A math paper just gave neural networks a rigorous shortcut for solving one of physics' oldest equations.
Researchers studied how to approximate solutions to the Poisson equation, a differential equation used across physics and engineering to model heat flow, electric fields, and fluid behavior, in bounded regions of arbitrary dimension. They show a modified walk on spheres Monte Carlo algorithm can compute these solutions efficiently, needing a number of samples that does not depend on the specific point being evaluated and scales only polynomially with dimension and desired accuracy. From that solver, they then constructively build a ReLU deep neural network that reproduces the same solution, with network size also growing only polynomially in dimension and error tolerance. The paper is a revised, fourth version of a preprint first posted in 2022.
Most work on neural networks solving differential equations is empirical: train a model, check if the loss curve looks good, ship it. This paper instead proves, from first principles, that a network of a specific size and structure solves the Poisson equation with high probability, which matters for anyone building simulation software that needs guarantees rather than a plausible-looking result. It also shows decades-old probabilistic techniques still have work to do justifying today's deep learning shortcuts.
The catch: a constructive existence proof is not a training recipe, so no engineer is handing this paper to PyTorch and getting a working solver by lunchtime.