If $\boldsymbol{A}$ is a rank deficient Hermitian matrix are the following true?

1) Is $<\boldsymbol{x}_j,\boldsymbol{x_k}>=0$ when $\lambda_j=\lambda_k$?

2) Is $<\boldsymbol{x}_j,\boldsymbol{x_k}>=0$ when $\lambda_j\neq\lambda_k$?

In the above two points $\lambda_i$ is an eigenvalue of $\boldsymbol{A}$ and $\boldsymbol{x}_i$ is its associated eigenvector.


Is in not the case that, whether $A$ is rank-deficient or not, we have

$\langle x, Ay \rangle = \langle A^{\dagger} x, y \rangle = \langle Ax, y \rangle? \tag{1}$

And by (1),

$\langle x_k, \lambda_j x_j \rangle = \langle x_k, Ax_j \rangle = \langle Ax_k, x_j \rangle = \langle \lambda_k x_k, x_j \rangle, \tag{2}$

which immediately leads to

$\lambda_j \langle x_k, x_j \rangle = \lambda_k \langle x_k, x_j \rangle, \tag{3}$


$(\lambda_k - \lambda_j) \langle x_k, x_j \rangle, \tag{4}$

and with $\lambda_k \ne \lambda_j$ we have

$\langle x_k, x_j \rangle = 0. \tag{5}$

That's half the story; I'll try to say more tomorrow, after I get some shut-eye.

Monday 2 December 2013 12:35 PM PST: Good Morning! I'm Baaaaack!!!

And now for the second part of our story, which was in fact the first question posed:

Again, the possible rank-deficiency of $A$ does not really bear on this issue.

If we have $\lambda_j = \lambda_k = \lambda$ for two linearly independent eigenvectors $x_j$ and $x_k$, they don't necessarily have to be orthogonal. Indeed, if they were orthogonal, $\langle x_j, x_k \rangle = 0$, we could always construct two linearly independent, non-orthogonal eigenvectors $y_j$, $y_k$ such that $\text{span}(y_j, y_k) = \text{span}(x_j, x_k)$; simply set $y_j = x_j$ and set $y_k = \alpha x_j + \beta x_k$ with the scalars $\alpha, \beta$ satisfying $\alpha \ne 0 \ne \beta$; then $\langle y_j, y_k \rangle = \langle x_j, \alpha x_j + \beta x_k \rangle = \alpha \langle x_j, x_j \rangle \ne 0$ and $\text{span}(y_j, y_k) = \text{span}(x_j, x_k)$. Furthermore, if $x_j$ and $x_k$ weren't orthogonal, $\langle x_j, x_k \rangle \ne 0$ then taking $y_j = x_j$ and $y_k = x_k - (\langle x_j, x_k \rangle / \langle x_j, x_j \rangle) x_j$ then $\langle y_j, y_k \rangle = 0$; I leave it to you to show $Ay_k = \lambda y_k$ and so forth, and that $\text{span}(y_j, y_k) = \text{span}(x_j, x_k)$ here as well. The point is that in the event of $\lambda_j = \lambda_k$, taking the eigenvectors orthogonal is an option, not a mathematical necessity.

Hope this helps. Cheerio,

and as always,

Fiat Lux!!!

  • $\begingroup$ Wow that definitely answers (2). Thanks, looking forward to the other half of the story. $\endgroup$ – trienko Dec 2 '13 at 9:25

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.