Fully nonlinear parabolic equations are notoriously difficult to solve numerically, especially in high dimensions. The paper introduces neural stochastic residual methods that use a quadratic residual term and exact diffusion sampling. This formulation enables both global training and backward regression with scalar networks, and the loss functions involve only spatial derivatives.
The authors provide an error analysis: under a uniform output regularity budget, the continuous residual norm is bounded by the square root of the global loss or the normalized backward-regression excess loss, plus an O(h^(1/2)) discretization term and terminal error. Residual stability then yields bounds on value, gradient, and Hessian errors. Under suitable assumptions, they construct admissible networks with O(h^(1/2)) bounds on approximation errors and on the square roots of both loss criteria.
In ten-dimensional backward-regression tests using a shared schedule tuned for the quadratic residual, mean relative Hessian errors were 5–6% with the quadratic residual and about 15% with the first-order residual. Tuned backward regression achieved Hessian accuracy comparable to physics-informed neural networks across the tested nonlinear coefficients. However, translation tests showed that the ratio of Hessian error to residual can grow sharply under a fixed sampling law, highlighting a practical caveat.