===== Negative Weight Mitigation (also for Higgs Signal/Background?) =====

**Code**: [[https://github.com/a-maier/cres|cres]]

== Specific Topics ==
  * Apply cell resampling ([[https://arxiv.org/abs/2109.07851|paper 1]] [[https://arxiv.org/abs/2303.15246|paper 2]] [[https://github.com/a-maier/cres|source]] [[https://cres.hepforge.org/|binaries]]) to Higgs signal/background (closure tests, measure of simplification e.g. #event reduction)
  * Graph of fraction of negative weights for Higgs processes/background (slide 5/9 [[https://phystev.cnrs.fr/wiki/_media/2023:cell_resampling.pdf|Jeppe talk]])

== Processes ==
  * di-photon NLO sample (m_gamgam centred around Higgs mass window)

== General Topics ==
  * Discussion on good metrics (what to compare before considering events close)
  * Measure the impact on reals vs virtuals, which events are most likely to be altered and by how much?
  * Try it on proper experimental samples (what metric is good, what issues arise?)
  * Applying this to NNLO (e.g. HighTea samples)
  * Try e.g. jpsi->leptons something very narrow (technical issues regarding IR sensitivity, modifying distributions...)
  * Plots of mean/median/width of the cell resampling bins, studying these distributions and their potential impact
  * Can narrow weight distribution for more efficient event unweighting

== Talk Discussion (winner: most questions per talk) ==
  * Q: why is w+5 improving more than z+3? 
    - initial event sample size? (no, both have 1e9)
    - dipole cut vs improvement? (check dipole cut used)
  * Q: calculating density of events to "pre-check" that this algorithm may work?
  * Q: metric needs to be compatible with how ps generator is populating ps, proof that this is not biasing anything?
  * Q: should there be a bias towards positive values (since you iteratively add nearest event until wi>=0)?
    - ps generator dependent because this alters which events are clustered, can this have an impact on the physics?
  * Q: plotting statistical uncertainties on the original sample 
    - more easily allows verifying that differences are within statistical uncertainty
  * Q: on which type of event does this have the biggest impact?
    - imagine reals are more impacted than virtual 
    - we were completely agnostic
  * Q: can we really prove that this does not alter distributions? (prove that you preserve distributional structure of the observables you compute)
    - we do not alter any of the event kinematics
    - there is no cross talk of events separated by more than the maximum allowed distance
    - IRC safe measure important, but this is not a sufficient condition, can you reproduce infrared sensitive observables (e.g. Sudakov shoulder, 0-bin of ptZ), can your smearing reproduce this feature in the limit that the smearing goes to 0?
  * Q: mean/median/width of the cell resampling bins
    - we have plots that we can examine