Recently, the concept of targeting weather observations has been tested by idealized model experiments and by deploying dropsonde observations from aircraft in Northern Hemisphere winter field programs. While the evaluation studies have so far provided convincing empirical evidence that targeting observations is an operationally attainable way to improve forecasts, the exact dynamical mechanisms through which targeted observations realize their beneficial forecast effects have not yet been explored.

The main differences between the targeting strategies proposed by the different teams are in the algorithms used to select to locations of the added observations. Nevertheless, there is one common element of these techniques; they are all linear inferences applied to a set of numerical forecasts. The main goal of this paper is to investigate why methods based on such a strong assumption can have skill in determining the optimal locations of the added observations. It will be argued that regions of local low dimensionality of the unstable subspace in the atmosphere (Patil et al., 2001) play an important role in the success of targeting.

The above goal is achieved by first generating a large experimental ensemble of forecasts using the T62 horizontal resolution version of the operational global model of the National Centers for Environmental Prediction (NCEP). A one month case study associated with the 2000 Winter Reconnaissance Program (Szunyogh et al. 2001) is prepared. Then the regions of local low dimensionality are determined. The relationship between locations of these regions, the local energetics of the baroclinic wave packets, and the propagation of the influence of the added observations is explored.