EIGENVALUES AND EIGENVECTORS

1. Diagonalizable linear transformations and matrices

Recall, a matrix, D, is diagonal if it is square and the only non-zero entries are
on the diagonal. This is equivalent to D~ei = λi~ei where here ~ei are the standard
vector and the λi are the diagonal entries. A linear transformation, T : Rn → Rn , is
diagonalizable if there is a basis B of Rn so that [T ]B is diagonal. This means [T ] is
similar to the diagonal matrix [T ]B . Similarly, a matrix A ∈ Rn×n is diagonalizable
if it is similar to some diagonal matrix D. To diagonalize a linear transformation
is to find a basis B so that [T ]B is diagonal. To diagonalize a square matrix is to
find an invertible S so that S −1 AS = D is diagonal.
Fix a matrix A ∈ Rn×n We say a vector ~v ∈ Rn is an eigenvector if
(1) ~v 6= 0.
(2) A~v = λ~v for some scalar λ ∈ R.
The scalar λ is the eigenvalue associated to ~v or just an eigenvalue of A. Geo-
metrically, A~v is parallel to ~v and the eigenvalue, λ. counts the stretching factor.
Another way to think about this is that the line L := span(~v ) is left invariant by
multiplication by A.
An eigenbasis of A is a basis, B = (~v1 , . . . , ~vn ) of Rn so that each ~vi is an
eigenvector of A.
Theorem 1.1. The matrix A is diagonalizable if and only if there is an eigenbasis
of A.
Proof. Indeed, if A has eigenbasis B = (~v1 , . . . , ~vn ), then the matrix
 
S = ~v1 | · · · | ~vn
satisfies  
λ1 0 ··· 0
 .. .. 
0 λ2 . . 
S −1 AS = 
.

 .. .. .. 
. . 0
0 ··· 0 λn
where each λi is the eigenvalue associated to ~vi . Conversely, if A is diagonalized by
S, then the columns of S form an eigenbasis of A. 

EXAMPLE: The the standard vectors ~ei form an eigenbasis of −In . Their eigen-
values are −1. More generally, if D is diagonal, the standard vectors form an
eigenbasis with associated eigenvalues the corresponding entries on the diagonal.
EXAMPLE: If ~v is an eigenvector of A with eigenvalue λ, then ~v is an eigenvector
of A3 with eigenvalue λ3 .
EXAMPLE: 0 is an eigenvalue of A if and only if A is not invertible. Indeed, 0
is an eigenvalue ⇐⇒ there is a non-zero ~v so A~v = ~0 true ⇐⇒ ~v ∈ ker A so ker A
is non-trivial ⇐⇒ A not invertible.
1
2 EIGENVALUES AND EIGENVECTORS

EXAMPLE: If ~v is an eigenvector of Q which is orthogonal, then the associated

eigenvalue is ±1. Indeed,
||~v || = ||Q~v || = ||λ~v || = |λ|||~v ||
as ~v 6= 0 dividing, gives |λ| = 1.
EXAMPLE: If A2 = −In , then there are no eigenvectors of A. To see this,
suppose ~v was an eigenvector of A. Then A~v = λ~v . As such
−~v = −In~v = A2~v = λ2~v
That is, λ2 = −1. There are no real numbers whose square is negative, so there
is no such ~v . This means A has no real eigenvalues (it does have have a comples
eigenvalues – see Section 7.5 of the textbook. This is beyond scope of this course).

2. Characteristic Equaiton
One of the hardest (computational) problems in linear algebra is to determine the
eigenvalues of a matrix. This is because, unlike everything else we have considered
so far, it is a non-linear problem. That being said, it is still a tractable problem
(especially for small matrices).
To understand the approach. Observe that if λ is an eigenvalue of A, then there
is a non-zero ~v so thatA~v = λ~v . That is,
A~v = (λIn )~v ⇐⇒ (A − λIn )~v = ~0 ⇐⇒ ~v ∈ ker(A − λIn ).
From which we conclude that A − λIn is not invertible and so det(A − λIn ) = 0.
In summary,
Theorem 2.1. λ is an eigenvalue of A if and only if
det(A − λIn ) = 0.
The equation det(A − λIn ) = 0 is called the characteristic equation of A.
EXAMPLE: Find the eigenvalues of
 
2 3
A= .
3 2
The characteristic equation is
 
2−λ 3
det(A − λI2 ) = det = λ2 − 4λ − 5 = (λ + 1)(λ − 5)
3 2−λ
Hence, the eigenvalues are λ = −1 and λ = 5. To find corresponding eigenvectors
we seek non-trivial solutions to
     
2 − (−1) 3 x1 2 − (5) 3 x1
= ~0 and = ~0
3 2 − (−1) x2 3 2 − (5) x2
By inspection the non-trivial solutions are
   
1 1
and .
−1 1
Hence,      
2 3 1 1 1 1 −1 0
=
3 2 −1 1 −1 1 0 5
So we have diagonalized A.
 EIGENVALUES AND EIGENVECTORS 3

EXAMPLE: Find eigenvalues of

 
1 −2 3
A = 0 2 0
0 0 3
So  
1−λ −2 3
det(A − λI3 ) = det  0 2−λ 0  = (1 − λ)(2 − λ)(3 − λ)
0 0 3−λ
Hence, eigenvalues are 1, 2, 3.
This example is a special case of a more general phenomena.
Theorem 2.2. If M is upper triangular, then the eigenvalues of M are the diagonal
entries of M .
EXAMPLE: When n = 2, the eigenvalues of
 
a b
A=
c d
so the characteristic equation is
 
a−λ b
det(A − λI2 ) = det = λ2 − (a + d)λ + (ad − bc) = λ2 − trAλ + det A
c d−λ
Using the quadratic formula we have the following:
(1) When tr(A)2 − 4 det A > 0, then two distinct eigenvalues
(2) When tr(A)2 − 4 det A = 0, exactly one eigenvalue 21 trA.
(3) When tr(A)2 − 4 det A < 0, then no (real) eigenvalues.

3. Characteristic Polynomial
As we say for a 2 × 2 matrix, the characteristic equation reduces to finding the
roots of an associated quadratic polynomial. More generally, for a n × n matrix A,
the characteristic equation det(A − λIn ) = 0 reduces to finding roots of a degree n
polynomial o fthe form
fA (λ) = (−1)n λn + (−1)n−1 (trA)λn−1 + · + det A
this is called the characteristic polynomial of A. To see why this is true observe
that if P is the diagonal pattern of A − λIn , then
prod(P ) = (a11 − λ) · · · (ann − λ) = (−1)n λn + (−1)n−1 (trA)λn−2 + RP (λ)
where RP (λ) is a polynomial of degree at most n − 2 that depends on P (and of
course also on A and λ). If P is some other pattern of A − λIn , then at least two
entries are not on the diagonal. Hence,
prod(P ) = RP (λ)
for some RP (λ) that is a polynomial of degree at most n − 2 that depends on P .
Hence,
fA (λ) = det(A − λIn ) = (−1)n λn + (−1)n−1 (trA)λn−1 + R(λ)
where R(λ) has degree at most n − 2. Finally, fA (0) = det(A) and so the constant
term of fA is det A as claimed.
4 EIGENVALUES AND EIGENVECTORS

EXAMPLE: If n is odd, there is always at least one real eigenvalue. Indeed, in

this case
lim fA (λ) = lim −λ3 = ∓∞
n→±∞ n→∞
That is for λ >> 1, fA (λ) < 0 and for λ << −1, fA (λ) > 0 and so by the interme-
diate value theorem from calculus (and the fact that polynomials are continuous),
fA (λ0 ) = 0 for some λ0 . This may fail when n is even.
An eigenvalue λ0 has algebraic multiplicity k if
fA (λ) = (λ0 − λ)k g(λ)
where g is a polynomial of degree n − k with g(λ0 ) 6= 0. Write almu(λ0 ) = k in this
case.
EXAMPLE: If  
2 0 1 1
0 1 1 0
A= 
0 0 −1 1
0 0 0 2
then fA (λ) = (2 − λ)2 (1 − λ)(−1 − λ) and so almu(2) = 2, while almu(1) =
almu(−1) = 1. Strictly speaking, almu(0) = 0, as 0 is not an eigenvalue of A and
it is sometimes convenient to follow this convention.
We say an eigenvalue, λ, is repeated if almu(λ) ≥ 2.
Algebraic fact, counting algebraic multiplicity, a n × n matrix has at most n real
eigenvalues. If n is odd, then there is at least one real eigenvalue. The fundamental
theorem of algebra ensures that, counting multiplicity, such a matrix always has
exactly n complex eigenvalues.
We conclude with a simple theorem
Theorem 3.1. If A ∈ Rn×n has eigenvalues λ1 , . . . , λn (listed counting multiplic-
ity):
(1) det A = λ1 λ2 · · · λn .
(2) trA = λ1 + λ2 + · · · + λn .
Proof. If the eigenvalues are λ1 , . . . , λn , then we have the complete factorization of
fA (λ)
fA (λ) = (λ1 − λ) . . . (λn − λ)
This means det(A) = fA (0) = λ1 . . . λn . (λ1 − λ) . . . (λn − λ) = (−1)n λn +
(−1)n−1 (λ1 + . . . + λn )λn−1 + R(λ) where R(λ) of degree at most n − 2. Comparing
the coefficient of the λn−2 term gives the result. 

4. Eigenspaces
Consider an eigenvalue λ of A ∈ Rn×n . We define the eigenspace associated to
λ to be
Eλ = ker(A − λIn ) = {~v ∈ Rn : A~v = λ~v } ⊂ Rn .
Observe that dim Eλ ≥ 1. All non-zero elements of Eλ are eigenvectors of A with
eigenvalue λ.  
1 0
EXAMPLE: A = has repeated eigenvalue 1. Clearly,
0 1
E1 = ker(A − I2 ) = ker(02×2 ) = R2
 EIGENVALUES AND EIGENVECTORS 5

 
1 2
Similarly, the matrix B = has one repeated eigenvalue 1. However,
0 1
   
0 2 1
ker(B − I2 ) = ker = span( ).
0 0 0
Motivated by this example, define the geometric multiplicity of an eigenvalue λ
of A ∈ Rn×n tobe
gemu(λ) = null(A − λIn ) = n − rank(A − λIn ) ≥ 1.

5. Diagonalizable Matrices
We are now ready to give a computable condition that will allow us to determine
an answer to our central question in this part of the course: When is A ∈ Rn×n
diagonalizable?

Theorem 5.1. A matrix A ∈ Rn×n is diagonalizable if and only if the sum of the
geometric multiplicities of all of the eigenvalues of A is n.

EXAMPLE: For which k is the following diagonalizable

 
1 k 0
0 1 0?
0 0 2
As this is upper triangular, the eigenvalues are 1 with almu(1) = 2 and 2 with
almu(2) = 1. It is not hard to see that gemu(1) = 1 when k 6= 0 and gemu(1) = 2
when k = 0. We always have gemu(2) = 1 Hence, according to the theorem the
matrix is diagonalizable only when it is already diagonal (that is k = 0) and is
otherwise not diagonalizable.
To prove this result we need the following auxiliary fact

Theorem 5.2. Fix a matrix A ∈ Rn×n and let ~v1 , . . . , ~vs be a set of vectors formed
by concatenating a basis of each non-trivial eigenspace of A. This set is linearly
independent (and so s ≤ n.)

To explain what I mean by concatenating. Suppose A ∈ R5×5 has exactly three

distinct eigenvalues λ1 = 2 and λ2 = 3 and λ3 = 4 If gemu(2) = 2 and
E2 = span(~a1 , ~a2 )
while gemu(3) = gemu(4) = 1 and

E3 = span(~b1 ) and E4 = span(~c1 ),

then their concatenation is the set of vectors
(~v1 , ~v2 , ~v3 , ~v4 ) = (~a1 , ~a2 , ~b1 , ~c1 ).
According to the auxiliary theorem this list is linearly independent. As s = 4 <
5 = n, if this was the complete set of eigenvalues, A would not be diagonalizable
by the main theorem. It would be if we had omitted one eigenvalue from our list.
Let us now prove the auxiliary theorem.
6 EIGENVALUES AND EIGENVECTORS

Proof. If the vectors ~vi are not linearly independent, then they at least one is
redundant. Let ~vm be the first redundant vector on the list That is for some
1 ≤ m ≤ s we can write
m−1
X
~vm = ci~vi
i=1
and cannot do this for any smaller m. This means ~v1 , . . . , ~vm−1 are linearly inde-
pendent.
Let λm be the eigenvalue associated to ~vm . Observe, there must be some 1 ≤ k ≤
m − 1 so that λk 6= λm and ck 6= 0 as otherwise we would have a non-trivial linear
relation for of a set of linearly independent vectors in Eλm (which is impossible).
Clearly,
m−1
X
~0 = (A − λm In )~vm = ci (λi − λm )~vi
i=1

Hence, ci (λi − λm ) = 0 for each i. This contradicts, λk 6= λm and ck 6= 0 and proves

the claim. 

The main theorem follows easily form this. Indeed, the hypotheses gives n lin
indep vectors all which are eigenvectors of A. That is, an eigenbasis of A.

6. Strategy for diagonalizing A ∈ Rn×n

We have the following strategy for diagonalizing a given matrix:
(1) Find eigenvalues of A by solving fA (λ) = 0 (this is a non-linear problem).
(2) For each eigenvalue λ find a basis of the eigenspace Eλ (this is a linear
problem).
(3) The A is diagonalizable if and only if the sum of the dimensions of the
eigenspaces is n. In this case, obtain an eigenbasis, ~v1 , . . . , ~vn , by concate-
nation.
(4) As A~vi = λi~vi , setting
 
λ1 0 · · · 0
.. 
 0 λ2 . . .

  −1 . 
S = ~v1 | · · · | ~vn , S AS = D =  .  
.. 
 .. . .. . .. . 
0 ··· 0 λn
EXAMPLE: Diagonalize (if possible)
 
2 1 0
A = 1 2 0
1 1 1
We compute
 
2−λ 1 0  
2−λ 1
fA (λ) = det(A − λI3 ) = det  1 2−λ 0  = (1 − λ) det
1 2−λ
1 1 1−λ

fA (λ) = (1 − λ)(λ2 − 4λ + 3) = (1 − λ)(λ − 3)(λ − 1) = (1 − λ)2 (3 − λ)

 EIGENVALUES AND EIGENVECTORS 7

Hence, eigenvalues are λ = 1 and λ = 3. Have almu(1) = 2 and almu(3) = 1. We

compute      
1 1 0 1 0
E1 = ker(A − I3 ) = ker 1 1 0 = span(−1 , 0)
1 1 0 0 1
Hence, gemu(1) = 2
Likewise,
   
−1 1 0 1
E3 = ker(A − 3I3 ) = ker  1 −1 0  = span(1)
1 1 −2 1
Hence, gemu(3) = 1. As gemu(1)+gemu(3) = 2+1 = 3 the matrix is diagonalizable.
Indeed, setting  
1 0 1
S = −1 0 1
0 1 1
We have  
1 0 0
S −1 AS = 0 1 0
0 0 3

7. Eigenvalues of linear transformations

Fix a linear space V and consider a linear transformation T : V → V . A scalar
λ is an eigenvalue of T , if
T (f ) = λf
for some nonzero (nonneutral) element f ∈ V . In general we refer to such f as a
eigenvector. If V is a space of functions, then it is customary to call f an eigen-
function, etc. If B = (f1 , . . . , fn ) is a basis of V and each fi is an eigenvector, then
we say B is an eigenbasis of T . If B is an eigenbasis of T , then T is diagonalizable
as  
λ1 · · · 0
[T ]B =  ... .. ..  .

. . 
0 · · · λn
EXAMPLE: If V = C , and D(f ) = f 0 , then
∞

d kx
D(ekx ) = e = kekx
dx
so each fk (x) = ekx is an eigenfunction and every scalar k ∈ R is an eigenvalue.
EXAMPLE: Consider the map T : P2 → P2 given by T (p) = p(2x + 1). Is T
diagonalizable? As usual it is computationally more convenient to work in some
basis. To that end, let U = (1, x, x2 ) be the usual basis of P2 . As
 
1
[T (1)]U = [1]U = 0
0
 
1
[T (x)]U = [2x + 1]U = 2
0
8 EIGENVALUES AND EIGENVECTORS

 
1
[T (x2 )]U = [4x2 + 4x + 1]U = 4 ,
4
the associated matrix is  
1 1 1
A = [T ]U = 0 2 4 .
0 0 4
This matrix is upper triangular, with distinct eigenvalues 1, 2 and 4 This means T
is also diagonalizable and has the same eigenvalues. We compute (for A)
   
0 1 1 1
E1 = ker(A − I3 ) = ker 0 1 4 = span(0)
0 0 3 0
   
−1 1 1 1
E2 = ker(A − 2I3 ) =  0 0 4 = span(1)
0 0 2 0
   
−3 1 1 1
E3 = ker(A − 4I3 ) =  0 −2 4 = span 2
0 0 0 1
 
1 1 1
Hence, A can be diagonalized by S = 0 1 2. Going back to T we check
0 0 1
T (1) = 1
T (1 + x) = 1 + (2x + 1) = 2(x + 1)
T (1 + 2x + x2 ) = 1 + 2(2x + 1) + (2x + 1)2 = 4(1 + 2x + x2 )
In particular, B = (1, 1 + x, 1 + 2x + x2 ) is an eigenbasis and
 
1 0 0
[T ]B = 0 2 0 .
0 0 4

8. Eigenvalues and Similarity

There is a close relationship between similar matrices and their eigenvalues. This
is clearest for diagonalizable matrices but holds more generally.
Theorem 8.1. If A is similar to B
(1) fA (λ) = fB (λ)
(2) rank(A) = rank(B) and null(A) = null(B)
(3) A and B have the same eigenvalues with the same algebraic and geometric
multiplicities (the eigenvectors are in general different)
(4) trA = trB and det(A) = det(B)
Proof. Proof of claim (1): A is similar to B means B = S −1 AS for some invertible
S. Hence,
fB (λ) = det(B − λIn ) = det(S −1 AS − λIn ) = det(S −1 (A − λIn )S)
= det(A − λIn ) = fA (λ).
Claim (2) was shown in a previous handout (the one on similar matrices).
 EIGENVALUES AND EIGENVECTORS 9

Claim (3) can be shown as follows: By claim (1), fA (λ) = fB (λ) and so A and
B have the same eigenvalues with the same algebraic multiplicities. Furthermore,
if λ is an eigenvalue of both A and B, then A − λIn is similar to B − λIn . Hence,
by claim (2)
gemuA (λ) = null(A − λIn ) = null(B − λIn ) = gemuB (λ)
so the geometric multiplicities are the same as well. Notice Eλ (A) 6= Eλ (B) in
general. Finally, claim (4) Follows from claim (1) and the observation that the
characteristic polynomial encodes the trace and determinant. 
   
2 3 3 2
EXAMPLE: The matrix is not similar to as their traces are
5 7 8 5
different.
Even if a matrix is not diagonalizable, it is still similar to one of a list of canonical
matrices. To do this is full generality is beyond the scope of this course. To illustrate
the idea we present the list for 2 × 2 matrices.
Theorem 8.2. Any A ∈ R2×2 is similar to one of the following:
 
λ1 0
(1) for λ1 , λ2 ∈ R. This occurs when A is diagonalizable.
0 λ2
 
λ 1
(2) . This occurs when A has repeated eigenvalue λ with gemu(λ) = 1.
0 λ
Such
 A is sometimes called  defective. 
λ1 −λ2 p cos θ − sin θ
(3) = λ21 + λ22 where λ2 > 0 and π > θ > 0.
λ2 λ1 sin θ cos θ
This corresponds to an A with no real eigenvalues (in this case the complex
eigenvalues of A are λ1 ± iλ2 ).