SWeissma/Fokker-Planck-particle-systems: First release

Stephanie Reich, Simon WeißmannarXiv 2021 · DOI · arXiv

TL;DR

This paper introduces a unifying framework for Bayesian inference using interacting particle systems derived from the Fokker–Planck equation. By combining preconditioning with a novel localized covariance approach, the authors enable gradient-free sampling that can accurately capture complex, multimodal posterior distributions while maintaining the computational efficiency of Ensemble Kalman-style methods.

What Problem Does This Paper Address?

Bayesian inference—the process of updating a "prior" belief with new data to form a "posterior" distribution—is a cornerstone of modern data science. However, for complex models, the posterior is often impossible to calculate directly.

Traditional Markov Chain Monte Carlo (MCMC) methods are robust but can be painfully slow. On the other hand, "ensemble" methods (like the Ensemble Kalman Filter) are fast and gradient-free, but they typically assume the distribution is nearly Gaussian (unimodal). When the truth is "multimodal" (having multiple peaks), these standard ensemble methods often collapse toward a single peak or average them out into a meaningless blur. The authors seek a method that is as efficient as an ensemble filter but flexible enough to handle the "curse of dimensionality" and multimodality.

Methods

The authors frame Bayesian inference as a dynamic process where a cloud of particles evolves from the prior to the posterior.

  • Fokker–Planck Dynamics: They use the Fokker–Planck equation (which describes how probability densities evolve under diffusion and drift) as a blueprint for particle motion.
  • Wasserstein Gradient Flow: The evolution is treated as a gradient flow that minimizes the Kullback–Leibler (KL) divergence between the particle distribution and the target posterior.
  • Preconditioning & Affine Invariance: By preconditioning the system with the empirical covariance matrix, the authors ensure "affine invariance." This means the algorithm's performance is robust regardless of how the parameters are scaled or rotated.
  • Localized Gradient-Free Approximation: This is the paper's core innovation. Standard gradient-free methods use a global covariance matrix to approximate derivatives. The authors instead calculate localized covariance matrices based on a particle's immediate neighbors. This allows different groups of particles to "sense" and track different modes of a distribution simultaneously.
  • Finite-Size Corrections: They introduce specific correction terms for stochastic (Langevin) systems to prevent ensemble collapse when using a relatively small number of particles.

Datasets

The paper evaluates the proposed methods on several synthetic but mathematically rigorous "toy" problems:

  1. 2D Unimodal Elliptic Problem: A nearly Gaussian benchmark used to verify basic convergence.
  2. 2D Bimodal Problem: A nonlinear forward map designed to produce a posterior with two distinct, separated peaks. This serves as the primary test for the localization strategy.
  3. High-Dimensional Toy Problem: A 1D elliptic boundary-value problem where the unknown is a permeability function. This is discretized using a Karhunen–Loève (KL) expansion, testing the system in up to 32 dimensions.

Key Insights and Ideas

  • Localization is the "Secret Sauce": Global covariance is a "blunt instrument" that forces particles to act as a single unit. Localization allows the ensemble to "break apart" and settle into different regions of the state space, which is essential for non-Gaussian inference.
  • Gradient-Free is Viable for Non-Gaussianity: Usually, gradient-free methods are synonymous with "Gaussian approximations." This paper proves that with the right localization, you can avoid calculating expensive derivatives of complex forward models even in multimodal settings.
  • The Deterministic vs. Stochastic Trade-off: Deterministic systems (based on Reproducing Kernel Hilbert Spaces) provide a direct density estimate but are highly sensitive to "kernel bandwidth" tuning. Stochastic systems (Langevin) are easier to set up but require careful correction terms to avoid variance underestimation.

Key Results

  • Multimodal Success: In the bimodal test case, standard gradient-free methods failed by pushing all particles toward a single peak. The localized version (3.53) successfully captured both peaks of the posterior.
  • Efficiency in High Dimensions: In the 32-dimensional permeability problem, the preconditioned deterministic system showed stable results, though it required a larger ensemble size ($M$) to avoid the "statistical curse of dimensionality."
  • Correction Term Impact: For the Langevin-based sampler, the authors demonstrated that their finite-size correction term (3.58) is critical. Without it, small ensembles (e.g., $M=16$) would collapse into a tiny area; with it, they accurately represented the spread of the target distribution.

Limitations and Future Work

  • Kernel Sensitivity: The deterministic Fokker–Planck approach relies heavily on the choice of the RKHS kernel and its bandwidth ($\alpha$). Choosing these parameters remains a "difficult task" and a general challenge in kernel density estimation.
  • Computational Stiffness: The resulting equations of motion can be "stiff," meaning they require very small time steps to remain stable, especially when particles start far from the target.
  • Future Directions: The authors suggest exploring more efficient time-stepping schemes and combining these particle systems with Stochastic Gradient Descent (SGD) or Stochastic Gradient Langevin Dynamics (SGLD) for even larger-scale applications.

Summary updated

Abstract

This is the first release.
BibTeX
@misc{reich2021sweissma,
  author = {Stephanie Reich and Simon Weißmann},
  title = {{SWeissma/Fokker-Planck-particle-systems: First release}},
  year = {2021},
  eprint = {1911.10832},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.10832},
}
Figure