eigenvalues and eigenvectors
/ EYE-gun-VAL-yooz and EYE-gun-VEK-turz /
When a matrix reshapes space, it shoves most vectors off in some new direction. But a few special vectors are stubborn: the transformation only stretches or shrinks them, never knocks them off their own line. Those stubborn directions are the eigenvectors, and the factor by which each one is stretched or shrunk is its eigenvalue. The German prefix 'eigen' means 'own' or 'characteristic' — these are the transformation's own private axes.
An eigenvalue of 2 means that vector gets twice as long but keeps pointing the same way; an eigenvalue of 0.5 means it shrinks by half; a negative eigenvalue means it flips to point backward along the same line. Together, the eigenvectors and eigenvalues reveal a transformation's skeleton: the natural directions along which it acts most simply, like finding the grain in a piece of wood so you know which way it splits cleanly.
This idea quietly powers a lot of data science. Principal component analysis, the classic tool for squeezing high-dimensional data down to its most important few directions, is built entirely on eigenvectors. Google's original PageRank found the most authoritative web pages as an eigenvector of the web's link structure. They also tell you whether a repeated process will blow up or settle down. One honest caveat: not every matrix has a tidy set of real eigenvectors, and computing them for huge matrices is heavy work — useful, but no free lunch.
Take the matrix [ [2, 0], [0, 3]]. A vector lying flat along the horizontal axis gets stretched to twice its length but never tilts — so it is an eigenvector with eigenvalue 2. A vector pointing straight up gets tripled — eigenvalue 3. A diagonal arrow, by contrast, gets swung to a new angle, so it is not an eigenvector of this matrix.
Eigenvectors are the directions a transformation can't knock sideways; the eigenvalue is just how much it stretches them.
An eigenvector marks a direction, so any nonzero multiple of it is the 'same' eigenvector — its length is not what counts. The concept applies only to square matrices; for non-square ones the closely related tool is the singular value decomposition.