Multilevel Hamiltonian Monte Carlo
Also known as: Multilevel HMC, MLHMC, multilevel HMC sampler, multilevel leapfrog MCMC
Multilevel Hamiltonian Monte Carlo (Multilevel HMC) combines the variance-reduction strategy of multilevel Monte Carlo with the efficient gradient-driven exploration of Hamiltonian Monte Carlo. By running coupled HMC chains at increasing levels of model fidelity or discretisation, it achieves accurate posterior estimates at a computational cost substantially lower than a single fine-level HMC chain.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use Multilevel HMC when the target posterior arises from a computationally expensive forward model that can be run at multiple fidelity levels (e.g., fine vs. coarse PDE solvers, deep vs. shallow neural surrogates, fine vs. coarse time-series discretisations), and when a standard HMC run at the finest level is prohibitively slow. It is particularly valuable in spatial statistics, continuum mechanics, and Bayesian inverse problems with PDE constraints. Do not use it when no natural hierarchy of model approximations exists, when the coarse-level model is a poor approximation of the fine level so corrections are large and expensive, or when the posterior is low-dimensional and a single HMC chain is already fast enough.
Strengths & limitations
- Substantially reduces total computational cost relative to a single fine-level HMC chain by concentrating gradient evaluations at cheap coarse levels.
- Inherits HMC's efficient exploration of high-dimensional posteriors through gradient-driven proposals.
- Asymptotically unbiased: the telescoping correction removes all approximation bias introduced by the coarse levels.
- Scales well to posteriors arising from expensive forward models such as PDE solvers or high-resolution simulators.
- Exploits modern automatic differentiation and probabilistic programming frameworks that already implement HMC.
- Requires a meaningful hierarchy of model fidelities; constructing well-coupled chains across levels demands significant modelling effort.
- Level-difference variance must be small enough that coarse corrections provide genuine savings; poorly coupled chains can eliminate the cost advantage.
- Implementation is substantially more complex than standard HMC, requiring synchronised random-number streams and careful tuning of step sizes and trajectory lengths at each level.
- The optimal number of levels and sample allocation across levels must be determined, often requiring a pilot run.
- Gradient computations are still needed at the fine level for correction chains, so the method does not eliminate fine-level cost entirely.
Frequently asked
How does Multilevel HMC differ from standard HMC?
Standard HMC runs a single Markov chain at the target model's full fidelity, evaluating the gradient of the log-posterior at each leapfrog step. Multilevel HMC runs coupled chains across multiple fidelity levels and combines their estimates via a telescoping sum, so the expensive fine-level gradient is evaluated far fewer times — this is where the cost saving comes from.
What does 'coupling' mean in this context, and why does it matter?
Coupling means the chains at two consecutive levels share the same stream of random numbers so their sample paths are as similar as possible. When the chains are tightly coupled, the variance of the level-difference correction is small, which is precisely what makes the multilevel estimator cheap. Poor coupling eliminates the variance reduction and with it the computational advantage.
Is a natural level hierarchy always available?
No, and this is the method's main practical limitation. The hierarchy must arise naturally from the model structure — mesh refinement, quadrature order, time-step size, or model complexity. If the forward model cannot be run meaningfully at a coarser resolution, Multilevel HMC is not applicable and standard HMC should be used instead.
How do I choose the number of levels and samples per level?
A pilot run is typically used to estimate the variance and cost at each level. The optimal sample allocation minimises total cost for a given mean-squared-error target; closed-form formulas analogous to those in Giles (2008) apply when cost and variance scale polynomially with level.
Can Multilevel HMC be implemented in standard probabilistic programming languages?
Partially. Frameworks such as Stan or NumPyro provide HMC, but the multilevel coupling and telescoping infrastructure must usually be coded manually. Research implementations exist in Python and Julia for specific model classes, but no general-purpose PPL supports Multilevel HMC out of the box as of 2026.
Sources
- Beskos, A., Jasra, A., Law, K., Tempone, R., & Zhou, Y. (2017). Multilevel sequential Monte Carlo samplers. Stochastic Processes and their Applications, 127(5), 1417–1440. DOI: 10.1016/j.spa.2016.08.004 ↗
- Neal, R. M. (2011). MCMC using Hamiltonian dynamics. In S. Brooks, A. Gelman, G. Jones, & X.-L. Meng (Eds.), Handbook of Markov Chain Monte Carlo (pp. 113–162). CRC Press. link ↗
How to cite this page
ScholarGate. (2026, June 3). Multilevel Hamiltonian Monte Carlo. ScholarGate. https://scholargate.app/en/bayesian/multilevel-hamiltonian-monte-carlo
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Hamiltonian Monte CarloBayesian↔ compare
- Hierarchical Hamiltonian Monte CarloBayesian↔ compare
- Markov Chain Monte CarloSimulation↔ compare
- Multilevel MCMCBayesian↔ compare
- Multilevel Variational InferenceBayesian↔ compare