What you are suggesting is close to what I am doing. There are two differences,
1) the number of rows does depend on k+p, but not directly. The number of rows should be as large as practical to make the operations more efficient. 2) Oops. You are right. The columns will need to be renormalized by the n (the number of blocks) On Tue, Apr 13, 2010 at 8:42 AM, Dmitriy Lyubimov <dlie...@gmail.com> wrote: > Why can't we say s=(k+p) and accumulate (k+p) x (k+p) A*Omega rows at one > time in the mapper and make sure we orthogonalize (k+p) rows of Q in one > block (by running a stock QR)? That way orthogonalization of the final Q > will be quaranteed (although columns of Q would be scaled at approx > [m/(k+p)] )? > > -Dmitriy > > On Mon, Apr 12, 2010 at 9:32 AM, Ted Dunning <ted.dunn...@gmail.com> > wrote: > > > I will be attaching Dmitry's pdf. > > > > The basic difference is that I have gone back to a form closer to the > > original paper to avoid computing an ill-conditioned SVD. > > > > Dmitriy also said: > > > > I guess at this point i don't understand the details of block QR. I guess > > > i'll need to read up on block QR. I guess you could indeed go into a > > little > > > more into details of that qr() call and perhaps give geometry of > > individual > > > Yi, Qi and Ri matrices in step 1 ( i assume they should have s x (k+p), > s > > x > > > ?, ? x (k+p) where s is the block height .) > > > > > > > There isn't anything magic about the block QR and I doubt you will find > any > > information on the variant that I am using. The basic idea is just that > I > > am doing QR decompositions on row-wise blocks of A Omega. > > > > On Sun, Apr 11, 2010 at 10:54 PM, Ted Dunning <ted.dunn...@gmail.com> > > wrote: > > > > > > > > Can you attach that to the JIRA? Attachments are stripped from the > > mailing > > > list, I think. > > > > > > On Sun, Apr 11, 2010 at 9:55 PM, Dmitriy Lyubimov <dlie...@gmail.com > > >wrote: > > > > > >> For reference, i am attaching summary of our previous discussion > which > > i > > >> took liberty of slightly reworking. > > > > > > > > > > > >
