🔬 physics

Positive-weight moment matching at NNLO and beyond

This paper introduces a novel method for generating positive-weight event samples that simultaneously reproduce high-precision IR-safe observables up to NNLO accuracy for processes like Drell-Yan and $Z+\text{jet}$ , thereby overcoming the statistical limitations caused by negative weights in state-of-the-art parton-shower simulations.

Original authors: Benoît Assi, Rene Poncelet, James Whitehead

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Benoît Assi, Rene Poncelet, James Whitehead

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

In the world of high-energy physics, scientists use massive particle colliders to smash protons together at nearly the speed of light. These collisions create a shower of new particles that fly out in all directions, carrying clues about the fundamental laws of nature. To understand what happens in these collisions, researchers rely on two main tools. The first is the detector, a giant machine that records the actual debris from the crash. The second is a computer simulation, a virtual model that predicts what the debris should look like based on our current theories. For these simulations to be useful, they must be incredibly precise, matching the theory to the highest possible level of detail. However, a persistent problem has long plagued these computer models: the way they generate data often produces "negative weights." In a physical sense, a negative weight is impossible; you cannot have negative probability. But in the math used to build these simulations, negative numbers appear as a necessary correction to fix errors. When a simulation contains too many of these negative numbers, the data becomes noisy and unreliable, requiring scientists to generate vastly more events to get a clear picture. This creates a bottleneck, making it difficult to produce the large, clean datasets needed for the next generation of experiments.

A team of researchers has now developed a new method to bypass this problem entirely. Instead of trying to fix the negative numbers within the simulation itself, they created a way to take a standard, positive-only simulation and adjust it so that it matches the most precise theoretical calculations available, without ever introducing a single negative weight. The researchers started with a basic simulation of proton collisions, specifically looking at a process called Drell–Yan, where protons collide to produce a pair of leptons, such as electrons or muons. This initial simulation was generated using a standard tool called Pythia, which is excellent at modeling how particles spray out after a collision but lacks the highest level of precision for the initial crash. The team then took the most advanced theoretical calculations available, known as next-to-next-to-leading order, which describe the collision with extreme accuracy but are computationally expensive and often come with the problematic negative weights.

The core of their new technique involves a mathematical process called maximum-entropy reweighting. Imagine you have a large bag of marbles representing the initial simulation, and you want to rearrange them so that their distribution matches a specific, highly accurate pattern. Instead of throwing marbles away or adding new ones, you simply change the "importance" or weight assigned to each marble. The researchers used a set of mathematical constraints, based on the properties of the particles like their speed and direction, to guide this adjustment. They calculated how the most precise theory predicted these properties should behave and then adjusted the weights of the marbles in their simulation until the simulation's average behavior matched those predictions perfectly. Crucially, the method they used guarantees that every marble keeps a positive weight. By doing this, they created a single, unified dataset that is both positive-weighted and accurate to the highest known theoretical order.

The results of this approach are striking. When the team applied their method to the Drell–Yan process, they achieved a level of accuracy previously impossible in a single, positive-weighted sample. They were able to simultaneously reproduce the precise predictions for the total number of events and the detailed shapes of distributions for particles with jets of energy, all while maintaining a clean dataset with no negative weights. They tested this against other existing methods that try to combine theory and simulation. Those older methods often struggle with the negative weights, which dilute the statistical power of the data, meaning scientists need to run the simulation for much longer to get the same level of certainty. In contrast, the new method produced a sample that was statistically just as powerful as the original simulation but with the precision of the most advanced theory. The researchers also checked their work against real data collected by the ATLAS experiment at the Large Hadron Collider. The adjusted simulation matched the real-world observations remarkably well, correctly describing the behavior of particles across a wide range of energies.

This work represents a significant shift in how physicists might approach the future of particle physics simulations. The method is flexible and can be applied to different types of collisions, different levels of theoretical precision, and even different particle generators. It allows scientists to take a cheap, fast simulation and upgrade it with expensive, high-precision theory without the computational cost of generating new data from scratch. The researchers demonstrated that this technique works not just for simple collisions but also for more complex scenarios involving Higgs bosons, another fundamental particle. By keeping the correlations between particles intact and only adjusting the weights to match the theory, the method preserves the natural behavior of the simulation while correcting its accuracy. This approach offers a practical solution to the growing demand for high-quality data in the coming years, particularly for the High-Luminosity Large Hadron Collider, which will produce data at a rate that current simulation methods may struggle to handle. The ability to generate large, positive-weighted samples with state-of-the-art accuracy means that future experiments can be analyzed with greater confidence and less computational waste, ensuring that the search for new physics remains on solid ground.

Technical Summary: Positive-weight Moment Matching at NNLO and Beyond

Problem Statement
State-of-the-art parton-shower simulations, particularly those aiming for Next-to-Next-to-Leading Order (NNLO) perturbative accuracy, frequently produce event samples containing negative event weights. While necessary for the consistency of certain matching schemes (e.g., MC@NLO, POWHEG, MiNNLOPS), these negative weights irreducibly increase the variance of observables. This "variance penalty" dilutes the statistical power of the simulation, rendering large-scale samples computationally prohibitive for detector simulation and unfolding, especially given the precision requirements of the High-Luminosity LHC (HL-LHC) program. Existing mitigation strategies, including weight-reduction techniques and alternative matching methods, have failed to combine state-of-the-art perturbative accuracy with strictly positive event weights in a single hadronised sample.

Methodology
The authors propose a method based on Maximum Entropy (MaxEnt) reweighting to generate positive-weighted event samples that reproduce high-precision fixed-order moments. The approach treats the generation of a high-accuracy sample as a constrained optimization problem:

Prior Sample: The method starts with an existing "prior" sample (e.g., a Leading-Order + Parton Shower [LO+PS] sample generated by Pythia) consisting of events $\Phi_i$ with positive weights $p_i$ .
Constraints: The target is defined by a set of fixed-order moments $\mu$ calculated to $N^k$ LO accuracy (e.g., NNLO) for specific IR-safe observables. These moments are derived from parton-level Monte Carlo codes (such as NNLOJET and STRIPPER) or cached event samples.
Optimization: The algorithm seeks new weights $q_i$ $q_{i}$ that minimize the relative entropy (Kullback–Leibler divergence) with respect to the prior $p_i$ $p_{i}$ , subject to the constraint that the reweighted sample reproduces the target moments $\mu$ $μ$ .
- The objective function is strictly convex: $\min \sum q_i \ln(q_i/p_i)$ .
- The solution takes the closed form $q_i = p_i e^{\lambda \cdot m_i} / Z(\lambda)$ , where $m_i$ are measurement functions and $\lambda$ are Lagrange multipliers.
- Because the exponential function is strictly positive, the resulting weights $q_i$ are guaranteed to be positive, provided the prior has support in the target region.
Moment Construction: To ensure smoothness and avoid binning artifacts, the method utilizes Chebyshev polynomial moments of observables mapped to $[-1, 1]$ . A window function $w_v$ blends the fixed-order calculation (valid in high-energy regions) with the prior (valid in soft/collinear regions) to maintain theoretical accuracy across the phase space.
Systematic Improvability: The method is designed to be systematically improvable. As higher-order moments or additional observables become available, existing samples can be rapidly refined by re-solving the optimization problem without regenerating the underlying events.

Key Results
The authors demonstrate the method using an LO+PS Pythia sample for Drell–Yan ( $pp \to \ell^+\ell^-$ ) production at $\sqrt{s} = 13$ TeV.

Simultaneous NNLO Accuracy: The reweighted sample achieves simultaneous NNLO accuracy for inclusive observables (e.g., dilepton mass $m_{\ell\ell}$ , rapidity $|y_{\ell\ell}|$ ) and NLO/NNLO accuracy for $Z + \text{jet}$ observables (e.g., dilepton transverse momentum $p_{\ell\ell}^T$ ), all within a strictly positive-weighted sample.
Held-Out Validation: The method was validated on "held-out" observables not used in the reweighting constraints, such as the leading-lepton transverse momentum ( $p_{\ell_1}^T$ ) and the angular variable $\phi^*_\eta$ . Despite the lack of explicit constraints, the reweighted sample closely reproduced the fixed-order NNLO distributions, demonstrating that the kinematic correlations inherent in the prior are successfully leveraged to propagate higher-order information.
Comparison with Data: The reweighted samples show good agreement with ATLAS 13 TeV data for $Z$ -boson transverse momentum and $\phi^*_\eta$ , reconciling fixed-order accuracy at high $p_T$ with the resummation properties of the parton shower at low $p_T$ .
Statistical Efficiency: Unlike standard NNLO matching methods (e.g., MiNNLOPS) which suffer from significant negative-weight fractions (e.g., $f_- \approx 23.5\%$ ), the MaxEnt method maintains $f_- = 0$ . Consequently, the effective sample size ( $N_{\text{eff}}$ ) remains nearly identical to the input sample size ( $N_{\text{eff}}/N \approx 99.9\%$ ), offering a massive computational advantage in terms of CPU cost per effective event.
Generalizability: The method was also applied to gluon-fusion Higgs production ( $H \to \gamma\gamma$ ), achieving similar results with positive weights and improved statistical power compared to MC@NLO and MiNNLOPS.

Significance and Claims
The paper claims to present a paradigm shift in Monte Carlo event generation by decoupling the generation of the phase-space topology (handled by the prior) from the imposition of high-order perturbative accuracy (handled by the reweighting).

Positive Weights: It is the first method to achieve simultaneous NNLO accuracy for inclusive and jet-associated observables in a strictly positive-weighted event sample.
Computational Efficiency: The method is computationally cheaper than existing NLO/NNLO matching techniques when moments are known externally, as it avoids the overhead of generating negative-weight events and the subsequent statistical penalty.
Flexibility: The approach is prior-agnostic and can be applied to any process or perturbative order where fixed-order calculations exist. It allows for the systematic refinement of samples as theoretical calculations improve, addressing the bottleneck of generating large, high-precision samples for the HL-LHC.
Theoretical Consistency: By preserving the correlations of the prior outside the constrained observables, the method retains the all-orders resummation properties of the parton shower while correcting the hard process to fixed-order accuracy.

The authors emphasize that the method does not require reformulating NNLO subtraction algorithms to be manifestly positive nor interfacing NNLO codes directly to parton showers, relying instead on the calculation of moments which can be cached or interpolated.

Technical Summary: Positive-weight Moment Matching at NNLO and Beyond

More like this