Nonlinear systems and linearization
Almost no nonlinear system of differential equations can be solved in formulas. How much does the matrix of first partial derivatives at a resting point say about the solutions that start near it?
Where you are 6 in Systems and transforms
Why this exists
Nonlinear systems, a swinging bell or two competing species, can almost never be solved in formulas, but a model usually asks something local: where can the system rest, and does a small push die out or grow? For one equation the derivative test of Autonomous equations and stability answered from the sign of f'(c). In the plane the Jacobian matrix replaces f'(c), and the classification of Phase portraits does the rest, except where it falls silent.
Definition
In plain words: near a resting point, replace the system by its best linear approximation and read the behavior from that.
Let f and g be of class C^1 (continuous first partial derivatives, Partial derivatives) on an open set D of the plane. The planar autonomous system
has unknowns x(t), y(t). Write \mathbf{x} = (x, y) and \mathbf{F} = (f, g), so the system is \mathbf{x}' = \mathbf{F}(\mathbf{x}). The theorem of Existence and uniqueness holds for it with |\cdot| read as length, so trajectories do not cross. We also use, without proof, the standard fact that a solution staying in a closed bounded subset of D for t \ge 0 is defined for all t \ge 0 (Hirsch, Smale and Devaney, Differential Equations, Dynamical Systems, and an Introduction to Chaos).
A point \mathbf{p} = (x_0, y_0) with f(x_0, y_0) = g(x_0, y_0) = 0 is an equilibrium point (or critical point): the constant \mathbf{x}(t) = \mathbf{p} is a solution. The Jacobian matrix there is
and with \mathbf{u} = (u, v) = (x - x_0, y - y_0), the system \mathbf{u}' = J(x_0, y_0)\,\mathbf{u} is the linearization at \mathbf{p}. The equilibrium is hyperbolic if no eigenvalue of J(x_0, y_0) has real part 0.
Stability is defined as in Phase portraits, with \mathbf{p} for the origin: \mathbf{p} is stable if for every \varepsilon > 0 there is \delta > 0 such that every solution with |\mathbf{x}(0) - \mathbf{p}| < \delta is defined for all t \ge 0 and stays within \varepsilon of \mathbf{p}; asymptotically stable if moreover such solutions tend to \mathbf{p}; unstable if not stable.
For a C^1 function V, the derivative of V along the system is DV = V_x f + V_y g. By the chain rule (The multivariable chain rule), \displaystyle \frac{d}{dt}V(\mathbf{x}(t)) = DV(\mathbf{x}(t)) along every solution.
Intuition
Magnify the plane about \mathbf{p}. Under enough magnification f and g cannot be told from their tangent planes (Tangent planes and linear approximation), which pass through zero, so the field becomes J\mathbf{u}. A robust linear picture (saddle, node, spiral) survives the small error; a center, destroyed by the slightest inward or outward drift, does not.
Results
Throughout, f and g are C^1 on the open set D, \mathbf{p} \in D is an equilibrium point, J is its Jacobian matrix, and \mathbf{u} = \mathbf{x} - \mathbf{p}.
1. The linear part. With \mathbf{R}(\mathbf{u}) = \mathbf{F}(\mathbf{p} + \mathbf{u}) - J\mathbf{u}, every solution satisfies
and |\mathbf{R}(\mathbf{u})|/|\mathbf{u}| \to 0 as \mathbf{u} \to \mathbf{0}.
2. Linearized stability. (a) If both eigenvalues of J have real part less than -\mu < 0, there are \rho > 0 and K \ge 1 such that every solution with |\mathbf{u}(0)| < \rho has |\mathbf{u}(t)| \le K|\mathbf{u}(0)|\,e^{-\mu t/2} for all t \ge 0; so \mathbf{p} is asymptotically stable. (b) If an eigenvalue of J has positive real part, \mathbf{p} is unstable.
3. The type survives (Hartman and Grobman). Let f and g have continuous second partial derivatives and \mathbf{p} be hyperbolic. (a) A continuous bijection with continuous inverse, from a neighborhood of \mathbf{p} onto one of \mathbf{0}, carries pieces of trajectories of the system onto pieces of trajectories of the linearization, keeping the direction of time. (b) If the linearization is a saddle, exactly two trajectories tend to \mathbf{p} as t \to \infty, from opposite sides, tangent to the eigenvector line of the negative eigenvalue (with \mathbf{p}, the stable curve), and exactly two tend to \mathbf{p} as t \to -\infty, tangent to the other line (the unstable curve). A node with distinct eigenvalues stays a node and a spiral a spiral, turning the same way; a double eigenvalue gives a node or a spiral.
4. Lyapunov functions. Let V be C^1 on a disk B \subseteq D about \mathbf{p}, with V(\mathbf{p}) = 0, V > 0 elsewhere on B, and DV \le 0 on B. Then \mathbf{p} is stable. If DV = 0 on B, no solution except the equilibrium solution tends to \mathbf{p}.
5. Without hyperbolicity the linearization decides nothing. For the center J = \begin{bmatrix} 0 & -1 \\ 1 & 0 \end{bmatrix} and for J = \operatorname{diag}(0, -1), systems with that J at \mathbf{p} can be asymptotically stable, stable but not asymptotically stable, or unstable there.
Why it is true
Result 1. As \mathbf{F}(\mathbf{p}) = \mathbf{0}, the first component of \mathbf{R}(\mathbf{u}) is f(\mathbf{p} + \mathbf{u}) - f(\mathbf{p}) - f_x(\mathbf{p})\,u - f_y(\mathbf{p})\,v. A C^1 function is differentiable (Result 2 of Tangent planes and linear approximation), so this divided by |\mathbf{u}| tends to 0; likewise for g. And \mathbf{u}' = \mathbf{x}' = \mathbf{F}(\mathbf{p} + \mathbf{u}). \square
Lemma. If both eigenvalues of J have real part less than -\mu < 0, some invertible P makes C = P^{-1}JP satisfy \mathbf{w} \cdot C\mathbf{w} \le -\mu|\mathbf{w}|^2 for every \mathbf{w}. If J has two independent real eigenvectors, take them as the columns of P: C = \operatorname{diag}(\lambda_1, \lambda_2) and \mathbf{w} \cdot C\mathbf{w} = \lambda_1 w_1^2 + \lambda_2 w_2^2 \le -\mu|\mathbf{w}|^2. For eigenvalues \alpha \pm \beta i, Complex eigenvalues and rotation-scaling matrices gives C = \begin{bmatrix} \alpha & -\beta \\ \beta & \alpha \end{bmatrix} and \mathbf{w} \cdot C\mathbf{w} = \alpha|\mathbf{w}|^2. Otherwise J has a double eigenvalue \lambda, J \ne \lambda I, and N = J - \lambda I has N^2 = 0 (Result 4(c) of Phase portraits). Pick \mathbf{w}_2 with \mathbf{w}_1 = N\mathbf{w}_2 \ne \mathbf{0}; then J\mathbf{w}_1 = \lambda\mathbf{w}_1, J\mathbf{w}_2 = \lambda\mathbf{w}_2 + \mathbf{w}_1, and the two are independent (\mathbf{w}_2 = k\mathbf{w}_1 would give N\mathbf{w}_2 = kN^2\mathbf{w}_2 = \mathbf{0}). Columns \mathbf{w}_1, \eta\mathbf{w}_2 with \eta > 0 give C = \begin{bmatrix} \lambda & \eta \\ 0 & \lambda \end{bmatrix}, so \displaystyle \mathbf{w} \cdot C\mathbf{w} = \lambda|\mathbf{w}|^2 + \eta w_1w_2 \le (\lambda + \frac{\eta}{2})|\mathbf{w}|^2. Take \eta = -2(\lambda + \mu) > 0. \square
Result 2(a). Take P, C from the lemma and put \mathbf{u} = P\mathbf{w}. With \lVert P \rVert the root of the sum of squared entries, |P\mathbf{w}| \le \lVert P \rVert\,|\mathbf{w}| by the Cauchy-Schwarz inequality. By Result 1, \mathbf{w}' = C\mathbf{w} + \mathbf{S}(\mathbf{w}) with \mathbf{S}(\mathbf{w}) = P^{-1}\mathbf{R}(P\mathbf{w}), and |\mathbf{S}(\mathbf{w})|/|\mathbf{w}| \to 0. Choose \sigma > 0 so that |\mathbf{w}| \le \sigma gives \mathbf{p} + P\mathbf{w} \in D and \displaystyle |\mathbf{S}(\mathbf{w})| \le \frac{\mu}{2}|\mathbf{w}|. While |\mathbf{w}| \le \sigma,
If |\mathbf{w}(0)| < \sigma, then |\mathbf{w}| never reaches \sigma, being nonincreasing until it would. So \mathbf{w} stays in a closed disk, is defined for all t \ge 0, and \big(e^{\mu t}|\mathbf{w}|^2\big)' \le 0 gives |\mathbf{w}(t)| \le |\mathbf{w}(0)|\,e^{-\mu t/2}. Take K = \lVert P \rVert\,\lVert P^{-1} \rVert and \rho = \sigma/\lVert P^{-1} \rVert; given \varepsilon, \delta = \min(\rho, \varepsilon/K) shows stability. \square
Result 2(b). If both eigenvalues have positive real part, pick \mu > 0 below both and apply the lemma to -J: some P gives \mathbf{w} \cdot C\mathbf{w} \ge \mu|\mathbf{w}|^2, and as above (|\mathbf{w}|^2)' \ge \mu|\mathbf{w}|^2 while |\mathbf{w}| \le \sigma. A start with 0 < |\mathbf{w}(0)| < \sigma then has |\mathbf{w}(t)|^2 \ge |\mathbf{w}(0)|^2e^{\mu t} until it leaves the disk, which it must, since by the standard fact it cannot stop existing inside. Otherwise the eigenvalues are real with \lambda_1 > 0 \ge \lambda_2; with eigenvectors as columns of P, C = \operatorname{diag}(\lambda_1, \lambda_2). Let W = w_1^2 - w_2^2. Then
Where W \ge 0, \displaystyle w_1^2 \ge \frac12|\mathbf{w}|^2, so choosing \sigma with \displaystyle |\mathbf{S}(\mathbf{w})| \le \frac{\lambda_1}{4}|\mathbf{w}| gives \displaystyle W' \ge \frac{\lambda_1}{2}|\mathbf{w}|^2 \ge \frac{\lambda_1}{2}W. Start at \mathbf{w}(0) = (s, 0), 0 < s < \sigma. Inside the disk W increases wherever W \ge 0, so W \ge s^2, and \displaystyle W' \ge \frac{\lambda_1}{2}s^2 pushes |\mathbf{w}|^2 \ge W past \sigma^2 in finite time. Either way, starts as near \mathbf{p} as we like reach |\mathbf{u}| \ge \sigma/\lVert P^{-1} \rVert. The cases are exhaustive, as complex eigenvalues share one real part. \square
Result 3 is stated without proof. Part (a) is proved in Hirsch, Smale and Devaney, and part (b) in Coddington and Levinson, Theory of Ordinary Differential Equations.
Result 4. Given \varepsilon > 0 with the closed \varepsilon-disk about \mathbf{p} inside B, let m > 0 be the minimum of V on its circle (extreme value theorem). Choose \delta < \varepsilon with V < m on |\mathbf{u}| < \delta. A solution starting there has V(\mathbf{x}(t)) nonincreasing while in B, so V < m and it never reaches the circle; it stays in the closed disk, so it exists for all t \ge 0. If DV = 0 on B and \mathbf{x}(t) \to \mathbf{p}, then for t \ge T, \mathbf{x}(t) \in B and V(\mathbf{x}(t)) is constant and tends to V(\mathbf{p}) = 0, so \mathbf{x}(t) = \mathbf{p}; by uniqueness \mathbf{x} is the equilibrium solution. \square
Result 5. Example 6 settles the center. For \operatorname{diag}(0, -1), Exercise 10 gives the unstable and asymptotically stable cases, and x' = 0, y' = -y, with solutions (x_0, y_0e^{-t}), is stable but not asymptotically stable. \square
Worked examples
Two equilibria, two pictures
Find and classify every equilibrium point of x' = x - y, y' = x^2 + y^2 - 8.
Solutiony = x and 2x^2 = 8 give (2, 2) and (-2, -2), and J(x, y) = \begin{bmatrix} 1 & -1 \\ 2x & 2y \end{bmatrix}.
At (2, 2), J = \begin{bmatrix} 1 & -1 \\ 4 & 4 \end{bmatrix} has \tau = 5, \Delta = 8, eigenvalues \displaystyle \frac{5 \pm i\sqrt7}{2}: unstable by Result 2(b), a spiral source by Result 3, counterclockwise since the lower-left entry 4 is positive (Result 5(b) of Phase portraits).
At (-2, -2), J = \begin{bmatrix} 1 & -1 \\ -4 & -4 \end{bmatrix} has \Delta = -8, eigenvalues \displaystyle \frac{-3 \pm \sqrt{41}}{2} \approx 1.70 and -4.70: a saddle, unstable. Eigenvectors are (1, 1 - \lambda), so the stable curve arrives tangent to the line of \displaystyle (2, 5 + \sqrt{41}).
Maya’s bell
Maya rings a tenor bell, swinging it through full circles. The angle \theta in radians of its mouth from straight down, t seconds after a stroke, satisfies \theta'' + \theta' + 6\sin\theta = 0, friction lumped into \theta'. With x = \theta, y = \theta', classify every equilibrium. Why must a bell resting mouth up be held, and how must it arrive to come to rest there?
SolutionThe system is x' = y, y' = -6\sin x - y, with equilibria (k\pi, 0), k an integer, and
For even k (mouth down), \tau = -1, \Delta = 6, eigenvalues \displaystyle \frac{-1 \pm i\sqrt{23}}{2}: asymptotically stable (Result 2(a)), a spiral sink turning clockwise (Result 3).
For odd k (mouth up), \Delta = -6, and \lambda^2 + \lambda - 6 = 0 gives \lambda = 2 and -3, with eigenvectors (1, \lambda): (1, 2) and (1, -3). A saddle: from balance, any push sends the bell away, so it must be held. By Result 3 only two trajectories reach (\pi, 0), tangent to the line of (1, -3): the bell must arrive with \theta' \approx -3(\theta - \pi), rising at about three times its remaining angle per second.
How near is near enough
Show that every solution of x' = -x + y^2, y' = -2y + x^2 starting at distance less than \displaystyle \sqrt2 from the origin tends to it.
SolutionJ(0, 0) = \operatorname{diag}(-1, -2), so Result 2(a) gives asymptotic stability, with a \rho it does not compute. Follow its proof with V = x^2 + y^2 = r^2:
Here -2x^2 - 4y^2 \le -2r^2, \displaystyle |xy| \le \frac12 r^2 and \displaystyle |x + y| \le \sqrt2\,r (Cauchy-Schwarz), so
If \displaystyle r(0) = r_0 < \sqrt2, then while r \le r_0, \displaystyle DV \le -(2 - \sqrt2\,r_0)V \le 0; so r never exceeds r_0 and \displaystyle V(t) \le V(0)\,e^{-(2 - \sqrt2 r_0)t} \to 0. The disk cannot grow much: x = y^2, y^3 = 2 gives a second equilibrium (2^{2/3}, 2^{1/3}) at distance \approx 2.027, a saddle since there \Delta = 2 - 2^{4/3}\cdot 2^{5/3} = -6.
A center the linearization cannot settle
For the softening spring x'' + 4x - x^3 = 0, classify the equilibria of x' = y, y' = -4x + x^3, decide the origin with \displaystyle E = \frac12 y^2 + 2x^2 - \frac14 x^4, and find the trajectories through the other equilibria.
SolutionThe equilibria are (0, 0) and (\pm 2, 0), and J = \begin{bmatrix} 0 & 1 \\ -4 + 3x^2 & 0 \end{bmatrix}. At (\pm 2, 0) the eigenvalues are \displaystyle \pm 2\sqrt2: saddles. At the origin they are \pm 2i, a center, and Result 2 is silent.
Along solutions DE = y(-4x + x^3) + (4x - x^3)y = 0. On the disk of radius 2, \displaystyle E = \frac12 y^2 + x^2\big(2 - \frac14x^2\big) > 0 except at the origin, so Result 4 makes the origin stable but not asymptotically stable. Each level set E = c, 0 < c < 4, |x| < 2, is a closed curve with no equilibrium on it, so nearby trajectories are closed (Figure 1).
At the saddles E = 8 - 4 = 4, and E = 4 means \displaystyle \frac12 y^2 = \frac14(x^2 - 4)^2:
two parabolas crossing at (\pm 2, 0) with slopes \displaystyle \pm 2\sqrt2, matching the eigenvectors \displaystyle (1, \pm 2\sqrt2). On the arc \displaystyle y = (4 - x^2)/\sqrt2, |x| < 2, x' = y > 0: one trajectory runs from (-2, 0) to (2, 0), and the lower arc runs back.
Diego’s barnacles
Diego surveys two species of barnacle competing for space on a seawall. With densities x and y in thousands per square meter and t in months, his model is
each species alone growing logistically (Logistic growth). Classify the equilibria in the closed first quadrant and decide who wins from (1.2, 0.9).
Solutionx' = 0 needs x = 0 or x + 2y = 3; y' = 0 needs y = 0 or x + y = 2. With
| point | J | eigenvalues, type |
|---|---|---|
| (0, 0) | \begin{bmatrix} 3 & 0 \\ 0 & 2 \end{bmatrix} | 3, 2: source node |
| (3, 0) | \begin{bmatrix} -3 & -6 \\ 0 & -1 \end{bmatrix} | -3, -1: sink node |
| (0, 2) | \begin{bmatrix} -1 & 0 \\ -2 & -2 \end{bmatrix} | -1, -2: sink node |
| (1, 1) | \begin{bmatrix} -1 & -2 \\ -1 & -1 \end{bmatrix} | \displaystyle -1 \pm \sqrt2: saddle |
Coexistence is unstable; each species alone at capacity is asymptotically stable. The axes are trajectories, so by uniqueness an absent species stays absent.
At (1, 1), \displaystyle -1 - \sqrt2 has eigenvector \displaystyle (\sqrt2, 1) and \displaystyle -1 + \sqrt2 has \displaystyle (\sqrt2, -1). The stable curve is tangent there to \displaystyle (\sqrt2, 1) and, in Figure 2, separates the two outcomes. The offset \displaystyle (0.2, -0.1) = a(\sqrt2, 1) + b(\sqrt2, -1) of (1.2, 0.9) has \displaystyle b = \frac{0.2/\sqrt2 + 0.1}{2} \approx 0.121 > 0, along \displaystyle (\sqrt2, -1), which points toward (3, 0). Numerically, at t = 30 the state is within 0.001 of (3, 0). The first species takes the wall.
A short proof: one linearization, three fates
Let a be real. Prove that x' = -y + ax(x^2 + y^2), y' = x + ay(x^2 + y^2) has a center as its linearization at the origin for every a, but that the origin is asymptotically stable for a < 0, stable but not asymptotically stable for a = 0, and unstable for a > 0.
SolutionThe cubic terms and their first partials vanish at the origin, so J(0, 0) = \begin{bmatrix} 0 & -1 \\ 1 & 0 \end{bmatrix}, eigenvalues \pm i. Along a solution, s = x^2 + y^2 has
If s(0) = s_0 > 0, s never vanishes (s \equiv 0 is the solution through 0), and (1/s)' = -2a gives
For a < 0, s decreases to 0: \delta = \varepsilon works and solutions tend to the origin. For a = 0 the trajectories are circles: stable, and none tends to the origin. For a > 0, s \to \infty as \displaystyle t \to \frac{1}{2as_0}, so every other start, however near, leaves every disk. \square
Boundaries
- Centers and zero eigenvalues. If no eigenvalue has positive real part, Results 2 and 3 are silent (Result 5). A Lyapunov function (Example 3, Exercise 15), a conserved quantity (Example 4, Exercises 6 and 8) or a direct computation (Example 6) must take over.
- Local only. Results 2 and 3 hold near \mathbf{p}, at an unknown scale. The basins of Example 5 and the closed trajectory of Exercise 11 are invisible to every linearization.
- Non-isolated equilibria. x' = x - y, y' = x^2 - y^2 rests at every point of the line y = x, where \det J = 0.
Where this goes next
Beyond linearization lie limit cycles (Exercise 11), bifurcations, where a parameter carries an eigenvalue across the imaginary axis, and chaos in three dimensions. The Laplace transform returns to linear equations with forcing.
Exercises
Equilibria and linearization
For Exercises 1 to 4, find every equilibrium point, its Jacobian matrix and eigenvalues, and what Results 2 and 3 say.
- x' = 1 - y, y' = x^2 - y
- x' = 4x - y^2, y' = x - y
- x' = \sin y, y' = x - y
- x' = x^2 + y^2 - 5, y' = xy - 2
- Find the linearization of x' = \sin(x + 2y), y' = e^x - 1 - 3y at the equilibrium point (0, 0) and classify it.
Models and borderline cases
- Lily’s pendulum clock loses almost no energy: \theta'' + 9\sin\theta = 0. With x = \theta, y = \theta', classify the equilibria where the linearization decides, show with \displaystyle E = \frac12y^2 + 9(1 - \cos x) that (0, 0) is stable but not asymptotically stable, and find the trajectories through (\pm\pi, 0).
- On another stretch of seawall Diego’s model is x' = x(4 - x - y), y' = y(6 - x - 3y). Classify the equilibria in the closed first quadrant and name the asymptotically stable outcome.
- For the predator-prey model x' = x(3 - y), y' = y(x - 2) with x, y > 0, classify the equilibria where linearization decides, show that H = x - 2\ln x + y - 3\ln y is constant along solutions, and deduce that (2, 3) is stable but not asymptotically stable.
- A bell obeys \theta'' + b\theta' + 4\sin\theta = 0 with b > 0. For which b is the linearization at (0, 0) a spiral, a node, a degenerate node, and what do Results 2 and 3 say? What is (\pi, 0)?
- Show that x' = x^2, y' = -y and x' = -x^3, y' = -y have the same linearization at the origin, solve both, and show that the origin is unstable for the first and asymptotically stable for the second.
- For x' = x - y - x(x^2 + y^2), y' = x + y - y(x^2 + y^2), classify the origin, find equations for \displaystyle r = \sqrt{x^2 + y^2} and the polar angle, show that the unit circle is a trajectory, and find r(t) when r(0) = r_0 > 0.
Proofs
- Let h be C^1, c > 0 and h(x_0) = 0. Prove that the equilibrium (x_0, 0) of x'' + cx' + h(x) = 0, with y = x', is asymptotically stable if h'(x_0) > 0 and a saddle, hence unstable, if h'(x_0) < 0.
- Let F have continuous second partial derivatives. Prove that at an equilibrium of the gradient system x' = -F_x, y' = -F_y the Jacobian matrix has real eigenvalues, so the linearization is never a spiral or a center, and that if F_{xx} > 0 and F_{xx}F_{yy} - F_{xy}^2 > 0 there, the equilibrium is asymptotically stable.
- Let H have continuous second partial derivatives. For the Hamiltonian system x' = H_y, y' = -H_x, prove that H is constant along solutions, that a hyperbolic equilibrium is a saddle, and that no equilibrium is asymptotically stable.
- Prove that the origin of x' = y - x^3, y' = -x - y^3 is asymptotically stable, although its linearization is a center. (Use V = x^2 + y^2 and show DV \le -V^2.)
Answers
- (1, 1): J = \begin{bmatrix} 0 & -1 \\ 2 & -1 \end{bmatrix}, eigenvalues \displaystyle \frac{-1 \pm i\sqrt7}{2}, asymptotically stable, a spiral sink turning counterclockwise. (-1, 1): eigenvalues 1 and -2 with eigenvectors (1, -1) and (1, 2), a saddle, unstable.
- (0, 0): J = \begin{bmatrix} 4 & 0 \\ 1 & -1 \end{bmatrix}, eigenvalues 4 and -1 with eigenvectors (5, 1) and (0, 1), a saddle. (4, 4): J = \begin{bmatrix} 4 & -8 \\ 1 & -1 \end{bmatrix}, eigenvalues \displaystyle \frac{3 \pm i\sqrt7}{2}, a spiral source. Both unstable.
- (k\pi, k\pi) for every integer k, J = \begin{bmatrix} 0 & \cos k\pi \\ 1 & -1 \end{bmatrix}. Even k: eigenvalues \displaystyle \frac{-1 \pm \sqrt5}{2}, a saddle. Odd k: eigenvalues \displaystyle \frac{-1 \pm i\sqrt3}{2}, asymptotically stable, a spiral sink turning counterclockwise.
- (x + y)^2 = 9 and (x - y)^2 = 1 give four points; J = \begin{bmatrix} 2x & 2y \\ y & x \end{bmatrix}. (2, 1): \displaystyle 3 \pm \sqrt3, a source node. (1, 2): \Delta = -6, \displaystyle \frac{3 \pm \sqrt{33}}{2}, a saddle. (-2, -1): \displaystyle -3 \pm \sqrt3, a sink node, asymptotically stable. (-1, -2): \displaystyle \frac{-3 \pm \sqrt{33}}{2}, a saddle.
- By the chain rule J(0, 0) = \begin{bmatrix} 1 & 2 \\ 1 & -3 \end{bmatrix}: u' = u + 2v, v' = u - 3v. \Delta = -5, eigenvalues \displaystyle -1 \pm \sqrt6: a saddle, unstable.
- (k\pi, 0) with J = \begin{bmatrix} 0 & 1 \\ -9\cos x & 0 \end{bmatrix}: odd k, eigenvalues \pm 3, saddles; even k, \pm 3i, undecided. DE = -9y\sin x + 9y\sin x = 0, and E > 0 on the disk of radius \pi except at (0, 0): Result 4. At (\pm\pi, 0), E = 18, and \displaystyle \frac12y^2 = 9 + 9\cos x = 18\cos^2\frac{x}{2} gives \displaystyle y = \pm 6\cos\frac{x}{2}, arcs joining (-\pi, 0) and (\pi, 0).
- (0, 0): 4, 6, a source node. (4, 0): \begin{bmatrix} -4 & -4 \\ 0 & 2 \end{bmatrix}, -4, 2, a saddle. (0, 2): \begin{bmatrix} 2 & 0 \\ -2 & -6 \end{bmatrix}, 2, -6, a saddle. (3, 1): \begin{bmatrix} -3 & -3 \\ -1 & -3 \end{bmatrix}, \displaystyle -3 \pm \sqrt3, a sink node. Coexistence at (3, 1) is the asymptotically stable outcome.
- (0, 0): eigenvalues 3, -2, a saddle. (2, 3): J = \begin{bmatrix} 0 & -2 \\ 3 & 0 \end{bmatrix}, \displaystyle \pm i\sqrt6, undecided. DH = (x - 2)(3 - y) + (y - 3)(x - 2) = 0. x - 2\ln x has second derivative \displaystyle \frac{2}{x^2} > 0 and a strict minimum at x = 2; likewise y - 3\ln y at y = 3. So V = H - H(2, 3) satisfies Result 4 with DV = 0.
- J(0, 0) = \begin{bmatrix} 0 & 1 \\ -4 & -b \end{bmatrix}, \tau^2 - 4\Delta = b^2 - 16: a spiral for 0 < b < 4, a node for b > 4, a degenerate node, \lambda = -2 double, for b = 4. As \tau = -b < 0, all are asymptotically stable (Result 2(a)), and for b = 4 Result 3 gives a node or a spiral. At (\pi, 0), \Delta = -4: a saddle for every b.
- Both have J(0, 0) = \begin{bmatrix} 0 & 0 \\ 0 & -1 \end{bmatrix} and y = y_0e^{-t}. First: \displaystyle x = \frac{x_0}{1 - x_0t}, so for any x_0 > 0, x \to \infty as \displaystyle t \to \frac{1}{x_0}: unstable. Second: \displaystyle x = \frac{x_0}{\sqrt{1 + 2x_0^2t}}; |x| \le |x_0|, |y| \le |y_0| give stability with \delta = \varepsilon, and both tend to 0.
- J(0, 0) = \begin{bmatrix} 1 & -1 \\ 1 & 1 \end{bmatrix}, eigenvalues 1 \pm i: a spiral source. rr' = xx' + yy' = r^2 - r^4, so r' = r - r^3; xy' - yx' = r^2, so the angle has derivative 1. At r = 1, r' = 0: (\cos(t + c), \sin(t + c)) solves. With s = r^{-2}, s' = 2 - 2s, so r(t)^2 = \dfrac{1}{1 + (r_0^{-2} - 1)e^{-2t}} for all t \ge 0, and r \to 1.
- J = \begin{bmatrix} 0 & 1 \\ -h'(x_0) & -c \end{bmatrix}, \tau = -c < 0, \Delta = h'(x_0). If \Delta > 0, complex eigenvalues have real part \displaystyle -\frac{c}{2}, and real ones have positive product and negative sum, so both are negative: Result 2(a). If \Delta < 0, the eigenvalues are real with negative product: a saddle, unstable by Result 2(b). \square
- J = -\begin{bmatrix} F_{xx} & F_{xy} \\ F_{xy} & F_{yy} \end{bmatrix} is symmetric (mixed partials agree), and \begin{bmatrix} a & b \\ b & d \end{bmatrix} has \tau^2 - 4\Delta = (a - d)^2 + 4b^2 \ge 0: real eigenvalues. Under the hypotheses F_{yy} > 0 too, so \tau < 0 < \Delta and Result 2(a) applies. \square
- DH = H_xH_y - H_yH_x = 0, and \tau = H_{yx} - H_{xy} = 0. Hyperbolic eigenvalues summing to 0 cannot be \pm\beta i, so they are real, nonzero, of opposite signs: a saddle. If \mathbf{p} were asymptotically stable, with every start in a disk B tending to \mathbf{p}, then H(\mathbf{q}) = H(\mathbf{x}(t)) \to H(\mathbf{p}) for each \mathbf{q} \in B; so H is constant on B, every point of B is an equilibrium, and a start \mathbf{q} \ne \mathbf{p} stays at \mathbf{q}, a contradiction. \square
- J(0, 0) has eigenvalues \pm i. DV = -2(x^4 + y^4) \le -(x^2 + y^2)^2 = -V^2, since the difference is (x^2 - y^2)^2 \ge 0. So V is nonincreasing: solutions exist for all t \ge 0, and \delta = \varepsilon gives stability. If V(0) > 0, then V > 0 throughout by uniqueness, (1/V)' \ge 1, and V(t) \le \dfrac{V(0)}{1 + V(0)t} \to 0. \square