betaFeedback

WorksMathematics

Critical bond and site percolation on the cubic lattice

Abstract

We prove that nearest-neighbor Bernoulli bond and site percolation on Z3\mathbb Z^3 have no infinite cluster at their respective critical parameters. The proof combines a finite connection inequality for independent hyperedges with a finite-scale extension estimate and an adaptive exploration.

Introduction

In Bernoulli bond percolation on the nearest-neighbor cubic lattice Z3\mathbb{Z}^3, each edge is independently open with probability pp. In Bernoulli site percolation, the independent open bits instead belong to the vertices, and an open path must have every vertex open, including its endpoints. Write Pp\mathbb{P}_p for the law of the model under consideration and define its critical parameter by

pc=inf⁡{p∈[0,1]:Pp(0 belongs to an infinite open cluster)>0}.p_c = \inf\{p \in[0,1] : \mathbb{P}_p(0\text{ belongs to an infinite open cluster}) > 0\}.

By translation invariance and a countable union over vertices, the same threshold is obtained by asking for the existence of any infinite cluster. The bond and site thresholds are defined separately throughout.

Theorem 1.1. For nearest-neighbor Bernoulli bond percolation on Z3\mathbb{Z}^3, there is almost surely no infinite open cluster at the bond critical parameter. The same assertion holds for nearest-neighbor Bernoulli site percolation at the site critical parameter.

Write θ(p)=Pp(0 belongs to an infinite open cluster)\theta(p)=\mathbb{P}_p(0\text{ belongs to an infinite open cluster}). The theorem gives continuity at the transition: θ(p)\theta(p) tends to zero as p↓pcp \downarrow p_c. To see this consequence directly, for n≥1n \ge1 let EnE_n be the event that the origin reaches the boundary of the box [−n,n]3∩Z3[-n,n]^3 \cap\mathbb{Z}^3 by an open path inside that box. Then Pp(En)↓θ(p)\mathbb{P}_p(E_n) \downarrow\theta(p) as n→∞n \to\infty. At pcp_c, the theorem makes this limit zero. Given ε>0\varepsilon> 0, first choose nn with Ppc(En)<ε\mathbb{P}_{p_c}(E_n) < \varepsilon; since EnE_n depends on finitely many bits, the same strict inequality holds for pp sufficiently close to pcp_c. Thus θ(p)≤Pp(En)<ε\theta(p) \le\mathbb{P}_p(E_n) < \varepsilon there. The proof below establishes critical extinction for both models without using an external criticality theorem.

History and the three-dimensional obstacle. Broadbent and Hammersley introduced percolation to study transport through a random medium [2]. In two dimensions, Harris’s critical nonpercolation theorem and Kesten’s identification of the square bond threshold established the bond result [9, 11]. Russo proved critical nonpercolation for nearest-neighbor site percolation on the square lattice [14], Theorem 2(b). These planar results use geometry that is unavailable for paths in three dimensions.

In sufficiently high dimensions, Hara and Slade established mean-field critical behavior through the lace expansion [8]. Fitzner and van der Hofstad obtained the nearest-neighbor bond range d≥11d \ge11 [6]. Hara and Slade also noted that their method applies to site percolation [8], p. 335. Heydenreich and Matzke later gave a detailed site lace-expansion treatment, proving the triangle condition in sufficiently high dimensions and its critical nonpercolation consequence [10], Theorem 1.1 and Section 1.3. The explicit bond dimension bound is not a site bound. Benjamini and Schramm formulated the broader criticality question for quasi-transitive graphs with pc<1p_c < 1 [1], Conjecture 4.

Grimmett and Marstrand proved convergence of slab thresholds to the full-lattice threshold and developed a dynamic renormalization argument using open seeds and sequentially controlled exploration [7]. Duminil-Copin, Sidoravicius, and Tassion subsequently proved critical nonpercolation for bond percolation on slabs [5]. As their discussion explains, extinction at every fixed slab threshold, together with convergence of those thresholds, does not by itself give critical extinction in the full cubic lattice. The proof here obtains the required connections between finite regions from a finite comparison, and fixes all scales before decreasing the percolation parameter.

Leder’s public account of a Lean formalization reports critical bond nonpercolation on every nearest-neighbor lattice Zd\mathbb{Z}^d, d≥2d \ge2 [13]. Its first-contact comparison [13], Section 3.4, (GEN) already implies the finite joint connection inequality for bonds: take the increasing cluster functional to indicate meeting the target set. The subsequent public Kozma–Nitzan exposition explicitly records this consequence and reports stronger finite comparisons [4], “Conjecture 1 from the Lean proof of Conjecture 3”. The related public site exposition reports the independent labelled-hyperedge comparison, a conditioned-covariance hierarchy, and critical site nonpercolation on Zd\mathbb{Z}^d for every d≥3d \ge3 [3], “One inequality for both models”, “Conditional covariance inequalities”, and “From (1) to the lattice”. These two public expositions describe work produced by AI systems prompted by Ahmed Bou-Rabee. Our purpose here is a self-contained cubic argument, including the finite comparison, explicit relay geometry, and exploration estimates.

Finite comparison and the proof method. Our main finite ingredient is a connection inequality for independent, possibly nonuniform hyperedges. For a node oo and nonempty node set AA, it asserts that

P(o↔A, o↔T)≥P(o↔A)min⁡a∈AP(a↔T).(1)\mathbb{P}(o \leftrightarrow A,\ o \leftrightarrow T) \ge\mathbb{P}(o \leftrightarrow A) \min_{a \in A} \mathbb{P}(a \leftrightarrow T). \tag*{(1)}

Here an open hyperedge joins all of its nodes. The inequality implies that if every possible relay in AA reaches a target TT with probability at least 1−b1-b, then the probability of reaching AA but missing TT is at most bb. This failure bound also holds for sites, and after any set of bits has been fixed while the remaining bits retain their product law. Thus it remains available during an exploration.

Taking TT to be a singleton in (1.1) and enlarging the event on the left gives the multiplicative gluing inequality of Kozma and Nitzan [12], Conjecture 1. They also record that van den Berg and Engelenburg independently considered this inequality. Section 4 of Kozma and Nitzan connects finite comparisons to criticality through renormalization. The proof of (1.1) uses increasing functions of clusters, a nonnegative cone of conditional connection columns, and alternating conditional cluster resampling. Conditional resampling has precedents in percolation correlation inequalities [15], Section 2.1. We prove the required comparison and every lattice relay estimate locally. The general bond criticality theorem does not imply the site assertion: independent site bits become independent hyperedges, and their endpoint convention must be preserved. The finite hyperedge comparison and its stability under fixed bits are useful independently of the lattice application.

Proof overview. Assume that an infinite cluster exists at pcp_c. Large seed boxes then connect with high probability to each of the 24 quarter-faces of a surrounding cube. Only finitely many radii are needed, so these estimates persist at one parameter q<pcq < p_c. The finite comparison lets us extend a connection through a fresh region. The seed and sequential-exploration architecture is related to the dynamic renormalization method of Grimmett and Marstrand [7], Sections 3–4. Here many well-separated entrance points provide independent chances for an open seed, while relay quality is defined using the interior bits alone. A sequence of thirteen elementary geometric moves transfers a connection from the inner box of one coarse cube to that of its neighbor. Finally, we explore cubes indexed by Z2\mathbb{Z}^2, retaining a high conditional connection probability for each pending cube. Each processed cube has a uniformly small conditional chance of failure. A planar boundary count then gives a positive chance of infinitely many successful cubes, and hence an infinite open cluster at q<pcq < p_c, a contradiction. Neither uniqueness of the infinite cluster nor independence of the successful coarse cubes is assumed.

Section 2 proves the finite comparison. Section 3 gives the planar count and fixes the scales. Sections 4 and 5 prove the extension and neighboring-box estimates. Section 6 constructs the exploration and completes the proof.

Connection conventions. For a vertex xx and vertex set TT, the notation x↔Tx \leftrightarrow T means that an open path in the specified graph joins xx to some vertex of TT. For two sets, at least one pair of their vertices must be so joined. We permit paths of zero edges. Such a path always gives a self-connection for bonds and hyperedges, but for sites its single vertex must be open. All graphs and hypergraphs in the finite comparison are undirected.

A finite connection comparison

The lattice argument will need to pass from a connection to a set of possible relays to a connection to a target. We prove the finite inequality that makes this passage possible, allowing different probabilities for the independent bits so that the result remains available after some bits have been fixed.

A finite hypergraph has a finite node set and a finite labelled family of nonempty subsets called hyperedges; distinct labels may represent the same subset. Each hyperedge is independently open with its own probability; an open hyperedge joins all its nodes. The cluster C(x)C(x) always contains xx, and C(S)=⋃x∈SC(x)C(S) = \bigcup_{x \in S} C(x). All connections in this section are in finite undirected graphs or hypergraphs.

Proposition 2.1 (Joint connection comparison). In independent percolation on a finite hypergraph, with arbitrary hyperedge probabilities in [0,1][0,1], let oo be a node and let A,TA,T be node sets with A≠∅A \ne\varnothing. Then

P(o↔A, o↔T)≥P(o↔A)min⁡a∈AP(a↔T).(2)\mathbb{P}(o \leftrightarrow A,\ o \leftrightarrow T) \ge\mathbb{P}(o \leftrightarrow A)\min_{a \in A}\mathbb{P}(a \leftrightarrow T). \tag*{(2)}

For bonds, subtraction immediately turns this inequality into the failure bound needed later. Sites require an additional check because a hypergraph node is always connected to itself, whereas a site self-connection requires an open vertex. We give that check at the end of the section.

The proof of Proposition 2.1 assigns each cluster meeting an ordered list to its first node in that list. Conditional connection probabilities form the columns of a triangular matrix. The positivity property below will bound the source’s joint connection probability from below by a nonnegative weighted sum of relay-to-target probabilities. The weights will total the probability of reaching the relay set, giving the proposition.

Deleting a node set ww means deleting its nodes and every hyperedge meeting it; write G∖wG \setminus w for the resulting hypergraph. On an exact-cluster event C(S)=wC(S) = w of positive probability, the hyperedges of G∖wG \setminus w retain their independent original laws. Indeed, the event asks that every node of ww connect to SS within ww, and that all hyperedges crossing from ww to its complement be closed. It tests no hyperedge disjoint from ww. The same observation applies after an earlier deletion.

Lemma 2.2 (Cluster columns). Let GG be a finite hypergraph whose hyperedges have independent open probabilities strictly less than one. Choose any ordered list of distinct nodes x1,…,xnx_1,\ldots,x_n; the list need not contain every node. Define

Sk={xi:i<k},Ek={C(xk)∩Sk=∅},dk=P(Ek)>0,S_k=\{x_i:i<k\},\qquad E_k=\{C(x_k)\cap S_k=\varnothing\},\qquad d_k=\mathbb{P}(E_k)>0,

and let μk\mu_k be the conditional law of C(xk)C(x_k) given EkE_k, regarded as a measure on

Dk={U:xk∈U, U∩Sk=∅}.\mathcal{D}_k=\{U:x_k\in U,\ U\cap S_k=\varnothing\}.

For a function FF on Dk\mathcal{D}_k, let

vkF(i)=Eμk[1{xi∈U}F(U)],hk=vk1,H=(h1 ⋯ hn).v_k^F(i)=\mathbb{E}_{\mu_k}\left[1_{\{x_i\in U\}}F(U)\right],\qquad h_k=v_k^1,\qquad H=(h_1\ \cdots\ h_n).

If FF is nonnegative and increasing under inclusion, then

vkF−μk(F)hk∈cone⁡{hj:k<j≤n}.v_k^F-\mu_k(F)h_k\in\operatorname{cone}\{h_j:k<j\le n\}.

Here the cone consists of all nonnegative linear combinations, including zero. In particular, H−1vkFH^{-1}v_k^F is coordinatewise nonnegative.

Proof. Closing every hyperedge incident to xkx_k proves dk>0d_k>0. The matrix HH is unit lower triangular: its kkth column vanishes above row kk and has entry one in row kk. We prove (2.2) by descending induction on kk. For k=nk=n, both vnFv_n^F and μn(F)hn\mu_n(F)h_n equal μn(F)en\mu_n(F)e_n, where ene_n is the nnth coordinate vector.

Fix k<nk<n and assume the assertion for every later column. Write s=xks=x_k, S=SkS=S_k, and W=C(S)W=C(S). For a set w⊇Sw\supseteq S avoiding ss, put

c(w)=E[F(CG∖w(s))],c(w)=\mathbb{E}\left[F(C^{G\setminus w}(s))\right],

where the expectation uses the original product law on the undeleted hyperedges. Restriction of the same configuration shows that c(w)c(w) is nonincreasing in ww.

A first-listed-node decomposition. Fix a row ii. For j>kj>k, let LijL_{ij} be the event that the first listed node of C(xi)C(x_i) is xjx_j. Condition on an exact value W=wW=w avoiding ss. Then the mean of F(C(s))F(C(s)) is c(w)c(w). If xi∉wx_i\notin w, its cluster contains the listed node xix_i and avoids all listed nodes preceding ss, which belong to ww. Its first listed node is therefore ss precisely when xi∈C(s)x_i\in C(s); otherwise exactly one of the events LijL_{ij} occurs. On LijL_{ij}, condition further on C(xj)=UC(x_j)=U. The set UU avoids ss, and the specification of LijL_{ij} adds only membership and avoidance conditions on UU. Exact-cluster conditioning therefore leaves the mean of F(C(s))F(C(s)) equal to c(w∪U)c(w\cup U). Partitioning the expectation c(w)c(w) and subtracting c(w)c(w) times the corresponding partition of total probability gives

E[1{xi∈C(s)}F(C(s))∣W=w]=c(w)P(xi∈C(s)∣W=w)+∑j>kE[1Lij{c(w)−c(w∪C(xj))}∣W=w].(3)\begin{aligned} \mathbb{E}\left[1_{\{x_i\in C(s)\}}F(C(s))\mid W=w\right] ={}&c(w)\mathbb{P}(x_i\in C(s)\mid W=w)\\ &+\sum_{j>k}\mathbb{E}\left[1_{L_{ij}}\{c(w)-c(w\cup C(x_j))\}\mid W=w\right]. \tag*{(3)} \end{aligned}

The expressions involving c(w∪C(xj))c(w\cup C(x_j)) are evaluated only on LijL_{ij}. If xi∈wx_i\in w, all LijL_{ij} are empty under the conditioning, and both sides vanish.

For UU disjoint from SS, let WU=CG\U(S)W_U=C^{G\backslash U}(S). Define, for U∈DkU\in \mathcal{D}_k,

(PF)(U)=E[c(WU)],(PF)(U)=\mathbb{E}[c(W_U)],

and, for UU disjoint from S∪{s}S\cup\{s\}, define

RF(U)=E[1{s∉WU}{c(WU)−c(WU∪U)}].(4)R_F(U)=\mathbb{E}\left[\mathbf{1}_{\{s\notin W_U\}}\{c(W_U)-c(W_U\cup U)\}\right]. \tag*{(4)}

with zero integrand when s∈WUs\in W_U. The operator PP preserves nonnegativity and monotonicity: enlarging UU shrinks WUW_U, whereas cc is nonincreasing.

We now average (3) conditional on EkE_k. Given C(s)=UC(s)=U avoiding SS, the conditional law of WW is that of WUW_U. Thus the first term on the right averages to

Eμk[1{xi∈U}(PF)(U)]=vkPF(i).\mathbb{E}_{\mu_k}\left[\mathbf{1}_{\{x_i\in U\}}(PF)(U)\right]=v_k^{PF}(i).

For a residual term, the exact event identity is

Ek∩Lij=Ek∩Ej∩{xi∈C(xj)}.E_k\cap L_{ij}=E_k\cap E_j\cap\{x_i\in C(x_j)\}.

Indeed, a cluster avoiding the earlier listed nodes also avoids their entire clusters. Given C(xj)=UC(x_j)=U under EjE_j, the conditional law of WW is again that of WUW_U, and the remaining restriction EkE_k is s∉WUs\notin W_U. Consequently the unnormalized residual expectation is

E[1Ek∩Lij{c(W)−c(W∪C(xj))}]=djEμj[1{xi∈U}RF(U)],\mathbb{E}\left[\mathbf{1}_{E_k\cap L_{ij}}\{c(W)-c(W\cup C(x_j))\}\right] =d_j\mathbb{E}_{\mu_j}\left[\mathbf{1}_{\{x_i\in U\}}R_F(U)\right],

where the expression on the left is evaluated only when its indicator is one. Dividing by dkd_k and summing over j>kj>k proves the column identity

vkF=vkPF+∑j>kdjdkvjRF.v_k^F=v_k^{PF}+\sum_{j>k}\frac{d_j}{d_k}v_j^{R_F}.

Here and below RFR_F is restricted to Dj\mathcal{D}_j in the jjth column. To complete the induction, we show that RFR_F is nonnegative and increasing and that PtFP^tF converges uniformly to μk(F)\mu_k(F).

The residual is increasing. Sample all hyperedge bits once, and use their restrictions after every deletion. For deterministic UU disjoint from S∪{s}S\cup\{s\}, the event WU=wW_U=w tests only bits of hyperedges disjoint from UU and meeting ww. Every full-graph hyperedge disjoint from ww therefore retains its independent original law after this conditioning, including hyperedges that meet UU. When s∉ws\notin w, the two clusters in the following difference consequently have the marginal laws defining c(w)c(w) and c(w∪U)c(w\cup U). Hence RF(U)=E[DU]R_F(U)=\mathbb{E}[D_U], where

DU=1{s∉WU}{F(CG\WU(s))−F(CG\(WU∪U)(s))}.(5)D_U=\mathbf{1}_{\{s\notin W_U\}}\left\{F\left(C^{G\backslash W_U}(s)\right)-F\left(C^{G\backslash(W_U\cup U)}(s)\right)\right\}. \tag*{(5)}

again interpreted as zero on the excluded event. All clusters at which FF is evaluated contain ss and avoid SS.

The first cluster contains the second, so DU≥0D_U\ge0. To compare two sets U⊆U′U\subseteq U', first note that WU′⊆WUW_{U'}\subseteq W_U. On {s∉WU}\{s\notin W_U\}, deleting the smaller set WU′W_{U'} makes the first cluster larger. The second cluster has a different monotonicity:

CG\(WU∪U)(s)=CG\U(s).C^{G\backslash(W_U\cup U)}(s)=C^{G\backslash U}(s).

Indeed, in G\UG\backslash U the source cluster is disjoint from the union WUW_U of the SS-clusters, so deleting WUW_U removes none of its open connections. The same equality holds for U′U', because s∉WU′s\notin W_{U'}. Thus the second cluster becomes smaller when UU is replaced by U′U'. Both changes increase the difference of the two values of FF, giving DU′≥DUD_{U'}\ge D_U. On {s∈WU}\{s\in W_U\}, this inequality follows instead from DU=0D_U=0 and DU′≥0D_{U'}\ge0. Taking expectations proves that RFR_F is nonnegative and increasing.

Conditional resampling and the limit. The operator PP is a Markov transition: from UU, first sample WUW_U, then independently sample CG∖WU(s)C^{G\setminus W_U}(s). Under the joint law of (C(s),W)(C(s),W) conditional on EkE_k, these are the successive conditional laws of WW given C(s)C(s) and of C(s)C(s) given WW. Hence μk\mu_k is stationary for PP. Alternating conditional cluster resampling under a separation condition is the method used by van den Berg, Häggström, and Kahn [15], Section 2.1, proof of Theorem 2.1; the cone conclusion here is proved by the following iteration.

Every transition assigns probability at least

a=∏e: s∈e(1−pe)>0a=\prod_{e:\,s\in e}(1-p_e)>0

to the singleton {s}\{s\}, since closing every incident hyperedge suffices to isolate ss in any deletion graph. Thus, for every real function gg on the finite state space Dk\mathcal{D}_k, subtracting this common mass gives

max⁡Pg−min⁡Pg≤(1−a)(max⁡g−min⁡g).\max Pg-\min Pg\leq(1-a)(\max g-\min g).

Iteration makes the oscillation of PtFP^tF tend to zero. Its μk\mu_k-mean remains μk(F)\mu_k(F) by stationarity, so

sup⁡U∈Dk∣(PtF)(U)−μk(F)∣≤(1−a)t(max⁡F−min⁡F)⟶0.\sup_{U\in\mathcal{D}_k}\left|(P^tF)(U)-\mu_k(F)\right|\leq(1-a)^t(\max F-\min F)\longrightarrow0.

This proves the required uniform convergence, including at states of Dk\mathcal{D}_k with zero μk\mu_k-mass.

By the induction hypothesis, each vjRFv_j^{R_F} is a nonnegative combination of hj,…,hnh_j,\ldots,h_n. Equation (2.5) therefore puts vkF−vkPFv_k^F-v_k^{PF} in the cone on the right of (2.2). All iterates PtFP^tF remain nonnegative increasing, so iteration puts vkF−vkPtFv_k^F-v_k^{P^tF} in that cone as well. The cone is closed, because its generators belong to the basis given by the columns of HH. Taking the uniform limit proves (2.2). Finally, adding μk(F)hk\mu_k(F)h_k expresses vkFv_k^F as a nonnegative combination of columns of HH, proving the last assertion. □

Proof of Proposition 2.1. First suppose all probabilities are below one. Use the notation of Lemma 2.2. For F(U)=1{xj∈U for some j>k}F(U)=1_{\{x_j\in U\text{ for some }j>k\}}, isolation of xkx_k gives μk(F)<1\mu_k(F)<1, and direct inspection of the rows gives

vkF−μk(F)hk=(1−μk(F))(hk−ek).v_k^F-\mu_k(F)h_k=(1-\mu_k(F))(h_k-e_k).

where eke_k is the kkth coordinate vector. Lemma 2.2 yields nonnegative coefficients Γjk\Gamma_{jk} such that

hk=ek+∑j>kΓjkhj.h_k=e_k+\sum_{j>k}\Gamma_{jk}h_j.

With Γ\Gamma zero on and above the diagonal, these column identities say H=I+HΓH=I+H\Gamma, or H−1=I−ΓH^{-1}=I-\Gamma. Also, for d=(d1,…,dn)Td=(d_1,\ldots,d_n)^{\mathsf T},

Hd=1,(I−Γ)1=d.(6)Hd=\mathbf{1},\qquad(I-\Gamma)\mathbf{1}=d. \tag*{(6)}

Indeed, the cluster of xix_i has first listed node xkx_k with probability dkhk(i)d_kh_k(i), and these events partition the sample space.

If o∈Ao\in A, (2) is immediate. Otherwise take the list to consist exactly of the nodes of AA, in any order, followed by xn=ox_n=o. Set

F(U)=1{U∩A≠∅, U∩T≠∅},yi=E[F(C(xi))].F(U)=1_{\{U\cap A\ne\varnothing,\ U\cap T\ne\varnothing\}},\qquad y_i=\mathbb{E}[F(C(x_i))].

Partitioning by the first listed node gives y=∑kdkvkFy=\sum_k d_kv_k^F, where FF is restricted to Dk\mathcal{D}_k in each summand. Every restriction is nonnegative increasing, so Lemma 2.2 gives (I−Γ)y≥0(I-\Gamma)y\geq0 coordinatewise. In the last row,

yn≥∑i<nΓniyi.y_n\geq\sum_{i<n}\Gamma_{ni}y_i.

For i<ni<n, yi=P(xi↔T)y_i=\mathbb{P}(x_i\leftrightarrow T), while yn=P(o↔A, o↔T)y_n=\mathbb{P}(o\leftrightarrow A,\ o\leftrightarrow T). By (6), the nonnegative weights in this row sum to

∑i<nΓni=1−dn=P(o↔A).\sum_{i<n}\Gamma_{ni}=1-d_n=\mathbb{P}(o\leftrightarrow A).

This proves (2). Finally, finite event probabilities are polynomials in the hyperedge parameters, and the minimum in (2) is finite. Continuity therefore allows probabilities equal to one as well.

Corollary 2.3 (Failure comparison). Consider independent bond or site percolation on a finite undirected graph, with arbitrary individual probabilities in [0,1][0,1]. For a vertex oo, vertex sets A,TA,T, and b∈[0,1]b\in[0,1], if

P(a↔T)≥1−b(a∈A),\mathbb{P}(a\leftrightarrow T)\ge1-b \qquad(a\in A),

then

P(o↔A, o↮T)≤b.(7)\mathbb{P}(o\leftrightarrow A,\ o\not\leftrightarrow T)\le b. \tag*{(7)}

The assertion therefore applies after any deterministic collection of bits has been fixed.

Proof. The empty-set case A=∅A=\varnothing is immediate. For bonds, Proposition 2.1 and subtraction give the stronger upper bound bP(o↔A)≤bb\mathbb{P}(o\leftrightarrow A)\le b.

For sites, use the incidence construction described in [3] (“One inequality for both models”). Create a node x∗x^* for every graph vertex xx and a connector node for every graph edge. For each site xx, form a hyperedge consisting of x∗x^* and all connector nodes corresponding to edges incident to xx; its independent open bit is the bit of site xx. For distinct vertices x,zx,z, connection of x∗x^* to z∗z^* is equivalent to an open site path from xx to zz. Hypergraph self-connection, however, always holds. Writing A∗,T∗A^*,T^* for the starred images, we consequently have

Phyp(a∗↔T∗)≥Psite(a↔T)≥1−b.\mathbb{P}_{\mathrm{hyp}}(a^*\leftrightarrow T^*)\ge\mathbb{P}_{\mathrm{site}}(a\leftrightarrow T)\ge1-b.

On the site event in (7), the origin oo is open. Its hypergraph connections to starred vertices therefore agree with its site connections, including a possible zero-edge path to itself. This gives the event inclusion

{o↔A, o↮T}site⊆{o∗↔A∗, o∗↮T∗}hyp.\{o\leftrightarrow A,\ o\not\leftrightarrow T\}_{\mathrm{site}}\subseteq\{o^*\leftrightarrow A^*,\ o^*\not\leftrightarrow T^*\}_{\mathrm{hyp}}.

Applying Proposition 2.1 and subtraction to the hypergraph bounds the latter event by bb, proving the site assertion. Fixed bits are exactly parameters equal to zero or one, already allowed in the statement.

Planar boundaries and finite scales

We now fix either the bond or the site model. This section supplies a planar boundary count and chooses all the finite scales needed to lower the parameter. The boundary count will be used first to show pc<1p_c<1, and then to prove that the adaptive exploration survives.

A deterministic boundary count

An edge of the dual square lattice Z2+(1/2,1/2)\mathbb{Z}^2+(1/2,1/2) crosses exactly one primal edge of Z2\mathbb{Z}^2. For a finite set X⊂Z2X\subset\mathbb{Z}^2, its dual edge boundary consists of the dual edges crossing primal edges with exactly one endpoint in XX.

Lemma 3.1 (Boundary cycles). There are deterministic families Cn\mathcal{C}_n of simple dual cycles of length nn, with ∣Cn∣≤2n3n|\mathcal{C}_n| \le2n3^n, such that the dual edge boundary of every finite X⊂Z2X \subset\mathbb{Z}^2 containing 00 includes a cycle in ⋃n≥1Cn\bigcup_{n\ge1} \mathcal{C}_n. For each cycle in Cn\mathcal{C}_n, one can fix at least n/7n/7 of its crossed primal edges with pairwise disjoint endpoints.

Proof. At a dual vertex, the boundary degree counts the membership changes around a primal square, and is therefore even. The finite dual boundary consequently decomposes into edge-disjoint simple cycles: extract a simple cycle whenever an edge remains, and remove its edges, preserving even degrees.

Along the positive horizontal ray from 00, the total number of boundary crossings is odd, since membership in XX starts at one and eventually becomes zero. At least one cycle in the decomposition has an odd number of positive-ray crossings. The total number of crossings of a closed dual cycle with the whole horizontal line is even, by counting changes of the sign of the height. The selected cycle therefore also crosses the negative ray.

If its length is nn, a positive crossing has horizontal coordinate t+1/2t+1/2 for some integer 0≤t<n0 \le t < n: the horizontal distance to a negative crossing is less than nn. Define Cn\mathcal{C}_n to be all simple length-nn cycles having such an anchored crossing and an odd number of positive-ray crossings. There are at most nn choices for the anchored edge, two orientations, and at most 3n3^n nonbacktracking continuations. This proves the asserted overcount. Finally, greedily select crossed primal edges in a fixed order, discarding all edges incident to either endpoint of each selection. Each selection discards at most seven edges, including itself, because the primal degree is four. The selected edges have disjoint endpoints and number at least n/7n/7.

Lemma 3.2. For both models on Z3\mathbb{Z}^3, pc<1p_c < 1.

Proof. It suffices to percolate on a coordinate copy of Z2\mathbb{Z}^2. In the bond model, if the origin cluster there is finite, all primal edges crossed by its dual boundary are closed. Lemma 3.1 and a union bound give

Pp(0 has a finite planar cluster)≤∑n≥12n3n(1−p)n<1\mathbb{P}_p(0\text{ has a finite planar cluster}) \le\sum_{n\ge1} 2n3^n(1-p)^n < 1

for p<1p < 1 sufficiently close to one.

For sites, first separate the event that 00 is closed. If its open planar cluster is finite, each edge crossed by the boundary cycle has a closed endpoint. For a fixed candidate cycle, select k≥n/7k \ge n/7 edges with disjoint endpoints as in Lemma 3.1. The union over choices of one closed endpoint of each edge has probability at most [2(1−p)]k[2(1-p)]^k. For pp close to one, 2(1−p)<12(1-p)<1, so

Pp(0 is not in an infinite planar open cluster)≤(1−p)+∑n≥12n3n[2(1−p)]n/7<1.\mathbb{P}_p(0\text{ is not in an infinite planar open cluster}) \le(1-p)+\sum_{n\ge1}2n3^n[2(1-p)]^{n/7}<1.

Indeed both displayed series are convergent geometric-power series tending to zero as p↑1p \uparrow1. This proves the lemma.

Seed boxes and quarter-faces

For u∈Z3u \in\mathbb{Z}^3 and a nonnegative integer bb, write

Λb(u)=u+([−b,b]3∩Z3),Λb=Λb(0).\Lambda_b(u)=u+([-b,b]^3\cap\mathbb{Z}^3),\qquad\Lambda_b=\Lambda_b(0).

A quarter-face of Λb(u)\Lambda_b(u) is obtained by setting one coordinate relative to uu equal to bb or −b-b, and restricting each remaining coordinate to either [0,b][0,b] or [−b,0][-b,0]. There are 24 such quarter-faces; overlaps on their boundaries do not matter. More explicitly, for a normal index i∈{1,2,3}i\in\{1,2,3\}, a sign σ∈{−1,+1}\sigma\in\{-1,+1\}, and signs εj∈{−1,+1}\varepsilon_j\in\{-1,+1\} for the two indices j≠ij\ne i, set

Fi,σ(εj)j≠i(u,b)={u+x:xi=σb, 0≤εjxj≤b (j≠i)}.(8)F^{(\varepsilon_j)_{j\ne i}}_{i,\sigma}(u,b)=\{u+x:x_i=\sigma b,\ 0\leq\varepsilon_jx_j\leq b\ (j\ne i)\}. \tag*{(8)}

The six choices of (i,σ)(i,\sigma) and the four choices of the transverse sign pair enumerate all 24 faces. Figure 1 illustrates this indexing. A bit is internal to a vertex set if it is a site in that set, or, for bonds, an edge with both endpoints in that set.

Quarter-faces of $\Lambda_b(u)$, shown schematically for $b>0$

Figure 1. Quarter-faces of Λb(u)\Lambda_b(u), shown schematically for b>0b>0. Each of the three visible faces is divided into four quarter-faces; the three opposite faces are divided in the same way. Thus 3 normal axes ×\times 2 normal signs ×\times 4 transverse sign pairs give 24 quarter-faces. Boundary overlaps are permitted. The shaded example F1,++,+F^{+,+}_{1,+} has x1−u1=bx_1-u_1=b and 0≤x2−u2,x3−u3≤b0\leq x_2-u_2,x_3-u_3\leq b.

Suppose, towards a contradiction, that at p=pcp=p_c an infinite cluster exists with positive probability. Then 0<p<10<p<1 by Lemma 3.2 and the absence of percolation at zero. In fact an infinite cluster exists almost surely. To see this without a uniqueness assumption, the existence event is invariant under changing finitely many bits. Deleting finitely many edges, or finitely many vertices in a locally finite graph, splits an infinite connected component into only finitely many components, at least one of which is infinite. The event is measurable by countably many finite-path tests, and is determined by the bits outside any finite set. Thus it is independent of every finite cylinder event. Finite cylinder events approximate all events in this countable product space in probability, so the existence event is independent of itself and has probability zero or one. It follows that

Pp(Λm meets an infinite cluster)⟶1(m⟶∞).(9)\mathbb{P}_p(\Lambda_m\text{ meets an infinite cluster})\longrightarrow1 \qquad(m\longrightarrow\infty). \tag*{(9)}

We also use Harris’s positive-association inequality [9] for increasing events under a finite product of independent bits, and of decreasing events. Here is an elementary proof. For two increasing functions, condition on the last bit. Induction bounds the conditional covariance below by zero. The conditional means are increasing functions of this last bit, whose covariance is nonnegative. The covariance decomposition proves the assertion by induction on the number of bits. The same argument works for decreasing functions, and iteration gives the intersection bound for any finite family of events.

The use of symmetric quarter-face estimates follows the renormalization framework of Kozma and Nitzan [12], Section 4, Lemmas 8–9. The next lemma fixes the geometry before decreasing pp. Its tolerance ff is reserved for the planar contour sum, while γ\gamma bounds the probability of missing one quarter-face. The intermediate tolerances distribute errors in the extension and exploration arguments. Once all scales have been chosen at pp, finite-event continuity will supply one q<pq<p for the three required radii.

Lemma 3.3 (Choice of scales). Under the preceding contradiction assumption, there are numbers f,δ,η,τ,γ∈(0,1)f,\delta,\eta,\tau,\gamma\in(0,1), positive integers m,K,R,rm,K,R,r, and q∈(p/2,p)q \in(p/2,p) with the following properties. First, f<1/2f < 1/2 and

∑n≥12n3n(2f)n/7<1.\sum_{n\ge1} 2n3^n(2f)^{n/7} < 1.

The error parameters satisfy

δ+52ηδ<f,τ<η4,γ<δ,2γτ<η4.(10)\delta+ \frac{52\eta}{\delta} < f,\qquad\tau< \frac{\eta}{4},\qquad\gamma< \delta,\qquad\frac{2\gamma}{\tau} < \frac{\eta}{4}. \tag*{(10)}

Let MM be the number of internal bits of Λm\Lambda_m, and set ρ0=(p/2)M+1\rho_0 = (p/2)^{M+1}. With L=K(8m+5)3L = K(8m+5)^3 and α=(1−p)L\alpha= (1-p)^L, we have

4Kρ0<η4,R>m,1α(R−m)<η4,r>100(R+1).(11)\frac{4}{K\rho_0} < \frac{\eta}{4},\qquad R > m,\qquad\frac{1}{\alpha(R-m)} < \frac{\eta}{4},\qquad r > 100(R+1). \tag*{(11)}

For every ℓ∈{2r,2r+1,10r}\ell\in\{2r,2r+1,10r\}, u∈Z3u \in\mathbb{Z}^3, and quarter-face FF of Λℓ(u)\Lambda_\ell(u), independent parameter-qq percolation satisfies

Pq(Λm(u)↔F within Λℓ(u))>1−γ.(12)\mathbb{P}_q\bigl(\Lambda_m(u) \leftrightarrow F\text{ within }\Lambda_\ell(u)\bigr) > 1-\gamma. \tag*{(12)}

Proof. The series in (3.3) tends to zero as f↓0f \downarrow0, so choose ff first. Choose δ<f\delta< f, then η\eta sufficiently small for the first inequality of (10), then τ\tau, and finally γ\gamma, in that order. By (9), choose m≥1m \ge1 so that the probability that Λm\Lambda_m meets no infinite cluster at parameter pp is less than (γ/2)24(\gamma/2)^{24}.

For every integer ℓ>m\ell> m, any infinite cluster meeting Λm\Lambda_m gives an open path from Λm\Lambda_m to the vertex boundary of Λℓ\Lambda_\ell inside that cube. Thus the probability of missing the whole boundary is less than (γ/2)24(\gamma/2)^{24}. By signed coordinate symmetry, all 24 quarter-face miss events have the same probability, say aℓa_\ell. They are decreasing events of a finite product, so positive association gives

aℓ24≤Pp(Λm misses every quarter-face within Λℓ)<(γ/2)24.a_\ell^{24} \le\mathbb{P}_p(\Lambda_m\text{ misses every quarter-face within }\Lambda_\ell) < (\gamma/2)^{24}.

Consequently aℓ<γ/2a_\ell< \gamma/2 for every ℓ>m\ell> m.

Now MM and ρ0\rho_0 are fixed and positive. Choose KK, define LL and α>0\alpha> 0, and choose RR and then rr to satisfy (11). All these choices precede the choice of qq. At each of the three radii 2r,2r+1,10r2r,2r+1,10r, a quarter-face connection event uses finitely many bits, so its probability is a polynomial in the parameter. The strict estimate at pp, with margin γ/2\gamma/2, therefore persists with margin γ\gamma at some single q∈(p/2,p)q \in(p/2,p), simultaneously for all 24 faces and all three radii. Translation invariance gives the same estimates for every center uu. This proves (12). □\square

Fix these choices for the rest of the proof. Every later connection estimate uses parameter qq on the indicated unrevealed region. Bits outside that region may already have been fixed, a possibility allowed by Corollary 2.3.

Extending a connection through a fresh region

The construction adapts the shell, separated-seed, and conditional-target strategy of Kozma and Nitzan [12] [Section 4, Lemma 10]. We use the hyperedge failure comparison for both models and define relay quality using the interior bits alone.

We retain the parameters chosen in Lemma 3.3. A rectangle is a nonempty set of the form

B=∏h=13([ah,bh]∩Z),ah,bh∈Z,B = \prod_{h=1}^{3} \bigl([a_h,b_h]\cap\mathbb{Z}\bigr),\qquad a_h,b_h\in\mathbb{Z},

and B(j)B^{(j)} denotes its expansion by jj in every coordinate direction. Bits internal to a vertex set mean sites in that set, or bonds with both endpoints in it, according to the model.

Lemma 4.1 (Extension estimate). Let GG be a finite subgraph of Z3\mathbb{Z}^{3} containing a rectangle DD and every lattice edge internal to DD. Give its bits independent laws, with parameter qq on all bits internal to DD and arbitrary parameters, including 00 and 11, elsewhere. Let o∈V(G)o \in V(G) and T⊆V(G)T \subseteq V(G). Suppose that a rectangle BB satisfies

B(R)⊆D,o∉B(R),B^{(R)} \subseteq D,\qquad o \notin B^{(R)},

and that for every u∈B(R)u \in B^{(R)} there is a radius ℓ∈{2r,2r+1,10r}\ell\in\{2r,2r+1,10r\} such that Λℓ(u)⊆D\Lambda_{\ell}(u) \subseteq D and some quarter-face of this cube is contained in TT. Then, for connections in GG,

P(o↔B, o↮T)<η.(13)\mathbb{P}(o \leftrightarrow B,\ o \not\leftrightarrow T) < \eta. \tag*{(13)}

Proof. We first choose a deterministic layer at which a connection to BB usually supplies many possible entry points. We then try disjoint open seed boxes at those points, and use Corollary 2.3 to pass from a successful seed to TT.

Choosing a layer. The disjoint closure events used here also appear in Grimmett and Marstrand [7], Section 3, Lemma 3. For m≤j<Rm \leq j < R, expose only the bits internal to the induced graph outside B(j)B^{(j)}. Let YjY_j be the exterior neighbors of B(j)B^{(j)} reachable from oo by an open path in this outside graph, and set Nj=∣Yj∣N_j = \lvert Y_j\rvert. Every exterior neighbor yy of a rectangle has a unique neighbor zyz_y inside it. Both endpoints and their edge lie in B(j+1)⊆DB^{(j+1)} \subseteq D. Since oo is outside B(R)B^{(R)}, any path from oo to BB has a first entry into B(j)B^{(j)}, and hence Nj≥1N_j \geq1.

On 1≤Nj<L1 \leq N_j < L, consider the event that all bonds {y,zy}\{y,z_y\}, y∈Yjy \in Y_j, are closed in the bond model, or that all corresponding sites zyz_y are closed in the site model. Given the outside bits, these are unexposed parameter-qq bits. There are at most NjN_j distinct bits; repeated sites zyz_y can only increase the closure probability. Thus its conditional probability is at least

(1−q)Nj≥(1−p)L=α.(1-q)^{N_j} \geq(1-p)^L = \alpha.

Such a closure prevents every first entry into B(j)B^{(j)}. In particular, Ni=0N_i=0 for all i<ji<j, because every exterior neighbor of B(i)B^{(i)} belongs to B(j)B^{(j)}. The events consisting of 1≤Nj<L1 \leq N_j < L and this closure are therefore pairwise disjoint as jj varies. It follows that

∑j=mR−1P(o↔B, Nj<L)≤∑j=mR−1P(1≤Nj<L)≤α−1.\sum_{j=m}^{R-1}\mathbb{P}(o \leftrightarrow B,\ N_j < L) \leq \sum_{j=m}^{R-1}\mathbb{P}(1 \leq N_j < L) \leq\alpha^{-1}.

Fix a deterministic j∈{m,…,R−1}j \in\{m,\ldots,R-1\} for which

P(o↔B, Nj<L)≤1α(R−m).(14)\mathbb{P}(o \leftrightarrow B,\ N_j < L) \leq\frac{1}{\alpha(R-m)}. \tag*{(14)}

Write C=B(j)C=B^{(j)}. Let ξ\xi be the configuration of bits internal to CC, and let ζ\zeta be the configuration of bits internal to the induced graph outside CC. Thus YjY_j is determined by ζ\zeta. For bonds, the remaining bits are the bonds crossing between CC and its complement. The configurations ξ\xi, ζ\zeta, and these crossing bits are independent. All bits of ξ\xi and all crossing bits have parameter qq, since they lie in B(j+1)⊆DB^{(j+1)} \subseteq D. For sites, ξ\xi and ζ\zeta already partition the bits.

We will use these two configurations for different purposes. The exterior ζ\zeta selects entrances that oo has reached, whereas ξ\xi determines which relay centers have high conditional probability of reaching TT. We first estimate relay quality for each fixed entrance, then use independence to apply that estimate to the entrances selected by ζ\zeta.

Defining good relays from ξ\xi. For every exterior neighbor yy of CC, choose vyv_y so that

Uy=Λm(vy)⊆C,zy∈Uy,∥vy−y∥∞≤m+1.U_y = \Lambda_m(v_y) \subseteq C,\qquad z_y \in U_y,\qquad\lVert v_y-y\rVert_\infty\le m+1.

Explicitly, clamp each coordinate of zyz_y to the corresponding coordinate interval of CC shrunk by mm at both ends. Those shrunken intervals are nonempty because j≥mj \ge m. Let OyO_y be the event that every internal bit of UyU_y is open. Then OyO_y is determined by ξ\xi and P(Oy)=qM\mathbb{P}(O_y)=q^M. Define

gy=P(vy↔T∣ξ),Aξ={vy:y is an exterior neighbor of C, gy>1−τ}.g_y=\mathbb{P}(v_y \leftrightarrow T \mid\xi),\qquad A_\xi=\{v_y:y\text{ is an exterior neighbor of }C,\ g_y>1-\tau\}.

In this definition the exterior configuration and the crossing bits are still averaged over according to their original product law. In particular, AξA_\xi is determined by ξ\xi alone.

For a fixed yy, apply the geometric hypothesis at u=vyu=v_y. The quarter-face estimate (12) supplies an increasing event of probability greater than 1−γ1-\gamma connecting UyU_y to TT inside the specified cube. On its intersection with OyO_y, the center vyv_y also connects to TT. Positive association gives

P(Oy, vy↮T)≤qMγ.\mathbb{P}(O_y,\ v_y\not\leftrightarrow T)\le q^M\gamma.

Since OyO_y is determined by ξ\xi, it follows that

τP(Oy, gy≤1−τ)≤E[1Oy(1−gy)]=P(Oy, vy↮T)≤qMγ.(15)\tau\mathbb{P}(O_y,\ g_y\le1-\tau)\le\mathbb{E}\left[\mathbf{1}_{O_y}(1-g_y)\right] =\mathbb{P}(O_y,\ v_y\not\leftrightarrow T)\le q^M\gamma. \tag*{(15)}

This bounds the chance that a fixed seed is open but its center fails the defining condition for AξA_\xi.

Selecting reachable entrances using ζ\zeta. Condition now on ζ\zeta. Whenever Nj≥LN_j\ge L, choose deterministically KK members of YjY_j at pairwise sup distances greater than 4m+24m+2. This is possible by greedy selection: each chosen vertex excludes at most (8m+5)3(8m+5)^3 candidates, while L=K(8m+5)3L=K(8m+5)^3. The corresponding centers have pairwise distances greater than 2m2m, so their boxes UyU_y are vertex-disjoint.

Call a selected entrance successful if OyO_y holds and, for bonds, the crossing bond {y,zy}\{y,z_y\} is open. For sites, yy is already open because it is reachable in the outside graph, and the open seed contains zyz_y. In either model a success therefore gives an actual open path from oo to vyv_y.

Given ζ\zeta, these successes are independent Bernoulli trials with common parameter

c={qM+1,for bonds,qM,for sites.c≥ρ0.c= \begin{cases} q^{M+1}, & \text{for bonds},\\ q^M, & \text{for sites}. \end{cases} \qquad c\ge\rho_0.

Indeed, the boxes use disjoint sets of bits of ξ\xi, and the selected crossing bonds are distinct and independent of ξ\xi and ζ\zeta. If SS counts successes, its conditional variance is at most KcKc. Chebyshev’s inequality gives

P(S<Kc2 | ζ)≤4Kc≤4Kρ0.(16)\mathbb{P}\left(S<\frac{Kc}{2}\,\middle|\,\zeta\right)\le\frac{4}{Kc}\le\frac{4}{K\rho_0}. \tag*{(16)}

Let ZZ count successful entrances whose centers satisfy gy≤1−τg_y\le1-\tau. For each selected yy, (15) gives

P(Oy, gy≤1−τ∣ζ)≤qMγτ.\mathbb{P}(O_y,\ g_y\le1-\tau\mid\zeta)\le\frac{q^M\gamma}{\tau}.

Indeed, after fixing ζ\zeta the selected vertices are fixed, and Oy∩{gy≤1−τ}O_y\cap\{g_y\le1-\tau\} is an event of ξ\xi, whose law has not changed. For bonds the additional crossing bit is independent and contributes a factor qq. Summing these bounds gives

E[Z∣ζ]≤Kcγτ,P(Z≥Kc2 | ζ)≤2γτ.\mathbb{E}[Z\mid\zeta]\le\frac{Kc\gamma}{\tau},\qquad \mathbb{P}\left(Z\ge\frac{Kc}{2}\,\middle|\,\zeta\right)\le\frac{2\gamma}{\tau}.

If S≥Kc/2S \ge Kc/2 and Z<Kc/2Z < Kc/2, some successful entrance has its center in AξA_\xi. Thus oo reaches AξA_\xi. Integrating over ξ\xi and including the low-entry event from (14), we conclude that

P(o↔B,o↮Aξ)≤1α(R−m)+4Kρ0+2γτ.(17)\mathbb{P}(o \leftrightarrow B, o \not\leftrightarrow A_\xi) \le\frac{1}{\alpha(R-m)}+\frac{4}{K\rho_0}+\frac{2\gamma}{\tau}. \tag*{(17)}

Passing from the relay to the target. The estimate (17) is now under the original law. To control the remaining failure event, condition on ξ\xi alone. Then AξA_\xi is a deterministic set, the remaining bits have their product law, and the conditional target probability of each vy∈Aξv_y \in A_\xi is the number gy>1−τg_y>1-\tau used in its definition. Corollary 2.3 therefore gives

P(o↔Aξ,o↮T∣ξ)≤τ,\mathbb{P}(o \leftrightarrow A_\xi, o \not\leftrightarrow T \mid\xi) \le\tau,

including the trivial case Aξ=∅A_\xi=\varnothing. Integrating this bound, and splitting the desired failure event according to whether oo reaches AξA_\xi, yields

P(o↔B,o↮T)≤1α(R−m)+4Kρ0+2γτ+τ<η.\mathbb{P}(o \leftrightarrow B, o \not\leftrightarrow T) \le\frac{1}{\alpha(R-m)}+\frac{4}{K\rho_0}+\frac{2\gamma}{\tau}+\tau<\eta.

Each of the four terms is less than η/4\eta/4 by (3.4) and (3.5).

Relaying between neighboring cubes

The contraction and translation strategy adapts the quarter-face targets of Kozma and Nitzan [12], Section 4, Lemmas 11–12. We specify thirteen moves that turn the extension estimate into a connection estimate between the present neighboring coarse boxes. The scales R,rR,r and the parameter qq are those of Lemma 3.3; in particular,

r>100(R+1).r>100(R+1).

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

cv=(20r+1)(v1,v2,0),Qv=Λ10r(cv),Bv=Λ4r(cv).c_v=(20r+1)(v_1,v_2,0),\qquad Q_v=\Lambda_{10r}(c_v),\qquad B_v=\Lambda_{4r}(c_v).

The cubes QvQ_v have disjoint vertex sets. If v,xv,x are nearest neighbors in Z2\mathbb{Z}^2, let IvxI_{vx} be the set of lattice edges with one endpoint in QvQ_v and the other in QxQ_x. Their union D=Qv∪QxD=Q_v\cup Q_x is a rectangular lattice prism: all edges internal to DD are precisely the edges internal to the two cubes together with IvxI_{vx}.

Lemma 5.1 (Neighboring-box relay). Let v,x∈Z2v,x\in\mathbb{Z}^2 be nearest neighbors. Let GG be a finite lattice subgraph containing D=Qv∪QxD=Q_v\cup Q_x and all lattice edges internal to DD. Consider independent bond or site percolation on GG, with parameter qq on all bits internal to DD and arbitrary parameters elsewhere. If o∈V(G)o\in V(G) lies outside DD, then

P(o↔Bv,o↮Bx)≤13η.(18)\mathbb{P}(o\leftrightarrow B_v, o\not\leftrightarrow B_x)\le13\eta. \tag*{(18)}

All connections in this statement are taken in GG.

Proof. By translation and a signed permutation of coordinates, we may suppose that cv=0c_v=0 and cx=(20r+1,0,0)c_x=(20r+1,0,0). Thus

D=([−10r,30r+1]×[−10r,10r]2)∩Z3.(19)D=\bigl([-10r,30r+1]\times[-10r,10r]^2\bigr)\cap\mathbb{Z}^3. \tag*{(19)}

Indeed, the first-coordinate intervals of the two cubes are [−10r,10r]∩Z[-10r,10r] \cap\mathbb{Z} and [10r+1,30r+1]∩Z[10r+1,30r+1] \cap\mathbb{Z}, whose union has no missing lattice vertices.

We construct rectangles P0,…,P13P_0,\ldots,P_{13}, starting with P0=BvP_0=B_v and ending with P13⊆BxP_{13} \subseteq B_x, such that Lemma 4.1 applies with B=PiB=P_i and T=Pi+1T=P_{i+1} for every 0≤i<130 \leq i < 13. At each translation, the required RR-expansion of the starting rectangle is accommodated by increasing every half-width by RR. We first shorten the rectangle in each of its three coordinate directions, leaving room for ten such increases before the final rectangle must fit inside BxB_x. Figure 2 shows these two stages, and Table 1 records the thirteen target rectangles and the radii and orientations used to choose the quarter-faces.

A schematic cross-section of the neighboring-box relay

Figure 2. A schematic cross-section of the neighboring-box relay, not to scale. The three contraction steps take P0P_0 to P3P_3; the marked centers then make ten translations to P13⊂BxP_{13} \subset B_x. Intermediate rectangles are omitted. The last translation has length 2r+12r+1, rather than 2r2r, to match the coarse spacing 20r+120r+1. The narrow gap represents the interface edges between disjoint lattice cubes, not a missing layer of lattice vertices.

TargetCenter’s first coordinateCoordinate half-widthsRadiusNormal
P1P_100(2r+R,4r+R,4r+R)(2r+R,4r+R,4r+R)2r2r11, inward
P2P_200(2r+2R,2r+2R,4r+2R)(2r+2R,2r+2R,4r+2R)2r2r22, inward
P3P_300(2r+3R,2r+3R,2r+3R)(2r+3R,2r+3R,2r+3R)2r2r33, inward
P4P_42r2r(2r+4R)1(2r+4R)\mathbf{1}2r2r11, forward
P5P_54r4r(2r+5R)1(2r+5R)\mathbf{1}2r2r11, forward
P6P_66r6r(2r+6R)1(2r+6R)\mathbf{1}2r2r11, forward
P7P_78r8r(2r+7R)1(2r+7R)\mathbf{1}2r2r11, forward
P8P_810r10r(2r+8R)1(2r+8R)\mathbf{1}2r2r11, forward
P9P_912r12r(2r+9R)1(2r+9R)\mathbf{1}2r2r11, forward
P10P_{10}14r14r(2r+10R)1(2r+10R)\mathbf{1}2r2r11, forward
P11P_{11}16r16r(2r+11R)1(2r+11R)\mathbf{1}2r2r11, forward
P12P_{12}18r18r(2r+12R)1(2r+12R)\mathbf{1}2r2r11, forward
P13P_{13}20r+120r+1(2r+13R)1(2r+13R)\mathbf{1}2r+12r+111, forward

Table 1. The thirteen interfaces Pj−1→PjP_{j-1}\to P_j in the relay. All centers have second and third coordinates zero, and 1=(1,1,1)\mathbf{1}=(1,1,1). Each listed radius is used about every point of Pj−1(R)P_{j-1}^{(R)}. “Inward” means toward zero in that normal coordinate; “forward” means the positive first direction. All transverse signs point toward zero. The proof verifies both target-face and full-cube containment.

Three contraction steps. For t=1,2,3t=1,2,3, let PtP_t be centered at zero with coordinate half-widths

at,h={2r+tR,h≤t,4r+tR,h>t,h∈{1,2,3}.(20)a_{t,h} = \begin{cases} 2r+tR, & h \leq t,\\ 4r+tR, & h > t, \end{cases} \qquad h \in\{1,2,3\}. \tag*{(20)}

Fix u∈Pt−1(R)u \in P_{t-1}^{(R)}. Its coordinate bounds are 2r+tR2r+tR in directions h<th<t and 4r+tR4r+tR in directions h≥th \geq t. In the cube Λ2r(u)\Lambda_{2r}(u), choose a quarter-face normal to coordinate tt, with its normal sign and both transverse signs directed toward coordinate zero. At a zero coordinate, either sign may be used.

In the normal direction, the face coordinate has absolute value at most

max⁡{2r,(4r+tR)−2r}=2r+tR.\max\{2r,(4r+tR)-2r\}=2r+tR.

In a transverse direction, the face coordinate ranges between a number aa and a−2rsign⁡(a)a-2r\operatorname{sign}(a), with either sign allowed when a=0a=0. Its absolute value is therefore at most max⁡{∣a∣,2r}\max\{|a|,2r\}, which fits the corresponding half-width in (20). The chosen quarter-face is consequently contained in PtP_t.

The full cube Λ2r(u)\Lambda_{2r}(u) has all coordinate absolute values at most 6r+3R<10r6r+3R<10r, by (5.1), and so lies in DD. In particular Pt−1(R)⊆DP_{t-1}^{(R)} \subseteq D. Since o∉Do \notin D, the remaining geometric hypothesis of Lemma 4.1 is also satisfied. This proves its applicability to all three contraction steps. The resulting rectangle is

P3=Λ2r+3R(0).P_3=\Lambda_{2r+3R}(0).

Ten translation steps. Define the step lengths and centers by

si={2r,1≤i≤9,2r+1,i=10,z0=0,zi=(s1+⋯+si,0,0).s_i= \begin{cases} 2r, & 1\le i\le9,\\ 2r+1, & i=10, \end{cases} \qquad z_0=0,\qquad z_i=(s_1+\cdots+s_i,0,0).

For 1≤i≤101\le i\le10, set

wi=2r+(3+i)R,P3+i=Λwi(zi).w_i=2r+(3+i)R,\qquad P_{3+i}=\Lambda_{w_i}(z_i).

The RR-expansion of the preceding rectangle, P2+i(R)P_{2+i}^{(R)}, is centered at zi−1z_{i-1} and has half-width wiw_i. Fix uu in that expansion. In Λsi(u)\Lambda_{s_i}(u) choose the quarter-face with forward normal in coordinate 1, and with both transverse signs directed toward zero. The normal coordinate on this face is u1+siu_1+s_i. Relative to the new center ziz_i, its offset is exactly u1−(zi−1)1u_1-(z_{i-1})_1, of absolute value at most wiw_i. Transverse absolute coordinates are at most max⁡{wi,si}=wi\max\{w_i,s_i\}=w_i. The equality also holds for the last step, because 13R≥113R\ge1. Hence this quarter-face is contained in P3+iP_{3+i}.

To check containment of the full radius-sis_i cube, note that 0≤(zi−1)1≤18r0\le(z_{i-1})_1\le18r, wi≤2r+13Rw_i\le2r+13R, and si≤2r+1s_i\le2r+1. Its first coordinate is therefore between

−4r−13R−1and22r+13R+1,(21)-4r-13R-1\quad\text{and}\quad22r+13R+1, \tag*{(21)}

and its transverse absolute coordinates are at most 4r+13R+14r+13R+1. Condition (5.1) gives 13R+1<6r13R+1<6r and 13R<8r13R<8r, so these bounds lie inside the prism (5.4). It follows again that the preceding expanded rectangle lies in DD and avoids oo. The radii used are 2r2r and 2r+12r+1, both allowed by Lemma 4.1. Thus all ten translation steps satisfy that Lemma’s hypotheses.

The final center is z10=(9(2r)+(2r+1),0,0)=cxz_{10}=(9(2r)+(2r+1),0,0)=c_x, and the final half-width is w10=2r+13R<4rw_{10}=2r+13R<4r. Consequently P13⊆BxP_{13}\subseteq B_x. If oo is connected to P0P_0 but not to P13P_{13}, then for at least one i∈{0,…,12}i\in\{0,\ldots,12\} it is connected to PiP_i but not to Pi+1P_{i+1}. A union bound and thirteen applications of Lemma 4.1 now give

P(o↔Bv, o↮Bx)≤P(o↔P0, o↮P13)≤∑i=012P(o↔Pi, o↮Pi+1)≤13η.\begin{aligned} \mathbb{P}(o\leftrightarrow B_v,\ o\not\leftrightarrow B_x) \le\mathbb{P}(o\leftrightarrow P_0,\ o\not\leftrightarrow P_{13}) \\ &\le\sum_{i=0}^{12}\mathbb{P}(o\leftrightarrow P_i,\ o\not\leftrightarrow P_{i+1})\le13\eta. \end{aligned}

No independence between these thirteen connection events is required.

Adaptive exploration and the final contradiction

We use the scales from Lemma 3.3 to construct an infinite open cluster at the parameter q<pcq < p_c. Recall the coarse cubes and their inner boxes:

cv=(20r+1)(v1,v2,0),Qv=Λ10r(cv),Bv=Λ4r(cv),v∈Z2.c_v = (20r + 1)(v_1,v_2,0), \qquad Q_v = \Lambda_{10r}(c_v), \qquad B_v = \Lambda_{4r}(c_v), \qquad v \in\mathbb{Z}^2.

For adjacent coarse vertices v,xv,x, the interface IvxI_{vx} consists of the lattice edges between QvQ_v and QxQ_x. Different cubes have disjoint vertex sets. We keep the source o=(0,0,0)o = (0,0,0) in the root cube Q0Q_0.

The exploration and its conditional promise

Retaining conditional connection predictions during adaptive exploration is the method used in Kozma and Nitzan’s proof of Theorem 6 [12], Section 4, equations (30)–(32). Here each pending cube’s unknown bits are reserved until that cube is processed.

Let I\mathcal{I} be the event that all bits internal to Q0Q_0 are open and, for bonds, that all bits on its four interfaces are open. This event has positive probability because it prescribes finitely many bits and q>0q > 0. Write

PI( ⋅ )=Pq( ⋅ ∣I).\mathbb{P}_{\mathcal{I}}(\,\cdot\,) = \mathbb{P}_q(\,\cdot\, \mid\mathcal{I}).

Under this law, all bits not prescribed by I\mathcal{I} are still independent Bernoulli(q)(q).

Declare the coarse root good and put its four neighbors in a queue, each with parent the root. A coarse vertex is visited once it has been put in the queue; the root is also visited. A queued vertex is pending until it is processed. Use a fixed ordering of neighbors whenever vertices are added to the queue, and always process the first pending vertex. No vertex is assigned a parent or queued more than once.

For each pending vertex we will retain a high conditional probability of connection from oo. The graph used for this prediction follows its parent chain, so that the only bits still needed to decide the connection will be in the pending cube and its incoming interface. Specifically, for a pending vertex vv, let G(v)G(v) consist of the cubes on its parent chain back to the root, all edges internal to those cubes, and the interfaces between successive cubes of the chain. We include no other edges, even when nonsuccessive cubes are adjacent. Set

Hv={o↔Bv in G(v)}.H_v = \{o \leftrightarrow B_v \text{ in } G(v)\}.

The construction will maintain the promise

PI(Hv∣revealed history)>1−δfor every pending v.(22)\mathbb{P}_{\mathcal{I}}(H_v \mid\text{revealed history}) > 1-\delta\quad\text{for every pending } v. \tag*{(22)}

A cube will acquire children only after its own connection is verified and those children satisfy the same promise.

The invariant holds initially. For a root neighbor vv, choose a quarter-face of QvQ_v facing Q0Q_0. Every vertex of that quarter-face has a neighbor in Q0Q_0. By (12), the probability that Λm(cv)\Lambda_m(c_v) connects to this quarter-face inside QvQ_v is greater than 1−γ1-\gamma. Here Λm(cv)⊆Bv\Lambda_m(c_v) \subseteq B_v. Such a connection joins BvB_v to oo: the root cube and the intervening edge are open in the bond model, while both endpoints of the intervening edge are open in the site model. Since γ<δ\gamma< \delta, this proves (22) for the initial queue.

To process a pending vertex vv, reveal the bits internal to QvQ_v and, for bonds, its incoming parent interface, except for bits already known. The ancestor cubes and the interfaces joining them have been revealed at earlier steps, so this determines HvH_v. For each unvisited neighbor xx of vv, extend G(v)G(v) by QxQ_x, its internal edges, and IvxI_{vx}, obtaining a graph G′G', and compute

bx=PI(o↔Bx in G′∣history after this reveal).(23)b_x = \mathbb{P}_{\mathcal{I}}(o \leftrightarrow B_x \text{ in } G' \mid\text{history after this reveal}). \tag*{(23)}

The number bxb_x is calculated from the revealed values and the product law of the remaining bits; its computation reveals no bits. Declare vv good if HvH_v holds and bx>1−δb_x > 1-\delta for every such neighbor. In this case, give each unvisited neighbor parent vv and add it to the back of the queue. Otherwise declare vv bad and add no vertices. Remove vv from the queue in either case, and continue until the queue is empty, if this ever occurs.

Lemma 6.1 (Freshness and persistence). Conditional on any positive-probability finite revealed history, all unrevealed bits are independent Bernoulli(q)(q). Moreover, (6.1) holds at every step of the exploration.

Proof. The initialization fixes finitely many bits and leaves the others independent with parameter qq. At every later step, the preceding revealed values determine which finite set of bits is exposed. Conditioning on the values of this selected set again leaves the remaining bits with their product law. The classifications, queue, parent assignments, and numbers bxb_x are all functions of revealed values. They therefore add no conditions on the remaining bits. Induction proves the first assertion.

When xx is assigned parent vv, its graph G(x)G(x) is precisely the graph G′G' used to compute bxb_x. The rule for declaring vv good therefore establishes (6.1) for xx at that time. To prove persistence, it suffices to verify that processing other vertices cannot expose any of the unknown bits of HxH_x.

All ancestor cubes and their connecting interfaces have already been revealed. Thus the unknown bits of HxH_x lie in QxQ_x or, for bonds, on IvxI_{vx}. The cubes are vertex-disjoint, so another vertex’s processing does not reveal an internal bit of QxQ_x. Each step also reveals only the processed vertex’s incoming parent interface. Distinct parent links use distinct interfaces. Moreover, a pending vertex has no children, since only processed good vertices acquire them. Hence IvxI_{vx} cannot be revealed by another step while xx is pending. If this interface was prescribed at initialization because vv is the root, its fixed values simply remain part of the history.

Consequently, until xx is processed, the known bits of HxH_x stay fixed and its unknown bits retain the same product law. Its conditional connection probability is unchanged. Together with the initial verification, this proves (6.1) at every step.

A uniform bound on bad vertices

Lemma 6.2. Immediately before any pending vertex is processed, its conditional probability of being declared bad is less than ff.

Proof. Condition on a positive-probability history immediately before processing vv. Let F\mathcal{F} denote this history and F+\mathcal{F}^{+} the history after the prescribed reveal at vv. The unvisited neighbors to be tested are already determined by F\mathcal{F}.

Fix one such neighbor xx and the corresponding graph G′G'. Under the conditioning on F\mathcal{F}, every bit internal to the prism D=Qv∪QxD=Q_v\cup Q_x has its fresh parameter-qq law. To check this carefully, neither cube has been processed: vv is pending and xx is unvisited. Their interface has not been revealed as a parent link, since that would require xx to have been visited or vv to have been processed. Neither vertex is the root, so their interface is not one prescribed by initialization. Although an initialized root interface may meet QvQ_v, it is external to DD. Lemma 6.1 gives the required independence.

The graph G′G' contains every edge internal to this prism, and oo lies outside it in the disjoint root cube. All exterior bits have independent laws, with probabilities 0 or 1 for known bits and qq for unknown bits. Thus Lemma 5.1 applies. If Cx={o↔Bx in G′}C_x=\{o\leftrightarrow B_x\text{ in }G'\}, it gives

PI(Hv∩Cxc∣F)≤13η,\mathbb{P}_{\mathcal{I}}(H_v\cap C_x^{\mathrm{c}}\mid\mathcal{F})\le13\eta,

because HvH_v implies a connection to BvB_v in the larger graph G′G'. Notice that the relay estimate is applied before revealing QvQ_v, not after conditioning on HvH_v. The event HvH_v is determined by F+\mathcal{F}^{+}, and bx=PI(Cx∣F+)b_x = \mathbb{P}_{\mathcal{I}}(C_x \mid\mathcal{F}^{+}). The tower property therefore yields

EI[1Hv(1−bx)∣F]≤13η.\mathbb{E}_{\mathcal{I}}[1_{H_v}(1-b_x) \mid\mathcal{F}] \le13\eta.

On Hv∩{bx≤1−δ}H_v \cap\{b_x \le1-\delta\} the integrand is at least δ\delta, so

PI(Hv∩{bx≤1−δ}∣F)≤13ηδ.\mathbb{P}_{\mathcal{I}}(H_v \cap\{b_x \le1-\delta\} \mid\mathcal{F}) \le\frac{13\eta}{\delta}.

There are at most four tests. The promise and a union bound now give

PI(v is declared bad∣F)≤δ+52ηδ<f,\mathbb{P}_{\mathcal{I}}(v\text{ is declared bad} \mid\mathcal{F}) \le\delta+ \frac{52\eta}{\delta} < f,

where the last inequality is (10). No independence between the tests is needed.

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

Proof. The assertion is immediate if the set contains the root, which is declared good. Otherwise, fix such a set SS. For 1≤j≤k1 \le j \le k, let TjT_j be the processing time of the jjth member of SS to be processed, and set Tj=∞T_j = \infty if fewer than jj members are ever processed. Let EjE_j be the event that Tj<∞T_j < \infty and these first jj members are all declared bad; put E0E_0 equal to the whole sample space.

For each finite time tt, the event

{Tj=t}∩Ej−1\{T_j=t\}\cap E_{j-1}

is determined before the reveal at time tt. Indeed, at that moment one knows the next vertex, how many members of SS have already been processed, and whether all of those members were bad. On {Tj=t}\{T_j=t\} there are exactly j−1j-1 such previous members, so these observations determine Ej−1E_{j-1} there. After termination we keep the history constant; events specifying a later processing time are empty. Applying Lemma 6.2 at time tt and summing gives

PI(Ej)≤f∑t≥1PI(Tj=t, Ej−1)≤fPI(Ej−1).\begin{aligned} \mathbb{P}_{\mathcal{I}}(E_j) \le f\sum_{t\ge1}\mathbb{P}_{\mathcal{I}}(T_j=t,\ E_{j-1}) \\ &\le f\mathbb{P}_{\mathcal{I}}(E_{j-1}). \end{aligned}

Iteration yields PI(Ek)≤fk\mathbb{P}_{\mathcal{I}}(E_k) \le f^k. Since no vertex is processed twice, EkE_k is exactly the event in the statement. The proof does not assume that every member of SS is eventually processed.

Contours and completion of the proof

Let XX be the set of good coarse vertices, including the root. If XX is finite, only finitely many vertices can ever enter the queue: new entries arise only from good vertices, with at most four per vertex. The queue must therefore empty after finitely many steps. Every vertex outside XX adjacent to XX has then been processed and declared bad. Indeed, the root’s neighbors were all queued at initialization. When any nonroot good vertex was processed, each of its neighbors was either already visited or assigned at that step. Thus every outside neighbor of XX is processed before termination and is not good, so it is bad.

Apply Lemma 3.1 to this finite set XX, which contains the origin of the coarse lattice. Its boundary contains a cycle of some length nn in the deterministic family of at most 2n3n2n3^n candidates. For each candidate, fix the matching of crossed primal edges supplied by that Lemma. It has k≥n/7k \ge n/7 edges with pairwise disjoint endpoints. If the candidate occurs as a boundary cycle of XX, every one of these edges has a bad endpoint outside XX.

For a fixed candidate there are 2k2^k ways to choose one endpoint from each matching edge. Each choice is a deterministic set of kk distinct vertices, so Lemma 6.3 bounds the probability that all chosen vertices are processed bad by fkf^k. Hence the candidate’s probability is at most

2kfk=(2f)k≤(2f)n/7.2^k f^k=(2f)^k\le(2f)^{n/7}.

using 2f<12f<1. We have not conditioned on the random set XX or on the occurrence of a cycle. Taking the union bound over the deterministic candidate families and using the choice of ff in Lemma 3.3 gives

PI(∣X∣<∞)≤∑n≥12n3n(2f)n/7<1.(24)\mathbb{P}_{\mathcal I}(|X|<\infty)\le\sum_{n\ge1}2n3^n(2f)^{n/7}<1. \tag*{(24)}

Thus, with positive probability under PI\mathbb{P}_{\mathcal I}, infinitely many coarse vertices are good. Every good nonroot vertex vv satisfies the actual connection event HvH_v. The boxes BvB_v are pairwise disjoint, so these connections put infinitely many lattice vertices in the open cluster of oo. Therefore

Pq(o belongs to an infinite open cluster)≥Pq(I)PI(∣X∣=∞)>0.\mathbb{P}_q(o\text{ belongs to an infinite open cluster})\ge\mathbb{P}_q(\mathcal I)\mathbb{P}_{\mathcal I}(|X|=\infty)>0.

This contradicts q<pcq<p_c. The assumption of critical percolation was therefore false. The construction and all its estimates apply to both bond and site percolation with their respective critical parameters, proving Theorem 1.1. The finite computations are not formal proof certificates.

References

  1. [1]Itai Benjamini and Oded Schramm. Percolation beyond Z^d, many questions and a few answers. Electronic Communications in Probability 1 (1996), 71–82. doi:10.1214/ECP.v1-978.DOI
  2. [2]S. R. Broadbent and J. M. Hammersley. Percolation processes. I. Crystals and mazes. Proceedings of the Cambridge Philosophical Society 53 (1957), no. 3, 629–641. doi:10.1017/S0305004100032680.
  3. [3]ChatGPT 5.6, prompted by Ahmed Bou-Rabee; exposition and formalization support by Claude Opus 5 and GLM 5.3. Percolation at criticality. Public research exposition, updated September 5, 2026. Public exposition.nitromannitol.github.io/site-percolation-7d3f2
  4. [4]ChatGPT 5.6 Sol and Claude Fable 5.1, prompted by Ahmed Bou-Rabee. Kozma–Nitzan inequalities: proofs and counterexamples. Public research exposition, updated September 5, 2026. Public exposition.nitromannitol.github.io/kn1-verification-b80e9
  5. [5]Hugo Duminil-Copin, Vladas Sidoravicius, and Vincent Tassion. Absence of infinite cluster for critical Bernoulli percolation on slabs. Communications on Pure and Applied Mathematics 69 (2016), no. 7, 1397–1411. doi:10.1002/cpa.21641.
  6. [6]Robert Fitzner and Remco van der Hofstad. Mean-field behavior for nearest-neighbor percolation in d > 10. Electronic Journal of Probability 22 (2017), paper no. 43, 1–65. doi:10.1214/17-EJP56.DOI
  7. [7]G. R. Grimmett and J. M. Marstrand. The supercritical phase of percolation is well behaved. Proceedings of the Royal Society of London. Series A 430 (1990), no. 1879, 439–457. doi:10.1098/rspa.1990.0100.DOI
  8. [8]Takashi Hara and Gordon Slade. Mean-field critical behaviour for percolation in high dimensions. Communications in Mathematical Physics 128 (1990), no. 2, 333–391. doi:10.1007/BF02108785.DOI
  9. [9]Theodore E. Harris. A lower bound for the critical probability in a certain percolation process. Proceedings of the Cambridge Philosophical Society 56 (1960), no. 1, 13–20. doi:10.1017/S0305004100034241.DOI
  10. [10]Markus Heydenreich and Kilian Matzke. Critical site percolation in high dimension. Journal of Statistical Physics 181 (2020), 816–853. doi:10.1007/s10955-020-02607-y.DOI
  11. [11]Harry Kesten. The critical probability of bond percolation on the square lattice equals 1/2. Communications in Mathematical Physics 74 (1980), 41–59. doi:10.1007/BF01197577.DOI
  12. [12]Gady Kozma and Shahaf Nitzan. A reduction of the θ(p_c) = 0 problem to a conjectured inequality. Preprint, 2024. arXiv:2401.12397v1.arxiv.org/abs/2401.12397v1
  13. [13]Justin Leder. θ(p_c) = 0 for Bernoulli bond percolation on Z^d in all dimensions d ≥ 2: a guide to the Lean formalization. Public research announcement, August 2026. Pinned repository version, commit 795efb86f191.github.com/anthropics/formal-math/blob/795efb86f191735c5481675763537cfb4ff37e55/percolation/summary.pdf
  14. [14]Lucio Russo. On the critical percolation probabilities. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete 56 (1981), no. 2, 229–237. doi:10.1007/BF00535742.DOI
  15. [15]J. van den Berg, O. Häggström, and J. Kahn. Some conditional correlation inequalities for percolation and related processes. Random Structures & Algorithms 29 (2006), no. 4, 417–435. doi:10.1002/rsa.20102.

Paper details

vlnc:W_2ey7ynex
Version
v1 · latest
Published
License
LicenseRef-OpenAI-public
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

unreleased internal OpenAI model

View statement
Looking for related papers…
Contents