OctoInformBeta

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

Proposition 2.2: Three-atom targets and a quantitative lower bound for the distance between slab maps

The one-third exponent is sharp on cubes.

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

For a uniform source on a fixed cube, the stability question compares the L2L^2 distance between quadratic optimal maps with the Wasserstein distance between their targets. The targets considered here have three atoms, with one site moving in a direction transverse to the other two.

Let E=ℝdE=\mathbb R^d, with coordinates indexed by 0,…,d−10,\ldots,d-1 and coordinate vectors eie_i.✦ The cube KK, also used as the target container YY, and its uniform measure ρ\rho are defined in (1), where ℒd\mathcal L^d is Lebesgue measure: K=Y={x∈E:|xi|≤1 for every i},ρ=ℒd↾Kℒd(K).K=Y=\{x\in E: |x_i|\le1\text{ for every }i\}, \qquad \rho=\frac{\mathcal L^d\!\restriction_K}{\mathcal L^d(K)}.(1) A measure μ\mu on EE is supported in a set YY when μ(Yc)=0\mu(Y^c)=0.✦ A three-atom probability measure μ\mu on EE supported in YY is a measure of the form (2), with pairwise distinct sites yk∈Yy_k\in Y: μ=∑k=13wkδyk,wk>0,∑k=13wk=1.\mu=\sum_{k=1}^3 w_k\delta_{y_k}, \qquad w_k>0, \qquad \sum_{k=1}^3 w_k=1.(2)✦

A coupling π∈Π(μ,ν)\pi\in\Pi(\mu,\nu) of measures μ,ν\mu,\nu on EE is a probability measure on E×EE\times E with first marginal μ\mu and second marginal ν\nu.✦ The graph transport plan πT\pi_T of a map T:E→ET:E\to E from ρ\rho, and the quadratic transport cost 𝒞(π)\mathcal C(\pi) of a measure on E×EE\times E, are specified by (3): πT=(id,T)#ρ,𝒞(π)=∫E×E∥x−y∥2dπ(x,y)∈[0,∞].\pi_T=(\mathrm{id},T)_\#\rho, \qquad \mathcal C(\pi)=\int_{E\times E}\|x-y\|^2\,d\pi(x,y)\in[0,\infty].(3) 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 larger than any coupling in Π(ρ,μ)\Pi(\rho,\mu).✦ A quadratic optimal map TT from ρ\rho to μ\mu is unique when every coupling in Π(ρ,μ)\Pi(\rho,\mu) of cost at most 𝒞(πT)\mathcal C(\pi_T) equals its graph transport plan πT\pi_T.✦

For targets μ,ν\mu,\nu and maps T,S:E→ET,S:E\to E, the quadratic Wasserstein distance W2W_2 between the targets and the L2(ρ)L^2(\rho) distance between the maps are given by (4), where real\operatorname{real} takes the real value of a finite extended nonnegative number and takes the value zero at infinity: W2(μ,ν)=real⁡(infπ∈Π(μ,ν)𝒞(π)),∥T−S∥L2(ρ)=(∫E∥T(x)−S(x)∥2dρ(x))1/2.W_2(\mu,\nu) =\sqrt{\operatorname{real}\!\left( \inf_{\pi\in\Pi(\mu,\nu)}\mathcal C(\pi)\right)}, \qquad \|T-S\|_{L^2(\rho)} =\left(\int_E\|T(x)-S(x)\|^2\,d\rho(x)\right)^{1/2}.(4)

The Slab Family

For coordinate indices i,ji,j and real parameters a,ba,b, the central transport cell Pcij(a,b)P_c^{ij}(a,b) and the outer slope selection Oi:E→EO_i:E\to E are defined by (5): Pcij(a,b)={x∈E:|xi|≤a+bxj},Oi(x)={−ei,xi<0,ei,xi≥0.P_c^{ij}(a,b)=\{x\in E:|x_i|\le a+bx_j\}, \qquad O_i(x)= \begin{cases} -e_i,&x_i<0,\\ e_i,&x_i\ge0. \end{cases}(5) The slab map Sa,bij:E→ES_{a,b}^{ij}:E\to E selects the slope of a maximizing affine branch of the convex potential in (6), with a fixed selection at ties.✦ The slopes corresponding to the branches in that order are −ei,ei,bej-e_i,e_i,be_j: ua,bij(x)=max⁡{−xi,xi,a+bxj}.u_{a,b}^{ij}(x)=\max\{-x_i,\ x_i,\ a+bx_j\}.(6)

In dimension d≥2d\ge2, the first coordinate is 00 and the transverse coordinate is 11. The target sites yk(b)y_k(b), target masses wk(a)w_k(a) and target measure ηa,b\eta_{a,b} are the families in (7), indexed by k∈{−,+,c}k\in\{-,+,c\}; here r+=max⁡{r,0}r_+=\max\{r,0\}: (y−(b),y+(b),yc(b))=(−e0,e0,be1),(w−(a),w+(a),wc(a))=((1−a2)+,(1−a2)+,a+),ηa,b=∑k∈{−,+,c}wk(a)δyk(b).\begin{gathered} (y_-(b),y_+(b),y_c(b))=(-e_0,e_0,be_1),\\ (w_-(a),w_+(a),w_c(a)) =\left(\left(\frac{1-a}{2}\right)_+, \left(\frac{1-a}{2}\right)_+,a_+\right), \qquad \eta_{a,b}=\sum_{k\in\{-,+,c\}}w_k(a)\delta_{y_k(b)}. \end{gathered}(7) The sharpness specialization is b=a/2b=a/2.

The Claim

Theorem 1. Let d≥2d\ge2, with KK and ρ\rho as in (1). For any distinct coordinate indices i,ji,j and parameters a≥0a\ge0, b>0b>0 with a+b/2≤1a+b/2\le1, the slab maps satisfy the lower bound (8): b16≤∥Sa,0ij−Sa,bij∥L2(ρ).\sqrt{\frac b{16}} \le \|S_{a,0}^{ij}-S_{a,b}^{ij}\|_{L^2(\rho)}.(8)✦ For 0<a<1/20<a<1/2, there is a sharpness certificate at aa: distinct three-atom probability measures μa,νa\mu_a,\nu_a supported in KK, with unique quadratic optimal maps Tμa,TνaT_{\mu_a},T_{\nu_a} from ρ\rho, satisfying (9): 0<W2(μa,νa)≤a(a/2)2,a32≤∥Tμa−Tνa∥L2(ρ).0<W_2(\mu_a,\nu_a)\le\sqrt{a(a/2)^2}, \qquad \sqrt{\frac a{32}} \le\|T_{\mu_a}-T_{\nu_a}\|_{L^2(\rho)}.(9)✦ For every real α>1/3\alpha>1/3 and C>0C>0, there are distinct three-atom probability measures μ,ν\mu,\nu supported in KK, with unique quadratic optimal maps Tμ,TνT_\mu,T_\nu from ρ\rho, for which (10) holds: 0<W2(μ,ν),CW2(μ,ν)α<∥Tμ−Tν∥L2(ρ).0<W_2(\mu,\nu), \qquad C\,W_2(\mu,\nu)^\alpha <\|T_\mu-T_\nu\|_{L^2(\rho)}.(10)✦

What Is Checked

The claims are stated over the normalized Lebesgue source on KK, with arbitrary distinct coordinate directions for the slab bound and three-atom targets with unique quadratic optimal graph plans for sharpness. For d≥2d\ge2 and 0<a<1/20<a<1/2, the cube family with transverse slopes zero and a/2a/2 supplies an explicit sharpness certificate at aa.