9
$\begingroup$

Let $B$ be a cyclic upper-triangular nonnegative matrix,

$$B = \begin{bmatrix} 0 & b_1 & 0& \dots &\dots &0 \\ 0 & 0 & b_2 & 0& \dots & 0\\ \vdots &\vdots&\vdots &\ddots &\ddots &\vdots \\ 0& 0 & 0 & 0 &\dots & b_{n-1}\\ b_n& 0 & 0 & 0 &\dots & 0 \end{bmatrix}$$

with $b_i \in [0,1]$. For any $\alpha > 0$, let us define the matrix

$$C(\alpha)=\alpha B +\alpha^{-1}B^T,$$

where $\alpha$ controls the degree of asymmetry, $\alpha=1$ giving a symmetric $C(\alpha)$ .

I am interested in controlling the localization of the normalized eigenvalues of $C(\alpha)$ on the complex plane. Formally, given any $C(\alpha)$ with eigenvalues $\{\lambda_1,\dots,\lambda_n\}$, let us define the normalized eigenvalues as $$\bar \lambda_i=\frac{\lambda_i}{\max_i |\lambda_i |}$$ By definition, the normalized eigenvalue fall within a circle of radius $1$ on the complex plane. Moreover, when $\alpha=1$ and $C(\alpha)$ is symmetric, the spectrum is real so the normalized eigenvalues fall on the real axis in the interval $[-1,1]$. For larger but finite $\alpha$, I have found numerically that the normalized eigenvalues fall within an ellipse of height $\tanh(\ln \alpha)$. This ellipse is achieved when all $b_i$ are equal so that $C(\alpha)$ is a circulant matrix (as pointed out in the partial answer by @Charr).

Any ideas for how to prove this for a general $B$?


Here's a plot of normalized eigenvalues from many random matrices (random $b_i$), $\alpha=1.2$, with the ellipse shown:

normalized eigenvalues


Edit: Let $v$ be the eigenvector associated with the largest (in magnitude) eigenvalue $\lambda^*$ of $C(\alpha)$. By Perron-Frobenius, we may assume $\lambda^*$ and $v$ are real-valued and nonnegative, so \begin{align}\lambda^* = v^T C(\alpha)v = (\alpha + 1/\alpha) v^T B v \end{align} Then, consider any other eigenvalue $\lambda_i$ with right eigenvector $u$ ($\Vert u \Vert =1$). It may be helpful to express $\bar{\lambda_i}$ as $$\begin{align} \bar{\lambda}_i = \frac{1}{\lambda^*}{u^* C(\alpha) u} &= \frac{1}{\lambda^*}\left[(\alpha + 1/\alpha) \,u^*\frac{B+B^T}{2}u + (\alpha -1/\alpha) u^*\frac{B-B^T}{2}u\right]\\ &= \mathrm{Re} x + \mathrm{i} \tanh(\ln \alpha)\mathrm{Im}\, x \end{align}$$ where for convenience I defined $x= u^* B u/v^T B v$. Thus, the problem reduces to showing that \begin{align}\vert x\vert =\left\vert \frac{u^* B u}{v^T B v}\right\vert \le 1 \end{align} where $u,v$ are two eigenvectors of $C(\alpha)$, with $v$ being the largest (Perron-Frobenius) one.

$\endgroup$
0

1 Answer 1

7
$\begingroup$

Here's a partial answer. Consider the case where all $b_{i}=1$. Then $C$ is a circulant matrix, whose eigenvalues are given by $\alpha ^{-1}\omega ^{-j}+\alpha \omega^{-(n-1)j}$, $j=0,\ldots ,n-1$, see, e.g., Wikipedia. The $j=0$ eigenvalue, $\alpha ^{-1}+\alpha $ is real and defines the length of the semimajor axis. For $n$ a multiple of 4, there is an eigenvalue on the imaginary axis at $(\alpha -\alpha ^{-1})i$, which defines the length of the semi-minor axis. The ratio is $$ \frac{\alpha -\alpha ^{-1}}{\alpha +\alpha ^{-1}}=\frac{\alpha ^{2}-1}{% \alpha ^{2}+1}=\tanh (\ln \alpha ) $$ It remains to show that for $0\leq b_{i}\leq 1$, the eigenvalues are inside this region.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.