Signal The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity

Summary

The paper examines masking diffusion, a discrete-sampling technique used in generative modeling, and introduces a new path-based measure of data geometry called unmasking growth complexity, or UGC. The authors show that local increments of this measure directly bound the Kullback-Leibler discretization error. This yields a single theoretical framework covering two different unmasking schemes, one based on Bernoulli subsets and one on fixed cardinality. Working in log-reveal-odds coordinates, the structure produces certified-optimal schedules whose optimality is proven rather than found through empirical tuning. It is a purely theoretical contribution to generative-model training dynamics with no stated application domain.