Carlos Hernandez | 7faaa9f | 2014-08-05 17:53:32 -0700 | [diff] [blame] | 1 | This unsupported package is about a divide and conquer algorithm to compute SVD. |
| 2 | |
| 3 | The implementation follows as closely as possible the following reference paper : |
| 4 | http://www.cs.yale.edu/publications/techreports/tr933.pdf |
| 5 | |
| 6 | The code documentation uses the same names for variables as the reference paper. The code, deflation included, is |
| 7 | working but there are a few things that could be optimised as explained in the TODOBdsvd. |
| 8 | |
| 9 | In the code comments were put at the line where would be the third step of the algorithm so one could simply add the call |
| 10 | of a function doing the last part of the algorithm and that would not require any knowledge of the part we implemented. |
| 11 | |
| 12 | In the TODOBdcsvd we explain what is the main difficulty of the last part and suggest a reference paper to help solve it. |
| 13 | |
| 14 | The implemented has trouble with fixed size matrices. |
| 15 | |
| 16 | In the actual implementation, it returns matrices of zero when ask to do a svd on an int matrix. |
| 17 | |
| 18 | |
| 19 | Paper for the third part: |
| 20 | http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf |
| 21 | |