1: \begin{abstract}

2: \noindent

3: The unordered eigenvalues of a Hermitian matrix function depending on one parameter 

4: analytically is analytic. Ordering these eigenvalues yields piece-wise analytic functions. 

5: For multi-variate Hermitian matrix functions depending on $d$ parameters analytically 

6: the ordered eigenvalues are piece-wise analytic along lines in the $d$-dimensional space.

7: These classical results imply the boundedness of the second derivatives of the pieces 

8: defining the eigenvalue functions along any direction. We derive an algorithm 

9: based on the boundedness of these second derivatives for the global minimization of an 

10: eigenvalue of an analytic Hermitian matrix function. The algorithm is globally convergent.

11: It determines the globally minimal value of a piece-wise quadratic under-estimator for the 

12: eigenvalue function repeatedly. In the multi-variate case determining this globally minimal value is 

13: equivalent to the solution of a quadratic program. The derivatives of the eigenvalue functions 

14: are used to construct quadratic models yielding rapid global convergence as compared to 

15: traditional global optimization algorithms. The applications that we have in mind include 

16: the $H_{\infty}$ norm of a linear system, numerical radius, distance to uncontrollability 

17: and distance to the nearest defective matrix. \\

18: 

19: \noindent

20: \textbf{Key words.} 

21: Hermitian eigenvalues, analytic, global optimization, perturbation of eigenvalues, quadratic program \\

22: 

23: 

24: \noindent

25: \textbf{AMS subject classifications.}

26: 65F15,  90C26 

27: \end{abstract}

28: