优化问题
R
n
R^n
Rn
m
i
n
f
0
(
x
)
f
i
(
x
)
⩽
0
,
i
=
1
,
2
,
.
.
.
,
m
h
i
(
x
)
=
0
,
i
=
1
,
2
,
.
.
.
,
p
D
=
⋂
m
d
o
m
f
i
⋂
p
d
o
m
h
i
minf_0(x) \\ f_i(x)\leqslant{0},i=1,2,...,m \\ h_i(x)=0,i=1,2,...,p \\ D=\overset{m}{\bigcap}domf_i\overset{p}{\bigcap}domh_i
minf0(x)fi(x)⩽0,i=1,2,...,mhi(x)=0,i=1,2,...,pD=⋂mdomfi⋂pdomhi
拉格朗日量
R
n
+
m
+
p
R^{n+m+p}
Rn+m+p
L
(
x
,
λ
,
v
)
=
f
0
(
x
)
+
∑
i
=
1
m
λ
i
f
i
(
x
)
+
∑
i
=
1
p
v
i
h
i
(
x
)
L(x,\lambda,v)=f_0(x)+\displaystyle\sum_{i=1}^m\lambda_if_i(x)+\displaystyle\sum_{i=1}^pv_ih_i(x)
L(x,λ,v)=f0(x)+i=1∑mλifi(x)+i=1∑pvihi(x)
拉格朗日对偶函数
R
m
+
p
R^{m+p}
Rm+p
g
(
λ
,
,
v
)
=
i
n
f
L
(
x
,
λ
,
v
)
=
i
n
f
f
0
(
x
)
+
∑
i
=
1
m
λ
i
f
i
(
x
)
+
∑
i
=
1
p
v
i
h
i
(
x
)
g(\lambda,,v)=infL(x,\lambda,v)=inff_0(x)+\displaystyle\sum_{i=1}^m\lambda_if_i(x)+\displaystyle\sum_{i=1}^pv_ih_i(x)
g(λ,,v)=infL(x,λ,v)=inff0(x)+i=1∑mλifi(x)+i=1∑pvihi(x)
对偶问题的重要性质:可以为原问题提供下界。
若限制
λ
i
⩾
0
\lambda_i\geqslant0
λi⩾0, 则
g
(
λ
,
v
)
⩽
p
∗
g(\lambda,v)\leqslant{p^*}
g(λ,v)⩽p∗
证明:
g
(
λ
,
v
)
⩽
f
0
(
x
)
+
∑
i
=
1
m
λ
i
f
i
(
x
)
+
∑
i
=
1
p
λ
i
f
i
(
x
)
+
∑
i
=
1
p
v
i
h
i
(
x
)
=
f
0
(
x
)
+
∑
i
=
1
m
+
p
λ
i
f
i
(
x
)
⩽
f
0
(
x
)
g(\lambda,v)\leqslant f_0(x)+\displaystyle\sum_{i=1}^m\lambda_if_i(x)+\displaystyle\sum_{i=1}^p\lambda_if_i(x)+\displaystyle\sum_{i=1}^pv_ih_i(x)\\ =f_0(x)+\displaystyle\sum_{i=1}^{m+p}\lambda_if_i(x) \\ \leqslant f_0(x)
g(λ,v)⩽f0(x)+i=1∑mλifi(x)+i=1∑pλifi(x)+i=1∑pvihi(x)=f0(x)+i=1∑m+pλifi(x)⩽f0(x)
对偶问题
m
i
n
f
0
(
x
)
s
.
t
.
λ
i
⩾
0
minf_0(x) \\ s.t. \lambda_i\geqslant0
minf0(x)s.t.λi⩾0
应用:最小向量范数,最大熵问题
对偶性质
\medspace
假设
m
i
n
f
0
∗
(
x
)
=
p
∗
,
m
a
x
g
0
(
λ
∗
,
v
∗
)
=
d
∗
minf_0^*(x)=p^*, maxg_0(\lambda^*,v^*)=d^*
minf0∗(x)=p∗,maxg0(λ∗,v∗)=d∗
\medspace
弱对偶性(总成立):
d
∗
⩽
p
∗
d^*\leqslant p^*
d∗⩽p∗
\medspace
强对偶性:
d
∗
=
p
∗
d^*=p^*
d∗=p∗
\medspace
\medspace
几乎所有的凸优化问题都满足强对偶性/若存在一个可行域中的点
x
x
x,使得
f
i
(
x
)
<
0
f_i(x)\lt 0
fi(x)<0,那么该凸优化满足强对偶性条件。