Lawrence Berkeley National Laboratory

We present a detailed study of truncated singular value decomposition (SVD) for column-partitioned matrices. In particular, we analyze the elation between the truncated SVD of a matrix and the truncated SVDs of its submatrices. We give necessary and sufficient conditions under which truncated SVD of a matrix can be constructed from that of its submatrices. We also present perturbation analysis to show that an approximate truncated SVD can still be computed even if the given necessary and sufficient conditions are only approximately satisfied.