Same histograms and same code but different fit results plotted in timelines

[hauenst]

During RGB pass0 timelines an interesting bug/feature/issue was observed. The two timelines https://clas12mon.jlab.org/rgb/pass0/v28.21_sqlite/dc/timeline/ and https://clas12mon.jlab.org/rgb/pass0/v28.21_mini/dc/timeline/ have the same histogram inputs but some of the DC timelines are different (dc_residuals_sec_sl mean/sigma, dc_residuals_sec mean/sigma). This is due to a different fit result stored in the files. However, the same monitoring/timeline code was used by Zhiwen and the underlying histograms are identical.

[drewkenjo]

This is related to the fitting stabilities of groot. I think @Sangbaek did the best he could with guessing starting parameters, and performing fits multiple times to choose best chi2 result.
If you have any suggestions how to improve it we can discuss it.

[hauenst]

We do perform the fit multiple times in DCfitter for the single Gaussian distribution with the same set of starting parameters. The best chi2 result is chosen. I am still surprised that this doesnt work in the same way twice on the same histogram.
But if it is a problem of groot then we cannot do anything besides documenting it so that no-one else is surprised in the future

[drewkenjo]

I would even argue it's a problem of Minuit implementation in Java.
@raffaelladevita what do you think?

[baltzell]

I remember with HPS thinking I found using jminuit directly to be always reproducible given the same starting conditions, but maybe not.

If we have a minimal test case people can easily run, doing a fit multiple times with the same initial parmaeters and getting different results, that might be helpful.

[raffaelladevita]

My experience with groot and jminuit is also that results are reproducible. Yes, we need an easy way to run this fit and do some debugging

[drewkenjo]

@hauenst could you provide such an example for people to test?

[hauenst]

well my groovy knowledge is rather poor. I am not sure where to start to also include the reading of the monitoring histograms. I think dc_residuals_sec_sl.groovy should be a simple example. One can just replace the storing of sigma and means in a list to storing it in a text file or printing them out and probably need to add a string of the input file name.

Like Nathan and Raffaella wrote, I would think jminuit itself is reproducible on the same histogram but the interplay between reading histograms for multiple sec/sl, fitting them and storing it might cause the problem