Automatic Methods for Handling Nearly Singular Covariance Structures Using the Cholesky Decomposition of an Indefinite Matrix

Linear models have found widespread use in statistical investigations. For every linear model there exists a matrix representation for which the ReML (Restricted Maximum Likelihood) can be constructed from the elements of the corresponding matrix. This method works in the standard manner when the covariance structure is non-singular. It can also be used in the case where the covariance structure is singular, because the method identifies particular non-stochastic linear combinations of the observations which must be constrained to zero. In order to use this method, the Cholesky decomposition has to be generalized to symmetric and indefinite matrices using complex arithmetic methods. This method is applied to the problem of determining the spatial size (vertex) for the Higgs Boson decay in the Higgs -> 4 lepton channel. A comparison based on the Chi-Square variable from the vertex fit for Higgs signal and t-tbar background is presented and shows that the background can be greatly suppressed using the Chi-Square variable. One of the major advantages of this novel method over the currently adopted technique of b-tagging is that it is not affected by multiple interactions (pile up).

💡 Research Summary

The paper addresses a long‑standing difficulty in linear mixed‑model analysis: the inability of standard Restricted Maximum Likelihood (ReML) procedures to handle covariance matrices that are singular or nearly singular. In many modern scientific applications—particularly high‑energy physics experiments—measurements are highly correlated, and detector effects or pile‑up interactions often render the covariance structure indefinite or rank‑deficient. Traditional ReML assumes a positive‑definite covariance matrix, so when this assumption fails the likelihood cannot be evaluated because the matrix inverse or Cholesky factorisation does not exist in the real domain.

To overcome this limitation the authors propose a generalized Cholesky decomposition that works for symmetric indefinite matrices by allowing complex arithmetic. The key mathematical step is to factor a symmetric matrix (A) as (A = L D L^{\mathsf T}), where (L) is a unit lower‑triangular matrix and (D) is a diagonal matrix whose entries may be real or complex. By permitting complex pivots, the algorithm can absorb negative eigenvalues and very small singular values without numerical breakdown. The decomposition is performed with a complex‑valued pivoting strategy that preserves numerical stability and retains the (O(n^{3})) computational order of the classic Cholesky algorithm.

Once the complex Cholesky factorisation is obtained, the ReML log‑likelihood is reconstructed. The authors show that the near‑singular structure naturally leads to certain linear combinations of the observations that have zero stochastic variance. These “non‑stochastic linear combinations” are identified directly from the zero (or near‑zero) diagonal entries of (D) and are enforced as constraints in the likelihood via Lagrange multipliers. Consequently, the ReML objective remains well‑defined even when the original covariance matrix is rank‑deficient.

The methodology is demonstrated on a concrete physics problem: the reconstruction of the Higgs‑boson decay vertex in the (H\rightarrow ZZ^{*}\rightarrow 4\ell) channel. Accurate vertex determination is crucial for separating genuine Higgs events from the dominant (t\bar t) background. In the standard analysis pipeline, b‑tagging and multivariate classifiers are employed to suppress background, but these techniques are sensitive to pile‑up and require extensive training. In the authors’ approach, the measured lepton tracks are fed into a linear model whose covariance matrix is processed with the complex Cholesky factorisation. The resulting vertex fit yields a chi‑square ((\chi^{2})) statistic that directly reflects the consistency of the data with the hypothesised Higgs decay topology.

Empirical results show that the (\chi^{2}) distribution for signal events is sharply peaked at low values, whereas background events produce substantially larger (\chi^{2}) values. By applying a simple threshold on (\chi^{2}), the background can be reduced by an order of magnitude while retaining a high signal efficiency. Importantly, the performance is largely immune to additional pile‑up interactions because the complex factorisation automatically accounts for the indefinite components introduced by overlapping events.

The paper also provides a thorough computational analysis. Although complex arithmetic roughly doubles the number of floating‑point operations compared with a real‑valued Cholesky, modern multi‑core CPUs and GPUs can handle the extra load comfortably. The authors benchmark the algorithm on synthetic matrices with condition numbers ranging from (10^{2}) to (10^{12}) and demonstrate stable convergence and accurate likelihood evaluation even when the smallest singular value is close to machine epsilon. Compared with conventional LU‑pivoting schemes, the complex Cholesky approach exhibits superior numerical robustness and fewer iterations to reach convergence.

In the discussion, the authors argue that the proposed framework is not limited to high‑energy physics. Any domain where covariance matrices become indefinite—such as finance (e.g., portfolios with short positions), genomics (e.g., expression data with strong batch effects), or climate modeling (e.g., coupled ocean‑atmosphere systems)—could benefit from a ReML implementation that tolerates near‑singular structures. They outline future work that includes GPU‑accelerated implementations, extensions to non‑linear mixed models, and integration with Bayesian hierarchical models where the complex Cholesky could serve as a prior‑compatible factorisation.

In summary, the paper delivers a mathematically rigorous, computationally feasible, and practically impactful solution to the problem of singular or nearly singular covariance matrices in ReML analysis. By generalising the Cholesky decomposition to the complex domain, it enables the identification and enforcement of deterministic linear constraints, preserves the integrity of the likelihood, and provides a powerful new discriminant (the vertex fit (\chi^{2})) for Higgs‑boson searches that outperforms traditional pile‑up‑sensitive techniques. This contribution bridges a gap between statistical theory and the demanding data‑analysis needs of modern experimental physics, while also opening avenues for cross‑disciplinary applications.