2
$\begingroup$

In sparse optimization, I am trying to solve the problem $$ \min_{x\in \mathbb R^{n}} \quad f(x) + \|x\|_1 $$ and at optimality, $x^*$ may be sparse. If I define the sparse manifold as $\mathcal M = \{z : z_i = 0 \text{ whenever } x_i^* = 0\}$, there exist methods in which, for certain cases of $f$, can identify the sparse manifold; e.g. $x^{(k)} \to x^*$, there exist no finite $k_0$ for which $x^{(k)} = x^*$, but there does exist a finite $k_0$ for which, for all $k > k_0$, $x^{(k)} \in \mathcal M$.

Now is it possible to extend this to low-rank matrices? Specifically, now I am trying to solve some problem $$ \min_{X\in \mathbb R^{m\times n}} \quad f(X) + \|X\|_* $$ where $\|X\|_* = $ the sum of the singular values of $X$, and at optimality, $\text{rank}(X) = r< \min\{m,n\}$. Then I define the low rank manifold as $$ \mathcal M = \left\{\sum_{i=1}^r s_i u_iv_i^T: s_i \in \mathbb R\right\} $$ where $X^* = \sum_{i=1}^r \sigma_i u_i v_i^T$ is the ``economic" SVD of $X^*$.

Are there any existing methods in which $X^{(k)} \to X^*$, and for some finite $k_0$, for all $k \geq k_0$, $X^{(k)}\in \mathcal M$?

$\endgroup$

1 Answer 1

1
$\begingroup$

Yes. You can see results like this in the thesis of Jingwei Liang titled "Convergence Rates of First-Order Operator Splitting Methods" where they are called finite identification results (ch. 6, 7, 8).

The nuclear norm is a partly smooth function with respect to the manifold you've identified; the thesis provides several manifold identification results for different algorithms applied to problems involving this class of functions.

$\endgroup$
1
  • $\begingroup$ Thanks, actually that is a great line of work. I believe you are right, though I suppose their followup work also contains some similar goodies. $\endgroup$
    – Y. S.
    Commented Feb 21, 2022 at 16:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.