Theorem

Eigenvectors of a unitary matrix associated with distinct eigenvalues are

mutually orthogonal.

Proof
Let U be a unitary matrix, and u, v eigenvectors of U associated with
distinct eigenvalues λ, µ ∈ C, respectively. Then

λ∗ µhu|v i = hλu|µv i = hUu|Uv i = hu|v i.

If hu|v i =
6 0, then we can divide by hu|v i, giving

λ∗ µ = 1.

Since |λ| = 1, multiplying both sides by λ gives

µ = |λ|2 µ = λλ∗ µ = λ,

a contradiction. Hence, hu|v i = 0, i.e., u ⊥ v .

1 / 20
Exercise
Prove the following:
1 If U and V are unitary, then UV is unitary.
2 If U is unitary, then U H is unitary.
3 If A is Hermitian and U is unitary, then U H AU is Hermitian.

2 / 20
5.6 Similarity Transformations

Definition
Let A be a square matrix. A transformation of the type
A S −1 AS
is called a similarity transformation. If
B = S −1 AS,
then A is said to be similar to B, and we will write
A ∼ B.

Note
(Reflexivity) A ∼ A
(Symmetry) A ∼ B =⇒ B∼A
(Transitivity) A ∼ B and B ∼ C =⇒ A∼C

3 / 20
Change of Basis
The matrix representation of a linear transformation

T :V →V

with respect to a basis α undergoes a similarity transformation when the

basis is changed α β:

[T ]α [T ]β = [IdV ]βα [T ]α [IdV ]αβ = S −1 [T ]α S

IdV IdV $
V /V T /V /V
ρβ ρα ρα ρβ
∼

∼
 [IdV ]α  [T ]α  [IdV ]β 
Cn
β
/ Cn / Cn
α
/ n
9C
[T ]β

4 / 20
Change of Basis (cont’d.)
Conversely, the similarity transformation

A S −1 AS

may be interpreted as a result of the change of basis

e α,

where α is the basis consisting of the columns of S:

A = [A]e [A]α = [IdCn ]αe [A]e [IdCn ]eα = S −1 AS

Note
The diagonalization of a matrix is the result of changing the standard
basis to an eigenvector basis.

5 / 20
Invariance of Determinant under Similarity Transformation
From
det Q −1 AQ = det A,
we see that
A∼B =⇒ det A = det B

Invariance of Spectrum under Similarity Transformation

From
pQ −1 AQ (λ) = pA (λ),
we see that
A∼B =⇒ Spec A = Spec B

6 / 20
Theorem
There is an isomorphism of eigenspaces
∼
→ Eλ (Q −1 AQ),
Eλ (A) − x 7→ Q −1 x .

Proof
If x ∈ Eλ (A), then Ax = λx , so
(Q −1 AQ)(Q −1 x ) = Q −1 Ax = λ(Q −1 x ).
Hence,
Q −1 x ∈ Eλ (Q −1 AQ),
and there is a function
Eλ (A) → Eλ (Q −1 AQ), x 7→ Q −1 x , (∗)
which is clearly linear. Similarly, there is a linear transformation
Eλ (A) ← Eλ (Q −1 AQ), Qx ←[ x ,
which is the inverse of (∗).
7 / 20
Corollary
If x is an eigenvector of A, then Q −1 x is an eigenvector of Q −1 AQ
associated with the same eigenvalue.

x ∈ Eλ (A) =⇒ Q −1 x ∈ Eλ (Q −1 A)

8 / 20
Exponentiation and Similarity Transformation
From
−1 AQ
eQ = Q −1 e A Q,
we see that
A∼B =⇒ eA ∼ eB .

9 / 20
Theorem
Tr AB = Tr BA

Proof
n
X n X
X n n X
X n
Tr AB = (AB)ii = Aij B ji = B ji Aij
i=1 i=1 j=1 i=1 j=1
Xn Xn Xn
= B ji Aij = (BA)jj = Tr BA
j=1 i=1 j=1

Corollary
Tr Q −1 AQ = Tr A

Remark
It folllows that the trace is invariant under similarity transformation:

A∼B =⇒ Tr A = Tr B
10 / 20
Schur’s Lemma
For every square matrix A, there exists a unitary matrix U such that

U H AU

is an upper triangular matrix.

Note
Every square matrix is unitarily similar to an upper triangular matrix.

Issai Schur (1875 – 1941) was a Russian mathematician who worked in Germany.

11 / 20
Proof
Let λ1 , . . . , λn be the eigenvalues of A, and v1 an eigenvector of A
associated with λ1 . We may extend v1 to a basis v1 , . . . , vn , and apply
Gram-Schmidt to convert it to an orthonormal basis u1 , . . . , un .
v1 v1 , . . . , vn u1 , . . . , un
Au1 = λ1 u1 , and Au2 , . . . , Aun are linear combinations of u1 , . . . , un , so
   
| | | | " #
 λ1 ∗
A u1 · · · un  = u1 · · · un 
  
0 B
| | | |
for some (n − 1) × (n − 1) matrix B. Now,
 
| |
U := u1 · · · un 
 
| |
is unitary, and
12 / 20
Proof (cont’d.)
" #
λ ∗
U AU = 1
H
.
0 B
It follows that
pU H AU (λ) = (λ1 − λ) det(B − λI).
On the other hand,

pU H AU (λ) = pA (λ) = (λ − λ1 ) · · · (λ − λn ),

so
pB (λ) = det(B − λI) = (λ − λ2 ) · · · (λ − λn ).
Hence, the eigenvalues of B are λ2 , . . . , λn . By the same argument,
" #
λ ∗
V H BV = 2 ,
0 C

where V is a unitary matrix, and C is an (n − 2) × (n − 2) matrix.

13 / 20
Proof (cont’d.)
Then
" # " # " #" #" #
1 0 1 0 1 0 λ1 ∗ 1 0
U H AU =
0 VH 0 V 0 VH 0 B 0 V

" #
λ ∗
= 1
0 V H BV

 
λ1 ∗ ∗
=  0 λ2 ∗  .
 
0 0 C

Note that
" #H " # " # " #−1
1 0 1 0 1 0 1 0
= = = .
0 V 0 VH 0 V −1 0 V
14 / 20
Proof (cont’d.)
Hence, " #
1 0
V̂ :=
0 V
is unitary. Note that
 
λ1 ∗ ∗ " # " #
1 0 H 1 0
 0 λ2 ∗  = U AU
 
0 VH 0 V
0 0 C

H
= V̂ U H AU V̂ = (U V̂ )H A(U V̂ ),
and U V̂ is unitary. We continue this way, until we obtain
λ1 ∗ ∗
 
 ..
W AW =  . . . .
H
∗ ,


0 ··· λn
where W is a unitary matrix.
15 / 20
Corollary
det exp A = exp Tr A

Remark
We know this is true for upper triangular matrices. (Week 12, THR)

Proof
By Schur’s lemma,
A∼T
for some upper triangular matrix T , so

exp A ∼ exp T ,

and therefore

det exp A = det exp T = exp Tr T = exp Tr A.

16 / 20
Note
If A is a real matrix with only real eigenvalues, then the proof of Schur’s
lemma goes through, producing a real upper triangular matrix of the form

Q T AQ,

where Q is an (real) orthogonal matrix.

Theorem
Every real symmetric matrix is orthogonally diagonalizable.

Proof
Let A be a real symmetric matrix. Then it is Hermitian, so its eigenvalues
are real. Hence, by the preceding note, there is a real upper triangular
matrix of the form
Q T AQ,
where Q is (real) orthogonal. But Q T AQ is symmetric, so it must be
diagonal.
17 / 20
Normal Matrices

Definition
An n × n matrix is said to be normal if it has n orthonormal eigenvectors.

Note
Every normal matrix is complete.

Note
A is normal if and only if it can be diagonalized by a unitary matrix U:

U H AU = Λ

We say that normal matrices are unitarily diagonalizable.

18 / 20
Theorem (Spectral Theorem)
If A is normal, then X
A= λ pEλ (A) .
λ∈Spec A

Proof
Let A be an n × n normal matrix, and u1 , . . . , un ∈ Cn its orthonormal
eigenvectors. Then
I = pCn = u1 u1H + · · · + un unH ,
so
A = Au1 u1H + · · · + Aun unH = λ1 u1 u1H + · · · + λn un unH

n
X X
= λi phui i = λ pEλ (A) .
i=1 λ∈Spec A

19 / 20
Example
 
4 0 0
A = 0 4 0
 
0 0 5
has orthonormal eigenvectors e1 , e2 , e3 associated with eigenvalues 4, 4, 5,
respectively. The eigenspaces are

E4 (A) = he1 , e2 i and E5 (A) = he3 i,

and the projections onto them are

pE4 (A) = e1 e1H + e2 e2H and pE5 (A) = e3 e3H .

Hence,

A = 4e1 e1H + 4e2 e2H + 5e3 e3H

= 4(e1 e1H + e2 e2H ) + 5e3 e3H
= 4pE4 (A) + 5pE5 (A) .
20 / 20