[FieldTrip] ICA and covariance matrix for CSP

Johanna Zumer johanna.zumer at donders.ru.nl
Tue Apr 23 17:16:28 CEST 2013


Hi Bethel,

Removing a component with ICA will reduce the rank (number linear
independent components) of your data.  The covariance matrix itself
probably can be computed, but I suspect the warning you are getting is when
you are computing the inverse.   The inverse of a rank-deficient matrix
cannot be computed (and so the warnings you get are expected).

Regularisation (e.g. adding some 'noise' to the diagonal of the covariance
matrix, roughly equal to the amount of estimated sensor noise) prior to
inversion is the common way around this problem.   In FieldTrip, this is
cfg.lambda when calling ft_sourceanalysis, which you can set to a
percentage, for example '5%' is common, of the trace of the covariance
matrix.

Best,
Johanna



2013/4/23 Bethel Osuagwu <b.osuagwu.1 at research.gla.ac.uk>

> Hi Everyone,
>
> I want to ask a general mathematical question relating to ICA and common
> spatial pattern. Hopefully someone will help me out.
>
> I have noticed that after using ICA to remove artifact in my signal, I can
> no longer compute the covariance matrices needed for common spatial pattern
> without matlab giving me warning regarding the accuracy of the estimated
> matrix. Indeed the estimated matrix is not accurate. This does not happen
> if I use only a small subset of my channels after removing noise with ICA.
>  If I do not use ICA at all the warning disappears.
>
> I want to ask if anybody knows why ICA causes this issue and how I can fix
> it.
>
> Thanks a lot.
> Bethel Osuagwu
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20130423/3c750feb/attachment-0002.html>


More information about the fieldtrip mailing list