Sampo Niskanen wrote:
> I haven't compared the AR and MA methods either. =A0I'm no DSP expert
and
> once I had a working filter that produced good enough results I went
> back to my original project. =A0:) =A0In fact, I decided to use only a
> two-pole filter in my application since I didn't want the values
> deviating from zero for long periods of time.
Hi Sampo
This sounds a bit contradicting. Either you are intersted in 1/f^{5/3}
noise, or you are worried about values in the series deviating from
zero for long periods.
You can't have both.
This deviaton for longer periods from zero is one of the very
characteristics that makes 1/f^alpha noise "statistically difficult"
to handle. This phenomenon was called the "Joseph Effect" by
Mandelbrot, because of Joseph's dream that Egypt was to experience
seven years of plentiful harvests followed by seven years of famine
(long periods of deviation from the zero mean). Particularly
climatological and hydrological time series can be modeled very well
by 1/f^alpha noise. The Nile river data has an estimated Hurst
exponent of H =3D 0.91, which translates to an alpha of 0.82. Note that
your alpha is > 1, indicating non-stationary long range correlation.
For that case long deviations from zero is the rule, not the
exception.
Just a thought.
> Anyway, I hope the code is useful to others as well.
I'm sure it is :-).
Regards,
Andor
|
