Sampling

Starlord uses thin wrapper classes around samplers provided by other libraries. This is to allow Starlord to automatically initialize and run the codes while still flexibly passing settings to the samplers themselves. Settings can be passed to the samplers in the [sampling] section of the TOML files. [sampler_name_init].[option] entries are passed to the sampler initializer, and [sampler_name]_run.[option] entries are passed to the run command. If you are calling Starlord from within Python, you have the additional option to retrieve and use the underlying sampler sampler, if you wish.

The [sampling] is also where you may set the values of any constants you used during model specification. The syntax is just const.constant_name = value, where value is a float. The Fitting Stars page has an example of this. However, note that constants may also be set at the command line via the repeatable argument -s constant_name=value; if a constant is specified in both places, the CLI argument gets priority.

This page lists some common sampling parameters you may wish to set; however, it is not an exhaustive list of the sampler options.

Emcee Sampler

The Emcee sampler is a popular MCMC sampler in astronomy which uses a collection of walkers to propose jumps that are invariant to affine transformations of the distribution being sampled. It’s a good choice so long as your posterior isn’t too complex (donuts, bananas, holes, etc).

Initialization Parameters

nwalkers:

The number of walkers, which must be at least twice the number of dimensions; more complex posteriors can benefit from higher walker counts.

Run Parameters

initial_state:

The initial walker states – if not set Starlord will draw this from the prior PPF.

nsteps:

The number of steps per walker to take during the run.

progress:

Whether to print the progress of the sampler as it runs.

Dynesty Sampler

The Dynesty sampler is an alternative sampler that can handle more complex posteriors. Technically it is not an MCMC algorithm. This version is the static sampler; a Starlord wrapper the dynamic sampler is not yet implemented.

Initialization Parameters

nlive:

The number of live points, larger values give larger posterior samples and better convergence, but take longer.

bound:

The method used to approximately bound the prior based on the live points: ‘none’, ‘single’, ‘multi’, ‘balls’, or ‘cubes’. See the Dynesty documentation for more information.

sample:

The method used to sample within the likelihood constraint: ‘auto’, ‘unif’, ‘rwalk’, ‘slice’, or ‘rslice’. See the Dynesty documentation for more information.

update_interval:

How often to update the posterior distribution estimate, which can be useful for very fast posterior evaluations.

Run Parameters

maxiter_init:

The maximum number of iterations to use during the run (default is no limit).

dlogz_init:

Determines the estimated uncertainty at which the sampler will complete.

n_effective:

The number of effective posterior samples at which the sampler will complete.

print_progress:

Whether to print the progress of the sampler as it runs.