Robust low-rank data matrix approximations