6
$\begingroup$

Prove or disprove:

if $A^2 \in M_{3}(\mathbb{R})$ is diagonalizable then so is $A$.

I'm pretty confident this is not true, but I've tried and tried to find a counter example without success. If someone contradicts this, I'd appreciate if you can outline your thought process when constructing the matrix $A$.

Also, another question in the same batch, asks:

Let $p(t)=t(t-0.25)(t-1)$ be the characteristic polynomial of $A^2$, is $A$ diagonalizable?

I'm thinking this is suppose to answer the previous question, assuming the answer here is false.

Here I know $A^2$ is diagonalizable, but I haven't made any substantial progress other than that.
I know I can go from the eigenvalues of $A$ to the eigenvalues of $A^k$, but not the other way around.

  • 0
    Daniel, I have changed the notation to $M_{3}(\mathbb{R})$ which is the standard notation for set of all $3 \times 3$ real matrices. I hope that isn't a problem. If it is, then please rollback.2011-07-06
  • 0
    @Chandru, not at all, thank you.2011-07-06
  • 0
    This is not true for matrices with a nilpotent Jordan component. However, if $A$ is non-singular, and we switch the field to complex numbers, then it is true. A slightly more general case (some power known to be diagonalizable) of this was recently discussed in http://fora.xkcd.com/viewtopic.php?f=17&t=711072011-07-06

4 Answers 4

9

Try $A=\begin{pmatrix} 0&0&1 \\ 0&0&0 \\ 0&0&0\end{pmatrix}$. The idea is that the null matrix is diagonalizable, but there exist matrices which satisfy $A^2=0$ and $A$ is not diagonalizable.

For the second question, note that the eigenvalues of $A^2$ are distinct, and therefore the eigenvalues of $A$ are distinct, which implies that $A$ is diagonalizable.

If $P_A$ denotes the minimal polynomial of $A$ then $P_{A^2}(X^2)$ is an annulating polynomial for $A$. This means that $P_A|P_{A^2}(X^2)$. If $A^2$ is diagonalizable then $P_{A^2}$ has simple roots. If zero is not root of $P_{A^2}$ then all roots of $P_{A^2}(X^2)$ are simple so all roots of $P_A$ are simple and $A$ is diagonizable. (here we need to be sure that the roots of $P_A$ are all in the field of elements of $A$; if we work on $\Bbb{C}$ then we are fine)

So if zero is not an eigenvalue of $A$ then $A^2$ diagonizable implies $A$ diagonizable.

4

The answer to the second problem is yes. In general, the (in general complex) eigenvalues of $A^2$ are the squares of the eigenvalues of $A$. In this case $A^2$ has the eigenvalues $0$, $\frac{1}{4}$, and $1$. Therefore $0$ is an eigenvalue of $A$, either $\frac{1}{2}$ or $-\frac{1}{2}$ is an eigenvalue of $A$, and either $1$ or $-1$ is an eigenvalue of $A$. This means that $A$ has $3$ distinct real eigenvalues, which implies it is diagonalizable. The key is that the eigenvalues of $A^2$ are distinct and nonnegative, which implies that the eigenvalues of $A$ are distinct and real.

Edit: I accidentally overlapped somewhat with Beni Bogosel's earlier answer. (More was added after I started writing.)

  • 0
    So if $A^2$ had a negative eigenvalue, say $-1$, the same argument would not work? Because then it's possible that $A$ characteristic polynomial has $i$ as one of its roots, which is not an eigenvalue of $A$.2011-07-06
  • 0
    @daniel.jackson: Yes, except you can strengthen your statements. If $A^2$ has a negative eigenvalue, then $A$ necessarily has a nonreal eigenvalue, which implies it is not diagonalizable in $M_n(\mathbb R)$.2011-07-06
  • 0
    Why must it have a nonreal eigenvalue?2011-07-06
  • 0
    It follows from the fact mentioned above (of which I have not provided a proof) that the eigenvalues of $A^2$ are precisely the squares of the eigenvalues of $A$, while squares of real numbers are nonnegative real numbers.2011-07-06
  • 0
    I see. Could you be so kind to provide a proof (a general idea will be fine too)?2011-07-06
  • 0
    Let $\lambda$ be an eigenvalue of $A$. Then $A-\lambda I$ is not invertible, which implies $(A-\lambda I)(A+\lambda I)=A^2-\lambda^2 I$ is not invertible. Therefore $\lambda^2$ is an eigenvalue of $A^2$. Next suppose that neither of the square roots of a complex number $\lambda$, say $\pm w$, is an eigenvalue of $A$. Then $A-w I$ and $A+wI$ are invertible, so $(A-wI)(A+wI)=A^2-\lambda I$ is invertible. This implies that $\lambda$ is not an eigenvalue of $A^2$.2011-07-06
  • 0
    See also the related question http://math.stackexchange.com/questions/28654/spectrum-of-the-operator2011-07-06
1

Try $A=\begin{pmatrix} 0&-1&0 \\ 1&0&0 \\ 0&0&1\end{pmatrix}$.
I found it by considering the complex matrix $\begin{pmatrix} 1&0&0 \\ 0&i&0 \\ 0&0&-i\end{pmatrix}$ and finding its characteristic polynomial, then using rational canonical form over $\mathbb{R}$.

  • 0
    The idea here is that over $\mathbb{C}$, being diagonalizable is equivalent to having distinct eigenvalues. So, over $\mathbb{C}$, your question is true: $A^2$ diagonalizable implies $A$ diagonalizable. So to find a counterexample, you need a diagonal matrix over $\mathbb{C}$ whose eigenvalues are not all real, but whose squares are all real.2011-07-06
  • 4
    uh? having distinct eigenvalues is sufficient but not necessary for being diagonalizable2011-07-06
  • 1
    Ah, yes, sorry, I meant distinct roots in its minimal polynomial.2011-07-06
  • 0
    No, $A^2$ diagonalizable does not imply $A$ diagonalizable. It doesn't matter what field this is over. See Beni Bogosel's answer.2011-07-06
  • 0
    I have been working with invertible matrices the whole time!!2011-07-06
0

To the second question: yes. If all eigenvalues of $A^2$ are pairwise distinct, then so are the eigenvalues of $A$, because if $Av=\lambda v$ then $A^2 v = \lambda ^2 v$, so $\ker(A-\lambda I) \subset \ker(A^2 - \lambda^2 I)$.

A matrix with pairwise distinct eigenvalues is diagonalizable. That's a simple and well known theorem.