OctoInformBeta

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

τ, the operator A, and the tests W and u (4.2), (4.4): The target Hessian τ\tau, the operator AA, the flux τ∇f\tau\nabla f, the weighted divergence and the source and target measures

The moment operator and its test-function fields in target coordinates.

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 φ\varphi be a convex function on ℝn\mathbb{R}^n whose gradient map ∇φ\nabla\varphi is a homeomorphism of ℝn\mathbb{R}^n onto itself and pushes the source measure ν=e−φdz\nu = e^{-\varphi}\,dz forward to a constant multiple of the target measure μ=Z−1e−Vdx\mu = Z^{-1}e^{-V}\,dx, where VV is a real function on ℝn\mathbb{R}^n, the potential, and Z=∫e−VdxZ = \int e^{-V}\,dx. This is the moment measure picture behind the Monge-Ampère equation log⁡det⁡D2φ=V(∇φ)+c−φ\log\det D^2\varphi = V(\nabla\varphi) + c - \varphi, where cc is a real constant. Source coordinates zz carry ν\nu, and target coordinates x=∇φ(z)x = \nabla\varphi(z) carry μ\mu.

Weighted integration by parts under μ\mu happens in xx, while φ\varphi and its Hessian are known at zz. The objects that pass between the two are the Hessian τ\tau of φ\varphi read as a function of xx, the second-order operator AA it defines on test functions of xx, and the flux of a test function, obtained by applying τ\tau to its gradient. For a test function ff there are also a matrix field BB and a vector field WW, the weighted divergence of BB under μ\mu; both are built in source coordinates, and a function of source coordinates is read in target coordinates by composing it with the inverse of the gradient map.

Coordinates, Derivatives and Weights

The coordinate space ℝι\mathbb{R}^\iota of a finite index set ι\iota is the real vector space of all functions from ι\iota to ℝ\mathbb{R}: a point has one real coordinate for each index, and points are added and scaled coordinatewise.✦ It is ℝn\mathbb{R}^n when ι\iota has nn elements. The standard basis vector eie_i of ℝι\mathbb{R}^\iota has coordinate 11 at the index ii and 00 at every other index.✦

The directional derivative DvfD_v f of a function ff on a real normed space EE along a vector vv of EE is the function x↦Df(x)vx \mapsto Df(x)\,v, where Df(x)Df(x) is the Fréchet derivative of ff at xx.✦ On ℝι\mathbb{R}^\iota, the partial derivative ∂if\partial_i f is DeifD_{e_i} f.✦ The gradient ∇f(x)\nabla f(x) of a function ff on ℝι\mathbb{R}^\iota at a point xx is the vector of its partial derivatives, and its Hessian D2f(x)D^2 f(x) is the real ι×ι\iota\times\iota matrix of its second partial derivatives: ∇f(x)=(∂if(x))i∈ι,\nabla f(x) = \bigl(\partial_i f(x)\bigr)_{i\in\iota},(1)✦ D2f(x)ij=∂j∂if(x)(i,j∈ι).D^2 f(x)_{ij} = \partial_j\partial_i f(x) \qquad (i, j \in \iota).(2)✦

The weight of a potential VV on EE is the positive function e−Ve^{-V}.✦ For a vector vv of EE, the weighted adjoint Dv*D_v^* of DvD_v for this weight sends a function gg on EE to Dv*g=(DvV)g−DvgD_v^* g = (D_v V)\,g - D_v g.✦ The weighted divergence δVY\delta_V Y of a family Y=(Yi)i∈ιY = (Y_i)_{i\in\iota} of functions on EE, along a family (vi)i∈ι(v_i)_{i\in\iota} of vectors of EE, is the function ∑iDvi*Yi\sum_i D_{v_i}^* Y_i: δVY(x)=∑i∈ι(DviV(x)Yi(x)−DviYi(x)).\delta_V Y(x) = \sum_{i\in\iota}\bigl(D_{v_i}V(x)\,Y_i(x) - D_{v_i}Y_i(x)\bigr).(3)✦

Fix a measure mm on ℝι\mathbb{R}^\iota; in the setting it is Lebesgue measure. The source measure ν\nu of a function φ\varphi on ℝι\mathbb{R}^\iota is the measure with density e−φe^{-\varphi} with respect to mm, and the target measure μ\mu of a potential VV and a constant ZZ is the measure with density Z−1e−VZ^{-1}e^{-V} with respect to mm: dν(z)=e−φ(z)dm(z),d\nu(z) = e^{-\varphi(z)}\,dm(z),(4)✦ dμ(x)=Z−1e−V(x)dm(x).d\mu(x) = Z^{-1}e^{-V(x)}\,dm(x).(5)✦

The Definition

Definition 1. Let φ\varphi and ff be real functions on ℝι\mathbb{R}^\iota, and let Φ\Phi be a homeomorphism of ℝι\mathbb{R}^\iota onto itself; in the setting, Φ\Phi is the gradient map ∇φ\nabla\varphi. The target Hessian of φ\varphi is the matrix field τ(x)=D2φ(Φ−1(x)).\tau(x) = D^2\varphi\bigl(\Phi^{-1}(x)\bigr).(6)✦ The moment operator AA sends ff to the function Af(x)=x⋅∇f(x)−τ(x):D2f(x),Af(x) = x\cdot\nabla f(x) - \tau(x) : D^2 f(x),(7)✦ where τ:D2f=∑i,jτij(D2f)ij\tau : D^2 f = \sum_{i,j}\tau_{ij}\,(D^2 f)_{ij}. The flux τ∇f\tau\nabla f of ff is the vector field (τ∇f)(x)=τ(x)∇f(x).(\tau\nabla f)(x) = \tau(x)\,\nabla f(x).(8)✦ In source coordinates, for indices aa and kk, let Pak(z)=∑iD2φ(z)aiD2f(∇φ(z))ikP_{ak}(z) = \sum_i D^2\varphi(z)_{ai}\,D^2 f(\nabla\varphi(z))_{ik} be the entries of the product of D2φ(z)D^2\varphi(z) with the Hessian of ff pulled back by ∇φ\nabla\varphi, and let Pa=(Pak)k∈ιP_a = (P_{ak})_{k\in\iota}.✦ The weighted divergence δφPa\delta_\varphi P_a of PaP_a for the potential φ\varphi, taken as in (3) along the standard basis, is (δφPa)(z)=∑k∈ι(∂kφ(z)Pak(z)−∂kPak(z)).(\delta_\varphi P_a)(z) = \sum_{k\in\iota}\bigl(\partial_k\varphi(z)\,P_{ak}(z) - \partial_k P_{ak}(z)\bigr).(9)✦ Let Bˆ(z)=D2φ(z)D2f(∇φ(z))D2φ(z)\hat B(z) = D^2\varphi(z)\,D^2 f(\nabla\varphi(z))\,D^2\varphi(z).✦ A function gg of source coordinates is read in target coordinates as g∘Φ−1g\circ\Phi^{-1}✦. The matrix field BB is Bˆ\hat B read in target coordinates, and the vector field W=(Wa)a∈ιW = (W_a)_{a\in\iota} has as components the functions δφPa\delta_\varphi P_a read in target coordinates: B(x)=Bˆ(Φ−1(x)),B(x) = \hat B\bigl(\Phi^{-1}(x)\bigr),(10)✦ Wa(x)=(δφPa)(Φ−1(x)).W_a(x) = (\delta_\varphi P_a)\bigl(\Phi^{-1}(x)\bigr).(11)✦

What Is Checked

The definitions are stated over the coordinate space ℝι\mathbb{R}^\iota of a finite index set ι\iota, with the weighted adjoint and the weighted divergence over any real normed space and every derivative a Fréchet derivative, and their data are a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota onto itself, real functions φ\varphi, ff and VV, a real constant ZZ and a measure mm on ℝι\mathbb{R}^\iota.