OctoInformBeta

OpenAI preprint, September 2026 · §2. Constructing the signed walk · §2

Lemma 2.5 [Gaussian correlation for path constraints]: A finite intersection of closed symmetric convex sets has at least the product of their probabilities under a linear image of a standard Gaussian

Gaussian correlation for symmetric convex sets in finite dimensions.

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

The construction of the signed walk runs a stationary Gaussian autoregressive chain Xt=rXt−1+1−r2ξtX_t=rX_{t-1}+\sqrt{1-r^2}\,\xi_t, taken coordinatewise in mm coordinates, over the steps 0,…,n0,\dots,n, driven by a standard Gaussian sample in ℝ(n+1)×m\mathbb{R}^{(n+1)\times m}, and needs a lower bound on the probability that its path satisfies many constraints at once. Each constraint asks that a linear image of the sample lie in a closed convex set symmetric about the origin, and the bound wanted is the product of the probabilities of the single constraints. For a centered Gaussian measure this is Royen’s Gaussian correlation inequality: symmetric convex events are positively correlated.

The library states the inequality in finite dimensions in two forms: for quadratic boxes, the sets cut out by one quadratic inequality in each coordinate, under a centered Gaussian measure with an arbitrary covariance matrix, and for closed symmetric convex sets under linear images of a standard Gaussian measure.

Gaussian Measures, Boxes and Constraint Sets

Let ι\iota be a finite index set. For a positive semidefinite ι×ι\iota\times\iota matrix Σ\Sigma, 𝒩(0,Σ)\mathcal{N}(0,\Sigma) is the centered Gaussian measure on ℝι\mathbb{R}^\iota with covariance matrix Σ\Sigma. For t∈ℝιt\in\mathbb{R}^\iota, the quadratic box BtB_t is the subset of ℝι\mathbb{R}^\iota Bt={x∈ℝι:xi2/2≤ti for every i∈ι},B_t=\{x\in\mathbb{R}^\iota : x_i^2/2\le t_i \text{ for every } i\in\iota\},(1)✦ and for a set S⊆ιS\subseteq\iota of coordinates, BtSB_t^S is the set of the x∈ℝιx\in\mathbb{R}^\iota with xi2/2≤tix_i^2/2\le t_i for every i∈Si\in S.

The standard Gaussian measure γE\gamma_E of a finite-dimensional real inner product space EE is the probability measure on EE under which the coordinates in an orthonormal basis are independent standard normal variables. For a linear map A:E→FA\colon E\to F into a real normed space FF, the image measure A#γEA_\#\gamma_E is the measure on FF with A#γE(K)=γE(A−1(K))A_\#\gamma_E(K)=\gamma_E(A^{-1}(K)). A subset KK of a real vector space is centrally symmetric when −x∈K-x\in K for every x∈Kx\in K.

For a family of continuous linear maps Ai:E→FA_i\colon E\to F between real normed spaces EE and FF and real numbers ρi\rho_i, indexed by a set II, the norm constraint set C(A,ρ)C(A,\rho) is the subset of EE C(A,ρ)={x∈E:∥Aix∥≤ρi for every i∈I}.C(A,\rho)=\{x\in E : \|A_i x\|\le\rho_i \text{ for every } i\in I\}.(2)✦

The Claim

Theorem 1. Let ι\iota be a finite index set, Σ\Sigma a positive semidefinite ι×ι\iota\times\iota matrix, t∈ℝιt\in\mathbb{R}^\iota and S⊆ιS\subseteq\iota. Then 𝒩(0,Σ)(BtS)𝒩(0,Σ)(Btι\S)≤𝒩(0,Σ)(Bt).\mathcal{N}(0,\Sigma)\bigl(B_t^S\bigr)\,\mathcal{N}(0,\Sigma)\bigl(B_t^{\iota\setminus S}\bigr) \le\mathcal{N}(0,\Sigma)(B_t).(3)✦ Let EE be a finite-dimensional real inner product space. If KK and LL are closed, convex, centrally symmetric subsets of EE, then γE(K)γE(L)≤γE(K∩L).\gamma_E(K)\,\gamma_E(L)\le\gamma_E(K\cap L).(4)✦ Let A:E→FA\colon E\to F be a linear map into a real normed space FF. If KK and LL are closed, convex, centrally symmetric subsets of FF, then A#γE(K)A#γE(L)≤A#γE(K∩L),A_\#\gamma_E(K)\,A_\#\gamma_E(L)\le A_\#\gamma_E(K\cap L),(5)✦ and if (Kj)j∈J(K_j)_{j\in J} is a finite family of closed, convex, centrally symmetric subsets of FF, then ∏j∈JA#γE(Kj)≤A#γE(⋂j∈JKj).\prod_{j\in J}A_\#\gamma_E(K_j)\le A_\#\gamma_E\Bigl(\bigcap_{j\in J}K_j\Bigr).(6)✦

In the paper, the result reads: if XX is a centered Gaussian process on [0,T][0,T] with continuous paths and each ℰi\mathcal{E}_i is a finite intersection of events {AijX(s)∈Kij for all s∈Iij}\{A_{ij}X(s)\in K_{ij}\text{ for all } s\in I_{ij}\}, with AijA_{ij} linear, KijK_{ij} closed, convex and centrally symmetric, and IijI_{ij} a closed interval, then ℙ(⋂iℰi)≥∏iℙ(ℰi)\mathbb{P}(\bigcap_i\mathcal{E}_i)\ge\prod_i\mathbb{P}(\mathcal{E}_i).

In discrete time, a path event of the chain is a finite intersection of norm constraint sets (2) in its standard Gaussian sample, to which (6) applies.

What Is Checked

The inequalities (3) to (6) are stated over the centered Gaussian measure 𝒩(0,Σ)\mathcal{N}(0,\Sigma) of a positive semidefinite matrix Σ\Sigma and the quadratic boxes (1), and over the standard Gaussian measure γE\gamma_E of a finite-dimensional real inner product space, its images A#γEA_\#\gamma_E under linear maps, and closed, convex, centrally symmetric sets.