4
$\begingroup$

I am trying to understand singular value decomposition. I get the general definition and how to solve for the singular values of form the SVD of a given matrix however, I came across the following problem and realized that I did not fully understand how SVD works:

Let $0\ne u\in \mathbb{R}^{m}$. Determine an SVD for the matrix $uu^{*}$.

I understand that $uu^{*}$ has rank 1 and thus only has one singular value i.e.

$$\Sigma = \begin{pmatrix} \sigma_1 & \ldots & 0\\ 0& \ldots & 0 \end{pmatrix} \in\mathbb{R}^{m\times m}$$

and I realize since $uu^{*}\in\mathbb{R}^{m\times m}$ then for $uu^{*}=U\sum V^{*}$ that $U,\Sigma,V\in\mathbb{R}^{m\times m}$. Additionally, I realize that the columns of $U$ and $V$ are orthonormal.

I guess my question is how do you determine U and V from $uu^{*}$?

$\endgroup$
5
$\begingroup$

The short SVD of $uu^*$ is $v\sigma v^*$ with $v=\frac{u}{\|u\|}$ and $σ=\|u\|^2$.

If you compute the bisector $w=u+\frac{u_1+0}{|u_1|+0}\|u\|e_1$ of $u$ and $e_1$, then you can get a full unitary matrix $V$ as the reflection matrix $I-2\frac{ww^*}{\|w\|^2}$, so that $uu^*=V\Sigma V^*$

$\endgroup$
  • $\begingroup$ How did you get that $\sigma=\|u\|^{2}$? $\endgroup$ – user133541 Mar 8 '14 at 15:48
  • $\begingroup$ Because $v$ has norm $1$ and $uu^*=vσv^*$ has to be an identity. Note that this works in general for $ab^*=uσv^*$ with $u=\frac{a}{\|a\|}$, $v=\frac{b}{\|b\|}$ and $σ=\|a\|\,\|b\|$. $\endgroup$ – LutzL Mar 8 '14 at 15:52
  • $\begingroup$ So, this implies that the singular value for a rank 1 matrix given by $ab^{*}$ is always $\sigma =\|a\|\|b\|$? $\endgroup$ – user133541 Mar 8 '14 at 15:56
  • $\begingroup$ Yes, there is only one positive singular value and that has to coincide with the operator norm $\|ab^*\|_{op}=\sup_{\|x\|=1}\|ab^*x\|_2=\sup_{\|x\|=1}\sqrt{x^*ba^*ab^*x}=\|a\| \ \sup_{\|x\|=1}|b^*x|$... $\endgroup$ – LutzL Mar 8 '14 at 16:00
  • $\begingroup$ Oh. Okay. I completely get it now. Thank you so much! $\endgroup$ – user133541 Mar 8 '14 at 16:14

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.