OctoInformBeta

OpenAI preprint, September 2026 · §3. From low energy to shifted signs · §3

Lemma 3.2 [Convexity under Minkowski averages]: For bounded open convex domains AA, BB and C⊇(1−t)A+tBC \supseteq (1-t)A+tB, with Prékopa’s inequality for Gaussian images and section rates convex in the height

The compressed heat norm is log-concave under Minkowski averages.

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

Let ι\iota be a finite set of coordinates and v=(vi)i∈ιv = (v_i)_{i \in \iota} a weight, a vector of nonnegative reals. The heat flow on ℝι\mathbb{R}^\iota run for unit time, with diffusion viv_i along the coordinate ii, is a Gaussian Fourier multiplier on L2(ℝι)L^2(\mathbb{R}^\iota), and its compression to the functions supported in a bounded open convex domain DD has a norm that measures how much of the flow DD retains.

Given two such domains AA and BB and 0≤t≤10 \le t \le 1, the lemma compares this norm for a convex domain containing the Minkowski average (1−t)A+tB(1-t)A + tB with the norms for AA and for BB, as Prékopa’s log-concavity does for the Gaussian measure of convex sets.

A second use is inside one convex body KK. For a direction ee and a height s≥0s \ge 0, the lemma describes the heat norm of the shifted section of KK at height ss, defined below, as a function of ss.

Heat Operators Compressed to a Domain

Let XX be a measure space with measure μ\mu. For a measurable complex function mm on XX and a constant C≥0C \ge 0 with |m|≤C|m| \le C almost everywhere, the multiplication operator MmM_m is the bounded linear operator on L2(μ)L^2(\mu) that sends ff to the product mfmf.✦ The mask PDP_D of a measurable set D⊆XD \subseteq X is the multiplication operator M𝟏DM_{\mathbf{1}_D} by its indicator, which restricts a function to DD.✦

Let EE be a Euclidean space, a finite-dimensional real inner product space, with Lebesgue measure, and let ℱ\mathcal{F} be the unitary Fourier transform on the complex space L2(E)L^2(E). For a measurable real function mm on EE with |m|≤1|m| \le 1, the Fourier multiplier operator with multiplier mm is the operator ℱ−1Mmℱ\mathcal{F}^{-1} M_m \mathcal{F} on L2(E)L^2(E).✦ For a measurable function w≥0w \ge 0 on EE, the heat operator HwH_w is the Fourier multiplier operator with multiplier e−we^{-w}✦, and the compressed heat operator TDwT^w_D of ww on a measurable set D⊆ED \subseteq E is TDw=PDHwPD=PDℱ−1Me−wℱPD.T^w_D = P_D\, H_w\, P_D = P_D\, \mathcal{F}^{-1} M_{e^{-w}} \mathcal{F}\, P_D .(1)✦

The Gaussian kernel symbol σv\sigma_v of a weight vv on ι\iota is the function on ℝι\mathbb{R}^\iota given by σv(ξ)=12∑i∈ιvi(2πξi)2,\sigma_v(\xi) = \frac{1}{2} \sum_{i \in \iota} v_i \, (2\pi \xi_i)^2 ,(2) and TDvT^v_D stands for TDσvT^{\sigma_v}_D, the heat flow of the weight vv compressed to DD.

Shifted Sections of a Convex Body

Let KK be a subset of ℝι\mathbb{R}^\iota, e∈ℝιe \in \mathbb{R}^\iota a direction and ℓ\ell a real number. The shifted section Ke,ℓ(s)K_{e,\ell}(s) of KK at height s∈ℝs \in \mathbb{R} is the subset of ℝι\mathbb{R}^\iota Ke,ℓ(s)={y:y+(c−s)e∈K and y+(c+s)e∈K for some c∈ℝ with |c|+s<ℓ}.K_{e,\ell}(s) = \bigl\{\, y : y + (c-s)e \in K \text{ and } y + (c+s)e \in K \text{ for some } c \in \mathbb{R} \text{ with } |c| + s < \ell \,\bigr\}.(3)✦ The set Se,ℓ(K)S_{e,\ell}(K) of active heights of KK is the set of real numbers s≥0s \ge 0 for which Ke,ℓ(s)K_{e,\ell}(s) is nonempty.✦

For KK open and a weight vv, the section norm of KK at height ss is the operator norm of TKe,ℓ(s)vT^v_{K_{e,\ell}(s)}✦, and the section rate ρ\rho of KK is its negative logarithm, the function of the height ρ(s)=−log⁡∥TKe,ℓ(s)v∥.\rho(s) = -\log \bigl\| T^v_{K_{e,\ell}(s)} \bigr\| .(4)✦

The Claim

Theorem 1. Let vv be a weight on ι\iota with vi>0v_i > 0 for every ii.

  1. Let AA, BB and CC be bounded open convex subsets of ℝι\mathbb{R}^\iota with AA and BB nonempty, and let 0≤t≤10 \le t \le 1 be such that (1−t)a+tb∈C(1-t)a + tb \in C for all a∈Aa \in A and b∈Bb \in B. Then ∥TAv∥1−t∥TBv∥t≤∥TCv∥.\|T^v_A\|^{1-t} \, \|T^v_B\|^{t} \le \|T^v_C\| .(5)✦

  2. Let EE be a Euclidean space with its standard Gaussian measure γE\gamma_E, let LL be a continuous linear map from EE to a real normed space FF, and let μ=γE∘L−1\mu = \gamma_E \circ L^{-1} be the image of γE\gamma_E under LL. Let AA, BB and CC be convex Borel subsets of FF with 0∈A0 \in A and 0∈B0 \in B, and let 0≤α≤10 \le \alpha \le 1 be such that (1−α)a+αb∈C(1-\alpha)a + \alpha b \in C for all a∈Aa \in A and b∈Bb \in B. Then μ(A)1−αμ(B)α≤μ(C).\mu(A)^{1-\alpha} \, \mu(B)^{\alpha} \le \mu(C) .(6)✦

  3. Let KK be a bounded open convex subset of ℝι\mathbb{R}^\iota, e∈ℝιe \in \mathbb{R}^\iota and ℓ∈ℝ\ell \in \mathbb{R}. Then the set Se,ℓ(K)S_{e,\ell}(K) of active heights is convex, and for all s0,s1∈Se,ℓ(K)s_0, s_1 \in S_{e,\ell}(K) and 0≤τ≤10 \le \tau \le 1 the section rate satisfies ρ((1−τ)s0+τs1)≤(1−τ)ρ(s0)+τρ(s1).\rho\bigl((1-\tau)s_0 + \tau s_1\bigr) \le (1-\tau)\,\rho(s_0) + \tau\,\rho(s_1) .(7)✦

In the paper, the result reads: for nonempty bounded open convex sets A,B⊂ℝιA, B \subset \mathbb{R}^\iota symmetric about zero, a positive definite diagonal matrix QQ and 0≤t≤10 \le t \le 1, the diagonal Dirichlet energy λQ\lambda_Q satisfies λQ((1−t)A+tB)≤(1−t)λQ(A)+tλQ(B)\lambda_Q\bigl((1-t)A + tB\bigr) \le (1-t)\lambda_Q(A) + t\lambda_Q(B).

With the weight vv in the place of QQ and −log⁡∥TDv∥-\log\|T^v_D\| in the place of λQ(D)\lambda_Q(D), the case C=(1−t)A+tBC = (1-t)A + tB of (5) is this inequality, for domains not required to be symmetric about zero.

What Is Checked

The inequalities (5) and (7) are stated over the compressed heat operator (1) of the Gaussian kernel symbol (2), and (7) also over the shifted sections (3) and the section rate (4) on the active heights. The inequality (6) is stated over the image of the standard Gaussian measure of a Euclidean space under a continuous linear map.

About these constructions the library proves that the symbol σv\sigma_v is continuous✦ and nonnegative✦, and that for a measurable complex function mm with |m|≤C|m| \le C almost everywhere and every f∈L2(μ)f \in L^2(\mu), the product mfmf lies in L2(μ)L^2(\mu)✦ and has norm at most C∥f∥C\|f\|✦.