Zeitschrift fhr

Z. Wahrscheinlichkeitstheorie verw. Gebiete Wahrscheinlichkeitstheorie

43, 3 9 - 4 8 (1978) und verwandte Gebiete
9 by Springer-Verlag 1978

A Note on Percolation

Lucio Russo
C.N.R., Gruppo Nazionale per la Fisica Matematica,
Istituto di Fisica Teorica detl'Universitg di Napoli,
Mostra d'Oltremare, Pad. 19, 1-80125 Napoli

Summary. An improvement of Harris' theorem on percolation is obtained; it

implies relations between critical points of matching graphs of the type of the
one stated by Essam and Sykes. As another consequence, it is proved that
the percolation probability, as a function of the probability of occupation of
a given site, is infinitely differentiable, except at most in the critical point.

1. Introduction

Percolation problems have been considerably studied in the literature (for an

extensive bibliography see references quoted in [1-5]). Nevertheless, even in the
case of Bernoulli measures, there are very few rigorous results; among them
Harris' theorem ([6, 7]) on non coexistence of infinite clusters o f " opposite type"
in the two-dimensional lattices plays a central role; it implies the relation Pc
+ p * > 1 between the critical probabilities of two "matching graphs". The same
relation was stated as an equality by Essam and Sykes ([8]) on the grounds of
very plausible arguments based on an analogy with the case of Bethe lattices
where exact results are known.
Harris proved his theorem by showing that if in a two-dimensional graph
there is a.s. an infinite cluster, say, of +, one can find an increasing sequence
{A,} of squares centered in a given site of the graph such that in A,+ I \ A , there
is a closed chain of + with probability greater than some constant not
depending on n. This implies that a.s. some of these chains actually occur,
forcing the clusters of opposite type to be finite.
In this note by using essentially the same techniques as in [6] we comple-
ment Harris' proof and show that one can control the size of the squares A~ by
proving that the finite clusters are forced to have also finite mean size. In this
way one gets a statement which can be inverted so that the inequality Pc + P* > 1
can be replaced by some equalities. We have not been able, however, to prove
the relation P ~ + P c * -- 1.

0044-3719/78/0043/0039/$02.00
40 L. Russo

In Section 2 we define the quantities of interest. The main statement is

proved in Section 3. In Section 4 we draw some consequences and prove that the
percolation probability is infinitely many times derivable except at most in the
critical point. Section 5 contains some final remarks.

2. Definitions

Two points in 7/2 which differ only by one unit in one coordinate are called
adjacent; they are called *adjacent if they are adjacent or such that both their
coordinates differ by one unit ~. A finite sequence (x~, ..., xn) of distinct points in
7/2 is called a (self-avoiding) chain if x i and xj are adjacent if and only if li-j]
=1 and a circuit if (x 2,...,x.) is a chain and xj is adjacent to x 2 and to x,;
*chains and *circuits are defined in an analogous way. A subset y ~ 7 / a is
connected [*connected] if, for all pairs x, y of points in Y,, there is a chain
[*chain] made up of points in Y, having x, y as terminal points. The boundary
[*boundary] of a given subset y ~ 7 / z is the set 0Y[~* Y] of all points in 7 / Z \ y
that are adjacent [*adjacent] to at least one point in I4.
Remark. The external boundary of a connected set is *connected and the
external *boundary of a *connected set is connected 2.
We consider the configuration space f2 = { - 1, 1}z2. If co e f2, the (+)clusters
[(+)*clusters] of co are the maximal connected [*connected] components of
co-~ (1); (-)clusters and (-)*clusters are defined in the same way. A (+)chain in
co is a chain included in co-t(1) 3. Two sets A, BeT/2 are (+)connected in co if
there is in o9 a (+)chain starting in A and ending in B.
If g is a translation invariant probability measure on f2, the #-probability
that a given site of the lattice belongs to a (+)cluster of size k is

pk(#)= ~, I~(C.[) (2.1)

I~l=k

where for any finite connected subset 7 of ;~2 containing the origin, C~+ is the set
of all configurations co such that 7 is a (+)cluster in co, and JTI means number of
sites in Y-
We put:

Pk(/_t)= ~ pn(/.t). (2.2)

n=k+l

The percolation probability, P~(#)[P*(~t)], is defined as the #-probability

that the origin belongs to an infinite (+)cluster [( +)*cluster].
The mean cluster size, S(#), is the expected value for the measure # of the
function n(co) which is defined as the size of the finite (+)cluster to which the
origin belongs if it belongs to any (+)cluster at all and zero otherwise.

In [8] a definition of matching graphs is given, which, in particular, applies to the graphs G
= {~2, L}, G* = {712, L*}, where L[L*] is the set of bonds between adjacent [*adjacent] points in 7/2
2 A statement equivalent to this remark is proven in [8] for any pair of matching graphs. All the
following can be restated in this more general case
3 Here and in the following we will not distinguish between chains and their images
A Note on Percolation 41

We have

s( t= 2 kpk( t= Y,
k=l k=O
(2.3t
We denote with a bar, - , the functions analogous to Pk(~), P~(#), S(#)
referred to as (-)*clusters.
We are particularly interested in the case of the measures:

#x = [ I G (2.4)
i~Z 2

where x~[O, 1] and v~ is the measure on { - 1 , 1} which assigns weights x and 1

- x respectively to 1 and - 1 . In the case of the measures (2.4) we shall write
simply pk(x), ~(x) .... instead of Pk(#x), Pk(#x),--.
We have

pk(x)= ~ xt>'l(1-x) la~l, (2.5)

171=k
Po(x) + P~ (x) = x. (2.6)

If A is a square in 7/2 and we call PA(X) the #x-probability that the origin is
(+)connected with aA, we have:

Poo(x) = Inf PA(X). (2.7)

A 0

(2.7) implies that Poo(x) is an upper semicontinuous function; this and the
obvious non-decreasing property of P~ (x) imply that P~o(x) is a right-continuous
function.
The critical points are defined by:

pc=Sup{x[Poo(x)=O}, p*=Sup{xIP*(x)=O}.
Note that if we call E~o the event that there exists in co an infinite (+)cluster,
the zero-one law implies that Poo(x)>0 if and only if #x(E~)= 1, so that Pc can
also be defined as the separation element between the x's such that #x(E~)=0
and those such that #x(Eoo)--1.
It is easy to check that 0 <Pc < 1: one can get crude bounds on Pc by applying
the Peierls argument.

3. An Improvement of Harris' Theorem

Our starting point is the following

Theorem 1. If fioo(x) > O, then P~ (x) = O.
This theorem was first proved by Harris ([6]) in the particular case of the
bond problem on the square lattice, but, as Fisher has remarked ([73), the proof
(for two-dimensional graphs) works quite in general. Recently it has been
extended, in a slightly weaker form, to the case of equilibrium measures for the
two-dimensional Ising model ([9, 103).
42 L. Russo

Theorem 1 is equivalent to the relation Pc+p* >1. In order to get exact

relations between the critical points one should require, besides Theorem 1,
some existence theorem for infinite clusters. On the other hand it is clear that
Harris' theorem cannot be inverted, because it implies that in those cases (bond
problem in the square lattice or site problem in the triangular lattice) in which
connection and *connection coincide both Po~ and/5~ must be zero in x = 1/2.
We observe, however, that the following proposition holds:
Proposition 1. I f # is a translation-invariant measure on f2, P~ (#)= 0 and S(#) < oo,
then fi~ (#) > O.
Proof. We consider the event

Ek= {coef2[in co (k, 0) is (+)connected with the 2-axis}.

We have

~(E~)__<~ P~(~)=s(~)< ~.
k=0 k=0

Hence the Borel-Cantelli lemma implies that #-a.e. at most a finite number of
the events E k can occur. For a given configuration co, if none of the Ek's occurs
the origin (0, 0) ~ 7Z2 belongs to an infinite ( -)*cluster. Also, if a finite, non-zero,
number of Ek's occur and n is the largest integer such that co ~ E,, it is clear that
co((n + 1, 0))= - 1 , and the point (n + 1, 0) cannot belong to a finite ( - ) * d u s t e r ?,
because in this case 6" ? should intersect the 1-axis in some point (N, 0) such that
N > n + 1, co~E N. Hence #-a.e. co contains an infinite (-)*cluster.
Proposition I means that if there are neither infinite (+)clusters, nor infinite
(-)*clusters, both S(x) and S(x) must diverge. This is intuitively clear, because
in this case each site in 7Z: is "surrounded" by infinitely many finite clusters of
both types. Actually, for the measures #~, this is the only case in which the mean
cluster size can diverge. In order to prove this we need some additional lemmas.
Lemma 1 ( F K G inequalities). I f we consider in f~ the partial ordering < defined
setting co1 <co2 if and only if col(x )_<co2(x) for all x s;g2 and we call an event A
positive [negative_] if its characteristic function is increasing [decreasing], then if
A and B are both positive (or both negative)

Vx~[0, 1] I~x(AC~B)>_#~(A)#~(B). (3.1)

The inequalities (3.1), known as F K G inequalities in the general case of

ferromagnetic measures, were first proved in the present case by Harris in [63.
If L is a positive integer, we call Rr.1 (x) the #~-probability that two given
opposite sides of the square

AL= {(xl, x2)~Z211xll <g, Ix21<g}

are connected in A/~ by a (+)chain.
Lemma 2. I f Po~(x) > O, then lirn Rr, 1(x) = 1.
L~o9
A Note on Percolation 43

Proof If Po~(x)>0, Theorem 1 implies that #x-a.e. there are infinitely many
(+)circuits surrounding the origin (otherwise, if n is the largest integer such that
(n, 0) belongs to such a circuit, the site (n+ 1, 0) should belong to an infinite
(-)*cluster). #x-a.e. all but at most a finite number of these circuits must belong
to the same infinite (+)cluster (otherwise no infinite (+)cluster should exist).
Hence for any e > 0 there is L 0 such that for all L > L o the event A L "there is in
A L a (+)circuit surrounding the origin and included in an infinite (+)cluster"
has #x-measure bigger than 1 - e, We call A~ (i = 1,... 4) the event that there is in
AL a (+)circuit surrounding the origin and (+)connected in A L with the i-th side
of AL. The symmetry properties of #x and the inequalities (3.1) imply:

4 z

(here and in the following " ~ " means complementation). On the other hand it
is obvious that
4
('~
A~iL c A L~ , .
i=1

hence #x(AL)<e
-i 1/4 and by using again inequalities (3.1)
i j 1/4 2
#x(ALc'~AL)>(1-e ).

If i and j are referred to two opposite sides we get

RL, 1(x) >=#x(AiL ~ AJL)>=(1 - - e 1 / 4 ) 2,

and this ends the proof.

The proof of the following lemma is based on arguments similar to the ones
used in [6] in the proof of Theorem 1.
We call RL,2(x ) the #x-probability that the two opposite smaller sides of the
rectangle

AL, 2 = {(x 1, x2)e 7]2J [x, [<2L, I.x2]<L}

are connected in AL, 2 by a (+)chain.
Lemma 3. I f P, (x) > 0, then lira RL, 2 ( X ) = 1.
L+co

Proof Given e>0, we choose two integers n, L o such that, for all L>=Lo, with
#x-probability bigger than 1 - g:
a) there is in A, a (+)circuit surrounding the origin and belonging to the
infinite ( +)cluster.
b) there is in A L a (+)chain connecting two given opposite sides of A L.
c) there is a (+)circuit surrounding the origin and contained in A L \ A n.
The existence of n, L 0 follows from Lemma 2 and the arguments used in its
proof. Let L > L o. Besides A n and A L, we consider the two translates of A L

Bl={(x1,x2)~2[-2L=<xl <0; IXe[~L};

Be={(x ,,xz) e~z[0<=xl-<2L; ]xzf<=L}.

44 L. Russo

We call sf(A) (i= 1, ...,4) the four sides of a square A, where sl(A ) is the "left
side" and the others are numbered clockwise. Using b) and the inequalities (3.1)
in the same way as in the proof of Lemma 2 we get that with #x-probability
bigger than 1 - e 1/z there is in B i a (+)chain connecting sl(B1) with s3(B1) and
having as endpoint a site of the 2-positive-half-axis; using (3.1) and c) we get that
with #x-probability bigger than (1-~1/2)(1-e) such a chain can be found in
B I \ A .. We call C the set of the chains of this type. If c e C we call c' the chain
obtained by reflecting c with respect to the 2-axis and we call S c[Sc,] the set of
sites in BI[B2] ',below c[c']". We order C by putting cl<c 2 if S~1~Sc2. It is
easy to see that if in a given configuration co there is a non-empty set 1 of
(+)chains belonging to C, then there is in I only one maximal element, namely
the chain c I = (~* ([,_)S~))c~B 1.
c~I
We define
Dc={coef21in co c is the maximal (+)chain connecting sl(Bx)
with s3(B~) and it is (+)connected in B i with s2(Bi) }.

We have:

#x( ~ Oc) = ~, #x(O~) > (1 - e)2 ( 1 - e1/2) (3.2)

c~C c~C

where we have used the fact that if there is a (+)chain in B 1 connecting s2(B1)
with s4(B1), then c is certainly (+)connected with s z(B1).
We consider now the following events:

F = {co E ~2[in co there is in A, a (+)circuit surrounding the origin and

(+)connected in A L with s2(AL) }.

E~ [E~,] = {co~ Olin co c [c'] is ( +)connected in S~u S~,

with a (+)circuit surrounding the origin}.

a) and the inequalities (3.1) imply that # x ( F ) > l - e 1/4. On the other hand, for
any c ~ C , / ~ c~/~c' ~ ~ and #~(E~) = #x(E~'); hence we have

so that

Vc~ C #x(E~) > 1 - e ~/8. (3.3)

We consider now the event A = ~ (/)r c~Ej. (3.2), (3.3) and the remark that for
c~C
any c ~ C De and E~ are independent events imply:

#~(A) = ~ #x(Dr #x(E~) >=(1 - e)2 (1 - ~l/z)(1 - ~i/8).

cEC

We call A' the event obtained from A by reflection with respect to the 2-axis. A
and A' are positive events; hence
#x (A c~A') > (1 - ~)4(1 - el/2)2 (1 - el/8)2.
A Note on Percolation 45

The lemma is proved by the remark that if A c~A' occurs, then there is a
(+)chain in BlwB 2 connecting sl(B 0 with $3(B2).

Lemma 4. If P~o(x) > O, then lim RL, a(x) = 1 (where RL,3(x) is defined in an
obvious way). L~
Proof This lemma is an easy corollary of Lemma 3; it suffices to observe that if
we consider a rectangle R of sides L and 3L as the union of three adjacent
squares Qz, Qz, Q3 and if there are (+)chains connecting the opposite smaller
sides of the rectangles Qx ~ Q2 and Q2 C.)Q3 and there is a (+)chain connecting
the two sides of Q2 adjacent to/~ (note that all these are positive events, so that
(3.1) can be applied), then there is also in R a (+)chain connecting the smaller
sides of R.
Theorem 2. Po~(x)> 0 if and only if _fi~(x)= 0 and S(x)< oo.

Proof. The "if" part of the theorem is proved by Proposition 1.

Suppose P~(x)>0, given e>0, L e m m a 4 implies that we can choose an
integer Lo(x ) such that for any L>Lo(x ), Rr,3(x)>l-e. We consider the
sequence of squares AL,, Ln=3"Lo(x). By applying once more the inequalities
(3.1) it is easy to check that for any n > 1, a (+)circuit surrounding the origin is
contained in ALn\AL,_I with #x-probability bigger than (1-e) ~. The origin can
belong to a (-)*cluster of size greater than an=4.32"(Lo(X)) 2 only if for all
i~{l ..... n} such a circuit does not exist in AL,\AL, .
Hence we have:

/~.(x) __<(1 - (1 - e)~)" (3.4)

(3.4) and the remark that/~(x) is a non-increasing sequence implies:

oo ao oo an + 1

k~O k=O n=O k=an+l

<4(Lo(x))2+ ~ (a,+l--an)fla.(X)
n=O

i
n=0

The last series is convergent if e is small enough. This, together with Theorem 1,
proves Theorem 2.

4. The Percolation Probability

In this section we consider the consequences of Theorem2 on the regularity

properties of the function P~o(x). For this we need a slight improvement of
Theorem 2.
46 L.Russo

If in the configuration co there is an infinite ( +)cluster C, we call "hole" of co

a maximal *connected component of 772\C. It is easy to see that, if H is a hole,
the set B(H)=Hc~3C is included in a (-)*cluster and IB(H)t2>IH[ 4
We consider the function h(co) defined as the size of the hole to which the
origin belongs if it belongs to any hole at all and zero otherwise, and we put

qk(x)=px(h-l(k)); Qk(X)= ~ q,(x). (4.1)

n=k+l

The expected value for the measure #~ of the function h(co) is

H(x)= ~ kqk(X)= ~ Qk(X). (4.2)

k=O k=0

The following proposition holds as a corollary of Theorem 2:

Proposition 2. If P~o(x) > 0 , then H(x) < pc.
Proof. We use the same notations as in the proof of Theorem 2 and rewrite (3.4)
in the form:

/~n(x) _<0". (4.3)

We consider the square Aa~.. (4.3) and the inequalities (3.1) imply that the event
"no site in A,~ belongs to a (-)*cluster of size greater than a," has a #~-
probability greater than

(1 - 0") (2~+ 1)~ > (1 - 0") 5~.

On the other hand if the origin belongs to a hole H of size between a,2 and a,+
2 1,
then B(H) must intersect Aa~" and all sites in B(H) belong to a (-)*cluster
greater than a,. Hence we get:
a2+1
qk(X) _--<1-- (1-- 0")Sa,"= 1 -- (1 -- 0")S(2Co(x))838
2
k=ana+ t

_-<5 (2Lo(x))S (3 s 0)',

Q~ ~ qk(x)=~ ot,+~qk(X) < lO(2 Lo(x))S(3S O)" (4.4)

k=a2+ 1 h=n k=a~+l

(where we have supposed 3s0 < 1/2, as it is certainly true if Lo(x ) is big enough).
The bound (4.4) together with the remark that Qk(X) is a non-increasing
sequence imply the convergence of (4.2).
Proposition 3. If P~ (x ) > O, then S (x ) < c~.
Proof. It is an immediate corollary of Proposition 2.
Proposition 4. I f 0 < x < 1, x =gPc, then Po~ is infinitely many times derivable in x.

This inequality can be proven by observing that if we put

L = Sup Sup[xi-yll , then [HI<L 2, [B(H)[>L

i= 1,2 x,yeH
A Note on Percolation 47

P r o @ If x < G the statement is trivially true. Suppose x > Pc. We observe that

eoo (x) = x - ~ pk(x) = x - Z xl'l (1 - x) j~ (4.5)

k=l y

where 7 runs over the family of all finite connected subsets of ~2 containing the
origin. Differentiating term by term we get the series:

1 - ( l / x ) ~ kpk(x ) + (1/(1 - x)) ~ h rh(X ) (4.6)

k=l h=l

where rh(x ) is the #x-probability that the b o u n d a r y of the ( + ) c l u s t e r to which

the origin belongs has size h. Using the b o u n d (4.4) and taking into account that
Qaa is a non-increasing function of x, it is easy to check the uniformity of the
convergence of the series (4.6) in any interval (x~, xz), x~ >Pc, x2 < 1. In the same
way one can check the existence of the higher order derivatives.

5. Concluding Remarks

If one defines, besides Pc and p*, the other two "critical points"

G=Sup{xlPoo(x)=O, S ( x ) < oe},

gc* = S u p { x l P * ( x ) = O , S*(x) < Go}.

T h e o r e m 2 can be written

Pc + G* =~c+p~*=l. (5.1)

If one can prove, as one expects, that the relation p c + p * = 1 holds, then (5.1)
should imply that Pc--no, so that the m e a n cluster size should diverge in at most
one point. We r e m a r k that conversely if one is able to p r o v e that p c = G s, then
the relation Pc + P * = 1 should follow directly from Proposition 1.
The p r o b l e m above is related to the other open p r o b l e m of the continuity
properties of P~ in G. In fact the right-continuity of P~ implies that it is
continuous in Pc if and only if P~(pc)=0; at least for the triangular lattice by
T h e o r e m 1 this is certainly true if pc=no (so that both are equal to 1/2) because
in this case Poo(Pc) = / 5 (G).

Acknowledgements. I would like to thank N. Kuiper and D. Ruelie for their kind hospitality at the
I.H.E.S. at Bures-sur-Yvette, where most of this work was written. I am also indebted to H. Kunz
and J. Lebowitz for their interest and encouragement and especially to M. Pulvirenti for many
helpful discussions.

5 This should be an assumption very similar to the one made in [8] that the function "mean
number of clusters per site" has Pc as unique singular point
48 L. Russo

References

1. Frisch, H.L., Hammersley, J.M.: J. Soc. Industr. Appl. Math. 11, 894 (1963)
2. Shante, V.K.S., Kirkpatrick, S.: Advances Phys. 20, 325 (1971)
3. Essam, J.W.: In Phase Transitions and Critical Phenomena, eds. C. Domb and M.S. Green. New
York: Academic Press 1973
4. Coniglio, A., Nappi, C.R., Peruggi, F., Russo, L.: J. Physics A, II. Ser. Math. Gen., 10, 205 (1977)
5. Lebowitz, J., Penrose, O.: Cluster and percolation inequalities for lattice systems with in-
teractions (preprint)
6. Harris, T.E.: Proc. Cambridge Philos. Soc. 56, 13 (1960)
7. Fisher, M.E.: J. Mathematical Phys. 2, 620 (1961)
8. Sykes, M.F., Essam, J.W.: J. Mathematical Phys. 5, 1117 (1964)
9. Miyamoto, M.: Comm. Math. Phys. 44, 169 (1975)
10. Coniglio, A., Nappi, C.R., Peruggi, F., Russo, L.: Comm. Math. Phys. 51, 315 (1976)

Received February 15, 1977; in revised form August 1, 1977