Re: more on eof calculus and missing values

[Mick Spillane <spillane@xxxxxxxxxxxxx>]

Hi Antonio,
	The total absence of data at a grid point (as in the case of land)
is not a problem (as Ansley has just pointed out.).  As far as I know
either the covariance matrix or svd approachs would both begin with the
selection of the time series, assigning them to an arbitrary 1-D "space"
grid.  Knowledge of the spatial arrangement of the data only becomes an
issue when the graphical representation of the eigenvectors is done.
	The issue of some of the time series having gaps is more serious.
I'm not an expert on eof/principal component analysis but I think the
standard singular value decomposition algorithms would need gap free
arrays.  So to use svd one would need to bridge any gaps in the individual
time series beforehand.  If the gappiness of the data were slight and an
appropriate time series degapping were done one would hope that the two
methods would provide similar results.  But if there were extensive
gappiness one could see that gap filling involving neighboring stations
might bias the eigenvectors that result - different kind of bias than
that involved in accepting an imperfect covariance matrix. The saying
"you pays your money and takes your chances" comes to mind.
	I hope someone with more knowledge of the issue will contribute
and if someone has an SVD algorithm as an external function they can
share ... eof_space et al, while easy to use, can be slow for large
grids.

Mick

|____Mick.Spillane@noaa.gov____|
|__Room 2070 Bldg#3 NOAA/PMEL__|
|____Phone_:_(206)526-6780_____|