Reducing negative weights in Monte Carlo event generation with Sherpa

Abstract

An increase in theoretical precision of Monte Carlo event generators is typically accompanied by an increased need for computational resources. One major obstacle are negative weighted events, which appear in Monte Carlo simulations with higher perturbative accuracy. While they can be handled somewhat easily in fixed-order calculations, they are a major concern for particle level event simulations. In this article, the origin of negative weights in the S-Mc@Nlo method is reviewed and mechanisms to reduce the negative weight fraction in simulations with the Sherpa event generator are presented, with a focus on V +jets and t t+jets simulations.

Modelling and computational improvements to the simulation of single vector-boson plus jet processes for the ATLAS experiment

“…In the Sherpa 2.2.11 configurations, the MC@NLO matching does not account for subleading gluon spin and colour correlations known to have small effects on final-state observables [42]. This results in an overall reduction in the fraction of negative weights, while preserving the essential features of the matched prediction [43]. In addition, the LO contributions are reweighed with a differential NLO/LO K-factor using a 2 → 2 core process obtained through an inverse parton-shower clustering.…”

Higher-order EW corrections in ZZ and ZZj production at the LHC

“…(3.7) † It is interesting to note that [102] adds an additional Sudakov factor with respect to the n-parton configuration to the H-events. This is done to, among other objectives, reduce the impact of negative weights.…”

Unbiased elimination of negative weights in Monte Carlo samples

“…relative to the required number N (0) in the absence of negative weights, see also [16]. Consequently, there have been a number of recent efforts to reduce the fraction r − of the negative weight contribution.…”

“…Consequently, there have been a number of recent efforts to reduce the fraction r − of the negative weight contribution. One avenue that is being explored is to optimise the event generation itself [16][17][18][19][20]. An alternative approach is to remove negative weights from the generated samples, while taking care that observables are not affected [21][22][23].…”