OctoInformBeta

OpenAI preprint, September 2026 · Section 3. Retaining more variance than the degree cost

Section 3.1 (setting): the exceptional-event cancellation identity: The selected failure indicator equals a soft weight on deterministic finite tables

Exceptional events have a cell degree bound.

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 deterministic table applied to reports on a sign cube can single out a failing leaf, while the Fourier degrees of the report cells measure the input cost.

For a finite set XX and a function h:X→ℝh:X\to\mathbb R, the uniform mean 𝔼Xh\mathbb E_X h is the real number given by (1), with value zero when XX is empty: 𝔼Xh={|X|−1∑x∈Xh(x),X≠⌀,0,X=⌀.\mathbb E_X h= \begin{cases} |X|^{-1}\displaystyle\sum_{x\in X}h(x),&X\ne\varnothing,\\ 0,&X=\varnothing. \end{cases}(1) For a finite coordinate set Γ\Gamma, the sign cube is CΓ={−1,1}ΓC_\Gamma=\{-1,1\}^{\Gamma}.✦ For T⊆ΓT\subseteq\Gamma and f:CΓ→ℝf:C_\Gamma\to\mathbb R, the Fourier character χT:CΓ→ℝ\chi_T:C_\Gamma\to\mathbb R is a function, the real Fourier-Walsh coefficient fˆ(T)\widehat f(T) is a real number, and the ordinary real multilinear Fourier degree deg⁡f\deg f is a nonnegative integer, each given by (2): χT(x)=∏k∈Txk,fˆ(T)=𝔼CΓ(fχT),deg⁡f=max⁡({0}∪{|T|:T⊆Γ,fˆ(T)≠0}).\begin{aligned} \chi_T(x)&=\prod_{k\in T}x_k, &\widehat f(T)&=\mathbb E_{C_\Gamma}(f\chi_T),\\ \deg f&=\max\bigl(\{0\}\cup \{|T|:T\subseteq\Gamma,\ \widehat f(T)\ne0\}\bigr). \end{aligned}(2) This degree is distinct from polynomial threshold degree and degree over a finite field.

Let β\beta be a label set, and let 𝟏Q\mathbf1_Q denote the real indicator of a condition QQ, equal to 11 when QQ holds and 00 otherwise. For any map F:X→βF:X\to\beta and label b∈βb\in\beta, the cell indicator cF,b:X→ℝc_{F,b}:X\to\mathbb R is the fiber indicator in (3): cF,b(x)=𝟏{F(x)=b}.c_{F,b}(x)=\mathbf1_{\{F(x)=b\}}.(3)✦

Let JJ be a finite leaf index set and β\beta a finite label set. A selector S:β→JS:\beta\to J is a map that chooses a leaf from the distinguished label. Let ℓj:β→{0,1}\ell_j:\beta\to\{0,1\} be the Boolean indicator of the jjth leaf test, with value 11 for true and 00 for false. For a table point (u,z)∈β×βJ(u,z)\in\beta\times\beta^J, the unique failing leaf event A⊆β×βJA\subseteq\beta\times\beta^J is defined by (4): A={(u,z):ℓS(u)(zS(u))=0,ℓj(zj)=1for every j≠S(u)}.A=\bigl\{(u,z):\ell_{S(u)}(z_{S(u)})=0,\quad \ell_j(z_j)=1\ \text{for every }j\ne S(u)\bigr\}.(4)✦

For real selector weights sj:β→ℝs_j:\beta\to\mathbb R and leaf weights wj:β→ℝw_j:\beta\to\mathbb R, the soft event weight Hs,w:β×βJ→ℝH_{s,w}:\beta\times\beta^J\to\mathbb R is the function in (5): Hs,w(u,z)=∑i∈Jsi(u)∏j∈J\{i}wj(zj)−∏j∈Jwj(zj).H_{s,w}(u,z)= \sum_{i\in J}s_i(u)\prod_{j\in J\setminus\{i\}}w_j(z_j) -\prod_{j\in J}w_j(z_j).(5)✦

The Claim

Theorem 1. Let SS and ℓ\ell define the event AA in (4), and let si(u)=𝟏{S(u)=i}s_i(u)=\mathbf1_{\{S(u)=i\}}. For every (u,z)∈β×βJ(u,z)\in\beta\times\beta^J, the event indicator satisfies (6): 𝟏A(u,z)=Hs,ℓ(u,z).\mathbf1_A(u,z)=H_{s,\ell}(u,z).(6)✦ Let F0:CΓ→βF_0:C_\Gamma\to\beta and Fj:CΓ→βF_j:C_\Gamma\to\beta for j∈Jj\in J be reports, and let D∈ℕD\in\mathbb N. Suppose deg⁡cF0,b≤D\deg c_{F_0,b}\le D and deg⁡cFj,b≤D\deg c_{F_j,b}\le D for every label b∈βb\in\beta and leaf j∈Jj\in J. The induced event on the cube satisfies the degree bound (7): deg⁡(x↦𝟏A(F0(x),(Fj(x))j∈J))≤|J|D.\deg\Bigl(x\longmapsto \mathbf1_A\bigl(F_0(x),(F_j(x))_{j\in J}\bigr)\Bigr)\le |J|D.(7)✦

In the paper, the result reads: for m≥3m\ge3 intervals IiI_i and a selector SS with finitely many cut points, on inputs (u,z1,…,zm)(u,z_1,\ldots,z_m) the events B={(u,z):zj∈Ijfor every j}B=\{(u,z):z_j\in I_j\ \text{for every }j\}, P={(u,z):zj∈Ijfor every j≠S(u)}P=\{(u,z):z_j\in I_j\ \text{for every }j\ne S(u)\} and A=P\BA=P\setminus B satisfy (8), with each summand using at most mm coordinates and the refined report WW returning only a tag on AA, all leaves on BB, and uu together with the unselected leaves on PcP^c: 𝟏A(u,z)=∑i=1m𝟏{S(u)=i}∏1≤j≤mj≠i𝟏Ij(zj)−∏j=1m𝟏Ij(zj).\mathbf1_A(u,z)= \sum_{i=1}^m\mathbf1_{\{S(u)=i\}} \prod_{\substack{1\le j\le m\\j\ne i}}\mathbf1_{I_j}(z_j) -\prod_{j=1}^m\mathbf1_{I_j}(z_j).(8)

What Is Checked

The statements concern deterministic finite tables with a selector and Boolean leaf tests; the degree statement concerns reports from a finite sign cube whose cell indicators have a common Fourier degree bound.