OctoInformBeta

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

Finitely many slabs (2.4): A variance bound on soft slab laws gives |P(λ)|≥|P(0)|1−λ|P(1)|λ|P(\lambda)| \ge |P(0)|^{1-\lambda}|P(1)|^{\lambda} for slab bodies of finite volume with geometrically interpolated widths

From soft slab integrals to the volume of slab bodies.

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

Work in the Euclidean space ℝn\mathbb{R}^n, with the inner product ⟨x,y⟩=∑jxjyj\langle x,y\rangle = \sum_j x_j y_j and the norm |x||x|; |K||K| is the Lebesgue volume of KK. A slab body is the intersection of finitely many symmetric slabs |⟨pi,x⟩|≤hi|\langle p_i,x\rangle| \le h_i, with normals pip_i and positive widths hih_i. Fix the normals and two vectors of positive widths h0h^0 and h1h^1, and move the widths geometrically from h0h^0 to h1h^1: the question is how the volume of each intermediate body compares with the volumes of the two bodies at the ends.

The library states this volume inequality under a variance bound for soft slab laws, the Gibbs measures in which each slab becomes a smooth weight, together with the limits that lead from integrals of soft slabs back to volumes.

Slab Bodies

The slab body P(ℓ)P(\ell) of linear forms ℓ1,…,ℓm\ell_1,\dots,\ell_m on ℝn\mathbb{R}^n is the subset of ℝn\mathbb{R}^n P(ℓ)={x∈ℝn:|ℓi(x)|≤1 for i=1,…,m}.P(\ell) = \{x \in \mathbb{R}^n : |\ell_i(x)| \le 1 \text{ for } i = 1,\dots,m\}.(1)✦

For vectors p1,…,pmp_1,\dots,p_m in ℝn\mathbb{R}^n and positive widths h1,…,hmh_1,\dots,h_m, the normalized forms are ℓi(x)=⟨pi,x⟩/hi\ell_i(x) = \langle p_i,x\rangle/h_i.✦ The geometric interpolation of two vectors of positive widths h0=(h10,…,hm0)h^0 = (h_1^0,\dots,h_m^0) and h1=(h11,…,hm1)h^1 = (h_1^1,\dots,h_m^1) is the family of vectors h(t)=(h1(t),…,hm(t))h(t) = (h_1(t),\dots,h_m(t)) with entries hi(t)=(hi0)1−t(hi1)t,t∈ℝ,h_i(t) = (h_i^0)^{1-t}(h_i^1)^{t}, \qquad t \in \mathbb{R},(2)✦ and P(t)P(t) is the slab body of the normalized forms ⟨pi,x⟩/hi(t)\langle p_i,x\rangle/h_i(t), so that P(0)P(0) and P(1)P(1) have the widths h0h^0 and h1h^1.

A finite set of directions gives slab bodies of its own. The finite slab body PD[f]P_D[f] of a finite set DD of vectors in ℝn\mathbb{R}^n and a real function ff on ℝn\mathbb{R}^n is the subset of ℝn\mathbb{R}^n PD[f]={x∈ℝn:|⟨x,u⟩|≤f(u) for every u∈D}.P_D[f] = \{x \in \mathbb{R}^n : |\langle x,u\rangle| \le f(u) \text{ for every } u \in D\}.(3)✦

Soft Slabs

For linear forms ℓ1,…,ℓm\ell_1,\dots,\ell_m on ℝn\mathbb{R}^n, a real number ε\varepsilon and an integer k≥0k \ge 0, the soft slab density of ℓ\ell is the function on ℝn\mathbb{R}^n gε,kℓ(x)=exp⁡(−ε2|x|2−∑i=1mℓi(x)2k+2),g^{\ell}_{\varepsilon,k}(x) = \exp\left(-\frac{\varepsilon}{2}|x|^2 - \sum_{i=1}^m \ell_i(x)^{2k+2}\right),(4)✦ and Zε,k(ℓ)=∫ℝngε,kℓ(x)dxZ_{\varepsilon,k}(\ell) = \int_{\mathbb{R}^n} g^{\ell}_{\varepsilon,k}(x)\,dx is its integral. The Gaussian weight of parameter ε\varepsilon is the function γε\gamma_\varepsilon on ℝn\mathbb{R}^n given by γε(x)=e−(ε/2)|x|2\gamma_\varepsilon(x) = e^{-(\varepsilon/2)|x|^2}.✦

The Soft Slab Law

Let ψ1,…,ψm\psi_1,\dots,\psi_m be real polynomials on ℝn\mathbb{R}^n, b1,…,bmb_1,\dots,b_m real rates and ε\varepsilon a real number. The regularized potential VtV_t at the parameter t∈ℝt \in \mathbb{R} is the function on ℝn\mathbb{R}^n Vt(x)=ε2|x|2+∑i=1mebitψi(x),V_t(x) = \frac{\varepsilon}{2}|x|^2 + \sum_{i=1}^m e^{b_i t}\psi_i(x),(5)✦ the Gibbs density of VtV_t is the function e−Vte^{-V_t} on ℝn\mathbb{R}^n✦, and the partition function of VtV_t is the number Zt=∫ℝne−Vt(x)dxZ_t = \int_{\mathbb{R}^n} e^{-V_t(x)}\,dx✦. The Gibbs law of VtV_t is the probability measure μt=Zt−1e−Vtdx\mu_t = Z_t^{-1}e^{-V_t}\,dx on ℝn\mathbb{R}^n; the mean and the variance of a real function FF on ℝn\mathbb{R}^n under μt\mu_t are EμtF=1Zt∫ℝnF(x)e−Vt(x)dx,VarμtF=Eμt(F−EμtF)2.\mathrm{E}_{\mu_t}F = \frac{1}{Z_t}\int_{\mathbb{R}^n} F(x)\,e^{-V_t(x)}\,dx, \qquad \mathrm{Var}_{\mu_t}F = \mathrm{E}_{\mu_t}\left(F - \mathrm{E}_{\mu_t}F\right)^2.(6) The derivatives of the potential in tt are the polynomials ∂tjVt(x)=∑i=1mbijebitψi(x),j≥1.\partial_t^j V_t(x) = \sum_{i=1}^m b_i^j e^{b_i t}\psi_i(x), \qquad j \ge 1.(7)

For normals p1,…,pmp_1,\dots,p_m in ℝn\mathbb{R}^n, positive widths h0h^0, h1h^1, an integer k≥0k \ge 0, the even exponent q=2k+2q = 2k+2 and a real number ε\varepsilon, the soft slab law μt\mu_t is the Gibbs law of the polynomials ψi\psi_i and the rates bib_i given by ψi(x)=(⟨pi,x⟩hi0)q,bi=−qlog⁡hi1hi0,ebitψi(x)=(⟨pi,x⟩hi(t))q,\psi_i(x) = \left(\frac{\langle p_i,x\rangle}{h_i^0}\right)^{q}, \qquad b_i = -q\log\frac{h_i^1}{h_i^0}, \qquad e^{b_i t}\psi_i(x) = \left(\frac{\langle p_i,x\rangle}{h_i(t)}\right)^{q},(8) so that μt\mu_t is the soft slab law of the widths hi(t)h_i(t). The variance bound for the normals pip_i and the widths h0h^0, h1h^1 is the condition that for every ε>0\varepsilon > 0, every integer k≥0k \ge 0 and every t∈ℝt \in \mathbb{R}, Varμt(∂tVt)≤Eμt(∂t2Vt).\mathrm{Var}_{\mu_t}(\partial_t V_t) \le \mathrm{E}_{\mu_t}(\partial_t^2 V_t).(9)

The Claim

Theorem 1. (i) For linear forms ℓ1,…,ℓm\ell_1,\dots,\ell_m on ℝn\mathbb{R}^n and ε>0\varepsilon > 0, limk→∞Zε,k(ℓ)=∫P(ℓ)γε(x)dx.\lim_{k\to\infty} Z_{\varepsilon,k}(\ell) = \int_{P(\ell)} \gamma_\varepsilon(x)\,dx.(10)✦

(ii) For a set S⊂ℝnS \subset \mathbb{R}^n of finite volume and a sequence of numbers εk≥0\varepsilon_k \ge 0 tending to 00, limk→∞∫Sγεk(x)dx=|S|.\lim_{k\to\infty} \int_S \gamma_{\varepsilon_k}(x)\,dx = |S|.(11)✦

(iii) Let ℓ0\ell^0, ℓ1\ell^1 and ℓλ\ell^\lambda be three families of mm linear forms on ℝn\mathbb{R}^n whose slab bodies have finite volume, and let λ∈[0,1]\lambda \in [0,1]. If Zε,k(ℓ0)1−λZε,k(ℓ1)λ≤Zε,k(ℓλ)Z_{\varepsilon,k}(\ell^0)^{1-\lambda} Z_{\varepsilon,k}(\ell^1)^{\lambda} \le Z_{\varepsilon,k}(\ell^\lambda) for every ε>0\varepsilon > 0 and every integer k≥0k \ge 0, then |P(ℓ0)|1−λ|P(ℓ1)|λ≤|P(ℓλ)|.|P(\ell^0)|^{1-\lambda}\,|P(\ell^1)|^{\lambda} \le |P(\ell^\lambda)|.(12)✦

(iv) Let p1,…,pmp_1,\dots,p_m be vectors in ℝn\mathbb{R}^n, let h0h^0 and h1h^1 be positive widths, and let λ∈[0,1]\lambda \in [0,1]. If P(0)P(0), P(1)P(1) and P(λ)P(\lambda) have finite volume and the variance bound (9) holds, then |P(0)|1−λ|P(1)|λ≤|P(λ)|.|P(0)|^{1-\lambda}\,|P(1)|^{\lambda} \le |P(\lambda)|.(13)✦

(v) Let DD be a finite set of vectors in ℝn\mathbb{R}^n containing the coordinate vectors e1,…,ene_1,\dots,e_n, let f0f_0 and f1f_1 be real functions on ℝn\mathbb{R}^n that are positive on DD, and let λ∈[0,1]\lambda \in [0,1]. If the variance bound (9) holds for the normals u∈Du \in D with the widths f0(u)f_0(u) and f1(u)f_1(u), then |PD[f0]|1−λ|PD[f1]|λ≤|PD[f01−λf1λ]|.|P_D[f_0]|^{1-\lambda}\,|P_D[f_1]|^{\lambda} \le |P_D[f_0^{1-\lambda} f_1^{\lambda}]|.(14)✦

In the paper, the result reads: for nonzero vectors p1,…,pmp_1,\dots,p_m spanning ℝn\mathbb{R}^n, |P(λ)|≥|P(0)|1−λ|P(1)|λ|P(\lambda)| \ge |P(0)|^{1-\lambda}|P(1)|^{\lambda}, by letting q→∞q \to \infty (dominated convergence) and then ε→0\varepsilon \to 0 (monotone convergence).

What Is Checked

The claim is stated over the slab bodies (1) of linear forms and the finite slab bodies (3), the soft slab densities (4) with the Gaussian weights, and the soft slab laws (8) along the geometric widths (2), whose means, variances (6) and potential derivatives (7) make up the variance bound (9).