Linear prediction theory. A mathematical basis for adaptive systems (Q1188618)

The development of appropriate algorithms for parameter estimation and tracking based on linear prediction approaches is the main topic of this book. In the case of a least squares solution, the system parameters are determined by a special set of linear equations termed normal equations. Properties of the normal equations are considered in Chapter 2. The normal equations have an important interpretation as a discrete form of the Wiener-Hopf equation. Classical algorithms for inversion of the normal equations are treated in Chapter 3. A situation of particular interest arises when the parameter set of a current time step can be computed in the presence of the parameter set of the previous time step. In this case the linear prediction problem has a recursive solution. The main goal in the recursive least squares (RLS) algorithm is the recursive updating of the inverse covariance matrix of the observed process. An algorithm for recursive inversion of the normal equations with a computational complexity that depends linearly on the model order became known as the fast Kalman algorithm. The algebraic derivation of the fast Kalman algorithm succeeds by introducing a scheme in the Sherman-Morrison identity, which is treated in Chapter 5. The RLS algorithms of the Kalman type are based on an algebraic solution of the normal equations. Using a geometrical model, the least square estimation problem can be stated more conveniently. A consequence of the orthogonalization procedure for solving the normal equations is that this leads to a completely new structure of the prediction error filter. Chapter 6 is devoted to this alternative structure of the prediction error filter, which has been termed the ladder structure. There are three classes of ladder algorithms. The first class, treated in Chapter 7, are the algorithms of Levinson type. The possibly most prominent member of this class of algorithms is the Levinson-Durbin algorithm for a block-wise calculation of prediction error filter parameters. A second class of ladder algorithms is formed by the class of methods which seek to replace the Levinson recursion by some type of inner-product recursions. These methods are discussed in Chapter 8. Fast estimation algorithms can also be derived in the case of the ladder structure. They constitute the third class of ladder algorithms, treated in Chapter 9. After introducing the most important concepts and algorithms of linear prediction theory for the simplest case of a scalar AR model, Chapter 10 presents ways how these techniques can be extended and applied to more complicated estimation problems. Such extensions of the signal model are the joint process estimation case. Other extensions deal with the case of a vector autoregressive process model. Chapter 11 discusses several aspects of implementation, such as numerical accuracy and stability. A large number of references are provided to enable the interested reader to find links to the application-oriented literature.