You are on page 1of 13

MATH 201 COMPLEX MATRICES

CRAIG VAN COEVERING

Abstract. These notes cover important results for complex ma-


trices. We discuss the notion of similar matrices, something we
have considered already in diagonalization. Spectral Theorem for
Hermitian and more generally normal matrices is proved, which
give an orthogonal diagonalization for such matrices. This give a
diagonalization by an unitary matrix. This result is a consequence
of Schur’s Lemma, which says that every matrix is similar by a
unitary matrix to an upper triangular matrix.
The Jordan decomposition is presented and a proof is sketched.
The Jordan form is a canonical form for general complex mantrices.
In general, complex matrix is not diagonalizable, so this form give
a canonical form which involves non-diagonal pieces.
Both of these canonical forms has important applications, such
as in solving linear systems of differential equations.

1. Similar matrices
Let A be an n × n matrix. Recall that A defines a linear transfor-
mation A : Cn → Cn . Suppose that {v1 , . . . , vn } ⊂ Cn is another basis
of Cn . Thus the n × n matrix with these vectors as columns
 
S = v1 · · · vn
is invertible. The matrix B = (bij of this linear transformation in terms
of the basis {v1 , . . . , vn } is defined by
n
X
Avj = bij vi , j = 1, . . . , n
i=1

which in matrix form is AS = SB. So we have B = S −1 AS.


Similar matrices: n × n matrices A and B are similar if there
is an invertible n × n matrix S so that
B = S −1 AS.

Date: December 24, 2021.


1
2 CRAIG VAN COEVERING

We have seen this before when diagonalizing matrices. If each vj is an


eigenvector with eigenvalue λj ,
Avj = λj vj , j = 1, . . . , n
then B = Λ
0
 
λ1
Λ=
 .. 
. 
0 λn
One can consider the similar matrix B to be either the matrix given
by the relation B = S −1 AS or the matrix of the linear transforma-
tion in terms of the alternative basis {v1 , . . . , vn }. The diagonalization
problem can be phrased in two ways. Firstly, a matrix A is diagonal-
izable if it is similar to a diagonal matrix Λ. Secondly, the matrix A is
diagonalizable if there is a basis of eigenvectors of the transformation
A : Cn → Cn .
Matrices being similar is an equivalence relation. This means that
the relation satisfies three properties, where we denote the relation of
A being similar to B, B = S −1 AS for some invertible S, by A ∼ B:
(1) Reflexive: A ∼ A.
(2) Symmetric: If A ∼ B, then B ∼ A.
(3) Transitive: If A ∼ B and B ∼ C, then A ∼ C.
The first is clear; just take S = I. For the second, suppose B = S −1 AS,
then A = SBS −1 . For the third, suppose that B = S −1 AS and C =
T −1 BT , then C = T −1 (S −1 AS)T = (ST )−1 A(ST ).
A property of an equivalence relation is that the set of n×n matrices
is partitioned into equivalence classes. An equivalence class is a subset
in which all the matrices are similar to one another. Note for example,
that I is only similar to itself, because S −1 IS = S −1 S = I.
A canonical form of a matrix A is a particularly simple form of
a matrix B which A is similar to. It can be considered to be the
same linear transformation in a different basis. A diagonal matrix is
a very simple form, but as we have seen, not every matrix is similar
to a diagonal matrix. A matrix must have enough eigenvalues to be
diagonalizable. For each eigenvalue λ the geometric multiplicity of λ
must equal the algebraic multiplicity of λ.
Example 1.1.  
1 1 0 0
 1 2 0
A= 
 1 3
0 1
MATH 201 COMPLEX MATRICES 3

A has the eigenvalue λ = 1 with algebraic multiplicity 4, but with


geometric multiplicity 1.
0 1 0 0
 
 0 2 0
A−I =
 0 3

0 0
has 1 dimensional nullspace. ♦

But we have seen that if the eigenvalues are all distinct then it it
diagonalizable. In the next section we will see other cases in which a
matrix is diagonalizable.
Proposition 1.2. Similar matrices have the same characteristic poly-
nomials. That is, if B = S −1 AS, then PA (z) = PB (z).
In particular, similar matrices have the same eigenvalues.
This is not surprising, as we have seen that similar matrices can
be considered as matrices of the same linear transformation, but for
different bases.
Proof.
PB (z) = det(B − zI) = det(S −1 AS − zI)
= det(S −1 (A − zI)S)
= det(S −1 ) det(A − zI) det(S)
= det(A − zI)
= PA (z)

One can easily check that if λ is an eigenvalue of A with eigenvector
v, then S −1 v is an eigenvector of B with eigenvalue λ.

2. Spectral theorem
In this section we will consider matrices up to unitary similarity,
B = U ∗ AU where U is an unitary matrix. Under a restriction on A,
that it be normal, we will prove that A is unitary similar to a diagonal
matrix.
Lemma 2.1 (Schur’s lemma). Let A be an n × n matrix with complex
entries. There exists an unitary matrix U and an upper triangular
matrix T , with complex entries, so that
U ∗ AU = T.
4 CRAIG VAN COEVERING

Note that since A is similar to T , they have the same eigenvalues,


and must be the diagonal entries of T .
Proof. A has an eigenvalue λ1 and corresponding eigenvector u1 , so
Au1 = λ1 u1 . Add vectors u2 , . . . , un so that {u1 , u2 , . . . , un } is an
orthonormal basis. So
 
U1 = u1 · · · un
is unitary. Then
λ1 ∗ · · · ∗
 
 0
U1∗ AU1 = 

.
 .. A1 
0
Then the (n − 1) × (n − 1) matrix A1 must have an eigenvalue λ2 with
eigenvector v2 . Add vectors v3 , . . . , vn ∈ Cn−1 to make a unitary basis
v2 , v3 , . . . , vn of Cn−1 . Then
1 0 ··· 0
 
 0   
U2 =  .
 .. v2 · · · vn 

0
is an unitary matrix. And
λ1 ∗ ∗ · · · ∗
 
 0 λ2 ∗ · · · ∗ 
U2∗ U1∗ AU1 U2 0 0
 
= 
 .. .. A 
 . . 2 
0 0
Continue by finding an eigenvalue λ3 and eigenvector of the (n − 2) ×
(n − 2) matrix A2 . In the end we get unitary matrices U1 , U2 , . . . Un ,
and if we set U = U1 U2 · · · Un , U ∗ AU = T is upper triangular. 
The Spectral theorem will apply to a class of complex matrices more
general than Hermitian matrices.
Normal matrix: A complex n × n matrix A is normal if A∗ A =
AA∗ .
In other words, A and A∗ commute.
Many of the classes of matrices we have seen are normal.
(1) Hermitian matrices: If A∗ = A, then A∗ A = AA∗ clearly
holds.
(2) Skew-Hermitian matrices: This means A∗ = −A.
MATH 201 COMPLEX MATRICES 5

(3) Unitary matrices: We have U ∗ = U −1 , so U ∗ U = I and


U U ∗ = I.
Of course if A has real entries then these cases correspond to symmet-
ric, skew-symmetric, and orthogonal matrices respectively, which are
normal. However, not all normal matrices are of these types.
Example 2.2. One can check that
 
1 1 0
A = 0 1 1
1 0 1
is normal. ♦

Theorem 2.3 (Spectral theorem). A complex matrix A is normal if


and only if it is unitary similar to a diagonal matrix, that is there is
an unitary U so that
U ∗ AU = Λ, where
0
 
λ1
Λ=
 ... 

0 λn
Note that we can describe the above three types of normal matrices
by the eigenvalues appearing in Λ. A is Hermitian if and only if its
eigenvalues λi are real; A is skew-Hermitian if and only if its eigenvalues
are imaginary; and A is unitary if and only if each eigenvalue has length
one, |λi | = 1.
Proof. By Schur’s lemma, there is an unitary U so that U ∗ AU = T
and upper triangular matrix. Since A is normal, T is a normal matrix
thus T ∗ T = T T ∗ . The 1, 1 entry of T ∗ T is the length squared of the
first column of T which is |λ1 |2 . But this is the same as the 1, 1 entry
of T T ∗ which is the length squared of the first row of T
Xn
2
|λ1 | + |t1j |2 .
j=2

Therefore t12 = t13 = · · · = t1n = 0. Continue by comparing the 2, 2


entries of T ∗ T which is λ2 |2 and of T T ∗ which is
n
X
2
|λ2 | + |t2j |2
j=3

so t23 = · · · = t2n = 0. Continuing, this proves that all entries of T


above the diagonal are zero, so T = Λ. 
6 CRAIG VAN COEVERING

Example 2.4. A from the previous example is neither Hermitian,


skew-hermitian, nor unitary. So we know that its eigenvalues cannot
be all real in particular. Since it has real entries we know that it must
have one real eigenvalue λ1 and a conjugate pair of complex (non-real)
eigenvalues λ2 , λ2 .
1−λ 1 0
PA (λ) = 0 1−λ 1 = (1 − λ)3 + 1.
1 0 1−λ
Substitute t = 1 − λ, so we need to solve t3 + 1 = 0. Its solutions are
the 3 cube roots of −1 = eiπ , which are
π 2π
t = ei( 3 + 3 k) , k = 0, 1, 2.
So √ √
1 3 1 3
λ = 2, + i , −i
2 2 2 2
The Spectral Theorem tells us that there is a unitary U with

2 √
0 
U ∗ AU =  1
2
+ i 23 √

0 1
2
− i 23
We already knew that A was diagonalizable because it has distinct
eigenvalues. But this is stronger, because it is diagonalized by a unitary
matrix. ♦

The following follows from the Spectral theorem but we can give a
direct proof also.
Proposition 2.5. Let A be a normal matrix. If v, w ∈ Cn are eigen-
vectors
Av = λv, Aw = µw, λ 6= µ
then v ⊥ w.
Lemma 2.6. If A is normal, then kAxk = kA∗ xk for all x ∈ Cn .
In particular, Null(A) = Null(A∗ ).
Proof.
kAk2 = (Ax, Ax)
= (A∗ Ax, x)
= (AA∗ x, x)
= (A∗ x, A∗ x) = kA∗ xk2

MATH 201 COMPLEX MATRICES 7

Lemma 2.7. If A is normal, then Av = λv implies that A∗ v = λv.


Proof. Note that A − λIn is a normal matrix. Since (A − λIn )∗ =
A∗ − λIn , the previous lemma proves that
Null(A − λIn ) = Null(A∗ − λIn ).
Recall that Av = λv is equivalent to v ∈ Null(A − λIn ). 
Proof of proposition. The argument is now identical to the case of an
Hermitian matrix. Suppose Av = λv, so
w∗ Av = λw∗ v.
But from the lemma we have
w∗ Av = (A∗ w)∗ v = (µw)∗ v = µw∗ v,
so we must have w∗ v = 0. 
Second proof of the Spectral theorem. A must have an eigenvalue λ1 and
a unit eigenvector u1 , Au1 = λ1 u1 . Let W1 = {u1 }⊥ ⊂ Cn be the sub-
space perpendicular to u1 .
Now we show that A restricts to a transformation
A1 := A|W1 : W1 → W1 .
Let w ∈ W1 , so v ∗ w = 0. Then by Lemma 2.7
v ∗ Aw = (A∗ v)∗ w = (λv)∗ w = λv ∗ w = 0
so Aw ∈ W1 . Then the transformation A1 has an eigenvalue λ2 and
eigenvector u2 , A1 u2 = λ2 u2 . Since A1 is just the restriction of A to
the subspace W1 , we have
Au2 = λ2 u2 .
Continuing the argument, we get orthonormal u1 , u2 , . . . , un with cor-
responding eigenvalues λ1 , λ2 , . . . , λn . Then if we set
 
U = u1 · · · un
we have U ∗ AU = Λ. 

3. Jordan canonical form


We next consider a canonical form for an arbitrary complex n × n
matrix. Since a matrix need not be diagonalizable the canonical form
will need to be more complicated than a diagonal matrix. A matrix
with real entries might not have real eigenvalues. It will have complex
eigenvalues, but can still fail to be diagonalizable. Nevertheless, the
Jordan form still makes use of the complex numbers.
8 CRAIG VAN COEVERING

Jordan block: The m × m Jordan block, with eigenvalue λ, is


the m × m matrix
 
λ 1
.
 λ .. 
 
Jλ,m =  ...  ,
 1 
λ
with λ on the main diagonal, 1s just above the diagonal, and
all other entries are 0s.
Note that Jλ,m has characteristic polynomial PJλ,m (z) = (−1)m (z−λ)m ,
so its only eigenvalue is λ with multiplicity m. And its eigenspace
Eλ = Span{e1 } has dimension 1.
Theorem 3.1 (Jordan canonical form). A complex n × n matrix A is
similar to a matrix J which is block diagonal
J1
 
J2
S −1 AS = 
 
.. 
 . 
J`
where each block Jk , k = 1, . . . , ` is a Jordan block. The matrix J is
unique up to a reordering of the blocks Jk , k = 1, . . . , `.
Note that different blocks Jk may have the same eigenvalue. And
note that J, and thus A is diagonalizable if and only if each Jordan
block Jk is a 1 × 1 matrix. This is because there is one 1-dimensional
eigenspace for each Jordan block Jk .
The Jordan Canonical form determines a complex matrix up to sim-
ilarity.
Example 3.2. What are the possible matrices up to similarity with
characteristic polynomial
PA (z) = (z − 1)2 (z − 2)3 ?
In other words, we need to find the similarity classes of 5×5 matrices
with eigenvalues 1, of multiplicity 2, and 2, of multiplicity 3.
The Jordan form J consists of Jordan blocks J1,m1 , . . . , J1,mi , J2,n1 , . . . , J2,nj ,
with eigenvalues 1 and 2, with the ranks of the blocks for each eigen-
value adding up to the multiplicity of that eigenvalue. For λ = 1 the
possible blocks are either J1,1 + J1,1 , 2 Jordan blocks making up a di-
agonal 2 × 2 block, or a single J1,2 . For λ = 2, we have J2,1 + J2,1 + J2,1 ,
J2,2 + J2,1 , or J2,3 . This is six total possibilities. One possibility for
example consists of the blocks J1,2 and J2,3
MATH 201 COMPLEX MATRICES 9

 
1 1
 1 
 

 2 1 
 2 1
2
This has a 1-dimensional eigenspace for λ = 1 and for λ = 2. ♦

Let A be an n × n matrix.
Generalized eigenspace: A vector v is a generalized eigenvec-
tor of A with corresponding eigenvalues λ if
N
A − λI v = 0
for some N ≥ 1. If (A − λ)N −1 v 6= 0, then the generalized
eigenvector v has rank N .
N
Note if v is a generalized eigenvalue then A − λI v = 0 for some
1 ≤ N ≤ n, and v is an ordinary eigenvector if and only if satisfied for
N = 1.
The importance of this concept comes from the fact that A has n
linearly independent generalized eigenvectors, in contrast to ordinary
eigenvectors. Let Gλ denotes the generalized eigenspace of λ ∈ C
N
Gλ = v ∈ Cn | A − λI v = 0 for some N >> 0


for λ = λi , i = 1, . . . k the eigenvalues of A. Then the spaces span Cn


Gλ1 + · · · + Gλk ,
and the spaces are independent, that is if vi ∈ Gλi and
v1 + v2 + · · · + vk = 0
then v1 = v2 = · · · = vk = 0. This usually written with the notation
k
M
Gλi = Cn .
i=1

Jordan chain: If pr is a generalized eigenvector of rank r, then


{pr , pr−1 , . . . , p1 } where
r−j
pj = A − λI pr
is a Jordan chain generated by pr .
One can check that a Jordan chain {pr , pr−1 , . . . , p1 } is linearly inde-
pendent.
10 CRAIG VAN COEVERING

Proof of Jordan canonical form. We will prove that the n × n matrix


A has a Jordan decomposition by induction on n. The 1 × 1 case is
immediate. Let λ be an eigenvalue of A The range of A − λI, denoted
Ran(A−λI) is a subspace invariant by A. Also, since λ is an eigenvalue
of A, Ran(A−λI) has dimension r < n. Let A0 be the restriction of A to
Ran(A − λI). By the induction hypothesis, there is a basis {p1 , . . . , pr }
such that A0 in this basis is in Jordan canonical form.
We consider the subspace
 
W = Ran A − λI ∩ Null A − λI .
 
If W = {0} then we have Cn = Ran A − λI ⊕ Null A − λI , giving
the desired Jordan decomposition.
Suppose W 6= {0}, and the dimension of W is s ≤ r. Each vector in
W is an eigenvector of A0 , so there are s independent Jordan chains of
A0 . So we can choose a basis {p1 , . . . , pr } of Ran(A − λI) so that the s
vectors {pr−s+1 , . . . , pr } ⊂ W are lead vectors in Jordan chains of A0 .
We can extend these Jordan chains by taking the preimages of these
lead vectors, that is, qi , i = r − s + 1, . . . , r so that

A − λI qi = pi

Note that no linear combination of the qi lie in Null A − λI , nor can
it lie in Ran A − λI as it would contradict the assumption that the pi
are lead vectors of the Jordan chains. The {qr−s+1 , . . . , qr } are linearly
independent, being preimages of a linearly independent set.
Then choose a linearly independent set {u1 , . . . , ut } that spans

Null A − λI /W.
The union of the three sets {p1 , . . . , pr }, {qr−s+1 , . . . , qr }, and {u1 , . . . , ut }
is linearly independent, is either an eigenvector of generalized eigenvec-
tor of A, and is n vectors by the Rank-Nullity theorem. Therefore A
is in Jordan canonical form in this basis.
We need to show that the Jordan canonical form is unique up to
the order of the Jordan blocks. Note that knowing the algebraic and
geometric multiplicities of the eigenvalues is not sufficient to determine
the Jordan form of A. Let mλ be the algebraic multiplicity of an
eigenvalue. The structure of the Jordan form can be determined from
the ranks of the powers
j
A − λI , 1 ≤ j ≤ mλ .
Suppose A has only one eigenvalue λ, so mλ = n. The smallest integer
k1 such that
k
A − λI 1 = 0
MATH 201 COMPLEX MATRICES 11

is the size of the largest Jordan block. The rank of


k −1
A − λI 1
is the number of Jordan blocks of size k1 . Similarly, the rank of
k −2
A − λI 1
is twice the number of Jordan blocks of size k1 plus the number of
Jordan blocks of size k1 − 1. Proceeding, we can determine the number
of Jordan blocks of each size.
So if J and J 0 are two Jordan normal forms of A. First they have
the same eigenvalues with multiplicities. The procedure just given can
determine the Jordan blocks corresponding to each eigenvalue. 

4. Applications
The Jordan canonical form has applications to problems where we
used diagonalization. Since not every matrix is diagonalizable is may
be necessary to just the Jordan form.
One application of matrix algebra is in differential equations. Sup-
pose we have an n × n linear system ordinary differential equation
ẋ(t) = Ax(t)
where A is an n × n matrix and
x1 (t)
 

x(t) =  ... 
xn (t)
We have seen that if u is an eigenvector of eigenvalue λ then
x(t) = ceλt u
is a solution to the system. Thus we get an independent solution for
each independent eigenvector. This gives a complete set of solutions
if A is diagonalizable. But in general, we must look for other types of
solutions to get a complete set of n solutions.
The matrix exponential will give a complete set of solutions.
∞ k k
tA
X t A A2
e = = I + tA + t2 + ···
k=0
k! 2
can be shown to converge and give a differentiable function of t. Then
x(t) = etA u0
satisfies ẋ(t) = AetA u0 = Ax(t) with x(0) = u0 .
12 CRAIG VAN COEVERING

We have seen that we can find etA by diagonalizing. If A = SΛS −1 ,


then
−1
etA = eStΛS = SetΛ S −1 .
And etΛ can be computed as
0
 
etλ1
etΛ =
 ... 

0 etλn
In the general case we can use the Jordan form
A = SJS −1
So
etA = SetJ S −1 ,
and the problem is reduced to finding etJλ,n for a Jordan block. Note
that Jλ,n = λIn + N , where
0 1
 
 0 1 
N =
 .. 
.
 0 . 
 1
0
Note that N n = 0, N is nilpotent, and λIn , N commute. We have
∞ k ∞ k  
tJλ,n
X t (λI + N )k X tk X k k−j j
e = = λ N
k=0
k! k=0
k! j=0 j
∞ X ∞
tk k k−j j
X  
= λ N
j=0 k=j
k! j
∞ X ∞
tj tk
 
X k+j k j
= λ N
j=0 k=0
(k + j)! j
∞ k k
n−1 j X
X t t λ
= Nj
j=0
j! k=0
k!
n−1 j
X t
= etλ N j
j=0
j!
tj tλ
which gives the entries of etJλ,n , i.e. the function j!
e is in the jth
super-diagonal.
MATH 201 COMPLEX MATRICES 13

Example 4.1. Consider Jλ,4 , the 4×4 Jordan block. The computation
above gives  2 3
1 t t2 t6
0 1 t t2 
etJλ,4 = eλt  2
0 0 1 t 
0 0 0 1
Note that the columns give independent solutions to the correspond-
ing differential equation.

Department of Mathematics, Boğaziçi University, 34342 Bebek, Is-


tanbul, Turkey
Email address: craig.coevering@boun.edu.tr

You might also like