OctoInformBeta

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

All directions and dimension one: For even continuous ff, slab bodies over dense directions recover W[f]W[f] and its volume; on the line the inequality is an equality

The log-Brunn-Minkowski inequality from slab bodies and a variance bound.

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

Theorem 1.1 of the preprint, the log-Brunn-Minkowski inequality, states that in every dimension n≥1n \ge 1, for origin-symmetric convex bodies KK and LL in ℝn\mathbb{R}^n and every λ∈[0,1]\lambda \in [0,1], the volume of the log combination K1−λLλK^{1-\lambda}L^{\lambda}, the Wulff body of the function u↦hK(u)1−λhL(u)λu \mapsto h_K(u)^{1-\lambda}h_L(u)^{\lambda} with hKh_K the support function of KK, is at least |K|1−λ|L|λ|K|^{1-\lambda}|L|^{\lambda}, all volumes |⋅||\cdot| being Lebesgue volumes taken in [0,∞][0,\infty]. Section 2 of the preprint, from a variance bound to volume, works with finitely many directions in place of the whole unit sphere.

How are a Wulff body and its volume recovered from the bodies cut out by finitely many symmetric slabs, as the directions fill the sphere? Which inequality for such bodies, or which variance bound, gives Theorem 1.1, and what does Theorem 1.1 say on the line? Every finite set of directions considered contains the coordinate vectors e1,…,ene_1,\dots,e_n of ℝn\mathbb{R}^n.

Slab Bodies and Wulff Bodies

The Euclidean space ℝn\mathbb{R}^n is the space of real nn-tuples with the inner product ⟨x,y⟩=∑ixiyi\langle x,y\rangle = \sum_i x_iy_i and the norm |x|=⟨x,x⟩1/2|x| = \langle x,x\rangle^{1/2}.✦ Its unit sphere is Sn−1={u∈ℝn:|u|=1}S^{n-1} = \{u \in \mathbb{R}^n : |u| = 1\}, and e1,…,ene_1,\dots,e_n are the coordinate vectors.✦ Volumes |⋅||\cdot| are Lebesgue volumes.

A convex body in ℝn\mathbb{R}^n is a compact convex subset of ℝn\mathbb{R}^n with nonempty interior.✦ A set K⊂ℝnK \subset \mathbb{R}^n is origin-symmetric if −x∈K-x \in K exactly when x∈Kx \in K.✦ The support function of a set K⊂ℝnK \subset \mathbb{R}^n is the function hKh_K on ℝn\mathbb{R}^n given by hK(u)=supx∈K⟨x,u⟩.h_K(u) = \sup_{x \in K} \langle x,u\rangle .(1)✦

The Wulff body of a real function ff on ℝn\mathbb{R}^n is the set W[f]={x∈ℝn:⟨x,u⟩≤f(u) for every u∈Sn−1}.W[f] = \{x \in \mathbb{R}^n : \langle x,u\rangle \le f(u) \text{ for every } u \in S^{n-1}\}.(2)✦ The interpolated support function of sets K,L⊂ℝnK, L \subset \mathbb{R}^n with a real parameter λ\lambda is the function hK1−λhLλh_K^{1-\lambda}h_L^{\lambda} on ℝn\mathbb{R}^n sending uu to hK(u)1−λhL(u)λh_K(u)^{1-\lambda}h_L(u)^{\lambda}.✦ The log combination of KK and LL with parameter λ\lambda is the Wulff body of the interpolated support function, K1−λLλ=W[hK1−λhLλ].K^{1-\lambda}L^{\lambda} = W[h_K^{1-\lambda}h_L^{\lambda}].(3)✦

The slab body of a finite set D⊂ℝnD \subset \mathbb{R}^n and a real function ff on ℝn\mathbb{R}^n is the intersection of one symmetric slab for each vector of DD, SD[f]={x∈ℝn:|⟨x,u⟩|≤f(u) for every u∈D}.S_D[f] = \{x \in \mathbb{R}^n : |\langle x,u\rangle| \le f(u) \text{ for every } u \in D\}.(4)✦ For a sequence (vk)k≥0(v_k)_{k \ge 0} in Sn−1S^{n-1} and an integer m≥0m \ge 0, the set DmD_m of test directions is the finite subset of ℝn\mathbb{R}^n made of the coordinate vectors and the first mm terms of the sequence, Dm={e1,…,en}∪{v0,…,vm−1}.D_m = \{e_1,\dots,e_n\} \cup \{v_0,\dots,v_{m-1}\}.(5)✦

The Soft Variance Inequality

The feature polynomial of a vector p∈ℝnp \in \mathbb{R}^n, a width hh and an exponent qq is the polynomial ψ(x)=(⟨p,x⟩/h)q\psi(x) = (\langle p,x\rangle/h)^q on ℝn\mathbb{R}^n.✦ Vectors p1,…,pN∈ℝnp_1,\dots,p_N \in \mathbb{R}^n and widths h1,…,hNh_1,\dots,h_N give the feature family ψ1,…,ψN\psi_1,\dots,\psi_N, with ψi\psi_i the feature polynomial of pip_i, hih_i and a common exponent qq.✦

With a regularization ε\varepsilon, the soft slab law of the feature family ψ1,…,ψN\psi_1,\dots,\psi_N is the probability measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}dx on ℝn\mathbb{R}^n, where the regularized potential is V(x)=ε2|x|2+∑i=1Nψi(x),V(x) = \frac{\varepsilon}{2}|x|^2 + \sum_{i=1}^N \psi_i(x),(6)✦ e−Ve^{-V} is the Gibbs density✦ and Z=∫ℝne−V(x)dxZ = \int_{\mathbb{R}^n} e^{-V(x)}\,dx is the partition function.✦

The mean 𝔼μF\mathbb{E}_\mu F and the variance VarμF\mathrm{Var}_\mu F of a real function FF on ℝn\mathbb{R}^n under μ\mu are 𝔼μF=1Z∫ℝnF(x)e−V(x)dx,VarμF=𝔼μ(F−𝔼μF)2.\mathbb{E}_\mu F = \frac{1}{Z}\int_{\mathbb{R}^n} F(x)\,e^{-V(x)}\,dx, \qquad \mathrm{Var}_\mu F = \mathbb{E}_\mu\big(F - \mathbb{E}_\mu F\big)^2 .(7)

Let n≥2n \ge 2 and N≥1N \ge 1, let q≥2q \ge 2 be even and ε>0\varepsilon > 0, let p1,…,pN∈ℝnp_1,\dots,p_N \in \mathbb{R}^n be nonzero vectors and h1,…,hNh_1,\dots,h_N positive widths, and let μ\mu be the soft slab law of the corresponding feature family. The soft variance inequality, Proposition 2.1 of the preprint, says that every choice of real coefficients b1,…,bNb_1,\dots,b_N satisfies Varμ(∑i=1Nbiψi)≤𝔼μ(∑i=1Nbi2ψi).\mathrm{Var}_\mu\Big(\sum_{i=1}^N b_i\psi_i\Big) \le \mathbb{E}_\mu\Big(\sum_{i=1}^N b_i^2\psi_i\Big).(8)✦

The Claim

Theorem 1. (a) Let (vk)k≥0(v_k)_{k \ge 0} be a dense sequence in Sn−1S^{n-1} with test directions DmD_m, and let ff be an even real function on ℝn\mathbb{R}^n whose restriction to Sn−1S^{n-1} is continuous. Then ⋂m≥0SDm[f]=W[f]\bigcap_{m \ge 0} S_{D_m}[f] = W[f](9)✦ and limm→∞|SDm[f]|=|W[f]|.\lim_{m \to \infty} |S_{D_m}[f]| = |W[f]| .(10)✦

(b) Let n≥1n \ge 1, let KK and LL be origin-symmetric convex bodies in ℝn\mathbb{R}^n, and let λ∈[0,1]\lambda \in [0,1]. Suppose that every finite set D⊂Sn−1D \subset S^{n-1} containing e1,…,ene_1,\dots,e_n satisfies |SD[hK]|1−λ|SD[hL]|λ≤|SD[hK1−λhLλ]|.|S_D[h_K]|^{1-\lambda}\,|S_D[h_L]|^{\lambda} \le |S_D[h_K^{1-\lambda}h_L^{\lambda}]| .(11) Then |K|1−λ|L|λ≤|K1−λLλ|.|K|^{1-\lambda}\,|L|^{\lambda} \le |K^{1-\lambda}L^{\lambda}| .(12)✦

(c) Let KK and LL be origin-symmetric convex bodies in ℝ\mathbb{R}, and let λ∈[0,1]\lambda \in [0,1]. Then |K1−λLλ|=|K|1−λ|L|λ.|K^{1-\lambda}L^{\lambda}| = |K|^{1-\lambda}\,|L|^{\lambda} .(13)✦

(d) Suppose the soft variance inequality (8) holds. Let n≥1n \ge 1, let KK and LL be origin-symmetric convex bodies in ℝn\mathbb{R}^n, and let λ∈[0,1]\lambda \in [0,1]. Then |K|1−λ|L|λ≤|K1−λLλ|.|K|^{1-\lambda}\,|L|^{\lambda} \le |K^{1-\lambda}L^{\lambda}| .(14)✦

What Is Checked

The claims (9) to (14) are stated over the Euclidean space ℝn\mathbb{R}^n and its unit sphere, the test directions DmD_m and the slab bodies SD[f]S_D[f], Wulff bodies, support functions and log combinations of origin-symmetric convex bodies, and the soft variance inequality (8) for the soft slab law of a feature family.