Extreme Points and Extreme Rays

原文链接： Describing Polyhedra by Extreme Points and Extreme Rays

原文作者：John Mitchell

Let P : = { x ∈ R n : A x ≤ b } \bm{P:}=\{x\in \mathcal{R}^n: Ax \leq b \} P:={x∈Rn:Ax≤b}, where $A$ is an $m\times n$ matrix, $x$ is an $n$-vector, and $b$ is an $m$-vector. Assume rank($A$)=$n$ and $\mathbf{P}\ne \varphi$. We look at the extreme points and extreme rays of $P$.

1）先描述 Extreme Points 的定义：

说明： 针对有界的线性约束，只需考虑 Extreme Points (极点) 即可

• Definition 1
  A point $x\in P$ is an extreme point of $P$, if there do not exist point $x^1$ and $x^2$ in $P$ and $a$ scalar $\lambda$ (with $x^1\ne x^2$ and $0<\lambda <1$ ) such that $x=\lambda x^1+(1-\lambda )x^2$.
  （解读：extreme point 就是有界可行集组成的多面体的边界点）

• Proposition 1
  $x\in P$ is an extreme point of $P$ if and only if $x$ is a zero-dimensional face${}^{标注[1]}$ of $P$.
  （解读：性质1与定义1讲得是一个意思，所谓的 extreme point 就是多面体的顶点）

2）其次描述 Extreme Rays 的定义：

说明： 针对无界的线性约束，则需考虑 Extreme Points (极射线)

• Definition 2
  Let P 0 : = { r ∈ R n : A r ≤ 0 } \bm{P^0:}=\{r\in \mathcal{R}^n: Ar \leq 0 \} P0:={r∈Rn:Ar≤0}. Any r ∈ P 0 \ { 0 } r\in P^0 \backslash\{0\} r∈P0\{0} is a ray of $P$.
  Note: ${\mathbf{P}}^0$ is a convex cone. A set $K$ is a cone if $x\in K,\lambda >0$ implies $\lambda x\in K$.
  A point $r\in {\mathcal{R}}^n$ is a ray of P if and only if for any point $x\in P$ the set { y ∈ R n : y = x + λ r , λ ≥ 0 } ⊆ P \{y\in \mathcal{R}^n:y=x+\lambda r, \lambda \geq 0\}\subseteq P {y∈Rn:y=x+λr,λ≥0}⊆P
  （解读：定义2就是为了引出定义3。其中，${\mathbf{P}}^0$是凸锥，这里给出了锥$K$的定义，且最后一行写为 $y-x=\lambda r$ 会更容易理解一些，因为 $x$ 相当于原始的偏置了）

• Definition 3
  A ray $r$ of $P$ is an extreme ray if there do not exist rays $r^1$, $r^2\in P^0$ and a scalar $\mu$ (with $r^1\ne r^2$ for any $\lambda >0$ and $0<\mu <1$) such that $r=\mu r^1+(1-\mu )r^2$.
  （解读：extreme ray 就是在无界可行集中，由两个超平面相交后得到的相交线）

• Proposition 2
  A vector $r$ is an extreme ray of $P$ if and only if { λ r : r ≥ 0 } \{\lambda r : r \geq0\} {λr:r≥0} is a one-dimensional face${}^{标注[2]}$ of $P^0$.
  （解读：性质2与定义3讲得是一个意思，所谓的 extreme ray 就是无界多面体的边）

3）最后描述三个定理：

• Theorem 1
  If max { c T x : x ∈ P } \{c^Tx:x\in P\} {cTx:x∈P} is finite then there is an optimal solution that is an extreme point.
  （解读：有界的可行集，线性规划的最优解一定取在 extreme point 上，这是显然的…）

• Theorem 2
  If max { c T x : x ∈ P } \{c^Tx:x\in P\} {cTx:x∈P} is unbounded then $P$ has an extreme ray $r^{\ast }$ with $c^T{r}^{\ast }>0$.

• Theorem 3
  The polyhedron $P$ can be represented as
  P = { x ∈ R n : x = ∑ i ∈ K λ k x k + ∑ j ∈ J μ j r j   w i t h   ∑ i ∈ K λ k = 1 , λ k ≥ 0   ∀ k ∈ K , μ j ≥ 0   ∀ j ∈ J } \begin{aligned} P=\{x\in \mathcal{R}^n:x=\sum_{i\in K} \lambda_kx^k+\sum_{j\in J}\mu_jr^j \ with \ \sum_{i\in K} \lambda_k=1, \lambda_k\geq0 \ \forall k \in K, \mu_j \geq 0 \ \forall j\in J \} \end{aligned} P={x∈Rn:x=i∈K∑​λk​xk+j∈J∑​μj​rj with i∈K∑​λk​=1,λk​≥0 ∀k∈K,μj​≥0 ∀j∈J}​
  where { x k } k ∈ K \{ x^k \}_{k\in K} {xk}k∈K​ is the set of extreme points of $P$ and { r j } j ∈ J \{r^j\}_{j\in J} {rj}j∈J​ is the set of extreme rays of $P$.
  This theorem is the basis for decomposition algorithms for linear programming. Note that if we drop the assumption about the rank of A, we get the following proposition:

• Proposition 3
  If $P\ne \varphi$ and $rank(A)=n-k$ then $P$ has a face of dimension $k$ and no proper face of lower dimension.

标注[1]： Zero-dimensional faces are known as polyhedron vertices (nodes)
标注[2]： One-dimensional faces are known as polyhedron edges
标注资料网址： Face-Wolfram MathWorld

更多优化内容，欢迎关注本人微信公众号：优化与博弈的数学原理