The Annals of Probability

2002, Vol. 30, No. 1, 1–61

A CONVERGENCE RESULT FOR CRITICAL

REVERSIBLE NEAREST PARTICLE SYSTEMS

B Y T HOMAS M OUNTFORD
University of California, Los Angeles
We consider critical reversible nearest particle systems. We assume
that the associated renewal measure has large moments as well as some
regularity conditions. It is shown that such processes, started from a nontrivial
ergodic translation invariant distribution, converge in distribution to the upper
invariant measure.

0. Introduction. In this paper we consider critical nearest particle systems

and their distributional properties. A nearest particle system (NPS) is a spin system
on {0, 1}Z with flip rates given by


1, if η(x) = 1,
c(x, η) =
f (lx , rx ), if η(x) = 0,
where lx = x − sup{y < x : η(y) = 1} and rx = inf{y > x : η(y) = 1} − x (either or
both possibly ∞).
Of particular interest are the so-called reversible NPSs. These are systems
where f (l, r) is of the form β(l)β(r)/β(l + r), f (l, ∞) = f (∞, l) = β(l),
f (∞, ∞) = 0. If a NPS is reversible in the classical sense, then [see Liggett
(1985)] f (·, ·) must be of this form. These processes are of mathematical interest
because there is an array of reversible Markov chain techniques with which to
analyze them. This paper considers reversible processes; we will also require the
condition
β(n)
(∗) ↓1 as n → ∞.
β(n + 1)
The convergence of the quotient to 1 is equivalent to the process being Feller,
a naturally desirable property. The assumption of monotonic convergence down
to 1 ensures that the process is attractive.
This makes the process much more
mathematically tractable. We also assume β(n) < ∞.
We will assume that the process ηt is generated by a given Harris system (which
will also generate auxiliary comparison processes as well): We suppose that we are
given for each x ∈ Z two Poisson processes Dx and Bx , independent of each other
and independent over Z, with Dx of rate 1 and Bx of rate M = (β(1))2 /β(2),
the largest possible flip rate. The value (or spin) at site x can only change for t
in either Dx or in Bx . The process Dx corresponds to flips of 1’s or deaths of

Received January 2000; revised February 2001.

AMS 2000 subject classifications. 82C22, 60K35.
Key words and phrases. Nearest particle system, critical, reversible.
1
2 T. MOUNTFORD

particles at site x and is simple. If t ∈ Dx then ηt (x) = 0, irrespective of its value

immediately preceding t. The Poisson process Bx corresponds to flips of a 0 to a 1
at site x and associated with each t ∈ Bx is a random variable Ut uniform on [0, 1].
At time t ∈ Bx , we have ηt (x) = 1 either if ηs (x) was equal to 1 immediately
prior

to t or if Ut ≤ c(x, ηt− )/M, for ηt− the stateimmediately before time t.
If x η0 (x) < ∞, then the summability condition, x β(x) < ∞, ensures that,
for all time t, x ηt (x) < ∞ andthat the update times for the process form a.s.
a discrete set. If x>0 η0 (x) = x≤0 η0 (x) = ∞, then a.s. for any T > 0 there
exist xiT → ∞, x−i T → −∞ as i → ∞ so that, for all i, η (x T ) = η (x T ) = 1
0 i 0 −i
and Dx T , Dx T have no points on time interval [0, T ]. It is clear that if we fix all
i −i
sites outside spatial interval [−n, n], while updating sites in (−n, n) according
to the above rules, then we obtain a process ηsn . The existence of the xiT , x−i T ’s,

above, shows that almost surely for every T > 0, N > 0, there exists n0 = n0 (ω)
n
so that ∀n ≥ n0 , ηsn (x) = ηs 0 (x) for |x| ≤ N, 0 ≤ s ≤ T . We can thus define the
infinite

process
 t
η as the limit of the ηtn ’s. The remaining case, where one of
x>0 η0 (x), x≤0 η0 (x) is finite and the other infinite, can be treated in a similar
manner.
The attractiveness condition ensures that if the same Harris system generates
two NPSs ηt and ξt and ξ0 ≤ η0 with respect to the natural partial order, then
ξt ≤ ηt for all times t. It also implies monotonicity relations between η and various
finite comparison Markov chains, as detailed

in Section 2.

We say that 
η ∈ {0, 1} is finite if x η(x) < ∞, that it is semiinfinite if one of
Z

x<0 η(x), x>0 η(x) is infinite but not both and that it is infinite if both these
terms are infinite.

If η0 is infinite, then for all t, ηt must be infinite, while the
finiteness of β(n) ensures (as already noted) that if η0 is finite (semiinfinite),
then it will remain so for all subsequent times. We can thus speak of finite,
semiinfinite and infinite NPSs. 
We say that a NPS is supercritical if β(n) > 1. In this case the finite and
infinite systems survive. In the finite case this means that if τ = inf{t : ηt ≡ 0},
then P {0} [τ = ∞] > 0, where P η denotes the probability for a process starting
from initial configuration η. [See Griffeath and Liggett (1982), Liggett (1985).] For
infinite systems survival means that there exists a nontrivial invariant distribution
for the process, ν̄. In fact we have ν̄ [see Mountford (1997)], so that, for each η0 ,
ηt → P η0 (τ = ∞)ν̄ + P η0 (τ < ∞)δ0
in distribution. In particular all infinite systems must converge in distribution to
the upper invariant distribution. 
We say that a NPS is subcritical if β(n) < 1. In this case finite systems must
die out: P η0 (τ = ∞) = 0 for all finite η0 . Under the additional assumption of
attractiveness, infinite subcritical systems die out in the sense that ηt → δ0 as t
tends to infinity. In fact, under minimal assumptions, Mountford (1995) showed
that for attractive subcritical systems P (ηt (0) = 1) tended to zero exponentially
fast.
 CRITICAL NEAREST PARTICLE SYSTEMS 3

The remaining case, β(n) = 1, is called the critical case and is the subject of
this paper. Griffeath and Liggett (1982) prove that in all cases the finite critical
case

NPS dies out. The infinite process, however, will survive if and only if
nβ(n) < ∞. If this condition is satisfied, then the upper invariant distribution
is the renewal measure on {0, 1}Z corresponding to probability measure on the
integers, β(·). Subsequently we will denote this renewal measure by Ren(β).
We show:

T HEOREM 1. Let ηt be a critical, attractive, Feller, reversible NPS satisfying

(∗∗) below with η0 ergodic, translation invariant and of positive density. Then
ηt → Ren(β),
the upper invariant distribution (the renewal measure corresponding to β).

R EMARK 1. We note that, to show the above result, it will suffice to show
that, for any fixed cylinder, increasing f and any positive ε arbitrarily small, that
E[f (ηt )] will exceed Ren(β), f   − ε for t large, where for any measure ν and
continuous function f, ν, f  = f (ω)ν(dω).

R EMARK 2. Liggett (1985) notes that, since Bernoulli of sufficiently high

density stochastically dominates the associated renewal measure, an attractive,
reversible NPS with such an initial distribution must converge to the upper
invariant distribution.

R EMARK 3. By the ergodic theorem it immediately follows that ηt converges

in distribution to Ren(β) as t becomes large whenever η0 has a translation-
invariant distribution such that P (η0 ≡ 0) = 0.

The condition (∗∗) below is a more extreme form of the assumptions used in
Mountford and Sweet (1998). The high order moment condition is by no means
optimal, but we feel that it is better to give a clear exposition showing the ideas
than to become involved in technical issues which may obscure the central idea.
We assume
 
β(n)
(∗∗) n − 1 → k ≥ 120.
β(n + 1)
If k = ∞ we also assume that, for some α ∈ (0, 1),
   
β(n) β(n)
0 < lim inf nα − 1 ≤ lim sup nα − 1 < ∞.
n→∞ β(n + 1) n→∞ β(n + 1)
These conditions (and in particular those for k = ∞) are far from necessary for
the arguments in this paper to work. We use these simple cases to avoid undue
technicalities in a paper already overburdened with technicalities.
4 T. MOUNTFORD

It is easy to check the following lemma.

L EMMA 0.1. The assumptions (∗∗) (including the extra assumption for the
case k = ∞) imply
 β(l)β(r)
(∗∗∗) sup ≤ V < ∞.
n
l+r=n
β(n)

Observe that if k is finite, then for any δ > 0 we have for large enough x that

β(x)  x −(k−δ) and β̄(x) = β(y)  x −(k−1−δ) .
y≥x

The proofs in this paper assume tacitly that the constant k above is finite. The
statements are all still valid for infinite k, and the proofs are typically easier. We
leave the minimal changes to the reader.
As detailed above, our process (and auxiliary processes) is constructed via a
system of Poisson processes. Repeatedly throughout the paper we will be using
large deviation bounds for Poisson random variables; given this and the elementary
nature of the following lemma we feel it is appropriate to introduce it here.

L EMMA 0.2. For fixed c > 0 and α ∈ (1/2, 1), there exists a constant K so
that, for y ≥ 1, a Poisson random variable Xy of mean y satisfies

P (|Xy − y| ≥ cy α ) < Ke−c

2 y 2α−1 /3
.

The proof follows from Stirling’s formula and, for reasons of space as well as
its elementary nature, is left to the reader.
We now give a brief sketch of the plan of attack.
The basic picture comes from Schinazi (1992), who considered semiinfinite
critical reversible NPSs. A right-sided NPS is one for which, for all time t,
 
x<0 ηt (x) = ∞ and x>0 ηt (x) < ∞. For such processes we denote the position
of the rightmost particle at time t[sup{x : ηt (x) = 1}] by rt (and similarly for
left-sided semiinfinite processes the leftmost particle is denoted by lt ). We say
a semiinfinite process is in equilibrium if, seen from its rightmost particle
(respectively leftmost particle), the configuration is distributed as a renewal
measure to the left (resp. right). It is easy to check that, for the one-sided process
continually shifted so that its rightmost particle occupies the origin, this measure
is invariant. We say a right semiinfinite process ηR is supported by (−∞, x] at
time 0 if its rightmost particle at time 0 is at site x; η0R is in equilibrium supported
on (−∞, x] if it is vacant on interval (x, ∞) and if on (−∞, x] its distribution is
renewal measure conditioned on there being a particle at x.
 CRITICAL NEAREST PARTICLE SYSTEMS 5

Schinazi (1992) showed the following theorem.

T HEOREM 0.1. Let ηR be a semiinfinite process in equilibrium supported by

(−∞, 0] at time 0. Then rN 2 t /N tends in distribution to a nontrivial Brownian
motion as N tends to infinity.

Purely to avoid too much notation we will make the following assumption which
will be in force throughout the sequel:
(∗∗∗∗) the diffusion constant for the limiting Brownian motion is 1.
Theorem 0.1 is the starting point for this paper. It suggests the following picture
of the evolution of an infinite critical NPS: After a short time the NPS ηt can be
seen as a succession of intervals alternately in the upper invariant regime and in
the lower regime (i.e., vacant). The interfaces between these intervals perform like
independent coalescing Brownian motions, each coalescence corresponding to the
disappearance of a particular interval. That is, we might expect the NPS to behave
like a one-dimensional voter model in which intervals of 1’s and 0’s compete,
though in our case the intervals of 1’s would not be intervals of fully occupied
sites but rather intervals in the “upper regime.”
Various problems present themselves: First, the initial distribution could be
far from an alternation of intervals in the upper regime and intervals completely
vacant; second, we have to be more precise about the evolution of finite intervals in
the “upper regime”; and third, this picture cannot give us what we want. The one-
dimensional voter model is symmetric between 0’s and 1’s and if we start with a
nontrivial translation invariant system, the system will converge in distribution to
a nontrivial mixture of all 1’s and all 0’s.
The first problem is not too difficult: First, note that we are dealing with an
attractive system and we are trying to show convergence to the upper invariant
distribution; second, if we run the process for any amount of time, the process will
have the possibility of forming large fully occupied intervals. This follows from
the irreducibility of the finite systems. We can then remove all particles that do not
belong to a fully occupied interval of length N say. In fact we can suppose that
after a fixed time our configuration consists of disjoint occupied intervals of length
precisely N .
The second problem presents the bulk of the work of this paper. Our main tool,
which was not available to previous workers in the area, is the method of estimating
the spectral gap of reversible Markov chains developed by Jerrum and Sinclair
(1990); see Diaconis and Stroock (1991) for a clear exposition. This technique
was used in Mountford and Sweet (1998), where a useful comparison finite state
Markov chain was introduced. This enables us to clumsily reprove the result of
Schinazi (1992) (in a dramatically reduced setting) in a way that can be used for
finite intervals. It also enables us to establish good regeneration properties of finite,
semiinfinite and infinite NPSs.
6 T. MOUNTFORD

The third objection prompts us to add to our picture: In Section 5 we detail how
the attractiveness of the system gives a slight drift outward for the boundaries of
an interval in the “upper regime.” This enables us to “grow” new “upper regime”
intervals and tips the process toward the “upper regime.”
The paper runs as follows. Section 1 establishes some crude bounds for the
deviations of the rightmost particle for one-sided processes starting in equilibrium
and starting with full occupancy to the left of the origin and full vacancy to the
right. These inequalities are built on in Section 2, which gives the first application
of the comparison processes in Mountford and Sweet (1998) and various coupling
arguments to obtain good large deviations estimates for one-sided processes.
Section 3 uses more coupling to give a central limit theorem for the rightmost
particle and to give a measure of closeness to normality for the edge fluctuations
of a semiinfinite process in equilibrium. Section 4 applies results of Section 3 to
finite NPSs and establishes a technical “regeneration” result. It is also shown that
for a sufficiently “rich” finite system the two edge fluctuations can essentially be
thought of as being those of two independent semiinfinite processes. Section 5
gives the key technical result for “growing new upper equilibrium” intervals.
Section 6 introduces an (unfortunately) complicated system of rules for upper
equilibrium intervals; those intervals which fail are to be killed. Good bounds are
given for the various probabilities of an interval being killed. The following section
establishes that these killing rules do not handicap the upper regime intervals too
much. Section 8 shows that upper regime intervals must eventually dominate. The
proof of Theorem 1 is completed in Section 9.
Throughout the paper we adopt the convention that nonspecific constants may
change their value from line to line or even from one side of an inequality to
another so that we may write, for example, for x ≥ 0, C(x 2 + 1)e−cx ≤ Ce−cx .

1. We consider some simple inequalities fora one-sided NPS, ηt with η0 ≡ 0

on (0, ∞). Let rt = sup{x : ηt (x) = 1}. Let µ = n nβ(n).

L EMMA 1.1. For η0 ≡ 0 on (0, ∞), rt satisfies

 
C
P sup rs ≥ 2µt ≤
s≤t t (k−3)/2
for t large.

P ROOF. By attractiveness we need only consider a NPS with all deaths

suppressed and with η0 ≡ 1 on (−∞, 0]. We can w.l.o.g. take t to be equal to n,
an integer. For this modified process it is immediate that E[rt ] = tµ. It is almost
 j
immediate that E[r1k−2 ] < ∞; rn is equal in distribution to nj=1 r1 , the sum of n
independent copies of r1 . We suppose further (taking the worst case) that k − 2 is
an odd integer. Then
 k−3
E rn − E(rn ) ≤ Ck−3 n(k−3)/2 .
 CRITICAL NEAREST PARTICLE SYSTEMS 7

So, by Chebyshev, P (rn ≥ 2µn) is less than or equal to

 k−3
1 C
P (|rn − E[rn ]| ≥ µn) ≤ Ck−3 n(k−3)/2 ≤ .

µn n(k−3)/2

The corresponding inequality for infs≤t rs over configurations with η0 (0) = 1,

is false. However, we have, by reversibility [and attractiveness for part (ii)], the
following lemma.

L EMMA 1.2. (i) For η0 distributed according to renewal measure supported

on (−∞, 0] and null on (0, ∞) we have
 
C
P sup |rs | ≥ 2µt ≤ .
s≤t t (k−3)/2
(ii) Starting from η0 ≡ 1 on (−∞, 0], we have
 
C
P sup |rs | ≥ 2µt ≤ .
s≤t t (k−3)/2

P ROOF. (i) The position of the rightmost particle of ηs is less than the sum
of all jumps to the right of the rightmost particle by time s, which in turn is less
than the sum of all jumps to the right by time t. However, this latter quantity is
treated in Lemma 1.1. We thus have
 
C
P sup rs ≥ 2µt ≤ (k−3)/2 ,
s≤t t
where C is the constant of Lemma 1.1. The time reversal of η as seen from
the rightmost particle is equal in distribution to η. From this we see that, in
equilibrium, the sum of the magnitude of jumps leftward of the rightmost particle
over the time interval [0, t] is equal in distribution to the sum of the magnitudes
of the rightward jumps of the rightmost particle. Therefore as we have that −rs is
dominated by the sum of the magnitudes of leftward jumps in time interval [0, s],
we have
C
P inf rs ≤ −2µt ≤ (k−3)/2 .
s≤t t
The result follows with C = 2C.
(ii) We have already seen in Lemma 1.1 that in this case
 
C
P sup rs ≥ 2µt ≤ .
s≤t t (k−3)/2
On the other hand by attractiveness we have that

P inf rs ≤ −2µt
s≤t
8 T. MOUNTFORD

is less than the corresponding probability for the rightmost particle of a process
started in equilibrium. Thus by (i) above we have
C
P inf rs ≤ −2µt ≤ .

s≤t t (k−3)/2

We now introduce the notion of a gap for a configuration η. This is a vacant

interval. We say that η has a gap of length b in interval I (or originating in I ) if
for some interval J of length b we have η ≡ 0 on J and J intersects I .
A natural question is at what rate will a one-sided process on (−∞, 0] feel
the effects of its infinite tail. We gain some crude bounds on this (which can
later be refined). Consider η1 , the NPS starting fully occupied on (−∞, 0],
9/4
completely vacant on (0, ∞). Let ηn be the (finite) NPS starting with the
occupied sites being precisely those in interval [−n9/4 , 0]. We assume that the
processes are generated by the same Harris system and therefore that, for all
9/4 9/4
times t, ηtn ≤ ηt1 . Let dt = sup{x : ηt1 (x) = ηtn (x)}. Necessarily we must have
9/4
ηt1 (dt ) = 1, ηtn (dt ) = 0.

L EMMA 1.3. P (sups≤n2 ds ≥ −2n9/4/3) ≤ Cn5 /n(k−2)/4 .

P ROOF. We majorize ds by a nondecreasing process Ds , where Ds is a jump

process, initially at −n9/4 − 1, that jumps forward by x units at time t if and only
if:
1 ≡ 0 on (D , D + x];
(i) ηt− t− t−
and also:
(ii) there is an attempted birth at site Dt + x (i.e., if y is the distance to the
right of site x until there is an occupied site in ηt1 , then we have t ∈ BDt +x and
Ut ≤ β(x)β(y)
β(x+y) /M).

By assumption (∗∗∗), we have that Ds jumps at rate bounded by V , uniformly

9/4
over all joint configurations. We easily see that ηsn and ηs1 agree on random
space interval (Ds , ∞) so


sup ds ≥ −2n9/4 /3 ⊂ B(1) ∪ B(2),
s≤n2

where
 
B(1) = ∃t ≤ n2 for which ηt1 has an n1/4 /6V gap in [−n9/4 , −2n9/4 /3] ,
 
B(2) = Ds has ≥ 2V n2 jumps in [0, n2 ] .
This corresponds to the fact that if Dn2 is large, then either D made a large number
of jumps or some of the jumps were large.
 CRITICAL NEAREST PARTICLE SYSTEMS 9

The jumps of Ds occur at a rate always bounded by V [recall our choice of V

in (∗∗∗)]. Accordingly the probability of event B(2) is bounded by the probability
that a Poisson random variable of mean V n2 exceeds 2V n2 . By Lemma 0.2 with
α = 3/4, say, we have P (B(2)) ≤ e−Cn .
To bound P (B(1)) we introduce a comparison one-sided process η, generated
by the same Harris system as η1 but with initial configuration equal to a renewal
process supported on (−∞, 0]. We denote its rightmost particle’s position at time t
by rt . By attractiveness, ηs ≤ ηs1 ∀s. Therefore to bound the probability of B(1) it
is enough to bound the probability of
 
B(1) = ∃t ≤ n2 for which ηt has an n1/4 /6V gap in [−n9/4 , −2n9/4 /3] .
This event is easily seen to be contained in the union
B(1a) ∪ B(1b)
 
= ∃t ≤ n2 for which ηt has an n1/4 /6V gap starting in [rt − 2n9/4, rt ]


∪ sup |rs | ≥ n9/4 /3 .
s≤n2
By Lemma 1.2, the second event, B(1b) above, has probability bounded above
by C/nk−3 . By invariance the process ηt seen from rt is always a one-sided
renewal process in distribution. From elementary theory of i.i.d. random variables,
the probability a renewal sequence with a particle at the origin has an n1/4 /6V gap
in the interval [−2n9/4 , 0] is bounded by Cn3 /n(k−2)/4 . Therefore
 n2 +1
Y= Iηt has an n1/4 gap in [rt −2n9/4 ,rt ] dt
0
has expectation bounded by C(n2 + 1)n3 /n(k−2)/4 . Furthermore should such a gap
appear it will remain in this condition for time 1 with probability at least e−(V +1) .
Thus by Chebyshev we have

P B(1a) ≤ e(1+V ) Cn5 /n(k−2)/4 .
Thus we have
 
  
P sup ds ≥ −2n9/4 /3 ≤ P B(1a) + P B(1b) + P B(2)
s≤n2
≤ Cn5 /n(k−2)/4 .

In the same way we can prove the following lemma.

L EMMA 1.4. Let η, η be two one-sided NPSs supported on (−∞, 0] at time 0.

Suppose also that η is in one-sided equilibrium and that η0 ≡ η0 on [−n9/4 , 0].
Then outside probability Cn5 /n(k−2)/4 ,

ηs ≡ ηs on − 23 n9/4 , ∞ ∀0 ≤ s ≤ n2 .
10 T. MOUNTFORD

2. The basic coupling. In dealing with infinite NPSs we find it useful to

introduce three comparison NPS-like continuous Markov chains (hereafter MC’s)
on {0, 1}[−n,n] . These chains were discussed in Mountford and Sweet (1998) and
spectral bounds calculated for them:
1. For Ztn , 1’s are fixed at −n and n; otherwise the sites flip according to NPS flip
rates; Ztn,1 is Ztn with Z0n ≡ 1 on [−n, n].
2. For Ytn , 1’s are fixed at −n and n, 0’s are fixed on (−n, −n + n1/3 ) and
(n − n1/3 , n) but otherwise sites flip according to NPS rates.
3. Process Xtn is the same as Ytn except that the Markov chain is not allowed to hit
the configuration which is null on (−n, n). We tacitly assume that X0n ≡ 0 on
(−n, n).
If η0 is an infinite NPS and ηt evolves according to a Harris system H , then
we can use the same Harris system to generate versions of the above MC’s on
{0, 1}[−n,n] .
The comparison involving process Z n with η (both using the same Harris
system) is the most straightforward: if
Z0n |(−n,n) ≥ η0 |(−n,n)
(where η|A denotes configuration η restricted to subset A), then from attractiveness
we have immediately that the following holds:
(a) ∀t, ηt |[−n,n] ≤ Ztn |[−n,n] ≤ Ztn,1 |[−n,n] .
The comparison involving Y n is not so straightforward in that for Y n we are not
simply fixing 1’s but both 0’s and 1’s. The large imposed intervals of 0’s for process
Y n “should” result in a process “below” η; however, any domination may break
down when η itself is vacant on interval (−n, −n+n1/3 ) or (n−n1/3 , n). However,
if we define τ = inf{s > 0 : ηs ≡ 0 on either (−n, −n + n1/3 ) or (n − n1/3 , n)} and
if Y0n |(−n+n1/3 ,n−n1/3 ) ≤ η0 |(−n+n1/3 ,n−n1/3 ) , then the following holds:
(b) ∀t ≤ τ, ηt |(−n,n) ≥ Ytn |(−n,n) .
The comparison involving Xn is even more involved in that Xn may jump over η
on (−n, n) by virtue of the fact that deaths of Xn particles may be suppressed. We
define σ = inf{s : Ysn ≡ 0 on (−n, n)}. We have if
Y0n |(−n+n1/3 ,n−n1/3 ) = X0n |(−n+n1/3 ,n−n1/3 ) ≤ η0 |(−n+n1/3 ,n−n1/3 ) ,
then the following holds:
(c) ∀t ≤ τ ∧ σ, ηt |(−n,n) ≥ Xtn |(−n,n) .
Thus Z n always approximates η from above on (−n, n); Y n approximates η
from below until a big gap for η appears; Xn approximates η from below until
either a big η-gap appears or Y n becomes vacant on (−n, n).
 CRITICAL NEAREST PARTICLE SYSTEMS 11

The reason for introducing process Xn as well as Y n is that process Xn (like Z n )

has a spectral gap of size greater than or equal to c/n2 some c > 0, whereas the
gap for Y n is typically much smaller than this [see Mountford and Sweet (1998)
n
for details]. We introduce notation P ξ,V to refer to probabilities for process
Vn (= Xn , Y n or Z n ) starting from configuration ξ . We say a configuration ξ on
{0, 1}[−n,n] is bad if

P ξ ,Y (σ ≤ n8 ) ≥ n−k/6 ,
n

where ξ  is equal to ξ except within n1/3 of the endpoints n, −n where ξ  is vacant.

The following lemmas are easy to establish. See Mountford and Sweet (1998).

L EMMA 2.1. Let π n be the equilibrium probability of Xn and let π n,Y be that
for Y n . Then for ξ ∈ {0, 1}[−n,n] \ {0},
0 ≤ π n (ξ ) − π n,Y (ξ ) ≤ Cπ n (ξ )/nk/3 .

L EMMA 2.2. Let ηt be a right semiinfinite NPS in equilibrium. Let rt be the

rightmost particle. The chance that either:
(i) for t ≤ T there is a gap of size n1/3 within S of rt ;
(ii) for t ≤ T ηt ◦ θ−rt +x |[−n,n] is bad for n ≤ x ≤ S;
is bounded by 3(V + 1)(T + 1)(S + 1)n · n−k/6 .

L EMMA 2.3. Under equilibrium, the probability that Xn has a gap of size n/5
or larger is bounded by C/nk/2 for C not depending on n. Under equilibrium, the
probability that Y n hits 0 in time n8 is bounded by K/nk/3 .

We now get comparisons between the stationary probabilities of π n,1 , the

equilibrium probability for Z n , and π n , the equilibrium probability for Xn . Define
π l,r to be the renewal measure restricted to [l, r] and conditioned to have 1’s at l
and r. (So π n,1 = π −n,n .)

L EMMA 2.4. For x ∈ [−4n/5, 4n/5],

   
0 ≤ π n,1 {ξ : ξ(x) = 1} − π n {ξ : ξ(x) = 1}  ≤ C/nk/3 .

P ROOF. It follows from attractiveness that π n,1 stochastically dominates

π n,Y .So the upper bound for π n ({ξ : ξ(x) = 1}) − π n,1 ({ξ : ξ(x) = 1}) follows
from Lemma 2.1. It also follows from attractiveness that, for −n ≤ l ≤ r ≤ n,
the restriction of π n,1 to [l, r] is stochastically dominated by π l,r . Given η ∈
{0, 1}[−n,n] , define L to equal inf{y > −n : η(y) = 1} and R to be sup{y < n :
12 T. MOUNTFORD

η(y) = 1}. It follows from the Markov property of renewal measures that, for
x ∈ [−4/5n, 4/5n],
   
π n {η : η(x) = 1} ≥ π l,r {η : η(x) = 1} π n {L = l, R = r}
l≤−4/5n,r≥4/5n
 
≥ π n,1 {η : η(x) = 1} − 2π n {L > −4/5n} .
The result now follows from Lemma 2.3.

For the following note that for some ν ∈ (0, ∞) and not depending on n we
have for all n and ξ ∈ {0, 1}[−n,n] \ 0 that π n ({ξ }), π n,1 ({ξ }) ≥ e−νn .

C OROLLARY 2.1. For X0n arbitrary in {0, 1}[−n,n] \ 0, if Xtn and Ztn,1 are
derived from the same Harris system, then, outside probability Cn/nk/3 ,
Znn,1
4 (x) = Xn4 (x)
n
∀x ∈ [−4n/5, 4n/5].

4
P ROOF. First consider the auxiliary process Z n /2 , which is derived from the
same Harris system as the other two processes, whose dynamics are those of Z n,1
but which starts at time n4 /2 at full occupancy. By attractiveness of NPS we have
n4 /2 n4 /2
that, for all t ≥ n4 /2, Zt ≥ Ztn,1 . In particular Zn4 ≥ Znn,1
4 . However, also by
the spectral gaps for these Markov chains we have that
n4 /2  n4 /2
P Zn4 = Znn,1
4 ≤ P Zn4 (x) = Znn,1
4 (x)
x

≤ Kn eνn e−(c/n
2 )n4 /2 1/2

≤ Ke−n .
4
Thus we may compare instead process Z n /2 with Xn . Now by the usual spectral
gap arguments applied to process Xn , we have that the Radon–Nykodym derivative
of the distribution of Xnn4 /2 with respect to the equilibrium distribution of Y n
is bounded by 2 for n large. Thus by Lemma 2.3 we have that the chance that
process Xn at some time in interval [n4 /2, n4 ] has a jump to all 0’s blocked is
n4 /2
bounded by K/nk/3 . So outside this probability we have that Xnn4 ≤ Zn4 . It
n4 /2
remains to bound the probability that Xnn4 < Zn4 .
However,
n4 /2
P Zn4 (x) = 1, Xnn4 (x) = 0
 
 n4 /2  n4 /2
≤ P Zn4 (x) = 1 − P Xnn4 (x) = 1  + P Zn4 (x) = 0, Xnn4 (x) = 1

≤ C  /nk/3 for some C  ,

 CRITICAL NEAREST PARTICLE SYSTEMS 13

by standard spectral theory and Lemma 2.4. The result follows by summing over
the (of order n) x.

From similar (but easier) arguments one has the following corollary:

C OROLLARY 2.2. For Z0n arbitrary in {0, 1}[−n,n] , if Ztn and Ztn,1 are derived
from the same Harris system, then outside probability Ke−n ,
Znn,1
7/2 (x) = Zn7/2 (x)
n
∀x ∈ [−n, n].

C OROLLARY 2.3. For η any infinite or semiinfinite NPS let η1 be a NPS

starting from full occupancy and generated by the same Harris systems as η and
let Y n be the Markov chain described at the start of the section with Y0n agreeing
with η0 on interval (−n + n1/3 , n − n1/3 ). If:
(i) σ = inf{t : ηt ≡ 0 on [−n, −n + n1/3 ] or [n − n1/3 , n]};
(ii) τ = inf{t : Ytn ≡ 0 on (−n, n)};
then
 Cn
P ηn14 (x) = ηn4 (x), ∀x ∈ (−4n/5, 4n/5) ≥ 1 − − P (τ ≤ n4 ) − P (σ ≤ n4 ).
nk/3

The following result is a key to the paper but as important 

are the techniques
employed in the proof which are used in Section 4. Recall µ = n nβ(n).

L EMMA 2.5 (A first coupling). Consider two processes (generated by the

same Harris system):
(a) ηt where η0 is in (semiinfinite) equilibrium, supported on (−∞, 0] and
vacant on (0, ∞);
(b) η1 started from all sites on (−∞, −4µn4/5 ] occupied and vacant on
(−4µn4/5 , ∞).
Then outside of probability (Cn13/4 /nk/30 ), rt ≥ rt1 ∀t ∈ [0, n2 ], where
rt = sup{x : ηt (x) = 1},
rt1 = sup{x : ηt1 (x) = 1}.
Also ηt1 ≤ ηt on [−n9/4 /2, ∞) for n4/5 ≤ t ≤ n2 .

P ROOF. It is clear that we need only prove the inequality for n large.
Lemma 1.2 implies that, outside of probability bounded by Kn2(k−3)/5 , on the
time interval [0, n4/5 ],
rs ≥ −2µn4/5, rs1 ≤ −2µn4/5 .
14 T. MOUNTFORD

This deals with the conclusions of the lemma in the case t ∈ [0, n4/5 ]. We introduce
the process γ , run with the same Harris system, where
γ0 (x) = Ix∈[−n9/4 −4µn4/5 ,−4µn4/5 ] .
Since by Lemma 1.3 we have (outside probability Kn5 /n(k−2)/4 ) that

∀0 ≤ t ≤ n2 , ηt1 ≤ γt on − 23 n9/4 , ∞ ,
our lemma will follow if we can show that (outside of probability Cn13/4 /nk/30 )
γn4/5 ≤ ηn4/5 .
Now note that by Lemma 1.1 (applied to the leftmost particle of γ , l γ , as well as
to the rightmost, r γ ), we have that, outside of probability bounded by Kn2(k−3)/5 ,
for t ∈ [0, n4/5 ], γt is vacant outside interval [−n9/4 − 6µn4/5 , −2µn4/5]. Divide
the interval [−n9/4 − 6µn4/5, −2µn4/5 ] into m (≤Kn9/4−1/5) equal (w.l.o.g.)
intervals of length 2n1/5 + 1, disjoint except at the endpoints, I0 , I1 , . . . , Im .
Now let J0 , J1 , . . . , Jm+1 be disjoint (except at the endpoints) intervals of length
2n1/5 + 1 such that the left endpoint of Ii is the center of Ji and the right endpoint
of Im is the center of Jm+1 .
Let us define Zti to be NPS on {0, 1}Ii with 1’s fixed at the endpoints of Ii and
such that Z0i ≡ 1 on Ii , and define Yti to be the NPS on {0, 1}Ii with 1’s fixed at the
endpoints of Ii , 0’s fixed within n1/15 of the endpoints and (subject to the above)
Y0i ≡ η0 on Ii . Let Xi be the process on {0, 1}Ii which has X0i = Y0i and which
evolves like Y i except that it is forbidden to hit all zeros on the interior of Ii .
Then we have the following:
Outside of probability K(2n9/4 + 1)(n4/5 + 1)n1/5 n−k/30 [by Lemma 2.2(i)],
there is no n1/15 gap of ηt for 0 ≤ t ≤ n4/5 , within 2n9/4 of rt .
Outside of probability Kn−2(k−3)/5 (by Lemma 1.1),
|rt | ≤ 2µn4/5 ∀0 ≤ t ≤ n4/5 ,
γ
rt ≤ −2µn4/5 ∀t ≤ n4/5 ,
γ
lt ≥ −n9/4 − 6µn4/5 ∀t ≤ n4/5 .
Outside of probability K(2n9/4 + 1)n1/5 n−k/30 + Kn9/4−1/5n−k/30 [by Lemma
2.2(ii)],
∀i, ∀0 ≤ t ≤ n4/5 , Yti is not identically zero on the interior of Ii .
Outside of probability n(9/4)−(1/5)n−(k−3)/15 (by Corollary 2.1),
∀i, ∀x in the central 4/5 of Ii , Xni 1/4 (x) = Zni 1/4 (x).
The sum of these probabilities is bounded by Kn13/4 /nk/30 .
For the x’s that are in the central four-fifths of an Ii , we have (outside of the
sum of the above probabilities) that
ηni 4/5 (x) ≥ Yni 4/5 (x) = Xni 4/5 = Zni 4/5 (x) ≥ γn4/5 (x).
 CRITICAL NEAREST PARTICLE SYSTEMS 15

Repeating this argument with the J ’s replacing the I ’s we get, outside of

probability (Cn13/4 /nk/30 ),
ηn4/5 (x) ≥ γn4/5 (x) ∀x.
This completes the proof that (outside of the contracted probability), at time n4/5
and therefore (by attractiveness) for all t ≥ n4/5 ,
γ ≤ η.

In the same way (but using Lemma 1.4 instead of Lemma 1.3) we obtain the
following lemma.

L EMMA 2.6. Consider two semiinfinite right NPSs in equilibrium, η, η , so

that both are generated by same Harris system:
(a) η0 is in (semiinfinite) equilibrium supported on (−∞, 0];
(b) η is supported on (−∞, −4µn4/5 ];
(c) η0 and η0 are independently distributed.
Then, outside of probability (Cn13/4 /nk/30 ), rt ≥ rt ∀t ∈ [0, n2 ], where
rt = sup{x : ηt (x) = 1},
rt = sup{x : ηt (x) = 1}.
Also ηt ≤ ηt on [−n9/4 /2, ∞) for n4/5 ≤ t ≤ n2 .

We will be (two sections hence) getting further couplings but first we use these
lemmas to obtain some bounds on the probabilities of large deviations for rt and
similar processes. For now we give the following corollary.

C OROLLARY 2.4. Let η01 (x) = Ix≤0 . Let rt1 = sup{x : ηt1 (x) = 1}. Then as
t → ∞,

√
P sup rs1 ≥ λ t → 2=(−λ),
s≤t
where = is the distribution function for a standard normal random variable.

P ROOF. We need only treat t of the form n2 . Let rtb be the rightmost particle
of an equilibrium NPS which is initially supported on (−∞, 0] and is run by the
same Harris system as η1 . From attractiveness we have immediately that

√ 
√
P sup rs1 ≥ λ t ≥ P sup rsb ≥ λ t .
s≤t s≤t
From Schinazi (1992) and the reflection principle we have that

√
P sup rsb ≥ λ t → 2=(−λ),
s≤t
as n becomes large.
16 T. MOUNTFORD

It remains to achieve an upper bound. Let rtu be the rightmost particle of an

equilibrium NPS which is initially supported on (−∞, 4µn4/5 ] and is run by the
same Harris system as η1 . Lemma 2.5 yields that

√ 
√
P sup rs1 ≥ λ t ≤ P sup rsb ≥ λ t + Cn13/4 /nk/30 .
s≤t s≤t

However, again via [Schinazi (1992)] and the reflection principle, as n becomes
large the right-hand side converges to 2=(−λ).

3. The purpose of this section is to prove large deviations bounds on the

behaviour of max{rs : 0 ≤ s ≤ 2n } for a right semiinfinite NPS supported in
equilibrium on (−∞, 0], either in equilibrium or starting from full occupancy on
this interval.
Our first tool is:

L EMMA 3.1. Let ηt1 be a semiinfinite NPS, with η01 equal to 1 on (−∞, 0] and
equal to 0 elsewhere. Let its rightmost particle at time t be denoted by rt1 . For n
sufficiently large we have
   n/2
1
P sup rs1 ≥ n2n/2 ≤ C for n large,
s≤2n 2
 
K2n
P sup rs1 ≥ n2 2n/2 /4 ≤ for n large.
s≤2n 2n/2(k−2)

P ROOF. From Corollary 2.4, we have that, for t large enough,


√ 1
P sup rs1 ≥2 t ≤ .
s≤t 16
Now we use a simple Hoffman–Jørgensen √ argument: if we denote by τm the first
time that rs1 is greater than or equal to t(2m + (2m−1 − 1)/t 1/3 ), then for m > 1
we have {τm ≤ t} ⊂ {τm−1 ≤ t} and that either
√
√  m−1 2m−2 − 1  t
rτm−1 ≤ t 2 + 1/3
+ 1/3
t t
or

√
t
∃0 ≤ s ≤ t : rs − rs− ≥ 1/3 .
t
(Here rs− is the limit of r as s is approached from the left.) The probability of
the latter event is bounded by t/(t 1/2 /t 1/3 )k−2 for t large by assumption (∗∗). By
 CRITICAL NEAREST PARTICLE SYSTEMS 17

attractiveness and the strong Markov property applied at time τ1 we have


√ 1
  2
1
 1/2 −(k−2)
t
P sup rs1 ≥ t 4 + 1/3 ≤ + t 1/3
s≤t t 16 t
 2  1/2 −(k−2)
1 t
≤ + 2t
16 t 1/3

√ 3
  2
1 −(k/6−2)
2
⇒P sup rs1 ≥ t 8 + 1/3 ≤ + 2t + t −(k/6−2)
s≤t t 16
 4
1
≤ + 2t −(k/6−2)
16
[where for the last inequality we used (x + y)2 = x 2 + (2x + y)y ≤ x 2 + y/2 if
2x + y ≤ 1/2]. This proves, after iteration of the argument, the inequality

√   2 m−1
2m−1 − 1 1
P sup rs1 ≥ t 2m + 1/3
≤ + 2t −(k/6−2) ;
s≤t t 16
thus (with t = 2n , m = #log2 (n)$ − 1, where # $ represents the integer part of), we
have for n large enough
      n/2
2m−1 − 1 1
P sup rs1 ≥ n2n/2 ≤ P sup rs1 ≥ 2n/2 2m + ≤2
s≤2n s≤2n 2n/3 2
(recall k is large). For the second inequality, we first define
 
for m ≥ 0 σm = inf t : rt1 ≥ n2n/2 3m .
Repeating the Hoffman–Jørgensen argument we have (on the event {σm ≤ 2n }),
either
rσ1m−1 ≤ 2n2n/2 3m−1
or
∃0 ≤ s ≤ 2n : rs − rs− ≥ n2n/2 3m−1 .
Using the strong Markov property at σm−1 as before, we conclude that

P (σm < 2n ) ≤ P (σm−1 < 2n ) + 2n (n2n/2 3m−1 )−(k−2) .
2

Provided that n was fixed sufficiently large we have for all m that

P (σm−1 < 2n ) > (2n 2−n(k−2)/2 )2/3
3/2
⇒ P (σm < 2n ) ≤ P (σm−1 < 2n ) .
18 T. MOUNTFORD

However, it is easy to see that (again provided that n was fixed large) for m =
log2 (n)/2 that
 (3/2)m
P (σ0 < 2n ) ≤ 2n (n2n/2 3m−1 )−(k−2) .
We conclude that for some m ≤ log2 (n)/2 we have
P (σm < 2n ) ≤ (2n 2−n(k−2)/2)2/3 .
We then consider for example σm+1 and conclude
 
n2 2n/2 K2n
P sup rs1 ≥ ≤ n/2(k−2) for K < ∞ not depending on n.

s≤2n 4 2
From this, attractiveness and reversibility we obtain:

P ROPOSITION 3.1. Let η0 be in semiinfinite equilibrium with r0 = 0:

 
 1 n/2
P sup |rs | ≥ 2n2n/2 ≤ C 2 for n large;
s≤2n
 
n2 K2n
P sup |rs | ≥ 2n/2 ≤ n/2(k−2) for n large.
s≤2n 2 2

P ROOF. The inequalities

 
 1 n/2
P sup rs ≥ n2n/2 ≤C 2 for n large,
s≤2n
 
n2 n/2 K2n
P sup rs ≥ 2 ≤ n/2(k−2) for n large
s≤2n 4 2
follow directly from Lemma 3.1 and attractiveness. By reversibility we also have
 1 n/2
P (r2n ≤ −n2n/2 ) ≤ C 2 for n large,
 
n2 K2n
P r2n ≤ − 2n/2 ≤ n/2(k−2) for n large.
4 2
Now let us define stopping time τ by τ = inf{t : rt ≤ −2n2n/2 }. By Lemma 3.1
and attractiveness (or simply by Lemma 2.5 and Schinazi’s theorem) and the strong
Markov property we have
 1 n/2
P (r2n > −n2n/2 |Fτ ) ≤ C 2

on {τ < 2n },
where Fτ is the σ -field generated by the Harris system Poisson
processes up to stopping time τ .
 CRITICAL NEAREST PARTICLE SYSTEMS 19

Therefore
 
 1 n/2
P infn rs ≤ −2n2 n/2
≤ P (r2n ≤ −n2n/2 ) + C 2 ,
s≤2

from which it easily follows that P (infs≤2n rs ≤ −2n2n/2) ≤ C(1/2)n/2 . We

conclude
 
 1 n/2
P sup |rs | ≥ 2n2n/2 ≤ C 2 .
s≤2n

The second inequality follows in an entirely similar manner.

We can now “improve” Lemma 2.6:

P ROPOSITION 3.2. Consider two semiinfinite right NPSs in equilibrium

η, η so that both are generated by same Harris system:
(a) η0 is supported on (−∞, 0];
(b) η0 is supported on (−∞, −2m ].
Let the rightmost occupied site at time t for these processes be respectively
rt , rt . Then, outside of probability (C241m/8 /2km/15 ), rt ≥ rt ∀t ∈ [0, 23m ] and

∀28m/5 ≤ t ≤ 23m , ηt ≥ ηt

on [2−25m/8 /2, ∞).

P ROOF. Our approach follows that for Lemma 2.5 closely, so we will pass
quickly over the areas which are essentially the same and stress the differences.
Divide interval [−22−25m/8 , −2m /2] into (w.l.o.g.) intervals I0 , I1 , . . . , Iv of
length 222m/5 + 1, disjoint except for the endpoints. Take J0 , J1 , . . . , Jv+1 so that,
for i ≤ v, the midpoint of Ji is the left endpoint of Ii and so that the midpoint of
Jv+1 is the right endpoint of Iv .
Let us define Zti to be NPS on {0, 1}Ii with 1’s fixed at the endpoints of Ii and
such that Z0i ≡ 1 on Ii ; define Yti to be the NPS on {0, 1}Ii with 1’s fixed at the
endpoints of Ii , 0’s fixed within 22m/15 of the endpoints and (subject to the above)
Y0i ≡ η0 on Ii . Let Xi be the process on {0, 1}Ii which has X0i = Y0i and which
evolves like Y i except that it is forbidden to hit all zeros on the interior of Ii . Let
Y i , Xi be the processes defined when η is replaced by η .
Then outside of probability K241m/8 2−mk/15 we have for the x’s that are in the
central four-fifths of an Ii ,
η28m/5 (x) ≥ Y2i8m/5 (x) = X2i 8m/5 = Z2i 8m/5 (x) ≥ η28m/5 (x),
η2 8m/5 (x) ≥ Y2i8m/5 (x) = X2i8m/5 = Z2i 8m/5 (x) ≥ η2 8m/5 (x)
20 T. MOUNTFORD

and so
η28m/5 (x) = η2 8m/5 (x).
Repeating this argument with the J  s replacing the I  s we get, outside of the above
probability,
η28m/5 ≤ η2 8m/5 on [−2225m/8 , ∞).
We now argue as in Lemma 1.3. We define the bad set B to be the union of the
following unlikely events:


 
sup rt − r0  ≥ 22m ,
t≤23m


 
sup rt − r0  ≥ 22m ,
t≤23m
 
for s ≤ 23m , ηs has a 2m/8 /6V gap in [rs − 3225m/8, rs ] .
By Proposition 3.1 and Lemma 2.2 we have that P (B) ≤ K241m/82−mk/15 .
Now define nondecreasing process Ds , s ≥ 28m/5 , by D0 = −2225m/8 and, at
time t, D jumps forward by x units if and only if:
 ≡ 0 on (D , D + x], and also
(i) ηt− t− t−
(ii) there is an attempted birth at site Dt + x (i.e., if y is the distance to the
right of site x until there is an occupied site in ηt , then we have t ∈ BDt +x and
Ut ≤ β(x)β(y)
β(x+y) /M).

Then we have that the following hold (until Ds = rs ) on event B c :

1. ηs ≥ ηs on (Ds , ∞);
2. D jumps forward by at most 2m/8 /6V ;
3. D jumps forward at a rate bounded by V .
Thus, if
η28m/5 ≥ η2 8m/5 on [−2225m/8 , ∞),
then
∀28m/5 ≤ t ≤ 23m , ηt ≥ ηt on [−225m/8 , ∞),
unless either B occurs or D has at least 2V 23m jumps in time interval [28m/5 ,
23m ]. The result follows from the bounds on the tails of Poisson random variables
as supplied by Lemma 0.2.

P ROPOSITION 3.3. Let ηs be a semiinfinite NPS supported on (−∞, 0], either

in equilibrium on this interval or in full occupancy. For all 1√≤ l < k − 1 there
exists a constant Cl so that, for all t sufficiently large, E[(|rt |/ t)l ] ≤ Cl < ∞.
 CRITICAL NEAREST PARTICLE SYSTEMS 21

P ROOF. We treat the case of full occupancy, leaving the equilibrium case to
the reader. Fix g so that l < g < k − 1. Choose x0 > 4 so that
 ∞ −u2 /2
e 1
2 √ du < g .
x0 2π 2x0
Now by Schinazi’s result (or Corollary 2.4) we have that there exists t0 so that for
all t ≥ t0 we have

√ 1
P sup rs ≥ x0 t ≤ g ,
s≤t x0
and [by assumption (∗∗)]
√  1
∀y ≥ 4 t, β(z) < .
z≥y yg

Employing our Hoffman–Jørgensen argument we have


√   1 2 √
P sup rs ≥ 3x0 t ≤ g + t (x0 t)−g
s≤t x0
 
1 3g 1 1
≤ g g + (g−2)/2 ≤ .
(3x0 ) x0 t (3x0 )g
Applying this argument repeatedly, we find for all x ≥ x0 that
√ 3g
P (rt ≥ x t) ≤ g .
x
However, we have
√ √
P (rt ≤ −x t) ≤ P (rt ≥ x t)
from reversibility so we are done.

C OROLLARY 3.1. There exists constant C so that, for a NPS beginning in

equilibrium supported on (−∞, 0],

1 − C 2−n/6 + (241n/24 2−nk/45)1−3/k
E[r22n ] 
≤ n
≤ 1 + C 2−n/6 + (241n/242−nk/45 )1−3/k .
2

P ROOF. The proof for upper and lower bounds is essentially the same so we
will only consider the left-hand inequality:
E[r22n+1 ] = E[(r2n + r2n+1 − r2n )2 ],
22 T. MOUNTFORD

but if rt = r2n+1 − r2n , then


E[r22n+1 ] = E[(r2n + r2 n )2 ] = 2E[r22n ] + 2E r2n E η2n [r2n ]
 
= 2E[r22n ] + 2E r2n E η2n [r2n ]Ir2n >0 + 2E r2n E η2n [r2n ]Ir2n <0
so
E[r22n+1 ] E[r22n ] E[r2n E η2n [r2n ]Ir2n >0 ] E[r2n E η2n [r2n ]Ir2n <0 ]
≤ + + .
2n+1 2n 2n 2n
We analyze the term E[r2n E η2n [r2n ]Ir2n >0 ]; the third term is handled in a manner
entirely similar. Introduce a comparison NPS γt , t ≥ 2n , which is generated by the
same Harris system as η on t ≥ 2n , which at time 2n is in equilibrium supported on
(−∞, r2n +2n/3 ] but whose distribution at time 2n is otherwise independent of η2n .
γ
Let A be the event that r2n+1 > r2n+1 , where r γ denotes the rightmost particle of
process γ . (By Proposition 3.2 we have that the probability of event A is bounded
by C241n/24 /2nk/45 .) We have
  γ
E r2n E η2n [r2n ]Ir2n >0 ≤ E r2n [r2n+1 − r2n ]Ir2n >0

+ E r2n [r2n+1 − r2n ]Ir2n >0 IA
 γ
− E r2n [r2n+1 − r2n ]Ir2n >0 IA .
γ
The term E[r2n [r2n+1 − r2n ]Ir2n >0 ] is actually equal to E[r2n Ir2n >0 2n/3 ], by the
conditional independence of γ2n and the independent increments property of the
Poisson processes, and thus bounded by K25n/6 .
By Hölder’s inequality we have

E r2n E η2n [r2n ]Ir2n >0 IA
 3/k
≤ E[(r2n )k/3 (r2n+1 − r2n )k/3 ] P (A)1−3/k
 3/2k  3/2k
≤ E[(r2n )2k/3 ] E[(r2n+1 − r2n )2k/3 ] P (A)1−3/k ,
by Cauchy–Schwarz. By Propositions 3.2 and 3.3 we have that this last bound
is less than C(2n )(241n/242−nk/45)1−3/k . We similarly bound E[r2n (r2n+1 − r2n )×
γ

Ir2n >0 IA ].
So
E[r22n+1 ] E[r22n ]
≤ + C2−n/6 + C(241n/242−nk/45 )1−3/k .
2n+1 2n
Iterating and using the limit value of E[rt2 ]/t we have
E[r22n ] 
n
≥1− C2−m/6 + C(241m/24 2−mk/45 )1−3/k
2 m≥n
≥ 1 − C2−n/6 + C(241n/242−nk/45 )1−3/k .

 CRITICAL NEAREST PARTICLE SYSTEMS 23

P ROPOSITION 3.4. Let η be a semiinfinite NPS in equilibrium on (−∞, 0].

Then there exist independent semiinfinite NPSs in equilibrium on (−∞, 0],
2n/7
η1 , . . . , η2 , such that, with probability 1 − C245n/14 /24nk/105 ,
2n/7
2
r22n ≥ r2i 12n/7 − 26n/7 .
i=1
2n/7
There exist independent semiinfinite NPSs on (−∞, 0], η1 , . . . , η2 , such that,
 2n/7
with probability at least 1 − C245n/14 /24nk/105 , r22n ≤ 2i=1 r2i 12n/7 + 26n/7 .

P ROOF. Since η is in equilibrium, for all i = 0, 1, . . . , 22n/7 , we have (by

Proposition 3.2) that, outside of probability C241n/14 /24nk/105,
 
r(i+1)212n/7 − ri212n/7 ≥ r(i+1)2
i
12n/7 − ri212n/7 − 2
i 4n/7
,

where r i , is the rightmost particle for process ηi , the one sided NPS starting
12n/7 supported on (−∞, ri212n/7 − 2
at time i212n/7 with ηi2 i 4n/7 ] but otherwise

independent of Fi212n/7 and in equilibrium on (−∞, ri212n/7 − 24n/7 ], and run with
the same Harris system as η on time interval [i212n/7 , (i + 1)212n/7]. The first
i
assertion is proven by taking ηsi to be ηi2 12n/7 +s . The second follows similarly.

Let r2i 12n/7 be independent copies of r212n/7 (for a NPS initially supported on
(−∞, 0] in equilibrium). It is easily seen, by standard embedding methods, that
we can find a standard Brownian motion, B (possibly with enlarged filtration) and
stopping times τi ≤ τi+1 , i = 0, 1, 2, . . . , so that the following hold:
1. τ0 = 0;
 j
2. for i ≥ 1, B(τi ) is equal in distribution to ij =1 r212n/7 ;
3. the random variables τi+1 − τi are independent identically distributed random
variables with [for p ≤ (k − 3)/2],
E[(τi−1 − τi )p ] = Cp E[(r212n/7 )2p ].
Therefore using Proposition 3.3 we have that, for even p ≤ (k − 2)/2,
 p  22n/7 p  p
τ22n/7 − E[τ22n/7 ]  τi − τi−1 − E[τ1 ] 1
E =E .
22n 1
212n/7 22n/7

By Proposition 3.3, this is less than

 p
1 1
Hp 2 np/7
≤ Hp .
22n/7 2pn/7
(Here Hp is a constant not depending on n.)
24 T. MOUNTFORD

Thus for any even integer p ≤ (k − 2)/2 we have by Chebyshev that

  1
P τ22n/7 − E[τ22n/7 ] ≥ 22n−n/14 ≤ Hp .
2np/14
By taking p to be greater than 28 (recall k ≥ 100), this 
last term is bounded by
2−2n for n large. We therefore obtain after division by E[τ22n/7 ] that, for any
c ≥ 0,
 22n/7 i  
 i=1 r212n/7 B(E[τ 2n/7 ]) 
P   −  2 ≥c
E[τ ]
22n/7 E[τ ]  22n/7
 
≤ 2−2n + P sup |B(t) − B(1)| ≥ c ,
|t−1|≤K2−n/14

where K = 22n /E[τ2n/2 ]. By Proposition 3.3, K ≤ 2 for n large and we have that,
for large n,
 22n/7 i  
 i=1 r212n/7 B(E[τ22n/7 ])  −n/30

P   −  ≥2 ≤ 2 × 2−2n .
E[τ ] E[τ
22n/7 ]  22n/7

The above and Proposition 3.4 yield the existence of standard normal random
variables Z1 , Z2 so that for large n, outside of probability K245n/14 / 24nk/105 +
4 × 2−2n , we have
r22n r22n
 < 2−n/30 + Z1 ,  > −2−n/30 + Z2 .
E[τ22n/7 ] E[τ22n/7 ]

Clearly, outside of probability K245n/14/24nk/105 + 4 × 2−2n , we have Z1 + 2×

2−n/30 ≥ Z2 .
By Cauchy–Schwarz, we have that, for any event A,
 1/2  1/2  1/2
E[(Z2 − Z1 )IA ] ≤ P (A) E[(Z2 − Z1 )2 ] ≤ 2 P (A) .

Therefore we have

2 × 2−n/30 = E[Z2 + 2 × 2−n/30 − Z1 ]

≥ E[(Z2 + 2 × 2−n/30 − Z1 )IZ2 >Z1 +2−n/60 ]
− 2(K245n/14/24nk/105 + 4 × 2−2n )1/2 .

So, by Chebyshev, we have


P (Z2 > Z1 + 2−n/60 ) ≤ 2n/60 2 × 2−n/30 + (K245n/14/24nk/105 + 4 × 2−2n )1/2 .

Therefore, putting this all together we have the following proposition.

 CRITICAL NEAREST PARTICLE SYSTEMS 25

P ROPOSITION 3.5. For n large there is a standard normal r.v. Z so that

 
 r 
  22n
 E[τ − Z  ≤ 2−n/60
22n/7 ]
outside of a set of probability

2n/60 3 × 2−n/30 + (K245n/14/24nk/105)1/2 + 4 × 2−2n
for K not depending on n.

4. This section is unfortunately heavily technical and repetitive. It serves to

assemble many results which will be used in later sections. The basic approach
is always to argue (as in Section 2) that if a NPS does not have large gaps over
a certain interval and time period and has a reasonable initial configuration, and
if the time period is large, then at the end of it the NPS should look like renewal
measure on the spatial interval.
As part of this program we wish to introduce “finite NPSs in equilibrium” and
record some “regenerations results.” We have already abused definitions by talking
of one-sided NPSs “in equilibrium.” We take this one step further by introducing
“finite NPSs in equilibrium.” As has already been noted, finite NPSs must die out
so the term is nonsense. Nevertheless we believe the suggestiveness of the term
outweighs this consideration.
A finite NPS in equilibrium (ηt )t≥T , supported on (spatial) interval [x, y], is
generated by one-sided equilibrium NPSs ηR , ηL over (time) interval [S, T ] if the
following hold:
1. ηSL is supported on [x, ∞);
2. ηSR is supported on (−∞, y];
3. (ηtL , ηtR )t≥S are generated by the same Harris system;
4. ηT ≡ ηTL on (−∞, x+y x+y
2 ), ηT ≡ ηT on [ 2 , ∞);
R

5. (ηt )t≥T is generated by the same Harris system as (ηtL , ηtR )t≥S .
Note that ηT may have occupied sites outside of interval [x, y] and that sites x, y
may or may not be occupied. This is in contrast to the useage of “η is supported”
for semi infinite processes.
The following result shows that, for “sensible” choices of time interval [S, T ]
given interval length y − x, the definition of η is not so arbitrary.

L EMMA 4.1. Let ηR be a one-sided NPS in equilibrium supported by

(−∞, N ] at time 0 and ηL a one-sided NPS in equilibrium supported by [0, ∞)
at time 0. Let η0R , η0L be independent but let the same Harris system generate the
two processes. Outside of probability C(N logp (N ) + 1)(2N + 1)(N logp (N ))1/4 /
(N logp (N ))k/24 ≤ KN 5/2 /N k/24 , we have that, at time N logp (N ) for |p| ≤
7, ηR is equal to ηL on [N/10, 9N /10]. The constant K may be taken independent
of p.
26 T. MOUNTFORD

P ROOF. As usual we only need treat the case where N is large. The proof
closely follows that of Lemma 2.5 and Proposition 3.2. We denote the rightmost
particle of ηtR by rt and the leftmost particle of ηtL by lt . First note that by
Proposition 3.1 (which applies equally to ηL ) we have that, outside probability
KN logp (N )/(N logp (N ))(k−2)/2 ,
(i)
sup {|rt − r0 |, |lt − l0 |} ≤ N 2/3 .
0≤t≤N logp (N)

We also have by Lemma 2.2(i) (which again applies to ηL ) that, outside

probability K(N logp (N ) + 1)(2N + 1)(N logp (N ))1/4 /(N logp (N ))k/24 :
(ii)
0 ≤ t ≤ N logp (N ) so that ηtR has an (N logp (N ))1/12 gap in
[rt − 2N, rt ];
(iii)
0 ≤ t ≤ N logp (N ) so that ηtL has an (N logp (N ))1/12 gap in [lt , lt +2N ];
and that by Lemma 2.2(ii) (outside of the same probability):
(iv)
x ∈ [−N, N − (N logp (N ))1/4 ) so that

η0R ◦ θx [−(N logp (N))1/4 ,(N logp (N))1/4 ] is bad;

(v)
x ∈ [(N logp (N ))1/4 , 2N ] so that


η0L ◦ θx [−(N logp (N))1/4 ,(N logp (N))1/4 ] is bad.
Let B (for bad) be the union of the complements of events 1–5. The above may
be summarized by the bound
 1/4  k/24
P (B) ≤ K N logp (N ) + 1)(2N + 1)(N logp (N ) N logp (N ) .
We divide the interval [N/10, 9N/10] into disjoint (except at the endpoints)
intervals of length 2(N logp (N ))1/4 + 1, . . . , I1 , I2 , . . . , IV , where V ≤ N . As with
Lemma 2.5 we also take intervals J0 , J1 , . . . , JV of the same length so that, for
i < V , Ji has midpoint equal to the left endpoint of Ii+1 , while JV has midpoint
equal to the right endpoint of IV . As with Lemma 2.5 we introduce processes
  
Y i,L , Y i,R , Z i on {0, 1}Ii and Y i,L , Y i,R , Z i, on {0, 1}Ji , where:
(a) Z i is the finite nearest particle system (generated by the Harris system
for ηR , ηL ) which has 1’s fixed at the endpoints of Ii and initially has all sites
occupied;
(b) Y i,L , Y i,R are the finite NPSs (again generated by the given Harris system)
which have 1’s fixed at the endpoints of Ii , 0’s fixed within (N logp (N ))1/12 of the
endpoints of Ii and (subject to this) Y0i,L = η0L on Ii , Y0i,R = η0R on Ii .
Similarly for the primed processes with Ii replaced by Ji . Then as with Lemma 2.5,
by Lemma 2.2 and Corollary 2.1, we have that, for all sites x in the middle four-
fifths of Ii ,
YNi,Llogp (N) (x) = YNi,Rlogp (N) (x) = ZN
i
logp (N) (x),
 CRITICAL NEAREST PARTICLE SYSTEMS 27

outside of probability
KN 1
p p
(N log (N ))1/4 (N log (N ))(k−3)/12
 1/4  k/24
+ K N logp (N ) + 1)(2N + 1)(N logp (N ) N logp (N ) .
However, on the event B we have that, for such x,
YNi,Llogp (N) (x) ≤ ηN
L
logp (N) (x) ≤ ZN logp (N) (x)
i

and
YNi,Rlogp (N) (x) ≤ ηN
R
logp (N) (x) ≤ ZN logp (N) (x).
i

Similarly for sites in the middle four-fifths of intervals Ji . We deduce that outside
of the contracted probability

logp (N) (x) = ηN logp (N) (x) ∀x ∈ [N/10, 9N/10].

R L
ηN

Using essentially the same proof we may show the following lemma.

L EMMA 4.2. Let η be a finite NPS with rightmost particle at time t, rt and
leftmost lt and so that r0 − l0 = N . Let B(p) (|p| ≤ 7) be the union of the following
events:
(i) sup0≤t≤N logp (N) {|rt − r0 | + |lt − l0 |} ≤ N 2/3 /4;
(ii) there does not exist t ∈ [0, N logp (N )] so that ηt has an (N logp (N ))1/12
gap on [lt , rt ];
(iii) ∀l0 + (N logp (N ))1/4 ≤ x ≤ r0 − (N logp (N ))1/4 ,
η0 ◦ θx |[−N logp (N))1/4 ,N logp (N))1/4 ] is good.

Then, outside of probability P (B) + KN 5/2 /N k/24 , we have

ηN logp (N) (x) = ηN
1
logp (N) (x) for x ∈ [l0 + N 2/3 /2, r0 − N 2/3 /2],

where η1 is the NPS run with η s Harris system and starting from full occupancy.

The following result is crucial as it enables one to transfer arguments and results
for one-sided processes to arguments and results for large finite systems.

L EMMA 4.3. Let η be a NPS on {0, 1}[−n,n] with 1’s fixed at −n, n. Let η
be a NPS on {0, 1}[−n,n] with 1’s fixed at −n, n and 0’s fixed at (−n, −n + n1/3 ),
(n − n1/3 , n). Suppose both are generated by the same Harris process and η0 , η0
are derived from a renewal process γ on Z (or a renewal process γ conditioned to
28 T. MOUNTFORD

have γ (x) = 1 for some fixed x with |x| ≥ n) as follows:

(i) η0 ≡ γ on (−n, n);
(ii) η0 ≡ γ on (−n + n1/3 , n − n1/3 ).
Then
 c
P
t ≤ n8 such that ηt (x) > ηt (x) for x ∈ (−4n/5, 4n/5) ≤ .
nk/3−10
P ROOF. The proofs are the same for the two cases of γ so we simply treat the
first.
We observe that by attractiveness ηt ≤ ηt ∀t and that by the strong Markov
property (for the coupled processes) if, for any x, ηt (x) = 1, ηt (x) = 0, then this
state will persist for time 1 with probability at least e−m for m = 1 + β(1)2 /β(2).
Therefore, using these two facts, we deduce that
 
4n
P
t ≤ n such that
8
ηt (x) > ηt (x)for|x| ≤
5
 n8 +1   
≤ em P ηt (x) = 1 − P ηt (x) = 1 dt.
0 |x|≤4n/5

The equilibrium measure π [−n,n] for η is renewal measure on {0, 1}[−n,n]

conditioned to have 1’s at sites −n, n. Therefore this distribution is stochastically
above renewal measure restricted to interval [−n, n], which is the distribution
of η0 . Thus η0 is stochastically below measure π [−n,n] and so ηt is also
stochastically below π [−n,n] for each t ≥ 0. That is, ∀x, t, P (ηt (x) = 1) ≤
π [−n,n] {η : η(x) = 1}; in particular this is true for every t ∈ [0, n8 + 1] and
x ∈ (−4n/5, 4n/5).
Also conditional on γ being nonidentically zero on both (−n, −n + n1/3 ) and

(n − n1/3 , n), η0 is distributed above ηt ’s natural equilibrium measure π [−n,n] and
so, ∀x, t,
  
P ηt (x) = 1 ≥ π [−n,n] ({η : η(x) = 1}) − 2P γ ≡ 0 on (0, n1/3)
and so the integral on the right-hand side of (†) is bounded by

Kn9 P γ ≡ 0 on (0, n1/3)
 
+ Kn8 π [−n,n] ({η : η(x) = 1}) − π [−n,n] ({η : η (x) = 1})
x∈(−4n/5,4n/5)

Kn9 Cnn8 Cn10

≤ + < .
n(k−2)/3 n(k−2)/3 nk/3
by Lemmas 2.4 and 2.1.

Lemma 4.3 enables us to relate the behavior of finite NPSs in equilibrium over
a time interval of order N 2 to that of semiinfinite processes in equilibrium.
 CRITICAL NEAREST PARTICLE SYSTEMS 29

L EMMA 4.4. Fix p with |p| ≤ 7. Let η be a finite NPS in equilibrium on [0, N ]
generated by ηR , ηL on time interval [0, N logp (N )]. Then, outside of probability
KN 7/2 /N k/24 , we have, for N logp (N ) ≤ t ≤ N 2 / log5 (N ),
ηt |[N/5,∞) = ηtR |[N/5,∞)
and
ηt |(−∞,4N/5] = ηtL |(−∞,4N/5] .

P ROOF. As usual we need only concern ourselves with large N . We al-

ready know from Lemma 4.1 that, outside probability KN 5/2 /N k/24 , ηN R
logp (N)
≡ ηNL
logp (N) on [N/10, 9N/10]. By definition we have η p
N log (N) ≡ η R
N logp (N) on
[N/2, ∞) and ηN logp (N) ≡ ηN logp (N) on [−∞, N/2). Therefore we have, out-
L

side probability KN 5/2 /N k/24 , that ηN logp (N) |[N/10,∞) = ηNR

logp (N) |[N/10,∞) and
ηN log (N) |(−∞,9N/10] = ηN logp (N) |(−∞,9N/10] . Let us denote this event by A.
p L

We choose intervals I, J of length 2N 1/4 + 1 so that the endpoint of I is N/5

and the midpoint of J is 4N/5. As with Lemma 4.1, for H ∈ {I, J }, we define
processes Y H,L , Y H,R to be finite NPSs on {0, 1}H , where 1’s are fixed at the
endpoints of H , 0’s are fixed within N 1/12 of the endpoints but otherwise the
processes evolve as NPSs generated by our Harris system. For k = R, L, H ∈
{I, J }, Y0H,k ≡ η0k on H , subject to the above constraints.
For k = R, L we have processes Z H,k which are NPSs on {0, 1}H run with
the same Harris system, with 1’s fixed at the endpoints and with (subject to this)
η0k ≡ Z0H,k on H .
By Lemma 4.3 we have, for k ∈ {R, L}, H ∈ {I, J }, for all 0 ≤ t ≤ N 2 and all x
in the middle four-fifths of H , outside probability KN 5/2 /N k/12 , that

(1) YtH,k (x) = ZtH,k (x)

for H ∈ {I, J }; but also, by Corollary 2.2, we have that, outside probability
Ce−N , for H ∈ {I, J },
1/4

H,1 H,R H,L

(2) ZN logp (N) ≡ ZN logp (N) ≡ ZN logp (N) ,

where Z H,1 is the finite NPS on {0, 1}H with 1’s fixed at the endpoints and so
that initially all sites are occupied. Thus we have ZtH,R ≡ ZtH,L ∀t ≥ N logp (N )
outside of this probability. Let B be the union of events (1) and (2) above. Then
P (B c ) ≤ KN 5/2 /N k/24 and on event B we have for H ∈ {I, J } that, for all
N logp (N ) ≤ t ≤ N 2 and all x in the middle four-fifths of H ,

YtH,R (x) = ZtH,R (x) = ZtH,L (x) = YtH,L (x).

30 T. MOUNTFORD

Let event C be the union of the following events:

0 ≤ t ≤ N 2 / log5 (N ) so that |rt − N | ≥ N/20, or |lt | ≥ N/20;

t ∈ [0, N 2 ] so that ηR has an N 1/12 gap in [rt − 2N, rt ];

t ∈ [0, N 2 ] so that ηL has a N 1/12 gap in [lt , lt + 2N ].

Then by Lemma 2.2 and Proposition 3.1 we have P (C c ) ≤ KN 2 2N N 1/4/
N −k/24 ≤ KN 7/2 /N −k/24 . By attractiveness we have that, on event C, Y H,k ≤
ηk ≤ Z H,k on interval H ∈ {I, J }. We conclude that, on event A ∩ B ∩ C, ηtR ≡ ηtL
on the middle four-fifths of H for N logp (N ) ≤ t ≤ N 2 / log5 (N ).
We claim that a.s.
ηt ≡ ηtR on [N/5, ∞) for N logp (N ) ≤ t ≤ N 2 / log5 (N )
and
ηt ≡ ηtL on (−∞, 4N/5] for N logp (N ) ≤ t ≤ N 2 / log5 (N )
on event A ∩ B ∩ C, an event of probability greater than 1 − KN 7/2 /N k/24 .
We first note that, since Harris death points unambiguously cause deaths at
occupied sites, the time
 
τ = inf t ≥ N logp (N ) : ηt = ηR on [N/5, ∞) or ηt = ηL on [−∞, 4N/5)
must occur at a potential birth point for the Harris system. Similarly
 
σ = inf t ≥ N logp (N ) : ηt has an N 1/4 gap in I ∪ [N/5, 4N/5] ∪ J
can only occur at a death point of the Harris system. Therefore we have that a.s.
τ = σ . We will see that no other relation is possible if τ < N 2 / log5 (N ).
First note that for t < τ we have that ηtR ≡ ηt on [N/5, ∞) and so (on event
A ∩ B ∩ C) ηt has no N 1/4 gap on interval [N/5, 9N/10]. Similarly ηt has no
N 1/4 gap on [N/10, 4N/5] (if A ∩ B ∩ C holds). We deduce that a.s. σ ≥ τ .
However conversely for N logp (N ) ≤ t < σ we have that, on interval I ,

YtI,L ≤ ηt ≤ Z I,L
and on J ,
YtJ,R ≤ ηt ≤ Z J,R .
Therefore for such t (on event A ∩ B ∩ C) we have that, for x in the middle four-
fifths of I and J ,
ηt (x) = ηtR (x) = ηtL (x).
 CRITICAL NEAREST PARTICLE SYSTEMS 31

But if immediately prior to τ, η, ηR , ηL have no N 1/4 gaps on [N/10, 9N/10],

then at time τ we must have that either there exists x in the middle four-fifths of I
so that
ητR (x) = ητ (x)
or there exists x in the middle four-fifths of J so that
ητL (x) = ητ (x).
However, as we have seen on event A ∩ B ∩ C this is impossible before time σ .
This contradiction implies that ∀N logp (N ) ≤ t ≤ N 2 / log5 (N ),
ηt |[(N/5),∞) = ηtR |[(N/5),∞) and ηt |(−∞,(4N/5)] = ηtL |(−∞,(4N/5)] ,
outside of probability KN 7/2 /N k/24 .

We are now ready to begin establishing results for “finite edges processes in
equilibrium.” Lemma 4.4 lets us regard finite NPSs as close to semiinfinite NPSs
in equilibrium which in turn are locally (at appropriate places) close to equilibrium
NPSs. The following lemma makes this a little more concrete.

L EMMA 4.5. Let f be a fixed cylinder function and let δ be a constant strictly
greater than 0. There exists N0 = N0 (f, δ) so that, for all N ≥ N0 , if ηN is a
finite NPS, in equilibrium on [0, N ], generated by ηR,N , ηL,N over time interval
[0, N logp (N )] with |p| > 7 and if x N satisfies
N − x N ≤ N/ log(N )
and T N satisfies
N2
N logp (N ) ≤ T N ≤ ,
log40 (N )
then
|E[f (ηT N ◦ θx N )] − Ren(β), f | ≤ δ.

P ROOF. Either
N N N 2N
≤ xN ≤ N − or ≤ xN ≤ ,
3 log(N ) log(N ) 3
or both. We assume without loss of generality that the former holds. By Lemma 4.4
it will suffice to show
E[f (ηTR,N
N ◦ θx N )] − Ren(β), f  → 0.

We introduce infinite NPS η run with the same Harris system as ηR,N and in
equilibrium. By the meaning of equilibrium, we have
E[f (ηT N ◦ θx N )] = Ren(β), f .
32 T. MOUNTFORD

Thus to show the lemma it will suffice to show that with probability tending to 1
as N → ∞, ηT N = ηTR,N N on [x
N − m, x N + m], where [−m, m] is the support of

cylinder function f .
We argue as in the previous lemma. Let I be the interval, centred at x, of length
2(T N )1/4 + 1. Let ZtI be the finite NPS on {0, 1}I with 1’s fixed at the endpoint
of I and starting from full occupancy at time 0.
Let Y I (Y I  ) be the finite NPSs on {0, 1}I with 1’s fixed at the endpoints, 0’s
fixed within (T N )1/12 of the endpoints but otherwise Y0I = η0R,N (Y0I  = η0 ). Then
we have with probability tending to 1 as N → ∞ that, for all y in the central half
of I ,
YTI N (y) ≤ ηTR,N
N (y) ≤ ZT N (y),
I

YTI N (y) ≤ ηT N (y) ≤ ZTI N (y)

and
YTI N (y) = YTI N (y) = ZTI N (y).

We conclude that as N → ∞, for y ∈ [x N − m, x N + m],


ηTR,N
N (y) = ηT N (y).

This proves the lemma.

C OROLLARY 4.1. Let η be a finite NPS in equilibrium on [0, N ] generated by

ηR , ηL during time interval [0, N logp (N )] (|p| ≤ 7). Let T be a fixed, nonrandom
time on time interval [N logp (N ), N 2 / log5 (N ) − N logq (N )] for |q| ≤ 7, and let
(ηt1,T )t≥T be the one-sided NPS generated by the same Harris system as η such
that
ηT1,T (x) = Ix≤rT −N 3/5 /5 .
Then
ηT1,T+N logq (N) ≤ ηT +N logq (N) on [N/5, ∞),

outside of probability KN 5 /N k/12 .

C OROLLARY 4.2. Let η be a finite NPS in equilibrium on [0, N ] generated

by ηR , ηL during time interval [0, N logp (N )] (|p| ≤ 7). Then, outside of
probability KN N 3c N −ck/2 + KN 7/2 /N k/24 , for all t satisfying N logp (N ) ≤ t ≤
N 2 / log5 (N ), there is no N c gap within ηt . Equally the chance that for some t
satisfying N logp (N ) ≤ t ≤ N 2 / log5 (N ) and for some x within the interval of
support of ηt and at least N c away from its boundary we have ηt ◦ θx |[−N c ,N c ] is
bad is bounded by KN N c N −ck/6 + KN 7/2 /N k/24 .
 CRITICAL NEAREST PARTICLE SYSTEMS 33

P ROOF. Let N logp (N ) ≤ t ≤ N 2 / log5 (N ). By Lemma 4.4, the probability

that there is an N c gap “within” ηt is bounded by KN 7/2/N k/24 + P (ηtR
has an N c gap within [rt − 2N, rt ]) + P (ηtL has an N c gap within [lt , lt +
2N ]) + P (supt≤N 2 / log5 (N) {|rt − r0 | + |lt |} ≥ N/2). The result now follows from
Lemma 2.2(i).
The second part of the corollary follows similarly [with Lemma 2.2(i) replaced
by Lemma 2.2(ii)].

From this corollary we immediately deduce (arguing as in Lemma 4.2) the

following proposition.

P ROPOSITION 4.1 (Regeneration proposition).. Let p, q satisfy 0 ≤ p, |q|,

|q| ≤ 7. Let η be a finite process in equilibrium on [0, N ] generated over time
η η
interval [0, N/ logp (N )]. Let lt (resp. rt ) be the position of the leftmost (resp.
rightmost) particle of η at time t for t ≥ N/ logp (N ). Let T be a nonrandom
time in the interval [N (log|q| (N ) + logp (N )), N 2 / log5 (N )]. Let γtL , γtR for t ≥
T − N log|q| (N ) be two one-sided NPSs in equilibrium. At time T − N log|q| (N ),
η 
γ L will be supported on [lT −N log|q| (N) + N 2/3 logq (N ), ∞] and γ R will be
η 
supported on (−∞, rT −N log|q| (N) − N 2/3 logq (N )], but otherwise γTR−N log|q| (N)
and γTL−N log|q| (N) are independent of each other and of η at this time. Let γ R , γ L
evolve according to the same Harris system as generated η. Then outside of
probability KN 7/2 /N k/24 we have that the finite NPS generated by γ R , γ L over
time interval [T − N log|q| (N ), T ] is dominated by η.

The importance of this proposition is that it enables us to replace periodically

finite NPSs in equilibrium by finite NPSs whose initial support is close to the
region of occupancy of η at the time but which are otherwise independent. The
price is that a slight shrinkage is involved and that (with very small probability)
there could be a breakdown.
We also have the following proposition which states that the motion of the
rightmost particle for a semiinfinite NPS in equilibrium does not depend on the
generating Harris system far from the rightmost site.

P ROPOSITION 4.2. Let η be a one-sided NPS in equilibrium (initially

supported on (−∞, 0]); let η be a (possibly finite) NPS with η0 |[−3N 2/3 ,∞) =
η0 |[−3N 2/3 ,∞) . Suppose that η (resp. η ) are generated by Harris systems H
(resp. H  ) so that
H |[−3N 2/3 ,∞) = H  |[−3N 2/3 ,∞) .
Then

P ηt |[−N 2/3 ,∞) = ηt |[−N 2/3 ,∞) ∀0 ≤ t ≤ N 16/3 ∧ τ ≥ 1 − KN 20/3 /n2k/9
34 T. MOUNTFORD

for τ = inf{s : ηs ≡ 0 on [−3N 2/3 , −3N 2/3 + N 2/9 ] or ηs ≡ 0 on [−3N 2/3 ,
−3N 2/3 + N 2/9 ] or ηs ≡ 0 on [−N 2/3 − N 2/9, −N 2/3 ]}.

P ROOF. Consider the finite NPSs on {0, 1}[−3N ,−N ] , Z and Y generated
2/3 2/3

by H (and therefore by H  ), where Z has 1’s fixed at endpoints −3N 2/3, −N 2/3 ,
while Y has 1’s fixed at these endpoints but also 0’s fixed within N 2/9 of the
endpoints. Initially both coincide with η0 (and therefore η ) subject to these
constraints (as in the preceding lemma).
Let B be the union of the following events:
1. Yt has an N 2/3 /2 gap for some 0 ≤ t ≤ N 16/3 ;
2. ∃0 ≤ t ≤ N 16/3, − 14
5N
2/3 ≤ x ≤ − 6 N 2/3 so that Z (x) > Y (x).
5 t t

The probability of event 2 occurring but not event 1 is, by preceding lemma,
bounded by KN 9 /N 2k−/9 . Given the large value of k we have that P (B) ≤
KN 5/3 /N (k−1)/18 .
We suppose that B does not occur. Let τ  = τ ∧ inf{ηs ≡ 0 on [−N 2/3 − N 2/9 ,
−N 2/3 ]}. Now as all processes on interval [−3N 2/3 , −N 2/3 ] are generated by the
same Harris system we have that, for all times t ≤ τ  ∧ N 16/3 and x in the interval
[−3N 2/3 , −N 2/3 ],
Yt (x) ≤ ηt (x), ηt (x) ≤ Zt (x).
As B does not occur we have
Yt (x) = Zt (x) for − 14
5N
2/3
≤ x ≤ − 65 N 2/3 .
This implies that for the relevant t and x that ηt (x) = ηt (x). Since (by the
nonoccurrence of event 1 above) there will always be an occupied site for η (and
therefore η ) in the interval [− 14
5N
2/3 , − 6 N 2/3 ], we cannot have the spontaneous
5
appearance of discrepancies of the processes to the right of − 65 N 2/3 before time τ .
That is (on event B),
  
τ  ∧ N 16/3 ≤ σ = inf t : ηt = ηt on − 65 N 2/3 , ∞ .
We are done since the first time we have a discrepancy between η and η on an
interval must correspond to a potential birth time for our Harris system while τ
must correspond to a death time for the Harris system and thus we must have
 
σ > inf t : ηt is vacant on [−N 2/3 − N 2/9 , −N 2/3 ] .

Proposition 4.2 and Lemma 4.4 yield:

C OROLLARY 4.3. Let one-sided equilibrium processes ηR , ηL generate η on

 
[0, N ] at time N logp (N ). Let ηR be a right one-sided NPS such that η0R = η0R

and ηR is generated by a Harris system H  such that H |[2N/3,∞) = H  |[2N/3,∞) .
 CRITICAL NEAREST PARTICLE SYSTEMS 35
  
Similarly, let ηL be a left one-sided process such that η0L = η0L and such that ηL
is generated by H  , where H  |(−∞,N/3] = H |(−∞,N/3] . Then outside probability
 
KN 7/2 /N k/24 we have ηt = ηtR on [4N /5, ∞) and ηt = ηtL on (−∞, N/5] for
N logp (N ) ≤ t ≤ N 2 / log 5 (N ).
 
R EMARK . Of course ηR and ηL may be taken to be independent. This
result says that edge fluctuations of a finite NPS in equilibrium are essentially
independent.

5. The purpose of this section is to establish a weak convergence result for the
behavior between boundaries of multiple intervals alternating between vacancy
and the upper regime.
For a configuration γ , we say that it is L-good on interval I if
there are no L1/3 gaps for γ within interval I
and
for every x, so that [x − L, x + L] ⊂ I , we have γ ◦ θx |[−L,L] is good.
Obviously for a process γt when we write γt is L-good on It for t ∈ [S, T ]
we mean that, simultaneously for all t ∈ [S, T ], the configuration γt is L-good for
interval It .
We wish to show:

T HEOREM 5.1. Let finite NPS ξ N be such that ξ0N ≡ 0 on (−∞, −λN ) ∪
(0, N ) ∪ ((1 + κ)N ,∞); ξ0N is “in equilibrium” on [−λN, 0] and on [N, (1 + κ)N ]
N,γ N,γ
where κ, λ > 0. There are associated processes lt , rt , ltN = inf{x : ξtN (x) = 1}
and rtN = sup{x : ξtN (x) = 1} and random variable τ N , so that for τ defined below
N,γ N,γ
we have for t ≤ τ N and v ≥ 18/k that ξtN is N v -good on [ltN , lt ] ∪ [rt , r N ]
and
 N
lN 2 t rN 2 t rNN2 t 
N,γ N,γ
l N 2t
, , , → (Bt1 , Xt1 , Xt2 , Bt2 )t≤τ ,
N N N N t≤τN /N 2
where (B01 , X01 , X02 , B02 ) = (−λ, 0, 1, 1 + κ) and (B 1 , B 2 ) is a standard two-
dimensional Brownian motion independent of (X1 , X2 ) and where
Xt1 = Wt1 + µt and Xt2 = Wt2 − µt for X2 > X1
and τ = 1 ∧ inf{t : |Bt1 + λ| > λ3 or |Bt2 − (1 + κ)| > κ3 or |Xt1 | > λ3 ∧ 13 or
|Xt2 − 1| > κ3 ∧ 13 }. The random variable τ N is the analogous stopping time
N,γ
for ξ N ; τ N = inf{t : |ltN + λN | > N λ/3 or |rtN − (1 + κ)N | > κ/3 or |lt | >
N,γ
(λ/3 ∧ 1/3)N or |rt | > (κ/3 ∧ 1/3)N }.
36 T. MOUNTFORD

To establish this we first consider a simpler model which is semiinfinite. Let

ξtN be a semiinfinite process on {0, 1}Z, ξ0N is Renewal measure on (−∞, 0],
conditioned to have a 1 at site 0, 0 on (0, ∞),

 1, if ξ N (x) = 1,
c (x, ξ ) = β(H, r),
N N
if r < ∞, ξ N (x) = 0,

1 + 1/N β(H), if r = ∞, ξ N (x) = 0.
Let rtN denote the position of the rightmost particle of ξtN .

R EMARK . Note that for this perturbed process, ξ N , the process seen from the
rightmost particle is in equilibrium and as such the bounds of Lemma 2.2 apply to
the configurations of these processes as seen from the extreme particle. It should
also be noted, however, that this new process is neither reversible nor attractive.
However, we may couple together ξ N and η, a one-sided semiinfinite process, so
that ξ0N = η0 and ∀t ≥ 0, ηt ≤ ξtN .

In the following we will consider both the regular equilibrium semiinfinite

NPS and our perturbed semiinfinite NPS described above. For probabilities and
expectations for the former we will use P or the suffix P ; for the latter we
will use Q. Let Y be the number of jumps to the right of the rightmost particle
in time interval [0, N 3/2 ]. Under Q, Y is Poisson(N 3/2 + N 1/2 ), under P it is
Poisson(N 3/2). Obviously the two distributions are very close in absolute variation
norm. Let us consider under probability Q, the random quantity
XN = rNN3/2 I|Y −N 3/2 |≤N 5/6 .
I wish to show:

L EMMA 5.1. The quantity XN satisfies the following:

 
(i) E Q [XN ] = µN 1/2 + O e−gN
1/6
for µ = nβ(n)n and some positive g
not depending on N ;
(ii) E Q [(XN )2 ] = N 3/2 (1 − O(1/N 1/10 )); and
(iii) E Q [(XN )4 ] ≤ KN 3 for some K not depending on N .

R EMARK . From the proof it will be clear that the conclusions remain valid if
time N 3/2 is replaced by some fixed, nonrandom time equal to N 3/2 (1 + o(1)).

P ROOF OF L EMMA 5.1. We first treat (i):

E [X ] = E
Q N Q
[rNN3/2 ] − E Q [rNN3/2 IY ≤N 3/2 −N 5/6 ] − E Q [rNN3/2 IY ≥N 3/2 +N 5/6 ].

The first term on the right-hand side is precisely equal to µN 1/2 for µ = nβ(n)
so it only remains to bound the absolute value of the other two right-hand-
side terms appropriately. The second term on the right-hand side is bounded in
 CRITICAL NEAREST PARTICLE SYSTEMS 37

magnitude by
  
E Q |rNN3/2 |IY ≤N 3/2 −N 5/6 = E Q |rNN3/2 | | Y = i Q(Y = i).
i<N 3/2 −N 5/6
However,
 
E Q |rNN3/2 | | Y = i = E P |rNN3/2 | | Y = i
for all i while for i’s in this range
 i
Q(Y = i) e−(N
3/2 +N 1/2 )
(N 3/2 + N 1/2)i 1
= e−N
1/2
= 1+ ≤ 1.
P (Y = i) e−N
3/2
(N 3/2 )i N
So we have
 
E Q |rNN3/2 |IY ≤N 3/2 −N 5/6 ≤ E P |rNN3/2 |IY ≤N 3/2 −N 5/6 .
This latter bound is, by Cauchy–Schwarz, bounded by
 P N 2 1/2 
< Ke−cN
1/2 1/6
E [|rN 3/2 | ] P (Y ≤ N 3/2 − N 5/6 )
for large N by Proposition 3.3 and Lemma 0.2. For the third term we split it up as
   
E Q |rNN3/2 |Ir N >0 IY ≥N 3/2 +N 5/6 + E
Q
|rNN3/2 |Ir N <0 IY ≥N 3/2 +N 5/6 .
N 3/2 N 3/2

The second term can be taken care of by the observation above that ξ N can be
coupled with η, an equilibrium semiinfinite NPS, initially supported on (−∞, 0],
with ηt ≤ ξtN ∀t ≥ 0 and so, in particular, sup{x : ηt (x) = 1} = rt ≤ rtN for all t. It
is easy to see by Cauchy–Schwarz that
 
E Q |rNN3/2 |Ir N <0 IY ≥N 3/2 +N 5/6
N 3/2
 
≤ E Q |rN 3/2 |Ir N <0 IY ≥N 3/2 +N 5/6
N 3/2

≤ Ce−cN
1/2 1/6
≤ (E Q [rN2 3/2 ])1/2 Q({Y ≥ N 3/2 + N 5/6 })
by Proposition 3.3 and Lemma 0.2.
By considering a NPS with no deaths, we have that

E Q |rNN3/2 |Ir N >0 | Y ≤ µY.
N 3/2

Together these two results (and the above bounds for tails of Poisson random
variables) yield

E Q |rNN3/2 |IY ≥N+N 5/6 ≤ Ce−gN
1/6
.
For (ii) we note that, by Proposition 3.3 and Cauchy–Schwarz,

E (rNN3/2 )2 I|Y −N 3/2 |>N 5/6 < e−gN
1/6
.
38 T. MOUNTFORD

Thus
   
P 1
E (rNN3/2 )2 | Y = i P (Y = i) = 1 − o ,
N 1/6
|i−N|≤N 5/6
P (Y =i)  
but it is easily seen that for i in this range  Q(Y 
=i) − 1 ≤
1
N 1/10
. This yields (ii).
Part (iii) follows in the same way.

Lemma 5.1, the Berry–Esseen theorem and standard weak convergence argu-
ments [see Durrett (1996) page 126], show the following:

L EMMA 5.2. Let ZiN be i.i.d. random variables with distribution equal to
rNN3/2 for a perturbed NPS initially in equilibrium, supported on (−∞, 0];
1  N
XtN = Zi
N √
i≤ Nt

tends in distribution as N tends to infinity to

X0 = 0,
Xt = Wt + µt,
where W is a standard Brownian motion.

L EMMA 5.3. For T ≤ 2N 3/2 we have for process ξ N , started in equilibrium

on (−∞, 0] with rightmost particle r N , that
 √ 
Q sup |rtN | ≥ log2 (T ) T ≤ CT /T (k−1)/2 .
t≤T

P ROOF. Since we can couple ξ N and a one-sided NPS in equilibrium η, so

that ηt ≤ ξtN for all t, by Proposition 3.1, we need only show
 √ 
Q sup rtN ≥ log2 (T ) T ≤ CT /T (k−1)/2 .
t≤T

Let random variable Y be the number of rightward jumps by the rightmost

particle of ξ N in the time interval [0, T ]; Y is a Poisson(T (1 + 1/N )) random
variable. Accordingly, by Lemma 0.2, we have that Q(|Y − T | ≥ T 7/12 ) ≤
CT /T (k−1)/2 .
On the other hand we have that if pλ (j ) is the probability that a Poisson(λ)
random variable is equal to j , then
 
 pT (j ) 
sup  − 1 = 1 + o(1) as T → ∞.
p (j )
|j −T |≤T 7/12 T (1+1/N)
 CRITICAL NEAREST PARTICLE SYSTEMS 39

Therefore we have that

√
Q sup rtN ≥ log2 (T ) T
t≤T
 √
≤ CT /T (k−1)/2 + pT (1+1/N) (j )P sup rtN ≥ log2 (T ) T |Y = j
|j −T |≤T 7/12 t≤T

 √
≤ CT /T (k−1)/2 + 2 pT (j )P sup rt ≥ log2 (T ) T |Y = j ,
|j −T |≤T 7/12 t≤T

where rt is the position of the rightmost particle of a one-sided NPS started from
(−∞, 0] in equilibrium and, again, Y is the number of rightward jumps made by its
rightmost particle. By Proposition 3.1, this latter sum is bounded by CT /T (k−1)/2 .
We are done.

C OROLLARY 5.1. Let ξ0N be in equilibrium supported on (−∞, 0]. Let ξtN,R
(R denoting restriction) be the process with the same initial configuration,
√ run
with the same Harris system but so that no births outside (−∞, log (T ) T ] are
2

permitted;
∀0 ≤ t ≤ T , ξtN = ξtN,R ,
outside probability CT /T (k−1)/2 .

We now check that a major tool in analyzing semiinfinite NPSs is still in force
for our perturbed NPSs.

L EMMA 5.4. There exists N0 so that, for all N ≥ N0 , if ξ N is a perturbed

NPS as above and H is an interval of length bounded by N, then, for all t ≥ 0,
ξtN ≤ ZtH,1 on H,
where Z H,1 be the finite NPS on {0, 1}H with 1’s fixed at the endpoints of H ,
generated by the same Harris system as ξ N and starting from full occupancy.

P ROOF. It will not be true in general that for all H , of arbitrary size, and all
t that ZtH,1 ≥ ξtN on H ; however, it will be true if H is of small length. By our
assumption (∗∗), on β(n)/β(n + 1) we have for N large that
 
β(N − 1) 1
≥ 1+ .
β(N ) N
Therefore there exists N0 so that for all N ≥ N0 and all n ≤ N we have
 
β(n − 1) β(N − 1) 1
≥ ≥ 1+ .
β(n) β(N ) N
40 T. MOUNTFORD

To show our result, we must simply consider the flip rate for ξtN at a site x ∈ H ,
vacant for ξtN and check that if x is also vacant for ZtH,1 and ξtN ≤ Z H,1 on H ,
then the flip rate for Z H,1 exceeds that for ξ N . If site x is to the left of the rightmost
particle of ξ N , then this is immediate. Accordingly we suppose that x, is l sites to
the right of rtN , the rightmost particle of ξtN . Let the distance to the left of x to a
Z H,1 -occupied site be l  (necessarily at most l); to the right let the distance be r.
Then the flip rate at x for process ξ N is
 
1
1+ β(l),
N
while for Z H,1 it is
β(l  )β(r  ) β(l)β(r  ) β(l)β(l  + r  − 1)
≥ ≥ .
β(l  + r  ) β(l  + r  ) β(l  + r  )
Now l  + r  is at most N , so by attractiveness
β(l  + r  − 1) β(N − 1)
≥ .
β(l  + r  ) β(N )
By our choice of N0 and our assumption that N ≥ N0 ,
 
β(N − 1) 1
≥ 1+ .
β(N ) N
Therefore we have that, for all intervals H of length bounded above by N ,
ZtH,1 ≥ ξtN ∀t ≥ 0.

L EMMA 5.5. Let ξ N,A , ξ N,B be two perturbed NPSs in equilibrium

and generated by the same Harris system. Suppose that ξ0N,A is supported on
(−∞, 0] and ξ0N,B is supported on (−∞, −N 1/3 ]. Then, outside of probability
KN 53/20 / N k/40 ,
ξtN,A ≥ ξtN,B on [−N, ∞),
for t ∈ [N 3/5 , N 3/5 + N 3/2 ].

P ROOF. The proof follows that of Lemma 2.5 closely; a slight difference is
the lack of attractiveness, which requires some care. We denote the rightmost
particles at time t by rtN,A and rtN,B , respectively. Let G be the event that, for
all t ∈ [0, N 3/2 + N 3/5 ],
ξtN,A is N 3/20-good on [rtN,A − 3N, rtN,A ] and |rtN,A | ≤ N.
By Lemma 2.2 (applied to perturbed NPSs) and Lemma 5.3 we have P (G) ≥
1 − KN N 3/2N 3/20 /N k/20 .
 CRITICAL NEAREST PARTICLE SYSTEMS 41

For any interval H , Z H,1 will be (as before) the finite NPS on {0, 1}H , starting
from full occupancy and generated from the Harris system for η.
We now proceed as in Proposition 3.2. We choose ≤ KN 1−3/20 equal disjoint
intervals I1 , I2 , . . . , IR of length 2N 3/20 + 1 covering [−3N, −N 1/3 /3] and
then we choose intervals J0 , J1 , . . . , JR (of equal length) as in Lemma 2.5 or
Proposition 3.2. As in the proofs 
of these

lemmas we argue that if (as will be the
case on event G), for all H ∈ Ii ∪ Ji ,
ξ0N,A ◦ θH |[−N 3/20 ,N 3/20 ] is good
(where θH is the shift that centers H on the origin), then outside of probability
bounded by KN 1−3/20/N k/40 we have that, for all sites x in the middle four-fifths
of H ,
ξNN,A H,1
3/5 (x) = ZN 3/5 (x).

By Lemma 5.4 this equality for all H implies that, outside of probability
KN N 3/2 N 3/20/N k/40 + KN 1−3/20/N k/40 ,
ξNN,A N,B
3/5 (x) ≥ ξN 3/5 (x) ∀x ∈ [−3N, −N 1/3 /3].

By Lemma 5.3 (this time applied to the motion of r N,B ), we have that, outside a
further set of probability bounded by KN 3/5 /N 3(k−1)/10 ,
ξNN,A N,B
3/5 (x) ≥ ξN 3/5 (x) ∀x ∈ [−3N, ∞].

To summarize, outside of a set of probability KN N 3/5/N (k−1)/20 , the above

inequality holds. We now wish to consider times t ∈ (N 3/5 , N 3/2 + N 3/5 ].
We argue as in Lemma 1.3. Define Dt = sup{x : ξtN,B (x) > ξtN,A (x)} ∧ −2N .
(So we have shown that, with high probability as N tends to infinity, D0 = −2N .)
We examine how D can jump forward at time t.
N,B
We will treat the case where Dt− < rt− : D can only jump forward to a site x
N,A
at time t if ξ N,B has a birth at x at time t while ξt− (x) = ξtN,A (x) = 0. Given
that the flip rates for perturbed NPSs are the same as NPSs except at sites to the
right of the rightmost site, we have that D cannot jump forward to x if there
N,B N,B
exists Dt− < y < x < z with ξt− (y) = ξt− (z) = 1 [and so, by definition of
N,A N,A
D, ξt− (y) = ξt− (z) = 1]. This leaves only two possibilities:
N,B
1. Dt− < rt− < x;
N,B N,B
2. Dt− < x < rt− and ξt− ≡ 0 on (Dt− , x].
We will show that case 1 is not possible if event G occurs and that, while case 2
can occur if event G also occurs, case 2 is restricted to x < Dt− + N 1/20 .
N,A
We first treat case 1. We can assume that ξt− (x) = 0. There are two
N,A
possibilities. If rt− < x, then the flip rate at x at t− for ξ N,A exceeds that for
42 T. MOUNTFORD

N,A
ξ N,B and D jumping to site x at time t is impossible. If x < rt− and B occurs,
N,A N,A
then at time t−, there exist y, z so that x ∈ (y, z) and ξt− (y) = ξt− (z) = 1
N,B
and z − y < N 1/20 . Since Dt− < rt− we have, from the definition of D, that
N,A N,B N,B
ξt− (rt− ) = 1, so we can take y to be at least rt− . Thus the flip rate at time t−
at x for ξ N,A is at least
N,B  
β(x − y)β(z − x) β(x − rt− )β(N 1/20 − 1) 1 N,B
≥ ≥ 1+ β(x − rt− )
β(z − y) β(N 1/20 ) N
for N large. So again a jump of D to x is impossible.
For case 2, by a similar argument, we have that if B occurs, then for x >
Dt− + N 1/20 the flip rate for ξ N,A must exceed that for ξ N,B .
Thus we have that if event G occurs, then Dt can only jump forward by at
most N 1/20 . Its jump rate forward is bounded uniformly by V so we have, for all
t ∈ [N 3/5 , N 3/2 + N 3/5 ],
ξtN,A ≥ ξtN,B on [−N, ∞)
outside of probability

  
KN N 3/5 N N 19/20
+P inf rtN,B <− +P X>
N (k−1)/20 t≤N 3/2 +N 3/5 2 3
for X a Poisson random variable with mean V N 3/5 . We are done, by Lemma 0.2.

P ROPOSITION 5.1. Let ξ N be a perturbed NPS in equilibrium so that ξ0N is

supported on (−∞, 0]. Let rtN be the site of the rightmost site at time t. Then, as
N → ∞,

N
r N 2t
⇒ {Xt }0≤t≤1 ,
N 0≤t≤1
where Xt is the diffusion starting at 0 at time 0,
Xt = Wt + µt

for W a standard Brownian motion and µ = n nβ(n).

P ROOF. First we get a lower estimate on rtN for 0 ≤ t ≤ N 2 .

Define a process ξtN with rightmost particle rtN as follows:
(a) for 0 ≤ t < N 3/2 , ξtN = ξtN (and so rtN = rtN );
(b) for i ≥ 1 and iN 3/2 ≤ t < (i + 1)N 3/2 , ξtN = ξtN,i , where ξtN,i , t ≥ iN 3/2 ,
is a perturbed NPS in equilibrium generated by the same Harris system as ξ N
so that ξNN,i3/2 is supported on (−∞, riN 3/2 − N
N 1/3 ] but is otherwise independent
N
of ξiN 3/2 .
 CRITICAL NEAREST PARTICLE SYSTEMS 43

We apply Lemma 5.5 repeatedly at times iN 3/2 . This and the large deviations
bounds for the behavior of rightmost particles provided by Lemma 5.3 entail that,
with probability tending to 1 as N tends to infinity,
rtN ≤ rtN ∀0 ≤ t ≤ N 2 .
We similarly introduce ξTN , rtN :
(a) for 0 ≤ t ≤ N 3/2 , ξtN = ξtN (and so rtN = rtN );
(b) for i ≥ 1 and iN 3/2 ≤ t < (i + 1)N 3/2 , ξtN = ξtN,i , where ξtN,i , t ≥ iN 3/2 ,
is a perturbed NPS in equilibrium generated by the same Harris system as ξ N
so that ξNN,i
3/2 is supported on (−∞, riN 3/2 + N
N 1/3 ] but is otherwise independent
N
of ξiN 3/2 .

Then analogously one has that with probability tending to 1 as N tends to

infinity
rtN ≥ rtN ∀0 ≤ t ≤ N 2 .
However, by Lemma 5.2 one has that
 1 N
XtN = r 3/2 − riN 3/2 ⇒ Xt .
N
√ N (i+1)N −
i≤t N

Now since Lemma 5.3 easily implies that

|rtN − riN
N
3/2 |
sup sup →0
i t∈[iN 3/2 ,(i+1)N 3/2 ) N

in probability we have that YtN , 0 ≤ t ≤ 1, converges in distribution to Xt where,

for s ∈ [iN 3/2 , (i + 1)N 3/2 ),
 1
2 = − rjN + (rsN − riN
N N N
Ys/N r(j +1)N 3/2 − N 3/2 3/2 ).
j <i
N

Similarly process ZtN , 0 ≤ t ≤ 1, tends in distribution to Xt where, for s ∈ [iN 3/2 ,

(i + 1)N 3/2 ),
 1
2 = − rjN + (rsN − riN
N N N
Zs/N r(j +1)N 3/2 − N 3/2 3/2 ).
j <i
N

But we have that, with probability tending to 1 as N tends to infinity, for all
0 ≤ t ≤ 1,
√ √
YtN − N N 1/3 /N ≤ crNN2 t /N ≤ ZtN + N N 1/3 /N.
The result follows.

From this result we obtain by simple coupling the following corollary.

44 T. MOUNTFORD

C OROLLARY 5.2. Let N η0 be a NPS with initial distribution equal to that of

renewal measure conditioned to have N η0 (0) = N η0 (N ) = 1− N η0 (i) ∀0 < i < N .
Then there exist processes N,L ξt , N,R ξ so that the following hold:
(i) N,L ξt is a perturbed NPS in equilibrium initially supported on (−∞, 0]
with rightmost particle HN t at time t;
(ii) N,R ξt is a perturbed NPS in equilibrium initially supported on [N, ∞) with
leftmost particle rtN at time t;
(iii) for
τ N = inf{iN 3/2 : |riN
N
3/2 − N | ≥ N/2} ∧ inf{iN
3/2
: |liN
N
3/2 | ≥ N/2} ∧ N
2

 lN rNN2 t∧τ N 
N 2 t∧τ N
, ⇒ (Xt∧τ
1 2
, Xt∧τ ),
N N
where X01 = 0 = 1 − X02 and Xt1 = Wt1 + µt and Xt2 = Wt2 − µt for W i
independent standard Brownian motions; τ = inf{t : |Xti − X0i | ≥ 1/2 for i = 1
or 2} ∧ 1.
With probability tending to 1 as N tends to infinity N,L ξ and N,R ξ are
t
dominated by N ηt for all t ≤ τ N .

We do not give a complete proof of Theorem 5.1 since this would involve a
large amount of repetition and requires no really new ideas.

Sketch of Proof of Theorem 5.1. Introduce four processes run with H (the same
Harris system as generates η), ηN,i , i = 1, 2, 3, by taking η0N,i to be mutually
independent, with the following:
1. ηN,1 is a semiinfinite NPS in equilibrium with initial support on [−λN +
N 1/3 , ∞);
2. ηN,2 is a perturbed NPS in equilibrium, initially supported on (−∞, −N 1/3 ];
3. ηN,3 is a perturbed NPS in equilibrium, initially supported on [N + N 1/3 , ∞);
4. ηN,4 is a semiinfinite NPS in equilibrium with initial support on (−∞, (1 +
κ)N − N 1/3 ].
We define the following:
1. ltN is the position of the leftmost particle of ηN,1 ;
N,γ
2. lt is the position of the rightmost particle of ηN,2 ;
N,γ
3. rt is the position of the leftmost particle of ηN,3 ;
4. rtN is the position of the rightmost particle of ηN,4 ;
N,γ
5. τ N = inf{t : |ltN + λN | > N λ/3 or |rtN − (1 + κ)N | > κ/3 or |lt | > (λ/3 ∧
N,γ
1/3)N or |rt | > (κ/3 ∧ 1/3)N }.
 CRITICAL NEAREST PARTICLE SYSTEMS 45

We then show that with high probability as N → ∞, for all t ≤ τ N ,

1. ηt dominates ηtN,1 on (−∞, − 35 λN ];
2. ηt dominates ηtN,2 on [− 25 λN, ∞);
3. ηt dominates ηtN,3 on (−∞, (1 + 25 κ)N ];
4. ηt dominates ηtN,4 on [(1 + 35 κ)N, ∞).
N,γ
From this it follows easily via Lemma 2.2 that ηt is N v -good on [ltN , lt ],
N,γ N
[rt , rt ] for all t ≤ τ N with probability tending to 1 as N ⇒ ∞.
To achieve the necessary independence to apply Corollary 5.2 and Schinazi’s
theorem we introduce Harris systems H i , i = 1, 2, 3, 4, mutually independent and
independent of H , the Harris system generating η. We then consider the following
independent processes:
1. ηN,1, , a NPS with η0N,1, = η0N,1 and generated by the Harris system equal to
H on (−∞, −λN/2]; equal to H 1 elsewhere;
2. ηN,2, , a perturbed NPS with η0N,2, = η0N,2 and generated by the Harris system
equal to H on (−λN/2, N/2]; equal to H 2 elsewhere;
3. ηN,3, , a perturbed NPS with η0N,3, = η0N,3 and generated by the Harris system
equal to H on (N/2, N (1 + κ/2)]; equal to H 3 elsewhere;
4. ηN,4, , a NPS with η0N,4, = η0N,4 and generated by the Harris system equal to
H on (N (1 + κ/2), ∞); equal to H 4 elsewhere.
These processes are clearly independent but using Proposition 4.2 we can show
that with probability tending to 1 as N ⇒ ∞ we have that

ηtN,1, = ηN,1 on −∞, − 25 λN ,

ηtN,2, = ηN,2 on − 35 λN, ∞ ,
 
ηtN,3, = ηN,3 on −∞, 1 + 25 κ N ,

ηtN,4, = ηN,4 on 1 + 35 κ N, ∞ .
Thus we obtain the desired weak convergence for
 N
lN 2 t rN 2 t rNN2 t 
N,γ N,γ
l N 2t
, , , .
N N N N t≤τN

6. After three, regrettably, highly technical sections we are now ready to start
employing the accrued results toward a proof of Theorem 1. Recall from the first
remark following the statement of Theorem 1 in the Introduction that it will suffice
to show that, for fixed δ > 0 and fixed, increasing, cylinder f with f (0) = 0,
46 T. MOUNTFORD

E[f (ηt )] ≥ Ren(β), f  − δ for t large. We consider δ fixed now. We choose ε

small but positive (in a way that depends on δ and will be specified in Section 9.
Accordingly from now on ε will be considered fixed; N is a large number to be
fully specified in the following work. It will depend on the ε chosen. We consider
the evolution of finite intervals. By attractiveness we may (and shall) consider
that the process ηt begins  z
at time 22N and that at this time the occupied sites
are precisely equal to z I , a collection of disjoint intervals of length exactly
equal to 2N ; N is not yet precisely specified but will be very large. We will
detail processes Itz : t ≥ 22N with I0z = I z . The processes will be called upper
intervals, since at all times t the process ηt will be, roughly speaking, in upper
equilibrium on Itz , though by no means fully occupied. Given interval I z , we will
denote by xtz , ytz the left and right endpoints of I z at time t. Of key relevance
will be R = {tin = 22n + 3i22n /n100 : n ≥ N , i = 0, 1, . . . , n100 }, the collection of
“regeneration times" of the intervals (of course tnn100 = t0n+1 ). We denote by τnz the
first time in R that |I z | is greater than 2n . Our first control rules for the evolution
of intervals are as follows:

C ONDITION 1. Interval I z is killed at time τnz if

τnz − τn−1
z
< 22n /n.

C ONDITION 2. z
Interval I z is killed at time τn−1 + n22n if

τnz ≥ τn−1
z
+ n22n .

C ONDITION 3. Interval I z is killed at time t ∈ R if

t > τnz and |Itz | < 2n /n4 .

Conditions 1–3 ensure that up to a small power of n, all intervals alive at time
tin have size of order 2n . More specifically we have:

L EMMA 6.1. There exists a constant K not depending on n so that if interval

I is still alive at time tin = 22n + 3i22n /n100 , then 2n /(Kn5 ) ≤ |Itin | ≤ Kn2n .

We will also specify a small growth condition:

C ONDITION 4. Interval I z is killed at time tin ∈ R, i > 0, if for some s ∈

n
(ti−1 , tin ],
   
 z   
xs − xtzn  ≥ 4 × 2n /n48 or ysz − ytzn  ≥ 4 × 2n /n48
i−1 i−1

(we have yet to specify the dynamics of x z , y z ).

 CRITICAL NEAREST PARTICLE SYSTEMS 47

We also need a medium growth condition:

C ONDITION 5. Interval I z is killed at time tin ∈ R, i > 0, if for some tjn , j < i,
   
 z   
xt n − xtzn  ≥ 2n /n6 or ytzn − ytzn  ≥ 2n /n6 but tin − tjn < 22n /n17 .
j i j i

In addition to these conditions on the growth of the endpoints of I we will

require that within the interval I no bad exceptional configurations may evolve
that would make regeneration difficult.
The evolution of an interval I z in the time interval [ti−1 n , t n ] and in particular
i
n n
the change in its endpoints from time ti−1 to ti will depend on whether the
interval I z is attached at both sides, at a single side or at no side to other
intervals. Accordingly we must specify these terms. Initially all intervals are
isolated (meaning nonattached) at either side. Interval I z becomes attached to
interval I w at time tin ∈ R if at this time |I z ∩ I w | is at least 2n /n47 . At this point
the common points are divided equally between the two intervals so that they are
disjoint intervals. Thereafter they will remain attached until one or both intervals
die. They will also remain disjoint. Given the size consequences of the growth
constraints, it follows that an interval can be attached to at most one interval at
either side.
n
We first suppose that at time ti−1 interval I z is isolated on both sides (implicitly
z n
we assume that I exists at time ti−1 ). Associated with I z at this time will
be a finite NPS ηz,n,i−1 which is generated by the Harris system and which is
dominated by η, our original process. The endpoints of the support of ηz,n,i−1
will be denoted by Xz,n,i−1 and Y z,n,i−1 , respectively. If I z is isolated then these
endpoints will coincide with x z , y z over time interval [ti−1 n
+ 2n , tin ] but not for I z
attached.
We now define our corresponding process ηtz,n,i defined for t ≥ ti−1 n + 2n .
z,n,i,l z,n,i,r
Introduce semiinfinite equilibrium processes γt , γt , t ≥ ti−1
n
, as follows:
n
1. at time ti−1 , γ z,n,i,l , γ z,n,i,r are independent of η and each other given that they
are supported on [xtzn + 22n/3 , ∞) and (∞, ytzn − 22n/3], respectively;
i−1 i−1
2. the processes evolve according to H , the Harris system for η.

Process ηtz,n,i t ≥ ti−1

n
+ 2n is the finite equilibrium NPS generated by the two
semiinfinite NPSs over the time interval [ti−1 n n
, ti−1 + 2n ].
Our interval Iz will have boundaries given by those of ηz,n,i on the time interval
[ti−1 + 2n , tin ]. On time interval [ti−1
n n n
, ti−1 + 2n ] it will have boundary given by the
rightmost particle of γ z,n,i,r and the leftmost particle of γ z,n,i,l . Accordingly our
intervals will die if they violate the growth stipulations given in Conditions 1–5.
48 T. MOUNTFORD

In addition we kill our interval I z :

n + 2n if ηz,n,i−1 does not dominate ηz,n,i ;
(A) at time ti−1
(B) at time ti if, for some u ∈ [xtzn + 2n/4 , ytzn − 2n/4 ], ηz,n,i ◦ θu |[−2n/4 ,2n/4] is bad;
n
i i
(C) at time tin if the conditional probability that a 2n/12 gap appears within 3ηz,n,i
or that the boundaries change by 22n/3/4 during time interval [tin , tin + 2n ] is
less than 2−nk/48 ;
(D) at time tin if the boundaries of ηz,n,i−1 have changed by 2n /n48 in time interval
[ti−1
n
+ 2n , tin ].
These conditions will be called the regeneration conditions.
n , I z is attached to another interval.
We now detail the differences if, at time ti−1
In this case the difference is that our associated finite process ηz,n,i will not define
the boundaries of I z on the time interval [ti−1
n
+ 2n , tin ]. In this case as before ηz,n,i
is generated by two semiinfinite processes γtz,n,i,l , γtz,n,i,r , t ≥ ti−1
n
; the difference
is that they are supported on (−∞, Y z,n,i−1 −2 2n/3 ] and on [X z,n,i−1 + 22n/3 , ∞),
z,n,i−1 z z
where Y is equal to yt n if I is not attached to the right and equal
i−1
to ytzn + 122n /n48 if I z is attached to the right at time ti−1
n
, and similarly
i−1
for Xz,n,i−1 . During time interval [ti−1 n
, tin ] the boundary between two attached
intervals fluctuates as an independent symmetric simple random walk of rate 1.
The natural analogues of the extinction rules (A)–(D) are enforced in this case:
Thus for (B) we consider sites u within the interval of ηz,n,i but 2n/4 away from
the boundary. It should be noted that we kill interval I z either because of extreme
fluctuations in its boundary or because of extreme fluctuations in the boundary
of ηz,n,i . For (A) we now kill our interval if ∪ηu,n,i−1 does not dominate ηz,n,i ,
where the union is taken over u equal to z or such that I u is attached to I z .
The next order of business is to get bounds on the various probabilities of an
n
interval being killed in interval (ti−1 , tin ] because of stipulations (A)–(D).

P ROPOSITION 6.1. Conditional on I z being alive at time ti−1 n , the chance

n n
that the interval is killed in interval (ti−1 , ti ] due to failing stipulations (A)–(D) is
bounded by K213n/12/2−nk/48 .

P ROOF. All cases are essentially the same so we will simply treat the case
n
that at time ti−1 interval I z is isolated on the left and attached on the right to
interval I w . First we consider (A), the event that ηz,n,i−1 ∪ ηw,n,i−1 does not
n + 2n . By Lemma 4.2 and Corollary 4.2 this probability
dominate ηz,n,i at time ti−1
is dominated by K/2 nk/48 and no further discussion is required.
We now deal with (B). By Corollary 4.2 this is bounded by K2n 2−nk/24 .
For (C) we note, by Lemma 2.2 and Proposition 3.1, that the chance that, for a
right one-sided process in equilibrium, in time interval [0, 2n ] either a 2n/12 gap
 CRITICAL NEAREST PARTICLE SYSTEMS 49

appears within Kn2 2n of the rightmost particle or the rightmost particle fluctuates
by more than n2 2n/2 is bounded by n2 2n 2n 2−kn/24 . Thus the chance that the
conditional probability of this at time 0 is at least K213n/122−nk/48 is bounded
by K213n/122−nk/48 . By Lemma 4.4 this bound extends to cover the chance of
corresponding boundary fluctuations or gap appearances for ηz,n,i over the time
interval in question. We similarly treat (D).

So far in this section we have listed our rules for the evolution of an interval
in the upper regime and bounded the probability of a regeneration failing. In the
remainder of the section we examine the chance of an interval dying because of
failing the growth stipulations given by Conditions 1, 2, 4 and 5. We also wish
to show that as the time or interval length becomes large most of the conditions
become “irrelevant” except for the condition that the interval is killed if its size
falls to below 2n /n4 after achieving length 2n . We also wish to show that, after
achieving length 2n , an interval is essentially as likely to achieve size 2n+1 as to
disappear before gaining this size.
Already we have good bounds on the failure of an interval to “regenerate”
from 22n + 3i22n /n100 to 22n + 3(i + 1)22n /n100 . Let these be called n-order
regenerations. By Proposition 6.1 the chance of such a regeneration failing is
bounded by K213n/122−nk/48 . Suppose that our interval I z achieves size 2n
at time τnz , after an r-order regeneration. By our size constraints on intervals
(Conditions 1–3) we must have that r ∈ [n − 6 log(n), n + 2 log(n)]. Now we
define stopping times Sn0 , Sn1 , . . . by taking Sn0 to be τnz and Sni = inf{t > Sni−1 : t
is a regeneration time and t − Sni−1 ≥ 22n }; it is easily checked via Corollary 3.1
that

z
P τn+1 or τDz ∈ (Sni−1 , Sni ] | τDz , τn+1
z
> Sni−1 > d > 0,
where d does not depend on n and τDz denotes the death time of I z .
This and the strong Markov property immediately yield:

L EMMA 6.2. Let τiz be as above. Then

P (n22n /2 < τn+1
z
− τnz < ∞ | Fτnz ) ≤ cn ,
for some c < 1 not depending on n.

We now consider the probability that after τnz the interval I z is killed because it
z
grows too quickly, that is, because τn+1 ≤ τnz + 22n /n. For simplicity only, we will
suppose that [τn , τn + 2 /n] ∈ [2 , 22(m+1) ] for some m. As already noted, τnz
z z 2n 2m

must be less than 3n2n . Therefore |Iτzz | < 2n + 7 × 2n /n47 . For τn+1
z
− τnz ≤ 22n /n
n
we must have that either
2n+1 − (2n + (7 × 2n )/n47 )
xτzz − xτzz ≥
n+1 n 2
50 T. MOUNTFORD

or
2n+1 − (2n + (7 × 2n )/n47 )
yτzz − yτzz ≥ ,
n+1 n 2
or both. (Recall x z and y z denote respectively the left and right endpoints of I z .)
So, by symmetry, we have

z

v
P (τn+1 − τnz ≤ 2 /n | F ) ≤ 2P
2n
τnz sup (xtzm − xtzm )
2n 2m j +1 j
v : τnz <tvm <τnz + 2 n + 32100 j : tjm =τnz
m


2n+1 − (2n + 7 × 2n /n47 )
≥ .
2

Now given Ftjm , (xtzm − xtzm ) is either stochastically less than r3×22m /m100 for a
j +1 j
one-sided equilibrium NPS supported on (−∞, 0] at time 0 or stochastically less
than S3×22m /m100 for S a rate-1 symmetric nearest neighbor random walk. Thus we
have that
z
P (τn+1 − τnz ≤ 22n /n)
 

v
j 2n+1 − (2n + (7 × 2n )/n47 )
≤ 2P sup r22m /m100 ≥
22n /n 4
1≤v≤ +1 j =1
22m /m100

 
2n+1 − (2n + (7 × 2n )/n47 )
+P sup Ss ≥ ,
s≤22n /n+22m /m100 4

for S a rate-1 symmetric nearest neighbor random walk and r i independent copies
of r. Then it is easily seen by arguments of Proposition 3.1 that this probability is
bounded by (1/2)n for n large. We have proven:

L EMMA 6.3. Let τnz be as above. Then

z
P (τn+1 − τnz ≤ 22n /n | Fτnz ) ≤ C(1/2)n .

We similarly prove:

L EMMA 6.4. Given τDz > 22n , the probability that for some i, j ∈ 0, 1, . . . ,
n100 , j − i < n81 , the rightmost edge of I2z2n +3i22n /n100 is less than the rightmost
edge I22n +3j 22n /n100 − 2n /3n9 is bounded by C(1/2)n .
 CRITICAL NEAREST PARTICLE SYSTEMS 51

Now we consider the probability that an interval attains length 2n+1 given that
it attains length 2n . By our growth conditions τnz ∈ [22n /n, 2 × 22n n]. Suppose
τnz = tim . Then we must have n − log2 (4n)/2 ≤ m ≤ n + log2 (n)/2. By our
growth condition, we must have 2n ≤ |Iτzz | ≤ 2n + 4 × 2m /m48 ≤ 2n + 2n /n47 .
n
Since the overshoot of 2n is bounded by 2n /n47  2n /n4 by growth stipulation
of Condition 4 and Lemma 6.1, we immediately have that the probability that
z
τn+1 < ∞ given that τnz < ∞ is less than 1/2. We wish to show that it is close
to 1/2.
We couple our interval with an evolving interval which has the same increments
as our interval when our interval is not killed but whose interval increments will
be independently generated after our interval is killed and where the increments
for an r-order regeneration are conditioned to be less than 4 × 2r /r 48 . Our interval
will start with the same length as I z at time τnz . Then after N regenerations the
comparison interval will have length

N
SN = |I |τnz + Xi − di ,
i=1

where Xi are independent mean-zero random variables bounded by 22n /n46 . The
di are equal to 2 × 22m/3 when i corresponds to an m-order regeneration and are
thus bounded by 4n22n/3 for i ≤ n102 . Let τ be the first time our comparison
interval has value greater than or equal to 2n+1 or less than 2n /n4 .

L EMMA 6.5. P (Sτ ≥ 2n+1 ) ≥ 1/2 − 2/n4 .

This result and the bounds on our real interval being killed before either
attaining length 2n+1 or having its length decline to below 2n /n4 immediately
yield the following corollary.
z
C OROLLARY 6.1. For an interval I z , P (τn+1 < ∞ | τnz < ∞) ≥ 1/2 − K/n4
for some K not depending on n. There exists K ∈ (0, ∞) so that, for n ≥ N and
all z,
1 N−n
2 ≤ P (τnz < ∞) ≤ K2N−n .
K
C OROLLARY 6.2. There exists K ∈ (0, ∞) so that, for n ≥ N and all
z, 1/K ≤ 2n−N P (I2z2n is alive) ≤ K.

P ROOF. It is easy to check (via Corollary 3.1 and bounds on killing

probabilities given in Lemmas 6.2–6.5 and Proposition 6.1) that, for some strictly
positive d not depending on n,
  
P {τnz < ∞} ∩ I z does not die in [τnz , τnz + 22n ] ≥ d2N−n .
52 T. MOUNTFORD

So the lower bound follows easily. On the other hand,

∞

P (I z is alive at time 22n ) = P (τmz ≤ 22n < τm+1
z
∧ τDz )
m=N


n−C
≤ P (τmz ≤ 22n < τm+1
z
∧ τDz ) + K2C 2N−n
m=N

for C fixed but large, by Corollary 6.1. Now n−C z
m=N P (τm ≤ 2 < τm+1 ∧ τD )
z 2n z
n−C z
is less than P ( m=N (τm+1 ∧ τDz − τmz )+ ≥ 22n ). In turn this is bounded by
n−C−N
i=0
z
P ((τn−C−i+1 ∧τDz −τn−C−i
z
)+ ≥ 22n 2−i−1 ). However, by the arguments
giving Lemma 6.2 we have that there exists a d < 1 not depending on n or C so
 2C i
that this is bounded by i=0 K2N−n+i d 2 2 ≤ 2N−n if C were fixed sufficiently
large.


P ROOF OF L EMMA 6.5. Consider the martingale MN = |I |τnI + N i=1 Xi =
N
SN + i=1 di for 1 ≤ N ≤ n . Then for all N, E[MN∧τ ] = E[M0 ] ≥ 2n . On
102

{τ > N }, MN∧τ ≤ 2n+1 + N 4n22n/3 < 2n+1 (1 + 1/n6 ). On {τ ≤ N, Sτ ≥ 2n+1 },

we have similarly that MN∧τ ≤ 2n+1 + 2n /n46 + N 4n22n/3 ≤ 2n+1 (1 + 1/n6 ),
while on {τ ≥ N, Sτ ≤ 2n /n4 }, we have that MN∧τ ≤ 2n /n4 + N 4n22n/3 ≤
2n (1/n4 + 1/n6 ). Thus we have

2n ≤ M0 ≤ E[MN∧τ ]
  
1
≤ P (τ ≤ N, Sτ ≥ 2n+1 ] 2n+1 1 +
n6
     
1 1 1
+ P (τ ≤ N, Sτ ≤ 2 /n ] 2
n 4 n
4
+ 6 + P (τ > N ) 2 n+1
1+ 6 .
n n n
However, by the argument of Lemma 6.2, we have P (τ > N ) ≤ cn . The result now
follows from routine manipulations.

7. In this section we consider the effect of our killing rules and regeneration
corrections on the density of points in ∪z I z at time t. The object is to show that
this density is essentially the same as the initial density; that is, the picture is close
to our “voter model” picture discussed in the Introduction.
n
As we move from time ti−1 to time tin the size of Iz (assuming that it is alive
n
at ti−1 ) changes to

0 if the interval dies in [ti−1

n
, tin ]
 CRITICAL NEAREST PARTICLE SYSTEMS 53

and otherwise to

|Itzn | + ξ z,n,i,l + ξ z,n,i,r − 22n/3 II zn is left isolated + II zn is right isolated

i−1 ti−1 ti−1

+ C z,n,i,l + C z,n,i,r ,

where, for event A, IA represents the usual indicator function (we apologize for
the dual use of I ) and the ξ are either edge fluctuations of equilibrium one-sided
NPSs or fluctuations of a simple symmetric random walk (provided both are less
than 4 × 2n /n48 ).
Here C z,n,i,l represents the loss of sites to I z if I z becomes attached at the left
at time tin ; similarly for C z,n,i,r .
Thus we can write |Itzn | as a martingale minus a decreasing process which (of
i
course) stops decreasing once I z dies (that is to say |I z | becomes zero). For the
rest of this section we try to bound this decreasing part. Since it is decreasing we
consider its expectation at ∞; this is equal to

 n −1
100

(A) 22n/3 II zn is left isolated + II zn is right isolated II zn is alive

ti−1 ti−1 ti
n=N i=0
100 −1
 n
(B) + (C z,n,i,l + C z,n,i,r )II zn is alive
ti
n=N i=0
100 −1
 n
(C) + L(n, i)II z dies at time tin ,
n=N i=0

where L(n, i) is the loss resulting from the death of interval I z at time tin . Note
that positive and negative large deviations of interval edges can be thought of as
cancelling each other out so that L(n, i) ≤ Kn2n + 8 × 2n /n48 ≤ Kn2n .

We first bound (A). A simple bound for (A) is n≥N 22n/3n100 I{I z is alive} . By
22n
Corollary 6.2 this is bounded in expectation by
 
K n100 22n/3 2N−n = K2N n100 2−n/3 .
n≥N n≥N

If N was chosen sufficiently large, then this is less than ε2N /10.
For (B) note first that by the growth stipulation of Condition 4 we have that
C z,n,i,l is less than 2n /n47 and likewise for C z,n,i,r . Second, it should be noted
that the growth stipulation of Condition 5 ensures that if |j − j  | ≤ n83 , then
  100
C z,n,j,l and C z,n,j ,l cannot both be nonzero. Thus nj =0 C z,n,j,l ≤ n17 2n /n47 =

2n /n30 . Therefore E[(B)] ≤ K n≥N 2n /n30 2N−n . Again provided N was fixed
sufficiently high this is bounded by ε2N /10.
54 T. MOUNTFORD

It remains to treat (C). As has already been noted, L(n, i) can be at most Kn2n .
So (C) is bounded by

K n2n I{I z dies due to failure to regenerate at tin for some i}
n≥N

+K n2n I{I z fails Condition 1, 2, 4 and 5 for t∈R∩[22n ,22(n+1) ]}
n≥N

 2n
+K I{I z fails Condition 3 after τnz before τn+1
z
}.
n≥N
n4

By Proposition 6.1 and Corollary 6.2 the expectation of the first term is bounded
by
 ε2N
K n2n 2N−n 213n/122−nk/48 ≤
n≥N
10

for N sufficiently large. Lemmas 6.1–6.3 and Corollary 6.2 similarly bound
the expectation of the second term. Lemma 6.4 bounds the third term by
 n 4 N−n ≤ ε2N /10 for N large.
n≥N (2 /n )2
Let us call the intervals I z the original upper equilibrium intervals in contrast
to the new or created upper equilibrium intervals that we will describe in the next
section.
We conclude from the above bounds:

P ROPOSITION 7.1. If ρt , t ≥ 22N , is the density of sites in an original upper

equilibrium interval at time t, then for all such t we have ρt ≥ (1 − ε)ρ22N .

We equally could say that the expectation of site loss for any original interval is
bounded by 2N ε.

8. In this section we introduce our mechanism for creating upper equilibrium

intervals and then prove that the intervals in the upper regime become dominant as
t → ∞. Recall we have fixed ε very small but positive.
We need to show that the density of vacant intervals goes to zero. We consider
various types of vacant interval. First there are big vacant intervals, where by
big, at time tin , we will mean large relative to 2n . Lemma 8.1 shows this density
is small. Then there are vacant intervals corresponding to growths of intervals
originally vacant at time 22n of “reasonable” size at time tin (again relative to
size 2n ). Proposition 8.1 below will show that if the density of such points in
such intervals is significant at time 22n then the density of “upper regime” points
must significantly increase over interval (22n , 22(n+1) ]. Accordingly this density
must decrease to zero. The density of points in vacant intervals created by the
 CRITICAL NEAREST PARTICLE SYSTEMS 55

deaths of “upper regime” intervals (including those created, as will be detailed) is

essentially bounded by Proposition 7.1. This will leave us with showing that the
density of small intervals tends to zero. This is done in Theorem 8.1.
By a simple but tedious 2-moment argument, we have:

L EMMA 8.1. Given ε there is an M, greater than or equal to 1/ε, so that for
all n ≥ N the chance that there is no I z alive at time 22n , of size 2n or more at this
time and contained in [0, M2n ], is less than ε2 /6.

S KETCH OF P ROOF. We associate with each Iz two Gaussian processes Wtz,x

z,y z,y
and Wt , t ∈ R, the set of regeneration times. We let W2z,x z z
2N = x22N , W22N = y22N ,
z,y z,y
and so that, for tin ∈ R, Wtz,x z,x
n − Wt n , Wt n − Wt n are independent Gaussian
i i−1 i i−1
random variables so that (see Proposition 3.5)
  2n
 
z,x z,x
 Wt n − Wt n − xtzn − xtzn  < K2−n/60 50 ,
i i−1 i i−1 n
outside of probability 2n/60 (3.2−n/30 + (K245n/14/24nk/105)1/2 ).
We may also choose these Gaussian increments so that we have independence
for intervals that are separated by at least 2n /n47 at time tin .
z,y z,y
We kill W z,x , W z,y if Wt − Wtz,x < 2n /n3 for t > τ n = inf{t ∈ R : Wt −
z,x
Wt ≥ 2 }. n

For z such that I2z2N is contained in ((j − 1/2)2n , (j + 1/2)2n ) for even j , we
let event A(z) be that
W z,x , W z,y are still alive at time 22n
and
z,y
W2z,x
2n ≤ (j − 1/2)2 ,
n
W22n ≥ (j + 1/2)2n
z,y 
for all t (∈ R) ≤ 22n , Wtz,x , Wt ∈ (j − 2/3)2n , (j + 2/3)2n .

Then we have P (A(n)) ≥ c2N−n and for such j we have P (A(z) ∩ A(z )) ≤
P (A(z))P (A(z)). So given ε, N we can choose M = M(ε, N ) so that
 
ε2
P A(z) ≥ 1 − .
12
z : I z2N ∈[0,M2n ]
2

We then note that if we had fixed N sufficiently large, then

 
! ε2
P A(z) B(z) ≤ ,
12
z : I z2N ∈[0,M2n ] z : I z2N ∈[0,M2n ]
2 2
56 T. MOUNTFORD

where B(z) is the event


at time 22n , I z is alive, of length at least 2n and
 
contained in the interval (j − 2/3)2n , (j + 2/3)2n .

This has the following immediate corollary:

C OROLLARY 8.1. Let M be as in Lemma 8.1. For all n the chance that the
origin is contained in a vacant interval of length 2M22n or greater is bounded
by ε2 /3.

The importance of Corollary 8.1 for this paper is that it means we do not have
to worry about very big vacant intervals in showing that for large time the density
of points in vacant intervals is small.
Before describing how “upper equilibrium” intervals are created we need to
develop a tagging system for vacant intervals and to introduce a distinction
between original vacant intervals and vacant intervals that arise from the killing of
“upper equilibrium” intervals. We have described the system of (possibly slightly
overlapping) intervals in the upper regime, their killing and creation. We now
introduce a system of labeling for the vacant intervals. We will divide these
intervals into those that correspond to an initial vacant interval and those that
correspond to a vacant interval created by the killing of an upper regime interval.
Initially we take as our intervals those natural maximal occupied intervals of
our “initial” configuration η22N . We will stipulate that immediately after every
regeneration time t all abutting vacant intervals coalesce into one large vacant
interval which is taken to correspond to one of the previous intervals, the other
intervals being deemed to be killed off. The “surviving interval is chosen in
proportion to the respective intervals. More concretely, if at regeneration time t
j
a maximal collection of vacant abutting intervals is Vt1 , Vt2 , . . . , Vt , then we
randomly (and independently of all other interval choices and of the Harris system)
choose k equal to i ∈ {1, 2, . . . , j } with probability
|Vti |
 .
|Vth |
Then at t+ all intervals
 l
V h for h = k are deemed to die, while interval V k expands
from Vt to Vt+ = l Vt . This step is called consolidation.
k k

We now detail the evolution of a null interval J from tin + = 22n + 3i22n /n100
n
to ti+1 = 22n + 3(i + 1)22n /n100 . Necessarily at time tin + the interval is abutted
by two upper intervals. If during the coming interval these two intervals do not
n
die, then at time ti+1 the vacant interval endpoints are defined via the endpoints of
these two upper intervals. However, if, say, the right abutting interval dies during
this interval then the new vacant interval is given by the neighbor of the rightmost
 CRITICAL NEAREST PARTICLE SYSTEMS 57

site of the surviving left interval and the initial leftmost site of the dead upper
interval at time tin minus 4 × 2n /n48 . If this interval is null, then this interval is
deemed to have died. We have similar rules for when the left upper interval dies or
when both surrounding intervals die. As a consequence for vacant intervals that are
alive at time tin death of surrounding upper intervals is never a good thing. On the
other hand this “redistribution” of sites from preexisting vacant intervals to newly
created vacant intervals only constitutes an additional small fraction of the number
of sites given to the new interval by the destroyed upper interval.
We now detail our mechanism for the creation of intervals in upper equilibrium.
Suppose that, at time 22n , V k is a vacant interval corresponding to an original
vacant interval of size 2n in the interval [ε2n , 2M2n ]. Match it with the first
original interval of size 2n to the right and with the first original interval of size 2n
to the left. Denote these intervals by I z(k,l) and I z(k,r) respectively. If either such
interval is further than M2n away from V k at this time we say that the match is
lost. If not we say the match holds. It is easy to see that with probability at least
p(ε, M) > 0 we have that at time 222n we have that I z(k,l) , I z(k,r) abut V k and that
all three intervals have size at least 2n and at V k has size at most 22n .
Now, given this situation, consider (in addition to the regular edge fluctuations
N,γ N,γ
of the occupied intervals), the subordinate processes lt , rt so that (as
described in Theorem 5.1) if
N,γ N,γ
rtN 2 − ltN 2
XtN = ,
N
then “locally” XtN → Xt , where
√
X(t) = 2W (t) − 2µt
and W is a Brownian motion.
It follows, using Theorem 5.1, that we can find some g, d, q > 0 (not depending
on n) so that with probability at least q we have at time 222n + d22n (assumed to
be tjn ) that there exists x k,j,n such that:
z(k,r) z(k,r)
1. xt n − 2g2n < x k,j,n < xt n − g2n ;
j j
z(k,l)
2. yt n + 2g2n < x k,j,n ;
j

3. if we consider η restricted to interval [x k,j,n , xtz(k,r)

n ] as an interval attached to
j
I z(k,r) to the right and unattached to the left, then the conditional probability
of successful regeneration for this interval and of interval I z(k,r) at time
222n + d22n exceeds the lower bound of Proposition 6.1.
z(k,r)
If these conditions are satisfied, then the interval [x k,j,n , xt n ] has been
j
created at time tjn . Thereafter the system evolves as if the newly created interval is
58 T. MOUNTFORD

an isolated interval. (So another vacant interval has been created between the new
interval and I z(k,r) .) If these conditions are not satisfied, we forget about creating
an upper equilibrium interval from V k and the intervals evolve without the extra
mechanism. We have shown:

P ROPOSITION 8.1. There exists h = h(ε, M) so that, given a vacant inter-

val V k matched at time 22n , with probability at least h a new upper equilibrium
interval is created during [22n , 22(n+1) ] of length at least g2n , where g is as chosen
above.

It is immediate that the bounds on original intervals being killed in Proposi-

tion 6.1 still hold. It is also clear that while this added mechanism for creating
new particles may result in additional sites of I z(k,r) being lost due to an extra
attachment of I z(k,r) the previous bounds on this loss still hold.
Proposition 7.1 can be applied to the created intervals (the slight difference
resulting from starting from initial length approximately c2n rather than exactly
2n notwithstanding). We thus have:

P ROPOSITION 8.2. The density of sites inside an interval in upper equilibrium

(whether original or created) at time tin is at least 1 − ε times
  
ρ22n + ρ(m) ,
tkm <tin

where ρ(m) is the density of sites in upper equilibrium sites created during
[22m , 22(m+1) ] at the time of creation.

Finally Proposition 8.1 yields the following:

P ROPOSITION 8.3. If at time 22n the density of sites inside original vacant
intervals of length between ε2n and 2M2n is at least f , then ρ(n) ≥ fgh/(2M).

T HEOREM 8.1. The density of sites in an upper invariant interval is eventually

more than 1 − 6ε.

P ROOF. From Proposition 7.1 (and then Proposition 8.2) it follows that the
density of sites in a created vacant interval is bounded by 2ε. By " Lemma 8.1 the
density of sites in an original vacant interval of size at least 2M tin is less than ε.
Propositions 8.2 and 8.3 force the density of sites at time tin having size in between
ε2n and 2M2n to go to zero. Thus it remains only" to consider the density of sites in
an original vacant intervals of size less than ε tin . To show this it suffices to show
that the density of distinct original vacant intervals is eventually less than 14 2−n at
time 22n .
 CRITICAL NEAREST PARTICLE SYSTEMS 59

Let the density of original intervals (not sites in original intervals) at time 22n
be d(n). It is easily seen that, provided ε was fixed sufficiently small (in a way
not depending on N ), a vacant interval alive at time 22n of length less than ε2n has
chance at least 34 of dying before time 22(n+1) . Also the density of original intervals
2 −n
of length at least ε2n is for large n less than 3M 2 by the above discussion. Thus
 
2 −n 1 2 −n
d(n + 1) ≤ d(n) − 2 · + 2 .
3M 4 3M
6 −n
Therefore if d(n) ≥ M 2 , then d(n + 1) ≤ d(n)(1 − 56 · 34 ) ≤ d(n) 38 . It follows
that 2n d(n) decreases with n by a factor of 2 · 38 = 34 , until d(n)2n ≤ M
6
. Thus we
12 −n
have eventually that d(n) ≤ M 2 . This completes the proof.

9. Proof of Theorem 1. All that remains is to show that the eventual

dominance of “upper equilibrium” intervals over vacant intervals translates into the
desired convergence in distribution. This part of the proof is close to the argument
for convergence employed in Mountford and Sweet (1998) but is given fully here
for completeness. As already noted, to show the theorem it will suffice to show
that E[f (ηt )] tends to Ren(β), f  for f an increasing cylinder function so that
f (0) is equal to zero. Without loss of generality we will assume that 'f '∞ ≤ 1.
Let its support be [−m, m].
Since for any such function f we have
lim sup E[f (ηt )] ≤ Ren(β), f ,
t→∞
it will suffice to show that, for fixed f as above and δ > 0,
E[f (ηt )] ≥ Ren(β), f  − δ.
Accordingly let us fix such f, δ. By Proposition 8.1 we may fix ε, N so that,
with these parameters defining our comparison process, the density of “upper
equilibrium” intervals exceeds 1 − δ/10 for all large times.
Let us fix t so large that if t ∈ [22v , 22(v+1) ] the density of “upper equilibrium”
2v
intervals at time t − 12 3 v2100 and subsequent times exceeds 1 − δ/10. Eventually
2n
of course we will let t go off to infinity. Choose tin so that tin + 3 n2100 ≤ t ≤
2n
tin + 2 × 3 n2100 . By slight abuse of notation let the two points in R following tin
be denoted by ti+1 n and t n , even though i may equal n100 or n100 − 1.
i+2
Consider the following events:
n
(A) At time tin , [−m, m] is within 9 2n8 of a vacant interval.
n
(B) At times ti+1 n
or ti+2 , [−m, m] intersects a vacant interval.
At this time all such intervals have length at least K2n /nn5 by Lemma 6.1. So the
density of points in an upper interval but within m + 9 2n8 of a vacant interval is
60 T. MOUNTFORD

m+9(2n /n8 )
bounded by K2n /n5
. Thus for t large the chance of event (A) above occurring
2
is bounded by 10 δ, and similarly for event (B).
Event (B) was introduced because, by Condition 4 introduced in Section 6, if
event (A) above does not occur but the upper interval in which the origin is within
at time tin dies at ti+1
n n
or ti+2 , then event (B) must occur.
Suppose that event A does not occur and that at time tin the origin is contained
n
within I z , at least 9 n28 from the boundary. Let (γT )t≥tin +2n be the finite NPS
generated by γ z,n,i,r , γ z,n,i,l over time interval [tin , tin + 2n ]. (Recall the definitions
in Section 6.)
As remarked above, if at time tin + 2n (and therefore for all subsequent times)
η does not dominate γ , then event (B) must occur. By Lemma 4.5 (and the growth
conditions to which I z is subject), if (A) does not occur (and n is sufficiently large)
then
E[f (γt )] ≥ Ren(β), f  − δ/10.
Putting these observations together we have for large t that
E[f (ηt )] ≥ E[f (ηt )IAc ∩B c ] − P (A) − P (B)
≥ E[f (γt )IAc ∩B c ] − P (A) − P (B)
≥ E[f (γt )] − 2P (A) − 2P (B),
which by our assumption on t being large is at least Ren(β), f  − δ.

REFERENCES
D URRETT, R. (1996). Probability: Theory and Examples. Duxbury, Belmont.
D IACONIS , P. and S TROOCK , D. (1991). Geometric bounds for eigenvalues of Markov chains. Ann.
Appl. Probab. 1 36–61.
G RIFFEATH , D. and L IGGETT, T. (1982). Critical phenomena for Spitzer’s reversible nearest-particle
systems. Ann. Probab. 10 881–895.
J ERRUM , M. and S INCLAIR , D. (1990). Approximate counting, uniform generation, and rapidly
mixing Markov chains. Inform. and Comput. 80 93–133.
L IGGETT, T. (1985). Interacting Particle Systems. Springer, New York.
L IGGETT, T. (1991). L2 rates of convergence for attractive reversible nearest particle systems: the
critical case. Ann. Probab. 19 935–959.
M OUNTFORD , T. (1995). Exponential convergence for attractive subcritical nearest particle systems.
Stochastic Process. Appl. 59 235–249.
M OUNTFORD , T. (1997). A complete convergence theorem for attractive reversible nearest particle
systems. Canad. J. Math. 49 321–337.
M OUNTFORD , T. and S WEET, T. (1998). Finite approximations to the critical reversible nearest
particle system. Ann. Probab. 26 1751–1780.
S CHINAZI , R. (1992). Brownian fluctuations of the edge for critical reversible nearest-particle
systems. Ann. Probab. 20 194–205.
S CHONMANN , R. (1994). Slow droplet-driven relaxation of stochastic Ising models in the vicinity
of the phase coexistence region. Comm. Math. Phys. 161 1–49.
 CRITICAL NEAREST PARTICLE SYSTEMS 61

S PITZER , F. (1976). Stochastic time evolution of one dimensional infinite particle systems. Bull.
Amer. Math. Soc. 13 880–890.
S WEET, T. (1997). One dimensional spin systems. Ph.D. dissertation, Univ. California, Los Angeles.

D EPARTEMENT DE M ATHEMATIQUES
E COLE P OLYTECHNIQUE F EDERALE DE L AUSANNE
E CUBLENS CH1015
S WITZERLAND
E- MAIL : [email protected]