In the search for physics beyond the Standard Model (BSM), researchers often encounter the "curse of dimensionality." New Physics (NP) models, such as Effective Field Theories (EFTs), introduce numerous unknown parameters called Wilson coefficients. These coefficients alter the kinematic distributions of particle decays, but simulating every possible parameter combination to test against experimental data is computationally prohibitive.
A recent study published in JHEP introduces ClusterKinG, a Python-based framework designed to automate the identification of representative "benchmark points" (BPs) within these high-dimensional spaces. By clustering similar theoretical predictions, ClusterKinG simplifies the interface between complex theory and experimental analysis.
Methodology: From Parameter Space to Clusters
The core philosophy of ClusterKinG is to group parameter points that produce statistically indistinguishable kinematic distributions. The workflow follows a structured pipeline:
- Sampling and Generation: The tool samples the parameter space (Wilson coefficients) and generates kinematic distributions (histograms) using the
flaviophysics library. - Metric Definition: To compare distributions, the authors employ a $\chi^2$ metric that accounts for experimental realities, including Poisson statistical noise and systematic uncertainties.
- Hierarchical Clustering: Using agglomerative clustering, the algorithm merges points based on their similarity.
- Stopping Criterion: Merging stops when the distance between clusters exceeds a threshold (typically $\chi^2/N > 1$). This ensures that all points within a cluster are indistinguishable given the assumed experimental precision.
- Benchmark Selection: A single representative point is chosen for each cluster, minimizing the distance to all other members.
Application to B-Meson Anomalies
The authors demonstrated the utility of ClusterKinG using semileptonic $B$ decays, specifically $\bar{B} \to D^{(*)}\tau^-\bar{\nu}_\tau$. These channels are of significant interest due to long-standing anomalies observed at LHCb and Belle.
Key Insights
- Resolution-Driven Clustering: The number of clusters is not an arbitrary hyperparameter but a function of experimental resolution. As signal yields increase (e.g., moving from Belle to Belle II), the "resolution" improves, and the algorithm naturally identifies more distinct clusters.
- Sensitivity Mapping: The tool helps identify which parameters most significantly impact the "shape" of the data. For instance, the tensor coefficient ($C_T$) was found to be a dominant driver of cluster formation in the studied decays.
- Dimensionality Reduction: In a test case involving the $\cos\theta_\tau$ distribution in $D^*$ decays, a space of 1,000 parameter points was reduced to just six benchmark points, assuming 700 signal events and 1% systematic error.
Validating and Extending Previous Research
The study re-evaluated a classic BaBar analysis of the Two-Higgs-Doublet Model (2HDM). While the original analysis manually identified two distinct regions, ClusterKinG provided a rigorous, automated confirmation and revealed that the transition region between these models is more nuanced than previously thought.
Furthermore, the authors showed that the number of required benchmark points grows logarithmically with signal yield. This provides a clear roadmap for how experimental sensitivity will evolve in future high-luminosity environments.
Limitations and Future Directions
While powerful, the current implementation of ClusterKinG has areas for growth:
- Sampling Efficiency: Moving from uniform grid sampling to adaptive sampling will improve performance in high-dimensional spaces where distributions change rapidly.
- Multi-dimensional Distributions: Future versions aim to handle 2D distributions (e.g., $q^2$ vs. angular variables) to capture more correlation data.
- Complex Coefficients: Expanding the framework to include complex Wilson coefficients will allow for the study of CP-violating effects.
ClusterKinG is available as an open-source package on GitHub, offering a bridge for theorists and experimentalists to collaborate more effectively on BSM searches.
Summary updated