考虑线性方程组
A
x
=
b
\mathbf{Ax}=\mathbf{b}
Ax=b
其中
A
\mathbf{A}
A非奇异
当 A \mathbf{A} A是大型稀疏矩阵的时候,用迭代法解比较好
前置知识
谱半径:https://blog.csdn.net/qq_39942341/article/details/126238464?spm=1001.2014.3001.5501
矩阵微积分:https://blog.csdn.net/qq_39942341/article/details/126255882?spm=1001.2014.3001.5501
对角占优矩阵
设
A
∈
R
n
×
n
\mathbf{A}\in\mathbb{R}^{n\times n}
A∈Rn×n
严格对角占优矩阵:如果
A
\mathbf{A}
A的元素满足
∣
a
i
i
∣
>
∑
j
=
1
j
≠
i
n
∣
a
i
j
∣
,
i
=
1
,
2
,
⋯
,
n
\left|a_{ii}\right|>\sum_{j=1\atop j\neq i}^{n}\left|a_{ij}\right|,\quad i=1,2,\cdots,n
∣aii∣>j=ij=1∑n∣aij∣,i=1,2,⋯,n
严格对角占优矩阵:如果
A
\mathbf{A}
A的元素满足
∣
a
i
i
∣
≥
∑
j
=
1
j
≠
i
n
∣
a
i
j
∣
,
i
=
1
,
2
,
⋯
,
n
\left|a_{ii}\right|\ge \sum_{j=1\atop j\neq i}^{n}\left|a_{ij}\right|,\quad i=1,2,\cdots,n
∣aii∣≥j=ij=1∑n∣aij∣,i=1,2,⋯,n
且上式至少由一个不等式严格成立
可约于不可约矩阵
如果存在置换矩阵
P
\mathbf{P}
P使得
P
T
A
P
=
(
A
11
A
12
0
A
22
)
\mathbf{P}^T\mathbf{A}\mathbf{P}=\begin{pmatrix} \mathbf{A}_{11}&\mathbf{A}_{12}\\ \mathbf{0}&\mathbf{A}_{22}\\ \end{pmatrix}
PTAP=(A110A12A22)
其中
A
11
\mathbf{A}_{11}
A11为
r
r
r阶方阵,
A
22
\mathbf{A}_{22}
A22为
n
−
r
n-r
n−r阶方阵,则称
A
\mathbf{A}
A为可约矩阵
如果不存在这样的置换矩阵,则称
A
\mathbf{A}
A为不可约矩阵
对角占优定理
如果
A
\mathbf{A}
A为严格对角占优矩阵或不可约弱对角占优矩阵,则
A
\mathbf{A}
A非奇异
证明:
设
A
\mathbf{A}
A为严格对角占优矩阵
假设
∣
A
∣
=
0
\left|\mathbf{A}\right|=0
∣A∣=0,则
A
x
=
0
\mathbf{Ax}=\mathbf{0}
Ax=0有非零解,则
∣
x
k
∣
=
max
1
≤
i
≤
n
∣
x
i
∣
≠
0
\left|x_k\right|=\max\limits_{1\le i\le n}\left|x_i\right|\neq 0
∣xk∣=1≤i≤nmax∣xi∣=0
则
∑
j
=
1
n
a
k
j
x
j
=
0
\sum_{j=1}^{n}a_{kj}x_j=0
j=1∑nakjxj=0
∣
a
k
k
x
k
∣
=
∣
a
k
k
x
k
−
∑
j
=
1
n
a
k
j
x
j
∣
=
∣
∑
j
=
1
j
≠
k
n
a
k
j
x
j
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
∣
x
j
∣
≤
∣
x
k
∣
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
\left|a_{kk}x_k\right|=\left|a_{kk}x_k-\sum_{j=1}^{n}a_{kj}x_j\right|=\left|\sum_{j=1\atop j\neq k}^n a_{kj}x_j\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|\left|x_j\right|\le \left|x_k\right| \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|
∣akkxk∣=∣
∣akkxk−j=1∑nakjxj∣
∣=∣
∣j=kj=1∑nakjxj∣
∣≤j=kj=1∑n∣akj∣∣xj∣≤∣xk∣j=kj=1∑n∣akj∣
即
∣
a
k
k
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
\left|a_{kk}\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|
∣akk∣≤j=kj=1∑n∣akj∣
矛盾,所以
A
\mathbf{A}
A非奇异
设
A
\mathbf{A}
A为不可约弱对角占优矩阵
假设
∣
A
∣
=
0
\left|\mathbf{A}\right|=0
∣A∣=0,则
A
x
=
0
\mathbf{Ax}=\mathbf{0}
Ax=0有非零解
J
=
{
k
∣
∣
x
k
∣
≥
∣
x
i
∣
,
i
=
1
,
2
,
⋯
,
n
;
∃
j
,
∣
x
k
∣
>
∣
x
j
∣
}
J=\left\{k|\left|x_k\right|\ge\left|x_i\right|,i=1,2,\cdots,n;\exists j,\left|x_k\right|>\left|x_j\right|\right\}
J={k∣∣xk∣≥∣xi∣,i=1,2,⋯,n;∃j,∣xk∣>∣xj∣}
如果
J
=
∅
J=\empty
J=∅,则
x
=
l
e
(
l
≠
0
)
\mathbf{x}=l\mathbf{e}\left(l\neq 0\right)
x=le(l=0)
对于
k
=
1
,
2
,
⋯
,
n
k=1,2,\cdots,n
k=1,2,⋯,n
∑
j
=
1
n
a
k
j
x
j
=
0
⇒
∑
j
=
1
n
a
k
j
=
0
\sum_{j=1}^{n}a_{kj}x_j=0\Rightarrow \sum_{j=1}^{n}a_{kj}=0
j=1∑nakjxj=0⇒j=1∑nakj=0
∣
a
k
k
∣
=
∣
a
k
k
−
∑
j
=
1
n
a
k
j
∣
=
∣
∑
j
=
1
j
≠
k
n
a
k
j
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
\left|a_{kk}\right|=\left|a_{kk}-\sum_{j=1}^{n}a_{kj}\right|=\left|\sum_{j=1\atop j\neq k}^n a_{kj}\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|
∣akk∣=∣
∣akk−j=1∑nakj∣
∣=∣
∣j=kj=1∑nakj∣
∣≤j=kj=1∑n∣akj∣
矛盾,所以
J
≠
∅
J\neq \empty
J=∅
∣
a
k
k
x
k
∣
=
∣
a
k
k
x
k
−
∑
j
=
1
n
a
k
j
x
j
∣
=
∣
∑
j
=
1
j
≠
k
n
a
k
j
x
j
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
∣
x
j
∣
⇒
∣
a
k
k
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
∣
x
j
∣
∣
x
k
∣
\left|a_{kk}x_k\right|=\left|a_{kk}x_k-\sum_{j=1}^{n}a_{kj}x_j\right|=\left|\sum_{j=1\atop j\neq k}^n a_{kj}x_j\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|\left|x_j\right|\Rightarrow \left|a_{kk}\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|\frac{\left|x_j\right|}{\left|x_k\right|}
∣akkxk∣=∣
∣akkxk−j=1∑nakjxj∣
∣=∣
∣j=kj=1∑nakjxj∣
∣≤j=kj=1∑n∣akj∣∣xj∣⇒∣akk∣≤j=kj=1∑n∣akj∣∣xk∣∣xj∣
因为
∣
a
k
k
∣
≥
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
\left|a_{kk}\right|\ge \sum\limits_{j=1\atop j\neq k}^{n}\left|a_{kj}\right|
∣akk∣≥j=kj=1∑n∣akj∣
于是
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
≤
∑
j
=
1
j
≠
k
n
∣
a
k
j
∣
∣
x
j
∣
∣
x
k
∣
\sum\limits_{j=1\atop j\neq k}^{n}\left|a_{kj}\right|\le \sum_{j=1\atop j\neq k}^n \left|a_{kj}\right|\frac{\left|x_j\right|}{\left|x_k\right|}
j=kj=1∑n∣akj∣≤j=kj=1∑n∣akj∣∣xk∣∣xj∣
也就是说
∀
k
∈
J
,
j
∉
J
\forall k\in J,j\not\in J
∀k∈J,j∈J,有
a
k
j
=
0
a_{kj}=0
akj=0
取置换矩阵
P
\mathbf{P}
P,使得
P
A
P
T
\mathbf{P}\mathbf{A}\mathbf{P}^T
PAPT为分块上三角矩阵
(
a
k
j
)
\left(a_{kj}\right)
(akj)换到左下角
矛盾,所以
A
\mathbf{A}
A非奇异
迭代法
定义
对于给定的线性方程组
x
=
B
x
+
f
\mathbf{x}=\mathbf{Bx}+f
x=Bx+f,(
A
=
I
−
B
\mathbf{A}=\mathbf{I}-\mathbf{B}
A=I−B)用公式
x
(
k
+
1
)
=
B
x
(
k
)
+
f
,
k
=
0
,
1
,
2
,
⋯
,
\mathbf{x}^{(k+1)}=\mathbf{B}\mathbf{x}^{(k)}+\mathbf{f},\quad k=0,1,2,\cdots,
x(k+1)=Bx(k)+f,k=0,1,2,⋯,
逐步迭代求近似解的方法称为迭代法
如果 lim k → ∞ x ( k ) \lim\limits_{k\to\infty}\mathbf{x}^{(k)} k→∞limx(k)存在(记为 x ∗ \mathbf{x}^* x∗),称此迭代法收敛,显然 x ∗ \mathbf{x}^* x∗就是此方程组的解,否则称为此迭代法发散
ϵ
(
k
+
1
)
=
x
(
k
+
1
)
−
x
∗
\mathbf{\epsilon}^{(k+1)}=\mathbf{x}^{(k+1)}-\mathbf{x}^*
ϵ(k+1)=x(k+1)−x∗
ϵ
(
k
)
=
B
ϵ
(
k
−
1
)
=
⋯
=
B
k
ϵ
(
0
)
\mathbf{\epsilon}^{(k)}=\mathbf{B}\mathbf{\epsilon}^{(k-1)}=\cdots=\mathbf{B}^k\mathbf{\epsilon}^{(0)}
ϵ(k)=Bϵ(k−1)=⋯=Bkϵ(0)
要想
{
x
(
k
)
}
\left\{\mathbf{x}^{(k)}\right\}
{x(k)}收敛,就要
lim
k
→
∞
ϵ
(
k
)
=
0
\lim\limits_{k\to\infty}\mathbf{\epsilon}^{(k)}=\mathbf{0}
k→∞limϵ(k)=0,就需要研究
B
\mathbf{B}
B满足什么条件时
B
k
→
0
(
k
→
∞
)
\mathbf{B}^k\to\mathbf{0}\left(k\to\infty\right)
Bk→0(k→∞)
建立
设
A
∈
R
n
×
n
\mathbf{A}\in\mathbb{R}^{n\times n}
A∈Rn×n非奇异
将
A
\mathbf{A}
A分裂为
A
=
M
−
N
\mathbf{A}=\mathbf{M}-\mathbf{N}
A=M−N
其中
M
\mathbf{M}
M非奇异,且使
M
x
=
d
\mathbf{M}\mathbf{x}=\mathbf{d}
Mx=d容易求解,一般选择为
A
\mathbf{A}
A的某种近似,称
M
\mathbf{M}
M为分裂矩阵
M
x
=
N
x
+
b
\mathbf{M}\mathbf{x}=\mathbf{N}\mathbf{x}+\mathbf{b}
Mx=Nx+b
所以求解
A
x
=
b
\mathbf{Ax}=\mathbf{b}
Ax=b等价于求解
x
=
M
−
1
N
x
+
M
−
1
b
\mathbf{x}=\mathbf{M}^{-1}\mathbf{N}\mathbf{x}+\mathbf{M}^{-1}\mathbf{b}
x=M−1Nx+M−1b
即求解
x
=
B
x
+
f
\mathbf{x}=\mathbf{Bx}+\mathbf{f}
x=Bx+f
从而可以构造一阶定常迭代法
{
x
(
0
)
初始向量
x
(
k
+
1
)
=
B
x
(
k
)
+
f
k
=
0
,
1
,
⋯
\begin{cases} \mathbf{x}^{(0)}&\text{初始向量}\\ \mathbf{x}^{(k+1)}=\mathbf{B}\mathbf{x}^{(k)}+\mathbf{f}&k=0,1,\cdots \end{cases}
{x(0)x(k+1)=Bx(k)+f初始向量k=0,1,⋯
其中
B
=
M
−
1
N
=
M
−
1
(
M
−
A
)
=
I
−
M
−
1
A
,
f
=
M
−
1
b
\mathbf{B}=\mathbf{M}^{-1}\mathbf{N}=\mathbf{M}^{-1}\left(\mathbf{M}-\mathbf{A}\right)=\mathbf{I}-\mathbf{M}^{-1}\mathbf{A},\mathbf{f}=\mathbf{M}^{-1}\mathbf{b}
B=M−1N=M−1(M−A)=I−M−1A,f=M−1b
称 B = I − M − 1 A \mathbf{B}=\mathbf{I}-\mathbf{M}^{-1}\mathbf{A} B=I−M−1A为迭代法的迭代矩阵
收敛性
迭代法收敛的充要条件时矩阵
B
\mathbf{B}
B的谱半径
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1
证明:
充分性:设
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1,
易知
A
=
I
−
B
≻
0
\mathbf{A}=\mathbf{I}-\mathbf{B}\succ 0
A=I−B≻0
所以
A
x
=
f
\mathbf{Ax}=\mathbf{f}
Ax=f有唯一解,记为
x
∗
\mathbf{x}^*
x∗,则
x
∗
=
B
x
∗
+
f
\mathbf{x}^*=\mathbf{B}\mathbf{x}^*+\mathbf{f}
x∗=Bx∗+f
误差向量
ϵ
(
k
)
=
x
(
k
)
−
x
∗
=
B
k
ϵ
(
0
)
,
ϵ
(
0
)
=
x
(
0
)
−
x
∗
\mathbf{\epsilon}^{(k)}=\mathbf{x}^{(k)}-\mathbf{x}^*=\mathbf{B}^k\mathbf{\epsilon}^{(0)},\quad\mathbf{\epsilon}^{(0)}=\mathbf{x}^{(0)}-\mathbf{x}^*
ϵ(k)=x(k)−x∗=Bkϵ(0),ϵ(0)=x(0)−x∗
由
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1,有
lim
k
→
∞
B
k
=
0
\lim\limits_{k\to\infty}\mathbf{B}^{k}=\mathbf{0}
k→∞limBk=0
于是
lim
k
→
∞
ϵ
k
=
0
\lim\limits_{k\to\infty}\mathbf{\epsilon}^k=\mathbf{0}
k→∞limϵk=0,即
lim
k
→
∞
x
(
k
)
=
x
∗
\lim\limits_{k\to\infty}\mathbf{x}^{(k)}=\mathbf{x}^*
k→∞limx(k)=x∗
必要性:设对于任意
x
(
0
)
\mathbf{x}^{(0)}
x(0)有
lim
k
→
∞
x
(
k
)
=
x
∗
\lim\limits_{k\to\infty}\mathbf{x}^{(k)}=\mathbf{x}^*
k→∞limx(k)=x∗
其中
x
(
k
+
1
)
=
B
x
(
k
)
+
f
\mathbf{x}^{(k+1)}=\mathbf{B}\mathbf{x}^{(k)}+\mathbf{f}
x(k+1)=Bx(k)+f,显然
x
∗
=
B
x
∗
+
f
\mathbf{x}^{*}=\mathbf{B}\mathbf{x}^{*}+\mathbf{f}
x∗=Bx∗+f
ϵ
(
k
)
=
x
(
k
)
−
x
∗
=
B
k
ϵ
(
0
)
→
0
(
k
→
∞
)
\mathbf{\epsilon}^{(k)}=\mathbf{x}^{(k)}-\mathbf{x}^*=\mathbf{B}^k\mathbf{\epsilon}^{(0)}\to\mathbf{0}\left(k\to\infty\right)
ϵ(k)=x(k)−x∗=Bkϵ(0)→0(k→∞)
故
lim
k
→
∞
B
k
=
0
\lim\limits_{k\to\infty}\mathbf{B}^k=\mathbf{0}
k→∞limBk=0
所以
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1
充分条件
如果
B
\mathbf{B}
B的某一种算子范数
∥
B
∥
=
q
<
1
\|\mathbf{B}\|=q<1
∥B∥=q<1,则
(1)迭代法收敛
(2)
∥
x
∗
−
x
(
k
)
∥
≤
q
k
∥
x
∗
−
x
(
0
)
∥
\|\mathbf{x}^*-\mathbf{x}^{(k)}\|\le q^k\|\mathbf{x}^*-\mathbf{x}^{(0)}\|
∥x∗−x(k)∥≤qk∥x∗−x(0)∥
(3)
∥
x
∗
−
x
(
k
)
∥
≤
q
1
−
q
∥
x
(
k
)
−
x
(
k
−
1
)
∥
\|\mathbf{x}^*-\mathbf{x}^{(k)}\|\le \frac{q}{1-q}\|\mathbf{x}^{(k)}-\mathbf{x}^{(k-1)}\|
∥x∗−x(k)∥≤1−qq∥x(k)−x(k−1)∥
(4)
∥
x
∗
−
x
(
k
)
∥
≤
q
k
1
−
q
∥
x
(
1
)
−
x
(
0
)
∥
\|\mathbf{x}^*-\mathbf{x}^{(k)}\|\le \frac{q^k}{1-q}\|\mathbf{x}^{(1)}-\mathbf{x}^{(0)}\|
∥x∗−x(k)∥≤1−qqk∥x(1)−x(0)∥
证明:
(1)显然
(2)
x
∗
−
x
(
k
+
1
)
=
B
(
x
∗
−
x
(
k
)
)
\mathbf{x}^*-\mathbf{x}^{(k+1)}=\mathbf{B}\left(\mathbf{x}^*-\mathbf{x}^{(k)}\right)
x∗−x(k+1)=B(x∗−x(k))
x
(
k
+
1
)
−
x
(
k
)
=
B
(
x
(
k
)
−
x
(
k
−
1
)
)
\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}=\mathbf{B}\left(\mathbf{x}^{(k)}-\mathbf{x}^{(k-1)}\right)
x(k+1)−x(k)=B(x(k)−x(k−1))
于是
∥
x
∗
−
x
(
k
+
1
)
∥
≤
q
∥
x
∗
−
x
(
k
)
∥
\|\mathbf{x}^*-\mathbf{x}^{(k+1)}\|\le q\|\mathbf{x}^*-\mathbf{x}^{(k)}\|
∥x∗−x(k+1)∥≤q∥x∗−x(k)∥
∥
x
(
k
+
1
)
−
x
(
k
)
∥
≤
q
∥
x
(
k
)
−
x
(
k
−
1
)
∥
\|\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\|\le q\|\mathbf{x}^{(k)}-\mathbf{x}^{(k-1)}\|
∥x(k+1)−x(k)∥≤q∥x(k)−x(k−1)∥
所以
∥
x
∗
−
x
(
k
)
∥
≤
q
k
∥
x
∗
−
x
(
0
)
∥
\|\mathbf{x}^*-\mathbf{x}^{(k)}\|\le q^k\|\mathbf{x}^*-\mathbf{x}^{(0)}\|
∥x∗−x(k)∥≤qk∥x∗−x(0)∥
(3)
∥
x
(
k
+
1
)
−
x
(
k
)
∥
=
∥
x
∗
−
x
(
k
)
−
(
x
∗
−
x
(
k
+
1
)
)
∥
≥
∥
x
∗
−
x
(
k
)
∥
−
∥
x
∗
−
x
(
k
+
1
)
∥
≥
(
1
−
q
)
∥
x
∗
−
x
(
k
)
∥
\begin{aligned} \|\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\|&=\|\mathbf{x}^{*}-\mathbf{x}^{(k)}-\left(\mathbf{x}^*-\mathbf{x}^{(k+1)}\right)\|\\ &\ge \|\mathbf{x}^{*}-\mathbf{x}^{(k)}\|-\|\mathbf{x}^{*}-\mathbf{x}^{(k+1)}\|\\ &\ge \left(1-q\right) \|\mathbf{x}^{*}-\mathbf{x}^{(k)}\| \end{aligned}
∥x(k+1)−x(k)∥=∥x∗−x(k)−(x∗−x(k+1))∥≥∥x∗−x(k)∥−∥x∗−x(k+1)∥≥(1−q)∥x∗−x(k)∥
即
∥
x
∗
−
x
(
k
)
∥
≤
1
1
−
q
∥
x
(
k
+
1
)
−
x
(
k
)
∥
≤
q
1
−
q
∥
x
(
k
)
−
x
(
k
−
1
)
∥
\|\mathbf{x}^{*}-\mathbf{x}^{(k)}\|\le\frac{1}{1-q} \|\mathbf{x}^{(k+1)}-\mathbf{x}^{(k)}\|\le \frac{q}{1-q}\|\mathbf{x}^{(k)}-\mathbf{x}^{(k-1)}\|
∥x∗−x(k)∥≤1−q1∥x(k+1)−x(k)∥≤1−qq∥x(k)−x(k−1)∥
(4)反复使用(3)
对称下的收敛
一下没找到,再说吧
Jacobi迭代
设
A
=
D
−
L
−
U
\mathbf{A}=\mathbf{D}-\mathbf{L}-\mathbf{U}
A=D−L−U
其中
D
=
(
a
11
a
22
⋱
a
n
n
)
,
L
=
(
0
−
a
21
0
⋮
⋮
⋱
−
a
n
−
1
,
1
−
a
n
−
1
,
2
⋯
0
−
a
n
1
−
a
n
2
⋯
−
a
n
,
n
−
1
0
)
,
U
=
(
0
−
a
12
⋯
−
a
1
,
n
−
1
,
−
a
1
n
0
⋯
−
a
2
,
n
−
1
−
a
2
n
⋱
⋮
⋮
0
−
a
n
−
1
,
n
0
)
\mathbf{D}=\begin{pmatrix} a_{11}&&&\\ &a_{22}&&\\ &&\ddots&\\ &&&a_{nn} \end{pmatrix},\mathbf{L}=\begin{pmatrix} 0&&&&\\ -a_{21}&0&&&\\ \vdots&\vdots&\ddots&&\\ -a_{n-1,1}&-a_{n-1,2}&\cdots&0&\\ -a_{n1}&-a_{n2}&\cdots&-a_{n,n-1}&0\\ \end{pmatrix},\mathbf{U}=\begin{pmatrix} 0&-a_{12}&\cdots&-a_{1,n-1},&-a_{1n}\\ &0&\cdots&-a_{2,n-1}&-a_{2n}\\ &&\ddots&\vdots&\vdots\\ &&&0&-a_{n-1,n}\\ &&&&0\\ \end{pmatrix}
D=⎝
⎛a11a22⋱ann⎠
⎞,L=⎝
⎛0−a21⋮−an−1,1−an10⋮−an−1,2−an2⋱⋯⋯0−an,n−10⎠
⎞,U=⎝
⎛0−a120⋯⋯⋱−a1,n−1,−a2,n−1⋮0−a1n−a2n⋮−an−1,n0⎠
⎞
设
a
i
i
≠
0
(
i
=
1
,
2
,
⋯
,
n
)
a_{ii}\neq 0\left(i=1,2,\cdots,n\right)
aii=0(i=1,2,⋯,n),选取
M
=
D
,
A
=
D
−
N
\mathbf{M}=\mathbf{D},\mathbf{A}=\mathbf{D}-\mathbf{N}
M=D,A=D−N
则
{
x
(
0
)
初始向量
x
(
k
+
1
)
=
B
x
(
k
)
+
f
k
=
0
,
1
,
⋯
\begin{cases} \mathbf{x}^{(0)}&\text{初始向量}\\ \mathbf{x}^{(k+1)}=\mathbf{B}\mathbf{x}^{(k)}+\mathbf{f}&k=0,1,\cdots \end{cases}
{x(0)x(k+1)=Bx(k)+f初始向量k=0,1,⋯
其中
B
=
I
−
D
−
1
A
=
D
−
1
(
L
+
U
)
=
J
,
f
=
D
−
1
b
\mathbf{B}=\mathbf{I}-\mathbf{D}^{-1}\mathbf{A}=\mathbf{D}^{-1}\left(\mathbf{L}+\mathbf{U}\right)=\mathbf{J},\mathbf{f}=\mathbf{D}^{-1}\mathbf{b}
B=I−D−1A=D−1(L+U)=J,f=D−1b
称
J
\mathbf{J}
J为解
A
x
=
b
\mathbf{Ax}=\mathbf{b}
Ax=b的Jacobi迭代法的迭代矩阵
由迭代公式
D
x
(
k
+
1
)
=
(
L
+
U
)
x
(
k
)
+
b
\mathbf{D}\mathbf{x}^{(k+1)}=\left(\mathbf{L}+\mathbf{U}\right)\mathbf{x}^{(k)}+\mathbf{b}
Dx(k+1)=(L+U)x(k)+b
或者
a
i
i
x
i
(
k
+
1
)
=
−
∑
j
=
1
i
−
1
a
i
j
x
j
(
k
)
−
∑
j
=
i
+
1
n
a
i
j
x
j
(
k
)
+
b
i
,
i
=
1
,
2
,
⋯
n
a_{ii}x_i^{(k+1)}=-\sum_{j=1}^{i-1}a_{ij}x_j^{(k)}-\sum_{j=i+1}^{n}a_{ij}x_j^{(k)}+b_i,\quad i=1,2,\cdots n
aiixi(k+1)=−j=1∑i−1aijxj(k)−j=i+1∑naijxj(k)+bi,i=1,2,⋯n
于是,计算公式为
x
i
(
k
+
1
)
=
b
i
−
∑
j
=
1
j
≠
i
n
a
i
j
x
j
(
k
)
a
i
i
x_i^{(k+1)}=\frac{b_i-\sum\limits_{j=1\atop j\neq i}^{n}a_{ij}x_j^{(k)}}{a_{ii}}
xi(k+1)=aiibi−j=ij=1∑naijxj(k)
或者
x
(
k
+
1
)
=
D
−
1
(
L
+
U
)
x
(
k
)
+
D
−
1
b
\mathbf{x}^{(k+1)}=\mathbf{D}^{-1}\left(\mathbf{L}+\mathbf{U}\right)\mathbf{x}^{(k)}+\mathbf{D}^{-1}\mathbf{b}
x(k+1)=D−1(L+U)x(k)+D−1b
收敛性
对角占优矩阵
(1)若
A
\mathbf{A}
A为严格对角占优阵,则Jacobi迭代收敛
(2)若
A
\mathbf{A}
A为不可约弱对角占优矩阵,则Jacobi迭代收敛
证明:
显然
a
i
i
≠
0
a_{ii}\neq 0
aii=0
(1)设
A
\mathbf{A}
A为严格对角占有矩阵
J
=
D
−
1
(
L
+
U
)
\mathbf{J}=\mathbf{D}^{-1}\left(\mathbf{L}+\mathbf{U}\right)
J=D−1(L+U)
∣
λ
I
−
J
∣
=
∣
λ
I
−
D
−
1
(
L
+
U
)
∣
=
∣
D
∣
∣
λ
D
−
(
L
+
U
)
∣
\left|\lambda \mathbf{I}-\mathbf{J}\right|=\left|\lambda \mathbf{I}-\mathbf{D}^{-1}\left(\mathbf{L}+\mathbf{U}\right)\right|=\left|\mathbf{D}\right|\left|\lambda\mathbf{D}-\left(\mathbf{L}+\mathbf{U}\right)\right|
∣λI−J∣=∣
∣λI−D−1(L+U)∣
∣=∣D∣∣λD−(L+U)∣
∣
D
∣
≠
0
\left|\mathbf{D}\right|\neq 0
∣D∣=0
若
∣
λ
∣
≥
1
\left|\lambda\right|\ge 1
∣λ∣≥1
令
C
=
λ
D
−
(
L
+
U
)
=
(
λ
a
11
a
12
⋯
a
1
n
a
21
λ
a
22
⋯
a
2
n
⋮
⋮
⋮
a
n
1
a
n
2
⋯
λ
a
n
n
)
\mathbf{C}=\lambda\mathbf{D}-\left(\mathbf{L}+\mathbf{U}\right)=\begin{pmatrix} \lambda a_{11}&a_{12}&\cdots&a_{1n}\\ a_{21}&\lambda a_{22}&\cdots&a_{2n}\\ \vdots&\vdots&&\vdots\\ a_{n1}&a_{n2}&\cdots&\lambda a_{nn} \end{pmatrix}
C=λD−(L+U)=⎝
⎛λa11a21⋮an1a12λa22⋮an2⋯⋯⋯a1na2n⋮λann⎠
⎞
∣
c
i
i
∣
=
∣
λ
∣
∣
a
i
i
∣
>
∣
λ
∣
∑
j
=
1
j
≠
i
n
∣
a
i
j
∣
≥
∑
j
=
1
j
≠
i
n
∣
a
i
j
∣
=
∑
j
=
1
j
≠
i
n
∣
c
i
j
∣
\begin{aligned} \left|c_{ii}\right|&=\left|\lambda\right|\left|a_{ii}\right|\\ &> \left|\lambda\right|\sum_{j=1\atop j\neq i}^{n}\left|a_{ij}\right|\\ &\ge \sum_{j=1\atop j\neq i}^{n}\left|a_{ij}\right|\\ &=\sum_{j=1\atop j\neq i}^{n}\left|c_{ij}\right|\\ \end{aligned}
∣cii∣=∣λ∣∣aii∣>∣λ∣j=ij=1∑n∣aij∣≥j=ij=1∑n∣aij∣=j=ij=1∑n∣cij∣
所以
C
\mathbf{C}
C也是对角占优矩阵,由对角占优定理,
∣
λ
D
−
(
L
+
U
)
∣
=
0
\left|\lambda\mathbf{D}-\left(\mathbf{L}+\mathbf{U}\right)\right|=0
∣λD−(L+U)∣=0
所以
∣
λ
∣
<
1
\left|\lambda\right|< 1
∣λ∣<1,即
ρ
(
J
)
<
1
\rho\left(\mathbf{J}\right)<1
ρ(J)<1,收敛
(2)与上面的证明类似,也可以证明
C
\mathbf{C}
C是不可约弱对角占优矩阵,然后就收敛了
充要条件
若 A \mathbf{A} A对称,且 a i i > 0 a_{ii}>0 aii>0,则收敛的充要条件为 A , 2 D − A ≻ 0 \mathbf{A},2\mathbf{D}-\mathbf{A}\succ 0 A,2D−A≻0
证明:
因为
a
i
i
>
0
a_{ii}>0
aii>0,所以
D
≻
0
\mathbf{D}\succ 0
D≻0
B
=
I
−
D
−
1
A
=
D
−
1
2
(
I
−
D
−
1
2
A
D
−
1
2
)
D
1
2
⇒
B
∼
I
−
D
−
1
2
A
D
−
1
2
\mathbf{B}=\mathbf{I}-\mathbf{D}^{-1}\mathbf{A}=\mathbf{D}^{-\frac{1}{2}}\left(\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\right)\mathbf{D}^{\frac{1}{2}}\Rightarrow \mathbf{B}\sim \mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}
B=I−D−1A=D−21(I−D−21AD−21)D21⇒B∼I−D−21AD−21
必要性:
设
D
−
1
2
A
D
−
1
2
\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}
D−21AD−21的特征值为
μ
\mu
μ,则
B
\mathbf{B}
B的特征值为
1
−
μ
1-\mu
1−μ
因为
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1,所以
μ
∈
(
0
,
2
)
\mu\in \left(0,2\right)
μ∈(0,2)
对于
∀
x
≠
0
\forall \mathbf{x}\neq \mathbf{0}
∀x=0,
x
T
D
−
1
2
A
D
−
1
2
x
=
(
D
−
1
2
x
)
T
A
(
D
−
1
2
x
)
>
0
⇒
A
≻
0
\mathbf{x}^T\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\mathbf{x}=\left(\mathbf{D}^{-\frac{1}{2}}\mathbf{x}\right)^T\mathbf{A}\left(\mathbf{D}^{-\frac{1}{2}}\mathbf{x}\right)>0\Rightarrow \mathbf{A}\succ 0
xTD−21AD−21x=(D−21x)TA(D−21x)>0⇒A≻0
2
D
−
A
=
D
1
2
(
2
I
−
D
−
1
2
A
D
−
1
2
)
D
1
2
⇒
2
D
−
A
∼
2
I
−
D
−
1
2
A
D
−
1
2
⇒
2
D
−
A
≻
0
2\mathbf{D}-\mathbf{A}=\mathbf{D}^{\frac{1}{2}}\left( 2\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\right)\mathbf{D}^{\frac{1}{2}}\Rightarrow 2\mathbf{D}-\mathbf{A}\sim 2\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\Rightarrow 2\mathbf{D}-\mathbf{A}\succ 0
2D−A=D21(2I−D−21AD−21)D21⇒2D−A∼2I−D−21AD−21⇒2D−A≻0
充分性:
A
≻
0
⇒
D
−
1
2
A
D
−
1
2
≻
0
\mathbf{A}\succ0\Rightarrow \mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\succ 0
A≻0⇒D−21AD−21≻0
所以
I
−
D
−
1
2
A
D
−
1
2
\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}
I−D−21AD−21的特征值小于1
所以
B
\mathbf{B}
B的特征值也小于1
−
B
=
−
D
−
1
2
(
I
−
D
−
1
2
A
D
−
1
2
)
D
1
2
=
D
−
1
2
(
I
−
D
−
1
2
(
2
D
−
A
)
D
−
1
2
)
D
1
2
\begin{aligned} -\mathbf{B} &=- \mathbf{D}^{-\frac{1}{2}}\left(\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\mathbf{A}\mathbf{D}^{-\frac{1}{2}}\right) \mathbf{D}^{\frac{1}{2}}\\ &=\mathbf{D}^{-\frac{1}{2}}\left(\mathbf{I}-\mathbf{D}^{-\frac{1}{2}}\left(2\mathbf{D}-\mathbf{A}\right)\mathbf{D}^{-\frac{1}{2}}\right) \mathbf{D}^{\frac{1}{2}}\\ \end{aligned}
−B=−D−21(I−D−21AD−21)D21=D−21(I−D−21(2D−A)D−21)D21
其特征值小于1(没看懂),所以
ρ
(
B
)
<
1
\rho\left(\mathbf{B}\right)<1
ρ(B)<1,收敛
Gauss-Seidel迭代法
取
M
=
D
−
L
\mathbf{M}=\mathbf{D}-\mathbf{L}
M=D−L(下三角矩阵)
A
=
M
−
N
\mathbf{A}=\mathbf{M}-\mathbf{N}
A=M−N于是
{
x
(
0
)
初始向量
x
(
k
+
1
)
=
B
x
(
k
)
+
f
k
=
0
,
1
,
⋯
\begin{cases} \mathbf{x}^{(0)}&\text{初始向量}\\ \mathbf{x}^{(k+1)}=\mathbf{B}\mathbf{x}^{(k)}+\mathbf{f}&k=0,1,\cdots \end{cases}
{x(0)x(k+1)=Bx(k)+f初始向量k=0,1,⋯
其中
B
=
I
−
(
D
−
L
)
−
1
A
=
(
D
−
L
)
−
1
U
=
G
,
f
=
(
D
−
L
)
−
1
b
\mathbf{B}=\mathbf{I}-\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{A}=\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{U}=\mathbf{G},\mathbf{f}=\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{b}
B=I−(D−L)−1A=(D−L)−1U=G,f=(D−L)−1b
称
G
\mathbf{G}
G为解
A
x
=
b
\mathbf{Ax}=\mathbf{b}
Ax=b的Gauss-Seidel迭代法的迭代矩阵
(
D
−
L
)
x
(
k
+
1
)
=
U
x
(
k
)
+
b
\left(\mathbf{D}-\mathbf{L}\right)\mathbf{x}^{(k+1)}=\mathbf{U}\mathbf{x}^{(k)}+\mathbf{b}
(D−L)x(k+1)=Ux(k)+b
或者
D
x
(
k
+
1
)
=
L
x
(
k
+
1
)
+
U
x
(
k
)
+
b
\mathbf{D}\mathbf{x}^{(k+1)}=\mathbf{L}\mathbf{x}^{(k+1)}+\mathbf{U}\mathbf{x}^{(k)}+\mathbf{b}
Dx(k+1)=Lx(k+1)+Ux(k)+b
即
a
i
i
x
i
(
k
+
1
)
=
b
i
−
∑
j
=
1
i
−
1
a
i
j
x
j
(
k
+
1
)
−
∑
j
=
i
+
1
n
a
i
j
x
j
(
k
)
,
i
=
1
,
2
,
⋯
,
n
a_{ii}x_i^{(k+1)}=b_i-\sum_{j=1}^{i-1}a_{ij}x_j^{(k+1)}-\sum_{j=i+1}^{n}a_{ij}x_j^{(k)},\quad i=1,2,\cdots,n
aiixi(k+1)=bi−j=1∑i−1aijxj(k+1)−j=i+1∑naijxj(k),i=1,2,⋯,n
所以计算公式为
x
i
(
k
+
1
)
=
b
i
−
∑
j
=
1
i
−
1
a
i
j
x
j
(
k
+
1
)
−
∑
j
=
i
+
1
n
a
i
j
x
j
(
k
)
a
i
i
x_i^{(k+1)}=\frac{b_i-\sum_{j=1}^{i-1}a_{ij}x_j^{(k+1)}-\sum_{j=i+1}^{n}a_{ij}x_j^{(k)}}{a_{ii}}
xi(k+1)=aiibi−∑j=1i−1aijxj(k+1)−∑j=i+1naijxj(k)
收敛性
对角占优矩阵
(1)若
A
\mathbf{A}
A为严格对角占优阵,则Gauss-Seidel迭代法收敛
(2)若
A
\mathbf{A}
A为不可约弱对角占优矩阵,则Gauss-Seidel迭代法收敛
证明:
显然
a
i
i
≠
0
a_{ii}\neq 0
aii=0
(1)设
A
\mathbf{A}
A为严格对角占有矩阵
G
=
(
D
−
L
)
−
1
U
\mathbf{G}=\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{U}
G=(D−L)−1U
∣
λ
I
−
G
∣
=
∣
λ
I
−
(
D
−
L
)
−
1
U
∣
=
∣
(
D
−
L
)
−
1
∣
∣
λ
(
D
−
L
)
−
U
∣
\left|\lambda \mathbf{I}-\mathbf{G}\right|=\left|\lambda \mathbf{I}-\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{U}\right|=\left|\left(\mathbf{D}-\mathbf{L}\right)^{-1}\right|\left|\lambda\left(\mathbf{D}-\mathbf{L}\right)-\mathbf{U}\right|
∣λI−G∣=∣
∣λI−(D−L)−1U∣
∣=∣
∣(D−L)−1∣
∣∣λ(D−L)−U∣
∣
(
D
−
L
)
−
1
∣
≠
0
\left|\left(\mathbf{D}-\mathbf{L}\right)^{-1}\right|\neq 0
∣
∣(D−L)−1∣
∣=0
若
∣
λ
∣
≥
1
\left|\lambda\right|\ge 1
∣λ∣≥1
令
C
=
λ
(
D
−
L
)
−
U
=
(
λ
a
11
a
12
⋯
a
1
n
λ
a
21
λ
a
22
⋯
a
2
n
⋮
⋮
⋮
λ
a
n
1
λ
a
n
2
⋯
λ
a
n
n
)
\mathbf{C}=\lambda\left(\mathbf{D}-\mathbf{L}\right)-\mathbf{U}=\begin{pmatrix} \lambda a_{11}&a_{12}&\cdots&a_{1n}\\ \lambda a_{21}& \lambda a_{22}&\cdots&a_{2n}\\ \vdots&\vdots&&\vdots\\ \lambda a_{n1}&\lambda a_{n2}&\cdots&\lambda a_{nn} \end{pmatrix}
C=λ(D−L)−U=⎝
⎛λa11λa21⋮λan1a12λa22⋮λan2⋯⋯⋯a1na2n⋮λann⎠
⎞
∣
c
i
i
∣
=
∣
λ
∣
∣
a
i
i
∣
>
∣
λ
∣
(
∑
j
=
1
i
−
1
∣
a
i
i
∣
+
∑
j
=
i
+
1
n
∣
a
i
i
∣
)
=
∣
λ
∣
∑
j
=
1
i
−
1
∣
a
i
i
∣
+
∣
λ
∣
∑
j
=
i
+
1
n
∣
a
i
i
∣
≥
∑
j
=
1
i
−
1
∣
λ
a
i
i
∣
+
∑
j
=
i
+
1
n
∣
a
i
i
∣
=
∑
j
=
1
j
≠
i
n
∣
c
i
j
∣
\begin{aligned} \left|c_{ii}\right|&=\left|\lambda\right|\left|a_{ii}\right|\\ &> \left|\lambda\right|\left(\sum_{j=1}^{i-1}\left|a_{ii}\right|+\sum_{j=i+1}^{n}\left|a_{ii}\right|\right)\\ &= \left|\lambda\right|\sum_{j=1}^{i-1}\left|a_{ii}\right|+\left|\lambda\right|\sum_{j=i+1}^{n}\left|a_{ii}\right|\\ &\ge \sum_{j=1}^{i-1}\left|\lambda a_{ii}\right|+\sum_{j=i+1}^{n}\left|a_{ii}\right|\\ &=\sum_{j=1\atop j\neq i}^{n}\left|c_{ij}\right|\\ \end{aligned}
∣cii∣=∣λ∣∣aii∣>∣λ∣(j=1∑i−1∣aii∣+j=i+1∑n∣aii∣)=∣λ∣j=1∑i−1∣aii∣+∣λ∣j=i+1∑n∣aii∣≥j=1∑i−1∣λaii∣+j=i+1∑n∣aii∣=j=ij=1∑n∣cij∣
所以
C
\mathbf{C}
C也是对角占优矩阵,由对角占优定理,
∣
λ
(
D
−
L
)
−
U
∣
=
0
\left|\lambda\left(\mathbf{D}-\mathbf{L}\right)-\mathbf{U}\right|=0
∣λ(D−L)−U∣=0
所以
∣
λ
∣
<
1
\left|\lambda\right|< 1
∣λ∣<1,即
ρ
(
G
)
<
1
\rho\left(\mathbf{G}\right)<1
ρ(G)<1,收敛
(2)与上面的证明类似,也可以证明
C
\mathbf{C}
C是不可约弱对角占优矩阵,然后就收敛了
对称正定
A
\mathbf{A}
A对称,若
A
≻
0
\mathbf{A}\succ 0
A≻0,则收敛
证明:
A
=
D
−
L
−
L
T
\mathbf{A}=\mathbf{D}-\mathbf{L}-\mathbf{L}^T
A=D−L−LT
G
=
(
D
−
L
)
−
1
L
T
\mathbf{G}=\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{L}^T
G=(D−L)−1LT
(
D
−
L
)
−
1
L
T
x
=
λ
x
⇒
L
T
x
=
λ
(
D
−
L
)
x
⇒
x
T
L
T
x
=
λ
x
T
(
D
−
L
)
x
\left(\mathbf{D}-\mathbf{L}\right)^{-1}\mathbf{L}^T \mathbf{x}=\lambda \mathbf{x}\Rightarrow \mathbf{L}^T\mathbf{x}=\lambda \left(\mathbf{D}-\mathbf{L}\right)\mathbf{x}\Rightarrow \mathbf{x}^T\mathbf{L}^T\mathbf{x}=\lambda \mathbf{x}^T\left(\mathbf{D}-\mathbf{L}\right)\mathbf{x}
(D−L)−1LTx=λx⇒LTx=λ(D−L)x⇒xTLTx=λxT(D−L)x
A
≻
0
⇒
p
=
x
T
D
x
>
0
\mathbf{A}\succ 0\Rightarrow p=\mathbf{x}^T\mathbf{D}\mathbf{x}>0
A≻0⇒p=xTDx>0
令
a
=
x
T
L
T
x
a=\mathbf{x}^T\mathbf{L}^T\mathbf{x}
a=xTLTx
x
T
A
x
=
x
T
(
D
−
L
−
L
T
)
x
=
p
−
2
a
>
0
\mathbf{x}^T\mathbf{A}\mathbf{x}=\mathbf{x}^T\left(\mathbf{D}-\mathbf{L}-\mathbf{L}^T\right)\mathbf{x}=p-2a>0
xTAx=xT(D−L−LT)x=p−2a>0
(
p
−
a
)
2
=
p
2
−
2
a
p
+
a
2
=
p
(
p
−
2
a
)
+
a
2
\left(p-a\right)^2=p^2-2ap+a^2=p\left(p-2a\right)+a^2
(p−a)2=p2−2ap+a2=p(p−2a)+a2
λ
=
x
T
L
T
x
x
T
(
D
−
L
)
x
=
a
p
−
a
\lambda =\frac{\mathbf{x}^T\mathbf{L}^T\mathbf{x}}{\mathbf{x}^T\left(\mathbf{D}-\mathbf{L}\right)\mathbf{x}}=\frac{a}{p-a}
λ=xT(D−L)xxTLTx=p−aa
λ
2
=
a
2
p
2
−
2
p
a
+
a
2
=
a
2
p
(
p
−
2
a
)
+
a
2
<
1
\lambda^2=\frac{a^2}{p^2-2pa+a^2}=\frac{a^2}{p\left(p-2a\right)+a^2}<1
λ2=p2−2pa+a2a2=p(p−2a)+a2a2<1
所以收敛
SOR迭代
M
=
1
ω
(
D
−
ω
L
)
\mathbf{M}=\frac{1}{\omega}\left(\mathbf{D}-\omega\mathbf{L}\right)
M=ω1(D−ωL)
其中
ω
>
0
\omega>0
ω>0为可选的松弛因子
L
ω
=
I
−
ω
(
D
−
ω
L
)
−
1
A
=
(
D
−
ω
L
)
−
1
(
(
1
−
ω
)
D
+
ω
U
)
\mathbf{L}_{\omega}=\mathbf{I}-\omega\left(\mathbf{D}-\omega\mathbf{L}\right)^{-1}\mathbf{A}=\left(\mathbf{D}-\omega\mathbf{L}\right)^{-1}\left(\left(1-\omega\right)\mathbf{D}+\omega\mathbf{U}\right)
Lω=I−ω(D−ωL)−1A=(D−ωL)−1((1−ω)D+ωU)
迭代为
x
(
k
+
1
)
=
L
ω
x
(
k
)
+
f
\mathbf{x}^{(k+1)}=\mathbf{L}_{\omega}\mathbf{x}^{(k)}+\mathbf{f}
x(k+1)=Lωx(k)+f
其中
L
ω
=
(
D
−
ω
L
)
−
1
(
(
1
−
ω
)
D
+
ω
U
)
,
f
=
ω
(
D
−
ω
L
)
−
1
b
\mathbf{L}_{\omega}=\left(\mathbf{D}-\omega\mathbf{L}\right)^{-1}\left(\left(1-\omega\right)\mathbf{D}+\omega\mathbf{U}\right),\mathbf{f}=\omega\left(\mathbf{D}-\omega\mathbf{L}\right)^{-1}\mathbf{b}
Lω=(D−ωL)−1((1−ω)D+ωU),f=ω(D−ωL)−1b
收敛条件
必要条件
若SOR迭代收敛,则 0 < ω < 2 0<\omega<2 0<ω<2
证明:
ρ
(
L
ω
)
<
1
\rho\left(L_{\omega}\right)<1
ρ(Lω)<1
∣
∣
L
ω
∣
∣
=
∣
λ
1
⋯
λ
n
∣
≤
(
ρ
(
L
ω
)
)
n
\left|\left|\mathbf{L}_{\omega}\right|\right|=\left|\lambda_1\cdots\lambda_n\right|\le\left(\rho\left(\mathbf{L}_{\omega}\right)\right)^n
∣∣Lω∣∣=∣λ1⋯λn∣≤(ρ(Lω))n
所以
∣
∣
L
ω
∣
∣
1
n
≤
ρ
(
L
ω
)
<
1
\left|\left|\mathbf{L}_{\omega}\right|\right|^{\frac{1}{n}}\le \rho\left(L_{\omega}\right)<1
∣∣Lω∣∣n1≤ρ(Lω)<1
∣
∣
L
ω
∣
∣
=
∣
D
−
ω
L
∣
∣
(
1
−
ω
)
D
+
ω
U
∣
=
∏
i
=
1
n
1
a
i
i
∏
i
=
1
n
(
1
−
ω
)
a
i
i
=
(
1
−
ω
)
n
\left|\left|\mathbf{L}_{\omega}\right|\right|=\left|\mathbf{D}-\omega\mathbf{L}\right|\left|\left(1-\omega\right)\mathbf{D}+\omega\mathbf{U}\right|=\prod_{i=1}^n\frac{1}{a_{ii}}\prod_{i=1}^n\left(1-\omega\right)a_{ii}=\left(1-\omega\right)^n
∣∣Lω∣∣=∣D−ωL∣∣(1−ω)D+ωU∣=i=1∏naii1i=1∏n(1−ω)aii=(1−ω)n
从而
∣
∣
L
ω
∣
∣
1
n
=
∣
1
−
ω
∣
≤
ρ
(
L
ω
)
<
1
⇒
0
<
ω
<
2
\left|\left|\mathbf{L}_{\omega}\right|\right|^{\frac{1}{n}}=\left|1-\omega\right|\le \rho\left(L_{\omega}\right)<1\Rightarrow 0<\omega <2
∣∣Lω∣∣n1=∣1−ω∣≤ρ(Lω)<1⇒0<ω<2
充分条件1
如果 A \mathbf{A} A对称正定, 0 < ω < 2 0<\omega<2 0<ω<2,则收敛
充分条件2
如果 A \mathbf{A} A为严格对角占优矩阵或弱对角占优不可约矩阵, 0 < ω ≤ 1 0<\omega\le 1 0<ω≤1,则收敛
最佳松弛因子
min
0
<
ω
<
2
ρ
(
L
ω
)
=
ρ
(
L
ω
o
p
t
)
\min\limits_{0<\omega<2}\rho\left(\mathbf{L}_\omega\right)=\rho\left(\mathbf{L}_{\omega_{opt}}\right)
0<ω<2minρ(Lω)=ρ(Lωopt)
ω
o
p
t
=
2
1
+
1
−
ρ
(
J
)
2
\omega_{opt}=\frac{2}{1+\sqrt{1-\rho\left(\mathbf{J}\right)^2}}
ωopt=1+1−ρ(J)22
其中
J
=
D
−
1
(
L
+
U
)
\mathbf{J}=\mathbf{D}^{-1}\left(\mathbf{L}+\mathbf{U}\right)
J=D−1(L+U)
SSOR迭代法
M = 1 ω ( 2 − ω ) ( D − ω L ) D − 1 ( D − ω U ) \mathbf{M}=\frac{1}{\omega\left(2-\omega\right)}\left(\mathbf{D}-\omega\mathbf{L}\right)\mathbf{D}^{-1}\left(\mathbf{D}-\omega \mathbf{U}\right) M=ω(2−ω)1(D−ωL)D−1(D−ωU)