Ref:
- “Convex Optimization” by Boyd, Vandenberghe.
- “Convex Optimization” by Boyd. Lecture-2. Link to lecture: Lecture.
[Dual cones]
Note that cones induce inequalities. Link to Stackexchange post: Link.
Consider ${ \mathbb{R} ^n . }$ Let ${ K \subseteq \mathbb{R} ^n }$ be a cone.
Consider the dual space
\[{ (\mathbb{R} ^n)^{\ast} := \lbrace x ^T (\cdot) : x \in \mathbb{R} ^n \rbrace . }\]Note that ${ (\mathbb{R} ^n) ^{\ast} }$ has the norm
\[{ \lVert x ^T (\cdot) \rVert := \sup \lbrace x ^T (y) : \lVert y \rVert = 1 \rbrace . }\]Note that
\[{ \lVert x ^T (\cdot) \rVert = \lVert x \rVert . }\]Q) What are all the elements of ${ (\mathbb{R} ^n) ^{\ast} }$ which respect ${ \preceq _K }$?
Equivalently
Q) What is
\[{ \lbrace \lambda ^T (\cdot) : x {\preceq _K} y \implies \lambda ^T (x) \leq \lambda ^T (y) \rbrace . }\]Note that the set is
\[{ {\begin{aligned} &\, \lbrace \lambda ^T (\cdot) : x {\preceq _K} y \implies \lambda ^T (x) \leq \lambda ^T (y) \rbrace \\ = &\, \lbrace \lambda ^T (\cdot) : \lambda ^T (y - x) \geq 0 \text{ for all } y - x \in K \rbrace \\ = &\, \lbrace \lambda ^T (\cdot) : \lambda ^T z \geq 0 \text{ for all } z \in K \rbrace . \end{aligned}} }\]We call
\[{ \boxed{ K ^{\ast} = \lbrace \lambda : \lambda ^T z \geq 0 \text{ for all } z \in K \rbrace } }\]as the dual cone of ${ K . }$
Note that geometrically, ${ \lambda \in K ^{\ast} }$ means ${ (-\lambda) }$ is a normal vector of a supporting hyperplane of ${ K }$ at ${ 0 . }$
Thm: Consider ${ \mathbb{R} ^n . }$ Let ${ K \subseteq \mathbb{R} ^n }$ be a proper cone. Then
\[{ x {\preceq _K} y \iff \lambda ^T (x) \leq \lambda ^T (y) \, \, \text{ for all } \lambda \in K ^{\ast} . }\]Pf: ${ \underline{\implies} }$ Note that it is true by definition of ${ K ^{\ast} . }$
${ \underline{\impliedby} }$ Suppose
\[{ \lambda ^T (y - x) \geq 0 \quad \text{ for all } \lambda \in K ^{\ast} . }\]Hence
\[{ (y - x) \in K ^{\ast \ast} . }\]Note that by the below theorem
\[{ K ^{\ast \ast} = K . }\]Hence
\[{ (y - x) \in K }\]as needed. ${ \blacksquare }$
[${ K ^{\ast \ast} = K }$]
Thm: Let ${ K \subseteq \mathbb{R} ^n }$ be a proper cone. Then
\[{ K ^{\ast \ast} = K . }\]Pf: Note that ${ K }$ is the intersection of all homogeneous half spaces containing ${ K . }$
Hence
\[{ {\begin{aligned} &\, K \\ = &\, \bigcap _{\lambda \in K ^{\ast}} \lbrace x : (- \lambda) ^T x \leq 0 \rbrace \\ = &\, \bigcap _{\lambda \in K ^{\ast}} \lbrace x : \lambda ^T x \geq 0 \rbrace \\ = &\, \lbrace x : \lambda ^T x \geq 0 \text{ for all } \lambda \in K ^{\ast} \rbrace \\ = &\, K ^{\ast \ast} . \end{aligned}} }\]Hence
\[{ K ^{\ast \ast} = K }\]as needed. ${ \blacksquare }$