People generally believe that a PC can handle SVD for matrices only up to thousands by thousands. Textbooks also suggest that it is not wise to compute singular vectors one by one. **In this post, I’ll refute both statements.** In particular, on a PC within 10 seconds without even parallelism, we can find the top 20 eigenvectors of a 260k x 260k matrix with 1.2 million entries. Moreover, the algorithm is simply to invoke rank-1 SVD 20 times. Continue reading

### Blogroll

### Archives

- January 2017 (1)
- November 2016 (1)
- July 2016 (2)
- June 2016 (3)
- May 2016 (1)
- June 2012 (3)
- May 2012 (1)
- May 2011 (1)
- March 2011 (1)
- January 2011 (5)
- December 2010 (2)
- November 2010 (2)

### Categories

- Readings (13)
- Research (22)
- Algorithm (9)
- Cryptogaphy (2)
- Game Theory (1)
- Learning Theory (6)
- Optimization (8)

- Story (2)