OctoInformBeta

OpenAI preprint, September 2026 · §4. Adjoints in moment coordinates

The integration formula (4.3) and the gradient identity (4.6): Under the Monge-Ampère equation, AfAf is the weighted divergence of τ∇f\tau\nabla f, and WW read at xx is τ∇(A−1)f\tau\nabla(A-1)f, with weighted divergence A(A−1)fA(A-1)f

Integration by parts and flux identities for the moment operator.

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

Moment coordinates use two copies of ℝn\mathbb{R}^n. The source, with coordinates zz, carries a smooth function φ\varphi and the measure ν=e−φdz\nu = e^{-\varphi}\,dz; the target, with coordinates xx, carries a smooth potential VV and the measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx for a constant Z>0Z > 0. The two sides are tied, in terms of the gradient ∇φ\nabla\varphi and the Hessian D2φD^2\varphi of φ\varphi, by 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),(1) with cc a constant and det⁡D2φ\det D^2\varphi nowhere zero, and by the gradient map Φ=∇φ\Phi = \nabla\varphi, which sends zz to x=∇φ(z)x = \nabla\varphi(z) and is a homeomorphism of ℝn\mathbb{R}^n with smooth inverse.

On the target side the Hessian of φ\varphi, carried over by Φ\Phi, supplies the coefficients of a second-order operator, the moment operator AA defined below. Against μ\mu, AA is computed in two ways, by integration by parts and as the weighted divergence of a vector field, and the second computation is repeated one order up, for a field WW built from the second derivatives of a function.

Operators in Moment Coordinates

Derivatives are taken in coordinates: ∂if=Deif\partial_i f = D_{e_i}f is the directional derivative of ff along the standard basis vector eie_i of ℝn\mathbb{R}^n, where Dvf(x)=Df(x)vD_vf(x) = Df(x)\,v. The gradient ∇f\nabla f is the vector of the ∂if\partial_i f✦, and the Hessian D2fD^2f is the matrix with entries ∂j∂if\partial_j\partial_i f✦.

The moment matrix τ\tau is the field of n×nn\times n matrices on the target given by the Hessian of φ\varphi at the source point of xx, τ(x)=D2φ(Φ−1(x)),\tau(x) = D^2\varphi(\Phi^{-1}(x)),(2)✦ and in the same way a function gg on the source is read in target coordinates as g∘Φ−1g\circ\Phi^{-1}✦.

For a field σ\sigma of n×nn\times n matrices on ℝn\mathbb{R}^n, the operator AσA_\sigma sends a smooth function ff to the function Aσf(x)=∑ixi∂if(x)−∑i,jσij(x)∂i∂jf(x)=x⋅∇f(x)−σ(x):D2f(x),A_\sigma f(x) = \sum_i x_i\,\partial_i f(x) - \sum_{i,j}\sigma_{ij}(x)\,\partial_i\partial_j f(x) = x\cdot\nabla f(x) - \sigma(x) : D^2f(x),(3)✦ and the gradient form of σ\sigma takes two smooth functions hh and ff to the function Γσ(h,f)(x)=∇h(x)⋅σ(x)∇f(x)\Gamma_\sigma(h,f)(x) = \nabla h(x)\cdot\sigma(x)\nabla f(x)✦. For σ=τ\sigma = \tau, the operator AτA_\tau is the moment operator, written AA, and (A−1)f=Af−f(A-1)f = Af - f.✦

The flux τ∇f\tau\nabla f of a smooth function ff is the vector field on the target with components (τ∇f)i=∑jτij∂jf(\tau\nabla f)_i = \sum_j \tau_{ij}\,\partial_j f.✦

The weighted adjoint of the directional derivative DvD_v for the weight e−Ve^{-V} is the operator Dv*g=(DvV)g−DvgD_v^*g = (D_vV)\,g - D_vg✦, and the weighted divergence of a vector field Y=(Y1,…,Yn)Y = (Y_1,\dots,Y_n) on the target is δμY=∑jDej*Yj=∑j(∂jVYj−∂jYj).\delta_\mu Y = \sum_j D_{e_j}^*Y_j = \sum_j \big(\partial_jV\,Y_j - \partial_jY_j\big).(4)✦ For a vector field YY on the source, the weighted divergence δνY\delta_\nu Y for the weight e−φe^{-\varphi} is given by (4) with φ\varphi in place of VV.

The moment coefficients of a smooth function ff on ℝn\mathbb{R}^n are the entries of the matrix field PP on the source given by P(z)=D2φ(z)D2f(∇φ(z)),Pak(z)=∑i∂i∂aφ(z)∂k∂if(∇φ(z)).P(z) = D^2\varphi(z)\,D^2f(\nabla\varphi(z)), \qquad P_{ak}(z) = \sum_i \partial_i\partial_a\varphi(z)\,\partial_k\partial_i f(\nabla\varphi(z)).(5)✦ For each index aa, the moment divergence WaW_a of ff is the function on the source given by the weighted divergence of the aa-th row of PP, Wa=δν(Pa1,…,Pan)=∑k(∂kφPak−∂kPak),W_a = \delta_\nu(P_{a1},\dots,P_{an}) = \sum_k \big(\partial_k\varphi\,P_{ak} - \partial_kP_{ak}\big),(6) The vector field W=(W1,…,Wn)W = (W_1,\dots,W_n) on the source collects these functions.

The Claim

Theorem 1. Let φ\varphi and VV be smooth functions on ℝn\mathbb{R}^n such that det⁡D2φ\det D^2\varphi vanishes nowhere and the Monge-Ampère equation (1) holds for a constant cc, let the gradient map Φ=∇φ\Phi = \nabla\varphi be a homeomorphism of ℝn\mathbb{R}^n with smooth inverse, and let ff be a smooth function on ℝn\mathbb{R}^n. If ff has compact support, then for every smooth function hh and every Z>0Z > 0, ∫hAfdμ=∫Γτ(h,f)dμ.\int h\,Af\,d\mu = \int \Gamma_\tau(h,f)\,d\mu.(7)✦ For every smooth ff, compactly supported or not, AfAf is the weighted divergence of the flux of ff, δμ(τ∇f)=Af,\delta_\mu(\tau\nabla f) = Af,(8)✦ the field WW read in target coordinates is the flux of (A−1)f(A-1)f, Wa∘Φ−1=(τ∇(A−1)f)afor each index a,W_a\circ\Phi^{-1} = \big(\tau\nabla(A-1)f\big)_a \quad\text{for each index } a,(9)✦ and its weighted divergence is δμ(W∘Φ−1)=A(A−1)f.\delta_\mu(W\circ\Phi^{-1}) = A(A-1)f.(10)✦ The identity (9) holds without the Monge-Ampère equation and without the condition on det⁡D2φ\det D^2\varphi.

What Is Checked

The four identities are stated over smooth functions φ\varphi, VV, ff and hh on ℝn\mathbb{R}^n, with det⁡D2φ\det D^2\varphi nowhere zero, the Monge-Ampère equation (1) and the gradient map Φ\Phi a homeomorphism with smooth inverse, and over the moment matrix (2), the operator and gradient form of (3), the flux τ∇f\tau\nabla f, the weighted divergence (4) for the weights e−Ve^{-V} and e−φe^{-\varphi}, the measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx and the field WW of (6).