Event Classification & Background Rejection Bill Atwood & Friends
Recap Since DC1 DC1 Closeout February 2004 Begin Recon-Rewrite, May 2004 First up: TkrRecon Participants: Tracy Usher, Leon Rochester, Johann Cohen-Tanugi, WBA Next up: CalRecon Participants: Tracy Usher, Philippe Breul, Pol d'Avezac, Fred Piron, Eric Nuss, David Chamont, WBA Next up: AcdRecon Participants: Heather Kelly, Eric Charles, WBA Development of Alternative High Level Analysis Toby Burnett & Riccardo Rando et al Begin DC2 Analysis, May 2005 First Energy Analysis
Energy Methods Method % Computed % Best Est. Parametric 100 63.6 Profile 49.9 24.7 Last Layer 23.4 5.5 Tracker 16.5 6.3 4 Methods* 3 Only cover a part of Glast Phase Space *Pol d'Avezac has combined Tracker & Last Layer Parametric Tracker Last Layer Profile Only Parametric Available: 37.7% This tends to be the Local Land Fill (City Dump!) Unfortunately there are too many events here to simply throw out. Compare each Method against a "standard" defining: Energy Resolution Model Select Method giving the Highest CT Prob. Data Set: All Gamma (GR-HEAD1.615) sModel = .02+.6/(McLogEnergy)2.5 + .005*(McLogEnergy-2.)2
E4: Unbalanced CTs - 2s on s=.05+.72/(log(E))3 Data Set: AG1617-mod
How to Tell What You Have: In an analysis you may wish to check which energy method your "signal" is relying on. In the IM environment this is captured by a set of CATAGORICAL Variables and life is easy... however this was not captured in the translation to GLEAM. Here's how to tell: Compare: CTBParamProb CTBProfileProb CTBLastLayerProb CTBTrackerProb CTBBestEnergyProb The one which matches is the method used for the event in question. CTBBestEnergy, CTBLogEnergy, and CTBBestEnergyProb summarize the final results
PSF Analysis IM "glue" to put events bact together and produce Split according to Thick and Thin Layers: Tkr1FirstLayer 2-5 (Thick) or 6-17 (Thin) IM "glue" to put events bact together and produce CTBBestXDir,YDir,ZDir and CTBCORE (the "good" PSF probability variable) CTs: Of Events with Valid VTX choose to use 1Tkr or VTX Solution: CTBVTX Split according to VTX Solution (VtxAngle > 0 && CTBVTX > .5 CTs: PSF Image Sharpening Produces CTBCORE Variable
PSF Analysis Results What CTBCOR Does On Axis can vary PSF by ~ 30% (at the expense of Aeff) The 95/68 Ratio improves significantly cos(q) < -.95 cos(q) < -.2
How to Tell What You Have: To tell which analysis branch you have - THICK or THIN? THICK == Tkr1FirstLayer < 5 THIN == !THICK Note: Tracks can not originate in Layers 0 and 1 - VTX or 1TKR ? VTX == VtxAngle > 0. && CTBVTX > .5 1TKR == !VTX Note: We could have used VtxStatus bits instead of VtxAngle CTBBestXDir,YDir,ZDir, CTBCORE summarize the results THIN Layers THICK Layers Cuts: CTBGAM > .5 && CTBBestZDir < -.95 && CTBBestEnergyProb > .35
Back Ground Rejection Evolution - Use of only Classification Trees not adequate (Typ. Rejection Power 10 -40) - Improve Input to CTs via "typical" HEP Style cuts (Implemented as PreFilters) - Divide sample according type 1) Topology (VTX – 1Tkr) 2) Obs. Cal. Energy (use 6 ~log(E) bins) 3) Location (Lowest energy only) (Gives 14 Separate catagories) What's left is what CTs are poor at finding or Irreducible (start here) Results in a g Bkg. Rejection Prob. CTBGAM Class 3
First: Irreducible Back Grounds Correlated Uncorrelated e+ Blanket Conversion Proton - Blanket Interaction > 60% Remaining background are from gs produced locally They are Irreducible
Second: Weakness in CT Analysis Subsequent Search for "Hot Spots" in GLAST Phase space revealed - Correlated Events in the Top Elimination of Events which project back to Ribbons Cut Summary Bkg: 964 - 4.7% All Gam: 3046 - 2.9% Top Ribbons More Ribbons Holes at the Top of Side Ribbons
Correlations among Variables Heavy Ion Filter CTBBestEnergy > 1000 & ((CalTransRms - 1.5*Tkr1ToTTrAve) < 5) & CTBGAM > .5 What's Eliminated Cut Summary Bkg: 43 - .22% All Gam: 14 - .014% Scrambled Tracks Filter (Tkr1FirstLayer - Tkr2FirstLayer) <= 0 & Tkr2FirstLayer > 2 & Tkr2TkrHDoca > 10 & (CTBGAM+.16*CTBBestLogEnergy) < 1.32 Cut Summary Bkg: 1284 - 6.6% All Gam: 435 - .42%
These Post-Processing Cuts can be found at: https://confluence.slac.stanford.edu/display/DC2/Trigger+and+Filter+settings+and+Event+classes This leaves an Aeff vs Energy with features (not good...). Smooth Response by "boosting" low energy CTBGAM : CTBGAM = min(1., CTBGAM*min(2.0, (3.5/min(CTBBestLogEnergy, 3.5))^2)) Base Class 3 – No Post-Processing Filters CTBGAM > .5 Base Class 3 – Post-Processing Filters & CTBGAM Boost at Low Energy CTBGAM > .5
Class 3 Class 3 Post Process Cuts + Careful: Order of Plots different Top to Bottom
Demonstrations with DC2 Data from Galactic Anti Center Region (thanks to Julie) Ooppsss - Almost forgot – The Limb g issue.... Rocked Backgrounds What's cut out: CTBGAM > .5 2216 evts Evts. in Common 277 evts Evts. Missed by BestZDir > -.3 127 evts Julie's Cut Bill's Original Cut Cut Used + ZDir < -.3 Evts. Missed by ZenithTheta > 100 32 evts. Residual Bkg. Evts. 436 Total Evts.
On with the show... Galactic Anti-Center: Class A – No Post Proc. Cuts
Class A Events + Post-Proc. Cuts
What was killed by Post-Proc. Cuts The Fog of Back Ground...
Image Resolution The Geminga Pulsar Low Res. Good Res. Best Res.
Estimating Residual Backgrounds The ~20k Background events remaining after PreFilters As with previous experiments – look for orbit associated modulation in signal. Use this as a measure of remaining contamination. Seems to be ~ 10% Need to "fold" all the orbits on top of each other – Hard due to multiple components and SAA (holes). Steve's suggestion: Profile residuals events vs the Trigger Rate
Summary DC2 Background Rejection gets us close... however There is no factor of 2 left – the Irreducible Component Dominates Background Rejections is a "work-in-progress" We just ran out of time. We have made 3 complete passes so far and a Pass 4 is in the works. Most Analysis are not that sensitive to Background Contamination Thankfully residual backgrounds are essentially spread ~ uniformally across the sky. As such contamination in a few sq. degrees is small (part in 10-4) Work will continue and perhaps a "reprocessed" DC2 Dataset will appear prior to the DC2 Closeout (similar to real life) The instrument "hand-off" is approaching and we need to meet the Science Requirements by then.