WorksPhysics and Astronomy
Loop-induced at NLO+PS for LHC simulations: exact, fast scale and PDF event weights, theory uncertainties, and the limits of rescaled LO+PS
Abstract
We study loop-induced production with charged-lepton and neutrino decays at next-to-leading order (NLO) in QCD matched to parton showers, using public software at and 13.6 TeV. An exact loop-amplitude cache reduces event-generation time by a factor of 26 with 450 scale and PDF weights, while Born suppression improves boosted-event statistics without changing the weighted prediction. We present inclusive results for both channels and charged-lepton differential and STXS Stage 1.3 predictions with theory uncertainties. At 13.6 TeV, fixed-order top-quark mass-scheme variations transferred to NLO+PS dominate the statistically precise central-Higgs STXS bins for GeV, reaching 32-50%. An inclusive factor does not reproduce the boosted and jet-dependent NLO+PS spectra of the tested LO+PS samples, and the differences depend strongly on the shower prescription. For specified CMS- and ATLAS-style vetoed sources at 13.6 TeV, we construct separate charged-lepton and neutrino corrections using ensembles of gradient-boosted classifiers, with calibration, target-theory variations and transfer uncertainties. Independent tests show substantial improvement over inclusive rescaling in most tested distributions, but not uniform closure. In charged-lepton calibration comparisons, the median STXS uncertainty from transported NLO scales and transfer terms is 17% for both the CMS- and ATLAS-style corrections, compared with LO scale uncertainties of 28% and 26% at fixed . These comparisons do not include the complete target-theory budget. The corrections do not restore the depleted hard-recoil population; the highest ATLAS-style bin stays below NLO+PS even at the upper edge of its band. The setup, reference results, and source- and channel-specific models support new NLO+PS generation and reuse of existing samples within stated limits.
Introduction
Higgs-boson production in association with a boson receives contributions from quark-initiated processes and from the loop-induced gluon-fusion channel, . Although the latter first appears at , it constitutes a sizeable part of the inclusive cross section and carries a much larger perturbative uncertainty than the quark-initiated contribution. At 13 TeV, the YR4 reference prediction assigns of 0.8839 pb (13.9%) to gluon fusion, with a scale uncertainty of [1, 2]. At 13.6 TeV, the corresponding contribution is 0.1360 of 0.9439 pb (14.4%) [3]. The top-quark loop also shapes the kinematics. The relative contribution grows in the boosted regime [4, 5], which is widely used in LHC analyses.
The calculation of beyond leading order (LO) is challenging because the NLO QCD correction requires two-loop amplitudes with massive quarks. Early NLO predictions therefore relied on the heavy-top limit and on expansions in complementary kinematic regimes [6, 7, 8, 9, 10]. The full top-mass dependence was subsequently obtained numerically and through improved analytic expansions, leading to complete NLO predictions and their comparison and combination [2, 11, 12, 13, 14]; soft-gluon resummation is also available at NLL [15, 16, 17].
The amplitudes of Refs. [18, 19] are now implemented in the public code ggxy [20] and interfaced to Powheg-Box-V2 [21, 22, 23]. This provides NLO QCD predictions matched to parton showers (ggxy ggZH, Ref. [24]).
LHC simulations of this channel have so far commonly relied on LO samples showered with Pythia 8 and normalized to a higher-order inclusive cross section, as in the measurements of ATLAS [25] and CMS [26]. The treatment of hard radiation depends on the shower configuration. For this colorless Born final state, the unvetoed Pythia 8 default allows emissions up to the kinematic limit (“power shower” [27, 28]). The CMS- and ATLAS-style sources studied here instead apply a Powheg veto using (Sections IX A and IX G); their predictions must be distinguished from the unvetoed default.
Large NLO+PS samples present a separate practical challenge. In the unmodified implementation, scale and parton-distribution-function (PDF) weights repeat expensive loop-amplitude evaluations. Inclusive generation also provides few events in the boosted Simplified Template Cross Section (STXS) bins. We therefore ask whether NLO+PS samples with the statistics and theory weights needed for LHC analyses can be generated efficiently, and how reliably existing LO+PS samples can be corrected to reproduce their kinematics.
We address both questions using only public software. We introduce an exact amplitude cache for ggxy ggZH to reduce the cost of scale and PDF event weights (Section III E), together with a Born-suppression factor that enhances the sampling of the high- region without changing the physical prediction (Section III B). We validate the resulting setup and present NLO and NLO+PS predictions at 13.6 and 13 TeV with a detailed theory uncertainty budget, including predictions in the STXS Stage 1.3 bins (§ 2 to VII). We compare inclusively rescaled LO+PS predictions with NLO+PS across several shower-start prescriptions to isolate their effect on boosted and jet-dependent observables (Section VIII). Separately, we test generator-level multivariate reweighting for the experiment-inspired sources and identify where the tested corrections succeed or remain limited (Section IX). The complete run definitions, reference results, and validation scripts are provided in the public package [29], with the reproduction procedure summarized in Section C.
Calculation, parameters, and codes
Process and calculation
We consider loop-induced production through gluon fusion at NLO QCD with full top-quark mass dependence, as implemented in ggxy [20] and in Powheg-Box-V2 through ggxy-ggZH [24]. The main results use with , summed over the three flavors. The boson is off shell and decays at amplitude level, preserving spin correlations and off-shell effects, with GeV. The upper boundary follows the validity range of the two-loop virtual amplitudes, whose expansion in external virtualities extends to approximately 150 GeV [24]. The Higgs boson is on shell in the hard process and is decayed by Pythia 8 for showered predictions (Section IV); none of the reported observables depends on its decay mode.
Born and real-emission one-loop matrix elements are evaluated with Recola [30] and Collier [31]. The two-loop virtual correction is provided by the ggxy amplitudes, using Padé-improved expansions [18, 24]; their accuracy is discussed in § 2 E. The real contribution includes the and channels and -boson radiation from initial-state quarks (extZ 1). Real-emission partons below 0.1 GeV in transverse momentum are removed by the implementation-level infrared cutoff.
We repeat the calculation for , summed over the three neutrino flavors, with the same settings apart from the decay mode (vdecaymode 12), the corresponding invariant-mass window, and independent random seeds. Inclusive results and reference comparisons for both decay channels are given in Table V and Table VI; dedicated neutrino reweighting results appear in Section IX F. The neutrino differential and STXS outputs and reweighting models are included in the data record (Section C). Unless stated otherwise, the differential and STXS predictions displayed in the paper refer to the charged-lepton channel.
Table V. Fixed-order cross sections in fb with PDF4LHC21, compared with Ref. [24].
| (fb) | Scale | PDF (%) | (%) | scheme | MC (%) | |
| LO, 13.6 TeV | 7.0225(10) | 1.79 | 1.07 | 0.01 | ||
| NLO, 13.6 TeV | 13.014(4) | 1.77 | 1.41 | 0.03 | ||
| LO, 13 TeV | 6.3643(9) | 1.82 | 1.08 | 0.01 | ||
| NLO, 13 TeV | 11.8044(35) | 1.80 | 1.43 | 0.03 | ||
| LO, 13.6 TeV | 13.8800(20) | 1.79 | 1.07 | 0.01 | ||
| NLO, 13.6 TeV | 25.704(8) | 1.77 | 1.41 | 0.03 | ||
| LO, 13 TeV | 12.5791(18) | 1.82 | 1.08 | 0.01 | ||
| NLO, 13 TeV | 23.296(7) | 1.80 | 1.43 | 0.03 |
Table VI. Inclusive fixed-order cross sections in fb with PDF4LHC21, summed over the three charged-lepton or neutrino flavors. Scale denotes the seven-point envelope; PDF and () are obtained from event weights; is the envelope of the predictions with relative to the on-shell result; MC is the relative integration uncertainty. For , the relative band is transferred from the charged-lepton calculation; the explicit check gives shifts of at LO and at NLO.
Although the Born contribution to is loop induced rather than tree level, we refer to it as LO. “LO” and “NLO” denote fixed-order predictions, while “LO+PS” and “NLO+PS” denote predictions matched to a parton shower. At LO, the and Higgs bosons are back to back in the transverse plane, so and no QCD parton is present. At NLO, one additional real-emission parton is possible. In LO+PS, all QCD radiation is generated by the shower. In NLO+PS, Powheg generates the hardest emission from the NLO real-emission kernel and the subsequent shower is vetoed against harder emissions; the matching and shower settings are described in Section IV.
Parameters
Input parameters follow Ref. [24], reproducing Tables 1 and 3 of that work at both center-of-mass energies (§ 2.3). Cross-code validation (§ 2.3) was performed with NNPDF3.1 NNLO [34] (NNPDF31_nnlo_as_0118_mc_hessian_pdfas, LHAPDF ID 325300) at 13.6 TeV; NNPDF3.1 variations are also retained as event weights for comparison. Using the same PDF set at LO and NLO ensures that the differential -factor, , isolates higher-order ME corrections.
Codes and versions
The nominal NLO and NLO+PS calculations use the public ggxy and ggxy-ggZH implementations in Powheg-Box-V2. Showering and analysis use Pythia 8 and Rivet, including the LHC Higgs Working Group STXS classifier. We use MCFM for an independent LO cross-check. The exact software commits, library versions, patches, and build toolchains are listed in Table XV and fixed by the public reproduction package [29].
| Component | Version | Reference |
| ggxy | commit 9edd1d5 (bundles Recola 1.4.1, Collier 1.2.8, CutTools, OneLOop) | [20] |
| Powheg-Box-V2 | commit 6654ea3 | [23] |
| ggxy_ggZH | commit 0fbaccf (includes the amplitude cache, Section III E) | [24, 29] |
| Patch | Born suppression in (Section III B) | [29] |
| Generator toolchain | LCG 107 (gcc/gfortran 11.3, LHAPDF 6.5.5, Boost 1.87) | |
| MCFM (LO check) | 10.3.1 with a patch selecting only; LCG 106 (gcc 13) | [58] |
| Pythia 8 | 8.315 | [35] |
| Rivet, YODA | 4.1.2, 2.1.2 | [37] |
| HepMC3, FastJet | 3.3.1, 3.5.0 | [36, 59] |
| Shower toolchain | LCG 108a (gcc 13) | |
| STXS classification | LHC Higgs WG Rivet routine SimplifiedTemplateCrossSections, commit 38eea57: Stage 1.3; framework and Stages 1.1, 1.2 | [42] [43] [1, 60] |
Table XV. Software. Compilers and libraries are those of the LCG software releases LCG 107 and LCG 108a; any equivalent installation can be used.
Fixed-order calculations
The nominal fixed-order LO and NLO predictions are obtained with ggxy, including the seven-point scale variations and the differential distributions used in this work. Two independent LO calculations provide cross-checks. A Born-only Powheg-Box-V2 stage-2 integration tests the implementation used for event generation, while MCFM evaluates stable- production with an independent analytic one-loop amplitude. The integration statistics, random-seed treatment, and binning used for these calculations are given in Section C.
Event generation
We generate the NLO+PS and LO+PS samples with ggxy-ggZH in Powheg-Box-V2 [21, 22, 23]. This section describes the choices that affect the physics content or the efficiency of the samples; the complete run settings and event statistics are given in Section C.
Powheg matching
The real-emission matrix element is separated into singular and finite-remnant components, , with
We use GeV, following Ref. [24]. The singular component enters the Sudakov exponent, while is generated as a finite, unexponentiated contribution. In the hard-emission limit, the Sudakov-generated term is proportional to , where is the NLO underlying-Born weight and is the Born weight [23]. The finite remnant carries no such enhancement. This local ratio can enhance hard radiation through terms beyond fixed NLO; it need not equal the inclusive factor. The choice of controls this partition and constitutes a matching ambiguity. We probe it with dedicated samples, including the change in the generated radiation (§ 2 D).
Born suppression for the high- tail
The physical spectrum falls steeply in the boosted region: at 13.6 TeV, the bin-averaged fixed-order NLO differential cross section decreases by a factor of between the 125–150 and 500–600 GeV bins. Inclusive generation therefore provides few events in the highest STXS bins. We enhance this region with a Born-suppression factor [23],
using GeV, , and . Here is the Higgs-boson transverse momentum in the underlying Born configuration, where it balances the lepton pair. The generation density is multiplied by , and the event receives the compensating factor . The weighted sample therefore retains the physical cross section while preferentially sampling large .
The suppression depends only on the underlying Born kinematics, so it does not alter the real-emission matrix element or the scale and PDF variations. Above GeV, a sample generated with the suppression has 2.1 times the effective number of events of an unweighted sample of the same size at 13.6 TeV (2.9 at 13 TeV). Closure against unsuppressed generation and fixed-order NLO is shown in § 2.3.
LO events
For the LO+PS reference, we generate Born events with the Powheg real emission disabled. These events are unweighted and carry SCALUP , the partonic center-of-mass energy. This scale sets the shower limit when pTmaxMatch = 1 is selected, or the veto threshold when the POWHEG hooks use pThard = 0. For this colorless Born final state, the unvetoed default shower (pTmaxMatch = 0) instead extends to the kinematic limit (Section VIII A). The mean event weight reproduces the fixed-order LO cross section at all scale points (§ 2.3); the corresponding run-card settings are listed in Section C.
Event weights
Each production event carries 155 alternative weights evaluated at the same phase-space point. They cover renormalization- and factorization-scale variations, PDF4LHC21 members and its variations, and an NNPDF3.1 set retained for validation. The complete weight list and identifiers are given in Section D. Cross sections are obtained from the mean signed event weight; the normalization and statistical convention are also specified there.
Exact amplitude cache
Alternative POWHEG weights normally trigger a new evaluation of at the same phase-space point. For ggxy_ggZH, this repeats the one-loop Born and real amplitudes from Recola and the two-loop virtual amplitude from ggxy, making large weight lists far more expensive than the nominal event generation. Without caching, an event with 450 weights takes 40 times as long as an event without alternative weights (Figure 1).
Figure 1. Event-generation cost as a function of the number of alternative weights, measured with the same 50 events at 13.6 TeV. The upper panel shows the time per event relative to generation without alternative weights or caching; bands give the range across 3 runs. The lower panel shows the cache gain in generation time and one-loop amplitude evaluations. Dashed lines extrapolate the measured linear trends from 450 to 884 weights.
The amplitudes do not depend on the PDF or on , and their powers of can be factored from the kinematic dependence. We therefore developed an exact amplitude cache for ggxy_ggZH, which has been merged upstream (commit 1a6dbc6). The cache keys each amplitude by the parton flavors, the bit representation of the momenta, and the mass-dependent inputs; the two-loop virtual term also includes . The corresponding power of is restored for the target scale, and the Born spin-correlation tensor is cached together with the Born amplitude. A cached value is reused only when all keyed inputs agree; otherwise the amplitude is evaluated again. Including the mass inputs prevents reuse of a nominal-mass result for a different mass, but does not turn that result into a mass-scheme variation. The closure tests reported here cover PDF and scale weights in the on-shell setup; no NLO+PS sample is generated. Implementation controls and memory settings are documented in Section C and in the public package [29].
With identical random seeds, cached and uncached generation gives identical event kinematics and alternative weights in the closure samples. The verification covers 150 events and 16 800 weights, with additional production checks of 150 events and 23 250 weights at each energy.
Figure 1 shows the computational cost as the number of alternative weights is increased from zero to 450. Without the cache, the number of one-loop amplitude evaluations grows linearly with the weight list; with the cache it remains 185 per event. At 450 weights, the cache reduces the number of amplitude evaluations by a factor 28 and the generation time by a factor 26. A linear extrapolation to 884 weights, representative of larger LHC production weight lists, gives factors of 53 and 41, respectively.
Parton shower and analysis
Pythia8 shower and Powheg veto
Generated LHE events are showered with Pythia 8 [35], including multiparton interactions and hadronization. The Higgs boson is decayed to by Pythia 8, and the resulting events are passed through HepMC3 [36] to Rivet [37]. LHE and shower-variation weights are propagated through the same event record.
For NLO+PS, the shower is allowed to explore the full kinematic phase space, while the Pythia 8 Powheg hooks veto QCD shower emissions above the event’s reference hardness. In the nominal setup, this hardness is SCALUP (pThard = 0), and shower emissions are tested with the Powheg transverse-momentum definition (pTdef = 1). This preserves the separation between the Powheg emission and subsequent QCD shower radiation. Alternative hardness definitions enter the matching uncertainty in § 2 D. The complete interface settings are listed in Section C.
For the shower-start comparison in Section VIII A, Born-only LO events are showered without a Powheg emission veto. Those configurations vary the allowed shower phase space directly. The CMS- and ATLAS-style sources instead use the veto prescriptions specified in Sections IX A and IX G; they are separate from the unvetoed comparison samples.
Tunes
We use the Monash tune [38] as the nominal shower configuration and compare it with the ATLAS A14 [39] and CMS CP5 [40] tunes (Table II). Each tune uses its native PDF set for initial-state radiation and multiparton interactions. In this production-sample comparison, the hard matrix element uses PDF4LHC21; the experiment-inspired sources have the PDF choices specified separately. The tune comparison probes shower couplings and PDFs as well as nonperturbative parameters.
| Tune | Shower/MPI PDF |
| Monash (nominal) | NNPDF2.3 QCD+QED LO |
| A14 | NNPDF2.3 LO |
| CP5 | NNPDF3.1 NNLO, |
Table II. Pythia 8 tunes used for the shower comparison.
The exact tune settings used in the production, including the CP5 parameters and energy-dependent cross-section inputs, are recorded in Section C.
Shower uncertainties
We evaluate renormalization-scale variations in initial- and final-state radiation with Pythia 8’s automated variation framework [41]. The ISR and FSR scales are varied independently by factors of and 2. For each component, the upward and downward shifts define an envelope; the ISR and FSR envelopes are then added in quadrature, separately upwards and downwards. Variations of the non-singular splitting-kernel terms are not included. We combine these scale variations with the Monash–A14–CP5 tune comparison when constructing the matching and shower uncertainty in § 2 D.
Analysis and STXS classification
Events are classified into STXS Stage 1.3 bins using the official LHC Higgs WG Rivet routine SimplifiedTemplateCrossSections [42, 43]. After separating the forward-Higgs bin , the fine classification assigns the remaining events to 18 bins: six intervals with boundaries 0, 75, 150, 250, 400, 600, GeV, each split into . We also use the standard eight-bin grouping, which keeps separate 0- and -jet bins in the 150–250 and 250–400 GeV intervals.
Jets are clustered with the anti- algorithm [44], radius , and counted for GeV. Stable descendants of the Higgs boson and of the decaying boson are excluded from the jet inputs, as required by the STXS definition. The vector-boson transverse momentum is evaluated from the hard-process after shower recoil, so for both charged-lepton and neutrino decays.
A companion Rivet analysis records the differential observables used in this work, including , , , , , jet multiplicity, and leading-jet . The validation of the shower and STXS analysis chain is presented in § 2.3; detailed event-record conventions are collected in Section D.
Uncertainties
We consider four uncertainty sources: perturbative scales, PDFs and , the top-quark mass scheme, and matching and shower effects. Fixed-order scale bands come from the integrations; scale, PDF, and variations of the generated samples use event weights. The mass-scheme study uses separate fixed-order calculations, while matching and shower variations use dedicated samples, reshowers, or shower weights as specified below.
Scales
Perturbative QCD scale uncertainties are defined by the standard seven-point envelope
at fixed PDF, following Refs. [2, 24]. Generated events contain the full nine scale combinations, from which this envelope is selected.
For rescaled LO+PS, the scale variation is separated into shape and normalization components. Each LO+PS variation is normalized to its own inclusive cross section and therefore changes only the shape. This envelope is added in quadrature to the inclusive NLO scale uncertainty carried by the normalization factor . The two components are treated as uncorrelated.
PDF and
PDF and strong-coupling uncertainties follow the PDF4LHC21 prescription [32]:
Both are evaluated directly from event weights. The quoted variation corresponds to , matching Ref. [2]; the standard PDF4LHC21 choice is obtained by rescaling this component by 1.5. PDF and uncertainties are added in quadrature. Each variation uses its associated PDF set, so the change is applied consistently to the PDF evolution and the matrix element.
Top-quark mass scheme
The baseline calculation uses the on-shell top-quark mass. We estimate the mass-scheme uncertainty by comparing it with predictions obtained using the running mass at , following Refs. [2, 45]. The envelope of these predictions relative to the on-shell result defines the scheme uncertainty. It estimates missing higher-order dependence on the mass definition and is distinct from the experimental uncertainty on .
We evaluate this variation at fixed order with ggxy at LO and NLO, inclusively and differentially, at both collision energies. No NLO+PS samples are generated in the scheme. For the central-Higgs STXS bins, we therefore transfer the fixed-order NLO /on-shell ratio in each interval to the corresponding NLO+PS bins. The same ratio is applied to all jet multiplicities within that interval. For the forward-Higgs bin , we use the inclusive fixed-order ratio, without a forward-region selection. These are transfer prescriptions rather than direct NLO+PS variations; they provide neither jet-dependent scheme information nor a separate calculation of the forward-region dependence.
ME–PS matching and shower uncertainties
The separation of radiation between the NLO calculation and the parton shower is not unique. We probe three components of this ambiguity at 13.6 TeV:
Matching: is compared with the nominal GeV to vary how much of the real emission is exponentiated by Powheg. We also vary the Pythia 8 Powheg-veto hardness definitions with pThard = 1 and pTdef = 2 relative to the nominal pThard = 0, pTdef = 1 setup (Section IV A).
Shower scales: ISR and FSR renormalization scales are varied as described in Section IV C.
Tune: CP5 and A14 are compared with the nominal Monash tune (Section IV B).
The matching contribution is the envelope of the and veto variations. We evaluate the shower-scale component and tune spread separately, and treat the three components as uncorrelated when adding them in quadrature, separately upwards and downwards, in each STXS bin. Table III gives the decomposition and total uncertainty.
TABLE III. Matching and shower uncertainties of the NLO+PS prediction in the STXS Stage 1.3 fine bins at 13.6 TeV. “Matching” is the envelope of the and Powheg-veto variations; “Shower scale” is the ISR/FSR scale component; “Tunes” is the CP5/A14 spread relative to Monash. The total is their quadrature sum, evaluated separately upwards and downwards. MC gives the statistical precision of the comparison between the independent and nominal samples. The veto variants use the nominal events and shower seeds. Variations are not resolved in bins where the nominal MC uncertainty exceeds 10%.
| (GeV), jets | Matching (%) | MC (%) | Shower scale (%) | Tunes (%) | Total (%) |
| 1.7 | |||||
| 0–75, 0 jets | 1.7 | ||||
| 75–150, 0 jets | 0.5 | ||||
| 150–250, 0 jets | 0.5 | ||||
| 250–400, 0 jets | 1.4 | ||||
| 400–600, 0 jets | 6.5 | ||||
| , 0 jets | 36.8 | ||||
| 0–75, 1 jet | 1.8 | ||||
| 75–150, 1 jet | 0.6 | ||||
| 150–250, 1 jet | 0.6 | ||||
| 250–400, 1 jet | 1.3 | ||||
| 400–600, 1 jet | 5.0 | ||||
| , 1 jet | 19.4 | ||||
| 0–75, jets | 1.8 | ||||
| 75–150, jets | 0.8 | ||||
| 150–250, jets | 0.7 | ||||
| 250–400, jets | 1.3 | ||||
| 400–600, jets | 3.7 | ||||
| , jets | 12.2 |
Table III.
Accuracy and scope of the calculation
The LO amplitudes retain the exact top-quark mass dependence. The two-loop virtual amplitudes use Padé-improved expansions in the forward and high-energy limits [18, 24]. Ref. [2] finds agreement with numerical virtual amplitudes containing the exact top-mass dependence [13] at better than the per-mille level. The expansion in external virtualities sets the upper validity range used throughout this work.
The calculation isolates the loop-induced gluon-fusion contribution. The interference with Drell–Yan-type amplitudes belongs to the quark-initiated production channel and is not included.
Validation
We validate the calculation at three levels: consistency among the fixed-order and event-generation implementations, comparison with the published ggxy–ggZH results, and closure of the event-level analysis and Born-suppressed production.
Fixed-order and event-generation consistency
Table IV compares integration and event-generation results at 13.6 TeV with NNPDF3.1. Standalone ggxy, the Powheg stage-2 integration, and the generated event weights are statistically compatible at LO and NLO. These routes share matrix-element ingredients and test their integration and event-generation interfaces. The LO event weights reproduce the fixed-order prediction at all seven scale points. At NLO, the standalone calculation covers
| stable | ||
| LO, ggxy | 7.0324(10) | 70.705(7) |
| LO, Powheg stage 2 | 7.0306(19) | |
| LO, event weights | 7.037(31) | |
| LO, MCFM | 71.111(18) | |
| NLO, ggxy | 13.043(4) | |
| NLO, Powheg grids | 13.031(7) | |
| NLO, event weights | 13.05(6) |
Table IV. Cross sections in fb at 13.6 TeV with NNPDF3.1 and . is summed over with GeV. The scale bands are the seven-point envelopes.
seven scale points; separate regression checks compare all nine stored scale weights and the PDF and variations with reference event-weight results. The automated suite passes 66/66 generator-level checks, which do not resolve the independent MCFM comparison below.
The independent MCFM stable- calculation lies 0.57% above ggxy, well beyond the quoted integration uncertainties. The same relative difference is observed with NNPDF3.1 and PDF4LHC21 and at both collision energies. Its origin remains unresolved. This comparison therefore does not establish agreement between the independent LO implementations, and no correction is applied to either prediction.
Agreement with the published ggxy_ggZH calculation
Table V compares our fixed-order results with Tables 1 and 3 of Ref. [24]. The charged-lepton, neutrino, and stable- cross sections agree within the quoted Monte Carlo precision at both 13 and 13.6 TeV. The stable- results are also consistent with the updated NLO predictions of Ref. [2].
Event-level and Born-suppression closure
At parton level, the production LO events reproduce the fixed-order LO spectrum bin by bin, with at 13.6 TeV and at 13 TeV. On the NNPDF3.1 validation sample, the integrated Rivet cross section after showering matches the LHE normalization for every scale and PDF weight. Every event is assigned an STXS bin, and the official Stage 1.3 implementation agrees with the independent bin assignment used in our validation.
We test the Born suppression with 50,000 events generated with identical seeds, once without suppression and once with GeV. The inclusive cross sections agree with pull ; all nine scale variations and the PDF and uncertainties close within the validation tolerances. After showering, the suppressed and unsuppressed samples agree in the Stage 1.3 fine bins and in , with and , respectively.
The unsuppressed validation sample has limited statistics above about GeV. We therefore also compare the larger Born-suppressed production sample at parton level with fixed-order NLO. At 13.6 TeV, the ratio above 400 GeV spans 0.95–1.03, with Monte Carlo uncertainties of 4–17%. Between 125 and 400 GeV, the deviations span to . A similar structure appears without Born suppression: the ratio in the 125–150 GeV bin is 0.946(18) in the unsuppressed validation sample and 0.945(7) in the production sample. With , the 13.6 TeV deviations over 125–400 GeV span to . Their persistence without suppression and dependence on point to higher-order matching effects in the POWHEG hardest emission, rather than a bias introduced by the Born-suppression factor. Agreement with fixed order in the tail is a consistency check, not a substitute for suppressed-versus-unsuppressed closure.
Predictions
Inclusive rates
Table VI gives the fixed-order cross sections and their uncertainty budgets. At 13.6 TeV, the charged-lepton channel increases from fb to fb, corresponding to . The neutrino channel receives a nearly identical relative NLO correction. Its cross section is approximately twice that of the charged-lepton channel, consistent with the branching-fraction expectation for the two -boson decay modes.
The NLO correction reduces the seven-point scale band from to . At NLO, the PDF and uncertainties are 1.77% and 1.41%, while the top-quark mass-scheme band is . Scale variation remains the largest inclusive uncertainty. The mass-scheme band is comparable to its downward variation and substantially larger than the PDF and uncertainties. The 13 TeV results show the same hierarchy.
Differential structure at NLO and NLO+PS
Figure 2 shows the main differential distributions at 13.6 TeV. The NLO correction is strongly kinematics dependent and cannot be represented by a single inclusive factor. In , the differential factor is 3.0 in the low-mass 200–225 GeV interval and decreases to 1.5–1.6 above 600 GeV. The behavior is opposite in the boosted transverse-momentum tails: is 1.7–2.1 below 400 GeV and rises to 2.6 at 600–800 GeV and 4.0 at 800–1000 GeV.
Figure 2. Differential predictions for at 13.6 TeV with PDF4LHC21: (a) , (b) , (c) , (d) , and (e) jet multiplicity for GeV. Fixed-order LO and NLO bands are seven-point scale envelopes. NLO+PS uses the Monash tune; its bands combine the scale envelope, PDF, and uncertainties in quadrature, without the mass-scheme or matching variations. The lower panels show ratios to LO for , (logarithmic scale), and , to NLO for , and to the central NLO+PS prediction for jet multiplicity.
The NLO real emission also separates the - and Higgs-boson transverse momenta. At LO they are back to back and satisfy ; at NLO, a hard jet can recoil predominantly against the Higgs boson while the remains softer. Consequently the tail receives an even larger correction, with reaching 21 at 800–1000 GeV. This hard-recoil topology was also observed in merged 0 + 1-jet LO studies [46].
NLO+PS leaves the spectrum essentially unchanged relative to fixed-order NLO and produces additional hardening in the and tails. By contrast, the nonzero
The spectrum is only LO accurate in the fixed-order NLO calculation and is more sensitive to matching. NLO+PS exceeds the fixed-order prediction by a factor of 1.6–1.9 between 60 and 300 GeV. The Sudakov-generated real contribution can receive the local enhancement discussed in Section III A, whereas the finite remnant does not. Varying changes this partition and probes the resulting higher-order matching dependence (§ 2 D). The differential fixed-order factors at 13 and 13.6 TeV are statistically compatible; the energy comparison and its Monte Carlo precision are discussed in Section A.
Top-quark mass-scheme dependence
Figure 3 shows the fixed-order NLO prediction in the scheme relative to the on-shell result for the three choices of defined in § 2 C. The inclusive NLO envelope is , but the effect is much more pronounced near the top-pair threshold and in the boosted region. Percentage shifts quoted below refer to cross sections integrated over each stated region; their ranges span the three choices of . The shift is 0 to +5 % for 200 300 GeV and to % for GeV. Near the top-pair threshold, the finite-order mass-scheme predictions also differ in shape: the largest single-bin /on-shell ratio between 300
Figure 3. Top-quark mass-scheme dependence at fixed order for at 13.6 TeV with PDF4LHC21. Shown is the /on-shell cross-section ratio for , , and in (a) , (b) , and (c) . Markers show NLO results with Monte Carlo errors from the two integrations combined in quadrature; the grey band is the envelope of the corresponding LO ratios.
and 350 GeV is 1.8, over the three choices.
For the cross section integrated over GeV, the shift is to , whereas for GeV it is to . The corresponding shift for GeV is to . The mass-scheme dependence therefore becomes a leading uncertainty in the boosted phase space targeted by the high- STXS bins.
STXS Stage 1.3 predictions and uncertainties
Table VII gives the NLO+PS Stage 1.3 fine-bin predictions at 13.6 TeV. Scale, PDF, and uncertainties are evaluated directly from the event weights, while matching and shower uncertainties use the variations of § 2 D. No NLO+PS sample is generated. For the central-Higgs bins, the fixed-order NLO mass-scheme ratio in each interval is transferred to all jet bins in that interval. The forward-Higgs bin instead receives the inclusive fixed-order ratio, without a forward-region selection (§ 2 C).
| (GeV), jets | (fb) | MC (%) | Scale (%) | PDF+ (%) | (%) | Match.+PS (%) |
| 0.476 | 1.2 | 7.5 | ||||
| 0–75, 0 jets | 0.985 | 1.2 | 2.2 | |||
| 75–150, 0 jets | 2.84 | 0.3 | 2.2 | |||
| 150–250, 0 jets | 1.17 | 0.4 | 2.2 | |||
| 250–400, 0 jets | 0.0975 | 1.0 | 2.7 | |||
| 400–600, 0 jets | 0.00358 | 4.5 | 3.6 | |||
| , 0 jets | 9.21e-05 | 29.2 | 6.1 | |||
| 0–75, 1 jet | 0.778 | 1.3 | 2.2 | |||
| 75–150, 1 jet | 2.03 | 0.4 | 2.2 | |||
| 150–250, 1 jet | 1.13 | 0.4 | 2.2 | |||
| 250–400, 1 jet | 0.137 | 0.9 | 2.6 | |||
| 400–600, 1 jet | 0.00712 | 3.4 | 3.3 | |||
| , 1 jet | 0.000455 | 13.1 | 5.1 | |||
| 0–75, jets | 0.61 | 1.3 | 2.2 | |||
| 75–150, jets | 1.47 | 0.6 | 2.2 | |||
| 150–250, jets | 1.05 | 0.5 | 2.3 | |||
| 250–400, jets | 0.226 | 0.9 | 2.6 | |||
| 400–600, jets | 0.0196 | 2.8 | 3.5 | |||
| , jets | 0.00163 | 7.8 | 5.4 |
Table VII. NLO+PS (Monash) cross sections at 13.6 TeV in the STXS Stage 1.3 fine bins of , summed over . The table gives the relative Monte Carlo uncertainty, seven-point scale envelope, PDF+ uncertainty (), top-mass scheme band from the fixed-order transfer of § 2 C, and the matching+shower uncertainty of § 2 D. The mass-scheme transfer uses the -interval ratio for central-Higgs bins and the inclusive ratio for the forward-Higgs bin. The 13 TeV predictions are given in Table XIV.
In the 16 bins with Monte Carlo uncertainties below 5%, scale uncertainties span 9–24%, PDF+ uncertainties 2.2–7.5%, top-mass scheme bands 2–50%, and matching+shower uncertainties 3–18%. For asymmetric bands, these ranges refer to the larger absolute excursion.
Scale variation is largest in the low- bins and the forward-Higgs region. With the stated transfer prescription, the mass-scheme band is the largest uncertainty in every statistically precise central-Higgs jet bin from GeV onward, reaching 32–50%. The common mass-scheme band across jet multiplicities is imposed by this prescription; its jet dependence has not been calculated.
Rescaled LO+PS versus NLO+PS
An inclusive factor corrects the normalization of an LO+PS sample but leaves its kinematic shapes unchanged. We test whether this rescaling reproduces the NLO+PS prediction in boosted and jet-dependent observables, and isolate the dependence on the shower starting scale. Differential reweighting of experiment-inspired samples is studied separately in Section IX.
Earlier studies compared Born-only showers with merged zero- and one-jet matrix elements and found sizeable shower-start dependence in hard-recoil observables [46]. The two Powheg+Pythia 8 configurations in that study used a restricted start or a vetoed power shower, rather than the unvetoed power shower considered below. Related studies of and examined the sensitivity of pair-recoil spectra to showering and matching [47, 48]. Here we use the NLO+PS calculation of Section II as the reference and quantify the differences in STXS Stage 1.3 bins.
Comparison setup and shower-start prescriptions
The comparison uses charged-lepton events at each order and collision energy, with PDF4LHC21 and the hard-process scales of Table I. We use the Monash tune as the baseline and repeat the comparison with CP5 and A14. The NLO events are showered with the Powheg veto of Section IV A. For each LO shower-start prescription, the same Born events are showered without that veto. These variations isolate shower-start effects rather than reproduce the experiment-inspired vetoed sources of Sections IX A and IX G.
| Parameter | Value |
| 13.6 TeV and 13 TeV (; beam energies 6800 and 6500 GeV) | |
| , | 91.1876 GeV, 2.4955 GeV |
| , | 80.385 GeV, GeV |
| 125 GeV, on shell | |
172.5 GeV, on-shell scheme (mtscheme 0); variations in § 2 C | |
| 1 | |
| 15–150 GeV | |
| , the invariant mass of the system | |
PDF4LHC21_40_pdfas [32] (LHAPDF [33] 93300), also for LO | |
0.118, from the PDF set (alphas_from_pdf 1) | |
| 250 GeV (Section III A) |
Table I. Parameters, identical at both energies except for the beam energy.
For the colorless Born final state, the default ISR setting SpaceShower:pTmaxMatch = 0 allows evolution up to the kinematic limit [28, 35]. This is the unvetoed “power shower” [27]; it is not restricted by the event’s SCALUP value. Setting pTmaxMatch = 1 instead uses SCALUP = as the evolution limit, optionally reduced by pTmaxFudge. We also test smooth damping of hard emissions without restricting the available phase space [28]. The settings
applied to ISR and FSR are given in Table VIII. The limits refer to the Pythia 8 evolution variable, not directly to the reconstructed leading-jet transverse momentum.
| Label | Settings | Shower evolution limit |
| default | pTmaxMatch = 0 | kinematic limit (colorless Born state) |
| power | pTmaxMatch = 2 | kinematic limit |
pTmaxMatch = 1 | SCALUP = | |
pTmaxMatch = 1, pTmaxFudge = 0.5 | ||
pTmaxMatch = 1, pTmaxFudge = 0.25 | ||
| damped | pTmaxMatch = 0, pTdampMatch = 1 | kinematic limit, damping scale |
Table VIII. LO+PS shower-start prescriptions, applied to the same Born events without a Powheg veto. The listed settings are applied in both the SpaceShower and TimeShower namespaces. The damped configuration uses pTdampFudge = 1 and is evaluated only at 13.6 TeV with Monash.
We multiply LO+PS by , using the inclusive cross sections of the corresponding event samples: at 13.6 TeV and 1.856 at 13 TeV. With this common normalization, ratios of rescaled LO+PS to NLO+PS cross sections are also ratios of the corresponding cross-section fractions. The rescaled LO+PS scale band combines its seven-point shape variations with the inclusive NLO normalization uncertainty in quadrature, as defined in Section V A. The shower-start dependence is studied separately from these hard-process scale variations.
Boosted spectra and STXS bins
The unvetoed default start produces a growing excess in the boosted region (Figure 4 and Table IX). Across Monash, CP5, and A14, the ratio of cross-section fractions above GeV is 5.4–6.1 at 13.6 TeV, rising to 14.2–17.4 above 600 GeV. The corresponding ranges at 13 TeV are 5.6–6.4 and 15.7–18.5. In the 18 central STXS fine bins, the excess is concentrated in bins with jets and increases with . For Monash it reaches 19.9(16) above 600 GeV with at least two jets, while the zero-jet bins below 400 GeV lie below NLO+PS (ratios 0.73–0.97).
Figure 4. Rescaled LO+PS divided by NLO+PS at 13.6 TeV with Monash for the unvetoed prescriptions of Table VIII. Left: the 18 central STXS Stage 1.3 fine bins, excluding the forward-Higgs bin. Markers show the ratios with Monte Carlo errors; boxes show the rescaled LO+PS scale bands of Section V A. Open markers identify bins with an NLO+PS Monte Carlo uncertainty above 10%. Right: the same comparison per bin, with scale bands and Monte Carlo errors. The blue band is the NLO+PS seven-point scale envelope relative to its central value. The power configuration coincides with the default and is not shown separately.
| LO shower start | GeV | ||||
| default = power | 1.03 | 1.26 | 2.27 | 6.11 | 17.35 |
| 1.02 | 1.19 | 1.68 | 2.33 | 2.52 | |
| 1.02 | 1.11 | 1.22 | 1.24 | 1.19 | |
| 1.02 | 1.04 | 0.99 | 0.90 | 0.79 | |
| damped | 1.02 | 1.18 | 1.66 | 2.79 | 4.33 |
Table IX. Ratio of LO+PS to NLO+PS cross-section fractions above each threshold at 13.6 TeV with Monash. With the inclusive rescaling used here, these are also ratios of rescaled LO+PS to NLO+PS cross sections in the selected regions. The shower-start prescriptions are those of Table VIII; the fractions include all Higgs rapidities.
Reducing or damping the shower start improves some distributions but does not reproduce all STXS bins. Among bins with an NLO+PS Monte Carlo uncertainty below 10%, the ratios span 0.75–1.50 for the start and 0.85–4.56 for the damped start. For the default start, the rescaled LO+PS scale band contains the NLO+PS central prediction in 5 of 16 bins; the start gives the largest count, 12 of 16. For the default start, none of the 5 statistically precise bins with jets and GeV is covered. This remains true for that start when using either times the full LO+PS scale envelope or correlated variations of the LO+PS shape and NLO normalization. The hard-process scale band therefore does not account for the observed shower-start dependence.
Origin of the excess and implications for corrections
Table X separates three stages: the independent fixed-order calculation, the generated events before showering, and the same events after showering. The LO events reproduce the fixed-order spectrum. The NLO events already include the Powheg emission and differ from fixed-order NLO by the matching effects discussed in § 2.3. Comparing each event sample before and after showering isolates the additional shower effect.
| GeV | |||||
| LO, fixed order | 81.8 | 29.0 | 3.15 | 0.175 | 0.00909 |
| LO events, parton level | 81.9 | 28.9 | 3.16 | 0.179 | 0.00859 |
| NLO, fixed order | 80.4 | 27.8 | 3.26 | 0.206 | 0.0132 |
| NLO events, parton level | 80.4 | 29.2 | 3.45 | 0.211 | 0.0132 |
| NLO+PS | 80.7 | 30.1 | 3.82 | 0.250 | 0.0167 |
| LO+PS, start default = power | 82.7 | 38.1 | 8.67 | 1.53 | 0.289 |
| LO+PS, start | 81.9 | 31.3 | 3.76 | 0.225 | 0.0132 |
| LO+PS, start damped | 82.3 | 35.6 | 6.34 | 0.698 | 0.0723 |
| NLO+PS / NLO events | 1.00 | 1.03 | 1.10 | 1.19 | 1.26 |
| LO+PS / LO events, start default = power | 1.01 | 1.32 | 2.74 | 8.52 | 33.67 |
| LO+PS / LO events, start | 1.01 | 1.24 | 2.02 | 3.25 | 4.89 |
| LO+PS / LO events, start | 1.00 | 1.16 | 1.47 | 1.74 | 2.30 |
| LO+PS / LO events, start | 1.00 | 1.08 | 1.19 | 1.26 | 1.53 |
| LO+PS / LO events, start damped | 1.01 | 1.23 | 2.01 | 3.89 | 8.41 |
Table X. Cross-section fractions (in %) above each threshold at 13.6 TeV with Monash, including all Higgs rapidities. Upper part: fixed-order predictions, generated events before showering, and showered predictions. Lower part: after-shower divided by before-shower fractions, using the same underlying event sample in each ratio.
Above GeV, showering raises the NLO event fraction by about a fifth, whereas the unvetoed default LO shower increases it by more than a factor of eight (Table X). The start reduces the LO shower enhancement to a level closer to the NLO one, but agreement in this projection does not imply agreement in the jet bins. Figure 5 shows the same contrast in the normalized and spectra: the default LO shower hardens the tails much more than the matched NLO shower.
Figure 5. Normalized (a) and (b) distributions after showering divided by those before showering, using the same events at 13.6 TeV with Monash. Curves show NLO events with the matched shower and LO events with the unvetoed default and starts. The plotted errors combine the two histogram Monte Carlo errors in quadrature; the covariance between the shared-event distributions is not included.
The large default-start excess is thus generated when showering the Born events; it is not removed by correcting the inclusive normalization. This behavior is consistent with extrapolating soft and collinear radiation kernels into hard-recoil regions, where they do not replace the process-specific real-emission matrix element [46, 47]. The NLO+PS reference includes that matrix element for the hardest emission, although its matching and shower uncertainties remain relevant (§ 2 D).
Invariant mass and recoil
For the tested configurations at 13.6 TeV, the largest relative change in the normalized distribution after showering is , evaluated in bins containing more than 1% of the before-shower cross section. This does not by itself exclude an improvement in from a weight depending only on : correlations between the two variables allow such a weight to change the marginal distribution. A mass-only weight does, however, leave the conditional recoil distribution at fixed unchanged. It therefore cannot repair a discrepancy in that conditional distribution. This distinction motivates the tests of radiation-sensitive reweighting variables in Section IX.
Reweighting LO+PS to NLO+PS
An event-dependent correction can transfer NLO+PS information to existing LO+PS samples without repeating their detector simulation. We construct corrections for specified CMS- and ATLAS-style samples in the charged-lepton and neutrino channels. Each correction has its own trained ensemble and calibration, together with target-theory variations and uncertainties in the transfer. Independent samples test the corrected boson distributions and STXS yields, while the source population in hard-recoil regions determines the principal limitation.
CMS-style source and target samples
We first define the CMS-style construction; the ATLAS-style application is presented in Section IX G. The CMS-style source and target each contain events per decay channel at 13.6 TeV. The CMS-style LO source follows the public production card [49]: Born-only generation with NNPDF3.1 NNLO PDFs [34], , and SCALUP . The charged-lepton mass window is GeV. Showering uses Pythia 8 with CP5 [40] and the CMS Powheg interface [50]: veto = 1, pThard = 0, pTdef = 1, vetoCount = 100, and pTmaxMatch = 2 for ISR and FSR. The shower explores the full phase space, but emissions harder than SCALUP in the Powheg measure are vetoed. This is distinct from the unvetoed shower-start comparisons of Section VIII.
The NLO+PS target uses the production setup of Section III, including Born suppression with GeV and the hard-process scale . We select its NNPDF3.1 central weight and use the same CP5 tune and shower-interface settings as for the source. Both the perturbative order and the hard-process scale therefore change between source and target.
The and Higgs momenta are taken from their final event-record copies. The decay products are ordered in transverse momentum before final-state radiation. Jets use anti- clustering [44] with and GeV, excluding all and Higgs decay descendants. The STXS jet count has no rapidity restriction; and are instead the leading jets within . This latter definition is used for the jet inputs and .
Whole event files are assigned in a fixed pseudo-random order to disjoint training, calibration, and evaluation samples. Calibration and evaluation each use events per class; the remaining files form the training pool. A further confirmation sample, described in Section IX E, tests the residual-uncertainty prescription.
Phase-space domain
The source shower veto suppresses hard recoil, although it does not impose a strict bound on reconstructed jet transverse momentum. In the charged-lepton samples, events with account for 9% of the NLO+PS cross section selected by GeV, but only 1.4% of LO+PS. For GeV, the shares are 25% and 1.5%, respectively. The source is severely depleted, but not empty, in this region.
We restrict the learned correction to or no jet within the acceptance used to define . The same restriction is applied to both training classes. Only 0.49% of the inclusive NLO+PS cross section and 0.21% of LO+PS lie outside this domain in the charged-lepton evaluation sample, but the boosted selections above enhance their importance. Outside the domain, the nominal additional weight remains one, retaining LO+PS; a separate variation describes the population difference (Section IX D). This separates the correction in well-populated regions from the uncertainty associated with sparse hard recoil.
Nominal correction
We estimate the NLO+PS-to-LO+PS density ratio with gradient-boosted classifiers [51, 52, 53], using the 17 inputs in Table XI. With equal total training weight for the two classes, the classifier odds estimate the density ratio in the training domain. Twenty classifiers use separately drawn subsets of events per class from the training pool and different random seeds; their subsets may overlap. LO events carry equal training weights, while NLO generator weights enter the loss with negative values set to zero. This positive-weight training approximation is assessed through closure against the NLO+PS reference.
| Inputs | Kinematic information |
| , , | mass, recoil, longitudinal boost |
| , (0, 1, 2, ) | STXS kinematics |
| azimuthal recoil geometry | |
| virtuality | |
| , , | leading jet |
| , , | second jet |
| , | decay products |
Table XI. The 17 classifier inputs, evaluated with the definitions of Section IX A. Missing-jet inputs take fixed values outside their physical ranges. In the neutrino channel, the decay-product inputs refer to neutrinos and is replaced by .
The classifiers use HistGradientBoostingClassifier in scikit-learn 1.2.2 [54], with binary cross entropy, learning rate 0.05, at most 63 leaves, at least 500 events per leaf, and regularization 1.0. Training stops after at most 1000 iterations or 20 iterations without improvement on a 20% internal validation subset. These validation events belong to the training pool, not to the independent evaluation sample.
The nominal ensemble combines the odds geometrically,
On the calibration sample, the ratio of normalized target and reweighted-source distributions defines a multiplicative factor . Starting from 25 GeV bins, adjacent bins are merged until both distributions have relative statistical uncertainties below 2%. For the charged-lepton channel, the 23 factors range from 0.93 to 1.03. The product is normalized on the calibration sample to give unit-mean shape weights, with weight one outside the domain. The overall cross section is set separately to the NLO value. Thus the learned correction changes the kinematics of LO+PS, not its nominal inclusive normalization.
Uncertainty prescription
We distinguish uncertainty in the transfer from theory uncertainty in the target prediction. The transfer prescription combines classifier variation, residual leading-jet dependence, and a binwise residual term. These components are treated as separate nuisance sources and combined in quadrature, not as a product of simultaneous weight shifts.
a. Classifier ensemble. Each calibrated classifier supplies alternative event weights that are propagated through the analysis selection. Following [55], the resulting histogram bins are standardized across the twenty predictions before principal-component analysis. Components accounting for 99% of the standardized variance are retained and transformed back to bin contents. Each defines a correlated shape variation; the compression is performed at histogram level, rather than assigning independent classifier errors to events.
b. Leading-jet dependence. Inside the domain, the calibration-sample closure ratio defines the alternatives and . Its bins are merged until the relative statistical uncertainty is below 5%; the charged-lepton values range from 0.85 to 1.06. Outside the domain, the alternatives multiply the nominal unit weight by the target-to-source ratio of cross-section shares, or its inverse, in bins of . This ratio ranges from 1.1 below 150 GeV to 20 above 600 GeV in the charged-lepton calibration sample. It describes the depleted hard-recoil population in these bins; it does not establish recovery of its multidimensional kinematics.
c. Residual differences. The ensemble spread does not capture biases shared by its members. Residual differences on the evaluation sample motivate an additional uncertainty of 1% of each final histogram bin content, treated as uncorrelated between bins. This empirical term is added after histogramming, not supplied as an event weight. Its validation on a separate confirmation sample and the remaining exceptions are reported in Section IX E.
d. Target-theory variations. For each seven-point NLO scale variation, the calibration target defines in cells of and , where and are the varied and nominal generator weights. These factors multiply the corrected source weights. PDF4LHC21 PDF and variations [32] are transferred in the same cells from the -event NLO+PS sample carrying those weights. The improved perturbative scale description therefore comes from the NLO+PS reference; the transfer uncertainties quantify the additional approximation made by reweighting.
Validation and improvement over inclusive rescaling
On the independent evaluation samples, reweighting brings the tested boson and jet observables substantially closer to NLO+PS than inclusive rescaling alone (Figure 6 and Figure 7). These shape comparisons are restricted to the training domain. Agreement within the transfer uncertainty holds over most of the tested bins, including and , which are not direct classifier inputs. The effective sample size after reweighting is about 90% of the event count in both decay channels.
Figure 6. CMS-style charged-lepton evaluation sample inside the training domain. Upper panels: normalized NLO+PS (solid), LO+PS (dotted), and reweighted LO+PS (dashed) distributions. The NLO+PS band shows shape-only scale variations. Lower panels: ratios to NLO+PS; bars combine source and target statistical errors, while the gray band combines ensemble, leading-jet, residual, and reweighted-source statistical uncertainties. Arrows indicate LO+PS ratios outside the displayed range. Adjacent presentation bins are merged to contain at least 1000 source events. Panels (b) and (g) are not direct classifier inputs.
Figure 7. As Figure 6, for the independently trained CMS-style neutrino correction; panel (h) shows .
A further events per class, disjoint from training, calibration, and the sample used to choose the residual term, test the uncertainty prescription. For bins with at least 100 source events, pulls combine the source and target Monte Carlo errors with the transfer uncertainty; target scale uncertainty is excluded. The diagnostic requires at most 5% of bins beyond two standard deviations and none beyond three. The charged-lepton correction satisfies it: 3.8% of 521 bins exceed two, with a maximum pull of 2.9. For neutrinos, 4.0% of 424 bins exceed two, but one reaches 3.6; the strict criterion is therefore not satisfied in
that channel. These checks support the uncertainty prescription for most tested bins, rather than unconditional closure under arbitrary selections.
Figure 8 and Figure 9 compare the full-phase-space predictions and uncertainty sizes on the calibration samples, retaining rate changes in every variation. Here the LO+PS band varies the original LO scale weights at fixed nominal ; it is not the shape-plus-NLO-normalization prescription of § 2 A. The reweighted band combines transported NLO scale variations with the three transfer components. Neither band includes Monte Carlo statistics. PDF, , mass-scheme, and matching variations are not included in this comparison, so these bands are not the complete theory budget of § 2.
Figure 8. CMS-style charged-lepton calibration sample in the full phase space, with all nominal predictions normalized to and divided by NLO+PS. Variations retain their changes in the total rate. Hatched: LO+PS with the original LO seven-point scale envelope, symmetrized to its larger excursion. Shaded: reweighted LO+PS with transported NLO scale, ensemble, leading-jet, and 1% residual uncertainties; outside the domain its nominal correction is one. The reweighted band retains separate upward and downward excursions. Light: the symmetrized NLO+PS scale envelope. No Monte Carlo statistical uncertainty is included.
Figure 9. As Fig. 8, for CMS-style .
The median uncertainties in Table XII are smaller after reweighting, while the central predictions move towards NLO+PS, except in the highest bins discussed below. The transferred prediction is narrower in 76–100% of the bins across these observables; the large outside-domain variation limits the improvement in the boosted bins above about 400 GeV. The neutrino results give similar median sizes, within one percentage point.
| Observable | LO+PS LO scale | Reweighted NLO scale + transfer | Transfer |
| 30 | 17 | 2.8 | |
| 29 | 17 | 2.6 | |
| 30 | 13 | 1.4 | |
| STXS | 28 | 17 | 2.1 |
Table XII. Median relative uncertainty sizes (%) in the charged-lepton calibration sample.
The high- tail remains a limitation. Above 450 GeV, the in-domain reweighted cross section exceeds NLO+PS by 5–6% for charged leptons and 10–12% for neutrinos on the evaluation and confirmation samples. For the full-phase-space calibration comparison, the corresponding ratios are 0.97 and 0.99, whereas LO+PS gives 3.7 and 3.7, respectively. The LO discrepancy lies far outside its scale band (−24/+34% for charged leptons). The near-unity full-phase-space ratios partly reflect compensation between an in-domain excess and an outside-domain deficit. They demonstrate improvement in this projection, not closure of its separate recoil components.
TABLE XII. Median relative uncertainty sizes (%) in the charged-lepton calibration sample. For each bin, a component’s size is its larger absolute excursion; components are combined in quadrature. “Transfer” includes ensemble, leading-jet, and residual terms and is already included in the reweighted column. These medians summarize the bands in Figure 8, not symmetric confidence intervals.
The tail shows the missing hard-recoil population directly. Inside the training domain, the reweighted cross section above GeV is 0.97 of NLO+PS for charged leptons (1.00 for neutrinos), whereas LO+PS is 1.5 (1.6) times NLO+PS. In the full phase space, however, the reweighted prediction falls below NLO+PS above about 350 GeV and reaches 0.69 (0.69) above 525 GeV. Above GeV, 25% of the charged-lepton NLO+PS cross section has : the Higgs boson recoils against a jet harder than , while the boson remains softer. The vetoed source contains only 1.5% of such events, and their weight remains one outside the domain, so reweighting cannot restore this contribution. The outside-domain variation, binned in , reaches NLO+PS in this bin (upper edges 1.04 and 1.00). LO+PS lies closer to NLO+PS there (1.07 and 1.08) only because its excess inside the domain offsets the same missing events.
CMS-style neutrino channel
The neutrino correction uses independent -event source and target samples, with GeV. Training, calibration, and uncertainty construction are performed separately with the same prescription, except for the PDF and ratios, which are taken from the charged-lepton sample. Its inputs use generator-level neutrino momenta; it is not obtained by substituting reconstructed missing transverse momentum into the chargedlepton correction. The parallel results above test the method in the two decay channels while retaining their distinct residual limitations.
ATLAS-style sources
We test the same reweighting prescription with ATLAS-style sources and separately trained corrections for both decay channels. The source follows the public ATLAS job options [56]: Born-only ggxy-ggZH generation with PDF4LHC21, , and SCALUP , followed by the A14 shower [39] with the ATLAS Powheg veto. The target uses the NLO production setup, with , the PDF4LHC21 central weight, and A14 showering. This tests the prescription under different hard-process and shower settings, rather than transferring the CMS-trained models.
Training, calibration, the domain restriction, and uncertainty construction follow Sections IX C and IX D. The 1% residual is retained without adjustment and was fixed before examining the ATLAS-style evaluation samples. The charged-lepton samples contain 3.5 million source and 3.4 million target events; the neutrino samples contain 3.3 and 3.3 million. The smaller training pools imply expected pairwise overlaps of 66–71% and 75–76%, respectively, compared with 14–24% for the CMS-style ensembles.
On the independent evaluation samples, fewer than 5% of the tested bins have pulls larger than two standard deviations in magnitude (Table XIII). Each channel nevertheless has one bin beyond three, with maximum absolute pulls of 3.1 and 3.4; neither satisfies the strict diagnostic of Section IX E. The pull means are and , with standard deviations 0.92 and 0.98. These diagnostics test the combined transfer prescription, including the leading-jet and residual terms, not the ensemble spread alone. The in-domain tail ratios in Table XIII and the distributions in Section B show the performance on events excluded from training and calibration.
| Quantity | ||
| Tested bins | 515 | 421 |
| (%) | 3.5 | 3.8 |
| Maximum | 3.1 | 3.4 |
| LO+PS : LO scale | 26 | 26 |
| Reweighted: NLO scale + transfer | 17 | 17 |
Table XIII. ATLAS-style corrections in the charged-lepton and neutrino channels. The upper block uses independent evaluation samples inside the training domain; pulls are evaluated in bins with at least 100 source events. is the reweighted-to-NLO+PS ratio for GeV, with Monte Carlo errors. The lower block gives median relative uncertainty sizes (%) in on the full-phase-space calibration samples, defined as in Table XII.
In the full-phase-space calibration comparison (Figure 10), the median uncertainty decreases from 26% to 17% for charged leptons and from 26% to 17% for neutrinos. The transfer contribution has median size 2.1% and 2.2%, respectively. These compare the original LO scale band at fixed with transported NLO scale variations plus transfer uncertainties, not the complete theory budget. Above GeV, the full-phase-space reweighted-to-NLO+PS ratios are 1.00 and 0.96, compared with 3.1 and 3.0 for inclusive rescaling alone. As in the CMS-style comparison, agreement in this projection does not establish closure of its separate recoil components.
Figure 10. ATLAS-style calibration samples in the full phase space: (a–c) charged leptons and (d–f) neutrinos, for , , and the STXS bins. Normalization and uncertainty definitions follow Figure 8. The reweighted bands combine transported NLO scale variations with transfer uncertainties; Monte Carlo statistics and the other target-theory uncertainties are not included. The highest bin retains an uncovered deficit in both channels.
The high- deficit is more restrictive. In the last displayed bin, beginning at 525 GeV, the full-phase-space reweighted predictions are only 0.66 and 0.64 of NLO+PS for charged leptons and neutrinos. Even the upper edges of the combined scale-plus-transfer bands reach only 0.92 and 0.91. Thus the present prescription does not cover the NLO+PS central prediction in that bin. The outside-domain variation must not be interpreted as a general correction for the depleted hard-recoil population.
Application and model release
For samples matching one of the stated source configurations, its channel-specific correction and alternative weights can be propagated to reconstructed analysis histograms using their generator-level inputs. The ensemble variations are then compressed for the chosen histograms, and the binwise residual is added separately. The calibration normalization is retained when applying analysis selections, so that the correction can change acceptance as well as shape.
The method improves the tested boson distributions and STXS yields relative to inclusive rescaling, but does not replace NLO+PS generation for analyses sensitive to the depleted hard-recoil configurations. In particular, agreement in a jet-inclusive projection does not validate a selection that isolates or vetoes those configurations. The uncovered ATLAS-style Higgs-tail deficit provides an explicit example. Applications with different source settings, selections, or binning require their own closure checks; the empirical residual is not a universal error bound.
For each source configuration and decay channel, the accompanying data record (Section C) contains the twenty ONNX classifiers, calibration and theory-transfer tables, and an inference script for the nominal and alternative event weights. It also provides the histogram-level principal-component procedure, reference events with expected weights, and a model card specifying the source, target, inputs, domain, normalization, and residual prescription.
Recommendations and conclusions
For the tested LO+PS samples, an inclusive factor does not reproduce the NLO+PS prediction in boosted and jet-dependent production. Using the public ggxy-ggZH NLO+PS implementation, we have evaluated inclusive charged-lepton and neutrino predictions at 13 and 13.6 TeV, together with differential and STXS Stage 1.3 predictions for the charged-lepton channel. At 13.6 TeV, the NLO correction reduces the inclusive seven-point scale band from to . Its kinematic dependence and the contribution of hard real radiation also change the distributions used in LHC measurements.
Scale variation remains the largest inclusive uncertainty, but the hierarchy changes in the boosted region. At 13.6 TeV, the top-quark mass-scheme band is largest in the central-Higgs STXS bins from GeV onward with Monte Carlo errors below 5%, reaching 32–50% (Section VII D). These bands use fixed-order mass-scheme ratios transferred to NLO+PS, with the same ratio for every jet multiplicity in a given interval. They are not direct NLO+PS variations. Scale, PDF+, mass-scheme, and matching and shower effects must therefore be assessed separately rather than represented by the inclusive scale band alone.
The LO+PS comparison isolates a strong dependence on the treatment of shower radiation. With the unvetoed default start at the kinematic limit, inclusive rescaling overestimates the cross section above GeV by factors of 14.2–17.4 at 13.6 TeV across the three tunes. Restricting or damping the shower start reduces the excess, but none of the tested prescriptions describes all statistically precise STXS bins within its scale band (Section VIII). This controlled comparison is distinct from the experiment-inspired vetoed sources used for reweighting.
For the specified CMS- and ATLAS-style sources at 13.6 TeV, we provide separately trained charged-lepton and neutrino corrections with calibration and transfer uncertainties. They improve the tested boson distributions and STXS yields over inclusive rescaling in most of the populated phase space. In the charged-lepton calibration comparisons, the median STXS uncertainty changes from 28% to 17% for CMS-style samples and from 26% to 17% for ATLAS-style samples. These compare the original LO scale band at fixed with transported NLO scales plus transfer uncertainties, not the complete theory budget (Sections IX E and IX G). The improvement therefore combines the NLO reference prediction with an additional, separately evaluated uncertainty in the transfer.
Independent evaluation and confirmation samples test this prescription beyond calibration. The charged-lepton CMS-style correction satisfies the stated pull diagnostic on its confirmation sample, while CMS-style neutrinos and both ATLAS-style channels retain bins beyond its strict threshold. For ATLAS-style samples, the 1% residual was fixed before evaluation. The hard-recoil limitation is also visible directly in the full-phase-space Higgs spectrum. In the last ATLAS-style bin, beginning at 525 GeV, the corrected charged-lepton and neutrino predictions are 0.66 and 0.64 of NLO+PS; even the upper band edges reach only 0.92 and 0.91. Near-unity ratios in other projections do not remove this deficit or establish closure of the separate recoil components.
For new simulation, we recommend NLO+PS generation, with Born suppression to improve statistics in the boosted region and the uncertainty components of § 2. The amplitude cache makes large theory-weight lists practical: at 450 weights, it reduces the measured generation time by a factor of 26, while cached and uncached samples retain identical event kinematics and alternative weights in the closure tests (Section III E). At 13.6 TeV, Born suppression increases the effective statistics above GeV by a factor of 2.1 at fixed event count, without changing the weighted prediction (Section III B).
For existing samples matching one of the specified source configurations, its calibrated correction provides an alternative to inclusive rescaling without repeating detector simulation. Its use requires the documented event-weight variations, histogram-level ensemble treatment, and residual uncertainty, with closure checked for the analysis selection and binning. Agreement in a jet-inclusive projection does not validate selections that isolate the source’s depleted hard-recoil region. Such applications should use NLO+PS events rather than rely on large weights or compensating discrepancies in different recoil regions.
The public production package provides the generation setup, reference results, and validation scripts [29]. The reweighting deliverable consists of source- and channel-specific ONNX ensembles, calibration and theory-transfer tables, inference and uncertainty procedures, reference events, and model cards defining the source, target, and applicability (Sections C and IX H). Together, these provide an NLO+PS reference for new production and a calibrated correction with stated limitations for reuse of existing LO+PS samples.
Acknowledgments
This work was supported in part by the U.S. Department of Energy under Grant No. DE-SC0010072. We thank the authors of ggxy_ggZH for reviewing and merging the amplitude cache. We thank Weitao Wang (ATLAS) for providing the ATLAS generation configuration and Sitian Qian (CMS) for fruitful discussions. The authors used OpenAI Codex and Anthropic Claude Code under their direction for manuscript language polishing, software-development assistance, monitoring computational runs and recording their execution times. We acknowledge OpenAI for providing complimentary access to ChatGPT through the Chat-GPT for Academic Researchers program. The software used in this work was validated by the authors, with the ggxy components validated by the ggxy authors. The authors retain full responsibility for the scientific content and conclusions.
Distributions at 13 TeV
Figure 11 shows the distributions of Figure 2 at . The differential fixed-order factors are statistically compatible with those at : , , and for , , and , respectively. This comparison concerns the fixed-order corrections, not equality of the matched distributions at the two energies.
Figure 11. As Figure 2, at 13 TeV: differential distributions of (PDF4LHC21) in (a) , (b) , (c) , (d) and (e) the number of jets; fixed-order LO and NLO with their seven-point scale bands, NLO+PS (Monash) with the scale band combined in quadrature with the PDF and uncertainties.
The qualitative matching and shower trends are similar, but the precision differs between bins. The largest central deviation of the parton-level Powheg spectrum from fixed-order NLO over is , in the – bin. The fixed-order integration error dominates the uncertainty of this ratio; that bin is statistically compatible with its counterpart and does not establish an energy-dependent matching effect.
Table XIV gives the STXS Stage 1.3 fine bins at ; the matching variations were made at only (Table III). The top-mass scheme ratios at are compatible with those at ( over the inclusive cross section and the six ranges for the three ).
| (GeV), jets | (fb) | MC (%) | Scale (%) | PDF+ (%) | (%) |
| 0.407 | 1.3 | 7.9 | |||
| 0–75, 0 jets | 0.912 | 1.2 | 2.2 | ||
| 75–150, 0 jets | 2.61 | 0.3 | 2.2 | ||
| 150–250, 0 jets | 1.07 | 0.4 | 2.3 | ||
| 250–400, 0 jets | 0.0882 | 1.0 | 2.8 | † | |
| 400–600, 0 jets | 0.00342 | 4.4 | 4.0 | ||
| > 600, 0 jets | 0.000151 | 20.0 | 5.0 | ||
| 0–75, 1 jet | 0.718 | 1.3 | 2.2 | ||
| 75–150, 1 jet | 1.85 | 0.4 | 2.2 | ||
| 150–250, 1 jet | 1.02 | 0.4 | 2.3 | ||
| 250–400, 1 jet | 0.121 | 0.9 | 2.7 | † | |
| 400–600, 1 jet | 0.00599 | 3.5 | 3.7 | ||
| > 600, 1 jet | 0.000279 | 15.1 | 5.4 | ||
| 0–75, jets | 0.55 | 1.4 | 2.2 | ||
| 75–150, jets | 1.31 | 0.6 | 2.2 | ||
| 150–250, jets | 0.936 | 0.5 | 2.3 | ||
| 250–400, jets | 0.197 | 0.9 | 2.7 | † | |
| 400–600, jets | 0.0163 | 2.5 | 3.7 | ||
| > 600, jets | 0.00128 | 8.6 | 5.3 |
Table XIV. As Table VII, at 13 TeV: NLO+PS (Monash) cross sections in the STXS Stage 1.3 fine bins in fb, with the relative MC error, the seven-point scale envelope, the PDF+ uncertainty and the top-mass scheme band. †: the MC error of this band exceeds 5% (250–400 GeV: 8%); the 13.6 TeV value of Table VII is compatible with it.
ATLAS-style closure
Figure 12 and Figure 13 show the closure of the ATLAS-style corrections on their evaluation samples, in the layout of Figure 6.
Figure 12. As Figure 6, for the ATLAS-style charged-lepton correction.
Figure 13. As Figure 6, for the ATLAS-style neutrino correction; panel (h) shows .
Reproduction
All results in this paper are generated with the public package [29], release v3.0.0 (commit 91a9aa8, DOI 10.5281/zenodo.24316545). The package specifies the software versions and commits listed in Table XV and includes workflow scripts, reference data tables, and automated checks. The driver script production/paper.sh defines the run parameters, job allocations, event statistics, random seeds, and input configurations for seven datasets: 13600 and 13000 (charged-lepton production at the two energies), vv13600 and vv13000 (the corresponding neutrino channels), verify (validation suite using NNPDF3.1 at 13.6 TeV and Born-suppression tests), atlas13600 (ATLAS-style charged-lepton reweighting at 13.6 TeV), and common (software build recipes). The CMS-style charged-lepton source and its CP5 NLO+PS target reshower belong to 13600, rather than a separate dataset. The command
production/paper.sh list <set> lists the runs, while the subcommands show, inputs, and outputs document job configurations. Each execution records the originating Git commit hash. Random seeds are deterministic functions of the job index, and shower seeds are mapped directly from corresponding LHE seeds, making the runs reproducible across batch systems.
Software and fixed-order integration settings
The exact software environment used for the production is summarized in Table XV. Build scripts fix the repository commits, apply the required patches, and ensure that the modified libraries bundled with ggxy take precedence over libraries in the host environment. The standalone ggxy LO calculation evaluates scalar one-loop integrals with OneLOop [57].
a. Fixed-order integration. The fixed-order calculations use the following independent integration setups; these details are recorded here to make the numerical results reproducible. ggxy Monte Carlo.: Integration programs based on ggxy examples for and on-shell production, configured with the parameters in Table I. A single integration run evaluates the seven scale variations of Section V A. Differential distributions in , , , and auxiliary observables use 25 GeV binning aligned with STXS bin boundaries. Sampling statistics per energy comprise points at LO (after warmup points for phase-space grid optimization) and points at NLO ( warmup points); independent seeds are combined via unweighted averaging.
Powheg stage 2.: Fixed-order integration in Powheg-Box-V2 evaluated with Born-only matrix elements (bornonly 1), with one run per scale point for each of 8 independent seeds and calls on a common importance grid ( calls across three grid iterations). Results are combined via variance-weighted averaging.
MCFM.: Standalone LO calculation of with stable bosons using MCFM’s analytic one-loop amplitudes (gg_zh), providing a cross-check independent of Recola and ggxy expansions. A dedicated patch adds an input switch to isolate the gluon-fusion contribution.
Parton-shower settings
For NLO+PS, SpaceShower:pTmaxMatch = 2 and TimeShower:pTmaxMatch = 2 open the shower to the full phase space, while the Pythia 8 Powheg hooks veto harder emissions. The nominal POWHEG:* settings are nFinal = 3, veto = 1, pThard = 0, pTemt = 0, pTdef = 1, emitted = 0, and vetoCount = 100. Other POWHEG:* settings keep their defaults; MPI and QED emissions are not vetoed. The matching variations replace pThard = 0 or pTdef = 1 as described in Section V D.
The nominal tune is Monash (Tune:pp = 14). A14 uses Tune:pp = 21. CP5 is initialized from Tune:pp = 14 and uses the parameters of Table 3 of Ref. [40]: MPI ecmPow = 0.03344, bProfile = 2, pT0Ref = 1.41, coreRadius = 0.7634, coreFraction = 0.63, ecmRef = 7000, ColourReconnection:range = 5.176, and SpaceShower:rapidityOrder = on. The CP5 shower/MPI PDF is selected with PDF:pSet = LHAPDF6:NNPDF31_nnlo_as_0118. In the CP5 card, ISR, FSR, MPI, and SigmaProcess use with two-loop running; these Pythia 8 settings do not replace the hard-event weights in the LHE file. The CP5 total and elastic cross-section inputs are 101.037/22.08 mb at 13.6 TeV and 100.309/21.89 mb at 13 TeV.
Event-generation settings
The Powheg grids are integrated with 16 independent seeds. Per seed, the importance grid uses ncall1 = 2·10^{4} and itmx1 = 3, the integration uses ncall2 = 4·10^{4} and itmx2 = 2, and the radiation upper-bound step uses nubound = 2 · 10^{4}. The folding parameters are (foldcsi, foldy, foldphi) = (1, 1, 1) and negative weights are retained (withnegweights 1). The grid-level cross section is the error-weighted mean of the 16 stage-2 integrations. Events are generated seed by seed from the merged grids.
LO samples use the same process with bornonly 1 and LOevents 1; eight integration seeds are used before generating unweighted Born events. The Born-suppression patch is inactive when bornsuppfact is omitted.
The amplitude cache can be disabled with GGXY_AMPCACHE=0. Its entry limit is controlled with GGXY_AMPCACHE_MAX; the default is entries, corresponding to about 100 MB. Closure tests compare cached and uncached jobs with identical random seeds. The stored event kinematics and alternative weights are identical. The only observed difference is an internal #rwgt comment at the relative level from removing and restoring the power of ; this affects only a later rwl_add re-reweighting of the LHE file. The runs use the following steps:
Builds:
generator/build.sh <prefix>(ggxy, Powheg-Box-V2 with ggxy--ggZH and the Born-suppression patch, the ggxy fixed-order programs) andshower/build.sh <prefix>(Pythia 8 driver, Rivet analyses including the STXS routine).
Fixed order:
generator/run/fo.sh <lo|nlo> <out> <first seed> <n seeds> <points> <optimization points> <PDF> <fo-ggllh|fo-ggvvh|fo-ggzh> <sqrt(s)>; combination withanalysis/fo_combine.py.
Powheg LO at the seven scale points:
run/powheg_lo.sh; combination withanalysis/powheg_stat.py.
Grids:
run/powheg_grids.shwith the card settings of the run (PDF, beam energy,bornsuppfact, orbornonly/LOeventsfor LO).
Events, one seed per job:
run/powheg_events.shwith the weight listcards/pwg-rwl-pdf4lhc21.dat; cross sections and variations withanalysis/lhe_xsec.py(per job, then combined).
Cache closure:
run/cache_ab.shandanalysis/compare_lhe.py.
MCFM:
mcfm/run_mcfm.shandanalysis/mcfm_xsec.py.
Showers:
shower/run/shower_batch.sh(shower/run/shower.shper event file); exact sums over files and jobs withanalysis/yoda_summary.py;analysis/compare_orders.pyfor Section VIII.
Checks:
generator/check.py(with--paper-energy 13600|13000against Ref. [24]) andshower/check_shower.py.
Table XVI lists the runs, with their statistics and use. The vv sets repeat the decayed- fixed-order, grid, event, and shower runs for with independent seeds and the same statistics, including the 13.6 TeV matching variations and reweighting chain; the scan is replaced by one check at , at 13.6 TeV only. Stable- and code-validation runs are shared. The CMS-style charged-lepton source uses NNPDF3.1, , and CP5 (Section IX A). Its grids-lo-cms, events-lo-cms, and table-lo-cms runs are recorded under 13600. The table-nlo-cp5 target reuses the production NLO events with the NNPDF3.1 central weight and a CP5 reshower; it is not an additional independently generated NLO sample. The atlas13600 set generates LO events with the scale, checks the scale weights against independent LO integrations, and reshowers production NLO events with A14, for earlier five- and six-input classifiers. The tables-atlas-zll-v1 and tables-atlas-zvv-v1 sets produce the ATLAS-style samples of Section IX G, and ensemble-atlas-zll-v2 and ensemble-atlas-zvv-v2 their classifiers, tests, and models.
| Run | Statistics | Use |
| ggxy LO / NLO | / points | fixed order, 7 points, |
| ggxy stable LO / NLO | / points | comparison with [24], Table I |
| Powheg LO (stage 2) | 7 points 8 seeds | code comparison |
| NLO grids | 16 seeds, GeV | NLO events |
| LO grids | 8 seeds | LO events |
| NLO events | weighted, 155 weights | NLO+PS |
| LO events | unweighted, 155 weights | LO+PS |
| Cache closure | events, cache on and off | Section III E |
| MCFM LO, stable | 7 points | independent LO check |
| ggxy LO / NLO, | as on-shell, | scheme (§ 2 C) |
| At 13.6 TeV only: | ||
| NLO, NNPDF3.1 | 16 seeds; events, unweighted and GeV | Born suppression (§ 2.3) |
| NLO, | 16 seeds; events | matching (§ 2 D) |
| ggxy stable LO / NLO, | as on-shell, | comparison with [2] |
| CMS-style LO+PS, NNPDF3.1, CP5 | events | source (Section IX A) |
| NLO+PS, NNPDF3.1 weight, CP5 | reused events | target (Section IX A) |
Table XVI. Production and validation samples. The first block is generated at both 13.6 and 13 TeV with PDF4LHC21. The second block is specific to 13.6 TeV, with PDF choices shown where they differ. The CMS-style target reuses the nominal NLO events rather than adding a new independent sample.
The regression suite compares fixed-order integrations and generated event weights with stored reference results. The corresponding numerical comparisons and their statistical precision are reported in § 2.3; they do not identify the compiler as the cause of any run-to-run difference.
a. Computing cost. The base charged-lepton production runs took about 1360 CPU hours at 13.6 TeV and 1260 at 13 TeV (user and system time of the successful jobs), with 39 % spent on the parton-shower runs (three tunes, five LO+PS starts). The NLO events with 155 weights took 90 CPU hours with the amplitude cache (Section III E). The variations made at 13.6 TeV only (PDF members, cache scan, , Powheg-veto variants, damped LO+PS start, stable- runs) added 790 CPU hours, and the NNPDF3.1 validation set 670. These accounting categories exclude the experiment-inspired reweighting sample preparation added to Table XVI.
b. Data. The data record [61] (DOI 10.5281/zenodo.24315535) holds the LHE files of the generated samples of Table XVI with their event weights (155 per nominal production event: scale, PDF4LHC21 and NNPDF3.1 members), the Powheg grids, with which more events of any sample can be generated without a new integration, the per-event tables, weights, and reusable models of the reweighting study (Section IX), including the neutrino samples and their own ggzh-bdt17-zvv-r5.tgz (charged leptons: ggzh-bdt17-zll-r5.tgz), the ATLAS-style samples with their own ggzh-bdt17-atlas-zll-r5.tgz and ggzh-bdt17-atlas-zvv-r5.tgz, and the analysis outputs: the Rivet histograms of every shower and parton-level sample with all weight variations, and every result file behind the numbers of this paper. Each file has a checksum; production/data_release.sh of the package assembles the record from the products of production/paper.sh. The model cards of the four models are also included directly in the source repository under results/13600/rwt-bdt17/.
Conventions
a. Weight identifiers (cards/pwg-rwl-pdf4lhc21.dat, 155 weights):
b. Event-weight normalization. For a sample of generated events with signed weights , the cross section is evaluated as
For Born-suppressed samples the event-weight magnitude contains the compensating factor . Event weights are stored to five significant figures.
Shower weights. Pythia 8 names: isrRedHi, fsrRedHi, isrRedLo, fsrRedLo ( factor , ), isrDefHi, fsrDefHi, isrDefLo, fsrDefLo (, 2); in Rivet the LHE weights appear as MUR1.0 MUF2.0, . . . and the PDF weights by their identifier.
Fixed-order normalization. Standalone ggxy outputs differential distributions per lepton flavor in nanobarns; yields are scaled by 3 to represent . Seeds with identical sample sizes are combined via unweighted averaging, with statistical uncertainties evaluated from the standard error of the mean. Fixed-order Powheg cross sections represent the variance-weighted mean of stage-2 integration results across all seeds.
STXS event-record conventions. The Higgs boson is identified as its last copy before decay. The hard-process boson is followed to its last copy before decay and defines . Jets are built from stable final-state particles after removing descendants of the Higgs boson and of the decaying boson, including products of decays. No jet rapidity requirement is applied. The 30 GeV jet threshold is used throughout; the alternative 25 GeV threshold available in the STXS routine is not used. The routine treats separately, while charged-lepton and neutrino decays enter the gluon-fusion bins used here.
References
- [1]D. de Florian et al. (LHC Higgs Cross Section Working Group), Handbook of lhc higgs cross sections: 4. deciphering the nature of the higgs sector (2016), arXiv:1610.07922 [hep-ph].DOI
- [2]B. Campillo Aveleira, L. Chen, J. Davies, G. Degrassi, P. P. Giardino, R. Gröber, G. Heinrich, S. Jones, M. Kerner, J. Schlenk, M. Steinhauser, and M. Vitti, gg → ZH: updated predictions at NLO QCD, SciPost Phys. Comm. Rep. , 16 (2026), arXiv:2508.09905 [hep-ph].DOI
- [3]A. Karlberg et al. (LHC Higgs Working Group), Ad interim recommendations for the Higgs boson production cross sections at √s = 13.6 TeV (2024), arXiv:2402.09955 [hep-ph].arxiv.org/abs/2402.09955
- [4]C. Englert, M. McCullough, and M. Spannowsky, Gluon-initiated associated production boosts Higgs physics, Phys. Rev. D 89, 013013 (2014), arXiv:1310.4828 [hep-ph].DOI
- [5]B. Hespel, F. Maltoni, and E. Vryonidou, Higgs and Z boson associated production via gluon fusion in the SM and the 2HDM, JHEP 2015 (06), 065, arXiv:1503.01656.arxiv.org/abs/1503.01656
- [6]L. Altenkamp, S. Dittmaier, R. V. Harlander, H. Rzehak, and T. J. E. Zirke, Gluon-induced Higgs-strahlung at next-to-leading order QCD, JHEP 2013 (02), 078, arXiv:1211.5015 [hep-ph].DOI
- [7]A. Hasselhuhn, T. Luthe, and M. Steinhauser, On top quark mass effects to gg → ZH at NLO, JHEP 2017 (01), 073, arXiv:1611.05881 [hep-ph].
- [8]J. Davies, G. Mishima, and M. Steinhauser, Virtual corrections to gg → ZH in the high-energy and large-m_t limits, JHEP 2021 (03), 034, arXiv:2011.12314 [hep-ph].DOI
- [9]L. Alasfar, G. Degrassi, P. P. Giardino, R. Gröber, and M. Vitti, Virtual corrections to qg → ZH via a transverse momentum expansion, JHEP 2021 (05), 168, arXiv:2103.06225 [hep-ph].DOI
- [10]R. Bonciani, G. Degrassi, P. P. Giardino, and R. Gröber, Analytical Method for Next-to-Leading-Order QCD Corrections to Double-Higgs Production, Phys. Rev. Lett. 121, 162003 (2018), arXiv:1806.11564 [hep-ph].DOI
- [11]L. Chen, G. Heinrich, S. P. Jones, M. Kerner, J. Klappert, and J. Schlenk, ZH production in gluon fusion: two-loop amplitudes with full top quark mass dependence, JHEP 2021 (03), 125, arXiv:2011.12325 [hep-ph].DOI
- [12]G. Wang, X. Xu, Y. Xu, and L. L. Yang, Next-to-leading order corrections for gg → ZH with top quark mass dependence, Phys. Lett. B 829, 137087 (2022), arXiv:2107.08206 [hep-ph].DOI
- [13]L. Chen, J. Davies, G. Heinrich, S. P. Jones, M. Kerner, G. Mishima, J. Schlenk, and M. Steinhauser, ZH production in gluon fusion at NLO in QCD, JHEP 2022 (08), 056, arXiv:2204.05225 [hep-ph].DOI
- [14]G. Degrassi, R. Gröber, M. Vitti, and X. Zhao, On the NLO QCD corrections to gluon-initiated ZH production, JHEP 2022 (08), 009, arXiv:2205.02769 [hep-ph].DOI
- [15]R. V. Harlander, A. Kulesza, V. Theeuwes, and T. Zirke, Soft gluon resummation for gluon-induced Higgs Strahlung, JHEP 2014 (11), 082, arXiv:1410.0217 [hep-ph].DOI
- [16]G. Das, C. Dey, M. C. Kumar, and K. Samanta, Soft gluon resummation for gluon fusion ZH production, Phys. Rev. D 113, 014024 (2026), arXiv:2501.10330 [hep-ph].DOI
- [17]P. Banerjee, C. Dey, N. Koirala, M. C. Kumar, and V. Pandey, Associated ZH production in gluon fusion process at NLO+NLL (2026), arXiv:2605.24131 [hep-ph].arxiv.org/abs/2605.24131
- [18]J. Davies, D. Grau, K. Schönwald, M. Steinhauser, D. Stremmer, and M. Vitti, Two-loop QCD corrections to ZH and off-shell Z boson pair production in gluon fusion, JHEP 2026 (03), 200, arXiv:2509.07072 [hep-ph].arxiv.org/abs/2509.07072
- [19]J. Davies, D. Grau, K. Schönwald, M. Steinhauser, and D. Stremmer, Three-loop corrections to gg → ZH in the large top quark mass limit, JHEP 2026 (02), 191, arXiv:2512.00156 [hep-ph].arxiv.org/abs/2512.00156
- [20]J. Davies, K. Schönwald, M. Steinhauser, and D. Stremmer, ggxy: A flexible library to compute gluon-induced cross sections, Comput. Phys. Commun. 320, 109933 (2026), https://gitlab.com/ggxy/ggxy-release, arXiv:2506.04323 [hep-ph].arxiv.org/abs/2506.04323
- [21]P. Nason, A new method for combining NLO QCD with shower Monte Carlo algorithms, JHEP 2004 (11), 040, arXiv:hep-ph/0409146.DOI
- [22]S. Frixione, P. Nason, and C. Oleari, Matching NLO QCD computations with parton shower simulations: the POWHEG method, JHEP 2007 (11), 070, arXiv:0709.2092.DOI
- [23]S. Alioli, P. Nason, C. Oleari, and E. Re, A general framework for implementing NLO calculations in shower Monte Carlo programs: the POWHEG BOX, JHEP 2010 (06), 043, arXiv:1002.2581 [hep-ph].DOI
- [24]J. Davies, K. Schönwald, M. Steinhauser, and D. Stremmer, gg → zh at nlo matched to parton showers with ggxy and powheg (2026), arXiv:2603.15762 [hep-ph].arxiv.org/abs/2603.15762
- [25]G. Aad et al. (ATLAS), Measurements of WH and ZH production in the H → bb̄ decay channel in pp collisions at 13 TeV with the ATLAS detector, Eur. Phys. J. C 81, 178 (2021), arXiv:2007.02873 [hep-ex].
- [26]A. Tumasyan et al. (CMS), Measurement of simplified template cross sections of the Higgs boson produced in association with W or Z bosons in the H → bb̄ decay channel in proton-proton collisions at √s = 13 TeV, Phys. Rev. D 109, 092011 (2024), arXiv:2312.07562 [hep-ex].
- [27]T. Plehn, D. Rainwater, and P. Z. Skands, Squark and gluino production with jets, Phys. Lett. B 645, 217 (2007), arXiv:hep-ph/0510144.DOI
- [28]R. Corke and T. Sjostrand, Improved Parton Showers at Large Transverse Momenta, Eur. Phys. J. C 69, 1 (2010), arXiv:1003.2384 [hep-ph].DOI
- [29]L. Zhang, ggxy-ggzh-ampcache: reproduction package for gg → zh at nlo+ps and lo+ps reweighting to nlo+ps, https://github.com/De-Cristo/ggxy-ggZH-ampcache (2026), version 3.0.0.
- [30]S. Actis, A. Denner, L. Hofer, J.-N. Lang, A. Scharf, and S. Uccirati, Recola: Recursive computation of one-loop amplitudes, Comput. Phys. Commun. 214, 140 (2017), arXiv:1605.01090.DOI
- [31]A. Denner, S. Dittmaier, and L. Hofer, Collier: a fortran-based complex one-loop library in extended regularizations, Comput. Phys. Commun. 212, 220 (2017), arXiv:1604.06792.DOI
- [32]R. D. Ball et al. (PDF4LHC Working Group), The PDF4LHC21 combination of global PDF fits for the LHC Run III, J. Phys. G 49, 080501 (2022), arXiv:2203.05506 [hep-ph].DOI
- [33]A. Buckley, J. Ferrando, S. Lloyd, K. Nordström, B. Page, M. Rüfenacht, M. Schönherr, and G. Watt, LHAPDF6: parton density access in the LHC precision era, Eur. Phys. J. C 75, 132 (2015), arXiv:1412.7420 [hep-ph].DOI
- [34]NNPDF Collaboration, Parton distributions from high-precision collider data, Eur. Phys. J. C 77, 663 (2017), arXiv:1706.00428 [hep-ph].
- [35]C. Bierlich et al., A comprehensive guide to the physics and usage of PYTHIA 8.3, SciPost Phys. Codeb. 2022, 8 (2022), arXiv:2203.11601 [hep-ph].arxiv.org/abs/2203.11601
- [36]A. Buckley, P. Ilten, D. Konstantinov, L. Lönnblad, J. Monk, W. Pokorski, T. Przedzinski, and A. Verbytskyi, The HepMC3 event record library for Monte Carlo event generators, Comput. Phys. Commun. 260, 107310 (2021), arXiv:1912.08005 [hep-ph].DOI
- [37]C. Bierlich, A. Buckley, J. M. Butterworth, C. Gutschow, L. Lonnblad, T. Procter, P. Richardson, and Y. Yeh, Robust independent validation of experiment and theory: Rivet version 4 release note, SciPost Phys. Codeb. , 36 (2024), arXiv:2404.15984 [hep-ph].arxiv.org/abs/2404.15984
- [38]P. Skands, S. Carrazza, and J. Rojo, Tuning PYTHIA 8.1: the Monash 2013 Tune, Eur. Phys. J. C 74, 3024 (2014), arXiv:1404.5630 [hep-ph].DOI
- [39]ATLAS Collaboration, ATLAS Pythia 8 tunes to 7 TeV data, Tech. Rep. ATL-PHYS-PUB-2014-021 (CERN, 2014).
- [40]CMS Collaboration, Extraction and validation of a new set of CMS PYTHIA8 tunes from underlying-event measurements, Eur. Phys. J. C 80, 4 (2020), arXiv:1903.12179.DOI
- [41]S. Mrenna and P. Skands, Automated Parton-Shower Variations in Pythia 8, Phys. Rev. D 94, 074005 (2016), arXiv:1605.08352 [hep-ph].DOI
- [42]LHC Higgs Working Group, STXS classification: Rivet routine SimplifiedTemplateCrossSections (2026), commit 38eea57.
- [43]LHC Higgs Working Group, Simplified template cross sections, Stage 1.3 (recommended binning for Run 3) (2026), accessed 2026-10-11; a document of Stage 1.3 is in preparation.
- [44]M. Cacciari, G. P. Salam, and G. Soyez, The anti-k_t jet clustering algorithm, JHEP 2008 (04), 063, arXiv:0802.1189 [hep-ph].DOI
- [45]J. Baglio, F. Campanario, S. Glaus, M. Mühlleitner, J. Ronca, and M. Spira, gg → HH : Combined uncertainties, Phys. Rev. D 103, 056002 (2021), arXiv:2008.11626 [hep-ph].DOI
- [46]S. Amoroso et al., Les Houches 2019: Physics at TeV Colliders: Standard Model Working Group Report, in 11th Les Houches Workshop on Physics at TeV Colliders: PhysTeV Les Houches (2020) arXiv:2003.01700 [hep-ph].
- [47]S. Jones and S. Kuttimalai, Parton Shower and NLO-Matching uncertainties in Higgs Boson Pair Production, JHEP 2018 (02), 176, arXiv:1711.03319 [hep-ph].DOI
- [48]S. Alioli, F. Caola, G. Luisoni, and R. Röntsch, ZZ production in gluon fusion at NLO matched to parton-shower, Phys. Rev. D 95, 034042 (2017), arXiv:1609.09719 [hep-ph].DOI
- [49]CMS Collaboration, POWHEG card gghz_hanythingj_nnpdf31_13p6tev_m125_-vleptonic.input, genproductions repository, https://github.com/cms-sw/genproductions (2022), bin/Powheg/production/Run3/13p6TeV/Higgs/ggHZ_HanythingJ_NNPDF31_13p6TeV, commit 7625d99.
- [50]CMS Collaboration, CMSSW configuration pythia8powhegemissionvetosettings_cfi.py, https://github.com/cms-sw/cmssw (2024), configuration/Generator/python, release CMSSW_14_0_19.
- [51]K. Cranmer, J. Pavez, and G. Louppe, Approximating Likelihood Ratios with Calibrated Discriminative Classifiers (2015), arXiv:1506.02169 [stat.AP].arxiv.org/abs/1506.02169
- [52]A. Rogozhnikov, Reweighting with Boosted Decision Trees, J. Phys. Conf. Ser. 762, 012036 (2016), arXiv:1608.05806 [physics.data-an].DOI
- [53]A. Andreassen and B. Nachman, Neural Networks for Full Phase-space Reweighting and Parameter Tuning, Phys. Rev. D 101, 091901 (2020), arXiv:1907.08209 [hep-ph].DOI
- [54]F. Pedregosa et al., Scikit-learn: Machine Learning in Python, J. Machine Learning Res. 12, 2825 (2011), arXiv:1201.0490 [cs.LG].arxiv.org/abs/1201.0490
- [55]C. Palmer and B. Kronheim, Improving statistical precision in Monte Carlo samples with negative weights via reweighting and uncertainty quantification (2025), arXiv:2510.16217.DOI
- [56]ATLAS Collaboration, Public Run-3 gg → zh production and shower job options, https://gitlab.cern.ch/atlas-physics/pmg/mcjoboptions (2026), datasets 602411 and 603335; accessed 19 October 2026.
- [57]A. van Hameren, OneLOop: For the evaluation of one-loop scalar functions, Comput. Phys. Commun. 182, 2427 (2011), arXiv:1007.4716 [hep-ph].DOI
- [58]J. Campbell and T. Neumann, Precision Phenomenology with MCFM, JHEP 2019 (12), 034, arXiv:1909.09117 [hep-ph].DOI
- [59]M. Cacciari, G. P. Salam, and G. Soyez, FastJet User Manual, Eur. Phys. J. C 72, 1896 (2012), arXiv:1111.6097 [hep-ph].DOI
- [60]N. Berger et al., Simplified Template Cross Sections – Stage 1.1 and 1.2, SciPost Phys. Comm. Rep. , 15 (2026), arXiv:1906.02754 [hep-ph].
- [61]L. Zhang and C. Palmer, Data record for “loop-induced gg → zh at nlo+ps for lhc simulations”: event samples, powheg grids, analysis outputs and lo+ps → nlo+ps reweighting models (2026).












