Finding a particular low-rank subspace via optimization

73 Views Asked by At

I have two [$n \times k$] matrices $X$ and $Y$, and a symmetric [$m \times m$] matrix $B$.

At a high-level, my goal is to find a low-rank subspace, $A$, in which $X$ and $Y$ are related by the linear mapping $B$. To find $A$, I've come up with the following optimization problem:

$$ \min_{A \in \mathbb{R}^{k \times m}\mbox{, }A^TA = I} \| YA - XAB \|_F^2 $$

where $\|C\|_F^2 = \sum_{i,j} (C_{i,j})^2$ is the Frobenius norm.

I have a couple of questions:

  • Is this problem tractable? Any ideas for how I might decompose this somehow into a more convenient form?

  • Is this objective convex? (I realize that my domain, the set of rank-m projection matrices, isn't a convex set. But I'm curious to know if the objective function itself is convex.)

Thanks for any advice!