Annealed Langevin Monte Carlo (ALMC)

• Annealed Langevin Monte Carlo (ALMC) is a sampling method that interpolates between a tractable base and a complex target distribution via annealed steps.
• It leverages Langevin updates combined with optimal transport and diffusion processes to ensure both global exploration and precise local refinement.
• Practical implementations depend on optimal scheduling, step size control, and accurate score approximations, making ALMC essential in modern generative modeling and Bayesian inference.

Annealed Langevin Monte Carlo (ALMC) refers to a class of Markov chain Monte Carlo methodologies designed to efficiently sample from high-dimensional, multimodal, or non-log-concave distributions by employing a sequence of intermediate "annealed" distributions that interpolate between a tractable base and the target law. The protocol integrates the Langevin Monte Carlo (LMC) update at each intermediate stage and builds upon theoretical tools from optimal transport, diffusion processes, and statistical physics to ensure both global exploration and local refinement. ALMC has become foundational in modern statistical sampling, score-based generative modeling, and Bayesian posterior inference.

1. Mathematical Formulation and Motivation

The target is to sample from a probability density $\pi$ on $\mathbb{R}^d$ specified only up to normalization: $\pi(x) \propto \exp(-V(x)),$ where $V:\mathbb{R}^d \to \mathbb{R}$ is a $\beta$-smooth potential. Classical LMC converges quickly when $V$ is strongly convex (log-concave), but for non-log-concave and multimodal targets, LMC suffers from poor mixing due to local trapping and exponential time-scale mode exploration.

Annealing constructs a family of intermediate distributions $\pi_{\beta}(x) \propto \exp(-\beta V(x))$ for $\beta \in [0,1]$, evolving from a "hot" (flatter, nearly uniform) density at small $\beta$ to the original target at $\beta=1$. More generally, one can define a curve of measures $\mathbb{R}^d$0 that interpolates between an easy-to-sample $\mathbb{R}^d$1 and the target $\mathbb{R}^d$2 using, for example, diffusion paths or convex combinations in log-density space (Guo et al., 2024, Cordero-Encinar et al., 13 Feb 2025).

The annealing mechanism "bridges" separated modes at high temperature, enabling transitions inaccessible to direct LMC, before gradually cooling to refine sampling in the correct local geometry.

2. Core Algorithmic Structure

The canonical ALMC proceeds by discretizing the annealing path and performing Langevin updates at each step. For a fixed schedule $\mathbb{R}^d$3, the update is

$\mathbb{R}^d$4

with step sizes $\mathbb{R}^d$5 determined by a total time parameter $\mathbb{R}^d$6 (Guo et al., 2024). The initial sample $\mathbb{R}^d$7 is typically drawn from $\mathbb{R}^d$8, often chosen as a standard Gaussian.

Extensions to score-based generative models and Bayesian settings replace the explicit potential with neural score networks $\mathbb{R}^d$9 approximating $\pi(x) \propto \exp(-V(x)),$0, allowing ALMC to function as the core sampler for these architectures (Parulekar et al., 11 Aug 2025, Cordero-Encinar et al., 13 Feb 2025). Posterior inference employs a "tilt" via the measurement model gradient.

A schematic pseudocode for the general annealed scheme:

$\beta$2

When applying to flows (e.g., ODE sampling), ALMC is used to approximate importance weights or velocity fields that govern deterministic transports from base to target (Huang, 21 Apr 2026).

3. Theoretical Guarantees and Oracle Complexity

The analysis of ALMC relies on path-wise control of Kullback-Leibler divergence via Girsanov's theorem and the action integral of the interpolation path. Under $\pi(x) \propto \exp(-V(x)),$1-smoothness of $\pi(x) \propto \exp(-V(x)),$2 and the existence of a path $\pi(x) \propto \exp(-V(x)),$3 with finite $\pi(x) \propto \exp(-V(x)),$4-Wasserstein action

$\pi(x) \propto \exp(-V(x)),$5

the following non-asymptotic guarantee holds (Guo et al., 2024): $\pi(x) \propto \exp(-V(x)),$6 oracle calls are sufficient, where $\pi(x) \propto \exp(-V(x)),$7 is the total number of gradient evaluations and $\pi(x) \propto \exp(-V(x)),$8 encapsulates the path difficulty. This result holds without requiring log-Sobolev or isoperimetric properties of $\pi(x) \propto \exp(-V(x)),$9 and applies to highly multimodal, non-log-concave cases.

For diffusion-based annealing paths (Gaussian or Student's $V:\mathbb{R}^d \to \mathbb{R}$0 convolution) as widely used in score-based models, similar polynomial-time error bounds for convergence in KL and Wasserstein-2 distance are established, depending on the second moment, Lipschitz smoothness, and the log-Sobolev constants of the path marginals (Cordero-Encinar et al., 13 Feb 2025).

4. Practical Implementations and Scheduling

Effective adoption of ALMC requires careful design of the annealing schedule, discretization step sizes, and, in score-based models, the accuracy of score network approximation.

• Annnealing schedule: Common choices include geometric spacing (for noise levels $V:\mathbb{R}^d \to \mathbb{R}$1) or smooth functions such as $V:\mathbb{R}^d \to \mathbb{R}$2 for time index $V:\mathbb{R}^d \to \mathbb{R}$3 (Cordero-Encinar et al., 13 Feb 2025, Parulekar et al., 11 Aug 2025).
• Step size: Typically $V:\mathbb{R}^d \to \mathbb{R}$4 per stage is required to control discretization error (Guo et al., 2024).
• Score network accuracy: Mean-squared score error must be below an explicit $V:\mathbb{R}^d \to \mathbb{R}$5 threshold along the entire path (Cordero-Encinar et al., 13 Feb 2025).

For measure-tilted targets (e.g., posteriors), the update incorporates both prior and measurement score terms, with step sizes scaled to noise or model uncertainty (Parulekar et al., 11 Aug 2025).

When ALMC is used as a component of flow-ODE sampling or importance weighting, the output particles are reweighted via schemes such as the Jarzynski identity to compensate for non-equilibrium bias and discretization (Huang, 21 Apr 2026).

5. Connections to Generative Models and Posterior Sampling

ALMC is foundational in score-based generative modeling, where the diffusion path interpolates between a tractable base (Gaussian or Student's $V:\mathbb{R}^d \to \mathbb{R}$6) and the empirical data distribution. The score function along this path is learned by neural networks via denoising-score-matching (Cordero-Encinar et al., 13 Feb 2025). Convergence results indicate that sample quality, mode coverage, and robustness depend on the interplay between the score approximation, annealing schedule, and step size.

In posterior sampling tasks (e.g., image super-resolution or MRI reconstruction), ALMC provides theoretical and empirical performance guarantees over direct single-level samplers. Empirically, ALMC yields improved PSNR, LPIPS, and NMSE metrics, with better sample diversity and measurement consistency, and is robust to conservative hyperparameter settings (Parulekar et al., 11 Aug 2025).

6. Empirical Performance and Benchmarking

ALMC and its adaptations (e.g., ALMC-ODE, Diffusion-ALMC) consistently outperform vanilla LMC and traditional Hamiltonian Monte Carlo in exploring highly multimodal, non-log-concave distributions—including high-dimensional Gaussian mixtures and non-convex physical models. Measured by energy distance, MMD, sliced Wasserstein, and kernelized Stein discrepancy, ALMC-ODE demonstrates orders-of-magnitude improved mixing, faithful mode visitation, and lower estimator variance (Huang, 21 Apr 2026). In generative modeling, ALMC robustly interpolates between the base and data distributions, yielding high-fidelity samples under both Gaussian and heavy-tailed smoothing approaches (Cordero-Encinar et al., 13 Feb 2025).

7. Summary of Key Formulas and Algorithmic Variants

Formula/Notation	Description	Reference
$V:\mathbb{R}^d \to \mathbb{R}$7	Target density	(Guo et al., 2024)
$V:\mathbb{R}^d \to \mathbb{R}$8	ALMC update	(Guo et al., 2024)
$V:\mathbb{R}^d \to \mathbb{R}$9	Curve action	(Guo et al., 2024)
$\beta$0	Oracle-bound	(Guo et al., 2024)
$\beta$1	DALMC step	(Cordero-Encinar et al., 13 Feb 2025)

In summary, Annealed Langevin Monte Carlo encompasses a widely adopted and rigorously guaranteed methodology for sampling and generative inference in complicated distributions, achieving polynomial-time mixing even for non-log-concave, multimodal, or heavy-tailed targets. The theoretical analyses delineate explicit error-income tradeoffs, while empirical evidence confirms its superiority over standard MCMC in challenging regimes (Guo et al., 2024, Cordero-Encinar et al., 13 Feb 2025, Parulekar et al., 11 Aug 2025, Huang, 21 Apr 2026).