Prev Next Top DualTheorem

The Duality Theorem

Notation
Let  A \in \R^{m \times n} ,  b \in \R^m , and  c \in \R^n define the primal problem with extended objective  f(x) and the dual problem with extended objective  g(y) .

Tableau Form

Original
We define  Q \in \R^{m \times m} to be the identity matrix and the slack vector  s \in \R_m are defined by  \[
     b =  A x + Q s 
\] 
The scalar  e \in \R and the coefficient vector  y \in \R^m are defined by  e = 0 ,  y = 0 . The primal problem has the following tableau form:  \[
\begin{array}{llclcl}
{\rm maximize}      & z
\\
{\rm subject \; to} & b = A x   & + & Q s
\\
                    & e = c^T x & - & y^T s     & - & z
\\
                    & x \geq 0  & , & s \geq 0
\end{array}
\] 


Final
The Simplex method is used to obtain the following equivalent tableau:  \[
\begin{array}{llclcl}
{\rm maximize}      & z
\\
{\rm subject \; to} & \hat{b} = \hat{A} x   & + & \hat{Q} s
\\
                    & \hat{e} = \hat{c}^T x & - & \hat{y}^T s & - & z
\\
                    & x \geq 0  & , & s \geq 0
\end{array}
\] 
In other words,  x \in \R^n ,  s \in \R^m , and  z \in \R satisfies the constraints in the tableau above if and only if they satisfy the constraints in the original tableau.

Theorem
Suppose that  \hat{c} \leq 0 ,  \hat{y} \geq 0 , and there is a feasible pair  ( \hat{x} , \hat{s} ) such that  \[
0 = \hat{c}^T \hat{x}  -  \hat{y}^T \hat{s}
\] 
It follows that  \hat{x} solves the primal problem and  \hat{y} solves the dual problem.

Proof
We have the equivalence of the two tableaus and  Q is the identity matrix and  e = 0 . Thus for all  x \in \R_+^n , by setting  s = b - A x we have  \[
\begin{array}{rcl}
\sum_{j=1}^n c_j x_j 
& = &  
- \hat{e} + \sum_{j=1}^n \hat{c}_j x_j - \sum_{i=1}^m \hat{y}_i s_i
\\
& = &  
- \hat{e} 
+ \sum_{j=1}^n \hat{c}_j x_j 
- \sum_{i=1}^m \hat{y}_i \left( b_i - \sum_{j=1}^n A_{i,j} x_j \right) 
\\
\sum_{i=1}^m b_i \hat{y}_i  + \hat{e}
& = &
\sum_{j=1}^n \left( 
     - c_j + \hat{c}_j  +  \sum_{i=1}^m \hat{y}_i A_{i,j} 
\right) x_j
\hspace{1cm} (\star)
\end{array}
\] 
From equation  (\star) , choosing  x = 0 , we obtain  \[
\begin{array}{rcl}
\sum_{i=1}^m b_i \hat{y}_i  + \hat{e} & = & 0
\\
\sum_{i=1}^m b_i \hat{y}_i            & = &  - \hat{e} 
\end{array}
\] 
Form equation  (\star) , choosing  x = 0 except for the j-th component which is one, we obtain the following equation for  j = 1 , \ldots , n :  \[
\begin{array}{rcl}
0   & = &  - c_j + \hat{c}_j  +  \sum_{i=1}^m \hat{y}_i A_{i,j}
\\
c_j & \leq &  \sum_{i=1}^m \hat{y}_i A_{i,j}
\end{array}
\] 
Note that we used  \hat{c}_j \leq 0 to obtain the inequality above. It follows that  \hat{y} is dual feasible,  \hat{x} is primal feasible, and  \[
f( \hat{x} ) = 
- \hat{e} + \hat{c}^T \hat{x}  -  \hat{y}^T \hat{s}
=
- \hat{e}
=
b^T \hat{y}
=
g( \hat{y} )
\] 
The conclusion of the theorem now follows from the Duality Gap Lemma .
Input File: DualTheorem.omh