Eigenvectors and Eigenvalues
When a matrix transforms the plane, most vectors get knocked clean off their original line —
they rotate as well as stretch, ending up pointing somewhere new. But a few special directions
refuse to budge: the transformation only stretches or shrinks them, leaving
them pointing exactly the same way (or exactly opposite). Those rare, unmoved directions are
the eigenvectors of the matrix, and the stretch factor is the
eigenvalue.
It sounds like a narrow, technical curiosity — yet this "stays-on-its-own-line" behaviour turns
out to unlock an astonishing range of real problems. The original Google search engine ranked
the entire web by finding one eigenvector. Engineers predict whether a bridge will
shake itself apart in the wind by finding the eigenvalues of its vibration equations. Both are
the same idea underneath: find the directions a transformation treats specially, and the rest
of the system's behaviour falls out for free.
For an eigenvector \vec{v}, applying A is
the same as multiplying by a single number \lambda — its eigenvalue:
A\vec{v} = \lambda\vec{v}.
-
For a square matrix A, a nonzero vector
\vec{v} is an eigenvector of
A if A\vec{v} is a scalar multiple of
\vec{v} itself: A\vec{v} = \lambda\vec{v}.
-
The scalar \lambda is the corresponding
eigenvalue — it can be positive, negative, or zero.
-
Geometrically: A is allowed to stretch, shrink, or flip
\vec{v}, but it may never rotate it off its own line.
Hunt for the unmoved directions
Spin the input vector \vec{v} (faint) and watch its image
A\vec{v} (bold) for the matrix
A = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix}. Most of the time
the two point different ways. But at a couple of special angles they snap onto the
same line — there \vec{v} is an eigenvector, and the
readout shows its eigenvalue \lambda (how many times longer
A\vec{v} is). Watch for the two faint dashed lines: they mark the
eigenvector directions, and we'll use this very matrix in the worked example below.
Worked example 1 — verify an eigenvector by hand
Is \vec{v} = \begin{bmatrix} 1 \\ 1 \end{bmatrix} an eigenvector of
A = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix} — the same matrix
from the diagram above? Just multiply and see what comes out:
A\vec{v} = \begin{bmatrix} 2 & 1 \\ 1 & 2 \end{bmatrix}\begin{bmatrix} 1 \\ 1 \end{bmatrix} = \begin{bmatrix} 2(1) + 1(1) \\ 1(1) + 2(1) \end{bmatrix} = \begin{bmatrix} 3 \\ 3 \end{bmatrix}.
Is (3,3) a scalar multiple of (1,1)?
Yes: (3,3) = 3\cdot(1,1). So \vec{v}
is an eigenvector of A, with eigenvalue
\lambda = 3 — this is exactly the direction where the diagram's two
arrows line up. The whole check is mechanical: compute A\vec{v},
then ask "is this just a number times the original \vec{v}?"
Worked example 2 — the easy case: a diagonal matrix
For a diagonal matrix, the eigenvectors are almost embarrassingly obvious. Take
D = \begin{bmatrix} 3 & 0 \\ 0 & 7 \end{bmatrix}. Multiply it by
\hat{\imath} = \begin{bmatrix}1\\0\end{bmatrix}:
D\hat{\imath} = \begin{bmatrix} 3 & 0 \\ 0 & 7 \end{bmatrix}\begin{bmatrix} 1 \\ 0 \end{bmatrix} = \begin{bmatrix} 3 \\ 0 \end{bmatrix} = 3\,\hat{\imath}.
So \hat{\imath} is an eigenvector with eigenvalue
3 — exactly the diagonal entry sitting in its own column. The same
happens for \hat{\jmath}: D\hat{\jmath} = 7\hat{\jmath},
eigenvalue 7. In general, a diagonal matrix's eigenvectors
are simply the standard basis vectors, and its eigenvalues are just its diagonal
entries, read straight off with no work at all. This trivial case is exactly why
diagonalization
is so desirable later on: if you can find a basis in which a matrix becomes diagonal, its
eigen-structure becomes this easy to read.
Worked example 3 — a reflection, geometrically
Eigenvectors don't always need algebra — sometimes geometry hands them to you directly.
Consider the matrix that reflects the plane across the line y = x:
R = \begin{bmatrix} 0 & 1 \\ 1 & 0 \end{bmatrix}.
Any point on the mirror line, like (1,1), reflects
to itself: R\begin{bmatrix}1\\1\end{bmatrix} = \begin{bmatrix}1\\1\end{bmatrix},
an eigenvector with eigenvalue +1 — reflecting it does nothing at
all. But a point perpendicular to the mirror, like
(1,-1), flips to exactly the opposite side:
R\begin{bmatrix}1\\-1\end{bmatrix} = \begin{bmatrix}-1\\1\end{bmatrix} = -1\begin{bmatrix}1\\-1\end{bmatrix},
an eigenvector with eigenvalue -1. Every reflection has exactly this
shape: +1 along the mirror, -1
perpendicular to it — you can often name a reflection's eigenvectors just by looking at the
picture, no matrix arithmetic required.
An eigenvalue of exactly 0 is not a mistake. If
A\vec{v} = 0\cdot\vec{v} = \vec{0}, that simply means
A squashes the direction \vec{v} flat
onto the zero vector — a real, important eigenvector with eigenvalue 0,
not a broken calculation. It happens whenever A collapses some
direction entirely, for instance a matrix that projects everything onto a single line: any
vector perpendicular to that line gets flattened to nothing, eigenvalue 0.
The other classic trap: eigenvectors are only defined up to scaling. If
\vec{v} is an eigenvector of A with
eigenvalue \lambda, then so is 2\vec{v},
-\vec{v}, or c\vec{v} for
any nonzero c — always with the very same
eigenvalue. Check it for Worked Example 1: A(2,2) = (6,6) = 3(2,2),
still eigenvalue 3. An eigenvector names a
direction, not a specific length — don't be surprised when two textbooks list
"different" eigenvectors for the same matrix that are really just scaled copies of each other.
Why anyone cares
Eigenvectors are the "natural axes" of a transformation — the directions along which it does
nothing but scale. Find them, as in the three worked examples above, and a tangled matrix
becomes simple: it's just stretching along its own private set of axes. That insight powers
diagonalization,
the stability of physical systems, the ranking behind Google's PageRank, and the
principal
directions of a dataset. First, though, we need a general way to compute
eigenvectors and eigenvalues for a matrix where the answer isn't as obvious as it was above —
that's next.
In the late 1990s, Larry Page and Sergey Brin needed a way to rank billions of web pages by
importance. Their idea, PageRank, builds a giant matrix where each entry
records whether one page links to another, and treats a link as a small "vote" of importance
passed along. A page is important if important pages link to it — a circular definition that
sounds impossible to pin down.
The resolution is exactly the idea on this page: the ranking of every page on the web is the
eigenvector of that link matrix corresponding to eigenvalue
1. Importance flowing around the web settles into the one
distribution of "scores" that the link matrix leaves completely unchanged — an eigenvector, by
definition. Instead of solving billions of equations directly, early search engines found this
eigenvector by repeatedly multiplying an importance guess by the matrix, watching it settle
towards the true ranking, page after page after page.
Every bridge, building, and aircraft wing has its own natural rhythms — frequencies at which it
loves to vibrate, the same way a wine glass rings at one particular pitch when you flick it.
Engineers find those rhythms by writing down the structure's equations of motion as a matrix
and computing its eigenvalues: each eigenvalue corresponds to one natural
resonant frequency, and its eigenvector describes the exact shape the structure takes as it
vibrates at that frequency (a "mode shape").
In 1940, the Tacoma Narrows Bridge in Washington State began twisting itself violently in
moderate wind, and collapsed within hours — famous footage shows the deck rippling like a
piece of cloth. The wind was feeding energy into the bridge at a rate that matched one of the
structure's own eigenvalue-derived resonant frequencies, pumping the vibration bigger and
bigger until the bridge tore itself apart. Modern bridges are deliberately engineered so their
eigenvalues stay well clear of the frequencies that wind, traffic, or earthquakes are likely to
produce — a direct, life-or-death application of finding a matrix's eigenvectors and
eigenvalues.
See it explained