Very interesting, Dirk. Thanks! On Mon, Jan 31, 2022 at 12:18 PM Dirk Eddelbuettel <e...@debian.org> wrote:
> > On 31 January 2022 at 09:13, Zé Vinícius wrote: > | Yes, quoting the paper on which ‘solve’ is based on ( > | http://arma.sourceforge.net/armadillo_solver_2020.pdf): > | > | “The SVD-based solver uses the xGELSD set of functions, which find a > | minimum-norm solution to a linear least squares problem.” > > Yes. A longer story though on how this is not what may matter. I should > start > by saying that I never found really decent documentation describing this > 'split' in the world view as this problem is, in essence, seen differently > by > numerical analysis specialists (as for example the LAPACK authors of the > *GELSD functions) and the statistical users which differs in what they > focus > on. Which is why R uses a modified version of LINPACK (as I recall going > back > to by Ross Ihaka) when computing lm(). And not the LAPACK routines. > > Doug was always adamant about this when I wrote the different simple > fastLm() > approaches (in RcppGSL, RcppArmadillo, ...) which do _not_ properly account > for rank-deficiency (as R's lm() would) -- Doug also wrote the nicest > fastLm > example in RcppEigen. > > There is a little bit more in the help pages for the various lmFast() > versions as well as an explicit example (also due to Doug). > > Now, I should add that whenever I tried to construct an example on a more > real-world-alike regression problem, I could not come anywhere close to > actually seeing the rank deficiency. But it is unmistakenly there is the > appropriately created case. So buyer beware. > > Dirk > > -- > https://dirk.eddelbuettel.com | @eddelbuettel | e...@debian.org > -- Zé Vinícius https://mirca.github.io
_______________________________________________ Rcpp-devel mailing list Rcpp-devel@lists.r-forge.r-project.org https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
