Parametrization via secant length and application to path following (Q1819543)

This paper deals with a specific algorithm for finding paths of solutions to nonlinear systems of n equations in \(n+1\) unknowns of the form \[ (1)\quad H(u)=0,\text{ where } H: R^{n+1}\to R^ n. \] The above task arises, for example, when a family of solutions to a parameter-dependent system is required; algorithms to solve it are the well-known continuation methods. In a continuation method, a point \(u^{k+1}\) satisfying (1) is obtained from a point \(u^ k\) satisfying (1) by solving an augmented system; the \((n+1)\)-st equation is \(N_ k(u,\tau)=0\), and defines a local parametrization in terms of \(\tau\). To date, popular N have been of the form \[ N_ k(u,\tau)=(r^ k)^ T(u- u^ k)-\tau, \] where \(r^ k\) is either the tangent direction at \(u^ k\) or a heuristically chosen coordinate vector, and where \(\tau\) is adaptively chosen at each step. H. Keller proposed an alternate parametrization defined by \[ (2)\quad N_ k(u,\tau)=\| u-u^ k\|^ 2-\tau^ 2, \] which could be numerically more desirable. This paper studies properties of the secant parametrization defined by (2). It then proposes an iteresting algorithm for solving the system defined by (1) and (2). This algorithm is based on writing the general solution to the linearization of (1) at the present estimate for \(u^{k+1}\) as a sum of a particular solution and a constant times the homogeneous solution. The constant is then determined by requiring the solution to satisfy (2). We then iterate the step, as in Newton's method. The authors prove convergence results.