OctoInformBeta

OpenAI preprint, September 2026 · §1. Introduction

§1 (setting): uniform source, target laws and quadratic optimal maps: Optimal graph plans for compactly supported laws, with Wasserstein distances and distances between maps

Uniform sources and quadratic transport maps.

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

In dimension dd, let K⊂ℝdK\subset\mathbb R^d be a compact convex body of positive Lebesgue volume, and let Y⊂ℝdY\subset\mathbb R^d be compact. The source law is normalized volume on KK, and the target laws are probability measures supported on YY. Quadratic transport compares the cost of moving this source by a map with the costs of all couplings having the same marginals.

The Definition

Definition 1. The ambient space EE is the Euclidean space in (1), with its usual norm and Borel sets: E=ℝd.E=\mathbb R^d.(1)✦

Write λd\lambda_d for Lebesgue measure on EE. Write ρ\rho for the source law in (2): ρ=λd(K)−1λd↾K.\rho=\lambda_d(K)^{-1}\lambda_d\!\restriction_K.(2)✦

A measure μ\mu on EE is supported on YY when it satisfies (3): μ(E\Y)=0.\mu(E\setminus Y)=0.(3)✦

A probability coupling π\pi of measures μ\mu and ν\nu on EE is a probability measure on E×EE\times E with marginals μ\mu and ν\nu.✦ For measures μ\mu and ν\nu on EE, the set Γ(μ,ν)\Gamma(\mu,\nu) of probability couplings is the set in (4), where p1,p2:E×E→Ep_1,p_2:E\times E\to E are the coordinate projections: Γ(μ,ν)={π:π(E×E)=1,(p1)#π=μ,(p2)#π=ν}.\Gamma(\mu,\nu) =\bigl\{\pi:\ \pi(E\times E)=1,\quad (p_1)_\#\pi=\mu,\quad (p_2)_\#\pi=\nu\bigr\}.(4)✦

The quadratic transport cost 𝒞(π)\mathcal C(\pi) of a measure π\pi on E×EE\times E is the extended nonnegative integral in (5): 𝒞(π)=∫E×E∥x−y∥2dπ(x,y)∈[0,+∞].\mathcal C(\pi)=\int_{E\times E}\|x-y\|^2\,\mathrm d\pi(x,y) \ \in[0,+\infty].(5)✦

For r∈[0,+∞]r\in[0,+\infty], the real representative rℝr_{\mathbb R} is the real number specified by rℝ=rr_{\mathbb R}=r when rr is finite and (+∞)ℝ=0(+\infty)_{\mathbb R}=0. The quadratic Wasserstein distance W2(μ,ν)W_2(\mu,\nu) between measures μ\mu and ν\nu on EE is the real-valued quantity in (6): W2(μ,ν)=(infπ∈Γ(μ,ν)𝒞(π))ℝ.W_2(\mu,\nu) =\sqrt{\left(\inf_{\pi\in\Gamma(\mu,\nu)} \mathcal C(\pi)\right)_{\mathbb R}}.(6)✦

The graph transport plan πT\pi_T of a map T:E→ET:E\to E and a measure ρ\rho on EE is the pushforward in (7): πT=(id⁡,T)#ρ.\pi_T=(\operatorname{id},T)_\#\rho.(7)✦

A quadratic optimal map Tμ:E→ET_\mu:E\to E from ρ\rho to μ\mu is a measurable map with pushforward μ\mu whose graph plan minimizes quadratic cost, as in (8): Tμis measurable,(Tμ)#ρ=μ,𝒞(πTμ)≤𝒞(π)(π∈Γ(ρ,μ)).T_\mu\ \text{is measurable},\qquad (T_\mu)_\#\rho=\mu,\qquad \mathcal C(\pi_{T_\mu})\leq\mathcal C(\pi) \quad\bigl(\pi\in\Gamma(\rho,\mu)\bigr).(8)✦

A unique quadratic optimal map Tμ:E→ET_\mu:E\to E from ρ\rho to μ\mu is a quadratic optimal map for which every coupling of no greater cost is its graph plan, as specified by (9): {π∈Γ(ρ,μ):𝒞(π)≤𝒞(πTμ)}⊆{πTμ}.\bigl\{\pi\in\Gamma(\rho,\mu): \mathcal C(\pi)\leq\mathcal C(\pi_{T_\mu})\bigr\} \subseteq\{\pi_{T_\mu}\}.(9)✦

The L2(ρ)L^2(\rho) distance between maps T,S:E→ET,S:E\to E is the real-valued quantity in (10): ∥T−S∥L2(ρ)=∫E∥T(x)−S(x)∥2dρ(x).\|T-S\|_{L^2(\rho)} =\sqrt{\int_E\|T(x)-S(x)\|^2\,\mathrm d\rho(x)}.(10)✦

What Is Checked

These definitions concern measures and maps on finite-dimensional Euclidean space, with the source measure given by normalized Lebesgue restriction to KK.