Skip to content

Gaussian Variogram Positive Definite?

2 messages · Brian J. Lopes, Paulo Justiniano Ribeiro Jr

#
Hello All:

I've been banging my head against the wall about this for quite some 
time now, and I can't seem to find any reference on the matter.  I'm 
trying to calculate MLE estimates for the Gaussian variogram, but it 
seems that I consistently reach the point where the covariance matrix is 
not positive definite, as dictated by the Cholesky decomposition, even 
though the range parameter is indeed positive (note that I am also 
incorporating a nugget to sill ratio as well).  Has anybody else 
experienced this problem?  Better yet, does anybody have any references 
that discuss the situation, or how I can avoid it?

The data is a bit large, so if an example is necessary I'll try to see 
if I can come up with something reasonable.

Thanks,
Brian
#
Brian

Gaussian variograms are known to generate numeric problems
in case you have the nugget parameter equals to zero.
This occours because the almost flat
and with points very close to each other the covariance matrix will be
nearly-singular -- numerically singular.

Soma alternatives are:
1. choose another covariance model:
   for instance a Matern model with smoothness parameter  to ensure a
behaviour which is similar to the gaussian (e.g. kappa = 4 in the
parametrisation used in geoR

2. add a small nugget to the model to make the covariance matrix
diagonally dominant

hope this helps

best
P.J.




Paulo Justiniano Ribeiro Jr
LEG (Laborat?rio de Estat?stica e Geoinforma??o)
Universidade Federal do Paran?
Caixa Postal 19.081
CEP 81.531-990
Curitiba, PR  -  Brasil
Tel: (+55) 41 3361 3573
Fax: (+55) 41 3361 3141
e-mail: paulojus AT  ufpr  br
http://www.leg.ufpr.br/~paulojus
On Fri, 31 Aug 2007, Brian J. Lopes wrote: