Eigenvalues and Eigenvectors: The Directions a Transformation Can't Turn

Apply a matrix to a vector and, in general, two things happen at once: the vector gets stretched (or shrunk) and it gets rotated to point in a new direction. The combination of these two effects is what makes matrices feel complicated. A transformation that twists every arrow into a different orientation is hard to picture all at once.

But for almost any transformation there exist a few special directions where the rotation simply doesn’t happen. Vectors pointing along these directions come out pointing the same way — only longer or shorter, possibly flipped. These privileged directions are the eigenvectors of the transformation, and the factor by which each one is stretched is its eigenvalue.

Finding them is like finding the grain in a piece of wood. Once you know which way the grain runs, the whole block becomes easy to work with. This article is about what eigenvectors are, how to find them, and why they turn out to be one of the most useful ideas in all of applied mathematics.

The defining equation

The whole concept fits in one short equation. A vector $\mathbf{v}$ (not the zero vector) is an eigenvector of a square matrix $A$ if

$A\mathbf{v} = \lambda\,\mathbf{v}$

for some number $\lambda$. The left side says “apply the transformation.” The right side says “just scale the vector by $\lambda$.” When the two sides agree, the transformation did nothing to $\mathbf{v}$ except resize it. The scaling factor $\lambda$ is the eigenvalue attached to that eigenvector.

A picture makes the distinction vivid. Apply a transformation to two different vectors: a generic one and an eigenvector.

On the left, the transformation has knocked the vector off its line — its angle is different. On the right, the eigenvector lands exactly on the line it started on, just farther out. That “staying on the line” is the entire content of the eigenvector idea.

How to find them

Rearrange the defining equation. Writing $\lambda\mathbf{v}$ as $\lambda I \mathbf{v}$, where $I$ is the identity matrix, gives

$A\mathbf{v} - \lambda I \mathbf{v} = (A - \lambda I)\,\mathbf{v} = \mathbf{0}.$

We want a nonzero vector $\mathbf{v}$ that the matrix $A - \lambda I$ sends to zero. A matrix can only crush a nonzero vector down to zero if it is singular — that is, if its determinant vanishes. So the eigenvalues are exactly the numbers $\lambda$ satisfying

$\det(A - \lambda I) = 0.$

This is the characteristic equation. Expanding the determinant produces a polynomial in $\lambda$ — the characteristic polynomial — whose roots are the eigenvalues. For a $2 \times 2$ matrix it is a quadratic; for an $n \times n$ matrix it has degree $n$, so by the fundamental theorem of algebra there are always $n$ eigenvalues if we allow complex numbers.

A concrete example. Take

$A = \begin{pmatrix} 2 & 1 \\ 1 & 2 \end{pmatrix}.$

Then

$\det(A - \lambda I) = \det\begin{pmatrix} 2-\lambda & 1 \\ 1 & 2-\lambda \end{pmatrix} = (2-\lambda)^2 - 1 = \lambda^2 - 4\lambda + 3.$

This factors as $(\lambda - 1)(\lambda - 3)$, so the eigenvalues are $\lambda = 1$ and $\lambda = 3$. Substituting each back into $(A - \lambda I)\mathbf{v} = \mathbf{0}$ gives the eigenvectors: $\lambda = 3$ belongs to the direction $(1, 1)$, and $\lambda = 1$ belongs to $(1, -1)$. Apply $A$ to anything pointing along $(1,1)$ and it triples in length; apply it to anything along $(1,-1)$ and it stays exactly the same size. Every other vector is some mixture of these two behaviors.

Why this is the natural coordinate system

Here is the payoff. Suppose a transformation has a full set of eigenvectors. Then those eigenvectors form a basis — a coordinate system custom-built for that transformation. In ordinary coordinates the matrix is a tangle of cross-terms. In the eigenvector coordinate system the same transformation becomes a diagonal matrix: it simply multiplies each coordinate by the corresponding eigenvalue, with no mixing.

This is called diagonalization, written $A = PDP^{-1}$, where $D$ holds the eigenvalues on its diagonal and $P$ holds the eigenvectors as columns. Once you have it, hard problems collapse. To compute $A^{100}$ — normally a brutal calculation — you only need $A^{100} = PD^{100}P^{-1}$, and raising a diagonal matrix to a power just means raising each diagonal entry to that power. Decades of repeated transformation reduce to a handful of multiplications.

This is exactly why eigenvalues govern the long-term behavior of any repeated process. If a system evolves by multiplying its state by $A$ at each step, then after many steps the term with the largest eigenvalue dominates everything else. The biggest eigenvalue decides whether the system grows without bound, settles to a steady state, or decays to nothing — and its eigenvector tells you the shape the system relaxes into.

Where they show up

The reach of this one idea is hard to overstate. Google’s original PageRank algorithm ranks web pages by finding the dominant eigenvector of a giant matrix describing the link structure of the web — the steady-state of a random surfer. Markov chains reach their long-run distribution as the eigenvector belonging to eigenvalue 1. In mechanical engineering, the eigenvalues of a structure are its natural frequencies of vibration; matching an external frequency to one of them is resonance, the phenomenon that can shake a bridge apart.

In statistics and machine learning, principal component analysis finds the eigenvectors of a data set’s covariance matrix to identify the directions of greatest variation, which is the workhorse behind dimensionality reduction and data compression. In quantum mechanics the entire theory is built on eigenvalues: the measurable values of energy, momentum, and spin are precisely the eigenvalues of the corresponding operators, and “quantization” — the fact that energy comes in discrete levels — is the statement that those eigenvalues form a discrete set. The word physicists use, eigenstate, is the same eigen.

What unites these wildly different fields is a single geometric observation: every linear transformation, however complicated it looks, has a few preferred directions along which it does nothing but stretch. Find those directions, and the transformation becomes transparent. That is why eigenvalues and eigenvectors are not just a topic in a linear algebra course — they are one of the small handful of ideas that the modern technical world genuinely runs on.

Frequently asked

What does the word 'eigen' mean?

It's German for 'own' or 'characteristic' — as in 'its own value' or 'proper value.' The terminology comes from David Hilbert's work in the early 1900s. An eigenvector is a vector belonging characteristically to a particular transformation; it captures something intrinsic about how that transformation behaves. English kept the German prefix rather than translating it, which is why the words look unusual.

Can a matrix have complex eigenvalues even if all its entries are real?

Yes, and it happens whenever the transformation involves rotation. A pure rotation in the plane has no real eigenvectors — no real direction is preserved, since everything turns. The eigenvalues come out complex, and the imaginary part encodes the rotation angle. This is why complex numbers are unavoidable in linear algebra even when you start with purely real data.

How many eigenvalues does an n×n matrix have?

Exactly n, counted with multiplicity, if you allow complex values. They are the roots of the characteristic polynomial det(A − λI) = 0, which has degree n. Some may repeat, and some may be complex, but the count is always n by the fundamental theorem of algebra.

Why are eigenvalues useful in practice?

They reveal the long-term behavior of repeated processes. Raising a matrix to a high power, solving a differential equation, ranking web pages, analyzing the stability of a structure, or compressing data with principal component analysis — all of these reduce to finding eigenvalues. The eigenvalue tells you whether a quantity grows, shrinks, or oscillates over time.

Continue reading

Blog

The Cayley–Hamilton Theorem: Every Matrix Solves Its Own Equation

Compute the characteristic polynomial of a square matrix, then substitute the matrix itself into the polynomial. The result is zero — every single time. A strange, useful algebraic identity.

Read →Blog

Determinants: The Number That Measures How Much a Matrix Stretches Space

A determinant looks like an arbitrary recipe of multiply-and-subtract. Geometrically it is something simple and beautiful: the factor by which a transformation scales area or volume.

Read →Blog

What Matrices Really Are

Most people meet matrices as grids of numbers to multiply in a specific way. That description is technically correct and deeply misleading about why matrices matter.

Read →