Further results on discrimination with autocorrelated observations

摘要

This paper considers the estimation of the actual error rates of Fisher's linear discriminant function formed from training data which are serially correlated. The asymptotic bias of the so-called plug-in error rate is given for multivariate training data following a first order autoregressive model. Also, the authors' earlier work on the performance of Fisher's linear discriminant function relative to its version using maximum likelihood estimates obtained under the appropriate correlation model, is reconsidered. A correction is noted to their results on the differences in the individual error rates of these two linear discriminant functions.