BRISK: An accelerated wide-field imager for radio astronomy

Abstract

Wide-field interferometric imaging with non-coplanar arrays is computationally expensive, particularly as we move towards imaging large datasets from next-generation telescopes such as the Square Kilometre Array (SKA). The dominant imaging cost arises during the gridding and degridding steps, where large $w$-terms require correction. In $w$-projection, this correction uses $w$-dependent convolution kernels whose support grows rapidly with field of view (FoV) and maximum $|w|$. In $w$-stacking, visibilities are partitioned into discrete $w$-layers, each requiring a 2D Fourier transform and image-domain phase correction; the required number of layers increases with $|w|_{\max}$ and FoV. We present BRISK (Baseline-aware Rank-compressed Interferometric Sky-imaging Toolkit), a wide-field interferometric imaging algorithm that accelerates $w$-projection by exploiting the low-rank structure of its convolution kernels. For each $w$-plane, the full 2D gridding kernel is compressed using a truncated randomised singular value decomposition (SVD), yielding a compact representation from which the SVD-compressed kernel can be reconstructed on the fly while preserving the dominant kernel energy. We benchmark memory usage, runtime, and accuracy against CASA's $w$-projection and WSClean's $w$-stacking, $w$-gridding and image-domain gridding algorithms on a simulated $15 \times 15$ grid of point sources, as well as the SKA1-Low GLEAM simulated measurement sets. BRISK runs up to 2443 times faster than CASA, and up to 132 times faster than WSClean, while maintaining a peak-flux bias of less than 1% at the largest image sizes and comparable peak memory usage.