Let $H$ be a $n \times n$ hermitian matrix which satisfies

$$ \sum_i^n |H_{ij}|^2 \leq 1,~~ \sum_j^n |H_{ij}|^2 \leq 1 . $$ This means that the norm of each column or row vector of $H$ is less than 1.

Then my question is, "Does this condition imply the maximum modulus of its eigenvalues $|\lambda_i|$ to be less than some constant?"

For similar case, the modulus of all the eigenvalues of Stochastic matrix $P$ which satisfies $\sum_i P_{ij}=1$ is less than 1.

2 by 2 case

I first tried $2 \times 2$ matrix. The general $2 \times 2$ hermitian matrix $M$ can be written as

$$ M= \begin{bmatrix} a+d & b+ic\\ b-ic & a-d \end{bmatrix} $$ where $a,b,c,d$ are real numbers. Then above condition implies

$$ (a+d)^2+b^2+c^2 \leq 1 \\ (a-d)^2+b^2+c^2 \leq 1 \\ $$

Also, the two eigenvalues for $M $ is

$$ \lambda_{\pm} = a \pm \sqrt{b^2+c^2+d^2} $$ Since $\pm a$ gives same maximum modulus (if $a>0$, $\lambda_{+}$ will give the maximum modulus, and vise versa), I only consider the maximum value of $\lambda_{+}$ with $a>0$. Squaring both side, we get

$$ (\lambda_+ -a)^2 = b^2+c^2+d^2 $$ Using the two inequality, I get

$$ (\lambda_+ -a)^2 = b^2+c^2+d^2 \leq 1-a^2-2ad\\ (\lambda_+ -a)^2 = b^2+c^2+d^2 \leq 1-a^2+2ad\\ $$ Both condition gives the same lower bound for $\lambda_{+}$ (if the sign of $a,d$ are same the first inequality gives lower bound and for opposite sign, the second inequality gives lower bound, and they gives the same lower bound)

So without loss of generality, I choose $a,d \geq 0$. Then tabulating the first inequality, I get

$$ \lambda_+^2 - 2a\lambda_+ - 1+2a^2+2ad \leq 0 $$

This means that the maximum value of $\lambda_+^{max}$ occurs at the solution of

$$ \lambda_+^2 - 2a\lambda_+ - 1+2a^2+2ad = 0 \\ \Rightarrow \lambda_+ = a \pm \sqrt{1-a^2-2ad} $$

So the maximum value of $\lambda_+^{max}$ would occur at $d=0$ and $\frac{d \lambda}{da}=0 \Rightarrow$ $\lambda_+^{max}=\sqrt{2}$ with $a=\frac{1}{\sqrt{2}}$

So there is a modulus maximum for all the eigenvalues. Does this generally hold for larger Hermitian matrix? If there's any flaw in my derivation, please let me know.


For a square matrix $A$ over $\Bbb C$ let $\|A\|=\left(\sum_{|a_{ij}^2|} \right)^{1/2} $ be its Euclidean norm. If $A=\|a_{ij}\|$ put $A^*=\|\bar a_{ji}\|$. Matrix $A$ is Hermitian iff $A^*=A$.

Theorem. Let $A$ be an $n\times n$ square matrix over $\Bbb C$ with eigenvalues $\mu_1,\dots,\mu_n$. Then

$\sum_{r=1}^n |\mu_r|^2\le \|A\|$, $\sum_{r=1}^n |\operatorname{Re}(\mu_r)|^2\le \|B\|$, $\sum_{r=1}^n |\operatorname{Im}(\mu_r)|^2\le \|B\|$,

where $B=\frac 12 \left(A+A^*\right)$ and $B=\frac 12 \left(A-A^*\right)$. The equality in any one of these inequalities implies the equality in all three and is realized iff $A$ is normal.

See P. Lankaster, “Theory of Matrices”, Academic Press, 1969 (Theorem 7.3.1 in Russian translation, Moskow, Nauka, 1978). Attributed, I expect, to I. Schur, Math. Ann 66 (1909), 488-510.


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.