In theory, an initial-time norm based on the analysis-error covariances (AEC) should be used to compute singular vectors (SV) that are associated with the largest forecast errors. However, due to the absence of complete AEC matrices, the total-energy norm is the most commonly used in practice. In this study, we measure the impact of these two types of norms on the properties of SVs generated using the Canadian Global Environmental Multiscale (GEM) model.

To construct the AEC norm, we use flow-dependent analysis error covariances estimated from the output of a quasi-operational configuration of the ensemble Kalman filter. Different approaches for modelling the correlations of analysis error are used, including the spatially localized sample correlations. The final-time norm, defined as the total energy over North America, is the same in all experiments. The influence of moist physical parametrizations and the choice of optimization time interval are also considered. SVs are computed every day from 1 to 14 December, 2003. The comparison of results is based on diagnostics of dynamical and physical properties of the SVs, such as energy spectra, spatial distribution and potential vorticity structure. We also compare the fraction of the forecast error energy that can be explained by each set of SVs at various lead times.