Summary: Eigenvectors and eigenvalues

When a transformation moves the plane, most vectors get knocked off their own line. A few stubborn ones do not: they stay on their line and only get scaled. The whole lesson reduces to this: eigenvectors are the directions a transformation only scales, the eigenvalue is the scale factor, and in the eigenvector basis the transformation becomes a diagonal matrix of those factors, the simplest description it has. This is the scan-it-in-five-minutes version.

Core ideas

An eigenvector is a nonzero vector v that a matrix merely scales: M · v = λ · v. The eigenvalue λ is the scaling factor. For an eigenvector, the entire matrix acts like a single number, no rotation or shear, just stretch (or flip if λ is negative, or collapse if λ = 0).
Geometrically, sweep a vector around the circle and watch its image: at most directions the output is off the input’s line, but at the eigenvector directions it snaps back onto the same line.
Some are readable by eye: a stretch [[2,0],[0,3]] has eigenvectors i-hat (λ=2) and j-hat (λ=3); a pure rotation has no real eigenvectors; a shear [[1,1],[0,1]] has only one ([1,0], λ=1). A full set of independent eigenvectors is common but not guaranteed.
For the rest, use the characteristic equation det(M - λI) = 0 to find the eigenvalues, then find each eigenvector as the null space of (M - λI). This works because (M - λI)·v = 0 has a nonzero solution only when the matrix collapses, the zero-determinant condition.
Worked anchor: M = [[3,1],[0,2]] gives det((3-λ)(2-λ)) = 0, so λ = 3 and 2, with eigenvectors [1,0] and [1,-1] (both verified against M·v = λ·v).
The payoff is diagonalization, the change-of-basis lesson’s promise delivered. With the eigenvectors as the columns of P, the sandwich D = P^-1 · M · P is diagonal with the eigenvalues on it. For [[3,1],[0,2]], D = [[3,0],[0,2]]: the messy matrix is revealed as “stretch by 3 along one eigenvector, by 2 along the other.” Eigenvectors are the natural coordinate system of a transformation.
This is central across machine learning. PCA finds the eigenvectors of a data covariance matrix (the principal components, directions of greatest variance, with eigenvalues measuring the spread captured). Exploding and vanishing gradients are an eigenvalue story (a signal scaled repeatedly by a matrix’s eigenvalues blows up above 1 and decays below 1). Spectral graph methods and PageRank are eigenvector problems.

What changes for you

Before this lesson, “eigenvector” was probably an intimidating word attached to a formula you may have computed without seeing the point. Now it has a clear picture: the directions a transformation leaves on their own line, scaled by the eigenvalue, and the basis in which the transformation turns transparent (a clean diagonal). That picture is the key to a whole tier of techniques you will keep meeting (PCA, gradient stability analysis, spectral methods), all of which amount to finding the directions a matrix simply scales. The final lesson steps back from arrows entirely and asks which of this track’s ideas survive when “vector” means a function or a polynomial, something that is not an arrow at all.