2
$\begingroup$

I have a large matrix $A \in \mathbb{R}^{n \times m}$ and would like to subtract a sparse matrix $B \in \mathbb{R}^{n \times m}$ with less than $c (n+m)$ non-zero entries, where $c > 0$ is a constant one is free to choose, such that the singular values of $A-B$ decay as quickly as possible.

I have no idea how to approach this problem. Any thoughts?

$\endgroup$
2
  • 1
    $\begingroup$ How would you quantify "as quickly as possible"? The singular values are a just a finite tuple, so they are bounded by any $Ki^{-s}$ or $Kq^i$… $\endgroup$ Commented Jun 11, 2018 at 19:04
  • $\begingroup$ Good question: I think a reasonable choice would be to fix a number $r\leq \min\{m,n\}$ and try to minimize $\sum_{i=r+1}^{\min\{m,n\}} |\sigma_i|^2$, where $\sigma_i$ are the singular values of $A-B$ in decreasing order $\endgroup$ Commented Jun 12, 2018 at 10:13

1 Answer 1

1
$\begingroup$

Maximizing the "decay" of the singular values could be thought of as minimizing the (numerical) rank. Hence, I believe that the original problem could be rephrased as follows:

Given $\mathrm A \in \mathbb R^{m \times n}$, find a sparse matrix $\mathrm X \in \mathbb R^{m \times n}$ such that $\mbox{rank} (\mathrm X - \mathrm A)$ is minimized.

which is hard. Convex proxies$^\dagger$ for sparsity and rank are the entry-wise 1-norm and the nuclear norm, respectively. Hence, relaxing the original problem, we obtain the following convex program

$$\begin{array}{ll} \underset{\mathrm X \in \mathbb R^{m \times n}}{\text{minimize}} & \| \mathrm X \|_1 + \gamma \| \mathrm X - \mathrm A \|_*\\ \end{array}$$

where $\gamma > 0$. Varying parameter $\gamma$, if we find a matrix $\rm X$ that is sufficiently sparse, we are done.


$\dagger$ Maryam Fazel, Recovering simultaneously structured objects, Simons Institute, September 2014.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.