OctoInformBeta

OpenAI preprint, September 2026 · §7. The variance inequality (proof of Proposition 2.1)

Homogeneity and the trace moment (7.1), (7.2): Even powers (⟨p,x⟩/h)q(\langle p,x\rangle/h)^q are convex with positive mean, and ∫tr(τD2ψ)dμ=q∫ψdμ\int\mathrm{tr}(\tau D^2\psi)\,d\mu=q\int\psi\,d\mu for ψ\psi of degree qq

Euler relations for slab powers and the trace moment identity.

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

The soft slab law μ\mu on ℝn\mathbb{R}^n of a regularization ε>0\varepsilon > 0, vectors p1,…,pkp_1,\dots,p_k and widths h1,…,hkh_1,\dots,h_k is the Gaussian measure with covariance ε−1I\varepsilon^{-1}I reweighted by the soft indicators e−ψie^{-\psi_i} of the slabs |⟨pi,x⟩|≤hi|\langle p_i,x\rangle| \le h_i, where ψi(x)=(⟨pi,x⟩/hi)q\psi_i(x) = (\langle p_i,x\rangle/h_i)^q and q≥2q \ge 2 is even. The functions ψi\psi_i are the features of μ\mu, and each is homogeneous of degree qq. The item asks what homogeneity gives for a feature and its derivatives, and how the Hessians of the features and of the potential VV of μ\mu, and the means Eψi\mathrm{E}\,\psi_i of the features under μ\mu, are bounded below.

Behind μ\mu stands a moment measure: a convex function φ\varphi of source coordinates zz solves the Monge-Ampère equation for VV, and its gradient sends ν=e−φdz\nu = e^{-\varphi}dz to target coordinates x=∇φ(z)x = \nabla\varphi(z), where μ\mu lives; τ\tau is the Hessian of φ\varphi read at xx. The trace moment of a function ψ\psi is Etr(τD2ψ)\mathrm{E}\,\mathrm{tr}(\tau D^2\psi), and the item asks for it when ψ\psi is homogeneous of degree qq, in particular for the features.

Slab Powers and Moment Coordinates

Points of the coordinate space ℝn\mathbb{R}^n are x=(x1,…,xn)x = (x_1,\dots,x_n), and e1,…,ene_1,\dots,e_n is its standard basis. For a function ff on ℝn\mathbb{R}^n, Dvf(x)=Df(x)vD_vf(x) = Df(x)\,v is the directional derivative along a vector vv✦, ∂if=Deif\partial_if = D_{e_i}f the partial derivative✦, ∇f=(∂1f,…,∂nf)\nabla f = (\partial_1f,\dots,\partial_nf) the gradient✦ and D2fD^2f the Hessian, the matrix with entries (D2f)ij=∂j∂if(D^2f)_{ij} = \partial_j\partial_if✦.

The slab power with direction p∈ℝnp \in \mathbb{R}^n, real width hh and exponent q∈ℕq \in \mathbb{N} is the function on ℝn\mathbb{R}^n given by ψ(x)=(⟨p,x⟩h)q,\psi(x) = \Bigl(\frac{\langle p,x\rangle}{h}\Bigr)^{q},(1) where ⟨p,x⟩=∑ipixi\langle p,x\rangle = \sum_i p_ix_i is the linear functional of pp.

The soft slab potential with regularization ε∈ℝ\varepsilon \in \mathbb{R}, vectors p1,…,pk∈ℝnp_1,\dots,p_k \in \mathbb{R}^n, real widths h1,…,hkh_1,\dots,h_k and exponent q∈ℕq \in \mathbb{N} is the function on ℝn\mathbb{R}^n given by V(x)=ε2|x|2+∑i=1k(⟨pi,x⟩hi)q.V(x) = \frac{\varepsilon}{2}\,|x|^2 + \sum_{i=1}^{k}\Bigl(\frac{\langle p_i,x\rangle}{h_i}\Bigr)^{q}.(2)✦

For a function VV on ℝn\mathbb{R}^n, a constant ZZ and a reference measure dxdx on ℝn\mathbb{R}^n, the target measure Z−1e−VdxZ^{-1}e^{-V}dx is the measure whose density with respect to dxdx is the weight e−Ve^{-V} divided by ZZ, taken in [0,∞][0,\infty].✦ The soft slab law μ\mu of the data in (2) is the target measure of the soft slab potential VV with Z=∫e−V(y)dyZ = \int e^{-V(y)}\,dy.

The matrix field τ\tau of a function φ\varphi on ℝn\mathbb{R}^n and a homeomorphism Φ\Phi of ℝn\mathbb{R}^n onto itself is given at a point xx by τ(x)=D2φ(Φ−1(x)).\tau(x) = D^2\varphi\bigl(\Phi^{-1}(x)\bigr).(3)✦

The Claim

Theorem 1. Let ψ\psi be a slab power (1) of exponent qq. If q≥1q \ge 1, then at every point xx of ℝn\mathbb{R}^n x⋅∇ψ(x)=qψ(x).x \cdot \nabla\psi(x) = q\,\psi(x).(4)✦ If q≥2q \ge 2, then at every point D2ψ(x)x=(q−1)∇ψ(x),D^2\psi(x)\,x = (q-1)\,\nabla\psi(x),(5)✦ and if moreover qq is even, the Hessian of ψ\psi is positive semidefinite at every point: D2ψ(x)≽0.D^2\psi(x) \succeq 0.(6)✦

Let μ\mu be the soft slab law of a regularization ε>0\varepsilon > 0, any vectors p1,…,pkp_1,\dots,p_k and widths h1,…,hkh_1,\dots,h_k, and an even exponent qq. Then every slab power with nonzero direction vv, nonzero width rr and even exponent mm has positive mean under μ\mu: ∫ℝn(⟨v,x⟩r)mdμ(x)>0.\int_{\mathbb{R}^n}\Bigl(\frac{\langle v,x\rangle}{r}\Bigr)^{m}\,d\mu(x) > 0.(7)✦ For every real ε\varepsilon, all vectors p1,…,pkp_1,\dots,p_k and widths h1,…,hkh_1,\dots,h_k, and every even exponent q≥2q \ge 2, the soft slab potential VV satisfies at every point xx D2V(x)≽εI.D^2V(x) \succeq \varepsilon I.(8)✦

Let φ\varphi, VV and ψ\psi be smooth functions on ℝn\mathbb{R}^n, and let cc and qq be real numbers. Suppose that at every point zz the Hessian D2φ(z)D^2\varphi(z) is positive definite and D2φ(z)≼ΛID^2\varphi(z) \preceq \Lambda I for a constant Λ≥0\Lambda \ge 0; that φ\varphi solves the Monge-Ampère equation log⁡det⁡D2φ(z)=V(∇φ(z))+c−φ(z)\log\det D^2\varphi(z) = V(\nabla\varphi(z)) + c - \varphi(z); and that ∇φ\nabla\varphi is a homeomorphism of ℝn\mathbb{R}^n onto itself with smooth inverse. Let τ\tau be the field (3) with Φ=∇φ\Phi = \nabla\varphi, and μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}dx the target measure of a constant Z>0Z > 0. Suppose further that ψ\psi satisfies Euler’s relation x⋅∇ψ(x)=qψ(x)x\cdot\nabla\psi(x) = q\,\psi(x) at every point, and that ∂iψ\partial_i\psi, xi∂iψx_i\,\partial_i\psi and ∂i∂jψ\partial_i\partial_j\psi are μ\mu-integrable for all ii and jj. Then tr(τD2ψ)\mathrm{tr}(\tau D^2\psi) is μ\mu-integrable and ∫ℝntr(τ(x)D2ψ(x))dμ(x)=q∫ℝnψdμ.\int_{\mathbb{R}^n}\mathrm{tr}\bigl(\tau(x)\,D^2\psi(x)\bigr)\,d\mu(x) = q\int_{\mathbb{R}^n}\psi\,d\mu.(9)✦

For a feature ψ=ψi\psi = \psi_i of the soft slab law, with Hi=D2ψiH_i = D^2\psi_i, the trace moment mi=Etr(τHi)m_i = \mathrm{E}\,\mathrm{tr}(\tau H_i) of ψi\psi_i equals qEψiq\,\mathrm{E}\,\psi_i by (9), the identity the item asks for.

What Is Checked

The claims are stated over the slab powers (1), the soft slab potential (2) with its law, the target measure Z−1e−VdxZ^{-1}e^{-V}dx of the weight e−Ve^{-V} and the field τ\tau of (3), with derivatives, gradients and Hessians taken in the coordinates of ℝn\mathbb{R}^n.