1: \begin{abstract}

2: In many applications, linear systems arise where the

3: coefficient matrix takes the special form 

4: $\MI + \MK + \ME$, where $\MI$ is the identity matrix of

5: dimension $n$, ${\rm rank}(\MK) = p \ll n$, and 

6: $\|\ME\| \leq \epsilon < 1$. GMRES convergence rates for

7: linear systems with coefficient matrices of the forms 

8: $\MI + \MK$ and $\MI + \ME$ are guaranteed by well-known

9: theory, but only relatively weak convergence bounds

10: specific to matrices of the form $\MI + \MK + \ME$

11: currently exist. In this paper, we explore the

12: convergence properties of linear systems with such coefficient

13: matrices by considering the pseudospectrum of

14: $\MI + \MK$.\ \ We derive a bound for the GMRES residual in

15: terms of $\epsilon$ when approximately solving the linear

16: system $(\MI + \MK + \ME)\Vx = \Vb$ and identify the eigenvalues

17: of $\MI + \MK$ that are sensitive to perturbation. In

18: particular, while a clustered spectrum away from the origin

19: is often a good indicator of fast GMRES convergence, that

20: convergence may be slow when some of those eigenvalues are

21: ill-conditioned. We show there can be at most $2p$

22: eigenvalues of $\MI + \MK$ that are sensitive to

23: small perturbations. We present numerical results when using GMRES to solve a sequence of linear systems of the form 

24: $(\MI + \MK_j + \ME_j)\Vx_j = \Vb_j$

25:  that arise from the application of Broyden's method to solve a nonlinear partial differential equation.

26: \end{abstract}

27: