I've recently spend a lot of time trying to understand the algorithm used
in MNE software, majorly from this MNE link: http://martinos.org/mne/stable/manual/mne.html
I'm having a hard time understanding how the Source Covariance matrix (R)
is calculated, as the source is unknown. For example, these equations
involved known R to calculate the inverse operator:
[image: \tilde{M} = R \tilde{G}^T (\tilde{G} R \tilde{G}^T + I)^{-1}\ ,]
[image: \tilde{M} = R^{^1/_2} V \Gamma U^T]
It would be great if you can point me to literature references explaining
this algorithm. Thank you in advance.
the source cov is parametrized by the depth and the loose paramaters.
you have a look at the MNE Python code to know all the details.
regarding papers you have a look at:
F. H. Lin, J. W. Belliveau, A. M. Dale, and M. S. Hamalainen,
?Distributed current estimates using cortical orientation
constraints,? Hum Brain Mapp, vol. 27, pp. 1-13, 2006.
Hi,
If you think of the MNE inverse solution in the Bayesian context, the
source covariance matrix is part of the "prior".. i.e., it is something
that is assumed based on previous knowledge..
The standard assumptions that are used that determine R are as follows:
(i) Apriori all sources are mean 0, and given equal weight to be active
and independent of each other: this means that R is a diagonal matrix with
all the elements on the diagonal being equal to the same number..
(ii) The SNR you get in your measurement is 3:1. This, along with the
estimate of the noise covariance matrix helps determine what that constant
on the diagonal ends up being..
If you have say a map of weights you want to give to the sources instead
of the above (say based on a prior fMRI experiment), then you would make
corresponding entries along the diagonal of R higher or lower to match
your prior.