Re: Uncertainty on error access ..

Hello,

  By default the errors in Minuit are obtained from the inverse of the Hessian matrix. This is correct if the function has a quadratic form (a parabola) around the minimum. If not, in case of high non-linearity you should use Minos.
The errors from MINOS are correct in the limit of high statistics, i.e. when the profile likelihood ratio is distributed as a chi2 distribution.
For the interpretation of the Minos error see for example http://seal.web.cern.ch/seal/documents/minuit/mnerror.pdf

In general the uncertainty in the error can be obtained by generating several toy experiments, and fitting each one of them and then look at the distribution of the obtained errors.

In your case, the correct approach is to leave floating the other parameters and use as error the one obtained with the floating parameters.

  Best Regards

  Lorenzo
On Apr 16, 2009, at 10:38 PM, Arthur E. Snyder wrote:

> Hello,
>
> I guess this is really, underneath a minuit problem:
>
> How to you get access to the uncertainy in the error printed on the
> fit output?
>
> This can be substantial even when status assigned to the error
> matrix is "3" indicating all is well.
>
> I can't find any function that will let me obtain the uncertainty on
> the error within my root program (that's scanning fit).
>
> -Art S.
>
> PS. this wouldn't matter much normally, but in this case I'm taking
> the quadrature difference between nominal fit and fit with some
> other parameters released in order to get the contribution to the
> error from the released parameters. When the difference is small a
> 5-10% error can matter a lot ..
>
>
>
> A.E. Snyder, Group EC \!c*p?/
> SLAC Mail Stop #95 ((. .))
> Box 4349 |
> Stanford, Ca, USA, 94309 '\|/`
> e-mail:snyder_at_slac.stanford.edu o
> phone:650-926-2701 _
> http://www.slac.stanford.edu/~snyder BaBar
> FAX:650-926-2657 Collaboration
>
>
Received on Wed Apr 22 2009 - 10:10:18 CEST