Information-Theoretic Bounds to Accuracy of Object Classification in Representation Spaces with Given Distances

Lower bounds to error probability of object classification with fixed amounts of processed information in spaces of object representations with given distances are considered. The bounds are specified by strictly decreasing functions of minimum average mutual information between the objects and estimates of their classes of error probability. The inverse functions give the lower bounds to error probability for the classification with fixed amounts of processed information. Numerical implementations of these bounds have been obtained for tree-structured and vector-based object representations. The lower bound of error probability in the space of vector-based object representations has been shown to be smaller as compared to the similar bound in the space of tree-structured representations. The bound of error probability can be lowered by fusion of object representations with various metrics.