SO3C: Lorentz-covariant jet tagging with complexified-SO(3) geodesic flows

SO3C, a jet tagger built on the isomorphism so(3, C) =∼ so(3, 1) between the complexified rotation algebra and the Lorentz algebra. Each constituent is lifted, through the bivector it forms with the jet momentum, to a complex 3-vector on which a Lorentz transformation acts as a complex rotation. A message-passing round forms the pair invariants za · zb, builds a covariant connection from invariant-weighted neighbour states, and moves every state along it in closed form with a complex Rodrigues formula that conserves z · z; no ODE solver is involved. With the two beam directions as reference vectors, the output is covariant when jets and beams are transformed together and invariant under rotations aboutthe beam axis, which we check numerically in float64.On the Kasieczka et al. top-tagging reference dataset (canonical split, 64 leading con-stituents) the best configuration, with 67,874 parameters, reaches AUC 0.98504 ± 0.00009 and 1/εB = 1286 at εS = 0.3 over two seeds. Averaging seven runs of three configurations gives 0.98575 / 1553, and the two-run ensemble of a wider configuration gives the best re-jection, 1683. This is below PELICAN (0.9870 / 2250): the seven-run ensemble is 0.0012 AUC and a factor 1.45 in rejection behind it, and PELICAN’s own 11,000-parameter model (0.9858 / 1879) is ahead of every single SO3C run. We do not claim state-of-the-art perfor-mance.The largest gains come from the beam reference vectors and from ParT-style pair features written in Minkowski products (+0.0010 AUC; on a short protocol, ln ∆ and ln z alone carry them). Wider networks add +0.00046, a PELICAN-style rank-2 pair latent +0.00039, and their combination +0.00016 over width alone; of these three, only the width gain meets the pre-registered criterion. Code, per-jet scores, checkpoints and logs for every reported run are archived.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-12
DOI
https://doi.org/10.5281/zenodo.24165396
Primary Topic
Particle physics theoretical and experimental studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

SO3C: Lorentz-covariant jet tagging with complexified-SO(3) geodesic flows

Alexandr A. Panchenko
Zenodo (CERN European Organization for Nuclear Research)
Particle physics theoretical and experimental studies
preprint

SO3C: Lorentz-covariant jet tagging with complexified-SO(3) geodesic flows

Alexandr A. Panchenko
preprint en

Abstract

SO3C, a jet tagger built on the isomorphism so(3, C) =∼ so(3, 1) between the complexified rotation algebra and the Lorentz algebra. Each constituent is lifted, through the bivector it forms with the jet momentum, to a complex 3-vector on which a Lorentz transformation acts as a complex rotation. A message-passing round forms the pair invariants za · zb, builds a covariant connection from invariant-weighted neighbour states, and moves every state along it in closed form with a complex Rodrigues formula that conserves z · z; no ODE solver is involved. With the two beam directions as reference vectors, the output is covariant when jets and beams are transformed together and invariant under rotations aboutthe beam axis, which we check numerically in float64.On the Kasieczka et al. top-tagging reference dataset (canonical split, 64 leading con-stituents) the best configuration, with 67,874 parameters, reaches AUC 0.98504 ± 0.00009 and 1/εB = 1286 at εS = 0.3 over two seeds. Averaging seven runs of three configurations gives 0.98575 / 1553, and the two-run ensemble of a wider configuration gives the best re-jection, 1683. This is below PELICAN (0.9870 / 2250): the seven-run ensemble is 0.0012 AUC and a factor 1.45 in rejection behind it, and PELICAN’s own 11,000-parameter model (0.9858 / 1879) is ahead of every single SO3C run. We do not claim state-of-the-art perfor-mance.The largest gains come from the beam reference vectors and from ParT-style pair features written in Minkowski products (+0.0010 AUC; on a short protocol, ln ∆ and ln z alone carry them). Wider networks add +0.00046, a PELICAN-style rank-2 pair latent +0.00039, and their combination +0.00016 over width alone; of these three, only the width gain meets the pre-registered criterion. Code, per-jet scores, checkpoints and logs for every reported run are archived.

Zenodo (CERN European Organization for Nuclear Research)
ITMO University (RU)
Particle physics theoretical and experimental studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon