0
$\begingroup$

Let $A = uv^{T} \in \mathbb{R}^{m \times n}$. Find the (condensed) SVD decomposition of $A$.

Theorem (Condensed SVD decomposition)

Let $A \in \mathbb{R}^{n \times m}$ be a non-zero matrix of rank $r$. Then there exists $\hat{U} \in \mathbb{R}^{n \times r}, \hat{\Sigma} \in \mathbb{R}^{r \times r} , \hat{V} \in \mathbb{R}^{r \times m}$ such that $\hat{U}$ and $\hat{V}$ are isometries , $\hat{\Sigma}$ is a diagonal matrix with main diagonal entries $\sigma_{1} \ge \dots \ge \sigma_r > 0$ and $$ A = \hat{U} \hat{\Sigma}\hat{V}^{T}.$$

Attempt at solution

Here the rank of $A$ is $1$ since each column is linearly dependent on $u$. So by the outer product form of the SVD decomposition , we get $$ A = \sum_{j=1}^{r} \sigma_j u_j v_{j}^{T} = \sigma_1 u_1 v_j^{T} = \hat{U}\hat{\Sigma}\hat{V}^{T}.$$ It follows from the theorem above that $$ A = \underbrace{\begin{pmatrix} u_1 \dots u_m\end{pmatrix}}_{\widehat{U}} \underbrace{\begin{pmatrix} \sigma_1 \end{pmatrix}}_{\widehat{\Sigma}} \underbrace{\begin{pmatrix} v_1 \\ \vdots \\ v_n\end{pmatrix}}_{\widehat{V}^{T}}.$$ To satisfy the initial equality ,we set $\sigma_1 =1$.

Could someone confirm my reasoning ? I feel like I am missing a step or something is wrong.

$\endgroup$
1
  • $\begingroup$ Vectors in the DVD are normalized. You'll adjust $\sigma$ to scale the product. $\endgroup$ Jun 27, 2021 at 0:35

1 Answer 1

1
$\begingroup$

When $r=1$, then $\hat{U}$ is a single column and $\hat{V}^\top$ is a single row. You might try $A=uv^\top = u \hat{\Sigma} v^\top$ where $\hat{\Sigma}$ is just the $1\times 1$ matrix $1$, but the problem is that $u$ and $v$ aren't isometries because their columns are not normalized.

However, there is an easy fix. The SVD of $A=uv^\top$ is $\hat{U} \hat{\Sigma} \hat{V}^\top$ where $\hat{U} = u/\|u\|$, $\hat{V} = v/\|v\|$, and the single entry of $\hat{\Sigma}$ is $\|u\|\|v\|$.

$\endgroup$

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.