4
$\begingroup$

I've shown that for a matrix $A\in \mathbb{R}^{n\times n}$ and an arbitrary $v\in \mathbb{R}^n$, we have the inequality $$\|Av\|\geq \lvert \langle u, v\rangle\rvert\cdot \|A\|$$ where $\|\cdot\|$ is the $2$-norm and $u$ satisfies $\|u\| = 1$ and $\|Au\| = \|A\|$ (the operator norm of $A$ corresponding to the vector $2$-norm). To do this, I used the singular value decomposition. However, I was wondering if there's a different proof that relies on more general properties of Hilbert spaces (under the assumption that $u$ exists in the more general setting where $\langle \cdot, \cdot\rangle$ and $\|\cdot\|$ are the Hilbert space inner product and norm). My own short proof is below:

We first handle the case where $A = \operatorname{diag}(d_1, \ldots, d_n)$ for some $d_1\geq \cdots\geq d_n\geq 0$. Then, $u = e_1$, so $$\lvert \langle u, v\rangle\rvert^2\|A\|^2 = d_1^2v_1^2\leq \sum_{i=1}^n d_i^2v_i^2 = \|Av\|^2$$ Now, we consider general $A$, which has a singular value decomposition $A = U\Sigma V^{\mathrm{T}}$ for $U$ and $V$ orthogonal and $\Sigma$ diagonal in the above form. First, we recall that multiplication by orthogonal matrices preserves the dot product, i.e. $\langle Qx, Qy\rangle = \langle x, y\rangle$ for orthogonal $Q$. This allows us to write $$\|Av\|^2 = \langle U\Sigma V^{\mathrm{T}}v, U\Sigma V^{\mathrm{T}}v\rangle = \langle \Sigma V^{\mathrm{T}}v, \Sigma V^{\mathrm{T}}v\rangle = \|\Sigma V^{\mathrm{T}}v\|^2$$ for arbitrary $v\in \mathbb{R}$. Furthermore, as $\|Au\|^2 = \|\Sigma V^{\mathrm{T}}u\|^2$ and $\|V^{\mathrm{T}}u\|^2 = \|u\|^2 = 1$, we have that $\|A\| = \|\Sigma\|$ and that $y = V^{\mathrm{T}}u$ satisfies $\|\Sigma y\| = \|\Sigma\|$. Combining our earlier results, $$\|Av\|^2 = \|\Sigma V^{\mathrm{T}}v\|^2\geq \lvert\langle y, V^{\mathrm{T}}v\rangle\rvert^2\|\Sigma\|^2 = \lvert\langle V^{\mathrm{T}}u, V^{\mathrm{T}}v\rangle\rvert^2\|A\|^2 = \lvert\langle u, v\rangle\rvert^2\|A\|^2$$

$\endgroup$
1
$\begingroup$

An easier proof is to use the fact that for any $y\in\mathbb R^n$, if $y\perp u$, then $Ay\perp Au$. Now we can write $v=\langle u,v\rangle u+y$ for some $y\perp u$. Let $x_1=\langle u,v\rangle Au$ and $x_2=Ay$. Then we have that $x_1\perp x_2$ and $Av=x_1+x_2$. So $\|Av\|\ge\|x_1\|=|\langle u,v\rangle|\|Au\|=|\langle u,v\rangle|\|A\|$.

Hint for proof of the fact: Suppose that there is some $y\perp u$ with $\langle Ay,Au\rangle\neq0$. We may assume that $\|y\|=1$. Show that there exists some $\theta$ such that $u':=u\cos\theta+y\sin\theta$ satisfies $\|Au'\|>\|Au\|$. This will give us a contradiction.

$\endgroup$
1
  • $\begingroup$ Perfect, thank you for the pointers! I think this achieves a much higher level of generality. $\endgroup$ – Michael L. Mar 7 '19 at 5:59

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.