Why this matters for simulation: Dutch roll — an aircraft's wobbly yaw-roll oscillation — is a complex eigenvalue pair.
By the end of this lesson you can…
Turn the complex solution $e^{\lambda t}\mathbf v$ for $\lambda = \alpha \pm i\beta$ into two real solutions using Euler's formula, and fit initial data with a $2\times2$ linear solve.
Read the frequency $\beta$, the period $2\pi/\beta$ and the decay per turn $e^{2\pi\alpha/\beta}$ from the eigenvalue, and say what $\mathbf a = \mathrm{Re}\,\mathbf v$ and $\mathbf b = \mathrm{Im}\,\mathbf v$ do to the shape of the spiral.
Separate the two repeated-eigenvalue cases: $A = \lambda I$ (star node) and the defective case with one eigenvector.
Find a generalized eigenvector $\mathbf w$ with $(A - \lambda I)\mathbf w = \mathbf v$ and derive the solution $e^{\lambda t}(c_1\mathbf v + c_2(t\mathbf v + \mathbf w))$.
Locate the transient hump of $t e^{\lambda t}$ (peak at $t = 1/|\lambda|$) and recognise critical damping as a defective node.
Measure how the eigenvector matrix becomes ill-conditioned as a defective matrix is approached, and say why codes prefer Schur or expm to eig there.
7.2 Linear Systems & the Eigenvalue Method solved x′=Ax for real, distinct eigenvalues, and 7.3 Phase Portraits & the Trace–Determinant Plane classified every 2×2 picture from T and D, taking on trust that a complex pair gives a spiral at rate α and speed β. This lesson cashes both cheques: the real form of the spiral solutions, and the missing second solution when an eigenvalue is repeated. Both cases matter for simulation. A decaying swirl, a lightly damped aircraft wobble and a vortex are complex pairs, and critical damping, the best-behaved oscillator, is a repeated eigenvalue.
▶ VideoComplex eigenvectors become spirals; a missing eigenvector becomes a factor of tcaptions only
Transcript
Last lesson took a spiral on trust. Here is the proof. The matrix with rows 0.8, minus 2 and 1, minus 1.2 has eigenvalues minus 0.2 plus or minus i and eigenvector 2, 1 minus i: real part a, 2, 1; imaginary part b, 0, minus 1.
The complex function e to the lambda t times v solves the system. Euler's formula splits it into cosine plus i sine, and since A is real, the real and imaginary parts are each a solution: x one starts at a, x two at b.
Starting from 1, 1 we need c one a plus c two b to match. Two c one is 1 and c one minus c two is 1, so c is 0.5, minus 0.5. The solution spirals inward, passing minus 0.2466, 0.4424 at t equals 1.
Each solution rotates inside a decay envelope. After one period, 2 pi over beta, x of P is e to the alpha P times x of 0, so the factor per turn is e to the 2 pi alpha over beta: 0.28461 here. The case minus 0.5, 2 gives 0.20788; alpha 0 gives exactly 1.
The shape lives in a and b. A acts on them as a rotation-scaling block, and a cos s plus b sin s is an ellipse through a and b. Last lesson's matrix gives circles; Example 1 is tilted. The factor stays 0.28461.
Now a repeated eigenvalue, T squared equals 4 D. The null space of A minus lambda I decides. If A equals lambda I, every trajectory is a straight ray: a star node. Otherwise there is one eigenline and a missing solution: a defective node.
Guess a factor t: x equals e to the lambda t times t v plus w. The t terms match, and the rest needs A minus lambda I times w equals v, a generalized eigenvector. So x is e to the lambda t times c one plus c two t, v, plus c two w.
Example 2: trace minus 4, determinant 4, so lambda is minus 2 twice. One eigenvector, v equals 1, 1, and w equals 0, 1. Starting at w, x is e to the minus 2 t times t, t plus 1: 0.18394, 0.55182 at t 0.5; 0.13534, 0.27067 at t 1.
The term t e to the lambda t starts at zero and decays, so it must rise first, peaking at t equals 1 over the size of lambda with height 1 over lambda e: 0.5 and 0.18394 for minus 2. From x zero equal to 1, 0, there is a trough of minus 0.02489 at t 1.5.
In the basis v, w, A v is lambda v and A w is v plus lambda w, so A is P J P inverse. The exponential of J is e to the lambda t times 1, t, 0, 1: the same t. For Example 2 at 0.5: e to the minus 1 times 1, 0.5, 0, 1.
Critical damping is a defective node. For y double prime plus 2 y prime plus y, lambda is minus 1 twice, v is 1, minus 1, w is 1, 0, and y is 1 plus t times e to the minus t. y of 1 is 0.73576, y of 2 is 0.40601; the velocity humps at 0.36788.
Perturb Example 2 by delta: lambda is minus 2 plus or minus root delta, with nearly parallel eigenvectors. The condition number of V is 19.95, 200 and 2000 for delta 10 to the minus 2, minus 4, minus 6: about 2 over root delta, and about 10 to the 8 when defective. From memory, codes prefer Schur or expm.
To sum up. A pure imaginary pair is a vortex; the Dutch roll, from memory, is a lightly damped pair with factor near 1. A complex pair gives real and imaginary parts, factor 0.28461 here. A repeated eigenvalue is a star or defective node; critical damping is defective. Next: the matrix exponential.
Complex eigenvalues: from one complex solution to two real ones
A real matrix with T2<4D has eigenvalues λ=α±iβ and complex eigenvectors, and the algebra is the same as before: (A−λI)v=0. Write v=a+ib with real vectors a,b. The complex function eλtv solves x′=Ax, but we want real solutions. Euler’s formula (see 1.7 Complex Eigenvalues & Rotation) gives eλt=eαt(cosβt+isinβt), so
eλtv=eαt[(acosβt−bsinβt)+i(asinβt+bcosβt)],
since i⋅ibsinβt=−bsinβt. Because A is real, A(Rexc)=Re(Axc) and likewise for Im, so the real part and the imaginary part of any complex solution are each a real solution. The two real solutions are
x1=eαt(acosβt−bsinβt),x2=eαt(asinβt+bcosβt).
The conjugate eigenvalue λˉ with vˉ only returns the conjugate solution, whose real and imaginary parts are x1 and −x2: nothing new. They are independent because at t=0 they equal a and b, and these are independent: if b=ka then v=(1+ik)a, a complex multiple of a real vector, which would make λ real. So the general real solution is x=c1x1+c2x2, and since x(0)=c1a+c2b, fitting initial data is the 2×2 solve [ab]c=x0.
Example 1. Take A=(0.81−2−1.2), with T=−0.4, D=1.04 and T2−4D=−4, so λ=−0.2±i (the same eigenvalues as the previous lesson’s spiral). Row 1 of A−λI is (1−i,−2), so (1−i)v1=2v2 and v=(2,1−i); row 2 checks: 2+(−1−i)(1−i)=2−2=0. So a=(2,1), b=(0,−1), and
For x0=(1,1): 2c1=1 and c1−c2=1 give c=(0.5,−0.5), and the combination simplifies to x(t)=e−0.2t(cost−sint,cost). At t=1 this is (−0.2466,0.4424), which block A reproduces from the formula, from eAt and from RK4.
Frequency, decay and shape
Each real solution is a rotation inside a decay envelope. The angular frequency is β, the period is P=2π/β, and the number of turns per unit time is β/2π (0.15915 for β=1). After exactly one period the cosines and sines are back at their starting values, so x(P)=eαPx(0), a vector identity and not just a statement about length. The factor per turn is therefore
ρ=e2πα/β.
For α=−0.2, β=1 this is 0.28461, the number the previous lesson measured, and here x(2π)=0.28461(1,1), exactly parallel to x0=(1,1); block A and the sim’s yellow dot both show (0.28461,0.28461).
There is a one-line rule behind this. Comparing real and imaginary parts of A(a+ib)=(α+iβ)(a+ib) gives Aa=αa−βb and Ab=βa+αb: in the basis (a,b) the matrix acts as the rotation-scaling block C=(α−ββα), so A=PCP−1 with P=[ab], the real form of the linear-algebra lesson. For Example 1, Aa=(−0.4,0.8) and −0.2(2,1)−(0,−1)=(−0.4,0.8). The sim prints both residuals (zero) for every setting of its sliders.
The rotation is not a circle in general. The curve acoss+bsins (the orbit at α=0) is an ellipse through a and b, and the spiral is that ellipse shrunk by eαt as it is traced: the motion goes from a at t=0 to −b a quarter turn later. The matrix of the previous lesson, (−0.21−1−0.2), has a=(1,0) and b=(0,−1), orthogonal and equal in length, so its ellipses are circles. Our Example 1 has a=(2,1) and b=(0,−1), and the spiral is tilted and stretched. The shape lives in a,b; the speed and decay live in β,α; and ρ depends on the shape not at all. To see it, load the lab at tilt 0 and tilt 1 and read the same 0.28461.
◆ InteractiveSpirals, defective nodes and star nodes: drag the start point and read the eigen-structure
Loading simulation…
What to try. (1) Complex mode: move the tilt slider from 0 to 1 and watch the dashed ellipse and the spiral change shape while the readout’s factor per turn stays 0.28461. (2) Put α=−0.5, β=2: the factor per turn is 0.20788 and the period 3.1416. (3) Put α=0, β=2, tilt 0: closed circles, factor 1.00000, the vortex u=(−2y,2x). (4) Defective and critical modes: drag the start point on and off the dashed line.
Repeated eigenvalues
If λ is a double root of λ2−Tλ+D (so T2=4D), there are two cases, decided by the null space of A−λI. Case 1: diagonalizable. If A−λI=0, i.e. A=λI, every vector is an eigenvector, x(t)=eλtx0, and every trajectory is a straight ray: the star node. Case 2: defective. Otherwise A−λI is a nonzero matrix with a one-dimensional null space: one eigenline, a single solution eλtv of the usual kind, and a missing second one.
Where the second solution comes from. Guess that the extra solution has a factor t: try x=eλt(tv+w). Then x′=eλt[λ(tv+w)+v], while Ax=eλt(tλv+Aw) because Av=λv. The terms in t match on their own, and the rest requires
(A−λI)w=v.
Such a w is a generalized eigenvector. It exists because for a 2×2 matrix with a double root, Cayley–Hamilton gives (A−λI)2=0: the image of N=A−λI lies inside its null space, which is the line spanned by v, and the image is a line since N=0, so v is in the image. Any w off the eigenline works; adding a multiple of v to w gives another valid choice and only reshuffles the constants. Then x1=eλtv and x2=eλt(tv+w) are independent (v,w are, at t=0), and
Example 2.A=(−3−11−1): T=−4, D=4, T2−4D=0, so λ=−2 twice. A+2I=(−1−111) is nonzero, so v=(1,1) is the only eigenvector. Solve (A+2I)w=v: −w1+w2=1, and w=(0,1) works (block B prints (A+2I)w=(1,1) and (A+2I)2=0). Starting from x0=w=(0,1) gives c=(0,1) and x=e−2t(t,t+1). At t=0.5 this is (0.18394,0.55182) and at t=1 it is (0.13534,0.27067); formula, eAt and RK4 agree in block B.
The picture. Since x≈c2teλtv at late times, for λ<0 every trajectory eventually runs into the origin tangent to the single eigenline, because the teλtv term dominates: that is the defective node of the previous lesson. Backwards in time the tv term dominates again, so the curves come in from far away along the same line, on the opposite side, with a bend in between. The sim’s defective mode draws v dashed, w as an arrow, and the curves bending onto the line.
The hump.teλt is zero at t=0 and decays at the end, so it must rise first. Setting its derivative (1+λt)eλt to zero gives the peak at t=1/∣λ∣ with height 1/(∣λ∣e). For λ=−2 that is t=0.5 and height 0.18394, the value in the dashed curve of the sim and in block B. A stable defective system can therefore get larger before it decays: a transient hump. In general the v-coordinate is (c1+c2t)eλt, with its extremum at t∗=1/∣λ∣−c1/c2 if that is positive; the readout states it, and for x0=(1,0) in Example 2 (c=(1,−1)) it is a trough of −0.02489 at t=1.5.
Growing and stable cases. Nothing above used the sign of λ. For λ>0 the same formulas describe a defective node that expands, which is the stable one traced backwards in time (A→−A sends λ→−λ, as the previous lesson noted); the hump is then simply monotone growth. For the star node the decay is the plain scalar one: after the e-folding time 1/∣λ∣ (the sim prints it) every vector has shrunk by e−1, and no direction is special. The defective node is the opposite extreme: one direction is special, and everything else leans toward it.
Jordan form. In the basis (v,w) we have Av=λv and Aw=v+λw, so A=PJP−1 with P=[vw] and
J=(λ01λ),eJt=eλt(10t1).
The second formula is stated here and derived in 7.5 The Matrix Exponential & Fundamental Matrices; the t in the corner is the same t as in tv+w. Block B finds P−1AP=J for Example 2 and prints eJ⋅0.5=(0.3678800.183940.36788), which is e−1 times (100.51).
Coupled oscillators, previewed. Two masses on springs give a four-dimensional system whose eigenvalues come in pairs ±iωk, one pair per normal mode: each mode is a center (α=0, β=ωk) in the sense above, and damping would add α<0 to each. That is the subject of 14.1 Coupled Oscillators & Normal Modes.
Critical damping is a defective node
The scalar lesson 5.2 Constant Coefficients & the Characteristic Equation solved y′′+2y′+y=0 with the repeated root r=−1 and the unexpected solution tert. Here is why. With x=(y,y′) the companion matrix is (0−11−2), with T=−2, D=1, double eigenvalue −1. A+I=(1−11−1) has rank 1 (block C), so there is one eigenvector v=(1,−1), and w=(1,0) solves (A+I)w=v. For y(0)=1, y′(0)=0 we get c=(0,1) and y=(1+t)e−t, y′=−te−t: the t is the defective factor. y(1)=0.73576 and y(2)=0.40601 match eAt and RK4. The velocity makes the hump of the previous paragraph, peaking in magnitude at t=1/∣λ∣=1 with 1/e=0.36788; this is the sim’s critical preset, whose readout gives the same extremum. In the (T,D) plane this is the point where the damping parameter crosses the parabola, which the last lesson found fragile: slightly more damping gives a node, slightly less an oscillating spiral.
Numerics: nearly repeated eigenvalues
Perturb the defective matrix of Example 2 to Aδ=(−3−1+δ1−1): now T2−4D=4δ, and for δ>0 the eigenvalues are −2±δ with eigenvectors (1,1±δ), nearly parallel. Block D measures the condition number of the unit-column eigenvector matrix: 19.95, 200.00 and 2000.00 for δ=10−2,10−4,10−6, close to 2/δ, and about 108 for the exactly defective matrix, where np.linalg.eig also splits the double eigenvalue into −2±2×10−8, a split that is rounding noise (this exact figure is platform dependent). The expansion coefficients blow up in the same way: with v±=(1,1±δ) and start (0,1), c1=−c2=1/(2δ)=5,50,500 (block D), and the solution is the difference of two large, nearly equal terms,
x1(t)=2δe(−2+δ)t−e(−2−δ)t→te−2t(δ→0),
the derivative of eλt with respect to λ: the t factor is the limit of two merging exponentials. In this small example the computed VeΛtV−1 still agreed with expm to about 10−14 (block D), so no disaster is shown; the worst-case bound grows like the condition number, and from memory, this is why production codes prefer the Schur decomposition (or a Padé expm), which uses orthogonal transformations and never forms a near-singular eigenvector matrix. The next lesson builds eAt without eigenvectors.
Computing it
🐍 PythonLab: complex solution against expm and RK4; defective solution, hump and Jordan form; critical damping; the conditioning of nearly repeated eigenvalues (blocks A to D)
⌘/Ctrl + Enter
Now do the central step yourself. Everything is real here, so the grader compares real numbers.
🧪 Coding exerciseExercise: generalized eigenvector, the defective solution formula, and an RK4 check✓ passed
⌘/Ctrl + Enter
💡 Everything is real: M = A - lam*np.eye(2) is singular but consistent, so lstsq returns a valid w. Reuse defective_x inside max_gap.
Concept checks
✅ Concept check
A system has eigenvalues $-0.5 \pm 2i$. By what factor does the radius shrink per full turn? (Give a number.)
Why: rho = e^(2 pi alpha / beta) = e^(-pi/2) = 0.20788. The sim shows it for alpha = -0.5, beta = 2.
✅ Concept check
Two real matrices both have eigenvalues $-0.2 \pm i$ but different eigenvectors. What differs between their phase portraits?
Why: The eigenvalue fixes the clock and the envelope; the eigenvector fixes the shape of the loop.
✅ Concept check
$A = \begin{pmatrix}-2&1\\0&-2\end{pmatrix}$ has the eigenvector $\mathbf v = (1, 0)$ and a double eigenvalue $-2$. Why is $e^{-2t}(c_1\mathbf v + c_2 t\mathbf v)$ not the general solution?
Why: With only v available, every solution stays on one line; w supplies the second dimension.
Problems
✏️ Pen & paperComplex pair by hand●●○
Solve x′=Ax, x(0)=(1,0) for A=(34−2−1). Give the period and the growth factor per turn.
Show worked solution
T=2, D=−3+8=5, λ=1±2i. Row 1 of A−λI: (2−2i)v1=2v2, so v=(1,1−i): a=(1,1), b=(0,−1). Fit c1a+c2b=(1,0): c1=1, c1−c2=0, so c2=1. Then x=et[(cos2t+sin2t)a+(cos2t−sin2t)b]=et(cos2t+sin2t,2sin2t). Check: x′(0)=(3,4)=Ax0. Period π, factor e2π⋅1/2=eπ=23.141 (a growing spiral; computed by hand and not from the sim).
✏️ Pen & paperA defective node●●○
For A=(11−13) find λ, v, a generalized w, and the solution with x(0)=(1,0).
Show worked solution
T=4, D=4, so λ=2 twice. A−2I=(−11−11)=0: v=(1,−1) only. (A−2I)w=v needs −w1−w2=1; take w=(−1,0). Fit c1(1,−1)+c2(−1,0)=(1,0): c1=0, c2=−1. So x=−e2t(tv+w)=e2t(1−t,t). Check: x′(0)=(1,1)=A(1,0). Growing λ: the same picture run backwards in time. (Hand computation; the exercise cell above verifies this matrix.)
✏️ Pen & paperCritical damping for any stiffness●●●
Show that y′′+2ky′+k2y=0 (k>0) is always a defective node, and find the largest speed ∣y′∣ for y(0)=1, y′(0)=0.
Show worked solution
The companion matrix is (0−k21−2k), with T=−2k, D=k2, T2−4D=0: double λ=−k. A+kI=(k−k21−k) has the entry 1=0, so it is nonzero: one eigenvector (1,−k) and a defective node. Take w=(1/k,0): (A+kI)w=(1,−k)=v. Then x0=(1,0)=kw, so c=(0,k) and y=(1+kt)e−kt, y′=−k2te−kt. This is largest in magnitude at t=1/k (the hump), where ∣y′∣=k/e. For k=1: 0.36788 (the sim’s critical preset).
Key Takeaways
For λ=α±iβ and v=a+ib: x1=eαt(acosβt−bsinβt) and x2=eαt(asinβt+bcosβt) are real solutions (real and imaginary parts of eλtv), and c1a+c2b=x0 fits the data.
The period is 2π/β and the factor per turn is e2πα/β (0.28461 for −0.2±i), independent of a,b, which set the tilt and stretch of the ellipse.
A repeated eigenvalue gives a star node (A=λI, every vector an eigenvector) or a defective node (one eigenvector).
In the defective case solve (A−λI)w=v; then x=eλt((c1+c2t)v+c2w), because substituting eλt(tv+w) leaves exactly that equation.
teλt peaks at t=1/∣λ∣ with height 1/(∣λ∣e) (0.5 and 0.18394 for λ=−2): a stable defective system has a transient hump.
The 2×2 Jordan block (λ01λ) has eJt=eλt(10t1) (derived in the matrix-exponential lesson); critical damping is a defective node with y=(1+t)e−t.
Near a defective matrix the eigenvector matrix has condition number about 2/δ (19.95, 200.00, 2000.00); Schur or expm is preferred (from memory).