noise covariance

✉️ Mailing List Archive (read-only)
1

hey all

So, I inherited several resting state datasets but empty room data were not collected for any of them. I was wondering about your experiences analyzing resting state in source space using an identity noise covariance matrix.

Some questions that come to mind:

- any suggestions for regularization levels?
- if I'm interested in frequency domain, what about constructing my noise matrix with the filtered signals of a non-interesting band (i.e. band-stop on the frequency band of interest)?
- say I'm interested in the usual 4-5 frequency bands, maybe using the same noise covariance that's constructed using high frequency signals (> 100Hz, outside of the bands of interest)?
- how would the answers above change in the context of a MNE vs LCMV beamformer analysis?

Thanks in advance,

G

2

Hey Gus,

So, I inherited several resting state datasets but empty room data were not collected for any of them. I was wondering about your experiences analyzing resting state in source space using an identity noise covariance matrix.

I have no such experience.

Some questions that come to mind:

- any suggestions for regularization levels?

I would take a noise cov with that machine from another experience
and make it diagonal. So the order of magnitudes should be ok.

- if I'm interested in frequency domain, what about constructing my noise matrix with the filtered signals of a non-interesting band (i.e. band-stop on the frequency band of interest)?

I would clearly not do this. The amplitude/structure of noise depends
on frequencies.
If I can and make sense for the question, I tend to use a noise cov per band.

- say I'm interested in the usual 4-5 frequency bands, maybe using the same noise covariance that's constructed using high frequency signals (> 100Hz, outside of the bands of interest)?

there is very little noise at these frequencies. I doubt it would be do the job.

- how would the answers above change in the context of a MNE vs LCMV beamformer analysis?

good question.

HTH
Alex

3

Hi Gustavo,
   I don't have any experience attempting localization without a noise
covariance.. Perhaps people who have attempted resting state
localization with EEG (where there is no good analogous "empty brain")
might have some insights.. However, I thought I'll loosely comment on
some of your questions.. Please see inline responses below.

Otherwise, my recommendation would be to get hold of some empty-room data
set from the particular scanner as close as possible to the dates on which
the data were acquired.

If that's not possible, perhaps finding a way to limit the rank of the
noise-cov to some reasonable but conservative number (like 50-80 for the
Neuromag system) might help. I'm not sure what would be the nicest/easiest
way to do that.. Maybe others have some insights..

Hari

hey all

So, I inherited several resting state datasets but empty room data were
not collected for any of them. I was wondering about your experiences
analyzing resting state in source space using an identity noise covariance
matrix.

Some questions that come to mind:

- any suggestions for regularization levels?

I'm assuming you mean regularization of the inverse (i.e., lambda^2),
rather than regularization of the noise-covariance? This could just be
whatever SNR assumptions you normally make when you do have access to
empty-room data..

- if I'm interested in frequency domain, what about constructing my noise
matrix with the filtered signals of a non-interesting band (i.e. band-stop
on the frequency band of interest)?

Intuitively, this doesn't seem like a good idea because the noise in the
different frequency bands don't have any particular reason to be related..
In fact, if the noise is stationary (wide-sense), the frequency domain
values you get in one band are necessarily uncorrelated with the frequency
domain values in another band. Also, in practice, often, high frequency
noise is less correlated spatially than low frequency noise.. This is
apart from high-frequency noise being much smaller than low-frequency
noise.

- say I'm interested in the usual 4-5 frequency bands, maybe using the
same noise covariance that's constructed using high frequency signals (>
100Hz, outside of the bands of interest)?

See above..

- how would the answers above change in the context of a MNE vs LCMV
beamformer analysis?

I don't have much experience using the LCMV beamformer, but I can't think
of why the choice of noise covariance should be different..