Systems of First-Order ODEs and the Phase Plane
Pages 9 and 10 of the class notes, in their order: systems and their existence theorem, linear systems with the Wronskian and the fundamental set, elimination and the eigenvalue method, and the three cases of a $2\times2$ matrix drawn by a change of basis $\mathbf z = P^{-1}\mathbf x$ — the six orderings of two real eigenvalues and a zero eigenvalue, complex eigenvalues through the complex diagonalisation of the notes, and a repeated eigenvalue with the construction of the generalised eigenvector. Robinson material beyond the notes is labelled as a complement.
We have $n$ dependent variables $x_1, x_2, \dots, x_n$ and one independent variable $t$. They satisfy a system of ODEs, $\dot x_1 = f_1(x_1, \dots, x_n, t),\ \dots,\ \dot x_n = f_n(x_1, \dots, x_n, t)$. Notation: $\mathbf x = (x_1, \dots, x_n)^T$ is a function of $t$ and $\mathbf f = (f_1, \dots, f_n)^T$ a function of $\mathbf x$ and $t$, so that $\dot{\mathbf x} = \mathbf f(\mathbf x, t)$.
If $\mathbf f$ and $D\mathbf f$ (the Jacobian matrix) are continuous functions of $\mathbf x$ and $t$ in an open neighbourhood of $\mathbf x_0$ and $t_0$, then the IVP $\dot{\mathbf x} = \mathbf f(\mathbf x, t)$, $\mathbf x(t_0) = \mathbf x_0$ has a unique local solution.
Robinson states the same theory as follows.
A system of $n$ coupled first-order equations [Robinson, Eq. (25.1), p. 249]
$$\dot x_i = f_i(x_1, \dots, x_n, t), \qquad i = 1, \dots, n, \tag{25.1}$$is written compactly with $\mathbf x = (x_1, \dots, x_n)^T$ and $\mathbf f = (f_1, \dots, f_n)^T$ as
$$\frac{d\mathbf x}{dt} = \mathbf f(\mathbf x, t).$$For $n = 1$ this is the scalar equation of Topic 2, and the theory is the same with vectors in place of numbers.
A solution of $\dot{\mathbf x} = \mathbf f(\mathbf x, t)$, $\mathbf x(t_0) = \mathbf x_0 \in \mathbb R^n$ (25.2) on an open interval $I \ni t_0$ is a differentiable function $\mathbf x : I \to \mathbb R^n$ with $\mathbf x(t_0) = \mathbf x_0$ and $\dot{\mathbf x}(t) = \mathbf f(\mathbf x(t), t)$ for all $t \in I$.
The role of $\partial f/\partial x$ is now played by the matrix of all partial derivatives, $D\mathbf f$ (usually called the Jacobian; Robinson just says "the matrix of partial derivatives") [Eq. (25.3), p. 250]:
$$D\mathbf f = \begin{pmatrix} \partial f_1/\partial x_1 & \cdots & \partial f_1/\partial x_n \\ \vdots & & \vdots \\ \partial f_n/\partial x_1 & \cdots & \partial f_n/\partial x_n\end{pmatrix}. \tag{25.3}$$If $\mathbf f(\mathbf x, t)$ and $D\mathbf f(\mathbf x, t)$ are continuous for $\mathbf x \in U = (a_1, b_1)\times\dots\times(a_n, b_n)$ and $c < t < d$, then for any $\mathbf x_0 \in U$ and $t_0 \in (c, d)$ the IVP (25.2) has a unique solution on some open interval containing $t_0$.
In practice: check that every $f_i$ and every partial derivative $\partial f_i/\partial x_j$ is continuous.
For $\mathbf f = \big(x(4 - 2x - y),\ y(9 - 3x - 3y)\big)$, $D\mathbf f = \begin{pmatrix} 4 - 4x - y & -x \\ -3y & 9 - 3x - 6y\end{pmatrix}$. All entries are polynomials, hence continuous on all of $\mathbb R^2$: every initial condition has a unique local solution, and in particular two trajectories of Figure 1.1 of Topic 1 can never cross.
1.1 Higher-order equations as systems [Robinson, Sec. 25.1, pp. 251–252]
ODEs of order $n \ge 2$ can be converted into a system of $n$ first-order ODEs. Example: $\ddot x = f(\dot x, x, t)$ becomes $\dot x_1 = x_2$, $\dot x_2 = f(x_2, x_1, t)$ with $x_1 = x$, $x_2 = \dot x$.
For $\ddot x = f(\dot x, x, t)$ introduce $x_1 = x$ and $x_2 = \dot x$. Then
$$\dot x_1 = x_2, \qquad \dot x_2 = \ddot x = f(x_2, x_1, t), \qquad\text{i.e.}\qquad \frac{d\mathbf x}{dt} = \mathbf g(\mathbf x, t) = \begin{pmatrix} x_2 \\ f(x_2, x_1, t)\end{pmatrix}. \tag{25.4}$$The vector initial condition $\mathbf x(t_0) = \mathbf x_0$ is exactly the pair $x(t_0)$, $\dot x(t_0)$, and $D\mathbf g = \begin{pmatrix} 0 & 1 \\ \partial f/\partial x_1 & \partial f/\partial x_2\end{pmatrix}$. Applying Theorem 25.2 gives Theorem 25.3, which is Theorem 11.1 of Topic 3. An $n$th-order equation becomes an $n$-dimensional system in the same way, with $x_k = x^{(k-1)}$.
$\ddot x + \lambda\dot x + \omega^2x = 0$ with $x_1 = x$ (position) and $x_2 = \dot x$ (velocity) becomes
$$\frac{d}{dt}\begin{pmatrix} x_1 \\ x_2\end{pmatrix} = \begin{pmatrix} 0 & 1 \\ -\omega^2 & -\lambda\end{pmatrix}\begin{pmatrix} x_1 \\ x_2\end{pmatrix}.$$The plane of $(x, \dot x)$ is the phase plane; we return to this example in Section 9.
2. Matrix form, superposition and the fundamental matrix
The homogeneous system $\dot x = ax + by$, $\dot y = cx + dy$ (27.1) is, with $\mathbf x = (x, y)^T$ [Eq. (27.2), p. 259],
$$\dot{\mathbf x} = A\mathbf x, \qquad A = \begin{pmatrix} a & b \\ c & d\end{pmatrix}. \tag{27.2}$$If $\mathbf x_1$ and $\mathbf x_2$ solve (27.2), so does $\mathbf x = \alpha\mathbf x_1 + \beta\mathbf x_2$: $\ \dot{\mathbf x} = \alpha\dot{\mathbf x}_1 + \beta\dot{\mathbf x}_2 = \alpha A\mathbf x_1 + \beta A\mathbf x_2 = A(\alpha\mathbf x_1 + \beta\mathbf x_2) = A\mathbf x$.
$\dot{\mathbf x} = A(t)\,\mathbf x + \mathbf f(t)$, $\mathbf x(t_0) = \mathbf x_0$, where $A$ is an $n\times n$ matrix (depending on $t$) and $\mathbf f$ a vector function of $t$; if $\mathbf f = \mathbf 0$ the system is homogeneous. Existence and uniqueness are guaranteed if $A$ and $\mathbf f$ are continuous.
- Notation. For two solutions write $\mathbf x_1 = (x_{11}, x_{21}, \dots, x_{n1})^T$, $\mathbf x_2 = (x_{12}, x_{22}, \dots, x_{n2})^T$.
- Superposition. If $\mathbf x_1, \mathbf x_2, \dots, \mathbf x_k$ solve $\dot{\mathbf x} = A(t)\mathbf x$, so does every linear combination $c_1\mathbf x_1 + c_2\mathbf x_2 + \dots + c_k\mathbf x_k$.
- Theorem. The solutions $\mathbf x_1, \dots, \mathbf x_k$ of $\dot{\mathbf x} = A(t)\mathbf x$ are linearly independent if and only if the Wronskian $W(\mathbf x_1, \dots, \mathbf x_k) = \det\big(x_{ij}\big) \neq 0$ for all $t \in I$.
- $n = 2$. $\begin{pmatrix} \dot x_1 \\ \dot x_2\end{pmatrix} = \begin{pmatrix} a_{11}(t) & a_{12}(t) \\ a_{21}(t) & a_{22}(t)\end{pmatrix}\begin{pmatrix} x_1 \\ x_2\end{pmatrix} + \begin{pmatrix} f_1(t) \\ f_2(t)\end{pmatrix}$. A fundamental set of solutions is a pair $\mathbf x_1(t)$, $\mathbf x_2(t)$ with $W(\mathbf x_1, \mathbf x_2) \neq 0$.
- Theorem. The general solution of $\dot{\mathbf x} = A(t)\mathbf x + \mathbf f(t)$ is $\mathbf x = \mathbf x_h + \mathbf x_p$, where $\mathbf x_h$ is the general solution of $\dot{\mathbf x} = A(t)\mathbf x$ and $\mathbf x_p$ a particular solution of the complete system.
Robinson does not develop this general theory for systems; the nearest book material is the scalar version (Sec. 11.3–11.4 and 14.1). The statements below make the notes precise for constant $A$; the proofs mirror Topic 3.
- Vector functions $\mathbf x_1(t), \mathbf x_2(t)$ are linearly independent on $I$ if $\alpha\mathbf x_1(t) + \beta\mathbf x_2(t) = \mathbf 0$ for all $t \in I$ forces $\alpha = \beta = 0$.
- Their Wronskian is the determinant of the matrix whose columns are the solutions: $W(t) = \det\big[\mathbf x_1(t)\ \ \mathbf x_2(t)\big] = x_{11}x_{22} - x_{12}x_{21}$, where $\mathbf x_j = (x_{1j}, x_{2j})^T$.
- Two independent solutions form a fundamental set; the matrix $\Phi(t) = [\mathbf x_1\ \ \mathbf x_2]$ is a fundamental matrix. It satisfies $\dot\Phi = A\Phi$, column by column.
- Two solutions of $\dot{\mathbf x} = A\mathbf x$ are linearly independent if and only if $W(t) \neq 0$; and then $W(t) \neq 0$ for all $t$.
- If $\Phi$ is a fundamental matrix, every solution is $\mathbf x(t) = \Phi(t)\,\mathbf c = c_1\mathbf x_1(t) + c_2\mathbf x_2(t)$, and the IVP $\mathbf x(t_0) = \mathbf x_0$ has $\mathbf c = \Phi(t_0)^{-1}\mathbf x_0$.
If $W(t_0) \neq 0$, the columns $\mathbf x_1(t_0)$, $\mathbf x_2(t_0)$ are independent vectors, so $\alpha\mathbf x_1(t_0) + \beta\mathbf x_2(t_0) = \mathbf 0$ already forces $\alpha = \beta = 0$. Conversely, if $W(t_0) = 0$, there are $(\alpha, \beta) \neq (0, 0)$ with $\alpha\mathbf x_1(t_0) + \beta\mathbf x_2(t_0) = \mathbf 0$. Then $\mathbf z = \alpha\mathbf x_1 + \beta\mathbf x_2$ solves the system (superposition) with $\mathbf z(t_0) = \mathbf 0$; so does $\mathbf 0$, and by uniqueness (Theorem 25.2) $\mathbf z \equiv \mathbf 0$: the solutions are dependent. For (2): given any solution $\mathbf x$, the system $\Phi(t_0)\mathbf c = \mathbf x(t_0)$ is solvable because $\det\Phi(t_0) = W(t_0) \neq 0$; then $\Phi\mathbf c$ and $\mathbf x$ are solutions with the same value at $t_0$, hence equal for all $t$.
Liouville's formula — Abel's identity for systems (complementary)
Write the columns as $\mathbf x_j = (p_j, q_j)^T$, so $\dot p_j = ap_j + bq_j$, $\dot q_j = cp_j + dq_j$ and $W = p_1q_2 - q_1p_2$. By the product rule
$$\dot W = \dot p_1q_2 + p_1\dot q_2 - \dot q_1p_2 - q_1\dot p_2 = (ap_1 + bq_1)q_2 + p_1(cp_2 + dq_2) - (cp_1 + dq_1)p_2 - q_1(ap_2 + bq_2).$$The $b$ terms ($bq_1q_2 - bq_1q_2$) and the $c$ terms ($cp_1p_2 - cp_1p_2$) cancel, leaving
$$\dot W = (a + d)(p_1q_2 - q_1p_2) = (\operatorname{tr}A)\,W \quad\Longrightarrow\quad W(t) = W(t_0)\,e^{(\operatorname{tr}A)(t - t_0)}.$$(The computation never uses that $a, b, c, d$ are constant; with time-dependent entries $W(t) = W(t_0)\exp\int_{t_0}^t\operatorname{tr}A(s)\,ds$.) For a second-order equation turned into a system, $A = \begin{pmatrix} 0 & 1 \\ -q & -p\end{pmatrix}$ has trace $-p$, the solution columns are $(x_j, \dot x_j)^T$, and $W$ is exactly the scalar Wronskian $x_1\dot x_2 - x_2\dot x_1$: Abel's identity of Topic 3 is the special case.
2.1 Inhomogeneous systems
If $\mathbf x_p$ is one solution, every solution is $\mathbf x = \Phi(t)\mathbf c + \mathbf x_p(t)$.
If $\mathbf x$ solves the inhomogeneous system, then $\frac{d}{dt}(\mathbf x - \mathbf x_p) = (A\mathbf x + \mathbf f) - (A\mathbf x_p + \mathbf f) = A(\mathbf x - \mathbf x_p)$, so $\mathbf x - \mathbf x_p$ solves the homogeneous system and equals $\Phi\mathbf c$. Conversely $\frac{d}{dt}(\Phi\mathbf c + \mathbf x_p) = A\Phi\mathbf c + A\mathbf x_p + \mathbf f$.
Variation of parameters for systems (complementary). As in Topic 3, let the constants vary: $\mathbf x_p = \Phi(t)\mathbf c(t)$. The product rule gives $\dot{\mathbf x}_p = \dot\Phi\mathbf c + \Phi\dot{\mathbf c} = A\Phi\mathbf c + \Phi\dot{\mathbf c}$; this equals $A\mathbf x_p + \mathbf f$ exactly when $\Phi\dot{\mathbf c} = \mathbf f$. Since $\det\Phi = W \neq 0$:
$$\mathbf x_p(t) = \Phi(t)\int\Phi(s)^{-1}\,\mathbf f(s)\,ds.$$Take $\dot x = x + y$, $\dot y = 4x - 2y + 4e^{-2t}$ (Robinson Example 26.1, solved by elimination in Section 3). The homogeneous solutions are $e^{2t}(1, 1)^T$ and $e^{-3t}(1, -4)^T$ (Section 5), so $\Phi = \begin{pmatrix} e^{2t} & e^{-3t} \\ e^{2t} & -4e^{-3t}\end{pmatrix}$, $W = -4e^{-t} - e^{-t} = -5e^{-t}$ (check: $\operatorname{tr}A = -1$, so $W \propto e^{-t}$ ✓), and $\Phi^{-1} = \frac{1}{W}\begin{pmatrix} -4e^{-3t} & -e^{-3t} \\ -e^{2t} & e^{2t}\end{pmatrix} = \frac15\begin{pmatrix} 4e^{-2t} & e^{-2t} \\ e^{3t} & -e^{3t}\end{pmatrix}$. With $\mathbf f = (0, 4e^{-2t})^T$: $\Phi^{-1}\mathbf f = \tfrac45\big(e^{-4t},\ -e^{t}\big)^T$, whose antiderivative is $\big(-\tfrac15e^{-4t},\ -\tfrac45e^{t}\big)^T$ (for the first component substitute $s = -4t$, $ds = -4\,dt$: $\int\tfrac45e^{-4t}dt = -\tfrac15\int e^{s}ds = -\tfrac15e^{-4t}$; for the second, $\int -\tfrac45e^{t}dt = -\tfrac45e^{t}$). Then
$$\mathbf x_p = \Phi\begin{pmatrix} -\tfrac15e^{-4t} \\ -\tfrac45e^{t}\end{pmatrix} = \begin{pmatrix} -\tfrac15e^{-2t} - \tfrac45e^{-2t} \\ -\tfrac15e^{-2t} + \tfrac{16}{5}e^{-2t}\end{pmatrix} = \begin{pmatrix} -e^{-2t} \\ 3e^{-2t}\end{pmatrix},$$the same particular solution that elimination produces in Section 3 (Example 26.1) ✓.
3. Linear systems and the elimination method
A first-order linear system of two equations with constant coefficients is [Robinson, Eq. (26.1), p. 253]
$$\dot x = ax + by + f(t), \qquad \dot y = cx + dy + g(t). \tag{26.1}$$It is homogeneous if $f = g = 0$ and inhomogeneous otherwise. If $b = c = 0$ the equations are uncoupled and each is solved by an integrating factor.
For $\dot x = ax + by$, $\dot y = cx + dy$ (with $b \neq 0$):
- From the first equation, $y = \dfrac{\dot x - ax}{b}$, so $\dot y = \dfrac{\ddot x - a\dot x}{b}$.
- Substitute into the second equation: $\dfrac{\ddot x - a\dot x}{b} = cx + d\,\dfrac{\dot x - ax}{b}$.
- Multiply by $b$: $\ddot x - a\dot x = bc\,x + d\dot x - ad\,x$, that is, a second-order equation for $x$:
The system $x' = 2x - y$, $y' = -x + 3y$ of Topic 1 gives $\ddot x - 5\dot x + 5x = 0$ this way.
The boxed result of the notes reads $\ddot x - (a + b)\dot x + (ad - bc)x = 0$. The steps give $a + d$ (the trace of the matrix), not $a + b$. Check with the Topic 1 system ($a = 2$, $b = -1$, $c = -1$, $d = 3$): direct elimination gives $\ddot x - 5\dot x + 5x = 0$, and $a + d = 5$ while $a + b = 1$.
Robinson's version also carries forcing terms:
3.1 Reduction to one second-order equation
Assume $b \neq 0$ (otherwise exchange the roles of $x$ and $y$).
- Solve the first equation for $y$: $\ y = \dfrac{\dot x - ax - f(t)}{b}$ (26.2).
- Differentiate: $\ \dot y = \dfrac{\ddot x - a\dot x - \dot f(t)}{b}$.
- Substitute both into the second equation: $\ \dfrac{\ddot x - a\dot x - \dot f}{b} = cx + d\,\dfrac{\dot x - ax - f}{b} + g$.
- Multiply by $b$ and collect terms:
Solve (26.3) by the methods of Topic 3, then recover $y$ from (26.2). The coefficients $a + d$ and $ad - bc$ are the trace and the determinant of the matrix $\begin{pmatrix} a & b \\ c & d\end{pmatrix}$ — they will reappear in the eigenvalue method.
- Here $a = 1$, $b = 1$, $c = 4$, $d = -2$, $f = 0$, $g = 4e^{-2t}$. From the first equation $y = \dot x - x$ (26.4), so $\dot y = \ddot x - \dot x$.
- Substituting: $\ddot x - \dot x = 4x - 2(\dot x - x) + 4e^{-2t}$, i.e. $\ddot x + \dot x - 6x = 4e^{-2t}$ (26.5) — as (26.3) predicts: $a + d = -1$, $ad - bc = -6$, right side $b\,g = 4e^{-2t}$.
- Complementary function: $k^2 + k - 6 = 0$, $k = 2, -3$. Particular integral $Ce^{-2t}$: $(4 - 2 - 6)C = 4$, $C = -1$. So $x = Ae^{2t} + Be^{-3t} - e^{-2t}$.
- $y = \dot x - x = (2A - A)e^{2t} + (-3B - B)e^{-3t} + (2 + 1)e^{-2t} = Ae^{2t} - 4Be^{-3t} + 3e^{-2t}$.
- Initial data: $A + B - 1 = 0$ and $A - 4B + 3 = -1$. Subtracting, $5B = 5$: $B = 1$, $A = 0$.
Check: $\dot x = -3e^{-3t} + 2e^{-2t}$ and $x + y = 2e^{-2t} - 3e^{-3t}$ ✓; $x(0) = 0$, $y(0) = -1$ ✓.
$y = \tfrac15(\dot x - 2x)$, so $\tfrac15(\ddot x - 2\dot x) = -2x$, i.e. $\ddot x - 2\dot x + 10x = 0$, with roots $k = 1 \pm 3i$: $x = e^t(A\cos 3t + B\sin 3t)$ and $y = \tfrac{e^t}{5}\big[(3B - A)\cos 3t - (3A + B)\sin 3t\big]$. The data give $A = 5$, $B = -5$: $x = 5e^t(\cos 3t - \sin 3t)$, $y = -e^t(4\cos 3t + 2\sin 3t)$ — oscillations with exponentially growing amplitude.
For the scalar equation $\dot x = ax$ the solutions are $e^{at}x_0$. By analogy try a solution that keeps a fixed direction $\mathbf v$ and changes only in size [Eq. (27.4), p. 260]:
$$\mathbf x(t) = e^{\lambda t}\mathbf v, \qquad \mathbf v \text{ constant}, \ \mathbf v \neq \mathbf 0. \tag{27.4}$$Then $\dot{\mathbf x} = \lambda e^{\lambda t}\mathbf v$ and $A\mathbf x = e^{\lambda t}A\mathbf v$; equating and dividing by $e^{\lambda t} \neq 0$:
$$A\mathbf v = \lambda\mathbf v. \tag{27.5}$$So $e^{\lambda t}\mathbf v$ is a solution exactly when $\lambda$ is an eigenvalue of $A$ and $\mathbf v$ a corresponding eigenvector.
Look for solutions of $\dot{\mathbf x} = A\mathbf x$ of the form $\mathbf x(t) = e^{\lambda t}\mathbf v$, with $\mathbf v \neq \mathbf 0$ a vector and $\lambda$ a number, both to be determined. Then $\lambda e^{\lambda t}\mathbf v = Ae^{\lambda t}\mathbf v$, and cancelling $e^{\lambda t}$, $A\mathbf v = \lambda\mathbf v$: the same as an eigenvalue problem, with $\mathbf v$ the associated eigenvector. We need to solve $\det(A - \lambda I) = 0$.
4.1 Finding eigenvalues and eigenvectors [Sec. 27.2, pp. 260–261]
- Rewrite (27.5) as $(A - \lambda I)\mathbf v = \mathbf 0$ (27.7). If $A - \lambda I$ were invertible, the only solution would be $\mathbf v = \mathbf 0$. A non-zero $\mathbf v$ therefore requires $\det(A - \lambda I) = 0$.
- Expand: $\det\begin{pmatrix} a - \lambda & b \\ c & d - \lambda\end{pmatrix} = (a - \lambda)(d - \lambda) - bc$, giving the characteristic equation $$\lambda^2 - (a + d)\lambda + (ad - bc) = 0, \qquad\text{i.e.}\qquad \lambda^2 - (\operatorname{tr}A)\lambda + \det A = 0. \tag{27.8}$$
- For each root $\lambda$, solve the (singular) system $(A - \lambda I)\mathbf v = \mathbf 0$: one row is a multiple of the other, so one equation fixes the direction of $\mathbf v$. Any non-zero multiple is also an eigenvector, $A(c\mathbf v) = c\lambda\mathbf v$.
Equation (27.8) is the characteristic equation (26.3) of the eliminated second-order equation: both methods see the same numbers. Its roots satisfy $\lambda_1 + \lambda_2 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = \det A$.
- $(2 - \lambda)(3 - \lambda) - 12 = \lambda^2 - 5\lambda - 6 = (\lambda - 6)(\lambda + 1) = 0$: $\lambda_1 = 6$, $\lambda_2 = -1$. (Check: sum $5 = \operatorname{tr}A$, product $-6 = \det A$ ✓.)
- $\lambda = 6$: $(A - 6I)\mathbf v = \begin{pmatrix} -4 & 2 \\ 6 & -3\end{pmatrix}\mathbf v = \mathbf 0$ gives $-4v_1 + 2v_2 = 0$, $v_2 = 2v_1$: $\mathbf v_1 = (1, 2)^T$. Check: $A\mathbf v_1 = (6, 12)^T = 6\mathbf v_1$ ✓.
- $\lambda = -1$: $\begin{pmatrix} 3 & 2 \\ 6 & 4\end{pmatrix}\mathbf v = \mathbf 0$ gives $3v_1 + 2v_2 = 0$: $\mathbf v_2 = (2, -3)^T$. Check: $A\mathbf v_2 = (-2, 3)^T = -\mathbf v_2$ ✓.
Reading note: the notes label this line "$\lambda = 1$" while solving $(A + I)\mathbf v = \mathbf 0$; the eigenvalue is $-1$, as the factor $(\lambda + 1)$ shows.
For $\begin{pmatrix} \lambda_1 & b \\ 0 & \lambda_2\end{pmatrix}$ the characteristic polynomial is $(\lambda_1 - \lambda)(\lambda_2 - \lambda) - b\cdot 0$, so the eigenvalues are the diagonal entries, and $(1, 0)^T$ is an eigenvector for $\lambda_1$. The same holds for lower-triangular matrices (with $(0, 1)^T$ for $\lambda_2$) and, with both eigenvectors along the axes, for diagonal matrices. Robinson simply treats matrices of the form $\begin{pmatrix} \lambda_1 & b \\ 0 & \lambda_2\end{pmatrix}$ without naming them; the lecture calls them triangular, and the notes record that for diagonal and upper-triangular matrices the eigenvalues are the diagonal entries [Notes, p. 9].
The roots of (27.8) fall into three cases, each with its own chapter [p. 261]; the class notes call them Case 1, Case 2 and Case 3 (pp. 9–10):
| Discriminant $(\operatorname{tr}A)^2 - 4\det A$ | Eigenvalues | Example | Treated in |
|---|---|---|---|
| $> 0$ | Real, distinct | 27.1: $6, -1$ | Ch. 28 — Sections 5–6 |
| $= 0$ | One repeated real | 27.4: $A = \begin{pmatrix} 5 & -4 \\ 1 & 1\end{pmatrix}$, $(\lambda - 3)^2 = 0$, $\mathbf v = (2,1)^T$ only | Ch. 30 — Section 8 |
| $< 0$ | Complex pair $\rho \pm i\omega$ | 27.5: $A = \begin{pmatrix} 2 & 5 \\ -2 & 0\end{pmatrix}$, $\lambda = 1 \pm 3i$ | Ch. 29 — Section 7 |
5. Distinct real eigenvalues: solution and change of basis
$\mathbf x_1(t) = e^{\lambda_1t}\mathbf v_{\lambda_1}$ and $\mathbf x_2(t) = e^{\lambda_2t}\mathbf v_{\lambda_2}$ are both solutions of $\dot{\mathbf x} = A\mathbf x$, and the general solution is
$$\mathbf x(t) = C_1e^{\lambda_1t}\mathbf v_{\lambda_1} + C_2e^{\lambda_2t}\mathbf v_{\lambda_2},$$where we need to check that $\mathbf v_{\lambda_1}$, $\mathbf v_{\lambda_2}$ are linearly independent, and $C_1$, $C_2$ are fixed by $\mathbf x_0$. Therefore $\{\mathbf x_1, \mathbf x_2\}$ is a fundamental set of solutions. The next step is to draw the general solution in a phase portrait, using a change of basis.
If $\lambda_1 \neq \lambda_2$ are real with eigenvectors $\mathbf v_1$, $\mathbf v_2$, then $e^{\lambda_1t}\mathbf v_1$ and $e^{\lambda_2t}\mathbf v_2$ are solutions, and by superposition [Eq. (28.1), p. 270]
$$\mathbf x(t) = Ae^{\lambda_1t}\mathbf v_1 + Be^{\lambda_2t}\mathbf v_2. \tag{28.1}$$This is the general solution because the two solutions are independent. Their Wronskian at $t = 0$ is $\det[\mathbf v_1\ \mathbf v_2]$, non-zero by the next result.
If $A\mathbf v_1 = \lambda_1\mathbf v_1$, $A\mathbf v_2 = \lambda_2\mathbf v_2$ with $\mathbf v_1, \mathbf v_2 \neq \mathbf 0$ and $\lambda_1 \neq \lambda_2$, then $\mathbf v_1$, $\mathbf v_2$ are linearly independent.
Suppose $\alpha\mathbf v_1 + \beta\mathbf v_2 = \mathbf 0$ (B.3). Applying $A$ gives $\alpha\lambda_1\mathbf v_1 + \beta\lambda_2\mathbf v_2 = \mathbf 0$ (B.4). Multiply (B.3) by $\lambda_1$ and subtract (B.4):
$$\beta(\lambda_1 - \lambda_2)\mathbf v_2 = \mathbf 0.$$Since $\lambda_1 \neq \lambda_2$ and $\mathbf v_2 \neq \mathbf 0$, $\beta = 0$; then (B.3) reads $\alpha\mathbf v_1 = \mathbf 0$, so $\alpha = 0$. (Robinson argues by dividing by $\alpha$ and $\lambda_1$; subtracting as here avoids having to treat $\alpha = 0$ or $\lambda_1 = 0$ separately.)
$A = \begin{pmatrix} 1 & 1 \\ 4 & -2\end{pmatrix}$: $(1 - \lambda)(-2 - \lambda) - 4 = \lambda^2 + \lambda - 6 = 0$, $\lambda = 2, -3$. $\lambda = 2$: $(A - 2I) = \begin{pmatrix} -1 & 1 \\ 4 & -4\end{pmatrix}$, $\mathbf v_1 = (1, 1)^T$. $\lambda = -3$: $(A + 3I) = \begin{pmatrix} 4 & 1 \\ 4 & 1\end{pmatrix}$, $\mathbf v_2 = (1, -4)^T$. Independence directly: $\alpha(1,1) + \beta(1,-4) = \mathbf 0$ gives $\alpha + \beta = 0$, $\alpha - 4\beta = 0$, so $\alpha = \beta = 0$.
$$\mathbf x(t) = Ae^{2t}\begin{pmatrix} 1 \\ 1\end{pmatrix} + Be^{-3t}\begin{pmatrix} 1 \\ -4\end{pmatrix}.$$This is the homogeneous part of Example 26.1, as it should be.
5.1 Changing coordinates: diagonalisation [Sec. 28.2, pp. 271–274]
Because $\mathbf v_1$, $\mathbf v_2$ are independent, every vector can be written in the eigenvector basis, $\mathbf x = \tilde x\mathbf v_1 + \tilde y\mathbf v_2$ (28.3). In matrix form, with $P = [\mathbf v_1\ \mathbf v_2]$ (eigenvectors as columns),
$$\mathbf x = P\tilde{\mathbf x}, \qquad \tilde{\mathbf x} = P^{-1}\mathbf x. \tag{28.4–28.5}$$If $A$ has a basis of eigenvectors $\mathbf v_{\lambda_1}$, $\mathbf v_{\lambda_2}$, we can rewrite it as $A = PDP^{-1}$, where $P = (\mathbf v_{\lambda_1}, \mathbf v_{\lambda_2})$ and $D = \begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2\end{pmatrix} = \operatorname{diag}(\lambda_1, \lambda_2)$. Make the change of variable $\mathbf z = P^{-1}\mathbf x$. Then
$$\dot{\mathbf z} = P^{-1}\dot{\mathbf x} = P^{-1}A\mathbf x = P^{-1}\big(PDP^{-1}\big)\mathbf x = DP^{-1}\mathbf x = D\mathbf z,$$so $\mathbf z = (z_1, z_2)^T$ satisfies $\dot z_1 = \lambda_1z_1$, $\dot z_2 = \lambda_2z_2$, with solutions $z_1(t) = C_1e^{\lambda_1t}$, $z_2(t) = C_2e^{\lambda_2t}$. Now graph the solution $(z_1, z_2)$ in the plane (Section 6).
(The lecture writes $\mathbf z$ for Robinson's $\tilde{\mathbf x}$.) In Robinson's words: differentiate the second relation; $P^{-1}$ is constant:
$$\frac{d\tilde{\mathbf x}}{dt} = P^{-1}\frac{d\mathbf x}{dt} = P^{-1}A\mathbf x = P^{-1}AP\,\tilde{\mathbf x}. \tag{28.7}$$Now compute $P^{-1}AP$ column by column, using $A[\mathbf v_1\ \mathbf v_2] = [A\mathbf v_1\ A\mathbf v_2]$:
$$AP = [\lambda_1\mathbf v_1\ \ \lambda_2\mathbf v_2] = [\mathbf v_1\ \mathbf v_2]\begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2\end{pmatrix} = P\Lambda \quad\Longrightarrow\quad P^{-1}AP = \Lambda = \begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2\end{pmatrix}.$$So in the new coordinates the system is uncoupled [Eqs. (28.8)–(28.9)]:
$$\dot{\tilde x} = \lambda_1\tilde x, \quad \dot{\tilde y} = \lambda_2\tilde y \quad\Longrightarrow\quad \tilde x = Ae^{\lambda_1t}, \quad \tilde y = Be^{\lambda_2t},$$and $\mathbf x = P\tilde{\mathbf x} = Ae^{\lambda_1t}\mathbf v_1 + Be^{\lambda_2t}\mathbf v_2$ recovers (28.1).
$P = \begin{pmatrix} 1 & 1 \\ 1 & -4\end{pmatrix}$, $\det P = -4 - 1 = -5 \neq 0$, and with $\begin{pmatrix} a & b \\ c & d\end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a\end{pmatrix}$ (Appendix B), $P^{-1} = \frac15\begin{pmatrix} 4 & 1 \\ 1 & -1\end{pmatrix}$, i.e. $\tilde x = \tfrac15(4x + y)$, $\tilde y = \tfrac15(x - y)$. Then
$$P^{-1}AP = \frac15\begin{pmatrix} 4 & 1 \\ 1 & -1\end{pmatrix}\begin{pmatrix} 1 & 1 \\ 4 & -2\end{pmatrix}\begin{pmatrix} 1 & 1 \\ 1 & -4\end{pmatrix} = \frac15\begin{pmatrix} 4 & 1 \\ 1 & -1\end{pmatrix}\begin{pmatrix} 2 & -3 \\ 2 & 12\end{pmatrix} = \frac15\begin{pmatrix} 10 & 0 \\ 0 & -15\end{pmatrix} = \begin{pmatrix} 2 & 0 \\ 0 & -3\end{pmatrix}\ ✓.$$Robinson stresses that this is not the efficient way to solve the system — (28.1) is — but it is the key to drawing it.
6. Phase portraits: nodes, saddles and a zero eigenvalue
A phase portrait shows the curves traced in the $(x, y)$ plane by solutions, with arrows for increasing $t$. The origin is always a stationary solution of $\dot{\mathbf x} = A\mathbf x$ (and the only one when $\det A \neq 0$). Draw first the uncoupled system, then "twist" it.
6.1 The uncoupled system [Sec. 28.3, pp. 276–278]
Solutions of $\dot{\tilde x} = \lambda_1\tilde x$, $\dot{\tilde y} = \lambda_2\tilde y$ are $\tilde x = \tilde x_0e^{\lambda_1t}$, $\tilde y = \tilde y_0e^{\lambda_2t}$ (28.12). To find the curve they trace, eliminate $t$: $e^{t} = (\tilde x/\tilde x_0)^{1/\lambda_1}$, so
$$\tilde y = \tilde y_0\,e^{\lambda_2t} = \tilde y_0\left(\frac{\tilde x}{\tilde x_0}\right)^{\lambda_2/\lambda_1} = K\,\tilde x^{\lambda_2/\lambda_1}, \qquad K = \tilde y_0\,\tilde x_0^{-\lambda_2/\lambda_1}.$$(Equivalently, by the chain rule $\dfrac{d\tilde y}{d\tilde x} = \dfrac{\dot{\tilde y}}{\dot{\tilde x}} = \dfrac{\lambda_2\tilde y}{\lambda_1\tilde x}$, a separable equation.) The axes themselves are trajectories: data on an axis stay on it.
- Stable node ($\lambda_1 < \lambda_2 < 0$): every solution tends to $\mathbf 0$. Since $|\lambda_2| < |\lambda_1|$, the $\tilde y$ component decays more slowly, so trajectories arrive tangent to the slow eigendirection.
- Unstable node ($0 < \lambda_1 < \lambda_2$): the same picture with arrows reversed; solutions leave the origin tangent to the eigendirection of smaller $|\lambda|$.
- Saddle ($\lambda_2 < 0 < \lambda_1$): $\tilde x$ grows while $\tilde y$ decays; the curves $\tilde x\,\tilde y^{\,r} = K$ ($r = -\lambda_1/\lambda_2 > 0$) are hyperbola-like. Only solutions starting on the $\tilde y$-axis reach the origin.
With $z_1 = C_1e^{\lambda_1t}$ and $z_2 = C_2e^{\lambda_2t}$, write $e^{\lambda_2t} = \big(e^{\lambda_1t}\big)^{\lambda_2/\lambda_1}$:
$$z_2(t) = C_2e^{\lambda_2t} = C_2\big(e^{\lambda_1t}\big)^{\lambda_2/\lambda_1} = C_2\,C_1^{-\lambda_2/\lambda_1}\big(C_1e^{\lambda_1t}\big)^{\lambda_2/\lambda_1} = \tilde C\,\big(z_1(t)\big)^{\lambda_2/\lambda_1},$$with $\tilde C = C_2C_1^{-\lambda_2/\lambda_1}$ (for $C_1 > 0$; for $C_1 < 0$ use $|z_1|$ and $|C_1|$, by symmetry). The exponent $r = \lambda_2/\lambda_1$ decides the shape.
| Ordering (notes) | $r = \lambda_2/\lambda_1$ | Origin | Orbits near $\mathbf 0$ |
|---|---|---|---|
| $0 < \lambda_1 < \lambda_2$ | $r > 1$ | unstable node | leave tangent to the $z_1$-axis, like $z_2 = \tilde Cz_1^{\,r}$ (flat at $0$) |
| $0 < \lambda_2 < \lambda_1$ | $0 < r < 1$ | unstable node | leave tangent to the $z_2$-axis |
| $\lambda_2 < \lambda_1 < 0$ | $r > 1$ | stable node | enter tangent to the $z_1$-axis |
| $\lambda_1 < \lambda_2 < 0$ | $0 < r < 1$ | stable node | enter tangent to the $z_2$-axis |
| $\lambda_1 < 0 < \lambda_2$ | $r < 0$ | saddle | $z_1$-axis in, $z_2$-axis out; hyperbola-like curves $z_2 = \tilde C|z_1|^{-|r|}$ |
| $\lambda_2 < 0 < \lambda_1$ | $r < 0$ | saddle | $z_1$-axis out, $z_2$-axis in |
In words: at a node the orbits are tangent, at the origin, to the axis of the eigenvalue with the smaller modulus (the slow direction), and far away they become parallel to the other axis. (For $r > 1$, $dz_2/dz_1 = r\tilde Cz_1^{\,r-1} \to 0$ as $z_1 \to 0$; for $0 < r < 1$ it tends to infinity.)
6.2 Back to the original coordinates [Sec. 28.4, p. 279]
The map $\mathbf x = P\tilde{\mathbf x}$ sends the $\tilde x$-axis to the line through $\mathbf v_1$ and the $\tilde y$-axis to the line through $\mathbf v_2$. Straight lines stay straight and the origin stays fixed, so the picture is the uncoupled one, sheared and rotated. Practical recipe:
Example of the notes. Find the solution and draw the phase diagram of $\dot x = x + y$, $\dot y = 4x - 2y$, i.e. $A = \begin{pmatrix} 1 & 1 \\ 4 & -2\end{pmatrix}$. Eigenvalues: $\lambda^2 + \lambda - 6 = (\lambda - 2)(\lambda + 3) = 0$, $\lambda_1 = 2$, $\lambda_2 = -3$. Eigenvectors: $\mathbf v_1 = (1, 1)^T$ and $\mathbf v_2 = (1, -4)^T$ (Example 28.1). General solution $\mathbf x(t) = C_1e^{2t}(1, 1)^T + C_2e^{-3t}(1, -4)^T$. Phase diagram: in $\mathbf z = P^{-1}\mathbf x$, since $\lambda_2 < 0 < \lambda_1$, a saddle (last panel of Figure 4.1 (b)); to go back to $\mathbf x$, $\mathbf x = P\mathbf z = (\mathbf v_1, \mathbf v_2)\mathbf z$ sends the $z_1$-axis to the line of $\mathbf v_1$ (out) and the $z_2$-axis to the line of $\mathbf v_2$ (in): Figure 4.2.
- Find the eigenvalues and eigenvectors.
- Draw the eigenvector lines; put arrows pointing away from the origin if $\lambda > 0$ and towards it if $\lambda < 0$.
- Fill in the other trajectories consistently with the type (node or saddle), tangent to the slow direction near the origin and parallel to the fast one far away.
6.3 Stable and unstable manifolds [Sec. 28.5, pp. 281–282]
The stable manifold $W^s(\mathbf 0)$ is the set of initial points whose trajectories tend to the origin as $t \to +\infty$; the unstable manifold $W^u(\mathbf 0)$ is the set whose trajectories tend to the origin as $t \to -\infty$.
| $W^s(\mathbf 0)$ | $W^u(\mathbf 0)$ | |
|---|---|---|
| Stable node | the whole plane | $\{\mathbf 0\}$ |
| Saddle | the eigenvector line of the negative eigenvalue | the eigenvector line of the positive eigenvalue |
| Unstable node | $\{\mathbf 0\}$ | the whole plane |
At a saddle the stable manifold is a separatrix: points on one side are carried to infinity along one branch of $W^u$, points on the other side along the opposite branch. In Example 28.3, initial data above the line $y = -4x$ end up with $x, y \to +\infty$; data below it with $x, y \to -\infty$.
According to the lecture map (ODE_lecture_theory_to_Robinson_chapters.txt), the lecture mentioned a "saddle node"; in the phase-portrait chapters Robinson says only "saddle point" or "saddle" and reserves "node" for stable and unstable nodes, so
the intended term is almost certainly "saddle". (In bifurcation theory, Topic 2, a saddle-node is something else.)
6.4 Special case: a zero eigenvalue [Notes, p. 9]
The notes end Case 1 with a special case: one eigenvalue is $0$. Say $\lambda_1 = 0 \neq \lambda_2$. Then $A\mathbf v_1 = \mathbf 0$, so every point $c\,\mathbf v_1$ of the line $\operatorname{span}\mathbf v_1$ is an equilibrium ($\mathbf x(t) \equiv c\mathbf v_1$ solves $\dot{\mathbf x} = A\mathbf x$, since both sides vanish). With $e^{0t} = 1$ the general solution of Case 1 becomes
$$\mathbf x(t) = C_1\mathbf v_1 + C_2e^{\lambda_2t}\mathbf v_2.$$Reading note: the notes write this case as $\mathbf x(t) = C_2\mathbf v_2e^{\lambda_2t}$ (reading of the handwriting), which keeps only the solutions with $C_1 = 0$. The term $C_1\mathbf v_1$ is needed: it is a solution on its own because $A\mathbf v_1 = 0\cdot\mathbf v_1 = \mathbf 0$, and without it the initial value $\mathbf x(0) = C_1\mathbf v_1 + C_2\mathbf v_2$ could not be arbitrary.
- Each orbit is a straight half-line parallel to $\mathbf v_2$, through the point $C_1\mathbf v_1$ of the line of equilibria.
- $\lambda_2 < 0$: every solution tends to $C_1\mathbf v_1$; $\lambda_2 > 0$: every solution leaves the line. The origin is no longer an isolated equilibrium ($\det A = \lambda_1\lambda_2 = 0$).
- In the $\mathbf z$-plane: $z_1 = C_1$ is constant and $z_2 = C_2e^{\lambda_2t}$, so the orbits are vertical lines and the $z_1$-axis is made of equilibria.
- $\det(A - \lambda I) = (-1 - \lambda)^2 - 1 = \lambda^2 + 2\lambda = \lambda(\lambda + 2) = 0$: $\lambda_1 = 0$, $\lambda_2 = -2$.
- $\lambda_1 = 0$: $-a + b = 0$, $\mathbf v_1 = (1, 1)^T$. $\lambda_2 = -2$: $(A + 2I)\mathbf v = \mathbf 0$ gives $a + b = 0$, $\mathbf v_2 = (1, -1)^T$.
- $\mathbf x(t) = C_1(1, 1)^T + C_2e^{-2t}(1, -1)^T$. From $\mathbf x(0) = (x_{10}, x_{20})^T$: $C_1 + C_2 = x_{10}$, $C_1 - C_2 = x_{20}$, so $C_1 = \tfrac12(x_{10} + x_{20})$, $C_2 = \tfrac12(x_{10} - x_{20})$.
- As $t \to \infty$, $\mathbf x(t) \to C_1(1, 1)^T$: the solution slides along the direction $(1, -1)^T$ onto the diagonal, keeping $x_1 + x_2$ constant (indeed $\dot x_1 + \dot x_2 = 0$).
Check: $A(1, 1)^T = (0, 0)^T$ ✓ and $A(1, -1)^T = (-2, 2)^T = -2(1, -1)^T$ ✓. Exercise 11 of the Chapter 4 (part 1) sheet treats this case with other matrices.
7. Complex eigenvalues: spirals and centres
Let $A$ be a real $2\times2$ matrix whose characteristic polynomial has a negative discriminant, $\Delta = (\operatorname{tr}A)^2 - 4\det A < 0$. Its eigenvalues are then a complex-conjugate pair
$$\lambda_\pm = \rho \pm i\omega, \qquad \rho = \tfrac12\operatorname{tr}A,\quad \omega = \tfrac12\sqrt{-\Delta} > 0, \tag{7.1}$$and the eigenvector of $\lambda_-$ is the complex conjugate of the eigenvector of $\lambda_+$ (because $A$ is real: conjugate $A\mathbf v = \lambda\mathbf v$ to get $A\bar{\mathbf v} = \bar\lambda\bar{\mathbf v}$). Write the eigenvector of $\lambda_+$ through its real and imaginary parts, $\mathbf v = \mathbf u + i\mathbf w$ with $\mathbf u, \mathbf w \in \mathbb R^2$.
Suppose $A$ has an eigenvalue $\lambda = \rho \pm i\omega$. Then it has eigenvectors $\mathbf v_\pm = \mathbf v_1 \pm i\,\mathbf v_2$ (linear algebra). The notes reach the canonical form through a complex diagonalisation (Section 7.1) and give
$$\text{general solution:}\quad \mathbf x(t) = C_1\operatorname{Re}\big\{\mathbf v e^{\lambda t}\big\} + C_2\operatorname{Im}\big\{\mathbf v e^{\lambda t}\big\},$$ $$\text{particular solution:}\quad \mathbf x(t) = r(0)\,e^{\rho t}\Big(\cos\big(\theta(0) - \omega t\big)\,\mathbf v_1 + \sin\big(\theta(0) - \omega t\big)\,\mathbf v_2\Big).$$Both are derived below (Sections 7.3 and 7.8).
The class notes and Robinson both write $\mathbf v_1 + i\mathbf v_2$ for the eigenvector; Robinson writes $P$ for the change-of-basis matrix and $\tilde{\mathbf x}$ for the new variable, while the notes use $Q$ and $\mathbf z$, and so does this section: $\mathbf u \equiv \mathbf v_1$, $\mathbf w \equiv \mathbf v_2$, $Q \equiv P$ (Robinson) and $\mathbf z \equiv \tilde{\mathbf x}$. Section 5 called the matrix of eigenvectors $P$; from here on every change of basis is called $Q$.
7.1 The route of the notes: a complex diagonalisation [Notes, p. 10]
The notes reach the same matrix $J$ from the complex eigenvectors. With $\mathbf v_\pm = \mathbf u \pm i\mathbf w$ and $\lambda_\pm = \rho \pm i\omega$, $A$ has the complex diagonalisation
$$A = PDP^{-1}, \qquad P = (\mathbf v_+ \mid \mathbf v_-), \qquad D = \operatorname{diag}(\lambda_+, \lambda_-) = \begin{pmatrix} \rho + i\omega & 0 \\ 0 & \rho - i\omega\end{pmatrix}.$$- Real vectors from the complex ones. Adding and subtracting $\mathbf v_\pm = \mathbf u \pm i\mathbf w$: $\ \mathbf u = \dfrac{\mathbf v_+ + \mathbf v_-}{2}$ and $\mathbf w = -i\,\dfrac{\mathbf v_+ - \mathbf v_-}{2}$ (check: $-i\cdot\frac{2i\mathbf w}{2} = \mathbf w$).
- The matrix $Q$ in terms of $P$. Each column of $Q = (\mathbf u \mid \mathbf w)$ is a combination of the columns of $P$: $$Q = P\,M, \qquad M = \begin{pmatrix} 1/2 & -i/2 \\ 1/2 & i/2\end{pmatrix},$$ since $P\,(1/2, 1/2)^T = \tfrac12(\mathbf v_+ + \mathbf v_-) = \mathbf u$ and $P\,(-i/2, i/2)^T = -\tfrac i2(\mathbf v_+ - \mathbf v_-) = \mathbf w$.
- Its inverse. $\det M = \tfrac12\cdot\tfrac i2 - \big(-\tfrac i2\big)\tfrac12 = \tfrac i2$, so $M^{-1} = \dfrac{1}{i/2}\begin{pmatrix} i/2 & i/2 \\ -1/2 & 1/2\end{pmatrix} = \begin{pmatrix} 1 & 1 \\ i & -i\end{pmatrix}$, and $P = Q\begin{pmatrix} 1 & 1 \\ i & -i\end{pmatrix}$, as in the notes.
- The new variable. With $\mathbf z = Q^{-1}\mathbf x$, $\ \dot{\mathbf z} = Q^{-1}A\mathbf x = Q^{-1}PDP^{-1}Q\,\mathbf z = M^{-1}DM\,\mathbf z$, because $Q^{-1}P = M^{-1}$ and $P^{-1}Q = M$.
- Multiply. $DM = \begin{pmatrix} \frac{\rho + i\omega}{2} & -i\,\frac{\rho + i\omega}{2} \\ \frac{\rho - i\omega}{2} & i\,\frac{\rho - i\omega}{2}\end{pmatrix}$, and then $$M^{-1}DM = \frac12\begin{pmatrix} (\rho + i\omega) + (\rho - i\omega) & -i(\rho + i\omega) + i(\rho - i\omega) \\ i(\rho + i\omega) - i(\rho - i\omega) & (\rho + i\omega) + (\rho - i\omega)\end{pmatrix} = \frac12\begin{pmatrix} 2\rho & 2\omega \\ -2\omega & 2\rho\end{pmatrix} = \begin{pmatrix} \rho & \omega \\ -\omega & \rho\end{pmatrix}.$$ (Top right: $-i\rho + \omega + i\rho + \omega = 2\omega$; bottom left: $i\rho - \omega - i\rho - \omega = -2\omega$.)
So, with $\mathbf z = Q^{-1}\mathbf x$, $\ \dot{\mathbf z} = J\mathbf z$ with $J = \begin{pmatrix} \rho & \omega \\ -\omega & \rho\end{pmatrix}$. Sections 7.4–7.5 reach the same matrix with real vectors only (Robinson's route, Eq. (7.6)), and Section 7.6 solves $\dot{\mathbf z} = J\mathbf z$ in polar coordinates, as the notes do: $z_1 = r\cos\theta$, $z_2 = r\sin\theta$, $r^2 = z_1^2 + z_2^2$, $\theta = \arctan(z_2/z_1)$ [Notes, p. 10].
7.2 A guided example: what to expect before the theory [Examples 27.5 and 29.1, pp. 265–266 and 286]
Here $A = \begin{pmatrix} 2 & 5 \\ -2 & 0\end{pmatrix}$, $\operatorname{tr}A = 2$, $\det A = 2\cdot 0 - 5\cdot(-2) = 10$.
- Eigenvalues. $\det(A - \lambda I) = (2 - \lambda)(0 - \lambda) - (5)(-2) = \lambda^2 - 2\lambda + 10 = 0$, so $\lambda = \dfrac{2 \pm \sqrt{4 - 40}}{2} = 1 \pm 3i$. Hence $\rho = 1$, $\omega = 3$ and $\Delta = -36 < 0$.
- Type. $\rho = 1 > 0$: an unstable spiral (radius grows like $e^{t}$).
- Eigenvector for $\lambda = 1 + 3i$. The first row of $(A - \lambda I)\mathbf v = \mathbf 0$ is $(2 - 1 - 3i)v_1 + 5v_2 = (1 - 3i)v_1 + 5v_2 = 0$. Take $v_1 = 5$; then $v_2 = -(1 - 3i) = -1 + 3i$, and $\mathbf v = \begin{pmatrix} 5 \\ -1 + 3i\end{pmatrix} = \underbrace{\begin{pmatrix} 5 \\ -1\end{pmatrix}}_{\mathbf u} + i\underbrace{\begin{pmatrix} 0 \\ 3\end{pmatrix}}_{\mathbf w}$. Check with the second row: $-2\cdot 5 + (0 - \lambda)(-1 + 3i) = -10 - (1 + 3i)(-1 + 3i) = -10 - (-1 - 9) = 0$. ✓
- Sense of rotation, without any change of basis. Evaluate the vector field at $(0, 1)^T$ (top of the plane): $\dot{\mathbf x} = A(0,1)^T = (5, 0)^T$, which points to the right. At the top, motion to the right is clockwise. Confirm at $(1, 0)^T$: $\dot{\mathbf x} = (2, -2)^T$ points down on the right-hand side, again clockwise.
- Sketch. A spiral that turns clockwise and moves outward, drawn as a distorted spiral because it lives in the $(x, y)$ basis (Fig. 4.5, right). The rest of this section explains why it is distorted exactly as it is, and gives its equation.
7.3 The real general solution [Sec. 29.1, pp. 285–286]
As for distinct real eigenvalues, $e^{\lambda_+t}\mathbf v$ and $e^{\lambda_-t}\bar{\mathbf v}$ solve $\dot{\mathbf x} = A\mathbf x$, and so does every combination $Ce^{\lambda_+t}\mathbf v + De^{\lambda_-t}\bar{\mathbf v}$ with complex $C, D$. To get a real solution choose $D = \bar C$: the second term is then the conjugate of the first and the sum is twice a real part, because $z + \bar z = 2\operatorname{Re}z$:
$$\mathbf x(t) = 2\operatorname{Re}\Big[Ce^{(\rho + i\omega)t}(\mathbf u + i\mathbf w)\Big]. \tag{7.2}$$Write $C = \alpha + i\beta$ and expand with Euler's formula $e^{i\omega t} = \cos\omega t + i\sin\omega t$:
- $Ce^{i\omega t} = (\alpha + i\beta)(\cos\omega t + i\sin\omega t) = (\alpha\cos\omega t - \beta\sin\omega t) + i(\beta\cos\omega t + \alpha\sin\omega t)$.
- Multiply by $\mathbf u + i\mathbf w$ and keep only the real part ($i\cdot i = -1$): $\operatorname{Re}\big[(a + ib)(\mathbf u + i\mathbf w)\big] = a\mathbf u - b\mathbf w$, with $a = \alpha\cos\omega t - \beta\sin\omega t$ and $b = \beta\cos\omega t + \alpha\sin\omega t$.
- Therefore $\mathbf x(t) = 2e^{\rho t}\big[(\alpha\cos\omega t - \beta\sin\omega t)\mathbf u - (\beta\cos\omega t + \alpha\sin\omega t)\mathbf w\big]$.
- Put $a_1 = 2\alpha$ and $b_1 = -2\beta$ (two arbitrary real constants):
This is Robinson's (29.2) with $\mathbf v_1 = \mathbf u$, $\mathbf v_2 = \mathbf w$. It already contains everything, but it does not show how the solution moves. The change of basis does.
The form of the notes. The real and imaginary parts of the single complex solution $\mathbf v e^{\lambda t}$ are themselves real solutions (the real and imaginary parts of $\dot{\mathbf y} - A\mathbf y = \mathbf 0$ vanish separately because $A$ is real). By Euler's formula,
$$\mathbf v e^{\lambda t} = e^{\rho t}(\mathbf u + i\mathbf w)(\cos\omega t + i\sin\omega t) = e^{\rho t}\Big[\big(\cos\omega t\,\mathbf u - \sin\omega t\,\mathbf w\big) + i\big(\sin\omega t\,\mathbf u + \cos\omega t\,\mathbf w\big)\Big],$$so
$$C_1\operatorname{Re}\big\{\mathbf v e^{\lambda t}\big\} + C_2\operatorname{Im}\big\{\mathbf v e^{\lambda t}\big\} = e^{\rho t}\Big[\big(C_1\cos\omega t + C_2\sin\omega t\big)\mathbf u + \big(C_2\cos\omega t - C_1\sin\omega t\big)\mathbf w\Big],$$which is (7.3) with $a_1 = C_1$ and $b_1 = C_2$ [Notes, p. 10].
7.4 How $A$ acts on $\mathbf u$ and $\mathbf w$ [Sec. 29.2, p. 287, Eq. (29.3)]
Separate the real and imaginary parts of $A\mathbf v = \lambda\mathbf v$. The left side is $A(\mathbf u + i\mathbf w) = A\mathbf u + iA\mathbf w$ (both real vectors). The right side is
$$(\rho + i\omega)(\mathbf u + i\mathbf w) = \rho\mathbf u + i\rho\mathbf w + i\omega\mathbf u + i^2\omega\mathbf w = (\rho\mathbf u - \omega\mathbf w) + i(\omega\mathbf u + \rho\mathbf w).$$Two complex vectors are equal exactly when their real parts and their imaginary parts are equal:
$$\boxed{A\mathbf u = \rho\mathbf u - \omega\mathbf w, \qquad A\mathbf w = \omega\mathbf u + \rho\mathbf w.} \tag{7.4}$$So $A$ maps the plane $\operatorname{span}\{\mathbf u, \mathbf w\}$ into itself: it scales by $\rho$ and mixes $\mathbf u$ with $\mathbf w$ through the rotation angle $\omega$. There is no real eigenvector, but there is an invariant basis.
If $\omega \neq 0$, the real vectors $\mathbf u, \mathbf w$ form a basis of $\mathbb R^2$.
Suppose they were dependent. Then the eigenvector $\mathbf v = \mathbf u + i\mathbf w$ would be a complex multiple of one real vector $\mathbf r \neq \mathbf 0$ (if $\mathbf w = c\mathbf u$ then $\mathbf v = (1 + ic)\mathbf u$; if $\mathbf u = \mathbf 0$ then $\mathbf v = i\mathbf w$): $\mathbf v = \kappa\mathbf r$ with $\kappa \in \mathbb C\setminus\{0\}$. Then $A\mathbf v = \kappa A\mathbf r$ and $\lambda\mathbf v = \lambda\kappa\mathbf r$, so $A\mathbf r = \lambda\mathbf r$. The left side is a real vector, so $\lambda\mathbf r$ is real and, as $\mathbf r \neq \mathbf 0$ is real, $\lambda$ must be real. This contradicts $\operatorname{Im}\lambda = \omega \neq 0$.
7.5 The change of basis $\mathbf z = Q^{-1}\mathbf x$ [Sec. 29.2, pp. 287–288, Eqs. (29.4)–(29.7)]
Because $\mathbf u, \mathbf w$ are a basis, every $\mathbf x$ has unique coordinates $(z_1, z_2)$: $\mathbf x = z_1\mathbf u + z_2\mathbf w$. In matrix form, with the columns in the order $\mathbf u$, then $\mathbf w$,
$$Q = [\,\mathbf u \mid \mathbf w\,], \qquad \mathbf x = Q\mathbf z, \qquad \mathbf z = Q^{-1}\mathbf x. \tag{7.5}$$$Q$ is constant, so differentiating $\mathbf z = Q^{-1}\mathbf x$ gives $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = Q^{-1}AQ\,\mathbf z$. Compute $Q^{-1}AQ$ column by column, using $AQ = [A\mathbf u \mid A\mathbf w]$ and (7.4):
$$AQ = [\rho\mathbf u - \omega\mathbf w \mid \omega\mathbf u + \rho\mathbf w] = [\,\mathbf u \mid \mathbf w\,]\begin{pmatrix} \rho & \omega \\ -\omega & \rho\end{pmatrix} = Q\,J.$$The last equality is the rule "$Q M$ has as its $j$-th column the combination of the columns of $Q$ with the coefficients of the $j$-th column of $M$": the first column of $J$ is $(\rho, -\omega)^T$, giving $\rho\mathbf u - \omega\mathbf w$, and the second is $(\omega, \rho)^T$, giving $\omega\mathbf u + \rho\mathbf w$. Multiplying on the left by $Q^{-1}$:
$$\boxed{Q^{-1}AQ = J = \begin{pmatrix} \rho & \omega \\ -\omega & \rho\end{pmatrix}, \qquad \dot z_1 = \rho z_1 + \omega z_2,\quad \dot z_2 = -\omega z_1 + \rho z_2.} \tag{7.6}$$With $Q = [\mathbf u \mid \mathbf w]$ (Robinson's choice) the canonical matrix is $J = \begin{pmatrix} \rho & \omega \\ -\omega & \rho\end{pmatrix}$ and the rotation in the $\mathbf z$-plane is clockwise. With the opposite order $Q' = [\mathbf w \mid \mathbf u]$ the same computation gives $J' = \begin{pmatrix} \rho & -\omega \\ \omega & \rho\end{pmatrix} = J^{T}$ and an anticlockwise rotation. Both are correct; this page always uses $[\mathbf u \mid \mathbf w]$, so the $\mathbf z$-plane always turns clockwise.
7.6 Solving $\dot{\mathbf z} = J\mathbf z$ in polar coordinates [Sec. 29.2, pp. 288–290]
The two equations (7.6) are coupled, but they say "rotate and scale", which polar coordinates $z_1 = r\cos\theta$, $z_2 = r\sin\theta$ express directly.
- The radius. $r^2 = z_1^2 + z_2^2$. Differentiate with the chain rule: $2r\dot r = 2z_1\dot z_1 + 2z_2\dot z_2$. Substitute (7.6): $2z_1(\rho z_1 + \omega z_2) + 2z_2(-\omega z_1 + \rho z_2) = 2\rho z_1^2 + 2\omega z_1z_2 - 2\omega z_1z_2 + 2\rho z_2^2 = 2\rho r^2$. Hence $\dot r = \rho r$.
- The angle. $\theta = \arctan(z_2/z_1)$. By the chain rule, with $\dfrac{d}{ds}\arctan s = \dfrac{1}{1 + s^2}$, and the quotient rule for $s = z_2/z_1$: $$\dot\theta = \frac{1}{1 + (z_2/z_1)^2}\cdot\frac{z_1\dot z_2 - z_2\dot z_1}{z_1^2} = \frac{z_1\dot z_2 - z_2\dot z_1}{z_1^2 + z_2^2}.$$ Substitute (7.6): $z_1\dot z_2 - z_2\dot z_1 = z_1(-\omega z_1 + \rho z_2) - z_2(\rho z_1 + \omega z_2) = -\omega z_1^2 - \omega z_2^2 = -\omega r^2$. Hence $\dot\theta = -\omega$.
- Integrate the two decoupled equations $\dot r = \rho r$ and $\dot\theta = -\omega$. For the first, $\dfrac{d}{dt}\big(e^{-\rho t}r\big) = e^{-\rho t}(\dot r - \rho r) = 0$, so $e^{-\rho t}r$ is constant, and Barrow's rule between $0$ and $t$ gives $e^{-\rho t}r(t) - r(0) = 0$. For the second, $\theta(t) - \theta(0) = \int_0^t(-\omega)\,ds = -\omega t$. $$r(t) = r_0e^{\rho t}, \qquad \theta(t) = \theta_0 - \omega t.$$
- Back to Cartesian $\mathbf z$. With $z_{10} = r_0\cos\theta_0$, $z_{20} = r_0\sin\theta_0$ and the subtraction formulas $\cos(\theta_0 - \omega t) = \cos\theta_0\cos\omega t + \sin\theta_0\sin\omega t$, $\sin(\theta_0 - \omega t) = \sin\theta_0\cos\omega t - \cos\theta_0\sin\omega t$: $$\mathbf z(t) = e^{\rho t}\begin{pmatrix} \cos\omega t & \sin\omega t \\ -\sin\omega t & \cos\omega t\end{pmatrix}\begin{pmatrix} z_{10} \\ z_{20}\end{pmatrix}. \tag{7.7}$$
- Check by substitution. $\dot z_1 = \rho z_1 + e^{\rho t}\omega(-z_{10}\sin\omega t + z_{20}\cos\omega t) = \rho z_1 + \omega z_2$ ✓, and likewise $\dot z_2 = -\omega z_1 + \rho z_2$ ✓.
A shortcut worth knowing: the complex variable $\zeta = z_1 + iz_2$ satisfies $\dot\zeta = (\rho - i\omega)\zeta$, so $\zeta = \zeta_0e^{(\rho - i\omega)t}$, whose real and imaginary parts are exactly (7.7).
7.7 Phase portrait in the $\mathbf z$-plane: spirals and centre [Sec. 29.2, pp. 289–291]
Eliminate $t$ using $t = (\theta_0 - \theta)/\omega$: the orbits are the logarithmic spirals $r = r_0\exp\big(\rho(\theta_0 - \theta)/\omega\big)$. The motion is always clockwise because $\dot\theta = -\omega < 0$; the sign of $\rho$ decides the radius:
| Sign of $\rho$ | Radius $r(t) = r_0e^{\rho t}$ | Origin | Stability |
|---|---|---|---|
| $\rho < 0$ | $\to 0$ as $t \to +\infty$ (and $\to \infty$ as $t \to -\infty$) | stable spiral | asymptotically stable |
| $\rho > 0$ | $\to \infty$ as $t \to +\infty$ ($\to 0$ as $t \to -\infty$) | unstable spiral | unstable |
| $\rho = 0$ | constant, $r(t) = r_0$ | centre: closed circles, period $T = 2\pi/\omega$ | stable, not attracting |
The solution of a spiral keeps oscillating: each component is an exponentially modulated sine/cosine, and one full turn takes $2\pi/\omega$ whatever $\rho$ is. A larger ratio $|\rho|/\omega$ gives a faster-collapsing (or exploding) spiral per turn: the radius changes by the factor $e^{2\pi\rho/\omega}$ in one revolution.
7.8 Back to $\mathbf x = Q\mathbf z$ [Sec. 29.3, pp. 291–293]
Write $\mathbf x = Q\mathbf z = z_1\mathbf u + z_2\mathbf w$ and insert (7.7):
$$\mathbf x(t) = e^{\rho t}\Big[\big(z_{10}\cos\omega t + z_{20}\sin\omega t\big)\mathbf u + \big(z_{20}\cos\omega t - z_{10}\sin\omega t\big)\mathbf w\Big]. \tag{7.8}$$This is (7.3) with $a_1 = z_{10}$ and $b_1 = z_{20}$: the arbitrary constants are the initial coordinates, $(z_{10}, z_{20})^T = Q^{-1}\mathbf x_0$. In components, if $\mathbf u = (u_1, u_2)^T$ and $\mathbf w = (w_1, w_2)^T$, $x_1 = z_1u_1 + z_2w_1$ and $x_2 = z_1u_2 + z_2w_2$: each coordinate of $\mathbf x$ is a sum of $e^{\rho t}\cos\omega t$ and $e^{\rho t}\sin\omega t$.
The particular solution of the notes. In polar form $\mathbf z = r(\cos\theta, \sin\theta)^T$ with $r = r(0)e^{\rho t}$ and $\theta = \theta(0) - \omega t$, so $\mathbf x = z_1\mathbf u + z_2\mathbf w$ reads [Notes, p. 10]
$$\mathbf x(t) = r(0)\,e^{\rho t}\Big(\cos\big(\theta(0) - \omega t\big)\,\mathbf v_1 + \sin\big(\theta(0) - \omega t\big)\,\mathbf v_2\Big), \qquad \mathbf v_1 = \mathbf u,\ \mathbf v_2 = \mathbf w.$$Geometry. $Q$ is a linear map: it sends the origin to the origin and straight lines to straight lines, and it deforms circles into ellipses.
- The $z_1$-axis is mapped to the line through $\mathbf u$ and the $z_2$-axis to the line through $\mathbf w$. Neither line is invariant (the flow rotates), so, unlike the real-eigenvalue case, they are a coordinate frame and not trajectories.
- A circle $|\mathbf z| = r_0$ becomes an ellipse through $\pm r_0\mathbf u$ and $\pm r_0\mathbf w$, with $\mathbf u, \mathbf w$ as conjugate semi-diameters. For $\rho = 0$ these ellipses are the orbits; for $\rho \neq 0$ the spirals of the $\mathbf z$-plane become elliptical spirals, tilted along $\mathbf u$ and $\mathbf w$.
- Straight lines and ratios are preserved, so the qualitative picture (spiral in, spiral out, closed loops) does not change: this is why the phase portrait can be drawn from the eigenvalues alone [Sec. 29.3, p. 291].
7.9 The sense of rotation in the $\mathbf x$-plane [Sec. 29.3, Example 29.3, pp. 291–293]
In the $\mathbf z$-plane the rotation is clockwise by construction. A linear map with $\det Q > 0$ preserves orientation, so the rotation in the $\mathbf x$-plane is also clockwise; if $\det Q < 0$ the map reflects the plane and the rotation becomes anticlockwise. Rather than computing $\det Q$, use the vector field itself:
Evaluate $\dot{\mathbf x} = A\mathbf x$ at $(0, 1)^T$: $A(0,1)^T = (a_{12}, a_{22})^T$. At the top of the plane, motion to the right ($a_{12} > 0$) is clockwise; motion to the left ($a_{12} < 0$) is anticlockwise. Cross-check at $(1, 0)^T$: $A(1,0)^T = (a_{11}, a_{21})^T$; on the right-hand side, $a_{21} < 0$ (down) is clockwise.
The test never gives $a_{12} = 0$ for a spiral or centre: the discriminant is $\Delta = (a_{11} - a_{22})^2 + 4a_{12}a_{21} < 0$, which forces $a_{12}a_{21} < 0$, so $a_{12}$ and $a_{21}$ are non-zero and have opposite signs, and the two checks always agree.
7.10 Two examples solved from start to finish
From 7.2: $\lambda = 1 \pm 3i$ ($\rho = 1$, $\omega = 3$), $\mathbf u = (5, -1)^T$, $\mathbf w = (0, 3)^T$.
- Check (7.4). $A\mathbf u = \begin{pmatrix} 2\cdot5 + 5\cdot(-1) \\ -2\cdot5\end{pmatrix} = \begin{pmatrix} 5 \\ -10\end{pmatrix}$ and $\rho\mathbf u - \omega\mathbf w = \begin{pmatrix} 5 \\ -1\end{pmatrix} - 3\begin{pmatrix} 0 \\ 3\end{pmatrix} = \begin{pmatrix} 5 \\ -10\end{pmatrix}$ ✓. $A\mathbf w = \begin{pmatrix} 15 \\ 0\end{pmatrix}$ and $\omega\mathbf u + \rho\mathbf w = 3\begin{pmatrix} 5 \\ -1\end{pmatrix} + \begin{pmatrix} 0 \\ 3\end{pmatrix} = \begin{pmatrix} 15 \\ 0\end{pmatrix}$ ✓.
- The matrices. $Q = \begin{pmatrix} 5 & 0 \\ -1 & 3\end{pmatrix}$, $\det Q = 15 - 0 = 15 > 0$, and $Q^{-1} = \dfrac1{15}\begin{pmatrix} 3 & 0 \\ 1 & 5\end{pmatrix}$ (swap the diagonal, change the sign of the off-diagonal, divide by the determinant). Check: $Q^{-1}Q = \dfrac1{15}\begin{pmatrix} 15 & 0 \\ 5 - 5 & 15\end{pmatrix} = I$ ✓.
- Canonical form. $AQ = [A\mathbf u \mid A\mathbf w] = \begin{pmatrix} 5 & 15 \\ -10 & 0\end{pmatrix}$ and $$Q^{-1}AQ = \frac1{15}\begin{pmatrix} 3 & 0 \\ 1 & 5\end{pmatrix}\begin{pmatrix} 5 & 15 \\ -10 & 0\end{pmatrix} = \frac1{15}\begin{pmatrix} 15 & 45 \\ 5 - 50 & 15\end{pmatrix} = \begin{pmatrix} 1 & 3 \\ -3 & 1\end{pmatrix} = J. ✓$$
- Solution in $\mathbf z$ (7.7): $z_1 = e^{t}(z_{10}\cos3t + z_{20}\sin3t)$, $z_2 = e^{t}(z_{20}\cos3t - z_{10}\sin3t)$. Radius $r_0e^{t}$, angle $\theta_0 - 3t$ (one turn every $2\pi/3 \approx 2.09$, radius multiplied by $e^{2\pi/3} \approx 8.1$ per turn).
- Back to $\mathbf x$. $x_1 = 5z_1$ and $x_2 = -z_1 + 3z_2$, so $$x_1 = 5e^{t}\big(z_{10}\cos3t + z_{20}\sin3t\big), \qquad x_2 = e^{t}\big[(3z_{20} - z_{10})\cos3t - (3z_{10} + z_{20})\sin3t\big].$$ With $z_{10} = A$ and $z_{20} = B$ this is Robinson's formula for Example 29.1. The initial point is $\mathbf x_0 = (5z_{10},\ -z_{10} + 3z_{20})^T$, i.e. $(z_{10}, z_{20}) = \big(x_{10}/5,\ (x_{10} + 5x_{20})/15\big)$.
- Two explicit trajectories. $\mathbf z_0 = (1, 0)^T$ ($\mathbf x_0 = (5, -1)^T$): $\mathbf x(t) = e^{t}\big(5\cos3t,\ -\cos3t - 3\sin3t\big)^T$. $\mathbf z_0 = (0, 1)^T$ ($\mathbf x_0 = (0, 3)^T$): $\mathbf x(t) = e^{t}\big(5\sin3t,\ 3\cos3t - \sin3t\big)^T$. Quarter-turn check: for the second, $\mathbf x(\pi/6) = e^{\pi/6}(5, -1)^T = e^{\pi/6}\mathbf u$. The orbit passes from $\mathbf w$ (at $t = 0$) to $e^{\pi/6}\mathbf u$ a quarter of a turn later, clockwise in the $\mathbf z$-plane: $z_2$-axis to $z_1$-axis.
- Sense of rotation. $\det Q = 15 > 0$ and the test of 7.9 agrees ($a_{12} = 5 > 0$): clockwise in both planes.
- Step 1, eigenvalues. $\det(A - \lambda I) = (2 - \lambda)(-\lambda) + 10 = \lambda^2 - 2\lambda + 10 = 0$, $\lambda = \dfrac{2 \pm \sqrt{-36}}{2} = 1 \pm 3i$.
- Step 2, eigenvectors. $\left(\begin{smallmatrix} 1 \mp 3i & 5 \\ -2 & -1 \mp 3i\end{smallmatrix}\right)\mathbf v = \mathbf 0$ reduces to one row, giving $\mathbf v_\pm = \begin{pmatrix} -5 \\ 1 \mp 3i\end{pmatrix} = -\left[\begin{pmatrix} 5 \\ -1\end{pmatrix} \pm i\begin{pmatrix} 0 \\ 3\end{pmatrix}\right]$, a multiple of the eigenvector of 7.2, so $\mathbf v_1 = (5, -1)^T$, $\mathbf v_2 = (0, 3)^T$. (Reading note: the notes write $\mathbf v_\pm = (-5,\ 1 \mp 3i)^T = (5, -1)^T \pm i(0, 3)^T$; the two sides differ by the factor $-1$. Both are eigenvectors, and the page uses the right-hand side, as the notes do from Step 3 on.)
- Step 3, particular solution. $\mathbf x(t) = r(0)\,e^{t}\Big(\cos\big(\theta(0) - 3t\big)\begin{pmatrix} 5 \\ -1\end{pmatrix} + \sin\big(\theta(0) - 3t\big)\begin{pmatrix} 0 \\ 3\end{pmatrix}\Big)$.
- Step 4, phase diagram. In $\mathbf z = Q^{-1}\mathbf x$ with $Q = (\mathbf v_1 \mid \mathbf v_2)$, $\dot{\mathbf z} = \begin{pmatrix} 1 & 3 \\ -3 & 1\end{pmatrix}\mathbf z$, whose solution $\mathbf z(t) = r(0)e^{t}\big(\cos(\theta(0) - 3t),\ \sin(\theta(0) - 3t)\big)^T$ is an unstable spiral turning clockwise ($\rho = 1 > 0$, $\omega = 3 > 0$). In the $(x_1, x_2)$ axes it is still an unstable spiral; for the direction go back to the original system: on the positive $x_2$-axis $\dot x_1 = 2x_1 + 5x_2 = 5x_2 > 0$, so the motion is to the right at the top: clockwise.
$A = \begin{pmatrix} 1 & -2 \\ 1 & -1\end{pmatrix}$, $\operatorname{tr}A = 0$, $\det A = -1 + 2 = 1$.
- Eigenvalues. $\lambda^2 + 1 = 0$, $\lambda = \pm i$: $\rho = 0$, $\omega = 1$. A centre.
- Eigenvector for $\lambda = i$. First row: $(1 - i)v_1 - 2v_2 = 0$. Take $v_1 = 2$: $v_2 = 1 - i$. So $\mathbf v = (2, 1 - i)^T = \underbrace{(2, 1)^T}_{\mathbf u} + i\underbrace{(0, -1)^T}_{\mathbf w}$. Check (7.4): $A\mathbf u = (2 - 2,\ 2 - 1)^T = (0, 1)^T = -\mathbf w$ ✓ and $A\mathbf w = (2, 1)^T = \mathbf u$ ✓ (here $\rho = 0$, $\omega = 1$).
- Matrices. $Q = \begin{pmatrix} 2 & 0 \\ 1 & -1\end{pmatrix}$, $\det Q = -2 < 0$, $Q^{-1} = \dfrac{1}{-2}\begin{pmatrix} -1 & 0 \\ -1 & 2\end{pmatrix} = \begin{pmatrix} 1/2 & 0 \\ 1/2 & -1\end{pmatrix}$. Then $AQ = [A\mathbf u \mid A\mathbf w] = \begin{pmatrix} 0 & 2 \\ 1 & 1\end{pmatrix}$ and $$Q^{-1}AQ = \begin{pmatrix} 1/2 & 0 \\ 1/2 & -1\end{pmatrix}\begin{pmatrix} 0 & 2 \\ 1 & 1\end{pmatrix} = \begin{pmatrix} 0 & 1 \\ -1 & 0\end{pmatrix} = J. ✓$$
- Solution. $z_1 = z_{10}\cos t + z_{20}\sin t$, $z_2 = z_{20}\cos t - z_{10}\sin t$ (circles of radius $r_0$, period $2\pi$); $x_1 = 2z_1$, $x_2 = z_1 - z_2$.
- The orbits as curves in $\mathbf x$. Invert $Q$: $z_1 = x_1/2$, $z_2 = z_1 - x_2 = x_1/2 - x_2$. The circle $z_1^2 + z_2^2 = r_0^2$ is the ellipse $$\tfrac14x_1^2 + \big(\tfrac12x_1 - x_2\big)^2 = r_0^2 \iff \tfrac12x_1^2 - x_1x_2 + x_2^2 = r_0^2.$$ Check that $V = \tfrac12x_1^2 - x_1x_2 + x_2^2$ is conserved: $\dot V = (x_1 - x_2)\dot x_1 + (2x_2 - x_1)\dot x_2$, and with $\dot x_1 = x_1 - 2x_2$, $\dot x_2 = x_1 - x_2$: $(x_1 - x_2)(x_1 - 2x_2) + (2x_2 - x_1)(x_1 - x_2) = (x_1 - x_2)\big[(x_1 - 2x_2) - (x_1 - 2x_2)\big] = 0$ ✓.
- Sense of rotation. In the $\mathbf z$-plane: clockwise. But $\det Q = -2 < 0$. Test: $A(0,1)^T = (a_{12}, a_{22})^T = (-2, -1)^T$ points left at the top, so the motion in the $\mathbf x$-plane is anticlockwise. The reflection inside $Q$ has reversed the orientation, which is why one must test the vector field and not copy the sign of $\omega$.
7.11 Practice: a stable spiral
Practice item written for this page to consolidate the complex case (it is not in the class notes). Solve the system and sketch the phase portrait in both planes.
- Eigenvalues. $A = \begin{pmatrix} 0 & -5 \\ 1 & -2\end{pmatrix}$, $\operatorname{tr}A = -2$, $\det A = 5$: $\lambda^2 + 2\lambda + 5 = 0$, $\lambda = -1 \pm 2i$ ($\rho = -1$, $\omega = 2$). A stable spiral.
- Eigenvector for $\lambda = -1 + 2i$. Second row of $(A - \lambda I)\mathbf v = \mathbf 0$: $v_1 + (-2 - \lambda)v_2 = v_1 - (1 + 2i)v_2 = 0$. Take $v_2 = 1$: $v_1 = 1 + 2i$. So $\mathbf v = (1 + 2i, 1)^T = (1, 1)^T + i(2, 0)^T$: $\mathbf u = (1, 1)^T$, $\mathbf w = (2, 0)^T$. First row check: $(0 - \lambda)(1 + 2i) - 5 = (1 - 2i)(1 + 2i) - 5 = 5 - 5 = 0$ ✓.
- Matrices. $Q = \begin{pmatrix} 1 & 2 \\ 1 & 0\end{pmatrix}$, $\det Q = -2$, $Q^{-1} = \dfrac1{-2}\begin{pmatrix} 0 & -2 \\ -1 & 1\end{pmatrix} = \begin{pmatrix} 0 & 1 \\ 1/2 & -1/2\end{pmatrix}$. $AQ = [A\mathbf u \mid A\mathbf w] = \begin{pmatrix} -5 & 0 \\ -1 & 2\end{pmatrix}$ and $$Q^{-1}AQ = \begin{pmatrix} 0 & 1 \\ 1/2 & -1/2\end{pmatrix}\begin{pmatrix} -5 & 0 \\ -1 & 2\end{pmatrix} = \begin{pmatrix} -1 & 2 \\ -5/2 + 1/2 & -1\end{pmatrix} = \begin{pmatrix} -1 & 2 \\ -2 & -1\end{pmatrix}. ✓$$
- Solution. $\mathbf z(t) = e^{-t}\big(z_{10}\cos2t + z_{20}\sin2t,\ z_{20}\cos2t - z_{10}\sin2t\big)^T$, and $\mathbf x = Q\mathbf z = (z_1 + 2z_2,\ z_1)^T$. For $\mathbf x_0 = (1, 1)^T$: $\mathbf z_0 = Q^{-1}\mathbf x_0 = (1, 0)^T$ and $$\mathbf x(t) = e^{-t}\big(\cos2t - 2\sin2t,\ \cos2t\big)^T.$$ Check: $\dot x_1 = -5x_2$ requires $\dot z_1 + 2\dot z_2 = (-z_1 + 2z_2) + 2(-2z_1 - z_2) = -5z_1$ ✓, and $\dot x_2 = \dot z_1 = -z_1 + 2z_2 = x_1 - 2x_2$ ✓.
- Sense of rotation. Clockwise in $\mathbf z$, but $\det Q = -2 < 0$. Test: $a_{12} = -5 < 0$, so at $(0,1)^T$ the velocity is $(-5, -2)^T$, leftwards: anticlockwise in the $\mathbf x$-plane.
8. A repeated eigenvalue: stars and improper nodes
Suppose $A$ has only one eigenvalue $\lambda$. Two subcases: 3.1, two different (independent) eigenvectors $\{\mathbf v_1, \mathbf v_2\}$; 3.2, only one eigenvector $\{\mathbf v\}$, so $A$ cannot be diagonalised.
8.1 Complement: algebraic and geometric multiplicity [Sec. 30, p. 295 and Example 27.4, p. 265]
The characteristic polynomial of a $2\times2$ matrix has a double root exactly when $\Delta = (\operatorname{tr}A)^2 - 4\det A = 0$. Then
$$\lambda_1 = \lambda_2 = \lambda = \tfrac12\operatorname{tr}A, \qquad \det(A - \mu I) = (\mu - \lambda)^2.$$- The algebraic multiplicity of $\lambda$ is the multiplicity of the root: $m_a = 2$.
- The geometric multiplicity is the number of independent eigenvectors: $m_g = \dim\ker(A - \lambda I)$. It is at least $1$ (an eigenvalue has an eigenvector) and at most $m_a$, so $m_g \in \{1, 2\}$.
The two values of $m_g$ produce two very different phase portraits, and no other possibility exists for a real $2\times2$ matrix with $\Delta = 0$.
8.2 Subcase 3.1, $m_g = 2$: the stars [Sec. 30.1, p. 295]
We can diagonalise $A$ as $A = PDP^{-1}$, where $P = (\mathbf v_1, \mathbf v_2)$ and $D = \begin{pmatrix} \lambda & 0 \\ 0 & \lambda\end{pmatrix} = \lambda I$. Then $A = P(\lambda I)P^{-1} = \lambda PP^{-1} = \lambda I$, so $(A - \lambda I)\mathbf v = 0\,\mathbf v = \mathbf 0$ for every $\mathbf v \in \mathbb R^2$: every vector is an eigenvector, and $\mathbf x(t) = \mathbf x_0e^{\lambda t}$ is a particular solution for every $\mathbf x_0$. Phase diagrams: unstable star ($\lambda > 0$) and stable star ($\lambda < 0$); every line through the origin is a solution.
If $\lambda$ is a repeated eigenvalue of the $2\times2$ matrix $A$ and $m_g = 2$, then $A = \lambda I = \begin{pmatrix} \lambda & 0 \\ 0 & \lambda\end{pmatrix}$.
Let $\mathbf v_1, \mathbf v_2$ be independent eigenvectors, $A\mathbf v_i = \lambda\mathbf v_i$. Put $B = [\mathbf v_1 \mid \mathbf v_2]$, invertible. Then $AB = [\lambda\mathbf v_1 \mid \lambda\mathbf v_2] = \lambda B$, and multiplying on the right by $B^{-1}$ gives $A = \lambda BB^{-1} = \lambda I$. (Equivalently, $\ker(A - \lambda I)$ is a two-dimensional subspace of $\mathbb R^2$, so it is all of $\mathbb R^2$, and $(A - \lambda I)\mathbf x = \mathbf 0$ for every $\mathbf x$ means $A - \lambda I = O$.)
So every vector is an eigenvector, and the system uncouples in every basis: $\dot x_1 = \lambda x_1$, $\dot x_2 = \lambda x_2$. For any change of basis, $Q^{-1}(\lambda I)Q = \lambda I$, so there is no separate $\mathbf z$-plane. Integrate each equation, $\dfrac{d}{dt}(e^{-\lambda t}x_i) = 0$, and then
$$\mathbf x(t) = e^{\lambda t}\mathbf x_0 = e^{\lambda t}(c_1\mathbf v_1 + c_2\mathbf v_2) \quad\text{for any basis }\mathbf v_1, \mathbf v_2. \tag{8.1}$$The orbit through $\mathbf x_0$ stays on the straight line through the origin and $\mathbf x_0$: every line through the origin is invariant, and the point moves along it with speed proportional to its distance from the origin.
- Stable star, $\lambda < 0$: all solutions go to $\mathbf 0$ along straight rays (the origin is asymptotically stable).
- Unstable star, $\lambda > 0$: all solutions leave $\mathbf 0$ along straight rays.
- For $\lambda = 0$ we get $A = O$: every point is an equilibrium (Robinson Exercise 30.5).
8.3 Subcase 3.2, $m_g = 1$: the construction of the notes [Notes, p. 10; Robinson, Sec. 30.2, pp. 295–297]
Now $A \neq \lambda I$, there is a single eigendirection $\mathbf v$, and (8.1) cannot give the general solution (it has only one free constant). The notes build a second basis vector as follows [Notes, p. 10].
- Choose and define. Choose any $\mathbf v_2$ linearly independent of $\mathbf v$, and define $\mathbf v_1 = (A - \lambda I)\mathbf v_2$. $\mathbf v_1 \neq \mathbf 0$: otherwise $\mathbf v_2$ would be a second, independent eigenvector.
- Write $\mathbf v_1$ in the basis $\{\mathbf v_2, \mathbf v\}$ of $\mathbb R^2$: $\ \mathbf v_1 = \alpha\mathbf v_2 + \beta\mathbf v$.
- Apply $A - \lambda I$. Since $(A - \lambda I)\mathbf v = \mathbf 0$, $$(A - \lambda I)\mathbf v_1 = (A - \lambda I)(\alpha\mathbf v_2 + \beta\mathbf v) = \alpha(A - \lambda I)\mathbf v_2 = \alpha\mathbf v_1 \quad\Longrightarrow\quad \big(A - (\alpha + \lambda)I\big)\mathbf v_1 = \mathbf 0.$$
- Conclude $\alpha = 0$. $\mathbf v_1 \neq \mathbf 0$ is then an eigenvector for $\lambda + \alpha$; but the only eigenvalue is $\lambda$, so $\alpha = 0$. Hence $\mathbf v_1 = \beta\mathbf v$ is proportional to $\mathbf v$ (an eigenvector, $(A - \lambda I)\mathbf v_1 = \mathbf 0$), and $\{\mathbf v_1, \mathbf v_2\}$ is a basis.
- The relation between them. By definition $(A - \lambda I)\mathbf v_2 = \mathbf v_1$, i.e. $$A\mathbf v_2 = \mathbf v_1 + \lambda\mathbf v_2. \tag{8.2}$$
- The matrix. Let $P = (\mathbf v_1, \mathbf v_2)$. Then $AP = (A\mathbf v_1, A\mathbf v_2) = (\lambda\mathbf v_1,\ \mathbf v_1 + \lambda\mathbf v_2) = (\mathbf v_1, \mathbf v_2)\begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix}$, so $$A = P\begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix}P^{-1}.$$
$\mathbf v_1$ is the eigenvector and $\mathbf v_2$ a generalised eigenvector. Robinson builds the pair in the same order (p. 296). An equivalent route, often quicker by hand, starts from the eigenvector and solves for $\mathbf v_2$; it rests on the following lemma (complement).
If $\det(A - \mu I) = (\mu - \lambda)^2$, then $N = A - \lambda I$ satisfies $N^2 = O$.
The Cayley–Hamilton theorem for $2\times2$ matrices says that $A$ satisfies its own characteristic equation $\mu^2 - (\operatorname{tr}A)\mu + \det A = 0$, i.e. $A^2 - (\operatorname{tr}A)A + (\det A)I = O$. Here $\operatorname{tr}A = 2\lambda$ and $\det A = \lambda^2$, so $A^2 - 2\lambda A + \lambda^2I = (A - \lambda I)^2 = N^2 = O$.
The eigenvector-first route produces the same two vectors, linked as a chain:
- An eigenvector $\mathbf v_1 \neq \mathbf 0$: $N\mathbf v_1 = \mathbf 0$, i.e. $A\mathbf v_1 = \lambda\mathbf v_1$.
- A generalised eigenvector $\mathbf v_2$ with $N\mathbf v_2 = \mathbf v_1$, i.e. $(A - \lambda I)\mathbf v_2 = \mathbf v_1$, equivalently $$A\mathbf v_2 = \lambda\mathbf v_2 + \mathbf v_1. \tag{8.2}$$
Why $\mathbf v_2$ exists. Since $N^2 = O$, every vector $N\mathbf y$ satisfies $N(N\mathbf y) = \mathbf 0$, so the image of $N$ lies inside its kernel. Both are one-dimensional ($\dim\ker N = m_g = 1$ and $\dim\operatorname{im}N = 2 - 1 = 1$), so they coincide: $\operatorname{im}N = \ker N$. Hence the eigenvector $\mathbf v_1 \in \ker N$ is in the image of $N$ and the system $N\mathbf v_2 = \mathbf v_1$ is solvable (its two equations are proportional: $N$ has rank 1). The solution is not unique: $\mathbf v_2 + s\mathbf v_1$ works for every $s$, because $N\mathbf v_1 = \mathbf 0$. Rescaling $\mathbf v_1$ rescales $\mathbf v_2$ in the same way.
Why $\mathbf v_1, \mathbf v_2$ are a basis. If $a\mathbf v_1 + b\mathbf v_2 = \mathbf 0$, apply $N$: $bN\mathbf v_2 = b\mathbf v_1 = \mathbf 0$, so $b = 0$; then $a\mathbf v_1 = \mathbf 0$ gives $a = 0$.
The notes and Robinson (p. 296) pick any $\mathbf v_2$ not parallel to the eigenvector and set $\mathbf v_1 = (A - \lambda I)\mathbf v_2$; with the lemma, $N\mathbf v_1 = N^2\mathbf v_2 = \mathbf 0$ shows at once that $\mathbf v_1$ is an eigenvector (the notes prove it with the $\alpha = 0$ argument instead). The eigenvector-first order, used in Example 8.1 and in the algorithm at the end, gives the same chain up to the choice of $\mathbf v_2$.
8.4 The change of variable $\mathbf z = Q^{-1}\mathbf x$ and the Jordan block [Sec. 30.2, pp. 296–297, Eqs. (30.2)–(30.3)]
Take $Q = [\,\mathbf v_1 \mid \mathbf v_2\,]$ (eigenvector first). Column by column, with (8.2):
$$AQ = [A\mathbf v_1 \mid A\mathbf v_2] = [\lambda\mathbf v_1 \mid \mathbf v_1 + \lambda\mathbf v_2] = [\mathbf v_1 \mid \mathbf v_2]\begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix} = Q\,J.$$(First column of $J$: $(\lambda, 0)^T$ gives $\lambda\mathbf v_1$. Second column: $(1, \lambda)^T$ gives $1\cdot\mathbf v_1 + \lambda\mathbf v_2$.) Multiplying by $Q^{-1}$ on the left:
$$\boxed{Q^{-1}AQ = J = \begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix}, \qquad \dot z_1 = \lambda z_1 + z_2,\quad \dot z_2 = \lambda z_2.} \tag{8.3}$$This is partially uncoupled: the second equation involves only $z_2$, the first is driven by it. The notes call the matrix $P$ and define $\mathbf z = P^{-1}\mathbf x$, so that $\dot{\mathbf z} = P^{-1}A\mathbf x = P^{-1}P\begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix}P^{-1}\mathbf x = \begin{pmatrix} \lambda & 1 \\ 0 & \lambda\end{pmatrix}\mathbf z$ [Notes, p. 10].
8.5 Solving for $\mathbf z$ [Sec. 30.2, p. 297, Eq. (30.4)]
- The second equation. $\dot z_2 = \lambda z_2$. Multiply by $e^{-\lambda t}$: $\dfrac{d}{dt}\big(e^{-\lambda t}z_2\big) = e^{-\lambda t}(\dot z_2 - \lambda z_2) = 0$, so $e^{-\lambda t}z_2$ is constant: $z_2(t) = c_2e^{\lambda t}$ with $c_2 = z_2(0)$.
- The first equation becomes a forced linear equation. Substituting $z_2$: $\dot z_1 - \lambda z_1 = c_2e^{\lambda t}$. The integrating factor is $e^{-\lambda t}$ (its derivative is $-\lambda e^{-\lambda t}$, which is exactly the coefficient needed).
- Multiply and recognise a derivative. $e^{-\lambda t}\dot z_1 - \lambda e^{-\lambda t}z_1 = \dfrac{d}{dt}\big(e^{-\lambda t}z_1\big)$ by the product rule, and the right side is $e^{-\lambda t}c_2e^{\lambda t} = c_2$. So $$\frac{d}{dt}\big(e^{-\lambda t}z_1\big) = c_2.$$
- Integrate with Barrow's rule between $0$ and $t$: $\big[e^{-\lambda s}z_1(s)\big]_{s=0}^{s=t} = \displaystyle\int_0^tc_2\,ds = c_2t$, i.e. $e^{-\lambda t}z_1(t) - z_1(0) = c_2t$. With $c_1 = z_1(0)$: $$z_1(t) = (c_1 + c_2t)e^{\lambda t}, \qquad z_2(t) = c_2e^{\lambda t}. \tag{8.4}$$
- Check. $\dot z_1 = c_2e^{\lambda t} + \lambda(c_1 + c_2t)e^{\lambda t} = \lambda z_1 + z_2$ ✓.
In matrix form $\mathbf z(t) = e^{\lambda t}\begin{pmatrix} 1 & t \\ 0 & 1\end{pmatrix}\mathbf z_0$ (this is $e^{Jt}$, since $J = \lambda I + \begin{pmatrix} 0 & 1 \\ 0 & 0\end{pmatrix}$ and the second matrix squares to zero). The factor $t$ in $z_1$ is the vector version of the $te^{\lambda t}$ of the repeated characteristic root in Topic 3.
8.6 The general solution in $\mathbf x$ [Sec. 30.2, p. 297, Eq. (30.4)]
Undo the change of basis, $\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$:
$$\mathbf x(t) = (c_1 + c_2t)e^{\lambda t}\mathbf v_1 + c_2e^{\lambda t}\mathbf v_2 \;=\; \boxed{\,e^{\lambda t}\Big[c_1\mathbf v_1 + c_2\big(t\,\mathbf v_1 + \mathbf v_2\big)\Big]\,} \tag{8.5}$$- $c_1e^{\lambda t}\mathbf v_1$: the pure eigenmode, a motion along the line $\operatorname{span}\mathbf v_1$.
- $c_2e^{\lambda t}\mathbf v_2$: the generalised-eigenvector mode; on its own it is not a solution.
- $c_2te^{\lambda t}\mathbf v_1$: the secular term. It couples the two modes and is the only reason the orbit is not a straight line: as $t \to \pm\infty$ it dominates, so the orbits become parallel to $\mathbf v_1$ at both ends.
- The constants are the initial coordinates in the chain basis: $(c_1, c_2)^T = Q^{-1}\mathbf x_0$.
The notes write the same result with $C = c_2$ and $D = c_1$: $\ \mathbf x = P\mathbf z = (\mathbf v_1, \mathbf v_2)\,\mathbf z = (Ct + D)\,e^{\lambda t}\mathbf v_1 + C\,e^{\lambda t}\mathbf v_2$ [Notes, p. 10].
Direct check. With $\mathbf x = e^{\lambda t}\mathbf s(t)$, $\mathbf s = c_1\mathbf v_1 + c_2(t\mathbf v_1 + \mathbf v_2)$: $\dot{\mathbf x} = \lambda\mathbf x + e^{\lambda t}c_2\mathbf v_1$. And by (8.2), $A\mathbf s = c_1\lambda\mathbf v_1 + c_2\big(t\lambda\mathbf v_1 + \lambda\mathbf v_2 + \mathbf v_1\big) = \lambda\mathbf s + c_2\mathbf v_1$, so $A\mathbf x = \lambda\mathbf x + e^{\lambda t}c_2\mathbf v_1 = \dot{\mathbf x}$ ✓. The two solutions $e^{\lambda t}\mathbf v_1$ and $e^{\lambda t}(t\mathbf v_1 + \mathbf v_2)$ have Wronskian $e^{2\lambda t}\det[\mathbf v_1\ \mathbf v_2] \neq 0$, so they form a fundamental set.
8.7 The improper node in the $\mathbf z$-plane [Sec. 30.2, pp. 297–298]
Eliminate $t$ from (8.4) for $c_2 \neq 0$: since $z_2 = c_2e^{\lambda t}$, $t = \tfrac1\lambda\ln(z_2/c_2)$ and $c_2e^{\lambda t} = z_2$, so
$$z_1 = \Big(\frac{c_1}{c_2} + t\Big)z_2 = z_2\Big[K + \frac1\lambda\ln|z_2|\Big], \qquad K = \frac{c_1}{c_2} - \frac1\lambda\ln|c_2|. \tag{8.6}$$(The same curve from the slope: $\dfrac{dz_1}{dz_2} = \dfrac{\dot z_1}{\dot z_2} = \dfrac{\lambda z_1 + z_2}{\lambda z_2} = \dfrac{z_1}{z_2} + \dfrac1\lambda$, so $\dfrac{d}{dz_2}\dfrac{z_1}{z_2} = \dfrac1{\lambda z_2}$ and $\dfrac{z_1}{z_2} = \dfrac1\lambda\ln|z_2| + K$.)
- The $z_1$-axis is invariant. For $c_2 = 0$: $z_2 \equiv 0$ and $z_1 = c_1e^{\lambda t}$, two half-lines moving away from the origin if $\lambda > 0$ (towards it if $\lambda < 0$). This is the single eigendirection, $\operatorname{span}\mathbf v_1$.
- Where the orbits turn. The notes start the sketch from the line where $\dot z_1 = 0$: $\lambda z_1 + z_2 = 0$, i.e. $z_2 = -\lambda z_1$ [Notes, p. 10]. There the velocity $(0, \lambda z_2)$ is vertical, so every S-shaped orbit crosses this line with a vertical tangent, where $z_1$ reaches its extreme value.
- Direction of motion. $\dot z_1 = \lambda z_1 + z_2$ equals $z_2$ on the $z_2$-axis. So in the upper half-plane ($z_2 > 0$) orbits cross the $z_2$-axis moving right and in the lower half-plane moving left.
- Behaviour as $t \to \pm\infty$ (take $\lambda > 0$): for $t \to -\infty$ the orbit tends to the origin and $z_1/z_2 = c_1/c_2 + t \to -\infty$, so it arrives tangent to the $z_1$-axis. For $t \to +\infty$ the orbit escapes to infinity with $z_1/z_2 \sim t \to +\infty$: it becomes parallel to the $z_1$-axis. For $\lambda < 0$ everything is reversed in time.
- Shape. Near the origin $z_1 \approx (z_2/\lambda)\ln z_2$, which tends to $0$ with infinite slope $dz_1/dz_2 \to -\infty$ (for $\lambda>0$): the orbit leaves the origin horizontally, curls over the $z_2$-axis and flattens again. Upper and lower half-planes are point-symmetric ($\mathbf z \mapsto -\mathbf z$ maps solutions to solutions), giving an S-shaped pair of curves. The approach to the axis is only logarithmic ($z_1/z_2 = K + \tfrac1\lambda\ln|z_2|$ grows very slowly as $z_2 \to 0$), so at the scale of Fig. 4.9 the portrait looks like a fan of almost straight rays that bend sharply only close to the origin.
8.8 The improper node in the $\mathbf x$-plane [Sec. 30.2, p. 298]
The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line $\operatorname{span}\mathbf v_1$, the only invariant line, and the $z_2$-axis to $\operatorname{span}\mathbf v_2$ (a reference direction, not an orbit). Straight lines go to straight lines, so the properties of the $\mathbf z$-portrait that are stated with lines survive:
- All orbits are tangent to $\operatorname{span}\mathbf v_1$ at the origin, and become parallel to it far away, bending in an S. In the words of the notes: the same as in $\mathbf z$, but with the trajectories drawn towards $\mathbf v_1$; the exact shape needs an explicit $A$ [Notes, p. 10].
- The role of $\mathbf v_2$ is to say which side and which way. Write $\mathbf x = z_1\mathbf v_1 + z_2\mathbf v_2$: the sign of $z_2$ tells on which side of the line $\operatorname{span}\mathbf v_1$ the point lies (the side where $\mathbf v_2$ points if $z_2 > 0$). The motion on that side is along $+\mathbf v_1$ for $z_2 > 0$, along $-\mathbf v_1$ for $z_2 < 0$ (because $\dot z_1 = \lambda z_1 + z_2$).
- Practical rule, independent of the algebra: $\dot{\mathbf x} - \lambda\mathbf x = (A - \lambda I)\mathbf x = N\mathbf x = z_2\mathbf v_1$. At any point off the eigenline, compute $N\mathbf x$: it is a multiple of $\mathbf v_1$ whose sign says how the orbit drifts along the eigendirection there.
- Because $Q$ may reflect ($\det Q < 0$) the S can look mirrored with respect to the $\mathbf z$-plane; the rule above is what fixes the orientation in the $\mathbf x$-plane.
8.9 A complete worked example [Example 30.1, p. 298]
$A = \begin{pmatrix} 5 & -4 \\ 1 & 1\end{pmatrix}$.
- Eigenvalue. $\operatorname{tr}A = 6$, $\det A = 5 + 4 = 9$, so $\lambda^2 - 6\lambda + 9 = (\lambda - 3)^2 = 0$: $\lambda = 3$ with $m_a = 2$ ($\Delta = 36 - 36 = 0$).
- Geometric multiplicity. $N = A - 3I = \begin{pmatrix} 2 & -4 \\ 1 & -2\end{pmatrix} \neq O$ has rank 1 (the rows are proportional), so $m_g = 1$: an improper node, not a star.
- Eigenvector. $N\mathbf v = \mathbf 0$: $2a - 4b = 0$, $a = 2b$. Take $\mathbf v_1 = (2, 1)^T$.
- Generalised eigenvector. $N\mathbf v_2 = \mathbf v_1$: $2a - 4b = 2$ and $a - 2b = 1$ (the same equation). Take $b = 0$, $a = 1$: $\mathbf v_2 = (1, 0)^T$. (General solution $\mathbf v_2 = (1, 0)^T + s(2, 1)^T$.)
- Matrices. $Q = \begin{pmatrix} 2 & 1 \\ 1 & 0\end{pmatrix}$, $\det Q = 0 - 1 = -1$, $Q^{-1} = \dfrac1{-1}\begin{pmatrix} 0 & -1 \\ -1 & 2\end{pmatrix} = \begin{pmatrix} 0 & 1 \\ 1 & -2\end{pmatrix}$. Check: $Q^{-1}Q = \begin{pmatrix} 0 + 1 & 0 \\ 2 - 2 & 1\end{pmatrix} = I$ ✓.
- Jordan form. $A\mathbf v_1 = (10 - 4,\ 2 + 1)^T = (6, 3)^T = 3\mathbf v_1$ ✓, $A\mathbf v_2 = (5, 1)^T = \mathbf v_1 + 3\mathbf v_2 = (2 + 3,\ 1 + 0)^T$ ✓. So $AQ = \begin{pmatrix} 6 & 5 \\ 3 & 1\end{pmatrix}$ and $$Q^{-1}AQ = \begin{pmatrix} 0 & 1 \\ 1 & -2\end{pmatrix}\begin{pmatrix} 6 & 5 \\ 3 & 1\end{pmatrix} = \begin{pmatrix} 3 & 1 \\ 6 - 6 & 5 - 2\end{pmatrix} = \begin{pmatrix} 3 & 1 \\ 0 & 3\end{pmatrix} = J. ✓$$
- Solution in $\mathbf z$ (8.4): $z_1 = (c_1 + c_2t)e^{3t}$, $z_2 = c_2e^{3t}$.
- Solution in $\mathbf x$ (8.5): $\mathbf x(t) = e^{3t}\Big[c_1\begin{pmatrix} 2 \\ 1\end{pmatrix} + c_2\Big(t\begin{pmatrix} 2 \\ 1\end{pmatrix} + \begin{pmatrix} 1 \\ 0\end{pmatrix}\Big)\Big]$, i.e. $$x_1 = e^{3t}\big[2c_1 + c_2(2t + 1)\big], \qquad x_2 = e^{3t}\big[c_1 + c_2t\big].$$ Initial data: $\mathbf x_0 = (2c_1 + c_2,\ c_1)^T$, so $c_1 = x_{20}$ and $c_2 = x_{10} - 2x_{20}$.
- Direct check. $\dot x_1 = e^{3t}\{3[2c_1 + c_2(2t + 1)] + 2c_2\} = e^{3t}[6c_1 + c_2(6t + 5)]$, and $5x_1 - 4x_2 = e^{3t}[10c_1 + c_2(10t + 5) - 4c_1 - 4c_2t] = e^{3t}[6c_1 + c_2(6t + 5)]$ ✓. $\dot x_2 = e^{3t}[3(c_1 + c_2t) + c_2]$ and $x_1 + x_2 = e^{3t}[3c_1 + c_2(3t + 1)]$ ✓.
Explicit trajectories.
| $(c_1, c_2)$ | $\mathbf x_0$ | $\mathbf x(t)$ | Orbit |
|---|---|---|---|
| $(1, 0)$ | $(2, 1)$ | $e^{3t}(2, 1)^T$ | half-line of $\operatorname{span}\mathbf v_1$, outward |
| $(-1, 0)$ | $(-2, -1)$ | $-e^{3t}(2, 1)^T$ | opposite half-line, outward |
| $(0, 1)$ | $(1, 0)$ | $e^{3t}(2t + 1,\ t)^T$ | crosses $x_1 = 0$ at $t = -\tfrac12$, at $e^{-3/2}(0, -\tfrac12)$ |
| $(0, -1)$ | $(-1, 0)$ | $-e^{3t}(2t + 1,\ t)^T$ | point-symmetric to the previous one |
| $(1, 1)$ | $(3, 1)$ | $e^{3t}(2t + 3,\ t + 1)^T$ | same side of $\operatorname{span}\mathbf v_1$ as $\mathbf v_2$ |
Asymptotics. For $c_2 \neq 0$, as $t \to -\infty$: $\mathbf x(t) \approx c_2te^{3t}\mathbf v_1 \to \mathbf 0$ (tangent to $\operatorname{span}\mathbf v_1$, from the side $-\operatorname{sign}(c_2)\mathbf v_1$); as $t \to +\infty$: $|\mathbf x| \to \infty$ with direction $\to +\operatorname{sign}(c_2)\mathbf v_1$. For example, $\mathbf x(t) = e^{3t}(2t + 1, t)^T$ arrives at the origin along the direction $-\mathbf v_1$ and leaves along $+\mathbf v_1$.
Orientation. The side of the eigenline is given by $c_2 = x_1 - 2x_2$ (positive below the line $x_1 = 2x_2$, on the side where $\mathbf v_2$ lies) and the drift by $N\mathbf x = c_2\mathbf v_1$. At $(0, 1)^T$: $N(0,1)^T = (-4, -2)^T = -2\mathbf v_1$, leftwards, in agreement with $\dot x = -4y < 0$ on the positive $y$-axis.
The stable twin. Reversing time turns an improper node into its stable twin: $A' = -A = \begin{pmatrix} -5 & 4 \\ -1 & -1\end{pmatrix}$ has $\lambda' = -3$, the same eigenvector $\mathbf v_1 = (2, 1)^T$, and $N' = A' + 3I = -N$, so $N'\mathbf v_2' = \mathbf v_1$ is solved by $\mathbf v_2' = -\mathbf v_2 = (-1, 0)^T$. Then $Q' = [\mathbf v_1 \mid -\mathbf v_2]$, $J' = \begin{pmatrix} -3 & 1 \\ 0 & -3\end{pmatrix}$ and the orbits are the previous curves with all arrows reversed.
- $\lambda = \tfrac12\operatorname{tr}A$. Compute $N = A - \lambda I$.
- If $N = O$: $A = \lambda I$, a star (stable if $\lambda < 0$, unstable if $\lambda > 0$); $\mathbf x(t) = e^{\lambda t}\mathbf x_0$.
- Otherwise: $\mathbf v_1$ spans $\ker N$; solve $N\mathbf v_2 = \mathbf v_1$; $Q = [\mathbf v_1 \mid \mathbf v_2]$.
- $\mathbf x(t) = e^{\lambda t}\big[c_1\mathbf v_1 + c_2(t\mathbf v_1 + \mathbf v_2)\big]$ with $(c_1, c_2)^T = Q^{-1}\mathbf x_0$; improper node, stable if $\lambda < 0$, unstable if $\lambda > 0$.
- Draw $\operatorname{span}\mathbf v_1$ with arrows (outward if $\lambda > 0$), then the S-shaped orbits tangent to it; fix the side and the orientation with $N\mathbf x$ at one point off the line.
9. Classification of linear phase portraits
Robinson's summary [Ch. 31, p. 301]: compute the eigenvalues of $A$; if they are real, compute the eigenvectors.
| Eigenvalues | Condition | Origin is a | Behaviour |
|---|---|---|---|
| Real, distinct | $\lambda_1 < \lambda_2 < 0$ | stable node | All trajectories approach $\mathbf 0$, tangent to the eigenvector of $\lambda_2$ (smaller $|\lambda|$) |
| $\lambda_1 > \lambda_2 > 0$ | unstable node | All trajectories leave $\mathbf 0$, tangent to the eigenvector of $\lambda_2$ | |
| $\lambda_1 < 0 < \lambda_2$ | saddle | Only trajectories on the stable eigenvector reach $\mathbf 0$ | |
| Complex $\rho \pm i\omega$ | $\rho < 0$ | stable spiral | Spiral in; sense from the sign of $\dot x$ on $x = 0$ |
| $\rho > 0$ | unstable spiral | Spiral out | |
| $\rho = 0$ | centre | Closed periodic orbits | |
| Repeated $\lambda$ | $A = \lambda I$ | star | Straight lines; stable if $\lambda < 0$ |
| $A \neq \lambda I$ | improper node | S-shaped; stable if $\lambda < 0$ |
9.1 The trace–determinant shortcut (complementary)
Since the eigenvalues solve $\lambda^2 - T\lambda + D = 0$ with $T = \operatorname{tr}A = \lambda_1 + \lambda_2$ and $D = \det A = \lambda_1\lambda_2$, the type can be read off $T$ and $D$ without computing the eigenvalues:
- $D < 0$: real eigenvalues of opposite sign — saddle.
- $D > 0$ and $T^2 > 4D$: real, same sign as $T$ — node (stable if $T < 0$).
- $D > 0$ and $T^2 < 4D$: complex with $\rho = T/2$ — spiral (stable if $T < 0$), or centre if $T = 0$.
- $T^2 = 4D$: repeated eigenvalue — star or improper node.
For $A = \begin{pmatrix} 0 & 1 \\ -\omega^2 & -\lambda\end{pmatrix}$ (Section 1), $T = -\lambda$ and $D = \omega^2 > 0$, and the characteristic equation $\mu^2 + \lambda\mu + \omega^2 = 0$ is the auxiliary equation of Topic 3. The damping regimes are exactly the portrait types:
| Damping | Condition | Phase portrait in $(x, \dot x)$ |
|---|---|---|
| none | $\lambda = 0$ ($T = 0$) | centre: energy ellipses $\tfrac12\dot x^2 + \tfrac12\omega^2x^2 = E$ |
| under-damped | $\lambda^2 < 4\omega^2$ | stable spiral, clockwise (on $x = 0$, $\dot x > 0$ means $x$ increases) |
| critical | $\lambda^2 = 4\omega^2$ | stable improper node ($A \neq \lambda I$ always, since the top-left entry is $0$) |
| over-damped | $\lambda^2 > 4\omega^2$ | stable node |
The same classification underlies flight dynamics: linearising the longitudinal equations of an aircraft about trimmed flight gives a matrix whose complex eigenvalue pairs are the short-period and phugoid modes, and their real parts decide whether the aircraft returns to trim after a gust.
Sections 1–6 follow page 9 of the class notes and the lectures up to Oct 8 (Ch. 25–28, Sec. 28.4). Sections 7 (complex eigenvalues, Ch. 29, Oct 13) and 8 (a repeated eigenvalue, Ch. 30, Oct 15) follow page 10. Sec. 28.5 (stable and unstable manifolds) and Ch. 31 (summary and Jordan form) are included here so that the topic is complete.