ABSTRACT
Biometric identification of persons depends on variation among persons. The more complex and random the variation, the better, because this increases uniqueness and enables greater discriminating power. However, although maximal between-person variability is desirable, it is also desirable to have minimal within-person variability in the chosen biometric features, over time and across changing conditions of capture. If the first variability (between-person) is too small, the likelihood of false matches between different persons is increased. If the second variability (within-person) is too large, then the likelihood of false nonmatches is increased. In a sense, the entire mathematical and statistical science of pattern recognition can be reduced to questions about the relationship between these two variabilities. Classification is only reliable if the diameters of the clusters of data corresponding to the different classes are smaller than the distances between the clusters.
