Diagonalisable and Non-Diagonalisable Matrices

Not all square matrices can be diagonalised. For example, consider the matrix

A=\left(\begin{array}{ccc}0&-6&-4\\5&-11&-6\\-6&9&4\end{array}\right).

Its eigenvalues are -2, -2 and -3. Now, it's certainly possible to find a matrix S with the property that

A\,S\,=\,S\,D,

where D is the diagonal matrix of eigenvalues. One such is

\left(\begin{array}{ccc}0&0&2\\0&2&-1\\0&-3&3\end{array}\right);

it's easy to check that

\left(\begin{array}{ccc}0&-6&-4\\5&-11&-6\\-6&9&4\end{array}\right)\left(\begin{array}{ccc}0&0&2\\0&2&-1\\0&-3&3\end{array}\right)=\left(\begin{array}{ccc}0&0&2\\0&2&-1\\0&-3&3\end{array}\right)\left(\begin{array}{ccc}-2&0&0\\0&-2&0\\0&0&-3\end{array}\right)=\left(\begin{array}{ccc}0&0&-6\\0&-4&3\\0&6&-9\end{array}\right).

However, the trouble is that S is singular. It turns out that there is no non-singular S with the property that

A\,S\,=\,S\,D,

and therefore no S with the property that

A\,=\,S\,D\,S^{-1}.

You may have spotted that two of the eigenvalues of A were repeated, and you may be wondering whether this has anything to do with why A can't be diagonalised. Indeed it does, but it can't be the whole story. For example, consider the matrix

B=\left(\begin{array}{ccc}4&8&-2\\-3&-6&1\\9&12&-5\end{array}\right).

Like A, B has two repeated eigenvalues: the eigenvalues of B are also -2, -2 and -3. However, B can certainly be diagonalised. It's not hard to show that if

S=\left(\begin{array}{ccc}4&1&2\\-3&0&-1\\0&3&3\end{array}\right),

then

B\,S=S\,D=\left(\begin{array}{ccc}-8&-2&-6\\6&0&3\\0&-6&-9\end{array}\right).

Moreover, S is not singular (in fact, \text{det}(S)=-3), and therefore S^{-1} exists, meaning that

B=S\,D\,S^{-1}.

The mystery deepens slightly when we realise that our choice of S here was far from unique. We could just as easily, for example, have used the matrix

S=\left(\begin{array}{ccc}5&3&2\\-3&-3&-1\\3&-3&3\end{array}\right).

So what's going on here? What is the difference between A and B? How come they have the same eigenvalues, each with one repeat, and yet A isn't diagonalisable yet B is?

The answer is revealed when obtain the eigenvectors of the two matrices.

The eigenvectors of A

(i) For the eigenvalue -3, we have

\left(\begin{array}{ccc}3&-6&-4\\5&-8&-6\\-6&9&7\end{array}\right)\left(\begin{array}{c}X\\Y\\Z\end{array}\right)=\left(\begin{array}{c}0\\0\\0\end{array}\right),

which straightforwardly gives the eigenvector

\left(\begin{array}{c}2\\-1\\3\end{array}\right).

(ii) For the repeated eigenvalue -2, we have

\left(\begin{array}{ccc}2&-6&-4\\5&-9&-6\\-6&9&6\end{array}\right)\left(\begin{array}{c}X\\Y\\Z\end{array}\right)=\left(\begin{array}{c}0\\0\\0\end{array}\right),

which equally straightforwardly gives the eigenvector

\left(-\begin{array}{c}0\\2\\3\end{array}\right).

And that's it. We can't go any further, because there are no more eigenvalues. So we don't have three eigenvectors with which to form the diagonalising matrix S.

Contrast this with B.

The eigenvectors of B

(i) For the eigenvalue -3, we have

\left(\begin{array}{ccc}7&8&-2\\-3&-3&1\\9&12&-2\end{array}\right)\left(\begin{array}{c}X\\Y\\Z\end{array}\right)=\left(\begin{array}{c}0\\0\\0\end{array}\right),

which, as before, straightforwardly gives the eigenvector

\left(\begin{array}{c}2\\-1\\3\end{array}\right).

(ii) This time, for the repeated eigenvalue -2, we have

\left(\begin{array}{ccc}6&8&-2\\-3&-4&1\\9&12&-3\end{array}\right)\left(\begin{array}{c}X\\Y\\Z\end{array}\right)=\left(\begin{array}{c}0\\0\\0\end{array}\right).

Now, here things are different, because all three of the rows of this matrix may be reduced to the equation

3X+4Y-Z=0.

This represents a plane in 3D space, and any vector in this plane is an eigenvector. We may therefore form our diagonalising matrix S out of

\left(\begin{array}{c}2\\-1\\3\end{array}\right)

together with any two non-parallel vectors of the form

\left(\begin{array}{c}X\\Y\\Z\end{array}\right)

that satisfy

3X+4Y-Z=0;

that is, that are perpendicular to the vector

\left(\begin{array}{c}3\\4\\-1\end{array}\right).

Both of the choices

S=\left(\begin{array}{ccc}4&1&2\\-3&0&-1\\0&3&3\end{array}\right),
S=\left(\begin{array}{ccc}5&3&2\\-3&-3&-1\\3&-3&3\end{array}\right)

will work fine, as will infinitely many others.

General considerations

1. In general, any 3 by 3 matrix whose eigenvalues are distinct can be diagonalised.

2. If there is a repeated eigenvalue, whether or not the matrix can be diagonalised depends on the eigenvectors.

(i) If there are just two eigenvectors (up to multiplication by a constant), then the matrix cannot be diagonalised.

(ii) If the unique eigenvalue corresponds to an eigenvector {\bf e}, but the repeated eigenvalue corresponds to an entire plane, then the matrix can be diagonalised, using {\bf e} together with any two vectors that lie in the plane.

3. If all three eigenvalues are repeated, then things are much more straightforward: the matrix can't be diagonalised unless it's already diagonal.

Created with the ExportAsWebPage package in Wolfram Mathematica 7.0