[Lapackpp-devel] Re: LaEigSolveVecIP (was: Complex Matrix Support)

[Christian S. <sti...@tu...>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Dear Andrew,

thanks for this detailed bug report. Indeed the output of the
LaEigSolveVecIP function is not what we would expect.

However, the documentation of that function immediately explains what is
going on: http://lapackpp.sourceforge.net/html/laslv_8h.html#a16 and the
laslv.h header file says: "This function calculates all eigenvalues and
eigenvectors of a symmetric matrix A, not a general matrix A!", and
since you passed a non-symmetric matrix into that function, it is clear
that it won't calculate what you expected. Instead, you have to use the
function LaEigSolve() and this will do what you expected. I've modified
your test file so that you can see what is going on. The output of
LaEigSolve() will fortunately match up with the MATLAB output again.

The LaEigSolve() function will return potentially complex-valued
eigenvalues, but this is perfectly correct for a general (non-symmetric)
matrix. Only in your particular case all eigenvalues happened to be
real-valued.

As for the "Additionally, for each run, the values do not match each
other" remark: This is also correct, because the function name suffix
"IP" stands for "in-place" (and also the matrix argument is non-const)
which means the given matrix M will be modified during calculation. This
is useful if you don't need the matrix any longer, because then lapack
internally doesn't have to allocate an extra copy of the matrix. If you
don't want to have your original matrix modified, either use a non-"IP"
version of that function (if it exists), or create a temporary copy and
pass that one into the function (which is just what the non-IP functions
will do for you). You can read the manual page of the underlying lapack
function "dsyev" for more details by typing the command "man dsyev" (if
the package "lapack-manpages" is installed).

Thank you again for this detailed bug report. This made it very clear to
me how to explain the problem here. I'm sorry if the documentation was
not clear enough about this function. I should probably change the
function name into something less ambiguous, like
"LaEigSolveVeryMisleadingOnlyForSymmetricMatrixIP" :-) but I will try to
improve the function documentation.

Regards,

Christian Stimming

Andrew Oh schrieb:
> Hey
> 
> Attached is the C++ code and the M file I used. The C++ code has some
> more detailed results in the commenting.
> 
> The machine I ran this on is a SuSe 10.0 from Novell. This is what I
> get when I type "uname -a"
> 
> Linux mnp 2.6.13-15.8-smp #1 SMP Tue Feb 7 11:07:24 UTC 2006 i686 i686
> i386 GNU/Linux
> 
> But the RPMs I installed for SuSe all had the i586 suffix on them.
> 
> These are the versions of the relevant compilers/rpms:
> 
> gcc-4.0.2_20050901-3
> gcc-fortran-4.0.2_20050901-3
> f2c-0.11-1052
> 
> When I compile test.cpp I type:
> 
> g++ test.cpp -I/usr/include -I/usr/include/lapackpp -L/usr/lib -llapackpp
> 
> and execute by typing "./a.out".
> 
> The m-file is just two simple lines and you can see the results in
> MATLAB and compare to the values printed out by test.cpp. Thanks for
> your help.
> 
> Andy
> ------------------------------------------------------------------------
> 
> M = [1,2,3,4;5,6,7,8;9,10,11,12;13,14,15,16];
> eig(M)

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.2.1 (MingW32)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iQCVAwUBRF76DGXAi+BfhivFAQLDFgP+Pk794zU4TTl2qKP4fgao8SPhr7p9jESj
sac5X/Tc8okSt1l/oP4acisdp5J7pVmoUYPm0ygZTJO+U/xMLEcWlxRZ17FBlh/c
ynYjT+lHq1KSAq2okBglQxL0L7ELQyzJ/zC9XqXbw/RJ73CpS6nN1cLbjDjcgML9
+GtUDp3gRRY=
=Loys
-----END PGP SIGNATURE-----