Spring 2026 MATH 235 Lecture 22

Minsik Han (University of Rochester)

Apr 8, 2026

Review
Let V be a finite-dimensional vector space over a field F of dimension n and T : V → V be a
linear operator on V . Also, let A ∈ Mn (F ).

• v ∈ V is an eigenvector of T if v is nonzero and T (v) = λv for some λ ∈ F .

• v ∈ F n is an eigenvector of A if it is an eigenvector of LA , or equivalently v is nonzero

and Av = λv for some λ ∈ F .

• T is diagonalizable if [T ]β is diagonal for an ordered basis β for V . For such an ordered

basis β, every vector in β is an eigenvector of T .

• A square matrix A ∈ Mn (F ) is diagonalizable if LA is diagonalizable.

• A is diagonalizable if and only if there is an invertible matrix Q such that Q−1 AQ is a

diagonal matrix. Moreover, every column of Q is an eigenvector of A.

• The characteristic polynomial of A is fA (t) = det(A − tIn ) . λ ∈ F is an eigenvalue of A

if and only if
det(A − λIn ) = 0.

• The characteristic polynomial of T is fT (t) = det([T ]β − tIn ) for an ordered basis β for
V , which does not depend on the choice of β. λ ∈ F is an eigenvalue of T if and only if

det([T ]β − λIn ) = 0.

Today’s topics
1. Diagonalizability

1
1 Diagonalizability
In today’s notes, we assume that F is a field, V is a finite-dimensional vector space over F of dimen-
sion n, and T : V → V be a linear operator on V .

We studied in MATH 165 how to determine if a given square matrix is diagonalizable (or, in an
older term, nondefective) by solving SLEs for each eigenvalue. The purpose of this section is to
prove that it really works. Because square matrices are interchangeable with linear operators on
finite-dimensional vector spaces, all results will be stated in terms of linear operators; however,
every result can be converted to a ‘matrix version’ by replacing T with A ∈ Mn (F ).
We first define eigenspaces. Let λ ∈ F be an eigenvalue of T . Then a nonzero vector v ∈ V is an
eigenvector of T corresponding to λ if and only if v satisfies

(T − λIV )(v) = T (v) − λv = 0V .

Therefore, all eigenvectors corresponding to a single eigenvalue λ, along with the zero vector,
form the null space of the linear operator T − λIV . Of course, it is a subspace of V .

Definition (Eigenspace)

For each eigenvalue λ ∈ F of T , the subspace Eλ = N (T − λIV ) of V is called the

eigenspace of T corresponding to λ.

Every eigenspace has a positive dimension, because it has at least one nonzero eigenvector.
On the other hand, we define the multiplicity of an eigenvalue using the characteristic polyno-
mial. See Lecture 0 notes (Section 5) for a review of the factorization of polynomials.

Definition (Multiplicity of an eigenvalue)

For each eigenvalue λ ∈ F of T , the multiplicity of λ is defined as the largest positive

integer m for which (t − λ)m is a factor of the characteristic polynomial of T .

For example, if the characteristic polynomial of T is factored as

fT (t) = (−1)n (t − λ1 )m1 (t − λ2 )m2 · · · (t − λk )mk ,

then mi is the multiplicity of λi for each i = 1, 2, . . . , k. If a polynomial is factored as the product

of linear polynomials, like the above one, then we say that the polynomial splits.

2
The key fact on determining diagonalizability is that the dimension of any eigenspace of T is
bounded by the multiplicity of the corresponding eigenvalue.

Proposition 1
 
A B
Let X =  be a block matrix such that A and C are square matrices (so that X is
O C
also a square matrix) and O is the zero matrix (of the appropriate size). Then
det(X) = det(A) det(C).

Proof. See HW 10, Problem 2.

Theorem 1
Let λ ∈ F be an eigenvalue of T with multiplicity m. Then 1 ≤ dim(Eλ ) ≤ m.

Proof. We already observed 1 ≤ dim(Eλ ).

For the other side, let p = dim(Eλ ) and choose any ordered basis {v1 , v2 , . . . , vp } for Eλ . This
ordered basis can be extended to an ordered basis β = {v1 , v2 , . . . , vp , vp+1 , . . . , vn } for V .
Let A = [T ]β . For each i = 1, 2, . . . , p, vi is an eigenvector of T corresponding to λ, so

[T (vi )]β = λvi = λei .

It follows that A can be written as the block matrix

 
λIp B
A=  ,
O(n−p)×p C

for some matrices B and C.

By Proposition 1, the characteristic polynomial of T is
 
(λ − t)Ip B
det([T ]β − tIn ) = det(A − tIn ) = det  
O(n−p)×p C − tIn−p
( ) ( )
= det (λ − t)Ip det C − tIn−p

= (−1)p (t − λ)p det(C − tIn−p ) .

Because (t − λ)p is already a factor of the characteristic polynomial of T , the multiplicity m of λ

must be at least p, i.e., p ≤ m, as desired.

3
By definition, the sum of multiplicities of all eigenvalues is at most n, the degree of the character-
istic polynomial. Therefore, Theorem 1 implies that the sum of dimensions of all eigenspaces is
also at most n. We now prove that T is diagonalizable if and only if this sum is equal to n.

Proposition 2
Let λ1 , λ2 , . . . , λk be distinct eigenvalues of T . For each i = 1, 2, . . . , k, let Si be a finite LI
set of eigenvectors of T corresponding to λi . Then the union S1 ∪ S2 ∪ · · · ∪ Sk is also LI.

Proof. We use mathematical induction on k.

If k = 1, there is nothing to prove because S1 is already LI.
Now assume that the result holds for k − 1 and prove that it also holds for k. Let S1 , S2 , . . . , Sk
be any LI sets described in the statement.
Label the vectors as Si = {vi1 , vi2 , . . . , vini } for each i = 1, 2, . . . , k, and suppose that

∑
k ∑
ni
aij vij = 0V
i=1 j=1

for some aij ∈ F . We claim that aij = 0 for all i = 1, 2, . . . , n and j = 1, 2, . . . , ni .

To show this, we apply the linear operator T − λk IV to both sides. Then
 
∑k ∑
ni
0V = (T − λk IV )(0V ) = (T − λk IV ) aij vij 
i=1 j=1
k ∑
∑ ni
= aij ( T (vij ) − λk vij )
i=1 j=1

∑
k ∑
ni
= aij (λi − λk )vij
i=1 j=1

∑∑
k−1 ni
= aij (λi − λk )vij .
i=1 j=1

Note that λi − λk ̸= 0 for all i = 1, 2, . . . , k − 1.

Meanwhile, the induction hypothesis implies that S1 ∪ S2 ∪ · · · ∪ Sk−1 is LI. Therefore, we have
aij = 0 for all i, j unless i = k . However, it then follows that

∑
nk
akj vkj = ak1 vk1 + ak2 vk2 + · · · + aknk vknk = 0V .
j=1

Because Sk is also LI, we have aij = 0 even if i = k. This completes the proof.

4
 Theorem 2
Suppose that the characteristic polynomial of T is factored as
fT (t) = (−1)n (t − λ1 )m1 (t − λ2 )m2 · · · (t − λk )mk .
That is, we assume that:
• the characteristic polynomial of T splits;
• T has k distinct eigenvalues λ1 , λ2 , . . . , λk ; and
• each eigenvalue λi has multiplicity mi .
Then T is diagonalizable if and only if dim(Eλi ) = mi for all i = 1, 2, . . . , k.

Proof. Let di = dim(Eλi ) for each i = 1, 2, . . . , k. Note that Theorem 1 implies di ≤ mi for all i.
First, suppose that T is diagonalizable. Then there is an ordered basis β for V consisting of
eigenvectors of T . Each vector in β corresponds to exactly one eigenvalue, so β can be partitioned
into β1 , β2 , . . . , βk where βi = β ∩ Eλi for each i. Because βi is an LI subset of Eλi , there are at
most di vectors in βi . Therefore, we have

∑
k ∑
k ∑
k
n = #β = #βi ≤ di ≤ mi = n.
i=1 i=1 i=1

Thus, every equality must hold, and particularly we have di = mi for all i = 1, 2, . . . , k.
Conversely, suppose that di = mi for all i = 1, 2, . . . , k and choose any ordered basis βi for each
Eλi . Let β = β1 ∪ β2 ∪ · · · ∪ βk . Then

∑
k ∑
k ∑
k
#β = #βi = di = mi = n
i=1 i=1 i=1

and β is LI by Proposition 2. Therefore, β is an ordered basis for V consisting of eigenvectors of

T . It follows that T is diagonalizable.

Remark. If T is diagonalizable and β is an ordered basis for V defined in the proof of Theorem 2,
then we have  
λ1 Im1 O ··· O
 

 O λ2 Im2 · · · O  
[T ]β = 
 .. .. . . 
 . . .. .. 
 
O O ··· λk Imk
where each O denotes the zero matrix (with the appropriate size).

5
 Example 1
 
1 1
Let A =  ∈ M2 (R). We computed the characteristic polynomial of A as
4 1
det(A − tI2 ) = (t − 3)(t + 1),
so that A has two eigenvalues λ1 = 3 and λ2 = −1, both with multiplicity 1. Because
[LA ]γ = A where γ is the standard ordered basis for R2 , the left-multiplication transfor-
mation LA : R2 → R2 has the same eigenvalues with the same multiplicities.
Let E1 and E2 be the eigenspace of LA corresponding to the eigenvalues λ1 and λ2 , respec-
tively. Then Theorem 1 says
dim(E1 ) = dim(E2 ) = 1.
Therefore, by Theorem 2, LA is diagonalizable (and so is A).
To find a basis for R2 such that [LA ]β is diagonal, we solve two homogeneous SLEs
(A − λ1 I2 )x = 0 and (A − λ2 I2 )x = 0
to find a basis
   
 1   1 
β1 =   and β2 =  
 2   −2 

for N (LA − λ1 IR2 ) and N (LA − λ2 IR2 ), respectively. Then

   
 1 1 
β = β1 ∪ β2 =  ,  
 2 −2 
is a basis for R2 consisting of eigenvectors of LA , and it follows that
   
λ1 0  3 0 
[LA ]β =  = .
0 λ2 0 −1
   
1 1 3 0
Note that if Q = [IR2 ]γβ =  , then D = Q−1 AQ =  is diagonal.
2 −2 0 −1

Example 1 suggests the following result, given without proof (but not hard to prove).

Proposition 3
Suppose that T has n distinct eigenvalues. Then T is diagonalizable.

6
We finish with a remark on the relationship between diagonalizability and the base field. One
of the assumptions in Theorem 2 was that the characteristic polynomial of T splits, which is not
always the case. For example, if
 
0 1
A= ∈ M2 (R),
−1 0

then the characteristic polynomial of LA ∈ L(R2 ) is

 
−t 1 
det(A − tI2 ) = det  = t2 + 1,
−1 −t

which cannot be factored as a product of linear polynomials with real coefficients. Consequently,
there is no λ ∈ R with det(A − λI2 ) = 0, so LA (or A) has no real eigenvalues and therefore is
not diagonalizable. We can even prove that, if the characteristic polynomial of T does not split,
then T is not diagonalizable.
However, it is known that every polynomial splits if we allow factors with complex coefficients.
For instance, the same polynomial t2 + 1 is factored as

t2 + 1 = (t + i)(t − i).

It follows that A is diagonalizable when it is regarded as a complex matrix; equivalently, LA is

diagonalizable as a linear operator on C2 instead of R2 . Because of this, we may omit the first
assumption in Theorem 2 when the base field is C.