[Rd] eigen returns NAs from a real matrix

[Prof Brian Ripley]

On Mon, 19 Mar 2007, Martin Maechler wrote:

>>>>>> "Duncan" == Duncan Murdoch <murdoch at stats.uwo.ca>
>>>>>>     on Mon, 19 Mar 2007 09:01:39 -0400 writes:
>
>    Duncan> On 3/19/2007 6:48 AM, Martin Maechler wrote:
>    >>>>>>> "BDR" == Prof Brian Ripley <ripley at stats.ox.ac.uk>
>    >>>>>>> on Sun, 18 Mar 2007 06:51:58 +0000 (GMT) writes:
>    >>
>    BDR> On Sat, 17 Mar 2007, Spencer Graves wrote:
>    >> >> Hi, All:
>    >> >> Attached please find a symmetric, indefinite matrix for which
>    >> >> 'eigen(...)$vectors' included NAs:
>    >> >>> load("eigenBug.Rdata")
>    >> >>> sum(is.na(eigen(eigenBug)$vectors))
>    >> >> [1] 5670
>    >> >>> sessioninfo()
>    >> >> Error: could not find function "sessioninfo"
>    >> >>> sessionInfo()
>    >> >> R version 2.4.1 (2006-12-18)
>    >> >> i386-pc-mingw32
>    >> >>
>    >> >> locale:
>    >> >> LC_COLLATE=English_United States.1252;LC_CTYPE=English_United
>    >> >> States.1252;LC_MONETARY=English_United
>    >> >> States.1252;LC_NUMERIC=C;LC_TIME=English_United States.1252
>    >> >>
>    >> >> attached base packages:
>    >> >> [1] "stats"     "graphics"  "grDevices" "utils"     "datasets"  "methods" [7]
>    >> >> "base"
>    >> >> Using EISPACK does NOT return NAs: > sum(is.na(eigen(eigenBug,
>    >> >> EISPACK=TRUE)$vectors))
>    >> >> [1] 0
>    >> >>
>    >> >>
>    >> >> I traced the problem to the following line in 'eigen':
>    >> >> z <- if (!complex.x)
>    >> >> .Call("La_rs", x, only.values, PACKAGE = "base")
>    >> >>
>    >> >>
>    >> >> Comments?
>    >>
>    BDR> Nothing appeared as an attachment.
>    >>
>    >> it was a zip file which is not among the allowed ones
>    >> (.tar.gz / *.tgz  would work).
>    >>
>    >> Spencer has also contacted me in private,
>    >> so you now can use something like :
>    >>
>    >> F <- tempfile()
>    >> download.file("ftp://stat.ethz.ch/U/maechler/R/eigenBug.Rdata", F)
>    >> load(F)
>
>    Duncan> This is unrelated to Spencer's problem, but this
>    Duncan> sequence failed for me.
>
>    Duncan> By default it will treat the file as a text file,
>    Duncan> and convert EOL markers for Windows.  Adding
>    Duncan> "mode='wb'" to the download.file() call is necessary.
>
> ok;  Thank you, Duncan.
>
> {For me, that's yet another reason why I think it would be  ``nice''
> if both  load() and source()  would work with general
> connections the same way as read.table(), and I could just say
>
>     load("ftp://stat.ethz.ch/U/maechler/R/eigenBug.Rdata")
> }

The help page for load() does have an example of using a url() connection 
here, and that works in this example on Windows, as does

load(url("ftp://stat.ethz.ch/U/maechler/R/eigenBug.Rdata"))

Not so hard, surely?  (The reason load() is different is that it can 
transparently handle compressed files, unlike read.table.)

>
>    >> str(eigenBug)# tons of dinames; get rid of them:
>    >> M <- unname(eigenBug)
>    >>
>    >> sum(is.na(eigen(M)$vectors))
>    >> # 0 (for MM, on Linux);  '5670' for Spencer
>
>    Duncan> I get the same as Spencer in 2.4.1, but zero in
>    Duncan> R-devel, in Windows.
>
> ahh, that's good news.
>
> Maybe Spencer could install a revcent version R-devel on his
> Windows machine, and try the example?

My belief is that the LAPACK update has solved this on Windows: it does 
not happen on my Linux systems, and does happen in R-patched on Windows.


>    Duncan> I haven't tried R-patched yet: there were network
>    Duncan> problems here this morning, and the build failed.
>
>    Duncan> Duncan Murdoch
>
>    >> sum(is.na(eigen(M, EISPACK=TRUE)$vectors))
>    >> # 0
>    >>
>    >> ## Guess about what kind of BLAS/Lapack I'm using :
>    >> file.info(list.files(R.home("lib"), full=TRUE))[,1:3]
>    >> ##                                        size isdir mode
>    >> ## ....64-linux-inst/lib/libRblas.so    422558 FALSE  755
>    >> ## ....64-linux-inst/lib/libRlapack.so 3144596 FALSE  755
>    >>
>    >>
>    >> I.e. I don't get any NA's
>    >> which shows to me that the bug is most probably in the Lapack -
>    >> Version that Spencer's version of R is linked against.
>    >>
>    >> Martin Maechler, ETH Zurich
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

-- 
Brian D. Ripley,                  ripley at 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 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595