On Thu, July 20, 2006 17:35, Talha Yalta wrote: > Professor Cottrell: > During my testing of gretl using the StRD linear regressions test > suit, I found that QR decomposition performs better than the Cholesky > decomposition and sent you a pdf file containing a table comparing the > two methods. QR method mostly creates higher number of accurate digits > and is able to produce a solution for the Flip data set, where > Cholesky fails. > > In the light of this evidence I wrote my paper and prepared the > summary tables assuming the new default for linear regressions would > be the QR decomposition. I see that the new snapshots still have > Cholesky as the default. If Cholesky will stay as the default in the > new version, please let me know so that I can update my tables. > > I am attaching the pdf file containing the comparisons.
IIRC, Sven brought this up some time ago. I did a little testing, and QR is about 10-15% slower than Cholesky. This is the price you have to pay for greater accuracy. So, it all boils down to choosing speed over precision. I would go for QR myself, but in 99.99% of the cases the difference in precision is not even noticeable: the test cases you give are artificial datasets especially designed to be very ill-conditioned (but, to be honest, I did stumble once into a real-life dataset where Cholesky couldn't cut it and QR would). Besides, with the CPUs we have today, a few microseconds are nothing. What's other people's opinion? Riccardo "Jack" Lucchetti Dipartimento di Economia Facoltà di Economia "G. Fuà" Ancona
