2
$\begingroup$

I am attempting to prove that if $A$ is a non-zero complex $n \times n$ matrix, then $r(A) = \sqrt{\operatorname{tr}(AA^*)}$ if and only if $A$ is normal of rank 1, where $r(A)$ denotes the numerical radius of $A$ defined as $r(A) = \max \{ |xAx^*| \ | \ xx^* = 1 \}$ (I use row vectors here). This is my attempt thus far; the doubt I have is whether $\lVert u^*v \rVert_2 = \lVert u \rVert \lVert v \rVert$ implies $u = kv$ for some scalar $k$ (the very last paragraph of my attempt). Note that this is not the classical Cauchy-Schwarz inequality because I am using row vectors, not column vectors.

Let $\lVert A \rVert_F$ denote the Frobenius norm of $A$. If $A$ has singular values $\sigma_1 \geq \sigma_2 \geq \cdots \geq \sigma_n \geq 0$. Then $$r(A) \leq \lVert A \rVert_2 = \sigma_1 \leq \sqrt{\sum_{j=1}^n \sigma_j^2} = \lVert A \rVert_F = \sqrt{\operatorname{tr}(AA^*)}.$$ In the case of equality, $A$ is given by $\sigma_1u^*v$ for some row vectors $u,v$ with $\lVert u \rVert = \lVert v \rVert = 1$. Now, observe that $$r(A) = r(\sigma_1 u^*v) = \sigma_1 r(u^*v) \leq \sigma_1 \lVert u^*v \rVert_2 \leq \sigma_1 \lVert u \rVert \lVert v \rVert = \sigma_1.$$ Since $r(A) = \sigma_1$, we have $\lVert u^*v \rVert_2 = \lVert u \rVert \lVert v \rVert$, and so $u = kv$ for some scalar $k$ satisfying $|k|=1$. We deduce that $A = k\sigma_1 u^*u$ is a normal matrix of rank 1, and the proof the complete?

$\endgroup$
  • 1
    $\begingroup$ no, it's not complete since $\|u^*v\|=\|u\|\|v\|$ is true for any couple of vectors (try $u = [0,1]$ and $v=[1,0]$) $\endgroup$ – Exodd Dec 5 '17 at 15:56
3
$\begingroup$

You are right to doubt the last point. In fact it is false:

$$ \|u^*v\|_2^2 = \rho(v^*u u^*v) = \|u\|^2_2 \rho(v^*v) = \|u\|^2_2\|v\|^2_2 $$ where $\rho$ is the spectral radius.

The real fact is the opposite:

$\|uv^*\| = \|u\|\|v\| \iff \exists a,b\in \mathbb C: au = bv$

and you can use this to prove your theorem:

$$r(A)=\sigma_i \implies r(u^*v)=1 \implies \max \{ |xu^*vx^*| \ | \ xx^* = 1 \} = 1 \implies |xu^*| = |vx^*| = 1 \implies\exists a,b,c\in \mathbb C : au=bx=cv $$

$\endgroup$

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.