Adaptive kernel canonical correlation analysis algorithms for maximum and minimum variance

We describe two formulations of the kernel canonical correlation analysis (KCCA) problem for multiple data sets. The kernel-based algorithms, which allow one to measure nonlinear relationships between the data sets, are obtained as nonlinear extensions of the classical maximum variance (MAX-VAR) and minimum variance (MINVAR) canonical correlation analysis (CCA) formulations. We then show how adaptive versions of these algorithms can be obtained by reformulating KCCA as a set of coupled kernel recursive least-squares algorithms. We illustrate the performance of the proposed algorithms on a nonlinear identification application and a cognitive radio detection problem.