Flow Lab

Syllabus / Module 7: Systems of ODEs & Dynamical Systems

Lesson 7.7

Nonlinear Systems & Linearization

Equilibria of nonlinear systems, the Jacobian, Hartman–Grobman, nullclines, and the nonlinear pendulum.

Written byClaude Sonnet

🌊
Why this matters for simulation: Every aircraft stability analysis linearizes the nonlinear flight equations about trim.

By the end of this lesson you can…

  • Find the equilibria of $x' = f(x,y),\ y' = g(x,y)$ by solving $f = g = 0$ and sketch the nullclines $f = 0$ and $g = 0$, reading the arrow directions in each region between them.
  • Derive the linearization $\mathbf u' = J\mathbf u$ from the two-variable Taylor expansion and compute the Jacobian $J$ at each equilibrium.
  • Classify an equilibrium from the trace $T$ and determinant $D$ of $J$, and state what Hartman–Grobman does and does not guarantee.
  • Show with numbers, using $r' = ar^3$, that a center in the linearization can be a stable or an unstable spiral in the nonlinear system.
  • Draw the full pendulum phase portrait: a center at $(0,0)$ proved by energy, a saddle with eigenvalues $\pm1$ at $(\pi,0)$, and the separatrix $E = 1$.
  • Measure how the linearization error scales with the distance from the equilibrium, and implement Newton's method with a numerical Jacobian to find and classify equilibria.
On this page
·
  1. Equilibria and nullclines
  2. Linearization: the Jacobian
  3. When the linearization is right, and when it lies
  4. The nonlinear pendulum
  5. How good is the linearization?
  6. Numerics: Newton’s method and the classifier
  7. Computing it
  8. Concept checks
  9. Problems
  10. Key Takeaways

A flight simulator integrates nonlinear equations, but nearly everything we can prove about stability is linear: eigenvalues, TT and DD, the factor 1+hλ1 + h\lambda. In 7.3 Phase Portraits & the Trace–Determinant Plane we classified x′=Ax\mathbf x' = A\mathbf x completely. This lesson is the bridge back to the real equations. Near an equilibrium a smooth system looks linear (the zoom-in idea of 2.2 Linearization, the Chain Rule & the Jacobian), so the Jacobian there is a matrix AA and the whole (T,D)(T, D) plane applies. We find where it works, where it fails (centers), and how to compute it: Newton’s method finds the equilibria and an eigenvalue test classifies them.

▶ Video Zooming in on an equilibrium until the nonlinear flow and its Jacobian flow coincidecaptions only
Transcript
  1. Two competing species: x prime equals x times three minus x minus two y, and y prime equals y times two minus x minus y. The nullclines are x equals zero, x plus two y equals three, y equals zero, and x plus y equals two. They meet at four equilibria: zero zero, zero two, three zero, one one.
  2. Between nullclines each sign is fixed. Below both lines, x prime and y prime are positive: arrows point up and right. Above both, both are negative. Between the lines the arrows point up-left or down-right.
  3. Write x equals x star plus u and expand in a Taylor series. Drop the remainder, of order size squared, and u prime equals J u, with J the Jacobian. Zoom in on the saddle, and the nonlinear flow merges with the linear one.
  4. At zero zero, T is five, D is six, eigenvalues three and two: an unstable node. At three zero, minus four, three: a stable node. At zero two, minus three, two: a stable node. At one one, minus two, minus one, eigenvalues minus one plus or minus root two: a saddle.
  5. The story is bistable. Either species alone is stable; coexistence at one one is a saddle. From two two the flow ends at zero two; from two, one point five, at three zero. The saddle's stable curve divides the basins, a global fact.
  6. Hartman Grobman, stated, not proved. Near a hyperbolic equilibrium, with no eigenvalue of real part zero, a continuous change of coordinates carries nonlinear orbits onto those of u prime equals J u. Saddles stay saddles; sinks and sources stay. It can bend angles. A center, or a zero eigenvalue, is where linearization is silent.
  7. Add cubic terms with coefficient a. At the origin the Jacobian is the same for every a: T zero, D one, a center. But r prime equals a r cubed. From one, zero to t equals four, r is 0.74536 for a minus 0.1, one for a zero, 2.23607 for a plus 0.1.
  8. The pendulum has equilibria at n pi, zero, with trace zero and determinant cosine x star. At the origin the eigenvalues are plus or minus i: the linearization cannot tell. But the energy, y squared over two minus cosine x, is conserved, so level sets are closed ovals: a true center.
  9. At pi, zero the eigenvalues are plus and minus one: a saddle. Its energy is one, and E equals one separates oscillation from rotation. Start on it at zero, two and the orbit creeps toward pi: RK4 gives pi minus x of ten equals 1.82 times ten to the minus four.
  10. Halving the offset d divides the species error at t equals one by 3.86, 3.92, 3.96, tending to four: order two. The pendulum error drops by 8.00 each time: order three, since sine x minus x has no quadratic term.
  11. Newton's method: x n plus one equals x n minus J inverse F. From one point five, point five, the errors against one one are 1.77 times ten to the minus one, then minus two, 2.16 times ten to the minus four, 3.29 times ten to the minus eight, and 7.1 times ten to the minus sixteen: digits double.
  12. Newton fails if J is singular, and finds roots, not stable roots; the eigenvalues of J decide stability. A real part of zero is inconclusive, opposite signs a saddle, equal signs a sink or source. Aircraft stability linearizes about a trim state; the fluid form is from memory.
  13. Recap. Solve f equals g equals zero, take the Jacobian, classify by T and D. Hartman Grobman covers hyperbolic points; centers can lie, as r prime equals a r cubed shows. Errors shrink like d squared, d cubed for the pendulum. Next: conservative systems and Lyapunov functions.

Equilibria and nullclines

An autonomous planar system is x′=f(x,y)x' = f(x,y), y′=g(x,y)y' = g(x,y). An equilibrium satisfies f=g=0f = g = 0: the velocity vanishes and the state sits still. The nullclines are the curves f=0f = 0 (where x′=0x' = 0, so arrows are vertical) and g=0g = 0 (where y′=0y' = 0, arrows horizontal). Equilibria are exactly the points where an xx-nullcline meets a yy-nullcline, and between nullclines each of x′x' and y′y' keeps one sign, so the arrows keep one quadrant of direction.

Example: two competing species. Take

x′=x(3−x−2y),y′=y(2−x−y).x' = x(3 - x - 2y), \qquad y' = y(2 - x - y).

Each species limits itself and hurts the other. The xx-nullclines are x=0x = 0 and the line x+2y=3x + 2y = 3; the yy-nullclines are y=0y = 0 and x+y=2x + y = 2. Intersect them in pairs:

  • x=0x = 0 with y=0y = 0: (0,0)(0,0); x=0x = 0 with x+y=2x + y = 2: (0,2)(0,2);
  • x+2y=3x + 2y = 3 with y=0y = 0: (3,0)(3,0); the two lines with each other: (1,1)(1,1).

Four equilibria. In the first quadrant the two lines cross at (1,1)(1,1) and cut it into four regions. Below both lines both x′x' and y′y' are positive (arrows point up and right); above both, both are negative; between them the arrows point up-left or down-right. Block A of the lab cell prints these sign patterns and the sim draws both nullclines (one color for x′=0x' = 0, another for y′=0y' = 0) over the arrows.

Linearization: the Jacobian

Let (x∗,y∗)(x_*, y_*) be an equilibrium and write x=x∗+ux = x_* + u, y=y∗+vy = y_* + v. Since x∗x_* is constant, u′=x′=f(x∗+u,y∗+v)u' = x' = f(x_* + u, y_* + v). Expand in a two-variable Taylor series about the equilibrium, where f(x∗,y∗)=0f(x_*, y_*) = 0:

u′=fx u+fy v+Rf,v′=gx u+gy v+Rg,u' = f_x\,u + f_y\,v + R_f, \qquad v' = g_x\,u + g_y\,v + R_g,

with the partial derivatives evaluated at (x∗,y∗)(x_*, y_*) and the remainders RR made of terms of order ∣u∣2|\mathbf u|^2 and higher. Dropping them gives the linearization

u′=Ju,J=(fxfygxgy)∣(x∗,y∗).\mathbf u' = J\mathbf u, \qquad J = \begin{pmatrix} f_x & f_y \\ g_x & g_y \end{pmatrix}\Big|_{(x_*, y_*)}.

JJ is the Jacobian of the vector field, the matrix of its tangent map from 2.2 Linearization, the Chain Rule & the Jacobian. The one-variable version, 3.8 Autonomous Equations, Equilibria & Stability, is the same statement: u′=f′(x∗)uu' = f'(x_*)u, stable if f′(x∗)<0f'(x_*) \lt 0. Now the eigenvalues, TT and DD of JJ give the portrait from the last lesson.

For the competing species, J=(3−2x−2y−2x−y2−x−2y)J = \begin{pmatrix}3 - 2x - 2y & -2x\\ -y & 2 - x - 2y\end{pmatrix}. At each equilibrium (block A, and the sim’s readout):

EquilibriumJJTTDDλ\lambdaType
(0,0)(0,0)(3002)\begin{pmatrix}3&0\\0&2\end{pmatrix}55663, 23,\ 2unstable node
(3,0)(3,0)(−3−60−1)\begin{pmatrix}-3&-6\\0&-1\end{pmatrix}−4-433−3, −1-3,\ -1stable node
(0,2)(0,2)(−10−2−2)\begin{pmatrix}-1&0\\-2&-2\end{pmatrix}−3-322−1, −2-1,\ -2stable node
(1,1)(1,1)(−1−2−1−1)\begin{pmatrix}-1&-2\\-1&-1\end{pmatrix}−2-2−1-1−1±2-1 \pm \sqrt2saddle

The story is bistable. Either species alone is a stable state, and coexistence (1,1)(1,1) is a saddle, so it is unstable. Which species wins depends on which side of the saddle’s stable curve you start: the sim’s default seed (2,2)(2,2) ends at (0,2)(0,2), while (2,1.5)(2, 1.5) ends at (3,0)(3,0) (click it). Linearization decides local questions only; the basins are a global fact it cannot see.

When the linearization is right, and when it lies

An equilibrium is hyperbolic if no eigenvalue of JJ has real part zero.

The change of coordinates is only continuous, so it can bend angles: it preserves “everything flows in”, not exactly how it winds. When the eigenvalues are real and distinct, or genuinely complex, smooth terms also keep node versus spiral (from memory). The borderline T2=4DT^2 = 4D (star or degenerate node) is where the finer shape can change (from memory), and the real failures are the non-hyperbolic cases: a center (T=0T = 0, D>0D \gt 0) and a zero eigenvalue (D=0D = 0), where the linearization is silent.

Why it fails there is a competition of sizes. The remainder RR is of order ∣u∣3|\mathbf u|^3 or ∣u∣2|\mathbf u|^2, tiny next to JuJ\mathbf u unless JuJ\mathbf u has no net radial push. With Re λ≠0\mathrm{Re}\,\lambda \neq 0 the linear part grows or shrinks the radius at a rate proportional to rr and wins near the equilibrium. With Re λ=0\mathrm{Re}\,\lambda = 0 it contributes nothing to r′r', and the nonlinear term decides.

A center that is not a center. Consider

x′=−y+a x(x2+y2),y′=x+a y(x2+y2).x' = -y + a\,x(x^2 + y^2), \qquad y' = x + a\,y(x^2 + y^2).

At the origin J=(0−110)J = \begin{pmatrix}0&-1\\1&0\end{pmatrix} for every aa: T=0T = 0, D=1D = 1, λ=±i\lambda = \pm i, a center. But with r2=x2+y2r^2 = x^2 + y^2 we get r r′=xx′+yy′=ar4r\,r' = x x' + y y' = a r^4, so

r′=ar3,r(t)=r01−2ar02t.r' = a r^3, \qquad r(t) = \frac{r_0}{\sqrt{1 - 2a r_0^2 t}}.

(The angle moves at the constant rate θ′=1\theta' = 1.) The sign of aa chooses the fate. Starting from (1,0)(1, 0) and integrating with RK4 to t=4t = 4 (block B; the sim’s readout shows the same):

aameasured r(4)r(4)formula
−0.1-0.10.745360.745361/1.81/\sqrt{1.8}
001.000001.0000011
+0.1+0.12.236072.236071/0.21/\sqrt{0.2}

For a<0a \lt 0 the origin is a stable spiral, for a>0a \gt 0 an unstable one, and a=0a = 0 is the true center. The linearization is identical in all three. The center lies on the line T=0T = 0, and the nonlinear terms act like a perturbation of unknown sign.

The nonlinear pendulum

Write x′′=−sin⁡xx'' = -\sin x as x′=yx' = y, y′=−sin⁡xy' = -\sin x. Equilibria: y=0y = 0 and sin⁡x=0\sin x = 0, so (nπ,0)(n\pi, 0). The Jacobian is J=(01−cos⁡x0)J = \begin{pmatrix}0&1\\-\cos x&0\end{pmatrix}, T=0T = 0 always, D=cos⁡x∗D = \cos x_*.

  • (0,0)(0,0): D=1D = 1, λ=±i\lambda = \pm i. Non-hyperbolic: the linearization cannot tell. But the energy E=12y2−cos⁡xE = \tfrac12 y^2 - \cos x satisfies E′=y(−sin⁡x)+(sin⁡x) y=0E' = y(-\sin x) + (\sin x)\,y = 0, and near the origin E+1≈12(x2+y2)E + 1 \approx \tfrac12(x^2 + y^2), so its level sets are closed ovals. So it is a true center, thanks to conservation and not to the linearization (energy methods are the subject of 7.8 Conservative Systems & Lyapunov Functions).
  • (π,0)(\pi, 0): D=−1D = -1, λ=±1\lambda = \pm 1 with eigenvectors (1,1)(1, 1) and (1,−1)(1, -1). A saddle, and hyperbolic, so it persists.

The saddle sits at energy E=0−cos⁡π=1E = 0 - \cos\pi = 1. Orbits with E<1E \lt 1 oscillate, orbits with E>1E \gt 1 rotate all the way round, and E=1E = 1 separates them: y2=2(1+cos⁡x)y^2 = 2(1 + \cos x), i.e. y=±2cos⁡(x/2)y = \pm 2\cos(x/2), the separatrices. Start exactly on one, (0,2)(0, 2), where E=2−1=1E = 2 - 1 = 1. The orbit creeps toward (π,0)(\pi, 0) along the stable direction and never reaches it in finite time: with u=π−xu = \pi - x on this curve, u′=−2sin⁡(u/2)≈−uu' = -2\sin(u/2) \approx -u, so u∼Ce−tu \sim Ce^{-t} (the eigenvalue −1-1). RK4 gives π−x(10)=1.82×10−4\pi - x(10) = 1.82\times10^{-4} (block C), consistent with 4e−10=1.82×10−44e^{-10} = 1.82\times10^{-4} from the exact solution u=4arctan⁡e−tu = 4\arctan e^{-t}. In the sim the default seed is this orbit; click just below and just above it and the readout flips between oscillation and rotation, with EE printed. Because xx is an angle, the picture repeats every 2π2\pi (the true phase space is a cylinder).

How good is the linearization?

Compare the nonlinear solution from x∗+d\mathbf x_* + \mathbf d with the linear solution from d\mathbf d at t=1t = 1, with the offset d=∣d∣d = |\mathbf d| in the direction (1,1)/2(1,1)/\sqrt2 (block D; the sim’s inset prints this error, so set W=2dW = 2d and select the equilibrium):

dderror, species at (1,1)(1,1)error, pendulum at (0,0)(0,0)
0.20.22.427×10−32.427\times10^{-3}1.028×10−31.028\times10^{-3}
0.10.16.293×10−46.293\times10^{-4}1.285×10−41.285\times10^{-4}
0.050.051.604×10−41.604\times10^{-4}1.606×10−51.606\times10^{-5}
0.0250.0254.050×10−54.050\times10^{-5}2.007×10−62.007\times10^{-6}

Halving dd divides the species error by 3.86,3.92,3.963.86, 3.92, 3.96 (tending to 44) and the pendulum error by 8.008.00 each time. So the observed orders are 22 and 33. The mechanism is the remainder: the species equations contain x2x^2 and xyxy terms, so R=O(d2)R = O(d^2) and over a fixed time the error is O(d2)O(d^2). The pendulum has sin⁡x−x≈−x3/6\sin x - x \approx -x^3/6, odd in xx, so no quadratic term exists and R=O(d3)R = O(d^3). So the error vanishes faster than the distance, but at d=0.2d = 0.2 it is already visible in the third digit.

Numerics: Newton’s method and the classifier

To find equilibria we solve F(x)=0F(\mathbf x) = \mathbf 0 with F=(f,g)F = (f, g). Linearize FF at the current guess: F(xn+δ)≈F(xn)+J(xn)δF(\mathbf x_n + \boldsymbol\delta) \approx F(\mathbf x_n) + J(\mathbf x_n)\boldsymbol\delta, set it to zero, and step:

xn+1=xn−J(xn)−1F(xn).\mathbf x_{n+1} = \mathbf x_n - J(\mathbf x_n)^{-1}F(\mathbf x_n).

(In code, solve the 2×22\times2 system; never form the inverse.) It is the same Jacobian that classifies the equilibrium. For the species from (1.5,0.5)(1.5, 0.5) the errors against (1,1)(1,1) are (block E):

step12345
error1.77×10−11.77\times10^{-1}1.77×10−21.77\times10^{-2}2.16×10−42.16\times10^{-4}3.29×10−83.29\times10^{-8}7.1×10−167.1\times10^{-16}

Correct digits go 0.75,1.75,3.7,7.5,150.75, 1.75, 3.7, 7.5, 15: they roughly double each step, which is quadratic convergence, with en+1/en2≈0.57,0.69,0.71e_{n+1}/e_n^2 \approx 0.57, 0.69, 0.71 (step 5 is at rounding level). Two cautions. It fails when JJ is singular, e.g. from (0.5,1.5)(0.5, 1.5), where D=0D = 0. And it finds roots, not stable roots: stability is a separate question answered by the eigenvalues of JJ at the answer. The sim does exactly this: Newton from a 17×1317\times13 grid of starts, remove duplicates, classify each. For the classifier: compute λ\lambda; if some ∣Re λ∣≤|\mathrm{Re}\,\lambda| \le tol the answer is “inconclusive”; otherwise opposite signs mean saddle, and equal signs mean a sink or source, a spiral when the imaginary part is nonzero.

◆ Interactive Nonlinear phase plane with nullclines, auto-found classified equilibria, and the linearized inset

Loading simulation…

What to try. (1) Competing species: the readout lists four equilibria; select (1,1)(1,1) and see T=−2T = -2, D=−1D = -1, saddle. Click (2,2)(2,2) then (2,1.5)(2,1.5). (2) Set W=0.4,0.2,0.1,0.05W = 0.4, 0.2, 0.1, 0.05 and compare the inset error with the table. (3) Pendulum: select (3.142,0.000)(3.142, 0.000) for the saddle, and (0.000,0.000)(0.000, 0.000), flagged non-hyperbolic. (4) Center that is not a center: a=−0.1a = -0.1, 00, +0.1+0.1 and read r(4)r(4).

Computing it

🐍 PythonLab: equilibria, signs, the counterexample, the pendulum, error scaling, Newton (blocks A to E)
⌘/Ctrl + Enter

Now build the whole pipeline: a numerical Jacobian, Newton, and the classifier.

🧪 Coding exerciseExercise: Newton's method with a numerical Jacobian, and eigenvalue classification
⌘/Ctrl + Enter

Concept checks

✅ Concept check

The pendulum's linearization at $(0,0)$ is a center. Why is the nonlinear pendulum really a center there?

✅ Concept check

For the pendulum linearized at $(0,0)$ the error at $t = 1$ is $1.285\times10^{-4}$ for $d = 0.1$. By what factor does the error drop when $d$ is halved to $0.05$ (measured table)?

✅ Concept check

At an equilibrium the Jacobian has $T = -3$ and $D = -2$. What does Hartman–Grobman let you conclude?

✅ Concept check

Newton's method on the species system has error $3.29\times10^{-8}$ (about 7.5 correct digits). About how many correct digits does the next iterate have? (Quadratic convergence; table above.)

Problems

✏️ Pen & paperLotka–Volterra equilibria●○○

Classify the equilibria of x′=x(1−y)x' = x(1 - y), y′=y(x−1)y' = y(x - 1).

Show worked solution

f=g=0f = g = 0 gives (0,0)(0,0) and (1,1)(1,1). J=(1−y−xyx−1)J = \begin{pmatrix}1 - y & -x\\ y & x - 1\end{pmatrix}. At (0,0)(0,0): J=diag(1,−1)J = \mathrm{diag}(1, -1), a saddle. At (1,1)(1,1): J=(0−110)J = \begin{pmatrix}0&-1\\1&0\end{pmatrix}, T=0T = 0, D=1D = 1, λ=±i\lambda = \pm i: inconclusive. The invariant H=x−ln⁡x+y−ln⁡yH = x - \ln x + y - \ln y is conserved (the sim shows it unchanged along the orbit), and has a minimum at (1,1)(1,1), so it is a true center. (The conservation of HH is checked in the sim; the algebra is by hand.)

✏️ Pen & paperNonlinear terms decide●●○

Show that x′=−y−x3x' = -y - x^3, y′=x−y3y' = x - y^3 has a stable spiral at the origin although its linearization is a center.

Show worked solution

J=(0−110)J = \begin{pmatrix}0&-1\\1&0\end{pmatrix} (the cubic terms have zero derivative at 00), so λ=±i\lambda = \pm i. Then r r′=x(−y−x3)+y(x−y3)=−(x4+y4)r\,r' = x(-y - x^3) + y(x - y^3) = -(x^4 + y^4). Since x4+y4≥r4/2x^4 + y^4 \ge r^4/2 (equality on the diagonals), r′≤−r3/2<0r' \le -r^3/2 \lt 0: the radius strictly decreases while the angle turns, so it spirals in. (By hand; not in the sim, which does the +a+a and −a-a versions.)

✏️ Pen & paperTwo Newton steps by hand●●●

For the competing-species system take x0=(2,2)\mathbf x_0 = (2,2). Compute FF, JJ and the Newton step.

Show worked solution

F(2,2)=(2(3−2−4), 2(2−4))=(−6,−4)F(2,2) = (2(3 - 2 - 4),\ 2(2 - 4)) = (-6, -4). J(2,2)=(−5−4−2−4)J(2,2) = \begin{pmatrix}-5&-4\\-2&-4\end{pmatrix}. Solve Jδ=−F=(6,4)J\boldsymbol\delta = -F = (6, 4): subtracting the rows, −3δ1=2-3\delta_1 = 2, so δ1=−2/3\delta_1 = -2/3, and −4δ2=4+2δ1-4\delta_2 = 4 + 2\delta_1 gives δ2=−2/3\delta_2 = -2/3. So x1=(4/3,4/3)=(1.3333,1.3333)\mathbf x_1 = (4/3, 4/3) = (1.3333, 1.3333). Block E prints the next ones, converging to the saddle (1,1)(1,1) with errors 0.4710.471, 0.09430.0943, 0.005550.00555, 2.2×10−52.2\times10^{-5} in order (step 5: 3×10−103\times10^{-10}). Note that Newton from (2,2)(2,2) lands on the unstable equilibrium, although the flow from (2,2)(2,2) goes to (0,2)(0,2).

Key Takeaways

  • Equilibria solve f=g=0f = g = 0; the nullclines f=0f = 0 and g=0g = 0 cut the plane into regions with a fixed sign pattern of (x′,y′)(x', y'). For the species system they meet at (0,0),(0,2),(3,0),(1,1)(0,0), (0,2), (3,0), (1,1).
  • The linearization is u′=Ju\mathbf u' = J\mathbf u, from the Taylor expansion with remainder O(∣u∣2)O(|\mathbf u|^2); classify by TT, DD exactly as in the last lesson. The species types: two stable nodes, one unstable node, and the saddle (1,1)(1,1) with λ=−1±2\lambda = -1 \pm \sqrt2.
  • Hartman–Grobman: at a hyperbolic equilibrium (no Re λ=0\mathrm{Re}\,\lambda = 0) the topology of the flow matches the linearization. Centers and D=0D = 0 are non-hyperbolic: the linearization is silent.
  • With x′=−y+ax(x2+y2)x' = -y + ax(x^2 + y^2) the same center gives r′=ar3r' = ar^3: after t=4t = 4 the radius is 0.7450.745 for a=−0.1a = -0.1 and 2.2362.236 for a=+0.1a = +0.1.
  • Pendulum: (0,0)(0,0) is a center because of energy conservation, (π,0)(\pi, 0) is a saddle with λ=±1\lambda = \pm1, and E=1E = 1 separates oscillation from rotation (the separatrix orbit reaches π−x(10)=1.82×10−4\pi - x(10) = 1.82\times10^{-4}).
  • Measured, not assumed: the linearization error at t=1t = 1 falls by about 44 per halving of dd for the species (quadratic terms) and by 88 for the pendulum (cubic terms).
  • Newton’s method xn+1=xn−J−1F\mathbf x_{n+1} = \mathbf x_n - J^{-1}F doubles the correct digits each step (errors 1.8×10−1,1.8×10−2,2.2×10−4,3.3×10−8,7×10−161.8\times10^{-1}, 1.8\times10^{-2}, 2.2\times10^{-4}, 3.3\times10^{-8}, 7\times10^{-16}) and uses the same JJ that classifies the equilibrium.
  • Linearization about a base state is how aircraft and flow stability are analyzed; next, 7.8 Conservative Systems & Lyapunov Functions proves stability with energy-like functions where linearization fails.
Finished this lesson?

Mark it complete to track your progress (saved in this browser). Reading every section does this for you.

Where this leads

These later lessons build directly on this one: