second order condition
equality constraints
First consider the case of only equality constraints
min x ∈ R n f ( x ) s . t . ci ( x ) = 0 , i = 1 , 2 , ⋯ , m \begin{aligned} \min_{x\in\mathbb {R}^n} ~~& f(x) \\ \mathrm{st} ~~& c_i(x) = 0,i = 1,2,\cdots,m \end{aligned}x∈Rnmin s.t. f(x)ci(x)=0,i=1,2,⋯,m
Suppose the optimal solution x ∗ x^*x∗ exists, and at pointx ∗ x^*x∗ Directional quantityai ∗ ( i ∈ E ) a_i^* ~(i \in\mathcal{E})ai∗ (i∈E ) Linear independence (that is, the LICQ constraint specification holds). For the sequence feasible directionp ∈ F ∗ p \in \mathcal{F}^*p∈F∗ , there exists a feasible sequencex ( k ) x^{(k)}x( k ) and the corresponding direction sequencep ( k ) → pp^{(k)} \to pp(k)→For p
, let f ( x ( k ) ) = f ( x ∗ + δ kp ( k ) ) = L ( x ∗ + δ kp ( k ), λ ∗ ) f(x^{(k)} ) = f(x^* + \delta_k p^{(k)}) = \mathcal{L}(x^* + \delta_k p^{(k)}, \lambda^*)f(x(k))=f(x∗+dkp(k))=L(x∗+dkp(k),l∗ )
becausex ∗ x^*x∗ isL \mathcal{L}The stable point of L ( ∇ L = 0 \nabla \mathcal{L} = 0∇L=0 ), soL \mathcal{L}L atx ∗ x^*x∗ Specialized Taylor equation
f ( x ∗ + δ kp ( k ) ) = L ( x ∗ + δ kp ( k ) , λ ∗ ) = f ∗ + 1 2 δ k 2 p ( k ) TW ∗ p ( k ) + o ( δ k ) \begin{aligned} f(x^* + \delta_k p^{(k)}) &= \mathcal{L}(x^* + \delta_k p^{(k). )}, \lambda^*)\\ &=f^* + \frac{1}{2} \delta_k^2 {p^{(k)}}^TW^* p^{(k)} + o (\delta_k) \end{aligned}f(x∗+dkp(k))=L(x∗+dkp(k),l∗)=f∗+21dk2p(k)TW∗p(k)+o ( dk)
For
W ∗ = ∇ x 2 L ( x ∗ , λ ∗ ) = ∇ 2 f ( x ∗ ) + ∑ i = 1 m λ i ∗ ∇ 2 ci ( x ∗ ) W^* = \nabla_x^2 \mathcal{ L}(x^*,\lambda^*) = \nabla^2f(x^*) + \sum_{i=1}^m \lambda_i^*\nabla^2c_i(x^*)W∗=∇x2L(x∗,l∗)=∇2f(x∗)+i=1∑mli∗∇2 ci(x∗ )
means the Lagrange function aboutxxHessian matrix of x .
For example,if
f ( x ∗ + δ kp ( k ) ) δ k 2 = f ∗ δ k 2 + 1 2 p ( k ) TW ∗ p ( k ) + o ( δ k ) δ k 2 \frac{f( x^* + \delta_k p^{(k)})}{\delta_k^2} =\frac{f^*}{\delta_k^2} + \frac{1}{2} {p^{(k )}}^TW^* p^{(k)} + \fraction(\delta_k)}{\delta_k^2}dk2f(x∗+dkp(k))=dk2f∗+21p(k)TW∗p(k)+dk2o ( dk)
And k → ∞ k \to \inftyk→∞ , andx ∗ x^*x∗ is a local minimum point,
p ( k ) TW ∗ p ( k ) ≥ 0 {p^{(k)}}^TW^* p^{(k)}\geq 0p(k)TW∗p(k)≥0
Theorem (second-order necessary condition) Let x ∗ x^*x∗ is the local minimum point of the problem and satisfies the KKT condition, and the corresponding Lagrange multiplier isλ ∗ \lambda^*l∗ , then reverse ppis feasible for any sequencep ,if
p TW ∗ p ≥ 0 p^TW^* p \geqpTW∗p≥0
inferenceLetx ∗ x^*x∗ is the local minimum point of the problem and satisfies the KKT condition, and the corresponding Lagrange multiplier isλ ∗ \lambda^*l∗,若 F ∗ = F ∗ \mathcal{F}^* = F^* F∗=F∗ ,if
p TW ∗ p ≥ 0 , ∀ p ∈ F ∗ p^TW^* p \geq 0, \forall p \in F^*pTW∗p≥0,∀p∈F∗
Theorem (second-order sufficient condition) Let x ∗ x^*x∗ is the KKT point of the problem, and the corresponding Lagrange multiplier isλ ∗ \lambda^*l∗ , if condition
p TW ∗ p > 0 , ∀ p ∈ F ∗ p^TW^*p > 0, \forall p \in F^*pTW∗p>0,∀p∈F∗
is established, thenx ∗ x^*x∗ is a strict local minimum of the problem.
Example Consider the problem
min f ( x ) = 1 2 ( x 1 − 1 ) 2 + 1 2 x 2 2 s . t . c ( x ) = x 1 − β x 2 2 = 0 \begin{aligned} \min ~~& f(x) = \frac{1}{2}(x_1 - 1)^2 + \frac{1}{2}x_2^2 \\ \mathrm{st}~~&c(x) = x_1 - \beta x_2^2 = 0 \end{aligned}min s.t. f(x)=21(x1−1)2+21x22c(x)=x1−βx22=0
Discussion parameter β \betaWhen β takes any value,x ∗ = ( 0 , 0 ) T x^*=(0,0)^Tx∗=(0,0)T is a local minimum point?
Solution Since g ∗ = ( − 1 , 0 ) T g^* = (-1,0)^Tg∗=(−1,0)T, a ∗ = ( 1 , 0 ) T a^* = (1,0)^T a∗=(1,0)T , so the first-order condition is satisfied,x ∗ x^*x∗ is a KKT point, andλ ∗ = 1 \lambda^* = 1l∗=1, W ∗ = ( 1 0 0 1 − 2 β ) T W^* = \begin{pmatrix} 1 & 0 \\ 0 &1- 2\beta \end{pmatrix}^T W∗=(1001−2 b)T
又 F ∗ = { p = ( 0 , p 2 ) T : p 2 ≠ 0 } F^* = \{p=(0,p_2)^T:p_2\neq0\} F∗={ p=(0,p2)T:p2=0 } , sop TW ∗ p = ( 1 − 2 β ) p 2 2 p^TW^*p = (1-2\beta)p_2^2pTW∗p=(1−2β)p22. Thus, there are
- 当 β < 1 2 \beta < \frac{1}{2} b<21, x ∗ x^* x∗ is a strict local minimum point
- 当β > 1 2 \beta > \frac{1}{2}b>21, x ∗ x^* x∗ Not a local minimum point
- 当β = 1 2 \beta = \frac{1}{2}b=21, x ∗ x^* x∗ is a strict local minimum point
Weak Positive Constraint vs. Strong Positive Constraint
Let x ∗ x^*x∗ is a KKT point,λ ∗ \lambda^*l∗ is the corresponding Lagrange multiplier.
Define
KaTeX parse error: Undefined control sequence: \or at position 41: …in\mathcal{E} ~\̲o̲r̲~ \lambda_i^* >…
as a set of strong positive constraints. That is, from A ∗ \mathcal{A}^*A∗ removes weak positive constraints, that is,λ i ∗ = 0 , i ∈ I ∗ \lambda^*_i = 0,i\in\mathcal{I}^*li∗=0,i∈I∗ , getA + ∗ \mathcal{A}^*_+A+∗。定义
G ∗ = { p ∈ F ∗ : c i ( x ( k ) ) = 0 , ∀ i ∈ A + ∗ } \mathcal{G}^* = \{p\in\mathcal{F}^*:c_i(x^{(k)}) = 0,\forall i \in \mathcal{A}^*_+\} G∗={ p∈F∗:ci(x(k))=0,∀i∈A+∗}G ∗ = { p ∈ R n : p ≠ 0 , a i ∗ T p = 0 , i ∈ A + ∗ , a i ∗ T p ≤ 0 , i ∈ A ∗ \ A + ∗ } G^* = \{p\in \mathbb{R}^n: p \neq 0, {a_i^*}^Tp = 0, i \in \mathcal{A}^*_+, {a_i^*}^Tp \leq 0, i \in \mathcal{A}^* \backslash \mathcal{A}_+^*\} G∗={ p∈Rn:p=0,ai∗Tp=0,i∈A+∗,ai∗Tp≤0,i∈A∗\A+∗}
In fact, there are
- G ∗ ⊂ G ∗ \mathcal{G}^* \subset G^* G∗⊂G∗
- 设 p ∈ F ∗ p \in F^* p∈F∗ , thenp T g ∗ = 0 p^T g^* = 0pTg∗=0 if and only ifp ∈ G ∗ p \in G^*p∈G∗
Regularity assumption 2: G ∗ = G ∗ \mathcal{G}^* = G^*G∗=G∗
general questions
First, the general constraint optimization problem
min x ∈ R n f ( x ) s . t . ci ( x ) = 0 , i ∈ E ci ( x ) ≤ 0 , i ∈ I \begin{aligned} \min_{x\in\ mathbb{R}^n} ~~& f(x) \\ \mathrm{st} ~~& c_i(x) = 0,i \in \mathcal{E}\\ & c_i(x) \leq 0, i \in \mathcal{I}\\ \end{aligned}x∈Rnmin s.t. f(x)ci(x)=0,i∈Eci(x)≤0,i∈I
There are the following second-order conditions
Theorem (second-order necessary condition) Let x ∗ x^*x∗ is the local minimum point of the problem and satisfies the KKT condition, and the corresponding Lagrange multiplier isλ ∗ \lambda^*l∗ , if regularization condition 2 holds, there must be
p TW ∗ p ≥ 0 , ∀ p ∈ G ∗ p^TW^* p \geq 0,\forall p \in G^*pTW∗p≥0,∀p∈G∗
Theorem (second-order sufficient condition) Let x ∗ x^*xThere is a Lagrange multiplier at ∗ so that the KKT condition is true, and the corresponding Lagrange multiplier isλ ∗ \lambda^*l∗ , if condition
p TW ∗ p > 0 , ∀ p ∈ G ∗ p^TW^*p > 0, \forall p \in G^*pTW∗p>0,∀p∈G∗
is established, thenx ∗ x^*x∗ is a strict local minimum of the constrained problem.
references
[1] Liu Hongying, Xia Yong, Zhou Yongsheng. Fundamentals of Mathematical Programming, Beijing, 2012.