Abstract
Langevin dynamics are widely used in sampling high-dimensional, non-Gaussian distributions whose densities are known up to a normalizing constant. In particular, there is strong interest in unadjusted Langevin algorithms (ULA), which directly discretize Langevin dynamics to estimate expectations over the target distribution. We study the use of transport maps that approximately normalize a target distribution as a way to precondition and accelerate the convergence of Langevin dynamics. We show that in continuous time, when a transport map is applied to Langevin dynamics, the result is a Riemannian manifold Langevin dynamics (RMLD) with metric defined by the transport map. We also show that applying a transport map to an irreversibly-perturbed ULA results in a geometry-informed irreversible perturbation (GiIrr) of the original dynamics. These connections suggest more systematic ways of learning metrics and perturbations, and also yield alternative discretizations of the RMLD described by the map, which we study. Under appropriate conditions, these discretized processes can be endowed with non-asymptotic bounds describing convergence to the target distribution in 2-Wasserstein distance. Illustrative numerical results complement our theoretical claims.
Introduction
Sampling from high-dimensional, non-Gaussian distributions is a cornerstone of modern statistics and machine learning, particularly in Bayesian inference. The Unadjusted Langevin Algorithm (ULA) is a popular tool for this; however, it faces two significant challenges: slow convergence in distributions with complex geometries (like long, thin funnels) and discretization bias, where the discrete steps required for computation prevent the algorithm from perfectly sampling the target distribution.
To address these, researchers often employ preconditioning via Riemannian Manifold Langevin Dynamics (RMLD). However, selecting an appropriate metric for RMLD is notoriously difficult and often relies on hand-tuned heuristics. This paper introduces the Transport Map Unadjusted Langevin Algorithm (TMULA), which leverages transport maps to systematically precondition complex distributions for faster and more accurate sampling.
The TMULA Framework
The core idea of TMULA is to use Transport Maps—deterministic functions $S$ that transform a complex "target" distribution $\pi$ into a simple "reference" distribution $\eta_{ref}$ (typically a standard Gaussian).
Instead of sampling the difficult target distribution directly, TMULA learns an approximate transport map $S$, transforms the target into a simpler space, runs a standard ULA on that simplified distribution, and then maps the resulting samples back to the original space. This "straightens" the geometry of the problem, allowing for more efficient exploration.
Continuous vs. Discrete Equivalence
A major theoretical contribution of this work is the bridge between transport maps and manifold dynamics. The authors rigorously prove that in continuous time, applying a transport map to Langevin dynamics is mathematically equivalent to RMLD, where the transport map's Jacobian defines the Riemannian metric.
However, a critical "discretization gap" exists. While the processes are equivalent in continuous time, their discrete-time implementations—TMULA and Euler-Maruyama RMLD (EMRMLD)—are not. TMULA discretizes the process in the simplified reference space, which handles the distribution's curvature more gracefully than discretizing the manifold equations directly. This leads to significantly lower asymptotic bias; for instance, in tests on a 2D "Banana" distribution, TMULA's bias constant was nearly 60 times smaller than that of standard RMLD discretization.
Key Results and Insights
The authors evaluated TMULA on several challenging benchmarks, including the Funnel distribution and a 7D Hybrid Rosenbrock distribution. Key findings include:
- Improved Stability: For distributions with sharp features, an implicit version (TMUILA) remains stable where standard ULA would fail.
- Better Tail Exploration: In the Funnel distribution, TMULA accurately captured the narrow "neck" that standard samplers often miss.
- Learning Metrics: The framework allows for learning Riemannian metrics by training transport maps (e.g., via normalizing flows) on initial samples, rather than guessing them.
- Irreversible Perturbations: The method extends to non-reversible dynamics, creating "Geometry-Informed Irreversible Perturbations" (GiIrr) that further accelerate convergence.
Conclusion
TMULA provides a robust, theoretically grounded method for sampling complex distributions by combining the geometric intuition of RMLD with the practical stability of transport maps. While the quality of the sampler is bounded by the expressivity of the learned map, the approach offers a systematic path toward more efficient high-dimensional inference.
Summary updated