Wasserstein Gradient Flows and Forward-Only Diffusion Are Not Enough for Multimodal Sampling
What it is
This paper argues that Wasserstein gradient flows (WGF) and forward-only diffusion processes (FODP) suffer from fundamental metastability issues when sampling multimodal distributions. By using the Jordan-Kinderlehrer-Otto scheme and spectral analysis, it demonstrates that these local, gradient-driven methods can require exponentially long mixing times to move probability mass between well-separated modes, regardless of annealing schedules.
The problem it solves
The misleading interpretation of theoretical convergence guarantees for WGF and FODP methods, which fails to account for the exponential mixing times caused by metastability in well-separated multimodal distributions.
What you could use it for
- Evaluating the theoretical limits of diffusion-based generative models
- Designing new sampling algorithms that incorporate nonlocal transport mechanisms
- Understanding the slow-mixing behavior of score-based samplers in complex energy landscapes
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.