OctoInformBeta

OpenAI preprint, September 2026 · §6. Adaptive exploration and the final contradiction

§6 (proof of Theorem 1.1): The finite good set estimate and percolation at a lower density when 0<p<10<p<1

Survival and a lower percolating parameter.

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 coarse exploration on ℤ2\mathbb Z^2 follows connections from the origin through finite subgraphs of the nearest-neighbor lattice ℤ3\mathbb Z^3. Its good vertices record successful coarse boxes. The question is whether this set can remain finite, and which lattice densities support an infinite origin cluster.

Fix a model MM, either bond or site. A lattice bond (x,i)(x,i) joins xx to x+eix+e_i, where i∈{0,1,2}i\in\{0,1,2\} and eie_i is the corresponding coordinate unit vector.✦ The variable set EME_M consists of these bonds in the bond model and of lattice vertices in the site model.✦ A configuration ω\omega assigns an open or closed state to every variable.✦ Under the Bernoulli law ℙpM\mathbb P_p^M, the variables are independent with opening probability p‾=max⁡{0,min⁡{1,p}}\bar p=\max\{0,\min\{1,p\}\}.✦ The clipped density of a real parameter pp is p‾=max⁡{0,min⁡{1,p}}\bar p=\max\{0,\min\{1,p\}\}, a value in [0,1][0,1].

Two lattice vertices are nearest neighbors when one is a positive coordinate step of the other.✦ In a configuration ω\omega, open bond adjacency between lattice vertices xx and yy means that their joining bond is open, and open site adjacency means that xx and yy are nearest neighbors and both are open. An open connection in a configuration ω\omega is a finite path of open bond adjacencies in the bond model or open site adjacencies in the site model. A bond path of length zero connects a vertex to itself, whereas a site path of length zero requires that vertex to be open. The cluster C(x)C(x) is the set of vertices connected to xx in the chosen model. Define the origin probability θM\theta_M and the model’s critical parameter pcMp_c^M by (1): θM(p)=ℙpM(|C(0)|=∞),pcM=inf⁡{p∈[0,1]:θM(p)>0}.\theta_M(p)=\mathbb P_p^M\bigl(|C(0)|=\infty\bigr), \qquad p_c^M=\inf\{p\in[0,1]:\theta_M(p)>0\}.(1) For either model, the infinite-cluster event at a lattice vertex xx is the set of configurations in which C(x)C(x) is infinite.✦

A finite lattice graph GG consists of a finite vertex set and a finite set of bonds whose endpoints both belong to that set.✦ The underlying graph of a finite lattice graph GG is the graph on its vertex set in which two vertices are adjacent when a bond of GG has those endpoints, in either order.✦ Connections within GG use its bonds and vertices, with the same open path conventions. For a target set TT of lattice vertices, the notation 0↔GT0\leftrightarrow_G T means that some vertex of GG in TT is connected to the origin within GG.✦ For a finite lattice graph GG, the variable set IM(G)I_M(G) is the set of its edge variables in the bond model and its vertex variables in the site model.✦

Finite Scale Parameters

A lattice box is a closed coordinate interval between ordered corners in ℤ3\mathbb Z^3.✦ Its vertices are the lattice points in that interval, and its bonds have both endpoints there. For a lattice box BB, the variable set IM(B)I_M(B) consists of its bonds in the bond model and its vertices in the site model.✦ For a lattice center uu and a nonnegative integer radius bb, the cube Λb(u)\Lambda_b(u) is the box of points xx satisfying |xi−ui|≤b|x_i-u_i|\le b in each coordinate.✦ For a seed radius mm, the seed variable count NM(m)N_M(m) is |IM(Λm(0))||I_M(\Lambda_m(0))|.✦

For a coarse vertex v=(v1,v2)∈ℤ2v=(v_1,v_2)\in\mathbb Z^2 and integer scale r≥0r\ge0, the center cr(v)c_r(v) is a lattice vertex, and the relay cube BvB_v and surrounding cube QvQ_v are lattice boxes, given by (2). cr(v)=((20r+1)v1,(20r+1)v2,0),Bv=Λ4r(cr(v)),Qv=Λ10r(cr(v)).c_r(v)=\bigl((20r+1)v_1,(20r+1)v_2,0\bigr),\qquad B_v=\Lambda_{4r}(c_r(v)),\qquad Q_v=\Lambda_{10r}(c_r(v)).(2)

A quarter-face index consists of a coordinate ii and signs σ∈{−1,1}3\sigma\in\{-1,1\}^3.✦ For a lattice center cc and radius nn, the quarter-face F(c,n;i,σ)F(c,n;i,\sigma) is the set of lattice vertices xx satisfying xi=ci+σinx_i=c_i+\sigma_i n and 0≤σj(xj−cj)≤n0\le\sigma_j(x_j-c_j)\le n for j≠ij\ne i.✦ There are 24 such signed-coordinate quarter-faces, with boundary overlaps permitted.

A vertex is available unconditionally in the bond model, and when it is open in the site model.✦ Let 𝒮M(c;m,n,T)\mathcal S_M(c;m,n,T) be the event that an available vertex in Λm(c)\Lambda_m(c) is joined to a vertex of TT by open steps whose endpoints belong to Λn(c)\Lambda_n(c).✦

Finite scale parameters consist of densities p,qp,q, allowances δ,η,τ,γ\delta,\eta,\tau,\gamma, and integers m,K,R,rm,K,R,r satisfying (3). Here mm is the seed radius and KK is the trial count. 0<p<1,q∈[0,1],p/2<q<p,δ,η,τ,γ∈(0,1),m,K,R,r∈ℕ,m>0,K>0,m<R,100(R+1)<r.\begin{gathered} 0<p<1,\qquad q\in[0,1],\qquad p/2<q<p,\qquad \delta,\eta,\tau,\gamma\in(0,1),\\ m,K,R,r\in\mathbb N,\qquad m>0,\quad K>0,\quad m<R,\quad 100(R+1)<r. \end{gathered}(3) The contour error tolerance ε\varepsilon and the required error budgets are given in (4): ε=12(1100)7,δ+52ηδ<ε,τ<η4,γ<δ,2γτ<η4,4K(p/2)NM(m)+1<η4,1(1−p)K(8m+5)3(R−m)<η4.\begin{gathered} \varepsilon=\frac12\left(\frac1{100}\right)^7,\qquad \delta+\frac{52\eta}{\delta}<\varepsilon,\qquad \tau<\frac\eta4,\qquad \gamma<\delta,\qquad \frac{2\gamma}{\tau}<\frac\eta4,\\ \frac{4}{K(p/2)^{N_M(m)+1}}<\frac\eta4,\qquad \frac{1}{(1-p)^{K(8m+5)^3}(R-m)}<\frac\eta4. \end{gathered}(4) These parameters also satisfy the quarter-face requirement (5) for every center cc, every quarter-face index (i,σ)(i,\sigma) and each n∈{2r,2r+1,10r}n\in\{2r,2r+1,10r\}: ℙqM(𝒮M(c;m,n,F(c,n;i,σ)))>1−γ.\mathbb P_q^M\bigl(\mathcal S_M(c;m,n,F(c,n;i,\sigma))\bigr)>1-\gamma.(5)

The Exploration State

A finite observation history h=(J,ξ)h=(J,\xi) consists of a finite known set J⊂EMJ\subset E_M and a state assignment ξ\xi, whose values on JJ record the observations.✦ For a finite observation history h=(J,ξ)h=(J,\xi), the measure ℙq,h\mathbb P_{q,h} is the independent Bernoulli product measure whose opening probability is qq outside JJ, one at an observed open variable, and zero at an observed closed variable. Pinned parameters retain unobserved opening probabilities and fix observed variables to their observed states. The history-conditioned Bernoulli parameters at density qq are the pinned parameters obtained from the constant family ae=qa_e=q.

Let 𝒟={(0,−1),(1,0),(0,1),(−1,0)}\mathcal D=\{(0,-1),(1,0),(0,1),(-1,0)\} be the four coarse directions. For coarse vertices v,wv,w, let LM(v,w)L_M(v,w) be the bonds with one endpoint in QvQ_v and the other in QwQ_w in the bond model, and the empty set in the site model.✦ For coarse vertices v,uv,u and a finite coarse set DD, define the root variables A0A_0, the processed region variables A(D)A(D), and the box and interface variables H(v,u)H(v,u) by (6). A0=IM(Q0)∪⋃d∈𝒟LM(0,d),A(D)=A0∪⋃v∈DIM(Qv)∪⋃v,w∈DLM(v,w),H(v,u)=IM(Qv)∪LM(v,u).\begin{aligned} A_0&=I_M(Q_0)\cup\bigcup_{d\in\mathcal D}L_M(0,d),\\ A(D)&=A_0\cup\bigcup_{v\in D}I_M(Q_v) \cup\bigcup_{v,w\in D}L_M(v,w),\\ H(v,u)&=I_M(Q_v)\cup L_M(v,u). \end{aligned}(6)

An exploration packet t=(v,u,G)t=(v,u,G) assigns a coarse vertex vv and its parent uu to a finite parent-chain graph GG. This graph contains the origin and contains QvQ_v as a subgraph, including all its vertices and bonds. The root connection event ℰt\mathcal E_t of an exploration packet t=(v,u,G)t=(v,u,G) is 0↔GBv0\leftrightarrow_G B_v.✦

An exploration state ss consists of a finite observation history hh, finite sets Ds,𝒢s,ℬsD_s,\mathcal G_s,\mathcal B_s of coarse vertices designated processed, good and bad, and an ordered queue of exploration packets.✦ The visited set VsV_s of an exploration state ss is its processed set DsD_s together with the vertices in its queue.✦ A colored exploration state ss has its processed set DsD_s partitioned into its good set 𝒢s\mathcal G_s and its bad set ℬs\mathcal B_s. Closure of an exploration state ss means that every coarse neighbor of every vertex of its good set 𝒢s\mathcal G_s belongs to its visited set VsV_s.✦ Coverage of an exploration state ss means that each vertex of its visited set VsV_s is good or is a coarse neighbor of a good vertex.✦

An exploration state ss with history h=(J,ξ)h=(J,\xi) is valid at q,δq,\delta when the origin is processed, queued vertices are distinct and unprocessed, their parents are processed, its history satisfies J⊂A(Ds)J\subset A(D_s), and every queued packet t=(v,u,G)t=(v,u,G) satisfies IM(G)⊂J∪H(v,u)I_M(G)\subset J\cup H(v,u) and ℙq,h(ℰt)>1−δ\mathbb P_{q,h}(\mathcal E_t)>1-\delta.

Let sn(ω)s_n(\omega) be the state after nn successive exploration steps at q,δq,\delta, starting from ss and using the same configuration ω\omega throughout; s0(ω)=ss_0(\omega)=s.✦ For an initial exploration state ss and a configuration ω\omega, the eventual good set 𝒢∞(s,ω)\mathcal G_\infty(s,\omega) is the subset of ℤ2\mathbb Z^2 given by (7). 𝒢∞(s,ω)=⋃n≥0𝒢sn(ω).\mathcal G_\infty(s,\omega) =\bigcup_{n\ge0}\mathcal G_{s_n(\omega)}.(7)✦

The Claim

Theorem 1. Fix either model MM and finite scale parameters satisfying (3), (4) and (5). Let ss be a valid, colored, closed and covered state at q,δq,\delta, with history hh and no bad vertices. The finite good set estimate is (8): ℙq,h(|𝒢∞(s,ω)|<∞and0∈𝒢∞(s,ω))≤25144.\mathbb P_{q,h}\bigl( |\mathcal G_\infty(s,\omega)|<\infty \ \text{and}\ 0\in\mathcal G_\infty(s,\omega)\bigr) \le\frac{25}{144}.(8)✦ The weighted trail series has the sum (9): ∑n=0∞4n4n(2ε)n/7=25144.\sum_{n=0}^{\infty}4n\,4^n(2\varepsilon)^{n/7}=\frac{25}{144}.(9)✦ For every choice of these finite scale parameters, (10) holds: θM(q)>0.\theta_M(q)>0.(10)✦ For every density p′∈(0,1)p'\in(0,1) with θM(p′)>0\theta_M(p')>0, there is a density q′q' with the properties in (11): q′∈[0,1],q′<p′,θM(q′)>0.q'\in[0,1],\qquad q'<p',\qquad \theta_M(q')>0.(11)✦

In the paper, the result reads: if 𝒢∞\mathcal G_\infty is finite then all its outside neighbors have been processed bad, a deterministic contour union bound gives conditional probability less than one for its finiteness, and infinitely many good coarse boxes give infinitely many vertices of C(0)C(0) at a density q<pcMq<p_c^M, a contradiction.

What Is Checked

The statements concern the Bernoulli bond and site laws on ℤ3\mathbb Z^3, the finite scale parameters, and the history-conditioned exploration states with the stated closure, coverage and validity conditions. The underlying adjacency of a finite lattice graph is loopless.