A method is sought to decompose errors in numerical forecasts of the atmosphere into components that are uncorrelated. This can simplify the representation of the probability density function of forecast errors so that it can be used in data assimilation. A new method based on potential vorticity (PV) is used to partition errors into balanced and unbalanced variables that are thought to be mutually uncorrelated. The effectiveness of the PV method is compared with a simpler method. A toy model and an operational forecasting model are used to show that the PV-based variables are usually less correlated than those of the simpler approach. Copyright © 2007 John Wiley & Sons, Ltd.