Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Decorrelation of neural networks from particle lifetimes in the LHCb topological $b$ trigger

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Adding a quadratic moment-decomposition penalty to LHCb's topological b-trigger neural networks flattens their efficiency for long-lived b-hadrons without losing classification power.

desk verdict Useful simulation study of lifetime decorrelation for the LHCb topological b trigger, but the claimed MoDe superiority is not statistically supported. read the letter →

arxiv 2607.22281 v1 pith:4QHKXYEY submitted 2026-07-24 hep-ex

Johannes Albrecht , Alessandro Bertolin , James Connaughton , Jonathan Davies , Blaise Delaney , Agnieszka Dziurda , Conor Fitzpatrick , Maciej Giza
show 8 more authors
Vava V. Gligorov James A. Gooding Nicole Schulte Nicole Skidmore Mika Vesterinen Mike Williams Shunan Zhang Valeriia Zhovkovska
This is my paper
classification hep-ex
keywords topologicalbtriggerlifetimedecorrelationLHCbHLT2monotonicLipschitzneuralnetworkspile-upbackgroundMoDeDisCo
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

The paper argues that in Run 3 of the LHC, with about five visible proton-proton collisions per bunch crossing, LHCb's topological b trigger suffers from a new background: vertices made from tracks originating in different collisions that mimic long-lived b decays. This background makes the neural-network trigger less efficient for genuinely long-lived b hadrons, which would bias time-dependent measurements. The paper presents two ways to train the trigger networks to decorrelate their score from lifetime and shows in simulation that a moment-decomposition penalty (MoDe) with strength lambda=0.2 flattens the efficiency at large lifetimes better than the distance-correlation (DisCo) penalty. If correct, this keeps the trigger unbiased for particles living longer than about 4-5 ps, preserving sensitivity to mixing and CP-violation measurements.

What carries the argument

The MoDe penalty expands the cumulative distribution of network scores in bins of lifetime into Legendre polynomials and penalises the difference between the actual distribution and its polynomial fit; with order l=2 and monotonicity constraints it allows the sharp efficiency turn-on at small lifetimes while enforcing flatness at large lifetimes. DisCo uses distance correlation between score and lifetime as a penalty. The comparison is made by fitting linear slopes to the efficiency as a function of true lifetime beyond 4 ps.

What would settle it

Take a real Run 3 data sample where the topological b trigger did not fire, e.g., selected by a muon trigger, reconstruct B+ to J/psi K+ candidates, and measure the trigger efficiency as a function of reconstructed B+ decay time for tau>4 ps; if the fitted slope is significantly more negative than the predicted -11 ns^-1, the MoDe-lambda=0.2 claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that the decay-time bias introduced by pile-up-related track misassociation can be removed by adding a quadratic penalty based on Legendre decomposition of the classifier score distribution as a function of lifetime. Measured on simulated B+ to J/psi K+ and B0 to D- pi+ events passing HLT2, the MoDe-penalised models (lambda=0.2) reduce the negative slope of trigger efficiency versus true lifetime for tau>4 ps from about -25 ns^-1 to about -11 ns^-1 (2-body) and from about -22 ns^-1 to about -7 ns^-1 (3-body, consistent with zero), while preserving precision, recall and ROC AUC. The paper concludes MoDe is better suited than DisCo for lifetime decorrelation in the topolog

Load-bearing premise

Everything is demonstrated in simulated events; if the simulation does not reproduce the Run 3 rate of multiple simultaneous collisions and their track-association properties, the observed flattening may not hold for real data.

Editorial extensions

If this is right

  • Time-dependent CP and mixing analyses using the topological b trigger will no longer lose statistical sensitivity from a sharp drop in acceptance for b hadrons living beyond about 4 ps.
  • The optimal MoDe working point (lambda=0.2) can be adopted for both 2- and 3-body trigger models without retuning, as it preserves precision and ROC AUC.
  • The identified PV-misassociation background, prominent at high pile-up, becomes a standard systematic to check in Run 3 trigger and offline selections.
  • The linear-slope criterion over tau>4 ps provides a simple benchmark for future trigger decorrelation studies.

Reading between the lines

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

  • If the simulation's pile-up model is representative, the same MoDe recipe should transfer to other LHCb inclusive triggers that use the same MLNN architecture, not only the two- and three-body lines.
  • The method's reliance on true lifetime in training means it is best applied when signal simulation is trustworthy; an analogous data-driven version could use reconstructed decay-time sidebands to estimate the lifetime distribution.
  • A natural next test is to verify that the efficiency flattening survives when the trigger is run in real Run 3 conditions, where track-association algorithms may behave differently under fluctuating pile-up.
  • Since MoDe with l=2 explicitly enforces monotonicity at large lifetimes, it could be combined with the monotonic Lipschitz architecture to give a formal guarantee of no lifetime sculpting beyond a chosen scale.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper addresses a lifetime-correlation problem in the LHCb topological b trigger, introduced by the higher pile-up conditions of Run 3 (mu = 5.3). The authors identify a background arising from tracks produced in different primary vertices being combined into a candidate, which creates a long-lifetime tail that can bias the trigger efficiency at large decay times. To mitigate this, they train monotonic Lipschitz neural networks for the 2-body and 3-body topological lines using two decorrelation penalties: distance correlation (DisCo) and moment decomposition (MoDe). They compare the resulting trigger efficiencies as a function of true b-hadron lifetime in simulated B+ -> J/psi K+ and B0 -> D- pi+ samples, fit linear slopes above 4 ps for the B0 sample, and conclude that MoDe is better suited for lifetime decorrelation in the topological b trigger.

Significance. If the central claim is established, the paper is practically significant: it identifies a concrete Run 3 trigger limitation and demonstrates two workable mitigation strategies, with potential to preserve large-lifetime acceptance for time-dependent LHCb analyses. The paper is empirically grounded: the authors train and evaluate actual MLNN models in full Pythia/EvtGen/Geant4 simulation, include a no-decorrelation baseline that makes the problem visible, and provide slope fits with uncertainties. The identification of the PV-misassociation background and the demonstration that both penalties reduce the negative large-lifetime slope are useful contributions. However, the headline conclusion that MoDe is superior is currently supported only by point estimates with overlapping uncertainties and an unmatched operating-point comparison; that conclusion needs stronger statistical and methodological support before it is suitable for publication.

major comments (3)
  1. [Sec. 5, Table 1] The claim that MoDe is "better suited" is not statistically established. For the 2-body models, the fitted slopes are -11.23 +/- 1.6 ns^-1 (MoDe) and -12.9 +/- 4.7 ns^-1 (DisCo); for 3-body, -6.6 +/- 5.5 and -8.8 +/- 1.2 ns^-1. In both cases the uncertainties overlap substantially, and for 3-body the MoDe slope is consistent with both zero and the DisCo value. No test of the slope difference is provided, and the fits are independent rather than joint, so the stated ranking rests on point estimates. Please add a quantitative comparison of the slopes (e.g., a difference with propagated uncertainty, a chi-square/confidence-interval statement, or a test that accounts for the correlation between models) before claiming superiority.
  2. [Sec. 5, Figs. 3 and 4] Slopes are reported only for B0 -> D- pi+ (Fig. 4 and Table 1), while the conclusion is stated generally for the topological b trigger. The B+ -> J/psi K+ efficiencies in Fig. 3 are shown only as curves, with no fitted slopes or uncertainties, so the reader cannot judge whether the MoDe-vs-DisCo ranking holds for that decay mode. Either fit and report slopes for both decay modes, or explicitly limit the quantitative claim to B0 -> D- pi+ and provide qualitative evidence for the other sample.
  3. [Sec. 5 and Appendix B] The comparison uses lambda = 0.8 for DisCo and lambda = 0.2 for MoDe, selected from Fig. 5 based on precision/recall/ROC AUC rather than on decorrelation strength. Since the penalty strength controls the degree of decorrelation, differences in slopes could reflect the unmatched lambda values rather than an intrinsic difference between the methods. To support the ranking, the authors should either match the methods by an achieved decorrelation level (e.g., equal dCorr or equal large-lifetime slope) or show that the conclusion is robust to reasonable choices of lambda on both sides. As written, a different lambda for either method could plausibly change the ordering, especially for the 3-body case.
minor comments (4)
  1. [General] The phrase "Appendix Appendix A" in Sec. 2 should read "Appendix A". Similarly, "Table. 1" in Sec. 5 should be "Table 1". There are also minor grammatical issues, e.g., "the version of MoDe used not penalizing the turn-on curve" should be rephrased.
  2. [Figs. 3 and 4] The efficiency points are shown without uncertainty bands or error bars. Since Table 1 gives slope uncertainties, the figure would be clearer if the data points included statistical uncertainties, particularly at large lifetime where the event counts decrease.
  3. [Table 3] The structure of Table 3 is hard to read: the DisCo 3-body row appears to have no penalty strengths listed, and the MoDe 3-body row is not aligned with the 2-body row. Please reformat using separate sub-tables or explicit column entries.
  4. [Appendix B] The text says the 3-body MoDe model has a maximum precision at lambda = 0.1 and significant degradation beyond, but then declares lambda = 0.2 optimal. This is a defensible choice, but the reasoning would be clearer if the trade-off were quantified, e.g., by stating how much each metric changes between lambda = 0.1 and 0.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the lifetime-decorrelation comparison is an empirical simulation measurement, and the self-citations to MLNN/MoDe method papers are not load-bearing.

full rationale

The paper's central claim ('We thus deem the MoDe approach to be better suited to achieving lifetime-decorrelation in the topological b trigger') is supported by direct simulation measurements: the MLNNs are trained with two independent penalty terms (DisCo and MoDe), then their efficiencies are evaluated as functions of true b-hadron lifetime in B+→J/ψK+ and B0→D−π+ samples (Figs. 3-4, Table 1). No equation in the paper defines the measured slope in terms of the penalty term by construction; the slope is an empirical output of the trained model. The methods themselves (MLNNs, DisCo, MoDe) are taken as inputs from prior work, some of which has overlapping authors, but the paper does not rely on those citations to prove the decorrelation performance; it evaluates the models directly. There is no fitted parameter renamed as a prediction, no imported uniqueness theorem, no ansatz smuggled via citation that is load-bearing for the conclusion. The noted weaknesses—overlapping slope uncertainties for the MoDe-vs-DisCo comparison, unmatched λ working points, and the reliance on simulation—are statistical or external-validity concerns, not circularity. The derivation chain is therefore self-contained for the claims it makes.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central comparison rests on the fidelity of LHCb simulation, on the proxy relationship between reconstructed and true lifetime, and on the imported DisCo/MoDe implementations. It introduces no new entities; the only tuned numbers are penalty strengths and analysis ranges.

free parameters (5)
  • DisCo penalty strength lambda (2- and 3-body) = 0.8 for both 2- and 3-body models
    Chosen as the strongest decorrelation without significant reduction in classification power (App. B); a hand-selected hyperparameter affecting the decorrelation-efficiency trade-off.
  • MoDe penalty strength lambda (2- and 3-body) = 0.2 for both 2- and 3-body models
    Selected from precision/recall/ROC curves (Fig. 5); for the 3-body model lambda=0.1 had better metrics but was deemed too weak to decorrelate, so lambda=0.2 was used. This choice directly affects the headline comparison.
  • MoDe Legendre order ell = 2
    ell=2 is used based on prior work [32] and the expectation of a controlled quadratic dependence; not re-derived here.
  • Efficiency slope fit range = tau > 4 ps
    Slopes in Table 1 are estimated only for bins with true lifetime above 4 ps; the choice of range affects the slope values and the conclusion of flattening.
  • Trigger threshold operating points = Not specified; set to match output bandwidth
    Thresholds were chosen so each model returns the same output bandwidth, but exact values are not given. This affects absolute efficiencies, though not the large-lifetime trend.
assumptions (4)
  • domain assumption The LHCb simulation chain (Pythia, EvtGen, Photos, Geant4) faithfully models Run 3 pileup and the PV-misassociation background.
    All training and evaluation uses simulation (Secs. 2, 5); no data validation is shown.
  • domain assumption Decorrelation from the reconstructed/composite lifetime is sufficient to flatten efficiency versus true b-hadron lifetime.
    The penalty decorrelates the NN from candidate lifetime tau, while the evaluation bins on true B lifetime (Figs. 3-4). The transfer is assumed and tested only in simulation.
  • domain assumption The DisCo and MoDe penalties are correctly implemented according to Refs. [31,32].
    The paper imports these methods without re-deriving them; implementation fidelity is trusted.
  • domain assumption The monotonic Lipschitz NN architecture (Refs. [5-7]) trains stably with the additional loss terms.
    The paper states 'all converging on a minimised loss' (App. B) but provides no detailed convergence checks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decorrelation of neural networks from particle lifetimes in the LHCb topological $b$ trigger." pith.science (2026). https://pith.science/paper/4QHKXYEY

@misc{pith2026260722281,
  author       = {Pith},
  title        = {Pith review of: Decorrelation of neural networks from particle lifetimes in the LHCb topological $b$ trigger},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QHKXYEY}},
  note         = {Machine review of arXiv:2607.22281}
}
abstract

The LHCb topological beauty trigger is the primary set of algorithms for selecting collision events containing $b$-hadrons in the fully software-based LHCb trigger. The algorithms apply monotonic Lipschitz neural networks (NNs) to select vertices of charged particles consistent with the distinct topology of a $b$ decay, i.e., those with large lifetimes and transverse momentum. Many analyses of the events recorded require that the selection must be unbiased with respect to the $b$-hadron lifetime at large lifetimes. Accurate reconstruction is challenging in busier detector environments, in which several visible proton-proton collisions occur simultaneously per bunch crossing, such that misassociation of decay products can result in vertices with artificially large measured lifetimes. This paper presents two approaches to mitigate correlations between NN scores and candidate lifetimes at large lifetime, and evaluates the performance of the resulting models.

Figures

Figures reproduced from arXiv: 2607.22281 by the authors.

Figure 1
Figure 1. The composite lifetime and minimum log(χ 2 IP) of constituent tracks are plotted against one another for 2- and 3-body candidates in minimum bias back￾ground and B+→ J/ψ(µ +µ −)K+ signal training sam￾ples. Each are normalised according to the number of candidates in the sample. dependent studies, this results in a sparse population of large-lifetime candidates, reducing sensitivity to the cosh term of the b-hadron d… view at source ↗
Figure 2
Figure 2. Fraction of 2- and 3-body candidates in the back [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Efficiencies of (top) 2- and (bottom) 3- [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Performance of (left) 2- and (right) 3-body MLNNs, trained with DisCo and MoDe penalty terms of varying [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 6 canonical work pages

  1. [1]

    Alves Jr., et al., JINST3(LHCb-DP-2008-001), S08005 (2008)

    A.A. Alves Jr., et al., JINST3(LHCb-DP-2008-001), S08005 (2008). DOI 10.1088/1748-0221/3/08/S08005

    doi:10.1088/1748-0221/3/08/s08005 2008

  2. [2]

    Williams, V.V

    M. Williams, V.V. Gligorov, C. Thomas, H. Dijkstra, J. Nardulli, P. Spradlin, The HLT2 Topological Lines. Tech. rep., CERN, Geneva (2011). URLhttps://cds. cern.ch/record/1323557

    arXiv 2011

  3. [3]

    Gligorov, C

    V.V. Gligorov, C. Thomas, M. Williams, The HLT in- clusive B triggers. Tech. rep., CERN, Geneva (2011). URLhttps://cds.cern.ch/record/1384380. LHCb- INT-2011-030

    arXiv 2011

  4. [4]

    Aaij, et al., JINST14(LHCb-DP-2019-001), P04013 (2019)

    R. Aaij, et al., JINST14(LHCb-DP-2019-001), P04013 (2019). DOI 10.1088/1748-0221/14/04/P04013

    doi:10.1088/1748-0221/14/04/p04013 2019

  5. [5]

    Kitouni, N

    O. Kitouni, N. Nolte, M. Williams, Machine Learning: Science and Technology4(3), 035020 (2023). DOI 10. 1088/2632-2153/aced80. URLhttp://dx.doi.org/10. 1088/2632-2153/aced80

    2023

  6. [6]

    Schulte, B.R

    N. Schulte, B.R. Delaney, N. Nolte, G.M. Ciezarek, J. Al- brecht, M. Williams. Development of the topological trig- ger for lhcb run 3 (2023). URLhttps://arxiv.org/abs/ 2306.09873 8 Table 2: Input features of the two- and three-body MLNNs of the topologicalbtrigger. The MLNNs were required to increase monotonically in features marked with✓(∼) for all of R...

    arXiv 2023

  7. [7]

    Delaney, N

    B. Delaney, N. Schulte, G. Ciezarek, N. Nolte, M. Williams, J. Albrecht. Applications of lipschitz neural networks to the run 3 lhcb trigger system (2024). DOI 10.1051/epjconf/202429509005. URLhttps://doi.org/ 10.1051/epjconf/202429509005

    arXiv 2024

  8. [8]

    Aaij, et al., JINST19, P05065 (2024)

    R. Aaij, et al., JINST19, P05065 (2024). DOI 10.1088/ 1748-0221/19/05/P05065

    2024

Show all 33 references
  1. [9]

    Aaij, et al., JINST14, P04006 (2019)

    R. Aaij, et al., JINST14, P04006 (2019). DOI 10.1088/ 1748-0221/14/04/P04006

    2019

  2. [10]

    Aaij, et al., Comput

    R. Aaij, et al., Comput. Softw. Big Sci.4(1), 7 (2020). DOI 10.1007/s41781-020-00039-7

    2020 doi

  3. [11]

    Dujany, B

    G. Dujany, B. Storaci, J. Phys. Conf. Ser.664, 082010 (2015). DOI 10.1088/1742-6596/664/8/082010

    2015 doi

  4. [12]

    Mathad, M

    A. Mathad, M. Ferrillo, S. Barr´ e, P. Koppenburg, P. Owen, G. Raven, E. Rodrigues, N. Serra, Com- put. Softw. Big Sci.8(1), 6 (2024). DOI 10.1007/ s41781-024-00116-1

    2024

  5. [13]

    Abdelmotteleb, A

    A. Abdelmotteleb, A. Bertolin, C. Burr, B. Couturier, E. Eckstein, D. Fazzini, N. Grieser, C. Haen, R. O’Neil, E. Rodrigues, N. Skidmore, M. Smith, A.R. Wiederhold, S. Zhang, Computing and Software for Big Science9(1) (2025). DOI 10.1007/s41781-025-00144-5. URLhttp: //dx.doi.o...

    2025 doi

  6. [14]

    Abe, et al., Phys

    K. Abe, et al., Phys. Rev. Lett.80, 660 (1998). DOI 10.1103/PhysRevLett.80.660

    1998 doi

  7. [15]

    Gligorov, M

    V.V. Gligorov, M. Williams, JINST8, P02013 (2013). DOI 10.1088/1748-0221/8/02/P02013

    2013 doi

  8. [16]

    Likhomanenko, et al., J

    T. Likhomanenko, et al., J. Phys. Conf. Ser.664, 082025 (2015). DOI 10.1088/1742-6596/664/8/082025

    2015 doi

  9. [17]

    CERN-LHCC-2013-021, Geneva (2013)

    LHCb collaboration, LHCb VELO Upgrade Technical Design Report. CERN-LHCC-2013-021, Geneva (2013)

    2013

  10. [18]

    CERN-LHCC-2014-001, Geneva (2014)

    LHCb collaboration, LHCb Tracker Upgrade Technical Design Report. CERN-LHCC-2014-001, Geneva (2014)

    2014

  11. [19]

    Dziurda, et al., Eur

    A. Dziurda, et al., Eur. Phys. J. C85(6), 609 (2025). DOI 10.1140/epjc/s10052-025-14225-7

    2025 doi

  12. [20]

    Paszke, et al., in33rd Conference on Neural Informa- tion Processing Systems(2019)

    A. Paszke, et al., in33rd Conference on Neural Informa- tion Processing Systems(2019)

    2019

  13. [21]

    Kingma, J

    D.P. Kingma, J. Ba, in3rd International Conference on Learning Representations(2014)

    2014

  14. [22]

    Sj¨ ostrand, S

    T. Sj¨ ostrand, S. Mrenna, P. Skands, Comput. Phys. Com- mun.178, 852 (2008). DOI 10.1016/j.cpc.2008.01.036

    2008 doi

  15. [23]

    Sj¨ ostrand, S

    T. Sj¨ ostrand, S. Mrenna, P. Skands, JHEP05, 026 (2006). DOI 10.1088/1126-6708/2006/05/026

    2006 doi

  16. [24]

    Belyaev, et al., J

    I. Belyaev, et al., J. Phys. Conf. Ser.331, 032047 (2011). DOI 10.1088/1742-6596/331/3/032047

    2011 doi

  17. [25]

    Lange, Nucl

    D.J. Lange, Nucl. Instrum. Meth.A462, 152 (2001). DOI 10.1016/S0168-9002(01)00089-4

    2001 doi

  18. [26]

    Davidson, T

    N. Davidson, T. Przedzinski, Z. Was, Comp. Phys. Comm.199, 86 (2016). DOI https://doi.org/10.1016/ j.cpc.2015.09.013

    2016

  19. [27]

    Allison, K

    J. Allison, K. Amako, J. Apostolakis, H. Araujo, P. Dubois, et al., IEEE Trans.Nucl.Sci.53, 270 (2006). DOI 10.1109/TNS.2006.869826

    2006

  20. [28]

    Agostinelli, et al., Nucl

    S. Agostinelli, et al., Nucl. Instrum. Meth.A506, 250 (2003). DOI 10.1016/S0168-9002(03)01368-8

    2003 doi

  21. [29]

    Clemencic, et al., J

    M. Clemencic, et al., J. Phys. Conf. Ser.331, 032023 (2011). DOI 10.1088/1742-6596/331/3/032023

    2011 doi

  22. [30]

    Dunietz, J.L

    I. Dunietz, J.L. Rosner, Phys. Rev. D34, 1404 (1986). DOI 10.1103/PhysRevD.34.1404

    1986 doi

  23. [31]

    Kasieczka, D

    G. Kasieczka, D. Shih, Physical Review Letters125(12) (2020). DOI 10.1103/physrevlett.125.122001. URLhttp: //dx.doi.org/10.1103/PhysRevLett.125.122001 9 0.0 0.2 0.4 0.6 0.8 1.0 Decorrelation strength, 0.80 0.85 0.90 0.95 1.00Precision LHCb Simulation DisCo MoDe 0.0 0.2 0.4 0.6...

    2020 doi

  24. [32]

    Kitouni, B

    O. Kitouni, B. Nachman, C. Weisser, M. Williams, Jour- nal of High Energy Physics2021(4) (2021). DOI 10.1007/jhep04(2021)070. URLhttp://dx.doi.org/10. 1007/JHEP04(2021)070

    2021 doi

  25. [33]

    Sz´ ekely, M.L

    G.J. Sz´ ekely, M.L. Rizzo, N.K. Bakirov, The An- nals of Statistics35(6), 2769 (2007). DOI 10.1214/ 009053607000000505. URLhttps://doi.org/10.1214/ 009053607000000505

    2007

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.