make check on DU4 with R-1.1.0 snapshot
On 2 Jun 2000, Peter Dalgaard BSA wrote:
Albrecht Gebhardt <albrecht.gebhardt@uni-klu.ac.at> writes:
abs(X - s$u %*% D %*% t(s$v)) - Eps
[,1] [,2] [1,] -2.109424e-15 -2.220446e-16 [2,] -1.998401e-15 -8.881784e-16 [3,] -2.220446e-15 -1.776357e-15 [4,] -1.998401e-15 -1.332268e-15 [5,] -1.998401e-15 -1.332268e-15 [6,] -1.998401e-15 4.440892e-16 [7,] -1.332268e-15 -2.220446e-15
abs(D - t(s$u) %*% X %*% s$v) - Eps
[,1] [,2] [1,] 3.108624e-15 -8.881784e-16 [2,] -1.165734e-15 -2.220446e-15 4.440892e-16 and 3.108624e-15 Eps was:
Eps
[1] 2.220446e-15
...
eps: 2.22044605E-16
So one of the calculations end up at about 5.3e-15 which is over 20 times the machine epsilon. OK, Hilbert matrices are nasty and AFAIR Alpha hardware doesn't have the extended precision of Intel FPUs but does this look reasonable enough that we should just use a bigger Eps?
Other hardware with standard IEEE arithmetic succeeds. But I think Eps = 100 * .Machine$double.eps would be adequate. My Solaris box gets up to 1.77e-15. I've just committed 100 * .Machine$double.eps. Brian
Brian D. Ripley, ripley@stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272860 (secr) Oxford OX1 3TG, UK Fax: +44 1865 272595 -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._