OctoInformBeta

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

Lemma 4.2 [Elementary pilot bounds]: After kk steps, pilot and actual corrections have entries in [0,kf][0,kf] and row sums at most kk; probed values only grow

Corrections of a bounded forward matrix stay bounded and forward.

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

The copy hierarchy works with a finite set SS of nodes and a matrix CC of transition weights between them. The case of interest is a matrix that is nonnegative and strictly forward, each positive weight leading from a node to a larger one, so that the weighted graph on SS has no cycles, and whose rows sum to at most 11, as in a substochastic transition matrix. The exact correction builds from CC a correction matrix step by step, in a pilot run steered by the probed values of a detector, a score computed at every node from random affine channels, and in an actual run that reuses the last probed values.

The question is what the corrections keep of the shape of CC after kk steps, and how the actual corrections compare with the pilot ones.

Matrices and Corrections

Nodes are natural numbers, and a node set is a finite set of them.✦ A real vector gg is a function assigning a real number g(x)g(x) to each node xx, and a real matrix DD is a function assigning a real entry D(x,y)D(x,y) to each pair of nodes xx and yy. The product over a node set SS of real matrices DD and FF is the real matrix DFDF with (DF)(x,y)=∑z∈SD(x,z)F(z,y)(DF)(x,y) = \sum_{z\in S} D(x,z)\,F(z,y), and the row weight of DD at a node xx over SS is ∑y∈SD(x,y)\sum_{y\in S} D(x,y).✦

Write κ(t)=min⁡(1,max⁡(0,t))\kappa(t) = \min(1,\max(0,t)) for the clamp of a real number tt to [0,1][0,1]✦, and A=1/4096A = 1/4096 for the cutoff scale.✦ The threshold profile QQ is the shoulder function with threshold and gap AA, the real function Q(t)=Aκ(t−AA)4+max⁡(0,t−2A),Q(t) = A\,\kappa\Bigl(\frac{t-A}{A}\Bigr)^{4} + \max(0,\,t-2A),(1)✦ which vanishes for t≤θ1=At \le \theta_1 = A and equals t−At-A for t≥θ2=2At \ge \theta_2 = 2A.

The entrywise correction at scale ff of a real matrix GG, with column weights from a real vector gg, is the real matrix Φf,g(G)\Phi_{f,g}(G) built from the profile (1): 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).(2)✦ For a node set SS, a scale ff, a real vector gg and a real matrix CC, the iterated correction of CC over SS is the sequence of real matrices FkF_k with F0=0,Fk+1=Φf,g(C+FkC),F_0 = 0, \qquad F_{k+1} = \Phi_{f,g}\bigl(C + F_k C\bigr),(3) where C+FkCC + F_kC, with the product over SS, is the combined matrix of CC and FkF_k.

For node sets SS and S′S' and reals ff and kk, a real matrix DD is a bounded support matrix with bound kk on S′S', over SS at scale ff, written D∈ℬk(S,S′,f)D \in \mathcal{B}_k(S,S',f), when for all nodes xx and yy 0≤D(x,y)≤kf,∑z∈SD(x,z)≤k,D(x,y)>0only ifx,y∈S′andx<y.0 \le D(x,y) \le kf, \qquad \sum_{z\in S} D(x,z) \le k, \qquad D(x,y) > 0 \ \text{only if}\ x, y \in S' \ \text{and}\ x < y.(4)✦

The Detector

Fix a prime nn and write Vn=(ℤ/n)4V_n = (\mathbb{Z}/n)^4.✦ A channel over nn is a pair s=(a,b)s = (a,b) of a nonzero vector aa of VnV_n and a vector bb of VnV_n✦; its point at a residue yy modulo nn is s(y)=b+yas(y) = b + ya, the point at parameter yy of an affine line.✦ The rate of a level ll is fl=2−4lf_l = 2^{-4l}✦, and the box BlB_l of level ll is the set of vectors of VnV_n whose four coordinates, read as integers in {0,…,n−1}\{0,\dots,n-1\}, are smaller than the side ml=⌊(n−1)/2l⌋+1m_l = \lfloor (n-1)/2^l \rfloor + 1.✦ Given a node set SS, the retained vertices of a channel ss at level ll form the subset Rl(s)R_l(s) of SS of the nodes xx with s(xmod⁡n)∈Bls(x \bmod n) \in B_l.✦

For a finite set RR, a real function hh on RR and a natural number KK, the padded rank of hh over RR is rk⁡K(R,h)=max⁡({0}∪{h(x):x∈R,#{x′∈R:h(x′)≥h(x)}>K}).\operatorname{rk}_K(R,h) = \max\Bigl(\{0\} \cup \bigl\{h(x) : x \in R,\ \#\{x' \in R : h(x') \ge h(x)\} > K\bigr\}\Bigr).(5)✦

The fade is the real function ϕ(t)=12(κ(2t)4+1−κ(2−2t)4)\phi(t) = \tfrac12\bigl(\kappa(2t)^4 + 1 - \kappa(2-2t)^4\bigr), equal to 00 for t≤0t \le 0 and to 11 for t≥1t \ge 1✦, and the ramp from aa to bb is ra,b(v)=ϕ((v−a)/(b−a))r_{a,b}(v) = \phi\bigl((v-a)/(b-a)\bigr).✦ With δ=A/400\delta = A/400 and η(z)=z/1000\eta(z) = z/1000, the cutoff is γ=1−r1/2,4/5\gamma = 1 - r_{1/2,\,4/5}, equal to 11 up to 1/21/2 and to 00 from 4/54/5 on✦; the light ramp is λ=rδ,2δ\lambda = r_{\delta,\,2\delta}✦, and the gate at a real zz is gz=1−rη(z),2η(z)g_z = 1 - r_{\eta(z),\,2\eta(z)}.✦

Let GG be a real matrix, ll a level and KK, JJ natural numbers. The mark μt(x,y)\mu_t(x,y) of an entry (x,y)(x,y) of GG at a channel tt is the light ramp of the scaled entry times the throttle of row xx, the gate at δ\delta of the padded rank (5) of that row over Rl(t)R_l(t): μt(x,y)=λ(G(x,y)fl)gδ(rk⁡K(Rl(t),z↦G(x,z)fl)).\mu_t(x,y) = \lambda\Bigl(\frac{G(x,y)}{f_l}\Bigr)\, g_\delta\Bigl(\operatorname{rk}_K\Bigl(R_l(t),\ z \mapsto \frac{G(x,z)}{f_l}\Bigr)\Bigr).(6) The score sc⁡y(s,t)\operatorname{sc}_y(s,t) of a pair of channels (s,t)(s,t) at a node yy is the padded rank (5) of the marks of column yy over the retained vertices of ss: sc⁡y(s,t)=rk⁡J(Rl(s),x↦μt(x,y)).\operatorname{sc}_y(s,t) = \operatorname{rk}_J\bigl(R_l(s),\ x \mapsto \mu_t(x,y)\bigr).(7) The averaged score of GG at a node yy is the mean of the score over all channels ss, averaged again over the channels tt whose point at yy lies in the box of level ll, with ss and tt uniform over the channels over nn: UG(y)=𝔼t[𝔼ssc⁡y(s,t)|t(ymod⁡n)∈Bl].U_G(y) = \mathbb{E}_{t}\Bigl[\,\mathbb{E}_{s}\,\operatorname{sc}_y(s,t) \;\mathrel{\Big|}\; t(y \bmod n) \in B_l\Bigr].(8)✦

The Pilot Run

A stage PP consists of parameter constants, a level ii, a node set SS and a real matrix CC of transition weights✦; its rate is f=fif = f_i.✦ The parameter constants are four natural numbers: an iteration count TT that is admissible, which means T>0T > 0 and (5/A)T/T!≤1/1024(5/A)^T/T! \le 1/1024, a number KK with 128(1+32(T+1)/η(δ))≤K+1128\,\bigl(1 + 32(T+1)/\eta(\delta)\bigr) \le K+1, a number of copies M>16(T+1)M > 16(T+1), and a number JJ with 64M(T+1)2≤δ(J+1)64M(T+1)^2 \le \delta(J+1).✦

A pilot pair (V,Eˆ)(V,\hat E) is a real vector VV together with a real matrix Eˆ\hat E.✦ Over a prime nn, the update of a pilot pair (V,Eˆ)(V,\hat E) is the pilot pair 𝒰(V,Eˆ)=(V′,Eˆ′)\mathcal{U}(V,\hat E) = (V',\hat E') with G=C+EˆC,V′(x)=max⁡(V(x),UG(x)),Eˆ′=Φf,γ∘V′(G),G = C + \hat E C, \qquad V'(x) = \max\bigl(V(x),\,U_G(x)\bigr), \qquad \hat E' = \Phi_{f,\,\gamma\circ V'}(G),(9)✦ where the product is over SS and UGU_G is the averaged score (8) at the level ii, with the stage’s KK and JJ.

The pilot run is the sequence of pilot pairs (Vk,Eˆk)(V_k,\hat E_k) that starts at the zero vector and the zero matrix and moves by the update: (V0,Eˆ0)=(0,0),(Vk+1,Eˆk+1)=𝒰(Vk,Eˆk).(V_0,\hat E_0) = (0,0), \qquad (V_{k+1},\hat E_{k+1}) = \mathcal{U}(V_k,\hat E_k).(10)✦ The probed values of the pilot run are the vectors VkV_k, its pilot corrections are the matrices Eˆk\hat E_k, and its pilot inputs are the test matrices Gk=C+EˆkCG_k = C + \hat E_k C.

The actual corrections EkE_k are the iterated correction (3) of CC over SS at scale ff with the column weights g(y)=γ(VT(y))g(y) = \gamma(V_T(y)), the cutoff of the probed values at the last step TT; the matrices C+EkCC + E_kC are the actual inputs.✦

The Claim

Theorem 1. Let PP be a stage with iteration count TT, node set SS, matrix CC and rate ff, and let nn be a prime. For steps k≤k′k \le k' the probed values satisfy Vk(x)≤Vk′(x)for every node x.V_k(x) \le V_{k'}(x) \quad \text{for every node } x.(11)✦ Let moreover S′S' be a node set such that C∈ℬ1(S,S′,f)C \in \mathcal{B}_1(S,S',f). Then for every step kk the pilot correction, the pilot input and the actual correction satisfy Eˆk∈ℬk(S,S′,f),\hat E_k \in \mathcal{B}_k(S,S',f),(12)✦ Gk∈ℬk+1(S,S′,f),G_k \in \mathcal{B}_{k+1}(S,S',f),(13)✦ Ek∈ℬk(S,S′,f),E_k \in \mathcal{B}_k(S,S',f),(14)✦ and for every step k≤Tk \le T the actual correction lies below the pilot correction: Ek(x,y)≤Eˆk(x,y)for all nodes x and y.E_k(x,y) \le \hat E_k(x,y) \quad \text{for all nodes } x \text{ and } y.(15)✦

In the paper, the result reads: all pilot and actual inputs and corrections are nonnegative and strictly forward; for 1≤k≤T1 \le k \le T their row sums are at most kk and their entries at most kfkf; Ek−1≤Ek≤EˆkE_{k-1} \le E_k \le \hat E_k and Gk′≤GkG'_k \le G_k; and their row and column supports lie in R(C)R(C) and T(C)T(C), with kk and TT standing for the paper’s ii and k*k_*, and Gk′G'_k, R(C)R(C) and T(C)T(C) being the paper’s own symbols for the actual input and for the row and column index sets of CC.

What Is Checked

The claim is stated over a stage, with its parameter constants, level ii, rate ff, node set SS and matrix CC, a prime nn and a node set S′S'; over the pilot run (10) with its update (9), probed values, pilot corrections and pilot inputs; over the actual corrections, the iterated correction (3) with weights γ(VT(y))\gamma(V_T(y)); and over the bounded support matrices (4).