Walk-on-Spheres Monte Carlo and deep neural network approximations of elliptic PDEs with drift and killing

What the math gives to ML

The paper provides an explicit randomized representation of elliptic solutions in terms of Brownian exit times, boundary evaluations, and accumulated source terms, together with a Walk-on-Spheres discretization whose error can be controlled uniformly. The transferable asset is not merely the PDE application, but the ability to replace expensive grid-based targets with stochastic computation whose cost is governed by boundary distance and rollout depth. A practical neural-network use is a stochastic target layer for PDE learning: generate controlled-bias targets by recursively jumping to random sphere boundaries, then train a network against those targets. The explicit error decomposition also enables principled allocation of rollout count and depth instead of choosing Monte Carlo budgets heuristically.