Langevin MCMC Methods

• Langevin MCMC is a family of methods that discretize continuous-time Langevin diffusions to sample from target distributions, forming the basis for algorithms like ULA and MALA.
• These techniques incorporate adaptations such as stochastic gradients, proximal operators, and manifold geometry to tackle high-dimensional, non-smooth inference problems.
• Extensions in the framework include distributed inference, multilevel Monte Carlo, and acceleration methods that enhance mixing rates and correct for discretization bias.

Langevin Markov chain Monte Carlo (MCMC) is a family of sampling methods that build Markov chains by discretizing a continuous-time stochastic differential equation whose invariant distribution is the target posterior or Gibbs measure (Giles et al., 2016, Durmus et al., 2016). In its classical form, the target density is written either as $\pi(x)\propto e^{-U(x)}$ or $\mu(\theta)\propto \exp(-U(\theta))$, and the resulting algorithms include the overdamped and underdamped Langevin schemes, the Unadjusted Langevin Algorithm (ULA), and Metropolis-adjusted variants such as MALA. Subsequent work extends the same core construction to stochastic gradients, adaptive scaling, non-smooth convex targets, manifold geometry, distributed inference, deep latent-variable models, privacy guarantees, and several notions of acceleration (Basak et al., 2012, Yao et al., 24 Mar 2025).

1. Diffusion formulation and classical discretizations

The canonical overdamped Langevin diffusion for $\pi(x)\propto e^{-U(x)}$ is

$dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$

and under mild conditions it has a unique invariant distribution with density proportional to $e^{-U(x)}$ (Durmus et al., 2016). In Bayesian notation, a corresponding posterior-targeting diffusion can be written as

$$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$

whose invariant law is the posterior $\pi(\theta\mid X)$ under suitable conditions (Giles et al., 2016).

Euler–Maruyama discretization yields the Unadjusted Langevin Algorithm. In one common form,

$$X_{k+1} = X_k - \gamma \nabla U(X_k) + \sqrt{2\gamma}\,Z_{k+1}, \qquad Z_{k+1}\sim\mathcal N(0,I_d),$$

while in posterior form one writes

$$\theta_{k+1} = \theta_k + h\Big( \nabla \log \pi(\theta_k) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_k) \Big) + \sqrt{2h}\,\xi_k.$$

ULA is biased: its invariant distribution approximates, but does not equal, the target (Durmus et al., 2016, Giles et al., 2016).

Metropolis-adjusted Langevin algorithms correct that discretization bias by using one Langevin step as a proposal inside Metropolis–Hastings. The resulting chain has the target posterior exactly as invariant distribution and is the standard correction from ULA to MALA (Giles et al., 2016). In state-space filtering, the same principle appears as preconditioned MALA and manifold MALA, with proposal covariance either fixed as $\epsilon^2\Sigma$ or state dependent through a metric $\mu(\theta)\propto \exp(-U(\theta))$0 (Septier et al., 2015).

Underdamped or kinetic Langevin dynamics enlarge the state by a momentum variable. A representative form is

$\mu(\theta)\propto \exp(-U(\theta))$1

which preserves an augmented Gibbs law in $\mu(\theta)\propto \exp(-U(\theta))$2 and often improves exploration relative to overdamped dynamics (Yao et al., 24 Mar 2025). This kinetic formulation is the basis of several later accelerations, including high-order, high-resolution, and regime-switching constructions.

2. Stochastic gradients and large-data Langevin sampling

When the target posterior factorizes over a large dataset, exact gradient evaluation is often the dominant cost. Stochastic Gradient Langevin Dynamics (SGLD) replaces the full gradient by a minibatch estimator: $\mu(\theta)\propto \exp(-U(\theta))$3 where $\mu(\theta)\propto \exp(-U(\theta))$4 is a random minibatch of size $\mu(\theta)\propto \exp(-U(\theta))$5 (Giles et al., 2016). With decreasing step sizes, SGLD can asymptotically sample from the true posterior, but its root mean square error scales as $\mu(\theta)\propto \exp(-U(\theta))$6 in computational cost $\mu(\theta)\propto \exp(-U(\theta))$7, in contrast to the $\mu(\theta)\propto \exp(-U(\theta))$8 rate of standard MCMC methods such as MALA (Giles et al., 2016).

Several Langevin variants target the extra variance introduced by stochastic gradients. “Laplacian Smoothing Stochastic Gradient Markov Chain Monte Carlo” (Wang et al., 2019) introduces the preconditioned SDE

$\mu(\theta)\propto \exp(-U(\theta))$9

and its discretization

$\pi(x)\propto e^{-U(x)}$0

The continuous-time process preserves the same Gibbs measure, and the paper proves that LS-SGLD achieves strictly smaller discretization error in $\pi(x)\propto e^{-U(x)}$1-Wasserstein distance, although its mixing rate can be slightly slower (Wang et al., 2019).

“Langevin Markov Chain Monte Carlo with stochastic gradients” (Matthews et al., 2018) instead develops an underdamped integrator, NOGIN, for noisy gradients. Its central theoretical claim is second-order weak consistency with respect to an underdamped Langevin SDE in which the friction is modified to

$\pi(x)\propto e^{-U(x)}$2

where $\pi(x)\propto e^{-U(x)}$3 is the covariance of the stochastic gradient estimator. The same paper also proves an exactness result in the Gaussian case: when the target is Gaussian and the stochastic gradient noise is Gaussian, the marginal invariant law in $\pi(x)\propto e^{-U(x)}$4 is exact (Matthews et al., 2018).

A different route is to incorporate local curvature. “Stochastic Quasi-Newton Langevin Monte Carlo” (Şimşekli et al., 2016) proposes HAMCMC, which uses L-BFGS approximations of inverse Hessians inside SG-MCMC. The algorithm maintains linear time and memory complexity in the dimension, uses dense approximations of the inverse Hessian, and is shown to be asymptotically unbiased and consistent with posterior expectations (Şimşekli et al., 2016). This places quasi-Newton preconditioning inside the Langevin framework without resorting to full Riemannian metrics.

3. Geometry, non-smoothness, and manifold adaptations

A recurrent limitation of classical ULA and MALA is their reliance on differentiability and globally Lipschitz gradients. For high-dimensional log-concave but non-smooth targets, especially in imaging, proximal Langevin methods replace the raw potential $\pi(x)\propto e^{-U(x)}$5 by a Moreau–Yoshida regularization $\pi(x)\propto e^{-U(x)}$6, where

$\pi(x)\propto e^{-U(x)}$7

This yields MYULA,

$\pi(x)\propto e^{-U(x)}$8

which targets a regularized density $\pi(x)\propto e^{-U(x)}$9 and comes with asymptotic and non-asymptotic total-variation bounds (Durmus et al., 2016). A common misunderstanding is that Langevin MCMC is intrinsically limited to smooth potentials; proximal constructions show that the framework extends to non-smooth log-concave posteriors once the drift is regularized through proximal operators (Durmus et al., 2016).

State-dependent geometry enters through manifold Langevin proposals. In sequential filtering, manifold MALA uses a position-specific metric $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$0 and proposal

$dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$1

with simplified variants omitting the $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$2 correction when its computation is too costly (Septier et al., 2015). In practice, the metric is often chosen from a negative Hessian or Fisher information construction tailored to the likelihood and prior (Septier et al., 2015).

A more explicitly differential-geometric development appears in “Geometrically adapted Langevin dynamics for Markov chain Monte Carlo simulations” (Mamajiwala et al., 2022). That work uses stochastic development on Riemannian manifolds, the metric tensor $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$3, and Christoffel symbols $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$4 to construct a geometrically adapted Langevin dynamics and the corresponding GALA algorithm. The proposal uses covariance $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$5 together with a nontrivial geometric drift term, and the paper reports strong empirical performance on anisotropic and high-dimensional targets, including logistic regression (Mamajiwala et al., 2022).

4. Adaptive limits, multilevel constructions, and acceleration

One line of work derives Langevin dynamics as a limit of adaptive Metropolis procedures. “Langevin type limiting processes for Adaptive MCMC” (Basak et al., 2012) studies an adaptive random-walk Metropolis scheme with proposal scale $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$6, continuous-time embedding with step size $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$7, proposal variance scaled like $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$8, and a sped-up time parameter. Under that regime, the embedded chain converges weakly to a coupled SDE for $dX_t = -\nabla U(X_t)\,dt + \sqrt{2}\,dB_t,$9, whose $e^{-U(x)}$0-component is

$e^{-U(x)}$1

while the non-adaptive limit reduces to the standard one-dimensional Langevin diffusion with invariant density $e^{-U(x)}$2 (Basak et al., 2012). The result gives a rigorous bridge between adaptive random-walk Metropolis and Langevin-type MCMC.

A different objective is estimator complexity rather than single-chain mixing. “Multilevel Monte Carlo for Scalable Bayesian Computations” (Giles et al., 2016) embeds SGLD in a multilevel Monte Carlo hierarchy with step sizes $e^{-U(x)}$3, coupled Brownian paths, and coupled subsampling indices. Its antithetic ML-SGLD construction achieves empirical variance decay of order $e^{-U(x)}$4 and recovers the $e^{-U(x)}$5 RMSE–cost scaling of standard MCMC without Metropolis–Hastings correction (Giles et al., 2016).

Acceleration has also been formalized directly at the level of continuous-time empirical measures. “Accelerating Langevin Monte Carlo Sampling: A Large Deviations Analysis” (Yao et al., 24 Mar 2025) introduces a generalized Langevin SDE framework

$e^{-U(x)}$6

covering overdamped, underdamped, nonreversible, mirror, high-order, and Hessian-free high-resolution dynamics. The paper establishes large deviation principles for empirical measures, derives explicit rate functions, and uses rate-function domination as an acceleration criterion relative to overdamped Langevin dynamics (Yao et al., 24 Mar 2025).

Two further accelerations use randomized or higher-order continuous dynamics. “Regime-Switching Langevin Monte Carlo Algorithms” (Wang et al., 31 Aug 2025) studies regime-switching overdamped and kinetic Langevin dynamics, together with discretizations RS-LMC, RS-KLMC, and FRS-KLMC, which can be interpreted as LMC and KLMC with random step sizes or random friction coefficients and admit $e^{-U(x)}$7-Wasserstein non-asymptotic guarantees under strong log-concavity (Wang et al., 31 Aug 2025). “High-Order Langevin Diffusion Yields an Accelerated MCMC Algorithm” (Mou et al., 2019) introduces a third-order Langevin diffusion and a tailored splitting integrator, proving $e^{-U(x)}$8 mixing for a broad class of $e^{-U(x)}$9-dimensional ridge-separable targets and $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$0 under $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$1-th order smoothness (Mou et al., 2019). These papers use different performance criteria—Wasserstein mixing, RMSE–cost scaling, and large-deviation rate functions—so their acceleration statements are complementary rather than interchangeable.

5. Distributed, sequential, and amortized Langevin formulations

Langevin proposals also serve as modular components inside larger inference architectures. “DG-LMC: A Turn-key and Scalable Synchronous Distributed MCMC Algorithm via Langevin Monte Carlo within Gibbs” (Plassier et al., 2021) starts from an asymptotically exact data augmentation factorization

$$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$2

which makes the conditional $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$3 independent across workers. Each worker then runs local LMC updates for its $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$4, while the master samples $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$5 from a Gaussian conditional. The resulting synchronous distributed chain has explicit high-dimensional $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$6 mixing-time guarantees under strong log-concavity (Plassier et al., 2021).

In nonlinear state-space models, Langevin updates appear inside Sequential MCMC. “Langevin and Hamiltonian based Sequential MCMC for Efficient Bayesian Filtering in High-dimensional Spaces” (Septier et al., 2015) samples the current latent state $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$7 conditional on a selected past trajectory using MALA, preconditioned MALA, manifold MALA, or simplified manifold MALA. The target at time $$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$8 is the conditional

$$d\theta_t = \Big( \nabla \log \pi(\theta_t) + \sum_{i=1}^N \nabla \log \pi(x_i \mid \theta_t) \Big)\,dt + \sqrt{2}\,dW_t,$$9

and the paper reports substantial improvements over particle-filter baselines in high-dimensional filtering examples (Septier et al., 2015).

A different extension replaces datapoint-wise latent Langevin chains by a shared encoder. “Langevin Autoencoders for Learning Deep Latent Variable Models” (Taniguchi et al., 2022) defines an encoder $\pi(\theta\mid X)$0 and a potential over the last-layer parameters,

$\pi(\theta\mid X)$1

Langevin dynamics are then run in $\pi(\theta\mid X)$2-space rather than separately for each latent $\pi(\theta\mid X)$3. Under a rank condition on the feature matrix $\pi(\theta\mid X)$4, the induced stationary law over the encoded latents equals the product of the true posteriors for the datapoints in the batch, so the amortized Langevin dynamics remain a valid MCMC algorithm (Taniguchi et al., 2022). This suggests that Langevin MCMC can operate not only on latent variables or model parameters, but also on shared inference-network parameters when the induced Markov structure is controlled.

6. Invariance, stability, and contemporary guarantees

The Langevin framework is anchored by invariant-measure calculations and SDE well-posedness conditions. For fixed diffusion coefficient, the stationary density $\pi(\theta\mid X)$5 solves the stationary Fokker–Planck equation associated with the one-dimensional Langevin SDE, and standard linear-growth and local-Lipschitz conditions guarantee existence, uniqueness, and non-explosion of the resulting SDEs (Basak et al., 2012). For MYULA, convexity and Lipschitz continuity of $\pi(\theta\mid X)$6 yield geometric ergodicity of the Markov kernel and explicit total-variation error bounds that separate mixing error, discretization bias, and Moreau–Yoshida approximation bias (Durmus et al., 2016).

A contemporary extension concerns privacy. “Differential privacy guarantees of Markov chain Monte Carlo algorithms” (Bertazzi et al., 24 Feb 2025) establishes differential privacy and Rényi differential privacy guarantees for MCMC outputs under assumptions on convergence of the underlying Markov chain. For ULA and SGLD specifically, the paper uses Girsanov’s theorem and a perturbation trick to prove uniform in $\pi(\theta\mid X)$7 privacy guarantees when the state of the chain after $\pi(\theta\mid X)$8 iterations is released, and also derives bounds for the privacy of the entire chain trajectory (Bertazzi et al., 24 Feb 2025). A central structural conclusion is that if an asymptotically exact MCMC algorithm is uniformly private in time, then the target posterior itself must be differentially private (Bertazzi et al., 24 Feb 2025).

Across these strands, Langevin MCMC functions less as a single algorithm than as a design principle: specify a diffusion with the desired invariant law, choose a discretization with acceptable bias and stability, and adapt the geometry, noise model, communication pattern, or auxiliary variables to the structure of the target and the computational environment. The resulting family now spans overdamped and kinetic samplers, stochastic-gradient and multilevel schemes, proximal and manifold methods, adaptive diffusion limits, distributed Gibbs constructions, amortized latent-variable inference, and privacy-aware sampling (Giles et al., 2016, Durmus et al., 2016, Yao et al., 24 Mar 2025).