Weakly supervised anomaly detection with event-level variables

TL;DR

The paper extends weakly supervised anomaly detection to a di-object + X resonance topology, framing background learning via sideband interpolation with the CATHODE method and classifying SR events using event-level phase-space features. By encoding multi-particle kinematics with simplex phase-space coordinates and supplementing with global event observables, the approach achieves robust background modeling and enhanced sensitivity. Across di-$\mu$ and di-$\tau$ final states, the method yields discovery-level significance for moderate signal strengths ($\leq 2\sigma$ injections) that would be missed by traditional bump-hunt strategies, demonstrating the potential of anomaly detection to broaden LHC search reach. The work also highlights avenues for improvement, including full phase-space representations (hypersphere coordinates) and non-Euclidean normalizing flows, to further boost sensitivity in resonance + X searches.

Abstract

We introduce a new topology for weakly supervised anomaly detection searches, di-object plus~X. In this topology, one looks for a resonance decaying to two standard model particles produced in association with other anomalous event activity (X). This additional activity is used for classification. We demonstrate how anomaly detection techniques which have been developed for di-jet searches focusing on jet substructure anomalies can be applied to event-level anomaly detection in this topology. To robustly capture event-level features of multi-particle kinematics, we employ new physically motivated variables derived from the geometric structure of a collision's phase space manifold. As a proof of concept, we explore the application of this approach to several benchmark signals in the di-$τ$ and di-$μ$ plus~X final states. We demonstrate that our anomaly detection approach can reach discovery-level significances for signals that would be missed in a conventional bump-hunt approach.

Figures (8)

• Figure 1: Feynman diagrams of the $t\bar{t}\phi$ (left), $T'\bar{T'}$ (middle), and $X \to YH$ (right) signal models. The $\ell$ symbol denotes the $\tau$ or $\mu$ leptons. The $S^0$ decays inclusively according to its branching ratios. The $H$ boson in the $X \to YH$ model decays invisibly.
• Figure 2: Invariant visible mass distribution of the di-$\tau$ system (left) and the di-$\mu$ system (right)
• Figure 3: A comparison of the distributions of true background events in the signal region and the synthetic ones generated by our normalizing flow for di-$\mu$. Simplex coordinate 10 is omitted for brevity, but has similar distributions.The last bin in each histogram shows overflow entries beyond the range of the histogram axis. The flow is seen to model most of the distributions well. Some minor discrepancies are seen in the modeling close to variable boundaries at zero. The feature distributions and agreement between synthetic and true are similar in the di-$\tau$ channel.
• Figure 4: Significance improvement curves showing classification performance on across our different signal models for both di-$\mu$(left) and di-$\tau$(right) for signal injections of 1$\sigma$ (top), and 1.5$\sigma$ (bottom). The shaded bands represent the statistical uncertainty (±1 standard deviation) in the measured average significance improvement from the 5 different runs.
• Figure 5: A summary of the sensitivity enhancement of our anomaly detection approach for both di-$\mu$ (left) and di-$\tau$ (right). The x-axis shows the size of the signal, quantified as $\frac{S}{\sqrt{B}}$ prior to any anomaly detection selection. The y-axis shows the approximate signal significance achieved after the application of an anomaly score cut that is 1% efficient on the background. Our procedure is able to enhance the signal significance above the discovery threshold ($5 \sigma$) for moderately sized signal injections $\leq 2 \sigma$ for all considered signal models.