Re: dependence of fit with previous one

From: Lorenzo Moneta <Lorenzo.Moneta_at_cern.ch>
Date: Tue, 1 Mar 2011 14:58:34 +0100

On Mar 1, 2011, at 2:26 PM, Marc Escalier wrote:

> sorry,
> i'm not sure to understand
> 
> i call exactly the same function for the two fits and i reinitialize the parameters and their errors
> only the histogram is not the same

I don't understand, if the histogram is different, the likelihood function will be different....

> 
> but i see that the digits are not exactly the same
> -->
> how to fix the "numerical error in the function evaluation" ?
> 
> i mean the result is not reproducible if one does a previous fit : it sounds problematic, isn't it ?

It is not a problem as long as the numerical error is much smaller that your statistical errors. For example you can get a different likelihood value if you are using a different order in summing the contributions, although the contributions are all the same.

> 
> >It could be also you are using a MC integration in the function evaluation. In this case this can happen.
> sorry : what is a Monte-Carlo Integration ?

See http://en.wikipedia.org/wiki/Monte_Carlo_integration In this case, since random number are used the result of the integral will be different if you are using different random seeds.

>
> (i just used : myhisto->Fit("LV");

Fine, but what is your function ?? It could be a very complicated object doing integrals or whatever....

Lorenzo

> 
> thank for any help
> 
> =======
> 
> Lorenzo Moneta a écrit :
>> Hi Marc,
>> 
>> The fit is re-initialized correctly when you set the parameters and the errors. In your case, it is probably your function which returns a different result given the same parameters probably due to a numerical error in the function evaluation. 
>> I can see that you have a numerical problem in evaluating your function to minimize by seeing  this error message in the log file:  MIGRAD FAILS TO FIND IMPROVEMENT
>> MACHINE ACCURACY LIMITS FURTHER IMPROVEMENT.
>> 
>> It could be also you are using a MC integration in the function evaluation. In this case this can happen.
>> 
>> Cheers, 
>> Lorenzo
>> On Mar 1, 2011, at 1:08 PM, Marc Escalier wrote:
>> 

>>> Hello,
>>>
>>> i observed a dependence of a given fit the previous one, *even* when i reinitialize each of the parameters and their errors
>>>
>>> -->is there a way to "reinitialize" the fitter to have reproducibility one one do some previous (or not) fits ?
>>>
>>> thanks a lot
>>>
>>> -->here is a log of the fits
>>> http://users.lal.in2p3.fr/escalier/ProblemRoot/
>>>
>>> it begins to change here :
>>> with only one fit :
>>> 2 CB_mean 1.40000e+02 6.00000e-01 2.01358e-01 2.51098e+02
>>>
>>> with *a* previous fit before (and after having reiniatzed the parameter by SetParameter and SetParError) :
>>> 2 CB_mean 1.40000e+02 6.00000e-01 2.01358e-01 2.51095e+02
>>>
>>> at the end of the fit : it gives
>>> with only one fit :
>>>
>>> 1 A 6.19000e+02 fixed 2 CB_mean 1.39795e+02 3.64629e-02 3 CB_sigma 1.95380e+00 3.00964e-02 4 CB_alpha 1.22909e+00 4.12433e-02 5 CB_n 1.00000e+01 fixed 6 Gauss_mean 1.44687e+02 7.63561e-01 7 Gauss_sigma 2.57589e+00 4.29212e-01 8 frac_CB 9.75035e-01 5.66139e-03
>>> with *a* previous fit before (and after having reiniatzed the parameter by SetParameter and SetParError) :
>>>
>>> 1 A 6.19000e+02 fixed 2 CB_mean 1.39795e+02 3.72340e-02 3 CB_sigma 1.95381e+00 3.07984e-02 4 CB_alpha 1.22910e+00 4.12501e-02 5 CB_n 1.00000e+01 fixed 6 Gauss_mean 1.44688e+02 7.55733e-01 7 Gauss_sigma 2.57555e+00 4.30838e-01 8 frac_CB 9.75039e-01 5.48416e-03
>>> -->some digits are not exactly the same
>>>
>>> thanks
>>>

>>
> Received on Tue Mar 01 2011 - 14:58:39 CET