MTH 311: Advanced Linear Algebra

Semester 1, 2024-2025

Prahlad Vaidyanathan (substituting for Rohit Holkar)

I. Canonical Forms
1. Review
Definition 1.1. A matrix A = (ai,j ) ∈ Mn (F ) is said to be upper triangular if

ai,j = 0

whenever i > j.

Proposition 1.2. Let T ∈ L(V ) and B = {v1 , v2 , . . . , vn } be an ordered basis of V .

Then, the matrix [T ]B is upper triangular if and only if

T (vk ) ∈ span{v1 , v2 , . . . , vk }

for all 1 ≤ k ≤ n.

Definition 1.3.

(i) A matrix A ∈ Mn (F ) is said to be block diagonal if it can be written in the form

 
A1 0 0 ... 0 0
 0 A2 0 ... 0 0 
 
A =  0 0 A3 ... 0 0
 

 .. .. 
 . . 
0 0 0 . . . 0 Ak

where each Ai is an square matrix.

(ii) An operator T ∈ L(V ) is said to be block diagonalizable if there is an ordered

basis B of V such that [T ]B is block diagonal.

Note: For convenience, we may sometimes write such a matrix as

 
Ar1 ×r1 0 0
A= 0 Br2 ×r2 0 
0 0 Cr3 ×r3

where A ∈ Mr1 (F ), B ∈ Mr2 (F ) and C ∈ Mr3 (F ).

2
Definition 1.4. Let V be a vector space and W1 , W2 , . . . , Wk be subspaces of V . We
say that V is a direct sum of {W1 , W2 , . . . , Wk } if every v ∈ V can be expressed uniquely
in the form
v = w1 + w2 + . . . + wk
where wi ∈ Wi for all 1 ≤ i ≤ k. If this happens, we write
k
M
V = W1 ⊕ W2 ⊕ . . . ⊕ Wk = Wi .
i=1

Proposition
Lk 1.5. Let V be a vector space and W1 , W2 , . . . , Wk be subspaces of V . Then,
V = i=1 Wi if and only if the following two conditions are met:

(i) V = W1 + W2 + . . . + Wk .

(ii) For each 1 ≤ i ≤ k,

Wi ∩ (W1 + W2 + . . . + Wi−1 + Wi+1 + . . . + Wk ) = {0}.

Corollary 1.6. For T ∈ L(V ), there is an ordered basis B of V such that A = [T ]B is

block diagonal if and only if there are T -invariant subspaces W1 , W2 , . . . , Wk such that
k
M
V = Wi .
i=1

Lk
Corollary 1.7. If V = i=1 Wi , then
k
X
dim(V ) = dim(Wi ).
i=1

Moreover, if Bi is a basis for Wi , then the {Bi : 1 ≤ i ≤ k} are mutually disjoint and

B = tki=1 Bi

is a basis for V .

2. Diagonalizable and Triangulable Operators

Definition 2.1.

(i) An operator T ∈ L(V ) is said to be diagonalizable if there is a basis B such that

[T ]B is diagonal.

Note: [T ]B is diagonal if and only if every vector in B is an eigenvector.

3
 (ii) A matrix A ∈ Mn (F ) is said to be diagonalizable if it is similar to a diagonal
matrix.
Proposition 2.2. Let T ∈ L(V ) and λ1 , λ2 , . . . , λk be distinct eigenvalues of T . If
v1 , v2 , . . . , vk are any corresponding eigenvectors, then {v1 , v2 , . . . , vk } is linearly inde-
pendent.
Corollary 2.3. If V is an n-dimensional vector space, then any operator T ∈ L(V ) has
atmost n distinct eigenvalues.
Proof. Any linearly independent set can have atmost n elements.
Corollary 2.4. Let V be an n-dimensional vector space. If T ∈ L(V ) has n distinct
eigenvalues, then T is diagonalizable.
Proof. Write c1 , c2 , . . . , cn for the eigenvalues and v1 , v2 , . . . , vn for some (any) corre-
sponding eigenvectors. Then, B = {v1 , v2 , . . . , vn } is linearly independent. Since dim(V ) =
n, it must be a basis.
Example 2.5.
(i) Let V = R2 and T ∈ L(V ) be the linear transformation

T (x, y) = (−y, x).

In other words, if B denotes the standard ordered basis for V , then

 
0 −1
A := [T ]B =
1 0

We claim that T is not diagonalizable. Suppose it was, then there would be a

basis B 0 = {v, w} consisting of eigenvectors of T . Now, the eigenvalues of T are
the roots of the polynomial
 
−x −1
fT (x) = fA (x) = det(A − xI) = det = x2 + 1.
1 −x

This polynomial does not have any roots in R, so T cannot be diagonalizable.

(ii) If V := C2 , however and T ∈ L(V ) is given exactly as above, then T is diagonal-

izable because it has two distinct eigenvalues i and −i.

(iii) Let V = R2 and T ∈ L(V ) be the linear transformation

T (x, y) = (y, 0).

In other words, if B denotes the standard ordered basis for V , then

 
0 1
A := [T ]B =
0 0

4
 We claim that T is not diagonalizable. Suppose it was, then there would be a
basis B 0 = {v, w} consisting of eigenvectors of T . Now, the eigenvalues of T are
the roots of the polynomial
 
−x 1
fT (x) = fA (x) = det(A − xI) = det = x2 .
0 −x

Hence, the only eigenvalue of T is zero. In particular, it must happen that T (v) =
0 = T (w). In that case,  
0 0
B := [T ]B0 =
0 0
In turn, this would imply that A is similar to the zero matrix. But since A 6= 0,
this is impossible. Therefore, T is not diagonalizable.

Remark 2.6. By Corollary 2.4, if an operator has n distinct eigenvalues, then it is

diagonalizable. Indeed, the matrix is of the form
 
λ1 0 0 ... 0
0 λ2 0 . . . 0 
A =  ..
 
.. .. .. .. 
. . . . .
0 0 0 . . . λn

where the λi are all distinct. However, this is not necessary because it can happen that
the eigenvalues of a diagonal matrix are repeated. In that case, we would have a matrix
of the form  
λ1 I1 0 0 ... 0
 0 λ2 I2 0 . . . 0 
A =  ..
 
.. .. .. .. 
 . . . . . 
0 0 0 . . . λk Ik
where Ij are identity matrices of various sizes.

(i) In this case, consider the characteristic polynomial of A.

 
(x − λ1 )I1 0 0 ... 0
 0 (x − λ2 )I2 0 . . . 0 
det(xI − A) = det 
 
.. .. .. .. .. 
 . . . . . 
0 0 0 . . . (x − λk )Ik
= (x − λ1 )d1 (x − λ2 )d2 . . . (x − λk )dk

So,
cT (x) = cA (x) = (x − λ1 )d1 (x − λ2 )d2 . . . (x − λk )dk

5
 (ii) Moreover, consider the matrix
 
0d1 0 0 ... 0
 0 (λ1 − λ2 )I2 0 ... 0 
λ1 I − A =  ..
 
.. .. .. .. 
 . . . . . 
0 0 0 . . . (λ1 − λk )Ik

Clearly,
B1 = {e1 , e2 , . . . , ed1 } ⊂ ker(λ1 I − A)
Moreover, if X = (x1 , x2 , . . . , xn ) ∈ ker(λ1 I − A), then

(λ1 − λ2 )xd1 +1 = 0 ⇒ xd1 +1 = 0.

Similarly, xj = 0 for all j > d1 . Hence, X = (x1 , x2 , . . . , xd1 , 0, 0, . . . , 0). In other

words,
ker(λ1 I − A) = span{B1 }.
In particular, dim(λ1 I − A) = d1 . Similarly, we see that

dim(ker(λi I − A)) = di

for all 1 ≤ i ≤ k.

(End of Day 1)

Definition 2.7. Let T ∈ L(V ) and λ ∈ F be an eigenvalue of T .

(i) The algebraic multiplicity of λ is the number of times (x − λ) occurs as a factor

of cT (x), the characteristic polynomial of T .

(ii) The geometric multiplicity of λ is the dimension of ker(λI − T ).

We denote these by a(T, λ) and g(T, λ) respectively. We may define these terms for a
matrix A ∈ Mn (F ) as well, in an analogous fashion.

Lemma 2.8. If  
Ar×r Cr×(n−r)
X=
0 B(n−r)×(n−r)
is a block upper triangular matrix, then

det(X) = det(A) det(B).

Proof.

6
 (i) Suppose first that A = Ir×r . Then, by expanding along the first column

det(X) = det(X 0 )

where
0
 
0 I(r−1)×(r−1) C(r−1)×(n−r)
X =
0 B(n−r)×(n−r)
By induction on r, it follows that det(X) = det(B).
(ii) Now suppose B = In−r×n−r , then the same argument shows that det(X) = det(A).
 
Ar×r Cr×(n−r)
(iii) For the general case, suppose X = . We consider two cases:
0 B(n−r)×(n−r)
(a) If det(A) = 0, then A is not invertible. So the columns of A are linearly
dependent. In that case, the columns of X are also linearly dependent, so X
is not invertible. Hence, det(X) = 0 as well.
(b) If det(A) 6= 0, then A is invertible. In that case,

Ir×r A−1 C
  
A 0
X=
0 In−r×n−r 0 B

Since det is multiplicative, we use part (i) and (ii) to obtain

Ir×r A−1 C
   
A 0
det(X) = det det = det(A) det(B).
0 In−r×n−r 0 B

Lemma 2.9. For any T ∈ L(V ) and eigenvalue λ ∈ F , g(T, λ) ≤ a(T, λ).
Proof. Suppose r = g(T, λ) = dim(ker(λI − T )), then there is a basis B of ker(λI − T )
with |B| = r. This is a linearly independent set in V , so it may be extended to form a
basis B 0 of V . Consider
A = [T ]B0
and observe that A has the form
 
λIr×r Cr×(n−r)
A=
0(n−r)×r Dn−r×(n−r)
Hence,  
(xI − λI)r×r Cr×n−r
(xI − A) =
0 (xI − D)
To the characteristic polynomial of A (by Lemma 2.8) is

cT (x) = cA (x) = (x − λ)r det(xI − D).

Hence, r ≤ a(T, λ) as required.

7
Theorem 2.10. An operator T ∈ L(V ) is diagonalizable if and only if the following two
conditions hold:

(i) The characteristic polynomial of T is of the form

cT (x) = (x − λ1 )d1 (x − λ2 )d2 . . . (x − λk )dk

(i.e. The characteristic polynomial splits).

(ii) For each 1 ≤ i ≤ k, g(T, λi ) = a(T, λi ).

Proof.

(i) If T is diagonalizable, then the two conditions hold by Remark 2.6.

(ii) Conversely, suppose these two conditions hold. Write Wi = ker(T −λi I). We claim
that
M k
V = Wi .
i=1

To do this, we verify the two conditions of Proposition 1.5.

(a) We first verify the second condition: Fix 1 ≤ i ≤ k, and suppose

v ∈ Wi ∩ (W1 + W2 + . . . + Wi−1 + Wi+1 + . . . + Wk ).

Then write v = w1 + w2 + . . . + wi−1 + wi+1 + . . . + wk with wj ∈ Wj . More-

over, we may assume that all these terms are non-zero (otherwise, drop them).
However, each such vector is an eigenvector associated to a different eigen-
value, so the set {v, w1 , w2 , . . . , wi−1 , wi+1 , . . . , wk } is linearly independent by
Proposition 2.2. This is impossible unless v = wj = 0. Hence,

Wi ∩ (W1 + W2 + . . . + Wi−1 + Wi+1 + . . . + Wk ) = {0}.

(b) Now consider W = W1 + W2 + . . . + Wk . This is a subspace of V . Moreover,

by part (a),
Mk
W = Wi .
i=1

By Corollary 1.7,
k
X k
X
dim(W ) = dim(Wi ) = di = deg(cT (x)) = n = dim(V ).
i=1 i=1

Hence, W = V . Therefore,
k
M
V = Wi .
i=1

8
 Now choose a basis Bi for each Wi , and set
k
G
B= Bi .
i=1

This is a basis of V and [T ]B is diagonal.

Example 2.11. Let T ∈ L(R3 ) be the linear operator represented in the standard basis
by the matrix  
5 −6 −6
A = −1 4 2
3 −6 −4

(i) We first compute the characteristic polynomial of A: cA (x) = det(xI − A)

 
(x − 5) 6 6
det  1 x − 4 −2 
−3 6 x+4

Subtracting column 3 from column 2 gives us a new matrix with the same deter-
minant. Hence,
   
x−5 0 6 x−5 0 6
f = det  1 x − 2 −2  = (x − 2) det  1 1 −2 
−3 2 − x x + 4 −3 −1 x + 4

Now adding row 2 to row 3 does not change the determinant, so

 
x−5 0 6
cA (x) = (x − 2) det  1 1 −2 
−2 0 x+2

Expanding along column 2 now gives

 
x−5 6
cA (x) = (x − 2) det = (x − 2)(x2 − 3x + 2) = (x − 2)2 (x − 1)
−2 x + 2

(ii) Hence, the characteristic values of T are 1 and 2. Moreover,

a(A, 2) = 2 and a(A, 1) = 1.

(iii) We wish to determine the dimensions of the characteristic spaces, W1 and W2 .

9
 (a) Consider the case c = 1, and the matrix
 
4 −6 −6
(A − I) = −1 3 2
3 −6 −5
Row reducing this matrix gives
   
4 −6 −6 4 −6 −6
0 3 1 
7→ 0 32 1 
=: B
2 2 2
−3 −1
0 2 2
0 0 0
Hence, rank(A − I) = 2 so by the Rank-Nullity theorem,
g(A, 1) = 1.

(b) Consider the case c = 2: We know that

rank(A − 2I) ≥ 1
since (A − 2I) is non-zero. Furthermore, we know that
2 + rank(A − 2I) = rank(A − I) + rank(A − 2I) ≤ dim(R3 ) = 3
So it follows that rank(A − 2I) = 1. As before, this implies that
g(A, 2) = 2.

Hence, we conclude by Theorem 2.10 that T is diagonalizable. Indeed, A is similar

to the matrix  
1 0 0
D = 0 2 0
0 0 2
(iv) We now determine a basis consisting of characteristic vectors:
(a) Consider the case c = 1: Using the matrix B above, we solve the system of
linear equations BX = 0. This gives a solution
α1 = (3, −1, 3)

(b) Consider the case c = 2, and the matrix

 
3 −6 −6
(A − 2I) = −1 2 2
3 −6 −6
Row reducing gives  
3 −6 −6
C = 0 0 0
0 0 0
So solving the system CX = 0 gives two solutions
α2 = (2, 1, 0) and α3 = (2, 0, 1)

10
 (v) Thus, we get an ordered basis

B = {(3, −1, 3), (2, 1, 0), (2, 0, 1)}

consisting of characteristic vectors of T . Furthermore,

 
1 0 0
[T ]B = 0 2 0 := D
0 0 2

(vi) Furthermore, if P is the matrix

 
3 2 2
P = −1 1 0
3 0 1

Then
P −1 AP = D

(End of Day 2)

Example 2.12. Let V = R3 and T ∈ L(V ) be the map T (v) = A(v) where
 
3 1 −1
A = 2 2 −1
2 2 0

(i) Now, cT (x) = cA (x) and

 
3−x 1 −1
cA (x) = − det(A − xI) = − det  2 2 − x −1 
2 2 −x
= (x − 3)[(2 − x)(−x) + 2] + 2[(−x) + 2] − 2[−1 + (2 − x)]
= x3 − 5x2 + 8x − 4
= (x − 1)(x − 2)2 .

So a(A, 1) = 1 and a(A, 2) = 2. Therefore, λ = 1 and λ = 2 are the eigenvalues of

T.

(ii) Since g(A, 1) ≤ 1 and g(A, 1) 6= 0, it follows that

g(A, 1) = 1.

11
(iii) To find eigenvectors associated to c = 2: We need to solve the equation (A −
2I)(X) = 0. Observe that
 
1 1 −1
(A − 2I) = 2
 0 −1
2 2 −2

Writing X = (x, y, z), we wish to solve the system

x+y−z =0
2x − z = 0
2x + 2y − 2z = 0

This yields that X is a scalar multiple of β = (1, 1, 2). Hence,

g(A, 2) = 1.

In particular, g(A, 2) = 1 < a(A, 2). Hence, A is not diagonalizable.

Definition 2.13.
(i) An operator T ∈ L(V ) is said to be triangulable if there is a basis B such that
[T ]B is upper triangular.

(ii) A matrix A in Mn (F ) is said to be triangulable if it is similar to an upper triangular

matrix.
Remark 2.14. Suppose T ∈ L(V ) is triangulable, then there is a basis B such that
 
a1,1 ∗ ∗ ... ∗ ∗
 0 a2,2 ∗ ... ∗ ∗ 
A = [T ]B =  
 0 0 a3,3 ... ∗ ∗ 
0 0 0 . . . 0 an,n

is upper triangular. Therefore,

fT (x) = fA (x) = det(xI − A) = (x − a1,1 )(x − a2,2 ) . . . (x − an,n ).

Therefore, the entries {a1,1 , . . . , an,n } are the eigenvalues of T . Combining like terms,
we may write
fT (x) = (x − λ1 )d1 (x − λ2 )d2 . . . (x − λk )dk .
Theorem 2.15. An operator T ∈ L(V ) in triangulable if and only if the characteristic
polynomial of T is of the form

fT (x) = (x − λ1 )d1 (x − λ2 )d2 . . . (x − λk )dk .

(i.e. the characteristic polynomial splits).

12
Proof. If T is triangulable, then fT (x) splits by Remark 2.14.

Conversely, suppose the characteristic polynomial splits as above. We induct on n :=

dim(V ).

(i) If n = 1, there is nothing to prove.

(ii) If n > 1, assume that the result is true for any vector space over F with dimension
≤ n − 1. Now suppose fT (x) splits, then T has an eigenvalue λ ∈ F . Let x1 ∈ V
be an eigenvector corresponding to λ and let

W := hx1 i

Then, W is a T -invariant subspace, so T induces an operator

T̂ : V /W → V /W given by x + W 7→ T (x) + W.

Moreover, cT̂ (x) divides cT (x), so cT̂ (x) also splits over F . Note that

dim(V /W ) = dim(V ) − dim(W ) ≤ n − 1.

So by the induction hypothesis, V /W has a basis B 0 = {x2 +W, x3 +W, . . . , xn +W }

such that [T̂ ]B0 is upper triangular. In other words, for each 2 ≤ k ≤ n,

T̂ (xk + W ) ∈ span{x2 + W, x3 + W, . . . , xk + W }.

Fix 2 ≤ k ≤ n and scalars {αk,2 , αk,3 . . . , αk,k } so that

T̂ (xk + W ) = αk,2 (x2 + W ) + αk,3 (x3 + W ) + . . . + αk,k (xk + W ).

Since T̂ (xk + W ) = T (xk ) + W , it follows that

k
X
T (xk ) − αk,i xi ∈ W = hx1 i
i=2

So there exists αk,1 ∈ F such that

k
X
T (xk ) = αk,i xi ∈ span{x1 , x2 , . . . , xk }
i=1

This is true for each 2 ≤ k ≤ n, and T (x1 ) ∈ span{x1 }, so T is triangulable.

Corollary 2.16. If V is a vector space over C, then every T ∈ L(V ) is triangulable.

(End of Day 3)

13
3. The Jordan Canonical Form
Definition 3.1. For t ∈ N and λ ∈ F , define Jt (λ) ∈ Mt (F ) to be the matrix
 
λ 1 0 ... 0 0
0 λ 1 . . . 0 0
 
 .. .. .. .. .. .. 
Jt (λ) =  . . . . . . 
 
0 0 0 . . . λ 1
0 0 0 ... 0 λ

Such a matrix is called Jordan matrix (or block) of size t.

The main theorem we wish to discuss is the following.
Theorem 3.2 (Jordan Canonical Form). Let V be a complex vector space and T ∈ L(V ).
Let {λ1 , λ2 , . . . , λk } be the set of distinct eigenvalues of T . Then, there is a basis B of
V such that  
A1 0 0 . . . 0
 0 A2 0 . . . 0 
[T ]B =  ..
 
.. .. .. .. 
 . . . . . 
0 0 0 . . . Ak
is a block diagonal matrix, and each Ai is a block diagonal matrix of the form
 
Jti1 (λi ) 0 0 ... 0
 0 Jti2 (λi ) 0 . . . 0 
Ai =  ..
 
.. .. .. .. 
 . . . . . 
0 0 0 . . . Jtisi (λi )

Moreover, for each 1 ≤ i ≤ k.

ti1 + ti2 + . . . + tisi = dim(Vλi ) = a(T, λi ) and

si = dim(ker(T − λi I)) = g(T, λi ).

Such a matrix is called a Jordan matrix, or the Jordan form of T .

Corollary 3.3. Let V be a complex vector space and T ∈ L(V ). Let {λ1 , λ2 , . . . , λk }
be the set of eigenvalues of T . If a(T, λi ) = g(T, λi ) for all 1 ≤ i ≤ k, then T is
diagonalizable.
Proof. Consider the Jordan Canonical form of A as above. Then,

ti1 + ti2 + . . . + tisi = dim(Vλi ) = a(T, λi ) and

si = dim(ker(T − λi I)) = g(T, λi ).

So if g(T, λi ) = a(T, λi ), it follows that ti1 = ti2 = . . . = tisi = 1. Hence, each Ai is a

diagonal matrix, so A is a diagonal matrix as well.

14
Example 3.4. The following matrices are in Jordan form:

•  
1 0 0 0
0 2 0 0
 
0 0 3 0
0 0 0 4
Here, fA (x) = (x − 1)(x − 2)(x − 3)(x − 4) and a(A, λ) = g(A, λ) = 1 for all
λ ∈ {1, 2, 3, 4}.

•  
2 1
0 2
Here, fA (x) = (x − 2)2 , so a(A, 2) = 2. Moreover,
 
0 1
(A − 2I) = 6= 0
0 0

to g(A, 2) 6= 2. Since 1 ≤ g(A, 2) ≤ 2, it follows that g(A, 2) = 1.

•  
2 1 0
0 2 0
0 0 3
Here, fA (x) = (x − 2)2 (x − 3), so a(A, 2) = 2, a(A, 3) = 1. Moreover, g(A, 2) = 1
and g(A, 3) = 1.

•  
0 1 0 0
0 0 0 0
 
0 0 3 0
0 0 0 3
Here, fA (x) = x2 (x − 3)2 so a(A, 0) = 2, a(A, 3) = 2. Moreover, g(A, 0) = 1 and
g(A, 3) = 2.

Example 3.5. Find the Jordan form of the matrix

 
2 2 3
A= 1 3 3
−1 −2 −2

(i) Find the characteristic polynomial:

 
x − 2 −2 −3
fA (x) = det  −1 x − 3 −3  = (x − 1)3 .
1 2 x+2

15
 (ii) Find eigenvalues and algebraic multiplicity: λ = 1 with a(A, 1) = 3.

(iii) Guess the possible Jordan forms: Hence, the Jordan form of A is of the form
 
Jt1 (1) 0 ... 0
B= 0 Jt2 (1) ... 0 
0 0 . . . Jtk (1)

such that t1 + t2 + . . . + tk = 3 and k = g(A, 1). The only possibilities are

     
1 1 0 1 1 0 1 0 0
B1 := 0 1 1 or B2 := 0 1 0 or B3 := 0 1 0
0 0 1 0 0 1 0 0 1

(iv) Determine the eigenspaces: Consider

 
1 2 3
(A − I) =  1 2 3
−1 −2 −3

Row reducing gives us a row equivalent matrix

 
1 2 3
C = 0 0 0
0 0 0

Solving B(X) = 0 gives us two linearly independent solutions (−2, 1, 0) and

(−3, 0, 1). Hence,
g(A, 1) = 2.
Since g(A, 1) < a(A, 1), A is not diagonalizable.

(v) Determine the Jordan form: Since g(A, 1) = 2, the Jordan form of A has two
blocks. Hence, A is similar to
 
1 1 0
B2 = 0 1 0
0 0 1

Example 3.6. Find the Jordan canonical form for

 
2 −1 0 1
0 3 −1 0
A= 0 1

1 0
0 −1 0 3

16
 (i) Find the characteristic polynomial:
 
x−2 1 0 −1
 0 x−3 1 0 
fA (x) = det  
 0 −1 x − 1 0 
0 1 0 x−3
 
x−3 1 0
= (x − 2) det  −1 x − 1 0 
1 0 x−3
 
x−3 1
= (x − 2)(x − 3) det
−1 x − 1
= (x − 2)(x − 3)[(x − 3)(x − 1) + 1]
= (x − 2)(x − 3)[x3 − 4x + 4]
= (x − 2)3 (x − 3)

(ii) Find the eigenvalues and algebraic multiplicities:

• λ = 2 and a(A, 2) = 3.
• λ = 3 and a(A, 3) = 1.
(iii) Guess the possible Jordan forms: It must be one of the following:
     
3 0 0 0 3 0 0 0 3 0 0 0
0 2 0 0 0 2 1 0 0 2 1 0
B1 = 
0 0 2 0 , B2 = 0 0
   or B3 =  
2 0 0 0 2 1
0 0 0 2 0 0 0 2 0 0 0 2

(iv) Determine the eigenspaces:

(a) λ = 2: Consider  
0 −1 0 1
0 1 −1 0
(A − 2I) = 
0 1 −1 0


0 −1 0 1
Row reducing, gives  
0 −1 0 1
0 1 −1 0
C=
0 0

0 0
0 0 0 0
Hence, the solution space to (A − 2I)(X) = 0 is the same as the solution
space to C(X) = 0, which is spanned by {(1, 0, 0, 0), (0, 1, 1, 1)}. Hence,

g(A, 2) = 2 < a(A, 2).

Thus, A is not diagonalizable.

17
 (b) λ = 3: Since a(A, 3) = 1 it follows that g(A, 3) = 1 as well.
(v) Determine the Jordan form for A: Since g(A, 2) = 2, the Jordan block associated
to λ = 2 has two sub-blocks. Therefore, A is similar to
 
3 0 0 0
0 2 1 0
B2 = 
0 0 2 0


0 0 0 2

(End of Day 4)

4. Generalized Eigenspaces
Definition 4.1. Let T ∈ L(V ) and λ ∈ F be an eigenvalue of T .
(i) The generalized eigenspace associated to λ is the set
∞
[
Vλ := {v ∈ V : (T − λI)j (v) = 0 for some j ≥ 1} = ker(T − λI)j
j=1

(ii) A vector v ∈ Vλ is called a generalized eigenvector of T associated to λ.

For convenience, we write Vλj := ker(T − λI)j .
Example 4.2.
(i) Let A ∈ M2 (R) be the matrix  
1 1
A=
0 1
Then λ = 1 is an eigenvalue. Solving the equation
  
0 1 x
(A − I)X = 0 ⇒ = 0 ⇒ y = 0.
0 0 y
Hence, ker(A − I) = span{e1 }. However,
  
2 0 1 0 1
(A − I) = = 0.
0 0 0 0
Hence, ker(A − I)2 = R2 . Therefore, all vectors in R2 are generalized eigenvectors
of A.
(ii) Let A ∈ M3 (R) be the matrix
 
3 1 0
A = 0 3 0
0 0 2

18
 (a) Then, λ = 3 is an eigenvalue and

ker(A − 3I) = span{e1 }.

Moreover, as above  
0 0 0
(A − 3I)2 = 0 0 0
0 0 4
More generally,  
0 0 0
(A − 3I)j = 0 0 0 
0 0 2j
for all j ≥ 2. Hence,

ker(A − 3I)j = ker(A − 3I)2 = span{e1 , e2 }

for all j ≥ 2.
(b) Finally, λ = 2 is an eigenvalue and
 
1 1 0
A − 2I = 0 1 0
0 0 0

Therefore,
ker(A − 2I) = span{e3 }.
Now observe that
    
1 1 0 1 1 0 1 2 0
(A − 2I)2 = 0 1 0 0 1 0 = 0 1 0
0 0 0 0 0 0 0 0 0

More generally,  
1 j 0
(A − 2I)j = 0 1 0 .
0 0 0
Therefore,
ker(A − 2I)j = ker(A − 2I) = span{e3 }
for all j ≥ 1. Hence, all generalized eigenvectors are eigenvectors.

Lemma 4.3. Fix T ∈ L(V ) and λ ∈ F be an eigenvalue of T .

(i) For each j ∈ N, Vλj ⊂ Vλj+1 .

(ii) Vλ is a T -invariant subspace of V .

19
(iii) There exists r ∈ N such that Vλ = ker(T − λI)r . In other words,

Vλr = Vλr+1 = Vλr+2 = . . . .

(iv) Let Wλ := Range(T − λI)r . Then, V is a direct sum

V = Vλ ⊕ Wλ

and both subspaces are T -invariant.

(v) If S := T |Wλ ∈ L(Wλ ), then λ is not an eigenvalue of S.

Proof.

(i) If v ∈ Vλj then (T − λI)j (v) = 0. Hence, (T − λI)j+1 (v) = 0 as well. Hence,
v ∈ Vλj+1 .

(ii) Since λ is an eigenvalue, Vλ 6= {0}. If α, β ∈ Vλ and c ∈ F , then there exists

i, j ∈ N such that v ∈ Vλi and w ∈ Vλj . Assume i ≤ j, then by part (i), v ∈ Vλj .
However, Vλj is a subspace, so (cv + w) ∈ Vλj ⊂ Vλ . Therefore, Vλ is a subspace.

Moreover, if v ∈ Vλ , then (T − λI)j (v) = 0 for some j ∈ N. Hence,

(T − λI)j T (v) = T (T − λI)j (v) = T (0) = 0.

Therefore, T (v) ∈ Vλ as well so Vλ is T -invariant.

(iii) Choose a basis B = {v1 , v2 , . . . , vk } of Vλ . For each 1 ≤ i ≤ k, there exists si ∈ N

such that vi ∈ Vλsi . Let r = max{s1 , s2 , . . . , sk }. Then by part (i),

vi ∈ Vλr

for all 1 ≤ i ≤ k. Hence, Vλ ⊂ Vλr ⊂ Vλ .

(iv) If R := (T − λI)r , then we wish to show that

V = ker(R) ⊕ Range(R).

We verify the conditions of Proposition 1.5.

(a) We first show that ker(R)∩Range(R) = {0}: Suppose v ∈ ker(R)∩Range(R),
then there exists w ∈ V such that

v = R(w) = (T − λI)r (w).

Since v ∈ ker(R), we have (T − λI)2r (w) = 0, so w ∈ Vλ2r . By part (ii),

Vλ2r = Vλr , so
v = (T − λI)r (w) = 0.

20
 (b) Note that ker(R) and Range(R) are both subspaces of V . Hence, W :=
ker(R) + Range(R) is a subspace of V . By the Rank-Nullity theorem,
dim(W ) = dim(V ).
Hence, W = V .
(v) Let S := T |Wλ : Wλ → Wλ . Suppose v ∈ Wλ such that S(v) = λv. Then,
(T − λI)(v) = 0.
Hence, v ∈ ker(T − λI) ⊂ Vλ . But by part (iv),
Vλ ∩ Wλ = {0}.
Hence, v = 0, so this is impossible.

Theorem 4.4. Let V be a complex vector space and T ∈ L(V ). Let {λ1 , λ2 , . . . , λk } be
the distinct eigenvalues of T . Then,
k
M
V = Vλi
i=1

Proof. We induct on k = the number of distinct eigenvalues of T (note that k ≤ dim(V )

by Corollary 2.3). If k = 1, then T has only one eigenvalue λ. By Lemma 4.3,
V = Vλ ⊕ Wλ
where Wλ = Range(T − λI)r . Moreover, T |Wλ has no eigenvalues (since T has only one
eigenvalue). Therefore, it must happen (by the Fundamental Theorem of Algebra) that
Wλ = {0}. Hence,
V = Vλ .
Now suppose k ≥ 2 and assume that the result is true for any operator R on a vector
space W with ≤ k − 1 eigenvalues. By the Fundamental Theorem of Algebra, T has an
eigenvalue λ1 , so by Lemma 4.3,
V = Vλ1 ⊕ Wλ1 .
Now consider S := T |Wλ1 ∈ L(Wλ1 ). Note that any eigenvalue of S is also an eigenvalue
of T and λ1 is not an eigenvalue of S (by Lemma 4.3). Hence, S has ≤ k − 1 distinct
eigenvalues. By induction, we may express
Wλ1 = Vλ2 ⊕ . . . ⊕ Vλk .
Therefore,
k
M
V = Vλi .
i=1

21
Remark 4.5. In F the above theorem, each Vλi is T -invariant. Therefore, if Bi is a basis
for Vλi , then B = ki=1 Bi is a basis for V and

[T ]B

is block diagonal. It turns out that we can choose Bi in such a way that [T ]B is in Jordan
form.

(End of Day 5)

22