SciPost Submission Page
Two Invertible Networks for the Matrix Element Method
by Anja Butter, Theo Heimel, Till Martini, Sascha Peitzsch, Tilman Plehn
This is not the latest submitted version.
This Submission thread is now published as
Submission summary
| Authors (as registered SciPost users): | Theo Heimel · Tilman Plehn |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2210.00019v3 (pdf) |
| Date submitted: | 2023-02-20 17:32 |
| Submitted by: | Heimel, Theo |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
Abstract
The matrix element method is widely considered the ultimate LHC inference tool for small event numbers, but computationally expensive. We show how a combination of two conditional generative neural networks encodes the QCD radiation and detector effects without any simplifying assumptions, and allows us to efficiently compute likelihoods for individual events. We illustrate our approach for the CP-violating phase of the top Yukawa coupling in associated Higgs and single-top production. Currently, the limiting factor for the precision of our approach is jet combinatorics.
List of changes
We would like to thank the referees for their comments and provide an updated version addressing their requests. A detailed list of changes can be found in our replies to the referees.
Current status:
Reports on this Submission
Report #1 by Sebastien Wertz (Referee 1) on 2023-3-17 (Invited Report)
- Cite as: Sebastien Wertz, Report on arXiv:2210.00019v3, delivered 2023-03-17, doi: 10.21468/SciPost.Report.6919
Report
Note: this review is based on the "v4" posted on Arxiv.
I thank the authors for answering my questions and implementing my suggestions (and providing a handy "diff"). I find that the new version improves in clarity. I have a few small remaining comments. Once they are answered, I will be happy to recommend this paper to be accepted for publication.
Requested changes
1- Now that the introduction puts the focus on the modelling of the transfer function, which the method is indeed mainly about, I think the abstract could be rephrased to reflect that: E.g. remove the "but computationally expensive", and explain that in practice we have had to make many assumptions about the form of the transfer function for QCD and detector effects, a limitation which your method lifts without increasing significantly the computational complexity.
2- Typo on p.3 ("to not (yet)")
3- About the statement added regarding the use of the missing transverse momentum: in that particular case (no ISR) you are right, but I just want to point out that in general the MET can provide additional information, e.g. to limit the jet combinatorics if you had included ISR in the leptonic channel.
4- Regarding the clarification added about Eq. (14), I would suggest to rephrase as "our xhard-integration would become trivial if we could sample from..." to make clear that the integration only becomes trivial if that sampling is exact, which won't be the case in practice.
5- Shouldn't the schematic in Figure 4 also be updated with the extra 1/p(r) factor added in the v4?
Anonymous on 2023-03-30 [id 3525]
I am satisfied with the authors responses to my questions and recommend publication after a minor update described below is implemented.
I also note that the authors found a mistake in Eq 16 in the paper which was also included in the code upon which the paper results were based. This correction is fully implemented in v4 of the manuscript and the authors inform that the changes were small so there was no need to make any changes to the discussion of the results.
I thank the authors for significant updates to the introduction section after the recommendations. It reads much better. However, there one small update i would like to see implemented before the article is published. The use of deep learning to address the computational challenges of the MEM calculation was first described (see page 20 in the section on "Sustainable Matrix Element Method" at https://inspirehep.net/files/ca2037001073f35f718b719fa3bddf8c) in "A Roadmap for HEP Software and Computing R&D for the 2020s" which was published in Comput.Softw.Big Sci. 3 (2019) 1, 7. This should be referenced in your introduction in an update before publication.
I would suggest (you should decide on the exact wording for resubmission)
"The connection between the MEM and modern ML-methods was pioneered in Ref. [67]..."
be changed to something like:
"The connection between the MEM and modern ML-methods was first described in Ref.[67] which proposed using deep neural networks to overcome the computation challenges inherent to the MEM. In Ref. [68], the use of a deep neural network built by regression of the MEM integral was first demonstrated."
where
[67] HEP Software Foundation Collaboration; Johannes Albrecht(Dortmund U.) et al., "A Roadmap for HEP Software and Computing R&D for the 2020s; ", Comput.Softw.Big Sci. 3 (2019) 1, 7 , arXiv:1712.06982 [hep-ex]
[68] F. Bury and C. Delaere, Matrix element regression with deep neural networks — Breaking the CPU barrier, JHEP 04 (2021) 020, arXiv:2008.10949 [hep-ex].
It is the opinion of this referee that with this change that your paper would accurately describe the when and where the key ideas of deep learning for MEM were first published. Your novel ideas and approaches described in your manuscript then build (significantly, in my opinion) upon these publications.