It is known that given corrupt measurements $y = Af+e \in \mathbb{R}^m$ with $f \in \mathbb{R}^n$ and $\|f\|_0 < m < n$, one can recover $f$ exactly by solving a convex optimization problem. What if $f$ is instead a square matrix? Can we recover a matrix from corrupt measurements instead of just a vector?
$\begingroup$
$\endgroup$
3
-
9$\begingroup$ Isn't a matrix just a special case of a vector? That is, can't you "flatten" an $n\times n$ matrix into a vector in $\mathbb{R}^{n^2}$ and then solve the problem by the usual means? $\endgroup$– Noah SteinCommented Nov 23, 2011 at 3:00
-
1$\begingroup$ I'm not sure I agree with the assertion in the first sentence. It seems to depend strongly on the nature of corruption, and on the precision of measurements. $\endgroup$– S. Carnahan ♦Commented Nov 23, 2011 at 5:31
-
$\begingroup$ See epubs.siam.org/doi/abs/10.1137/070697835 $\endgroup$– rodmsCommented Jun 5, 2014 at 21:19
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
First, I emphasize S. Carnahan's comment: Exact recovery from noisy measurements is not that simple. "Exact recovery" usually means "recovery of the exact support of $f$". Moreover, sparsity assumptions for $f$ and special assumptions for $A$ and the size of $e$ are needed.
To address your question: This again depends on a lot of things. Of course you can view this as $n$ multiple instances of the original problem and basically use the previous theory. Other structural assumption lead to other results (e.g. having a "joint sparsity pattern in the columns of $f$"). Buzzwords here are "joint sparsity" or "multiple measurement vectors".
