2.6.5: Uncertainty in TopDrawer file doesn't match the one in summary

Asked by Andrej Saibel

Dear all,

Coming back to https://answers.launchpad.net/mg5amcnlo/+question/673019

I'm using 2.6.5 FO analysis (LO) with TopDrawer, 0.0001 = req_acc_FO, and 4 cores. The uncertainties in the summary of the run and the TopDrawer don't match.

--------------------------------------------------------------
      Final results and run summary:
      Process p p > t t~ h [QCD]
      Run at p-p collider (6500.0 + 6500.0 GeV)
      Total cross section: 4.078e-01 +- 3.5e-05 pb
   --------------------------------------------------------------

while in MADatNLO.top, it says:

 ( total rate
 ( INT= 7.049E-02 ENTRIES= 33532736
    0.100000D+01 0.407780D+00 0.275229D-04
    0.200000D+01 0.000000D+00 0.000000D+00
    0.300000D+01 0.000000D+00 0.000000D+00
    0.400000D+01 0.000000D+00 0.000000D+00
    0.500000D+01 0.000000D+00 0.000000D+00

Could you please have a look?

Kind regards,

Andrej

Question information

Language:
English Edit question
Status:
Solved
For:
MadGraph5_aMC@NLO Edit question
Assignee:
Rikkert Frederix Edit question
Solved by:
Andrej Saibel
Solved:
Last query:
Last reply:
Revision history for this message
Rikkert Frederix (frederix) said :
#1

Dear Andrej,

Thanks for your message.
The behaviour you are seeing is to be expected. The way the uncertainty is estimated in the topdrawer plots is slightly different from the one you get from normal MG5_aMC running. Hence, there can be differences in the estimation of the uncertainties.

Best,
Rikkert

Revision history for this message
Andrej Saibel (asaibel) said :
#2

Dear Rikkert,

Thanks a lot for the answer and the explanations!

Could you please elaborate on how different the distributions are and where the difference comes from? I need to be certain that the distributions and uncertainties that I plot are indeed correct.

Kind regards,

Andrej

Revision history for this message
Rikkert Frederix (frederix) said :
#3

Dear Andrej,

The total cross section and distributions should be computed correctly. It's just that the statistical uncertainty that is computed in slightly different ways in the plots and in the main code. The uncertainty quoted by MG5_aMC on the total cross section is accumulated by considering the variance of the weights of all phase-space points considered in the phase-space integration. In the plots, the statistical uncertainty is by considering the variance of the results from consecutive iterations in the phase-space integration. In the limit of large number of events and iterations, the two should be identical. I guess that in your case there have not been that many iterations (it's actually not so easy to get this number out of the code and check how many iterations there have been) and therefore the statistical uncertainty as given in the topdrawer plots is sometimes a slight underestimate of the true value.

Note that when using the HwU plots instead, the error estimation follows more closely the one used to compute the central value as given by MG5_aMC should therefore not be affected by any of these issues.

Finally, you should keep in mind that these are just estimations of the true uncertainty and should therefore not taken too seriously.

Best,
Rikkert

Revision history for this message
Andrej Saibel (asaibel) said :
#4

Dear Rikkert,

thank you very much for the detailed answer!

Kind regards,
Andrej

Revision history for this message
Olivier Mattelaer (olivier-mattelaer) said :
#5

I will add a pure mathematical point.

The estimator of the variance is not ALWAYS converging trough the actual variance even in the infinite number of points/iterations. You can proof that such estimator correctly converge ONLY for convex functions.

This is not the case for the estimator of the integral which is ensure to converge towards the correct value.

Cheers,

Olivier
> On 26 Mar 2019, at 16:17, Rikkert Frederix <email address hidden> wrote:
>
> Question #679425 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/679425
>
> Status: Open => Answered
>
> Rikkert Frederix proposed the following answer:
> Dear Andrej,
>
> The total cross section and distributions should be computed correctly.
> It's just that the statistical uncertainty that is computed in slightly
> different ways in the plots and in the main code. The uncertainty quoted
> by MG5_aMC on the total cross section is accumulated by considering the
> variance of the weights of all phase-space points considered in the
> phase-space integration. In the plots, the statistical uncertainty is by
> considering the variance of the results from consecutive iterations in
> the phase-space integration. In the limit of large number of events and
> iterations, the two should be identical. I guess that in your case there
> have not been that many iterations (it's actually not so easy to get
> this number out of the code and check how many iterations there have
> been) and therefore the statistical uncertainty as given in the
> topdrawer plots is sometimes a slight underestimate of the true value.
>
> Note that when using the HwU plots instead, the error estimation follows
> more closely the one used to compute the central value as given by
> MG5_aMC should therefore not be affected by any of these issues.
>
> Finally, you should keep in mind that these are just estimations of the
> true uncertainty and should therefore not taken too seriously.
>
> Best,
> Rikkert
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.