TV Distance Estimation for Graphical Models

Develop efficient algorithms to estimate the total variation distance between probability distributions represented as graphical models (e.g., Markov random fields or Bayesian networks), with formal performance guarantees analogous to those obtained for multivariate Gaussians.

Background

The authors' main contribution addresses multivariate Gaussians via reductions and discretization techniques. They explicitly list graphical models as an open direction for TV-distance estimation.

Graphical models are widely used to encode dependencies among variables; establishing computational methods for TV estimation in this setting would expand practical relevance in probabilistic inference.