We describe a novel approach to evaluate the performance of global atmospheric ensemble prediction systems. This approach is based on the observation that beyond a 2-3-day forecast time (i) the errors in the extratropics in a numerical weather prediction are dominated by synoptic scale structures (ii) the spatial error patterns in a 1000-km radius local neighborhood of location l can be described by a local linear space, S(l), and (iii) ensemble prediction systems usually provide a good representation of S(l). The performance of an ensemble prediction system that satisfies (iii) can be assessed by investigating the efficiency of the ensemble in representing the errors in S(l). This approach is applied to the operational ensembles included in the THORPEX Interactive Grand Global Ensemble (TIGGE) data set. We show that the different ensemble systems, which use a variety of techniques for the simulation of the effects of the initial condition uncertainties and the random component of the forecast errors due to model deficiency, satisfy different optimality conditions when evaluated by our approach.