Blog · Collection
Advanced statistical computing
Modern Monte Carlo: Hamiltonian and Langevin methods, adaptive MCMC, tempering, sequential Monte Carlo, normalizing constants, non-reversible samplers, and flows.
This overview surveys the computational methods behind modern Bayesian statistics, drawing on the research literature. Many of them are implemented in the MCMC playground, so you can try the ideas below.
Better proposals: gradients, geometry and adaptation
Random-walk proposals scale badly with dimension, so much of the field is about smarter moves. Langevin methods (MALA) and Hamiltonian Monte Carlo use the gradient of to make long, directed moves. The No-U-Turn Sampler removes HMC’s hardest tuning choice, the path length, by extending each trajectory until it starts to double back. Riemannian manifold methods go further and shape each move to the local curvature of the target, which is what lets them cope with funnel-shaped posteriors. Adaptive MCMC learns a good proposal from the chain’s own history, which remains valid as long as the adaptation fades over time.
Multimodality and tempering
A chain that cannot cross between modes can look converged while missing much of the distribution. Tempering flattens the target to with . Parallel tempering runs chains at several temperatures and swaps their states, and simulated tempering lets a single chain wander up and down the temperature ladder.
Populations of particles
Sequential Monte Carlo moves a population of weighted particles through a sequence of distributions, reweighting, resampling and jittering them at each stage. SMC samplers apply this to a fixed target by tempering from an easy distribution to the posterior, and they estimate the normalizing constant along the way. Particle MCMC and the pseudo-marginal method go one step further: they plug an unbiased estimate of the likelihood into an MCMC algorithm, which is enough to keep the exact target distribution.
Normalizing constants
Comparing models requires the marginal likelihood , which ordinary MCMC does not provide. Bridge sampling, path sampling (thermodynamic integration), stepping-stone sampling and nested sampling all estimate by connecting the prior to the posterior through a family of intermediate distributions.
New directions
Non-reversible, continuous-time samplers such as the Zig-Zag and bouncy particle samplers move in straight lines and change direction at random event times, which avoids the back-and-forth of reversible chains. Function-space methods such as preconditioned Crank–Nicolson stay efficient as the dimension grows without bound. Normalizing flows and diffusion models learn transport maps from simple distributions to complicated ones, and they are increasingly used to build better proposals for samplers.