Ostman: I have not been working on that problem neither on 
how to estimate the parameters nor on the use of 
the parameters. The subject of my paper was for a 
given description of the terrain, for instance its 
autocorrelation function, to derive the most sig- 
nificant parameters. 
Jacobi: I think you are right. These methods drove from 
the theory that the terrain surface is a Gaussian 
process and the break lines can't be considered in 
this. I think we will have to have some structural 
models based on geo-morphology to take care of break 
lines. 
Ligterink: But I think in the terrain with break lines that 
those points are just the most important and most 
studies there are being made of the form of the 
terrain and so on. 
Rauhala : There is an interesting connection between Karhunen 
- Loewe expansion and finite element modeling that 
I have been using in my array approach to the compu- 
tational solution of the finite elements.  Namely, 
if you utilize Karhunen - Loewe expansion in two 
dimensions then the continuity constraints of the 
finite element method can be diagonalized. The 
resulting diagonal system is a function of wellknown 
eigenvalues of a tri-linear matrix. My approach 
for an adaptive and automated terrain classification 
is to use this property of K-L expansion for the 
average sample spacing and putting the deviating 
terms to the right hand side. Then this fast two- 
dimensional K-L solution can be utilized in each 
iteration cycle and the resulting system will be 
very closely related to what Prof. Ebner is going 
to report in the multi-grid method. There are also 
other applications for K-L expansion, e.g., in 
physical geodesy. I used K-L expansion for making 
the kernel function of Bjerhammar problem in physi- 
cal geodesy separable and then used the separable 
array solution to the computational least squares 
problem. In recent years I have more or less dis- 
carded Fourier transformations in favour of K-L, 
because there is an extremely efficient way of 
implementing K-L in software and hardware. The 
resulting computer performs one billion arithmetic 
operations per second for an adaptive finite element 
solution of millions of parameters in a second. 
Fórstner: I have a question and a comment. Did you have to 
take the trend of your terrain into account or did 
you work with the original data? And a comment to 
the break lines. I think there are two aspects of 
describing the terrain, one is the deterministic 
part and the other the stochastic part. The de- 
terministic, that is to say the break lines which 
have to be structured and this is more or less a 
63