Occupation-weighted Hessian contraction for PINNs
Implementation & benchmark of arXiv:2608.16119 — Fully nonlinear parabolic equations under a fixed reference diffusion:weighted $L^2$ Hessian estimates and well-posedness
Source paper: Fully nonlinear parabolic equations under a fixed reference diffusion:weighted $L^2$ Hessian estimates and well-posedness arXiv:2608.16119 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Train a neural PDE solver using collocation points sampled from a fixed reference diffusion and a time weight that compensates for the point-start singularity. Replace the Euclidean Hessian by the intrinsic tensor Gθ=σD²uθσ, and use source Picard updates so that nonlinear curvature coupling is iterated under an explicit contraction target.
Formulas
Mathematical statement
The reference diffusion is dX_t=b(t,X_t)dt+σ(t,X_t)dW_t, with covariance a=σσ⊤ and generator L_t=b·D_x+(1/2)a:D_x². The paper defines the intrinsic derivatives Z u=σD_xu and G u=σD_x²uσ, and the backward operator P=-(∂_t+L_t). Let μ_t be the law of X_t and let ρ_w(dt,dx)=w(t)dtμ_t(dx), where w(t)>0 is a time weight. For the zero-terminal linear problem Pv=q, v(T,·)=0, C_w(X) is the optimal source-to-intrinsic-Hessian norm satisfying ||Gv||_{ρ_w}≤C_w(X)||q||_{ρ_w}. If the nonlinear driver f(t,x,u,Zu,Gu) is L_H-Lipschitz in its Hessian argument, source Picard iteration contracts when C_w(X)L_H<1. In the Brownian case the supplied benchmark is C_α^Br=sup_{n≥2}2√n(n-1)/(n-1+α), for a weight of order α.
Implementation notes
(1) Integration point: modify the residual sampler and curvature residual of a terminal-value PINN. Let uθ(t,x) be an MLP and let fφ(t,x,u,z,g) be the PDE driver. Compute D_tuθ, D_xuθ, and D_x²uθ by automatic differentiation, then form zθ=σD_xuθ and gθ=σD_x²uθσ. Sample x at time t from trajectories of the fixed reference SDE instead of from a uniform spatial grid. Use w(t)=t^{α-1}, initially α=1/2, with importance correction if samples are drawn from another time proposal.
(2) Pseudocode: simulate X[j] with Euler-Maruyama from x₀; sample time index j with probability proportional to w(t[j]); evaluate r=fφ(t[j],X[j],uθ,zθ,gθ)+D_tuθ+b·D_xuθ+(1/2)a:D_x²uθ; minimize the weighted loss w(t[j])|r|² plus terminal loss |uθ(T,x)-ψ(x)|². For source Picard, initialize q⁰=0, fit a network vᵏ to Pvᵏ=qᵏ, compute qᵏ⁺¹=fφ(t,x,vᵏ,Zvᵏ,Gvᵏ) with vᵏ detached while fitting the next source, and stop when the weighted source difference is below tolerance.
(3) Compute the intrinsic derivatives and occupation norm directly. Compute C_α^Br from the displayed supremum. Estimate L_H by drawing matched tuples (t,x,u,z) and two Hessians g₁,g₂, using L̂_H=max ||fφ(...,g₁)-fφ(...,g₂)||/(||g₁-g₂||+ε). Track the margin m=1-C_α^BrL̂_H. If m≤0, reduce the g-channel scale or apply spectral normalization to that channel; do not interpret this as a proof for arbitrary learned drivers. Compare against uniform collocation, Euclidean Hessians, and direct residual training without Picard iteration.
(4) First experiment: solve a two-dimensional Allen-Cahn or HJB terminal-value PDE with a small tanh MLP. Use identical width, parameter count, collocation budget, and optimizer for uniform PINN and occupation-weighted PINN, with and without Picard source updates. Evaluate dense-grid relative solution error, weighted PDE residual, terminal error, optimizer steps to a fixed error, and the empirical Picard ratio ||qᵏ⁺¹-qᵏ||/|qᵏ-qᵏ⁻¹|. Success means improved behavior near t=0, a ratio below one when the estimated threshold holds, and at least 1.5x fewer steps or lower error at equal collocation cost.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.