-1

Let $D = [c,d]\times [c,d]\subseteq \mathbb{R}^2$ and let $A$ be the set of all closed subsets of $D$. For $a \in D$ and $B\in A,$ define $d(a,B) := \min\{d(a,b) | b\in B\},$ where the $d$ inside the min is the Euclidean metric on $\mathbb{R}^2$. This is defined as closed and bounded subsets of $\mathbb{R}^2$ are compact and the function $b\mapsto d(a,b)$ is continuous so the image of $D$ under this map attains its minimum value. For $B,C \in A,$ let $d_A(B,C) = \max\{\max_{b\in B} d(b,C), \max_{c\in C} d(c,B)\}.$

  1. Show that $d$ satisfies the triangle inequality.
  2. Show that $(A,d_A)$ is complete.

To show that $d$ satisfies the triangle inequality, let $A_0, B_0,C_0\in A.$ Suppose without loss of generality that $d_A(A_0,C_0) = \max_{a\in A_0} d_A(a,C_0)$. For any $a_0\in A_0, b_0\in B_0, c_0\in C_0, d(a_0,c_0) \leq d(a_0,b_0) + d(b_0, c_0)\Rightarrow d(a_0, C_0) \leq d(a_0,b_0) + d(b_0, C_0).$ Since $b_0$ is arbitrary, this implies $d(a_0, C_0)\leq \min_{b_0\in B_0} (d(a_0,b_0) + d(b_0,C_0)) \leq d(a_0, B_0) + d(b,C_0),$ where $b$ is such that $d(a_0,b) = \min_{b_0\in B_0} d(a_0,b_0)$. Then $d_A(A_0,C_0) = \max_{a_0\in A_0} d(a_0, C_0)\leq \max_{a_0\in A_0} d(a_0, B_0) + d(b,C_0)\leq d_A(A_0, B_0) + d_A(b,C_0).$

However, I'm unsure how to show that $(A,d_A)$ is complete. My first step was to let $(B_n)$ be a Cauchy sequence in $A.$ Then for any $\epsilon > 0,$ for sufficiently large $m$ and $n, d_A(B_n, B_m) < \epsilon.$ I'm not sure how to use the definition of $d_A$ to come up with a closed subset $C$ of $D$ that one can show is the limit of the sequence $(B_n).$

Is the proof that $d_A$ satisfies the triangle inequality correct?

user3472
  • 1,195
  • I assume what you mean is $d_A(B,C) = \max{\max_{b\in B} d(b,C), \max_{c\in C} d(c,B)}$. Correct? Also the notation is not really clear, since $a$ and $b$ used in the definition of $D$ have nothing to do with $a$ and $b$ used in the following definitions of distances. – dfnu Jul 20 '21 at 11:16
  • Your question is about completeness of Hausdorff distance. It was asked and answered many times at MSE in much greater generality: https://math.stackexchange.com/questions/2493757/hausdorff-distance-prove-that-if-e-d-is-complete-then-mathcalke-m, https://math.stackexchange.com/questions/1309827/completeness-of-the-hausdorff-distance,... – Moishe Kohan Jul 20 '21 at 22:43
  • I would look either at the accepted answer here or at Theorem $3$-$3$ here; both arguments are pretty straightforward. – Brian M. Scott Jul 24 '21 at 00:09

1 Answers1

0

Here is a later review of my previous answer, as promised (the original, left in the bottom, had a flaw, as you pointed out in comment).

I know from the comments there is already a clear reference to other answers, but I did not want to leave here an incomplete (accepted and rewarded) answer.


I will need the two following Lemmas.

Lemma 1. If $C_1 \supset C_2$ are elements of $\mathcal A$, then $d_{\mathcal A}(C_1,C_2) = \max_{x \in C_1}[\min_{y \in C_2}(x,y)]$.

Proof. $\forall y \in C_2$ we have that $\min_{x \in C_1}(x,y) = 0$. Therefore $\max_{y\in C_2}[\min_{x\in C_1}(x,y)]=0$. $\square$


Lemma 2. If $C_1 \supset C_2\supset \cdots$ is a nested sequence of elements of $\mathcal A$, and $$C= \bigcap_{n=1}^\infty C_n,$$then the sequence $(C_n)$ converges to $C$, that is $$\lim_{n\to \infty} d_{\mathcal A}(C_n,C) = 0.$$

Proof. (Note that $C$ is not empty.) Suppose, by contradiction, that there exists $\varepsilon >0$ such that, for all $n>0$, we have $d_A(C_{p_n},C)> \varepsilon$, for some $p_n>n$. By Lemma 1, this means that there exists an element $c_{p_n} \in C_{p_n}$ such that $\min_{y \in C}d(c_{p_n},y) > \varepsilon$. This implies that $$d(c_{p_n},y)> \varepsilon, \ \ \forall y\in C.$$ Since the sequence $(c_{p_n})$ is bounded, it must have a converging subsequence. Let us call it $(\gamma_n)$. Suppose $(\gamma_n) \to \gamma$. What we have stated guarantees that $\gamma \not \in C$, that is $$\gamma \in U =\overline{\bigcap_{n=N}^\infty C_n}$$ for some $N>0$. $U$ is an open set. Thus there exists an open neighborhood of $\gamma$ entirely contained in $U$. This, however, implies that $(\gamma_n)\not \to \gamma$, a contradiction. $\square$


Now we have what we need to prove that $(\mathcal A, d_{\mathcal A})$ is complete.

Consider a Cauchy sequence $(B_n)$. We want to construct a subsequence that is convergent. This will imply convergence of $(B_n)$.

At this aim, let $B_{p_1}$ an element of the sequence such that $$d_{\mathcal A}(B_{p_1},B_m) < \frac12, \ \ \forall m> p_1,$$ and let $$\tilde C_1 = \mbox{cl}\left(\bigcup_{x\in B_{p_1}}N_{\frac12}(x)\right),$$ where $\mbox{cl}(\cdot)$ denotes closure and $N_r(x) = \{y \in D : d(x,y) < r\}$ is the open neighbourhood of $x$ with radius $r$. Observe that $$B_m \subset \tilde C_1, \ \ \forall m > p_1.$$

Now choose $p_2 > p_1$ in such a way that $$d_{\mathcal A}(B_{p_2},B_m)< \frac14, \ \ \forall m>p_2, $$ and, correspondingly $$\tilde C_2 = \mbox{cl}\left(\bigcup_{x\in B_{p_2}}N_{\frac14}(x)\right).$$ We have now that $$B_m \subset \tilde C_1 \cap \tilde C_2, \ \ \forall m>p_2.$$

Proceed iteratively like that for $n=1,2,3,\dots$. Therefore, select $B_{p_n}$ so that $p_n>p_{n-1}$ and $$d_A(B_{p_n},B_m)< \frac1{2^n}, \ \ \forall m>p_n$$ and construct closed $$\tilde C_n = \mbox{cl}\left(\bigcup_{x\in B_{p_n}}N_{\frac1{2^n}}(x)\right),$$ so that $$B_m \subset C_n = \bigcap_{k=1}^n \tilde C_n, \ \ \forall m>p_n.$$ Observe that, by construction, $$d_{\mathcal A}(B_{p_n},C_n) \leq \frac1{2^n},$$ and therefore, by triangle inequality, $$d_{\mathcal A}(B_{p_m},C_n) \leq d_{\mathcal A}(B_{p_m},B_{p_n}) + d_{\mathcal A}(B_{p_n},C_n) < \frac1{2^{n-1}}, \ \ \forall p_m> p_n.$$

Recall also that, by Lemma 2, $$\lim_{n\to \infty} d_{\mathcal A}(C_n,C) = 0,$$ where $$C = \bigcap_{n=1}^\infty C_n.$$

We want to show that $(B_{p_n})$ converges to $C$. Fix $\varepsilon > 0$. For $k> N_1$, $$d_{\mathcal A}(C_k,C)< \frac{\varepsilon}{2}.$$ For $k>2-\log_2\varepsilon$, $$d_{\mathcal A}(B_{p_n},C_k) < \frac{\varepsilon}{2}, \ \ \forall p_n > k.$$ Thus $$d_{\mathcal A}(B_{p_n},C) \leq d_{\mathcal A}(B_{p_n},C_k) + d_{\mathcal A}(C_k,C) < \varepsilon,$$ for $p_n > \max(N_1,2-\log_2\varepsilon)$. Since $(B_{p_n})$ converges to $C$ and it is a subsequence of a Cauchy sequence, it follows that $(B_n)$ converges also to $C$. The thesis follows. $\square$


ORIGINAL ANSWER.

I think your proof is correct.

As for the completeness of $(A,d_A)$, I would proceed as follows. I am pretty sure there are simpler proofs. Have a look; and let us see, in the meantime, what the community suggests.

Consider, as you said, a Cauchy sequence $(B_n)$. We want to construct a subsequence that is convergent. This will imply convergence of $(B_n)$.

At this aim, let $B_{p_1}$ an element of the sequence such that $$d_A(B_{p_1},B_m) < \frac12, \ \ \forall m> p_1,$$ and let $$\tilde C_1 = \mbox{cl}\left(\bigcup_{x\in B_{p_1}}N_{\frac12}(x)\right),$$ where $\mbox{cl}(\cdot)$ denotes closure and $N_r(x) = \{y \in D : d(x,y) < r\}$ is the open neighbourhood of $x$ with radius $r$. Observe that $$B_m \subset \tilde C_1, \ \ \forall m > p_1.$$

Now choose $p_2 > p_1$ in such a way that $$d_A(B_{p_2},B_m)< \frac14, \ \ \forall m>p_2, $$ and, correspondingly $$\tilde C_2 = \mbox{cl}\left(\bigcup_{x\in B_{p_2}}N_{\frac14}(x)\right).$$ We have now that $$B_m \subset \tilde C_1 \cap \tilde C_2, \ \ \forall m>p_2.$$

Proceed iteratively like that for $n=1,2,3,\dots$. Therefore, select $B_{p_n}$ so that $p_n>p_{n-1}$ and $$d_A(B_{p_n},B_m)< \frac1{2^n}, \ \ \forall m>p_n$$ and construct closed $$\tilde C_n = \mbox{cl}\left(\bigcup_{x\in B_{p_n}}N_{\frac1{2^n}}(x)\right),$$ so that $$B_m \subset C_n = \bigcap_{k=1}^n \tilde C_n, \ \ \forall m>p_n.$$ Observe that, by construction, $$d_A(B_{p_n},C_n) \leq \frac1{2^n},$$ and, by triangle inequality, $$d_A(B_{p_m},C_n) \leq d_A(B_{p_m},B_{p_n}) + d_A(B_{p_n},C_n) < \frac1{2^{n-1}}, \ \ \forall p_m> p_n\tag{1}\label{tri}$$ Since the sets $C_n$ are closed (and therefore compact) and $C_1 \supset C_2 \supset C_3 \dots$, we have that $$C = \bigcap_{i=1}^\infty C_i \neq \emptyset.$$

We want now to show that the sequence $(B_{p_n})$ converges to $C$. Suppose, by contradiction, that there exists $\varepsilon >0$ such that, for all $N>0$, there is $p_m> N$ for which $d_A(B_{p_m},C)>\varepsilon$, implying, by definition, $\max_{c \in C} [\min_{b\in B_{p_m}}(b,c)]> \varepsilon$. This means that there exists $c\in C$ such that $\min_{b\in B_{p_m}}d(b,c)>\varepsilon$. Select $n$ large enough that $\varepsilon > \frac1{2^{n-1}}$. We have that $c \in C_n$, and, thus $$d_A(B_{p_m},C_n) = \max_{c\in C_n}[\min_{b\in B_{p_m}}(b,c)]>\varepsilon > \frac1{2^{n-1}}.$$ Recall that there exists such $p_m >N$ for arbitray choice of $N$. So taking $N>p_n$ contradicts \eqref{tri}. Hence $$(B_{p_n}) \to C,$$ and, since $(B_n)$ is Cauchy, $$(B_n)\to C.$$

dfnu
  • 7,528