A Discrete-Time Random Feature Method for Nonlinear Evolution Equations with Implicit-Explicit Runge-Kutta Time Stepping

Haoran Zhou, Zhaohui Fu, Yangshuai Wang, Xinlong FengarXiv 2026 · DOI

TL;DR

This paper introduces a novel discrete-time Random Feature Method (RFM) for solving nonlinear partial differential equations (PDEs). By combining RFM with an implicit-explicit Runge-Kutta (IMEX-RK) time-stepping scheme, the method efficiently handles both spatial and temporal discretization, reducing complex nonlinear problems to a series of linear least-squares problems. This approach achieves higher accuracy and significantly lower computational cost compared to existing methods like Physics-Informed Neural Networks (PINNs).

What Problem Does This Paper Address?

Solving nonlinear partial differential equations (PDEs) is a fundamental challenge in many scientific and engineering fields. Traditional numerical methods often struggle with the computational cost and complexity associated with these equations, especially when dealing with nonlinearities and stiff time dynamics. Existing machine learning-based approaches, such as Physics-Informed Neural Networks (PINNs), often treat time as an additional input variable, which can lead to issues with causality and efficiency. Furthermore, applying methods like Random Feature Methods (RFM) directly to nonlinear evolutionary PDEs can result in computationally expensive nonlinear least-squares problems. This paper aims to bridge this gap by developing a discrete-time RFM that efficiently handles nonlinear evolutionary PDEs, overcoming the limitations of continuous-time formulations and avoiding nonlinear optimization at each time step.

Methods

The core of this work lies in a discrete-time Random Feature Method (RFM) combined with an implicit-explicit Runge-Kutta (IMEX-RK) time-stepping scheme.

Here's a breakdown of the key methodological aspects:

  • Discrete-Time Approach: Unlike continuous-time methods that treat time as an input, this approach advances the solution step-by-step, ensuring temporal causality. Each time level's solution is computed from previously available states.
  • Spatial Discretization with RFM: The spatial domain is decomposed into subdomains. On each subdomain, a local RFM approximation is constructed using random feature basis functions. These local approximations are then assembled into a global trial function using a partition-of-unity (PoU) mechanism. A key advantage of RFM is that its "hidden" parameters are fixed after random initialization, and the unknown coefficients enter linearly, allowing for efficient least-squares solves.
  • Implicit-Explicit Runge-Kutta (IMEX-RK) Time Stepping: The authors employ an IMEX-RK scheme, specifically a four-stage, third-order accurate method. This scheme treats the stiff linear part of the PDE implicitly and the nonlinear part explicitly. This splitting is crucial because it ensures that each stage of the IMEX-RK scheme remains a linear least-squares problem in terms of the RFM coefficients, thus avoiding computationally intensive nonlinear optimization at each time step.
  • Linear Least-Squares Solves: By combining the RFM spatial representation with the IMEX-RK time stepping, each stage of the nonlinear problem reduces to a linear least-squares formulation. This is a significant advantage over methods that would require solving nonlinear systems at each time step.
  • Global Error Analysis: The paper provides a theoretical analysis of the global error, decomposing it into contributions from the RFM approximation, perturbations in the least-squares coefficients, and the temporal discretization error of the IMEX scheme.

Datasets

Not discussed in the paper. The paper focuses on numerical experiments using well-established benchmark PDEs.

Key Insights and Ideas

  • Bridging RFM and Time-Stepping: The most significant insight is the successful integration of RFM for spatial discretization with a discrete-time IMEX-RK scheme for temporal evolution. This combination elegantly addresses the challenges of nonlinear time-dependent PDEs.
  • Linearization of Nonlinear Problems: The IMEX splitting is key to transforming a potentially intractable nonlinear problem into a sequence of linear least-squares problems at each time step. This is a major computational advantage.
  • Causality and Efficiency: The discrete-time approach inherently enforces temporal causality, which is physically more representative and computationally more efficient than continuous-time space-time formulations for evolutionary problems.
  • RFM's Advantage in Linearity: The paper highlights how RFM's structure, where coefficients enter linearly, is preserved even within the stages of the IMEX-RK scheme, making it computationally attractive.
  • Error Decomposition: The theoretical error analysis provides a clear understanding of the different sources of error (spatial approximation, coefficient perturbation, temporal discretization), which is crucial for guiding future improvements.

Key Results

The proposed discrete-time IMEX-RFM method demonstrated strong performance across a range of nonlinear evolution equations:

  • High Accuracy: Achieved relative L²-errors as low as 5.97 × 10⁻⁶ for the 1D Allen-Cahn equation.
  • Third-Order Convergence: Numerical results consistently showed convergence rates consistent with the third-order IMEX-RK scheme.
  • Significant Speedup: The IMEX-RFM method was substantially faster than a comparable IMEX-PINN variant, with speedup factors ranging from approximately 2.8× to 15.1× across different test cases.
  • Robustness: The method was tested on various PDEs, including Allen-Cahn, Burgers, Korteweg-De Vries (KdV), and Cahn-Hilliard equations, in both one and two spatial dimensions, demonstrating its versatility.
  • Comparison with PINNs: In all tested cases, the IMEX-RFM method outperformed the IMEX-PINN variant in terms of both accuracy and computational cost.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Conditional Stability: The current analysis does not guarantee unconditional stability for the RFM discretization, meaning very stiff problems might still require small time steps.
  • Random Feature Variance: The performance can be sensitive to the random initialization of feature parameters. Different random seeds can lead to variations in errors, as observed in non-monotone convergence curves.
  • Absence of Structure Preservation: The current formulation does not explicitly enforce problem-specific invariants like energy dissipation or mass conservation.
  • Limited Higher-Dimensional Testing: While the method is designed for higher dimensions, its performance and scaling with spatial dimension require more systematic study.

Future work could focus on:

  • Systematic studies in higher dimensions to assess accuracy and cost scaling.
  • Developing adaptive strategies for selecting RFM parameters (M, Jn, Q, τs) to reduce manual tuning.
  • Extending the framework to enforce physical invariants.
  • Investigating variance reduction techniques, such as ensemble averaging or quasi-random sampling, to mitigate sensitivity to random initialization.

Summary updated

BibTeX
@misc{zhou2026discrete,
  author = {Haoran Zhou and Zhaohui Fu and Yangshuai Wang and Xinlong Feng},
  title = {{A Discrete-Time Random Feature Method for Nonlinear Evolution Equations with Implicit-Explicit Runge-Kutta Time Stepping}},
  year = {2026},
  eprint = {2604.25502},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.25502},
}
Figure