I need to construct a Hemintian Matrix with some constraint of its eigenvectors.

Suppose the eigenvectors of the matrix are $u_1=[u_{1,1},u_{1,2},u_{1,3}]$, $u_2=[u_{2,1},u_{2,2},u_{2,3}]$, and $u_3=[u_{3,1},u_{3,2},u_{3,3}]$, and I want $|u_{n,m}|=1$ for all $n=1,2,3$ and $m=1,2,3$. How can I construct these kind of matrices?


  • $\begingroup$ Did you really mean to have two $u_1's$? $\endgroup$ – Amzoti Jul 6 '13 at 2:06
  • $\begingroup$ Sorry for the typo, I've corrected it, the last one should be $u_3$. $\endgroup$ – user2008790 Jul 6 '13 at 2:54
  • $\begingroup$ The eigenvectors of a Hermitian matrix are orthogonal, no? So that's another constraint on your eigenvectors --- not easy to find three of them with unit entries and mutually perpendicular. Maybe they have to be (essentially) $(1,1,1)$, $(1,\omega,\omega^2)$, and $(\omega^2,1,\omega)$, where $\omega$ is a complex cube root of 1. $\endgroup$ – Gerry Myerson Jul 6 '13 at 3:00
  • $\begingroup$ @GerryMyerson Thanks! $\endgroup$ – user2008790 Jul 6 '13 at 17:43
  • 1
    $\begingroup$ @awl, is it obvious that there is an orthogonal basis for, say, ${\bf R}^{36}$ in which the all 36 components of all 36 vectors have modulus 1? Also, I think OP is after examples beyond the identity matrix. $\endgroup$ – Gerry Myerson Jul 7 '13 at 9:53

Let $U=\begin{pmatrix}1&e^{-i\pi/3}&e^{i\pi/3}\\1&e^{i\pi}&-1\\e^{2\pi i/3}&e^{\pi i/3}&-1\end{pmatrix}$. Then, since $1+e^{2i\pi/3}=e^{i\pi/3}$, we find that $V\overline V^T=I_3$, where $V=U/\sqrt3$. Now let $A=V\begin{pmatrix}\lambda_1&0&0\\0&\lambda_2&0\\0&0&\lambda_3\end{pmatrix}\overline V^T$, where $\lambda_1, \lambda_2, \lambda_3\in \mathbb C$.

It is then immediate that $A$ is a Hermitian matrix, with eigen-vecotrs of the required type, since the eigen-spaces still contain the vectors in $U$.
Essentially it is to find a set of three orthonormal vectors with unit entries. And we can assign any eigen-value that we want to the matrix. And I just find three orthogonal such vectors.

The following is about the verification of the properties of $V$. Skip it if needed.
Let me show that $U$ is orthogonal:
$(1,e^{-i\pi/3},e^{i\pi/3})\cdot(1,e^{i\pi},-1)=1+e^{2i\pi/3}-e^{i\pi/3}=0$, $(1,e^{-i\pi/3},e^{i\pi/3})\cdot(e^{2\pi/3},e^{i\pi/3},-1)=e^{2i\pi/3}-e^{i\pi/3}+1=0$,
$(1,e^{i\pi},-1)\cdot(e^{2i\pi/3},e^{i\pi/3},-1)=e^{2i\pi/3}-e^{i\pi/3}+1=0$. (Notice that $e^{i\pi}+1=0$)
In the end, to see the claimed equality $e^{2i\pi/3}+1=e^{i\pi/3}$, see the following (rough) picture:



In fact, the space of eigen-vectors of the required Hermitian matrix is unique, and we shall show it in the following:
Suppose $u_1\cdot u_2=0$ where $u_1, u_2$ have unit entries. Then we obtain an identity $x_1+x_2+x_3=0$, where each $x_i$ is of the form $e^{i\theta}$. By factoring out a constant, which affects not the eigen-space, we might assume that $x_1=1$. Now it is easily verified that such an equality $1+e^{i\theta_1}+e^{i\theta_2}=0$ is of the above type. It is now a simple but tedious matter to verify that the choice of three vectors which satisfy the identities right above the picture is unique, in the sense that the eigen-space is unchanged. Thus the requirement defines a unique eigen-space in $\mathbb C^3$.

Hope the above is right, and that this helps.

  • $\begingroup$ Thanks a lot! It is exactly I need to construct. $\endgroup$ – user2008790 Jul 6 '13 at 17:50
  • $\begingroup$ I was think if there is a general method to construct such matrix if the matrix is arbitrary dimension? By the way, even for the $3 \times 3$ matrix, what is the intuition to find the vectors to be orthogonal? $\endgroup$ – user2008790 Jul 6 '13 at 18:01
  • 1
    $\begingroup$ user, did you not check my remark about the eigenvectors of Hermitian matrices being orthogonal? $\endgroup$ – Gerry Myerson Jul 7 '13 at 3:23
  • 1
    $\begingroup$ As for $n\times n$ matrices: if $n$ is prime, then there will be a similar construction with $n$th roots of unity. If there is a Hadamard matrix of size $n$, its columns have entries $\pm1$ and are mutually orthogonal, so will serve as eigenvectors. E.g., $n=4$: $(1,1,1,1)$, $(1,1,-1,-1)$, $(1,-1,1,-1)$, $(1,-1,-1,1)$. I suspect there's a roots-of-unity construction for every $n$. $\endgroup$ – Gerry Myerson Jul 7 '13 at 3:27
  • $\begingroup$ @GerryMyerson Thanks a lot. Yes I checked it, sorry for the confusion, I understand that the vectors must be orthogonal. What I didn't understand is that how did you find such construction that will lead to orthogonality. To me seems that it involves some fundamental theory about Algebra, now I kind understand maybe it is an open question and looking at roots-of-unity seems to be a good way. $\endgroup$ – user2008790 Jul 7 '13 at 19:35

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.