Improved training for stochastic flow maps via Legendre polynomials

The method addresses non-differentiability in Brownian paths by representing them with a shifted Legendre polynomial expansion. Using a Wong-Zakai styled objective, the approach enables derivative-based losses for training stochastic differential equation solvers.