Kalman Filtering with Partially Diffuse Initial Conditions

In this paper a square root algorithm is proposed for estimating linear state space models. A particular feature of the approach is that it contains special provisions for nonstationary time series with incompletely specified initial conditions. It differs from earlier approaches to the problem in that an additional property of the covariance matrix of the state estimation error vector is exploited to further reduce storage requirements and computatational loads in computer implementations.