4

Is it true that the largest eigenvalue of $A+A^{\rm T}$ is always (in some sense) bigger than the largest eigenvalue of $A$? For example, does it always hold that $$ \quad {\rm Re \,} \lambda_{\rm max}(A) \le {\rm Re \,} \lambda_{\rm max}(A+A^{\rm T}) \quad ? $$

Zeekless
  • 1,599
  • 10
  • 18

2 Answers2

5

EDIT. Your second version is very different from the first one and has the advantage of being correct. Here $\lambda_{max}(A+A^T)$ is its largest eigenvalue, not $\rho(A+A^T)$.

I recall the $2$ keys of the proof: i) if $\lambda$ is an eigenvalue of $A$ associated to $v$, then $\lambda=v^*Av/v^*v$. Beware, $\max_{complex \; vectors} |\dfrac{x^*Ax}{x^*x}|\not= \max_{real \; vectors} |\dfrac{x^TAx}{x^Tx}|$ as the following example shows: $A=\begin{pmatrix}0&1\\-1&0\end{pmatrix}$.

ii) Note that $1/2(x^*Ax+x^*A^Tx)=Re(x^*Ax)$ and not $x^*Ax$.

We can slightly strengthen the final proposition.

$\textbf{Proposition}$. Let $A\in M_n(\mathbb{R})$. If $\lambda(A)\in spectrum(A)$, then

$1/2\lambda_{min}(A+A^T)\leq Re(\lambda(A))\leq 1/2\lambda_{max}(A+A^T)$.

  • Thank you very much for your answer! I didn't know that $1/2(x^Ax+x^A^Tx)={\rm Re} ; (x^*Ax)$. I have edited my answer - is it correct now? Again thank you very much. – Zeekless Jul 17 '18 at 06:18
4

I am proposing a new version of the answer to my own question based on the deleted answer of @Surb, the answer of @TZakrevskiy (how to prove the relationship about spectral radius, numerical radius and matrix two norm?) and the answer of @loup blanc, who pointed out some major mistakes in my original answer.

Let $\lambda(A)$ be an arbitrary eigenvalue of matrix $A \in M_n(\mathbb{R})$ and let $v$ denote the corresponding unit eigenvector ($v^*v=1$).

Then $$ {\rm Re} \, \lambda(A)={\rm Re} \, \frac{v^* \lambda(A) v}{v^*v} ={\rm Re} \, \frac{v^* A v}{v^*v} \le \sup_{x\ne 0}\left( {\rm Re} \, \frac{x^* A x}{x^*x}\right)= $$ $$= \sup_{x\ne 0}\left( \frac{\frac{1}{2}x^* (A + A^{\rm T}) x}{x^*x}\right)= \frac{1}{2}\lambda_{\rm max}(A + A^{\rm T}).$$

Thus we have $$ {\rm Re} \, \lambda(A) \le \frac{\lambda_{\rm max}(A + A^{\rm T})}{2} $$ for any eigenvalue $\lambda(A)$ of real-valued matrix $A$, which gives the answer to the original question.

Zeekless
  • 1,599
  • 10
  • 18
  • 1
    yep that looks good! – Surb Jul 10 '18 at 15:12
  • Thank you, @Surb! – Zeekless Jul 10 '18 at 15:14
  • I think you should accept your own answer (I will delete mine anyway) – Surb Jul 10 '18 at 18:10
  • 1
    @Surb, I will do it when it is possible ("in 2 days", the site says). Your answer was of great help, thank you very much for it! – Zeekless Jul 10 '18 at 18:31
  • 1
    I don't see how this answers your original question. Your question is unclear in the first place -- as loup blanc has pointed out, the meaning of $\lambda_\max(A)$ is ambiguous. And in your question, there isn't a factor of 2 in the denominator. Since $\lambda_\max(A+A^T)$ can be negative. That $\operatorname{Re}(\lambda)\le\lambda_\max(A+A^T)/2$ doesn't imply that $\operatorname{Re}(\lambda)\le\lambda_\max(A+A^T)$. A simple counterexample is given by $A=\pmatrix{-1&-1\ 1&-1}$. The real parts of its two eigenvalues are equal to -1, which are greater than $\lambda_\max(A+A^T)=-2$. – user1551 Jul 17 '18 at 22:18
  • @user1551, your comment has completed it for me, thank you! The first part of the question has the words "in some sense", that's why I took the obtained inequality as an answer. So $|{\rm Re} , \lambda(A)| \le \lambda_{\rm big}(A + A^{\rm T})$, where $\lambda_{\rm big} = {\rm max}(|\lambda_{\rm min}|, |\lambda_{\rm max}|)$? This is even better. Thank you! – Zeekless Jul 17 '18 at 22:56