Skip to content

ParKalman Filter issue with high momentum, large slope tracks (e.g. Z-> l l)

As pointed out by the bandwidth division team, there is an issue with the chi2 and momentum estimate returned by the ParKalman filter for some of the samples used in the bandwidth division and specifically the SingleHighPt lines for muons and electrons, in the following I'll discuss the case of Z->mu mu.

The observation of the issue is the following: image Here we see that the SingleHighPt line is less efficient for very large pt with the ParKF. The observed drop in selected events in this test is 5%. We also see that the line selects more high-pt tracks than there really are, i.e. ghost and the the ParKF does a better job of rejecting these than the velo-only kf.

Further study shows, that the ParKF seems to break down for Z-> mumu tracks especially if the initial absolute track slope in y is large (abs(ty)>0.1), and to some amount the same is also true for tx. image image

However, it's worth point out that it's not generally true that these slope values are problematic, rerunning these test with a Bs->mumu sample does not yield the same issue, so probably a combination of large slopes and high momentum.

Proposed solutions:

  1. As a fairly easy work around we can also attach the chi2/ndof of only the Velo (still with parKF) to the states and also the momentum estimate of the tracking algorithms. This largely mitigates the observed issue in Zmumu. See MR (In my test this still reduces the number of events selected in the line 2%)
  2. Another idea was to regenerate the parameter set with a sample containing a sample of Z-> mu mu samples. This did not yield the wanted results. ( I did achieve minor improvements for all track, so we should still do this once new sim is available.)
  3. Rework the parametrisation [current TODO].

Updates:

  1. [31.03]: the issue really seems to be with the very low assumed scattering noise for very large momentum tracks.
  2. [31.03] Current working theory: For very large momentum tracks, the MS is small enough that the intrinsic extrapolation/parametrisation error is exposed. Since this is not accounted for in the ParKF, chi2 become very large, momentum estimates erratic etc. By implementing a lower cap on the noise matrix The following result can be recovered, with CHANGES being the aptly name version to look at. There is no loss of tracks compared to the velo_only kalman filter. image Further test and studies are needed see MR

fyi: @dovombru @decianm @aiwieder @ahedes

Edited by Lennart Uecker