0
$\begingroup$

I'm given a space of $X \in R^{mxn}$ with inner product $\langle X_1, X_2 \rangle = tr(X_1^T X_2)$ and another space of random vectors $Y \in RV^m$ with inner product $\langle y_1,y_2 \rangle = E[y_1^Ty_2]$. Additionally an operator that maps between these spaces $A: X \rightarrow Y $ is known to be $A(X) = Xb(\omega)$. Where $b \in RV^n$ and $E[bb^T] = I$. I believe the last identity is only necessary in proving that $A(X)$ is one to one.

The goal is to solve the least squares problem for the case where A is one to one. Therefore I know it's necessary to determine the adjoint operator, $A^* : Y \rightarrow X$, which satisfies the equation $$ \langle y,A(X)\rangle = \langle A^*(y),X \rangle$$

But with different definitions for the inner products of the two spaces, I am stuck in manipulating the expectation into a function with the trace. I have...

$$ \langle y,A(X)\rangle$$ $$E[y^TXb(\omega)]$$ $$.$$ $$.$$ $$.$$ $$tr((something)^TX)$$ $$\langle A^*(y),X \rangle$$

If anyone has any references on dealing with such a combination of different inner products in determining the adjoint operator, especially in a stochastic framework, I would be very nice to see. General literature with examples of adjoint determination would be greatly appreciated as well.

$\endgroup$
1
$\begingroup$

Using the cyclic property of trace and linearity of expectation/trace we have: \begin{eqnarray} \langle y, A X \rangle &=& E [ y^T X b ] \\ &=& E [ \operatorname{tr} ( y^T X b )] \\ &=& E [ \operatorname{tr} ( b y^T X )] \\ &=& \operatorname{tr} ( E [ b y^T ] X ) \\ &=& \operatorname{tr} (E[y b^T]^T X ) \\ &=& \langle E[y b^T] , X \rangle \end{eqnarray} Hence $A^*y = E[y b^T]$.

$\endgroup$
  • $\begingroup$ Of course... because $y^TXb$ is scalar valued, and $X$ is itself not a random vector/matrix. It just had to be something this obvious. :) $\endgroup$ – Falimond Nov 8 '15 at 5:22
  • $\begingroup$ Well, I would say that it is elementary but not obvious :-). Many things become obvious in retrospect. $\endgroup$ – copper.hat Nov 8 '15 at 5:26
  • $\begingroup$ Yes, hindsight is 20/20. Oh how convenient that inner products be scalar valued by definition! $\endgroup$ – Falimond Nov 8 '15 at 5:39

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.