Critically Sampled Subband Adaptive Filtering Using Special Structure Of The Input Sample Correlation Matrix

In many applications, subband adaptive filter structures based on the gradient algorithms have been shown to be superior computationally and performancewise. However, in practice, the convergence improvement gained by increasing the number of subbands is ultimately limited by nonideal filter banks and finite-word-length effects. In many adaptive algorithms, direct matrix inversion (DMI) algorithm provides with excellent convergence performance, whereas the huge computational complexity of matrix inverse is unacceptable in practice. This paper presents a subband direct matrix inversion algorithm suitable for use within a recently proposed adaptive filter structure employing critically sampled filter banks. This new method reduces the computational complexity by using the block tridiagonal structure of the input sample correlation matrix, and at the same time keeps the property of fast convergence. Experimental results show that the output residue power of the subband DMI algorithm is around 3dB upon the optimum value after only 2K updating of the adaptive subfilters, where K is the dimension of the adaptive subfilters .