Rylan Schaeffer

Kernel Papers

Complete Random Measures

Disclaimer: Most of this comes from Tamara Broderick’s excellent paper “Posteriors, conjugacy, and exponential families for completely random measures.”

Review of BNP Models

Bayesian nonparametric (BNP) models revolve around collections of pairs of (traits, frequencies/rates). The principle challenge of Bayesian nonparametrics is how, starting with a countable infinity of traits and frequencies in the prior, to integrate over the infinite possibilities to compute a finite posterior over traits and frequencies based on data. More specifically, we have traits \(\{\psi_k \in \Psi\}\) and frequencies or rates \(\theta_k\). A BNP model starts with a discrete measure on \(\Psi\):

\[\Theta := \sum_{k=1}^K \theta_k \delta_{\psi_k}\]

where \(K\) can be finite or countably infinite. The \(n\)th datum \(X_n$ is another discrete measure on\)\Psi$$:

\[X_n := \sum_{k=1}^{K_n} x_{n,k} \delta_{\psi_{n,k}}\]

where \(x_{n,k} \in \mathbb{R}_+\) is the degree to which the \(n\)th datum possesses the trait \(\psi_{n,k}\). Each \(\psi_{n,k} \in \{\psi_k \}\) but different data can possess different traits.

Using a BNP model requires specifying a prior distribution \(p(\Theta)\) and a likelihood \(p(X_n|\Theta)\).

Random Measures

A random measure is a random element whose values are measures. More formally, let \(\Sigma_{\Psi}\) be the sigma-algebra of some space \(\Psi\). For a measure \(\Theta\) over \(\Psi\) to be random, for any measurable set \(A \in \Sigma_{\Psi}\), the quantity \(\Theta(A)\) must be a random variable.

Completely Random Measures

A completely random measure (CRM) is a random measure that satisfies 1 additional property: for any disjoint, measurable sets \(A_1, ..., A_k \in \Sigma_{\Psi}\), the random variables \(\Theta(A_1), ..., \Theta(A_k)\) are independent.


Kingman 1967 shows that CRMs can always be split into 3 measures:

\[\Theta = \Theta_{det} + \Theta_{fix} + \Theta_{ord}\]

Each measure is explained in more detail below:

Deterministic Component Measure

\(\Theta_{det}\) is a deterministic measure.

Fixed Locations Measure

\(\Theta_{fix}\) is the “fixed locations” measure.

\[\Theta_{fix} = \sum_{k=1}^{K_{fix}} \theta_{fix, k} \delta_{\psi_{fix}, k}\]

where \(\theta_{fix,k} \in \mathbb{R}_{\geq 0}\) are random weights and \(\delta_{\psi_{fix}, k}\) are fixed locations. Note that, by the independence property of CRMs, the \(\theta_{fix,k}\) must be independent random variables.

Ordinary Measure

\(\Theta_{ord}\) is the “ordinary” measure. Explaining this requires some familiarity with Poisson point processes. To generate an ordinary component, start with a Poisson point process on \(\mathbb{R}_{\geq 0} \times \Psi\) characterized by some rate measure \(\mu(d\Theta \times d\Psi)\). The ordinary component is

\[\Theta_{ord} = \sum_{k=1}^{K_{ord}} \theta_{ord, k} \delta_{\psi_{ord, k}}\]