OctoInformBeta

OpenAI preprint, September 2026 · §5. Moving sites with fixed masses

Proposition 5.1: Derivative energy bounds give endpoint potentials and an atomic criterion for one-third stability

Fixed masses control differences of potentials.

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

Moving the atoms of a target measure at fixed masses changes the intercepts of its finite affine potential. The estimates concern potential differences along such paths and the resulting stability of optimal transport maps.

Let K⊂ℝdK\subset\mathbb R^d be compact and convex with nonempty interior. The uniform measure ρ\rho on KK is normalized Lebesgue measure, as in (1); let R≥0R\geq0 bound ‖x‖\lVert x\rVert for ρ\rho-almost every xx.✦ ρ=ℒd↾Kℒd(K).\rho=\frac{\mathcal L^d\!\restriction K}{\mathcal L^d(K)}.(1)✦

For a finite nonempty index set II, sites yi∈ℝdy_i\in\mathbb R^d and intercepts ci∈ℝc_i\in\mathbb R, the finite affine potential uy,cu_{y,c} is the function in (2). The closed Laguerre cell PiP_i is the subset of KK where the iith affine function attains the maximum, including ties; its normalized cell mass MiM_i is its ρ\rho-mass, as in (3). uy,c(x)=maxi∈I(x⋅yi+ci).u_{y,c}(x)=\max_{i\in I}(x\cdot y_i+c_i).(2)✦ Pi={x∈K:x⋅yj+cj≤x⋅yi+ci for every j∈I},Mi=ρ(Pi).P_i=\{x\in K:x\cdot y_j+c_j\leq x\cdot y_i+c_i\text{ for every }j\in I\}, \qquad M_i=\rho(P_i).(3)

For weights mim_i, the weighted intercept mean ℓm:ℝI→ℝ\ell_m:\mathbb R^I\to\mathbb R is the linear functional in (4); its zero-mean hyperplane HmH_m is the subspace on which it vanishes. For positive masses with ∑imi=1\sum_i m_i=1, the canonical normalized intercept vector c*(y)c^*(y) is a vector in HmH_m assigning the prescribed masses at distinct sites and equal to zero when sites repeat, as in (5). ℓm(c)=∑imici,Hm={c∈ℝI:ℓm(c)=0}.\ell_m(c)=\sum_i m_i c_i, \qquad H_m=\{c\in\mathbb R^I:\ell_m(c)=0\}.(4) c*(y)∈Hm,{Mi(K,y,c*(y))=mi(i∈I),if the sites are distinct,c*(y)=0,if sites repeat.c^*(y)\in H_m,\qquad \begin{cases} M_i(K,y,c^*(y))=m_i\quad(i\in I),&\text{if the sites are distinct},\\ c^*(y)=0,&\text{if sites repeat}. \end{cases}(5)

A supporting pair (u,T)(u,T) on ℝd\mathbb R^d consists of a function u:ℝd→ℝu:\mathbb R^d\to\mathbb R and a subgradient selection T:ℝd→ℝdT:\mathbb R^d\to\mathbb R^d satisfying (6). For L≥0L\geq0, a supporting pair (u,T)(u,T) on ℝd\mathbb R^d is LL-bounded when TT is measurable and ‖T(x)‖≤L\lVert T(x)\rVert\leq L everywhere. u(x)+(z−x)⋅T(x)≤u(z)(x,z∈ℝd).u(x)+(z-x)\cdot T(x)\leq u(z) \qquad(x,z\in\mathbb R^d).(6)✦

A coupling π\pi of probabilities α,β\alpha,\beta on ℝd\mathbb R^d is a probability measure on ℝd×ℝd\mathbb R^d\times\mathbb R^d with these two marginals; write Π(α,β)\Pi(\alpha,\beta) for their couplings.✦ The graph transport plan ΓT\Gamma_T is the pushforward measure in (7), and 𝒬\mathcal Q is the quadratic cost functional of a plan. For measures α,β\alpha,\beta on ℝd\mathbb R^d, the quadratic Wasserstein distance W2(α,β)W_2(\alpha,\beta) is the square root of the real value of the infimum of their attainable quadratic coupling costs. ΓT=(id,T)#ρ,𝒬(π)=∫‖x−z‖2dπ(x,z),W2(α,β)=(infπ∈Π(α,β)𝒬(π))1/2.\Gamma_T=(\mathrm{id},T)_\#\rho,\qquad \mathcal Q(\pi)=\int\lVert x-z\rVert^2\,d\pi(x,z),\qquad W_2(\alpha,\beta)= \left(\inf_{\pi\in\Pi(\alpha,\beta)}\mathcal Q(\pi)\right)^{1/2}.(7)

For a probability μ\mu on ℝd\mathbb R^d, a quadratic optimal map TT from ρ\rho to μ\mu is a measurable map with T#ρ=μT_\#\rho=\mu whose graph plan minimizes quadratic cost among couplings.✦ Uniqueness of a quadratic optimal map TT from ρ\rho to μ\mu means that every coupling of no greater cost equals ΓT\Gamma_T; denote such a map by TμT_\mu.✦ A measure μ\mu on ℝd\mathbb R^d is carried by a set Y⊂ℝdY\subset\mathbb R^d when it assigns zero mass to YcY^c.✦ For maps T,S:ℝd→ℝdT,S:\mathbb R^d\to\mathbb R^d, the map distance ‖T−S‖L2(ρ)\lVert T-S\rVert_{L^2(\rho)} is the square root of ∫‖T(x)−S(x)‖2dρ(x)\int\lVert T(x)-S(x)\rVert^2\,d\rho(x).✦

For site and intercept velocities vi(t),ri(t)v_i(t),r_i(t), the energy bounds with parameters w,κ≥0w,\kappa\geq0 mean (8) at every 0<t<10<t<1. ∑imi‖vi(t)‖2≤w2,∑imiri(t)2≤κR2w2.\sum_i m_i\lVert v_i(t)\rVert^2\leq w^2, \qquad \sum_i m_i r_i(t)^2\leq\kappa R^2w^2.(8)

The Claim

Theorem 1. Let ziz_i be a second family of sites in ℝd\mathbb R^d. Suppose mi>0m_i>0, ∑imi=1\sum_i m_i=1, and the linear site path yi(t)=(1−t)yi+tziy_i(t)=(1-t)y_i+tz_i has distinct sites for 0<t<10<t<1. Assume its canonical intercepts c*(y(t))c^*(y(t)) are differentiable there and satisfy (8) with vi=zi−yiv_i=z_i-y_i and ri(t)=ddtci*(y(t))r_i(t)=\frac{d}{dt}c_i^*(y(t)). For Ut=uy(t),c*(y(t))U_t=u_{y(t),c^*(y(t))}, the interior estimate (9) holds whenever 0<a<b<10<a<b<1: ‖Ub−Ua‖L2(ρ)≤2(1+κ)Rw(b−a).\lVert U_b-U_a\rVert_{L^2(\rho)} \leq\sqrt{2(1+\kappa)}\,Rw(b-a).(9)✦

For the endpoint assertion, KK may be any set whose uniform measure ρ\rho is finite and satisfies the same almost everywhere norm bound. Let mi>0m_i>0. Suppose the site paths yi:ℝ→ℝdy_i:\mathbb R\to\mathbb R^d are continuous, with distinct sites on (0,1)(0,1), and let ci:ℝ→ℝc_i:\mathbb R\to\mathbb R be intercept paths. Assume yi′=viy_i'=v_i, ci′=ric_i'=r_i on (0,1)(0,1), where vi,riv_i,r_i are continuous on (0,1)(0,1). Assume (8) and ∑imici(t)=0\sum_i m_i c_i(t)=0 on (0,1)(0,1). Suppose a measurable selector A:ℝ×ℝd→IA:\mathbb R\times\mathbb R^d\to I attains the affine maximum for every xx at each interior time, and ρ{x:A(t,x)=i}=mi\rho\{x:A(t,x)=i\}=m_i there. There exist intercepts c‾i\bar c_i whose potentials u‾t=uy(t),c‾(t)\bar u_t=u_{y(t),\bar c(t)} satisfy (10): c‾i∈C(ℝ),c‾i=ci on (0,1),c‾i′=ri on (0,1)(i∈I),∑imic‾i(t)=0(0≤t≤1),‖u‾1−u‾0‖L2(ρ)≤2(1+κ)Rw,∫(u‾1−u‾0)2dρ≤2(1+κ)R2w2.\begin{gathered} \bar c_i\in C(\mathbb R),\qquad \bar c_i=c_i\text{ on }(0,1),\qquad \bar c_i'=r_i\text{ on }(0,1)\quad(i\in I),\\ \sum_i m_i\bar c_i(t)=0\quad(0\leq t\leq1),\\ \lVert\bar u_1-\bar u_0\rVert_{L^2(\rho)} \leq\sqrt{2(1+\kappa)}\,Rw,\qquad \int(\bar u_1-\bar u_0)^2\,d\rho \leq2(1+\kappa)R^2w^2. \end{gathered}(10)✦

Let KK again be the compact convex source body. Retain the endpoint path hypotheses, replacing the selector assumptions by measurability of each cic_i and Mi(K,y(t),c(t))=miM_i(K,y(t),c(t))=m_i for 0<t<10<t<1. Assume also ∑imi=1\sum_i m_i=1 and ‖yi(0)‖,‖yi(1)‖≤L\lVert y_i(0)\rVert,\lVert y_i(1)\rVert\leq L. There are LL-bounded supporting pairs (u0,T0),(u1,T1)(u_0,T_0),(u_1,T_1) with the endpoint properties in (11): (T0)#ρ=∑imiδyi(0),(T1)#ρ=∑imiδyi(1),∫(u0−u1)2dρ≤2(1+κ)R2w2.(T_0)_\#\rho=\sum_i m_i\delta_{y_i(0)},\qquad (T_1)_\#\rho=\sum_i m_i\delta_{y_i(1)},\qquad \int(u_0-u_1)^2\,d\rho\leq2(1+\kappa)R^2w^2.(11)✦

Let Y⊂ℝdY\subset\mathbb R^d be compact and nonempty, with ‖z‖≤L\lVert z\rVert\leq L on YY, and let P∈ℝP\in\mathbb R. Suppose every finite positive mass vector of total mass one and every pair of site families y,zy,z satisfying ‖yi‖,‖zi‖≤L\lVert y_i\rVert,\lVert z_i\rVert\leq L, whose linear interpolation has distinct sites at every interior time, admit LL-bounded supporting pairs (u,T),(v,S)(u,T),(v,S) sending ρ\rho to ∑imiδyi\sum_i m_i\delta_{y_i} and ∑imiδzi\sum_i m_i\delta_{z_i}, respectively, and a real constant hh satisfying (12): ∫(u−v−h)2dρ≤P2∑imi‖yi−zi‖2.\int(u-v-h)^2\,d\rho \leq P^2\sum_i m_i\lVert y_i-z_i\rVert^2.(12) There is C≥0C\geq0 such that every pair of probabilities μ,ν\mu,\nu carried by YY has unique quadratic optimal maps from ρ\rho obeying (13): ‖Tμ−Tν‖L2(ρ)≤CW2(μ,ν)1/3.\lVert T_\mu-T_\nu\rVert_{L^2(\rho)} \leq C\,W_2(\mu,\nu)^{1/3}.(13)✦

In the paper, the result reads: for the uniform source on K⊂B(0,R)K\subset B(0,R) and any finitely supported probabilities μ,ν\mu,\nu in B(0,L)B(0,L), there are convex LL-Lipschitz finite maxima u0,u1u_0,u_1 of affine functions, with ∇u0=Tμ\nabla u_0=T_\mu and ∇u1=Tν\nabla u_1=T_\nu ρ\rho-almost everywhere, such that ‖u0−u1‖L2(ρ)≤(1+162)RW2(μ,ν)\lVert u_0-u_1\rVert_{L^2(\rho)}\leq(1+\sqrt{162})R\,W_2(\mu,\nu).

What Is Checked

The estimates (9), (10) and (11) are stated over finite moving-site paths and normalized Lebesgue measure, while (13) quantifies over all probabilities carried by the compact target container YY. For the source body KK and distinct sites, the normalized intercept vector in (5) assigning prescribed positive masses of total mass one exists uniquely.✦