Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

This paper argues that a compact optimal-transport embedding of jet substructure, reduced to a few principal components and added to standard observables, nearly doubles anomaly-detection significance at low signal fractions on collider ben

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 18:30 UTC pith:6XCQPO2H

load-bearing objection Useful OT feature idea, but the PCA basis is fit with signal labels—the weakly-supervised claim needs a background-only test. the 3 major comments →

arxiv 2512.04839 v3 pith:6XCQPO2H submitted 2025-12-04 hep-ph hep-ex

Optimal Transport Event Representation for Anomaly Detection

Tianji Cai , Aditya Bhargava , Benjamin Nachman This is my paper
classification hep-ph hep-ex
keywords optimal transportweakly supervised anomaly detectionlinearized Wasserstein distancejet substructureprincipal component analysisresonant new physicsevent representationLHC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the limiting factor in weakly supervised anomaly detection at low signal fractions is not classifier capacity but the event representation. It proposes encoding each jet as a linearized optimal-transport embedding—a geometric map of transverse-momentum flow—then extracting a handful of principal components to add to jet mass and n-subjettiness. At about 0.5% signal injection, this augmented set reaches a maximum significance improvement near 25, roughly twice the standard high-level observable set and far above full-phase-space deep learning or a large pretrained foundation model. The gains persist across two signal types and both classifier families tested.

Core claim

The discovery is that a linearized 2-Wasserstein embedding of jet substructure, compressed by PCA, is a powerful physics-based feature set for weak-supervision anomaly detection. Treating each jet as a distribution of transverse momentum over the rapidity–azimuth plane, the method embeds it into the tangent space of a fixed reference jet, yielding a 400-dimensional vector per event; the first few principal components, appended to standard observables, saturate performance. In the ultra-low signal regime (signal-to-background below 0.5%), this raises the maximum significance improvement to around 25 for the two-pronged benchmark, roughly 65% above standard high-level observables and more than

What carries the argument

The central object is the linearized 2-Wasserstein (LinW2) embedding. It measures, for each reference particle, the pT-weighted barycenter of where that particle is transported under an optimal transport plan from a fixed reference jet to the event jet; each event becomes a Euclidean vector in 400 dimensions. PCA compresses these embeddings, and the first 3–6 components carry almost all the discriminating power, acting as a geometric complement to jet mass and n-subjettiness.

Load-bearing premise

The paper evaluates an idealized anomaly detector that assumes perfect background interpolation; if real sideband-to-signal modeling has nontrivial error, the reported significance gains are upper bounds and may shrink.

What would settle it

Retrain the same classifiers with a learned background model, for example a density estimator fitted to sideband events, replacing the perfect-interpolation assumption; if the maximum significance improvement at 0.5% signal injection falls to the level of the standard high-level feature set, the claimed advantage would not survive realistic background uncertainty.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • At low signal fractions, one can roughly double anomaly-detection significance without heavy end-to-end models, making the method computationally cheap to deploy in searches.
  • The gains hold for both two-pronged and three-pronged jet substructures, suggesting the representation captures general morphology rather than one signal shape.
  • Because the OT representation is infrared and collinear safe by construction, its benefits are expected to be more transferable from simulation to detector data.
  • Only a handful of features are needed; expanding to 100 PCA modes degrades performance in the ultra-low signal regime, underscoring the value of feature selection.
  • In the high-signal regime, full phase-space methods still win, so the OT representation is complementary rather than a replacement for end-to-end learning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If these gains survive with realistic background interpolation, the same few-feature recipe could be added to existing dijet searches with minimal overhead and improve sensitivity to a broad class of resonances; this is an editorial extrapolation, not a claim in the paper.
  • The saturation of performance at 3–6 PCA modes hints that the OT embedding's discriminative manifold is very low-dimensional; identifying that manifold directly, without PCA, could yield further gains.
  • The IRC safety of the representation suggests it may be a more robust input than raw four-momenta for training foundation models, potentially improving transfer across pileup and detector conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces an intermediate event representation for weakly supervised resonant anomaly detection based on the linearized 2-Wasserstein optimal transport (LinW2) embedding of the two leading jets, followed by PCA dimensionality reduction. The OT-derived features are added to standard high-level observables (jet masses and n-subjettiness ratios) and used to train BDT (and MLP) classifiers under an idealized anomaly detector (IAD) that assumes perfect background interpolation. On the LHCO R&D1 dataset, the authors report that at 0.5% signal injection the OT-augmented feature set achieves a maximum significance improvement of about 25, roughly double the standard observables and well above full-phase-space and foundation-model baselines; qualitatively similar but smaller gains are reported on R&D2. The paper includes ablations, classifier robustness checks, and a public code repository.

Significance. If the central result holds, the paper makes a useful contribution: it shows that a compact, physics-motivated representation can outperform both engineered high-level variables and expensive end-to-end deep learning in the ultra-low-signal regime, with negligible computational overhead. The study is transparent in using the established LHCO benchmarks, includes multiple signal fractions, two classifier families, and an ablation isolating the OT contribution, and it ships reproducible code. The main caveat is that the PCA basis used to build the OT features is fit on a 50% signal-enriched sample, which is not a weakly supervised setting; a background-only PCA test is needed before the headline comparison can be trusted as a statement about unsupervised/weakly supervised representation learning. The IAD idealization is explicitly acknowledged, but its impact on the quantitative claims should be assessed with at least one realistic background-model comparison.

major comments (3)
  1. [Optimal Transport Representation and Features] The PCA basis is fit on '5k background and 5k signal events randomly selected from R&D1 or R&D2'. This is a 50% signal fraction, roughly 100 times the 0.5% injection used in the main results. Because the paper frames the method as weakly supervised, using event-level signal labels to orient the PCA modes is not legitimate in the intended deployment scenario and could artificially inflate the reported SI by encoding signal-specific variance into the features. The authors should (i) refit the PCA on background-only data (e.g., A1 or SR sidebands) and show whether the OT_k gains persist, and (ii) state explicitly whether the PCA training sample overlaps with A2, B1, or B2; if it does, the test results are affected by leakage.
  2. [Datasets and Weak Supervision Framework] The paper assumes 'perfect background interpolation' and evaluates an idealized anomaly detector (IAD) [15]. The reported SI values are therefore upper bounds for a real weak-supervision pipeline, since learned sideband interpolation (CATHODE, density-estimation, etc.) introduces additional error. This is clearly stated, but the abstract and conclusions present the factor-of-two gain without this qualification. A concrete test with at least one realistic background model (e.g., CATHODE or a density-estimation surrogate) at S/B=0.5–1% on R&D1 would establish whether the gains survive in a deployable setup.
  3. [Results] On R&D2 the comparison is incomplete: there is no full-phase-space or foundation-model baseline, and the statement that 'comparable trends are likely' is speculative. Since one of the paper's central claims is that OT features outperform end-to-end deep learning, the absence of these baselines on R&D2 weakens the claim that the gains 'persist across signal types.' The authors should either add the low-level benchmarks for R&D2 or explicitly restrict the comparison to standard high-level observables in the conclusions.
minor comments (4)
  1. [Results] The significance improvement (SI) is used throughout but never defined. Please provide the definition (e.g., SI = (ε_s/√ε_b) for a fixed working point or as a function of the true-positive rate) and explain how the 'maximum SI' is obtained from the scan over S/B and over OT feature counts.
  2. [Optimal Transport Representation and Features] Fig. 1 caption should state whether the variance-explained curve is computed on the signal-enriched PCA sample or on a separate sample; this is relevant to the 'only a few features suffice' claim.
  3. [Datasets and Weak Supervision Framework] The text says 'we aggregate over 50 independently trained BDT classifiers (10 for MLPs as in [15])' but it is unclear whether the 50 BDTs are averaged to one score per event and then SI is computed, or SI is computed per BDT and averaged. Clarify the ensemble procedure.
  4. [Appendix] The axis label 'T PR /(s)' in Figs. 3–5 appears garbled; presumably it should read 'signal efficiency' or 'true positive rate'. Please fix the typesetting.

Circularity Check

0 steps flagged

No construction-level circularity; the reported SI is a measured benchmark rather than a quantity reducible to the OT/PCA fit.

full rationale

The derivation chain is empirical: W2 (Eq. 1) and LinW2 (Eq. 2) define a representation; PCA reduces its dimension; a BDT is trained on A1 vs A2 and evaluated on B1/B2 to obtain SI. No equation reconstructs SI from the fitted PCA modes or from the LinW2 reference, so none of the paper's claimed predictions is equivalent to its inputs by construction. The idealized anomaly detector (IAD) is explicitly labeled an upper-bound idealization ('We therefore assume perfect background interpolation and adopt an idealized anomaly detector (IAD) [15]'), and the low-level/foundation-model comparisons are external benchmarks from Refs. [17,21]. The principal caveats are methodological rather than circular: the PCA basis is fit on a 50/50 signal/background sample ('5k background and 5k signal events randomly selected from R&D1 or R&D2'), which uses event-level labels unavailable in weak supervision, and the claim that 'only 3–5 PCA components suffice' is calibrated on the same scan that reports the SI. These are leakage/selection concerns that would require a background-only PCA control, not reductions of the result to the input. The uniform-reference choice is justified by self-citations ([25,26,50]), but it is a design detail with stated robustness analysis and is not the load-bearing step that produces the benchmark numbers. Overall, no circular step is identifiable by construction, so the score is low.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claim is an empirical benchmark result. It rests on simulated data, an idealized background model, and several representation hyperparameters (reference grid, PCA dimension). No new physical entities are introduced; the OT machinery is standard from prior work. The PCA fit sample composition is a potential data-handling assumption.

free parameters (3)
  • Reference jet grid size and pT normalization = 10x10 grid, each particle pT = 1/100
    The LinW2 embedding is defined relative to this hand-chosen reference; the paper cites prior insensitivity studies but does not optimize or justify it from first principles in this analysis.
  • Number of PCA modes k (OT_k) = k=3-6 optimal; k=100 degrades at S/B<0.6%
    Selected by scanning k and observing SI; the 'few features suffice' claim is calibrated on this scan, and OT100's degradation is reported post hoc.
  • PCA fitting sample size/composition = 5,000 background + 5,000 signal events
    Used to determine variance and construct the PCA transform; the paper does not state whether this sample overlaps the training/test partitions, which is an unresolved methodological choice.
axioms (6)
  • domain assumption Pythia 8 + Delphes simulation accurately models detector response for this benchmark
    All events are simulated; no real-data validation is provided. Invoked in the Datasets and Weak Supervision Framework section.
  • domain assumption Perfect background interpolation (idealized anomaly detector)
    Explicitly assumed in the Datasets section and used for all main results; central performance numbers are upper bounds.
  • domain assumption Two leading anti-kT R=1 jets contain the X/Y decay products
    Event preprocessing discards all other particles; this holds for the benchmark but limits generality to more complex topologies.
  • domain assumption Uniform 10x10 reference jet gives stable LinW2 features
    Adopted from Refs. [25,26,50]; the paper does not test sensitivity to this choice within its own setup.
  • standard math CWoLa / Neyman-Pearson framework for weak supervision
    Used to justify training classifiers on mixed samples; taken from Refs. [6,45].
  • standard math IRC safety of W2 and LinW2
    Claimed 'by definition' in Eq. (1) and after Eq. (2); standard result from Ref. [25].

pith-pipeline@v1.3.0-alltime-deepseek · 9659 in / 17837 out tokens · 156837 ms · 2026-08-03T18:30:20.199207+00:00 · methodology

0 comments
read the original abstract

We introduce optimal transport (OT) as a physics-based intermediate event representation for weakly supervised anomaly detection. With only $0.5\%$ injection of resonant signals in the LHC Olympics benchmark datasets, the OT-augmented feature set achieves nearly twice the significance improvement of the standard high-level observables using an idealized setup, while end-to-end deep learning on low-level four-momenta is less effective in this low-signal regime. The observed gains persist across signal types and classifiers considered in this study, suggesting that structured, physics-informed representations can provide a useful complement to existing approaches for anomaly detection.

Figures

Figures reproduced from arXiv: 2512.04839 by Aditya Bhargava, Benjamin Nachman, Tianji Cai.

Figure 1
Figure 1. Figure 1: FIG. 1: Total variance explained by increasing numbers [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Maximum Significance Improvement (SI) for the R&D1 ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Significance Improvement (SI) curves for the R&D1 ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Anomaly detection for the R&D1 dataset using MLP ensembles as the classifier. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: The SI curves for the R&D1 dataset using BDT ensembles as the classifier at S/B= 0 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Machine learning fully hadronic events with spectral functions

    hep-ph 2026-06 unverdicted novelty 6.0

    Spectral functions from two-point correlations serve as multiplicity-independent ML inputs and improve expected gluino mass reach by 150-250 GeV in a fully hadronic ttbar vs gluino benchmark.

  2. Machine learning fully hadronic events with spectral functions

    hep-ph 2026-06 conditional novelty 6.0

    Spectral-function features fed to a dense network improve expected gluino-mass reach by ~150 GeV over ATLAS Gtt-0L-C and ~250 GeV over kinematics-only ML for the fully hadronic channel.

Reference graph

Works this paper leans on

51 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kasieczka andet al., The lhc olympics 2020 a com- munity challenge for anomaly detection in high energy physics, Reports on Progress in Physics84, 124201 (2021), arXiv:2101.08320

    G. Kasieczka andet al., The lhc olympics 2020 a com- munity challenge for anomaly detection in high energy physics, Reports on Progress in Physics84, 124201 (2021), arXiv:2101.08320. 5

    Pith/arXiv arXiv 2020

  2. [2]

    T. Aarrestadet al., The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider, Sci- Post Phys.12, 043 (2022), arXiv:2105.14027 [hep-ph]

    Pith/arXiv arXiv 2022

  3. [3]

    Karagiorgi, G

    G. Karagiorgi, G. Kasieczka, S. Kravitz, B. Nachman, and D. Shih, Machine learning in the search for new fun- damental physics (2021), arXiv:2112.03769 [hep-ph]

    Pith/arXiv arXiv 2021

  4. [4]

    Belis, P

    V. Belis, P. Odagiu, and T. K. Aarrestad, Machine learn- ing for anomaly detection in particle physics, Reviews in Physics12, 100091 (2024)

    2024

  5. [5]

    Hern´ andez-Gonz´ alez, I

    J. Hern´ andez-Gonz´ alez, I. Inza, and J. A. Lozano, Weak supervision and other non-standard classification prob- lems: A taxonomy, Pattern Recognition Letters69, 49 (2016)

    2016

  6. [6]

    E. M. Metodiev, B. Nachman, and J. Thaler, Classifica- tion without labels: Learning from mixed samples in high energy physics, JHEP10, 174, arXiv:1708.02949 [hep- ph]

    Pith/arXiv arXiv

  7. [7]

    Collins, K

    J. Collins, K. Howe, and B. Nachman, Anomaly detection for resonant new physics with machine learning, Phys. Rev. Lett.121, 241803 (2018)

    2018

  8. [9]

    G. e. a. Aad (ATLAS Collaboration), Weakly supervised anomaly detection for resonant new physics in the dijet fi- nal state using proton-proton collisions at √s= 13 TeV with the atlas detector, Phys. Rev. D112, 072009 (2025)

    2025

  9. [10]

    Gambhir, R

    R. Gambhir, R. Mastandrea, B. Nachman, and J. Thaler, Isolating unisolated upsilons with anomaly detection in cms open data, Phys. Rev. Lett.135, 021902 (2025)

    2025

  10. [11]

    T. A. Vami and D. Zhang, Search for black holes and sphalerons using novel machine learning techniques at cms (2025), arXiv:2511.10662 [hep-ex]

    arXiv 2025

  11. [12]

    D. Shih, M. R. Buckley, L. Necib, and J. Tamanas, via machinae: Searching for stellar streams us- ing unsupervised machine learning, Monthly No- tices of the Royal Astronomical Society509, 5992 (2021), https://academic.oup.com/mnras/article- pdf/509/4/5992/41764058/stab3372.pdf

    2021

  12. [13]

    D. Shih, M. R. Buckley, and L. Necib, Via machinae 2.0: Full-sky, model-agnostic search for stellar streams in gaia dr2, Monthly Notices of the Royal Astronomical Society529, 4745 (2024), https://academic.oup.com/mnras/article- pdf/529/4/4745/57162125/stae446.pdf

    2024

  13. [14]

    Hallin, D

    A. Hallin, D. Shih, C. Krause, and M. R. Buckley, Via Machinae 3.0: A search for stellar streams in Gaia with the CATHODE algorithm, (2025), arXiv:2509.08064 [astro-ph.GA]

    Pith/arXiv arXiv 2025

  14. [15]

    Hallin, J

    A. Hallin, J. Isaacson, G. Kasieczka, C. Krause, B. Nach- man, T. Quadfasel, M. Schlaffer, D. Shih, and M. Som- merhalder, Classifying anomalies through outer den- sity estimation, Physical Review D106, 10.1103/phys- revd.106.055006 (2022), arXiv:2109.00546

    Pith/arXiv arXiv 2022

  15. [16]

    Benkendorfer, L

    K. Benkendorfer, L. Le Pottier, and B. Nach- man, Simulation-assisted decorrelation for resonant anomaly detection, Phys. Rev. D104, 035003 (2021), arXiv:2009.02205

    Pith/arXiv arXiv 2021

  16. [17]

    Buhmann, C

    E. Buhmann, C. Ewen, G. Kasieczka, V. Mikuni, B. Nachman, and D. Shih, Full phase space resonant anomaly detection, Phys. Rev. D109, 055015 (2024), arXiv:2310.06897 [hep-ph]

    Pith/arXiv arXiv 2024

  17. [18]

    Nachman and D

    B. Nachman and D. Shih, Anomaly detection with den- sity estimation, Physical Review D101, 10.1103/phys- revd.101.075042 (2020)

    doi:10.1103/phys- 2020

  18. [19]

    Stein, U

    G. Stein, U. Seljak, and B. Dai, Unsupervised in- distribution anomaly detection of new physics through conditional density estimation (2020), arXiv:2012.11638 [cs.LG]

    Pith/arXiv arXiv 2020

  19. [20]

    C. L. Cheng, G. Singh, and B. Nachman, Incorporating physical priors into weakly supervised anomaly detection, Phys. Rev. Lett.135, 021801 (2025)

    2025

  20. [21]

    Mikuni and B

    V. Mikuni and B. Nachman, Method to simultaneously facilitate all jet physics tasks, Phys. Rev. D111, 054015 (2025), arXiv:2404.16091

    arXiv 2025

  21. [22]

    Peyr´ e and M

    G. Peyr´ e and M. Cuturi, Computational optimal trans- port (2020), arXiv:1803.00567 [stat.ML]

    Pith/arXiv arXiv 2020

  22. [23]

    P. T. Komiske, E. M. Metodiev, and J. Thaler, Metric Space of Collider Events, Phys. Rev. Lett.123, 041801 (2019), arXiv:1902.02346 [hep-ph]

    Pith/arXiv arXiv 2019

  23. [24]

    Cai,Optimal Transport for High Energy Physics, Ph.D

    T. Cai,Optimal Transport for High Energy Physics, Ph.D. thesis, UC, Santa Barbara (main) (2023)

    2023

  24. [25]

    T. Cai, J. Cheng, N. Craig, and K. Craig, Linearized optimal transport for collider events, Phys. Rev. D102, 116019 (2020)

    2020

  25. [26]

    T. Cai, J. Cheng, K. Craig, and N. Craig, Which metric on the space of collider events?, Phys. Rev. D105, 076003 (2022)

    2022

  26. [27]

    T. Cai, J. Cheng, N. Craig, G. Koszegi, and A. J. Larkoski, The phase space distance between collider events (2024), arXiv:2405.16698 [hep-ph]

    Pith/arXiv arXiv 2024

  27. [28]

    P. T. Komiske, E. M. Metodiev, and J. Thaler, The Hidden Geometry of Particle Collisions, JHEP07, 006, arXiv:2004.04159 [hep-ph]

    Pith/arXiv arXiv 2004

  28. [29]

    A. J. Larkoski and J. Thaler, A Spectral Metric for Col- lider Geometry, (2023), arXiv:2305.03751 [hep-ph]

    Pith/arXiv arXiv 2023

  29. [30]

    Gambhir, A

    R. Gambhir, A. J. Larkoski, and J. Thaler, SPECTER: efficient evaluation of the spectral EMD, JHEP12, 219, arXiv:2410.05379 [hep-ph]

    Pith/arXiv arXiv

  30. [31]

    Onyisi, D

    P. Onyisi, D. Shen, and J. Thaler, Comparing point cloud strategies for collider event classification, Physical Re- view D108, 10.1103/physrevd.108.012001 (2023)

    work page doi:10.1103/physrevd.108.012001 2023

  31. [32]

    T. Cai, N. Craig, K. Craig, and X. Lin, Multiscale opti- mal transport for complete collider events, Phys. Rev. D 112, 036021 (2025), arXiv:2501.10681

    Pith/arXiv arXiv 2025

  32. [33]

    G. e. a. Aad (ATLAS), A continuous calibration of the ATLAS flavour-tagging classifiers via optimal trans- portation maps, Eur. Phys. J. C85, 1272 (2025), arXiv:2505.13063 [hep-ex]

    arXiv 2025

  33. [34]

    Craig, J

    N. Craig, J. N. Howard, and H. Li, Exploring Opti- mal Transport for Event-Level Anomaly Detection at the Large Hadron Collider, (2024), arXiv:2401.15542 [hep- ph]

    Pith/arXiv arXiv 2024

  34. [35]

    Brennan, T

    L. Brennan, T. A. Vami, O. Amram, S. Sekhar, Y. Taka- hashi, L. Moureaux, M. Sommerhalder, P. Maksimovic, T. Cai, and N. Craig, Weakly supervised anomaly detec- tion with event-level variables, Phys. Rev. D112, 055040 (2025), arXiv:2504.13249

    arXiv 2025

  35. [36]

    R. T. D’Agnolo, A. Glioti, G. Rigo, and A. Valenti, The intrinsic dimension of collider events and model- independent searches in 100 dimensions (2025), arXiv:2511.20760 [hep-ph]

    arXiv 2025

  36. [37]

    Kasieczka, B

    G. Kasieczka, B. Nachman, and D. Shih, R&D Dataset for LHC Olympics 2020 Anomaly Detection Challenge, 10.5281/zenodo.6466204 (2022)

    work page doi:10.5281/zenodo.6466204 2020

  37. [38]

    Shih, Additional QCD Background Events for LHCO2020 R&D (signal region only), 10.5281/zen- 6 odo.8370758 (2023)

    D. Shih, Additional QCD Background Events for LHCO2020 R&D (signal region only), 10.5281/zen- 6 odo.8370758 (2023)

    doi:10.5281/zen- 2023

  38. [39]

    Sjostrand, S

    T. Sjostrand, S. Mrenna, and P. Z. Skands, PYTHIA 6.4 Physics and Manual, JHEP05, 026, arXiv:hep- ph/0603175 [hep-ph]

    arXiv

  39. [40]

    Sjostrand, S

    T. Sjostrand, S. Mrenna, and P. Z. Skands, A Brief Intro- duction to PYTHIA 8.1, Comput. Phys. Commun.178, 852 (2008), arXiv:0710.3820 [hep-ph]

    Pith/arXiv arXiv 2008

  40. [41]

    de Favereau, C

    J. de Favereau, C. Delaere, P. Demin, A. Giammanco, V. Lemaitre, A. Mertens, and M. Selvaggi, DELPHES 3: A modular framework for fast simulation of a generic col- lider experiment, JHEP02, 057, arXiv:1307.6346 [hep- ex]

    Pith/arXiv arXiv

  41. [42]

    M. Selvaggi, DELPHES 3: A modular framework for fast- simulation of generic collider experiments, inProceed- ings, 15th International Workshop on Advanced Comput- ing and Analysis Techniques in Physics Research (ACAT 2013): Beijing, China, May 16-21, 2013, Vol. 523 (2014) p. 012033

    2013

  42. [43]

    A. Mertens, New features in Delphes 3, inProceedings, 16th International Workshop on Advanced Computing and Analysis Techniques in Physics (ACAT 14): Prague, Czech Republic, September 1-5, 2014, Vol. 608 (2015) p. 012045

    2014

  43. [44]

    Cacciari, G

    M. Cacciari, G. P. Salam, and G. Soyez, Fastjet user manual, The European Physical Journal C72, 10.1140/epjc/s10052-012-1896-2 (2012)

    doi:10.1140/epjc/s10052-012-1896-2 2012

  44. [45]

    Neyman, E

    J. Neyman, E. S. Pearson, and K. Pearson, Ix. on the problem of the most efficient tests of statistical hypothe- ses, Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character231, 289 (1933)

    1933

  45. [46]

    Andreassen, B

    A. Andreassen, B. Nachman, and D. Shih, Simulation assisted likelihood-free anomaly detection, Phys. Rev. D 101, 095004 (2020), arXiv:2001.05001

    Pith/arXiv arXiv 2020

  46. [47]

    J. A. Raine, S. Klein, D. Sengupta, and T. Golling, CUR- TAINs for your sliding window: Constructing unobserved regions by transforming adjacent intervals, Front. Big Data6, 899345 (2023), arXiv:2203.09470 [hep-ph]

    Pith/arXiv arXiv 2023

  47. [48]

    Finke, M

    T. Finke, M. Hein, G. Kasieczka, M. Kr¨ amer, A. M¨ uck, P. Prangchaikul, T. Quadfasel, D. Shih, and M. Som- merhalder, Tree-based algorithms for weakly supervised anomaly detection, Phys. Rev. D109, 034033 (2024), arXiv:2309.13111

    Pith/arXiv arXiv 2024

  48. [49]

    Freytsis, M

    M. Freytsis, M. Perelstein, and Y. C. San, Anomaly de- tection in the presence of irrelevant features, Journal of High Energy Physics2024, 220 (2024)

    2024

  49. [50]

    T. Cai, J. Cheng, B. Schmitzer, and M. Thorpe, The linearized hellinger–kantorovich distance, SIAM Journal on Imaging Sciences15, 45 (2022), arXiv:2102.08807

    Pith/arXiv arXiv 2022

  50. [51]

    Cesarotti and J

    C. Cesarotti and J. Thaler, A robust measure of event isotropy at colliders (2020), arXiv:2004.06125 [hep-ph]

    Pith/arXiv arXiv 2020

  51. [52]

    Craig, N

    K. Craig, N. G. Trillos, and D. Nikoli´ c, Vector valued optimal transport: from dynamic to static formulations (2025), arXiv:2505.03670 [math.AP]. 7 APPENDIX 0.0 0.2 0.4 0.6 0.8 1.0 T PR /uni00A0( s) 0 1.3 2 4 6 8 10 12 14 16 18 20 22 24 26SI 0 3 |/uni00A0 6 1,/uni00A02,/uni00A0100 O T 0 (High/uni00ADlevel) O T 1 O T 2 O T 3 O T 4 O T 5 O T 6 O T 100 PS/...

    Pith/arXiv arXiv 2025