[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Suggestions for gstat
(On Nov 5 I got the following mail from Konstantin; I'll cc:
the answer to gstat-info as well)
Konstantin Malakhanov wrote:
> I have two suggestions for gstat.
> 1) Because of UNISYS LZW patent, gnuplot is no more support gif output
> via gd library. Instead of gifs gd library produce now png files. You
> cand read about it at . Probably that mean that gstat should be
> changed at this point too.
I agree -- I will change gif support into PNG support. PNG is superior
anyway, and all major web browsers support it nowadays. Now let's see,
should this go into 2.1.1 or 2.2.0? Mmmm...
> 2) today I've made (countless time again ;-) a stupid error of using
> gaussian variogram without nugget effect. I know that it leads to
> numerical problems (singular matrix). But gstat says "use average,
> increase zero" etc. So that I was trying, without of course no success
> until I got an an idea to put a nugget effect into play. Maybe one can
> extent singular matrix message with something like that:
> "Another reason for this error could be using of Gaussing model
> without nugget effect. This have a theoretical background and could be
> solved by adding small nugget effect to your variogram".
> This problem is mentioned in GSLIB (with two references). But
> references are quire large and it would be akward to put it in error
> message. But maybe the same remark with references could be put into
> app. B2 of gstat manual.
A good suggestion again. I wasn't too happy with the current error
message (`hint:..') on a singular matrix. In 2.1.1. I will print a
message that only refers to the relevant section in the manual, and
I will try to make this section cover all possible options, not only
the most frequent ones. For instance, UK (or trend estimation) leads
to singularities when working with small neighbourhoods, when the
covariates (X) become dependent.
I have a bit trouble with uploading new versions, but I will try in
the coming weeks.