Parallel faceted imaging in radio interferometry via proximal splitting (Faceted HyperSARA)

Wideband radio-interferometric (RI) imaging consists in estimating images of the sky across a whole frequency band from incomplete Fourier data. In practice, powerful prior information is needed to regularize the associated inverse imaging problem. At the extreme resolution and dynamic range of interest to modern telescopes, image cubes will far exceed Terabyte sizes, with data volumes orders of magnitude larger, making image estimation a challenging task. The computational cost and memory requirements of corresponding iterative image recovery algorithms are extreme and call for high parallelism. To address this issue, a data-splitting strategy was recently introduced to parallelize computations over data blocks within a primal-dual algorithm (HyperSARA), showing a significantly improved image reconstruction quality compared to the state-of-the-art CLEAN algorithm. Building on a similar algorithmic structure, we further extend the splitting functionality to decompose the image cube into spatio-spectral facets with their own prior. All these terms can be handled in parallel, thereby reducing the computational bottleneck of the original algorithm from full image cube to facet size. Simulation results on synthetic image cubes confirm that faceting can provide a major increase in parallelization capability when compared to the original HyperSARA approach, at a negligible cost in imaging quality. A proof-of-concept reconstruction of a 15 GB image of Cygnus A from 7.4 GB of JVLA data further illustrates the reconstruction performance of the proposed approach.