Representative observations are useful to express explicitly the distributional variation of the data by a few selected observations corresponding to the quantiles in the univariate situation. Jones and Rice(1992) extended it to the multidimensional case by the principal component based method. This study introduces a modified version of Jones and Rice exploiting the self-consistency of principal components in expressing the chosen observation vectors. Compared to that of Jones and Rice, the suggested method tends to provide with less susceptible representative observations to the sampling variation of the data and the resulted vectors benefits from the self-consistency.