OctoInformBeta

OpenAI preprint, September 2026 · §4. The exact correction and copy hierarchy · §4

Proposition 4.5 [One-stage density and active-count reduction]: The weighted projection intertwines the copy matrix with the residual, and for 2i≤n2^i\le n active vertices are rare and the node count at most doubles

Splitting into copies keeps a step matrix, at rate f/16f/16..

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

This item belongs to the exact correction and copy hierarchy of Section 4, a sequence of stages indexed by a level ii. At level ii the walk of interest is substochastic and only moves forward, from a vertex to a larger one, as a clocked randomized machine does on its configuration graph; its vertices form a finite set of natural numbers, and no single transition weight exceeds the rate 2−4i2^{-4i}.

From probed values at its vertices, a stage builds a correction and with it a residual matrix DD. It passes to level i+1i+1 by splitting each vertex into MM copies, where MM is one of the stage’s parameter constants, and spreading the weight of each entry of DD over copies of its endpoints, in amounts set by the probed values and by the size of the entry against the rate. This single step is the copy stage of a walk at level ii of the shape above, and the residual is seen on the copies through a weighted projection.

Stages and Their Residuals

Vertices are natural numbers, and a finite domain is a finite set of them.✦ A real vector assigns a real number to each natural number✦, and a matrix of reals assigns one to each pair of natural numbers.✦ Over a finite domain SS, for matrices of reals FF and CC and a real vector vv, the product FCFC of the two matrices and the product CvCv of a matrix with a vector are (FC)(x,y)=∑w∈SF(x,w)C(w,y),(Cv)(x)=∑w∈SC(x,w)v(w),(FC)(x,y)=\sum_{w\in S}F(x,w)\,C(w,y),\qquad (Cv)(x)=\sum_{w\in S}C(x,w)\,v(w),(1) and the row weight of CC at xx over SS is ∑w∈SC(x,w)\sum_{w\in S}C(x,w).✦

A stochastic step matrix on a finite domain SS at rate ff is a matrix of reals CC such that, for all natural numbers xx and yy, 0≤C(x,y)≤f,∑w∈SC(x,w)≤1,C(x,y)>0⇒x,y∈Sandx<y;0\le C(x,y)\le f,\qquad \sum_{w\in S}C(x,w)\le 1,\qquad C(x,y)>0\ \Longrightarrow\ x,y\in S\ \text{and}\ x<y;(2)✦ we write 𝒮(S,f)\mathcal{S}(S,f) for the set of them. More generally, for finite domains SS and S′S', a rate ff and a real number kk, a bounded support matrix CC is a matrix of reals with entries in [0,kf][0,kf], row weights over SS at most kk, and positive entries C(x,y)C(x,y) only for x<yx<y in S′S'; the stochastic step matrices are the bounded support matrices with S′=SS'=S and k=1k=1.✦

A stage PP consists of parameter constants, a level ii, a finite domain SS of vertices and a matrix of reals CC of transition weights✦; its rate is f=2−4if=2^{-4i}, the retention rate of level ii.✦ An iteration count TT is admissible when it is positive and (5/A)T/T!≤1/1024(5/A)^T/T!\le 1/1024✦, where A=1/4096A=1/4096 is the cutoff scale.✦ The parameter constants are natural numbers TT, KK, MM and JJ, with TT admissible and 128(1+32(T+1)/η(δ))≤K+1,16(T+1)<M,64M(T+1)2≤δ(J+1),128\big(1+32(T+1)/\eta(\delta)\big)\le K+1,\qquad 16(T+1)<M,\qquad 64M(T+1)^2\le\delta(J+1),(3)✦ where δ=A/400\delta=A/400✦ and η(z)=z/1000\eta(z)=z/1000✦.

A stage is considered at a field size nn, a prime, the modulus of the group (ℤ/n)4(\mathbb{Z}/n)^4 on which the channels of the hierarchy act. A pilot pair is a pair (V,E)(V,E) of a real vector VV, the probed values, and a matrix of reals EE, the correction.✦ The update map of a stage PP is the map from pilot pairs to pilot pairs that gives, from the probed values and the correction of one round, those of the next. The pilot sequence of PP at field size nn is the sequence of pilot pairs that starts from the zero vector and the zero matrix, each later term being the image of the one before under the update map of PP; the probed vector VkV_k of PP after kk rounds is the vector of the term of index kk of the pilot sequence.

The clamp of a real number uu to [0,1][0,1] is c(u)=min⁡(1,max⁡(0,u))c(u)=\min(1,\max(0,u))✦, and the smooth fade is the real function φ(x)=12(c(2x)4+1−c(2−2x)4),\varphi(x)=\tfrac12\big(c(2x)^4+1-c(2-2x)^4\big),(4)✦ a continuously differentiable ramp that vanishes for x≤0x\le0 and equals 11 for x≥1x\ge1. The smooth rising ramp from aa to bb is ρa,b(x)=φ((x−a)/(b−a))\rho_{a,b}(x)=\varphi\big((x-a)/(b-a)\big); for a<ba<b it vanishes up to aa and equals 11 from bb on.✦ The cutoff γ\gamma, the boost β\beta and the slice σ\sigma are the functions γ=1−ρ1/2,4/5,β=ρ1/4,1/2,σ=ρ1/128,1/64.\gamma=1-\rho_{1/2,\,4/5},\qquad \beta=\rho_{1/4,\,1/2},\qquad \sigma=\rho_{1/128,\,1/64}.(5)

The entrywise correction at scale ff with weights gg is the operation on matrices of reals that replaces each entry of a matrix by a value determined by that entry, the scale ff and the weights gg. For a finite domain SS, a scale ff, a real vector gg and a matrix of reals CC, the corrected iteration (Fk)k≥0(F_k)_{k\ge0} is the sequence of matrices of reals with F0=0F_0=0 and Fk+1F_{k+1} obtained from the combined matrix C+FkCC+F_kC by the entrywise correction at scale ff with weights gg.✦ The stage correction matrix FTF_T of a stage PP is the term of index TT of the corrected iteration for the domain SS, the scale ff, the matrix CC and the weights g(x)=γ(VT(x))g(x)=\gamma(V_T(x)) of PP, and the corrected residual DD of PP is D=C+FTC−FT,D=C+F_TC-F_T,(6) the products taken over SS.

The Copy Stage

Let PP be a stage at field size nn, with level ii, domain SS, matrix CC, rate ff and constants TT and MM. The copies of a vertex xx are the numbers z+Mxz+Mx with z<Mz<M, the copy with z=0z=0 being its primary copy✦, and the copy domain of SS is S+={z+Mx:x∈S,z<M}.S^+=\{\,z+Mx : x\in S,\ z<M\,\}.(7)✦ The boost factor of copy zz of xx is r0(x)=1r_0(x)=1 on the primary copy and rz(x)=β(VT(x))r_z(x)=\beta(V_T(x)) for z≠0z\neq0.✦ The share of copy zz at a real number tt is s0(t)=1−σ(t)+σ(t)/M,sz(t)=σ(t)/M(z≠0).s_0(t)=1-\sigma(t)+\sigma(t)/M,\qquad s_z(t)=\sigma(t)/M\quad(z\neq0).(8)✦

The copy matrix of the stage is the matrix of reals C+C^+ with C+(z+Mx,z′+My)=rz(x)D(x,y)sz′(D(x,y)/f)(x,ynatural,z,z′<M),C^+(z+Mx,\,z'+My)=r_z(x)\,D(x,y)\,s_{z'}\big(D(x,y)/f\big)\qquad (x,y\ \text{natural},\ z,z'<M),(9)✦ and the copy stage P+P^+ has the same constants, level i+1i+1, domain S+S^+ and matrix C+C^+.✦ The weighted projection Π\Pi takes a real vector vv to the real vector (Πv)(z+Mx)=rz(x)v(x)(xnatural,z<M).(\Pi v)(z+Mx)=r_z(x)\,v(x)\qquad(x\ \text{natural},\ z<M).(10)✦

The active set Act(P)\mathrm{Act}(P) of the stage is the set of vertices x∈Sx\in S whose probed value VT(x)V_T(x) exceeds 1/41/4.✦ The node set NS(C)N_S(C) of a matrix of reals CC on a finite domain SS is the set of vertices x∈Sx\in S with C(x,y)>0C(x,y)>0 or C(y,x)>0C(y,x)>0 for some y∈Sy\in S.✦

The Claim

Theorem 1. Let PP be a stage at field size nn, with level ii, domain SS, matrix CC, rate f=2−4if=2^{-4i} and constants TT and MM, and let DD, S+S^+, C+C^+ and Π\Pi be its corrected residual, copy domain, copy matrix and weighted projection. Suppose that CC is a stochastic step matrix on SS at rate ff. Then C+∈𝒮(S+,2−4(i+1)),C^+\in\mathcal{S}\big(S^+,\,2^{-4(i+1)}\big),(11)✦ and every entry of C+C^+ satisfies C+(u,v)≤2−4(i+1)for all natural numbers u,v.C^+(u,v)\le 2^{-4(i+1)}\qquad\text{for all natural numbers } u,v.(12)✦ For natural numbers xx, yy and j≠0j\neq0 with sj(D(x,y)/f)≠0s_j\big(D(x,y)/f\big)\neq0, the vertex yy satisfies VT(y)>12andrj(y)=1.V_T(y)>\tfrac12\qquad\text{and}\qquad r_j(y)=1.(13)✦ If moreover 2i≤n2^i\le n, then M|Act(P)|≤|NS(C)|M\,|\mathrm{Act}(P)|\le|N_S(C)|(14)✦ and |NS+(C+)|≤2|NS(C)|.|N_{S^+}(C^+)|\le 2\,|N_S(C)|.(15)✦ For every real vector vv, with the product by C+C^+ taken over S+S^+ and the product by DD over SS, C+(Πv)=Π(Dv).C^+(\Pi v)=\Pi(Dv).(16)✦

In the paper, the result reads: for J≥2Cπ(M−1)T(T+1)/δ0J\ge 2C_\pi(M-1)T(T+1)/\delta_0 and the paper’s R*R_*, the copy matrix C+C^+ is strictly forward and substochastic, C+(z+Mx,z′+My)≤f/qC^+(z+Mx,\,z'+My)\le f/q, |NS+(C+)|≤2|NS(C)||N_{S^+}(C^+)|\le 2|N_S(C)|, and C+Π=ΠDC^+\Pi=\Pi D for the weighted projection Π(z+Mx,x)=rz(x)\Pi(z+Mx,\,x)=r_z(x), where MM, TT, NN and Π\Pi stand for the paper’s D0D_0, k*k_*, AA and PrP^r, the copy z+Mxz+Mx for its xzx_z, and CπC_\pi, δ0\delta_0 and qq are the paper’s constants.

What Is Checked

The claim is stated over a stage at field size nn with its parameter constants, the stochastic step matrices (2) at the rates 2−4i2^{-4i} and 2−4(i+1)2^{-4(i+1)}, the probed vector VTV_T of the pilot sequence, the corrected residual (6) with its correction FTF_T, the copy domain (7), the copy matrix (9) with its boost factors and shares, the weighted projection (10), the active set and the node sets.