View on GitHub

memo

Linear Programming

Linear Programming

Theorem1 Farkas

Then following statemetns are selective. (i.e. If (1) holds, then (2) cannot hold. Vice versa if (2) hold, then (1) cannot hold)

(1)

\[\begin{eqnarray} \exists x \in \mathbb{R}^{n}, \ \text{ s.t. } \quad \langle c, x \rangle & > & 0, \nonumber \\ Ax & \le & 0, \nonumber \end{eqnarray}\]

(2)

\[\begin{eqnarray} \exists y \in \mathbb{R}^{m}, \ \text{ s.t. } \ & & A^{\mathrm{T}}y = c \nonumber \\ & & y \ge 0, \nonumber \end{eqnarray}\]

proof.

(1) \(\Rightarrow\) not (2)

Let

\[S := \{ x \in \mathbb{R}^{n} \mid x = A^{\mathrm{T}}y, \ y \ge 0 \} .\]

$S$ is closed convex set. Suppose that (2) does not hold. By separation theorem between a point and a set, there exist \(\alpha \in \mathbb{R}\) and \(c \in \mathbb{R}^{n}\) such that

\[\begin{eqnarray} & & \langle a, c \rangle > \alpha \nonumber \\ \forall x \in S, \quad & & \langle a, x \rangle \le \alpha . \end{eqnarray}\]

Since $0 \in S$, substituing $x=0$ into the equation above we have

\[\langle a, 0 \rangle = 0 \le \alpha\]

It follows that $\langle a, c \rangle > 0$. Next, for all $y \ge 0$,

\[\begin{eqnarray} \alpha & \ge & \langle a, x \rangle \nonumber \\ & \ge & \langle a, A^{\mathrm{T}}y \rangle \nonumber \\ & = & \langle Aa, y \rangle . \nonumber \end{eqnarray}\]

Then \(Aa \le 0\). Indeed, if \(Aa > 0\), there exists \(i \in \{1, \ldots, m\}\) such that \((Aa)_{i} > 0\). This is contradiction since

\[\forall y \in \mathbb{R}^{n}, \quad \langle y, Aa \rangle \le \alpha .\]

not (1) \(\Leftarrow\) (2)

Suppose there exists \(x \in \mathbb{R}^{n}\) such that \(Ax \le 0\). Then

\[\begin{eqnarray} \langle c, x \rangle & = & \langle A^{\mathrm{T}}y, x \rangle \nonumber \\ & = & \langle y, Ax \rangle \le 0 \quad (\because Ax \le 0,\ y \ge 0) . \nonumber \end{eqnarray}\]

Hence (1) cannot hold.

$\Box$

Remark2

farkas’s theorem is equivalent that not (1) if and only if (2), that is,

\[\begin{eqnarray} \forall x \in \mathbb{R}^{n}, \quad ( \langle c, x \rangle > 0 \Rightarrow Ax \ge 0 ) \text{ or } ( \langle c, x \rangle \le 0 \Rightarrow Ax < 0 ) \end{eqnarray}\]

Lemma2 Homogeneous Farkas Lemma

Then following statmeents are equivalent:

(1) Inequality system below has no solution in $\mathbb{R}^{n}$

\[\begin{eqnarray} \langle c, x \rangle & < & 0 \nonumber \\ \langle a^{i}, x \rangle & \ge & 0 \quad i = 1, \ldots, m \end{eqnarray}\]

(2)

\[\exists y \ge 0, \ \text{ s.t. } \ c = \sum_{i=1}^{m} y_{i}a^{i}\]

proof

\[A := \left( \begin{array}{c} a^{1} \\ \vdots \\ a^{m} \end{array} \right)\]

where \(a^{i} \in \mathbb{R}^{n}\) is $i$ th row vector.

\[Ax = \left( \begin{array}{c} (a^{1})^{\mathrm{T}}x \\ \vdots \\ (a^{m})^{\mathrm{T}}x \end{array} \right)\]

Then

\[A^{\mathrm{T}}y = \sum_{i=1}^{n} y^{i}a^{i}\] \[\langle c, x \rangle < 0 \Rightarrow \langle a^{j}, x \rangle < 0 \quad \langle a^{i}, x \rangle \ge 0 \Rightarrow \langle c, x \rangle \ge 0\]

Since $x \in \mathbb{R}^{n}$, (1) is equivalent that inequality system below has no solution

\[\begin{eqnarray} \langle c, x \rangle & > & 0 \nonumber \\ \langle a^{i}, x \rangle & \le & 0 \quad i = 1, \ldots, m \nonumber \end{eqnarray}\]
$\Box$

Definition 3

Primal LP problem is defined as

\[\begin{align} \min_{x \in \mathbb{R}^{n}} & & & c^{\mathrm{T}}x \label{mathmatical_programming_problem_lp_primal} \\ \mathrm{subject\ to} & & & Ax - b \ge 0 . \nonumber \end{align}\]

Dual of LP problem is defined as

\[\begin{align} \min_{y \in \mathbb{R}^{m}} & & & b^{\mathrm{T}}y \label{mathmatical_programming_problem_lp_dual} \\ \mathrm{subject\ to} & & & A^{\mathrm{T}}y - c \ge 0 \nonumber \\ & & & y \ge 0 \nonumber . \end{align}\]

Theorem

(1) Dual of dual probelm is a primal problem

(2) Let $y^{\prime}$ be a feasible solution of \(\eqref{mathmatical_programming_problem_lp_dual}\). Let $x^{\prime}$ be a feasible solution of \(\eqref{mathmatical_programming_problem_lp_primal}\).

\[\begin{eqnarray} b^{\mathrm{T}}y^{\prime} \le c^{\mathrm{T}}x^{\prime} \nonumber \end{eqnarray}\]

(3)

proof

$\Box$

Reference