On the power method for quaternion right eigenvalue problem (Q1789685): Difference between revisions

Let \(\mathbb{H}\) denote the quaternion skew-field and \(\mathbb{H}^{n ,n}\) the set of all \(n \times n\) matrices over \(\mathbb{H}\). For a quaternion matrix, the authors propose a power method for computing the right eigenvalue with the maximum norm and its associated eigenvector. First, they discuss the power method for a quaternion Hermitian matrix. Let \(\lambda _{1} ,\lambda _{2} ,\ldots ,\lambda _{n}\) be the right eigenvalues of a Hermitian matrix \(A \in \mathbb{H}^{n ,n}\) satisfying \(\left | \lambda _{1}\right | >\left | \lambda _{2}\right | \geq \cdots \geq \left | \lambda _{n}\right | \). In analogy with the complex matrix case, the authors propose the following iterative algorithm \[ \begin{cases} y_{k +1} =Au_{k}, \\ u_{k +1} =\frac{y_{k +1}}{\left \| y_{k +1} \right \| _{2}}, \end{cases}\tag{1} \] for \(k =0 ,1 ,2 ,\ldots \). Here, \(\left \| \,\right \| _{2}\) denotes the Euclidean norm and \(u_{0} \in \mathbb{H}^{n}\) is any given initial vector with \(\left \| u_{0}\right \| _{2} =1\). Suppose that \(u_{0} \in \mathbb{H}^{n}\) has a non-zero component in the direction of the right eigenvector \(x_{1}\) associated with the right eigenvalue \(\lambda _{1}\). According to (1), the power method produces two sequences of quaternion vectors \(y_{k +1}\) and \(u_{k} .\) The authors show that if \(k\) is sufficiently large then \(u_{k}^{H}y_{k +1} \approx \lambda _{1}\), i.e., \(u_{k}^{H}y_{k +1}\) is a good approximation of \(\lambda _{1}\) where \(u_{k}^{H}\) is the conjugate transpose of \(u_{k}\). They also compute an approximation of the right eigenvector \(x_{1}\) of \(A\) associated with \(\lambda _{1}\). The authors also present a power method for the right eigenvalue problem for non-Hermitian matrices. They briefly discuss the inverse power method and the shift inverse power method. The authors conclude the paper with some numerical examples.