We apply RLS algorithm, LMS algorithm and the proposed algorithm to linearize the PA and compare their performance in terms of convergence speed, spectral regrowth suppression and constellation diagrams. The normalized mean square error (NMSE) is observed to determine the required iteration number of each algorithm. If the difference between the maximum and the minimum of the 100 consecutive NMSE values is smaller than 0.2dB, the algorithm is regarded as
converged. In all power spectral density (PSD) lines, output PSDs are normalized with respect to input PSD for easy visual comparison. Fig. 5(a) shows the convergence curves of RLS algorithm, LMS algorithm and the proposed algorithm. In Fig. 5(a), the green solid line is the convergence curve of RLS algorithm, the green dash line is the convergence curve of RLS algorithm with conventional polynomials model [14], the black solid line is the convergence curve of LMS algorithm (δ = 0.05), the black dash line is the convergence curve of LMS algorithm (δ = 0.005), the blue solid line is the convergence curve of the proposed algorithm (1/n), and the blue dash line is the convergence curve of the proposed algorithm (δ(n)). Fig. 5(a) indicates that RLS algorithm converges quickly and steady. In addition, the RLS algorithm with orthonormal polynomials converges faster than that with conventional polynomials as the orthonormal basis functions can improve the numerical stability. LMS algorithm converges quickly but unsteady when the step size δ =0.05. On the contrary, LMS algorithm converges steady but slowly when the step size δ =0.005. The proposed algorithm achieves almost the same performance as that of the RLS algorithm. Moreover, the proposed algorithm with δ(n) converges a bit slowly than the proposed algorithm with 1/n does. Theoretically, the convergence speed of the proposed algorithm with 1/n should be the same as that of the RLS algorithm. However, in practice, the new basis functions ψ may not be exactly orthonormal for a given set of data samples. Therefore, some performance degradation in convergence speed happens in our simulation results, which validate our analysis in Section III-B. Table IV shows the details of all the algorithms discussed above