Stick-Breaking and Random Measures

Stick-breaking gives an explicit recipe for constructing the random discrete measures that underlie Bayesian nonparametric priors, making them simulable and computable.

Definition

A stick-breaking construction builds a random discrete probability measure by successively breaking off fractions of a unit-length stick to form the weights and assigning each weight a location drawn from a base measure, providing an explicit representation of nonparametric priors such as the Dirichlet process.

Scope

This topic covers Sethuraman's stick-breaking construction of the Dirichlet process, the resulting weight distribution, generalizations such as the Pitman-Yor process and other stick-breaking priors, completely random measures, and the truncated and slice-sampling algorithms these representations enable.

Core questions

• How does stick-breaking construct the weights of a Dirichlet process?
• How do Pitman-Yor and other stick-breaking priors generalize the construction?
• What are completely random measures and how do they generate nonparametric priors?
• How do truncation and slice sampling exploit these representations for inference?

Key concepts

• stick-breaking construction
• GEM distribution
• Pitman-Yor process
• completely random measure
• truncation
• slice sampling
• atoms and weights

Key theories

Stick-breaking representation

Sethuraman showed the Dirichlet process can be written as an infinite weighted sum of point masses, with weights formed by independent Beta-distributed stick-breaks, making the prior explicit and simulable.

Stick-breaking inference

Truncated and slice-sampling Gibbs methods built on the stick-breaking form give general algorithms for posterior inference under broad classes of stick-breaking priors.

Clinical relevance

Stick-breaking representations underpin practical algorithms for fitting nonparametric mixture and clustering models, enabling their use in genomics, topic modeling, and other large-scale applications.

History

Sethuraman's 1994 stick-breaking construction gave the Dirichlet process an explicit, computable form. Ishwaran and James's 2001 sampling methods and the Pitman-Yor generalization extended this to a broad family of stick-breaking priors central to modern nonparametric Bayesian computation.

Key figures

• Jayaram Sethuraman
• Hemant Ishwaran
• Lancelot James
• Jim Pitman

Seminal works

• sethuraman1994
• ishwaran2001

Frequently asked questions

Why is the stick-breaking construction useful?

It turns an abstract prior over distributions into an explicit, simulable sum of weighted point masses, which makes it possible to draw from the prior and to design Gibbs and slice samplers for posterior inference.