🚀 go-pugleaf

In article <1at2faINN9he@network.ucsd.edu>jasmith@ucsd.edu (jerry smith)
writes : 
> .....
>Covariance-based methods are often referred to as "auto-regressive"
models
>(or AR-spectra;  I can't say why this name has stuck).  Marple reviews
the
>relative advantages of AR versus "periodogram" (fft) and a few others
>(e.g.  Proni's method).
> .....
I would not use AR methods on many oceanographic data sets. AR methods
work well for relatively high resolution detection of close peaks.
For instance, identifying both a tidal and an inertial peak in data
collected at mid-latitudes. An AR(1) filter or Markov model can either
act as a high or low pass filter, while an AR(2m) model is used for
detecting m sinusoids. Many oceanographic data sets are red and have
significant energy in a broad mesoscale range. AR methods which use
the sample covariance/autocorrelation were not formulated for this
type of spectra. They work well in speech applications and for detecting
sinusoids in white noise. To quote from Thomson (1990), "If your
spectrum
estimate explicitly requires sample autocorrelations you are almost
certaintly doing something wrong." My suggestion is to read this
paper and his earlier work before you do anything.  cheers, arthur

Thomas, D.J., 1990: Quadratic-inverse spectrum estimates: applications
to palaeoclimatology, Phil. Trans. R. Soc. Lond., 332, 539-597.


-- 
*****************R.I.P., Bill of Rights, 1791-1991********************
Arthur J. Mariano INET: mariano@sgrmag.rsmas.miami.edu 
Dept. of Meteorology and Physical Oceanography        305-361-4193 
Rosenstiel School of Marine and Atmospheric Science,  U. of Miami
4600 Rickenbacker Causeway, Miami, FL 33149-1098, USA