Wasserstein Contraction of Stochastic Nonlinear Systems

Jake Bouvrie, Jean-Jacques SlotinearXiv 2019 · DOI · arXiv

Abstract

We suggest that the tools of contraction analysis for deterministic systems can be applied towards studying the convergence behavior of stochastic dynamical systems in the Wasserstein metric. In particular, we consider the case of Itô diffusions with identical dynamics but different initial condition distributions. If the drift term of the diffusion is contracting, then we show that the Wasserstein distance between the laws of any two solutions can be bounded by the Wasserstein distance between the initial condition distributions. Dependence on initial conditions exponentially decays in time at a rate governed by the contraction rate of the noise-free dynamics. The choice of the Wasserstein metric affords several advantages: it captures the underlying geometry of the space, can be efficiently estimated from samples, and advances a viewpoint which begins to bridge the gap between somewhat distinct areas of the literature.

Introduction

Analyzing the stability of nonlinear systems is a fundamental challenge in control theory, and the introduction of stochasticity—representing sensor noise, environmental disturbances, or inherent system variability—compounds this difficulty. In deterministic settings, contraction analysis has emerged as a powerful tool. It demonstrates that trajectories of a system will eventually "forget" their initial conditions and converge toward one another if the system's dynamics are sufficiently contractive.

However, when moving to stochastic systems, such as those governed by Itô diffusions, we shift our focus from individual trajectories to the evolution of probability distributions. Traditional metrics for measuring the distance between these distributions, such as Total Variation (TV) or mean-square distance, often fail to account for the underlying physical geometry of the state space. For instance, TV distance is insensitive to the actual distance between supports if they do not overlap. This paper addresses these limitations by applying contraction analysis within the framework of the Wasserstein metric.

Methodology

1. Contraction Analysis

The authors assume the system's "drift" term—the deterministic component of the dynamics—is contracting. Formally, this implies the existence of a metric where the system's Jacobian is uniformly negative definite. This property ensures that, in the absence of noise, any two trajectories would be forced together over time.

2. Stochastic Modeling

The system is modeled using Itô Stochastic Differential Equations (SDEs) of the form: $$dX_t = f(X_t)dt + \sigma(X_t, t)dB_t$$ where $f$ represents the contracting drift and $\sigma$ represents the noise intensity driven by Brownian motion $B_t$.

3. Optimal Transport and Couplings

To bridge the gap between deterministic contraction and stochastic evolution, the authors utilize Optimal Transport theory. By employing "couplings"—comparing two different solutions starting from distinct initial distributions and driven by independent noise processes—they derive bounds on the evolution of the Wasserstein distance ($W_2$). The $W_2$ distance, often called the "Earth Mover’s Distance," is particularly suited for this analysis as it measures the minimum "work" required to transform one distribution into another, inherently respecting the geometry of the space.

Key Results: Theorem 2

The central contribution of the paper is Theorem 2, which provides a concrete upper bound for the Wasserstein distance between two evolving distributions $\mu_t$ and $\nu_t$:

$$W_2(\mu_t, \nu_t) \le \alpha^{-1/2} \left( e^{-\beta t} W_2(\mu_0, \nu_0) + \sqrt{C_\sigma/\beta} \right)$$

In this expression:

  • $\beta$ represents the contraction rate of the deterministic drift.
  • $C_\sigma$ represents the noise intensity.
  • $\alpha$ relates to the bounds of the transformation matrix.

This result reveals two critical behaviors. First, the influence of the initial state ($W_2(\mu_0, \nu_0)$) vanishes exponentially at a rate determined by the deterministic dynamics. Second, the noise introduces a "steady-state" floor (the $\sqrt{C_\sigma/\beta}$ term), meaning that while distributions converge, they will never perfectly overlap due to independent stochastic perturbations.

To validate this theoretical bound, the authors apply it to an Ornstein-Uhlenbeck (OU) process. The analytical results for this linear SDE confirm that the proposed bound accurately predicts the exponential convergence observed in exact solutions.

Discussion and Future Work

This work provides a geometrically intuitive way to guarantee the stability of noisy nonlinear systems, with significant implications for machine learning—particularly in understanding Langevin dynamics for sampling and non-convex optimization.

However, several avenues for future research remain:

  • Conservatism: The current bound may be conservative because it assumes independent noise processes. Analyzing the Fokker-Planck equation directly might yield tighter bounds.
  • State-Dependent Metrics: While this proof utilizes time-varying metrics, extending the framework to state-dependent metrics $M(x, t)$ would broaden its applicability to complex robotic control.
  • Networked Systems: The authors suggest that these Wasserstein bounds could be instrumental in analyzing the stability of large-scale interconnected stochastic systems, such as neural networks or power grids.

Summary updated

Abstract

We suggest that the tools of contraction analysis for deterministic systems can be applied towards studying the convergence behavior of stochastic dynamical systems in the Wasserstein metric. In particular, we consider the case of Ito diffusions with identical dynamics but different initial condition distributions. If the drift term of the diffusion is contracting, then we show that the Wasserstein distance between the laws of any two solutions can be bounded by the Wasserstein distance between the initial condition distributions. Dependence on initial conditions exponentially decays in time at a rate governed by the contraction rate of the noise-free dynamics. The choice of the Wasserstein metric affords several advantages: it captures the underlying geometry of the space, can be efficiently estimated from samples, and advances a viewpoint which begins to bridge the gap between somewhat distinct areas of the literature.
BibTeX
@misc{bouvrie2019wasserstein,
  author = {Jake Bouvrie and Jean-Jacques Slotine},
  title = {{Wasserstein Contraction of Stochastic Nonlinear Systems}},
  year = {2019},
  eprint = {1902.08567},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1902.08567},
}

Loading related content...

Figure