ParKF: set pt limit on noise matrix scaling to fix Z-> ll Problem
Related to issue 590.
Diagnosis.
The problems observed in the issue stem from the fact that the addition of the noise matrix doesn't properly account for the extrapolation errors in the V->UT step.
When a state is predicted in a Kalman filter we expect. x_{k} = f(x_{k-1}) + sigma_k. Where sigma_k is the result of multiple scattering and generally gaussian and consecutive scatterings are independent. sigma_k behaves like 1/p and accounted for by adding a noise matrix Q after each prediction step to the state covariance.
In the parKF there is an additional error x_{k} = f_par(x_{k-1}) + sigma_k + err_kpar. err_kpar is in general not gaussian and consecutive errors are highly correlated. In general we expect err_kpar to be large for large x,y and tx,ty, pt i.e. the empty regions of the detector phase-space.
The ParKF relies on the fact that the parametrisation is good enough for err_kpar to be dwarfed by sigma_k and thus not relevant.
For these high momentum tracks at large slopes, sigma_k ~1/p vanished an the ParKF is confronted with underestimated covariances and non-gaussian errors.
Proposed solution: For track with large pt, here > 10_000 MeV, a lower limit is set on the elements of the noise matrix Q to keep coverage over the extrapolation errors.
The Good:
The vanished Z-> mu mu tracks are recovered. In fact the new shape is closer to the truth than the selection from the default sequence.
The momentum resolution in general in unaffected in p< 100_000 MeV, slight changes above.
General improvement in momentum pulls for muons from Z.
The not so good:
The Chi2/ndof distributions for mu from Z specifically still show some not nice behaviour in the rare cases, but generally following a reasonable distribution. Please note, for a minbias or B sample all distribution here look good.
Merge request reports
Activity
mentioned in issue #590
added RTA label
added hlt1-throughput-decreased label
requested review from @thboettc
removed review request for @thboettc
removed hlt1-throughput-decreased label
requested review from @thboettc
