OctoInformBeta

OpenAI preprint, September 2026 · §6. From potentials to maps

Corollary 6.2: For a fixed compact convex source with nonempty interior and a compact target set in dimension at least two

Uniform constants in one-third transport stability.

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 stability question compares quadratic optimal maps from a fixed uniform source when their target measures vary within a compact set. Let ℝd\mathbb R^d be Euclidean space, with dd a nonnegative integer, and let K⊂ℝdK\subset\mathbb R^d be compact and convex with nonempty interior. The uniform measure ρ\rho on the source body K⊂ℝdK\subset\mathbb R^d is the normalized restriction of Lebesgue measure in (1), where |K||K| is the Lebesgue volume of KK: ρ=1|K|ℒd↾K.\rho=\frac{1}{|K|}\,\mathcal L^d\!\restriction_K.(1) A measure μ\mu on ℝd\mathbb R^d is supported in a set YY when μ(ℝd\Y)=0\mu(\mathbb R^d\setminus Y)=0.✦

A coupling π\pi of measures α,β\alpha,\beta on ℝd\mathbb R^d is a probability measure on ℝd×ℝd\mathbb R^d\times\mathbb R^d with first marginal α\alpha and second marginal β\beta; write Π(α,β)\Pi(\alpha,\beta) for the set of these couplings.✦ The quadratic transport cost 𝒬(π)\mathcal Q(\pi) of a measure π\pi on ℝd×ℝd\mathbb R^d\times\mathbb R^d is the nonnegative extended real integral in (2): 𝒬(π)=∫ℝd×ℝd|x−y|2dπ(x,y).\mathcal Q(\pi)=\int_{\mathbb R^d\times\mathbb R^d}|x-y|^2\,d\pi(x,y).(2)✦ The quadratic Wasserstein distance W2(α,β)W_2(\alpha,\beta) between measures α,β\alpha,\beta on ℝd\mathbb R^d is the square root of the real value of the infimum of 𝒬(π)\mathcal Q(\pi) over π∈Π(α,β)\pi\in\Pi(\alpha,\beta).✦

The graph transport plan γT\gamma_T of a map T:ℝd→ℝdT:\mathbb R^d\to\mathbb R^d from ρ\rho is the pushforward in (3): γT=(id⁡,T)#ρ.\gamma_T=(\operatorname{id},T)_\#\rho.(3)✦ A quadratic optimal transport map TT from ρ\rho to μ\mu is a measurable map with T#ρ=μT_\#\rho=\mu whose graph plan minimizes 𝒬\mathcal Q over Π(ρ,μ)\Pi(\rho,\mu).✦ A unique quadratic optimal transport map TT from ρ\rho to μ\mu is a quadratic optimal map such that every π∈Π(ρ,μ)\pi\in\Pi(\rho,\mu) with 𝒬(π)≤𝒬(γT)\mathcal Q(\pi)\le\mathcal Q(\gamma_T) equals γT\gamma_T.✦

The L2(ρ)L^2(\rho) distance between maps T,S:ℝd→ℝdT,S:\mathbb R^d\to\mathbb R^d is the square root of the integral in (4): ∥T−S∥L2(ρ)=(∫ℝd∥T(x)−S(x)∥2dρ(x))1/2.\|T-S\|_{L^2(\rho)} =\left(\int_{\mathbb R^d}\|T(x)-S(x)\|^2\,d\rho(x)\right)^{1/2}.(4)✦

A subgradient selection TT of a function u:ℝd→ℝu:\mathbb R^d\to\mathbb R is a map T:ℝd→ℝdT:\mathbb R^d\to\mathbb R^d satisfying (5): u(x)+⟨z−x,T(x)⟩≤u(z)(x,z∈ℝd).u(x)+\langle z-x,T(x)\rangle\le u(z) \qquad(x,z\in\mathbb R^d).(5)✦

The Claim

Theorem 1. For every L,P∈ℝL,P\in\mathbb R, there is a constant Cint≥0C_{\mathrm{int}}\ge0 with the following property. Let u,v:ℝd→ℝu,v:\mathbb R^d\to\mathbb R be continuous, and let T,ST,S be measurable subgradient selections of u,vu,v, respectively, with ∥T(x)∥,∥S(x)∥≤L\|T(x)\|,\|S(x)\|\le L for every x∈ℝdx\in\mathbb R^d. If w≥0w\ge0 and some c∈ℝc\in\mathbb R satisfies ∫(u−v−c)2dρ≤P2w2\int (u-v-c)^2\,d\rho\le P^2w^2, then the normalized interpolation estimate is ∥T−S∥L2(ρ)≤Cintw1/3.\|T-S\|_{L^2(\rho)}\le C_{\mathrm{int}}\,w^{1/3}.(6)✦

Let D,w,A,BD,w,A,B be nonnegative real numbers. If D2≤Aw2/h2+BhD^2\le Aw^2/h^2+Bh for every h>0h>0, then the slice stability estimate is D≤A+Bw1/3.D\le\sqrt{A+B}\,w^{1/3}.(7)✦

Suppose d≥2d\ge2 and Y⊂ℝdY\subset\mathbb R^d is compact and nonempty. There is a constant C≥0C\ge0, depending only on d,K,Yd,K,Y, such that every pair of probability measures μ,ν\mu,\nu supported in YY admits unique quadratic optimal transport maps Tμ,TνT_\mu,T_\nu from ρ\rho satisfying ∥Tμ−Tν∥L2(ρ)≤CW2(μ,ν)1/3.\|T_\mu-T_\nu\|_{L^2(\rho)}\le C\,W_2(\mu,\nu)^{1/3}.(8)✦

In the paper, the result reads: for the global source and target setting, with R,L≥1R,L\ge1 bounding the norms on K,YK,Y, respectively, PKP_K the perimeter of KK, and C0=162C_0=162, every pair of finitely supported probabilities μ,ν\mu,\nu on YY satisfies (9): C*=(12dR2(1+C0)2+28L2PK|K|)1/2,∥Tμ−Tν∥L2(ρ)≤C*W2(μ,ν)1/3.\begin{aligned} C_*&=\left(12dR^2(1+\sqrt{C_0})^2 +28L^2\frac{P_K}{|K|}\right)^{1/2},\\ \|T_\mu-T_\nu\|_{L^2(\rho)}&\le C_*\,W_2(\mu,\nu)^{1/3}. \end{aligned}(9)

What Is Checked

The assertions (6) and (7) concern continuous potentials with globally bounded measurable subgradient selections and a scalar inequality at every positive slice width, respectively, while (8) is stated for all probability measures supported in a fixed compact YY and their unique quadratic optimal maps from ρ\rho.