A note on the Levenberg-Marquardt parameter (Q1002289)

The authors consider the problem of determining efficient Levenberg-Marquardt (LM) parameters for systems of nonlinear equations (1) \(F(x)= 0\), where \(F: \mathbb{R}^n\to\mathbb{R}^n\) is a continuously differentiable function, provided \(\| F(x)\|\) satisfies a local error bound condition which is weaker than nonsingularity. The authors present a more general choice of the LM parameters which not only includes all the LM parameters choices of the type \(\lambda_k=\theta\| F(\chi_k)\|+ (1-\theta)\| J^T_k F_k\|\) with \(\theta\) being a constant in \([0,1]\) assuring the quadratic convergence for (1) but also takes into account the computational efficiency of the LM algorithm. They discuss both the global and local convergence and also the relationship between the new LM method and trust region method. Conditions under which the new LM algorithm has superlinear or quadratic convergence, respectively, are given. Finally, some numerical results are presented to illustrate the efficiency of the new method.