On the fallacy of the method commonly employed in finding the probable error of a function of two or more quantities whose adjusted values have been derived from the same least-square solution