Robust Hamiltonian Monte Carlo
Also known as: Robust HMC, heavy-tailed HMC, geometric-ergodic HMC, outlier-robust HMC
Robust Hamiltonian Monte Carlo (Robust HMC) is a family of extensions to standard HMC designed to maintain geometric ergodicity and sampling efficiency when the posterior has heavy tails, strong curvature variation, or near-degenerate geometry. By modifying the kinetic energy, mass matrix, or proposal mechanism, these methods ensure reliable exploration of difficult posteriors that defeat the standard NUTS/HMC sampler.
Key highlights
- Maintains geometric ergodicity and valid inference even for heavy-tailed or geometrically difficult posteriors where standard HMC fails.
- Gradient-based proposals retain high acceptance rates and low autocorrelation, outperforming random-walk Metropolis on continuous parameters.
- Divergent-transition monitoring provides an automatic diagnostic: zero divergences after warm-up is a strong signal of reliable sampling.
- Riemannian and adaptive mass-matrix variants are scale-invariant and require less manual prior rescaling.
- Modular: robust kinetic energy or proposal modifications can be layered onto standard HMC/NUTS without changing the overall MCMC workflow.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use Robust HMC when standard NUTS reports divergent transitions, very low effective sample sizes, or high R-hat values despite extended warm-up — signals that the posterior geometry is challenging. It is particularly valuable for posteriors with heavy tails (e.g. Student-t likelihoods, hierarchical models with weakly identified variance components) or sharp curvature changes. Avoid it when a well-specified standard NUTS run converges cleanly, because the additional complexity — heavier computational cost per step and more tuning parameters — offers no benefit for well-behaved posteriors.
Strengths & limitations
- Maintains geometric ergodicity and valid inference even for heavy-tailed or geometrically difficult posteriors where standard HMC fails.
- Gradient-based proposals retain high acceptance rates and low autocorrelation, outperforming random-walk Metropolis on continuous parameters.
- Divergent-transition monitoring provides an automatic diagnostic: zero divergences after warm-up is a strong signal of reliable sampling.
- Riemannian and adaptive mass-matrix variants are scale-invariant and require less manual prior rescaling.
- Modular: robust kinetic energy or proposal modifications can be layered onto standard HMC/NUTS without changing the overall MCMC workflow.
- Higher computational cost per iteration than standard HMC, especially with a dense or Riemannian mass matrix.
- More tuning parameters (step size, mass matrix type, kinetic energy family) that interact in non-obvious ways.
- Riemannian HMC requires computing the Fisher metric and its derivatives, which may be expensive or unstable for some models.
- When the posterior is genuinely non-identifiable, no robust sampler can rescue the inference — the model must be reparameterised.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
What is a divergent transition and why does it matter?
A divergent transition occurs when the leapfrog integrator loses energy conservation — typically because the step size is too large relative to the posterior curvature. Divergences indicate that the sampler is not exploring that region correctly, so parameter estimates near divergent regions are biased. In Stan, the default recommendation is to obtain zero divergent transitions after warm-up before trusting the posterior.
How is Robust HMC different from standard NUTS?
NUTS is an adaptive HMC that automatically sets the trajectory length, but it uses a standard Gaussian kinetic energy and a diagonal or dense mass matrix. Robust HMC modifies the kinetic energy (e.g. to Student-t) or the proposal mechanism (e.g. Barker) to maintain ergodicity for posteriors that violate the light-tail assumptions implicit in standard HMC.
Does reparameterisation help before trying Robust HMC?
Yes. Non-centred parameterisations (e.g. replacing a hierarchical parameter mu + sigma*z for the raw z) often resolve funnel geometry without requiring a more complex sampler. Robust HMC is most useful when reparameterisation is not straightforward or insufficient.
How do I know if Robust HMC has converged?
Check R-hat values below 1.01 for all parameters, effective sample size (ESS) above 100 per chain (or 400 total), zero divergent transitions after warm-up, and stable trace plots across chains. For Riemannian HMC also inspect acceptance rates, which should be near the target (typically 0.65–0.90).
Which software implements Robust HMC?
Stan (via NUTS with step-size adaptation) covers most practical cases. The Barker proposal is available in research packages such as MCMCglmm and custom implementations. Riemannian HMC is available in PyMC and custom Stan models via user-defined metrics. NumPyro and BlackJAX also support custom kinetic energies.
Sources
- 1.Livingstone, S. & Zanella, G. (2022). The Barker proposal: combining robustness and efficiency in gradient-based MCMC. Journal of the Royal Statistical Society: Series B, 84(2), 496–523.
- 2.Betancourt, M. (2017). A conceptual introduction to Hamiltonian Monte Carlo. arXiv preprint arXiv:1701.02434.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Hamiltonian Monte Carlo. ScholarGate. https://scholargate.app/bayesian/robust-hamiltonian-monte-carlo