Probabilistic image reconstruction for radio interferometers

Abstract

We present a novel, general-purpose method for deconvolving and denoizing images from gridded radio interferometric visibilities using Bayesian inference based on a Gaussian process model. The method automatically takes into account incomplete coverage of the uv-plane, signal mode coupling due to the primary beam and noise mode coupling due to uv sampling. Our method uses Gibbs sampling to efficiently explore the full posterior distribution of the underlying signal image given the data. We use a set of widely diverse mock images with a realistic interferometer set-up and level of noise to assess the method. Compared to results from a proxy for point source-based clean method we find that in terms of rms error and signal-to-noise ratio our approach performs better than traditional deconvolution techniques, regardless of the structure of the source image in our test suite. Our implementation scales as |$\mathcal {O}(n_{\rm p} \log n_{\rm p})$| provides full statistical and uncertainty information of the reconstructed image, requires no supervision and provides a robust, consistent framework for incorporating noise and parameter marginalizations and foreground removal.