Minimizing Misclassifications in Linear Discriminant Analysis*