On the linear combination of observations and the general theory of least squares