How to make CLEAN variants faster? Using clustered components informed by the autocorrelation function

Abstract

Deconvolution, imaging and calibration of data from radio interferometers is a challenging computational (inverse) problem. The upcoming generation of radio telescopes poses significant challenges to existing, and well proven data reduction pipelines due to the large data sizes expected from these experiments, and the high resolution and dynamic range. In this manuscript, we deal with the deconvolution problem. A variety of multiscalar variants to the classical CLEAN algorithm (the de-facto standard) have been proposed in the past, often outperforming CLEAN at the cost of significantly increasing numerical resources. In this work, we aim to combine some of these ideas for a new algorithm, Autocorr-CLEAN, to accelerate the deconvolution and prepare the data reduction pipelines for the data sizes expected by the upcoming generation of instruments. To this end, we propose to use a cluster of CLEAN components fitted to the autocorrelation function of the residual in a subminor loop, to derive continuously changing, and potentially non-radially symmetric, basis functions for CLEANing the residual. Autocorr-CLEAN allows for the superior reconstruction fidelity achieved by modern multiscalar approaches, and their superior convergence speed. It achieves this without utilizing any substep of super-linear complexity in the minor loops, keeping the single minor loop and subminor loop iterations at an execution time comparable to CLEAN. Combining these advantages, Autocorr-CLEAN is found to be up to a magnitude faster than the classical CLEAN procedure. Autocorr-CLEAN fits well in the algorithmic framework common for radio interferometry, making it relatively straightforward to include in future data reduction pipelines. With its accelerated convergence speed, and smaller residual, Autocorr-CLEAN may be an important asset for the data analysis in the future.