0

Let $K_1,K_2$ be disjoint convex sets in $\mathbb C$. Let $z_1\in\partial K_1,z_2\in\partial K_2$ be the minimizers of $\mbox{dist}(K_1,K_2)=\inf{|x-y|}$ where $x\in K_1$ and $y\in K_2$. Is it true that we can pick the separating hyperplane in this case to be orthogonal to the line connecting $z_1$ and $z_2$?

chhro
  • 2,226
  • 8
  • 14

1 Answers1

0

Edit: never mind, your sets are disjoint, so they can be separated properly. Then you can use the theorem stating that: if $C,D$ are non-empty sets in $\mathbb{R}^n$ then there exists a hyperplane separating $C,D$ properly if and only if there exists a vector v such that: $$\inf\{x\cdot v, x\in C\}\geq \sup\{y\cdot v, y\in D\} \wedge \sup\{x\cdot v, x\in C\}>\inf\{y\cdot v, y\in D\}$$ Then the hyperplane separating them has $v$ as normal.

So yes you can pick it to be orthogonal to that line, but you can also pick a hyperplane that is not orthogonal and it still may separate them porperly as long as it fulfills the above requirements in the theorem (basically what you require is unnecessarily strong).

To see why your assumption is correct, you can use the projection theorem. Then $(z_2-z_1) \cdot (x-z_1)\leq 0,\forall x\in K_1$ and $(z_1-z_2) \cdot (y-z_2)\leq 0,\forall y\in K_2$. Thus if you pick any point $z = (1-\lambda)z_1+\lambda z_2$, and use $z_1-z_2$ as the hyperplane normal, then $(y-z)\cdot(z_1-z_2)\leq 0$ and $(x-z)\cdot(z_1-z_2)\geq 0$ and the inequality is strict for $\lambda \in (0,1)$ since the sets are disjoint.

Edit2: Fixed proper/strong separation mistake.

Edit3: Proof clarification: You have the minimizers $z_1 \in \partial K_1 ,z_2 \in \partial K_2$, but from their definition and the definition of projection, you have that the projection of $z_1$ onto $cl(K_2)$ is $z_2$ and the projection of $z_2$ onto $cl(K_1)$ is $z_1$. Applying the projection theorem to $z_1,z_2$ we get: $$(z_2-z_1)\cdot(x - z_1)\leq 0, \forall x \in cl(K_1)$$ $$(z_1-z_2)\cdot(y - z_2)\leq 0, \forall y \in cl(K_2)$$

Having these results we want to show that $(z_2-z_1)\cdot(x-z) \leq 0, \forall x \in cl(K_1)$: $$(z_2-z_1)\cdot(x-z) = (z_2-z_1)\cdot(x-(1-\lambda)z_1 - \lambda z_2) = $$ $$(z_2-z_1)\cdot(x-z_1) + \lambda(z_2-z_1)\cdot(z_1-z_2) = $$ $$(z_2-z_1)\cdot(x-z_1) - \lambda||z_2-z_1||^2 \leq 0 $$ Where I have used in the last line that $(z_2-z_1)\cdot(x-z_1) \leq 0$ and the norm being non-negative and $\lambda \geq 0$. Similarly $(z_1-z_2)\cdot(y-z) \leq 0, \forall y \in cl(K_2)$: $$(z_1-z_2)\cdot(y-z) = (z_1-z_2)\cdot(y-(1-\lambda)z_1 - \lambda z_2) =$$ $$(z_1-z_2)\cdot(y-z_2) + (1-\lambda)(z_1-z_2)\cdot(z_2-z_1) = $$ $$(z_1-z_2)\cdot(y-z_2) - (1-\lambda)||z_1-z_2||^2 \leq 0$$ Where I have used in the last line that $(z_1-z_2)\cdot(y-z_2) \leq 0$ and the norm being non-negative and $\lambda \geq 0$. Thus $(z_1-z_2)\cdot(y-z) \leq 0, \forall y \in cl(K_2)$ and $(z_2-z_1)\cdot(x-z) \leq 0, \forall x \in cl(K_1)$ or equivalently $(z_1-z_2)\cdot(x-z) \geq 0, \forall x \in cl(K_1)$. However these two conditions simply mean that the plane defined as $(z_1-z_2)\cdot w = (z_1-z_2)\cdot z$ separates $cl(K_1)$ and $cl(K_2)$. Since the plane shares a point with $cl(K_1)$ or $cl(K_2)$ only when $\lambda$ is zero or one, and the or is exclusive, then it separates the two sets properly.

lightxbulb
  • 2,047
  • Disjoint convex sets can't be separated strongly in general. Take K_1=upper region of the branch of 1/x in Quadrant 1 and K_2=quadrant 2.

    But I'm still looking at your response...

    – chhro Mar 05 '19 at 22:32
  • Ah ok, you need to have $0\ne cl(K_1-K_2)$. Then the theorem above holds for proper separation if you replace the first strict inequality with a non-strict one. Proper separation holds though since $ri(K_1)\cap ri(K_2) = \emptyset$. – lightxbulb Mar 05 '19 at 22:43
  • @chhro If you need extra clarification regarding the projection theorem just tell me which part is not clear and I'll explain it. – lightxbulb Mar 05 '19 at 22:45
  • Could you point me to references about these two results you invoked?

    Is $\wedge$ minimum here?

    – chhro Mar 06 '19 at 00:38
  • @chhro The wedge in the first result should be read as 'and', that is there are two conditions. I cited the first result just to show you that you don't need such a strong requirement like orthogonality. The part with the projection theorem is really the answer to your question. You can find both the projection theorem and the necessary and sufficient condition for proper separability in Rockafella's convex analysis. – lightxbulb Mar 06 '19 at 01:00
  • @lightbulb I will accept the answer while I try to understand your argument. Cheers! – chhro Mar 08 '19 at 16:20
  • @chhro You could just ask if something's unclear. I'll try to add an edit to clarify it a bit more, but I can't be sure what part you do not understand, so do clarify that. – lightxbulb Mar 08 '19 at 17:01