OctoInformBeta

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

The soft slab potential and its law (2.1): The potential V(x)=ε2|x|2+∑i(⟨pi,x⟩/hi)qV(x)=\frac{\varepsilon}{2}|x|^2+\sum_i(\langle p_i,x\rangle/h_i)^q, its law μ=Z−1e−Vdx\mu=Z^{-1}e^{-V}dx, and mean and variance when weights ebite^{b_i t} move the widths

The soft slab potential and its Gibbs 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

A slab in ℝn\mathbb{R}^n is a region {x:|⟨p,x⟩|≤h}\{x : |\langle p,x\rangle| \le h\} between two parallel hyperplanes, given by a direction pp and a width h>0h > 0. The slabs of directions p1,…,pmp_1, \dots, p_m and widths h1,…,hmh_1, \dots, h_m meet in the symmetric convex set P={x:|⟨pi,x⟩|≤hi for every i}P = \{x : |\langle p_i,x\rangle| \le h_i \text{ for every } i\}, a polytope when the pip_i span ℝn\mathbb{R}^n.

For an even exponent q≥2q \ge 2, the soft slab law multiplies the factors e−(⟨pi,x⟩/hi)qe^{-(\langle p_i,x\rangle/h_i)^q} with a Gaussian factor e−ε|x|2/2e^{-\varepsilon|x|^2/2}, ε>0\varepsilon > 0, and normalizes the product to a log-concave probability measure on ℝn\mathbb{R}^n.

Given positive widths hi0h_i^0 and hi1h_i^1, the widths move with a parameter t∈[0,1]t \in [0,1] along hi(t)=(hi0)1−t(hi1)th_i(t) = (h_i^0)^{1-t}(h_i^1)^t, and μt\mu_t is the soft slab law at the widths hi(t)h_i(t). At width hi(t)h_i(t) the term (⟨pi,x⟩/hi(t))q(\langle p_i,x\rangle/h_i(t))^q is (⟨pi,x⟩/hi0)q(\langle p_i,x\rangle/h_i^0)^q multiplied by ebite^{b_i t}, where bi=qlog⁡(hi0/hi1)b_i = q\log(h_i^0/h_i^1). The partition function and the mean and variance under μt\mu_t are taken for the potential in which each term is a polynomial ψi\psi_i multiplied by a weight ebite^{b_i t}.

Linear Forms and Weighted Sums

Points of ℝn\mathbb{R}^n are written by their coordinates x=(x1,…,xn)x = (x_1, \dots, x_n), and |x|2=x12+…+xn2|x|^2 = x_1^2 + \dots + x_n^2. For a vector p∈ℝnp \in \mathbb{R}^n, the linear functional ⟨p,⋅⟩\langle p,\cdot\rangle is the continuous linear map from ℝn\mathbb{R}^n to ℝ\mathbb{R} given by ⟨p,x⟩=p1x1+…+pnxn\langle p,x\rangle = p_1x_1 + \dots + p_nx_n.✦ The linear polynomial of pp is the polynomial ℓp=p1X1+…+pnXn\ell_p = p_1X_1 + \dots + p_nX_n of ℝ[X1,…,Xn]\mathbb{R}[X_1, \dots, X_n], whose value at xx is ⟨p,x⟩\langle p,x\rangle.✦

For polynomials ψ1,…,ψm∈ℝ[X1,…,Xn]\psi_1, \dots, \psi_m \in \mathbb{R}[X_1, \dots, X_n], real weights b1,…,bmb_1, \dots, b_m, a real parameter tt and a natural number kk, the kk-th derivative polynomial Ψt(k)\Psi^{(k)}_t is the polynomial Ψt(k)=∑i=1mbikebitψi,\Psi^{(k)}_t = \sum_{i=1}^m b_i^k\, e^{b_i t}\, \psi_i ,(1)✦ the kk-th derivative in tt of Ψt(0)=∑iebitψi\Psi^{(0)}_t = \sum_i e^{b_i t}\psi_i. Its value at a point xx of ℝn\mathbb{R}^n is written Ψt(k)(x)\Psi^{(k)}_t(x).✦

The Definition

Definition 1. Let ε\varepsilon be a real number, qq a natural number, p,p1,…,pmp, p_1, \dots, p_m vectors of ℝn\mathbb{R}^n and h,h1,…,hmh, h_1, \dots, h_m real numbers.

The slab summand of pp and hh is the function σp,h\sigma_{p,h} on ℝn\mathbb{R}^n with σp,h(x)=(⟨p,x⟩h)q.\sigma_{p,h}(x) = \Bigl(\frac{\langle p,x\rangle}{h}\Bigr)^q .(2)✦ The soft slab potential of the directions pip_i and the widths hih_i is the function V(x)=ε2|x|2+∑i=1mσpi,hi(x),V(x) = \frac{\varepsilon}{2}|x|^2 + \sum_{i=1}^m \sigma_{p_i,h_i}(x),(3)✦ and the soft slab law is the measure μ\mu on ℝn\mathbb{R}^n with dμ(x)=e−V(x)Zdx,Z=∫ℝne−V(y)dy.d\mu(x) = \frac{e^{-V(x)}}{Z}\,dx, \qquad Z = \int_{\mathbb{R}^n} e^{-V(y)}\,dy .(4)✦

The feature polynomial of pp and hh is the polynomial ψp,h=(h−1ℓp)q∈ℝ[X1,…,Xn],\psi_{p,h} = \bigl(h^{-1}\ell_p\bigr)^q \in \mathbb{R}[X_1, \dots, X_n],(5)✦ and the slab features of the directions pip_i and the widths hih_i form the family of polynomials ψi=ψpi,hi,i=1,…,m.\psi_i = \psi_{p_i,h_i}, \qquad i = 1, \dots, m .(6)✦

For polynomials ψ1,…,ψm∈ℝ[X1,…,Xn]\psi_1, \dots, \psi_m \in \mathbb{R}[X_1, \dots, X_n] (the slab features, for instance), real weights b1,…,bmb_1, \dots, b_m and a real parameter tt, the regularized potential VtV_t is the function Vt(x)=ε2|x|2+Ψt(0)(x).V_t(x) = \frac{\varepsilon}{2}|x|^2 + \Psi^{(0)}_t(x).(7)✦ The Gibbs density ρt\rho_t of the regularized potential VtV_t is ρt(x)=e−Vt(x).\rho_t(x) = e^{-V_t(x)}.(8)✦ The partition function Z(t)Z(t) of the regularized potential VtV_t is Z(t)=∫ℝnρt(x)dx.Z(t) = \int_{\mathbb{R}^n} \rho_t(x)\,dx.(9)✦ For a function F:ℝn→ℝF : \mathbb{R}^n \to \mathbb{R}, the mean 𝔼tF\mathbb{E}_t F and the variance Var⁡tF\operatorname{Var}_t F of FF under the Gibbs density ρt\rho_t are 𝔼tF=1Z(t)∫ℝnF(x)ρt(x)dx,\mathbb{E}_t F = \frac{1}{Z(t)} \int_{\mathbb{R}^n} F(x)\,\rho_t(x)\,dx,(10)✦ Var⁡tF=𝔼t[(F−𝔼tF)2].\operatorname{Var}_t F = \mathbb{E}_t\bigl[(F - \mathbb{E}_t F)^2\bigr].(11)✦

In the paper, the definition reads: ψi(x)=(⟨pi,x⟩/hi)q\psi_i(x) = (\langle p_i,x\rangle/h_i)^q, V(x)=ε2|x|2+∑iψi(x)V(x) = \frac{\varepsilon}{2}|x|^2 + \sum_i \psi_i(x) and dμ=Z−1e−Vdxd\mu = Z^{-1}e^{-V}\,dx, with 𝔼μ\mathbb{E}_\mu and Var⁡μ\operatorname{Var}_\mu the mean and the variance under μ\mu, the paper’s ψi(x)\psi_i(x) being the page’s σpi,hi(x)\sigma_{p_i,h_i}(x).

When the ψi\psi_i are the slab features (6) and t=0t = 0, the regularized potential V0V_0 is the soft slab potential VV of (3), and 𝔼0\mathbb{E}_0 and Var⁡0\operatorname{Var}_0 are the mean and the variance under the soft slab law μ\mu of (4).

What Is Checked

The definitions are stated over the coordinate space ℝn\mathbb{R}^n, for a real ε\varepsilon, vectors pip_i, real widths hih_i and a natural exponent qq, and, for the regularized potential VtV_t, for arbitrary real polynomials ψi\psi_i in nn variables, real weights bib_i, a real parameter tt and real functions FF on ℝn\mathbb{R}^n.