OctoInformBeta

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

Lemma 2.1: Normalized sums of a centered variance-one finite observable exceed the fixed gain threshold at some positive sample size

Gaussian moment limits yield finite soft gain.

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

Finite amplification compares weighted moments of normalized finite sums with Gaussian moments. The weights here are fixed tent functions and selector weights, and the gain is a function of their moments of orders zero, one and two.

Let γ\gamma be the standard normal probability law. For a measurable space Ω\Omega, a measure PP on it, a real observable Z:Ω→ℝZ:\Omega\to\mathbb R and a family of real weights wi:ℝ→ℝw_i:\mathbb R\to\mathbb R indexed by a set II, the weighted moment array MP,ZwM_{P,Z}^{w} is defined by (1). MP,Zw(k,i)=∫Ωwi(Z(x))Z(x)kdP(x),k∈{0,1,2},i∈I.M_{P,Z}^{w}(k,i)=\int_\Omega w_i(Z(x))Z(x)^k\,dP(x), \qquad k\in\{0,1,2\},\quad i\in I.(1)✦ The weighted means μi(l)\mu_i(l) and weighted variances vi(l)v_i(l) of a moment array l=(lk,i)l=(l_{k,i}) are given by (2), with nonzero zeroth moments where the quotients are used: μi(l)=l1,il0,i,vi(l)=l2,il0,i−μi(l)2.\mu_i(l)=\frac{l_{1,i}}{l_{0,i}},\qquad v_i(l)=\frac{l_{2,i}}{l_{0,i}}-\mu_i(l)^2.(2)

The Fixed Gaussian Weights

The clipping function C:ℝ→[0,1]C:\mathbb R\to[0,1] and, for real parameters a,ha,h, the tent function τa,h:ℝ→ℝ\tau_{a,h}:\mathbb R\to\mathbb R are defined in (3). C(t)=max⁡{0,min⁡{1,t}},τa,h(t)=max⁡{0,1−|t−a|h}.C(t)=\max\{0,\min\{1,t\}\},\qquad \tau_{a,h}(t)=\max\left\{0,1-\frac{|t-a|}{h}\right\}.(3) For a grid half-size N∈ℕN\in\mathbb N and spacing δ>0\delta>0, let IN={0,…,2N}I_N=\{0,\ldots,2N\}, whose middle index is NN. The grid centers cic_i, ramp differences HiH_i and selector weights WiW_i are the functions in (4): ci(N,δ)=1+δ(i−N),HiN,δ(t)=C(t/δ+N−i+1)−C(t/δ+N−i),WiN,δ(t)=HiN,δ(t)+𝟏{i=N}(1−∑j∈INHjN,δ(t)).\begin{aligned} c_i(N,\delta)&=1+\delta(i-N),\\ H_i^{N,\delta}(t)&=C(t/\delta+N-i+1)-C(t/\delta+N-i),\\ W_i^{N,\delta}(t)&=H_i^{N,\delta}(t) +\mathbf 1_{\{i=N\}}\left(1-\sum_{j\in I_N}H_j^{N,\delta}(t)\right). \end{aligned}(4) The leaf index set is I=INI=I_N for the fixed grid parameters in (5), where RR is the grid radius and hh is the leaf tent width: N=(2128)8,R=96,δ=RN,h=1(2128)12.N=(2^{128})^8,\qquad R=96,\qquad \delta=\frac{R}{N},\qquad h=\frac{1}{(2^{128})^{12}}.(5) For each leaf index i∈Ii\in I, the leaf test ϕi:ℝ→ℝ\phi_i:\mathbb R\to\mathbb R and selector test ψi:ℝ→ℝ\psi_i:\mathbb R\to\mathbb R are the functions defined in (6): ϕi(t)=τci(N,δ),h(t),ψi(t)=WiN,δ(t),i∈I.\phi_i(t)=\tau_{c_i(N,\delta),h}(t),\qquad \psi_i(t)=W_i^{N,\delta}(t),\qquad i\in I.(6)

The moment gain 𝒢(l,s)\mathcal G(l,s) of two arrays on a finite index set is the expression in (7): 𝒢(l,s)=(∏il0,i)(1+∑ivi(l)−∑is2,i+2(1−μi(l))s1,i+((1−μi(l))2+∑j≠ivj(l))s0,il0,i).\mathcal G(l,s)=\left(\prod_i l_{0,i}\right) \left(1+\sum_i v_i(l) -\sum_i\frac{s_{2,i}+2(1-\mu_i(l))s_{1,i} +\left((1-\mu_i(l))^2+\sum_{j\ne i}v_j(l)\right)s_{0,i}}{l_{0,i}}\right).(7)✦ The Gaussian leaf and selector moment arrays ℓG,sG\ell^G,s^G and the gain threshold η\eta are defined by (8); id\mathrm{id} denotes the identity observable on ℝ\mathbb R: ℓG=Mγ,idϕ,sG=Mγ,idψ,η=12𝒢(ℓG,sG).\ell^G=M_{\gamma,\mathrm{id}}^\phi,\qquad s^G=M_{\gamma,\mathrm{id}}^\psi,\qquad \eta=\tfrac12\mathcal G(\ell^G,s^G).(8) For observables TnT_n on probability spaces (Ωn,Pn)(\Omega_n,P_n), write ℓn=MPn,Tnϕ\ell_n=M_{P_n,T_n}^\phi and sn=MPn,Tnψs_n=M_{P_n,T_n}^\psi.

Finite Sums and Soft Events

For a finite set AA and a function f:A→ℝf:A\to\mathbb R, the uniform mean 𝔼Af\mathbb E_A f is the real number defined in (9); for nonempty AA, its uniform probability measure UAU_A assigns every measurable subset B⊆AB\subseteq A the mass shown there: 𝔼Af={|A|−1∑x∈Af(x),A≠⌀,0,A=⌀,UA(B)=|B||A|(A≠⌀).\mathbb E_A f= \begin{cases}|A|^{-1}\sum_{x\in A}f(x),&A\ne\varnothing,\\ 0,&A=\varnothing, \end{cases} \qquad U_A(B)=\frac{|B|}{|A|}\quad(A\ne\varnothing).(9) For T:A→ℝT:A\to\mathbb R and n≥1n\ge1, the normalized sum Sn:An→ℝS_n:A^n\to\mathbb R is defined in (10); averages on product sets are uniform: Sn(x)=1n∑j=1nT(xj).S_n(x)=\frac{1}{\sqrt n}\sum_{j=1}^n T(x_j).(10)✦

For a real observable ZZ on a finite set BB and weights wi:B→ℝw_i:B\to\mathbb R, the finite weighted moment array RZ,wR_{Z,w} is defined by (11): RZ,w(k,i)=𝔼x∈B[wi(x)Z(x)k],k∈{0,1,2}.R_{Z,w}(k,i)=\mathbb E_{x\in B}\bigl[w_i(x)Z(x)^k\bigr], \qquad k\in\{0,1,2\}.(11)✦ The soft event weight associated with families si,wi:B→ℝs_i,w_i:B\to\mathbb R is the function qs,w:B×BI→ℝq_{s,w}:B\times B^I\to\mathbb R in (12): qs,w(x,y)=∑i∈Isi(x)∏j∈I\{i}wj(yj)−∏j∈Iwj(yj).q_{s,w}(x,y)=\sum_{i\in I}s_i(x)\prod_{j\in I\setminus\{i\}}w_j(y_j) -\prod_{j\in I}w_j(y_j).(12) For a real observable ZZ on a finite set BB, a real shift bb and a weight q:B×BI→ℝq:B\times B^I\to\mathbb R, the quadratic deficit DZ(b,q)D_Z(b,q) is the real number given by (13): DZ(b,q)=𝔼(x,y)∈B×BI[q(x,y)(1−(Z(x)+∑i∈IZ(yi)−b)2)].D_Z(b,q)=\mathbb E_{(x,y)\in B\times B^I} \left[q(x,y)\left(1-\left(Z(x)+\sum_{i\in I}Z(y_i)-b\right)^2\right)\right].(13) For the normalized sum, let wn,i=ϕi∘Snw_{n,i}=\phi_i\circ S_n and un,i=ψi∘Snu_{n,i}=\psi_i\circ S_n. The finite leaf moment array rnr_n, shift bnb_n and soft event weight qnq_n associated with SnS_n are given by (14): rn=RSn,wn,bn=∑i∈Iμi(rn)−1,qn=qun,wn.r_n=R_{S_n,w_n},\qquad b_n=\sum_{i\in I}\mu_i(r_n)-1,\qquad q_n=q_{u_n,w_n}.(14)

The Claim

Theorem 1. Let TnT_n converge in distribution under PnP_n to the standard normal law γ\gamma. For every continuous compactly supported function f:ℝ→ℝf:\mathbb R\to\mathbb R, the compact-test limit is (15): ∫Ωnf(Tn(x))dPn(x)→∫ℝf(t)dγ(t).\int_{\Omega_n}f(T_n(x))\,dP_n(x) \longrightarrow\int_{\mathbb R}f(t)\,d\gamma(t).(15)✦ For every k∈{0,1,2}k\in\{0,1,2\} and leaf index i∈Ii\in I, the leaf moment limit is (16): ℓn(k,i)→ℓG(k,i).\ell_n(k,i)\longrightarrow\ell^G(k,i).(16)✦

Suppose additionally that every TnT_n is measurable, that TnkT_n^k is integrable for k∈{0,1,2}k\in\{0,1,2\}, and that ∫TnkdPn→∫tkdγ(t)\int T_n^k\,dP_n\longrightarrow\int t^k\,d\gamma(t) for each such kk. For every k∈{0,1,2}k\in\{0,1,2\} and i∈Ii\in I, the selector moment limit is (17): sn(k,i)→sG(k,i).s_n(k,i)\longrightarrow s^G(k,i).(17)✦

On any finite index set, let λn,σn\lambda_n,\sigma_n be moment arrays converging entrywise to λ,σ\lambda,\sigma, respectively, with λ0,i≠0\lambda_{0,i}\ne0 at every index. Their gain limit is (18): 𝒢(λn,σn)→𝒢(λ,σ).\mathcal G(\lambda_n,\sigma_n)\longrightarrow\mathcal G(\lambda,\sigma).(18)✦

Under the additional moment hypotheses on TnT_n, there is an index n0n_0 for which (19) holds: ℓn(0,i)>0(i∈I),𝒢(ℓn,sn)>η(n≥n0).\ell_n(0,i)>0\quad(i\in I),\qquad \mathcal G(\ell_n,s_n)>\eta\qquad(n\ge n_0).(19)✦

Let AA be finite and nonempty, and let T:A→ℝT:A\to\mathbb R satisfy 𝔼AT=0\mathbb E_A T=0 and 𝔼AT2=1\mathbb E_A T^2=1. Its normalized sums have the moment limits (20), for k∈{0,1,2}k\in\{0,1,2\}: ∫AnSn(x)kdUAn(x)→∫ℝtkdγ(t).\int_{A^n} S_n(x)^k\,dU_{A^n}(x) \longrightarrow\int_{\mathbb R}t^k\,d\gamma(t).(20)✦ There is a positive integer nn for which the finite soft gain satisfies (21): 𝔼Anwn,i>0(i∈I),DSn(bn,qn)>η.\mathbb E_{A^n}w_{n,i}>0\quad(i\in I),\qquad D_{S_n}(b_n,q_n)>\eta.(21)✦

In the paper, the result reads: if centered variance-one variables ZjZ_j converge in distribution to a standard normal with sup⁡j𝔼Zj4<∞\sup_j\mathbb E Z_j^4<\infty, then for fixed rr, an rr-tuple 𝒁j\mathbf Z_j of independent copies and a standard Gaussian rr-tuple 𝑮\mathbf G satisfy 𝔼g(𝒁j)→𝔼g(𝑮)\mathbb E g(\mathbf Z_j)\to\mathbb E g(\mathbf G) for every measurable g:ℝr→ℝg:\mathbb R^r\to\mathbb R whose discontinuities are Gaussian-null and for which |g(z)|≤K(1+∑i=1rzi2)|g(z)|\le K(1+\sum_{i=1}^r z_i^2) with a fixed constant KK.

What Is Checked

The transfer statements concern observables on probability spaces and the fixed Gaussian weights, with the additional integrability and moment convergence hypotheses for selector tests, while (20) and (21) concern centered variance-one observables on finite uniform spaces.