Width of the transition band when applying a Hilbert transform

✉️ Mailing List Archive (read-only)
1

Hello,

My question relates more to a signal processing issue, so apologies in
advance if it is not appropriate for the forum.
I am currently doing an exploratory EEG power analysis using the
filter-Hilbert transform and would appreciate some advice on the
parameterization of the band-pass filter. For each epoch and time step, I
want to compute the power of each frequency from 1 to 30 Hz (in steps of
1).
I am thus bandpass filtering the data prior to the application of the
Hilbert transform (so I bandpass filter from 1 to 2Hz, then from 2 to 3hz
etc. ).
Mne's FIR filter implemented via mne.filter.filter_data is perfect for this
purpose, but I am wondering if I should use the default automatic width of
transition band, or whether i should set l_trans_bandwidth and
h_trans_bandwidth
to some fixed value x across all frequency bands. If so, how should I set x
to achieve a good balance between frequency/timing precision ?

Best,
J.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20180120/c5561245/attachment.html

2

hi J,

may I ask what you are aiming to do afterwards with these filtered data?

Alex

3

Hello,
I mainly want to estimate theta, alpha, and beta power in the 200-600ms
postimulus window. Note that I could bandpass filter the data in the
corresponding ranges (e.g., theta: 4-7).

averaged power in a specific temporal window (200 600ms poststimulus) for
theta bvand, alpha band, and beta band.

4

I would fix the transition band widths otherwise you will have different
filters per band. In one of our examples we do something similar.

Denis

5

hi,

how about starting with something like this:

http://martinos.org/mne/stable/auto_examples/time_frequency/plot_time_frequency_global_field_power.html

?

HTH
Alex

6

Yes, this is the example I had in mind.

7

Hi J,

Note that 1 Hz wide filters are very narrow, especially if you want to
analyze only a 400 ms window. The higher your frequency resolution, the
lower your temporal resolution.

Best,
Ross

8

Hi Ross,

I agree. I set the transition band to 2 Hz, but I don't know if that will
provide sufficient temporal resolution. Is there some analytic way to
determine the temporal/spectral precision based on the width of the
transition band?
Best,
J

9

Unfortunately there's no hard and fast rule. In general you should filter
as "gently" as possible, since impulse responses can get shockingly long
with steeper slopes in the frequency domain. The best way to know how
spread in time your filtered signal is going to be is probably to look
directly at the filter impulse response (or the envelope of the IR). That
will give you and idea of how much the estimate at one time point depends
on the time points nearby.

Another good thing to do is to make some synthetic, known input signals and
process them to see how your analysis affects them.

Hope that helps,
Ross