OctoInformBeta

OpenAI preprint, September 2026 · §2. A one-sided coupling estimate

Lemma 2: For U,VU,V with the same law on a finite probability space and V−U≤aV-U\le a, 𝔼(U−V)2≤8a𝔼(U−m)2\mathbb{E}(U-V)^2\le 8a\sqrt{\mathbb{E}(U-m)^2} for every real mm

Equal laws and one-sided steps bound the mean square gap.

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

Let UU and VV be real random variables on the same finite probability space, with the same distribution. For a real function FF on the cube Ω={−1,1}n\Omega = \{-1,1\}^n with the uniform measure and a coordinate ii, U(x)=F(x)U(x) = F(x) and V(x)=F(x⊕i)V(x) = F(x^{\oplus i}), where x⊕ix^{\oplus i} is xx with its ii-th coordinate reversed, are such a pair. Suppose that VV exceeds UU by at most a≥0a \ge 0 wherever the probability is positive, with no bound on U−VU - V. The question is how large the mean square of U−VU - V can be in terms of aa and the spread of UU, measured by its second moment σ2\sigma^2 about a point mm.

Definitions

A finite probability space (Λ,μ)(\Lambda, \mu) is a finite set Λ\Lambda with a weight μ(λ)≥0\mu(\lambda) \ge 0 at each point λ∈Λ\lambda \in \Lambda, the weights summing to 11. A real random variable YY on a finite probability space (Λ,μ)(\Lambda, \mu) is a function Y:Λ→ℝY \colon \Lambda \to \mathbb{R}, and the expectation of YY is 𝔼Y=∑λ∈Λμ(λ)Y(λ).\mathbb{E}\,Y = \sum_{\lambda \in \Lambda} \mu(\lambda)\,Y(\lambda).(1)

Two real random variables UU and VV on Λ\Lambda have the same distribution if 𝔼φ(U)=𝔼φ(V)\mathbb{E}\,\varphi(U) = \mathbb{E}\,\varphi(V) for every function φ:ℝ→ℝ\varphi \colon \mathbb{R} \to \mathbb{R}. An inequality between random variables on Λ\Lambda holds almost surely if it holds at every point λ\lambda with μ(λ)>0\mu(\lambda) > 0. The second moment of UU about a point m∈ℝm \in \mathbb{R} is 𝔼(U−m)2\mathbb{E}\,(U - m)^2, the variance of UU when m=𝔼Um = \mathbb{E}\,U.

For a point m∈ℝm \in \mathbb{R}, the coupling potential about mm is the function g:ℝ→ℝg \colon \mathbb{R} \to \mathbb{R} given by g(t)=12(t−m)|t−m|.g(t) = \tfrac{1}{2}\,(t - m)\,|t - m|.(2)✦

The Claim

Theorem 1. (a) Let (Λ,μ)(\Lambda, \mu) be a finite probability space and let UU and VV be real random variables on it with the same distribution. Let m∈ℝm \in \mathbb{R}, σ≥0\sigma \ge 0 and a≥0a \ge 0, and suppose that the second moment of UU about mm is σ2\sigma^2 and that V−U≤aV - U \le a almost surely. Then 𝔼(U−V)2≤8aσ.\mathbb{E}\,(U - V)^2 \le 8a\sigma.(3)✦

(b) Let m∈ℝm \in \mathbb{R} and let gg be the coupling potential about mm. For all real numbers u≤vu \le v, (v−u)24≤g(v)−g(u)≤v−u2(|u−m|+|v−m|),\frac{(v - u)^2}{4} \le g(v) - g(u) \le \frac{v - u}{2}\,\bigl(|u - m| + |v - m|\bigr),(4)✦ and for all real numbers uu and vv, (u−v)2≤4|g(u)−g(v)|.(u - v)^2 \le 4\,|g(u) - g(v)|.(5)✦

In the paper, the result reads: if UU and VV are real random variables with the same distribution, finite mean mm and finite variance σ2\sigma^2, and V−U≤aV - U \le a almost surely, where a≥0a \ge 0, then 𝔼(U−V)2≤8aσ\mathbb{E}\,(U - V)^2 \le 8a\sigma.

What Is Checked

The claim is stated over a finite probability space given by nonnegative weights summing to one, two real random variables on it whose expectations 𝔼φ(U)\mathbb{E}\,\varphi(U) and 𝔼φ(V)\mathbb{E}\,\varphi(V) agree for every function φ:ℝ→ℝ\varphi \colon \mathbb{R} \to \mathbb{R}, and the coupling potential (2).