Lipschitz-Constrained Neural Networks Catch Leptons in Real Time ✛
Context
This work stands apart: the nearest neighbors in the archive are either doing spectroscopy and searches for rare decays, or machine learning for offline analysis, but no one has deployed neural networks in the hottest spot, the streaming GPU trigger. There are already neural networks for Higgs boson decays and laser diagnostics, but they run offline or on CPUs, where there is no tight budget on time and memory. This paper fills the gap between the accuracy of machine learning and the harsh requirements of real time, where every extra byte per track and every fraction of a percent of GPU load are critical.
Methods
Two separate fully connected neural networks with a Lipschitz constraint were developed: one for muons (5 layers of width 10) and one for electrons (4 layers of width 12). The constraint is imposed via weight normalization: in the first layer, each weight is individually constrained; in subsequent layers, the sum of absolute weights entering a neuron is constrained. GroupSort activations do not increase the Lipschitz constant, so the sensitivity of the output to input changes is strictly bounded. The networks were trained on simulated Pythia events with J/ψ and Υ(1S) decays into lepton pairs; background for muons is protons, for electrons pions and kaons. Input features are scaled to the range [0,1]. For muons, five variables are used: the chi-squared of the track fit to muon station hits, the sum of squared residuals, and three track slope differences between subdetectors. For electrons, six variables are used: two estimates of the energy-to-momentum ratio, the squared distance between the extrapolated track and the cluster center, and three shower shape moments in the calorimeter.
Results
Compared with the previous threshold algorithms, the neural networks provide better background rejection at the same signal efficiency in all studied kinematic regions. For muons, the gain is largest at moderate and high momenta, where most tracks pass the kinematic selection; at low momenta the difference is small. For electrons, the improvement is most noticeable at low momenta, where the calorimeter energy resolution is worse and the single E/p feature is weak. The network is nearly independent of pseudorapidity and transverse momentum. Both networks contain about 500 parameters (roughly 2 kilobytes in single precision) and add 24–28 bytes of data per track. Profiling a typical HLT1 sequence shows that the muon and electron identification kernels each consume 0.5–0.9% of the total GPU time, which is less than the threshold operators they replace. The algorithms are deployed in the live trigger and run stably on data.
Implications
Unlike neighboring works, where neural networks are applied offline or on CPUs, this work shows for the first time that strictly Lipschitz-constrained networks can live inside a streaming GPU trigger with a budget of fractions of a percent of time and kilobytes of memory. This is not just an improvement in accuracy, but a proof that controlled models are suitable for the hottest pipelines in high-energy physics. The work sets a template: how to squeeze trainable discrimination into the tight constraints of real time without losing reliability.
Future development
The method could extend to other trigger tasks: hadron identification, photon identification, track reconstruction, or whole-event selection. Lipschitz-constrained architectures allow robustness to be guaranteed in advance, which is valuable as detector conditions change over time. Deeper or convolutional variants may appear if a small footprint can be maintained. The approach could also be transferred to other experiments with hard real-time constraints, such as neutrino telescopes or future colliders.
Impact
This will affect the entire LHCb program for studying rare decays with leptons in the final state, increasing statistical precision without increasing the data rate. The technique is applicable to any real-time system in particle physics and beyond, where fast and reliable classification is needed.
Next steps
Conduct a detailed comparison of efficiency on real Run 3 data, including an assessment of systematic uncertainties. Investigate transfer learning to other particle types and adaptation to changing detector conditions.
Key open problems
In high-energy physics, the problem of optimally using the limited trigger bandwidth remains unsolved: how to retain the maximum number of rare signals under a hard limit on the output rate. This work directly addresses this issue by offering a way to improve background rejection without losing efficiency and without increasing computational cost. Another open question is the robustness of machine learning algorithms to the discrepancy between simulation and real data; the Lipschitz constraint here acts as a regularizer that mitigates this problem.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
You have developed and deployed two compact neural networks in the LHCb first-level trigger for real-time identification of muons and electrons. These networks, constrained by Lipschitz continuity and trained on simulated events, outperform the previous threshold-based algorithms in background rejection across all kinematic regions, while using only a fraction of a percent of GPU time and just 2 kilobytes of memory.
Where its strength is
The main strength is the demonstration that strictly constrained neural networks can operate under the demanding conditions of a streaming GPU trigger, combining guaranteed robustness with extremely low computational cost.
Where it could go next
- It might be interesting to look at the EveNet experience: there, the neural network was trained on 500 million collisions, and perhaps a similarly large-scale approach to training on diverse data could further improve the generalization of your compact models.based on: EveNet: How a Neural Network Learns from 500 Million Collisions arXiv:2601.17126
- In the archive, there is a nearby paper on full reconstruction of B meson production at LHCb: your networks could potentially be integrated into a broader event selection context if considered as part of a holistic trigger strategy.based on: All tracks at once: how LHCb reads the birth of B mesons in full arXiv:2508.20180
- It would be interesting to compare with the approach from the paper on neural networks for Higgs boson decays: there, the network learns to identify complex signatures, and perhaps similar architectural techniques or training methods could help extend your approach to other particle types or decay channels.based on: Neural Network Learns to Spot Higgs Boson Decays arXiv:2604.09809
Worth trying
- Test whether training on a wider set of simulated events (in the spirit of EveNet) provides additional background suppression without losing simplicity.
- Assess whether your networks can serve as building blocks for more complex trigger decisions, for example, selecting events with B mesons, building on experience with full track reconstruction.
- Investigate whether methods from the paper on neural networks for Higgs decays could be applied to improve electron identification in regions with poor energy resolution.
Why this may matter
This work paves the way for the systematic use of strictly constrained neural networks in online systems of high-energy physics, where simple threshold algorithms have previously dominated. Showing that it is possible to fit within tight time and memory budgets may encourage broader adoption of machine learning in triggers of other experiments.
Nearby works in our archive
- All tracks at once: how LHCb reads the birth of B mesons in fullsimilarity 0.52
- How to Catch a Rare Signal in a Sea of Datasimilarity 0.52
- Digital Conductor of Gravitational Waves: Neural Network Conducts Black Hole Mergerssimilarity 0.51
- EveNet: How a Neural Network Learns from 500 Million Collisionssimilarity 0.50
- Heavy Higgs Twin Hides in Four Jetssimilarity 0.50
- Teaching Cameras to See Faster and Cheaper: On-Chip Neuromorphic Video Processingsimilarity 0.50
- Neural Network Learns to Spot Higgs Boson Decayssimilarity 0.49
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
The muon identification network takes up less memory than a single 32×32 color PNG image, yet it processes thousands of tracks per second on a GPU.
Key numbers
- input event rate: 30 MHz
- GPU time for both networks: 0.5–0.9% of total
- parameters in the network: ~500 (about 2 kB)
- additional data per track: 24–28 bytes
- layers in the muon network: 5 (width 10)