Adaptive deep density approximation for fractional Fokker-Planck equations

Li Zeng, Xiaoliang Wan, Tao Zhou

Abstract

In this work, we propose adaptive deep learning approaches based on normalizing flows for solving fractional Fokker-Planck equations (FPEs). The solution of a FPE is a probability density function (PDF). Traditional mesh-based methods are ineffective because of the unbounded computation domain, a large number of dimensions and the nonlocal fractional operator. To this end, we represent the solution with an explicit PDF model induced by a flow-based deep generative model, simplified KRnet, which constructs a transport map from a simple distribution to the target distribution. We consider two methods to approximate the fractional Laplacian. One method is the Monte Carlo approximation. The other method is to construct an auxiliary model with Gaussian radial basis functions (GRBFs) to approximate the solution such that we may take advantage of the fact that the fractional Laplacian of a Gaussian is known analytically. Based on these two different ways for the approximation of the fractional Laplacian, we propose two models, MCNF and GRBFNF, to approximate stationary FPEs and MCTNF to approximate time-dependent FPEs. To further improve the accuracy, we refine the training set and the approximate solution alternately. A variety of numerical examples is presented to demonstrate the effectiveness of our adaptive deep density approaches.

Introduction

The Fokker-Planck equation (FPE) is a fundamental tool in physics and biology used to describe how the probability density of a system evolves over time. While standard FPEs deal with smooth Gaussian noise (Brownian motion), fractional FPEs account for Lévy noise—random movements that include sudden, large jumps. Solving these equations is notoriously difficult due to several factors: the solution must strictly satisfy probability density function (PDF) constraints (non-negativity and unit integration), the domains are often unbounded, and the fractional Laplacian is a nonlocal operator. Furthermore, high-dimensional systems suffer from the "curse of dimensionality," making traditional grid-based methods computationally prohibitive.

Methodology

To overcome these challenges, we propose a framework that represents the PDF using a flow-based generative model, specifically a simplified version of KRnet.

Normalizing Flows and KRnet

Instead of predicting values directly, the model learns an invertible mapping that transforms a simple base distribution (like a Gaussian) into the complex target PDF. This ensures the output is always a valid probability density function without requiring additional penalty terms in the loss function.

The architecture of the simplified KRnet, featuring rotation, scale/bias, affine coupling, squeezing, and nonlinear layers.

Approximating the Fractional Laplacian

We introduce two primary models to handle the nonlocal fractional operator:

  • MCNF (Monte Carlo Normalizing Flow): Uses Monte Carlo sampling to approximate the difficult nonlocal integral. This method is highly flexible and robust for high-dimensional systems.
  • GRBFNF (Gaussian Radial Basis Function NF): Uses an auxiliary model made of Gaussians. Since the fractional Laplacian of a Gaussian is known analytically, this avoids noisy sampling, though it is more computationally intensive for high dimensions.

Adaptive Sampling Strategy

The "secret sauce" of our approach is an adaptive sampling loop. The model starts with a random set of training points and, after initial training, uses the KRnet to generate new training points in regions where the probability density is highest. This self-refining process allows the model to ignore empty space and focus computational resources on the most relevant areas of the domain.

Numerical Results

The framework was evaluated on several synthetic benchmarks, including 2D bimodal distributions and high-dimensional FPEs (4D, 6D, and 8D).

  • Scalability: The MCNF model successfully solved an 8-dimensional FPE, maintaining high accuracy where traditional finite-element methods would fail.
  • Adaptivity Gains: The adaptive sampling strategy improved accuracy by at least one order of magnitude compared to static sampling methods.
  • Complex Topologies: The model accurately captured bimodal (split) distributions, demonstrating its ability to handle complex PDF shapes.
  • Temporal Stability: For time-dependent problems, the MCTNF model utilized a modified temporal coupling layer to ensure initial conditions were met exactly, leading to stable training over time.

Conclusion

This work demonstrates that adaptive deep density approximation using normalizing flows provides a powerful, mesh-free alternative for solving fractional FPEs in high dimensions. While the MCNF method introduces some statistical noise and the GRBFNF method has higher computational costs for specific functions, the overall framework effectively bypasses the curse of dimensionality and strictly enforces PDF constraints. Future work will explore variance reduction techniques and further optimization of the nonlocal operator evaluation.

Summary updated

Abstract

In this work, we propose adaptive deep learning approaches based on normalizing flows for solving fractional Fokker-Planck equations (FPEs). The solution of a FPE is a probability density function (PDF). Traditional mesh-based methods are ineffective because of the unbounded computation domain, a large number of dimensions and the nonlocal fractional operator. To this end, we represent the solution with an explicit PDF model induced by a flow-based deep generative model, simplified KRnet, which constructs a transport map from a simple distribution to the target distribution. We consider two methods to approximate the fractional Laplacian. One method is the Monte Carlo approximation. The other method is to construct an auxiliary model with Gaussian radial basis functions (GRBFs) to approximate the solution such that we may take advantage of the fact that the fractional Laplacian of a Gaussian is known analytically. Based on these two different ways for the approximation of the fractional Laplacian, we propose two models, MCNF and GRBFNF, to approximate stationary FPEs and MCTNF to approximate time-dependent FPEs. To further improve the accuracy, we refine the training set and the approximate solution alternately. A variety of numerical examples is presented to demonstrate the effectiveness of our adaptive deep density approaches.
BibTeX
@article{zeng2023adaptive,
  author = {Li Zeng and Xiaoliang Wan and Tao Zhou},
  title = {{Adaptive deep density approximation for fractional Fokker-Planck equations}},
  journal = {J. Sci. Comput.},
  year = {2023},
  doi = {10.1007/s10915-023-02379-z},
}
Figure