Come analyze HEASARC, IRSA, and MAST data in the cloud! The Fornax Initiative is now welcoming all interested beta users.

Choosing a sampler

The following table summarizes the trade-offs between the four samplers.

	MH	GW	HMC	nest
Tuning effort	Moderate	Low	Moderate	Low
Per-sample cost	Cheapest	Low	High	Highest
Effective samples per unit time, smooth posterior	Low to moderate	Moderate	High	Moderate
Robustness to multi-modality	Poor	Poor	Poor	Good
Robustness to correlated parameters	Poor (without good covariance)	Good	Excellent (with gradients)	Good
Returns evidence ?	No	No	No	Yes
Requires gradients?	No	No	Yes	No
Warm-start from MLE?	Helpful	Helpful	Essential	Ignored

A useful decision flow:

Few parameters ( $D \le 5$ ), smooth posterior, single source — start with GW (the default). It is the lowest-friction option for routine work.
Moderate to many parameters ( $D \ge 10$ ), gradients available for the model components in use — HMC. The gain over GW becomes decisive past about .
Multi-modal posterior, or a model-comparison question that needs Bayes factors — nested sampling. Accept the higher elapsed time as the cost of reliability and the bonus of .
Custom proposal needed (e.g. exploiting an exact conditional) or working with a model class that defeats both HMC and nest — MH with a user-defined proposal class through initpackage.