OctoInformBeta

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

Log-concavity of the partition function (2.3): Hence Z(t)≥Z(0)1−tZ(1)tZ(t)\ge Z(0)^{1-t}Z(1)^t on [0,1][0,1]; for slab features of even degree with widths moving geometrically, the soft variance inequality supplies the bound

A variance bound makes the partition function log-concave.

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

Finitely many nonzero vectors pi∈ℝnp_i\in\mathbb{R}^n and widths hi>0h_i>0 cut out a symmetric polyhedron, the intersection of the slabs |⟨pi,x⟩|≤hi|\langle p_i,x\rangle|\le h_i. The soft slab law replaces its indicator by a Gibbs weight. Each slab contributes a factor e−ψie^{-\psi_i}, with ψi\psi_i a high even power of ⟨pi,x⟩/hi\langle p_i,x\rangle/h_i, so that the factor is close to 11 inside the slab and to 00 outside, and a Gaussian factor e−ε|x|2/2e^{-\varepsilon|x|^2/2} with a fixed ε>0\varepsilon>0 makes the weight integrable. Its total mass ZZ is a smoothed volume of the polyhedron.

Let the widths move geometrically from hi0h_i^0 at t=0t=0 to hi1h_i^1 at t=1t=1, as support values do in the logarithmic Brunn-Minkowski inequality. Is the smoothed volume Z(t)Z(t) log-concave in tt? The answer passes through a general Gibbs family, whose potential depends on tt through exponential weights ebite^{b_it} with real rates bib_i on fixed nonnegative polynomials, and a bound of a variance by a mean along that family, which for slab features is the soft variance inequality.

The Gibbs Family

Fix ε>0\varepsilon>0, a finite set II, polynomials ψi\psi_i on ℝn\mathbb{R}^n and real numbers bib_i, the rates, for i∈Ii\in I; the Gibbs family determined by these data is the family of probability measures μt\mu_t, indexed by a real parameter tt, defined below. For an integer k≥0k\ge0 and a real parameter tt, the variation of order kk is the polynomial Vt(k)(x)=∑i∈Ibikebitψi(x),V_t^{(k)}(x)=\sum_{i\in I}b_i^k\,e^{b_it}\,\psi_i(x),(1)✦ the kkth derivative in tt of ∑iebitψi\sum_i e^{b_it}\psi_i. The regularized potential of the family is the function VtV_t on ℝn\mathbb{R}^n given by Vt(x)=ε2|x|2+Vt(0)(x),V_t(x)=\frac{\varepsilon}{2}\,|x|^2+V_t^{(0)}(x),(2)✦ and we write V̇t=Vt(1)\dot V_t=V_t^{(1)} and V̈t=Vt(2)\ddot V_t=V_t^{(2)} for its first two derivatives in tt.

The Gibbs density of the family at tt is the function e−Vte^{-V_t} on ℝn\mathbb{R}^n✦, and its partition function is the function ZZ of tt given by Z(t)=∫ℝne−Vt(x)dx;Z(t)=\int_{\mathbb{R}^n}e^{-V_t(x)}\,dx;(3)✦ the Gibbs law of the family at tt is the probability measure μt=Z(t)−1e−Vtdx\mu_t=Z(t)^{-1}e^{-V_t}\,dx on ℝn\mathbb{R}^n. The Gibbs mean and the Gibbs variance of a real function FF on ℝn\mathbb{R}^n are EμtF=1Z(t)∫ℝnFe−Vtdx,Var⁡μtF=Eμt(F−EμtF)2.E_{\mu_t}F=\frac{1}{Z(t)}\int_{\mathbb{R}^n}F\,e^{-V_t}\,dx, \qquad \operatorname{Var}_{\mu_t}F=E_{\mu_t}\bigl(F-E_{\mu_t}F\bigr)^2.(4) The first and second partition derivatives are the integrals of the first and second tt-derivatives of the Gibbs density: Z′(t)=−∫ℝnV̇te−Vtdx,Z″(t)=∫ℝn(V̇t2−V̈t)e−Vtdx.Z'(t)=-\int_{\mathbb{R}^n}\dot V_t\,e^{-V_t}\,dx, \qquad Z''(t)=\int_{\mathbb{R}^n}\bigl(\dot V_t^{\,2}-\ddot V_t\bigr)\,e^{-V_t}\,dx.(5)

Slab Features

For a vector p∈ℝnp\in\mathbb{R}^n, a width h>0h>0 and an integer q≥0q\ge0, the feature polynomial of pp and hh is the qqth power of the normalized linear form x↦⟨p,x⟩/hx\mapsto\langle p,x\rangle/h.✦ The feature family of vectors pi∈ℝnp_i\in\mathbb{R}^n and widths hi>0h_i>0, i∈Ii\in I, is the family of their feature polynomials, ψi(x)=(⟨pi,x⟩hi)q.\psi_i(x)=\Bigl(\frac{\langle p_i,x\rangle}{h_i}\Bigr)^q.(6)✦ For two widths hi0,hi1>0h_i^0,h_i^1>0, the decay rate bib_i of exponent qq is a real number and the interpolated width hi(t)h_i(t) at tt a positive one, given by bi=−qlog⁡hi1hi0,hi(t)=(hi0)1−t(hi1)t.b_i=-q\log\frac{h_i^1}{h_i^0}, \qquad h_i(t)=(h_i^0)^{1-t}(h_i^1)^t.(7)

The soft slab law μ\mu of nonzero vectors pip_i and widths hi>0h_i>0, with an even exponent q≥2q\ge2, is the law μ0\mu_0 of the Gibbs family of their feature family: μ=Z−1e−Vdx\mu=Z^{-1}e^{-V}dx with V(x)=ε2|x|2+∑iψi(x)V(x)=\frac{\varepsilon}{2}|x|^2+\sum_i\psi_i(x).

The soft variance inequality is the following claim. Let n≥2n\ge2, let II be a finite nonempty set, let q≥2q\ge2 be an even integer and ε>0\varepsilon>0, let pi∈ℝnp_i\in\mathbb{R}^n, i∈Ii\in I, be nonzero vectors, hi>0h_i>0 widths and bib_i real numbers, and let μ\mu be the soft slab law of the pip_i and hih_i. Then Var⁡μ(∑i∈Ibiψi)≤Eμ(∑i∈Ibi2ψi).\operatorname{Var}_\mu\Bigl(\sum_{i\in I}b_i\psi_i\Bigr) \le E_\mu\Bigl(\sum_{i\in I}b_i^2\psi_i\Bigr).(8)✦

The Claim

Theorem 1. Let ε>0\varepsilon>0 and let II be a finite set.

  1. Let ψi\psi_i, i∈Ii\in I, be polynomials on ℝn\mathbb{R}^n that are nonnegative everywhere, and let bib_i be real numbers. For every real tt, Z″(t)Z(t)−Z′(t)2Z(t)2=Var⁡μt(V̇t)−Eμt(V̈t).\frac{Z''(t)\,Z(t)-Z'(t)^2}{Z(t)^2} =\operatorname{Var}_{\mu_t}(\dot V_t)-E_{\mu_t}(\ddot V_t).(9)✦ If moreover Var⁡μt(V̇t)≤Eμt(V̈t)\operatorname{Var}_{\mu_t}(\dot V_t)\le E_{\mu_t}(\ddot V_t) for every real tt, then for all s,u∈ℝs,u\in\mathbb{R} and θ∈[0,1]\theta\in[0,1] log⁡Z((1−θ)s+θu)≥(1−θ)log⁡Z(s)+θlog⁡Z(u),\log Z\bigl((1-\theta)s+\theta u\bigr)\ge(1-\theta)\log Z(s)+\theta\log Z(u),(10)✦ and for every t∈[0,1]t\in[0,1] Z(0)1−tZ(1)t≤Z(t).Z(0)^{1-t}\,Z(1)^t\le Z(t).(11)✦

  2. Let pi∈ℝnp_i\in\mathbb{R}^n be vectors, hi0,hi1>0h_i^0,h_i^1>0 widths and q≥0q\ge0 an integer; let ψi\psi_i be the feature family of the pip_i and the hi0h_i^0, and bib_i the decay rates of exponent qq. For every integer k≥0k\ge0, every real tt and every x∈ℝnx\in\mathbb{R}^n, Vt(k)(x)=∑i∈Ibik(⟨pi,x⟩hi(t))q.V_t^{(k)}(x)=\sum_{i\in I}b_i^k\Bigl(\frac{\langle p_i,x\rangle}{h_i(t)}\Bigr)^q.(12)✦

  3. Assume the soft variance inequality. Let n≥2n\ge2 and II be nonempty, and let ψi\psi_i and bib_i be as in (b), with nonzero vectors pip_i and an even exponent q≥2q\ge2. For every real tt, Var⁡μt(V̇t)≤Eμt(V̈t).\operatorname{Var}_{\mu_t}(\dot V_t)\le E_{\mu_t}(\ddot V_t).(13)✦

The left side of (9) is the expression of (log⁡Z)″(\log Z)'' through Z′Z' and Z″Z'', and (13) is the hypothesis of (10) and (11) for the feature family with the decay rates.

What Is Checked

The claim is stated over the Gibbs family of the regularized potential (2), with its variations, partition function, partition derivatives, Gibbs mean and Gibbs variance, over the feature family, the decay rates and the interpolated widths, and, in (13), under the soft variance inequality as a hypothesis.