Classifier ensembles have been shown to outperform single classifier systems. An apparent necessary condition for ensembles to out?perform single systems is that the classifier systems exhibit a reasonable degree of "diversity". It has also been demonstrated that diversity is an important predictive factor for the improvement. However, in lack of a universally accepted definition, various diversity measures have been proposed and applied in the literature. A natural question then follows: How can we compare, and hence choose among, various diversity mea?sures? This work exploits analytically the relationships among several well-accepted diversity measures. These different diversity measures are proved to be closely related, which facilitates further research on classi?fier ensembles since the effective number of diversity measures is reduced by such close relationships.