betaFeedback

WorksMathematics

No infinite cluster at criticality for long-range percolation with fast-decaying interactions

Abstract

We consider independent long-range bond percolation on Zd\mathbb{Z}^d in which xx and yy are joined with probability 1−e−βJ(x−y)1-e^{-βJ(x-y)}. Let d≥2d\ge2 and let JJ be invariant under coordinate permutations and sign changes, positive on nearest-neighbour pairs, and satisfy lim inf⁡n→∞nd∑∥x∥>nJ(x)=0\liminf_{n\to\infty} n^d\sum_{\|x\|>n}J(x)=0. We prove that there is no infinite cluster at the critical parameter. Together with Berger's theorem for heavy tails, this gives continuity of the phase transition for J(x)≍∥x∥−d−αJ(x)\asymp\|x\|^{-d-α} for every α>0α>0 except α=dα=d, in every dimension d≥2d\ge2, and for every exponentially decaying kernel satisfying these hypotheses. Previously the fast-decaying case was known for nearest-neighbour and uniform finite-range kernels, in d=2d=2 for finite-range and exponential kernels, and in high dimension for sufficiently spread-out kernels. The proof adapts a recent renormalization argument for nearest-neighbour percolation on Z3\mathbb{Z}^3 that is driven by a gluing inequality conjectured by Kozma and Nitzan. Long edges enter in two places: they can jump over the surfaces that a nearest-neighbour path must cross, and they let a cluster outside a box enter it from far away. We control the first with thick outer faces and the tail condition, and the second by measuring entrances through the total edge weight between a cluster and a box rather than through the number of boundary vertices.

Introduction

Long-range percolation and the main result

Let d≥1d \ge1 and let J:Zd∖{0}→[0,∞)J:\mathbb{Z}^d \setminus\{0\}\to[0,\infty) be a kernel with J(x)=J(−x)J(x)=J(-x) and ∑xJ(x)<∞\sum_x J(x)<\infty. In long-range percolation at parameter β>0\beta>0, each unordered pair {x,y}\{x,y\} of distinct vertices of Zd\mathbb{Z}^d is open independently with probability

pxy(β)=1−e−βJ(x−y),p_{xy}(\beta)=1-e^{-\beta J(x-y)},

and closed otherwise; pairs with J(x−y)=0J(x-y)=0 are never open. Write Pβ\mathbb{P}_{\beta} for the law, θ(β)=Pβ(0↔∞)\theta(\beta)=\mathbb{P}_{\beta}(0\leftrightarrow\infty) for the probability that the origin lies in an infinite open cluster, and βc=inf⁡{β:θ(β)>0}\beta_c=\inf\{\beta:\theta(\beta)>0\}. The function θ\theta is nondecreasing, and the basic question is whether it is continuous at βc\beta_c, that is, whether θ(βc)=0\theta(\beta_c)=0.

For nearest-neighbour percolation, θ(pc)=0\theta(p_c)=0 was proved in d=2d=2 by Harris and Kesten [11, 14] and in high dimensions by the lace expansion [10, 7]. Kozma and Nitzan [15] reduced the remaining dimensions to a correlation inequality for independent percolation on finite graphs, which has since been proved in several forms [16, 17, 4, 8]. We use it in the joint form of Theorem 2.1, which holds for arbitrary edge probabilities; all graphs to which we apply it are finite.

For long-range models the answer depends on the tail of JJ. Writing

T(n)=∑∥x∥>nJ(x),∥x∥=max⁡i∣xi∣,T(n)=\sum_{\lVert x\rVert>n}J(x),\qquad\lVert x\rVert=\max_i |x_i|,

we use the following hypotheses.

(Sym) JJ is invariant under coordinate permutations and sign changes of coordinates.

(Pos) J(e1)>0J(e_1)>0, where e1e_1 is the first unit vector.

(Tail) lim inf⁡n→∞ndT(n)=0\liminf_{n\to\infty} n^dT(n)=0.

Theorem 1.1. Let d≥2d\ge2 and let JJ satisfy (Sym), (Pos) and (Tail). Then 0<βc<∞0<\beta_c<\infty and Pβc(0↔∞)=0\mathbb{P}_{\beta_c}(0\leftrightarrow\infty)=0.

Combined with Berger’s theorem for heavy-tailed kernels [3 Theorem 2.6], which we recall in § 8, this settles the power-law case away from the borderline exponent.

Corollary 1.2. Let d≥2d\ge2, α>0\alpha>0, α≠d\alpha\ne d, and let JJ satisfy (Sym) and c∥x∥−d−α≤J(x)≤C∥x∥−d−αc\lVert x\rVert^{-d-\alpha}\le J(x)\le C\lVert x\rVert^{-d-\alpha} for all x≠0x\ne0, with 0<c≤C<∞0<c\le C<\infty. Then θ(βc)=0\theta(\beta_c)=0.

Corollary 1.3. Let d≥2d\ge2 and let JJ satisfy (Sym), (Pos) and J(x)≤Ae−a∥x∥J(x)\le Ae^{-a\lVert x\rVert} for some A,a>0A,a>0. Then θ(βc)=0\theta(\beta_c)=0. The same holds for every finite-range JJ satisfying (Sym) and (Pos).

Corollary 1.2 covers both regimes of α\alpha around the exceptional value dd. For α<d\alpha<d it is Berger’s theorem; for α>d\alpha>d it follows from Theorem 1.1, since then T(n)≤C′n−αT(n)\le C'n^{-\alpha}. The value α=d\alpha=d is excluded: there ndT(n)n^dT(n) is bounded away from zero, so edges of length comparable to nn leave the cube [−n,n]d[-n,n]^d with probability of order one at every scale. This is the mechanism behind the discontinuous transition of the one-dimensional model with J(x)≍∣x∣−2J(x)\asymp|x|^{-2} [1, 5], and Theorem 1.1 says nothing about it.

Earlier results

Berger [3, Theorem 2.6] proved that θ(βc)=0\theta(\beta_c)=0 whenever lim inf⁡∥k∥→∞P(0↔k directly)∥k∥s>0\liminf_{\lVert k\rVert\to\infty}\mathbb{P}(0\leftrightarrow k\text{ directly})\lVert k\rVert^s>0 for some s<2ds<2d, in every d≥1d\ge1; Hutchcroft [13] made this quantitative. In the opposite regime the kernel decays fast enough that the model should behave like nearest-neighbour percolation, and Hutchcroft writes that “it is widely believed that the phase transition should be continuous for all α>0\alpha>0 and d≥2d\ge2” [13 p. 2]. The known cases of this regime are the following.

  • In d=2d=2, Duminil-Copin, Sidoravicius and Tassion [6] proved continuity for finite-range kernels invariant under the symmetries of Z2\mathbb{Z}^2, and Angel, de la Riva, Panagiotis and Shi [2] extended this to kernels with exponential decay; they state the case α>2\alpha>2 of power laws in d=2d=2 as a hope [2 §1.4]. Both arguments are planar in spirit: they glue clusters that cross as subsets of R2\mathbb{R}^2.

  • In high dimensions, the lace expansion gives mean-field behaviour, and in particular θ(βc)=0\theta(\beta_c)=0, for sufficiently spread-out kernels when d>3(α∧2)d>3(\alpha\wedge2) [12].

  • For kernels that are uniform on a finite range, Theorem 1.1’s conclusion follows from the recent preprint on quasi-transitive graphs [18].

Theorem 1.1 treats every d≥2d\ge2 and requires neither planarity nor high dimension nor a spread-out kernel. It contains the theorem of [2], whose kernels satisfy (Sym) (in d=2d=2, invariance under π/2\pi/2-rotations and reflections), (Pos) and exponential decay.

Idea of the proof

We follow the self-contained proof of θ(pc)=0\theta(p_{\mathrm{c}})=0 for nearest-neighbour percolation on Z3\mathbb{Z}^{3} in [17], which builds on the renormalization scheme of Kozma and Nitzan [15 §4] and the seed-and-exploration method of Grimmett and Marstrand [9]. Figure 1 shows the structure. Assume, for a contradiction, that θ(β)>0\theta(\beta)>0 at β=βc\beta=\beta_{\mathrm{c}}. The argument has four steps.

Diagram showing how Theorem 1.1 and its corollaries are proved

Figure 1. How Theorem 1.1 and its corollaries are proved. A solid arrow means that its source is used in the proof of its target; only the main dependencies are drawn. The dashed arrow means that the theorem of [2] is contained in the case d=2d=2 of Corollary 1.3. Box styles give provenance; grey labels give the source of each adapted step and where it is proved here. The modifications are listed in § 1.3.

  1. Faces at β\beta. A large box Λm(u)\Lambda_{m}(u) connects with high probability, inside a surrounding cube Λℓ(u)\Lambda_{\ell}(u), to each of the d2dd2^{d} quarter-faces of that cube (a face cut in half in every transverse coordinate). Only three radii ℓ\ell are needed, so these finitely many estimates persist at some q<βq<\beta.

  1. Extension. If a vertex oo outside a box BB reaches BB, then it reaches, with almost the same probability, any target set TT that every point near BB can hit through a quarter-face. The cluster of oo enters a neighbourhood of BB at many possible points; at each one a small seed box, if all its edges are open, gives a relay point that reaches TT with high probability. The gluing inequality (Theorem 2.1) turns “every relay point reaches TT with high probability” into “oo reaches TT if it reaches a relay point”.

  1. Relay. A fixed sequence of extensions moves a connection from the centre of one coarse cube

to the centre of a neighbouring one.

  1. Exploration. Cubes indexed by Z2\mathbb{Z}^2 are explored one at a time, keeping a high conditional connection probability for each pending cube; a planar contour count then yields an infinite cluster at q<βcq < \beta_c, a contradiction.

Long edges break two nearest-neighbour features of this argument.

First, a path that leaves a box need not pass through its boundary: its first step outside can be a long jump. In Step 1 we therefore use outer thick quarter-faces, slabs of width ww just outside a face of the cube (Figure 2), and we require the last edge of the connection to have length at most ww. Jumps longer than ww out of Λℓ\Lambda_\ell have total probability at most β(2ℓ+1)dT(w)\beta(2\ell+ 1)^d T(w). The tail condition makes this small for a ww that is a small fixed fraction of ℓ\ell, so the thick faces still fit inside the relay geometry of Step 3.

Lemma 3.2 in a planar cross-section

Figure 2. Lemma 3.2 in a planar cross-section. An open path from Λm(u)\Lambda_m(u) that leaves Λℓ(u)\Lambda_\ell(u) by an edge of length at most ww lands in the shell Λℓ+w(u)∖Λℓ(u)\Lambda_{\ell+w}(u) \setminus\Lambda_\ell(u) (light grey), which the NFN_F outer thick quarter-faces cover; one is shown in dark grey. A path can also leave by a long edge (dashed), which a nearest-neighbour argument would not see; such edges are rare when ℓdT(w)\ell^d T(w) is small.

Second, in Step 2 the nearest-neighbour proof selects an expansion B′=B(j)B' = B^{(j)} of BB such that the cluster of oo outside B′B' has many exterior neighbours, using the fact that closing the single edge from each exterior neighbour into B′B' disconnects oo from B′B'. With long edges, every vertex of the outside cluster has edges into B′B'. We replace the count of exterior neighbours by the crossing weight

W=∑x∈C∑z∈B′βJ(x−z),W = \sum_{x\in C} \sum_{z\in B'} \beta J(x-z),

where CC is the cluster of oo in the complement of B′B'; the probability that all crossing edges are closed is exactly e−We^{-W} (Figure 3). If WW is small this disconnection is cheap, and disjointness of the corresponding events over well-separated layers shows that a small weight is rare at some layer. If WW is large, the landing points of the crossing edges support many independent attempts to open a seed box, because a single seed box Λm(v)\Lambda_m(v) can absorb at most β(2m+1)d∥J∥1\beta(2m+1)^d\lVert J\rVert_1 of the weight. Entrances that jump over a shell of width ww around B′B' are again controlled by the tail condition. The other steps of [17] need only bookkeeping: dd contraction steps instead of three in the relay between neighbouring cubes, margins enlarged by ww, and an exploration whose root cube is opened without fixing the edges that leave it.

The extension lemma in a planar cross-section

Figure 3. The extension lemma (Lemma 4.1) in a planar cross-section. The cluster C\mathcal{C} of oo outside C=B(j)\mathcal{C}=B^{(j)} reaches the shell SjS_j. All crossing edges from C\mathcal{C} into C\mathcal{C} are closed with probability exactly e−Wje^{-W_j}, which is cheap when WjW_j is small; when WjW_j is large, the landing points zz in one residue class carry disjoint seed boxes U(z)U(z), which give many independent chances to reach a good relay centre. Entrances that jump over the shell are rare by the tail condition.

The scales are chosen in the order m≤w≪R≪rm \le w \ll R \ll r, where RR and rr are large constant multiples of ww and the tail condition makes βrdT(w)\beta r^d T(w) small (Lemma 3.3). Only one such choice is needed; this is why a lim inf suffices.

Adapted results. Lemma 3.2 and Lemma 3.3 adapt [17 Lemma 3.3] (outer thick faces and short exits); Lemma 4.1 adapts [17 Lemma 4.1], itself adapted from [15 Lemma 10] (crossing weight, layers spaced by ww, seeds at landing points); Lemma 5.1 adapts [17 Lemma 5.1] to Zd\mathbb{Z}^d and to the margins R+wR+w; Section 6 adapts [17 §6, Lemmas 6.1–6.3] (long-range edges between neighbouring cubes, a hypothesis on edges entering the union of two cubes, and the root). The planar contour count is used as in [17 Lemma 3.1], and we include its proof.

Organization

Section 2 collects the inputs. Sections 3–6 carry out Steps 1–4, and Section 7 proves the corollaries and records where each hypothesis is used.

Preliminaries

Notation. Λb(u)=u+([−b,b]d∩Zd)\Lambda_b(u) = u + ([-b,b]^d \cap\mathbb{Z}^d) and Λb=Λb(0)\Lambda_b = \Lambda_b(0). A rectangle is a nonempty product of integer intervals, and B(j)B^{(j)} denotes the expansion of a rectangle BB by jj in every coordinate direction. Throughout, ∥J∥1=∑xJ(x)\lVert J\rVert_1 = \sum_x J(x); it is finite under (Tail), since T(n0)<∞T(n_0) < \infty for some n0n_0 and the remaining sum is finite.

Product laws on finite graphs. In Sections 4–6 we work on finite graphs G=(V,E)G = (V,E) with V⊂ZdV \subset\mathbb{Z}^d, whose edges are pairs {x,y}\{x,y\} with J(x−y)>0J(x-y) > 0, with independent bits of individual probabilities pe∈[0,1]p_e \in[0,1]. We put λe=−log⁡(1−pe)∈[0,∞]\lambda_e = -\log(1-p_e) \in[0,\infty], so that P(e open)≤λe\mathbb{P}(e\text{ open}) \leq\lambda_e and the probability that all edges of a set SS are closed is exp⁡(−∑e∈Sλe)\exp(-\sum_{e\in S}\lambda_e). A bit is internal to a vertex set UU if both endpoints lie in UU. For q>0q > 0 we say that (Fq(U))(F_q(U)) holds if every pair {x,y}⊂U\{x,y\} \subset U with $J(x-y) > 0 is an edge of GG with λxy=qJ(x−y)\lambda_{xy} = qJ(x-y).

The gluing inequality

Theorem 2.1 [17, 4, 8]. Consider independent bond percolation on a finite graph with arbitrary edge probabilities in [0,1][0,1]. For vertices s,ts,t and a nonempty vertex set AA,

P(s↔A, s↔t)≥P(s↔A)min⁡a∈AP(a↔t).\mathbb{P}(s \leftrightarrow A,\ s \leftrightarrow t) \geq\mathbb{P}(s \leftrightarrow A)\min_{a\in A}\mathbb{P}(a \leftrightarrow t).

This is a joint form of Conjecture 1 of Kozma and Nitzan [15], in which the event {s↔t}\{s \leftrightarrow t\} alone appears on the left. Theorem 2.1 is proved for hypergraphs, with tt replaced by a vertex set, in [17 Proposition 2.1], and in the form above in [8 Theorem 1], where the extension from (0,1)(0,1) to [0,1][0,1] is by continuity; it also follows from [4 Theorem 1] together with Harris’s inequality. The Lean formalization of Leder [16] proves the additive form: if P(a↔t)≥1−b\mathbb{P}(a \leftrightarrow t) \geq1-b for every a∈Aa \in A, then P(s↔t)≥P(s↔A)−b\mathbb{P}(s \leftrightarrow t) \geq\mathbb{P}(s \leftrightarrow A)-b [16 Theorem 2.2]. Either form gives the following corollary, which is [17 Corollary 2.3]; we include the short proof from Theorem 2.1.

Corollary 2.2 (failure comparison). Consider independent bond percolation on a finite graph GG with arbitrary edge probabilities in [0,1][0,1], let A,TA,T be vertex sets and b∈[0,1]b \in[0,1]. If P(a↔T)≥1−b\mathbb{P}(a \leftrightarrow T) \geq1-b for every a∈Aa \in A, then P(o↔A,o↮T)≤b\mathbb{P}(o \leftrightarrow A, o \not\leftrightarrow T) \leq b for every vertex oo. The same holds after any deterministic set of bits has been fixed.

Proof. If o∈To \in T, A=∅A = \varnothing or T=∅T = \varnothing (which forces b=1b = 1 when A≠∅A \ne\varnothing) there is nothing to prove. Otherwise let G′G' be obtained from GG by identifying the vertices of TT to a single vertex tt, removing loops, and merging the parallel edges e1,…,eke_1,\ldots,e_k between two vertices into one edge, open if and only if one of the eie_i is open; this couples the two models, and the merged edge has probability 1−∏i(1−pei)1-\prod_i(1-p_{e_i}). Under the coupling, o↮To \not\leftrightarrow T in GG if and only if o↮to \not\leftrightarrow t in G′G', and then the open clusters of oo in GG and G′G' coincide. Let A′A' be the image of AA in G′G'. Hence PG(o↔A,o↮T)=PG′(o↔A′,o↮t)\mathbb{P}_G(o \leftrightarrow A, o \not\leftrightarrow T) = \mathbb{P}_{G'}(o \leftrightarrow A', o \not\leftrightarrow t). In G′G' every a∈A′a \in A' satisfies PG′(a↔t)≥1−b\mathbb{P}_{G'}(a \leftrightarrow t) \geq1-b: for a∉Ta \notin T this is PG(a↔T)≥1−b\mathbb{P}_G(a \leftrightarrow T) \geq1-b, and t↔tt \leftrightarrow t always. Theorem 2.1 in G′G' gives PG′(o↔A′,o↮t)=PG′(o↔A′)−PG′(o↔A′,o↔t)≤bPG′(o↔A′)≤b\mathbb{P}_{G'}(o \leftrightarrow A', o \not\leftrightarrow t) = \mathbb{P}_{G'}(o \leftrightarrow A')-\mathbb{P}_{G'}(o \leftrightarrow A', o \leftrightarrow t) \leq b\mathbb{P}_{G'}(o \leftrightarrow A') \leq b. Fixing a bit means giving it probability 0 or 1. □\square

We also use Harris’s inequality [11]: increasing events of a finite product of independent bits are positively correlated, and so are decreasing events.

A planar boundary count. The following lemma is [17 Lemma 3.1]; we include the short proof.

Lemma 2.3. There are families CnC_n of simple cycles of length nn in the dual lattice Z2+(12,12)\mathbb{Z}^2 + \left(\frac{1}{2},\frac{1}{2}\right), with ∣Cn∣≤2n3n|C_n| \leq2n3^n, such that the dual edge boundary of every finite X⊂Z2X \subset\mathbb{Z}^2 containing 0 contains a cycle in ⋃nCn\bigcup_n C_n. For each cycle in CnC_n one can fix at least n/7n/7 of its crossed primal edges with pairwise disjoint endpoints.

Proof. Each dual vertex has even degree in the dual boundary of XX (it counts membership changes around a primal square), so the boundary decomposes into edge-disjoint simple cycles. Along the positive horizontal ray from 00 the number of boundary crossings is odd, so some cycle of the decomposition crosses it an odd number of times; a closed cycle crosses the whole horizontal line an even number of times, so this cycle also crosses the negative ray. If its length is nn, some positive crossing has horizontal coordinate t+12t+\frac{1}{2} with 0≤t<n0 \le t < n. Let CnC_n be the set of simple cycles of length nn with such an anchored crossing and an odd number of positive crossings: there are at most nn anchors, two orientations and 3n3^n non-backtracking continuations. Finally select crossed primal edges greedily, discarding those incident to an endpoint of a selected edge; each selection discards at most seven edges, since primal degrees are four.

Elementary facts.

Lemma 2.4. Under (Sym), (Pos) and ∥J∥1<∞\lVert J\rVert_1 < \infty, if d≥2d \ge2 then 0<βc<∞0 < \beta_c < \infty. If θ(β)>0\theta(\beta) > 0, then an infinite cluster exists almost surely, and

Pβ(Λm meets an infinite cluster)→1(m→∞).(1)\mathbb{P}_{\beta}(\Lambda_m\text{ meets an infinite cluster}) \to1 \qquad(m \to\infty). \tag*{(1)}

Proof. The nearest-neighbour pairs in a coordinate plane are open independently with probability 1−e−βJ(e1)1-e^{-\beta J(e_1)}, by (Sym) and (Pos), and planar bond percolation with parameter close to 11 percolates by Lemma 2.3 and a union bound over dual cycles; so βc<∞\beta_c < \infty. The cluster of 00 is dominated by a Galton–Watson tree with mean offspring ∑y(1−e−βJ(y))≤β∥J∥1\sum_y(1-e^{-\beta J(y)}) \le\beta\lVert J\rVert_1, so βc≥1/∥J∥1\beta_c \ge1/\lVert J\rVert_1. Removing an edge splits a connected component into at most two components, so the event that an infinite cluster exists is unchanged by modifying finitely many bits and is a tail event; by Kolmogorov’s 0–1 law it has probability 1 when θ(β)>0\theta(\beta) > 0. The events that Λm\Lambda_m meets an infinite cluster increase to this event.

Faces at a percolating parameter, and the choice of scales

Definition 3.1. Let u∈Zdu \in\mathbb{Z}^d and ℓ,w≥1\ell,w \ge1. For i∈{1,…,d}i \in\{1,\ldots,d\}, σ∈{±1}\sigma\in\{\pm1\} and ε=(εj)j≠i∈{±1}d−1\varepsilon=(\varepsilon_j)_{j\ne i}\in\{\pm1\}^{d-1}, the outer thick quarter-face Fi,σ,ε(u,ℓ)F_{i,\sigma,\varepsilon}(u,\ell) is

{u+x: σxi∈(ℓ,ℓ+w], 0≤εjxj≤ℓ+w (j≠i)}.\{u+x:\ \sigma x_i\in(\ell,\ell+w],\ 0\le\varepsilon_jx_j\le\ell+w\ (j\ne i)\}.

There are NF=d2dN_F=d2^d of them. For m<ℓm<\ell and such a face FF, let X(u,ℓ,F)X(u,\ell,F) be the event that there is an open path x0,…,xkx_0,\ldots,x_k with x0∈Λm(u)x_0\in\Lambda_m(u), x0,…,xk−1∈Λℓ(u)x_0,\ldots,x_{k-1}\in\Lambda_\ell(u), xk∈Fx_k\in F and ∥xk−xk−1∥≤w\lVert x_k-x_{k-1}\rVert\le w. The event X(u,ℓ,F)X(u,\ell,F) is increasing. It depends only on the bits internal to Λℓ(u)\Lambda_\ell(u) and on the pairs between Λℓ(u)\Lambda_\ell(u) and FF, all of which are internal to Λℓ+w(u)\Lambda_{\ell+w}(u).

Lemma 3.2 (faces at a percolating parameter). Assume (Sym) and θ(β)>0\theta(\beta)>0. For ℓ>m\ell>m, w≥1w\ge1, every uu and every outer thick quarter-face FF of (u,ℓ)(u,\ell),

Pβ(X(u,ℓ,F)c)≤[Pβ(Λm meets no infinite cluster)+β(2ℓ+1)dT(w)]1/NF.\mathbb{P}_{\beta}(X(u,\ell,F)^c)\le\left[\mathbb{P}_{\beta}(\Lambda_m\text{ meets no infinite cluster})+\beta(2\ell+1)^dT(w)\right]^{1/N_F}.

Proof. Take u=0u=0. Let ElongE_{\mathrm{long}} be the event that some open edge {x,y}\{x,y\} has x∈Λℓx\in\Lambda_\ell and ∥x−y∥>w\lVert x-y\rVert>w; then P(Elong)≤∑x∈Λℓ∑∥z∥>w(1−e−βJ(z))≤β(2ℓ+1)dT(w)\mathbb{P}(E_{\mathrm{long}})\le\sum_{x\in\Lambda_\ell}\sum_{\lVert z\rVert>w}(1-e^{-\beta J(z)})\le\beta(2\ell+1)^dT(w). If Λm\Lambda_m meets an infinite cluster, take a finite open path from Λm\Lambda_m to a vertex outside Λℓ\Lambda_\ell, and let yy be its first vertex outside Λℓ\Lambda_\ell and xx its predecessor. Off ElongE_{\mathrm{long}} we have ∥x−y∥≤w\lVert x-y\rVert\le w, so y∈Λℓ+w∖Λℓy\in\Lambda_{\ell+w}\setminus\Lambda_\ell. Choose ii with ∣yi∣>ℓ|y_i|>\ell, σ=sign⁡yi\sigma=\operatorname{sign} y_i and

εj=sign⁡yj\varepsilon_j = \operatorname{sign} y_j (either sign if yj=0y_j = 0); then y∈Fi,σ,ε(0,ℓ)y \in F_{i,\sigma,\varepsilon}(0,\ell) and the path up to yy witnesses X(0,ℓ,Fi,σ,ε)X(0,\ell,F_{i,\sigma,\varepsilon}). Hence P(⋂FX(0,ℓ,F))≤P(Λm meets no infinite cluster)+P(Elong)\mathbb{P}\left(\bigcap_{F} X(0,\ell,F)\right) \leq\mathbb{P}(\Lambda_m\text{ meets no infinite cluster}) + \mathbb{P}(E_{\mathrm{long}}). The events X(0,ℓ,F)cX(0,\ell,F)^c are decreasing and depend on finitely many bits. The signed coordinate permutations fixing 00 act transitively on the faces and, by (Sym), preserve the law, Λm\Lambda_m and Λℓ\Lambda_\ell; so all NFN_F complements have the same probability aa. Harris’s inequality gives aNF≤P(⋂FX(0,ℓ,Fc))a^{N_F} \leq\mathbb{P}\left(\bigcap_F X(0,\ell,F^c)\right). □

Lemma 3.3 (choice of scales). Assume (Sym), (Tail) and θ(β)>0\theta(\beta) > 0. There are f,δ,η,τ,γ∈(0,1)f,\delta,\eta,\tau,\gamma\in(0,1), positive integers m,K,w,R,rm,K,w,R,r, constants ρ0,A,cA,L>0\rho_0,A,c_A,L > 0 and q∈(β/2,β)q \in(\beta/2,\beta) such that:

(S1) f<12f < \frac{1}{2} and ∑n≥12n3n(2f)n/7<1\sum_{n\geq1} 2n3^n(2f)^{n/7} < 1;

(S2) δ+4(d+10)η/δ<f\delta+ 4(d+10)\eta/\delta< f, τ<η/4\tau< \eta/4, γ<δ\gamma< \delta and 2γ/τ<η/42\gamma/\tau< \eta/4;

(S3) Pβ(Λm meets no infinite cluster)<12(γ/2)NF\mathbb{P}_\beta(\Lambda_m\text{ meets no infinite cluster}) < \frac{1}{2}(\gamma/2)^{N_F};

(S4) ρ0=∏e(1−e−(β/2)J(e))\rho_0 = \prod_e(1-e^{-(\beta/2)J(e)}), the product over the pairs ee internal to Λm\Lambda_m with J(e)>0J(e)>0; A=β∥J∥1A=\beta\lVert J\rVert_1; cA=(1−e−A)/Ac_A=(1-e^{-A})/A; 4/(Kρ0)<η/44/(K\rho_0)<\eta/4; and L=K(4m+3)d/cAL=K(4m+3)^d/c_A;

(S5) β(41r+1)dT(w)≤min⁡{12(γ/2)NF,η/8}\beta(41r+1)^dT(w) \leq\min\{\frac{1}{2}(\gamma/2)^{N_F},\eta/8\};

(S6) w≥mw \geq m, R≥m+w+8eLw/ηR \geq m+w+8e^{Lw/\eta}, r>4(d+12)(R+w+1)r>4(d+12)(R+w+1) and r>m+wr>m+w;

(S7) for ℓ∈L:={2r,2r+1,10r−w}\ell\in\mathcal{L}:=\{2r,2r+1,10r-w\}, every uu and every outer thick quarter-face FF of (u,ℓ)(u,\ell), Pq(X(u,ℓ,F))>1−γ\mathbb{P}_q(X(u,\ell,F))>1-\gamma.

Proof. Choose ff, then δ,η,τ,γ\delta,\eta,\tau,\gamma in this order; then mm by (1); then ρ0,A,cA,K,L\rho_0,A,c_A,K,L. None of these depends on ww or rr. By (Tail) there are arbitrarily large w≥mw\geq m with wdT(w)w^dT(w) as small as we like. For such ww put R=m+w+⌈8eLw/η⌉R=m+w+\lceil8e^{Lw/\eta}\rceil and r=4(d+12)(R+w+2)r=4(d+12)(R+w+2); then (S6) holds and r≤C0wr\leq C_0w with C0C_0 depending only on d,m,L,ηd,m,L,\eta, so β(41r+1)dT(w)≤β(42C0)dwdT(w)\beta(41r+1)^dT(w)\leq\beta(42C_0)^d w^dT(w) and (S5) holds for a suitable such ww. Every ℓ∈L\ell\in\mathcal{L} satisfies ℓ>m\ell>m and 2ℓ+1≤41r+12\ell+1\leq41r+1, so Lemma 3.2, (S3) and (S5)

::: give Pβ(X(u,ℓ,F)c)<γ/2\mathbb{P}_{\beta}(X(u,\ell,F)^{\mathrm c})<\gamma/2. These finitely many events (up to translation) depend on finitely many bits whose probabilities are continuous in the parameter, so (S7) holds for some q∈(β/2,β)q\in(\beta/2,\beta). □

We fix these choices from now on, and write ρ=∏e(1−e−qJ(e))≥ρ0\rho=\prod_e(1-e^{-qJ(e)})\geq\rho_0 (over the same pairs) for the probability that all pairs ee internal to a translate of Λm\Lambda_m with J(e)>0J(e)>0 are open.

Extension through a fresh region

Lemma 4.1 (extension). Let G=(V,E)G=(V,E) be a finite graph as in § 2, D⊂VD\subset V a rectangle with ∣D∣≤(41r+1)d|D|\leq(41r+1)^d and (Fq(D))(F_q(D)), BB a rectangle with B(R)⊂DB^{(R)}\subset D, o∈V∖B(R)o\in V\setminus B^{(R)} and T⊂VT\subset V. Assume:

(E2)(E2) every edge {x,z}\{x,z\} of GG with z∈B(R)z\in B^{(R)} and x∉Dx\notin D has λxz≤qJ(x−z)\lambda_{xz}\leq qJ(x-z);

(E3)(E3) for every u∈B(R)u\in B^{(R)} there are ℓ∈L\ell\in\mathcal{L} and an outer thick quarter-face FF of (u,ℓ)(u,\ell) with Λℓ+w(u)⊂D\Lambda_{\ell+w}(u)\subset D and F⊂TF\subset T.

Then P(o↔B, o↮T)<η\mathbb{P}(o\leftrightarrow B,\ o\not\leftrightarrow T)<\eta.

Proof. By (Fq(D))(F_q(D)) and (E2), every edge {x,z}\{x,z\} with z∈B(R)z\in B^{(R)} and x∉B(R)x\notin B^{(R)} has λxz≤qJ(x−z)\lambda_{xz}\leq qJ(x-z); hence ∑xλxz≤q∥J∥1≤A\sum_x\lambda_{xz}\leq q\lVert J\rVert_1\leq A for z∈B(R)z\in B^{(R)}.

Step 1: a layer. For an integer j∈[m,R)j\in[m,R) let ζj\zeta_j be the bits internal to V∖B(j)V\setminus B^{(j)}, let Cj\mathcal{C}_j be the open cluster of oo in the graph on V∖B(j)V\setminus B^{(j)} with these bits, and set

Wj=∑x∈Cj∑z∈B(j)λxz,Sj=B(j+w)∖B(j),W_j=\sum_{x\in\mathcal{C}_j}\sum_{z\in B^{(j)}}\lambda_{xz},\qquad S_j=B^{(j+w)}\setminus B^{(j)},
Lj={Wj<L, Cj∩Sj≠∅},Clj={every edge between Cj and B(j) is closed}.L_j=\{W_j<L,\ \mathcal{C}_j\cap S_j\neq\varnothing\},\qquad\mathrm{Cl}_j=\{\text{every edge between }\mathcal{C}_j\text{ and }B^{(j)}\text{ is closed}\}.

Here Wj≤∣B(j)∣A<∞W_j\leq|B^{(j)}|A<\infty. The crossing bits between V∖B(j)V\setminus B^{(j)} and B(j)B^{(j)} are independent of ζj\zeta_j, so P(Clj∣ζj)=e−Wj≥e−L\mathbb{P}(\mathrm{Cl}_j\mid\zeta_j)=e^{-W_j}\geq e^{-L} on LjL_j. On Clj\mathrm{Cl}_j the open cluster C(o)\mathcal{C}(o) of oo in GG equals Cj\mathcal{C}_j, because every other edge leaving Cj\mathcal{C}_j is internal to V∖B(j)V\setminus B^{(j)} and therefore closed by maximality.

Let J={m+kw:k≥0}∩[m,R)\mathcal{J}=\{m+kw:k\geq0\}\cap[m,R). If j′<jj'<j both lie in J\mathcal{J}, then j′+w≤jj'+w\leq j; on Lj∩CljL_j\cap C_{l_j} we have C(o)=Cj⊂V∖B(j)C(o)=C_j\subset V\setminus B^{(j)}, hence Cj′=C(o)C_{j'}=C(o), while Sj′⊂B(j)S_{j'}\subset B^{(j)}, so Cj′∩Sj′=∅C_{j'}\cap S_{j'}=\varnothing and Lj′L_{j'} fails. The events Lj∩CljL_j\cap C_{l_j}, j∈Jj\in\mathcal{J}, are therefore pairwise disjoint, and

e−L∑j∈JP(Lj)≤∑j∈JP(Lj∩Clj)≤1.e^{-L}\sum_{j\in\mathcal{J}}\mathbb{P}(L_j)\leq\sum_{j\in\mathcal{J}}\mathbb{P}(L_j\cap C_{l_j})\leq1.

Since ∣J∣≥(R−m)/w|\mathcal{J}|\geq(R-m)/w, we can fix j∈Jj\in\mathcal{J} with P(Lj)≤eLw/(R−m)≤η/8\mathbb{P}(L_j)\leq e^{Lw}/(R-m)\leq\eta/8, by (S6).

Let ElongE_{\mathrm{long}} be the event that some open edge {x,z}\{x,z\} has z∈B(R)z\in B^{(R)} and ∥x−z∥>w\|x-z\|>w. Then

P(Elong)≤∑z∈B(R)∑∥x−z∥>wλxz≤q∣D∣T(w)≤η/8\mathbb{P}(E_{\mathrm{long}})\leq\sum_{z\in B^{(R)}}\sum_{\|x-z\|>w}\lambda_{xz}\leq q|D|T(w)\leq\eta/8

by (S5). If o↔Bo\leftrightarrow B, Wj<LW_j<L and Cj∩Sj=∅C_j\cap S_j=\varnothing, then an open path from oo enters B(j)B^{(j)} for the first time at some zz from a predecessor x∈Cj∖B(j+w)x\in C_j\setminus B^{(j+w)}, so ∥x−z∥>w\|x-z\|>w and ElongE_{\mathrm{long}} occurs. Hence

P(o↔B, Wj<L)≤P(Lj)+P(Elong)≤η/4.(2)\mathbb{P}(o\leftrightarrow B,\ W_j<L)\leq\mathbb{P}(L_j)+\mathbb{P}(E_{\mathrm{long}})\leq\eta/4. \tag*{(2)}

Step 2: relay quality. Put C=B(j)C=B^{(j)} and let ξ\xi be the bits internal to CC. Every side of CC has at least 2j+1≥2m+12j+1\geq2m+1 points. For z∈Cz\in C let v(z)v(z) be obtained from zz by clamping each coordinate to the corresponding side of CC shrunk by mm at both ends, let U(z)=Λm(v(z))⊂CU(z)=\Lambda_m(v(z))\subset C (so z∈U(z)z\in U(z)), let O(z)O(z) be the event that all bits internal to U(z)U(z) are open, so P(O(z))=ρ\mathbb{P}(O(z))=\rho, and put

g(z)=P(v(z)↔T∣ξ),Aξ={v(z):z∈C, g(z)>1−τ}.g(z)=\mathbb{P}(v(z)\leftrightarrow T\mid\xi),\qquad A_\xi=\{v(z):z\in C,\ g(z)>1-\tau\}.

AξA_\xi is a function of ξ\xi. Fix zz. By (E3) at u=v(z)∈B(R)u=v(z)\in B^{(R)} there are ℓ∈L\ell\in\mathcal{L} and FF with Λℓ+w(v(z))⊂D\Lambda_{\ell+w}(v(z))\subset D and F⊂TF\subset T; the event X(v(z),ℓ,F)X(v(z),\ell,F) uses only bits internal to DD, so by (Fq(D))(F_q(D)) and (S7) its probability exceeds 1−γ1-\gamma. By (Pos) and (Sym) an open box is connected through its nearest-neighbour pairs, so on O(z)∩X(v(z),ℓ,F)O(z)\cap X(v(z),\ell,F) we have v(z)↔Tv(z)\leftrightarrow T. Harris’s inequality gives P(O(z),v(z)↮T)≤ργ\mathbb{P}(O(z),v(z)\not\leftrightarrow T)\leq\rho\gamma, and since O(z)O(z) is ξ\xi-measurable,

τP(O(z),g(z)≤1−τ)≤E[1O(z)(1−g(z))]=P(O(z),v(z)↮T)≤ργ.(3)\tau\mathbb{P}(O(z),g(z)\leq1-\tau)\leq\mathbb{E}\left[\mathbf{1}_{O(z)}(1-g(z))\right]=\mathbb{P}(O(z),v(z)\not\leftrightarrow T)\leq\rho\gamma. \tag*{(3)}

Step 3: entrances. Condition on ζ=ζj\zeta=\zeta_j; write C=CjC=C_j and az=∑x∈Cλxza_z=\sum_{x\in C}\lambda_{xz} for z∈Cz\in C, so that Wj=∑z∈CazW_j=\sum_{z\in C}a_z and az≤Aa_z\leq A. Among the (4m+3)d(4m+3)^d residue classes of CC modulo 4m+34m+3, choose (as a function of ζ\zeta) a class ZZ with ∑z∈Zaz≥Wj/(4m+3)d\sum_{z\in Z}a_z\geq W_j/(4m+3)^d. Distinct z,z′∈Zz,z'\in Z satisfy ∥z−z′∥≥4m+3\|z-z'\|\geq4m+3, and clamping moves each coordinate by at most mm, so ∥v(z)−v(z′)∥≥2m+3\|v(z)-v(z')\|\geq2m+3 and the boxes U(z)U(z), z∈Zz\in Z, are disjoint.

Let Σz\Sigma_z be the event that some edge between CC and zz is open and O(z)O(z) holds. Given ζ\zeta, the crossing bits are independent of ξ\xi, edges into distinct zz are distinct, and the events O(z)O(z), z∈Zz\in Z, use disjoint bits of ξ\xi; hence the events Σz\Sigma_z, z∈Zz\in Z, are independent with P(Σz∣ζ)=ρ(1−e−az)\mathbb{P}(\Sigma_z\mid\zeta)=\rho(1-e^{-a_z}). Let S=∑z∈Z1ΣzS=\sum_{z\in Z}\mathbf{1}_{\Sigma_z} and μ=∑z∈Z(1−e−az)\mu=\sum_{z\in Z}(1-e^{-a_z}). Since a↦(1−e−a)/aa\mapsto(1-e^{-a})/a is decreasing, 1−e−az≥cAaz1-e^{-a_z}\geq c_Aa_z and μ≥cAWj/(4m+3)d\mu\geq c_AW_j/(4m+3)^d, which is at least KK on {Wj≥L}\{W_j\geq L\}. By Chebyshev’s inequality, on {Wj≥L}\{W_j\geq L\},

P(S<ρμ/2∣ζ)≤4ρμ≤4ρ0K<η4.\mathbb{P}(S<\rho\mu/2\mid\zeta)\leq\frac{4}{\rho\mu}\leq\frac{4}{\rho_0K}<\frac{\eta}{4}.

Let Zbad=∑z∈Z1Σz1{g(z)≤1−τ}Z_{\mathrm{bad}}=\sum_{z\in Z}\mathbf{1}_{\Sigma_z}\mathbf{1}_{\{g(z)\leq1-\tau\}}. By the independence of the crossing bits from ξ\xi and (3),

E[Zbad∣ζ]=∑z(1−e−az)P(O(z),g(z)≤1−τ)≤μργ/τ,\mathbb{E}[Z_{\mathrm{bad}}\mid\zeta]=\sum_{z}(1-e^{-a_z})\mathbb{P}(O(z),g(z)\leq1-\tau)\leq\mu\rho\gamma/\tau,

so P(Zbad≥ρμ/2∣ζ)≤2γ/τ<η/4\mathbb{P}(Z_{\mathrm{bad}}\geq\rho\mu/2\mid\zeta)\leq2\gamma/\tau<\eta/4. If S≥ρμ/2>ZbadS\geq\rho\mu/2>Z_{\mathrm{bad}}, some z∈Zz\in Z has Σz\Sigma_z and g(z)>1−τg(z)>1-\tau: then oo reaches some x∈Cx\in C, the open edge {x,z}\{x,z\}, and v(z)∈Aξv(z)\in A_\xi through the open box U(z)U(z). Together with (2),

P(o↔B, o↮Aξ)≤34η.\mathbb{P}(o\leftrightarrow B,\ o\not\leftrightarrow A_\xi)\leq\frac{3}{4}\eta.

Step 4: the target. Condition on ξ\xi. Then AξA_\xi is fixed, the remaining bits keep their product law, and P(a↔T∣ξ)>1−τ\mathbb{P}(a\leftrightarrow T\mid\xi)>1-\tau for every a∈Aξa\in A_\xi. Corollary 2.2, applied with the bits of ξ\xi fixed, gives P(o↔Aξ, o↮T∣ξ)≤τ<η/4\mathbb{P}(o\leftrightarrow A_\xi,\ o\not\leftrightarrow T\mid\xi)\leq\tau<\eta/4. Splitting according to whether o↔Aξo\leftrightarrow A_\xi gives P(o↔B, o↮T)<η\mathbb{P}(o\leftrightarrow B,\ o\not\leftrightarrow T)<\eta. □\square

Relay between neighbouring coarse cubes

For v=(v1,v2)∈Z2v=(v_1,v_2)\in\mathbb{Z}^2 let

cv=(20r+1)(v1,v2,0,…,0)∈Zd,Qv=Λ10r(cv),Bv=Λ4r(cv)c_v=(20r+1)(v_1,v_2,0,\ldots,0)\in\mathbb{Z}^d,\qquad Q_v=\Lambda_{10r}(c_v),\qquad B_v=\Lambda_{4r}(c_v).

The cubes QvQ_v are vertex-disjoint. For neighbours v,xv,x in Z2\mathbb{Z}^2, let IvxI_{vx} be the finite set of pairs {y,y′}\{y,y'\} with y∈Qvy\in Q_v, y′∈Qxy'\in Q_x and J(y−y′)>0J(y-y')>0. The union Dvx=Qv∪QxD_{vx}=Q_v\cup Q_x is a rectangle with ∣Dvx∣=(40r+2)(20r+1)d−1≤(41r+1)d|D_{vx}|=(40r+2)(20r+1)^{d-1}\leq(41r+1)^d, whose internal bits are those internal to QvQ_v, those internal to QxQ_x, and IvxI_{vx}. Let Dvx∘D^\circ_{vx} be the set of points of DvxD_{vx} at distance more than 2r2r from Zd∖Dvx\mathbb{Z}^d\setminus D_{vx}.

Lemma 5.1 (relay). Let v,xv,x be neighbours in Z2\mathbb{Z}^2 and GG a finite graph as in § 2 with D=Dvx⊂VD=D_{vx}\subset V and (Fq(D))(F_q(D)). Assume

(E2∘)(E2^\circ) every edge {y,z}\{y,z\} of GG with z∈Dvx∘z\in D^\circ_{vx} and y∉Dy\notin D has λyz≤qJ(y−z)\lambda_{yz}\leq qJ(y-z).

If o∈V∖Do\in V\setminus D, then P(o↔Bv, o↮Bx)≤(d+10)η\mathbb{P}(o\leftrightarrow B_v,\ o\not\leftrightarrow B_x)\leq(d+10)\eta.

Proof. By (Sym) we may assume cv=0c_v=0 and cx=(20r+1)e1c_x=(20r+1)e_1, so that D=[−10r,30r+1]×[−10r,10r]d−1D=[-10r,30r+1]\times[-10r,10r]^{d-1}. Put R′=R+wR'=R+w. We define rectangles P0,…,Pd+10P_0,\ldots,P_{d+10} and apply Lemma 4.1 with B=Pk−1B=P_{k-1} and T=PkT=P_k; the radius is 2r2r for k≤d+9k\leq d+9 and 2r+12r+1 for k=d+10k=d+10. Let P0=BvP_0=B_v. For t=1,…,dt=1,\ldots,d let PtP_t be centred at 00 with half-width 2r+tR′2r+tR' in directions h≤th\leq t and 4r+tR′4r+tR' in directions h>th>t. For i=1,…,10i=1,\ldots,10 put si=2rs_i=2r (i≤9i\leq9), s10=2r+1s_{10}=2r+1, zi=(s1+⋯+si)e1z_i=(s_1+\cdots+s_i)e_1, wi=2r+(d+i)R′w_i=2r+(d+i)R', and Pd+i=Λwi(zi)P_{d+i}=\Lambda_{w_i}(z_i).

Contraction steps. Let 1≤t≤d1\leq t\leq d and u∈Pt−1(R)u\in P_{t-1}^{(R)}. Then ∣uh∣≤2r+tR′−w|u_h|\leq2r+tR'-w for h<th<t and ∣uh∣≤4r+tR′−w|u_h|\leq4r+tR'-w for h≥th\geq t. Take F=Ft,σ,ε(u,2r)F=F_{t,\sigma,\varepsilon}(u,2r) with σ=−sign⁡ut\sigma=-\operatorname{sign}u_t and εj=−sign⁡uj\varepsilon_j=-\operatorname{sign}u_j (with sign⁡0=+1\operatorname{sign}0=+1). A point y∈Fy\in F has yt=ut+σsy_t=u_t+\sigma s with s∈(2r,2r+w]s\in(2r,2r+w], so ∣yt∣≤∣ut∣−2r≤2r+tR′|y_t|\leq|u_t|-2r\leq2r+tR' if ∣ut∣≥2r+w|u_t|\geq2r+w and ∣yt∣≤2r+w|y_t|\leq2r+w otherwise. For j≠tj\ne t, yjy_j lies between uju_j and uj−sign⁡(uj)(2r+w)u_j-\operatorname{sign}(u_j)(2r+w), so ∣yj∣≤max⁡{∣uj∣,2r+w}|y_j|\leq\max\{|u_j|,2r+w\}, which is at most the half-width of PtP_t in direction jj. Thus F⊂PtF\subset P_t. The coordinates of Λ2r+w(u)\Lambda_{2r+w}(u) are at most 6r+dR′+w<8r6r+dR'+w<8r in absolute value, by (S6), so Λ2r+w(u)⊂Qv\Lambda_{2r+w}(u)\subset Q_v and Pt−1(R)⊂Dvx∘P_{t-1}^{(R)}\subset D^\circ_{vx}. After the last contraction, Pd=Λ2r+dR′(0)P_d=\Lambda_{2r+dR'}(0).

Translation steps. Let 1≤i≤101\leq i\leq10 and u∈Pd+i−1(R)u\in P_{d+i-1}^{(R)}, so ∥u−zi−1∥≤wi−1+R\lVert u-z_{i-1}\rVert\leq w_{i-1}+R with w0=2r+dR′w_0=2r+dR'. Take F=F1,+,ε(u,si)F=F_{1,+,\varepsilon}(u,s_i) with εj=−sign⁡uj\varepsilon_j=-\operatorname{sign}u_j. For y∈Fy\in F, y1−(zi)1=(u1−(zi−1)1)+(y1−u1−si)y_1-(z_i)_1=(u_1-(z_{i-1})_1)+(y_1-u_1-s_i) with y1−u1−si∈(0,w]y_1-u_1-s_i\in(0,w], so ∣y1−(zi)1∣≤wi−1+R+w=wi|y_1-(z_i)_1|\leq w_{i-1}+R+w=w_i, and ∣yj∣≤max⁡{∣uj∣,si+w}≤wi|y_j|\leq\max\{|u_j|,s_i+w\}\leq w_i for j≥2j\geq2. Thus F⊂Pd+iF\subset P_{d+i}. The first coordinates of Λsi+w(u)\Lambda_{s_i+w}(u) lie in [−4r−(d+11)R′−1,22r+(d+11)R′+1][-4r-(d+11)R'-1,22r+(d+11)R'+1] and the others have absolute value at most 4r+(d+11)R′+14r+(d+11)R'+1; by (S6) these lie in DD at distance more than 2r2r from its complement. Finally z10=cxz_{10}=c_x and w10=2r+(d+10)R′<4rw_{10}=2r+(d+10)R'<4r, so Pd+10⊂BxP_{d+10}\subset B_x.

Each Pk−1(R)P_{k-1}^{(R)} lies in Dvx∘D^\circ_{vx}, so (E2∘)(E2^\circ) gives (E2), and (E3) was just verified; also o∉Do\notin D. If o↔P0o\leftrightarrow P_0 but o↮Pd+10o\not\leftrightarrow P_{d+10}, then o↔Pk−1o\leftrightarrow P_{k-1} and o↮Pko\not\leftrightarrow P_k for some kk, and the union bound over the d+10d+10 applications of Lemma 4.1 proves the claim. □\square

Exploration and the proof of Theorem 1.1

We follow the exploration of [17 §6], with the long-range interfaces IvxI_{vx}. Let o=0∈Q0o=0\in Q_0, and let I\mathcal{I} be the event that all bits internal to Q0Q_0 are open. It prescribes finitely many bits, each of probability in (0,1)(0,1), so Pq(I)>0\mathbb{P}_q(\mathcal{I})>0, and under PI=Pq(⋅∣I)\mathbb{P}_{\mathcal{I}}=\mathbb{P}_q(\mathord\cdot\mid\mathcal{I}) all other bits are independent with their parameter-qq law.

The exploration. The coarse root 0∈Z20 \in\mathbb{Z}^{2} is declared good and visited, but is never queued or processed; its four neighbours are put in a queue with parent 00. A coarse vertex is visited once it has been queued, and pending until it is processed; vertices are processed in queue order, and none is queued twice. For a pending vv, let G(v)G(v) be the graph consisting of the cubes on the parent chain of vv back to the root, the bits internal to these cubes, and the interfaces II between successive cubes of the chain, and no other pairs. Put Hv={o↔Bv in G(v)}H_v=\{o \leftrightarrow B_v\text{ in }G(v)\}. To process vv, reveal the bits internal to QvQ_v and those of IpvI_{pv}, where pp is the parent of vv. For each unvisited neighbour xx of vv, let G′=G(v)∪Qx∪IvxG'=G(v)\cup Q_x\cup I_{vx} (with the bits internal to QxQ_x) and compute

bx=PI(o↔Bx in G′∣ history after this reveal),b_x=\mathbb{P}_{\mathcal{I}}(o\leftrightarrow B_x\text{ in }G'\mid\text{ history after this reveal}),

which reveals nothing. Declare vv good if HvH_v holds and bx>1−δb_x>1-\delta for every such xx, and in that case queue these xx with parent vv; otherwise declare vv bad. We maintain

PI(Hv∣revealed history)>1−δfor every pending v.(4)\mathbb{P}_{\mathcal{I}}(H_v\mid\text{revealed history})>1-\delta\qquad\text{for every pending }v. \tag*{(4)}

Lemma 6.1 (freshness and the initial promise). Conditional on any revealed history of positive probability, the bits that are neither prescribed by I\mathcal{I} nor revealed are independent with their parameter-qq law, and (4) holds at every step.

Proof. Each step reveals a finite set of bits determined by earlier revealed values, and all decisions are functions of revealed values; induction gives the first claim. The unknown bits of HxH_x for a pending xx with parent vv are internal to QxQ_x or lie in IvxI_{vx}. The cubes are vertex-disjoint, each step reveals the internal bits of one cube and its own parent interface, and G(x)G(x) contains no other pairs, so these bits stay unrevealed while xx is pending; together with the definition of good, this keeps (4) for every queued xx once it holds when xx is queued.

It remains to check (4) for the root’s neighbours. Let vv be one; by symmetry Q0Q_0 lies in direction −e1-e_1 from QvQ_v. Let ℓ′=10r−w∈L\ell'=10r-w\in\mathcal{L}, u=cv−we1u=c_v-we_1 and F=F1,−,(+,…,+)(u,ℓ′)F=F_{1,-,(+,\ldots,+)}(u,\ell'). Then Λℓ′(u)⊂Qv\Lambda_{\ell'}(u)\subset Q_v; Λm(u)⊂Bv\Lambda_m(u)\subset B_v since m+w≤4rm+w\le4r; and F⊂Q0F\subset Q_0, since its first coordinates lie in [(cv)1−10r−w,(cv)1−10r−1][(c_v)_1-10r-w,(c_v)_1-10r-1] and its other coordinates in [0,10r][0,10r]. On X(u,ℓ′,F)X(u,\ell',F), the box Λm(u)\Lambda_m(u) is joined inside QvQ_v to a vertex with an open edge of length at most ww into Q0Q_0, and this edge lies in I0vI_{0v}, so X(u,ℓ′,F)∩I⊂HvX(u,\ell',F)\cap\mathcal{I}\subset H_v (the cube Q0Q_0 is open and connected by (Pos)). The event X(u,ℓ′,F)X(u,\ell',F) depends only on bits internal to QvQ_v and on pairs of I0vI_{0v}, which are independent of I\mathcal{I}; hence PI(Hv)≥Pq(X(u,ℓ′,F))>1−γ>1−δ\mathbb{P}_{\mathcal{I}}(H_v)\ge\mathbb{P}_{q}(X(u,\ell',F))>1-\gamma>1-\delta by (S7) and (S2). □\square

Lemma 6.2. Immediately before a pending vv is processed, the conditional probability that vv is declared bad is less than ff.

Proof. Fix an unvisited neighbour xx of vv and let D=DvxD=D_{vx}. Neither vv (pending) nor xx (unvisited) has been processed and IvxI_{vx} is not a parent interface yet, so every bit internal to DD is fresh with parameter qq, and (Fq(D))(F_q(D)) holds in G′G'. The pairs of G′G' with one endpoint in DD and the other outside are those of IpvI_{pv}, which is still unrevealed, with λ=qJ\lambda=qJ; so (E20^{0}) holds. Since o∉Do\notin D, Lemma 5.1 gives PI(Hv∩{o↮Bx in G′}∣history)≤(d+10)η\mathbb{P}_{\mathcal{I}}(H_v\cap\{o\not\leftrightarrow B_x\text{ in }G'\}\mid\text{history})\le(d+10)\eta, and the tower property gives PI(Hv,bx≤1−δ∣history)≤(d+10)η/δ\mathbb{P}_{\mathcal{I}}(H_v,b_x\le1-\delta\mid\text{history})\le(d+10)\eta/\delta. With at most four such xx and (4), PI(v bad∣history)≤δ+4(d+10)η/δ<f\mathbb{P}_{\mathcal{I}}(v\text{ bad}\mid\text{history})\le\delta+4(d+10)\eta/\delta<f by (S2). □\square

Lemma 6.3. For every deterministic set of kk distinct coarse vertices, the PI\mathbb{P}_{\mathcal{I}}-probability that all of them are processed and declared bad is at most fkf^k.

Proof. The root is good, so assume it is not in the set SS. Let τj\tau_j be the time at which the jjth member of SS is processed (τj=∞\tau_j = \infty if there is none) and EjE_j the event that τj<∞\tau_j < \infty and the first jj processed members are bad. For each finite tt the event {τj=t}∩Ej−1\{\tau_j = t\} \cap E_{j-1} is determined before the reveal at time tt, so Lemma 6.2 gives PI(Ej)≤f∑tPI(τj=t,Ej−1)≤fPI(Ej−1)\mathbb{P}_{\mathcal{I}}(E_j) \le f \sum_t \mathbb{P}_{\mathcal{I}}(\tau_j = t, E_{j-1}) \le f\mathbb{P}_{\mathcal{I}}(E_{j-1}), and induction finishes the proof.

Proof of Theorem 1.1. By Lemma 2.4, 0<βc<∞0 < \beta_c < \infty. Suppose that θ(βc)>0\theta(\beta_c) > 0 and choose the parameters of Lemma 3.3 with β=βc\beta= \beta_c. Let X⊂Z2X \subset\mathbb{Z}^2 be the set of good coarse vertices. If XX is finite, the queue empties after finitely many steps, and every coarse vertex outside XX adjacent to XX has been processed and declared bad. Apply Lemma 2.3 to XX: its boundary contains a cycle of some length nn from a family of at most 2n3n2n3^n candidates, each with k≥n/7k \ge n/7 crossed edges with disjoint endpoints, each of which has a bad endpoint outside XX. Choosing one endpoint per edge in 2k2^k ways, Lemma 6.3 bounds the probability of a given candidate by (2f)k≤(2f)n/7(2f)^k \le(2f)^{n/7}. By (S1), PI(∣X∣<∞)≤∑n2n3n(2f)n/7<1\mathbb{P}_{\mathcal{I}}(|X| < \infty) \le\sum_n 2n3^n(2f)^{n/7} < 1. Every good non-root vertex vv satisfies HvH_v, and the boxes BvB_v are disjoint, so with positive PI\mathbb{P}_{\mathcal{I}}-probability the open cluster of oo in the long-range model is infinite. Hence θ(q)≥Pq(I)PI(∣X∣=∞)>0\theta(q) \ge\mathbb{P}_q(\mathcal{I})\mathbb{P}_{\mathcal{I}}(|X| = \infty) > 0 with q<βcq < \beta_c, a contradiction.

The corollaries, and where the hypotheses enter

Proof of Corollary 1.2. If α>d\alpha> d then T(n)≤C′n−αT(n) \le C'n^{-\alpha}, so (Tail) holds, and (Pos) holds by the lower bound; apply Theorem 1.1. If α<d\alpha< d, put s=d+α<2ds = d + \alpha< 2d and Pk=1−e−βJ(k)P_k = 1 - e^{-\beta J(k)} at a parameter β\beta with θ(β)>0\theta(\beta) > 0. Berger’s theorem [3 Theorem 2.6] states: if {Pk}\{P_k\} is percolating and lim inf⁡∥k∥→∞Pk∥k∥s>0\liminf_{\|k\|\to\infty} P_k\|k\|^s > 0 for some s<2ds < 2d, then {(1−ϵ)Pk}\{(1-\epsilon)P_k\} is percolating for some ϵ>0\epsilon> 0 (his standing assumption is 0≤Pk=P−k<10 \le P_k = P_{-k} < 1, which holds here). Monotonicity and 1−e−t≥t/21-e^{-t} \ge t/2 on [0,1][0,1] give Pk≥(βc/2)∥k∥−s≥(βc/2)∥k∥1−sP_k \ge(\beta_c/2)\|k\|^{-s} \ge(\beta_c/2)\|k\|_1^{-s} for large ∥k∥\|k\|. Concavity of t↦1−e−tt \mapsto1-e^{-t} gives 1−e−(1−ϵ)βJ≥(1−ϵ)(1−e−βJ)1-e^{-(1-\epsilon)\beta J} \ge(1-\epsilon)(1-e^{-\beta J}), so θ((1−ϵ)β)>0\theta((1-\epsilon)\beta) > 0. Thus θ(β)>0\theta(\beta) > 0 forces β>βc\beta> \beta_c.

Proof of Corollary 1.3. Exponential decay and finite range both give ndT(n)→0n^dT(n) \to0.

Remark 7.1 (where the hypotheses enter). The tail condition is used only in (S5), at one pair of scales (w,r)(w,r) with rr a large constant multiple of ww fixed by (S6): in Lemma 3.2, for exits from Λℓ\Lambda_\ell by edges longer than ww, and in the event ElongE_{\mathrm{long}} of Lemma 4.1, for entrances that jump over the shell. The layer argument needs R≫eLwR \gg e^Lw and the relay needs r≫R+wr \gg R+w; both hold with r≤C0wr \le C_0w, and (Tail) then makes β(41r+1)dT(w)≤β(42C0)dwdT(w)\beta(41r+1)^dT(w) \le\beta(42C_0)^dw^dT(w) small. Summability gives the uniform bound az≤Aa_z \le A, which turns a large crossing weight into many seed attempts. (Sym) gives the equal probabilities of the faces in Lemma 3.2 and the orientation reductions. (Pos) gives βc<∞\beta_c < \infty and the connectivity of open boxes. The dimension d≥2d \ge2 enters through the coordinate plane in Lemma 2.4 and the coarse lattice Z2\mathbb{Z}^2 with its contour count.

Remark 7.2 (the borderline). For J(x)≍∥x∥−2dJ(x) \asymp\|x\|^{-2d} the quantity ndT(n)n^dT(n) stays bounded away from zero, and nothing in the argument guarantees (S5) and (S6) together: long edges at scale nn leave the cube [−n,n]d[-n,n]^d with probability of order one at every scale. In d=1d=1 and J(x)≍∣x∣−2J(x) \asymp|x|^{-2} the transition is discontinuous [1, 5], and the argument above also lacks the coarse plane. Apart from sufficiently spread-out kernels in d≥7d \ge7, where d>6=3(α∧2)d > 6 = 3(\alpha\wedge2) and [12] applies, whether θ(βc)=0\theta(\beta_c) = 0 for α=d≥2\alpha= d \ge2 remains open.

References

  1. [1]M. Aizenman and C. M. Newman. Discontinuity of the percolation density in one-dimensional 1/|x − y|² percolation models. Comm. Math. Phys., 107(4):611–647, 1986.
  2. [2]O. Angel, D. de la Riva, C. Panagiotis, and Y. Shi. Continuity of the phase transition for long-range Bernoulli percolation with exponential decay on ℤ². Preprint, arXiv:2610.35233, 2026.arxiv.org/abs/2610.35233
  3. [3]N. Berger. Transience, recurrence and critical behavior for long-range percolation. Comm. Math. Phys., 226:531–558, 2002. doi:10.1007/s002200200617, arXiv:math/0110296.DOI
  4. [4]A. Blanc-Renaudie. Notes on Kozma and Nitzan’s proof of θ(p_c) = 0. Preprint, arXiv:2610.10502, 2026.
  5. [5]H. Duminil-Copin, C. Garban, and V. Tassion. Long-range models in 1D revisited. Preprint, arXiv:2011.04642, 2020.
  6. [6]H. Duminil-Copin, V. Sidoravicius, and V. Tassion. Absence of infinite cluster for critical Bernoulli percolation on slabs. Comm. Pure Appl. Math., 69(7):1397–1411, 2016.
  7. [7]R. Fitzner and R. van der Hofstad. Mean-field behavior for nearest-neighbor percolation in d > 10. Electron. J. Probab., 22:43, 2017. doi:10.1214/17-EJP56.DOI
  8. [8]J.-B. Goûéré. On the proof of the Kozma–Nitzan conjecture. Preprint, arXiv:2610.30063, 2026.arxiv.org/abs/2610.30063
  9. [9]G. R. Grimmett and J. M. Marstrand. The supercritical phase of percolation is well behaved. Proc. Roy. Soc. London Ser. A, 430(1879):439–457, 1990.DOI
  10. [10]T. Hara and G. Slade. Mean-field critical behaviour for percolation in high dimensions. Comm. Math. Phys., 128(2):333–391, 1990.
  11. [11]T. E. Harris. A lower bound for the critical probability in a certain percolation process. Proc. Cambridge Philos. Soc., 56:13–20, 1960.DOI
  12. [12]M. Heydenreich, R. van der Hofstad, and A. Sakai. Mean-field behavior for long- and finite range Ising model, percolation and self-avoiding walk. J. Stat. Phys., 132:1001–1049, 2008. doi:10.1007/s10955-008-9580-5.DOI
  13. [13]T. Hutchcroft. Power-law bounds for critical long-range percolation below the upper-critical dimension. Probab. Theory Related Fields, 181:533–570, 2021. arXiv:2008.11197.arxiv.org/abs/2008.11197
  14. [14]H. Kesten. The critical probability of bond percolation on the square lattice equals 1/2. Comm. Math. Phys., 74:41–59, 1980.DOI
  15. [15]G. Kozma and S. Nitzan. A reduction of the θ(p_c) = 0 problem to a conjectured inequality. Preprint, arXiv:2401.12397, 2024.arxiv.org/abs/2401.12397
  16. [16]J. Leder. θ(p_c) = 0 for Bernoulli bond percolation on ℤ^d in all dimensions d ≥ 2, in Lean 4/Mathlib, 2026. https://github.com/anthropics/formal-math/tree/795efb86f191735c5481675763537cfb4ff37e55/percolation.
  17. [17]OpenAI. Critical bond and site percolation on the cubic lattice, 2026. https://github.com/openai/math/tree/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Critical-bond-and-site-percolation-on-the-cubic-lattice-September-24-2026.DOI
  18. [18]OpenAI. No percolation at criticality on quasi-transitive graphs, 2026. https://github.com/openai/math/tree/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/No-percolation-at-criticality-on-quasi-transitive-graphs-September-24-2026.DOI

Paper details

vlnc:W_mu6bavfa
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents