Abstract
In the numerical analysis of stochastic processes, establishing strong convergence at a fixed time point is a standard benchmark. However, many applications require a more robust understanding of how the entire path of an approximation converges to the true solution. This paper establishes a theoretical framework showing that if a sequence of piecewise affine linear processes converges strongly at a given rate, it also converges in stronger Hölder norms. The resulting convergence rate is essentially the original rate reduced by the Hölder exponent. This result is applied to derive pathwise convergence rates for spectral Galerkin approximations of various stochastic partial differential equations (SPDEs) and to optimize multilevel Monte Carlo (MLMC) methods in Banach spaces.
Introduction
The study of stochastic differential equations (SDEs) and stochastic partial differential equations (SPDEs) often focuses on "strong convergence," typically defined as the error between the approximation and the exact solution at a specific terminal time $T$. While useful, this metric is insufficient for path-dependent problems, such as pricing Asian options in finance or analyzing the long-term behavior of random dynamical systems.
Proving pathwise convergence—convergence across the entire time interval $[0, T]$—is significantly more challenging. In finite dimensions, one might rely on Doob’s maximal inequality, but this approach frequently fails for SPDEs because their solutions often lack the semi-martingale property. This paper addresses this gap by providing a general methodology to translate discrete-time error bounds into continuous-time pathwise bounds using Hölder norms.
Theoretical Methodology
The authors' approach centers on the regularity of the stochastic process rather than its martingale properties. The core methodology involves three primary components:
- Piecewise Affine Linear Interpolation: The framework considers numerical approximations that are linear between discrete grid points. This is a common structure in many time-stepping schemes (e.g., Euler-Maruyama).
- Error Decomposition and Interpolation: The error in a Hölder norm—which quantifies the smoothness and "wiggliness" of a function—is bounded by combining the maximum error at grid points with the inherent temporal regularity of the process.
- Kolmogorov-Chentsov Extension: The authors adapt the classical Kolmogorov-Chentsov continuity theorem to the setting of Banach-space-valued processes. This allows them to establish continuity and Hölder regularity for numerical approximations in infinite-dimensional spaces.
The Rate-Smoothness Trade-off
A central insight of this work is the quantified trade-off between the strength of the norm and the resulting convergence rate. If a process converges at a rate $\beta$ in a standard $L^p$ sense, its convergence rate in a Hölder norm with exponent $\alpha$ is approximately $\beta - \alpha$.
This "penalty" is a fundamental characteristic of pathwise analysis. Using Brownian motion as a benchmark, the authors demonstrate that this reduction in the convergence rate is essentially unavoidable. To gain information about the smoothness of the path (the Hölder norm), one must accept a slightly slower rate of convergence.
Applications to SPDEs and Monte Carlo Methods
The theoretical results are applied to two major areas of computational mathematics:
Spectral Galerkin Approximations
The paper derives pathwise convergence rates for spectral Galerkin approximations of several non-linear SPDEs, including:
- Stochastic Burgers Equation: Modeling turbulent fluid flow.
- Stochastic Cahn-Hilliard-Cook Equation: Describing phase separation in binary alloys.
- Stochastic Ginzburg-Landau Equation: Used in the study of superconductivity and phase transitions.
A significant technical achievement here is the treatment of "non-globally Lipschitz" coefficients. Many SPDEs involve terms that grow faster than linearly, which typically complicates error analysis. The Hölder-norm framework proves robust enough to handle these complexities.
Multilevel Monte Carlo (MLMC) in Banach Spaces
In infinite-dimensional settings, the efficiency of Monte Carlo methods is tied to the geometric "type" of the Banach space. Many common spaces, such as the space of continuous functions $C([0,T], V)$, have poor "type" properties, which usually suggests poor Monte Carlo performance. However, the authors show that by leveraging path regularity and Hölder convergence, one can derive strong convergence rates for MLMC approximations of expectations. This is vital for high-dimensional uncertainty quantification, where MLMC significantly reduces the computational burden compared to standard Monte Carlo.
Conclusion and Limitations
This research provides a powerful tool for numerical analysts, bypassing the need for semi-martingale structures in pathwise error estimation. However, the framework assumes the target process is strongly Hölder continuous. Consequently, it may not be directly applicable to processes with jumps (such as those driven by Lévy noise) or extremely irregular paths. Future research may extend these Hölder-norm results to other spatial discretization methods, such as finite element or finite difference schemes, further broadening the applicability of this pathwise convergence theory.
Summary updated