Bayesian Blocks in Two or More Dimensions: Image Segmentation and Cluster Analysis

TL;DR

This work generalizes Bayesian Blocks to higher-dimensional data by partitioning the data space into regions described by Poisson counts, with the region posterior $\Phi(N,V)$ depending only on the region's count $N$ and volume $V$ and with $\lambda_n$ marginalized. It initializes from a Voronoi tessellation and greedily merges blocks using the Bayes merge factor, $P(\text{Merge})$, to produce adaptive, shape-agnostic blocks and potential clusters, effectively performing joint segmentation and density estimation in multiple dimensions. The approach provides dimension-agnostic, nonparametric structure discovery suitable for image-like data and high-dimensional catalogs, avoiding binning of raw data and accommodating slowly varying backgrounds. It also frames inference in an information-theoretic light, discussing priors, model averaging, and the interpretation of results as information about the signal.

Abstract

This paper describes an extension, to higher dimensions, of the Bayesian Blocks algorithm for estimating signals in noisy time series data (Scargle 1998, 2000). The mathematical problem is to find the partition of the data space with the maximum posterior probability for a model consisting of a homogeneous Poisson process for each partition element. For model M_{n}, attributing the data within region n of the data space to a Poisson process with a fixed event rate lambda_{n}, the global posterior is: P(M_{n}) = Phi(N,V) = Gamma(N+1)Gamma(V-N+1) / Gamma(V+2) = N!(V-N)!/(V+1)! . Note that lambda_{n} does not appear, since it has been marginalized, using a flat, improper prior. Other priors yield similar formulas. This expression is valid for a data space of any dimension. It depends on only N, the number of data points within the region, and V, the volume of the region. No information about the actual locations of the points enters this expression. Suppose two such regions, described by N_{1},V_{1} and N_{2},V_{2}, are candidates for being merged into one. From the above equation, construct a Bayes merge factor, giving the ratio of posteriors for the two regions merged and not merged, respectively: P(Merge) = Phi(N_{1}+N_{2},V_{1}+V_{2}) / Phi(N_{1},V_{1}) Phi(N_{2},V_{2}) . Then collect data points into blocks with a greedy cell coalescence algorithm.