6

Let $X$ be a Banach space, and denote by $B_r (x)$ the closed ball of radius $r > 0$ around $x \in X$. Furthermore, let $A \subset X$ be compact and $N \in \Bbb{N}$. I am interested in "optimally" covering $A$ by $N$ balls, i.e,. with as small radius as possible. More precisely, define $$ r_0 := \inf \Big\{ r > 0 \colon \exists \, x_1,\dots,x_N \in X \text{ such that } A \subset \bigcup_{i=1}^N B_r (x_i) \Big\}, $$ and assume that $r_0 > 0$.

I would like to know whether there necessarily exist $x_1,\dots,x_N \in X$ such that $A \subset \bigcup_{i=1}^N B_{r_0} (x_i)$. In other words, I would like to know if the infimum above is actually a minimum.

Note I am taking the $x_i$ from the "surrounding" space $X$, not from the compact set $A$.

I can prove the claim in case that $X$ is reflexive (even only assuming that $A$ is bounded), but I am not sure whether it is true for more general Banach spaces. I will give my proof for the reflexive case below, in case one can either generalize it, or use it to get an idea for a counterexample.



Proof for the reflexive case: Choose a sequence $r_n \to r_0$ such that for each $n$ there are $x_1^n,\dots,x_N^n \subset X$ satisfying $A \subset \bigcup_{i=1}^N B_{r_n}(x_i^n)$. If $B_{r_n} (x_i^n) \cap A = \emptyset$ for some $i,n$, replace $x_i^n$ by zero. Note that this retains the property $A \subset \bigcup_{i=1}^N B_{r_n}(x_i^n)$.

Since $A$ and the sequence $(r_n)_{n}$ are bounded, there is $R > 0$ such that $\| x \| \leq R$ and $r_n \leq R$ for all $x \in A$ and $n \in \Bbb{N}$: There are now two cases for each $i,n$: 1) There is some $x \in A \cap B_{r_n} (x_i^n)$, and hence $\| x_i^n \| \leq \| x_i^n - x \| + \| x \| \leq r_n + R \leq 2R$. 2) There is no $x \in A \cap B_{r_n} (x_i^n)$, and hence $x_i^n$, whence $\| x_i^n \| \leq 2R$.

Therefore, each of the sequences $(x_i^n)_{n \in \Bbb{N}} \subset X$ is bounded. Since $X$ is assumed to be reflexive, we can choose a common subsequence (which I will ignore in the notation below) such that $x_i^n \to x_i$ weakly for all $i = 1,\dots,N$.

Now, let $x \in A$ be arbitrary. For each $n \in \Bbb{N}$, there is $i_n \in \{1,\dots,N\}$ satisfying $\| x - x_i^n \| \leq r_n$. Next, there is $\ell \in \{1,\dots,N\}$ such that $i_n = \ell$ for infinitely many $n \in \Bbb{N}$, say for $n = n_m$ with $n_m \to \infty$. Since $x - x_i^{n_m} \to x - x_i$ weakly and since the norm is lower semicontinuous with respect to weak convergence, we see that $\| x - x_\ell \| \leq \liminf_{m \to \infty} \| x - x_i^{n_m} \| \leq \liminf_{m \to \infty} r_{n_m} = r_0$. Since this holds for any $x \in A$, we get $A \subset \bigcup_{\ell=1}^N B_{r_0} (x_\ell)$, as desired.

PhoemueX
  • 36,211
  • 2
    Have you figured out $N=1$? – mathworker21 Jul 17 '19 at 20:22
  • @mathworker21: No, even that would be interesting. – PhoemueX Jul 18 '19 at 19:13
  • @mathworker21 In any metric space $\text{diam}A$, the diameter of $A:=\sup { d(a_1,a_2) \mid a_1, a_2 \in A }$, is well defined. In a complete metric space I believe this $\sup$ will be achieved. If so, in the case $N=1$, the optimal $r$ should be $r_0=\text{diam}A / 2$. – Selrach Dunbar Jul 20 '19 at 09:22
  • @Selrach: I don't think this is actually true; see https://math.stackexchange.com/questions/2734808/smallest-ball-to-contain-a-subset-of-diameter-d-in-mathbbrn – PhoemueX Jul 20 '19 at 10:21
  • @PhoemueX Aha, so Jung's Theorem gives us an upper bound on $r_0$. That means for $A$, a compact subset of $\mathbb{R}^n$, $\text{diam}A/2$ is just a lower bound. Combining the bounds we have: $$ \text{diam}A \cdot\frac{1}{2}\leq r_0 \leq \text{diam}A\sqrt{\frac{n}{2(n+1)}}$$ [Reason for the lower bound is $A \subseteq B_{r_0}$ which implies $ \text{diam}A \leq \text{diam}B_{r_0}=2r_0 $.] – Selrach Dunbar Jul 22 '19 at 12:50
  • @PhoemueX doesn't my answer below answer your question? – mathworker21 Jan 16 '20 at 17:44
  • 1
    @mathworker: Sorry, forgot to click that button. Thanks again for your answer! – PhoemueX Jan 16 '20 at 17:46

2 Answers2

4

I don't feel too bad posting a "proof by references" since I worked on this problem for quite a while but failed, before realizing the $N=1$ case seems pretty google-able. In any event, it seems Konyagin proved that any non-reflexive Banach space admits an equivalent norm and some 3 points so that there is no smallest ball with respect to the equivalent norm containing those 3 points. A good buzzword is "Chebyshev center". Below is a link to a strengthening of Konyagin's result (and a mention of Konyagin's result). I couldn't find a free version of Konyagin's paper online.

https://dml.cz/bitstream/handle/10338.dmlcz/119230/CommentatMathUnivCarolRetro_42-2001-1_11.pdf

mathworker21
  • 35,247
  • 1
  • 34
  • 88
0

Unfortunately, the infimum above is not actually a minimum. To see this let $A$ be any two point set in $\mathbb{R}^2$. Then $r_0=0$ whenever $N \geq 2$.