Header and Body 2

Problem Addressed

Other low-rank matrix approximation algorithms that perform SVD/rank-k SVD on a nxm matrix A (n>>m) require a O(nm2) runtime and O(nm) run space to compute SVD, which can be infeasible requirements for large matrices. The algorithms the Inventors discover do not require saving all existing data in the past. Instead, given a low-rank approximation of existing data, they update approximation on new data without recursively recomputing SVD on the entire dataset, saving storage and computational cost.