On Fri, Sep 25, 2009 at 12:11 PM, Ted Dunning <[email protected]> wrote:
a) the random vectors are multiplied by A simultaneously rather than > sequentially as in Lanczos > > b) no risk of losing orthgonality > I haven't read the papers in detail yet, have you seen how these two points are done? In a parallel environment, how do you maintain orthogonality while looking for multiple projection directions? Whats to keep you from finding the same directions? d) it looks like it is possible to compute the decomposition in one or two > passes over A. This point alone is huge > Yeah, that is huge! -jake
