Pith. sign in

REVIEW 3 major objections 5 minor 6 references

Lepton and photon energy scale and resolution corrections based on the minimization of an analytical likelihood: IJazZ2.0

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read The paper argues that an analytical, error-function-based likelihood can replace random smearing in Z->ll calibration, recovering scale and resolution parameters with unbiased estimates and large CPU gains.

desk verdict Useful calibration tool, but the 'exact' claim is overstated: Eq. 9 is a first-order approximation and Eq. 15 as printed is dimensionally wrong; still worth a serious referee. read the letter →

arxiv 2602.17300 v2 pith:KPAYV4PY submitted 2026-02-19 hep-ex physics.data-an

classification hep-exphysics.data-an
keywords leptonenergyscaleresolutionsmearinganalyticallikelihoodZ->llcalibrationerror-functionconvolutionrelativepTcategorisationphotonautomaticdifferentiation
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 aims to show that the standard calibration step of matching simulated Z->ll invariant-mass shapes to data does not need random-number smearing of each simulated event. Instead, it builds an exact analytical expression, based on error functions, for the probability that a smeared event lands in a given mass bin, making the likelihood smooth and differentiable. This lets calibration parameters—per-category energy scale and resolution—be fitted with automatic differentiation, giving a large speed gain and stable minimization. The authors validate that fitted parameters recover injected values in toy and realistic simulations, and they extend the same likelihood to photon energy scale using Z->mumu-gamma decays. If the method holds, it gives experiments a fast, unified calibration tool for leptons and photons.

What carries the argument

The key object is the tensor alpha_ijc (Eq. 11): for each di-lepton category c, the probability that a simulated event in fine bin j migrates to data bin i after applying scale r and resolution sigma, expressed with error functions. It turns the binned likelihood into a smooth, differentiable function of r and sigma, so automatic differentiation can compute gradients exactly. For photons, the observable V_DY (Eq. 24) replaces the invariant mass, making the same machinery applicable to Z->mumu-gamma events. The relative-pT variable r_pT = pT/m_ll is the workaround that removes category-migration biases when calibration categories depend on lepton transverse momentum.

What would settle it

A concrete check: generate Z->ll events with a much larger injected resolution (for example sigma_l = 5-10%) and fit the parameters; if the recovered scale or resolution drifts beyond the reported uncertainty, the Gaussian approximation of Eq. (9) is falsified. Alternatively, compute the exact distribution of m_ll from the product of two independent Gaussian-smearing factors and compare it to the assumed Gaussian shape; any visible deviation at the resolution values used in real experiments would bound the method's validity.

Watch

Extended reading notes

Core claim

The central claim is that Eqs. (9) and (11), which model each simulated di-lepton event's smeared mass as a Gaussian with mean r_ll times the Monte Carlo mass and width r_ll sigma_ll times that same mass, and which bin this Gaussian via error functions, yield a likelihood whose fitted parameters are unbiased and whose covariance is accurate. The error-function tensor alpha_ijc turns the binned convolution into a dense but differentiable operation, so gradients can be computed exactly. A relative-pT variable r_pT = pT/m_ll is introduced to avoid biases when categories depend on lepton pT, and a two-step fit (scale first, then resolution) recovers injected parameters in realistic Drell-Yan sim

Load-bearing premise

The whole fit rests on approximating the smeared di-lepton invariant mass as Gaussian with mean r_ll times the Monte Carlo mass and width r_ll sigma_ll times that mass (Eq. 9); because m_ll is a product of two Gaussian-smeared pT values, this is only exact in a small-resolution limit, and the paper does not quantify where it breaks.

Editorial extensions

If this is right

  • Lepton energy scale and resolution can be simultaneously extracted per category with exact gradients, making large-scale multi-parameter calibrations (100+ parameters, 20M events) run in minutes instead of days.
  • Finite-MC-statistics uncertainties are propagated to the fitted parameters through an analytic linear-response formula, so the fit reports correct total uncertainties even when simulation size is limiting.
  • pT-dependent calibrations become usable without bias by rebinning in r_pT rather than absolute pT, and correcting scale before measuring resolution.
  • Photon energy scale can be measured in Z->mumu-gamma events using the same likelihood, with a small iterative correction for the V_DY approximation.
  • The same analytical machinery can run on any framework supporting automatic differentiation, including GPUs, making it well suited for the precision-calibration environment of a large collider experiment.

Reading between the lines

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

  • The Gaussian form of Eq. (9) is a first-order approximation; if the lepton pT resolution is large, the exact product-of-Gaussians distribution will deviate, so the claimed unbiasedness likely has a regime bound in resolution fraction.
  • The V_DY bias for photon resolution smearing above ~2% (Sections 5.4.1-5.4.2) suggests a two-pass strategy: fit resolution from Z->ee events, then smear the simulation and re-fit the photon scale, rather than extracting both simultaneously.
  • The method's speed opens the possibility of per-event calibration as a function of many variables at once, something that was previously prohibitive with random smearing.
  • A natural stress test is to run the same likelihood with a non-Gaussian smearing model (e.g., a double-sided Crystal Ball); the error-function integral approach may be generalizable beyond Gaussian shapes.
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 paper presents an analytical likelihood method (IJazZ2.0) for extracting per-category lepton energy scale and resolution corrections from Z→ℓℓ decays, replacing random-smearing convolution with a fully differentiable error-function-based likelihood. The central formulas are the per-lepton Gaussian smearing model (Eq. 4), the combination rules for dilepton scale and resolution (Eq. 6), the Gaussian form of the smeared invariant mass (Eq. 9), and the tensor α_ijc that maps finely binned simulation to data bins (Eq. 11). The method is validated with a toy Breit–Wigner MC and a realistic Pythia Z→ℓℓ sample, showing unbiased parameter recovery at σ_sim=1.5%. A relative-pT categorization is introduced to mitigate migration biases. The approach is extended to photon energy calibration using Z→µµγ decays via a new observable V_DY (Eq. 24), with an iterative fitting procedure and a documented set of biases for large resolutions and pT cuts. The paper also derives an analytical formula for MC-statistical uncertainties (Eqs. 15–19) and reports large CPU gains from automatic differentiation.

Significance. If the derivation were exact as claimed, the method would be a valuable, fast, and numerically stable calibration tool for LHC precision measurements. The paper has concrete strengths: it ships reproducible software on PyPI, demonstrates the CPU gain quantitatively, validates the likelihood against random-smearing, and explicitly documents some limitations in the photon case (Sec. 5.4). The analytical treatment of finite-MC-statistics uncertainties is an ambitious and useful contribution. However, the central Gaussian propagation step (Eq. 9) is only a first-order approximation, and the exactness claim is not supported; the MC-uncertainty derivative (Eq. 15) appears to be missing a normalization factor. These issues affect the core claim of unbiased parameter recovery and accurate uncertainties. With corrections and additional validation, the method could become a solid contribution.

major comments (3)
  1. [§2.3, Eq. (9)] Eq. (9) is presented as a consequence of Eqs. (4) and (6), but it is not exact. From Eq. (1), m_ll² ∝ p_T1 p_T2; under independent Gaussian smearing of each p_T (Eq. 4), the conditional distribution of m_ll is that of a constant times the square root of a product of two Gaussians, which is not Gaussian. To second order in σ, the mean is shifted by ≈ −m_mc sqrt(r1 r2)(σ1²+σ2²)/8 and the distribution is skewed. Eq. (9) is therefore a first-order delta-rule approximation that is never stated or quantified. Since the α tensor (Eq. 11) and the likelihood (Eq. 12) inherit this Gaussian form, the extracted r_b and σ_b and their Hessian uncertainties carry the neglected bias. The Pythia validation uses σ_sim=1.5%, where the bias is ~5×10⁻⁵ relative and invisible; it does not test larger resolutions (high-|η| electrons, high-pT muons). Please either prove the Gaussian form or explicitly introduce
  2. [§3, Eq. (15)] The derivative of the NLL with respect to the MC bin content m_jc is missing a normalization factor. With p_ic = A_ic/S_c, A_ic = Σ_j α_ijc m_jc, S_c = Σ_i A_ic, the correct derivative is δnll/δm_jc = (1/S_c)[ n_c Σ_i α_ijc − Σ_i n_ic α_ijc / p_ic ]. Eq. (15) as written lacks the 1/S_c factor; the stated definitions n_c ≡ Σ_i n_ic and p_c ≡ Σ_i p_ic (=1) do not repair the dimensional inconsistency. Because Eq. (19) and the validation in Fig. 4 build on this derivative, please correct Eq. (15), verify against the code, and confirm that Fig. 4 was produced with the corrected formula.
  3. [§4, Fig. 7] The claim of unbiased parameter recovery is qualified only by the first-pT-bin bias, but the relative-pT recasting procedure itself is heuristic: the absolute-pT bin edges are reconstructed from average relative-pT values, and the two-step fit fixes scale before extracting smearing. The paper asserts this 'closely reproduces' the original categories without a quantitative comparison. Given that the method is intended for differential calibration, please provide a quantitative check of the recast bin accuracy (e.g., bin-by-bin differences in injected parameters or in fitted parameters versus the original absolute-pT categorization).
minor comments (5)
  1. [Abstract and Introduction] The phrase 'exact analytical treatment' is too strong. Consider wording such as 'analytic treatment within a Gaussian-smearing approximation'.
  2. [Eq. (6)] The notation for the resolution combination is garbled: 'σℓℓc =0.5× q σℓ2 b1 +σ ℓ2 b2' should be typeset as σℓℓc = 0.5√(σℓ_b1² + σℓ_b2²).
  3. [Sec. 5.4, Fig. 10] The 'naive bias prediction' is described only verbally. Please provide the exact formula used to produce the dashed curve so the reader can reproduce it.
  4. [Sec. 4, second fit] The statement that the second fit converges to the same smearing parameters as the first 'because no correlations are introduced' in the absolute-pT case could be expanded: a short explanation of why relative-pT introduces correlations would help.
  5. [References] Reference [3] is the software repository; it would be useful to mention the version number and a DOI if available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the free likelihood parameters are fitted against independent injected/pseudo-data decalibrations, and the only self-citation is the non-load-bearing software reference.

full rationale

The derivation chain is self-contained and does not reduce any predicted quantity to an input of the fit. The scale and resolution parameters r_lb and sigma_lb appear as free parameters of the multinomial negative log-likelihood (Eqs. 7-8), with the expected probabilities built from the analytic error-function smearing tensor alpha_ijc (Eqs. 10-12). These parameters are then extracted by maximum likelihood and compared to separately injected decalibration and smearing values in toy and Pythia-based pseudo-data; the validation is therefore a standard closed-loop parameter-recovery check, not a recycling of fitted quantities as predictions. The V_DY construction for photons (Eqs. 22-25) is derived from event kinematics and a first-order Taylor expansion, and its agreement with simulation is a check of that approximation rather than an output that was used to define the fit. The self-citation [3] is only the PyPI software distribution and carries no theoretical load. The only substantive concern is non-circular: Eq. 9 propagates single-lepton Gaussian pT smearing to m_ll as a Gaussian with width r_ll sigma_ll m_mc, which is a first-order approximation not stated or quantified in the paper; this affects the accuracy of the method at large resolutions or kinematic cuts, but it is a modeling approximation, not a case where a claimed prediction is identical to a fitted input by construction.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

No new physical entities are postulated: V_DY is a constructed kinematic variable, not an independent physical entity. The main hidden load-bearing choices are the Gaussian smearing ansatz, the first-order propagation to m_ll, and the relative-pT recasting procedure.

free parameters (6)
  • Per-category scale parameters r_lb = fitted by likelihood
    These are the targets of the method; their number is a user choice (N_B).
  • Per-category smearing parameters sigma_lb = fitted by likelihood
    Extracted simultaneously with r_lb; identifiability relies on line-shape and binning.
  • Lepton category binning N_B and bin edges
    User-chosen discretization; too-fine categories create migration and low-statistics bias; no objective criterion is derived.
  • Adaptive binning defaults (min 100 events/category; N_bins=min(N^(1/3), DeltaM/0.5 GeV); fine MC width <0.2 GeV) = defaults
    Heuristics to avoid zero-probability bins and reduce bias; not derived from first principles.
  • Relative-pT bin mapping and recast to absolute pT
    The paper defines bins by pT/m_ll and then recasts using average pT per bin; this mapping is ad hoc and not proven optimal.
  • Photon V_DY fit window and pT-gamma selection = pT_gamma>25 GeV; 80<m_mu_mu_gamma<100 GeV
    Results depend on these selections; residual biases from sigma_gamma and the pT_gamma cut are documented but not corrected in the software.
assumptions (7)
  • standard math Analytical Gaussian-bin integrals via error functions and covariance from inverse Hessian are valid.
    Relies on standard calculus and probability; not an issue.
  • domain assumption Residual data/MC energy differences follow a Gaussian smearing model (Eq. 4).
    Assumes MC resolution is better than data and that both leptons' energies are perturbed by normal distributions; no non-Gaussian tails or correlated material effects.
  • domain assumption Delta_eta and Delta_phi are perfectly measured and contribute no smearing or bias to m_ll.
    Stated in Sec. 2; ignores angular resolution and possible track or cluster direction biases.
  • domain assumption Single-lepton corrections combine to a Gaussian di-lepton mass distribution via Eqs. 6 and 9.
    This is a first-order propagation of relative scale and resolution; the exact distribution of sqrt(pT1*pT2) under Gaussian pT smearing is not Gaussian, but this is not stated in the paper.
  • domain assumption The binned data are multinomially distributed and the MC histogram m_jc is a faithful template (Eqs. 7 and 12).
    Assumes no data/MC shape differences other than the parameterized scale and smearing, and ignores MC statistical noise in p_ic during the main fit.
  • ad hoc to paper V_DY is a valid photon-calibration observable with an additive scale shift (Eqs. 24-25).
    Constructed variable normalized by mZ; relies on a first-order Taylor expansion and on muons being well calibrated; an iterative correction is required.
  • ad hoc to paper Relative pT = pT/m_ll categorization removes migration biases, and the recast absolute-pT binning is representative.
    Introduced to fix the pT-migration problem; no proof is given that the two-step scale-then-smear procedure is unbiased for all spectra.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lepton and photon energy scale and resolution corrections based on the minimization of an analytical likelihood: IJazZ2.0." pith.science (2026). https://pith.science/paper/KPAYV4PY

@misc{pith2026260217300,
  author       = {Pith},
  title        = {Pith review of: Lepton and photon energy scale and resolution corrections based on the minimization of an analytical likelihood: IJazZ2.0},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPAYV4PY}},
  note         = {Machine review of arXiv:2602.17300}
}
abstract

We present a novel method to determine lepton energy scale and resolution corrections by means of an analytical likelihood maximization applied to Drell--Yan \(Z \to \ell\ell\) events. The approach relies on an exact analytical treatment of the Gaussian energy smearing model, avoiding random-number-based convolution techniques. This formulation results in a fully differentiable likelihood enabling the use of automatic differentiation algorithms, and thus a substantial reduction in computational cost. The method, implemented in the \ijazz software, allows the simultaneous extraction of scale and resolution parameters across multiple lepton categories defined by detector or kinematic variables. We validate the technique using toy Monte Carlo studies and realistic Pythia-based simulations, demonstrating unbiased parameter recovery and accurate uncertainty estimates. Particular attention is given to categorizations involving lepton transverse momentum, for which a relative-\(p_T\) strategy is introduced to mitigate biases induced by category migration and kinematic correlations. The method is further adapted to photon-energy scale measurement in \(Z \to \mu^+\mu^-\gamma\) decays. Compared to conventional approaches, the analytical method improves numerical stability, robustness of the minimization, and computational performance, making it well suited for large-scale precision calibration tasks at the LHC.

Figures

Figures reproduced from arXiv: 2602.17300 by the authors.

Figure 1
Figure 1. Left: the original MC distribution (Breit–Wigner) from 10,000 generated events. Right: the smeared MC distribution obtained using a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Validation of the method using a naive MC simulation. A subset of events is decalibrated and smeared according to known functions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. adaptive binning: The bin width is adapted so that the total number of events in the simulation is the same in each bin, the total number [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Validation of the statistical uncertainties due to limited MC statistics for the response parameters (left) and smearing parameters (right). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Measured scale (left) and smearing (right) parameters using the Pythia-based DY simulation. The dashed line corresponds to the injected [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Left: reconstructed di-lepton mass for the absolute [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Measured scale (left) and smearing (right) parameters using the relative [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Validation of Eq. 25 using a Pythia-based simulation of the [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Validation of the method using a Pythia-based [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Potential bias induced by a difference in energy resolution between data and simulation. Left: retrieved value of the photon energy scale shift δrγ as a function of the injected σγ value. A small bias is observed for large values of σγ. A naive prediction of this bias…
Figure 11
Figure 11. Figure 11: Potential bias induced by the steeply falling photon transverse-momentum spectrum combined with energy-resolution mismodeling. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 2 linked inside Pith

  1. [1]

    Navas, et al

    S. Navas, et al. (Particle Data Group), Review of particle physics, Phys. Rev. D 110 (2024) 030001. doi:10. 1103/PhysRevD.110.030001

  2. [2]

    A. G. Baydin, B. A. Pearlmutter, A. A. Radul, J. M. Siskind, Automatic differentiation in machine learning: a sur- vey, Journal of Machine Learning Research 18 (2018) 1–43. URL:http://jmlr.org/papers/v18/17-468. html

  3. [3]

    Couderc, P

    F. Couderc, P. Gaigne, O. Sahin, IJazZ: I Just AnalyZe the Z, 2025. URL:https://pypi.org/project/ ijazz/, python package distributed via PyPI

  4. [4]

    doi:10.1088/1748-0221/16/05/P05014.arXiv:2012.06888

    CMS Collaboration, Electron and photon reconstruction and identification with the CMS experiment at the CERN LHC, JINST 16 (2021) P05014. doi:10.1088/1748-0221/16/05/P05014.arXiv:2012.06888

  5. [5]

    Abadi, A

    M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y . Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Lev- enberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V . Vanho...

  6. [6]

    Bierlich, S

    C. Bierlich, S. Chakraborty, N. Desai, L. Gellersen, I. Helenius, P. Ilten, L. Lönnblad, S. Mrenna, S. Prestel, C. T. Preuss, T. Sjöstrand, P. Skands, M. Utheim, R. Verheyen, A comprehensive guide to the physics and usage of pythia 8.3, 2022. URL:https://arxiv.org/abs/2203.11601.arXiv:2203.11601. 17

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.