Abstract: In this paper, we explore the varing range of the error distributions within which the LSE is the best linear estimator in the sense of minimizing the MSE matrix. We also give the maximal class of distributions of error term for which the LSE possesses a certain kind of robustness.