Professional Documents
Culture Documents
on Zd
Author(s): Richard Arratia
Source: The Annals of Probability , Dec., 1981, Vol. 9, No. 6 (Dec., 1981), pp. 909-936
Published by: Institute of Mathematical Statistics
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
BY RICHARD ARRATIA
See Griffeath (1979) for an exposition of this and other basic results about interacting
particle systems.
Since St Ed o it is natural to consider spatial rescalings
Pt aett,
Thus, for any compact convex B C Rd having Lebesgue measure m(B) > 0,
on Rd having an atom of mass one at the rescaled location of each particle. Statement (2)
shows that the family { Vt, t - 0) is tight, so every sequence Vt, with ti -X o must
subsequence which converges to a limiting point process It on R dIn general, for arbitrar
p, we can show that any limit ,u is a simple point process (i.e. /A has no multiple po
having Lebesgue measure m as its intensity. What limits ,u are possible?
This question can be answered completely when p is simple random walk on Zd: for
each d = 1, 2, * , there is a point process jid on Rd such that
Theorem 1 states that for d>- 2, the limit above exists and is the basic Poisson point
process. For d = 1, the limiting point process pu on the line is not Poisson (Arratia 1979);
formula (20) given in Section 2 specifies i,1 in terms of its zero function. The proof of
Theorem 1 depends upon knowing the asymptotic behavior of Pt = P(0 E Ct), obtained for
simple random walks on Zd, d>- 2, in Bramson and Griffeath (1980b):
Pt (Vt)-112 d = 1
where Yd is the probability that a d-dimensional simple random walk never returns to its
origin. The other key ingredient for Theorem 1 is a negative correlation result, that
which we prove for arbitrary random walk p, using Harris's correlation inequality (Harris
1977).
What relation is there between the systems t, coalescing random walks, and qt,
annihilating random walks? There is a coupling such that, for all t>- 0, for every w,
(5) qt C Ot.
Recall that we start with all sites occupied: qjo = 40 = Zd. The coupling is easy to construct
directly: in each system, particles undergo the same random walks, and when two particles
collide, one of them disappears in it, while both of them disappear in qt. It may be bra
to suggest that since twice as many particles vanish per collision in Tt compared with it,
then the ratio of the density of particles in the two systems should go to one half as t goes
to infinity:
(6) P(0 E 7t)/P( E it) = P(0 E IOtl? E it) '/2 ast-3 mo.
Indeed, by the standard duality of coalescing and annihilating random walks with the finite
voter model t starting with a single individual at x, (6) is equivalent to
p a non-nearest neighbor walk on the integers with finite expectation, relation (7) remains
unproved.
For a large class of random walks p, Theorem 3 reduces the problem of finding the
asymptotic density of particles in the annihilating system qt to a problem abou
survival probability for the finite voter model tt. Partial results describing P(t
appear in Sudbury (1976), Kelly (1977), and Sawyer (1979). For simple random walk on
Zd, d 2 2, asymptotics forpt = P(O E At) = P(t2 # p) were finally established by Bramson
and Griffeath (1980b). Thus, for qit = ,td, the system of annihilating simple random walks
starting from all sites occupied,
log tl(2irt) d = 2
1/ (2 Ydt 0 d 2 3.
The case with d = 1 above is in Griffeath (1979); the cases with d 2 2 are an immediate
consequence of Theorem 3 and the asymptotics (4) found by Bramson and Griffeath.
The one half density relation suggests that for large t, rjt may be approximately a "one
half thinning" of At, i.e. a subset of At obtained by tossing a fair coin for each particle in At
to decide whether to retain or delete that particle. Theorem 4 gives a precise version of
this: if p is a multidimensional random walk on Zd, or a nearest neighbor random walk on
Z1, then for any compact K C Rd, with Kt given by (1),
where IK denotes the restriction of a measure to K. This implies a weaker result, Cor
4: for multidimensional random walk, or nearest neighbor random walk on the line,
d = 1: (Ot)-1/2 atd 8d
Here 61, is the one half thinning of the point process i,' specified by formula (20). For d
>- 2, the one half thinning O/id of the intensity one Poisson process jd on Rd is the Poisson
point process with intensity one half.
The one half thinnings O/id that are the rescaled limits for rescaled annihilating simple
random walks are examples of a "compound point process." In general, if 8, ,81, ,82, * are
i.i.d., R '-valued random variables which are independent of a point process ,u = E x,,, then
the 13-compound of ji is the random measure E 813,&. In the one half thinning example
above, 13 is the fair coin variable with P(,8 = 1) = P(,8 = 0) = 1/2. For d 2 2, another
example of a 13-compound of the Poisson point process jid on Rd, with 1 exponentially
distributed, arises as the limit of rescalings of a system yt of coalesing simple random walks
on Zd in which mass is preserved. Start with a particle of mass 1 at each site x E Zd.
Whenever two particles in yt collide, they coalesce into a single particle whose ma
sum of the colliding masses. All particles, regardless of their masses, still undergo identical
random walks, independent apart from the coalescing interference. The state space now is
(Z+)Zd, where yt(X) = n means that there is a particle of mass n at site x E Zd, n = 0
(representing no particle present), 1, 2, .... This system arises naturally in studying
coalescing and annihilating random walks; the coupling (5) may be achieved by taking
having atoms of mass Pt, 2pt, ..., carried on the lattice atZd. Extending a moment
calculation by Sawyer (1979), Bramson and Griffeath (1980b) show for simple random
walk on Zd, d>-2, that for any a-0,
Theorem 2 says that in a rescaled limit, these exponential masses are independent of the
particle locations and each other. More precisely, for simple random walk on Zd with d -
2, /It converges in distribution to the fl-compound of the simple, intensity one Poisson point
process, where ,/ is exponential with mean one. In terms of Laplace transforms, Theorem
1 and 2 are equivalent to, and are proved by showing: for any continuous non-negative f
having compact support in Rd,
2.1 Arbitrary Random Walks. We continue with the notation Vt = atft, where Pt
- P(0 E it), at = (pt)l/d, and It = (Zd is a system of coalescing random walks on Zd, startin
with all sites occupied, based on an arbitrary random walk p. This resealing is appropriate
in the sense that the family { Vt, t 2 0) of point processes is tight, with respect to the vague
topology.
[Here are the details: Let Be denote the E-neighborhood of B in the Euclidean metric on
Rd. We have, for any t-0, and B bounded
which is equivalent to tightness for the family {Vt}. See Kallenberg (1975) for a general
reference on random measures.]
Consider a limiting point process ,I on Rd, that is, suppose that Vt, Id M for som
sequence t -- oo. It is easily seen that for any compact, convex B C Rd,
vt(B) are not uniformly integrable, which indicates clustering in It. Another way to detect
clustering in It, relative to the scale of distance at = P(O E t)1/d, would be through the
existence of multiple points-atoms of mass 2, 3, ..., in a limiting point process ,u. Lemma
1 will be the key to showing that no such clustering occurs-the full strength of Lemma 1
is not needed here; an estimate such as: for x # y E Zd
LEMMA 1. For any A C Zd, and for an arbitrary random walk p on Zd, let (A be the
system of coalescing random walks starting with a particle at each site x E A. For any
X # y E Zd,
PROOF. Let (VB, B C Zd) be the family of voter models based on p, all constructed
together via a single random substructure Y. (B9 is a collection of Poisson flows T(x, y) on
[0, oo) having rate p(x, y) = p(O, y - x); the event times of T(x, y) tell the voter at x when
to discard his opinion and adopt the opinion held by the voter at y.) Thus for any B, C C
Zd, for each w, ~tBUC = gt U tc. By the usual duality between coalescing random walks and
the voter model, (12) is equivalent to
with the ordering (B, C) c (B', C') iff B C B' and C C C'. Define the process starting at
(B, C) E E to be
X(BC) = B CF;
this process is monotone and has only jumps up and down. For an initial distribution
having positive correlations, take the deterministic configuration ({x}, Zd _ {y}) which
has zero correlations; thus we are considering the process
The conclusion, that for every t - 0 the distribution of Xt has positive correlations, says
that for any increasing functions f, g on E,
A slight generalization of this is needed for Theorem 2; Lemma 1 is exactly the special
case m = n = 1 of the following lemma.
LEMMA 2. For an arbitrary random walk p on Zd, A C Zd, let yA be the system of
coalescing random walks with mass conserved, starting with a particle of mass one at
eachsitexEA. ThenVtmn>-OxyEZd,
PROOF. In terms of the usual coupling with the family of voter models, (17) is
equivalent to
El(B) = m(B),
i.e. the intensity of any limiting point process ,A is Lebesgue measure m on Rd.
PROOF. With Bt given by (1), vt(B) = I (t n Bt 1. Identify (t with its indicator function
i.e. write (t(x) = 1 if x E it; (t(x) = 0 otherwise. Lemma 1 says that E( t(x)(t(y)) s if x
# y. Thus
m(aB) = 0 implies EI(aB) = 0 and hence 1i(OB) = 0 a.s., B is a ,i-continuity set. Thus as
random variables, vtB ad ,IB, and formula (18) shows that the vtB are uniformly
integrable, so that EiB = lim Evt B = m(B). 0
COROLLARY 2. With the same hypotheses as Corollary 1, any limiting point process
IL is simple (or orderly), i.e.
P(IL({x}) > 1 for some x E Si) = 0.
PROOF. It follows easily from Lemma 1 that for any bounded B C Rd,
Write Sr = {x E Rd: I X I < r). There is a constant c depending only on d, such that fo
r E (0, 1), Si can be covered by nfr s c/m(Sr) translates U1, U2, ***, U, of Sr. Now
P(,I({x}) > 1 for some x E Si) < P(1I(Uj) > 1 for somej, 1 cj c nr)
2.2 Simple Random Walks. For the remainder of this section we will consider only
the case where p is simple random walk on Zd, i.e. p(x, y) = 1/(2d) if I x - y I = 1, p(x, y)
= 0 otherwise. Thus, when a particle moves, it chooses any one of the 2d neighboring sites
in the lattice Zd with equal probability. For the system of coalescing simple random walks,
the asymptotic behavior of pt = P(0 E at) is known in all cases (d = 1 in Bramson and
Griffeath 1980, d>- 2 in Bramson and Griffeath 1980b). The dth root of these asymptotics
is:
at (rt)-1/2 d = 1
at (ydt)-ld d - 3.
P(/Id(B) = 0) = e-m(B)
vt attt ad Id,
PROOF. To show that vt = at6 is approximately Poisson, we will run the system
for a while with the collision mechanism suspended-particles will follow independent
random walks with no interference, and lattice sites may be multiply occupied. For any
t - e (= 2.7 ... ), set
= tl/2+1/d d - 3
and let at be the system of coalescing random walks with the collision mechanism suspended
from time s to time t = s + At. Write tt(x) = 0, 1, 2, ... for the number of particles at
E Zd; for each t there is a coupling such that V w, V X E Zd,
Let
vt at 6 E>xezd Xt(X)8ax.
so that
E(-it(B) l s) = Mt(B).
(To check that this is really a random measure, having Mt(B) < 00 a.s. for compact
convex B, compute
Lf =E exp(-f f(x)v(dx))
Thus v = PM, the mixture of Poisson processes directed by the random measure M.
We want to show that v is the simple Poisson process with intensity m, i.e. that M = m
(a.s.). Take B to be compact and convex, and consider the random variables v(B) and A
- M(B). The distribution of v(B) is a mixture of ordinary Poisson distributions with
parameter A directed by A, so that Ev(B)2 = E(A2) + EA. We know that EA = Ev(B) =
m(B); the bound (18) of Corollary 1, together with vt(B) Ed v(B) yields
Ev(B)2 ' lim sup E(vtB)2 < lim sup[(EvtB)2 + E vtB] = m(B)2 + m(B).
Thus E(A2) ' (EA)2, so A = m(B), (a.s.). We have M(B) = m(B) a.s., for each compact
convex B, so M = m (a.s.). Thus v = Pm, the simple Poisson point process on Rd with
intensity one. Since this limit is obtained along a subsequence of an arbitrary convergent
sequence (with t4 - oo) from the tight family {Pt, t ' 0), the theorem is proved. E
Let yt be the system of coalescing random walks on Zd with mass conserved, startin
with a particle of mass one at each lattice site. Write yt(x) = n to indicate that there is a
particle of mass n = 0, 1, 2, * * * at x E Zd; interpret yt(X) = 0 as "no particle present." In
terms of the usual coupling of the family (Q, A C Zd) of coalescing random walks, which
(t = t t,
The particle locations for yt are the same as those for t: for each w,
Thus we rescale both the spatial locations and the masses of the particles; let
(27) vt *d PmXfl,
g(x, a) = af(x).
Now
= Lp (-log L3 of
The expression (28) for the Laplace transform of the limiting random measure can be
simplified; since L,8(t) = (1 + t)-1,
To prove the Poisson convergence (27), take s and t as given by (21). Let ,t* be the
system ,t* run with the collision mechanism suspended from time s to t, writing (Zx, n)
= 0, 1, 2, ... for the number of particles at (x, n), x E Zd, n = 1, 2, * .. The corresponding
point process on Rd x R+ is
t= ,t(x, n)8(atx,ptn).
We need to show that there are couplings of Ct* and ,t* such that
(30) {w: xt*(x n) #4 (t*(x n) for some n} = {(t(x) # (t(x)) U {y}&t(x) - 2).
Compute
Now
so
hBtis 2(posth () h0 as to l.
so that
| Bt I e-aps 8 m(B)ea
=(m xf3)(B x I).
Note that E( t* (B x Re)) = E(i-t(B)) - c < oo for compact B, so that the familes {t*, t
0) and {Mt*, t - 0) of random measures on Rd X R' are tight. Starting with any convergen
sequence of v*, say t* v -d Vie take a subsequence along which Mt* converges, say Mt* d
M*. A Laplace transform calculation similar to (26ff.) shows that v* is the mixture of
Poisson processes on Rd X R' directed by the random measure M*. Lemma 2 is now used
to compare first and second moments of the random variable v* (B x [a, 00)) and show that
M* = m x ,8 almost surely. This establishes (27) and concludes the proof of Theorem 2. 0
3. Annihilating Random Walks. The one half thinning relationship between the
system qt of annihilating random walks and the system it of coalescing random walks
natural when viewed in terms of the dual system, the family of voter models (t, A c Zd).
(See Griffeath (1979) for an exposition of all the material in this paragraph.) The voter
model it is the spin flip system on Zd in which the voter at any site x changes opinion at
a rate equal to the proportion of his neighbors (weighted by p) who hold the opposite
opinion; equivalently, the flip rates are
Identify the state space tO, 1}Zd for this spin system with A= fall subsets of Zd}, and write
Ot for the voter model starting with opinion 1 held by the voters at sites x E A, opinion 0
held everywhere else. When A = {x}, write gt' for the voter model tX} starting with a lone
dissenting opinion at x. Thus {w: rt #4 4 is the event that this dissenting opinion survives
until time t, and I gtx I is the size at time t of the dynasty of converts to that dissenting
opinion. For each t - 0, there is a coupling, based on a random substructure B of event
times, of (t, the system of coalescing random walks starting with all sites occupied, jqt, the
corresponding system of annihilating random walks, and (t, A C Zd), the family of voter
models started at each possible initial configuration, such that for every W,
Thus the one half density result of Theorem 3 is equivalent to a statement about the
parity of the finite voter model:
-- 1/2 as t a) o.
This is the only case of p for which the one half density relation (32) had been established.
Theorem 3 extends this to arbitrary multidimensional random walks p, and to p on the
integers Z having E I x I p(O, x) = mo.
For any random walk p, the size of the finite voter model, conditional on survival, tends
in probability to infinity: for any m,
Combining this with (34) yields, for these random walks, that the border size of the voter
model tO, conditional on survival, tends in probability to infinity: for any ro,
FIRST PROOF OF THEOREM 3. We want to show that (32) holds. GivenE eE (0, 1), set
(37) ro = 8E-Iog(1/E)
(39) At = E3/16.
For any c - 1,
(40) c P(t? has at least Ec/2 jumps during [t, t + At] I r(gt) = c)
c P(X ? Ec/2), where X is a mean At(l + E)C Poisson random variable
E(X2) [At(l + E)C][At(l + E)C + 1] 16At(At + 1/(2c) 16At
-< ~~~ ~ ~ ~ = E.
(Ec/2)2 (Ec/2)2 E 2 E2
For a continuous time Markov chain Y whose states can be partitioned into "odd" and
"even" states which alternate with each other, and with all jump rates lying in the interval
[(1 - E)C, (1 + E)C] it is easily proved that
[Here is a proof of (41). Let f(t) = P(Yt is odd), and write [a, b] for the interval of rates.
Then
f(t) a (1 -e-(a+b)t
from which (41) follows.] Using ro and At as specified by (37) and (39), we get, for any c
> ro,
Thus
and this together with (38) yields, for any t > to,
This establishes (32) and concludes the first proof of Theorem 3. The second proof also
starts from statement (35); the heart of this second proof is presented as Lemma 6.
A natural generalization of the one half density relation (32) between annihilating and
coalescing random walks is that such a relation should hold independently at each of n
sites. Let A = {X1, X2, * * , x4 C Zd, IA I = n; it should be the case that for each B C A,
P(ll:-mVxEAIlt#t)VxEA)--*as t-*oo.
Intuitively, the worst case is when the set A is a cluster; a proof along the same lines as the
proof of Lemma 3 requires a lower bound on the probability of the conditioning event of
the form: for every a > 0,
A different way of generalizing the one half density result (32) involves the notion of a
one half thinning Ot of It: independently toss a fair coin for each particle in It to decide
whether or not to retain that point. Equivalently, if v112 is a random subset of Zd whose
distribution is product measure with density one half, and V1/2 independent of It, then
t = It n v112,
or any random set with this distribution, will be called a one half thinning of It. With the
usual notion of convergence in distribution for random elements of {tO 1) Zd, the statement
that Tt gets close to a one half thinning of It as t gets large is: for any finite K C Zd,
-*0 if A=+,
Then at f and atq' converge to the simple Poisson point processes on Rd having intensiti
respectively 1 and 1/2, but for any version of 0ft, for any compact convex K C Rd having
Lebesgue measure m (K) > 0,
LEMMA 3. For the voter model based on an arbitrary nontrivial random walk p, for
every m
PROOF. The jump rate for At, conditional on Ut # p, is at least c = 2(1 - p (0, 0)) > 0.
By comparison to a process which always jumps at rate c, then exists At such that
p(tA has at least m - 1 jumps during [0, At], or t' = 4) > 1/2,
for all finite A C Zd. With probability 2-m1-) the first m - 1 jumps are down, so
By (46),
so that
With this notation, formula (33) for the jump rates of the voter model is
Bi :A S n aA
x -* (x + kxi, x + (k + 1)xi),
where k is the maximal interger such that x + kx, E A. The map x -3 (Bi(x), B2(x)) from
A into (S n aA)2 is 1 -1 since each of the pairs Bi determines a line in Rd, and these two
lines intersect exactly at x. Thus
PROOF. We may assume without loss of generality thatp is symmetric and irreducible.
Thus there is a set K = {x1, X2, * **, Xk) whose elements are positive integers with greate
common divisor 1, such that
Let 1= max(x1, ** , Xk). Since 1 = a1x1 + * * * + akxk for some integer coeffi
it follows for any integers a < b that there is a path
>X=exp(0, x) 2 r.
If, on the other hand, Ai x = 4 for some i, x with 1{ x ' m, 0 ' i < x, then choose any
integer j E [0, 3x) such that I Aj, 3x I I A I /(3x). The assumption Ai x = 4 guarantees, for
every n such that a = 3nx + j E A, that b = 3nx + i X A. A path of the form (50) leads to
an edge en = (cA, dn) CE A having p(cn, dn) 2 A, and
This last relation is needed in order to conclude that for different n, the edges en are really
distinct. Now if I A I > 3xr/A, then
for any random walk p on Zd, for any finite A C Zd, rp (A) 2 r implies
PROOF. We use a randomization-take a flow of jump times having double the usual
rate, and then select with a fair coin which times will be used to run the process. Such a
device appears in Griffeath (1979) to prove that "cancellative systems having pure births
are exponentially ergodic." This randomization can be viewed as a modification of the
usual coupling proof (run two copies of the process independently, starting one of them in
equilibrium, until they reach the diagonal, then run both together along the diagonal) that
a finite state irreducible Markov process converges to its equilibrium distribution. In the
context here, our proof is a modification of the coupling proof that a two state (even, odd)
Markov chain which flips back and forth between even and odd at rate r each way, starting
from a pure state Seven or 8odd can be coupled to the chain in equilibrium (1/2) (8even + 8odd)
by time A t with probability exactly 1 - (1/2)exp(-2rAt).
Construct the voter models DB for every B c Z d simultaneously from a substructure BA
consisting of independent exponential alarm clocks T(Xy) which ring at rates p (x, y). When
the clock for the pair (x, y) rings, the voter at site x is influenced by the voter at site y: the
configuration either flips at x or else remains unchanged, so that its components at x and
y agree.
Write a = (x, y) for a pair of sites in Z d, and let Ta = T(x y) be the random set of times
at which the clock for (x, y) rings, with
To say that Ta is a Poisson flow or Poisson point process on [0, mo) with rate or intensity
p (x, y) is to say that ta(1), ta(2) - ta(1), ta(3) - ta(2), . . . are independent and exponentially
distributed with mean 1/p (x, y). Thus, the substructure ?v is a family of independent
random flows Ta of event times, indexed by pairs a = (x, y) E (Zd)2.
Construct 9 as the one half thinning of another substructure 9' having double the
usual rates. That is, let s"' = (T'xy), x, y E Zd), where T'(x, y) = {tt (1) < t' (2) < ...) is
a Poisson flow with intensity 2p(x, y), let W = (c(x,y)(i); x, y E Zd i = 1, 2, ...) be a
collection of fair coin tossing variable (P(ca(i) = 0) = P(ca(i) = 1) = 1/2) independent of
BA, and set
When ca(i) = 0 so that t' (i) is not an event time in the substructure bA, we say "there is a
phantom effect at time t' (i)," and when ca(i) = 1 we call the effect at time t' (i) "real." For
any C C (Zd)2 having rp(C) -(xy)ecp(x, y) < co, write
Tc = UaEc Ta
for the set of all event times of .4' for edges in C; T'c is a Poisson flow with intensity
2rp(C). Let
be the time and place of the first effect in bA' on some edge in OA. Let
G = {w: ca(1) = 0)
be the event, having probability 1/2, that the effect at time T in B4' is a phantom effect.
Thus, for all w,
Although G and H are not independent, there is an event F having P(F) > 1 - E,
Na = Nx U Ny.
Note that rp(Nx) c 2 and rp(NO) c 4. For each a E dA, define the event
F = UaEOA Foa
Here is the argument that P( F) > 1 - 2E. For any a E dA and t E [0, A t],
P (FaI T = t, at = a)
Since bA' has double the usual rates, and rp(dA) 2 r-(4/E)log(1/E), and At
= /2P(FfnH).
The same argument using the complements of G or H establishes the claims of indepen-
dence.
Finally,
= l/2P(FfnHC) + 1/2P(FfnH)
and similarly
THEOREM 3. Let p be any random walk on Z d for which statement (35) about the
voter model holds. In particular, p may be any genuinely multidimensional random walk,
or a random walk on Z having l I x I p (0, x) = co. Then
PROOF. Given E E (0, 1), set r = (4/E)log(1/E) and At = E/8. Lemma 6 guarantees that
for finite A C Zd, rp (A) 2 r implies
Since the flip rates for the voter model at any site are at most 1,
By relation (35), (which follows easily for those particular cases of p mentioned by using
Lemmas 3, 4, and 5,) there exists to such that t >to implies
The same inequality holds for P( I 't+&t I is even, gt+At # 4), so that
Thus P( | ~t is odd I #to 4 >) 1/2 as t a0, which is equivalent to (51) by duality. 1
Now we embark on a proof of Theorem 4. The main part of the argument lies in Lemma
8, which generalizes the technique of Lemma 6.
LEMMA 7. Letp be a genuinely multidimensional random walk on Zd; let r > 0 and
n > 0 be given. There exists m such that for any disjoint finite sets A1, A2, * , An C
I Ai1 I m for i = 1 to n implies the existence of Ci, C2, * * Cn C (Zd)2 satisfying, for
i, j c n,
PROOF. Continue with the notations xl, x2, A, S from the proof of Lemma 4. Let
ml= r ro/ -,
(m1 is the number of pairs from S needed in each Ci to yield a rate ?ro), let
m2 = 6min,
(m2 is the number of pairs from S needed for sets Cl so that subsets Cl of size exactly ml
may be chosen to satisfy the "no interference" condition (52c)), and take
m = (m2n2)2
for the value m whose existence this lemma asserts. Assume now that we are given disjoint
sets Ai C Zd, satisfying m < I Ai I < o for i = 1 to n. Let
Ao = Zd _ Ui=i Ai,
N= {O,1, ,n}.
each Bij has I Bij < M2, since the choice of I as a connected component means (i, j) 0 G.
This partition involves at most n2 sets Bij, giving I dA n s J < n2m2 in contradiction to (5
Thus, 0 is an element of every connected component of G.
Since G is connected, there is a subgraph T C G which is a connected tree on N. Fix
such a T and define a map
s:N- {O)} N
s (i) = the unique vertex: (i, s (i)) is the first edge of a path in T from i to 0.
Let el, e2, * , e, be the standard basis for (Z/2)' and let eo = 0. For i = 1 to n, define
di = ei + e.(i).
To see that {d1, d2, * *, ch) is a basis for (Z/2)', check that each ei is in its linear span.
Just follow the path in Tfrom i to 0, i.e., with h(i) = min{k:sk(i) = 0),
el= =k L0 d-k()
For i = 1 to n, let
Ci = Bi.,(i),
so that the C1 are disjoint subsets of S, and the Ci in place of the Ci satisfy (52b), and i Ci |
2 M2 = 6min. Observe that for any pair a EC S,
LEMMA 8. For any genuinely multidimensional random walk p, E E (0, 1), and
positive interger no there exists m with the following property:
Given disjoint finite subsets Al, A2, * * , An C Zd, n S no, with each I Ai I 2 m, define the
random joint parity vector in (Z/2)':
ri = T'c,
to be the time of the first effect in the double rate substructure 7' to involve a pair in Ci.
Define the random pair
A = (xi, yi) E Ci
to be the pair a such that ri = iX ' (1), i.e., the place at which the effect at ri occurs. Define
the 0, 1-valued variable
ci= cf3,(1)
to be the value of the coin in W which determines whether the effect at time -r is real (ci
= 1) or phantom (cl = 0). The vector of these ci is a random element of (Z/2)n
Note that I B I = 2n since i #j implies ,/i M NqB. Define a modified parity vector Qt*, similar
to Qt, but which "resets" the opinions on B to what they were initially, just before counting
up the ttA, (mod 2):
FaW = {/31= at, i-i ' At, (T' -{rj) n [O, At] = o ,
which says that ri -- At is caused by an effect in 7' at a, and that sites x and y are not
involved in any other effects in 7' during [0, At]. Thus, on Fa there cannot be any
interaction before At between what happens at time Ti (a real or phantom effect, indicated
by cM(O) = 1 or 0) and the rest of the evolution of the family of voter models. Define F,
Taking D to be the n x n matrix over Z/2 whose rows are d1, d2, dn, this may be
rewritten as
= 2-nP(Q t = qj, F)
= 2-nP(F).
- H exp(-2r,(N,4)At)
2 exp(-8nAt) > 1 - E.
For each i,
so that
To show the independence result (55), start with any choices of a, E Ci, and q1, q2 E
(Z/2) . On the event nfl1 Faon c is determined by coin tossing variables ca,(1), i = 1 to n.
The event n Fa, is measurable with respect to 7' (without using any of the information in
W, saying which of the event times in B?' are real and which are phantom). On the event
nfFa,, QIt is measurable with respect to B' and that part of ' which excludes the first
ca, (1) for each pair of ai. Thus, conditional on nFl ga,, Qut and c are independent:
Summing the outer equality over all choices of the ai E C, yields (55) and completes the
proof. ?
The counterpart to Lemma 8, for a nearest neighbor random walk p on the line, is easy
because the jump rate (36) for this voter model is constant. In contrast to Lemma 8, where
At = E/8no, in Lemma 8' below At must be taken very large to accommodate either E small
or no large.
LEMMA 8'. Let p be a nearest neighbor random walk on Z1, and let E> 0 and no be
given. There exist m and At > 0 such that, for any n - no and ao < a, < ... < an having
ai- ai- 2 m for i = 1 to n, the joint voter model parity
Up until time
Q and R can be coupled so that Qt = Rt. Take m so large that P(Dk m) = 4 ) < E/(4no), and
thus P(T c At) < E/4. The distribution PRt-l of Rt converges to rn, having mass 2-n at
each point of (Z/2) n. Take At so large that the total variation distancel PR-1 -O ,l1, which
is the same regardless of which pure state R is started in, is less than E/2. Now
so that
PROOF. We show that (57) holds for each choice of c < Xo, uniformly in arbitrary choices
Kt C Zd satisfying (56); the geometric structure of the Kt does not enter into the argument,
except in the case d = 1 where it is necessary to assume that each Kt is an interval.
For the family of voter models (tA , A C Zd) with the standard additive coupling induced
by a substructure 9:
At = nxe-A aXt
introduce abbreviation for the set of individuals whose initial opinions survive until time
t:
St= {x E Zd: Dx )
and for the set of those whose dynasty of followers at time t has odd cardinality:
(59) S= t, 7t = St
Fix c < Xo and sets Kt c Zd satisfying (56); in d = 1 also require that the Kt be intervals.
Introduce the abbreviation, for finite A C Zd, t > 0
Since each gt (A) is the total variation distance between two probability measures on the
collection of all subsets of A,
O-gt(A) - 2.
(61) P(IKtfnStI>no)<E.
s = t - At.
For n _ no and any disjoint finite Al, A2, * * , An C Zd satisfying I Ai1 2 m [requir
the d = 1 case, that the Ai as well as UAi be intervals] Lemma 8 or Lemma 8' says
Since the part of the substructure 9 up to time s is independent of the part of 9 Rafter s, for
gt(A) <-6E,
E,,k=f{ljxj-kVxEKtnSs, KtnSs=KtnSt},
(66) P(Etc)-C infk:k::m ,P(E t'k) C (C/Pt) infk k::,[P(0 <1| s l< k) +P(| Isl| 2k, t = ?)].
Since the flip rate at any site in the voter model is at most 1,
P(I DSI | 2k, Dt = 4) C P(| I | k) SUPA:JAJ:k P(D 5 = 4)) <Ps SUPA:JAJ~k P(IA = ) ).
Thus, continuing with (66),
P(Etc) < cent infk klm[P( IS I < k I Is 3 4)) + supAjAIsk P(kAt = 4))].
Lemma 3 says that the first term above goes to zero as t -X 00, for any fixed k. To see that
the second term can be made arbitrarily small by choosing k large enough, note that the
embedded Markov chain for MA must pass through a sequence Ak, Aki1, *.-, Al of
configurations having I i = i before reaching 4. The jump rate out of Ai is rp (dAt) ' I) A I
- i. For a pure death process which starts at k and jumps from i to i - 1 at rate i, the time
until extinction has mean Z i=i (1/i) and variance Z 1/i2; the probability that this process,
starting at k, is extinct by time At tends to 0 as k increases. A comparison of MA with this
process shows that
PROOF. When the sum in (57) is c, there exist versions of qt and Edt such that
P(qt n Kt 3# (Jt n Kt) = E.
Let S8 be the sphere of radius n centered at the origin in Rd, and fix an increasing sequence
tn -X00 such that for t 2 tn, with K = Sn,, the sum in (57) is less than 1/n. For t E [t,, tn+1)
select versions of mt and 0-~t such that
at, t, Ed AL,
then
at,7qt, Ed OI,
where the point process 01i is the one half thinning of the point process it on Rd.
PROOF. Write C+(Rd) for the class of all nonnegative continuous functions on Rd
having compact support. The Laplace transform L4 a random measure ju on Rd is defined
by LLf = E (exp(-f f d,)) for f E C+. The one half thinning OI of a random measure is
characterized by
Suppose f E C+ with support contained in a compact set K. Let g = -log(l - 1/2(1 -e-f)),
so that Le, f = LLg. Take versions of -qt and O(t as in Corollary 3. We have
REFERENCES
[1] ARRATIA, R. (1979). Coalescing Brownian Motions on the Line. Ph.D. thesis, University of
Wisconsin at Madison.
[2] BRAMSON, M. and GRIFFEATH, D. (1979). Renormalizing the 3-dimensional voter model. Ann.
Probability 7 418-432.
[3] BRAMSON, M. and GRIFFEATH, D. (1980). Clustering and dispersion rates for some interacting
particle systems on ZV. Ann. Probability 8 183-213.
[4] BRAMSON, M. and GRIFFEATH, D. (1980b). Asymptotics for interacting particle systems on Zd.
Z. Warsch. verw. Gebiete 53 183-196.
[5] GRIFFEATH, D. (1979). Additive and Cancellative Interacting Particle Systems. Lecture Notes
in Math. 724. Springer, New York.
[6] HARRIS, T. (1977). A correlation inequality for Markov processes in partially ordered state
spaces. Ann Probability 5 451-454.
[7] HOLLEY, R. and STROOCK, D. (1979). Central limit phenomena of various interacting systems.
Ann. of Math. 110 333-393.
[8] KALLENBERG, 0. (1975). Random Measures. Akademie-Verlag, Berlin.
[9] KELLY, F. P. (1977). The asymptotic behavior of an invasion process. J. Appl. Probability 14
584-590.
[10] LIGGETT, T. (1978). Random invariant measures for Markov chains, and independent particle
systems. Z. Wahrsch. verw. Gebiete 45 297-313.
[11] SAWYER, S. (1979). A limit theorem for patch sizes in a selectively-neutral migration model. J.
Appl. Probability 16 482-495.
[12] SUDBURY, A. (1976). The size of the region occupied by one type in an invasion process. J. Appl.
Probability 13 355-356.
DEPARTMENT OF MATHEMATICS
UNIVERSITY OF SOUTHERN CALIFORNIA
LOS ANGELES, CALIFORNIA 90007