# Shifter Power Method

If we know all the eigenvalues of a matrix A except the largest one. We want to apply shifted power iteration to get the largest eigenvalue. Something like $(A-\alpha I)$ . Then what should be the value of shift ($\alpha$) to make it converge as fast as possible ?

-

If the eigenvalues of $A$ are all real:
$$\lambda_1 \gt \lambda_2 \gt \ldots \gt \lambda_k$$
then the power iterations of the shifted matrix $A - \alpha I$ converge at the optimal rate when $\alpha = \frac{\lambda_2 + \lambda_k}{2}$.
I am sorry I forgot to mention about the real eigenvalues. So for shifted power method we can say convergence will happen at the rate of $\mod(\frac{\lambda_2 - \alpha}{\lambda_1 - \alpha})$. Which means we want to make $\alpha$ as close as possible to $\lambda_2$. And above condition will ensure that. Is the understanding correct ? –  Anonym Dec 2 '12 at 17:41
You want to put $\alpha$ halfway between $\lambda_2$ and the eigenvalue at the other end of the spectrum, which I labelled $\lambda_k$. If you put $\alpha$ any closer to $\lambda_2$, then the shifted other end of the spectrum $\lambda_k - \alpha$ starts to get bigger (in absolute value) than $\lambda_2 - \alpha$. –  hardmath Dec 2 '12 at 17:46