Source-linked AI summary
Event Generation with Normalizing Flows
Christina Gao, Stefan Hoeche, Joshua Isaacson, Claudius Krause, Holger Schulz
TL;DR
Precision event generation faces computational-resource demands associated with declining unweighting efficiency. The paper combines Normalizing Flows with recursive multichannel phase-space integration; in simple processes, the tuned integrator exceeds conventional methods by a factor of 2–3, while high-multiplicity processes remain a more difficult regime.
Problem
Projected computing resources fall short of precision event-generation needs, and declining unweighting efficiency drives much of the demand for multijet simulations.
Source: §I · p. 1Method
The integrator combines a Normalizing Flow with recursive multichannel phase-space integration in Sherpa-based event generation.
Source: §I · p. 1, §III · p. 4, §IV · p. 5Results
In simple processes, the tuned integrator’s unweighting efficiency exceeds conventional methods by a factor of 2–3; at high multiplicity, traditional techniques perform similarly well.
Source: §V · p. 7, C. Comparison to existing approaches · p. 7Takeaways & Limitations
The approach improves unweighting efficiency in simple processes, while traditional techniques perform similarly in high-multiplicity processes and require fewer computing resources.
Source: §V · p. 7At high multiplicity, traditional techniques tend to perform similarly well while requiring fewer computing resources.
Source: §V · p. 7
Abstract
from arXiv · showhide
We present a novel integrator based on normalizing flows which can be used to improve the unweighting efficiency of Monte-Carlo event generators for collider physics simulations. In contrast to machine learning approaches based on surrogate models, our method generates the correct result even if the underlying neural networks are not optimally trained. We exemplify the new strategy using the example of Drell-Yan type processes at the LHC, both at leading and partially at next-to-leading order QCD.
I. INTRODUCTION
Precision event generation is increasingly constrained by computational resources, with declining unweighting efficiency driving much of the demand. The paper proposes a Normalizing-Flow integrator combined with recursive multichannel sampling and tests it on Drell–Yan processes at the LHC.
I. INTRODUCTION
- Declining unweighting efficiency largely determines the computing needs of multijet simulations, while projected resources fall short of precision event-generation demands.Dedicated integrators require detailed knowledge of the physics problem; adaptive Monte Carlo methods are presented as the general-purpose alternative.Source: §I · p. 1
- The paper’s stated motivation is the growing precision-simulation demand associated with increasing LHC data and the high-luminosity upgrade.The manuscript proceeds from Monte Carlo and Normalizing Flow background to the integrator, computing performance and applications, then an outlook.Source: §I · p. 1
- Neural-network variable transformations can improve integration efficiency, but computing their gradients at every inference point makes them costly for high-multiplicity processes.Other machine-learning event-generation approaches use training data from traditional generators and can yield incorrect results if their networks do not approximate the integrand well.Source: §I · p. 1
- The proposed integrator replaces standard adaptive algorithms such as Vegas with a Normalizing Flow and combines it with recursive multichannel sampling.It is tested on Drell–Yan processes at leading and partially next-to-leading order, with event-generation efficiency compared to Comix.Source: §I · p. 1
II. NORMALIZING FLOWS
Monte Carlo integration estimates an integral by sampling and weighting points, while importance sampling can reduce its variance. A Normalizing Flow uses a bijective sequence of transformations to map between distributions, with tractable Jacobians shaping the design.
II. NORMALIZING FLOWS
- Monte Carlo integration estimates an integral from sampled, integrand-weighted points, and importance sampling or stratified sampling can reduce its variance.The sampling distribution is uniform in x before the change of variables.Source: §II · p. 2
- The sampling density g(x) should resemble the integrand to reduce variance, while remaining integrable and invertible for efficient sampling.In the limit g(x) → f(x)/I, the stated estimate has vanishing uncertainty.Source: §II · p. 2
- A Normalizing Flow maps a variable drawn from a flat distribution through a bijection so that the transformed variable follows a target probability distribution.Composing transformations Gk defines a bijective mapping between the distributions of the input and final variables.Source: §II · p. 2
- Flow transformations are often kept simple to make Jacobian determinants easy to compute, while coupling layers or autoregressive models increase the complexity that can be modeled.This study focuses on a generalization of coupling layers.Source: §II · p. 2
A. Coupling Layers
A coupling layer partitions its input and transforms one subset using a bijective map whose parameters are produced from the other subset. This construction keeps the Jacobian computation tractable and can be compounded into a complete flow.
A. Coupling Layers
- A coupling layer partitions the input into xA and xB, leaves xA unchanged, and transforms xB with C parameterized by a neural network of xA.The figure caption identifies the network output m as defining the coupling transform applied to xB.Source: A. Coupling Layers · p. 2, Figure 1 · p. 3
- The network m takes xA as input and outputs parameters for the invertible coupling transform C applied to xB.The inverse map is also specified in the section.Source: A. Coupling Layers · p. 2
- The coupling-layer Jacobian avoids computing the gradient of m, and its diagonal transform makes determinant computation scale linearly with dimension.The paper states that computing the gradient of m would scale as O(D^3).Source: A. Coupling Layers · pp. 2–3
- A complete flow compounds coupling layers, choosing which input dimensions to transform at each layer; the implementation requires a neural network, an integrand, and a loss function.The paper states that at most 2⌈log2 D⌉ layers suffice to express arbitrarily complicated nonseparable integrand structures.Source: A. Coupling Layers · p. 3
B. Piecewise Polynomial and Rational Quadratic Spline Coupling Transforms
The coupling transform is designed as a flexible, monotonic approximation to a cumulative distribution function on the unit hypercube. The paper describes piecewise linear, quadratic, and rational-quadratic choices with adjustable bin parameters.
B. Piecewise Polynomial and Rational Quadratic Spline Coupling Transforms
- The coupling transform can be interpreted as the cumulative distribution function of the variable, modeled by a monotonically increasing polynomial in each bin.Each dimension is divided into K bins, and the coupling layer maps the unit hypercube to itself.Source: B. Piecewise Polynomial and Rational… · p. 3
- Piecewise linear and quadratic transforms are described, with variable bin widths in the quadratic implementation to increase flexibility.Source: B. Piecewise Polynomial and Rational… · p. 3
- The rational-quadratic variant allows bin heights, widths, and inter-bin derivatives to vary, with these parameters predicted by the neural network.It generalizes the piecewise quadratic transform using a monotonically increasing rational-quadratic function in each bin.Source: B. Piecewise Polynomial and Rational… · p. 3
III. PHASE-SPACE INTEGRATION
The integrator combines recursive multichannel phase-space decomposition with a Normalizing Flow in place of Vegas. The recursion reduces factorial scaling to exponential scaling, while adjustable weights and a shared network support integration across channels.
III. PHASE-SPACE INTEGRATION
- Recursive phase-space computation uses building blocks corresponding to t- and s-channel processes, with scaling reduced from factorial to exponential by dynamic programming.The phase-space weights can be computed recursively, with the same exponential scaling as hard matrix-element computation.Source: §III · pp. 3–4
- The complete phase space is decomposed into t- and s-channel decay processes and an s-channel production process that enforces overall momentum conservation.The phase-space recursion factorizes the phase space and constructs its integration from these building blocks.Source: §III · p. 4
- Averaging over ordered partitions and assigning adjustable weights to the splittings defines the recursive multichannel integrator.The weights are vertex-specific, normalized, and can be adapted to optimize the integrator.Source: §III · p. 4
- The method replaces Vegas with a Neural Network plus Normalizing Flow, includes multichannel weights, and uses one network for all channels.Other integrator components remain the same, enabling direct performance comparisons with existing algorithms.Source: §III · p. 4
IV. NUMERICAL RESULTS
The numerical study uses Sherpa with Comix matrix elements and tests combined color-momentum training for the Neural Network plus Normalizing Flow integrator.
IV. NUMERICAL RESULTS
- Combined color-momentum training was tested but showed no advantage for processes beyond W/Z + 1j.The study used Sherpa, with matrix elements provided by Comix; Comix samples color degrees of freedom with Monte Carlo.Source: §IV · p. 5
A. Definition of efficiency
The paper assesses unweighting efficiency while accounting for outliers and the differing maximum weights encountered in distributed event-generation jobs. It estimates a robust maximum from bootstrap replicas and permits partial event weights when weights exceed that estimate.
A. Definition of efficiency
- The basic unweighting efficiency is the average event weight divided by the maximum weight, and its inverse is the average number of draws before unit-weight acceptance.If the weight distribution has no sharp upper edge, the maximum depends on sample size and efficiency decreases as more events are generated.Source: A. Definition of efficiency · p. 5
- Efficiency is based on points from the final optimization iteration because distributed jobs cannot continue adaptive optimization in coordination, and weighted samples can be prone to outliers.A bootstrap is used to remove possible outliers from the estimate.Source: A. Definition of efficiency · p. 5
- The bootstrap estimates the total maximum weight as the median of maxima from replicas, each containing Nopt events.It resamples nNopt events into m replicas, then takes the median of their individual maxima.Source: A. Definition of efficiency · p. 5
- Events exceeding the bootstrap maximum are retained with relative weights, allowing the sample to be partially weighted against that statistically estimated maximum.The method can in principle reach arbitrary precision using bootstrap and jackknife techniques.Source: A. Definition of efficiency · p. 5
B. Hyperparameter Optimization
Hyperparameters are tuned by quasirandom search using unweighting efficiency as the figure of merit. The scan favors a high learning rate, more epochs, and fewer bins, with separate-scan coverage limited for higher jet multiplicities.
B. Hyperparameter Optimization
- Sobol sequences populate the hyperparameter space because training is computationally costly and computing resources are limited.The search follows a quasirandom strategy for tuning neural-network hyperparameters.Source: B. Hyperparameter Optimization · pp. 5–6
- Table I reports the hyperparameter ranges, prior distributions, and optimal scan point, and defines LR as learning rate.Source: Table I · p. 6
- Unweighting efficiency is the figure of merit, with wmax set to the 99.9th percentile of nonzero absolute weights for tuning.This choice was tested for stability against random-seed variation and gave results qualitatively compatible with the earlier definition.Source: B. Hyperparameter Optimization · p. 6
- The scan suggests that a high learning rate with many epochs and a small number of bins is beneficial.Figure 2 marks the best-performing configuration with a star; the scan also found a comparable best configuration and similar efficiency for W + 2 jets.Source: B. Hyperparameter Optimization · p. 6, Figure 2 · p. 6
- No separate hyperparameter scan was performed for W+ ≥3 jets because computing resources were limited.The reported W + 2 jets scan found a best configuration comparable to W + 1 jet and similar unweighting efficiency.Source: B. Hyperparameter Optimization · p. 6
C. Comparison to existing approaches
Against Sherpa, the normalizing-flow integrator has much higher unweighting efficiency in few-jet processes at both LO and NLO, but performs similarly in higher-multiplicity processes. Its results depend on the integration structure and training resources, with particularly low efficiencies reported for Z-boson production.
C. Comparison to existing approaches
- The comparison considers W± and Z production in proton-proton collisions at 14 TeV, using NNPDF3.0 NNLO PDFs and the corresponding strong coupling.Jets use kT clustering with R = 0.4, pT,j > 20 GeV, and |ηj| < 6; Z/γ* production also applies 66 < mll < 116 GeV.Source: C. Comparison to existing approaches · p. 6, Table II · p. 7
- The flow integrator has much higher unweighting efficiency than Sherpa for few-jet LO and NLO processes, but performs similarly for processes with more final-state jets.Table II compares unweighting efficiencies for LHC production at √s = 14 TeV.Source: Table II · p. 7, C. Comparison to existing approaches · p. 7
- The neural-network method generally performs better without multichannel initial-state integration, and its lower Z-boson efficiencies arise when combining 1/ŝ and Breit-Wigner mappings.The authors interpret this as the flow approximating smooth integrand structures more efficiently than differentiating between effectively independent integration domains.Source: C. Comparison to existing approaches · p. 7
- The method is best suited to low-multiplicity problems, while possible gains at high multiplicity require more training epochs and samples than the study could verify for W±/Z+4j.The authors could not test that high-multiplicity claim because computing resources were restricted.Source: C. Comparison to existing approaches · p. 7
- Future accelerator implementations of matrix-element generators could alter the high-multiplicity performance comparison.Source: C. Comparison to existing approaches · p. 7
V. CONCLUSIONS
The Sherpa add-on normalizing-flow integrator exceeds conventional methods in unweighting efficiency by a factor of 2–3 for simple processes, while traditional techniques perform similarly in high-multiplicity processes. The authors expect accelerator implementations to change that comparison.
V. CONCLUSIONS
- For simple processes, the integrator's unweighting efficiency exceeds conventional methods by a factor of 2–3, while traditional techniques perform similarly in high-multiplicity processes.Source: §V · p. 7
- Implemented as a Sherpa add-on, the integrator works with the Comix and Amegic internal matrix-element generators, and its neural-network hyperparameters were tuned by quasirandom search.Source: §V · p. 7
- The authors expect accelerator implementations such as GPUs and TPUs to change the high-multiplicity performance picture, where traditional techniques currently require fewer computing resources.Source: §V · p. 7