5
$\begingroup$

I am trying to derive the conditions under which $(K\cap L)^+\subseteq K^+ + L^+$ where $K^+$ (and $L^+$) denotes the dual cone of convex cone $K$ (and $L$), and $K+L$ denotes the Minkowski sum of $K$ and $L$. (I've read somewhere that the condition is $K\cap \textrm{int} L\neq \emptyset.$)

To start off, suppose $x\in(K\cap L)^+ $. Then, $\langle\ x, \phi\rangle \geq 0,\ \forall \phi\in K\cap L$. In order for $x$ to be in $K^+ + L^+$, it should be decomposable into $x=x_1+x_2$ such that $x_1\in K^+,\ x_2\in L^+$. One way I can think of to find such a decomposition is to project $x $ on to $K^+$ and $L^+$ and (since dual cones are cones themselves) use appropriate scalars on the respective projections to get $x_1$ and $x_2$ such that their sum is equal to $x$. For these projections to add up nicely to $x$, I'm thinking that:

  1. either $x\in K^+ \cap L^+$ or
  2. if $ x\notin K^+ \cap L^+$, $x$ should make an acute angle with the closest boundaries/faces of the dual cones.

Where am I going wrong?

$\endgroup$
2
+50
$\begingroup$

It is straightforward that $P^+\cap Q^+ \subset (P+Q)^+$. Conversely if $x \in (P+Q)^+$, assuming that $P$ and $Q$ are non-empty cones you have that $0\in Cl(P)$ and $0 \in Cl(Q)$ whence $x\in P^+$ and $x\in Q^+$.

So we have $P^+\cap Q^+=(P+Q)^+$ for any two non-empty cones $P, Q$.

Taking the dual cone we get that:

$$Cl(P+Q) = (P^+ \cap Q^+)^+$$ so if you let $P=K^+$ and $Q=L^+$ (which are two closed clones) and using the fact that the bidual is the closure you get that: $$K^+ + L^+ = (Cl(K) \cap Cl(L))^+ $$

So the question you are left with is: can the dual cone of $Cl(K)\cap Cl(L)$ be smaller than that of $K\cap L$?

It can be the case for instance if $K$ is a half-line on the boundary of $L$ (open), in which case $K\cap L=\emptyset$ and $Cl(K)\cap Cl(L)=K$.

A sufficient condition is obviously that $K$ and $L$ are closed.

I don't think the condition $K\cap Int(L) \not= \emptyset$ is necessary, for instance you could have $K$ and $L$ two distinct half-lines - they are closed but their interior is empty.

A necessary condition is that $Cl(K)\cap Cl(L)=Cl(K\cap L)$, using again the fact that the bidual is the closure. Because the dual of $K\cap L$ and that of its closure are the same, this condition is also sufficient. Not sure we can make this condition more explicit.

$\endgroup$
  • $\begingroup$ What is Adh(P)? $\endgroup$ – Teodorism Jul 15 at 16:43
  • 1
    $\begingroup$ I mean the closure, sorry for the frenchism. Will edit. $\endgroup$ – FXV Jul 15 at 16:45
1
$\begingroup$

This is only a partial answer.

First, the claim is obviously true if $K\subset L$, which implies $L^+\subset K^+$, and $(K\cap L)^+ = K^+ = K^++L^+$.

Second, the claim is true if $K\cap L=\{0\}$ and $K,L$ are closed convex cones. Assume the claim is not true. Then there is $f \not\in K^++L^+$. We can separate $f$ from $K^++L^+$, i.e., there exists non-zero $x$ such that $$ f^Tx \le k'^Tx + l'^Tx \quad \forall k'\in K^+, l'\in L^+. $$ Since $K^+$ and $L^+$ are cones, it follows $k'^Tx\ge 0$ and $l'^Tx\ge0$ for all $k'\in K^+$, $l'\in L^+$. Hence, $x\in K^{++} \cap L^{++}=K\cap L = \{0\}$, contradiction.

I am not sure how to 'interpolate' between those extreme cases. The separation argument is going nowhere if $K\cap L \ne\{0\}$. (It yields $f^Tx=0$)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.