Introduction to Bernoulli percolation

Hugo Duminil-Copin∗

October 7, 2018

Contents
1 Phase transition in Bernoulli percolation 2

2 Everyone’s toolbox 4
2.1 Increasing coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Harris-FKG inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 The van den Berg Kesten and the Reimer inequalities . . . . . . . . . . . . . . . . . 7
2.4 Margulis-Russo formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 The non-critical phases 11

3.1 Uniqueness of the infinite connected component . . . . . . . . . . . . . . . . . . . . 11
3.2 Continuity properties of θ(p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Exponential decay in the subcritical regime . . . . . . . . . . . . . . . . . . . . . . . 14

4 Critical percolation on Z2 18
4.1 Kesten’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 The Russo-Seymour-Welsh theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Conformal invariance of two-dimensional percolation . . . . . . . . . . . . . . . . . 25

Notation A graph G is given by a set V whose elements are called vertices, and a subset E
of unordered pairs of elements in V , called edges. We usually denote an element of E by e or
xy, where x and y are understood as being the two endpoints of e. We will mostly work on the
infinite graph (V, E), called the d-dimensional hypercubic lattice (when d = 2, we speak of the
square lattice), defined as follows. The vertex set is defined as the set V ∶= Zd of elements of Rd
with integer coordinates, and the edge set E composed of edges xy with endpoints x and y (in
Zd ) satisfying ∥x − y∥ = 1. Below, we use the notation Zd to refer both to the lattice and its
vertex set. Also introduce Λn ∶= [−n, n]d for every integer n ≥ 1.
For a subgraph G = (V, E) of Zd , we introduce the boundary of G defined by

∂G ∶= {x ∈ V ∶ ∃y ∈ Zd such that xy ∈ E ∖ E}.

∗
[email protected] Institut des Hautes Études Scientifiques and Université de Genève
These lecture notes describe the content of a class given at IHES at a master level. We believe that percolation
theory is best learned by exercising on problems. We therefore included a number of them in the notes and
recommend that the reader tries them. This teaching was funded by an IDEX Chair from Paris Saclay and by
the NCCR SwissMap from the Swiss NSF.

1
1 Phase transition in Bernoulli percolation
A percolation configuration ω = (ωe ∶ e ∈ E) on G = (V, E) is an element of {0, 1}E . If ωe = 1,
the edge e is said to be open, otherwise e is said to be closed. A configuration ω can be seen as
a subgraph of G with vertex-set V and edge-set {e ∈ E ∶ ωe = 1}. A cluster will be a connected
component of ω.
A percolation model is given by a distribution on percolation configurations on G. The
simplest example of percolation model is provided by Bernoulli percolation: each edge is open
with probability p, and closed with probability 1 − p, independently of the states of other edges.
This model was introduced by Broadbent and Hammersley in 1957 [9] and has been one of the
most studied probabilistic models. We refer to [20, 6, 13] for books on the subject.
We will often define Bernoulli percolation on the infinite lattice Zd . We therefore consider
the probability space ({0, 1}E , F, Pp ), where {0, 1}E is the probability space of percolation con-
figurations on Zd , F is the σ-algebra generated by events depending on finitely many edges, and
Pp is the corresponding product measure, where each coordinate is a Bernoulli random variable
of parameter p (its expectation is denoted by Ep ).
The first question of interest in Bernoulli percolation is the existence or not of an infinite
cluster. Note that this event is measurable with respect to F since
∞
{x is in an infinite cluster} = ⋂ {the cluster of x contains a vertex at distance n of x}
n=0
{there exists an infinite cluster} = ⋃ {x is in an infinite cluster}.
x∈Zd

Define the parameter

θ(p) ∶= Pp [0 is in an infinite cluster]
and define
pc ∶= inf{p ∈ [0, 1] ∶ θ(p) > 0}.
Note that for p < pc , the probability that x is in an infinite cluster is zero by invariance under
translations. The union bound (on countably many vertices) implies that there is almost surely
no infinite cluster.

Theorem 1.1 For d ≥ 2, we have that 0 < pc (d) < 1.

A self-avoiding path of length n is a sequence of edges e1 , . . . , en with ei ≠ ej for i ≠ j, such that

ei and ei+1 share an endpoint for every 1 ≤ i < n. Let Ωn be the set of self-avoiding paths of
length n starting from the origin.

Proof Fix n > 0. If 0 is connected to infinity, there exists a path of open edges of length n
starting from 0. Therefore

θ(p) ≤ Pp [there exists γ ∈ Ωn such that ωe = 1 for every e ∈ γ]

≤ ∑ Pp [ωe = 1 for every e ∈ γ]
γ∈Ωn

≤ ∣Ωn ∣ pn ≤ (2dp)n .

In the second line, we used the union bound and independence for edges, and in the last one,
the fact that ∣Ωn ∣ ≤ (2d)n .
When p < 2d 1
, the quantity tends to 0 and therefore the origin is connected to infinity with
0 probability. In other words p ≤ pc (d). We therefore proved that pc (d) ≥ 2d
1
.

2
 The argument showing that pc < 1 is more elaborated. We need to prove that when p is
close to 1, then θ(p) > 0. Any percolation on Zd contains a copy of Bernoulli percolation on Z2
(simply look at the restriction on ω ∩ Z2 ). Hence, if the probability of 0 ←→ ∞ on Z2 is strictly
positive, then θ(p) > 0 on Zd , and therefore pc (d) ≤ pc (2). It is sufficient to show the result for
d = 2.
Consider the graph (Z2 )∗ defined as the copy of Z2 translated by the vector (1/2, 1/2). For
every configuration ω on Z2 , define a configuration ω ∗ , called the dual configuration of ω, as
follows. Every edge e of Z2 is naturally associated with an edge e∗ of (Z2 )∗ intersecting it in its
middle. Set ωe∗∗ = 1 − ωe . With words, an edge of (Z2 )∗ is open if the corresponding edge of Z2 is
closed. Note that the law of ω ∗ is a translate by (1/2, 1/2) of P1−p . An example of configuration
of ω and ω ∗ are represented on Fig. 1.

Figure 1: Two configurations ω and ω ∗ .

For the origin not to be in an infinite cluster (see Fig. 2), there must be a circuit in (Z2 )∗
of open edges in ω ∗ surrounding the origin. The circuit must intersect {(n + 12 , 0) ∶ n ∈ N}.

Figure 2: A configuration for which the cluster of the origin is finite, together with a circuit of
open edges in ω ∗ surrounding the origin.

Therefore,

1 − θ(p) ≤ ∑ Pp [ω ∗ contains an open circuit surrounding 0 and passing through (n + 12 , 0)]

n≥1
≤ ∑ Pp [ω ∗ contains an open path of length 2n + 4 passing through (n + 12 , 0)]
n≥1
≤ ∑ (4(1 − p))2n+4 . (1.1)
n≥1

As in the previous argument, we used that the probability that a path of length k is composed
of open edges in ω ∗ has probability (1 − p)k . Furthermore, there are less than 4k such paths.
When p is closed enough to 1, this sum is strictly smaller than 1. We deduce that θ(p) > 0. We
deduce that pc (d) < 1. ◻

Remark. The graph (Z2 )∗ is called the dual graph of Z2 . This notion will be often used when
working with planar percolation (see below).

3
 Exercise 1 Show that any measurable event can be approximated by events depending on finitely many edges, in
the sense that for any A in the product σ-algebra, there exists a sequence (Bn ) such that Bn is measurable in terms
of the ωe for e in the box of size n, and Pp [A∆Bn ] tends to 0 as n tends to infinity.

Exercise 2 Show that pc (Z) = 1. Show that pc (Z × {0, . . . , n}) = 1.

Exercise 3 Show that pc (d) ≤ 3/4 for every d ≥ 2.

Exercise 4 Consider the infinite tree of degree d + 1. Connect the cluster of the origin to a Galton-Watson tree.
Use this connexion to compute the critical point.

Exercise 5 Show that ∣Ωn+m ∣ ≤ ∣Ωn ∣ ⋅ ∣Ωm ∣. Deduce that ∣Ωn ∣1/n converges as n tends to infinity to µc ∈ (1, ∞). In
dimension d, what is the best bound you can come up with for µc .

Exercise 6 What site percolation on the square (hexagonal and triangular) lattice could mean? Show that 0 < pc < 1.
Show that bond percolation on a graph corresponds to site percolation on a modified graph. Show that

pc (bond) ≤ pc (site) ≤ 1 − (1 − pc (bond))d

where d is the degree of the graph.

Exercise 7 Consider a graph G for which every vertex has degree smaller than or equal to d. A finite connected
subset of a graph G is called a lattice animal. For x ∈ G and k, n two integers let a(n, x) be the number of lattice
animals with n vertices. Using percolation to show that

∞
dn
∑ [p(1 − p)] a(n, x) ≤ 1.
n=0

Deduce that for every x and n, a(n, x) ≤ 4dn . Show that one can replace pd (1 − p)d by p(1 − p)d in the previous
bound.

Exercise 8 The goal of this exercise is to implement Peierls’ argument in dimension d without using the coupling.
A minimal blocking surface is a set of edges E such that every path of adjacent edges from the origin to infinity
intersects E, and any strict subset of E does not have this property.
1. Show that any minimal blocking surface is a finite set.
2. Consider such a minimal blocking surface E and assume that it is contained in the box of size n around the origin.
Let x be outside of the box. We partition E into two non-empty sets E1 and E2 . Show that there exists two paths
γ1 and γ2 of edges from 0 to x such that γ1 does not intersect E1 , and γ2 does not intersect E2 . From now on, we
call γ the concatenation of γ1 and the time-reversal of γ2 .
3. Show that the space of loops on Zd is a Z2 -vector space. Show that the set of elementary loops, meaning loops of
length 4 is generating this vector space. Deduce that there exists a set L of elementary loops such that

γ = ∑ `.
`∈L

4. By studying the parity of the number of intersections of γ1 , γ2 , γ and each ` with E, show that there must exist
a loop ` ∈ L intersecting both E1 and E2 .
5. Deduce that a minimal surface is connected in the graph with vertex-set given by the edges of Zd , and edge-set
given by pairs of edges of Zd that are on the same loop of length 4.
6. Use the previous study to implement a version of Peierls’ argument in dimension d.

2 Everyone’s toolbox
We will need a number of tools in the following sections.

2.1 Increasing coupling

We wish to show that in a certain sense, measures Pp can be coupled in an increasing way. It will
also be interesting to consider Bernoulli percolation of different parameters at the same time. If
one considers two independent copies ω and ω ′ of Bernoulli percolation with parameters p < p′ ,
then each edge has probability p(1 − p′ ) to be open in ω. Obviously, one would like to argue that

4
there are more and more open edges, and that it should be possible to couple different Bernoulli
percolations to guarantee that ω ≤ ω ′ almost surely.
A configuration ω is smaller than or equal to ω ′ (denoted by ω ≤ ω ′ ) if ωe ≤ ωe′ for every
e ∈ Zd . The function f ∶ Ω = {0, 1}E → R is increasing if f (ω) ≤ f (ω ′ ) for every ω ≤ ω ′ . The event
A on the product σ-algebra is said to be increasing if 1A is non-decreasing. In other words, A is
increasing if ω, ω ′ ,
ω ∈ A and ω ≤ ω ′ ⇒ ω ′ ∈ A.
An event A is decreasing if and only if Ac is increasing.
Examples of increasing events:
• The edge e is open.
• The sites x and y are connected in ω (i.e. that there exists a path of open edges from x to
y), denoted by x ←→ y.
• There exists a path connecting A to B (where A and B are two sets of vertices), denoted
by A ←→ B.
• The site x is in an infinite connected component (denoted by x ←→ ∞).
• The number of open edges in F exceeds k.

Proposition 2.1 (increasing coupling) Fix p < p′ . There exists a measure P on [0, 1] ←→
d
{0, 1}Z with marginales Pp , such that

P[(ω p , ω p ) ∶ ω p ≤ ω p ] = 1.
′ ′

Proof Consider a family of iid uniform [0, 1] random variables (Ue )e∈Zd . Construct two con-
d
figurations ω, ω ′ ∈ {0, 1}Z as follows: for every e ∈ Zd , set
⎧
⎪
⎪1 if Ue ≤ p,
ωep ∶= ⎨
⎪
⎩0 otherwise.
⎪

It is easy to check that the law of ω p is Pp . Indeed, each edge is open with probability p, and
closed with probability 1 − p, independently of the state of the other edges. By construction, we
have that ω p ≤ ω p . Define P to be the law of p ↦ ω p . ◻
′

Corollary 2.2 Consider an increasing event A. Then p ↦ Pp [A] is non-decreasing.

Proof Since ω p ≤ ω p , if ω p ∈ A, then ω p ∈ A. Hence,

′ ′

Pp [A] = P[ω p ∈ A] ≤ P[ω p ∈ A] = Pp [A].

′

In particular, p ↦ θ(p) is non-decreasing and

pc = sup{p ∈ [0, 1] ∶ θ(p) = 0}.

5
2.2 Harris-FKG inequality
To estimate or compute the probability of events, it is often necessary to compute the probability
of intersections. Unfortunately, when two events are not independent, it is rarely possible to
compute the probability exactly. In the case of percolation, it is nonetheless possible to bound
this probability for a large class of events.
Proposition 2.3 (Harris inequality) Let A and B be two increasing events, then

Pp [A ∩ B] ≥ Pp [A]Pp [B].

More generally, if f and g are two bounded increasing functions,

Ep (f g) ≥ Ep (f )Ep (g).

This inequality is often called the FKG (Fortuin, Kasteleyn, Ginibre) inequality. There three
authors proved a similar inequality in a more general context. Intuitively, this inequality means
that the two increasing events A and B are positively correlated. Indeed, the inequality can be
restated as
Pp [A∣B] ≥ Pp [A],
i.e. that the conditional probability of A on B is larger or equal to the probability of A. This is
not so surprising, the fact that B occurs suggest that many edges are in fact open, which helps
the occurrence of A.

Proof It suffices to show the inequality for increasing functions since the function on events
can be obtained by setting f = 1A and g = 1B . Define {ei , i ≥ 1} to be the set of edges E
and ωek = ωk . For any increasing function, fn = Ep (f ∣ω1 , . . . , ωn ) and gn = Ep (g∣ω1 , . . . , ωn )
converge to f and g respectively (by the martingale convergence theorem). Since one also has
f g = lim fn gn , we deduce that the Harris inequality follows from the inequality

Ep (fn gn ) ≥ Ep (fn )Ep (gn ),

i.e. from Harris inequality for variables depending on finitely many edges. We prove the later
by induction on n.
For n = 1, the functions f and g depend on ω1 only, or if we prefer, are functions of {0, 1}
to R. It suffices to show the result for f (0) = g(0) = 0, since adding a constant to f and/or g
does not change the inequality. In such case, f (1) ≥ 0 and g(1) ≥ 0 since f and g are increasing.
Then,

Ep (f g) − Ep (f )Ep (g) = [pf (1)g(1) + (1 − p) ⋅ 0] − [pf (1) + (1 − p) ⋅ 0][pg(1) + (1 − p) ⋅ 0]

= pf (1)g(1) − p2 f (1)g(1) ≥ 0.

Let us now consider n ≥ 2 and let us assume that the result is true for n − 1. Fix first
ω1 , . . . , ωn−1 . The definition of conditional expectation implies that

Ep (f g∣ω1 , . . . , ωn−1 ) = pf (ω1 , . . . , ωn−1 , 1)g(ω1 , . . . , ωn−1 , 1)

+ (1 − p)f (ω1 , . . . , ωn−1 , 0)g(ω1 , . . . , ωn−1 , 0)
= Eωn (f (ω1 , . . . , ωn−1 , ⋅)g(ω1 , . . . , ωn−1 , ⋅)),

where Pωn is the law of ωn (i.e. a Bernoulli percolation of parameter p), and where f (ω1 , . . . , ωn−1 , ⋅)
and g(ω1 , . . . , ωn−1 , ⋅) depend only on ωn . Similarly,

Eωn (f ∣Fn−1 )(ω1 , . . . , ωn−1 ) = Eωn (f (ω1 , . . . , ωn−1 , ⋅)) and

Eωn (g∣Fn−1 )(ω1 , . . . , ωn−1 ) = Eωn (g(ω1 , . . . , ωn−1 , ⋅)).

6
For each (ω1 , . . . , ωn−1 ), the induction hypothesis applied to Eωn and the increasing functions
f (ω1 , . . . , ωn−1 , ⋅) and g(ω1 , . . . , ωn−1 , ⋅), which depend on ωn only, gives that

Eωn (f (ω1 , . . . , ωn−1 , ⋅)g(ω1 , . . . , ωn−1 , ⋅)) ≥ Eωn (f (ω1 , . . . , ωn−1 , ⋅)) Eωn (g(ω1 , . . . , ωn−1 , ⋅))

or in other words

Ep (f g∣ω1 , . . . , ωn−1 ) ≥ Ep (f ∣ω1 , . . . , ωn−1 ) Ep (g∣ω1 , . . . , ωn−1 ).

Now,

Ep (f g) = Ep (Ep (f g∣ω1 , . . . , ωn−1 )

≥ Ep [Ep (f ∣ω1 , . . . , ωn−1 )Ep (g∣ω1 , . . . , ωn−1 )]
≥ Ep [Ep (f ∣ω1 , . . . , ωn−1 )]Ep [Ep (g∣ω1 , . . . , ωn−1 )]
= Ep (f )Ep (g),

where we used the induction hypothesis to the variables ω1 , . . . , ωn−1 and the conditional expec-
tations Ep (f ∣ω1 , . . . , ωn−1 ) and Ep (g∣ω1 , . . . , ωn−1 ), which depend on n − 1 variables only.
◻

Exercise 9 Show that Harris inequality holds for decreasing events.

Exercise 10 Show that increasing events span the product σ-algebra.

Exercise 11 (Square-root-trick) Show that for n increasing events A1 , . . . , An . Show that

1/n
max{Pp [Ai ], i ≤ n} ≥ 1 − (1 − Pp [A1 ∪ ⋯ ∪ An )] . (2.1)

The previous inequality, called the (2.1), is very useful. It shows that when the probability of
the union of increasing events is close to 1, then so is the maximal probability of events.

2.3 The van den Berg Kesten and the Reimer inequalities
Let A be an event. Given a configuration ω in A, there is a set I ⊂ E is a witness of A for ω
(given ωI ∈ A) if ω ∈ A and any other configuration ω ′ coinciding with ω on I is also in A.
We say that two events A and B are realized disjointly if there exist two witnesses I = I(ω)
and J = J(ω) of A and B that are disjoint. This event is denoted by A ○ B. We insist on the
fact that I and J may depend on ω.
If A and B depend on two deterministic disjoint sets of edges I and J, then these sets
can be considered as witnesses, and we get that A ○ B = A ∩ B (in this case we obtain that
Pp [A ○ B] = Pp [A]Pp [B]. The BK-Reimer inequality generalizes this fact to events that a priori
depend on the same set of edges. In this case, we obtain only an inequality (see below).
Example. Note that A ∩ B and A ○ B are in general different. For A = {x ←→ x′ } and B = {y ←→
y ′ }, the event A ○ B means that there are two disjoint paths connecting x to x′ and y to y ′ .

Proposition 2.4 (BK-Reimer inequality) If A and B are events depending on finitely many
edges, then
Pp [A ○ B] ≤ Pp [A]Pp [B].

Intuitively, the probability that A ○ B knowing B is smaller than or equal to A. This is not
surprising, certain edges are used by A and cannot be used by B, When A and B are increasing,
the inequality was proved by van den Berg and Kesten. We insist on the fact that this inequality
holds only for systems depending on finitely many edges.

7
 The proof of the BK inequality is not the simplest, and we believe that it is best understood
when done by oneself. We therefore choose to postpone it to Exercise 12. Also, we will in
fact always propose an alternative to using the BK inequality, and therefore this inequality is
not really necessary for the next sections. Reimer’s inequality is even harder to use and really
necessary, we therefore omit the proof as a whole.
An important application is the following inequality, which is known in the context of other
classical models of statistical physics (such as the Ising model) under the name of Simon-Lieb
inequality.
S
We say that 0 ←→ x occurs if 0 is connected to x by edges with both endpoints in S ⊂ Zd .
Also, let ∂S be the (vertex-)boundary of S given by the set of vertices of S that are connected
in Zd to a vertex outside S.

Corollary 2.5 Consider a finite set S containing the origin and x ∉ S, then
S
Pp [0 ←→ x] ≤ ∑ Pp [0 ←→ y]Pp [y ←→ x].
y∈∂S

Proof Consider n ≥ 1. If 0 is connected to x in Λn , there exists a self-avoiding path γ of

open edges from 0 to x. Let y be the first vertex in ∂S on this path. By definition, we have
S Λn
{0 ←→ y} ○ {y ←→ x} since the piece of γ between 0 and y is a witness for the first event, and
the remaining of γ a witness for the second one. We deduce that
Λn S Λn
Pp [0 ←→ x] ≤ ∑ Pp [{0 ←→ y} ○ {y ←→ x}].
y∈∂S

The BK inequality (the events depend on finitely many edges) implies that
Λn S Λn
Pp [0 ←→ x] ≤ ∑ Pp [0 ←→ y]Pp [y ←→ x].
y∈∂S

The claim follows by letting n tend to infinity. ◻

The previous inequality can also be put in parallel with the following inequality for random
walk. Let GS (⋅, ⋅) be the Green function of the random walk on Zd , killed upon exiting S. We
have that for every finite set S containing the origin and every x ∉ S,

GZd (0, x) ≤ ∑ GS (0, y)GZd (y, x).

y∈∂S

As often, the BK inequality is used slightly too quickly and an alternative proof can replace
the argument using the BK inequality. Usually, this simpler proof exists when the witness of one
of the two events can be explored algorithmically. To illustrate this fact, let us take the example
of the previous corollary and provide an elementary proof not relying on the BK inequality.

Proof of Corollary 2.5 (without the BK inequality) Introduce the random variable
S
C ∶= {y ∈ S ∶ y ←→ 0} corresponding to the cluster of 0 in S. Since x ∉ S, one can find an edge
S Cc
y ∈ ∂S such that 0 ←→ y and y ←→ x. Using the union bound, and then decomposing on the

8
possible realizations of C, we find
S Cc
Pp [0 ←→ x] ≤ ∑ ∑ Pp [{0 ←→ y} ∩ {C = C} ∩ {y ←→ x}]
y∈∂S C⊂S
S Cc
≤ ∑ ∑ Pp [{0 ←→ y} ∩ {C = C}] ⋅ Pp [y ←→ x]
y∈∂S C⊂S
S
≤ ∑ ∑ Pp [{0 ←→ y} ∩ {C = C}]Pp [y ←→ x]
y∈∂S C⊂S
S
≤ ∑ Pp [0 ←→ y]Pp [y ←→ x].
y∈∂S

Cc S
In the second line, we used that {y ←→ x} and {0 ←→ y} ∩ {C = C} are independent. Indeed,
these events depend on disjoint sets of edges: the first one on edges with both endpoints outside
of C, the second one on edges between vertices of S with at least one endpoint in C. In the
third line, we used that
Cc
Pp [y ←→ x] ≤ Pp [y ←→ x].
S S
In the fourth line, we used that the events {0 ←→ y} ∩ {C = C} partition the event 0 ←→ y. ◻

Exercise 12 Consider A and B to be two increasing events depending on edges in E = {e1 , . . . , en } only. Consider
the duplicated set {e1 , e2 , . . . , en , e′1 , e′2 , . . . , e′n } and the product measure P where each coordinate is a Bernoulli
random variable. For j ≤ n, set ω j = (ωe1 , . . . , ωej−1 , ωe′ , . . . , ωe′n ) and
j

Aˆj = {ω̂ ∶ ω j ∈ A} and Bˆ = {ω̂ ∶ ω n+1 ∈ B}.

1. Show that P[Aˆ1 ○ B]

ˆ = Pp [A]Pp [B] and P[Aˆn+1 ○ B]
ˆ = Pp [A ○ B].

2. We wish to show that j ↦ P[Aˆj ○ B] ˆ is decreasing in j by constructing a measure-preserving injection ω̂ ↦ s(ω̂)

from Aˆj+1 ○ Bˆ to Aˆj ○ B.
ˆ Let ω̂ be a configuration in Âj+1 ○ B̂. Let I and J be witnesses of Aˆj+1 and Bˆ (respectively)
for ω̂. Assume that there exists I not containing ej . In this case, simply set s(ω) = ω. Assume now that any I
contains ej . Define s(ω̂) obtained by exchanging ej and e′j . Check that s is one-to-one and that P[ω̂] = Pp [s(ω̂)].
3. Deduce the BK inequality.

Exercise 13 Find two events A and B depending on finitely many edges for which

Pp [A ○ B] < Pp [A]Pp [B].

Exercise 14 Consider bond percolation on p = 1/2 and set R ∶= [−n, n] × [0, 2n − 1].
1. Let Hn be the event that there is an open path from left to right in R. Show that P1/2 (Hn ) = 1/2. Deduce that

P1/2 (0 ←→ ∂Λ2n ) ≥ 1/(2n).

′ be the event that there exists a self-avoiding path from left to right in R. What can be said on P
2. Let Hn ′
1/2 (Hn )?

3. Let Cn (x) be the event that there exists an open path from x to x + ∂Λn . Show that

P1/2 [Cn (0) ○ Cn (0)] ≥ 1/(2n).

What can we say about P1/2 [Cn (0)]?

S
Exercise 15 Using Corollary 2.5, show that if there exists a set S with ψp (S) ∶= ∑ Pp [0 ←→ y] < 1, then there
y∈∂S
exists c > 0 such that for every n ≥ 1,
Pp [0 ←→ ∂Λn ] ≤ exp(−cn).

2.4 Margulis-Russo formula

In this section, G is a finite graph. For a boolean function f ∶ {0, 1}E → {0, 1} and p ∈ [0, 1], set
f (p) ∶= Ep [f ].

9
Proposition 2.6 For any p ∈ [0, 1] and E finite, we have that for every f ∶ {0, 1}E → {0, 1},

f ′ (p) = 1
p(1−p) ∑ Cov[f , ωe ].
e∈E

Proof Set ∣ω∣ = ∑e∈E ωe . Differentiating f (p) = ∑ω f (ω)p∣ω∣ (1 − p)∣E∣−∣ω∣ with respect to p
immediately gives

f ′ (p) = p1 Ep [f (ω)∣ω∣] − 1−p Ep [f (ω)(∣E∣ − ∣ω∣)]

1
= 1
p(1−p) ∑ Ep [f (ω)(ωe − p)]. (2.2)
e∈E

When f is the indicator function of an increasing event, the previous formula has a more geo-
metric interpretation in terms of so-called pivotal points. For a configuration ω and an edge e,
we set ω (e) (resp. ω(e) ) for the configuration coinciding with ω except at e, where it is equal to
1 (resp. 0). We say that ω ∈ Pive (A) if ω (e) ∈ A and ω(e) ∉ A. The edge e is said to be pivotal
for A if ω ∈ Pive (A).
Example. Consider the event A ∶= {ωe = 1}. For this event, Pivf (A) is either empty if f ≠ e,
and to the full space if f = e. Note that the event that e is pivotal for A is independent of ωe .
In fact, one can easily check that this is always the case: Pive (A) is independent of ωe .
Example. Consider the event {0 ←→ ∂Λn } on Z2 . In this case, Pive (A) is the event that both
endpoints of e are connected respectively to 0 and ∂Λn in ω; furthermore, the endpoints of e∗
are connected by a circuit (minus e∗ ) which is open in ω ∗ that surrounds the origin.

Proposition 2.7 (Margulis [27], Russo [30]) For any p ∈ [0, 1] and any increasing event A
depending on finitely many edges, we have

f ′ (p) = ∑ Pp [Pive (A)].

e∈E

In words, this formula means that the derivative of the probability of an increasing event is equal
to the expected number of pivotal points for the event.

Proof Set f ∶= 1A . Note that

Ep [f (ω)(ωe − p)1Pive (A)c ] = Ep [ωe − p] ⋅ Ep [f (ω)1Pive (A)c ] = 0

(in the first equality, we used that A ∩ Pive (A)c depends on edges different from e only). There-
fore,
Ep [f (ω)(ωe − p)] = Ep [f (ω)(ωe − p)1Pive (A) ].
For ω ∈ Pive (A), the fact that f is increasing implies that f (ω) = 0 is ωe = 0, and f (ω) = 1 if
ωe = 1. We deduce that

Ep [f (ω)(ωe − p)] = (1 − p)Pp [Pive (A) ∩ {ωe = 1}].

Since ωe and Pive (A) are independent, we conclude that

Ep [f (ω)(ωe − p)] = p(1 − p)Pp [Pive (A)].

Inserting this expression in (2.2) implies the claim. ◻

10
 Exercise 16 Prove Russo’s formula using the increasing coupling between the percolations of parameters p and p + ε
(where ε converges to 0).

2.5 Ergodicity
Let τx be a translation of the lattice by x ∈ Zd . This translation induces a shift on the space
of configurations {0, 1}E . Define τx A ∶= {ω ∈ {0, 1}E ∶ τx−1 ω ∈ A}. An event A is invariant
under translations if for any x ∈ Zd , τx A = A. A measure µ is invariant under translations if
µ[τx A] = µ[A] for any event A and any x ∈ Zd . The measure is said to be ergodic if any event
invariant under translation has probability 0 or 1.
Examples. The existence of infinitely many edges, of an infinite connected component, of k
infinite connected components are three examples of events that are invariant under translations.

Lemma 2.8 The measure Pp is invariant under translations and ergodic.

Proof Let A be an increasing event depending on finitely many edges, and x ∈ Zd . We

obviously have Pp [A] = Pp [τx A] for events depending on finitely many edges. Since events
depending on finitely many edges span the σ-algebra of measurable events, we obtain that Pp is
invariant under translations.
Let us now show that the probability of an event A which is invariant under translations is
either 0 or 1. For that, we prove that Pp [A] ≤ Pp [A]2 .
Fix ε > 0. By definition of F, A can be approximated by events depending on finitely
many edges (see Exercise 10). Therefore, choose B depending on finitely many edges such that
Pp [A∆B] ≤ ε, where A∆B ∶= (B ∖ A) ∪ (A ∖ B) is the symmetric difference between A and
B. Since B depends on a finite set E of edges, there exists x large enough so that E does not
intersect the translate of E by x, so that

Pp [B ∩ τx B] = Pp [B]Pp [τx B] = Pp [B]2 .

We deduce that

Pp [A] = Pp [A ∩ A] = Pp [A ∩ τx A]
≤ Pp [B ∩ τx B] + 2ε = Pp [B]2 + 2ε ≤ Pp [A]2 + 4ε.

By letting ε tend to 0, we deduce that Pp [A] ≤ Pp [A]2 which implies that Pp [A] ∈ {0, 1}. ◻

Corollary 2.9 For p > pc , there exists an infinite cluster in ω almost surely.

Note that we may have used Kolmogorov’s law to deduce that the infinite cluster in ω exists
almost surely when p > pc since this event is in the asymptotic σ-algebra. Ergodicity will be in
fact crucial in the next section where Kolmogorov’s law will not be sufficient anymore.
Exercise 17 Show that the existence of two edge-disjoint self-avoiding paths of open edges starting from 0 has a
probability which is either 0 or 1.

3 The non-critical phases

3.1 Uniqueness of the infinite connected component
Theorem 3.1 If p ∈ [0, 1] is such that θ(p) > 0, then Pp [∃ a unique infinite cluster] = 1.

This result was first proved by Aizenman, Kesten and Newman in [2]. It was later obtained
via different types of arguments. The beautiful argument presented here is due to Burton and
Keane [10].

11
 y Λn

z
0

Figure 3: Construction of a trifurcation at the origin starting from three disjoint infinite clusters
(in gray) intersecting Λn . The three paths inside Λn are vertex-disjoint, except at the origin.

Proof of Theorem 3.1 We present the proof in the case of wired boundary conditions and
for p ∈ (0, 1) (the result is obvious for p equal to 0 or 1). Let E≤1 , E<∞ and E∞ be the events that
there are no more than one, finitely many and infinitely many infinite clusters respectively. Since
having no infinite cluster is an event which is invariant under translations, it has probability 0
or 1 by ergodicity, and it is therefore sufficient to prove that Pp [E≤1 ] = 1.
Let us start by showing that Pp [E<∞ ∖ E≤1 ] = 0. By ergodicity, E<∞ and E≤1 both have
probability equal to 0 or 1. Since E≤1 ⊂ E<∞ , we only need to prove that Pp [E<∞ ] > 0 implies
Pp [E≤1 ] > 0. Let F be the event that all (there may be none) the infinite clusters intersect Λn .
Since F is independent of En , we get that

Pp [F ∩ {ωe = 1, ∀e ∈ En }] ≥ Pp [F] p ∣En ∣

.

Now, assume that Pp [E<∞ ] > 0. Since any configuration in the event on the left contains zero
or one infinite cluster (all the vertices in Λn are connected), choosing n large enough that
Pp [F]≥ 12 Pp [E<∞ ] > 0 implies that Pp [E≤1 ] > 0.

We now exclude the possibility of an infinite number of infinite clusters. Consider n > 0 large
enough that
Pp [K infinite clusters intersect the box Λn ] ≥ 21 Pp [E∞ ], (3.1)
where K = K(d) is large enough that three vertices x, y, z of ∂Λn at a distance at least three of
each other that are connected to infinity in ω∣E∖En . Using these three vertices, one may modify1
the configuration in En as follows:
1. Choose three paths in Λn intersecting each other only at the origin, and intersecting ∂Λn
only at one point, which is respectively x, y and z.
2. Open all the edges of these paths, and close all the other edges in En .
We deduce from this construction that leaves

Pp [T0 ] ≥ [p(1 − p)] ∣En ∣

⋅ 21 Pp [E∞ ], (3.2)

where T0 is the following event: Zd ∖ {0} contains three distinct infinite clusters which are
connected to 0 by an open edge. A vertex x ∈ Zd is called a trifurcation if τx T0 =∶ Tx occurs.
1
Note that one may wish to pick K = 3 in (3.1) instead of a (a priori) larger K, but that this choice would make
the construction of the trifurcations described below more difficult due to the fact that the three clusters may
arrive very close to each other on the corner of Λn , and therefore prevent us from “rewiring them” to construct
a trifurcation at the origin.

12
 Fix n≥1 and denote the number of trifurcations in Λn by T. By invariance under translation,
Pp [Tx ] = Pp [T0 ] and therefore
Ep [T] = Pp [T0 ] × ∣Λn ∣. (3.3)
Let us now bound deterministically T. In order to do this, first perform the following two
“peelings” of the set F0 ∶= {e1 , . . . , er } of edges in En that are open in ω.
• For each 1 ≤ i ≤ r, if ei is on a cycle formed by edges in Fi−1 , set Fi = Fi−1 ∖ {ei }, otherwise,
set Fi = Fi−1 . In the end, the set F̃0 ∶= Fr = {f1 , . . . , fs } is a forest.

• For each 1 ≤ j ≤ s, if F̃j−1 ∖ {fj } contains a cluster not intersecting ∂Λn , then set F̃j to
be F̃j−1 ∖ {fj } and the cluster in question. Otherwise, set F̃j = F̃j−1 . At the end, F̃s is a
forest whose leaves belong to ∂Λn .
Since the trifurcations are vertices of degree at least three in this forest, we deduce that T is
smaller than the number of leaves in the forest, i.e. T ≤ ∣∂Λn ∣. This gives

(3.3) Ep [T] ∣∂Λn ∣

Pp [T0 ] = ≤ Ð→ 0 as n → ∞.
∣Λn ∣ ∣Λn ∣
Combined with (3.2), this implies that Pp [E∞ ] = 0. The claim follows. ◻

Exercise 18 We say that an (countable) infinite locally finite transitive graph G is amenable if

∣∂G∣
inf = 0.
G⊂G ∣G∣

Show that Theorem 3.1 still holds in this context. What about graphs which are not amenable (give some examples
of such graphs), do we always have uniqueness of the infinite cluster?

3.2 Continuity properties of θ(p)

Proposition 3.2 The map p ↦ θ(p) is continuous on [0, 1] ∖ {pc } and right continuous at pc .

Proof Set θn (p) = Pp (0 ←→ ∂Λn ). Since 0 ←→ ∂Λn depends only on the edges in Λn , the
function θn (p) is polynomial and therefore continuous. Moreover, θn (p) decreases to θ(p) as
n tends to infinity. The function θ is therefore the decreasing limit of continuous increasing
functions. This implies that θ is right continuous (see Exercise 19).
To prove the left continuity, we use the uniqueness of the infinite cluster. For p < pc , the
result is obvious. Fix p0 > pc and consider the increasing coupling P. We wish to show that

lim θ(p) = θ(p0 ).

p→p0 −

In this case,

θ(p0 ) − lim θ(p) = P(0 ∈ Cp0 ) − P(0 ∈ Cp for a certain p < p0 )

p→p0 −

= P(0 ∈ Cp0 and 0 ∉ Cp for all p < p0 ).

Choose p1 ∈ (pc , p0 ). Almost surely, ωp1 has a (unique) infinite connected component Cp1 . Let
us assume that 0 ∈ Cp0 but that 0 ∉ Cp for every p < p0 . By uniqueness of the infinite connected
component in ωp0 , there exists a finite open set from 0 to Cp1 , but that this is not the case for
p < p0 . This implies that Ue = p0 for one of the edges in this path (where Ue is the uniform
random variable used to define e). Then,

P(∃x ∈ Zd ∶ x ∈ Cp0 and x ∉ Cp for all p < p0 ) = P(∃e ∈ Zd ∶ Ue = p0 ) = 0.

13
Thus, lim θ(p) = θ(p0 ). ◻
p→p0 −

There is one question that the previous theorem does not answer: is θ left continuous at
pc ? This is equivalent to asking that θ(pc ) = 0. We will show this fact when d = 2. For Zd
with d ≥ 3, the absence of an infinite cluster at criticality was proved using lace expansion for
d ≥ 19 by Hara and Slade[22] (it was recently improved to d ≥ 11 [29]). The technique involved
in the proof is expected to work until d ≥ 6. For d ∈ {3, 4, 5}, the strategy will not work and the
following conjecture remains one of the major open questions in our field.
Conjecture 1 For any d ≥ 2, θ(pc ) = 0.
Some partial results going in the direction of this conjecture were obtained in the past decades.
For instance, it is known that the probability at pc of an infinite cluster in N × Z2 is zero [3].
Let us also mention that Ppc (Z2 ×G) [0 ←→ ∞] was proved to be equal to 0 on graphs of the form
Z2 × G, where G is finite; see [15], and on graphs with exponential growth in [4] and [23] (see
also Exercise 27).
Exercise 19 Show that a decreasing limit of continuous increasing functions is right continuous.

Exercise 20 Show that p ↦ θ(p) is strictly increasing when p > pc .

3.3 Exponential decay in the subcritical regime

Theorem 3.3 (Exponential decay in diameter) Fix d ≥ 2. For every p < pc , there exists
cp > 0 such that for all n ≥ 1, Pp [0 ↔ ∂Λn ] ≤ exp(−cp n). Furthermore, there exists c > 0 such
that for p > pc , θ(p) ≥ c(p − pc ).
Theorem 3.3 was first proved by Aizenman and Barsky [1] and Menshikov [28] (these two proofs
are presented in [19]). Here, we choose to present a new argument from [17, 18]. We also refer
to [14] for an alternative proof. The second inequality is called the mean-field lower bound.
Of course, the probability of {0 ←→ ∂Λn } is at least pn (since pn is the probability that the
path goes straight in one direction is open). We deduce that the decay cannot be faster.

S
Proof Define θn (p) ∶= Pp [0 ↔ ∂Λn ]. Recall that 0 ←→ x means that 0 is connected to x using
only edges between vertices of S. Denote the edge-boundary of S by ∆S = {xy ⊂ E ∶ x ∈ S, y ∉ S}.
For p ∈ [0, 1] and S ⊂ Zd , define
S
ϕp (S) ∶= p ∑ Pp [0 ←→ x] (3.4)
xy∈∆S

(note that ϕp (S) = 0 if 0 ∉ S). The proof will be based on the following two claims.

Claim 1. If there exists a finite set S containing 0 such that ϕp (S) < 1, then there exists cp > 0
such that for every n ≥ 1, θn (p) ≤ exp(−cp n).

Claim 2. For n ≥ 1, introduce S ∶= {z ∈ Λn ∶ z ←→

/ ∂Λn } to be the set of points not connected to
the boundary of the box of size n. For every p ∈ (0, 1), we have that
θn′ (p) = p(1−p) Ep [ϕp (S)].
1
(3.5)
Remark 3.4 On the one hand, Claim 2 motivates the introduction of the quantity ϕp (S). As we
will see in the proof, ϕp (S) can be interpreted (up to constant) as the expected number of closed
pivotal points for the event 0 ←→ ∂Λn conditionally on the fact that the connected component of
the boundary in Λn is equal to S. On the other hand, ϕp (S) < 1 for a single finite set S ∋ 0 is a
sufficient condition to imply exponential decay (this observation was actually made in the case
of S = Λn in one of the first papers on Bernoulli percolation, namely Hammersley’s paper from
1957 [21]), which provides a second motivation for the introduction of ϕp (S).

14
 Before proving these two claims, let us show how they imply the proof. Set

p̃c ∶= sup {p ∈ [0, 1] ∶ ∃ finite set S containing 0 with ϕp (S) < 1}. (3.6)

By Claim 1, for any p < p̃c , there exists cp > 0 such that θn (p) ≤ exp(−cp n) for every n ≥ 1. In
particular, θ(p) = 0 and therefore p ≤ pc . As a consequence p̃c ≤ pc . We therefore only need to
prove that p̃c ≥ pc to conclude the proof of the theorem, or equivalently that θ(p) > 0 for every
p > p̃c . In order to do this, fix p > p̃c , for which ϕp (S) ≥ 1 for every S containing 0, and ϕp (S) = 0
for every S not containing 0. The identity (3.5) becomes

θn′ (p) ≥ p(1−p) Pp [0

1
∈ S] = p(1−p) (1 − θn (p))
1
(3.7)

which can be rewritten as

p
[ log ( 1−θ )] ≥ [ log ( 1−p )] .
1 ′ ′
n

p−p̃c
Integrating between p̃c and p implies that for every n ≥ 1, θn (p) ≥ p(1−p̃ c)
. By letting n tend to
infinity, we obtain the desired lower bound on θ(p). We now prove the two claims.

Proof of Claim 1. Choose S such that ϕp (S) < 1 and L > 0 such that S ⊂ ΛL−1 . A proof similar
to the proof of Corollary 2.5 (see Exercise 23) implies that
S
θkL (p) ≤ p ∑ Pp [0 ←→ x]Pp [y ←→ ∂ΛkL ] ≤ ϕp (S)θ(k−1)L (p). (3.8)
xy∈∆S

(In the second inequality, we used that y is at a distance at least (k − 1)L of ∂ΛkL . Induction
on k gives θkL (p) ≤ ϕp (S)k , thus proving exponential decay at p.

Proof of Claim 2. Let En be the set of edges between vertices of Λn . Russo’s formula (Proposi-
tion 2.7) implies that

θn′ (p) = ∑ Pp [e is pivotal for 0 ↔ ∂Λn ] (3.9)

e∈En
= 1
(1−p) ∑ Pp [ωe = 0 and e is pivotal for 0 ↔ ∂Λn ], (3.10)
e∈En

where in the second equality we used that e is pivotal is independent of the state of e. Now, the
edge e is pivotal and closed if
P1 one of the endpoints of e is connected to 0,
P2 the other one is connected to ∂Λn ,
P3 0 is not connected to ∂Λn .
We deduce that

θn′ (p) = 1
(1−p) ∑ Pp [0 ←→ x, y ←→ ∂Λn , 0 ←→
/ ∂Λn ]. (3.11)
x,y∈Λn
xy∈En

Now, fix a set S. The intersection of {S = S} with the event on the right-hand side of (3.11)
can be interpreted nicely. Indeed, the conditions get rephrased as 0 and x belong to S and y
does not belong to S. This can be rewritten as xy ∈ ∆S and 0 is connected to x in S. Thus,
partitioning the event on the right of (3.11) into the possible values of S gives
S
θn′ (p) = 1
(1−p) ∑ ∑ Pp [0 ←→ x, S = S]
S⊂Λn xy∈∆S
S
= 1
(1−p) ∑ ( ∑ Pp [0 ←→ x]) ⋅ Pp [S = S]
S⊂Λn xy∈∆S

= (1−p) Ep [ p ϕp (S)],
1 1

15
 S
where in the second line we used that 0 ←→ x is measurable in terms of edges with both endpoints
in S, and S = S is measurable in terms of the other edges. This concludes the proof of Claim 2.
◻

Remark 3.5 Since ϕp ({0}) = 2dp, we find pc (d) ≥ 1/2d. Also, pc (d) ≤ pc (2) = 12 .

Remark 3.6 The set of parameters p such that there exists a finite set 0 ∈ S ⊂ Zd with ϕp (S) < 1
is an open subset of [0, 1]. Since this set is coinciding with [0, pc ), we deduce that ϕpc (Λn ) ≥ 1
for any n ≥ 1. As a consequence, the expected size of the cluster of the origin satisfies at pc ,

∑ Ppc [0 ←→ x] ≥ 1
dpc ∑ ϕpc (Λn ) = +∞.
x∈Zd n≥0

In particular, Ppc [0 ↔ x] cannot decay faster than algebraically (see Exercise 26 for more detail).

This theorem implies that when n is very large and p < pc , the largest clusters are of order
log n, i.e. they are much smaller than n (see Exercise 22). The previous result implies the
following exponential decay in volume.

Theorem 3.7 (Exponential decay volume) Let C be the connected component of the origin.
For every p < pc , there exists c′p > 0 such that for every n,

Pp [∣C∣ ≥ n] ≤ exp(−c′p n).

The proof of this result is a perfect example of the use of coarse graining. We also present an
alternative proof based on the BK inequality in Exercice 25. In the next proof, Λn (x) denotes
the box of size n around k.

Proof Let k be an integer to be chosen in (3.12). Define the graph Gk as being the graph
with vertex set kZd , and edges between two vertices x, y ∈ Gk if the boxes Λ2k (x) and Λ2k (y)
intersect. The graph is of degree D, where D is independent of k. We will be considering the
set A(m) of animals of cardinality m in Gk . For each animal in A(m), let T (A) be a maximal
set of sites containing at most one endpoint of each edge of Gk .
A site of Gk is good (in ω) if there exists a path from Λk (x) to Λ2k (x) in ω. Note that every
site x of Gk such that C ∩ Λk (x) ≠ ∅ and 0 ∉ Λ2k (x) is necessarily good in ω. In particular, if
n−∣Λ ∣
∣C∣ ≥ n, then there exists a connected animal in Gk of good sites of cardinality m ≥ ∣Λk2k ∣ . We
deduce that

Pp [∣C∣ ≥ n] ≤ Pp [∃A ∈ A(m) such that ∀x ∈ A, x is good]

≤ ∑ Pp [∀x ∈ A, x is good]
A∈A(m)

≤ ∑ Pp [∀x ∈ T (A), x is good]

A∈A(m)

≤ ∑ Pp [Λk ←→ ∂Λ2k ]∣T (A)∣ ,

A∈A(m)

where in the last inequality we have used that the states (good or not good) of sites in T (A)
are independent since the boxes of size 2k centered on these sites are all disjoint. By Exercise 7,
∣A(m)∣ ≤ 4Dm . Also, one can clearly find T (A) of cardinality larger than or equal to m/D (this
fact can be proved by induction). Now, by picking k large enough, one can choose k in such a
way that
2
Pp [Λk ←→ ∂Λ2k ] ≤ ∣∂Λk ∣ exp(−ck) < 1e ⋅ 4−D . (3.12)

16
Altogether, we find that

n − ∣Λ2k ∣
Pp [∣C∣ ≥ n] ≤ ∣A(m)∣ Pp [Λk ←→ ∂Λ2k ]m/D ≤ exp ( − ).
D∣Λk ∣
◻

Exercise 21 Show (3.5) using Russo formula.

Exercise 22 Show that for every p < pc , there exists Cp > 0 such that the probability that there exists a cluster of
radius larger than Cp log n in the box of size Λn tends to 0 as n tends to infinity.

Exercise 23 Provide two proofs of (3.8), one with and one without the BK inequality.

Exercise 24 (Percolation with long-range interactions) Consider a family (Jxy )x,y∈Zd of non-negative cou-
pling constants which is invariant under translations, meaning that Jxy = J(x − y) for some function J. Let Pβ be
the bond percolation measure on Zd defined as follows: for x, y ∈ Zd , {x, y} is open with probability 1 − exp(−βJxy ),
and closed with probability exp(−βJxy ).

1. Define the analogues βc , β̃c and ϕβ (S) of pc , p̃c and ϕp (S) in this context.

2. Show that there exists c > 0 such that for any β ≥ β̃c , Pβ [0 ←→ ∞] ≥ c(β − β̃c ).
3. Show that if the interaction is finite range (i.e. that there exists R > 0 such that J(x) = 0 for ∥x∥ ≥ R), then for
any β < β̃c , there exists cβ > 0 such that Pβ [0 ←→ ∂Λn ] ≤ exp(−cβ n) for all n.

4. In the general case, show that for any β < β̃c , ∑ Pβ [0 ←→ x] < ∞.
x∈Zd
Λn ∣S∣
Hint. Consider S such that ϕβ (S) < 1 and show that for n ≥ 1 and x ∈ Λn , ∑ Pβ [x ←→ y] ≤ .
y∈Λn 1 − ϕβ (S)

Exercise 25 The goal of this exercise is to show Theorem 3.7. In this exercise, p < pc is fixed.
1) What is the best bound for Pp (∣C∣ ≥ n) given by Theorem 3.3?

2) Show that it is sufficient to show that Ep (eε∣C∣ ) < ∞ for some ε > 0 sufficiently small.
3) Show that Ep (∣C∣) = ∑x∈Zd Pp (0 ←→ x) < ∞ and more generally

Ep [∣C∣n ] = ∑ Pp (0 ←→ x1 ←→ ⋯ ←→ xn ) < ∞.
x1 ,...,xn ∈Zd

4) Show that if 0, x and y are in the same connected component, then there exists u ∈ Zd such that

{u ←→ 0} ○ {u ←→ x} ○ {u ←→ y}.

Deduce that Ep [∣C∣2 ] ≤ Ep [∣C∣]3 .

In other words, the preceding reasoning shows that there exists a skeleton, in this case a tree, in Zd such that all the
edges are open, which contains the points 0, x and y, and u is the unique vertex of degree three.
5) Try to generalize this notion of a skeleton to show that Ep [∣C∣3 ] ≤ 3Ep [∣C∣]5 .
6) More generally, show that Ep (∣C∣n ) ≤ An Ep (∣C∣)2n−1 where An denotes the number of skeletons with n leaves
x1 , . . . , x n .
7) Show that An ≤ 2n n! and that Ep (∣C∣n ) ≤ 2n Ep (∣C∣)2n−1 n!.
8) Conclude.

The next two exercises illustrate the use of a very simple, yet very powerful technique: sub-additivity (and its
corresponding notion sub-multiplicativity).

Exercise 26 (Definition of the correlation length) Fix d ≥ 2 and set e1 = (1, 0, . . . , 0).
1. Prove that, for any p ∈ [0, 1] and n, m ≥ 0, Pp [x0 ←→ (m + n)e1 ] ≥ Pp [x0 ←→ me1 ] ⋅ Pp [x0 ←→ ne1 ]
−1
1
2. Deduce that ξ(p) = ( lim − n log Pp [0 ←→ ne1 ]) and that Pp [0 ←→ ne1 ] ≤ exp(−n/ξ(p)).
n→∞

17
 3. Assume that 0 ←→ ∂Λn+m , show that there exists x ∈ ∂Λn such that {0 ←→ x} ○ {x ←→ x + ∂Λm }. Deduce that for
all n, m,
Pp (0 ←→ ∂Λm+n ) ≤ ∣∂Λn ∣Pp (0 ←→ ∂Λn )Pp (0 ←→ ∂Λm ).
Show that
Pp (0 ←→ ∂Λn ) ≥ e−n/ξp /(2d d(2n + 1)d−1 ).
Show that for every n ∈ N and x ∈ ∂Λn ,
n
ξp ≥ .
Λn
− log Pp [0 ←→ x]
Deduce that ξp tends to infinity as p ↗ pc . Show that p ↦ ξp is continuous on [0, pc ).
4. Prove that for any x ∈ ∂Λn ,
Pp [0 ←→ 2ne1 ] ≥ Pp [0 ←→ x]2 .
Deduce that c
Pp [0 ↔ x] ≥ 2d(d−1)
exp(−∥x∥∞ /ξp ).
∥x∥∞
c
5. Deduce that for every x ∈ Zd , Ppc [0 ↔ x] ≥ 2d(d−1) .
∥x∥∞

Much more precise (and more difficult) estimates are known for Pp (0 ←→ ne1 ). These estimates, known as Orstein-
Zernike estimates, state that there exists c = c(p) > 0 such that

c
Pp [0 ←→ ne1 ] = exp(−n/ξp ) ⋅ (1 + o(1)).
n(d−1)/2

This exercise presents the beautiful proof due to Tom Hutchcroft of absence of percolation at criticality for
amenable locally finite transitive graphs with exponential growth. We say that G has exponential growth if there
exists cvg > 0 such that ∣Λn ∣ ≥ exp(cvg n).

Exercise 27 (Ppc [0 ↔ ∞] = 0 for amenable Cayley graphs with exponential growth) Let G be an amenable
infinite locally finite transitive graphs with exponential growth.
1. Use amenability to prove that θ(pc ) > 0 Ô⇒ inf{Ppc [x ↔ y], x, y ∈ G} > 0. Hint: use Exercise 18.
2. Use the FKG inequality to prove that un (p) = inf{Ppc [x ↔ 0], x ∈ ∂Λn } satisfies that for every n and m,

un+m (p) ≥ un (p)um (p).

3. Adapt Step 1 of the proof of Theorem 3.3 (see also Question 4 of Exercise 24) to get that for any p < pc ,
∑ Pp [0 ←→ x] < ∞.
x∈G

4. Use the two previous questions to deduce that for any p < pc , un (p) ≤ exp(−cvg n) for every n ≥ 1.
5. Conclude.

4 Critical percolation on Z2
4.1 Kesten’s theorem
In this section, we focus on the case d = 2. Recall the definition of the dual lattice (Z2 )∗ ∶=
( 12 , 12 ) + Z2 of the lattice Z2 defined by putting a vertex in the middle of each face, and edges
between nearest neighbors. Each edge e ∈ E is in direct correspondence with an edge e∗ of
the dual lattice crossing it in its middle. For a finite graph G = (V, E), let G∗ be the graph
with edge-set E ∗ = {e∗ , e ∈ E} and vertex-set given by the endpoints of the edges in E ∗ . The
configuration ω is naturally associated with a dual configuration ω ∗ : every edge e which is closed
(resp. open) in ω corresponds to an open (resp. closed) edge e∗ in ω ∗ . More formally,
ωe∗∗ ∶= 1 − ωe ∀e ∈ E.
Note that if ω is sampled according to Pp , then ω ∗ is sampled according to (a translate of) P1−p .
This duality relation suggests that the critical point of Bernoulli percolation on Z2 is equal to
1/2. We discuss different levels of heuristic leading to this prediction.

Heuristic level 0 The simplest non-rigorous justification of the fact that pc = 1/2 invokes the
uniqueness of the phase transition, i.e. the observation that the model should undergo a single

18
Figure 4: The rectangle Rn together with its dual Rn∗ (the green edges on the boundary are
irrelevant for the crossing, so that we may consider only the black edges, for which the dual
graph is isomorphic to the graph itself (by rotating it). The dual edges (in red) of the edge
boundary of the cluster of the right boundary in ω (in blue) is a cluster in ω ∗ crossing from top
to bottom in Rn∗ .

change of macroscopic behavior as p varies. This implies that pc must be equal to 1 − pc , since
otherwise the model will change at pc (with the appearance of an infinite cluster in ω), and
at 1 − pc (with the disappearance of an infinite cluster in ω ∗ ). Of course, it seems difficult to
justify why there should be a unique phase transition. This encourages us to try to improve our
heuristic argument.

Heuristic level 1 One may invoke a slightly more subtle argument. On the one hand, assume
for a moment that pc < 1/2. In such case, for any p ∈ (pc , 1 − pc ), there (almost surely) exist
infinite clusters in both ω and ω ∗ . Since the infinite cluster is unique almost surely, this seems
to be difficult to have coexistence of an infinite cluster in ω and an infinite cluster in ω ∗ , and it
therefore leads us to believe that pc ≥ 1/2. On the other hand, assume that pc > 1/2. In such
case, for any p ∈ (1 − pc , pc ), there (almost surely) exist no infinite cluster in both ω and ω ∗ .
This seems to contradict the intuition that if clusters are all finite in ω, then ω ∗ should contain
an infinite cluster. This reasoning is wrong in general (there may be no infinite cluster in both
ω and ω ∗ ), but it seems still believable that this should not occur for a whole range of values of
p. Again, the argument is fairly weak here and we should improve it.

Heuristic level 2 Consider the event, called Hn , corresponding to the existence of a path of
open edges of ω in Rn ∶= [0, n] × [0, n − 1] going from the left to the right side of Rn . Observe
that the complement of the event Hn is the event that there exists a path of open edges in ω ∗
going from top to bottom in the graph Rn∗ ; see Fig. 5. Using the rotation by π/2, one sees that
at p = 1/2, these two events have the same probability, so that

P1/2 [Hn ] = 1
2 ∀n ≥ 1. (4.1)

Now, one may believe that for p < pc , the clusters are so small that the probability that one
of them contains a path crossing Rn from left to right tends to 0, which would imply that the

19
Figure 5: Construction in the proof of Proposition 4.1. One path connects the left side of Rn (in
blue) to the blue hatched area. The other one on the right side of Rn (in red) to the red hatched
area. The two paths must be in the same cluster (of Rn ) by uniqueness, which therefore must
contain a path from left to right.

probability of Hn would tend to 0, and therefore that pc ≤ 1/2. On the other hand, one may
believe that for p > pc , the infinite cluster is so omnipresent that it contains with very high
probability a path crossing Rn from left to right, thus implying that the probability of Hn would
tend to 1. This would give pc ≥ 1/2. Unfortunately, the first of these two claims is difficult to
justify. Nevertheless, the second one can be proved as follows.

Proposition 4.1 Assume that θ(p) > 0, then lim Pp [Hn ] = 1.

n→∞

Proof Fix n ≥ k ≥ 1. Since a path from Λk to Λn ends up either on the top, bottom, left or
right side of Λn , the square root trick implies that

Pp [Λk is connected in Λn to the left of Λn ] ≥ 1 − Pp [Λk ←→

/ ∞]1/4 .

Set n′ = ⌊(n − 1)/2⌋. Consider the event An that (n′ , n′ ) + Λk is connected in Rn to the left of
Rn , and (n′ + 2, n′ ) + Λk is connected in Rn to the right of Rn . We deduce that

Pp [An ] ≥ 1 − 2Pp [Λk ←→

/ ∞]1/4 .

The uniqueness of the infinite cluster implies2 that

lim inf Pp [Hn ] = lim inf Pp [An ] ≥ 1 − 2Pp [Λk ←→

/ ∞]1/4 .
n→∞ n→∞

Letting k tend to infinity and using that the infinite cluster exists almost surely, we deduce that
Pp [Hn ] tends to 1. ◻

2
The event An ∖ Hn is included in the event that there are two distinct clusters in Rn going from Λk to ∂Rn .
The intersection of the latter events for n ≥ 1 is included in the event that there are two distinct infinite clusters,
which has zero probability. Thus, the probability of An ∖ Hn goes to 0 as n tends to infinity.

20
 Exercise 28 (Zhang argument) 1. Show that

/ ∞]1/4 .
P1/2 [top of Λn is connected to infinity outside Λn ] ≥ 1 − P1/2 [Λn ←→

2. Deduce that the probability of the event Bn that there exist infinite paths in ω from the top and bottom of Λn to
infinity in Z2 ∖ Λn , and infinite paths in ω ∗ from the left and right sides to infinity satisfies

/ ∞]1/4 .
P1/2 [Bn ] ≥ 1 − 4P1/2 [Λn ←→

3. Using the uniqueness of the infinite cluster, prove that P1/2 [Λn ↔
/ ∞] cannot tend to 0.

Exercise 29 Consider Bernoulli percolation on Z2 . Use the Borel-Cantelli lemma to show that for every p < pc ,
there exists an infinite connected component in ω ∗ almost surely. Deduce that pc ≤ 1/2.

This proposition together with (4.1) implies the following corollary

Corollary 4.2 There is no infinite cluster at p = 1/2. In particular, pc ≥ 1/2.
We are in a position to state Kesten’s theorem.
Theorem 4.3 (Kesten [24]) For Bernoulli percolation on Z2 , pc is equal to 1/2. Furthermore,
there is no infinite cluster at pc .
As mentioned above, the last thing to justify rigorously is the fact that for p < pc , Pp [Hn ] tends
to 0.

Proof By Theorem 3.3, for any p < pc , there exists cp > 0 such that for all n ≥ 1,
Pp [0 ←→ ∂Λn ] ≤ exp(−cp n).
Then, Pp [Hn ] tends to 0 as n tends to infinity since
n−1
Pp [Hn ] ≤ ∑ Pp [(0, k) is connected to the right of Rn ]
k=0
≤ n Pp [0 ←→ ∂Λn ] ≤ n exp(−cp n).
By (4.1), this implies that p < 1/2. ◻

Remark 4.4 Note that we just proved that θ(pc ) = 0 on Z2 .

B
Exercise 30 (pc (G) + pc (G∗ ) = 1) In this exercise, we use the notation A ←→ C the event that A and C are con-
nected by a path using vertices in B only. Consider Bernoulli percolation on a planar lattice G embedded in such a
way that Z2 acts transitively on G. We do not assume any symmetry of the lattice. We call the left, right, top and
bottom parts of a rectangle Left, Right, Top and Bottom. Also, H(n, k) and V(n, k) are the events that [0, n] × [0, k]
is crossed horizontally and vertically by paths of open edges.
1. Use the Borel-Cantelli lemma and Theorem 3.3 (one may admit the fact that the theorem extends to this context)
to prove that for p < pc (G), there exists finitely many open circuits surrounding a given vertex of G∗ . Deduce that
pc (G) + pc (G∗ ) ≤ 1.
We want to prove the converse inequality by contradiction. From now on, we assume that both p > pc (G) and
p∗ > pc (G∗ ).
2. For s > 0 and x ∈ Z2 , define Sx = x + [0, s]2 . Prove that for any rectangle R, there exists x = x(R) ∈ R ∩ Z2 such
that there exists x′ and x′′ neighbors of x in Z2 satisfying

R R R R
Pp [Sx ←→ Bottom] ≥ Pp [Sx ←→ Top] Pp [Sx ←→ Left] ≥ Pp [Sx ←→ Right], (4.2)
R R R R
Pp [Sx′ ←→ Top] ≥ Pp [Sx′ ←→ Bottom] Pp [Sx′′ ←→ Right] ≥ Pp [Sx′′ ←→ Left]. (4.3)

3. Set H ∶= R+ × R, `+ ∶= {0} × R+ , `− ∶= {0} × R− and ` = `− ∪ `+ . Prove that there exists x = x(m) with first coordinate
equal to m satisfying

H H H H
Pp [Sx ←→ `− ] ≥ Pp [Sx ←→ `+ ] and Pp [Sx+(0,1) ←→ `− ] ≤ Pp [Sx+(0,1) ←→ `+ ].

21
 4. Using the square root trick, deduce that

H √ H
√
Pp [Sx ←→ `− ] ≥ 1 − / `]
Pp [Sx ←→ and Pp [Sx+(0,1) ←→ `+ ] ≥ 1 − / `].
Pp [Sx+(0,1) ←→

5. Using the fact that there exists a unique infinite cluster in ω almost surely, prove that the probability that {0}×[0, 1]
is connected in ω ∗ ∩ H to infinity is tending to 0.
6. Prove that the distance between x(R) and the boundary of R is necessarily tending to infinity as min{n, k} tends
to infinity.
7. Using x(R), prove that max{Pp [V(n, k)], Pp [H(n, k + 1)]} tends to 1 and min{Pp [V(n, k)], Pp [H(n, k)]} tends to
0 as min{k, n} tends to infinity. Hint. Use the square root trick and the uniqueness criterion like in the previous
questions.
8. By considering the largest integer k such that Pp [V(n, k)] ≥ Pp [H(n, k)], reach a contradiction. Deduce that
pc (G) + pc (G∗ ) ≥ 1.

Exercise 31 (Critical points of the triangular and hexagonal lattices) Define p such that p3 + 1 = 3p and
set pc for the critical parameter of the triangular lattice.
1. Consider a graph G and add a vertex x inside the triangle u, v, w. Modify the graph F by removing edges uv, vw
and wu, and adding xu, xv and xw. The new graph is denoted G′ . Show that the Bernoulli percolation of parameter
p on G can be coupled to the Bernoulli percolation of parameter p on G′ in such a way that connections between
different vertices of G are the same.

w w
p p 1−p

p
1−p x 1−p
u v u v

2. Using exponential decay in subcritical for the triangular lattice, show that if p < pc , the percolation of parameter
1 − p on the hexagonal lattice contains an infinite cluster almost surely. Using the transformation above, reach a
contradiction.
3. Prove similarly that p ≤ pc (T).
4. Find a degree three polynomial equation for the critical parameter of the hexagonal lattice.

4.2 The Russo-Seymour-Welsh theory

We saw that the probability of crossing (almost) squares was equal to 1/2. This raises the
question of probabilities of crossing more complicated shapes, such as a rectangle with an aspect
ratio ρ ≠ 1. While this could look like a technical question, studying crossing probabilities is
instrumental in the study of critical random cluster models.
We begin with some general notation. For a rectangle R ∶= [a, b] × [c, d], introduce the event
H(R) that R is crossed horizontally, i.e. that the left side {a} × [c, d] is connected by a path
in ω ∩ R to the right side {b} × [c, d]. Similarly, define V(R) be the event that R is crossed
vertically, i.e. that the bottom side [a, b] × {c} is connected by a path in ω ∩ R to the top side
[a, b] × {d}. When R = [0, n] × [0, k], we rather write V(n, k) and H(n, k).

Theorem 4.5 (Box-crossing property) Let ρ > 0, there exists c = c(ρ) > 0 such that for
every n ≥ 1,
c ≤ P1/2 [H(ρn, n)] ≤ 1 − c.

Note that as soon as we have to our disposal a uniform lower bound (in n) for some ρ > 1 on
crossing horizontally rectangles of the form [0, n] × [0, ρn], then one can easily combine crossings
in different rectangles to obtain a uniform lower bound for any ρ′ > 1. Indeed, set ε = ρ − 1 and
define (for integers i ≥ 0) the rectangles Ri ∶= [iεn, (iε+ρ)n]×[0, n] and the squares Si ∶= Ri ∩Ri+1 .
Then,
(Harris)
P1/2 [H(ρ′ n, n)] ≥ P1/2 [ ⋂ (H(Ri ) ∩ V(Si ))] ≥
′
c(ρ)⌊ρ /ε⌋ .
i≤ρ′ /ε

22
One may even prove lower and upper bounds for crossing probabilities in arbitrary topological
rectangles (see Exercise 32 below).
Note that combining crossings in squares is much harder. This will in fact be the major ob-
stacle: the main difficulty of this theorem lies in passing from crossing squares with probabilities
bounded uniformly from below to crossing rectangles in the hard direction with probabilities
bounded uniformly from below. A statement claiming that crossing a rectangle in the hard
direction can be expressed in terms of the probability of crossing squares is called a Russo-
Seymour-Welsh type theorem.

Theorem 4.6 (Russo-Seymour-Welsh) For every n ≥ 1,

P1/2 [H(3n, 2n)] ≥ 1

128 .

For Bernoulli percolation on the square lattice, such a result was first proved in [30, 31] (maybe
with a different lower bound on the right-hand side). Since then, many proofs have been pro-
duced, among which [7, 5, 8, 33, 32]. We present a recent proof [7], which is the shortest one
(for the square lattice) we are aware of.

Proof Let us introduce the three rectangles

R ∶= [−n, 2n] × [−n, n] S ∶= [0, n]2 S ′ ∶= [−n, n]2 .

Define A = H(S) and B the event that there exists a horizontal crossing of S which is connected
to the left L of S ′ . For a path γ from top to bottom in S, and σ(γ) the reflection of this path
with respect to Z × {0}, define the set V (γ) of vertices x ∈ S ′ on the left of γ ∪ σ(γ). Now, on
A, condition on the right-most crossing Γ of S. We find that

P1/2 [B] ≥ ∑ P1/2 [B ∣ A ∩ {Γ = γ}] P1/2 [A ∩ {Γ = γ}]

γ
V (γ)
≥ ∑ P1/2 [γ ←→ L] P1/2 [{Γ = γ} ∩ A]
γ

≥ 4 ∑ P1/2 [{Γ
1
= γ} ∩ A] = 41 P1/2 [A] ≥ 18 .
γ

In the third line, to deduce the lower bound 1/4 we used the facts that conditioned on A∩{Γ = γ},
the configuration in V (γ) is a Bernoulli percolation of parameter 1/2 (since A ∩ {Γ = γ} is
measurable with respect to edges on γ or above γ), the symmetry and the fact that the probability
of an open path from left to right in V (γ) is larger than 1/2 (by (4.1) applied to S ′ ). In the last
inequality, we used (4.1) applied to S ′ .
Now, the event H(R) occurs if the three events V(S), B and B ′ occur, where B ′ is the
event that there exists a horizontal crossing of S which is connected to the right in [0, 2n]2 . By
symmetry,
P1/2 [B ′ ] = P1/2 [B] ≥ 81 .
Harris inequality (used in the second inequality) implies that
1
P1/2 [H(R)] ≥ P1/2 [V(S) ∩ B ∩ B ′ ] ≥ P1/2 [V(S)]P1/2 [B]P1/2 [B ′ ] ≥ .
128
◻

Corollary 4.7 There exists α ∈ (0, ∞) such that for every n ≥ 1,

1 1
≤ P1/2 [0 ←→ ∂Λn ] ≤ α .
2n n

23
Proof For the lower bound, simply observe that

nP1/2 [0 ←→ ∂Λn ] ≥ P1/2 [H([0, n] × [0, n − 1])] = 1/2.

For the upper bound, introduce the event A ∶= V([−3n, 3n] × [2n, 3n]). If ∂Λn is connected to
∂Λ4n , then one of the four rotated versions of the event A must also occur (where the angles of
the rotation are π2 k with 0 ≤ k ≤ 3). Therefore, the FKG inequality implies that

P1/2 [∂Λn ←→ ∂Λ4n ] ≤ 1 − P1/2 [Ac ]4 ≤ 1 − c,

where c > 0 exists thanks to Russo-Seymour-Welsh. Using the independence, we obtain that

P1/2 [0 ←→ ∂Λn ] ≤ ∏ P1/2 [∂Λ4k−1 ←→ ∂Λ4k ] ≤ (1 − c)⌊log4 n⌋ ≤ n−α . (4.4)

4k ≤n

Exercise 32 Consider a simply connected domain with a smooth boundary Ω with four distinct points a, b, c and
d on the boundary. Let (Ωδ , aδ , bδ , cδ , dδ ) be the finite graph with four marked points on the boundary defined as
follows: Ωδ is equal to Ω ∩ δZ2 (we assume here that it is connected and of connected complement, so that the
boundary is a simple path) and aδ , bδ , cδ , dδ be the four points of ∂Ωδ closest to a, b, c and d.
Prove that there exists c = c(q, Ω, a, b, c, d) > 0 such that for any δ > 0,

Ωδ
φ1pc ,q [(aδ bδ ) ←→ (cδ dδ )] ≥ c,

where (aδ bδ ) and (cδ dδ ) are the portions of ∂Ωδ from aδ to bδ , and from cδ to dδ , when going counterclockwise
around ∂Ωδ .

Exercise 33 Consider Bernoulli percolation (of parameter p) on a planar transitive locally finite infinite graph with
π/2 symmetry.
1. Using the rectangles R1 = [0, n] × [0, 2n], R2 = [0, n] × [n, 3n], R3 = [0, n] × [2n, 4n], R4 = [0, 2n] × [n, 2n] and
R5 = [0, 2n] × [2n, 3n], show that
Pp [H(n, 4n)] ≤ 5P[H(n, 2n)] .
2. Deduce that u2n ≤ 25u2n where un = Pp [H(n, 2n)]. Show that (un ) decays exponentially fast as soon as there
1
exists n such that un < 25 .
1
3. Deduce that un ≥ 25
for every n or (EXPp ). What did we prove at pc ?

Exercise 34 1. Prove that there exists c > 0 such that P1/2 [0 ←→ ∂Λn ] ≤ cP1/2 [0 ←→ ∂Λ2n ].
2. Prove that there exist c1 , c2 > 0 such that for any x ∈ ∂Λn ,

c1 P1/2 [0 ←→ ∂Λn ]2 ≤ P1/2 [0 ←→ x] ≤ c2 P1/2 [0 ←→ ∂Λn ]2 .

3. (quasi-multiplicativity) Prove that there exists c1 > 0 such that for any 1 ≤ n ≤ N /2,

P1/2 [0 ←→ ∂ΛN ] P1/2 [0 ←→ ∂ΛN ]

≤ P1/2 [Λn ←→ ∂ΛN ] ≤ c1 .
P1/2 [0 ←→ ∂Λn ] P1/2 [0 ←→ ∂Λn ]

Exercise 35 1. show that the probability that there exists a horizontal crossing of [0, 4n] × [−n, 0] in ω and a
horizontal crossing of [− 12 , 4n + 21 ] × [ 12 , n − 12 ] in ω ∗ .
2. Use reasoning similar to Exercise 13 applied to the upper boundary of a cluster crossing [0, 4n] × [−n, 0] to
get that P1/2 [0 ←→ ∂Λn ] ≥ cn−1/3 for all n ≥ 1.

Exercise 36 Consider a measure µ on {0, 1}E which is invariant under the graph isomorphisms of Z2 onto itself.
We further assume that µ satisfies the FKG inequality. We assume that inf n µ[H(n, n)] > 0 . The goal of this exercise
is to prove that
lim sup µ[H(3n, n)] > 0 . (4.5)
n

1. Let En be the event that the left side of [−n, n]2 is connected to the top-right corner (n, n). Use the FKG inequality
to prove that lim supn µ[En ] > 0 implies (4.5).

24
 2. Assume the limit superior above is zero. Now, for any −n ≤ α < β ≤ n, define the event Fn (α, β) to be the
existence of a crossing from the left side of [−n, n]2 to the segment {n} × [α, β]. We consider the function

hn (α) = µ[Fn (0, α)] − µ[Fn (α, n)] .

Show that hn is an increasing function, and that there exists c0 > 0 such that hn (n) > c0 for all n.
3. Assume that hn (n/2) < c0 /2. Use (FKG) to prove that (4.5).
4. Assume that hn (n/2) > c0 /2, and let αn = inf{α ∶ h(α) > c0 /2}. Define the event Xn (α) by the existence of a
cluster in [−n, n]2 connecting the four segments {−n} × [−n, −α], {−n} × [α, n], {n} × [−n, −α], and {n} × [α, n].
Prove that there exists a constant c1 > 0 independent of n such that µ[Xn (α)] ≥ c1 .
5. Prove that, for infinitely many n’s, αn < 2α2n/3 .
6. Prove that, whenever αn < 2α2n/3 , there exists a constant c2 such that µ[H(8/3n, 2n)] > c2 . Conclude.

4.3 Conformal invariance of two-dimensional percolation

In 1992, the observation that properties of interfaces should also be conformally invariant led
Langlands, Pouliot and Saint-Aubin ([26]) to publish numerical values in agreement with the
conformal invariance in the scaling limit of crossing probabilities in the percolation model. More
precisely, consider a Jordan domain Ω with four points A, B, C and D on the boundary. The 5-
tuple (Ω, A, B, C, D) is called a topological rectangle. The authors checked numerically that the
probability C δ (Ω, A, B, C, D) of having a path of adjacent open sites between the boundary arcs
AB and CD of a finite piece of a lattice of mesh size δ converges as δ goes to 0 towards a limit
which is the same for (Ω, A, B, C, D) and (Ω′ , A′ , B ′ , C ′ , D′ ) if they are images of each other
by a conformal map. Notice that the existence of such a crossing property can be expressed in
terms of properties of a well-chosen interface, thus keeping this discussion in the frame proposed
earlier.
The paper [26], while only numerical, attracted many mathematicians to the domain. The
authors attribute the conjecture on conformal invariance of the limit of crossing probabilities
to Aizenman. The same year (1992), Cardy [11] proposed an explicit formula for the limit. In
2001, Smirnov proved Cardy’s formula rigorously for critical site percolation on the triangular
lattice, hence rigorously providing a concrete example of a conformally invariant property of the
model. In this section, we switch our attention to site percolation on the triangular lattice. Note
that the Russo-Seymour-Welsh is also true in this context.
Below, a domain always denotes a finite piece Ω of H cut from H by taking the finite
connected component of H ∖ Γ, where Γ is a self-avoiding polygon on T ∶= H∗ . The set ∂Ω will
denote the set of mid-edges on Γ. A topological rectangle is a domain with four points a, b, c, d
on ∂Ω. The arc (ab) denotes the counterclockwise arc of ∂Ω from a to b. We extend this
definition to subdomains of the triangular lattice δT with mesh size δ > 0. In this case, we set
(Ωδ , aδ , bδ , cδ , dδ ).
We see site percolation on the triangular lattice as a coloring in black and white of the
hexagons of H or δH. In this context, we extend the measure Pp to δT.

Set j = e2πi/3 . Let T be the equilateral triangle with vertices 1, j and j 2 .

Theorem 4.8 (Smirnov) Consider a family of topological rectangles (Ωδ , aδ , bδ , cδ , dδ )δ>0 con-
verging to (Ω, a, b, c, d) in the Caratheodory sense. Then,

Ωδ
lim P1/2 [(aδ bδ ) ←→ (cδ dδ )] = x,
δ→0

where d is mapped to x + j 2 (1 − x) by the conformal map from Ω to T mapping a to 1, b to j

and c to j 2 .

Below, we present a version of Smirnov’s proof of this theorem due to Mikhail Khristoforov.

25
 Let us highlight a connection between site percolation on the triangular lattice and the
so-called loop O(1) model on the hexagonal lattice. A loop configuration on a domain Ω is
a collection of non-intersecting loops. Equivalently, it is an even subgraph of Ω, i.e. a graph
with even degree everywhere. We denote the set of loop configurations by EΩ∅ . Each percolation
configuration on H can be seen as a loop configuration by considering all the edges of H separating
two faces with different colors. A percolation configuration on a finite domain Ω can be seen as a
percolation configuration on H by coloring all the hexagons outside of Ω in white. We therefore
can associate a percolation configuration in Ω with a loop configuration in EΩ∅ . We deduce that
∣EΩ∅ ∣ = 2N where N is the number of faces in Ω.
We extend the definition of loop configurations to include paths from mid-edges to mid-edges.
More precisely, let EΩuv (resp. EΩuv,rs ) be the set of configurations with loops together with one
self-avoiding path from u to v (resp. and one self-avoiding path from r to s) avoiding the loops.
Before diving into the proof, let us introduce an object which will be important. Fix z a
mid-edge of Ω and three mid-edges a, b, c on ∂Ω. Define Fa (z) ∶= 2−N card(EΩaz,bc ) and similarly
Fb and Fc . Finally, define
F (z) ∶= Fa (z) + jFb (z) + j 2 Fc (z).
The function F is called the parafermionic observable with three marked points in the domain
Ω. Among other things, these parafermionic observables have been used to prove conformal
invariance of the Ising model [12] and to compute the connective constant of the hexagonal
lattice [16].
Remark 4.9 We have that Fa (z) + Fb (z) + Fc (z) = 1 since
card(EΩaz,bc ) + card(EΩbz,ca ) + card(EΩcz,ab ) = card(EΩ∅ ) = 2N .
The first identity is given by the fact that for any η0 ∈ EΩaz,bc ∪ EΩbz,ca ∪ EΩcz,ab , the map η ↦ η∆η0
is a bijection from EΩaz,bc ∪ EΩbz,ca ∪ EΩcz,ab to EΩ∅ .
Lemma 4.10 For any vertex v in Ω, if p, q and r denote the three mid-edges around v, indexed
in counterclockwise order,
(p − v)F (p) + (q − v)F (q) + (r − v)F (r) = 0. (4.6)
We propose two proofs of this statement.

Proof (number 1) Fix a path γ from b to c and let EΩaz,bc [γ] and EΩbc [γ] be the sets of loop
configurations in EΩaz,bc and EΩbc with path γ from b to c. We obtain that
⎧
⎪
az,bc ⎪0 if there is no lattice path from a to z disjoint from γ
card(EΩ [γ]) = ⎨ (4.7)
⎪
⎩card(EΩ [γ]) otherwise.
bc
⎪
Indeed, the first line is trivial and the second simply follows from the fact that for any path ψ
from a to z in the complement of γ, the map η ↦ η∆ψ is an involution from EΩaz,bc [γ] to EΩbc [γ].
Together with the observation that if γ does not go through v, it either separates simultaneously
p, q and r from v, or it does not, we deduce that
(p − v)Fa (p) + (q − v)Fa (q) + (r − v)Fa (r) = (p − v)Ta (p) + (q − v)Ta (q) + (r − v)Ta (r), (4.8)
where Ta (p) if equal to 2−N times the number of configurations in E az,bc for which the path from
b to c visits v.
But any configuration in EΩap,bc is naturally in bijection with configurations in EΩbq,ca and
EΩcr,ab by simply rotating the two mid-edges incident to v. We deduce that
(p − v)Ta (p) + (q − v)jTb (q) + (r − v)j 2 Tc (r) = 0.
The proof follows by doing the same for b and c instead of a, and by inserting this in (4.8). ◻

26
Proof (number 2) Let c(ω) ∶= τ (ω)(z − v)2−N be the contribution of the configuration ω
to the left-hand side of (4.6), where τ (ω) is equal to 1, j or j 2 depending on which mid-edge of
the boundary is paired with z.
We group configurations in triplets (ω1 , ω2 , ω3 ) in such a way that c(ω1 ) + c(ω2 ) + c(ω3 ) = 0.
In the next cases, we assume that ω1 contains a path ending at p. There are three cases:
• If no edge in ω1 is incident to v, then add pv and vq to obtain ω2 , and pv and vr to obtain
ω3 . One easily obtains that c(ω1 ) = j 2 c(ω2 ) = jc(ω3 ).

• If a loop of ω1 passes through v, then add pv and remove vq to obtain ω2 , and add pv and
remove vr to obtain ω3 . Again, one obtains that c(ω1 ) = j 2 c(ω2 ) = jc(ω3 ).

• If the path from boundary to boundary passes through v, then add pv and remove vq
to obtain ω2 , and add pv and remove vr to obtain ω3 . We find τ (ω2 ) = jτ (ω1 ) and
τ (ω3 ) = j 2 τ (ω1 ) so that c(ω1 ) = jc(ω2 ) = j 2 c(ω3 ).
If ω1 contains a path ending at q or r, one can adapt the previous construction by permuting
the indices. Note that any configuration is in one of these triplets, so that the claim follows. ◻

Remark 4.11 The result of Lemma 4.10 can be understood in the following way. Coefficients
in (4.6) are three cubic roots of unity multiplied by p − v, so that the left-hand side can be seen as
a discrete integral along an elementary contour on the dual lattice in the following sense. For a
closed path z = (zi )i≤n of vertices in the triangular lattice T dual to H, define the discrete integral
of a function F on mid-edges by
n−1
∮ F (z)dz ∶= ∑ F ( 2 ) (zi+1 − zi ).
zi +zi+1
(4.9)
z i=0

Equation (4.6) at v implies that the discrete contour integral going around the face of T corre-
sponding to v is zero. Decomposing a closed contour into a sum of elementary triangles (this
can always be done by the definition of Ω) gives that the discrete integral along any closed path
vanishes.

Remark 4.12 Note that (4.7) implies that Fa (z) = Ga (z), where Ga (z) is equal to 2−N times
the number of configurations in EΩbc for which the path from b to c does not disconnect a from z.

Proof of Theorem 4.8 Consider the domain Ωδ and three mid-edges aδ , bδ , cδ on ∂Ωδ . Define
Faδ , Fbδ , Fcδ and F δ as before. Extend F δ to the interior of the polygon ∂Ωδ by convexly
interpolating between mid-edges. Remark 4.9 implies that F δ maps Ωδ to T. Furthermore
Fbδ (z) (resp. Fcδ (z) and Faδ (z)) is obviously equal to 0 on (cδ aδ ) (resp. (aδ bδ ) and (bδ cδ )). As a
consequence F δ (z) ∈ [j 2 , 1] (resp. [1, j] and [j, j 2 ]) for z ∈ (cδ aδ ) (resp. (aδ bδ ) and (bδ cδ )). In
other words, ∂Ωδ is mapped to ∂T.
We also have that for z ∈ (cδ aδ ),

Ωδ
Faδ (z) = P1/2 [(aδ bδ ) ←→ (cδ dδ )] = 1 − Fcδ (z).

(The first identity comes from the fact that EΩaz,bc is in bijection with percolation configurations
having a crossing from (ab) to (cd), and the second from Remark 4.9 again.) Using Lemma 4.13
below, we may consider a sub-sequential limit F of (F δ )δ>0 . By Lemma 4.10 and Morera’s
theorem, F is holomorphic. The previous observation gives us that F is the holomorphic map
from Ω to T mapping ∂Ω in an injective way (this follows from the fact that F δ (z) maps ∂Ω to
∂T in an injective way). Furthermore, a is mapped to 1, b to j, c to j 2 . From standard results
of complex analysis (more precisely the principle of corresponding boundaries, see for instance

27
[25, Theorem 4.3]), we deduce that F is the conformal map from Ω to T mapping a, b and c to
1, j and j 2 respectively. We deduce that

F (d) = x + j 2 (1 − x) = lim [Faδ (dδ ) + j 2 (1 − Faδ (dδ ))],

δ→0

and the result follows readily. ◻

Lemma 4.13 The family (F δ )δ>0 is precompact for the uniform convergence on any compact
subset of Ω.

Proof We prove that (Faδ )δ>0 is precompact. By Remark 4.12, Faδ = Gδa . We deduce that

Faδ (z) = Gδa (z) = P1/2 [γ does not disconnect z from a],

where γ is obtained as the interface between the black clusters touching (bδ cδ ) and the white
clusters connecting (cδ bδ ), and disconnecting means that there exists no self-avoiding path from
z to a not intersecting γ.
By the Russo-Seymour-Welsh theorem, one can then prove that

∣Faδ (z) − Faδ (z ′ )∣ ≤ P1/2 [γ disconnects exactly one of z or z ′ from a] ≤ C∣z − z ′ ∣α .

Indeed, if there exists a circuit of black hexagons surrounding both z and z ′ , or a path of black
hexagons from boundary to boundary and disconnecting both z and z ′ from a, then the interface
must either disconnect simultaneously z and z ′ , or disconnect none of them.
This implies that (Faδ ) is equicontinuous and the result follows by Arzela-Ascoli. ◻

References
[1] M. Aizenman and D. J. Barsky. Sharpness of the phase transition in percolation models.
Comm. Math. Phys., 108(3):489–526, 1987.

[2] M. Aizenman, H. Kesten, and C. M. Newman. Uniqueness of the infinite cluster and
continuity of connectivity functions for short and long range percolation. Comm. Math.
Phys., 111(4):505–531, 1987.

[3] D. J. Barsky, G. R. Grimmett, and Charles M. Newman. Percolation in half-spaces: equality

of critical densities and continuity of the percolation probability. Probab. Theory Related
Fields, 90(1):111–148, 1991.

[4] I. Benjamini, Russell Lyons, Y. Peres, and Oded Schramm. Critical percolation on any
nonamenable group has no infinite clusters. Ann. Probab., 27(3):1347–1356, 1999.

[5] Béla Bollobás and Oliver Riordan. The critical probability for random Voronoi percolation
in the plane is 1/2. Probab. Theory Related Fields, 136(3):417–468, 2006.

[6] Béla Bollobás and Oliver Riordan. Percolation. Cambridge University Press, New York,
2006.

[7] Béla Bollobás and Oliver Riordan. A short proof of the Harris-Kesten theorem. Bull.
London Math. Soc., 38(3):470–484, 2006.

[8] Béla Bollobás and Oliver Riordan. Percolation on self-dual polygon configurations. In An
irregular mind, volume 21 of Bolyai Soc. Math. Stud., pages 131–217. János Bolyai Math.
Soc., Budapest, 2010.

28
 [9] S. R. Broadbent and J. M. Hammersley. Percolation processes. I. Crystals and mazes. Proc.
Cambridge Philos. Soc., 53:629–641, 1957.

[10] R. M. Burton and M. Keane. Density and uniqueness in percolation. Comm. Math. Phys.,
121(3):501–505, 1989.

[11] John L. Cardy. Critical percolation in finite geometries. J. Phys. A, 25(4):L201–L206, 1992.

[12] Dmitry Chelkak and Stanislav Smirnov. Universality in the 2D Ising model and conformal
invariance of fermionic observables. Invent. Math., 189(3):515–580, 2012.

[13] H. Duminil-Copin. Lectures on the Ising and Potts models on the hypercubic lattice.
arXiv:1707.00520.

[14] H. Duminil-Copin, A. Raoufi, and V. Tassion. Sharp phase transition for the random-cluster
and Potts models via decision trees. arXiv:1705.03104, 2017.

[15] H. Duminil-Copin, V. Sidoravicius, and V. Tassion. Absence of infinite cluster for crit-
ical Bernoulli percolation on slabs. Communications in Pure and Applied Mathematics,
69(7):1397–1411, 2016.

[16] √
H. Duminil-Copin and S. Smirnov. The connective constant of the honeycomb lattice equals
√
2 + 2. Ann. of Math. (2), 175(3):1653–1665, 2012.

[17] H. Duminil-Copin and V. Tassion. A new proof of the sharpness of the phase transition
for Bernoulli percolation and the Ising model. Communications in Mathematical Physics,
343(2):725–745, 2016.

[18] H. Duminil-Copin and V. Tassion. A new proof of the sharpness of the phase transition for
Bernoulli percolation on Zd . Enseignement Mathématique, 62(1-2):199–206, 2016.

[19] G. Grimmett. Percolation, volume 321 of Grundlehren der Mathematischen Wissenschaften

[Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, second edition,
1999.

[20] G. R. Grimmett. Inequalities and entanglements for percolation and random-cluster models.
In Perplexing problems in probability, volume 44 of Progr. Probab., pages 91–105. Birkhäuser
Boston, Boston, MA, 1999.

[21] J. M. Hammersley. Percolation processes: Lower bounds for the critical probability. Ann.
Math. Statist., 28:790–795, 1957.

[22] Takashi Hara and Gordon Slade. Mean-field critical behaviour for percolation in high
dimensions. Comm. Math. Phys., 128(2):333–391, 1990.

[23] Tom Hutchcroft. Critical percolation on any quasi-transitive graph of exponential growth
has no infinite clusters. Comptes Rendus Mathematique, 354(9):944–947, 2016.

[24] H. Kesten. The critical probability of bond percolation on the square lattice equals 2.
1

Comm. Math. Phys., 74(1):41–59, 1980.

[25] Serge Lang. Complex analysis, volume 103 of Graduate Texts in Mathematics. Springer-
Verlag, New York, fourth edition, 1999.

[26] R. Langlands, P. Pouliot, and Y. Saint-Aubin. Conformal invariance in two-dimensional

percolation. Bull. Amer. Math. Soc. (N.S.), 30(1):1–61, 1994.

29
[27] Grigorii A. Margulis. Probabilistic characteristics of graphs with large connectivity. Prob-
lemy Peredaci Informacii, 10(2):101–108, 1974.

[28] M. V. Menshikov. Coincidence of critical points in percolation problems. Dokl. Akad. Nauk
SSSR, 288(6):1308–1311, 1986.

[29] Fitzner R. and van der Hofstad R. Mean-field behavior for nearest-neighbor percolation in
d>10. arXiv:1506.07977.

[30] L. Russo. A note on percolation. Z. Wahrscheinlichkeitstheorie und Verw. Gebiete, 43(1):39–

48, 1978.

[31] P. D. Seymour and D. J. A. Welsh. Percolation probabilities on the square lattice. Ann.
Discrete Math., 3:227–245, 1978. Advances in graph theory (Cambridge Combinatorial
Conf., Trinity College, Cambridge, 1977).

[32] Vincent Tassion. Planarité et localité en percolation. PhD thesis, ENS Lyon, 2014.

[33] Vincent Tassion. Crossing probabilities for voronoi percolation. Annals of Probability,
44(5):3385–3398, 2016. arXiv:1410.6773.

30