Lorentz-Equivariance without Limitations
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: We're looking at the paper: "Lorentz-Equivariance without Limitations".
Mira: Here's the gist of the paper. Lorentz Local Canonicalization (LLoCa) ensures exact Lorentz-equivariance for arbitrary neural networks with minimal computational overhead The paper introduces LLoCa, a framework that predicts local reference frames for each particle and propagates any-order tensorial information between them It is applied to graph networks and transformers, showcasing cutting-edge performance on amplitude regression, end-to-end event generation, and jet tagging Lorentz Local Canonicalization Mechanism LLoCa builds upon the concept of local canonicalization and extends it to Lorentz symmetry by predicting local reference frames for each input particle The construction of a local frame is achieved using four orthonormal unit vectors eaµ with the definition L = (3) and the transformation rule L'aµ = L aν (Λ−1) ν µ (5) Local frames constructed as in Eq.
Kai: (4) satisfy the required transformation behaviour of Eq.
Mira: (5) since L aµ = e aµ Λ→ L′aµ = e′aµ = e aν (Λ−1) ν µ = L a ν (Λ−1) ν µ, as by definition of the transformation rule of covectors The transformation behavior of Lorentz tensors in the LLoCa formalism is read off from their index structure, for example xµ Λ→ Λµν xν xµ Λ→ xν (Λ−1) ν fµν λ Λ→ Λµ σΛ ν κ fσγ (Λ−1) γλ x aL L(x L)aΛ→ (x L)a (fL) ab c Λ→ (fL) ab c Equivariance and Message Passing Lorentz-equivariance is achieved by transforming the input vector xµ into the local frame yielding the invariant object x aL = L aµ x µ (8) The output vector yµ is then transformed back to the global frame as yµ = (L−1) µa y aL (9) This leads to the output vector satisfying the correct transformation behavior yµ Λ→ y′µ = Λµν (L−1) ν a y aL = Λµν y ν For message passing networks, the tensorial message passing is generalized to f updated i,Li= ψ f i,Li N □ j=1 φ(mjj) (14) The received message mj,Li is related to the sender's frame by mj,Li = ρm(Li L−1j)mj,Lj (15) This generalization leads to the updated message f updated i,Li= ψ f i,Li N □ j=1 φ ρm(Li L−1j)mj,Lj (17) The scaled dot-product attention in the LLoCa framework is a special case of Eq.
Kai: (15), where the queries qLi are computed from latent features in the respective local frame and keys kLj are transformed by Li L−1j Symmetry Breaking Strategies Lorentz symmetry breaking is necessary when symmetry is realized only under a subgroup R of the full Lorentz group This can be done through two main approaches: 1.
Title and authors: Kai: Okay, so we're diving into the title and authors of this paper, "Lorentz-Equivariance without Limitations." The core idea here is that they are tackling the problem of achieving exact Lorentz equivariance for arbitrary neural networks while keeping the computational cost low.
Mira: I think the main point is that they introduce LLoCa, which is a framework designed to predict local reference frames for every particle and then propagate any order of tensorial information between them. It applies this to graph networks and transformers, showing it works well on amplitude regression, event generation, and jet tagging.
Lev: From my side in error correction terms, the part that interests me is the claim about minimal computational overhead because those frame conversions are usually where real hardware bottlenecks show up when you try to run these things.
Kai: That’s right. It’s not just about getting symmetry; it's about how they handle the transformation rules for these local frames and how they affect tensorial messages between particles, which is what makes the mechanism work.
Mira: And that tensorial message passing is quite powerful because it generalizes to message passing networks in a way that relates the received message back to what the sender's frame was by transforming it using rho m Li L minus one j mj,Lj.
Lev: If you take that generalization, the updated message f updated i,Li N squared j equals psi f i,Li N squared j plus phi rho m Li L minus one j mj,Lj and you see how it directly incorporates the frame transformation into the message update itself.
Kai: Precisely. And that leads to the output vector satisfying the correct transformation behavior under Lorentz transformations because of how everything is set up in equation nine.
Mira: So, they show that this setup ensures y mu lambda equals lambda mu nu L minus one a y a L, which is the correct transformation rule for the output vector when you transform it back to the global frame.
Lev: That’s crucial because without that correct transformation behavior, any symmetry you build up in the network architecture would be meaningless if it doesn't respect relativity.
Kai: So they’re showing that this structure is robust enough to handle arbitrary architectures, meaning you don't have to design a network specifically for this new method.
Mira: And they also address how to break the symmetry if you only care about a subgroup R of the full Lorentz group, which is where it gets more practical than just aiming for full symmetry all the time.
Lev: That’s where my experience comes in; it's much more useful when you have constraints on what your hardware can actually run, so knowing how to break equivariance down to a residual SO(two) group might be a practical way forward <ref:2508.14898#pg1>.
Kai: So they are not just proving the full symmetry thing, but providing tools for practical applications where you need to manage the complexity of Lorentz transformations in a constrained setting.
Mira: And that’s exactly what they seem to be doing with this paper, providing both the exact solution and a strategy for managing those subgroups.
The paper's summary: Kai: Now let's talk about what the paper actually summarizes in "Lorentz-Equivariance without Limitations." It boils down to LLoCa being a framework that uses learnable reference frames to build Lorentz-equivariant adaptations of any network architecture.
Mira: Essentially, it predicts these local frames for each particle and then propagates any order of tensorial information between those frames. This is applied to graph networks and transformers, and the results they showcase on amplitude regression, event generation, and jet tagging are very high quality.
Lev: I’d focus on the fact that they show that this method respects the structure of particle data by embedding these fundamental symmetries directly into the ML networks.
Kai: And they do this by using a learnable reference frame for each particle, which is what makes it local canonicalization, and then propagating those tensorial messages between those frames.
Mira: They are showing that this isn't just a theoretical curiosity; it works on concrete LHC problems like jet tagging where performance surpasses non-equivariant models.
Lev: And the key thing they point out is that they provide methods to handle the symmetry breaking, which means you can modify the network architecture or add specific inputs to control the residual symmetry group.
Kai: So, it’s not just a theoretical proof of full equivariance; it’s a practical method for embedding physics into machine learning models in a way that's flexible enough for real-world use.
Mira: The paper is about showing that you can embed fundamental symmetries in machine learning networks without sacrificing performance on the tasks we care about.
The paper's improvements: Lev: Regarding the improvements, they suggest two main ways to handle symmetry breaking, which we already talked about in detail earlier—architectural breaking and input-level breaking.
Kai: And I think one key improvement is that they allow for a fair comparison between different Lorentz-symmetry breaking approaches, which is something previous work hadn't really managed.
Mira: That's important because it lets users see how data augmentation emerges naturally as a byproduct of defining those reference frames, which you don't have to think about separately.
Lev: And for event generation, they say that having an architecture equivariant under the unbroken subgroup of the Lorentz group is sufficient for that task.
Kai: For jet tagging, they confirm that including symmetry breaking information at the input level is needed when you want optimal performance there.
Mira: So it’s not a one-size-fits-all solution; it depends entirely on whether you are doing event generation or jet tagging, right?
Conclusion: Kai: To wrap up, the main takeaway from "Lorentz-Equivariance without Limitations" is that LLoCa provides exact Lorentz equivariance at minimal computational cost for any network architecture.
Mira: It gives you a straightforward way to break or even remove it when symmetry is only realized under a subgroup of the Lorentz group, allowing for a fair comparison between exact equivariant networks and data augmentation.
Lev: For event generation, you just need an architecture equivariant under the unbroken subgroup of the Lorentz group for that task.
Kai: And for jet tagging, including symmetry breaking information at the input level seems necessary to get top performance there according to their findings when you look at the results.
Mira: So, as we conclude this discussion on "Lorentz-Equivariance without Limitations," it’s a framework that gives us exact Lorentz equivariance without adding massive computational costs for any network architecture.
Lev: It’s a framework that lets us compare standard and Lorentz-equivariant implementations fairly, and it enables users to easily modify the equivariance group.
Université catholique de Louvain Institute für Theoretische Physik Universität Heidelberg Interdisciplinary Center for Scientific Computing CERN
hep-ph, hep-ex
Submitted: 2025-08-20
Updated: 2026-11-04
Comments: 33 pages, 9 figures, 8 tables. v2: refined and additional results, added appendix B and C, 38 pages, 9 figures, 10 tables; v3: final version accepted by SciPost, 41 pages, 13 figures, 10 tables
Journal ref: SciPost Phys. Core 9, 058 (2026)
DOI: 10.21468/SciPostPhysCore.9.3.058
Code: https://github.com/heidelberg-hepml/lloca
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Lorentz Local Canonicalization (LLoCa) ensures exact Lorentz-equivariance for arbitrary neural networks with minimal computational overhead The paper introduces LLoCa…
Key concepts
- Local Reference Frames
- LLoCa predicts a specific local coordinate system for every input particle. This frame is constructed using four orthonormal unit vectors that define the local geometry of each particle's spacetime location. This allows the network to understand interactions relative to each particle's immediate environment.
- Lorentz Transformation Rule
- The paper defines how physical quantities, like vectors and tensors, must change when moving between different reference frames in special relativity. LLoCa uses this rule to ensure that the network's internal representations correctly transform when particles are moved or observed from different perspectives.
- Symmetry Breaking Strategies
- Sometimes a full Lorentz symmetry is too restrictive for the problem. LLoCa offers two ways to handle this: changing the network's architecture so it only respects a subgroup of the symmetry, or by adding specific input features that are only invariant under that smaller group.
Terminology
Summary
Lorentz Local Canonicalization (LLoCa) ensures exact Lorentz-equivariance for arbitrary neural networks with minimal computational overhead The paper introduces LLoCa, a framework that predicts local reference frames for each particle and propagates any-order tensorial information between them It is applied to graph networks and transformers, showcasing cutting-edge performance on amplitude regression, end-to-end event generation, and jet tagging
Lorentz Local Canonicalization Mechanism
LLoCa builds upon the concept of local canonicalization and extends it to Lorentz symmetry by predicting local reference frames for each input particle The construction of a local frame is achieved using four orthonormal unit vectors eaµ with the definition L = (3) and the transformation rule L'aµ = L aν (Λ−1) ν µ (5) Local frames constructed as in Eq. (4) satisfy the required transformation behaviour of Eq. p. 2 (5)
Local frames constructed as in Eq. (4) satisfy the required transformation behaviour of Eq.Where the paper says it
Equivariance and Message Passing
Lorentz-equivariance is achieved by transforming the input vector xµ into the local frame yielding the invariant object x aL = L aµ x µ (8) The output vector yµ is then transformed back to the global frame as yµ = (L−1) µa y aL (9) This leads to the output vector satisfying the correct transformation behavior yµ Λ→ y′µ = Λµν (L−1) ν a y aL = Λµν y ν For message passing networks, the tensorial message passing is generalized to f updated i,Li= ψ f i,Li N □ j=1 φ(mjj) (14) The received message mj,Li is related to the sender's frame by mj,Li = ρm(Li L−1j)mj,Lj (15) This generalization leads to the updated message f updated i,Li= ψ f i,Li N □ j=1 φ ρm(Li L−1j)mj,Lj (17) The scaled dot-product attention in the LLoCa framework is a special case of Eq. (15), where the queries qLi are computed from latent features in the respective local frame and keys kLj are transformed by Li L−1j
Symmetry Breaking Strategies
Lorentz symmetry breaking is necessary when symmetry is realized only under a subgroup R of the full Lorentz group This can be done through two main approaches:
-
Lorentz symmetry breaking through architecture: This involves changing the architecture such that Eq. p. 6
(13) only holds for Λ ∈ R, by fixing a subset of reference directions vk in the construction of local frames in Eq. p. 6Where the paper says it
The flexibility of the LLoCa framework allows for a fair comparison between different Lorentz-symmetry breaking approaches, which is difficult to achieve with previous Lorentz-equivariant architectures. Lorentz symmetry breaking through architecture First, we change the architecture such that Eq.
(21) Examples include setting v0 = (1, 0, 0, 0) for R = SO(3), or using auxiliary inputs like energy E or transverse momentum pT to break equivariance down to a residual SO(2) groupWhere the paper says it
(13) only holds for Λ ∈ R. While this is challenging to achieve in Lorentz-equivariant architectures with specialized layers [22,23,25], LLoCa allows to reduce the symmetry group by fixing a subset of reference directions vk in the construction of local frames in Eq.
-
Lorentz symmetry breaking through input: This breaks Lorentz-equivariance at the input level by adding extra inputs which are only invariant under R, such as reference vectors r, with similar assignments to the residual symmetry group R For jet tagging, scalar features like E or pT can be included to break equivariance down to SO(2)beam
Performance Across Applications
LLoCa has been tested on several LHC applications, including amplitude regression, event generation with high-multiplicity final states, and jet tagging In amplitude regression for Z + gluons, LLoCa networks consistently outperform their non-equivariant baselines across different architectures like MLP-I and L-GATr For event generation of high-multiplicity events, the conditional flow matching approach using LLoCa generators is used to learn the normalized phase space density directly In jet tagging on the JetClass dataset, LLoCa versions of established taggers like ParticleNet and ParT achieve superior performance compared to non-equivariant counterparts The LLoCa-Transformer achieves an AUC of 0.9882 on the TopTagXL dataset, demonstrating state-of-the-art results for top tagging
Computational Efficiency and Comparison
LLoCa enables efficient Lorentz-equivariance by reducing FLOPs by 10× and training time by 4× compared to the state-of-the-art L-GATr architecture while achieving the same performance The computational overhead of LLoCa is primarily due to frame-to-frame conversions in Sec. 2.3 and local frame construction in Sec. p. 9 The
computational overhead of LLoCa comes from the frame-to-frame conversions in Sec. 3 and
from the construction of local frames in Sec.Where the paper says it
Conclusion and Future Directions
LLoCa provides exact Lorentz-equivariance at minimal computational costs for any network architecture and offers a straightforward way to break and even remove Lorentz-equivariance The framework allows for a fair comparison between exact Lorentz-equivariant networks and data augmentation, where data augmentation emerges as a byproduct of the definition of reference frames For event generation, it is sufficient to have an architecture that is equivariant under the unbroken subgroup of the Lorentz group For jet tagging, including symmetry-breaking information at the input level is needed for optimal performance The LLoCa implementation is available at https://github.com/heidelberg-hepml/lloca, and all experiment code is available at https://github. p. 0 SciPost Physics Submission
Lorentz-Equivariance without Limitations
Luigi Favaro1
, Gerrit Gerhartz2
, Fred A. Hamprecht3
, Peter Lippmann3
,
Sebastian Pitz2
, Tilman Plehn2,3, Huilin Qu4
, and Jonas Spinner2
1 CP3, Université catholique de Louvain, Louvain-la-Neuve, Belgium
2 Institut für Theoretische Physik, Universität Heidelberg, Germany
3 Interdisciplinary Center for Scientific Computing (IWR), Universität Heidelberg, Germany
4 CERN, EP Department, Geneva, Switzerland
November 11, 2025
Abstract
Lorentz Local Canonicalization (LLoCa) ensures exact Lorentz-equivariance for arbitrary neural networks with minimal computational overhead.Where the paper says it
Training Specifications
The Frames-Net architecture uses a simple MLP ϕk acting on Lorentz-invariant features to predict vectors vµi,k To stabilize training, it is essential to constrain this network to predict only timelike vectors using Eq. (21) The timelike nature of the input particle four-momenta leads to the specific configuration of Eq. p. 6 Simpler choices such as exp and softplus yield a comparable but slightly worse
performance. Similarly, rescaling the particle four-momenta as (pi +pj
)/(∥pi +pj∥+ε) slightly
improves numerical stability compared to a simple factor of pj
.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Exact Lorentz equivariance for arbitrary neural networks can be achieved with minimal computational overhead using Lorentz Local Canonicalization (LLoCa). This allows systems to
predict local reference frames for each particle and propagate any-order tensorial information between them.
-
Systems can perform amplitude regression, end-to-end event generation, and jet tagging with state of the art performance by elevating a non-equivariant network to its Lorentz equivariant counterpart. LLoCa can reduce FLOPs by 10 times and training time by 4 times compared to state of the art architectures while achieving the same performance.
-
The framework allows for a straightforward way to break and even remove Lorentz equivariance, which is useful when symmetry is only realized under a subgroup of the Lorentz group. This flexibility enables
a fair comparison between different Lorentz-symmetry breaking approaches.
Abstract
Lorentz Local Canonicalization (LLoCa) ensures exact Lorentz-equivariance for arbitrary neural networks with minimal computational overhead. For the LHC, it equivariantly predicts local reference frames for each particle and propagates any-order tensorial information between them. We apply it to graph networks and transformers. We showcase its cutting-edge performance on amplitude regression, end-to-end event generation, and jet tagging. For jet tagging, we introduce a large top tagging dataset to benchmark LLoCa versions of a range of established benchmark architectures and highlight the importance of symmetry breaking.
Sources
- MadNIS -- Neural Multi-Channel Importance Sampling
- The Path to Proton Structure at One-Percent Accuracy
- AtlFast3: the next generation of fast simulation in ATLAS
- The Machine Learning Landscape of Top Taggers
- CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation
- Model-agnostic search for dijet resonances with anomalous jet substructure in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Search for new physics in final states with semi-visible jets or anomalous signatures using the ATLAS detector
- Classifying Anomalies THrough Outer Density Estimation (CATHODE)
- Residual ANODE
- Semi-visible jets, energy-based models, and self-supervision
- The Landscape of Unfolding with Machine Learning
- Generative Unfolding with Distribution Mapping
- OmniFold: A Method to Simultaneously Unfold All Observables
- How to Unfold Top Decays
- Jet Diffusion versus JetGPT -- Modern Networks for the LHC
- Extrapolating Jet Radiation with Autoregressive Transformers
- ParticleNet: Jet Tagging via Particle Clouds
- Particle Transformer for Jet Tagging
- The Physics Behind ML-based Quark-Gluon Taggers
- Deep-learned Top Tagging with a Lorentz Layer
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses