6
$\begingroup$

Exact low rank matrix completion using nuclear norm minimization can be formulated as a semidefinite program (SDP). Following the notation in the paper, a convex problem for noisy matrix completion can be:

$$\text{minimize} \,\, \|X\|_* \quad \text{subject to} \quad \|\mathcal{P}_\Omega(X)-\mathcal{P}_\Omega(M)\|_2\le\epsilon \tag{*}$$

where $\mathcal{P}_\Omega(X)$ is the projection of $X$ onto the set of observed entries $\Omega$. This problem can be formulated as the SDP below SDP in $X, W_1, W_2$

$$\begin{array}{rl} \text{minimize} & \text{trace}(W_1)+\text{trace}(W_2)\\ \text{subject to} & \left[ \begin{array}{cc} \epsilon I & \mathcal{P}_\Omega(X-M) \\ \mathcal{P}_\Omega(X-M)^T & \epsilon I \end{array} \right] \succeq 0\\ & \left[ \begin{array}{cc} W_1 & X \\ X^T & W_2 \end{array} \right] \succeq 0\end{array}$$

The question is when the $2$-norm constraint of problem (*) is replaced by the Frobenius norm constraint, i.e.,

$$\text{minimize} \,\, \|X\|_* \quad \text{subject to} \quad \|\mathcal{P}_\Omega(X)-\mathcal{P}_\Omega(M)\|_F \le \epsilon$$

it can still be cast as a SDP as stated in here and many other papers. Could anyone tell me how to formulate this problem as a SDP?


Edit: follow up question.

$\endgroup$

1 Answer 1

8
$\begingroup$

The Frobenius norm does not cause a problem. Remember, the Frobenius norm of a matrix $X$ is actually nothing more than the 2-norm of the vector formed by stacking the columns of $X$ on top of each other. And for vectors, the vector 2-norm and the matrix 2-norm coincide. So if $\mathcal{Q}$ implements this "stacking" isomorphism, we have $$\|\mathcal{P}_\Omega(X-M)\|_F=\|\mathcal{Q}(\mathcal{P}_\Omega(P-M))\|_2$$ and $$\|\mathcal{P}_\Omega(X-M)\|_F\leq\epsilon \quad\Longleftrightarrow\quad \begin{bmatrix} \epsilon I & \mathcal{Q}(\mathcal{P}_\Omega(X-M)) \\ \mathcal{Q}(\mathcal{P}_\Omega(X-M))^T & \epsilon \end{bmatrix} \succeq 0$$ In fact, handling the Frobenius norm is actually somewhat easier than the spectral norm in practice, because most solvers handle vector 2-norms more directly and efficiently. See second-order cone programming or "semidefinite quadratic linear programming" (SQLP) for more background.

$\endgroup$
1

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.