Suppose $A$ and $B$ are conjugate invertible real $n \times n$-matrices. Does there always exist a path from $A$ to $B$ inside their conjugacy class?

I thought I had an easy proof for odd $n$ which goes as follows, but it was incorrect as pointed out in this answer. To show where my misunderstanding arised, here is the wrong argument.

Suppose there exists a real matrix $P$ such that $B = PAP^{-1}$. By replacing $P$ with $-P$ if necessary, we can assume that $\det P > 0$ (this is what goes wrong in even dimensions, see this question). Then we have that $P = e^Q$ for some real matrix $Q$ (since the image of the exponential map is the path-component of the identity in $\text{GL}_n(\mathbb R)$). But now the path $$t \mapsto e^{tQ}Ae^{-tQ}$$ is a path connecting $A$ to $PAP^{-1} = B$.


This edit comes a bit late, but the way I see it is a bit different than the other answers so I'll write it anyway.

Again, I use the example from the question that you link to: $A$ is the rotation by $\frac \pi 2$ in the euclidean plane, and $B$ is the rotation by $-\frac \pi 2$.

Now any conjugate of $A$ by a matrix with positive determinant will correspond to a linear map $\varphi$ such that for any non-zero vector $v$, the vectors $v$ and $\varphi (v)$ in this order make a positive basis of the plane.

Conversely, a conjugate of $A$ by a matrix with negative determinant (which is the case of $B$) will correspond to a linear map $\varphi$ such that for any non-zero vector $v$, the vectors $v$ and $\varphi (v)$ in this order make a negative basis of the plane.

A path from $A$ to $B$ has to cross the set of matrices that have real eigenvalues, and such a matrix cannot be conjugate to $A$.

  • $\begingroup$ I like this geometric approach, this is the solution I will not forget. $\endgroup$ – Levi Nov 4 '19 at 0:50
  • $\begingroup$ @Levi Thanks :) $\endgroup$ – Arnaud Mortier Nov 4 '19 at 0:51

We can use the counterexample from your other question to answer this one. Let $A=\begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}$. Matrices $\begin{pmatrix} a & b \\ c & d \end{pmatrix}$ must have trace $a+d$ equal to $0$ and determinant $ad-bc=-a^2-bc$ equal to $1$. This implies that such a matrix cannot have $b=0$, since $-a^2\leq 0$. Therefore the set of conjugates of $A$ is disjoint union of open sets defined by $b>0$ and $b<0$. Neither of those sets is empty, since one contains $A$ and the other contains $-A$. Thus the conjugacy class of $A$ is disconnected.

  • $\begingroup$ This is much nicer than my argument! (But slightly less detailed, whence I'm leaving mine up.) $\endgroup$ – darij grinberg Nov 4 '19 at 0:30
  • $\begingroup$ This is an awesome way of seeing this. This particular example also shows that I have been trying to prove the wrong thing when answering the other question. So I am particularly grateful! $\endgroup$ – Levi Nov 4 '19 at 0:33
  • 1
    $\begingroup$ Sorry for unaccepting your answer, but I cannot resist going with coordinate-freeness :) $\endgroup$ – Levi Nov 4 '19 at 0:50

Other people have given counterexamples, so I would like to demonstrate that counterexamples are somewhat rare. Here is an attempt at a conceptual explanation for both why this is not true in general and also when it is true. First, we note that if we have a path $P(t)$ in $GL_n(\mathbb R)$ with $P(0)=P_0$ and $P(1)=P_1$, then $P(t)AP(t)^{-1}$ gives us a path inside the conjugacy class of $A$. Since $GL_n(\mathbb R)$ has two path components (given by sign of determinant), this shows that the conjugacy class of $A$ is the union of two path connected subsets (conjugating by things with positive or negative determinant), and it will be path connected if these two subsets intersect.

If $PAP^{-1}=QAQ^{-1}$ where $\det(P)>0$ and $\det(Q)<0$, we have $A=(P^{-1}Q)A(P^{-1}Q)^{-1}$, so $A$ commutes with a matrix with negative determinant. The converse is also true, so we have the following result.

Lemma: The conjugacy class of $A$ is path connected if and only if $A$ commutes with some matrix with negative determinant.

This gives several conditions that would ensure conjugacy classes are path connected.

  • If $n$ is odd, since then $\det(-I_n)=-1$
  • If $\det(A)<0$
  • If $\mathbb R^n$ as a direct sum of two $A$-invariant subspaces where one summand is odd dimensional, (e.g, if the Jordan normal form of $A$ has a block of odd size with a real eigenvalue).
  • If $\mathbb R^{n}$ is the direct sum of two $A$-invariant subspaces, and the restriction of $A$ to either subspace has negative determinant.

One can check that Jordan blocks only commute with matrices that are upper triangular and have only a single eigenvalue, and if $n$ is even and that eigenvalue is real, such a matrix could not have negative determinant. So these give counterexamples.

I suspect one could give a nice characterization of all counterexamples in terms of JNF. However, I have not worked out the details.



Here is a counterexample: Let $n=2$, $A=\left( \begin{array} [c]{cc} 0 & -1\\ 1 & 0 \end{array} \right) $ and $B=-A=\left( \begin{array} [c]{cc} 0 & 1\\ -1 & 0 \end{array} \right) $. Then, $A$ and $B$ are conjugate, since $B=C^{-1}AC$ for the invertible matrix $C=\left( \begin{array} [c]{cc} 1 & 0\\ 0 & -1 \end{array} \right) $. Thus, there exists a conjugacy class $\mathcal{C}$ in $\operatorname*{GL}\nolimits_{2}\left( \mathbb{R}\right) $ that contains both $A$ and $B$. However, there exists no path from $A$ to $B$ in $\mathcal{C}$. Why not?

Probably there is a nice conceptual reason [EDIT: yes, and @Wojowu explains it in his answer], but you can just as well brute-force it: I claim that every matrix in $\mathcal{C}$ has a nonzero $\left( 1,2\right) $-th entry. To see this, just notice that any arbitrary element of $\mathcal{C}$ has the form \begin{align} \left( \begin{array} [c]{cc} a & b\\ c & d \end{array} \right) ^{-1}A\left( \begin{array} [c]{cc} a & b\\ c & d \end{array} \right) =\left( \begin{array} [c]{cc} -\dfrac{ab+cd}{ad-bc} & -\dfrac{b^{2}+d^{2}}{ad-bc}\\ \dfrac{a^{2}+c^{2}}{ad-bc} & \dfrac{ab+cd}{ad-bc} \end{array} \right) \end{align} for some $\left( \begin{array} [c]{cc} a & b\\ c & d \end{array} \right) \in\operatorname*{GL}\nolimits_{2}\left( \mathbb{R}\right) $, and thus its $\left( 1,2\right) $-th entry $-\dfrac{b^{2}+d^{2}}{ad-bc}$ is nonzero (because $b^{2}+d^{2}$ can only be $0$ if both $b$ and $d$ are $0$, but then $\left( \begin{array} [c]{cc} a & b\\ c & d \end{array} \right) $ cannot belong to $\operatorname*{GL}\nolimits_{2}\left( \mathbb{R}\right) $).

Thus, every matrix in $\mathcal{C}$ has a nonzero $\left( 1,2\right) $-th entry. But if there was a path from $A$ to $B$ in $\mathcal{C}$, then some point on this path would be a matrix in $\mathcal{C}$ with zero $\left( 1,2\right) $-th entry (since the $\left( 1,2\right) $-th entries of $A$ and $B$ have opposite signs). Thus, there cannot be a path from $A$ to $B$ in $\mathcal{C}$.

  • $\begingroup$ There is also a geometric way to see it if you're interested. $\endgroup$ – Arnaud Mortier Nov 4 '19 at 0:48
  • $\begingroup$ @ArnaudMortier I'm interested. $\endgroup$ – Abhimanyu Pallavi Sudhir Nov 29 '19 at 13:22
  • $\begingroup$ @AbhimanyuPallaviSudhir Thanks! You will find it in my answer which is currently the accepted one I think. $\endgroup$ – Arnaud Mortier Nov 29 '19 at 13:47
  • $\begingroup$ Ah I missed that, thanks. $\endgroup$ – Abhimanyu Pallavi Sudhir Nov 29 '19 at 15:04

Notice that the assertion "since the image of the exponential map is the path-component of the identity in $GLn(R)$" is false.

Firstly, the path-component of the identity in $GLn(R)$ is the set of matrices with $>0$ determinant. Thus (inside this component) there is a path linking your $P$ and $I$ (in fact, there is nothing to prove).

Secondly, $diag(-1,-2)$ has $>0$ determinant and is not in the image of the real matrices by the exponential map.


Let $SC(A)$ be the conjugacy class of $A\in M_n(\mathbb{R})$.

*In dimension $2$, $SC(A)$ is non-connected in 2 cases

a. The eigenvalues of $A$ are non-zero conjugate complex; then $SC(A)$ is homeomorphic to a hyperboloid of two sheets.

b. $A$ is non-zero and non-diagonalizable; then $SC(A)$ is homeomorphic to a conical surface with the apex cut off.

*In dimension $4$, we consider the matrices $A_1=diag(U,U)$ where $U=\begin{pmatrix}0&1\\0&0\end{pmatrix}$ and $A_2=diag(V,V)$ where $V=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$. Then, using the Aaron's test, we can prove that $SC(A_1),SC(A_2)$ are non-connected (it's easy for $A_1$ and more difficult for $A_2$).

*Assume that we randomly choose $A\in M_n(\mathbb{R})$ -the $(a_{i,j})$ follow iid normal laws- We deduce that follows

$\textbf{Proposition}.$ When $n\rightarrow +\infty$, the probability that $SC(A)$ is connected tends to $1$.

$\textbf{Proof}$. A random matrix $A$ has distinct complex eigenvalues with probabiity $1$. Then, up to a real change of basis, $A=diag(a_1I_2+b_1V,\cdots,a_pI_2+b_pV,\lambda_1,\cdots,\lambda_q)$, where $2p+q=n$, the $(b_i)$ are non-zero and the $(\lambda_j)$ are real distinct.

Thus $SC(A)$ is connected iff $q\not=0$ (Aaron's test). When $n$ tends to $+\infty$, the mean of the number of real zeroes of a polynomial of degree $n$ is in $\Omega(\sqrt{n})$; we can deduce that the probability that $A$ has, at least, one real eigenvalue tends to $1$ when $n$ tends to $+\infty$. $\square$

  • $\begingroup$ Sorry for not reacting to this sooner. I have now finally managed to convince myself that what you say is true. Thanks for pointing it out! $\endgroup$ – Levi Nov 29 '19 at 13:09
  • $\begingroup$ This statement also struck me when I first read it but then I focused on answering the question and forgot about it. Thanks for this. $\endgroup$ – Arnaud Mortier Nov 29 '19 at 13:47
  • $\begingroup$ @Levi . More precisely, the real matrices that can be written in the form $e^Q$ are the squares of real matrices. $\endgroup$ – loup blanc Nov 29 '19 at 14:02
  • $\begingroup$ @Arnaud Mortier . Thanks Arnaud. I wrote this because I was surprised that no one would react. $\endgroup$ – loup blanc Nov 29 '19 at 14:05

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.