Underdamped Diffusion Samplers Outperform Traditional Methods: Researchers from Karlsruhe Institute of Technology, NVIDIA, and Zuse Institute Berlin Introduce a New Framework for Efficient Sampling from Complex Distributions with Degenerate Noise

Underdamped Diffusion Samplers Outperform Traditional Methods: Researchers from Karlsruhe Institute of Technology, NVIDIA, and Zuse Institute Berlin Introduce a New Framework for Efficient Sampling from Complex Distributions with Degenerate Noise

Diffusion processes have emerged as promising approaches for sampling from complex distributions but face significant challenges when dealing with multimodal targets. Traditional methods based on overdamped Langevin dynamics often exhibit slow convergence rates when navigating between different modes of a distribution. While underdamped Langevin dynamics have shown empirical improvements by introducing an additional momentum variable, fundamenta…