The Loss Floor of Denoising Score Matching: Fisher Geometry from Schrödinger Bridges
This paper establishes that the irreducible training loss floor in denoising score matching is exactly the integrated trace of the Fisher–Rao metric of the conditional endpoint family, derived via a Schrödinger bridge variational principle, thereby revealing that information geometry is an intrinsic component of diffusion model training and explaining why raw loss values may not consistently rank models across different noise schedules.