Commit d889bd1e authored by Niklas Rieken's avatar Niklas Rieken

up to complementary slackness, MnSymbol(?)

parent b44dd2e7
No preview for this file type
......@@ -9,6 +9,7 @@
......@@ -783,5 +784,136 @@ Next, we look at another example for the case when there is no feasible solution
Now observe that there are no positive reduced cost w.r.t. $M$ but the artificial variable $x_5$ is still in the optimal basis. Hence, the original problem is infeasible. If there would be a feasible solution then there exists basic solution with basic variables drawn from $\{x_1, x_2, x_3, x_4\}$ with objective value $0$ w.r.t. $M$, i.e. basic solutions with artificial variables are never optimal.
Now we make a sharp break and move on to a very enlightning concept of mathematical optimization: \emph{duality}. The motivation is to find an estimate on how large the optimum of an LP (that is to be maximized) can be, i.e. find a (strong) \emph{upper bound}. Note that fo \emph{lower bounds}, we can evaluate any feasible solution. We use the constraints to find such upper bounds. Every \emph{cone combination} (i.e. a linear combination with only non-negative coefficients) of constraints yields to an upper bound if the sum of these coefficients (also called \emph{Lagrange multipliers}) are at least as large as the objective coefficients for each variable.
Consider the following LP.
\text{maximize } & 2x_1 + 4x_2 - 3x_3\\
\text{subject to } & x_1 + 3x_2 - x_3 \leq 4\\
& x_1 + 2x_2 - 2x_3 \leq 3\\
& x_1, x_2, x_3 \geq 0.
From the first constraint, we can derive an upper bound on theobjective function:
z(x) = 2x_1 + 4x_2 - 3x_3 \leq 2(x_1 + 3x_2 - x_3) \leq 2 \cdot 4 = 8.
If we use both constraints, we can obtain an even tighter bound:
z(x) = 2x_1 + 4x_2 - 3x_3 \leq (x_1 + 3x_2 - x_3) + (x_1 + 2x_2 - 2x_3) \leq 4 + 3 = 7.
If we want to find the best upper bound, consider the following:
z(x) = 2x_1 + 4x_2 - 3x_3 &\leq \lambda_1(x_1 + 3x_2 - x_3) + \lambda_2(x_1 + 2x_2 - 2x_3)\\
&\leq (\lambda_1 + \lambda_2)x_1 + (3\lambda_1 + 2\lambda_2)x_2 + (-\lambda_1 - 2\lambda_2)x_3\\
&\leq 4\lambda_1 + 3\lambda_2.
This chain of inequalities hold if $\lambda_1 + \lambda_2 \geq 2 (= c_1)$, $3\lambda_1 + 2\lambda_2 \geq 4 (= c_2)$, $-\lambda_1 - 2\lambda_2 \geq -3 (= c_3)$, and of course $\lambda_1, \lambda_2 \geq 0$. We obtain the best (i.e. smallest) upper bound by minimizing $4\lambda_1 + 3\lambda_2$ subject to these four constraints.
Note that the minimization problem from the example is again an LP, which we call the \emph{dual LP} of the LP we started with (\emph{primal LP}). More formally we have
\item[primal LP] $\max \{c^Tx \mid \underbrace{Ax \leq b, x \geq 0}_{\eqqcolon P}\}$
\item[dual LP] $\min \{\lambda^Tb \mid \underbrace{\lambda^TA \geq c^T, \lambda \geq 0}_{\eqqcolon D}\}$
By construction of the dual LP, we have the following property.
\begin{theorem}[weak duality]
Let $P, D$ denote the polyhedra of a primal and a corresponding dual LP.
\item If $x \in P$ and $\lambda \in D$, then $c^Tx \leq \lambda^Tb$.
\item If $P \neq \emptyset$, $c^Tx$ unbounded from above iff $D = \emptyset$.
\item If $D \neq \emptyset$, $\lambda^Tb$ unbounded from below iff $P = \emptyset$.
\item $c^Tx \leq \lambda^TAx \leq \lambda^Tb$.\qedhere
For each primal constraint, there is a dual variable; for each primal variable, there is a dual constraint. We do not need that the primal LP is in canonical form. By the following rules, we can dualize every LP.
primal & dual\\
maximize & minimize\\
objective coeffiecients & right hand side\\
right hand side & objective coefficients\\
$A$ & $A^T$\\
$a_i^T x \leq b$ & $\lambda_i \geq 0$\\
$a_i^T x = b$ & $\lambda_i$ free\\
$a_i^T x \geq b$ & $\lambda_i \leq 0$\\
$x_j \geq 0$ & $\lambda^TA_j \geq c_j$\\
$x_j$ free & $\lambda^TA_j = c_j$\\
$x_j \leq 0$ & $\lambda^TA_j \leq c_j$\\
In fact, we can also read the table above from right to left to dualize minimization problem. This yields to the following theorem.
The dual of a dual is the primal.
What makes duality interesting is that every dual solution is a certificate for optimality (or rather maximal gap to optimality). Especially if we have $x^\ast \in P, \lambda^\ast \in D$ with $c^Tx^\ast = \lambda^{\ast T}b$, then $x^\ast$ is optimal for the primal LP and $\lambda^\ast$ is optimal for the dual LP. Note that the concept of dual problems is not restricted to linear optimization; it can be applied to arbitrary non-linear problems, too. However, the following theorem is special and not true in arbitrary non-linear settings but always holds for linear programs.
\begin{theorem}[strong duality]
If the primal LP has a finite optimal solution $x^\ast$, then the dual LP has a finite optimal solution $\lambda^\ast$. Moreover,
c^Tx^\ast = \lambda^{\ast T}b.
Let $B$ be an optimal basis for the primal LP, i.e. $x_B$ solves $A_B x_B = b \iff x_B = A_B^{-1}b$. Since $B$ is optimal, we have that the reduced cost are non-positive:
\bar{c}_N^T = c_N^T - \underbrace{c_B^T A_B^{-1}}_{\eqqcolon \lambda^T}A_N \leq 0
We claim that $\lambda^T$ is optimal dual solution. For feasibility just note that $\lambda^T A \geq c^T$ because of the reduced cost. Moreover,
c_B^T x_B = c_B^T A_B^{-1}b = \lambda^T b,
i.e. the objective values match.
From duality we derive
\item primal optimality implies dual feasibility,
\item primal feasibility implies dual optimaliyt.
Note that dual variables $\lambda^T = c_B^T A_B^{-1}$ are implicitly computed in the Simplex algorithm: Reduced cost $\bar{c}_N^T = c_N^T - \lambda^TA_N$. For slack variables $x_{n+i}$ is $c_{n+i} = 0$ and $A_{n+i}$ is the $i$-th unit vector.
There is also a very common economic interpretation of dual variables; ofthen they are called \emph{shadow prices} or \emph{opportunity costs}. A constraint often models the availability of some resource $i$. That is, given a primal LP
\text{maximize } & c^T x\\
\text{subject to } & Ax \leq b + \epsilon\\
& x \geq 0,
where we consider a right hand side that is \emph{relaxed} by some $\epsilon \geq 0$, the interpretation of a dual variable $\lambda_i$ for the $i$-th constraint is the possible increment of the objective value if one more unit of the $i$-th resource is available.
\lambda^T(b+\epsilon) = \lambda^Tb + \lambda^T\epsilon.
\begin{theorem}[complementary slackness]
A primal solution $x$ and a dual solution $\lambda$ are optimal iff
\item $a_i^T x = b_i$ or $\lambda_i = 0$ for all $i \in \{1, \ldots, m\}$. \label{thm:cs:1}
\item $\lambda^TA_j = c_j$ or $x_j = 0$ for all $j \in \{1, \ldots, n\}$. \label{thm:cs:2}
By strong duality: $c^T x = \lambda^TAx = \lambda^T b$. Then we have by the first equation:
(\underbrace{c^T - \lambda^TA}_{\geq 0}) \underbrace{x}_{\geq 0} = 0
providing the proof for \ref{thm:cs:2} and for \ref{thm:cs:1}, we use the second equation for the analogous argument
\underbrace{\lambda^T}_{\geq 0} (\underbrace{Ax - b}_{\geq 0}) = 0.\qedhere
We now want to analyze soundness and runtime of the Simplex algorithm. In the worst case the algorithm still needs an exponential number (in $m$) of iterations. Note that there are also algorithm that guarantee a runtime in polynomial time for solving LPs but it turns out that the Simplex algorithm is better in practice where it usually visits $\mathcal{O}(m)$ vertices.
Markdown is supported
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment