Hadronic radiation mapped onto a sphere: the inside-outside asymmetry around a colour-singlet Higgs candidate tightens expected top-quark rare-decay limits by about an order of magnitude
Synopsis
At the Large Hadron Collider, flavour-changing neutral top-Higgs signals share reconstructed objects with their backgrounds, and analyses have leaned on invariant masses and angular kinematics. This work maps hadronic constituents onto the unit sphere with a modified inverse stereographic projection centred on the midpoint of the two b-tagged jets forming the Higgs candidate, so radiation inside and outside the cone lands in opposite hemispheres, and a graph neural network built with geodesic distances reads that asymmetry. In simulation at 14 TeV with an assumed 3000 fb⁻¹, expected 95% confidence-level limits improve by about an order of magnitude over a baseline graph network built in the plane.
Interpretation
Rare-process searches begin to treat the colour structure of soft quantum chromodynamics radiation as usable information, not only the reconstructed jet and lepton kinematics. Signal and dominant background contain the same reconstructed objects, such as several jets and several b-tagged jets, so invariant masses and angular separations alone separate them poorly. On a simulated benchmark of flavour-changing neutral top-Higgs interactions, two graph neural networks share the same kinematic branch and classification head, and the spherical representation gives the stronger expected exclusion.
Each event's hadronic constituents are placed on a sphere centred on the Higgs candidate, so inside and outside the cone become the southern and northern hemispheres. The mapping uses a modified inverse stereographic projection whose centre is the midpoint of the two b-tagged jets in the η-φ plane and whose reference radius is half their separation. Graph neighbours and edge attributes are defined with the geodesic distance on the sphere, the shorter great-circle arc between two points rather than the straight chord through the sphere.
At 14 TeV with an assumed 3000 fb⁻¹ and a per-jet b-tagging efficiency of 70%, the main analysis gives expected limits on the top-Higgs flavour-changing neutral couplings. Under the same conditions, the baseline graph network built in the plane gives expected limits weaker by about an order of magnitude. The main result comes from a profile-likelihood analysis with a systematic-uncertainty model and agrees within 10% with an exact Poisson treatment.
Perspective
The demonstrated setting is simulated 14 TeV proton-proton collisions with fast detector simulation, leading-order background generation and a per-jet b-tagging efficiency of 70%, reported as expected limits; the same representation extends to other colour-singlet resonances decaying to a quark-antiquark pair and to colour-singlet decays into two gluons.
The most sensitive working points sit at false positive rates between 10⁻³ and 10⁻⁴, where the background estimate rests on one or two simulated events; uncertainties from parton-shower and hadronisation modelling, including colour reconnection, are not yet included in the assessment of classifier acceptance.
