全部笔记的汇总贴(视频也有传送门):中科大-凸优化
min f 0 ( x ) s . t . f i ( x ) ≤ 0 , i = 1 , ⋯ , m h i ( x ) = 0 , i = 1 , ⋯ , P x ∈ R n , D = ∩ i = 0 m d o m f i ∩ ∩ i = 1 P d o m h i P ∗ o p t i m a l c a l u e \min f_0(x)\\s.t.\;f_i(x)\le0,i=1,\cdots,m\\h_i(x)=0,i=1,\cdots,P\\x\in\R^n,D=\cap_{i=0}^m dom\;f_i\cap\cap_{i=1}^P dom\;h_i\\P^*\;optimal\;calue minf0(x)s.t.fi(x)≤0,i=1,⋯,mhi(x)=0,i=1,⋯,Px∈Rn,D=∩i=0mdomfi∩∩i=1PdomhiP∗optimalcalue
拉格朗日函数(Lagrangian function/Lagrangian)
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_if_i(x)+\sum_{i=1}^P v_ih_i(x) L(x,λ,v)=f0(x)+i=1∑mλifi(x)+i=1∑Pvihi(x)
对偶函数(Lagrange Dual Function/Dual Function)
g ( λ , v ) = inf x ∈ D L ( x , λ , v ) g(\lambda,v)=\inf_{x\in D}L(x,\lambda,v) g(λ,v)=x∈DinfL(x,λ,v)
Lagrange Multiplier/Multipler
性质:
- 对偶函数为凹函数
sup x ∈ D L ( x , λ , v ) 凸 → inf x ∈ D L ( x , λ , v ) 凹 \sup_{x\in D}L(x,\lambda,v)凸\rightarrow\inf_{x\in D}L(x,\lambda,v)凹 x∈DsupL(x,λ,v)凸→x∈DinfL(x,λ,v)凹 - ∀ λ ≥ 0 , ∀ v , g ( λ , v ) ≤ P ∗ \forall \lambda\ge0,\forall v,g(\lambda,v)\le P^* ∀λ≥0,∀v,g(λ,v)≤P∗
证明:设 x ∗ x^* x∗是原问题的最优解,则必可行。
则 f i ( x ∗ ) ≤ 0 , h i ( x ∗ ) = 0 f_i(x^*)\le0,h_i(x^*)=0 fi(x∗)≤0,hi(x∗)=0
当 ∀ λ ≥ 0 , ∀ v \forall \lambda\ge0,\forall v ∀λ≥0,∀v,有 ∑ i = 1 m λ i f i ( x ∗ ) + ∑ i = 1 P v i h i ( x ∗ ) ≤ 0 \sum_{i=1}^m\lambda_if_i(x^*)+\sum_{i=1}^Pv_ih_i(x^*)\le0 ∑i=1mλifi(x∗)+∑i=1Pvihi(x∗)≤0
L ( x ∗ , λ , v ) = f 0 ( x ∗ ) ∑ i = 1 m λ i f i ( x ∗ ) + ∑ i = 1 P v i h i ( x ∗ ) ≤ P ∗ g ( λ , v ) ≤ P ∗ L(x^*,\lambda,v)=f_0(x^*)\sum_{i=1}^m\lambda_if_i(x^*)+\sum_{i=1}^Pv_ih_i(x^*)\le P^*\\g(\lambda,v)\le P^* L(x∗,λ,v)=f0(x∗)i=1∑mλifi(x∗)+i=1∑Pvihi(x∗)≤P∗g(λ,v)≤P∗
例:
min X T X s . t . A x = b , x ∈ R n , b ∈ R P , A ∈ R P ∗ n ⇒ L ( x , v ) = X T X + V T ( A x − b ) ⇒ g ( v ) = inf x ∈ D L ( x , v ) = inf x ∈ D X T X + V T A x − V T b 对 x 偏 导 2 x + A T V = 0 , x = − A T y 2 代 回 去 求 最 优 值 1 4 ( V T A A T V ) − 1 2 ( V T A A T V ) − V T b = − 1 4 V T A A T V − b T V 凹 ( 其 中 A A T 半 正 定 ) \min X^TX\\s.t.\;Ax=b,x\in\R^n,b\in\R^P,A\in\R^{P*n}\\\Rightarrow L(x,v)=X^TX+V^T(Ax-b)\\\Rightarrow g(v)=\inf_{x\in D}L(x,v)=\inf_{x\in D}X^TX+V^TAx-V^Tb\\对x偏导\;\;2x+A^TV=0,x=-\frac{A^Ty}2代回去求最优值\\\frac14(V^TAA^TV)-\frac12(V^TAA^TV)-V^Tb=-\frac14V^TAA^TV-b^TV\;\;\;凹(其中AA^T半正定) minXTXs.t.Ax=b,x∈Rn,b∈RP,A∈RP∗n⇒L(x,v)=XTX+VT(Ax−b)⇒g(v)=x∈DinfL(x,v)=x∈DinfXTX+VTAx−VTb对x偏导2x+ATV=0,x=−2ATy代回去求最优值41(VTAATV)−21(VTAATV)−VTb=−41VTAATV−bTV凹(其中AAT半正定)
例:
min C T x s . t . A x ≥ b x ≥ 0 ( A x − b = 0 − x ≤ 0 ) ⇒ L ( x , λ , v ) = C T x − λ T x + V T ( A x − b ) = − b T V + ( C + A T V − λ ) x ⇒ g ( λ , v ) = inf x ∈ D L ( x , λ , v ) = { − b t V C T + A T V − λ = 0 + ∞ o t h e r w i s e \min C^Tx\\s.t.\;Ax\ge b\;x\ge0\\(Ax-b=0\;\;-x\le0)\\\Rightarrow L(x,\lambda,v)=C^Tx-\lambda^Tx+V^T(Ax-b)\\=-b^TV+(C+A^TV-\lambda)x\\\Rightarrow g(\lambda,v)=\inf_{x\in D}L(x,\lambda,v)=\left\{ \begin{array}{l} -b^tV\;\;C^T+A^TV-\lambda=0 \\ \\+\infty\;\;\;otherwise \end{array} \right. minCTxs.t.Ax≥bx≥0(Ax−b=0−x≤0)⇒L(x,λ,v)=CTx−λTx+VT(Ax−b)=−bTV+(C+ATV−λ)x⇒g(λ,v)=x∈DinfL(x,λ,v)=⎩⎨⎧−btVCT+ATV−λ=0+∞otherwise
例:
min X T W X s . t . X i = ± 1 , i = 1 , ⋯ , m X i 2 − 1 = 0 ( 二 次 约 束 ) ⇒ L ( x , λ ) = X T W X + ∑ i = 1 m v i ( x i 2 − 1 ) = X T ( W + d i a g { v } ) x − 1 T v ⇒ g ( v ) = inf x ∈ D X T ( W + d i a g { v } ) X − 1 T v = { − 1 T v W + d i a g { v } ⪰ 0 − ∞ o t h e r w i s e \min X^TWX\\s.t.\;X_i=\pm 1,i=1,\cdots,m\\X_i^2-1=0(二次约束)\\\Rightarrow L(x,\lambda)=X^TWX+\sum_{i=1}^mv_i(x_i^2-1)\\=X^T(W+diag\{v\})x-1^Tv\\\Rightarrow g(v)=\inf_{x\in D}X^T(W+diag\{v\})X-1^Tv=\left\{ \begin{array}{l} -1^Tv\;\;\;W+diag\{v\}\succeq0 \\ \\-\infty\;\;\;otherwise \end{array} \right. minXTWXs.t.Xi=±1,i=1,⋯,mXi2−1=0(二次约束)⇒L(x,λ)=XTWX+i=1∑mvi(xi2−1)=XT(W+diag{ v})x−1Tv⇒g(v)=x∈DinfXT(W+diag{ v})X−1Tv=⎩⎨⎧−1TvW+diag{ v}⪰0−∞otherwise
下一章传送门:中科大-凸优化 笔记(lec30)-Lagrange对偶(二)