1 of 13

LDMX Software-dev meeting:�EoT Estimates

�24/01/2024

2 of 13

Background

  • LDCS Production of enhanced kaon sample
    • Needed to determine how many events to generate
    • How to best calculate EoT from a biased sample
  • Started with the simple case of an Ecal PN sample
    • Went down a rabbit hole

3 of 13

  • The number of selected events (M) in a sample should be binomially distributed with parameters: Number of tried events (N) and probability (p)
  • To make an EoT estimate from a biased sample with N events, we need to know how the probability in the biased sample differs from one in an inclusive sample
  • Many (most?) LDMX EoT estimates:
    • p(biased) = B p(inclusive)
    • With B being the biasing factor?
    • We can test it!
    • Alternative: p(biased) = W * p(inclusive),
      • With W ratio of average event weight of the two samples
      • For inclusive, the average event weight is just 1 :)

4 of 13

Binomial basics

  • Valid for distribution corresponding to N binary yes/no questions
  • Sample with M selected events and N generated, the probability estimate is just M/N
  • So we want C = p(biased)/p(inclusive)
  • Ratio of two probability parameters isn’t usually well behaved
    • Except if the distribution is binomial :)
    • 95% Confidence interval for this can be reliably calculated
    • CI[(ln(C)] = 1.96 sqrt{1/Ni - 1/Mi + 1/Nb - 1/Mb}
  • For a given C, we can extrapolate the 95% CI to a large enough sample size

5 of 13

6 of 13

EcalPN

  • At 1e8, both hypotheses are OK
    • But the flat bias looks like it is in trouble
  • So we extrapolate from the ratio at 1e8 to 1e9
    • Note: loglog scale

7 of 13

8 of 13

EcalPN

  • At 1e8, both hypotheses are OK
    • But the flat bias looks like it is in trouble
  • So we extrapolate from the ratio at 1e8 to 1e9
    • Note: loglog scale
  • At 1e9 events, flat bias is ruled out
    • Overestimates the biasing by 5% or so
    • We may need to generate larger samples
  • Weight estimate is still spot on (also holds for larger samples!)

9 of 13

Kaon samples

  • We can always estimate the probability ratio correctly given two samples
    • But producing a large inclusive sample kills the point of biasing :)
  • For more complicated biasing procedures, we don’t have a good guess for what the biasing factor should be
    • Ideally it factorises with the PN bias but… it might not
  • The only handle we have is the event weight
    • Two sources:
      • PN bias and pair production down-bias:
      • Resampling photonuclear interactions
    • Weights are multiplied during an event

10 of 13

11 of 13

Kaon samples

  • Weights are overestimating the bias :(
    • Issue with how we assign them?
  • For the LDCS production of the kaon sample, we can start a production with biasing factor being ~34
    • On top of the regular PN biasing
    • Has some uncertainty

12 of 13

Appendix

13 of 13