When KF uses track parameter and stub residuals from f/w the z0 resolution degrades by a factor of ~1.5 over all eta.
Creating fake seed stubs (seed stub on tracklet seed parameter at average barrel layer radius / disk z position) alone degrades z0 resolution noticeably, which means track parameter are corrupted. Looks for me like +-1 errors on integer values for z0 and tanL. Due to the roughness of those parameter the +-1 errors become significant.
Looking at Stub residuals I see more than +-1 errors (where I compare with simple calculated residuals). Actually the tracklet code uses higher order corrections so that one expects to see errors but those should improve z0 resolution and not degrade...
When KF uses track parameter and stub residuals from f/w the z0 resolution degrades by a factor of ~1.5 over all eta.
Creating fake seed stubs (seed stub on tracklet seed parameter at average barrel layer radius / disk z position) alone degrades z0 resolution noticeably, which means track parameter are corrupted. Looks for me like +-1 errors on integer values for z0 and tanL. Due to the roughness of those parameter the +-1 errors become significant. Looking at Stub residuals I see more than +-1 errors (where I compare with simple calculated residuals). Actually the tracklet code uses higher order corrections so that one expects to see errors but those should improve z0 resolution and not degrade...
The code is now in our dev branch, by default the s/w takes the residuals from TrackBuilder, one can use instead recalculated resiudals by settings this handle to true: https://github.com/cms-L1TK/cmssw/blob/L1TK-dev-12_0_0_pre4/L1Trigger/TrackFindingTracklet/python/Producer_cfi.py#L20