Prev Next Top DualIneq

Duality With Inequality Constraints

Primal Problem
Given  A \in \R^{m \times n} ,  b \in \R^m , and  c \in \R^n , the corresponding primal problem is  \[
\begin{array}{ll}
{\rm maximize}      & c^T x  \; {\rm with \; respect \; to} \; x \in \R_+^n
\\
{\rm subject \; to} & A x \leq b
\end{array}
\] 
The extended primal objective function  f : \R_+^n \rightarrow \R \cup \{ - \infty \} is defined by  \[
f(x) = \left\{ \begin{array}{ll}
     c^T x     & {\rm if} \; b - A x \geq 0 \\
     - \infty  & {\rm otherwise}
\end{array} \right. 
\] 
Solving the primal problem is equivalent to finding an argument  x that maximizes the value of  f(x) with respect to  x \in \R_+^n .

Lagrangian
The Lagrangian corresponding to the problem above  L : \R_+^n \times \R_+^m \rightarrow \R is defined by  \[
\begin{array}{rcl}
L(x, y) 
& = & c^T x + y^T ( b - A x ) 
\\
& = & b^T y + x^T ( c - A^T y ) 
\end{array}
\] 
It follows that  \[
     f(x) = \inf \; L(x, y) \; {\rm with \; respect \; to} \; y \in \R_+^m
\] 


Dual Problem
The extended dual objective function  g : \R_+^m \rightarrow \R \cup \{ + \infty \} is defined by  \[
     g(y) = \sup \; L(x, y) \; {\rm with \; respect \; to} \; x \in \R_+^n
\] 
It follows that  \[
g(y) = \left\{ \begin{array}{ll}
     b^T y     & {\rm if} \; c - A^T y \leq 0 \\
     + \infty  & {\rm otherwise}
\end{array} \right. 
\] 
The dual problem is to find an argument value  y that minimizes  g(y) with respect to  y \in \R_+^m . This is equivalent to the following problem:  \[
\begin{array}{ll}
{\rm minimize}      & b^T y  \; {\rm with \; respect \; to} \; y \in \R_+^m
\\
{\rm subject \; to} & A^T y \geq c
\end{array}
\] 


Duality Gap Lemma
Given  \hat{x} \in \R_+^n and  \hat{y} \in \R_+^m , the corresponding duality gap is  g( \hat{y} ) - f( \hat{x} ) . The duality gap is always non-negative and if it is zero,  \hat{x} solves the primal problem and  \hat{y} solves the dual problem.

Proof
 \[
\begin{array}{rcl}
g( \hat{y} ) 
& = & 
\sup L( x , \hat{y} ) \; {\rm with \; respect \; to} \; x \in \R_+^n 
\\
& \geq &
L( \hat{x} , \hat{y} ) 
\\
& \geq &
\inf \; L( \hat{x} , y ) \; {\rm with \; respect \; to} \; y \in \R_+^m 
\\
& = & 
f( \hat{x} )
\end{array}
\] 
Thus we conclude that the duality gap is non-negative. Further if it equals zero then for all  x \in \R_+^n and  y \in \R_+^m  \[
\begin{array}{rcccl}
g( y )       & \geq & f( \hat{x} ) & =    & g( \hat{y} )
\\
f( \hat{x} ) & =    & g( \hat{y} ) & \geq & f( x ) 
\end{array}
\] 
which completes the proof of the lemma.
Input File: DualIneq.omh