Least square problem constrained to projection matrices

241 Views Asked by At

Some times in engineering, it is important to find an optimum subspace in which projecting on it satisfies some properties.

Let known matrices $A$ and $B$ belong to $\mathbb{R}^{p\times n}$ and $\|\cdot\|_F$ be Frobenius norm. How can I find the best subspace in which projecting $A$ on it is as close as possible to $B$? In other words how can I find a solution to the following constrained optimization problem? \begin{eqnarray} &&\min_P \|PA-B\|_F^2 \\ &&\mathrm{s.t. \ }P^T=P, \, P^2=P \end{eqnarray}


Update:

I have incorporated the symmetry property ($P^T=T$) in objective function as follows. Since $P$ is symmetric, $P$ can decompose as $P=Y+Y^T$ where $Y\in \mathbb{R}^{n\times n}$. Now, the optimization problem reduces to \begin{eqnarray} &&\min_Y \|C(Y+Y^T)-D\|_F^2 \\ &&\mathrm{s.t. \ }\, (Y+Y^T)^2=Y+Y^T, \end{eqnarray} Where $C=A^T$ and $D=B^T$.

Further, using "vec" operator, we get vec$(Y^T)=$$K$vec$(Y)$, where $K\in \mathbb{R}^{n\times n}$ is a unique and known matrix. Using "Kronecker product", our optimization problem will be reduced to \begin{eqnarray} &&\min_y \|(I\otimes C)(I+K)y-d\|_2^2 \\ &&\mathrm{s.t. \ }\, (Y+Y^T)^2=Y+Y^T, \end{eqnarray} where $y=$ vec$(Y)$ and $d=$ vec$(D)$.