瑞利商
\qquad
首先我们给出瑞利商(瑞利商是一个标量)的定义:
R
(
A
,
x
)
=
x
T
A
x
x
T
x
R(A,x)=\frac{x^TAx}{x^Tx}
R(A,x)=xTxxTAx
\qquad
其中
A
A
A为
n
×
n
n\times n
n×n的对称矩阵,
x
x
x为维度为
n
n
n的向量,我们记
A
A
A的从小到大排序的特征值和对应的特征向量为
λ
1
,
λ
2
,
λ
3
.
.
.
λ
n
;
v
1
,
v
2
,
v
3
.
.
.
v
n
\lambda_1,\lambda_2,\lambda_3...\lambda_n;v_1,v_2,v_3...v_n
λ1,λ2,λ3...λn;v1,v2,v3...vn,满足:
λ
m
i
n
=
λ
1
≤
λ
2
≤
.
.
.
≤
λ
n
=
λ
m
a
x
\lambda_{min}=\lambda_1\leq\lambda_2\leq...\leq\lambda_n=\lambda_{max}
λmin=λ1≤λ2≤...≤λn=λmax
\qquad
瑞利商有性质:
m
a
x
x
R
(
A
,
x
)
=
λ
n
\mathop{max}\limits_{x}R(A,x)=\lambda_n
xmaxR(A,x)=λn
m
i
n
x
R
(
A
,
x
)
=
λ
1
\mathop{min}\limits_{x}R(A,x)=\lambda_1
xminR(A,x)=λ1
证明:
\qquad
由于A为对称矩阵,可以相似对角化为"正交阵-特征对角阵-正交阵"的性形式:
A
=
U
Λ
U
T
A=U\Lambda U^T
A=UΛUT
\qquad
其中
Λ
=
d
i
a
g
(
λ
1
,
λ
2
,
λ
3
.
.
.
λ
n
)
\Lambda=diag(\lambda_1,\lambda_2,\lambda_3...\lambda_n)
Λ=diag(λ1,λ2,λ3...λn)特征对角阵,
U
=
(
v
1
,
v
2
,
c
v
3
.
.
.
v
n
)
U=(v_1,v_2,cv_3...v_n)
U=(v1,v2,cv3...vn)为特征向量阵,且为正交阵满足
U
U
T
=
I
UU^T=I
UUT=I。由此瑞利商可以转换为:
R
(
A
,
x
)
=
x
T
A
x
x
T
x
R(A,x)=\frac{x^TAx}{x^Tx}
R(A,x)=xTxxTAx
=
x
T
U
Λ
U
T
x
x
T
U
U
T
x
=\frac{x^TU\Lambda U^Tx}{x^TUU^Tx}
=xTUUTxxTUΛUTx
\qquad
我们令
y
=
U
T
x
y=U^Tx
y=UTx,则瑞利商变为:
R
(
A
,
x
)
=
y
T
Λ
y
y
T
y
R(A,x)=\frac{y^T\Lambda y}{y^Ty}
R(A,x)=yTyyTΛy
=
∑
i
n
λ
i
y
i
2
∑
i
n
y
i
2
=\frac{\mathop{\sum}\limits_{i}^n \lambda_iy_i^2}{\mathop{\sum}\limits_{i}^n y_i^2}
=i∑nyi2i∑nλiyi2
\qquad
由特征值的大小关系显然有:
λ
1
∑
i
n
y
i
2
≤
∑
i
n
λ
i
y
i
2
≤
λ
n
∑
i
n
y
I
2
\lambda_1\mathop{\sum}\limits_{i}^n y_i^2 \leq \mathop{\sum}\limits_{i}^n\lambda_i y_i^2 \leq \lambda_n \mathop{\sum}\limits_{i}^n y_I^2
λ1i∑nyi2≤i∑nλiyi2≤λni∑nyI2
\qquad
于是有:
λ
1
∑
i
n
y
i
2
∑
i
n
y
i
2
≤
∑
i
n
λ
i
y
i
2
∑
i
n
y
i
2
≤
λ
n
∑
i
n
y
I
2
∑
i
n
y
i
2
\frac{\lambda_1\mathop{\sum}\limits_{i}^n y_i^2}{\mathop{\sum}\limits_{i}^n y_i^2} \leq \frac{\mathop{\sum}\limits_{i}^n\lambda_i y_i^2}{\mathop{\sum}\limits_{i}^n y_i^2} \leq \frac{\lambda_n \mathop{\sum}\limits_{i}^n y_I^2}{\mathop{\sum}\limits_{i}^n y_i^2}
i∑nyi2λ1i∑nyi2≤i∑nyi2i∑nλiyi2≤i∑nyi2λni∑nyI2
λ
1
≤
R
(
A
,
x
)
≤
λ
n
\lambda_1 \leq R(A,x) \leq \lambda_n
λ1≤R(A,x)≤λn
\qquad
由此证得瑞利商的上界和下界,下面说明这也是瑞利商的上下确界。
\qquad
不论对向量
x
x
x进行何种放缩,瑞利商不变,对任意非零实数c,对x进行放缩有:
R
(
A
,
c
x
)
=
c
x
T
A
c
x
c
x
T
c
x
=
c
2
x
T
A
x
c
2
x
T
x
=
R
(
A
,
x
)
R(A,cx)=\frac{cx^TA cx}{cx^T cx}=\frac{c^2 x^T Ax}{c^2x^Tx}=R(A,x)
R(A,cx)=cxTcxcxTAcx=c2xTxc2xTAx=R(A,x)
\qquad
因此不妨令
x
T
x
=
1
x^Tx=1
xTx=1,此举不会影响到
x
x
x的任意性
\qquad
由于
v
1
,
v
2
,
v
3
.
.
.
v
n
v_1,v_2,v_3...v_n
v1,v2,v3...vn是n维线性空的一组正交基,利用这组正交基对向量
x
x
x进行线性表出,设系数为
a
1
,
a
2
,
a
3
.
.
.
a
n
a_1,a_2,a_3...a_n
a1,a2,a3...an,满足:
x
=
∑
i
n
a
i
v
i
x=\mathop{\sum}\limits_{i}^n a_iv_i
x=i∑naivi
\qquad
由于
x
T
x
=
1
x^Tx=1
xTx=1,则有:
x
T
x
=
∑
i
n
a
i
2
=
1
x^Tx=\mathop{\sum}\limits_{i}^na_i^2=1
xTx=i∑nai2=1
\qquad
此时考虑到
U
=
(
v
1
,
v
2
,
v
3
.
.
.
v
n
)
U=(v_1,v_2,v_3...v_n)
U=(v1,v2,v3...vn),且有
v
1
,
v
2
,
v
3
.
.
.
v
n
v_1,v_2,v_3...v_n
v1,v2,v3...vn相互正交:
R
(
A
,
x
)
=
(
∑
i
n
a
i
v
i
)
T
U
Λ
U
T
(
∑
i
n
a
i
v
i
)
x
T
x
R(A,x)=\frac{(\mathop{\sum}\limits_{i}^n a_iv_i)^T U \Lambda U^T (\mathop{\sum}\limits_{i}^n a_iv_i)}{x^Tx}
R(A,x)=xTx(i∑naivi)TUΛUT(i∑naivi)
=
(
∑
i
n
a
i
v
i
)
T
U
Λ
U
T
(
∑
i
n
a
i
v
i
)
=(\mathop{\sum}\limits_{i}^n a_iv_i)^T U \Lambda U^T (\mathop{\sum}\limits_{i}^n a_iv_i)
=(i∑naivi)TUΛUT(i∑naivi)
=
∑
i
n
a
i
2
λ
i
=\mathop{\sum}\limits_{i}^n a_i^2\lambda_i
=i∑nai2λi
\qquad
显然也有:
λ
1
≤
R
(
A
,
x
)
≤
λ
n
\lambda_1 \leq R(A,x) \leq \lambda_n
λ1≤R(A,x)≤λn
\qquad
并且由于
∑
i
n
a
i
2
=
1
\mathop{\sum}\limits_{i}^na_i^2=1
i∑nai2=1,可以看出:
\qquad
\qquad
当
a
1
2
=
1
,
a
2
=
a
3
=
.
.
.
=
a
n
=
0
a_1^2=1,a_2=a_3=...=a_n=0
a12=1,a2=a3=...=an=0时,此时
x
=
v
1
x=v_1
x=v1,瑞利商取得最小值
λ
1
\lambda_1
λ1
\qquad
\qquad
当
a
n
2
=
1
,
a
1
=
a
2
=
.
.
.
=
a
n
−
1
=
0
a_n^2=1,a_1=a_2=...=a_{n-1}=0
an2=1,a1=a2=...=an−1=0时,此时
x
=
v
n
x=v_n
x=vn,瑞利商取得最大值
λ
n
\lambda_n
λn
\qquad
瑞利商性质证毕。