3
$\begingroup$

Focusing on following least squares problem:

$$\min\limits_{V} \lVert Z - WV \rVert _{_F}^2$$

$$Z∈{R}^{m\times n},\quad W∈{R}^{m\times k},\quad V∈{R}^{k\times n},\quad k\lt m\lt n $$

This problem can be easily solved given $Z$ and $W$. I assume the solution of this problem as $\overline{V}$.

My question is that, how can we recover $W$ by $\overline{V}$ and $Z$?

I found something called inverse optimization, but cannot figure out the relationship between them.

$\endgroup$
1
$\begingroup$

The solution to the initial least squares problem can be written in terms of the pseudoinverse $${\bar V} = W^+Z$$ Then several days later, you still have $({\bar V},Z)$ but have somehow lost the definition of $W$.
You can try the following $$\eqalign{ W^+ &= {\bar V}Z^+ \cr W &= ({\bar V}Z^+)^+ \cr }$$

 

$\endgroup$
  • $\begingroup$ Thanks for your answer. In your answer you assume $${W^+} = W^+ZZ^+$$ I am not sure if it is true. $\endgroup$ – lee Jun 6 '18 at 1:18
  • $\begingroup$ It'll be true so long as $\,{\rm rank}(Z) = m\,\,$ $\endgroup$ – greg Jun 6 '18 at 4:06
  • $\begingroup$ Thank you. @greg Your answer is very valuable. What if rank($Z$) <$m$? $\endgroup$ – lee Jun 6 '18 at 9:44
  • $\begingroup$ @greg are you sure? $(e_1)^+ = e_1^T$, but $w^T = w^Te_1e_1^T$ does not hold for all vectors $w$. $\endgroup$ – Federico Poloni Jun 6 '18 at 16:20
  • $\begingroup$ @FedericoPoloni The question stipulates that $Z\in{\mathbb R}^{m\times n}$ and $m<n$. Your counter-example is $e_1\in{\mathbb R}^{m\times 1}$, which does not satisfy the second criteria since $m>1\,\,$ $\endgroup$ – greg Jun 6 '18 at 18:11

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.