OctoInformBeta

OpenAI preprint, September 2026 · §4. A uniform estimate for cell moments

Lemma 4.1: For fixed sites, with intrinsic face derivatives and convex body midpoint couplings

A uniform estimate for cell moment derivatives.

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

Changing intercepts in a finite convex potential moves mass and first moment between its Laguerre cells while the sites remain fixed.

Work in Euclidean space ℝd\mathbb{R}^d with a finite nonempty index set II, a compact convex set KK, sites yi∈ℝdy_i\in\mathbb{R}^d and intercepts ci∈ℝc_i\in\mathbb{R}. The finite affine potential uu associated with the sites yiy_i and intercepts cic_i is the real-valued function on Euclidean space obtained by taking the maximum of the affine functions with those slopes and intercepts. The closed Laguerre cell Pi(y,c)P_i(y,c) of the potential uu on KK is the subset of KK where the affine function with slope yiy_i and intercept cic_i attains the maximum defining uu, including ties: u(x)=maxi∈I(x⋅yi+ci),Pi(y,c)={x∈K:x⋅yj+cj≤x⋅yi+ci for every j∈I}.u(x)=\max_{i\in I}(x\cdot y_i+c_i),\qquad P_i(y,c)=\{x\in K:x\cdot y_j+c_j\le x\cdot y_i+c_i \text{ for every }j\in I\}.(1)✦

Let λ\lambda denote Lebesgue measure on ℝd\mathbb{R}^d. For λ(K)>0\lambda(K)>0, the uniform measure ρ\rho on KK is the normalized restriction of Lebesgue measure to KK.✦ The normalized mass Mi(y,c)M_i(y,c) and first moment Bi(y,c)B_i(y,c) of the cell (1) are ρ=λ(K)−1λ|K,Mi(y,c)=ρ(Pi(y,c)),Bi(y,c)=∫Pi(y,c)xdρ(x).\rho=\lambda(K)^{-1}\lambda|_K,\qquad M_i(y,c)=\rho(P_i(y,c)),\qquad B_i(y,c)=\int_{P_i(y,c)}x\,d\rho(x).(2) Write PiP_i, MiM_i and BiB_i for their values at the fixed sites and intercepts.

For a linear subspace W⊂ℝdW\subset\mathbb{R}^d and a basepoint a∈ℝda\in\mathbb{R}^d, the intrinsic affine plane measure λa+W\lambda_{a+W} is the translation of Lebesgue measure λW\lambda_W on WW: λa+W=(w↦a+w)#λW.\lambda_{a+W}=(w\mapsto a+w)_\#\lambda_W.(3)✦ The canonical hyperplane basepoint z(n,h)z(n,h) for a normal n∈ℝdn\in\mathbb{R}^d and an offset h∈ℝh\in\mathbb{R} is the vector in (4). The intrinsic hyperplane measure σn,h\sigma_{n,h} is the affine plane measure (3) on z(n,h)+(ℝn)⟂z(n,h)+(\mathbb{R}n)^\perp: z(n,h)={−h∥n∥2n,n≠0,0,n=0,σn,h=λz(n,h)+(ℝn)⟂.z(n,h)= \begin{cases} -\dfrac{h}{\|n\|^2}n,&n\ne0,\\ 0,&n=0, \end{cases} \qquad \sigma_{n,h}=\lambda_{z(n,h)+(\mathbb{R}n)^\perp}.(4)✦

The intrinsic interface measure σij\sigma_{ij} is the hyperplane measure (4) with normal yi−yjy_i-y_j and offset ci−cjc_i-c_j.✦ For distinct sites yiy_i and yjy_j, the canonical face coefficients gijg_{ij} and bijb_{ij} are the normalized mass and first moment of the common cell face Pi∩PjP_i\cap P_j with respect to the intrinsic interface measure σij\sigma_{ij}: σij=σyi−yj,ci−cj,gij=σij(Pi∩Pj)λ(K)∥yi−yj∥,bij=1λ(K)∥yi−yj∥∫Pi∩Pjxdσij(x)(i≠j),gii=0,bii=0.\begin{aligned} \sigma_{ij}&=\sigma_{y_i-y_j,c_i-c_j},\\ g_{ij}&=\frac{\sigma_{ij}(P_i\cap P_j)}{\lambda(K)\|y_i-y_j\|},\qquad b_{ij}=\frac{1}{\lambda(K)\|y_i-y_j\|} \int_{P_i\cap P_j}x\,d\sigma_{ij}(x)\quad(i\ne j),\\ g_{ii}&=0,\qquad b_{ii}=0. \end{aligned}(5) For a real coefficient array a=(aij)a=(a_{ij}), the cell Laplacian ℒa\mathcal{L}_a is the linear operator on ℝI\mathbb{R}^I whose value at a vector q∈ℝIq\in\mathbb{R}^I is defined by (ℒaq)i=∑j∈Iaij(qi−qj).(\mathcal{L}_a q)_i=\sum_{j\in I}a_{ij}(q_i-q_j).(6)

For an intercept direction q∈ℝIq\in\mathbb{R}^I and site velocities vi∈ℝdv_i\in\mathbb{R}^d, the directional mass and moment derivatives are the two-sided derivatives along the affine path (y+sv,c+sq)(y+s v,c+s q), whenever they exist: DMi[q,v]=dds|s=0Mi(y+sv,c+sq),DBi[q,v]=dds|s=0Bi(y+sv,c+sq).DM_i[q,v]=\left.\frac{d}{ds}\right|_{s=0}M_i(y+s v,c+s q),\qquad DB_i[q,v]=\left.\frac{d}{ds}\right|_{s=0}B_i(y+s v,c+s q).(7)

A coupling π\pi of measures μ\mu and ν\nu on ℝd\mathbb{R}^d is a probability measure on ℝd×ℝd\mathbb{R}^d\times\mathbb{R}^d whose coordinate marginals are μ\mu and ν\nu.✦ For compact convex sets U,VU,V of positive volume, write ρU,ρV\rho_U,\rho_V for their uniform measures and m(x,z)=12(x+z)m(x,z)=\tfrac12(x+z) for the midpoint map.

The Claim

Theorem 1. Suppose KK has nonempty interior, the sites are pairwise distinct, and Mi>0M_i>0 for every i∈Ii\in I. Let R≥0R\ge0 satisfy ∥x∥≤R\|x\|\le R for ρ\rho-almost every xx. Assume that every compact convex pair U,VU,V of positive volume admits a coupling of their uniform measures whose midpoint pushforward, multiplied by the geometric mean of their volumes, is bounded above by Lebesgue measure. For an intercept direction qq, suppose the stationary derivatives Ḃi=DBi[q,0]\dot B_i=DB_i[q,0] and Ṁi=DMi[q,0]\dot M_i=DM_i[q,0] exist. Then ∑i∈I∥Ḃi∥2Mi≤162R2∑i∈I|Ṁi|2Mi.\sum_{i\in I}\frac{\|\dot B_i\|^2}{M_i} \le 162R^2\sum_{i\in I}\frac{|\dot M_i|^2}{M_i}.(8)✦ For real coefficients gˆij\widehat g_{ij} and vector coefficients bˆij∈ℝd\widehat b_{ij}\in\mathbb{R}^d satisfying Ṁi=(ℒgˆq)i\dot M_i=(\mathcal{L}_{\widehat g}q)_i and Ḃi=∑j(qi−qj)bˆij\dot B_i=\sum_j(q_i-q_j)\widehat b_{ij}, with the Laplacian defined by (6), one has ∑i∈I∥∑j∈I(qi−qj)bˆij∥2Mi≤162R2∑i∈I|(ℒgˆq)i|2Mi.\sum_{i\in I}\frac{\left\|\sum_{j\in I}(q_i-q_j)\widehat b_{ij}\right\|^2}{M_i} \le 162R^2\sum_{i\in I}\frac{|(\mathcal{L}_{\widehat g}q)_i|^2}{M_i}.(9)✦

In the paper, the result reads: for a compact convex body KK with ∥x∥≤R\|x\|\le R on KK, pairwise distinct sites yiy_i and intercepts cic_i with positive uniform cell masses MiM_i satisfy ∑i∥DBi[q,0]∥2/Mi≤162R2∑i|DMi[q,0]|2/Mi\sum_i\|DB_i[q,0]\|^2/M_i\le162R^2\sum_i|DM_i[q,0]|^2/M_i for every intercept direction qq.

What Is Checked

The estimates (8) and (9) are stated over the normalized cell masses and first moments (2) and the directional derivatives (7).

Theorem 2. Let KK be compact and convex, with pairwise distinct sites. Suppose a measure η\eta absolutely continuous with respect to λ\lambda gives positive mass to every closed cell. For the canonical face coefficients (5), the stationary moment derivative exists and is given by (10): DBi[q,0]=∑j∈I(qi−qj)bij.DB_i[q,0]=\sum_{j\in I}(q_i-q_j)b_{ij}.(10)✦ For every intercept direction r∈ℝIr\in\mathbb{R}^I and site velocity v∈(ℝd)Iv\in(\mathbb{R}^d)^I, the mass derivative exists and is given by (11): DMi[r,v]=∑j∈I[gij(ri−rj)+⟨bij,vi−vj⟩].DM_i[r,v]=\sum_{j\in I} \left[g_{ij}(r_i-r_j)+\langle b_{ij},v_i-v_j\rangle\right].(11)✦ Every compact convex pair U,V⊂ℝdU,V\subset\mathbb{R}^d of positive volume admits a coupling π\pi of ρU\rho_U and ρV\rho_V such that λ(U)λ(V)m#π≤λ.\sqrt{\lambda(U)\lambda(V)}\,m_\#\pi\le\lambda.(12)✦