Imaging and uncertainty quantification in radio astronomy via convex optimization : when precision meets scalability
Abstract
Upcoming radio telescopes such as the Square Kilometre Array (SKA) will provide sheer amounts
of data, allowing large images of the sky to be reconstructed at an unprecedented resolution and
sensitivity over thousands of frequency channels. In this regard, wideband radio-interferometric
imaging consists in recovering a 3D image of the sky from incomplete and noisy Fourier data, that
is a highly ill-posed inverse problem. To regularize the inverse problem, advanced prior image
models need to be tailored. Moreover, the underlying algorithms should be highly parallelized to
scale with the vast data volumes provided and the Petabyte image cubes to be reconstructed for
SKA. The research developed in this thesis leverages convex optimization techniques to achieve
precise and scalable imaging for wideband radio interferometry and further assess the degree of
confidence in particular 3D structures present in the reconstructed cube.
In the context of image reconstruction, we propose a new approach that decomposes the image
cube into regular spatio-spectral facets, each is associated with a sophisticated hybrid prior image
model. The approach is formulated as an optimization problem with a multitude of facet-based
regularization terms and block-specific data-fidelity terms. The underpinning algorithmic structure benefits from well-established convergence guarantees and exhibits interesting functionalities
such as preconditioning to accelerate the convergence speed. Furthermore, it allows for parallel processing of all data blocks and image facets over a multiplicity of CPU cores, allowing the
bottleneck induced by the size of the image and data cubes to be efficiently addressed via parallelization. The precision and scalability potential of the proposed approach are confirmed through
the reconstruction of a 15 GB image cube of the Cyg A radio galaxy.
In addition, we propose a new method that enables analyzing the degree of confidence in
particular 3D structures appearing in the reconstructed cube. This analysis is crucial due to the
high ill-posedness of the inverse problem. Besides, it can help in making scientific decisions on
the structures under scrutiny (e.g., confirming the existence of a second black hole in the Cyg A
galaxy). The proposed method is posed as an optimization problem and solved efficiently with
a modern convex optimization algorithm with preconditioning and splitting functionalities. The
simulation results showcase the potential of the proposed method to scale to big data regimes.