A modification on Mahalanobis distance on samples of limited size by compensation for errors of non-dominant eigenvalues and asymmetrical distribution on dominant principle components is proposed.By the introduction of modified Mahalanobis distance

one can compute efficiently the nearest neighbor in transformed space

with compensation for errors of non-dominant eigenvalues and more accurate characterization of the projection distribution of feature vector on every dominant principal component.Experimental results on UCI dataset for handwritten digit recognition indicate that modified algorithm is effective to improve the recognition performance.