Performance loss of multivariate detection algorithms due to covariance estimation
Charles E. Davidson, Avishai Ben‐David · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
Performance of the matched filter and anomaly detection algorithms relies on the quality of the inverse sample covariance matrix, which depends on sample size (number of vectors). The "RMB rule" provides the number of vectors required to achieve a specific average performance loss of the matched filter. In this paper we extend the RMB rule to provide the number of vectors needed to ensure a minimum performance loss (within a certain confidence). We also review a general metric for covariance estimation accuracy based on the Wishart distribution and discuss anomaly detector performance loss.