[Rcpp-devel] trans() changed in latest RcppArmadillo
On 31 May 2011 05:02, Savitsky, Terrance <savitsky at rand.org> wrote:
Dr. Sanderson, I've been able to verify that my issue resides in the inv() function (that I typically apply to small p x p matrices, where p = 3 - 10). In particular, the new version/algorithm induces numerical instability.
This is a bit strange, as the underlying algorithm for inv() hasn't changed. For matrix sizes <= 4x4, a fast version is used (and has been since the early days of Armadillo). For sizes >= 5x5, Lapack is used. Can you provide an example matrix which provides different results with the new Armadillo ? btw, in the new Armadillo the solve() function now calls inv() for matrix sizes <= 4x4. Otherwise it uses Lapack.
I've not yet tested that the results of the inv() computation are generally accurate, only that they are not numerically robust in comparison to 0.2.19. I'm up against a deadline, so I switched back to 0.2.19, which resolves my problem, for now, but will provide a reproducible example when finished with my work.
That would be great -- as the issue affects you, it's likely to affect other people. Cheers, Conrad -- Dr Conrad Sanderson : Sr Research Scientist : NICTA : http://arma.sf.net/cs