In my research I came across the following problem. Let $A$ be a symmetric and $\Gamma$ be a diagonal $n\times n$ matrices. The eigenvalues of $A$ are known $\lambda_1,\ldots\lambda_n$. The traces $\mathrm{Tr}(A^k\,\Gamma)=t_k$, $k\in\mathbb{N}$ are also known, $\Gamma$ is given. Can $A$ be found based on this information? If yes, how? While I am curious about the general $n$ case, information on $n=3$ would be most valuable. Any reference would be greatly appreciated.


  • From the answer of Francesco Polizzi it is known that the reconstruction is not possible when there is an orthogonal matrix $M$ that commutes with $\Gamma$. Fortunately, in the case of interest, this situation can be excluded. Specifically, it is known that diagonal entries of $\Gamma$ are all distinct and positive. By this answer, $\Gamma$ then commutes only with diagonal matrices.
  • user44191 suggested that $B=M^{-1}AM$ ($M$-nonsingular) will have the same traces as $A$. However, apart from a trivial $\pm1$ possibility, which can be excluded by a proper sign convention, the new matrix $B$ is no longer symmetric. Thus, it leaves the question open.
  • 4
    $\begingroup$ It is not true for $\Gamma=I$. Since $Tr(A^k)=\sum\limits_{i=1}^n \lambda_{i}^k$, no additional information can be retrieved besides the eigenvalues (which are already given). $\endgroup$ – Josiah Park Apr 30 at 8:45
  • 1
    $\begingroup$ If $B$ is a nonsingular matrix that commutes with $\Gamma$, then $B^{-1}AB$ will have the same traces; assuming that $\Gamma$ has distinct eigenvalues, these $B$ are arbitrary diagonal matrices. It may still be possible to determine the $C_G(\Gamma)$-conjugacy class $A$ belongs to in $\mathbb{M}_n$ (though I haven't checked fully). $\endgroup$ – user44191 Apr 30 at 8:45
  • $\begingroup$ @user44191 Can you please elaborate about the conjugacy classes? $\endgroup$ – yarchik Apr 30 at 8:57
  • $\begingroup$ @JosiahPark Indeed $\Gamma=I$ is somewhat a special case. The problem may also be ill posed for $\Gamma$ in the vicinity of $I$. It is my hope, however, solutions may be found for some $\Gamma$. $\endgroup$ – yarchik Apr 30 at 9:01
  • $\begingroup$ @yarchik Allowing for multiple $\Gamma$'s allows for a positive answer. For instance for $n=3$ if one takes $\Gamma_{i}=e_{i}e_{i}^{T}$, $i=1,\dots,3$ one can recover $A$. $\endgroup$ – Josiah Park Apr 30 at 9:02

Unfortunately even the generic situation is bad. Since we know the eigenvalues, we should search for the orthonormal system of eigenvectors $v_i$ of $A$. We have ($e_i$ is the standard basis) $$ Tr(A^k\Gamma)=\sum_i\left[\sum_j\gamma_j \langle v_i,e_j\rangle^2\right]\lambda_i^k $$ so, in effect, you have the knowledge of $\sum_j\gamma_j \langle v_i,e_j\rangle^2$ ($i=1,2,3$). However, these three values are not independent: since the matrix $(\langle v_i,e_j\rangle^2)_{ij}$ is bistochastic, their sum is just $Tr \Gamma$, so you have only $2$ independent equations, while the orthogonal matrices form a 3D manifold. Thus in the generic case you should expect a continuous 1-parametric family of solutions.

  • 1
    $\begingroup$ Yes, this fully answers my question. Just a short clarification question, do you have any insight on what kind of 1-parameter freedom would that be? If we talk about the orthogonal matrices realizing a rotation in 3d space, would it be possible to say something like, the axis of rotation can be determined, but not the angle? $\endgroup$ – yarchik Apr 30 at 14:13
  • $\begingroup$ One may expect a 1-parametric family of solutions, however, that is because you allow them to be complex. $A$ is symmetric and, therefore, eigenvectors should be real. Imposing this additional condition may lead to a finite number of solutions in some cases. Do you agree with this? $\endgroup$ – yarchik May 1 at 9:46
  • $\begingroup$ @yarchik No. I did purely real analysis. Nothing was complex. The real orthogonal matrices are a 3D manifold and the 2 equations were real as well. $\endgroup$ – fedja May 1 at 11:18

I do not think you can reconstruct $A$ just from this information.

Take $\Gamma=I_n$, let $M$ be any orthogonal $n \times n$ matrix and set $B={}^tMA M$.

Then $A$ and $B$ are two similar (symmetric) matrices, and so all their positive powers $A^k$ and $B^k$ have the same eigenvalues (and in particular the same trace).

  • $\begingroup$ $\Gamma=I_n$ is somewhat special. What if we do not assume from onset that gamma is the identity matrix? $\endgroup$ – yarchik Apr 30 at 8:54
  • 1
    $\begingroup$ In the general case, take as $M$ a (non-identical) orthogonal matrix commuting with $\Gamma$. Then $${}^tM(A^k \Gamma) M = B^k ({}^tM \Gamma M) = B^k \Gamma,$$ so $A^k \Gamma$ and $B^k \Gamma$ are similar and the same argument applies. $\endgroup$ – Francesco Polizzi Apr 30 at 9:06
  • $\begingroup$ Can one always find a (non-identical) orthogonal matrix $M$ that commutes with $\Gamma$? For me it is not obvious, scaling and rotation typically do not commute. $\endgroup$ – yarchik Apr 30 at 9:33
  • 1
    $\begingroup$ Ok, let's say that if the centralizer of $\Gamma$ contains a (non-identical) orthogonal matrix, then the reconstruction of $A$ is not possible. I did not check whether this is always the case for all diagonal matrices. $\endgroup$ – Francesco Polizzi Apr 30 at 9:42
  • 1
    $\begingroup$ @yarchik A diagonal matrix with only $\pm1$ entries is orthogonal and commutes with any diagonal matrix. For a diagonal matrix with only distinct entries, I suspect that these are the only orthogonal matrices that commute. $\endgroup$ – M. Winter Apr 30 at 12:03

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.