Bayesian Multiscale Methods for Poisson Count Data

摘要

We present an overview of recent work on a flexible framework for multiscale modeling of Poisson count data, such as is encountered regularly in the field of high-energy astrophysics, that allows for intuitive, easily interpretable, computationally efficient implementations of Bayesian inference for standard tasks like smoothing, deconvolution, and segmentation. At the foundation of this approach is a multiscale factorization of the Poisson likelihood, which can be viewed formally as deriving from a blending of concepts from the literatures on wavelets, recursive partitioning, and graphical models.