[R] Estimated covariance matrix with tgp package

[Jochen Fiedler]

Hello everyone,

at the moment I'm using the tgp package for modelling a nonstationary 
data set on a two dimensional area D and I'm interested in the 
prediction and the estimated covariance matrix. For this purpose I'm 
using the function btgp. As far as I understand, btgp uses a MCMC 
algorithm to split up D along lines parallel to the coordinate axes and 
estimates independent Gaussian processes on each resulting region, 
conditional on the predefined class of covariance functions (like the 
Matern class) and returns a tgp class object, which I call 'out'.
Now I'm wondering if out$Zp.s2 gives the estimated covariance matrix. In 
my example I have a matrix of locations X and data set Z. When I run the 
btgp function it yields a tree of hight two, indicating that the 
algorithm splits up the area D into two independent regions and 
correspondingly the matrix of locations X. But if I look at out$Zp.s2 
the matrix has no zeros, which implies dependence of both regions. So I 
don't understand what Zp.s2 gives exactly and how I can get the 
estimated covariance matrix.

A similar question is what ZZ.s2 gives? The manual says that this yields 
the predictive covariance matrix at some predictive locations XX. But if 
I put X=XX it doesn't yield the same as Zp.s2.

So I would be grateful if someone could help me.

Best regards

Jochen Fiedler