OctoInformBeta

OpenAI preprint, September 2026 · §7. The variance inequality (proof of Proposition 2.1)

The dual bound (7.6), density, and the conclusion: A dual bound for the tests A(A−1)fA(A-1)f, and Var⁡μ(∑ibiψi)≤𝔼μ∑ibi2ψi\operatorname{Var}_\mu(\sum_ib_i\psi_i)\le\mathbb{E}_\mu\sum_ib_i^2\psi_i for soft slab laws on ℝn\mathbb{R}^n, n≥2n\ge2

The dual bound and the soft slab variance inequality.

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

Fix ε>0\varepsilon>0, an even integer q≥2q\ge2, nonzero vectors p1,…,pmp_1,\dots,p_m of the coordinate space ℝn\mathbb{R}^n and widths h1,…,hm>0h_1,\dots,h_m>0. The features ψi\psi_i are the polynomials obtained as powers of the linear forms ⟨pi,x⟩\langle p_i,x\rangle; the soft slab potential VV adds them to a Gaussian term, and the soft slab law μ\mu is the probability measure with Gibbs density e−Ve^{-V} normalized by the partition function ZZ: ψi(x)=(⟨pi,x⟩hi)q,V(x)=ε2|x|2+∑i=1mψi(x),dμ(x)=e−V(x)Zdx,Z=∫ℝne−Vdx.\psi_i(x)=\Bigl(\frac{\langle p_i,x\rangle}{h_i}\Bigr)^{q},\qquad V(x)=\frac{\varepsilon}{2}|x|^2+\sum_{i=1}^m\psi_i(x),\qquad d\mu(x)=\frac{e^{-V(x)}}{Z}\,dx,\qquad Z=\int_{\mathbb{R}^n}e^{-V}\,dx.(1) Means and variances are taken under μ\mu: 𝔼μF=1Z∫ℝnFe−Vdx,Var⁡μF=𝔼μ(F−𝔼μF)2.\mathbb{E}_\mu F=\frac1Z\int_{\mathbb{R}^n}F\,e^{-V}\,dx,\qquad \operatorname{Var}_\mu F=\mathbb{E}_\mu\bigl(F-\mathbb{E}_\mu F\bigr)^2.(2)

The question is how much a combination ∑ibiψi\sum_ib_i\psi_i of the features, with real coefficients bib_i, can fluctuate under μ\mu, measured against the means 𝔼μψi\mathbb{E}_\mu\psi_i of the features themselves; Proposition 2.1 of the OpenAI preprint on the logarithmic Brunn-Minkowski conjecture answers it with a variance inequality. On the dual side, a function uu is tested against the features through its correlations 𝔼μuψi\mathbb{E}_\mu u\psi_i; the dual bound is the bound (9) on these correlations for the functions u=A(A−1)fu=A(A-1)f. Here ff is a test function, a smooth function with compact support, and AA is a second-order operator defined below from moment coordinates for μ\mu.

Moment Coordinates

The partial derivative ∂if\partial_if of a function ff on ℝn\mathbb{R}^n is its directional derivative DeifD_{e_i}f along the standard basis vector eie_i, where Dvf(x)=Df(x)vD_vf(x)=Df(x)\,v✦; the gradient ∇f\nabla f collects the ∂if\partial_if✦, and the Hessian D2fD^2f is the matrix with entries ∂j∂if\partial_j\partial_if.✦

Moment coordinates for μ\mu are a smooth function φ\varphi on ℝn\mathbb{R}^n and a homeomorphism Φ\Phi of ℝn\mathbb{R}^n with smooth inverse such that Φ=∇φ\Phi=\nabla\varphi; at every point zz the Hessian D2φ(z)D^2\varphi(z) is positive definite and (4/ε)I−D2φ(z)(4/\varepsilon)I-D^2\varphi(z) is positive semidefinite; φ\varphi solves the Monge-Ampère equation log⁡det⁡D2φ(z)=V(∇φ(z))+log⁡Z−φ(z);\log\det D^2\varphi(z)=V(\nabla\varphi(z))+\log Z-\varphi(z);(3) and ∇φ\nabla\varphi pushes the weighted source measure dν(z)=e−φ(z)dzd\nu(z)=e^{-\varphi(z)}\,dz forward to μ\mu.

The source variable zz carries ν\nu and the target variable x=Φ(z)x=\Phi(z) carries μ\mu; a function gg of zz is read in target coordinates as the target function g∘Φ−1g\circ\Phi^{-1}.✦ The moment matrix τ\tau of moment coordinates φ\varphi, Φ\Phi is the matrix-valued function on ℝn\mathbb{R}^n given by the Hessian of φ\varphi read in target coordinates, τ(x)=D2φ(Φ−1(x))\tau(x)=D^2\varphi(\Phi^{-1}(x)).✦

The Operator and Its Tests

An operator term TT on ℝn\mathbb{R}^n is a pair made of a smooth real function aa on ℝn\mathbb{R}^n, the coefficient, and a finite list of vectors v1,…,vkv_1,\dots,v_k of ℝn\mathbb{R}^n, the directions; it acts on a function ff as a∂v1⋯∂vkfa\,\partial_{v_1}\cdots\partial_{v_k}f, the coefficient times the iterated directional derivative, where ∂v\partial_v stands for DvD_v.✦ The differential operator of a finite list T1,…,TrT_1,\dots,T_r of terms is the operator on functions that sends ff to the sum of the actions of T1,…,TrT_1,\dots,T_r on ff.✦ The operator AA is the differential operator of the drift terms xi∂ix_i\,\partial_i, one for each coordinate, and the principal terms −τij∂i∂j-\tau_{ij}\,\partial_i\partial_j, one for each pair of coordinates: Af=x⋅∇f−τ:D2f=∑ixi∂if−∑i,jτij∂i∂jf.Af=x\cdot\nabla f-\tau:D^2f=\sum_ix_i\,\partial_if-\sum_{i,j}\tau_{ij}\,\partial_i\partial_jf.(4)

The test functions on ℝn\mathbb{R}^n form a linear space.✦ On it, A(A−1)A(A-1) sends ff to A(Af−f)A(Af-f)✦, and the range of A(A−1)A(A-1) in L2(μ)L^2(\mu) is the linear subspace ℛ={[A(Af−f)]:f a test function}⊂L2(μ),\mathcal{R}=\bigl\{\,[A(Af-f)]\;:\;f\text{ a test function}\,\bigr\}\subset L^2(\mu),(5)✦ where [g][g] is the class in L2(μ)L^2(\mu) of a test function gg.

Weighted Divergence and the Field WW

For a potential UU on ℝn\mathbb{R}^n, with weight e−Ue^{-U}, the weighted adjoint of DvD_v sends a function gg to Dv*g=DvUg−DvgD_v^*g=D_vU\,g-D_vg✦, and the weighted divergence of a vector field Y=(Y1,…,Yn)Y=(Y_1,\dots,Y_n) is δUY=∑jDej*Yj=∑j(∂jUYj−∂jYj).\delta_UY=\sum_jD_{e_j}^*Y_j=\sum_j\bigl(\partial_jU\,Y_j-\partial_jY_j\bigr).(6)✦ For U=VU=V the weighted divergence δUY\delta_UY is the divergence δμY\delta_\mu Y of μ\mu, and for U=φU=\varphi it is the divergence δνY\delta_\nu Y of ν\nu.

Given a test function ff, the moment coefficients PakP_{ak} are the entries of the matrix D2φ(z)D2f(∇φ(z))D^2\varphi(z)\,D^2f(\nabla\varphi(z)), Pak(z)=∑i∂a∂iφ(z)∂i∂kf(∇φ(z)).P_{ak}(z)=\sum_i\partial_a\partial_i\varphi(z)\,\partial_i\partial_kf(\nabla\varphi(z)).(7)✦ The moment divergence W̃a=δν(Pa1,…,Pan)\widetilde W_a=\delta_\nu(P_{a1},\dots,P_{an}) is the divergence under ν\nu of the aath row✦, and the transported field WW is its reading in target coordinates, Wa=W̃a∘Φ−1W_a=\widetilde W_a\circ\Phi^{-1}.✦ The function δμW\delta_\mu W is A(A−1)fA(A-1)f.

The Claim

Theorem 1. Let φ\varphi and Φ\Phi be moment coordinates for the soft slab law μ\mu. For every test function ff, the function δμW\delta_\mu W built from ff satisfies ∑i=1m(𝔼μ[ψiδμW])2𝔼μψi≤𝔼μ(δμW)2,\sum_{i=1}^m\frac{\bigl(\mathbb{E}_\mu[\psi_i\,\delta_\mu W]\bigr)^2}{\mathbb{E}_\mu\psi_i} \le\mathbb{E}_\mu(\delta_\mu W)^2,(8)✦ and every uu in the range ℛ\mathcal{R} of A(A−1)A(A-1) in L2(μ)L^2(\mu) satisfies ∑i=1m(𝔼μuψi)2𝔼μψi≤𝔼μu2.\sum_{i=1}^m\frac{(\mathbb{E}_\mu u\psi_i)^2}{\mathbb{E}_\mu\psi_i}\le\mathbb{E}_\mu u^2.(9)✦

Theorem 2. Let HH be a real inner product space, ψ1,…,ψm\psi_1,\dots,\psi_m vectors of HH and a1,…,am>0a_1,\dots,a_m>0. If ∑i⟨u,ψi⟩2/ai≤∥u∥2\sum_i\langle u,\psi_i\rangle^2/a_i\le\|u\|^2 for every uu in a dense subset of HH, then for all real b1,…,bmb_1,\dots,b_m ∥∑i=1mbiψi∥2≤∑i=1mbi2ai.\Bigl\|\sum_{i=1}^mb_i\psi_i\Bigr\|^2\le\sum_{i=1}^mb_i^2a_i.(10)✦ Let ρ\rho be a measure, ψ1,…,ψm\psi_1,\dots,\psi_m square integrable functions under ρ\rho, a1,…,am>0a_1,\dots,a_m>0 and b1,…,bmb_1,\dots,b_m real, and let F=∑ibiψiF=\sum_ib_i\psi_i, with centered function F‾=F−∫Fdρ\bar F=F-\int F\,d\rho. If ∑i(∫F‾ψidρ)2/ai≤∫F‾2dρ\sum_i\bigl(\int\bar F\psi_i\,d\rho\bigr)^2/a_i\le\int\bar F^2\,d\rho, then ∫F2dρ−(∫Fdρ)2≤∑i=1mbi2ai.\int F^2\,d\rho-\Bigl(\int F\,d\rho\Bigr)^2\le\sum_{i=1}^mb_i^2a_i.(11)✦

Theorem 3. Let μ\mu be the soft slab law (1) and b1,…,bmb_1,\dots,b_m real numbers. If φ\varphi and Φ\Phi are moment coordinates for μ\mu, then 𝔼μ(∑i=1mbiψi)2−(𝔼μ∑i=1mbiψi)2≤∑i=1mbi2𝔼μψi.\mathbb{E}_\mu\Bigl(\sum_{i=1}^mb_i\psi_i\Bigr)^2-\Bigl(\mathbb{E}_\mu\sum_{i=1}^mb_i\psi_i\Bigr)^2 \le\sum_{i=1}^mb_i^2\,\mathbb{E}_\mu\psi_i.(12)✦ For the soft slab law μ\mu on ℝn\mathbb{R}^n, n≥2n\ge2, and real numbers b1,…,bmb_1,\dots,b_m, Var⁡μ(∑i=1mbiψi)≤∑i=1mbi2𝔼μψi.\operatorname{Var}_\mu\Bigl(\sum_{i=1}^mb_i\psi_i\Bigr)\le\sum_{i=1}^mb_i^2\,\mathbb{E}_\mu\psi_i.(13)✦

What Is Checked

The statements are made over the soft slab law (1) with its features, mean and variance (2); over moment coordinates φ\varphi, Φ\Phi and the moment matrix τ\tau; over the operator AA of (4) on test functions, the range ℛ\mathcal{R} of A(A−1)A(A-1) in L2(μ)L^2(\mu), the weighted divergence (6) and the transported field WW; and, in (10) and (11), over any real inner product space and any measure.

About these constructions the library proves that τ\tau has smooth entries whenever φ\varphi and Φ−1\Phi^{-1} are smooth✦, and that AA is a linear map of the test functions into themselves.✦ It also proves that a test function is square integrable under any measure that is finite on compact sets.