OctoInformBeta

OpenAI preprint, September 2026 · §3. Probability bounds for the cells

Lemma 3.1: Quadratic estimates for measurable cells and stronger bounds for minimal optimal partitions above the propeller value

Bounds for Gaussian cell radii and probabilities.

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 relation between a cell’s Gaussian probability and the length of its centroid restricts the geometry of partitions whose energy exceeds the propeller value. Let γd\gamma_d be standard Gaussian probability measure on Euclidean space ℝd\mathbb R^d, with mean zero and identity covariance.✦ The unnormalized Gaussian centroid z(D)z(D) of a set D⊆ℝdD\subseteq\mathbb R^d is the vector in (1): z(D)=∫Dxdγd(x).z(D)=\int_D x\,d\gamma_d(x).(1) For a family A=(Ai)i=1kA=(A_i)_{i=1}^k of sets in ℝd\mathbb R^d, its Gaussian centroid energy F(A)F(A) is the sum in (2): F(A)=∑i=1k∥z(Ai)∥2.F(A)=\sum_{i=1}^k\|z(A_i)\|^2.(2)✦ Write B=3/(22π)B=3/(2\sqrt{2\pi}), so the propeller value is B2=9/(8π)B^2=9/(8\pi).

A Gaussian measurable partition A=(Ai)i=1kA=(A_i)_{i=1}^k of ℝd\mathbb R^d is a family of Borel sets such that γd\gamma_d-almost every point belongs to exactly one cell.✦ Empty cells are allowed. The active label set I(A)I(A) of a family A=(Ai)i=1kA=(A_i)_{i=1}^k of sets in ℝd\mathbb R^d is the set of labels of its positive-mass cells, and the number n(A)n(A) of active cells is the cardinality of I(A)I(A), as in (3): I(A)={i∈{1,…,k}:γd(Ai)>0},n(A)=|I(A)|.I(A)=\{i\in\{1,\ldots,k\}:\gamma_d(A_i)>0\}, \qquad n(A)=|I(A)|.(3)

An optimal partition AA of ℝd\mathbb R^d with kk labels is a Gaussian measurable partition for which F(A)F(A) is at least the energy of every Gaussian measurable partition in the same space with the same labels.✦ A minimal optimal partition AA of ℝd\mathbb R^d with kk labels is an optimal partition whose number n(A)n(A) of active cells is no larger than that of any optimal partition of ℝd\mathbb R^d with kk labels.✦

For a partition AA of positive energy, set C=F(A)C=F(A). The normalized centroid radii rir_i and cell probabilities PiP_i of the partition A=(Ai)i=1kA=(A_i)_{i=1}^k of ℝd\mathbb R^d are the quantities in (4), with zi=z(Ai)z_i=z(A_i): ri=∥zi∥C,Pi=γd(Ai),∑i=1kri2=∑i=1kPi=1.r_i=\frac{\|z_i\|}{\sqrt C},\qquad P_i=\gamma_d(A_i),\qquad \sum_{i=1}^k r_i^2=\sum_{i=1}^k P_i=1.(4)

The Claim

Theorem 1. Let D⊆ℝdD\subseteq\mathbb R^d be Borel measurable and let C>B2C>B^2. Then (5) and (6) hold: ∥z(D)∥C<23.\frac{\|z(D)\|}{\sqrt C}<\frac23.(5)✦ 0.884(∥z(D)∥C)2≤γd(D).0.884\left(\frac{\|z(D)\|}{\sqrt C}\right)^2\le\gamma_d(D).(6)✦ Let A=(Ai)i=1kA=(A_i)_{i=1}^k be a minimal optimal partition whose cells all have positive Gaussian measure, and suppose C=F(A)>B2C=F(A)>B^2. When k=5k=5, every cell satisfies (7): 0.929ri2≤Pi.0.929r_i^2\le P_i.(7)✦ When k≥2k\ge2, every cell satisfies (8): 0.415ri+0.15ri2<Pi.0.415r_i+0.15r_i^2<P_i.(8)✦

In the paper, the result reads: under its preceding extremizer hypotheses, with C>B2C>B^2 and m≥5m\ge5 cells, put a=0.929a=0.929 if m=5m=5 and a=0.884a=0.884 if m≥6m\ge6; every cell satisfies ri≤2/3r_i\le2/3, Pi≥ari2P_i\ge ar_i^2 and Pi≥0.415ri+0.15ri2P_i\ge0.415r_i+0.15r_i^2.

What Is Checked

The set bounds are stated over Borel subsets of Euclidean space with an independently chosen C>B2C>B^2; the partition bounds are stated over minimal optimal partitions with positive-mass cells, normalized by their own centroid energy.