On an Approximation of Mutual Information

An approximation of mutual information which is often used for feature evaluation in pattern recognition is shown to be more closely related to Bayes error probability than mutual information is when equal a priori probabilities are assumed.