Variational inference-based polychromatic imaging for optical interferometry

Abstract

Image reconstruction in optical interferometry remains a fundamentally ill-posed inverse problem due to the typical sparsity of its Fourier data. In this work, we present an image reconstruction algorithm for optical interferometric data from arrays such as CHARA and the VLTI, built on a Bayesian framework rooted in Information Field Theory. Our approach introduces a compactness-enforcing prior model. This prior is flexible and general-purpose, capable of modeling compact astrophysical sources such as Gaussian components, ring-like structures, and sparse multicomponent fields. It is also suitable for young stellar object fields where compact central emission and surrounding structures may both be present. Additionally, the prior supports polychromatic imaging by handling all wavelength channels simultaneously. The prior combines three key elements: (i) a physically motivated limb-darkened radial weight that defines compact support and suppresses artifacts outside the source region, (ii) a smoothed log-intensity field that enforces spatial smoothness, and (iii) a wavelength-dependent component that allows chromatic variations across spectral channels. We integrate this prior model with a forward operator that models complex visibilities, squared visibilities, and bispectra. Reconstruction is performed using Metric Gaussian Variational Inference (MGVI) implemented in the Numerical Information Field Theory (NIFTy) environment. We evaluate the method on synthetic measurements generated from simulated Gaussian and ring-like source models. We further demonstrate its performance on real observations by reconstructing the interacting binary β Lyrae and the Herbig Be star HD 190073. Finally, we show that the same prior model can recover sparse fields by reconstructing the cluster of point-like sources from the 2016 Interferometric Imaging Beauty Contest data.