Eigenvalues and eigenvectors

A matrix can rotate, shear, and stretch all at once. Most input arrows leave pointing a new way. Somewhere, if you are lucky, there are special arrows that the map only scales: longer, shorter, or flipped, but still on the same line. Those arrows are the map's simplest confession of what it does.

The finite-dimensional story has deep nineteenth-century roots. MacTutor's matrices topic credits Cauchy with major spectral results for real symmetric matrices (including reality of eigenvalues in that setting). David Hilbert (1862-1943) later helped make spectral language central for operators and infinite-dimensional spaces; MacTutor sketches the career. Steal the ethic carefully across both eras: find directions where a complicated action becomes simple scaling. Do not flatten Cauchy into Hilbert or invent a single "Eureka, eigenvalues!" night.

For a modern rereader, $A\mathbf{v}=\lambda\mathbf{v}$ is the sentence. Geometry first; characteristic polynomials second.

A nonzero vector $\mathbf{v}$ is an eigenvector of square $A$ with eigenvalue $\lambda$ when

Geometrically: $A$ stretches $\mathbf{v}$ by $|\lambda|$ and flips it if $\lambda<0$, without rotating it off its line. The zero vector is excluded by convention; it would satisfy the equation for every $\lambda$ and teach nothing.

Rearrange: $(A-\lambda I)\mathbf{v}=\mathbf{0}$. Nontrivial solutions exist when $A-\lambda I$ is singular, i.e.

Roots of that characteristic polynomial are the eigenvalues (in $\mathbb{C}$, counting multiplicity).

If enough independent eigenvectors exist, they form a basis in which $A$ acts by pure scaling on each axis: diagonalization. Not every matrix earns that luxury (Jordan blocks appear when geometric multiplicity falls short). Symmetric real matrices are the friendly case: real eigenvalues and an orthonormal eigenbasis (spectral theorem in the finite-dimensional Euclidean setting).

Watch a linear transformation of the plane. Most arrows twist. Eigenarrows are the stubborn ones that stay on their rails. $\lambda=2$ means "double along this rail." $\lambda=1/2$ means "halve." $\lambda=-1$ means "flip."

A pure rotation by $90^{\circ}$ has no real eigenvectors: nothing stays on its line. Over complex numbers it still has eigenvalues on the unit circle. The field matters.

Powers $A^k$ become easy on eigen-directions: multiply by $\lambda^k$. Long-run behavior of $\mathbf{x}_{t+1}=A\mathbf{x}_t$ is dominated by the largest $|\lambda|$.

Algebrica's eigenvalues page draws the stretch picture; the diagonalization page explains when a full eigenbasis exists.

Eigenvalues are the diagonal entries always. Only in special bases (triangular form reveals eigenvalues on the diagonal). In a random basis they hide.

Every matrix has a full set of real eigenvectors. False. Real rotations are the classroom counterexample. Complex eigenvalues come in conjugate pairs for real matrices.

"The" eigenvector for $\lambda$. Eigenspaces can be higher-dimensional. Any nonzero vector in the eigenspace works; scaling an eigenvector stays an eigenvector for the same $\lambda$.

Large eigenvalue means "important feature" in every application. In dynamics, large $|\lambda|$ dominates growth. In PCA (next chapter), large eigenvalues of covariance mean high variance directions. Context sets the moral.

Diagonalization always succeeds. Need a full basis of eigenvectors. Defective matrices need a broader Jordan story.

Eigenvectors must be orthogonal. Not in general. Symmetric (or Hermitian) cases give orthogonal eigenbases; general matrices need not.

Cauchy developed key spectral facts while studying quadratic forms; MacTutor's matrices topic is the careful pointer. Hilbert's later operator-theoretic work made "spectrum" a keyword of twentieth-century analysis and mathematical physics. The finite-dimensional $A\mathbf{v}=\lambda\mathbf{v}$ picture is the doorway; Hilbert spaces are the mansion.

Steal the shared instinct: complicated linear action becomes intelligible once invariant directions are named.

A vibrating string or bridge has modal shapes; frequencies relate to eigenvalues of an operator or matrix model. Principal stress axes align with eigenvectors of a stress tensor. Population models $\mathbf{n}_{t+1}=A\mathbf{n}_t$ grow like the dominant eigenvalue when that mode is excited.

Anytime you ask "what does this linear rule do if I wait a long time?", eigenvalues answer.