Positive definiteness (PDness) is an essential property of covariance matrix estimators, yet many high-dimensional regularized estimators fail to maintain it due to their elementwise construction and lack of eigenvalue control. In this paper, we prop...
Positive definiteness (PDness) is an essential property of covariance matrix estimators, yet many high-dimensional regularized estimators fail to maintain it due to their elementwise construction and lack of eigenvalue control. In this paper, we propose LPD (Linear Positive Definitization), a unified linear-shrinkage framework for modifying any symmetric non–PD matrix into a positive definite one. Extending the line of work of (Choi et al., 2019; Cho et al., 2021; Park et al., 2024), we derive explicit optimal solutions for LPD under a broad class of matrix norms, including the spectral, scaled Frobenius, infinity, and element-wise maximum norms.
We further present a geometric interpretation of LPD and demonstrate, through a generalized Marshall’s law perspective, that LPD-based covariance estimators exhibit an error contraction property relative to the initial estimator across all considered norms. When applied to sparse covariance estimation, the LPD modification preserves the convergence rate of the initial estimator while ensuring positive definiteness. Extensive simulation studies validate these theoretical findings.
Finally, we introduce a new application of LPD to estimating the optimal projection direction in high-dimensional two-sample mean testing and show that incorporating LPD into existing procedures yields substantial improvements in recovering the true optimal direction.