TensorWriter: tame statistically-empty-bin systematics (clip / zero-low-neff)#146
Merged
Merged
Conversation
Near-empty bins whose nominal is a mixed-sign NLO weight-cancellation residual have n_eff = sumw**2/sumw2 << 1. There a log_normal systematic ratio logk = log(syst/nom) is floating-point noise reaching |logk| ~ 40-55, so the varied yield nom*exp(logk) blows up to ~1e15-1e20. This is what debug_inputdata flags and what makes rabbit_plot_inputdata explode. Two opt-in TensorWriter knobs (both off by default): - clip_syst_variations=x: clip each bin's up/down factor to [1/x, x]. Also fixes a real bug in get_logk: the clip was applied to _logk but the function returns _logk_view, so it never took effect for the returned value; now clip the returned array. - zero_syst_low_neff=thr (+ zero_syst_low_neff_procs): at write() time, zero every systematic's logk in any (bin, process) with n_eff < thr. Keeps the unbiased nominal, removes only the meaningless systematic lever, and cleans the card so debug/plot tools come out clean. Handles dense and sparse storage; restrictable to a list of processes (default: all). Co-Authored-By: Claude Opus 4.8 <[email protected]>
lucalavezzo
added a commit
to lucalavezzo/WRemnants
that referenced
this pull request
Jul 21, 2026
Add two opt-in CLI flags (both off by default) that forward to the rabbit TensorWriter knobs of the same name: - --clipSystVariations X: clip each bin's log_normal up/down factor to [1/X, X]. - --zeroSystLowNeff X (+ --zeroSystLowNeffProcs): zero every systematic's logk in any (bin, process) with n_eff = sumw^2/sumw2 < X. Tames the ~1e15-1e20 blow-ups from mixed-sign NLO weight-cancellation-residual bins in thin backgrounds (Ztautau/PhotonInduced/Other) that break debug/plot tools. Requires the corresponding rabbit TensorWriter change (WMass/rabbit#146). Co-Authored-By: Claude Opus 4.8 <[email protected]>
lucalavezzo
added a commit
to lucalavezzo/WRemnants
that referenced
this pull request
Jul 21, 2026
Add two opt-in CLI flags (both off by default) that forward to the rabbit TensorWriter knobs of the same name: - --clipSystVariations X: clip each bin's log_normal up/down factor to [1/X, X]. - --zeroSystLowNeff X (+ --zeroSystLowNeffProcs): zero every systematic's logk in any (bin, process) with n_eff = sumw^2/sumw2 < X. Tames the ~1e15-1e20 blow-ups from mixed-sign NLO weight-cancellation-residual bins in thin backgrounds (Ztautau/PhotonInduced/Other) that break debug/plot tools. The new knobs are forwarded to TensorWriter only when actually set, so this script remains compatible with rabbit versions predating them. Using the flags requires a rabbit with these TensorWriter arguments (WMass/rabbit#146). Co-Authored-By: Claude Opus 4.8 <[email protected]>
lucalavezzo
added a commit
to lucalavezzo/WRemnants
that referenced
this pull request
Jul 21, 2026
WMass/rabbit#146 is merged (5a01cb4), adding the TensorWriter clip_syst_variations / zero_syst_low_neff arguments that the new setupRabbit flags forward to. Bumps the pin 948a94a -> 5a01cb4 (only WMass#146 in between) so --clipSystVariations / --zeroSystLowNeff are usable. Co-Authored-By: Claude Opus 4.8 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
In dense log-normal datacards, thin backgrounds have bins whose nominal is a mixed-sign NLO weight-cancellation residual, so the effective sample size
n_eff = sumw**2/sumw2is<< 1(down to ~1e-6). There the systematic ratiologk = log(syst/nom)is pure floating-point noise and reaches|logk| ~ 40–55, so the varied yieldnom*exp(logk)blows up to ~1e15–1e20. This is whatdebug_inputdataflags and what makesrabbit_plot_inputdataexplode.Changes (both opt-in, off by default)
clip_syst_variations=x— clip each bin's up/down factor to[1/x, x]. Also fixes a real bug inget_logk: the clip was applied to the local_logkbut the function returns_logk_view, so the clip never took effect for the returned value (there was even aFIXMEnoting this). Now the returned array is clipped.zero_syst_low_neff=thr(+zero_syst_low_neff_procs) — atwrite()time, zero every systematic'slogkin any(bin, process)withn_eff < thr. Keeps the unbiased nominal, removes only the meaningless systematic lever, and — unlike clipping — cleans the card itself sodebug_inputdata/rabbit_plot_inputdatacome out clean. Handles dense and sparse storage; restrictable to a list of processes (default: all).Validation
On a real 4D Z datacard (49920 bins × 4 procs × 3746 systs), a build with
zero_syst_low_neff=1produces a card byte-identical to an independent post-hoc reference across every dataset (hnorm/hsumw/hsumw2/hdata_obs/hlogk). Only backgrounds are flagged (Ztautau/PhotonInduced/Other); the signal has no low-n_effbins, so the all-process default is safe.Note: the WRemnants
setupRabbit.pyCLI that exposes these (--clipSystVariations,--zeroSystLowNeff) is a separate WRemnants PR and depends on this one.🤖 Generated with Claude Code