1: \begin{abstract}

2: One of the most widely used methods for eigenvalue computation

3: is the $QR$ iteration with Wilkinson's shift:

4: here the shift $s$ is the eigenvalue of the bottom $2\times 2$

5: principal minor closest to the corner entry.

6: It has been a long-standing conjecture that the rate

7: of convergence of the algorithm is cubic.

8: In contrast, we show that there exist matrices for which

9: the rate of convergence is strictly quadratic.

10: More precisely, 

11: let $T_\xxxx$ be the $3 \times 3$ matrix having only two nonzero

12: entries $(T_\xxxx)_{12} = (T_\xxxx)_{21} = 1$

13: and let $\ILa$ be the set of real, symmetric

14: tridiagonal matrices with the same spectrum as $T_\xxxx$.

15: There exists a neighborhood $\bUu \subset \ILa$ of $T_\xxxx$

16: which is invariant under Wilkinson's shift strategy

17: with the following properties.

18: For $T_0 \in \bUu$, the sequence of iterates $(T_k)$ exhibits

19: either strictly quadratic or

20: strictly cubic convergence to zero of the entry $(T_k)_{23}$.

21: In fact, quadratic convergence occurs exactly when

22: $\lim T_k = T_\xxxx$.

23: Let $\bXx$ be the union of such quadratically convergent sequences $(T_k)$:

24: the set $\bXx$ has Hausdorff dimension $1$ and is

25: a union of disjoint arcs $\bXx^\sigma$ meeting at $T_\xxxx$,

26: where $\sigma$ ranges over a Cantor set.

27: \end{abstract}

28: