G-estimation of the regular Mahalanobis distance under the condition that the dimension of the sample space is comparable with the size of the learning samples is analysed. It is proved for the first time that properties of consistency and asymptotic normality of this estimation hold, if the distribution of observations is not normal.