Consider a $N \times N$ matrix $A=B_1+B_2$, where $B_1$ is a diagonal matrix with all the diagonal entries between $0$ and $1$ and $B_2$ is a skew symmetric matrix which can be written as

$$B_2= \begin{bmatrix} 0_{N-1} & f \\ -f^T & 0 \end{bmatrix}$$

where $0_{N-1}$ is the $N-1 \times N-1$ zero matrix, $f$ is a $N-1$ by $1$ vector. So $B_2=-B_2^T$.

My question is how many complex eigenvalues (non-zero imaginary part) $A$ has?

I have run some numerical simulations. It seems that $A$ can at most have one pair of complex eigenvalues. And it is plausible to me because the skew symmetric $B_2$ is only rank $2$.

How can we prove this mathematically or come up with a counter example? If needed, we can put an upper bound on $|f|_2$. But I don't know that if it is necessary, as indicated from my simulations.

  • $\begingroup$ I don't know if it helps, but using the Schur complement, the characteristic polynomial of $A$ is$$p_A(z) = \prod_{k=1}^{n-1}(z_n-z) + \sum_{j=1}^{n-1}\left(\prod_{k=1,k\neq j}^{n-1}(z_k-z)\right)f_j^2,$$where $z_k$ is the $k$-th enty in the diagonal of $B_1$. $\endgroup$ – Friedrich Philipp Mar 18 '16 at 15:43
  • $\begingroup$ I meant $z_k$ instead of $z_n$ in the first product. $\endgroup$ – Friedrich Philipp Mar 18 '16 at 15:58
  • $\begingroup$ Does $p_A(z)$ depend on $z_N$, last entry of $B_1$? $\endgroup$ – user322840 Mar 18 '16 at 16:01
  • $\begingroup$ The first product ranges to $n$. So, it is $\prod_{k=1}^n(z_k-z)$. Sorry for the confusion. $\endgroup$ – Friedrich Philipp Mar 18 '16 at 16:06
  • $\begingroup$ What you could try is the argument principle en.wikipedia.org/wiki/Argument_principle. But this means you would have to derive $p_A$ and consider $p_A'/p_A$. Not that nice... $\endgroup$ – Friedrich Philipp Mar 18 '16 at 16:08

We only need $B_1$ to be diagonal. Positive definiteness is not needed. Let $A=\pmatrix{D&f\\ -f^T&a}$. If some entries of $f$ is zero, we can reduce the problem to case with a smaller-sized $A$. Assume $f$ is entrywise nonzero. As Friedrich Philipp suggests in his comment, we may make use of Schur complement: \begin{align} \det(tI-A)&=\det(tI-D)\left(t-a+f^T(tI-D)^{-1}f)\right)\\ &=\det(tI-D)\left(t-a+\sum_{j=1}^{n-1} \frac{f_j^2}{t-d_j}\right). \end{align} Hence it suffices to show that $g(t)=t-a+\sum_{j=1}^{n-1} \frac{f_j^2}{t-d_j}$ has at least $n-2$ real roots.

By permutation, we may assume that the diagonal entries are arranged in strictly ascending order. Since $g(d_j^-)=-\infty$ and $g(d_j^+)=+\infty$, if the diagonal elements of $D$ are distinct, $g$ changes sign on every interval $(d_j,d_{j+1})$. Consequently, $g$ changes sign at least $n-2$ times on $\mathbb R\setminus\{d_1,d_2,\ldots,d_{n-1}\}$ and $A$ has at least $n-2$ (distinct) real eigenvalues. By continuity of eigenvalues in the matrix entries, $A$ also has at least $n-2$ real eigenvalues when some diagonal elements of $D$ are the same.

| cite | improve this answer | |
  • $\begingroup$ Thank you and @FriedrichPhilipp so much. I think your answer solves the problem! If two of the diagonal elements are the same, one eigenvalue of the A matrix will be the same as that diagonal element. So we are back to the distinct eigenvalue case. $\endgroup$ – user322840 Mar 20 '16 at 15:40
  • $\begingroup$ @user322840 Huh? No, that's not what I mean. Although some diagonal entries of $D$ may be equal to each other, $D$ is always the limit of a sequence of diagonal matrices that have distinct diagonal entries. In fact, if we define $D_k=D+\operatorname{diag}(k^{-1},k^{-2},\ldots,k^{-n})$, then $\lim_{k\to\infty}D_k=D$ and $D_k$ has distinct diagonal elements when $k$ is sufficently large. Let $A_k$ be the matrix obtained by replacing the $D$ in $A$ by $D_k$. Then each $A_k$ has at least $n-2$ real eigenvalues. Therefore, $A$, the limit of $A_k$, also has at least $n-2$ real eigenvalues. $\endgroup$ – user1551 Mar 20 '16 at 16:02
  • $\begingroup$ Thanks, @user1551, for your detailed explanation. Here is what I meant: let $d_{1}=d_{2}\neq d_j, \forall j\neq 1,2$. We then have $\det (tI-A) = \Pi_{i=1}^{n-1}(t-d_i)\times (t-a+\sum_{i=1}^2\frac{f_1^2+f_2^2}{t-d_1}+\sum_{j=3}^{n-1}\frac{f_j^2}{t-d_j})=\Pi_{i=2}^{n-1}(t-d_i)\times ((t-a)(t-d_1)+{f_1^2+f_2^2}+(t-d_1)\sum_{j=3}^{n-1}\frac{f_j^2}{t-d_j})$. Then in addition to an eigenvalue at $t=d_1=d_2$, $A$ has at least $n-3$ other real eigenvalues. $\endgroup$ – user322840 Mar 20 '16 at 16:40
  • $\begingroup$ Probably not as simple as your method. $\endgroup$ – user322840 Mar 20 '16 at 16:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.