Eigenvalues and Eigenvectors: An Intuitive Introduction

Take any matrix and multiply it by a vector. What happens to the vector?

Usually, two things at once: it gets stretched, and it gets rotated. It ends up pointing somewhere new, with a new length. That's just what matrices do, they twist space around.

But here's the question worth sitting with: is there ever a direction that only stretches, and never rotates?

Turns out, yes. Almost every matrix has a few of these special directions. Find them, and the matrix stops looking like a tangle of numbers and starts looking like something you can actually picture.

Picture the matrix as a stretchy sheet

Imagine a rubber sheet with a woven grain in it, like fabric. Grab the sheet and pull it diagonally. Most of the threads tilt as well as stretch. But the threads running exactly along the grain lines? They just get longer. They don't tilt at all.

A matrix does the same thing to space. Most vectors get pushed off their own line. But a handful of special directions, the grain of the transformation, only get scaled.

We call those directions eigenvectors. We call the amount they get scaled by the eigenvalue.

Let's find them by hand, with a real matrix.

Why most directions get knocked off course

Take the matrix

Pick a vector $v$ and compute $Av$. Do this for a few random directions and you'll notice $Av$ almost never points along the same line as $v$. It swings off to one side.

Drag $v$ all the way around the circle. Watch the angle between $v$ and $Av$ in the readout below. For most of the circle, that angle is some annoying nonzero number.

But at exactly four points (really just two lines through the origin), the angle snaps to zero. At those points, $Av$ lands right back on top of $v$'s own line. That's not a coincidence, and it's not special to this matrix. Every square matrix has directions like this (sometimes complex ones, more on that later).

Naming what just happened

So we write the special condition like this:

In words: applying $A$ to $v$ gives you back the same vector $v$, just scaled by some number $\lambda$. No rotation, no shearing, just a stretch (or a shrink, or a flip if $\lambda$ is negative).

$v$ is the eigenvector. $\lambda$ is the eigenvalue. For our matrix $A$, the two eigen-directions you found by dragging are $(1,1)$ and $(1,-1)$, with eigenvalues $3$ and $1$. Check it yourself: $A(1,1) = (3,3) = 3(1,1)$. Exactly what the readout showed.

What if you don't already know the eigenvector?

Dragging until the angle hits zero works, but it's not exactly a method. So how do we find eigenvectors without guessing?

Here's a neat trick: start with literally any vector, and just keep applying $A$ to it, over and over.

Press play. The starting vector isn't an eigenvector, it's nowhere close. But watch what happens after a few steps: it gets pulled toward the $(1,1)$ direction anyway. Every application of $A$ nudges it a little closer to the grain, because the stretch along that direction (factor of $3$) dominates the stretch along the other eigen-direction (factor of $1$).

Notice the estimated eigenvalue in the readout climbing toward $3.0$ as $k$ grows. This is called power iteration, and it's not just a classroom trick. Search engines, recommendation systems, and vibration analysis in bridges all lean on this exact idea: repeat a transformation enough times, and whatever survives is the dominant eigenvector.

Finding eigenvalues without a picture

Dragging and iterating both work, but they're visual. What's the actual algebra?

Rearrange $Av = \lambda v$ like this:

For this to have a solution where $v$ isn't just the zero vector, the matrix $(A - \lambda I)$ has to be "squishing," meaning it collapses some direction to zero. That only happens when its determinant is zero:

Expand that determinant for a 2x2 matrix and you get a plain quadratic in $\lambda$, called the characteristic polynomial. Its roots are the eigenvalues.

Drag $d$, one entry of the matrix. Watch the parabola shift, and watch its two roots slide along the $\lambda$-axis. Those roots are the eigenvalues. No matrix algebra required to see them, just where a curve crosses zero.

And notice something else: no matter how far you drag $d$, this parabola always crosses zero twice. It never lifts off the axis entirely. That's not luck, it's because this matrix is symmetric (the numbers mirror across the diagonal), and symmetric matrices are guaranteed real eigenvalues. Not every matrix gets that guarantee.

When there's no real grain at all

So what happens with a matrix that isn't symmetric? Specifically, what happens with a matrix that rotates?

A pure rotation matrix turns every vector by the same angle. If it turns every vector, there's no direction left that stays on its own line. No real eigenvector survives.

Set $\theta$ to $0$ and the trail sits still on one point, run through the exact same story as before. Now slide $\theta$ up. The trail stops converging to a line and starts spiraling instead, because every step both rotates and scales the vector.

The eigenvalues haven't vanished, they've just gone complex: $\lambda = a \pm bi$. The real part $a$ tells you how the spiral scales, the imaginary part $b$ tells you how fast it turns. There's no contradiction here, just a reminder that "eigenvector" sometimes means a direction in the complex plane, not a direction you can literally draw with a pencil.

The short version

An eigenvector is a direction a matrix doesn't rotate, it only stretches. The eigenvalue is exactly how much stretching happens along that direction.

You can find these directions by dragging until $Av$ lines up with $v$. You can find them by repeatedly applying the matrix and watching where any vector ends up, the dominant one always wins. Or you can find them algebraically, as the roots of the characteristic polynomial, $\det(A - \lambda I) = 0$.

And sometimes, when a matrix only rotates, there's no real direction left to find at all, just a pair of complex eigenvalues describing a spiral.

Once you see eigenvectors as "the directions that don't rotate," a huge amount of linear algebra stops being symbol pushing and starts being geometry.

All visualizations are interactive React components running entirely in your browser, built with plain SVG and no external libraries. The power-iteration and complex-eigenvalue animations use requestAnimationFrame to step forward; everything else responds instantly to drags and slider changes.