A new method is developed for the state estimation of linear discrete-time stochastic systems in the presence of an unknown disturbance. The filter obtained is optimal in the unbiased minimum variance sense. The necessary and sufficient conditions for the existence and the stability of the filter are given.