Eigenvalues and Eigenvectors

Suppose a linear operator acts on a vector space by stretching, rotating, and shearing space in some complicated way. Is there ever a direction that the operator leaves fixed? Such a direction would only be scaled, not bent elsewhere. This question — does an operator have an invariant line — turns out to be the single most productive question in linear algebra. Its answer is organized by the notion of an eigenvalue and eigenvector. Eigenvectors matter because they replace an entangled, coordinate-dependent description of an operator with something simpler. They give the simplest possible local picture: multiplication by a scalar. Once we find enough independent eigenvectors, we can use them as a basis. In this basis, the operator's matrix becomes diagonal. This collapses a complicated system of coupled equations into independent one-dimensional problems. This is the mechanism behind solving linear differential equations, analyzing the long-run behavior of Markov chains, understanding vibration modes of physical systems, and compressing data via principal component analysis. The big picture is this: every linear operator $T$ on a vector space $V$ decomposes $V$ . This decomposition is at least partial, and complete when things go well. It splits $V$ into Invariant Subspaces on which $T$ acts by pure scaling. The one-dimensional invariant subspaces are exactly the spans of eigenvectors. So eigenvalues and eigenvectors are the atomic building blocks of this decomposition. Understanding them precisely is a necessary first step. We must know what equation they satisfy, why they need not exist over every field, and how to compute them. This is essential groundwork toward diagonalization and the structure theory of Linear Operators. This article gives the precise definition of eigenvalue and eigenvector. It proves the basic structural facts: uniqueness of the scalar, and linear independence of eigenvectors for distinct eigenvalues. It also builds the intuition needed for later results. These include the characteristic polynomial and the existence of eigenvalues over the complex numbers.

Key takeaways

An eigenvector of $T$ is a nonzero vector $v$ with $Tv = \lambda v$ for some scalar $\lambda$ , called the eigenvalue.
The scalar $\lambda$ associated to a given eigenvector is unique; a single eigenvector cannot have two eigenvalues.
$\lambda$ is an eigenvalue of $T$ if and only if $T - \lambda I$ is not injective, i.e. $\ker(T-\lambda I) \neq \{0\}$ .
Eigenvectors corresponding to distinct eigenvalues are always linearly independent — this is the key combinatorial fact.
Not every operator has an eigenvalue over every field; real rotations of the plane are the classic counterexample.
Key idea: an eigenvector spans a one-dimensional invariant subspace, the simplest possible nontrivial building block for understanding a linear operator.

Where this lives today

Eigenvalues and eigenvectors are indispensable across mathematics and its applications. In pure mathematics, they underlie the entire theory of Diagonalization and the structure theorems for linear operators, including the Jordan form and spectral theorems for self-adjoint and normal operators. In differential equations, the eigenvectors of a linear system's coefficient matrix give the "normal modes" along which solutions evolve by simple exponential scaling, turning coupled systems into independent scalar equations. In physics, eigenvalues of the Hamiltonian operator are the allowed energy levels of a quantum system, and eigenvectors are the corresponding stationary states — this is the direct descendant of Heisenberg's matrix mechanics. In engineering, eigenvalues of a system matrix determine the stability of feedback control systems and the resonant frequencies of vibrating structures such as bridges and airplane wings. In statistics and data science, the eigenvectors of a covariance matrix give the directions of maximal variance used in Principal Component Analysis, a core dimensionality-reduction technique. In computer science, the PageRank algorithm ranks web pages using the dominant eigenvector of a giant transition matrix, and eigenvalues govern convergence rates of iterative numerical algorithms. In cryptography and coding theory, eigenvalue bounds on the adjacency matrices of graphs (spectral graph theory) control the mixing properties of random walks used in expander-graph constructions.

Ask the tutor: How does the characteristic polynomial let us compute all eigenvalues of a matrix at once, instead of guessing eigenvectors one at a time?

V

Intuition

Picture a linear operator $T$ as a machine that takes in a vector and outputs another vector, generally rotating and rescaling it in the process. For almost every input vector $v$ , the output $Tv$ points in a different direction than $v$ did. An eigenvector is a special input direction that the machine refuses to rotate: it only stretches or shrinks it (and possibly flips it, if $\lambda < 0$ ), never bending it off its original line. This is why eigenvectors are naturally tied to Invariant Subspaces: if $v$ is an eigenvector with eigenvalue $\lambda$ , then the one-dimensional subspace $\mathrm{span}(v)$ is invariant under $T$ , because $T$ sends every scalar multiple $cv$ to $c\lambda v \in \mathrm{span}(v)$ . Conversely, every one-dimensional invariant subspace of $T$ is the span of some eigenvector. So finding eigenvectors is precisely the search for the smallest, simplest nontrivial invariant pieces of $V$ under $T$ . The picture below shows this concretely for a shear-like operator on $\mathbb{R}^2$ : most vectors get pushed off their line, but along two special directions the arrows stay on the same line through the origin, only rescaled.

Figure

Two eigenvectors of a

2\times 2

operator: directions preserved, lengths rescaled

The green vector $u_1$ and its image $Tu_1$ lie on the same ray through the origin, scaled by factor $2$ ; the orange vector $u_2$ and its image $Tu_2$ lie on the same line but reversed, scaled by $-1$ . The blue vector $v$ is generic: its image $Tv$ points somewhere else entirely.

V

Tv = \lambda v

V

T

Worked examples

Example 1: a $3\times 3$ matrix with an asymmetric structure. Let $T : \mathbb{R}^3 \to \mathbb{R}^3$ be given by the matrix

A = \begin{pmatrix} 4 & 1 & 0 \\ 0 & 2 & 0 \\ -2 & -1 & 3 \end{pmatrix}.

We verify directly, without yet invoking the characteristic polynomial machinery, that $\lambda = 2$ and $\lambda=3$ are eigenvalues, and separately determine whether $4$ is one, by exhibiting eigenvectors. Testing $\lambda = 2$ : we seek nonzero $v = (x,y,z)$ with $Av = 2v$ , i.e.

4x + y = 2x, \qquad 2y = 2y, \qquad -2x - y + 3z = 2z.

The first equation gives $2x + y = 0$ , i.e. $y = -2x$ . The second equation is automatically satisfied for any $y$ . The third becomes $-2x - y + z = 0$ , and substituting $y=-2x$ : $-2x + 2x + z = 0$ , so $z = 0$ . Taking $x=1$ gives $y = -2$ , $z=0$ : the vector $v_2 = (1,-2,0)$ satisfies $Av_2 = 2v_2$ . Check directly:

A(1,-2,0) = (4\cdot 1 + 1\cdot(-2) + 0,\ 0 + 2\cdot(-2)+0,\ -2\cdot1 -1\cdot(-2)+0) = (2,-4,0) = 2(1,-2,0). \checkmark

Testing $\lambda = 3$ : seek $Av = 3v$ :

4x+y = 3x, \qquad 2y = 3y, \qquad -2x-y+3z = 3z.

The second equation forces $y = 0$ . The first becomes $x + y = 0$ , so with $y=0$ we get $x=0$ . The third becomes $-2x - y = 0$ , automatically true when $x=y=0$ , for any $z$ . So $v_3 = (0,0,1)$ works. Check:

A(0,0,1) = (0,0,3) = 3(0,0,1). \checkmark

Testing $\lambda = 4$ : seek $Av=4v$ :

4x+y=4x \Rightarrow y = 0, \qquad 2y = 4y \Rightarrow y = 0 \ (\text{consistent}), \qquad -2x - y + 3z = 4z \Rightarrow -2x - y - z = 0.

With $y=0$ , the third equation is $-2x - z = 0$ , i.e. $z = -2x$ . Taking $x=1$ : $v_1=(1,0,-2)$ . Check:

A(1,0,-2) = (4, 0, -2-6) = (4,0,-8) = 4(1,0,-2). \checkmark

So $4$ , $2$ , $3$ are all eigenvalues, with respective eigenvectors $(1,0,-2)$ , $(1,-2,0)$ , $(0,0,1)$ — three distinct eigenvalues of a $3\times 3$ matrix, consistent with the corollary's bound of "at most $n=3$ ." One can further check these three eigenvectors are linearly independent (their $3\times3$ coefficient matrix has nonzero determinant $-2$ ), exactly as guaranteed by the theorem, and in fact they form a basis of $\mathbb{R}^3$ in which $A$ acts diagonally — the phenomenon that motivates diagonalization.

Example 2: eigenvectors of a differentiation-like operator on a polynomial space, with a numerically asymmetric twist. Let $V = \mathbb{R}_{\le 2}[x]$ , the space of real polynomials of degree at most $2$ , and define $T: V \to V$ by

T(p)(x) = 3xp'(x) - p(x) + 2p(0).

This is linear: for polynomials $p,q$ and scalars $a$ , $T(ap + q)(x) = 3x(ap+q)'(x) - (ap+q)(x) + 2(ap+q)(0) = a\,T(p)(x) + T(q)(x)$ , using linearity of differentiation and evaluation at $0$ . We look for eigenvectors among $1, x, x^2$ directly (an efficient guess given the operator's form), then confirm eigenvalues by direct computation — no shortcuts, no unproved claims. Try $p(x) = 1$ (constant). Then $p'(x) = 0$ , $p(0) = 1$ , so

T(1) = 3x\cdot 0 - 1 + 2\cdot 1 = 1.

So $T(1) = 1 \cdot 1$ : the constant polynomial $1$ is an eigenvector with eigenvalue $\lambda = 1$ . Try $p(x) = x$ . Then $p'(x) = 1$ , $p(0) = 0$ , so

T(x) = 3x\cdot 1 - x + 2\cdot 0 = 3x - x = 2x.

So $T(x) = 2x$ : eigenvector $x$ with eigenvalue $\lambda = 2$ . Try $p(x) = x^2$ . Then $p'(x) = 2x$ , $p(0)=0$ , so

T(x^2) = 3x(2x) - x^2 + 0 = 6x^2 - x^2 = 5x^2.

So $T(x^2) = 5x^2$ : eigenvector $x^2$ with eigenvalue $\lambda=5$ . We have found three distinct eigenvalues $1, 2, 5$ of $T$ on a $3$ -dimensional space $V$ , again saturating the bound of the corollary. By the theorem, the eigenvectors $1, x, x^2$ — already visibly a basis of $V$ — are automatically linearly independent, which we can double check: the only way $a\cdot 1 + b\cdot x + c\cdot x^2 = 0$ as a polynomial (identically zero) is $a=b=c=0$ , confirming independence directly as well as via the theorem. Note the asymmetry deliberately built into this example: the operator mixes a derivative term, a multiplication term, and an evaluation term, so the eigenvectors were not assumed in advance to be the monomials — they were verified to satisfy $Tp = \lambda p$ term by term.

\mathbb{R}

Scope, edge cases & common traps

Situation	What is true	What is not true
$\lambda = 0$	$0$ is an eigenvalue iff $T$ is not injective (has nontrivial kernel)	$0$ being an eigenvalue does not mean $T$ is the zero map
Eigenvector $v$	determines a unique eigenvalue $\lambda$	a single vector can be an eigenvector for two different $\lambda$ 's only if $v = 0$ , which is excluded
Distinct eigenvalues $\lambda_1,\dots,\lambda_m$	corresponding eigenvectors are linearly independent	this says nothing about eigenvectors sharing the same eigenvalue — those can be dependent or span a whole [[Eigenspaces	eigenspace]]
Finite dimension $n$	at most $n$ distinct eigenvalues	operators need not have $n$ distinct eigenvalues, or any at all (see the rotation example)
Field $\mathbb{F}$	over $\mathbb{C}$ , every operator on a nonzero finite-dimensional space has at least one eigenvalue	over $\mathbb{R}$ this can fail entirely

The mathematician behind it

Origins · 1 minAugustin-Louis Cauchy and how this was discovered

Augustin-Louis Cauchy

1789–1857 · early 19th-century French mathematical analysis and algebra

Cauchy was among the first to systematically study what we now call eigenvalues, in the context of quadratic forms and the principal axes of ellipsoids arising in mechanics and the theory of elasticity. His work connected the algebraic problem of characteristic roots to the geometric problem of finding invariant directions of a symmetric transformation, laying groundwork later formalized by Cauchy, Jacobi, and Sylvester into the general spectral theory used today.

History

- 1826 — Cauchy studies "characteristic values" of quadratic forms related to axes of symmetry in mechanics.

- 1858 — Cayley's memoir on matrices introduces systematic matrix algebra, setting up the language for eigenvalue equations.

- 1904 — Hilbert names the eigenvalue problem ("Eigenwertproblem") in the context of integral equations, coining the now-standard German-rooted term.

- 1925 — Heisenberg's matrix mechanics makes eigenvalues the central objects representing observable quantities in quantum theory.

- 1970s — Numerical linear algebra (QR algorithm, power iteration) turns eigenvalue computation into a cornerstone of computational science.

Does every linear operator have at least one eigenvector?

Not necessarily, and this is a genuinely important subtlety. Over $\mathbb{R}$ , an operator can have no real eigenvalue at all, as shown by the $90°$ rotation of the plane: every nonzero vector is sent to a perpendicular vector, so no direction is preserved even up to sign. However, over the complex numbers $\mathbb{C}$ , it is a deep and important theorem — relying on the fundamental theorem of algebra — that every linear operator on a nonzero finite-dimensional complex vector space has at least one eigenvalue. This existence result is proved separately from the basic definitions given here, and is the natural next topic once eigenvalues and eigenvectors are understood.

Can the zero vector ever be an eigenvector?

No, by definition eigenvectors are required to be nonzero. This is not a minor technicality: without it, the equation $Tv = \lambda v$ would be satisfied by $v = 0$ for every scalar $\lambda$ whatsoever, since $T0 = 0 = \lambda \cdot 0$ always holds. Allowing $v=0$ would make every scalar in the field an "eigenvalue" trivially, destroying all the useful content of the definition. Excluding zero is exactly what makes the eigenvalue associated to an eigenvector unique, as proved in the uniqueness theorem above.

Can zero itself be an eigenvalue?

Yes. The scalar $\lambda = 0$ is an eigenvalue of $T$ precisely when there is a nonzero vector $v$ with $Tv = 0$ , that is, precisely when $T$ has a nontrivial kernel (is not injective). This is a perfectly legitimate and common situation — for example, any non-invertible matrix has $0$ as an eigenvalue. Do not confuse " $0$ is an eigenvalue" with " $T$ is the zero operator"; the former only says some nonzero vector is annihilated by $T$ , while the latter says every vector is.

If two eigenvectors have the same eigenvalue, are they always dependent?

No — this is exactly where the distinct-eigenvalues hypothesis in the theorem matters. If $v_1$ and $v_2$ both satisfy $Tv_i = \lambda v_i$ for the same $\lambda$ , they can easily be linearly independent; together with all other eigenvectors sharing that eigenvalue (and the zero vector), they span the corresponding eigenspace $\ker(T - \lambda I)$ , which can have dimension greater than one. The independence theorem proved here specifically requires the eigenvalues attached to the vectors to be pairwise distinct; it says nothing about vectors within a single eigenspace.

How do eigenvalues relate to invertibility of a matrix?

A square matrix $A$ (or the linear operator it represents) is invertible if and only if $0$ is not one of its eigenvalues. Indeed, $A$ fails to be invertible exactly when $Av = 0$ has a nonzero solution $v$ , which by definition means $0$ is an eigenvalue with eigenvector $v$ . So checking whether $0$ appears among the eigenvalues is equivalent to checking invertibility — this connection becomes very concrete once eigenvalues are computed via the characteristic polynomial, where $\det(A) = 0$ if and only if $0$ is an eigenvalue.

Why do we require

v_1,\dots,v_{k-1}

independent in the induction, rather than just assuming the full list is dependent?

The proof isolates the smallest index $k$ at which dependence first appears, which guarantees the shorter preceding list $v_1,\dots,v_{k-1}$ is independent (otherwise a smaller $k$ would have worked). This lets us legitimately conclude that a vanishing linear combination of $v_1,\dots,v_{k-1}$ must have all-zero coefficients — a fact that is only available for linearly independent lists. Without pinning down this minimal $k$ , we could not extract the clean contradiction $a_i(\lambda_k - \lambda_i)=0$ for every $i$ , since a longer, possibly-dependent list admits multiple representations and no unique coefficients to compare.

Is the definition the same for infinite-dimensional vector spaces?

The definition $Tv = \lambda v$ with $v \neq 0$ makes sense verbatim for any vector space, including infinite-dimensional ones such as function spaces, and the uniqueness-of-eigenvalue and independence-of-eigenvectors theorems proved here hold with identical proofs in that generality since nothing in the argument used finite dimension. What changes in infinite dimensions is mainly analytic: operators like differentiation on smooth functions have infinitely many eigenvalues (e.g. $\frac{d}{dx}e^{\lambda x} = \lambda e^{\lambda x}$ for every $\lambda$ ), and questions of convergence, boundedness, and spectral theory (rather than pure linear algebra) become central, which is the subject of functional analysis.