1: \begin{abstract}

2: By extending the classical analysis techniques due to 

3: Samokish, Faddeev and Faddeeva, and Longsine and McCormick among others, 

4: we prove the convergence 

5: of preconditioned steepest descent with implicit deflation (PSD-id) method

6: for solving Hermitian-definite generalized eigenvalue problems.  

7: Furthermore, we derive a nonasymptotic estimate of the rate of convergence

8: of the \psdid method.  We show that with the proper choice of the shift,

9: the indefinite shift-and-invert preconditioner

10: is a locally accelerated preconditioner, and is

11: asymptotically optimal that leads to superlinear convergence. 

12: Numerical examples are presented to verify the theoretical results 

13: on the convergence behavior of the \psdid method for solving 

14: ill-conditioned Hermitian-definite generalized eigenvalue problems arising from 

15: electronic structure calculations.  

16: While rigorous and full-scale convergence proofs of 

17: preconditioned block steepest descent methods in practical use 

18: still largely eludes us, we believe the theoretical results presented 

19: in this paper sheds light on an improved understanding of the

20: convergence behavior of these block methods.

21: 

22: \vskip 2mm

23: 

24: \noindent {\bf Key words.}

25: eigenvalue problem, steepest descent method, preconditioning, 

26: superlinear convergence.

27: 

28: \vskip2mm

29: 

30: \noindent {\bf MSC.}  65F08, 65F15, 65Z05, 15A12.

31: \end{abstract}

32: