betaFeedback

WorksPhysics and Astronomy

Positive-weight moment matching at NNLO and beyond

Abstract

Parton-shower simulations at state-of-the-art perturbative accuracy carry many negative event weights, which dilute their statistical power. For many purposes, this makes them unusable. We formulate a method for the generation of positive-weighted event samples which reproduce, up to a controllable approximation, suites of IR-safe observables with the highest perturbative accuracy of any available calculation. This method is arbitrarily differential in phase space, and free from unphysical artefacts such as binning-dependence. Applied to a leading-order Pythia sample, the method gives simultaneous NNLO accuracy for Drell–Yan and Z+jetZ+\text{jet} observables in a positive-weighted event sample, for the first time. The method generalises immediately to higher orders of perturbation theory, higher multiplicities, and other processes, observables, and event generators.

The High-Luminosity LHC programme requires modelling uncertainties that match its anticipated experimental precision, and large simulated samples for detector simulation and unfolding. At present, event samples with the best calculable perturbative accuracy are computationally prohibitive to process at scale, because of their large fraction of negative event weights. Negative weights contribute irreducibly to the variance, diluting the statistical power of a given simulation budget, potentially catastrophically [1].

Efforts to address this bottleneck include mitigation strategies within established NLO matching methods, MC@NLO [2] ([3, 4, 5, 6]) and POWHEG [7, 8] ([9, 10]), alternative matching methods [11, 12, 13, 14, 15], and post-hoc methods to repair or remove negative weights within event samples [16, 17, 18, 19, 20, 21, 22, 23]. Despite these innovations, no existing method combines state-of-the-art perturbative accuracy with strictly positive event-weights in a single hadronised sample. Existing NNLO matching schemes [24, 25, 26, 27, 28, 29] all carry negative weights.

For processes with resolved jets in the final state, the ‘NNLO revolution’ of two-loop calculations and local subtraction algorithms [30, 31, 32, 33, 34] implemented in parton-level Monte Carlo codes [35, 36] now enables state-of-the-art precision calculations with up to two additional final-state emissions modelled at lower perturbative orders. In parallel, the ‘NLL revolution’ has led to parton showers which achieve unprecedented formal accuracy in their all-orders resummation [37, 38, 39, 40, 41, 42, 43, 44]. For collider phenomenology, these cannot yet be combined.

The characterisation of positive probability measures from their complete sequence of polynomial moments is a classical problem [45, 46]. On a compact domain, a measure is uniquely determined by its complete collection of polynomial moments, known as the Hausdorff moment problem [47, 48]. With finitely many moment constraints, the problem is generally underdetermined. Among all reweightings of an input sample that reproduce those moments, the principle of maximum entropy [49] selects the unique event-by-event differential reweighting that achieves this while changing the sample the least. Provided that the system of constraints is consistent and can be solved, any solution is positive by construction, and so preserves the positivity of any positive-definite input sample.

Maximum-entropy reweighting of parton shower samples (‘priors’) to reproduce precision moments, with strictly positive weights, was introduced for e+e−e^+e^- event shapes [50, 51] to augment the leading-logarithmic accuracy achieved by the shower algorithm to reproduce the state-of-the-art NNLL′^\prime resummation available analytically, and has since been applied to resummed Drell–Yan spectra [52].

We reformulate this approach with constraints from parton-level Monte Carlo codes. Because the constraints are calculated independently of their application to showered, hadronised event samples, there is no requirement to reformulate NNLO subtraction calculations in a manifestly positive way, nor to interface NNLO codes to parton showers directly. As well as new runs of NNLO codes, moments can be computed from cached event samples [53, 54], or approximately from the interpolation grids used for PDF fitting [55, 56, 57, 58]. There is no material limit to the number of observables, perturbative order, multiplicity or number of higher moments that can be incorporated, beyond their availability to within the requisite statistical uncertainty. We return to this point below.

We emphasise that the reweighting process is systematically improvable. When further multiplicities, observables, or moments become known to higher perturbative accuracies or statistical precision, existing samples can be refined quickly and cheaply to match. This is a paradigm shift from the current approach, in which the accuracy of large samples is locked-in when they are generated.

We demonstrate the method using an LO+PS Pythia Drell–Yan event sample, achieving simultaneous moment-matched NNLO accuracy for inclusive observables and for Z+Z+ jet observables. Rather than saturate the dimensionality of the fixed-order phase space, we match only a subset of the observables of interest, in order to validate the method on held-out observables not used in the reweighting. We benchmark the computational cost and statistical power of the resulting sample against alternative methods.

Method. Let the prior {Φi,pi}\{\Phi_i,p_i\} consist of events Φi\Phi_i with weights pi>0p_i>0, ∑ipi=1\sum_i p_i=1, and let m(Φ)\mathbf{m}(\Phi) be a vector of measurement functions, mi≡m(Φi)\mathbf{m}_i\equiv\mathbf{m}(\Phi_i), whose fixed-order expectations μ\boldsymbol{\mu} are calculated to Nk^{k}LO. Among all normalized weights qi≥0q_i\geq0 that reproduce μ\boldsymbol{\mu}, we take the one closest to the prior in relative entropy [59, 60],

min⁡q∑iqiln⁡qipiwith∑iqimi=μ,∑iqi=1.(1)\min_{\mathbf{q}}\sum_i q_i\ln\frac{q_i}{p_i}\quad\text{with}\quad\sum_i q_i\mathbf{m}_i=\boldsymbol{\mu},\quad\sum_i q_i=1. \tag*{(1)}

The objective is strictly convex, so a solution, where one exists, is unique, and has the closed form

qi=pieλ⋅miZ(λ),Z(λ)=∑jpjeλ⋅mj,(2)q_i=\frac{p_i e^{\boldsymbol{\lambda}\cdot\mathbf{m}_i}}{Z(\boldsymbol{\lambda})},\qquad Z(\boldsymbol{\lambda})=\sum_j p_j e^{\boldsymbol{\lambda}\cdot\mathbf{m}_j}, \tag*{(2)}

i.e. is manifestly positive for every multiplier vector λ\boldsymbol{\lambda}. We call qi/piq_i/p_i the tilt. The multipliers follow from a strictly convex dual problem, which may be solved cheaply by Newton iteration (End Matter). As for any multiplicative reweighting, the prior must have support throughout the target region, or the multipliers diverge. For the Drell–Yan results reported here, the tilts qi/piq_i/p_i lie between 0.97 and 1.79 for the NLO Z+jZ+j input and between 0.96 and 3.95 for the NNLO one.

As constraints we use Chebyshev polynomial moments of observables v(Φ)v(\Phi), mapped monotonically to [−1,1][-1,1] by uvu_v, calculated event-by-event. The map uvu_v is chosen linear for rapidity variables, logarithmic for steep spectra, and the Breit–Wigner coordinate for resonances (End Matter). A window wvw_v restricts each input calculation to its region of theoretical accuracy. Each constrained observable contributes a tower of moments, whose targets blend the prior outside the window with the fixed order inside it,

m⊃{Tn(uv)}n=1Nv,\mathbf{m}\supset\left\{T_n(u_v)\right\}_{n=1}^{N_v},
μa=(1−rv)⟨ma⟩p,1−wv+rv⟨ma⟩FO,wv,(3)\mu_a=(1-r_v)\langle m_a\rangle_{p,1-w_v}+r_v\langle m_a\rangle_{\mathrm{FO},w_v}, \tag*{(3)}

where TnT_n are the Chebyshev polynomials, aa indexes the moments, rvr_v is the fixed-order share of the cross-section inside the window, and the averages are weighted by 1−wv1-w_v over the prior, and wvw_v over the fixed-order moments. Only expectations in the span of the imposed measurement functions are fixed directly. Other predictions depend on the finite-moment approximation and on conditional distributions inherited from the prior. Where wvw_v vanishes the target is the prior’s own moment, where it is one the fixed-order moment. Since

ln⁡(qipi)=∑vPv(uv)−ln⁡Z(4)\ln\left(\frac{q_i}{p_i}\right)=\sum_v P_v(u_v)-\ln\mathcal{Z} \tag*{(4)}

with PvP_v a polynomial of degree NvN_v, the tilt is smooth by construction, and depends only on the constrained observables. For fixed values of these observables, the conditional distribution of the remaining event degrees of freedom remains that of the prior. Where a correlation is itself a perturbative target, observables can be constrained jointly through cross-moments (End Matter).

Uncertainty. If moments are assumed to be known exactly (i.e., taking the large-statistics limit), the convergence of the reweighted distributions to the target is governed by how well polynomials can approximate the log-ratio of target to prior densities [61]. Soft-gluon radiation, as generated by a parton shower, smooths the local discontinuities of fixed-order calculations [62]. Provided that kinematic thresholds are mapped to the endpoint of the Chebyshev domain, the target and prior are both smooth and strictly positive, and thus the logarithm of their ratio is real and smooth. This is the function approximated by MaxEnt as a polynomial, and so the reweighted distribution converges uniformly to the target as more moments are included, and under stronger regularity assumptions on the prior, geometrically. In particular, provided the prior has support everywhere the target does, the Kullback–Leibler divergence [63] between the target and the MaxEnt reweight with Chebyshev moments truncated at order NvN_v converges as O(ρ−2Nv)O(\rho^{-2N_v}) under favourable regularity assumptions. The total variation dTVd_{\mathrm{TV}} converges as O(ρ−Nv)O(\rho^{-N_v}). Here ρ>1\rho>1 labels the Bernstein ellipse EρE_\rho, in whose neighbourhood the log-ratio of target to prior is analytic (End Matter).

In practice, higher moments are harder to resolve statistically and so the moment-truncation uncertainty must be balanced against the statistical and scale uncertainty on the higher moments. We incorporate moment constraints up to NvN_v in each tower, the highest order that is resolved at the level of the target moments’ statistical convergence. Order nn is considered resolved when the shift it incurs exceeds the input’s combined Monte Carlo and scale uncertainty σFO\sigma_{\mathrm{FO}}, i.e., when its signal-to-noise ratio

SNRn=∣μnFO−μnprior∣σFO(μn)>1,(5)\mathrm{SNR}_n=\frac{\left|\mu_n^{\mathrm{FO}}-\mu_n^{\mathrm{prior}}\right|}{\sigma_{\mathrm{FO}}(\mu_n)}>1, \tag*{(5)}

with μn\mu_n the tower’s nthn^{\mathrm{th}} moment normalised to its window, on the fixed-order points and on the prior.

Numerical results. We use moments calculated with NNLOJET [35] for the inclusive Drell–Yan observables, choosing μR=μF=mℓℓ\mu_R=\mu_F=m_{\ell\ell}, and with STRIPPER [33, 34, 36] for the Z+jZ+j observables, with μR=μF=mT,ℓℓ\mu_R=\mu_F=m_{T,\ell\ell}. We vary the renormalisation and factorisation scales by the standard seven-point prescription as a heuristic estimate. of missing higher-order uncertainties, and propagate it through the MaxEnt moment-matching.

The prior used here is a sample of 2.1×1062.1 \times10^{6} events generated as pp→ℓ+ℓ−pp \to\ell^{+}\ell^{-} at leading-order within MG5_aMC [65], showered by Pythia 8.312 [66], with multiparton interactions and hadronisation on and matrix-element corrections off, of which 7.5×1057.5 \times10^{5} pass the ATLAS fiducial cuts [64], pTℓ>27p_{\mathrm{T}}^{\ell} > 27 GeV, ∣ηℓ∣<2.5|\eta^{\ell}| < 2.5 and 66<mℓℓ<11666 < m_{\ell\ell} < 116 GeV. All samples are generated with a mass window only, and the fiducial cuts are applied after the shower, so that the shower recoil can move leptons across the fiducial boundaries. For the input moments and the hard process of the prior we use s=13\sqrt{s} = 13 TeV, the CT18NNLO PDF set [67], and for the input moments we use αs(MZ)=0.118\alpha_{\mathrm{s}}(M_Z) = 0.118.

the CT18NNLO PDF set [67], and for the input moments we use αs(MZ)=0.118\alpha_{\mathrm{s}}(M_Z) = 0.118.

As references we use MC@NLO in MG5_aMC [65] and POWHEG at NLO, and MiNNLOPS_{\mathrm{PS}} [24, 25] at NNLO, the last two in POWHEG-BOX-V2 [68], all with the same PDF set, GμG_{\mu}-scheme electroweak inputs and no folding.

We compare with the ATLAS 13 TeV measurement [64] of the ZZ-boson transverse momentum, precise to about 0.2% below 30 GeV, and held-out (but kinematically correlated) observable ϕη∗\phi_{\eta}^{*} [69].

For the inclusive observables we apply no window beyond the fiducial cuts, i.e. wv≡1w_v \equiv1. For resolved Z+jZ+j observables wvw_v rises from zero at the matching threshold vmv_m to one at 2vm2v_m, and is cut off above 500 GeV (End Matter). For pTℓℓp_{\mathrm{T}}^{\ell\ell} we choose vmv_m according to the breakdown of fixed-order perturbation theory in the unresolved pTℓℓ→0p_{\mathrm{T}}^{\ell\ell} \to0 limit, since the resummation of the parton shower prior provides the more accurate description. We require αs(v)ln⁡2(Q/v)<0.2\alpha_{\mathrm{s}}(v)\ln^{2}(Q/v) < 0.2, which holds above 28 GeV for Q=mZQ = m_Z and accordingly set vm=30v_m = 30 GeV. Below this threshold the spectrum remains the shower prior’s, up to the effect of the other constraints.

From the NNLO DY moment calculation, the inclusion criterion of Eq. (5) selects 12 moments of mℓℓm_{\ell\ell}, 11 of ∣yℓℓ∣|y_{\ell\ell}|, and 20 moments of pTℓℓp_{\mathrm{T}}^{\ell\ell}, with no joint moments Eq. (8) (in the End Matter). The calculation achieves NNLO accuracy for the inclusive observables, and NLO accuracy for pTℓℓp_{\mathrm{T}}^{\ell\ell} (which vanishes in the DY Born kinematics). We additionally calculate the pTℓℓp_{\mathrm{T}}^{\ell\ell} moments at NNLO accuracy in a separate NNLO Z+jZ+j calculation. In this case, the criterion of Eq. (5) selects 9 moments.

Figure 1 illustrates the ratio of the methods to fixed-order NNLO for two inclusive observables, dσ/d∣yℓℓ∣d\sigma/d|y_{\ell\ell}| (constrained within MaxEnt) and dσ/dpTℓ1d\sigma/dp_{\mathrm{T}}^{\ell_1} (unconstrained). For the moment-constrained observable, the fixed-order distribution is reproduced to within the propagated statistical uncertainty, despite the starkly different behaviour of the prior at large rapidities. From the held-out observable pTℓ1p_{\mathrm{T}}^{\ell_1}, both MaxEnt samples can be seen to closely reproduce the fixed-order distribution except in the low-pTℓ1p_{\mathrm{T}}^{\ell_1} threshold region, which is resummation-sensitive. Despite the absence of explicit constraints, the MaxEnt result closely follows the other matched calculations there, and agrees with the fixed-order calculation for larger values of pTℓ1p_{\mathrm{T}}^{\ell_1}. This is in part attributable to the low number of free dimensions in phase space. At Born level pTℓ1p_{\mathrm{T}}^{\ell_1} is fixed by mℓℓm_{\ell\ell} and ∣cos⁡θ∗∣|\cos\theta^{*}| in the Collins–Soper frame [70], so the prior’s pre-existing correlations are combined with the additional higher-order information carried by the tilt, moving pTℓ1p_{\mathrm{T}}^{\ell_1} towards the NNLO prediction. Additionally imposing the joint moments of mℓℓm_{\ell\ell} and ∣cos⁡θ∗∣|\cos\theta^{*}| leaves pTℓ1p_{\mathrm{T}}^{\ell_1} essentially unchanged (End Matter). In principle this can be approached fully systematically to identify a complete set of kinematic variables characterising the fixed-order phase-space, and including higher-order correlation information calculable perturbatively, such as with Energy-Flow Polynomials (EFPs) [51, 71].

Figure 1. Ratios to NNLO of the differential cross-sections of two observables, rapidity ∣yℓℓ∣|y_{\ell\ell}| (constrained, top) and leading-lepton transverse momentum pTℓ1p_{\mathrm{T}}^{\ell_1} (unconstrained, bottom). The reweighted LO+PS sample (MaxEnt, thick red and orange) reproduces the input NNLO distribution to the subpercent level. On the held-out observable pTℓ1p_{\mathrm{T}}^{\ell_1} the reweighting procedure corrects the prior to good agreement with the NNLO calculation, despite the absence of an explicit constraint. Here and throughout, bands represent seven-point scale-variation envelopes, and bars statistical errors (End Matter).

Figure 2 similarly shows the distributions dσ/dpTℓℓ\mathrm{d}\sigma/\mathrm{d}p_{T}^{\ell\ell} (constrained within MaxEnt, at NLO and NNLO respectively) and dσ/dϕη∗\mathrm{d}\sigma/\mathrm{d}\phi_{\eta}^{*} (unconstrained), this time alongside ATLAS data. Again, the underlying kinematic correlations between observables present in the prior propagate perturbative information from pTℓℓp_{T}^{\ell\ell} into ϕη∗\phi_{\eta}^{*} automatically. In the last bins, ϕη∗≳2\phi_{\eta}^{*}\gtrsim2, the MaxEnt samples and MiNNLOPS lie above the data and the NNLO Z+jZ+j prediction. Above 2vm2v_{\mathrm{m}}, where wvw_{\mathrm{v}} reaches one, the MaxEnt samples follow their fixed-order input within its uncertainty and describe the data as well as the matched generators. Below vmv_{\mathrm{m}} they keep the shower’s description and avoid the divergence of fixed order in the singular limit.

Figure 2. MaxEnt, the LO+PS prior, fixed-order NNLO DY and Z+jZ+j, NLO-matched MC@NLO and POWHEG, and NNLO-matched MiNNLOPS predictions, compared to 13 TeV ATLAS data [64], for the constrained Z+jZ+j observable pTℓℓp_{T}^{\ell\ell} (left) and the unconstrained observable ϕη∗\phi_{\eta}^{*} (right). The legends give the order of each sample for the inclusive process and for the recoil. Cross sections are absolute, the data being the published shape times the measured fiducial cross section, 736±17736 \pm17 pb. Both MaxEnt variants, with or without the NNLO Z+jZ+j moment constraints, describe the data well across the kinematic range. They reconcile fixed-order accuracy above the transition window, whose lower edge is the dashed vertical line, with the resummation of the parton shower prior below it. On the right only the NNLO Z+jZ+j calculation is shown, from ϕη∗=0.07\phi_{\eta}^{*}=0.07, where its 5 GeV cut on pTℓℓp_{T}^{\ell\ell} leaves it complete. Below ϕη∗=0.15\phi_{\eta}^{*}=0.15 its bins are merged in pairs.

Performance comparison. The variance penalty of event samples of non-uniform weight is given by the Kish effective sample size NeffN_{\mathrm{eff}} [72],

Neff=(∑iqi)2∑iqi2≤N(1−2f−)2,(6)N_{\mathrm{eff}}=\frac{\left(\sum_i q_i\right)^2}{\sum_i q_i^2}\leq N\left(1-2f_{-}\right)^2, \tag*{(6)}

where f−f_{-} is the fractional contribution of negative-weighted events,

f−=−∑iΘ[qi<0]qi∑i∣qi∣.(7)f_{-}=-\frac{\sum_i\Theta[q_i<0]q_i}{\sum_i\lvert q_i\rvert}. \tag*{(7)}

For weights of equal magnitude f−f_{-} is simply the fraction of negative-weight events, and the bound is saturated. The tilt keeps f−=0f_{-}=0 by construction, and the reweighting leaves the effective sample size of the NNLO MaxEnt sample essentially unchanged.

To compare the alternative methods on a like-for-like basis we compare them on a CPU-per-NeffN_{\mathrm{eff}} basis, within the fiducial DY and Z+jZ+j phase spaces used above. This measures the cost of a statistically-equivalent generation campaign. We show this comparison in Figure 3. If the moments are known externally, MaxEnt is both computationally cheaper, and delivers a more statistically-powerful sample, than other methods. This computational advantage is distinct from the theoretical advantage of incorporating higher-accuracy perturbative information for higher-multiplicity observables. If they must also be calculated de novo, this fixed cost is amortised by Neff∼108N_{\mathrm{eff}} \sim10^8 relative to existing NLO matching methods, and 10610^6 for NNLO methods. If an NNLO fixed-order calculation is in any case anticipated, the moment calculation may be included at negligible additional cost.

Figure 3. Performance comparison of alternative generation methods (a) per marginal additional effective event, and (b) for a generation campaign in which NNLO moments must also be calculated de novo. The diagonal lines mark constant per-generated-event cost. On the left, we additionally show the effect of weight-reduction techniques such as MC@NLO-Δ\Delta [4] and folding in POWHEG [10]. On the right, we show the statistical cost of the NNLO DY campaign with target uncertainties down to 0.5%.

Acknowledgments

We thank the organisers of the ‘Negative-weights suppression in Monte Carlo samples’ workshop held at CERN, at which this work was conceived. Research at the Perimeter Institute is supported by the Government of Canada through the Department of Innovation, Science and Industry Canada, and by the Province of Ontario through the Ministry of Colleges and Universities. B.A. was supported in part by the Canadian AI Safety Institute Research Program at CIFAR. This work was performed in part at the Aspen Center for Physics, with support for B.A. by a grant from the Simons Foundation (1161654, Troyer). R.P. is funded by the European Union (ERC, STAPLE, 101222253). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. This work was performed in part using the Cambridge Service for Data Driven Discovery (CSD3), part of which is operated by the University of Cambridge Research Computing on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The DiRAC component of CSD3 was supported by STFC grants ST/P002307/1, ST/R002452/1 and ST/R00689X/1. J.W. acknowledges the support of grant 2025/57/B/ST2/04034 of the National Science Centre, Poland and the Polish high-performance computing infrastructure PLGrid (HPC Centre: ACK Cyfronet AGH) for providing facilities and support within computational grant PLG/2026/062031.

Data availability. The MaxEnt code and all NNLO moments used here are openly available at https://github.com/benleo12/maxent_NNLO. The event samples are available from the authors upon reasonable request.

End Matter

Solver. The multipliers minimize the strictly convex dual g(λ)=ln⁡Z(λ)−μ⋅λ+ϵ2∥λ∥2g(\boldsymbol{\lambda})=\ln Z(\boldsymbol{\lambda})-\boldsymbol{\mu}\cdot\boldsymbol{\lambda}+\frac{\epsilon}{2}\lVert\boldsymbol{\lambda}\rVert^2, with measurement functions standardized under the prior, by safeguarded Newton iteration. The ridge ϵ=10−4\epsilon=10^{-4} makes the Hessian positive definite, relaxing each moment by ϵλa\epsilon\lambda_a in prior standard deviations (significantly below MC errors). Those errors enter through the depth of each tower via Eq. (5), and the bootstrap re-solves below.

Maps, window, and targets. The Drell–Yan moment constraint set used is

M={Tn(uBW(mℓℓ))}n=112∪{Tn(ulin(∣yℓℓ∣))}n=111∪{Tn(ulog⁡(pTℓℓ))}n=120(8)\mathcal{M}=\left\{T_n(u_{\mathrm{BW}}(m_{\ell\ell}))\right\}_{n=1}^{12}\cup\left\{T_n(u_{\mathrm{lin}}(\lvert y_{\ell\ell}\rvert))\right\}_{n=1}^{11}\cup\left\{T_n(u_{\log}(p_{\mathrm{T}}^{\ell\ell}))\right\}_{n=1}^{20} \tag*{(8)}

with the recoil targets blended through the window ww of Eq. (10), as in Eq. (3), and the maps, for ulinu_{\mathrm{lin}}, ulog⁡u_{\log}, and uBWu_{\mathrm{BW}} respectively,

uv=2h(v)−h(vlo)h(vhi)−h(vlo)−1,u_v=2\frac{h(v)-h(v_{\mathrm{lo}})}{h(v_{\mathrm{hi}})-h(v_{\mathrm{lo}})}-1,
h(v)=v, ln⁡v, arctan⁡v2−mZ2mZΓZ,(9)h(v)=v,\ \ln v,\ \arctan\frac{v^2-m_Z^2}{m_Z\Gamma_Z}, \tag*{(9)}

on the intervals [vlo,vhi]=[66,116][v_{\mathrm{lo}},v_{\mathrm{hi}}]=[66,116] GeV for mℓℓm_{\ell\ell}, [0,2.4][0,2.4] for ∣yℓℓ∣\lvert y_{\ell\ell}\rvert, and [30,2500][30,2500] GeV for pTℓℓp_{\mathrm{T}}^{\ell\ell}. A value outside its interval is set to the nearest edge. In the Breit–Wigner coordinate a pure resonance is flat. Any monotonic map reaches the same tilt in the limit, but not under truncation. The window is the C2C^2 profile

w(v)=(6s5−15s4+10s3) Θ(vc−v),(10)w(v)=(6s^5-15s^4+10s^3)\,\Theta(v_c-v), \tag*{(10)}

with s=min⁡{1,max⁡[0,log⁡2(v/vm)]}s=\min\{1,\max[0,\log_2(v/v_m)]\} and Θ\Theta the step function, rising from 0 at vm=30v_m=30 GeV to 1 at 2vm2v_m and closing at vc=500v_c=500 GeV, above which the prior’s shape is kept. The threshold 0.2 in the rule αs(v)ln⁡2(Q/v)<0.2\alpha_s(v)\ln^2(Q/v)<0.2 is a scheme choice that in principle induces an uncertainty.

Moment depth and truncation. Moments are booked to the 12th12^{\mathrm{th}} order for inclusive observables and to the 20th20^{\mathrm{th}} for pTℓℓp_{\mathrm{T}}^{\ell\ell}. We impose all orders up to the highest deemed resolved. In the NNLO Z+Z+jet tower, dropping the unresolved 3rd3^{\mathrm{rd}} and 6th6^{\mathrm{th}} moments moves them by less than their error and changes the recoil above vmv_m by less than half a percent in the median bin, well inside the scale band. We estimate the effect of unknown moments by solving again with two and four fewer orders and taking the change as the truncation uncertainty. For the inclusive observables it is comparable to the residual deviations, which therefore sit at the noise floor of the input. For pTℓℓp_{\mathrm{T}}^{\ell\ell}, cutting the tower to the nine orders of the NNLO Z+Z+jet tower moves the medians of Table I by less than their errors.

f−f_- (%)Neff/NN_{\mathrm{eff}}/N (%)pTℓℓp_{\mathrm{T}}^{\ell\ell} (%) <400pTℓℓp_{\mathrm{T}}^{\ell\ell} (%) >vm>v_mϕn∗\phi_n^* (%) allϕn∗\phi_n^* (%) >0.33>0.33σfid\sigma_{\mathrm{fid}} (pb)
LO+PS prior010010.511.83.48.3616
MaxEnt, NLO(ZjZ_j)099.96.08.04.07.7710
MaxEnt, NNLO(ZjZ_j)a^a099.93.84.13.75.5710
MiNNLOPS_{\mathrm{PS}}23.527.26.97.76.17.8699
MC@NLO5.280.410.524.25.131.7677
POWHEG1.295.24.58.04.613.4705

Table I. Negative-weight fraction f−f_-, Neff/NN_{\mathrm{eff}}/N, median deviation from the data of the normalised pTℓℓp_{\mathrm{T}}^{\ell\ell} (below 400 GeV and above vmv_m) and ϕn∗\phi_n^* (all bins and above 0.33), and the fiducial cross section per lepton flavour, measured as 736±17736\pm17 pb.

Convergence and moment-truncation uncertainty. We sketch the argument underpinning the convergence claim. Suppose pp is the prior probability distribution and rr the target.

Let ℓ=log⁡(r/p)\ell=\log(r/p) be the exact reweight and suppose that ℓN∈PN\ell_N\in\mathcal{P}_N is the polynomial of degree NN or less that best uniformly approximates ℓ\ell, with approximation error ϵN=∥ℓ−ℓN∥∞\epsilon_N=\lVert\ell-\ell_N\rVert_\infty. Let qNq_N be the MaxEnt fit matching all polynomial moments up to degree NN. Since log⁡(qN/p)∈PN\log(q_N/p)\in\mathcal{P}_N, and the moments up to NN must match,

4dTV(r,qN)2≤DKL(r∥qN)+DKL(qN∥r)=∫(r−qN)(ℓ−ℓN)≤ϵN∫∣r−qN∣=2ϵNdTV(r,qN),(11)\begin{aligned} 4d_{\mathrm{TV}}(r,q_N)^2\le D_{\mathrm{KL}}(r\Vert q_N)+D_{\mathrm{KL}}(q_N\Vert r) \\ &=\int(r-q_N)(\ell-\ell_N)\le\epsilon_N\int\lvert r-q_N\rvert=2\epsilon_Nd_{\mathrm{TV}}(r,q_N), \tag*{(11)} \end{aligned}

applying Pinsker’s inequality [73, 74] for the lower bound. It follows that the total variation dTV(r,qN)≤12ϵNd_{\mathrm{TV}}(r,q_N)\le\frac{1}{2}\epsilon_N and

DKL(r∥qN)+DKL(qN∥r)≤ϵN2.(12)D_{\mathrm{KL}}(r\Vert q_N)+D_{\mathrm{KL}}(q_N\Vert r)\le\epsilon_N^2. \tag*{(12)}

Thus the distributional deviation between reweight and target is bounded by the quality of the degree-NN polynomial approximation to ℓ\ell. For Chebyshev polynomials, a standard approximation theory result [75, Theorem 8.3] states that, if ℓ\ell extends analytically to a neighbourhood of the Bernstein ellipse EρE_\rho, ρ>1\rho>1, ϵN=O(ρ−N)\epsilon_N=\mathcal{O}(\rho^{-N}), and so

dTV(r,qN)=O(ρ−N),DKL(r∥qN)=O(ρ−2N),(13)d_{\mathrm{TV}}(r,q_N)=\mathcal{O}(\rho^{-N}),\qquad D_{\mathrm{KL}}(r\Vert q_N)=\mathcal{O}(\rho^{-2N}), \tag*{(13)}

i.e. converge geometrically.

The analyticity properties of ℓ\ell required here are not unreasonable, since common kinematic, resonant, etc. factors between the prior and target cancel in the ratio, but should not be expected to hold for arbitrary IR-safe observables. Even without geometric convergence, the bound on the deviation of MaxEnt distributions in terms of the best polynomial approximation error persists. This is the case for the windowed moments for pTℓℓp_{\mathrm{T}}^{\ell\ell}, for which the target is here C2C^2 through Eq. (10), so ϵN\epsilon_N falls as N−3N^{-3} rather than geometrically [75, Theorem 7.2].

Note that this applies to the exact MaxEnt problem in one dimension, without the (small) numerical regularisation ridge used in practice. With additional dimensions and correlations, εN\varepsilon_N corresponds to the best approximation of ℓ\ell achievable within the sum of the single-variable Chebyshev constraints considered here, and cannot vanish without cross-moments where the target distribution has correlations beyond those accessible within the considered class (see Cross moments below).

Leptons and showers. Leptons are taken before QED final-state radiation, the convention of the published data and of the QCD-only fixed-order input. POWHEG and MiNNLOPS events are showered with POWHEG emission as starting scale, and MC@NLO with the settings Pythia recommends for MG5_aMC.

Uncertainty propagation. The scale uncertainty band plotted for the MaxEnt sample is the envelope of six re-solves at the standard seven-point scale variations of the target moments, each warm-started from the central multipliers. On the mℓℓm_{\ell\ell} distribution (Figure 4), the band closely follows that of the NNLO calculation. Error bars add two uncertainties in quadrature: the spread of 20 re-solves, with targets rebuilt from the fixed-order runs (resampled with replacement), and the reweighted sample’s per-bin Monte Carlo error. For the NNLO Z+Z+jet sample the re-solves also redraw the Z+Z+jet rate within its Monte Carlo error. The Z+Z+jet shape moments keep their central values, so their statistical errors do not enter the error bars.

Ratio to NNLO of the dilepton mass, constrained, with the curves, bands and bars of Fig. 1.

Figure 4. Ratio to NNLO of the dilepton mass, constrained, with the curves, bands and bars of Fig. 1.

Cross moments. The construction accepts joint constraints from cross-moments w Tn(uv) Tn′(uv′)w\,T_n(u_v)\,T_{n'}(u_{v'}). These carry information about the correlations between observables that is missing from one-dimensional moment targets. Unlike multiply-differential histogram binning, cross-moments can resolve correlations without requiring statistically-independent integrals to converge in each bin. For Drell–Yan we book the first four orders of mℓℓm_{\ell\ell} and ∣cos⁡θ∗∣|\cos\theta^*| jointly, 16 cross-moments. On average they lie 4.1 σFO4.1\,\sigma_{\mathrm{FO}} from fixed order in the prior and 2.0 σFO2.0\,\sigma_{\mathrm{FO}} after the tilt on the marginals alone, with σFO\sigma_{\mathrm{FO}} as in Eq. (5). Imposing all 16 reproduces them at no cost in NeffN_{\mathrm{eff}}.

Higgs boson production. We further apply the MaxEnt method to gluon-fusion Higgs production with decay to two photons, again with LO+PS prior. Here we impose the NNLO Higgs rapidity and, above a matching scale, the NLO H+H+jet recoil from NNLOJET [35]. Each gluon-fusion sample is normalised to the inclusive N3^3LO cross section [77] for comparison with the data. Other production modes are then added from LO+PS samples. Figure 5 compares the diphoton transverse momentum with ATLAS data [76] and MiNNLOPS [24] and MC@NLO [2, 65] predictions, which for the displayed distribution have NLO and LO accuracy respectively. The MaxEnt tilt qi/piq_i/p_i lie between 0.27 and 2.0, with Neff/N=88%N_{\mathrm{eff}}/N=88\%. MiNNLOPS and MC@NLO have negative-weight fractions f−=14%f_- = 14\% and 9% and keep Neff/N=8%N_{\mathrm{eff}}/N=8\% and 69%.

Figure 5. Gluon-fusion Higgs production: the diphoton transverse momentum in the ATLAS fiducial region [76], with the curves, bands and bars of Fig. 2. These samples have been normalised to the inclusive N3^3LO cross section [77], whose uncertainty is included in the bands.

References

† [email protected]

‡ [email protected]

References

  1. [1]S. Amoroso et al. (HSF Physics Event Generator WG), Comput. Softw. Big Sci. 5, 12 (2021), arXiv:2004.13687 [hep-ph].
  2. [2]S. Frixione and B. R. Webber, JHEP 2002 (06), 029, arXiv:hep-ph/0204244.DOI
  3. [3]K. Danziger, S. Höche, and F. Siegert, Reducing negative weights in Monte Carlo event generation with Sherpa (2021), arXiv:2110.15211 [hep-ph].arxiv.org/abs/2110.15211
  4. [4]R. Frederix, S. Frixione, S. Prestel, and P. Torrielli, JHEP 2020 (07), 238, arXiv:2002.12716 [hep-ph].DOI
  5. [5]R. Frederix and P. Torrielli, Eur. Phys. J. C 83, 1051 (2023), arXiv:2310.04160 [hep-ph].arxiv.org/abs/2310.04160
  6. [6]S. El Farkh, R. Frederix, and M. Gouighiri, Eur. Phys. J. C 86, 751 (2026), arXiv:2602.18124 [hep-ph].
  7. [7]P. Nason, JHEP 2004 (11), 040, arXiv:hep-ph/0409146.DOI
  8. [8]S. Frixione, P. Nason, and C. Oleari, JHEP 2007 (11), 070, arXiv:0709.2092 [hep-ph].DOI
  9. [9]P. Nason, MINT: A Computer program for adaptive Monte Carlo integration and generation of unweighted distributions (2007), arXiv:0709.2085 [hep-ph].arxiv.org/abs/0709.2085
  10. [10]S. Alioli, P. Nason, C. Oleari, and E. Re, JHEP 2011 (01), 095, arXiv:1009.5594 [hep-ph].DOI
  11. [11]S. Jadach, W. Placzek, S. Sapeta, A. Siódmok, and M. Skrzypek, JHEP 2015 (10), 052, arXiv:1503.06849 [hep-ph].DOI
  12. [12]P. Sarmah, A. Siódmok, and J. Whitehead, JHEP 2025 (01), 062, arXiv:2409.16417 [hep-ph].DOI
  13. [13]M. van Beekveld, S. Ferrario Ravasio, J. Helliwell, A. Karlberg, G. P. Salam, L. Scyboz, A. Soto-Ontoso, G. Soyez, and S. Zanoli, JHEP 2025 (10), 038, arXiv:2504.05377 [hep-ph].DOI
  14. [14]P. Nason and G. P. Salam, JHEP 2022 (01), 067, arXiv:2111.03553 [hep-ph].DOI
  15. [15]Y. Cheng and R. Frederix, Eur. Phys. J. C 86, 870 (2026), arXiv:2605.22108 [hep-ph].
  16. [16]B. Nachman and J. Thaler, Phys. Rev. D 102, 076004 (2020), arXiv:2107.11586 [hep-ph].
  17. [17]J. R. Andersen, C. Gütschow, A. Maier, and S. Prestel, Eur. Phys. J. C 80, 1007 (2020), arXiv:2005.09375 [hep-ph].DOI
  18. [18]J. R. Andersen and A. Maier, Eur. Phys. J. C 82, 433 (2022), arXiv:2109.07851 [hep-ph].DOI
  19. [19]J. R. Andersen, A. Maier, and D. Maître, Eur. Phys. J. C 83, 835 (2023), arXiv:2303.15246 [hep-ph].DOI
  20. [20]J. R. Andersen, E. Cole, and A. Maier, Precision Cell Resampling with a Relative and Resonant Aware Metric (2026), arXiv:2605.13237 [hep-ph].arxiv.org/abs/2605.13237
  21. [21]R. Doherty, L. Hay, R. Jain, M. LeBlanc, J. Marrinan, C. Mauceri, and J. Roloff, Optimal-Transport-Based Cell Resampling for Negative and Pathological Event Weights (2026), arXiv:2607.08723 [hep-ph].arxiv.org/abs/2607.08723
  22. [22]C. Palmer and B. Kronheim, Phys. Rev. D 113, 012003 (2026), arXiv:2510.16217 [hep-ex].DOI
  23. [23]P. Shyamsundar, ARCANE Reweighting: A Monte Carlo Technique to Tackle the Negative Weights Problem in Collider Event Generation (2025), arXiv:2502.08052 [hep-ph].arxiv.org/abs/2502.08052
  24. [24]P. F. Monni, P. Nason, E. Re, M. Wiesemann, and G. Zanderighi, JHEP 2020 (05), 143, [Erratum: JHEP 2022 (02), 031], arXiv:1908.06987 [hep-ph].DOI
  25. [25]P. F. Monni, E. Re, and M. Wiesemann, Eur. Phys. J. C 80, 1075 (2020), arXiv:2006.04133 [hep-ph].DOI
  26. [26]K. Hamilton, P. Nason, E. Re, and G. Zanderighi, JHEP 2013 (10), 222, arXiv:1309.0017 [hep-ph].DOI
  27. [27]A. Karlberg, E. Re, and G. Zanderighi, JHEP 2014 (09), 134, arXiv:1407.2940 [hep-ph].DOI
  28. [28]S. Alioli, C. W. Bauer, C. Berggren, F. J. Tackmann, J. R. Walsh, and S. Zuberi, JHEP 2014 (06), 089, arXiv:1311.0286 [hep-ph].DOI
  29. [29]S. Höche, Y. Li, and S. Prestel, Phys. Rev. D 91, 074015 (2015), arXiv:1405.3607 [hep-ph].
  30. [30]A. Gehrmann-De Ridder, T. Gehrmann, and E. W. N. Glover, JHEP 2005 (09), 056, arXiv:hep-ph/0505111.DOI
  31. [31]A. Daleo, T. Gehrmann, and D. Maître, JHEP 2007 (04), 016, arXiv:hep-ph/0612257.DOI
  32. [32]J. Currie, E. W. N. Glover, and S. Wells, JHEP 2013 (04), 066, arXiv:1301.4693 [hep-ph].DOI
  33. [33]M. Czakon, Phys. Lett. B 693, 259 (2010), arXiv:1005.0274 [hep-ph].DOI
  34. [34]M. Czakon and D. Heymes, Nucl. Phys. B 890, 152 (2014), arXiv:1408.2500 [hep-ph].DOI
  35. [35]A. Huss et al. (NNLOJET), SciPost Phys. Codeb. 2026, 69 (2026), arXiv:2503.22804 [hep-ph].DOI
  36. [36]M. Czakon, A. van Hameren, A. Mitov, and R. Poncelet, JHEP 2019 (10), 262, arXiv:1907.12971 [hep-ph].
  37. [37]M. Dasgupta, F. A. Dreyer, K. Hamilton, P. F. Monni, G. P. Salam, and G. Soyez, Phys. Rev. Lett. 125, 052002 (2020), arXiv:2002.11114 [hep-ph].DOI
  38. [38]J. R. Forshaw, J. Holguin, and S. Plätzer, JHEP 2020 (09), 014, arXiv:2003.06400 [hep-ph].arxiv.org/abs/2003.06400
  39. [39]F. Herren, S. Höche, F. Krauss, D. Reichelt, and M. Schönherr, JHEP 2023 (10), 091, arXiv:2208.06057 [hep-ph].arxiv.org/abs/2208.06057
  40. [40]M. van Beekveld, S. Ferrario Ravasio, K. Hamilton, G. P. Salam, A. Soto-Ontoso, G. Soyez, and R. Verheyen, JHEP 2022 (11), 020, arXiv:2207.09467 [hep-ph].DOI
  41. [41]B. Assi and S. Höche, Phys. Rev. D 109, 114008 (2024), arXiv:2307.00728 [hep-ph].DOI
  42. [42]S. Höche, F. Krauss, and D. Reichelt, Phys. Rev. D 111, 094032 (2025), arXiv:2404.14360 [hep-ph].arxiv.org/abs/2404.14360
  43. [43]C. T. Preuss, JHEP 2024 (07), 161, arXiv:2403.19452 [hep-ph].arxiv.org/abs/2403.19452
  44. [44]J. Holguin, S. Plätzer, M. H. Seymour, and S. Sule, JHEP 2026 (08), 146, arXiv:2605.02622 [hep-ph].arxiv.org/abs/2605.02622
  45. [45]E. K. Haviland, American Journal of Mathematics 58, 164 (1936).
  46. [46]K. Schmüdgen, The Moment Problem, 1st ed., Graduate Texts in Mathematics (Springer International Publishing, Cham, Switzerland, 2017).DOI
  47. [47]F. Hausdorff, Mathematische Zeitschrift 9, 74 (1921).
  48. [48]T. H. Hildebrandt and I. J. Schoenberg, Annals of Mathematics 34, 317 (1933).
  49. [49]E. T. Jaynes, Phys. Rev. 106, 620 (1957).
  50. [50]B. Assi, S. Höche, K. Lee, and J. Thaler, Phys. Rev. Lett. 135, 131901 (2025), arXiv:2501.17219 [hep-ph].arxiv.org/abs/2501.17219
  51. [51]B. Assi, K. Lee, and J. Thaler, JHEP 2026 (09), 191, arXiv:2604.00084 [hep-ph].arxiv.org/abs/2604.00084
  52. [52]B. Assi, E. Bothmann, J. Campbell, C. Gütschow, S. Höche, W.-L. Ju, M. Knobbe, M. Schönherr, J. Thaler, and M. L. Wagman, Drell-Yan lepton pair production from low to high transverse momentum (2026), arXiv:2610.02169 [hep-ph].arxiv.org/abs/2610.02169
  53. [53]Z. Bern, L. J. Dixon, F. Febres Cordero, S. Höche, H. Ita, D. A. Kosower, and D. Maître, Comput. Phys. Commun. 185, 1443 (2014), arXiv:1310.7439 [hep-ph].DOI
  54. [54]M. Czakon, Z. Kassabov, A. Mitov, R. Poncelet, and A. Popescu, J. Phys. G 51, 115002 (2024), arXiv:2304.05993 [hep-ph].arxiv.org/abs/2304.05993
  55. [55]T. Carli, D. Clements, A. Cooper-Sarkar, C. Gwenlan, G. P. Salam, F. Siegert, P. Starovoitov, and M. Sutton, Eur. Phys. J. C 66, 503 (2010), arXiv:0911.2985 [hep-ph].DOI
  56. [56]S. Carrazza, E. R. Nocera, C. Schwan, and M. Zaro, JHEP 2020 (12), 108, arXiv:2008.12789 [hep-ph].DOI
  57. [57]D. Britzger et al., Eur. Phys. J. C 82, 930 (2022), arXiv:2207.13735 [hep-ph].arxiv.org/abs/2207.13735
  58. [58]T. Ježo, E. R. Nocera, T. R. Rabemananjara, C. Schwan, T. Sharma, and J. Wissmann, PineAPPLv1: fast and flexible theory predictions for present and future colliders (2026), arXiv:2606.17134 [hep-ph].arxiv.org/abs/2606.17134
  59. [59]J. E. Shore and R. W. Johnson, IEEE Trans. Inf. Theory 26, 26 (1980), [Correction: IEEE Trans. Inf. Theory 29, 942 (1983)].
  60. [60]I. Csiszár, Ann. Probab. 3, 146 (1975).
  61. [61]J. M. Borwein and A. S. Lewis, SIAM Journal on Optimization 1, 191 (1991).
  62. [62]S. Catani and B. R. Webber, JHEP 1997 (10), 005, arXiv:hep-ph/9710333.DOI
  63. [63]S. Kullback and R. A. Leibler, The Annals of Mathematical Statistics 22, 79 (1951).
  64. [64]G. Aad et al. (ATLAS), Eur. Phys. J. C 80, 616 (2020), arXiv:1912.02844 [hep-ex].
  65. [65]J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S. Shao, T. Stelzer, P. Torrielli, and M. Zaro, JHEP 2014 (07), 079, arXiv:1405.0301 [hep-ph].DOI
  66. [66]C. Bierlich et al., SciPost Phys. Codeb. 2022, 8 (2022), arXiv:2203.11601 [hep-ph].arxiv.org/abs/2203.11601
  67. [67]T.-J. Hou et al., Phys. Rev. D 103, 014013 (2021), arXiv:1912.10053 [hep-ph].DOI
  68. [68]S. Alioli, P. Nason, C. Oleari, and E. Re, JHEP 2010 (06), 043, arXiv:1002.2581 [hep-ph].DOI
  69. [69]A. Banfi, S. Redford, M. Vesterinen, P. Waller, and T. R. Wyatt, Eur. Phys. J. C 71, 1600 (2011), arXiv:1009.1580 [hep-ex].DOI
  70. [70]J. C. Collins and D. E. Soper, Phys. Rev. D 16, 2219 (1977).
  71. [71]P. T. Komiske, E. M. Metodiev, and J. Thaler, JHEP 2018 (04), 013, arXiv:1712.07124 [hep-ph].DOI
  72. [72]L. Kish, Survey Sampling (John Wiley & Sons, New York, 1965).
  73. [73]M. S. Pinker, Information and Information Stability of Random Variables and Processes (Holden-Day, San Francisco, 1964).
  74. [74]I. Csiszár, Studia Sci. Math. Hungar. 2, 299 (1967).
  75. [76]G. Aad et al. (ATLAS), JHEP 2022 (08), 027, arXiv:2202.00487 [hep-ex].
  76. [77]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, Tech. Rep. (CERN, 2017) arXiv:1610.07922 [hep-ph].DOI

Paper details

vlnc:W_7dkau3uf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents