A Deep Second-Order Stochastic Residual Method for Fully Nonlinear Parabolic PDEs

arXiv:2607.16730 2026 Training 1 ideas extracted · analyzed Aug 30, 2026

What the math gives to ML

The paper provides a constructive way to train one scalar network whose automatic derivatives simultaneously represent a PDE solution, its gradient, and its Hessian, rather than fitting separate derivative networks. Its key transferable asset is a second-order Brownian one-step residual: centered quadratic Gaussian increments expose Hessian information through a stochastic Taylor identity while avoiding explicit third- and fourth-order PDE residuals. The most promising neural-network transfer is a Hessian-aware stochastic residual objective for high-dimensional parabolic PDE solvers, with terminal gradient supervision and occupation-weighted full-jet regularization.

Ideas from this paper

Failed on benchmark 2026

Second-Order Brownian Jet Residual

Replace pointwise high-order PINN residuals with a stochastic one-step residual evaluated on Brownian transitions. A single scalar network produces the value, gradient, and Hessian by automatic differentiation, and the quadratic centered increment supplies a stochastic probe of the Hessian. Add a terminal gradient penalty so the learned full jet is constrained at the terminal boundary, not only the scalar value.

Useful8/10
Difficulty5/10
Novelty7/10
Paper: A Deep Second-Order Stochastic Residual Method for Fully Nonlinear Parabolic PDEs arXiv:2607.16730