[FieldTrip] mutual information on LCMV virtual channels

Stenroos Matti matti.stenroos at aalto.fi
Fri Jan 29 10:19:53 CET 2021

Hi all,

I do not remember how mutual information is computed but maybe this is still informative:

Maybe the simplest form of signal similarity is correlation. Beamformer, in theoretical conditions i.e. as described in van Veen 1997 and the Sekihara--Nagarajan book, does not see correlated activity. So if our signal is produced by two sources whose time-courses are partially correlated, and we have a perfect measurement of the signal, the LCMV estimate shows only those parts of the two source time-series that are uncorrelated with the other source. In real world, noise, regularization, model errors, under-determination (more true sources than degrees-of-freedom in the measurement) etc. enter the game, but the theoretical property is still buried there.

So I am not surprised that LCMV estimate does not reproduce an MI effect.



Lähettäjä: fieldtrip <fieldtrip-bounces at science.ru.nl> käyttäjän Schoffelen, J.M. (Jan Mathijs) <jan.schoffelen at donders.ru.nl> puolesta
Lähetetty: 28. tammikuuta 2021 14:49:21
Vastaanottaja: FieldTrip discussion list
Aihe: Re: [FieldTrip] mutual information on LCMV virtual channels

Hi Xavier,

It’s not clear what in a resting-state context is meant with ‘baseline’. Neither what is meant with an increase in mutual information in the frontal region. Typically, mutual information is computed between 2 things, so what’s the other thing?

Yet, regarding the question relating to source vs. electrode space: mutual information is not immune for volume conduction effects, so what you observe at the channel level might reflect a difference in power, rather than a difference in connectivity. (Partially) undoing the volume conduction through source reconstruction might  result in a reduction in the signal leakage, and hence a  reduction in the MI.

Best wishes,

On 25 Jan 2021, at 08:39, Schoffelen, J.M. (Jan Mathijs) <jan.schoffelen at donders.ru.nl<mailto:jan.schoffelen at donders.ru.nl>> wrote:

Hi Xavier,

I am forwarding your question to the list, this is a better venue to ask your question. More people can chime in and think along.

I am aware that this question has been asked before on the list, and apparently it has not been answered. Perhaps now there is somebody who has time to think along.

Best wishes,

Begin forwarded message:

From: Xavier Vrijdag <x.vrijdag at auckland.ac.nz<mailto:x.vrijdag at auckland.ac.nz>>
Subject: mutual information on LCMV virtual channels
Date: 25 January 2021 at 02:48:06 CET
To: "jan.schoffelen at donders.ru.nl<mailto:jan.schoffelen at donders.ru.nl>" <jan.schoffelen at donders.ru.nl<mailto:jan.schoffelen at donders.ru.nl>>

Hello Jan-Mathijs,

I am using the mutual information analysis from the IBTB toolbox as incorporated in fieldtrip to run connectivity analysis on resting-state EEG data (32 channels) of participants exposed to narcotic gases. I found an increase in mutual information in the frontal region compared to baseline. I am interested in localizing this effect further, so I set up an LCMV beamformer source localization. However, in the derived time series of the virtual channels, the mutual information effect is completely gone. Could someone help me explain this effect? Could it be that LCMV, in its attempt to calculate the spatial filters, cancels out the mutual information in the EEG signals?

If helpful, I can share the Matlab scripts of the analysis pipeline.


Xavier Vrijdag

Xavier Vrijdag, MSc

Department of Anaesthesiology │ School of Medicine
Faculty of Medical & Health Sciences │ The University of Auckland
Private Bag 92019 │ Auckland 1142 │ New Zealand

M +64 21 0230 4558
E x.vrijdag at auckland.ac.nz<mailto:x.vrijdag at auckland.ac.nz>


fieldtrip mailing list

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20210129/24905c29/attachment.htm>

More information about the fieldtrip mailing list