Trigger Configuration Study for U(3s)

Samples

This study used the signal MC L1TNtuples generated by Steve Sekula for U(3s)->U(1s)+X  where X can be p⁺p^-,p⁰p⁰ or gg while U(1s) goes to invisible decays. For L1 rate studies, some old Run 6 ntuples for logged L3 events were used.      

U(3s)->U(1s) p⁺p^-

MC truth fiducial is defined as both truth p⁺p^- are within 20^o<q<150^o. This corresponds to the SVT tracking fiducial. Tracks exiting DCH at SL5 is only slight wider than this but close enough to use the SVT fidcial (beyond which the material spoiling of tracking may become serious anyway). Both truth p⁺p^-  also should have Pt>60MeV/c which is also the tracking limit. ~78% of all generated events are in the truth fiducial. The various kinematic observables of the events are shown below (black line=all event; blue filled=MC truth fiducial):       

One could also further look at the correlation of the truth max and min Pt for all events (black) and fiducial events (blue filled):

The fortunate part is that the kinematics of events prefer balanced Pt between the two pions which is easier to trigger than the one high one low case. Only a small fraction of events falling in the corner of both tracks below 250 MeV/c. Higher Pt track cut lowering from 600 MeV/c to 400-500 MeV/c can cover a fair number of events independent of whether the lower Pt track is found.      

We can then further select analysis reco fiducial events by requiring at least one +ve charged track and at least one -ve charge track, with |d0|<2cm; |Z0-Zip|<5cm. 63% of the truth fiducial events pass the analysis reco fiducial (or 49% of all generated events). For the existing L1 & L3 events, 52% of the reco fiducial events pass L1, 16.6% passing both L1 and L3. The absolute total efficiency of triggering (both L1&L3) for all generated events is therefore ~8.1% for the existing trigger as a starting point.    

L1 Trigger

Let's first looking at the L1 efficiency. Majority of the events relied on the 2Zt&1A&1M trigger to get through. The main choking criterion  is the 2Zt. Various alternatives are examined:

For the analysis reco fiducial events, the following results were found by trying the various combos of the test triggers:

The emulated L1 line pattern for the last scenario of ORing T1|T2|T3|T4 to the existing 2Zt&1A&1M and displacing several old lines to accommodate T1-T4 is shown below:

For displacing old lines, it's worth noting that the line 3M+D2+1Z is close to a subset of T3 and T4 is a much better deal than the prescaled M*&1z. The lesson from looking at the trigger lines is that the 2-particle triggers generally cannot be totally efficient, not even D2 or 2M, which makes this a rather tricky task. One has to scramble with a combination of many partial coverage solutions. The GLT object counts for all reco fiducial events (open block line) and those would fail the new L1 trigger (filled blue):

The remaining untriggered events are becoming to hard to grab. Some additional combos may get another 1-2% at best.

To test the new trigger line rates live, Rainer and I looked at the test lines and their common denominators to come up with the following configuration changes for a short rate test run 77924 on Saturday Jan/05/08 (L1 lines are reordered also) before beam lost due to thunderstorm:

The overall L1 rate was 77924 (Init Lumi=7.0E33) was 2474 Hz. A previous run 77922 (init Lumi=6.5E33) with current configuration was  2108 Hz. Judging from the raw rates,  T0 alone may be indeed too high a rate to be affordable as initially suspected. We therefore decided to try the existing 2Zt&1A&1M|T1|T2|T3|T4 solution as a real physics configuration which should be affordable for the current luminosity at least. This was brought into physics running since run 77946 at a luminosity ~7E33. The L1 rate detail snapshot can be seen here and the stripchart of L1 rate at the initial operations can be seen here.  

L3 Trigger   

To examine the L3 inefficiencies for the reco fiducial events:

One can see that  for those events passed existing L1 to get L3 a chance to process on them, roughly half of them failed L3 T0. Unfortunately the actual T0 value and kinematics are not very different between the events succeeded and events failed the L3 T0 finding. This cause for the T0 finding problem may be somewhat subtle. However, once the T0 is found, L3 is mostly reconstructing 2 tracks and one can further reduce the 1-track Pt cut down from 600 MeV/c so that it is hopeful that the solving the T0 finding problem may quickly recover most of the efficiency.   An interesting question: is the L3 T0 finding efficiency different for events passing the old 2Zt&1A&1M trigger rather different from the extra L1 events picked up by the new test L1 triggers ? They turn out to be essentially identical: 47.8% vs 47.5% respectively using the new ntuples generated by Steve with the test L1 trigger configuration. This means that the L1 trigger efficiency gain will directly translate to same total efficiency gain before changing L3. However, this further deepens the mystery as to what's governing  the L3 T0 finding failures.  

Before solving the T0 finding problem, we can try to solve the remaining L3DCH algorithm inefficiency first. For the reco fiducial events triggered by L1 and with successful L3 T0 found, it turns out we are still losing ~44% of them. The L3 2-prong filtter has a Pt cut of 250 MeV/c, while the L3 tracking can actually go down to ~230MeV/c. So we can simply drop the Pt cut on tracks for the 2-prong filter, which will recover ~25% of these events failed default L3DCH. We can then try to dial down the Pt cut for the 1-prong filter. If it is dialed all the way down to 200 MeV/c (effectively no cut), 97% of the previously failed events will be recovered.     

However, dialing the 1-prong Pt cut all the way down surely will blow up the L3 logging rate? Apparently not. Testing the same loosening on the L1 pass through sample from run 67092 from Run 6 (only on the L1 physics lines - excluding 1Y&1B): 

The loosening of all Pt cuts apparently will only add 11% to the L3 output rate. This seems to be too good to be true, but the single IP track in L3 was the original L3DCH algorithm. Why did we spent the trouble to make it 1+2 prong  IPtrack filter ? I could not find the original record of arguments for the change, but it could be due to a combination of reasons e.g. offline processing bottleneck in the early days can be helped for even 10% rate reduction ? or we simply try to meet our 120Hz at 3E33 exactly ? If this is true then it is great news and change should be totally trivial: just edit two pt cut numbers in the trigger config tcf.  

So in principle, just the L1 change 50%->80%, L3DCH pt cuts loosening can takeL3 eff 47.6%->98.5%. These two combine to a factor of 2.8 increase in efficiency already before solving the L3 T0 problem which may be close to another factor of 2. So we have the prospect of a total of a factor 5 gain in trigger efficiency to boost the absolute efficiency to 40%.  

U(3s)->U(1s) p⁰p⁰  

For simplicity we take the starting point as the events with both p⁰ decaying into gg, ignoring the 1.8% events containing a Dalitz decay of p⁰->eeg. The MC truth fiducial selection requirement of all 4 photons in the EMC volume has a geometrical efficiency of ~67%. 96.7% of the truth fiducial events have at least 4 reconstructed EMC candidates. This gives a total fiducial event selection efficiency of ~64.5%. The basic kinematics is similar to U(3s)->U(1s) p⁺p^- mode as shown below (shade blue is for truth fiducial events with all 4 photons in EMC):

while for the dominant decays of 4 photons in the final state, the least energetic photon (E4) is unfortunately mostly below the L1 M and L3 EMC cluster threshold of ~120MeV so that we have to work with 3 photons or less in general. 

L1 Trigger 

The existing L1 trigger primarily relies on the 3M&M* line and only has an efficiency of 27.7% for the fiducial events. A detailed look at the MC signal event GLT counts gives some clues for possible improvements: 

While M* seems to cost a fair a bit efficiency, 1G appears to be rather good for efficiency and anything can be down with M=2 can be quite beneficial. Testing some lines on the MC signal sample and L1 pass through of  77968 (already with the new U(3s) trigger config, at L=7E33 and L1 rate of 2.5Khz). For testing the L1 rate effect under the condition of 77968, the study is based on the 2M triggers as a base to derive rate wrt the 2M rate (3.7Khz for 77968):

Note: the exclusive rates for T0-T5 refers to the configuration of adding just this single line on the original L1 config. For the last 2 rows,  the combine OR is treated as one `single line' to test their combined effect so that the exclusive rate is the overall addition of L1 rate from these two lines together. The L1 rate here is strictly for for condition of 77968. More recent runs at L=>9E33 is running in L1 rate ~1.6 higher of 3.5-4.5 Khz already. The corresponding additional rates scales similarly. At 4.5Khz. one is losing a few percent deadtime. A detaliled look at the T3|T4 L1 pattern compared to previous L1 config and looking at the exclusive rates of each line in the new L1 config:          

Ideally, one would like to run T2, but we don't know if L3 can utilize the additional events and it may be push the deadtime at the edge. The backup is T3|T4. Although T5 is slight more efficient, the no BLT 0B requirement is a veto and less preferable compared to 1G. The combined lower efficiency and branching ratio for 2pi0 mode may only carry ~1/5 of the overall statistics while the pi+pi- mode carrying the bulk. A 10% efficiency in 2pi0 mode is only worth getting if it is not costing more than 2% deadtime. This is a sensitive decision to be made.

L3 Trigger 

For the all neutral events, there are unfortunately not a lot of handles and the filter previously designed for saving low W 2-photon events can be extended to do the same job here, with the following filter cuts:

• L3 reconstructed track = 0    (note: we don't want off IP track either which can generate extra EMC clusters)
• L3 EMC clusters >= 3
• Highest energy EMC cluster E*<3.5 GeV   (reject crummy Bhabha/Gamgam background)
• Require the event mass from all EMC clusters to be > mass cut

Some additional kinematic information for the signal MC and the L1 passthrough data can be seen below:

Note that only events triggered L1 can have L3 EMC clusters produced to go into these plots. The non-flat cosq* in the signal MC is due to the more favorable lab energy to pass EMT M in the forward region. The total CM energy is heavily correlated with the event mass and the separation of signal and background with Pt balance is also marginal (modeling of Pt balance is also rather complex). So one might not be able to do much better than the simple scheme of cutting on event mass. Proceeding with just the event mass cut, the resulting signal efficiency and L3 rate as a function of the mass cut:     

A compromise choice may be a event ass cut of 400 MeV for an L3 efficiency of ~64% and L3 rate increase of 17%. However, this is still only for the old L1 trigger configuration. It is unclear with the new EMT lines how the efficiency and rate would scale. The new L3 EMC neutral filter rate has very little overlap with existing filters so that it is nearly all exclusive. 

Appendix: L3 EMC Threshold Check and Energy Scale

For various L3 neutral selections, there is a concern whether the MC signal efficiency is reliable due to the known large difference in L3/Reco EMC energy scale. L3 uses the EMC frontend calibration for energy scale and that is not even in the EMC simulation. This is expected to cause an average level shift. L3 EMC clustering is taking crystals at a much higher threshold than reco so that there is another scale effect depending the cluster crystal composition varing cluster to cluster. In particular this is expected to be different between charge particle and neutrals due to the different cluster shapes. 

This study used run 64100 (On peak Apr/22/2006) comparing to rather old release 12 MC U(4s) hadronic mix back from 2002 (I don't have ntuples with L3 info beyond R12 unfortunately). However, we should not have changed anything for L3 EMC since 2002. A hadronic selection is applied to both data and MC. Reco EMC candidates are selected requiring an isolation of no other EMC candidates within 0.2rad. This large removes possible ambiguity of matching L3 cluster to Recon and track matching. Clusters are further divided into "charged" or "neutral" depending if there is a matching reco track (this is of course not pure, just an enriched selection). Closest L3 cluster is then searched for around each EMC candidate, requiring the match df<0.15rad; dq<0.1rad. The L3 EMC threshold turn on are shown below:                

The effective threshold is ~130MeV for either charged or neutral clusters, with the charged turn on somewhat sluggish. The agreement between data and MC is reasonable (would more recent MC look different ?). For the matched clusters, one can then look at the energy ratio of L3/Reco:

There is a significant spread from cluster to cluster, worse at low energy, expected from the high crystal threshold in L3. Further quantifying the mean energy ratio on a refined sacle: 

The general features of the variations seem to be modeled reasonably well by the MC, e.g. bending low pT tracks going sideways with a longer trail in the EMC is seen to dump up the energy at low E. The variation as function of energy is significant and the largest scale shift at low E is as expected. The data vs MC discrepancies are at 1-2% level which may still matter systematic wise for some of the physics signals sitting very much on edge. 

Su Dong
 
Last modified:   Tue Jan 22 10:05:13 PST 2008