Pith. sign in

REVIEW 4 major objections 8 minor 34 references

Polarization fraction measurement in ZZ scattering using deep learning

T0 review · 4 major / 8 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A particle-based deep neural network that preprocesses inputs with a power transformation and decorrelates its outputs with PCA raises the expected significance for measuring the longitudinal ZZ polarization fraction at the HL-LHC to 1.74…

desk verdict A clean fast-simulation study showing a modest but real gain for longitudinal ZZ extraction via a particle-based DNN with PCA, yet the HL-LHC projection rests on a no-pileup assumption that could easily erase the improvement. read the letter →

arxiv 1908.05196 v2 pith:ADAOT3QP submitted 2019-08-14 hep-ph hep-ex

Junho Lee , Nicolas Chanon , Andrew Levin , Jing Li , Meng Lu , Qiang Li , Yajun Mao This is my paper · ORCID
classification hep-phhep-ex
keywords vectorbosonscatteringZZpolarizationlongitudinalfractiondeepneuralnetworkprincipalcomponentanalysisYeo-JohnsontransformationHL-LHChelicity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a deep-learning pipeline can measure the longitudinally polarized fraction of ZZ vector-boson scattering at the High-Luminosity LHC with an expected significance of about 1.7 standard deviations, using the full planned dataset of 3000 inverse femtobarns. The pipeline feeds the four-momenta of the four leptons and two jets into a per-particle neural network, preprocesses the inputs with a power transformation plus standardization, then applies principal-component analysis to the network's five class scores and fits the leading three components in three dimensions. This matters because the longitudinal (LL) component of vector-boson scattering is the piece tied directly to unitarity restoration through the Higgs mechanism and to possible new physics; a 1.7-sigma expectation would improve on the boosted-decision-tree baseline of about 1.4 sigma and would make a combined measurement by the two HL-LHC experiments likely to exceed 2 sigma.

What carries the argument

The central object is a particle-based deep neural network: each lepton and jet enters as its own small input block built from its four-momentum, the blocks are merged hierarchically, and the network ends in five output nodes scoring the classes LL, TL, TT, qqZZ, and ggZZ. Two preprocessing steps carry much of the gain: the Yeo-Johnson power transformation, which makes skewed input distributions nearly Gaussian and handles negative values, followed by standardization computed on the LL training sample. The final mechanism is PCA on the five output scores; the leading three principal components, which together explain about 96% of the variance, are fitted simultaneously in three dimensions. This multidimensional fit extracts more signal than a cut on any single discriminant because the five class scores carry correlated shape information that PCA concentrates into a few decorrelated axes.

What would settle it

Re-run the same Monte Carlo analysis with pileup included at 140 to 200 collisions per bunch crossing and compare the DNN-PCA significance with the BDT baseline; if the DNN-PCA significance drops to the BDT's 1.41 sigma or below, the claimed improvement is an artifact of the pileup-free simulation, while staying above about 1.6 sigma would confirm the central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that in the fully leptonic ZZ channel, the expected significance for the LL polarization fraction at 3000 inverse femtobarns can be raised from 1.41 sigma with a two-step boosted decision tree to 1.74 sigma with a particle-based DNN whose five class outputs are decorrelated by PCA and fitted in three dimensions; including a flat 10% systematic uncertainty on signal and background, the DNN-PCA still gives 1.66 sigma. The paper compares several discriminants and finds that the particle-based DNN with a power transformation plus standardization has the best single-classifier separation. It then shows that treating the five output scores as a vector, rotating them with PCA, and fitting the leading three principal components yields more significance than either a single discriminant or a two-step DNN. The paper presents this as evidence that this machine-learning pipeline makes the longitudinal VBS measurement feasible at the HL-LHC and transferable to other helicity-fraction measurements.

Load-bearing premise

The projection relies on neglecting pileup—the detector simulation is configured for the HL-LHC but with the extra proton-proton collisions per bunch crossing turned off—so if pileup degrades lepton isolation or jet properties enough to reduce the network's separation power, the quoted 1.74-sigma significance would be optimistic.

Editorial extensions

If this is right

  • At 3000 inverse femtobarns the longitudinal ZZ fraction could be extracted with an expected significance near 1.7 sigma, better than the 1.41-sigma BDT baseline.
  • Combining the measurements of the two HL-LHC experiments would push the expected significance above 2 sigma, moving from sensitivity toward evidence for the longitudinal component.
  • Because the pipeline uses only per-particle four-momenta and multiclass scores, the same recipe can be applied to helicity-fraction measurements in other vector-boson-scattering channels.
  • Since the LL component grows strongly at high mass when Higgs or gauge couplings deviate from the standard model, the same DNN-PCA analysis could serve as a model-independent probe of Higgs couplings and anomalous gauge couplings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The PCA-on-outputs step suggests a general recipe for multiclass classifiers in high-energy physics: instead of thresholding one output score, keep the full score vector, decorrelate it, and fit the leading components; this can be tested on any multiclass tagger already in use.
  • If pileup at the HL-LHC degrades performance, the preprocessing is the first place to re-adapt: the power transformation is fit to the LL training sample, so a pileup-contaminated sample would shift both the transformation and the PCA axes.
  • A significance of 1.74 sigma still falls short of discovery; the method's real payoff may come from combining the ZZ channel with WW and WZ scattering, where the same longitudinal-fraction question has higher rates, or from using the PC distributions as inputs to a broader Higgs-coupling fit.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. This paper studies the extraction of the longitudinal (LL) polarization fraction in vector-boson-scattering ZZ -> 4l events at the HL-LHC using fast simulation. The authors generate LL, TL, TT, qqZZ, and ggZZ samples with MadGraph/Pythia/Delphes, apply a VBS-like event selection, and compare a BDT, a dense DNN, and a particle-based DNN with preprocessing by standardization and Yeo-Johnson transformation, followed by principal-component analysis of the DNN outputs. They report expected significances from the Asimov formula, with the best result being 1.74 sigma (1.66 sigma with a flat 10% systematic uncertainty) at 3000 fb^-1 using a three-dimensional fit to PC1-PC3, compared with 1.41 sigma for their BDT baseline. The paper concludes that the method improves the LL sensitivity and is applicable to other helicity-fraction measurements.

Significance. If the projection is robust, the paper offers a useful, relatively simple improvement over the previous CMS BDT projection of about 1.4 sigma for a challenging measurement, and the comparison across preprocessing and model variants is informative. Its strengths are that all significance values are listed in Table I and Fig. 8, the Asimov significance procedure is standard, and the BDT baseline is explicitly validated against the CMS result. The main weaknesses are that the simulation neglects pileup, which directly affects the inputs on which the event selection and the DNN rely, and that the systematic uncertainty is an unvalidated flat 10% assumption; both make the quantitative HL-LHC projection, rather than the qualitative methodological conclusion, uncertain.

major comments (4)
  1. [Simulation and event selection paragraph] The simulation paragraph explicitly states that 'pileup' has been neglected as in ref. [24]. This is load-bearing for the central claim because the event selection requires at least two jets with pT > 25 GeV, |eta| < 4.7, mjj > 400 GeV, |Delta eta_jj| > 2.4, no b-tagged jets, and four isolated leptons, and the DNN consumes exactly these lepton and jet four-momenta. At the HL-LHC with 140-200 pileup interactions per bunch crossing, pileup can add jets, alter the selected jet pair, break the b-veto, and degrade lepton isolation, shifting the DNN scores and the PC templates used in the 1.74 sigma fit. Because the improvement over the BDT baseline (1.41 sigma to 1.74 sigma) is modest, even a moderate pileup-induced degradation could erase the claimed gain. Please rerun with pileup overlay, or at least with a robustness test against isolation and jet-threshold variations, or clearly restate the result as a no-pileup feasibility projection.
  2. [Table I and preceding text] The 'Stat. & syst.' row applies a flat 10% uncertainty to both signal and background yields, but the paper gives no derivation or validation for this value. The quoted significance with systematics is therefore not a measured property of the method but an input assumption; different correlations or uncertainties for qqZZ, ggZZ, and signal would materially change the 1.66 sigma number. The authors should justify the 10% value, test its variation (for example 5% and 20%), or present the result as statistical-only with the systematic treated as an illustrative prescription.
  3. [PC fit paragraph and Fig. 7] The paper reports 1.65 sigma for a PC1-PC2 fit and 1.74 sigma for a PC1-PC3 fit but does not define the test statistic, binning, or likelihood used for the Asimov calculation on multidimensional histograms. Without this information, a reader cannot reproduce the central number or assess whether the gain over using PC1 alone (1.55 sigma) is meaningful. Please specify the exact fit procedure (for example, binned Poisson likelihood, number of bins per dimension, and nuisance parametrization) and, if feasible, show the sensitivity of the significance to the binning choice.
  4. [Table I and Fig. 8] The final significance is selected as the maximum over several preprocessing schemes, DNN variants, and fit dimensionalities. The paper reports all variants, which is commendable, but it does not account for this post-hoc selection in the significance statement. At minimum, state that the 1.74 sigma is a post-hoc optimum and discuss the look-elsewhere effect, or pre-specify the PC3 fit as the analysis plan.
minor comments (8)
  1. [Title] The title contains 'd eep learning' with a spurious space; it should read 'deep learning'.
  2. [Abstract and throughout] The phrase 'principle component analysis' is used repeatedly; the correct statistical term is 'principal component analysis'.
  3. [Fig. 7] The labels in Fig. 7 are garbled: the LL panel appears twice with '(a)' and the TL panel is also labeled '(b)' twice; the intended labels should be (a) LL, (b) TL, and (c) qqZZ.
  4. [Fig. 1] The legend entry 'DNN ( article-based)' should be 'DNN (particle-based)'.
  5. [General text] There are several typos: 'Prbing' should be 'Probing', 'gaugue boson' should be 'gauge boson', and 'Comparision' should be 'Comparison'.
  6. [PACS line] The PACS line lists keywords ('HL-LHC, vector boson scatter, electroweak') rather than PACS codes; either provide actual PACS codes or remove the line.
  7. [Reference [29]] Reference [29] is a Wikipedia article; a standard textbook or review reference for principal component analysis would be more appropriate.
  8. [Reproducibility] The paper does not state whether the code or trained models will be made available. A data-availability statement would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DNN-PCA significance is derived from independent simulated templates and an Asimov fit, with only a non-load-bearing self-citation.

full rationale

The central claim, an expected 1.74 sigma significance for the longitudinal ZZ fraction using DNN-PCA at 3000 fb^-1, is computed from Monte Carlo samples generated with MadGraph/Pythia/Delphes and an Asimov likelihood fit to principal-component templates. No target quantity is defined in terms of a fitted parameter: the LL fraction is extracted from template shapes, and the DNN itself is trained on generator-level labels and retested on held-out events, with the paper reporting that 'the loss values are comparable for the training and test samples'. The BDT baseline is reproduced internally (1.41 sigma) and matches the external CMS projection of about 1.4 sigma from ref. [15], so the comparison is anchored outside the paper's own fitted values. The particle-based DNN architecture is taken from ref. [14] by the same authors, but for this ZZ study the network is retrained, preprocessed with standardization and Yeo-Johnson transformation, and evaluated on a different final state; the prior paper is a method reference, not a logical premise of the new significance. The explicit statement that 'pileup has been neglected as in ref. [24]' is a modeling limitation that could bias the absolute projection, but it is not a circular step because the claimed significance is not defined in terms of that assumption. No equation or fitted parameter reduces by construction to an input, so no circularity is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central sensitivity rests on simulation and modeling assumptions rather than on a measurement. The DNN architecture and preprocessing choices are drawn from prior work by the same team, and the systematic uncertainty is a flat ad hoc number. No new physical entities are introduced.

free parameters (3)
  • Flat 10% systematic uncertainty on signal and background yields = 10%
    Applied to both signal and background when computing the 'stat. & syst.' significances; no derivation from detector studies or data is given.
  • DNN hyperparameters and architecture = Learning rate 0.001, Adam, ReLU, layer sizes as listed in text
    Chosen by hand or standard practice; no hyperparameter search or cross-validation is reported, yet performance depends on these choices.
  • Number of principal components used in the final fit = 3
    Selected after comparing PC1, PC1+PC2, and PC1+PC2+PC3; the reported best significance uses the 3D choice, introducing selection bias.
assumptions (5)
  • domain assumption The MadGraph/Pythia/Delphes simulation chain correctly models ZZ vector boson scattering, its polarization decomposition, and the qqZZ and ggZZ backgrounds.
    Invoked throughout the event generation paragraph; no comparison to data or to an independent generator is shown.
  • domain assumption Delphes fast simulation is a sufficient proxy for the CMS HL-LHC detector response for the discriminants studied.
    Used for all reconstructed objects; fast simulation may differ from full Geant4 simulation, especially for jet substructure and lepton isolation.
  • domain assumption Neglecting pileup does not materially change the ranking or significance of the discriminants.
    Explicitly stated in the detector simulation paragraph; pileup at HL-LHC could affect jets and leptons.
  • domain assumption The Asimov significance with a flat 10% yield uncertainty estimates the expected measurement sensitivity.
    Used for all quoted significances; real systematics are more complex and correlated.
  • domain assumption The Monte Carlo samples are large enough that statistical fluctuations in training do not affect the quoted significance.
    Sample sizes after selection are given (40k to 240k events), but no error bars on the significances are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Polarization fraction measurement in ZZ scattering using deep learning." pith.science (2026). https://pith.science/paper/ADAOT3QP

@misc{pith2026190805196,
  author       = {Pith},
  title        = {Pith review of: Polarization fraction measurement in ZZ scattering using deep learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADAOT3QP}},
  note         = {Machine review of arXiv:1908.05196}
}
read the original abstract

Measuring longitudinally polarized vector boson scattering in the ZZ channel is a promising way to investigate unitarity restoration with the Higgs mechanism and to search for possible new physics. We investigated several deep neural network structures and compared their ability to improve the measurement of the longitudinal fraction Z_L Z_L. Using fast simulation with the Delphes framework, a clear improvement is found using a previously investigated 'particle-based' deep neural network on a preprocessed dataset and applying principle component analysis to the outputs.A significance of around 1.7 standard deviations can be achieved with the integrated luminosity of 3000 fb-1 that will be recorded at the High-Luminosity LHC.

Figures

Figures reproduced from arXiv: 1908.05196 by the authors.

Figure 1
Figure 1. FIG. 1: [left:] Comparision of 4 lepton invariant mass distr [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: Comparison of different data preprocessing methods. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3: Distribution of the LL score of the particle-based DN [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FIG. 4: Comparison of data preprocessing methods on the disc [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5: Leading three principle component distribution for [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6: Two dimensional histograms based on PC1 and PC2. (a), [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7: Three dimensional histograms based on PC1, PC2, and P [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8: Expected signal significances obtained with the vari [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages

  1. [24]

    Alwall et al

    J. Alwall et al. , JHEP 1407, 079 (2014) doi:10.1007/JHEP07(2014)079 [arXiv:1405.03 01 [hep-ph]]

    doi:10.1007/jhep07(2014)079 2014

  2. [1]

    VBS (BDT1) and then cla ssifies between LL and all backgrounds (BDT2)

    The two-step BDT first classifies QCD vs. VBS (BDT1) and then cla ssifies between LL and all backgrounds (BDT2). BDT2 was trained on the events left after a selection on th e output of BDT1, which maximizes S/ √ B. We obtain a significance of 1.41 σ when only statistical uncertainties are considered, and a significan ce of 1.23 σ when a 10% uncertainty is appl...

    work page

  3. [2]

    The first DNN structure is the previously described particle-b ased DNN, while a shallow dense NN is used as second DNN

    Similarly to the two-step BDT, in the two-step DNN the second DNN is applied on the output of the first DNN. The first DNN structure is the previously described particle-b ased DNN, while a shallow dense NN is used as second DNN. Since there are only five output nodes on the first DNN , which represents LL, TL, TT, qqZZ, and ggZZ score, the second DNN is shal...

    work page

  4. [3]

    The PCA algorithm rotates the original axis of the features into a new axis containing decorrela ted features

    The DNN-PCA has only one particle-based DNN, the same particle- based model as was used in the first part of the two-step DNN, but with principle component analysis (PCA) implement ed on its output. The PCA algorithm rotates the original axis of the features into a new axis containing decorrela ted features. More specifically, given n-dimension target-data ...

    work page

  5. [4]

    A. M. Sirunyan et al. [CMS Collaboration], Phys. Rev. Lett. 120, no. 8, 081801 (2018) doi:10.1103/PhysRevLett.120.08180 1 [arXiv:1709.05822 [hep-ex]]

    arXiv 2018

  6. [5]

    Aaboud et al

    M. Aaboud et al. [ATLAS Collaboration], arXiv:1906.03203 [hep-ex]

    arXiv 1906

  7. [6]

    A. M. Sirunyan et al. [CMS Collaboration], Phys. Lett. B 795, 281 (2019) doi:10.1016/j.physletb.2019.05.042 [arXiv:1901.04060 [hep-ex]]

    arXiv 2019

  8. [7]

    The ATLAS collaboration [ATLAS Collaboration], ATLAS- CONF-2018-033

    work page 2018

Show all 34 references
  1. [8]

    Khachatryan et al

    V. Khachatryan et al. [CMS Collaboration], Phys. Lett. B 770, 380 (2017) doi:10.1016/j.physletb.2017.04.071 [arXiv:1702.03025 [hep-ex]]

    2017 arXiv

  2. [9]

    CMS Collaboration [CMS Collaboration], CMS-PAS-SMP-1 8-007

  3. [10]

    ATLAS Collaboration [ATLAS Collaboration], ATLAS-CON F-2019-033

    2019

  4. [11]

    A. M. Sirunyan et al. [CMS Collaboration], Phys. Lett. B 774, 682 (2017) doi:10.1016/j.physletb.2017.10.020 [arXiv:1708.02812 [hep-ex]]

    2017 arXiv

  5. [12]

    Chang, K

    J. Chang, K. Cheung, C. T. Lu and T. C. Yuan, Phys. Rev. D 87, 093005 (2013) doi:10.1103/PhysRevD.87.093005 [arXiv:1303.6335 [hep-ph]]

    2013 arXiv

  6. [13]

    S. J. Lee, M. Park and Z. Qian, Phys. Rev. D 100, no. 1, 011702 (2019) doi:10.1103/PhysRevD.100.011702 [arXiv:1812.02679 [hep-ph]]

    2019 arXiv

  7. [14]

    Quigg, Rev

    C. Quigg, Rev. Accel. Sci. Tech. 10 (2019) no.01, 3 doi:10.1142/9789811209604 0002, 10.1142/S1793626819300020 [arXiv:1808.06036 [hep-ph]]

    2019 arXiv

  8. [15]

    Brooijmans et al

    G. Brooijmans et al. , arXiv:1405.1617 [hep-ph]

    arXiv

  9. [16]

    CMS Collaboration [CMS Collaboration], CMS-PAS-FTR- 18-005

  10. [17]

    J. Lee, N. Chanon, A. Levin, J. Li, M. Lu, Q. Li and Y. Mao, P hys. Rev. D 99, no. 3, 033004 (2019) doi:10.1103/PhysRevD.99.033004 [arXiv:1812.07591 [hep -ph]]

    2019 arXiv

  11. [18]

    CMS Collaboration [CMS Collaboration], CMS-PAS-FTR- 18-014

  12. [19]

    B. P. Roe, H. J. Yang, J. Zhu, Y. Liu, I. Stancu and G. McGre gor, Nucl. Instrum. Meth. A 543, no. 2-3, 577 (2005) doi:10.1016/j.nima.2004.12.018 [physics/0408124]

    2005 arXiv

  13. [20]

    Hocker, A. et al. TMV A - Toolkit for Multivariate Data An alysis. PoS ACAT, 040 (2007)

    2007

  14. [21]

    Chollet et al., https://github.com/fchollet/kera s

    F. Chollet et al., https://github.com/fchollet/kera s

  15. [22]

    Software available from tensorflow.org

    Mart ˜An Abadi et al., TensorFlow: Large-scale machine learning o n heterogeneous systems, 2015. Software available from tensorflow.org

    2015

  16. [23]

    I. Yeo, R. A. Johnson, Biometrika 87, no. 4, 954-959 (2000) doi:10.1093/biomet/87.4.954

    2000 doi

  17. [25]

    Sjostrand, L

    T. Sjostrand, L. Lonnblad, S. Mrenna and P. Z. Skands, he p-ph/0308153

  18. [26]

    de Favereau et al

    J. de Favereau et al. [DELPHES 3 Collaboration], JHEP 1402, 057 (2014) [arXiv:1307.6346 [hep-ex]]

    2014 arXiv

  19. [27]

    Searcy, L

    J. Searcy, L. Huang, M. A. Pleier and J. Zhu, Phys. Rev. D 93, no. 9, 094033 (2016) doi:10.1103/PhysRevD.93.094033 [arXiv:1510.01691 [hep-ph]]

    2016 arXiv

  20. [28]

    K. He, X. Zhang, S. Ren and J. Sun CoRR abs/1502.01852 (2015) [arXiv:1502.01852 [cs]]

    2015 arXiv

  21. [29]

    Kingma, D.P., and Ba, J.: 2014, arXiv e-prints , arXiv:1412.6980

    2014 arXiv

  22. [30]

    G. E. P. Box and D. R. Cox, Journal of the Royal Statistica l Society. Series B (Methodological) 26, no. 2, 211-252 (1964)

    1964

  23. [31]

    Cowan, The European Physical Journal C 71, no

    G. Cowan, The European Physical Journal C 71, no. 2, 1554 (2011) doi:10.1140/epjc/s10052-011-1554-0

    2011 doi

  24. [32]

    Wikipedia contributors, Principal component analysi s — Wikipedia, The Free Encyclopedia (2019) ”https://en.wikipedia.org/w/index.php?title=Principal_component_analysis&oldid=910232284”,

    2019

  25. [33]

    Brehmer, K

    J. Brehmer, K. Cranmer, G. Louppe and J. Pavez, Phys. Rev . D 98 (2018) no.5, 052004 doi:10.1103/PhysRevD.98.052004 [arXiv:1805.00020 [hep-ph]]

    2018 arXiv

  26. [34]

    Brehmer, K

    J. Brehmer, K. Cranmer, G. Louppe and J. Pavez, Phys. Rev . Lett. 121 (2018) no.11, 111801 doi:10.1103/PhysRevLett.121.111801 [arXiv:1805.00013 [hep-ph]]

    2018 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.