alexander-held / coffeahzzanalysis

CoffeaHZZAnalysis

(Note: CoffeaHZZAnalysisWithServiceX.ipynb as well as the Coffea backend uption for CabinetryHZZAnalysis.ipynb do not work on Binder. See the comments in those notebooks for how you can run them locally.)

Based on https://github.com/atlas-outreach-data-tools/notebooks-collection-opendata/blob/master/13-TeV-examples/uproot_python/HZZAnalysis.ipynb.

Also makes use of code from https://github.com/alexander-held/PyHEP-2021-cabinetry and https://github.com/gordonwatts/pyhep-2021-SX-OpenDataDemo.

Runtime Comparisons:

• Original (data from https): 16 minutes
• Original (data stored locally): 33 seconds
• Coffea Processor (data from https): 46 seconds (3 minutes on Binder, 50 seconds on coffea-casa)
• Coffea Processor (data stored locally): 6 seconds (18 seconds on Binder)
• Coffea Processor (data from ServiceX, uncached): 33 seconds
• Coffea Processor (data from ServiceX, cached): 2 seconds
• Cabinetry with Uproot Backend (data stored locally): 4 seconds (5 seconds on Binder, 4 seconds on coffea-casa)
• Cabinetry with Coffea Backend (data stored locally): 10 seconds

---

This section is the runtimes with the input data files multiplied by 10.

• Original (data from https): Over 80 minutes (estimate, I did not finish running this)
• Original (data stored locally): 6 minutes
• Coffea Processor (data from https): 5 minutes
• Coffea Processor (data stored locally): 29 seconds
• Coffea Processor (data from ServiceX, uncached): 3 minutes
• Coffea Processor (data from ServiceX, cached): 3 seconds
• Cabinetry with Uproot Backend (data stored locally): 35 seconds
• Cabinetry with Coffea Backend (data stored locally): 1 minute

---

• Coffea Processor (with list comprehension): ~20 minutes
• Coffea Processor (with cuts vectorized): ~11 minutes
• Coffea Processor (with all vectorized functions except calc_weight): ~2 minutes
• Coffea Processor (with all vectorized functions): ~3 seconds
• Coffea Processor (using hist package instead of coffea.hist): ~4 seconds

---

Cut order and method does not seem to significantly affect runtime.

• Cut on Charge, then Type: 5.0 seconds
• Cut on Type, then Charge: 5.3 seconds
• Cut on Charge and Type Simultaneously with Boolean Operators: 4.8 seconds
• Cut on Charge and Type Simultaneously with a Single Function: 5.2 seconds
• Cut on Charge, then Type Directly in Processor without a Function: 4.6 seconds

Discovery Significance Comparisons:

Bin Width is 5 GeV except where stated otherwise.

Bounds are 80-250 GeV except where stated otherwise.

• Original Bins
  • Observed Significance: 2.345
  • Expected Significance: 1.301
• Bin Width: 10 GeV
  • Observed Significance: 2.090
  • Expected Significance: 1.253
• Bin Width: 2 GeV
  • Observed Significance: 1.887
  • Expected Significance: 1.509
• Bin Width: 17 GeV
  • Observed Significance: 1.303
  • Expected Significance: 0.996
• Bin Width: 1 GeV
  • Observed Significance: 2.674
  • Expected Significance: 1.688
• Bin Width: 34 GeV
  • Observed Significance: 0.013
  • Expected Significance: 0.481
• Bin Width: 1 GeV between 110 GeV and 140 GeV and 10 GeV elsewhere
  • Observed Significance: 2.244
  • Expected Significance: 1.624
• Bin Width: 10 GeV between 110 GeV and 140 GeV and 1 GeV elsewhere
  • Observed Significance: 2.676
  • Expected Significance: 1.344
• Bin Width: 30 GeV between 110 GeV and 140 GeV and 1 GeV elsewhere
  • Observed Significance: 1.429
  • Expected Significance: 0.743
• Bounds: 110-140 GeV
  • Observed Significance: 2.743
  • Expected Significance: 1.044
• Bounds: 95-195 GeV
  • Observed Significance: 2.223
  • Expected Significance: 1.253
• Bounds: 30-300 GeV
  • Observed Significance: 2.112
  • Expected Significance: 1.282
• Bin Width: 1 GeV, Bounds: 110-140 GeV
  • Observed Significance: 2.044
  • Expected Significance: 1.335
• Bin Width: 10 GeV, Bounds: 110-140 GeV
  • Observed Significance: 2.476
  • Expected Significance: 1.020