Let $[\theta_1,\theta_2, \dots, \theta_N]^\mathrm{T} \, \in \mathbb{R}^N$. The angles are not all identical (on the circle), i.e. $[\theta_1,\theta_2, \dots, \theta_N] \not \equiv c [1,1,\dots, 1]^\mathrm{T}\,\, \mathrm{mod}\,\, 2\pi$. Define matrices $C$ and $S$ as:

\begin{align} \begin{split} [C]_{jl}&= 1 \,\, \mathrm{if}\,\, j =l \\ [C]_{jl}&= \cos(\theta_j-\theta_l)\,\, \mathrm{if} \,\, j \neq l \end{split} \end{align}


\begin{align} \begin{split} [S]_{jl}&= 0 \,\, \mathrm{if}\,\, j =l \\ [S]_{jl}&= \sin(\theta_j-\theta_l)\,\,\mathrm{if} \,\, j \neq l\,. \end{split} \end{align}

Is it true that any $v$ that belongs to the nullspace of S also belongs to the nullspace of C?

My repeated simulations in MATLAB with randomly generated $[\theta_1,\theta_2, \dots, \theta_N]^\mathrm{T}$ seems to suggest that $C$ and $S$ are rank $2$ and any eigenvectors with $0$ eigenvalue for one is an eigenvector with $0$ eigenvalue for the other.

  • 2
    $\begingroup$ Could you explain the motivation behind the question? This looks like it could make for a really nice problem on a take-home linear algebra final; how did you stumble across it "in the wild"? $\endgroup$ Dec 13 '16 at 23:47
  • 1
    $\begingroup$ With the updated version of the problem, you don't need to separate the $j=l$ and $j \neq l$ cases any more, since they are both subsumed by the $j \neq l$ formula. Also, I'm wondering whether all the matrices in the family $\left(S_\phi\right)_{\phi\in\mathbb{R}}$, where the $\left(j,l\right)$-th entry of $S_\phi$ is $\sin\left(\theta_j-\theta_l+\phi\right)$, have the same nullspace. $\endgroup$ Dec 14 '16 at 1:15
  • $\begingroup$ @ElenaYudovina. I am trying to solve a nonlinear dynamics problem. Lyapunov analysis lead me to such kind of matrices. $\endgroup$
    – Mohit
    Dec 14 '16 at 4:17
  • $\begingroup$ What happens if $\theta_1=\pi$ and all other $\theta_i=0$? In that case, $S=0$ and $C\neq 0$, if I do not misunderstand. $\endgroup$ Dec 14 '16 at 7:26

You must have made a mistake.

$E = -C + 2I + i S$ has entries $\exp(i (\theta_j - \theta_l)) = \exp(i\theta_j) \exp(-i \theta_l)$ and thus is the product $V V^*$ where $V$ is the column vector with entries $\exp(i \theta_j)$, so it has rank $1$.
Then $S = (E - \overline{E})/(2i)$ has rank at most $2$ (it would be $0$ e.g. if all $\theta_j$ are multiples of $\pi$). Similarly, $-C+2I$ has rank at most $2$, so $C$ has rank at least $N-2$ (that part at least you got right). But "generically" $C$ will have rank $N$.

EDIT: OK, now with the corrected question we have $E = C + iS$. Again $S = (E - \overline{E})/(2i)$ and now $C = (E + \overline{E})/(2i)$ have rank at most $2$. Moreover, $\ker(E) \cap \ker(\overline{E}) = (V^*)^\perp \cap (\overline{V^*})^\perp$ is contained in both $\ker(S)$ and $\ker(C)$. Note that the following are equivalent:

  1. $\theta_j - \theta_l \not \equiv 0 \mod \pi$ for some $j,l$.
  2. $V$ and $\overline{V}$ are linearly independent.
  3. $(V^*)^\perp$ and $(\overline{V^*})^\perp$ are distinct linear subspaces.
  4. $\ker(E) \cap \ker(\overline{E})$ has dimension exactly $N-2$.
  • $\begingroup$ Thanks for your answer. I realized I made a mistake in writing my question. Actually my matrix $C$ had $\cos (\theta_j -\theta_l)$ as entries, without the negative sign. Also, I saw they were rank $2$ with $N-2$ zero eigenvalues as your argument points out. I have edited my question. I apologize for my sloppiness. $\endgroup$
    – Mohit
    Dec 13 '16 at 22:16
  • $\begingroup$ Your answer can be used to solve the corrected problem. Let me write $n$ for your $N$, out of habit. Let us use the notation $\operatorname{Ker}_{\mathbb{R}} U$ for "the intersection of the kernel of an $n\times n$-matrix $U \in \mathbb{C}^{n\times n}$ with the real vector space $\mathbb{R}^n$". With the correction in place, we have $E = C + iS$. But from $E = VV^\ast$, we obtain $\operatorname{Ker}E = \operatorname{Ker} \left(V^\ast\right)$ (by a standard argument using the positivity of the inner product). Hence, ... $\endgroup$ Dec 14 '16 at 1:59
  • $\begingroup$ ... we get $\operatorname{Ker}_{\mathbb{R}} E = \operatorname{Ker}_{\mathbb{R}} \left(V^\ast\right) = \left\{\left(x_1, x_2, \ldots, x_n\right)^T \in \mathbb{R}^n \mid \sum\limits_{k=1}^n e^{-i\theta_k}x_k = 0\right\}$ $ = \left\{\left(x_1, x_2, \ldots, x_n\right)^T \in \mathbb{R}^n \mid \sum\limits_{k=1}^n \left(\sin\theta_k\right) x_k = 0 \text{ and } \sum\limits_{k=1}^n \left(\cos\theta_k\right) x_k = 0\right\} $ (because the sum $\sum\limits_{k=1}^n e^{-i\theta_k}x_k$ has real part $\sum\limits_{k=1}^n \left(\cos\theta_k\right) x_k = 0 $ and ... $\endgroup$ Dec 14 '16 at 2:01
  • $\begingroup$ ... imaginary part $\sum\limits_{k=1}^n \left(\sin\theta_k\right) x_k = 0 $). This is the intersection of two hyperplanes in $\mathbb{R}^n$. Okay, now I am seeing that the problem is missing an assumption: We want the two hyperplanes to be distinct. But they will coincide if we have $\theta_1 \equiv \theta_2 \equiv \cdots \equiv \theta_n \mod \pi$ (not just $2 \pi$). That said, I hope that if we require this not to happen, the hyperplanes will be distinct, and thus $E$ (being their intersection) will be an $\left(n-2\right)$-dimensional real subspace of $\mathbb{R}^n$. This ... $\endgroup$ Dec 14 '16 at 2:05
  • $\begingroup$ ... said, I have no idea how to finish this off now, but a few minutes ago I saw it clearly that $\operatorname{Ker}_{\mathbb{R}} E = \operatorname{Ker}_{\mathbb{R}} C = \operatorname{Ker}_{\mathbb{R}} S$, at least assuming that the latter two kernels also have dimension $n-2$. $\endgroup$ Dec 14 '16 at 2:13

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.