1: \begin{abstract}

2: A common algorithm for the computation of eigenvalues

3: of real symmetric tridiagonal matrices 

4: is the iteration of certain special maps $F_\sigma$

5: called \textit{shifted $QR$ steps}.

6: Such maps preserve spectrum and

7: a natural common domain is $\ILa$,

8: the manifold of  real symmetric tridiagonal matrices

9: conjugate to the diagonal matrix $\Lambda$.

10: More precisely, a (generic) \textit{shift} $s \in \RR$

11: defines a map $F_s: \ILa \to \ILa$.

12: A strategy $\sigma: \ILa \to \RR$ specifies the shift to be applied at $T$

13: so that $F_\sigma(T) = F_{\sigma(T)}(T)$.

14: Good shift strategies should lead to fast \textit{deflation}:

15: some off-diagonal coordinate tends to zero,

16: allowing for reducing of the problem to submatrices.

17: For topological reasons,

18: continuous shift strategies do not obtain fast deflation;

19: many standard strategies are indeed discontinuous.

20: Practical implementation only gives rise systematically

21: to \textit{bottom deflation},

22: convergence to zero of the lowest off-diagonal entry $\bbb(T)$.

23: For most shift strategies, convergence to zero of $\bbb(T)$ is

24: cubic, $|\bbb(F_\sigma(T))| = \Theta(|\bbb(T)|^k)$ for $k = 3$.

25: The existence of arithmetic progressions in the spectrum of $T$

26: sometimes implies instead quadratic convergence, $k = 2$.

27: The complete integrability of the Toda lattice

28: and the dynamics at non-smooth points

29: are central to our discussion.

30: The text does not assume knowledge of numerical linear algebra.

31: \end{abstract}