## LMSVD
## Introduction
Consider a real matrix ( without loss of generality) and a given positive integer . The task is to find two orthogonal matrices and , and a diagonal matrix whose diagonal entries are the largest singular values of , say, , such that The approximation is called the -th dominant singular value decomposition of . ## AuthorsZaiwen Wen, BICMR, Peking University
## PublicationXin Liu, Zaiwen Wen and Yin Zhang, In the case that LMSVD is helpful in your published work, please make a reference to the above paper. ## Download and install## DownloadLMSVD is distributed under the terms of the GNU General Public License. Version 1.1 (Download)
## InstallationUnzip the distribution. This will create a directory called LMSVD. We'll refer to this directory as <lmsvd-root>. Start Matlab and add <lmsvd-root> to your path:
>> addpath <lmsvd-root> To verify the LMSVD installation, execute the following command within Matlab:
>> demo_lmsvd ## A Typical Demo Result
(The figure in the left-hand side is the distribution of relevant singular values of a random test matrix in our demo, and the right-hand side is the runtime comparison between Matlab 'svds’ and 'lmsvd’ in calculating 20 to 100 dominant singular values of the test matrix.) ## Known Issues
In Matlab, dense linear algebra operations have been generally well optimized by using BLAS and LAPACK tuned to the CPU processors in use. On the other hand, we have observed that some sparse linear algebra operations in Matlab (up to version 2013b) seem to have not been as highly optimized. In particular, when doing multiplications between a sparse matrix and a dense vector or matrix (denoted by “SpMM’’) the performance of Matlab's own version of SpMM can differ significantly from that of the corresponding routine in Intel's Math Kernel Library (MKL), which is named ”mkl_dcscmm’’.
Comparison Results from running a Matrix Completion code called SVT are given below. LMSVD with SpMM-MKL is faster than with SpMM-Matlab.
Detailed explanations are given in Section 5.2 of our paper “X. Liu, Z. Wen, Y. Zhang, ## Related Software and Links |