Nonparametric Drift Estimation for Multidimensional Stochastic Differential Equations under Censoring

What the math gives to ML

The paper gives a mathematically justified way to learn latent diffusion drift fields when trajectories are censored by projection onto time-varying convex regions. Its key transferable asset is that projection-induced finite-variation corrections contribute no mass on visible times, so stochastic estimating equations can be formed from projected observations without reconstructing the hidden state. This suggests a censoring-aware neural diffusion or world-model objective that trains only on visibility-weighted stochastic increments, rather than treating boundary projections as ordinary latent states. The approach is most useful for robotics, medical monitoring, and physical systems with saturation or field-of-view constraints.