Interacting Particle Systems -

An Introduction

Thomas M. Liggett*

Mathematics Department, University of California, Los Angeles, USA

Lectures given at the

School and Conference on Probability Theory
Trieste, 13 - 31 May, 2002

LNS0417001

'tmlmath.ucla.edu
 Abstract
Interacting particle systems is a large and growing field of proba-
bility theory that is devoted to the rigorous analysis of certain types of
models that arise in statistical physics, biology, economics, and other
fields. In these notes, we provide an introduction to some of these
models, give some basic results about them, and explain how certain
important tools are used in their study. The first chapter describes con-
tact, voter and exclusion processes, and introduces the tools of coupling
and duality. Chapter 2 is devoted to an analysis of translation invariant
linear voter models, using primarily the duality that is available in that
case. Chapters 3-5 are concerned with the exclusion process, beginning
with the symmetric case, in which one can also use duality, and fol-
lowing with asymmetric systems, which are studied using coupling and
other monotonicity techniques. At the end, we report on some very
recent work on the stationary distributions of one dimensional systems
with positive drift.
Contents
1 Introduction 5
1.1 Examples 6
1.2 Coupling 10
1.3 Duality 15

2 Linear Voter Models 20

2.1 The recurrent case 22
2.2 The transient case 23

3 Symmetric Exclusion Processes 32

3.1 The recurrent case 33
3.2 The transient case 34
3.3 The general (non-translation invariant) case 38

4 Translation Invariant Exclusion Processes 40

4.1 Stationary product measures 40
4.2 Coupling 42
4.3 Shift invariant stationary measures 43
4.4 Some extensions 46

5 One Dimensional Exclusion Processes 48

5.1 The nearest neighbor case 48
5.2 The general case 51

References 56
 Interacting Particle Systems - An Introduction 5

1 Introduction
These notes are intended to provide a first exposure to the area of interacting
particle systems. More comprehensive treatments of this field can be found
in the author's 1985 and 1999 books, as well as in the references in the latter
book. The main prerequisite for reading these notes is a good background
in measure theoretic probability theory.
The processes we will discuss here are continuous time Feller processes
t]t on the compact configuration space {0,1} S , where S is a countable set.
(Recall that a Feller process is a strong Markov process whose transition
measures are weakly continuous in the initial state.) A very simple such
process is that in which the coordinates r]t(x) evolve according to indepen-
dent two state Markov chains with transitions from 0 to 1 and from 1 to 0
at rate 1. Even this simple example illustrates some of the important differ-
ences between particle systems and more classical Feller processes such as
Brownian motion on Rd. The distribution of Brownian motion at any pos-
itive time is equivalent to Lebesgue measure, and hence these distributions
at different positive times are equivalent to each other. In our example of
independent two state Markov chains, the situation is entirely different. For
example, if r/o = 1, then the distribution at every time is a homogeneous
product measure with time dependent density, so the distributions of the
process at different times are mutually singular with respect to one another.
The main issues to be considered here are:
(a) a description of the class X of stationary distributions for the process,
and
(b) limit theorems for the distribution of r]t as t —> oo.
Settling these issues is naturally very easy in the example of independent
two state Markov chains. The only stationary distribution is the product
measure with density \, and there is convergence to this limit for every initial
distribution.
In general, all that can be said is that since {0,1} S is compact and
the process is Feller, X is always nonempty. Typically, there will be some
stationary distributions that can be written down explicitly, and the task is
to determine whether or not these exhaust all of X. If not, these additional
stationary distributions must usually be constructed by some type of limiting
argument. The set X is convex, and we will denote its extreme points by Xe.
The process r]t is usually described by specifying the rates at which tran-
sitions occur. If S is finite, saying that the transition 77 —> £ (for 77 ^ £)
6 T.M. Liggett

occurs at rate c means that

PfJ(% = C) = c« + o(*) (1)

as t J. 0. If S is infinite, the probability on the left of (1) is typically 0, but
the intuitive meaning of "rate" is similar. One simply replaces the event on
the left of (1) by {qt — ( on A} for large finite sets A C S.
The precise relation between the process and its transition rates is pro-
vided by the infinitesimal generator fi of 7ft. It is a (typically unbounded)
operator defined on an appropriate dense subset of C({0,1}S); £1 is deter-
mined by its values on the cylinder functions, i.e., functions that depend on
finitely many coordinates. For cylinder functions, it takes the form
(2)

where c(r], () is the rate at which transitions occur from 77 to (. Our choices of
rates will guarantee that the series in (2) converges uniformly for cylinder / ' s .
The stationary distributions of the process are determined by the generator
as follows:
Theorem 1. A probability measure [/, on {0,1} 5 is stationary for the process
if and only if
Qfdfi = 0

for all cylinder functions f on {0,1} 5 .

Of course, assumptions on the transition rates must be made in order for
statements of this sort to hold. General conditions under which they do hold
can be found in Chapter I of Liggett (1985). Our approach in these notes
is to make assumptions for each model considered which will guarantee that
statements such as those in Theorem 1 are correct.
1.1 Examples
We next turn to some of the most important models considered in this
field. To describe their transition rates, we need the following notation: If
77 G {0,1} S and x,y € S, then r]x, j]x<Sj € {0,1} S are defined by

{ j](u)
r](y)
r](x)
iiu^x,y,
iiu — x,
ifu = y.
 Interacting Particle Systems - An Introduction 7

Thus f)x is obtained from 77 by changing its value at x, while f)x<Sj is obtained
by interchanging the values at x and y. In the latter case, if r](x) ^ f](y),
the transition 77 —> r]Xjy can be interpreted as moving a particle from x to y
or vice versa.
Example 1. The contact process. Here S is a graph whose vertices have
bounded degree, and A is a positive parameter. Use the notation x ~ y to
mean that the vertices x and y are connected by an edge. Then for each
x G S,
fl
f] —> Vx a t r a t e <
\\\{y~x:V(y) = l}\
Here |yl| denotes the cardinality of the set A. The interpretation is that
sites with r](x) — 1 are infected, while sites with r](x) — 0 are healthy.
Infected sites recover from the infection after an exponential time of rate 1,
while healthy sites become infected at a rate proportional to the number of
infected neighbors. The only "trivial" (in the sense that it can be found by
inspection) stationary distribution is the pointmass 5o on the configuration
77 = 0 .
While we have described the contact process as a model for the spread
of infection, it arises in other contexts as well. For example, it is related to
Reggeon Field Theory in high energy physics, and it is a building block for
more complex models in biology.
Here are some of the answers to our basic questions in the case of the
contact process. They illustrate some of the variety of behavior that even
relatively simple systems can exhibit.
If S is finite, then I — {do}, and r]t is eventually = 0 for any initial
configuration. This follows from finite state Markov chain theory.
If S — Zd, the d-dimensional integer lattice, then there is a critical value

(i) A < \(d) implies that I — {5Q} and r\t —> So weakly for any initial
configuration,
and
(ii) A > \(d) implies that Ie — {5Q, U} for some u ^ 5Q, and r\t —> v weakly
for any initial configuration with infinitely many infected sites.
Remark. \i d — 1, A = A(l), and 770 = 1, then 77^ —> 5o weakly, yet with
probability 1, for every x, r]t(x) — 1 for arbitrarily large i's - see Theorem
8 T.M. Liggett

3.10 of Chapter VI of Liggett (1985) for the latter statement. This illustrates
the fact that weak convergence results often miss interesting a.s. behavior.
If S — Td (d > 2), the tree in which every vertex has d + 1 neighbors,
then there are two critical values satisfying

so that
(i) A < \i(d) implies that I — {5o} and r]t —> So weakly for any initial
configuration,
(ii) Ai (d) < A < A2 (d) implies that Ie is infinite, and
(iii) A > \2(d) implies that le — {60, v} for some v ^ 60, and rjt -> v weakly
for any initial configuration with infinitely many infected sites.

In case (ii), if the initial configuration 77 has finitely many infected sites, then

P^irjt # 0 V* > 0) > 0, but Mx G S, P " ( 3 T so that r]t{x) = 0 V* > T) = 1.

That is, the set of infected sites does not become empty (with positive prob-
ability), but it "wanders out to 00", so that every site is eventually and
thereafter healthy.
The proofs of these results when S — Zd or S — T^ are quite involved,
and can be found in Part I of Liggett (1999). After describing two more
examples, we will return to the contact process to illustrate the use of two
important techniques: coupling and duality.
Example 2. The linear voter model. Here S is an arbitrary countable
set, and p(x,y) are the transition probabilities for a Markov chain on S:

p(x,y)>0 and ^ p ( s , y ) = 1.
y

The transition rates are now given by

77 -> rjx at rate ^ p(x,y).

The interpretation is that sites are individuals who at any time can have one
of two opinions (denoted by 0 and 1) on an issue. At exponential times of
rate 1, the individual at x chooses a y with probability p(x,y), and adopts
 Interacting Particle Systems - An Introduction 9

y's opinion. An alternate interpretation is in terms of spatial conflict. Two

nations control the areas {x : r](x) — 0} and {x : r](x) — 1} respectively. A
flip from 0 to 1 at x, for example, then represents an invasion of x by the
second nation.
The trivial stationary distributions for the linear voter model are the
pointmasses 5o and 5\ on r] = 0 and 77 = 1 respectively. We will see in
Chapter 2 that there is a close connection between the existence of nontrivial
stationary distributions for the voter model and the recurrence properties
and harmonic functions of the Markov chain with transition probabilities
p(x,y)-
Example 3. The exclusion process. Again S is a general countable set
and p(x,y) are the transition probabilities for a Markov chain on S. Now
we must also assume that p(x,y) satisfies

supV)p(a;,y) < 00 (3)

y
x

in order to guarantee that r\t is well defined as a Feller process. The transition
rates are given by

f] ~> Vx,y at rate p(x, y) if rj(x) = 1, rj(y) = 0.

The interpretation of r](x) — 0 or 1 is that site x is vacant in the first case,

and is occupied by a particle in the second case. In the evolution, a particle
at x waits a unit exponential time, and then chooses a y with probability
p(x,y). If y is vacant, it moves to y, while if y is occupied, the particle
remains at x. The generator takes the following form for cylinder functions

«/(»/)= E p(x,v)[f(vx,y)-f(v)]-
r,(x)=l,r)(y)=0
Note that (3) implies that this series converges uniformly, and hence defines
a continuous function on {0,1} S .
Again, 60 and 61 are stationary, and it is often possible to produce other
stationary distributions explicitly. To see that one should expect to have
more stationary distributions for the exclusion process than for the linear
voter model, consider the case in which S is finite. Assuming irreducibility
of the Markov chain with transition probabilities p(x, y), it is easy to see
that Ie — {5o, S\} in the case of the linear voter model, while for the exclu-
sion process, Ie has one element for each 0 < n < \S\ — it is the stationary
10 T.M. Liggett

distribution for the exclusion process restricted to configurations of n par-

ticles. Chapters 3-5 are devoted to a description of I in many cases, but
there are still many interesting open problems in this context. For example,
what analogues of Theorems 3-5 of Chapter 5 can be proved in dimensions
greater than 1? See the discussion at the end of Chapter 4 and the paragraph
following the statement of Theorem 4 of Chapter 5 as well.
The remainder of this chapter is devoted to a discussion of two important
tools that we will use repeatedly, coupling and duality. Their use will be
illustrated in two contexts - Markov chains and the contact process. At the
end, we will derive the duality relations for the voter and exclusion processes
that we will use in the later chapters.
1.2 Coupling
A coupling is a joint definition of two or more stochastic processes on a com-
mon probability space. It is a surprisingly powerful tool. Before discussing
its use in the context of particle systems, we consider two applications of
this technique to Markov chains.
1.2a. Coupling for Markov chains. Suppose first that Xt is a birth and
death chain on the integers. This is a Markov chain that can only move to
nearest neighbors. Let (Xt,Yt) be two copies of this chain that are coupled
in the following way:
(a) Xt and Yt move independently until the first time r (if ever) that
Xt = Yt, and
(b) Xt and Yt move together after time r.
This coupling has the property that XQ < YQ implies that Xt < Yt for all
t > 0. It follows that if / is any bounded increasing function on the integers,
then
x -+ Exf(Xt)
is again an increasing function of x. We will exploit the analogous property
for particle systems many times, beginning with our discussion of the contact
process below.
Since harmonic functions will play an important role in Chapters 2 and
3, our second example is one in which coupling is used to prove that certain
Markov chains have no nonconstant bounded harmonic functions. Recall
that a function / : S —> Rl is said to be harmonic for the Markov chain Xt
if Exf{Xt) — f(x) for all x G S and all t > 0. Implicit in this definition
is the requirement that the expected value be well defined, e.g., that / be
bounded or nonnegative.
 Interacting Particle Systems - An Introduction 11

Theorem 2. Suppose that Xl,...,X™ are independent irreducible Markov

chains on the countable set S with the following property: For each i, two
independent copies of X\ will eventually meet with probability 1. If f is a
bounded harmonic function for the Markov chain Xt — (Xj, ...,X") on Sn,
then f is constant.
Proof. Define a coupling of two copies, Xt and Yt, of the Markov chain in
the following way:
(a) Xt and Yt evolve independently until the first time T\ that X\ — Ytl.
(This will happen eventually by our assumption.)
(b) For t > T\, X\ and Y^ run together, and the other n — 1 coordinates
run independently, until the first time r-i that Xf — Y2.
(c) Continuing in this way, we find a finite stopping time T — Tn so that
for all t>T,Xt = Yt.
Now fix x, y G Sn and use the initial states Xo — x and Yo — y. It follows
that

\f(x) - f(y)\ = \Ef(Xt) - Ef(Yt)\ < E\f(Xt) - f(Yt)\ < 2||/||0OP(r > t).
To complete the proof, let t —> oo.
R e m a r k s , (a) The assumption of Theorem 2 is satisfied, for example, by
any irreducible recurrent random walk on Zl or Z2. To see this, let pt(x,y)
be the transition probabilities for Xf, let X\ and Y% be two independent
copies of the random walk, and write

By recurrence, the integral of the left side above is infinite. Therefore, the
integral of the right side is infinite, and hence the random walk X\ — Y"/ is
recurrent.
(b) We will use repeatedly the fact that irreducible random walks on Zd
have no nonconstant bounded harmonic functions. For many random walks,
this fact follows from Theorem 2. A coupling proof for general random walks
can be found on pages 69-70 of Liggett (1985).
12 T.M. Liggett

1.2b. Coupling for the contact process. Turning to applications to

particle systems, define a partial order on the set of probability measures on
{0,1} 5 by saying that /i < v if

//*</ fdu

for all bounded increasing functions / on {0,1} 5 . This is equivalent to the

existence of a probability measure 7 on {0,1} 5 x {0,1} 5 with marginals /i
and v that satisfies

(Theorem 2.4 in Chapter II of Liggett (1985)). The partial order on {0,1} S

that we are using here is coordinatewise: 77 < ( iff r](x) < ((x) for all x G S.
The contact process is an example of an attractive process, i.e., one for
which two copies started from 77 and C with 77 < ( can be coupled in such a
way that 77^ < (t for all t > 0. (We saw earlier that birth and death chains
are attractive.) The property of the rates that allows for this coupling is the
following: The rate of the transition 77 -> r]x is an increasing function of 77
if 77(3:) = 0 and a decreasing function if 77(3:) = 1. Specifically, writing the
C coordinate above the 77 coordinate, we can use the coupling that has the
following transitions at site x:

at rate 1,

at rate \\{y ~ x : r](y) — 1}|

(
< 0
at rate 1,

!
at rate \\{y ~ x : 77(2/) = 1}|
1
< 1
at rate \\{y ~ a; : C(y) = 1, »?(l/) = 0}|.
0

It follows from the above coupling that if / is increasing on {0,1} 5 , then

so is the function 77 —> Evf(r]t) for any t > 0. To see this, take 77 < £ and
use the coupling to write

Vt < tt < f((t) Ef(m) < Ef(Ct).

 Interacting Particle Systems - An Introduction 13

A consequence of this is that the relation /i < v is preserved by the evolu-

tion: If [it and vt are the distributions of the process at time t with initial
distributions /io and VQ respectively, then

Mo < ^o => Vt< vt (4)

for all t > 0. Again this is immediate, since for increasing / ,

j fdnt = J WfimWo < J E*f{m)dvQ = Jfdut.

In order to develop some consequences of these ideas, it is convenient
to introduce some notation. We will write fiS(t) for the distribution of the
process at time t when the initial distribution is /i. The reason for this
choice is the following. The semigroup S(t) of the process f)t is defined for
continuous functions / by S(t)f(r]) — E^f^t)- We then have

E»f(nt) = J&HTHW = f s{t)fdn = J fd[pS(t)].

Therefore we use the same notation S(t) for the dual objects that operate
on functions or measures respectively.
Since any probability measure /i satisfies /i < #i, it follows from (4) that
/iS(t) < 6iS(t). Applying this to fj, — 6iS(s) and using the Markov property,
it follows that
s) = 6iS(s)S(t) < 6iS(t).
Therefore 5iS(t) J. in t, so that the weak limit

v = lim 6iS(t) (5)

t—>00

exists. (For this, we use the fact that every cylinder function can be written
as a finite linear combination of increasing cylinder functions.) This v is
the distribution that is referred to in our discussion of Example 1 above. It
is often called the upper invariant measure, since it is the largest possible
stationary distribution.
Of course, it is possible that v — So- In fact, that is what happens for
small values of A, as we will see below in Theorem 3. If that is the case,
another application of coupling gives the following limit theorem:
14 T.M. Liggett

as t f oo for all initial distributions /i. To see this, use (4) to write
£o < /i < Si =4- 6Q = $oS(t) < fiS(t) < 6iS(t),
and pass to the limit as t f oo.
We would like to know that the set of A's for which v — So is an interval, so
that we can define the critical value Ac as the right endpoint of that interval.
But this again an easy consequence of coupling. To see this, write v\ for
the limit in (5). We would like to know that v\ is increasing in A. So, take
Ai < A2, and couple copies r]t and Ct of the processes with these parameter
values respectively, and initial distribution S\ so that r]t < Ct for all t > 0.
(This is possible because the processes are attractive, and the transition
rates are monotone functions of A.) It follows that the distributions of the
two processes at time t are ordered, and hence so are their limits as t f 00:

As a final application of coupling to the contact process, we will prove

the following result. Note that it gives the lower bound for \\{d) for the
process on T^ that was stated in Example 1 above. For the process on Zd,
it gives a slightly weaker lower bound for \(d).
Theorem 3. Suppose the degree of every vertex of the graph S is at most
d. Then the critical value Ac satisfies
1
Ac>-.
d'

Proof. Let r]t be the contact process on S, and let Ct be the branching
random walk on S that is defined as follows. The value Ct(x) is a nonnegative
integer that represents the number of particles at x at time t. The possible
transitions are
C(x) -> C(x) - 1 at rate C(%),
and
C(x) -> C(x) + 1 at rate

In other words, each particle dies at rate 1, and for neighboring vertices x
and y, each particle at y gives birth to a new particle at x at rate A. Now
couple these processes with initial configurations 770 = 1 and Co = 1) so that
Vt(x) < Ct(x) for allt > 0 and x £ 5 . Then
 Interacting Particle Systems - An Introduction 15

which, as we will see, implies that E(t(x) < e^" 1 )* for all x G S and
t > 0. (Note that this is true with equality on a homogeneous graph of
constant degree d, since then f(t) — E(t(x) is independent of a; and satisfies
/'(*) = (\d- l)/(t).) So, if Ad < 1, it follows that E(t(x), and hence Erjt(x)
tends to 0 as t f oo.
To check E(t(x) < e ^ " 1 ^ , one can proceed as follows. Rewrite the
expression for the derivative above as

Integrate this from 0 to t, and then take suprema appropriately to get

rt
M{t) < e~l + Xd f
Jo
where
M ( t ) = sup E(t(x).
X

1
We need to show that M(t) < e^" )* for all t > 0. Elementary estimates
imply that M(t) is bounded on bounded t sets. Suppose that C > 1 and
T > 0 satisfy
M(t) < Ce^'1^
for 0 < t < T. (By local boundedness, such a C exists for every choice of
T.) Using the above integral inequality, it follows that
M(t) < c V A d - ^
for 0 < t < T, with C = C - (C - l)e~ A d r . Iterating this argument gives
M(t) < e^- 1 )*
for 0 < t < T as required.
1.3 Duality
Suppose H(r], () is a nonnegative continuous function of two variables. The
Markov processes f)t and (t are said to be dual with respect to H if

t) (6)
for all 77, ( and all t > 0. For reasonable choices of if, this relation means
that probabilities related to one of the processes can be expressed in terms
16 T.M. Liggett

of probabilities related to the other. Subtracting H(r], () from both sides

of (6), dividing by t and letting t J. 0, one formally obtains the following
relation between the generators £l\ and Q2 of the processes r]t and (t-

«!#(., C)fa) = fi2Jf fa,-HO- (7)

Under weak conditions, which are satisfied by all the examples we will con-
sider, one can prove that (6) and (7) are equivalent.
1.3a. Duality for Markov chains. To illustrate the use of duality, sup-
pose that f)t is a birth and death chain on {0,1,2,...} that is irreducible,
except for the fact that 0 is a trap. In other words, r]t has transitions

n —> n + 1 at rate f3(n)

and
n —> n — 1 at rate 6(n)
for n > 1, where /3(n) > 0 and 6(n) > 0 for all n > 1. Letting H(rj,() =
' w e c a n c o m P u t e the left side of (7) as follows:

= m [H(V, c - 1 ) - ^ fa, 0 ] + <^(c+1) [H(V, c+1) - H(V, 0]

= fi2fffa,.)(C), (8)

where Q2 is the generator of the birth and death process with transitions

n —> n — 1 at rate f3(n)

and
n —> n + 1 at rate 5(n + l).
Note that this chain is irreducible, and the roles of the /3's and £'s have been
reversed. Writing (6) as

and passing to the limit as t f 00, we see that f)t has positive probability of
escaping to 00 if and only if (t is positive recurrent, and in this case,

is absorbed at 0) =
 Interacting Particle Systems - An Introduction 17

where n is the stationary distribution of the chain (f

1.3b. Duality for the contact process. The computation in (8) can
be repeated in other contexts. The trick is to find an appropriate duality
function H. One should look for relatively simple ones, since otherwise it is
difficult to compute with them.
For example, consider the contact process r\t on the graph S, and let £1
be its generator. For 77 G {0,1} S and a finite subset A of S, let

1 J1 if 77 = 0 on A,
I0 otherwise.

Then

{ if (77, A\{x})
-if(77, A)

0
if x G A and 77(3:) = 1,
iixcA and rj(x) = 0,

so
[H(r,x,A)-H(r,,A)]

+A Y [H(r}x,A)-H(r],A)]
t)(a!)=O,t)(»)=l

H(n,A\{x})-\
xeA,r)(x)=l

xEA

Note that the right side of (9) is simply the generator of the process with
transitions
A -> A\{x} at rate 1 if x G A,
(10)
A -> A U {y} at rate \\{x G A : x ~ y}| if y

applied to the function H(r], •). But this process, At, can be thought of as
the contact process itself on configurations with finitely many infected sites,
18 T.M. Liggett

with the identification A — {x : r](x) — 1}. Thus the contact process is

self-dual, and (6) becomes
P11 (rjt = 0 on A) ^ PA (rj = 0 on At). (11)
One might think that is is not useful, since it simply relates the contact
process to itself. However, (11) has important implications. For example,
take 77 = 1 in (11), and pass to the limit as t f oo to obtain

U{TI : ri = 0 on A} = lim PA(At = 0) = PA(At = 0 eventually),

t—>00

where v is the upper invariant measure. In particular, v — 5o if and only

if the "finite" contact process At is absorbed at the empty set a.s. for any
initial state A. This property might appear to be obvious, but it does not
hold for some relatives of the contact process, such as critical, reversible
nearest particle systems. (See Chapter VII of Liggett (1985).)
1.3c. Duality for the linear voter model. This time, for 77 € {0,1} S
and finite A C S, let

. . fl if 77 = I o n A . .
(x) = { ' . (12)
IU otherwise,

let fi be the generator of the linear voter model, and write

p(x,y)[H(Vx,A)-H(r,,A)]
p(x,y)H(r,,A\{x})[l-2r,(x)]{r,(x)[l-r,(y)]+r,(y)[l-r,(x)]}

p(x,y)[H(r,,(A\{x})U{y})-H(r,,A)].
xeA,yes
The right side above can be seen as the result of applying to H(r],A), as
a function of A, the generator of the process in which points in At move
according to the Markov chain with transition probabilities p(x, y), with the
proviso that a point that moves to an occupied site coalesces with the point
at that site. Thus At is called the coalescing Markov chain process. The
duality relation (6) becomes

= I on A) = PA{r] = I on At). (13)

 Interacting Particle Systems - An Introduction 19

This will play an essential role in our analysis of the linear voter model in
the next chapter.
3d. Duality for the symmetric exclusion process. Again take H(r], A)
as in (12), and let fi be the generator of the exclusion process. Let

{ A

A\{x}U{y}
if x,y € A or x,y ^ A,

iix£A,y£A,
A\{y}U{x} ifyeA,x£A,
and write
QH(;A)(r})= £ ) ri(x)[l-Ti(yMx,v)[H(Tlx,y,A)-H(Ti,A)]
x,yeS

x,yeS
£) p(x,y)[l -r,(y)]H(r,,A)

xeA,y<£A
+\p(y, x) - p(x, y)]H(v, A U {y})}.

In order to get a duality statement, we need to recognize the right side above
as the generator of a Markov process, applied to H(r], A) as a function of
A. This is easy to do if p(x,y) — p(y,x) for all x, y, and in this case, we
conclude that

p(x,y)[H(r,,AXty)-H(r,,A)].

Thus we see that in the symmetric case, the exclusion process is self-dual,
and hence
PV (r]t = 1 on A) = PA(r] = Ion At), (14)
where At is the finite version of the exclusion process. We will use (14)
extensively in Chapter 3. The lack of duality in the asymmetric case makes
the process more challenging to analyze, but we are rewarded with more
interesting and diverse behavior. We will address the asymmetric exclusion
process in Chapters 4 and 5.
20 T.M. Liggett

Considering the nature of the duals in the three cases, we see that for the
symmetric exclusion process, the size of the dual does not change with time.
By contrast, the dual of the linear voter model has a cardinality that can
decrease with time, while in the case of the contact process, the cardinality
of the dual can both increase and decrease. These differences have a large
effect on how the duality relation is exploited.

2 Linear Voter Models

Linear voter models were introduced in Example 2 of Chapter 1. In this
chapter, we will explain how duality can be used to give a complete analysis
of the limiting behavior of this process, mainly following the approach of
Holley and Liggett (1975). In order to simplify the presentation, we will
restrict our discussion to the case in which S — Zd, and p(x, y) — p(0, y — x)
are the transition probabilities for an irreducible random walk on Zd. The
general case is treated in Chapter V of Liggett (1985).
We will discover that the behavior of the process depends heavily on
the recurrence or transience properties of the symmetrized random walk
X(t) —Y(t), where X(t) and Y(t) are independent random walks with unit
exponential jump times and transition probabilities ]?(•,•)• We begin with
an informal description of the way in which this comes about. Suppose
we want to determine the opinion r]t(x) of the voter at a; at a large time
t. His opinion came from some other voter at x\ at some earlier time t\.
Continuing backwards in this way, we find that r]t(x) — r]o(X(t)) for some
random X(t). Note that the process X(t) is simply a random walk with
transition probabilities p(-, •) and initial point -X"(0) = x.
We can do a similar backward construction for the evolution of the opin-
ion r]t(y) for y ^ x, and find that i]t(y) — r]o(Y(t)) for a random walk Y(s)
with y(0) = y. However, X(s) and Y(s) are no longer independent - they
are independent until the first time r at which X(T) — Y(T), but after that
time, they evolve together: Y(s) — X(s) for s > r. (These are the coalescing
random walks that we found in our discussion of the voter model duality in
Chapter 1.) Thus, r]t(x) and r]t(y) can agree for two different reasons: (i)
t > T, or (ii) t < T and r]o(X(t)) — r]o(Y(t)). If for the independent random
walks, X(s) — Y(s) is recurrent, then the coalescing random walks will agree
eventually with probability 1, and hence r]t(x) — r]t(y) with large probability
for large t.
 Interacting Particle Systems - An Introduction 21

Remark. It might appear that this would show that for each x and y,
r]t(x) — r]t(y) from some time on. This is not necessarily the case, since
changing the t in the argument changes the random walks X(s) and Y(s).
An easy example is the nearest neighbor symmetric linear voter model in 1
dimension. If the initial configuration is • • • 1 1 1 0 0 0 • • •, for example,
then the configuration at later times has the same form, with the boundary
between l's and 0's moving like a simple symmetric random walk. Therefore,
every site changes opinion infinitely often.
We will now formalize the above argument using duality directly. Let
At be the dual process of coalescing random walks. It will be convenient to
define
g(A) = PA(\At\ < \A\ for some t > 0).
We will regard g(A) as a measure of how spread apart the points in A are
— g(A) small corresponds to A being spread out.
Lemma 1. (a) IfAcB, then g(A) < g(B).
(b) For any nonempty A,

g(A) <
BCA,\B\=2

Proof. Given A C B, At and Bt with AQ — A and Bo — B can be coupled

together in such a way that At C Bt for all t. To do so, if x €E At use the
same exponential times and the same jump choices for x viewed as a point in
At and for x viewed as a point in Bt. With this coupling \At\ < \AQ\ implies
\Bt\ < \BQ\, and hence g(A) < g(B). This proves part (a).
For part (b), let X\(t),...,XiAi(t) be independent random walks with the
distribution of X(t), and initial states {Xi(0), . . . ^ ^ ( O ) } = A. Then

g(A) = P(Xi(t) = Xj(t) for some i ^ j , t > 0)

P(Xi(t) = Xj(t) for some t > 0)

Exercise. Write down the transition rates for the Markov chain (At, Bt) on
{(A, B) : A C B} used in the proof of part (a) of Lemma 1.
22 T.M. Liggett

2.1 The recurrent case

In this section, we assume that for the independent random walks X(t) and
Y(t), X(t) — Y(t) is recurrent. As we will see, it follows that the process
tends to a consensus in this case.
Lemma 2. For every finite, nonempty A C Zd,
PA(\At\ = 1 eventually) = 1. (1)

Proof. The recurrence assumption implies that g(A) — 1 if \A\ — 2. In

particular, (1) holds for all A of size 2. We will now argue by induction on
the size of A. By Lemma l(a), g(A) — 1 for all A with at least two points.
Now take A with \A\ > 1, and let r = inf{t > 0 : \At\ < \A\}, which is then
finite a.s. By the strong Markov property,
PA(\At\ = 1 eventually) = EAPAr(\At\ = 1 eventually).
Since \AT\ — \A\ — 1, if (1) holds for all A of size n, it holds for all A of size
n + 1 (with n > 2).
Theorem 3. (a) For every 77 € {0,1} 5 and every x,y € S,

(c) If fj,{r] : r](x) — 1} — X for all x € S, then

lim nS(t) = X61 + (1 - X)6Q.

t—too

Proof. For part (a), let X(t) and Y(t) be coalescing random walks starting at
x and y respectively, and let r be the coalescing time. By duality (equation
(13) of Chapter 1),

P^imix) + m(y)) = P(v(x(t)) + rj(Y(t))) < P(T >t)^o,

since r < 00 a.s.
For part (b), note that / / £ l implies that

=J
 Interacting Particle Systems - An Introduction 23

which tends to 0 as t —> oo by part (a). Therefore /i concentrates on the two

constant configurations 77 = 0 and 77 = 1.
For part (c), use duality again to write for finite A ^ 0

fiS(t){r] :r] = lonA}= f P^rjt = 1 on A)n(drj) = f PA(r] = 1 on At)n(drj)

= f PA{r, = 1 on AU \At\ > l))fi(dr,) + '£/PA(At = MM*/ • M = 1}-

J
y

Therefore,
\nS(t){v : V = 1 on A} - \\ < 2PA(\At\ > 1),
which tends to 0 as t -> 00 by Lemma 2.
Exercise. Is it the case that the weak limit of /J,S(t) exists for every initial
distribution /i?
2.2 The transient case
In this section, we assume that for the independent random walks X(t) and
Y(t), X(t) — Y(t) is transient. In this case, the behavior of the process f)t is
quite different. We begin by showing how to construct a nontrivial stationary
distribution fj,p for each density 0 < p < 1. Let vp be the homogeneous
product measure on {0,1} S with density p:
vp\r\ : 77 = 1 on A} — p\A\
for each finite A C S.
By duality,

upS{t){q : 77 = 1 on A} = f Pv(r]t = 1 on A)dvp = f PA(r] = 1 on At)dup

(x)dvp = EAp^. (2)

J
B xeB
Since \At\ is nonincreasing in t, it follows that the limit of upS(t){r] : 77 =
1 on A} exists for every finite A C S, and hence the weak limit
(3)
exists. Since fip is the limit of the distribution of the process at time t as
t f 00, it is automatically stationary, which we see as follows:

f fdfipS(t) = I S{t)fdnp= Urn I S(t)fdupS(s)

= ^lirn^ / fdi>pS{t + s)= I fdfip.

24 T.M. Liggett

Before going further, we need to develop some properties of the function

g. Recall that g(A + x) = g{A) for all x € 5.
Lemma 4. (a^ lima;_HX)<7({0,;E}) = 0.
^ If Xi(t), ...,Xn(t) are independent copies of the random walk X(t),
then
-+ 0o.«.
as i —> oo.
(c^ g>(-At) —> 0 a.s. /or any initial set A^ty.
Proof. For part (a), let Z(t) — X(t) — Y(t), which is a symmetric random
walk. Using this symmetry, the Markov property, and the Schwarz inequal-
ity, write
P*(Z(2t)=y) =

(4)

= [Px(Z(2t) = x)Py(Z(2t) = y)]1/2 = P°(Z(2t) = 0).

Now let POO

G(x,y)=/ Px(Z(t) =
Jo
be the Green function for the random walk Z(t). By the strong Markov
property applied at the hitting time of y,
G(x,y) = Px(Z(t) = y for some t > 0)G(y,y).
Similarly, for T > 0,

Px{Z(t) = O)dt = Px(Z(t) = 0 for some t > T)G(0,0),

/IT
JT
which by (4), takes its largest value at x — 0. Therefore,
g({0, x}) = Px(Z(t) = 0 for some t > 0)
< Px(Z(t) = 0 for some t<T) + Px(Z(t) = 0 for some t > T)
< Px(Z(t) = 0 for some t <T) + P°(Z(t) = 0 for some t > T).
Since Z(t) is transient, the second term on the right above can be made
small by taking T large. Then the first term on the right can be made small
by taking \x\ large.
 Interacting Particle Systems - An Introduction 25

For part (b), use part (a) and Lemma l(b), together with the fact
that Xi(t) — Xj(t) —> oo a.s. as t —> oo for i ^ j by the transience as-
sumption. Part (c) follows from part (b) and Lemma l(a), since At and
(Xi(t),..., Xn{t)) can be coupled together in such a way that the contain-
ment
^ C { X i (*),...,*„(*)}
is maintained.
We are now in a position to develop some properties of the stationary
distributions [ip we have constructed. For the statement of part (b) below,
recall that a shift invariant probability measure on {0, \}z is said to be
ergodic if it assigns probability 0 or 1 to every shift invariant subset of
{0,1} Z . These turn out to be the extremal shift invariant distributions.
Theorem 5. For each 0 < p < 1, the measure [ip defined in (3) has the
following properties:
(a) The coordinates are asymptotically independent in the sense that

for all A,
(b) Up is translation invariant and spatially ergodic,
(c) np{n : r](x) = 1} = p,
and
(d)

where G is the Green function of Z(t).

Proof. Passing to the limit in (2), we see that

Hp{n : r? = 1 on A} = Ep\A°° L (5)

(Of course, this is an abuse of notation since the limit of At does not exist;
by \Aool, we mean the limit of \At\, which does exist by monotonicity.)
Therefore, part (a) follows from

\EAp\A~\ - p\A\\ = EA(p\A°°\ - pW, 1^1 < \A\) < g{A).

The translation invariance statement in part (b) is immediate, since vpS{t)

is translation invariant for every t. For the ergodicity statement, let At, A\
26 T.M. Liggett

and A2 be copies of the coalescing random walks process that are coupled
so that
(a) A\ and A2 are independent,
and
(b) At = A\ U A\ for t < r = inf{s > 0 : A] n A2 ^ 0}.
Given disjoint sets .A1 and .A2, start the coupled chains with initial states
A\ = A1, Al = A2, and Ao = A1 U A2. Since |4x>| = | ^ | + 1^1 on the
event {r = oo}, (5) implies that

\[ip{rj : i) = 1 on A1 U A2} - nP{n • f] = 1 on Al}np{n : f] = 1 on A2}\

\\ <P(r <oo)<

Replacing A2 by A2 + z, it then follows from Lemma 4(a) that

lim fJ,p{r] : 77 = 1 on yl 1 U(yl 2 +2;)} = ^p{f] '• f] = 1 on Al}[/,p{r] : 77 = 1 on yl 2 },

and this implies that [ip is spatially ergodic.

Part (c) is an immediate consequence of (5). Part (d) also follows from
(5) as follows:

e?(o,o)

Exercise. Write down the transition rates for the Markov chain (A], A2, At)
used in the proof of part (b) of Theorem 5.

Next we check that we have captured all the extremal stationary distri-
butions via the construction in Theorem 5.

Theorem 6. (a) I is the closed convex hull of {/ip : 0 < p < 1}.
(b) le = {HP : 0 < p < 1}.
 Interacting Particle Systems - An Introduction 27

Proof. One containment is clear, since ^ e l and 1 is convex and closed.

For the other, take / i G l . Then by duality,

H{q •• V = 1 on A} = iiS(t){q : r/ = 1 on A} = /p"(r/t E I on A)dfi

= f PA(r] = 1 on At)dn = £ ) P A ( A t = B)n{n • V = 1 on B}. (6)

J
B

To simplify the notation, let h(A) — fi{r] : 77 = 1 on A},

Vtf(A) = EAf(At), and Utf(xu...,xn) = Ex^~'x" f {X^t), ...,Xn(t)),

where Xi(t),...,Xn(t) are independent versions of our basic random walk

Then (6) can be written as h — Vth. When applied to a function of a set A
instead of a vector (xi, ...,xn), Utf(A) is interpreted as EXl>->Xnf({Xi(t),...,
x
n(t)}), where A - {xi,...,xn}.
Using the coupling between At and {Xi(t), ...,Xn(t)} that we used in the
proof of Lemma 4(b), we see that for any function / satisfying |/(^4)| < 1
for all A,
\Vtf(A)-Utf(A)\<g(A). (7)
Applying this to h gives

\h(A)-Uth(A)\<g(A). (8)

Therefore,
\U,h(A) -Ut+.h(A)\ <U,g(A).
The right side tends to zero as s —> 00 by Lemma 4(b), so lim^oo Ush
exists, and is harmonic for the irreducible random walk (Xi(t),..., Xn(t)) on
Sn. Such harmonic functions are constant (see the second remark following
the proof of Theorem 2 of Chapter 1), so we conclude that there are constants
cn so that
lim Ush(A) — C\A\
for every A. Of course, cn depends on h, and therefore on the stationary
measure /i. Passing to the limit in (8), we conclude that

\h(A)-clAl\<g(A). (9)

Note that by Lemmas l(b) and 4(a), g(A) is small if A is spread out.
Therefore, (9) says that if A is spread out and has cardinality n, then h(A)
28 T.M. Liggett

is approximately equal to cn. We are trying to show that /i is a mixture of

the /ip's. By Theorem 5(a), fip{r] : 77 = 1 on A} and vp\j] : r] = 1 on A} are
approximately equal if yl is spread out. Therefore, we would expect cn to be
a mixture of vp\y\ : r] = 1 on .A} = p" for p € [0,1].
Now we will carry out the program described in the last paragraph. To
do so we need to show that {cn} is a moment sequence, i.e., there exists a
probability measure 7 on [0,1] so that

C = / pn7(dp).
Jo
A necessary and sufficient condition for this is that

0 (10)

for all nonnegative integers n, m. (See Theorem 2 of Section VII.3 of Feller

(1966).) Note that the necessity is clear from

= fpm{l-p)n1{dp).

To check (10), fix m, n, and let A% be a sequence of sets of size m + n so that

<7(.Aj) —> 0. Such a sequence exists by Lemma 4(b). Write A% — Bi U Cj,
where |i?j| = m and |Cj| = n. Note that by Lemma l(a), g(Bi) —>• 0 and
<7(Cj) —> 0 as well, since these are subsets of Ai. By inclusion-exclusion,

: r? = 1 on ^ , r? = 0 on C»} = ^ (-l)l F l/i(Bi U F ) .

Applying (9), it then follows that

lim /i{?7 : rj = 1 on Bu rj = 0 on C»} = V ( l ) ffeec fe+

) ((-l)

which gives (10).

Now define

— and h*(A) — n*{q : 77 = 1 on A}.

Jo
 Interacting Particle Systems - An Introduction 29

Since /i* G 1, (6) can be applied to /i* to conclude that h* — Vth*. By

Theorem 5(a),
\h*(A)-c\A\\<g(A),
and comparing with (9),

\h*(A)-h(A)\<2g(A).

Applying Vt to this and using the harmonicity for Vt of both h and h* gives

\h*(A)-h(A)\<2Vtg(A).

Letting t —> oo and using Lemma 4(c) gives h* = h, and hence /i* = /i. Thus

H = / Vplidp), (11)
./o
which completes the proof of part (a).
To prove part (b), note first that fip G Xe for each p by Theorem 5(b),
since the ergodic measures are extremal in the class of all translation invari-
ant measures. So,
W : 0 < p < 1} C Xe.
For the reverse containment, take | i E l e . By part (a), /i is a mixture of
/ip's. Since /i is extremal, it must be one of the /ip's.
Exercise. Use the fact that (10) is a necessary and sufficient condition for cn
to be a moment sequence to prove de Finetti's Theorem, which states that
every exchangeable measure /i on {0,1} S , where S is infinite, can be written
as a mixture of the homogeneous product measures vp. (See Section VII.4 of
Feller (1966). An exchangeable measure is one for which /i{rj : rj = 1 on A}
depends on A only through its cardinality.) We will use de Finetti's Theorem
in Chapters 3 and 4.
Next is a convergence theorem for translation invariant initial distribu-
tions.
Theorem 7. Suppose /i is translation invariant and spatially ergodic. Then

lim
t—>00

where p — fi{r] : r/(0) = 1 } .

30 T.M. Liggett

Proof. Let h(A) — fi{r] : r/ = 1 on A}, hi(A) — fj,p{r] : r/ = 1 on A}, and

h2(A) = V/>{V : V= Ion A} = p \ A \ .

Define Vt and Ut as in the proof of Theorem 6. By duality, we need to show

that
(12)
for each A. We will prove instead that

Uth(x) ->• pn (13)

for each x G Sn. To see that (13) implies (12), use (7) to write

\Vt+sh(A)-hl(A)\ < \Vs(Vth-Uth)(A)\ + \VsUth(A)-hl(A)\

\VaUth(A)-hi(A)\

which implies by (13) that

limsup \Vth(A) - hi(A)\ < Vsg(A) + \Vsh2(A) - tn(A)\. (14)

t^oo
Since Vsh,2 —> hi as s —> oo by (2) and (3), and Vsg —> 0 by Lemma 4(c), the
right side of (14) tends to zero, and hence (12) holds.
Now we will check (13) using some elementary Fourier analysis. Since
the covariance of the random field with distribution /i is positive definite,
Bochner's theorem implies that there is a measure 7 on [—n, n)d so that

: r,(x) = 1, r,(y) = 1} - p2 = J e

( < • , • > is the usual Rd inner product.) Let

be the characteristic function of the random walk jumps. Since the random
walk is irreducible, \4>{0)\ — 1 if and only if 9 — 0. Note that
Exei<Xt,9> _ ei<x,0>e-t[

Now define
 Interacting Particle Systems - An Introduction 31

Then
2
J[Wt(x,r,)-Wt(x,r,)]2fi(dr,) = 7(d0),

which tends t o zero as s,t —> oo. Therefore

W(x, 77) = lim Wt(x,r])

exists in L2(fi) for each x.

By the definition of Wt(x,r]) and the Chapman-Kolmogorov equations,

= v)Wt(y,T}),

so that by passing to the limit as s —> oo, we have

W(x,ri) = V P * ( X ( * ) = y)W(y,ri) a.s.(/i).

Since all bounded harmonic functions for the random walk are constant, we
conclude that
W(x,rj) = W(0,rj) a.s.(/i) (15)

for each x. Using the definition of Wt(x,r]) again, we have

Wt{x + u,rj) = Wt{x,Turj),

where r u is the shift: Tur](y) — r](y + u). Therefore, W(x + u, rf) — W(x,Tur])
a.s., so by (15), W(0,77) is an a.s. invariant random variable. Since /i is
ergodic, W(0,77) is a.s. constant, and since it has mean p, we conclude that

Wt(x,rj) -^ p

in L2(n) as t —> 00 for every x. But

Uth{{xu...,xn})= f f[Wt{xi,r,)dn ^ (T

as t —> 00. This verifies (13) as required.

32 T.M. Liggett

3 Symmetric Exclusion Processes

Exclusion processes were introduced in Example 3 of Chapter 1. In this
chapter, we will explain how coupling and duality can be used to give a
complete analysis of the limiting behavior of this process in the symmetric
case. The main reason for the symmetry assumption is that there is no
useful duality for asymmetric systems. As in the previous chapter, we will
restrict our discussion to the translation invariant case. Thus we assume that
S — Zd, and p(x,y) — p(0, y — x) — p(y, x) are the transition probabilities for
a symmetric irreducible random walk on Zd. The general case is treated in
Chapter VIII of Liggett (1985) - the main results of that chapter are stated
without proof in Section 3 of this chapter.
The principal result of this chapter is the following.
Theorem 1. (a) Xe — {up : 0 < p < 1}.
(b) If /i is translation and spatially ergodic, then
lim fiS{t) = up,

where p — fi{r] : r/(0) = 1}.

We will only prove part (a), since the proof of part (b) is similar to the
proof of Theorem 7 of Chapter 2. Using duality, part (a) of Theorem 1
follows from the following statement.
Theorem 2. / / / is a bounded harmonic function for the finite exclusion
process At, then f is constant on {A : \A\ — n} for each n > 1.
Proof of Theorem l(a). Recall that a probability measure /i on {0,1} 5 is
said to be exchangeable if
: 77 = 1 on A}
depends on A only through its cardinality. By de Finetti's Theorem (see
the exercise preceding Theorem 7 in Chapter 2), if S is infinite, then every
exchangeable measure is a mixture of the homogeneous product measures
Up. Therefore, Theorem l(a) is equivalent to the statement that I agrees
with the set of exchangeable probability measures. To check this, use duality
(equation (14) of Chapter 1) to write

nS(t){n : T} = 1 on A} = f P^fa = 1 on A)dn = f PA(v = 1 on At)dn

r / = l on B}. (1)
 Interacting Particle Systems - An Introduction 33

Since the cardinality of At does not change with t, this shows that every
exchangeable measure is stationary. Conversely, suppose /i is stationary.
Then fiS(t) — /i for each t, so (1) implies that the function /(-A) = /i{r/ : 77 =
1 on A} is harmonic for At. By Theorem 2, / depends on A only through
its cardinality, so /i is exchangeable.
The proof of Theorem 2 depends very much on whether the random walk
X(t) with transition probabilities p(x,y) is transient or recurrent. We will
explain the proof in these two cases in the next two sections.
3.1 The recurrent case
To prove Theorem 2 in this case, it suffices to construct a coupling of two
copies At and Bt of the finite exclusion process for any initial states AQ, BQ
of cardinality n so that \A0 n Bo\ — n — 1 with the following property:

P(At — Bt for all t beyond some point) = 1. (2)

(A coupling with property (2) is said to be successful.) To see that this is

sufficient, we argue as in the proof of Theorem 2 of Chapter 1. Suppose that
/ is bounded and harmonic for the finite exclusion process At. Then using
the coupling,

\f(A0)-f(B0)\ = \Ef(At)-Ef(Bt)\ < E\f(At)-f(Bt)\ < 2\\f\\P(At + Bt).

The right side of this expression tends to zero as t —> 00 by (2), and hence
f(Ao) — f(Bo). Since any two sets of the same cardinality have the property
that one can be transformed into the other by successively moving one site
at a time, it follows that f(A) — f(B) for any two sets A, B with the same
cardinality.
Next, we must construct a coupling with property (2). If n — 1, this is
very simple: simply let At and Bt evolve independently until they meet, and
make them evolve together thereafter. To see that they will eventually meet,
write At = {X(t)} and Bt = {Y(t)}. Then Z(t) = X(t) - Y(t) is a random
walk with exponential holding times of rate 2 and transition probabilities
p(x,y), until it first hits 0. (Z(t) moves from u to v if either X(t) has an
increment of v — u or Y(t) has an increment of u — v; the rate of this is
p(0, v — u) +p(0,u — v). Now use symmetry.) Therefore Z(t) is recurrent,
and hence it will hit 0 eventually.
For general n, the proof is similar, but relies on a crucial observation.
After At and Bt agree, they will move together. Until that time, call it r,
34 T.M. Liggett

they will always satisfy | ^ n ^ | = n - 1 . So, for t < r, write At — Ctll{X(t)}

and Bt — CtU {Y(t)}, where Ct — At D B<. Before r, the Markov process
(Ct,X(t),Y(t)) has the following transitions from state (C,x,y). In each
case, u E C and v ^ C U {x, y}. The transitions are then to

{Cu,v,X,y) at rate p{u,v),

(C,v, y) at rate p{x,v),
(C,x, v) at rate p(y,v),
(C,x, x) at rate p(y,x),
(C,y, y) at rate p{x,y),
(Cux,u, y) at rate p{u,x),
{Cu,y,X,u) at rate p(u,y).

Note that the marginal processes have the following rates: The pair (X(t), Y(t))
has the transition rates for two independent random walks with transition
probabilities p(-, •) and Ct evolves like an exclusion process, until the first
time r that X(t) — Y(t). This again uses the symmetry of ]?(•,•)• To see
this, write a configuration at some time as

* * * * *
Bt: * * * * *
ct X(t) Y(t)

If u € Ct and x — X(t), then X(t) moves from x to u at rate p(u,x),

when it should make this transition at rate p(x,u). But by symmetry, these
are the same. Now by the recurrence assumption, X(t) and Y(t) will meet
eventually, and hence P(T < oo) = 1, as required.
3.2 The transient case
The proof in case n — 1 is immediate, since irreducible random walks have
no nonconstant bounded harmonic functions. So, fix n > 1, and let

T-{x-(xi,..., xn) € Sn :xi^ Xj for all i ^ j}.

Then we can regard the exclusion process with n particles as a Markov chain
on T. Let Vt be its semigroup, and let Ut be the semigroup for the system
of n independent random walks X(t) — (Xi(t), ...,Xn(t)). We need to prove
that Vtf — f implies / is constant if / is a bounded symmetric function
 Interacting Particle Systems - An Introduction 35

on T. Since Ut is the semigroup for a random walk on Sn, we know that

Utf — f implies / is constant if / is a bounded function on Sn. Thus we
need to compare Ut and Vt in some way. To do so, define g on Sn by

g(x) = Px(X(t) i T for some t > 0).

This function will play a role similar to that of the g in Chapter 2.

By coupling the exclusion process and the independent random walk
system so that they agree until the first time the latter process hits T, we
see that if 0 < / < 1, then

\Vtf(x)-Utf(x)\<g(x), x£T. (3)

By Lemma 4(b) of Chapter 2 and the fact that Px(X(t) e T) ->• 1,

lim Utg(x) = 0, x G Sn. (4)

t-KX>

(Px(X(t) G T) —> 1 is needed here because Lemma 4(b) refers to the set

rather than the vector (Xi(t),...,Xn(t)). Thus, for example, g({u,u}) — 0,

but g(u,u) — 1.)
We will argue in the following way. Suppose / is a symmetric function
on T satisfying 0 < / < 1 and Vtf = f for all t > 0. Extend / to Sn by
setting / = 0 on Sn\T. By (3) and the fact that g = 1 on Sn\T,

\f(x)-Utf(x)\<g(x), zeS". (5)

Applying Us to this inequality gives

\U.f(x) - Us+tf(x)\ < U.g(x), x € Sn.

By (4), Usf has a limit as s —> oo. Since this limit is harmonic for Ut, it is
a constant; call it C. We conclude that

lim Utf(x) = C, xeSn.

Passing to the limit in (5) gives

\f(x)-C\<g(x), xeSn.
36 T.M. Liggett

Applying Vt to this and using Vtf — f again leads to

\f(x)-C\ = \Vtf(x)-C\<Vtg(x), x£T. (6)

To complete the proof that / is constant on T, it then suffices to show that

the right side of (6) tends to zero as t —> oo.
In view of (4), it would be enough to show that

Vtg(x) < Utg(x), xeT. (7)

The intuition behind this inequality is the following: g(x) is small if the
coordinates of x are widely separated. Since the difference between the in-
dependent system and the interacting system is that the latter evolution
keeps the coordinates of the process apart, one might expect that the co-
ordinates are generally further apart in the interacting system than in the
independent system. This relation of being "further apart" cannot be main-
tained in a pathwise sense, but (7) would say that the intuition is correct in
an expected value sense.
It is easier to consider this issue for another function on Sn defined by

where
A(u) = Pu(Z(t) = 0 for some t > 0).
Here Z(t) is the underlying random walk on S. Since

(4) implies that

KmUtg*(x) = 0, xeSn, (8)
t—>00

and we need only prove

Vtg*(x) < Utg*{x), xeT. (9)

We begin the proof of (9) by writing the integration by parts formula

Ut-Vt= f Vt-.(U - V)Usds, (10)

Jo
 Interacting Particle Systems - An Introduction 37

where U and V are the generators of Ut and Vt respectively. (To check this,
integrate the following identity from 0 to t:

^-Vt-sUs = Vt-sUUs - Vt-sVUs,

as
recalling that, while the ?7's and V s do not commute with each other, each
semigroup commutes with its generator.) It follows from (10) that it will be
enough to show that (U — V)Utg* > 0 on T, since Vt maps functions that
are nonnegative on T to functions that are nonnegative on T. Since the two
generators agree except for transitions that lead out of T,

(U -V)Utg*{x) —

-Utg*(x)]
1 "
= g ^2p(xi,Xj)[Utg*(xi,...,Xi-i,Xj,Xi+i,...,xn)

i,...,xn) -2Utg*(x)],

where we have used the symmetry of p(-, •) again. To check that each sum-
maud on the right above is nonnegative, it is therefore enough to check that

Utg*(u,v,x3,...,xn)P(u)P(v) >0 (11)

for all choices of X3,..., xn G S if

The /?(•) that we are using in this application is

{ +1
-1
0
ifu = xi,
ifu = a;2,
otherwise.

To compute the left side of (11), let

At(u) = E(°'u)A(Xi(t) - X2(t)) = Pu(Z(s) = 0 for some s > 2t).

38 T.M. Liggett

Then
Utg*(x)=

so the left side of (11) is

The nonnegativity of this expression follows from the following computation,

for which we take 0 < t\ < • • • < tn:

= X2(U) for some 1 < i < n)

u,v
n
~ ~ X^U) = X2(U), X^tj) + X2(tj) V j > i)

Pu(Xl(ti) = w)Pv(X2(U) = w)

X2(tj - U) V j > i)]

L
j=l w u
Xi(tj - U) # X2{tj - U) V j > i) > 0.

3.3 The general (non—translation invariant) case

We conclude this chapter by stating the results corresponding to Theorems
1 and 2 when S is a general countable set, and p(x, y) are the transition
probabilities for a general symmetric irreducible Markov chain on S. The
proofs can be found in Section VIII. 1 of Liggett (1985).
Define
U = L : S -+ [0,1] : ^2p(x,y)a(y) = a(x)\.
^ y >

For a € 7i, let va be the product measure on {0,1} 5 with marginals given
by a:
va{n • v(x) = 1} = a(x).
 Interacting Particle Systems - An Introduction 39

Finally, let pt(x,y) be the transition probabilities for the continuous time
Markov chain with jump probabilities p(x, y) and exponential holding times
with mean 1.
Theorem 3. (a) For every a € %,

Ha - lim uaS(t)
tHX>

exists.
(b) Halv '• vix) — 1} — &{x) for all x £ 5 , and
= 1} < a(x)a(y)

for all x ^ y €E S.
(c) na is a product measure if and only if a is a constant.
(d) le = {tia:a£ %}.
(e) If the probability measure fi on {0,1} 5 satisfies

lim S^Pt(x, y)^{n • r](y) — 1} — a(x)

for every x £ S , and

lim
2/1,2/2

for every X\,X2 £ S, then a 6 W and

lim

Note that Theorem 1 is a special case of Theorem 3. For part (a), this
uses the fact that random walks on Zd have only constant bounded harmonic
functions, so H consists only of the constants in [0,1] in this case. To deduce
part (b) of Theorem 1 from Theorem 3, use the argument at the end of the
proof of Theorem 7 of Chapter 2.
Here is an example in which H is very large. Let S — T2, the tree in
which each vertex has 3 neighbors. Let p(x, y) — 1/3 if a; and y are neighbors,
and p(x,y) — 0 otherwise. To construct a nonconstant a £ ? f , proceed as
follows: Fix two adjacent vertices a, b, and let a(a) — 1/3 and a(b) — 2/3. If
£ is a distance n from a and n + 1 from 6, let a(x) — (l/3)2~", while if a; is a
40 T.M. Liggett

distance n from b and n+1 from a, let a(x) — 1 —(l/3)2~™. Let T* be the half
of the tree made up of sites that are closer to b than to a. Then the harmonic
function we have defined is a(x) — Px(X(t) €E T* from some time on). By
Theorem 3, there is an extremal stationary distribution for the corresponding
exclusion process with these marginals. It is not known explicitly what this
measure looks like. There are of course many other elements of %.

4 Translation Invariant Exclusion Processes

In the last chapter, we saw how duality can be used to analyze symmetric
exclusion processes. There is no useful duality for asymmetric systems, so
other techniques must be used. In fact, the behavior of asymmetric systems
is more varied and interesting, so it is not just the tools that are different -
the processes themselves are fundamentally different.
In the first two sections, we will let S and p(x,y) be general. However,
the main objective here is to treat translation invariant systems, so that
in the rest of the chapter, we will consider the translation invariant case:
5 = Zd and p(x, y) = p(0, y-x).
4.1 Stationary product measures
In this section, we will see when (possibly inhomogeneous) product measures
are stationary for an exclusion process. If a : S —> [0,1], let va be the product
measure on {0,1} 5 with marginals given by va{j] : rj(x) — 1} — a{x) for each
xeS.
Theorem 1. (a) If p(x,y) is doubly stochastic, then va € X for every
constant a € [0,1]. In particular, this is true in the translation invariant
case.
(b) If n(x) > 0 and n(x)p(x,y) — ir(y)p(y,x) for all x,y € S, (i.e., n is
reversible with respect to p(x, y)) then va € X, where

Remark. If p(x,y) is symmetric, then it satisfies both (a) and (b) (with
IT ^constant). In this case, we checked that the homogeneous product mea-
sures are stationary using duality in Chapter 3.
 Interacting Particle Systems - An Introduction 41

Proof. By Theorem 1 of Chapter 1, it suffices to check that

fnfdva = o
for all cylinder functions / . In the interest of simplicity, we will carry out
the computation in the case in which S is finite, so that all the sums below
are finite. The general case follows by approximation, or by a more careful
handling of the sums involved. (See Theorem 2.1 of Chapter VIII of Liggett
(1985).) Then

P(x,y)[f(r]x,y) - f(ri)]dva
x,y:r)(x)=l,T)(y)=O

,V) I _ _ [f(rix,y)-f(ri)]dva.

It suffices to assume that 0 < a(x) < 1 for all x. Making the change of
variables 77 —> r]x>y in the integral, write

2p(x,y) /
,y Jrr.ri(x)=l,

(2)

In case (a), a(x) is constant, so we get

x,y

But in the doubly stochastic case,

} ^ { i){x)} = 0 .
x,y x,y

(Recall that these sums are finite, since we are taking S finite in this com-
putation.)
Turning to case (b) and interchanging the roles of x and y, (2) becomes

f :fj(»)=O,fj(*)=l
42 T.M. Liggett

Using the reversibility assumption, this becomes

x,y J T]-.T](y)=0,T](x)=l

so that / £lfdva — 0.
Remark. Note that in case (b), va is not only stationary, but is reversible as
well. This means that the stationary process f)t obtained by using va as the
initial distribution (when extended to t € (—00,00)) has the property that
{fhi t € (—00, 00)} and {rj-t, t € (—00, 00)} have the same joint distributions.
Analytically, it means that the semigroup of the process is symmetric on

J fS(t)gdva = JgS(t)fdua
for f,g€. L2(va)-
4.2 Coupling
The basic coupling of two copies of the exclusion process is the process (ijt, Ct)
with the following transitions at rate p(x,y):

(»7>C) -> (Vx,y,(x,y) Krj(x) = ((x) = 1 and rj(y) = ((y) = 0,

(V, 0 -+ (Vx,y, 0 if V(x) = 1, V(y) = 0 and (((x) = 0 or ((y) = 1)
(V, 0 -+ (V, Cx,y) if C(x) = 1, C(V) = 0 and (r](x) = 0 or r/(y) = 1).

In other words, particles move together whenever they can. A discrepancy

is a site x at which r](x) ^ ((x). An important property of the coupling is
that, while discrepancies can move and disappear, they cannot be created.
Another important property is that 770 < Co implies r]t < Ct for all t > 0,
which shows that the process is attractive. An immediate consequence of
this is that every extremal stationary distribution for the coupled process
assigns probability zero or one to the sets {77 < (}, {77 = £} and {£ < 77}.
Let 1* be the set of stationary distributions for the coupled process.
More generally, * will denote objects related to the coupled process. The
next result provides some connections between 1* and 1. By the translation
invariant case, we will mean the case in which S — Zd and p(x, y) — p(y — x)
for some probability density p(x) on Zd. In this case, we will let S and
S* denote the class of shift invariant probability measure on {0,1} 5 and
{0,1} S x {0,1} S respectively.
 Interacting Particle Systems - An Introduction 43

Theorem 2. (a) If v* € X* then its marginals are in I.

(b) If v\,V2 £ X, then there is a v* El* with marginals v\ and V2 •
(c) If v\, V2 Gl e , then the v* in part (b) can be taken to be in I*.
(d) In parts (b) and (c), if v\ < V2, then v* can be taken to concentrate
on {77 < C}-
(e) In the translation invariant case, parts (a)-(d) hold if I and I* are
replaced by IDS and I* D S* respectively.
Proof. Part (a) follows from the fact that the marginal processes r]t and (t
are versions of the exclusion process. For part (b), let /i* be the product
measure v\ x V2. Then v* can be taken to be the limit of a weakly convergent
sequence
1 Ctn
— I n*S{s)ds
*n JO
of Cesaro averages, with tn f 00. Such a limit exists by compactness of the
state space, and is in I*. Turning to part (c), take the v* obtained in part
(b), and write it as a mixture of elements of 1*. The marginals v\ and V2
are the corresponding mixtures of the marginals of these elements of 1*, so
by the extremality assumption on v\ and V2, almost every such element has
marginals v\ and v-i. Take any one of them to be the required element of 1*.
For part (d), instead of taking /i* to be the product measure viX v2, take it
to be any measure with marginals v\ and V2 that concentrates on {77 < C}-
For the last statement, carry out the above proofs with the replacements
given.
Remark. The extension of (d) to the translation invariant setting is some-
what more subtle than the others. If v\ < V2 and both Vi are shift invari-
ant, one needs to find a coupling measure for them that concentrates on
1(^)0 : V < C}> a n d is also shift invariant. To do so, take any coupling
measure v for v\ and V2 that concentrates on {(77, C) : f] < C} ~ it n e e d not
be shift invariant. Then take a limit of Cesaro averages of shifts of v as
the coupling measure to be used in the proof. To check that this limit is
shift invariant, one needs to use the fact that the boundary of a cube of side
length N in Zd has a size that is of smaller order than the volume of the
cube. Thus this proof would not work if Zd is replaced by the homogeneous
tree T^, for example.
4.3 Shift invariant stationary measures
In this section, we consider the irreducible translation invariant case, and
explain how to use coupling to show that the stationary distributions that are
44 T.M. Liggett

shift invariant are exactly the exchangeable measures. Since all exchangeable
measures are mixtures of homogeneous product measures by de Finetti's
Theorem (see the exercise preceding Theorem 7 in Chapter 2), the fact that
every exchangeable measure is stationary follows from Theorem l(a). We
begin with a basic fact about the coupled process - in a shift invariant
equilibrium situation, discrepancies of opposite type cannot coexist. This is
a consequence of the fact that discrepancies cannot be created in the coupled
process.
Theorem 3. If v* 6 l * n S * , then

u*{(rj, C) : v(x) = C(V) = 0, v(y) = C(x) = 1} = 0

for every x,y €E S

Proof. For any probability measure v* on {0,1} S x {0,1} S ,

= ((x) =

,C) : r,(x) = C(x) =

,C) : r,(x) ± CWMv) = CM = 0}

= CM = 1}
y

> , y) + p(y, x)]v*{(V, Q : v(x) = C(y) + v(y) = C(x)}. (3)

The first two terms on the right of (3) correspond to a discrepancy moving
from y to x, the third and fourth terms correspond to a discrepancy moving
from x to y, and the final term corresponds to the destruction of the discrep-
ancy at x. If v* 6 l * , then the left side of (3) is zero. If v* € S*, then the
^-probabilities appearing on the right are functions of y — x. Therefore in
this case, the first and fourth terms cancel, and the second and third terms
cancel. This implies that

v*{(ri, 0 • r](x) = C(V) + V(V) = C(x)} = 0 (4)

 Interacting Particle Systems - An Introduction 45

for all x, y such that p(x, y) +p(y, x) > 0. Using the irreducibility of p(x, y)
and the stationarity of v* again, it then follows that (4) holds for all x,y as
required. To check this, one shows by induction on n that if
x — £0,2:1, ...,xn — y satisfy p(xi,Xi+i) > 0 for each i, (5)
then (4) holds for that pair x, y. We have proved the basis step n — 1
already. For the induction step, assume this is true for n — 1 (and all x, y
which can be joined by a path of length n — 1 as in (5)), and take x, y joined
by a path of length n as in (5). Use the notation:

= C(v) = 0, V(v) = C(u) = 1},

for example. Then write

1 0 0
0 0 1
X Xl y
1 0 0
0 1 1
X Xl y
The last two terms on the right are zero by the induction hypothesis.
We will check that the second term is zero; the first term is treated in a
similar manner. Suppose that at time zero, the process is in the situation
1 0 0
0 0 1
X Xl y
and that by time t, there has been an attempted transition from x — XQ to
xi, but no other attempted transitions involving XQ, ...,xn. Then at time t,
the situation will be
0 1 0
0 0 1.
x xi y

>v*
46 T.M. Liggett

Since v* is stationary for the coupled process, the left side above is zero by
the induction hypothesis, and therefore the right side is also zero as required.

Theorem 4. (X n S)e — {vp, 0 < p < 1}.

Proof. By Theorem l(a), vp € X, and of course vp € S. Furthermore,
Up € <Se, since it is spatially ergodic. Therefore, ^ 6 ( I n <S)e.
For the converse, take v € (Xn«S)e. By Theorem 2(e), for any 0 < p < 1,
there is a v* € (X* D <S*)e with marginals fp and f. By Theorem 3,

"*{(ri,Q 'V<C,V^C} + "*{(ri,Q • C < W + C} + ^*{(»7,C) : »7 = C} = 1-

Since the three sets above are closed for the evolution and translation in-
variant, and since v* is extremal, it follows that one of the three sets has full
measure. Therefore, for every 0 < p < 1, either v < vp or vp < v. It follows
that v — VpQ where po is determined by

v < up for p > pQ,

v > vp for p < po-

4.4 Some extensions

The technique used in the proof of Theorem 4 can be used in some other
situations. Here is an example.

Theorem 5. Ifp(x,y) —p(0,y — x) is irreducible on Zl,

^ |a;|p(0, x) < oo, and ^ xp(O, x) — 0,

X X

then all stationary distributions are exchangeable.

The proof of this result can be found in Chapter VIII of Liggett (1985).
Here we describe only the idea of the proof. The main point is again to prove
an analogue of Theorem 3, without assuming that v* is shift invariant. Once
that is done, one can proceed as in the proof of Theorem 4. Instead of con-
sidering the rate of change of the probability of having a discrepancy at x
in (3) (which is negative only because of the cancellation that occurs in the
translation invariant case), we now need to look at the rate of change of the
expected number of discrepancies in a large interval [—N, N]. This quan-
tity can change in both directions - discrepancies can enter [—N, N] from
 Interacting Particle Systems - An Introduction 47

outside the interval, leading to an increase in discrepancies in the interval.

This problem is handled by showing that for large N, the rate at which dis-
crepancies cross N or — N is small. When looked at near ±00, the coupling
measure v* is almost translation invariant (this requires a Cesaro argument).
But Theorem 4 tells us that the translation invariant stationary distributions
for r]t are exchangeable. Suppose, for example, that v* were shift invariant
and had marginals v\ and vp respectively, with \ < p. Then by Theorem 3,
one might expect that v* is the product measure with marginals

It would follow that the net rate of flow of discrepancies to the right across
N, for example, is
P(x, vV{(ri, 0 : V(x) = 0, ((x) = 1, r}(y) = C(y) = 0}
x<N<y

+ Y P(V> sK{fa> 0 : V(x) = 0, C(x) = 1, r,(y) = ((y) = 1}

x<N<y

- E P(y> x>*{(v, 0 • v(x) = C(x) = 0, ri(v) = 0, C(v) = 1}

x<N<y
- E p(x, y)u*{(V, C) : r](x) = C(x) = 1, v(y) = 0, C(y) = 1}
x<N<y

which vanishes by the mean zero assumption. This is not quite the way the
proof goes, but this computation does show how the mean zero assumption
arises.
Remark. We are very far from understanding the structure of Ie in the
general translation invariant case on Zd. In fact, for d > 2, Ie has been
completely described only in the symmetric case covered in Chapter 3. Cer-
tainly it is often not the case that
Ie = {vP,0<p< 1}.
For example, suppose there is a nonzero v € Zd so that
< >
x ) (6)
48 T.M. Liggett

for all x. Then n(x) — ce<x'v> is reversible for p(x,y), so Theorem l(b)
provides examples on stationary distributions that are not shift invariant.
An instructive example is the following. Let {ei,..., e^} be the standard
basis vectors in Zd, and take

p(x, x + e^ — an, p(x, x — e^ — fa

with all other p(x, y)'s zero. Assume that all the <Vs and /%'s are strictly
positive, and set
, OLi
t>j = l O g — .
Pi
Then (6) holds with v — (vi,..., vj). The corresponding stationary distribu-
tion v satisfies
^ (7)

Note that this is constant on hyperplanes that are orthogonal to v, while one
might have expected it to be constant on hyperplanes that are orthogonal
to the the mean vector

It would be very interesting to know that all the extremal stationary distri-
butions in this case are the homogeneous product measures and the product
measures with marginals given by (7).
The picture is much more complete in the one dimensional case, however.
That is the topic of the next chapter.

5 One Dimensional Exclusion Processes

In this chapter, we will focus on translation invariant exclusion processes on
S — Z1. The first section deals with the nearest neighbor case. The more
general case is described in Section 2. As we will see, the behavior of the
process is much more interesting when the particles have a drift.
5.1 The nearest neighbor case
In this section, we take p(x, x + 1) — p and p(x, x — 1) — q where p + q —
1,0 <p< 1. Letting
PX
n(x) — I - I and a(x) —
 Interacting Particle Systems - An Introduction 49

we see from Theorem l(b) of Chapter 4 that va G X. Note that by the

Borel-Cantelli Lemma, if p > ^, then va assigns full measure to

E=L € {0,1} S : ^ri(x) < oo, £)[1 - ri(x)] < oo j .

Decompose £ in the following way:

1
{ nr\ ^ i-" • \
// C i—i . /
T\\ T\
IfyjUJ
N
/
1
[X
T?I'7'l >
/V /J | *

Then the process 7ft remains in En if it begins in it, so r]t is simply an

irreducible, countable state Markov chain on S n . But the conditional distri-
bution

is stationary for this Markov chain, so the chain is positive recurrent.

Using coupling, much as it was used in Chapter 4, one can prove that
we have now found all the stationary distributions for the process. The
main part of the proof is an analogue of Theorem 3 of Chapter 4. That
theorem says that discrepancies of opposite type cannot coexist in a trans-
lation invariant equilibrium for the coupled process. This is not the case
in the present context, since if one couples the stationary measures v\ and
2

/io, for example, there must be discrepancies of one type on the negative
half line, and of the other type on the positive half line. This is the worst
that can happen, though. It turns out that every element of X* concentrates
on configurations that have all discrepancies of one type to the left of all
discrepancies of the other type. (See Liggett (1976) for details.) Once this
is established, the proof of the following result follows the lines of the proof
of Theorem 4 of Chapter 4.
Theorem 1. If p > \, then Xe — {up, 0 < p < 1} U {/in, -oo < n < oo}.
Thus we see that the asymmetric case allows for the existence of more sta-
tionary distributions than does the symmetric case.
Next, we look at the limiting behavior of the process when the initial
distribution is highly non-translation invariant. For 0 < \,p < 1, let V\JP be
the product measure with

(A ifz<0
:
vxAv V{x) = 1} = <
[p if x > 0.
50 T.M. Liggett

By Theorem 3(e) of Chapter 3, if p — ^ then

By contrast, in the asymmetric case, we have the following quite different

statement. (See Liggett (1977) and Andjel, Bramson and Liggett (1988) for
the proof.)
Theorem 2. Ifp>^, then

v\ji if A > 1/2 and p < 1/2,

up if p> 1/2 and A + p > 1,
lim v\,pS(t) = vx if A< 1/2 and A + p < 1,
if 0 < A < p and A + p = :
if A = 0 and p—1.

To help understand why this might be limit in the various cases, we

describe a heuristic argument. Take as an example the case p—1, and write
/itfor/iS(t). Then

jtVt{n • V{x) = 1} = IH{T} •• v(x-i) = l,f}(x) = 0}-IH{T} : f}(x) = l,r]{x+l) = 0}.

Letting u(x,t) — utiv '• v(x) — 1} a n d pretending that [it is a product

measure (which it is not), this can be seen as a discrete approximation to
Burgers' partial differential equation

Consider this differential equation with initial condition

(A ifx<0,
u(x,0) = \
[p if x > 0.
Then the (entropic) solution is given by

ifz<(l-2A)t,
if (1 - 2A)t < x < (1 - 2p)t,
ifx>(l-2p)t,
 Interacting Particle Systems - An Introduction 51

provided that A > p, and

u(x, t) — u(x — vt, 0), where v — 1 —A — p

if A < p. Therefore, except in the case A + p = l , A < p , the limit in Theorem

2 is the product measure with density given by

lim u(x, t)

(which is independent of x).

5.2 The general case
In this section, we assume that S — Z1, p(x, y) — p(0, y — x), Y^x \X\P(9I x) <
oo and the corresponding random walk is irreducible. In this case, we know
that Xe = {vp,0 < p < 1} if Y,XXP(°IX) = ° (Theorem 5 of Chapter 4),
and that the analogues of the first three cases of Theorem 2 above hold if
Y^x xp(0, x) > 0 (Theorem 1.1 of Liggett (1977)). For the rest of this section,
we focus on analogues of Theorem 1 above. Without loss of generality, we
may assume from here on that Ylx xPi^ix) ^ 0-
A probability measure on S will be called a blocking measure, while a
probability measure /i on {0,1} 5 that satisfies

lim n{r] : r](x) — 1} = 0 and lim fi{r] : r](x) — 1} = 1

X—>•—OO X—>-00

will be called a profile measure. Every blocking measure is a profile measure,

but not conversely. The following results were proved recently in Bramson,
Liggett and Mountford (2002). That paper was an outgrowth of two other
recent papers: Bramson and Mountford (2002) and Ferrari, Lebowitz and
Speer (2001).
T h e o r e m 3. Either (a) Te — {up, 0 < p < 1}, or
(b) there is a profile invariant measure /i so that

%e — {vp, 0 < p < 1} U {fj,n, —oo < n < oo},

where /in is the shift of /i by a distance n.

We know that both situations can occur: (a) happens when p(0, •) has
mean zero (Theorem 5 of Chapter 4), and (b) happens in the nearest neighbor
case with positive drift (Theorem 1 in this chapter). Thus we would like to
know exactly when each of (a) and (b) occurs. The next two theorems
provide some answers to this question.
52 T.M. Liggett

Theorem 4. / / Ylx<o x2p(0,x) — °°> then no stationary blocking measures

exist.
The idea behind this result is that if the negative tails of p(0, •) are very
large, and if there were a blocking equilibrium (in which case all sites to the
right of some point would be occupied), then there would have to be infinitely
many particles to the left of the origin. Note that Theorem 4 does not rule
out the existence of a stationary profile measure. It is not known whether
nonblocking stationary profile measures exist. The following complements
Theorem 4.
Theorem 5. Suppose (a) p(0, x) andp(0, —x) are decreasing in x for x > 1,
(b) p(0,x) > p(0,—x) for all x > 1, with a strict inequality for some
x>l,
and
(c) EaKO z2P(0, x) <oo.
Then a stationary blocking measure exists.
The proof of this theorem is too long to explain here. However, we will
describe the three main steps in the proof, including a computation that
shows explicitly how the second moment assumption enters into the proof.
Outline of part 1 of the proof. Here the objective is to show that stationary
blocking measures exist when the negative tails of p are small. Suppose, for
example, that
Y,\x\5p(0,x) <oo,
x<0
and choose e > 0 so small that

(l + e\x\)4p(0,x) <
x<0 x>0

This is possible by the fifth moment assumption on the negative tails, and
the fact the drift is strictly positive. Define the product measure va as in
Section 3.3, with a(x) chosen to be

J 1 , „ !,I_I^ if x < 0
a(x) —

Note that va is a blocking measure. However, there is no reason for it to be

stationary. However, because of assumption (a) and the choice of e, one can
 Interacting Particle Systems - An Introduction 53

use a nonstandard coupling argument to show that

x<u
is nonincreasing in t for every u. The limiting measure as t —> oo is then a
stationary blocking measure.
An important step in checking this is the following symmetry: For 77 €
{ 0 , l } z \ define 77* €E {0,1 } z ' by
rf{x) = l-rj(-x).
Then 77^ also evolves like the exclusion process. To see this, note that if a
particle in rjt moves from it to w, then a nonparticle (a site at which r]t(x) — 0)
moves from v to u. Since va is invariant under the transformation 77 —>• 77*,
so is vaS(t).
Outline of part 2 of the proof. To remove the moment assumption that
we imposed in part 1, we will need to take the p that we are interested
in, whose negative tails have only a second moment, and truncate it to
obtain an approximating sequence pk whose negative tails have a finite fifth
moment. We will need to have some sort of a priori bound that will allow
us to maintain control of the corresponding sequence of stationary blocking
measures for pk as k —> 00. To do so, suppose v is a stationary measure that
satisfies
{q : T](x) = 1} < 00, ^ v{q : r]{x) = 0} < 00, (1)
x<0 x>0

(which implies it is a blocking measure) and Ylx<o x2p{^ix) < °°- While
(1) is stronger than being blocking, note that the stationary measures con-
structed in part 1 have this stronger property. The result of this part will
be applied to p^ for each k.
Let
00
M(n) — Y^ u{r] : r](x) — l,r](x+ n)— 0}.
x=—00
l
for n € Z . For n — 1, this is the expected number of 10's in the configuration
77. Note that M(n) < 00 for all n by (1). For n > 1,
M(-ri) - M(n) = ^2 [v{n(x) = 0,77(3: + n) = 1} - v{q(x) = 1,77(3: + n) = 0}]
X

{x) = 1}] = n,
54 T.M. Liggett

since the latter series telescopes. Therefore,

M(-n) = M(n)+n. (2)

Next, note that for a stationary measure, the net rate at which particles
go from the left of x to the right of x is independent of x. If the stationary
measure is also a profile measure, this rate must be zero (since it tends to
zero at ±00). Therefore, for fixed x,

p(y,u)v{r](u) = 0,rj(y) = 1}.

u<x<v u<x<v

Summing over all x yields

00

)M(n) = ^ n p ( 0 , - n ) M ( - n ) .
n=l

Using (2) in this expression gives

00

n2p(0, -n) = £ ) nM{n)[p(0, n) - p(0, - n ) ] . (3)

By assumption (b) of Theorem 5, the coefficients of M(n) on the right

are nonnegative, and one of them is strictly positive. Therefore, the sec-
ond moment of the negative tail of p controls the size of some M(n). For
simplicity, suppose that it is M(l).
Outline of part S of the proof. Now take p satisfying the assumptions of the
theorem, and approximate it by a sequence pk, each of which satisfies those
assumptions, and in addition has negative tails with a finite fifth moment.
Let Hk be the corresponding stationary blocking measures that concentrate
on So- Their existence is guaranteed by part 1 of the proof. By part 2 of
the proof, we may assume that

:r](x) = l,r/(x + l) = 0} < 00. (4)

Informally speaking, there cannot be many 10's in /i&. The stationary block-
ing measure /i we are trying to produce will be a limit point of /i& as k —> 00.
It is easy to check that such a limit point is stationary (Proposition 2.14 of
Chapter I of Liggett (1985).) The problem is to show that it is a blocking
measure. In particular, we need to rule out the possibility that /i = 5
 Interacting Particle Systems - An Introduction 55

So, suppose that /i& converges weakly to ^6Q + ^6\. Then for large k,
relative to /i&, there will be a large number of l's to the left of the origin
with probability close to \. Since /i& concentrates on So, there must also
be a large number of O's to the right of the origin. These O's may occur in
a large clump, or in many small clumps. The latter possibility is ruled out
by (4). So, there must be something close to a long interval of l's followed
by a long interval of O's in the configuration. By Theorem 2, the limiting
distribution of the exclusion process beginning with the configuration

•••1 1 1 1 0 0 0 0 •••

is v\/2- (Recall that the first three parts of Theorem 2 have been proved in
the nonnearest neighbor context.) But /i& is stationary, and fi/2 has many
10's, so this again contradicts (4). Thus n ^ \b~o + \b~\.

Acknowledgments
(a) The author would like to thank J. Steif for his very helpful corrections
and comments on the first draft of these notes.
(b) Preparation of these notes was supported in part by NSF Grant
DMS-00-70465.
56 T.M. Liggett

References
[1] E.D. Andjel, M. D. Bramson, and T. M. Liggett, Shocks in the asym-
metric exclusion process, Probab. Th. Rel. Fields 78 (1988), 231-247.

[2] M.D. Bramson and T. Mountford, Blocking measures for one dimen-
sional nonzero mean exclusion processes, Ann. Probab. (2002).

[3] M.D. Bramson, T.M. Liggett, and T. Mountford, Characterization

of stationary measures for one dimensional exclusion processes, Ann.
Probab. (2002).

[4] W. Feller, An Introduction to Probability and its Applications, Vol. II,

Wiley, New York, 1966.

[5] P.A. Ferrari, J.L. Lebowitz, and E. Speer, Blocking measures for asym-
metric exclusion processes via coupling, Bernoulli (2001).

[6] R. Holley and T.M. Liggett, Ergodic theorems for weakly interacting
infinite systems and the voter model, Ann. Probab. 3 (1975), 643-663.

[7] T.M. Liggett, Coupling the simple exclusion process, Ann. Probab. 4
(1976), 339-356.

[8] T.M. Liggett, Ergodic theorems for the asymmetric simple exclusion
process II, Ann. Probab. 5 (1977), 795-801.

[9] T.M. Liggett, Interacting Particle Systems, Springer, New York, 1985.

[10] T.M. Liggett, Stochastic Interacting Systems: Contact, Voter and Ex-
clusion Processes, Springer, Berlin, 1999.