OctoInformBeta

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

§3: Normalized radii and normal functions: Active cell constraints and the Gaussian functions that express deletion loss

Normalized radii and Gaussian cell profiles.

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 centroid lengths and probabilities of the cells of a Gaussian partition describe its energy and the distribution of its mass. Normalizing the centroid lengths removes the scale of the energy, while one-dimensional Gaussian tails give scalar cell profiles.

Let γd\gamma_d be standard Gaussian probability measure, with mean zero and identity covariance, on Euclidean space ℝd\mathbb R^d.✦ The Gaussian centroid z(E)z(E) of a set E⊆ℝdE\subseteq\mathbb R^d is the unnormalized vector in (1); the centroid energy F(A)F(A) of a finite family A=(Ai)i=1kA=(A_i)_{i=1}^k is the sum of squared centroid lengths in that equation. z(E)=∫Exdγd(x),F(A)=∑i=1k∥z(Ai)∥2.z(E)=\int_E x\,d\gamma_d(x), \qquad F(A)=\sum_{i=1}^k\|z(A_i)\|^2.(1)

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 finite family AA consists of the labels of its positive-mass cells, and its active count mm is the cardinality in (2). I(A)={i∈{1,…,k}:γd(Ai)>0},m=|I(A)|.I(A)=\{i\in\{1,\ldots,k\}:\gamma_d(A_i)>0\}, \qquad m=|I(A)|.(2) The probability PiP_i of a cell AiA_i and, for positive energy C=F(A)C=F(A), its normalized active radius rir_i are defined by (3). Pi=γd(Ai),ri=∥z(Ai)∥C(i∈I(A),C>0).P_i=\gamma_d(A_i), \qquad r_i=\frac{\|z(A_i)\|}{\sqrt C}\quad(i\in I(A),\ C>0).(3)

An optimal Gaussian partition AA is a Gaussian measurable partition whose energy is at least that of every Gaussian measurable partition with the same labels in the same space.✦ A minimal optimal partition AA is an optimal partition whose active count is no larger than that of any other optimal partition with those labels in that space.✦

One-dimensional Tails

The unnormalized Gaussian density hh on ℝ\mathbb R and its upper tail TT are the functions in (4). Write c=(2π)−1/2c=(2\pi)^{-1/2} for the Gaussian normalizing constant.✦ h(t)=exp⁡(−t2/2),T(x)=∫x∞h(t)dt.h(t)=\exp(-t^2/2), \qquad T(x)=\int_x^\infty h(t)\,dt.(4)

The Claim

Definition 1. For a finite family A=(Ai)i=1kA=(A_i)_{i=1}^k of subsets of ℝd\mathbb R^d with positive energy C=F(A)C=F(A), the normalized active radii satisfy (5). ∑i∈I(A)ri2=1.\sum_{i\in I(A)}r_i^2=1.(5)✦ For any Gaussian measurable partition AA, the active probabilities satisfy (6). ∑i∈I(A)Pi=1.\sum_{i\in I(A)}P_i=1.(6)✦ For a minimal optimal partition AA with k≥2k\geq2 and positive energy, each active cell satisfies (7). 0<Pi≤12(i∈I(A)).0<P_i\leq\frac12\qquad(i\in I(A)).(7)✦

The standard Gaussian density φ\varphi and distribution function Φ\Phi on ℝ\mathbb R are defined by (8) and (9). φ(t)=ch(t).\varphi(t)=c\,h(t).(8)✦ Φ(t)=∫−∞tφ(s)ds.\Phi(t)=\int_{-\infty}^t\varphi(s)\,ds.(9)✦ The Gaussian quantile q:ℝ→ℝq:\mathbb R\to\mathbb R is the inverse of the standard Gaussian distribution function on (0,1)(0,1), as in (10), and is zero outside that interval. q(P)=Φ−1(P)(0<P<1).q(P)=\Phi^{-1}(P)\qquad(0<P<1).(10) On ℝ\mathbb R, (11) and (12) define the Gaussian upper tail pp and inverse Mills ratio λ\lambda. p(x)=cT(x)=Φ(−x).p(x)=c\,T(x)=\Phi(-x).(11)✦ λ(x)=h(x)T(x)=φ(x)Φ(−x).\lambda(x)=\frac{h(x)}{T(x)} =\frac{\varphi(x)}{\Phi(-x)}.(12)✦ The inverse Mills derivative 𝒢\mathcal G in the upper-tail variable xx and its reflection GG in the variable t=−xt=-x are defined by (13). 𝒢(x)=λ(x)(λ(x)−x),G(t)=𝒢(−t).\mathcal G(x)=\lambda(x)\bigl(\lambda(x)-x\bigr), \qquad G(t)=\mathcal G(-t).(13)✦ The Gaussian loss profile loss⁡:ℝ→ℝ\operatorname{loss}:\mathbb R\to\mathbb R is the function defined by (14). loss⁡(x)=p(x)2𝒢(x)=p(x)2G(−x).\operatorname{loss}(x)=p(x)^2\mathcal G(x)=p(x)^2G(-x).(14)✦ For P∈(0,1)P\in(0,1) and x∈ℝx\in\mathbb R with p(x)=Pp(x)=P, its quantile expression is (15). loss⁡(x)=φ(q(P))(q(P)P+φ(q(P))).\operatorname{loss}(x) =\varphi(q(P))\bigl(q(P)P+\varphi(q(P))\bigr).(15)✦

In the paper, the result reads: for a hypothetical minimal optimal partition (Ki)i=1m(K_i)_{i=1}^m with m≥5m\geq5 active cells and value C>B2C>B^2, where B=3/(22π)B=3/(2\sqrt{2\pi}), in its (m−1)(m-1)-dimensional centroid span, the quantities ri=∥z(Ki)∥/Cr_i=\|z(K_i)\|/\sqrt C and Pi=γm−1(Ki)P_i=\gamma_{m-1}(K_i) satisfy ri>0r_i>0, 0<Pi≤1/20<P_i\leq1/2 and ∑iri2=∑iPi=1\sum_i r_i^2=\sum_i P_i=1, with φ(t)=exp⁡(−t2/2)/2π\varphi(t)=\exp(-t^2/2)/\sqrt{2\pi}, Φ(t)=∫−∞tφ(s)ds\Phi(t)=\int_{-\infty}^t\varphi(s)\,ds and G(t)=φ(t)(tΦ(t)+φ(t))/Φ(t)2G(t)=\varphi(t)(t\Phi(t)+\varphi(t))/\Phi(t)^2.

What Is Checked

The radius identity is stated for finite families of subsets of Euclidean space with positive centroid energy; the probability identities concern Gaussian measurable partitions, with minimal optimality, at least two labels and positive centroid energy for the upper bound. The quantile loss identity is stated for an arbitrary real tail threshold whose probability lies in (0,1)(0,1).

The Gaussian distribution function in (9) is a strictly increasing bijection from ℝ\mathbb R onto (0,1)(0,1).