Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I have a matrix $Z$ of the form $Z = \left[Q^{-1}-Q^{-1}A^T\left(AQ^{-1}A^T\right)^{-1}AQ^{-1}\right]\Phi$


  • $\Phi$ is a diagonal matrix of real non-negative values.
  • $\Theta$ (not explicitly in the equation above) is another diagonal matrix of real non-negative values.
  • $Q = \Phi + \Theta$ such that $Q$ is a diagonal matrix of real positive values.
  • $A$ is a rectangular matrix of the form $\left[J \quad -I\right]$ where the elements of $J$ are real values and $I$ is the identity matrix of appropriate size.

Matrix dimensions are as follows:

  • $\Phi, \Theta, Q$ are of dimension $(n+m)\times (n+m)$.
  • $A$ is of dimension $m\times (n+m)$.
  • $J$ is of dimension $m\times n$.
  • $I$ is of dimension $m\times m$.

$Z$ is the plant matrix of a discrete time linear time-invariant state space equation, and I have reason to believe that, given the properties above, $Z$ is always convergent. As the state space equation is in discrete time, the conditions for a convergent plant are that all eigenvalues must have magnitude $\leq 1$. Also, should there be an eigenvalue with magnitude $= 1$, its multiplicity must be 1 to be at least Lyapunov stable.

I would appreciate any help in either proving or disproving my claim.

So far, I have been able to prove that $Q^{-1}\Phi$ and $Q^{-1}A^T\left(AQ^{-1}A^T\right)^{-1}AQ^{-1}\Phi$ are both symmetric positive semi-definite. I have tried applying the Sherman–Morrison–Woodbury formula only to show that $Z$ is singular. I can also show that $Q^{-1}\Phi$ has a spectral radius $\leq 1$ but not the same for $Q^{-1}A^T\left(AQ^{-1}A^T\right)^{-1}AQ^{-1}\Phi$.

share|cite|improve this question
Sorry, didn't see the dimensions. – nonlinearism Aug 20 '13 at 18:40
up vote 2 down vote accepted

In general, for (possibly noninvertible) complex square matrices $A$ and $B$, the spectra of $AB$ and $BA$ are identical. Therefore the spectrum of $Z$ is identical to that of the real symmetric matrix \begin{align*} \widetilde{Z} :=&\Phi^{1/2}\left[Q^{-1}-Q^{-1}A^T\left(AQ^{-1}A^T\right)^{-1}AQ^{-1}\right]\Phi^{1/2}\\ =&\Phi^{1/2}Q^{-1/2}\left[I - Q^{-1/2}A^T(AQ^{-1}A^T)^{-1}AQ^{-1/2} \right]Q^{-1/2}\Phi^{1/2}\\ =&\Phi^{1/2}Q^{-1/2}\left(I-P\right)Q^{-1/2}\Phi^{1/2}, \end{align*} where $P := Q^{-1/2}A^T(AQ^{-1}A^T)^{-1}AQ^{-1/2}$ is an orthogonal projection (because it is real symmetric and $P^2=P$). Therefore, $I-P$ is also an orthogonal projection. As $A$ has full rank, it follows that the eigenvalues of $I-P$ are $0$ (of multiplicity $m$) and $1$ (of multiplicity $n$). Since all diagonal entries of $\Phi^{1/2}Q^{-1/2}$ do not exceed $1$, the spectral norm and in turn the spectral radius of $\widetilde{Z}$ are also $\le1$.

Now, $1$ is a repeated eigenvalue of $\widetilde{Z}$ (or $Z$) if and only if the eigenspaces of $\Phi^{1/2}Q^{-1/2}$ and $I-P$ corresponding to the eigenvalue $1$ have a two- or higher dimensional intersection. Such undesired cases can happen, for instances, when (a) $\Phi=Q$ or (b) when $J=0,\,n\ge2$ and some two among the first $n$ entries of $\Phi^{1/2}Q^{-1/2}$ are equal to $1$. However, if $n=1$ or $\Phi^{1/2}Q^{-1/2}$ has at most one diagonal entry equal to $1$, then it is guaranteed that $1$ is not a repeated eigenvalue of $Z$.

share|cite|improve this answer
Wow that is amazing. I didn't realize that the second term in the brackets was the form of a projection. Thank you! – user91124 Aug 21 '13 at 18:00
@user91124 The second term in your square bracket is not a projection. It only becomes one after you pull out $Q^{-1/2}$ from both left and right. – user1551 Aug 21 '13 at 18:18
Thanks again, I see that now. Do you have a source or citation for your proof? or did you come up with it? – user91124 Aug 21 '13 at 20:10
I noticed you edited the proof. I am wondering where the $\Phi$ outside the bracket went in the first equation? – user91124 Aug 21 '13 at 20:46
@user91124 The $\Phi$ does not go away. It becomes $\Phi^{1/2}\Phi^{1/2}$ and I move the second $\Phi^{1/2}$ in front of the square bracket term. Now $Z=[\cdots]\Phi^{1/2}\Phi^{1/2}$ and $\widetilde{Z}=\Phi^{1/2}[\cdots]\Phi^{1/2}$ have identical spectra because this is true for $AB$ and $BA$ for general complex square matrices $A$ and $B$. – user1551 Aug 22 '13 at 14:46

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.