Performance-Guaranteed Dimensionality Reduction of Large-Scale Data for Adaptive Filtering

In this paper, the mean-square performance of the normalized least-mean-square (NLMS) learning algorithm is elaborately analyzed to carefully adjust the sampling size of large-scale data. Based on the performance analysis of the NLMS learning algorithm, a steady-state criterion is derived to determine whether its learning performance reaches the steady state or not. From this criterion, a performance-guaranteed adjusting method of the sampling size is proposed to reduces the redundant computations at the steady state. Furthermore, a fast matrix inverse method also proposed with respect to the NLMS algorithm for large-scale data.