Quark-Gluon Tagging: Machine Learning vs Detector

First of all, we would like to thank the referees for their careful
reading and for their comments. We agree that this paper essentially
has a negative bottom line, and we also agree that we cannot
enthusiastically report the impact of quark-gluon tagging on a
reference process. However, we believe that these finding should be
out in the public, and if only to encourage others to do better or to
drop quark-gluon tagging from the problems worth pursuing for LHC
searches.

Because many of the referees addressed the same points, we have collected them together for a combined response.

• We have changed the title to "Quark-Gluon Tagging: Machine Learning vs Detector".
• We added many references on quark-gluon tagging and on IR safety.
• We now refer to 1711.02633 for a similar study.
• We specified LoLa in the abstract and the introduction.
• We slightly modified the introduction of the two reference processes.
• We added a reference to the difference of q-g tagging based on Pythia vs Sherpa.
• We streamlined Sec.2.1.
• We expanded the discussion of Fig.2, relating our findings to the available literature.
• We expanded the discussion of Fig.3, including the relation to the incomplete and non-basis of observables.
• We added discussion of IR safety at the end of Sec.2.1 and now explicitly mention girth and C2 as safe.
• We now refer to the detailed comparison of architectures from 1810.05165 and make it clear that our new focus is on detector effects. We do not believe that a comparison to more than the CNN of 1612.01551 would add to the conclusions of the paper.
• We clarified the caption of Fig.7, how we extract MC truth from our simiulation, and the discussion of Fig.8.
• We unfortunately have no way of estimating the exact effect of quark-gluon tagging on a specific di-jet resonance search, but we relate it to event-level observables and a similar analysis.
• We have added a (blunt) bottom line to the mono-jet discussion, but we believe that the discussion should be kept in spite of the negative conclusion.
• We added a couple of references and clarifications as requested by the referees, including on pile-up.

A couple of points we could not change are:

• In the discussion of Fig.9 we already discuss the fact that for a stiff MET cut the quark-gluon tagging performance suffers.
• We are sorry, but adding reliable rejection efficiencies to Tab.1 would require us to use much more GPU power than we have. But an example number is 1/FPR=9.3 @ TPR=0.3.