2
$\begingroup$

I know that a symmetric matrix with positive eigenvalues (i.e. a positive-definite matrix) is diagonalizable. But what if the matrix is not symmetric? Can we still diagonalize it? i.e. if $A$ is a matrix with all positive eigenvalues, can we write it as $A=P^{-1}DP$ where $D$ is a diagonal matrix with eigenvalues on the diagonal?

$\endgroup$
  • 4
    $\begingroup$ Nope. Example: $\begin{pmatrix}1 & 1\\0 & 1\end{pmatrix}$. $\endgroup$ – amsmath Oct 10 '17 at 15:31
  • 2
    $\begingroup$ It is true, however, that every symmetric matrix is diagonalizable. $\endgroup$ – Omnomnomnom Oct 10 '17 at 15:34
  • 2
    $\begingroup$ And if the eigenvalues are distinct, the matrix can always be diagonalized as well. $\endgroup$ – Peter Oct 10 '17 at 15:35
  • 2
    $\begingroup$ Yes, $(I + AB)$ is similar to $I + B^{1/2} A B^{1/2}$, which is self adjoint, so diagonalizable. So $I + A B$ is also diagonlizable. $\endgroup$ – Orest Bucicovschi Oct 10 '17 at 16:03
  • 2
    $\begingroup$ @Mah Note that adding $I$ to a matrix has absolutely no effect on diagonalizabilty, so your particular case boils down to diagonalizing $AB$ and has nothing to do with positivity of the eigenvalues. $\endgroup$ – Erick Wong Oct 10 '17 at 20:55
1
$\begingroup$

The answer is no. As pointed out by amsmath in his comment, the matrix

$A = \begin{bmatrix} 1 & 1 \\ 0 & 1 \end{bmatrix} \tag 1$

cannot be diagonalized, and a wider class of examples is provided by size $n$ square matrices of the form

$B = \lambda I + T, \tag 2$

where $I$ is the $n \times n$ identity matrix, and $T \ne 0$ is an $n \times n$ strictly upper triangular matrix, that is, $T$ takes the form

$T = [t_{ij}] \ne 0, \tag 3$

where

$t_{ij} = 0, \; i \le j. \tag 4$

We note that

$T^n = 0, \tag 5$

and this implies that the only eigenvalue of $B$ is $\lambda$, since if

$B \vec v = \mu \vec v \tag 6$

for some $\vec v \ne 0$, by (2) we would have

$(\mu - \lambda) \vec v= (B - \lambda I)\vec v = T\vec v, \tag 7$

and this in turn yields

$(\mu - \lambda)^n \vec v = (B - \lambda)^n \vec v = T^n \vec v = 0; \tag 8$

since $\vec v \ne 0$ we conclude that

$(\mu - \lambda)^n = 0, \tag 9$

whence

$\mu = \lambda. \tag{10}$

Since $\lambda$ is the only eigenvalue of $B$, if some matrix $P$ diagonalized $B$ we would have

$\Lambda = PBP^{-1}, \tag{11}$

$\Lambda$ being the diagonal matrix

$\Lambda = \lambda I; \tag {12}$

inserting $B$ as in (2) into (11) yields

$\Lambda = P(\lambda I + T)P^{-1} = P(\lambda I)P^{-1} + PTP^{-1}$ $= \lambda PIP^{-1} + PTP^{-1} = \lambda I + PTP^{-1} = \Lambda + PTP^{-1}, \tag {14}$

or

$PTP^{-1} = 0, \tag{15}$

so that in fact

$T = 0. \tag{16}$

We have proved that if a matrix of the form (2) is diagonalizable then $T = 0$; thus, by contraposition, $T \ne 0$ implies $B$ cannot be diagonalied; in this way a great many instances of non-diagonalizable matrices may be constructed. Note that we do not require $\lambda > 0$.

Of course, if the eigenvalues of any matrix are all distinct, then it may be diagonalized, as is well-known.

Finally, it appears upon scrutiny of the above argument that we may relax the condition that $T$ is strictly upper triangular to the more general hypothesis that $T$ is nilpotent; that is

$T^m = 0 \tag{17}$

for some positive integer $m$. The essential details of the proof are left undisturbed by this assumption.

$\endgroup$

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.