Computing solution landscape of nonlinear space-fractional problems via fast approximation algorithm

Bing Yu, Xiangcheng Zheng, Pingwen Zhang, Lei ZhangJournal of Computational Physics 2021 · DOI · arXiv

Space-fractional partial differential equations (PDEs) are essential for modeling "anomalous" physical processes, such as the seepage of pollutants through heterogeneous soil or the evolution of sharp interfaces in fluid dynamics. Unlike standard calculus, fractional derivatives can be of any order, allowing for the representation of non-local effects. However, mapping the "solution landscape"—the complete set of stable and unstable stationary states—of these systems is computationally intensive.

The Computational Bottleneck

The primary challenge in simulating space-fractional systems is the non-locality of the operators. In discrete models, this results in dense matrices that lead to quadratic computational complexity ($O(M^2)$). Furthermore, while finding stable energy minimizers is standard, identifying unstable "saddle points" is necessary to understand system transitions but is significantly more difficult due to the complex hierarchy of states.

Innovation: Fast Approximation and Saddle Dynamics

Researchers have developed a high-efficiency framework to address these challenges through two main innovations:

  1. Fast Approximation Algorithm: By treating the variable fractional order as a power function and applying a series expansion, the spatial and frequency dependencies are decoupled. Specifically, by expanding the Fourier multiplier around a base order, the spatial dependency of the variable fractional order is separated from the frequency dependency of the Laplacian operator. This allows the use of Fast Fourier Transforms (FFT), reducing complexity to near-linear $O(M \ln^2 M)$.
  2. Saddle Dynamics: This method navigates the solution landscape using "upward" and "downward" searches to systematically identify the hierarchy of stationary states, from high-index unstable saddles to stable minimizers. This integrated approach allows for the discovery of complex patterns that were previously too expensive to compute.

Key Results

The algorithm demonstrates exceptional accuracy and scalability across various test cases.

Convergence of the $L^2$ error as a function of the number of expansion terms $S$ for different test functions. The error reaches machine precision ($10^{-16}$) with approximately 25 terms across different grid sizes.

As shown in the figure above, the fast method achieves machine-level precision ($10^{-16}$ error) with as few as 25 terms in the series expansion. This performance is consistent across various grid sizes, ensuring the algorithm scales effectively for large, high-dimensional problems. In 1D and 2D benchmarks, the method maintained near-linear growth in CPU time compared to the quadratic growth of direct methods, enabling the simulation of systems with millions of grid points.

Insights into Fractional Landscapes

The study reveals a profound connection between fractional and local models. The solution landscape of a spectral fractional Laplacian can be reconfigured to resemble a standard integer-order model by varying diffusion coefficients. Additionally, in phase-field models, decreasing the fractional order $\alpha$ acts similarly to decreasing the diffusion coefficient, resulting in sharper interfaces between phases. The authors successfully mapped the 1D and 2D landscapes of fractional Allen-Cahn models, identifying how new "stripe" and "square" stationary states emerge as the fractional order is manipulated.

Conclusion

This new algorithmic approach enables the efficient exploration of complex solution landscapes in fractional-order systems. While currently optimized for periodic boundary conditions, future work will focus on extending these techniques to more complex boundaries and diverse fractional derivative definitions, such as Riemann-Liouville or Caputo. The physical implications of non-gradient landscapes in variable-order systems remain an open area for exploration.

Summary updated

BibTeX
@article{yu2021computing,
  author = {Bing Yu and Xiangcheng Zheng and Pingwen Zhang and Lei Zhang},
  title = {{Computing solution landscape of nonlinear space-fractional problems via fast approximation algorithm}},
  journal = {Journal of Computational Physics},
  year = {2021},
  doi = {10.1016/j.jcp.2022.111513},
}
Figure