RD: fixes to radinclusive lines
Description
- Updates the BDT used in HHG radiative inclusive line so that it doesn't rely on CorrMass -> very high priority from rad because right now it selects candidates biased on their mass and compromises the whole utility of the line
- Fixes CHI2 inputs to the BDTs since in run3 for the moment F.CHI2 on a combination of a neutral gives 0 instead of the CHI2 on the next real vertex in the decaytree (what happened in run 2)
- adds BDT output monitoring
BW impact
TLDR: this brings down the rate in full from 69.2 kHz to 68.7 kHz
PR tests
Results of the PR tests can be found here
| Line | Ref Rate (kHz) | New Rate (kHz) | Change (%) |
|---|---|---|---|
| Hlt2RD_BToHHGamma_InclDecision | 0.752 | 0.912 | +21% |
| Hlt2RD_BToHHGamma_GammaToEE_InclDecision | 0.262 | 0.49 | +87% |
| Hlt2RD_BToHHHGamma_GammaToEE_InclDecision | 0.16 | 0.171 | +7% |
| Hlt2RD_BToHHHGamma_InclDecision | 0.821 | 0.0912 | -89% |
| Parameter | Old | New |
|---|---|---|
| Full stream rate (kHz) | 69.2 | 68.7 |
| Full stream DstData Bandwidth (GB/s) | 7.76 | 7.68 |
| RD full pseudo-stream rate (kHz) | 1.8 | 1.5 |
| RD full pseudo-stream DstData Bandwidth (GB/s) | 0.5 | 0.4 |
| Overlap of RD with topo | 27% | 43% |
The +21% rate in HHGamma was a conscious choice and is physics motivated as described below.
The -89% rate in HHHG is a bit drastic and comes from the fix of the chi2 input. At some point I would like to update the BDT and loosen up a bit the working point (wp) to gain back signal efficiencies.
The effect on the converted photons lines is quite reasonable.
I'm not sure if the increase of overlap with topo comes from the fact that the HHHG line is reduced a lot, from the change in the HHG BDT or a combination of both (probably).
Local Tests
I get a rate of 1kHz for HHG instead of the 0.75kHz it has now, so potentially a 0.218 GB/s increase to full. I will update it when PR tests results run. I understand we are quite tight for BW increases now but it is quite compensated with the rate decrease in the converted photons counterparts. It was indeed designed that HHG and HHHG had higher rates than the converted photons ones because the stats are lower on the second, so this converges to that situation. This used to be the case and it changed over the samples used and reconstruction versions, I'm not sure why
Efficiencies
The plot below shows BDT efficiencies for different channels by comparing the amount of events selected after each BDT cut with the total amount of events selected at preselection level. As expected the BDT performs better on channels with two opposite charged tracks in the final states with respect to channels with V0s in the final state and partially reconstructed channels where a pi0 is missing.
This plot should not be trusted completely: since the tuples were produced before the neutral truth matching fix in LHCb!4453 (merged) only the hadronic part was matched so the reference signal sample is not 100% free of background (especially true for pi0 modes). Furthermore the usual reconstructible requirement on the photon to correct the denominator couldn't be applied either. Lastly, the ROC is obtained by tupling at preselection level and looking at the efficency of the BDT per possible working point and accounting for multiplicity but apparently the expected rates per working point by extrapolation don't match precisely to the actual rates when running that cut (up to 30% mismatch).
All that said, even with the aforementioned uncertainties it's quite safe to assume that the signal efficiencies are quite sensible to the BDT working point in the range of possible realistic values that meet BW requirements. A rate of 1kHz was chosen to stick to the original expectations when this lines were first implemented. If the situation with BW makes it too difficult to accept an increment now I would offer to tighten the BDT working point to 0.29 which would recover the rate in master, and merge anyway since the most urgent mater is to fix the bias in the present BDT.
