Using multiobjective optimization to reconstruct interferometric data. Part I

Abstract

Context. Imaging in radioastronomy is an ill-posed inverse problem. However, with increasing sensitivity and capabilities of telescopes, several strategies have been developed in order to solve this challenging problem. In particular, novel algorithms have recently been proposed using (constrained) nonlinear optimization and Bayesian inference. Aims. The Event Horizon Telescope (EHT) Collaboration convincingly investigated the fidelity of their image reconstructions with large surveys, solving the image reconstruction problem with different optimization parameters. This strategy faces a limitation for the existing methods when imaging active galactic nuclei: Large and expensive surveys solving the problem with different optimization parameters are time-consuming. We present a novel nonconvex, multiobjective optimization modeling approach that gives a different type of claim and may provide a pathway to overcome this limitation. Methods. To this end, we use a multiobjective version of the genetic algorithm (GA): the Multiobjective Evolutionary Algorithm Based on Decomposition, or MOEA/D. The GA strategies explore the objective function by evolutionary operations to find the different local minima and to avoid becoming trapped in saddle points. Results. First, we tested our algorithm (MOEA/D) using synthetic data based on the 2017 EHT array and a possible EHT plus next-generation EHT configuration. We successfully recover a fully evolved Pareto front of nondominated solutions for these examples. The Pareto front divides into clusters of image morphologies representing the full set of locally optimal solutions. We discuss approaches to find the most natural guess among these solutions and demonstrate its performance on synthetic data. Finally, we apply MOEA/D to observations of the black hole shadow in Messier 87 with the EHT data in 2017. Conclusions. The MOEA/D is very flexible and faster than any other Bayesian method, and it explores more solutions than regularized maximum likelihood methods. We have written two papers to present this new algorithm. In the first, we explain the basic idea behind multiobjective optimization and MOEA/D, and we use MOEA/D to recover static images. In the second paper, we extend the algorithm to allow dynamic and (static and dynamic) polarimetric reconstructions.