The Langevin equation is an equation of motion in which a system experiences both systematic forces and a randomly fluctuating influence. Originally introduced to describe Brownian motion, it provides a trajectory-based account of how microscopic interactions produce observable randomness. Its simplest form supplements Newton’s equation with a frictional force and a rapidly varying random force representing collisions with a surrounding fluid.
The equation connects classical mechanics, statistical mechanics, and the mathematics of stochastic differential equations. It explains how apparently irregular particle motion leads to predictable statistical laws, including diffusion, thermal velocity distributions, and relaxation toward equilibrium. Generalizations describe systems with memory, state-dependent fluctuations, many interacting particles, and quantum environments.
“Langevin equation” therefore names both a particular model of a Brownian particle and a broad modeling framework. Its validity depends on how the resolved variables are chosen, which environmental influences have been eliminated, and whether the assumed noise and dissipation accurately represent those influences.
Historical origin and physical construction
Paul Langevin introduced his treatment of Brownian motion in 1908, following the diffusion-based theories developed by Albert Einstein in 1905 and Marian Smoluchowski in 1906. Einstein’s approach related the statistical spreading of suspended particles to measurable transport coefficients. Langevin instead began with a mechanical equation for an individual particle and separated the fluid’s action into an average resistance and a fluctuating remainder. This distinction made the connection between mechanical trajectories and statistical behavior especially direct.
For a particle moving along one coordinate, the standard inertial equation is
Here is the particle’s mass, its velocity, a systematic external or interaction force, a drag coefficient, and a random force. The term describes linear resistance relative to a stationary medium. If the fluid moves with velocity , the corresponding drag is usually proportional to , rather than to the laboratory-frame particle velocity.
For a spherical particle of radius moving slowly through an unbounded Newtonian fluid with dynamic viscosity , Stokes’ law gives
This expression assumes creeping flow, a no-slip particle surface, and negligible boundary effects. Nearby walls, nonspherical geometry, slip, or appreciable fluid inertia change the drag. The Langevin framework does not require Stokes’ formula, but the formula supplies a particularly useful microscopic interpretation for colloidal particles.
The physical construction is a form of coarse-graining. A suspended particle interacts with enormous numbers of fluid molecules, but their individual positions and velocities are not followed. Their collective influence is represented by a small number of statistical quantities. Friction summarizes the average transfer of momentum to the environment, while the random force represents fluctuations around that average. These are not generally independent physical mechanisms: in a thermal environment, both arise from the same microscopic interactions.
The simplest model assumes that environmental correlations decay much faster than the particle’s resolved motion. The random force can then be approximated as instantaneous in its correlations. This approximation is useful over an intermediate range of times: long compared with microscopic collision correlations, but not necessarily long compared with the particle’s velocity relaxation time. At sufficiently short times, molecular structure and fluid memory invalidate the idealization. At larger scales, omitted boundaries, interactions, or external driving may become important.
Thermal noise and the fluctuation–dissipation relation
The standard equilibrium model specifies a zero-mean random force,
with correlation
Here is the absolute temperature, is the Boltzmann constant, angle brackets denote an ensemble average, and is the Dirac delta distribution. A Gaussian random force with this correlation is called Gaussian white noise. In several spatial dimensions, independent Cartesian components satisfy the corresponding relation with an additional Kronecker delta.
The noise amplitude is fixed by thermal consistency, not chosen independently of friction. For a free particle, the stationary velocity variance must obey the equipartition theorem,
for each Cartesian component. Suppose the force correlation is instead written as . The velocity variance evolves according to
Its stationary value is , so equipartition requires .
This relationship is a basic example of the fluctuation–dissipation theorem. Dissipation removes mechanical energy from the resolved particle, while thermal fluctuations continually return energy to it. Friction without noise would cool the particle toward rest. Noise without friction would increase its velocity variance without limit. Their equilibrium balance produces a finite stationary distribution rather than either outcome.
The numerical coefficient depends on the definition of the friction parameter. Some authors write the drag as , where has units of inverse time. Their force covariance is then . Others write an equation for velocity with a noise measured in acceleration units. These are equivalent conventions when , but interchanging their amplitudes without changing units gives an incorrect model.
White noise has a flat power spectrum and no finite instantaneous variance in the ordinary sense. This does not imply that a physical fluid exerts forces of literally infinite magnitude. It means that the idealized force is defined through its time integrals and correlations. A real environment has a finite correlation time and a high-frequency cutoff. Moreover, a delta-function covariance alone does not establish Gaussianity: the Gaussian assumption supplies information about higher-order statistics. Thermal white Gaussian noise is therefore a specific approximation, not a universal property of all random forcing.
Stochastic interpretation and trajectory-level meaning
The mathematically precise form of the ordinary Langevin equation uses a Wiener process :
A Wiener increment has zero mean and variance , and increments over disjoint time intervals are independent. Informally, white noise is written as , but Wiener trajectories are almost surely nowhere differentiable. Consequently, the white-noise-driven equation is an integral equation rather than an ordinary differential equation with an unusually irregular but conventional force.
For example, integrating the velocity equation from to gives
taking . In the inertial model, velocity is continuous but not ordinarily differentiable. Position, being the time integral of a continuous velocity, has a different degree of regularity. In the overdamped approximation, position itself is directly driven by Wiener increments and loses this differentiability.
The general stochastic equation
contains a drift vector and a noise-amplitude matrix . Its instantaneous diffusion matrix is
The factor of one-half is conventional but important. For scalar motion written as , the mean-square random displacement over a short interval is .
When the noise amplitude depends on the current state, the stochastic integral requires a convention. The two most widely used are Itô calculus and Stratonovich calculus. Itô integration evaluates the noise coefficient at the beginning of each infinitesimal step; Stratonovich integration corresponds to a midpoint limit. They generally assign different drifts to the same symbolic equation. With constant noise amplitude, as in the elementary inertial model, the two conventions coincide.
A further distinction concerns the type of solution. A strong solution constructs trajectories relative to a specified Wiener process and initial condition. A weak solution specifies a probability law satisfying the stochastic dynamics without requiring that particular construction. Under suitable regularity assumptions on the coefficients, both existence and uniqueness are well controlled; singular forces, vanishing diffusion, or rapid coefficient growth can require additional analysis. Thus, writing a Langevin equation is only part of defining a model: initial distributions, boundaries, noise interpretation, and coefficient assumptions also determine its meaning.
Exact solutions, relaxation, and measurable correlations
For a free particle, the velocity follows the Ornstein–Uhlenbeck process. Defining the inertial relaxation time
the solution is
An initially imposed mean velocity decays as . For a fixed initial velocity, the conditional variance grows from zero to , approaching its equilibrium value on the scale for the variance’s exponential relaxation.
If the initial velocity is already drawn from its thermal equilibrium distribution, the stationary velocity autocorrelation is
This correlation quantifies how long the particle retains memory of its direction and speed. Although individual trajectories remain irregular, averaging over many trajectories produces a smooth and reproducible decay. Its integral determines the long-time diffusion coefficient:
This is the Einstein relation between diffusion and mobility.
Integrating the velocity correlation gives the one-dimensional mean-square displacement for thermally distributed initial velocities:
At short times, , expansion of the exponential gives
This is the ballistic regime: the particle approximately preserves its initial velocity. At long times, , the leading behavior is , the usual diffusive law. In three independent dimensions, the corresponding long-time displacement is .
A harmonic restoring force, , produces a thermally driven damped oscillator:
Its deterministic relaxation is underdamped when and overdamped when . Equilibrium gives
Thus, confinement limits positional fluctuations even though the random force continues to act.
With the two-sided angular-frequency spectral convention, the position spectrum is
This expression connects a stochastic equation to experimentally measured frequency-dependent fluctuations. Optical trapping and micromechanical measurements use such correlations and spectra to infer drag, stiffness, and relaxation times. Finite sampling, detector noise, and deviations from instantaneous drag must be distinguished from the intrinsic dynamics.
Probability evolution and the overdamped limit
A Langevin equation describes individual stochastic paths, while the Fokker–Planck equation describes the evolution of their probability density. For a particle in a potential , with , the density satisfies
This phase-space equation is often called the Kramers equation. Its first term transports probability through position space; the remaining terms describe acceleration, frictional contraction in velocity space, and noise-induced velocity spreading.
For a time-independent confining potential, the equilibrium solution is
where is the normalization factor. This is the Boltzmann distribution for mechanical energy. A normalizable stationary distribution requires appropriate confinement or boundaries: a free particle on the infinite line has a stationary velocity distribution but no normalizable stationary position distribution.
When is much shorter than the timescale of interest, velocity can be eliminated. With constant drag and temperature, the resulting overdamped equation is
Its position density obeys the Smoluchowski equation,
The associated probability current is
At thermal equilibrium in a conservative potential, the drift and diffusion contributions cancel and .
The overdamped limit is not equivalent to assuming that the particle physically lacks mass. It is a reduced description that omits a rapidly relaxing variable. It reproduces long-time position statistics but cannot describe the ballistic regime, the full velocity distribution, or inertial resonances. Simply dropping the acceleration term can also fail when friction varies with position or when several rapidly evolving variables are coupled.
Boundary conditions are part of either probability formulation. Reflecting boundaries prohibit probability flow through a wall, while absorbing boundaries remove trajectories when they reach a specified region. The latter are central to first-passage problems, including reaction times and escape from potential wells. In 1940, Hendrik Kramers used a Brownian-particle model to analyze thermally activated escape. Such calculations show how an activation barrier and environmental friction jointly determine rates, connecting continuous stochastic motion with chemical kinetics.
Multiplicative noise and nonequilibrium thermodynamics
Noise is called multiplicative when its amplitude depends on the system’s state. This occurs naturally when mobility varies near a surface, when different coordinates experience different resistance, or when fluctuations scale with a population or concentration. In the scalar equation
the same written functions and generally describe different processes under Itô and Stratonovich interpretations. The conversion is
An explicit convention therefore carries physical content unless the drift has already been adjusted to represent the same underlying process.
A useful equilibrium example is overdamped motion with position-dependent mobility at uniform temperature. The local diffusion coefficient is . The Itô equation consistent with the equilibrium positional distribution is
The additional drift is often called noise-induced or thermal drift. It is required because spatial variation of fluctuations contributes to probability transport.
The corresponding probability current is
Substitution of makes this current vanish. Omitting the additional Itô drift instead produces an unwanted mobility-dependent equilibrium weighting. This example illustrates why stochastic conventions cannot be selected solely for algebraic convenience. They must match the limiting physical model, including the variables eliminated in obtaining the overdamped description.
Langevin trajectories also support stochastic thermodynamics, which assigns fluctuating work and heat to individual realizations. For an overdamped particle in a potential , where is an externally controlled parameter, the Stratonovich chain rule gives
The first term is work done on the system through parameter changes. If heat delivered to the environment is defined by , then
The circle indicates Stratonovich integration. An Itô formulation is also possible, but its energy balance includes the corresponding second-derivative correction.
External nonconservative forces, time-dependent driving, or multiple thermal reservoirs can maintain nonequilibrium steady states with persistent probability currents. In such systems, equilibrium noise–friction relations may apply separately to individual thermal reservoirs without implying global equilibrium. Entropy production can be related to the relative probabilities of forward and time-reversed trajectories under specified assumptions. Conversely, an arbitrary fitted random force need not have a meaningful temperature, and matching a stationary variance alone does not establish thermal equilibrium.
Memory effects and generalized Langevin equations
Instantaneous friction assumes that the environment responds without appreciable delay. When eliminated degrees of freedom retain correlations, the force on a particle depends on its past motion. A common generalized Langevin equation is
where is a memory kernel and is a fluctuating force. The kernel determines how earlier velocities contribute to the present drag. A slowly decaying kernel can produce behavior qualitatively different from ordinary exponential relaxation.
For a suitable equilibrium preparation and a consistent projection of the microscopic dynamics, the classical fluctuation–dissipation relation takes the form
This equation uses the convention in which appears directly in the force convolution above. Alternative definitions factor out the mass or use a different endpoint convention for delta-function kernels. Initial conditions matter: preparing the particle and environment independently can introduce transient terms that are absent when their equilibrium correlations are retained.
Systematic foundations were developed through projection-operator methods, notably in work by Robert Zwanzig and Hazime Mori in the 1960s. These methods separate selected observables from the remaining microscopic degrees of freedom. The exact reduced equation contains a memory term and an orthogonal fluctuating force. Under suitable timescale separation, rapidly decaying memory can be approximated by local friction. Thus, an apparently simple Langevin equation can arise as an approximation to deterministic many-body dynamics rather than as a fundamental postulate of randomness.
Fluid inertia provides a concrete source of memory. A moving particle creates a flow field that takes time to diffuse through the surrounding fluid. The resulting hydrodynamic interactions produce history-dependent resistance and long-time tails in velocity correlations. Viscoelastic materials introduce additional delayed responses because their internal stresses relax over a range of times. Depending on the kernel and observation window, a particle’s mean-square displacement may scale approximately as with , a form of anomalous diffusion. Such scaling is not unique evidence for a particular memory mechanism; trapping and heterogeneous dynamics can produce similar observations.
Quantum Langevin equations extend the framework to operators governed by quantum mechanics. Environmental fluctuations must then satisfy operator correlations and commutation constraints, not merely classical noise covariances. The equilibrium fluctuation spectrum depends on frequency and includes zero-point contributions as well as thermal ones. In cavity optics and related open-system models, input-noise operators accompany damping so that the dynamics preserve the required commutators. At high temperature and sufficiently low frequencies, appropriate quantum fluctuation relations recover their classical counterparts, but a classical white-noise replacement is generally inadequate at low temperature or high frequency.
Numerical integration and statistical sampling
Computer simulation replaces continuous trajectories with discrete updates. For the overdamped equation
the Euler–Maruyama method gives
where the are independent standard normal random variables. The factor is essential: Wiener increments have variance proportional to elapsed time. Scaling the random displacement linearly with would cause the simulated diffusion to vanish as the step size decreases.
For constant-coefficient velocity relaxation, an exact finite-time update is available:
It preserves the correct equilibrium velocity variance at any step size for this isolated Ornstein–Uhlenbeck subproblem. Simultaneously updating position exactly requires a correlated Gaussian displacement; using an independent positional random increment generally gives incorrect joint statistics. An exact noise substep also does not make a full simulation exact when nonlinear forces are approximated.
In molecular dynamics, Langevin dynamics commonly serves as a thermostat. Deterministic motion under interatomic forces is combined with damping and thermal random forcing, often through operator-splitting schemes. One widely used arrangement, BAOAB, alternates force, position, and Ornstein–Uhlenbeck updates in a symmetric order. Such methods can give favorable equilibrium configurational sampling, although accuracy depends on timestep, force stiffness, friction, and the observable being measured.
Correct equilibrium sampling and correct dynamics are separate requirements. A large damping coefficient may efficiently control temperature while substantially altering diffusion coefficients, barrier-crossing rates, and time correlations. Conversely, weak damping can preserve some dynamical features but equilibrate slowly. Independent particle thermostats also remove momentum locally and need not preserve the hydrodynamic behavior of a momentum-conserving fluid. The intended use of a simulation therefore determines which artifacts matter.
The same mathematical structure underlies Langevin Monte Carlo. For a target density , the dimensionless diffusion
has as its invariant density under appropriate confinement and regularity conditions. A discrete unadjusted algorithm generally samples a biased approximation because of finite-step error. A Metropolis acceptance step can correct that discretization bias, producing the Metropolis-adjusted Langevin algorithm, though its accepted and rejected moves are no longer the original physical dynamics. This use connects stochastic mechanics with Bayesian inference and machine learning without requiring the simulated variables to represent actual particle coordinates.
Applications, model identification, and limits
Colloidal physics remains a central application because suspended particles provide directly observable stochastic trajectories. Combining the Einstein relation with Stokes drag gives
For a sphere of radius in water near room temperature, taking gives . The value changes with temperature, particle size, boundaries, and surface conditions. Tracking such motion can reveal material properties, while trapped-particle fluctuations allow local measurements of mechanical response.
Biomolecular models often use Langevin dynamics to represent a selected molecule while treating the solvent implicitly. This reduces computational cost compared with following every solvent molecule. Coarse coordinates such as molecular extension, domain separation, or a protein-folding coordinate may then evolve on an effective free-energy landscape. The appropriate drift is determined by that landscape and by the coordinate-dependent mobility. A scalar reaction coordinate can nevertheless hide slow internal variables, producing memory or an apparently position-dependent diffusion coefficient even when the underlying molecular dynamics is more complex.
Electrical circuits provide a closely related realization. For a series resistor–inductor circuit at classical thermal equilibrium,
where is inductance, is resistance, and is thermal voltage noise. In the two-sided time-domain convention,
The correspondence with mechanical velocity relaxation is and . The stored magnetic energy then has the same equilibrium quadratic structure as mechanical kinetic energy.
In spatially extended systems, Langevin-type field equations describe fluctuating order parameters, concentration fields, and interface heights. Whether the variable is conserved determines how both dissipation and noise enter. A conserved concentration requires random fluxes whose divergence changes the local concentration without changing the total amount in a closed system. A nonconserved order parameter can instead relax locally. These distinctions are important near a phase transition, where slow collective fluctuations can dominate the dynamics.
Inferring a Langevin model from data requires more than fitting an irregular trajectory. For a scalar Itô diffusion with sufficiently resolved observations, conditional short-time increments identify the drift and diffusion through
Real measurements have finite exposure times, localization errors, and incomplete access to relevant variables. Measurement noise can imitate rapid fluctuations, while unresolved inertia or memory can invalidate an overdamped Markov model. Reliable identification therefore examines several observables—increment distributions, stationary densities, correlation functions, and responses to applied forces—rather than treating agreement with a single mean-square-displacement curve as a unique determination of the mechanism.