Paper 3, Section II, A

Numerical Analysis | Part II, 2017

Let AA be a real symmetric n×nn \times n matrix with real and distinct eigenvalues 0=λ1<<λn1=1<λn0=\lambda_{1}<\cdots<\lambda_{n-1}=1<\lambda_{n} and a corresponding orthogonal basis of normalized real eigenvectors (wi)i=1n\left(\mathbf{w}_{i}\right)_{i=1}^{n}.

To estimate the eigenvector wn\mathbf{w}_{n} of AA whose eigenvalue is λn\lambda_{n}, the power method with shifts is employed which has the following form:

y=(AskI)x(k),x(k+1)=y/y,skR,k=0,1,2,\mathbf{y}=\left(A-s_{k} I\right) \mathbf{x}^{(k)}, \quad \mathbf{x}^{(k+1)}=\mathbf{y} /\|\mathbf{y}\|, \quad s_{k} \in \mathbb{R}, \quad k=0,1,2, \ldots

Three versions of this method are considered:

(i) no shift: sk0s_{k} \equiv 0;

(ii) single shift: sk12s_{k} \equiv \frac{1}{2};

(iii) double shift: s2s0=14(2+2),s2+1s1=14(22)s_{2 \ell} \equiv s_{0}=\frac{1}{4}(2+\sqrt{2}), s_{2 \ell+1} \equiv s_{1}=\frac{1}{4}(2-\sqrt{2}).

Assume that λn=1+ϵ\lambda_{n}=1+\epsilon, where ϵ>0\epsilon>0 is very small, so that the terms O(ϵ2)\mathcal{O}\left(\epsilon^{2}\right) are negligible, and that x(0)\mathbf{x}^{(0)} contains substantial components of all the eigenvectors.

By considering the approximation after 2m2 m iterations in the form

x(2m)=±wn+O(ρ2m)(m)\mathbf{x}^{(2 m)}=\pm \mathbf{w}_{n}+\mathcal{O}\left(\rho^{2 m}\right) \quad(m \rightarrow \infty)

find ρ\rho as a function of ϵ\epsilon for each of the three versions of the method.

Compare the convergence rates of the three versions of the method, with reference to the number of iterations needed to achieve a prescribed accuracy.

Typos? Please submit corrections to this page on GitHub.