a couple of days ago I posted a message and just one person responded. Maybe I didn't point at the problem straight forward enough.
Here it comes again:
I have a matrix "C_22" which is real and 6x6. (I included it at the end.) It is supposed to have all singular values close to 1. When I load this matrix into Matlab and try to compute the singular values or the spectral norm, I get errors like
>> svd(C_22) ??? Error using ==> svd Solution will not converge.
>> norm(C_22) ??? Error using ==> norm Solution will not converge.
Unfortunately, this error terminates the computation and I have some jobs running for hours. I assume that svd or norm are heavily used inside Matlab and I do not know when this error will occur.
It seems to be caused libc.so or libm.so. I reported this problem to TMW and their workaround was to add some noise to the matrix. Unfortunately this does not work for me.
I am running the update of Matlab for Linux which can be found on ftp.mathworks.com. The binary on the CD-ROM shipped by TMW segfaults under my Linux kernel. I use Linux kernel 2.0.0 (2.0.20 does not solve the problem) from the (german) "S.U.S.E.ÃÂ Linux 4.2" distribution. The error occurrs regardless of the processor Matlab is running on. (I verified 486DX/2-66, P5-133 (new one - not with the fdiv bug), P5-150 (new as well), and a P6-200.)
Here are my questions:
1) Is there anybody out there running Linux 2.0.0 (or newer) AND the update from TMW who can compute the singular values of C_22? If so, please tell me what distribution of Linux and combination of kernel and libraries (libm.so, libc.so...) are used.
2) Is there anybody with Linux 2.0.* running the ORIGINAL binary from the installation CD-ROM? Under Linux 1.2.13 this was running smoothly and svd/norm worked...
3) This one is for TMW specially! Is it possible (and if, how) to replace a built-in function like svd or norm by a MEX-file? The svd-solver from LAPACK-2.0 converges and I would like to replace the built-in solver by the LAPACK solver. (A straight forward approach didn't work.)
Here the uuencoded matrix comes:
At the moment downgrading to an out-of-date 1.2.13 distribution and using the original binary seems to be the only way to get a working Matlab.
Thank you for your efforts!
ps: Even on C_22(4:6,4:6) svd or norm does not work! On our Sparc20 I get the right answers.