P14 保证函数凸性
保证函数凸性
非负加权和
f
i
.
.
.
f
m
f_i...f_m
fi...fm 为凸,则
f
=
∑
i
=
1
m
w
i
f
i
f=\sum_{i=1}^m w_i f_i
f=∑i=1mwifi 为凸 ,
w
i
≥
0
,
∀
i
w_i \geq 0 , \forall i
wi≥0,∀i
证:
设
f
,
g
f,g
f,g都是凸函数,有
f
,
g
f,g
f,g的二阶导数都是非负的;根据导数的运算法则有:
(
f
+
g
)
(f+g)
(f+g) 的二阶导数等于
f
,
g
f,g
f,g 的二阶导数相加,因此
(
f
+
g
)
(f+g)
(f+g) 的二阶导数是非负的,所以
(
f
+
g
)
(f+g)
(f+g) 是凸函数。
若
f
(
x
,
y
)
f(x,y)
f(x,y), 对任意
y
∈
A
y \in A
y∈A
f
(
x
,
y
)
\quad f(x,y)
f(x,y) 均为凸
(
x
,
y
)
(x,y)
(x,y) jointly convex
设
w
(
y
)
≥
0
,
∀
y
∈
A
w(y) \geq 0, \forall y \in A
w(y)≥0,∀y∈A
g
(
x
)
=
∫
y
∈
A
w
(
y
)
f
(
x
,
y
)
d
y
g(x) = \int_{y \in A} w(y) f(x,y) dy
g(x)=∫y∈Aw(y)f(x,y)dy 为凸
仿射映射
f
:
R
n
↦
R
A
∈
R
n
∗
m
b
∈
R
n
f:R^n \mapsto R \quad A \in R^{n*m} \quad b \in R^n
f:Rn↦RA∈Rn∗mb∈Rn
g
(
x
)
=
f
(
A
x
+
b
)
d
o
m
g
=
{
x
∣
A
x
+
b
∈
d
o
m
f
}
g(x)=f(Ax+b) \quad dom g = \lbrace x|Ax+b \in dom f \rbrace
g(x)=f(Ax+b)domg={x∣Ax+b∈domf}
证:
令
x
,
y
∈
d
o
m
g
0
≤
θ
≤
1
x,y \in dom g \quad 0 \leq \theta \leq 1
x,y∈domg0≤θ≤1
g
(
θ
x
+
(
1
−
θ
)
y
)
=
f
(
θ
A
x
+
(
1
−
θ
)
A
y
+
b
)
=
f
(
θ
(
A
x
+
b
)
+
(
1
−
θ
)
(
A
y
+
b
)
)
≤
θ
f
(
A
x
+
b
)
+
(
1
−
θ
)
f
(
A
y
+
b
)
=
θ
g
(
x
)
+
(
1
−
θ
)
g
(
y
)
\begin{aligned} g(\theta x + (1-\theta)y)&=f(\theta Ax + (1-\theta)Ay + b)\\ &=f(\theta(Ax+b)+(1-\theta)(Ay+b)) \\ & \leq \theta f(Ax+b) + (1-\theta)f(Ay +b) \\ &= \theta g(x) + (1-\theta)g(y) \end{aligned}
g(θx+(1−θ)y)=f(θAx+(1−θ)Ay+b)=f(θ(Ax+b)+(1−θ)(Ay+b))≤θf(Ax+b)+(1−θ)f(Ay+b)=θg(x)+(1−θ)g(y)
所以:仿射函数为凸函数
f
i
R
n
↦
R
,
i
=
1...
m
f_i \quad R^n \mapsto R,i=1...m
fiRn↦R,i=1...m为凸
A
∈
R
n
,
b
∈
R
A \in R^n,b \in R
A∈Rn,b∈R
g
(
x
)
=
A
T
[
f
1
(
x
)
.
.
.
.
f
m
(
x
)
]
T
+
b
g(x) = A^T[f_1(x)....f_m(x)]^T+b
g(x)=AT[f1(x)....fm(x)]T+b
不能保证A都是正的,不是凸函数
两个函数的极大值函数
f
1
,
f
2
f_1,f_2
f1,f2为凸,定义
f
(
x
)
=
m
a
x
{
f
1
(
x
)
,
f
2
(
x
)
}
f(x)=max \lbrace f_1(x),f_2(x) \rbrace
f(x)=max{f1(x),f2(x)}
d
o
m
f
=
d
o
m
f
1
⋂
d
o
m
f
2
dom f=dom f_1 \bigcap dom f_2
domf=domf1⋂domf2
设:
x
,
y
∈
d
o
m
f
0
≤
θ
≤
1
x,y \in dom f \quad 0\leq\theta \leq 1
x,y∈domf0≤θ≤1
f
(
θ
x
+
(
1
−
θ
)
y
)
=
m
a
x
{
f
1
(
θ
x
+
(
1
−
θ
)
y
)
,
f
2
(
θ
x
+
(
1
−
θ
)
y
)
}
≤
m
a
x
{
θ
f
1
(
x
)
+
(
1
−
θ
)
f
1
(
y
)
,
θ
f
2
(
x
)
+
(
1
−
θ
)
f
2
(
y
)
}
≤
m
a
x
{
θ
f
1
(
x
)
,
θ
f
2
(
x
)
}
+
m
a
x
{
(
1
−
θ
)
f
1
(
y
)
,
(
1
−
θ
)
f
2
(
y
)
}
=
θ
f
(
x
)
+
(
1
−
θ
)
f
(
y
)
\begin{aligned} f(\theta x + (1-\theta)y) &=max \lbrace f_1(\theta x + (1-\theta)y), f_2(\theta x + (1-\theta)y) \rbrace \\ &\leq max \lbrace \theta f_1(x) + (1-\theta)f_1(y),\theta f_2(x) + (1-\theta)f_2(y) \rbrace \\ &\leq max \lbrace \theta f_1(x), \theta f_2(x) \rbrace + max \lbrace (1-\theta)f_1(y), (1-\theta)f_2(y) \rbrace \\ &=\theta f(x) + (1-\theta)f(y) \end{aligned}
f(θx+(1−θ)y)=max{f1(θx+(1−θ)y),f2(θx+(1−θ)y)}≤max{θf1(x)+(1−θ)f1(y),θf2(x)+(1−θ)f2(y)}≤max{θf1(x),θf2(x)}+max{(1−θ)f1(y),(1−θ)f2(y)}=θf(x)+(1−θ)f(y)
例:
向量中r个最大元素的和,
x
∈
R
n
x \in R^n
x∈Rn
x
[
i
]
x[i]
x[i] 第i大元素
f
(
x
)
=
∑
i
=
1
r
x
[
i
]
f(x)=\sum_{i=1}^r x[i]
f(x)=∑i=1rx[i]
f
(
x
)
=
max
{
x
i
1
+
.
.
.
+
x
i
r
∣
1
≤
i
1
.
.
.
≤
i
r
≤
n
}
f(x)=\max \lbrace x_{i1} + ... + x_{ir} |1 \leq i_1... \leq i_r \leq n \rbrace
f(x)=max{xi1+...+xir∣1≤i1...≤ir≤n}
无限个凸函数极大值
f
(
x
,
y
)
f(x,y)
f(x,y)对于x为凸,
∀
y
∈
A
\forall y \in A
∀y∈A
g
=
s
u
p
y
∈
A
f
(
x
,
y
)
g=sup_{y \in A}f(x,y)
g=supy∈Af(x,y)
例:
实对称阵的最大特征值
f
(
x
)
=
λ
m
a
x
(
x
)
d
o
m
f
=
S
m
f(x) = \lambda max (x) \quad dom f = S^m
f(x)=λmax(x)domf=Sm
x
y
=
λ
y
xy = \lambda y
xy=λy
y
T
x
y
=
y
T
λ
y
y^T xy = y^T \lambda y
yTxy=yTλy
y
T
x
y
=
λ
∣
∣
y
∣
∣
2
2
y^T xy = \lambda ||y||_2^2
yTxy=λ∣∣y∣∣22
λ
=
y
T
x
y
∣
∣
y
∣
∣
2
2
\lambda = \frac{y^T xy }{||y||_2^2}
λ=∣∣y∣∣22yTxy
λ
=
y
T
x
y
∣
∣
y
∣
∣
2
=
1
\lambda = y^T xy \quad ||y||_2=1
λ=yTxy∣∣y∣∣2=1
λ
m
a
x
(
x
)
=
s
u
p
{
y
T
x
y
∣
∣
∣
y
∣
∣
2
=
1
}
\lambda max(x)=sup \lbrace y^T xy \;| \;||y||_2=1 \rbrace
λmax(x)=sup{yTxy∣∣∣y∣∣2=1}