A randomized algorithm for tensor singular value decomposition using an arbitrary number of passes (Q6143819)

This paper deals with image restoration when the transmitted or received pixel data has large sections with lost, damaged or unreadable pixel contents. This problem is approached as a tubular tensor singular value problem and a new algorithm is proposed that succeeds with fewer randomized passes than current methods. It experiments with 70 or 80 \% of pixel information loss images and shows its efficiency there. The paper also develops theoretical error bounds for the tensor SVD and the tensor QR decomposition using the Frobenius norm throughout. Its theoretical high speed gains are verified in several experiments. The observed 2--7 fold speed-up of the proposed algorithm is based on its use of lower tubal rank approximations than were previously needed. The authors also offer heuristic advice on how to find an optimal tubal rank for any given restoration problem. The bibliography of 45 works is extensive but unfortunately almost completely useless -- by listing the occurrence of each quoted work in the order of its first appearance in the text. A year-of-appearance order should have been used instead.