OctoInformBeta

OpenAI preprint, September 2026 · §5. All normalized scales and all centers

Proposition 5.2 [Raw failure estimate; weaker concrete Lean event]: Avoidance at every positive dyadic index, with separate window and routing budgets

Small failure probability at stable centers.

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

A finite tree of periodic dyadic tables assigns a terminal bit to each real point. The points assigned the bit one form the random routed set.✦ Fix natural numbers MM and dd, allowing zero. The ordered children are 0,…,M0,\ldots,M, with default child MM; the paper’s parameter for the total number of children is therefore M+1M+1.✦ A decision node PP is a path of length |P|<d|P|<d, and a terminal path LL has length dd. Write 𝒫\mathcal P and ℒ\mathcal L for these two sets, ⌀\varnothing for the root, and PiPi for the path obtained by appending child ii.

For a real mesh parameter hh, the periodic mesh key κh\kappa_h is the floor function in (1).✦ At a natural level bb, the finite dyadic grid key kbk_b takes values in {0,…,2b+2−1}\{0,\ldots,2^{b+2}-1\}: κh(y)=⌊h{y}⌋,kb(y)=κ2b+2(y),{y}=y−⌊y⌋.\kappa_h(y)=\lfloor h\{y\}\rfloor, \qquad k_b(y)=\kappa_{2^{b+2}}(y), \qquad \{y\}=y-\lfloor y\rfloor.(1) The dyadic cells are half-open and include their left boundaries.

Fix selector resolutions bS(P,i)b_S(P,i) for P∈𝒫P\in\mathcal P and i<Mi<M, and terminal resolutions bT(L)b_T(L) for L∈ℒL\in\mathcal L. The selector and terminal address sets S,TS,T, and their queried addresses aS,aTa_S,a_T, are given in (2): S={(P,i,j):P∈𝒫,i<M,0≤j<2bS(P,i)+2},T={(L,j):L∈ℒ,0≤j<2bT(L)+2},aS(P,i,y)=(P,i,kbS(P,i)(y)),aT(L,y)=(L,kbT(L)(y)).\begin{aligned} S&=\{(P,i,j):P\in\mathcal P,\ i<M,\ 0\le j<2^{b_S(P,i)+2}\},\\ T&=\{(L,j):L\in\mathcal L,\ 0\le j<2^{b_T(L)+2}\},\\ a_S(P,i,y)&=(P,i,k_{b_S(P,i)}(y)),\qquad a_T(L,y)=(L,k_{b_T(L)}(y)). \end{aligned}(2) A Boolean selector table is a function σ:S→{0,1}\sigma:S\to\{0,1\}, and a terminal Boolean table is a function τ:T→{0,1}\tau:T\to\{0,1\}.

For 0≤p≤10\le p\le1, let βp\beta_p be the Bernoulli bit law assigning masses pp and 1−p1-p to one and zero, respectively.✦ The fair bit law is β1/2\beta_{1/2}.✦ The selector law and terminal law are the independent products over SS and TT; their joint outcome law on ω=(σ,τ)\omega=(\sigma,\tau) is ℙp=(⨂s∈Sβ1/2)⊗(⨂u∈Tβp).\mathbb P_p= \left(\bigotimes_{s\in S}\beta_{1/2}\right) \otimes\left(\bigotimes_{u\in T}\beta_p\right).(3)✦

Under (3), the extended selector value Vσ(P,i,y)V_\sigma(P,i,y) reads the queried bit when PP is a decision node and is zero otherwise.✦ The pointwise first-success child cσ(P,y)c_\sigma(P,y) is the first nondefault child with selector value one, or the default child when every nondefault selector fails: Vσ(P,i,y)={σ(aS(P,i,y)),|P|<d,0,|P|≥d,cσ(P,y)=min⁡({i<M:Vσ(P,i,y)=1}∪{M}).\begin{aligned} V_\sigma(P,i,y)&= \begin{cases} \sigma(a_S(P,i,y)),&|P|<d,\\ 0,&|P|\ge d, \end{cases}\\ c_\sigma(P,y)&=\min\bigl(\{i<M:V_\sigma(P,i,y)=1\}\cup\{M\}\bigr). \end{aligned}(4) At a fixed center xx, the choice cσ(P,x)c_\sigma(P,x) in (4) is the center first-success child.✦

For any child selection c(P,y)c(P,y), its iterated routing path Rkc(P,y)R_k^c(P,y) makes kk successive choices starting from PP.✦ The recursion and the routed terminal path Lσ(y)L_\sigma(y) are R0c(P,y)=P,Rk+1c(P,y)=Rkc(P,y)c(Rkc(P,y),y),Lσ(y)=Rdcσ(⌀,y).R_0^c(P,y)=P,\qquad R_{k+1}^c(P,y)=R_k^c(P,y)\,c(R_k^c(P,y),y),\qquad L_\sigma(y)=R_d^{c_\sigma}(\varnothing,y).(5)

For a selector space AA, an address space KK, an address map a:A×ℝ→Ka:A\times\mathbb R\to K and z=(u,τ)∈A×{0,1}Kz=(u,\tau)\in A\times\{0,1\}^K, the address-selected set 𝒮a(z)\mathcal S_a(z) is the set of real points whose selected bit is one, as in (6).✦ For the routed path (5), the routed terminal address aσ(y)a_\sigma(y) and the random routed set BωB_\omega are 𝒮a(u,τ)={y∈ℝ:τ(a(u,y))=1},aσ(y)=aT(Lσ(y),y),Bω=𝒮(σ,y)↦aσ(y)(σ,τ).\begin{aligned} \mathcal S_a(u,\tau)&=\{y\in\mathbb R:\tau(a(u,y))=1\},\\ a_\sigma(y)&=a_T(L_\sigma(y),y),\qquad B_\omega=\mathcal S_{(\sigma,y)\mapsto a_\sigma(y)}(\sigma,\tau). \end{aligned}(6)

A center exposure Xx(σ)X_x(\sigma) is the assignment of the selector bits queried at xx to all pairs (P,i)(P,i) with P∈𝒫P\in\mathcal P and i<Mi<M.✦ The zero-filled completion σˆx,ξ\widehat\sigma_{x,\xi} of a center exposure ξ\xi at xx is a selector table that assigns ξ(P,i)\xi(P,i) to the address queried at xx for each pair (P,i)(P,i) with P∈𝒫P\in\mathcal P and i<Mi<M, and zero to every other address.✦ The center exposure atom Ax(ξ)A_x(\xi) is the set of outcomes with exposure ξ\xi: Xx(σ)(P,i)=σ(aS(P,i,x)),σˆx,ξ(P,i,j)={ξ(P,i),j=kbS(P,i)(x),0,j≠kbS(P,i)(x),Ax(ξ)={(σ,τ):Xx(σ)=ξ}.\begin{aligned} X_x(\sigma)(P,i)&=\sigma(a_S(P,i,x)),\\ \widehat\sigma_{x,\xi}(P,i,j)&= \begin{cases} \xi(P,i),&j=k_{b_S(P,i)}(x),\\ 0,&j\ne k_{b_S(P,i)}(x), \end{cases}\\ A_x(\xi)&=\{(\sigma,\tau):X_x(\sigma)=\xi\}. \end{aligned}(7) An exposure ξ\xi is good if the completed table in (7) routes xx through the default child at some stage k<dk<d.✦

For natural parameters g,r0g,r_0, separated window data WW consist of a length function r:ℕ0→ℕ0r:\mathbb N_0\to\mathbb N_0 and natural-valued assignments a,b,b*a,b,b^* on finite paths. An admissible edge is a nonempty path PP with |P|≤d|P|\le d; its window is the integer interval [a(P),b(P)][a(P),b(P)].✦ Put RP=r(d+1−|P|)R_P=r(d+1-|P|). Write P≼QP\preceq Q for the prefix relation and P≺QP\prec Q for a proper prefix. The separation and padding conditions are (8), for admissible paths wherever they occur and any finite continuation UU: RP≥max⁡{1,r0},a(P)≥3,b(P)+1=a(P)+RP,b*(P)+1≤a(P)+2RP,b(P)≤b(Q)≤b*(P)(P≼Q),b(P)+g<a(Q)(P≺Q),b(Pi)+g<a(PjU)(i<j).\begin{aligned} R_P&\ge\max\{1,r_0\},\qquad a(P)\ge3,\qquad b(P)+1=a(P)+R_P,\\ b^*(P)+1&\le a(P)+2R_P,\\ b(P)&\le b(Q)\le b^*(P) &&(P\preceq Q),\\ b(P)+g&<a(Q) &&(P\prec Q),\\ b(Pi)+g&<a(PjU) &&(i<j). \end{aligned}(8)✦ The nondefault child edge at (P,i)(P,i) is PiPi.✦ Its selector window endpoint, and the terminal window endpoint at LL, give the resolutions bS(P,i)=b(Pi),bT(L)=b(L).b_S(P,i)=b(Pi),\qquad b_T(L)=b(L).(9)

A stable center x∈ℝx\in\mathbb R for WW is a point satisfying (10) for the resolutions (9), every selector table σ\sigma, every P∈𝒫P\in\mathcal P and i<Mi<M, and every natural exponent n>b(Pi)+gn>b(Pi)+g: Vσ(P,i,x)=Vσ(P,i,x+t2−n)(1≤t≤2).V_\sigma(P,i,x)=V_\sigma(P,i,x+t2^{-n}) \qquad(1\le t\le2).(10)✦

For a set C⊆ℝC\subseteq\mathbb R, an index set N⊆ℕ0N\subseteq\mathbb N_0 and a real sequence α\alpha, the exceptional-center set E(C,N;α)E(C,N;\alpha) consists of the centers of normalized dilates avoiding CC at every index in NN.✦ For a family Cω⊆ℝC_\omega\subseteq\mathbb R, the dyadic failure event FN(C,x)F_N(C,x) is the set of outcomes for which xx belongs to the exceptional-center set in (11); write ℕ+={1,2,…}\mathbb N_+=\{1,2,\ldots\} and F+(C,x)=Fℕ+(C,x)F_+(C,x)=F_{\mathbb N_+}(C,x): E(C,N;α)={x:∃t∈[1,2],x+tαn∉C for every n∈N},FN(C,x)={ω:x∈E(Cω,N;(2−n)n≥0)}.\begin{aligned} E(C,N;\alpha)&=\{x:\exists t\in[1,2], x+t\alpha_n\notin C\text{ for every }n\in N\},\\ F_N(C,x)&=\{\omega:x\in E(C_\omega,N;(2^{-n})_{n\ge0})\}. \end{aligned}(11)

The Claim

Theorem 1. Let 0≤p≤10\le p\le1 and let WW be separated window data with M≥1M\ge1. Suppose the window budget J(1−p/2)MR≤p(R,J∈ℕ0,R≥r0,J≤1+MR22R+2)J(1-p/2)^{MR}\le p \quad\bigl(R,J\in\mathbb N_0,\ R\ge r_0, J\le1+MR\,2^{2R+2}\bigr)(12) holds. Let xx be stable for WW and let ξ\xi be a good exposure. For the routed set (6) with endpoint resolutions (9), ℙp(F+(B,x)∩Ax(ξ))≤pℙp(Ax(ξ)).\mathbb P_p\bigl(F_+(B,x)\cap A_x(\xi)\bigr) \le p\,\mathbb P_p\bigl(A_x(\xi)\bigr).(13)✦

For arbitrary natural numbers M,dM,d, take selector and terminal resolutions on a tree with MM nondefault children and depth dd, any family Cω⊆ℝC_\omega\subseteq\mathbb R, any N⊆ℕ0N\subseteq\mathbb N_0 and any center xx. Suppose the routing tail and every good exposure atom satisfy (1−2−M)d≤p,ℙp(FN(C,x)∩Ax(ξ))≤pℙp(Ax(ξ))(ξ good).\begin{aligned} (1-2^{-M})^d&\le p,\\ \mathbb P_p\bigl(F_N(C,x)\cap A_x(\xi)\bigr) &\le p\,\mathbb P_p\bigl(A_x(\xi)\bigr) \qquad(\xi\text{ good}). \end{aligned}(14) Then ℙp(FN(C,x))≤2p.\mathbb P_p\bigl(F_N(C,x)\bigr)\le2p.(15)✦

In the paper, the result reads: for every fixed stable center xx, the probability that there exists t∈[1,2]t\in[1,2] with x+t2−n∉Bx+t2^{-n}\notin B for every positive index nn in the finite union NN of the edge windows is at most 2p2p.

What Is Checked

The estimates (13) and (15) are stated over the finite independent table law (3), with selector stability and the window budget (12) for the concrete routed set, and the routing tail and hypotheses on good exposure atoms (14) for an arbitrary family of sets.