ON THE LEAST-SQUARES FITTING OF CORRELATED DATA - REMOVING THE CORRELATION

In the most general case of linear least-squares fitting, the input data may be mutually correlated. This correlation is accommodated through a weight matrix W which contains off-diagonal elements. The existence of strong correlation in the data makes it difficult to interpret the fit residuals in terms of ''good'' and ''bad'' data. However, it is possible to transform the original correlated problem into an equivalent uncorrelated, unweighted fit through a simple linear transformation. This transformation is readily obtained from the orthogonal transformation which diagonalizes W. (C) 1994 Academic Press, Inc.