Let $S$ denote the space of skew-symmetric $n\times n$ real matrices, where every element $A\in S$ satisfies $A^T+A = 0$.

Let $M$ denote an orthogonal $n\times n$ matrix, and $L_M$ denotes the linear map $$L_M: A\mapsto MAM^{-1}.$$ What is the determinant of $L_M$ as an endomorphism of $S$?

  • 1
    $\begingroup$ I'm not sure that $MAM^{-1}$ is skew-symmetric for skew-symmetric $A$ and general invertible $M$: $(MAM^{-1})^T = (M^{-1})^TA^TM^T = -(M^{-1})^TAM^T$. If $M$ is orthogonal, that is, $M^{-1} = M^T$, then $-(M^{-1})^TAM^T = -MAM^T$, so things work. But I have my doubts when $M$ is not orthogonal. Does it work for $M= M^T$, that is, $M$ symmetric? $\endgroup$ – Robert Lewis Jul 28 '14 at 18:32
  • $\begingroup$ For $M = M^T$, $(MAM^{-1})^T = (M^{-1})^TA^TM^T = - (M^T)^{-1}AM = -M^{-1}AM$; this doesn't look like it will work for skew-symmetric $A$ and general symmetric $M$! $\endgroup$ – Robert Lewis Jul 28 '14 at 19:15
  • $\begingroup$ See also math.stackexchange.com/questions/275925/… where the case of conjugation on the whole of $M_n(K)$ is discussed $\endgroup$ – martini Jul 28 '14 at 19:56
  • $\begingroup$ @RobertLewis You are right. I edited the question. $\endgroup$ – mez Jul 28 '14 at 20:11

Proposition For any orthogonal $n \times n$ matrix $M$, the conjugation map $L_M: S \to S$, $M \mapsto M A M^{-1}$ has determinant $1$ if $M \in SO(n)$ and/or $n$ is odd, and $-1$ if $M \in O(n) - SO(n)$ and $n$ is even.

Lemma For any orthogonal $n \times n$ matrix $M$, $\Phi: M \mapsto \det L_M$ is a homomorphism $O(n) \to \mathbb{R}^*$, that is, $\Phi$ satisfies:

  1. $\Phi(I_n) = 1$,
  2. $\Phi(MN) = \Phi(M) \Phi(N)$ for all $M, N \in O(n)$, and
  3. $\Phi(M^{-1}) = \Phi(M)^{-1}$ for all $M \in O(n)$.

Proof of Lemma This is immediate, because $\Phi$ is a composition of the determinant representation $GL(S) \to \mathbb{R}^*$ and the adjoint representation $O(n) \to GL(S)$, but for readers who aren't familiar with these ideas, we can verify these claims directly.

  1. We have $L_{I_n}(A) = I_n A I_n^{-1} = A$, so $L_{I_n}$ is the identity map on $S$, and thus $\Phi(I_n) = \det L_{I_n} = 1$.
  2. By definition, $L_{MN}(A) = (MN)A(MN)^{-1} = M(NAN^{-1})M^{-1} = L_M(L_N(A))$, so $\Phi(MN) = \det L_{MN} = \det (L_M \circ L_n) = \det L_M \det L_N = \Phi(M) \Phi(N)$.
  3. (Follows from (1) and (2).) $\square$

Proof of Proposition Since $\Phi$ is continuous and $O(n)$ is compact, the image of $\Phi$ is compact, and hence it is contained in $\{\pm 1\}$. (Suppose not; then there is an element $A$ with $\Phi(A) \not\in \{\pm 1\}$, and by inverting [and using (3)] if necessary we can assume $|\Phi(A)| > 1$. Then, by (2) the image under $\Phi$ of the sequence $(A, A^2, A^3, \ldots)$ is $(\det A, (\det A)^2, (\det A)^3, \ldots)$, which is unbounded, a contradiction with the compactness of the image of $\Phi$.) Since $\Phi$ is continuous with discrete components, it is constant on connected components, and by (1) in the Lemma, $\Phi(M) = \Phi(I_n) = \det L_{I_n} = 1$ for $M$ in the connected component of $I_n$ in $O(n)$, namely $SO(n)$. (Notice how quickly we arrive at this result just using a few easy observations.)

Since $O(n) - SO(n)$ is connected and $\Phi$ is constant on connected components, to compute $\det L_M$ for $M$ in this set, it suffices to compute it for a single element (for each $n$). A simple choice is

$Z := \left(\begin{array}{cc} -1 & 0 \\ 0 & I_n \end{array}\right)$.

We compute $\det L_Z$ directly by computing the matrix representation of $L_Z$ in the basis $(F_{ij})_{i < j}$ of $S$, where $F_{ij}$ is the matrix with $(i, j)$ entry $1$, $(j, i)$ entry $-1$, and all other entries zero, so that

$(F_{ij})_{kl} = \delta_{ik} \delta_{jl} - \delta_{il} \delta_{jk}$.

Then, $L_Z(F_{ij}) = Z F_{ij} Z^{-1}$ has entries

$(L_Z(F_{ij}))_{ab} = \sum_{k, l} Z_{ak} (F_{ij})_{kl} Z_{lb} = Z_{ai} Z_{jb} - Z_{aj} Z_{ib}$,

and by construction this is just the $((a, b), (i, j))$ entry $(L_Z)_{(a, b), (i, j)}$ of the matrix representation of $L_Z$ in the basis $(F_{ij})_{i < j}$. We show that this matrix is in fact diagonal, using that $Z_{11} = -1$, $Z_{aa} = 1$ for $a > 1$, and $Z_{ab} = 0$ for $a \neq b$:

When $a = i = 1$ (so that $b, j > 1$), we have

$(L_Z)_{(1, b), (1, j)} = Z_{11} Z_{jb} - Z_{1j} Z_{1b} = -\delta_{jb}$;

when $i = 1$ and $a > 1$ (or, by symmetry of notation, $a = 1$ and $i > 1$, we have

$(L_Z)_{(a, b), (1, j)} = Z_{a1} Z_{jb} - Z_{aj} Z_{ib} = 0$;

when $a, i > 1$, we have

$(L_Z)_{(a, b), (i, j)} = Z_{ai} Z_{jb} - Z_{aj} Z_{ib} = \delta_{ai} \delta_{jb} - \delta_{aj} \delta_{ib} = \delta_{ai} \delta_{jb}$,

where the last equality comes from the fact that if $\delta_{aj} \neq 0$, then $a = j$, so that $i < j = a < b$, so $\delta_{ib} = 0$, that is the latter term in the second-to-last expression in the previous display is always zero.

In summary, the above computations give us that in the basis $(F_{ij})_{i < j}$, the matrix representation of $L_Z$ is diagonal, the $n - 1$ diagonal entries for the basis elements $F_{1j}$ are $-1$, and the remaining diagonal entries are all $1$. Thus, $\det L_Z = (-1)^{n - 1}$ as claimed. $\square$

  • 1
    $\begingroup$ @ Travis, $\det(L_M)$ can be $-1$. For instance, take, for $n=2$, $M=\begin{pmatrix}0&1\\1&0\end{pmatrix}\in O^{-}(2)$. $\endgroup$ – loup blanc Aug 24 '14 at 10:34
  • $\begingroup$ loup, you're right. The error was that elements in $O^-(n)$ cannot be written as $\exp N$ for any $N \in S$, which is a consequence of the fact that $O(n)$ is not connected. I've added a short argument for handling elements in $O^-(n)$ to the end of the reply, which reduces the problem to computing the answer explicitly just once (for each matrix size anyway). $\endgroup$ – Travis Aug 24 '14 at 13:12
  • 1
    $\begingroup$ @ Travis, I also thought that $\det(L_M)=\det(M)$. Unfortunately, it is false: take, for $n=3$, $M=diag(1,1,-1)$ ; then $L_M=diag(1,-1,-1)$. $\endgroup$ – loup blanc Aug 24 '14 at 15:35
  • $\begingroup$ Thanks, loup, I didn't check your remark and hedged in my answer accordingly, and included the natural guess. I am somewhat surprised that parity arises in this way, perhaps there is a conceptual explanation for it? $\endgroup$ – Travis Aug 24 '14 at 15:46
  • $\begingroup$ I've rewritten the solution from the scratch, it's now much cleaner and rather easier. $\endgroup$ – Travis Aug 25 '14 at 5:57

If we stack the entries of the matrices row by row, then $L_M=M\bigotimes M^{-T}:M_n\rightarrow M_n$ is a Kronecker product. The eigenvalues of $L_M$ are the $(\lambda_j/\lambda_k)_{j,k}$ where $\operatorname{spectrum}(M)=(\lambda_j=\exp(i\theta_j))_j$. Finally $\operatorname{spectrum}(L_M)=(\exp(i(\theta_j-\theta_k))_{j,k}$ and each eigenvalue has modulus $1$.

Here we consider $\phi={L_M}_{|S}$; then $\det(\phi)$ is a product (the result is real) of a part of the eigenvalues of $L_M$. Thus $\det(\phi)=\pm 1$. For example, let $n=2$ ; if $M\in O^+(2)$ then $\det(\phi)=1$ ($\phi(s)=s$) and if $M\in O^-(2)$, then $\det(\phi)=-1$ ($\phi(s)=-s$).

EDIT: on can slightly improve the previous result: ${L_M}^T=M^T\bigotimes M^T$ implies that $L_M{L_M}^T=MM^T\bigotimes MM^T=I_{n^2}$ ; moreover $\det(L_M)=\det(M)^{2n}=1$. Then $L_M\in O^+(n^2)$ and $\phi\in O^+(S)$ or $O^-(S)$.

  • $\begingroup$ What do you mean by $M^{-T}$? $\endgroup$ – user26857 Aug 23 '14 at 19:25
  • $\begingroup$ @ user26857, $M^{-T}$ is $(M^{-1})^T$. $\endgroup$ – loup blanc Aug 24 '14 at 1:17

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.