OctoInformBeta

OpenAI preprint, September 2026 · Section 2. Finite observations and amplification

Section 2 (setting): normalized independent copies: Conditional scores on grouped sign cubes and their first two normalized moments

Independent copies admit a Gaussian normalization.

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

An observation of a sign cube retains a conditional sum of coordinates. Grouping independent observations places these conditional scores on disjoint coordinate blocks. Let Γ\Gamma be a finite coordinate set, and write QΓ={−1,1}ΓQ_\Gamma=\{-1,1\}^{\Gamma} for its sign cube with the uniform law. An observation F:QΓ→BF:Q_\Gamma\to B is a function that reports a value in a set BB.

For a real-valued function gg on a finite set AA, the uniform mean 𝔼Ag\mathbb E_Ag is the average in (1), with value zero on the empty set.✦ 𝔼Ag=|A|−1∑a∈Ag(a).\mathbb E_Ag=|A|^{-1}\sum_{a\in A}g(a).(1) Here and in the scalar normalizations below, 0−1=00^{-1}=0. For nonempty AA, its uniform probability measure μA\mu_A assigns the masses in (2) to measurable subsets EE of AA.✦ μA(E)=|E||A|.\mu_A(E)=\frac{|E|}{|A|}.(2)

The cell indicator cb:QΓ→ℝc_b:Q_\Gamma\to\mathbb R of an observation F:QΓ→BF:Q_\Gamma\to B at a value b∈Bb\in B is the indicator of the fiber F−1(b)F^{-1}(b); its cell probability pbp_b is its uniform mean, as in (3). cb(x)=𝟏{F(x)=b},pb=𝔼QΓcb.c_b(x)=\mathbf 1_{\{F(x)=b\}},\qquad p_b=\mathbb E_{Q_\Gamma}c_b.(3) A cell degree bound DD for an observation F:QΓ→BF:Q_\Gamma\to B is a common bound on the ordinary real multilinear Fourier degrees of its cell indicators cbc_b, for b∈Bb\in B, with constants and the zero function assigned degree zero.

For an observation F:QΓ→BF:Q_\Gamma\to B, a value b∈Bb\in B and a function g:QΓ→ℝg:Q_\Gamma\to\mathbb R, the restricted moment mg(b)m_g(b) is the real number given by the uniform mean of cbgc_bg, and the fiber conditional mean CFg(b)C_Fg(b) is the real number obtained by dividing mg(b)m_g(b) by the cell probability pbp_b. The conditional average 𝔼[g∣F]:QΓ→ℝ\mathbb E[g\mid F]:Q_\Gamma\to\mathbb R of a function g:QΓ→ℝg:Q_\Gamma\to\mathbb R given an observation F:QΓ→BF:Q_\Gamma\to B evaluates the fiber conditional mean CFgC_Fg at F(x)F(x), as in (4).✦ mg(b)=𝔼QΓ(cbg),CFg(b)=mg(b)pb,𝔼[g∣F](x)=CFg(F(x)).\begin{aligned} m_g(b)&=\mathbb E_{Q_\Gamma}(c_bg),& C_Fg(b)&=\frac{m_g(b)}{p_b},\\ \mathbb E[g\mid F](x)&=C_Fg(F(x)). \end{aligned}(4) The fiber conditional mean is zero when the cell probability is zero.

For an observation F:QΓ→BF:Q_\Gamma\to B, the coordinate sum SΓ:QΓ→ℝS_\Gamma:Q_\Gamma\to\mathbb R is the function that sums the coordinates, and its conditional score τF:QΓ→ℝ\tau_F:Q_\Gamma\to\mathbb R is the conditional average of this sum given FF, as in (5). The retained variance v(F)v(F) of an observation F:QΓ→BF:Q_\Gamma\to B is the second moment of its centered conditional score τF\tau_F under the uniform law on QΓQ_\Gamma.✦ SΓ(x)=∑i∈Γxi,τF=𝔼[SΓ∣F],v(F)=𝔼QΓτF2.S_\Gamma(x)=\sum_{i\in\Gamma}x_i,\qquad \tau_F=\mathbb E[S_\Gamma\mid F],\qquad v(F)=\mathbb E_{Q_\Gamma}\tau_F^2.(5) For an observation F:QΓ→BF:Q_\Gamma\to B and a value b∈Bb\in B, the standardized conditional value uF(b)u_F(b) is the real number obtained by dividing the fiber conditional coordinate sum CFSΓ(b)C_FS_\Gamma(b) by v(F)\sqrt{v(F)}. The standardized score tF:QΓ→ℝt_F:Q_\Gamma\to\mathbb R of an observation F:QΓ→BF:Q_\Gamma\to B evaluates the standardized conditional value uFu_F at F(x)F(x), as in (6).✦ uF(b)=CFSΓ(b)v(F),tF(x)=uF(F(x)).u_F(b)=\frac{C_FS_\Gamma(b)}{\sqrt{v(F)}},\qquad t_F(x)=u_F(F(x)).(6)

For a finite block index set KK, the block coordinate map ℬK:QK×Γ→(QΓ)K\mathcal B_K:Q_{K\times\Gamma}\to(Q_\Gamma)^K is the bijection in (7).✦ The grouped observation FK:QK×Γ→BKF_K:Q_{K\times\Gamma}\to B^K applies FF to each block.✦ (ℬKx)j(i)=x(j,i),FK(x)j=F((ℬKx)j).(\mathcal B_Kx)_j(i)=x_{(j,i)},\qquad F_K(x)_j=F((\mathcal B_Kx)_j).(7)

Write [n]={1,…,n}[n]=\{1,\ldots,n\} for a nonnegative integer nn, with [0]=⌀[0]=\varnothing. For a function T:A→ℝT:A\to\mathbb R and a nonnegative integer nn, the normalized finite sum NnT:A[n]→ℝN_nT:A^{[n]}\to\mathbb R is the function in (8).✦ NnT(a)=1n∑j∈[n]T(aj).N_nT(a)=\frac{1}{\sqrt n}\sum_{j\in[n]}T(a_j).(8) For an observation F:QΓ→BF:Q_\Gamma\to B and a nonnegative integer nn, the grouped score scale aF(n)a_F(n) is the real number obtained by multiplying the square roots of the block count nn and the retained variance v(F)v(F). The normalized grouped score ZF,n:Q[n]×Γ→ℝZ_{F,n}:Q_{[n]\times\Gamma}\to\mathbb R is the normalized finite sum of the standardized block scores, as in (9).✦ aF(n)=nv(F),ZF,n=NntF∘ℬ[n].a_F(n)=\sqrt n\sqrt{v(F)},\qquad Z_{F,n}=N_nt_F\circ\mathcal B_{[n]}.(9)

The Claim

For a probability measure ν\nu and a real-valued random variable YY, write ℒν(Y)\mathcal L_\nu(Y) for the distribution of YY under ν\nu, and ⇒\Longrightarrow for weak convergence of probability measures. The standard normal law on ℝ\mathbb R is denoted by 𝒩(0,1)\mathcal N(0,1).

Theorem 1. Let μ\mu be a probability measure on a measurable space AA, and let T:A→ℝT:A\to\mathbb R be measurable with ∫Tdμ=0\int T\,d\mu=0 and ∫T2dμ=1\int T^2\,d\mu=1. Under the product laws μ⊗n\mu^{\otimes n}, ℒμ⊗n(NnT)⇒𝒩(0,1)(n→∞).\mathcal L_{\mu^{\otimes n}}(N_nT) \Longrightarrow\mathcal N(0,1)\qquad(n\to\infty).(10)✦ For a nonempty finite set AA and a function T:A→ℝT:A\to\mathbb R with 𝔼AT=0\mathbb E_AT=0 and 𝔼AT2=1\mathbb E_AT^2=1, the uniform laws satisfy ℒμA[n](NnT)⇒𝒩(0,1)(n→∞).\mathcal L_{\mu_{A^{[n]}}}(N_nT) \Longrightarrow\mathcal N(0,1)\qquad(n\to\infty).(11)✦ Let F:QΓ→BF:Q_\Gamma\to B be an observation. For every n≥0n\geq0, its normalized grouped score has mean 𝔼Q[n]×ΓZF,n=0.\mathbb E_{Q_{[n]\times\Gamma}}Z_{F,n}=0.(12)✦ If v(F)>0v(F)>0 and n≥1n\geq1, its grouped conditional score satisfies τF[n]=aF(n)ZF,n,\tau_{F_{[n]}}=a_F(n)Z_{F,n},(13)✦ and the normalized second moment is 𝔼Q[n]×ΓZF,n2=1.\mathbb E_{Q_{[n]\times\Gamma}}Z_{F,n}^2=1.(14)✦

In the paper, the result reads: for v=v(F)>0v=v(F)>0 and a cell degree bound DD, MM independent copies T1,…,TMT_1,\ldots,T_M of the conditional score give ZM=(T1+⋯+TM)/MvZ_M=(T_1+\cdots+T_M)/\sqrt{Mv}, the joint observation has cell degree bound MDMD and score MvZM\sqrt{Mv}\,Z_M, and ZMZ_M converges in distribution to 𝒩(0,1)\mathcal N(0,1).

What Is Checked

The limit statements (10) and (11) concern product probability measures and uniform laws on finite Cartesian powers; the score identities (12), (13) and (14) concern observations on finite sign cubes, with positive retained variance and a positive block count required for the latter two identities.