoving traffic. at any time whilst crossing. how long on average does it take to cross over safely? [Consider the time at which the first car passes.3 Arrivals of the Number 1 bus form a Poisson process of rate one bus per hour. .2. a car would pass in either direction). respectively. are independent exponential random variables of parameters qi. its holding times 8 1 . . and arrivals of the Number 7 bus form an independent Poisson process of rate seven buses per hour.1.4. conditional on X o == i.] How long on average does it take to cross two similar lanes (a) when one must walk straight across (assuming that the pedestrian will not cross if. } U {oo} is a birth process of rates (qj : j ~ 0) if.2. What.4. (a) What is the probability that exactly three buses pass by in one hour? (b) What is the probability that exactly three Number 7 buses pass by while I am waiting for a Number I? (c) When the maintenance depot goes on strike half the buses break down before they reach my stop..2 Show directly from the infinitesimal definition that the first jump time J 1 of a Poisson process of rate A has exponential distribution of parameter A. Suppose the number of vehicles that have passed by time t is a Poisson process of rate A.. A Geiger counter placed near the source records a fraction p of the particles emitted.4. and suppose it takes time a to walk across the lane. Assuming that the pedestrian can foresee correctly the times at which vehicles will pass by. is the probability that I wait for 30 minutes without seeing a single bus? 2. The data for a birth process consist of birth rates 0 ~ qj < 00. 2.82 . where j == 0.
82
2. Continuoustime Markov chains I
ql
o
~
1
2
. . ..
q2
~
q3
3
4
.
The flow diagram is shown above and the Qmatrix is given by:
Q=
Example 2.5.1 (Simple birth process) Consider a population in which each individual gives birth after an exponential time of parameter X, all independently. If i individuals are present then the first birth will occur after an exponential time of parameter iA. Then we have i + 1 individuals and, by the memoryless property, the process begins afresh. Thus the size of the population performs a birth process with rates qi = iX. Let X t denote the number of individuals at time t and suppose X o = 1. Write T for the time of the first birth. Then E(Xt ) = E(XtlT~t)
=
it
+ E(Xt lT>t)
,xeASlE(X
tI
T = s )ds
+ e At •
Put Jl(t)
= E(Xt ), then E(Xt IT = s) = 2Jl(t  s), so
p,(t)
=
it
2,xe AS p,(t  s)ds + e At
and setting r
=t s
By differentiating we obtain
so the mean population size grows exponentially:
2.5 Birth processes
83
Much of the theory associated with the Poisson process goes through for birth processes with little change, except that some calculations can no longer be made so explicitly. The most interesting new phenomenon present in birth processes is the possibility of explosion. For certain choices of birth rates, a typical path will make infinitely many jumps in a finite time, as shown in the diagram. The convention of setting the process to equal 00 after explosion is particularly appropriate for birth processes!
8
· ~
. . : :
:.. ~ .:•..................
7
6
;
·
:
.
:
:
:.. : .•..................
:.. ~ •...................
; ·
. .
:
5
4 3
:
· ~
:
:
"
:
:
:.. ~
~.:::
.
.
.
· :
~
. .
:
......0. : .:.:
2
..............• lIQo"" .:.. : .:.:: ~
.
1 ....................••.....e~
: :.. :.::
.
t
041~e'""'"""&....&o&_..
Jo == 0
In fact, Theorem 2.3.2 tells us exactly when explosion will occur. Theorem 2.5.2. Let starting from o. (i) If
(Xt)(~O
be a birth process of rates
(qj
:
j
> 0),
I: 00 00
1
< 00, then P(( < 00)
= 00,
= 1.
j==O qj
1
(ii) If:E j==O qj
then P(( = 00) = 1.
D
Proof. Apply Theorem 2.3.2 to the sequence of holding times 8 1 ,82 ,. . ..
The proof of the Markov property for the Poisson process is easily adapted to give the following generalization.
84
2. Continuoustime Markov chains I
Theorem 2.5.3 (Markov property). Let (Xt)(~O be a birth process of rates (qj : j 2:: 0). Then, conditional on X s = i, (Xs+t)(~O is a birth process of rates (qj : j ~ 0) starting from i and independent of (X r : r ~ s).
We shall shortly prove a theorem on birth processes which generalizes the key theorem on Poisson processes. First we must see what will replace the Poisson probabilities. In Theorem 2.4.3 these arose as the unique solution of a system of differential equations, which we showed were essentially the forward equations. Now we can still write down the forward equation
P'(t)
or, in components
= P(t)Q,
P(O)
= I.
and, for j = 1, 2, ...
Moreover, these equations still have a unique solution; it is just not as explicit as before. For we must have
which can be substituted in the equation
P~ 1 ( t) = PiO (t )qo  Pi 1 ( t )q1 ,
Pi 1 ( 0)
= 8i 1
and this equation solved to give
Now we can substitute for Pi1(t) in the next equation up the hierarchy and find an explicit expression for Pi2(t), and so on.
Theorem 2.5.4. Let (Xt)t~O be an increasing, rightcontinuous process with values in {O, 1,2, ... } U {oo}. Let 0 ~ qj < 00 for all j ~ o. Then the following three conditions are equivalent: (a) (jump chain/holding time definition) conditional on X o = i, the holding times 8 1 ,82 , ... are independent exponential random variables of parameters qi, qi+1, . .. respectively and the jump chain is given by Y n = i + n for all n;
2.5 Birth processes
85
(b) (infinitesimal definition) for all t, h ~ 0, conditional on X t = i, X t+h is independent of (X s : s ~ t) and, as h ! 0, uniformly in t,
IF(Xt + h lP(Xt + h
= i I X t = i) = 1  qih + o(h), = i + 1 I X t = i) = qih + o(h);
(c) (transition probability definition) for all n = 0,1,2, ... , all times to ~ ... ~ tn+l and all states io, ... ,in+l
°
~
where (pij(t) : i,j = 0,1,2, ... ) is the unique solution of the forward equations.
If (Xt)t~O satisfies any of these conditions then it is called a birth process of rates (qj : j ~ 0).
Proof. (a) => (b) If (a) holds, then, by the Markov property for any t, h ~ 0, conditional on X t = i, X t + h is independent of (X s : s ~ t) and, as h ! 0, uniformly in t,
lP(Xt + h ~ i
+ 1 I X t = i) = lP(Xh
~ h I X o = i)
~ i
= IF(Jl
and
=
1
+ 1 I X o = i) e qih = qih + o(h),
IF(Xt + h ~ i
+ 2 I X t = i) = IF(Xh
~ i
+ 2 I X o = i)
~ h
= IF( J2 ~ h I X o = i) ~ IF(Sl ~ hand S2 = (1  e qih )(l  e qi + 1h ) = o(h),
which implies (b). (b) => (c) If (b) holds, then certainly for k = i
I X o = i)
+ 2, i + 3, ...
IF(Xt + h
Set Pij(t)
= k I X t = i) = o(h)
as h
! 0, uniformly in t.
1,2, ...
= IF(Xt = j I X o = i).
j
Then, for j
=
Pij(t
+ h) = IF(Xt + h = j I X o = i)
=
LJP>(Xt = k I X o = i)JP>(XHh = j
k==i
IX t
= k)
= Pij(t)(l 
qjh + o(h))
+ Pi,jl(t)(qjlh + o(h)) + o(h)
86
so
2. Continuoustime Markov chains I
As in the proof of Theorem 2.4.3, we can deduce that Pij(t) is differentiable and satisfies the differential equation
By a simpler argument we also find
Thus (Pij(t) : i,j = 0,1,2, ... ) must be the unique solution to the forward equations. If (Xt)t;~o satisfies (b), then certainly
but also
(Xtn+t)t;~o
satisfies (b), so
by uniqueness for the forward equations. Hence (c) =* (a) See the proof of Theorem 2.4.3.
Exercise
(Xt)t;~o
satisfies (c).
D
2.5.1 Each bacterium in a colony splits into two identical bacteria after an exponential time of parameter X, which then split in the same way but independently. Let X t denote the size of the colony at time t, and suppose X o = 1. Show that the probability generating function ¢(t) = E(zX t ) satisfies
Make a change of variables u = t  s in the integral and deduce that d¢/dt = X¢(¢  1). Hence deduce that, for q = 1  e At and n = 1,2, ...
which will occupy the remainder of this chapter and the whole of the next.6. 0 This procedure is best thought of row by row. (iii) L qij = 0 for all i. This provides the most direct mathematical description.6 Jump chain and holding times 87 This section begins the theory of continuoustime Markov chains proper.2. JEI We will sometimes find it convenient to write qi or q( i) as an alternative notation for qii. (ii) qij ~ 0 for all i =I j. which we shall describe. This is only impossible when the offdiagonal entries are all 0. where possible. Let I be a countable set. The jump matrix IT = (7rij : i. It also makes possible a number of constructive realizations of a given Markov chain. We are going to describe a simple procedure for obtaining from a Qmatrix Q a stochastic matrix IT. Recall that a Qmatrix on I is any matrix Q = (qij : i. The approach we have chosen is to introduce continuoustime chains in terms of the joint distribution of their jump chain and holding times. then we leave them alone and put a 1 on the diagonal. j E I) of Q is defined by 7rij 7rii = = qij / qi { 0 if j if j =I i =I i and qi and qi =I 0 = 0. the offdiagonal entries in the ith rovJ of Q and scale them so they add up to 1. For each i E I we take. j E I) which satisfies the following conditions: (i) 0 ~ qii < 00 for all i.1 The Qmatrix Q = (~2 ~1 ~) 2 1 3 . the associated diagram transforms into a discretetime Markov chain diagram simply by rescaling all the numbers on any arrows leaving a state so they add up to 1. {O 1 ~f qi 1= If qi = O. Example 2.6 Jump chain and holding times 2. As you will see in the following example. The basic data for a continuoustime Markov chain on I are given in the form of a Qmatrix. putting a 0 on the diagonal. and which underlie many applications.
q(Yn1) respectively.x and generator matrix Q if its jump chain (Yn)n~O is discretetime Markov(. .Sn are independent exponential random variables of parameters q(Yo).x. its holding n times Sl. Q) for short. . . conditional on yo. IT) and let T 1 ..2. IT) and if for each n ~ 1.see Section 2.1 .x. A minimal rightcontinuous process (Xt)t~O on I is a Markov chain with initial distribution .. We can construct such a process as follows: let (Yn)n~O be discretetime Markov(. ••• be independent exponential random . Recall that a minimal process is one which is set equal to 00 after any explosion . .88 has diagram: 2.Y . T 2 . Continuoustime Markov chains I 1 3 1 2 The jump matrix IT of Q is given by II = ( ~ 2/3 1/2 o 1/3 and has diagram: 1 3 3 1 2 Here is the definition of a continuoustime Markov chain in terms of its jump chain and holding times. We say (Xt)t~O is Markov(.x.
2. inductively for n = 0.2. You will need to understand these constructions in order to identify processes in applications which can be modelled as Markov chains. . Sn+l = y: Jr'I. as required. In more mathematical terms.1.1. conditional on Y n = i.8n . j E I) of independent exponential random variables of parameter 1.. Then. and with an array (T~ : n ~ 1. . if Y n = i we set 8~+1 = T~+l/qij. Y n+1 has distribution (7rij : j E I). 89 = T n /q(Yn. This construction n shows why we call qi the rate of leaving i and qij the rate of going from i to j. .x... n 1 00 00.6 Jump chain and holding times variables of parameter 1. Set 8 n I n = 8 1 + . 8 n+1 is exponential of parameter qi = Ej:j=i qij. the random variables 8~+1 are independent exponentials of parameter qij for all j =I i.. nj I n+1 = inf{t > I n : =I Nj: j for some j #lyn } ° Nr Yn + 1 = j { if I + < . 00 ~~ S~+l' n+1  _ {j 'l . independent of (Yn)n~O. 1 if 8~+1 = 8 n + 1 < ·f S n+1 = 00.x is based on the Poisson process. + 8 n and Yn if I n ~ t < I n+1 for some n X t = { 00 otherwise.2.3. by Theorem 2. . Suppose the door giving access to chamber j from chamber i opens at the jump times of a Poisson process of rate qij and you take every chance to move that you can.. then you will perform a Markov chain with Qmatrix Q. for j =I i. (N..3.. Here is the second construction.1 ). and 8 n + 1 and Y n + 1 are independent. rather than proceeding first via the jump matrix. . Imagine the statespace I as a labyrinth of chambers and passages. and with a family of independent Poisson processes {(N. and Njnj n+l =I Njnj n 'l If I n + 1 = . . We shall now describe two further constructions.i =I j}. each passage shut off by a single door which opens briefly from time to time to allow you through in one direction only.j)t~o : i.x. So. .j E I.. We begin with an initial state X o = Yo with distribution . Then set J o = and define inductively for n = 0. . Then (Xt)t~O has the required properties. conditional on Yn = i.. . Our third and final construction of a Markov chain with generator matrix Q and initial distribution . Both constructions make direct use of the entries in the Qmatrix.Y and 8 1 . we begin with an initial state X o = Yo with distribution . and independent of yo.j)t~o having rate qij. Then.
are independent E(l) and independent of (Yn)n~O. then T 1 . Set Tn = q(Yn. moren over. q = SUPi qi < 00 and 00 .. conditional on T < 00 and XT = i.Y and 8 1 .j)t~o is exponential of parameter qij. Y1 has distribution (7rij : j E I). .7 Explosion We saw in the special case of birth processes that. 8 2 . Then explode if anyone of the following conditions holds: (i) I is finite. . Proof. especially when the statespace is infinite.l) =1= (i. J 1 is exponential of parameter qi = Lj#i qij.7. one can run through a sequence of states with shorter and shorter holding times and end up taking infinitely many jumps in a finite time.j : s ~ t)..j. In cases (i) and (ii).j). So. conditional on I n < 00 and Y n = i. and J 1 and Y1 are independent. Y n+1 has distribution (7rij : j E I). .. Recall the notation of Section 2. In particular. Let (Xt)t~O be Markov(X. . Now suppose T is a stopping time of (Xt)t~o.l) =1= (i. and 8 n +1 and Y n +1 are independent.j : s ~ T). Q). 8 n+1 is exponential of parameter qi. the explosion time ( is given by Sn n==l Theorem 2. then {T ~ t} depends only on (N. Hence.. Q) and. So. .j := N!}+t . Continuoustime Markov chains I The first jump time of (N. we can take T = I n to see that. o (= supJn = n 2: 00 (Xt)t~O does not iEI (iii) X o = i. by Theorem 2.8n · Hence (Xt)t~O is Markov(X.1. and holding times 8 1 . If we condition on X o and on the processes (Ntkl)t~O for (k. T 2 . . J 1. . (XT+t)t~O has the same distribution as (Xt)t~O and is independent of (X s : s ~ T). (Xt)t~O has the strong Markov property.90 2. although each holding time is strictly positive.. by the strong Markov property of the Poisson process N.. and independent of yo. conditional on Yo = i. (ii) sup qi < 00.2: for a process with jump times Jo. . which are independent of N. This phenomenon is called explosion..3.3. J2.N!} is a Poisson process of rate qij independent of (N.j). so we shall not rely on this third construction in the development of the theory.. and independent of X o and (Ntkl)t~O for (k. The conditioning on which this argument relies requires some further justification..1 )Sn. . 2. and i is recurrent for the jump chain.
+! = 00 We say that a Qmatrix Q is explosive if. • •• . The result just proved gives simple conditions for nonexplosion and covers many cases of interest.1. by the Markov property of the jump chain at time n = 1. Here as in Chapter 1 we denote by Pi the conditional probability Pi(A) = P(AIXo = i).2. then for all i. The time and place of the first jump are independent. conditional on XJ 1 = k.Q) and independent of J 1 • So and . Theorem 2. (XJl+t)t~O is Markov(8k.7.7. Q). for the associated Markov chain Pi (( < 00) > 0 for some i E I. Proof. we know that often. ifz also satisfies (i) and (ii).2. Then Z = (Zi : i E I) satisfies:  (i) IZil ~ 1 for all i. Then qi( ~ with probability 1. but these are not as easy to apply as Theorem 2. J 1 is E(qi) and Moreover. It is a simple consequence of the Markov property for (Yn)n>O that under Pi the process (Xt)t>o is Markov(8i . Zi ~ Zi (ii) Qz = Oz. N 2 . Condition on X o = i. say. In case (iii). As a corollary to the next result we shall obtain necessary and sufficient conditions for Q to be explosive. Fix () > 0 and set Zi = Ei(e(}(). at times N 1 . 7 Explosion with probability 1. Otherwise Q is nonexplosive. D (Yn)n~O 91 visits i infinitely L m==l 00 TN". Moreover. Let (Xt)t~O be a continuoustime Markov chain with generator matrix Q and write ( for the explosion time of (Xt)t>o.
D Corollary 2. in particular for all i. By the theorem. Then qikZk (0 .9() = 0 for all i. then. Note that the same argument also shows that Suppose that z also satisfies (i) and (ii).9 () = o.7. then by the theorem E i (e. For each (J > 0 the following are equivalent: (a) Q is nonexplosive.qii)Zi =L k#i so OZi = Lqikzk kEI and so z satisfies (i) and (ii). On the other hand. D . By monotone convergence as n ~ 00.9 (). Suppose inductively that z. Proof. If (a) holds then JP>i(( = 00) = 1 so E i (e.9Jn ) ~ ~ then.9Jn ) for all n. (b) Qz = (Jz and IZil ~ 1 for all i imply z = o. < E·(e. so Zi ~ Zi for all i. since z satisfies (ii) Hence Zi ~ E i (e. Qz = (Jz and Izl ~ 1 imply Zi ~ E i (e. if (b) holds.3.92 2. Continuoustime Markov chains I qi Recall that = qii and q(Jrik = qik. by symmetry z ~ 0. and hence (b) holds. so JP>i(( = 00) = 1 and (a) holds. hence z ~ 0.
write down a necessary and sufficient condition for (Xt)t. For example. Theorem 2. starting from 0. In fact. Let (Xt)t~O be Markov(A. In this section we shall obtain two more ways of characterizing a continuoustime Markov chain. we shall give the strong Markov property as this is a fundamental result and the proof is not much harder.~o to be explosive. Q) and let T be a stopping time of (Xt)t~o.2. starting from 0. that X t hits i. A > Jl and A < Jl and E~l 1/ qi = 00. As for Poisson processes and birth processes. conditional on T < 00 and XT = i. E~ll/qi = 00.il ~ 2. (XT+t)t~O is Markov(8i . qi. the proof of both results really requires the precision of measure theory.8. Why is this condition necessary for (Xt)t~O to be explosive for all Jl E [0. (b) the expected total time spent in state i.00] is a stopping time of (Xt)t~O if for each t E [0. Then.1 Let (Xt )t. However.~o be a Markov chain on the integers with transition rates and qij = if Ij .00) the event {T ~ t} depends only on (X s : s ~ t). We shall present first a version for the case of finite statespace.8 Forward and backward equations Although the definition of a continuoustime Markov chain in terms of its jump chain and holding times provides a clear picture of the process. (Xt)t~O is nonexplosive if and only if one of the following conditions holds: ° qi. we might wish to calculate IPi(Xt = j). where there is a . 2. where A + Jl = 1 and qi > for all i.il = Jlqi ° (i) (ii) (iii) A = Jl.1/2)? Show that. so we have deferred it to Section 6. in general.7. Q) and independent of (Xs : s ~ T). it does not answer some basic questions.1 (Strong Markov property). Theorem 2. Find for all integers i: (a) the probability.1 on the Poisson process gives the main idea in a simpler context.4.5.8 Forward and backward equations Exercise 93 2. We come to the key result for continuoustime Markov chains. which will in particular give us a means to find IPi(Xt = j). In the case where Jl = 0.i+l = Aqi. the first step is to deduce the Markov property from the jump chain/holding time definition. Recall that a random variable T with values in [0. If you want to understand what happens.
. Then the following three conditions are equivalent: (a) (jump chain/holding time definition) conditional on X o = i. Y1 = j. just as for the Poisson process. . is the distribution of X o. 2.eqih)7rijeqjh = qijh + o(h). P(o) = I. where (Pij (t) : i. IT) and for each n ~ 1. Let (Xt)t. h ~ 0. Let Q be a Qmatrix on I with jump matrix II. .. . as h ! 0.8. (b) (infinitesimal definition) for all t. t ~ 0) is the solution of the forward equation P'(t) = P(t)Q.~o be a rightcontinuous process with values in a finite set I. for all j (c) (transition probability definition) for all n = 0. uniformly in t.. . (a) =* (b) Suppose (a) holds.94 2. . Thus for every state j there is an inequality .Y .8n n are independent exponential random variables of parameters q(Yo). If (Xt)t.. j E I. . Xt+h is independent of (X s : s ~ t) and. . the jump chain (Yn)n~O of (Xt)t~O is discretetime Markov(8 i . conditional on X t = i.. . Continuoustime Markov chains I simpler proof.2.. and for j =I i we have JP>i(Xh = j) ~ JP>(J1 ~ h. ~ t n +1 and all states io. the holding times 8 1 ... Theorem 2...in +1 °::....1 ) respectively.. 8 2 > h) = (1 . .1 . then.~o satisfies any of these conditions then it is called a Markov chain with generator matrix Q. conditional on yo. . Proof. In this case there are three alternative definitions. Q) for short.. all times to ~ t 1 ~ .q(Yn . 1. We say that (Xt)(~O is Markov(>. as h ! 0. where>.
. ! 0.Pij(t) = LPik(t)qkj + O(h) kEI so. then and. Indeed. then differentiable on the left. Then by the Markov property. kEI Pij(O) = 8ij .3. as h = lFi(Xt = j) = IF(Xt = j I X o = i). and satisfies the forward equations ° p~/t) = LPik(t)qkj. kEI Since I is finite we have Pij(t + h~ . as h ! 0. X t + h is independent of (X s : s ~ t) and. Pij(t) is then the unique solution by Theorem 2. h ~ 0. moreover.2. hence differentiable.1 that for I finite the forward and backward equations have the same solution.h in the above and let h ! to see first that Pij(t) is continuous on the left.1. if (b) holds. proving (c).1.4.1. letting h ! 0. by the above argument. there is a slight variation of the argument from (b) to (c) which leads directly to the backward equation. D We know from Theorem 2. for any t. Then by uniformity we can replace t by t . uniformly in t (b) :::} (c) Set Pij(t) for all t. then Pij(t + h) = LlPi(Xt = k)IP(XHh = j = I X t = k) LPik(t)(8kj + qkjh + o(h)). uniformly in t kEI If (b) holds. (c) :::} (a) See the proof of Theorem 2. Also. conditional on X t = i. h ~ 0. So in condition (c) of the result just proved we could replace the forward equation with the backward equation. (b) holds for (Xtn+t)(~O so. we see that Pij(t) is differentiable on the right.8 Forward and backward equations 95 and by taking the finite sum over j we see that these must in fact be equalities. Since I is finite.
1 no longer apply. the qualifications minimal and nonnegative being understood. t + h) : i. Note that if I is finite we must have P(t) = etQ by Theorem 2.96 2.t) (I+t~ +o(~))n. On the other hand. t n 4 eq as n 4 00.j E I) of differentiable functions satisfying this system of differential equations. Then the backward equation P'(t) = QP(t).8.4. Some care is needed in making this precise. t ~ O.1. Continuoustime Markov chains I The deduction of (c) from (b) above can be seen as the matrix version of the following result: for q E lR we have (1 + ~ + o( ~) ) Suppose (b) holds and set then P(t.~). Theorem 2. are not a priori identical. . P(O) =I only now we have an infinite system of differential equations P~j(t) = L qikPkj(t). t) = etQ . + h) + h) = P(Xt +h = j I X t = i). A solution to the backward equation is any matrix (pij(t) : i.3. Let Q be a Qmatrix.j E I) satisfies P(t.1. This solution forms a matrix semigroup P(s)P(t) = P(s + t) for all s. We call (P(t) : t ~ 0) the minimal nonnegative semigroup associated to Q. We turn now to the case of infinite statespace. since the o(h) terms.. though uniform in t. We shall prove this result by a probabilistic method in combination with Theorem 2. Here is the key result for Markov chains with infinite statespace.t)=p(o. or simply the semigroup of Q. The backward equation may still be written in the form P'(t) = QP(t).8.. p((n~l)t. in (c) we see that P(O. There are just two alternative definitions now as the infinitesimal characterization become problematic for infinite statespace. P(O) =I has a minimal nonnegative solution (P(t) : t 2: 0). kEI Pij(O) = 15ij and the results on matrix exponentials given in Section 2. = (Pij(t. t and Pij(t.~)p(~. t + h) = I + Qh + o(h) = P(O.
So and JP>i(J1 ::. XJl k#i = k. . II) and for each n ~ 1. . .1) = eQit8ij + 2: k#i Jo t qieQiS7rikPkj(t .1 ) respectively. ..3 and 2.s)ds.4. t < J1) + LJP>i(J1 ::. Proof of Theorems 2.2.. . Let (Xt)(~O be a minimal rightcontinuous process with values in I.. We know that there exists a process (Xt)t~O satisfying (a). X t = j) (2.1 . Q).1. Then conditional on J 1 = sand X J1 = k we have (X8+t)t~O rv Markov(8k.s)ds.8. .s in each of the integrals.8. Step 1.. (b) (transition probability definition) for all n = 0. t.. il. . the jump chain (Yn)n~O of (Xt)t~O is discretetime Markov(8i. Let Q be a Qmatrix on I with jump matrix IT and semigroup (P(t) : t ~ 0).. Make a change of variable u = t ..2. Then the following conditions are equivalent: (a) (jump chain/holding time definition) conditional on X o = i. XJI = k. all times 0 ~ to ~ tl ~ .Y ..8n n are independent exponential random variables of parameters q(Yo). We say that (Xt)t~O is Markov (A. t.8 Forward and backward equations 97 Theorem 2. ~ t n+l and all states io. .i n+l If (Xt)t~O satisfies any of these conditions then it is called a Markov chain with generator matrix Q. Conditional on X o = i we have J 1 rv E(ql) and X J1 rv (7rik : k E I). JP>i(Xt = j.q(Yn. . conditional on yo. Q) for short. .4.2) .8. So let us define P(t) by Pij(t) =JP>i(Xt =j).. (2. .. X t = j) = Therefore Pij(t) = I t qieQiS7rikPkj(t . where A is the distribution of X o. the holding times 8 1 .. interchange sum and integral by monotone convergence and multiply by eqit to obtain eQitpij(t) = 8ij + Jo tL k#i qieQiU7rikPkj(u)du. We show that P(t) satisfies the backward equation.
Then. it also satisfies the integral form: If P(t) 2 0. j. Let us suppose inductively that for all i. Secondly.3) P~j(t) = L qikPkj(t) kEI so P(t) satisfies the backward equation. and hence Pij (t) is differentiable in t and satisfies eqit(qiPij(t) + P~j(t)) = qieqit7rikPkj(t). L k#i Recall that obtain qi = qii and qik = qi'lrik for k =I i.5) we have for all i.1) to (2. The integral equation (2. Step 2. then.1) is called the integral form of the backward equation.t < I n +l) = eqit15ij +L k=li Jo rt qieQiS7rikJP>k(Xts = j. We show that if P(t) is another nonnegative solution of the backward equation. . then P(t) ~ P(t). firstly. (2.98 2. by reversing the steps from (2. j and t.3).4) On the other hand. j and t. The argument used to prove (2. if P(t) satisfies the backward equation. hence P(t) is the minimal nonnegative solution. that Pij (t) is continuous in t for all i.4) and (2.1) also shows that JP>i(Xt =j. on rearranging. the integrand is then a uniformly converging sum of continuous functions. t  s < In)ds. j and t. then for all i. Continuoustime Markov chains I This equation shows. then by comparing (2. hence continuous. we (2.
by a probabilistic argument resembling Step 1 of the proof of Theorems 2.8..8. The forward equation may still be written P'(t) = P(t)Q. We shall show that the semigro~p (P( t) : t ~ 0) of Q does satisfy the forward equations. . The argument is a little longer because there is no reversetime Markov property to give the conditional distribution. as we have throughout. Since (Xt)t~O does not return from 00 we have Pij(S + t) = lPi(Xs +t = j) = LlPi(XS +t = j kEI I X s = k)lPi(Xs = k) = :ElPi(Xs = k)lPk(Xt = j) = :EPik(S)Pkj(t) kEI kEI by the Markov property.3 and 2.4.2.. Hence (P(t) : t ~ 0) is a matrix semigroup. This completes the proof of Theorem 2. Remember that the finite statespace results of Section 2. kEI Pij(O) = 8ij .3.3. 99 Step 3. instead of conditioning on the first event. This time. we condition on the last event before time t.4 by the usual argument that (b) must now imply (a) (see the proof of Theorem 2.4. Then.4. P(O) = I. Suppose.j E I) of differentiable functions satisfying this system of equations.1 are no longer valid. . Step 4. (c) :::} (a)).tn) so (Xt)t~O satisfies (b). that by the Markov property (Xt)t~O satisfies (a). j and t. We need the following timereversal identity. Hence for all i. .8.8 Forward and backward equations and the induction proceeds. D So far we have said nothing about the forward equation in the case of infinite statespace. a simple version of which was given in Theorem 2. We complete the proof of Theorem 2. A solution is then any matrix (pij(t) : i.8.3.Xtn = in) = Pin (Xtn+ltn = in+1) = Pi nin+l (tn+l . P(Xtn +1 = i n+1 IXto = io. now understood as an infinite system of differential equations p~/t) = LPik(t)qkj.
100
2. Continuoustime Markov chains I
Lemma 2.8.5. We have
qinJP(Jn ~ t < I n+ 1 I Yo = io, Y 1 = i 1, ... ,Y = in) n = qioJP(Jn ~ t < I n+1 I Yo = in,· .. , Y n  1 = iI, Y n
= io).
Proof· Conditional on Yo = i o, ... ,Y = in, the holding times 8 1 , ... ,8n+ 1 n are independent with 8k rv E(qik_l). So the lefthand side is given by
Jt:!..(t)
{
qi n exp{ qi n (t 
Sl 
· ·· 
sn)}
k=l
IT
qikl exp{ qikl Sk}dsk
where Ll(t) = {(SI, ... ,sn) : SI + ... + Sn ~ t and SI, ... ,8 n ~ O}. On making the substitutions Ul = t  81  ...  Sn and Uk = 8 n k+2, for k = 2, ... ,n, we obtain
qinJP(Jn ~ t < I n+ 1 I Yo
=
= i o,·
.. ,Y = in) n
Jt:!..(t)
(
qio exp{ qio(t  U1  · ..  Un)}
~
k=l
IT
qink+l exp{ qink+l Uk}duk
= io).
= qioJP(Jn
t < I n+ 1 I Yo
= in, . ..
,Y  1 = iI, Y n n
D
Theorem 2.8.6. The minimal nonnegative solution (P(t) : t 2: 0) of the backward equation is also the minimal nonnegative solution of the forward equation P'(t) = P(t)Q, P(O) = I.
Proof. Let (Xt)t~O denote the minimal Markov chain with generator matrix Q. By Theorem 2.8.4
00
= L
LJP>i(Jn ::; t < In+i, Y n 1 = k, Y n = j).
~
n==O k=j:j
Now by Lemma 2.8.5, for n
1, we have
JPi(Jn ~ t < I n+ 1 I Y n  1 = k, Y n = j) = (qi/qj)JPj(Jn ~ t < I n+ 1 I Y1 = k, Y n = i)
= (qi/Qj)
= qi
I
it
qjeQjBJP>k(Jn_1 ::; t 
S
< I n I Yn1 = i)ds
t
e QjB (qk/qi)JP>i (In1 ::; t  s < I n I Yn  1 = k)ds
2.8 Forward and backward equations
101
where we have used the Markov property of (Yn)n~O for the second equality. Hence
Pij(t) = 8ijeqit
+
f
00
'2: i t JP>i(Jnl ::; t 0
X
s
1
< I n I Yn 
1
= k)
n=lki=i
IPi(Yn 
= k, Yn = j)Qkeqj8ds
= 8ijeqit + L
= 8ijeqit
L
io
0
ft
JP>i(Jnl ::; t 
s < I n , Yn  1 =
k)qk'lrkjeqjSds
n=lki=i
+
io
t '2:Pik(t k#j
s)qkjeqjSds
(2.6)
where we have used monotone convergence to interchange the sum and integral at the last step. This is the integral form of the forward equation. Now make a change of variable u = t  s in the integral and multiply by eqjt to obtain (2.7) We know by equation (2.2) that eqitpik(t) is increasing for all i, k. Hence either '2:Pik(U)qkj converges uniformly for u E [0, t]
ki=i
or
LPik(U)qkj =
ki=j
00
for all u ~ t.
The latter would contradict (2.7) since the lefthand side is finite for all t, so it is the former which holds. We know from the backward equation that Pii (t) is continuous for all i, j; hence by uniform convergence the integrand in (2.7) is continuous and we may differentiate to obtain
P~j(t) + Pij(t)qj
= '2:Pik(t)qkj. ki=i
Hence P(t) solves the forward equation. To establish minimality let us suppose that Pij(t) is another solution of the forward equation; then we also have
102
2. Continuoustime Markov chains I
~
A small variation of the argument leading to (2.6) shows that, for n
0
Pi(Xt = j, t < I n + 1 )
= 8ijeqit +
If P(t) ~ 0, then
L
(t IP\(Xt = j, t < In)qkjeQjBds.
(2.8)
k#jJo
P(Xt = j, t < J o) = 0
Let us suppose inductively that
~ Pij (t)
for all i, j and t.
then by comparing (2.7) and (2.8) we obtain
and the induction proceeds. Hence
Exercises 2.8.1 Two fleas are bound together to take part in a ninelegged race on the vertices A, B, C of a triangle. Flea 1 hops at random times in the clockwise direction; each hop takes the pair from one vertex to the next and the times between successive hops of Flea 1 are independent random variables, each with with exponential distribution, mean 1/ A. Flea 2 behaves similarly, but hops in the anticlockwise direction, the times between his hops having mean 1/ Il. Show that the probability that they are at A at a given time t > 0 (starting from A at time t = 0) is
2.8.2 Let
(Xt)t;~o
Iln = nil, and assume that X o = 1. Show that h(t) =
be a birthanddeath process with rates An = nA and P(Xt = 0) satisfies
2.9 Nonminimal chains
and deduce that if A =I Jl then
103
2.9 Nonminimal chains
This book concentrates entirely on processes which are rightcontinuous and minimal. These are the simplest sorts of process and, overwhelmingly, the ones of greatest practical application. We have seen in this chapter that we can associate to each distribution A and Qmatrix Q a unique such process, the Markov chain with initial distribution A and generator matrix Q. Indeed we have taken the liberty of defining Markov chains to be those processes which arise in this way. However, these processes do not by any means exhaust the class of memoryless continuoustime processes with values in a countable set I. There are many more exotic possibilities, the general theory of which goes very much deeper than the account given in this book. It is in the nature of things that these exotic cases have received the greater attention among mathematicians. Here are some examples to help you imagine the possibilities.
Example 2.9.1
Consider a birth process (Xt)t~O starting from 0 with rates qi = 2i for i ~ We have chosen these rates so that
00 00
o.
Lq:;l = L2i < 00
i=O i=O
which shows that the process explodes (see Theorems 2.3.2 and 2.5.2). We have until now insisted that X t = 00 for all t 2 (, where ( is the explosion time. But another obvious possibility is to start the process off again from o at time (, and do the same for all subsequent explosions. An argument based on the memoryless property of the exponential distribution shows that for 0 ~ to ~ . . . ~ t n + 1 this process satisfies
for a semigroup of stochastic matrices (P(t) : t ~ 0) on I. This is the defining property for a more general class of Markov chains. Note that the chain is no longer determined by A and Q alone; the rule for bringing (Xt)t~O back into I after explosion also has to be given.
104 Example 2.9.2
2. Continuoustime Markov chains I
We make a variation on the preceding example. Suppose now that the jump chain of (Xt)t~O is the Markov chain on Z which moves one step away from o with probability 2/3 and one step towards 0 with probability 1/3, and that Yo = O. Let the transition rates for (Xt)t~O be qi = 21il for i E Z. Then (Xt)t~O is again explosive. (A simple way to see this using some results of Chapter 3 is to check that (Yn)n~O is transient but (Xt)t~O has an invariant distribution  by solution of the detailed balance equations. Then Theorem 3.5.3 makes explosion inevitable.) Now there are two ways in which (Xt)t~O can explode, either X t ~ 00 or X t ~ 00. The process may again be restarted at 0 after explosion. Alternatively, we may choose the restart randomly, and according to the way that explosion occurred. For example if X t if X t
~ 00 ~ 00
i( as t i (
as t
where Z takes values ±1 with probability 1/2. Example 2.9.3 The processes in the preceding two examples, though no longer minimal, were at least rightcontinuous. Here is an altogether more exotic example, due to P. Levy, which is not even rightcontinuous. Consider for n
~
0
and set I = UnD n . With each i E D n\Dn 1 we associate an independent exponential random variable Si of parameter (2 n )2. There are 2n  1 states in (Dn\Dnl) n [0,1), so, for all i E I
and
Now define if
L Sj :::; t < L Sj for some i E I
j<i
j~i
otherwise.
Between any two distinct states i and j it makes infinitely many visits to 00. 1967). IQl + Q21 ~ IQll + IQ21 and IQIQ21 ~ IQIIIQ21· Proof. 2. Lemma 2. San Francisco. Chichester. v#O IIQlloo = sup Iqijl· i. We shall show that the two norms are equivalent and that IQI is well adapted to sums and products of matrices. 1994). In particular. . C. For further reading. Rogers and D. with 1 in the jth place. We have (a) (b) IIQlloo ~ IQI ~ NIIQlloo. 2nd edition. 2nd edition. we have IQcjl ~ IQI. Williams (Wiley. for the vector Cj = (0.j Obviously. or Diffusions. G. for 0 ~ to ~ .L... at j say. Freedman (HoldenDay.1. Vol 1: Foundations by L. IIQlloo is finite for all Q and controls the size of the entries in Q. It remains in i E D n \Dn .10 Appendix: matrix exponentials Define two norms on the space of realvalued N x N matrices IQI = sup IQvl/lvl. .. which IIQlloo is not. .2. . or else Markov Chains with Stationary Transition Probabilities by K. so IIQII~ ~ L(qij)2 = IQcjl2 i ~ IQI 2...0). The Lebesgue measure of the set of times t when X t = 00 is zero. The details may be found in Markov Chains by D.1 for an exponential time of parameter 1.10 Appendix: matrix exponentials 105 This process runs through all the dyadic rationals i E I in the usual order. . 1971). Markov Processes and Martingales. Berlin. We hope these three examples will serve to suggest some of the possibilities for more general continuoustime Markov chains. (a) For any vector v we have IQvl ~ IQllvl.. There is a semigroup of stochastic matrices (P(t) : t ~ 0) on I such that. see Freedman's book.10. P(Xt = 00) = 0 for all t ~ O. Chung (Springer. The supremum defining IIQlloo is achieved.1. ~ tn+l In particular.
.E(m)lloo ~ IE(n) . so Qk < Since '" .k!· LJ k=m+l IQI ~ NllQlloo < 00. IQIQ2 V ~ IQI11Q2 V l ~ IQIIIQ21Ivl· l D Now for n = 0.1.n ~ 00. by the CauchySchwarz inequality (b) For any vector v we have I(QI + Q2)vl ~ IQI V I + IQ2 V l ~ (IQll + IQ21)1vl. n k=m+l ' " . Continuoustime Markov chains I IQvI 2 = ~ (~qijVj) 2 ~ ~ ( ~ IIQllooIVjl) = 2 NIIQII~ ( ~ IVjl) 2 and. .k! LJ IQl k ~O as m. and m ~ = L k=O n Qk kf' n.. consider the finite sum E(n) For each i and j.2. .E(m)1 n k=m+l n Lkf IQl k E~o IQl k /k! converges by the ratio test. we have I(E(n) .106 On the other hand 2.E(m))iil ~ IIE(n) . .
for two commuting matrices Q1 and Q2 we have .)  ~ (tQ)fj L.. Finally.2.j. _ '1. proving that Qk e Q = L kf k=O CX) is well defined and. that the power series (e t Q ) .10 Appendix: matrix exponentials 107 Hence each component of E(n) forms a Cauchy sequence. indeed. which therefore converges.J k! k=O has infinite radius of convergence for all i.
. this chapter should look reassuringly familiar. (iii) L qij = 0 for all i. Note that II is a stochastic matrix. if qi if qi =I 0 = o. Recall that a Qmatrix on I is a matrix Q = (qij : i. All the facts from Chapter 2 that are necessary to understand this synthesis are reviewed in Section 3. 3.1. (ii) qij 2 0 for all i =I j.3 Continuoustime Markov chains II This chapter brings together the discretetime and continuoustime theories. You will require a reasonable understanding of Chapter 1 here. for continuoustime chains. jEI We set qi IT = (7rij : = q(i) = qii.j E I) satisfying the following conditions: (i) 0 ~ qii < 00 for all i. _ ~~  {o qij / qi 0 1 if j if j =I i =I i and qi =I 0 and qi = 0.1 Basic properties Let I be a countable set. Associated to any Qmatrix is a jump matrix i. .. Exercises remain an important part of the text. allowing us to deduce analogues. given such an understanding.j E I) given by 1r. of all the results given in Chapter 1. _ { ~J 7r. but.
. forgetting what has gone before. being held in these states for times 8 1 . The first.. The Qmatrix is known as the generator matrix of (Xt)t~O and determines how the process evolves from its initial state. in terms of jump chain and holding times. given that the chain starts at i. It then starts afresh. 1 jEI for all Associated to any Qmatrix is a semigroup (P(t) : t 2 0) of substochastic matrices P(t) = (Pij(t) : i. .j E I). E I) with non LPij ::. Thus I n = 8 1 + . . We established in Section 2. Briefly... it waits there for an exponential time of parameter qi and then jumps to a new state. lim L. . .· . The distribution X gives the initial distribution. + 8 n .. .. As the name implies we have P(s)P(t) = P(s + t) for all s. (b) conditional on Yo = i o. The discretetime process (Yn)n~O is the jump chain.Y .1 = inI. states that (a) (Yn)n~O is Markov(X. a rightcontinuous process (Xt)t~O runs through a sequence of states Yo.. .J n=l For a minimal process we take a new state 00 and insist that X t = 00 for all t 2 (. jump times. (8n )n>1 are the holding times and (In )n>1 are the jump times. .. IT).. jump chain and explosion. t 2 o. ways to describe how the process evolves. . but equivalent. . The data for a continuoustime Markov chain (Xt)t~O are a distribution X and a Qmatrix Q. holding times.3. the distribution of X o. Y2 .. choosing state j with probability 1rij.82 .83 .2: minimal rightcontinuous random process. Y1 .1 Basic properties A substochastic matrix on I is a matrix P negative entries and such that 109 = (Pij : i. An important point is that a minimal rightcontinuous process is determined by its jump chain and holding times. .. You will need to be familiar with the following terms introduced in Section 2. Put more simply.qin 1 . J 2 . the holding times 8 1 .Sn n are independent exponential random variables of parameters qi o ' . J 3 . The explosion time ( is given by 00 ( = '""" 8 n = n+oo I n . respectively and jumping to the next state at times J1..8 that there are two different. j i..
.. .. Continuoustime Markov chains II The second description. But there is no danger of confusion in using the simpler notation. that is.i n + 1 ~ to ~ tl ~ . ~ tn+l and all states Again. P(t) is simply the matrix exponential etQ . where ~ and P(h) are defined on IU{oo}.7. The semigroup is also the minimal nonnegative solution of the forward equation P'(t) = P(t)Q.x and P(h) by ~oo = 0 and Pooj(h) = O. Note that we have not yet said how the semigroup P(t) is associated to the Qmatrix Q. . It then starts afresh. in the explosive case. states that the finitedimensional distributions of the process are given by (c) for all n = 0. all times 0 i o. we should say Markov(~. In the case where I is finite. This description implies that for all h > 0 the discrete skeleton (Xnh)n~O is Markov(. P(O) = I which reads in components P~j(t) = L qikPkj(t).1. The information coming from these two descriptions is sufficient for most of the analysis of continuoustime chains done in this chapter. . . in terms of the semigroup. So we recall from Section 2. .:EPij(t) > 0 jEI the chain is found at 00 with probability Pi 00 (t). except via the process! This extra information will be required when we discuss reversibility in Section 3. In the case where Pioo(t) := 1 .110 3. when P(t) is strictly substochastic. forgetting what has gone before.. iI. and is the unique solution of the backward and forward equations. put more simply. The semigroup P(t) is referred to as the transition matrix of the chain and its entries Pij (t) are the transition probabilities. .2. P(O) = I. P(h)). P(h)).x. kEI Pij(O) = 8ij . by time t it is found in state j with probability Pij (t).8 that the semigroup is characterized as the minimal nonnegative solution of the backward equation P'(t) = QP(t). extending. given that the chain starts at i. Strictly..
.. We say that i leads to j and write i ~ j if lPi(Xt =j for some t ~ 0) > o. as discussed in Section 1. . If (ii) holds.2 Class structure 111 A first step in the analysis of a continuoustime Markov chain (Xt)t~O is to identify its class structure. absorbing state and irreducibility are inherited from the jump chain.2.. '1ri n li n > 0. Theorem 3. then Pij (t) ~ Pioil (tin) . Proof. = i. then by Theorem 1. and (iv) holds. (iii) Q oilqili2 . (iv) Pij(t) > 0 for all t > 0.1. so if (iii) holds.2.in with i o i in = j. in = j and '1rioi l '1rili2 .2.in with io = i. and then only periodically. Pinlin (tin) > 0 for all t > 0. . For distinct states i and j the following are equivalent: (i) i ~j. Implications (iv) => (v) => (i) => (ii) are clear. but only after a certain length of time.3 Hitting times and absorption probabilities Let (Xt)t>o be a Markov chain with generator matrix Q. there are states io. We emphasise that we deal only with minimal chains. The notions of communicating class. .2 Class structure 3. 3... D Condition (iv) of Theorem 3.1. closed class.3. Then the class structure is simply the discretetime class structure of the jump chain (Yn)n~O. then for all t > 0. . The hitting time of a subset A of I is the random variable D A defined by . (v) Pij(t) > 0 for some t > o.. If qij > 0. i I . (ii) i ~ j for the jump chain. those that die after explosion. Qinli n > 0 for some states io. where it may be possible to reach a state.1 shows that the situation is simpler than in discretetime.2...iI. We say i communicates with j and write i ~ j if both i ~ j and j ~ i... which implies (iii).
112 3. 1 2 2 (Xt)t~O to reach A is given 1 1 2 3 3 3 4 . Continuoustime Markov chains II with the usual convention that inf 0 = 00. D The average time taken.3) in terms ofQ. then  {H A < oo} and on this set we have = {D A < oo} D A = JHA.3. starting from i. Theorem 3. So if H A is the hitting time of A for the jump chain. Apply Theorem 1. We emphasise that (Xt)t>o is minimal.1.3. Since the hitting probabilities are those of the jump chain we can calculate them as in Section 1. starting from i.3. for by In calculating kf we have to take account of the holding times so the relationship to the discretetime case is not quite as simple. hf is called the absorption probability.2 to the jump chain and rewrite (1. that (Xt)t~O ever hits A is then When A is a closed class. The vector of hitting probabilities h A = (hf : i E I) is the minimal nonnegative solution to the system of linear equations Proof. The probability.
then D A = 0.Lqijk1 JEI = 1. k3 = + ~k1 + ~k2. Then kf = Yi = 0 for i E A. The proof follows the same lines as Theorem 1. A.2 Consider the Markov chain (Xt)t~O with the diagram given on the preceding page. The vector of expected hitting times k A = (kt : i E I) is the minimal nonnegative solution to the system of linear equations kt { =0 1 for i E A .1) Proof. then D A ~ J 1 . If X o = i E A.1).1).3. Assume that qi > 0 for all i fj.l+L j#i j#i 'lrij k 1 and so . First we show that k A satisfies (3. On starting in 1 we spend an average time ql = 1/2 in 1. Suppose now that Y = (Yi : i E I) is another solution to (3. A.5. then Yi = qi + ~ 'lrijYj = qi + ~ 'lrij j~A j~A 1 '"' 1 '"' (1 + '"' ) qj ~ 'lrjkYk k~A = Ei(Sl) + Ei (S2 1{HA:2:2}) + L L'lrij'lrjkYk. A. Thus 1 k1 = ~ + ~ k 2 + ~ k3 ~ and similarly k2 = ~ + ~k 1 + ~k3. so kt = o. so by the Markov property of the jump chain so kt = Ei(D A ) = Ei(J1 )+ LE(DAJ1 I Y1 = j)IP\(Yl = j) = q. (3.LjEI qijkf = for i fj.3. On solving these linear equations we find k 1 = 17/12. How long on average does it take to get from 1 to 4? Set ki = E i (time to get to 4). then jump with equal probability to 2 or 3. Here is the general result. Theorem 3.3 Hitting times and absorption probabilities 113 Example 3. j~A k~A .3.3. If X o = i fj. A.3. Suppose i fj.
3.4 Recurrence and transience Let (Xt)t~O (Xt)t~O be Markov chain with generator matrix Q. 2. We say a state i is recurrent if JP>i({t ~ 0: X t = i} is unbounded) = 1. like class structure.. Yi ~ E i (D A) = Exercise kf.1 Consider the Markov chain on {I. Now so. by monotone convergence. Continuoustime Markov chains II By repeated substitution for y in the final term we obtain after n steps Yi = lEi(Sd + . · 1rindnYin· So.114 3. D 3.3. Recall that we insist be minimal. .. 3... 4} with generator matrix Q= (0~ 1/6 o 1/2 1/2 1/2 o o o o 1/3 1/6 1~4) o. recurrence and transience are determined by the jump chain. Note that if (Xt)t~O can explode starting from i then i is certainly not recurrent. if y is nonnegative where we use the notation HA /\ n for the minimum of H A and n. + lEi (Sn 1{HA:2:n}) + L . as required.. Calculate (a) the probability of hitting 3 starting from 1. We say that i is transient if JP>i( {t ~ 0 : Xt = i} is unbounded) = O. The next result shows that. L jl~A jn~A 1riil . (b) the expected time to hit 4 starting from 1.
(iii) Apply Theorem 1. Also X(Jn ) = Y n = i infinitely often. D The next result gives continuoustime analogues of the conditions for recurrence and transience found in Theorem 1. If qi = 0.4.3. then i is recurrent and (ii) ifqi > 0 and IPi(Ti < 00) < 1.2.1 and so that i is recurrent if and only if Pii(t)dt = the corresponding result for the jump chain. Write 1r}j) for the (i. (Yn)n~O. Jo oo 00.4 Recurrence and transience Theorem 3. because (Yn : n ~ N) cannot include an absorbing state. (iv) recurrence and transience are class properties. The following dichotomy holds: (i) if qi = 0 or IPi(Ti < 00) = 1. which is finite. (i) Suppose i is recurrent for (Yn)n~O. 115 then i is recurrent for (ii) if i is transient for the jump chain. n=O 00 (n) 1rii (3.5. defined by Ti(w) = inf{t ~ J 1 (w) : Xt(w) = i}. Proof.5. Theorem 3.2) by Theorem 3. We denote by T i the first passage time of (Xt)t~O to state i. If X o = i then (Xt)t~O does not explode and I n ~ 00 by Theorem 2.3 to the jump chain. so {t ~ 0 : X t = i} is unbounded.7.4. Let N i denote for all t. by Theorem 3.4. Proof. (iii) every state is either recurrent or transient.1 and the corresponding result for the jump chain. (iv) Apply Theorem 1.4 to the jump chain. . Then Jo IPi(Ni < 00) = IPi(Ti < 00) so i is recurrent if and only if IPi(Ti < 00) = 1. Suppose then that qi > o. We shall show that 1 o 00 1 pii(t)dt . and the first passage time of the jump chain (Yn)n~O to state i.4. (ii) Suppose i is transient for (Yn)n~O.1. with probability 1.:_ L q1. Pii(t) = 1 oo Pii(t)dt = 00.5.3. Pii(t)dt < 00. N = sup{ n ~ 0 : Yn = i} < so {t ~ 0: X t = i} is bounded by J(N +1). so i is recurrent. If X o = i then 00.1. then i is transient for (Xt)t~O. with probability 1. then (Xt)t~O cannot leave i. then i is transient and Jo Jo oo oo Pii(t)dt = 00.j) entry in rrn. We have: (i) if i is recurrent for the jump chain (Xt)t~O.
5. Proof. state A signifying that he is 'in attendance' and state B that he is having a teabreak.4.1 Customers arrive at a certain queue in a Poisson process of rate A.4): = lEi 00 n=O L 00 Sn+l 1{Yn=i} 1 = L lEi (Sn+l I Y n = i)JP>i(Yn = i) = :n=O ~ n=O L 00 1I"i~)· 0 Finally. he fluctuates between these states as a Markov chain Y on {A.4.2. (i) If i is recurrent for (Xt)(~O then i is recurrent for (ii) If i is transient for (Xt)t~O then i is transient for (Zn)n~O. The single 'server' has two states A and B. Continuoustime Markov chains II To establish (3. .3. The total service time for any customer is exponentially distributed with parameter Jl and is independent of the chain Y and of the service times of other customers.3 and 3. B} with Qmatrix (a a) (3 (3 . Theorem 3. Claim (ii) is obvious. Independently of how many customers are in the queue.2) we use Fubini's theorem (see Section 6.4. Then. To prove (i) we use for nh estimate ~ t < (n + l)h the which follows from the Markov property. (Zn)n~O. Let h > 0 be given and set Zn = X nh . we show that recurrence and transience are determined by any discretetime sampling of (Xt)(~o. by monotone convergence and the result follows by Theorems 1. D Exercise 3.116 3.
Proof· We have qi ('lrij .1. Theorem 3.5 Invariant distributions Describe the system as a Markov chain X with statespace 117 An signifying that the server is in state A and there are n people in the queue (including anyone being served) and B n signifying that the server is in state B and there are n people in the queue.2.xiqij iEI iEI = (.xQk D This tieup with measures invariant for the jump matrix means that we can use the existence and uniqueness results of Section 1. and explain why this is intuitively obvious. Explain why. the notions of invariant distribution and measure play an important role in the study of continuoustime Markov chains. . 3.1]. j. prove that X is transient if 1l(3 < A(a+(3). The following are equivalent: (i) A is invariant. We say that A is invariant if AQ=O. so (Jl(II .3.7 to obtain the following result.1. (ii) Il IT = Il where Ili = Aiqi. .8ij ) = :E.1) f (fJ) = 0.I))j = :EJli(1rij . . where By considering f(l) or otherwise..5 Invariant distributions Just as in the discretetime theory.5. Show that (fJ . Let Q be a Qmatrix with jump matrix IT and let A be a measure.bij) = qij for all i.. and k = 0. for some fJ in (0.
(iii) Q is nonexplosive and has an invariant distribution A. Then.J n=O Nil = q.5.2.n<Ni } . Proof. D Recall that a state i is recurrent if qi = 0 or lPi (Ti < 00) = 1.L. we can take Ai = jji/qi to obtain an invariant measure unique up to scalar multiples. by Theorem 3. It is obvious that (i) implies (ii). Moreover.118 3. Let us exclude the trivial case I = {i}. = Ei(Ti 1\ (). As in the discretetime case positive recurrence is tied up with the existence of an invariant distribution. Theorem 3.n<Ni } ) n=O 00 = qJ ~lE· '""" 1{Yn=J. then irreducibility forces qi > 0 for all i. L jEI f. denotes the minimum of T i and (. Proof. by Theorems 1.7. By Theorems 3. If qi = 0 or the expected return time mi = Ei(Ti) is finite then we say i is positive recurrent.L~ = Ei 10 where T i /\ ( fTil\( 1 {xs=j}ds. Otherwise a recurrent state i is called null recurrent. : j E I) by f.4. (ii) some state i is positive recurrent.5 and 1.2. By Fubini's theorem 00 f. Let Q be an irreducible Qmatrix. II has an invariant measure jj.5. By monotone convergence.6. then irreducibility forces qi > 0 for all i. ~ L.l Ei L n=O l{Yn=j} = . So.7.j/qj . when (iii) holds we have mi = l/(Aiqi) for all i.1. which is unique up to scalar multiples. II is irreducible and recurrent.L.n<Nd = LEi(Sn+l I Y n =j)Ei (l{Yn=j. Suppose that Q is irreducible and recurrent. = Ei L n=O 00 Sn+l 1{Yn=j. Denote by N i the first passage time of the jump chain to state i.1 and 3. Continuoustime Markov chains II Theorem 3. Let us exclude the trivial case I = {i}.1. Then the following are equivalent: (i) every state is positive recurrent. Define Jli = (Jl.5.3. Then Q has an invariant measure A which is unique up to scalar multiples.
So mi = LJl.. where qi > 0 for all i and where 0 < A = 1 . for all j by Theorem 1.5.7.6.1.5 Invariant distributions 119 "i where.7../mi. suppose (iii) holds. then Vi = 1 and vll = v by Theorem 3. Vj = .3. = L ..~ . . is the expected time in j between visits to i for the jump chain... Suppose (ii) holds.4 Consider the Markov chain (Xt)(~O on Z+ with the following diagram. On the other hand.. and mi = l/(Aiqi) for all i.. by Theorem 2. . To compute an invariant measure v it is convenient to use the detailed balance equations for all i.1.. D The following example is a caution that the existence of an invariant distribution for a continuoustime Markov chain is not enough to guarantee positive recurrence..1 iI i i+1 The jump chain behaves as a simple random walk away from 0..5. so the jump chain is recurrent.7.. To complete the proof we return to the preceding calculation armed with the knowledge that Q is recurrent.. so Vj ~ . then i is certainly recurrent..5. But Jli has finite total mass L Jl.7.. We know that "ill = by Theorem 1. = Ei(Ti) = jEI mi so we obtain an invariant distribution A by setting Aj = Jl.. Example 3./qj ~ L Vj/Qj jEI jEI jEI = 2: Aj/(Aiqi) = l/(Aiqi) < 00 jEI showing that i is positive recurrent... j. or even recurrence. hence 11 is recurrent.1.. and Q is nonexplosive.Jl < 1: AqO o .5.. so JliQ = 0 by Theorem 3. in the notation of Section 1.. Fix i E I and set Vj = Ajqj/(Aiqi). so (Xt)(~O is recurrent if A ~ Jl and transient if A > Jl.
Consider.3. and from the proof of Theorem 3. If the jump rates qi are constant then v can be normalized to produce an invariant distribution precisely when A < Il. and P(s) is recurrent by Theorem 3. (ii) AP(S) = A. Thus it suffices to show IlP( s) = Il. the only possibility is that (Xt)t~O is explosive. Let Q be irreducible and recurrent. Hence any A satisfying (i) or (ii) is unique up to scalar multiples. Given Theorem 3.7.120 3.5.l(A/Il)i.5.3. Then v has finite total mass so (Xt)t~O has an invariant distribution. In this case the nonzero equations read for all i. but (Xt)t~O is also transient. Theorem 3.3. Continuoustime Markov chains II Look ahead to Lemma 3.7.5.5. on the other hand.P'(s) = >.1. it is nonexplosive by Theorem 2. so IlQ = O. the case where qi = 2i for all i and 1 < AI Il < 2. Let s > 0 be given. Proof. if we fix i and set then IlQ = o.4. For infinite statespace. the interchange of differentiation with the summation involved in multiplication by A is not justified and an entirely different proof is needed.QP(s) so AQ Il P (s) d = 0 implies AP(S) = AP(O) = A for all s. The following are equivalent: (i) AQ = 0. P(s) is also recurrent. so = Il implies that Il is proportional to A. There is a very simple proof in the case of finite statespace: by the backward equation ds>'P(s) = >. and let A be a measure. So a solution is given by Vi = q:. By the strong Markov property at T i (which is a simple consequence of the strong Markov property of the jump chain) . Since Q is recurrent. The next result justifies calling measures A with AQ = 0 invariant.2 to see that any solution is invariant.
only there is no longer any possibility of periodicity. conditional on X s Markov(8i . We have Ipij(t + h)Pij(t)1 = I:EPik(h)Pkj(t) .Pij(t)1 kEI = !:EPik(h)Pkj(t) . Let Q be an irreducible nonexplosive Qmatrix having an invariant distribution A.5. lF i (J1 :s. Q). D .qih . Then.6.1.»i(X kEI i = k.5. D Theorem 3.pii(h) :s. Lemma 3.e. for all i. t L(lE kEI Jo (Ti l{Xt=k}dt)Pk j (s) :E /£kPkj(S) as required. using Fubini's theorem.(1 .6 Convergence to equilibrium We now investigate the limiting behaviour of Pij (t) as t ~ 00 and its relation to invariant distributions.pii(h))pij(t) I k#i :s. IF(Xs = i) = (AP(S))i = Ai so. If (Xt)(~O is Markov(A.6. h) = 1 . D 3.6 Convergence to equilibrium Hence. s +Ti /£j = = 121 1 Jo kEI lEi l 00 s l{X t=j}dt = j. t t JP\(Xs +t < Ti)dt < Ti)Pkj(s)dt = = = roo '2:)J. By Theorem 3.5.3. h 2: 0 Proof. We shall need the following estimate of uniform continuity for the transition probabilities. Proof. Let Q be a Qmatrix with semigroup P(t). 1 . for all t. You will see that the situation is analogous to the case of discretetime. Q) then so is (Xs+t)(~O for any S 2: o. by the Markov property.
we fill in the gaps using uniform continuity.2 (Convergence to equilibrium).2. D The complete description of limiting behaviour for irreducible chains in continuous time is provided by the following result. So. It follows from Theorem 1.5 by the same argument we used in the preceding result. Let Q be an irreducible nonexplosive Qmatrix with semigroup P(t).6. and having an invariant distribution A.5. by discretetime convergence to equilibrium. so that for n 2: N.8.j so P(h) is irreducible and aperiodic. for all i.1. Then for all states i.4 so (Zn)n~O is discretetime Markov(8i . . P(h)). For t 2: Nh we have nh ~ ~ s ~ h t < (n + l)h for some n 2: Nand by Lemma 3. By Theorem 3. By Theorem 3. By Theorem 2. Hence Pij(t)~Aj as n~oo. j we have Proof Let (Xt)(~O be Markov(8i . j Thus we have a lattice of points along which the desired limit holds.8. A is invariant for P(h).1 irreducibility implies pij(h) > 0 for all i. Given € > 0 we can find h > 0 so that for 0 and then find N. We do not give the details. Q).6. Fix h > 0 and consider the hskeleton Zn = Xnh. Continuoustime Markov chains II Theorem 3. Fix a state i.122 3.5.
6. 7 Time reversal 123 Theorem 3.3. Each customer has a service requirement distributed as the sum of two independent exponential random variables of parameter Jl. Write down the generator matrix Q of a continuoustime Markov chain which models this.4}: (a) (c) (11 (1 2 1 1 0 0 1 1 0 0 1 1 1 0 0 0 2 2 11) ]2) (b) (d) (1 (1 2 1 1 0 0 2 1 1 1 1 0 0 2 1 0 2 1 0 ~) ~) 3.7 Time reversal Time reversal of continuoustime chains has the same features found in the discretetime case. Then as t ~ 00 for all j E I where mj is the expected return time to state j. and deduce that the chain is positive recurrent if and only if AI Il < 1/2. Suppose that (Xt)(~O is Markov(v. 3.1. Service requirements are independent of one another and of the arrival process.3 Customers arrive at a singleserver queue in a Poisson stream of rate A. 3.2 In each of the following cases. Exercises 3.6. Reversibility provides a powerful tool in the analysis of Markov chains.1 Find an invariant distribution A for the Qmatrix Q = (~2 2 !4 1 3 ~) and verify that limt~ooPII(t) = Al using your answer to Exercise 2. compute limt~oo P(Xt = 21Xo = 1) for the Markov chain (Xt)t~O with the given Qmatrix on {1.1. explaining what the states of the chain represent.3. Calculate the essentially unique invariant measure for Q. Let Q be an irreducible Qmatrix and let v be any distribution.6. as we shall see in Section 5.2.6. Note in the following .2.3. Q).
Let T E (0.5. and P(t) is then its minimal nonnegative solution.8.. P(O) = I. < t n = T and Sk = tk . the semigroup (P(t) : t 2: 0) of Q is the minimal nonnegative solution of the forward equation P'(t) = P(t)Q. which is itself a Qmatrix. . We could if we wished redefine the timereversed process to equal its right limit at the jump times. Hence Q is irreducible and nonexplosive and has invariant distribution A.. this is unimportant. for 0 = to < .. Pinlin (Sn) SO.. This echoes our proof of the forward equation. by Theorem 2. Finally.k~v(A. then P(t) is an irreducible stochastic matrix with invariant distribution A.00) be given and let (Xt)O<t<T be Markov(A.7.8. For the processes we consider.e Proof. .. Also. and forget about the problem. Continuoustime Markov chains II result how time reversal interchanges the roles of backward and forward equations.Xtn = in) = P(XTto = io. P(Xto = io.8. By Theorem 2.6. But this is the backward equation for Q. for all t > 0. by Theorem 2. Q). A small technical point arises in time reversal: rightcontinuous processes become leftcontinuous processes.8. Let Q be irreducible and nonexplosive and suppose that Q has an invariant distribution A.. P(t) is an irreducible stochastic matrix with invariant distribution A.. Theorem 3.. which rested on the time reversal identity of Lemma 2. A Qmatrix Q and a measure A are said to be in detailed balance if for all i. and we can rewrite the forward equation transposed as P'(t) = QP(t) . Pi l io (Sl) = AioPioi l (Sl) . j. Mo. where Q = (~j : i.4 again. Define P(t) by AjPJi(t) = AiPij(t). thus obtaining again a rightcontinuous process. . Q is also irreducible and nonexplosive with invariant distribution A.tn = in) = AinPininl (sn) .j E 1) is given by Aj~i = Aiqijo over. Set Xt = X T . Q). D The chain (Xt)O~t~T is called the timereversal of (Xt)O~t~T. . We shall suppose implicitly that this is done. Q).. Then the process (Xt)O<t<T is Ma. .4 we have .124 3.tk1. (Xt)O~t~T is Markov(A.t ..1.XT .
Proof· We have (AQ)i = EjEI Ajqji = EjEI Aiqij = 0.7. Find the equilibrium distribution of the number of buses in service. t] has Poisson distribution of mean At. Theorem 3. and the lengths of calls are independent exponential random variables of parameter Jl. set up a Markov chain model for this process. (b) Q and A are in detailed balance. given that n are in use at time 0. Assuming that infinitely many telephone lines are available. is neJLt + A(l . Then both (a) and (b) are equivalent to the statement that Q = Q in Theorem 3.3. (XTt)OstsT is also Markov(A.7. Here is the result. We say that (Xt)(~O is reversible if.2. Q). Show that. Find the mean length of the busy periods during which at least one line is in use. in equilibrium. D Let (Xt)(~O be Markov(A.8 Ergodic theorem 125 Lemma 3. Show that the expected number of lines in use at time t. D Exercise 3. Show that for large t the distribution of the number of lines in use at time t is approximately Poisson with mean AI Jl. with Q irreducible and. for all T > 0. nonexplosive. IfQ and A are in detailed balance then A is invariant for Q.1 Consider a fleet of N buses.7. 3. and for similar reasons. Both (a) and (b) imply that A is invariant for Q.8 Ergodic theorem Longrun averages for continuoustime chains display the same sort of behaviour as in the discretetime case. Let Q be an irreducible and nonexplosive Qmatrix and let A be a distribution. when it is sent to the depot for repair.7. Suppose that (Xt)(~O is Markov(A. Is (Nt)t~o a Poisson process? 3. Proof.eJLt)1 Jl. Then the following are equivalent: (a) (Xt)(~O is reversible.3.1.2 Calls arrive at a telephone exchange as a Poisson process of rate A. . the number Nt of calls finishing in the time interval [0. The repair shop can only repair one bus at a time and each bus takes an exponential time of parameter A to repair.7. Each bus breaks down independently at rate Jl. Q). Q).
. then ]p> (~ t t io l{x s =i}ds ~ _1_ as t miqi ~ 00) = 1 where mi = IEi(Ti ) is the expected return time to state i. are independent and identically distributed with mean mi.. it suffices to consider the case v = bi. t t 1 {~ Jo l{x s =i}ds ~ 0 = 1 mi · Suppose then that Q is recurrent and fix a state i.8.2. So... ~ ~ ~ By the strong Markov property (of the jump chain) at the stopping times Tin for n ~ 0 we find that L}. Denote by Mi the length of the nth visit to i. Moreover.. for any bounded function f : I ~ lR we have where and where (Ai: i E I) is the unique invariant distribution. are independent and identically distributed with mean l/qi... If Q is transient then the total time spent in any state i is finite.+l = inf{t > Tin: X t =1= i} . +M!L 1 ~~qi asn~oo . Then (Xt)(~O hits i with probability 1 and the longrun proportion of time in i equals the longrun proportion of time in i after first hitting i. setting Tp = 0. so t 1 Jo l{x s =i}ds::. by the strong law of large numbers (see Theorem 1. + Lr:t ~ n n ~ mi asn~oo M~ ~ + . and that Ml. .10. . Thus for n = 0. Q). Continuoustime Markov chains II Theorem 3..1) L~ ~ + .Tin T in + 1 = inf{t > Tin + M. . by the strong Markov property (of the jump chain). in the positive recurrent case.1. Proof. Hence. M?. Let Q be irreducible and let v be any distribution. L~ .1 (Ergodic theorem). we have M.T!" . . by Tin the time of the nth return to i and by Li the length of the nth excursion to i..+l : X t = i} L ~+l = T?1'+l .126 3. . If (Xt)(~O is Markov(v.
for Tin ~ t < T in+ 1 we have T!L M~ ~ T!"+l L~ ~ ~ ~ + ···+ M!L <1 ~ + ···+ Lr:t ... by the same argument as was used in the proof of Theorem 1. + L..8 Ergodic theorem and hence M~ ~ 127 + . In particular. D .. we note that Tin /Tin+ 1 ~ 1 as n ~ with probability 1. + M!L ~ L} + ···+ Li ~ 1 asn~oo miqi 00 with probability 1.2.3. with probability 1 In the positive recurrent case we can write where Ai = 1/(miqi).10.. We conclude that 1 t it 0 f(Xs)ds ~ 1 ast~oo with probability 1.t ~ it 0 1 {Xs=~} ·ds<~ T!"+l M~ ~ T!L L~ ~ ~ + ···+ M!"+l ~ + . Now.:+l ~ so on letting t ~ 00 we have.
This is a sort of balancing property. which we shall make precise shortly.1 Martingales A martingale is a process whose average value remains constant in a particular strong sense. the overall structure is. The further theory inevitably involves more sophisticated techniques which. electrical networks and Brownian motion. Often. namely. 4.4 Further theory In the first three chapters we have given an account of the elementary theory of Markov chains. The idea is that the Markov chain case serves as a guiding metaphor for more complicated processes. can obscure the overall structure. already present in the elementary theory. This already covers a great many applications. On the other hand. but is just the beginning of the theory of Markov processes. to a large extent. although having their own interest. We begin with a simple example. further insight is gained into Markov chains themselves. which is a Markov chain with the following diagram . the identification of martingales is a crucial step in understanding the evolution of a stochastic process. So the reader familiar with Markov chains may find this chapter helpful alongside more general higherlevel texts. We therefore thought it worth while to discuss some features of the further theory in the context of simple Markov chains. At the same time. potential theory. Consider the simple symmetric random walk (Xn)n~O on Z. martingales.
. for n E(Xn  ~ m.· . X m I X o = io.. . A third formulation of the martingale property involves another notion of conditional expectation.Xn=in }' . The sequence (Fn)n~o is called the filtration of (Xn)n~O and we think of F n as representing the state of knowledge. An adapted integrable process (Mn)n~o is called a martingale if for all A E F n and all n.. indeed it has the stronger property that the average value of the walk at some future time is always simply the current value. . ..··· .in lE(Y I X o = i o. This stronger property says that (Xn)n~O is in fact a martingale. .. . Let us fix for definiteness a Markov chain and write F n for the collection of all sets depending only on X o.Xn . ... In precise terms we have and the stronger property says that.. . Since the collection F n consists of countable unions of elementary events such as this martingale property is equivalent to saying that for all i o...in and all n. Given an integrable random variable Y. we define lE(Y I F n ) = L io.4.Xn = i n )l{xo=io. .1 Martingales 1 1 129 2 iI 2 i I( • • i+l The average value of the walk is constant. . Here is the general definition.Xm (Xn)n~O = i m ) = o. of the chain up to time n. A process (Mn)n~o is called integmble if EIMnl < 00 for all n.Xn . . . or history. A process (Mn)n~o is called adapted if M n depends only on X o. .
M r nl 1) + ···+ (M1  Mo) = :2) Mk +l k=O Mk)lk<T. } u {(X)} is a stopping time if {T = n} E F n for all n < 00. Recall from Section 1. so if we complete the process and average the conditional expectation we should get the full expectation E(E(Y I F n )) = E(Y). We shall prove one general result about martingales. Proof. Let (Mn)n~o be a martingale and let T be a stopping time. n for some n.4 that all sorts of hitting times are stopping times. for a martingale so. It is easy to check that an adapted integrable process (Mn)n~o is a martingale if and only if Conditional expectation is a partial averaging. . 1. Theorem 4. what we do is to replace on each elementary event A E F n .1. by induction This was already clear on taking A = n in our original definition of a martingale. 2. n} E F n for all n < 00. Further theory The random variable E(Y I F n ) is called the conditional expectation of Y given F n . In particular. the random variable Y by its average value E(Y I A). Assume that (i) holds.. (ii) T < 00 and IMnl ~ C whenever n Then lEMr = lEMo. In passing from Y to E(Y I F n ). ~ T. then see how it explains some things we know about the simple symmetric random walk. An equivalent condition is that {T ::. .1 (Optional stopping theorem).M o = (Mr .130 4. It is easy to check that this formula holds. Recall that a random variable T :n ~ {O. Suppose that at least one of the following conditions holds: (i) T::.. Then M r .
Thus condition (ii) of the optional stopping theorem applies with M n = X n and C = a V b. bEN are given. and so since (Mk)k~O is a martingale. then the preceding argument applies to the stopping time T 1\ n.MTAnl ~ 2CJP>(T > n) for all n.4. So what? Well. We have X T = a with probability p and X T = b with probability 1 . We deduce that IEXT = EXo = O. Markovian.3 the counterintuitive case of a gambler who keeps on playing a fair game against an infinitely rich casino. But the intuition behind the result EXT = 0 is very clear: a gambler.IEMTAnl ~ IEIMT . way to compute p. now we can compute p = JP>(Xn hits a before b). so EMT = EMo.4.p)b giving p = b/(a + b). the average gain should be zero. with the certain outcome of ruin. Hence nl EMT . since the game is fair. If we do not assume (i) but (ii).< T} = {T ~ Martingales 131 k}C E Fk since T is a stopping time. so that IEMTAn = IEMo. D (Xn)n~O. leaves the casino once losses reach a or winnings reach b. There is an entirely different. Then T is a stopping time and T < 00 by recurrence of finite closed classes. so o = EXT = p( a) + (1 . Returning to the simple symmetric random walk X o = 0 and we take T suppose that = inf {n ~ 0: Xn = a or Xn = b} where a.EMo = L E[(Mk+l k=O Mk)lk<T] = O. We discussed in Section 1. whichever is sooner. playing a fair game. using the methods of Section 1. But JP>(T > n) ~ 0 as n ~ 00. This game ends at the finite stopping time T = inf{ n ~ 0 : Xn = a} . Then IIEMT .1 Now {k.IEMol = IIEMT .p.
s~p llil jEI J IM~I ~ 2(n+1)suplfjl < j 00 showing that M~ is integrable for all n. Nevertheless. By the Markov property . Write (Fn)n~o for the filtration of (Xn)n~O. . indeed a martingale is necessarily realvalued and we do not in general insist that the statespace I is contained in JR. Further theory where a is the gambler's initial fortune.1. The example just discussed was rather special in that the chain (Xn)n~O itself was a martingale. Since XT = a we have EXT = a =I 0 = EXo but this does not contradict the optional stopping theorem because neither condition (i) nor condition (ii) is satisfied. Let f be a bounded function.. Obviously.. to every Markov chain is associated a whole collection of martingales. the following process is a martingale: nl (i) (Xn)n~O M~ = f(Xn )  f(Xo)  L (P . this is not true in general. Thus. Then I(PJ)(i)1 = so ILPij1i1 ::. while intuition might suggest that EXT = EXo is rather obvious. given a function f : I ~ JR and a Markov chain (Xn)n~O with transition matrix P. We recall that. some care is needed as it is not always true. Suppose (i) holds. and these martingales characterize the chain. we have (pn J)(i) = Lp~j) Ii = lEi (J(Xn )). m=O Proof.2. (ii) for all bounded functions f : I ~ JR. Let (Xn)n~O be a random process with values in I and let P be a stochastic matrix.132 4.Xn = in}. . This is the basis of a deep connection between martingales and Markov chains. jEI Theorem 4.I)f(X m ). Then the following are equivalent: is a Markov chain with transition matrix P. Let A = {X o = i o.
X o)] (Pf)(n + 1.1)2/2 . for all n ~ 0. Suppose that a function f : N x I ~ lR satisfies.1. Then.Xn D = in) = = E in [f(n + 1.· . On the other hand.i). i) = (i . We consider f (i) = i and g( n.1)2/2 + (i .i) Then M n = LPijf(n+ 1. if (ii) holds. We have assumed that M n is integrable for all n. starting from O. (Pg)(n + 1. both and (Pf)(n+ 1.3. X n ) is a martingale. Xl) . Let (Xn)n~O be a Markov chain with transition matrix P. Theorem 4. in) .1 so Martingales 133 (Pf)(Xn ) E(M~+1 . Let us see how this theorem works in the case where (Xn)n~O is a simple random walk on Z. Hence both X n = f(X n ) and Yn = g(n. . On taking f = l{i n +l} we obtain so (Xn)n~O is Markov with transition matrix P.1)/2 + (i + 1)/2 = i = f(i).n. Since IXnl ~ n for all n. then for all bounded functions f. D Some more martingales associated to a Markov chain are described in the next result. .4. i) = i 2 . X n ) are martingales.. in) = O.n = g(n.j) JEI = f(n. Notice that we drop the requirement that f be bounded. So (Mn)n~o is a martingale.(n + 1) = i2 .M! I A) = E[f(Xn + l )  I A] = 0 and so (Ml)n~o is a martingale.f(n. i).f(n. Proof. we have Also (P f)(i) = (i . = f(n. by the Markov property E(Mn +1  M n I X o = i o.
1. 1991). gives rise to a gravitational potential ¢. By the optional stopping theorem Hence On letting n + 00. 4. For an excellent introduction to martingales and their applications we recommend Probability with Martingales by David Williams (Cambridge University Press. Exercise 4. consider again the stopping time T = inf{n ~ 0 : Xn = a or Xn = b} where a. Show that M n = h(Xn ) is a martingale. Some more will appear in later sections. fluid flow and the diffusion of heat. of density p say. where ~ = 8 2 /8x 2 + 8 2 /8 y 2 through its gradient + 82 / 8z 2 . which is known as potential theory. In gravity. The potential ¢ is felt physically '\l¢ = (8¢ 8¢ 8¢) 8x' 8y' 8z .2 Potential theory Several physical theories share a common mathematical framework. but potential theory is also relevant to electrostatics. the left side converges to E(T).134 4. Set T = inf{n ~ 0: X n E A} and hi = IPi(Xn E A for some n ~ 0) = IPi(T < 00). bEN. One example is Newton's theory of gravity.1 Let (Xn)n~O be a Markov chain on I and let A be an absorbing set in I. and the right side to E(Xf) by bounded convergence. Further theory In order to put this to some use. a distribution of mass. So we obtain We have given only the simplest examples of the use of martingales in studying Markov chains. by monotone convergence. which in suitable units satisfies the equation Da¢ = p.
Obviously. This is a unifying idea. we shall briefly run through some of the main features of potential theory. What is the fair price to move from state 3 to state 4? The fair price is always the difference in the expected total cost. containing within it other notions previously considered such as hitting probabilities and expected hitting times. where space is discrete. but the advantage of wider applicability. An indirect link is provided by Brownian motion. In this section we are going to consider potential theory for a countable statespace.2. Markov chains. This is the easiest way to appreciate the general structure of potential theory. Suppose that on each visit to states i = 1. which at each step choses among the allowable transitions with equal probability. cPs = 0 and . In these examples the potential cP has the interpretation of expected total cost.4. and by showing how to calculate these potentials.4. which has the disadvantage that one loses the intuitive picture of the process. where Ci = i.2 Potential theory 135 which gives the force of gravity acting on a particle of unit mass. obviously have no direct link with this theory. It also finds application when one associates costs to Markov chains in modelling economic activity: see Section 5.1 Consider the discretetime random walk on the directed graph shown above. explaining their significance in terms of Markov chains. which we shall discuss in Section 4. Before we embark on a general discussion of potentials associated to a Markov chain. We denote by cPi the expected total cost starting from i.2. which has much of the structure of the continuum version. The basic ideas of boundary theory for Markov chains will also be introduced. Once we have established the basic link between a Markov chain and its associated potentials. in which space is a continuum. 1 4 2 5 Example 4. here are two simple examples.4 a cost Ci is incurred.3.4. We shall begin by introducing the idea of potentials associated to a Markov chain. unobscured by technical difficulties. This discrete theory amounts to doing Markov chains without the probability.
q4 = 1. Further theory by considering the effect of a single step we see that ¢l = 1 + ¢2. What now is the fair price to move from 3 to 4? The expected cost incurred on each visit to i is given by Ci/ qi and ql = 1. and suppose cost is incurred at rate Ci = i in state i for i = 1. Thus the total cost is now Tl n=O L c(X n ) + f(X T ) . ¢3 = 3 + ~¢l + ~¢4' ¢4 = 4. First suppose our process is. transitions are allowed in both directions.136 4. q3 = 3. states 1 and 5 are absorbing. as before ¢l = 1 + ¢2.3.4. q2 = 1.2. So we see. Obviously. Hence ¢3 = 8 and the fair price to move from 3 to 4 is 4. 4. Thus the total cost is now 00 1 c(Xs)ds. where Ii = i. and a final cost Ii on arrival at i = 1 or 5. the continuoustime random walk (Xt)t~O on the same directed graph which makes each allowable transition at rate 1. instead. ¢2 = 2 + ¢3. Hence ¢3 i + ~¢4' = 5 and the fair price to move from 3 to 4 is 1. We impose a cost Ci = i on each visit to i for i = 2. 1 4 2 5 In the second variation we consider the discretetime random walk (Xn)n~O on the modified graph shown above. ¢2 = 2 + ¢3. ¢3 = + ~¢l ¢4 = 4. 3. We shall now consider two variations on this problem. Where there is no arrow.
as the total cost is a sum over infinitely many times. On solving these equations we obtain cP2 = 7. 5}. Let us look for a finite solution: then so cPo = c cc2 p . and takes negative values in between. we know that the walk X n ~ 00 with probability 1. as before. So in cPo = 00. =9 and cP4 cP4 = 4 + ~(cP3 + cPs).c2pq) = { 00 if c E (q/p. l) otherwise. cP3 = 3 + ~(cPl + cP2 + cP4 + cPs). On moving one step to the right. cP3 this case the fair price to move from 3 to 4 is 2. Let c > 0 and suppose that a cost ci is incurred every time the walk visits state i.4. Hence the expected total cost is given by cPo c/(c.2.c + q has roots at q/p and 1.q .i+l. we see cPo = 1 + PcPl + qcPl = 1 + (cp + q/c)cPo. Then cPl = 1. cPs = 5 and cP2 = 2 + ~(cPl + cP3). Example 4. so we must have By considering what happens on the first step.2 Potential theory 137 where T is the hitting time of {I. Note that cPo = 00 always satisfies this equation.il = q < P = Pi. so for c ~ 1 we shall certainly have = 11.2 Consider the simple discretetime random walk on Z with transition probabilities Pi. We shall see in the general theory that cPo is the minimal nonnegative solution. all costs are multiplied by c. What is the expected total cost cPo incurred by the walk starting from O? We must be prepared to find that cPo = 00 for some values of c. Write. The quadratic c2p . Indeed. Let cPi denote the expected total cost starting from i. . cPi for the expected total cost starting from i.
Nevertheless. Further theory It was clear at the outset that cPo = 00 when c ~ 1. when and if it hits the boundary. Ci ~ 0 for all i E D and Ii ~ 0 for all i E aD. or even that the boundary is nonempty. In the examples just discussed we were able to calculate potentials by writing down and solving a system of linear equations. we shall assume for the most part that c and I are nonnegative. This situation is familiar from hitting probabilities and expected hitting times. Note that we do not assume the chain will hit the boundary. We shall discuss the cases of discrete and continuous time sidebyside. and will reveal also what happens when the linear equations do not have a unique solution. we are now going to give some general results on potentials. We shall consider the associated potential. The most obvious interpretation of these potentials is in terms of cost: the chain wanders around in D until it hits the boundary: whilst in D. at state i say. the expected cost incurred to the left of 0 is infinite. a final cost Ij is incurred. and (Xt)t~O for a continuoustime chain with generator matrix Q. As the examples show. To be sure that the sums and integrals here are well defined. we shall write (Xn)n~O for a discretetime chain with transition matrix P. we call aD the boundary. In the explosive case we always set c( 00) = 0. Throughout. More generally. defined by in discrete time. It is interesting that cPo = 00 also when c is too small: in this case the costs rapidly become large to the left of 0. it incurs a cost Ci per unit time. one does not really need a general theory to write down the linear equations. Let us partition the statespace I into two disjoint sets D and aD. . We suppose that functions (Ci : i E D) and (Ii: i E aD) are given. we insist that (Xt)(~O be minimal. that is. so this assumption is not too restrictive. so no further costs are incurred after explosion. and although the walk eventually drifts away to the right. ¢ is the difference of the potentials associated with the positive and negative parts of c and I. and in continuous time where T denotes the hitting time of aD. As usual. these are simple examples of potentials for Markov chains. Indeed. These will help to reveal the scope of the ideas used in the examples.138 4. at j say.
(ii) if'l/J = ('l/Ji : i E I) satisfies 'l/J { 'l/J ~ ~ P'l/J+c in D in aD I (4. Suppose that (Ci : i E D) and (Ii: i E aD) are nonnegative. For i E D by the Markov property IEi ( L l~n<T c(Xn ) + f(X T )IT<oo Xl c(Xn ) = j) j) = IEj so we have (L + f(XT )lT <OO) = ¢j n<T <Pi = Ci + LpijlE ( L jEI l~n<T c(Xn ) + f(X T )l T <oo Xl = =Ci+ LPWPj jEI as required. (ii) Consider the expected cost up to time n: . (i) Obviously.3. then (4.4. (iii) if IPi(T < 00) solution.2 Potential theory 139 Theorem 4.1) has at most one bounded Proof.2) and 'l/Ji ~ 0 for all i. Then (i) the potential ¢ = (¢i : i E I) satisfies (4. Set where T denotes the hitting time of aD.1) ¢ = P¢+ C in D { ¢=I in aD. then 'l/Ji ~ ¢i for all i.2. ¢ = f on aD. = 1 for all i.
by repeated substitution for 'ljJi 1/J on the right ~ Ci + L jEaD Pijli + LPijCj JED + ···+ 2: jlED . and hence 1/J ~ ¢. Similarly and by induction. in the case of equality..2). .2).2) then with equality if equality holds in (4..2) and 1/J ~ 0 = ¢(O). L jlED Pijl · · · Pjnlin 'ljJjn jnED E i (c(XO)lT>O + !(XdlT=l + C(Xdlr>l + ···+ C(Xn1)lT>nl + !(Xn)lT=n + 'ljJ(Xn)lT>n) = ¢i(n  1) + IEi (1/J(Xn)lT~n) D as required. then asn~oo so 1/J = limn~oo ¢(n) = ¢. 1/J ~ ¢(n) for all n. with equality when equality holds in (4. Further theory i ¢i as n ~ By monotone convergence. (iii) We shall show that if 1/J satisfies (4. if l1/Ji I ~ M and Pi (T < 00) = 1 for all i. by the argument ¢(n + 1) = c + P¢(n) in D in aD. so 1/J ~ ¢(1). Also. For i E D we have 'ljJi ~ Ci + L jEaD Pij!j + LPij'ljJj JED and. But also.. This is another proof of (ii). proving (iii). ¢i(n) used in part (i).140 4. we find 00. { ¢(n + 1) = f Suppose that 1/J satisfies (4.. Then 1/J ~ P1/J + c ~ P¢(O) + c = ¢(1) in D and 1/J ~ f = ¢(1) in aD. 2: jnIED Piil'" Pjn2jnl Cjn_l PijI'" Pjnlin !jn + L jlED '" L L jnIEDjnEaD + = 2: .
2). If the linear equations have a finite nonnegative solution on J.3) may involve subtraction of infinities. and that (Ci : i E D) and (Ii: i E aD) are nonnegative. Then minimal nonnegative solution to cP = (cPi : i E I). which forces cPi = 00. Nevertheless. Consider nl Mn = L c(Xk)lk<T + f(XT )IT<n + 'ljJ(Xn )ln::. For continuoustime chains there is a result analogous to Theorem 4.4. then (cPj : j E J) is the minimal such solution. If not.2. Assume that (Xt)t~O is minimal. We have to state it slightly differently because when cP takes infinite values the equations (4.2).2.3. then cPj = 00 for some j E J. Set where T is the hitting time of aD. Although the conclusion then appears to depend on the finiteness of cP. We note that M n is not necessarily integrable.T' nl k==O Then lE(Mn+l I F n ) = L c(Xk)lk<T + f(XT )lT<n k==O + (P1jJ + c)(Xn )lT>n + I(Xn )lT==n :::.3) . Theorem 4. since i leads to j. we can still use the result to determine cPi in all cases.4.Mn with equality if equality holds in (4. (4. it still follows that with equality if equality holds in (4. and therefore not make sense.2 Potential theory 141 It is illuminating to think of the calculation we have just done in terms of martingales. if finite. which is a priori unknown. is the { QcP = C cP=1 in D in aD. To do this we restrict our attention to the set of states J accessible from i.
4) which equations have at most one bounded solution if IPi(N < 00) = 1 for all i.5. .3) has no finite nonnegative solution. if IPi(T < 00) == 1 for all i. and by IT the jump matrix.142 4. Proof. 1) or rate A E (0.3). then (4. Moreover.2. and since N is finite whenever T is finite. Theorem 4. Potentials with discounted costs may also be calculated by linear equations. ¢ is therefore the minimal nonnegative solution to ¢ { ¢ = IT¢ + c == f in D in aD. Suppose that (Ci : i E I) is bounded.82 .4) are exactly the finite solutions of (4. this proves the result.2.3. We have lE ( ) Sn+l I Yn c(Yn so. . D It is natural in some economic applications to apply to future costs a discount factor a E (0. Set ¢i = lEi then ¢ n==O L anc(Xn ) 00 = (¢i :i E I) is the unique bounded solution to ¢ = aP¢+c. by Fubini's theorem = J") = __ = Cj { 0 Cj / qj if Cj >0 if Cj = 0. ¢i = lEi(L c(Y ) + !(YN)lN<oo). corresponding to an interest rate. Further theory If ¢i = 00 for some i. indeed the discounting actually makes the analysis easier. the jump chain and holding times of (Xt)t~O.. n n<N By Theorem 4. then (4. Denote by (Yn)n>O and 8 1 . and where we use the convention o x 00 = 0 on the right. Then iT c(Xt)dt + !(XT)lT<oo = o L n<N c(Yn)Sn+l + !(YN)lN<oo where N is the first time (Yn)n~O hits aD.3) has at most one bounded solution. 00 ). (4. Since the finite solutions of (4..
4. We have a similar looking result for continuous time.6.¢il. which however lies a little deeper. C for all i. D Hence M ::. (4. = aP(1/J  ¢) I'l/Ji  <Pi I :::.a) so ¢ is bounded. Set (Xt)t~O is nonexplosive. suppose that 1/J is bounded and also that 1/J Set M = sUPi l1/Ji . ¢ = so c+aP¢. aM.. Assume that (Ci : i E I) is bounded. = (¢i : i E I) is the unique bounded solution to (A . But 1/J . Theorem 4. which forces M = 0 and 1/J = ¢.tc(Xt)dt.¢ so = c + aP1/J. then 00 143 I<Pi I :::. Suppose that ICil ::. then M < 00. a LPijl'l/Jj jEI <pjl :::.2.5) . aM.Q)¢ = c.2 Potential theory Proof. On the other hand. because it really corresponds to a version of the discretetime result where the discount factor may depend on the current state. CLan = n==O C/(l. Suppose that <Pi = lEi then ¢ 1 00 e>. By the Markov property Then 00 <Pi = lEi L anc(Xn ) n==O = ci + a LPijlE jEI jEI (f n==l anIc(Xn ) Xl = j) = Ci + a LPij<Pj.
C for all i. so 1jJ¢ is the unique bounded solution for the case when c = 0. then so ¢ is bounded. We have so. and then at the role of the boundary.so ¢ is bounded. T qa = O. We shall now take a brief look at some structural aspects of the set of all potentials of a given Markov chain.Q)¢ = c. by Theorem 4. What we describe is just the simplest case of a structure of great generality. which is O. Then (Xt)t~O is a Markov chain on I U {a} with modified transition rates Qi = qi + X.{Xt a for t < for t ~ T. Also T is the hitting time of a. and wished to calculate it. Finally.¢.tc±(Xt)dt = (±c) V o.144 4. ¢ is the unique bounded solution to Q¢ = c. and is finite with probability 1. = Ei l T c(Xt)dt.2. Then ¢ = ¢+ . then (XQ)(1jJ¢) = 0. Assume for now that c is nonnegative. Introduce a new state a with Ca = O. D The point of view underlying the last four theorems was that we were interested in a given potential associated to a Markov chain. First we shall look at the Green matrix. where ct = Ei 1 00 e>. Hence. if 1jJ is bounded and (XQ)1jJ = c. subtracting (X .5).4. . Let T be an independent E(X) random variable and define Xt = . Further theory Proof. which is the same as (4. When C takes negative values we can apply the preceding argument to the potentials <P. Qia = X. By Fubini's theorem <Pi Suppose Ci :::.
We have 9ij = lEi L n==O 00 lXn=j in discrete time. in continuous time ¢ = Gc. j E I) is the Green matrix 00 Similarly. in discrete time . and without boundary.1) and (4.2 Potential theory 145 Let us consider potentials with nonnegative costs c. The potential is defined by ¢i = lEi L c(X n==O 00 n) in discrete time.3).5 and 2. we have explicit expressions for all potentials of the Markov chain. and in continuous time By Fubini's theorem we have 00 00 ¢i = L n==O lEic(Xn ) = L(pnC)i = (GC)i n==O where G = (9ij : i. with G= 1 00 P(t)dt. The jth column (9ij : i E I) is itself a potential. The Green matrix is also called the fundamental solution of the linear equations (4.4. Thus.11: we know that 9ij = 00 if and only if i leads to j and j is recurrent. Indeed. These quantities have already appeared in our discussions of transience and recurrence in Sections 1. once we know the Green matrix. and in continuous time Thus 9ij is the expected total time in j starting from i.
and bility for j. .\ E (0.146 4. The formula for continuous time is hi /j is the return proba For potentials with discounted costs the situation is similar: in discrete time <Pi where = lEi L anc(Xn ) = L anlEic(Xn ) = n=O n=O 00 00 (RaC)i and in continuous time where R>.00)) the resolvent of the Markov chain. Our object now is to examine the relation between nonnegative functions.1)) and (R>. Further theory where is the probability of hitting j from i.t P(t)dt. We call (R Q: E (0. = 1 00 e>. harmonic in D. Here are two examples. and the boundary aD. with boundary aD.. for finite statespace we have Q : and We return to the general case. Unlike the Green matrix the resolvent is always finite... Any bounded function (¢i : i E I) for which ¢ = P¢ in D is called harmonic in D. : . Indeed.
3 we find 1/2 5/12) 1/2 1/3 1/2 0 where we have written the vector h a as a matrix. The linear equations for the vector h a read ha = Ph a { h~ in D = 1.~ a . . By the method of Section 1. where each allowable transition is made with equal probability. In fact. starting from i.4.2. b}. .. Let hi denote the absorption probability for a.. corresponding in an obvious way to the statespace.... States a and b are absorbing. harmonic in D.7 Consider the random walk (Xn)n~O on the above graph.. Thus we can find two nonnegative functions ha and hb .  b Example 4.2 Potential theory 147 ... hg = o.~ . We set aD = {a. the most general nonnegative harmonic function ¢ in D satisfies ¢=P¢ { ¢=f in D in aD where fa' fb 2:: 0.. and this implies Thus the boundary points a and b give us extremal generators h a and h b of the set of all nonnegative harmonic functions. but with different boundary values.
. In the preceding example the generators of C were in onetoone correspondence with the points of the boundary . 2. f+ ~ 0 and where hi = h~i and +_ { ~ hi 1 + ~ (1 .. .q. except that at o it jumps to lor 1 with probability 1/2.2. 1.the possible places for the chain to end up. . for i = 1.. We choose p > q so that the walk is transient. . .2. For we have The suggestion of this example.148 Example 4. . starting from 0. 2.(q/p)t for i = 1.8 4... 2 . which is fully developed i~ other works. but the generators of C still correspond to the two possibilities for the longtime behaviour of the chain. We must have cPi = PcPi+l + qcPil cPo = ~cPl + ~cPl. In fact.(q/p)i) 1( ") for i = 0. ~ +B O.. . at speed p . we can show that (Xn)n~O is equally likely to end up drifting to the left or to the right. Let us consider the problem of determining for (Xn)n~O the set C of all nonnegative harmonic functions cP. .. . Further theory Consider the random walk (Xn)n~O on Z which jumps towards 0 with probability q and jumps away from 0 with probability p = 1 .q.1. .2. .2 1 . This forces A = A' and B + B' = O. 1. Similarly. . 2.(q/p)i) which is nonnegative provided A has general solution for i = 0. In this example there is no boundary. cPi = qcPi+l + PcPil The first equation has general solution for i = 1. .2.. . . nonnegative provided A' + B' ~ O.. is that the set of nonnegative harmonic functions may be used to identify a . . It follows that all nonnegative harmonic functions have the form where f.. the third equation cPi = A' + B' (1 . To obtain a general harmonic function we must match the values cPo and satisfy cPo = (cPl + cPl)/2. . cPi = A + B(l .(q/p)i) for i = 0.
and since (4.3 we have h8 = Ph 8 in D (4.7.2. Indeed. 1984). Revuz (NorthHolland. See.2.7) is the minimal solution.3.2 Potential theory 149 generalized notion of boundary for Markov chains. as we showed in Theorem 4.6) has a unique bounded solution. L jE8D where . We cannot begin to give the general theory corresponding to Example 4. by monotone convergence <Pi = lEi (J(XT )) = L jE8D hJI»i(XT=j).6) has a unique bounded solution then hf = JP>i(T < 00) = 1 for all i. Markov Chains by D. indeed <P = !i hi . (4. if Pi (T < 00) = 1 for all i.2.3 when the situation is more like Example 4. for example. Set h? = JP>i(T < 00) where T is the hitting time of aD. but we can draw some general conclusions from Theorem 4. Conversely. we showed more generally that this condition implies that ¢ = P¢ + c in D { ¢= f in aD has at most one bounded solution. Note that hf = 1 for all i always gives a possible solution.4.6) { h8 = 1 in aD. then. but more generally corresponds to the varieties of possible limiting behaviour for X n as n ~ 00. Let ¢ be a bounded nonnegative function. harmonic in D. which sometimes just consists of points in the statespace. Then.7). with boundary values ¢i = Ii for i E aD.8. any bounded solution is given by (4. Hence every bounded harmonic function is determined by its boundary values and. Hence if (4. Suppose we have a Markov chain (Xn)n~O with absorbing boundary aD. Then by the methods of Section 1. Amsterdam.2. Suppose from now on that JP>i(T < 00) = 1 for all i.
8 that 4. Check that <Pi = E i L n<T c(Xn ) = Ei L c(X n=O 00 n) where T = T + 1 and where we set Ci = Ii on aD and Ca = O.h. where T is the hitting time of aD and a is a new state.2.2 Prove the fact claimed in Example 4.2. for all i E I.150 4.2. Exercises 4. Set ifn if n ~ T > T.=0 in aD also have a unique bounded solution. .3 Let (Ci : i E I) be a nonnegative function. and that 11/" 0/'1. Show that (Xn)n~o is a Markov chain and determine its transition matrix. without boundary at all.2. Can you find a similar reduction for continuoustime chains? 4. Partition I as D U aD and suppose that the linear equations ¢ = P¢ +c in D { ¢=0 in aD have a unique bounded solution. Show that the Markov chain with transition matrix P is certain to hit aD. indeed. the hitting probabilities for boundary states form a set of extremal generators for the set of all bounded nonnegative harmonic functions.1 Consider a discretetime Markov chain (Xn)n~O and the potential ¢ with costs (Ci : i E D) and boundary values (Ii: i E aD). where D~ D.  < 0/'1.7.2. This shows that a general potential may always be considered as a potential with boundary value zero or. Show that the 'l/J = P'l/J + c in 15_ { 1f. . Further theory Just as in Example 4. Consider now a new partition linear equations (Xn)n~O D u aD.
current may also enter or leave the network through aD . j E I) we shall write 'Yi T'i for the total flow from i to = 2: JEI 'Yij' In equilibrium the charge at each node is constant.4. Some nodes are joined by wires. but here it is not the current but the potential which is determined externally. subject to given external conditions. each node i having a capacity 1ri > O.¢j) = 9i. There is a simple correspondence between electrical networks and reversible Markov chains in continuoustime. which determines the conductivity as the reciprocal of its resistance.8) { ¢i = fi' for i E aD. (4.3 Electrical networks 151 An electrical network has a countable set I of nodes.3 Electrical networks 4. Therefore. The case where gi = 0 corresponds to a node with no external connection. The nodes are partitioned into two sets D and aD. given by for i =I j. the wire between i and j having conductivity aij = aji 2:: o. j E I) with T'ij = T'ji. Physically it is found that the current T'ij from i to j obeys Ohm's law: Thus charge flows from nodes of high potential to nodes of low potential. which determines its potential ¢i by A current or flow of charge is any matrix (T'ij : i. by Ohm's law. In equilibrium. The first problem in electrical networks is to determine equilibrium flows and potentials. . each 'wire' contains a resistor. These have the effect that each node i E aD is held at a given potential Ii. and that a given current 9i enters the network at each node i ED. External connections are made at the nodes in aD and possibly at some of the nodes in D. Where no wire joins i to j we take aij = O. Given a flow the network: (T'ij : i. Each node i holds a certain charge Xi. so T'i = 9i for i E D. In practice. any equilibrium potential ¢ = (¢i : i E I) must satisfy for i E D LjEIaij(¢i .
The new problem is equivalent to the old.9) where Ci = 9i/7ri. the case where aD is empty may be dealt with as follows: we must have or there is no possibility of equilibrium. that the network is connected. (4. Then. because ct and f have the same physical dimensions. In fact. However.152 4. and the symmetry of aij corresponds to the detailed balance equations. pick one node k. This is enough to ensure uniqueness of potentials. The equations for an equilibrium potential may now be written in a form familiar from the preceding section: Q¢ { =c ¢= f in D in aD. set aD = {k}. and replace the condition ~k = 9k by ¢i = O. the equilibrium potential is given by ¢i = Ei (iT c(Xt)dt + f(XT )) L9i=O iEI (4. Further theory We shall assume that the total conductivity at each node is finite: ai = Laij < 00.2. We know that these equations may fail to have a unique solution. we shall assume that I is finite. It is natural that c appears here and not 9. by Theorem 4.10) where T is the hitting time of aD. j#i Then ai = 7riqi = 7riqii. indicating the interesting possibility that there may be more than one equilibrium potential. and that aD is nonempty. The capacities 7ri are the components of an invariant measure. A 1 2 2 1 B 2 2 1 c D E F .4. but now aD is nonempty. to keep matters simple here.
Different node capacities result in different Markov chains. but the same jump chain and hence the same hitting probabilities. By symmetry.cPe.3 Electrical networks Example 4. ~Be = 1/2. (b) The unique equilibrium flow ~ with ~A = 1 and ~B = 1 is given by where r ij is the number of times that (Xt)t~O jumps from i to j before hitting B.cPF) + 2(cPe . by Ohm's law.4. Note that the node capacities do not affect the problem we considered.no scaling was necessary.3. subject to XB = 0.10).2.cPB and cPD = 1. cPE = 1. = 1/2 and cPe = 1/4. (a) The unique equilibrium potential cP with cPA = 1 and cPB = 0 is given by where T A and TB are the hitting times of A and B.3. F}.cPE) + 2(cPB . Theorem 4. Let us set cPA = 1 and cPF = O. Let us arbitrarily assign to each node a capacity 1. Here is a general result expressing equilibrium potentials.cPe) 2(cPe . and the associated flow is given by ~AB = 1/2. Then. The conductivities are shown on the diagram. is given by .1 153 Determine the equilibrium current in the network shown on the preceding page when unit current enters at A and leaves at F. This will result in some flow from A to F. since the total current leaving Band C must vanish (cPB . which we can scale to get a unit flow. cPB = 0. flows and charges in terms of the associated Markov chain. the equilibrium potential is given by cPi = Ei (lxT=A) = JP>i(XT = A) where T is the hitting time of {A. In fact. ~eF = 1/2. Then there is an associated Markov chain and. (c) The charge X associated with ~. according to (4.cPB) Hence.cPA) + (cPB . = O. and the associated Markov chain (Xt)t~o. Consider a finite network with external connections at two nodes A and B. we were lucky . ~BE = O.
154 4.. Observe that if X o = A then if i = A if i ~ {A. D ~ is the equilibrium unit flow and X the associated charge. j I(r) = l L .n< N} B '1. We shall prove (b) and (c) together. Yn +! = j. and consider the associated potential'l/Ji = 00 Then Xiqij so = XiQi 1rij = LIPA(Yn = i..jEI i.) _0. by the lEA(rij) = LIPA(Y n=O 00 00 n = i.jEI (<Pi .n < NB) 1rij = lEA(rij) n=O ('l/Ji 'l/Jj)aij = Xiqij  Xjqij = ~ij· Hence'l/J = ¢.f aijl.~n+l=). Set Xi = lEA 10 fTB l{Xt=i}dt Xi/1ri." LJ ~n'I. We define for a potential ¢ = (¢i : i E I) and a flow ~ = (~ij : i. The formula for ¢ is a special case of (4. Further theory Proof.10).B} if i = B so if ~ij = EA(r ij  = r ji ) then ~ is a unit flow from A to B. . j E I) E(<p) = lL i. as The interpretation of potential theory in terms of electrical networks makes it natural to consider notions of energy. 00 We have roo = '""'1{." 0 n=O where N B is the hitting time of B for the jump chain Markov property of the jump chain (Yn)n~O. n < NB) = i. where c = 0 and f l{A}. required..<pj)2aij . So.n = LIPA(Y n n=O < NB)1rij.
where c = (ci : i E I) with Ci = 0 for i E aD. as heat.j E I) L (¢i . The equilibrium potential and flow may be determined as follows.4.jEI (a) Denote by ¢ = (¢i : i E I) and by ~ = (~ij : i.j E I) the equilibrium potential and flow.3. i. we shall see that certain equilibrium potentials and flows determined by Ohm's law minimize these energy functions. . Theorem 4. Then L i. = 9i (b) The equilibrium flow r = (rij : i. We can write any potential in the minimization problem in the form ¢ + c.j EI (Ei .Ej)rij = 2 LEi'Yi = 0 i.¢j)aij = L (Ei .3 Electrical networks The 1/2 means that each wire is counted once.¢j)rij = 2 L iEI ¢i'Yi.3.j EI iEI so E(¢ + c) = E(¢) + E(c) ~ E(¢) with equality only if c = o. j E I) with current sources ri for i E D and boundary potential zero is the unique solution to minimize subject to Proof. For any potential ¢ we have I (~ ) ~i = 9i for i E D. (a) The equilibrium potential ¢ = (¢i : i E I) with boundary values ¢i = Ii for i E 8D and no current sources in D is the unique solution to minimize E (¢ ) subject to ¢i = fi for i E aD. Moreover. We have ~i = 0 for i E D. by the network.Ej)(¢i . = (¢i : i E I) and any flow ~ = (~ij : i. When ¢ and by Ohm's law we have E(¢) = ~ 155 are related l L (¢i i.jEI ¢j)rij = I(r) and E( ¢) is found physically to give the rate of dissipation of energy. This characteristic of energy minimization can indeed replace Ohm's law as the fundamental physical principle.
4. 2 1(8) = ~ L i. for a potential I 8 = (8ij : i.J i.j E I) is a flow with 8i = 0 for i E D. a) as for (I. Corollary 4.j E J) is an effective conductivity on J if. The following reformulation of part (a) of the preceding result states that harmonic functions minimize energy.jEI(¢i .¢j )8ij = 2 L ¢i 8i = 0. E J. Suppose that ¢ = (¢i : i E I) satisfies Q¢ = 0 in D { cP = I in aD. We know that I determines an equilibrium potential ¢ = (cPi : i E I) by for ~ ~ J cPi . where 8 = (8 ij : i. for i E J we have L. Then a is an effective conductivity if.jEJ . We have cPi = 0 for i E aD. In fact. for all potentials I = (Ii : i E J). Then cP is the unique solution to minimize subject to E (¢ ) ¢ = f in aD.jEI iEI I(~) so I(~ + 8) = D + 1(8) ~ 1(8) with equality only if 8 = o.3. the external currents into J when J is held at potential I are the same for (J.Ij) aij.¢j)aij : 0 { L(¢i . j E I) the equilibrium potential and flow. given any network.jEJ 8fj u i?· . j E J) we set  = (Ii : i E J) and a flow E(/) = 1 " (Ii 2 L. Then L i. a). We can write any flow in the minimization problem in the form ~ + 8. An important feature of electrical networks is that networks with a small number of external connections look like networks with a small number of nodes altogether. Let J ~ I.Ji) aij' jEJ For a conductivity matrix a on J. for all I. i. equivalent in its response to external flows and potentials. there is always another network of wires joining the externally connected nodes alone. We say that a = (aij : i.jEI 'Yij8ijai/ = L (¢i . Further theory (b) Denote by cP = (cPi : i E I) and by ~ = (~ij : i.Ii for 'I.156 4.¢j)aij JEI = :E(Ji .
q) = aij + L aik<P{ k~J = (¢~ : i E I) is the potential defined by = 0 ~kEI(<p1 . There is a unique effective conductivity a given by 157 aij where for each j E J. and. by Corollary 4. and also by the Thompson variational principle inf Oi==gi on J1(8) = "Yi== inf {J gi on I(~).J { Moreover.3 Electrical networks Theorem 4.3. o off J Proof.3.4. J for i E J. (4.11) a is characterized by the Dirichlet variational principle <Pi==fi on J E(/) = inf E(¢).<p{)ai~ for i tf. Z:=aij<Pj = z:=aijfj + z:=z:=aik<p{fj = z:=aijfjo JEI jEJ k~JjEJ jEJ In particular.4. ¢~ = 8ij for i E J. ¢ solves minimize subject to We have. Given I = (Ii: i E J). define ¢ = (¢i : i E I) by <Pi = L 1i<p1 jEJ then ¢ is the equilibrium potential given by ~jEI aij(¢i { ¢j) =0 ¢i = Ii for i fj. for i E J E (¢ ) ¢i = Ii for i E J.5. taking I == 1 we obtain Laij JEI = Laijo jEJ .
jEJ 2 1 g··a·· • ~J ~J o Effective conductivity is also related to the associated Markov chain (Xt)t~O in an interesting way. if gij = (Ii  Ij )aij and ~ij = (¢i .jEJ i. J. with jump matrix IT given by for i.158 4.jEI = = L (cPi .Ii )Uij · jEJ Moreover. we have = (8ij : i.cPj)2aij i.j E J.cPj)aij JEI = 2 L Ii LUi !i)Uij = L Ui _!i)2Uij . i.jEI L (Ii .3.jEI L ~2 1 v··a·· ~J ~J > L i.¢j )aij. we also have equality of energies: L (cPi .3.j E I) with 8i = gi for i E J i.jEI = 2L iEI cPi L(cPi . iEJ jEJ i. This is really a transformation of the jump chain. Define the time spent in J At = it l{X s EJ}ds and a timechanged process (X t)t~O by X t = Xr(t) where r(t) = inf{s ~ 0 : As > t}. By applying the strong Markov property to the jump chain we find that (X t)t~O is itself a Markov chain. then L "Yljai/ i. 1'fij = 'lrij + L'lrik¢k k~J . by Theorem 4. Further theory Hence we have equality of external currents: 2)cPi JEI cPj )aij = 'r)1i .cPj)2 aij i.jEJ so. We obtain (Xt)t~O by observing (Xt)t~O whilst in J.Ii )2 Uij = L gfjUi/. and stopping the clock whilst (Xt)t~O makes excursions outside J. for any flow 8 and bi = 0 for i ¢.jEJ Finally.
Hence has Qmatrix given by Qij = qij (Xt)t~O + L qik¢{ k~J Since ¢i = (¢{ : k E I) is the unique solution to (4.4. Doyle and J.4 Brownian motion Imagine a symmetric ran'dom walk in Euclidean space which takes infinitesimal jumps with infinite frequency and you will have some idea of Brownian motion. G. you should see Random Walks and Electrical Networks by P. We shall show that the scaleddown and speededup process X t(c)  c 1/2Xct is a good approximation to Brownian motion.4. The mathematical object now called Brownian motion was actually discovered by Wiener. Snell (earus Mathematical Monographs 22. The simple symmetric random walk (Xn)n~O on Zd is a Markov chain which is by now quite familiar.4 Brownian motion where 159 ¢{ = JP>k(XT = j) and T denotes the hitting time of J. A discrete approximation to Euclidean space ]Rd is provided by where c is a large positive number. Also.4. for example in tying up the nonequilibrium behaviour of Markov chains and electrical networks. This provides an elementary way of thinking about Brownian motion. 4. Mathematical Association of America. methods coming from one theory one provide insights into the other. For an entertaining and illuminating account of the subject. this shows that so (Xt)t~O is the Markov chain on J associated with the effective conductivitya. Moreover. 1984). There is much more that one can say. It is named after a botanist who observed such a motion when looking at pollen grains under a microscope. See Example 1. and is also called the Wiener process.11). L. it makes it reasonable to .
For ct E Z+ we have xi so the squareroot scaling gives which is independent of c.. for all bounded continuous functions f. Let Xl. Further theory suppose that some properties of the random walk carryover to Brownian motion. that is to say. A continuous realvalued process (Bt)t.00). Why is space rescaled by the squareroot of the timescaling? Well. The fundamental role of Gaussian distributions in probability derives from the following result. and then show that this is not an empty definition.160 4. ••• be a sequence of independent and identically distributed realvalued random variables with mean 0 and variance t E (0.1 (Central limit theorem).+ Xt(w) is continuous}) = 1. Brownian motions exist.4. We begin by defining Brownian motion.:::o is called a Brownian motion if B o = 0 and for all 0 = to < tl < ..:::o is said to be continuous if lP({w : t. Then. as n ~ 00 we have We shall take this result and a few other standard properties of the Gaussian distribution for granted in this section. A realvalued random variable is said to have Gaussian distribution with mean 0 and variance t if it has density function cPt(X) = (27T"t)1/2 exp{ _x 2/2t}. Theorem 4. X 2 . we will at least want IE[lxi ) 12 ] to converge to a nondegenerate limit. There are many introductory texts on probability which give the full details. if c ) converges in some sense as c ~ 00 to a nondegenerate we hope that c limit. < t n the increments . At the end of this section we state some results which confirm that this is true to a remarkable extent.. A realvalued process (Xt)t.
For N = 0. . However.1 and define Zt = 2(N+l)/2yt.00).2.00). Let us say that (B t : t E D N ) is a Brownian motion indexed by D N if B o = 0 and for all o = to < tl < . B s . Brownian motion exists.N S = t + 2.N in [0.1 • For t E DN\D N . t E D N . Suppose we have constructed (Bt : t E D N .4. Theorem 4.1 set r = t . . For t E Do = Z+ set then (B t : t E Do) is a Brownian motion indexed by Do. B t = ~(Br We obtain two new increments: + B s) + Zt· Br ) Bt . .Zt. We shall show how to extend this process successively to Brownian motions (Bt : tEDN ) indexed by D N . an independent Gaussian random variable yt of mean 0 and variance 1. it is not obvious that there is any such process. is uniquely determined.2.4.B r ) . a Brownian motion indexed by D N .. denote by D N the set of integer multiples of 2.. for each tED.Br = ~ (B s = ~(Bs  + Zt.Bt . To put it properly. which is a measure on the set of continuous paths.1 ). Then (B t : tED) is a Brownian motion indexed by D.1. Next we shall show that (Bt : tED) extends continuously to t E [0. Proof. and denote by D the union of these sets. < t n in D N the increments are independent Gaussian random variables of mean 0 and variance We suppose given..4 Brownian motion are independent Gaussian random variables of mean 0 and variance 161 The conditions made on (Bt)t~o are enough to determine all the probabilities associated with the process.. and finally check that the extension is a Brownian motion. the law of Brownian motion.N so that s.2 (Wiener's theorem). We need the following result.
162 We compute E[(Bt . Set MN = sup IziN)I.l] There are 2 N  I points in (DN\D N .Xp1IP(IY11 > 'x)d'x = 2N . N For t E D N . ~2(NI) .I .2(N+I) = o. Hence (Bt : tEDN) is a Brownian motion indexed by D N. being constructed from B s . Further theory = E[(B s   B t )2] E[(Bt Br)(B s . we obtain a Brownian motion (B t : tED). by induction.I] Then.Xp.1E(IY1IP) . tE[O.B r )2] 4.BiNI). are therefore independent and of the required variance.1].1 1 00 p. Hence. as required. For t E DN\DN . Moreover.1 IP(2(N+l)/2 M N > 'x)d'x ~ 2N .I we have zi ) = O. For each N denote by (BiN))t~O the continuous process obtained by linear interpolation from (Bt : t E D N ). ~2(NI) The two new increments. by our construction we have with yt Gaussian of mean 0 and variance 1. tE(DN \DN l)n[o. s).N . So for A > 0 we have For a random variable X ~ 0 and p > 0 we have the formula Hence 2P(N+l)/2E(M~) = 1 00 p. since (ZiN))t~O interpolates linearly between its values on D N . they are certainly independent of increments over intervals disjoint from (r. Also. being Gaussian. we obtain MN = sup 2(N+I)/2Iytl.I ) n [0.B r and yt.B t )] = = + 2(N+I) = 2. set zi N ) = BiN) .
as claimed. Let (Xn)n. as N N=O BiN) = Bi O + zi 1 ) ) + .c 1/2Xct where the value of X ct when ct is not an integer is found by linear interpolation.. for all bounded continuous functions f : jRm ~ jR and all 0 ~ tl < . Now BiN) eventually equals B t for any tED and the uniform limit of continuous functions is continuous.?:o have the required joint distribution. < t m .· .. and by a similar argument uniformly for t in any bounded interval. very/much as the Gaussian distribution does for sums of independent random variables.4 Brownian motion and hence. It remains to show that the increments of (Bt)t.. with probability 1.?:O be a discretetime realvalued random walk with steps of mean 0 and variance 0'2 E (0. (B t : tED) has a continuous extension (Bt)t. for any p > 2 00 00 163 ELMn= LE(MN ) N=O N=O : :. Set to = to = o.. using continuity of (Bt)t. .. Then. But given 0 < t 1 < .?:o.aBtm )] where (Bt). . we now want to show how it appears as a universal scaling limit of random walks.Xi~)] tE[f(aBh as c ~ 00..?:o we can let m distribution for the increments to obtain the desired D Having shown that Brownian motion exists. .3.. Therefore. for all m.?:o is a Brownian motion.1]. + zi N) converges uniformly in t E [0. We know that the increments are Gaussian of mean 0 and variance tr .t.4. E(IY1IP)1/p ~ 00 L 00 (2(p2)/2 p )N < 00. · · · . < t n we can find sequences (t''k)mEN in D such that 0 < t 1 < . Theorem 4.t(f. < t~ for all m and tf: ~ tk for all k. we have  E[f(Xi~). L 00 E(MKr)l/P :::. N=O It follows that.4..t~  t~l· ~ 00 Hence. For c > 0 consider the rescaled process X t(c) ..··· .00).
..X(e) .tk_l)1/2. .. and (C. . Further theory Proof The claim is that (xi~). Hence U~c) converges weakly to Zk. · . .. . .tk_l)1/2 Zk.. so it suffices to prove the claim with .. There now follows a series of related remarks..L [et n] +1 'V:)\ t rn t1 t rn .. X(e)) . But e (Ui ). .. . it suffices to show that U~e) converges weakly to Zk for each k.. Then. We call (Bt)t~o a Brownian . The right side converges weakly e to 0. X(e))\ :::.. Consider now the increments (e) Uk .U~)) converges weakly to (ZI.X(e) tk tkl' for k Z k = a (B tk .164 4.1/ 2Nk(C)1/2)Nk(C)1/2(Yi + . .t (tk .. D To summarize the last two results. Given d independent Brownian motions (Bi )t~o.B tkl ) increments are independent. that there is a continuous process (Bt)t~o with stationary independent increments and such that B t is Gaussian of mean 0 and variance t.(Bt)t~o. . we established a sort of convergence to Brownian motion.. . .Zm). .i~e) replacing xi ).. . Then \(xt(e) 1 . . First define (e) Xt  e I/2X[et] where (et] denotes the integer part of ct. we have shown. using special properties of the Gaussian distribution... By the central limit theorem Nk(e)I/2(YI + . That was Wiener's theorem. e I/2\('V:[et 1 ] +1. having stationary independent increments and such that X t is Poisson of parameter At for each t ~ O. . for each t ~ o. m.[etkI]./ [etk] L Yn rv (C. Then since both sets of = 1. as required. In the proof we shall take for granted some basic properties of weak convergence. let us consider the JRdvalued process B t = (Bi.aB trn ) as c ~ 00..1 / 2 Nk(C)1/2) .Bt). ... .. using the central limit theorem applied to the increments of a rescaled random walk. Since .(X(e) . + YN(c») n=[etk_l]+1 where rv denotes identity of distribution and Nk(e) = [etk] .ide) = Bo = 0 it suffices to show that U~c) = 12 C.L where Y n denotes the nth step of (Xn)n~O. .. . + YN(e)) converges weakly to (tk . . Note the similarity to the definition of a Poisson process as a rightcontinuous integervalued process (Xt)t~O starting from 0... xi~) converges weakly to (a Btl.
We might take (Xn)n~O to be the simple symmetric random walk in Z3.3 suggests the following scaling invariance property of Brownian motion (Bt)t>o. then V = ~I. For any c > 0 the proces.4. to anything else. By analogy with continuoustime Markov chains we look for a transition semigroup (Pt)t~O . what we have proved is strong enough to ensure that (Xie))t~o does not converge. The sense in which we have shown that (Xie))t~o converges to Brownian motion is very weak. given a random walk with a complicated step distribution. Thus Brownian motion appears as a fixed point of the scaling transformation. which attracts all other finite variance symmetric random walks as c ~ 00. (Bic)k.4. but rather about processes with independent increments.4. the result shows that in the scaling limit they behave asymptotically the same. Here are two examples. with no essential change in the proof. in the same sense. once the difference in variance is taken out.4 Brownian motion motion in ]Rd. 165 There is a multidimensional version of the central limit theorem which leads to a multidimensional version of Theorem 4. The discussion to this point has not really been about the Markov property. in which case V = I.BO)t~o is a Brownian motion (starting from 0). Alternatively. To remedy this we must first define Brownian motion starting from x: this is simply any process (Bt)t~o such that B o = x and (B t . As a limit of Markov chains it is natural to look in Brownian motion for the structure of a Markov process.3. we might take the components of (Xn)n~O to be three independent simple symmetric random walks in Z. All other aspects of the step distribution become irrelevant as c ~ 00. However. then for all bounded continuous functions as c ~ 00 we have f : (]Rd)m ~ ]R. Thus if (Xn)n~O is a random walk in ]Rd with steps of mean 0 and covariance matrix and if V is finite. Although these are different random walks.~o defined by B t(e)  C 1/2Bet is a Brownian motion. The scaling used in Theorem 4. which is also easy to check from the definition. More generally. it is useful to know that on large scales all one needs to calculate is the variance (or covariance matrix). and one can with effort prove stronger forms of convergence.
Further theory and a generator G.B s . x. This is the transition density for Brownian motion and the transition semigroup is given by (Ptf)(x) = ( p(t. For any bounded measurable function have f : ]Rd ~ ]R we lEx[J(Bt )] = lEo[J(x + B t )] = { f(x + Y)¢t(Yd··· ¢t(Yd)dYl . if aXl2 + · · · + ax d 2· JRd a2 f has two bounded derivatives. by analogy with continuoustime chains. x. x. y) where = !~xp(t. JRd To check the semigroup property PsPt = Ps+t we note that Ex[f(B s+t )] = Ex [f(B s + (B s+t . y)f(y)dy JRd = ( JRd p(t. . x. a term we have not defined precisely. should be given by G=!~. . y) = (27rt)d/2 exp{ Iy . x. y)( !~J)(y)dy = Ex[(~~f)(Bt)] ~ ~~f(x) as t ! O. x. y)f(y)dy = { !~yp(t. y)f(y)dy JRd where p(t. that the generator. y) a2 ~x = Hence..166 4. x. x. dYd JRd = ( p(t. we have 8 at (Ptf)(x) = { !~xp(t. y)f(y)dy = lEx [f(Bt )]. This suggests.xI 2/2t}. For t > 0 it is easy to check that :t p(t.B s ))] = Ex [Ptf(B s )] = (PsPtf)(x) where we first took the expectation over the independent increment B s +t .
N.4.I / 2Xl) . You will notice some weasel words creeping in. relative to the corresponding results for chains.I / 2 X Nt . For further details see. But. 1993). .2. f(x . Roger. t N(pi. f has two bounded derivatives then.N.f)(x) ~~f(x). liN. G. this aside.2/N. required to have various degrees of local regularity. continuous and differentiable. the main ideas are very similar..I / 2 ) so  2f(x) + f(x + N. t == 0.I / 2 )}. now we have linear operators on functions: Pt is an integral operator. is (N) /N N ( PI/Nf . W. now there are functions f : jRd ~ jR.. . such as measurability and differentiability.. . For a bounded continuous function f : jRd ~ jR. 2nd edition 1994). We would like to explain the appearance of the Laplacian ~ by reference to the random walk approximation. . C. Volume 1: Foundations by L. here is a result on recurrence and transience. These are various sorts of local regularity for functions defined on the statespace jRd. for example.I / 2 ) . They did not appear for Markov chains because a discrete statespace has no local structure.f ) ) . G is a differential operator.s and David Williams (Wiley. Chichester.. and a proper measuretheoretic basis for the probability.I / 2 ) == NI(~f(x) + o(N)).2f(x) + f(x + N. You might correctly guess that the proofs would require additional real analysis. f ((N))] Xo NIE N l/2 x [f(N. Stroock (Cambridge University Press.4 Brownian motion 167 Where formerly we considered vectors (fi : i E I). First. 2/N. Markov Processes and Martingales.I / 2X O)]  = (N/2){f(x .. by Taylor's theorem. We finish by stating some results about Brownian motion which emphasise how much of the structure of Markov chains carries over.Jf .NIE x [f (X I(N) ) (x _ == If we assume that as N ~ 00. Probability Theory . set X (pt(N) f)(x) == IEx[f(Xi N ))]. Where formerly we considered matrices Pt and Q.I/N. .I / 271 d • The closest thing we have to a derivative in t at 0 for (pt(N))t=O.f(N. such as measurable. Denote by (Xn)n~O the simple symmetric random walk in tl d and consider for N == 1.. or Diffusions.an analytic view by D. E N. the rescaled process N xi ) == N.
168 4.8. then for any N < 00 00 00) JP>(I B t I~ = 1. + z) = 00. The generator ~ of Brownian motion in jRd reappears as it should in the following martingale characterization of Brownian motion. that in Z and Z2 the simple symmetric random walk is recurrent. Theorem 4.4. as t ~ we have t lEx[f(Bt )] and. The invariant measure for Brownian motion is Lebesgue measure dx. So the projected process is positive recurrent and we can expect convergence to equilibrium and ergodic results corresponding to Theorems 1. If we accept this latter notion of recurrence the correspondence extends to dimension two. for any € > 0) = 0 >0 JP>( {t ~ 0 : IBt I < €} is unbounded) = 1. Let (Bt)t~o be a Brownian motion in (i) If d = 1. then JP>(Bt = 0 for some t but.4.l]d f(z)dz JI»x (~ it f(Bs)ds t 1 ast t 00) = 1. In dimension one this just means wrapping the line round a circle of circumference 1. Theorem 4.5. whereas in Z3 it is transient.2. This has infinite total mass so in dimensions one and two Brownian motion is only null recurrent. we shall consider Brownian motion in jRd projected onto the torus T d = jRd /Zd. Let (Bt)t>o be a Brownian motion in f : jRd ~ jR be a continuous periodic function. = 2. then JP>({t ~ 0: B t (ii) If d = O} is unbounded) = 1. so that jRd and let f(x Then for all x E jRd. The invariant measure remains Lebesgue measure but this now has total mass 1. In dimension two we see the fact that for continuous statespace it makes a difference to demand returns to a point or to arbitrarily small neighbourhoods of a point.3 and 1.6. So that we can state some results for the positive recurrent case. f(x) for all z E Zd. ! .10. as t ~ (iii) If d = 3. It is natural to compare this result with the facts proved in Section 1.4. Further theory jRd. The results correspond exactly in dimensions one and three. moreover 1= [ J[O.
then 'l/J ~ ¢. Theorem 4.2. Write (Ft )t2::o for the filtration of (Xt)t~o. as we remarked in Section 4.4 Brownian motion 169 Theorem 4.12) has no finite solution. corresponding very closely to Theorem 4.4. . These potentials are identical to those appearing in Newton's theory of gravity.1. then (4.7.00) be continuous. if JEIMtl < 00 for all t. (iii) if ¢(x) = 00 for some x. and whenever s :::. a continuous time process (Mt )t2::o is a martingale if it is adapted to the given filtration (Ft )t2::o. (ii) for all bounded functions f which are twice differentiable with bounded second derivative. Let c : D ~ [0. In case you are unsure.4. (iv) if JPx(T < 00) = 1 for all x. Let D be an open set in ]Rd with smooth boundary aD.12) (ii) if 1/J E C 2 (D) n C(D) and satisfies { ~~1/J ~ c in D in aD 1/J ~ f and 'l/J ~ 0.4. Let (Xt)t~O be a continuous ]Rdvalued random process. 00) be measurable and let f : aD ~ [0. (4.3 for Markov chains.6.2. Set <jJ(x) = JEx [I T c(Bt)dt + f(XT )IT<oo] where T is the hitting time of aD. t and A E F 8 • We end with a result on the potentials associated with Brownian motion. Then the following are equivalent: (i) (Xt )t2::o is a Brownian motion.2.12) has at most one bounded solution in C 2 (D) n C(D). Then (i) ¢ if finite belongs to C 2 (D) n C(D) and satisfies in D in aD. the following process is a martingale: This result obviously corresponds to Theorem 4. then (4.
In each case our aim is to provide an introduction rather than a systematic account or survey of the field.for hitting probabilities and expected rewards. Exercise 1. Markov decision processes and Markov chain Monte Carlo. Example 1. Some have already appeared to illustrate the theory.1. They are often easy to spot. We are now going to give . See Example 1.5. queueing models.1.we know. from calculating the fair price for a random reward to calculating the probability that an absentminded professor is caught without an umbrella. such as are often found in biology. that every state is either recurrent or transient. and for longrun behaviour via invariant distributions. We have already illustrated some aspects of the theory by simple models with a biological interpretation.4 (birthanddeath chain) and Exercise 2. Once a Markov chain is identified. In this chapter we shall look at five areas of application in detail: biological models. 5. References to books for further reading are given in each section.1 (bacteria).3.5 (virus).5 Applications Applications of Markov chains arise in many different areas.1 Markov chains in biology Randomness is often an appropriate model for systems of high complexity. resource management models. for example. from games of chance to the evolution of populations. There are also good computational methods .6 (octopus). In a realworld problem involving random processes you should always look for Markov chains. there is a qualitative theory which limits the sorts of behaviour that can occur .
Suppose. they wished to determine the probability that after n generations an individual had no male descendents. and also to the study of chain reactions in chemistry and nuclear fission. so for i 2: 1. } with absorbing state o. We deduce that with probability 1 either X n = 0 for some n or X n ~ 00 as n ~ 00. Example 5. who dies and is replaced at time n = 1 by a random number of offspring N. The process (Xn)n~O is a Markov chain on I = {O. .:)kEN. by providing quantitative understanding of various phenomena they can provide a useful contribution to science. We can construct the process by taking for each n E N a sequence of independent random variables (N. by setting X o = 1 and defining inductively.. Consider the probability generating function </J(t) = E(t N ) = :E tkJP(N = k==O 00 k). having the same distribution as N. The basic branching process model has many applications to problems of population growth. each independently. .. and every state i 2: 1 is transient. The case where P(N = 1) = 1 is trivial so we exclude it. + N Xn _ • 1 Then X n gives the size of the population in the nth generation. hence i is transient in any case. Nevertheless. even in a growing population.1 Markov chains in biology 171 some more examples where Markov chains have been used to model biological processes. and j does not lead to i.1 (Branching processes) The original branching process was considered by Galton and Watson in the 1870s while seeking a quantitative explanation for the phenomenon of the disappearance of family names. Further information on (Xn)n~O is obtained by exploiting "the branching structure. Under the assumption that each male in a given family had a probability Pk of having k sons. 1..5. and so on. next. that these offspring also die and are themselves replaced at time n = 2. i leads to j for some j > i. each with the same distribution as N. by a random number of further offspring. in the study of population growth. If P(N = 0) = 0 then P(N 2: 2) > 0. epidemics and genetic inheritance. The solution to this problem is explained below. Suppose at time n = 0 there is one individual.2..1. for n 2: 1 X n = Nf + . We have P(Xn = 0 I X n  1 = i) = P(N = O)i so if P(N = 0) > 0 then i leads to 0. It should be recognised from the start that these models are simplified and somewhat stylized in order to make them mathematically tractable.
00 = ¢( q) so also q 2:: r.1) we must have q = 1. also so.1] : ¢(t) = t}. Hence. we have q = P(Xn = 0 for some n) = lim ¢(n)(o). n+oo Now ¢(t) is a convex function with ¢(1) = 1. 0 ¢.1 ) . ¢(n)(o) ~ r for all n. so if X n = k then it takes k steps to obtain X n + l . til t til t til where Jl = E(N). by induction. and if ¢'(1) ~ 1 then since either ¢" = 0 or ¢" > 0 everywhere in [0. we find that E(t Xn ) = ¢(n)(t). this gives the entire distribution of X n . where ¢(n) is the nfold composition ¢ 0 . then ¢(r) = r by continuity.00 = lim ¢( ¢(n) (0)) n +. Suppose that in each generation we replace individuals by their offspring one at a time. . Applications t ~ 1. by induction. In principle. .172 defined for 0 ~ 5. On the other hand q = lim ¢(n+l) (0) n +. we have ¢(O) ~ r and. There is a nice connection between branching processes and random walks. Since ¢ is increasing and 0 ~ r. where Jl is the mean of the offspring distribution. Conditional on X n  l = k we have Xn so =Nf+···+Nr and so 00 lE(t Xn ) = 'LlE(t Xn I X n k==O 1 = k)JP>(Xn . hence q ~ r.1 = k) = lE(</>(t)xn . Hence q = r.n. Some quantities are easily deduced: we have E(Xn ) = lim dd lE(t Xn ) = lim dd </>(n)(t) = (lim </>'(t)r = p. though ¢(n) may be a rather complicated function. If ¢'(1) > 1 then we must have q < 1. Let us set r = inf{t E [0. We have shown that the population survives with positive probability if and only if Jl > 1. . since 0 is absorbing.
whether infected or not. E.4. These two possibilities have identical consequences for the progress of the epidemic. (Xn)n>O hits 0 if and only if (Ym)m>O hits O.5.1. In an idealized population we might suppose that all pairs of individuals make contact randomly and independently at a common rate. individuals themselves become infective and remain so for an exponential random time. for n 2: 0 Observe that X n = YTn for all n.1. these naive explanations leave unanswered many quantitative questions that are important in predicting the behaviour of epidemics.2 (Epidemics) Many infectious diseases persist at a low intensity in a population for long periods. New York. This idealized model is obviously unrealistic. The decline of an epidemic can also be explained by the eventual decline in the number of individuals susceptible to infection. Occasionally a large number of cases arise together and form an epidemic. However. For an idealized disease we might suppose th~t on contact with an infective. Harris (Dover. if qi then qi = P(Ym = 0 for some m I Yo = i) = qf for all i and so ql = P(N = 0) + L qfP(N = i) = ¢(ql)' k==l 00 Now each nonnegative solution of this equation provides a nonnegative solution of the hitting probability equations.1 Markov chains in biology 173 The population size then performs a random walk (Ym)m~O with step distribution N . The classic work in this area is The Theory of Branching Processes by T. Define stopping times To = 0 and. .3 to see that. but it is the simplest mathematical model to incorporate the basic features of an epidemic. in agreement with the generating function approach. after which they either die or recover. so we deduce that ql is the smallest nonnegative root of the equation q = ¢(q). This behaviour is to some extent explained by the observation that the presence of a large number of infected individuals increases the risk to the rest of the population. 1989). Moreover we can use the strong Markov property and a variation of the argument of Example 1. and since (Ym)m~O jumps down by at most 1 each time. Example 5. as infectives either die or recover and are then resistant to further infection.
~ ~ . N . rather. It can be shown that as N ~ 00 the process (sf'. q(s. then N . Of greater concern is the behaviour of an epidemic in a large population.x.i+l) = Asi.(St. Consider the case where So = N 1. if') converges to the solution (St. and the rate at which the rumour is passed on is qi = i(N . Since St + It does not increase. How long does it take until everyone knows the rumour? If i people know the rumour. they in turn pass the rumour on to everyone they meet. Let us consider the proportions sf = St/N and if' = It/N and suppose that .x = v/N. The expected time until everyone knows the rumour is then Nl i=l 1 Nl i=l N Nl(l i=l '" qi = '" ~(N .10 = 1. indeed all the communicating classes are singletons.174 5. X t = (St. if') .i)(si. We assume that each individual meets another randomly at the jump times of a Poisson process of rate 1.~) = '" . it) I] ~ o for all t 2:: o. of size N. In the idealized model.O) for s E Z+ are all absorbing and all the other states are transient. ~ =~ 1) = +. .i)/N. say.~ Nl i=l 2 ' " =~ ~ 1 rv 210gN as N ~ 00. we effectively have a finite statespace. an asymptotic equivalence. i o). We can calculate these probabilities explicitly when So + 10 is small. /l E (0. This is not a limit as above but.i)(s. The fact that the expected time grows with N is related to the fact . The states (s. It) performs a Markov chain on (Z+)2 with transition rates q(s. The epidemic must therefore eventually die out. it) of the differential equations (d/dt)st (d/dt)i t = vstit = vStit .i do not.il) = /li for some . Here convergence means that E[I( sf'./lit starting from (so. where v is independent of N. Applications We denote the number of susceptibles by St and the number of infectives by It. but will give an example of another easier asymptotic calculation. We will not prove this result. and the absorption probabilities give the distribution of the number of susceptibles who escape infection. 00 ). This has the following interpretation: a rumour is begun by a single individual who tells it to everyone she meets.x = l/N and /l = O. Consider now a sequence of models as N ~ 00 and choose s~ ~ So and i~ ~ i o.
and in particular make no requirement that parents be of opposite sexes. for example. (The word gene refers to a particular chromosomal locus. so the possibilities are AA. then (Xn)n~O is a Markov chain with the above transition probabilities. Let us take m to be even with m = 2k. We emphasise that we present only the very simplest examples in a rich theory. the proportion of 'infectives' changes very slowly. . Ewens (Springer. for which we refer the interested reader to Mathematical Population Genetics by W. the varieties of genetic material that can be present at such a locus are known as alleles. all independent. many more questions of a statistical nature have arisen. for example. 1979).m} with transition probabilities In each generation there are m alleles. some of type A and some of type a.) This sort of study was motivated in the first place by a desire to find mathematical models of natural selection. Berlin. They represent an attempt to understand quantitatively the consequences of randomness in genetic inheritance. in sexual reproduction the choice of partner. and thereby to discriminate between various competing accounts of the process of evolution. get aa aA Aa AA AA. The final two examples come from population genetics. If X n denotes the number of alleles of type A in generation n. More recently.5. We have to allow that both parents may be the same. 1. We might. Example 5. or the choice of parents' alleles retained by their offspring. as scientists have gained access to the genetic material itself. The types of alleles in generation n+ 1 are found by choosing randomly (with replacement) from the types in generation n. Suppose that individuals in the next generation are obtained by mating randomly chosen individuals from the current generation and that offspring inherit one allele from each parent.1 Markov chains in biology 175 that we do not scale /0 with N: when the rumour is known by very few or by almost all.. AA aA AA AA aa. We suppose that each individual has two genes. Then if the generation n is. This can be viewed as a model of inheritance for a particular gene with two alleles A and a. J. .1. Aa and aa.. then each gene in generation n + 1 is A with probability 7/10 and a with probability 3/10. The randomness here might derive from the choice of reproducing individual.3 (WrightFisher model) This is the discretetime Markov chain on {O. .
The structure of pairs of genes is irrelevant to the Markov chain which simply counts the number of alleles of type A. j==o According to this model. . Suppose A mutates to a with probability u and a mutates to A with probability v. The communicating classes of (Xn)n~O are {O}.p) log(l . Firstly.i)jm)2 a(ijm)2 and the transition probabilities are Secondly.176 5. that. Then the probability of choosing A when X n = i is given by ¢i = {i(l . for p E (0. known. as m ~ 00 lEpm(T) ~ 2m{(1 . Applications (Xn)n~O. ~ + (lj2)(3i(m . {I..m .1} is transient..1).i)jm 2 + .((m . so in a large population diversity does not disappear quickly. This means that the probability of choosing allele A when X n = i is given by 'l/J. Some modifications are possible which model other aspects of genetic theory. (3. States 0 and m are absorbing and {I.. {m}. ~ > 0 respectively.. . however.p) It is + plogp} where T is the hitting time of {O. The hitting probabilities for state m (pure AA) are given by This is obvious when one notes that one can check that (Xn)n~O m is a martingale. alternatively hi = LPijh j .m . m}. we may allow genes to mutate. aa have a relative selective advantage given by lX. .u) and + (m  i)v}/m . genetic diversity eventually disappears. .i)jm 2 a(ijm)2 + (3i(m .I}. it may be that the three genetic types AA. Aa.
. the Moran model cannot be interpreted in terms of a species where genes come in pairs. Now. Pi. not the whole population. The hitting probabilities are given by IPi(Xn = m for some n) = i/m. so attention shifts from hitting probabilities to the invariant distribution 7r.. absorbing states 0 and m and transient class {I.m .5. in which case there is no change in the makeup of the population. a and A. {I.1 Markov chains in biology 177 With u. . the basic Markov chain structure is the same..1. or where individuals have more than one parent. . There is an exact calculation for the mean of 7r: we have p.m .I}. like the WrightFisher model. . v > 0. if X n denotes the number of type A individuals in the population at time n. Example 5. Pii = (i 2 + (m  i)2)/m 2. We can also calculate explicitly the mean time to absorption . and add a new individual of the same type. then we choose. then (Xn)n~O is a Markov chain with transition matrix P.i+l = i(m  i)/m2. = L i7ri = E 7r (XI) = L 7riEi(X1) i=O m m m m i=O = 2: m7ri<Pi = 2:{i(l i=O i=O u) + (m . {m }.il = i(m  i)/m 2. is a martingale.4 (Moran model) The Moran model is the birthanddeath chain on {O.i)V} 7ri = (1  u)p. The Moran model is reversible. one individual from the population at time n and remove it. There are some obvious differences from the WrightFisher model: firstly. 1. again randomly. both to give birth and to die. . secondly in the Moran model we only change one individual at a time. and. with communicating classes {O}..I}. the states 0 and m are no longer absorbing. in fact the chain is irreducible. The same individual may be chosen each time. so we obtain the population at time n + 1. + mv  vp. . .. we choose randomly one individual from the population at time n. so that JL = mv/(u + v). . Here is the genetic interpretation: a population consists of individuals of two types.m} with transition probabilities Pi. However.
m2 . = m.p) . .k pm = (1.2 + .1 kg = kin = 0.p) + plogp} which has the same functional form in p and differs by a factor of m/2. + L m J . starting from i.i1 k t1 + Piikt + Pi.( mj j j j(mj) which gives k.. kf where = for i ~ j ((mi)/(mj))k. Then.. before absorption: kf kf = bij + (Pi.) m1.m . This factor is partially explained by the fact that the Moran model deals . Hence ki=Lkf=m L j=1 m1 { i ( m=z. for i = 1.m .k~ = ..p) + plogp}... The simplest method is first to fix j and write down equations for the mean time spent in j.p) 10g(1 .j ..p) j==1 L m_ ~ 00 mp 1 m1 .. For the WrightFisher model we claimed that Epm(T) ~ 2m{(1 . . is determined by m.1 so that . J +P j==mp+1 J L ~ 1 t (1.. as m Epm(T) ~ m 2 {(1 . } j=1 j=i+1 J ~ · As in the WrightFisher model. {(i/j)k. So. m}. and i = pm for some P E (0. for i ~j k.. . Applications where T is the hitting time of {O. Then 2 m. one is really interested in the case where m is large...l jl).1).i+1 k t+1) for i = 1.178 5. .p) 10g(1 .p) log(1.plogp as m ~ 00.
in the notation of Example 5. Aa or aa with probabilities 1/4. How long on average does it take to achieve a pure strain. it is assumed that the times between arrivals are independent random variables of the same distribution. Show that. n + Nx nl + In where (In)n'?:o is a sequence of independent Z+valued random variables with common generating function 'ljJ(t) = E(t 1n ).2 Queues and queueing networks Queues form in many circumstances and it is important to be able to predict their behaviour.1. 1/2 and 1/4. of some other distribution. that is. as compared to a type A individual.queues runs as follows: there is a succession of customers wanting service. A single plant of genotype Aa is crossed with itself.i)). The basic mathematical model for.1. so that the offspring has genotype AA. on arrival each customer must wait until a server is free. if X o = 1.5.3 In the Moran model we may introduce a selective bias by making it twice as likely that a type a individual is chosen to die. find the longrun proportion of time during which the population is zero.1 Consider a branching process with immigration. and where P(N = 0) = 1 .2 A species of plant comes in three genotypes AA. In the case where the number of immigrants in each generation is Poisson of parameter A. AA or aa? Suppose it is desired to breed an AA plant.1. Thus in a population of size m containing i type A individuals. and the times taken to serve customers are also independent random variables. Aa and aa. then nl k==O lE(t X n) = ¢(n) ( t) II 1/J (¢ (k) ( t) ) . What should you do? How many crosses would your procedure require.1. Exercises 5. 5. on average? 5.. The main .P and P(N = 1) = p. the probability that some type A is chosen to die is now i/(i + 2(m . This is defined. What is the probability that eventually the whole population is of type A? 5.1. whereas the WrightFisher model changes all m at once. Suppose we begin with just one type A. giving priority to earlier arrivals. by n X n = N 1 + .2 Queues and queueing networks 179 with one individual at a time..
where > O.180 5. an analysis is still possible. and a certain discretetime Markov chain embedded in the queue.I if T < A. 3. with probabilities Jl/(A+Jl) and A/(A+Jl) respectively. Similarly.1 (M/M/I queue) This is the simplest queue of all. We shall find conditions for the stability of the queue. In cases where interarrival times and service times have exponential distributions. Then the number of customers in the queue (Xt)t~O evolves as a Markov chain with the following diagram: Jl III( A Jl •• 111( A i i +1 •• i To see this.2. by exploiting the memorylessness of the Poisson process of arrivals. Denote by T the time taken to serve the first customer and by A the time of the next arrival. If the interarrival times only are exponential. If we condition on J 1 = T. This is explained in the final two examples. We shall also look at the random times that customers spend waiting and the length of time that servers are continuously busy. XJl = i + 1 if T > A. The code means: memoryless interarrival times/memoryless service times/one server. conditional on XJl = j. which events are independent of J 1 . (Xt)t~O . (Xt)t~O turns out to be a continuoustime Markov chain. T . which is exponential of parameter A + JL. and in the stable case find means to compute the equilibrium distribution of queue length.6. Some further variations on queues of this type have already appeared in Exercises 3. This is the context of our first six examples. 3.7.2. and XJl = i . Example 5.3. Applications quantity of interest is the random process (Xt)t~O recording the number of customers in the queue at time t. Let us suppose that the interarrival times are exponential of parameter A. Hence.4.J 1 is exponential of parameter JL and independent of J 1 . The case where i = 0 is simpler as there is no serving going on. suppose at time 0 there are i customers in the queue. In each example we shall aim to describe some salient features of the queue in terms of the given data of arrivaltime and servicetime distributions. so we can answer many questions about the queue.1 and 3. and the service times are exponential of parameter Jl. then AJ1 is exponential of parameter A and independent of J1: the time already spent waiting for an arrival is forgotten.7. conditional on J 1 = A.1. Then the first jump time J 1 is A 1\ T. This is always taken to include both those being served and those waiting to be served.
this is (i + 1) I Jl. Another quantity of interest is the mean waiting time for a typical customer. when A < Jl and the queue is in equilibrium. that is X t ~ 00 as t ~ 00. This sort of argument should by now be very familiar and we shall not spell out the details like this in later examples.::: i) = L{.. Even in more complicated examples where exact calculation is limited.transience and recurrence. except that it does not take jumps below o. Thus. . longrun averages. or we multiply the mean waiting time by the expected number of customers At. and so on.A).2 Queues and queueing networks 181 begins afresh from j at time Jl. convergence to equilibrium. The first calculation is exact but we have not fully justified the sec9nd.\/JL)i = . the mean time to return to 0 is given by so the mean length of time that the server is continuously busy is given by rnO  (llqo) = 1/(Jl . We deduce that if . Thus if A > Jl the queue grows without limit in the long term.\/{JL . Either way we get Atl Jl .\ > Jl then (Xt)t~O is transient. (Xt)t~O is positive recurrent with invariant distribution So when A < Jl the average number of customers in the queue in equilibrium is given by 00 00 lE7r {Xt ) = LJP>7r{Xt i=l . The MIMll queue thus evolves like a random walk.5. once the queue size is identified as a Markov chain. When A < Jl. so the overall mean waiting time is A rough check is available here as we can calculate in two ways the expected total time spent in the queue over an interval of length t: either we multiply the average queue length by t.~o is the claimed Markov chain. Conditional on finding a queue of length i on arrival.A. its behaviour is largely understood. i=l Also. once the Markovian character of the queue is noted we know what sort of features to look for .\). It follows that (Xt)t.
.~o performs a Markov chain with the following diagram: •• ••• ••• 012 A Jl A 2Jl A SJl • s•• ••• s+l A SJl A So this time we obtain a birthanddeath chain.2.. To find an invariant measure we look at the detailed balance equations Hence for i for i = 0.S = s + 1. The number of arrivals by time t is a Poisson process of rate A. Let us suppose that A < SJl.2 (M/M/s queue) This is a variation on the last example where there is one queue but there are S servers. (Xt)O~t~T .182 5../JL so that and the invariant distribution is Poisson of parameter AI Jl. .. The detailed balance equations hold and (Xt)t~O is nonexplosive. The total service rate increases to a maximum SJl when all servers are working. so by Theorem 3. It is transient in the case A > SJl and otherwise recurrent.1. and consider the queue in equilibrium. By an argument similar to the preceding example. and each departure to a decrease.. S + 2. We emphasise that the queue size includes those customers who are currently being served. The queue is therefore positive recurrent when A < SJl. Let us assume that the arrival rate is A and the service rate by each server is Jl. Applications Example 5.2. so there is an invariant distribution. . There are two cases when the invariant distribution has a particularly nice form: when S = 1 we are back to Example 5. the first service is completed at the minimum of i independent exponential times of parameter Jl.7. The first service time is therefore exponential of parameter iJl. Each arrival corresponds to an increase in X t .1 and the invariant distribution is geometric of parameter AI Jl: When S = 00 we normalize 1r by taking 1ro = e)... the queue size (Xt)t.3 for any T > 0. Then if i servers are occupied.
and hence the longrun proportion of calls that are lost. and then by server . This is slightly counterintuitive.2. This time we get a truncated Poisson distribution By the ergodic theorem. Compare this example with the bus maintenance problem in Exercise 3.7.2 Queues and queueing networks 183 and (XTt)O~t~T have the same law. additional calls are lost.3 (Telephone exchange) A variation on the M/M/s queue is to turn away customers who cannot be served immediately. is given by This is known as Erlang's formula. as one might imagine that the departure process runs in fits and starts depending on the number of servers working. Example 5. the number of departures by time t is also a Poisson process of rate A.5. the longrun proportion of time that the exchange is full. in equilibrium. Example 5. This might serve as a simple model for a telephone exchange. It follows that. in equilibrium.1. Instead. The maximum queue size or buffer size is s and we get the following modified Markov chain diagram: A • JL I( A • 2JL I( A • 012 sl s We can find the invariant distribution of this finite Markov chain by solving the detailed balance equations.4 (Queues in series) Suppose that customers have two service requirements: they arrive as a Poisson process of rate A to be seen first by server A. as in the last example.2. where the maximum number of calls that can be connected at once is s: when the exchange is full. it turns out that the process of departures. is just as regular as the process of arrivals.
184
5. Applications
B. For simplicity we shall assume that the service times are independent exponentials of parameters Q and j3 respectively. What is the average queue length at B?
Let us denote the queue length at A by (Xt)t>o and that by B by (¥t)t>o. Then (Xt)(~O is simply an MIMll queue. If A > Q, then (Xt)(~O is transient so there is eventually always a queue at A and departures form a Poisson process of rate a. If A < a, then, by the reversibility argument of Example 5.2.2, the process of departures from A is Poisson of rate A, provided queue A is in equilibrium. The question about queue length at B is not precisely formulated: it does not specify that the queues should be in equilibrium; indeed if A ~ Q there is no equilibrium. Nevertheless, we hope you will agree to treat arrivals at B as a Poisson process of rate Q /\ A. Then, by Example 5.2.1, the average queue length at B when Q /\ A < j3, in equilibrium, is given by (Q /\ A) 1((3  (Q /\ A)). If, on the other hand, Q /\ A > (3, then (¥t)t~O is transient so the queue at B grows without limit. There is an equilibrium for both queues if A < Q and A < j3. The fact that in equilibrium the output from A is Poisson greatly simplifies the analysis of the two queues in series. For example, the average time taken by one customer to obtain both services is given by
1/(a  A)
+ 1/(j3 
A).
Example 5.2.5 (Closed migration process)
Consider, first, a single particle in a finite statespace I which performs a Markov chain with irreducible Qmatrix Q. We know there is a unique invariant distribution Jr. We may think of the holding times of this chain as service times, by a single server at each node i E I. Let us suppose now that there are N particles in the statespace, which move as before except that they must queue for service at every node. If we do not care to distinguish between the particles, we can regard this as a new process (Xt)t~O with statespace Y= N1 , where X t = (ni : i E I) if at time t there are ni particles at state i. !n fact, this new process is ~lso ~ Markov chain. To describe its Qmatrix Q we define a function bi : I + I by
Thus bi adds a particle at i. Then for i =I j the nonzero transition rates are given by
n
E
I,
i,j E I.
5.2 Queues and queueing networks
Observe that we can write the invariant measure equation 1rQ form
1ri
185
= 0 in the
(5.1)
L qij = L 1rjqji·
j#i j#i
For n
=
(ni : i E I) we set
1f(n)
=
II
iEI
1rfi.
Then
1f(8i n)
L q(8 n, 8 n) II 1r~k (1riLqji)
i
j
=
j#i
kEI
j#i
(L:: 1rjqji)
j#i
=
L 1f(8 n)q(8 n, 8m).
j j
j#i
Given mEl we can put m = 8i n in the last identity whenever mi summing the resulting equations we obtain
~
1. On
1f(m)
L
n#m
q(m, n) =
L 1f(n)q(n, m)
n#m
so 7r is an invariant measure for Q. The total number of particles is conserved so Q has communicating classes
and the unique invariant distribution for the Nparticle system is given by normalizing 7f restricted to eN.
Example 5.2.6 (Open migration process)
We consider a modification of the last example where new customers, or particles, arrive at each node i E I at rate Ai. We suppose also that customers receiving service at node i leave the network at rate /li. Thus customers enter the network, move from queue to queue according to a Markov chain and eventually leave, rather like a shopping centre. This model includes the closed system of the last example and also the queues
186
5. Applications
in series of Example 5.2.4. Let X t = (X; : i E I), where X; denotes the number of customers at node i at time t. Then (Xt)t>o is a Markov chain in Y= N I and the nonzero transition rates are given by
q(n, bin)
=
for n E I and distinct i and /lj > 0 for some j; then Q is irreducible on I. The system of equations (5.1) for an invariant measure is replaced here by
1ri
= /lj states i, LEI. We shall ass':,me that Ai > 0 for some
Ai, q(bi n , bjn) q(bjn, n)
= qij,
(f.1,i
+ L qi j )
j#i
= Ai
+ L 1rjqji·
j#i
This system has a unique solution, with 1ri > 0 for all i. This may be seen by considering the invariant distribution for the extended Qmatrix Q on I U {8} with offdiagonal entries
q8j
=
Aj,
qij
= qij,
qi8
= /li·
On summing the system over i E I we find
L
iEI
1rif.1,i = L
iEI
Ai.
As in the last example, for n = (ni : i E I) we set
1f(n) =
II 1r~i.
iEI
Transitions from m E I may be divided into those where a new particle is added and, for each i E I with mi 2:: 1, those where a particle is moved from i to somewhere else. We have, for the first sort of transition
1f(m) Lq(m,8j m) = 1f(m) LAj
JEI JEI
= 1f(m)
L 1rjf.1,j = L 1f(8j m)q(8j m, m)
JEI JEI
and for the second sort
1f(8i n) (q(8in, n)
=
+ L q(8i n, 8j n))
j#i
II 1r~k (1ri (f.1,i + L
kEI j#i kEI j#i
qij))
=
II 1r~k ( Ai + L 1rjqj i)
+ L 1f(8j n)q(8j n, 8i n).
j#i
= 1f(n)q(n, 8i n)
5.2 Queues and queueing networks
On summing these equations we obtain
1f(m)
187
L: q(m, n) = L: 1f(n)q(n, m)
n#m n#m
so 7r is an invariant measure for Q. If 1ri < 1 for all i then 7r has finite total mass niEI(l1ri), otherwise the total mass if infinite. Hence, Qis positive recurrent if and only if 1ri < 1 for all i, and in that case, in equilibrium, the individual queue lengths (Xi: i E I) are independent geometric random variables with
Example 5.2.7 (M/G/! queue)
As we argued in Section 2.4, the Poisson process is the natural probabilistic model for any uncoordinated stream of discrete events. So we are often justified in assuming that arrivals to a queue form a Poisson process. In the preceding examples we also assumed an exponential servicetime distribution. This is desirable because it makes the queue size into a continuoustime Markov chain, but it is obviously inappropriate in many realworld examples. The service requirements of customers and the duration of telephone calls have observable distributions which are generally not exponential. A better model in this case is the MIGll queue, where G indicates that the servicetime distribution is general. We can characterize the distribution of a service time T by its distribution function
F(t) = JP>(T
or by its Laplace transform
~
t),
(The integral written here is the LebesgueStieltjes integral: when T has a density function f(t) we can replace dF(t) by f(t)dt.) Then the mean service time Jl is given by
Jl
= E(T) = £'(0+).
To analyse the MIGll queue, we consider the queue size X n immediately following the nth departure. Then
(5.2)
188
5. Applications
where Yn denotes the number of arrivals during the nth service time. The case where X n = 0 is different because then we get an extra arrival before the (n + 1)th service time begins. By the Markov property of the Poisson process, Y1 , Y 2 , • •• are independent and identically distributed, so (Xn)n~O is a discretetime Markov chain. Indeed, except for visits to 0, (Xn)n~O behaves as a random walk with jumps Yn  1. Let Tn denote the nth service time. Then, conditional on Tn = t, Y n is Poisson of parameter At. So
and, indeed, we can compute the probability generating function A(z) = E(zYn ) =
=
1
1
00
E(zYn
I Tn
=
t)dF(t)
00
eAt(lZ)dF(t) = £('\(1  z)).
Set p = E(Yn ) = AJl. We call p the service intensity. Let us suppose that p < 1. We have
X n = X o + (Y1 + · · · + Yn )  n + Zn
where Zn denotes the number of visits of X n to 0 before time n. So
E(Xn ) = E(Xo)  n(l  p)
Take X o = 0, then, since X n
~
+ E(Zn).
0, we have for all n
o<
1  p ~ E(Znln).
+ 00
By the ergodic theorem we know that, as n
E(Znln)
where mo is the mean return time to
+
limo
Hence
00
o.
mo ~ 1/(1  p) <
showing that (Xn)n~O is positive recurrent. Suppose now that we start (Xn)n~O with its equilibrium distribution Set G(z) = E(zX n ) =
Jr.
I:
i==O
00
1riZi
p) = p + A2£"(0+ )/2(1 .z)/(l. The fact that generating functions work well here is due to the additive structure of (5.A).p. we can now obtain. as we found in Example 5.z)G(z) (5. the full equilibrium distribution. G(z) we must therefore have 1t"o = 1 .z) + A(z) .z) . in principle. rno = 1/(1 .z)A(z)/(A(z) .p.A'(l) = 1. To obtain the mean queue length we differentiate (5. so (A(z) . Since A is given explicitly in terms of the servicetime distribution.1.)/2(1 .l)G(z) = (1  p){ A'(z)(l .5.p) = p + A2E(T 2)/2(1  p).2 Queues and queueing networks then 189 = lE(ZYn+l) (1rOZ + L 1r Z i 00 i ) i==l = A(z) (1t"OZ + G(z) 1t"o) = 1t"oA(z)(l z). . as z i 1 + (A(z) . Since G(l) and = 1 = A(l).2.z) 1.z) (lp)A(z) { (A' (z) 1) (1. In the case of the M/M/1 queue p = AIJl. E(T 2) = 2/ Jl2 and E(Xn ) = p/(l .2(1p)2 · Hence E(Xn ) = G'(l) = p + A"(l.A(z)}. then substitute for G(z) to obtain G' (z) = (lp )A'(z) (1.z} .2).z).3) By I'Hopital's rule. (A(z) .3) (A(z) .p) = (1  p)(l .z) (A(z) _ Z)2 By l'Hopital's rule: lim (A'(z)l)(lz) zjl + A(z)z lim (A(Z)Z)2  zjl A"(z)(lz) _ _ _A_"_(l__) 2(A'(z)1) (A(z)z) .p) = A/(Jl .z)G'(z) = (A'(z) .
82 .t(lB(w»dF(t) = L(w + .B(w))).t.. as above. . since the customers in the queue at time 0 are precisely those who arrived during the queueing and service times of the departing customer. Consider the queue (Xn)nEZ in equilibrium. conditional on Q + T = t.190 5.. Consider the Laplace transform Let T denote the service time of the first customer in the busy period.. we have 8=t+81 + . Then.L(w))).WS IT = t)dF(t) 00 ewte>. . +8N.z)) where M is the Laplace transform On substituting for G(z) we obtain the formula M(w) = (1 .x(1 .z))L(A(l..A(Q+T)(lZ)) = M(A(l. Hence G(z) = E(e. which is Poisson of parameter At. with the same distribution as 8. Differentiation and l'Hopital's rule. Hence B(w) = = 1 1 00 E(e.xL" (0+) 2 (1 + AL' (0+)) 2 We now turn to the busy period 8.x(1. lead to a formula for the mean queueing time E(Q) = M'(O+) = . and time T being served.p)w/(w .the queueing time of a typical customer and the busy periods of the server. Suppose that the customer who leaves at time 0 has spent time Q queueing to be served. X o is Poisson of parameter >. where N is the number of customers arriving while the first customer is served. and where 8 1 .. are independent.. Then conditional on T = t. Applications We shall use generating functions to study two more quantities of interest .
. service is incomplete at time t with probability 1 1 p=JP>(T>s)ds=tot it it 0 (lF(s))ds. so all customers in fact receive service at once./(n .An are independent and uniformly distributed on the interval [0.6. What.p)tt. Hence. is the distribution of the number X t being served at time t? The number Nt of arrivals by time t is a Poisson random variable of parameter At. say. X t is binomial of parameters nand p. We condition on Nt = n and label the times of the n arrivals randomly by AI.. . conditional on Nt = n.t(>. then. with a common distribution function F(t) = lP(T ~ t). Example 5.pt)k /k! L('\(1 . . Then.8 (M/G/oo queue) Arrivals at this queue form a Poisson process.F(s))ds. The analysis here is simpler than in the last example because customers do not interact. by Theorem 2.AB'(O+)) = Jl(l + AE(S)) so the mean length of the busy period is given by E(S) = Jl/(l . t]. For each of these customers.k)! n==k 00 eAPt(Apt)k /k! So we have shown that X t is Poisson of parameter ..\ it (1 .2. .4.p).An. Service times are independent.2 Queues and queueing networks 191 Although this is an implicit relation for B(w). of rate A. There are infinitely many servers.5. . AI. we can obtain moments by differentiation: E(S) = B'(O+) = £'(0+)(1 .. Then P(Xt = k) = L P(Xt = k I Nt = n)P(Nt = n) n==O 00 = = k e>. Suppose there are no customers at time O. .
Chichester. Given that X n = i. The warehouse manager restocks to capacity for the beginning of period n + 1 whenever X n ~ m. the queue size has a limiting distribution. P. For further reading see Reversibility and Stochastic Networks by F. which is Poisson of parameter . . for some threshold m. Here we present some examples involving the management of a resource: either the stock in a warehouse.3 Markov chains in resource management Management decisions are always subject to risk because of the uncertainty of future events.. See also Exercise 3. If one can quantify that risk.1 (Restocking a warehouse) A warehouse has a capacity of c units of stock. excepting some peculiar demand structures.xE(T). Example 5.3.c}. then (Xn)n~O is a Markov chain.. often involving a Markov chain. Kelly (Wiley. and.10.7. .192 Recall that 5. Thus (Xn)n~O satisfies if X n ~ m ifm < X n ~ c. there is a demand for D n units of stock. or the water in a reservoir. ••• are independent and identically distributed. D 2 . c.1 on the maintenance of unreliable equipment. Hence (Xn)n~O has a unique invariant distribution 7r which determines the longrun proportion of time in each state. In each time period n. which is met if possible. 1978). Applications Hence if E(T) < 00. is irreducible on {O. the expected unmet demand in period n + 1 is given by if i ~ m if m < i :s. We denote the residual stock at the end of period n by X n . . or the reserves of an insurance company. 5. Let us assume that D 1 . 1. perhaps on the basis of past experience. The statistical problem of estimating transition rates for Markov chains has already been discussed in Section 1. then the determination of the best action will rest on the calculation of probabilities.
1. The warehouse manager may want to compute these quantities in order to optimize the longrun cost ar(m) + bu(m) where a is the cost of restocking and b is the profit per unit. by 1/8 1/2 ( 1/4 1/8 1/8 1/4 1/2) (1/8 1/8 1/4 1/2) (1/8 1/8 1/2 0 0 1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/2 0 1/4 1/4 1/2 0 1/8 1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/2 1/8 1/8 1/4 1/4 1/4 1/4 1/2) 1/2 1/2 1/2 with invariant distributions (1/4.3 Markov chains in resource management Hence the longrun proportion of demand that is unmet is u(m) = 193 L 7l"i i· U c i==O The longrun frequency of restocking is given by r(m) = L7l"i. We shall discuss in detail a special case where the calculations work out nicely.5.1/4. 00 lE((D . u(m) decreases and r(m) increases. Hence (1/6. (1/8.2. .i .2.1. k=l k=l The transition matrices for m = 0.1/4. i==O m Now as m increases. so possible threshold values are m = 0.1/4).i)+ 2: k) = LIP(D 2: i + k) = 2..i)+) = LIP((D . and that the demand satisfies P(D ~ i) = 2. it is a relatively simple matter to write down the (c + 1) x (c + 1) transition matrix P for (Xn)n~O and solve 1rP = 1r subject to E~==o 1ri = 1.1/8. respectively. u(O) = 1/4. u(2) = 1/8 .1/3). Suppose that the capacity c = 3. There is no general formula for 1r.2 are given.1/4.1/2).i Then 00 for i = 0. u(l) = 1/6..1/3.1.. Suppose that the profit per unit b = 1. but once the distribution of the demand is known.1/6.
So we would like to calculate 7r. 1.z) .1)+ + A n+l . to minimize the longrun cost ar(m) 2 m = + u(m) we should take { 1 o if 1/4 < a ~ 1 if 1 < a.2. Example 5.z)A(z)j(A(z) . Hence we know that the longrun behaviour of (Xn)n~O is controlled by its unique invariant distribution 7r. If we assume that An.z)A(z)j(A(z) . by the argument used in Example 5..discrete time) We are concerned here with a storage facility.2 (Reservoir model . then is positive recurrent and the invariant distribution 7r satisfies (Xn)n~O L where A(z) 00 7riZi = (1 . B n and c are integervalued and that AI. no water can be supplied. For example. ••• are independent and identically distributed. whether or not E(A n ) < 1.EA n )(1 . In each time period n. 00 In fact. A 2 . r(l) = 1/3. of finite capacity c. . if a ~ 1/4 r(2) = 1/2.3. A simplifying assumption which makes some calculations possible is to assume that consumption in each period is constant. and that our units are chosen to make this constant 1. .c}. Therefore. likewise B I . ••• . We assume that newly available resources cannot be used in the current time period. surplus water is lost. B 2 . whose transition probabilities may be deduced from the distributions of An and B n . assuming irreducibility. When the reservoir is full. Hence. Then the infinite capacity model satisfies a recursion similar to the M/G/l queue: X n+1 = (Xn . the longrun proportion of time that the reservoir is empty is simply 7ro. Applications r(O) = 1/4. An units of resource are available to enter the facility and B n units are drawn off.7. the equation lIiZi L i=O = (1 .z) i=O = E(zA n ). When the reservoir is empty.. for example a reservoir. Then the quantity of water X n in the reservoir at the end of period n satisfies X n+1 = ((Xn Bn+I)+ + A n+l ) Ac. then (Xn)n~O is a Markov chain on {O.194 and 5. if E(A n ) < 1.
we can deduce for the finitecapacity model that the longrun proportion of time in state i is given by vi/(vo +. Note that (Wt)t~O can enter [0. by the strong Markov property. Hence in the positive recurrent case where AE(8n ) < 1. As in the preceding example we can obtain the finite capacity model by observing (Wt)t~O whilst in [0. Example 5.. Hence c should be chosen large enough to make ° lIO/(lIo + . c] only through c. For a reservoir of infinite capacity.xIE(Sn)). € > say.. The periods when the reservoir is empty correspond to idle periods of the queue.3.. because it illustrates a surprising and powerful connection between reflected random walks and the maxima . and the level c equation is redundant. c]. We assume that there is a continuous demand for water of rate 1. Note that (Xn)n~O can only enter {O.. but we shall not pursue this here..z and equate powers of z: the resulting equations are the equilibrium equations for (Xn)n~O: Vo = lIo(ao l/i + a1) + V1 aO + = l/i+l a O L j=O i l/j a ij+1.5.3 (Reservoir model . 1... The quantities of water 8 1 . +ve ).. . In the case where IE(A n) < 1. The problem faced by the water company is to keep the longrun proportion of time 1ro(c) that the reservoir is empty below a certain acceptable fraction.. multiply by A(z) .1. The next example is included. To see this. so II must have infinite total mass. for i ~ 1 where ai = P(A n = i). which is true if E(A n ) > 1. . (Xn)n~O observed whilst in {O. Hence. in the long run.. 1. . In fact.3 Markov chains in resource management 195 serves to define a positive invariant measure v for (Xn)n~O. this is true in general as the equilibrium equations for the finitecapacity model coincide with those for v up to level c . Then (Xn)n~O is transient. it is to be hoped that. supply will exceed demand.c} is simply the finitecapacity model.82 . . In reality..82 . arriving each time are assumed independent and identically distributed.. . the quantity of water held (Wt)t~O is just the stored work in an M/G/I queue with the same arrival times and service times 8 1 .. the longrun proportion of time that the reservoir is empty is given by IE(Sn)/(l . + lie) < c which is always possible in the transient case.. in part. .c} through c.continuous time) Consider a reservoir model where fresh water arrives at the times of a Poisson process of rate A.
4 (Ruin of an insurance company) An insurance company receives premiums continuously at a constant rate.. We choose units making this rate 1. by induction. Hence.196 5.3.P per unit time. By the strong law of large numbers as n ~ 00. the claims YI. M n has the same distribution as Zn. We know by Example 5. we have + X n+l )+. However. Y 2 . The maximum loss that the company will have to sustain is M= lim M n n+oo where By the argument given above. so does the ..2. Applications of random walks.. where X n = Y n . X 2 . + X n and define (Zn)n~O by Zo = 0 and Zn+l = (Zn Then.T n )+. where Zo = 0 and Zn+l = (Zn + Yn .Tn and Tn is the nth interarrival time. Let Xl. Set p = XE(Yi) and assume that p < 1. • •• denote a sequence of independent. which we now explain. Then in the long run the company can expect to make a profit of 1 .7 that the queuelength distribution converges to equilibrium. identically distributed random variables. there is a danger that large claims early on will ruin the company even though the longterm trend is good. But Zn is the queueing time of the nth customer in the M/G/1 queue with interarrival times Tn and service times Yn . Set Sn = Xl + . Denote by Sn the cumulative net loss following the nth claim. ••• being independent and identically distributed. The company pays claims at the times of a Poisson process of rate X. · + X n . Thus Sn = Xl + . so Zn has the same distribution as M n where Example 5.
. } jumps one step to the right with probability p and one step to the left with probability q = 1 . .10 and Section 4. However. then the choice p = 1 may be best. but incurs a cost c(p) = l/p. Starting from i we must first hit iI.5. incurring no further costs. 1. this may now be obtained by inverting the Laplace transform.2p) 00 Thus for c(p) = lip the choice p = 1/4 is optimal.4 Markov decision processes 197 queueingtime distribution.2 how to calculate in these circumstances the longrun average cost or the expected total cost.p. we know the Laplace transform of the equilibrium queueingtime distribution. then i . The question arises as to how best to do this to minimize our expected costs. Given the lack of memory in the model. ¢(p) = { c(p)/(l . The expected total cost starting from 2 is 2¢(p) since we must first hit 1. In principle. and seek to minimize ¢(p).. Any value of p E (0. Hence ¢(p) so that = c(p) + 2p¢(p) for p < 1/2 for p ~ 1/2. The general discussion which follows will make rigorous what we claimed was reasonable.4 Markov decision processes In many contexts costs are incurred at a rate determined by some process which may best be modelled as a Markov chain. Suppose now that we are able to choose the transition probabilities for each state from a given class and that our choice determines the cost incurred. with expected cost 8.7. Hence The probability of eventual bankruptcy is P(M > a). this makes it reasonable to pick the same value of p throughout. If we are only concerned with minimizing costs over the first few time steps. 2.4. in the long run the only way to avoid an infinite total cost is to get to O. Example 5. We have seen in Section 1. . Also by Example 5. and so on. The walker on reaching 0 stays there.1 A random walker on {O. where a denotes the initial value of the company's assets.2.2. the expected total cost starting from 1.1] may be chosen. 5.
. . . the probability law pu makes (Xn)n>O Markov. a transition matrix P(a) = (Pij(a) : i.Xn )). a). .j E I) and a cost function c( a) = (Ci (a) : i E I).. let us suppose given some distribution A = (Ai: i E I) and. Each policy u determines a probability law pu for a process (Xn)n~O with values in I by (i) PU(Xo = io) = Aio ... a) ~ 0 for all i and a.Xn =i n ) =Pi n in +l(u n (i o .Un(Xo. jE! . We abuse notation and write u also for the associated policy given by Under a stationary policy u. . Applications Generally. . Define also the value function V*(i) = infVU(i) U which is the minimal expected total cost starting from i. given by 00 VU(i) =E U LC(Xn. A stationary policy u is a function u : I ~ A. The basic problem of Markov decision theory is how to minimize expected costs by our choice of policy.1. To get a process we must choose a policy. with transition probabilities P"tj = Pij (u(i))· We suppose that a cost c(i. (ii) PU(Xn+ 1 =in+11 X o =io. for each action a E A. Then we associate to a policy u an expected total cost starting from i. though so far we have no process and when we do it will not in general be Markov.. .. a Then the minimum expected cost incurred before time n = 2 is V2 (i) = i~{ c(i. a policy U is a sequence of functions n=O. . n=O So that this sum is well defined.. that is. a way of determining actions by our current knowledge of the process.198 5. a) = ci(a) is incurred when action a is chosen in state i.. These are the data for a Markov decision process.in)). a) + LPij(a)V1 (j)}.. The minimum expected cost incurred before time n = 1 is given by VI (i) = inf c(i. Formally. we assume that c(i.2.
possibly infinite. We make three technical assumptions to ensure this.00) are continuous.1. is in fact the value function V*(i). for all but finitely many j. Voo(i) :::. This is not always true.4) It is easy to see by induction that Vn (i) ~ Vn +l (i) for all i. A simple case where (i) and (ii) hold is when A is a finite set.6) Proof. i~{ c(i.5. so Vn (i) increases to a limit Voo(i).1]. a) + LPij(a)VnU)}. being the limit of minimal expected costs over finite time intervals.4 Define inductively Markov decision processes 199 Vn+l (i) = i~f{ c(i. u(i)) + LPij (u(i))VooU)· jEI (5.5) It is a reasonable guess that Voo(i). with A = (0. letting n ~ 00 and then minimizing over a.4. (iii) for each i. for all a E A we have Pij (a) = 0. jEI (5.a if j = iI { otherwise. so let us assume that Voo(i) ~ B < 00. jEI (5. unless we can show that the inequality (5. There is a stationary policy u such that Voo(i) = c(i. j the functions Ci : A ~ [0.a) + LPij(a)Vn(j)} jEJ . We have Vn+l(i) :::.5) is actually an equality. ci(a) = l/a and a ifj=i+1 1 Pij ( a) = 0 . It is easy to check that the assumptions are valid in Example 5. We assume that (i) for all i. Then Vn+l(i) = !~ {C(i.00) and Pij : A ~ [0.4. c(i.2. Lemma 5.a) + LPij(a)VnU) jEI for all a so. (ii) for all i and all B < 00 the set {a : ci(a) ~ B} is compact. If Voo(i) = 00 there is nothing to prove. a) + LPij(a)VOOU)}. with obvious exceptions at i = O.
Let us suppose inductively that Vn(i) ~ V*(i) for all i.uo(i)) + :EPij(UO(i))vu[i](j) jEI where u[i] is the policy given by Hence we obtain VU(i) ~ i~{ c(i. we obtain (5.Xn )) c(i. a) + LPij (a)V* (j) }.8) we find Vn+ 1 (i) ~ V*(i) and the induction proceeds.3. (ii) if u * is any stationary policy such that a (i) Vn(i) i V*(i) as n ~ 00 = u * (i) minimizes c(i.Un (Xo.4) and (5. in the sense that v u * (i) = Proof. a) + LPij (a)V* (j) } jEI and. D Theorem 5. the infimum is attained and Vn+l(i) = c(i. on passing to the limit nk ~ 00 in (5. Then by substitution in the right sides of (5. Hence.200 5.7).8) Certainly. Vo(i) = 0 ~ V*(i).7) for some un(i) E K. O}. · " . Applications where K is the compact set {a : c( i. then u* is optimal. and. . VU(i) = = Ei L n=O c(Xn . by continuity. Hence Voo(i) ~ V*(i) for all i.6). jEI (5.un(i)) + :EPij(un(i))Vn(j) jEJ (5. For any policy u we have 00 V*(i) for all i. say. We have for all i. By compactness there is a convergent subsequence u nk (i) ~ u(i).4. on taking the infimum over u V*(i) ~ i~{ c(i. a) ~ B} and where J is the finite set {j : Pij ¢.a) + :EPij (a)V* (j) jEI for all i.
An alternative means of constructing nearly optimal policies is sometimes provided by the method of policy improvement.3.2. for example Vn for n large. We have (i) VeU(i) ~ VU(i) for all i.3 we have vu* (i) ~ Voo(i) for all i. . part (i) gives an explicit way of obtaining the value function V* and.u(i)) + LPij(u(i))VU(j) jEI ~ c(i.4 (Policy improvement).4 Markov decision processes Let u* be any stationary policy for which 201 Voo(i) ~ c(i. a) + :EPij(a)Vn(j) jEI we get a nearly optimal policy. D v u * (i) for all i The theorem just proved shows that the problem of finding a good policy is much simpler than we might have supposed. Moreover.4. (5. by Theorem 4.2.9) Proof. by Theorem 4.4. provided that for all i.u*(i)) + LPij(U*(i))Voo(j).Ou(i)) + :EPij(Ou(i))vu(j) jEI so VU(i) ~ VeU(i) for all i. (i) We have. We may then hope that.2. jEI Theorem 5.3 VU(i) = c(i. even less that this policy would be stationary. part (ii) identifies an optimal stationary policy.2. Given one stationary policy u we may define another Ou by the requirement that a = (Ou)(i) minimizes c(i. by choosing a = u(i) to minimize c(i. But V*(i) ~ vu* (i) for all i. once this is known. For it was not clear at the outset that there would be a single policy which was optimal for all i. (ii) VenU(i) ! V*(i) as n ~ 00 for all i. Then by Theorem 4.5.a) + :Epij(a)Vu(j). so Voo(i) = V*(i) = and we are done. jEI We know such a policy exists by Lemma 5. In practice we may know only an approximation to V*.
10) with u replaced by (INnl u ..a) + LPii(a)Vu(j) jEI for all i and a. . n=O Define the discounted value function V. which is only relevant to the transient case. then V 9N U(i) = IEy· (Mo) ~ IEY· (MN ) =Ef(VU(XN)) +Eu* ~V*(i) (~c(Xn.U*(Xk))' k=O Recall the notation for conditional expectation introduced in Section 4. It follows that EU· (Mn +1) ~ EU· (Mn ) for all n.9).Un(Xo.U*(Xk)) k=O ~Mn where we used (5. u(j)) > 0. We now seek to minimize the expected total discounted cost 00 V:(i) = EiLOnC(Xn.. c(i. U . u*(Xn )) * jEI nl + LC(Xk.202 5.Xn )).N the process nl M n = VONnU(Xn) + LC(Xk.. are transient. . Hence if we assume (5. accessible from i.(i) = infV:(i).10) Fix N ~ 0 and consider for n = 0.1. Applications (ii) We note from part (i) that VOU(i) :::. . . i = X n and a = u*(Xn ). . The recurrent case is also of practical importance and one way to deal with this is to discount costs at future times by a fixed factor Q E (0. the only states j with positive cost c(j. 1. We have been discussing the minimization of expected total cost. (5. This is because we will have V* (i) = 00 unless for some stationary policy u. We have EU (Mn+ 1 I :Fn ) = LPXni (U*(Xn))VoNnlu(j) + c(Xn .u*(Xn))) D asN~oo.1).
U) jEI for all i.a (i) = 0 and. Introduce VO. a) + a 2: Pij(a)V. (iv) For all stationary policies u we have as n Proof. jEI (5. given a stationary policy u. define another Oau by the requirement that a = (Oau)(i) minimizes c(i. ca(a) = O. Then u* is optimal in the sense that V:* (i) = VC:(i) for all i.5. (ii) The value function VC: is the unique bounded solution to 00 (i) We have Vn.o:(i) = i~f {c(i.a.a) +a 2:Pij(a)V. a) + a 2:Pij(a)V jEJ U U). a) + a 2:Pij (a)Vn.4. at some geometric time of parameter a. a) is uniformly bounded. it jumps to 8 and stays there. incurring no further costs.(i) = i~f{ c(i.11) (iii) Let u* be a stationary policy such that a = u*(i) minimizes c(i. for all i. With obvious notation we have ~ 00 for all i. Thus the new process follows the old until. Theorem 5.5. .U) }. ci(a) = ci(a).4 Markov decision processes 203 In fact. Pia(a) = 1 . the discounted case reduces to the undiscounted case by introducing a new absorbing state 8 and defining a new Markov decision process by Pij (a) = apij (a).o: (in jEJ and. Suppose that the cost function c(i.a(i) i VC:(i) as n ~ V. inductively Vn+l.
although this is sometimes valid.4.4. we use a martingale approach.4. We finish with a discussion of longrun average costs.11). JEI Then V = V~. or even stationary.4.: LC(Xk. This suggests that one approach to minimizing longrun costs might be to minimize L 7rjcU. . for all i. = anEf* (V~(Xn)) ~ Ba n /(l. which is more general.4. a)1 ~ B < 00 for all i and a. Then (}au = U so (iv) will show that u is optimal and V = VC:. We have c(i. but in general the sequence ~ (i) may fail to converge as n ~ 00. JEI But. Pi almost surely. except for the uniqueness claim in (ii). This forces IV: (i) I ~ B for all n.204 5.. with the same limit. by Theorem 4. we do not know in general that the optimal policy is positive recurrent. u(j)).0:) n=O iEf* (VU(Xn )) as n ~ 00. there is a stationary policy u such that V(i) = c(i. we know by the ergodic theorem that 1 nl . So ~ (i) does converge in this case by bounded convergence. So for any stationary policy u we have 00 V:(i) and so = Ei:E o:nc(xn . (ii) and (iii) follow directly from Theorems 5. Applications so parts (i). of We assume that Ic(i. Here we are concerned with the limiting behaviour. Instead. as n ~ 00. u(Xn )) :::. B/(l .2. a) ~ B for some B < 00.3 and 5.5. But given any bounded solution V to (5.a) ~ D 0 Hence (iv) also follows from Theorem 5.u(j)) JEI as n ~ 00. In the case of a stationary strategy u for which (Xn)n>O has a unique invariant distribution 1ru .U(Xk)) ~ L k=O 7r j c U.u(i)) +0: :Epij(u(i))V(j).
EU(Mn+l I F n ) = M n + {C(Xn . Un) + LPXni(Un)W(j)} jEI (V* + W(Xn )) with equality if u = u*. then 2: 1I"f (V* + W(i)) iEI ~ 2:1I"f (C(i'U(i)) + 2:Pii(U(i))W(j)) iEI jEI = 2:1I"fc(i. Firstly.12) Let u* be any stationary strategy such that a = u*(i) achieves the infimum in (5.6. . Therefore So we obtain v* ~ V~(i) + 2 sup IW(i)l/n. (ii) liminfn~oo (i) ~ V* for all i.4.t V* as n ..5. Fix a strategy u and set Un = un(Xo. D IW(i)l/n The most obvious point of this theorem is that it identifies an optimal stationary policy when the hypothesis is met. Then (i) (i) .t 00 for all i. Suppose we can find a constant V* and a bounded function W(i) such that V* + W(i) = i~f{c(i. if u is a stationary policy for which (Xn)n~O has an invariant distribution 1r u . . nl Consider M n = W(X n ) .a) + LPii(a)W(j)} jEI for all i.4 Markov decision processes 205 Theorem 5. v::* v:: Proof.. (5.Xn ). i This implies (ii) on letting n ~ 00. for all u. Two further aspects also deserve comment. When u 2 sup i = u* we also have v::* (i) ~ V* + and hence (i).u(i)) iEI + L1I"jW(j) jEI .12) for each i.nV* Then +L k=O C(Xk' Uk).
U21 U22 U 23 = 0.11.1]: .(i) = V* /(1 .206 so 5. Then we can show that as a i 1 we have V. u(i)) iEI with equality if we can take u = u* . Stochastic Models .a) + W(j) + 0(1 V* + W(i) = a))} i~f {C(i' a) + a LPij(a)W(j)} + 0(1 jEI a) which brings us back to (5. Chichester. Tijms.12) on letting a i 1.Ull U12 U 13 Ulm U2m U3m U2 = U3 O. uniformly distributed on [0. Applied Probability Models with Optimization Applications (HoldenDay. Secondly. Ross.11) we find v* /(1 = so a) + W(i) + 0(1 jEI a) i~f {C(i' a) + a LPij(a)(V* /(1. 1994) for more examples. C. Assume that I is finite and that P( a) is irreducible for all a. Applications v* ~ L 1rfc(i. The interested reader is referred to S. 1970) and to H. 5.5 Markov chain Monte Carlo Most computers may be instructed to provide a sequence of numbers Ul = O. results and references.U31 U32 U 33 written as decimal expansions of a certain length.an algorithmic approach (Wiley. there is a connection with the case of discounted costs. On substituting this into (5. San Francisco. which for many purposes may be regarded as sample values of a sequence of independent random variables.a) + W(i) + 0(1 .a).
Since EiEI Ai = 1 we can partition [0. if u E A ij . the generators of such pseudorandom numbers are in general as reliable an imitation as one could wish of U1(w).m and. We shall now describe one procedure to simulate a Markov chain (Xn)n~O with initial distribution A and transition matrix P. uniformly distributed on [0. U3 (w).1]. G(i. .1] into disjoint subintervals (A ij : j E I) such that Now define functions Go : [0. Nevertheless. . lP(Xn +1 = i n +1 I X o = io. . • •• is a sequence of independent random variables.5. Ul. are actually all integer multiples of 10.5 Markov chain Monte Carlo 207 We are cautious in our language because..Xn = in) = JP>(Un +1 E Ainin+l) = Pi ni n+l so (Xn)n~O is Markov(A. X n+1 Then = G(Xn . and set X o = Go(Uo). P). G : I x [0. U3. more seriously. lP(Xo = i) = lP(Uo E Ai) = Ai.u)=j Suppose that Uo..1] into disjoint subintervals (Ai: i E I) with lengths Similarly for each i E I. of course. we can partition [0.. . U1... U2 . U2. and then might exploit the observed properties of such processes. 1] ~ I by Go(U) =i if u E Ai. This simple procedure may be used to investigate empirically those aspects of the behaviour of a Markov chain where theoretical calculations become infeasible. 1] ~ I. U2 (w). they are usually derived sequentially by some entirely deterministic algorithm in the computer. This makes it worth while considering how one might construct Markov chains from a given sequence of independent uniform random variables... Un+1 ) for n ~ 0. .
X(m) takes values in 8 m . perhaps up to an unknown constant multiple. 1] random variables. and it is desired to compute the number (5. k=l The strong law of large numbers guarantees that this is a good approximation as n ~ 00 and. . where. An alternative approach would be to simulate a large number of independent random variables Xl.. Monte Carlo is another name for computer simulation so this sounds no different from the procedure just discussed... After a general discussion we shall give two examples.208 5.performing the sum (5. known as Markov chain Monte Carlo. unless 7r has product form 1r(X) = II 1r mEA m {x(m)).13) for some given function I = (Ii : i E I). one can obtain error estimates which indicate how large to make n in practice. We are given a distribution 7r = (7ri : i E I). Applications The remainder of this section is devoted to one application of the simulation of Markov chains. Markov chain Monte Carlo is sometimes the only way to simulate samples from 7r. and to approximate (5. the object of primary interest being the invariant distribution of the Markov chain and not the chain itself. for each site mEA. The essential point to understand is that A is typically a large set.. The context for Markov chain Monte Carlo is a statespace in product form I = 8m II mEA where A is a finite set. For the purposes of this discussion we shall also assume that each component 8 m is a finite set. . statistical physics and computer science. especially in statistics. moreover. It is the application which finds greatest practical use. However. A random variable X with values in I is then a family of component random variables (X (m) : mEA). L!(Xk). When 7r does not have product form. For recall that a computer just simulates sequences of independent U[O. making the statespace I very large indeed. simulation from the distribution 7r is also difficult.13) state by state for a start.Xn in I.13) by 1 n . each with distribution 7r. Then certain operations are computationally infeasible . . But what is really meant is simulation by means of Markov chains.
5. assuming aperiodicity and irreducibility. j we want There are many possible choices for P(m) satisfying these equations. When the chosen site is m. not one in the much larger space I. But why should simulating an entire Markov chain be easier than simulating a simple distribution Jr? The answer lies in the fact that the statespace is a product.2:Pij(m) ~ O. We would like Jr to be invariant for P( m). we simulate a new random variable X n + 1 (m) with values in 8 m according to a distribution determined by X n . Indeed. we know. where pij(m) = 0 unless i ~ j. by Theorem 1. Each component Xo(m) of the initial state X o is a random variable in 8 m . Thus at each step we have only to simulate a random variable in 8 m . that as n ~ 00 the distribution of X n converges to Jr. A sufficient condition is that the detailed balance equations hold: thus for all i. The process (Xn)n~O is made to evolve by changing components one site at a time. which is constructed to have invariant distribution Jr.8. given any stochastic matrix R( m) with rij(m) =0 unless i ~ j we can determine such a P(m) by for i =I j. and then pii(m) = 1.3.10. make all components independent. It does not matter crucially what distribution X o is given. Let us write i ~ j if i and j agree. Theorem 1. assuming only irreducibility.5 Markov chain Monte Carlo 209 The basic idea is to simulate a Markov chain (Xn)n~O. but we might. The law for simulating a new value at site m is described by a transition matrix P( m). Indeed. for example. except possibly at site m. j#i .2 shows that with probability 1. Then. and for k =I m we set X n+ 1 ( k) = X n ( k) .
mEA We shall stick with this second choice. Xn)n~O is a Markov chain in A x I. There are two commonly used special cases.1 :E P(m). provided we keep returning to every site. Then (m n . which is easy to ensure in the examples. It is particularly useful in Bayesian statistics.210 5. On taking rij(m) = rji(m) for all i and j we find for i ~ j. Applications This has the following interpretation: if X n = i we simulate a new random variable Y n so that Y n = j with probability rij(m). In practice this may not matter much. where the analysis is simpler to present. We have not yet specified a rule for deciding which site to visit when. generating a sequence of sites (mn)n~O. whereas if 7rj ~ 7ri we adopt the new value with probability 7rj/7ri.1) where N m == ISml. given the other components. if 7rj > 7ri. We might choose to visit every site once and then repeat. Alternatively. i =1= j. then we adopt the new value. Then (Xn)n~O is itself a Markov chain with transition matrix P= IAI. We know that . So we simply resample X n (m) according to the conditional distribution under 7r. This is called the Gibbs sampler. On taking for i ~ j we also find for i ~ j. This is called a Metropolis algorithm. we might choose a site randomly at each step. i =1= j rij(m) == l/(Nm . Let us assume that P is irreducible. This is called a Hastings algorithm. This amounts to choosing another value jm at site m uniformly at random. For definiteness we mention two possibilities. A particularly simple case would be to take for i ~ j. then if Y n = j we set with probability (7ririj(m)/7rjrji(m)) 1\ 1 otherwise.
7) is given by 7r(/l. Green. Example 5. 1987).80. cPo. is then given by Bayes' formula 7r(/l. 7) f (y I /l.2.( 0 )2 /2} exp { T t.. much more can be said. Thus the algorithm works eventually.1 . D. 7). we might assume that /l rv N( 00. Ripley (Wiley.t 7rdi n k=O iEI as n ~ 00 with probability 1. J. The parameters 00 .5. Hence. For example.1 (Bayesian statistics) In a statistical problem one may be presented with a set of independent observations Y1 . and Markov Chain Monte Carlo in practice by W. The recent survey article Bayesian computation and stochastic systems by J. 1996). The Bayesian approach to this problem is to assume that /l and 7 are themselves random variables. Spiegelhalter (Chapman and Hall. by Theorem 1.5. 340. . L 1 n1 L o The posterior density for (/l. and 7 has gamma distribution of parameters 0:0 and . . which is the conditional density given the observations.807}. In practice one is concerned with how fast it works. cPo 1 ). we have f(Xk) . which it is reasonable to assume are normally disn tributed. Then the prior density for (/l.Y . R. It should also be emphasised that there is an empirical side to simulation: with due caution informed by the theory. /l is normal of mean 00 and variance cPo 1 . with a given prior distribution.O )2/2}7 ao . 7 I y) ex 7r(/l. Mengersen (Statistical Science.80). Gilks. j. 7) ex exp{ cPO(/l.5 Markov chain Monte Carlo for all m and all i. . D.10.(Yi  p)2 /2 } T ao l+ n / 2 exp{ {jOT}. London. One then seeks to draw conclusions about /l and 7 on the basis of the observations. For further reading we recommend Stochastic Simulation by B. but useful information of this type cannot be gained in the present general context. 1995) contains many interesting references. 10 (1). P. 0:0 and . Besag. S. pp. We shall not pursue the matter here. .1 exp{ . 7 rv r( 0:0. that is to say. but with unknown mean /l and variance 7. the computer output gives a good idea of how well we are doing. .80 are known. We finish with two examples. Chichester. Higdon and K. Given more information on the structure of 8 m and the distribution 7r to be simulated. 7) ex exp{ ¢o(p. Richardson and D. so also 7riPij 211 = 7rjPji and so 7r is the unique invariant measure for P.
the full conditional distributions still have a simple form 1r(pl 1r(r Y.1). .t f(x)1r(x I y)dx as k .J r(a n . First we simulate X o. We may wish to compute probabilities and expectations. with . ¢. then set Xk+1 = (Jlk+1. p) + t. Then (Xk)k~O is a Markov chain in I with invariant measure 7r(Jl. T I y). We wish to simulate X = (Jl. numerical integration would also be feasible as the dimension is only two. Tk). 00) are not finite sets does not affect the basic idea. We now turn to an elaboration of this example where the Gibbs sampler is indispensible.( 0)2/2} exp { r t.(Yi  p)2 /2) } I'. In any case the computer will work with finite approximations to JR and [0. Applications Note that the posterior density is no longer in product form: the conditioning has introduced a dependence between Jl and T. Of course.Jlk+1). but you may find it reasonable at an intuitive level. I Y..00).8n) where n an = ao + n/2. T) with density 7r(Jl. To make the connection with our general discussion we set 1=81 X 8 2 = JR x [0. The model consists of m copies of the preceding one. say from the product form density 7r(Jl.(Yi ex r CYo l+ n / 2 exp { r (. for all bounded continuous functions f : I + JR.t 00 1 with probability 1.80 + 2)Yi i==l  p)2/2. T I y). with a rate of convergence depending on the smoothness of 7r and f. we first simulate Jlk+1 from 7r(Jl I y. Our final belief about Jl and T is regarded as measured by the posterior density. given X k = (Jlk. .8n = . T).212 5. r) ex exp{ ¢o(p. At the kth stage.80 p)2 /2 } I'. Tk) and then Tk+1 from 7r(T I y.00). This is not an immediate consequence of the ergodic theorem for discrete statespace. and one can show that k1 ~ ~ f(X j ) .J N(On. The fact that JR and [0.Tk+1). Nevertheless. Here the Gibbs sampler provides a particularly simple approach.
with means jjj and common variance 7. .. effect of speeding convergence to the equilibrium distribution. by a Gibbs sampler method.. Thus there are mn independent observations }!ij.m. . and j = 1. . so the Markov chain approach is the only one available. This has the.7 I y) is infeasible. . numerical integration of 7r(jj. just as in the case m = 1 discussed above. We take these parameters to be independent random variables as before. normally distributed. i=1 j=1 We can construct approximate samples from 7r(jj.8oT}. remain independent.( 0 )2 /2} .. n. Thus one can update all the means simultaneously in the Gibbs sampler. x exp { T t. for j = 1. ..m.. In cases where m is large. as is direct simulation from the distribution. .jjn). The prior density is given by and the posterior density is given by 1r(Jl.8n = f30 + L L(Yij .5. conditional on 7. .5 Markov chain Monte Carlo 213 different means but a common variance. T I y) ex exp { ¢o ~(Jlj . . Note that..Jlj)2/2.1 . with Let us write jj = (jj1. 7 I y). . ~(Yij Hence the full conditional distributions are where n m On = 00 + mn/2. where i = 1.Jlj)2 /2} TQoHmn/2 exp{ . . the means jjj..
5.1 . As (3 ! the weighting becomes uniform..0.= {(ml' m2) + m2 E A : ml + m2 E A : ml is even}.))2 where the sum is taken over all pairs {m.1. In fact.214 5. m / } ~ A with 1m . where simulation is still possible by simple modifications of the methods presented here. . First we describe a Gibbs sampler.. . as (3 i 00 the mass concentrates on configurations x where H(x) is small. . the Ising model is rather well understood theoretically.1. We write I+ and for each (3 > ° = {x E I : x(m) = 1 for all m E 8A} define a probability distribution (rr(x) : x E I+) by 1t"(x) ex e(3H(x). Note that H(x) is small when the values taken by x at neighbouring sites are predominantly the same. then ° In particular.. For x E A define H(x) = ! I:: (x(m) . whereas if sinh 2{3 > 1 there is a residual effect of the boundary values on X(O). A famous and deep result of Onsager says that if X has distribution 1t". is odd} . This is one of the fundamental models of statistical physics. Here we consider the problem of simulating the Ising model. whereas. Simulations may sometimes be used to guide further developments in the theory. called the Ising model.N}2 and the configuration space = AN\A N. if sinh 2(3 ~ 1.. uniformly in N.2 (Ising model and image analysis) Consider a large box A = AN in 71 2 A with boundary 8A = {N. .m'l = 1. but there are many related models which are not.x(m. Consider the sets of even and odd sites A+ = {(ml' m2) A. Applications Example 5. the fact that X is forced to take boundary values 1 does not significantly affect the distribution of X(O) when N is large. or even to detect phenomena quite out of reach of the current theory.
it is easy to simulate from 1r(x+ I x) and likewise from 1r(x. we change the sign of Xt(m) with probability p(m. to obtain X n+1 . independently for each m E A+\8A. Let us call the resulting configuration Y n . Next we apply the corresponding transformation to Y n. Given that X n = x.+l with distribution 1r(. given X.. the distribution of X n is approximately 1r. where 1r has an approximate product structure on large scales.Bx(m)s(m) 1\ 1 where x ~ x with x(m) = x(m). We may encode a digitized image on a twodimensional grid as a particular configuration (x(m) : mEA) E I. there is little to choose between them. I x+). We can exploit the fact that the conditional distribution 1r(x+ product form 1r(X+ I x) ex e/3x(m)s(m) I x) has II mEA+\8A where.8 in the Ising model.5 Markov chain Monte Carlo and for x E I set 215 x± = (x(m) : m E A±). where x(m) = 1 for a white pixel and x(m) = 1 for a black pixel. for large n. Convergence is fast in the subcritical case sinh 2.8 < 1. We can again exploit the even/odd partition. simulate X.( m) for the odd sites m E A\8A. we vary the tendency of black pixels .5. Then inductively. The process (Xn)n~O is then a Markov chain in 1+ with invariant distribution 1r. In a Bayesian analysis of twodimensional images. Choose now some simple initial configuration X o in 1+. simulate firstly X~+l with distribution 1r(. I x) and then given X~+l = x+. the Ising model is sometimes used as a prior.I x+). An alternative approach is to use a Metropolis algorithm. for mEA+\8A s(m) = L Im'ml=l x(m'). By varying the parameter . Then according to our general discussion. Note that we did not use the value of the normalizing constant Z = e/3H(x) L xEI+ wllich is hard to compute by elementary means when N is large.x) = (1r(x)/1r(x)) 1\ 1 = e2 . Therefore. Both methods we have described serve to simulate samples from 1r. = x.
black or white. the same methods work. thus obtaining a prior 7r(x).y) (1 . y) are the numbers of sites at which x and y agree and disagree respectively. Observations are now made at each site which record the true pixel. y) and d(x. I y). Thus (3 is a sort of texture parameter.y) where a(x. the same for white pixels. Call the resulting configuration X n + 1 / 2 . The posterior distribution for X given observations Y is then given by 7r(x I y) ex 7r(x)f(y I x) ex e. Although this is not exactly the Ising model.1). independently for each m E A+\8A. y) = (7r(x I Y)/7r(x I y)) = 1\ 1 e. . Then (Xn)n~O is a Markov chain in /+ with invariant distribution 7r(.216 5.6x(m)s(m)((1_ p)/pt(m)y(m) where x ~ x with x(m) = x(m). which we choose according to the sort of image we expect. change the sign of X:(m) with probability p(m. Applications to clump together.f3H (x)pa(x. We describe the appropriate Metropolis algorithm: given that X n = x.2 . x. 'Cleanedup' versions of the observed image Y may now be obtained by simulating from the posterior distribution. with probability p E (0.p)d(x. Next apply the corresponding transformation to X~+1/2 for the odd sites to obtain X n+ 1 .
or a bijection f : N ~ I. . One crucial result which we found impossible to discuss convincingly without measure theory is the strong Markov property for continuoustime chains.6.4. The proofs we have given may be read on two levels.3. which are often more convenient than aalgebras. When interpreted in terms of measure theory.6 Appendix: probability and measure Section 6. In either case we can enumerate all the elements of I . Two important results of measure theory. For much of the book we can do without explicit mention of more general aspects of measure theory. Finally. we discuss a general technique for determining probability measures and independence in terms of 1rsystems. This is because the statespace is at worst countable. 6. . are needed a number of times: these are discussed in Section 6. This is proved in Section 6. The basic framework of measure and probability is reviewed in Sections 6.. except an elementary understanding of Riemann integration or Lebesgue measure.1 Countable sets and countable sums A set I is countable if there is a bijection f : {I. the proofs are intended to be rigorous. the monotone convergence theorem and Fubini's theorem.5. . in Section 6. with or without a measuretheoretic background.n} ~ I for some n E N.1 contains some reminders about countable sets and the discrete version of measure theory.2 and 6.
Any countable union of countable sets is countable.n} for some n E N: this just corresponds to a particular choice of the bijection f. for any two enumerations of I ~1.218 6. . Any finite cartesian product of countable sets is countable. . There would have been no loss in generality had we insisted that all our Markov chains had statespace N or {I. n=l and the result follows on letting N ~ D Since the value of the sum does not depend on the enumeration we are justified in using a notation which does not specify an enumeration and write simply More generally. ~2. The set of all subsets of N is uncountable and so is the set of real numbers JR.. for example tl n for any n. then we can set where .1. We need the following basic fact.J 't n=l n=l 00.1. ~ 0 for all i E I. Given any N E N we can find M ~ Nand N' ~ M such that Then N M N' """ A· n < """ A· < """ A· n L. ••• . Appendix: probability and measure where in one case the sequence terminates and in the other it does not. ~3. Let I be a countably infinite set and let Ai Then. Any subset of a countable set is countable.J In L..J 't L. if we allow Ai to take negative values. we have 00 00 LAin n=l = LAin' n=l Proof. Lemma 6.
D Lemma 6. Then L (LAi j) = L (LAij). iEI jEJ jEJ iEI Proof.1. . Hence and the result follows by symmetry. .j3.1. jji 2 0 that I)Ai + Pi) L Ai + LPi.. = iEI jEJ jEJ iEI The following two results on sums are simple versions of fundamental results for integrals. We take the opportunity to prove these simple versions in order to convey some intuition relevant to the general case. Then LAi(n) i LAi as n iEI iEI t 00. Suppose for each i E I we are given an increasing sequence (Ai(n))n~O with limit Ai. Let I and J be countable sets and let Aij 2 0 for all i E I and j E J.discrete case).discrete case). = iEI iEI Aij iEI By induction. There is no difficulty in showing for Ai. for any finite set J and for 2 0. Let jl. we have L (LAi j) L (LAij). Then as n ~ 00.2 (Fubini's theorem . Lemma 6. .6.j2.1 Countable sets and countable sums 219 allowing that the sum over I is undefined when the sums over I+ and Iare both infinite.. be an enumeration of J. and that Ai ( n) 2 0 for all i and n.3 (Monotone convergence .
(~8i(k)) t.2 Basic facts of measure theory We state here for easy reference the basic definitions and results of measure theory. (ii) A E £ =* AC E £. A aalgebra £ on E is a set of subsets of E satisfying (i) 0 E £.8i(k)) = t. For such A we obtain a measure on the measurable space (I.00] which has the following countable additivity property: The triple (E. n E N with Un En = E and Jl(En ) < 00 for all n. If there exist sets En E £.1 if Ai E [0.00) for all i.I) by setting In fact.220 6. Example 6. Set 8i (1) = Ai(l) and for n ~ 2 set Then 8i (n) ~ 0 for all i and n.I).1 Let I be a countable set and denote by I the set of all subsets of I. Recall that A = (Ai: i E I) is a measure in the sense of Section 1. then we say Jl is afinite. by Fubini's theorem ~Ai(n) = ~ (t. Appendix: probability and measure Proof. A measure Jl on (E. Here AC denotes the complement E\A of A in E. The pair (E. £) is called a measurable space. (~8i(k)) i = L (f 8i (k)) LAi = iEI k=l iEI o D 6. Jl) is called a measure space. . £.2. Let E be a set. so as n ~ 00. we obtain in this way all afinite measures Jl on (I. £) is a function Jl : £ ~ [0. Thus £ is closed under countable set operations.n E N) =* Un An E £. (iii) (An E £.
This measure J. m£+ is closed under countable suprema: (fi E m£+.t(A) for all A E £. It can be shown that there is a unique map ji : m£+ ~ [0.0:.2. B) such that J. n E N) ~ ji(En fn) = En ji(fn). (ii) ji(o:f + (3g) = o:ji(f) + (3ji(f) for all f. where we take on [0. b) = b . (3 ~ 0) ~ o:f + {3g E m£+. Then m£ is a vector space. .6. 00] the aalgebra generated by the open intervals (a.2 221 Let A be any set of subsets of E. Also.b): a. We denote by m£+ the set of measurable functions f : E ~ [0. We denote by m£ the set of measurable functions f : E ~ JR. Then m£+ is a cone (f. 9 E m£+ . The collection of aalgebras containing A is therefore nonempty and its intersection is a aalgebra a(A). £1) and (E2 . which is called the aalgebra generated by A.3 In the preceding example take E = JR and A = {(a. The aalgebra B generated by A is called the Borel aalgebra of JR. The set of all subsets of E is a aalgebra containing A. When the range E 2 is a countable set I we take £2 to be the set of all subsets I by default.2 Basic facts of measure theory Example 6. i It follows that. b). both lim sUPn f nand lim inf n fn are in m£+. Example 6.t is called Lebesgue measure.i E I) ~ SUpfi E m£+. for a sequence of functions f n E m£+. 0:. A function f : E 1 ~ E 2 is measurable if f1(A) E £1 whenever A E £2.t(a. 00].a < b}.g E m£+. The intersection of any collection of aalgebras is again a aalgebra. Let (E 1 . £2) be measurable spaces. b. It can be shown that there is a unique measure J.t on (JR.b E JR. Let (E.2. £) be a measllrable space. (iii) (fn E m£+.a for all a. and so is limn fn when this exists.00] such that (i) ji(lA) = J. When the range E 2 = JR we take £2 = B by default. {3 ~ 0.
. F) is called a random variable.3 Probability spaces and expectation The basic apparatus for modelling randomness is a probability space (0. A 2 disjoint. Appendix: probability and measure For f E mE. Thus F is a aalgebra of subsets of 0 and P : F ~ [0. We use random variables Y : 0 ~ lR to model random quantities.{3 ~ 0. + (3Y) = oE(X) + (3E(Y) for X. given a countable statespace I.and IfI = f+ + f· If jt(lfl) < 00 then f is said to be integrable and we set We call ji(f) the integral of f. It is conventional to drop the tilde and denote the integral by one of the following alternative notations: p(J) = lE r fdp = lXEE f(x)p(dx). then f+. 6. o. (ii) P(AI n A 2 ) = P(A I ) + P(A 2 ) for AI. with distribution I Ai = P(X = i) = p({w : X(w) = i}). r jj. F. with Un An = A. This is simply a measure space with total mass P(O) = 1. (iii) P(A n ) i P(A) whenever An i A. . In (iii) we write An i A to mean Al ~ An ~ . A measurable function X defined on (0.1] satisfies (i) P(O) = 1.f.f.E m£+. In the case of Lebesgue measure one usually writes simply lXEJR r f(x)dx.. f = f+ .222 6. To every nonnegative or integrable realvalued random variable Y is associated an average value or expectation E(Y). Thus we have (i) E(IA) (ii) E(oX = P(A) for A E F. P). set f± = (±f) V 0. ~ Similarly. where for a Borel set B ~ lR the probability that Y E B is given by P(Y E B) = P({w: Y(w) E B}). a random variable X : 0 models a random state. which is the integral of Y with respect to P. Y E mF+.
in Probability with Martingales by D.00] is £2 measurable. (b) x ~ f yE E2 f(x. then we shall discuss some places where they are used. n E N. 6.Y)Jtl(dx~Jt2(dY). £2. as n ~ 00 (fn(x) i f(x) for all x E E) :::} Jt(fn) i Jt(f)· Theorem 6. Yn i Y) :::} IE(Yn ) i IE(Y). ') J l l ') . Theorem 6.00] is £1 measurable for all Y E E 2. £1.4 Monotone convergence and Fubini's theorem Here are the two theorems from measure theory that come into play in the main text. 00] satisfies (i) x ~ f(x. Williams (Cambridge University Press.6.4. Let (E 1 . (ii) Y ~ IXEE 1 f(x. For a realvalued random variable Y the probabilities are sometimes given by a measurable density function p in terms of Lebesgue measure: P(Y E B) = Then for any measurable function L p(y)dy. First we shall state the theorems.00] is £1 measurable. Jt) be a measure space and let (fn)n~l be a sequence of nonnegative measurable functions.1 (Monotone convergence).00] the LAdi iEI where A is the distribution of X. Proofs may be found.00] there is an explicit formula E(J(Y)) = L f(y)p(y)dy. Jl2) be two afinite measure spaces. £. y)Jt1(dx) : E 2 ~ [0. y) : E 1 ~ [0. 1991). y) : E 2 ~ [0.4.00] is £2 measurable for all x E E 1.2 (Fubini's theorem). Then. When X is a random variable with values in I and f expectation of Y = f(X) = foX is given explicitly by E(J(X)) = 223 : I ~ [0. Let (E. Suppose that f : E 1 x E 2 ~ [0. Then (a) y ~ f(x. Y)Jl2(dy) : E 1 ~ [0.y)Jt2(dy~Jtl(dx)=1yEE2 (rxEE f(X. f : lR ~ [0. Jl1) and (E2. for example. (c) r yEE2 JxEE (1 f(x.4 Monotone convergence and Fubini's theorem (iii) (Yn E mF+.
We used monotone convergence in Theorem 2. Jt2) to be the probability space with the measure Pi.2 to see that for random variables Sn ~ 0 we have E(LSn) = LE(Sn) n n and E(exp { .3. The problem lies with the . £1.2 as a defining property of the integral. Jt) is afinite: just take E 2 = {I.X) and if IE(Y) < 00 we can deduce IE(Xn ) ! IE(X). + 9n. £2.224 6.2 to see that Thus we have taken (E 1. 2.1 to see that for a nonnegative random variable Y we have IE(Y) = N+oo lim IE(Y /\ N). } and Jt2( {n}) = 1 for all n.fn = 91 +. But if 0 ~ X n ~ Y and X n ! X then Y . There is an equivalent formulation of monotone convergence in terms of sums: for nonnegative measurable functions 9n we have To see this just take .5 Stopping times and the strong Markov property The strong Markov property for continuoustime Markov chains cannot properly be understood without measure theory. Fubini's theorem is used in Theorem 3. This form of monotone convergence has already appeared in Section 6. n~N In the last application convergence is not monotone increasing but monotone decreasing.. £.L Sn}). 3. . provided that (E..10.. This is also a special case of Fubini's theorem.X.. Jt1) to be [0.X n ) i IE(Y .00) with Lebesgue measure and (E2.L Sn}) n n~N =J~= E(exp { .LSn}) =E(J~= exp { .4.X n i Y . So IE(Y . They are powerful results and very easy to use. 6. Appendix: probability and measure The measurability conditions in the above theorems rarely need much consideration. We used monotone convergence in Theorem 1.
that is to say.tl/n}EFt n forall t~O. We define for stopping times T FT = {A E F : A n {T ~ t} E F t for all t ~ O}. by all sets {X s = i} for s ~ t and i E I. Of course. in general.m ~ T < k2.1. Then both X T and {S ~ T} are FTmeasurable.m ~ t and X k2 rn = X T .5. This turns out to be the correct way to make precise the notion of sets which 'depend only on {Xt : t ~ T}'. which in measure theory is made precise as measurability with respect to some aalgebra. Lemma 6. Hence so X T is FTmeasurable.6. Since (Xt)t~O is rightcontinuous. and so {S ~ T} E Fr· . Note that this certainly implies {T<t}=U{T::. Let (Xt)t~O be a rightcontinuous process with values in a countable set I. Without measure theory the statement that a set A depends only on (X s : s ~ t) does not have a precise meaning.1)2. We have {S>T}n{T~t}= so {S U sEQ. if the dependence is reasonably explicit we can exhibit it. We say that a random variable T with values in [0. in what terms would you require the dependence to be exhibited? So in this section we shall give a precise measuretheoretic account of the strong Markov property. on {T < t} there exists an n ~ 0 such that for all m ~ n. for some k ~ 1. Denote by F t the aalgebra generated by {X s : s ~ t}. Proof.s~t ({T~s}n{S>s})EFt D > T} E Fr. (k . Let Sand T be stopping times of (Xt)t~o.5 Stopping times and the strong Markov property 225 notion of 'depending only on'.00] is a stopping time of (Xt)t~O if {T ~ t} E F t for all t ~ O. but then.
Jm ~ s}) n{t< Jm+d so {X s = i} E At. For all m (Xt)t~o.5. Assume inductively that J m is a stopping time. 0. by events of the form {Yk = i} for k ~ m and i E I or of the form {8k > s} for k ~ m and s > o. Am = then T m and Am are Qmmeasurable and T ml{T<J Tn U Am(t) tEQ +l} = suptlB (t)n{T<J Tn Tn +l} tEQ = (sup tl{T~t}) l{T<J tEQ Tn +l} = Tl{T<J Tn +l} .··· .s~t {Jm ~ s} n {X s =1= X J71J E F t D 0. Lemma 6. t} E F t and A(t) := An {T ~ t} E F t for all t ~ O. Then {Jm +! ~ t} for all t ~ = U sEQ. . Tn Proof. Let T be a stopping time of (Xt)t~O and let A EFT.5. Jo = 0 is a stopping time. the jump time J m is a stopping time of Proof.Y and 8 1 .8m .. n {T < Jm+ 1 }.. both measurable with respect to Qm. Appendix: probability and measure ~ Lemma 6. so is At. so J m + 1 is a stopping time and the induction proceeds. So we can find Bm(t). this implies that At = Ft· For T a stopping time and A E FT we have B(t) := {T :s. . Am(t) E Qm such that B(t) n {T < J m+ 1 } = Bm(t) A(t) n {T < Jm+ 1 } = Am(t) Set n {T < Jm+1}.2. Since these sets generate F t .3.Jk ~ s < Jk+l}U{Ym =i. Fix t ~ 0 and consider Since Qm is a aalgebra. such that T = T m and lA = lA on {T < J m + 1 }. For s ~ t we have {X s = = i} n {t < J m +1} (D\Y k=O k =i. Obviously.226 6. m that is. Then for all m ~ 0 there exist a random variable T m and a set Am. We denote by Qm the aalgebra generated by yo.
. .5.5.Sn ~ 0 IF( {Yo = io.6.5. . conditional on T < ( and X T = i. By Lemma 6. as shown in the diagram . Proof On {T < (} set X = X T +t and denote by (Yn)n~O the jump chain t and by (Sn)n~1 the holding times of (Xt)t~o. all io.:c: : 82 . 81 > S1..: o On {Jm ~ 8 m +1 : i T T < J m + 1 } we have.1 and 6.. D Theorem 6. . 8 1 > S1. Y n = in. Tn ..5. for all A EFT.in E I and all S1.4 (Strong Markov property). .3 we can write T = T m and 1A = 1A on {T < J m + 1 }. By Lemmas 6. Q) and let T be a stopping time of (Xt)(~o. (XT+t)(~O is Markov(8i . . Let (Xt)(~O be Markov(A.... ..8n > sn) It suffices to prove this with {T < (} replaced by {Jm ~ T < Jm + 1 } for all m ~ 0 and then sum over m.2. . . . {Jm ~ T} n {XT = i} E FT so we may assume without loss of generality that A ~ {Jm ~ T} n {XT = i}. . .8n > sn} nAn {T < (} n {XT = i}) = lFi(YO = io..5 Stopping times and the strong Markov property and 227 Am n {T < Jm+d = = U Am(t) n {T < Jm+d tEQ U (A n {T ~ t}) n {T < Jm+d = tEQ An {T < Jm+d as required. Then.. Yn = in.. X IF(A n {T < (} n {X T = i}). where T m and Am are Qmmeasurable.. Q) and independent of FT.. 81 :c: . We have to show that.
. Appendix: probability and measure Now. Consider . that our definitions determine the probabilities of all events depending on the process. A 7rsystem A on 0 is a collection of subsets of 0 which is closed under finite intersections. The constructive approach we have taken should make this seem obvious. 8 m+ 2 > 82. by the Markov property of the jump chain P( {Yo = io.6 Uniqueness of probabilities and independence of aalgebras For both discretetime and continuoustime Markov chains we have given definitions which specify the probabilities of certain events determined by the process.8n > 8 n )P(A n {Jm ~ T < Jm + 1 } n {XT = i}) as required.Y = in. in measuretheoretic terms. 81 > 81. for example hitting probabilities.. D 6. .Yn = in.J m and Am and.228 6. thus We denote as usual by a(A) the aalgebra generated by A. .8n > 8 n } nAn {Jm ~ T < Jm + 1 } n {XT = i}) = p( {Ym = io. F) which agree on a 7rsystem A generating :F.:F) be a measurable space. but it is illuminating to see what has to be done. Then PI = P 2 · Proof. = :F Theorem 6. In this section we shall show.6. If a(A) we say that A generates :F. 8m+ 1 > 81 + (Tm . Ym+n = in.J m }) = Pi(Y = io.J m ). From these specified probabilities we have often deduced explicitly the values of other probabilities. 8 m+ 1 is independent of gm and hence of T m . . Let (O. .8m+ n > 8 n } n Am n {8m+ 1 > T m . Let 0 be a set. .· .1. Let PI and P 2 be probability measures on (0. n o 8 1 > 81. by the memoryless property of the exponential Hence. conditional on Ym = i.
Let Al be a 1rsystem generating F 1 and let A2 be a 1rsystem generating F2. Then a(A) ~ V. You may easily check that any dsystem which is also a 1rsystem is necessarily a aalgebra. B E V and A ~ B) :::} B\A (iii) (An E V. Since V is the smallest dsystem containing A. Suppose that (0.6. Any collection of subsets having these properties is called a dsystem. . this shows VI = V. But this shows V is a 1rsystem. This we do in two stages. F.2 (Dynkin's 1rsystem lemma). E V. Suppose A ~ V.1. Consider first VI = {A E V : A n B E V for all B E A}. An i A) :::} A E V. Proof. Any intersection of dsystems is again a dsystem.3. (ii) (A.6 Uniqueness of probabilities and independence of aalgebras 229 We have assumed that A ~ V.because V is a dsystem. D The notion of independence used in advanced probability is the independence of aalgebras.6. Suppose that Then Fl and F 2 are independent. Since VI = V. since PI and P 2 are probability measures.6. You may easily check that VI is adsystem . Let A be a 1rsystem and let V be a dsystem. Theorem 6. Since A is a 1rsystem. Since A generates F. Moreover. D Lemma 6. You can easily check that V 2 is also adsystem. V has the following properties: (i) 0 E V. the result now follows from the following lemma. A ~ V 2 . We say that Fl and F 2 are independent if The usual means of establishing such independence is the following corollary of Theorem 6. Next consider V2 = {A E V : A n B E V for all B E V}. Hence also V 2 = V. A ~ VI. P) is a probability space and F 1 and F 2 are subaalgebras of F.6. so it suffices to show V is a 1rsystem. so we may without loss assume that V is the smallest dsystem containing A.
230 6. There are two steps..such as the event that (Xn)n~O visits a set of states A. The point about rightcontinuity is that without such an assumption an event such as {Xt = i for some t > O} which might reasonably be considered to depend on (Xt)t~O. Appendix: probability and measure Proof.1 justifies (a precise version) of this claim. Now events of the form {Xto = io. First fix A 2 E A 2 with P(A 2 ) the probability measure > 0 and consider jp> = P on :Fl. is not necessarily measurable with respect to a(Xt : t ~ 0). by Theorem 6. P =P on :F2 .6.1 and 6. .Xn = in} form a 1rsystem which generates the aalgebra a(Xn : n ~ 0).1. so. .1.. So Theorem 6.3 are relevant. An argument given in . In our general discussion of continuoustime random processes in Section 2. Hence:F1 and :F2 are independent.1 we showed that our definition of a discretetime Markov chain (Xn)n~O with initial distribution .6. by Theorem 6. > 0 and consider the probability measure We showed in ~e first step that P(A) = P(A) for all A E A2.6.2 we claimed that for a rightcontinuous process (Xt)t~O the probabilities of events of the form for all n ~ 0 determined the probabilities of all events depending on (Xt)t~o.6. Hence. by Theorem 6.Xtn = in} form a 1rsystem which generates the aalgebra a(Xt : t ~ 0). Next fix Al E :F1 with P(A 1) We have assumed that P(A) = P(A) for all A E AI. .1.6. our definition determines (in principle) the probabilities of all events in this aalgebra. D We now review some points in the main text where Theorems 6.1.\ and transition matrix P determines the probabilities of all events of the form But subsequently we made explicit calculations for probabilities of events which were not of this form ..6.. . We note now that the events {X o = io. so. In Theorem 1.
We conclude that. without some assumption like rightcontinuity. by Theorem 6. • eqio81 ••• eqin18n • Then.Xt = i)JP>(A for all C E F and A E FT.6 Uniqueness of probabilities and independence of aalgebras 231 Section 2. n~O so also F C g. for such events l1J)(B) .Yn = i n . .Yn = i n . and that JP>(C n A I T < (.XT = i) = JP>(C I T < (. On the set = {T < (} define Xt = X T +t and let j = a(Xt : t ~ 0). Assume that (Xt)t~O is Markov(.2 shows that this event is measurable in the rightcontinuous case.6. Then Lemmas 6.Sn > sn}. general continuoustime processes are unreasonable. By T~eore~ 6.6.1 and 6. C I'fr.x.5. Thus 9 ~ F where 9 = a((Yn)n~O' (Sn)n~1). Our jump chain/holding time definition of the continuoustime chain (Xt)t~O with initial distribution . Consider now the method of describing a minimal rightcontinuous process (Xt)t~O via its jump process (Yn)n~O and holding times (Sn)n~1. . · . .. Theorem 6. and coincide by the same argument as = g. X T 9 = a((Yn)n~O' (Sn)n~O).2 show that (Yn)n~O and (Sn)n~1 are Fmeasurable. . . A useful1rsystem generating 9 is given by sets of the form B = {Yo = i o. which is what we did in the proof of Theorem 6. · .1.A'tO"'tO'tl \.x and generator matrix Q may be read as stating that. this definition determines JP> on 9 and hence on F. Set B = {Yo = i o.XT = i) with B as above. Then the conclusion of the strong Markov property states that = i) = lPi(B) IT < (.5..4.5. write CYn)n~O and (Sn)n~O for the jump chain and holding times of (Xt)t~O and set n Thus for F F and 9 are aalgebras on n.· ..3 it suffices to prove the independence assertion for the case C = B.S1 > S1.. Q) and that T is a stopping time of (Xt)t~o. Let us take F = a(Xt : t ~ 0). On the other hand. Finally.4. ••• "'tnl't n I'fr.. JP>(B I T < (. for all i E I {Xt = i} = U {In ~ t < In+d n {Yn = i} E g.5.S1 > S1. we consider the strong Markov property.6.·.Sn > sn}.
Markov Chains with Stationary Transition Probabilities. 1971. S. Markov Chain Monte Carlo in Practice. Statistical Science 10 (1) (1995). W. Bayesian computation and stochastic systems. Carus Mathematical Monographs 22. J. Chichester. 2nd edition. Doyle and J. Markov Processes and Martingales. Wiley. Higdon and K. Chapman and Hall. Ripley.G. Vol 1: Foundations. Spiegelhalter.D. 2nd edition. Ewens. D. Berlin. 1989. Freedman.G. Revuz. Springer. D. S. Chichester. London. San Francisco. 1996. 1984. Wiley. P. Berlin.C. .Further reading We gather here the references for further reading which have appeared in the text. 1994.R. Dover. Stochastic Simulation. New York. Applied Probability Models with Optimization Applications. B. Amsterdam. HoldenDay. Wiley. F. Reversibility and Stochastic Networks. 1970. Richardson and D. 340. San Francisco.J. Snell. 1967. Rogers and D.L. L. This may provide a number of starting points for your exploration of the vast literature on Markov processes and their applications. 1987. W. Springer. D. The Theory of Branching Processes.P. Random Walks and Electrical Networks.J. Green.M. NorthHolland. Markov Chains. Gilks. Mathematical Population Genetics. Chung. Mathematical Association of America. Kelly. Besag. 1979. Ross. Markov Chains.E. Diffusions. Harris.L. Mengersen. P. Williams. 1978. Chichester. K. HoldenDay. T. 1984.
. 1994. Stroock. 1993. Cambridge University Press. Chichester. Stochastic Models .an Algorithmic Approach.C. D. H.Further reading 233 D.An Analytic View.W. Tijms. Wiley. 1991. Williams. Cambridge University Press. Probability Theory . Probability with Martingales.
112 action 198 adapted 129 alleles 175 aperiodicity 40 average number of customers 181 backward equation 62. 16. 82. 111 conditional expectation 129 conductivity 151 continuoustime Markov chains 87 construction 89 infinitesimal definition 94 jump chain/holding time definition 94. 111 closed migration process 184 communicating class 11.Index absorbing state 11. 96 Bayesian statistics 211 biological models 6. 111 absorption probability 12. 170 birth process 81 infinitesimal definition 85 jump chain/holding time definition 85 transition probability definition 85 birthanddeath chain 16 boundary 138 boundary theory for Markov chains 147 branching process 171 with immigration 179 Brownian motion 159 as limit of random walks 164 existence 161 in jRd 165 scaling invariance 165 starting from x 165 transition density 166 busy period 181 capacity 151 central limit theorem 160 charge 151 closed class 11. 97 transition probability definition 94. 168 countable set 217 coupling method 41 current 151 . 9. 97 continuoustime random process 67 convergence to equilibrium 41. 121.
118 explosion for birth processes 83 for continuoustime chains 90 explosion time 69 explosive Qmatrix 91 exponential distribution 70 Gibbs sampler 210 gravity 134. 145 harmonic function 146 Hastings algorithm 210 hitting probability 12 hitting time 12. 97 to j 118 Markov(:A. 176. 124 discounted value function 202 distribution 1 lE. 204 . 118 J. expected time in i between visits to j 35 GaltonWatson process 171 gambling 131 Gaussian distribution 160 generator 166 generator matrix 94. 222 integral form of the backward equation 98 integral form of the forward equation 101 interarrival times 180 invariant distribution 33. 100 for birth processes 84 for Poisson process 78 Fubini's theorem 223 discrete case 219 full conditional distributions 212 fundamental solution 145 1'1. 117 computation of 40 irreducibility 11. Q) 94.t{. 115 flow 151 forward equation 62. expected time in i between visits F n . 169 Green matrix 144. 126 mi. filtration 129 fair price 135 filtration 129 finitedimensional distributions 67 first passage decomposition 28 first passage time 19. 208 Markov decision process 197 expected total cost 198 expected total discounted cost 202 longrun average costs 204 Markov property 3 for birth processes 84 for continuoustime chains 93 for Poisson process 75 martingale 129. expected return time 37. 113 expected return time 37.Index detailed balance 48. 126. 168 Erlang's formula 183 excursions 24 expectation 222 expected hitting time 12. 141. 111 holding times 69 235 I. expectation 222 E(:A). P) 2 Markov(:A.97 Markov chain continuoustime 88 discretetime 2 Markov chain Monte Carlo 206. 111 Ising model 214 jump chain 69 jump matrix 87 jump times 69 last exit time 20 longrun proportion of time 53. exponential of Q 62 effective conductivity 156 electrical network 151 energy 154· epidemic model 173 equilibrium distribution 33. statespace 2 infective 173 integrable 129. 117 ergodic theorem 53. exponential distribution of parameter :A 70 eQ .
O(t). transition matrix of reversed chain 47 P(t). 114. 167 recurrence relations 57 reflected random walks 195 reservoir model 194. 125 rightcontinuous process 67 ruin gambler 15 insurance company 196 selective advantage 176 semigroup 96 semigroup property 62 service times 180 shopping centre 185 simple birth process 82 simulation 206 skeleton 122 statespace 1 . jump matrix 87 1rsystem 228 Poisson process 74 infinitesimal definition 76 jump chain/holding time definition 76 transition probability definition 76 policy 198 policy improvement 201 population genetics 175 population growth 171 positive recurrence 37. 176 nonminimal chains 103 null recurrence 37. 118 Q. rate of leaving i 61 qij. generator matrix of reversed chain 124 qi. transition semigroup 96 P(t). probability 222 P. rate of going from i to j 61 queue 179 MIGll 187 M/G/oo 191 MIMll 180 M/M/s 182 queueing network 183185 queues in series 183 o(t). nstep transition probability 5 II. order notation 63 Ohm's law 151 open migration process 185 optional stopping theorem 130 IP. transition matrix 2 [>.236 Index gravitational 134 in electrical networks 151 with discounted costs 142 potential theory 134 probability generating function 171 probability measure 222 probability space 222 Qmatrix 60 associated to a Markov chain 132 associated to Brownian motion 169 matrix exponentials 105 maximum likelihood estimate 56 measure 1 memoryless property 70 Metropolis algorithm 210 minimal nonnegative solution 13 minimal process 69 monotone convergence 223 discrete case 219 Moran model 177 mutation 6. 195 resolvent 146 resource management 192 restocking a warehouse 192 return probability 25 reversibility 48. 118 potential 138 associated to a Markov chain 138 associated to Brownian motion 169 random chessboard knight 50 random walk on tl d 29 on a graph 49 recurrence 24. semigroup of reversed chain 124 p~j).
number of visits to i before n 53 valency 50 value function 198 weak convergence 164 Wiener process 159 Wiener's theorem 161 WrightFisher model 175 (.policy 198 statistics 55. 167 transition matrix 2 irreducible 11 237 maximum likelihood estimate 56 transition semigroup 165 truncated Poisson distribution 183 unit mass 3 Vi(n). 123 transience 24. 215 stochastic matrix 2 stopping time 19 strong law of large numbers 52 strong Markov property 19. explosion time 69 .Index stationary distribution 33. 114. 93. 117 stationary increments 76 stationary. 227 successrun chain 38 susceptible 173 telephone exchange 183 texture parameter 216 time reversal 47. 211.