Correntropy as a nonlinear local similarity measure which is robust to the outliers has been extensively utilized in several applications. However, there is no analytical solution for maximum correntropy criterion (MCC), so iterative optimization methods are used to solve MCC problem. Fixed-point algorithm under MCC (FP-MCC) is an efficient way to solve a linear regression with correntropy cost function. Just recently, the sufficient condition to guarantee the convergence of FP-MCC has been obtained, but the convergence rate of this method has not been studied more. Since the main advantage of the Fixed-point method in comparison with gradient based method is its faster convergence rate, in this paper we analyze the convergence rate of FP-MCC method and prove with the proper selection of the kernel bandwidth that this method could converge to the optimal solution at least quadratically. In addition, we unveil a trade-off between convergence rate and accuracy of FP-MCC.

Keywords