Reweighting with large scaling of the cross section

Asked by Andrew Wightman on 2018-01-19

Hello,

I want to use reweighting in regions of parameter space that produce ~10x scaling of the cross section. Is this feasible, as long as I run with enough stats?

Cheers,
Andrew

Question information

Language:
English Edit question
Status:
Solved
For:
MadGraph5_aMC@NLO Edit question
Assignee:
No assignee Edit question
Solved by:
Andrew Wightman
Solved:
2018-02-01
Last query:
2018-02-01
Last reply:
2018-01-30

Hi Andrew,

Actually if all your event are scaled by a factor of 10, you do not need that much stats.
What require a lot of stats is when you have a huge variance in the weights.

But this being said, if you run with enough stats everything should be feasible (as long as you have the same phase-space and other equivalent restriction --see my paper on that topic--)

Cheers,

Olivier

Andrew Wightman (awightma) said : #2

Hi Olivier,

I guess my question falls in the scope of your 2nd point. Namely, as I scan over values in my parameter space, the cross section can vary between 1x and 10x the SM.

For example in: https://cp3.irmp.ucl.ac.be/projects/madgraph/wiki/Reweight

The approach fails for CWWW when the coupling value is set to 100, is it possible to get this to not fail when using reweighting?

Also, I notice that I get very different results depending on where I initially generate my events in parameter space. For example, if I start at SM coupling values and calculate cross sections at different values via reweighting, compared to starting at a non-SM value and calculating the same points (still via reweighting).

Coeff. : Cross Section
0.000 : 0.5636 +- 0.00052 *start*
0.025 : 0.6636 +- 0.09952
0.050 : 0.9595 +- 1.55127
0.075 : 1.4513 +- 7.8012

Coeff. : Cross Section
0.000 : 0.5811 +- 0.00213
0.025 : 0.5878 +- 0.00078
0.050 : 0.6482 +- 0.00062 *start*
0.075 : 0.7624 +- 0.00113

Finally, could you please point me to the paper you mentioned?

Cheers,
Andrew

Hi,

The approach fails for CWWW when the coupling value is set to 100, is it
possible to get this to not fail when using reweighting?

For this case, I would say yes but you will need very large stats.

The problem here is that you have the high energy behaviour which is (heavily) suppressed in the SM
while it is basically flat with cwww=100 (I would bet that you break unitarity for such value actually).
So you do not have a factor of 10 in this part of the phase-space but a factor of ~10^4 (maybe even more) while you have a factor of order 1 close to the threshold.

Also, I notice that I get very different results depending on where I
initially generate my events in parameter space. For example, if I start
at SM coupling values and calculate cross sections at different values
via reweighting, compared to starting at a non-SM value and calculating
the same points (still via reweighting).

This is a bad sign obviously and indicate that at least one re-weighting has large error associated to it. (i.e. that the variance of the re-weighting is very large)
Now it is normal that the re-weighting accuracy will depend of your starting point.
A (probably bit naive) idea is that it is better to start from a theory over-shouthing the tail than the opposite (i.e. having many weight lower than one hurts less that few event with very large weight).

Cheers,

Olivier

Finally, could you please point me to the paper you mentioned?

arXiv:1607.00763<http://arXiv.org/abs/arXiv:1607.00763>
I have added it in the wiki actually.

On 19 Jan 2018, at 22:07, Andrew Wightman <<email address hidden><mailto:<email address hidden>>> wrote:

Question #663279 on MadGraph5_aMC@NLO changed:
https://answers.launchpad.net/mg5amcnlo/+question/663279

   Status: Answered => Open

Andrew Wightman is still having a problem:
Hi Olivier,

I guess my question falls in the scope of your 2nd point. Namely, as I
scan over values in my parameter space, the cross section can vary
between 1x and 10x the SM.

For example in:
https://cp3.irmp.ucl.ac.be/projects/madgraph/wiki/Reweight

The approach fails for CWWW when the coupling value is set to 100, is it
possible to get this to not fail when using reweighting?

Also, I notice that I get very different results depending on where I
initially generate my events in parameter space. For example, if I start
at SM coupling values and calculate cross sections at different values
via reweighting, compared to starting at a non-SM value and calculating
the same points (still via reweighting).

Coeff. : Cross Section
0.000 : 0.5636 +- 0.00052 *start*
0.025 : 0.6636 +- 0.09952
0.050 : 0.9595 +- 1.55127
0.075 : 1.4513 +- 7.8012

Coeff. : Cross Section
0.000 : 0.5811 +- 0.00213
0.025 : 0.5878 +- 0.00078
0.050 : 0.6482 +- 0.00062 *start*
0.075 : 0.7624 +- 0.00113

Finally, could you please point me to the paper you mentioned?

Cheers,
Andrew

--
You received this question notification because you are an answer
contact for MadGraph5_aMC@NLO.

Andrew Wightman (awightma) said : #4

Hi Olivier,

Thanks for the response! I had another follow up question, is it possible to generate the re-weighted points in the weighted (instead of unweighted) events file?

Cheers,
Andrew

Hi,

Note that the latest version of the code does not create such weighted file anymore.
But they are no problem to start from a weighted sample which happens inside the workflow
when you generated weighted sample via the bias module.

You can also run the reweighting manually and there specify what is your inputfile.
(./bin/madevent reweight PATH_TO_FILE)

Cheers,

Olivier

> On 30 Jan 2018, at 12:42, Andrew Wightman <email address hidden> wrote:
>
> Question #663279 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/663279
>
> Status: Answered => Open
>
> Andrew Wightman is still having a problem:
> Hi Olivier,
>
> Thanks for the response! I had another follow up question, is it
> possible to generate the re-weighted points in the weighted (instead of
> unweighted) events file?
>
> Cheers,
> Andrew
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

Andrew Wightman (awightma) said : #6

Thanks Olivier!