RS-KLMC: Regime-Switching Langevin Sampling

• The paper introduces RS-KLMC, a method that discretizes a regime‐switching kinetic Langevin diffusion wherein a CTMC randomizes the effective timestep for sampling.
• RS-KLMC preserves the core KLMC framework while providing non-asymptotic 2-Wasserstein error bounds and improved dependence on dimension in strongly log-concave settings.
• The algorithm leverages CTMC-driven modulation of both drift and diffusion to enhance stability and efficiency compared to fixed-parameter underdamped samplers.

Searching arXiv for the core RS-KLMC paper and closely related KLMC/regime-switching references. Regime-Switching Kinetic Langevin Monte Carlo (RS-KLMC) is a Markov chain Monte Carlo method obtained by discretizing a regime-switching kinetic Langevin diffusion in which the effective time scale of the underdamped dynamics is modulated by a finite-state continuous-time Markov chain (CTMC). In the formulation proposed in "Regime-Switching Langevin Monte Carlo Algorithms" (Wang et al., 31 Aug 2025), the method targets densities of the form $\pi(x)\propto \exp(-f(x))$ and can be interpreted as a Kinetic Langevin Monte Carlo (KLMC) algorithm with randomized stepsizes governed by regime dynamics. The construction sits at the intersection of two lines of work: non-asymptotic KLMC theory for strongly log-concave targets (Dalalyan et al., 2018), and stochastic switching Langevin systems in which CTMC-driven modulation changes the effective dynamics in a singularly perturbed manner (Walker et al., 2021).

1. Definition and formal setup

In the regime-switching kinetic Langevin dynamics (RS-KLD) of (Wang et al., 31 Aug 2025), the target density is $\pi(x)\propto \exp(-f(x))$, and the continuous-time process is

$$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$

where $(B_t)_{t\ge 0}$ is standard Brownian motion and $\beta(t)$ is a positive CTMC taking values in $\{\bar\beta_1,\ldots,\bar\beta_N\}$ with generator matrix $\mathbf Q$ (Wang et al., 31 Aug 2025). The regime variable therefore modulates both drift and diffusion coefficients through a common multiplicative factor.

The invariant distribution of the joint process $(\beta(t),V(t),X(t))$ is stated to be

$$\psi(\beta)\otimes \mathcal N(0,I_d)\otimes \pi(x),$$

where $\psi$ is the stationary distribution of the CTMC (Wang et al., 31 Aug 2025). This product structure is important because it separates the auxiliary regime and momentum variables from the target positional marginal.

A useful conceptual point is that RS-KLMC is not merely a heuristic randomization of a discretization parameter. In (Wang et al., 31 Aug 2025), the randomization is induced by an underlying regime-switching diffusion with exponential ergodicity. This distinguishes the method from ad hoc stochastic stepsize rules. A broader stochastic-switching perspective is also consistent with the analysis of switching Langevin systems in (Walker et al., 2021), where force switching is modeled by a CTMC and coupled to Brownian noise through an explicit small-noise scaling.

2. Discretization and algorithmic structure

The RS-KLMC algorithm of (Wang et al., 31 Aug 2025) discretizes RS-KLD with a base step size $\pi(x)\propto \exp(-f(x))$0. Given $\pi(x)\propto \exp(-f(x))$1, the method first updates the regime and then performs a KLMC-type position-velocity update using the realized regime value.

The regime update uses first-order transition probabilities

$\pi(x)\propto \exp(-f(x))$2

so that $\pi(x)\propto \exp(-f(x))$3 is sampled from $\pi(x)\propto \exp(-f(x))$4 according to the CTMC over one discretization interval (Wang et al., 31 Aug 2025).

Conditioned on the current regime, the updates are

$\pi(x)\propto \exp(-f(x))$5

where $\pi(x)\propto \exp(-f(x))$6 and $\pi(x)\propto \exp(-f(x))$7 are appropriately constructed correlated Gaussian increments, and

$\pi(x)\propto \exp(-f(x))$8

(Wang et al., 31 Aug 2025).

This update is structurally aligned with the KLMC discretization studied in (Dalalyan et al., 2018), where the gradient is frozen at the beginning of each step and the resulting Ornstein-Uhlenbeck subdynamics is solved exactly. In that sense, RS-KLMC preserves the standard KLMC mechanism but replaces the deterministic step length by the random effective step $\pi(x)\propto \exp(-f(x))$9 (Wang et al., 31 Aug 2025).

A concise comparison of the algorithmic ingredients is useful.

Component	KLMC	RS-KLMC
Base dynamics	Kinetic Langevin diffusion	Regime-switching kinetic Langevin diffusion
Per-step scale	Fixed	Randomized by $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$0
Random mechanism	Gaussian noise	Gaussian noise plus CTMC regime update

This suggests that RS-KLMC can be viewed as a structured random-time-change version of KLMC, rather than a different momentum sampler in the Hamiltonian sense.

3. Relation to classical KLMC theory

The closest Euclidean precursor is the KLMC theory of "On sampling from a log-concave density using kinetic Langevin diffusions" (Dalalyan et al., 2018). There, for smooth and strongly log-concave targets on $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$1, the kinetic Langevin diffusion has a geometric mixing property in $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$2, and its discretization admits non-asymptotic Wasserstein error bounds (Dalalyan et al., 2018). In particular, for $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$3, the continuous semigroup contracts exponentially, and the discretized KLMC achieves a square-root improvement over overdamped LMC in dimension/precision scaling for moderate condition number (Dalalyan et al., 2018).

RS-KLMC inherits this underdamped sampling template but alters the effective coefficients through CTMC switching (Wang et al., 31 Aug 2025). The paper explicitly frames RS-KLMC as a KLMC algorithm with random stepsizes and provides non-asymptotic $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$4-Wasserstein guarantees analogous in spirit to the fixed-parameter theory of (Dalalyan et al., 2018). The dependence on dimension remains of root-$$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$5 type, which matches the underdamped advantage emphasized in standard KLMC analysis (Wang et al., 31 Aug 2025, Dalalyan et al., 2018).

The relationship is especially clear at the level of complexity statements reported in (Wang et al., 31 Aug 2025). RS-KLMC has iteration complexity

$$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$6

to reach $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$7-accuracy in $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$8, with step size $$\begin{aligned} dV(t) &= -\gamma \beta(t) V(t)\,dt - \beta(t)\nabla f(X(t))\,dt + \sqrt{2\gamma\beta(t)}\,dB_t, \ dX(t) &= \beta(t) V(t)\,dt , \end{aligned}$$9 (Wang et al., 31 Aug 2025). In the comparative table reported there, KLMC and RS-KLMC both have $(B_t)_{t\ge 0}$0 dependence, while overdamped LMC and RS-LMC have $(B_t)_{t\ge 0}$1 dependence (Wang et al., 31 Aug 2025). The data therefore present RS-KLMC as preserving the principal complexity order of KLMC while changing constants through the switching mechanism.

A further extension of the KLMC family appears in KLMC2, a second-order discretization for Hessian-Lipschitz targets (Dalalyan et al., 2018). That paper notes that quantitative dependence on curvature and higher-order smoothness can support adaptive or regime-switching designs. This suggests a natural design interpretation for RS-KLMC: switching can be used to encode geometry-dependent sampling phases, although that particular adaptive mechanism is not formalized in the supplied RS-KLMC definition (Dalalyan et al., 2018).

4. Non-asymptotic convergence guarantees

The central theoretical statement for RS-KLMC in (Wang et al., 31 Aug 2025) is a $(B_t)_{t\ge 0}$2-Wasserstein convergence bound under strong convexity and smoothness assumptions. Let $(B_t)_{t\ge 0}$3 and $(B_t)_{t\ge 0}$4 denote the maximum and minimum regime values. If the step size $(B_t)_{t\ge 0}$5 is sufficiently small, and the process is initialized with $(B_t)_{t\ge 0}$6 and $(B_t)_{t\ge 0}$7, then

$(B_t)_{t\ge 0}$8

where

$(B_t)_{t\ge 0}$9

and $\beta(t)$0 is a diagonal matrix of regime values (Wang et al., 31 Aug 2025).

Several features of this theorem are explicit in the paper. First, the decay rate $\beta(t)$1 depends jointly on the CTMC generator and the regime magnitudes (Wang et al., 31 Aug 2025). Second, the dimension dependence is only $\beta(t)$2, which the paper contrasts with the linear-$\beta(t)$3 dependence of overdamped LMC (Wang et al., 31 Aug 2025). Third, the condition-number dependence is linear in $\beta(t)$4 (Wang et al., 31 Aug 2025). These properties place RS-KLMC squarely within the quantitative underdamped-sampling paradigm developed for KLMC in (Dalalyan et al., 2018), but with a spectral term altered by the switching generator.

The role of the CTMC spectrum is one of the distinctive features of RS-KLMC. Since $\beta(t)$5 is determined by the real parts of eigenvalues of $\beta(t)$6, the switching schedule cannot be summarized only by an average regime value. The generator itself enters the convergence rate (Wang et al., 31 Aug 2025). This is compatible with the broader switching-systems lesson from (Walker et al., 2021), where proper asymptotic treatment of switching and noise is necessary, and naive averaging can mischaracterize effective barriers and transition behavior.

A related variant in (Wang et al., 31 Aug 2025) is frictional-regime-switching kinetic Langevin Monte Carlo (FRS-KLMC), in which the friction rather than the step scale switches according to a CTMC. The reported rate parameter becomes

$\beta(t)$7

and the paper states iteration complexity $\beta(t)$8 (Wang et al., 31 Aug 2025). This is a separate algorithm, but it clarifies that the regime-switching framework encompasses multiple loci of randomization within underdamped samplers.

5. Regime switching, stochastic switching, and effective dynamics

RS-KLMC belongs to a broader class of switching stochastic dynamics in which a finite-state Markov process modulates continuous evolution. The paper "Numerical computation of effective thermal equilibrium in Stochastically Switching Langevin Systems" (Walker et al., 2021) analyzes overdamped Langevin systems with switching force terms of the form

$\beta(t)$9

where the regime process $\{\bar\beta_1,\ldots,\bar\beta_N\}$0 evolves as a CTMC with rate matrix $\{\bar\beta_1,\ldots,\bar\beta_N\}$1 (Walker et al., 2021). In that setting, the interaction of switching force and Brownian motion can create an "effective thermal equilibrium" even though the system does not obey a potential function (Walker et al., 2021).

The paper derives a quasipotential $\{\bar\beta_1,\ldots,\bar\beta_N\}$2 using a WKB ansatz

$\{\bar\beta_1,\ldots,\bar\beta_N\}$3

leading to a Hamilton-Jacobi problem

$\{\bar\beta_1,\ldots,\bar\beta_N\}$4

with $\{\bar\beta_1,\ldots,\bar\beta_N\}$5 defined through the principal eigenvalue of a matrix combining advection, diffusion, and switching terms (Walker et al., 2021). Although this is not an RS-KLMC paper and the dynamics are overdamped rather than kinetic, it provides a mathematically relevant comparison point: switching modifies effective equilibrium structure in ways that are not captured by a mean-field replacement of the switching dynamics.

That broader lesson bears directly on how RS-KLMC should be interpreted. In (Walker et al., 2021), simple force-averaging is reported to overestimate the effective barrier, whereas the proper distinguished-limit analysis predicts the correct weaker emergent barrier. This suggests that CTMC randomization in RS-KLMC should not be treated as a cosmetic perturbation of a fixed-step KLMC. A plausible implication is that the regime process can alter exploration and metastable crossing behavior through dynamical effects that are genuinely spectral and pathwise, not merely average-parametric.

6. Numerical behavior, variants, and related extensions

The RS-KLMC paper reports numerical experiments on Bayesian linear regression and logistic regression, evaluating mean squared error or classification accuracy for RS-LMC, RS-KLMC, and FRS-KLMC against standard LMC and KLMC on synthetic and real datasets (Wang et al., 31 Aug 2025). The reported findings are that RS-LMC/KLMC accelerates convergence when the regime space is sufficiently wide or the CTMC mixes rapidly, that FRS-KLMC delivers the best $\{\bar\beta_1,\ldots,\bar\beta_N\}$6-iteration complexity, and that switching is especially beneficial for stability in small-sample regimes (Wang et al., 31 Aug 2025). The paper further states that properly chosen regime parameters and mixing matrices result in faster and/or stabler convergence of error or prediction-accuracy metrics (Wang et al., 31 Aug 2025).

Several adjacent works broaden the methodological context. The analysis of the stochastic exponential Euler discretization of KLMC in (Kim et al., 4 Oct 2025) revisits synchronous Wasserstein coupling and shows that the exponential integrator remains stable in the overdamped regime provided proper time acceleration is applied. That work reports contraction and asymptotic bias bounds from underdamped to overdamped regimes, including a phase transition at $\{\bar\beta_1,\ldots,\bar\beta_N\}$7 (Kim et al., 4 Oct 2025). This is not an RS-KLMC result, but it is directly relevant to regime-switching designs in which parameters may move across underdamped and overdamped regimes.

Another neighboring direction is the switching Hamiltonian framework of (Sharma, 11 Jun 2026), which introduces switching Hamiltonian Monte Carlo for finite mixture Boltzmann-Gibbs distributions and develops symmetric numerical integrators interlaced with Poisson jumps and CTMC switching. The paper proves geometric ergodicity and second-order bias for the proposed numerical integrators, and explicitly notes that the discrete Poisson equation approach can be generalized to other settings, for example, kinetic Langevin equations (Sharma, 11 Jun 2026). This suggests a possible analytical route for future bias analysis of higher-order RS-KLMC discretizations.

More distant, but structurally relevant, are two extensions of kinetic Langevin methodology beyond standard Euclidean fixed-parameter sampling. "Convergence of Kinetic Langevin Monte Carlo on Lie groups" (Kong et al., 2024) establishes exponential $\{\bar\beta_1,\ldots,\bar\beta_N\}$8 convergence for KLMC on compact Lie groups using a structure-preserving discretization, and "Langevin dynamics with general kinetic energies" (Stoltz et al., 2016) studies non-quadratic kinetic energies with hypocoercive convergence and Metropolized splitting schemes. These works do not define RS-KLMC, but they show that the kinetic Langevin paradigm admits substantial generalization in geometry, kinetic structure, and discretization.

7. Interpretation, scope, and common points of confusion

A common misconception is to equate RS-KLMC with a simple heuristic of drawing an i.i.d. random stepsize at each iteration. The formulation in (Wang et al., 31 Aug 2025) is more specific: the stepsize randomization is induced by a finite-state CTMC with generator $\{\bar\beta_1,\ldots,\bar\beta_N\}$9, and the convergence rate depends on the spectrum of a matrix involving both $\mathbf Q$0 and the regime values. Thus, the switching law is part of the model, not an incidental implementation detail.

Another possible confusion concerns the role of friction. In RS-KLMC, the regime variable $\mathbf Q$1 scales the drift, diffusion, and position equation in the underdamped dynamics (Wang et al., 31 Aug 2025). By contrast, FRS-KLMC randomizes the friction coefficient itself and obtains a different spectral rate and different $\mathbf Q$2-dependence in iteration complexity (Wang et al., 31 Aug 2025). The two methods belong to the same regime-switching family but are analytically distinct.

It is also important not to conflate regime switching with naive parameter averaging. The switching-systems analysis of (Walker et al., 2021) shows that averaging can fail to capture effective equilibrium quantities such as barriers. While that paper studies overdamped force switching rather than RS-KLMC, it reinforces the broader principle that stochastic switching can produce effective behavior not reducible to deterministic averaging.

Finally, RS-KLMC as currently formulated is situated in the strongly convex, smooth setting of non-asymptotic Wasserstein theory (Wang et al., 31 Aug 2025). Broader generalizations are plausible but remain distinct research topics. The surrounding literature indicates several directions: higher-order KLMC discretization under Hessian-Lipschitz assumptions (Dalalyan et al., 2018), stable simulation across underdamped and overdamped regimes (Kim et al., 4 Oct 2025), structure-preserving kinetic sampling on Lie groups (Kong et al., 2024), and generalized kinetic energies with Metropolized splitting (Stoltz et al., 2016). Together, these results place RS-KLMC within a larger program of designing underdamped samplers whose parameters, geometry, or auxiliary dynamics vary in a controlled and analyzable manner.