For both datasets of biometric objects given by images and an ensemble of the different modality datasets, the lower bounds on the error probability of person identification subject to a fixed amount of information have been investigated. The bounds are constructed using a probabilistic object classification model in metric spaces of the object representations. These bounds are independent of decision algorithms and are given by the inverses of the rate-distortion functions for discrete source coding with Hamming distortion, when the source letters are transmitted over a noisy channel. The difference between unity and any obtained lower bound produces an appropriate upper bound on the accuracy of person identification, depending on a given amount of processed information in a dataset of object representations. The obtained bounds are useful for estimating the efficiency of decision algorithms in terms of deviations of the algorithm error probability or accuracy relative to the boundary values, subject to a given average amount of information used for decision-making.
Lange et al. (Mon,) studied this question.