Learning high dimensional stochastic partial differential equations in a predictor-corrector framework

He Zhang, Ran Zhang, Tao ZhouarXiv 2022 · OpenReview

Abstract

Stochastic Partial Differential Equations (SPDEs) are fundamental tools for modeling complex systems subject to random noise, yet they are notoriously difficult to solve in high dimensions due to the "curse of dimensionality." This paper proposes a novel deep learning-based numerical method that utilizes a predictor-corrector splitting strategy. By decomposing the SPDE into a degenerate stochastic component and a deterministic partial differential equation, the framework leverages deep neural networks and Backward Stochastic Differential Equation (BSDE) representations to achieve accurate solutions in up to 100 dimensions with polynomial computational scaling.

Introduction

Stochastic Partial Differential Equations (SPDEs) play a critical role in diverse fields such as population genetics, fluid dynamics, and financial filtering. Despite their utility, traditional numerical methods—including finite element and finite difference methods—fail as the dimensionality ($d$) of the system increases. The computational cost of these methods grows exponentially with $d$, a phenomenon known as the curse of dimensionality.

While recent advancements in deep learning have successfully addressed high-dimensional deterministic PDEs, extending these techniques to the stochastic regime remains a significant challenge. This research bridges that gap by introducing a computationally efficient and mathematically robust algorithm designed for high-dimensional random systems.

Methodology

The Predictor-Corrector Framework

The core of the proposed method is an operator splitting strategy applied at each discrete time step. The original SPDE is decomposed into two distinct sub-problems:

  1. Prediction Step: A degenerate SPDE that captures the stochastic "noise" component of the system. This step is solved using a standard Euler scheme.
  2. Correction Step: A second-order deterministic PDE that accounts for the "drift" and diffusion components.

Deep Learning Integration

To solve the deterministic PDE in the correction step, the authors utilize the Backward Stochastic Differential Equation (BSDE) representation. At each time step, a pair of deep neural networks (DNNs) is trained to approximate the solution ($u$) and its gradient ($\nabla u$).

Unlike global optimization methods that attempt to solve the entire space-time trajectory simultaneously—requiring massive memory resources—this algorithm operates iteratively backward in time. This step-by-step training significantly reduces memory overhead, making the approach feasible for standard computing hardware.

Theoretical Convergence

Beyond the algorithmic implementation, the paper provides a rigorous theoretical foundation. The authors prove the convergence rate of the algorithm and establish error estimates under standard Lipschitz assumptions, ensuring that the deep learning approximations remain grounded in mathematical stability.

Experimental Results

Setup and Benchmarks

The framework was evaluated using synthetic benchmarks where analytical solutions are known, allowing for precise error measurement. The experiments covered dimensions ranging from $d=1$ to $d=100$. The networks were implemented in TensorFlow 2.7 using fully connected layers with Leaky ReLU activations and optimized via the Adam optimizer.

Performance and Scalability

The results demonstrate that the algorithm effectively overcomes the curse of dimensionality:

  • High-Dimensional Accuracy: In 100-dimensional tests, the algorithm achieved a relative $L^2$ error of approximately 0.0098.
  • Stability: The training process proved stable across multiple realizations of Brownian motion, with standard deviations in error ranging from 0.02 to 0.06 in the 100D case.
  • Efficiency: The number of parameters required by the neural networks grows only polynomially with the dimension $d$. This efficiency allowed 100-dimensional problems to be solved on a standard laptop, a feat impossible for traditional grid-based solvers.

Discussion and Future Work

The predictor-corrector framework offers a flexible and memory-efficient path forward for high-dimensional stochastic modeling. By separating stochastic noise from deterministic diffusion, the method allows for the specialized application of numerical tools to each component.

Future research directions include extending the framework to fully nonlinear SPDEs and applying the method to stochastic optimal control problems. Such applications are particularly relevant to quantitative finance and robotics, where high-dimensional, noise-driven decision-making is a central challenge.

Summary updated

BibTeX
@misc{zhang2022learning,
  author = {He Zhang and Ran Zhang and Tao Zhou},
  title = {{Learning high dimensional stochastic partial differential equations in a predictor-corrector framework}},
  year = {2022},
  url = {https://openreview.net/forum?id=z3MZE4yyCn},
}
Figure