Discretization and Statistical Consistency of Functional Flow Matching

What the math gives to ML

The paper provides a principled way to train flow-matching velocity fields when functions are observed through finite, scattered, adaptive, or non-nested sensor sets. Its transferable asset is that convergence of conditional velocity targets does not require nested sigma-algebras or a martingale argument; strong consistency of the finite-rank reconstruction is sufficient. This suggests sensor-randomized functional flow matching in which the same neural operator is trained across resolutions and sensor layouts, with each finite-dimensional target represented in the current sensor basis. The resulting model should be less sensitive to discretization and support adaptive sensing or variable-resolution generation without retraining a separate flow model.