-- ZebingWang - 2021-12-14

Search for dark photon in four lepton channel at 13 TeV

Color code:
* Answered
* Open discussion
* Not answered

comments v2 from Stephane

Abstract - add integrated lumi number

Introduction - you actually go down to mass of 1 GeV not 4 GeV, right?

Fig 5-8 - In general the data/MC agreement does look fine, but it is a bit hard to tell given the statistical variations and the large range in the ratio panel. Can you please remake the plots with fewer bins (factor 2 less perhaps?) and with the ratio panel significantly zoomed in? 0.5 to 1.5 at least if not more zoomed in - otherwise it is very difficult to assess the level of data/MC agreement in the bulk of the distribution. The main point from these plots for me should be to assess as precisely as possible the level of data/MC agreement in the regions where the bulk of the signal is located, so please try to adapt the plots a bit along these lines.

Similar for Figure 10, although make sure to show granularity at high score still comparable with the category boundary chosen, i.e. if the boundary is 0.95 then have bin sizes of at least 0.05 at high score.

As far as I understood you are essentially performing a mass parametrization for the BDT (nice!). The method looks good to me, but have you verified that the BDT classifier does not sculpt the mass? Basically I want to confirm that the tight BDT selection you apply does not change the shape of the background near 125 GeV in a way that would not be covered by the analytical background models you consider. Let’s consider a pathological case where for the mA = 30 GeV hypothesis when you give mA=30 GeV for the evaluation of the background you get a sharp mass peak at mA= 30 GeV, which gives you basically a peak at 125 GeV mimicking a signal. I realize that your MC statistics are very limited after applying the BDT selection, but maybe you can show the 2D profile plot of the BDT score vs. mA for the background MC under the different mA hypothesis? If that shows no significant correlation then it is probably enough.

Fig 9 - why is the signal KS test value 0? By eye it looks fine and the statistical fluctuations seem small so I was surprised that it is 0.

Fig 9 - what is the typical signal efficiency of the selections you apply for the BDT categories? You are imposing quite a tight selection and I am not sure that we really see this region on the ROC in this figure. Can you clarify the relevant region for the analysis and perhaps change it to sig eff. vs. log(1-bkg eff.) so that we can see the relevant region better?

Eq. 4 - why not apply a one or two sigma mass window selection to define S and B, rather than the inclusive signal region? Probably it does not make too much difference, but it could change slightly the optimal working point.

I did not quite follow exactly how you model the background for the signal extraction fit, is it: a) a single analytic function per category that was determined to be the best-fit prior to the fit b) a discrete profile per category, i.e. each category can fit a different functional form c) a discrete profile over functions but correlated across categories, i.e. same function for all categories ?

Systematic uncertainties in section 10 - what about any nuisances that affect the signal shape? i.e. photon momentum scale and resolution? Perhaps also the lepton momentum scale and resolution, I am not sure if these are negligible or not.

How exactly do you evaluate the photon efficiency uncertainty of 10%? Since it is your dominant systematic uncertainty I think it is worth expanding on that.

Fig 15 - how do you construct the uncertainty band on the background prediction in these plots? There were a few things that stood out to me on the plots: a) the nominal prediction is far from the center of the one sigma band b) there are clear kinks in the uncertainty band c) From 110 to 125 GeV the uncertainty band for the top left and bottom right plots looks a bit strange and different from each other. Can you explain further the shape of these uncertainty bands? Maybe it is just the Poissonian nature of these uncertainties and the limited number of events and/or limited toys in the determination of the uncertainty bands, but I think it would be good to check things further and make sure the uncertainty bands make sense.

Tab 21 - with what criteria did you decide on a sum of 4 Gaussians for the signal modeling? It’s clear that the single gaussian is insufficient but the gain after 2 looks relatively marginal.

Sec 9 - what is the difference between the two sets of signal mass plots? The selections applied? Because the sum of 4 gaussians by eye seems to fit much better in the second set of plots…Sorry if I missed the explanation in the text.

Results - Can we interpolate between mass points? Because at the moment the limit can change by as much as a factor of two between mass points on the grid. Maybe we can just parametrize the signal shape as a function of the mass and interpret the limits ~continuously?

Edit | Attach | Watch | Print version | History: r1 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r1 - 2021-12-14 - ZebingWang
    • 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