5

Let $M$ be a metric space and let $\mathscr B$ be a family of open balls in $M$ whose radii are bounded. Assuming that $\mathscr B$ is totally ordered by inclusion, is the union of all members of $\mathscr B$ an open ball?

I believe this is false for the rational numbers, but what if $M$ is complete? Does it hold in every normed space? Every Banach space? What about $\mathbb R^n$?


My attempt at the case $M=\mathbb R^n$:

Let $R$ be the set of real numbers formed by the radii of the balls in $\mathfrak{B}$ and let $r=\sup R$. If $r\in R$ then the answer is obvious, so let us suppose otherwise.

One may then find an increasing sequence $\{r_i\}_i$ in $R$ converging to $r$. The centers $c_i$ of the corresponding balls $B_i$ must form a converging sequence (this seems awful obvious, but might require a rather lengthy proof) so let $c$ be the limit.

I guess it should also be clear that $\bigcup \mathfrak B = \bigcup_i B_i$, and it seems reasonable to try to prove that this set coincides with the open ball $B_r(c)$, centered at $c$, with radius $r$. Regarding the inclusion $$ \bigcup_i B_i\subseteq B_r(c), $$ let $x$ belong to the left-hand-side set, so there exists some $i_0$ such that $x\in B_{i_0}$ and, since the balls are increasing, this should also hold for all $i>i_0$. In other words $$ \|x-c_{i}\|<r_{i},\quad \forall i\geq i_0. $$ Taking the limit as $i\to\infty$, one concludes that $$ \|x-c_{i_0}\|\leq r_{i_0}. $$ Among many, this is one of the main outstanding points! How to get "$<$" instead of "$\leq $"???

Black
  • 247
  • 1
  • 7
  • why do you think it is false for the rational numbers? do you have a counterexample? – Jackozee Hakkiuz Oct 04 '20 at 04:52
  • Totally ordered and increasing are different. Please make sure that the title matches your question. – Kavi Rama Murthy Oct 04 '20 at 06:00
  • 4
    @JackozeeHakkiuz $\mathbb Q\cap(\pi,3\pi)$ is the union of a chain of open balls in $\mathbb Q$. – bof Oct 04 '20 at 06:50
  • 1
    @Kavi, I am aware of the different meanings but (1) "increasing" is shorter and more intuitively appealing, (2) "totally ordered" is technically more precise and is the property I am interested in, and finally (3) it does not affect the substance of the question. – Black Oct 04 '20 at 12:37
  • @Bof, I got you on the solution to my attempt. Thanks – Black Oct 05 '20 at 12:38
  • @Bof, here is another philosophical comment: any analyst is well prepared to deal with lots of balls at the same time, but as long as the centers are fixed! – Black Oct 05 '20 at 12:44
  • @Bof, the trouble of course is to show that this is not a ball! – Black Oct 06 '20 at 15:15
  • @Bof, you are right about non complete spaces. See my answer below. – Black Oct 07 '20 at 14:17

3 Answers3

5

To answer your first question, it's not necessarily true in a complete metric space that the union of a chain of open balls is a ball. Here is a counterexample.

Let $M=\{a_i:i\in\mathbb N\}\cup\{b_i:i\in\mathbb N\}\cup\{c\}$ with the following metric:
$d(a_i,a_j)=1$ if $i\ne j$;
$d(b_i,b_j)=2$ if $i\ne j$;
$d(a_i,b_j)=1$ if $j\le i$;
$d(a_i,b_j)=2$ if $j\gt i$;
$d(a_i,c)=d(b_i,c)=2$.

The triangle inequality holds, since all nonzero distances are $1$ or $2$.

The metric is complete, since every Cauchy sequence is eventually constant.

Let $\mathscr B=\{B_n:n\in\mathbb N\}$ where
$B_n=\{x\in M:d(a_n,x)\le1\}=\{x\in M:d(a_n,x)\lt2\}=\{a_i:i\in\mathbb N\}\cup\{b_i:i\le n\}$.

S0 $\mathscr B$ is a chain of open balls and a chain of closed balls. The union $\bigcup\mathscr B=M\setminus\{c\}$ is not a ball because, for each point $x\ne c$, there is a point $y\ne c$ such that $d(x,y)=d(x,c)=2$.


Regarding your other questions. I'm going to guess that it's true for Banach spaces, false for incomplete normed spaces.

bof
  • 78,265
  • Thanks @bof! Any ideas, or at least guesses, for the other cases? – Black Oct 04 '20 at 12:39
  • Well, that is a very conservative guess :-) I agree that it should be true there and also that the center of the ball should be the limit of the centers of the balls in the chain. However I got in trouble proving this due to the fact that the key inequality is obtained thru a limit, and hence comes up as "$\leq$" rather than "$<$". – Black Oct 04 '20 at 14:16
  • That is an excellent idea but, in order to concentrate efforts, I instead edited the question with a description of my failed attempt. – Black Oct 04 '20 at 15:05
  • Done. I opted to unaccept your question after realizing that the loss of 15 reputation points will not be a huge blow in percentage terms! – Black Oct 05 '20 at 12:37
2

The answer is affirmative for Banach spaces:

Theorem. If $S$ is a nonempty bounded open set in a Banach space such that, for any positive number $d\lt\operatorname{diam}(S)$, the set $S$ contains a ball of diameter $d$, then $S$ is an open ball.

Proof. Let $d=\operatorname{diam}(S)$ and $r=\frac12d$. Choose a sequence $B_1,B_2,B_3,\dots$ of open balls $B_n\subseteq S$ such that $\operatorname{diam}(B_n)\to d$. Let $d_n=\operatorname{diam}(B_n)$, let $r_n=\frac12d_n$, and let $c_n$ be the center of $B_n$, so that $B_n=B_{r_n}(c_n)$. Note that $$r_m+\|c_m-c_n\|+r_n\le d,$$ i.e., $$\|c_m-c_n\|\le d-r_m-r_n\le\max\{d-d_m,d-d_n\}.$$ Hence $c_1,c_2,c_3,\dots$ is a Cauchy sequence and converges to a point $c$. I claim that $S=B_r(c)$.

Claim 1. $B_r(c)\subseteq S$.

Proof. Suppose $x\in B_r(c)$, so $\|x-c\|=r-\varepsilon\lt r$. Choose $n$ so that $\|c_n-c\|\lt\frac\varepsilon2$ and $r_n\gt r-\frac\varepsilon2$. Then $$\|x-c_n\|\le\|x-c\|+\|c-c_n\|\lt(r-\varepsilon)+\frac\varepsilon2=r-\frac\varepsilon2\lt r_n,$$ so $x\in B_n\subseteq S$.

Claim 2. $S\subseteq B_r(c)$.

Proof. Assume for a contradiction that $x\in S$ and $\|x-c\|\ge r$. Since $S$ is open, there is a point $y\in S$ with $\|y-c\|=r+\varepsilon\gt r$. Choose a point $z\in B_r(c)$, antipodal to $y$, with $\|z-c\|=r-\frac\varepsilon2$, so that $$\|y-z\|=\|y-c\|+\|z-c\|=(r+\varepsilon)+(r-\frac\varepsilon2)=d+\frac\varepsilon2\gt d.$$ So $\|y-z\|\gt d$. Since $y\in S$, and $z\in B_r(c)\subseteq S$ by Claim 1, this contradicts the fact that $\operatorname{diam}(S)=d$.

bof
  • 78,265
  • Very smart! So the key was that $r_m+|c_m-c_n|+r_n\le d$! I feel much better now after returning your deserved 15 rep points! – Black Oct 06 '20 at 15:14
1

Complementing the excellent answers by @bof, here is the final case:

Theorem. Let $X$ be a normed space. Then the following are equivalent:

i) $X$ is complete,

ii) For every totally ordered family $\mathscr B$ of open balls in $X$ with uniformly bounded radii, the union of the members of $\mathscr B$ is an open ball.

Proof (i) $\Rightarrow$ (ii) was already proven by @bof in their accepted answer.

(ii) $\Rightarrow$ (i): Assuming (ii), and arguing by contradiction, suppose that $X$ is not complete. Denoting by $\tilde X$ the completion of $X$, let $a$ be a point of $\tilde X$ which is not in $X$. Choose a sequence $\{c_n\}_{n=1}^\infty $ in $X$ such that $\|c_n-a\|<1/2^{n+1}$, so that $c_n\to a$, as $n\to\infty$.

It follows that $$ \|c_n-c_{n+1}\| \leq \|c_n-a\| + \|a-c_{n+1}\| \leq {1\over 2^{n+1}} + {1\over 2^{n+2}} < {1\over 2^n}. \tag 1 $$

Define $$ r_n = \sum_{k=0}^{n-1} {1\over 2^k}, \quad \text{and} \quad r = \sum_{k=0}^\infty {1\over 2^k} $$ (of course we could spell out the explicit values of $r_n$ and $r$, but the above expressions will turn out to be more convenient for us) and observe that $$ r_{n+1} = r_n + {1\over 2^n}, \tag 2 $$ and that $r_n\to r$, as $n\to\infty $. Setting $$ \tilde B_n=B^{\tilde X}_{r_n}(c_n), \quad \text{and}\quad B_n=B^X_{r_n}(c_n), $$ (where the superscript indicates the normed space under consideration for the purpose of defining a ball), it is clear that $\tilde B_n\cap X=B_n$, and we claim that $$ \tilde B_n\subseteq \tilde B_{n+1}, \tag 3 $$ for every $n$. In fact, given any $y$ in $\tilde B_n$, we have that $$ \|y-c_{n+1}\| \leq \|y-c_n\| + \|c_n-c_{n+1}\| < r_n + {1\over 2^n} = r_{n+1}, $$ by (1) and (2), thus proving the claim. It is not hard to prove that $$ \bigcup_{n=1}^\infty \tilde B_n = B^{\tilde X}_{r}(a), $$ so, if both sets above are intercepted with $X$, we deduce that $$ \bigcup_{n=1}^\infty B_n = B^{\tilde X}_{r}(a)\cap X. $$ Observing that $B_n\subseteq B_{n+1}$ by (3), our assumption (ii) implies that $B^{\tilde X}_{r}(a)\cap X$ is an open ball in $X$, but beware that $a$ is not its center because $a$ is not even in $X$!

Let us therefore write $$ B^{\tilde X}_{r}(a)\cap X = B^{X}_{r}(b), $$ for some $b$ in $X$, where we have retained the radius $r$ because the radius of a ball is half its diameter, and it is clear that the diameter of $B^{\tilde X}_{r}(a)\cap X$ is $2r$.

It follows that $B^{X}_{r_n}(c_n) = B_n \subseteq B^{X}_{r}(b)$ so, by @bof's Lemma (see below), we have $$ \|c_n-b\|\leq r-r_n \to 0, $$ as $n\to\infty$, so we see that $$ b=\lim_{n\to\infty } c_n = a, $$ contradicting the assumption that $a$ is not in $X$. This concludes the proof.


Lemma (@bof) Let $X$ be a normed space, and pick two elements $c$ and $d$ in $X$, as well as two positive real numbers $r$ and $s$. Assuming that $B_r(c)\subseteq B_s(d)$, one has that $r+\|c-d\|\leq s$.

Proof. This is obvious in case $c=d$. Otherwise note that, for every $t$ in the half-open interval $[0,r)$, one has that $$ x:= c+{t\over\|c-d\|}(c-d)\in B_r(c). $$ By assumption $x\in B_s(d)$, so $$ s>\|x-d\| = \left\|c-d+{t\over\|c-d\|}(c-d)\right\| = \|c-d\| + t. $$ Therefore $$ s\geq \lim_{t\to r_-} \|c-d\| + t = \|c-d\| + r, $$ concluding the proof.

Black
  • 247
  • 1
  • 7