By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

All papers

Claims

Findings

Midpoint lenses in segment spaces

Let YY be either the finite-height forest dual X=J∗X = J^* or the infinite-height coordinate predual B∞B_\infty. For every finite ancestral HH, every x,y∈Yx, y \in Y with x=PHxx = P_H x, and every R≥0R \ge 0 with ∥x+y∥,∥x−y∥≤R\|x + y\|, \|x - y\| \le R, we have ∥QHy∥≤2R2−∥x∥2\|Q_H y\| \le 2\sqrt{R^2 - \|x\|^2}.

In either the finite-height forest dual X=J∗X = J^* or the infinite-height coordinate predual B∞B_\infty, the averaged asymptotic midpoint modulus satisfies δ^(t)≥1+t2/4−12\widehat{\delta}(t) \ge \frac{\sqrt{1 + t^2/4} - 1}{2} for all t>0t > 0, so both norms are asymptotically midpoint uniformly convex.

Neither the infinite-height coordinate predual B∞B_\infty nor the finite-height forest dual X=J∗X = J^* admits an equivalent asymptotically uniformly convex (AUC) norm; if a∥⋅∥≤N≤b∥⋅∥a\|\cdot\| \le N \le b\|\cdot\| with 0<a≤b<∞0 < a \le b < \infty, then the one-sided asymptotic modulus δ‾N\overline{\delta}_N vanishes at a/(2b)a/(2b) and at a/ba/b.

The finite-height forest dual X=J∗X = J^* is infinite dimensional, separable, and reflexive, and the coordinate functionals have dense span in XX.

Using the primal cumulative-clipping method, for both B∞B_\infty and XX, the averaged asymptotic midpoint modulus satisfies δ^Z(t)≥1+t2/12−1\widehat{\delta}_Z(t) \ge \sqrt{1 + t^2/12} - 1 for all t>0t > 0.

Methods

Disjoint-segment square-sum norm construction on a countable forest of finite-height trees and on the infinite countably branching tree, defining the dual X=J∗X = J^* and coordinate predual B∞=span⁡‾{ev∗:v∈T∞}⊂J∞∗B_\infty = \overline{\operatorname{span}}\{e_v^* : v \in T_\infty\} \subset J_\infty^*.

Source excerpts

“ρ(u) = sup { (∑_{i=1}^m u(S_i)^2)^{1/2} : m ≥ 1, S_1, . . . , S_m pairwise disjoint segments }.” · Show in paper

“The second model uses the infinite tree T∞ = N<ω, including its root. Let J∞ be the completion for the same disjoint-segment norm, and let B∞ = span{e∗v : v ∈ T∞} ⊂ J∗∞.” · Show in paper

Finite atomic-ball representation and joint coefficient-and-slack budget, replacing atom coefficients by head-test sums and bounding opposite-sign crossing sums at gates via splitting the longer head segment.

Source excerpts

“At a gate with opposite nonzero sums b, b′, the two heads end at the same parent v. They are therefore nested and cannot coincide. Suppose the unprimed head [a, v] is longer than the primed head [a′, v]. Split the unprimed segment immediately before a′, retaining its own endpoint in the second piece.” · Show in paper

“We will replace each atom coefficient by a multiple of the head test’s sum on that segment. This produces two nonnegative errors: the squared change of coefficients and the unused squared head-test energy. Their sum is controlled by the squared-radius deficit of the center.” · Show in paper

Bounded-path obstruction to equivalent asymptotically uniformly convex norms using weakly null child increments along unit root-path functionals.

Source excerpts

“The equivalent norms have the same continuous linear functionals, so (zj ) is weakly null, and N(zj ) ≥ ac/(bM) = t. Lemma 8.3 gives a child for which N(us⌢j ) > (1 + γ)N(us). Starting at the root and repeating k times gives am(1 + γ)k < bM, which is impossible for sufficiently large k.” · Show in paper

Cumulative clipping of primal exterior tests into two inward-signed perturbations v=h1−h2v = h_1 - h_2 with quadratic norm cost ρ(f+shj)2≤ρ(f)2+3s2ρ(v)2\rho(f + sh_j)^2 \le \rho(f)^2 + 3s^2\rho(v)^2.

Source excerpts

“Lemma 10.2 (Two cumulative clips). For every s > 0 and finite tests f = PHf, v = QHv, there are h1, h2 supported on supp v such that v = h1 − h2, ρ(hj ) ≤ ρ(v), ρ(f + shj )2 ≤ ρ(f)2 + 3s2ρ(v)2 (j = 1, 2).” · Show in paper