Accelerating convergence using a shift

We are now in a position to understand how convergence can be accelerated. Assume we use $A-\lambda I$ instead of $A$. Then the evals are shifted by $\lambda$: ${\lambda }_i-\lambda$. This can be used to accelerate convergence.

For example, assume that $\lambda \approx {\lambda }_n$ and we shift. We get:

$$|\lambda -{\lambda }_n|\ll |\lambda -{\lambda }_{n-1}|.$$

The last eval will converge very rapidly. Shifting allows accelerating convergence. We will see that, if the shifting is done correctly, we get a quadratic convergence of the eval!

This method works well.

Note that the focus is now on the eigenvalues rather than $Q$. We need to efficiently compute $T_k$. It turns out that there is a simple algorithm to do that.