When a matrix warps space, most vectors get knocked off their original direction — rotated, sheared away. But a few special directions hold their line; they only get stretched or squashed. Those are the eigenvectors, and how much they stretch is the eigenvalue.

W v = λ v

The eigenvector v, transformed by W, gives back the SAME vector v — just scaled by the number λ (its eigenvalue). Same direction, new length.

Plain English: an eigenvector is a direction the transformation respects. If λ = 2, vectors along that line double. If λ = 0.5, they halve. If λ = −1, they flip but stay on the line. In Python: np.linalg.eig(W).

🧠 Callback: remember the exponential being an "eigenfunction" in calculus — the function differentiation only scaled? Same idea, one level down: an eigenvector is the vector a matrix only scales. Eigen-things are what survive an operation unchanged-except-for-scaling.

🤖 Everywhere in research: eigenvectors are the axes PCA rotates your data onto (the directions of most variance). Eigenvalues decide whether a recurrent network's activity blows up or settles (next quest). Google's original PageRank was one giant eigenvector. This is arguably the single most useful idea in applied math.