WorksPhysics and Astronomy
ESME stays positive: event generation with positive weights
Abstract
Fully differential simulations of hard-scattering processes are a cornerstone of precision phenomenology at collider experiments. For long, Monte Carlo event generators have been able to achieve next-to-leading order (NLO) accuracy, but their results typically suffer from the presence of negative-weight events. These degrade statistical power and complicate downstream analyses. In this contribution, we present Exponentiated Subtraction for Matching Events (ESME), a novel approach to NLO+PS matching that yields positive-definite event weights while preserving perturbative accuracy and the logarithmic structure of the shower. We discuss the formalism and its implementation within the PanScales framework.
Introduction
The precise simulation of high-energy particle collisions requires a consistent description of both the short-distance hard interaction and the subsequent pattern of QCD radiation. This can be achieved through the combination of fixed-order perturbative calculations with parton shower algorithms. General-purpose event generators provide such simulations in the form of fully differential ‘hard’ events, which can subsequently be processed by shower algorithms and hadronisation models.
Over the past decades, significant progress has been made in improving the perturbative accuracy of these simulations. On the resummation side, showers with next-to-next-to-leading logarithmic (NLL) accuracy are becoming the standard [1, 2, 3, 4, 5, 6, 7, 8, 9], and crucial steps towards achieving next-to-NLL are being made as well [10, 11, 12, 13]. On the fixed-order side, matching schemes such as MC@NLO and POWHEG allow the combination of next-to-leading order (NLO) calculations with parton showers [14, 15, 16, 17, 18, 19, 20], while more recent developments extend this to next-to-next-to-leading order (NNLO) [21, 22, 23, 24, 25, 26, 27]. A persistent issue affecting (N)NLO+PS approaches is the presence of negative-weight events. These arise as a consequence of subtraction procedures and matching prescriptions, and pose both conceptual and practical challenges.
The problem of negative weights
The presence of negative-weight events reduces the statistical efficiency of Monte Carlo simulations. If a fraction of events carries a negative weight, the effective number of events required to reach a given statistical precision increases according to
This implies a substantial increase in computational cost when is sizeable (see Figure 1 for negative-weight fractions found in typical collider processes).
Figure 1. Eq. (1) as a function of , the fraction of negative-weight events. Negative-weight fractions as obtained with MadGraph v3.6.3 for some selected collider processes are indicated as well.
Beyond higher computational costs, negative weights also complicate the use of modern analysis techniques, including machine learning methods that rely on probabilistic interpretations of event samples (see e.g. Refs. [28, 29, 30]). Negative weights are also profoundly unphysical. For these reasons, the theory community has been devoting considerable effort to reduce or eliminate negative weights in event generation (see e.g. Refs. [31, 32, 33, 34, 35, 36, 37, 38]).
Structure of NLO+PS event generation
At NLO accuracy, event generation proceeds by constructing a modified Born-level contribution that incorporates virtual corrections and the integral over real emission contributions. This results in a modified Born weight
where denotes the squared Born matrix element, the virtual contribution, the real-emission matrix element, and () the (integrated) subtraction term. These contributions depend on the phase space and , which describe the underlying Born configuration and the Born plus radiation phase space. The contribution is integrated over the radiation phase space, .
Negative weights can originate from several sources.1 First of all, the correction in can be genuinely large and negative, which happens e.g. when one tries to compute a fixed-order correction in a region of phase-space where resummation is important. Another source comes from the way the integral for is actually being computed in Monte Carlo algorithms. Since it is generally expensive to evaluate the full integral for each Born event fully, often only a (few) single call(s) are made for each Born event (whereas the complete integral in this case can be obtained by simply generating many events). Of course, positivity of the integrand is not guaranteed, as one may happen to sample a region of phase space where , rendering the total (i.e. differential in both the Born and radiation phase space) negative. It is precisely this issue that the ESME method aims to tackle.
Exponentiated Subtraction for Matching Events
The Exponentiated Subtraction for Matching Events (ESME) method provides a new way for NLO event generation that avoids negative weights altogether [42]. The idea relies on expressing as the expectation value of a stochastic variable, allowing the removal of negative weights at the cost of introducing higher-order corrections. One ingredient of the ESME method is an algorithm for Exponentiated Subtraction (ES). This algorithm is constructed so that it assigns a Born weight for each Born configuration whose average becomes
The ES algorithm proceeds by introducing an overestimate of the real-emission contribution and iteratively updating the weight through a sequence of probabilistic steps. Starting from an initial value , one samples radiation variables and applies accept-reject criteria based on the difference . Depending on the sign of this difference, the value of is incremented or decremented by a value of . The procedure is designed such that the expectation value of reproduces the NLO-correct weight (see Sec. 3.2 for more details).
In its basic form, the stochastic construction can yield negative values of . Such configurations can be removed by rejecting events with , thereby ensuring that all retained events carry a positive weight. This introduces the correction in Eq. (3).1 A key feature of the algorithm is that the order at which these higher-order corrections are introduced are controllable. As a result, the method preserves NLO accuracy while achieving strictly positive event weights.
The ES algorithm can be adapted to not only ensure NLO accuracy with positive-weights, but also combine it with real-emission generation and shower evolution, at which point it becomes the ESME algorithm (see Sec. 3.3 of [42]). An important feature of ESME is that it preserves the logarithmic accuracy of the parton shower without requiring additional modifications. This is in contrast to existing HEG-based approaches, where the hardest emission is generated using another phase-space coverage than that of the shower, and consequently care must be taken to preserve the resummation accuracy of the shower [43].
Implementation and results
The ESME method has been implemented in the PanScales framework [44] for a range of processes, including production, deep-inelastic scattering, and scatterings, achieving NLO+NLL accuracy. Comparisons with fixed-order predictions demonstrate that the method correctly reproduces NLO results for a variety of observables. For example, considering Drell-Yan production, the -boson rapidity and transverse-momentum distributions agree with those obtained from MCFM[45] (see Figure 2). From a computational perspective, the method offers a significant improvement in efficiency. In general, the absence of negative weights reduces the number of events required to reach a given statistical precision. The precise speed-up depends, of course, on the fraction of negative weights present in the original setup, which is relatively small for the colour-singlet processes considered here. Independently of this, the implementation of ESME in the PanScales framework achieves speed-ups of up to an order of magnitude compared to existing NLO tools when comparing the event generation time spent per event (see Fig. 12 of Ref. [42]). This demonstrates that eliminating negative weights does not require compromising computational speed as a trade-off.

Figure 2. NLO test of the ESME matching method for the process, with the rapidity of the lepton pair on the left, and their transverse momentum on the right.
Conclusions
We have presented a new approach to NLO+PS matching based on exponentiated subtraction, which enables the generation of positive-weight events while preserving the logarithmic accuracy of parton showers. The method has been validated for a range of processes and demonstrates both theoretical robustness and practical efficiency.
Future developments will focus on extending the method beyond colour-singlet production and towards higher perturbative orders. These directions are expected to both speed up the generation and enhance the precision of Monte Carlo event generators.
Acknowledgments
MvB thanks G. P. Salam and L. Scyboz for comments on the manuscript. The results presented in these proceedings have been obtained in collaboration with S. Ferrario Ravasio, J. Helliwell, A. Karlberg, G. P. Salam, L. Scyboz, A. Soto Ontoso, G. Soyez and S. Zanoli. This work was supported by the Dutch Research Council (NWO) under project number VI.Veni.232.190. AI tools were used to assist drafting this document.
References
- [1]Mrinal Dasgupta, Frédéric A. D. Dreyer, Keith Hamilton, Pier Francesco Monni, Gavin P. Salam, and Gregory Soyez. Parton showers beyond leading logarithmic accuracy. Phys. Rev. Lett., 125(5):052002, 2020.DOI
- [2]Zoltán Nagy and Davison E. Soper. Summations by parton showers of large logarithms in electron-positron annihilation. 11 2020.arxiv.org/abs/2011.04777
- [3]Jeffrey R. Forshaw, Jack Holguin, and Simon Plätzer. Building a consistent parton shower. JHEP, 09:014, 2020.arxiv.org/abs/2003.06400
- [4]Florian Herren, Stefan Höche, Frank Krauss, Daniel Reichelt, and Marek Schoenherr. A new approach to color-coherent parton evolution. JHEP, 10:091, 2023.arxiv.org/abs/2208.06057
- [5]Melissa van Beekveld, Silvia Ferrario Ravasio, Keith Hamilton, Gavin P. Salam, Alba Soto-Ontoso, Gregory Soyez, and Rob Verheyen. PanScales showers for hadron collisions: all-order validation. JHEP, 11:020, 2022.DOI
- [6]Melissa van Beekveld, Silvia Ferrario Ravasio, Gavin P. Salam, Alba Soto-Ontoso, Gregory Soyez, and Rob Verheyen. PanScales parton showers for hadron collisions: formulation and fixed-order studies. JHEP, 11:019, 2022.DOI
- [7]Melissa van Beekveld and Silvia Ferrario Ravasio. Next-to-leading-logarithmic PanScales showers for Deep Inelastic Scattering and Vector Boson Fusion. JHEP, 02:001, 2024.DOI
- [8]Stefan Höche, Frank Krauss, and Daniel Reichelt. alaric parton shower for hadron colliders. Phys. Rev. D, 111(9):094032, 2025.arxiv.org/abs/2404.14360
- [9]Christian T. Preuss. A partitioned dipole-antenna shower with improved transverse recoil. JHEP, 07:161, 2025.arxiv.org/abs/2403.19452
- [10]Melissa van Beekveld, Mrinal Dasgupta, Basem Kamal El-Menoufi, Jack Helliwell, and Pier Francesco Monni. Collinear fragmentation at NNLL: generating functionals, groomed correlators and angularities. JHEP, 05:093, 2024.arxiv.org/abs/2307.15734
- [11]Melissa van Beekveld, Mrinal Dasgupta, Basem Kamal El-Menoufi, Jack Helliwell, Pier Francesco Monni, and Gavin P. Salam. A collinear shower algorithm for NSL non-singlet fragmentation. JHEP, 03:209, 2025.arxiv.org/abs/2409.08316
- [12]Silvia Ferrario Ravasio, Keith Hamilton, Alexander Karlberg, Gavin P. Salam, Ludovic Scyboz, and Gregory Soyez. Parton Showering with Higher Logarithmic Accuracy for Soft Emissions. Phys. Rev. Lett., 131(16):161906, 2023.DOI
- [13]Melissa van Beekveld et al. New Standard for the Logarithmic Accuracy of Parton Showers. Phys. Rev. Lett., 134(1):011901, 2025.DOI
- [14]Christian Bierlich et al. A comprehensive guide to the physics and usage of PYTHIA 8.3. SciPost Phys. Codeb., 2022:8, 2022.arxiv.org/abs/2203.11601
- [15]Johannes Bellm et al. Herwig 7.0/Herwig++ 3.0 release note. Eur. Phys. J. C, 76(4):196, 2016.
- [16]Enrico Bothmann et al. Event generation with Sherpa 3. JHEP, 12:156, 2024.arxiv.org/abs/2410.22148
- [17]Simone Alioli, Paolo Nason, Carlo Oleari, and Emanuele Re. A general framework for implementing NLO calculations in shower Monte Carlo programs: the POWHEG BOX. JHEP, 06:043, 2010.DOI
- [18]Tomáš Ježo and Paolo Nason. On the Treatment of Resonances in Next-to-Leading Order Calculations Matched to a Parton Shower. JHEP, 12:065, 2015.DOI
- [19]J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H. S. Shao, T. Stelzer, P. Torrielli, and M. Zaro. The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations. JHEP, 07:079, 2014.DOI
- [20]R. Frederix, S. Frixione, V. Hirschi, D. Pagani, H. S. Shao, and M. Zaro. The automation of next-to-leading order electroweak calculations. JHEP, 07:185, 2018. [Erratum: JHEP 11, 085 (2021)].DOI
- [21]Simone Alioli, Christian W. Bauer, Calvin J. Berggren, Andrew Horning, Frank J. Tackmann, Christopher K. Vermilion, Jonathan R. Walsh, and Saba Zuberi. Combining Higher-Order Resummation with Multiple NLO Calculations and Parton Showers in GENEVA. JHEP, 09:120, 2013.
- [22]Simone Alioli, Christian W. Bauer, Alessandro Broggio, Alessandro Gavardi, Stefan Kallweit, Matthew A. Lim, Riccardo Nagar, Davide Napolitano, and Luca Rottoli. Matching NNLO predictions to parton showers using N3LO color-singlet transverse momentum resummation in geneva. Phys. Rev. D, 104(9):094020, 2021.DOI
- [23]Simone Alioli, Georgios Billis, Alessandro Broggio, Alessandro Gavardi, Stefan Kallweit, Matthew A. Lim, Giulia Marinelli, Riccardo Nagar, and Davide Napolitano. Double Higgs production at NNLO interfaced to parton showers in GENEVA. JHEP, 06:205, 2023.
- [24]Pier Francesco Monni, Paolo Nason, Emanuele Re, Marius Wiesemann, and Giulia Zanderighi. MiNNLO_PS: a new method to match NNLO QCD to parton showers. JHEP, 05:143, 2020. [Erratum: JHEP 02, 031 (2022)].
- [25]Daniele Lombardi, Marius Wiesemann, and Giulia Zanderighi. Advancing MiNNLO_PS to diboson processes: Zγ production at NNLO+PS. JHEP, 06:095, 2021.DOI
- [26]Javier Mazzitelli, Pier Francesco Monni, Paolo Nason, Emanuele Re, Marius Wiesemann, and Giulia Zanderighi. Next-to-Next-to-Leading Order Event Generation for Top-Quark Pair Production. Phys. Rev. Lett., 127(6):062001, 2021.DOI
- [27]Christian Biello, Javier Mazzitelli, Aparna Sankar, Marius Wiesemann, and Giulia Zanderighi. Higgs boson production in association with massive bottom quarks at NNLO+PS. JHEP, 04:088, 2025.arxiv.org/abs/2412.09510
- [28]Georges Aad et al. Search for tt̄H/A → tt̄tt̄ production in the multilepton final state in proton–proton collisions at √s = 13 TeV with the ATLAS detector. JHEP, 07:203, 2023.
- [29]Armen Tumasyan et al. Search for Higgs Boson Decay to a Charm Quark-Antiquark Pair in Proton-Proton Collisions at s=13 TeV. Phys. Rev. Lett., 131(6):061801, 2023.
- [30]Aram Hayrapetyan et al. Reweighting simulated events using machine-learning techniques in the CMS experiment. 11 2024.
- [31]S. Jadach, W. Placzek, S. Sapeta, A. Siódmok, and M. Skrzypek. Matching NLO QCD with parton shower in Monte Carlo scheme — the KrkNLO method. JHEP, 10:052, 2015.
- [32]Benjamin Nachman and Jesse Thaler. Neural resampler for Monte Carlo reweighting with preserved uncertainties. Phys. Rev. D, 102(7):076004, 2020.DOI
- [33]Paolo Nason and Gavin P. Salam. Multiplicative-accumulative matching of NLO calculations with parton showers. JHEP, 01:067, 2022.DOI
- [34]Jeppe R. Andersen and Andreas Maier. Unbiased elimination of negative weights in Monte Carlo samples. Eur. Phys. J. C, 82(5):433, 2022.DOI
- [35]Jeppe R. Andersen, Andreas Maier, and Daniel Maître. Efficient negative-weight elimination in large high-multiplicity Monte Carlo event samples. Eur. Phys. J. C, 83(9):835, 2023.DOI
- [36]Rikkert Frederix and Paolo Torrielli. A new way of reducing negative weights in MC@NLO. Eur. Phys. J. C, 83(11):1051, 2023.arxiv.org/abs/2310.04160
- [37]Prasanth Shyamsundar. ARCANE Reweighting: A Monte Carlo Technique to Tackle the Negative Weights Problem in Collider Event Generation. 2 2025.arxiv.org/abs/2502.08052
- [38]Benjamin Nachman and Dennis Noll. Stay Positive: Neural Refinement of Sample Weights. 6 2025.DOI
- [39]Paolo Nason. A New method for combining NLO QCD with shower Monte Carlo algorithms. JHEP, 11:040, 2004.DOI
- [40]Stefano Frixione, Paolo Nason, and Carlo Oleari. Matching NLO QCD computations with Parton Shower simulations: the POWHEG method. JHEP, 11:070, 2007.DOI
- [41]Stefano Frixione and Bryan R. Webber. Matching NLO QCD computations and parton shower simulations. JHEP, 06:029, 2002.DOI
- [42]Melissa van Beekveld, Silvia Ferrario Ravasio, Jack Helliwell, Alexander Karlberg, Gavin P. Salam, Ludovic Scyboz, Alba Soto-Ontoso, Gregory Soyez, and Silvia Zanoli. Logarithmically-accurate and positive-definite NLO shower matching. 4 2025.DOI
- [43]Keith Hamilton, Alexander Karlberg, Gavin P. Salam, Ludovic Scyboz, and Rob Verheyen. Matching and event-shape NNDL accuracy in parton showers. JHEP, 03:224, 2023. [Erratum: JHEP 11, 060 (2023)].DOI
- [44]Melissa van Beekveld et al. Introduction to the PanScales framework, version 0.1. SciPost Phys. Codeb., 2024:31, 2024.DOI
- [45]Radja Boughezal, John M. Campbell, R. Keith Ellis, Christfried Focke, Walter Giele, Xiaohui Liu, Frank Petriello, and Ciaran Williams. Color singlet production at NNLO in MCFM. Eur. Phys. J. C, 77(1):7, 2017.DOI
