Why introduce velocity?

It preserves directional information between movements and changes exploration dynamics.

Is the target normalization constant needed?

Often not; gradients of the log density remove a constant independent of position.

Can stochastic minibatch gradients be substituted freely?

No. Additional gradient noise requires its own analysis or algorithmic treatment.

AI Glossary / Probabilistic machine learning / Advanced

Underdamped Langevin dynamics

Definition

Underdamped Langevin dynamics is a stochastic process that evolves both position and velocity under a potential force, friction and random noise. In statistical sampling, the potential is chosen so that the position's equilibrium distribution matches a target density. Retaining velocity gives the process momentum, distinguishing it from overdamped Langevin dynamics, which evolves position alone.

What it means

A sampler must explore a distribution rather than simply find its highest point. A potential force guides movement toward higher-density regions, momentum carries movement across successive steps, friction dissipates velocity, and noise sustains random exploration.

The balance between friction and noise is essential. Adding arbitrary noise to a momentum optimizer does not automatically create a correct sampler. The continuous process, its numerical approximation and any acceptance correction are separate pieces of the method. Their properties must be distinguished when claiming that produced samples represent the intended target.

How it works

Use a simple unit-mass, unit-temperature convention with target density proportional to exp(minus U(x)). The continuous equations are dx = v dt and dv = minus gradient U(x) dt minus gamma v dt plus the square root of 2 gamma times dW. Here gamma is positive friction and W is standard Brownian motion.

Under suitable regularity and recurrence conditions, the joint invariant density is proportional to exp[minus U(x) minus squared velocity norm divided by two]. Its position marginal is the desired target, and its velocity marginal is standard normal. Other mass or temperature conventions require corresponding changes to the equations.

A computer approximates these continuous dynamics with finite steps. Different integrators handle force, movement and friction-noise updates differently. Discretization can distort the invariant distribution; small steps, suitable integrators or appropriately designed corrections address different aspects of that problem.

Assess exploration using multiple chains and sampling diagnostics. Theoretical acceleration results typically rely on stated assumptions, such as smoothness and strong log-concavity. They do not establish that every multimodal posterior will be explored efficiently.

Three practical examples

1. A Bayesian modeling researcher uses momentum-based stochastic dynamics to explore a continuous posterior with differentiable log density.

2. A simulation scientist studies how friction and thermal noise affect the motion of particles in a specified energy landscape.

3. A numerical-analysis team compares discretizations by checking their empirical equilibrium moments against a distribution with known moments.

Think of it this way

Think of a ball moving across a shaped surface while experiencing drag and continual small random kicks. The ball retains momentum, so it need not restart its direction at every instant. For sampling, however, the surface, drag and kicks must be coordinated so that time spent in each region matches a particular probability distribution.

Worked example

Take a one-dimensional standard normal target, so U(x) = x squared divided by two and its gradient is x. Use friction gamma = 0.5, step h = 0.1, current position x = 1 and velocity v = 0.4.

For an illustrative explicit Euler step, update position using the old velocity: x-new = 1 + 0.1 times 0.4 = 1.04. Update velocity using the old position and velocity: v-new = 0.4 minus 0.1 times 1 minus 0.1 times 0.5 times 0.4 plus square root of 0.1 times z, where z is standard normal.

The deterministic part is 0.28. If the sampled z is 0.5, the random contribution is approximately 0.1581, giving v-new approximately 0.4381. Conditional on the old state, the new velocity has mean 0.28 and variance 0.1.

For comparison, the exact friction-and-noise substep alone over this duration multiplies velocity by exp(minus 0.05) and adds noise with variance 1 minus exp(minus 0.1), approximately 0.09516. That substep excludes the potential force and position movement; it is not the complete exact dynamics.

The Euler calculation illustrates mechanics only. One step neither produces an independent target draw nor proves that this discretization has the exact standard normal invariant position distribution.

Illustrative case study

Fictional case: Juniper Analytics evaluates samplers for a Bayesian continuous-parameter model. An underdamped implementation produces rapidly moving trajectories, which initially look encouraging.

The team first tests it on a normal target with known mean and variance, then varies step size and friction. It finds that visually active trajectories can still produce biased moments at an unsuitable step size. Analysts compare effective sampling efficiency and agreement across chains before applying the method to the business model. Movement alone is not treated as evidence of successful inference.

Common misconceptions and mistakes

Momentum does not guarantee independent samples or rapid switching between distant modes. Low friction can preserve long trajectories but does not automatically solve exploration barriers.

Underdamped sampling is not ordinary gradient descent with momentum: its objective is a distribution, not merely a minimizer. Removing noise changes that objective, and changing noise without the corresponding friction balance changes the process.

An invariant distribution of the continuous equation is not automatically an invariant distribution of a finite-step implementation. Also, the term underdamped in sampling literature identifies the position-velocity formulation; it should not be interpreted as a universal promise of oscillatory behavior for every local potential and friction setting.

Business use

The business relevance lies in probabilistic modeling and uncertainty estimation when continuous posterior sampling is required. Benefits depend on useful samples per unit of computation, not raw update speed. Teams should account for gradient cost, step-size sensitivity, numerical stability and diagnostic results. Comparisons with established samplers on the actual model are more informative than a generic claim that momentum is faster.

Check your understanding

In the example, which state is used by the explicit Euler force?

Answer: The old position x = 1.

Does the exact friction-noise substep solve the whole sampling problem?

Answer: No. It covers only one component of the dynamics.

Related terms

Hamiltonian Monte Carlo
Stochastic gradient Hamiltonian Monte Carlo
Langevin dynamics
overdamped Langevin dynamics
Brownian motion
invariant distribution and discretization bias

Learn next

Next, verify equilibrium moments for a Gaussian target under several integrators. Then study mass matrices, splitting schemes and the assumptions behind convergence guarantees before adapting the method to complex posteriors.

Sources

Primary source: Xiang Cheng, Niladri S. Chatterji, Peter L. Bartlett and Michael I. Jordan, Underdamped Langevin MCMC: A Non-Asymptotic Analysis: https://proceedings.mlr.press/v75/cheng18a/cheng18a.pdf

The worked step uses an explicitly stated unit-mass convention and is original teaching material, as is the fictional case.

Independent review consulted Equation 1, the invariant-density discussion and stated smoothness/strong-concavity assumptions in the primary paper.

Previous: Undercoverage Bias Next: Underfitting