Edzer J Pebesma a ?crit :> Patrick Giraudoux wrote:
>
>> I wonder what is the default value for the argument 'cutoff'
when not
>> specified in the variogram.formula function of gstat. Computing
>> variogram envelops within gstat, I am comparing the results obtained
>> with variog in geoR and variogram in gstat, and it took me a while
>> before understanding that the cutoff default value is not the maximum
>> distance.
>>
>> Can Edzer tell us about it?
>
> Yes, of course :
>
> the default value is computed in the c code. Without checking
> (meaning: from >10 years memory) I do recall that gstat uses
> one third of the diagional of the rectangular (or block for 3D)
> that spans the data locations.
>
> Why? In time series you compute ACF's up to one half
> of the length of the series; after this things start to oscillate
> because you lack independent replication at large distance;
> look at what is meant by ergodicity for further reading.
> Variograms are basically flipped & unscaled acf's for higher
> dimensions.
>
> Some books (Journel & Huijbregts?) gave suggestions
> that half the max. distance in the data set is a good guideline,
> back in 1978. I used one third of the diagonal because
> I thought finding the maximum distance between any
> too point pairs may be expensive to find for large data
> sets. The parameter "one third" can be overridden by
> those who don't like it.
>
> Please keep us updated about your milage comparing
> gstat and geoR; I once spent an afternoon on this, trying
> to reproduce sample variogram across the packages and
> found this hard (but not impossible). I had the feeling
> it had to do with using < or <= to decide whether a
> point pairs falls in a distance interval or not, but didn't 100%
> assure myself.
> --
> Edzer
>