0
$\begingroup$

Let $A\in\mathbb{R}^{n\times n}$ and $B = UDV^T$ where $U\in\mathbb{R}^{n\times r}$, $V\in\mathbb{R}^{n\times r}, D\in\mathbb{R}^{r\times r}$ and $U^TU = V^TV = I_r$. Define two sets: $$S_1=\mathbb{R}^{r\times r}$$ $$S_2 = \{\text{All $r\times r$ diagonal matrices}\}\subset S_1$$ and $$D_1=\arg\min_{M\in S_1}\|UMV^T - A\|_F^2$$ $$D_2=\arg\min_{M\in S_2}\|UMV^T - A\|_F^2$$ where $\|\|_F$ is the Frobenius norm.

My questions are:

  1. Is this identity true? \begin{align*} \|B-A\|_F^2&=\|B - UD_1V^T\|_F^2+\|UD_1V^T-A\|_F^2\\ &=\|B - UD_2V^T\|_F^2+\|UD_2V^T-A\|_F^2 \end{align*}

  2. Can we express $\|UD_1V^T-A\|_F^2$ and $\|UD_2V^T-A\|_F^2$ explicitly as functions of $U,V,A$?

$\endgroup$
  • $\begingroup$ Is $D$ (defining $B$) a diagonal matrix as well? $\endgroup$ – user7440 Aug 14 at 3:42
0
$\begingroup$
  1. Note that the Frobenius norm $$ || X + Y ||_{F}^2 = tr\left( (X+Y)^T (X+Y) \right) = || X ||_{F}^2 + 2 tr\left( X^T Y \right) + || Y ||_{F}^2 $$ The minimization (defining $D_1$) yields that $$ tr\left( (UD_1 V^T - A)^T (U M V^T ) \right) = 0 $$ for any matrix $M$ in $S_1$. Next write $$ ‖𝐵−𝐴‖_F^2 = ||B - UD_1 V^T ||_F^2 + 2 tr\left( (UD_1 V^T - A)^T (B - UD_1 V^T ) \right) + || U D_1 V^T - A ||_F^2 $$ The middle term is $0$ because $B - UD_1 V^T = U( D - D_1 )V^T$.
  2. Write $$ || U D V^T - A ||_{F}^2 = || U ( D - U^T A V ) V^T ||_{F}^2 = || D - U^T A V ||_{F}^2 $$ So we get $D_1 = U^T A V$ and $D_2$ is the diagonal part of $U^T A V$.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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