OctoInformBeta

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

Lemma 4.1 [Fixed point and factorial truncation]: If CC is a step matrix at scale ff and 0≤g≤10 \le g \le 1, then ET+1−ET≤f(5/A)T/T!E_{T+1}-E_T \le f(5/A)^T/T!, and some TT has (5/A)T/T!≤1/1024(5/A)^T/T! \le 1/1024

Corrected iterates increase in factorially small steps.

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

Consider a random walk on finitely many nodes, numbered by natural numbers, that only moves forward: from a node xx it steps to a node y>xy > x with probability C(x,y)C(x,y) or stops, so each row of CC sums to at most 11, as for a walk on the configuration graph of a clocked machine. The entries of CC are at most a scale f>0f > 0. A correction matrix is built from CC by iterating a fixed correction map, starting from the zero matrix. The question is how far one round of this iteration moves an entry, and after how many rounds, independent of CC, that movement falls to a small fixed fraction of ff.

Forward Matrices and the Corrected Iteration

A domain is a finite set of nodes.✦ A matrix MM assigns a real number M(x,y)M(x,y) to each pair of nodes✦, a vector gg assigns a real number g(x)g(x) to each node✦, and matrices are compared entry by entry. For a domain SS, the row weight of MM at a node xx over SS is ∑y∈SM(x,y)\sum_{y \in S} M(x,y)✦, and the product MNMN over SS of two matrices MM and NN is the matrix with entries (MN)(x,y)=∑z∈SM(x,z)N(z,y).(MN)(x,y) = \sum_{z \in S} M(x,z)\,N(z,y).(1)✦

Given domains SS and S′S' and reals ff and kk, a matrix MM is a bounded support matrix on SS with active set S′S', scale ff and bound kk when, for all nodes xx and yy, 0≤M(x,y)≤kf,∑y′∈SM(x,y′)≤k,M(x,y)>0⟹x,y∈S′ and x<y.0 \le M(x,y) \le kf, \qquad \sum_{y' \in S} M(x,y') \le k, \qquad M(x,y) > 0 \implies x, y \in S' \text{ and } x < y.(2)✦ Write ℬS,S′(f,k)\mathcal{B}_{S,S'}(f,k) for the set of these matrices. A step matrix on SS at scale ff is a matrix in ℬS,S(f,1)\mathcal{B}_{S,S}(f,1).✦

Fix a domain SS, a scale ff, a vector gg and a matrix CC, and let Φf,g\Phi_{f,g} be the correction at scale ff with weights gg, a map on matrices that acts entry by entry. The iterated correction of CC is the sequence of matrices E0,E1,E2,…E_0, E_1, E_2, \dots given by E0=0E_0 = 0 and En+1=Φf,g(C+EnC),E_{n+1} = \Phi_{f,g}\bigl(C + E_n C\bigr),(3)✦ with the product taken over SS. The successive differences of the iterated correction are the matrices Dn=En+1−EnD_n = E_{n+1} - E_n.✦

The cutoff scale is the constant A=1/4096A = 1/4096.✦

The Claim

Theorem 1. Let SS and S′S' be domains, f>0f > 0 a scale, gg a vector with values in [0,1][0,1] and CC a matrix, and let EnE_n be the iterated correction of CC, with products over SS, and DnD_n its successive differences.

If CC is a bounded support matrix on SS with active set S′S', scale ff and bound 11, then En∈ℬS,S′(f,n)for every n≥0,E_n \in \mathcal{B}_{S,S'}(f,n) \qquad \text{for every } n \ge 0,(4)✦ and E0≤E1≤E2≤⋯.E_0 \le E_1 \le E_2 \le \cdots.(5)✦ If CC is a step matrix on SS at scale ff, then for every nn Dn+1≤5DnC,D_{n+1} \le 5\, D_n C,(6)✦ and for every TT and all nodes xx and yy DT(x,y)≤f(5/A)TT!.D_T(x,y) \le f\,\frac{(5/A)^T}{T!}.(7)✦ There is a positive integer TT, an admissible iteration count, with (5/A)TT!≤11024.\frac{(5/A)^T}{T!} \le \frac{1}{1024}.(8)✦

In the paper, the result reads: for every fixed column multiplier 0≤g≤10 \le g \le 1 there is a unique strictly forward nonnegative solution E*E_* of E*(x,y)=fQ((C+E*C)(x,y)/f)g(y)E_*(x,y) = fQ\bigl((C + E_*C)(x,y)/f\bigr)\,g(y), the iterates satisfy 0≤Ei≤E*0 \le E_i \le E_* and max⁡(E*−Ek)/f≤e1/θ1(LQ/θ1)k/k!\max (E_* - E_k)/f \le e^{1/\theta_1}(L_Q/\theta_1)^k/k!, and so a fixed k*k_* makes the right side at most q−1/32q^{-1}/32, uniformly in CC, ff, VV and gg, with θ1=A\theta_1 = A, k*k_* standing for TT, and QQ, LQL_Q, VV and qq the paper’s own quantities. The statement above bounds the step DTD_T from round TT to round T+1T+1 of the increasing iterates, where the paper bounds their distance E*−EkE_* - E_k to the fixed point.

What Is Checked

The claim is stated over domains SS and S′S', a scale f>0f > 0, a vector gg with values in [0,1][0,1], a bounded support matrix CC on SS with active set S′S', scale ff and bound 11 (a step matrix on SS in (6) and (7)), the iterated correction EnE_n of CC with its products over SS, its successive differences DnD_n, and the cutoff scale A=1/4096A = 1/4096 in the admissible iteration count.