OctoInformBeta

OpenAI preprint, September 2026 · §6. A tensor estimate for the divergence tests

Lemma 6.1: The transported Hessian energy 𝔼tr⁡(MBHB)\mathbb{E}\operatorname{tr}(MBHB) bounds 𝔼tr⁡((DxW)2)\mathbb{E}\operatorname{tr}((D_xW)^2) and 𝔼u2−𝔼W⊤HW\mathbb{E}u^2-\mathbb{E}W^{\top}HW from below

Tensor estimates for a convex function and its moment measure.

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 ι\iota be a finite index set and let φ\varphi be a smooth function on ℝι\mathbb{R}^\iota whose Hessian D2φD^2\varphi is positive definite at every point, so that φ\varphi is strictly convex. Its gradient map x=∇φ(z)x=\nabla\varphi(z) carries the measure ν=e−φdz\nu=e^{-\varphi}\,dz on the source side to a measure on the target side. When that measure is μ=Z−1e−Vdx\mu=Z^{-1}e^{-V}\,dx for a smooth potential VV and a normalizing constant ZZ, it is the moment measure of φ\varphi, and the Monge-Ampère equation behind it reads, 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) Here zz is the source point and xx the target point, τ\tau is D2φD^2\varphi read at xx, M=τ−1M=\tau^{-1} and H=D2VH=D^2V.

A test function ff, smooth with compact support, gives the matrix field B=τD2fτB=\tau\,D^2f\,\tau, a vector field WW obtained by one weighted divergence, and the function u=δμWu=\delta_\mu W, where δμ\delta_\mu is the weighted divergence for μ\mu. The question is how, with 𝔼\mathbb{E} the mean under μ\mu, the energy 𝔼tr⁡((DxW)2)\mathbb{E}\operatorname{tr}((D_xW)^2) of the Jacobian DxWD_xW of WW in target coordinates and the energy 𝔼u2\mathbb{E}u^2 compare with the transported Hessian energy 𝔼tr⁡(MBHB)\mathbb{E}\operatorname{tr}(MBHB), in which the Hessian HH of the potential enters.

Derivatives and Weighted Integrals

The coordinate space ℝι\mathbb{R}^\iota is the real vector space of all functions from ι\iota to ℝ\mathbb{R}, added and scaled coordinatewise✦, and its standard basis vector eie_i is the point with coordinate 11 at ii and 00 at every other index.✦

The directional derivative DvgD_vg of a function gg on a real normed space EE along a vector vv of EE is the function whose value at xx is the Fréchet derivative of gg at xx applied to vv.✦ The directional derivative DvAD_vA of a matrix field AA on EE is the matrix field whose entries are the directional derivatives of the entries of AA.✦

On ℝι\mathbb{R}^\iota, the partial derivative ∂ig\partial_ig of a function gg is the directional derivative DeigD_{e_i}g✦, the gradient ∇g(x)\nabla g(x) is the vector with coordinates ∂ig(x)\partial_ig(x)✦, and the Hessian D2g(x)D^2g(x) is the matrix with entry ∂j∂ig(x)\partial_j\partial_ig(x) in row ii and column jj.✦ The Jacobian DY(x)DY(x) of a vector field Y=(Ya)a∈ιY=(Y_a)_{a\in\iota} is the matrix with entry ∂lYa(x)\partial_lY_a(x) in row aa and column ll.✦ For an index ii, the Hessian derivative CiC_i of φ\varphi is the matrix field Dei(D2φ)D_{e_i}(D^2\varphi), with entries (Ci)ja=∂i∂a∂jφ(C_i)_{ja}=\partial_i\partial_a\partial_j\varphi.✦

The reference measure on ℝι\mathbb{R}^\iota is the measure written dzdz on the source side and dxdx on the target side. The weight of a potential VV is the positive function e−Ve^{-V}✦, and the weighted expectation of a function FF is the integral ∫Fe−Vdx\int F\,e^{-V}\,dx of FF against it.✦ For a constant ZZ, the normalized expectation of FF is 𝔼F=Z−1∫Fe−Vdx,\mathbb{E}F=Z^{-1}\int F\,e^{-V}\,dx,(2)✦ the mean of FF under μ\mu when Z=∫e−VdxZ=\int e^{-V}\,dx.

For a potential VV on EE and a vector vv, the weighted adjoint Dv*D_v^* is the operator that sends a function gg to Dv*g=(DvV)g−DvgD_v^*g=(D_vV)\,g-D_vg.✦ The weighted divergence of a family of functions YiY_i along a family of vectors viv_i is the function ∑iDvi*Yi\sum_iD_{v_i}^*Y_i; for a vector field YY on ℝι\mathbb{R}^\iota, along the standard basis, it is δμY=∑i(∂iVYi−∂iYi),\delta_\mu Y=\sum_i\bigl(\partial_iV\,Y_i-\partial_iY_i\bigr),(3)✦ and with φ\varphi in place of VV it is the divergence δν\delta_\nu for the source measure.

The Fields of a Test Function

Let ff be a real function on ℝι\mathbb{R}^\iota. In source coordinates, with τ=D2φ(z)\tau=D^2\varphi(z), the Hessian of ff pulled by the gradient is the matrix field S(z)=D2f(∇φ(z))S(z)=D^2f(\nabla\varphi(z))✦, and the moment coefficients are the entries of P=τSP=\tau S: Pak(z)=∑i∂a∂iφ(z)∂i∂kf(∇φ(z)).P_{ak}(z)=\sum_i\partial_a\partial_i\varphi(z)\,\partial_i\partial_kf(\nabla\varphi(z)).(4)✦ The moment divergence WW is the vector field whose component WaW_a is the divergence δν\delta_\nu of the row (Pak)k(P_{ak})_k, Wa=∑k(∂kφPak−∂kPak),W_a=\sum_k\bigl(\partial_k\varphi\,P_{ak}-\partial_kP_{ak}\bigr),(5)✦ and J=DWJ=DW is its Jacobian in source coordinates.✦ The pulled matrix is the matrix field B=τSτB=\tau S\tau✦ and the quadratic matrix is the matrix field Q=SτSQ=S\tau S.✦

For a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota, a function gg of the source point is read in target coordinates as the target function g∘Φ−1g\circ\Phi^{-1}.✦ The inverse Hessian field is the matrix field M(x)=D2φ(Φ−1(x))−1M(x)=D^2\varphi(\Phi^{-1}(x))^{-1} of the target point✦; the transported matrix field BB and the transported field WW are the pulled matrix and the moment divergence read in target coordinates, B(Φ−1(x))B(\Phi^{-1}(x)) and Wa∘Φ−1W_a\circ\Phi^{-1}, and DxWD_xW is the Jacobian of the latter. With H=D2V(x)H=D^2V(x), the transported Hessian energy of ff is the function tr⁡(MBHB)\operatorname{tr}(MBHB) of xx.✦

For a vector field YY on ℝι\mathbb{R}^\iota, the VV-Hessian energy of YY is Y⊤HY=∑i,j∂j∂iVYiYjY^\top HY=\sum_{i,j}\partial_j\partial_iV\,Y_iY_j✦ and the square trace of its Jacobian DYDY is tr⁡((DY)2)\operatorname{tr}((DY)^2)✦.

In source coordinates, the first trace term at a point zz is the trace of the product τ−1BHB\tau^{-1}BHB, where H=D2V(∇φ(z))H=D^2V(\nabla\varphi(z)).✦ The trace term LL at the same point zz is the trace of the square of the product Jτ−1J\tau^{-1} of the Jacobian J=DWJ=DW with the inverse Hessian of φ\varphi.✦

The cross term T1T_1, the derivative term T2T_2 of QQ and the cubic term T3T_3 are the following functions of the source point T1=∑a,k,l∂kSal∂lPak,T2=∑i,j,d,k∂dQij(τ−1)dk(Ci)jk,T3=∑i,jQijtr⁡(τ−1Ciτ−1Cj).\begin{aligned} T_1&=\sum_{a,k,l}\partial_kS_{al}\,\partial_lP_{ak},\qquad T_2=\sum_{i,j,d,k}\partial_dQ_{ij}\,(\tau^{-1})_{dk}\,(C_i)_{jk},\\ T_3&=\sum_{i,j}Q_{ij}\operatorname{tr}\bigl(\tau^{-1}C_i\,\tau^{-1}C_j\bigr). \end{aligned}(6)

The Claim

Theorem 1. Let φ\varphi, VV and ff be smooth functions on ℝι\mathbb{R}^\iota.

  1. Suppose that ff has compact support, that the Hessian of φ\varphi is positive definite at every point, that φ\varphi solves (1) for a constant cc, and that ∇φ\nabla\varphi is a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota. Then, with τ=D2φ\tau=D^2\varphi and H=D2V∘∇φH=D^2V\circ\nabla\varphi as fields of the source point, ∫tr⁡(τ−1BHB)e−φdz≤∫tr⁡((Jτ−1)2)e−φdz.\int\operatorname{tr}\bigl(\tau^{-1}BHB\bigr)\,e^{-\varphi}\,dz \le\int\operatorname{tr}\bigl((J\tau^{-1})^2\bigr)\,e^{-\varphi}\,dz.(7)✦

  2. Suppose in addition that Φ−1\Phi^{-1} is smooth and that ∇φ\nabla\varphi pushes ν\nu forward to the target measure Z−1e−VdxZ^{-1}e^{-V}\,dx for a constant Z>0Z>0. Then, with 𝔼\mathbb{E} the normalized expectation (2), MM, BB, WW and H=D2VH=D^2V read at the target point, and u=δμWu=\delta_\mu W, 𝔼tr⁡(MBHB)≤𝔼tr⁡((DxW)2)and𝔼tr⁡(MBHB)+𝔼W⊤HW≤𝔼u2.\mathbb{E}\operatorname{tr}(MBHB)\le\mathbb{E}\operatorname{tr}\bigl((D_xW)^2\bigr) \quad\text{and}\quad \mathbb{E}\operatorname{tr}(MBHB)+\mathbb{E}\,W^\top HW\le\mathbb{E}\,u^2.(8)✦

  3. For every vector field YY on ℝι\mathbb{R}^\iota whose components are smooth with compact support, with H=D2VH=D^2V, ∫(δμY)2e−Vdx=∫Y⊤HYe−Vdx+∫tr⁡((DY)2)e−Vdx.\int(\delta_\mu Y)^2\,e^{-V}\,dx =\int Y^\top HY\,e^{-V}\,dx+\int\operatorname{tr}\bigl((DY)^2\bigr)\,e^{-V}\,dx.(9)✦

  4. If the Hessian of φ\varphi is nonsingular at every point and φ\varphi solves (1) for a constant cc, then, with τ=D2φ\tau=D^2\varphi and H=D2V∘∇φH=D^2V\circ\nabla\varphi as fields of the source point, for all indices ii and jj and at every point, (τHτ−τ)ij=∑d(∂d(∑a(τ−1)da(Ci)ja)−∂dφ∑a(τ−1)da(Ci)ja)−tr⁡(τ−1Ciτ−1Cj).\begin{split} \bigl(\tau H\tau-\tau\bigr)_{ij} ={}&\sum_d\Bigl(\partial_d\Bigl(\sum_a(\tau^{-1})_{da}(C_i)_{ja}\Bigr) -\partial_d\varphi\,\sum_a(\tau^{-1})_{da}(C_i)_{ja}\Bigr)\\ &-\operatorname{tr}\bigl(\tau^{-1}C_i\,\tau^{-1}C_j\bigr). \end{split}(10)✦

  5. If the Hessian of φ\varphi is positive definite at every point, then at every point zz 2T1(z)+T2(z)+T3(z)≥0.2T_1(z)+T_2(z)+T_3(z)\ge0.(11)✦

In the paper, the result reads: for smooth probability measures dμ=Z−1e−Vdxd\mu=Z^{-1}e^{-V}\,dx and dν=e−φdzd\nu=e^{-\varphi}\,dz with D2φD^2\varphi positive definite and x=∇φ(z)x=\nabla\varphi(z) a smooth diffeomorphism pushing ν\nu to μ\mu, every real f∈Cc∞f\in C_c^\infty satisfies 𝔼tr⁡((DxW)2)≥𝔼tr⁡(MBHB)\mathbb{E}\operatorname{tr}((D_xW)^2)\ge\mathbb{E}\operatorname{tr}(MBHB) and 𝔼u2≥𝔼tr⁡(MBHB)+𝔼W⊤HW\mathbb{E}u^2\ge\mathbb{E}\operatorname{tr}(MBHB)+\mathbb{E}W^\top HW, with τ=D2φ\tau=D^2\varphi, M=τ−1M=\tau^{-1}, H=D2VH=D^2V, B=τD2fτB=\tau D^2f\tau, W=δμBW=\delta_\mu B and u=δμWu=\delta_\mu W.

What Is Checked

The claim is stated over smooth functions φ\varphi, VV and ff on the coordinate space ℝι\mathbb{R}^\iota, the Monge-Ampère equation (1), the gradient map of φ\varphi as a homeomorphism relating the source and target measures, the weighted divergence (3) with the weighted and normalized expectations, and the fields built from ff: the moment coefficients (4), the moment divergence (5), the pulled, quadratic and inverse Hessian matrices with their transported versions, and the terms (6).