Passive Langevin Dynamics

• Passive Langevin dynamics is a class of stochastic processes where non-self-propelled particles evolve under conservative forces, damping, and thermal noise.
• It underpins coarse-graining methods by projecting microscopic dynamics onto latent variables while preserving key equilibrium statistics like the Gibbs or Boltzmann law.
• Extensions of passive Langevin dynamics include non-Markovian memory effects and adaptations for inference and singular interaction geometries in both equilibrium and nonequilibrium settings.

Passive Langevin dynamics denotes a family of stochastic dynamics for degrees of freedom that are not self-propelled and are driven by conservative or mean forces, dissipation, and noise supplied by an environment. In equilibrium settings, it is the standard thermal dynamics whose invariant law is Gibbs or Boltzmann; in coarse-grained and latent descriptions, it denotes an autonomous effective process for reduced variables that reproduces equilibrium statistics of an underlying microscopic system without back-influence on the unresolved degrees of freedom; and in some inference settings, “passive” refers to Langevin samplers that consume exogenous gradients they do not control (Olivares-Rivas et al., 2011, Bruce et al., 23 Oct 2025, Krishnamurthy et al., 2020).

1. Canonical stochastic structure

In its standard underdamped form, passive Langevin dynamics is built from a Hamiltonian that is quadratic in momenta. For microscopic variables $\vec q,\vec p\in\mathbb R^n$, one writes

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$

with $M(\vec q)$ symmetric positive definite and $U(\vec q)$ a potential energy. The associated Langevin SDE is

$d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$

$$d\vec p=-\nabla_{\vec q}H(\vec q,\vec p)\,dt-\gamma M^{-1}(\vec q)\vec p\,dt+\sqrt{2\gamma k_BT}\,d\vec W_t,$$

and its invariant density is

$$\rho(\vec q,\vec p)\propto \exp\!\left(-\frac{H(\vec q,\vec p)}{k_BT}\right).$$

This quadratic-in-momenta structure is the basic condition under which standard underdamped Langevin dynamics preserves the Boltzmann distribution (Bruce et al., 23 Oct 2025).

The same passive structure persists in overdamped limits, but with geometry- or medium-dependent drift and diffusion. For a tagged particle moving along the $z$-direction in a non-homogeneous anisotropic fluid, the overdamped equation is

$\frac{dz(t)}{dt}=v_\gamma(z)+\sqrt{2D(z)}\,\xi(t),$

with $D(z)$ a position-dependent diffusion coefficient and $$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$0. Interpreted in the Itô sense, it yields a Fokker–Planck equation with multiplicative noise, and the fluctuation–dissipation relation becomes position dependent; in particular, the naïve Sutherland–Einstein relation $$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$1 does not hold globally in anisotropic, non-homogeneous systems (Olivares-Rivas et al., 2011).

A geometric formulation on curved surfaces makes the same point in intrinsic coordinates. For a rigid inclusion on a curved surface, the momentum equations are supplemented by linear friction and Stratonovich white noise,

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$2

and the motion is passive precisely when the generalized forces are conservative and the noise covariance satisfies

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$3

Under these conditions, the invariant law is Gibbsian on phase space (Németh et al., 2024).

Microscopic derivations from system–bath models provide a rigorous basis for the same structure. In a quantum-to-classical derivation with a weakly coupled harmonic bath, the effective reduced dynamics of the slow system is

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$4

with a friction matrix $$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$5 determined by the microscopic coupling and a noise covariance fixed by fluctuation–dissipation. In that setting, Langevin dynamics appears as a passive thermostat derived from the underlying Hamiltonian model rather than imposed ad hoc (Hoel et al., 2019).

2. Coarse-graining, latent variables, and effective closure

A central use of passive Langevin dynamics is the construction of reduced models. For a coarse-graining map

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$6

one seeks a coarse free energy of the form

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$7

so that the latent or coarse variables admit their own underdamped Langevin dynamics. The key structural condition is

$$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$8

with $$H(\vec q,\vec p)=\frac12\,\vec p^\top M^{-1}(\vec q)\vec p+U(\vec q),$$9 the Jacobian of $M(\vec q)$0. For this to define a genuine coarse mass matrix $M(\vec q)$1, the projected inverse mass $M(\vec q)$2 must be constant along each level set $M(\vec q)$3. Under that condition, the constrained free energy factorizes into quadratic kinetic and scalar potential parts, and the latent variables support a bona fide Langevin dynamics with invariant density $M(\vec q)$4 (Bruce et al., 23 Oct 2025).

This criterion substantially enlarges the admissible family of coarse variables. The isometric-invariant construction identifies embeddings for which configurations on the same level set are related by isometries preserving the microscopic mass matrix. Distance-based latent spaces furnish a concrete example, and the same framework is explicitly described as paving the way for Langevin dynamics on non-geometric coarse-graining representations such as principal components of time-lagged independent component analysis and neural latent embeddings (Bruce et al., 23 Oct 2025).

For passive coarse-graining in a broader sense, one may distinguish between the exact coarse-grained process and an effective Markov closure. In the overdamped and Langevin settings, the exact coarse-grained dynamics is obtained by taking conditional expectations of the raw projected coefficients, whereas the effective dynamics replaces them by equilibrium conditional expectations with respect to the Gibbs measure. The discrepancy between these two objects can be quantified in relative entropy and Wasserstein distance. The resulting bounds show that the quality of a coarse-graining map is controlled by functional inequalities on the conditional Gibbs measures along level sets, which encode the scale separation of the unresolved variables (Duong et al., 2017).

This establishes a recurring theme: passive Langevin dynamics on reduced spaces is not automatic. It requires a compatibility between the coarse variables, the kinetic structure, and the conditional equilibrium geometry. When that compatibility is available, the reduced process is autonomous and equilibrium-consistent; when it fails, the projected variables need not form a closed Markovian Langevin system.

3. Memory, geometry, and singular interactions

Passive Langevin dynamics is often generalized beyond the Markovian, constant-friction setting. In coarse-grained particle systems with incomplete time-scale separation, the appropriate model is a generalized Langevin equation,

$M(\vec q)$5

with a fluctuation–dissipation relation

$M(\vec q)$6

The memory kernel is decomposed into self- and pair-contributions, and in the generalized Langevin dynamics construction these kernels are iteratively reconstructed from dynamical correlation functions of an underlying fine-grained system. This yields a non-Markovian passive coarse-grained dynamics suitable for systems with incomplete time-scale separation, including hydrodynamic memory (Jung et al., 2018).

The same expansion of passive Langevin dynamics into more structured settings appears on manifolds. For rigid inclusions constrained to curved surfaces, the intrinsic Hamiltonian dynamics exhibits curvature-induced coupling between translational and angular momenta. Adding linear friction, white noise, and configuration-dependent forces produces an intrinsic Langevin equation in phase space whose overdamped limit is derived by adiabatic elimination of momenta from the Fokker–Planck equation. The nontrivial point is that surface curvature modifies the integrability conditions for the forces and torques, thereby distinguishing passive, potential-driven motion from active, non-potential driving (Németh et al., 2024).

Singular interactions require yet another reformulation. For hard spheres in a solvent, standard force-based Langevin integrators fail because hard-body collisions are not slowly varying on the integration time scale. Event-driven algorithms based on a splitting of the Kramers operator, or on an approximation of the two-body Green’s function, treat the collisions exactly at the event level while retaining the passive underdamped Langevin bath. In this setting, the passive dynamics is still defined by linear friction, Gaussian white noise, and fluctuation–dissipation, but its numerical realization must respect the singular interaction geometry (Scala, 2011).

At the opposite extreme, in the limit $M(\vec q)$7 the passive many-body Langevin dynamics of interacting particles becomes exactly reducible to a single one-dimensional effective stochastic equation with self-consistent kernels. In that representation, the local environment of a tagged degree of freedom becomes an effective bath with memory, and quantities such as pressure, shear stress, mean-square displacement, and response are computed from a self-consistent scalar process rather than the original $M(\vec q)$8-body system (Agoritsas et al., 2018). This shows that non-Markovianity in passive Langevin dynamics can arise either from coarse-graining or from exact mean-field reduction.

4. Passive dynamics away from equilibrium

Passive does not mean equilibrium. A passive particle may evolve out of equilibrium whenever the surrounding medium is spatially heterogeneous, externally driven, or itself active, provided the particle has no self-propulsion of its own.

In confined non-homogeneous fluids, a passive tagged particle is driven by a potential of mean force, linear friction, and thermal Gaussian noise. The observable of interest in that context is the persistence probability $M(\vec q)$9, the probability that a particle initially in a virtual layer remains in that layer at time $U(\vec q)$0. In the Langevin or Smoluchowski description this quantity is well approximated by a single exponential, $U(\vec q)$1, and the average persistence time is the mean first-passage time averaged over the equilibrium density in the layer. The continuum passive model captures the spatial variation of persistence times, while a simple scaling relation is required to match the absolute molecular-dynamics time scale (Olivares-Rivas et al., 2011).

A different nonequilibrium passive setting is a background flow field. For a large particle immersed in a bath whose local mean velocity is $U(\vec q)$2, the small-bath-mass limit yields

$U(\vec q)$3

Here the deterministic relaxation is toward the local fluid velocity, yet the friction coefficient and the noise amplitude are the same coefficients that appear in the equilibrium derivation, and the fluctuation–dissipation relation is preserved. This is the natural passive Langevin model for a tracer in an imposed incompressible flow (Dobson et al., 2012).

The strongest departure from equilibrium in the supplied literature occurs for passive rigid bodies submerged in a chiral active bath. In the adiabatic, large-mass limit, the effective rigid-body dynamics is underdamped and Langevin-like, with friction and noise tensors obtained from Kubo-type formulas. For a rotationally symmetric disk, odd diffusion and odd mobility satisfy an Einstein relation in the adiabatic regime; for a rod, translational and rotational sectors decouple but acquire distinct effective temperatures; and for a wedge, translation–rotation coupling renders the dynamics fully irreversible. The odd Einstein relation breaks down outside the adiabatic limit, and the second fluctuation–dissipation relation fails in the odd sector even when the disk has an effective equilibrium phase-space distribution (Hargus et al., 2024).

The infinite-dimensional mean-field formulation reinforces the same distinction. The bath may be thermal and passive at the microscopic level, yet the emergent dynamics of the interacting system may display aging, FDT violation at the level of correlation and response, and glassy state-following. In that framework, passive Langevin dynamics supports both equilibrium liquids and out-of-equilibrium glassy regimes within one exact self-consistent formalism (Agoritsas et al., 2018).

5. Learned, inferred, and reweighted passive Langevin models

Recent work extends passive Langevin ideas from physical coarse-graining to statistical modeling and inference. In neural latent-variable modeling, the latent state may itself be assigned an underdamped Langevin prior,

$U(\vec q)$4

with a time-independent learned potential $U(\vec q)$5. In the model called LangevinFlow, $U(\vec q)$6 is parameterized as a network of locally coupled oscillators, biasing the latent dynamics toward oscillatory and wave-like behavior. Once the potential is learned, the latent evolution is autonomous and is explicitly described as conceptually close to passive Langevin dynamics, because the time variation is generated by a fixed landscape, damping, and thermal-like noise rather than explicit controls (Song et al., 15 Jul 2025).

In inverse reinforcement learning, passive Langevin dynamics appears in a different sense. The inverse learner aims to reconstruct an unknown reward or cost function from observed stochastic-gradient iterates of a forward learner, but it does not control the points at which gradients are evaluated. In reward formulations, the passive Langevin algorithm is built so that its stationary distribution is proportional to $U(\vec q)$7; in cost formulations, the stationary law is $U(\vec q)$8. In both cases, the Langevin chain acts as a randomized sampler whose stationary measure encodes the objective function. Kernel weighting compensates for the fact that the observed gradients are exogenous and generally evaluated at states different from the inverse learner’s current state (Krishnamurthy et al., 2020, Snow et al., 2023).

The finite-sample theory of passive stochastic gradient Langevin dynamics makes this precise. Under smoothness, dissipativity, kernel regularity, and sampling-distribution assumptions, the law $U(\vec q)$9 of the passive chain after $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$0 steps satisfies an explicit 2-Wasserstein bound with respect to the Gibbs target $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$1. The proof decomposes the error into a discretization term and a diffusion-convergence term, and uses log-Sobolev, Poincaré, and weighted transportation inequalities for the limiting diffusion operator (Snow et al., 2023).

A separate computational theme concerns dynamic reweighting of passive Langevin trajectories. For Langevin dynamics propagated by a variant of the Langevin Leapfrog integrator, the exact path probability ratio $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$2 can be derived for potential perturbations, and this ratio matches the discrete stochastic integrator used in the simulation. A previously proposed approximate ratio $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$3 differs from the exact one only by $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$4. The same work also shows that path probability ratios derived for overdamped Euler–Maruyama dynamics do not apply to underdamped Langevin trajectories generated by this integrator, even in equilibrium passive settings (Kieninger et al., 2020).

6. Validity criteria, misconceptions, and recurring limitations

A common misconception is that passive Langevin dynamics is necessarily overdamped, Markovian, or equilibrium. The literature does not support any of these identifications. Passive rigid-body dynamics in a background flow is underdamped (Dobson et al., 2012); passive generalized Langevin models with memory are explicitly non-Markovian (Jung et al., 2018); and passive objects in chiral active baths or glassy passive many-body systems can be far from equilibrium (Hargus et al., 2024, Agoritsas et al., 2018).

A second misconception is that any projected or learned coordinate automatically supports a passive Langevin closure. In coarse-grained molecular dynamics, the central obstruction is kinetic: if $d\vec q=\nabla_{\vec p}H(\vec q,\vec p)\,dt,$5 is not constant on the level sets of the embedding, then the reduced free energy need not be quadratic in the coarse momenta and a standard latent Langevin equation is not available (Bruce et al., 23 Oct 2025). In quantitative coarse-graining, this mismatch is reflected in explicit relative-entropy and Wasserstein errors between the exact coarse-grained process and an effective Markov closure, with the size of the error governed by functional inequalities of the conditional Gibbs measures (Duong et al., 2017).

A third recurring issue is that passive modeling assumptions are often only asymptotically correct. In confined fluids, the overdamped Smoluchowski description reproduces relative variations of persistence times but not the absolute molecular-dynamics clock without calibration (Olivares-Rivas et al., 2011). In chiral active baths, odd mobility and odd diffusion satisfy an Einstein relation only in the adiabatic, large-mass limit, and that relation breaks at finite mass (Hargus et al., 2024). In underdamped trajectory reweighting, path-probability formulas must be integrator-consistent; otherwise even equilibrium passive dynamics is reweighted incorrectly (Kieninger et al., 2020).

These limitations do not undermine the concept. Rather, they define its domain of validity. Passive Langevin dynamics is most reliable when the drift is conservative or a well-defined mean force, the friction–noise structure is explicitly characterized, the reduced variables are compatible with the kinetic geometry, and the unresolved dynamics is either negligible or represented by memory kernels rather than discarded. Under those conditions, it provides a unifying language for equilibrium sampling, coarse-grained molecular simulation, nonequilibrium tracer dynamics, latent-variable modeling, and passive inference schemes across physics, chemistry, and machine learning.