Optimal Arrays for Compressed Sensing in Snapshot-Mode Radio Interferometry
Radio interferometry has always faced the problem of incomplete sampling of the Fourier plane. A possible remedy can be found in the promising new theory of compressed sensing (CS), which allows for the accurate recovery of sparse signals from sub-Nyquist sampling given certain measurement conditions. We provide an introductory assessment of optimal arrays for CS in snapshot-mode radio interferometry, using orthogonal matching pursuit (OMP), a widely used CS recovery algorithm similar in some respects to CLEAN. We focus on centrally condensed (specifically, Gaussian) arrays versus uniform arrays, and the principle of randomization versus deterministic arrays such as the VLA. The theory of CS is grounded in $a)$ sparse representation of signals and $b)$ measurement matrices of low coherence. We calculate a related quantity, mutual coherence (MC), as a theoretical indicator of arrays' suitability for OMP based on the recovery error bounds in (Donoho et al. 2006). OMP reconstructions of both point and extended objects are also run from simulated incomplete data. Optimal arrays are considered for object recovery through 1) the natural pixel representation and 2) the representation by the block discrete cosine transform (BDCT). We find that reconstructions of the pixel representation perform best with the uniform random array, while reconstructions of the BDCT representation perform best with normal random arrays. Slight randomization to the VLA also improves it hugely for CS with the pixel basis. In the pixel basis, array design for CS reflects known principles of array design for small numbers of antennas, namely of randomness and uniform distribution. Differing results with the BDCT, however, emphasize the importance of studying how sparsifying bases affect array design before CS can be optimized.
