If the measures at each of the N datapoints are uncorrelated, we can sum over the individual B values (obtained from (2)):
We get rid of any correlations by orthogonalising the covariance matrix by singular value decomposition. The residuals are calculated by projecting the correlated residuals along the axis of the eigenvectors of and the eigenvalues of give the probable errors at each point.
We used the script model_complexity_estimate.m to generate the results in figure 6.