OctoInformBeta

OpenAI preprint, September 2026 · §5. Density of the even tests

Lemma 5.1: For τ\tau read from moment coordinates with VV even; there, every g∈L2(μ)g\in L^2(\mu) orthogonal to all A(A−1)fA(A-1)f is affine

Centered even functions are limits of A(A−1)fA(A-1)f in L2(μ)L^2(\mu)..

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

In the moment-measure picture, the gradient of a smooth convex potential φ\varphi on source coordinates zz carries e−φdze^{-\varphi}\,dz to a law μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx on target coordinates x=∇φ(z)x = \nabla\varphi(z), where VV is a smooth function and Z>0Z > 0 a constant. The Hessian of φ\varphi, read at xx, is a field τ\tau of positive definite matrices, and τ\tau defines a second-order operator AA on functions of xx. The question is whether, when VV is even, every even function FF in L2(μ)L^2(\mu) with ∫Fdμ=0\int F\,d\mu = 0, a centered function, is a limit in L2(μ)L^2(\mu) of functions A(A−1)fA(A-1)f with ff smooth and compactly supported.

The Operator and the Moment Coordinates

Points of the coordinate space ℝι\mathbb{R}^\iota have one real coordinate xix_i for each index ii of a finite set ι\iota, and ⟨b,x⟩=∑ibixi\langle b, x\rangle = \sum_i b_i x_i. For a function ff on ℝι\mathbb{R}^\iota, ∂if\partial_i f is its directional derivative along the standard basis vector eie_i✦, ∇f=(∂if)i\nabla f = (\partial_i f)_i is its gradient✦ and D2f=(∂j∂if)ijD^2 f = (\partial_j\partial_i f)_{ij} its Hessian.✦ A function FF on ℝι\mathbb{R}^\iota is even when F(−x)=F(x)F(-x) = F(x) for every xx, and affine when F(x)=a+⟨b,x⟩F(x) = a + \langle b, x\rangle for some a∈ℝa \in \mathbb{R} and b∈ℝιb \in \mathbb{R}^\iota. A measure μ\mu on ℝι\mathbb{R}^\iota is symmetric when ∫q(−x)dμ(x)=∫q(x)dμ(x)\int q(-x)\,d\mu(x) = \int q(x)\,d\mu(x) for every real function qq on ℝι\mathbb{R}^\iota. A test function is a smooth function on ℝι\mathbb{R}^\iota with compact support; the test functions form a linear space Cc∞(ℝι)C_c^\infty(\mathbb{R}^\iota).✦

Let τ\tau be a field of real ι×ι\iota\times\iota matrices on ℝι\mathbb{R}^\iota. The operator AA of τ\tau is the second-order differential operator, with smooth coefficients when τ\tau is smooth, that is the sum of the drift part x⋅∇x\cdot\nabla and the principal part −τ:D2-\tau : D^2, where M:N=∑i,jMijNijM : N = \sum_{i,j} M_{ij}N_{ij} for matrices MM and NN: Af(x)=∑ixi∂if(x)−∑i,jτij(x)∂i∂jf(x)=x⋅∇f(x)−τ(x):D2f(x).Af(x) = \sum_i x_i\,\partial_i f(x) - \sum_{i,j}\tau_{ij}(x)\,\partial_i\partial_j f(x) = x\cdot\nabla f(x) - \tau(x) : D^2 f(x).(1) The gradient form of τ\tau sends two functions ff and gg to the function Γτ(f,g)(x)=∇f(x)⋅τ(x)∇g(x).\Gamma_\tau(f,g)(x) = \nabla f(x)\cdot\tau(x)\,\nabla g(x).(2)✦

Let μ\mu be a measure on ℝι\mathbb{R}^\iota. The field τ\tau satisfies the integration formula for μ\mu when all test functions ff and hh satisfy ∫hAfdμ=∫Γτ(h,f)dμ,\int h\,Af\,d\mu = \int \Gamma_\tau(h,f)\,d\mu,(3) and the Hessian identity for μ\mu when every test function ff satisfies ∫tr⁡(τD2fτD2f)dμ=∫(Af)2dμ−∫fAfdμ,\int \operatorname{tr}\bigl(\tau\,D^2 f\,\tau\,D^2 f\bigr)\,d\mu = \int (Af)^2\,d\mu - \int f\,Af\,d\mu,(4) where the integrand on the left is the Hessian energy density of ff.

For a smooth function VV on ℝι\mathbb{R}^\iota and a constant Z>0Z > 0, the target law μ\mu is the measure on target coordinates xx with density e−V/Ze^{-V}/Z with respect to Lebesgue measure✦; for a smooth function φ\varphi, the source measure ν\nu on source coordinates zz has density e−φe^{-\varphi}: dμ(x)=Z−1e−V(x)dx,dν(z)=e−φ(z)dz.d\mu(x) = Z^{-1}e^{-V(x)}\,dx, \qquad d\nu(z) = e^{-\varphi(z)}\,dz.(5)✦ Moment coordinates for μ\mu are a smooth function φ\varphi on ℝι\mathbb{R}^\iota and a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota such that VV is bounded below; μ\mu is finite and gives positive mass to every nonempty open set; for some Λ≥0\Lambda \ge 0, every Hessian D2φ(z)D^2\varphi(z) is positive definite with D2φ(z)≤ΛID^2\varphi(z) \le \Lambda I; for some constant cc, φ\varphi solves the Monge-Ampère equation log⁡det⁡D2φ(z)=V(∇φ(z))+c−φ(z)for every z;\log\det D^2\varphi(z) = V(\nabla\varphi(z)) + c - \varphi(z) \quad \text{for every } z;(6) Φ=∇φ\Phi = \nabla\varphi and Φ−1\Phi^{-1} is smooth; and ∇φ\nabla\varphi pushes ν\nu forward to μ\mu: (∇φ)#ν=μ.(\nabla\varphi)_\#\,\nu = \mu.(7) The moment matrix of (φ,Φ)(\varphi,\Phi) is the field τ\tau on target coordinates whose value at xx is the Hessian of φ\varphi at the source point of xx: τ(x)=D2φ(Φ−1(x)).\tau(x) = D^2\varphi\bigl(\Phi^{-1}(x)\bigr).(8)✦

The Claim

Theorem 1. Let (φ,Φ)(\varphi,\Phi) be moment coordinates for μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx, and let AA be the operator of their moment matrix τ\tau.

  1. If VV is even, every even function F∈L2(μ)F \in L^2(\mu) with ∫Fdμ=0\int F\,d\mu = 0 lies in the closure in L2(μ)L^2(\mu) of the range of A(A−1)A(A-1) on test functions: F∈{A(A−1)f:f∈Cc∞(ℝι)}¯.F \in \overline{\{A(A-1)f : f \in C_c^\infty(\mathbb{R}^\iota)\}}.(9)✦

  2. If g∈L2(μ)g \in L^2(\mu) satisfies ∫gA(A−1)fdμ=0\int g\,A(A-1)f\,d\mu = 0 for every test function ff, then for some a∈ℝa \in \mathbb{R} and b∈ℝιb \in \mathbb{R}^\iota, g(x)=a+⟨b,x⟩for μ-almost every x.g(x) = a + \langle b, x\rangle \quad \text{for $\mu$-almost every } x.(10)✦

Let μ\mu be a finite symmetric measure on ℝι\mathbb{R}^\iota and SS a linear subspace of L2(μ)L^2(\mu) whose orthogonal complement S⟂S^\perp consists of functions that are affine μ\mu-almost everywhere.

  1. Every even function F∈L2(μ)F \in L^2(\mu) with ∫Fdμ=0\int F\,d\mu = 0 satisfies F∈S¯.F \in \overline{S}.(11)✦

Let μ\mu be a finite measure on ℝι\mathbb{R}^\iota and τ\tau a smooth field of positive semidefinite matrices that satisfies the integration formula (3) for μ\mu.

  1. If η\eta is a test function with Γτ(η,η)≤C\Gamma_\tau(\eta,\eta) \le C everywhere for some constant CC and ww is a smooth function in L2(μ)L^2(\mu), then ∫η2Γτ(w,w)dμ≤2∫η2wAwdμ+4C∫w2dμ.\int \eta^2\,\Gamma_\tau(w,w)\,d\mu \le 2\int \eta^2\,w\,Aw\,d\mu + 4C\int w^2\,d\mu.(12)✦

Suppose further that μ\mu gives positive mass to every nonempty open set and that, for some Λ≥0\Lambda \ge 0, every τ(x)\tau(x) is positive definite with τ(x)≤ΛI\tau(x) \le \Lambda I; let ww, vv and gg be smooth functions in L2(μ)L^2(\mu).

  1. If Aw=0Aw = 0, then ww is constant: w(x)=w(y)for all x,y∈ℝι.w(x) = w(y) \quad \text{for all } x, y \in \mathbb{R}^\iota.(13)✦

When τ\tau also satisfies the Hessian identity (4) for μ\mu, the following hold.

  1. If Av=vAv = v, then vv is linear: for some b∈ℝιb \in \mathbb{R}^\iota, v(x)=⟨b,x⟩for every x.v(x) = \langle b, x\rangle \quad \text{for every } x.(14)✦

  2. If A(A−1)g=0A(A-1)g = 0, then gg is affine: for some a∈ℝa \in \mathbb{R} and b∈ℝιb \in \mathbb{R}^\iota, g(x)=a+⟨b,x⟩for every x.g(x) = a + \langle b, x\rangle \quad \text{for every } x.(15)✦

  3. If A(A−1)g=0A(A-1)g = 0 and gg is even, then gg is constant: g(x)=g(0)for every x.g(x) = g(0) \quad \text{for every } x.(16)✦

In the paper, the result reads: if dμ=ρdxd\mu = \rho\,dx is a probability measure with a smooth, strictly positive, even density ρ\rho and τ\tau is a smooth, even, symmetric matrix field with 0<τ≤ΛI0 < \tau \le \Lambda I that satisfies the integration formula and the Hessian identity for μ\mu, then the closure in L2(μ)L^2(\mu) of {A(A−1)f:f∈Cc∞(ℝι) even}\{A(A-1)f : f \in C_c^\infty(\mathbb{R}^\iota) \text{ even}\}, with AA the operator of τ\tau, is the space of even g∈L2(μ)g \in L^2(\mu) with 𝔼g=0\mathbb{E}\,g = 0, where 𝔼\mathbb{E} is the expectation under μ\mu.

Part (a) is the inclusion for the moment matrix of moment coordinates, with all test functions in place of the even ones.

What Is Checked

The theorem is stated over the coordinate space ℝι\mathbb{R}^\iota with its partial derivatives, gradients and Hessians, test functions, even and affine functions and symmetric measures; the operator AA of a matrix field, its gradient form Γτ\Gamma_\tau, the Hessian energy density, the integration formula and the Hessian identity; and, in (a) and (b), the laws μ\mu and ν\nu, moment coordinates, their moment matrix τ\tau and the range of A(A−1)A(A-1) on test functions in L2(μ)L^2(\mu).

The library proves of these constructions that the moment matrix is smooth whenever φ\varphi and Φ−1\Phi^{-1} are smooth✦, that for a smooth field τ\tau the operator AA maps test functions linearly to test functions, and that A(A−1)A(A-1) maps test functions linearly into L2(μ)L^2(\mu).✦