A new Geometric Setting for the Analysis of Partial Differential Equations

What the math gives to ML

The paper constructs a genuine metric on probability densities by coupling spatial transport with mass-preserving log-compositional reactions, rather than treating these effects as unrelated penalties. Its transferable asset is the dynamic action: a probability distribution can change by moving mass between nearby locations or by reallocating mass locally, while a centered source constraint preserves normalization. This suggests regularizing attention or MoE routing distributions with a geometry-aware proximal cost that distinguishes positional drift from compositional reshaping. The most practical first transfer is a discretized hybrid action used between successive attention distributions, with Wasserstein-only and KL-only penalties as controlled baselines.