OctoInformBeta

OpenAI preprint, September 2026 · Section 1. Introduction

Section 1 (setting): finite observations, score and retained variance: Retained variance and standardization for observations with positive natural-number cell costs

Scoring finite observations of the cube.

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

A finite observation of independent uniform signs records only a finite amount of information about their configuration. Its score describes the conditional coordinate sum associated with the recorded value, and its cell cost bounds the Fourier degrees of the events that distinguish its values.

Let II be a finite coordinate set and let ΩI={−1,1}I\Omega_I=\{-1,1\}^I be its sign cube, with the uniform distribution.✦ The coordinate sign functions on ΩI\Omega_I are the real-valued maps x↦xix\mapsto x_i, for i∈Ii\in I.

For a finite set AA and a function g:A→ℝg:A\to\mathbb R, the uniform mean 𝔼Ag\mathbb E_A g is the normalized sum in (1), including its value on the empty set. 𝔼Ag={1|A|∑x∈Ag(x),|A|>0,0,|A|=0.\mathbb E_A g= \begin{cases} \displaystyle\frac{1}{|A|}\sum_{x\in A}g(x),& |A|>0,\\ 0,& |A|=0. \end{cases}(1)✦ Write 𝔼\mathbb E for the uniform mean on ΩI\Omega_I.

For S⊆IS\subseteq I, the Fourier-Walsh character χS:ΩI→ℝ\chi_S:\Omega_I\to\mathbb R is the coordinate product in (2). χS(x)=∏i∈Sxi.\chi_S(x)=\prod_{i\in S}x_i.(2)✦ For a function f:ΩI→ℝf:\Omega_I\to\mathbb R and a subset S⊆IS\subseteq I, the real Fourier-Walsh coefficient fˆ(S)\widehat f(S) is the mean in (3). fˆ(S)=𝔼[fχS].\widehat f(S)=\mathbb E\bigl[f\chi_S\bigr].(3)✦ The ordinary real multilinear Fourier degree deg⁡f\deg f of a function f:ΩI→ℝf:\Omega_I\to\mathbb R is the nonnegative integer in (4), with degree zero for constants and the zero function. deg⁡f=max⁡({0}∪{|S|:S⊆I,fˆ(S)≠0}).\deg f=\max\bigl(\{0\}\cup\{|S|:S\subseteq I,\ \widehat f(S)\ne0\}\bigr).(4)✦

The Definition

Definition 1. Let AA be a finite set, let BB be a set, and let F:A→BF:A\to B be a map. For b∈Bb\in B, the cell indicator cF,b:A→ℝc_{F,b}:A\to\mathbb R is the function in (5). cF,b(x)=𝟏{F(x)=b}.c_{F,b}(x)=\mathbf 1_{\{F(x)=b\}}.(5)✦ The cell probability pF(b)p_F(b) is the uniform mass in (6). pF(b)=𝔼AcF,b.p_F(b)=\mathbb E_A c_{F,b}.(6)✦ For a map F:A→BF:A\to B, a function g:A→ℝg:A\to\mathbb R and a label b∈Bb\in B, the unnormalized cell moment mF,g(b)m_{F,g}(b) of gg on the cell F−1({b})F^{-1}(\{b\}) is the mean in (7). mF,g(b)=𝔼A[cF,bg].m_{F,g}(b)=\mathbb E_A\bigl[c_{F,b}g\bigr].(7)✦ Equation (8) defines the cell conditional mean μF,g:B→ℝ\mu_{F,g}:B\to\mathbb R. μF,g(b)={mF,g(b)pF(b),pF(b)>0,0,pF(b)=0.\mu_{F,g}(b)= \begin{cases} \displaystyle\frac{m_{F,g}(b)}{p_F(b)},&p_F(b)>0,\\ 0,&p_F(b)=0. \end{cases}(8)✦ The conditional average 𝔼A[g∣F]:A→ℝ\mathbb E_A[g\mid F]:A\to\mathbb R of gg given FF is the function that assigns the cell conditional mean to each point, as in (9). 𝔼A[g∣F](x)=μF,g(F(x)).\mathbb E_A[g\mid F](x)=\mu_{F,g}(F(x)).(9)✦

On the sign cube ΩI\Omega_I, the coordinate sign sum HI:ΩI→ℝH_I:\Omega_I\to\mathbb R is the function in (10). HI(x)=∑i∈Ixi.H_I(x)=\sum_{i\in I}x_i.(10)✦ For a map F:ΩI→BF:\Omega_I\to B, the observation score TF:ΩI→ℝT_F:\Omega_I\to\mathbb R is the conditional average in (11). TF=𝔼[HI∣F].T_F=\mathbb E[H_I\mid F].(11)✦ The retained variance v(F)v(F) is the second moment of the centered score, given in (12). v(F)=𝔼[TF2].v(F)=\mathbb E\bigl[T_F^2\bigr].(12)✦ For an observation F:ΩI→BF:\Omega_I\to B, equation (13) defines the standardized cell score sF:B→ℝs_F:B\to\mathbb R. sF(b)={μF,HI(b)v(F),v(F)>0,0,v(F)=0.s_F(b)= \begin{cases} \displaystyle\frac{\mu_{F,H_I}(b)}{\sqrt{v(F)}},&v(F)>0,\\ 0,&v(F)=0. \end{cases}(13)✦ The standardized observation score SF:ΩI→ℝS_F:\Omega_I\to\mathbb R is the function in (14). SF(x)=sF(F(x)).S_F(x)=s_F(F(x)).(14)✦

An observation with cell cost is a tuple 𝒪=(I,B,F,d)\mathcal O=(I,B,F,d), where II and BB are finite sets, F:ΩI→BF:\Omega_I\to B is a map and dd is a natural number satisfying (15). d>0,v(F)>0,deg⁡cF,b≤dfor every b∈B.d>0,\qquad v(F)>0,\qquad \deg c_{F,b}\le d\quad\text{for every }b\in B.(15)✦

In the paper, the result reads: for a finite-valued observation FF of N=|I|N=|I| independent uniform signs, HIH_I is their sum, TF=𝔼[HI∣F]T_F=\mathbb E[H_I\mid F] and v(F)=𝔼[TF2]v(F)=\mathbb E[T_F^2], the standardized score for v(F)>0v(F)>0 is TF/v(F)T_F/\sqrt{v(F)}, and a positive real DD is a cell degree bound when every indicator cF,bc_{F,b} has degree at most DD.

What Is Checked

The finite conditioning definitions are stated for maps from any finite uniform space to an arbitrary label set; observations with cell costs have finite coordinate and label sets and satisfy (15).