More on low rank approximation of a matrix

Abstract: We call a matrix algorithm superfast (aka running at sublinear cost) if it involves much fewer flops and memory cells than the matrix has entries. Using such algorithms is highly desired or even imperative in computations for Big Data, which involve immense matrices and are quite typically reduced to solving linear least squares problem and/or computation of low rank approximation of an input matrix. The known algorithms for these problems are not superfast, but we propose a novel ACA-like superfast randomized iterative refinement of a crude LRA. Like ACA iterations we output CUR LRA, which is a particular memory efficient form of LRA. Unlike ACA method we use sampling probability to reduce our task to recursive solution of generalized Linear Least Squares Problem (generalized LLSP) and to prove monotone convergence to a close LRA with a high probability under mild assumptions on an initial LRA. For crude initial LRA lying reasonably far from optimal we have consistently observed significant improvement in two or three iterations in our numerical tests.