OctoInformBeta

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

Proposition 6.3: Quantized limits under potential and stability hypotheses, with the same 1/31/3 bound beyond finite supports

Optimal maps for compact target laws.

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

A compactly supported target law need not have finitely many atoms. The question is whether finite target transport maps have a limit in the distance between maps, and whether their stability estimate persists for the limiting laws.

Work in Euclidean space ℝd\mathbb R^d, with coordinates indexed by 0,…,d−10,\ldots,d-1. For a compact source set K⊂ℝdK\subset\mathbb R^d, its uniform measure ρ\rho is the restriction of Lebesgue measure to KK, normalized as in (1): ρ=ℒd(K)−1ℒd↾K.\rho=\mathcal L^d(K)^{-1}\,\mathcal L^d\!\restriction_K.(1)✦ A measure μ\mu is supported in a target set Y⊂ℝdY\subset\mathbb R^d when μ(Yc)=0\mu(Y^c)=0.✦

A coupling π\pi of measures α\alpha and β\beta is a probability measure on ℝd×ℝd\mathbb R^d\times\mathbb R^d with marginals α\alpha and β\beta.✦ The graph transport plan γT\gamma_T of a map TT with source measure ρ\rho is the pushforward in (2).✦ The quadratic transport cost 𝒞(π)\mathcal C(\pi) of a measure π\pi on the Cartesian square is the nonnegative extended integral in (2): γT=(id,T)#ρ,𝒞(π)=∫ℝd×ℝd|x−y|2dπ(x,y).\gamma_T=(\mathrm{id},T)_\#\rho, \qquad \mathcal C(\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\alpha and β\beta is the square root of the real value of the infimum of 𝒞(π)\mathcal C(\pi) over couplings of α\alpha and β\beta.

A quadratic optimal map TT from ρ\rho to μ\mu is a measurable map with T#ρ=μT_\#\rho=\mu whose graph plan has cost no greater than that of any coupling of ρ\rho and μ\mu.✦ A unique quadratic optimal map TT from ρ\rho to μ\mu is a quadratic optimal map for which every coupling of ρ\rho and μ\mu with cost at most 𝒞(γT)\mathcal C(\gamma_T) equals γT\gamma_T.✦ For maps T,S:ℝd→ℝdT,S\colon\mathbb R^d\to\mathbb R^d, the distance between them is the L2(ρ)L^2(\rho) distance in (3): ∥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}.(3)✦

Quantization Data

For the limit statement, let KK be compact and convex, fix x0∈Kx_0\in K, and let μ\mu be a probability measure supported in a compact set YY. Take measurable maps Qn,Tn:ℝd→ℝdQ_n,T_n\colon\mathbb R^d\to\mathbb R^d satisfying (4): μn=(Qn)#μ,(Tn)#ρ=μn,Tn∈L2(ρ),Tn(x)∈Yfor ρ-almost every x.\mu_n=(Q_n)_\#\mu, \qquad (T_n)_\#\rho=\mu_n, \qquad T_n\in L^2(\rho), \qquad T_n(x)\in Y\quad\text{for $\rho$-almost every $x$}.(4) The displacement tolerances εn\varepsilon_n satisfy (5): εn≥0,εn→0,|Qn(y)−y|≤εnfor μ-almost every y.\varepsilon_n\geq0, \qquad \varepsilon_n\longrightarrow0, \qquad |Q_n(y)-y|\leq\varepsilon_n \quad\text{for $\mu$-almost every $y$}.(5) For a common constant L≥0L\geq0, let un:ℝd→ℝu_n\colon\mathbb R^d\to\mathbb R be convex and LL-Lipschitz on KK, with the normalization and subgradient condition (6): un(x0)=0,un(x)+⟨Tn(x),z−x⟩≤un(z)for ρ-almost every x and every z∈K.u_n(x_0)=0, \qquad u_n(x)+\langle T_n(x),z-x\rangle\leq u_n(z) \quad\text{for $\rho$-almost every $x$ and every $z\in K$}.(6) Assume also the pairwise stability bound (7) for a constant C≥0C\geq0: ∥Tm−Tn∥L2(ρ)≤CW2(μm,μn)1/3(m,n∈ℕ).\|T_m-T_n\|_{L^2(\rho)} \leq C\,W_2(\mu_m,\mu_n)^{1/3} \qquad(m,n\in\mathbb N).(7)

The Claim

Theorem 1. For a compact convex source KK with nonempty interior and a nonempty compact target set YY, every probability measure μ\mu supported in YY admits a map as in (8): Tμ:ℝd→ℝdis a unique quadratic optimal map from ρ to μ.T_\mu\colon\mathbb R^d\to\mathbb R^d \quad\text{is a unique quadratic optimal map from $\rho$ to $\mu$}.(8)✦

For the quantization data satisfying (4), (5), (6) and (7), there is a map satisfying (9): Tμ is a unique quadratic optimal map from ρ to μ,Tμ∈L2(ρ),Tμ(x)∈Yfor ρ-almost every x,limn→∞∥Tn−Tμ∥L2(ρ)=0.\begin{gathered} T_\mu\text{ is a unique quadratic optimal map from $\rho$ to $\mu$}, \qquad T_\mu\in L^2(\rho),\\ T_\mu(x)\in Y\quad\text{for $\rho$-almost every $x$}, \qquad \lim_{n\to\infty}\|T_n-T_\mu\|_{L^2(\rho)}=0. \end{gathered}(9)✦

Let KK be compact with nonempty interior, and let μ\mu be a probability measure supported on a finite set F⊂YF\subset Y. Suppose |y|≤L|y|\leq L for every y∈Yy\in Y, where L≥0L\geq0, and fix x0∈ℝdx_0\in\mathbb R^d. There are a potential uu and a map TμT_\mu satisfying (10): Tμ is a unique quadratic optimal map from ρ to μ,u:ℝd→ℝ is convex and L-Lipschitz,u(x0)=0,Tμ(ℝd)⊂Y,u(x)+⟨Tμ(x),z−x⟩≤u(z)(x,z∈ℝd).\begin{gathered} T_\mu\text{ is a unique quadratic optimal map from $\rho$ to $\mu$},\\ u\colon\mathbb R^d\to\mathbb R\text{ is convex and $L$-Lipschitz}, \qquad u(x_0)=0,\\ T_\mu(\mathbb R^d)\subset Y, \qquad u(x)+\langle T_\mu(x),z-x\rangle\leq u(z) \quad(x,z\in\mathbb R^d). \end{gathered}(10)✦

Let K,YK,Y be compact, and let μ,ν\mu,\nu be probability measures supported in K,YK,Y, respectively. For measurable maps Q,Q′:ℝd→ℝdQ,Q'\colon\mathbb R^d\to\mathbb R^d and constants ε,η≥0\varepsilon,\eta\geq0, suppose |Q(x)−x|≤ε|Q(x)-x|\leq\varepsilon for μ\mu-almost every xx and |Q′(y)−y|≤η|Q'(y)-y|\leq\eta for ν\nu-almost every yy. Then (11) holds: W2(Q#μ,Q#′ν)≤W2(μ,ν)+ε+η.W_2(Q_\#\mu,Q'_\#\nu) \leq W_2(\mu,\nu)+\varepsilon+\eta.(11)✦

Let KK be compact and convex with nonempty interior, let YY be nonempty and compact, and fix C≥0C\geq0. Suppose every pair of probability measures supported on finite subsets of YY admits unique quadratic optimal maps satisfying the bound in (12). For every pair of probability measures μ,ν\mu,\nu supported in YY, there are unique quadratic optimal maps Tμ,TνT_\mu,T_\nu from ρ\rho to the respective laws satisfying ∥Tμ−Tν∥L2(ρ)≤CW2(μ,ν)1/3.\|T_\mu-T_\nu\|_{L^2(\rho)} \leq C\,W_2(\mu,\nu)^{1/3}.(12)✦

In the paper, the result reads: every μ∈Prob⁡(Y)\mu\in\operatorname{Prob}(Y), where Prob⁡(Y)\operatorname{Prob}(Y) denotes the probability measures supported in YY, admits a quadratic optimal map TμT_\mu from ρ\rho to μ\mu, unique up to a ρ\rho-null set, and if Fn⊂YF_n\subset Y are finite εn\varepsilon_n-nets with εn→0\varepsilon_n\to0 and Qn:Y→FnQ_n\colon Y\to F_n are Borel nearest-point selections, then T(Qn)#μ→TμT_{(Q_n)_\#\mu}\to T_\mu in L2(ρ)L^2(\rho).

What Is Checked

The claims are stated over the uniform measure (1) and target probability laws, with uniqueness referring to the graph coupling and with the limit conditional on the stated approximation hypotheses.