对于一个最优化问题:
min f 0 ( x ) s.t. f i ( x ) ≤ 0 , i = 1 , ⋯ , m h i ( x ) = 0 , i = 1 , ⋯ , p \begin{array}{ll}\min & f_{0}(x) \\ \text {s.t.} & f_{i}(x) \leq 0, \quad i=1, \cdots, m \\ & h_{i}(x)=0, \quad i=1, \cdots, p\end{array} mins.t.f0(x)fi(x)≤0,i=1,⋯,mhi(x)=0,i=1,⋯,p
f 0 ( x ) f_{0}(x) f0(x)为目标函数, f i ( x ) ≤ 0 f_{i}(x)\leq0 fi(x)≤0为不等式约束, h i ( x ) = 0 h_{i}(x)=0 hi(x)=0为等式约束.
通过引入不等式约束和等式约束的lagranage乘子 λ i \lambda_i λi与 v i v_i vi,得到原问题的Lagrange函数为:
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)+\sum_{i=1}^{m} \lambda_{i} f_{i}(x)+\sum_{i=1}^{p} v_{i} h_{i}(x) L(x,λ,v)=f0(x)+i=1∑mλifi(x)+i=1∑pvihi(x)
定义Lagrange对偶函数为Lagrange 函数对 x x x求最小值,
g ( λ , v ) = inf x L ( x , λ , v ) = inf x ( f 0 ( x ) + ∑ i = 1 m λ i f i ( x ) + ∑ i = 1 p v i h i ( x ) ) g(\lambda, v)=\inf _{x } L(x, \lambda, v)=\inf _{x}\left(f_{0}(x)+\sum_{i=1}^{m} \lambda_{i} f_{i}(x)+\sum_{i=1}^{p} v_{i} h_{i}(x)\right) g(λ,v)=xinfL(x,λ,v)=xinf(f0(x)+i=1∑mλifi(x)+i=1∑pvihi(x))
根据对偶问题的性质,Lagrange对偶函数构成了原问题最优值 p ∗ p^* p∗的下界,即:
g ( λ , v ) ≤ p ∗ g(\lambda, v)\leq p^* g(λ,v)≤p∗
于是,我们再对Lagrange对偶函数求上界,便构成了Lagrange对偶问题:
max g ( λ , v ) s.t. λ i ≥ 0 , i = 1 , 2 , ⋯ , q \begin{array}{ll}\max & g(\lambda,v) \\ \text {s.t.} & \lambda_i \geq 0,i=1,2,\cdots,q\end{array} maxs.t.g(λ,v)λi≥0,i=1,2,⋯,q
我们用 d "表示 Lagrange 对偶问题的最优解,则有
d ∗ ≤ p ∗ d^{*} \leq p^{*} d∗≤p∗
即使原问题不是凸优化问题,这个不等式也成立。这个性质称作弱对偶性。
若 d ∗ = p ∗ d^{*} = p^{*} d∗=p∗成立,则成强对偶性成立。
若原问题为凸优化问题,则强对偶性通常成立。
当强对偶性成立时, KaTeX parse error: Expected '\right', got 'EOF' at end of input: \left.x^{*}是原问题的最优解, λ ∗ , v ∗ \lambda^{*}, v^{*} λ∗,v∗ 是对偶问题的最优解的一个必要条件是互补松他性条件:
∑ i = 1 m λ i ∗ f i ( x ∗ ) = 0 \sum_{i=1}^{m} \lambda_{i}^{*} f_{i}\left(x^{*}\right)=0 i=1∑mλi∗fi(x∗)=0
因为每一项都非正,因此有
λ i ∗ f i ( x ∗ ) = 0 , i = 1 , ⋯ , m \lambda_{i}^{*} f_{i}\left(x^{*}\right)=0, \quad i=1, \cdots, m λi∗fi(x∗)=0,i=1,⋯,m
当强对偶性成立时, x ∗ x^{*} x∗ 和 ( λ ∗ , v ∗ ) \left(\lambda^{*}, v^{*}\right) (λ∗,v∗) 分别是原问题和对偶问题的最优
解,则其必须满足 KKT 条件:
f i ( x ∗ ) ≤ 0 , i = 1 , ⋯ , m h i ( x ∗ ) = 0 , i = 1 , ⋯ , p λ i ∗ ≥ 0 , i = 1 , ⋯ , m λ i ∗ f i ( x ∗ ) = 0 , i = 1 , ⋯ , m ∇ f 0 ( x ∗ ) + ∑ i = 1 m λ i ∗ ∇ f i ( x ∗ ) + ∑ i = 1 p v i ∗ ∇ h i ( x ∗ ) = 0 \begin{array}{l} f_{i}\left(x^{*}\right) \leq 0, \quad i=1, \cdots, m \\ h_{i}\left(x^{*}\right)=0, \quad i=1, \cdots, p \\ \lambda_{i}^{*} \geq 0, \quad \quad\quad i=1, \cdots, m \\ \lambda_{i}^{*} f_{i}\left(x^{*}\right)=0, \quad i=1, \cdots, m \\ \nabla f_{0}\left(x^{*}\right)+\sum_{i=1}^{m} \lambda_{i}^{*} \nabla f_{i}\left(x^{*}\right)+\sum_{i=1}^{p} v_{i}^{*} \nabla h_{i}\left(x^{*}\right)=0 \end{array} fi(x∗)≤0,i=1,⋯,mhi(x∗)=0,i=1,⋯,pλi∗≥0,i=1,⋯,mλi∗fi(x∗)=0,i=1,⋯,m∇f0(x∗)+∑i=1mλi∗∇fi(x∗)+∑i=1pvi∗∇hi(x∗)=0
当原问题是凸优化问题时,满足KKT条件的点也是原、对偶问题的最优解。(证明后补)