Inconsistent HepMC files in some runs

[johannbrehmer]

Running the particle physics tutorial in the Delphes version with a clean install of MadGraph 2.6.5 and Pythia 8 installed from within MadGraph leads to a strange issue:

The HepMC file for run_01 (the one sampled from the SM) has an inconsistent HepMC file with events beginning with something like this:

E 0 -1 -1.0000000000000000e+00 -1.0000000000000000e+00 -1.0000000000000000e+00 0 0 142 1 2 0 6 9.8323591999999982e-07 5.4039015000000000e-07 1.2517713000000000e-07 1.0361000000000004e-07
N 6 "-247.8_#_ORIG:_MERGING=0.000" "229.0487_#_ORIG:_0.0
SET_PARAM_CARD_DIM6_5_272.0368_#_ORIG:_MERGING=0.000" "247.8_#_ORIG:_MERGING=0.000" "Weight_MERGING=0.000" 
U GEV MM

There should be six weights, as correctly given by N 6. But only four are given. This leads to Delphes not processing the HepMC file (ERROR: invalid event format).

Everything's fine in run_02:

E 0 -1 -1.0000000000000000e+00 -1.0000000000000000e+00 -1.0000000000000000e+00 0 0 561 1 2 0 6 4.3427632999999996e-06 2.2917435000000001e-05 2.3613610000000001e-05 1.3052637000000000e-09 1.2243924000000000e-05 4.0563000000000004e-06
N 6 "-247.8_#_ORIG:_MERGING=0.000" "0.0_#_ORIG:_247.8
SET_PARAM_CARD_DIM6_5_-247.8_#_ORIG:_MERGING=0.000" "0.0_#_ORIG:_247.8
SET_PARAM_CARD_DIM6_5_247.8_#_ORIG:_MERGING=0.000" "0.0_#_ORIG:_MERGING=0.000" "229.0487_#_ORIG:_247.8
SET_PARAM_CARD_DIM6_5_272.0368_#_ORIG:_MERGING=0.000" "Weight_MERGING=0.000" 
U GEV MM

The number of weights is consistent, and Delphes can process this file.

?!

[johannbrehmer]

Some more details:

• The LHE files seem fine in both cases.
• Run cards, param cards, and Pythia cards were identical between the two runs.
• The only difference between the two runs is which benchmark was used for sampling (see the tutorial notebook).

[johannbrehmer]

This was originally reported by Kevin Kwok with the following setup:

My MadGraph is on 2.6.5
My Pythia is just pythia8?
Delphes 3.4.1
I cloned the github repo and also pip install madminer
madminer 0.4.6
I am not using Docker
No custom MG-pythia patch

[johannbrehmer]

The only obvious difference between the LHE files I can find is a consequence of the different sampling benchmarks.

run_01 (sampled from benchmark "sm") has

<initrwgt>
<weightgroup name='mg_reweighting' weight_name_strategy='includeIdInWeightName'>
<weight id='w'>set param_card dim6 2 247.8 # orig: 0.0
</weight>
<weight id='neg_w'>set param_card dim6 2 -247.8 # orig: 0.0
</weight>
<weight id='ww'>set param_card dim6 5 247.8 # orig: 0.0
</weight>
<weight id='neg_ww'>set param_card dim6 5 -247.8 # orig: 0.0
</weight>
<weight id='morphing_basis_vector_5'>set param_card dim6 2 229.0487 # orig: 0.0
set param_card dim6 5 272.0368 # orig: 0.0
</weight>
</weightgroup>
</initrwgt>

On the other hand, run_02 (sampled from "w") has

<initrwgt>
<weightgroup name='mg_reweighting' weight_name_strategy='includeIdInWeightName'>
<weight id='sm'>set param_card dim6 2 0.0 # orig: 247.8
</weight>
<weight id='neg_w'>set param_card dim6 2 -247.8 # orig: 247.8
</weight>
<weight id='ww'>set param_card dim6 2 0.0 # orig: 247.8
set param_card dim6 5 247.8 # orig: 0.0
</weight>
<weight id='neg_ww'>set param_card dim6 2 0.0 # orig: 247.8
set param_card dim6 5 -247.8 # orig: 0.0
</weight>
<weight id='morphing_basis_vector_5'>set param_card dim6 2 229.0487 # orig: 247.8
set param_card dim6 5 272.0368 # orig: 0.0
</weight>
</weightgroup>
</initrwgt>

[johannbrehmer]

Same issue with MG v2.6.6, and also if we just do one run (sampling at the SM) and disable the second one.

[johannbrehmer]

Surprisingly, the old Delphes tutorial in MadMiner v0.3.0 runs fine., using the same MG + Pythia installation that failed for v0.4.6...

I went through a git diff between the two versions. I could not find any changes in madminer.core that seemed related to this. The run and Pythia cardsare also identical. The param and reweight card are structurally identical, but a different morphing basis means that the basis points are different.

I wonder if there's anything particular about the morphing basis in the 0.4.6 tutorial and sampling from the SM that leads to Pythia having issues...

[cranmer]

@alexander-held any chance you might have some time to help here?

[johannbrehmer]

Okay, I fixed the issue in the strangest way: by changing the morphing basis.

Here's a hypothetical explanation for what might be happening:

• The reweight card that MadMiner creates sets numerical values for the parameters depending on the morphing basis (benchmark points).
• MadGraph then writes a header into the LHE file that summarizes the effects of these weights, for instance as <weight id='ww'>set param_card dim6 5 247.8 # orig: 0.0
• Pythia (or the MG-Pythia interface) uses names for the weights internally that seem like some obfuscated version of what was in the LHE header: "247.8#_ORIG:_0.0 SET_PARAM_CARD_DIM6_5_272.0368_#_ORIG:_MERGING=0.000"
• When multiple benchmark points have identical values for the parameter points, it seems not too far fetched that these weight names in Pythia become identical, causing some confusion in Pythia
• And then the HepMC output is off

Sounds ridiculous? Yep. But when I changed the morphing basis from (0,0), (15, 0), (0,15), (-15, 0) etc. to something where no values appeared multiple times, the issue went away...