[R-pkg-team] Bug#994457: r-cran-gnm autopkgtest needs to be adapted for lapack 3.10

[Heather Turner]

Hi Andreas,

Thanks for forwarding this bug report and thanks to Sébastien for the detailed and accurate analysis. I will need to submit an update to CRAN, which should be feasible in the next week or two. Is there a deadline that I need to work to?

Best wishes,

Heather

On Mon, Sep 27, 2021, at 10:12 AM, Andreas Tille wrote:
> Control: tags -1 upstream
> Control: forwarded -1 Heather Turner <ht at heatherturner.net>
>
> Hi Heather,
>
> the Debian packaged gnm recieved a bug report about a failing test in
> connection with the upgrade to lapack 3.10.0 on the machine running the
> test.  Please read the bug report below.
>
> We admit we need your help to solve this issue that might affect
> other systems as well.
>
> Kind regards
>
>     Andreas.
>
> On Thu, Sep 16, 2021 at 10:59:19AM +0200, Sébastien Villemot wrote:
>> Package: r-cran-gnm
>> Version: 1.1-1-2
>> Severity: serious
>> Tags: sid bookworm
>> X-Debbugs-CC: debian-ci at lists.debian.org
>> User: debian-ci at lists.debian.org
>> Usertags: needs-update
>> 
>> Dear Maintainer,
>> 
>> Since the upload of lapack 3.10.0-1, the autopkgtest of r-cran-gnm
>> fails in unstable. See for example:
>> https://ci.debian.net/data/autopkgtest/unstable/amd64/r/r-cran-gnm/15155026/log.gz
>> 
>> More precisely, test-biplot.R fails, because some results have the
>> opposite sign compared to the one which is expected.
>> 
>> My understanding is that this comes from the SVD of barleyMatrix in
>> that test file, which is different between lapack 3.9 and 3.10.
>> Mathematically, the SVD is not unique, and lapack 3.10 returns a
>> different (still valid) solution. More precisely, I verified that one
>> of the right-singular vector of that matrix has the opposite sign in
>> lapack 3.10. I also verified that the decomposition is correct by
>> checking that:
>> 
>>   max(abs(barleySVD$u %*% diag(barleySVD$d) %*% t(barleySVD$v) - barleyMatrix))
>> 
>> is a small value (about 2e-14).
>> 
>> Also note that the hardcoded expected values already partially differ 
>> from those of the original research paper mentioned in that test
>> (Gabriel (1998): Generalised bilinear regression). More precisely, half
>> of the values were hardcoded with the opposite sign. It seems that now
>> all values need to be hardcoded with the opposite sign.
>> 
>> The testsuite of r-cran-gnm thus needs to be adapted, by being more
>> tolerant to such sign changes.
>> 
>> N.B. : when trying to reproduce the problem, please ensure that your
>> lapack alternative (as given by “update-alternatives --display
>> liblapack.so.3-x86_64-linux-gnu) points to /usr/lib/x86_64-linux-
>> gnu/lapack/liblapack.so.3, and not to the binary provided by either
>> openblas or atlas (because these two have not yet been recompiled
>> against lapack 3.10, and thus do not expose the problem).
>> 
>> Best regards,
>> 
>> -- 
>> ⢀⣴⠾⠻⢶⣦⠀  Sébastien Villemot
>> ⣾⠁⢠⠒⠀⣿⡁  Debian Developer
>> ⢿⡄⠘⠷⠚⠋⠀  https://sebastien.villemot.name
>> ⠈⠳⣄⠀⠀⠀⠀  https://www.debian.org
>> 
>
>
>
>> _______________________________________________
>> R-pkg-team mailing list
>> R-pkg-team at alioth-lists.debian.net
>> https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/r-pkg-team
>
>
> -- 
> http://fam-tille.de