Yong Zhang wrote:
> reasonable. The problem is, when looking at the plot, it is quite
> obvious that the second set of data has higher fluctuation than the
> first set, but the error bar provided by gnuplot was quite close for
> both fitted T and b for the two set of data. Could you please provide
> any explanation about this results?
For somebody
The fit parameter errors are *not* the only thing you should be looking
at to evaluate the quality of a fit. You must check the residuals, too,
i.e. what gnuplot displays as "stdfit", i.e. the sqrt(chisquare/ndf).
In the case at hand, the fit to f1.dat has a much better stdfit figure
of ~1.0, compared to the ~2.4 you get for f2.dat. In a different view
at the same result, 'fit' is telling you by these numbers what the
variance of the data relative to the fitted function turned out to be,
and in particular, that the second fit is indeed worse than the first.
gnuplot's "fit" command, unlike some other fitting programs, does not
scale parameter errors with "stdfit". You can see the difference if you
actually supply data errors to the fit, by scaling them up and down and
observing how various fitting programs change their reported results.
