OctoInformBeta

OpenAI preprint, September 2026 · §2. Affine labels with a prescribed tolerance

Corollary 2.2: The same affine family meets its tolerance throughout the exponent cube at all sufficiently small positive scales

Fixed certificates control active slopes uniformly.

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

For positive concentration vectors, the affine construction compares the slope of a minimizing branch with the exponent vector of the concentration. The error is measured in the supremum norm, with a tolerance that may depend on the selected slope.

Let mm be the exponent dimension, let a,b∈ℝa,b\in\mathbb R, and let E:ℝm→ℝE:\mathbb R^m\to\mathbb R be a tolerance function. The exponent cube [a,b]m[a,b]^m is the set in (1): [a,b]m={p∈ℝm:a≤pi≤b for every i}.[a,b]^m=\{p\in\mathbb R^m:a\le p_i\le b\text{ for every }i\}.(1)✦ For h∈ℝh\in\mathbb R, the power point Xh(p)X_h(p) of an exponent vector p∈ℝmp\in\mathbb R^m is the vector of coordinate powers defined by (2): Xh(p)=(hpi)i=1m.X_h(p)=(h^{p_i})_{i=1}^m.(2) The small-scale regime is 0<h<10<h<1, with h→0h\to0 from above.

A parameterized affine label L=(rL,cL)L=(r_L,c_L) in dimension mm is an affine branch with a fixed slope rL∈ℝmr_L\in\mathbb R^m and an offset cL:ℝ→ℝc_L:\mathbb R\to\mathbb R depending on the real parameter hh.✦ For a real vector x∈ℝmx\in\mathbb R^m and a real parameter hh, the value ℓL(h,x)\ell_L(h,x) of the affine label LL and the dot product are given by (3). ℓL(h,x)=rL⋅x+cL(h)=∑i=1mrL,ixi+cL(h).\ell_L(h,x)=r_L\cdot x+c_L(h) =\sum_{i=1}^m r_{L,i}x_i+c_L(h).(3) An active label LL of a finite family Λ\Lambda at (h,x)(h,x) is a minimizing branch as specified by (4), including every tied minimizer: L∈Λ,ℓL(h,x)≤ℓJ(h,x)for every J∈Λ.L\in\Lambda,\qquad \ell_L(h,x)\le\ell_J(h,x)\quad\text{for every }J\in\Lambda.(4)✦

Sequential slope approximation on [a,b]m[a,b]^m with tolerance EE is the property of a finite family Λ\Lambda of parameterized affine labels that the following condition holds: for each fixed L∈ΛL\in\Lambda, every sequence hn>0h_n>0 tending to zero and every sequence pn∈ℝmp_n\in\mathbb R^m converging to p0∈[a,b]mp_0\in[a,b]^m, with LL active at (hn,Xhn(pn))(h_n,X_{h_n}(p_n)) for every nn, obey the bound (5): ∥p0−rL∥∞<E(rL).\|p_0-r_L\|_\infty<E(r_L).(5)✦

A full affine approximation certificate 𝒞\mathcal C on [a,b]m[a,b]^m with tolerance EE is a finite nonempty family Λ\Lambda having sequential slope approximation and satisfying the slope and offset conditions (6) for each L∈ΛL\in\Lambda: rL∈[a,b]m,cL(h)=o(ha)(h→0+).r_L\in[a,b]^m,\qquad c_L(h)=o(h^a)\quad(h\to0^+).(6)✦

The Claim

Theorem 1 (Corollary 2.2). Let 𝒞\mathcal C be a full affine approximation certificate on [a,b]m[a,b]^m with tolerance EE and family Λ\Lambda. There is h0>0h_0>0 such that (7) holds for every 0<h<h00<h<h_0, every p∈[a,b]mp\in[a,b]^m, and every L∈ΛL\in\Lambda active at (h,Xh(p))(h,X_h(p)): ∥p−rL∥∞<E(rL).\|p-r_L\|_\infty<E(r_L).(7)✦

Suppose E(r)>0E(r)>0 for every r∈ℝmr\in\mathbb R^m, and let 𝒞\mathcal C be a full certificate on [−1,1]m[-1,1]^m with tolerance EE. There exist an integer n≥1n\ge1, slopes r1,…,rn∈ℝmr_1,\ldots,r_n\in\mathbb R^m, offsets cj:ℝ→ℝc_j:\mathbb R\to\mathbb R, and h0>0h_0>0 such that (8) holds for every 0<h<h00<h<h_0, every p∈[−1,1]mp\in[-1,1]^m, and every index jj attaining the minimum of rk⋅Xh(p)+ck(h)r_k\cdot X_h(p)+c_k(h) over 1≤k≤n1\le k\le n: |pi−rj,i|<E(rj)(1≤i≤m).|p_i-r_{j,i}|<E(r_j)\qquad(1\le i\le m).(8)✦

Suppose every everywhere positive tolerance on ℝm\mathbb R^m admits a full certificate on [−1,1]m[-1,1]^m. For each such tolerance EE, there exist an integer n≥1n\ge1, slopes r1,…,rn∈ℝmr_1,\ldots,r_n\in\mathbb R^m, offsets cj:ℝ→ℝc_j:\mathbb R\to\mathbb R, and h0>0h_0>0 such that (9) holds for every 0<h<h00<h<h_0, every p∈[−1,1]mp\in[-1,1]^m, and every index jj attaining the minimum of rk⋅Xh(p)+ck(h)r_k\cdot X_h(p)+c_k(h) over 1≤k≤n1\le k\le n: |pi−rj,i|<E(rj)(1≤i≤m).|p_i-r_{j,i}|<E(r_j)\qquad(1\le i\le m).(9)✦

In the paper, the result reads: for m≥1m\ge1, a<ba<b, and any positive tolerance EE on [a,b]m[a,b]^m as in Lemma 2.1, every finite nonempty family supplied by that lemma admits h0>0h_0>0, chosen after the family, such that every label LL active at Xh(p)X_h(p), for 0<h<h00<h<h_0 and p∈[a,b]mp\in[a,b]^m, satisfies ∥p−rL∥∞<E(rL)\|p-r_L\|_\infty<E(r_L).

What Is Checked

The uniform activity statement (7) concerns a fixed full certificate, while the indexed statements (8) and (9) concern, respectively, a full certificate for an everywhere positive tolerance and the existence of full certificates for every such tolerance.