Summary: This paper presents a new strategy for the design of the reduced-order Kalman filter in discrete-time stochastic linear systems. The problem is to estimate a part of the state vector in the case where none of the observations is assumed to be noise free. The proposed filter is obtained by minimizing the trace of the estimation error covariance matrix with respect to the remaining design of freedom after non-interesting state decoupling. The necessary and sufficient conditions for stability and convergence of the filter are established.