This is a subject I've been working on for a very long time now, but still did not manage to fully understand the interesting properties of this matrix. I have already asked a (viewed but unanswered) question about the same matrices (cf. here), but the question here is different.

First, let's define two matrices:

  • $\mathbf{N}$ is the following matrix: \begin{equation} \mathbf{N}=\begin{bmatrix} \mathbf{I}_n & \mathbf{0}_n \\ \mathbf{0}_n & \mathbf{P}^{-1}\begin{bmatrix}1 & && \\ & \ddots && \\ & & 1& \\ &&& -1 \end{bmatrix}\mathbf{P} \end{bmatrix}\in\mathbb{R}^{2n\times2n} \end{equation} where $\mathbf{P}\in\mathbb{R}^{n\times n}$ is any invertible matrix.

  • with $\omega_i>0$ and $t>0$, the block-diagonal matrix: \begin{equation} \mathbf{S}(t)=\begin{bmatrix} \begin{bmatrix} \cos(\omega_1t) & \\& \ddots & \\ & & \cos(\omega_n t) \end{bmatrix} & \begin{bmatrix} \dfrac{\sin(\omega_1t)}{\omega_1} & \\& \ddots & \\ & & \dfrac{\sin(\omega_nt)}{\omega_n} \end{bmatrix} \\ \begin{bmatrix} -\omega_1 \sin(\omega_1t) & \\& \ddots & \\ & & -\omega_n\sin(\omega_n t) \end{bmatrix} & \begin{bmatrix} \cos(\omega_1t) & \\& \ddots & \\ & & \cos(\omega_n t) \end{bmatrix}\end{bmatrix}\in\mathbb{R}^{2n\times2n} \end{equation}

The eigenvalues of $\mathbf N$ are of course 1 (multiplicity $2n-1$) and $-1$ (multiplicity $1$). The eigenvalues of $\mathbf{S}(t)$, which is an exponential matrix, are the $n$ couples of the complex conjugates $(\exp(i\omega_jt),\overline{\exp(i\omega_jt)})$.

Now, we can define $\forall t>0$, $\mathbf{A}(t)=\mathbf N\mathbf S(t)$. We know that the product of the eigenvalues of $\mathbf{A}(t)$ is the product of those of $\mathbf{N}$ and $\mathbf S(t)$, i.e. $-1$.

I observe an interesting property but can't prove where it stems from:

  • $1$ and $-1$ are eigenvalues of $\mathbf{A}(t)$ ($\forall t$);
  • $1$ and $-1$ are $\color{red}{\text{not}}$ eigenvalues of $\mathbf{A}(t_2)\mathbf{A}(t)$ ($\forall t,t_2$, except maybe for specific values of $\mathbf P$ and $\omega_k$);
  • $1$ and $-1$ are eigenvalues of $\mathbf{A}(t_3)\mathbf{A}(t_2)\mathbf{A}(t)$ ($\forall t,t_2,t_3$);
  • $1$ and $-1$ are $\color{red}{\text{not}}$ eigenvalues of $\mathbf{A}(t_4)\mathbf{A}(t_3)\mathbf{A}(t_2)\mathbf{A}(t)$ ($\forall t,t_2,t_3,t_4$, except maybe for specific values of $\mathbf P$ and $\omega_k$);
  • $\dots$

I managed to prove $1$ and $-1$ are eigenvalues of $\mathbf{A}(t)$ by considering $\mathbf{S}(t)\pm\operatorname{diag}(1,\dots,1,-1,\dots,1)$, calculating its kernel, and building the appropriate vectors (without having to calculate them explicitly).

Also, I understand that the product of the eigenvalues of $\mathbf{A}(t_2)\mathbf{A}(t)$ is 1, while that of $\mathbf{A}(t_3)\mathbf{A}(t_2)\mathbf{A}(t)$ is -1, but that does not prove anything.


1) Any suggestion to prove the framed observation would be very welcome: why are apparently 1 and -1 eigenvalues of $\prod_{i=1}^m A(t_i)$ if and only if $m$ is odd?

2) Also, I have the impression that there exists a powerful mathematical framework to study these matrices, but I can't figure out which one, as not being a mathematician; Lie algebra because $\mathbf S(t)$ is an exponential? Galois groups because the eigenvalues are complex conjugate? Zariksi topology because @loup blanc mentioned it (see end of answer)?

Edit A simple Mathematica file to reproduce the results is available here. Just play with the arguments of calculateEigenvals to change the dimension $n$ or/and the exponent $m$ (to prove: 1,-1 eigenvalues iff $m$ is odd).

  • $\begingroup$ @user1551 I've added a link to a more general Mathematica file (any dimension, any exponent) -- the observations still stand. $\endgroup$ – anderstood Feb 28 '15 at 17:07
  • $\begingroup$ You do mean the lower right corner of $\mathbf{N}\text{ is }\mathbf{P}^{-1}\begin{bmatrix}1 & && \\ & \ddots && \\ & & 1& \\ &&& -1 \end{bmatrix}\mathbf{P}=\mathbf{P}^{-1}(\mathbf{I}-\begin{bmatrix}0 & && \\ & \ddots && \\ & & 0& \\ &&& 2 \end{bmatrix})\mathbf{P}=\mathbf{I}-2\mathbf{P}^{-1}\begin{bmatrix}0 & && \\ & \ddots && \\ & & 0& \\ &&& 1 \end{bmatrix}\mathbf{P}$? $\endgroup$ – Mark Hurd Mar 3 '15 at 5:29
  • $\begingroup$ @MarkHurd That's right. Note that I got an answer on MO which I will give also here once I have understood everything. $\endgroup$ – anderstood Mar 3 '15 at 15:38
  • 1
    $\begingroup$ @user1551: That's not exactly what he say, but if you slightly modify the answer: consider $(D\oplus D)$, then $(D\oplus D)S(t)(D\oplus D)^{-1}=S(t)$ because diagonal blocks commute. What is not clear to me is why such a $(D\oplus D)$ matrix exists (on a dense set of matrices)---see question here. $\endgroup$ – anderstood Mar 5 '15 at 15:18
  • 1
    $\begingroup$ Cross-posted on MO under the same title. $\endgroup$ – user1551 Feb 18 '17 at 9:21

Answer by Terry Tao on MO:

First, one should conjugate all matrices by $$ \begin{pmatrix} \operatorname{diag}(\omega_1,\dots,\omega_n) & 0 \\ 0 & 1 \end{pmatrix} $$ as this converts $S(t)$ to a rotation matrix while leaving the reflection $N$ unchanged.

The matrix $P^{-1} \operatorname{diag}(1,\dots,1,-1) P$ has a line as its $-1$ eigenspace and a hyperplane as its $+1$ eigenspace. By a limiting argument we may take $P$ to be in general position, so that the $-1$ eigenspace is not contained in any coordinate hyperplane and the $+1$ eigenspace does not contain any coordinate line. Then after applying a further conjugation by a diagonal matrix in the lower $n \times n$ block, we can arrange for the $-1$ and $+1$ eigenspaces to be orthogonal, without affecting $A(t)$.

After all these conjugations, $S(t)$ is now orthogonal and orientation preserving, while $N$ is orthogonal and orientation reversing. Then it is now clear that the product of any odd number of the $A(t)$ will be orientation-reversing orthogonal matrices. Such matrices have spectrum on the unit circle, symmetric with respect to conjugation, and multiplying to $-1$, hence must have an odd number of eigenvalues at $+1$ and also at $-1$. (If instead one multiplies an even number of $A(t)$ together, one obtains an even number of eigenvalues at $-1$ and at $+1$, but usually one expects to get zero eigenvalues at either.)


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.