Linear systems with Complex Eigenvalues

Complex vectors

Definition

When the matrix $A$ of a system of linear differential equations \begin{equation} \dot\vx = A\vx \label{eq:linear-system} \end{equation} has complex eigenvalues the most convenient way to represent the real solutions is to use complex vectors. A complex vector is a column vector \[ \vv = \begin{bmatrix} v_1\\\vdots\\v_n \end{bmatrix} \] whose entries $v_k$ are complex numbers. Every complex vector can be written as $\vv = \va+i\vb$ where $\va$ and $\vb$ are real vectors. To do this write each entry $v_k=a_k+ib_k,$ with $a_k$ and $b_k$ the real and imaginary parts of $v_k$, and split the vector $\vv$ as follows: \begin{equation} \vv = \begin{bmatrix} v_1\\ \vdots\\ v_n \end{bmatrix} = \begin{bmatrix} a_1+ib_1\\ \vdots\\ a_n+ib_n \end{bmatrix} = \begin{bmatrix} a_1\\ \vdots\\ a_n \end{bmatrix} + i \begin{bmatrix} b_1\\ \vdots\\ b_n \end{bmatrix} \end{equation} The vectors $\va$ and $\vb$ are the real and complex parts of the vector $\vv=\va+i\vb$.

Complex conjugate

Every complex number $z=a+bi$ has a complex conjugate, \[ \bar z = a-bi. \] Addition and multiplication of complex numbers behaves nicely with respect to complex conjugation: \begin{equation} \overline{z+w} = \bar z + \bar w, \qquad \overline{zw} = \bar z\, \bar w,\qquad \overline{\left(\frac zw\right)} = \frac{\bar z}{\bar w}. \end{equation} The length or absolute value $|z| = \sqrt{a^2+b^2}$ of a complex number $z=a+bi$ satisfies \begin{equation} |z|^2 = z\bar z, \qquad \frac{1}{z} = \frac{\bar z}{|z|^2}. \end{equation} The following formulas allow us to recover the real and imaginary parts of a complex number $z=a+bi$ from $z$ and its complex conjugate: \begin{equation} z=a+bi \iff a=\frac{z+\bar z}{2} \text{ and } \quad b=\frac{z-\bar z}{2i}. \end{equation} By analogy we define the complex conjugate of a complex vector $\vv=\va+i\vb$ to be \[ \bar \vv = \overline{\va+i\vb} = \va - i\vb. \] Then we have \begin{equation} \overline{\vv+\vw} = \bar\vv + \bar \vw,\qquad \overline{z\vv} = \bar z \, \bar \vv \end{equation} for any complex vectors $\vv$, $\vw$, and any complex number $z$.

Complex conjugates of complex exponentials

By definition the exponential of a complex number $z=a+bi$ is \[ e^{a+bi} = e^{a} \bigl(\cos b + i \sin b\bigr). \] Replacing $b$ by $-b$, and using that $\cos(-b) = \cos b$, $\sin(-b)=-\sin b$, leads to \[ e^{a-bi} = e^{a} \bigl(\cos b - i \sin b\bigr). \] Thus for any complex number $z=a+bi$ one has \[ e^{\bar z} = \overline{e^z}. \]

Real matrices and complex vectors

Multiplying complex matrices

If $A$ is any matrix, possibly with complex entries, and $\vv$ is a complex vector, then one defines the matrix product $A\vv$ by exactly the same formula as in the real case, namely, \[ \begin{bmatrix} a_{11} & \dots & a_{1n} \\ \dots & & \vdots \\ a_{n1} & \dots & a_{nn} \end{bmatrix} \begin{bmatrix} v_1 \\ \vdots \\ v_n \end{bmatrix} = \begin{bmatrix} a_{11}v_1 + \cdots + a_{1n}v_n \\ \vdots \\ a_{n1}v_1 + \cdots + a_{nn}v_n \end{bmatrix}. \] If $A$ is a real matrix, i.e. a matrix all of whose entries $a_{kl}$ are real numbers, and if the complex vector is $\vv = \va+i\vb$, then the matrix product $A\vv$ satisfies \[ A\vv = A(\va+i\vb) =A\va + i A\vb. \] From this it follows that \[ \overline{A\vv} = A\bar\vv \]

Complex eigenvalues and eigenvectors of real matrices

Let $A$ be a real matrix. If $\lambda$ is a complex eigenvalue of $A$, then the corresponding eigenvectors will also be complex.
Theorem. If $\lambda$ is a complex eigenvalue of the real matrix $A$, and if $\vv$ is a corresponding complex eigenvector, then $\bar \lambda$ is also an eigenvalue, and \[ A\bar \vv = \bar \lambda\, \bar\vv, \] i.e. $\bar\vv$ is an eigenvector corresponding to the eigenvalue $\bar \lambda$.
Proof. This follows from \( A\bar\vv = \overline{A\vv} = \overline{\lambda\vv} = \bar \lambda \, \bar\vv. \)
q.e.d.

Exponentials of matrices with complex eigenvalues

The basic example

Consider the matrix \[ J= \begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix}. \] The eigenvalues of this matrix are $\lambda_\pm = \pm i$, and the corresponding eigenvectors are \[ \vv= \begin{bmatrix} 1 \\ i \end{bmatrix} = \begin{bmatrix} 1 \\ 0 \end{bmatrix} + i \begin{bmatrix} 0 \\ 1 \end{bmatrix} \text{ and } \bar\vv = \begin{bmatrix} 1 \\ - i \end{bmatrix} = \begin{bmatrix} 1 \\ 0 \end{bmatrix} - i \begin{bmatrix} 0 \\ 1 \end{bmatrix}. \] (Check for yourself that $J\vv = i\vv$ and $J\bar\vv = -i \bar \vv$.)
The matrix $J$ satisfies \[ J^2 = - I,\quad J^3=-J,\quad J^4=I, \ldots \] and thus we have \begin{align*} e^{tJ} &= I + tJ +\frac{t^2}{2!}J^2+ \frac{t^3} {3!}J^3+\cdots\\ &= I + tJ -\frac{t^2}{2!}I - \frac{t^3} {3!}J+\cdots\\ &= \Bigl(1-\frac{t^2} {2!} + \frac{t^4} {4!} - \cdots\Bigr)I +\Bigl(t-\frac{t^3} {3!}+\cdots\Bigr)J \\[1ex] &= \cos(t)I + \sin(t)J \end{align*} so that \[ e^{tJ} = \begin{bmatrix} \cos t & \sin t \\ - \sin t & \cos t \end{bmatrix} \]

Second example

The eigenvalues of the matrix \[ A = \begin{bmatrix} \alpha & \omega \\ - \omega & \alpha \end{bmatrix} \] are $\lambda=\alpha+i\omega$ and $\bar\lambda=\alpha-i \omega$. The corresponding eigenvectors are again \[ \vv = \begin{bmatrix} 1 \\ i \end{bmatrix} = \begin{bmatrix} 1 \\ 0 \end{bmatrix} + i \begin{bmatrix} 0 \\ 1 \end{bmatrix} \text{ and } \bar\vv = \begin{bmatrix} 1 \\ - i \end{bmatrix} = \begin{bmatrix} 1 \\ 0 \end{bmatrix} - i \begin{bmatrix} 0 \\ 1 \end{bmatrix}. \] For this matrix we can use the previous example to compute $e^{tA}$. Write \[ A = \alpha I + \omega J, \qquad I = \begin{bmatrix} 1 &0 \\ 0 &1 \end{bmatrix}, \quad J = \begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix}. \] The matrices $I$ and $J$ commute, so $e^{tA} = e^{\alpha tI+\omega tJ} = e^{\alpha tI} e^{\omega tJ} $. We have \[ e^{\alpha tI} = \begin{bmatrix} e^{\alpha t} & 0 \\ 0 & e^{\alpha t} \end{bmatrix}, \qquad e^{\omega tJ} = \begin{bmatrix} \cos \omega t & \sin \omega t \\ -\sin \omega t & \cos \omega t \end{bmatrix} \] and thus \[ e^{tA} = e^{\alpha t}\begin{bmatrix} \cos \omega t & \sin \omega t \\ -\sin \omega t & \cos \omega t \end{bmatrix} = \begin{bmatrix} e^{\alpha t}\cos \omega t & e^{\alpha t}\sin \omega t \\ -e^{\alpha t}\sin \omega t & e^{\alpha t}\cos \omega t \end{bmatrix}. \]

Diagonalizing complex matrices

Suppose that $A$ has $\ell$ real eigenvalues $\lambda_1, \ldots, \lambda_\ell$ and $k$ pairs of complex eigenvalues $ \alpha_1 \pm i\omega_1$, $\dots$, $\alpha_k\pm i\omega_k$, where we assume that $\omega_1, \dots, \omega_k\gt0$. Let $\vv_1$, … , $\vv_\ell$ be real eigenvectors corresponding to the real eigenvalues, and let $\vw_1$, $\bar\vw_1$, … , $\vw_k$, $\bar\vw_k$, be the complex eigenvectors that go with the complex eigenvalues. We write these complex eigenvectors as $\vw_j = \va_j+i\vb_j$, $\bar\vw_j = \va_j - i\vb_j$, and we define the matrix \[ V = \begin{bmatrix} | & | & & | & | & | & & | & |\\ \vv_1 & \vv_2 & \cdots & \vv_\ell & \va_1 & \vb_1 & \cdots& \va_k & \vb_k\\ | & | & & | & | & | & & | & |\\ \end{bmatrix} \]
Lemma.  $AV = VD$ where $D$ is the not–quite–diagonal matrix \[ D = \begin{pmatrix} \lambda_1 & & & & & & \\ & \ddots\\ & & \lambda_\ell & & & & & \\ & & & \alpha_1 & \omega_1 & & \\ & & & -\omega_1 & \alpha_1 & & \\ & & & & & \ddots & \\ & & & & & & \alpha_k & \omega_k \\ & & & & & & - \omega_k & \alpha_k \\ \end{pmatrix} \]

Proof

First consider the simplest case $\ell=0$, $k=1$, i.e. the case where $A$ is a real $2\times 2$ matrix with a complex eigenvalue $\alpha+ i\omega$ and eigenvectors $\vw = \va+i\vb$. It then follows from $A\vw=\lambda\vw$ that \[ A\va + iA\vb = A\bigl(\va+i\vb\bigr) = (\alpha+i\omega)(\va+i\vb) =\bigl(\alpha\va-\omega\vb\bigr) + i\bigl(\omega\va+\alpha\vb\bigr). \] Comparing real and imaginary parts we conclude \[ A\va = \alpha\va-\omega\vb, \qquad A\vb = \omega\va+\alpha\vb. \] If we now form the matrix $V=[\va\;\; \vb]$, and compute $AV$, then we get \[ AV = A[\va\;\; \vb] = [A\va\;\; A\vb] = [ \alpha\va-\omega\vb \;\; \omega\va+\alpha\vb] = [\va\;\; \vb] \begin{bmatrix} \alpha & \omega \\ -\omega & \alpha \end{bmatrix} =VD, \] where \[ D = \begin{bmatrix} \alpha & \omega \\ -\omega & \alpha \end{bmatrix}. \] The general case for larger matrices can be handled similarly: if \[ V= \begin{bmatrix} \vv_1 & \cdots & \vv_\ell & \va_1 & \vb_1 & \cdots& \va_k & \vb_k \end{bmatrix} \] then \begin{align*} AV &= \begin{bmatrix} A\vv_1 & \cdots & A\vv_\ell & A\va_1 & A\vb_1 & \cdots& A\va_k & A\vb_k \end{bmatrix}\\ &=\begin{bmatrix} \lambda_1\vv_1 & \cdots & \lambda_\ell\vv_\ell & \alpha_1\va_1-\omega\vb_1 & \omega_1\va_1+\alpha_1\vb_1 & \cdots& \alpha_k\va_k-\omega_k\vb_k & \omega_k\va_k+\alpha_k\vb_k \end{bmatrix}\\ &= \begin{bmatrix} \vv_1 & \cdots & \vv_\ell & \va_1 & \vb_1 & \cdots& \va_k & \vb_k \end{bmatrix} D\\ &=VD \end{align*} where $D$ is the matrix in the theorem.
q.e.d.

Computing $e^{tA}$

It follows from $AV=VD$ that $A = V D V^{-1}$, and therefore the matrix exponential $e^{tA}$ is given by $e^{tA} = Ve^{tD}V^{-1}$, so we must first find $e^{tD}$: \[ e^{tD} = \begin{pmatrix} e^{\lambda_1t} & & & & & & \\ & \ddots\\ & & e^{\lambda_\ell t} & & & & & \\ & & & e^{\alpha_1 t}\cos\omega_1t & e^{\alpha_1 t}\sin \omega_1 t & & \\ & & & -e^{\alpha_1 t}\sin\omega_1 t& e^{\alpha_1 t}\cos\omega_1 t & & \\ & & & & & \ddots & \\ & & & & & & e^{\alpha_k t}\cos\omega_kt & e^{\alpha_k t}\sin \omega_k t \\ & & & & & & -e^{\alpha_k t}\sin\omega_k t& e^{\alpha_k t}\cos\omega_k t \end{pmatrix} \]

Solving $\dot\vx = A\vx$ without the matrix exponential

Let $A$ be an $n\times n$ matrix with real eigenvalues \[ \lambda_m,\quad 1\le m\le\ell, \] and complex eigenvalues \[ \mu_m=\alpha_m + i\omega_m,\quad \bar\mu_m = \alpha_m-i\omega_m, \qquad 1\le m \le k, \] and let the corresponding eigenvectors be $\vv_m$ ($1\le m\le \ell$) and $\vw_m,\bar\vw_m$ ($1\le m\le k$).

Assume that $n=\ell+2k$ so that these are all the eigenvalues of $A$.

Theorem. The vectors $\{\vv_1, \dots, \vv_\ell, \vw_1, \bar\vw_1, \dots, \vw_m, \bar\vw_m\}$ are linearly independent. If $\vx$ is a real vector then it is of the form \[ \vx = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \tfrac12q_1\vw_1 + \tfrac12\bar q_1\bar\vw_1 + \cdots + \tfrac12q_k\vw_k + \tfrac12\bar q_k\bar\vw_k \] where $p_1$, …, $p_\ell$ are real and $q_1$, …, $q_k$ are complex numbers. The linear combination above is often written as \[ \vx = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \Re \bigl\{q_1\vw_1 + \cdots + q_k\vw_k\bigr\} \]

Proof

The vectors $\{\vv_1, \dots, \vv_\ell, \vw_1, \bar\vw_1, \dots, \vw_m, \bar\vw_m\}$ are eigenvectors corresponding to different eigenvalues, and therefore they are linearly independent. Since we have assumed that we have exactly $n=\ell+2k$ eigenvalues, the vectors $\{\vv_1, \dots, \vv_\ell, \vw_1, \bar\vw_1, \dots, \vw_m, \bar\vw_m\}$ also span $\C^n$, and thus they form a basis. This means that every vector $\vx\in\C^n$ is a linear combination \[ \vx = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \tfrac12 q_1\vw_1 + \tfrac12 r_1\bar \vw_1 + \cdots + \tfrac12 q_k\vw_k + \tfrac12 r_k\bar \vw_k \] for certain complex numbers $p_m, q_m, r_m$.

The vector $\vx$ is real if and only if $\vx = \bar \vx$. Since \[ \vx = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \tfrac12 q_1\vw_1 + \tfrac12 r_1\bar \vw_1 + \cdots + \tfrac12 q_k\vw_k + \tfrac12 r_k\bar \vw_k \] we have \[ \bar\vx = \bar p_1\vv_1 + \cdots + \bar p_\ell\vv_\ell + \tfrac12 \bar q_1\bar \vw_1 + \tfrac12 \bar r_1\vw_1 + \cdots + \tfrac12 \bar q_k\bar \vw_k + \tfrac12 \bar r_k\vw_k \] The vectors $\{\vv_1, \dots, \vv_\ell, \vw_1, \bar\vw_1, \dots, \vw_m, \bar\vw_m\}$ are linearly independent, so $\vx=\bar\vx$ holds if and only if \[ p_m=\bar p_m \quad(1\le m\le\ell), \text{ and } r_m=\bar q_m \quad(1\le m\le k). \] Thus the coefficients of the real vectors $\vv_m$ must be real, while the coefficients of the complex vectors $\{\vw_m, \bar \vw_m\}$ come in complex conjugate pairs $q_m, \bar q_m$.

Since $\bar q_j\bar \vw_j$ is the complex conjugate of the vector $q_j\vw_j$, one has \[ q_j\vw_j + \bar q_j\bar\vw_j = 2\Re \bigl(q_j\vw_j\bigr) \] which implies the second form for $\vx$ in the theorem.

q.e.d.
Theorem. The general real valued solution of $\dot \vx=A\vx$ is given by \begin{align*} \vx(t) &= p_1e^{\lambda_1t}\vv_1 + \cdots + p_\ell e^{\lambda_\ell}\vv_\ell + \Re\Bigl\{ q_1e^{\mu_1 t}\vw_1 + \cdots + q_k e^{\mu_kt}\vw_k\Bigr\} \\ &= p_1e^{\lambda_1t}\vv_1 + \cdots + p_\ell e^{\lambda_\ell}\vv_\ell + e^{\alpha_1 t}\Re\bigl\{ q_1e^{i\omega_1 t}\vw_1\bigr\} + \cdots + e^{\alpha_k t}\Re\bigl\{ q_ke^{i\omega_k t}\vw_k\bigr\}. \end{align*} Here the coefficients $p_j, q_j$ are determined by the initial value of the solution through \[ \vx(0) = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \Re \bigl\{q_1\vw_1 + \cdots + q_k\vw_k\bigr\} \]
The second form of the solution shows that the terms in the solution corresponding to the complex eigenvalues grow or decay according to the real parts $\alpha_j$ of the complex eigenvalues, while they oscillate with frequency $\omega_j$.

Proof

Each of the terms \[ e^{\lambda_j t}\vv_j, \quad e^{\mu_jt}\vw_j, \quad e^{\bar \mu_j t}\bar\vw_j \] satisfies $\dot\vx = A\vx$, and therefore any linear combination, such as \begin{align*} \vx(t) = p_1e^{\lambda_1t}\vv_1 &+ \cdots + p_\ell e^{\lambda_\ell t}\vv_\ell \\ & + \frac{q_1}{2}e^{\mu_1t}\vw_1 + \frac{r_1}{2}e^{\bar\mu_1t}\bar\vw_1 +\cdots + \frac{q_k}{2}e^{\mu_kt}\vw_1 + \frac{r_k}{2}e^{\bar\mu_kt}\bar\vw_1 \end{align*} also is a solution of $\dot\vx=A\vx$.

If we choose $p_j\in\R$ and $q_j\in\C$ so that \[ \vx(0) = p_1\vv_1 + \cdots + p_\ell\vv_\ell + \Re \bigl\{q_1\vw_1 + \cdots + q_k\vw_k\bigr\} \] holds, and set $r_j=\bar q_j$, then we find that \begin{align*} \vx(t) &= p_1e^{\lambda_1t}\vv_1 + \cdots + p_\ell e^{\lambda_\ell t}\vv_\ell + \frac12 \Bigl\{q_1e^{\mu_1t}\vw_1 + r_1e^{\bar\mu_1t}\bar\vw_1 +\cdots + q_ke^{\mu_kt}\vw_k + r_ke^{\bar\mu_kt}\bar\vw_k\Bigr\}\\ &= p_1e^{\lambda_1t}\vv_1 + \cdots + p_\ell e^{\lambda_\ell t}\vv_\ell + \frac12 \Bigl\{q_1e^{\mu_1t}\vw_1 + \overline{q_1e^{\mu_1t}\vw_1} +\cdots + q_ke^{\mu_kt}\vw_k + \overline{q_ke^{\mu_kt}\vw_k}\Bigr\}\\ &= p_1e^{\lambda_1t}\vv_1 + \cdots + p_\ell e^{\lambda_\ell t}\vv_\ell + \Re\Bigl\{ q_1e^{\mu_1 t}\vw_1 + \cdots + q_k e^{\mu_kt}\vw_k\Bigr\}. \end{align*} Using $\mu_j = \alpha_j + i\omega_j$ one can rewrite \[ \Re\bigl\{q_j e^{\mu_j t}\vw_j\bigr\} = \Re\bigl\{q_j e^{\alpha_j t}e^{i\omega_j t}\vw_j\bigr\} = e^{\alpha_j t}\Re\bigl\{q_j e^{i\omega_j t}\vw_j\bigr\}, \] because $e^{\alpha_jt}$ is real.

q.e.d.