Note: The pre-filters and MC samples in this analysis are consistent with the analysis Eric was performing; however, they are significantly different from the data sets and pre-filters Bill applied in designing the Pass6 and Pass7 analyses. For example, these data sets have no cut on FswGamState == 0 and the IrrFilter is different (and almost certainly having unintended consequences).
Introduction
I've gone back to Eric's old page with the hope of picking up where he left off. Of course, this leads to the always enjoyable task of replicating the work that he has already done. The working directory for the previous analysis is:
...
Name | Purpose | Gamma Efficiency | Bkg Efficiency | Cut | ||
---|---|---|---|---|---|---|
IrrFilter | Remove irreducible bkg events | 0% | 13% |
| ||
TkrCalCore | Take only "good" events | 6.2% | 2.2% |
| ||
AcdEngFilter | Only events with moderate signal in the ACD | 0% | 47% |
| ||
Prefilter | Only select relevant events | 6.2% | 1.2% |
|
| nSig | nBkg |
---|---|---|
Before Prefilters | ~9x10^5 | ~4x10^6 |
After Prefilters | 55015 | 48614 |
One very important thing to note here is how confusing the convention of "Cuts" and "Filters" is. There needs to be some standard convention for either keeping or rejecting events that return TRUE for a selection (I believe that Bill uses "Cuts" and "Vetos").
Pass 6 Performance
First defining some ACD variables:
...
These numbers seem to agree well with Eric's results. However, it naively seems like there are still some photons sneaking into the bkg sample. This is because selecting !(IrrFilter) will return True if McCharge == 0 (leaving about ~50 gamma events out of ~550). Putting a cut on McCharge, we now get a background rejection of 1.03%.
Pass 7 Performance
Moving on to Pass7, the ACD analysis has be modified a fair amount. I rebuilt GR v17r35p10 and generated roughly the same number of signal and background events (again using Eric allGamma.txt and background.txt scripts). Since Pass7 was just a re-design of the worksheet, the prefilters should have the same efficiency (Irreducible 13%, AcdEngCut 47%) with a slight change in CalTkrCore since it depends on CTBCore (2.2% for bkg and 6.4% for sig) leaving a total of 1.3% of the bkg and 6.4% of the gamma-ray events.
...
Name | Purpose | Gamma Efficiency | Bkg Efficiency | Cut | ||
---|---|---|---|---|---|---|
BasicTileCut | Reject events with track pointing at struck tile. | 94.2% | 6.18% |
| ||
RibbonCut | Reject events with track pointing at struck ribbon. | 93% | 5.85% |
| ||
TotalTileEnergyCut | Reject events with excess ACD total energy. | 88.1% | 2.08% |
| ||
CornerCut | Reject events in the corner gap of the ACD. | 86.9% | 2.06% |
| ||
TileEdgeCut | Reject events at tile edges with decreased signal. | 86.9% | 2.06% |
| ||
VetoTileCut | Reject events with ... | 86.6% | 1.71% |
|
Pass 8 Performance
Generating merit files off of the GlastRelease-HEAD1.1365 version (which came shortly after GlastRelease-v19r3p4), I started by looking at the effect of the Pass7 ACD cuts. Here the prefilters could have changed (since track finding etc. has changed). Using the same cut definitions, I find fairly similar results with slightly worse background rejection.
...
Name | Purpose | Gamma Efficiency | Bkg Efficiency | Cut | ||
---|---|---|---|---|---|---|
VetoHit1Cut | Reject events with best track pointing a struck tile/ribbon | 94.5% | 14.2% |
| ||
TotalTileEnergyCut | Reject events with excess ACD total energy | 86.5% | 2.44% |
| ||
VetoGap1Cut | Reject events in gaps of the ACD | 86.2% | 2.42% |
| ||
VetoHitCut | Reject events if other track points at hit tile | 86.1% | 2.35% |
|
Conclusions
Well, it doesn't seem like we are where we want to be (not a surprise). It looks like we would like to roughly double the background rejection (1/2 the efficiency) for the Pass 8 analysis with the AcdReconV2 variables. However, it appears that the simple cuts with AcdReconV2 are roughly comparable to those with the original AcdRecon, meaning that the degradation in background rejection may be in part due to upstream reconstruction changes (which in any case are not yet complete).
Looking Forward
- Can we easily modify the AcdV2 cuts to increase background rejection with the current reconstruction? Basically, can we optimize these cuts?
- Right now AcdV2 is using the G4 propagated covariance matrix, which really isn't right for cosmic-rays (since it uses electron hypothesis). Would changing this make an improvement?
- How do the CalOnly events look? Can we develop some rough background rejection for them?
- Obviously this will improve as the upstream reconstruction improves. Can we provide any guidance for that?
- CTs are always the after-burner to get the boosted rejection power that we need.