Single- and di-jet measurements from ATLAS and CMS

[enocera]

Status

This is a summary of the current implementation of single- and di-jet data from ATLAS and CMS. The following data sets are being considered:

• CMS_1JET_8TEV (CMS, single-jet production at 8 TeV, 2D diff. distrs.);
• CMS_2JET_7TEV (CMS, di-jet production at 7 TeV, 2D diff. distr.);
• CMS_2JET_3D_8TEV (CMS, di-jet production at 8 TeV, 3D diff. distr.);
• ATLAS_1JET_8TEV_R04 (ATLAS, single-jet production at 8 TeV, 2D diff. distr., R=0.4);
• ATLAS_1JET_8TEV_R06 (ATLAS, single-jet production at 8 TeV, 2D diff. distr., R=0.6);
• ATLAS_2JET_7TEV_R04 (ATLAS, di-jet production at 7 TeV, 2D diff. distr., R=0.4);
• ATLAS_2JET_7TEV_R06 (ATLAS, di-jet production at 7 TeV, 2D diff. distr., R=0.6).

The implementation of all these data sets is now finalised, pending final reviews (see PRs NNPDF/buildmaster#113 and #564 ), in particular:

• all the available information provided by the experimentalists has been taken into account; standard checks on the correctness of its implementation have been performed;
• fastNLO tables (NLO) have been generated from scratch with NLOjet++ and have been benchmarked against the numbers provided by the Zurich group; the choice of central scale is Ĥ_T for single-jets and m₁₂ for di-jets;
• FK tables for theories 52 and 53 have been produced and uploaded to the server; their numerical accuracy has been benchmarked against the native fastNLO tables;
• NNLO QCD K-factors have been implemented as per the numbers provided by the Zurich group.

Results

A data/theory comparison is linked in the table below; numbers correspond to the χ² per data point using the corresponding (NLO or NNLO) default NNPDF3.1 fitted charm PDF sets.

Experiment	Ndat	NLO (EW)	NNLO (EW)	NLO (reg)	NNLO (reg)
CMS_1JET_8TEV	185	6.67 (6.33)	8.60 (8.37)	1.56	0.93
CMS_2JET_7TEV	54	2.16 (1.92)	2.26 (2.09)	2.14	2.26
CMS_2JET_3D_8TEV	122	8.27 (8.46)	3.96 (3.75)	5.84	3.10
ATLAS_1JET_8TEV_R04	171	4.37 (3.78)	8.34 (7.88)	1.74	1.96
ATLAS_1JET_8TEV_R06	171	3.12 (2.81)	4.25 (4.09)	1.51	1.67
ATLAS_2JET_7TEV_R04	90	1.77 (1.50)	2.94 (2.63)	0.74	0.80
ATLAS_2JET_7TEV_R06	90	1.60 (1.35)	2.31 (1.96)	0.80	0.90

For the ATLAS di-jet data, three independent sets of systematic uncertainties (corresponding to three correlation models) are provided (nominal, weak, strong). They are all implemented. The χ² per data point changes as follows upon variation of the correlation model.

Experiment	NLO			NNLO
	nominal	weak	strong	nominal	weak	strong
ATLAS_2JET_7TEV_R04	1.77	1.81	1.84	2.94	3.07	3.11
ATLAS_2JET_7TEV_R06	1.60	1.58	1.61	2.31	2.32	2.37

For the CMS data, nonperturbative (NP) corrections are provided on top of experimental systematics as rescaling factors to the central value of the data. These are implemented as (optional) additional systematic uncertainties, determined as the difference between the nominal data point and its value rescaled by the provided factor. The χ² per data point changes as follows upon inclusion of the NP uncertainties.

Experiment	NLO		NNLO
	w/o NP	w/ NP	w/o NP	w/ NP
CMS_1JET_8TEV	6.67	4.32	8.60	6.06
CMS_2JET_7TEV	2.16	2.15	2.26	2.24
CMS_2JET_3D_8TEV	8.27	6.22	3.96	3.12

Comments

1. Upon inclusion of NNLO QCD corrections, the theory prediction becomes generally closer to the central value of the data points (w.r.t. the NLO prediction). However, this improvement in the data/theory "visual agreement" is not statistically mirrored in the chi2, which consistently increases from NLO to NNLO.
2. Varying the correlation scenario for the ATLAS di-jet data appears to have a mild effect. The nominal scenario always leads to the best chi2.
3. The impact of NP corrections in the CMS data sets can be relevant. We might want to study whether there is any difference between deweighting the data (the approach implemented now) and shifting the data (the approach whereby data points are rescaled), in analogy to what was attempted in the nuclear uncertainty study.
4. No information about correlations between different data sets reconstructed from the same measurement (namely single-jet and di-jet data at the same c.m. energy for the same experiment) is included. In the past, @juanrojochacon claimed that such an information should be available (somewhere, in some format). Certainly it is not available from the papers, nor hepdata, nor the ATLAS/CMS public web pages.
5. Information on different correlation models is available from hepdata only for ATLAS di-jet data. Therefore they are implemented only for this experiment. It seems that information on different correlation models is not available for the other data sets. Further action in this respect might be taken depending on the outcome of ATLAS and CMS correlation models for inclusive jet and dijet data #507.
6. Pre-computed EW corrections are (in principle) available for the following experiments:

• CMS_1JET_8TEV*
• CMS_2JET_3D_8TEV
• ATLAS_2JET_7TEV_R04**
• ATLAS_2JET_7TEV_R06**
  [*] available from hepdata only to people with a CERN account.
  [**] Joao Pires claims that they are available from hepdata, I claim that they aren't.
  It might be interesting to include the provided EW corrections within the formalism of theory uncertainties. However, the agreement in Varenna seemed different, i.e. to fit PDFs with EW corrections (by generating the appropriate FK tables with QCDxEW corrections). We might want to do both: the former for the phenomenological paper on jets with the Zurich group; the latter for NNDPF4.0.

To-do

It seems that the onwards natural workflow can be as follows

• review and merge PRs NNPDF/buildmaster#113 and Dijets #564 ;
• finalise PR NNPDF/external#51 (applgrid interface to fastNLO tables);
• revisit the χ² values in light of the covariance matrix regularisation (see Add tools to both fit and analyse fits with regularised covmats #541);
• [modify the COMMONDATA format to allow for arbitrary N-dim distributions];
• document the NLOjet++-fnlo-applgrid-apflecomb toolchain used to produce FK tables;
• implement CMS di-jet data at 5 TeV, when/if NNLO QCD K-factors will be provided;
• revisit the implementation of ATLAS and CMS single-jet data at 7 TeV;
• do some interesting physics;
• report to the Zurich group.

[juanrojochacon]

Thanks @enocera for all the work and for the excellent summary!

At least for the ATLAS data, the correlations between inclusive jets, dijets, and trijets were made available in a PUB note on CDS rather than in a full fledged publication. I will try to fish it but I am not sure we want to fit at the same time incljets and dijets - in any case first we need to understand the two datasets separately.

Concerning the documentation of the NLOjet++-fnlo-applgrid-apflecomb toolchain used to produce FK tables, perhaps this is something where @rabah-khalek can help - at the end of the day we will also need to use this chain to include the CMS 5 TeV dijet data into our nuclear fits, so this would be a useful exercise for everyone.

It is interesting that the chi2 goes up at NNLO, though of course this is at the fixed PDF level, we need to see what happens when the dijet data is used in the fit instead of the inclusive jets. It might be also worth playing a bit the game of the regularised covariance matrix that @wilsonmr @Zaharid presented at PDF4LHC: perhaps it is the problem rather than a serious inconsistency between data and theory.

Concerning your roadmap, I think that it would be most useful to be able to run a NNPDF3.1 fit without incl jets and with dijets at NLO and NNLO, to have an idea of whether or not the chi2 improves in this case and what is the pull on the gluon. Then we can do the other physics studies, but this will also provide useful feedback for the ongoing validation.

I will also take a look at the reports showing the data versus theory comparison. It is an interesting remark that while visually NNLO helps the chi2 becomes worse. So this is why we need to make sure that it is as in the ATLAS 7 TeV 2011 jets a problem with the experimental correlation matrix.

Let's discuss at the phone conf!

[juanrojochacon]

Hi @enocera thanks again for all the work with this project. Let me share some reflections and then I propose some action points. cc @stefanoforte comments welcome!

• I will now write to our Zurich friends to request

• if they would provide the K-factors for the inclusive jet cross-sections at 7 TeV with their new central scale choice
• if they can provide us with electroweak corrections, ideally obtained from the same calculation for all the same datasets

• I will also update them on the situation (chi2 comparisons at the pre-fit level) and mention them that now we can start with the fits, to try to understand better the relative impact of the incl jets vs dijet data.

• It is quite interesting that the effect of the regularisation is rather less marked for the dijets than for the inclusive jet cross-section measurements. This is a further indication that the former is a more robust observable than the latter. In any case it would be important to produce fits with and without the regularisation and study whether or not the results at the PDF level change by much.

• I would always include the NP corrections. I don't really trust the models used to evaluate them since they are at best MC tunes, so I think that the deweighting option is much safer than the shifting one. So I would stick to the approach we have been adopting up to now.

• I think that at this point we are ready to produce some fits - I won't count on the Zurich people to send us any more results, at least in the near future. We need to produce fits that can eventually be part of this pheno paper (which is mostly a proof-of-concept for the implementation of dijet cross-sections in the context of the global PDF fit) and if needed we can refine our choice for NNPDF4.0.

• My proposal for the fits would be the following

1. Start with a NNPDF3.1 baseline without top and without jets (so basically an unconstrained gluon at large-x)
2. Add the ATLAS and CMS dijet measurements at 7 TeV and 8 TeV, perhaps first the 7 TeV data and then adding the 8 TeV data
3. Starting from the baseline, add the inclusive jets at 8 TeV from ATLAS and CMS

Then we can compare 2) and 3) in terms of PDF impact: this is not exactly an apples with apples comparisons, but for the time being we cannot do much more. In any case 3) is the most interesting comparison of all, since it demonstrates that we can fit dijet data which is the main goal of the exercise.

Concerning the fits I would propose the following settings:
*) Both NLO and NNLO, to study the perturbative convergence
*) Both with and without regularisation, to eliminate the issues due to ill-defined correlation matrices
*) Include whenever possible NP corrections via the deweighting method.

Once we have the results of this first bunch of fits we can discuss and decide what other checks we want to carry out eg adding back the top quark pair production measurements.

Perhaps you can take care of coordinating the running of these fits and assigning them to people that can have access to enough computing resources? For example @scarrazza can we run fits in Milan now or we don't have enough resources there?

[enocera]

@juanrojochacon I agree on most counts - we can have a more extensive chat at the PC on Friday. As far as fits are concerned, you're proposing quite some work - for the time being I'd look only at the unregularised fits. Again, I would discuss a plan (and allocation of computational resources) at the PC next Friday. Thanks for liaising with the Zurich people.

[juanrojochacon]

sure, let's discuss on Friday. Perhaps at the code meeting @scarrazza and @Zaharid can summarise where do we stand in terms of computing resources, since in any case I expect that we will need to run a largish number of fits for this project. In the meantime I will check with our Zurich friends if we can expect anything from them - I don't think so, recall that several months ago they promised to compute the NLO fast grids as well and nothing happened!

[juanrojochacon]

ok email sent to our Zurich friends. I will also check separately with Joao via Skype, since he is typically reactive this way. Let's see.

In case they ask, is it possible to compare the covariance matrices that we use without and with the regularisation? For example can we plot the difference between rho(original) and rho(reg) for each of these datasets, and compare it with the original one? This should be possible using some of the vp2 tools that @wilsonmr used for the plots in his pdf4lhc talk

[stefanoforte]

s -> jr, ern Dear Joan (+ Emanuele) • My proposal for the fits would be the following 1. Start with a NNPDF3.1 baseline without top and without jets (so basically an unconstrained gluon at large-x) 2. Add the ATLAS and CMS dijet measurements at 7 TeV and 8 TeV, perhaps first the 7 TeV data and then adding the 8 TeV data 3. Starting from the baseline, add the inclusive jets at 8 TeV from ATLAS and CMS as mentioned in a previous email message of mine, it seems to me that we need four PDF sets: a) NNPDF3.1 dataset but with no jets b) NNPDF3.1 dataset but with dijets c) NNPDF3.1 dataset with single-jet inclusive, old scale choice d) NNPDF3.1 dataset with single-jet inclusive, new scale scale choice Comparing to your proposal there are two issues: a) when you say "add the inclusive jets" which scale do you mean? I understand that doing both might be ideal, but a bit too heavy. If we can do only one I'd go for the old scale choice, for various obvious reasons. b) should be remove top or not? I think that we should not, for two reasons: first, because if we really want unconstrained gluon at large x we'd have to remove both top and Z pt; second, because part of the purpose of the exercise is test the consistency of jets with the rest of the fit. At a later stage we could also check the effect of adding and removing top and Zpt with either jet observable, but this seems more like a second order question. Look at it this way: if the gluon at large x is u constrained, it is quite likely that we can get a good fit even with single jet despite it being perturbatively unstable, because th PDF will make up for missing pert. corections. We would then end up with two fits, with dijets and single jet inclusive, with similar quality but different gluons. It would then be difficult to argue which of the two is the good one. If instead the gluon is already constrained by something else, we then see rightaway which is the good jet observable: the one which leads to a good fit with no tensions. Cheers Stefano

[juanrojochacon]

Hi @stefanoforte (cc @enocera ) , concerning your suggestions

• I am happy to keep the top data, perhaps we can remove it at a later stage. For large-x gluons Z pt is not very relevant but it is fine by me to keep both in the baseline for the time being.

• Note that we cannot fit the inclusive jet data with the old scale choice: K-factors at NNLO are available only with the new choice. Also, I am not sure why is interesting to use a suboptimal choice if we now have evidence that the new one is better?

We should try to figure an action plan with the calculations we have received so far, we cannot really wait much more. This project has been around for a long time so it would be nice to be able to wrap it up relatively soon with a pheno paper

[stefanoforte]

s-> j Dear Joan, • Note that we cannot fit the inclusive jet data with the old scale choice: K-factors at NNLO are available only with the new choice. Also, I am not sure why is interesting to use a suboptimal choice if we now have evidence that the new one is better? well, first, we do not have any solid evidence that the new one is better in fact: the evidence that the new one is better is one of the goals of this study! The evidence that has been presented so far is very circumstantial, basically a table of desirable properties for which it seems that the nwe choice hs more desirable properties! Second, all PDF fits performed so far have used single-inclusive jets with the old scale choice. If we only do fits with the new scale choice we have no clue of the way results will change between new and old, and in fact even if we find that dijects are better (or no better) than single-inclusive jets we would not know whether this does or does not depend on the scale choice. So it is absolutely crucial to have a comparison with the old scale choice. Note that we could do this for just the old ATLAS and CMS 7 TEV: i.e. if we get the K-factors with new scale choice for those experiments we can first run a fit with just this jet data, old scale, or new scale. Then move to the wider set of measurements for which we have both single-inclusive and dijet observables. But I'd say that this part of the study is crucial! Cheers Stefano We should try to figure an action plan with the calculations we have received so far, we cannot really wait much more. This project has been around for a long time so it would be nice to be able to wrap it up relatively soon with a pheno paper — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.*

…

--