Abstract:In this paper we propose a new approach for sampling from probability measures in, possibly, high dimensional spaces. By perturbing the standard overdamped Langevin dynamics by a suitable Stratonovich perturbation that preserves the invariant measure of the original system, we show that accelerated convergence to equilibrium and reduced asymptotic variance can be achieved, leading, thus, to a computationally advantageous sampling algorithm. The new perturbed Langevin dynamics is reversible with respect to the target probability measure and, consequently, does not suffer from the drawbacks of the nonreversible Langevin samplers that were introduced in~[C.-R. Hwang, S.-Y. Hwang-Ma, and S.-J. Sheu, Ann. Appl. Probab. 1993] and studied in, e.g. [T. Lelievre, F. Nier, and G. A. Pavliotis J. Stat. Phys., 2013] and [A. B. Duncan, T. Lelièvre, and G. A. Pavliotis J. Stat. Phys., 2016], while retaining all of their advantages in terms of accelerated convergence and reduced asymptotic variance. In particular, the reversibility of the dynamics ensures that there is no oscillatory transient behaviour. The improved performance of the proposed methodology, in comparison to the standard overdamped Langevin dynamics and its nonreversible perturbation, is illustrated on an example of sampling from a two-dimensional warped Gaussian target distribution.

What problem does this paper attempt to address?

This paper attempts to address the problem of how to accelerate convergence to equilibrium and reduce asymptotic variance when sampling from probability measures in high-dimensional spaces. Specifically, the authors propose a new method to achieve this goal by introducing an appropriate Stratonovich perturbation into the standard overdamped Langevin dynamics. This method not only accelerates the system's convergence to equilibrium but also reduces the estimated asymptotic variance, thereby improving the computational efficiency of sampling. ### Main Contributions of the Paper: 1. **Accelerated Convergence**: By introducing Stratonovich perturbation, the system converges to equilibrium faster. 2. **Reduced Asymptotic Variance**: Compared to traditional overdamped Langevin dynamics, the new method significantly reduces the estimated asymptotic variance. 3. **Reversibility**: The new method maintains the reversibility of the dynamical system, avoiding the oscillatory transient behavior that may occur in non-reversible methods. ### Method Overview: - **Standard Overdamped Langevin Dynamics**: \[ dX(t) = -\nabla V(X(t)) dt + \sqrt{2} dW_t \] where \( V(x) \) is the potential function, and \( W_t \) is the standard Brownian motion. - **Langevin Dynamics with Stratonovich Perturbation**: \[ dX(t) = -\nabla V(X(t)) dt + g(X(t)) \circ \sqrt{2} d\beta_t + \sqrt{2} dW_t \] where \( g(x) \) is a divergence-free vector field (i.e., \( \text{div}(g\pi) = 0 \)), and \( \beta_t \) is a one-dimensional standard Brownian motion independent of \( W_t \). ### Theoretical Results: - **Reversibility**: The new dynamical system is reversible with respect to the target probability measure. - **Increased Spectral Gap**: The introduction of Stratonovich perturbation increases the spectral gap, accelerating the convergence rate. - **Reduced Asymptotic Variance**: The asymptotic variance of the new method is smaller than that of the standard overdamped Langevin dynamics. ### Numerical Experiments: The authors validated the theoretical results through numerical experiments, demonstrating the superior performance of the new method in terms of accelerated convergence and reduced asymptotic variance under different perturbation intensities. A 2-dimensional twisted Gaussian distribution was used as an example in the experiments, showing that the new method quickly converges in a short time and exhibits smaller asymptotic variance over a long time. ### Conclusion: This paper proposes a new Langevin dynamics method based on Stratonovich perturbation, which efficiently samples probability measures in high-dimensional spaces. The new method not only accelerates the convergence rate but also reduces the estimated asymptotic variance, holding significant theoretical and practical application value.

Accelerated convergence to equilibrium and reduced asymptotic variance for Langevin dynamics using Stratonovich perturbations

Optimizing the diffusion coefficient of overdamped Langevin dynamics

Gradient-adjusted underdamped Langevin dynamics for sampling

Tamed Langevin sampling under weaker conditions

Adaptive stepsize algorithms for Langevin dynamics

A Dynamical System View of Langevin-Based Non-Convex Sampling

Sampling as optimization in the space of measures: The Langevin dynamics as a composite optimization problem

Contraction and Convergence Rates for Discretized Kinetic Langevin Dynamics

Sampling with Adaptive Variance for Multimodal Distributions

Exact Langevin Dynamics with Stochastic Gradients

Sampling and estimation on manifolds using the Langevin diffusion

Reweighted Interacting Langevin Diffusions: an Accelerated Sampling Methodfor Optimization

Accelerating Approximate Thompson Sampling with Underdamped Langevin Monte Carlo

Non-asymptotic entropic bounds for non-linear kinetic Langevin sampler with second-order splitting scheme

Gradual convergence for Langevin dynamics on a degenerate potential

The velocity jump Langevin process and its splitting scheme: long time convergence and numerical accuracy

User-friendly guarantees for the Langevin Monte Carlo with inaccurate gradient

Explicit convergence rates of underdamped Langevin dynamics under weighted and weak Poincaré--Lions inequalities

Improving sampling by modifying the effective diffusion