Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

An Automated Bandwidth Division for the LHCb Upgrade Trigger

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An automated pseudo-$\chi^2$ optimizer divides LHCb's 1 MHz HLT1 budget across 80 physics channels, raising average signal efficiency by roughly 30% for beauty and 70% for charm and semileptonic decays while saving 200 kHz of output rate…

desk verdict A genuine engineering contribution—first automated bandwidth division for a full software trigger—but the headline efficiency gains are in-sample and lack a data-taking closure test; worth refereeing with requests for validation. read the letter →

arxiv 2502.09557 v2 pith:JEDCENYU submitted 2025-02-13 hep-ex

T. Evans , C. Fitzpatrick , J. Horswill This is my paper · ORCID
classification hep-ex
keywords LHCbHLT1bandwidthdivisiontriggeroptimisationpseudo-chi2figureofmeritAdamoptimizersoftwaresignalefficiency
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 reports an automated tool that divides the LHCb HLT1 trigger bandwidth equitably across the collaboration's physics programme. It does this by minimizing a pseudo-$\chi^2$ figure of merit that balances per-channel signal efficiency against a hard cap on total output rate. The automatically tuned thresholds, already used to collect 2024 data, improve average efficiency by about 30% for beauty channels and 70% for charm and semileptonic channels compared with the previous default thresholds, while using 200 kHz less of the 1 MHz budget. The work matters because LHCb's fully software trigger has hundreds of tunable selections, and manual tuning cannot keep pace with changing run conditions.

What carries the argument

The central object is the pseudo-$\chi^2$ figure of merit $\chi^2_{\mathrm{global}}(\mathbf{x}) = \sum_i \omega_i \bigl(1 - \epsilon_i(\mathbf{x})/\epsilon_i^{\max}\bigr)^2$, where $\epsilon_i(\mathbf{x})$ is the rate-penalized efficiency of channel $i$ and $\epsilon_i^{\max}$ is the best efficiency that channel could achieve if given the entire bandwidth. The rate penalty divides the signal efficiency by $\mathrm{OR}_{\mathrm{limit}}/\mathrm{OR}$ whenever the predicted output rate exceeds the cap, so the optimizer never simply loosens every line. The minimization runs on a discrete grid, first with the Adam gradient-based optimizer on a continuous relaxation, then a recursive search over neighboring discrete thresholds to find the best grid point. Warm restarts with scaled momentum are used to escape local minima, and the evaluation is parallelized over threads to keep a full retuning within minutes.

What would settle it

Compare the tool's predicted per-line output rates and per-channel efficiencies with the actual HLT1 decisions recorded during 2024 running: if the measured total output rate at nominal luminosity exceeds the 1 MHz cap, or if the measured average retention for a category falls well below the reported 93%/79%/97%/90% values, then the bandwidth division is not equitable in real data.

Watch

Extended reading notes

Core claim

The paper claims that a single ensemble of 16 tunable HLT1 thresholds, found by minimizing a rate-penalized pseudo-$\chi^2$ over 80 characteristic signal channels and 35 trigger lines, achieves near-optimal physics retention across the entire menu. Relative to the default manual thresholds, the optimized thresholds yield an average increase in signal efficiency of roughly 30% for beauty and 70% for charm and semileptonic channels, at a saving of 200 kHz in output rate. Relative to each channel's individually maximized efficiency, the optimized single working point retains on average 93% for beauty, 79% for charm, 97% for electroweak, and 90% for semileptonic channels. These thresholds were used to collect data in HLT1 during 2024.

Load-bearing premise

The whole division rests on the 80 analyst-chosen channels, each weighted equally, being a fair stand-in for LHCb's full physics programme, and on simulation efficiencies plus a 9-million-event minimally biased sample accurately predicting real trigger rates.

Editorial extensions

If this is right

  • At a fixed 1 MHz budget, LHCb analyses retain roughly 30% more beauty signal and 70% more charm and semileptonic signal than they would with the previous default thresholds.
  • The optimized thresholds save 200 kHz of output rate at the same or better efficiency, giving the 1 MHz budget headroom for looser inclusive lines or changed running conditions.
  • Because retuning takes only minutes, the tool can keep up with changes in luminosity, trigger menu, or the set of physics channels, which was not practical with the older method.
  • The thresholds chosen by the tool were actually used during 2024 data taking, so the claimed improvement is already encoded in recorded data rather than being a purely simulated projection.
  • Divisions at multiple output-rate limits (0.8, 1.0, 1.2 MHz) allow the collaboration to interpolate as computing throughput and buffer capacity evolve.

Reading between the lines

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

  • An implication the authors leave implicit is that the same pseudo-$\chi^2$ machinery could divide bandwidth at HLT2 or in other fully software triggers, since its inputs are only per-channel efficiencies and a total rate cap.
  • The equal channel weights ($\omega_i = 1$) are a policy choice; a collaboration that re-weighted toward rarer channels would systematically loosen the lines serving those channels, which the paper notes but does not explore.
  • A testable extension is to rerun the tool using actual recorded trigger decisions from 2024 data instead of the 9-million-event minimally biased sample, then compare predicted and observed per-line output rates; agreement would validate the sample as a rate predictor.
  • The paper finds Adam's advantage over a genetic algorithm shrinks as the number of tuned thresholds grows, so for even larger future trigger menus a hybrid optimizer may be worth investigating.
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

3 major / 5 minor

Summary. The manuscript describes an automated tool for dividing the 1 MHz output-rate budget of the LHCb HLT1 software trigger among the experiment's physics channels. The tool minimizes a pseudo-chi2 objective that combines per-channel signal efficiencies, estimated from simulated samples, with the total HLT1 output rate, estimated from a 9M-event minimally biased data sample. The optimization uses an adapted Adam algorithm on a discrete grid, with a warm-restart mechanism, and is compared with a genetic algorithm. The authors report that the automatically tuned thresholds yield average efficiency increases of about 30% for beauty channels and 70% for charm and semileptonic channels, at a 200 kHz output-rate saving relative to manual defaults, and state that the tuned thresholds were used for 2024 data taking.

Significance. If the reported gains are realized in production, the tool is a valuable and generalizable method for trigger bandwidth division in fully software-based triggers. The optimization formulation is clearly defined, the comparison with a genetic algorithm is useful, and the paper demonstrates practical integration with the Allen trigger framework. The principal caveat is that the headline improvements are point estimates from the same simulated and minimally biased samples used for tuning, without uncertainties or a comparison with actual 2024 data-taking conditions. Adding a closure test or systematic uncertainty estimate would substantially strengthen the central claim. The paper does not release code or data, which limits independent reproducibility, but the algorithmic description is sufficiently detailed for reimplementation.

major comments (3)
  1. [Section 3] The headline results (average efficiency increase of ~30% for beauty and ~70% for charm/semileptonic channels at a 200 kHz OR saving) are quoted as exact values, but they are point estimates from the same simulated signal samples and the same 9M-event OR sample used for the optimization. No statistical or systematic uncertainty is given for the efficiencies or for the OR, and no comparison is made with actual HLT1 rates or data-derived trigger efficiencies under the 2024 thresholds mentioned in Section 3. A closure test comparing predicted and measured OR and per-channel efficiencies, or at minimum an explicit enumeration of neglected uncertainties, is needed to establish that the gains are not simulation artefacts.
  2. [Section 3 and Figs. 9-11] The baseline for the central comparison is missing. The figures plot the optimized ('optimal') and per-channel maximum efficiencies, but not the efficiencies obtained with the default thresholds. The reader therefore cannot verify the claimed ~30% and ~70% improvements or the 200 kHz OR saving from the displayed data. Please add a table or an overlay with the default-threshold efficiencies and OR, and state explicitly whether the comparison is made at a common OR limit or at each threshold set's own OR.
  3. [Section 2.1, Eqs. (3)-(4)] The pseudo-chi2 objective treats the 80 channel weights as fixed and equal (omega_i = 1), and the paper does not assess how sensitive the optimized thresholds are to the choice of channel ensemble or weights. Since the 'equitable division' claim depends on this ensemble representing the full physics programme, a short sensitivity study (for example, varying a subset of weights or adding/removing a few channels) would materially support the claim. At minimum, this limitation should be stated explicitly rather than implied by the phrase 'chosen carefully'.
minor comments (5)
  1. [Section 2.2] The sentence 'Adam outperformed the GA by at least an order of magnitude in run time' appears to refer to an earlier five-threshold configuration, while Fig. 5 (with 16 thresholds) shows that the GA is on average about 30% faster. Please clarify which comparison is being reported and avoid mixing the two configurations.
  2. [Section 3] The sentence 'Between these rate limits, there is an average increase in efficiency of around 30%' is ambiguous; it should state the comparator (for example, default thresholds at the same rate limit) and the averaging procedure over channels.
  3. [Figure captions] Several figure captions contain typesetting artifacts: 'MV A' should be 'MVA' and 'Ge V /c' should be 'GeV/c'.
  4. [Section 1.2] The abstract and introduction mention O(100) HLT1 trigger selections, while Section 2.1 states that 35 trigger lines are optimized. Please reconcile these numbers or clarify that 35 is the tuned subset.
  5. [Section 2.1] The grid step sizes are said to be chosen to ensure statistical significance, but no numerical values or derivation are given. A one-sentence justification or a reference would help the reader assess the discretization choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the tuned thresholds are optimized against an external pseudo-chi2 loss and benchmarked against manually chosen defaults, not against the optimization target itself.

full rationale

The paper's derivation chain is self-contained. Section 2.1 defines a pseudo-chi2 figure of merit, chi2_global = sum omega_i (1 - eps_i/eps_i^max)^2, minimized over a fixed set of HLT1 thresholds using Adam and a discrete grid search. The headline improvements are comparisons between the optimizer's output and manually chosen default thresholds, so the result is not an identity or a fitted quantity renamed as a prediction. The normalization eps_i^max is obtained by a separate per-channel maximization, not by the global solution, so the reported ~93% retention is a genuine ratio rather than a tautology. The only self-citations (Allen, RTA model papers) support the reconstruction infrastructure and do not carry the optimization claim. The absence of a closure test against actual 2024 HLT1 rates is a validation limitation, not circularity: the simulation and minimally biased samples are independent inputs rather than outputs of the fit. In-sample evaluation is an overfitting concern, not a logical circularity, and no step reduces to its own input by construction.

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

The central claim rests on the choice of representative channels and weights, on the fidelity of simulation and short minimally-biased data samples to real running conditions, and on the heuristic optimizer's ability to find good minima. No new physical entities are introduced; the free parameters are engineering choices internal to the tool.

free parameters (4)
  • 16 floating HLT1 thresholds x = not reported
    The optimized trigger thresholds (pT, IP, and chi2_IP cuts for TrackMVA, TwoTrackMVA, and other lines) are the central tuning variables; the paper does not list their optimal values.
  • Channel weights omega_i = 1 (usual)
    Weights in chi2_global are set by LHCb analysts to represent physics priorities; omega_i=1 is an arbitrary default that determines the equitable division.
  • Grid step sizes for discretization = not quantified
    Thresholds are truncated to a discrete grid with step sizes chosen to be statistically significant, but the actual steps and statistical criteria are not specified; they affect which local minimum is found.
  • Warm-restart moment scaling factor = one order of magnitude
    Adam moments are heuristically scaled by 10x in the first epoch and during warm restarts to escape local minima; this hyperparameter is not derived from first principles.
assumptions (5)
  • domain assumption The 80 chosen channels with unit weights represent the full LHCb physics programme
    Section 2.1 states the channels and weightings 'were chosen carefully to best represent the physics interests of the collaboration'; if this ensemble is unrepresentative, the equitable division is biased.
  • domain assumption Simulated signal samples with reconstructibility requirements and a 9-million-event minimally biased sample predict real HLT1 rates and efficiencies
    Section 1.3 and Section 2.1 use about 9 million minimally biased events (about 0.5 seconds of 2024 data) to estimate OR and simulated samples for efficiencies; statistical and systematic stability of these estimates is assumed.
  • domain assumption The pseudo-chi2 with rate-penalty epsilon_i(x) is a valid proxy for physics retention
    Section 2.1 defines chi2_global and the ORlimit/OR penalty; the form of the penalty is chosen by the authors and is not derived from physics requirements.
  • domain assumption Adam converges to a good minimum on the discrete threshold grid
    Sections 2.2 and 2.3 assume gradient-based Adam with warm restarts finds near-optimal thresholds; no guarantee or benchmark against global search is provided.
  • domain assumption Efficiency and OR curves are smooth functions of thresholds
    Section 1.3 states smoothness 'significantly increases the convergence rate' but it is an empirical assumption about the response functions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Automated Bandwidth Division for the LHCb Upgrade Trigger." pith.science (2026). https://pith.science/paper/JEDCENYU

@misc{pith2026250209557,
  author       = {Pith},
  title        = {Pith review of: An Automated Bandwidth Division for the LHCb Upgrade Trigger},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEDCENYU}},
  note         = {Machine review of arXiv:2502.09557}
}
abstract

The upgraded Large Hadron Collider beauty (LHCb) experiment is the first detector based at a hadron collider using a fully software based trigger. The first `High Level Trigger' stage (HLT1) reduces the event rate from 30 MHz to approximately 1 MHz based on reconstruction criteria from the tracking system, and consists of $\mathcal{O}(100)$ trigger selections implemented on Graphics Processing Units (GPUs). These selections are further refined following the full offline-quality reconstruction at the second stage (HLT2) prior to saving for analysis. An automated bandwidth division has been performed to equitably divide this 1 MHz HLT1 Output Rate (OR) between the signals of interest to the LHCb physics programme. This was achieved by optimising a set of trigger selections that maximise efficiency for signals of interest to LHCb while keeping the total HLT1 readout capped to a maximum. The bandwidth division tool has been used to determine the optimal selection for 35 selection algorithms over 80 characteristic physics channels.

Figures

Figures reproduced from arXiv: 2502.09557 by the authors.

Figure 1
Figure 1. Online LHCb dataflow [4] with ORs obtained from LHCb technical and computing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. HLT1 ORs of the single (left) and two (right) track MVA lines as a function of their [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. B0 → D+D− efficiencies of the single (left) and two (right) track MVA lines as a function of their input thresholds. The TrackMVA algorithm performs a preselection of tracks to remove the most obvious fakes, and then applies a hyperbolic selection criterion to a two-dimensional plane of pT and χ 2 IP. This method captures most of the signal background discrimination and is simple to implement. The TwoTrackMVA algori… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Flow chart of the Adam algorithm adapted to search for the best set of truncated [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: A plot of the discrete χ 2 global vs. algorithm runtime for a number of Adam and GA minimisations starting with different seeds. Adam is computationally efficient and is well-suited to problems with large datasets, many parameters, and noisy gradients. All three of the…
Figure 6
Figure 6. Figure 6: Top - A χ 2 indiv scan for B0 → D+D−. Bottom - An analogous minimisation path. ∼ 70% smaller than that of Adam. The average GA runtime is also ∼ 30% shorter than the average Adam runtime. Adam is preferred due to the need to carefully choose the GA hyperparameters. Ano…
Figure 7
Figure 7. Figure 7: Minimisation of χ 2 indiv for B0 s → ϕϕ. After the first Adam minimisation, a ‘warm-restart’ mechanism is initiated. A warm restart involves beginning a second minimisation at the thresholds chosen by the first minimisation with the learning rate for each parameter res…
Figure 8
Figure 8. Figure 8: A demonstration of the ‘warm restart’ mechanism halfway through the minimisation [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: The optimal (blue) and maximum (red) efficiencies for each signal channel produced [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: The optimal (blue) and maximum (red) efficiencies for each signal channel produced [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: The optimal (blue) and maximum (red) efficiencies for each signal channel produced [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Track fitting at the full LHC collision rate

    hep-ex 2026-07 conditional novelty 5.0 of 10

    A parameterised Kalman filter on GPUs fits full tracks in LHCb's HLT1 trigger at 30 MHz, halving reconstructed D0 and J/ψ mass widths and more than halving ghost-track efficiency.

Reference graph

Works this paper leans on

21 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    LHCb Trigger and Online Upgrade Technical Design Report,

    The LHC Experiments Committee, “LHCb Trigger and Online Upgrade Technical Design Report,” tech. rep., CERN, May 2014. https://cds.cern.ch/record/ 1701361

    work page 2014

  2. [2]

    The LHCb Upgrade I,

    R. Aaij, S. Akar, and J. A. et al, “The LHCb Upgrade I,” JINST, vol. 19, no. 05, p. P05065, 2024. https://cds.cern.ch/record/2859353. 14

    work page arXiv 2024

  3. [3]

    Framework TDR for the LHCb Upgrade: Technical Design Report,

    I. Bediaga, H. Chanal, P. Hopchev, S. Cadeddu, S. Stoica, M. Calvo Gomez, S. T’Jampens, I. Machikhiliyan, Z. Guzik, A. Alves Jr, et al., “Framework TDR for the LHCb Upgrade: Technical Design Report,” tech. rep., LHCb-TDR-012, 2012. https://cds.cern.ch/record/1443882/files/LHCB-TDR-012.pdf

    work page arXiv 2012

  4. [4]

    RTA and DPA Dataflow Diagrams for Run 1, Run 2, and the Upgraded LHCb Detector,

    V. V. Gligorov and E. Rodrigues, “RTA and DPA Dataflow Diagrams for Run 1, Run 2, and the Upgraded LHCb Detector,” November 2020. http://cds.cern.ch/ record/2730181/

    work page arXiv 2020

  5. [5]

    Computing Model of the Upgrade LHCb experi- ment,

    The LHC Experiments Committee, “Computing Model of the Upgrade LHCb experi- ment,” tech. rep., CERN, Geneva, 2018. https://cds.cern.ch/record/2319756

    work page arXiv 2018

  6. [6]

    Allen: A High-Level Trigger on GPUs for LHCb,

    R. Aaij, J. Albrecht, and e. a. M. Belous, “Allen: A High-Level Trigger on GPUs for LHCb,” Computing and Software for Big Science, vol. 4, p. 7, April 2020. https: //doi.org/10.1007/s41781-020-00039-7

    doi:10.1007/s41781-020-00039-7 2020

  7. [7]

    A Comprehensive Real-time Analysis Model at the LHCb Experiment,

    R. Aaij, S. Benson, M. De Cian, A. Dziurda, C. Fitzpatrick, E. Govorkova, O. Lupton, R. Matev, S. Neubert, A. Pearce, H. Schreiner, S. Stahl, and M. Vesterinen, “A Comprehensive Real-time Analysis Model at the LHCb Experiment,” JINST, vol. 14, no. 04, p. P04006, 2019

    work page 2019

  8. [8]

    Allen Documentation

    R. Aaij and D. Vom Bruch, “Allen Documentation.” https://allen-doc.docs. cern.ch/, 2021

    work page 2021

Show all 21 references
  1. [9]

    Looking Forward: A High-Throughput Track Following Algorithm for Parallel Architectures,

    A. Bailly-Reyre, L. Bian, P. Billoir, D. H. C. P´ erez, V. V. Gligorov, F. Pisani, R. Quagliani, A. Scarabotto, and D. V. Bruch, “Looking Forward: A High-Throughput Track Following Algorithm for Parallel Architectures,” IEEE Access, vol. 12, pp. 114198–114211, 2024. http://dx....

    2024

  2. [10]

    Design and Performance of the LHCb Trigger and Full Real-time Reconstruction in Run 2 of the LHC,

    R. Aaij, S. Akar, and J. A. et al, “Design and Performance of the LHCb Trigger and Full Real-time Reconstruction in Run 2 of the LHC,” Journal of Instrumentation, vol. 14, pp. P04013–P04013, apr 2019. https://doi.org/10.1088%2F1748-0221% 2F14%2F04%2Fp04013

    2019

  3. [11]

    A Comprehensive Real-time Analysis Model at the LHCb Experiment,

    R. Aaij, S. Benson, M. D. Cian, A. Dziurda, C. Fitzpatrick, E. Govorkova, O. Lupton, R. Matev, S. Neubert, A. Pearce, H. Schreiner, S. Stahl, and M. Vesterinen, “A Comprehensive Real-time Analysis Model at the LHCb Experiment,” Journal of Instrumentation, vol. 14, pp. P04006–P...

    2019

  4. [12]

    ROOT - An Object Oriented Data Analysis Framework

    R. Brun and F. Rademakers, “ROOT - An Object Oriented Data Analysis Framework.” https://github.com/root-project/root/, 1997

    1997

  5. [13]

    OpenMP: an Industry Standard API for Shared-memory Programming,

    L. Dagum and R. Menon, “OpenMP: an Industry Standard API for Shared-memory Programming,” Computational Science & Engineering, IEEE, vol. 5, no. 1, pp. 46–55,

  6. [14]

    Mitchell, An Introduction to Genetic Algorithms

    M. Mitchell, An Introduction to Genetic Algorithms. The MIT Press, 1996. https: //doi.org/10.7551/mitpress/3927.001.0001

    1996 doi

  7. [15]

    Ketkar, Stochastic Gradient Descent, pp

    N. Ketkar, Stochastic Gradient Descent, pp. 113–132. Berkeley, CA: Apress, 2017. https://doi.org/10.1007/978-1-4842-2766-4_8 . 15

    2017 doi

  8. [16]

    Draw.io Online Diagram Editor,

    JGraph, “Draw.io Online Diagram Editor,” October 2021. https://www.diagrams. net/

    2021

  9. [17]

    MINUIT User’s Guide,

    F. James and M. Winkler, “MINUIT User’s Guide,” June 2004.https://inspirehep. net/literature/1258345

    2004

  10. [18]

    Adam: A Method for Stochastic Optimization,

    D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” in 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings(Y. Bengio and Y. LeCun, eds.),

    2015

  11. [19]

    Decoupled Weight Decay Regularization,

    I. Loshchilov and F. Hutter, “Decoupled Weight Decay Regularization,” in International Conference on Learning Representations , 2017. https://api. semanticscholar.org/CorpusID:53592270. 16

    2017

  12. [1998]

    https://doi.org/10.1109/99.660313

    doi

  13. [2015]

    https://api.semanticscholar.org/CorpusID:6628106

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.