Dear Isobel:

In the case of intrinsic random functions the covariance (non ergodic covariance) and 
the
correlogram are artifacts (see my example in ai-geostats).

I prefer to use the variogram. If you whish to have a symetric system, with no 
problems of pivoting,
use the "Matheron System", and your program runs faster!

Proof:

The kriging system is (take a as lambda, g as gamma):

a1g(1,1)+a2g(1,2)+...+aNg(1,N)+mu=g(1,0)............(1)
a1g(2,1)+a2g(2,2)+...+aNg(2,N)+mu=g(2,0)............(2)
........................................
a1g(N,1)+a2g(N,2)+...+aNg(N,N)+mu=g(N,0).........(N)
a1+a2+...+aN=1.......................................................(N+1)
______________________________________________________

Now I sustract the equation (N) from the others (mu vanishes):

a1(g(1,1)-g(N,1))+a2(g(1,2)-g(N,2))+...+aN(g(1,N)-g(N,N))=g(1,0)-g(N,0)
a1(g(2,1)-g(N,1))+a2(g(2,2)-g(N,2))+...+aN(g(2,N)-g(N,N))=g(2,0)-g(N,0)
......................................................................
a1(g(N,1)-g(N,1))+a2(g(N,2)-g(N,2))+...+aN(g(N,N)-g(N,N))=g(N,0)-g(N,0)
_______________________________________________________________________

The last equation is an identity, then you have only N-1 equations. Now
I take (to eliminate the weight  aN):

aN=1-a1-a2-...-aN-1 and replace in the others equations to get the
"Matheron System":

a1R(1,1)+a2R(1,2)+...+aN-1R(1,N-1)=B(1)...........................(1)
a1R(2,1)+a2R(2,2)+...+aN-1R(2,N-1)=B(2)...........................(2)
.......................................
a1R(N-1,1)+a2R(N-1,2)+...+aN-1R(N-1,N-1)=B(N-1).........(N-1)
_________________________________________________________

with: R(i,j)=g(N,j)-g(i,j)+g(i,N)-g(N,N)
      B(i)=g(N,0)-g(i,0)-g(N,N)+g(i,N)

(for the kriging variance you have a very nice formula, cf. (2))

The system is symetric R(i,j)=R(j,i), and the diagonal term
R(i,i) is never 0.

This is a system with only N-1 equations!! (not with N+1 equations).
If you use this system, your kriging program runs faster!(I remember
that the time to solve a linear system of n equations is exponential
with n, cf (1)).

References:

1. M. Alfaro: Nouveaux R�sultats sur la vitesse des calculs
g�ostatistiques. Journ�es de G�ostatistique, Fontainebleau, 1991.
2. G. Matheron. Cours de G�ostatistique. Fascicule 1. Ecole des
Mines de Par�s, 1965.
3. A. Mar�chal: El sistema de Matheron en Krigeage Universal.
Bolet�n de Geoestad�stica, Chile, 1972.


Best regards

Marco






Isobel Clark wrote:

> > It is well known that when inverting a
> > matrix it is much better (for numerical reasons)
> > that the higher values
> > are on the diagonal and the lower values far off the
> > diagonal.
> Have you not heard of "pivoting"?
>
> The computational "problems" of using a matrix based
> on the semi-variogram rather than the covariance are
> removed totally by putting the last equation first in
> the matrix. Of course, this means you cannot use a
> computational algorithm which demands a symmetric
> matrix, but so what?
>
> Alternatively, you pivot on the largest term in the
> first equation, then the second etc. I did a lot of
> experimentation with this around 15 years ago and
> found that the two (covariance and semi-variogram)
> sets of equations become identical after around the
> second or third 'pivot'.
>
> Please let us not confuse programming problems with
> geostatistical problems.
>
> There are a lot of packages out there which ask you to
> model the semi-variogram and then use a covariance for
> kriging. There are a lot of packages out there which
> model the semi-variogram and krige with the
> semi-variogram. Are there main stream geostatistical
> (as opposed to statistical or strict GIS) packages
> which model the covariance?
>
> Isobel Clark
> http://geoecosse.bizland.com
>
> ____________________________________________________________
> Do You Yahoo!?
> Get your free @yahoo.co.uk address at http://mail.yahoo.co.uk
> or your free @yahoo.ie address at http://mail.yahoo.ie
>
> --
> * To post a message to the list, send it to [EMAIL PROTECTED]
> * As a general service to the users, please remember to post a summary of any useful 
>responses to your questions.
> * To unsubscribe, send an email to [EMAIL PROTECTED] with no subject and 
>"unsubscribe ai-geostats" followed by "end" on the next line in the message body. DO 
>NOT SEND Subscribe/Unsubscribe requests to the list
> * Support to the list is provided at http://www.ai-geostats.org


--
* To post a message to the list, send it to [EMAIL PROTECTED]
* As a general service to the users, please remember to post a summary of any useful 
responses to your questions.
* To unsubscribe, send an email to [EMAIL PROTECTED] with no subject and "unsubscribe 
ai-geostats" followed by "end" on the next line in the message body. DO NOT SEND 
Subscribe/Unsubscribe requests to the list
* Support to the list is provided at http://www.ai-geostats.org

Reply via email to