4

I'm in a linear programming class and I'm trying to understand why the unit circle with its interior is not a polyhedron.

I know there is a proof by contradiction that the unit circle not including its interior is not a polyhedron. But I'm wondering why if we include the interior, the set is still not a polyhedron. What is the intuition behind why such a circle is not a polyhedron? Does it have to do with the fact that a circle has an infinite amount of extreme points? The set in consideration is actually $$\{(x,y) \in \mathbb{R}^2 ; x^2+y^2 \leq 1, x \geq 0, y \geq 0\},$$ and my definition of polyhedron is a set which is the intersection of finitely many half-planes.

forreal
  • 53
  • Question has been edited – forreal Sep 27 '19 at 05:02
  • 1
    A polyhedron has faces that are polygons, so has to be 3 dimensional, which a disk is not. And a polygon has to have edges that are straight lines, which a disk does not have. – Heath Winning Sep 27 '19 at 05:09
  • A bounded polyhedron is the convex hull of a finite set; a finite subset of the plane has a conventional $n$-sided polygon as its convex hull, not a disc. – Angina Seng Sep 27 '19 at 05:16
  • This has been discussed on many other threads. Here are a few of the best. For general discussion: https://math.stackexchange.com/q/31785 For the formal difficulties of going down the limit path: https://math.stackexchange.com/a/1164990 For the most generous interpretation, using nonstandard analysis: https://math.stackexchange.com/a/1644477 – Chris Culter Sep 27 '19 at 05:23
  • Hi @LordSharktheUnknown is there a way to think of this in terms of linear programming and extreme points. Does a circle having an infinite extreme points imply that its not a polyhedron? Are all sets with infinite extreme points not polyhedra? Is there a theorem about this? – forreal Sep 27 '19 at 05:27
  • I want to understand this in terms of optimization and linear programming where a polyhedron is defined as the intersection of a finite number of half spaces. – forreal Sep 27 '19 at 05:30
  • Can you prove that in 2D, an extreme point of a polyhedron must lie on the boundary of two half-spaces? Then if you have finely many half-spaces, how many intersections can there be at most? –  Sep 27 '19 at 06:18
  • 1
    The unit disk cannot be expressed as the intersection of a finite number of half spaces. – copper.hat Sep 27 '19 at 06:36
  • @ copper disk, Unit disk cannot be expressed as intersection of a finite number of half spaces, because there are infinite number of extreme points? I want to understand this intuitively why it cannot be intersection of a finite number of half spaces (I feel like I don't have good geometric intuition). – forreal Sep 27 '19 at 06:44
  • There can be two intersections at most? @Rahul, not sure if I follow, – forreal Sep 27 '19 at 06:45
  • One reason is that the boundary of the disc has nonzero curvature. The boundary of the intersection of finitely many half planes either has zero curvature or undefined curvature (where the half planes intersect). – Eric Sep 27 '19 at 08:21

3 Answers3

3

Let me use the standard terminology disc to refer to a circle union its interior. Also, let me use the standard terminology polygon to refer to a polyhedron in the plane.

Each circle is the boundary of its corresponding disc, and that boundary has the following special relation with the center point of the disc:

  1. Given a disc, its center is equidistant from all points on its boundary.

On the other hand,

  1. Given a polygon, there does not exist any point in the plane which is equidistant from all points on its boundary.

Once you are convinced of the truth of 2, then it is clear that a circle is not a polygon, because they do not satisfy the same geometric properties.

To prove 2, one may apply the fact that the boundary of a polygon is a union of finitely many line segments, such that any two of those segments are either disjoint or intersect only at a common endpoint.

Applying that fact, all that remains in order to prove 2 is to prove that for any given line segment, there does not exist any point in the plane which is equidistant from all points on that line segment. I will leave this final proof as an exercise that is easily solved with coordinate geometry (and easily solved with axiomatic Euclidean geometry as well).

Lee Mosher
  • 120,280
  • How can I prove the statement "The boundary of a polygon is a union of finitely many line segments, such that any two of those segments are either disjoint or intersect only at a common endpoint"? I can only show that the boundary is the union of finitely segments, I don't know what to do to have the rest condition. – RopuToran Mar 25 '21 at 14:06
  • That could perhaps be an interesting question on its own, if you could write it carefully so as to include the definition of polygon that you are using. But putting new questions in comments to answers to old questions, where the only person likely to see it is the answerer (me), is a poor use of this site. – Lee Mosher Mar 25 '21 at 14:33
2

Presumably you are using the Euclidean norm. (The unit ball with the $l_1, l_\infty$ norms are polyhedral.)

The Euclidean norm is strictly convex, so if $a \neq b$ then $\|ta+(1-t)b\|^2 < t \|a\|^2+(1-t)\|b\|^2$ for $t \in (0,1)$.

In particular, if $\|a\| \le 1, \|b\| \le 1$ and $a \neq b$, then $\|ta+(1-t)b\|^2 < 1$ for $t\in (0,1)$.

Suppose $P$ is a polyhedron and $P \subset \bar{B}$, the closed unit ball. Then $\bar{B} \setminus P$ is non empty and so we cannot express $\bar{B}$ as a polyhedron. To see this, consider the following:

Any compact polyhedron can be written as the convex hull of a finite number of points, so we can write $P = \operatorname{co} \{ x_1, \cdots , x_m\}$. We can assume that none of the points can be written as a convex combination of the remaining points. We must have $\|x_k\| \le 1$.

Any point $x \in P \setminus \{ x_1, \cdots , x_m\}$ can be written as $x = \sum_k t_k x_k$, with $t_k < 1$ and $\sum_k t_k = 1$.

Then $x = t_1 x_1 + (1-t_1) \sum_{k \neq 1} {t_k \over 1-t_1} x_k$ and since $\|x_1\| \le 1$ and $\| \sum_{k \neq 1} {t_k \over 1-t_1} x_k \| \le 1$, we see that $\|x\| < 1$.

In particular, the only points in $P$ that have norm one are $x_1,...,x_p$. Hence $P$ cannot equal $\bar{B}$.

copper.hat
  • 172,524
  • This invokes the fact that a bounded set which is the intersection of finitely many half-planes is the convex hull of a finite number of points. – Paul Frost Sep 28 '19 at 09:57
  • You need closure as well. That is why I mentioned compact above. – copper.hat Sep 28 '19 at 14:00
  • @copper.hat: why when $\bar{B}\backslash P$ is not empty we can say $\bar{B}$ is not a polyhedron? What is the mathematical condition for "We can assume that none of the points can be written as a convex combination of the remaining points."? Why any point in $P\backslash{x_1,\dots,x_m}$ has positive coefficient? – Saeed Jan 22 '22 at 05:22
  • @Sepide I had written $t_k>0$ when I should have had $t_k <1$. The mathematical condition is that the $x_k$ are extreme points of $P$. – copper.hat Jan 22 '22 at 05:29
  • @copper.hat: Can we prove it this way? pick finitely may points on the unite circle. Then the convex hull of these points is a polyhedron that contains only points that are with norm less than one and the ones with norm one are the ones we have already picked. But we know there are infinitely many points on the unit circle. – Saeed Jan 22 '22 at 05:33
  • @Sepide That is more or less what I have done above. If $P$ is a polyhedron contained in the closed unit ball then at most a finite number of points of $P$ lie on the boundary of the unit ball. – copper.hat Jan 22 '22 at 05:53
  • It is %99 your proof but if we pick points on the boundary of unit circle, how can we make sure none of the points can be written as a convex combination of the remaining points? And another question, $t_k<1$ just excludes $x_i$'s from the polyhedron, right? – Saeed Jan 22 '22 at 06:06
  • @Sepide I am not sure what you are asking. If a point lies in $P$ and is not one of the $x_k$ then it has norm $<1$. That is what the above is all about. All points in $P$ lie in the convex hull of the $x_k$. – copper.hat Jan 22 '22 at 06:08
0

We shall a give a proof which is (hopefully) intuitive but can be made precise if desired.

The "unit circle with its interior" is the set $D = \{(x,y) \in \mathbb{R}^2 ; x^2+y^2 \leq 1 \}$. Its boundary is the unit circle $C = \{(x,y) \in \mathbb{R}^2 ; x^2+y^2 = 1 \}$.

For each $c \in \mathbb R^2 \setminus \{ 0 \}$ let $L'(c)$ be the line line through $0$ and $c$, and $L(c)$ denote the line through $c$ which is perpendicular to $L'(c)$. By $H(c) $ we denote the half-plane which has $L(c)$ as its boundary and contains $0$. Clearly each half-plane $H$ with $0 \in H$ whose boundary line $L$ does not contain $0$ has the form $H = H(c)$ for a unique $c$. In fact, let $L'$ be the line through $0$ which is perpendicular to $L$ and let $c$ be the intersection point of $L$ and $L'$. Then $H = H(c)$.

Assume that $D$ is the intersection $I(c_1,\ldots,c_n)$ of finitely many half-planes $H(c_1),\ldots, H(c_n)$.

W.l.o.g. we may assume that all $c_i \in C$. In fact, the line $L'(c_i)$ intersects $C$ in a point $c^*_i$ between $0$ and $c_i$ ($c^*_i = c_i$ is possible). Then we have $D \subset H(c^*_i) \subset H(c_i)$ so that $I(c_1,\ldots,c_n) = I(c^*_1,\ldots,c^*_n)$. Note that the boundary line $L(c^*_i)$ of $H(c^*_i)$ is the tangent to $C$ at $c^*_i$.

If we add finitely points $c_i \in C$, $i= n+1,\ldots,m$, then clearly $I(c_1,\ldots,c_n) = I(c_1,\ldots,c_m)$. We may therefore w.l.o.g. assume that the points $(0,1)$ and $(1,0)$ belong to the $c_i$ and that the $c_i$ are pairwise distinct.

W.l.o.g. we may assume that $c_1 = (0,1)$ and $c_2$ is the first point reached if we travel counterclockwise along $C$ starting at $c_1$. The point $c_2$ lies on the quarter circle between $(1,0)$ and $(0,1)$. The lines $L_1 = L(c_1)$ and $L(c_2)$ intersect in a point $p$ lying on the line segment between $(1,0)$ and $(1,1)$. Clearly, $p \in H(c_1) \cap H(c_2)$, but $p \notin D$.

Any half-plane $H(c)$ where $c \in C$ does not belong to the arc between $c_1$ and $c_2$ contains $p$. This is intuitively clear: When we move counterclockwise along $C$ from $c_2$ to $(-1,0)$, then the intersection point $p_c$ of $L_1$ and $L(c)$ moves on $L_1$ from $p$ to $\infty$ and the half-line $L_1^-(p_c) \subset L_1)$ below $p_c$ (which has the property $p_c \in L_1^-(p_c)$) is contained in $H(c)$. When we reach $(-1,0)$, then $L_1$ and $L(-1,0)$ are parallel (i.e. do not intersect) and trivially $p \in H(-1,0)$. When we move further from $(-1,0)$ to $(1,0)$, the intersection point $p_c$ moves on $L_1$ from $-\infty$ to $(1,0)$ and the half-line $L_1^+(p_c) \subset L_1$ above $p_c$ (which has the property $p \in L_1^+(p_c)$) is contained in $H(c)$.

You should draw a picture - everything can be made absolutely precise if desired.

This shows that $p \in \bigcap_{i=1}^n H(c_i)$ which is a contradiction.

Paul Frost
  • 76,394
  • 12
  • 43
  • 125