2019:groups:tools:correlations
Differences
This shows you the differences between two versions of the page.
| Both sides previous revision Previous revision Next revision | Previous revision Last revision Both sides next revision | ||
|
2019:groups:tools:correlations [2019/06/25 13:41] sabine.kraml |
2019:groups:tools:correlations [2019/06/27 16:30] wolfgang.waltenberger [Quantifying overlaps between analysis search regions using ADLs] |
||
|---|---|---|---|
| Line 1: | Line 1: | ||
| ====== Study: correlations between signal regions ====== | ====== Study: correlations between signal regions ====== | ||
| - | //Members: Sophie, Wolfgang, Humberto, Benj, Andy, Sabine // | + | //Members: **Sophie**, Wolfgang, Humberto, Benj, Andy, Sabine, Sezen // |
| Problem statement: to identify pairs of signal regions of the analyses in the PAD database that can safely be treated as approximately uncorrelated. | Problem statement: to identify pairs of signal regions of the analyses in the PAD database that can safely be treated as approximately uncorrelated. | ||
| Line 10: | Line 10: | ||
| For the stats, we'll use bootstrapping rather than directly study which events fall into common signal regions. Procedure: in the analysis framework the set of populated SRs is reported for // each event//, e.g. as a line of N_SR 1s and 0s in an output file. We then process this: for each event (=line) we sample N_history = O(100-1000) Poisson(lambda=1) weights, and enter these into a set of N_history histograms (each histogram has N_SR bins). We then build a correlation matrix between the bins using the standard sample covariance cov_ij = <sumw_i sumw_j> - <sumw_i> <sumw_j> and corr_ij = cov_ij / sqrt( cov_ii cov_jj ). Finally, convert this to a binary "sufficiently independent SRs" matrix with a corr threshold: indep_ij = (|corr_ij| < thres). The 1s in each row (or column) of this binary matrix define a set of statistically independent SRs, which can be trivially combined in a likelihood. | For the stats, we'll use bootstrapping rather than directly study which events fall into common signal regions. Procedure: in the analysis framework the set of populated SRs is reported for // each event//, e.g. as a line of N_SR 1s and 0s in an output file. We then process this: for each event (=line) we sample N_history = O(100-1000) Poisson(lambda=1) weights, and enter these into a set of N_history histograms (each histogram has N_SR bins). We then build a correlation matrix between the bins using the standard sample covariance cov_ij = <sumw_i sumw_j> - <sumw_i> <sumw_j> and corr_ij = cov_ij / sqrt( cov_ii cov_jj ). Finally, convert this to a binary "sufficiently independent SRs" matrix with a corr threshold: indep_ij = (|corr_ij| < thres). The 1s in each row (or column) of this binary matrix define a set of statistically independent SRs, which can be trivially combined in a likelihood. | ||
| + | |||
| + | |||
| + | === MA5 package to be used for correlation studies === | ||
| + | |||
| + | - Any version of the code from v1.8.20 onwards, to be downloaded e.g. from {{ 2019:groups:tools:ma5_v1.8.20.tgz | here }}. | ||
| + | |||
| + | - This version contains two new dedicated functions (to be added in tools/PAD/Build/Main/main.cpp): | ||
| + | * void manager.DumpSR(std::ostream&): writes a series of 0 and 1. One entry for each considered signal region. One line per event. To be included in the event loop. | ||
| + | * void manager.HeadSR(std::ostream&): writes the header of the file: one comment line (starting with a hash) with the list of analysis-SR. To be included before entering the event loop/ | ||
| + | |||
| + | |||
| + | **Generic overview of [[RecastCodeComparison|what is implemented in which recast framework]]** | ||
| + | |||
| Line 21: | Line 34: | ||
| | CMS-SUS-16-033 | multijet + MET, 36 fb-1 | T1, T1bbbb, T1tttt(off), T2, T2bb, T2tt(off) | UL | | | CMS-SUS-16-033 | multijet + MET, 36 fb-1 | T1, T1bbbb, T1tttt(off), T2, T2bb, T2tt(off) | UL | | ||
| | CMS-SUS-16-039 | multilepton EWK | TChiWZ(off), TChiWH, TChipmSlep... | UL | | | CMS-SUS-16-039 | multilepton EWK | TChiWZ(off), TChiWH, TChipmSlep... | UL | | ||
| - | | CMS-SUS-16-052 | 1L stop, soft | T2bbWWoff, T6bbWWoff | UL, agg-EM | SModelS has only PAS version of this | | + | | CMS-PAS-SUS-16-052 | 1L stop, soft | T2bbWWoff, T6bbWWoff | UL, agg-EM | | |
| | CMS-SUS-17-001 | 2L stop | T2tt(off), T6bbWW | UL | | | CMS-SUS-17-001 | 2L stop | T2tt(off), T6bbWW | UL | | ||
| + | |||
| + | See also http://madanalysis.irmp.ucl.ac.be/wiki/PublicAnalysisDatabase | ||
| Line 42: | Line 57: | ||
| | CMS-SUS-16-039 | multilepton EWK | TChiWZ(off), TChiWH, TChipmSlep... | UL | | | CMS-SUS-16-039 | multilepton EWK | TChiWZ(off), TChiWH, TChipmSlep... | UL | | ||
| + | ==== Quantifying overlaps between analysis search regions using ADLs ==== | ||
| + | |||
| + | Members: Sezen, Wolfgang (, Harrison) | ||
| - | === CheckMate - SModelS correspondence (13 TeV): === | + | Find and visualize overlaps in a model-independent way, without generating events. Directly sample the event selection. Useful for analysis design phase, or quick comparisons within experiments (e.g. Run2 CMS SUSY pMSSM combination) |
| - | //... to be done ...// | + | * Start from the analysis description, which lists objects and event selections. |
| + | * Construct a feature space from all mathematically orthogonal "basic" variables (e.g. MET, jet1.pt, jet2.pt, electron1.eta, ...). | ||
| + | * Randomly sample the feature space for each analysis based on cuts on the feature space components (jet1.pt > 100, MET > 299, etc.). | ||
| + | * Use the sampled points to compute values for "composite" variables such as HT(jets), dphi(jets), MT(lepton, MET), etc. | ||
| + | * Compare feature spaces between analyses, find and visualize overlaps and exclusions. | ||
| + | * As a very simple first step, we simply check if two analyses are disjoint in any of the basic variables. | ||
2019/groups/tools/correlations.txt · Last modified: 2019/06/27 17:10 by sezen.sekmen
Page Tools
Except where otherwise noted, content on this wiki is licensed under the following license: Public Domain
