Can the matrix transpose be represented by $X^T = AXB$ for a given $A$ and $B$?

I think it is possible, but please correct me if I am wrong--please see my attempt below.

My thinking (not sure whether I am on the right path) is that one could employ $\mbox{vec}$ and Kronecker product operator. However, I am not sure whether this formally proves it, does it?

If I vectorize both sides, i.e., \begin{align} X^T &= AXB \\ \Downarrow \\ \mbox{vec}\left( X^T \right) &= \mbox{vec}\left( AXB \right) \\ \mbox{vec}\left( X^T \right) &= \underbrace{\left(B^T \otimes A\right)}_{\equiv \ \ T_{m,n}}\mbox{vec}\left( X \right) \\ \mbox{vec}\left( X^T \right) &= T_{m,n}\mbox{vec}\left( X \right), \\ \end{align} where $T_{m,n}$ is some permutation matrix (cf. http://www4.ncsu.edu/~pfackler/MatCalc.pdf).

Similar question: Solve $AXB=X^\top$

  • 2
    $\begingroup$ An interesting approach. Now the question is, can you find a decomposition of the form $$T_{m,n} = B^T\otimes A$$ If you can, then it would support your hypothesis. Of course, you can always find a Kronecker decomposition of the form $$T_{m,n} = \sum_{k=1}^{r}B^T_k\otimes A_k$$ but it is very rare to find a decomposition with $r=1$. $\endgroup$ – greg May 23 '18 at 8:19

No, there are no such matrices $A$ and $B$.

If there were, we would have $\operatorname{Id}=\operatorname{Id}^T=A\operatorname{Id}B=AB$ and therefore $B=A^{-1}$.

Now, is there an invertible $2\times2$ matrix $A=\left(\begin{smallmatrix}a&b\\c&d\end{smallmatrix}\right)$ such that$$\begin{pmatrix}0&1\\0&0\end{pmatrix}=A\begin{pmatrix}0&0\\1&0\end{pmatrix}A^{-1}?$$It is easy to see that the only such matrices are those of the type $\left(\begin{smallmatrix}0&1/c\\c&0\end{smallmatrix}\right)$.



I partially answer this question based on your thoughts and the comment given by @greg .

$\mbox{vec}(X^T)=T_{m,n} \mbox{vec}(X)$

Note that $T_{m,n}$ can be written as the following way (cf. https://en.wikipedia.org/wiki/Commutation_matrix $K^{(r,m)}$)

$T_{m,n}=\displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1}J^{ij}_{mn}\otimes (J^{ij}_{mn})^T$, where $(J^{ij}_{mn})^T=J^{ji}_{nm}$.

$J^{ij}_{mn}$ is the single-entry matrix (cf. https://en.wikipedia.org/wiki/Single-entry_matrix), with the matrix size $m\times n$. Hence,

\begin{align} \mbox{vec}(X^T) &= \left(\displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1}J^{ij}_{mn}\otimes (J^{ij}_{mn})^T\right) \mbox{vec}(X)\\ &= \displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1}\left(J^{ij}_{mn}\otimes J^{ji}_{nm}\right) \mbox{vec}(X)\\ &= \displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1}\mbox{vec}\left( J^{ji}_{nm} X (J^{ij}_{mn})^T \right)\\ &= \mbox{vec}\left( \displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1} J^{ji}_{nm} X J^{ji}_{nm} \right) \end{align}

Therefore, we have

\begin{align} X^T= \displaystyle\sum^m_{i=1}\displaystyle\sum^n_{j=1} J^{ji}_{nm} X J^{ji}_{nm} \end{align}

You can obtain this relationship via directly using the definition of single-entry matrix as well. And if you could do some tranformation to remove two summations, then we can completely answer your question. Hope it helps.


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.