Blind source recovery: Theoretical formulations, implementations and application to CDMA communication systems

Blind Source Recovery (BSR) is an autonomous (or unsupervised) stochastic adaptation approach that denotes recovering signals from measurements in environments that may include convolution, transients, and even possible nonlinearity. The primary goal of BSR is to recover original source signals, as best as possible, even in the absence of precise environment identifiability. A discrete-time optimization framework for BSR has been adopted based on minimization of the Kullback-Lieblar Divergence, subject to the constraints of a state-space representation, using the Riemannian contra-variant gradient adaptation.; The modeling of the environment and its representation is vital in the proposed framework. The adoption of the state space framework allows for the derivation of more general update laws capable of catering to most known filtering paradigms. For the interesting case of multi-variable linear time-invariant dynamic BSR, parametric update laws for discrete-time canonical feedforward and feedback state space configurations have been derived. Higher Order Statistics have been explored to obtain adaptation algorithms that are purely blind of actual source distributions. The BSR cases of undercomplete and overcomplete mixtures have also been investigated. A new two-stage BSR algebraic algorithm for sparse static overcomplete mixtures has been developed. Further, the BSR framework has been successfully formulated for multi-user detection in modern CDMA wireless communication networks. Promising new results clearly demonstrate the effectiveness and practicality of the formulated approach.; Computer Simulations have been widely conducted during the course of this research to evaluate the performance of all the developed algorithms for a variety of scenarios. In most cases, the performance verification has been done using actual speech and communication data in the presence of noise. In summary the main contributions of this thesis are (1) Info-theoretic update laws for a multi-variable dynamic state space network were investigated. Natural gradient based BSR update laws for the cases of minimum phase and non-minimum phase mixing environments as well as both feedforward and feedback canonical state space configurations were developed and implemented. (2) An exploration and extension of parametric source distribution models and derivation of adaptive score-functions (or non-linearities) for the BSR of sources with multiple source distributions. (3) The development of the Algebraic ICA (AICA) Algorithm, which is a new ICA algorithm for blind, sparse, static mixing matrix recovery. It enables the BSR from overcomplete mixtures of speech and other sparse distributions using a combination of AICA and interior point linear programming (IP-LP) techniques. (4) Development and simulation of new Blind Multi-user Detection algorithms for DS-CDMA and WCDMA wireless communication networks based on the BSR framework.