Schur triangularization: Let
A
A
A be a
n
×
n
n\times n
n×n matrix in the
complex field. We use
λ
1
,
λ
2
,
⋯
,
λ
n
\lambda_1, \lambda_2, \cdots, \lambda_n
λ1,λ2,⋯,λn to
denote the eigenvalues of
A
A
A. Then there is an unitary
U
U
U such that
U
∗
A
U
=
T
U^{*}AU=T
U∗AU=T, where
U
=
(
x
u
2
⋯
u
n
)
U=\left( \begin{array}{cccc} x & u_2 & \cdots & u_n \end{array} \right)
U=(xu2⋯un),
A
x
=
λ
1
x
Ax=\lambda_1x
Ax=λ1x,
∥
x
∥
=
1
\|x\|=1
∥x∥=1 and
T
T
T is an upper triangular
matrix with diagonal entries
t
i
i
=
λ
i
,
i
=
1
,
2
,
⋯
,
n
t_{ii}=\lambda_i, i=1,2,\cdots,n
tii=λi,i=1,2,⋯,n.(Note:
U
∗
U^{*}
U∗ is the conjugate
transpose of
U
U
U).
We can always contruct an unitary matrix from
x
x
x. we name it as
U
1
=
(
x
u
1
⋯
u
n
)
U_1=\left( \begin{array}{cccc} x & u_1 & \cdots & u_n \end{array} \right)
U1=(xu1⋯un). Then
U
1
∗
A
U
1
=
U
1
∗
(
A
x
A
u
1
⋯
A
u
n
)
=
U
1
∗
(
λ
1
x
A
u
1
⋯
A
u
n
)
=
(
x
∗
u
2
∗
⋮
u
n
∗
)
(
λ
1
x
A
u
1
⋯
A
u
n
)
=
(
λ
1
x
∗
x
x
∗
A
u
1
⋯
x
∗
A
u
n
λ
1
u
2
∗
x
u
2
∗
A
u
2
⋯
u
2
∗
A
u
n
⋮
⋮
⋱
⋮
λ
1
u
n
∗
x
u
n
∗
A
u
2
⋯
u
n
∗
A
u
n
)
=
(
λ
1
⋆
0
A
1
)
\left. \begin{aligned} U_1^{*}AU_1&=U_1^{*}\left( \begin{array}{cccc} Ax & Au_1 & \cdots & Au_n \end{array} \right)\\ &=U_1^{*}\left( \begin{array}{cccc} \lambda_1x & Au_1 & \cdots & Au_n \end{array} \right)\\ &=\left( \begin{array}{c} x^{*} \\ u^{*}_2 \\ \vdots \\ u_n^{*} \end{array} \right)\left( \begin{array}{cccc} \lambda_1x & Au_1 & \cdots & Au_n \end{array} \right)\\ &=\left( \begin{array}{cccc} \lambda_1x^{*}x & x^{*}Au_1 & \cdots & x^{*}Au_n \\ \lambda_1u_2^{*}x & u_2^{*}Au_2 & \cdots & u_2^{*}Au_n \\ \vdots & \vdots & \ddots & \vdots \\ \lambda_1u_n^{*}x & u^{*}_nAu_2 & \cdots & u_n^{*}Au_n \end{array} \right)\\ &=\left( \begin{array}{cc} \lambda_1 & \star \\ 0 & A_1 \end{array} \right) \end{aligned} \right.
U1∗AU1=U1∗(AxAu1⋯Aun)=U1∗(λ1xAu1⋯Aun)=⎝⎜⎜⎜⎛x∗u2∗⋮un∗⎠⎟⎟⎟⎞(λ1xAu1⋯Aun)=⎝⎜⎜⎜⎛λ1x∗xλ1u2∗x⋮λ1un∗xx∗Au1u2∗Au2⋮un∗Au2⋯⋯⋱⋯x∗Aunu2∗Aun⋮un∗Aun⎠⎟⎟⎟⎞=(λ10⋆A1) It is not difficult to realize that the eigenvalues of
A
1
A_1
A1 are
λ
2
,
λ
3
,
⋯
,
λ
n
\lambda_2, \lambda_3, \cdots, \lambda_n
λ2,λ3,⋯,λn.From the definition
of eigenvalues,
λ
1
,
λ
2
,
⋯
,
λ
n
\lambda_1, \lambda_2, \cdots, \lambda_n
λ1,λ2,⋯,λn are the
solutions of the following polynomial.
∣
U
∗
A
U
−
λ
I
∣
=
∣
U
∗
A
U
−
λ
U
∗
U
∣
=
∣
U
∗
∣
∣
A
−
λ
I
∣
∣
U
∣
=
∣
A
−
λ
I
∣
\left. \begin{aligned} |U^{*}AU-\lambda I|&=|U^{*}AU-\lambda U^{*}U|\\ &=|U^{*}||A-\lambda I||U|\\ &=|A-\lambda I| \end{aligned} \right.
∣U∗AU−λI∣=∣U∗AU−λU∗U∣=∣U∗∣∣A−λI∣∣U∣=∣A−λI∣ The last equation tells us that
U
∗
A
U
U^{*}AU
U∗AU and
A
A
A have
same eigenvalues.
∣
U
∗
A
U
−
λ
I
∣
=
∣
(
λ
1
⋆
0
A
1
)
−
λ
I
∣
=
∣
λ
1
−
λ
⋆
0
A
1
−
λ
I
n
−
1
∣
=
∣
λ
1
−
λ
∣
∣
A
1
−
λ
I
n
−
1
∣
\left. \begin{aligned} |U^{*}AU-\lambda I|&= \left| \left( \begin{array}{cc} \lambda_1 & \star \\ 0 & A_1 \end{array} \right)-\lambda I \right|\\ &= \left| \begin{array}{cc} \lambda_1-\lambda & \star \\ 0 & A_1-\lambda I_{n-1} \end{array} \right|\\ &=|\lambda_1-\lambda||A_1-\lambda I_{n-1}| \end{aligned} \right.
∣U∗AU−λI∣=∣∣∣∣(λ10⋆A1)−λI∣∣∣∣=∣∣∣∣λ1−λ0⋆A1−λIn−1∣∣∣∣=∣λ1−λ∣∣A1−λIn−1∣ Finally,
∣
λ
1
−
λ
∣
∣
A
1
−
λ
I
n
−
1
∣
=
∣
A
−
λ
I
∣
|\lambda_1-\lambda||A_1-\lambda I_{n-1}|=|A-\lambda I|
∣λ1−λ∣∣A1−λIn−1∣=∣A−λI∣ yields that
A
1
A_1
A1's eigenvalues are
λ
2
,
λ
3
,
⋯
,
λ
n
\lambda_2, \lambda_3, \cdots, \lambda_n
λ2,λ3,⋯,λn.
Similarly, we can find an unitary matrix
V
1
V_1
V1 such that
V
1
∗
A
1
V
1
=
(
λ
2
⋆
0
A
2
)
V_1^{*}A_1V_1=\left( \begin{array}{cc} \lambda_2 & \star \\ 0 & A_2 \end{array} \right)
V1∗A1V1=(λ20⋆A2) Using the notation
U
2
=
[
1
]
⊕
V
1
U_2=[1]\oplus V_1
U2=[1]⊕V1 to denote
U
2
=
[
1
]
⊕
V
1
=
(
I
0
0
V
1
)
U_2=[1]\oplus V_1=\left( \begin{array}{cc} I & 0 \\ 0 & V_1 \end{array} \right)
U2=[1]⊕V1=(I00V1), we can easily find that
U
2
∗
U
1
∗
A
U
1
U
2
=
(
λ
1
⋆
⋆
0
λ
2
⋆
0
0
A
2
)
U_2^{*}U_1^{*}AU_1U_2=\left( \begin{array}{ccc} \lambda_1 & \star & \star \\ 0 & \lambda_2 & \star \\ 0 & 0 & A_2 \end{array} \right)
U2∗U1∗AU1U2=⎝⎛λ100⋆λ20⋆⋆A2⎠⎞ It is obvious that
U
1
U
2
U_1U_2
U1U2 is another unitary matrix.
Continue this reduction to produce a series of
U
i
U_i
Ui. We get an unitary
matrix
U
=
U
1
U
2
⋯
U
n
U=U_1U_2\cdots U_n
U=U1U2⋯Un and
U
∗
A
U
U^{*}AU
U∗AU is upper triangular.
Lemma 1: Let
A
A
A be partitioned as
A
=
(
A
11
A
12
0
A
22
)
A=\left( \begin{array}{cc} A_{11} & A_{12} \\ 0 & A_{22} \end{array} \right)
A=(A110A12A22), in which
A
11
,
A
22
A_{11}, A_{22}
A11,A22 are square. Then
A
A
A is normal if
and only if
A
11
A_{11}
A11 and
A
22
A_{22}
A22 are normal and
A
12
=
0
A_{12}=0
A12=0.
Firstly, if
A
11
,
A
22
A_{11},A_{22}
A11,A22 are normal and
A
12
=
0
A_{12}=0
A12=0, we have
A
A
∗
=
(
A
11
0
0
A
22
)
(
A
11
∗
0
0
A
22
∗
)
=
(
A
11
A
11
∗
0
0
A
22
A
22
∗
)
=
(
A
11
∗
0
0
A
22
∗
)
(
A
11
0
0
A
22
)
becasue
(
A
11
A
11
∗
=
A
11
∗
A
11
,
A
22
A
22
∗
=
A
22
∗
A
22
)
=
A
∗
A
\left. \begin{aligned} AA^{*}&=\left( \begin{array}{cc} A_{11} & 0 \\ 0 & A_{22} \end{array} \right)\left( \begin{array}{cc} A_{11}^{*} & 0 \\ 0 & A_{22}^{*} \end{array} \right)\\ &=\left( \begin{array}{cc} A_{11}A_{11}^{*} & 0 \\ 0 & A_{22}A_{22}^{*} \end{array} \right)\\ &=\left( \begin{array}{cc} A_{11}^{*} & 0 \\ 0 & A_{22}^{*} \end{array} \right)\left( \begin{array}{cc} A_{11} & 0 \\ 0 & A_{22} \end{array} \right)\text{ becasue } (A_{11}A_{11}^{*}=A_{11}^{*}A_{11}, A_{22}A_{22}^{*}=A_{22}^{*}A_{22})\\ &=A^{*}A \end{aligned} \right.
AA∗=(A1100A22)(A11∗00A22∗)=(A11A11∗00A22A22∗)=(A11∗00A22∗)(A1100A22) becasue (A11A11∗=A11∗A11,A22A22∗=A22∗A22)=A∗A If
A
A
A is normal, we get
A
A
∗
=
(
A
11
A
11
∗
+
A
12
A
12
∗
A
12
A
22
∗
A
22
A
12
∗
A
22
A
22
∗
)
\left. \begin{aligned} AA^{*}&=\left( \begin{array}{cc} A_{11}A_{11}^{*}+A_{12}A_{12}^{*} & A_{12}A_{22}^{*} \\ A_{22}A_{12}^{*} & A_{22}A_{22}^{*} \end{array} \right)\\ \end{aligned} \right.
AA∗=(A11A11∗+A12A12∗A22A12∗A12A22∗A22A22∗) and
A
∗
A
=
(
A
11
∗
A
11
A
11
∗
A
12
A
12
∗
A
11
A
12
∗
A
12
+
A
22
∗
A
22
)
\left. \begin{aligned} A^{*}A=\left( \begin{array}{cc} A_{11}^{*}A_{11} & A_{11}^{*}A_{12} \\ A_{12}^{*}A_{11} & A_{12}^{*}A_{12}+A_{22}^{*}A_{22} \end{array} \right) \end{aligned} \right.
A∗A=(A11∗A11A12∗A11A11∗A12A12∗A12+A22∗A22) Because
A
A
A is normal, it is easy to see that
A
11
∗
A
11
=
A
11
A
11
∗
+
A
12
A
12
∗
A_{11}^{*}A_{11}=A_{11}A_{11}^{*}+A_{12}A_{12}^{*}
A11∗A11=A11A11∗+A12A12∗ and it follows
that
t
r
a
c
e
(
A
11
∗
A
11
)
=
t
r
a
c
e
(
A
11
A
11
∗
+
A
12
A
12
∗
)
=
t
r
a
c
e
(
A
11
A
11
∗
)
+
t
r
a
c
e
(
A
12
A
12
∗
)
=
t
r
a
c
e
(
A
11
∗
A
11
)
+
t
r
a
c
e
(
A
12
A
12
∗
)
\left. \begin{aligned} trace(A_{11}^{*}A_{11})&=trace(A_{11}A_{11}^{*}+A_{12}A_{12}^{*})\\ &=trace(A_{11}A_{11}^{*})+trace(A_{12}A_{12}^{*})\\ &=trace(A_{11}^{*}A_{11})+trace(A_{12}A_{12}^{*})\\ \end{aligned} \right.
trace(A11∗A11)=trace(A11A11∗+A12A12∗)=trace(A11A11∗)+trace(A12A12∗)=trace(A11∗A11)+trace(A12A12∗) Finalli we get
t
r
a
c
e
(
A
12
A
12
∗
)
=
0
trace(A_{12}A_{12}^{*})=0
trace(A12A12∗)=0. It means
that
A
12
=
0
A_{12}=0
A12=0.
Lemma 2: An upper triangular matrix is normal if and only if it is
diagonal.
Using Lemma 1 recursively, we can easily find that Lemma 2 is true.
Theorem: Hermitian matrix is unitarily diagonalizable.
Schur triangularization tells us that Hermitian matrix
A
A
A can be
unitarily transfomed into a upper triangular matrix with eigenvalues in
the diagonal.
U
∗
A
U
=
T
U^{*}AU=T
U∗AU=T Next,
T
T
∗
=
U
∗
A
U
U
∗
A
∗
U
=
U
∗
A
A
∗
U
=
U
∗
A
∗
A
U
=
U
∗
A
∗
U
U
∗
A
U
=
T
∗
T
\left. \begin{aligned} TT^{*}&=U^{*}AUU^{*}A^{*}U\\ &=U^{*}AA^{*}U\\ &=U^{*}A^{*}AU\\ &=U^{*}A^{*}UU^{*}AU\\ &=T^{*}T \end{aligned} \right.
TT∗=U∗AUU∗A∗U=U∗AA∗U=U∗A∗AU=U∗A∗UU∗AU=T∗T yields that
T
T
T is a normal matrix. From Lemma 2, it
implies that
T
T
T is a diagonal matrix. Hence, hermitian matrix is
unitarily diagonalizable.