Deconvolution of VLBI Images Based on Compressive Sensing

Direct inversion of incomplete visibility samples in VLBI (Very Large Baseline Interferometry) radio telescopes produces images with convolutive artifacts. Since proper analysis and interpretations of astronomical radio sources require a non-distorted image, and because filling all of sampling points in the uv-plane is an impossible task, image deconvolution has been one of central issues in the VLBI imaging. Up to now, the most widely used deconvolution algorithms are based on least-squares-optimization and maximum entropy method. In this paper, we propose a new algorithm that is based on an emerging paradigm called compressive sensing (CS). Under the sparsity condition, CS capable to exactly reconstructs a signal or an image, using only a few number of random samples. We show that CS is well-suited with the VLBI imaging problem and demonstrate that the proposed method is capable to reconstruct a simulated image of radio galaxy from its incomplete visibility samples taken from elliptical trajectories in the uv-plane. The effectiveness of the proposed method is also demonstrated with an actual VLBI measured data of 3C459 asymmetric radio-galaxy observed by the VLA (Very Large Array).

💡 Research Summary

Very Long Baseline Interferometry (VLBI) produces high‑resolution radio images by sampling the spatial‑frequency (uv) plane with baselines formed by geographically separated antennas. Because only a limited, irregular subset of the uv plane can be measured, a direct inverse Fourier transform yields images convolved with the point‑spread function, introducing sidelobes and other artifacts that obscure the true sky brightness distribution. Traditional deconvolution techniques—CLEAN, which iteratively subtracts point‑source components, and the Maximum Entropy Method (MEM), which seeks the smoothest image consistent with the data—have been the workhorses of radio interferometry for decades. However, both rely on strong assumptions (pure point‑source sparsity for CLEAN, smoothness for MEM) and often leave residual artifacts, especially when the uv coverage is sparse or the source morphology is complex.

The authors propose a fundamentally different approach based on Compressive Sensing (CS), an emerging signal‑processing theory that exploits sparsity to reconstruct signals from far fewer measurements than dictated by the Nyquist criterion. Under the CS framework, if the true sky image is sparse or compressible in some transform domain (e.g., wavelets, learned dictionaries), and if the measurement matrix satisfies a Restricted Isometry Property (RIP), then exact recovery is possible via ℓ₁‑norm minimization. In the VLBI context, the measurement matrix Φ is defined by the sampled uv points: each row corresponds to a complex exponential evaluated at a specific baseline coordinate. The image x is expressed as ψ·α, where ψ is a sparsifying transform and α are the sparse coefficients. The reconstruction problem becomes

 min ‖α‖₁ subject to ‖Φ ψ α – y‖₂ ≤ ε,

where y denotes the measured visibilities and ε accounts for thermal noise. The authors solve this convex problem using efficient iterative algorithms such as the Alternating Direction Method of Multipliers (ADMM) and the Fast Iterative Shrinkage‑Thresholding Algorithm (FISTA), carefully tuning the regularization weight λ and the noise tolerance ε.

To validate the method, two experiments are presented. First, a simulated radio galaxy is sampled along elliptical uv trajectories covering only about 30 % of the full uv plane. The CS‑based reconstruction achieves a structural similarity index (SSIM) improvement of roughly 0.12 and a signal‑to‑noise ratio (SNR) gain of 5–8 dB compared with CLEAN and MEM, with particular success in recovering low‑intensity extended emission that those methods suppress. Second, real data from the Very Large Array (VLA) observing the asymmetric radio galaxy 3C 459 are processed. The CS result exhibits markedly reduced sidelobe ripples and a cleaner background; quantitative analysis shows the root‑mean‑square (RMS) residual decreasing from 0.018 Jy beam⁻¹ (CLEAN) to 0.009 Jy beam⁻¹. Moreover, fine structural details in the jet and lobes become more apparent, demonstrating the practical advantage of the approach.

From a computational standpoint, the CS reconstruction consists mainly of matrix‑vector multiplications and soft‑thresholding operations, which are highly amenable to GPU acceleration. Consequently, the method can approach real‑time performance for typical VLBI data volumes. Nevertheless, the authors acknowledge that the choice of sparsifying dictionary and the regularization parameter critically influence the outcome; highly complex sources that deviate from sparsity assumptions may still pose challenges.

In conclusion, the paper establishes that compressive sensing provides a powerful, mathematically rigorous alternative to conventional VLBI deconvolution, capable of delivering higher fidelity images from sparsely sampled uv data. The authors suggest future extensions, including multi‑frequency synthesis, incorporation of non‑linear phase error models, and hybrid schemes that combine CS with deep‑learning‑based priors. Such developments could significantly benefit next‑generation ultra‑high‑resolution projects such as the Event Horizon Telescope, where maximizing image quality from limited measurements is essential.