Re: [Gnuplot-bugs] bugs using fit

[Hans-Bernhard B. <br...@ph...>]

On Fri, 4 Jun 2004, Wanderson Wanzeller wrote:

[...]
> Final set of parameters            Asymptotic Standard Error
> =======================            ==========================
> 
> a               = 0.000550667      +/- 0.001319     (239.5%)
> b               = 27.5771          +/- 4.699        (17.04%)
> c               = 0.0268601        +/- 0.009831     (36.6%)

These results contain hints of significant trouble already: notice the
relative error on 'a' being extremely large?  That's because 'a' and 'b'
are so hugely different (by 5 orders of magnitude) that it hurts the
performance of the fitting code.  You had better redefine the parameters
to make them have roughly the same size.

[...]
> gnuplot> fit [1:16] g(x) 'omegQ_25bt.dat' using 1:2:3 via a,b,c
[...]
> Final set of parameters            Asymptotic Standard Error
> =======================            ==========================
> 
> a               = 0.000550667      +/- nan          (nan%)
> b               = 27.5771          +/- nan          (nan%)
> c               = 0.0268601        +/- nan          (nan%)

The reason for this is evident if you look into the datafile: the first
few datapoints claim to have an error of exactly zero.  That's physically
impossible, and disrupts the result.  You'll have to remove those
datapoints from the fit, either by changing the file itself or by 
using datapoint selection methos offered by gnuplot (the 'every' option, 
a restriction of the x or y range, or a filtering 'using' specification).


-- 
Hans-Bernhard Broeker (br...@ph...)
Even if all the snow were burnt, ashes would remain.