Title:Improved Discrete-Time Kalman Filtering within Singular Value Decomposition

Abstract:This paper is concerned with the hidden dynamic state estimation in linear discrete-time stochastic systems in presence of Gaussian noises. The associated recursive estimator is known as the discrete-time Kalman filter. To improve a numerical stability of the estimator with respect to roundoff errors, a variety of implementation methods have been derived. The most popular techniques belong to the square-root or UD factorization-based approaches and imply the Cholesky decomposition of the corresponding error covariance matrix. Another important matrix factorization method is the singular value decomposition (SVD). The SVD-based filtering was shown to outperform all other implementation methods in accuracy, especially, when the error covariance matrix is nearly singular. However, previous works produced the SVD-based filter in conventional form where the n-by-n matrix inversion is required at each iteration step of the filter (with n being a dimension of the dynamic state to be estimated). For numerical stability and computational complexity reasons it is preferable to avoid the matrix inversion operation. In this paper, we suggest some new SVD-based implementations that resolve this problem. They are inherently more stable than the previously designed implementation and, additionally, allow for computing the corresponding log likelihood function value. The algorithms are developed in covariance form, and a set of numerical experiments are performed for comparative study.