A kernel compression method for distributed-order fractional partial differential equations

TL;DR

This paper introduces a kernel compression technique for efficiently solving distributed-order fractional PDEs by transforming the non-local problem into a local-in-time system with auxiliary variables, enabling high-dimensional simulations.

Contribution

The authors develop a novel kernel compression algorithm that approximates integral kernels with exponential sums, reducing computational complexity for distributed-order fractional PDEs.

Findings

Achieved exponential kernel approximation with fewer than 100 terms.

Successfully solved 2D and 3D nonlinear DOFPDEs with up to 40 million degrees of freedom.

Demonstrated optimal temporal discretization error decay using graded meshes.

Abstract

We propose a kernel compression method for solving Distributed-Order (DO) Fractional Partial Differential Equations (DOFPDEs) at the cost of solving corresponding local-in-time PDEs. The key concepts are (1) discretization of the integral over the order of the fractional derivative and (2) approximation of linear combinations of integral kernels with exponential sums, expressing the non-local history term as a sum of auxiliary variables that solve a weakly coupled, local in time system of PDEs. For the second step, we introduce an improved algorithm that approximates the occurring integral kernels with double precision accuracy using only a moderate number (<100) of exponential terms. After temporal discretization using implicit Runge--Kutta methods, we exploit the inherent structure of the PDE system to obtain the solution at each time step by solving a single PDE. At the same time, the auxiliary variables are computed by a linear update, not even requiring a matrix-vector multiplication. Choosing temporal meshes with a grading factor corresponding to the convergence order of the Runge--Kutta schemes, we achieve the optimal decay of the temporal discretization error. The flexibility and robustness of our numerical scheme are illustrated by recreating well-studied test cases and solving linear and nonlinear DOFPDEs in 2D and 3D with up to 40 million spatial degrees of freedom.