Convergence of Numerical Time-Averaging and Stationary Measures via Poisson Equations

Jonathan C. Mattingly, Andrew M. Stuart, Michael V. TretyakovSIAM J. Numer. Anal 2010 · OpenReview

TL;DR

This paper introduces a novel and unified approach using Poisson equations to rigorously analyze the long-term behavior of numerical methods for stochastic differential equations (SDEs). The authors demonstrate that these numerical methods accurately approximate the stationary distribution of the original SDE, providing error bounds that highlight the interplay between numerical integration and statistical errors.

What Problem Does This Paper Address?

Simulating stochastic differential equations (SDEs) is crucial for modeling phenomena in physics, finance, and engineering. A key challenge is understanding the long-term behavior of these systems, specifically their stationary (or invariant) measures, which describe the probability distribution of the system after a long time. Numerical methods are used to approximate SDEs, but it's essential to know how well these approximations capture the true stationary behavior. Specifically, the paper addresses:

  1. Convergence of Numerical Stationary Measures: Does the stationary measure of a numerical scheme converge to the stationary measure of the SDE as the time step goes to zero?
  2. Accuracy of Time-Averaging Estimators: How close are the averages of quantities computed along numerical trajectories to the true long-term averages of the SDE?

Existing work often focused on specific schemes or assumptions (like uniform ellipticity). This paper aims for a more universal and simpler approach applicable to a broader range of SDEs and numerical methods.

Methods

The core of the paper's methodology lies in leveraging the Poisson equation associated with the SDE. For a given function $\phi$, the Poisson equation is $L\psi = \phi - \bar{\phi}$, where $L$ is the generator of the SDE and $\bar{\phi}$ is the stationary average of $\phi$. The authors show that solving this equation provides a powerful tool for analyzing both continuous-time SDEs and their discrete-time numerical approximations.

Key aspects of their method include:

  • Poisson Equation Analysis: They establish the existence and regularity of solutions to the Poisson equation under general hypoelliptic and elliptic conditions.
  • Error Estimates for Time-Averaging: They derive explicit error bounds for estimators of stationary averages computed from numerical trajectories. These bounds relate the error to the time step ($\Delta$) and the total simulation time ($T$).
  • Connection to Weak Convergence: The analysis highlights how the convergence rate of numerical schemes on finite time intervals (weak convergence) directly translates to the convergence of stationary measures.
  • Universality: The approach is designed to be general, working for various explicit and implicit numerical schemes, including those using simple random variables, and for hypoelliptic SDEs.
  • Analogy to Stein's Method: The paper draws parallels between their Poisson equation approach and Stein's method for proving convergence of distributions, suggesting a deeper theoretical connection.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis of SDEs and numerical methods, not empirical evaluation on specific datasets.

Key Insights and Ideas

  • Poisson Equation as a Unifying Tool: The Poisson equation acts as a central analytical tool, bridging the gap between the continuous SDE dynamics and their discrete numerical approximations. Its solution effectively encodes information about the system's long-term behavior.
  • Simplicity and Universality: The approach offers a remarkably simple and general framework for analyzing a wide class of SDEs and numerical schemes, including those that are not uniformly elliptic. This is a significant advantage over more specialized techniques.
  • Direct Link Between Finite-Time and Infinite-Time Behavior: The paper demonstrates a principle: the accuracy of a numerical scheme on finite time intervals directly informs its accuracy in approximating the stationary distribution. This is facilitated by the stability of the SDE (geometric ergodicity).
  • Error Decomposition: The analysis implicitly decomposes the total error into components related to numerical integration (controlled by $\Delta$) and statistical error (controlled by $T$), providing insights into how to optimize simulations.
  • Hypoelliptic SDEs Handled: A notable achievement is the extension of these results to hypoelliptic SDEs, which are common in applications but often challenging to analyze.

Key Results

  • Error Estimates for Time-Averaging Estimators: For first-order schemes (like Euler-Maruyama), the bias of the time-averaging estimator is shown to be $O(\Delta + 1/T)$, and the $L^2$ error is $O(\Delta^2 + 1/T)$. For higher-order schemes, the bias improves to $O(\Delta^p)$ for a scheme of order $p$.
  • Convergence of Stationary Measures: The paper proves that any stationary measure of the numerical method is close to the true stationary measure of the SDE in a specific metric (based on smooth test functions). The distance is bounded by $O(\Delta^p)$ for a $p$-th order scheme.
  • Talay-Tubaro Expansion: An expansion of the global error in powers of the time step $\Delta$ is derived, analogous to the Talay-Tubaro result, providing a more detailed understanding of the error structure.
  • Quantitative Estimates: The results provide quantitative bounds on the errors, allowing for informed choices of simulation parameters ($\Delta$ and $T$) to achieve desired accuracy.

Limitations and Future Work

  • State Space Simplification: For expositional simplicity, the paper primarily considers SDEs on a torus and smooth test functions. Extensions to the whole space ($\mathbb{R}^d$) are anticipated but require additional control over the time spent outside compact regions.
  • Numerical Method Specifics: While the framework is general, specific error bounds and convergence rates depend on the properties of the chosen numerical method and the SDE coefficients.
  • Future Work: The authors suggest extending the approach to SPDEs (stochastic partial differential equations) using infinite-dimensional Poisson equations. They also mention the potential for applying the method to adaptive step-size schemes and exploring the connection to Stein's method more deeply. The analysis of mixing rates for numerical chains generated by more general methods is also left for future study.

Summary updated

Abstract

Numerical approximation of the long time behavior of a stochastic differential equation (SDE) is considered. Error estimates for time-averaging estimators are obtained and then used to show that the stationary behavior of the numerical method converges to that of the SDE. The error analysis is based on using an associated Poisson equation for the underlying SDE. The main advantages of this approach are its simplicity and universality. It works equally well for a range of explicit and implicit schemes, including those with simple simulation of random variables, and for hypoelliptic SDEs. To simplify the exposition, we consider only the case where the state space of the SDE is a torus, and we study only smooth test functions. However, we anticipate that the approach can be applied more widely. An analogy between our approach and Stein's method is indicated. Some practical implications of the results are discussed.MSC codes65C3060H3537H1060H10
BibTeX
@article{mattingly2010convergence,
  author = {Jonathan C. Mattingly and Andrew M. Stuart and Michael V. Tretyakov},
  title = {{Convergence of Numerical Time-Averaging and Stationary Measures via Poisson Equations}},
  journal = {SIAM J. Numer. Anal.},
  year = {2010},
  url = {https://openreview.net/forum?id=sXyPrUUXhi},
}
Figure