$ε$-Strong Simulation for Multidimensional Stochastic Differential Equations via Rough Path Analysis

José Blanchet, Xinyun Chen, Jing DongarXiv 2014 · DOI · arXiv

Abstract

Consider a multidimensional diffusion process $X=\{X(t) : t \in [0,1]\}$. Let $\epsilon > 0$ be a deterministic, user-defined tolerance error parameter. Under standard regularity conditions on the drift and diffusion coefficients of $X$, we construct a probability space supporting both $X$ and an explicit, piecewise constant, fully simulatable process $X_{\epsilon}$ such that $\sup_{0 \leq t \leq 1} \|X_{\epsilon}(t) - X(t)\|_{\infty} < \epsilon$ with probability one. Our construction requires a detailed study of continuity estimates of the Itô map using Lyons’ theory of rough paths. We approximate the underlying Brownian motion, jointly with the Lévy areas, with a deterministic $\epsilon$ error in the underlying rough path metric.

Introduction

In computational finance, physics, and biological modeling, Stochastic Differential Equations (SDEs) are the standard language for describing systems evolving under uncertainty. While numerical methods like the Euler-Maruyama scheme are widely used, they typically offer convergence in a weak sense (averages) or a strong $L^2$ sense (expected error). However, these methods do not guarantee that any single simulated trajectory remains within a fixed distance of the true mathematical solution.

This paper addresses the challenge of $\epsilon$-strong simulation (also known as Tolerance-Enforced Simulation), which aims to produce a path $X_\epsilon$ such that the maximum deviation from the true path $X$ is bounded by a user-defined $\epsilon$ with probability one. While this problem was previously solved for one-dimensional SDEs, the multidimensional case—specifically those with non-constant diffusion coefficients—has remained a significant hurdle due to the complexity of simulating iterated integrals, known as Lévy areas.

Theoretical Framework: The Itô-Lyons Map

The core innovation of this work lies in viewing the SDE solution not merely as a stochastic process, but as a deterministic functional of a "rough path." By employing Lyons’ Rough Path Theory, the authors treat the solution to an SDE as the image of the driving signal (Brownian motion and its associated Lévy areas) under the Itô-Lyons map $\Phi$.

Rough path theory provides explicit continuity estimates for this map. Specifically, if the input signal (the rough path) is approximated within a certain tolerance $\delta$ in the appropriate Hölder-type metric, the resulting output path $X_\epsilon$ is guaranteed to be within $\epsilon$ of the true solution $X$. The problem of simulating the SDE is thus transformed into the problem of simulating the driving Brownian motion and its Lévy areas to a specific, computable precision.

The "Record Breaker" Sampling Strategy

To simulate the driving signal to the required precision, the authors utilize the Lévy-Ciesielski construction, which represents Brownian motion as an infinite series of wavelets. Since an infinite series cannot be computed, the algorithm must determine when the remaining, unsimulated terms are small enough to be ignored without violating the $\epsilon$ guarantee.

The authors implement a "record breaker" algorithm. This strategy uses Bernoulli trials to answer whether the supremum of the remaining terms in the wavelet expansion will ever exceed a specific threshold. If the answer is "no," the series can be safely truncated. This approach allows for the joint simulation of:

  1. Brownian increments.
  2. Lévy areas (the iterated integrals).
  3. Hölder norm bounds required to satisfy the continuity estimates of the Itô-Lyons map.

Implementation and Results

The paper introduces Algorithm II, a recursive procedure that constructs the pathwise approximation. To validate the framework, the authors applied the algorithm to a 2D Geometric Brownian Motion (GBM). GBM serves as an ideal benchmark because its closed-form solution allows for a direct comparison between the $\epsilon$-strong approximation and the ground truth.

Key Findings:

  • Theorem 2.1: The authors establish an explicit, computable error bound $G$ for the multidimensional Euler scheme, expressed in terms of the Hölder norms of the driving rough path.
  • Theorem 2.2: The paper proves that a Monte Carlo procedure can simulate these bounds and path increments jointly in finite time.
  • Numerical Performance: In tests with $\epsilon = 0.1$, the algorithm successfully maintained the required error bound. Interestingly, the authors found that the theoretical constants used to guarantee the bound are quite pessimistic; in practice, the actual error was often significantly smaller than the prescribed $\epsilon$.

Discussion and Future Work

While this research provides a theoretical breakthrough by proving that $\epsilon$-strong simulation is possible for general multidimensional SDEs, it also highlights practical challenges. The computational cost is currently high; the constants required for the guarantee (e.g., $G \approx 10^3$) may necessitate simulating a large number of random variables (up to $2^{20}$ for certain configurations).

Future research will focus on optimizing these bounds and improving the efficiency of the sampling techniques, such as through parallelization or more refined "thinning" of the wavelet terms. Beyond simulation, this framework offers a promising path toward quantifying model uncertainty, as the derived bounds are uniform across large classes of drift and diffusion coefficients.

Summary updated

Abstract

Consider a multidimensional diffusion process $X=\{X\left(t\right) :t\in\lbrack0,1]\}$. Let $\varepsilon&gt;0$ be a \textit{deterministic}, user defined, tolerance error parameter. Under standard regularity conditions on the drift and diffusion coefficients of $X$, we construct a probability space, supporting both $X$ and an explicit, piecewise constant, fully simulatable process $X_{\varepsilon}$ such that \[ \sup_{0\leq t\leq1}\left\Vert X_{\varepsilon}\left(t\right) -X\left(t\right) \right\Vert_{\infty}0$ with probability one. Our construction requires a detailed study of continuity estimates of the Ito map using Lyon's theory of rough paths. We approximate the underlying Brownian motion, jointly with the Lévy areas with a deterministic $\varepsilon$ error in the underlying rough path metric.
BibTeX
@misc{blanchet2014strong,
  author = {José Blanchet and Xinyun Chen and Jing Dong},
  title = {{$ε$-Strong Simulation for Multidimensional Stochastic Differential Equations via Rough Path Analysis}},
  year = {2014},
  eprint = {1403.5722},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1403.5722},
}

Loading related content...

Figure