-- SamHigginbotham - 2022-05-02

SUS-TBT (Previously Higgs Exo) High Mass 2mu 2tau: AN-20-208

This twiki documents the analysis progress and documentation for the RunII High Mass Higgs Exotic search of pseudoscalar Higgs-like particles decaying to a pair of muons and tau leptons in the final state.

Useful Links

Related to the main analysis, note the branch change to "susyPAG" !

https://gitlab.cern.ch/tdr/notes/AN-20-208

https://gitlab.cern.ch/tdr/notes/AN-20-208/-/blob/susyPAG/AN-20-208_temp.pdf

Local paper documentation

https://github.com/samhiggie/SUSHAAPaper

datacards on lxplus:

/afs/cern.ch/user/s/shigginb/public/HAA_datacards/

discussion on statistical fluctuations and support on using Asymptotic Limit methods

unbinned fit discussion

discussion on shape based systematic uncertainties for unbinned parametric fit

shape systematic uncertainties in unbinned parametric fit

Past work and supporting documentation Previous 2016 result

https://cms.cern.ch/iCMS/jsp/db_notes/noteInfo.jsp?cmsnoteid=CMS%20AN-2017/173

Run II SM HTT Measurement

https://cms.cern.ch/iCMS/jsp/db_notes/noteInfo.jsp?cmsnoteid=CMS%20AN-2019/109

Background estimation methods

https://cms.cern.ch/iCMS/jsp/db_notes/noteInfo.jsp?cmsnoteid=CMS%20AN-2019/170

Preapproval Checklist

Currently updating plots and documentation with these fixes below
Task Description and Status
Fake Factors Updated Documentation to reflect using the SMHTT VH effort's fake rate measurements
Muon Energy Scale Uncertainty (combine) All systematic effects are now propagated to the shape directly
Hybrid New Limits Figuring out how to do this correctly in combine, 2016 analysis datacards aren't working in current implementation
Freeze ZZ bkg Norm Done
Freeze Fake Bkg Norm Done
Fake Background Norm Unc Taking from closure measurements for each channel and year
Fake background Shape Unc Now decorrelating the uncertainties taking VH SS relaxed region and analysis as justification of methodology
Other backgrounds that ZZ and FF Not considered and justified
Blinded Signal Region Distribution shown in note, documentation and below (note normalization is included in pdf shape and events are binned with final weight)
Comparison to 2016 DNN improvement and anti-discriminators cut down background events
Fix m4L cut not year depended Done! this was a misunderstanding initially consistent across years
ZZ fit quality depends on channel and year but chi square is computed and is acceptable
emu 2018 using fake rate and with updated cuts and criteria this is much better
Tau Pt using only taus above 20 GeV now, POG supported
Mu Pt using scale factors for low pt taus from Muon POG https://gitlab.cern.ch/cms-muonPOG/muonefficiencies/-/tree/master/Run2/preUL

October 20th and 25th 2022

General comments:

1) All the fake rates are those I (Cecile) have measured for the VH H->tautau analysis and all the FR-related text is a copy-paste of what I have written in the VH H->tautau AN. This needs to be referenced and explained clearly in your AN. Please make clear what is an exact copy-paste and add the reference to the VH AN as often as necessary.

2) Similarly you use a lot of material from the 2016-only analysis without referencing it correctly each time.

3) For HybridNew limits, having limits around 1 helps the numerical stability, so you could multiply the signal by an arbitrary factor to reach that level of limits and rescale after.

4) You need the blessing of POGs for non-standard object use. In particular you use taus with pT greater than 18.5 GeV instead of 20 GeV. Low pt muon scale factors need to be derived and blessed for 2017 and 2018. Not clear what trigger SFs you are using. The electron dxy and dz cuts seem non standard.

Samples:

- Are you using private signal samples? If so why and has this been blessed? The official samples must be available now

- Why do you use MuonEG datasets?

Corrections:

- Table 10 shows uncertainties and not corrections. Do you apply these numbers as corrections?

- You are reusing the 2016 low pt muon corrections. You should reference the source clearly and you should make the measurements for 2017 and 2018

Event selection:

- Can you write the explicit trigger requirements per year (HLT path names)?

- Not clear which working points are used for the tau discriminators (vs mu, e, jets)

- Do you apply different scale factors for the muons depending on the muon isolation threshold?

- There is no reason for the m4l cut to be year-dependent, please make it uniform. Your answer: "Was done in 2016 and I disagree. Some channels had terrible statistics otherwise and wouldn't converge in unbinned parametric fit". How can there be a year-dependent cut in the 2016-only analysis? In any case you cannot choose cuts based on statistical fluctuations of your samples. Loosen the cuts everywhere if you cannot deal with the stat of your samples.

Background estimation:

- Outline is wrong, no SS events are used

- the electron and muon fake rates are measured for pt> 10 GeV. How do you do since you have lower thresholds in your selection? The isolation WPs are not the same either as those you use

Systematic:

- This needs a lot of work! From the impacts, it looks like the normalisation of the backgrounds is still floating (and not correctly floating since it cannot go down), the scale uncertainties are all identical (impacts identical), and so many uncertainties are missing (eg muon ID is not negligible, trigger, tau ID, etc).

- As Adinda suggests you should parameterise the muon energy scale as uncertainties on the parameters of the functional form.

- Fake background normalisation uncertainties: evaluate the size of normalisation uncertainty (taking them directly from previous publication does not work because of different statistics and tau ID discriminators) Your answer: "Based on the closure plot we can assign a 20% uncertainty moving forward". This closure test is done only for one year and one final state. It needs to be expanded. In addition you should consider statistical uncertainties from the fake rate method because of the limited number of reweighted events.

- It is not clear what you are actually doing with the background shape uncertainties and how it is implemented. If you have uncertainties in the Bernsteing coefficients separately for eg c0 and c1, then the correlation is not taken into account.

Results:

- Limits with HybridNew: It seems like there is a 30% difference, which is not negligible. If you followup with Andrea and ask if you can use Asymptotic when you expect less than 0.1 event, I doubt he will agree. If you want to convince up the difference is negligible you should at least run a sufficient number of toys for a few mass points.

- For HybridNew limits, having limits around 1 help the stability, so you could multiply the signal by an arbitrary factor to reach that level of limits and rescale after.

- there is no plot showing the stacked normalised predicted backgrounds with the signal prediction scaled to some arbitrary BR. In the twiki you just show plots for the different processes separately and with a random normalisation (thousands of ZZ events).

- Can you add yield tables with uncertainties?

- Add more bins to the ZZ fits

September 27th 2022

email chain: comments in line (*)

To do before preapproval:

- Fake factors: verify which fake factors are used and correct the documentation, plot all the functions that are used (starting from the Root file you actually use in the analysis) and check they match those from the original documentation, check that your tau ID working points match those used to derive the original fake factors

updated the documentation using VH SM HTT

- Muon energy scale uncertainty: treat it as a shape systematic in Combine

Currently we do consider shape systematic uncertainties in combine. However, propagating the energy scale variations into combine as separate shapes is not possible please see: shape systematic uncertainties in unbinned parametric fit An alternative approach would be to widen the uncertainty in the spline or add additional errors on the shape parameters - an approach that shouldn't effect the expected values of the limit because the analysis is statically limited.

- Compute the expected limit with HybridNew for a few mass points to evaluate the difference wrt AsymptoticLimits (if there is a significant difference, HybridNew should be used for all mass points in the publication)

There is not a significant difference: please consider looking at the discussion here: unbinned fit discussion

- Fix the ZZ background normalisation to its predicted value, and use the corresponding uncertainty (no floating normalisation)

ok

- Fake background normalisation: if you want to use the predictions of the FF method, don’t float its normalisation and assign yield uncertainty

ok

- Fake background normalisation uncertainties: evaluate the size of normalisation uncertainty (taking them directly from previous publication does not work because of different statistics and tau ID discriminators)

Based on the closure plot we can assign a 20% uncertainty moving forward.

- Fake background shape uncertainties: Can you consider alternative shapes to fit the background? Do you decorrelate the coefficients of the functions before varying them by +/- 1 sigma? Can you show the background uncertainty band to demonstrate they cover the data uncertainties?

We chose to use Bernstein polynomials based on the 2016 result and their performance in the fit. In the end it shouldn't matter greatly what shape is chosen for the shallow sloping background that is practically constant in the fit range of the voigtian. Doesn't RooFit decorrelate the shape coefficients in the fit already by the way it changes the primary RooRealVars? The names are different and the distributions are different in the datacards and in the RooWorkspace so by default these should be uncorrelated in the construction of the likelihood - at least at that level. The background uncertainty is captured in the error on the fit parameters.

- Are the irreducible backgrounds other than ZZ neglected in the final fit? (It seems you only fit ZZ and fakes)

Yes, the rest of the standard model distributions are negligible.

- Can you show the final (blinded) signal region distributions?

Please take a look at the Fit Model and signal extraction section of the AN! These are the final signal region distributions. I've attached 2016 mmmt as an example. The unbinned PDF fit is overlayed with the events. To make it easier to read the signal (a40) and the two background categories (ZZ-irBkg and Datadriven-Bkg) are binned (signal has 25 bins, irBkg and Bkg 4 bins) the the PDF is overlayed on these bands.

40 Nominal a40 2016 mmmt Nominal.png
signal distribution a 40 in signal region
2 Nominal irBkg 2016 mmmt Nominal.png
ZZ irreducible background in signal region
2 Nominal Bkg 2016 mmmt Nominal.png
FF Datadriven reducible background in the signal region

- Can you compare your 2016 expected results to the 2016 published results and comment?

By limit the cases are better for the Run II analysis for all channels and most mass points expect for mmem 2016 (20% better in some cases).

Here are the event yields:

2016 Analysis     Run II
Bkg_2016      
mmmt_2016 4.0   mmmt_2016 2.65
mmet_2016 9.7   mmet_2016 0.20
mmtt_2016 1.2   mmtt_2016 0.30
mmem_2016 13.2   mmem_2016 0.06
irBkg_2016      
mmmt_2016 1.2   mmmt_2016 1.05
mmet_2016 0.5   mmet_2016 0.09
mmtt_2016 0.03   mmtt_2016 0.06
mmem_2016 1.5   mmem_2016 0.23
a40_2016    
mmmt_2016 6.8   mmmt_2016 8.75
mmet_2016 2.8   mmet_2016 2.36
mmtt_2016 1.4   mmtt_2016 2.99
mmem_2016 5.7   mmem_2016 2.05
Here are the limits:
2016limit.png
2016 combined limit
plotLimit aa 2016 all 2016.png
Run II 2016 combined limit

- There is no reason for the m4l cut to be year-dependent, please make it uniform

Was done in 2016 and I disagree. Some channels had terrible statistics otherwise and wouldn't converge in unbinned parametric fit

- ZZ fits: add more bins given the good statistics. The fit quality does not look good, do you have a p-value/chi2?

The ZZ distribution and fit is the same criterion as the other categories. The choice of binning is just for visualization purpose and 4-bins were chosen to make the plot easier to read. The main focus is the error on the parameters... the central value must be lower than the error which was found in the fit range with painstaking effort. Another remark: Goodness of Fit tests were also conducted here: unbinned fit discussion And those goodness of fit tests closed.

- emu: If you want to continue using the simulation for the fake background, you need to evaluate the systematics (jet->e and jet->mu poorly modelled in simulation) and treat correctly the weights of the backgrounds in your signal region (eg it seems you have weights of ˜5 for the DY background, which would mean bins without content in the signal region have an uncertainty of about 10 events). Given the large weights of the simulation, I don’t really see how you can gain using MC compared to not making any prediction at all and letting float a smooth background (not making a prediction is probably more correct actually…). The best way would be to measure jet->e/mu fake factors and estimate this background from data.

So 2018 was the only one that used MC ... we will go back to FakeFactor -datadriven methodology. It will likely not effect the limit that much.

Other more minor items to be addressed as soon as possible (but could be after preapproval):

- Evaluate and add the h->aa->4 tau signal - Evaluate and add the other Higgs production modes (VBF, WH, ZH) - Switch to “new” lumi recommendations (138 total) - Do you really select taus with pT>18.5 GeV? Is this POG-approved? Do you have special scale factors/ uncertainties?

September 23rd 2022

minutes from meeting - Cecile: I cannot understand how the method is applied since FFs are measured in 3 different kinds of backgrounds, which are not the backgrounds of your analysis - Sam: it looks like we are using something else than what is shown in the slides and documentation. We should be using the SM ZH fake factors. Will correct the documentation and check that what we are using matches the new documentation

- Cecile: how do you treat uncertainty on empty bins, especially in the case of large MC weights (eg in emu)? - Sam: No special treatment

- Caleb: s14, why are the scale factors 1, 1, and 2? - Sam : by definition

- Danyer: what is the contribution from h->aa->4tau signal? - Sam: it was included in published 2016 analysis, contribution small, didnt have time to include before defense - Danyer: are the samples available? - Sam: maybe not, can use private samples - Cecile: need to include them to be correct in the NMSSM model, it is not a question of sensitivity - Sam: can we just scale? - Danyer: shape is different, you need different signal samples

- Danyer: the muon energy scale is changing the shape of the distributions. Is it a shape uncertainty in the fit? - Sam: no, norm only at the moment - Danyer: needed for the signal, which is a narrow resonance - Sam: computationally expensive - Cecile: could be easy as parametric uncertainties

- Danyer: what is the uncertainty on normalization of backgrounds? Seem to be floating from the impacts? - Sam: background normalization is floating - Cecile: why not use ZZ cross section from theory prediction? - Sam: can correct

- Mario: slide 39, which component of the background are you trying to model? - Sam: separate ZZ and data-driven backgrounds - Mario: is it based on MC? - Sam: for ZZ yes, not for the fakes

- Mario: the Fisher test should be feasible to decide the order of the polynomials

- Caleb: slide 44, are these postfit? - Sam: it is Asimov dataset with expected signal strength of 1, still blinded

- Alexis: we need item list to get preapproval - Cecile: will come back to you after a discussion - Valentina: we can open a CADI line, but to move to preapproval there are some comments to address first

- Cecile: comparison with old 2016 analysis - Sam: better by up to 50%, probably thanks to DNN - Cecile: asymptotic approximation could be too optimistic too

comments after meeting: Need fake factor explanation for files that were given and used.

Parametric uncertainties are not "easy" to implement - there is no documentation on combine.io or cmsdas that shows how to implement shape systematic uncertainties for unbinned analyses. Is this required if we are searching for any deviation within the fit range?

Unclear about needing the 4 tau signal - is this just for combination why is the NMSSM model limits incorrect without it? To what magnitude is it incorrect?

June 20th 2022

Draft AN2020_208_v4.pdf

Version 4 has been sent to HIG conveners. This contains all the major comments and address the concerns from the previous iterations (please see below)

Particular concerns addressed:

1. low muon pt disagreement ( fixed by trigger bug and trigger matching )

2. energy scale systematics on signal template ( included in overall mean RooRealVar from the spline 5% for now)

3. scale of CL limits were off ( certain normalizations were missing from the RooWorkspace ... fully fixed now )

Comments and Answers:

Stephane:

Lepton energy scale uncertainty: I did not really follow here - I think the important point is just the uncertainty in the width of signal template, not so much the background. I did not follow what that has to do with the spline - you can essentially just add the X% in quadrature to the uncertainty in sigma you already have from the interpolation, so technically speaking I think you just need to change that number. No? So I would just fit the signal template for nominal vs up/down, and use the difference in sigma’s to evaluate the total uncertainty to include X. I think that should be quite straightforward.

- Investigating alternative possibilities for the systematic shifts in the fit model - challeneges expected because both the spline is used and it is not entirely clear how to include an entire PDF as a shape based systematic in an unbinned parametric fit model. For now, the 5% shift is used on the signal and log-normal for the overall background in ZZ (irreducible). Which is similar to what the previous analysis did.

Muon trigger efficiency SF’s: I did not go in detail through AN-17-173, but I had a quick look at Section 7 and the 5% quoted is for the hadronic tau reconstruction, no? Isn’t the muon ID + isolation reconstruction efficiency uncertainty 2%? If there is a plot showing that uncertainty vs. muon pT is flat I did not see it, but maybe I missed it. A flat Y% uncertainty is probably fine, but it would be nice to demonstrate this should be flat from the POG efficiency measurement plots, and also to motivate the choice of Y%.

- Will consider adding this in the impacts, however, right now the low pt trigger efficiency scale factor ( as well as standard pog scale factors ) are not fully considered in the uncertainty model as it is expected to be a "second order" error in a statistically limited analysis. Lets say it's a 5% efficieny which is included on leptons for that event.. then the error on that is probably 1% or less ... propagated to a distribution that would only have 5 events in all years and all channels combined for the irreducible background.

Keti:

1. it's a final state with a pair of muons and taus in the final state format taken from the previous paper What I meant in my comment is that the way it is state is not clear how pairs are formed? Moreover you mention "a pair" (i.e. singular), while there are two pairs. Then it is not clear whether muons are paired together and taus are paired together or there are some mixed pairs (which for example is the case in some other searches). So, some rephrasing and clarification is needed.

- will consider rephrasing

2. your guess is as good as mine ... however, I would imagine this is a complex balance between resolved vs boosted decays coupled with the fact that lepton object reconstruction also gets worse or better in the region as a function of pt. Is the reconstruction changing when moving from low to high masses? That would explain that reconstruction, which is optimal for low mass, is not optimal for mid masses anymore, while another reconstruction is not applied yet. And as soon high mass reco is used, the efficiency is recovered.. Is something like this happenening? If no, i.e. if reconstruction/id/iso are identical for all masses, then this trend is very surprising and

- this effect was also seen in the last analysis, but will consider what is happening here. Please note that this is not the overall signal efficiency, but the lepton pair matching efficiency in identification of the "a".

3. please see CMS not on the SM HTT RunII measurement (AN-19-109) I am familiar with the FR method used in the SM HTT analysis. The fact that I got this question probably suggests that the description you provide in your note is not complete and/or clear and prompted my question. Maybe it would help to clarify the description or explain what is misunderstanding. The same applies to the following comment about orthogonality of the regions. This AN should be self explanatory -- maybe you do not provide a very detailed description and validation material, but it should be clear on how the method works and is applied.

-The FR methods are quite complex and I will consider rewording some things to make it clear. In the meantime, what precisely is the issue with the method or it's presentation in the note?

4. this is likely the deviation of the PDF on the fit itself. Not quite sure what it means "deviation of the PDF on the fit itself". Please clarify and also add this explanation in the text so it is clear how the 1,2/ sigma uncertainties are derived.

- Cecile made these plots because it's the SM HTT fake rates that were measured, so I'd double check; however, from what I see in the plots it's a line so there are two parameters or degrees of freedom in the fit. Those have an error associated with the fit, so the other lines are the error (1 sigma) propagated to each fit variable (so 4 extra lines total because of the +- sigma) and then re-plotted.

May 24th 2022

Draft AN2020_208_v3.pdf

Comments and Answers:

Stephane:

where do you get 2 GeV from for uncertainty in the lepton energy scale? I think the procedure you mention toi add it to the intmean in quadrature makes sense, this is also basically what I was trying to say. Just want to double-check the number of 2 GeV.

- This 2GeV simply provides a window that is acceptable to allow the mean to shift. I noticed in the muon energy scale systematic shifts that when the dimuon mass is plotted, events can migrate to neighboring bins in the signal peak (please consider looking at figure 4). These bins are in units of 2GeV, thus I felt it appropriate to assign that as the magnitude of the shift within the shape-spline for the mean.

mll fine mmtt inclusiveDown.png
mmtt dimuon mass Down muon scale
mll fine mmtt inclusiveUp.png
mmtt dimuon mass Up muon scale

low-pt muon SF’s, okay nice to see that with the SFs applied it already looks better. Are there uncertainties on these SFs that should also be propagated? Also why does the SF stop at 10 GeV, shouldn’t it extend up to at least 20 GeV? Were these SFs measured with respect to the same set of triggers as what you are using?

- Above 10 GeV, typical scale factors are used from the Muon POG - it's the window under 10 GeV that requires private scale factors.

Keti:

Abstract does not explain well the final state.

- it's a final state with a pair of muons and taus in the final state format taken from the previous paper

Remind me, what is the plan about using 2017/2018 private sample

- as discussed on page 7, nanoAODv7 private samples are used for signal for 2017 and 2018

Sec 3.2 why electron energy scale discussion is needed in this analysis?

- electron energy scale is needed for the final states where the tau decays to an electron - like the mmet and mmem channels.

lines 60-63: says that Fig 3 shows difference in data-MC for different decay modes. While Fig3 caption says something different, at least not different decay modes. It also does not explain all three figures.

- this was taken from the Standard Model HTT paper. I'll try to reword to make it more clear and indeed the figure is just one of the decay modes (1prong + pi0)

L81: Sentence "Scale factors are measured ..." should come before the previous sentence, i.e. after "mismodelled in simulation."

- ok

Fig 4. What is a process 3alpha or 4alpha?

- I will provide a description; however, these are "rare" processes that often invole either 3 or 4 electroweak couplings in the interaction.

Table after line 161: What is the reason for noticeably lower efficiencies for mid mass points, while higher efficiency for very low and very high mass points?

- your guess is as good as mine ... however, I would imagine this is a complex balance between resolved vs boosted decays coupled with the fact that lepton object reconstruction also gets worse or better in the region as a function of pt.

L174-177: Do you ensure matching between selected muons and trigger objects? In particular for the phase space with low-pT offline muons.

- muons are selected and delta-R cleaned and should match with the trigger that used to select the event

Lines after L177: It is a bit confusing here. You already talked about muon pT>5 GeV and now you again are talking about muon pT > 15 GeV. Why electron reconstruction is relevant here? It does not harm, but I am just trying to understand what it helps...Also L 195-196.

- the muon threshold is for the trigger I'll place that in the description. And for electrons, we still need to apply criteria for the electrons coming from taus. Naturally, a description of muons, taus, and electrons are required for a full explanation of event selection.

Fig8. There two muon pT plots.. What l1 means in bottom left plot?

- leading lepton of highest pt muon, thanks adding it to the description!

L207: "events passing loose" loose what?

- added "loose identification"

Fr equation: While I am guessing you are doing correctly, the equation implies as if Prompt MC Bkg quantity in numerator and denominator are the same. You need to call those differently to distinguish quantities for loose and tight regions.

- nomenclature is difficult ... will add Tight and Loose to numerator and denomenator more explicitly

L215: Do you mean you measure FR individually for each process and then apply those FSs to respective individual control region? Why do you expect FR to differ between processes? Especially, how to ensure to measure individual process FRs in data?

- Following the SM fake rate measurements - I've used their measurements literally so they split it by process and indeed one can imagine that the production mode of the hadronic tau might induce events with differing levels of missing energy from the neutrinos so I think that is why they split it.

L225: I am bit confused. Usually FR is measured in a sample dominated by non-genune taus, i.e. fake taus, where you check what is the probability of those taus passing tight selection if they are already passing loose selecting. However, here you explicitly select ttbar sample with fake taus, as I understand.

- please see CMS not on the SM HTT RunII measurement (AN-19-109)

L229-235 (in relation to L215): How you ensure orthogonality of application regions for different bkgs? E.g. For WJ and TT, you subtract all MC predictions for genuine taus and apply (respective, I assume) FRs to the remaining sample. What happens to QCD, for which there is no MC prediction?

- please see CMS not on the SM HTT RunII measurement (AN-19-109) : basically for QCD that is the mis-modelled descrepancy in the determination regions!

L242: Which two figures do you mean here? Fig 24 does not seem to be the one.

- please consider taking a closer look at the caption there is the fit in data and simulation for et and for mt channels respectively.

FR Fit figures: How 1st and 2nd uncertainty +/- 1sigma fits are designed? I could not find where it is discussed.

- this is likely the deviation of the PDF on the fit itself.

Fig29: There is large discrepancy 2018 mmmt plots, and I believe it is coming from jet-fake bkg. Have you looked what causes this discrepancy?

- when combining all the data for Run II the descrepancy goes away because the 2016 under estimates it. Please refer to Figure 8. I believe it's consistent... similar things happen to the mmtt channel. Please note that this is not the fit variable and furthermore, extra signal extraction cuts are applied.

pt 1 combined.png
Run II leading muon P_t

329-331: Periods are missing from these sentences.

- Thanks! It's been noted and adressed.

May 10th 2022

Draft AN2020_208_v2.pdf

Comments and Answers:

Table 4 (and related): You consider many different background samples here - that is in principle fine, even if some do not significantly contribute in the selected region, but can you please confirm that the MC samples are non overlapping, and/or that you account for any overlap in samples? I just want to make sure that you are not double-counting any contributions. (I admittedly did not check through every sample on the list, it’s just a general comment).

- The SM Higgs to Tau Tau ZH measurement sample list was used. For samples that "double count" - for example W1Jets and WJets - an inclusive cross section and separate cross section is combined and used for these events. So with the correct normalization, I can use both. With the tails being super important in this analysis we try to include everything we can, lots of legacy samples are also used in the fake factor calculation through the prompt MC subtraction.

Section 4 - as far as I understood you primarily trigger with single and/or double muon triggers with > 20 GeV thresholds. But then you require an offline muon pT of 5 GeV right? Do you apply a trigger efficiency SF, or are you sure that you are nonetheless on the trigger efficiency plateau?

- Correct, an offline 5GeV threshold is used. To address these concerns, I've incorporated the following scale factors that were measured in the 2016 analysis: Muons with 5 < pT < 9 GeV 0.956 0.930 Muons with 9 < pT < 10 GeV 0.916 0.897 These are to address low Pt muons in the barrel and endcap that pass a higher threshold trigger. NEW results will be posted to the AN with these changes soon.

Figure 28: It looks to me like you are missing some contribution in the data at low muon pT here, the discrepancies are too large to be statistical. I think this is a concern, since this is also the region where you expect to see the signal. I think this could be due to the difference in off-line vs. online muon pT requirements, along the lines of what I ask in the comment above.

- With the previous tag and probe method done in the 2016 analysis, there is confidence that the low muon pT is corrected in a consistent manner. The disagreement is only prevalent in 2018. 2016 is low and 2017 closer to the mark; therefore, under combination this fluctuation will go away.

pt 1 mmmt inclusive.png
Run II mmmt leading muon P_t

Section 11 - doesn’t the lepton energy scale uncertainty affect the mean of the signal peak, more so than the signal acceptance? I would recommend including the lepton scale uncertainty as an uncertainty in the mean of the signal peak - I suspect this effect is larger than the 0.1% uncertainty you quote for the uncertainty in the interpolation of the mean. The lepton resolution uncertainty also affects the shape via the sigma, albeit this is much less than the 20% interpolation uncertainty so considering it should make no difference.

- We are following up on this point; however, I think technically speaking this should be covered in the intMean systematic. Even though param options are available it is not clear how this would be propagated to the datacard and correspondingly the likelihood function. Under inspection of the systematic shifts that are listed in the AN a 2GeV bump up and down affects the shape. This would directly effect the signal mass hypothesis so to address this point at the moment another systematic uncertainty will be added in quadrature to the mass hypothesis so instead of 0.1% I will consider 0.1% + 2GeV... which is mostly just the 2GeV bump ~ 5%.

Table 12: what signal cross section and BR are you assuming here?

- Added to the AN, the signal is ggF only for the yield so 48.37 pb and the branching ratio is 0.01%.

Results: why is the expected sensitivity more or less the same for each year. Isn’t this a statistically limited analysis? I would have expected something more like sqrt(L) improvements. Then the full Run-2 limit is an order of magnitude better, which seems too much - can you please check this?

- With the background and signal normalizations added to the RooWorkspace this has been addressed. The limit model was lacking the normalization scaling. A new updated version of the results are more reasonable:

plotLimit aa 2022 all 2022.png
RunII Expected Limit

Results - do you show anywhere the distribution you fit to extract the signal? I think it’s important that you show blinded plots of the m(mumu) in the SR, to demonstrate what you are doing with the signal extraction fit.

- Of course! Please see figures in the appendix for the raw templates and the splines for what is used in the likelihood model. And for the distribution of the fit variable (histogrammed) within the signal region with no data please section 3.10 Visualizing the Corrections and Intro to Uncertainty Model.

Topic attachments
I Attachment History Action Size Date Who Comment
PNGpng 2016limit.png r1 manage 112.1 K 2022-10-18 - 17:25 SamHigginbotham  
PNGpng 2_Nominal_Bkg_2016_mmmt_Nominal.png r1 manage 13.8 K 2022-10-14 - 12:02 SamHigginbotham  
PNGpng 2_Nominal_irBkg_2016_mmmt_Nominal.png r1 manage 16.0 K 2022-10-14 - 12:02 SamHigginbotham  
PNGpng 40_Nominal_a40_2016_mmmt_Nominal.png r1 manage 20.4 K 2022-10-14 - 12:02 SamHigginbotham  
PNGpng AMass_blinded_combined.png r1 manage 20.8 K 2022-06-03 - 02:29 SamHigginbotham  
PNGpng mll_fine_mmtt_inclusiveDown.png r1 manage 13.7 K 2022-06-03 - 02:35 SamHigginbotham  
PNGpng mll_fine_mmtt_inclusiveUp.png r1 manage 13.7 K 2022-06-03 - 02:35 SamHigginbotham  
PNGpng plotLimit_aa_2016_all_2016.png r2 r1 manage 19.0 K 2022-10-18 - 17:41 SamHigginbotham  
PNGpng plotLimit_aa_2022_all_2022.png r1 manage 19.7 K 2022-10-07 - 18:16 SamHigginbotham  
PNGpng pt_1_combined.png r1 manage 18.0 K 2022-06-03 - 02:29 SamHigginbotham  
PNGpng pt_1_mmmt_inclusive.png r1 manage 18.0 K 2022-05-18 - 14:36 SamHigginbotham Run II mmmt leading muon momentum ... discrepancy falls out when all stats are considered
Edit | Attach | Watch | Print version | History: r22 < r21 < r20 < r19 < r18 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r22 - 2022-11-15 - SamHigginbotham
 
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    Sandbox All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright &© 2008-2023 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback