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

Sharp One-Third Stability of Brenier Maps

For every integer d≥2d \ge 2, compact convex set K⊂RdK \subset \mathbb{R}^d with nonempty interior, uniform probability measure ρ=1K∣K∣dx\rho = \frac{1_K}{|K|} dx, and nonempty compact set Y⊂RdY \subset \mathbb{R}^d, there is a finite constant C(K,Y)C(K, Y) such that ∥Tμ−Tν∥L2(ρ)≤C(K,Y)W2(μ,ν)1/3\|T_\mu - T_\nu\|_{L^2(\rho)} \le C(K, Y) W_2(\mu, \nu)^{1/3} for all μ,ν∈Prob(Y)\mu, \nu \in \mathrm{Prob}(Y), where targets may be atomic, singular, or absolutely continuous, and the exponent 1/31/3 cannot be replaced by any larger exponent in general, even for K=Y=[−1,1]dK = Y = [-1, 1]^d and targets with exactly three atoms.

For every fixed d≥2d \ge 2, K=Y=[−1,1]dK = Y = [-1, 1]^d, ρ=2−d1Kdx\rho = 2^{-d} 1_K dx, and every α>1/3\alpha > 1/3, sup⁡μ,ν∈Prob(Y),μ≠ν∥Tμ−Tν∥L2(ρ)W2(μ,ν)α=+∞\sup_{\mu, \nu \in \mathrm{Prob}(Y), \mu \ne \nu} \frac{\|T_\mu - T_\nu\|_{L^2(\rho)}}{W_2(\mu, \nu)^\alpha} = +\infty, and the supremum remains infinite when both targets are required to have exactly three atoms.

For distinct sites y1,…,yNy_1, \dots, y_N and intercepts c1,…,cNc_1, \dots, c_N whose Laguerre cells have positive masses MiM_i, if K⊂{x:∣x∣≤R}K \subset \{x : |x| \le R\}, then every intercept direction q∈RNq \in \mathbb{R}^N satisfies ∑i=1N∣DBi[q,0]∣2Mi≤C0R2∑i=1N∣DMi[q,0]∣2Mi\sum_{i=1}^N \frac{|DB_i[q, 0]|^2}{M_i} \le C_0 R^2 \sum_{i=1}^N \frac{|DM_i[q, 0]|^2}{M_i} with C0=162C_0 = 162.

For finitely supported probabilities μ,ν\mu, \nu in B(0,L)B(0, L), there are convex LL-Lipschitz finite maxima u0,u1u_0, u_1 of affine functions such that ∇u0=Tμ\nabla u_0 = T_\mu and ∇u1=Tν\nabla u_1 = T_\nu ρ\rho-almost everywhere, and ∥u0−u1∥L2(ρ)≤(1+C0)RW2(μ,ν)\|u_0 - u_1\|_{L^2(\rho)} \le (1 + \sqrt{C_0}) R W_2(\mu, \nu) with C0=162C_0 = 162.

For any two finitely supported probabilities μ,ν\mu, \nu on YY, ∥Tμ−Tν∥L2(ρ)≤C∗W2(μ,ν)1/3\|T_\mu - T_\nu\|_{L^2(\rho)} \le C_* W_2(\mu, \nu)^{1/3}, where C∗=(12dR2(1+C0)2+28L2PK∣K∣)1/2C_* = \left(12 d R^2 (1 + \sqrt{C_0})^2 + \frac{28 L^2 P_K}{|K|}\right)^{1/2} with C0=162C_0 = 162.

On the open set of configurations with distinct sites and Mi>0M_i > 0 for every ii, the maps MiM_i and BiB_i are jointly C1C^1 in the intercepts and sites with explicit face-integral differentials, and the graph on {1,…,N}\{1, \dots, N\} with edges gij>0g_{ij} > 0 is connected.

For fixed positive weights m1,…,mNm_1, \dots, m_N with ∑imi=1\sum_i m_i = 1 and every configuration of distinct sites, there is a unique vector of intercepts cc satisfying Mi=miM_i = m_i (1≤i≤N1 \le i \le N) and ∑imici=0\sum_i m_i c_i = 0, and this normalized vector is a C1C^1 function of the sites.

For compact convex sets U,V⊂RdU, V \subset \mathbb{R}^d of positive volume, there is a coupling (X,Z)(X, Z) of the uniform probability measures on UU and VV such that Law(X+Z2)≤dx∣U∣∣V∣\mathrm{Law}\left(\frac{X+Z}{2}\right) \le \frac{dx}{\sqrt{|U||V|}}.

For convex, LL-Lipschitz functions u0,u1u_0, u_1 on KK and f=u0−u1f = u_0 - u_1, every h>0h > 0 satisfies ∥∇u0−∇u1∥L2(ρ)2≤12dh2∥f∥L2(ρ)2+28L2PK∣K∣h\|\nabla u_0 - \nabla u_1\|_{L^2(\rho)}^2 \le \frac{12d}{h^2}\|f\|_{L^2(\rho)}^2 + \frac{28 L^2 P_K}{|K|} h, where PK=∑j=1d∣projej⊥K∣d−1P_K = \sum_{j=1}^d |\mathrm{proj}_{e_j^\perp} K|_{d-1}.

For every λ∈Prob(Y)\lambda \in \mathrm{Prob}(Y), there exists a unique quadratic optimal map TλT_\lambda from ρ\rho to λ\lambda up to a ρ\rho-null set, and if Fn⊂YF_n \subset Y are finite ϵn\epsilon_n-nets with ϵn→0\epsilon_n \to 0 and Qn:Y→FnQ_n : Y \to F_n are Borel nearest-point selections, then T(Qn)#λ→TλT_{(Q_n)_\#\lambda} \to T_\lambda in L2(ρ)L^2(\rho).

Methods

Midpoint coupling of uniform probability measures on two compact convex sets via conditional-quantile (Knothe triangular) parameterization on the unit cube.

Source excerpts

“We give a direct proof by the conditional-quantile construction underlying Knothe’s triangular rearrangement [12, 4].” · Show in paper

“Given s = (s1, . . . , sd) ∈ (0, 1)d, construct X(s) recursively. For a previously constructed prefix ξ = (X1, . . . , Xk−1) ∈ Dk−1, let Xk be the quantile at level sk of the density t 7−→ pk(ξ, t) pk−1(ξ) .” · Show in paper

Linear site interpolation along an optimal coupling of two finite targets with C^1 intercepts preserving cell masses, combined with testing the graph-Laplacian derivative against an intercept direction whose mass derivative is the weighted intercept velocity.

Source excerpts

“Choose an optimal finite coupling of the targets. ... Set vi = y1 i − y0 i , yi(t) = (1 − t)y0 i + ty1 i , w = W2(µ, ν).” · Show in paper

“The range statement in Theorem 3.1, together with (5.4), permits a choice of q satisfying DMi[q, 0] = miri (1 ≤ i ≤ N). No bound on this q is needed.” · Show in paper

Coordinate-slice gradient interpolation for convex L-Lipschitz functions using one-dimensional difference quotients and orthogonal projection measures.

Source excerpts

“The argument is a coordinate-slice version of the convex-function interpolation inequality of Delalande and Mérigot [7, Proposition 4.1].” · Show in paper

“Kj,h = {x ∈ K : x + hej ∈ K}, Dj,hu(x) = u(x + hej ) − u(x) h (x ∈ Kj,h).” · Show in paper

Finite epsilon_n-net nearest-point discretization of arbitrary Borel targets in Prob(Y) combined with Arzela-Ascoli convergence of convex potentials and Fenchel-gap optimality.

Source excerpts

“If Fn ⊂ Y are finite ϵn-nets with ϵn → 0, and Qn : Y → Fn are Borel nearest-point selections, then T(Qn)#λ −→ Tλ in L2(ρ).” · Show in paper