Dear all:
It turns out that there is a problem with MNE version 2.6.1 on the Mac
OSX Intel platform. This problem may affects you if and only if *both*
of the following are true:
(i) You are using MNE on the Mac OSX Intel platform.
(ii) Your MNE version is 2.6.1, compiled on May 5, 2009 or later. You
can find this information in $MNE_ROOT/VERSION.
If both of the above are true, read on. Otherwise, you can ignore this
email.
The problem has to do with the fact that at this version many of the
MNE programs for Mac OSX Intel platform were compiled in 64-bit mode.
It turns out that in this case the BLAS matrix multiplication routine
sgemm *sometimes* produces incorrect results. I found this by looking
at output from mne_compute_raw_inverse. The workaround was to revert
to 32-bit compilation, which works correctly.
To get rid of the problem, download the newly compiled MNE 2.6.2 from
the website. With help of Matlab and comparisons between mne_analyze/
mne_make_movie/mne_compute_raw_inverse and direct tests of matrix
multiplication I verified that this version works correctly.
Again, you have nothing to worry about in this respect if you are
running MNE on LINUX or Mac PowerPC or have an MNE version prior to
May 5.
Thanks to Eliezer Kanal to bringing up the question about consistency
between mne_analyze/mne_make_movie and mne_compute_raw_inverse, which
lead to the discovery of this trouble.
Sorry for the inconvenience,
Matti
PS. The "dev" version common MNE software installation at the Martinos
Center has been updated to 2.6.2. The "standard" version is unaffected.