Uniform-in-time convergence bounds for Persistent Contrastive Divergence algorithms
This work reformulates Persistent Contrastive Divergence (PCD) as a continuous-time system of stochastic differential equations (SDEs). By modeling joint sampling and optimization across two time scales, the authors derive explicit uniform-in-time error bounds comparing PCD iterates to the Maximum Likelihood Estimation (MLE) solution. This provides guaranteed convergence for training energy-based models.