Score Matching

Score matching fits a model by matching its score—the gradient of log density—to data-derived or denoising targets, avoiding the need to evaluate a density’s normalizing constant. Its main families include classical and generalized score matching for exponential families and constrained domains, denoising score matching across noise scales in diffusion models, and methods adapted to manifolds or discrete data. These tools support generative sampling and also appear in causal discovery, inverse problems, scientific simulation, and statistical estimation. A central challenge is translating score accuracy into reliable outcomes: finite-sample error, optimization, boundary behavior, and reverse-diffusion stability can affect generated samples, while standard score-loss measures may not capture those effects.