Annealed Underdamped Langevin Dynamics
Abstract
In sampling, different flavours of annealing, tempering, or other successive approximation approaches are widely used and studied. In this work we investigate annealed underdamped Langevin sampling for logconcave distributions of the form π(x) ∝ e−Ψ(x) for a potential Ψ:ℝd → ℝ. That is, we assume access to the score of a general approximating family of distributions (πτ)τ∈ [0,1] with π0=π. This family is used in a time-inhomogeneous underdamped Langevin process with moving target πτ(t) and an appropriate annealing schedule t↦ τ(t) such that τ(t)↓ 0 as t→ T for some T>0. While the degeneracy of the Brownian motion prohibits conventional Girsanov arguments, we are able to circumvent this issue by relying on an appropriately designed auxiliary process. To the best of our knowledge these are the first (quantitative) results for sampling using time-inhomogeneous underdamped Langevin dynamics. Moreover, we obtain an iteration complexity of 𝒪(ε−2) in total variation distance in comparison to 𝒪(ε−6) for annealed overdamped Langevin sampling. While we currently assume convexity, our proof strategies are amenable to generalization for non-logconcave targets and enable for several future research directions.
Related
- The Fenchel Game of Underdamped Langevin Dynamics: Insights into Accelerated Convergence
- Underdamped Anchored Langevin Dynamics
- Sharp hypocoercive convergence estimates for underdamped Langevin dynamics with specular reflection
- Averaged Annealed Markov Chains Achieve the Algorithmic Threshold for Tensor PCA, Part II: Langevin Dynamics
- Theoretical guidelines for annealed Langevin dynamics in compositional simulation-based inference
BibTeX
@misc{nagler2026annealed,
title = {{Annealed Underdamped Langevin Dynamics}},
author = {Laurenz Nagler and Alexander Falk and Andreas Habring},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.32446},
eprint = {2609.32446},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.32446},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Annealed Underdamped Langevin Dynamics
AU - Laurenz Nagler
AU - Alexander Falk
AU - Andreas Habring
PY - 2026
DA - 2026/09/26
PB - arXiv
DO - 10.48550/arxiv.2609.32446
UR - https://arxiv.org/abs/2609.32446
AB - In sampling, different flavours of annealing, tempering, or other successive approximation approaches are widely used and studied. In this work we investigate annealed underdamped Langevin sampling for logconcave distributions of the form π(x) ∝ e^−Ψ(x) for a potential Ψ:ℝ^d → ℝ. That is, we assume access to the score of a general approximating family of distributions (π_τ)_τ∈ [0,1] with π_0=π. This family is used in a time-inhomogeneous underdamped Langevin process with moving target π_τ(t) and an appropriate annealing schedule t↦ τ(t) such that τ(t)↓ 0 as t→ T for some T>0. While the degeneracy of the Brownian motion prohibits conventional Girsanov arguments, we are able to circumvent this issue by relying on an appropriately designed auxiliary process. To the best of our knowledge these are the first (quantitative) results for sampling using time-inhomogeneous underdamped Langevin dynamics. Moreover, we obtain an iteration complexity of 𝒪(ε^−2) in total variation distance in comparison to 𝒪(ε^−6) for annealed overdamped Langevin sampling. While we currently assume convexity, our proof strategies are amenable to generalization for non-logconcave targets and enable for several future research directions.
ER -