Using multiobjective optimization to reconstruct interferometric data (I)

Imaging in radioastronomy is an ill-posed inverse problem. Particularly the Event Horizon Telescope (EHT) Collaboration investigated the fidelity of their image reconstructions convincingly by large surveys solving the problem with different optimization parameters. This strategy faces a limitation for the existing methods when imaging the active galactic nuclei (AGN): large and expensive surveys solving the problem with different optimization parameters are time-consumptive. 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. To this end we used a multiobjective version of the genetic algorithm (GA): the Multiobjective Evolutionary Algorithm Based on Decomposition, or MOEA/D. GA strategies explore the objective function by evolutionary operations to find the different local minima, and to avoid getting trapped in saddle points. First, we have tested our algorithm (MOEA/D) using synthetic data based on the 2017 Event Horizon Telescope (EHT) array and a possible EHT + next-generation EHT (ngEHT) configuration. We successfully recover a fully evolved Pareto front of non-dominated 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 (M87) with the EHT data in 2017. MOEA/D is very flexible, faster than any other Bayesian method and explores more solutions than Regularized Maximum Likelihood methods (RML).
Comment: to appear in A&A