An evaluation distance function is one of the most important factors that influences the accuracy of a handwritten character recognition system. Various evaluation distance functions have been proposed and investigated theoretically. City block distance, Euclidean distance, weighted Euclidean distance, sub-space method, multiple similarity method, Bayes decision method and Mahalanobis distance are known typical distance functions. Although observing the performance of each evaluation function in a large-scale handwritten character recognition system is quite important, there has been little research reported on this topic. In this paper, because the emphasis is on how to improve the accuracy of a recognition system, comparison experiments are carried out. The experimental results show that the Mahalanobis distance is the most effective of the seven typical evaluation distance functions. Considering the foregoing result and the properties of distribution on each axis, a modified Mahalanobis distance is proposed to construct a more accurate and faster system. Using the proposed modified Mahalanobis distance as the evaluation function, a recognition rate of 98.24 percent has been achieved for ETL9B, the largest public database of handwritten characters in Japan. In this paper, the behavior of seven typical evaluation functions is studied and a new evaluation distance, called the modified Mahalanobis distance, is proposed based on these results. © 1997 Scripta Technica, Inc. Syst Comp Jpn, 28 (1): 46–55, 1997