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.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

Subpolynomial dimension reduction in Lp

For every fixed 1<p<∞1 < p < \infty with p≠2p \neq 2 and every D>1D > 1, there is a constant Cp,D<∞C_{p,D} < \infty such that for every integer n≥2n \ge 2, every nn-point subset of a real LpL_p space embeds into ℓpd\ell_p^d with distortion at most DD and subpolynomial dimension dp(n,D)≤exp⁡(Cp,D(log⁡n)γ(p))d_p(n, D) \le \exp(C_{p,D}(\log n)^{\gamma(p)}), where γ(p)=2−p\gamma(p) = 2 - p for 1<p<21 < p < 2 and γ(p)=1−2/p\gamma(p) = 1 - 2/p for 2<p<∞2 < p < \infty; consequently, lim⁡n→∞log⁡dp(n,D)log⁡n=0\lim_{n\to\infty} \frac{\log d_p(n, D)}{\log n} = 0 for D>1D > 1.

For exact embeddings (D=1D = 1) with 1<p<∞1 < p < \infty and p≠2p \neq 2, ⌊n−14⌋2≤dp(n,1)≤(n2)\lfloor \frac{n-1}{4} \rfloor^2 \le d_p(n, 1) \le \binom{n}{2} for all n≥9n \ge 9, and consequently lim⁡n→∞log⁡dp(n,1)log⁡n=2\lim_{n\to\infty} \frac{\log d_p(n, 1)}{\log n} = 2.

For 1<p<∞1 < p < \infty, n≥2n \ge 2, and D≥1D \ge 1, the minimum embedding dimension satisfies dp(n,D)≥log⁡nlog⁡(1+2D)d_p(n, D) \ge \frac{\log n}{\log(1 + 2D)}.

Let 0<η<1/30 < \eta < 1/3 and b,K≥1b, K \ge 1. If for every strictly positive probability vector λ\lambda on EE there is a random normalized gradient F∈VF \in V with ∥F∥∞≤K\|F\|_\infty \le K almost surely and ∑e∈Eλe∣E∣Fe∣p−b∣≤ηb\sum_{e \in E} \lambda_e |\mathbb{E}|F_e|^p - b| \le \eta b (with b,Kb, K independent of λ\lambda), then the points admit an embedding into ℓpd\ell_p^d with distortion at most ((1+3η)/(1−3η))1/p((1+3\eta)/(1-3\eta))^{1/p} and d≤⌈Cmax⁡{1,K2p/(η2b2)}log⁡(2n)⌉d \le \lceil C \max\{1, K^{2p}/(\eta^2 b^2)\} \log(2n) \rceil.

For every strictly positive probability vector λ\lambda on EE, there is a probability vector σ\sigma with σe≥λe/2\sigma_e \ge \lambda_e / 2 such that the orthogonal projection PP onto VV in ℓ2(σ)\ell_2(\sigma) satisfies max⁡e∈E∑f∈E∣Pef∣≤H=4log⁡(2n)\max_{e \in E} \sum_{f \in E} |P_{ef}| \le H = 4\log(2n).

For 1<p<21 < p < 2, clipping and projecting the heavy-tailed random vector Y∈VY \in V yields parameters b,K≥1b, K \ge 1 independent of λ\lambda and the input points, satisfying the weighted moment condition (3.1) with log⁡K=Op,η((log⁡n)2−p)\log K = O_{p,\eta}((\log n)^{2-p}).

For 2<p<∞2 < p < \infty, overlapping magnitude truncations followed by projection, signed Poisson sampling, and event truncation yield parameters b,K≥1b, K \ge 1 independent of λ\lambda and the input points, satisfying the weighted moment condition (3.1) with log⁡K=Op,η((log⁡n)1−2/p)\log K = O_{p,\eta}((\log n)^{1-2/p}).

Methods

Exact discretization of any nn-point subset of a real LpL_p space (1≤p<∞1 \le p < \infty) into ℓp(n2)\ell_p^{\binom{n}{2}} by representing the normalized distance-power vector as a barycenter in the convex hull of normalized label-difference vectors and applying Carathéodory's theorem.

Source excerpts

“Lemma 2.1 (Exact discretization). Every n-point subset of a real Lp space embeds isometrically into ℓ_p(nchoose2)p^{(n choose 2)}. This holds for every 1 ≤ p < ∞.” · Show in paper

Reduction of simultaneous edge-moment matching to a per-weight average condition via finite-dimensional convex separation, followed by independent sampling (Maurey's empirical method) with a Hoeffding-type bounded-sum estimate and union bound.

Source excerpts

“Lemma 3.1 reduces this simultaneous requirement to a weighted-average condition for each strictly positive probability on the pairs. Crucially, the common moment b and magnitude bound K are uniform over those probabilities. Convex separation produces one distribution working for all pairs; independent sampling then supplies the coordinates of a finite embedding.” · Show in paper

Fixed-point reweighting (via Brouwer's fixed-point theorem) of the electrical-flow localization estimate of Gurel-Gurevich, Nachmias, and Sachdeva to construct an ℓ2(σ)\ell_2(\sigma)-orthogonal projection onto normalized gradients with ℓ∞\ell_\infty operator norm O(log⁡n)O(\log n).

Source excerpts

“Thus Φ is a continuous self-map of Σ, and Brouwer’s theorem [8, Corollary 2.15] provides a fixed point. At that fixed point ue(σ) ≤ σe; since σe > 0, this gives Re(σ) ≤ 2√2 log n ≤ H.” · Show in paper

Construction of random normalized gradients for 1<p<21 < p < 2 by dyadic signed Poisson sampling of homogeneous increments to form a heavy-tailed vector Y∈VY \in V, followed by coordinatewise clipping and projection.

Source excerpts

“We first construct a random vector Y ∈ V by symmetric Poisson sampling of the vector v against the intensity ν.” · Show in paper

“For T > 1, let [Y ]T denote coordinatewise clipping to [−T, T], and set F = P[Y ]T .” · Show in paper

Construction of random normalized gradients for 2<p<∞2 < p < \infty using overlapping dyadic magnitude truncations (creating a triangular multiplicity profile), projection, signed Poisson sampling with near-additive ppth moments, and event truncation.

Source excerpts

“The overlapping truncations below create a triangular multiplicity profile of height ℓ.” · Show in paper

“Discarding the entire vector on a bad event preserves membership in V , so F ∈ V and ∥F∥∞ ≤ K.” · Show in paper

Lower-bound obstruction for exact embeddings (D=1D = 1, p≠2p \neq 2) using an antipodal row-and-column configuration Xk\mathcal{X}_k in ℓpk2\ell_p^{k^2} and the pp-parallelogram defect equality criterion.

Source excerpts

“Consequently supp(f(ri)) meets supp(f(cj)) for every (i, j). A single coordinate can belong to at most one row support and at most one column support, so it can account for at most one of these k^2 intersections. It follows that d ≥ k^2.” · Show in paper