LazySVD: Even Faster SVD Decomposition Yet Without Agonizing Pain

We study k-SVD that is to obtain the first k singular vectors of a matrix A. Recently, a few breakthroughs have been discovered on k-SVD: Musco and Musco [1] proved the first gap-free convergence result using the block Krylov method, Shamir [2] discovered the first variance-reduction stochastic method, and Bhojanapalli et al. [3] provided the fastest O(nnz(A) + poly(1/\varepsilon))-time algorithm using alternating minimization. In this paper, we put forward a new and simple LazySVD framework to improve the above breakthroughs. This framework leads to a faster gap-free method outperforming [1], and the first accelerated and stochastic method outperforming [2]. In the O(nnz(A) + poly(1/\varepsilon)) running-time regime, LazySVD outperforms [3] in certain parameter regimes without even using alternating minimization.