"Karl" <karlknoblich at yahoo.de> wrote in message
news:235b9607.0401210500.3ebedda5 at posting.google.com...
>> I want to calculate the correlation dimension of a time serie.
>> What I have done
> I calculated the correlation integral C(r) (number of point having a
> distance smaller than r) for different embedding dimensions. Taking
> the slopes of the curve of log C(r) against log r for the different
> embedding dimensions and plotting them against the embedding dimension
> should result in a limes of the slopes: the correlation dimension.
>> My problem
> Which slope shall I take?
>> In examples I saw in text books there is a nice limit of the slopes
> with higher embedding dimensions. In my data I do not know which slope
> I should take because the slope of the curve varies. If I take the
> slope at a certain value of log r I can not get a limes.
>> My curves (log C(r) against log r) can be seen in
>http://karlknoblich.4t.com/korrdim.jpg>>> What to do? Does anybody knows such data and how to handle it?
>> Hope somebody can help!
Value of C.D. depends on several factors: level of data stationarity,
the length of a stationary segment, external noise, sampling
rate, and embedding dimension.
For example, if we choose "stationary" region of time series, the length of
region must be about 10^(D+2) points. Therefore to
determine D=10 we need about 10^6 points of suficiently stationary
I guess you are trying apply the method of C.D. to EEG series,
if so, you'll be disappointed unless you can find a subject with
sufficiently (see above)
long segment of quasi-stationary data.
Some time ago I got similar results for EEG series.
If you want, you can check the paper at:
Hope this help,