eigen in beta
Paul Gilbert wrote:
Prof Brian Ripley wrote:
On Wed, 11 Apr 2007, Paul Gilbert wrote:
Hmmm. It is a bit disconcerting that make check passes and I can get
a fairly seriously wrong answer. Perhaps this test could be added to
make check.
Well, you have learnt something new about software engineering! 'make
check' is supposed to test the operation of R, not compilers.
I'm not sure I would call this engineering. Engineers are suppose to worry about the integrity of the whole system. In any case, if it is not make check, then what is the system for checking that you have a good R build?
Well, there's make check-all...
That doesn't check everything either, though. The point is that there is
a limit to what we can check: We don't check whether the CPU gets
floating point operations wrong in the 5th decimal place on rare
occasions either. (Same thing applies to structural engineering: You
can't test the chemical composition of every steel rod. At some point
you have to shift reponsibility to your subcontractors.)
We do have regression checks though, i.e. we try to ensure that known
errors do not reappear, and compiler issues have been worked around
occasionally. E.g. (IIRC) the original qbeta() function was written so
as to get the very last bit of accuracy squeezed out of the solution to
a nonlinear equation, but modern optimizing compilers would reorder some
instructions, losing a few bits of accuracy and sending the code into
an infinite loop in some cases.
-pd
Whether or not the one answer is correct may be questionable, but
there is no question that
prod(eigen(z, symmetric = FALSE, only.values = TRUE)$values ) *
prod(eigen(solve(z), symmetric = FALSE, only.values = TRUE)$values )
[1] 1.01677-0i
is wrong. (The product of the determinants should equal the
determinant of the product, and the determinant of I is 1.)
On this machine I am using GNU Fortran (GCC 3.2.3 20030502 (Red Hat
Linux 3.2.3-54)) 3.2.3 20030502 (Red Hat Linux 3.2.3-54), which is a
bit old but not really old. (I am using gcc 4.1.1 on my home machine,
but the failing machine is suppose to be fairly new and "supported."
(There is an OS upgrade planned.) Should I really be thinking of
this as an old compiler? Is the
Yes, a known bad compiler, as searching the list archives would have
found out. Don't use anything older that gcc 3.4.x on x86_64 Linux.
Thanks, I'll bug the sys admin. Paul
compiler the most likely problem or is it possible I have a bad BLAS
configuration, or something else? Previous versions of R have
compiled without problems on this machine. (I am never very sure
where to find all the information to report for a problem like this.
Is there a simple way to get all the relevant information?)
Paul Gilbert
Prof Brian Ripley wrote:
All the systems I tried this on give the 'correct' answer, including
x86_64 Linux, FC5 (gcc 4.1.1)
i686 Linux, FC5
ix86 Windows (both gcc 3.4.5 and gcc pre-4.3.0)
Sparc Solaris, with gcc3, gcc4 and SunPro compilers.
Mainly with R 2.5.0 beta, some with R-devel (where the code is
unchanged).
We have seen problems specific to RHEL's Fortran compilers on x86_64
several times before. I would strongly recommend compiler updates.
On Tue, 10 Apr 2007, Peter Dalgaard wrote:
Paul Gilbert wrote:
Here is the example. Pehaps others could check on other platforms.
It is
only the first eigenvalue that is different. I am relatively sure
the
old values are correct, since I compare with an alternate calculation
using the expansion of a polynomial determinant.
z <- t(matrix(c(
0, 0, 0, 0, 0, 0, 0, 0, 0, -0.0064083373167516857,
-0.14786612501440565826, 0.368411802235074137,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0568624483195125444,
0.08575928008564302762, -0.101993668348446601,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0.0039684327579889069,
-0.00002857482925046247, 0.202241897806646448,
1, 0, 0, 0, 0, 0, 0, 0, 0, -0.0222834092601282285,
-0.09126708346036176145, 0.644249961695308682,
0, 1, 0, 0, 0, 0, 0, 0, 0, -0.0032676036920228878,
0.16985862929849462888, 0.057282326361118636,
0, 0, 1, 0, 0, 0, 0, 0, 0, 0.0148488735227452068,
-0.06175528918915401677, 0.109566197834008949,
0, 0, 0, 1, 0, 0, 0, 0, 0, -0.0392756265125193960,
0.04921079262665441212, 0.078176878215115805,
0, 0, 0, 0, 1, 0, 0, 0, 0, -0.0013937451966661973,
0.02009823693764142133, -0.207228935136287512,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0.0273358858605219357,
0.03830466468488327725, 0.224426004034737836,
0, 0, 0, 0, 0, 0, 1, 0, 0, -0.1456426235151105919,
0.28688029213315069388, 0.326933845656016908,
0, 0, 0, 0, 0, 0, 0, 1, 0, 0.0164670122082246559,
-0.21966261349875662590, 0.036404179329694988,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0.0146156940584119890,
0.07505490943478997090, 0.077660578370038813
), 12, 12))
R-2.5.0 gives
eigen(z, symmetric = FALSE, only.values = TRUE)$values
[1] 0.8465266+0.0000000i -0.0280087+0.6244992i
-0.0280087-0.6244992i
[4] -0.2908409+0.5522274i -0.2908409-0.5522274i
-0.6228929+0.0000000i
[7] 0.6177419+0.0000000i -0.5604582+0.1958709i
-0.5604582-0.1958709i
[10] 0.1458799+0.4909300i 0.1458799-0.4909300i
0.3378356+0.0000000i
R-2.4.1 and many, many previous versions gave
eigen(z, symmetric = FALSE, only.values = TRUE)$values
[1] 0.8794798+0.0000000i -0.0280087+0.6244992i
-0.0280087-0.6244992i
[4] -0.2908409+0.5522274i -0.2908409-0.5522274i
-0.6228929+0.0000000i
[7] -0.5604582+0.1958709i -0.5604582-0.1958709i
0.5847887+0.0000000i
[10] 0.1458799+0.4909300i 0.1458799-0.4909300i
0.3378356+0.0000000i
Sys.info()
sysname
release
"Linux"
"2.4.21-40.ELsmp"
version
nodename
"#1 SMP Thu Feb 2 22:13:55 EST 2006"
"mfa04559"
machine
"x86_64"
Paul Gilbert
Hmm, I don't get that
version$version.string
[1] "R version 2.5.0 beta (2007-04-10 r41105)"
eigen(z, symmetric = FALSE, only.values = TRUE)$values
[1] 0.8794798+0.0000000i -0.0280087+0.6244992i -0.0280087-0.6244992i
[4] -0.2908409+0.5522274i -0.2908409-0.5522274i -0.6228929+0.0000000i
[7] -0.5604582+0.1958709i -0.5604582-0.1958709i 0.5847887+0.0000000i
[10] 0.1458799+0.4909300i 0.1458799-0.4909300i 0.3378356+0.0000000i
Sys.info()
sysname
release
"Linux"
"2.6.20-1.2933.fc6"
version
nodename
"#1 SMP Mon Mar 19 11:38:26 EDT 2007" "titmouse2.kubism.ku.dk"
machine login
"i686" "pd"
user
"pd"
And
version$version.string
[1] "R version 2.5.0 beta (2007-04-09 r41098)"
eigen(z, symmetric = FALSE, only.values = TRUE)$values
[1] 0.8794798+0.0000000i -0.0280087+0.6244992i -0.0280087-0.6244992i
[4] -0.2908409+0.5522274i -0.2908409-0.5522274i -0.6228929+0.0000000i
[7] -0.5604582+0.1958709i -0.5604582-0.1958709i 0.5847887+0.0000000i
[10] 0.1458799+0.4909300i 0.1458799-0.4909300i 0.3378356+0.0000000i
Sys.info()
sysname
release
"Linux"
"2.6.18.8-0.1-default"
version
nodename
"#1 SMP Fri Mar 2 13:51:59 UTC 2007"
"viggo"
machine
login
"x86_64"
"pd"
user
"pd"
The latter should be the actual build used in the current beta tarball
(which is what you used, right?).
I would suspect one of the following:
- RHEL compilers
- over-optimizing compiler settings
- system blas/lapack libraries
- system glibc
but Brian probably has more concrete information.
Prof Brian Ripley wrote:
We are only aware of better behaviour from LAPACK 3.1 (which is
what I
suppose you are talking about, that is R compiled with its internal
LAPACK).
But in at least one case that means finding a complex set of
eigenvalues where previously a real one was found.
On Tue, 10 Apr 2007, Paul Gilbert wrote:
I am having some trouble with a case where eigen in R-beta
gives a
different largest value than in previous versions of R. Other
values
seem to be the same. Before I spend too much time, is anyone
aware of a
problem (symmetric = FALSE, only.values = TRUE).
Paul Gilbert
====================================================================================
La version fran?aise suit le texte anglais.
------------------------------------------------------------------------------------
This email may contain privileged and/or confidential
inform...{{dropped}}
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
====================================================================================
La version fran?aise suit le texte anglais.
------------------------------------------------------------------------------------
This email may contain privileged and/or confidential
inform...{{dropped}}
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
====================================================================================
La version fran?aise suit le texte anglais.
------------------------------------------------------------------------------------
This email may contain privileged and/or confidential information,
and the Bank of
Canada does not waive any related rights. Any distribution, use, or
copying of this
email or the information it contains by other than the intended
recipient is
unauthorized. If you received this email in error please delete it
immediately from
your system and notify the sender promptly by email that you have
done so.
------------------------------------------------------------------------------------
Le pr?sent courriel peut contenir de l'information privil?gi?e ou
confidentielle.
La Banque du Canada ne renonce pas aux droits qui s'y rapportent.
Toute diffusion,
utilisation ou copie de ce courriel ou des renseignements qu'il
contient par une
personne autre que le ou les destinataires d?sign?s est interdite. Si
vous recevez
ce courriel par erreur, veuillez le supprimer imm?diatement et
envoyer sans d?lai ?
l'exp?diteur un message ?lectronique pour l'aviser que vous avez
?limin? de votre
ordinateur toute copie du courriel re?u.
====================================================================================
La version fran?aise suit le texte anglais.
------------------------------------------------------------------------------------
This email may contain privileged and/or confidential inform...{{dropped}}
______________________________________________ R-devel at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel