OctoInformBeta

OpenAI preprint, September 2026 · §2. Sharpness on a fixed cube

Lemma 2.1: Strict maxima give derivatives, and supporting potentials give uniqueness for quadratic transport

Measurable affine maximizers and unique optimal transport.

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 finite maximum of affine functions supplies candidates for a transport map through its slopes. The question is which measurable choices of these slopes give the unique quadratic optimal transport plan.

Let E=ℝdE=\mathbb{R}^d with its Euclidean inner product.✦ For a source set K⊂EK\subset E, the uniform measure ρ\rho is the restriction of Lebesgue measure to KK, multiplied by the reciprocal of its volume: ρ=(ℒd(K))−1ℒd↾K.\rho=\bigl(\mathcal{L}^d(K)\bigr)^{-1} \mathcal{L}^d\!\restriction K.(1)✦

Let VV be a real inner product space, let II be a finite nonempty index set, and choose sites yi∈Vy_i\in V and intercepts ci∈ℝc_i\in\mathbb{R}. The finite affine potential u:V→ℝu:V\to\mathbb{R} associated with the sites yiy_i and intercepts cic_i is the function given by their maximum; for x∈Vx\in V, the active set A(x)⊂IA(x)\subset I consists of the indices whose affine functions attain this maximum: u(x)=maxi∈I(⟨x,yi⟩+ci),A(x)={i∈I:⟨x,yi⟩+ci=u(x)}.u(x)=\max_{i\in I}\bigl(\langle x,y_i\rangle+c_i\bigr), \qquad A(x)=\{i\in I:\langle x,y_i\rangle+c_i=u(x)\}.(2) An index i∈Ii\in I is a strict maximizer in (2) at a point x∈Vx\in V when ⟨x,yj⟩+cj<⟨x,yi⟩+ci\langle x,y_j\rangle+c_j<\langle x,y_i\rangle+c_i for every j≠ij\ne i.

Quadratic Transport

A coupling π\pi of measures ρ\rho and μ\mu on EE is a probability measure on E×EE\times E with first marginal ρ\rho and second marginal μ\mu; write Π(ρ,μ)\Pi(\rho,\mu) for these couplings.✦ The graph plan γT\gamma_T of a map T:E→ET:E\to E is the pushforward of ρ\rho by x↦(x,T(x))x\mapsto(x,T(x))✦, and the quadratic transport cost 𝒞(π)\mathcal{C}(\pi) is the extended nonnegative integral of squared distance: γT=(id,T)#ρ,𝒞(π)=∫E×E∥x−z∥2dπ(x,z).\gamma_T=(\mathrm{id},T)_\#\rho, \qquad \mathcal{C}(\pi)=\int_{E\times E}\|x-z\|^2\,d\pi(x,z).(3) A quadratic optimal map T:E→ET:E\to E from ρ\rho to μ\mu is a measurable map satisfying T#ρ=μ,𝒞(γT)≤𝒞(π)for every π∈Π(ρ,μ).T_\#\rho=\mu, \qquad \mathcal{C}(\gamma_T)\leq\mathcal{C}(\pi) \quad\text{for every }\pi\in\Pi(\rho,\mu).(4)✦ A unique quadratic optimal map T:E→ET:E\to E from ρ\rho to μ\mu is a quadratic optimal map in the sense of (4) for which every π∈Π(ρ,μ)\pi\in\Pi(\rho,\mu) with 𝒞(π)≤𝒞(γT)\mathcal{C}(\pi)\leq\mathcal{C}(\gamma_T) equals the graph plan γT\gamma_T in (3).✦ A measure μ\mu is supported in a target container Y⊂EY\subset E when μ(Yc)=0\mu(Y^c)=0.✦

The Claim

Theorem 1. For sites in EE and every source set K⊂EK\subset E, the slopes of (2) satisfy (5): #{yi:i∈A(x)}=1for ρ-almost every x.\#\{y_i:i\in A(x)\}=1 \qquad\text{for }\rho\text{-almost every }x.(5)✦ For VV equipped with its Borel σ\sigma-algebra, a selection exists as in (6): ∃T:V→VBorel measurable,T(x)∈{yi:i∈A(x)}(x∈V).\exists\,T:V\to V\ \text{Borel measurable},\qquad T(x)\in\{y_i:i\in A(x)\}\quad(x\in V).(6)✦ At a strict maximizer ii at x∈Vx\in V, the potential is Fréchet differentiable, with derivative (7): Du(x)[h]=⟨h,yi⟩(h∈V).Du(x)[h]=\langle h,y_i\rangle\qquad(h\in V).(7)✦ Suppose the sites belong to EE and are pairwise distinct, KK is compact, and ρ\rho is a probability measure. Let T:E→ET:E\to E be measurable and satisfy T(x)∈{yi:i∈A(x)}T(x)\in\{y_i:i\in A(x)\} for ρ\rho-almost every xx. Then (8) holds: T is a unique quadratic optimal map from ρ to T#ρ.T\text{ is a unique quadratic optimal map from }\rho\text{ to }T_\#\rho.(8)✦ Let KK be nonempty, compact and convex, let Y⊂EY\subset E be compact, and let ρ\rho and μ\mu be finite measures, with μ\mu supported in YY. Let u:E→ℝu:E\to\mathbb{R} be Lipschitz on KK and let T:E→ET:E\to E be measurable with T#ρ=μT_\#\rho=\mu. Suppose that, for ρ\rho-almost every xx, u(x)+⟨z−x,T(x)⟩≤u(z)u(x)+\langle z-x,T(x)\rangle\leq u(z) for every z∈Kz\in K. Then the supporting potential gives (9): T is a unique quadratic optimal map from ρ to μ.T\text{ is a unique quadratic optimal map from }\rho\text{ to }\mu.(9)✦

In the paper, the result reads: for any positive integer NN and finite maximum u(x)=max⁡1≤i≤N(x⋅yi+ci)u(x)=\max_{1\leq i\leq N}(x\cdot y_i+c_i) in the general source setting, with repeated slopes allowed, all maximizing indices have the same slope for ρ\rho-almost every xx, and choosing the smallest maximizing index gives a Borel map TT with T=∇uT=\nabla u almost everywhere whose graph plan is the unique quadratic optimal plan from ρ\rho to T#ρT_\#\rho.

What Is Checked

The claims are stated over finite affine families in real inner product spaces and, for transport, over the uniform measure (1) with the compactness and measure hypotheses of (8) and (9).