OctoInformBeta

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

Lemma 5.1 [Finite representatives for all local scale tests]: A deterministic set in [1,2][1,2] works simultaneously for every table outcome

Finitely many scales represent every local test.

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

At a fixed center, a normalized scale moves each dyadic point along a short interval. The local tests record selector bits and the terminal bit reached by routing that point. Their dependence on scale is measured on periodic dyadic grids.

Fix natural parameters M,d,g,r0M,d,g,r_0. The router has children 0,…,M0,\ldots,M, of which 0,…,M−10,\ldots,M-1 are nondefault and MM is default; its total number of children is q=M+1q=M+1.✦ A decision node is a path PP with |P|<d|P|<d✦, and a terminal path is a path LL with |L|=d|L|=d.✦ Write PiPi for appending the child ii to PP. An admissible edge is a nonempty path PP with |P|≤d|P|\le d, identifying the edge into its last child.✦

Separated window data WW consist of a length function r:ℕ→ℕr:\mathbb N\to\mathbb N and natural-valued assignments a,b,b*a,b,b^* on paths. Put h(P)=d+1−|P|h(P)=d+1-|P| for an admissible edge. Their defining conditions are (1), where PP and QQ are admissible in the first four rows and Pi,PjVPi,PjV are admissible in the last row. Here P≼QP\preceq Q means prefix inclusion, P≺QP\prec Q means proper prefix inclusion, and VV is a continuation: r(h(P))≥max⁡{1,r0},a(P)≥3,b(P)+1=a(P)+r(h(P)),b*(P)+1≤a(P)+2r(h(P)),b(P)≤b(Q)≤b*(P)(P≼Q),b(P)+g<a(Q)(P≺Q),b(Pi)+g<a(PjV)(i<j).\begin{aligned} r(h(P))&\ge\max\{1,r_0\}, & a(P)&\ge3,\\ b(P)+1&=a(P)+r(h(P)), & b^*(P)+1&\le a(P)+2r(h(P)),\\ b(P)&\le b(Q)\le b^*(P) &&(P\preceq Q),\\ b(P)+g&<a(Q) &&(P\prec Q),\\ b(Pi)+g&<a(PjV) &&(i<j). \end{aligned}(1)

Fix a decision node UU and a real center xx. The local windows and child grid levels are defined by (2); here hh is the height of UU and kk is the common window length: h=d−|U|,k=r(h),ai=a(Ui),Bi=b*(Ui),WU,i={ai,…,ai+k−1}(0≤i<M).\begin{gathered} h=d-|U|,\qquad k=r(h),\qquad a_i=a(Ui),\qquad B_i=b^*(Ui),\\ W_{U,i}=\{a_i,\ldots,a_i+k-1\}\quad(0\le i<M). \end{gathered}(2)

Grids and Routing

For y∈ℝy\in\mathbb R, write {y}=y−⌊y⌋\{y\}=y-\lfloor y\rfloor. The periodic mesh key κv(y)\kappa_v(y) at a real parameter vv and its dyadic specialization Kb(y)K_b(y) at a natural level bb are defined in (3). The latter indexes the 2b+22^{b+2} half-open cells of the periodic grid, including each cell’s left boundary. κv(y)=⌊v{y}⌋,Kb(y)=κ2b+2(y)∈{0,…,2b+2−1}.\kappa_v(y)=\lfloor v\{y\}\rfloor,\qquad K_b(y)=\kappa_{2^{b+2}}(y)\in\{0,\ldots,2^{b+2}-1\}.(3)

The selector window endpoint at a decision node PP and nondefault child ii is bS(P,i)=b(Pi)b_S(P,i)=b(Pi)✦; the terminal window endpoint at a terminal path LL is bT(L)=b(L)b_T(L)=b(L).✦ A selector table address consists of (P,i)(P,i) and a natural index below 2bS(P,i)+22^{b_S(P,i)+2}.✦ A terminal table address consists of LL and a natural index below 2bT(L)+22^{b_T(L)+2}.✦ A Boolean selector table ωS\omega_S and a terminal Boolean table ωT\omega_T assign bits to their respective addresses. The queried addresses at a point yy are those in (4). AS(P,i;y)=((P,i),KbS(P,i)(y)),AT(L;y)=(L,KbT(L)(y)).A_S(P,i;y)=\bigl((P,i),K_{b_S(P,i)}(y)\bigr),\qquad A_T(L;y)=\bigl(L,K_{b_T(L)}(y)\bigr).(4)

The extended selector value SωS(P,i;y)S_{\omega_S}(P,i;y) is ωS(AS(P,i;y))\omega_S(A_S(P,i;y)) when |P|<d|P|<d, and is 00 otherwise.✦ A routing rule FF assigns a child to a path and a vector of MM bits. Its selector-dependent choice cωSc_{\omega_S} is defined by (5). cωS(P,y)=F(P,(SωS(P,j;y))0≤j<M).c_{\omega_S}(P,y) =F\bigl(P,(S_{\omega_S}(P,j;y))_{0\le j<M}\bigr).(5)

For a child choice cc, the iterated routing path ℛnc(P;y)\mathcal R_n^c(P;y) is the path obtained by nn successive choices from PP, as in (6).✦ When |P|+ℓ=d|P|+\ell=d, the local routed leaf Lℓc(P;y)L_\ell^c(P;y) is the path obtained after ℓ\ell successive choices from PP at yy, as in (6).✦ ℛ0c(P;y)=P,ℛn+1c(P;y)=ℛnc(P;y)c(ℛnc(P;y),y),Lℓc(P;y)=ℛℓc(P;y).\begin{gathered} \mathcal R_0^c(P;y)=P,\qquad \mathcal R_{n+1}^c(P;y) =\mathcal R_n^c(P;y)c(\mathcal R_n^c(P;y),y),\\ L_\ell^c(P;y)=\mathcal R_\ell^c(P;y). \end{gathered}(6)

Choose a natural ℓ\ell with |Ui|+ℓ=d|Ui|+\ell=d for every i<Mi<M. For a table outcome ω=(ωS,ωT)\omega=(\omega_S,\omega_T), the concrete local test Qiω(y)Q_i^\omega(y) is the Boolean conjunction in (7), written as a product of bits. Qiω(y)=ωS(AS(U,i;y))ωT(AT(LℓcωS(Ui;y);y)).Q_i^\omega(y)=\omega_S(A_S(U,i;y))\, \omega_T\bigl(A_T(L_\ell^{c_{\omega_S}}(Ui;y);y)\bigr).(7)

For an outcome set Ω\Omega, let 𝒞Ω\mathcal C_\Omega be the collection of Boolean families Φiη:ℝ→{0,1}\Phi_i^\eta:\mathbb R\to\{0,1\}, indexed by η∈Ω\eta\in\Omega and i<Mi<M, that take equal values whenever the corresponding keys KBiK_{B_i} agree. Write zn(t)=x+t2−nz_n(t)=x+t2^{-n} for the point tested at normalized scale t∈[1,2]t\in[1,2].

The Claim

Theorem 1. Fix separated window data WW, a decision node UU, and a real center xx as above. Let routing depend on selector values through a rule FF as in (5), and let |Ui|+ℓ=d|Ui|+\ell=d for every i<Mi<M. For all table outcomes and real points y,zy,z, the concrete local tests of (7) have the grid factorization (8): Qiω(y)=Qiω(z)whenever KBi(y)=KBi(z),ω=(ωS,ωT),i<M.Q_i^\omega(y)=Q_i^\omega(z) \quad\text{whenever }K_{B_i}(y)=K_{B_i}(z), \qquad \omega=(\omega_S,\omega_T),\quad i<M.(8)✦ For every outcome set Ω\Omega, the universal window representatives satisfy (9): ∃R⊂[1,2] finite,|R|≤1+Mk22k+2,∀t∈[1,2]∃t′∈R∀Φ∈𝒞Ω∀η∈Ω∀i<M∀n∈WU,i,Φiη(zn(t′))=Φiη(zn(t)).\begin{gathered} \exists\,R\subset[1,2]\text{ finite},\qquad |R|\le1+Mk2^{2k+2},\\ \forall t\in[1,2]\ \exists t'\in R\quad\\ \forall\Phi\in\mathcal C_\Omega\ \forall\eta\in\Omega\ \forall i<M\ \forall n\in W_{U,i},\\ \Phi_i^\eta(z_n(t'))=\Phi_i^\eta(z_n(t)). \end{gathered}(9)✦ The concrete local test representatives satisfy (10): ∃R⊂[1,2] finite,|R|≤1+Mk22k+2,∀t∈[1,2]∃t′∈R∀ω=(ωS,ωT)∀i<M∀n∈WU,i,Qiω(zn(t′))=Qiω(zn(t)).\begin{gathered} \exists\,R\subset[1,2]\text{ finite},\qquad |R|\le1+Mk2^{2k+2},\\ \forall t\in[1,2]\ \exists t'\in R\quad\\ \forall\omega=(\omega_S,\omega_T)\ \forall i<M\ \forall n\in W_{U,i},\\ Q_i^\omega(z_n(t'))=Q_i^\omega(z_n(t)). \end{gathered}(10)✦

In the paper, the result reads: for a stable center xx and a full center-exposure atom ξ\xi whose first default node UU has height hh, with k=rhk=r_h and qq denoting the paper’s total number of children, there is a finite R(x,U)⊂[1,2]R(x,U)\subset[1,2], depending only on x,Ux,U and the deterministic windows, with |R(x,U)|≤20qk(1+22k+2)|R(x,U)|\le20qk(1+2^{2k+2}), such that every t∈[1,2]t\in[1,2] has a representative t′∈R(x,U)t'\in R(x,U) satisfying Qi(x+t2−n)=Qi(x+t′2−n)Q_i(x+t2^{-n})=Q_i(x+t'2^{-n}) for every nondefault i<qi<q and every n∈WU,in\in W_{U,i}, simultaneously for every realization consistent with ξ\xi.

What Is Checked

The statements are over separated window data on a finite router, arbitrary real centers and Boolean tables with selector-dependent routing; the universal window statement ranges over arbitrary outcome sets and the families in 𝒞Ω\mathcal C_\Omega.