Chapter 4 Problem Set, Part 1: Linear Systems, Elimination and Real Eigenvalues

Complete, step-by-step solutions to the eleven exercises of the first Chapter 4 sheet: higher-order equations as systems, superposition, Wronskian and general solution of linear systems, elimination with initial conditions, general solutions and phase portraits from eigen-data, decoupling of four matrices with real eigenvalues, the trajectories of an uncoupled system and a zero eigenvalue.

Source Document: sources/cuatrimestre-1/04-advanced-maths/unit-04-systems-of-odes/problemas/probls_ch4-i_2627_v2.pdf
Exercise 1

An n-th Order Equation as a First-Order System and Its Uniqueness

Original statement · probls_ch4-i_2627_v2.pdf
Reduction to a System • Existence and Uniqueness
Statement

By choosing an appropriate collection of new variables $x_1, \dots, x_n$, rewrite the $n$th order differential equation

$$ \frac{d^nx}{dt^n} = f\left(\frac{d^{n-1}x}{dt^{n-1}}, \dots, \frac{dx}{dt}, x, t\right) $$

as a set of $n$ coupled linear first order equations. Find the conditions on the function $f(x_n, \dots, x_1, t)$ for the original differential equation to have a unique solution.

Setup and hypotheses

Unknown: $x(t)$, $n$ times differentiable on an interval $I$. Data: $f(x_n, \dots, x_1, t)$ defined on a region $\Omega \subset \mathbb R^{n+1}$. The initial value problem prescribes the $n$ numbers $x(t_0), \dot x(t_0), \dots, x^{(n-1)}(t_0)$. The tool for uniqueness is the existence and uniqueness theorem for first-order systems $\dot{\mathbf x} = \mathbf F(\mathbf x, t)$ (Robinson Thm 25.2, Topic 4 Sec. 1).

Formulation: the new variables

Introduce one variable for the function and for each derivative below the highest order (Robinson Sec. 25.1):

$$ x_1 = x,\quad x_2 = \frac{dx}{dt},\quad x_3 = \frac{d^2x}{dt^2},\quad \dots,\quad x_n = \frac{d^{n-1}x}{dt^{n-1}}, \qquad\text{i.e. } x_k = x^{(k-1)}\ (k = 1, \dots, n) $$

Method of the class notes [Notes, p. 9]

The important point of the notes: an ODE of order $n \ge 2$ is a system of $n$ first-order ODEs; for $\ddot x = f(\dot x, x, t)$ they set $x_1 = x$, $x_2 = \dot x$, so that $\dot x_1 = x_2$ and $\dot x_2 = f(x_2, x_1, t)$. The variables above are the same choice for order $n$. Uniqueness then follows from the condition the notes state for systems: $\mathbf f$ and its Jacobian $D\mathbf f$ continuous near $(\mathbf x_0, t_0)$. Only the last component $f$ is not trivial, so the condition falls on $f$ and its partial derivatives $\partial f/\partial x_k$.

Derivation

The system

By construction $\dot x_k = x^{(k)} = x_{k+1}$ for $k = 1, \dots, n - 1$, and $\dot x_n = x^{(n)} = f(x_n, \dots, x_1, t)$ by the original equation:

$$ \begin{aligned} \dot x_1 &= x_2 \\ \dot x_2 &= x_3 \\ &\ \,\vdots \\ \dot x_{n-1} &= x_n \\ \dot x_n &= f(x_n, x_{n-1}, \dots, x_1, t) \end{aligned} \qquad\Longleftrightarrow\qquad \dot{\mathbf x} = \mathbf F(\mathbf x, t) = \begin{pmatrix} x_2 \\ x_3 \\ \vdots \\ x_n \\ f(x_n, \dots, x_1, t) \end{pmatrix} $$

The first $n - 1$ equations are linear in the unknowns (they say $\dot x_k = x_{k+1}$); the last one is linear exactly when $f$ is a linear function of $x_1, \dots, x_n$ (as for a linear ODE, where the whole system is linear, $\dot{\mathbf x} = A(t)\mathbf x + \mathbf b(t)$). The first-order system form is valid for any $f$.

Equivalence

(a) If $x$ solves the $n$th-order equation, then $\mathbf x = (x, \dot x, \dots, x^{(n-1)})^T$ solves the system. (b) Conversely, if $(x_1, \dots, x_n)$ solves the system, then $\dot x_1 = x_2$, $\ddot x_1 = \dot x_2 = x_3$, ..., $x_1^{(k)} = x_{k+1}$ for $k \leq n - 1$, and $x_1^{(n)} = \dot x_n = f(x_n, \dots, x_1, t) = f(x_1^{(n-1)}, \dots, \dot x_1, x_1, t)$: the function $x = x_1$ solves the original equation. The initial condition $\mathbf x(t_0) = (x_1^0, \dots, x_n^0)$ is exactly $x(t_0) = x_1^0$, $\dot x(t_0) = x_2^0$, ..., $x^{(n-1)}(t_0) = x_n^0$.

Conditions for a unique solution

Theorem 25.2 (Robinson p. 250) requires $\mathbf F$ and the matrix of partial derivatives $D\mathbf F$ to be continuous on an open set containing the initial point. Compute it row by row: for $F_k = x_{k+1}$ ($k < n$) the partial derivatives are $0$ or $1$, and for $F_n = f$ they are $\partial f/\partial x_j$:

$$ D\mathbf F = \begin{pmatrix} 0 & 1 & 0 & \cdots & 0 \\ 0 & 0 & 1 & \cdots & 0 \\ \vdots & & & \ddots & \vdots \\ 0 & 0 & 0 & \cdots & 1 \\ \dfrac{\partial f}{\partial x_1} & \dfrac{\partial f}{\partial x_2} & \dfrac{\partial f}{\partial x_3} & \cdots & \dfrac{\partial f}{\partial x_n} \end{pmatrix} $$

The first $n - 1$ rows are constant, hence continuous. Therefore all the conditions reduce to the last row:

Condition for uniqueness

$f(x_n, \dots, x_1, t)$ and its partial derivatives $\partial f/\partial x_j$, $j = 1, \dots, n$, must be continuous in an open region containing the initial point $(x(t_0), \dots, x^{(n-1)}(t_0), t_0)$. Then the IVP has exactly one solution on some interval around $t_0$ (local existence and uniqueness), and two solutions with the same $x(t_0), \dots, x^{(n-1)}(t_0)$ coincide wherever both exist.

Interpretation and checks
  • For $n = 2$ the system is $\dot x_1 = x_2$, $\dot x_2 = f(x_2, x_1, t)$ and the condition is Theorem 11.1 of Topic 3 (continuity of $f$, $\partial f/\partial x_1$, $\partial f/\partial x_2$); for $n = 1$ it is Picard's theorem.
  • No condition on $\partial f/\partial t$ is needed: the theorem asks continuity in $t$ but no differentiability in $t$.
  • Dependence on the $n$ derivatives: continuity of the partial derivatives $\partial f/\partial x_j$ gives a local Lipschitz bound in $(x_1, \dots, x_n)$, which is what excludes branching solutions (compare Exercise 2.14 of Chapter 2, where the Lipschitz condition fails).
  • For a linear equation $x^{(n)} + a_{n-1}(t)x^{(n-1)} + \dots + a_0(t)x = g(t)$, $f = g - \sum a_{k}x_{k+1}$ and $\partial f/\partial x_{k+1} = -a_k(t)$: the conditions reduce to continuity of the $a_k$ and $g$. For linear equations the solution in fact exists on the whole interval of continuity of the coefficients (a standard result for linear systems, not in Robinson Thm 25.2, which is local; see the continuation argument in Exercise 4).
  • Check with the mass-spring-damper (Topic 4 Sec. 1.1): $n = 2$, $f = -\lambda x_2 - \omega^2x_1$ has constant partial derivatives $-\omega^2$ and $-\lambda$: unique solutions for all data.
Exercise 2

Superposition for Homogeneous Linear Systems

Original statement · probls_ch4-i_2627_v2.pdf
Superposition • Linear Combination of Solutions
Statement

Show that if a homogeneous linear system

$$ \begin{cases} \dfrac{dx}{dt} = a_1(t)x + b_1(t)y, \\[4pt] \dfrac{dy}{dt} = a_2(t)x + b_2(t)y \end{cases} \tag{1} $$

has two solutions

$$ \begin{cases} x = x_1(t) \\ y = y_1(t) \end{cases} \quad\text{and}\quad \begin{cases} x = x_2(t) \\ y = y_2(t) \end{cases} \tag{2} $$

then any linear combination

$$ \begin{cases} x = c_1x_1(t) + c_2x_2(t) \\ y = c_1y_1(t) + c_2y_2(t) \end{cases} \tag{3} $$

with constants $c_1, c_2 \in \mathbb R$ is a solution as well.

Setup

Hypotheses: $a_i, b_i$ are functions of $t$ (not necessarily constant) defined on an interval $I$, and $(x_1, y_1)$, $(x_2, y_2)$ are differentiable on $I$ with

$$ \dot x_j = a_1x_j + b_1y_j, \qquad \dot y_j = a_2x_j + b_2y_j \qquad (j = 1, 2) \tag{H} $$

Formulation

To show: with $x = c_1x_1 + c_2x_2$ and $y = c_1y_1 + c_2y_2$ the two equations of (1) hold. Tool: differentiation is linear, $\frac{d}{dt}(c_1f + c_2g) = c_1\dot f + c_2\dot g$ for constants $c_1, c_2$. In vector notation: $\mathbf x = c_1\mathbf x_1 + c_2\mathbf x_2$ solves $\dot{\mathbf x} = A(t)\mathbf x$ because the right-hand side is linear in $\mathbf x$ (Robinson Eq. 27.3).

Method of the class notes [Notes, p. 9]

The notes state superposition for $\dot{\mathbf x} = A(t)\mathbf x$: combinations of solutions are solutions. This exercise is that statement written out in components, with $A(t) = \begin{pmatrix} a_1 & b_1 \\ a_2 & b_2 \end{pmatrix}$.

Derivation

First equation. Differentiate $x = c_1x_1 + c_2x_2$ and use (H):

$$ \dot x = c_1\dot x_1 + c_2\dot x_2 = c_1(a_1x_1 + b_1y_1) + c_2(a_1x_2 + b_1y_2) $$

Regroup the terms with $a_1$ and with $b_1$:

$$ \dot x = a_1(c_1x_1 + c_2x_2) + b_1(c_1y_1 + c_2y_2) = a_1\,x + b_1\,y $$

Second equation. In the same way

$$ \dot y = c_1\dot y_1 + c_2\dot y_2 = c_1(a_2x_1 + b_2y_1) + c_2(a_2x_2 + b_2y_2) = a_2(c_1x_1 + c_2x_2) + b_2(c_1y_1 + c_2y_2) = a_2\,x + b_2\,y $$

Both equations of (1) are satisfied by (3). $\blacksquare$

Matrix form

$$ \dot{\mathbf x} = c_1\dot{\mathbf x}_1 + c_2\dot{\mathbf x}_2 = c_1A\mathbf x_1 + c_2A\mathbf x_2 = A(c_1\mathbf x_1 + c_2\mathbf x_2) = A\mathbf x, \qquad A(t) = \begin{pmatrix} a_1 & b_1 \\ a_2 & b_2 \end{pmatrix} $$

Interpretation and checks
  • Linearity and homogeneity are both used: homogeneity guarantees that the right-hand side of (1) vanishes when $x = y = 0$ (a forcing term $f(t)$ would be multiplied by $c_1 + c_2$ instead of by $1$, and the combination would not be a solution unless $c_1 + c_2 = 1$ and so on).
  • Constant coefficients were not needed: only linearity of the derivative and of $(x, y) \mapsto (a_1x + b_1y, a_2x + b_2y)$.
  • The set of solutions is therefore a vector space; Exercises 3 and 4 show that it has dimension $2$.
  • Quick example: $\dot x = y$, $\dot y = -x$ has $(\cos t, -\sin t)$ and $(\sin t, \cos t)$ as solutions, and so $(c_1\cos t + c_2\sin t,\ -c_1\sin t + c_2\cos t)$ does.
Exercise 3

The Wronskian of a Linear System: Liouville's Formula and the Dichotomy

Original statement · probls_ch4-i_2627_v2.pdf
Wronskian • Liouville • Dichotomy
Statement

Show that given a homogeneous linear system (1) with continuous coefficients in $[a, b]$, then the Wronskian $W(t)$ of two solutions (2) satisfies the first order ODE

$$ \frac{dW}{dt} = \left(a_1(t) + b_2(t)\right)W. $$

Deduce that this Wronskian is either identically $0$ or nowhere zero in $[a, b]$.

Setup

System (1) of Exercise 2 with $a_1, b_1, a_2, b_2$ continuous on $[a, b]$. For two solutions $(x_1, y_1)$ and $(x_2, y_2)$ the Wronskian is the determinant of the matrix whose columns are the solutions (Topic 4 Sec. 3):

$$ W(t) = \begin{vmatrix} x_1 & x_2 \\ y_1 & y_2 \end{vmatrix} = x_1y_2 - x_2y_1 $$

Formulation

Differentiate $W$ with the product rule, substitute the equations (H) of Exercise 2 for each derivative, and recognise $W$. The resulting scalar linear ODE is solved by an integrating factor.

Class notes and complement [Notes, p. 9]

The notes state that $\mathbf x_1, \mathbf x_2$ are independent if and only if $W = \det(x_{ij}) \neq 0$ for all $t \in I$; the words for all $t$ are what this exercise proves. The equation $W' = (\operatorname{tr}A)\,W$ and its solution $W(t) = W(t_0)e^{\int_{t_0}^{t}\operatorname{tr}A}$ (Liouville's formula) are a complement (not in the class notes). For the system of a second-order equation $\operatorname{tr}A = -p$, and it becomes the $W' = -pW$ of Topic 3 (Exercise 6).

Derivation

Step 1: the product rule

$$ \dot W = \dot x_1y_2 + x_1\dot y_2 - \dot x_2y_1 - x_2\dot y_1 $$

Step 2: substitute the system

$$ \dot W = (a_1x_1 + b_1y_1)y_2 + x_1(a_2x_2 + b_2y_2) - (a_1x_2 + b_1y_2)y_1 - x_2(a_2x_1 + b_2y_1) $$

Expand and collect by coefficient:

$$ \dot W = a_1(x_1y_2 - x_2y_1) + b_2(x_1y_2 - x_2y_1) + b_1(y_1y_2 - y_2y_1) + a_2(x_1x_2 - x_2x_1) $$

The terms in $b_1$ and $a_2$ vanish identically ($y_1y_2 - y_2y_1 = 0$ and $x_1x_2 - x_2x_1 = 0$), and the remaining brackets are $W$:

$$ \boxed{\dot W = \left(a_1 + b_2\right)W} \qquad (= (\operatorname{tr}A)\,W) $$

Step 3: solve the scalar equation

Let $\tau(t) = \int_{t_0}^{t}\left[a_1(s) + b_2(s)\right]ds$, which exists and is differentiable because $a_1 + b_2$ is continuous, with $\tau'(t) = a_1(t) + b_2(t)$ (fundamental theorem of calculus). The chain rule gives $\frac{d}{dt}e^{-\tau(t)} = -\tau'(t)e^{-\tau(t)}$, so by the product rule $\frac{d}{dt}\left(e^{-\tau}W\right) = e^{-\tau}\dot W - \tau'e^{-\tau}W = e^{-\tau}(\dot W - \tau'W) = e^{-\tau}\left(\dot W - (a_1 + b_2)W\right) = 0$, so $e^{-\tau}W$ is constant on $[a, b]$. Barrow's rule, $[e^{-\tau(s)}W(s)]_{t_0}^{t} = 0$, gives $e^{-\tau(t)}W(t) = e^{-\tau(t_0)}W(t_0) = W(t_0)$ (as $\tau(t_0) = 0$):

$$ W(t) = W(t_0)\exp\left(\int_{t_0}^{t}\left[a_1(s) + b_2(s)\right]ds\right), \qquad t \in [a, b] $$

Step 4: the dichotomy

The exponential is finite and strictly positive, so $W(t) = 0$ exactly when $W(t_0) = 0$. Since $t_0 \in [a, b]$ was arbitrary: either $W(t_0) = 0$ for one (hence every) $t_0$, and then $W \equiv 0$ on $[a, b]$; or $W(t_0) \neq 0$ for one (hence every) $t_0$, and then $W$ does not vanish anywhere on $[a, b]$. There is no third possibility. $\blacksquare$

Interpretation and checks
  • Constant coefficients: $W(t) = W(0)e^{(a_1 + b_2)t} = W(0)e^{(\operatorname{tr}A)t}$ (Topic 4 Sec. 3, Liouville). Check with Example 28.1: eigenvalues $2$ and $-3$ have trace $-1$ and $W \propto e^{2t}e^{-3t} = e^{-t}$.
  • For the system of a second-order equation $\ddot x + p\dot x + qx = 0$, $a_1 = 0$ and $b_2 = -p$ and the result is Abel's identity $\dot W = -pW$ (Exercise 6).
  • Geometric meaning: $|W|$ is the area of the parallelogram spanned by the two solution vectors; $\operatorname{tr}A$ is the divergence of the flow, which rates the growth of areas. If $\operatorname{tr}A = 0$ areas are conserved ($W$ constant).
  • This dichotomy is what makes the Wronskian a test for independence: it is enough to evaluate it at a single instant (Exercise 4).
Exercise 4

A Non-Vanishing Wronskian Gives the General Solution

Original statement · probls_ch4-i_2627_v2.pdf
Fundamental Set • General Solution • Uniqueness
Statement

Show that if two solutions (2) of (1) have a Wronskian that does not vanish in $[a, b]$, then (3) gives the general solution of the system.

Setup and hypotheses

System (1) with coefficients continuous on $[a, b]$ (as in Exercise 3). $(x_1, y_1)$ and $(x_2, y_2)$ are solutions with $W(t) = x_1y_2 - x_2y_1 \neq 0$ for all $t \in [a, b]$. To show: for every solution $(x, y)$ of (1) there are constants $c_1, c_2$ with $x = c_1x_1 + c_2x_2$ and $y = c_1y_1 + c_2y_2$ on $[a, b]$. (By Exercise 2 every combination is a solution, so this shows that (3) describes all of them.)

Formulation

Tools: (a) uniqueness for linear systems with continuous coefficients: two solutions with the same value at one instant $t_0 \in [a, b]$ coincide on all of $[a, b]$. Robinson Thm 25.2 gives this locally (near $t_0$); the extension to $[a, b]$ is a continuation argument: the set $S$ of instants where the two solutions agree contains $t_0$, is closed (the solutions are continuous) and is open in $[a, b]$ (at each point of $S$ the local uniqueness applies, since the coefficients are continuous), and an interval is connected, so $S = [a, b]$; (b) Cramer's rule: a $2\times 2$ linear system with non-zero determinant has exactly one solution. Plan: match the given solution at one instant with a combination, then invoke uniqueness.

Method of the class notes [Notes, p. 9]

The notes: for $n = 2$ a fundamental set is a pair of solutions with $W(\mathbf x_1, \mathbf x_2) \neq 0$, and every solution is a combination of it. This exercise proves that statement. Matching the data at $t_0$ is the $2 \times 2$ system with determinant $W(t_0)$, as in the second-order example of the notes (p. 5), and uniqueness does the rest.

Derivation

Step 1: choose the constants at one instant

Let $(x, y)$ be any solution and fix $t_0 \in [a, b]$. We look for $c_1, c_2$ such that the combination equals $(x, y)$ at $t_0$:

$$ \begin{pmatrix} x_1(t_0) & x_2(t_0) \\ y_1(t_0) & y_2(t_0) \end{pmatrix}\begin{pmatrix} c_1 \\ c_2 \end{pmatrix} = \begin{pmatrix} x(t_0) \\ y(t_0) \end{pmatrix} $$

The determinant of the coefficient matrix is $W(t_0) \neq 0$ by hypothesis, so there is exactly one solution (Cramer):

$$ c_1 = \frac{x(t_0)y_2(t_0) - x_2(t_0)y(t_0)}{W(t_0)}, \qquad c_2 = \frac{x_1(t_0)y(t_0) - x(t_0)y_1(t_0)}{W(t_0)} $$

Step 2: uniqueness

Define $\tilde x = c_1x_1 + c_2x_2$, $\tilde y = c_1y_1 + c_2y_2$. By Exercise 2 $(\tilde x, \tilde y)$ solves (1), and by Step 1 $(\tilde x(t_0), \tilde y(t_0)) = (x(t_0), y(t_0))$. Both $(x, y)$ and $(\tilde x, \tilde y)$ solve the same IVP at $t_0$, so by uniqueness $(x, y) = (\tilde x, \tilde y)$ on $[a, b]$. $\blacksquare$

Where Exercise 3 enters

If the Wronskian were known not to vanish at only some $t_0$, the argument would work at that $t_0$; Exercise 3 shows that a non-zero value at one point already implies $W \neq 0$ everywhere, so one test is enough. Also, two solutions with $W \equiv 0$ cannot form a general solution: then the matrix above is singular at each instant, and not every initial vector can be matched.

Interpretation and checks
  • The solution space is a 2-dimensional vector space with basis $\{(x_1, y_1), (x_2, y_2)\}$ (a fundamental set; $\Phi = [\mathbf x_1\ \mathbf x_2]$ is a fundamental matrix and $\mathbf c = \Phi(t_0)^{-1}\mathbf x(t_0)$, Topic 4 Sec. 3).
  • Example: for $\dot x = x + y$, $\dot y = 4x - 2y$ the solutions $e^{2t}(1, 1)$ and $e^{-3t}(1, -4)$ have $W = e^{2t}e^{-3t}(-4 - 1) = -5e^{-t} \neq 0$, hence form a fundamental set (Topic 4 Example 28.1).
  • Check: if $x_2 = kx_1$ and $y_2 = ky_1$ (parallel solutions) then $W = kx_1y_1 - kx_1y_1 = 0$ and the argument rightly fails.
  • The same proof works for $n \times n$ systems with $\det\Phi(t_0) \neq 0$.
Exercise 5

General Solution of an Inhomogeneous System

Original statement · probls_ch4-i_2627_v2.pdf
Inhomogeneous System • Particular Solution
Statement

Suppose that we have an inhomogeneous system

$$ \begin{cases} \dfrac{dx}{dt} = a_1(t)x + b_1(t)y + f_1(t), \\[4pt] \dfrac{dy}{dt} = a_2(t)x + b_2(t)y + f_2(t) \end{cases} \tag{4} $$

and one particular solution

$$ \begin{cases} x = x_p(t) \\ y = y_p(t). \end{cases} \tag{5} $$

Show that, if $x_1(t)$ and $x_2(t)$ are linearly independent solutions of the homogeneous system, then the general solution of (4) is

$$ \begin{cases} x = c_1x_1(t) + c_2x_2(t) + x_p(t) \\ y = c_1y_1(t) + c_2y_2(t) + y_p(t) \end{cases} $$

Setup and hypotheses

Coefficients and forcings $f_1, f_2$ continuous on $[a, b]$. The homogeneous system is (1) of Exercise 2. $\mathbf x_1 = (x_1, y_1)$ and $\mathbf x_2 = (x_2, y_2)$ are linearly independent solutions of (1) (the statement writes "$x_1(t)$ and $x_2(t)$" for the two solution vectors). Their Wronskian is non-zero on $[a, b]$: by Exercise 3 it is either $\equiv 0$ or never zero, and if $W(t_0) = 0$ there would be $(c_1, c_2) \neq (0, 0)$ with $c_1\mathbf x_1(t_0) + c_2\mathbf x_2(t_0) = \mathbf 0$; the combination $c_1\mathbf x_1 + c_2\mathbf x_2$ solves (1) with zero data at $t_0$, so it is $\equiv \mathbf 0$ by uniqueness (Exercise 4, Step 2), contradicting independence. $\mathbf x_p = (x_p, y_p)$ solves (4).

Formulation

Two claims: (A) every function of the displayed form solves (4); (B) every solution of (4) has that form. The key idea (Robinson Sec. 14.1 for scalar equations, Topic 4 Sec. 3.1) is that the difference of two solutions of (4) solves the homogeneous system.

Method of the class notes [Notes, p. 9]

The last theorem of the notes for linear systems: $\mathbf x = \mathbf x_h + \mathbf x_p$. The proof is the one of the scalar case (Topic 3, notes p. 6): the difference of two solutions of the inhomogeneous system solves the homogeneous one.

Derivation

(A) The family solves (4)

Write $\mathbf x = c_1\mathbf x_1 + c_2\mathbf x_2 + \mathbf x_p$ and $\dot{\mathbf x} = A\mathbf x + \mathbf f$ with $A = \begin{pmatrix} a_1 & b_1 \\ a_2 & b_2 \end{pmatrix}$, $\mathbf f = (f_1, f_2)^T$. Using $\dot{\mathbf x}_j = A\mathbf x_j$ and $\dot{\mathbf x}_p = A\mathbf x_p + \mathbf f$:

$$ \dot{\mathbf x} = c_1A\mathbf x_1 + c_2A\mathbf x_2 + A\mathbf x_p + \mathbf f = A(c_1\mathbf x_1 + c_2\mathbf x_2 + \mathbf x_p) + \mathbf f = A\mathbf x + \mathbf f $$

(B) Every solution has this form

Let $\mathbf x = (x, y)$ be any solution of (4) and put $\mathbf z = \mathbf x - \mathbf x_p$. Then

$$ \dot{\mathbf z} = \dot{\mathbf x} - \dot{\mathbf x}_p = (A\mathbf x + \mathbf f) - (A\mathbf x_p + \mathbf f) = A(\mathbf x - \mathbf x_p) = A\mathbf z $$

so $\mathbf z$ solves the homogeneous system (1). Since $\mathbf x_1, \mathbf x_2$ are independent solutions of (1) (non-zero Wronskian), Exercise 4 gives constants $c_1, c_2$ with $\mathbf z = c_1\mathbf x_1 + c_2\mathbf x_2$. Hence $\mathbf x = c_1\mathbf x_1 + c_2\mathbf x_2 + \mathbf x_p$. $\blacksquare$

Interpretation and checks
  • The structure "general = complementary + particular" is the same as for scalar linear equations (Topic 3 Sec. 5 lemma). The two constants are fixed by an initial condition $\mathbf x(t_0) = \mathbf x_0$: $\mathbf c = \Phi(t_0)^{-1}\left[\mathbf x_0 - \mathbf x_p(t_0)\right]$.
  • Example (Topic 4 Example 26.1): $\dot x = x + y$, $\dot y = 4x - 2y + 4e^{-2t}$ has the particular solution $(-e^{-2t}, 3e^{-2t})$: check $\dot x_p = 2e^{-2t} = x_p + y_p = -e^{-2t} + 3e^{-2t}$ and $\dot y_p = -6e^{-2t} = 4(-e^{-2t}) - 2(3e^{-2t}) + 4e^{-2t} = -6e^{-2t}$.
  • A particular solution can be found by elimination (Exercise 7), undetermined coefficients, or variation of parameters $\mathbf x_p = \Phi\int\Phi^{-1}\mathbf f$ (Topic 4 Sec. 3.1).
  • Different particular solutions differ by a homogeneous solution, which is absorbed in $c_1, c_2$.
Exercise 6

A Second-Order Equation as a System: the Two Wronskians Coincide

Original statement · probls_ch4-i_2627_v2.pdf, p. 2
Second Order to System • Wronskian • Abel
Statement

Suppose that we have a second order linear ODE $x''(t) + p(t)x'(t) + q(t)x(t) = 0$, and two solutions $x_1(t)$ and $x_2(t)$. Write it as a linear system of two first order ODEs. Check that if

$$ \begin{cases} x = x_1(t) \\ y = y_1(t) \end{cases} \quad\text{and}\quad \begin{cases} x = x_2(t) \\ y = y_2(t) \end{cases} $$

are the corresponding solutions of the system, then the Wronskian as defined in this chapter coincides with the definition that we had before for second order ODEs.

Setup

$p, q$ continuous on an interval $I$. Two definitions to compare: the scalar Wronskian $W_2[x_1, x_2] = x_1x_2' - x_2x_1'$ (Topic 3 Sec. 2.1) and the Wronskian of vector solutions $W_{sys} = \det[\mathbf x_1\ \mathbf x_2]$ (Topic 4 Sec. 3).

Formulation: the system

Introduce $y = x'$ (Robinson Sec. 25.1: $x_1 \to x$, $x_2 \to x'$). Then $x' = y$ and, from the ODE, $y' = x'' = -p\,x' - q\,x = -q\,x - p\,y$:

$$ \begin{cases} x' = 0\cdot x + 1\cdot y \\ y' = -q(t)\,x - p(t)\,y \end{cases} \qquad \dot{\mathbf x} = \begin{pmatrix} 0 & 1 \\ -q & -p \end{pmatrix}\mathbf x $$

This has the form (1) with $a_1 = 0$, $b_1 = 1$, $a_2 = -q$, $b_2 = -p$ (continuous). A solution $x_j(t)$ of the ODE gives the solution $(x_j, y_j) = (x_j, x_j')$ of the system, and conversely the first component of a system solution solves the ODE.

Method of the class notes [Notes, p. 9]

This is the important point of the notes with $x_1 = x$ and $x_2 = \dot x$ (here $y$), applied to $x'' + px' + qx = 0$: $A = \begin{pmatrix} 0 & 1 \\ -q & -p \end{pmatrix}$. The Wronskian of the system, $\det\begin{pmatrix} x_1 & x_2 \\ y_1 & y_2 \end{pmatrix} = x_1y_2 - x_2y_1$ with $y_j = x_j'$, is then the Wronskian $x_1x_2' - x_2x_1'$ of the notes for second-order equations (p. 5).

Derivation

The corresponding solutions of the system are $\mathbf x_1 = (x_1, x_1')^T$ and $\mathbf x_2 = (x_2, x_2')^T$, so $y_j = x_j'$. Their system Wronskian is

$$ W_{sys} = \begin{vmatrix} x_1 & x_2 \\ y_1 & y_2 \end{vmatrix} = x_1y_2 - x_2y_1 = x_1x_2' - x_2x_1' = W_2[x_1, x_2] $$

The two definitions coincide identically. $\blacksquare$

Consistency with Exercise 3

Exercise 3 applied to this system ($a_1 + b_2 = 0 + (-p) = -p$) gives $\dot W = -pW$, i.e. $W(t) = W(t_0)\exp\left(-\int_{t_0}^{t}p\right)$, which is Abel's identity of Topic 3 (Robinson Exercise 11.2). Direct check from the scalar definition: $W_2' = x_1'x_2' + x_1x_2'' - x_2'x_1' - x_2x_1'' = x_1x_2'' - x_2x_1'' = x_1(-px_2' - qx_2) - x_2(-px_1' - qx_1) = -p(x_1x_2' - x_2x_1') = -pW_2$, the same equation.

Interpretation and checks
  • Therefore all the theory of systems (non-vanishing Wronskian $\Leftrightarrow$ independence, general solution, dichotomy) contains the scalar results of Topic 3 as the special case $A = \begin{pmatrix} 0 & 1 \\ -q & -p \end{pmatrix}$; the notation $\mathbf x(t_0) = (x(t_0), \dot x(t_0))$ for the initial data is the same pair as in Theorem 11.1.
  • Example: $\ddot x + x = 0$, $x_1 = \cos t$, $x_2 = \sin t$: $W_2 = \cos^2t + \sin^2t = 1$ and $W_{sys} = \begin{vmatrix} \cos t & \sin t \\ -\sin t & \cos t \end{vmatrix} = 1$; $p = 0$ so $W$ is constant, as Abel/Liouville predict.
  • The matrix $\begin{pmatrix} 0 & 1 \\ -q & -p \end{pmatrix}$ has trace $-p$ and determinant $q$; for constants it gives the characteristic polynomial $\mu^2 + p\mu + q$ of Topic 3.
Exercise 7

Seven Linear Systems Solved by Elimination

Original statement · probls_ch4-i_2627_v2.pdf, p. 2
Elimination Method • Seven IVPs • Undetermined Coefficients
Statement

Find the general solutions of the following differential equations by converting them into a single second-order equation. Also find the solution that satisfies the given initial conditions.

  • (i) $\dot x = 4x - y,\quad \dot y = 2x + y + t^2$, $\quad x(0) = 0,\ y(0) = 1$;
  • (ii) $\dot x = x - 4y + \cos 2t,\quad \dot y = x + y$, $\quad x(0) = 1,\ y(0) = 1$;
  • (iii) $\dot x = 2x + 2y,\quad \dot y = 6x + 3y + e^{t}$, $\quad x(0) = 0,\ \dot x(0) = 1$;
  • (iv) $\dot x = 5x - 4y + e^{3t},\quad \dot y = x + y$, $\quad x(0) = 1,\ y(0) = -1$;
  • (v) $\dot x = 2x + 5y,\quad \dot y = -2x + \cos 3t$, $\quad x(0) = 2,\ y(0) = -1$;
  • (vi) $\dot x = x + y + e^{-t},\quad \dot y = 4x - 2y + e^{2t}$, $\quad x(0) = 1,\ y(0) = -1$;
  • (vii) $\dot x = 8x + 14y,\quad \dot y = 7x + y$, $\quad x(0) = y(0) = 1$.
Setup and classification

Each system is linear with constant coefficients, of the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1), homogeneous when $f = g = 0$. In every item $b \neq 0$, so the first equation can be solved for $y$. Hypotheses: coefficients and forcings continuous on $\mathbb R$, so the IVP has a unique solution, which exists on all of $\mathbb R$ (linear system with continuous coefficients; Thm 25.2 gives local existence and uniqueness and the continuation argument of Exercise 4 extends it). Item (iii) gives $x(0)$ and $\dot x(0)$; the value of $y(0)$ then follows from the first equation.

Method: elimination (Topic 4 Sec. 2.1)

Solve the first equation for $y = (\dot x - ax - f)/b$ (26.2), differentiate, and substitute in the second equation. The result is one second-order equation for $x$ (26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Its coefficients are the trace and determinant of the matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}$, so its characteristic polynomial is the one of the eigenvalue method. It is solved by the methods of Topic 3 (complementary function plus undetermined coefficients). Then $y$ is recovered from (26.2) and the constants from the data.

Method of the class notes [Notes, p. 9]

Method 1 of the notes: for $\dot x = ax + by$, $\dot y = cx + dy$ with $b \neq 0$ (true in all seven items), $y = (\dot x - ax)/b$ and $\dot y = (\ddot x - a\dot x)/b$; substituting and multiplying by $b$ gives $\ddot x - (a + d)\dot x + (ad - bc)x = 0$, the homogeneous part of the equation above. The forcing terms are a complement (not in the class notes); the second-order equation is then solved as in Topic 3, $x = x_H + x_N$ with the educated guess.

Discrepancy with the notes: they box $\ddot x - (a + b)\dot x + (ad - bc)x = 0$, but the steps give $a + d$. Check with $x' = 2x - y$, $y' = -x + 3y$: elimination gives $\ddot x - 5\dot x + 5x = 0$, and $a + d = 5$ while $a + b = 1$.

Derivation of the seven solutions
(i) $\dot x = 4x - y,\quad \dot y = 2x + y + t^2$, $\ x(0) = 0,\ y(0) = 1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 4,\quad b = -1,\quad c = 2,\quad d = 1,\quad f(t) = 0,\quad g(t) = t^{2} $$

Elimination of $y$

Since $b = -1 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{-1}\left(\dot x - (4)x - (0)\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{-1}\left(\ddot x - (4)\dot x - (0)\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 5$, $ad - bc = (4)(1) - (-1)(2) = 6$ and $F(t) := \dot f + bg - df = (0) + (-1)(t^{2}) - (1)(0) = - t^{2}$, so

$$ \ddot{x} - 5\dot{x} + 6x = - t^{2} $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (4)(0) + (-1)(1) + (0) = -1$.

The homogeneous equation $\ddot{x} - 5\dot{x} + 6x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 5 k + 6 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-5)^2 - 4(1)(6) = 1$.

Factorising, $P(k) = \left(k - 3\right) \left(k - 2\right)$, so the roots are

$$ k = 2, \qquad k = 3 $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{2 t}$, $e^{3 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{2 t} + c_{2} e^{3 t} $$

The forcing $- t^{2}$ is a polynomial of degree 2; a polynomial is $e^{0t}$ times a polynomial, so the exponent to test is $s = 0$: $P(0) = 6$.

Since $P(s) \neq 0$, no term of the standard guess solves the homogeneous equation, so the standard guess is used as it stands.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} + K_{2} t + K_{3} t^{2} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= K_{2} + 2 K_{3} t \\ x_p''(t) &= 2 K_{3} \end{aligned} $$

Substitute into $L[x_p] = x_p'' - 5x_p' + 6x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = 6 K_{3}\,t^{2} + \left(6 K_{2} - 10 K_{3}\right)\,t + \left(6 K_{1} - 5 K_{2} + 2 K_{3}\right) $$

This must equal the forcing $F(t) = - t^{2}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} t^{2}: &\quad 6 K_{3} = -1 \\ t: &\quad 6 K_{2} - 10 K_{3} = 0 \\ 1: &\quad 6 K_{1} - 5 K_{2} + 2 K_{3} = 0 \end{aligned} $$

The linear system is

$$ \begin{aligned} 6 K_{3} &= -1 \quad (1) \\ 6 K_{2} - 10 K_{3} &= 0 \quad (2) \\ 6 K_{1} - 5 K_{2} + 2 K_{3} &= 0 \quad (3) \end{aligned} $$

$6 K_{3} = -1$ contains one unknown only, so $K_{3} = - \frac{1}{6}$.

Substitute in $6 K_{2} - 10 K_{3} = 0$: $6 K_{2} + \frac{5}{3} = 0$.

Substitute in $6 K_{1} - 5 K_{2} + 2 K_{3} = 0$: $6 K_{1} - 5 K_{2} - \frac{1}{3} = 0$.

$6 K_{2} + \frac{5}{3} = 0$ contains one unknown only, so $K_{2} = - \frac{5}{18}$.

Substitute in $6 K_{1} - 5 K_{2} - \frac{1}{3} = 0$: $6 K_{1} + \frac{19}{18} = 0$.

$6 K_{1} + \frac{19}{18} = 0$ contains one unknown only, so $K_{1} = - \frac{19}{108}$.

Hence $K_{1} = - \frac{19}{108}$, $K_{2} = - \frac{5}{18}$, $K_{3} = - \frac{1}{6}$.

Therefore

$$ x_p(t) = - \frac{t^{2}}{6} - \frac{5 t}{18} - \frac{19}{108} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{2 t} + c_{2} e^{3 t} - \frac{t^{2}}{6} - \frac{5 t}{18} - \frac{19}{108} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= 2 c_{1} e^{2 t} + 3 c_{2} e^{3 t} - \frac{t}{3} - \frac{5}{18} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} + c_{2} - \frac{19}{108} = 0 \\ x'(0) &= 2 c_{1} + 3 c_{2} - \frac{5}{18} = -1 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} + c_{2} - \frac{19}{108} &= 0 \quad (1) \\ 2 c_{1} + 3 c_{2} - \frac{5}{18} &= -1 \quad (2) \end{aligned} $$

Solve $c_{1} + c_{2} - \frac{19}{108} = 0$ for $c_{1}$: $c_{1} = \frac{19}{108} - c_{2}$.

Substitute in $2 c_{1} + 3 c_{2} - \frac{5}{18} = -1$: $c_{2} + \frac{2}{27} = -1$.

$c_{2} + \frac{2}{27} = -1$ contains one unknown only, so $c_{2} = - \frac{29}{27}$.

Back-substitution: $c_{1} = \frac{19}{108} - c_{2} = \frac{5}{4}$.

Hence $c_{1} = \frac{5}{4}$, $c_{2} = - \frac{29}{27}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = - \frac{t^{2}}{6} - \frac{5 t}{18} - \frac{29 e^{3 t}}{27} + \frac{5 e^{2 t}}{4} - \frac{19}{108}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = - \frac{t}{3} - \frac{29 e^{3 t}}{9} + \frac{5 e^{2 t}}{2} - \frac{5}{18} $$

$$ y(t) = \frac{1}{-1}\left(\dot x - (4)x - (0)\right) = - \frac{2 t^{2}}{3} - \frac{7 t}{9} - \frac{29 e^{3 t}}{27} + \frac{5 e^{2 t}}{2} - \frac{23}{54} $$

Check in the original system

$$ \begin{aligned} \dot x &= - \frac{t}{3} - \frac{29 e^{3 t}}{9} + \frac{5 e^{2 t}}{2} - \frac{5}{18} \\ ax + by + f &= - \frac{t}{3} - \frac{29 e^{3 t}}{9} + \frac{5 e^{2 t}}{2} - \frac{5}{18} \\[4pt] \dot y &= - \frac{4 t}{3} - \frac{29 e^{3 t}}{9} + 5 e^{2 t} - \frac{7}{9} \\ cx + dy + g &= - \frac{4 t}{3} - \frac{29 e^{3 t}}{9} + 5 e^{2 t} - \frac{7}{9} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 0$, $y(0) = 1$ reproduce the initial data.

(ii) $\dot x = x - 4y + \cos 2t,\quad \dot y = x + y$, $\ x(0) = 1,\ y(0) = 1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 1,\quad b = -4,\quad c = 1,\quad d = 1,\quad f(t) = \cos{2 t},\quad g(t) = 0 $$

Elimination of $y$

Since $b = -4 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{-4}\left(\dot x - (1)x - (\cos{2 t})\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{-4}\left(\ddot x - (1)\dot x - (- 2 \sin{2 t})\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 2$, $ad - bc = (1)(1) - (-4)(1) = 5$ and $F(t) := \dot f + bg - df = (- 2 \sin{2 t}) + (-4)(0) - (1)(\cos{2 t}) = - 2 \sin{2 t} - \cos{2 t}$, so

$$ \ddot{x} - 2\dot{x} + 5x = - 2 \sin{2 t} - \cos{2 t} $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (1)(1) + (-4)(1) + (1) = -2$.

The homogeneous equation $\ddot{x} - 2\dot{x} + 5x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 2 k + 5 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-2)^2 - 4(1)(5) = -16$.

Factorising, $P(k) = \left(k - 1 - 2 i\right) \left(k - 1 + 2 i\right)$, so the roots are

$$ k = 1 - 2 i, \qquad k = 1 + 2 i $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{t} \cos{2 t}$, $e^{t} \sin{2 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{t} \cos{2 t} + c_{2} e^{t} \sin{2 t} $$

The forcing $- 2 \sin{2 t} - \cos{2 t}$ is a sinusoidal term: it is the real or imaginary part of $e^{(2 i)t}$, so we test $s = 2 i$ in the characteristic polynomial: $P(2 i) = 1 - 4 i$.

Since $P(s) \neq 0$, no term of the standard guess solves the homogeneous equation, so the standard guess is used as it stands.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} \cos{2 t} + K_{2} \sin{2 t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= - 2 K_{1} \sin{2 t} + 2 K_{2} \cos{2 t} \\ x_p''(t) &= - 4 K_{1} \cos{2 t} - 4 K_{2} \sin{2 t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' - 2x_p' + 5x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = \left(K_{1} - 4 K_{2}\right)\,\cos 2t + \left(4 K_{1} + K_{2}\right)\,\sin 2t $$

This must equal the forcing $F(t) = - 2 \sin{2 t} - \cos{2 t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} \cos 2t: &\quad K_{1} - 4 K_{2} = -1 \\ \sin 2t: &\quad 4 K_{1} + K_{2} = -2 \end{aligned} $$

The linear system is

$$ \begin{aligned} K_{1} - 4 K_{2} &= -1 \quad (1) \\ 4 K_{1} + K_{2} &= -2 \quad (2) \end{aligned} $$

Solve $K_{1} - 4 K_{2} = -1$ for $K_{1}$: $K_{1} = 4 K_{2} - 1$.

Substitute in $4 K_{1} + K_{2} = -2$: $17 K_{2} - 4 = -2$.

$17 K_{2} - 4 = -2$ contains one unknown only, so $K_{2} = \frac{2}{17}$.

Back-substitution: $K_{1} = 4 K_{2} - 1 = - \frac{9}{17}$.

Hence $K_{1} = - \frac{9}{17}$, $K_{2} = \frac{2}{17}$.

Therefore

$$ x_p(t) = \frac{2 \sin{2 t}}{17} - \frac{9 \cos{2 t}}{17} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{t} \cos{2 t} + c_{2} e^{t} \sin{2 t} + \frac{2 \sin{2 t}}{17} - \frac{9 \cos{2 t}}{17} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= - 2 c_{1} e^{t} \sin{2 t} + c_{1} e^{t} \cos{2 t} + c_{2} e^{t} \sin{2 t} + 2 c_{2} e^{t} \cos{2 t} + \frac{18 \sin{2 t}}{17} + \frac{4 \cos{2 t}}{17} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} - \frac{9}{17} = 1 \\ x'(0) &= c_{1} + 2 c_{2} + \frac{4}{17} = -2 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} - \frac{9}{17} &= 1 \quad (1) \\ c_{1} + 2 c_{2} + \frac{4}{17} &= -2 \quad (2) \end{aligned} $$

$c_{1} - \frac{9}{17} = 1$ contains one unknown only, so $c_{1} = \frac{26}{17}$.

Substitute in $c_{1} + 2 c_{2} + \frac{4}{17} = -2$: $2 c_{2} + \frac{30}{17} = -2$.

$2 c_{2} + \frac{30}{17} = -2$ contains one unknown only, so $c_{2} = - \frac{32}{17}$.

Hence $c_{1} = \frac{26}{17}$, $c_{2} = - \frac{32}{17}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = - \frac{32 e^{t} \sin{2 t}}{17} + \frac{26 e^{t} \cos{2 t}}{17} + \frac{2 \sin{2 t}}{17} - \frac{9 \cos{2 t}}{17}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = - \frac{84 e^{t} \sin{2 t}}{17} - \frac{38 e^{t} \cos{2 t}}{17} + \frac{18 \sin{2 t}}{17} + \frac{4 \cos{2 t}}{17} $$

$$ y(t) = \frac{1}{-4}\left(\dot x - (1)x - (\cos{2 t})\right) = \frac{13 e^{t} \sin{2 t}}{17} + \frac{16 e^{t} \cos{2 t}}{17} - \frac{4 \sin{2 t}}{17} + \frac{\cos{2 t}}{17} $$

Check in the original system

$$ \begin{aligned} \dot x &= - \frac{84 e^{t} \sin{2 t}}{17} - \frac{38 e^{t} \cos{2 t}}{17} + \frac{18 \sin{2 t}}{17} + \frac{4 \cos{2 t}}{17} \\ ax + by + f &= - \frac{84 e^{t} \sin{2 t}}{17} - \frac{38 e^{t} \cos{2 t}}{17} + \frac{18 \sin{2 t}}{17} + \frac{4 \cos{2 t}}{17} \\[4pt] \dot y &= - \frac{19 e^{t} \sin{2 t}}{17} + \frac{42 e^{t} \cos{2 t}}{17} - \frac{2 \sin{2 t}}{17} - \frac{8 \cos{2 t}}{17} \\ cx + dy + g &= - \frac{19 e^{t} \sin{2 t}}{17} + \frac{42 e^{t} \cos{2 t}}{17} - \frac{2 \sin{2 t}}{17} - \frac{8 \cos{2 t}}{17} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 1$, $y(0) = 1$ reproduce the initial data.

(iii) $\dot x = 2x + 2y,\quad \dot y = 6x + 3y + e^{t}$, $\ x(0) = 0,\ \dot x(0) = 1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 2,\quad b = 2,\quad c = 6,\quad d = 3,\quad f(t) = 0,\quad g(t) = e^{t} $$

Elimination of $y$

Since $b = 2 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{2}\left(\dot x - (2)x - (0)\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{2}\left(\ddot x - (2)\dot x - (0)\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 5$, $ad - bc = (2)(3) - (2)(6) = -6$ and $F(t) := \dot f + bg - df = (0) + (2)(e^{t}) - (3)(0) = 2 e^{t}$, so

$$ \ddot{x} - 5\dot{x} - 6x = 2 e^{t} $$

Initial data. Here $x(0)$ and $\dot x(0)$ are given; the value of $y(0)$ follows from the first equation at $t = 0$: $y(0) = \dfrac{\dot x(0) - a\,x(0) - f(0)}{b} = \dfrac{1 - (2)(0) - (0)}{2} = \frac{1}{2}$.

The homogeneous equation $\ddot{x} - 5\dot{x} - 6x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 5 k - 6 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-5)^2 - 4(1)(-6) = 49$.

Factorising, $P(k) = \left(k - 6\right) \left(k + 1\right)$, so the roots are

$$ k = -1, \qquad k = 6 $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{-t}$, $e^{6 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{-t} + c_{2} e^{6 t} $$

The forcing $2 e^{t}$ is an exponential with exponent $s = 1$; we test it in the characteristic polynomial: $P(1) = -10$.

Since $P(s) \neq 0$, no term of the standard guess solves the homogeneous equation, so the standard guess is used as it stands.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} e^{t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= K_{1} e^{t} \\ x_p''(t) &= K_{1} e^{t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' - 5x_p' - 6x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = - 10 K_{1}\,e^{t} $$

This must equal the forcing $F(t) = 2 e^{t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} e^{t}: &\quad - 10 K_{1} = 2 \end{aligned} $$

The linear system is

$$ \begin{aligned} - 10 K_{1} &= 2 \quad (1) \end{aligned} $$

$- 10 K_{1} = 2$ contains one unknown only, so $K_{1} = - \frac{1}{5}$.

Hence $K_{1} = - \frac{1}{5}$.

Therefore

$$ x_p(t) = - \frac{e^{t}}{5} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{-t} + c_{2} e^{6 t} - \frac{e^{t}}{5} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= - c_{1} e^{-t} + 6 c_{2} e^{6 t} - \frac{e^{t}}{5} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} + c_{2} - \frac{1}{5} = 0 \\ x'(0) &= - c_{1} + 6 c_{2} - \frac{1}{5} = 1 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} + c_{2} - \frac{1}{5} &= 0 \quad (1) \\ - c_{1} + 6 c_{2} - \frac{1}{5} &= 1 \quad (2) \end{aligned} $$

Solve $c_{1} + c_{2} - \frac{1}{5} = 0$ for $c_{1}$: $c_{1} = \frac{1}{5} - c_{2}$.

Substitute in $- c_{1} + 6 c_{2} - \frac{1}{5} = 1$: $7 c_{2} - \frac{2}{5} = 1$.

$7 c_{2} - \frac{2}{5} = 1$ contains one unknown only, so $c_{2} = \frac{1}{5}$.

Back-substitution: $c_{1} = \frac{1}{5} - c_{2} = 0$.

Hence $c_{1} = 0$, $c_{2} = \frac{1}{5}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = \frac{e^{6 t}}{5} - \frac{e^{t}}{5}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = \frac{6 e^{6 t}}{5} - \frac{e^{t}}{5} $$

$$ y(t) = \frac{1}{2}\left(\dot x - (2)x - (0)\right) = \frac{2 e^{6 t}}{5} + \frac{e^{t}}{10} $$

Check in the original system

$$ \begin{aligned} \dot x &= \frac{6 e^{6 t}}{5} - \frac{e^{t}}{5} \\ ax + by + f &= \frac{6 e^{6 t}}{5} - \frac{e^{t}}{5} \\[4pt] \dot y &= \frac{12 e^{6 t}}{5} + \frac{e^{t}}{10} \\ cx + dy + g &= \frac{12 e^{6 t}}{5} + \frac{e^{t}}{10} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 0$, $y(0) = \frac{1}{2}$ reproduce the initial data.

(iv) $\dot x = 5x - 4y + e^{3t},\quad \dot y = x + y$, $\ x(0) = 1,\ y(0) = -1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 5,\quad b = -4,\quad c = 1,\quad d = 1,\quad f(t) = e^{3 t},\quad g(t) = 0 $$

Elimination of $y$

Since $b = -4 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{-4}\left(\dot x - (5)x - (e^{3 t})\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{-4}\left(\ddot x - (5)\dot x - (3 e^{3 t})\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 6$, $ad - bc = (5)(1) - (-4)(1) = 9$ and $F(t) := \dot f + bg - df = (3 e^{3 t}) + (-4)(0) - (1)(e^{3 t}) = 2 e^{3 t}$, so

$$ \ddot{x} - 6\dot{x} + 9x = 2 e^{3 t} $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (5)(1) + (-4)(-1) + (1) = 10$.

The homogeneous equation $\ddot{x} - 6\dot{x} + 9x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 6 k + 9 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-6)^2 - 4(1)(9) = 0$.

Factorising, $P(k) = \left(k - 3\right)^{2}$, so the roots are

$$ k = 3 \ \text{(multiplicity 2)} $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{3 t}$, $t e^{3 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{3 t} + c_{2} t e^{3 t} $$

The forcing $2 e^{3 t}$ is an exponential with exponent $s = 3$; we test it in the characteristic polynomial: $P(3) = 0$.

Since $P(s) = 0$ and $s$ is a root of multiplicity $2$, the standard guess would contain terms of the complementary function. By the rule of thumb (Robinson Sec. 14.5, p. 139) the guess is multiplied by $t^{2}$.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} t^{2} e^{3 t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= 3 K_{1} t^{2} e^{3 t} + 2 K_{1} t e^{3 t} \\ x_p''(t) &= 9 K_{1} t^{2} e^{3 t} + 12 K_{1} t e^{3 t} + 2 K_{1} e^{3 t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' - 6x_p' + 9x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = 2 K_{1}\,e^{3 t} $$

This must equal the forcing $F(t) = 2 e^{3 t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} e^{3 t}: &\quad 2 K_{1} = 2 \end{aligned} $$

The linear system is

$$ \begin{aligned} 2 K_{1} &= 2 \quad (1) \end{aligned} $$

$2 K_{1} = 2$ contains one unknown only, so $K_{1} = 1$.

Hence $K_{1} = 1$.

Therefore

$$ x_p(t) = t^{2} e^{3 t} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{3 t} + c_{2} t e^{3 t} + t^{2} e^{3 t} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= 3 c_{1} e^{3 t} + 3 c_{2} t e^{3 t} + c_{2} e^{3 t} + 3 t^{2} e^{3 t} + 2 t e^{3 t} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} = 1 \\ x'(0) &= 3 c_{1} + c_{2} = 10 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} &= 1 \quad (1) \\ 3 c_{1} + c_{2} &= 10 \quad (2) \end{aligned} $$

$c_{1} = 1$ contains one unknown only, so $c_{1} = 1$.

Substitute in $3 c_{1} + c_{2} = 10$: $c_{2} + 3 = 10$.

$c_{2} + 3 = 10$ contains one unknown only, so $c_{2} = 7$.

Hence $c_{1} = 1$, $c_{2} = 7$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = t^{2} e^{3 t} + 7 t e^{3 t} + e^{3 t}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = 3 t^{2} e^{3 t} + 23 t e^{3 t} + 10 e^{3 t} $$

$$ y(t) = \frac{1}{-4}\left(\dot x - (5)x - (e^{3 t})\right) = \frac{t^{2} e^{3 t}}{2} + 3 t e^{3 t} - e^{3 t} $$

Check in the original system

$$ \begin{aligned} \dot x &= 3 t^{2} e^{3 t} + 23 t e^{3 t} + 10 e^{3 t} \\ ax + by + f &= 3 t^{2} e^{3 t} + 23 t e^{3 t} + 10 e^{3 t} \\[4pt] \dot y &= \frac{3 t^{2} e^{3 t}}{2} + 10 t e^{3 t} \\ cx + dy + g &= \frac{3 t^{2} e^{3 t}}{2} + 10 t e^{3 t} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 1$, $y(0) = -1$ reproduce the initial data.

(v) $\dot x = 2x + 5y,\quad \dot y = -2x + \cos 3t$, $\ x(0) = 2,\ y(0) = -1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 2,\quad b = 5,\quad c = -2,\quad d = 0,\quad f(t) = 0,\quad g(t) = \cos{3 t} $$

Elimination of $y$

Since $b = 5 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{5}\left(\dot x - (2)x - (0)\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{5}\left(\ddot x - (2)\dot x - (0)\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 2$, $ad - bc = (2)(0) - (5)(-2) = 10$ and $F(t) := \dot f + bg - df = (0) + (5)(\cos{3 t}) - (0)(0) = 5 \cos{3 t}$, so

$$ \ddot{x} - 2\dot{x} + 10x = 5 \cos{3 t} $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (2)(2) + (5)(-1) + (0) = -1$.

The homogeneous equation $\ddot{x} - 2\dot{x} + 10x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 2 k + 10 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-2)^2 - 4(1)(10) = -36$.

Factorising, $P(k) = \left(k - 1 - 3 i\right) \left(k - 1 + 3 i\right)$, so the roots are

$$ k = 1 - 3 i, \qquad k = 1 + 3 i $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{t} \cos{3 t}$, $e^{t} \sin{3 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{t} \cos{3 t} + c_{2} e^{t} \sin{3 t} $$

The forcing $5 \cos{3 t}$ is a sinusoidal term: it is the real or imaginary part of $e^{(3 i)t}$, so we test $s = 3 i$ in the characteristic polynomial: $P(3 i) = 1 - 6 i$.

Since $P(s) \neq 0$, no term of the standard guess solves the homogeneous equation, so the standard guess is used as it stands.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} \cos{3 t} + K_{2} \sin{3 t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= - 3 K_{1} \sin{3 t} + 3 K_{2} \cos{3 t} \\ x_p''(t) &= - 9 K_{1} \cos{3 t} - 9 K_{2} \sin{3 t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' - 2x_p' + 10x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = \left(K_{1} - 6 K_{2}\right)\,\cos 3t + \left(6 K_{1} + K_{2}\right)\,\sin 3t $$

This must equal the forcing $F(t) = 5 \cos{3 t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} \cos 3t: &\quad K_{1} - 6 K_{2} = 5 \\ \sin 3t: &\quad 6 K_{1} + K_{2} = 0 \end{aligned} $$

The linear system is

$$ \begin{aligned} K_{1} - 6 K_{2} &= 5 \quad (1) \\ 6 K_{1} + K_{2} &= 0 \quad (2) \end{aligned} $$

Solve $K_{1} - 6 K_{2} = 5$ for $K_{1}$: $K_{1} = 6 K_{2} + 5$.

Substitute in $6 K_{1} + K_{2} = 0$: $37 K_{2} + 30 = 0$.

$37 K_{2} + 30 = 0$ contains one unknown only, so $K_{2} = - \frac{30}{37}$.

Back-substitution: $K_{1} = 6 K_{2} + 5 = \frac{5}{37}$.

Hence $K_{1} = \frac{5}{37}$, $K_{2} = - \frac{30}{37}$.

Therefore

$$ x_p(t) = - \frac{30 \sin{3 t}}{37} + \frac{5 \cos{3 t}}{37} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{t} \cos{3 t} + c_{2} e^{t} \sin{3 t} - \frac{30 \sin{3 t}}{37} + \frac{5 \cos{3 t}}{37} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= - 3 c_{1} e^{t} \sin{3 t} + c_{1} e^{t} \cos{3 t} + c_{2} e^{t} \sin{3 t} + 3 c_{2} e^{t} \cos{3 t} - \frac{15 \sin{3 t}}{37} - \frac{90 \cos{3 t}}{37} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} + \frac{5}{37} = 2 \\ x'(0) &= c_{1} + 3 c_{2} - \frac{90}{37} = -1 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} + \frac{5}{37} &= 2 \quad (1) \\ c_{1} + 3 c_{2} - \frac{90}{37} &= -1 \quad (2) \end{aligned} $$

$c_{1} + \frac{5}{37} = 2$ contains one unknown only, so $c_{1} = \frac{69}{37}$.

Substitute in $c_{1} + 3 c_{2} - \frac{90}{37} = -1$: $3 c_{2} - \frac{21}{37} = -1$.

$3 c_{2} - \frac{21}{37} = -1$ contains one unknown only, so $c_{2} = - \frac{16}{111}$.

Hence $c_{1} = \frac{69}{37}$, $c_{2} = - \frac{16}{111}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = - \frac{16 e^{t} \sin{3 t}}{111} + \frac{69 e^{t} \cos{3 t}}{37} - \frac{30 \sin{3 t}}{37} + \frac{5 \cos{3 t}}{37}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = - \frac{637 e^{t} \sin{3 t}}{111} + \frac{53 e^{t} \cos{3 t}}{37} - \frac{15 \sin{3 t}}{37} - \frac{90 \cos{3 t}}{37} $$

$$ y(t) = \frac{1}{5}\left(\dot x - (2)x - (0)\right) = - \frac{121 e^{t} \sin{3 t}}{111} - \frac{17 e^{t} \cos{3 t}}{37} + \frac{9 \sin{3 t}}{37} - \frac{20 \cos{3 t}}{37} $$

Check in the original system

$$ \begin{aligned} \dot x &= - \frac{637 e^{t} \sin{3 t}}{111} + \frac{53 e^{t} \cos{3 t}}{37} - \frac{15 \sin{3 t}}{37} - \frac{90 \cos{3 t}}{37} \\ ax + by + f &= - \frac{637 e^{t} \sin{3 t}}{111} + \frac{53 e^{t} \cos{3 t}}{37} - \frac{15 \sin{3 t}}{37} - \frac{90 \cos{3 t}}{37} \\[4pt] \dot y &= \frac{32 e^{t} \sin{3 t}}{111} - \frac{138 e^{t} \cos{3 t}}{37} + \frac{60 \sin{3 t}}{37} + \frac{27 \cos{3 t}}{37} \\ cx + dy + g &= \frac{32 e^{t} \sin{3 t}}{111} - \frac{138 e^{t} \cos{3 t}}{37} + \frac{60 \sin{3 t}}{37} + \frac{27 \cos{3 t}}{37} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 2$, $y(0) = -1$ reproduce the initial data.

(vi) $\dot x = x + y + e^{-t},\quad \dot y = 4x - 2y + e^{2t}$, $\ x(0) = 1,\ y(0) = -1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 1,\quad b = 1,\quad c = 4,\quad d = -2,\quad f(t) = e^{-t},\quad g(t) = e^{2 t} $$

Elimination of $y$

Since $b = 1 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{1}\left(\dot x - (1)x - (e^{-t})\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{1}\left(\ddot x - (1)\dot x - (- e^{-t})\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = -1$, $ad - bc = (1)(-2) - (1)(4) = -6$ and $F(t) := \dot f + bg - df = (- e^{-t}) + (1)(e^{2 t}) - (-2)(e^{-t}) = e^{2 t} + e^{-t}$, so

$$ \ddot{x} + \dot{x} - 6x = e^{2 t} + e^{-t} $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (1)(1) + (1)(-1) + (1) = 1$.

The homogeneous equation $\ddot{x} + \dot{x} - 6x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} + k - 6 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (1)^2 - 4(1)(-6) = 25$.

Factorising, $P(k) = \left(k - 2\right) \left(k + 3\right)$, so the roots are

$$ k = -3, \qquad k = 2 $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{-3 t}$, $e^{2 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{-3 t} + c_{2} e^{2 t} $$

The forcing is a sum of terms. By linearity (Robinson Eq. 14.7) a particular integral is the sum of particular integrals of the separate terms, each found in turn.

Particular integral for the term $e^{-t}$

The forcing $e^{-t}$ is an exponential with exponent $s = -1$; we test it in the characteristic polynomial: $P(-1) = -6$.

Since $P(s) \neq 0$, no term of the standard guess solves the homogeneous equation, so the standard guess is used as it stands.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} e^{-t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= - K_{1} e^{-t} \\ x_p''(t) &= K_{1} e^{-t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' + x_p' - 6x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = - 6 K_{1}\,e^{-t} $$

This must equal the forcing $F(t) = e^{-t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} e^{-t}: &\quad - 6 K_{1} = 1 \end{aligned} $$

The linear system is

$$ \begin{aligned} - 6 K_{1} &= 1 \quad (1) \end{aligned} $$

$- 6 K_{1} = 1$ contains one unknown only, so $K_{1} = - \frac{1}{6}$.

Hence $K_{1} = - \frac{1}{6}$.

Therefore

$$ x_p(t) = - \frac{e^{-t}}{6} $$

Particular integral for the term $e^{2 t}$

The forcing $e^{2 t}$ is an exponential with exponent $s = 2$; we test it in the characteristic polynomial: $P(2) = 0$.

Since $P(s) = 0$ and $s$ is a root of multiplicity $1$, the standard guess would contain terms of the complementary function. By the rule of thumb (Robinson Sec. 14.5, p. 139) the guess is multiplied by $t^{1}$.

Guess for the particular integral, with undetermined coefficients:

$$ x_p(t) = K_{1} t e^{2 t} $$

Differentiate term by term (product rule on $t^j e^{\sigma t}$, chain rule on $\cos\omega t$ and $\sin\omega t$):

$$ \begin{aligned} x_p'(t) &= 2 K_{1} t e^{2 t} + K_{1} e^{2 t} \\ x_p''(t) &= 4 K_{1} t e^{2 t} + 4 K_{1} e^{2 t} \end{aligned} $$

Substitute into $L[x_p] = x_p'' + x_p' - 6x_p$ and collect the terms that multiply the same function:

$$ L[x_p] = 5 K_{1}\,e^{2 t} $$

This must equal the forcing $F(t) = e^{2 t}$ for every $t$. The functions on the left are linearly independent, so the coefficient of each must match the coefficient of the same function in $F$:

$$ \begin{aligned} e^{2 t}: &\quad 5 K_{1} = 1 \end{aligned} $$

The linear system is

$$ \begin{aligned} 5 K_{1} &= 1 \quad (1) \end{aligned} $$

$5 K_{1} = 1$ contains one unknown only, so $K_{1} = \frac{1}{5}$.

Hence $K_{1} = \frac{1}{5}$.

Therefore

$$ x_p(t) = \frac{t e^{2 t}}{5} $$

Adding the particular integrals,

$$ x_p(t) = \frac{t e^{2 t}}{5} - \frac{e^{-t}}{6} $$

The general solution is the complementary function plus the particular integral:

$$ x(t) = x_h(t) + x_p(t) = c_{1} e^{-3 t} + c_{2} e^{2 t} + \frac{t e^{2 t}}{5} - \frac{e^{-t}}{6} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= - 3 c_{1} e^{-3 t} + 2 c_{2} e^{2 t} + \frac{2 t e^{2 t}}{5} + \frac{e^{2 t}}{5} + \frac{e^{-t}}{6} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} + c_{2} - \frac{1}{6} = 1 \\ x'(0) &= - 3 c_{1} + 2 c_{2} + \frac{11}{30} = 1 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} + c_{2} - \frac{1}{6} &= 1 \quad (1) \\ - 3 c_{1} + 2 c_{2} + \frac{11}{30} &= 1 \quad (2) \end{aligned} $$

Solve $c_{1} + c_{2} - \frac{1}{6} = 1$ for $c_{1}$: $c_{1} = \frac{7}{6} - c_{2}$.

Substitute in $- 3 c_{1} + 2 c_{2} + \frac{11}{30} = 1$: $5 c_{2} - \frac{47}{15} = 1$.

$5 c_{2} - \frac{47}{15} = 1$ contains one unknown only, so $c_{2} = \frac{62}{75}$.

Back-substitution: $c_{1} = \frac{7}{6} - c_{2} = \frac{17}{50}$.

Hence $c_{1} = \frac{17}{50}$, $c_{2} = \frac{62}{75}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = \frac{t e^{2 t}}{5} + \frac{62 e^{2 t}}{75} - \frac{e^{-t}}{6} + \frac{17 e^{-3 t}}{50}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = \frac{2 t e^{2 t}}{5} + \frac{139 e^{2 t}}{75} + \frac{e^{-t}}{6} - \frac{51 e^{-3 t}}{50} $$

$$ y(t) = \frac{1}{1}\left(\dot x - (1)x - (e^{-t})\right) = \frac{t e^{2 t}}{5} + \frac{77 e^{2 t}}{75} - \frac{2 e^{-t}}{3} - \frac{34 e^{-3 t}}{25} $$

Check in the original system

$$ \begin{aligned} \dot x &= \frac{2 t e^{2 t}}{5} + \frac{139 e^{2 t}}{75} + \frac{e^{-t}}{6} - \frac{51 e^{-3 t}}{50} \\ ax + by + f &= \frac{2 t e^{2 t}}{5} + \frac{139 e^{2 t}}{75} + \frac{e^{-t}}{6} - \frac{51 e^{-3 t}}{50} \\[4pt] \dot y &= \frac{2 t e^{2 t}}{5} + \frac{169 e^{2 t}}{75} + \frac{2 e^{-t}}{3} + \frac{102 e^{-3 t}}{25} \\ cx + dy + g &= \frac{2 t e^{2 t}}{5} + \frac{169 e^{2 t}}{75} + \frac{2 e^{-t}}{3} + \frac{102 e^{-3 t}}{25} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 1$, $y(0) = -1$ reproduce the initial data.

(vii) $\dot x = 8x + 14y,\quad \dot y = 7x + y$, $\ x(0) = y(0) = 1$

The system is linear with constant coefficients, in the form $\dot x = ax + by + f(t)$, $\dot y = cx + dy + g(t)$ (Robinson Eq. 26.1) with

$$ a = 8,\quad b = 14,\quad c = 7,\quad d = 1,\quad f(t) = 0,\quad g(t) = 0 $$

Elimination of $y$

Since $b = 14 \neq 0$, the first equation can be solved for $y$ (Eq. 26.2):

$$ y = \frac{\dot x - ax - f}{b} = \frac{1}{14}\left(\dot x - (8)x - (0)\right) $$

Differentiate with respect to $t$ (everything is differentiable):

$$ \dot y = \frac{\ddot x - a\dot x - \dot f}{b} = \frac{1}{14}\left(\ddot x - (8)\dot x - (0)\right) $$

Substitute both expressions into the second equation, $\dot y = cx + dy + g$, and multiply by $b$:

$$ \ddot x - a\dot x - \dot f = bc\,x + d\,(\dot x - ax - f) + bg $$

Collect the terms in $\ddot x$, $\dot x$ and $x$ (Eq. 26.3):

$$ \ddot x - (a + d)\dot x + (ad - bc)x = \dot f + bg - df $$

Numerically, $a + d = 9$, $ad - bc = (8)(1) - (14)(7) = -90$ and $F(t) := \dot f + bg - df = (0) + (14)(0) - (1)(0) = 0$, so

$$ \ddot{x} - 9\dot{x} - 90x = 0 $$

Initial data. The condition on $y$ is converted into one on $\dot x$ with the first equation at $t = 0$: $\dot x(0) = a\,x(0) + b\,y(0) + f(0) = (8)(1) + (14)(1) + (0) = 22$.

The homogeneous equation $\ddot{x} - 9\dot{x} - 90x = 0$ has constant coefficients, so we try $x = e^{kt}$ (Robinson Sec. 12.1 and 16.1): by the chain rule each derivative multiplies by $k$, and dividing by $e^{kt} \neq 0$ leaves the characteristic equation

$$ P(k) = k^{2} - 9 k - 90 = 0 $$

Its discriminant is $\Delta = b^2 - 4ac = (-9)^2 - 4(1)(-90) = 441$.

Factorising, $P(k) = \left(k - 15\right) \left(k + 6\right)$, so the roots are

$$ k = -6, \qquad k = 15 $$

A real root $k$ of multiplicity $m$ gives $t^j e^{kt}$ for $j = 0, \dots, m - 1$; a complex pair $\rho \pm i\omega$ gives $t^j e^{\rho t}\cos\omega t$ and $t^j e^{\rho t}\sin\omega t$ (Euler's formula, Robinson Sec. 12.3). The fundamental set is $e^{-6 t}$, $e^{15 t}$, so the complementary function is

$$ x_h(t) = c_{1} e^{-6 t} + c_{2} e^{15 t} $$

The equation is homogeneous, so there is no particular integral and the general solution is the complementary function:

$$ x(t) = x_h(t) = c_{1} e^{-6 t} + c_{2} e^{15 t} $$

Initial conditions. Differentiate the general solution (product and chain rules):

$$ \begin{aligned} x'(t) &= - 6 c_{1} e^{-6 t} + 15 c_{2} e^{15 t} \end{aligned} $$

Evaluate at $t = 0$, using $e^{0} = 1$, $\cos 0 = 1$ and $\sin 0 = 0$:

$$ \begin{aligned} x(0) &= c_{1} + c_{2} = 1 \\ x'(0) &= - 6 c_{1} + 15 c_{2} = 22 \end{aligned} $$

The linear system is

$$ \begin{aligned} c_{1} + c_{2} &= 1 \quad (1) \\ - 6 c_{1} + 15 c_{2} &= 22 \quad (2) \end{aligned} $$

Solve $c_{1} + c_{2} = 1$ for $c_{1}$: $c_{1} = 1 - c_{2}$.

Substitute in $- 6 c_{1} + 15 c_{2} = 22$: $21 c_{2} - 6 = 22$.

$21 c_{2} - 6 = 22$ contains one unknown only, so $c_{2} = \frac{4}{3}$.

Back-substitution: $c_{1} = 1 - c_{2} = - \frac{1}{3}$.

Hence $c_{1} = - \frac{1}{3}$, $c_{2} = \frac{4}{3}$.

Substituting the constants into the general solution,

$$ \boxed{x(t) = \frac{4 e^{15 t}}{3} - \frac{e^{-6 t}}{3}} $$

Recovering $y$

From Eq. 26.2, $y = (\dot x - ax - f)/b$. With the solution found,

$$ \dot x(t) = 20 e^{15 t} + 2 e^{-6 t} $$

$$ y(t) = \frac{1}{14}\left(\dot x - (8)x - (0)\right) = \frac{2 e^{15 t}}{3} + \frac{e^{-6 t}}{3} $$

Check in the original system

$$ \begin{aligned} \dot x &= 20 e^{15 t} + 2 e^{-6 t} \\ ax + by + f &= 20 e^{15 t} + 2 e^{-6 t} \\[4pt] \dot y &= 10 e^{15 t} - 2 e^{-6 t} \\ cx + dy + g &= 10 e^{15 t} - 2 e^{-6 t} \end{aligned} $$

Both pairs of expressions coincide, and $x(0) = 1$, $y(0) = 1$ reproduce the initial data.

Interpretation and checks
Item$x(t)$$y(t)$
(i)$- \frac{t^{2}}{6} - \frac{5 t}{18} - \frac{29 e^{3 t}}{27} + \frac{5 e^{2 t}}{4} - \frac{19}{108}$$- \frac{2 t^{2}}{3} - \frac{7 t}{9} - \frac{29 e^{3 t}}{27} + \frac{5 e^{2 t}}{2} - \frac{23}{54}$
(ii)$- \frac{32 e^{t} \sin{2 t}}{17} + \frac{26 e^{t} \cos{2 t}}{17} + \frac{2 \sin{2 t}}{17} - \frac{9 \cos{2 t}}{17}$$\frac{13 e^{t} \sin{2 t}}{17} + \frac{16 e^{t} \cos{2 t}}{17} - \frac{4 \sin{2 t}}{17} + \frac{\cos{2 t}}{17}$
(iii)$\frac{e^{6 t}}{5} - \frac{e^{t}}{5}$$\frac{2 e^{6 t}}{5} + \frac{e^{t}}{10}$
(iv)$t^{2} e^{3 t} + 7 t e^{3 t} + e^{3 t}$$\frac{t^{2} e^{3 t}}{2} + 3 t e^{3 t} - e^{3 t}$
(v)$- \frac{16 e^{t} \sin{3 t}}{111} + \frac{69 e^{t} \cos{3 t}}{37} - \frac{30 \sin{3 t}}{37} + \frac{5 \cos{3 t}}{37}$$- \frac{121 e^{t} \sin{3 t}}{111} - \frac{17 e^{t} \cos{3 t}}{37} + \frac{9 \sin{3 t}}{37} - \frac{20 \cos{3 t}}{37}$
(vi)$\frac{t e^{2 t}}{5} + \frac{62 e^{2 t}}{75} - \frac{e^{-t}}{6} + \frac{17 e^{-3 t}}{50}$$\frac{t e^{2 t}}{5} + \frac{77 e^{2 t}}{75} - \frac{2 e^{-t}}{3} - \frac{34 e^{-3 t}}{25}$
(vii)$\frac{4 e^{15 t}}{3} - \frac{e^{-6 t}}{3}$$\frac{2 e^{15 t}}{3} + \frac{e^{-6 t}}{3}$
  • Every pair $(x, y)$ was substituted in both original equations and in the initial data (shown in each item and re-checked symbolically in scripts/verify_topic3_topic4.py).
  • Characteristic exponents seen: (i) $2, 3$ (real); (ii) $1 \pm 2i$ (spiral) with the forcing $\cos 2t$ at a different frequency; (iii) $6, -1$ (saddle; the forcing $e^{t}$ has exponent $1$, which is not a root, so no extra factor $t$); (iv) $3$ double (improper-node type), and the forcing $e^{3t}$ resonates with the double root, giving $t^2e^{3t}$; (v) $1 \pm 3i$; (vi) $2, -3$ with the forcing exponent $2$ resonant (term $te^{2t}$) and $-1$ not; (vii) $15, -6$ (homogeneous saddle).
  • Cross-check of the homogeneous parts with Topic 4: (iv) and (v) have the same matrices as Examples 8.2 and 7.1; (vii) is the matrix of Exercise 9 (i).
  • Dimensions: with $t$ in seconds the matrix entries have units s$^{-1}$ and the forcings those of $\dot x$; the elimination keeps this consistent ($a + d$ in s$^{-1}$, $ad - bc$ in s$^{-2}$).
Exercise 8

General Solution and Phase Portrait from Eigenvalues and Eigenvectors

Original statement · probls_ch4-i_2627_v2.pdf, p. 2
Eigen-data • Nodes and Saddles • Eight Portraits
Statement

Write down the general solution and draw the phase portrait for the equation $\dot{\mathbf x} = A\mathbf x$, when the eigenvalues and eigenvectors of $A$ are as given in each case. You should take particular care with stable (or unstable) nodes to ensure that the trajectories approach (or move away from) the origin tangent to the correct eigenvector.

  • (i) $\lambda_1 = 1$, $\mathbf v_1 = (1, 1)$ and $\lambda_2 = 2$, $\mathbf v_2 = (1, -1)$;
  • (ii) $\lambda_1 = 1$, $\mathbf v_1 = (1, 0)$ and $\lambda_2 = -2$, $\mathbf v_2 = (1, 1)$;
  • (iii) $\lambda_1 = -2$, $\mathbf v_1 = (1, 2)$ and $\lambda_2 = -3$, $\mathbf v_2 = (2, -3)$;
  • (iv) $\lambda_1 = 3$, $\mathbf v_1 = (2, 3)$ and $\lambda_2 = -5$, $\mathbf v_2 = (0, 1)$;
  • (v) $\lambda_1 = 3$, $\mathbf v_1 = (1, 2)$ and $\lambda_2 = 1$, $\mathbf v_2 = (1, -3)$;
  • (vi) $\lambda_1 = 2$, $\mathbf v_1 = (0, 1)$ and $\lambda_2 = -3$, $\mathbf v_2 = (1, 5)$;
  • (vii) $\lambda_1 = 1$, $\mathbf v_1 = (1, 1)$ and $\lambda_2 = 2$, $\mathbf v_2 = (2, 1)$;
  • (viii) $\lambda_1 = -3$, $\mathbf v_1 = (1, 3)$ and $\lambda_2 = -1$, $\mathbf v_2 = (-3, 2)$;
Setup and classification

In every item the two eigenvalues are real and distinct, with linearly independent eigenvectors (so $\det[\mathbf v_1\ \mathbf v_2] \neq 0$, checked in each item). The general solution is then $\mathbf x = c_1e^{\lambda_1t}\mathbf v_1 + c_2e^{\lambda_2t}\mathbf v_2$ (Robinson Eq. 28.1) and the portrait is determined by the signs of $\lambda_1, \lambda_2$ (Topic 4 Sec. 6 and 9):

SignsOriginBehaviour
$\lambda_1, \lambda_2 < 0$stable nodeall orbits tend to $\mathbf 0$, tangent to the eigenvector of the smaller $\vert \lambda\vert $
$\lambda_1, \lambda_2 > 0$unstable nodeall orbits leave $\mathbf 0$, tangent to the eigenvector of the smaller $\vert \lambda\vert $
$\lambda_1\lambda_2 < 0$saddleonly the orbits on the stable eigenline reach $\mathbf 0$
Item$\lambda_1, \lambda_2$TypeSlow direction (smaller $\vert \lambda\vert $)
(i)$1, 2$unstable node$\mathbf v_1$ = $(1, 1)$
(ii)$1, -2$saddlenot applicable
(iii)$-2, -3$stable node$\mathbf v_1$ = $(1, 2)$
(iv)$3, -5$saddlenot applicable
(v)$3, 1$unstable node$\mathbf v_2$ = $(1, -3)$
(vi)$2, -3$saddlenot applicable
(vii)$1, 2$unstable node$\mathbf v_1$ = $(1, 1)$
(viii)$-3, -1$stable node$\mathbf v_2$ = $(-3, 2)$
Formulation: how each item is treated

For each item: (1) the general solution; (2) the matrix $A = Q\Lambda Q^{-1}$ reconstructed from the data, used to draw and to check; (3) the canonical system $\dot{\mathbf z} = \Lambda\mathbf z$ with $\mathbf z = Q^{-1}\mathbf x$ and its solution; (4) the portrait in the $\mathbf z$-plane and its image under $\mathbf x = Q\mathbf z$ in the $(x, y)$-plane, with the arrows and the tangent direction decided by the rules above.

Method of the class notes [Notes, p. 9]

Case 1 of the notes: $\mathbf x = C_1e^{\lambda_1t}\mathbf v_1 + C_2e^{\lambda_2t}\mathbf v_2$. With $P = (\mathbf v_1, \mathbf v_2)$ (the $Q$ of this page), $\mathbf z = P^{-1}\mathbf x$ gives $\dot{\mathbf z} = D\mathbf z$ and the orbits $z_2 = \tilde C z_1^{\lambda_2/\lambda_1}$. The table of the notes reads the portrait from $r = \lambda_2/\lambda_1$: eigenvalues of one sign give a node, tangent at the origin to the axis of the eigenvalue of smaller modulus ($z_1$-axis when $r > 1$, $z_2$-axis when $0 < r < 1$); opposite signs give a saddle. Back in $\mathbf x$ the $z$-axes become the eigenvector lines.

Derivation of the eight portraits
(i) $\lambda_1 = 1$, $\mathbf v_1 = (1, 1)^T$; $\lambda_2 = 2$, $\mathbf v_2 = (1, -1)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix}, \qquad \det Q = -2, \qquad Q^{-1} = \begin{pmatrix} \frac{1}{2} & \frac{1}{2} \\ \frac{1}{2} & - \frac{1}{2} \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot1 + 1\cdot0 = 1 \\ (Q\Lambda)_{12} &= 1\cdot0 + 1\cdot2 = 2 \\ (Q\Lambda)_{21} &= 1\cdot1 + (-1)\cdot0 = 1 \\ (Q\Lambda)_{22} &= 1\cdot0 + (-1)\cdot2 = -2 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 1 & 2 \\ 1 & -2 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 1\cdot\frac{1}{2} + 2\cdot\frac{1}{2} = \frac{3}{2} \\ (A)_{12} &= 1\cdot\frac{1}{2} + 2\cdot(- \frac{1}{2}) = - \frac{1}{2} \\ (A)_{21} &= 1\cdot\frac{1}{2} + (-2)\cdot\frac{1}{2} = - \frac{1}{2} \\ (A)_{22} &= 1\cdot\frac{1}{2} + (-2)\cdot(- \frac{1}{2}) = \frac{3}{2} \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} \frac{3}{2} & - \frac{1}{2} \\ - \frac{1}{2} & \frac{3}{2} \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 1 \\ 1 & -1 \end{pmatrix} $$

$\det Q = (1)(-1) - (1)(1) = -2$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-2}\begin{pmatrix} -1 & -1 \\ -1 & 1 \end{pmatrix} = \begin{pmatrix} \frac{1}{2} & \frac{1}{2} \\ \frac{1}{2} & - \frac{1}{2} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} \frac{3}{2} & - \frac{1}{2} \\ - \frac{1}{2} & \frac{3}{2} \end{pmatrix}\begin{pmatrix} 1 \\ 1 \end{pmatrix} = \begin{pmatrix} 1 \\ 1 \end{pmatrix} = 1\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} \frac{3}{2} & - \frac{1}{2} \\ - \frac{1}{2} & \frac{3}{2} \end{pmatrix}\begin{pmatrix} 1 \\ -1 \end{pmatrix} = \begin{pmatrix} 2 \\ -2 \end{pmatrix} = 2\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 1 & 2 \\ 1 & -2 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{1}{2}\cdot1 + \frac{1}{2}\cdot1 = 1 \\ (Q^{-1}AQ)_{12} &= \frac{1}{2}\cdot2 + \frac{1}{2}\cdot(-2) = 0 \\ (Q^{-1}AQ)_{21} &= \frac{1}{2}\cdot1 + (- \frac{1}{2})\cdot1 = 0 \\ (Q^{-1}AQ)_{22} &= \frac{1}{2}\cdot2 + (- \frac{1}{2})\cdot(-2) = 2 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix} = \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = z_1, \qquad \dot z_2 = 2z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{1t}, \qquad z_2(t) = c_2 e^{2t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{1t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2 e^{2t}\begin{pmatrix} 1 \\ -1 \end{pmatrix} $$

$$ x(t) = c_{1} e^{t} + c_{2} e^{2 t}, \qquad y(t) = c_{1} e^{t} - c_{2} e^{2 t} $$

Phase portrait

$\lambda_1 = 1$ and $\lambda_2 = 2$ are both positive: the origin is a unstable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{1t}$ and $z_2 = c_2e^{2t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{2}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = 2$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 1$, i.e. to the line through $(1, 1)$ (they leave from there (the limit $t \to -\infty$)), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point away from the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, -1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 1\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = 2\mathbf v_2$ points outward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (unstable node): invariant axes and orbits $z_2 = K|z_1|^{2}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (unstable node): dashed lines are the eigenlines $\mathbf v_1 = (1, 1)$ and $\mathbf v_2 = (1, -1)$.
Open the figure full size
(ii) $\lambda_1 = 1$, $\mathbf v_1 = (1, 0)^T$; $\lambda_2 = -2$, $\mathbf v_2 = (1, 1)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & -2 \end{pmatrix}, \qquad \det Q = 1, \qquad Q^{-1} = \begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot1 + 1\cdot0 = 1 \\ (Q\Lambda)_{12} &= 1\cdot0 + 1\cdot(-2) = -2 \\ (Q\Lambda)_{21} &= 0\cdot1 + 1\cdot0 = 0 \\ (Q\Lambda)_{22} &= 0\cdot0 + 1\cdot(-2) = -2 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 1 & -2 \\ 0 & -2 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 1\cdot1 + (-2)\cdot0 = 1 \\ (A)_{12} &= 1\cdot(-1) + (-2)\cdot1 = -3 \\ (A)_{21} &= 0\cdot1 + (-2)\cdot0 = 0 \\ (A)_{22} &= 0\cdot(-1) + (-2)\cdot1 = -2 \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} 1 & -3 \\ 0 & -2 \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} $$

$\det Q = (1)(1) - (1)(0) = 1$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{1}\begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 1 & -3 \\ 0 & -2 \end{pmatrix}\begin{pmatrix} 1 \\ 0 \end{pmatrix} = \begin{pmatrix} 1 \\ 0 \end{pmatrix} = 1\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 1 & -3 \\ 0 & -2 \end{pmatrix}\begin{pmatrix} 1 \\ 1 \end{pmatrix} = \begin{pmatrix} -2 \\ -2 \end{pmatrix} = (-2)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 1 & -2 \\ 0 & -2 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= 1\cdot1 + (-1)\cdot0 = 1 \\ (Q^{-1}AQ)_{12} &= 1\cdot(-2) + (-1)\cdot(-2) = 0 \\ (Q^{-1}AQ)_{21} &= 0\cdot1 + 1\cdot0 = 0 \\ (Q^{-1}AQ)_{22} &= 0\cdot(-2) + 1\cdot(-2) = -2 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 1 & 0 \\ 0 & -2 \end{pmatrix} = \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & -2 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = z_1, \qquad \dot z_2 = -2z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{1t}, \qquad z_2(t) = c_2 e^{-2t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{1t}\begin{pmatrix} 1 \\ 0 \end{pmatrix} + c_2 e^{-2t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} $$

$$ x(t) = c_{1} e^{t} + c_{2} e^{-2 t}, \qquad y(t) = c_{2} e^{-2 t} $$

Phase portrait

$\lambda_1 = 1$ and $\lambda_2 = -2$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{1t}$ and $z_2 = c_2e^{-2t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{-2}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = -2$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(1, 1)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(1, 0)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 0)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, 1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 1\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -2\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{-2}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (1, 0)$ and $\mathbf v_2 = (1, 1)$.
Open the figure full size
(iii) $\lambda_1 = -2$, $\mathbf v_1 = (1, 2)^T$; $\lambda_2 = -3$, $\mathbf v_2 = (2, -3)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & 2 \\ 2 & -3 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} -2 & 0 \\ 0 & -3 \end{pmatrix}, \qquad \det Q = -7, \qquad Q^{-1} = \begin{pmatrix} \frac{3}{7} & \frac{2}{7} \\ \frac{2}{7} & - \frac{1}{7} \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot(-2) + 2\cdot0 = -2 \\ (Q\Lambda)_{12} &= 1\cdot0 + 2\cdot(-3) = -6 \\ (Q\Lambda)_{21} &= 2\cdot(-2) + (-3)\cdot0 = -4 \\ (Q\Lambda)_{22} &= 2\cdot0 + (-3)\cdot(-3) = 9 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} -2 & -6 \\ -4 & 9 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= (-2)\cdot\frac{3}{7} + (-6)\cdot\frac{2}{7} = - \frac{18}{7} \\ (A)_{12} &= (-2)\cdot\frac{2}{7} + (-6)\cdot(- \frac{1}{7}) = \frac{2}{7} \\ (A)_{21} &= (-4)\cdot\frac{3}{7} + 9\cdot\frac{2}{7} = \frac{6}{7} \\ (A)_{22} &= (-4)\cdot\frac{2}{7} + 9\cdot(- \frac{1}{7}) = - \frac{17}{7} \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} - \frac{18}{7} & \frac{2}{7} \\ \frac{6}{7} & - \frac{17}{7} \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 2 \\ 2 & -3 \end{pmatrix} $$

$\det Q = (1)(-3) - (2)(2) = -7$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-7}\begin{pmatrix} -3 & -2 \\ -2 & 1 \end{pmatrix} = \begin{pmatrix} \frac{3}{7} & \frac{2}{7} \\ \frac{2}{7} & - \frac{1}{7} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} - \frac{18}{7} & \frac{2}{7} \\ \frac{6}{7} & - \frac{17}{7} \end{pmatrix}\begin{pmatrix} 1 \\ 2 \end{pmatrix} = \begin{pmatrix} -2 \\ -4 \end{pmatrix} = (-2)\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} - \frac{18}{7} & \frac{2}{7} \\ \frac{6}{7} & - \frac{17}{7} \end{pmatrix}\begin{pmatrix} 2 \\ -3 \end{pmatrix} = \begin{pmatrix} -6 \\ 9 \end{pmatrix} = (-3)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} -2 & -6 \\ -4 & 9 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{3}{7}\cdot(-2) + \frac{2}{7}\cdot(-4) = -2 \\ (Q^{-1}AQ)_{12} &= \frac{3}{7}\cdot(-6) + \frac{2}{7}\cdot9 = 0 \\ (Q^{-1}AQ)_{21} &= \frac{2}{7}\cdot(-2) + (- \frac{1}{7})\cdot(-4) = 0 \\ (Q^{-1}AQ)_{22} &= \frac{2}{7}\cdot(-6) + (- \frac{1}{7})\cdot9 = -3 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} -2 & 0 \\ 0 & -3 \end{pmatrix} = \Lambda = \begin{pmatrix} -2 & 0 \\ 0 & -3 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = -2z_1, \qquad \dot z_2 = -3z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{-2t}, \qquad z_2(t) = c_2 e^{-3t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{-2t}\begin{pmatrix} 1 \\ 2 \end{pmatrix} + c_2 e^{-3t}\begin{pmatrix} 2 \\ -3 \end{pmatrix} $$

$$ x(t) = c_{1} e^{-2 t} + 2 c_{2} e^{-3 t}, \qquad y(t) = 2 c_{1} e^{-2 t} - 3 c_{2} e^{-3 t} $$

Phase portrait

$\lambda_1 = -2$ and $\lambda_2 = -3$ are both negative: the origin is a stable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{-2t}$ and $z_2 = c_2e^{-3t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{\frac{3}{2}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = \frac{3}{2}$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 2$, i.e. to the line through $(1, 2)$ (they arrive there as $t \to +\infty$), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point towards the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 2)$ and the $z_2$-axis to the line through $\mathbf v_2 = (2, -3)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = -2\mathbf v_1$ points inward; on $\mathbf v_2$: $\dot{\mathbf x} = -3\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (stable node): invariant axes and orbits $z_2 = K|z_1|^{\frac{3}{2}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (stable node): dashed lines are the eigenlines $\mathbf v_1 = (1, 2)$ and $\mathbf v_2 = (2, -3)$.
Open the figure full size
(iv) $\lambda_1 = 3$, $\mathbf v_1 = (2, 3)^T$; $\lambda_2 = -5$, $\mathbf v_2 = (0, 1)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 2 & 0 \\ 3 & 1 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 3 & 0 \\ 0 & -5 \end{pmatrix}, \qquad \det Q = 2, \qquad Q^{-1} = \begin{pmatrix} \frac{1}{2} & 0 \\ - \frac{3}{2} & 1 \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 2\cdot3 + 0\cdot0 = 6 \\ (Q\Lambda)_{12} &= 2\cdot0 + 0\cdot(-5) = 0 \\ (Q\Lambda)_{21} &= 3\cdot3 + 1\cdot0 = 9 \\ (Q\Lambda)_{22} &= 3\cdot0 + 1\cdot(-5) = -5 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 6 & 0 \\ 9 & -5 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 6\cdot\frac{1}{2} + 0\cdot(- \frac{3}{2}) = 3 \\ (A)_{12} &= 6\cdot0 + 0\cdot1 = 0 \\ (A)_{21} &= 9\cdot\frac{1}{2} + (-5)\cdot(- \frac{3}{2}) = 12 \\ (A)_{22} &= 9\cdot0 + (-5)\cdot1 = -5 \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} 3 & 0 \\ 12 & -5 \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 2 & 0 \\ 3 & 1 \end{pmatrix} $$

$\det Q = (2)(1) - (0)(3) = 2$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{2}\begin{pmatrix} 1 & 0 \\ -3 & 2 \end{pmatrix} = \begin{pmatrix} \frac{1}{2} & 0 \\ - \frac{3}{2} & 1 \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 3 & 0 \\ 12 & -5 \end{pmatrix}\begin{pmatrix} 2 \\ 3 \end{pmatrix} = \begin{pmatrix} 6 \\ 9 \end{pmatrix} = 3\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 3 & 0 \\ 12 & -5 \end{pmatrix}\begin{pmatrix} 0 \\ 1 \end{pmatrix} = \begin{pmatrix} 0 \\ -5 \end{pmatrix} = (-5)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 6 & 0 \\ 9 & -5 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{1}{2}\cdot6 + 0\cdot9 = 3 \\ (Q^{-1}AQ)_{12} &= \frac{1}{2}\cdot0 + 0\cdot(-5) = 0 \\ (Q^{-1}AQ)_{21} &= (- \frac{3}{2})\cdot6 + 1\cdot9 = 0 \\ (Q^{-1}AQ)_{22} &= (- \frac{3}{2})\cdot0 + 1\cdot(-5) = -5 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 3 & 0 \\ 0 & -5 \end{pmatrix} = \Lambda = \begin{pmatrix} 3 & 0 \\ 0 & -5 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 3z_1, \qquad \dot z_2 = -5z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{3t}, \qquad z_2(t) = c_2 e^{-5t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{3t}\begin{pmatrix} 2 \\ 3 \end{pmatrix} + c_2 e^{-5t}\begin{pmatrix} 0 \\ 1 \end{pmatrix} $$

$$ x(t) = 2 c_{1} e^{3 t}, \qquad y(t) = 3 c_{1} e^{3 t} + c_{2} e^{-5 t} $$

Phase portrait

$\lambda_1 = 3$ and $\lambda_2 = -5$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{3t}$ and $z_2 = c_2e^{-5t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{- \frac{5}{3}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = - \frac{5}{3}$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(0, 1)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(2, 3)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (2, 3)$ and the $z_2$-axis to the line through $\mathbf v_2 = (0, 1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 3\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -5\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{- \frac{5}{3}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (2, 3)$ and $\mathbf v_2 = (0, 1)$.
Open the figure full size
(v) $\lambda_1 = 3$, $\mathbf v_1 = (1, 2)^T$; $\lambda_2 = 1$, $\mathbf v_2 = (1, -3)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & 1 \\ 2 & -3 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 3 & 0 \\ 0 & 1 \end{pmatrix}, \qquad \det Q = -5, \qquad Q^{-1} = \begin{pmatrix} \frac{3}{5} & \frac{1}{5} \\ \frac{2}{5} & - \frac{1}{5} \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot3 + 1\cdot0 = 3 \\ (Q\Lambda)_{12} &= 1\cdot0 + 1\cdot1 = 1 \\ (Q\Lambda)_{21} &= 2\cdot3 + (-3)\cdot0 = 6 \\ (Q\Lambda)_{22} &= 2\cdot0 + (-3)\cdot1 = -3 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 3 & 1 \\ 6 & -3 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 3\cdot\frac{3}{5} + 1\cdot\frac{2}{5} = \frac{11}{5} \\ (A)_{12} &= 3\cdot\frac{1}{5} + 1\cdot(- \frac{1}{5}) = \frac{2}{5} \\ (A)_{21} &= 6\cdot\frac{3}{5} + (-3)\cdot\frac{2}{5} = \frac{12}{5} \\ (A)_{22} &= 6\cdot\frac{1}{5} + (-3)\cdot(- \frac{1}{5}) = \frac{9}{5} \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} \frac{11}{5} & \frac{2}{5} \\ \frac{12}{5} & \frac{9}{5} \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 1 \\ 2 & -3 \end{pmatrix} $$

$\det Q = (1)(-3) - (1)(2) = -5$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-5}\begin{pmatrix} -3 & -1 \\ -2 & 1 \end{pmatrix} = \begin{pmatrix} \frac{3}{5} & \frac{1}{5} \\ \frac{2}{5} & - \frac{1}{5} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} \frac{11}{5} & \frac{2}{5} \\ \frac{12}{5} & \frac{9}{5} \end{pmatrix}\begin{pmatrix} 1 \\ 2 \end{pmatrix} = \begin{pmatrix} 3 \\ 6 \end{pmatrix} = 3\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} \frac{11}{5} & \frac{2}{5} \\ \frac{12}{5} & \frac{9}{5} \end{pmatrix}\begin{pmatrix} 1 \\ -3 \end{pmatrix} = \begin{pmatrix} 1 \\ -3 \end{pmatrix} = 1\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 3 & 1 \\ 6 & -3 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{3}{5}\cdot3 + \frac{1}{5}\cdot6 = 3 \\ (Q^{-1}AQ)_{12} &= \frac{3}{5}\cdot1 + \frac{1}{5}\cdot(-3) = 0 \\ (Q^{-1}AQ)_{21} &= \frac{2}{5}\cdot3 + (- \frac{1}{5})\cdot6 = 0 \\ (Q^{-1}AQ)_{22} &= \frac{2}{5}\cdot1 + (- \frac{1}{5})\cdot(-3) = 1 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 3 & 0 \\ 0 & 1 \end{pmatrix} = \Lambda = \begin{pmatrix} 3 & 0 \\ 0 & 1 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 3z_1, \qquad \dot z_2 = z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{3t}, \qquad z_2(t) = c_2 e^{1t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{3t}\begin{pmatrix} 1 \\ 2 \end{pmatrix} + c_2 e^{1t}\begin{pmatrix} 1 \\ -3 \end{pmatrix} $$

$$ x(t) = c_{1} e^{3 t} + c_{2} e^{t}, \qquad y(t) = 2 c_{1} e^{3 t} - 3 c_{2} e^{t} $$

Phase portrait

$\lambda_1 = 3$ and $\lambda_2 = 1$ are both positive: the origin is a unstable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{3t}$ and $z_2 = c_2e^{1t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{\frac{1}{3}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = \frac{1}{3}$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 1$, i.e. to the line through $(1, -3)$ (they leave from there (the limit $t \to -\infty$)), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point away from the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 2)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, -3)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 3\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = 1\mathbf v_2$ points outward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (unstable node): invariant axes and orbits $z_2 = K|z_1|^{\frac{1}{3}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (unstable node): dashed lines are the eigenlines $\mathbf v_1 = (1, 2)$ and $\mathbf v_2 = (1, -3)$.
Open the figure full size
(vi) $\lambda_1 = 2$, $\mathbf v_1 = (0, 1)^T$; $\lambda_2 = -3$, $\mathbf v_2 = (1, 5)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 0 & 1 \\ 1 & 5 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 2 & 0 \\ 0 & -3 \end{pmatrix}, \qquad \det Q = -1, \qquad Q^{-1} = \begin{pmatrix} -5 & 1 \\ 1 & 0 \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 0\cdot2 + 1\cdot0 = 0 \\ (Q\Lambda)_{12} &= 0\cdot0 + 1\cdot(-3) = -3 \\ (Q\Lambda)_{21} &= 1\cdot2 + 5\cdot0 = 2 \\ (Q\Lambda)_{22} &= 1\cdot0 + 5\cdot(-3) = -15 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 0 & -3 \\ 2 & -15 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 0\cdot(-5) + (-3)\cdot1 = -3 \\ (A)_{12} &= 0\cdot1 + (-3)\cdot0 = 0 \\ (A)_{21} &= 2\cdot(-5) + (-15)\cdot1 = -25 \\ (A)_{22} &= 2\cdot1 + (-15)\cdot0 = 2 \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} -3 & 0 \\ -25 & 2 \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 0 & 1 \\ 1 & 5 \end{pmatrix} $$

$\det Q = (0)(5) - (1)(1) = -1$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-1}\begin{pmatrix} 5 & -1 \\ -1 & 0 \end{pmatrix} = \begin{pmatrix} -5 & 1 \\ 1 & 0 \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} -3 & 0 \\ -25 & 2 \end{pmatrix}\begin{pmatrix} 0 \\ 1 \end{pmatrix} = \begin{pmatrix} 0 \\ 2 \end{pmatrix} = 2\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} -3 & 0 \\ -25 & 2 \end{pmatrix}\begin{pmatrix} 1 \\ 5 \end{pmatrix} = \begin{pmatrix} -3 \\ -15 \end{pmatrix} = (-3)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 0 & -3 \\ 2 & -15 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= (-5)\cdot0 + 1\cdot2 = 2 \\ (Q^{-1}AQ)_{12} &= (-5)\cdot(-3) + 1\cdot(-15) = 0 \\ (Q^{-1}AQ)_{21} &= 1\cdot0 + 0\cdot2 = 0 \\ (Q^{-1}AQ)_{22} &= 1\cdot(-3) + 0\cdot(-15) = -3 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 2 & 0 \\ 0 & -3 \end{pmatrix} = \Lambda = \begin{pmatrix} 2 & 0 \\ 0 & -3 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 2z_1, \qquad \dot z_2 = -3z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{2t}, \qquad z_2(t) = c_2 e^{-3t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{2t}\begin{pmatrix} 0 \\ 1 \end{pmatrix} + c_2 e^{-3t}\begin{pmatrix} 1 \\ 5 \end{pmatrix} $$

$$ x(t) = c_{2} e^{-3 t}, \qquad y(t) = c_{1} e^{2 t} + 5 c_{2} e^{-3 t} $$

Phase portrait

$\lambda_1 = 2$ and $\lambda_2 = -3$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{2t}$ and $z_2 = c_2e^{-3t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{- \frac{3}{2}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = - \frac{3}{2}$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(1, 5)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(0, 1)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (0, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, 5)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 2\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -3\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{- \frac{3}{2}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (0, 1)$ and $\mathbf v_2 = (1, 5)$.
Open the figure full size
(vii) $\lambda_1 = 1$, $\mathbf v_1 = (1, 1)^T$; $\lambda_2 = 2$, $\mathbf v_2 = (2, 1)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & 2 \\ 1 & 1 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix}, \qquad \det Q = -1, \qquad Q^{-1} = \begin{pmatrix} -1 & 2 \\ 1 & -1 \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot1 + 2\cdot0 = 1 \\ (Q\Lambda)_{12} &= 1\cdot0 + 2\cdot2 = 4 \\ (Q\Lambda)_{21} &= 1\cdot1 + 1\cdot0 = 1 \\ (Q\Lambda)_{22} &= 1\cdot0 + 1\cdot2 = 2 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} 1 & 4 \\ 1 & 2 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= 1\cdot(-1) + 4\cdot1 = 3 \\ (A)_{12} &= 1\cdot2 + 4\cdot(-1) = -2 \\ (A)_{21} &= 1\cdot(-1) + 2\cdot1 = 1 \\ (A)_{22} &= 1\cdot2 + 2\cdot(-1) = 0 \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} 3 & -2 \\ 1 & 0 \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 2 \\ 1 & 1 \end{pmatrix} $$

$\det Q = (1)(1) - (2)(1) = -1$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-1}\begin{pmatrix} 1 & -2 \\ -1 & 1 \end{pmatrix} = \begin{pmatrix} -1 & 2 \\ 1 & -1 \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 3 & -2 \\ 1 & 0 \end{pmatrix}\begin{pmatrix} 1 \\ 1 \end{pmatrix} = \begin{pmatrix} 1 \\ 1 \end{pmatrix} = 1\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 3 & -2 \\ 1 & 0 \end{pmatrix}\begin{pmatrix} 2 \\ 1 \end{pmatrix} = \begin{pmatrix} 4 \\ 2 \end{pmatrix} = 2\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 1 & 4 \\ 1 & 2 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= (-1)\cdot1 + 2\cdot1 = 1 \\ (Q^{-1}AQ)_{12} &= (-1)\cdot4 + 2\cdot2 = 0 \\ (Q^{-1}AQ)_{21} &= 1\cdot1 + (-1)\cdot1 = 0 \\ (Q^{-1}AQ)_{22} &= 1\cdot4 + (-1)\cdot2 = 2 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix} = \Lambda = \begin{pmatrix} 1 & 0 \\ 0 & 2 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = z_1, \qquad \dot z_2 = 2z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{1t}, \qquad z_2(t) = c_2 e^{2t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{1t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2 e^{2t}\begin{pmatrix} 2 \\ 1 \end{pmatrix} $$

$$ x(t) = c_{1} e^{t} + 2 c_{2} e^{2 t}, \qquad y(t) = c_{1} e^{t} + c_{2} e^{2 t} $$

Phase portrait

$\lambda_1 = 1$ and $\lambda_2 = 2$ are both positive: the origin is a unstable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{1t}$ and $z_2 = c_2e^{2t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{2}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = 2$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 1$, i.e. to the line through $(1, 1)$ (they leave from there (the limit $t \to -\infty$)), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point away from the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (2, 1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 1\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = 2\mathbf v_2$ points outward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (unstable node): invariant axes and orbits $z_2 = K|z_1|^{2}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (unstable node): dashed lines are the eigenlines $\mathbf v_1 = (1, 1)$ and $\mathbf v_2 = (2, 1)$.
Open the figure full size
(viii) $\lambda_1 = -3$, $\mathbf v_1 = (1, 3)^T$; $\lambda_2 = -1$, $\mathbf v_2 = (-3, 2)^T$

The statement gives the eigen-data, not the matrix. Since $A\mathbf v_j = \lambda_j\mathbf v_j$ for $j = 1, 2$ means $AQ = Q\Lambda$ with $Q = [\mathbf v_1 \mid \mathbf v_2]$, the matrix is $A = Q\Lambda Q^{-1}$ (it is needed to draw the portrait and to check the data):

$$ Q = \begin{pmatrix} 1 & -3 \\ 3 & 2 \end{pmatrix}, \qquad \Lambda = \begin{pmatrix} -3 & 0 \\ 0 & -1 \end{pmatrix}, \qquad \det Q = 11, \qquad Q^{-1} = \begin{pmatrix} \frac{2}{11} & \frac{3}{11} \\ - \frac{3}{11} & \frac{1}{11} \end{pmatrix} $$

$$ \begin{aligned} (Q\Lambda)_{11} &= 1\cdot(-3) + (-3)\cdot0 = -3 \\ (Q\Lambda)_{12} &= 1\cdot0 + (-3)\cdot(-1) = 3 \\ (Q\Lambda)_{21} &= 3\cdot(-3) + 2\cdot0 = -9 \\ (Q\Lambda)_{22} &= 3\cdot0 + 2\cdot(-1) = -2 \end{aligned} $$

$$ Q\Lambda = \begin{pmatrix} -3 & 3 \\ -9 & -2 \end{pmatrix} $$

$$ \begin{aligned} (A)_{11} &= (-3)\cdot\frac{2}{11} + 3\cdot(- \frac{3}{11}) = - \frac{15}{11} \\ (A)_{12} &= (-3)\cdot\frac{3}{11} + 3\cdot\frac{1}{11} = - \frac{6}{11} \\ (A)_{21} &= (-9)\cdot\frac{2}{11} + (-2)\cdot(- \frac{3}{11}) = - \frac{12}{11} \\ (A)_{22} &= (-9)\cdot\frac{3}{11} + (-2)\cdot\frac{1}{11} = - \frac{29}{11} \end{aligned} $$

$$ A = Q\Lambda Q^{-1} = \begin{pmatrix} - \frac{15}{11} & - \frac{6}{11} \\ - \frac{12}{11} & - \frac{29}{11} \end{pmatrix} $$

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & -3 \\ 3 & 2 \end{pmatrix} $$

$\det Q = (1)(2) - (-3)(3) = 11$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{11}\begin{pmatrix} 2 & 3 \\ -3 & 1 \end{pmatrix} = \begin{pmatrix} \frac{2}{11} & \frac{3}{11} \\ - \frac{3}{11} & \frac{1}{11} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} - \frac{15}{11} & - \frac{6}{11} \\ - \frac{12}{11} & - \frac{29}{11} \end{pmatrix}\begin{pmatrix} 1 \\ 3 \end{pmatrix} = \begin{pmatrix} -3 \\ -9 \end{pmatrix} = (-3)\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} - \frac{15}{11} & - \frac{6}{11} \\ - \frac{12}{11} & - \frac{29}{11} \end{pmatrix}\begin{pmatrix} -3 \\ 2 \end{pmatrix} = \begin{pmatrix} 3 \\ -2 \end{pmatrix} = (-1)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} -3 & 3 \\ -9 & -2 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{2}{11}\cdot(-3) + \frac{3}{11}\cdot(-9) = -3 \\ (Q^{-1}AQ)_{12} &= \frac{2}{11}\cdot3 + \frac{3}{11}\cdot(-2) = 0 \\ (Q^{-1}AQ)_{21} &= (- \frac{3}{11})\cdot(-3) + \frac{1}{11}\cdot(-9) = 0 \\ (Q^{-1}AQ)_{22} &= (- \frac{3}{11})\cdot3 + \frac{1}{11}\cdot(-2) = -1 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} -3 & 0 \\ 0 & -1 \end{pmatrix} = \Lambda = \begin{pmatrix} -3 & 0 \\ 0 & -1 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = -3z_1, \qquad \dot z_2 = -z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{-3t}, \qquad z_2(t) = c_2 e^{-1t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{-3t}\begin{pmatrix} 1 \\ 3 \end{pmatrix} + c_2 e^{-1t}\begin{pmatrix} -3 \\ 2 \end{pmatrix} $$

$$ x(t) = c_{1} e^{-3 t} - 3 c_{2} e^{-t}, \qquad y(t) = 3 c_{1} e^{-3 t} + 2 c_{2} e^{-t} $$

Phase portrait

$\lambda_1 = -3$ and $\lambda_2 = -1$ are both negative: the origin is a stable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{-3t}$ and $z_2 = c_2e^{-1t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{\frac{1}{3}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = \frac{1}{3}$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 1$, i.e. to the line through $(-3, 2)$ (they arrive there as $t \to +\infty$), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point towards the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 3)$ and the $z_2$-axis to the line through $\mathbf v_2 = (-3, 2)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = -3\mathbf v_1$ points inward; on $\mathbf v_2$: $\dot{\mathbf x} = -1\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (stable node): invariant axes and orbits $z_2 = K|z_1|^{\frac{1}{3}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (stable node): dashed lines are the eigenlines $\mathbf v_1 = (1, 3)$ and $\mathbf v_2 = (-3, 2)$.
Open the figure full size
Interpretation and checks
ItemGeneral solution
(i)$\mathbf x = c_1e^{1t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2e^{2t}\begin{pmatrix} 1 \\ -1 \end{pmatrix}$
(ii)$\mathbf x = c_1e^{1t}\begin{pmatrix} 1 \\ 0 \end{pmatrix} + c_2e^{-2t}\begin{pmatrix} 1 \\ 1 \end{pmatrix}$
(iii)$\mathbf x = c_1e^{-2t}\begin{pmatrix} 1 \\ 2 \end{pmatrix} + c_2e^{-3t}\begin{pmatrix} 2 \\ -3 \end{pmatrix}$
(iv)$\mathbf x = c_1e^{3t}\begin{pmatrix} 2 \\ 3 \end{pmatrix} + c_2e^{-5t}\begin{pmatrix} 0 \\ 1 \end{pmatrix}$
(v)$\mathbf x = c_1e^{3t}\begin{pmatrix} 1 \\ 2 \end{pmatrix} + c_2e^{1t}\begin{pmatrix} 1 \\ -3 \end{pmatrix}$
(vi)$\mathbf x = c_1e^{2t}\begin{pmatrix} 0 \\ 1 \end{pmatrix} + c_2e^{-3t}\begin{pmatrix} 1 \\ 5 \end{pmatrix}$
(vii)$\mathbf x = c_1e^{1t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2e^{2t}\begin{pmatrix} 2 \\ 1 \end{pmatrix}$
(viii)$\mathbf x = c_1e^{-3t}\begin{pmatrix} 1 \\ 3 \end{pmatrix} + c_2e^{-1t}\begin{pmatrix} -3 \\ 2 \end{pmatrix}$
  • In every item the reconstructed $A$ satisfies $A\mathbf v_j = \lambda_j\mathbf v_j$ (checked in the text) and $\det A = \lambda_1\lambda_2$, $\operatorname{tr}A = \lambda_1 + \lambda_2$.
  • Nodes: the tangent direction at the origin is the eigenvector of the smaller $|\lambda|$ because $|z_2|/|z_1|^{\lambda_2/\lambda_1}$ is constant: if $|\lambda_2| > |\lambda_1|$ the exponent $r = \lambda_2/\lambda_1$ exceeds $1$, so $z_2 = K|z_1|^{r}$ tends to $0$ faster than $z_1$ and the curve hugs the $z_1$-axis (the slower eigendirection) near the origin; if $r < 1$ it hugs the $z_2$-axis, which is then the slower one. The phase portraits drawn above show this numerically (curves integrated from the matrix $A$ itself, not from the formula, so the tangency is a genuine check).
  • Saddles (ii), (iv), (vi): the stable manifold is the eigenline of the negative eigenvalue.
  • Items (ii), (iv) and (vi) have an eigenvector with a zero entry ((ii) $\mathbf v_1 = (1, 0)$, (iv) $\mathbf v_2 = (0, 1)$, (vi) $\mathbf v_1 = (0, 1)$): that eigenline is an axis of the $(x, y)$-plane.
Exercise 9

Decoupling Four Systems with Real Eigenvalues

Original statement · probls_ch4-i_2627_v2.pdf, p. 3
Diagonalisation • Decoupling • Four Matrices
Statement

Find the eigenvalues and eigenvectors of the matrix on the right-hand side in each case, and hence find the coordinate transformation that will decouple the corresponding equations. Show that this transformation has the desired effect. You can also write down the general solution and draw the phase portrait for the equation.

  • (i) $\dfrac{d\mathbf x}{dt} = \begin{pmatrix} 8 & 14 \\ 7 & 1 \end{pmatrix}\mathbf x$;
  • (ii) $\dfrac{d\mathbf x}{dt} = \begin{pmatrix} 2 & 0 \\ -5 & -3 \end{pmatrix}\mathbf x$;
  • (iii) $\dfrac{d\mathbf x}{dt} = \begin{pmatrix} 11 & -2 \\ 3 & 4 \end{pmatrix}\mathbf x$;
  • (iv) $\dfrac{d\mathbf x}{dt} = \begin{pmatrix} 1 & 20 \\ 40 & -19 \end{pmatrix}\mathbf x$.
Setup

Each system is $\dot{\mathbf x} = A\mathbf x$ with a constant $2\times 2$ matrix. Plan (Topic 4 Sec. 4 and 5): (a) characteristic equation $\det(A - \lambda I) = 0$; (b) eigenvectors from $(A - \lambda I)\mathbf v = \mathbf 0$; (c) if the eigenvalues are real and distinct the eigenvectors are independent and $Q = [\mathbf v_1\ \mathbf v_2]$ defines $\mathbf z = Q^{-1}\mathbf x$ with $\dot{\mathbf z} = Q^{-1}AQ\,\mathbf z$; (d) show $Q^{-1}AQ = \Lambda$ by explicit multiplication.

Formulation

Since $\mathbf x = Q\mathbf z$ with $Q$ constant, $\dot{\mathbf x} = Q\dot{\mathbf z}$ and $A\mathbf x = AQ\mathbf z$, so $\dot{\mathbf z} = Q^{-1}AQ\,\mathbf z$. If $Q^{-1}AQ = \Lambda = \operatorname{diag}(\lambda_1, \lambda_2)$ the equations $\dot z_1 = \lambda_1z_1$, $\dot z_2 = \lambda_2z_2$ are uncoupled.

Method of the class notes [Notes, p. 9]

The diagonalisation of the notes: $A = PDP^{-1}$ with $P = (\mathbf v_1, \mathbf v_2)$ and $D = \operatorname{diag}(\lambda_1, \lambda_2)$; with $\mathbf z = P^{-1}\mathbf x$, $\dot{\mathbf z} = P^{-1}A\mathbf x = DP^{-1}\mathbf x = D\mathbf z$, so $z_1 = C_1e^{\lambda_1t}$ and $z_2 = C_2e^{\lambda_2t}$ are uncoupled. The $Q$ and $\Lambda$ of this page are the notes' $P$ and $D$. Item (i) is the matrix of Exercise 7 (vii): elimination and eigenvalues give the same characteristic polynomial, $\lambda^2 - 9\lambda - 90$.

Derivation
(i) $A = \begin{pmatrix} 8 & 14 \\ 7 & 1 \end{pmatrix}$

Characteristic equation $\det(A - \lambda I) = 0$ (Topic 4, Eq. 27.8):

$$ \det(A - \lambda I) = \begin{vmatrix} 8 - \lambda & 14 \\ 7 & 1 - \lambda \end{vmatrix} = (8 - \lambda)(1 - \lambda) - (14)(7) $$

$$ = \lambda^2 - (8 + 1)\lambda + (8\cdot1 - 14\cdot7) = \lambda^2 - (9)\lambda + (-90) = \lambda^{2} - 9 \lambda - 90 $$

$\operatorname{tr}A = 9$, $\det A = -90$, discriminant $\Delta = (\operatorname{tr}A)^2 - 4\det A = 81 - (-360) = 441$.

$$ \lambda = \frac{9 \pm \sqrt{441}}{2} = \frac{9 \pm 21}{2} \quad\Longrightarrow\quad \lambda_1 = 15,\ \lambda_2 = -6 $$

Check: $\lambda_1 + \lambda_2 = 9 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = -90 = \det A$.

Eigenvectors

For $\lambda = 15$: $(A - (15)I)\mathbf{v_1} = \mathbf 0$, that is

$$ \begin{pmatrix} -7 & 14 \\ 7 & -14 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} -7v_1 + 14v_2 = 0\quad (\text{row 1}) \\ 7v_1 - 14v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $-1$ times row 1) and express a single condition. Row 1 reads $-7v_1 + 14v_2 = 0$, so $v_1 = 2\,v_2$: every eigenvector is a multiple of $(2, 1)^T$. The smallest integer representative is

$$ \mathbf{v_1} = \begin{pmatrix} 2 \\ 1 \end{pmatrix} $$

Check: $A\mathbf{v_1} = \begin{pmatrix} 30 \\ 15 \end{pmatrix} = 15\mathbf{v_1}$.

For $\lambda = -6$: $(A - (-6)I)\mathbf{v_2} = \mathbf 0$, that is

$$ \begin{pmatrix} 14 & 14 \\ 7 & 7 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 14v_1 + 14v_2 = 0\quad (\text{row 1}) \\ 7v_1 + 7v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $\frac{1}{2}$ times row 1) and express a single condition. Row 1 reads $14v_1 + 14v_2 = 0$, so $v_1 = -v_2$: every eigenvector is a multiple of $(-1, 1)^T$. The smallest integer representative is

$$ \mathbf{v_2} = \begin{pmatrix} 1 \\ -1 \end{pmatrix} $$

Check: $A\mathbf{v_2} = \begin{pmatrix} -6 \\ 6 \end{pmatrix} = -6\mathbf{v_2}$.

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 2 & 1 \\ 1 & -1 \end{pmatrix} $$

$\det Q = (2)(-1) - (1)(1) = -3$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-3}\begin{pmatrix} -1 & -1 \\ -1 & 2 \end{pmatrix} = \begin{pmatrix} \frac{1}{3} & \frac{1}{3} \\ \frac{1}{3} & - \frac{2}{3} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 8 & 14 \\ 7 & 1 \end{pmatrix}\begin{pmatrix} 2 \\ 1 \end{pmatrix} = \begin{pmatrix} 30 \\ 15 \end{pmatrix} = 15\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 8 & 14 \\ 7 & 1 \end{pmatrix}\begin{pmatrix} 1 \\ -1 \end{pmatrix} = \begin{pmatrix} -6 \\ 6 \end{pmatrix} = (-6)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 30 & -6 \\ 15 & 6 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{1}{3}\cdot30 + \frac{1}{3}\cdot15 = 15 \\ (Q^{-1}AQ)_{12} &= \frac{1}{3}\cdot(-6) + \frac{1}{3}\cdot6 = 0 \\ (Q^{-1}AQ)_{21} &= \frac{1}{3}\cdot30 + (- \frac{2}{3})\cdot15 = 0 \\ (Q^{-1}AQ)_{22} &= \frac{1}{3}\cdot(-6) + (- \frac{2}{3})\cdot6 = -6 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 15 & 0 \\ 0 & -6 \end{pmatrix} = \Lambda = \begin{pmatrix} 15 & 0 \\ 0 & -6 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 15z_1, \qquad \dot z_2 = -6z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{15t}, \qquad z_2(t) = c_2 e^{-6t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{15t}\begin{pmatrix} 2 \\ 1 \end{pmatrix} + c_2 e^{-6t}\begin{pmatrix} 1 \\ -1 \end{pmatrix} $$

$$ x(t) = 2 c_{1} e^{15 t} + c_{2} e^{-6 t}, \qquad y(t) = c_{1} e^{15 t} - c_{2} e^{-6 t} $$

The decoupling transformation

In the original variables $\mathbf x = (x, y)^T$ the new coordinates are $\mathbf z = Q^{-1}\mathbf x$:

$$ z_1 = \frac{x}{3} + \frac{y}{3}, \qquad z_2 = \frac{x}{3} - \frac{2 y}{3} $$

Phase portrait

$\lambda_1 = 15$ and $\lambda_2 = -6$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{15t}$ and $z_2 = c_2e^{-6t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{- \frac{2}{5}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = - \frac{2}{5}$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(1, -1)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(2, 1)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (2, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, -1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 15\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -6\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{- \frac{2}{5}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (2, 1)$ and $\mathbf v_2 = (1, -1)$.
Open the figure full size
(ii) $A = \begin{pmatrix} 2 & 0 \\ -5 & -3 \end{pmatrix}$

Characteristic equation $\det(A - \lambda I) = 0$ (Topic 4, Eq. 27.8):

$$ \det(A - \lambda I) = \begin{vmatrix} 2 - \lambda & 0 \\ -5 & -3 - \lambda \end{vmatrix} = (2 - \lambda)(-3 - \lambda) - (0)(-5) $$

$$ = \lambda^2 - (2 + (-3))\lambda + (2\cdot(-3) - 0\cdot(-5)) = \lambda^2 - (-1)\lambda + (-6) = \lambda^{2} + \lambda - 6 $$

$\operatorname{tr}A = -1$, $\det A = -6$, discriminant $\Delta = (\operatorname{tr}A)^2 - 4\det A = 1 - (-24) = 25$.

$$ \lambda = \frac{-1 \pm \sqrt{25}}{2} = \frac{-1 \pm 5}{2} \quad\Longrightarrow\quad \lambda_1 = 2,\ \lambda_2 = -3 $$

Check: $\lambda_1 + \lambda_2 = -1 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = -6 = \det A$.

Eigenvectors

For $\lambda = 2$: $(A - (2)I)\mathbf{v_1} = \mathbf 0$, that is

$$ \begin{pmatrix} 0 & 0 \\ -5 & -5 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 0v_1 + 0v_2 = 0\quad (\text{row 1}) \\ -5v_1 - 5v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 1 is identically zero) and express a single condition. Row 2 reads $-5v_1 + (-5)v_2 = 0$, so $v_1 = -v_2$: every eigenvector is a multiple of $(-1, 1)^T$. The smallest integer representative is

$$ \mathbf{v_1} = \begin{pmatrix} 1 \\ -1 \end{pmatrix} $$

Check: $A\mathbf{v_1} = \begin{pmatrix} 2 \\ -2 \end{pmatrix} = 2\mathbf{v_1}$.

For $\lambda = -3$: $(A - (-3)I)\mathbf{v_2} = \mathbf 0$, that is

$$ \begin{pmatrix} 5 & 0 \\ -5 & 0 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 5v_1 + 0v_2 = 0\quad (\text{row 1}) \\ -5v_1 + 0v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $-1$ times row 1) and express a single condition. Row 1 reads $5v_1 + 0v_2 = 0$, so $v_1 = 0$ and $v_2$ is free; the simplest choice is

$$ \mathbf{v_2} = \begin{pmatrix} 0 \\ 1 \end{pmatrix} $$

Check: $A\mathbf{v_2} = \begin{pmatrix} 0 \\ -3 \end{pmatrix} = -3\mathbf{v_2}$.

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 0 \\ -1 & 1 \end{pmatrix} $$

$\det Q = (1)(1) - (0)(-1) = 1$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{1}\begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 2 & 0 \\ -5 & -3 \end{pmatrix}\begin{pmatrix} 1 \\ -1 \end{pmatrix} = \begin{pmatrix} 2 \\ -2 \end{pmatrix} = 2\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 2 & 0 \\ -5 & -3 \end{pmatrix}\begin{pmatrix} 0 \\ 1 \end{pmatrix} = \begin{pmatrix} 0 \\ -3 \end{pmatrix} = (-3)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 2 & 0 \\ -2 & -3 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= 1\cdot2 + 0\cdot(-2) = 2 \\ (Q^{-1}AQ)_{12} &= 1\cdot0 + 0\cdot(-3) = 0 \\ (Q^{-1}AQ)_{21} &= 1\cdot2 + 1\cdot(-2) = 0 \\ (Q^{-1}AQ)_{22} &= 1\cdot0 + 1\cdot(-3) = -3 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 2 & 0 \\ 0 & -3 \end{pmatrix} = \Lambda = \begin{pmatrix} 2 & 0 \\ 0 & -3 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 2z_1, \qquad \dot z_2 = -3z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{2t}, \qquad z_2(t) = c_2 e^{-3t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{2t}\begin{pmatrix} 1 \\ -1 \end{pmatrix} + c_2 e^{-3t}\begin{pmatrix} 0 \\ 1 \end{pmatrix} $$

$$ x(t) = c_{1} e^{2 t}, \qquad y(t) = - c_{1} e^{2 t} + c_{2} e^{-3 t} $$

The decoupling transformation

In the original variables $\mathbf x = (x, y)^T$ the new coordinates are $\mathbf z = Q^{-1}\mathbf x$:

$$ z_1 = x, \qquad z_2 = x + y $$

Phase portrait

$\lambda_1 = 2$ and $\lambda_2 = -3$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{2t}$ and $z_2 = c_2e^{-3t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{- \frac{3}{2}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = - \frac{3}{2}$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(0, 1)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(1, -1)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, -1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (0, 1)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 2\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -3\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{- \frac{3}{2}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (1, -1)$ and $\mathbf v_2 = (0, 1)$.
Open the figure full size
(iii) $A = \begin{pmatrix} 11 & -2 \\ 3 & 4 \end{pmatrix}$

Characteristic equation $\det(A - \lambda I) = 0$ (Topic 4, Eq. 27.8):

$$ \det(A - \lambda I) = \begin{vmatrix} 11 - \lambda & -2 \\ 3 & 4 - \lambda \end{vmatrix} = (11 - \lambda)(4 - \lambda) - (-2)(3) $$

$$ = \lambda^2 - (11 + 4)\lambda + (11\cdot4 - (-2)\cdot3) = \lambda^2 - (15)\lambda + (50) = \lambda^{2} - 15 \lambda + 50 $$

$\operatorname{tr}A = 15$, $\det A = 50$, discriminant $\Delta = (\operatorname{tr}A)^2 - 4\det A = 225 - 200 = 25$.

$$ \lambda = \frac{15 \pm \sqrt{25}}{2} = \frac{15 \pm 5}{2} \quad\Longrightarrow\quad \lambda_1 = 10,\ \lambda_2 = 5 $$

Check: $\lambda_1 + \lambda_2 = 15 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = 50 = \det A$.

Eigenvectors

For $\lambda = 10$: $(A - (10)I)\mathbf{v_1} = \mathbf 0$, that is

$$ \begin{pmatrix} 1 & -2 \\ 3 & -6 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 1v_1 - 2v_2 = 0\quad (\text{row 1}) \\ 3v_1 - 6v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $3$ times row 1) and express a single condition. Row 1 reads $1v_1 + (-2)v_2 = 0$, so $v_1 = 2\,v_2$: every eigenvector is a multiple of $(2, 1)^T$. The smallest integer representative is

$$ \mathbf{v_1} = \begin{pmatrix} 2 \\ 1 \end{pmatrix} $$

Check: $A\mathbf{v_1} = \begin{pmatrix} 20 \\ 10 \end{pmatrix} = 10\mathbf{v_1}$.

For $\lambda = 5$: $(A - (5)I)\mathbf{v_2} = \mathbf 0$, that is

$$ \begin{pmatrix} 6 & -2 \\ 3 & -1 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 6v_1 - 2v_2 = 0\quad (\text{row 1}) \\ 3v_1 - 1v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $\frac{1}{2}$ times row 1) and express a single condition. Row 1 reads $6v_1 + (-2)v_2 = 0$, so $v_1 = \frac{1}{3}\,v_2$: every eigenvector is a multiple of $(\frac{1}{3}, 1)^T$. The smallest integer representative is

$$ \mathbf{v_2} = \begin{pmatrix} 1 \\ 3 \end{pmatrix} $$

Check: $A\mathbf{v_2} = \begin{pmatrix} 5 \\ 15 \end{pmatrix} = 5\mathbf{v_2}$.

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 2 & 1 \\ 1 & 3 \end{pmatrix} $$

$\det Q = (2)(3) - (1)(1) = 5$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{5}\begin{pmatrix} 3 & -1 \\ -1 & 2 \end{pmatrix} = \begin{pmatrix} \frac{3}{5} & - \frac{1}{5} \\ - \frac{1}{5} & \frac{2}{5} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 11 & -2 \\ 3 & 4 \end{pmatrix}\begin{pmatrix} 2 \\ 1 \end{pmatrix} = \begin{pmatrix} 20 \\ 10 \end{pmatrix} = 10\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 11 & -2 \\ 3 & 4 \end{pmatrix}\begin{pmatrix} 1 \\ 3 \end{pmatrix} = \begin{pmatrix} 5 \\ 15 \end{pmatrix} = 5\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 20 & 5 \\ 10 & 15 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{3}{5}\cdot20 + (- \frac{1}{5})\cdot10 = 10 \\ (Q^{-1}AQ)_{12} &= \frac{3}{5}\cdot5 + (- \frac{1}{5})\cdot15 = 0 \\ (Q^{-1}AQ)_{21} &= (- \frac{1}{5})\cdot20 + \frac{2}{5}\cdot10 = 0 \\ (Q^{-1}AQ)_{22} &= (- \frac{1}{5})\cdot5 + \frac{2}{5}\cdot15 = 5 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 10 & 0 \\ 0 & 5 \end{pmatrix} = \Lambda = \begin{pmatrix} 10 & 0 \\ 0 & 5 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 10z_1, \qquad \dot z_2 = 5z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{10t}, \qquad z_2(t) = c_2 e^{5t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{10t}\begin{pmatrix} 2 \\ 1 \end{pmatrix} + c_2 e^{5t}\begin{pmatrix} 1 \\ 3 \end{pmatrix} $$

$$ x(t) = 2 c_{1} e^{10 t} + c_{2} e^{5 t}, \qquad y(t) = c_{1} e^{10 t} + 3 c_{2} e^{5 t} $$

The decoupling transformation

In the original variables $\mathbf x = (x, y)^T$ the new coordinates are $\mathbf z = Q^{-1}\mathbf x$:

$$ z_1 = \frac{3 x}{5} - \frac{y}{5}, \qquad z_2 = - \frac{x}{5} + \frac{2 y}{5} $$

Phase portrait

$\lambda_1 = 10$ and $\lambda_2 = 5$ are both positive: the origin is a unstable node (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{10t}$ and $z_2 = c_2e^{5t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{\frac{1}{2}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = \frac{1}{2}$ is positive, so the orbits are parabola-like curves through the origin. At the origin they are tangent to the eigenvector of the eigenvalue of smaller modulus, $|\lambda| = 5$, i.e. to the line through $(1, 3)$ (they leave from there (the limit $t \to -\infty$)), and far from the origin they become parallel to the other eigenvector (the fast direction). The arrows point away from the origin on both eigenlines.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (2, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, 3)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 10\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = 5\mathbf v_2$ points outward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (unstable node): invariant axes and orbits $z_2 = K|z_1|^{\frac{1}{2}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (unstable node): dashed lines are the eigenlines $\mathbf v_1 = (2, 1)$ and $\mathbf v_2 = (1, 3)$.
Open the figure full size
(iv) $A = \begin{pmatrix} 1 & 20 \\ 40 & -19 \end{pmatrix}$

Characteristic equation $\det(A - \lambda I) = 0$ (Topic 4, Eq. 27.8):

$$ \det(A - \lambda I) = \begin{vmatrix} 1 - \lambda & 20 \\ 40 & -19 - \lambda \end{vmatrix} = (1 - \lambda)(-19 - \lambda) - (20)(40) $$

$$ = \lambda^2 - (1 + (-19))\lambda + (1\cdot(-19) - 20\cdot40) = \lambda^2 - (-18)\lambda + (-819) = \lambda^{2} + 18 \lambda - 819 $$

$\operatorname{tr}A = -18$, $\det A = -819$, discriminant $\Delta = (\operatorname{tr}A)^2 - 4\det A = 324 - (-3276) = 3600$.

$$ \lambda = \frac{-18 \pm \sqrt{3600}}{2} = \frac{-18 \pm 60}{2} \quad\Longrightarrow\quad \lambda_1 = 21,\ \lambda_2 = -39 $$

Check: $\lambda_1 + \lambda_2 = -18 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = -819 = \det A$.

Eigenvectors

For $\lambda = 21$: $(A - (21)I)\mathbf{v_1} = \mathbf 0$, that is

$$ \begin{pmatrix} -20 & 20 \\ 40 & -40 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} -20v_1 + 20v_2 = 0\quad (\text{row 1}) \\ 40v_1 - 40v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $-2$ times row 1) and express a single condition. Row 1 reads $-20v_1 + 20v_2 = 0$, so $v_1 = v_2$: every eigenvector is a multiple of $(1, 1)^T$. The smallest integer representative is

$$ \mathbf{v_1} = \begin{pmatrix} 1 \\ 1 \end{pmatrix} $$

Check: $A\mathbf{v_1} = \begin{pmatrix} 21 \\ 21 \end{pmatrix} = 21\mathbf{v_1}$.

For $\lambda = -39$: $(A - (-39)I)\mathbf{v_2} = \mathbf 0$, that is

$$ \begin{pmatrix} 40 & 20 \\ 40 & 20 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 40v_1 + 20v_2 = 0\quad (\text{row 1}) \\ 40v_1 + 20v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 is identical to row 1) and express a single condition. Row 1 reads $40v_1 + 20v_2 = 0$, so $v_1 = - \frac{1}{2}\,v_2$: every eigenvector is a multiple of $(- \frac{1}{2}, 1)^T$. The smallest integer representative is

$$ \mathbf{v_2} = \begin{pmatrix} 1 \\ -2 \end{pmatrix} $$

Check: $A\mathbf{v_2} = \begin{pmatrix} -39 \\ 78 \end{pmatrix} = -39\mathbf{v_2}$.

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 1 \\ 1 & -2 \end{pmatrix} $$

$\det Q = (1)(-2) - (1)(1) = -3$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-3}\begin{pmatrix} -2 & -1 \\ -1 & 1 \end{pmatrix} = \begin{pmatrix} \frac{2}{3} & \frac{1}{3} \\ \frac{1}{3} & - \frac{1}{3} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} 1 & 20 \\ 40 & -19 \end{pmatrix}\begin{pmatrix} 1 \\ 1 \end{pmatrix} = \begin{pmatrix} 21 \\ 21 \end{pmatrix} = 21\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} 1 & 20 \\ 40 & -19 \end{pmatrix}\begin{pmatrix} 1 \\ -2 \end{pmatrix} = \begin{pmatrix} -39 \\ 78 \end{pmatrix} = (-39)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 21 & -39 \\ 21 & 78 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{2}{3}\cdot21 + \frac{1}{3}\cdot21 = 21 \\ (Q^{-1}AQ)_{12} &= \frac{2}{3}\cdot(-39) + \frac{1}{3}\cdot78 = 0 \\ (Q^{-1}AQ)_{21} &= \frac{1}{3}\cdot21 + (- \frac{1}{3})\cdot21 = 0 \\ (Q^{-1}AQ)_{22} &= \frac{1}{3}\cdot(-39) + (- \frac{1}{3})\cdot78 = -39 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 21 & 0 \\ 0 & -39 \end{pmatrix} = \Lambda = \begin{pmatrix} 21 & 0 \\ 0 & -39 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 21z_1, \qquad \dot z_2 = -39z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{21t}, \qquad z_2(t) = c_2 e^{-39t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{21t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2 e^{-39t}\begin{pmatrix} 1 \\ -2 \end{pmatrix} $$

$$ x(t) = c_{1} e^{21 t} + c_{2} e^{-39 t}, \qquad y(t) = c_{1} e^{21 t} - 2 c_{2} e^{-39 t} $$

The decoupling transformation

In the original variables $\mathbf x = (x, y)^T$ the new coordinates are $\mathbf z = Q^{-1}\mathbf x$:

$$ z_1 = \frac{2 x}{3} + \frac{y}{3}, \qquad z_2 = \frac{x}{3} - \frac{y}{3} $$

Phase portrait

$\lambda_1 = 21$ and $\lambda_2 = -39$ are of opposite signs: the origin is a saddle (Robinson Ch. 28, Topic 4 Sec. 6 and 9).

In the canonical $\mathbf z$-plane, eliminating $t$ from $z_1 = c_1e^{21t}$ and $z_2 = c_2e^{-39t}$ (for $c_1 \neq 0$): $|z_1|/|c_1| = e^{\lambda_1t}$, so $e^{t} = (|z_1|/|c_1|)^{1/\lambda_1}$ and $|z_2| = |c_2|e^{\lambda_2t} = |c_2|\,(|z_1|/|c_1|)^{\lambda_2/\lambda_1}$, i.e. $z_2 = K\,|z_1|^{\lambda_2/\lambda_1} = K\,|z_1|^{- \frac{13}{7}}$ with $K = c_2|c_1|^{-\lambda_2/\lambda_1}$ (equivalently, integrate the separable equation $\frac{dz_2}{dz_1} = \frac{\lambda_2z_2}{\lambda_1z_1}$ by the chain rule, as in Exercise 10). The axes $z_1 = 0$ and $z_2 = 0$ are invariant: a solution that starts on an axis stays on it.

The exponent $\lambda_2/\lambda_1 = - \frac{13}{7}$ is negative, so the orbits are hyperbola-like. Only solutions starting on the eigenline of the negative eigenvalue, the line through $(1, -2)$ (the stable manifold), reach the origin as $t \to +\infty$; the line through $(1, 1)$ (positive eigenvalue) is the unstable manifold. All other orbits come in close to the stable line, then leave along the unstable one.

The map $\mathbf x = Q\mathbf z$ sends the $z_1$-axis to the line through $\mathbf v_1 = (1, 1)$ and the $z_2$-axis to the line through $\mathbf v_2 = (1, -2)$. Straight lines stay straight and the origin stays fixed, so the $\mathbf x$-portrait is the canonical one sheared onto the eigenlines. On $\mathbf v_1$: $\dot{\mathbf x} = A\mathbf v_1 = 21\mathbf v_1$ points outward; on $\mathbf v_2$: $\dot{\mathbf x} = -39\mathbf v_2$ points inward.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane (saddle): invariant axes and orbits $z_2 = K|z_1|^{- \frac{13}{7}}$.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane (saddle): dashed lines are the eigenlines $\mathbf v_1 = (1, 1)$ and $\mathbf v_2 = (1, -2)$.
Open the figure full size
Interpretation and checks
  • Each $Q^{-1}AQ$ was multiplied out entry by entry above and equals $\Lambda$ exactly; the eigenvalues satisfy $\lambda_1 + \lambda_2 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = \det A$.
  • (i) $15$ and $-6$: saddle; (ii) $2$ and $-3$ (lower-triangular matrix: the eigenvalues are the diagonal entries): saddle; (iii) $10$ and $5$: unstable node, tangent at the origin to the eigenvector of $5$; (iv) $21$ and $-39$: saddle with very different rates.
  • (i) is the homogeneous part of Exercise 7(vii) and (ii) shows that for a triangular matrix the eigenvalues can be read off the diagonal (Topic 4 Examples 27.2 and 27.3).
  • The change of coordinates is not the efficient way to solve the system (the general solution follows directly from the eigenpairs), but it is what justifies the portrait: it reduces every case to the uncoupled one (Robinson Sec. 28.2).
  • Dimensional check: with $t$ in seconds the eigenvalues have units s$^{-1}$, and $Q$ is dimensionless (it mixes coordinates of equal units).
Exercise 10

Trajectories of an Uncoupled System by the Chain Rule

Original statement · probls_ch4-i_2627_v2.pdf, p. 3
Chain Rule • Separable Equation • Orbit Equation
Statement

Using the chain rule for $y = y(x(t))$,

$$ \frac{dy}{dt} = \frac{dy}{dx}\frac{dx}{dt} \;\Longrightarrow\; \frac{dy}{dx} = \frac{\dfrac{dy}{dt}}{\dfrac{dx}{dt}}. $$

Therefore, if

$$ \frac{dx}{dt} = \lambda_1x, \qquad \frac{dy}{dt} = \lambda_2y $$

we have

$$ \frac{dy}{dx} = \frac{\lambda_2y}{\lambda_1x}. $$

Solve the latter equation to determine the curves traced out by trajectories of the system of equations for $x(t)$ and $y(t)$.

Setup and hypotheses

Real constants $\lambda_1, \lambda_2$. The chain rule is valid where $x(t)$ is a differentiable function with $\dot x \neq 0$, so that $t \mapsto x$ is locally invertible and $y$ can be regarded as a function of $x$. Here $\dot x = \lambda_1x \neq 0$ requires $\lambda_1 \neq 0$ and $x \neq 0$. The cases $x = 0$ (the $y$-axis), $y = 0$ (the $x$-axis) and $\lambda_1 = 0$ are treated separately at the end.

Formulation

The orbit equation $\frac{dy}{dx} = \frac{\lambda_2}{\lambda_1}\frac{y}{x}$ has no explicit $t$ and is separable: all the $y$ on the left and all the $x$ on the right.

Method of the class notes [Notes, p. 9]

This is the step of the notes that turns Case 1 into a portrait. From $z_1 = C_1e^{\lambda_1t}$ and $z_2 = C_2e^{\lambda_2t}$ they eliminate $t$ (for $C_1 > 0$):

$$ z_2 = C_2\big(e^{\lambda_1t}\big)^{\lambda_2/\lambda_1} = C_2C_1^{-\lambda_2/\lambda_1}\big(C_1e^{\lambda_1t}\big)^{\lambda_2/\lambda_1} = \tilde C\,z_1^{\lambda_2/\lambda_1}. $$

The chain-rule route below reaches the same curves, $y = \tilde C\,\lvert x\rvert^{\lambda_2/\lambda_1}$ on each half-plane, without solving for $t$.

Derivation

Integration

Separate variables (for $y \neq 0$, $x \neq 0$) and integrate both sides with their antiderivatives, $\int\frac{dy}{y} = \ln|y|$:

$$ \frac{dy}{y} = \frac{\lambda_2}{\lambda_1}\frac{dx}{x} \;\Longrightarrow\; \ln|y| = \frac{\lambda_2}{\lambda_1}\ln|x| + C $$

Exponentiate and absorb $e^{C}$ in a constant $K = \pm e^{C}$:

$$ |y| = e^{C}|x|^{\lambda_2/\lambda_1} \;\Longrightarrow\; \boxed{y = K\,|x|^{\lambda_2/\lambda_1}}, \qquad K \neq 0 $$

Checking with the explicit solution

The equations are solved directly by $x = x_0e^{\lambda_1t}$, $y = y_0e^{\lambda_2t}$ (separation of variables, $\frac{d}{dt}(e^{-\lambda t}u) = 0$). For $x_0 \neq 0$, $|x|/|x_0| = e^{\lambda_1t}$ so $e^{t} = (|x|/|x_0|)^{1/\lambda_1}$ and

$$ y = y_0\left(e^{\lambda_1t}\right)^{\lambda_2/\lambda_1} = y_0\left(\frac{|x|}{|x_0|}\right)^{\lambda_2/\lambda_1} = \underbrace{y_0|x_0|^{-\lambda_2/\lambda_1}}_{K}|x|^{\lambda_2/\lambda_1} $$

which is the same family, with $K = y_0|x_0|^{-\lambda_2/\lambda_1}$ (the sign of $y$ is the sign of $y_0$, preserved since $y = y_0e^{\lambda_2t}$).

The axes

If $x_0 = 0$ then $x(t) \equiv 0$ and the orbit is on the $y$-axis (with $y = y_0e^{\lambda_2t}$); if $y_0 = 0$ then $y \equiv 0$ (that is $K = 0$) and the orbit lies on the $x$-axis. The origin is the equilibrium. If $\lambda_1 = 0$, $x$ is constant and the orbits are vertical lines.

Shapes of the curves

Exponent $\lambda_2/\lambda_1$Curves $y = K\vert x\vert ^{r}$Origin
$r > 1$ (same sign, $\vert \lambda_2\vert > \vert \lambda_1\vert $)parabola-like, tangent to the $x$-axis at $0$node (stable if $\lambda_i < 0$, unstable if $> 0$)
$r = 1$ ($\lambda_1 = \lambda_2$)straight lines $y = Kx$ (rays)star
$0 < r < 1$tangent to the $y$-axis at $0$node
$r < 0$ (opposite signs)hyperbola-like, $y\vert x\vert ^{-r} = K$, asymptotic to the axessaddle
$r = -1$rectangular hyperbolas $xy = \pm K$saddle
Interpretation and checks
  • This is the derivation behind the canonical portraits of Topic 4 Sec. 6.1: replace $(x, y)$ by $(z_1, z_2)$ and $\lambda_1, \lambda_2$ by the two eigenvalues to get $z_2 = K|z_1|^{\lambda_2/\lambda_1}$ (Robinson Eq. 28.12).
  • Examples: $\lambda_1 = 1$, $\lambda_2 = 2$: $y = Kx^2$ (parabolas); $\lambda_1 = 1$, $\lambda_2 = -1$: $y = K/x$ (hyperbolas).
  • Direction along the orbits is not given by the orbit equation: it is read from $\dot x = \lambda_1x$, $\dot y = \lambda_2y$ (away from $0$ along an axis if $\lambda > 0$, towards it if $\lambda < 0$).
  • Differentiating the answer: $\frac{dy}{dx} = K\frac{\lambda_2}{\lambda_1}|x|^{\lambda_2/\lambda_1 - 1}\operatorname{sign}(x) = \frac{\lambda_2}{\lambda_1}\frac{y}{x}$, which recovers the ODE (using $|x|^{r}/x = |x|^{r-1}\operatorname{sign}x$).
Exercise 11

A Zero Eigenvalue: Lines of Equilibria

Original statement · probls_ch4-i_2627_v2.pdf, p. 4
Zero Eigenvalue • Line of Equilibria • Parallel Orbits
Statement

Suppose that $A$ has eigenvalues $\lambda_1 = 0$, $\lambda_2 \neq 0$, with eigenvectors $\mathbf v_1$ and $\mathbf v_2$, respectively.

  • (i) Write down the general solution of $\dot{\mathbf x} = A\mathbf x$.
  • (ii) After changing to a coordinate system $\tilde{\mathbf x}$ referred to the eigenvectors, we obtain $\dfrac{d\tilde{\mathbf x}}{dt} = \begin{pmatrix} 0 & 0 \\ 0 & \lambda_2 \end{pmatrix}\tilde{\mathbf x}$. By solving these equations draw the phase portrait in the $\tilde{\mathbf x}$ system, and sketch the phase portrait for the original coordinates $\mathbf x$.
  • (iii) Draw the phase portrait for the equation $\dfrac{d\mathbf x}{dt} = \begin{pmatrix} -2 & 2 \\ 1 & -1 \end{pmatrix}\mathbf x$.
Setup and hypotheses

$A$ is $2\times 2$ real with eigenvalues $0$ and $\lambda_2 \neq 0$ (both real: $\lambda_1 = 0$ is real and $\lambda_2 = \operatorname{tr}A$ is real because $A$ is real; being distinct, they have independent eigenvectors; $\det A = 0$). $A\mathbf v_1 = \mathbf 0$, $A\mathbf v_2 = \lambda_2\mathbf v_2$. Because $\det A = 0$ the origin is not an isolated equilibrium: every point of $\ker A = \operatorname{span}\{\mathbf v_1\}$ is an equilibrium ($A\mathbf x = \mathbf 0$).

Formulation

Use the general result for distinct eigenvalues, $\mathbf x = c_1e^{\lambda_1t}\mathbf v_1 + c_2e^{\lambda_2t}\mathbf v_2$ (Robinson Eq. 28.1), valid whenever the eigenvalues differ, in particular if one of them is $0$; and $Q = [\mathbf v_1\ \mathbf v_2]$, $\mathbf z = Q^{-1}\mathbf x$.

Method of the class notes [Notes, p. 9]

Special case $\lambda_1 = 0$ of the notes: $A\mathbf v_1 = \mathbf 0$, so the line spanned by $\mathbf v_1$ is made of equilibria, and $\mathbf x = C_1\mathbf v_1 + C_2e^{\lambda_2t}\mathbf v_2$: straight orbits parallel to $\mathbf v_2$.

Reading of the handwriting: the notes write $\mathbf x(t) = C_2\mathbf v_2e^{\lambda_2t}$, which drops the line of equilibria $C_1\mathbf v_1$. The general solution needs both terms, since it must match any $\mathbf x(0)$.

Derivation
(i) General solution

With $\lambda_1 = 0$, $e^{\lambda_1t} = e^{0} = 1$:

$$ \mathbf x(t) = c_1\mathbf v_1 + c_2e^{\lambda_2t}\mathbf v_2 $$

Check: $\dot{\mathbf x} = c_2\lambda_2e^{\lambda_2t}\mathbf v_2$ and $A\mathbf x = c_1A\mathbf v_1 + c_2e^{\lambda_2t}A\mathbf v_2 = \mathbf 0 + c_2\lambda_2e^{\lambda_2t}\mathbf v_2$. They agree. The constant vector $c_1\mathbf v_1$ is an equilibrium and $c_2e^{\lambda_2t}\mathbf v_2$ the time-dependent part.

(ii) Canonical phase portrait and its image

In the coordinates $\tilde{\mathbf x} = (\tilde x, \tilde y)^T$ referred to $\mathbf v_1, \mathbf v_2$ the system is $\dot{\tilde x} = 0$, $\dot{\tilde y} = \lambda_2\tilde y$. The first gives $\tilde x = c_1$ (constant); the second is separable: $\frac{d\tilde y}{\tilde y} = \lambda_2\,dt$, $\ln|\tilde y| = \lambda_2t + C$, so $\tilde y = c_2e^{\lambda_2t}$.

$$ \tilde x(t) = c_1, \qquad \tilde y(t) = c_2e^{\lambda_2t} $$

  • Orbits: the vertical lines $\tilde x = c_1$, because $\tilde x$ never changes. Each such line is a pair of half-lines (for $\tilde y > 0$ and $\tilde y < 0$) and the point $(c_1, 0)$ is an equilibrium.
  • The $\tilde x$-axis ($\tilde y = 0$) consists entirely of equilibria (it is the image of $\ker A$).
  • Direction: if $\lambda_2 < 0$, $|\tilde y| = |c_2|e^{\lambda_2t}$ decreases: the arrows on every vertical line point towards the $\tilde x$-axis, which they approach as $t \to +\infty$ without reaching it. If $\lambda_2 > 0$ the arrows point away from the axis. The sign of $\dot{\tilde y} = \lambda_2\tilde y$ gives the direction at once.
  • Original coordinates: $\mathbf x = Q\tilde{\mathbf x} = \tilde x\mathbf v_1 + \tilde y\mathbf v_2$. The $\tilde x$-axis goes to the line $\operatorname{span}\{\mathbf v_1\}$ of equilibria and the vertical lines go to the lines $c_1\mathbf v_1 + s\mathbf v_2$, $s \in \mathbb R$, parallel to $\mathbf v_2$, drawn through the points of the line of equilibria. Straight lines remain straight: the picture is the canonical one sheared, with orbits along $\mathbf v_2$ towards/away from the line of equilibria.
x̃ ỹ z1 z2 z-plane, λ₂ < 0
Canonical $\tilde{\mathbf x}$-plane, $\lambda_2 < 0$: the $\tilde x$-axis (green) is a line of equilibria and the vertical orbits approach it.
Open the figure full size
x̃ ỹ z1 z2 z-plane, λ₂ > 0
Canonical $\tilde{\mathbf x}$-plane, $\lambda_2 > 0$: the vertical orbits leave the line of equilibria.
Open the figure full size
x y v1 v2 x-plane, example
Original $(x, y)$-plane for the example $\mathbf v_1 = (1, 0)$, $\mathbf v_2 = (1, 1)$, $\lambda_2 = -2$: equilibria on the line of $\mathbf v_1$, orbits parallel to $\mathbf v_2$.
Open the figure full size

(For the sketch in the original plane, take the example $\mathbf v_1 = (1, 0)$, $\mathbf v_2 = (1, 1)$, $\lambda_2 = -2$, whose matrix is $A = Q\Lambda Q^{-1} = \begin{pmatrix} 0 & -2 \\ 0 & -2 \end{pmatrix}$; any other eigen-data give the same picture up to a linear change of the axes.)

(iii) The matrix $\begin{pmatrix} -2 & 2 \\ 1 & -1 \end{pmatrix}$

Characteristic equation $\det(A - \lambda I) = 0$ (Topic 4, Eq. 27.8):

$$ \det(A - \lambda I) = \begin{vmatrix} -2 - \lambda & 2 \\ 1 & -1 - \lambda \end{vmatrix} = (-2 - \lambda)(-1 - \lambda) - (2)(1) $$

$$ = \lambda^2 - (-2 + (-1))\lambda + (-2\cdot(-1) - 2\cdot1) = \lambda^2 - (-3)\lambda + (0) = \lambda^{2} + 3 \lambda $$

$\operatorname{tr}A = -3$, $\det A = 0$, discriminant $\Delta = (\operatorname{tr}A)^2 - 4\det A = 9 - 0 = 9$.

$$ \lambda = \frac{-3 \pm \sqrt{9}}{2} = \frac{-3 \pm 3}{2} \quad\Longrightarrow\quad \lambda_1 = 0,\ \lambda_2 = -3 $$

Check: $\lambda_1 + \lambda_2 = -3 = \operatorname{tr}A$ and $\lambda_1\lambda_2 = 0 = \det A$.

Eigenvectors

For $\lambda = 0$: $(A - (0)I)\mathbf{v_1} = \mathbf 0$, that is

$$ \begin{pmatrix} -2 & 2 \\ 1 & -1 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} -2v_1 + 2v_2 = 0\quad (\text{row 1}) \\ 1v_1 - 1v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 equals $- \frac{1}{2}$ times row 1) and express a single condition. Row 1 reads $-2v_1 + 2v_2 = 0$, so $v_1 = v_2$: every eigenvector is a multiple of $(1, 1)^T$. The smallest integer representative is

$$ \mathbf{v_1} = \begin{pmatrix} 1 \\ 1 \end{pmatrix} $$

Check: $A\mathbf{v_1} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} = 0\mathbf{v_1}$.

For $\lambda = -3$: $(A - (-3)I)\mathbf{v_2} = \mathbf 0$, that is

$$ \begin{pmatrix} 1 & 2 \\ 1 & 2 \end{pmatrix}\begin{pmatrix} v_1 \\ v_2 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} $$

$$ \begin{aligned} 1v_1 + 2v_2 = 0\quad (\text{row 1}) \\ 1v_1 + 2v_2 = 0\quad (\text{row 2}) \end{aligned} $$

The coefficient matrix is singular ($\det(A - \lambda I) = 0$), so the rows are proportional (row 2 is identical to row 1) and express a single condition. Row 1 reads $1v_1 + 2v_2 = 0$, so $v_1 = -2\,v_2$: every eigenvector is a multiple of $(-2, 1)^T$. The smallest integer representative is

$$ \mathbf{v_2} = \begin{pmatrix} 2 \\ -1 \end{pmatrix} $$

Check: $A\mathbf{v_2} = \begin{pmatrix} -6 \\ 3 \end{pmatrix} = -3\mathbf{v_2}$.

Change of coordinates

The eigenvectors of distinct eigenvalues are linearly independent (Topic 4, Appendix B argument), so they form a basis. Put them as columns, $Q = [\mathbf v_1 \mid \mathbf v_2]$, and define the new coordinates by $\mathbf z = Q^{-1}\mathbf x$, i.e. $\mathbf x = Q\mathbf z$:

$$ Q = \begin{pmatrix} 1 & 2 \\ 1 & -1 \end{pmatrix} $$

$\det Q = (1)(-1) - (2)(1) = -3$, which is not zero, so $Q$ is invertible. For a $2\times 2$ matrix $\begin{pmatrix} a & b \\ c & d \end{pmatrix}^{-1} = \frac{1}{ad - bc}\begin{pmatrix} d & -b \\ -c & a \end{pmatrix}$ (Robinson, Appendix B):

$$ Q^{-1} = \frac{1}{-3}\begin{pmatrix} -1 & -2 \\ -1 & 1 \end{pmatrix} = \begin{pmatrix} \frac{1}{3} & \frac{2}{3} \\ \frac{1}{3} & - \frac{1}{3} \end{pmatrix} $$

Since $Q$ is constant, $\dot{\mathbf z} = Q^{-1}\dot{\mathbf x} = Q^{-1}A\mathbf x = (Q^{-1}AQ)\mathbf z$. Compute $AQ$ one column at a time, with $A\mathbf v_j = \lambda_j\mathbf v_j$:

$$ A\mathbf v_1 = \begin{pmatrix} -2 & 2 \\ 1 & -1 \end{pmatrix}\begin{pmatrix} 1 \\ 1 \end{pmatrix} = \begin{pmatrix} 0 \\ 0 \end{pmatrix} = 0\mathbf v_1, \qquad A\mathbf v_2 = \begin{pmatrix} -2 & 2 \\ 1 & -1 \end{pmatrix}\begin{pmatrix} 2 \\ -1 \end{pmatrix} = \begin{pmatrix} -6 \\ 3 \end{pmatrix} = (-3)\mathbf v_2 $$

$$ AQ = [A\mathbf v_1 \mid A\mathbf v_2] = \begin{pmatrix} 0 & -6 \\ 0 & 3 \end{pmatrix} $$

Now multiply by $Q^{-1}$ on the left, entry by entry:

$$ \begin{aligned} (Q^{-1}AQ)_{11} &= \frac{1}{3}\cdot0 + \frac{2}{3}\cdot0 = 0 \\ (Q^{-1}AQ)_{12} &= \frac{1}{3}\cdot(-6) + \frac{2}{3}\cdot3 = 0 \\ (Q^{-1}AQ)_{21} &= \frac{1}{3}\cdot0 + (- \frac{1}{3})\cdot0 = 0 \\ (Q^{-1}AQ)_{22} &= \frac{1}{3}\cdot(-6) + (- \frac{1}{3})\cdot3 = -3 \end{aligned} $$

$$ Q^{-1}AQ = \begin{pmatrix} 0 & 0 \\ 0 & -3 \end{pmatrix} = \Lambda = \begin{pmatrix} 0 & 0 \\ 0 & -3 \end{pmatrix} $$

The transformation has decoupled the system (the matrix is diagonal), as claimed.

Solution in the canonical plane

$$ \dot z_1 = 0, \qquad \dot z_2 = -3z_2 $$

Each equation is separable. If $z_i \neq 0$, $\frac{dz_i}{z_i} = \lambda_i\,dt$ and integrating, $\ln|z_i| = \lambda_i t + C_i$, so $z_i = c_i e^{\lambda_i t}$ with $c_i = \pm e^{C_i}$ (the solution $z_i \equiv 0$ corresponds to $c_i = 0$). Equivalently, $\frac{d}{dt}(e^{-\lambda_i t}z_i) = 0$.

$$ z_1(t) = c_1 e^{0t}, \qquad z_2(t) = c_2 e^{-3t} $$

Back to the original coordinates

$\mathbf x = Q\mathbf z = z_1\mathbf v_1 + z_2\mathbf v_2$, so the general solution is

$$ \mathbf x(t) = c_1 e^{0t}\begin{pmatrix} 1 \\ 1 \end{pmatrix} + c_2 e^{-3t}\begin{pmatrix} 2 \\ -1 \end{pmatrix} $$

$$ x(t) = c_{1} + 2 c_{2} e^{-3 t}, \qquad y(t) = c_{1} - c_{2} e^{-3 t} $$

Phase portrait

One eigenvalue is zero and the other is $\lambda_2 = -3 \neq 0$. In the canonical plane $\dot z_1 = 0$, so $z_1$ is constant along every solution (Topic 4, case $\lambda = 0$).

Every point of the line through $\mathbf v_1 = (1, 1)$ is an equilibrium, because $A\mathbf v_1 = \mathbf 0$. Every other solution moves on a straight line parallel to $\mathbf v_2 = (2, -1)$ with $\mathbf x - \mathbf x_{eq} \propto e^{-3t}$: it approaches the line of equilibria as $t \to +\infty$ (the equilibria are stable but each orbit ends at a different point). The arrows point towards the line of equilibria.

z1 z2 z1 z2 z-plane
Canonical $\mathbf z$-plane: the line of equilibria (green) and the trajectories parallel to the other axis.
Open the figure full size
x y v1 v2 x-plane
Original $(x, y)$-plane: equilibria on the line through $(1, 1)$ (green); trajectories parallel to $(2, -1)$, towards that line.
Open the figure full size
Interpretation and checks
  • (iii): $\operatorname{tr}A = -3$ and $\det A = (-2)(-1) - 2\cdot 1 = 0$, so the eigenvalues are $0$ and $-3$: the general solution is $\mathbf x = c_1(1, 1)^T + c_2e^{-3t}(2, -1)^T$, every orbit converges to the equilibrium $c_1(1, 1)^T$, i.e. $\mathbf x(t) \to \mathbf x_\infty$ on the line $y = x$ along the direction $(2, -1)$.
  • Explicit limit for $\mathbf x(0) = (x_0, y_0)$: $\mathbf x_\infty = c_1(1, 1)^T$ with $c_1 = (x_0 + 2y_0)/3$ (from $Q^{-1}$ computed above, first row $(\tfrac13, \tfrac23)$). Example: $\mathbf x(0) = (2, -1)$ gives $c_1 = 0$ and the orbit is the line $\mathbf x = e^{-3t}(2, -1)$ going to the origin.
  • The origin is no longer an isolated equilibrium: the node/saddle/spiral classification of Topic 4 Sec. 9 assumes $\det A \neq 0$ (see the trace-determinant discussion), and here $\det A = 0$, which is the borderline between a stable node ($\lambda_2 < 0$) and a saddle.
  • Conserved quantity: the combination $\tilde x = c_1$ is a first integral of the motion ($x + 2y$ in (iii), up to the factor $\tfrac13$): $\frac{d}{dt}(x + 2y) = (-2x + 2y) + 2(x - y) = 0$ $\checkmark$.