Computational Mean-field information dynamics associated with Reaction diffusion equations

Wuchen Li, Wonjun Lee, Stanley OsherJournal of Computational Physics 2021

Abstract

We formulate and compute a class of mean-field information dynamics for reaction-diffusion equations. Given a class of nonlinear reaction-diffusion equations and entropy-type Lyapunov functionals, we study their gradient flow formulations with generalized optimal transport metrics and mean-field control problems. We apply the primal-dual hybrid gradient (PDHG) algorithm to compute the mean-field control problems with potential energies. A byproduct of the proposed method is a new and efficient variational scheme for solving implicit-in-time schemes of mean-field control problems. Several numerical examples demonstrate the solutions of mean-field control problems.

Introduction

Reaction-diffusion equations (RDEs) are fundamental to modeling phenomena in biology, chemistry, and physics, describing how substances spread (diffusion) and interact (reaction). However, solving nonlinear RDEs numerically is challenging. Traditional optimal transport—specifically the Wasserstein metric—is excellent for modeling diffusion as a gradient flow of entropy, but it assumes mass is conserved. In RDEs, "reactions" mean mass is created or destroyed, which the standard Wasserstein framework cannot handle. While "unbalanced optimal transport" has emerged to bridge this gap, there remains a need for a robust, generalized variational framework and efficient numerical algorithms that can handle complex, nonlinear reaction terms without being crippled by the strict stability constraints (like the CFL condition) that plague standard explicit solvers.

Methodology: A Unified Variational Framework

The authors develop a framework called Mean-field Information Dynamics, which treats RDEs as gradient flows within a generalized metric space. This approach consists of three main pillars:

1. Metric Design and Gradient Flows

The authors define a metric space that combines the $L^2$-Wasserstein metric (for transport/diffusion) and the Fisher-Rao metric (for reactions). This allows RDEs to be viewed as gradient descent flows of a Lyapunov functional, such as entropy. Many famous equations (Heat equation, Fisher-KPP, Allen-Cahn) are shown to be specific instances of this general formulation, differing only in their "mobility" functions.

2. Mean-Field Control Formulation

The problem of solving a PDE is transformed into an optimization task: finding the "cheapest" path between two density distributions over time. This is a generalization of the Benamou-Brenier formula from optimal transport. By solving the RDE as an optimization task rather than through direct time-stepping, the authors circumvent the stability issues that usually force researchers to use tiny time steps.

The workflow of the proposed framework, illustrating the transition from reaction-diffusion equations to numerical solvers.

3. G-prox PDHG Algorithm

To solve the resulting saddle-point problem, the paper utilizes a specialized version of the Primal-Dual Hybrid Gradient (PDHG) algorithm. Crucially, they use a "G-prox" step—a preconditioning technique that uses a specific norm related to the differential operators. This makes the algorithm's step-size requirements independent of the spatial grid size, a major advantage over standard finite-difference methods.

Key Results and Insights

  • Numerical Robustness: 2D simulations demonstrate that the algorithm successfully captures complex behaviors, such as mass shifting across the domain while simultaneously growing or decaying based on the reaction term.
  • Efficiency: The use of Fast Fourier Transforms (FFT) within the algorithm allows it to maintain a computational complexity of $O(n \log n)$ per iteration, where $n$ is the number of grid points.
  • Implicit Scheme Byproduct: The method naturally yields a new variational implicit-in-time scheme, which is a generalization of the famous Jordan-Kinderlehrer-Otto (JKO) scheme used in pure diffusion.

Limitations and Future Work

While the framework is robust, for certain nonlinear reaction terms where the mobility function is non-convex, the algorithm is only guaranteed to find a local minimum. Additionally, the current formulation is first-order accurate in time. Future research may explore higher-order schemes and the application of these "natural gradient flows" to finite-dimensional parameterized models, such as neural networks, for AI inference and optimization.

Summary updated

Abstract

We formulate and compute a class of mean-field information dynamics for reaction-diffusion equations. Given a class of nonlinear reaction-diffusion equations and entropy type Lyapunov functionals, we study their gradient flows formulations with generalized optimal transport metrics and mean-field control problems. We apply the primal-dual hybrid gradient algorithm to compute the mean-field control problems with potential energies. A byproduct of the proposed method contains a new and efficient variational scheme for solving implicit in time schemes of mean-field control problems. Several numerical examples demonstrate the solutions of mean-field control problems.
BibTeX
@article{li2021computational,
  author = {Wuchen Li and Wonjun Lee and Stanley Osher},
  title = {{Computational Mean-field information dynamics associated with Reaction diffusion equations}},
  journal = {Journal of Computational Physics},
  year = {2021},
}
Figure