Abstract
This article summarizes the recent erratum for the ClusterKinG tool, a software package designed to cluster kinematic distributions in $B$ meson decays. The update addresses critical statistical oversights regarding degrees of freedom and bin correlations, ensuring more precise model differentiation in high-energy physics parameter spaces.
Introduction
In the search for physics beyond the Standard Model, researchers frequently utilize tools like ClusterKinG to manage vast theoretical parameter spaces. By grouping similar kinematic distributions (histograms) into clusters, physicists can identify distinct regions of parameter space that yield unique experimental signatures. However, a recent update (v1.1.0) was necessitated by the discovery of two statistical inaccuracies in the original distance metric used for clustering.
The Statistical Refinement
The core of the erratum involves the "distance" calculation between two histograms. The original implementation suffered from two primary issues:
- Degrees of Freedom ($r$): For a histogram with $N$ bins, the original tool assumed the $\chi^2$ distance followed a distribution with $N$ degrees of freedom. Because the histograms are normalized, one degree of freedom is inherently lost. The correct distribution follows $r = N - 1$.
- Correlated Data: The initial metric assumed bin independence. In practice, experimental and theoretical uncertainties often correlate bin contents. The updated metric now incorporates the inverse covariance matrix ($\Sigma^{-1}$) to account for these relationships.
As shown in Figure 1, the cutoff value $c$ used to define the stopping criterion for clustering is highly sensitive to the degrees of freedom $r$, especially at lower values typical of kinematic histograms.
Validation via Toy Experiments
To validate the corrected metric, the authors performed extensive "toy" experiments. They generated thousands of synthetic histograms based on the expected means and covariance matrices of $B$ meson decay variables, such as $\cos \theta_\tau$, $q^2$, and lepton energy $E_\ell$.
The results demonstrated that the empirical distance values from these toys align precisely with the theoretical $\chi^2_{N-1}$ distribution.
Impact and Key Results
The authors report that while these corrections are mathematically significant, their impact on previously published clustering results is minimal. The primary change is in the stopping criterion—the point at which the algorithm decides two clusters are sufficiently different to remain separate.
For instance, in a 10-bin example ($N=10$, so $r=9$), a normalized distance of $\chi^2/r = 1$ corresponds to a $p$-value of 34% with the correct $r=9$, compared to 32% with the incorrect $r=10$. This shift ensures that the "indistinguishability" of models is assessed with proper statistical rigor.
Conclusion
The release of ClusterKinG v1.1.0 provides a more robust framework for the high-energy physics community. By correctly accounting for normalization constraints and bin correlations, the tool now offers a mathematically sound basis for exploring the complex landscapes of $B$ meson decays and potential new physics.
Summary updated