Possible svd problem ?
Brought to you by:
rwhaley,
tonyc040457
Menu
▾
▴
#486 Possible svd problem ?
closed
5
2007-10-02
2007-09-03
No
The problem described here:
https://bugs.launchpad.net/ubuntu/+source/python-scipy/+bug/134277
Originally thought to be a scipy issue, appears very likely to be the result of changes made to atlas somewhere between versions 3.6.0-13 and 3.6.0-20.
Briefly, using svd to generate a kernel for some (but not all) large matrices fails such that:
k = kernel(A)
A.k != 0.
More complete description with test script and data from the link above.
Thanks for reading,
Logged In: YES
user_id=182470
Originator: NO
Hmm, this isn't near enough info for me to do anything. I don't have any idea of what the problem is, and/or any idea if ATLAS is causing it. The 3.7 series will become the new stable by the end of this month, so I am not supporting 3.6 much anyway.
So, I would ask the following:
(1) See if the problem occurs with the newest 3.7 release
(2) If so, can you reproduce it with a C test that I can run (I don't run scipi)?
(3) ATLAS doesn't supply the SVD, so I assume you are using LAPACK's, which would mean that if ATLAS is causing the problem it would likely come from the BLAS. See what ATLAS BLAS is causing the trouble, by linking to various F77 BLAS until problem goes away.
BTW, 3.7.37 passes all the lapack testers that f77 lapack & blas do . . .
Let me know if you get any/all of this info . . .
Cheers,
Clint
Logged In: YES
user_id=182470
Originator: NO
No response, closing. Reopen if problem still occurring and we have something to report.