{"id":"arxiv/2608.28949","url":"https://arxiv.org/abs/2608.28949","tags":["stat.ML","cs.AI","cs.LG","math.ST"],"score":0,"title":"The information geometry of product-reference discrete diffusion: Interaction growth complexity and optimal scheduling","domain":"stat.ML","source":"arxiv","authors":["Martin J. Wainwright"],"reviews":0,"subject":"stat.ML","summary":"We study a class of product-reference diffusion algorithms for sampling from a discrete distribution. We show that their sampling performance can be characterized using a path-based measure of data geometry that we call the interaction growth complexity (IGC). We show that a bivariate IGC kernel gives an exact representation of both the KL discretization error and a simple one-step upper bound. The simpler univariate IGC density can be used to study the effect of stepsize choices on the iteration complexity required to obtain $ε$-accurate samples in KL divergence. Samplers that traverse the…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-08-28","reactions":{}}