⚛️ phenomenology

Reweighting Underlying Event and Colour Reconnection parameter variations in Sherpa

This paper proposes and validates a computationally efficient method for tracing parameter variations in Sherpa's multi-parton interaction and colour reconnection models through event-by-event reweighting, enabling robust uncertainty quantification and tuning using LHC and Tevatron data.

Original authors: Moritz Pabst, Max Knobbe, Frank Krauss, Steffen Schumann

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Moritz Pabst, Max Knobbe, Frank Krauss, Steffen Schumann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the outcome of a chaotic, high-speed car race. You have a super-complex computer simulation that models every car, every tire friction, and every gust of wind. But to make the simulation match reality, you have to tweak dozens of "knobs" and "dials" (parameters) that represent things we can't calculate perfectly, like how the cars interact when they bump into each other or how the track surface changes.

In the world of particle physics, this simulation is called a Monte Carlo event generator (specifically, one named SHERPA). The "knobs" it needs to tweak are related to the Underlying Event (the messy background noise of particles created when protons collide) and Colour Reconnection (a quantum mechanical game of musical chairs where particles swap partners before they settle down).

The Problem: The "Re-Run" Bottleneck

Traditionally, if a physicist wanted to see what would happen if they turned one of these knobs slightly, they had to stop the simulation, change the setting, and run the entire massive calculation from scratch.

If you wanted to test 500 different combinations of these knobs to find the perfect setting, you would have to run the simulation 500 separate times. This is like asking a chef to cook a full banquet 500 times, changing just one spice in the soup each time, just to see which version tastes best. It takes forever, costs a fortune in computer power, and fills up hard drives with thousands of nearly identical files.

The Solution: The "Magic Weight" Trick

This paper introduces a clever new method called reweighting.

Instead of cooking the banquet 500 times, the authors found a way to cook it once (using a "central" or "nominal" setting). Then, they developed a mathematical "magic trick" that allows them to look at that single meal and instantly calculate what it would have looked like if they had used any of the other 500 spice combinations.

They do this by assigning a weight to every single particle collision in that one simulation.

If a specific particle collision would have been more likely with a different knob setting, its weight goes up.
If it would have been less likely, its weight goes down.
If it would have been impossible (like trying to fit a square peg in a round hole), its weight becomes zero.

By simply multiplying the data by these weights, they can generate the results for hundreds of different scenarios without ever running the heavy simulation again.

The Analogy: The "What-If" Photo Filter

Think of the simulation as taking a single, high-resolution photo of a crowded party.

The Old Way: To see what the party would look like if the lighting were warmer, or if everyone wore hats, you had to organize the party again, change the lights, and take a new photo. Do this 500 times.
The New Way: You take one photo. Then, you use a sophisticated app that applies a "filter" to every single person in the photo. The app calculates: "If the lighting were warmer, this person would look 10% brighter; if hats were mandatory, this person's head would be 20% larger." The app instantly generates 500 different versions of the party from that one photo, just by adjusting the numbers attached to each person.

What They Did in the Paper

Built the Trick: They wrote the code to calculate these "magic weights" for the specific physics of particle collisions (MPI and Colour Reconnection) inside the SHERPA software.
Tested the Trick: They ran the simulation normally, then used the trick to predict what would happen with different settings. They compared these predictions to actual "re-runs" (the old, slow way) and found they matched perfectly.
Found the Best Settings: They used this method to tune the SHERPA simulation against real data from the Large Hadron Collider (LHC). They tested 500 different combinations of knobs in the time it usually takes to test just a few.
The Result: They found that the simulation works best when "Colour Reconnection" is turned on. They also figured out how to adjust these settings for different energy levels (like moving from 7 TeV to 13 TeV collisions) without needing new data for every single step.

Why It Matters

This method is a massive time-saver. It reduces the computer time needed for these studies by a factor of roughly 250. It means physicists can now:

Tune their models much faster.
Understand how sensitive their results are to small changes in the physics (uncertainty quantification).
Do all this without filling up their hard drives with thousands of redundant files.

In short, they turned a process that required running a marathon 500 times into a process that requires running it once and then doing a very fast mental calculation to see the other 499 outcomes.

Technical Summary: Reweighting Underlying Event and Colour Reconnection Parameter Variations in SHERPA

Problem Statement
Precision phenomenology in high-energy physics relies on Monte Carlo Event Generators (MCEGs) to simulate the transition from perturbative partonic scatterings to non-perturbative hadronic final states. While uncertainties in perturbative regimes (e.g., scale variations, PDFs) are efficiently handled via on-the-fly event reweighting, significant theoretical uncertainties arise from non-perturbative (NP) modeling, specifically Multiple Parton Interactions (MPI) and Colour Reconnection (CR). These models depend on empirical parameters that must be "tuned" to experimental data. Currently, assessing the sensitivity of observables to these parameters or quantifying their uncertainties requires generating dedicated simulation runs for numerous points in the parameter space. This process is computationally prohibitive, often limiting uncertainty studies to superficial comparisons between only a few generator tunes or versions.

Methodology
The authors introduce and validate a new method for on-the-fly reweighting of MPI and CR parameters within the SHERPA event generator. The approach allows for the propagation of parameter variations without regenerating events, capturing the effects through relative event weights.

MPI Reweighting: The MPI model in SHERPA is based on the Sjöstrand–van-Zijl (SZ) formalism, utilizing a geometric overlap of hadronic matter distributions and a Sudakov form factor to regulate the probability of secondary scatters. The reweighting factor is decomposed into two components:
1. Geometric Weight ( $w_b$ ): Accounts for changes in the impact-parameter sampling due to variations in matter distribution parameters (radii $R_{1,2}$ and weights $\alpha_1$ ).
2. Sudakov Weight ( $w_{\text{Sudakov}}$ ): Accounts for variations in the acceptance probability of secondary scatters. This is constructed by comparing the true integrand of the varied parameter set against the original set for every accepted and rejected scattering in the event generation sequence.
- Constraint: Variations in the transverse-momentum cutoff ( $p_{\perp,\min}$ ) are unidirectional (only upward variations are possible) because events generated with a lower cutoff do not populate the phase space excluded by a higher cutoff.
CR Reweighting: The Colour Reconnection model rearranges colour lines between partons to minimize a distance measure in momentum space before hadronisation. The reweighting factor is calculated multiplicatively for every reconnection attempt, comparing the acceptance probability $P_{\text{acc}}$ of the varied parameters against the nominal parameters for both accepted and rejected swaps.
Tuning Framework: The method is integrated with the Apprentice framework to construct multivariate surrogate models (polynomial approximations) of the generator response. This enables a global fit of MPI and CR parameters to LHC data (7 TeV) using a single nominal event sample reweighted to 500 random parameter configurations.

Key Contributions

First Implementation: This work presents the first implementation of on-the-fly reweighting for MPI and CR models in SHERPA, extending existing perturbative reweighting capabilities to key non-perturbative components.
Computational Efficiency: By generating a single nominal sample and reweighting it to cover a 4-dimensional (MPI-only) or 5-dimensional (MPI+CR) parameter space, the authors reduce the computational cost of tuning by a factor of approximately 250 compared to generating dedicated samples for each point.
Validation: The method undergoes rigorous "closure tests." Predictions from reweighted samples are compared against dedicated generator runs for specific parameter points. The results show excellent agreement for intrinsic model variables (impact parameter, MPI multiplicity, CR multiplicity) and physical observables (charged-particle densities, transverse momentum sums) across Drell–Yan and jet production processes.
Statistical Characterization: The paper quantifies the "statistical dilution" caused by reweighting using the Kish effective sample size ( $N_{\text{eff}}$ ). It demonstrates that while $N_{\text{eff}}$ decreases for parameter points far from the nominal set, it remains sufficient for tuning within the studied ranges (e.g., $N_{\text{eff}} \approx 0.5N$ to $0.7N$ ).

Results

Tuning at 7 TeV: The authors performed combined tunes for SHERPA MPI and CR models using ATLAS and CMS data.
- MPI-CRoff: A tune with CR disabled yielded a best-fit set of parameters.
- MPI-CRon: A tune with CR enabled provided a significantly better description of the data, particularly for Drell–Yan processes where the $\chi^2/\text{ndf}$ was effectively halved compared to the CR-off scenario. This indicates that experimental data favors the inclusion of colour reconnections.
Energy Extrapolation: The authors calibrated the energy scaling of dimensionful MPI parameters ( $p_{\perp,0}$ and $p_{\perp,\min}$ ) using the scaling exponent $\eta$ . By reweighting a single sample to cover $\eta \in [0.0, 0.08]$ , they determined an optimal value of $\eta^* = 0.026$ . This single parameter successfully described underlying-event data from the Tevatron ( $\sqrt{s} = 1.96$ TeV) and the LHC ( $\sqrt{s} = 13$ TeV) simultaneously.
Performance: The reweighting runtime scales linearly with the number of variations but remains significantly faster than generating equivalent dedicated samples. The effective sample size remains robust enough to support high-precision tuning.

Significance and Claims
The paper claims that this reweighting method closes a significant gap in the capabilities of modern MCEGs, enabling efficient exploration of non-perturbative parameter spaces. The authors position this technique as essential for:

Parameter Sensitivity Studies: Allowing for the robust quantification of parametric uncertainties on-the-fly.
Tuning: Drastically reducing the computational cost of generator tuning, facilitating the use of advanced methods like History Matching (HM) which require dense sampling of parameter space.
Future Simulations: Providing a foundation for the high-luminosity LHC era, where precise control over non-perturbative uncertainties is critical.

The authors explicitly state that while the method is demonstrated for MPI and CR, it is designed to be adaptable to other event generators and other non-perturbative models (such as hadronisation), with future work planned to extend this to SHERPA's cluster fragmentation and minimum-bias simulations. The method is presented not as a replacement for dedicated runs in all cases, but as a powerful tool to alleviate the computational burden of parameter sensitivity studies and uncertainty quantification.

The Problem: The "Re-Run" Bottleneck

The Solution: The "Magic Weight" Trick

The Analogy: The "What-If" Photo Filter

What They Did in the Paper

Why It Matters

More like this