OctoInformBeta

OpenAI preprint, September 2026 · §2. Extremal partitions and their centroids

§2: The optimal value, scores and active cells: Gaussian width and score cones, with optimality refined by the number of cells of positive mass

Optimal partitions and their active cells.

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 Gaussian centroid energy measures a partition through the first moments of its cells. Let d≥0d\geq 0 and k≥1k\geq 1 be integers, and write Ik={0,…,k−1}I_k=\{0,\ldots,k-1\} for the ordered label set. The ambient space is Euclidean ℝd\mathbb{R}^d, with inner product ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle and standard Gaussian measure γd\gamma_d.

A Borel Gaussian kk-cell partition A=(Ai)i∈IkA=(A_i)_{i\in I_k} of ℝd\mathbb{R}^d is a family of Borel sets such that almost every point belongs to exactly one cell: Part⁡d,k(A)⇔(Ai is Borel for every i∈Ik)andγd({x:∑i∈Ik𝟏Ai(x)=1})=1.\operatorname{Part}_{d,k}(A) \quad\Longleftrightarrow\quad \bigl(A_i\text{ is Borel for every }i\in I_k\bigr) \ \text{and}\ \gamma_d\!\left(\left\{x: \sum_{i\in I_k}\mathbf{1}_{A_i}(x)=1\right\}\right)=1.(1)✦ Empty cells are allowed in (1).

The unnormalized Gaussian centroid z(D)z(D) of a set D⊆ℝdD\subseteq\mathbb{R}^d is its vector integral with respect to standard Gaussian measure, as in (2).✦ z(D)=∫Dxdγd(x).z(D)=\int_D x\,d\gamma_d(x).(2)✦ The centroid energy F(A)F(A) of a family A=(Ai)i∈IkA=(A_i)_{i\in I_k} of sets in ℝd\mathbb{R}^d is the sum of the squared norms of its Gaussian centroids, as in (3).✦ F(A)=∑i∈Ik∥z(Ai)∥2.F(A)=\sum_{i\in I_k}\|z(A_i)\|^2.(3)✦

The Definition

Definition 1. A Euclidean vector tuple v∈𝒯d,kv\in\mathcal{T}_{d,k} is a labelled family of vectors in ℝd\mathbb{R}^d, with the Euclidean product norm in (4): 𝒯d,k=(ℝd)Ik,∥v∥𝒯d,k2=∑i∈Ik∥vi∥2.\mathcal{T}_{d,k}=(\mathbb{R}^d)^{I_k}, \qquad \|v\|_{\mathcal{T}_{d,k}}^2=\sum_{i\in I_k}\|v_i\|^2.(4)✦ For v∈𝒯d,kv\in\mathcal{T}_{d,k} and x∈ℝdx\in\mathbb{R}^d, the maximum linear score mv(x)m_v(x) is the largest inner product in (5): mv(x)=maxi∈Ik⟨vi,x⟩.m_v(x)=\max_{i\in I_k}\langle v_i,x\rangle.(5)✦ The Gaussian width width⁡(v)\operatorname{width}(v) of a tuple v∈𝒯d,kv\in\mathcal{T}_{d,k} is the Gaussian integral of its maximum score, as in (6): width⁡(v)=∫ℝdmv(x)dγd(x).\operatorname{width}(v)=\int_{\mathbb{R}^d}m_v(x)\,d\gamma_d(x).(6)✦

For a tuple v∈𝒯d,kv\in\mathcal{T}_{d,k} and a label i∈Iki\in I_k, the score cell Si(v)S_i(v) is the set of points in ℝd\mathbb{R}^d where ii is the smallest label attaining the maximum score, specified by (7): Si(v)={x∈ℝd:⟨vj,x⟩≤⟨vi,x⟩for every j∈Ik,⟨vj,x⟩<⟨vi,x⟩for every j<i}.S_i(v)=\left\{x\in\mathbb{R}^d: \begin{array}{ll} \langle v_j,x\rangle\leq\langle v_i,x\rangle &\text{for every }j\in I_k,\\ \langle v_j,x\rangle<\langle v_i,x\rangle &\text{for every }j<i \end{array}\right\}.(7)✦ For a tuple v∈𝒯d,kv\in\mathcal{T}_{d,k} and a label i∈Iki\in I_k, the closed score cone Ki(v)⊆ℝdK_i(v)\subseteq\mathbb{R}^d is the set where label ii attains the maximum linear score, including ties, as in (8): Ki(v)={x∈ℝd:⟨vj,x⟩≤⟨vi,x⟩ for every j∈Ik}.K_i(v)=\{x\in\mathbb{R}^d: \langle v_j,x\rangle\leq\langle v_i,x\rangle \text{ for every }j\in I_k\}.(8)✦

An optimal partition AA of ℝd\mathbb{R}^d is a Borel Gaussian kk-cell partition whose centroid energy (3) is maximal among such partitions, as specified by (9): Opt⁡(A)⇔Part⁡d,k(A)andF(D)≤F(A)for every D with Part⁡d,k(D).\begin{split} \operatorname{Opt}(A)\quad\Longleftrightarrow\quad& \operatorname{Part}_{d,k}(A)\ \text{and}\\ &F(D)\leq F(A)\quad \text{for every }D\text{ with }\operatorname{Part}_{d,k}(D). \end{split}(9)✦ The active label set ℐ(A)⊆Ik\mathcal{I}(A)\subseteq I_k of a family of cells AA consists of the labels of cells with positive Gaussian measure, as in (10): ℐ(A)={i∈Ik:γd(Ai)>0}.\mathcal{I}(A)=\{i\in I_k:\gamma_d(A_i)>0\}.(10)✦ The active count a(A)a(A) of a family A=(Ai)i∈IkA=(A_i)_{i\in I_k} of cells in ℝd\mathbb{R}^d is the number of labels in its active label set ℐ(A)\mathcal{I}(A), as in (11): a(A)=#ℐ(A).a(A)=\#\mathcal{I}(A).(11)✦ A minimal optimal partition AA of ℝd\mathbb{R}^d with label set IkI_k is an optimal partition whose active count is least among optimal partitions of ℝd\mathbb{R}^d with label set IkI_k, as specified by (12): MinOpt⁡(A)⇔Opt⁡(A)anda(A)≤a(D)for every optimal partition D.\begin{split} \operatorname{MinOpt}(A)\quad\Longleftrightarrow\quad& \operatorname{Opt}(A)\ \text{and}\\ &a(A)\leq a(D)\quad\text{for every optimal partition }D. \end{split}(12)✦

In the paper, the result reads: fix n≥4n\geq4 and N=n+1N=n+1; with CC the supremum of FF over measurable NN-cell partitions of ℝn\mathbb{R}^n, B=3/(22π)B=3/(2\sqrt{2\pi}) and width⁡(v)=𝔼max⁡i⟨vi,g⟩\operatorname{width}(v)=\mathbb{E}\max_i\langle v_i,g\rangle for a standard Gaussian vector gg, an optimal partition attains CC, and a minimal-active-cell optimum minimizes the number of cells of positive Gaussian measure among optima.

What Is Checked

Optimality (9) and minimal optimality (12) are stated over Borel Gaussian partitions of ℝd\mathbb{R}^d with a fixed finite label set.