Undesirable effects of covariance matrix techniques for error analysis

Abstract

Regression with ${\chi }^2$ constructed from the covariance matrix should not be used for some combinations of covariance matrices and fitting functions. Using the technique for unsuitable combinations can amplify systematic errors. This amplification is uncontrolled, and can produce arbitrarily inaccurate results that might not be ruled out by a ${\chi }^2$ test. In addition, this technique can give incorrect (artificially small) errors for fit parameters. I give a test for this instability and a more robust (but computationally more intensive) method for fitting correlated data.

Comments: CERN-TH.6892/93 (revised), replaced because of major improvements (finding a test for stability of covariance matrix regression and an alternative method for fitting correlated data)