Numerical experience with Newton-like methods for nonlinear algebraic systems (Q2365380)

The authors report results of extensive numerical tests of several Newton-like iteration methods applied to a collection of 31 systems of nonlinear equations. The systems are all of the general form \(F(x)=0\) where \(F:\mathbb{R}^n\to \mathbb{R}^n\) and the size \(n\) can be regarded as a parameter. Hence the systems are examined for \(n=10,50,100\). The methods tested are of the general form \(x_{i+1}=x_i - p_i\) where \(p_i =H_iF(x_i)\) and \(H_i\) is an approximation to \(F'(x_i)^{-1}\). Among the methods tested are Newton's method, Broyden's two methods, two methods of Greenstadt, a class of ABS methods, and methods of Thomas and of Martinez. Besides varying \(n\), also two sets of starting values are used. The conclusions show Newton's method to have the greatest robustness and speed of convergence, however, at the price of a generally higher overhead cost than the update methods. Among the quasi-Newton methods, the method of Thomas appears to be the method of choice. A few examples portraying basins of attraction for some of the methods with \(n=2\) are given. As is to be expected, they exhibit a fractal-type of structure. Reviewer's remarks: The authors express surprise that the number of iterations required for satisfying the stopping condition varied little with respect to \(n\). However, in view of the mesh independence principle [see, e.g., \textit{E. L. Allgower, K. Böhmer, F. A. Potra} and \textit{W. C. Rheinboldt}, SIAM J. Numer. Anal. 23, 160-169 (1986; Zbl 0591.65043)] this is to be expected for systems of equations arising from discretizations of a large class of operator equations. A number of the examples fall precisely into this category. Some of the citations for the test functions appear to be inaccurate.