A recent arXiv paper introduces an algorithm for sampling from composite log-concave distributions, which have densities proportional to e^{-f-g} over R^d. The method is designed for settings where one component, f, is α_f-convex and its gradient is available, while the other component, g, is handled via a proximal approach.
The algorithm builds on proximal gradient techniques, commonly used in optimization, and adapts them for sampling. This is a promising direction because composite log-concave distributions arise in many Bayesian and machine learning applications, where direct sampling is often challenging.
The abstract is brief and does not detail convergence guarantees or experimental results, so the practical performance remains to be seen. Still, the theoretical framing suggests a useful bridge between optimization and sampling for structured densities.