TL;DR
This paper introduces "Cluster Scanning," a novel, model-independent method for searching for new physics resonances at the LHC. By using k-means clustering on low-level jet observables, it identifies anomalous clusters to create a signal-enriched region, outperforming traditional background fitting methods and reducing the number of signal events needed for discovery.
What Problem Does This Paper Address?
The search for new physics beyond the Standard Model (SM) is a major goal of particle physics. One common approach is the "bump hunt," where researchers look for a peak in the distribution of a mass-like observable, indicating a new resonant particle. However, traditional bump hunts often rely on assumptions about the shape of the background distribution, which can introduce uncertainties and limit sensitivity, especially when dealing with complex datasets or when the signal is subtle. Furthermore, these methods can become model-specific, requiring optimization for particular signal types. This paper aims to address these limitations by developing a more data-driven, model-independent method that leverages more information from the events themselves.
Methods
The core of the proposed method, Cluster Scanning (CS), involves the following steps:
- Training Region Selection: A narrow window of the dijet invariant mass ($m_{jj}$) is chosen to train the clustering algorithm. This region is expected to have a low signal-to-background ratio.
- K-means Clustering: A mini-batch k-means clustering algorithm is applied to the jet images (low-level event observables) of events within the training region. This groups similar jet images into a predefined number of clusters (k=50 in this study).
- Cluster Spectra Generation: For each of the 16 $m_{jj}$ bins in the evaluation region, the distribution of events falling into each cluster is calculated, forming "cluster spectra."
- Per-Bin Standardization: These cluster spectra are standardized using outlier-robust estimators to account for variations and identify potentially anomalous clusters.
- Anomalous Cluster Selection: Clusters showing a significant deviation (more than 3 standard deviations) from the norm in the positive direction are identified as "signal-rich."
- Signal-Rich and Signal-Depleted Regions: The signal-rich clusters are combined to form a signal-enriched region, while the remaining clusters form a signal-depleted region.
- Background Estimation: The background estimate is constructed from the signal-depleted clusters, normalized to match the total number of events in the signal-rich region.
- Test Statistic and Significance: The standard maximum local significance (MLS) is used as the test statistic, comparing the signal-enriched and signal-depleted regions. Calibration is performed using pseudo-experiments.
The novelty lies in using k-means clustering on jet images to create data-driven bins (clusters) and then analyzing the distribution of these clusters across different $m_{jj}$ bins to identify signal excesses, rather than relying on pre-defined functional forms for background fitting.
Datasets
The study utilizes the LHCO R&D dataset, which consists of:
- 1 million Standard Model (QCD) dijet background events.
- 100,000 signal events from a simulated Z' resonance decaying into quark-antiquark pairs.
These events were generated using PYTHIA 8.219 and simulated with DELPHES 3.4.1. The dataset is publicly available and commonly used for benchmarking anomaly detection techniques.
Key Insights and Ideas
- Data-Driven Binning: Cluster Scanning replaces fixed functional forms for background modeling with a data-driven approach using k-means clustering. This reduces reliance on assumptions about the background shape.
- Leveraging Low-Level Information: By clustering jet images, CS utilizes detailed, low-level event information that is often overlooked in traditional bump hunts, leading to increased sensitivity.
- Overdensity vs. Outlier Detection: Unlike many anomaly detection methods that focus on outliers (events far from the norm), CS searches for "overdensities" – regions where signal events are concentrated within specific clusters.
- Computational Efficiency: The k-means algorithm is relatively fast, allowing for efficient ensembling and calibration on synthetic data, which is crucial for robust statistical analysis.
- Model Independence: The method is designed to be model-independent, meaning it doesn't require prior knowledge of the specific properties of a potential new particle.
Key Results
- Performance Improvement: Cluster Scanning outperforms traditional 3- and 4-parameter functional background fitting procedures in a pseudo-analysis on the LHCO R&D dataset.
- Reduced Signal Event Requirement: CS reduces the number of signal events needed to reach a 3σ significant excess by a factor of 0.61 in the realistic analysis scenario.
- Significance Gain: In the transition region (where benchmark algorithms achieve 1σ to 3σ significance), CS improves the result by a factor of 2 or more for the realistic case.
- Idealized Case: Even in an idealized scenario where the background is perfectly known, CS still shows an improvement factor of up to 1.5 over an idealized functional fit.
- Speed: The method's speed allows for efficient calibration on synthetic data, which is a significant advantage.
Limitations and Future Work
- Assumption of Cluster Mass Independence: While CS aims to be less dependent on $m_{jj}$ assumptions than traditional methods, the paper acknowledges that cluster centroids are not perfectly independent of $m_{jj}$. This can lead to false positives, especially in the presence of signal.
- Training Region Impact: The performance of CS can be influenced by the choice of the training region. While the current study uses a region with minimal signal, training in a signal-rich region can further boost performance, though this is not feasible in a real analysis where the signal location is unknown.
- Future Work: The authors suggest exploring other clustering algorithms beyond k-means, developing alternative methods for selecting anomalous clusters, and unifying CS assumptions with those of other methods. They also note that further improvements could be made by refining the background estimation and cluster selection criteria.
Summary updated