FRS-KLMC: Frictional Regime Switching KLMC

• The paper introduces FRS-KLMC, a sampler that randomizes only the friction coefficient to balance momentum damping against stochastic forcing in kinetic Langevin sampling.
• FRS-KLMC leverages a finite-state Markov process for friction regime switching, providing explicit non-asymptotic convergence guarantees under strong convexity and smoothness assumptions.
• Empirical studies demonstrate that carefully designed friction regimes can accelerate convergence relative to fixed-friction and variable-step methods while preserving the target distribution.

Frictional-Regime-Switching Kinetic Langevin Monte Carlo (FRS-KLMC) is a regime-switching variant of kinetic Langevin Monte Carlo in which the friction coefficient evolves according to a finite-state Markov process; equivalently, it can be viewed as the KLMC algorithm with random frictional coefficients (Wang et al., 31 Aug 2025). It was introduced together with frictional-regime-switching kinetic Langevin dynamics (FRS-KLD) for sampling from targets of the form $\pi(x)\propto e^{-f(x)}$ under strong convexity and smoothness assumptions, with non-asymptotic guarantees stated in the $2$-Wasserstein distance (Wang et al., 31 Aug 2025). Within the Langevin literature, it occupies an intermediate position between fixed-friction KLMC (Dalalyan et al., 2018) and more general variable-friction or geometry-adaptive kinetic samplers (Lim et al., 2023).

1. Definition and conceptual placement

The basic sampling problem is to draw from a density on $\mathbb R^d$,

$\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$

where $f$ is assumed to be twice continuously differentiable, $m$-strongly convex, and $M$-smooth: $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$ The regime process is a finite-state irreducible continuous-time Markov chain (CTMC), so it admits a unique stationary distribution $\psi$ (Wang et al., 31 Aug 2025).

FRS-KLMC is best understood by contrast with two neighboring constructions. Standard KLMC uses a fixed friction parameter $\gamma$ and a fixed step size $2$0. Regime-switching KLMC (RS-KLMC) randomizes an effective time scale through a multiplicative regime variable $2$1, and is therefore interpreted as KLMC with random stepsizes. FRS-KLMC instead keeps the physical step size fixed and randomizes only the damping/noise coefficient $2$2, thereby altering the balance between momentum persistence, dissipation, and stochastic forcing (Wang et al., 31 Aug 2025).

Method	What varies	Interpretation
KLMC	fixed $2$3, fixed step $2$4	constant-friction kinetic Langevin discretization
RS-KLMC	regime variable $2$5	KLMC with random stepsizes
FRS-KLMC	regime variable $2$6	KLMC with random frictional coefficients

This distinction is structural rather than cosmetic. In RS-KLMC the regime variable rescales the full kinetic system, while in FRS-KLMC only the friction term and its matched noise amplitude switch. Consequently, FRS-KLMC is not merely a random-step discretization; it is a sampler with regime-dependent momentum damping.

2. Continuous-time model: FRS-KLD

The continuous-time dynamics underlying the algorithm are the frictional-regime-switching kinetic Langevin dynamics

$2$7

where $2$8 is standard $2$9-dimensional Brownian motion and $\mathbb R^d$0 is an independent positive CTMC taking values in $\mathbb R^d$1 (Wang et al., 31 Aug 2025). In this formulation, only the damping term $\mathbb R^d$2 and the noise amplitude $\mathbb R^d$3 switch; the force $\mathbb R^d$4 and transport $\mathbb R^d$5 do not.

The full generator on $\mathbb R^d$6 is the sum of a kinetic Langevin part and the CTMC generator: $\mathbb R^d$7 If $\mathbb R^d$8 denotes the invariant distribution of $\mathbb R^d$9, then

$\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$0

is an invariant distribution of the joint process, and the $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$1-marginal remains exactly $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$2 (Wang et al., 31 Aug 2025). The reason is that for each fixed friction value $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$3, the invariant kinetic law in $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$4 is the same Gibbs density, while the switching acts only on the regime coordinate.

The continuous-time contraction theorem makes the role of the regime process explicit. Under

$\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$5

with $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$6 and $\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$7,

$\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$8

where

$\pi(x)\propto e^{-f(x)},\qquad x\in\mathbb R^d,$9

This formula shows that convergence is governed jointly by the CTMC generator $f$0 and the inverse-friction matrix $f$1, rather than by a single scalar damping constant (Wang et al., 31 Aug 2025).

3. Discrete-time algorithmic form

FRS-KLMC discretizes FRS-KLD by combining a discrete approximation of the regime CTMC with a KLMC step whose coefficients depend on the current friction regime (Wang et al., 31 Aug 2025). If $f$2 is the current state and $f$3 is the step size, the regime update uses

$f$4

Conditional on $f$5, the position and velocity update is

$f$6

with

$f$7

The Gaussian vector $f$8 is centered in $f$9 with covariance

$m$0

The paper also gives an equivalent piecewise-constant-friction SDE interpretation: the iterate $m$1 has the same distribution as $m$2 for a kinetic Langevin system on each interval $m$3 with frozen friction $m$4 (Wang et al., 31 Aug 2025). Conceptually, each step is therefore a standard KLMC move, but with coefficients reparameterized by the currently active regime.

From an implementation standpoint, the dominant cost per iteration remains one gradient evaluation plus Gaussian sampling in $m$5 dimensions, with additional overhead from simulating the finite-state regime and recomputing $m$6 at the current friction value (Wang et al., 31 Aug 2025). In this respect FRS-KLMC is algorithmically close to standard KLMC, but not dynamically equivalent to it.

4. Non-asymptotic convergence and iteration complexity

The central discrete-time result is a recursive $m$7-Wasserstein bound for the position marginal $m$8. Under Assumptions 1 and 2, the friction lower bound

$m$9

and the step-size restriction

$M$0

the squared error satisfies

$M$1

where

$M$2

Equivalently,

$M$3

with

$M$4

(Wang et al., 31 Aug 2025).

The structure of the estimate is standard for unadjusted Langevin schemes but notable in its constants. The first term is geometric transient decay, controlled by the spectral quantity $M$5. The second is a stationary discretization bias of order $M$6 in $M$7, equivalently $M$8 in squared $M$9. In the paper’s comparison, this bias order is stronger than the $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$0 term reported for RS-KLMC (Wang et al., 31 Aug 2025).

The proof strategy conditions on the regime path $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$1, applies constant-friction KLMC contraction estimates stepwise, and then averages over the CTMC. The factors

$$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$2

are controlled through spectral analysis of the tilted matrix $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$3, which is the source of the parameter $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$4 (Wang et al., 31 Aug 2025).

For target accuracy $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$5, the paper states that it suffices to choose

$$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$6

and concludes a complexity of

$$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$7

(Wang et al., 31 Aug 2025). The source also notes a typographical inconsistency in the placement of the logarithm; a plausible reading of the displayed lower bound is

$$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$8

The substantive point is unchanged: the claimed dependence is better than the paper’s stated $$f(y)-f(x)-\langle \nabla f(x),y-x\rangle \ge \frac m2 \|y-x\|^2, \qquad \|\nabla f(y)-\nabla f(x)\|\le M\|y-x\|.$$9 behavior for RS-KLMC and $\psi$0 for RS-LMC (Wang et al., 31 Aug 2025).

5. Relation to fixed-friction, variable-friction, and geometry-adaptive Langevin methods

FRS-KLMC generalizes the fixed-friction kinetic Langevin framework developed for strongly log-concave targets. In the constant-friction setting,

$\psi$1

the near-optimal friction identified for mixing is

$\psi$2

or $\psi$3 when $\psi$4 (Dalalyan et al., 2018). This fixed-friction result is directly reflected in the FRS-KLMC admissibility condition

$\psi$5

which requires every regime to lie in a theorem-supported friction range (Wang et al., 31 Aug 2025).

A different but closely related line of work studies position-dependent matrix friction. The paper "Appropriate State-Dependent Friction Coefficient Accelerates Kinetic Langevin Dynamics" does not use the phrase “Frictional-Regime-Switching KLMC,” but its central message is that friction should be adapted to local curvature (Lim et al., 2023). Its recommended law,

$\psi$6

is a matrix-valued, geometry-adaptive analogue of critical damping, and it is proved to accelerate $\psi$7-convergence relative to any constant scalar friction for a broad class of strongly convex nonlinear potentials (Lim et al., 2023). This suggests that a finite collection of scalar friction regimes in FRS-KLMC can be interpreted as a coarse approximation to continuously state-dependent friction, although the theories are distinct: FRS-KLMC randomizes among finitely many scalar values, while the state-dependent theory uses a position-dependent matrix field.

The distinction becomes sharper in the small-mass limit. For variable friction $\psi$8, the naive overdamped limit is not generally correct; after regularization, the correct limit is Stratonovich,

$\psi$9

or, in isotropic Itô form,

$\gamma$0

(Freidlin et al., 2012). In one dimension the paper also derives explicit interface conditions for discontinuous friction. These effects do not appear in the basic FRS-KLMC formulation of (Wang et al., 31 Aug 2025), where $\gamma$1 is an independent CTMC rather than a spatially variable coefficient. They become relevant only when friction switching is tied to the state or when one seeks an overdamped reduction of a genuinely state-dependent scheme.

A further contextual result concerns the underdamped-to-overdamped transition for exponential-integrator KLMC. That analysis shows that the discretization remains stable in the overdamped regime provided proper time acceleration is used,

$\gamma$2

and identifies a numerical transition around

$\gamma$3

(Kim et al., 4 Oct 2025). This suggests that any high-friction phase in a friction-switching kinetic sampler should be interpreted jointly with timestep scaling rather than through $\gamma$4 alone.

6. Empirical behavior, limitations, and adjacent directions

The empirical behavior reported for FRS-KLMC is selective rather than uniform. In Bayesian linear regression, the paper studies two friction regime sets,

$\gamma$5

with a generator matrix chosen to have large spectral gap $\gamma$6. The reported outcome is that with small or narrow friction regimes, even a large spectral gap in $\gamma$7 does not provide acceleration, whereas with larger friction values FRS-KLMC accelerates convergence in this task (Wang et al., 31 Aug 2025). The stochastic-gradient analogue FRS-SGHMC is also reported to outperform SGHMC and RS-SGLD on synthetic and real logistic-regression datasets under the parameter settings used in the paper (Wang et al., 31 Aug 2025).

Several limitations are explicit. The main theory assumes strong convexity and smoothness of the potential. The friction lower bound

$\gamma$8

is restrictive, especially if one is interested in exploratory low-friction phases. The step-size condition depends on $\gamma$9, so very large friction values tighten the discretization budget. The theory does not provide an explicit optimal regime-design rule; it only shows how $2$00, $2$01, the stationary regime weights $2$02, and the CTMC generator $2$03 enter the bounds (Wang et al., 31 Aug 2025). These restrictions matter in practice because the experiments indicate that switching is beneficial only when the friction set itself is chosen meaningfully.

Adjacent switching work suggests one path beyond the present formulation. In switching Hamiltonian Monte Carlo, symmetric splitting schemes combined with exact simulation of the switching chain via uniformization or the stochastic simulation algorithm yield geometric ergodicity and second-order bias, and the paper states that its discrete-Poisson-equation approach can be generalized to other settings, for example, kinetic Langevin equations (Sharma, 11 Jun 2026). This suggests that future FRS-KLMC variants could combine regime-switching friction with symmetric operator splittings and higher-order invariant-measure analysis. Such a development would go beyond the current FRS-KLMC theory, but it is a natural extension of the existing switching-sampler framework.

Taken together, the literature places FRS-KLMC in a precise niche. It is not merely KLMC with noisy hyperparameters, nor is it the full state-dependent-friction program. It is a finite-state, Markovian randomization of the kinetic damping coefficient that preserves the Gibbs target marginal, admits explicit $2$04 non-asymptotic analysis, and empirically depends sensitively on the chosen friction regimes and switching law (Wang et al., 31 Aug 2025).