How does clustering signal events in parameter space differ from peak-hunting in single observable distributions?

# Clustering in Parameter Space vs. Peak-Hunting: A Key Distinction

The fundamental difference lies in dimensionality and information leverage.

Peak-Hunting: Single Observable

Traditional peak-hunting looks for a bump in one variable—say, the invariant mass distribution of reconstructed particles. You plot the histogram, fit a smooth background curve, and look for an excess above that curve. This is simple and interpretable, but it discards information: two events with the same invariant mass might look very different if you examined other properties (angles, energy ratios, etc.).

Cluster Scanning: Multi-Dimensional Signal Enrichment

By contrast, cluster scanning works in the full space of *multiple* observables simultaneously [AG-2024.02-1209]. Rather than looking for a bump in one histogram, the method uses algorithms like k-means to group events that are "close together" in this high-dimensional space. The idea is that signal events (e.g., from a new Z′ resonance) cluster together in ways that background events do not—not necessarily along a single axis, but as a cohesive pattern across many variables at once.

Once clusters are identified, you separate the anomalous cluster from the normal background, creating a signal-enriched region. *Then* you look for a peak in an observable like invariant mass—but now you're hunting within a curated subset of events rather than the full dataset.

Why it matters: Cluster Scanning reduced the signal sensitivity threshold by 39% (reaching 3σ significance with 61% fewer signal events) compared to traditional background-fitting approaches on a realistic LHC benchmark [AG-2024.02-1209]. The method is also fast enough to calibrate on synthetic data in real time.

Related Approaches

This principle extends beyond classical clustering. Quantum similarity learning [AG-2024.11-1249] and signal-aware latent spaces [AG-2026.03-1712] apply the same multi-dimensional logic: learn a representation where background and signal separate in a learned feature space, *then* search for anomalies. Even quantum clustering algorithms [AG-2024.06-2401] promise to scale this approach to very high-dimensional datasets where classical methods slow down.

The archive documents do not directly compare peak-hunting statistics to clustering-based search statistics in a single unified framework.