[FieldTrip] How to impove dics realfilter solution

[jan-mathijs schoffelen]

Dear Björn, caro Gio,

I don't agree with you, Gio, that one should use complex-valued filters. The reason for this is that they don't make sense from a biophysical point of view. The mixing from sources to channels (as represented in the leadfields) is strictly real-valued, because it is (near-)instantaneous. I think that the unmixing from channels to sources also should be real-valued, for the same argument. When you allow filter coefficients to be complex-valued, this is equivalent to allowing the sensor-level signals to be shifted in time with respect to one another. I consider this a bit odd. In Joachim's seminal paper (on which FieldTrip's initial DICS implementation is based), this was not really explicit in the methods description. However, based on personal communications with said Joachim, I at least distilled that the filters should be real-valued. The complex-valued filters were kept in (as default) for backward compatibilitiy reasons. Although mathematically possible (perhaps even more strictly answering to the minimum variance and unit gain constraints), and sometimes seemingly giving nicer subjective results, I'd advice against using them.

Cheers,

JM



On Sep 25, 2012, at 12:21 PM, Gio Piantoni wrote:

> Hi Björn,
> 
> It's not surprising that the results are pretty bad after using real
> filters. The fourier transform and DICS are intrinsically
> complex-valued, so using only the real part means throwing away half
> of the information (and getting meaningless results).
> 
> The realfilter option was implemented in order to allow to pass the
> time-domain signal through the complex-value spatial filter and obtain
> a "virtual electrode". A virtual electrode with complex values is
> "weird to say the least", see:
> http://mailman.science.ru.nl/pipermail/fieldtrip/2006-May/000510.html
> Robert considered it an error in the algorithm and fixed it by
> providing the realfilter option. However, to me it seems more correct
> to think that complex spatial filters requires complex input (i.e.
> frequency-domain signal), so don't use DICS spatial filters on
> time-domain real-valued MEG recordings to generate time-domain virtual
> electrodes. If you want time-domain virtual electrodes, use LCMV
> spatial filters centered at the location of interest. Other smart
> people might disagree though.
> 
> I suggest not to use realfilter option at all, but I'd be happy to
> hear other meaningful applications of this option.
> 
> Hope this helps,
> 
> Gio
> 
> On Fri, Sep 14, 2012 at 10:23 AM, Björn Herrmann <bherrmann at cbs.mpg.de> wrote:
>> Dear fieldtrip user,
>> 
>> I am using the DICS beamformer to localize a 3Hz FFT power topographical distribution (using a neuromag system). The topographies are very clear auditory in single subjects, thus it shouldn't be a problem finding the "right" solution. When I use the beamformer to obtain complex-value spatial filters, the solution looks very good (using NAI). No problem here. However, using cfg.dics.realfilter = 'yes', i.e. the spatial filter will have real values, the localization completely breaks down, sometimes even finding a strong peak in the opposite hemisphere than visible in the sensor topography (the head is centered in the sensor array).
>> 
>> Does someone have a suggestion how to improve the solution using realfilter = 'yes'. I use the following settings. They are rather standard I guess.
>> 
>> cfg = [];
>> cfg.method            = 'dics';
>> cfg.frequency         = 3;
>> cfg.dics.projectnoise = 'yes';
>> cfg.dics.lambda       = '5%';
>> cfg.dics.keepfilter   = 'yes';
>> cfg.dics.realfilter   = 'yes';
>> cfg.dics.fixedori     = 'no';
>> 
>> 
>> If someone has a suggestion, I would appreciate it.
>> Thanks in advance,
>> best,
>> 
>> Björn
>> 
>> 
>> 
>> _______________________________________________
>> fieldtrip mailing list
>> fieldtrip at donders.ru.nl
>> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip
> 
> _______________________________________________
> fieldtrip mailing list
> fieldtrip at donders.ru.nl
> http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

Jan-Mathijs Schoffelen, MD PhD 

Donders Institute for Brain, Cognition and Behaviour, 
Centre for Cognitive Neuroimaging,
Radboud University Nijmegen, The Netherlands

Max Planck Institute for Psycholinguistics,
Nijmegen, The Netherlands

J.Schoffelen at donders.ru.nl
Telephone: +31-24-3614793

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20120927/275a435f/attachment.html>