On Backtracking Failure in Newton–GMRES Methods with a Demonstration for the Navier–Stokes Equations

Abstract

In an earlier study of inexact Newton methods, we pointed out that certain counterintuitive behavior may occur when applying residual backtracking to the NavierStokes equations with heat and mass transport. Specifically, it was observed that a Newton-GMRES method globalized by backtracking (linesearch, damping) may be less robust when high accuracy is required of each linear solve in the Newton sequence than when less accuracy is required. In this brief discussion, we offer a possible explanation for this phenomenon, together with an illustrative numerical experiment involving the Navier-Stokes equations. c 2002 Elsevier Science (USA)

Jacobian-free Newton–Krylov methods: a survey of approaches and applications

“…Too large a value for results in less work for the Krylov method but more nonlinear iterations, whereas too small a v alue for results in more Krylov iterations per Newton iteration. The forcing function and the issue of oversolving" a Newton step has gained recent interest 148,159 . It has been demonstrated that in some situation the Newton connvergence may actually su er if is too small in early Newton iterations.…”

Newton-conjugate-gradient methods for solitary wave computations

“…Thus an effective strategy for minimizing oversolving is to use the accuracy of the approximate solution u n to determine adaptively the accuracy with which the linear Newton-correction Eq. (2.5) is solved [26]. The accuracy of u n can be measured by…”

“…But unfortunately this method works only for one-dimensional problems, or higher-dimensional problems which can be reduced to one-dimensional problems (through symmetry reduction). The Petviashvili method was (see [26,27] for instance). The Newton-CG and Newton-BCG methods proposed in this paper are two particular cases of the Newton-Krylov methods.…”

Composing Scalable Nonlinear Algebraic Solvers

“…BT defaults to taking the full step if that step is sufficient with respect to the Wolfe conditions [74], and it does no more work unless necessary. The BT line search may stagnate entirely for ill-conditioned Jacobians [69]. For a general step BT is not appropriate for a few reasons.…”