Run forward Euler on a frictionless pendulum and something unsettling happens: the pendulum swings a little higher every period, gaining energy from nowhere, until the simulation is visibly wrong. Or linearize the flight equations of an airliner and you find a slow, lazy oscillation in which the aircraft trades altitude for airspeed, again and again. Both behaviors are rotations, and the eigenvector machinery of 1.5 Eigenvectors & Eigenvalues and 1.6 Change of Basis & Diagonalization seems to break down on rotations: a matrix that turns the plane leaves no direction on its own line, so it has no real eigenvectors at all. The fix is to let the eigenvalues be complex numbers. Once we do, rotation, oscillation and spirals become just another eigenvalue story, and we can say exactly when a spiral grows or decays, and by how much per step.
▶ VideoPart 1: why rotations need complex numbers, and Euler's formulacaptions only
Transcript
Some matrices have no real eigenvectors at all. They rotate every direction. To understand them we need complex numbers, and the payoff is spirals.
Here is a matrix that rotates the plane by sixty degrees. An eigenvector would have to stay on its own line. Watch the amber image as I sweep v around.
It never lands back on the dashed line. Every direction gets turned, so there is no real eigenvector, and no real eigenvalue.
The algebra agrees. The characteristic equation is lambda squared minus lambda plus one equals zero, and its discriminant is negative. The roots are complex.
Quick refresher. A complex number a plus b i is just a point in the plane: a steps along the real axis, b steps up the imaginary axis. Here is three plus two i.
Every complex number also has a length, the modulus r, and an angle, the argument theta. So we can write it in polar form: r times cosine theta plus i sine theta.
The conjugate, z bar, flips the sign of the imaginary part. It is the mirror image across the real axis, with the same length and the opposite angle.
Now the key fact. Multiplying by i turns one into i, and i into minus one. It rotates the entire plane by ninety degrees.
Multiplying by one plus i does more. Its length is root two and its angle is forty five degrees, so it rotates the plane by forty five degrees and stretches it by root two.
In general, multiplying by a complex number multiplies the lengths and adds the angles. Every complex multiplication is a rotation combined with a scaling.
And here is the bridge to matrices. Multiply a plus b i by x plus i y, and read off the real and imaginary parts. That is exactly the matrix a, minus b, b, a, acting on the vector x, y.
The single most useful formula in this subject is Euler's formula: e to the i theta equals cosine theta plus i sine theta.
As theta grows, the point e to the i theta walks around the unit circle at unit speed. Its shadow on the real axis is cosine, its shadow on the imaginary axis is sine.
Why is it true? Take the Taylor series of e to the x from Calc Two and plug in i theta. Powers of i cycle through one, i, minus one, minus i.
Collect the real terms and you get the cosine series. Collect the imaginary terms and you get the sine series. That is Euler's formula.
So any complex number is r times e to the i theta, and multiplying by it scales by r and rotates by theta. Powers are easy: lambda to the n is r to the n, times e to the i n theta.
Let's recap. A complex number is a point with a length and an angle. Multiplying by it rotates and scales the plane, and Euler's formula packs the angle into an exponential. In part two, we put this to work on real matrices and the spirals they make.
When eigenvectors disappear
Take the matrix that rotates the plane counterclockwise by an angle θ:
R(θ)=[cosθsinθ−sinθcosθ].
Its columns are where ^ and ^ land after turning by θ (look back at 1.2 Matrices as Transformations if that reading of columns feels unfamiliar). Unless θ is 0 or π, every arrow gets knocked off its line, so there is no real eigenvector. The characteristic polynomial agrees:
det(R−λI)=(cosθ−λ)2+sin2θ=λ2−2cosθλ+1.
The discriminant is 4cos2θ−4=−4sin2θ, which is negative whenever sinθ=0. The quadratic formula still produces roots, but they involve −1:
λ=cosθ±isinθ.
We could declare “no eigenvalues” and move on. But those roots carry real information: cosθ and sinθ are right there. To read it, we need to be comfortable with complex numbers.
Complex numbers, refreshed
A complex number is z=a+bi with a,b real and i2=−1. We call a=Rez the real part and b=Imz the imaginary part. The key mental move: a complex number is a point in the plane, (a,b), drawn in the complex plane with the real axis horizontal and the imaginary axis vertical. So 3+2i sits at (3,2), i sits at (0,1) and −1 at (−1,0).
Arithmetic. Addition is vector addition: (3+2i)+(1−4i)=4−2i. Multiplication uses the distributive law and i2=−1:
(a+bi)(c+di)=ac+adi+bci+bdi2=(ac−bd)+(ad+bc)i.
Modulus and argument. Like any arrow, z has a length and a direction. The modulus is ∣z∣=r=a2+b2, and the argumentθ=argz is the angle from the positive real axis, so a=rcosθ and b=rsinθ. This gives the polar form
z=r(cosθ+isinθ).
For z=3+2i: r=13≈3.606 and θ=arctan(2/3)≈33.7∘. (Watch the quadrant: −3−2i has the same ratio b/a but θ≈213.7∘. In code, use atan2(b, a).)
Conjugate. The conjugatezˉ=a−bi is the mirror image across the real axis: same modulus, opposite angle. Two facts make it useful:
zzˉ=(a+bi)(a−bi)=a2+b2=∣z∣2,z1=∣z∣2zˉ.
The second is how you divide: 1+i1=21−i.
Multiplication is rotate-and-scale
Here is the fact that makes everything in this lesson work. Multiply two numbers in polar form:
The trig addition formulas from precalculus did all the work. So:
Now write that multiplication as a matrix. If w=a+bi and z=x+iy, then wz=(ax−by)+i(bx+ay), which in vector form is
[ax−bybx+ay]=[ab−ba][xy].
Multiplication by the complex number a+bi is the real linear map with matrix [ab−ba]. Keep this matrix in mind; it’s the star of the rest of the lesson.
✅ Concept check
Multiplying every complex number by $w = 2i$ does what to the complex plane?
Why: $1 \to 2i$ and $i \to 2i^2 = -2$: both basis arrows turn a quarter turn and double in length.
Euler’s formula
Polar form is clumsy to write. Euler’s formula packs it into one exponential:
There are two good ways to see why this is the right definition of an exponential with an imaginary exponent.
The real terms form the cosine series and the imaginary terms form the sine series. (The series converge absolutely for every θ, so regrouping is legal.)
Reason 2: motion on a circle. The defining property of ekt is dtdekt=kekt. Ask for that with k=i: a point z(θ) with z(0)=1 whose velocity is z′(θ)=iz(θ). Multiplying by i rotates by 90∘, so the velocity is always perpendicular to the position, and it has the same length. Motion that is always perpendicular to the radius is motion on a circle. Starting at 1 with unit speed, after time θ you’ve gone arc length θ around the unit circle, which puts you at (cosθ,sinθ). As θ grows, eiθ goes around the unit circle; its shadow on the real axis is cosθ and its shadow on the imaginary axis is sinθ.
Consequences. Every complex number can be written z=reiθ, and the multiplication rule becomes the law of exponents:
Powers are easy in polar form: raise the modulus to the nth power and multiply the angle by n. Adding eiθ and e−iθ also gives the handy formulas cosθ=21(eiθ+e−iθ) and sinθ=2i1(eiθ−e−iθ), which turn trig identities into algebra.
✏️ Pen & paperComplex arithmetic warm-up●○○
Let z=1+i3 and w=2i.
Find ∣z∣, argz, and write z in the form reiθ. Do the same for w.
Compute zw in polar form, then convert to a+bi. Check by multiplying directly.
Compute zˉ and 1/z.
Compute z6 without expanding.
Show worked solution
∣z∣=1+3=2 and argz=arctan(3)=π/3, so z=2eiπ/3. For w: ∣w∣=2, argw=π/2, so w=2eiπ/2.
zˉ=1−i3, and 1/z=zˉ/∣z∣2=41(1−i3). In polar form, 1/z=21e−iπ/3: half the length, opposite angle.
z6=26ei6π/3=64e2πi=64. Six turns of 60∘ bring you all the way around to the positive real axis.
Complex eigenvalues of real matrices
Back to matrices. For any real 2×2 matrix A with trace T=a11+a22 and determinant D=detA, you saw in 1.5 Eigenvectors & Eigenvalues that the characteristic polynomial is
λ2−Tλ+D=0⟹λ=2T±2T2−4D.
When T2<4D, the square root is of a negative number and
λ=2T±i24D−T2=a±bi.
The identity ∣λ∣2=D is worth remembering: the product of the eigenvalues is the determinant, and here the product is λλˉ=∣λ∣2. So for complex eigenvalues, ∣λ∣=detA, which you can read off without solving anything. The conjugate-pair rule holds for real matrices of any size, because the characteristic polynomial has real coefficients.
Worked example: eigenvalues from trace and determinant
Find the eigenvalues of A=[34−2−1] and their modulus and angle.
T=3+(−1)=2 and D=(3)(−1)−(−2)(4)=−3+8=5. Since T2=4<20=4D, the eigenvalues are complex:
λ=22±i220−4=1±2i.
Then r=∣λ∣=1+4=5≈2.236 (check: D=5 ✓), and θ=arctan(2/1)≈63.43∘. We’ll see shortly that this means: in a suitable coordinate system, A rotates by about 63.4∘ and stretches by 5. We’ll finish this example below by finding that coordinate system.
✅ Concept checkselect all that apply
Select every matrix with non-real (complex) eigenvalues.
Why: Compute T and D, then compare T² with 4D. No eigenvectors needed.
▶ VideoPart 2: rotation-scaling matrices, spirals, and the real form A = PCP⁻¹captions only
Transcript
In part one we saw that complex multiplication rotates and scales. Now we find real matrices that do exactly the same thing, and watch what happens when we iterate them.
Back to real matrices. The matrix a, minus b, b, a is r times a rotation by theta. Its eigenvalues are a plus and minus b i: the very complex number it imitates, and its conjugate.
For any real two by two matrix, the eigenvalues solve lambda squared minus trace times lambda plus determinant. When trace squared is less than four times the determinant, they come as a complex conjugate pair.
Now iterate. Start at a point and apply the rotation scaling matrix over and over. With r equal to point nine two, each step turns twenty five degrees and shrinks the distance by eight percent.
The orbit spirals inward. After n steps the distance is r to the n times the start, so magnitude less than one means decay to zero.
Make r one point zero five and the same spiral runs outward. Make r exactly one and the points stay on a circle forever. The modulus of the eigenvalue is the whole story.
A general real matrix with complex eigenvalues does the same thing, just in a tilted coordinate system. Here the orbit spirals around ellipses instead of circles.
The ellipses are circles seen through the matrix P, whose columns are the real and imaginary parts of a complex eigenvector. In that basis, A is a pure rotation scaling: A equals P C P inverse.
The real reason this matters is differential equations. For x prime equals A x with eigenvalues alpha plus or minus i beta, solutions look like e to the alpha t times cosine and sine of beta t.
Beta sets how fast it turns. Alpha sets whether it grows or decays. With alpha negative you get a smooth inward spiral, which is a damped oscillator: a mass on a spring with friction.
Aircraft have these modes too. The slow phugoid, where the plane trades height for speed, and the Dutch roll, a yawing and rolling wobble, are each a complex eigenvalue pair of the linearized flight equations.
Here is a connection to your simulations course. Take the undamped oscillator, whose true orbits are circles, and step it with forward Euler.
The Euler matrix is a rotation scaling matrix with eigenvalues one plus or minus i h. Their modulus is the square root of one plus h squared, which is bigger than one. So every step spirals outward.
That is the energy drift you saw when Euler made a pendulum swing higher and higher. It is not a bug in your code. It is an eigenvalue outside the unit circle.
Finally, fluids. Flow past a cylinder above a critical speed starts shedding vortices, alternating top and bottom, in the famous von Karman street.
Linearize the flow around the steady state and you find a complex eigenvalue pair. The imaginary part is the shedding frequency. The real part crossing zero is the moment the wake starts to oscillate.
To recap. Complex multiplication is rotation and scaling. Complex eigenvalues of a real matrix mean rotation, and their modulus decides between inward spirals, outward spirals, and closed orbits.
Next time, we let time flow continuously, and meet the matrix exponential, which turns these step by step spirals into smooth ones.
Rotation-scaling matrices
The simplest matrices with complex eigenvalues are the ones we met as “multiplication by a+bi”:
C=[ab−ba].
Write a=rcosθ and b=rsinθ, where r=a2+b2. Then
C=r[cosθsinθ−sinθcosθ]=rR(θ):
a rotation by θ followed by a uniform scaling by r. That’s why C is called a rotation-scaling matrix. Its eigenvalues are
det(C−λI)=(a−λ)2+b2=0⟹a−λ=±bi⟹λ=a∓bi=re∓iθ.
So the eigenvalues of C are exactly the complex number a+bi that C imitates, and its conjugate. The modulus of the eigenvalue is the scaling factor and the argument is the rotation angle. The eigenvectors are complex: you can check that C[1−i]=[a+bib−ai]=(a+bi)[1−i], and conjugating gives the eigenvector (1,i) for a−bi.
Powers of C are as easy as powers of complex numbers. Rotations compose by adding angles, so
Cn=rnR(nθ),
the matrix version of (reiθ)n=rneinθ.
✏️ Pen & paperRotations, powers and eigenvectors●●○
Show that the rotation matrix R(θ) has eigenvalues e±iθ, and that (1,−i) is an eigenvector for eiθ.
Compute M8 for M=[11−11] without multiplying eight matrices.
Why must a rotation-scaling matrix with θ=2π/k and r=1 satisfy Ck=I? What does the orbit of a point look like?
Show worked solution
From the first section, det(R−λI)=λ2−2cosθλ+1, with roots cosθ±isinθ=e±iθ. For the eigenvector, R[1−i]=[cosθ+isinθsinθ−icosθ]. The second entry is −i(cosθ+isinθ), because −icosθ−i2sinθ=sinθ−icosθ. So the result is eiθ(1,−i). ✓
M is the rotation-scaling matrix for 1+i=2eiπ/4, so M=2R(45∘) and M8=(2)8R(360∘)=16I. This is the matrix version of (1+i)8=16.
Ck=1kR(k⋅2π/k)=R(2π)=I. Every point returns to where it started after k steps, so the orbit is k equally spaced points on a circle: a regular k-gon. (You saw the 12-gon for θ=30∘ in the video.)
Spirals from iterating a matrix
Now iterate: pick a starting vector x0 and repeatedly apply C,
xn+1=Cxn⟹xn=Cnx0=rnR(nθ)x0.
Each step turns the point by θ and multiplies its distance from the origin by r, so
∥xn∥=rn∥x0∥.
The modulus of the eigenvalue is the whole story:
∣λ∣=r
orbit of xn+1=Cxn
r<1
spirals inward to the origin (stable)
r=1
stays on a circle forever; closed if θ/2π is rational
r>1
spirals outward to infinity (unstable)
The angle θ only sets how tightly wound the spiral is: about 360∘/θ steps per turn.
◆ InteractiveSpiral map: iterate xₙ₊₁ = A xₙ and watch the eigenvalues
Loading simulation…
Things to try:
In rotation-scaling mode, slide r slowly through 1. The orbit switches from an inward to an outward spiral exactly as the red eigenvalue dots in the inset cross the dashed unit circle. The bottom readout compares ∥xn∥ with rn∥x0∥; they agree exactly.
Press r = 1 and set θ=30∘, then θ=72∘. You get a 12-gon and a pentagon. Try θ=37∘: the orbit does not close until step 360, so within 200 steps it fills in the circle instead.
Switch to General A = P C P⁻¹ and change the tilt and length of the columns of P. Same eigenvalues, same inward or outward behavior, but the spiral now runs along ellipses. That’s the next section.
Switch to Any 2×2 matrix and drag a12 from negative values up past zero. When T2−4D turns positive, the eigenvalue dots collide on the real axis and split apart, and the spiral stops spiraling.
✅ Concept check
You iterate $\vec x_{n+1} = A\vec x_n$ with $A = \begin{bmatrix} 0.8 & -0.6 \\ 0.6 & 0.8 \end{bmatrix}$. What does the orbit do?
Why: det A = 1, so |λ| = 1: a pure rotation. The entries 0.8 and 0.6 are cos and sin of about 36.87°.
The real form: A = P C P⁻¹
A general real matrix with complex eigenvalues isn’t a rotation-scaling matrix; the [34−2−1] from before certainly doesn’t look like one. But in the right basis it is one. This is the complex-eigenvalue analogue of diagonalization, and like 1.6 Change of Basis & Diagonalization it’s a change of basis. The difference is that we can’t use the complex eigenvectors as a real basis directly. Instead we split one eigenvector into its real and imaginary parts.
Derivation. Write v=x+iy with x=Rev and y=Imv real vectors. Expand Av=λv:
Ax+iAy=(a−bi)(x+iy)=(ax+by)+i(−bx+ay).
Since A, x, y are real, match real and imaginary parts:
Ax=ax+by,Ay=−bx+ay.
Read column by column, that says A[xy]=[xy][ab−ba], i.e. AP=PC. Finally, x and y are linearly independent: if y=cx, then v=(1+ic)x would be a complex multiple of a real vector, x would itself be an eigenvector, and Ax=λx with Ax real forces λ real, which is a contradiction. (The case x=0 works the same way with y.) So P is invertible and A=PCP−1. ■
Geometric meaning.P−1 moves to coordinates along Rev and Imv. In those coordinates A acts as C, rotating by θ and scaling by r=∣λ∣, and P moves back. The circles that C rotates along become ellipses in the original coordinates. So every real 2×2 matrix with complex eigenvalues is a rotation-scaling in disguise. Iterating it gives spirals along a family of nested ellipses, inward if ∣λ∣<1 and outward if ∣λ∣>1. Also, An=PCnP−1=rnPR(nθ)P−1, so ∥xn∥ stays within a constant factor of rn∥x0∥. The constant is the ratio of the ellipses’ longest to shortest axes.
Worked example: the full real form
Write A=[34−2−1] as PCP−1 and describe the iteration xn+1=Axn.
Step 1: eigenvalues. From before, λ=1±2i. Take λ=1−2i, so a=1 and b=2.
Step 2: eigenvector. Solve (A−λI)v=0:
A−(1−2i)I=[2+2i4−2−2+2i].
The rows of a singular 2×2 matrix are proportional, so the first row is enough: (2+2i)v1−2v2=0 gives v2=(1+i)v1. Take v1=1:
v=[11+i].
Check with the second row: 4(1)+(−2+2i)(1+i)=4+(−2−2i+2i+2i2)=4−4=0. ✓
Step 5: interpret.C=5R(63.43∘). In the skewed coordinates along (1,1) and (0,1), every step rotates by 63.43∘ and stretches by 5. In the original plane, iterates spiral outward along ellipses, taking a little under 6 steps per turn. The video used A/2.4, which has the same P and the same angle but ∣λ∣=5/2.4≈0.93, so it spirals inward along the same ellipses.
✏️ Pen & paperReal form, start to finish●●○
Let A=[11−23].
Find the eigenvalues, r=∣λ∣ and θ.
Find P and C with A=PCP−1, and verify your answer by multiplying.
Describe the orbit of xn+1=Axn starting from any nonzero x0.
Show worked solution
T=4 and D=3−(−2)=5. Since 16<20, λ=2±i20−16/2=2±i. So r=5≈2.236 and θ=arctan(1/2)≈26.57∘.
Take λ=2−i (a=2, b=1). Then A−λI=[−1+i1−21+i]. The second row gives v1+(1+i)v2=0; with v2=1, v=(−1−i,1). Check the first row: (−1+i)(−1−i)−2=(1+i−i−i2)−2=2−2=0. ✓ So
Rev=(−1,1), Imv=(−1,0), and
P=[−11−10],P−1=[0−11−1],C=[21−12].
(detP=0−(−1)=1.) Verify: PC=[−32−1−1] and PCP−1=[0+10+1−3+12+1]=[11−23]. ✓ A different eigenvector scaling gives a different P, which is also correct; the check PCP−1=A is what matters.
Since r=5>1, the orbit spirals outward along ellipses centered at the origin. Each step advances 26.57∘ in P-coordinates (about 13.5 steps per turn), and the size grows by roughly 5 per step. The origin is an unstable spiral point for this map.
The imaginary part β is the angular frequency; the real part α is the growth or decay rate. In continuous time the dividing line is α=0 (the imaginary axis) instead of ∣λ∣=1 (the unit circle). That’s because one time-step of length h of this flow is the matrix eαhR(βh), with eigenvalue modulus eαh, which is less than 1 exactly when α<0.
The damped oscillator x′′+2ζωx′+ω2x=0 written as a system has characteristic equation λ2+2ζωλ+ω2=0, with roots λ=−ζω±iω1−ζ2 when ζ<1. In the (x,x′) phase plane, every underdamped mass-spring system, RLC circuit and pendulum near rest is a decaying spiral. That’s 5.3 Mechanical Vibrations: Spring–Mass–Damper.
✏️ Pen & paperSwirl or strain?●●●
In a 2D incompressible flow the velocity-gradient matrix G=∇u has trace zero (that’s ∇⋅u=0). Near a point, relative motion follows x′=Gx. For each G, find the eigenvalues and decide whether nearby fluid swirls (closed elliptical paths) or is strained (stretched along one direction and squeezed along another).
G1=[01−20],G2=[102−1],G3=[0010].
Then explain why G3, a simple shear flow u=(y,0), has nonzero vorticity but no swirl.
Show worked solution
With T=0, the characteristic equation is λ2+D=0, so the eigenvalues are complex exactly when D>0.
G1: D=0−(−2)(1)=2>0, so λ=±i2. Purely imaginary means α=0: particles circulate on closed ellipses (stretched along x1, since ∣−2∣>∣1∣) with angular frequency 2. Swirl. It’s a vortex core.
G2: D=−1<0, so λ=±1, real. Fluid is stretched along the eigenvector for +1, namely (1,0), and compressed along the eigenvector for −1, namely (1,−1). Strain, a saddle.
G3: D=0, so λ=0,0, real and repeated. Particles slide along horizontal lines at speed y; nobody goes around anything. No swirl.
For G3 the vorticity is ωz=∂v/∂x−∂u/∂y=0−1=−1=0. Its rotation part Ω does have eigenvalues ±i/2, but the strain part S=21(G+GT)=[01/21/20] is exactly strong enough to cancel the rotation. Vorticity measures the rotation part of G; swirl needs rotation to beat strain, which is the condition that G itself has complex eigenvalues. That’s why vortex-detection criteria use the eigenvalues of G rather than vorticity alone: boundary layers are full of vorticity but contain no vortices.
✏️ Pen & paperHow fast does Euler's energy drift?●●○
A 1 Hz oscillator (ω=2π) is integrated with forward Euler at h=0.01 s.
Show that the energy En=21(vn2+ω2xn2) satisfies En+1=(1+h2ω2)En exactly.
How many steps until the energy doubles? How many oscillation periods is that?
Show worked solution
Let yn=(ωxn,vn). The Euler update xn+1=xn+hvn, vn+1=vn−hω2xn becomes
[ωxn+1vn+1]=[1−hωhω1][ωxnvn],
a rotation-scaling matrix with r=1+h2ω2. So ∥yn+1∥=r∥yn∥, and since En=21∥yn∥2, En+1=r2En=(1+h2ω2)En.
hω=0.02π≈0.06283, so the factor is 1+0.003948=1.003948. Doubling needs n=ln2/ln(1.003948)≈0.6931/0.003940≈176 steps, which is 1.76 s, less than two periods. Halving h only roughly doubles that time (the per-step growth is O(h2) but there are 1/h steps per unit time), which is why Euler is hopeless for long oscillator runs.
Experiments in code
First, see the rn law numerically. Change r and theta and watch the spiral and the ratio.
🐍 PythonIterating a rotation-scaling matrix
⌘/Ctrl + Enter
Now implement the key computations yourself, from trace and determinant only.
🧪 Coding exerciseExercise: complex eigenvalues and the real form✓ passed
⌘/Ctrl + Enter
💡 T = A[0,0] + A[1,1], D = det. For real_form: lam = a - b*1j with a = T/2, b = sqrt(4D - T^2)/2. The first row of (A - lam I) v = 0 is (A[0,0] - lam) v1 + A[0,1] v2 = 0, so v = (A[0,1], lam - A[0,0]) works (A[0,1] is never 0 when the eigenvalues are complex).
Finally, measure Euler’s outward spiral and confirm the 1+h2ω2 factor.
🧪 Coding exerciseExercise: forward Euler on an oscillator spirals out✓ passed
⌘/Ctrl + Enter
💡 Loop n times with x, v = x + h*v, v - h*omega**2*x (update both from the old values). The amplitude is np.sqrt(omega**2 * xs**2 + vs**2).
✅ Concept check
Forward Euler on $x'' + x = 0$ with $h = 0.1$: by what factor does the amplitude $\sqrt{x^2 + v^2}$ grow each step? (3 decimals)
Why: The Euler matrix [[1, h], [−h, 1]] has eigenvalues 1 ± 0.1i, modulus √1.01 ≈ 1.00499.
Key Takeaways
A complex number a+bi=reiθ is a point in the plane. Multiplying by it rotates by θ and scales by r, and as a real map it is the matrix [ab−ba].
Euler’s formulaeiθ=cosθ+isinθ follows from the Taylor series, and from the fact that z′=iz means moving on the unit circle. It makes powers easy: (reiθ)n=rneinθ.
A real 2×2 matrix has complex eigenvalues exactly when T2<4D. They come in conjugate pairsa±bi with a=T/2 and ∣λ∣=D.
Rotation-scaling matricesC=rR(θ) have eigenvalues re±iθ, and Cn=rnR(nθ).
Iterating xn+1=Axn with complex eigenvalues gives spirals: inward if ∣λ∣<1, outward if ∣λ∣>1, closed orbits on a circle or ellipse if ∣λ∣=1.
Real form: with λ=a−bi and eigenvector v, A=PCP−1, where P=[RevImv]. Every such matrix is a rotation-scaling in skewed coordinates, and its circles become ellipses.
In continuous time, λ=α±iβ gives eαt(cosβt,sinβt): damped oscillators, phugoid and Dutch roll modes, and the onset of vortex shedding.
Forward Euler on an undamped oscillator has eigenvalues 1±ihω with ∣λ∣=1+h2ω2>1, which explains its energy drift.
Finished this lesson?
Mark it complete to track your progress (saved in this browser). Reading every section does this for you.