I know that $\|AB\|_2 \leq \|A\|_2\|B\|_2$. However, let suppose that I have a matrix A, that is symmetric and positive definite. So, there exist a factorization such that $A = LL^t$ (Cholesky). Suppose $\|A\|_2 = 4$. I am interested in knowing the value of $\|L^t\|_2$. My first guess is that it should be 2, because $\|L^t\|_2$ = $\|L\|_2$, using the fact that $\|LL^t\|_2 = \|A\|_2$ and asuming that $\|L\|_2$$\|L^t\|_2 = \|LL^t\|_2$. However I don't know when does this hold, or how to prove this in this case.
$\begingroup$
$\endgroup$
4
-
1$\begingroup$ Is there a difference between $\|A\|$ and $\|A\|_2$, or are you just abusing notation? $\endgroup$– Thomas AndrewsNov 9, 2016 at 20:29
-
$\begingroup$ In fact: for any matrix $M$, we have $\|M^TM\| = \|M\|^2$ $\endgroup$– Ben GrossmannNov 9, 2016 at 20:31
-
$\begingroup$ $||M||_2$ is the largest singular value $\sigma_{\rm max}$ of $M = U\Sigma V^T$ (the svd of $M$) and the largest singular value of $M^T M = V \Sigma^2 V^T$ is $\sigma_{\rm max}^2$. $\endgroup$– TobiasNov 9, 2016 at 20:40
-
$\begingroup$ Thomas, just abusing notation. $\endgroup$– dpalmaNov 10, 2016 at 4:14
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
As Omnomnomnom already mentioned in his comment $\|LL^T\|_2 = \|L\|_2^2$ is true for any matrix $L\in\mathbb{R}^{n\times n}$:
When calculating the 2-norm \begin{align*} \|L\|_2 = \max\{\|Lx\|_2 : x\in\mathbb{R}^n, \|x\|_2 = 1\} \end{align*} of $L$ we can apply any norm-preserving change of coordinates $U, V$ in the image space and the pre-image space of $L$, respectively ($\|Lx\|_2 = \|U^T Lx\|_2 = \|U^T L V y\|_2$ with $x= Vy$, $\|x\|_2 = \|Vy\|_2 = \|y\|_2$).
Let $L = U\Sigma V^T$ be the singular value decomposition of $L$. (That means $\Sigma$ is a diagonal matrix $\Sigma=\operatorname{diag}(\sigma_1,\ldots,\sigma_r,0,\ldots,0)$ with $\sigma_{\rm max}:=\sigma_1\geq \ldots \geq \sigma_r > 0$ where $r\in\mathbb{N}$ is the rank of $L$ and $U,V$ are orthonormal and therefore norm-preserving.)
The norm-preservation of $U,V$ leads to \begin{align*} \|L\|_2 = \|\Sigma\|_2 = \sigma_{\rm max} \end{align*} which implies $\|LL^T\|_2 = \|L\|_2^2$: \begin{align*} \|LL^T\|_2 = \|U\Sigma V^T V \Sigma U^T\|_2 = \|\Sigma^2\|_2 = \sigma_{\rm max}^2 = \|L\|_2^2. \end{align*}
