Generative Models on Phase Space for High-Energy Physics

A new arXiv preprint (2604.02415) introduces deep generative models that operate directly on the manifold of massless N-particle Lorentz-invariant phase space. These models, including diffusion and flow matching, are designed to exactly enforce physical constraints like energy and momentum conservation, rather than learning them approximately. This approach aims to improve the interpretability and reliability of generative models for high-energy physics data, which consists of relativistic energy-momentum 4-vectors. The method ensures that every step of the sampling trajectory remains confined to the physically allowed submanifold.