Which principal components are most sensitive to distributional changes?

Martin Tveten

PCA is often used in anomaly detection and statistical process control tasks. For bivariate data, we prove that the minor projection (the least varying projection) of the PCA-rotat…