OctoInformBeta

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

§4 (setting): stage inputs, the cutoff Q and the correction iteration: Strictly forward matrices at rate ff, a cutoff QQ between t−2At-2A and max⁡(0,t−A)\max(0,t-A), the corrections EiE_i and the residual DD

Stages and their corrections through a shoulder cutoff.

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

Number the configurations of a clocked randomized machine so that every step goes from a smaller number to a larger one, and weight each transition by its probability. The weights form a strictly forward substochastic matrix CC. The total weight of the forward paths from xx to yy is the entry E(x,y)E(x,y) of the finite sum E=C+C2+C3+⋯E = C + C^2 + C^3 + \cdots, and EE solves E=C+ECE = C + EC.

The exact correction and copy hierarchy starts from such a matrix on a finite set of nodes, with entries at most a rate ff. In place of EE it iterates the equation E=C+ECE = C + EC with a cutoff QQ at scale A=1/4096A = 1/4096 applied entry by entry, and a residual measures how far a matrix is from solving that equation.

Nodes and Matrices

A finite domain SS is a finite set of natural numbers, called nodes.✦ A real vector gg is a function assigning a real number g(y)g(y) to each node yy.✦ A matrix CC is a function assigning a real number C(x,y)C(x,y) to each pair of nodes, its entry in row xx and column yy.✦

The row weight of a matrix CC at a node xx over a finite domain SS is the sum of the entries of row xx in the columns of SS, wS(C,x)=∑y∈SC(x,y).w_S(C,x) = \sum_{y \in S} C(x,y).(1)✦

The product over a finite domain SS of matrices FF and CC is the matrix FCFC with entries (FC)(x,y)=∑z∈SF(x,z)C(z,y),(FC)(x,y) = \sum_{z \in S} F(x,z)\,C(z,y),(2)✦ and the combined matrix of CC and FF over SS is C+FCC + FC.✦

The Cutoff

The clamp of a real number uu to [0,1][0,1] is σ(u)=min⁡(1,max⁡(0,u))\sigma(u) = \min(1, \max(0,u)).✦ For reals aa and ww, the shoulder function with corner aa and width ww is the real function sh⁡a,w(t)=wσ(t−aw)4+max⁡(0,t−a−w).\operatorname{sh}_{a,w}(t) = w\,\sigma\Bigl(\frac{t-a}{w}\Bigr)^{4} + \max(0,\, t-a-w).(3)✦ The cutoff scale is the constant A=1/4096A = 1/4096.✦

Parameter Constants

The constant δ\delta is A/400A/400✦, and η\eta is the map z↦z/1000z \mapsto z/1000 on the reals.✦ An iteration count TT is admissible when it is a positive integer with (5/A)T/T!≤1/1024(5/A)^T/T! \le 1/1024.✦ The parameter constants p=(T,K,M,J)p = (T, K, M, J) are four natural numbers, an admissible iteration count TT, a number of copies MM and two parameters KK and JJ, such that 128(1+32(T+1)η(δ))≤K+1,16(T+1)<M,64M(T+1)2≤δ(J+1).128\Bigl(1 + \frac{32(T+1)}{\eta(\delta)}\Bigr) \le K + 1, \qquad 16(T+1) < M, \qquad 64\,M\,(T+1)^2 \le \delta\,(J+1).(4)✦

The Claim

Definition 1. For finite domains SS and S′S', the second called the active set, and reals ff and kk, a bounded support matrix for (S,S′,f,k)(S, S', f, k) is a matrix CC such that for all nodes xx and yy 0≤C(x,y)≤kf,wS(C,x)≤k,C(x,y)>0⟹x,y∈S′ and x<y;0 \le C(x,y) \le kf, \qquad w_S(C,x) \le k, \qquad C(x,y) > 0 \implies x, y \in S' \text{ and } x < y;(5)✦ we write ℬ(S,S′,f,k)\mathcal{B}(S, S', f, k) for the set of them. A step matrix on SS at rate ff is a bounded support matrix with active set SS and bound k=1k = 1: 𝒫(S,f)=ℬ(S,S,f,1).\mathcal{P}(S, f) = \mathcal{B}(S, S, f, 1).(6)✦ A stage is a quadruple (p,i,S,C)(p,\, i,\, S,\, C)(7)✦ of parameter constants pp, a level i∈ℕi \in \mathbb{N}, a finite domain SS and a matrix CC. The cutoff is the shoulder function with corner and width AA, Q(t)=sh⁡A,A(t)=Aσ(tA−1)4+max⁡(0,t−2A),Q(t) = \operatorname{sh}_{A,A}(t) = A\,\sigma\Bigl(\frac{t}{A} - 1\Bigr)^{4} + \max(0,\, t - 2A),(8)✦ and for every real tt it satisfies Q(t)≤max⁡(0,t−A),Q(t) \le \max(0,\, t - A),(9)✦ t−2A≤Q(t).t - 2A \le Q(t).(10)✦ For a rate ff and a vector gg, the correction of a matrix GG is the matrix Φf,g(G)(x,y)=fQ(G(x,y)/f)g(y).\Phi_{f,g}(G)(x,y) = f\,Q\bigl(G(x,y)/f\bigr)\,g(y).(11)✦ For a finite domain SS, a rate ff, a vector gg and a matrix CC, the iterated corrections are the matrices E0,E1,E2,…E_0, E_1, E_2, \dots given by E0=0,Ei=Φf,g(Gi′),Gi′=C+Ei−1C(i≥1).E_0 = 0, \qquad E_i = \Phi_{f,g}(G'_i), \quad G'_i = C + E_{i-1}C \quad (i \ge 1).(12)✦ The residual over SS of a matrix FF against CC is the matrix D=C−F+FC.D = C - F + FC.(13)✦

In the paper, the result reads: a stage has a finite ordered vertex set VV, a rate 0<f≤10 < f \le 1 and a nonnegative strictly forward matrix CC with C𝟏≤𝟏C\mathbf{1} \le \mathbf{1} and C(x,y)≤fC(x,y) \le f; QQ is a nondecreasing Lipschitz cutoff with max⁡(0,t−θ2)≤Q(t)≤max⁡(0,t−θ1)\max(0, t-\theta_2) \le Q(t) \le \max(0, t-\theta_1), its θ1\theta_1 and θ2\theta_2 being AA and 2A2A here; and E0=0E_0 = 0, Gi′=C+Ei−1CG'_i = C + E_{i-1}C, Ei(x,y)=fQ(Gi′(x,y)/f)g(y)E_i(x,y) = f\,Q(G'_i(x,y)/f)\,g(y) and D=C−E+ECD = C - E + EC. The paper’s conditions on CC are those of a step matrix in 𝒫(V,f)\mathcal{P}(V, f).

What Is Checked

The definitions and the two bounds on QQ are stated over finite domains of nodes, real vectors and matrices on the nodes, the row weight (1), the product and the combined matrix over a finite domain (2), the clamp σ\sigma and the shoulder function (3), the constant AA, and the parameter constants (4) with δ\delta, η\eta and the admissible iteration counts.