My impression is that the advantages over Lanczos include the following points:
a) the random vectors are multiplied by A simultaneously rather than sequentially as in Lanczos b) no risk of losing orthgonality c) the bounds are known so we know how many extra vectors we need to use. With Lanczos, you are never quite sure how many you need. d) it looks like it is possible to compute the decomposition in one or two passes over A. This point alone is huge On Fri, Sep 25, 2009 at 12:03 PM, Jake Mannix <[email protected]> wrote: > Those look very cool, I'd love to see how those compare with doing > plain-old Lanczos for SVD on Hadoop. > -- Ted Dunning, CTO DeepDyve
