MadGraph reweighting only works in some specific cases

Asked by Yue Xu

Dear MadGraph Group,
I generated the events with an external model which contains a heavy higgs ( hh ), and the process is:
  import model SMwithHeavyScalarDim4Dim6_NoDecay_UFO
 generate p p > z hh, z > l- l+, hh > j j j j
 add process p p > z hh, z > j j, hh > e- e+ j j
 add process p p > z hh, z > j j, hh > mu- mu+ j j
 add process p p > w- hh, w- > j j, hh > e- e+ j j
 add process p p > w- hh, w- > j j, hh > mu- mu+ j j
 add process p p > w+ hh, w+ > j j, hh > e- e+ j j
 add process p p > w+ hh, w+ > j j, hh > mu- mu+ j j

There are two parameters we can change. They are fw and fww that influence the coupling between heavy higgs and vector boson and the decay width of heavy higgs.
I would like to use MadGraph to reweight the generated events with a different setting of fw and fww from parameter point with fw=1200 and fww=2900. I always reweight from broad width point to narrow width point.
The reweight_card is:

change keep_ordering True
launch --rwgt_name=fw1000fww2412
 set DIM6COEFF 3 1000
 set DIM6COEFF 4 2412
 set DECAY 254 auto

launch --rwgt_name=fw800fww1930
 set DIM6COEFF 3 800
 set DIM6COEFF 4 1930
 set DECAY 254 auto

launch --rwgt_name=fwM800fwwM1930
 set DIM6COEFF 3 -800
 set DIM6COEFF 4 -1930
 set DECAY 254 auto

launch --rwgt_name=fw800fww0
 set DIM6COEFF 3 800
 set DIM6COEFF 4 0
 set DECAY 254 auto

I will use (fw,fww) to repreaent parameter points for convinent. The reweighted cross-section and cross-section calculated by MG5 directly for each points is:
     point, reweighted cs[pb] , calculated by MG5 directly[pb]
A: (1000, 2412) 0.000109022232265 +- 7.95662029691e-07, 0.0001079
B: (800,1930) 7.19540848184e-05 +- 9.65657147264e-07 , 7.143e-05
C: (800, 0) 5.54344901268e-05 +- 7.29610350547e-06 , 5.804e-05
D: (-800, -1930) 8.27684121487e-05 +- 9.03141792537e-06 , 8.0754e-05

They are obtained from E:(1200,2900).
I found when use reweighting between parameter points in the same radial direction with respect to (0,0) original point, the reweighting works good. For example, from E to A, B
But when reweighting between parameter points in the different radial direction, there is a big difference between the reweighted cs and and the MG5 cs. And the fluctuation of reweighted cs is very large. For example, from E to C, D.

Could you help me to figure out what cases this? A bug?

The model used is here:
https://github.com/xuyue1231/MG5-heavy-Higgs/tree/master/SMwithHeavyScalarDim4Dim6_NoDecay_UFO

The process is here:
 https://github.com/xuyue1231/MG5-heavy-Higgs/tree/master/LHC_HH_2lep_rw

Thanks,
Yue

Question information

Language:
English Edit question
Status:
Solved
For:
MadGraph5_aMC@NLO Edit question
Assignee:
No assignee Edit question
Solved by:
Yue Xu
Solved:
Last query:
Last reply:
Revision history for this message
Olivier Mattelaer (olivier-mattelaer) said :
#1

Hi,

Re-weighting is not a method which is expected to work in all scenario.
In some benchmark change, you would need very high statistic in the original sample to reach a small precision in the re-weighted sample. That loss of precision is linked to the variance of the re-weight factor which can be quite large if you have variation of the width of the particle.
(Now I do not know by how much your width is modified from one benchmark to the next).

For the radial part, I do not know what you are talking about. The fact that you have large estimated error in that case proofs that you have large variation in the weights and therefore that you are two benchmark are too far away from each other to offer a decent method of generation.

Cheers,

Olivier

PS: You can try to do "change helicity False" which changes the definition of the weight (using the average over helicity). If you have issue due to helicity this might fix the issue.

Revision history for this message
Yue Xu (yuexu) said :
#2

Hi Olivier,

I talked about the radial direction because of the lagrangian. The pt distributions of parameter points in the same radial direction are similar.
In the cases I described above, the heavy higgs width of A is 1.9GeV, width of B is 1.5GeV, width of D is 2.0GeV. If considering the width difference, the reweighting from E with width 2.4GeV to D should work well, but it's not.

And if I reweight from (0,9000) with width 32GeV to (0,5400) with width 12GeV, the reweighting works well. The reweighted cs is 0.000107 +- 1.793e-06 and the cs calculated by MG5 directly is 0.000107.
In same case, the variation of decay width is very large, but the reweighting is good. In other radial directions, I also found the same phenomenon.

I will try to do "change helicity False".
Thanks again,

Cheers,
Yue

Revision history for this message
Olivier Mattelaer (olivier-mattelaer) said :
#3

OK then the main culprint might not be the breitwigner but another kinematic.
In general, I prefer to consider the re-weighting method as NOT working but if prooves to work.
It is very difficult to know in advance what would be the precision of such method and we know that this method is not very robust by construction. So in itself having examples where it does not work is expected and not surprising.

Cheers,

Olivier

Revision history for this message
Yue Xu (yuexu) said :
#4

Hi Olivier,

We want to use reweighting technique to reduce the requested signal samples. But it seems reweighting doesn't work well in this case. It's quite hard to reduce hundreds signal samples to several samples by reweighting.

Thanks a lot for the help and answers!

Cheers,
Yue

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

Hi,

Indeed this method is not expected to work in all cases.
It is a method that can save you a lot of cpu time when working but it is quite difficult to know in advance when such method will be efficient enough or not.

Cheers,

Olivier

> On 22 Nov 2019, at 14:08, Yue Xu <email address hidden> wrote:
>
> Question #685986 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/685986
>
> Status: Answered => Solved
>
> Yue Xu confirmed that the question is solved:
> Hi Olivier,
>
> We want to use reweighting technique to reduce the requested signal
> samples. But it seems reweighting doesn't work well in this case. It's
> quite hard to reduce hundreds signal samples to several samples by
> reweighting.
>
> Thanks a lot for the help and answers!
>
> Cheers,
> Yue
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.