OctoInformBeta

OpenAI preprint, September 2026 · §2. A low-degree estimate on biased products

§2: The product noise estimate: A coefficient estimate and a function estimate on a finite biased cube

Product noise controls fourth moments.

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

Consider a finite coordinate set II and the cube {0,1}I\{0,1\}^I, with independent Bernoulli coordinates of common bias 0<p<10<p<1. The product noise estimate concerns the fourth moment of a Fourier sum after its coefficients are damped according to their degree. The library states it both for arbitrary real coefficient families and for the Fourier coefficients of a real function.

The Bernoulli mass mpm_p on {0,1}\{0,1\} assigns probability pp to 11, and its product probability mass wpw_p on the cube is given by (1): mp(b)={p,b=1,1−p,b=0,wp(x)=∏i∈Imp(xi).m_p(b)= \begin{cases} p,&b=1,\\ 1-p,&b=0, \end{cases} \qquad w_p(x)=\prod_{i\in I}m_p(x_i).(1) For a real function hh on the cube, its expectation 𝔼ph\mathbb E_p h and its LqL^q norm ∥h∥q\|h\|_q under this product measure are defined by (2): 𝔼ph=∑x∈{0,1}Iwp(x)h(x),∥h∥q=(𝔼p|h|q)1/q(q=2,4).\mathbb E_p h=\sum_{x\in\{0,1\}^I}w_p(x)h(x), \qquad \|h\|_q=\bigl(\mathbb E_p|h|^q\bigr)^{1/q} \quad(q=2,4).(2)

Put σ=p(1−p)\sigma=\sqrt{p(1-p)}. The biased coordinate character χi\chi_i on the cube is the centered coordinate divided by σ\sigma, and the product character χS\chi_S indexed by S⊆IS\subseteq I is the product in (3): χi(x)=xi−pσ,χS(x)=∏i∈Sχi(x).\chi_i(x)=\frac{x_i-p}{\sigma}, \qquad \chi_S(x)=\prod_{i\in S}\chi_i(x).(3) The biased Fourier coefficient gˆ(S)\widehat g(S) of a real function gg on the cube is its expected product with χS\chi_S, as in (4): gˆ(S)=𝔼p[gχS]=∑x∈{0,1}Iwp(x)g(x)χS(x).\widehat g(S)=\mathbb E_p[g\chi_S] =\sum_{x\in\{0,1\}^I}w_p(x)g(x)\chi_S(x).(4)✦

For a real coefficient family c=(cS)S⊆Ic=(c_S)_{S\subseteq I} and a real parameter ρ\rho, the damped Fourier synthesis AρcA_\rho c is a real function on the cube, and the product noise operator TρT_\rho sends a real function gg on the cube to the real function TρgT_\rho g, as defined in (5): (Aρc)(x)=∑S⊆Iρ|S|cSχS(x),(Tρg)(x)=∑S⊆Iρ|S|gˆ(S)χS(x).(A_\rho c)(x)=\sum_{S\subseteq I}\rho^{|S|}c_S\chi_S(x), \qquad (T_\rho g)(x)=\sum_{S\subseteq I}\rho^{|S|}\widehat g(S)\chi_S(x).(5)

The Claim

Theorem 1. Let II be finite, let 0<p<10<p<1, and set ρ=σ/4\rho=\sigma/4. For every real coefficient family c=(cS)S⊆Ic=(c_S)_{S\subseteq I}, the coefficient estimate is (6): 𝔼p[(Aρc)4]≤(∑S⊆IcS2)2.\mathbb E_p\bigl[(A_\rho c)^4\bigr] \leq \left(\sum_{S\subseteq I}c_S^2\right)^2.(6)✦ For every real function gg on {0,1}I\{0,1\}^I, the noise estimate is (7): 𝔼p[(Tρg)4]≤(𝔼p[g2])2.\mathbb E_p\bigl[(T_\rho g)^4\bigr] \leq \bigl(\mathbb E_p[g^2]\bigr)^2.(7)✦

In the paper, the result reads: for ρ=σ/4\rho=\sigma/4, TρχS=ρ|S|χST_\rho\chi_S=\rho^{|S|}\chi_S, and every real function gg on a finite biased cube satisfies ∥Tρg∥4≤∥g∥2\|T_\rho g\|_4\leq\|g\|_2.

What Is Checked

The coefficient estimate (6) ranges over all real families indexed by subsets of a finite coordinate set, and the noise estimate (7) ranges over all real functions on its Boolean cube, in each case at every bias 0<p<10<p<1 with the product masses and characters defined above.