Uniqueness of k-sparse solution in compressed sensing

245 Views Asked by At

This question has been bothering me for some time now. Please correct me if I've got anything wrong.

In compressed sensing, we try to find a unique k-sparse solution to an underdetermined system by using an RIP-satisfying measurement matrix $\Phi$ and $l_{1}$ minimization.

But what if the system doesn't have a unique k-sparse solution? I'm sure everybody familiar with this has seen the figure with the $l_{1}$ and $l_{2}$ ball showing why $l_{1}$ recovery works. Even in the $\mathbb{R}^{2}$ case, there are two 1-sparse solutions: at the point of intersection of $\Phi x=u$ with the axes. In general, $\Phi x=u$ will always be this hyperplane that intersects the coordinate axes at some point. So there are always multiple 1-sparse solutions. How is it that all these 1-sparse points give the same measurements? What difference would an RIP-satisfying $\Phi$ make in this picture? And why do we want the $l_{1}$ minimization to give us a unique k-sparse solution when the original problem does not have one?