1: \begin{abstract}

2: If the numerical range of a matrix is contained in the right half-plane, 

3: the GMRES algorithm for solving linear systems will make progress 

4: at every iteration.  

5: In his Ph.D.\ dissertation, Howard Elman derived a bound that

6: guarantees convergence.  

7: When all eigenvalues are in the right half-plane but the

8: numerical range contains the origin, GMRES need not make progress 

9: at every step, 

10: and Elman's bound does not apply.  

11: By solving a Lyapunov equation, one can construct an inner product 

12: in which the numerical range is contained in the right half-plane.

13: One can bound GMRES (run in the standard Euclidean norm) by applying

14: Elman's bound in this new inner product, at the cost of a multiplicative 

15: constant that characterizes the distortion caused by the change of inner 

16: product.  

17: Using Lyapunov inverse iteration, one can build a family of 

18: suitable inner products, trading off the location of the numerical

19: range with the size of constant. 

20: This approach complements techniques recently proposed by 

21: Greenbaum and colleagues for eliminating the origin from the 

22: numerical range for GMRES convergence analysis.

23: \end{abstract}

24: