A note on some critical thresholds of Bernoulli percolation

Pengfei Tang∗

March 2, 2023
arXiv:2012.01135v2 [math.PR] 1 Mar 2023

Abstract
Consider Bernoulli bond percolation on a locally finite, connected graph G and let pcut be
the threshold corresponding to a “first-moment method” lower bound. Kahn (Electron. Comm.
Probab. Volume 8, 184-187. (2003)) constructed a counter-example to Lyons’ conjecture of
pcut = pc and proposed a modification. Here we give a positive answer to Kahn’s modified
question. The key observation is that in Kahn’s modification, the new expectation quantity
also appears in the differential inequality of one-arm events. This links the question to a lemma
of Duminil-Copin and Tassion (Comm. Math. Phys. Volume 343, 725-745. (2016)). We also
study some applications for Bernoulli percolation on periodic trees.

1 Introduction
Let G = (V, E) be a locally finite (i.e., each vertex has finite degree), connected, infinite graph.
For p ∈ [0, 1], Bernoulli(p) bond percolation studies the random subgraph ω of G formed by
keeping each edge with probability p and removing otherwise, independently of each other. The
edges kept in ω are called open edges and the edges removed are called closed edges. The
connected components are called (open) clusters. For background on Bernoulli percolation, see
Chapter 7 of [12] or [5]. For p ∈ [0, 1], let Pp denote the law of Bernoulli(p) bond percolation and
Ep the corresponding expectation.
Let C(x) denote the open cluster of x in Bernoulli percolation. Let |C(x)|V , |C(x)|E denote the
number of vertices and edges in the cluster C(x) respectively. Let A ←→ B denote the event that
there is an open path connecting some vertex x ∈ A and y ∈ B. Let x ←→ ∞ denote the event
that the diameter of C(x) is infinite. The critical probability pc is defined as
pc = pc (G) := sup{p ≥ 0 : Pp (x ←→ ∞) = 0}.

Since for a locally finite graph G, the three events x ←→ ∞, |C(x)|V = ∞ and |C(x)|E = ∞
are actually the same event, one can also define

pc = pc (G) := sup{p ≥ 0 : Pp (|C(x)|V = ∞) = 0}

or
pc = pc (G) := sup{p ≥ 0 : Pp (|C(x)|E = ∞) = 0}.
Let x be a vertex in G. We say that ΠE is a edge cutset separating x from infinity, if ΠE is
a set of edges such that the connected component of x in G\ΠE is finite. Similarly one can define
vertex cutsets.
∗
Department of Mathematical Sciences, Tel Aviv University. Partially supported by the National Science
Foundation under grant DMS-1612363, DMS-1954086 and by ERC starting grant 676970 RANDGEOM. Email:
[email protected].

1
Definition 1.1. Suppose G is a locally finite, connected, infinite graph. Define
pcut,E = pcut,E (G) := sup{p ≥ 0 : inf Ep [|C(x) ∩ ΠE |] = 0},
ΠE

where the infimum is taken over all edge cutsets ΠE separating x from infinity and C(x) ∩ ΠE
denotes the intersection of the edge set of C(x) with ΠE .
Define
pcut,V = pcut,V (G) := sup{p ≥ 0 : inf Ep [|C(x) ∩ ΠV |] = 0},
ΠV
where the infimum is taken over all vertex cutsets ΠV separating x from infinity and C(x) ∩ ΠV
denotes the intersection of the vertex set of C(x) with ΠV .
For any edge (or vertex) cutset Π separating x from infinity, if the event {x ←→ ∞} occurs,
then C(x) ∩ Π is nonempty. Hence
Pp (x ←→ ∞) ≤ Pp (|C(x) ∩ Π| ≥ 1) ≤ Ep [|C(x) ∩ Π|]. (1.1)
Thus one has that
pcut,E ≤ pc and pcut,V ≤ pc . (1.2)
Historically another critical value pT is also of great interest (coincide with the notation pT,V
below).
Definition 1.2. Suppose G is a locally finite, connected, infinite graph. Define
pT,V = pT,V (G) := sup{p ≥ 0 : Ep [|C(x)|V ] < ∞}
and
pT,E = pT,E (G) := sup{p ≥ 0 : Ep [|C(x)|E ] < ∞}.
P∞
If p < pT,V , then n=1 Ep [|C(x) ∩ Πn |] ≤ Ep [|C(x)|V ] < ∞, where Πn := {y : dG (y, x) = n} is
the cutset consisting of vertices at graph distance n to x. Hence p < pT,V implies that p ≤ pcut,V .
Thus
pT,V ≤ pcut,V . (1.3)
Similarly one has that
pT,E ≤ pcut,E . (1.4)
It is easy to see that these critical values pc , pcut,E , pcut,V , pT,E , pT,V do not depend on the choice
of x by Harris’ inequality [12, Section 5.8].
By (1.2), (1.3) and (1.4) we now have
pT,E ≤ pcut,E ≤ pc .
and
pT,V ≤ pcut,V ≤ pc .
Lyons showed that pc = pcut,V holds for trees [11] and tree-like graphs [10] and pointed out
pcut,V = pc for transitive graphs in [11] since pT,V = pc for such graphs [1, 13]; and these results for
pcut,V applied equally to pcut,E on these graphs. In view of these examples Lyons conjectured that
pc = pcut,V for general graphs (lines 11–12 on page 955 of [11]).
Later Kahn [8] constructed a family of counterexamples to Lyons’ conjecture. P Kahn’s examples
exhibited a sequences of vertex cutsets Πn such that the quantity |C(x) ∩ Πn | = v∈Πn 1(x←→v) is
usually zero but has a large expectation for some p < pc (G). That was achieved by large correlation
among the events {x ←→ v} for v ∈ Πn , i.e., conditioned on the event that v is connected to x via
an open path, with high probability a lot of other vertices in Πn are also connected to x via v. In
light of this Kahn proposed the following modification of Lyons’ conjecture:

2
Question 1.3. Does pc (G) = p′cut,V (G) hold for every locally finite, connected, infinite graph G?

Here the notation p′cut,V = p′cut,V (G) from [8] (there it was denoted by p′cut ) is defined as follows.

Definition 1.4. Suppose G is a locally finite, connected, infinite graph. For x ∈ V , let ΠV be a
vertex cutset which separates x from infinity. For each v ∈ ΠV , let A(x, v, ΠV ) denote the event
that x is connected to v via an open path without using vertices in ΠV \{v}. Define
n X o
p′cut,V = p′cut,V (G) := sup p ≥ 0 : inf Pp [A(x, v, ΠV )] = 0 ,
ΠV
v∈ΠV

where the infimum is taken over all vertex cutsets ΠV separating x from infinity.
Similarly for an edge cutset ΠE separating x from infinity and e ∈ ΠE , let A(x, e, ΠE ) denote
the event that x is connected to e via an open path without using edges in ΠE \{e} (Here we assume
e itself is also open on A(x, e, ΠE ).) Define
n X o
p′cut,E = p′cut,E (G) := sup p ≥ 0 : inf Pp [A(x, e, ΠE )] = 0 ,
ΠE
e∈ΠE

where the infimum is taken over all edge cutsets ΠE separating x from infinity.

Similarly one can ask ([12, Question 5.16]):

Question 1.5. Does pc (G) = p′cut,E (G) hold for every locally finite, connected, infinite graph G?

Our main result is the following affirmative answer to Question 1.3 and 1.5 for Bernoulli bond
percolation.

Theorem 1.6. For Bernoulli bond percolation on every locally finite, connected, infinite graph G,
one has that
p′cut,E = p′cut,V = pc .

The same result holds for Bernoulli site percolation on a locally finite, connected, infinite graph
with bounded degree if one defines p′cut,E , p′cut,V accordingly using Bernoulli site percolation; see
Remark 5.1 and Conjecture 5.2 for more discussions.

2 Some relations of the critical thresholds

For any edge cutset Π separating x from infinity, if the event {x ←→ ∞} occurs, then there is at
least one edge e such that the event A(x, e, Π) occurs. Hence by union bounds,
X
Pp (x ←→ ∞) ≤ Pp [A(x, e, Π)] (2.1)
e∈Π

Thus one has that

p′cut,E ≤ pc . (2.2)
Similarly one has that
p′cut,V ≤ pc . (2.3)
P P
Also obviously for any edge cutset Π one has that e∈Π Pp [A(x, e, Π)] ≤ e∈Π Pp [e ∈ C(x)] =
Ep [|C(x) ∩ Π|]. Hence one has that
pcut,E ≤ p′cut,E . (2.4)

3
Similarly one has that
pcut,V ≤ p′cut,V . (2.5)
By (1.3), (1.4), (2.2),(2.3), (2.4) and (2.5) one has that

pT,V ≤ pcut,V ≤ p′cut,V ≤ pc (2.6)

and
pT,E ≤ pcut,E ≤ p′cut,E ≤ pc . (2.7)
We also have the following relations.
Lemma 2.1. Suppose G is a locally finite, connected, infinite graph. Then

pcut,E ≤ pcut,V (2.8)

If moreover G has bounded degree, then the equality holds in (2.8).

Lemma 2.2. Suppose G is a locally finite, connected, infinite graph. Then

pT,E ≤ pT,V (2.9)

If moreover G has bounded degree, then the equality holds in (2.9).

Proof of Lemma 2.2. For (2.9), if p > pT,V , then Ep [|C(x)|V ] = ∞. Since C(x) is connected,
|C(x)|E ≥ |C(x)|V − 1. Hence Ep [|C(x)|E ] = ∞. Therefore if p > pT,V , then p ≥ pT,E . Thus
pT,E ≤ pT,V as desired.
If G has bounded degree, i.e., D(G) := sup{deg(v) : v ∈ V } < ∞, then by |C(x)|E ≤
D(G)|C(x)|V one can get the other direction similarly. Hence if G has bounded degree, then
pT,V (G) = pT,E (G).

Example 2.3. Here we give an example G with unbounded degree and such that pT,E < pT,V . Let
M > 1 be an integer. Let Cn be a complete graph with M n vertices. Let o = (0, 0) be the origin of
Z2 and let (n, 0) ∈ Z2 , n ≥ 1 be the points on the x-axis. For each n ≥ 1, add an edge from (n, 0) to
each vertex of Cn . Let G be the graph obtained in this way; see Figure 1. Then obviously pc (G) =
pc (Z2 ) = 12 . Note that for p ∈ (0, pc ), Pp [o ←→ (n, 0) in G] = Pp [o ←→ (n, 0) in Z2 ] ≈ e−nϕ(p) ,
where ϕ(p) is the reciprocal of the correlation length (see Proposition 6.47 in [5] for example.) When
computing Ep [|Co |V ], each clique Cn contributes roughly p·e−nϕ(p) ·M n but when computing Ep [|Co |E ],
each clique Cn contributes roughly p2 · e−nϕ(p) · M 2n . Using the properties of ϕ(p) (Theorem 6.14
in [5]) it is easy to show that 0 < pT,E (G) = ϕ−1 (2 log M ) < pT,V (G) = ϕ−1 (log M ) < pc (G) and
we omit the details.
Before proving Lemma 2.1, we recall the definitions of boundaries of a set of vertices.
Definition 2.4. For a nonempty set of vertices K ⊂ V , we define its inner vertex boundary, outer
vertex boundary and edge boundary as follows. The inner vertex boundary ∂V in K is

in
∂V K := {y ∈ K : ∃ z ∈
/ K s.t. y ∼ z},

where y ∼ z denotes that y and z are neighbors in G. The outer vertex boundary ∂V K is

∂V K := {z ∈
/ K : ∃ y ∈ K s.t. y ∼ z}.

The edge boundary ∆K is defined as

∆K := {e = (y, z) ∈ E : y ∈ K, z ∈
/ K}.

4
 C1 C2 C3 C4
··· ··· ··· ···

(0, 0) (1, 0) (2, 0) (3, 0) (4, 0)

Figure 1: An example with 0 < pT,E < pT,V < pc < 1.

Lemma 2.5. If Π is a minimal vertex cutset separating x from infinity, then ∂V S(Π) = Π, where
S(Π) is the connected component of x in the subgraph G\Π. Similarly, if Π is a minimal edge cutset
separating x from infinity, then ∆S(Π) = Π, where S(Π) is the connected component of x in the
subgraph G\Π.

Definition 2.6. Suppose G is a locally finite, connected, infinite graph. Fix x ∈ V (G). Let
BE = BE (x) be the collection of edge cutsets that are the edge boundary of some finite cluster of
x. In other words,

BE = {∆S : S is a finite connected subgraph containing x}.

We call BE the family of boundary edge cutset. Similarly we denote by BV the collection of
vertex cutsets that arise as outer vertex boundary of some finite cluster of x and call BV the family
of boundary vertex cutset.

Proof of Lemma 2.5. Suppose Π is a minimal vertex cutset separating x from infinity. The con-
nected component S(Π) of x in G\Π is finite.
First we show that ∂V S(Π) ⊂ Π. For any z ∈ ∂V S(Π), by definition of ∂V S(Π), there is some
vertex y ∈ S(Π) such that y ∼ z. Then if z ∈
/ Π, then by definition of S(Π), then z can be connected
to x via a path from x to y in G\Π and the edge (y, z). This implies that z ∈ S(Π) if z ∈/ Π, which
contradicts with the choice of z ∈ ∂V S(Π). Hence ∂V S(Π) ⊂ Π.
On the other hand, since ∂V S(Π) is also a vertex cutset and Π is minimal with respect to
inclusion, one has that ∂V S(Π) ⊃ Π.
The case of minimal edge cutset can be proved similarly and we omit the details.

Remark 2.7. The reverse of Lemma 2.5 is not true. For example consider the half integer line G =
(N, E), where E = {(n, n + 1) : n ∈ N}. Let S = {10, 11, . . . , 100}, then ∆S = {(9, 10), (100, 101)}
is not minimal.

Proof of Lemma 2.1. For any edge cutset ΠE separating x from infinity, let S(ΠE ) be the connected
component of x in G\ΠE . Since ΠE is a cutset, S(ΠE ) is finite. Let ΠV = ΠV (ΠE ) be the endpoints
of edges in ∆S(ΠE ) that are not in S(ΠE ). Then ΠV is a vertex cutset, since every infinite path
from x to infinity has to leave S(ΠE ), the first vertex on the path that is not in S(ΠE ) must be a
vertex in ΠV . For each v ∈ ΠV , pick an arbitrary edge e = e(v) ∈ ∆S(ΠE ) such that e is incident
to v. Note that

1. ∆S(ΠE ) ⊂ ΠE ;

5
 2. for distinct v ∈ ΠV , the edges e(v) are also distinct (since each such edge e(v) has exactly
one endpoints not in S(ΠE ), i.e., v);

3. for such v and e = e(v), P[e ∈ C(x)] ≥ pPp [v ∈ C(x)] (By insertion-tolerance of Bernoulli
bond percolation, see [12, Exercise 7.1]).

u1 ev
x u2 v
u3
S(ΠE )

Figure 2: A systematic drawing of S(ΠE ), v ∈ ΠV (ΠE ) and e = e(v); edges in ∆S(ΠE ) are colored
brown.

Hence
X 1 X 1
Ep [|C(x) ∩ ΠV |] = Pp [v ∈ C(x)] ≤ Pp [e ∈ C(x)] = Ep [|C(x) ∩ ΠE |] (2.10)
p p
v∈ΠV e∈ΠE

If 0 < p < pcut,E , then inf ΠE Ep [|C(x) ∩ ΠE |] = 0. Thus by (2.10), for p < pcut,E

inf Ep [|C(x) ∩ ΠV |] = 0.
ΠV

Hence if p < pcut,E , then p ≤ pcut,V . This implies (2.8).

Now we assume that G has bounded degree. Let ΠV be a vertex cutset and without loss of
generality we assume that x ∈ / ΠV . Let S(ΠV ) denote the connected component of x in G\ΠV .
Since ΠV is a cutset, the connected component of S(ΠV ) is finite. Let ΠE = ΠE (ΠV ) be the edge
boundary ∆S(ΠV ). Now for each edge e ∈ ∆S(ΠV ), there is a unique vertex v = v(e) ∈ ΠV
associated to e: v is incident to e (the other endpoint of e is in S(ΠV ), which is disjoint from ΠV
by its definition). Note that

1. for each v ∈ ΠV , there are at most D = D(G) edges in ∆S(ΠV ) associated to it;

2. for each e ∈ ∆S(ΠV ) and its associated vertex v = v(E) ∈ ΠV , Pp [v ∈ C(x)] ≥ Pp [e ∈ C(x)].

Hence
X X
Ep [|C(x) ∩ ΠE |] = Pp [e ∈ C(x)] ≤ D Pp [v ∈ C(x)] = DEp [|C(x) ∩ ΠV |] (2.11)
e∈ΠE v∈ΠV

6
Thus when D < ∞, by (2.11) one has that
∀ p < pcut,V , inf Ep [|C(x) ∩ ΠE |] = 0.
ΠE

Hence when D < ∞, p < pcut,V ⇒ p ≤ pcut,E . Together with (2.8) one has the equality when
D < ∞.
When considering pcut,V and pcut,E , obviously it suffices to consider minimal cutsets (with respect
to inclusion). However a priori it is not clear whether it is sufficient to consider minimal cutsets for
p′cut,E and p′cut,V . This consideration for minimal cutsets together with Lemma 2.5 motivates us to
consider p′′cut,E and p′′cut,V in Definition 2.8 and it turns out they all coincides with each other; see
Theorem 2.9.
Definition 2.8. Suppose G is a locally finite, connected, infinite graph. Fix x ∈ V (G). Define
n X o
p′′cut,V = p′′cut,V (G) := sup p ≥ 0 : inf Pp [A(x, v, Π)] = 0 ,
Π∈BV
v∈Π

where the infimum is taken over all boundary vertex cutsets Π that separate x from infinity.
Similarly, we define
n X o
p′′cut,E = p′′cut,E (G) := sup p ≥ 0 : inf Pp [A(x, e, ΠE )] = 0 ,
ΠE ∈BE
e∈ΠE

where the infimum is taken over all boundary edge cutsets ΠE that separate x from infinity.
By Definition 1.4 and 2.8, and inequalities (2.2), (2.3) one has that
p′′cut,E ≤ p′cut,E ≤ pc and p′′cut,V ≤ p′cut,V ≤ pc . (2.12)
Theorem 1.6 is contained in the following more general theorem.
Theorem 2.9. For Bernoulli bond percolation on every locally finite, connected, infinite graph G,
one has that
p′′cut,E = p′cut,E = p′′cut,V = p′cut,V = pc .
Lemma 2.10. Suppose G is a locally finite, connected infinite graph. Then
p′′cut,E ≤ p′′cut,V . (2.13)
Proof. Let ΠE ∈ BE be a boundary edge cutset separating x from infinity. Let S be the finite
connected component of x in G\ΠE . By definition ΠE = ∆S. Let ΠV = ∂V S be the outer vertex
boundary of S. Then ΠV ∈ BV is a boundary vertex cutset separating x from infinity.
For each v ∈ ΠV , if the event A(x, v, ΠV ) occurs, then there is a self-avoiding open path γx,v
from x to v only using v in ΠV . Hence this path uses only one edge e in ∆S, namely the edge e on
γx,v that is incident to v. Hence the event A(x, e, ΠE ) occurs for this edge e on the path γx,v . Thus
X
Pp (A(x, v, ΠV )) ≤ Pp (A(x, e, ΠE )),
e∈∆S : e∼v

where e ∼ v denotes that v is an endpoint of e.

Note that for any two distinct vertices v, v ′ ∈ ΠV , the two sets {e ∈ ∆S : e ∼ v} and {e′ ∈
∆S : e′ ∼ v ′ } are disjoint. Hence summing the above inequality over v ∈ ΠV = ∂S, one has that
X X
Pp (A(x, v, ΠV )) ≤ Pp (A(x, e, ΠE )).
v∈ΠV e∈ΠE

From this we have that p < p′′cut,E ⇒ p ≤ p′′cut,V and then we have the desired inequality (2.13).

7
3 Proof of Theorem 2.9
Duminil-Copin and Tassion [3] gave a new proof of the sharpness of the phase transition [1, 13].
For our purpose, we just need to look at the short version for Bernoulli percolation [4].
For p ∈ [0, 1], x ∈ V and a finite set S with x ∈ S ⊂ V , define
X X S
ϕp (x, S) := p Pp (x ←→ y), (3.1)
y∈S z ∈S,(y,z)∈E
/

S
where {x ←→ y} denotes the event that there is an open path connecting x and y only using
vertices lying in S. Recall that the edge boundary ∆S of S is the set of edges that connect S
to its complement. So ϕp (x, S) is the expected number of open edges on the edge boundary ∆S
which has an endpoint is connected to x via an open path entirely lying in S. For transitive graphs,
Duminil-Copin and Tassion defined

pec := sup{p ≥ 0 : inf{ϕp (x, S) : x ∈ S, S is finite} < 1}

and showed that pec = pc for transitive graphs.

The main new ingredient of the proof of Theorem 2.9 is the following observation.

Proposition 3.1. For Bernoulli bond percolation on a locally finite, connected, infinite graph G
one has that X
inf Pp (A(x, e, Π)) = inf ϕp (x, S), (3.2)
Π∈BE S
e∈Π

where the infimum on the left hand side of (3.2) is over all the boundary edge cutsets separating
x from infinity and the infimum on the right is over all finite sets containing x.

We have that pc (G) = sup{p ≥ 0 : inf{ϕp (x, S) : x ∈ S, S is finite} = 0} for all locally finite,
connected, infinite graphs in light of Proposition 3.1 and Theorem 2.9.

Proof of Proposition 3.1. On the one hand, for any finite set S containing x, let S ′ be the connected
component of x in the induced subgraph of S. Then Π(S) := ∆S ′ is a boundary edge cutset
separating x from infinity. For each edge e = (y, z) ∈ ∆S ′ , say y ∈ S ′ , z ∈
/ S ′ , it is easy to see that
z∈/ S and
S
Pp (A(x, e, Π(S))) = p · Pp [x ←→ y].
Summing this over all edges e ∈ ∆S ′ , one has that
X X
inf Pp (A(x, e, Π)) ≤ Pp (A(x, e, Π(S))) = ϕp (x, S ′ ) = ϕp (x, S),
Π∈BE
e∈Π e∈Π(S)

where the last equality is a simple observation from the definition of S ′ .

Hence X
inf Pp (A(x, e, Π)) ≤ inf ϕp (x, S).
Π∈BE S
e∈Π

On the other hand, for any boundary edge cutset Π separating x from infinity, let S = S(Π)
be the connected component of x in the graph G\Π. By Definition 2.6, ∆S = Π. For each edge
e = (y, z) ∈ ∆S = Π, say y ∈ S, z ∈
/ S, one has that
S
Pp (A(x, e, Π)) = p · Pp [x ←→ y].

8
Summing this over all edges e ∈ ∆S, one has that for a boundary edge cutset Π and S = S(Π)
X
Pp (A(x, e, Π)) = ϕp (x, S(Π)) ≥ inf ϕp (x, S).
S
e∈Π

Hence one has the other direction

X
inf Pp (A(x, e, Π)) ≥ inf ϕp (x, S).
Π∈BE S
e∈Π

Next we recall a lemma from [4]. For a finite set Λ, let Λc denote its complement in V . Let Λn
denote the ball {y : d(x, y) ≤ n} of radius n centered at x, where d denotes the graph distance on
G.

Lemma 3.2 (Lemma 2.1 of [4]). For x ∈ V and ball Λn with n ≥ 1, one has
d 1
Pp (x ←→ Λcn ) ≥ · inf ϕp (x, S) · [1 − Pp (x ←→ Λcn )] (3.3)
dp p(1 − p) S⊂Λn ,x∈S

Proof of Theorem 2.9. By (2.12) and Lemma 2.10, it suffices to show p′′cut,E ≥ pc .
Suppose p′′cut,E < pc . Pick p0 , p1 such that p′′cut,E < p0 < p1 < pc .
By the definition of p′′cut,E and Proposition 3.1, there is a constant κ > 0 such that for any
p ∈ [p0 , p1 ],
inf ϕp (x, S) ≥ κ.
S

Write θx (p) := Pp (x ←→ ∞) and θx,n (p) := Pp (x ←→ Λcn ). By (3.3) one has that for p ∈ [p0 , p1 ],
′ (p)
θx,n κ
≥ .
1 − θx,n (p) p(1 − p)

Integrating this inequality from p0 to p1 , one has that

 κ  κ  κ
1 − p1 p0 1 − p1 p0 1 − p1 p0
θx,n (p1 ) ≥ 1 − · + θx,n (p0 ) · ≥ 1− · . (3.4)
p1 1 − p0 p1 1 − p0 p1 1 − p0
Letting n → ∞ one has that
 κ
1 − p1 p0
θx (p1 ) ≥ 1 − · > 0,
p1 1 − p0

which contradicts with the choice that p1 < pc . Hence p′′cut,E ≥ pc and we are done.

4 Percolation probability for subperiodic trees

To highlight the importance of Proposition 3.1, in this section we discuss some applications of it to
subperiodic trees.
For transitive graphs, Duminil-Copin and Tassion pointed out that inf{ϕp (x, S) : x ∈ S, S is finite} ≥
1 at p = pc . Using this they obtained a lower bound for percolation probability on transitive graphs
p−pc
with pc ∈ (0, 1): θ(p) ≥ p(1−p c)
for p ≥ pc . Here for transitive graphs, the percolation probability
θ(x, p) does not depend on x and we simply write it as θ(p). This lower bound can be extended to
0-subperiodic trees. We first adopt some notations and then recall the definitions of periodic and
subperiodic trees as in [12, Section 3.3].

9
Notation. Suppose T is an infinite, locally finite tree with a distinguished vertex o, called the
root of T . Write |x| for the graph distance from x to o; x ≤ y if x is on the shortest path from o to
y; x < y if x ≤ y and x 6= y; x → y if x ≤ y and |y| = |x| + 1 and in this case we call x the parent
of y; and T x for the subtree of T containing the vertices y ≥ x. For Bernoulli(p) percolation on the
tree T with root o, let θ(p) := Pp [o ←→ ∞] be the probability that o is in an infinite cluster.

Definition 4.1 (Definition on page 82 of [12]). Let N ≥ 0 be an integer. An infinite, locally

finite tree T with root x is called N -periodic (resp., N -subperiodic), if ∀x ∈ T there exists an
adjacency-preserving bijection (resp., injection) f : T x → T f (x) with |f (x)| ≤ N . A tree is periodic
(resp., subperiodic) if there is some N ≥ 0 for which it is N -periodic (resp., N -subperiodic).
P
Remark 4.2. For a 0-subperiodic tree T with root o, one has that inf Π e∈Π br(T )−|e| ≥ 1 (formula
(3.7) on page 85 of [12]). It is well-known that pc (T ) = 1/br(T ) for every locally finite infinite tree
T (for instance see [12, Theorem 5.15]). Hence
X X X
inf Ppc (A(o, e, Π)) = inf p|e|
c ≥ inf br(T )−|e| ≥ 1.
Π∈BE Π∈BE Π
e∈Π e∈Π e∈Π

Then by Proposition 3.1 one can set p0 = pc and κ = 1 in (3.4) and letting n → ∞ to get
p − pc
θ(p) = Pp (o ←→ ∞) ≥ , p ≥ pc
p(1 − pc )

for every 0-subperiodic tree T .

Actually this is true for all subperiodic trees with nontrivial pc .

Proposition 4.3. Consider Bernoulli percolation on a subperiodic tree T with pc (T ) < 1. Then
the lower right Dini derivative of the percolation probability θ(p) at pc belong to (0, ∞].
P
Proof. Theorem 3.8 in [12] says that inf Π e∈Π br(T )−|e| > 0 for a general subperiodic tree T with
pc (T ) < 1, where Π runs over all edge cutsets separating the root of T from infinity. Define

α(o, p) := inf{ϕp (o, S) : o ∈ S, S is finite}.

Then as before, one has that for a subperiodic tree T with root o and pc < 1,

Prop.3.1 X X X
α(o, pc ) = inf Ppc (A(o, e, Π)) = inf p|e|
c ≥ inf br(T )−|e| > 0.
Π∈BE Π∈BE Π
e∈Π e∈Π e∈Π

By the definition of ϕ in (3.1) it is obvious that α(o, p) is increasing in p. Setting p0 = pc in (3.4)

and letting n → ∞ one has that
 α(o,pc )
θ(p) − θ(pc ) 1−p pc
≥1− · .
1 − θ(pc ) p 1 − pc

This implies that the lower right Dini derivative of the percolation probability θ(p) at pc is positive:

θ(p) − θ(pc ) α(o, pc )(1 − θ(pc ))

D+ θ(pc) := lim inf ≥ > 0.
p→p+
c
p − pc pc (1 − pc )

10
 Without of the restriction of (sub)periodicity, it is possible to have the lower right Dini derivative
of the percolation probability being equal to zero. See the following two remarks.
Remark 4.4. It is easy to construct trees with the property that at p = pc ,

inf{ϕp (o, S) : o ∈ S, S is finite} = 0.

Indeed, we construct a spherically symmetric tree T with root o as follows. Let Tn denote the set
of vertices with graph distance n to the root o. If n = 2k for some k ≥ 0, let each vertex in Tn have
exactly one child; otherwise, let each vertex in Tn have exactly two children. Then it is easy to see
that
2n
|Tn | ≍ .
n
Here for two positive function f, g on Z+ , f (n) ≍ g(n) means that there exist constants c1 , c2 > 0
such that c1 g(n) ≤ f (n) ≤ c2 g(n) for all n > 0.
Hence br(T ) = lim inf n |Tn |1/n = 2 by Exercise 1.2 in [12]. Thus pc (T ) = 1/br(T ) = 1/2. Let
Sn be the ball of radius n and center x. Then at p = pc = 1/2,
1 c2
ϕp (o, Sn ) = |Tn+1 | · ≤ .
2n+1 n+1
Thus inf{ϕp (o, S) : o ∈ S, S is finite} = 0.
Remark 4.5. For the spherically symmetric tree T in Remark 4.4, one also has
1
2
θ(p) ≍ p − , p ≥ pc
2
θ(p)−θ(pc )
and in particular, the upper right Dini derivative D + θ(pc ) := lim supp→p+
c p−pc = 0.
−1 |e|
In fact, let c(e) = (1 − p) p be the conductance of edge e. Then formula (5.12) on page 142
of [12] is satisfied with P = Pp . Since T is spherically symmetric, for p > pc = 1/2, the effective
resistance is
X∞ X∞
(1 − p)n 1−p
R(o ←→ ∞) = (1 − p)p−n /|Tn | ≍ n
≍ .
n=1 n=1
(2p) (2p − 1)2
Then by Theorem 5.24 [12] one has
C (o ←→ ∞) 1 1
2
θ(p) ≍ = ≍ p− , p > pc .
1 + C (o ←→ ∞) 1 + R(o ←→ ∞) 2
Proposition 4.3 states that the lower right Dini derivative of the percolation probability on
a subperiodic tree with pc ∈ (0, 1) is positive at pc and it might equal to infinity in some cases
(Example 4.10). This leads us to the following question:
Question 4.6. What kind of subperiodic trees have the property that the right Dini derivatives
of θ(p) at pc are finite? What kind of subperiodic trees have the property that limp↓pc θ(p)−θ(p
p−pc
c)
∈
(0, ∞)?
The critical exponent β for Bernoulli percolation is characterized by θ(p) − θ(pc ) ≈ (p − pc )β .
5
For Z2 , it is conjectured that θ(p) − θ(pc) ≈ (p − pc )β for β = 36 [5, Table 10.1 on page 279], in
particular in this case the lower right Dini derivative at pc is infinite [9]. Indeed for site percolation
5
on the triangular lattice in the plane, one does have θ(p) − θ(pc ) = (p − pc ) 36 +o(1) [14, Theorem
1.1]. Question 4.6 asks what kind of subperiodic trees have β = 1.

11
 A partial answer for Question 4.6 is Theorem 4.7 which considers directed covers of strongly
connected graphs. An oriented graph G is called strongly connected if for any two vertices u, v
of G, there is a directed path in G from u to v. Suppose that G is a finite oriented graph and v is
any vertex in G. The directed cover of G based at v is the tree T whose vertices are the finite
paths of edges he1 , e2 , . . . , en i in G that start at v. We take the root of T to be the empty path and
we join two vertices in T by an edge when a path is an extension of the other path by one more
edge in G. Every periodic tree is a directed cover of a finite directed graph G; for a proof see pages
82-83 of [12]. Also not all periodic trees have finite right derivatives for θ(p) at pc ; see item 3 in
Example 4.10.

Theorem 4.7. If T is a periodic tree with root o and pc (T ) ∈ (0, 1) and it is the directed cover of
some strongly connected graph, then the right derivative of θ(p) exists at pc and this derivative is
positive and finite.

We first outline the ideas of the proof of Theorem 4.7. Suppose T is the directed cover of
some finite strongly connected directed graph G = (V, E) based at some vertex v1 ∈ V , where
V = {v1 , . . . , vn }. Let θi (p) be the percolation probability of the tree Ti that is the directed cover
of G based at vi . Then these quantities θi (p) are related by a family of algebraic equations (4.1).
We will then proceed as follows:

1. We first show that θi (p) are continuous at pc (Ti ) (Lemma 4.8).

2. Then we show that these trees Ti have the same critical probability pc and the upper right
Dini derivatives of θi (p) are finite at pc (Lemma 4.9).

3. We finish the preparation by showing that the right derivatives of θi (p) at pc exist via a
convexity/concavity argument in Lemma 4.12. The proof of Lemma 4.12 is a little bit involved
and we will need two more lemmas for its proof:

• the functions θi (p) are analytic on (pc , 1) (Lemma 4.13) and

• the uniqueness of θi (p) as solutions of (4.1) (Lemma 4.14).

4. We then finish the proof of Theorem 4.7: the existence is from Step 3 (Lemma 4.12), the
positiveness is from Proposition 4.3 and the finiteness is from Step 2 (Lemma 4.9).

Lemma 4.8. Suppose T is a periodic tree with root o and pc (T ) < 1. Then θ(pc ) = 0.
1 −1 |x|
Proof. At p = pc = br(T ) , if we put conductance c(e(x)) = (1 − pc ) pc , where e(x) is the edge
from x to its parent, then (5.12) on page 142 of [12] is satisfied. As noted on page 142 line 17 of
[12], these conductances correspond to the homesick random walk RWbr(T ) . If we put resistance
Φ(e(x)) = λ|x|−1 for the edge e(x) instead, then it is known that as λ ↑ λ∗ = p1c , the effective
resistance from the root to infinity of the corresponding network is tend to infinity [11, Theorem
5.1]. This implies that the homesick random walk RWbr(T ) is recurrent. Hence by Corollary 5.25
of [12] we know θ(pc ) = 0.

Now we restrict to a subset of periodic trees that are directed covers of finite strongly connected
oriented graphs.

Lemma 4.9. Suppose G = (V (G), E(G)) is a finite, strongly connected directed graph and V (G) =
{v1 , . . . , vn }. Let λ∗ be the largest eigenvalue of the adjacency matrix AG of G. Let Ti be the
directed cover of G based at vi and denote its root by oi . Then pc (T1 ) = · · · = pc (Tn ) = λ1∗ .

12
 Moreover if λ∗ > 1, then the upper right Dini derivative of θi (p) at pc is finite for every
i ∈ {1, · · · , n}, where θi (p) := Pp [oi ←→ ∞ in Ti ] denotes the probability that the root oi of Ti is in
an infinite open cluster.

Proof. The first part is a standard result. See the discussion on pages 83-84 of [12] for example.
Since G is strongly connected, AG is irreducible. Hence by the Perron-Frobenius theorem (e.g.
see [6, Theorem 8.4.4]) there is a left λ∗ -eigenvector v∗ = (v1 , . . . , vn ) all of whose entries are
positive. We also normalize v∗ such that its l2 -norm is 1.
Since oi 6↔ ∞ in Ti if and only if oi can’t connect to infinity via any of its children, we have the
following relations for these percolation probabilities:
n
Y
1 − θi (p) = [1 − pθj (p)]aij , i ∈ {1, . . . , n},
j=1

i.e.,
n
Y
θi (p) = 1 − [1 − pθj (p)]aij , i ∈ {1, . . . , n}, (4.1)
j=1

where aij is the (i, j)-entry of the matrix AG , i.e., the number of directed edges in G from vertex
vi to vj .
Denote by θmax (p) = max{θ1 (p), · · · , θn (p)}. Since G is strongly connected, there exists M > 0
such that there is a directed path with length at most M from vi to vj for any pair vi , vj ∈ V (G).
Hence θi (p) ≥ pM θmax (p) for all i = 1, . . . , n. Thus for p > pc ,

0 < θi (p) ≍ θmax (p), i ∈ {1, . . . , n}, . (4.2)

By Lemma 4.8 and the right continuity of θi (p) (e.g., see [12, Exercise 7.33]), one has that

0 < θi (p) = o(p − pc ), 0 < p − pc ≪ 1. (4.3)

Using (4.3) when 0 < p − pc ≪ 1 and i ∈ {1, . . . , n}, we can rewrite (4.1) as
n
X n 
X  X
2 aij 2
θi (p) = p aij θj (p) − p θj (p) − p2 2
aij aik θj (p)θk (p) + θmax (p) · o(1), (4.4)
2
j=1 j=1 j6=k


where we use the convention that a2ij = 0 if aij = 0, 1.
Multiplying vi on both sides of (4.4) and adding them up, one has that
n
X n
X n
X
2
vi θi (p) = p vi aij θj (p) + θmax (p) · o(1)
i=1 i=1 j=1
X X
n n  
aij 2 X 
−p2 vi θj (p) + aij aik θj (p)θk (p) (4.5)
2
i=1 j=1 j6=k

Since pc (Ti ) = λ1∗ < 1, there exists some i such that either aij ≥ 2 for some j or aij aik ≥ 1 for
some j 6= k. Therefore by (4.2) and (4.5) there exists c > 0 such that
n
X n
X n
X
vi θi (p) ≤ p vi aij θj (p) − cp2 θmax
2
(p), 0 < p − pc ≪ 1. (4.6)
i=1 i=1 j=1

13
 Since v∗ is a left λ∗ -eigenvector of AG , one has that
n
X n
X n
X
vi aij θj (p) = v∗ AG θ(p) = λ∗ v∗ · θ(p) = λ∗ vi θi (p), (4.7)
i=1 j=1 i=1

where θ(p) = (θ1 (p), · · · , θn (p))T

is the vector of percolation probabilities in Rn .
Plugging (4.7) into (4.6) and using λ∗ = p1c and (4.2) one has that for 0 < p − pc ≪ 1,
n
p − pc X
cp2 θmax
2
(p) ≤ vi θi (p) ≤ c′ θmax (p)(p − pc ),
pc
i=1

for some constant > 0. This implies that θi (p) ≤ θmax (p) ≤ c′′ (p − pc ) for 0 < p − pc ≪ 1 and
c′
then we have the desired result on the upper right Dini derivative.
The strongly connectedness of G is needed for the finiteness of the right Dini derivative. See
the following example.

v1 v2

v1 v1 v2

v4 v3

Ga Gb Gc

Figure 3: Directed graphs Ga , Gb , Gc from left to right.

Example 4.10. Let Ga , Gb , Gc be as illustrated in Figure 3. Let Ts , s ∈ {a, b, c} be the directed

cover of Gs based at v1 (Gs ).
1 1
1. The tree Ta is a binary tree with root o. It is easy to see that pc (Ta ) = 2 and for p ≥ 2,
θ(p) = 2p−1
p2
′ (p ) = 8 and θ(p) is concave on (p , 1).
. In this case θ+ c c

2. The tree Tb is a Fibonacci tree with root o and deg(o) = 2. See Figure 3.2 √on page 83 of
[12] for an illustration of the Fibonacci tree. It is easy to see that pc (Tb ) = 5−1
2 . Writing
2 +p−1
√
θ(p) = Pp [o ←→ ∞] and using (4.1) one has that θ(p) = p p3
for p ≥ 5−1
2 . Hence in this
√
′ (p ) = 5 + 5 and θ(p) is also concave on (p , 1).
case θ+ c c
√
5−1
3. The tree Tc also has pc (Tc ) = 2 . Actually if we define Ti , θi (p) as in Lemma 4.9, then it is
√
5−1
easy to check that pc (Ti ) = 2 , i = 1, 2, 3, 4. Solving (4.1) one can find that for p ∈ (pc , 1),
 √
 (1−2p)(p2 +p−1)+ (p2 +p−1)(−3p2 +5p−1)

 θ1 (p) = 2p2 (1−p2 )





 √

 p2 +p−1+ (p2 +p−1)(−3p2 +5p−1)
 θ2 (p) = 2p2
(4.8)



 θ3 (p) = p2 +p−1

 p2





 p2 +p−1
θ4 (p) = p3

14
 √
In particular, θ(p) = θ1 (p) = Θ( p − pc ) for 0 < p−pc ≪ 1 and thus the right Dini derivative
at pc is infinite. One can also check that θ1 (p) and θ2 (p) are concave on (pc , 1).

Example 4.10 and the fact that θ(p) − θ(pc ) ≈ (p − pc )5/36 on the triangle lattice on the plane
[14, Theorem 1.1] lead us to the following question:

Question 4.11. For a transitive graph or a periodic tree with root o, is the percolation probability
θ(p) concave on (pc , 1)?

We now proceed to show the right derivatives of θi (p) exist at pc .

Lemma 4.12. The right derivative of θi (p) exists at pc for all i ∈ {1, . . . , n}.

As mentioned earlier, the proof of Lemma 4.12 is somewhat long and we begin by showing that
θi (p) is analytic on (pc , 1) for all i ∈ {1, . . . , n}.

Lemma 4.13. Under the same assumptions as Lemma 4.9 one has that θi (p) is analytic on (pc , 1)
for all i ∈ {1, · · · , n}.

Proof. Recall that the percolation probabilities satisfy (4.1):

n
Y
θi (p) = 1 − [1 − pθj (p)]aij , i ∈ {1, . . . , n},
j=1

where aij is the number of directed edges in G from vertex vi to vj .

Define fi : R1+n → R for i ∈ {1, . . . , n} by
n
Y
T
fi ((x, y1 , . . . , yn ) ) = yi − 1 + [1 − xyj ]aij .
j=1

Write f = (f1 , · · · , fn )T . By (4.1), we know that (p, θ1 (p), . . . , θn (p)) is a positive solution of f = 0
when p > pc .
Note that when i 6= j,

∂fi Y (4.1) −paij (1 − θi (p))

(p, θ1 (p), . . . , θn (p)) = −paij [1 − pθj (p)]aij −1 · [1 − pθj ′ (p)]aij′ =
∂yj 1 − pθj (p)
′j 6=j

and
∂fi Y (4.1) paii (1 − θi (p))
(p, θ1 (p), . . . , θn (p)) = 1 − paii [1 − pθi (p)]aii −1 · [1 − pθj ′ (p)]aij′ = 1 − .
∂yi 1 − pθi (p)
j 6=i
′

 ∂fi 
Therefore the Jacobi matrix J = ∂yj (p, θ1 (p), . . . , θn (p)) 1≤i,j≤n can be written as

J = I − BC (4.9)

where I is the identity matrix and B is a diagonal matrix with bii = 1 − θi (p) and C is a matrix
pa
with (i, j)-entry cij = 1−pθijj (p) .
Notice that Bernoulli(p) percolation on Ti can also be viewed as a multi-type Galton–Watson
tree Z. Each vertex u on the tree corresponds to a directed path on G. If the endpoint of the path
is vj , then we say that u has type j. In particular, we view the root of Ti is of type i. The number of

15
type j children of a type i vertex has Binomial distribution Bin(aij , p). The percolation probability
θi (p) is just the non-extinction probability for such a n-type Galton–Watson tree started with a
single type i vertex. Let Ps and Es denote the probability measure and corresponding expectation
for such an n-type Galton-Watson tree started with a single ancestor with type s ∈ {1, . . . , n}.
Now let Ext denote the event that the n-type Galton–Watson tree is extinct. Then Pi [Ext] =
1 − θi (p). Let Z1j denote the number of children of type j of Z0 . For a nonnegative integer sequence
(t1 , . . . , tn ) with ti ≤ aij , one has that
 
Pi Z1j = tj , j = 1, 2, . . . , n Ext
Yn  
1 aij tj
= · p (1 − p)aij −tj · (1 − θj (p))tj
1 − θi (p) tj
j=1
Yn  
1 aij
= · (1 − p)aij −tj · (p − pθj (p))tj
1 − θi (p) tj
j=1
  1 − p aij −tj  p − pθj (p) tj
n
(4.1) Y aij
= ·
tj 1 − pθj (p) 1 − pθj (p)
j=1
n  
Y aij  p − pθj (p) tj  p − pθj (p) aij −tj
= · 1− . (4.10)
tj 1 − pθj (p) 1 − pθj (p)
j=1

By [7] we know conditioned on extinction, the n-type Galton-Watson tree is still a multi-type
Galton–Watson tree. Let Pe s and E
e s denote the probability measure and corresponding expectation
for the n-type Galton-Watson tree started with a single ancestor with type s conditioned on
extinction. By (4.10), conditioned on extinction, the number of type j children of a type i vertex
p−pθ (p)
has Binomial distribution Bin(aij , 1−pθjj (p) ). Hence the mean offspring matrix M has (i, j)-entry
p−pθ (p) paij
mij = aij 1−pθjj (p) = (1 − θj (p)) · 1−pθj (p) . Observe that

M = CB (4.11)

Let q := max1≤j≤n [1 − θj (p)] be the maximum of the extinction probability. For p > pc , we know
q < 1. Let Zk denote the size of k-th generation of the multi-type Galton-Watson tree. As the last
displayed inequality on page 547 of [7], one has that
1  
e s [Zk ] ≤
E · Es Zk q Zk → 0 as k → ∞.
1 − θs (p)

Hence the largest eigenvalue λ1 (M ) for the mean offspring matrix M satisfies λ1 (M ) < 1. By
[6, Theorem 1.3.22], the largest eigenvalue of BC satisfies that λ1 (BC) = λ1 (CB) = λ1 (M ) < 1.
Therefore by (4.9) the Jacobi matrix J is invertible for p ∈ (pc , 1). Hence by the analytic implicit
function theorem, we obtain that the functions θi (p) are analytic on (pc , 1).

Lemma 4.14. The solution (p, θ1 (p), . . . , θn (p)) of (4.1) in (pc , 1) × (0, 1)n is unique.

The following Proposition 4.15 will be needed for Lemma 4.14.

For p ∈ [0, 1], we define the operator Bp : [0, 1]n −→ [0, 1]n as given by (4.1):
n
Y
Bp (α)i = 1 − [1 − pαj ]aij , (4.12)
j=1

16
where α = (α1 , · · · , αn )T ∈ [0, 1]n . For example, Bp (0) = 0.
For α, β ∈ [0, 1]n , write α ≤ β if αi ≤ βi for all i ∈ {1, · · · , n} and write α ≺ β if α ≤ β and
α 6= β.

Proposition 4.15. We have the following properties for the operator Bp .

(a) The operator Bp is increasing in the sense that if α ≤ β, then Bp (α) ≤ Bp (β).

(b) Moreover, if p ∈ (0, 1), then Bp is strictly increasing in the sense that if α ≺ β, then
Bp (α) ≺ Bp (β).

(c) If 0 ≤ p1 < p2 ≤ 1 and α ∈ [0, 1]n , then Bp1 (α) ≤ Bp2 (α).

(d) Moreover, if α ∈ [0, 1]n and α 6= 0, then for 0 ≤ p1 < p2 < 1, one has that Bp1 (α) ≺ Bp2 (α).

(e) For p > 0, if 0 6= α ∈ [0, 1]n is a fixed point of Bp , i.e., Bp (α) = α, then αi > 0 for all
i ∈ {1, · · · , n}.

Proof of Proposition 4.15. The items (a) and (c) are obvious from the definition of Bp . For item
(b), suppose αj < βj for some j ∈ {1, · · · , n}. Since G is strongly connected, there exists some i
such that aij ≥ 1. Then
[1 − pαj ]aij > [1 − pβj ]aij
and [1 − pαj ′ ]aij′ ≥ [1 − pβj ′ ]aij′ ≥ [1 − p]aij′ > 0 for j ′ 6= j. Therefore Bp (α)i < Bp (β)i . Together
with item (a) we know Bp (α) ≺ Bp (β).
For item (d), the proof is similar to item (b) and we omit it.
For item (e), if aij ≥ 1, by (4.12) and the fact that α is a fixed point,

αi = Bp (α)i ≥ 1 − [1 − pαj ]aij ≥ 1 − [1 − pαj ] = pαj .

Repeating this argument, we get

αi = Bp (α)i ≥ pM αj ′ , ∀ j ′ ∈ {1, · · · , n},

where M is the maximum of the lengths of the shortest oriented paths connecting two points in G.
Therefore since αj > 0 for some j, then αi > 0 for all i ∈ {1, · · · , n}.

Proof of Lemma 4.14. Write θi,k (p) := Pp [oi is connected to level k of Ti ], where by level k we
mean the set of vertices in Ti with graph distance k to the root. Let θk (p) = (θ1,k (p), · · · , θn,k (p))T ∈
[0, 1]n . In particular, θ0 (p) = 1. Then as (4.1), one has that

θk+1 (p) = Bp (θk (p))

and thus θk (p) = Bp◦k (1). By the definition of θi,k (p), θi (p),

θ(p) = lim θk (p) = lim Bp◦k (1).

k→∞ k→∞

Suppose α ∈ [0, 1]n is some fixed point of Bp , i.e., Bp (α) = α. Then α ≤ 1. By item (a) of
Proposition 4.15, one has that α = Bp (α) ≤ Bp (1) = θ1 (p), and then α = Bp (α) ≤ Bp (θ1 (p)) =
θ2 (p), · · · In the end, we have
α ≤ lim θk (p) = θ(p),
k→∞

i.e., θ(p) is the largest fixed point for the operator Bp in [0, 1]n .

17
 Now suppose p ∈ (pc , 1) and we have some solution α ∈ [0, 1]n \{0} for (4.1), i.e., α is a
nonzero fixed point of Bp in [0, 1]n . Since α ∈ [0, 1]n is a fixed point of Bp , we have showed
that α ≤ θ(p). Since θ(p) ≥ α 6= 0, by item (e) in Proposition 4.15, α ∈ (0, 1)n . Define
p1 := sup{t ≥ pc : θ(t) ≤ α}.
Since α ≤ θ(p) and limt↓pc θ(t) → 0 (Lemma 4.8), one has that p1 ∈ (pc , p]. Since θ(p) is
infinite differentiable (Lemma 4.13) and increasing in (pc , 1), one has that θ(p1 ) ≤ α and for some
i ∈ {1, . . . , n}, θi (p1 ) = αi .
Since θ(p1 ) is a fixed point of Bp1 and α is a fixed point of Bp , one has that
n
Y n
Y
θi (p1 ) = 1 − [1 − p1 θj (p1 )]aij = αi = 1 − [1 − pαj ]aij . (4.13)
j=1 j=1

Since G is strongly connected, there is some j such that aij ≥ 1. As in the proof of item (e) in
Proposition 4.15, to satisfy (4.13), by p1 ≤ p and θ(p1 ) ≤ α one must have
1 − p1 θj (p1 ) = 1 − pαj .
Since 0 < θj (p1 ) ≤ αj ≤ θj (p) and pc < p1 ≤ p < 1, one must have p1 = p and αj = θj (p1 ).
Therefore by p1 = p and the continuity of θ one has the other direction α ≥ limt↑p1 θ(t) = θ(p1 ) =
θ(p). Hence α = θ(p) is the unique solution of (4.1) in (0, 1)n for p ∈ (pc , 1).

Now we are ready to prove Lemma 4.12.

Proof of Lemma 4.12. First by (4.1) and Lemma 4.14 we know that the set
{(p, θ1 (p), . . . , θn (p)) ∈ (pc , 1) × (0, 1)n }
is semi-algebraic (Definition 2.1.4 of [2]). By Theorem 2.2.1 of [2], its projection S := {(θ1 (p), . . . , θn (p)) : p ∈
(pc , 1)} is also semi-algebraic set and by Definition 2.2.5 the map p 7→ (θ1 (p), . . . , θn (p)) is a semi-
algebraic map from (pc , 1) to S in view of (4.1).
By Theorem 2.2.1 and Proposition 2.2.6 of [2], the maps p 7→ θi (p) are also semi-algebraic
functions on (pc , 1).
By Lemma 4.13 we already know that the functions p 7→ θi (p) are infinitely differentiable on
(pc , 1). Hence by Proposition 2.9.1 of [2], we know the second derivatives θi′′ (p) are also semi-
algebraic functions on (pc , 1), i.e., the sets {(p, θi′′ (p)) : p ∈ (pc , 1)} are semi-algebraic sets for all
i ∈ {1, . . . , n}.
Hence for every i ∈ {1, . . . , n}, {(p, θi′′ (p)) : p ∈ (pc , 1), θi′′ (p) = 0} is a semi-algebraic set since it
is the intersection of two semi-algebraic sets, {(p, θi′′ (p)) : p ∈ (pc , 1)} and {(p, 0) : p ∈ (pc , 1)}. Thus
by Theorem 2.2.1 of [2] for every i ∈ {1, . . . , n}, the projection {p : p ∈ (pc , 1), θi′′ (p) = 0} is a semi-
algebraic set. By Proposition 2.1.7 of [2], for every i ∈ {1, . . . , n}, the set {p : p ∈ (pc , 1), θi′′ (p) = 0}
is a finite union of points and open intervals. Hence there exists some εi > 0 such that θi′′ (p)
cannot change its sign on (pc , pc + εi ). Taking ε = min{εi : i = 1, 2, . . . , n} we have that the
function θi′′ (p) does not change its sign (i.e. remains nonnegative or nonpositive) on (pc , pc + ε) for
all i ∈ {1, 2, . . . , n}.
By Lemma 4.12 and the right continuity of the functions θi (p) at pc , we know these functions
θi (p) are either convex or concave on [pc , pc + ε). Hence limp↓pc θi (p)−θ p−pc
i (pc )
= limp↓pc θip−p
(p)−0
c
exists,
i.e., the right derivative of θi (p) at pc exists.

Proof of Theorem 4.7. The existence of the right derivative of θ(p) follows from Lemma 4.12. The
positiveness and finiteness of the right derivatives follow from Proposition 4.3 and Lemma 4.9
respectively.

18
5 Concluding remarks and questions
5.1 Remark on Bernoulli site percolation
For p ∈ [0, 1], if we instead keep each vertex with probability p and remove it otherwise. Call the
vertices kept open vertices and those removed closed vertices. Bernoulli(p) site percolation
studies the random subgraph ξ of G induced by the open vertices. For Bernoulli site percolation,
an edge is call open if and only if its two endpoints are open. When talking about Bernoulli site
percolation, we will use Psite
p and Psite
p to stress that.
Remark 5.1. If the connected graph G has bounded degree, say, with a upper bound D, then for
Bernoulli site percolation, the following analogue of Lemma 3.2 holds:
d site 1 inf S⊂Λn ,x∈S ϕp (x, S)
Pp (x ←→ Λcn ) ≥ min(1, ) · [1 − Psite c
p (x ←→ Λn )] (5.1)
dp 1−p D−1
P P S
where ϕp (x, S) := y∈S z ∈S,(y,z)∈E
/ Psite ′
p [x ←→ y]. If one defines pcut,site accordingly for site
percolation, one can prove pc,site = p′cut,site similarly as the bond percolation case.
This leads us to the following conjecture:
Conjecture 5.2. The answers for Question 1.3 and 1.5 are also positive for Bernoulli site perco-
lation.

5.2 Is there an example with pcut,E < pcut,V ?

We have seen examples with pT,E < pT,V (Example 2.3). One can ask the same question for pcut,E
and pcut,V :
Question 5.3. Is there a locally finite, connected, infinite graph G such that pcut,E < pcut,V ?
In view of Lemma 2.1, if there is a graph G with pcut,E < pcut,V , then it must have unbounded
degree and for a vertex cutset ΠV , for “most” v ∈ ΠV there should be a lot of edges in the
corresponding edge cutset ΠE = ∆S(ΠV ) incident to v. One might first want to consider certain
1-dimensional multigraphs. However there is no simple 1-dimensional example; see Proposition 5.5.
Definition 5.4. Let (an )n≥0 be a sequence of positive S integers. Let G = G((an )n≥0 ) be the graph
with vertex set V = N = {0, 1, 2, · · · } and edge set E = n En , where En = {en,j : j = 1, . . . , an } is
the set of an parallel edges from n to n + 1.
Proposition 5.5. There is no sequence of positive integers (an )n≥0 such that G = G((an )n≥0 ) has
the property of pcut,E (G) < pcut,V (G).
We need a technical lemma.
Lemma 5.6. There is no sequence of positive integers (an )≥0 that satisfies both
∞
X
(1 − p2 )ai = ∞ (5.2)
i=0

and
c
an ≥ Qn−1   , ∀ n ≥ 1. (5.3)
ai
i=0 1 − (1 − p1 )
for some constants 0 < p1 < p2 < 1 and c > 0.

19
Proof of Proposition 5.5 assuming Lemma 5.6. Notice that
n−1
Y  
Pp [0 ←→ n] = 1 − (1 − p)ai (5.4)
i=0

Thus
n−1
Y ∞
X
 
p < pc ⇒ lim 1 − (1 − p)ai = 0 ⇔ (1 − p)ai = ∞ (5.5)
n→∞
i=0 i=0
and
∞
X n−1
Y 
(1 − p)ai < ∞ ⇔ lim 1 − (1 − p)ai > 0 ⇒ p ≥ pc (5.6)
n→∞
i=0 i=0

Notice that the minimal vertex cutsets are {Πn } where Πn = n and the minimal edge cutsets
are En = {en,j : j = 1, · · · , an }, the an parallel
Qn−1 edges
 from n to n + 1.
Hence Ep [|C(0) ∩ Πn |] = Pp [0 ←→ n] = i=0 1 − (1 − p)ai . Thus pcut,V = pc .
Suppose there is some sequence (an )n≥0 such that pcut,E < pcut,V . Pick p1 , p2 such that pcut,E <
p1 < p2 < pcut,V .
Since p2 < pcut,V = pc , by (5.5) one has that
∞
X
(1 − p2 )ai = ∞
i=0

Since we choose p1 > pcut,E and noting that Ep [|C(0) ∩ En |] = pan Pp [0 ←→ n], one has that
n−1
Y  
inf p1 an 1 − (1 − p1 )ai > 0 (5.7)
n
i=0

i.e., there exists c > 0 s.t.

c
an ≥ Qn−1   , ∀ n ≥ 1.
ai
i=0 1 − (1 − p1 )
But this contradicts with Lemma 5.6 and hence Proposition 5.5 holds.

Proof of Lemma 5.6. First we reduce to the case of increasing sequence. If there is some sequence
satisfies both (5.2) and (5.3) for some 0 < p1 < p2 < 1, then
n−1
Y n−1
Y
 ai
   (5.2)
1 − (1 − p1 ) ≤ 1 − (1 − p2 )ai → 0 as n → ∞.
i=0 i=0

Thus by (5.3), one has that an → ∞. In the following for simplicity we write p = p1 .
Now we consider the sequence (a′n ), the rearrangement of an in the non-decreasing order. Ob-
viously, (a′n ) also satisfies (5.2). As for (5.3), let m = m(n) be the last index such that am ≤ a′n ,
i.e., m = max{k : ak ≤ a′n }. Obviously m ≥ n. Since ak → ∞ as k → ∞, m < ∞. Then we claim
that there exists a constant c0 > 0 such that
n−1
Y m−1
Y
 ′  (5.3)
a′n 1 − (1 − p)ai ≥ c0 am 1 − (1 − p)ai ≥ cc0 , (5.8)
i=0 i=0

Write A = {vi : v1 < v2 < · · · } for the all the values of the sequence (an ). For each v ∈ A, let
N (v) = |{j : aj = v}| ≥ 1 be the number of times the sequence taking the value v.

20
 Case one: am = a′n . By the definition of a′n , we assume a′n = vk and then
n−1
Y  ′  N (vk )−1 k−1
Y N (vi )
1 − (1 − p)ai ≥ 1 − (1 − p)vk × 1 − (1 − p)vi (5.9)
i=0 i=1

By the choice of m, the multi-set {a0 , · · · , am−1 } contains at least N (vk )− 1’s vk and all the N (vi )’s
vi for i < k. Hence
k
Y m−1
Y
 
vk −1
 
vi N (vi )
 
1 − (1 − p) × 1 − (1 − p) ≥ 1 − (1 − p)ai . (5.10)
i=1 i=0

By a′n = am and the above two inequalities (5.9),(5.10) we have (5.8) for all c0 ≤ 1.
Case two: am < a′n , say a′n = vk and am = vj for some j < k. Then by the choice of m, the
multi-set {a0 , · · · , am−1 } contains at least N (vj ) − 1’s vj and all the other N (vi )’s vi for i ≤ k, i 6= j.
Hence
Yk m−1
Y
 
vj −1
 
vi N (vi )

1 − (1 − p) × 1 − (1 − p) ≥ 1 − (1 − p)ai . (5.11)
i=1 i=0

By (5.9) and (5.11) we have that

Qn−1  a′i
  −1
a′ni=0 1 − (1 − p) vk 1 − (1 − p)vk vk
Qm−1  ≥  −1 ≥ p ≥ p.
a
am i=0 1 − (1 − p) i v
vj 1 − (1 − p) j vj

where in the second inequality we use the fact that

x x
f (x) = x
∈ (x, ], x ≥ 1.
1 − (1 − p) p

Hence in this case (5.8) holds with c0 = p = p1 .

Combining the two cases one has that (5.8) holds with c0 = p = p1 .
By the reduction, we can assume (an )n≥0 is increasing. Thus there is a strictly increasing
sequence (nj ) such that for n ∈ [nj , nj+1 − 1], an = vj . In particular, (5.2) becomes
∞
X
(nj+1 − nj )(1 − p2 )vj = ∞ (5.12)
j=1

and (5.3) becomes (only needs to look at times nj+1 − 1)

 
c 1 − (1 − p1 )vj
vj ≥ Q j   (5.13)
vi ni+1 −ni
i=1 1 − (1 − p1 )

By (5.13) one has that

j−1
Y
c n −n
  n −n −1 ≤ 1 − (1 − p1 )vi i+1 i ≤ 1.
vj 1 − (1 − p1 )vj j+1 j
i=1

Hence  n −n −1
vj 1 − (1 − p1 )vj j+1 j ≥ c.

21
Taking logarithm one has that

log vj + (nj+1 − nj − 1) log[1 − (1 − p1 )vj ] ≥ log c

Hence
log c − log vj c′ log vj
nj+1 − nj − 1 ≤ ≤ , when vj > 1.
log[1 − (1 − p1 )vj ] (1 − p1 )vj
But this contradicts with (5.12): (noting {vj } is a strictly increasing subsequence of N and 1 − p2 <
1 − p1 )
∞
X ∞
X ∞
X c′ log vj
(nj+1 − nj )(1 − p2 )vj ≤ (n2 − n1 )(1 − p2 )v1 + (1 − p2 )vj + (1 − p2 )vj < ∞.
(1 − p1 )vj
j=1 j=2 j=2

This contradiction implies Lemma 5.6.

Acknowledgments
We thank Russ Lyons for many helpful discussions and support. Most of the work was done
while the author was a graduate student at Indiana University. We also thank Wai-Kit Yeung for
discussions and the reference on semi-algebraic functions. We are also grateful to the anonymous
referee for his/her careful reading and helpful comments which improve the presentation a lot.

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

[2] Jacek Bochnak, Michel Coste, and Marie-Françoise Roy. Real algebraic geometry, volume 36
of Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related
Areas (3)]. Springer-Verlag, Berlin, 1998. Translated from the 1987 French original, Revised
by the authors.

[3] Hugo Duminil-Copin and Vincent Tassion. A new proof of the sharpness of the phase transition
for Bernoulli percolation and the Ising model. Comm. Math. Phys., 343(2):725–745, 2016.

[4] Hugo Duminil-Copin and Vincent Tassion. A new proof of the sharpness of the phase transition
for Bernoulli percolation on Zd . Enseign. Math., 62(1-2):199–206, 2016.

[5] Geoffrey Grimmett. Percolation, volume 321 of Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, sec-
ond edition, 1999.

[6] Roger A. Horn and Charles R. Johnson. Matrix analysis. Cambridge University Press, Cam-
bridge, second edition, 2013.

[7] Peter Jagers and Andreas N. Lagerås. General branching processes conditioned on extinction
are still branching processes. Electron. Commun. Probab., 13:540–547, 2008.

[8] Jeff Kahn. Inequality of two critical probabilities for percolation. Electron. Comm. Probab.,
8:184–187, 2003.

22
 [9] Harry Kesten and Yu Zhang. Strict inequalities for some critical exponents in two-dimensional
percolation. J. Statist. Phys., 46(5-6):1031–1055, 1987.

[10] Russell Lyons. The Ising model and percolation on trees and tree-like graphs. Comm. Math.
Phys., 125(2):337–353, 1989.

[11] Russell Lyons. Random walks and percolation on trees. Ann. Probab., 18(3):931–958, 1990.

[12] Russell Lyons and Yuval Peres. Probability on Trees and Networks, volume 42 of Cambridge
Series in Statistical and Probabilistic Mathematics. Cambridge University Press, New York,
2016. Available at http://rdlyons.pages.iu.edu/.

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

[14] Stanislav Smirnov and Wendelin Werner. Critical exponents for two-dimensional percolation.
Math. Res. Lett., 8(5-6):729–744, 2001.

23