Linear Algebra › Section E · Eigenvalues and eigenvectors › Topic E.5

Complex eigenvalues

A rotation turns every vector off its line, so it has no real eigenvectors. Allowing complex numbers recovers the eigenvalues, and their size and angle reveal a hidden rotation-and-scaling inside any real 2 × 2 matrix of this kind.

When real eigenvectors run out

The 90° rotation of the plane has standard matrix

R = \begin{bmatrix} 0 & -1 \\ 1 & 0 \end{bmatrix}

Its trace is 0 and its determinant is 1, so the characteristic equation from "The characteristic equation" is \(\lambda^2 + 1 = 0\), which has no real solution. Geometrically this is no surprise: every non-zero vector is turned through a right angle, so none stays on its own line. Over the complex numbers, though, the equation has two roots, \(\lambda = i\) and \(\lambda = -i\). The plan of this topic is to take those complex eigenvalues seriously and translate what they say back into real geometry.

Complex vectors and the arithmetic you need

A complex number is \(z = a + bi\) with \(i^2 = -1\); its conjugate is \(\bar{z} = a - bi\) and its modulus is \(|z| = \sqrt{a^2 + b^2}\), its distance from 0 in the complex plane. \(\mathbb{C}^n\) is the set of vectors with \(n\) complex entries, and matrices act on them by the usual rules, with complex scalars allowed. Every complex vector splits into a real part and an imaginary part: for \(\mathbf{v} = (1 - 2i, 1)\), \(\operatorname{Re}\mathbf{v} = (1, 1)\) and \(\operatorname{Im}\mathbf{v} = (-2, 0)\), so \(\mathbf{v} = \operatorname{Re}\mathbf{v} + i\operatorname{Im}\mathbf{v}\). The definitions of eigenvalue and eigenvector carry over unchanged.

Eigenvectors of the rotation

For \(\lambda = i\), row reduce \(R - iI\), whose rows are \((-i, -1)\) and \((1, -i)\). The second row says \(x_1 = ix_2\), and the first row is \(-i\) times the second, so it adds nothing new. Taking \(x_2 = -i\) gives \(x_1 = 1\) and the eigenvector \((1, -i)\). Check: \(R(1, -i) = (i, 1)\), and \(i(1, -i) = (i, -i^2) = (i, 1)\). For a 2 × 2 matrix \(A - \lambda I\) with a complex eigenvalue, the two rows are always complex multiples of each other, so one row is enough to read off the eigenvector. This shortcut saves a lot of complex arithmetic.

Conjugate pairs

When \(A\) has only real entries, its complex eigenvalues come in conjugate pairs, and so do their eigenvectors. Take the conjugate of both sides of \(A\mathbf{v} = \lambda\mathbf{v}\). Since \(\bar{A} = A\), this gives

A\bar{\mathbf{v}} = \bar{\lambda}\bar{\mathbf{v}}

So once you have one eigenpair, the other comes free: for the rotation, \(-i\) has eigenvector \((1, i)\). This also matches the characteristic polynomial, whose real coefficients force any non-real roots to pair up as \(a \pm bi\). The sum and product checks still hold: \(i + (-i) = 0\) is the trace, and \(i \cdot (-i) = 1\) is the determinant.

Rotation and scaling in one matrix

Matrices of the form below are the model for complex eigenvalues.

C = \begin{bmatrix} a & -b \\ b & a \end{bmatrix}

Its characteristic equation is \((a - \lambda)^2 + b^2 = 0\), so its eigenvalues are \(a \pm bi\). Let \(r = \sqrt{a^2 + b^2}\), the modulus of the eigenvalues, and let \(\varphi\) be the angle with \(\cos\varphi = a/r\) and \(\sin\varphi = b/r\). Factoring out \(r\) leaves a rotation matrix:

C = r\begin{bmatrix} \cos\varphi & -\sin\varphi \\ \sin\varphi & \cos\varphi \end{bmatrix}

So \(C\) rotates every vector through \(\varphi\) and scales it by \(r\). A pure rotation through \(\theta\) has \(r = 1\) and eigenvalues \(\cos\theta \pm i\sin\theta\).

Worked example: finding a complex eigenpair

Find the eigenvalues and an eigenvector of

A = \begin{bmatrix} 3 & -5 \\ 1 & 1 \end{bmatrix}

1. Trace 4 and determinant \(3 + 5 = 8\) give \(\lambda^2 - 4\lambda + 8 = 0\). 2. The quadratic formula gives \(\lambda = \frac{4 \pm \sqrt{16 - 32}}{2} = 2 \pm 2i\). 3. For \(\lambda = 2 - 2i\), the second row of \(A - \lambda I\) is \((1, -1 + 2i)\), so \(x_1 = (1 - 2i)x_2\), and \(x_2 = 1\) gives \(\mathbf{v} = (1 - 2i, 1)\). 4. Check the first entry: \(3(1 - 2i) - 5 = -2 - 6i\), and

(2 - 2i)(1 - 2i) = 2 - 6i + 4i^2 = -2 - 6i

5. By conjugation, \(2 + 2i\) has eigenvector \((1 + 2i, 1)\).

The real factorisation

A real 2 × 2 matrix with complex eigenvalues cannot be diagonalised with real numbers, but it is always similar to a rotation-scaling matrix. Let \(A\) have eigenvalue \(\lambda = a - bi\) with \(b \ne 0\) and eigenvector \(\mathbf{v}\). Then

A = PCP^{-1}

where \(P\) has columns \(\operatorname{Re}\mathbf{v}\) and \(\operatorname{Im}\mathbf{v}\), and \(C\) is the matrix of the previous card built from that same \(a\) and \(b\). The signs matter: the theorem is stated for \(\lambda = a - bi\), so read \(b\) from the eigenvalue whose eigenvector you used. In the coordinates given by the columns of \(P\), the action of \(A\) is a rotation followed by a scaling.

Worked example: rotation hidden in a matrix

Continue with \(A\) from the previous worked example, using \(\lambda = 2 - 2i\), so \(a = 2\) and \(b = 2\), and \(\mathbf{v} = (1 - 2i, 1)\). 1. \(\operatorname{Re}\mathbf{v} = (1, 1)\) and \(\operatorname{Im}\mathbf{v} = (-2, 0)\), so \(P\) has columns (1, 1) and (-2, 0). 2. \(C\) has rows (2, -2) and (2, 2). 3. Check \(AP = PC\): both have columns (-2, 2) and (-6, -2). 4. Interpret: \(r = \sqrt{2^2 + 2^2} = 2\sqrt{2} \approx 2.83\), and \(\cos\varphi = \sin\varphi = \frac{1}{\sqrt{2}}\), so \(\varphi = 45°\). In the new coordinates, \(A\) turns every vector through 45° and stretches it by about 2.83.

What repeated multiplication does

Because \(A = PCP^{-1}\), the powers are \(A^k = PC^kP^{-1}\), and \(C^k\) is a rotation through \(k\varphi\) with scaling \(r^k\). So the points \(\mathbf{x}_k\) of \(\mathbf{x}_{k+1} = A\mathbf{x}_k\) turn around the origin, spiralling outwards if \(r > 1\), inwards if \(r < 1\), and staying on a closed ellipse-shaped curve if \(r = 1\). For the matrix above, \(r = 2\sqrt{2}\): starting from (1, 0), the iterates (3, 1), (4, 4) and (-8, 8) wind round the origin and move steadily outwards. Measured in the coordinates given by \(P\), each step multiplies the length by exactly \(2\sqrt{2}\); in ordinary coordinates the ratio wobbles from step to step. "Discrete dynamical systems" uses exactly this picture.

Common mistakes to avoid

First, deciding that a matrix has no eigenvalues because the characteristic equation has no real roots; the eigenvalues are complex. Second, using \(b\) from \(a + bi\) while taking the eigenvector of \(a - bi\): the factorisation flips sign and fails, so keep the eigenvalue and its eigenvector together. Third, taking the columns of \(P\) in the wrong order; it is the real part first, then the imaginary part. Fourth, judging spiral behaviour by the real part alone; it is the modulus \(r = |\lambda|\) that decides growth or decay.

Key takeaways

A real matrix can have complex eigenvalues, found from the same characteristic equation. They come in conjugate pairs \(a \pm bi\), and the eigenvector for \(\bar{\lambda}\) is \(\bar{\mathbf{v}}\). The matrix with rows \((a, -b)\) and \((b, a)\) rotates through the angle \(\varphi\) and scales by \(r = \sqrt{a^2 + b^2}\). A real 2 × 2 matrix with eigenvalue \(a - bi\) and eigenvector \(\mathbf{v}\) factors as \(PCP^{-1}\), where \(P\) has columns \(\operatorname{Re}\mathbf{v}\) and \(\operatorname{Im}\mathbf{v}\). Repeated multiplication spirals outwards when \(r > 1\) and inwards when \(r < 1\).

Flashcards

Flashcard

90° rotation

Eigenvalues i and −i

λ² + 1 = 0 has no real roots.

Test your understanding

Question 1 of 5

What are the eigenvalues of the matrix with rows (1, −2) and (2, 1)?

Question 2 of 5

A real 2 × 2 matrix has eigenvalue 3 − 4i with eigenvector (1 + i, 2). Which is another eigenvalue with an eigenvector?

Question 3 of 5

The matrix C has rows (3, −4) and (4, 3). It acts as a rotation combined with scaling by which factor?

Question 4 of 5

A real 2 × 2 matrix A has eigenvalue λ = 1 − 3i with eigenvector v = (2 + i, 1). Which choice gives A = PCP⁻¹?

Question 5 of 5

In xₖ₊₁ = Axₖ, the real 2 × 2 matrix A has eigenvalues 0.6 ± 0.8i. What happens to the points xₖ?

← Previous topicNext topic →