Complex eigenvalues — rotation-scaling interpretation: Visual walkthrough

Step 1 — What "eigenvalue" even asks, and why a spinner has none

WHAT. A matrix $A$ is a machine: feed it an arrow $v$ (a "vector" — an arrow from the origin), it spits out a new arrow $A v$ . An eigenvector is a special input arrow whose output points along the same line — only stretched or flipped. The stretch factor is the eigenvalue $λ$ :

$A v = λ v .$

$A$ :: the machine (a table of numbers acting on arrows).
$v$ :: the input arrow, not allowed to be the zero arrow.
$λ$ :: how many times longer (and same/opposite direction) the output is.

WHY we care. If such a line exists, the machine is "boring" along it — it just slides arrows in and out. But some machines have no such line at all: a machine that spins every arrow. After spinning, no arrow points where it started, so there is no real $λ$ .

PICTURE. On the left, a stretch machine: the red arrow keeps its direction, only grows — that direction is an eigenvector. On the right, a spinner: every arrow tilts, so no arrow survives as an eigenvector.

Figure — Complex eigenvalues — rotation-scaling interpretation

Step 2 — Where the numbers come from: the characteristic equation

WHAT. To find $λ$ , rewrite $A v = λ v$ as $(A - λ I) v = 0$ , where $I$ is the identity (the "do-nothing" machine). A nonzero arrow $v$ gets crushed to zero only if the machine $A - λ I$ is degenerate — its determinant is $0$ :

$det (A - λ I) = 0.$

For $A = (a c b d)$ this multiplies out to

$squared term λ^{2} - trace tr A (a + d) λ + d e t A (a d - b c) = 0.$

$tr A = a + d$ :: the trace, sum of the diagonal — see Trace and Determinant.
$det A = a d - b c$ :: the determinant, the area-scaling of the machine.

WHY this shape. A $2 \times 2$ matrix gives a quadratic in $λ$ , so at most two eigenvalues. The quadratic formula gives them:

$λ = \frac{tr A \pm ( tr A ) ^{2} - 4 d e t A}{2} .$

PICTURE. The parabola $y = λ^{2} - (tr A) λ + det A$ . Where it crosses the horizontal axis are the real eigenvalues. When the parabola floats entirely above the axis it never crosses — no real roots — and that is the spinner.

Step 3 — Reading a complex $λ$ as a point, then as a spin-zoom

WHAT. When $Δ < 0$ , $Δ = - ∣Δ∣ = i ∣Δ∣$ , where $i$ is the invented number with $i^{2} = - 1$ . So the two eigenvalues are

$λ = α \pm i β, α = \frac{tr A}{2}, β = \frac{∣Δ∣}{2} .$

$α$ :: the real part (horizontal coordinate of the point $λ$ ).
$β$ :: the imaginary part (vertical coordinate).

WHY plot it. A complex number is a point in a plane. Every point in a plane can also be named by how far out it is and at what angle — its polar form (see Complex Numbers - Polar Form):

$r = ∣ λ ∣ = α^{2} + β^{2} (distance), θ = ar g λ = atan2 (β, α) (angle) .$

We use $α^{2} + β^{2}$ because that is the Pythagorean length of the arrow from the origin to $(α, β)$ — the right triangle with legs $α, β$ . We use $atan2$ (not plain $arctan$ ) because $arctan (β / α)$ cannot tell quadrant II from IV — it repeats every $180°$ . $atan2$ looks at the signs of both $α$ and $β$ to place the angle in the correct quadrant.

PICTURE. The point $λ = α + i β$ with its right triangle: horizontal leg $α$ , vertical leg $β$ , hypotenuse $r$ , opening angle $θ$ .

Step 4 — Splitting one complex equation into two real ones

WHAT. We have two eigenvalues, $α + i β$ and $α - i β$ . We deliberately grab the lower one, $λ = α - i β$ (the branch with the minus sign). Its eigenvector is also complex; write it as $v = x + i y$ where $x$ and $y$ are ordinary real arrows. Plug into $A v = λ v$ :

$A (x + i y) = (α - i β) (x + i y) .$

Expand the right side using $i^{2} = - 1$ :

$A x + i A y = (α x + β y) + i (α y - β x) .$

Because $A$ , $x$ , $y$ are all real, $A x$ and $A y$ are real. Two complex things are equal exactly when their real parts match and their imaginary parts match. That single complex equation becomes two real ones:

$A x = α x + β y, A y = - β x + α y .$

real part :: $A x = α x + β y$ — how the machine moves the arrow $x$ .
imaginary part :: $A y = - β x + α y$ — how it moves the arrow $y$ .

Intuition Why the

minus branch, and what the other one would do Nothing forces $α - i β$ — it is a choice, and here is exactly what it buys and what the other choice costs:

If we took $λ = α + i β$ instead, the very same expansion would flip the sign of $β$ throughout, giving $A x = α x - β y$ and $A y = β x + α y$ . The block would come out as $(α - β β α)$ — a rotation by $- θ$ instead of $+ θ$ . Same picture, spin reversed. It is only a relabelling ( $x \leftrightarrow$ mirror of $y$ ), not a different matrix.
The conjugate eigenvector is free. If $v = x + i y$ solves $A v = (α - i β) v$ , then its conjugate $\overset{v}{ˉ} = x - i y$ automatically solves $A \overset{v}{ˉ} = (α + i β) \overset{v}{ˉ}$ — you get the other eigenvector for free by flipping the sign of $i$ , because $A$ is real. So we never solve twice; one complex eigenvector hands us both real arrows $x, y$ .

We fix the minus branch once so that $C$ always comes out with $+ β$ in the lower-left (a positive, counter-clockwise spin) and stay consistent.

WHY this is the whole trick. Complex algebra just became two real geometric facts about two real arrows. This is the bridge from " $i$ " to "pictures."

PICTURE. The two real arrows $x$ (mint) and $y$ (coral), and where the machine $A$ sends each — a blend of $x$ and $y$ with weights $α, β$ .

Step 5 — Packaging the two facts into one matrix equation

WHAT. Put $x$ and $y$ side by side as the columns of a matrix $P = (x y)$ . Then "apply $A$ to each column" is just $A P$ , whose columns are $A x$ and $A y$ . Our two boxed facts say those columns are $α x + β y$ and $- β x + α y$ . But those are exactly what you get by multiplying $P$ on the right by a little $2 \times 2$ table:

$A P = P (α β - β α) .$

top-left column of the little table $(α, β)$ :: builds $α x + β y = A x$ . ✓
right column $(- β, α)$ :: builds $- β x + α y = A y$ . ✓

Intuition Why

$P$ really is invertible here — $x$ and $y$ can't be parallel We must be sure the columns $x, y$ of $P$ are not parallel, or $P^{- 1}$ would not exist. Suppose, to see the contradiction, that they were parallel: $y = k x$ for some real number $k$ . Then the complex eigenvector would be $v = x + i y = x + i (k x) = (1 + ik) x,$ a real arrow $x$ scaled by the complex number $1 + ik$ . But $A v = (α - i β) v$ with $β \neq = 0$ would force $A x = (α - i β) (1 + ik)^{- 1} \dots$ to have a genuinely complex direction — impossible, since $A x$ is a real arrow (real matrix times real arrow). Concretely: a real matrix can never send a real arrow to a complex multiple of itself unless the multiple is real, and $α - i β$ is not real when $β \neq = 0$ . So $y = k x$ is impossible: $x$ and $y$ must point in different directions. Therefore $P$ is invertible and $P^{- 1}$ exists. ✓

Multiply both sides on the right by $P^{- 1}$ (which we just proved exists):

$A = P C P^{- 1}, C = (α β - β α) .$

WHY. $A = P C P^{- 1}$ means: change into the $x, y$ frame ( $P^{- 1}$ ), do $C$ , change back ( $P$ ). This is exactly Diagonalization — but instead of a diagonal, we land on the block $C$ . All the interesting geometry lives in $C$ .

PICTURE. The sandwich: input frame → $P^{- 1}$ untwists to the $x, y$ coordinates → $C$ acts cleanly → $P$ retwists back.

Step 6 — Why $C$ is rotate-then-scale

WHAT. Look at $C = (α β - β α)$ . Recall $α = r cos θ$ and $β = r sin θ$ from Step 3's triangle. Substitute and pull out the common $r$ :

$C = (r cos θ r sin θ - r sin θ r cos θ) = r pure rotation by θ (cos θ sin θ - sin θ cos θ) .$

the matrix in the brace :: the standard rotation matrix — it turns every arrow by $θ$ without changing length (see Rotation Matrices).
the factor $r$ out front :: uniform scaling — every arrow's length multiplied by $r$ .

So $C =$ rotate by $θ$ , then scale by $r$ . And $r = ∣ λ ∣$ , $θ = ar g λ$ — the modulus and argument of the eigenvalue itself.

WHY it must factor. Any table of that special pattern (equal diagonal $α$ , opposite off-diagonals $\mp β$ ) is $r$ times a rotation, guaranteed, because $(α, β)$ can always be written in polar form. There is nothing to check case-by-case — the algebra forces it.

PICTURE. A grid rotated by $θ$ and blown up by $r$ in one shot: the unit circle becomes a larger circle turned by $θ$ .

Step 7 — Every case: $r < 1$ , $r = 1$ , $r > 1$ , and the degenerate corners

WHAT. The eigenvalue lives somewhere in the plane; its distance $r$ from the origin decides the shape of repeated application $A, A^{2}, A^{3}, \dots$ (a discrete Linear Dynamical Systems — its Phase Portraits):

| Where $λ$ sits | $r = ∣ λ ∣$ | Each step does | Orbit shape | |---|---|---|---| | Inside unit circle | $r < 1$ | spin + shrink | spiral inward | | On unit circle | $r = 1$ | spin only | closed ellipse | | Outside unit circle | $r > 1$ | spin + grow | spiral outward |

Degenerate corners:

$β = 0$ :: not complex at all — the two eigenvalues collapse to a single real $λ = α$ , and $C$ collapses to $λ I = (α 0 0 α)$ : pure scaling by $α$ along invariant lines, no coupling of $x$ and $y$ . If $α > 0$ every arrow just lengthens/shortens in place; if $α < 0$ each eigen-line is flipped through the origin (a $180°$ turn on that line) as well as scaled — the one "spin-like" corner, but it is a reflection through $0$ , not a genuine rotation of the plane.
$α = 0, β \neq = 0$ :: $θ = \pm 90°$ , pure quarter-turn scaled by $r$ (Worked Example 1 was $r = 1$ , $θ = 90°$ ).
$λ = 0$ :: $r = 0$ , the machine collapses arrows to the origin.

WHY $r$ and not $α$ for iteration. After $n$ steps, lengths multiply by $r^{n}$ . If instead you flow continuously ( $\overset{v}{˙} = A v$ ), growth is set by the real part $α = Re λ$ — a different question, don't confuse them.

PICTURE. Three orbits side by side: inward spiral ( $r < 1$ ), ellipse ( $r = 1$ ), outward spiral ( $r > 1$ ), each starting from the same seed arrow.

The one-picture summary

The complex eigenvalue $λ = α - i β$ , plotted as a point, is the instruction. Its length $r$ is how much each step scales; its angle $θ$ is how much each step spins. The matrix $P$ (columns $x, y$ from the eigenvector) is just the frame in which this clean spin-zoom happens; sandwiching gives $A = P C P^{- 1}$ .

Recall Feynman retelling — the whole walkthrough in plain words

We asked a matrix, "is there an arrow you only stretch?" A spinner said "no real one." So we solved for the stretch factor anyway and got a complex number — a point off the real line. We plotted that point. Its distance from the center told us how much the spinner grows or shrinks each turn; its angle told us how far it turns. To prove it, we wrote the complex eigenvector as two real arrows $x$ and $y$ , and the one complex equation split into two honest real ones about how the machine pushes $x$ and $y$ . Those two arrows can never point the same way (or the eigenvector would be a real arrow, which a real spinner can't have), so we can safely stack them into $P$ and undo it with $P^{- 1}$ — turning everything into $A = P C P^{- 1}$ , where $C$ is nothing but "rotate by the angle, scale by the distance." Then we walked the three fates — spiral in, loop forever, spiral out — plus the flat corners where the spin vanishes. That's the entire secret: a complex eigenvalue is a spin-and-zoom written in disguise.