A Milstein Scheme for SPDEs

Arnulf Jentzen, Michael RöcknerFound. Comput. Math 2015 · OpenReview

TL;DR

This paper introduces an infinite-dimensional generalization of the Milstein scheme, a powerful numerical method for solving stochastic ordinary differential equations (SODEs). The authors demonstrate that this generalized scheme can be efficiently simulated for a specific class of semilinear stochastic partial differential equations (SPDEs) with multiplicative trace-class noise, offering significant computational advantages over existing methods.

What Problem Does This Paper Address?

Solving stochastic partial differential equations (SPDEs) numerically is crucial for modeling complex systems in fields like physics, finance, and biology. While finite-dimensional stochastic ordinary differential equations (SODEs) have well-established numerical methods like the Milstein scheme, extending these to infinite-dimensional SPDEs presents significant challenges. The Milstein scheme for SODEs is known for its efficiency, particularly when the noise satisfies a commutativity condition. However, directly applying or generalizing this to SPDEs is difficult due to the infinite-dimensional nature of the state space and the complexities of iterated stochastic integrals. Existing methods for SPDEs often require a large number of computational operations and random variables, limiting their practical applicability. This paper aims to bridge this gap by developing and analyzing an infinite-dimensional Milstein-type scheme that retains the efficiency of its finite-dimensional counterpart for a relevant class of SPDEs.

Methods

The core contribution is the development of an infinite-dimensional analog of the Milstein scheme. This involves:

  • Generalizing the Commutativity Condition: The authors introduce an infinite-dimensional version of the commutativity condition that makes the finite-dimensional Milstein scheme efficient. They show that certain semilinear SPDEs with multiplicative trace-class noise naturally satisfy this condition.
  • Adapting the Milstein Algorithm: The paper derives a numerical scheme that incorporates this generalized Milstein approach. This involves careful handling of infinite-dimensional stochastic integrals and their approximations through spatial and temporal discretization.
  • Spatial and Temporal Discretization: Similar to existing methods, the SPDEs are discretized in both space (using spectral Galerkin methods) and time. The key innovation lies in how the Milstein scheme's structure is preserved and adapted to these discretized spaces.
  • Analysis of Convergence and Computational Cost: The authors provide theoretical guarantees for the convergence of their proposed scheme and analyze its computational complexity, comparing it to established methods like the linear implicit Euler scheme.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and numerical examples rather than training on specific datasets.

Key Insights and Ideas

  • Infinite-Dimensional Milstein is Possible and Efficient: The most significant insight is that the efficiency gains of the Milstein scheme can be realized for certain infinite-dimensional SPDEs, contrary to initial intuition that the complexity might be prohibitive.
  • Commutativity is Key: The generalization of the commutativity condition is crucial for simplifying the scheme and avoiding complex iterated stochastic integrals, which are a major hurdle in infinite dimensions.
  • Computational Savings are Substantial: The proposed Milstein-type scheme offers a significant reduction in computational operations and random variables needed to achieve a desired accuracy, making it more practical for complex simulations.
  • Trade-off Between Scheme Complexity and Computational Cost: The paper highlights how a more sophisticated numerical scheme (Milstein-type) can lead to lower overall computational cost by requiring fewer time steps or random variables for a given accuracy.

Key Results

  • Theoretical Convergence Guarantees: The paper provides a convergence theorem (Theorem 1) establishing the rate of convergence for the proposed Milstein-type scheme for a general class of SPDEs.
  • Significant Speedup: Numerical examples, particularly for a stochastic reaction-diffusion equation, demonstrate that the Milstein-type algorithm is orders of magnitude faster than the linear implicit Euler scheme for achieving a precision of three decimals. For instance, one example showed a hundred-fold speedup.
  • Reduced Computational Complexity: The analysis shows that the Milstein-type method requires approximately $O(\epsilon^{-2})$ computational operations and random variables, compared to $O(\epsilon^{-3})$ for the linear implicit Euler scheme, to achieve a precision $\epsilon$.

Limitations and Future Work

  • Specific Class of SPDEs: The developed Milstein scheme is applicable to a specific class of semilinear SPDEs with multiplicative trace-class noise that satisfy the generalized commutativity condition. The authors acknowledge that this is a restriction.
  • Noise Structure: The noise is assumed to be of trace-class type. The paper notes that the treatment of noise that is bounded but not trace-class is an open question.
  • Future Directions:
    • Investigating simpler mollifiers for the exponential terms in the scheme.
  • Combining the temporal approximation with other spatial discretization methods like finite elements to handle more complex domains.
  • Relaxing Lipschitz assumptions on nonlinearities to handle a broader range of SPDEs.
  • Exploring the combination with multilevel Monte Carlo methods for faster statistical quantity approximation.
  • Extending the method to non-Gaussian noise.

Summary updated

Abstract

This article studies an infinite-dimensional analog of Milstein’s scheme for finite-dimensional stochastic ordinary differential equations (SODEs). The Milstein scheme is known to be impressively efficient for SODEs which fulfill a certain commutativity type condition. This article introduces the infinite-dimensional analog of this commutativity type condition and observes that a certain class of semilinear stochastic partial differential equation (SPDEs) with multiplicative trace class noise naturally fulfills the resulting infinite-dimensional commutativity condition. In particular, a suitable infinite-dimensional analog of Milstein’s algorithm can be simulated efficiently for such SPDEs and requires less computational operations and random variables than previously considered algorithms for simulating such SPDEs. The analysis is supported by numerical results for a stochastic heat equation, stochastic reaction diffusion equations and a stochastic Burgers equation, showing significant computational savings.
BibTeX
@inproceedings{jentzen2015milstein,
  author = {Arnulf Jentzen and Michael Röckner},
  title = {{A Milstein Scheme for SPDEs}},
  booktitle = {Found. Comput. Math},
  year = {2015},
  url = {https://openreview.net/forum?id=OJj5eXxsNo},
}
Figure