OctoInformBeta

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

§6 (setting): bounded supporting maps and coordinate slices: Translated source overlaps enter the estimate through the measure of the lost region

Directional bounds for globally supporting 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

The passage from potentials to transport maps involves the directional components of supporting maps on a source region and its translates. Write FF for the ambient real inner product space, with inner product denoted by a dot, and μ\mu for the measure in the integrals.

Definition 1. A supporting map T:F→FT:F\to F for a potential u:F→ℝu:F\to\mathbb R is a subgradient selection satisfying the global supporting inequality (1): u(z)≥u(x)+T(x)⋅(z−x)(x,z∈F).u(z)\geq u(x)+T(x)\cdot(z-x) \qquad (x,z\in F).(1)✦

The directional slope sT(e,x)s_T(e,x) of a map T:F→FT:F\to F at x∈Fx\in F in a direction e∈Fe\in F is the component defined by (2): sT(e,x)=e⋅T(x).s_T(e,x)=e\cdot T(x).(2)✦ For a source region K⊂FK\subset F, a direction e∈Fe\in F and a step h>0h>0, the translated overlap Ke,hK_{e,h} is the subset of KK defined in (3): Ke,h={x∈K:x+he∈K}.K_{e,h}=\{x\in K:x+he\in K\}.(3)

The Claim

Theorem 1. Let u0,u1:F→ℝu_0,u_1:F\to\mathbb R have measurable supporting maps T0,T1:F→FT_0,T_1:F\to F, and let L∈ℝL\in\mathbb R satisfy ∥Ti(x)∥≤L\|T_i(x)\|\leq L for every x∈Fx\in F and i∈{0,1}i\in\{0,1\}. Let K⊂FK\subset F be measurable with μ(K)<∞\mu(K)<\infty, let h>0h>0, and let e∈Fe\in F satisfy ∥e∥≤1\|e\|\leq1. Suppose that f=u0−u1f=u_0-u_1 is measurable on FF and that f2f^2 is integrable over KK. Then the directional interpolation bound is (4): ∫K(sT0(e,x)−sT1(e,x))2dμ(x)≤12h2∫Kf(x)2dμ(x)+28L2μ(K\Ke,h).\int_K\bigl(s_{T_0}(e,x)-s_{T_1}(e,x)\bigr)^2\,d\mu(x) \leq \frac{12}{h^2}\int_K f(x)^2\,d\mu(x) +28L^2\mu(K\setminus K_{e,h}).(4)✦

In the paper, the result reads: in dimension dd with coordinate unit vectors eje_j, for convex LL-Lipschitz potentials u0,u1u_0,u_1 on KK, set f=u0−u1f=u_0-u_1, Aj=|proj⁡ej⟂K|d−1A_j=|\operatorname{proj}_{e_j^\perp}K|_{d-1} and PK=∑jAjP_K=\sum_j A_j, use Kj,h={x∈K:x+hej∈K}K_{j,h}=\{x\in K:x+he_j\in K\} for finite differences, and identify the relevant gradients through the supporting inequality u(z)≥u(x)+T(x)⋅(z−x)u(z)\geq u(x)+T(x)\cdot(z-x).

What Is Checked

The estimate (4) is stated for measurable, globally bounded supporting maps as in (1) and their directional slopes (2), integrated over a measurable source region of finite measure.