Re: [ROOT] Convergence of minuit

From: Topher Cawlfield (cawlfiel@uiuc.edu)
Date: Thu May 01 2003 - 05:47:59 MEST


Hi,

I don't have any experience with Minuit from root, but lots of 
experience with Minuit as a stand-alone.  I'm responding because your 
messages look very familliar, and often don't really indicate a problem. 
  Apart from the invalid errors, do the fitted parameters seem 
reasonable?  If so, you might try to recompute the errors some other 
way.  Often asking for Minos errors solves that problem, though there 
may be a faster way if you're worried about doing hundreds of thousands 
of fits.

I see that TMinuit::mnmnos() is provided as a hook to compute Minos 
errors, which can be retrieved by mnerrs(...).  Minos will scan your 
chi^2 function to find +1 unit from the minimum on both sides, and 
return asymmetric errors.  This is valuable for those cases where either 
the chi^2 really isn't parabolic at the minimum or, more often, when 
Minuit fixates on some tiny numerical roundoff glitch which wrecks the 
covariance matrix.

  - Topher

Michael Groys wrote:
> Dear rooters,
> I'm using TMinuit class with some complex theoretical function
> for minimization of chi2 for set of data points.
> Often I get following messages:
>   MIGRAD FAILS TO FIND IMPROVEMENT
>   MIGRAD TERMINATED WITHOUT CONVERGENCE.
> Changing of initial values and errors doesn't help.
> What can be done to avoid this problem?
> I'am interesting in parameter errors.
> But errors that minuit returns in this case are invalid.
> Thanks,
> 	Michael
> 
> 



This archive was generated by hypermail 2b29 : Thu Jan 01 2004 - 17:50:11 MET