Track fitting at the full LHC collision rate

Lennart H. Uecker; Michel De Cian; Pierre Billoir; Thomas Boettcher

REVIEW 3 major objections 5 minor 26 references

A full-track Kalman fitter now runs at the 30 MHz LHCb collision rate by replacing magnetic-field integration with analytic simulation-fitted functions, roughly halving D0 and J/psi mass widths.

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

2026-08-02 01:02 UTC pith:KISFBRGV

load-bearing objection A real production engineering milestone — full-track Kalman fitting on LHCb GPUs at 30 MHz — but the physics claims all rest on simulation and the parameters are not released. the 3 major comments →

arxiv 2607.14793 v2 pith:KISFBRGV submitted 2026-07-16 hep-ex physics.ins-det

Track fitting at the full LHC collision rate

Pierre Billoir , Thomas Boettcher , Michel De Cian , Lennart H. Uecker This is my paper

classification hep-ex physics.ins-det

keywords Kalman filtertrack fittinghigh-level triggerGPULHCbparameterised propagationmomentum resolutionghost tracks

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 reports the first Kalman-based full-track fit to run in production at the LHCb collision rate of 30 MHz, using about 500 GPUs in the first trigger stage. It makes this feasible by replacing the two expensive parts of a normal Kalman fit — numerical extrapolation through the magnetic field and per-step multiple-scattering noise — with analytic parameterised functions fitted to simulation. The filter uses hits from all three tracking detectors rather than only the vertex detector, at a cost of about 2% more processing time per event. On simulated samples the gain is a roughly factor-of-two improvement in D0 and J/psi invariant mass resolution, ghost-track efficiency dropping from 70% to 29% with slightly better genuine-track efficiency, and full recovery of momentum resolution under injected detector misalignment.

Core claim

The central claim is that a Kalman filter whose physics model is parameterised analytically can deliver full-track fits at the full LHC collision rate with accuracy close to the offline-quality reconstruction. The Kalman machinery — predictions, residuals, gains, covariance propagation, chi-squared — is unchanged; only the prediction step is replaced by closed-form functions whose parameters are derived from simulation. The paper shows that on simulated signal events this Parameterised Kalman Filter achieves momentum resolution close to the HLT2 offline fit, narrows the core Gaussian widths of the D0 mass peak from 10.3 to 5.8 MeV/c2 and the J/psi mass peak from 21.4 to 10.8 MeV/c2, reduces

What carries the argument

The load-bearing machinery is a set of vector-valued extrapolation functions f_{X->Y}, one for each pair of detector layers, together with a Gaussian multiple-scattering matrix Q. Each function is a closed-form expression in the state variables (position, slopes, charge/momentum) with coefficients fixed by minimising a Gaussian likelihood against high-precision Runge-Kutta extrapolations for propagation and against simulated true states for the noise parameters. The Jacobians of these functions are computed analytically so covariances propagate exactly as in a standard Kalman filter. The full parameter set comprises 240,466 floating-point values, with only 946 read per thread; because most o

Load-bearing premise

The load-bearing premise is that the simulation used to train the 240,466 parameters reproduces the real detector's material, field, and hit resolution closely enough that the parameterisation error remains below the physical multiple-scattering uncertainty; all quoted gains are computed in that same simulation, with no real-data validation presented.

What would settle it

Compare the D0 and J/psi invariant mass widths and the momentum-resolution curve obtained from real 2025 collision events with the simulation predictions for the same trigger lines; if the real-data widths do not approach the simulated factor-of-two improvement over the VELO-only fitter, the central claim is disconfirmed.

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

If this is right

HLT1 trigger lines can use roughly half the invariant-mass window for D0 and J/psi selections, reducing background while preserving signal.
The first trigger stage gets momentum estimates close to the offline-quality reconstruction, tightening kinematic and vertexing selections that previously had to wait for HLT2.
The improved chi-squared separation between real and ghost tracks lets the trigger operate at a working point that keeps more genuine tracks while removing more combinatorial background.
The fitter's tolerance of SciFi misalignment makes momentum resolution—and therefore trigger selection—much more stable over a long data-taking period.
Because the processing-time increase is only about 2%, the full 30 MHz collision rate remains reachable on the existing ~500-GPU farm, so this upgrade does not trade physics capability for throughput.

Where Pith is reading between the lines

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

If the approach generalises, the practical bottleneck shifts from GPU throughput to simulation fidelity: any detector whose material and field are accurately simulable could use the same parameterised filter in a real-time trigger.
The compressed 240k-parameter representation of field and material could be repurposed as a monitoring diagnostic—comparing parameters refitted at different times would reveal slow changes in detector material or magnetic field.
Since the current implementation omits outlier removal, adding a cheap outlier step while preserving warp-uniform memory access is a plausible next gain, likely recovering tracks that currently fail the fit.
The absence of real-data validation in the paper means the strongest test is not another simulation but the collision data already being recorded since 2025; a direct width comparison would settle transferability.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A real production engineering milestone — full-track Kalman fitting on LHCb GPUs at 30 MHz — but the physics claims all rest on simulation and the parameters are not released.

read the letter

You should know this is a production engineering paper, not a new-algorithm paper. The parameterised-propagation formalism is from Ref. [23] (two of the same authors) and the paper credits it. What is genuinely new is the GPU implementation in Allen, the specific choices for constant memory, single-precision arithmetic, and warp-synchronized loops, and the performance comparison against the VELO-only fitter it replaces.

The production story is the strongest part. The algorithm has run in HLT1 since the start of 2025 data-taking, the throughput measurement in Fig. 3 is real GPU processing, and the 2% cost in full HLT1 time is a concrete, believable number. The physics gains — roughly factor two in D0 and J/ψ mass widths, ghost efficiency down from 70% to 29% at a working point with higher genuine efficiency — are exactly what you would expect from adding UT and SciFi hits to a fit that previously used only the VELO. So the qualitative conclusions are not surprising; they are physically sound.

The soft spots are mostly about validation. Every performance number in §6 comes from the same Geant4 simulation used to fit the 240,466 parameters. The paper says the algorithm is deployed on real data since 2025, so a real-data closure test is feasible — and its absence is conspicuous. The mass widths and ghost efficiencies also carry no statistical uncertainties, which is sloppy for a paper making factor-of-two claims. And the 240k fitted parameter values are not shipped, so the model's fitted content is not independently reproducible; at minimum a commit hash or a parameter file should be provided.

The misalignment-robustness claim is the one place where I would push back. The paper shows full recovery of momentum resolution under injected SciFi misalignment, but gives no mechanism. Random module shifts should bias the measured bend, and a fit that uses those hits should see that bias. The result is not obviously wrong, but it needs more than a plot: a residual distribution versus misalignment magnitude, or an explanation of how the fit absorbs the shifts.

Who gets value from this: anyone working on real-time tracking, GPU reconstruction, or Kalman-filter approximations. The paper deserves a serious referee, and I would send it to review rather than desk-reject. But the referee should insist on real-data validation or an explicit pointer to an LHCb note with it, plus uncertainties and release of the parameter artifacts. With those, this is a solid contribution; without them, the headline physics claims remain unproven in real running conditions.

Referee Report

3 major / 5 minor

Summary. The paper presents the Parameterised Kalman Filter, a track-fitting algorithm deployed in LHCb's first-level trigger (HLT1) since the start of 2025 data-taking. It replaces numerical magnetic-field extrapolation with analytical parameterisations fitted to Geant4 simulation, while retaining the Kalman-filter formalism. The filter uses hits from VELO, UT, and SciFi, runs on approximately 500 GPUs at the full 30 MHz collision rate, and is compared with the previous VELO-only Kalman filter. The reported gains include a factor of two improvement in D0 and J/psi invariant mass resolutions, improved ghost-track rejection (ghost efficiency 70% to 29% at equal or better genuine efficiency), recovery of momentum resolution under injected SciFi misalignment, and only a roughly 2% increase in HLT1 processing time. All performance figures are obtained from simulation.

Significance. If the quantitative claims are confirmed on data, this is a significant result: it is the first full-track Kalman filter running at the full LHC collision rate, and it demonstrates that offline-quality track parameters are feasible in a real-time trigger. The engineering contribution is concrete and well described: parameterised extrapolations, GPU memory layout, constant-memory usage, single-precision arithmetic, and a production-level throughput measurement. The direction of the physics improvements is physically expected and internally consistent. The main weakness is that the quantitative performance claims rest entirely on the same simulation framework used to train the 240,466 parameters, with no real-data validation shown; the paper's significance is therefore conditional on such validation or on a clear re-scoping as a simulation-only study.

major comments (3)

[§6, Figs. 4–6] All performance metrics supporting the factor-of-two claim (mass widths in Fig. 6, ghost efficiencies in Fig. 5, misalignment robustness in Fig. 4) are computed from Geant4 simulation, and the same simulation framework is used to fit the 240,466 extrapolation and scattering parameters via Eqs. (12)–(13). This is a closed loop: the quoted widths (21.4→10.8 MeV/c2 for J/ψ, 10.3→5.8 MeV/c2 for D0) and ghost efficiencies (70%→29%) validate the parameterisation against the training simulation, not against the real detector. Since the fitter is deployed in 2025 data-taking, a comparison on collision data using the identical HLT1 lines is feasible and should be provided. Without it, the central performance claim is not established for the real detector.
[§4] The correctness condition is stated as δpar (parameterisation uncertainty) smaller than the physical state uncertainty from multiple scattering and detector resolution, but no quantitative evidence is given. The paper only asserts in §5 that the required accuracy is achieved. A plot of δpar versus z or momentum, together with a comparison to Q and detector-resolution terms, is needed for each extrapolation function. This is especially important for the UT-to-SciFi extrapolation, which contains 240,000 of the 240,466 parameters and dominates the parameter set.
[§5, Fig. 3] The throughput measurement is described as 'emulating the production environment' with 10 runs. Please state explicitly whether the input workload is simulated events or real 2024/2025 data. The '30 MHz' operation claim and the '2% increase in processing time' statement in the abstract and conclusion depend on the representativeness of this workload. If the measurement is simulation-based, this should be disclosed and the caveat reflected in the conclusions.

minor comments (5)

[Title page] Typo in the title: 'T rack fitting' should be 'Track fitting'.
[Eq. (9)] The shorthands f_tx and f_ty are used before being defined. Please define them explicitly and provide the full functional form or a clear reference to the appendix/Ref. [23].
[§4, Eq. (10)–(11)] The multiple-scattering model assumes Gaussian scattering θ ~ N(0, α/p). This is a reasonable approximation for hadrons and muons, but for electrons bremsstrahlung is a hard, non-Gaussian process. The paper should state explicitly that the quoted performance applies to hadrons/muons and discuss expected degradation for electrons, or note it as a limitation.
[§5, Eq. (13)] The empirical constant c in Eq. (13) is mentioned but its role is unclear. A sentence explaining why it is needed and how its value is chosen would improve reproducibility.
[Fig. 5] The ROC curves are compared via AUC values (0.576 vs 0.887) without uncertainties. Since the same simulated events are used, a bootstrap estimate of the AUC uncertainty would help assess whether the difference is statistically significant, though the separation appears large.

Circularity Check

0 steps flagged

No circularity: fitted simulation parameters do not encode the reported performance metrics.

full rationale

The paper's derivation chain is: choose empirical parameterisations for track extrapolation and multiple-scattering noise (Eqs. 9-11), fit their parameters to simulated Runge-Kutta extrapolations and simulated true states (Eqs. 12-13), then run a Kalman filter using those parameterisations and compare the resulting momentum resolution, mass widths, ghost rejection, and misalignment robustness against the previous VELO-only fitter (Section 6). None of the reported performance metrics is an objective function in the parameter fit; the parameters are optimised only to reproduce the extrapolation and scattering distributions, not to improve J/psi or D0 mass resolution or ghost efficiency. The mass-resolution and ghost-rejection improvements therefore are not forced by construction. The functional forms are attributed to Ref. [23], a prior paper with overlapping authors, but the present paper does not use that citation as proof of the performance claims; it supplies the relevant forms and evaluates them on simulation. The main caveat, explicitly acknowledged in the paper, is that the required parameterisation accuracy delta_par is only asserted to be smaller than the multiple-scattering uncertainty ('The required accuracy better than the expected MS contribution is achieved for all extrapolation functions', Section 5), and the entire performance evaluation is simulation-based with no real-data validation shown. That is an external-validity/validation concern, not circularity: no step reduces a predicted quantity to a fitted input by definition. Hence no significant circularity is found.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central demonstration is purchased with a very large fitted parameter set (240,466 numbers) trained on a single simulation, plus the standard Kalman and Gaussian-scattering assumptions. The reader is not asked to accept new physics entities, but is asked to accept that a simulation-trained empirical model transfers to the deployed detector — the paper's own text asserts the accuracy requirement (δpar below physical uncertainty) without displaying the evidence.

free parameters (3)

Extrapolation parameters p^i_{X→Y} per layer/region = 240,466 floating-point values (240,000 for UT→SciFi); not tabulated
Fitted by minimising Eq. 12 against a precise Runge-Kutta extrapolation computed in simulation (§5); these empirical functions replace field-map integration and define the propagation the filter relies on.
Multiple-scattering parameters p̃0–p̃5 per step = not quoted in paper
Extracted by minimising Eq. 13 against simulated true states with scattering; six per step (four in VELO). They set the Q matrix, i.e. the filter's uncertainty model.
Empirical constant c = not given
Added 'to help with convergence' of the scattering-parameter fit (Eq. 13, §5); an extra tuning knob in the training procedure.

axioms (6)

standard math Linearised Kalman filter with Gaussian measurement noise is the correct estimator for track parameters
Eqs. 1–8 (§3); standard and unobjectionable, but the filter's optimality is only approximate at this precision.
domain assumption Multiple-scattering angle is Gaussian N(0, α/p) with per-step fitted width
§4, Eq. 10; central-limit approximation for material scattering; the exponents and correlations are fitted, and the real scattering distribution has non-Gaussian tails.
domain assumption Energy loss can be absorbed into the parameterisation: f_{q/p}(q/p)=q/p
§4: 'energy loss is not directly accounted for... absorbed into the parameterisation of the propagation of the remaining state components'. Restricts validity to particles without hard bremsstrahlung; stated.
domain assumption Parameterisation uncertainty δpar is smaller than physical state uncertainty (MS + detector resolution)
§4 requirement; asserted to be satisfied ('achieved for all extrapolation functions') but the δpar distributions are not shown in this paper.
domain assumption Geant4/Pythia/EvtGen simulation faithfully represents the 2025 detector (material, field map, hit resolution, alignment)
§2 and §5: the parameter training and all performance evaluation use this simulation; the deployed real-data performance is not shown.
domain assumption The 'precise Runge-Kutta extrapolation' is an adequate ground truth, and the injected SciFi shifts are realistic 2024 misalignments
Eq. 12 uses the RK extrapolation as the target; §6 states misalignment shifts are 'realistic for the 2024 data-taking period' without quantitative description.

pith-pipeline@v1.3.0-alltime-deepseek · 11981 in / 20854 out tokens · 174744 ms · 2026-08-02T01:02:16.267286+00:00 · methodology

0 comments

read the original abstract

The LHCb experiment at the Large Hadron Collider underwent a major upgrade before the LHC Run 3 data taking period, employing an all-software approach in its trigger system. Here we present a fast implementation of a Kalman filter, used in the first trigger stage since the 2025 data taking period, allowing to determine parameter estimates of charged-particle trajectories at the full LHCb collision frequency of 30 MHz. This approach replaces computationally expensive magnetic field map lookups and numerical integration methods with fast analytical parameterisations while maintaining the mathematical framework of Kalman filtering. Implemented on approximately 500 GPUs within the first-level trigger, the algorithm has replaced the previous partial track fitting algorithm in the real-time trigger environment at the cost of a 2% increase in processing time. Compared to the previous fitter this parameterised Kalman filter shows a significantly improved momentum resolution, resulting in a factor of two improvement in the invariant mass resolutions for reconstructed D0 and J/{\psi} hadrons. It additionally demonstrates greater robustness against detector misalignment effects and substantially sharpens the discrimination between genuine particle trajectories and accidental background, more than doubling the rejection of the latter at no cost to genuine-track efficiency, for a standard selection working point.

Figures

Figures reproduced from arXiv: 2607.14793 by Lennart H. Uecker, Michel De Cian, Pierre Billoir, Thomas Boettcher.

**Figure 1.** Figure 1: Layout of the upgraded LHCb detector. [2] charged clusters in detectors with their spatial information, so-called hits, are combined to form particle trajectories using pattern recognition algorithms. Second, a track fit is performed to obtain the best estimate of the properties of the particle from the collection of associated hits. In LHCb this fit is performed using a Kalman filter. First published in 1… view at source ↗

**Figure 2.** Figure 2: Measured magnetic field strength in the y direction inside the LHCb detector as a function of z [24]. Overlaid are the positions of the three tracking detectors. of the involved layer, i.e. ⃗xk|k−1 = ⃗fX→Y (⃗xk−1|k−1; p X→Y,k−1 ) 4 . For example, the effect of the magnetic field is already sizeable between the UT layers, with a field strength in the y-component By between 0.1 T and 0.3 T. In this region th… view at source ↗

**Figure 3.** Figure 3: Comparison of throughput in a single production GPU (NVIDIA RTX A5000) running [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: The momentum resolution is compared using the same simulated events of the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: Discrimination power between real and fake (ghost) tracks, compared using the [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Two-particle invariant mass histograms taken from running HLT1 lines over simulation [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 9 linked inside Pith

[1]

LHCb collaboration, A. A. Alves Jr.et al.,The LHCb detector at the LHC, JINST3 (2008) S08005

2008
[2]

Aaijet al.,The LHCb Upgrade I, JINST19(2024) P05065, arXiv:2305.10515

LHCb collaboration, R. Aaijet al.,The LHCb Upgrade I, JINST19(2024) P05065, arXiv:2305.10515

arXiv 2024
[3]

Aaijet al.,LHCb detector performance, Int

LHCb collaboration, R. Aaijet al.,LHCb detector performance, Int. J. Mod. Phys. A30(2015) 1530022,arXiv:1412.6352

Pith/arXiv arXiv 2015
[4]

LHCb collaboration,LHCb SMOG Upgrade, CERN-LHCC-2019-005, 2019

2019
[5]

LHCb collaboration,LHCb VELO Upgrade Technical Design Report, CERN-LHCC- 2013-021, 2013

2013
[6]

LHCb collaboration,LHCb Tracker Upgrade Technical Design Report, CERN-LHCC- 2014-001, 2014

2014
[7]

Adinolfiet al.,Performance of the LHCb RICH detector at the LHC, Eur

M. Adinolfiet al.,Performance of the LHCb RICH detector at the LHC, Eur. Phys. J.C73(2013) 2431,arXiv:1211.6759

Pith/arXiv arXiv 2013
[8]

LHCb collaboration,LHCb PID Upgrade Technical Design Report, CERN-LHCC- 2013-022, 2013

2013
[9]

LHCb collaboration,LHCb Upgrade GPU High Level Trigger Technical Design Report, CERN-LHCC-2020-006, 2020

2020
[10]

Evans, C

T. Evans, C. Fitzpatrick, and J. Horswill,An automated bandwidth division for the LHCb upgrade trigger, Comput. Softw. Big Sci.9(2025) 7,arXiv:2502.09557

Pith/arXiv arXiv 2025
[11]

Aaijet al.,A comparison of CPU and GPU implementa- tions for the LHCb experiment Run 3 trigger, Comput

LHCb collaboration, R. Aaijet al.,A comparison of CPU and GPU implementa- tions for the LHCb experiment Run 3 trigger, Comput. Softw. Big Sci.6(2022) 1, arXiv:2105.04031

Pith/arXiv arXiv 2022
[12]

Aaijet al.,Allen: A high level trigger on GPUs for LHCb, Comput

R. Aaijet al.,Allen: A high level trigger on GPUs for LHCb, Comput. Softw. Big Sci.4(2020) 7,arXiv:1912.09161

Pith/arXiv arXiv 2020
[13]

Sj¨ ostrand, S

T. Sj¨ ostrand, S. Mrenna, and P. Skands,A brief introduction to PYTHIA 8.1, Comput. Phys. Commun.178(2008) 852, arXiv:0710.3820; T. Sj¨ ostrand, S. Mrenna, and P. Skands,PYTHIA 6.4 physics and manual, JHEP05(2006) 026, arXiv:hep-ph/0603175

Pith/arXiv arXiv 2008
[14]

Belyaevet al.,Handling of the generation of primary events in Gauss, the LHCb simulation framework, J

I. Belyaevet al.,Handling of the generation of primary events in Gauss, the LHCb simulation framework, J. Phys. Conf. Ser.331(2011) 032047

2011
[15]

D. J. Lange,The EvtGen particle decay simulation package, Nucl. Instrum. Meth. A462(2001) 152

2001
[16]

Davidson, T

N. Davidson, T. Przedzinski, and Z. Was,PHOTOS interface in C++: Technical and physics documentation, Comput. Phys. Commun.199(2016) 86, arXiv:1011.0937. 14

Pith/arXiv arXiv 2016
[17]

Allisonet al.,Geant4 developments and applications, IEEE Trans

Geant4 collaboration, J. Allisonet al.,Geant4 developments and applications, IEEE Trans. Nucl. Sci.53(2006) 270; Geant4 collaboration, S. Agostinelliet al.,Geant4: A simulation toolkit, Nucl. Instrum. Meth.A506(2003) 250

2006
[18]

Clemencicet al.,The LHCb simulation application, Gauss: Design, evolution and experience, J

M. Clemencicet al.,The LHCb simulation application, Gauss: Design, evolution and experience, J. Phys. Conf. Ser.331(2011) 032023

2011
[19]

R. E. Kalman,A new approach to linear filtering and prediction problems, Transactions of the ASME–Journal of Basic Engineering82(1960) 35

1960
[20]

W. D. Hulsbergen,Decay chain fitting with a Kalman filter, Nucl. Instrum. Meth. A552(2005) 566,arXiv:physics/0503191

Pith/arXiv arXiv 2005
[21]

Aiet al.,A gpu-based kalman filter for track fitting, Computing and Software for Big Science5(2021) 20

X. Aiet al.,A gpu-based kalman filter for track fitting, Computing and Software for Big Science5(2021) 20

2021
[22]

Cerati, Giuseppeet al.,Parallelized kalman-filter-based reconstruction of particle tracks on many-core processors and gpus, EPJ Web Conf.150(2017) 00006

2017
[23]

Billoir, M

P. Billoir, M. De Cian, P. A. G¨ unther, and S. Stemmle,A parametrized Kalman filter for fast track fitting at LHCb, Comput. Phys. Commun.265(2021) 108026, arXiv:2101.12040

Pith/arXiv arXiv 2021
[24]

Losassoet al.,Tests and Field Map of LHCb Dipole Magnet, Applied Supercon- ductivity, IEEE Transactions on16(2006) 1700

M. Losassoet al.,Tests and Field Map of LHCb Dipole Magnet, Applied Supercon- ductivity, IEEE Transactions on16(2006) 1700

2006
[25]

LHCb collaboration,HLT1 RICH decoding algorithm performance, LHCB-FIGURE- 2024-027, 2024

2024
[26]

LHCb collaboration,LHCb Framework TDR for the LHCb Upgrade II Opportunities in flavour physics, and beyond, in the HL-LHC era, CERN-LHCC-2021-012, 2022. 15

2021

[1] [1]

LHCb collaboration, A. A. Alves Jr.et al.,The LHCb detector at the LHC, JINST3 (2008) S08005

2008

[2] [2]

Aaijet al.,The LHCb Upgrade I, JINST19(2024) P05065, arXiv:2305.10515

LHCb collaboration, R. Aaijet al.,The LHCb Upgrade I, JINST19(2024) P05065, arXiv:2305.10515

arXiv 2024

[3] [3]

Aaijet al.,LHCb detector performance, Int

LHCb collaboration, R. Aaijet al.,LHCb detector performance, Int. J. Mod. Phys. A30(2015) 1530022,arXiv:1412.6352

Pith/arXiv arXiv 2015

[4] [4]

LHCb collaboration,LHCb SMOG Upgrade, CERN-LHCC-2019-005, 2019

2019

[5] [5]

LHCb collaboration,LHCb VELO Upgrade Technical Design Report, CERN-LHCC- 2013-021, 2013

2013

[6] [6]

LHCb collaboration,LHCb Tracker Upgrade Technical Design Report, CERN-LHCC- 2014-001, 2014

2014

[7] [7]

Adinolfiet al.,Performance of the LHCb RICH detector at the LHC, Eur

M. Adinolfiet al.,Performance of the LHCb RICH detector at the LHC, Eur. Phys. J.C73(2013) 2431,arXiv:1211.6759

Pith/arXiv arXiv 2013

[8] [8]

LHCb collaboration,LHCb PID Upgrade Technical Design Report, CERN-LHCC- 2013-022, 2013

2013

[9] [9]

LHCb collaboration,LHCb Upgrade GPU High Level Trigger Technical Design Report, CERN-LHCC-2020-006, 2020

2020

[10] [10]

Evans, C

T. Evans, C. Fitzpatrick, and J. Horswill,An automated bandwidth division for the LHCb upgrade trigger, Comput. Softw. Big Sci.9(2025) 7,arXiv:2502.09557

Pith/arXiv arXiv 2025

[11] [11]

Aaijet al.,A comparison of CPU and GPU implementa- tions for the LHCb experiment Run 3 trigger, Comput

LHCb collaboration, R. Aaijet al.,A comparison of CPU and GPU implementa- tions for the LHCb experiment Run 3 trigger, Comput. Softw. Big Sci.6(2022) 1, arXiv:2105.04031

Pith/arXiv arXiv 2022

[12] [12]

Aaijet al.,Allen: A high level trigger on GPUs for LHCb, Comput

R. Aaijet al.,Allen: A high level trigger on GPUs for LHCb, Comput. Softw. Big Sci.4(2020) 7,arXiv:1912.09161

Pith/arXiv arXiv 2020

[13] [13]

Sj¨ ostrand, S

T. Sj¨ ostrand, S. Mrenna, and P. Skands,A brief introduction to PYTHIA 8.1, Comput. Phys. Commun.178(2008) 852, arXiv:0710.3820; T. Sj¨ ostrand, S. Mrenna, and P. Skands,PYTHIA 6.4 physics and manual, JHEP05(2006) 026, arXiv:hep-ph/0603175

Pith/arXiv arXiv 2008

[14] [14]

Belyaevet al.,Handling of the generation of primary events in Gauss, the LHCb simulation framework, J

I. Belyaevet al.,Handling of the generation of primary events in Gauss, the LHCb simulation framework, J. Phys. Conf. Ser.331(2011) 032047

2011

[15] [15]

D. J. Lange,The EvtGen particle decay simulation package, Nucl. Instrum. Meth. A462(2001) 152

2001

[16] [16]

Davidson, T

N. Davidson, T. Przedzinski, and Z. Was,PHOTOS interface in C++: Technical and physics documentation, Comput. Phys. Commun.199(2016) 86, arXiv:1011.0937. 14

Pith/arXiv arXiv 2016

[17] [17]

Allisonet al.,Geant4 developments and applications, IEEE Trans

Geant4 collaboration, J. Allisonet al.,Geant4 developments and applications, IEEE Trans. Nucl. Sci.53(2006) 270; Geant4 collaboration, S. Agostinelliet al.,Geant4: A simulation toolkit, Nucl. Instrum. Meth.A506(2003) 250

2006

[18] [18]

Clemencicet al.,The LHCb simulation application, Gauss: Design, evolution and experience, J

M. Clemencicet al.,The LHCb simulation application, Gauss: Design, evolution and experience, J. Phys. Conf. Ser.331(2011) 032023

2011

[19] [19]

R. E. Kalman,A new approach to linear filtering and prediction problems, Transactions of the ASME–Journal of Basic Engineering82(1960) 35

1960

[20] [20]

W. D. Hulsbergen,Decay chain fitting with a Kalman filter, Nucl. Instrum. Meth. A552(2005) 566,arXiv:physics/0503191

Pith/arXiv arXiv 2005

[21] [21]

Aiet al.,A gpu-based kalman filter for track fitting, Computing and Software for Big Science5(2021) 20

X. Aiet al.,A gpu-based kalman filter for track fitting, Computing and Software for Big Science5(2021) 20

2021

[22] [22]

Cerati, Giuseppeet al.,Parallelized kalman-filter-based reconstruction of particle tracks on many-core processors and gpus, EPJ Web Conf.150(2017) 00006

2017

[23] [23]

Billoir, M

P. Billoir, M. De Cian, P. A. G¨ unther, and S. Stemmle,A parametrized Kalman filter for fast track fitting at LHCb, Comput. Phys. Commun.265(2021) 108026, arXiv:2101.12040

Pith/arXiv arXiv 2021

[24] [24]

Losassoet al.,Tests and Field Map of LHCb Dipole Magnet, Applied Supercon- ductivity, IEEE Transactions on16(2006) 1700

M. Losassoet al.,Tests and Field Map of LHCb Dipole Magnet, Applied Supercon- ductivity, IEEE Transactions on16(2006) 1700

2006

[25] [25]

LHCb collaboration,HLT1 RICH decoding algorithm performance, LHCB-FIGURE- 2024-027, 2024

2024

[26] [26]

LHCb collaboration,LHCb Framework TDR for the LHCb Upgrade II Opportunities in flavour physics, and beyond, in the HL-LHC era, CERN-LHCC-2021-012, 2022. 15

2021