Date: 2020/04/04
Editor:萧潇子(Jesse)
Contact: 1223167600@qq.com
1.仿射集合和凸集
在介绍凸集概念之前,我们先介绍一下基本的集合概念
1.1直线与线段
假设空间
R
n
R^n
Rn 中存在两个点
x
1
≠
x
2
x_1 \neq x_2
x1=x2,考虑如下方程式:
y
=
θ
x
1
+
(
1
−
θ
)
x
2
=
x
2
+
θ
(
x
1
−
x
2
)
\begin{aligned} y&= \theta x_1 + (1-\theta) x_2 \\&= x_2 + \theta (x_1-x_2) \end{aligned}
y=θx1+(1−θ)x2=x2+θ(x1−x2)
- θ ∈ R \ \theta \in R θ∈R,代表一条直线
- θ ∈ [ 0 , 1 ] \ \theta \in [0,1] θ∈[0,1],代表线段 (见下图深色阴影所示)
1.2仿射集合
定义:
一个集合是仿射集,若
∀
x
1
,
x
2
∈
C
\forall \ x_1,\ x_2\ \in\ C
∀ x1, x2 ∈ C 则连接
x
1
x_1
x1与
x
2
x_2
x2的直线也在集合内
把上述定义进行泛化:
设:
x
1
x
2
,
⋯
x
k
∈
C
x_1x_2,\cdots x_k \in C
x1x2,⋯xk∈C,
θ
1
,
θ
2
,
⋯
,
θ
k
∈
R
\theta_1,\theta_2,\cdots,\theta_k\in R
θ1,θ2,⋯,θk∈R 且
θ
1
+
θ
2
+
⋯
+
θ
k
=
1
\theta_1 + \theta_2 + \cdots + \theta_k=1
θ1+θ2+⋯+θk=1
泛化(仿射组合):
θ
1
x
1
+
θ
2
x
2
+
⋯
+
θ
k
x
k
∈
C
\theta_1x_1 + \theta_2x_2 + \cdots + \theta_kx_k \in C
θ1x1+θ2x2+⋯+θkxk∈C
证明:
假设存在三点
x
1
,
x
2
,
x
3
x_1,x_2,x_3
x1,x2,x3 ,
θ
1
+
θ
2
+
θ
3
=
1
\theta_1+\theta_2+\theta_3 = 1
θ1+θ2+θ3=1 首先考虑使用
x
1
,
x
2
x_1,x_2
x1,x2 两点构建在
C
C
C内的一点:
θ
1
θ
1
+
θ
2
x
1
+
θ
2
θ
1
+
θ
2
x
2
∈
C
\frac{\theta_1}{\theta_1+\theta_2}x_1 + \frac{\theta_2}{\theta_1+\theta_2}x_2 \in C
θ1+θ2θ1x1+θ1+θ2θ2x2∈C
加上
x
3
x_3
x3 构建:
(
θ
1
+
θ
2
)
(
θ
1
θ
1
+
θ
2
x
1
+
θ
2
θ
1
+
θ
2
x
2
)
+
(
1
−
θ
1
−
θ
2
)
x
3
∈
C
(\theta_1+\theta_2)(\frac{\theta_1}{\theta_1+\theta_2}x_1 + \frac{\theta_2}{\theta_1+\theta_2}x_2) + (1-\theta_1-\theta_2)x_3 \in C
(θ1+θ2)(θ1+θ2θ1x1+θ1+θ2θ2x2)+(1−θ1−θ2)x3∈C
展开上式得;
θ
1
x
1
+
θ
2
x
2
+
θ
3
x
3
∈
C
\theta_1x_1+\theta_2x_2+\theta_3x_3 \in C
θ1x1+θ2x2+θ3x3∈C
因此泛化正确.
性质:
存在特殊仿射集合
V
V
V 满足更一般的情况:
α
x
1
+
β
x
2
∈
V
α
β
∈
R
\alpha x_1 +\beta x_2 \in V \quad \alpha\ \beta \ \in R
αx1+βx2∈Vα β ∈R
如下图中绿色直线所示:
构造 V = C − x 0 = { x − x 0 ∣ x ∈ C } V = C -x_0=\{x-x_0|x\in C\} V=C−x0={x−x0∣x∈C} ∀ x 0 ∈ C \forall \ x_0 \in \ C ∀ x0∈ C
构造新的仿射集合 V V V是与 C C C相关的子空间( C C C平移 x 0 x_0 x0),子空间 V V V 经过原点
证明:
∀ v 1 , v 2 ∈ V \forall \ v_1,v_2 \in V ∀ v1,v2∈V ∀ α , β ∈ R ⇒ α v 1 + β v 2 ∈ V \forall \ \alpha, \ \beta \in R \ \Rightarrow \alpha v_1+\beta v_2 \in V ∀ α, β∈R ⇒αv1+βv2∈V
即: α v 1 + β v 2 + x 0 ∈ C \alpha v_1+\beta v_2 +x_0\in C αv1+βv2+x0∈C
⇐ α ( v 1 + x 0 ) ∈ C + β ( v 2 + x 0 ) ∈ C + ( 1 − α − β ) x 0 ∈ C ∈ C \Leftarrow \mathop{\alpha(v_1+x_0)}\limits_{\in C}+\mathop{\beta(v_2+x_0)}\limits_{\in C}+\mathop{(1-\alpha-\beta)x_0}\limits_{\in C}\ \in C ⇐∈Cα(v1+x0)+∈Cβ(v2+x0)+∈C(1−α−β)x0 ∈C
例子:
线性方程组的解集是仿射集(反之也成立)
C
=
{
x
∣
A
x
=
b
}
∀
∈
R
m
×
n
b
∈
R
m
x
∈
R
n
C=\{x|Ax=b\} \ \ \forall \in R^{m\times n} \ \ b\in R^m \ x\in R^n
C={x∣Ax=b} ∀∈Rm×n b∈Rm x∈Rn
构造新的子空间仿射集:
V
=
{
x
−
x
0
∣
x
∈
C
}
∀
x
0
∈
C
=
{
x
−
x
0
∣
A
x
=
b
}
A
x
0
=
b
=
{
x
−
x
0
∣
A
(
x
−
x
0
)
=
0
}
=
{
y
∣
A
y
=
0
}
→
化
零
空
间
\begin{aligned} V &= \{x-x_0|x\in C\}\quad \forall x_0 \in C\\ &= \{x-x_0|Ax=b\}\quad Ax_0=b \\ &= \{x-x_0|A(x-x_0)=0\} \\ &= \{y|Ay=0\}\rightarrow 化零空间 \end{aligned}
V={x−x0∣x∈C}∀x0∈C={x−x0∣Ax=b}Ax0=b={x−x0∣A(x−x0)=0}={y∣Ay=0}→化零空间
任意集合
C
C
C, 构造尽可能小的仿射集
仿射包
a
f
f
C
=
{
θ
1
x
1
+
⋯
+
θ
k
x
k
∣
∀
x
1
,
⋯
,
x
k
∈
C
,
∀
θ
1
+
⋯
+
θ
k
=
1
}
\mathop{aff}C=\{\theta_1 x_1+\cdots +\theta_k x_k\ |\ \forall x_1,\cdots,x_k\in C,\ \forall \theta_1+\cdots+\theta_k=1\}
affC={θ1x1+⋯+θkxk ∣ ∀x1,⋯,xk∈C, ∀θ1+⋯+θk=1}
1.3凸集
定义:
一个集合
C
C
C是凸集,当任意连接两点,空间的两点仍然在
C
C
C内
C
C
C is convex
⟺
\Longleftrightarrow
⟺
∀
x
1
,
x
2
∈
C
∀
θ
∈
[
0
,
1
]
θ
x
1
+
(
1
−
θ
)
x
2
∈
C
\forall x_1,x_2 \in C \ \ \ \forall \ \theta \in [0,1] \ \ \theta x_1+(1-\theta)x_2 \in C
∀x1,x2∈C ∀ θ∈[0,1] θx1+(1−θ)x2∈C (凸集是仿射集的特例)
设: x 1 x 2 , ⋯ x k ∈ C x_1x_2,\cdots x_k \in C x1x2,⋯xk∈C, θ 1 , θ 2 , ⋯ , θ k ∈ R \theta_1,\theta_2,\cdots,\theta_k\in R θ1,θ2,⋯,θk∈R & θ 1 + θ 2 + ⋯ + θ k = 1 \theta_1 + \theta_2 + \cdots + \theta_k=1 θ1+θ2+⋯+θk=1 & $\theta_1 \cdots \theta_k \in [0,1] $
泛化:
凸组合 θ 1 x 1 + θ 2 x 2 + ⋯ + θ k x k ∈ C \theta_1x_1 + \theta_2x_2 + \cdots + \theta_kx_k \in C θ1x1+θ2x2+⋯+θkxk∈C, ∀ x 1 , ⋯ , x k ∈ C , ∀ θ 1 ⋯ θ k ∈ [ 0 , 1 ] , ∀ θ 1 + ⋯ + θ k = 1 \forall x_1,\cdots,x_k\in C,\ \forall \theta_1\cdots\theta_k\in [0,1],\ \forall \theta_1+\cdots+\theta_k=1 ∀x1,⋯,xk∈C, ∀θ1⋯θk∈[0,1], ∀θ1+⋯+θk=1
C C C为凸集 ⇔ \Leftrightarrow ⇔任意元素凸组合 ∈ C \in C ∈C
凸包 C ∈ R n C\in R^n C∈Rn
C = { θ 1 x 1 + ⋯ + θ k x k ∣ ∀ x 1 , ⋯ , x k ∈ C , ∀ θ 1 ⋯ θ k ∈ [ 0 , 1 ] , ∀ θ 1 + ⋯ + θ k = 1 } C=\{\theta_1 x_1+\cdots +\theta_k x_k\ |\ \forall x_1,\cdots,x_k\in C,\ \forall \theta_1\cdots\theta_k\in [0,1],\ \forall \theta_1+\cdots+\theta_k=1\} C={θ1x1+⋯+θkxk ∣ ∀x1,⋯,xk∈C, ∀θ1⋯θk∈[0,1], ∀θ1+⋯+θk=1}
1.4锥 凸锥
定义:
C C C是锥 ⇔ \Leftrightarrow ⇔ ∀ x ∈ C \forall \ x\in C ∀ x∈C θ ≥ 0 \theta\ge0 θ≥0 有 θ x ∈ C \theta x\in C θx∈C
C
C
C是凸锥
⇔
\Leftrightarrow
⇔
∀
x
1
,
x
2
∈
C
\forall \ x_1,x_2\in C
∀ x1,x2∈C
θ
1
,
θ
2
≥
0
\theta_1,\theta_2\ge0
θ1,θ2≥0 有
x
1
θ
1
+
x
2
θ
2
∈
C
x_1\theta_1+x_2\theta_2\in C
x1θ1+x2θ2∈C
泛化:
凸锥组合 θ 1 x 1 + θ 2 x 2 + ⋯ + θ k x k \theta_1x_1 + \theta_2x_2 + \cdots + \theta_kx_k θ1x1+θ2x2+⋯+θkxk , θ 1 ⋯ θ k ≥ 0 {\theta_1 \cdots \theta_k \ge 0} θ1⋯θk≥0
凸锥包 C = { θ 1 x 1 + ⋯ + θ k x k ∣ ∀ x 1 , ⋯ , x k ∈ C , θ 1 ⋯ θ k ≥ 0 } C=\{\theta_1 x_1+\cdots +\theta_k x_k\ |\ \forall x_1,\cdots,x_k\in C,\theta_1 \cdots \theta_k \ge 0 \} C={θ1x1+⋯+θkxk ∣ ∀x1,⋯,xk∈C,θ1⋯θk≥0}
1.5总结
- 仿射组合: ∀ θ 1 , θ 2 , ⋯ , θ k ∈ R \forall \theta_1,\theta_2,\cdots,\theta_k\in R ∀θ1,θ2,⋯,θk∈R, θ 1 + θ 2 + ⋯ + θ k = 1 \theta_1 + \theta_2 + \cdots + \theta_k=1 θ1+θ2+⋯+θk=1
- 凸组合: ∀ θ 1 , θ 2 , ⋯ , θ k ∈ R \forall \theta_1,\theta_2,\cdots,\theta_k\in R ∀θ1,θ2,⋯,θk∈R, θ 1 + θ 2 + ⋯ + θ k = 1 \theta_1 + \theta_2 + \cdots + \theta_k=1 θ1+θ2+⋯+θk=1 , θ 1 ⋯ θ k ∈ [ 0 , 1 ] \theta_1\cdots\theta_k\in [0,1] θ1⋯θk∈[0,1]
- 凸锥组合: ∀ θ 1 , θ 2 , ⋯ , θ k ∈ R \forall \theta_1,\theta_2,\cdots,\theta_k\in R ∀θ1,θ2,⋯,θk∈R, θ 1 ⋯ θ k ≥ 0 \theta_1 \cdots \theta_k \ge 0 θ1⋯θk≥0
1.6参考
1、Stephen Boyd 、Lieven Vandenberghe——《Convex Optimization》)
2、中科大凌青凸优化 (https://www.bilibili.com/video/BV1Jt411p7jE?)