Phase portraits

How can every solution of a two-dimensional linear system be seen at once, and how do the trace and determinant of its matrix decide what the picture looks like?

Where you are 6 in Systems and transforms

Why this exists

Linear systems of differential equations gives a formula for every solution of \mathbf{x}' = A\mathbf{x}, but hides what a model asks: what does every solution do in the long run, and does a small disturbance die out? For two unknowns the answer is one picture, the phase portrait, whose shape the trace and determinant of A decide. It extends the phase line of Autonomous equations and stability to the plane.

Definition

In plain words: draw the plane of the two unknowns and, through each point, the path a solution follows from there, with arrows for the direction of time.

Let A be a constant real 2 \times 2 matrix and \mathbf{x} = (x_1, x_2) a function of t. Every solution of \mathbf{x}' = A\mathbf{x} is defined for all real t (Result 2 of Linear systems of differential equations). The plane with axes x_1, x_2 is the phase plane. The trajectory (or orbit) of a solution is the curve \{\mathbf{x}(t) : t \in \mathbb{R}\}, oriented by increasing t; a phase portrait draws enough trajectories, with arrows, to show how all of them run. A point \mathbf{p} with A\mathbf{p} = \mathbf{0} is an equilibrium point (or critical point): \mathbf{x}(t) = \mathbf{p} is a solution.

With |\cdot| the length of Dot product, length and orthogonality in n dimensions, the origin is stable if for every \varepsilon > 0 there is \delta > 0 such that every solution with |\mathbf{x}(0)| < \delta has |\mathbf{x}(t)| < \varepsilon for all t \ge 0; asymptotically stable if it is stable and every such solution tends to \mathbf{0} as t \to \infty; unstable if it is not stable. This “stable” is weaker than the one in Autonomous equations and stability, which asked nearby solutions to tend to the equilibrium; that property is “asymptotically stable” here.

Write \tau = \operatorname{tr} A and \Delta = \det A. The eigenvalues are the roots of \lambda^2 - \tau\lambda + \Delta = 0 (The characteristic equation and finding eigenvalues), so \lambda_1 + \lambda_2 = \tau and \lambda_1\lambda_2 = \Delta. When \Delta \ne 0, the origin is named by them:

eigenvaluesthe origin is a
real, opposite signssaddle
real, distinct, same signnode
real, equal, A \ne \lambda Idegenerate node
real, equal, A = \lambda Istar
\alpha \pm \beta i with \alpha \ne 0, \beta > 0spiral
\pm\beta i with \beta > 0center

A node, degenerate node, star or spiral is a sink when its eigenvalues have negative real part and a source when positive.

Intuition

The straight-line solutions e^{\lambda t}\mathbf{v} are the skeleton: on an eigenvector line, points move straight in when \lambda < 0 and out when \lambda > 0. Any other solution is a sum of two such terms, and the dominant one sets its course. In a node both shrink (or both grow), and the one with eigenvalue nearer 0 outlasts the other, so curves arrive along the slow eigenvector. In a saddle one shrinks and one grows. Complex eigenvalues rotate and scale in disguise (Complex eigenvalues and rotation-scaling matrices), so curves wind around the origin.

node saddle spiral degenerate node center star
Figure 1One phase portrait of each type, with the eigenvector lines of real eigenvalues drawn thick. Node: \begin{bmatrix} -2 & 1 \\ 2 & -3 \end{bmatrix} (Example 1). Saddle: \begin{bmatrix} 1 & 4 \\ 2 & -1 \end{bmatrix} (Example 2). Spiral: \begin{bmatrix} -1 & -4 \\ 1 & -1 \end{bmatrix} (Example 3). Degenerate node: \begin{bmatrix} 3 & -1 \\ 1 & 1 \end{bmatrix} (Example 4). Center: \begin{bmatrix} 1 & -2 \\ 1 & -1 \end{bmatrix}. Star: -I.

Figure 2 places every matrix in the (\tau, \Delta) plane. The parabola \tau^2 = 4\Delta separates real eigenvalues (below) from complex ones (above), the line \Delta = 0 separates saddles from the rest, and the sign of \tau separates sinks from sources.

saddle sink node source node spiral sink spiral source center −6 9 trace determinant
Figure 2The trace and determinant decide the type when \Delta \ne 0. The parabola \tau^2 = 4\Delta holds the degenerate nodes and stars, and the positive \Delta-axis the centers. The dashed segment is Omar’s pontoon of Example 5, \Delta = 9, \tau = -\gamma/400; the dots are \gamma = 2000 (spiral), 2400 (on the parabola) and 3000 (node).

Results

Throughout, A = \begin{bmatrix} a & b \\ c & d \end{bmatrix} is real, with \tau and \Delta as above.

1. Trajectories do not cross. Through each point passes exactly one trajectory, and two trajectories with a point in common are the same oriented curve. If \Delta \ne 0, the origin is the only equilibrium point, and no other trajectory reaches it.

2. Change of coordinates. Let P be invertible and B = P^{-1}AP. Then \mathbf{x} solves \mathbf{x}' = A\mathbf{x} exactly when \mathbf{y} = P^{-1}\mathbf{x} solves \mathbf{y}' = B\mathbf{y}. So the portrait of A is the image of that of B under \mathbf{y} \mapsto P\mathbf{y}, and A, B have the same trace, determinant and eigenvalues.

3. Classification. Let \Delta \ne 0. The origin is a saddle if \Delta < 0; a node if \Delta > 0 and \tau^2 > 4\Delta; a degenerate node or a star if \tau^2 = 4\Delta, a star exactly when \displaystyle A = \frac{\tau}{2}I; a spiral if \tau^2 < 4\Delta and \tau \ne 0; a center if \tau = 0 < \Delta. Nodes, degenerate nodes, stars and spirals are sinks when \tau < 0 and sources when \tau > 0.

4. Real eigenvalues.

  • (a) Node. Let 0 < |\lambda_1| < |\lambda_2|, with eigenvectors \mathbf{v}_1 (the slow direction) and \mathbf{v}_2. Every trajectory off the line of \mathbf{v}_2 tends to the origin tangent to the line of \mathbf{v}_1, as t \to \infty for a sink and t \to -\infty for a source. In the other time direction it is unbounded, with direction approaching \pm\mathbf{v}_2 if it is off both eigenvector lines.
  • (b) Saddle. Let \lambda_1 > 0 > \lambda_2. Solutions on the line of \mathbf{v}_2 tend to \mathbf{0} as t \to \infty, those on the line of \mathbf{v}_1 as t \to -\infty, and every other solution is unbounded in both time directions.
  • (c) Degenerate node. Let \lambda be a double eigenvalue, A \ne \lambda I and N = A - \lambda I. Then N^2 = 0, the eigenvectors are the nonzero vectors of the line \ker N, which contains every N\mathbf{x}_0, and the solution with \mathbf{x}(0) = \mathbf{x}_0 is
\mathbf{x}(t) = e^{\lambda t}\,(\mathbf{x}_0 + tN\mathbf{x}_0).

Every trajectory off \ker N tends to the origin tangent to \ker N, as t \to \infty if \lambda < 0 and t \to -\infty if \lambda > 0.

5. Complex eigenvalues and turning.

  • (a) Let A have eigenvalues \alpha \pm \beta i, \beta > 0, let \mathbf{v} be an eigenvector for \alpha - \beta i, and P = \begin{bmatrix} \operatorname{Re}\mathbf{v} & \operatorname{Im}\mathbf{v} \end{bmatrix}. The solutions are \mathbf{x} = P\mathbf{y} with \mathbf{y}(t) = e^{\alpha t}R_{\beta t}\,\mathbf{y}(0), where R_\theta = \begin{bmatrix} \cos\theta & -\sin\theta \\ \sin\theta & \cos\theta \end{bmatrix}: in the coordinates \mathbf{y}, a solution turns counterclockwise at angular speed \beta while its length is multiplied by e^{\alpha t}. For \alpha = 0 the trajectories are ellipses traced in time \displaystyle \frac{2\pi}{\beta}; for \alpha \ne 0 they are spirals.
  • (b) For any A, write a nonzero solution as \mathbf{x} = r(\cos\theta, \sin\theta) with r, \theta continuous. Then
r^2\theta' = c\,x_1^2 + (d - a)\,x_1x_2 - b\,x_2^2.

When the eigenvalues are not real, this form is definite with the sign of c \ne 0, so trajectories turn counterclockwise when c > 0 and clockwise when c < 0.

6. Stability. Let \Delta \ne 0. The origin is asymptotically stable exactly when \tau < 0 < \Delta (every eigenvalue has negative real part), stable but not asymptotically stable exactly when \tau = 0 < \Delta (a center), and unstable otherwise.

Why it is true

Result 1. Each initial value problem \mathbf{x}' = A\mathbf{x}, \mathbf{x}(0) = \mathbf{p} has a solution. If \mathbf{y} solves, so does \mathbf{z}(t) = \mathbf{y}(t + s) (chain rule), tracing the same oriented curve. If \mathbf{x}(t_1) = \mathbf{y}(t_2), take s = t_2 - t_1: then \mathbf{z}(t_1) = \mathbf{x}(t_1), so \mathbf{z} = \mathbf{x} by uniqueness. If \Delta \ne 0, A\mathbf{p} = \mathbf{0} forces \mathbf{p} = \mathbf{0}, and a solution vanishing at one time is the zero solution by uniqueness. \square

Result 2. \mathbf{y}' = P^{-1}A\mathbf{x} = P^{-1}AP\,\mathbf{y}, and back with \mathbf{x} = P\mathbf{y}. Also \det B = \det A and \operatorname{tr}(P^{-1}(AP)) = \operatorname{tr}((AP)P^{-1}) = \operatorname{tr} A, so the characteristic polynomials agree. \square

Result 3. The discriminant is \tau^2 - 4\Delta. If \Delta < 0 it is positive: real roots with negative product. If \Delta > 0 and \tau^2 > 4\Delta, real distinct roots with positive product share the sign of their sum \tau. If \tau^2 = 4\Delta, the root \displaystyle \frac{\tau}{2} \ne 0 has eigenvectors filling the plane exactly when \displaystyle A = \frac{\tau}{2}I. If \tau^2 < 4\Delta, the roots are \displaystyle \frac{1}{2}\big(\tau \pm i\sqrt{4\Delta - \tau^2}\big). \square

Result 4. (a) For a sink, \lambda_2 < \lambda_1 < 0, and by Result 6(b) of Linear systems of differential equations, \mathbf{x} = c_1e^{\lambda_1t}\mathbf{v}_1 + c_2e^{\lambda_2t}\mathbf{v}_2. If c_1 \ne 0,

\frac{\mathbf{x}(t)}{c_1e^{\lambda_1t}} = \mathbf{v}_1 + \frac{c_2}{c_1}e^{(\lambda_2 - \lambda_1)t}\mathbf{v}_2 \to \mathbf{v}_1 \quad (t \to \infty),

so \mathbf{x}/|\mathbf{x}| \to \pm\mathbf{v}_1/|\mathbf{v}_1|. As t \to -\infty, a nonzero \mathbf{x} is unbounded, since its coordinates c_ie^{\lambda_it} in the basis \mathbf{v}_1, \mathbf{v}_2 are; if c_2 \ne 0, dividing by c_2e^{\lambda_2t} gives the limit \mathbf{v}_2. For a source, \mathbf{x}(-t) solves \mathbf{z}' = -A\mathbf{z}, a sink. (b) The same computation with \lambda_2 < 0 < \lambda_1.

(c) Every 2 \times 2 matrix satisfies A^2 - \tau A + \Delta I = 0 (Cayley-Hamilton in this size): A^2 - \tau A has diagonal entries a^2 + bc - a(a + d) = -\Delta and d^2 + bc - d(a + d) = -\Delta, and off-diagonal entries b(a + d) - \tau b = 0 and c(a + d) - \tau c = 0. A double root has \tau = 2\lambda and \Delta = \lambda^2, so N^2 = A^2 - 2\lambda A + \lambda^2I = 0. As N \ne 0, its kernel is a line, and N(N\mathbf{x}_0) = \mathbf{0} puts every N\mathbf{x}_0 in it. Using N^2 = 0 and \lambda I + N = A,

\begin{aligned} \big(e^{\lambda t}(I + tN)\mathbf{x}_0\big)' &= e^{\lambda t}(\lambda I + \lambda tN + N)\mathbf{x}_0 \\ &= A\,e^{\lambda t}(I + tN)\mathbf{x}_0, \end{aligned}

and uniqueness makes this the solution. If N\mathbf{x}_0 \ne \mathbf{0}, then \mathbf{x}(t)/(te^{\lambda t}) = \mathbf{x}_0/t + N\mathbf{x}_0 \to N\mathbf{x}_0 as t \to \pm\infty, and te^{\lambda t} \to 0 at the end named. \square

Result 5. (a) By the uncovering result of Complex eigenvalues and rotation-scaling matrices, P is invertible and A = PCP^{-1} with C = \begin{bmatrix} \alpha & -\beta \\ \beta & \alpha \end{bmatrix} = \alpha I + \beta J, J = R_{\pi/2}. Since R_\theta = \cos\theta\,I + \sin\theta\,J and J^2 = -I, the derivative of R_{\beta t} is \beta JR_{\beta t}, so \big(e^{\alpha t}R_{\beta t}\mathbf{y}_0\big)' = C\,e^{\alpha t}R_{\beta t}\mathbf{y}_0; Result 2 and uniqueness give every solution. Rotations keep lengths, so |\mathbf{y}(t)| = e^{\alpha t}|\mathbf{y}(0)|; for \alpha = 0 the \mathbf{y}-trajectories are circles, and their images under P are ellipses.

(b) Differentiating x_1 = r\cos\theta, x_2 = r\sin\theta gives x_1x_2' - x_2x_1' = r^2\theta'; substitute x_1' = ax_1 + bx_2, x_2' = cx_1 + dx_2. The form’s discriminant is (d - a)^2 + 4bc = \tau^2 - 4\Delta < 0, so it has no zero off the origin and has the sign of its value c at (1, 0). \square

Result 6. Let \Phi(t) have as columns the solutions starting at (1, 0) and (0, 1), so \mathbf{x}(t) = \Phi(t)\mathbf{x}(0) and, by the Cauchy-Schwarz inequality, |\mathbf{x}(t)| \le \lVert\Phi(t)\rVert\,|\mathbf{x}(0)|, with \lVert\Phi\rVert the root of the sum of squared entries. If \tau < 0 < \Delta, Results 4 and 5 make every entry a combination of exponentials (times 1, t, \cos\beta t or \sin\beta t) with negative exponents, so \lVert\Phi\rVert \to 0 and is at most some K on [0, \infty): \delta = \varepsilon/K works and solutions tend to \mathbf{0}. If \tau = 0 < \Delta, \Phi is periodic, hence bounded, but nonzero solutions are periodic. Otherwise an eigenvalue has positive real part, and s times a straight-line solution, or s times the real solution e^{\alpha t}(\mathbf{a}\cos\beta t - \mathbf{b}\sin\beta t) of Linear systems of differential equations (equal to e^{\alpha t}\mathbf{a} at t = 2\pi k/\beta), starts as near \mathbf{0} as we like and is unbounded. \square

Worked examples

Example 1

A node

Classify the origin for \mathbf{x}' = \begin{bmatrix} -2 & 1 \\ 2 & -3 \end{bmatrix}\mathbf{x}, solve with \mathbf{x}(0) = (0, 3), and describe the trajectory.

Solution

Here \tau = -5, \Delta = 4 and \tau^2 - 4\Delta = 9 > 0: by Results 3 and 6, a sink node, asymptotically stable. The roots of \lambda^2 + 5\lambda + 4 = 0 are -1 and -4; A + I = \begin{bmatrix} -1 & 1 \\ 2 & -2 \end{bmatrix} gives \mathbf{v}_1 = (1, 1) and A + 4I = \begin{bmatrix} 2 & 1 \\ 2 & 1 \end{bmatrix} gives \mathbf{v}_2 = (1, -2). The start gives c_1 + c_2 = 0 and c_1 - 2c_2 = 3, so c_1 = 1, c_2 = -1 and

\mathbf{x} = e^{-t}\begin{bmatrix} 1 \\ 1 \end{bmatrix} - e^{-4t}\begin{bmatrix} 1 \\ -2 \end{bmatrix} = \begin{bmatrix} e^{-t} - e^{-4t} \\ e^{-t} + 2e^{-4t} \end{bmatrix}.

The e^{-4t} term dies first, so the trajectory enters the origin tangent to the slow line x_2 = x_1, from the first quadrant. Far back in time \mathbf{x} \approx e^{-4t}(-1, 2), parallel to (1, -2).

Example 2

A saddle and its separatrix

For \mathbf{x}' = \begin{bmatrix} 1 & 4 \\ 2 & -1 \end{bmatrix}\mathbf{x}, classify the origin, find curves that carry the trajectories, and find the starting points for which x_1(t) \to +\infty.

Solution

Here \tau = 0 and \Delta = -9 < 0: a saddle, unstable. From \lambda^2 - 9 = 0, \lambda = \pm 3; A - 3I = \begin{bmatrix} -2 & 4 \\ 2 & -4 \end{bmatrix} gives \mathbf{v}_1 = (2, 1) and A + 3I = \begin{bmatrix} 4 & 4 \\ 2 & 2 \end{bmatrix} gives \mathbf{v}_2 = (1, -1). So

x_1 = 2c_1e^{3t} + c_2e^{-3t}, \qquad x_2 = c_1e^{3t} - c_2e^{-3t}.

Then x_1 + x_2 = 3c_1e^{3t} and x_1 - 2x_2 = 3c_2e^{-3t}, so

(x_1 + x_2)(x_1 - 2x_2) = 9c_1c_2.

Every trajectory lies on a hyperbola of this family or, when c_1c_2 = 0, on an asymptote: the eigenvector lines x_1 = 2x_2 and x_1 + x_2 = 0. As t \to \infty, \mathbf{x} \approx c_1e^{3t}(2, 1), so x_1 \to +\infty exactly when c_1 > 0, that is, when x_1(0) + x_2(0) > 0. The line x_1 + x_2 = 0 of starts that approach the origin is a separatrix, parting solutions that leave up and right from those that leave down and left.

Example 3

A spiral through its rotation-scaling form

Solve \mathbf{x}' = \begin{bmatrix} -1 & -4 \\ 1 & -1 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (2, 0), decide the direction of turning, and find how long one turn takes and where the solution is after it.

Solution

Here \tau = -2, \Delta = 5 and \tau^2 - 4\Delta = -16: a spiral sink with \lambda = -1 \pm 2i. For \lambda = -1 - 2i, the second row of A - \lambda I = \begin{bmatrix} 2i & -4 \\ 1 & 2i \end{bmatrix} gives x_1 = -2ix_2, so \mathbf{v} = (-2i, 1) = (0, 1) + i(-2, 0) and

P = \begin{bmatrix} 0 & -2 \\ 1 & 0 \end{bmatrix}, \qquad C = \begin{bmatrix} -1 & -2 \\ 2 & -1 \end{bmatrix},

with AP = PC = \begin{bmatrix} -4 & 2 \\ -1 & -2 \end{bmatrix}. Then \mathbf{y}(0) = P^{-1}\mathbf{x}(0) = (0, -1), Result 5(a) gives \mathbf{y}(t) = e^{-t}R_{2t}(0, -1) = e^{-t}(\sin 2t, -\cos 2t), and

\mathbf{x}(t) = P\mathbf{y}(t) = e^{-t}\begin{bmatrix} 2\cos 2t \\ \sin 2t \end{bmatrix}.

So \displaystyle \frac14 x_1^2 + x_2^2 = e^{-2t}: the trajectory crosses each ellipse \displaystyle \frac14 x_1^2 + x_2^2 = k, the image under P of a circle, once, inward. Since c = 1 > 0 it turns counterclockwise. One turn takes \displaystyle \frac{2\pi}{\beta} = \pi, and \mathbf{x}(\pi) = e^{-\pi}(2, 0) \approx (0.0864, 0): each turn multiplies the crossing point on the positive x_1-axis by e^{-\pi} \approx 0.0432.

Example 4

A degenerate node

Solve \mathbf{x}' = \begin{bmatrix} 3 & -1 \\ 1 & 1 \end{bmatrix}\mathbf{x}, \mathbf{x}(0) = (1, 0), and describe how the trajectory leaves the origin.

Solution

Here \tau = 4 = \Delta, so \tau^2 = 4\Delta: the one eigenvalue is \lambda = 2, and A \ne 2I, so the origin is a degenerate source, unstable. Then N = A - 2I = \begin{bmatrix} 1 & -1 \\ 1 & -1 \end{bmatrix} has N^2 = 0, \ker N is the line of (1, 1), and N\mathbf{x}_0 = (1, 1). By Result 4(c),

\mathbf{x}(t) = e^{2t}\left(\begin{bmatrix} 1 \\ 0 \end{bmatrix} + t\begin{bmatrix} 1 \\ 1 \end{bmatrix}\right) = e^{2t}\begin{bmatrix} 1 + t \\ t \end{bmatrix}.

As t \to -\infty, \mathbf{x} \approx te^{2t}(1, 1) with t < 0, so the trajectory leaves the origin along the eigenvector line on the side of -(1, 1); it crosses the negative x_2-axis at t = -1, and its direction approaches (1, 1) as t \to \infty. Result 5(b) gives r^2\theta' = (x_1 - x_2)^2 \ge 0: the curve swings counterclockwise through a half turn in all.

Example 5

Omar’s pontoon

Omar is testing a floating pontoon for a ferry landing. After a wake passes, its height y meters above rest level, t seconds later, satisfies

400y'' + \gamma y' + 3600y = 0,

with mass 400 kilograms, buoyancy acting as a spring of constant 3600 newtons per meter, and damping \gamma \ge 0 newton-seconds per meter (Springs, damping and resonance). Classify the origin of the (y, y') phase plane for every \gamma \ge 0. For \gamma = 3000, find y when the pontoon is held 0.2 meters up and released from rest, and show that it never drops past its rest level.

Solution

With x_1 = y and x_2 = y' (Result 5 of Linear systems of differential equations),

\mathbf{x}' = \begin{bmatrix} 0 & 1 \\ -9 & -\gamma/400 \end{bmatrix}\mathbf{x}, \qquad \tau = -\frac{\gamma}{400}, \quad \Delta = 9.

Then \tau^2 - 4\Delta = \gamma^2/160{,}000 - 36 vanishes at \gamma = 2400. So \gamma = 0 gives a center, period \displaystyle \frac{2\pi}{3} seconds; 0 < \gamma < 2400 a spiral sink, clockwise since c = -9 < 0; \gamma = 2400 a degenerate sink, \lambda = -3; and \gamma > 2400 a sink node. In Figure 2 the family runs left along \Delta = 9.

For \gamma = 3000, \lambda^2 + 7.5\lambda + 9 = 0 gives \lambda = -1.5 and -6, with eigenvectors (1, \lambda). The start (0.2, 0) gives c_1 + c_2 = 0.2 and -1.5c_1 - 6c_2 = 0, so \displaystyle c_1 = \frac{4}{15}, \displaystyle c_2 = -\frac{1}{15}, and

y = \tfrac{4}{15}e^{-1.5t} - \tfrac{1}{15}e^{-6t} = \tfrac{1}{15}e^{-1.5t}\big(4 - e^{-4.5t}\big).

For t \ge 0, e^{-4.5t} \le 1 < 4, so y > 0: the pontoon settles without passing its rest level.

Example 6

A short proof: a conserved quantity when the trace is zero

Let A = \begin{bmatrix} a & b \\ c & -a \end{bmatrix} with \Delta = -a^2 - bc \ne 0. Prove that H = c\,x_1^2 - 2a\,x_1x_2 - b\,x_2^2 is constant along every solution of \mathbf{x}' = A\mathbf{x}, and deduce that the trajectories lie on ellipses when \Delta > 0 and on hyperbolas or lines when \Delta < 0.

Solution

Along a solution, x_1' = ax_1 + bx_2 and x_2' = cx_1 - ax_2, so x_1'x_2 + x_1x_2' = cx_1^2 + bx_2^2 and

\begin{aligned} H' &= 2cx_1x_1' - 2a(x_1'x_2 + x_1x_2') - 2bx_2x_2' \\ &= (2ac\,x_1^2 + 2bc\,x_1x_2) - (2ac\,x_1^2 + 2ab\,x_2^2) \\ &\quad + (2ab\,x_2^2 - 2bc\,x_1x_2) = 0. \end{aligned}

The form H has discriminant (2a)^2 + 4bc = -4\Delta. If \Delta > 0 it is definite, so a nonzero solution, which never reaches \mathbf{0} (Result 1), stays on an ellipse H = h \ne 0. If \Delta < 0 it is indefinite, with level sets hyperbolas and, for h = 0, two lines through the origin. In Example 2, H = 2(x_1 + x_2)(x_1 - 2x_2). \square

Boundaries

  • \det A = 0. Every point of \ker A is an equilibrium. If \tau \ne 0, \mathbf{x} = c_1\mathbf{v}_1 + c_2e^{\tau t}\mathbf{v}_2 (eigenvalues 0 and \tau) moves along lines parallel to \mathbf{v}_2, toward the line of equilibria when \tau < 0. If \tau = 0 and A \ne 0, then A^2 = 0 and \mathbf{x} = \mathbf{x}_0 + tA\mathbf{x}_0 moves along lines parallel to \ker A (Exercise 13). None of these is among the six types.
  • Linear systems only. Near an equilibrium of a nonlinear system the picture matches its linear part’s for a saddle, node or spiral, but not in general for a center: small nonlinear terms can turn a center into a spiral. Nonlinear systems and linearization takes this up.
  • Two dimensions only. For n \ge 3 there is no trace-determinant plane; the long-run behavior needs every eigenvalue.

Where this goes next

Nonlinear systems and linearization reads these pictures near each equilibrium of a nonlinear system. The Laplace transform turns forced systems into algebra.

Exercises

Classifying the origin

For Exercises 1 to 6, classify the origin of \mathbf{x}' = A\mathbf{x} (type and stability) and give the eigenvalues, with the eigenvector lines for real ones and the direction of turning for complex ones.

  1. A = \begin{bmatrix} -5 & 2 \\ 1 & -4 \end{bmatrix}
  2. A = \begin{bmatrix} 1 & -5 \\ 2 & 3 \end{bmatrix}
  3. A = \begin{bmatrix} 2 & 3 \\ 4 & 1 \end{bmatrix}
  4. A = \begin{bmatrix} -1 & 2 \\ -2 & -5 \end{bmatrix}
  5. A = \begin{bmatrix} 1 & -2 \\ 5 & -1 \end{bmatrix}
  6. A = \begin{bmatrix} -1 & -1 \\ 2 & -3 \end{bmatrix}

Solutions and trajectories

  1. For the matrix of Exercise 1, solve with \mathbf{x}(0) = (3, 0) and say along which line the trajectory enters the origin.
  2. For the matrix of Exercise 3, find the starting points whose solutions stay bounded as t \to \infty, and those for which x_1(t) \to +\infty.
  3. For the matrix of Exercise 4, solve with \mathbf{x}(0) = (1, 0) and find the lowest point of the trajectory for t \ge 0.
  4. For the matrix of Exercise 5, solve with \mathbf{x}(0) = (1, 0), find the ellipse its trajectory runs on, and its period.
  5. For the matrix of Exercise 2, solve with \mathbf{x}(0) = (1, 0) and find where the solution next crosses the positive x_1-axis.

Families and models

  1. Naomi tests a lighter pontoon with 250y'' + \gamma y' + 4000y = 0, in the units of Example 5. Find the \gamma > 0 at which the type of the origin changes, and classify the origin, with its eigenvalues, for \gamma = 1000, 2000 and 2500.
  2. For which real k is the origin of \mathbf{x}' = \begin{bmatrix} 1 & k \\ 1 & -1 \end{bmatrix}\mathbf{x} a center, and for which a saddle? Describe the trajectories when k = -1.
  3. Find the matrix A for which e^{2t}(1, 0) and e^{-t}(1, 2) both solve \mathbf{x}' = A\mathbf{x}, and name the type of its origin.
  4. Without finding eigenvectors, classify the origin for A = \begin{bmatrix} -2 & -3 \\ 2 & -1 \end{bmatrix}, decide the direction of turning, and find \theta' at the points (1, 0) and (0, 1).

Proofs

  1. Prove that \mathbf{x}' = -A\mathbf{x} has the same trajectories as \mathbf{x}' = A\mathbf{x} with the arrows reversed, so that a sink of one is a source of the other.
  2. Let \det A < 0. Prove that the eigenvalues are real with opposite signs, and that the solutions bounded on [0, \infty) are exactly those on the eigenvector line of the negative eigenvalue.
  3. Let \operatorname{tr} A = 0, \det A = \Delta > 0 and \displaystyle \omega = \sqrt{\Delta}. Prove that A^2 = -\Delta I, that the solution with \mathbf{x}(0) = \mathbf{x}_0 is \displaystyle \cos(\omega t)\,\mathbf{x}_0 + \frac{1}{\omega}\sin(\omega t)\,A\mathbf{x}_0, and that every solution has period \displaystyle \frac{2\pi}{\omega}.
Answers
  1. Sink node, asymptotically stable: \tau = -9, \Delta = 18, \tau^2 - 4\Delta = 9. Eigenvalues -3 with line of (1, 1) (slow) and -6 with line of (2, -1).
  2. Spiral source, unstable: \tau = 4, \Delta = 13, eigenvalues 2 \pm 3i, turning counterclockwise since c = 2 > 0.
  3. Saddle, unstable: \Delta = -10. Eigenvalues 5 with line of (1, 1) and -2 with line of (3, -4).
  4. Degenerate sink, asymptotically stable: \tau = -6, \Delta = 9, double eigenvalue -3, and A + 3I = \begin{bmatrix} 2 & 2 \\ -2 & -2 \end{bmatrix} \ne 0 has kernel the line of (1, -1).
  5. Center, stable but not asymptotically stable: \tau = 0, \Delta = 9, eigenvalues \pm 3i, turning counterclockwise since c = 5 > 0.
  6. Spiral sink, asymptotically stable: \tau = -4, \Delta = 5, eigenvalues -2 \pm i, turning counterclockwise since c = 2 > 0.
  7. \mathbf{x} = (e^{-3t} + 2e^{-6t},\ e^{-3t} - e^{-6t}), from c_1 = c_2 = 1. It enters tangent to the line of (1, 1), from the first quadrant.
  8. Writing \mathbf{x}(0) = c_1(1, 1) + c_2(3, -4) gives \displaystyle c_1 = \frac{1}{7}\big(4x_1(0) + 3x_2(0)\big). Bounded exactly when c_1 = 0, on the line 4x_1 + 3x_2 = 0 of (3, -4); x_1 \to +\infty exactly when 4x_1(0) + 3x_2(0) > 0.
  9. \mathbf{x} = e^{-3t}(1 + 2t,\ -2t), since N\mathbf{x}_0 = (2, -2). Then x_2' = e^{-3t}(6t - 2) vanishes at \displaystyle t = \frac13: the lowest point is \displaystyle \big(\frac{5}{3e}, -\frac{2}{3e}\big) \approx (0.613, -0.245).
  10. \displaystyle \mathbf{x} = \big(\cos 3t + \frac13\sin 3t,\ \frac53\sin 3t\big), on the ellipse 5x_1^2 - 2x_1x_2 + 2x_2^2 = 5 (Example 6 with a = 1, b = -2, c = 5), with period \displaystyle \frac{2\pi}{3}.
  11. \displaystyle \mathbf{x} = e^{2t}\big(\cos 3t - \frac13\sin 3t,\ \frac23\sin 3t\big), from \mathbf{v} = (-1 + 3i, 2) for 2 + 3i. Here x_2 = 0 at \displaystyle t = \frac{\pi}{3} with x_1 < 0, then at \displaystyle t = \frac{2\pi}{3}, at (e^{4\pi/3}, 0) \approx (65.94, 0).
  12. \Delta = 16 and \tau = -\gamma/250, so the type changes at \gamma = 2000. For \gamma = 1000: spiral sink, \displaystyle -2 \pm 2\sqrt3\,i. For \gamma = 2000: degenerate sink, -4 double. For \gamma = 2500: sink node, -2 and -8.
  13. \tau = 0 and \Delta = -1 - k: a center for k < -1, a saddle for k > -1. For k = -1, A^2 = 0 and A\mathbf{x}_0 = (x_1 - x_2)(1, 1): the line x_1 = x_2 is all equilibria, and other points move at constant speed parallel to (1, 1), forward exactly where x_1 > x_2.
  14. \displaystyle A = \begin{bmatrix} 2 & -\frac32 \\ 0 & -1 \end{bmatrix}, from A = PDP^{-1} with P = \begin{bmatrix} 1 & 1 \\ 0 & 2 \end{bmatrix} and D = \operatorname{diag}(2, -1); a saddle.
  15. \tau = -3, \Delta = 8, \tau^2 - 4\Delta = -23: a spiral sink. Result 5(b) gives r^2\theta' = 2x_1^2 + x_1x_2 + 3x_2^2, positive since c = 2 > 0: counterclockwise, with \theta' = 2 at (1, 0) and \theta' = 3 at (0, 1).
  16. If \mathbf{x}' = A\mathbf{x}, then \mathbf{z}(t) = \mathbf{x}(-t) has \mathbf{z}'(t) = -\mathbf{x}'(-t) = -A\mathbf{z}(t), and conversely: the same point sets, opposite orientation. The eigenvalues of -A are those of A negated, so a sink becomes a source. \square
  17. \tau^2 - 4\Delta > 0, so the eigenvalues are real, and their product \Delta is negative. With \lambda_1 > 0 > \lambda_2, \mathbf{x} = c_1e^{\lambda_1t}\mathbf{v}_1 + c_2e^{\lambda_2t}\mathbf{v}_2. If c_1 \ne 0, then for t \ge 0, where e^{\lambda_2t} \le 1, |\mathbf{x}| \ge |c_1||\mathbf{v}_1|e^{\lambda_1t} - |c_2||\mathbf{v}_2| \to \infty. If c_1 = 0, \mathbf{x} = c_2e^{\lambda_2t}\mathbf{v}_2 is bounded. \square
  18. With A = \begin{bmatrix} a & b \\ c & -a \end{bmatrix}, A^2 = (a^2 + bc)I = -\Delta I. The given \mathbf{x} has \mathbf{x}(0) = \mathbf{x}_0 and \mathbf{x}' = -\omega\sin(\omega t)\,\mathbf{x}_0 + \cos(\omega t)\,A\mathbf{x}_0, which equals \displaystyle A\mathbf{x} = \cos(\omega t)\,A\mathbf{x}_0 + \frac{1}{\omega}\sin(\omega t)\,A^2\mathbf{x}_0 since A^2\mathbf{x}_0 = -\omega^2\mathbf{x}_0. By uniqueness it is the solution, of period \displaystyle \frac{2\pi}{\omega}. \square
math.wiki · CC-BY-SA-4.0