On Wed, Feb 17, 2010 at 02:20:55PM +0100, Steven De Herdt wrote: > Package: octave3.2 > Version: 3.2.4-1 > Severity: normal > > While working with complex matrices, I noticed strange things going on. > Some searching showed that there is definitely something wrong: > > octave3.2:2> a=orth(randn(3)); > octave3.2:3> a'*a > ans = > > 1.0000e+00 -1.0742e-16 2.2264e-16 > -1.0742e-16 1.0000e+00 -3.1537e-17 > 2.2264e-16 -3.1537e-17 1.0000e+00 > > octave3.2:4> b=orth(randn(3)+J*randn(3)); > octave3.2:5> b'*b > ans = > > 1.06482 + 0.00000i 0.10484 + 0.51574i 0.14882 + 0.71172i > 0.10484 - 0.51574i 0.90395 + 0.00000i -0.07914 + 0.08895i > 0.14882 - 0.71172i -0.07914 - 0.08895i 1.06476 + 0.00000i > > The result is Hermitian, but not quite unitary. > This also happens on octave3.0, but not on amd64 (either version). > > A look through Octave's own bug reports produces e.g. > http://old.nabble.com/wrong-output-for-complex-scalar-product-to26766409.html > and > http://old.nabble.com/Wrong-results-for-svd-on-complex-numbers-to27079479.html > , > so this kind of problem is not unheard of. > > Depending on one's needs, this may be an 'important' bug.
Hmm, this looks like http://bugs.debian.org/569436 And indeed, I can reproduce the example from there in a sid chroot. However, according to the rest of your mail, you have ATLAS installed, which is strange. > ii libatlas3gf-base 3.6.0-24 Automatically Tuned Linear > Algebra > ii libatlas3gf-sse2 3.6.0-24 Automatically Tuned Linear > Algebra Could you use the example from the above bug report and see if uninstalling the optimized libraries helps? ============================================================================== A = [1 2 3; 4 5 6; 7 8 9] * 1j; [U, S, V] = svd(A); disp(U); The correct U is: -0.000000 - 0.214837i -0.000000 - 0.887231i -0.402839 - 0.066237i 0.000000 - 0.520587i 0.000000 - 0.249644i 0.805678 + 0.132474i 0.000000 - 0.826338i 0.000000 + 0.387943i -0.402839 - 0.066237i But in the chroot, I get: -0.00000 - 0.88741i -0.00000 - 0.88723i -0.39566 - 0.10061i 0.00000 - 0.25007i -0.00000 - 0.24964i 0.79131 + 0.20122i 0.00000 + 0.38726i 0.00000 + 0.38794i -0.39566 - 0.10061i ============================================================================== Thanks Thomas -- To UNSUBSCRIBE, email to [email protected] with a subject of "unsubscribe". Trouble? Contact [email protected]
