Attention
In this section we will assume the reader is familiar with complex numbers and their properties. If preferred, a review of complex numbers can be found in Chapter 10.
6.4. Complex eigenvalues and eigenvectors#
6.4.1. Introduction#
In the previous sections we hinted at the possibility to allow eigenvalues to be complex numbers. For an \(n\times n\)-matrix \(A\) the eigenvalues are the zeros of the characteristic polynomial \(p_A(\lambda)\) of \(A\). Even if the matrix is real, these zeros may be complex. We start with an example to explore this until now unknown territory.
Example 6.4.1
Consider the matrix \(A = \left(\begin{array}{cc} 1 & -2 \\ 1 & 3 \end{array}\right)\).
We first compute the characteristic polynomial
Completing the square we see that
So the eigenvalues are complex numbers. Can we still find eigenvectors? We sure can, but they will not be real vectors. To find an eigenvector \(\vect{v}_1\) for \(\lambda_1 = 2+i\), as before we have
So we have to solve a homogeneous system of linear equations, where now the coefficient matrix contains complex numbers. This slightly complicates the computation, but for this \(2 \times 2\)-matrix things don’t get too bad.
where the first row operation we invoke is: add the second row \((1+i)\) times to the first row. The blue \(0\) is the result of the evaluation of
We can read off a solution (i.e., complex eigenvector): \(\vect{v}_1 = \left(\begin{array}{c} -1+i \\1 \end{array}\right)\).
To check that we have indeed an eigenvector is also slightly more involved than in the real case.
and that agrees with
Alternatively, to simplify the augmented matrix
we could have used the entry \(-2\) in the first row of the augmented matrix to create a zero in the second row. More specifically, if we add the first row \(\dfrac{1-i}{2}\) times to the second row, we get
Then the vector \(\vect{u}_1 = \left(\begin{array}{c} -2 \\ 1+i \end{array}\right) \) is a natural candidate for an eigenvector.
At first sight it seems that we have found two linearly independent eigenvectors for the eigenvalue \(\lambda_1 = 2+i\). However, closer inspection shows that
so the two vectors are complex multiples of each other, and hence are not linearly independent.
For the other eigenvalue we can proceed in the same manner and find (for instance) the eigenvector \(\vect{u}_2 =\left(\begin{array}{c} -2 \\ 1-i \end{array}\right) \).
Note that the two eigenvalues are each others complex conjugate, and that the same holds for the corresponding eigenvectors, that is, if we define the complex conjugate of a vector component wise.
Remark 6.4.1
Row reduction of the augmented matrix
from Example 6.4.1 can even be performed in a third way: first multiply the first row by the complex conjugate of the pivot in the first row:
This first row operation guarantees that the first column only contains real numbers, which makes the next row operation easier to perform, since we can now use the first row to create a zero in the second row.
6.4.2. Vectors and matrices with complex entries#
We need a few definitions to settle matters a bit more formally. In the remainder of this section matrices (so in particular vectors) are allowed to have complex numbers as entries. If the entries are supposed to be real numbers we explicitly state this by speaking of a real matrix (or a real vector).
Definition 6.4.1
The complex conjugate \(\overline{A}\) of a matrix \(A\) is defined component wise:
if \(\quad A = \left(\begin{array}{ccc} a_{11} & \cdots & a_{1n} \\ \vdots & &\vdots \\ a_{m1} & \cdots & a_{mn} \end{array}\right) \quad\) then \(\quad\overline{A} = \left(\begin{array}{ccc} \overline{a_{11}} & \cdots & \overline{a_{1n}} \\ \vdots & &\vdots \\ \overline{a_{m1}} & \cdots & \overline{a_{mn}} \end{array}\right) \).
Proposition 6.4.1
-
If \(A\) and \(B\) are two \(m\times n\)-matrices, then \(\overline{A+B} = \overline{A}+\overline{B}\).
-
If \(A\) and \(C\) are two matrices for which the product \(AC\) exists,
then \(\overline{AC} = \overline{A}\) \(\overline{C}\).
Proof of Proposition 6.4.1
The statements follow immediately from the definitions of the sum and product of two matrices, and of the corresponding rules of complex arithmetic that say
With this we can put the outcomes in Example 6.4.1 in a broader perspective.
Proposition 6.4.2
Suppose \(A\) is a real matrix, and \(\lambda = \alpha + \beta i\), with \(\beta \neq 0\), is an eigenvalue of \(A\). Then the following properties hold:
-
\(\overline{\lambda} = \alpha - \beta i\) is an eigenvalue too.
-
If \(\vect{v} = \vect{u}+i\vect{w}\), where \(\vect{u}\) and \(\vect{w}\) are real vectors, is an eigenvector for \(\lambda\), then \(\overline{\vect{v}} = \vect{u}-i\vect{w}\) is an eigenvector for \(\overline{\lambda}\).
Proof of Proposition 6.4.2
Suppose \(A(\vect{u}+i\vect{w}) = (\alpha + \beta i)(\vect{u}+i\vect{w})\).
Complex conjugation yields that
since \(A\) is a real matrix, and \(\vect{u}\) and \(\vect{v}\) are supposed to be real. At the same time
Equating the last expressions of both derivations yields the desired result.
This states exactly that \(A\) has the eigenvalue \(\overline{\lambda} = \alpha - \beta i\) with the corresponding eigenvector \(\vect{u}-i\vect{w}\).
Things look especially simple in the next example.
Example 6.4.2
Let \(A\) be the matrix \(A= \left(\begin{array}{cc} a & -b \\ b & a \end{array}\right)\), with \(b \neq 0\).
The characteristic polynomial of \(A\) is then given by \(p_A(\lambda) = (a-\lambda)^2 + b^2\).
So \(A\) has the complex eigenvalues
From row reduction of the augmented matrix
we see that \(\vect{v} = \left(\begin{array}{c} 1 \\ -i \end{array}\right)\) is an eigenvector for \(\lambda_1 = a+bi\). By taking conjugates (according to Proposition 6.4.2) we can conclude that \(\vect{w} = \left(\begin{array}{c} 1 \\ i \end{array}\right)\), i.e. the conjugate of \(\vect{v}\), will be an eigenvector for \(\lambda_2= a-bi\).
There is a nice geometric interpretation for matrices of the form \(\left(\begin{array}{cc} a & -b \\ b & a \end{array}\right)\).
Or rather, for the corresponding linear transformation.
Proposition 6.4.3
The linear transformation \(T:\R^2\to\R^2\) given by
can be described as a rotation followed by a “scaling”.
In fact, it holds that
Proof of Proposition 6.4.3
Both columns of \(A\) have length \(r = \sqrt{a^2 + (\pm b)^2} = \sqrt{a^2 + b^2}\).
If we take out this factor, we get
The first column \(\left(\begin{array}{c} a/r \\ b/r \end{array}\right)\) has length one, so it lies on the unit circle, and hence must be of the form
and this makes that
So the ‘action’ of the matrix \(A = \left(\begin{array}{cc} a & -b \\ b & a \end{array}\right)\) on a vector \(\vect{x}\) is a (anticlockwise) rotation over the angle \(\varphi\) followed by a scaling (stretching/dilatation) with a factor \(r\).
The scaling factor \(r\) and the angle \(\varphi\) are exactly the polar coordinates of the eigenvalue \(\lambda = a + bi\). That is,
where
Proposition 6.4.3 can be generalised as follows. If a real \(n \times n\)-matrix \(A\) has a non-real eigenvalue, there is always a rotation ‘hidden’ in the transformation \(T: \vect{x} \mapsto A\vect{x}\).
Proposition 6.4.4
Suppose the real \(n \times n\)-matrix \(A\) has a complex eigenvalue \(\lambda = \alpha - \beta i\), with \(\beta \neq 0\). Then there exist two linearly independent real vectors \(\vect{u}\) and \(\vect{w}\) for which
That means that the two-dimensional subspace \(S = \Span{\vect{u},\vect{w}}\) is invariant under the linear transformation \(T\) that has \(A\) as a standard matrix.
Proof of Proposition 6.4.4
Let \(\vect{v}\) be an eigenvector for \(\lambda=\alpha - \beta i\). So, \(\vect{v} = \vect{u}+i\vect{w}\), for real vectors \(\vect{u}\) and \(\vect{w}\). Note that \(\vect{v}\) cannot be a real vector.
\(\vect{u}\) and \(\vect{w}\) must be linearly independent, because if \(\vect{w}= \vect{0}\), then \(\vect{v}\) would be a real eigenvector, and if \(\vect{u}= k\vect{w}\) for some \(k\) in \(\R\), then \(\vect{v} = (k+i)\vect{w}\), and then \(\vect{w}\) would be a real eigenvector for the complex eigenvalue \(\lambda\) of the real matrix \(A\). And that is unheard of.
So we conclude that \(\vect{u}\) and \(\vect{w}\) must be linearly independent.
We now have
Comparing real and imaginary parts we conclude that indeed
If we apply the above to the case \(n = 2\), we can rewrite Equation (6.4.1) as
So if we define \(P\) to be the matrix \((\,\vect{u}\,\, \vect{w}\,)\) then we have \(AP = PC\), where
This more or less settles the following proposition.
Proposition 6.4.5
Suppose \(A\) is a \(2 \times 2\)-matrix with eigenvalues \(\alpha \pm \beta i\), with \(\beta \neq 0\).
Then \(A\) can be written as
The matrix \(P\) is of the form
where \(\vect{u} + i \vect{w}\) is an eigenvector for \(\lambda = \alpha - \beta i\).
Remark 6.4.2
Proposition 6.4.3 states that \(C\) can be written as
We can interpret \(\left(\begin{array}{cc} \cos(\varphi) & -\sin(\varphi) \\ \sin(\varphi) & \cos(\varphi) \end{array}\right)\) as a hidden rotation in \(A\).
Let us illustrate matters with the following example.
Example 6.4.3
The matrix \(A = \left(\begin{array}{cc} 1 & -2 \\ 1 & 3 \end{array}\right)\) of Example 6.4.1 has the eigenvalues \(\lambda_{1,2} = \alpha \pm \beta i = 2 \pm i\). An eigenvector for \(2-i\) is given by
Let us establish the identity
According to Proposition 6.4.5 we can take
with
We then get
We conclude this section by reconsidering diagonalisability of real matrices if we allow complex eigenvalues.
6.4.3. Complex diagonalisability#
Let us first generalise Definition 6.3.2 to the complex case:
Definition 6.4.2
A matrix is \(A\) is called complex diagonalisable if it can be written in the form
where \(D\) is a diagonal matrix, and \(P\) and \(D\) may contain complex entries. We then say that \(PDP^{-1}\) is a complex diagonalisation of \(A\).
Remark 6.4.3
Note that if a matrix \(A\) has a diagonalisation from Definition 6.3.2, then it is automatically a complex diagonalisation, since the entries of \(P\) and \(D\) are real numbers, which are also complex numbers.
Just like in the real case diagonalisability has all to do with the existence of enough (possibly complex) eigenvectors. The derivation is the same as in Section 6.3, we only repeat the conclusion.
Proposition 6.4.6
An \(n\times n\)-matrix is complex diagonalisable if and only if there exists a basis of eigenvectors for \(\mathbb{C}^n\).
In that case, if \(\vect{v}_1, \ldots, \vect{v}_n\) are \(n\) linearly independent eigenvectors for the eigenvalues \(\lambda_1, \ldots, \lambda_n\), then
Example 6.4.4
The matrix \(A = \left(\begin{array}{cc} 1 & -2 \\ 1 & 3 \end{array}\right)\) of Example 6.4.1 has the eigenvalues \(\lambda_1 = 2 + i\), \(\lambda_2 = 2 - i\), with corresponding eigenvectors \(\vect{v}_1 = \left(\begin{array}{c} -1+i \\1 \end{array}\right)\), \(\vect{v}_2 = \left(\begin{array}{c} -1-i \\1 \end{array}\right)\).
It follows that
Which you are challenged to check by a careful calculation.
Remark 6.4.4
-
When the context requires it, we will specify whether we mean diagonalisable or complex diagonalisable.
-
The definition also makes sense for matrices with complex numbers as entries. However, we will not pursue that track.
For a matrix to be diagonalisable Theorem 6.3.1 states that two conditions must be satisfied. One is that the characteristic polynomial of \(A\) must have \(n\) real roots, counting multiplicities. The second is that for each eigenvalue the geometric multiplicity must be equal to the algebraic multiplicity. The Fundamental Theorem of Algebra guarantees that each polynomial of degree \(n\) has \(n\) roots. So if we allow complex eigenvalues, the first condition is automatically satisfied. We thus find the following criterion for complex diagonalisability of a matrix \(A\).
Proposition 6.4.7
A matrix \(A\) is complex diagonalisable if and only if for each eigenvalue the geometric multiplicity is equal to the algebraic multiplicity.
6.4.4. Grasple exercises#
Grasple Exercise 6.4.1
Given a complex eigenvector of a \(2\times 2\)-matrix, find the corresponding eigenvalue.
Click to show/hide
Grasple Exercise 6.4.2
Given a complex eigenvector of a \(2\times 2\)-matrix, find the corresponding eigenvalue.
Click to show/hide
Grasple Exercise 6.4.3
To find the (complex) eigenvalues of a \(2 \times 2\)-matrix \(A\).
Click to show/hide
Grasple Exercise 6.4.4
To find the (complex) eigenvalues of a \(2 \times 2\)-matrix \(A\).
Click to show/hide
Grasple Exercise 6.4.5
To find the (complex) eigenvalues and bases of their eigenspaces of a \(2 \times 2\)-matrix \(A\).
Click to show/hide
Grasple Exercise 6.4.6
Given a complex eigenvector of a real \(2 \times 2\)-matrix, point out other eigenvectors.
Click to show/hide
Grasple Exercise 6.4.7
Given a complex eigenvector of a real \(2 \times 2\)-matrix, point out other eigenvectors.
Click to show/hide
Grasple Exercise 6.4.8
To find an invertible matrix \(P\) and a scaling-rotation matrix \(C\) for which \(A = PCP^{-1}\).
Click to show/hide
Grasple Exercise 6.4.9
To find the complex eigenvalues of a \(2 \times 2\)-matrix (of a special form).
Click to show/hide
Grasple Exercise 6.4.10
True or false? Every real \(5 \times 5\)-matrix has at least one real eigenvalue.
Click to show/hide
Grasple Exercise 6.4.11
To find a complex diagonalisation of a real \(2\times 2\)-matrix (eigenvalues given).
Click to show/hide
Grasple Exercise 6.4.12
To find a complex diagonalisation of a real \(3\times 3\)-matrix (one eigenvalue given).
Click to show/hide
Grasple Exercise 6.4.13
To mark all correct statements about real/complex eigenvalues/eigenvectors of a real \(2\times 2\)-matrix.