1: \begin{abstract}

2: Using a new analysis approach, we establish a general

3: convergence theory of the Shift-Invert Residual Arnoldi (SIRA)

4: method for computing a simple eigenvalue nearest to a given target $\sigma$ and

5: the associated eigenvector. In SIRA, a subspace expansion vector at each step

6: is obtained by solving a certain inner linear system. We prove that the

7: inexact SIRA method mimics the exact SIRA well, that is,

8: the former uses almost the same outer iterations to achieve

9: the convergence as the latter does if all the inner linear systems

10: are iteratively solved with {\em low} or {\em modest} accuracy during

11: outer iterations. Based on the theory, we design practical stopping criteria

12: for inner solves. Our analysis is on one step expansion of subspace and the approach

13: applies to the Jacobi--Davidson (JD) method with the fixed target $\sigma$ as well,

14: and a similar general convergence theory is obtained for it.

15: Numerical experiments confirm our theory and demonstrate that the inexact SIRA and JD

16: are similarly effective and are considerably superior to the inexact SIA.

17: 

18: \smallskip

19: 

20: {\bf Keywords.} Subspace expansion, expansion vector, inexact,

21: low or modest accuracy, the SIRA method, the JD method,

22: inner iteration, outer iteration.

23: 

24: \smallskip

25: 

26: {\bf AMS subject classifications.} 65F15, 15A18, 65F10.

27: \end{abstract}