1: \begin{abstract}
2: In this \ej{paper} we take a \ej{quasi-Newton} approach to
3: nonlinear eigenvalue problems (NEPs) of the type \ej{$M(\lambda)v=0$}, where $M:\CC\rightarrow\CC^{n\times n}$ is a holomorphic function.
4: We investigate which types of approximations of
5: the \gm{Jacobian matrix} lead to competitive algorithms, and provide
6: convergence theory.
7: The convergence analysis is based on theory for quasi-Newton methods
8: and Keldysh's theorem for NEPs. We derive new algorithms and
9: also show that several well-established methods for NEPs
10: can be interpreted as quasi-Newton methods,
11: and thereby provide insight to their convergence behavior.
12: In particular, we establish quasi-Newton interpretations of
13: Neumaier's residual inverse iteration and Ruhe's method of successive linear
14: problems.
15:
16: %of equations aim to improve the performance
17: %of Newton's method by approximating the \gm{Jacobian matrix}.
18: %We study quasi-Newton methods for
19: %nonlinear eigenvalue problems
20: %of the type $M(\lambda)v=0$,
21: %where $M: \CC \rightarrow \CC^{n\times n}$ is an analytic function.
22: %We characterize the differences between several choices
23: %of quasi-Newton methods that can be
24: %carried out by solving one linear system $M(\sigma)x=b$
25: %per iteration, where $\sigma$ is a constant (shift).
26: %We show that the well-established method
27: %called the residual inverse iteration
28: %is a quasi-Newton method in disguise.
29: %It is illustrated how the naive quasi-Newton method,
30: %corresponding to a fixed \gm{Jacobian matrix}
31: %is worse than both the residual inverse iteration
32: %and a quasi-Newton method where only one
33: %block of the \gm{Jacobian matrix} is kept constant.
34: %The differences between the quasi-Newton methods are
35: %characterized in terms of
36: %asymptotic convergence rate.
37: %Moreover, by the application of Keldysh's theorem we show
38: %how eigenvalue clustering influences the convergence.
39: %Our theory is supported by
40: %illustrations with large-scale examples.
41: \end{abstract}
42: