The paper gives a new optimal property of the classical method of multi-dimensional scaling when the distance matrix is non-Euclidean. We also examine robustness of the method under a linear model. A technique to estimate missing values is also given.