Langevin dynamics

Langevin dynamics refers to the stochastic differential equations that describe the motion of particles in a fluid under the combined influence of deterministic forces and random collisions. Named after Paul Langevin, who introduced the framework in 1908 to extend Newton's equations to Brownian motion, Langevin dynamics captures the essential physics of systems that are too large for quantum mechanics and too small for purely classical treatment: colloids, polymers, biomolecules, and any system where thermal fluctuations compete with structured forces.

In Markov Chain Monte Carlo, Langevin dynamics inspires the Metropolis-adjusted Langevin algorithm (MALA), which uses gradient information to propose states that follow the local geometry of the target distribution. Unlike the random walk proposals of standard Metropolis-Hastings, MALA proposals are biased toward regions of higher probability, dramatically improving mixing in well-behaved targets. The algorithm can be seen as a discretization of continuous Langevin diffusion, with a Metropolis accept-reject step correcting for the discretization error.

The connection between Langevin dynamics and MCMC reveals a deeper principle: systems that combine directed motion (gradients, forces) with random perturbation (thermal noise, proposal randomness) can efficiently explore complex landscapes without getting trapped. This principle appears not only in statistical sampling but in optimization algorithms, reinforcement learning, and evolutionary biology. The Langevin equation is the mathematical expression of a universal strategy: use structure to guide exploration, use noise to escape traps.

Langevin dynamics is the physics of compromise: not deterministic enough to be predictable, not random enough to be structureless. It is the mathematical description of what it feels like to be pushed by forces you can measure and buffeted by forces you cannot.