OctoInformBeta

OpenAI preprint, September 2026 · §2. From a variance bound to volume

Proposition 2.1: Var⁡μ(∑ibiψi)≤∑ibi2𝔼μψi\operatorname{Var}_\mu(\sum_i b_i\psi_i)\le\sum_i b_i^2\,\mathbb{E}_\mu\psi_i for n≥2n\ge 2, and in every dimension for μ\mu with moment coordinates

A variance bound for the soft slab law.

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\ge 2, nonzero vectors p1,…,pmp_1,\dots,p_m in the coordinate space ℝn\mathbb{R}^n of real nn-tuples, and widths h1,…,hm>0h_1,\dots,h_m>0. Each pair (pi,hi)(p_i,h_i) cuts out the slab {x:|⟨pi,x⟩|≤hi}\{x : |\langle p_i,x\rangle|\le h_i\}, and the mm slabs meet in a symmetric convex polyhedron. The feature of the iith slab is the polynomial ψi(x)=(⟨pi,x⟩hi)q,⟨pi,x⟩=∑j=1npijxj,\psi_i(x)=\Bigl(\frac{\langle p_i,x\rangle}{h_i}\Bigr)^{q},\qquad \langle p_i,x\rangle=\sum_{j=1}^n p_{ij}\,x_j,(1) and e−ψie^{-\psi_i} is a soft indicator of the slab, sharper as qq grows.

The soft slab law is the measure μ\mu on ℝn\mathbb{R}^n whose density is proportional to the product of these soft indicators with the Gaussian factor e−ε|x|2/2e^{-\varepsilon|x|^2/2}: its soft slab potential VV, its partition function ZZ and μ\mu itself are V(x)=ε2|x|2+∑i=1mψi(x),Z=∫ℝne−V(x)dx,μ=Z−1e−V(x)dx.V(x)=\frac{\varepsilon}{2}|x|^2+\sum_{i=1}^m\psi_i(x),\qquad Z=\int_{\mathbb{R}^n}e^{-V(x)}\,dx,\qquad \mu=Z^{-1}e^{-V(x)}\,dx.(2)

The question is how much a linear combination ∑ibiψi\sum_i b_i\psi_i of the features, with real coefficients bib_i, can fluctuate under μ\mu, measured against the means of the features. The library bounds this fluctuation for n≥2n\ge 2, and in every dimension when μ\mu has moment coordinates, a notion defined in the section Moment Coordinates below.

Means and Variances

The mean 𝔼μF\mathbb{E}_\mu F and the variance Var⁡μF\operatorname{Var}_\mu F of a real function FF on ℝn\mathbb{R}^n under μ\mu are the numbers 𝔼μF=∫Fdμ=1Z∫ℝnF(x)e−V(x)dx,Var⁡μF=𝔼μ(F−𝔼μF)2.\mathbb{E}_\mu F=\int F\,d\mu=\frac{1}{Z}\int_{\mathbb{R}^n}F(x)\,e^{-V(x)}\,dx,\qquad \operatorname{Var}_\mu F=\mathbb{E}_\mu\bigl(F-\mathbb{E}_\mu F\bigr)^2.(3)

For real polynomials P1,…,PmP_1,\dots,P_m on ℝn\mathbb{R}^n, real rates r1,…,rmr_1,\dots,r_m and a real parameter tt, the regularized potential is the function x↦ε2|x|2+∑ieritPi(x)x\mapsto\frac{\varepsilon}{2}|x|^2+\sum_i e^{r_it}P_i(x), and its Gibbs density, partition function, Gibbs mean and Gibbs variance are formed from it as e−Ve^{-V}, ZZ, 𝔼μ\mathbb{E}_\mu and Var⁡μ\operatorname{Var}_\mu are formed from VV. For Pi=ψiP_i=\psi_i and all rates zero, the Gibbs density, partition function, Gibbs mean and Gibbs variance of the regularized potential are e−Ve^{-V}, ZZ, 𝔼μ\mathbb{E}_\mu and Var⁡μ\operatorname{Var}_\mu.

For a nonnegative integer kk, the derivative polynomial of order kk of the polynomials PiP_i with rates rir_i and parameter tt is the polynomial ∑irikeritPi\sum_i r_i^k e^{r_it}P_i on ℝn\mathbb{R}^n, the kkth derivative in tt of ∑ieritPi\sum_i e^{r_it}P_i.✦ The polynomial variation of order kk at a point xx is the value at xx of the derivative polynomial of order kk.✦

Moment Coordinates

For a function ff on ℝn\mathbb{R}^n and a vector vv, the directional derivative Dvf(x)D_vf(x) is the derivative of ff at xx applied to vv✦, and the partial derivative ∂if=Deif\partial_if=D_{e_i}f is taken along the iith standard basis vector eie_i✦. The gradient of ff at xx is the vector ∇f(x)=(∂1f(x),…,∂nf(x))\nabla f(x)=(\partial_1f(x),\dots,\partial_nf(x))✦, and the Hessian of ff at xx is the matrix D2f(x)=(∂j∂if(x))i,j=1nD^2f(x)=(\partial_j\partial_if(x))_{i,j=1}^n✦.

The weighted measure of a function φ\varphi on ℝn\mathbb{R}^n is ν=e−φ(z)dz\nu=e^{-\varphi(z)}\,dz, Lebesgue measure in the source coordinates zz with the weight e−φe^{-\varphi}.✦ The function φ\varphi solves the Monge-Ampère equation for μ\mu when log⁡det⁡D2φ(z)=V(∇φ(z))+log⁡Z−φ(z)for every z∈ℝn.\log\det D^2\varphi(z)=V\bigl(\nabla\varphi(z)\bigr)+\log Z-\varphi(z) \qquad\text{for every } z\in\mathbb{R}^n.(4) A smooth function φ\varphi on ℝn\mathbb{R}^n gives moment coordinates for μ\mu when it solves (4), its 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 at every point zz, and its gradient map z↦x=∇φ(z)z\mapsto x=\nabla\varphi(z) is a homeomorphism of ℝn\mathbb{R}^n onto itself, with smooth inverse, that pushes ν\nu forward to μ\mu, making μ\mu the moment measure of φ\varphi in the target coordinates xx.

The Claim

Theorem 1 (Variance bound). Let ε>0\varepsilon>0, let q≥2q\ge 2 be even, let p1,…,pmp_1,\dots,p_m be nonzero vectors in ℝn\mathbb{R}^n and h1,…,hmh_1,\dots,h_m positive widths, and let ψi\psi_i and μ\mu be the features and the soft slab law of (1) and (2). Let b1,…,bmb_1,\dots,b_m be real numbers.

  1. (Soft variance inequality.) If n≥2n\ge 2 and m≥1m\ge 1, then Var⁡μ(∑i=1mbiψi)≤𝔼μ(∑i=1mbi2ψi).\operatorname{Var}_\mu\Bigl(\sum_{i=1}^m b_i\psi_i\Bigr) \le\mathbb{E}_\mu\Bigl(\sum_{i=1}^m b_i^2\psi_i\Bigr).(5)✦

  2. (Moment variance bound.) If a function φ\varphi on ℝn\mathbb{R}^n gives moment coordinates for μ\mu, then 𝔼μ[(∑i=1mbiψi)2]−(𝔼μ∑i=1mbiψi)2≤∑i=1mbi2𝔼μψi.\mathbb{E}_\mu\Bigl[\Bigl(\sum_{i=1}^m b_i\psi_i\Bigr)^2\Bigr] -\Bigl(\mathbb{E}_\mu\sum_{i=1}^m b_i\psi_i\Bigr)^2 \le\sum_{i=1}^m b_i^2\,\mathbb{E}_\mu\psi_i.(6)✦

What Is Checked

Both bounds are stated over the features ψi\psi_i and the linear forms ⟨pi,x⟩\langle p_i,x\rangle of (1); the soft variance inequality (5) is stated over the Gibbs mean and Gibbs variance of the regularized potential with all rates zero, with its Gibbs density, partition function and polynomial variation, and the moment variance bound (6) over the soft slab potential VV, the law μ\mu, and a function φ\varphi giving moment coordinates, with its gradient, Hessian and weighted measure ν\nu.