Questions and Answers

The following questions are meant to address the bad/worrying features of the regression.

• The questions refer to the slides in this link, from E/γ/DPG talk.
• The answers are included in this updated version of the slides ECAL DPG talk of 24 of September

Globally: what is the energy taken for the super-cluster ? Do you get it from ele->superCluster()->energy() [general SC corrections] or with ele->ecalEnergy() [class-dependent SC corrected energy]

• in the labels through all the slides the convention is:
  • raw SC energy: electron→superCluster()→rawEnergy()
  • std SC energy: electron→ecalEnergy()
  • regression enerrgy: raw SC energy with the regression applied

Can you cross-check that it is the same definition which is used throughout the slides ?

• yes, it is always like this.

Slide 3: can you tell us again what is the sample used for the training of the regression ? It is written DY but at the last presentation in e/g it was said that HZZ was used from training.

• right. Since there was a slight loss of performance due to the training in 52X and application in 53X (because of some changes in reco, like the supercluster algorithm in the EE) we retrained the regression 1 in 53X DY sample (Madgraph one).
• The performances in the slides are shown on a sample of H→ZZ→4l, mH=125 GeV (53X).

Slide 4-6: the fits are not very good. To determine the peak position (slide 6), it would be better to restrict the fit range.

• done. Restricted to -0.1 - 0.1, since we are interested in the core resolution (when estimated peak position and sigma. Asymmetry / tails are accounted by the mean/RMS of the distribution). The plots are updated. We show also a larger collections of fits in the slides

Slide 4. For the 7-10 GeV the (Ereco-Etrue)/Etrue is not representing the bias on the mean showed in the following slide 5 (first bin on the upper left plot). The mean is 0.01 (slide 4) why it should be 0.05 (slide 5). Could you show the (Ereco-Etrue)/Etrue plot for the first bin on the upper left plot in slide 5. It would be nice to see this plot in bin of eta.

• beacuse the fit examples in slide 4 were done with (p-ptrue)/ptrue (where p is the combination of P(GSF)-E), which is not in the scope for this presentation. Removed them and shown the ones for the E(regression)

Slide 4. The (Ereco-Etrue)/Etrue for upper-right plot shows a bad asymmetry on the right, which is shifting the mean towards higher values. This most probably is larger on the lower pt bin (7-10) GeV.

• this is not particularly created by the regression. It is very similar to what the standard-corrected ECAL energy show. It is mostly for showering electrons.

Could we explain/understand the origin of these tails on the right?

• the right tails are for low pT (both barrel and endcap, especially for showering electrons). Both the standard and regression corrections show it. Regression tail is less pronounced, it is very visible outside the training phase space. To avoid border effects we should train probably from pT> 4 GeV

Slide 5: upper left plot: There is a clear problem with the lowest pT bin where there is a significant overshoot which must come from the tails. Can you please distribute the plot (probably in log scale) with a range sufficient to see the tails ? It could point to a problem in the training of the regression. What is the electron selection applied before training the regression ?

• indeed the very first pT bin (7-10 GeV) shows a larger right tail than the standard ECAL corrections. This could be a border effect of the training of the regression (or due to limited statistics of the DY MC). To be checked, but it has not impact on the p (after the E-p combination) because in that bin it is dominated by p. The plot in log scale is included now in the presentation.

Then, on the upper right plot, one can see that the mean value is ~0.02 while from the pT-dependent plot, only the first bin has a value greater than 0.02. Can you check that this systematic shift is due only to the low pT electrons ? (by doing this eta-dependent plot for several different pT bins: 7-10; 10-15, 15-20, 20-)

• Yes. Indeed, divinding the eta plot in bins of pT, the resolution decreases drastically. Already in the 10-15 GeV bin the resolution is at ~1% for most of the barrel and ~3% in the endcap. All these estimates are after having aspplied the extra smearings to the MC.

Slide 7: can you adjust the bounds of the plot so that one can see all the points ?

• Done, in the next version of the slides.

Slide 10-11: the events "moved" away from the true value should be investigated, as their fraction is not negligible. Practically, one should first understand if the migration comes from a given category or from a given pT range. Do you have enough statistics to do the plot by [class] x [pT range] ? (same ranges as above). If not, do the plots by class and the by pT range.

• they are concentrated in the pT bins, in particular in the bin 7-10 GeV. I need the fraction on each class to conclude quantitatively

In addition, it is important to quantify the migration: one convenient variable is |E_corr-Etrue| - |E_regress-Etrue| Do the plot of this variable for all electrons, for the electrons for which the regression is improving the energy, and for the electrons for which the regression is degrading the energy.

• done. There are no visible difference in the distribution for the worsened and improved electrons. we need the "all electron set", even if the most significant are there

Hopefully, in the process, the origin of the well measured electrons with the standard corrections but overcorrected by the regression (slide 10, bottom right plot, |Esc-Etrue|<0.1 && |Ereg-Etrue|>0.3 ) should appear and cure can be found; if not, please isolate these events and plot their eta, phi, E_raw distribution and other relevant variables.

• we should see if a retraining with an increased pT range cures the very low pT bin

Slide 5-11: looking at the upper right plot, one would say that the mean value of the E_SC is higher than E_regression which contradicts the plots from slide 5.

• it is again one effect of being inclusive in pT. If one just excludes the 7-10 GeV electrons, the regression gives the mean and resolution peak less biased in almost all the eta range. Included in the new version of the slides.

-- EmanueleDiMarco - 27-Sep-2012