本文探讨了拉格朗日对偶问题，包括最大化问题的求解，以及弱对偶性和强对偶性的概念及其应用。通过实例演示了线性规划和二次规划中的对偶问题，并介绍了斯莱特约束条件。文章重点讲解了如何利用斯莱特条件确保在凸问题中弱对偶等于强对偶，以及如何验证严格可行条件下的对偶最优性质。
Navigator
Lagrange dual problem
max g ( λ , ν ) s . t . λ ≥ 0 \begin{aligned} &\max &g(\lambda, \nu)\\ &s.t. &\lambda\geq 0 \end{aligned} maxs.t.g(λ,ν)λ≥0
- Find the best lower bound p ∗ p^* p∗
- CVX problem, the optimal value is d ∗ d^* d∗
- λ , ν \lambda, \nu λ,ν are dual feasible of λ ≥ 0 , ( λ , ν ) ∈ d o m g \lambda\geq 0, (\lambda, \nu)\in dom g λ≥0,(λ,ν)∈domg
Weak dual and Strong dual
weak duality:
d
∗
≤
p
∗
d^*\leq p^*
d∗≤p∗.
strong duality:
d
∗
=
p
∗
d^*=p^*
d∗=p∗. usually hold in convex problem. Add some constrained qualification, make strong duality hold.
illustration
min
f
0
(
x
)
s
.
t
.
f
i
(
x
)
≥
0
\begin{aligned} &\min &f_0(x)\\ &s.t. & f_i(x)\geq 0 \end{aligned}
mins.t.f0(x)fi(x)≥0
Define
g
(
λ
)
=
inf
(
u
,
t
)
∈
G
(
t
+
λ
u
)
,
G
=
{
(
f
1
(
x
)
,
f
0
(
x
)
)
∣
∃
x
∈
D
}
g(\lambda)=\inf_{(u, t)\in G}(t+\lambda u), G=\{(f_1(x), f_0(x))| \exist x\in D\}
g(λ)=inf(u,t)∈G(t+λu),G={(f1(x),f0(x))∣∃x∈D}. Expand the direction of u and t:
A
=
{
(
u
,
t
)
∣
f
1
(
x
)
≤
u
,
f
0
(
x
)
≤
t
,
for some
x
∈
D
}
A=\{(u, t)| f_1(x)\leq u, f_0(x)\leq t, \text{for some } x\in D\}
A={(u,t)∣f1(x)≤u,f0(x)≤t,for some x∈D}
Proof the assertion: if
f
1
(
x
)
f_1(x)
f1(x) and
f
0
(
x
)
f_0(x)
f0(x) are convex. Let
(
u
1
,
t
1
)
,
(
u
2
,
t
2
)
∈
A
(u_1, t_1), (u_2, t_2)\in A
(u1,t1),(u2,t2)∈A.
θ
u
1
+
(
1
−
θ
)
u
2
∈
A
\theta u_1+(1-\theta)u_2\in A
θu1+(1−θ)u2∈A.
∃
x
1
f
1
(
x
1
)
≤
u
1
∃
x
2
f
1
(
x
2
)
≤
u
2
→
f
(
θ
x
1
+
(
1
−
θ
)
x
2
)
≤
θ
f
(
x
1
)
+
(
1
−
θ
)
f
(
x
2
)
≤
θ
u
1
+
(
1
−
θ
)
u
2
\begin{aligned} &\exist x_1 & f_1(x_1)\leq u_1\\ &\exist x_2 & f_1(x_2)\leq u_2\\ &\rightarrow &f(\theta x_1+(1-\theta)x_2)\leq \theta f(x_1)+(1-\theta)f(x_2)\leq \theta u_1+(1-\theta)u_2 \end{aligned}
∃x1∃x2→f1(x1)≤u1f1(x2)≤u2f(θx1+(1−θ)x2)≤θf(x1)+(1−θ)f(x2)≤θu1+(1−θ)u2
Slater’s constraint condition
String duality holds for a convex problem:
min
f
0
(
x
)
s
.
t
.
f
i
(
x
)
≤
0
i
=
1
,
2
,
…
,
m
A
x
=
b
\begin{aligned} &\min & f_0(x)\\ &s.t. & f_i(x)\leq 0 & i=1,2,\dots, m\\ &&Ax=b \end{aligned}
mins.t.f0(x)fi(x)≤0Ax=bi=1,2,…,m
If it is strictly feasible i.e.;
∃
x
∈
i
n
t
D
,
f
i
(
x
)
<
0
,
A
x
=
b
\exists x\in int D, f_i(x)<0, Ax=b
∃x∈intD,fi(x)<0,Ax=b. The following conclusions are hold:
- also guarantees that dual optimal d ∗ d^* d∗ is attained of p ∗ > − ∞ p^*>-\infty p∗>−∞.
- ReInt C= { x ∈ A f f C ∣ B ( x , r ) ∩ A f f C ∈ C for some r > 0 } \{x\in Aff C| B(x, r)\cap Aff C \in C \text{for some }r>0\} {x∈AffC∣B(x,r)∩AffC∈Cfor some r>0}.
- linear inequality do not need to hold with strict inequality.
Demo: Inequality form LP
primal problem:
min
c
′
x
s
.
t
.
A
x
≤
b
\begin{aligned} &\min & c'x\\ &s.t.& Ax\leq b \end{aligned}
mins.t.c′xAx≤b
dual function:
g
(
λ
)
=
inf
x
(
(
c
+
A
′
λ
)
x
−
b
′
λ
)
=
{
−
b
′
λ
,
A
′
λ
+
c
=
0
−
∞
,
O
.
W
.
g(\lambda)=\inf_x ((c+A'\lambda)x-b'\lambda)= \left\{ \begin{aligned} &-b'\lambda, &A'\lambda + c=0\\ &-\infty, &O.W. \end{aligned} \right.
g(λ)=xinf((c+A′λ)x−b′λ)={−b′λ,−∞,A′λ+c=0O.W.
from Slater’s conditon:
p
∗
=
d
∗
p^*=d^*
p∗=d∗.
Demo: Quadratic Programming
primal problem:
min
x
′
P
x
s
.
t
.
A
x
≤
b
\begin{aligned} &\min & x'Px\\ &s.t.& Ax\leq b \end{aligned}
mins.t.x′PxAx≤b
The dual problem is:
g
(
λ
)
=
inf
x
(
x
′
P
x
+
λ
′
(
A
x
−
b
)
)
=
−
1
4
λ
′
A
P
−
1
A
λ
−
b
′
λ
max
λ
g
(
λ
)
s
.
t
.
λ
≥
0
\begin{aligned} &g(\lambda)=\inf_x (x'Px+\lambda'(Ax-b))=-\frac{1}{4}\lambda'AP^{-1}A\lambda-b'\lambda\\ &\max_\lambda g(\lambda)\\ &s.t. \lambda\geq 0 \end{aligned}
g(λ)=xinf(x′Px+λ′(Ax−b))=−41λ′AP−1Aλ−b′λλmaxg(λ)s.t.λ≥0
Reference
B站 许志钦
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
