In article <1991Nov1.013436.23545 at meteor.wisc.edu> stvjas at meteor.wisc.edu (Steve Jascourt) writes:
>In article <1991Oct28.051227.4611 at frodo.cc.flinders.edu.au>
>>We wish to assess the power of 40Hz activity in short period (100-200msec)
>>EEG records sampled at 1khz. The only spectral analysis tool available to
>>us is FFT based power spectrum calculation. I'm not at all confident
>>that this is the best approach, but have no way to assess this. Could
>>anyone *please* help us out ? I guess what I'm asking is whether
>>there exists a technique for measuring the power in a limited band, given
>>a restricted number of samples ?
>>-----------------------------------------------------------------------
>> _--_|\ James Tizard
>>The "maximum entropy" methods works wonders with short samples. It is
>outstanding at picking out key frequencies, but one shouldn't put too much
>stock in its exact values for power. In contrast, the FFT methods tend to
>be better at broad-band power and show broader features in their spectral
>estimates. The maximum entropy method basically is done as follows, with
>a fair number of variations on the theme used by various people:
> Take the autocorrelation of your time series.
> Do a (least squares) fit to the autocorrelation of an autoregressive random
> process. You will definitely not need an AR process higher than order 30,
> perhaps order 10 or 15 will suffice -- you can try it for various orders.
> Now plug into the formula for the spectrum of an AR process -- you just
> found the AR coefficients in the previous step.
>>Talk to people in signal processing (that's in electrical engineering). They
>spend their life solving the kinds of problems you present.
>>Stephen Jascourt
Look at, for example, Ronald R. Coifman and M. Victor Wickerhauser, _Best-
adapted wave packet bases_, Yale university, 1990. This sort of thing may
help. Remember that you have to contend with Heisenberg's uncertainty
principle; that is, for any measured signal, the uncertainty in time *
the uncertainty in frequency >= 1/2.
Matt Lundberg
