I am looking for general solution of the form $B = XAX^{-1}$ with the following constraints


where $X$ is unknown matrix and $A,\,B,\, X$, are $3\times 3$ matrices.

The problem appears peculiar and thus I wish to simplify it further to more practical problem X is rotation matrix ($R$), and $A$ is plane equation matrix, B is resulted homography in the sense of $B =XAX^{-1} = XAX^{T}= R(I-tn^T/d)R^{T}$ ,

where $t$ is translation vector, $n$ is plane normal vector, and $d$ is signed distance, all three are known

My dirty solution so far is as follows : Finding the Jordan Canonical form:


and then $X$ is obviously

$X = N^{-1}M$

but this works only if the canonical forms are the same, otherwise cannot be utilized for solution.

The second dirty trick is to solve a system of Linear equations in terms of

$AX = BX$

but i cannot figure out how to program this in the software (e.g. Matlab linsolve). Could you help me to clarify how to write system of linear equations in the shape of 9 rows?

The trick is how to notate them on the "right side" to use in ready to use functions for solving SLE.

Thank, you

  • $\begingroup$ well, let's simplify: X is always invertible, orthogonal and X' = X^{1} and is also 3x3 matrix $\endgroup$ – Misha Bolgarskiy Nov 14 '18 at 13:32
  • $\begingroup$ Are $A$ and $B$ symmetric then ? $\endgroup$ – nicomezi Nov 14 '18 at 13:32
  • $\begingroup$ This will not work. Take $A=0$ and $B=I$, the identity. $\endgroup$ – Dietrich Burde Nov 14 '18 at 13:33
  • $\begingroup$ No, A and B are not symmetric. For the symmetric case, there is already a valuable thread. Even I can generalize a solution of the kind MAM' = J = NBN' (finding the jordan canonical form) - if the canonical form is same, then there is solution, if not - no solution, then X = N^{-1}M $\endgroup$ – Misha Bolgarskiy Nov 14 '18 at 13:38

Case 1. $A,B$ are not orthogonally similar; no solutions.

Case 2. $A,B$ are orthogonally similar; $A=PBP^T$, where $P\in O(n)$. Let $C(A)=\{X\in M_n(\mathbb{R});AX=XA\}$. Then, it is not difficult to show

$\textbf{Proposition}$ $\{X\in O(n);A=XBX^T\}=(C(A)\cap O(n))P$.

It remains to obtain $C(A)\cap O(n)$. When $n=3$, $C(A)$ is a vector space of dimension $3,5$ or $9$ (when $A$ is a scalar matrix).

Generically (randomly choose $A$), $A$ is cyclic


and $C(A)=\{aI+bA+cA^2;a,b,c\in \mathbb{R}\}$. Generically, the eigenspaces of $A$ are not orthogonal and, consequently, $C(A)\cap O(n)=\pm I_n$ and $X=\pm P$.

EDIT. $A,B$ are orthogonally similar iff the couples $(A,A^T)$ and $(B,B^T)$ are similar. Then, practically, one studies the linear system

$(S)$ $AX=XB,A^TX=XB^T$.

If $A,B$ are not orthog. similar, then the sole solution is $X=0$.

If $A,B$ are orthog. similar, then, for a generic $A$, $(S)$ admits a vector space of solutions of dimension $1$: $X=uX_0$. It remains to calculate $u$ so that $||uX_0||_2=1$.

  • $\begingroup$ thank you, I need time to digest this.. $\endgroup$ – Misha Bolgarskiy Nov 14 '18 at 19:03
  • $\begingroup$ you are somehow, point towards similarity transformations..? $\endgroup$ – Misha Bolgarskiy Nov 14 '18 at 19:04
  • $\begingroup$ Maybe I could simplify further the problem, A is plane equation matrix in the shape of $i-pn/h$ and $X$ is a rotation matrix. $\endgroup$ – Misha Bolgarskiy Nov 15 '18 at 1:38
  • $\begingroup$ @Misha Bolgarskiy , If you want to change the problem, then open a new file. Anyway, I did the job; now it's your business. $\endgroup$ – loup blanc Nov 15 '18 at 11:08
  • $\begingroup$ I am really sorry Mr. Blanc, i am pretty newbie here $\endgroup$ – Misha Bolgarskiy Nov 15 '18 at 17:03

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.