OctoInformBeta

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

The pairings and the two Cauchy-Schwarz bounds (7.3) to (7.5): The pairing d=𝔼ψδμWd = \mathbb{E}\,\psi\,\delta_\mu W for a convex ψ\psi and the field WW of a test function, ψ\psi homogeneous in the second bound

Three forms of one pairing and two Cauchy-Schwarz bounds.

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 setting is that of §7, the proof of the variance inequality of Proposition 2.1. The gradient map x=∇φ(z)x = \nabla\varphi(z) of a potential φ\varphi takes source coordinates, where ν=e−φdz\nu = e^{-\varphi}\,dz lives, to target coordinates, where the target measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx with potential VV and a constant Z>0Z > 0 lives, and φ\varphi is tied to VV by the Monge-Ampère equation, for a constant cc, log⁡det⁡D2φ(z)=V(∇φ(z))+c−φ(z).\log\det D^2\varphi(z) = V(\nabla\varphi(z)) + c - \varphi(z).(1) A test function ff in target coordinates gives a vector field WW, built from the Hessians of φ\varphi and ff, and its weighted divergence u=δμWu = \delta_\mu W. The item pairs uu with a second smooth function ψ\psi: the integral of ψu\psi u against μ\mu has two other expressions, and its square obeys two Cauchy-Schwarz bounds, the second for ψ\psi homogeneous of degree qq.

Weighted Calculus

The coordinate space ℝι\mathbb{R}^\iota of a finite index set ι\iota is the real vector space of functions from ι\iota to ℝ\mathbb{R}, added and scaled coordinatewise✦; its standard basis vector eie_i has coordinate 11 at ii and 00 elsewhere.✦ For a function gg on ℝι\mathbb{R}^\iota and a vector vv, the directional derivative Dvg(x)=Dg(x)vD_v g(x) = Dg(x)\,v is the Fréchet derivative of gg at xx applied to vv✦, and the partial derivative ∂ig\partial_i g is DeigD_{e_i} g.✦ The gradient ∇g(x)\nabla g(x) is the vector with coordinates ∂ig(x)\partial_i g(x).✦ The Hessian D2g(x)D^2 g(x) is the matrix with ∂j∂ig(x)\partial_j\partial_i g(x) in row ii and column jj.✦

A potential on ℝι\mathbb{R}^\iota is a function VV on ℝι\mathbb{R}^\iota, and its weight is e−Ve^{-V}.✦ The weighted adjoint of the directional derivative DvD_v for VV is the operator Dv*g=DvVg−DvgD_v^* g = D_vV\,g - D_v g on functions gg on ℝι\mathbb{R}^\iota.✦ The weighted divergence δμY\delta_\mu Y of a vector field Y=(Yi)Y = (Y_i) on ℝι\mathbb{R}^\iota, for the target measure μ\mu with density e−V/Ze^{-V}/Z, is the function δμY=∑iDei*Yi=∑i(∂iVYi−∂iYi),\delta_\mu Y = \sum_i D_{e_i}^* Y_i = \sum_i \bigl(\partial_i V\,Y_i - \partial_i Y_i\bigr),(2)✦ and δν\delta_\nu denotes the same operator with φ\varphi in place of VV. For a function gg, the expectation 𝔼g\mathbb{E}\,g is ∫gdμ\int g\,d\mu.

The Transported Fields

Let Φ\Phi be a homeomorphism of ℝι\mathbb{R}^\iota, sending a source point zz to the target point x=Φ(z)x = \Phi(z); in the claim, Φ=∇φ\Phi = \nabla\varphi. A function gg of zz is read at xx through its target function g∘Φ−1g\circ\Phi^{-1}.✦ The Hessian of φ\varphi read at xx is τ(x)=D2φ(Φ−1x)\tau(x) = D^2\varphi(\Phi^{-1}x)✦, and the inverse Hessian field is M(x)=τ(x)−1M(x) = \tau(x)^{-1}.✦

The matrix field BB is the target function of a matrix built at the source point from φ\varphi and the Hessian of ff pulled by the gradient, z↦D2f(∇φ(z))z \mapsto D^2 f(\nabla\varphi(z)): B(x)=D2φ(z)D2f(∇φ(z))D2φ(z),z=Φ−1x,B(x) = D^2\varphi(z)\,D^2 f(\nabla\varphi(z))\,D^2\varphi(z), \qquad z = \Phi^{-1}x,(3)✦ so that B=τD2fτB = \tau\,D^2 f\,\tau when Φ=∇φ\Phi = \nabla\varphi.

The moment coefficients of φ\varphi and ff are the entries of D2φ(z)D2f(∇φ(z))D^2\varphi(z)\,D^2 f(\nabla\varphi(z)), Pak(z)=∑i∂a∂iφ(z)∂i∂kf(∇φ(z)),P_{ak}(z) = \sum_i \partial_a\partial_i\varphi(z)\,\partial_i\partial_k f(\nabla\varphi(z)),(4)✦ and the moment divergence Wa=δνPa=∑k(∂kφPak−∂kPak)W_a = \delta_\nu P_a = \sum_k (\partial_k\varphi\,P_{ak} - \partial_k P_{ak}) is the weighted divergence of the row PaP_a for the source law ν\nu. The field WW is the vector field on the target side with components Wa∘Φ−1W_a\circ\Phi^{-1}, and u=δμWu = \delta_\mu W is its weighted divergence for VV.

Two energies enter the bounds. For a function ψ\psi, the Hessian trace energy is the function tr⁡(MBD2ψB)\operatorname{tr}(M\,B\,D^2\psi\,B)✦, and the Hessian quadratic form energy of a vector field YY is Y⊤D2ψY=∑i,j∂j∂iψYiYjY^{\top} D^2\psi\,Y = \sum_{i,j} \partial_j\partial_i\psi\,Y_i Y_j.✦

The Claim

Theorem 1. Let φ\varphi, VV, ff and ψ\psi be smooth functions on ℝι\mathbb{R}^\iota, with ff compactly supported, and let φ\varphi satisfy (1) at every point for a constant cc. Let the gradient map ∇φ\nabla\varphi be a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota with smooth inverse, and let Z>0Z > 0 be a constant for which μ\mu is a finite measure. Write d=𝔼ψud = \mathbb{E}\,\psi u, where u=δμWu = \delta_\mu W.

  1. If det⁡D2φ\det D^2\varphi vanishes nowhere, then d=𝔼∇ψ⋅W=𝔼tr⁡(BD2ψ).d = \mathbb{E}\,\nabla\psi\cdot W = \mathbb{E}\operatorname{tr}\bigl(B\,D^2\psi\bigr).(5)✦

  2. If D2φD^2\varphi is positive definite and D2ψD^2\psi is positive semidefinite at every point, and tr⁡(τD2ψ)\operatorname{tr}(\tau\,D^2\psi) is integrable for μ\mu, then d2≤𝔼tr⁡(τD2ψ)⋅𝔼tr⁡(MBD2ψB).d^2 \le \mathbb{E}\operatorname{tr}\bigl(\tau\,D^2\psi\bigr)\cdot \mathbb{E}\operatorname{tr}\bigl(M\,B\,D^2\psi\,B\bigr).(6)✦

  3. If D2φD^2\varphi is positive definite and D2ψD^2\psi is positive semidefinite at every point, ψ\psi is integrable for μ\mu, and for a real number qq the function ψ\psi satisfies Euler’s identities x⋅∇ψ(x)=qψ(x)x\cdot\nabla\psi(x) = q\,\psi(x) and D2ψ(x)x=(q−1)∇ψ(x)D^2\psi(x)\,x = (q-1)\,\nabla\psi(x) at every xx, then (q−1)2d2≤((q−1)q𝔼ψ)⋅𝔼W⊤D2ψW.(q-1)^2 d^2 \le \bigl((q-1)\,q\,\mathbb{E}\,\psi\bigr)\cdot \mathbb{E}\,W^{\top} D^2\psi\,W.(7)✦

A smooth function homogeneous of degree qq satisfies both of Euler’s identities, so (7) holds for every such ψ\psi that meets the other hypotheses.

What Is Checked

The three claims are stated over the target measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx and its weighted divergence δμ\delta_\mu, the fields τ\tau, MM, BB and WW carried to target coordinates by the homeomorphism Φ=∇φ\Phi = \nabla\varphi, and the two Hessian energies of ψ\psi, for a potential φ\varphi solving the Monge-Ampère equation (1).