Dantzig-Wolfe Decomposition

Some important concepts

A polyhedron is a set that can be described in the form { x ∈ R n ∣ A x ≥ b } \{\mathcal{x}\in \mathcal{R^n}|\bm{Ax\ge b}\} {x∈Rn∣Ax≥b}, where $\mathbf{A}$ is an $m\times n$ matrix and $\mathbf{b}$ is a vector in ${\mathcal{R}}^m$.

Let $P$ be a polyhedron. A vector $\mathbf{x}\in P$ is an extreme point of P if we cannot find two vectors $\mathbf{y},\mathbf{z}\in P$, both different from $\mathbf{x}$, and a scalar $\lambda \in [0,1]$, such that $\mathbf{x}=\lambda \mathbf{y}+(1-\lambda )\mathbf{z}$.

A nonzero element $\mathbf{x}$ of a polyhedron cone $C\subset {\mathcal{R}}^n$ is called an extreme ray if there are n-1 linearly independent constraints that are active at $\mathbf{x}$.

Representation of polyhedra

This is one of the fundamental results of linear programming theory.

Theorem (Resolution theorem) Let
P = { x ∈ R n ∣ A x ≥ b } P=\{\mathcal{x}\in \mathcal{R^n}|\bm{Ax\ge b}\} P={x∈Rn∣Ax≥b}
be a nonempty polyhedron with at least one extreme point. Let ${\mathbf{x}}^1,...,{\mathbf{x}}^{\mathbf{k}}$ be the extreme points, and let ${\mathbf{w}}^1,...,{\mathbf{w}}^{\mathbf{r}}$ be a complete set of extreme rays of $P$. Let
Q = { ∑ i = 1 k λ i x i + ∑ j = 1 r θ j w j ∣ λ i ≥ 0 , θ j ≥ 0 , ∑ i = 1 k λ i = 1 } . Q=\{\sum_{i=1}^{k}\lambda_i\bm{x^i}+\sum_{j=1}^r\theta_j\bm{w^j}|\lambda_i\ge0,\theta_j\ge0,\sum_{i=1}^{k}\lambda_i=1\}. Q={i=1∑k​λi​xi+j=1∑r​θj​wj∣λi​≥0,θj​≥0,i=1∑k​λi​=1}.
Then, $Q=P$.

Dantzig-Wolfe decomposition

详见 Introduction to Linear Optimization (Dimitris Bertsimas, John N. Tsitsiklis, 1997) 中的 Dantzig-Wolfe decomposition。