Solving linear parabolic rough partial differential equations

Christian Bayer, Denis Belomestny, Martin Redmann, Sebastian Riedel, John SchoenmakersJournal of Mathematical Analysis and Applications 2020 · DOI · arXiv

Abstract

Linear parabolic partial differential equations (PDEs) driven by rough signals, such as fractional Brownian motion, present significant challenges for classical numerical solvers due to the non-smooth nature of the driving paths. This paper introduces a novel numerical framework that bridges the gap between abstract rough path theory and practical computation. By integrating a Feynman-Kac stochastic representation with a regression-based Monte Carlo method, the authors develop a pathwise solution that is both computationally efficient and theoretically robust. The approach is validated through convergence analysis and synthetic benchmarks, demonstrating high precision in high-dimensional settings.

Introduction

In disciplines ranging from quantitative finance to stochastic physics, systems are frequently influenced by noise that deviates from the standard Brownian motion model. These systems often exhibit "roughness"—characterized by high-frequency oscillations and low Hölder regularity—and are mathematically modeled using Rough Partial Differential Equations (RPDEs). While the theoretical understanding of RPDEs has advanced through the development of rough path theory, numerical implementations have lagged behind. Standard grid-based PDE solvers often fail when confronted with the non-differentiable nature of rough drivers, and classical Itô or Stratonovich calculus is frequently inapplicable.

This research addresses the need for an efficient, implementable numerical scheme for linear parabolic RPDEs. The proposed method leverages the regularity of the solution in space to apply regression techniques, providing a stable approximation even when the driving signal is highly irregular.

Methodology

The proposed algorithm rests on three methodological pillars designed to handle the interplay between deterministic roughness and stochastic diffusion.

1. Feynman-Kac Representation

The core of the approach is the transformation of the RPDE into a stochastic problem. The solution $u(t, x)$ is represented as the expected value of a functional of a "hybrid" stochastic process $X$. This process is driven by two distinct sources: a standard Brownian motion (representing internal diffusion) and an external deterministic rough path $W$. This representation allows the problem to be framed as a regression task over simulated trajectories.

2. Regression-Based Monte Carlo

To avoid the curse of dimensionality associated with traditional grid-based methods, the authors approximate the solution using a linear combination of basis functions, such as Legendre polynomials. The algorithm employs Monte Carlo simulations to generate training paths. A "pseudo-regression" technique is introduced to optimize performance; by pre-calculating the Gram matrix of the basis functions, the authors significantly reduce the computational overhead typically associated with high-dimensional regression.

3. Simplified Euler Scheme for Rough Paths

Simulating the underlying rough differential equations (RDEs) requires specialized tools. The authors utilize a simplified Euler-type scheme that incorporates higher-order "signature" terms of the rough path. These terms capture the essential geometric information of the driver, ensuring that the simulation remains stable and accurate despite the roughness of the input signal.

Theoretical Insights

Spatial Regularity and Stability

A critical theoretical finding of this work is the proof of spatial regularity. Even when the driving path $W$ is rough, the solution $u(t, x)$ remains smooth with respect to the spatial variable $x$. This regularity provides the mathematical justification for using smooth basis functions in the regression step. Furthermore, the method exhibits pathwise stability, meaning the solution is continuous relative to the driving noise—a property that ensures robustness against small perturbations in the input signal.

Separation of Noise

The algorithm effectively separates the deterministic "roughness" of the external driver from the stochastic "diffusion" of the internal process. This separation allows researchers to adapt standard Monte Carlo variance-reduction tools to the rough path setting, which was previously a significant hurdle in RPDE numerics.

Results and Discussion

The algorithm was evaluated using synthetic benchmarks, including fractional Brownian motion (fBm) with a Hurst index of $H = 0.4$ and systems involving nilpotent matrices.

  • Accuracy: In a 3D nilpotent system where analytical solutions are available for comparison, the algorithm achieved a relative error of less than 1% using $10^6$ samples and 64 basis polynomials.
  • Convergence: The authors established explicit error bounds, showing that the approximation error is a function of the sample size ($M$), the number of basis functions ($K$), and the inherent smoothness of the solution.
  • Temporal Sensitivity: Numerical experiments highlighted that the algorithm is sensitive to the temporal step size ($h$). Capturing the high-frequency behavior of the rough path requires a sufficiently fine time grid, which remains a primary computational constraint.

Limitations and Future Directions

While the current framework provides a robust solution for equations in the whole space ($\mathbb{R}^n$), several challenges remain:

  1. Boundary Conditions: Extending the method to Dirichlet problems on bounded domains requires handling "stopped" rough processes, which is non-trivial in a rough path context.
  2. Variance Reduction: While the regression is effective, the Monte Carlo variance can be substantial. Future work will focus on identifying optimal variance-reduction parameters.
  3. Random Drivers: The current study treats the rough path as deterministic. Adapting the method for fully random rough drivers would require selecting basis functions within the "path space" (using the signature of the path), adding a new layer of complexity to the regression model.

In conclusion, this regression-based approach provides a scalable and theoretically sound path toward solving complex RPDEs, offering a new tool for practitioners dealing with non-standard noise in physical and financial systems.

Summary updated

BibTeX
@article{bayer2020solving,
  author = {Christian Bayer and Denis Belomestny and Martin Redmann and Sebastian Riedel and John Schoenmakers},
  title = {{Solving linear parabolic rough partial differential equations}},
  journal = {Journal of Mathematical Analysis and Applications},
  year = {2020},
  doi = {10.1016/j.jmaa.2020.124236},
}
Figure