OctoInformBeta

OpenAI preprint, September 2026 · §4. An extension estimate

Lemma 4.1: An upper bound for a source reaching a box without reaching a target

An extension estimate for missed targets.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

The event of interest is that a source connects to an inner rectangle while having no connection to a prescribed target. The surrounding region has a constant opening density, although opening probabilities elsewhere may vary.

The cubic lattice ℤ3\mathbb{Z}^3 consists of integer triples.✦ Write eie_i for the iith coordinate unit vector, so that the positive coordinate step from xx is x+eix+e_i.✦ Two vertices are nearest neighbors when their difference is eie_i or −ei-e_i for some ii.✦ A lattice bond (x,i)(x,i) is the edge with endpoints xx and x+eix+e_i✦; incidence with two vertices means that they are these endpoints, in either order.✦

A finite lattice subgraph GG has finite vertex and bond sets V(G)V(G) and E(G)E(G), with both endpoints of every bond in V(G)V(G).✦ The underlying graph of a finite lattice subgraph GG is the graph with vertex set V(G)V(G) in which two vertices are adjacent when they are incident with a bond in E(G)E(G).✦ A graph GG contains a region D⊆ℤ3D\subseteq\mathbb{Z}^3, including a lattice box, when it contains every vertex of DD and every lattice bond with both endpoints in DD.✦

Fix a model MM, either bond or site percolation.✦ The variables of a percolation model MM are lattice bonds in the bond model and lattice vertices in the site model.✦ A configuration ω\omega of the model MM is an assignment of an open or closed state to each variable of MM.✦ For an assignment π\pi of opening probabilities in [0,1][0,1], let ℙπ\mathbb{P}_{\pi} be the independent Bernoulli product measure with opening probability π(e)\pi(e) at variable ee; write ℙq\mathbb{P}_q when π\pi is constantly qq.✦ An internal variable ee of a region DD is a lattice bond with both endpoints in DD in the bond model, or a lattice site belonging to DD in the site model.✦ The assignment π\pi has uniform density qq on DD when π(e)=q\pi(e)=q for every internal variable of DD.✦

Open bond adjacency means incidence with an open bond.✦ On a graph, open site adjacency means adjacency with both vertices open.✦ Open lattice adjacency uses these relations on the nearest-neighbor lattice✦; within GG, bond adjacency uses only bonds of E(G)E(G) and site adjacency uses its underlying graph. The notation x↔Gyx\leftrightarrow_G y means that a finite open path within GG joins the two vertices.✦ A bond path of length zero connects a vertex to itself; a site path, including one of length zero, requires all its vertices to be open. For o∈V(G)o\in V(G) and T⊆ℤ3T\subseteq\mathbb{Z}^3, the target connection event is {o↔GT}={ω:some y∈T∩V(G) satisfies o↔Gy}.\{o\leftrightarrow_G T\} =\{\omega:\text{some }y\in T\cap V(G)\text{ satisfies } o\leftrightarrow_G y\}.(1)✦

Seeds and Scales

A lattice box BB is a rectangle with integer lower and upper corners l,ul,u satisfying li≤uil_i\le u_i in each coordinate.✦ The vertices of a lattice box BB with lower and upper corners l,ul,u are all lattice points between ll and uu, and the bonds of BB are all lattice bonds with both endpoints among these vertices. For nonnegative integers j,nj,n, the expansion B(j)B^{(j)} and the cube Λn(c)\Lambda_n(c) of coordinate radius nn centered at cc are the boxes in (2). B={x∈ℤ3:li≤xi≤ui for every i},B(j)={x∈ℤ3:li−j≤xi≤ui+j for every i},Λn(c)={x∈ℤ3:ci−n≤xi≤ci+n for every i}.\begin{aligned} B&=\{x\in\mathbb{Z}^3:l_i\le x_i\le u_i\text{ for every }i\},\\ B^{(j)}&=\{x\in\mathbb{Z}^3:l_i-j\le x_i\le u_i+j\text{ for every }i\},\\ \Lambda_n(c)&=\{x\in\mathbb{Z}^3:c_i-n\le x_i\le c_i+n \text{ for every }i\}. \end{aligned}(2)

For a nonnegative integer aa, the seed variable count NM(a)N_M(a) is the number of internal variables of Λa(0)\Lambda_a(0).✦ For a model MM and a nonnegative integer aa, the seed success exponent κM(a)\kappa_M(a) is the integer specified in (3), where ℐM(A)\mathcal{I}_M(A) denotes the internal variables of a region AA. NM(a)=|ℐM(Λa(0))|,κM(a)={NM(a)+1,M is the bond model,NM(a),M is the site model.N_M(a)=|\mathcal{I}_M(\Lambda_a(0))|, \qquad \kappa_M(a)= \begin{cases} N_M(a)+1,&M\text{ is the bond model},\\ N_M(a),&M\text{ is the site model}. \end{cases}(3)

An available vertex xx of the cubic lattice is a vertex that is open in the site model, with every lattice vertex available in the bond model. Adjacency restricted to a region requires both endpoints to belong to that region.✦ For a nonnegative integer bb, the confined seed connection event EM(c;a,b,T)E_M(c;a,b,T) consists of configurations with an available vertex x∈Λa(c)x\in\Lambda_a(c) and a vertex y∈Ty\in T joined by a finite path of open lattice adjacencies restricted to Λb(c)\Lambda_b(c), with a path of length zero allowed.✦

A quarter-face index f=(i,ϵ)f=(i,\epsilon) consists of a coordinate direction ii and signs ϵ∈{−1,1}3\epsilon\in\{-1,1\}^3.✦ The quarter-face F(c,n;f)F(c,n;f) of Λn(c)\Lambda_n(c) is the set in (4). There are 2424 such signed-coordinate quarter-faces, and their boundary overlaps are permitted. F(c,n;i,ϵ)={x∈ℤ3:xi=ci+ϵin,0≤ϵj(xj−cj)≤n for j≠i}.F(c,n;i,\epsilon)= \{x\in\mathbb{Z}^3:x_i=c_i+\epsilon_i n,\quad 0\le\epsilon_j(x_j-c_j)\le n\text{ for }j\ne i\}.(4)✦

A signed lattice symmetry SS of ℤ3\mathbb{Z}^3 is a map determined by a coordinate permutation σ\sigma, a choice of signs ϵ∈{−1,1}3\epsilon\in\{-1,1\}^3, and a translation t∈ℤ3t\in\mathbb{Z}^3, acting as in (5).✦ S(x)i=ti+ϵixσ(i).S(x)_i=t_i+\epsilon_i x_{\sigma(i)}.(5)✦

Renormalization scale data for MM consist of densities p,qp,q, error allowances δ,η,τ,γ\delta,\eta,\tau,\gamma, and nonnegative integers m,K,R,rm,K,R,r satisfying (6) and (7).✦ Here mm is the seed radius, KK the trial count, RR the intermediate radius, and rr the large radius. The contour error tolerance is ε=12(1100)7\varepsilon=\frac12(\frac1{100})^7.✦ 0<p<1,p/2<q<p,δ,η,τ,γ∈(0,1),0<m<R,K>0,100(R+1)<r,δ+52ηδ<ε,γ<δ,τ<η4,2γτ<η4,4K(p/2)NM(m)+1<η4,1(1−p)K(8m+5)3(R−m)<η4.\begin{gathered} 0<p<1,\qquad p/2<q<p,\qquad \delta,\eta,\tau,\gamma\in(0,1),\\ 0<m<R,\qquad K>0,\qquad 100(R+1)<r,\\ \delta+\frac{52\eta}{\delta}<\varepsilon,\qquad \gamma<\delta,\qquad \tau<\frac{\eta}{4},\qquad \frac{2\gamma}{\tau}<\frac{\eta}{4},\\ \frac{4}{K(p/2)^{N_M(m)+1}}<\frac{\eta}{4},\qquad \frac{1}{(1-p)^{K(8m+5)^3}(R-m)}<\frac{\eta}{4}. \end{gathered}(6) The quarter-face requirement in (7) concerns every center cc and quarter-face index ff at each of the three indicated radii. ℙq(EM(c;m,n,F(c,n;f)))>1−γ,n∈{2r,2r+1,10r}.\mathbb{P}_q\bigl(E_M(c;m,n,F(c,n;f))\bigr)>1-\gamma, \qquad n\in\{2r,2r+1,10r\}.(7)

The Claim

Theorem 1. Let GG be a finite lattice subgraph, π\pi an opening probability assignment, and 0<q≤p0<10<q\le p_0<1. Take nonnegative integers a,K,Ra,K,R with K>0K>0 and a<Ra<R, a lattice box BB, a source o∈V(G)o\in V(G), and a target T⊆ℤ3T\subseteq\mathbb{Z}^3. Suppose GG contains B(R)B^{(R)}, o∉B(R)o\notin B^{(R)}, and π\pi has uniform density qq on B(R)B^{(R)}. Let γ≥0\gamma\ge0 and τ>0\tau>0. Suppose every c∈B(R)c\in B^{(R)} admits an integer ℓ≥a\ell\ge a such that GG contains Λℓ(c)\Lambda_\ell(c) and ℙπ(EM(c;a,ℓ,T))≥1−γ\mathbb{P}_\pi(E_M(c;a,\ell,T))\ge1-\gamma. Then the extension failure bound is (8). ℙπ(o↔GB,o↮GT)≤1(1−p0)K(8a+5)3(R−a)+4KqκM(a)+2γτ+τ.\mathbb{P}_\pi(o\leftrightarrow_G B,\ o\not\leftrightarrow_G T) \le \frac{1}{(1-p_0)^{K(8a+5)^3}(R-a)} +\frac{4}{Kq^{\kappa_M(a)}}+\frac{2\gamma}{\tau}+\tau.(8)✦

For renormalization scale data satisfying (6) and (7), suppose GG contains a region DD and π\pi has uniform density qq on DD. Take o∈V(G)\Do\in V(G)\setminus D and a box BB with B(R)⊆DB^{(R)}\subseteq D. Suppose every u∈B(R)u\in B^{(R)} admits n∈{2r,2r+1,10r}n\in\{2r,2r+1,10r\}, a signed lattice symmetry SS, a center cc, and a quarter-face index ff such that S(c)=uS(c)=u, Λn(u)⊆D\Lambda_n(u)\subseteq D, and S(F(c,n;f))⊆TS(F(c,n;f))\subseteq T. Then the scale-specialized bound is (9). ℙπ(o↔GB,o↮GT)<η.\mathbb{P}_\pi(o\leftrightarrow_G B,\ o\not\leftrightarrow_G T)<\eta.(9)✦

In the paper, the result reads: with the scales and qq fixed by Lemma 3.3, let GG be a finite subgraph of ℤ3\mathbb{Z}^3 containing a rectangle DD and every lattice edge internal to DD, give the variables independent laws with parameter qq on all variables internal to DD and arbitrary parameters, including zero and one, elsewhere, and take o∈V(G)o\in V(G) and T⊆V(G)T\subseteq V(G); if a rectangle BB satisfies B(R)⊆DB^{(R)}\subseteq D and o∉B(R)o\notin B^{(R)}, and for every u∈B(R)u\in B^{(R)} some ℓ∈{2r,2r+1,10r}\ell\in\{2r,2r+1,10r\} has Λℓ(u)⊆D\Lambda_\ell(u)\subseteq D and a quarter-face F(u,ℓ;f)⊆TF(u,\ell;f)\subseteq T, then ℙπ(o↔GB,o↮GT)<η\mathbb{P}_\pi(o\leftrightarrow_G B,\ o\not\leftrightarrow_G T)<\eta.

What Is Checked

The claims (8) and (9) are stated over independent Bernoulli bond or site configurations on finite lattice subgraphs, with uniform density on the specified regions and the stated seed or quarter-face hypotheses. The underlying graph adjacency is symmetric and loopless, and each signed lattice action in (5) is a bijection of ℤ3\mathbb{Z}^3.