[RNG-156] Dirichlet distribution sampler

• Type: New Feature
• Status: Closed
• Priority: Minor
• Resolution: Implemented
• Affects Version/s: 1.4
• Fix Version/s: 1.4
• Component/s: sampling
• Assignee: Unassigned
• Reporter: Alex Herbert
• Created: 23/Jul/21 13:55
• Updated: 13/Sep/21 10:41
• Resolved: 25/Jul/21 09:12

Description

Add sampling from a Dirichlet distribution.

Samples can be produced by generating Gamma(shape=alpha, scale=1) deviates for each category with concentration alpha, summing the deviates and dividing by the sum.

Proposed public API.

public abstract class DirichletSampler
    implements SharedStateObjectSampler<double[]> {

    public double[] sample();
    public DirichletSampler withUniformRandomProvider(UniformRandomProvider rng);

    public static DirichletSampler of(UniformRandomProvider rng,
                                      double... alpha);
    public static DirichletSampler symmetric(UniformRandomProvider rng,
                                             int k,
                                             double alpha);
}

The symmetric case is a special case where the alpha concentration parameters are the same for all categories. This can be implemented with a single sampler which reduces memory footprint. This is especially true when alpha > 1 as a AhrensDieterMarsagliaTsangGammaSampler holds two precomputed values and a normalised Gaussian sampler.

An example of using a symmetric case is a unit Simplex which uses dimension k and alpha=1.

 

Issue Links

• relates to: RNG-132 Package rng.sampling.shape for sampling coordinates from shapes