Nils Friess,Alexander D. Gilbert,Robert Scheichl
Abstract:We introduce a new Projected Rayleigh Quotient Iteration aimed at improving the convergence behaviour of classic Rayleigh Quotient iteration (RQI) by incorporating approximate information about the target eigenvector at each step. While classic RQI exhibits local cubic convergence for Hermitian matrices, its global behaviour can be unpredictable, whereby it may converge to an eigenvalue far away from the target, even when started with accurate initial conditions. This problem is exacerbated when the eigenvalues are closely spaced. The key idea of the new algorithm is at each step to add a complex-valued projection to the original matrix (that depends on the current eigenvector approximation), such that the unwanted eigenvalues are lifted into the complex plane while the target stays close to the real line, thereby increasing the spacing between the target eigenvalue and the rest of the spectrum. Making better use of the eigenvector approximation leads to more robust convergence behaviour and the new method converges reliably to the correct target eigenpair for a significantly wider range of initial vectors than does classic RQI. We prove that the method converges locally cubically and we present several numerical examples demonstrating the improved global convergence behaviour. In particular, we apply it to compute eigenvalues in a band-gap spectrum of a Sturm-Liouville operator used to model photonic crystal fibres, where the target and unwanted eigenvalues are closely spaced. The examples show that the new method converges to the desired eigenpair even when the eigenvalue spacing is very small, often succeeding when classic RQI fails.
What problem does this paper attempt to address?
This paper attempts to solve the convergence problems encountered by the classical Rayleigh Quotient Iteration (RQI) when calculating internal eigenvalues, especially in the case where eigenvalues are closely arranged. Specifically, although the classical RQI has local cubic convergence for Hermitian matrices, its global behavior is unpredictable and may converge to eigenvalues far from the target, even if the initial conditions are very accurate. This problem is particularly serious when the eigenvalue spacing is very small.
To solve this problem, the author proposes a new Projected Rayleigh Quotient Iteration (PRQI), which improves the convergence behavior of the classical RQI by introducing complex - valued projections at each step. The core idea of the new method is to perform a complex projection on the original matrix at each step, so that the unwanted eigenvalues are lifted into the complex plane, while the target eigenvalue remains close to the real axis, thereby artificially increasing the spacing between the target eigenvalue and other eigenvalues. This method enables the new algorithm to converge more reliably to the correct target eigenpair and performs well for a wider range of initial vectors.
The following are the main steps and formulas of this method:
1. **Introducing complex projection**:
Suppose we have a Hermitian matrix \(A\in\mathbb{C}^{n\times n}\), and a given approximation of the target eigenvector \(u\approx v\). We perturb the matrix \(A\) as follows:
\[
A\rightarrow A + i\gamma(I - uu^*)
\]
where \(i\) is the imaginary unit, \(u^*\) represents the conjugate transpose of \(u\), \(I - uu^*\) is the matrix projected onto the orthogonal complement space of \(u\), and \(\gamma> 0\) is a real scalar that controls the size of the perturbation.
2. **Updating the perturbed matrix**:
In each iteration, use the previous eigenvector approximation \(u\) to update the perturbed matrix and gradually reduce the size of the perturbation. Specifically, replace \(\gamma\) with a non - negative decreasing sequence \((\gamma(k))_{k\geq0}\subset\mathbb{R}\), so that as \(k\rightarrow\infty\), the perturbation gradually becomes smaller and finally returns to the original unperturbed problem.
3. **Simplified version**:
Through lemma proof, the algorithm can be simplified. Only an imaginary part offset needs to be added in each iteration. The simplified algorithm is as follows:
```markdown
Algorithm 3.2 Projected Rayleigh Quotient Iteration (Simplified Version)
Input: \(x(0)\in\mathbb{C}^n\) and \(\|x(0)\| = 1\)
for \(k = 0,1,2,\cdots\), until the stop condition is met do
Calculate the Rayleigh quotient \(\mu(k)=(x(k))^*Ax(k)\)
Update the scaling factor \(\gamma(k)\)
Solve the linear equation system \([A - (\mu(k)-i\gamma(k))I]z(k + 1)=x(k)\) to obtain \(z(k + 1)\)
Normalize \(x(k + 1)=z(k + 1)/\|z(k + 1)\|
end for
```
4. **Convergence analysis**:
By choosing an appropriate sequence of \(\gamma(k)\), the local quadratic or cubic convergence of the algorithm can be guaranteed. Specifically, assume that \(\gamma(k)\) satisfies:
\[
0 <\gamma(k)\leq C_\gamma\tan^{1 + q}(\theta(x(