Abstract
An analysis of variational data assimilation schemes for linear dynamical forecast models shows that the penalty functional must include an explicit contribution from the initial conditions in order to ensure a unique, low-noise forecast. The noise level is related to the effective number of data being assimilated.