Compactness

Equivalent and non-equivalent notions of compactness, and their proofs.

Definition. A topological space is compact (C) if every open cover has a finite subcover. It is Lindelöf (L) if every open cover has a countable subcover. It is countably compact (CC) if every countable open cover has a finite subcover.

Proposition 1. \: C     \iff (CC and L)

Proof. Evident. \quad\square

Definition. A topological space is FIP if every family of closed sets with the finite intersection property has a nonempty intersection.

Proposition 2. \: C     \iff FIP

Proof. Any {Fi}\{F_i\} family of closed sets is dual to {Vi}\{V_i\} by Vi=FiCV_i = F_i^C. {Vi}\{V_i\} is a cover iff Fi=\bigcap F_i = \emptyset and {Vi}\{V_i\} has a finite subcover iff {Fi}\{F_i\} does not have the finite intersection property. \quad\square

Definition. XX is CM if for any YY, the projection π2:X×YY\pi_2 : X \times Y \to Y is a closed map.

Proposition 3. \: C     \iff CM

Proof. (    \implies) Let FX×YF \subseteq X \times Y closed and yπ2(F)y \notin \pi_2(F). Then FCF^C contains the tube X×{y}X \times \{y\}. By the tube lemma, there is some neighborhood VV of yy such that X×VFCX \times V \subseteq F^C, and Vπ2(F)=V \cap \pi_2(F) = \emptyset. So π2(F)C\pi_2(F)^C is open. (    \impliedby) If XX not compact, take {Un}\{U_n\} a countable cover with no finite subcover. Let Y={0}{1/n:nN}RY = \{0\} \cup \{1/n : n \in \N\} \subseteq R. Let Vn=Y[0,1/n]V_n = Y \cap [0, 1/n] so that {Vn}\{V_n\} is a countable basis for 0 and V=nN(Un×Vn)X×YV = \bigcup_{n \in \N} (U_n \times V_n) \subseteq X \times Y, is open. Choose A=π2(VC)A = \pi_2(V^C). Clearly 0A0 \notin A yet any neighborhood WW of 0 contains {1/n:nN}\{1/n : n \ge N\}. If WA=W \cap A = \emptyset, then n=1NUn\bigcup_{n=1}^N U_n would be a cover of XX, so 0A0 \in \overline{A}. \quad\square

Definition. A directed set II is any set with a preorder (partial order without antisymmetry) such that any finite set has an upper bound (in particular, for a,bIa,b \in I, there is some cIc \in I with aca \le c and bcb \le c). A net {xi}iI\{x_i\}_{i \in I} is any family indexed by a directed set. A subnet {xf(i)}\{x_{f(i)}\} is a net defined by f:JIf : J \to I which is order-preserving and cofinal (for any iIi \in I, there’s a jJj \in J with if(j)i \le f(j)). A net is eventually in SS if for some iIi \in I, any upper bound of ii is in SS. A net converges to a point xx, and we write xixx_i \to x if it is eventually in VV for every neighborhood VV of xx. We call xx the limit of the net. A net is cofinally in SS if for all ii, there’s a jj with iji \le j and jSj \in S. A point xx is said to be a cluster point if the net is cofinally in VV for all VV neighborhoods of xx.

Proposition 4. \: C     \implies Every net has a cluster point     \implies Every net has a converging subnet     \implies C (uses A.C.)

Proof. (1st     \implies) Suppose {xi}iI\{x_i\}_{i \in I} has no cluster point. Let V(x)={V neighborhood of x : iIji  xjV}\mathcal{V}(x) = \{ V \text{ neighborhood of } x \ : \ \exists i \in I \forall j \ge i \ \ x_j \notin V \}. Since no xx is a cluster point, V(x)\mathcal{V}(x) \ne \emptyset. Then V=xXV(x)\mathcal{V} = \bigcup_{x \in X} \mathcal{V}(x) is an open cover. Take {Vn}V\{V_n\} \subseteq \mathcal{V} finite subcover. For each nn, there is some ini_n for which jin    xVnj \ge i_n \implies x \notin V_n. Take any upper bound jj of {in}\{i_n\}.

(2nd     \implies) Let xx be cluster point of {xi}iI\{x_i\}_{i \in I} and Vx\mathcal{V}_x a basis for xx. Then D={(i,V)I×Vx : xiV}D = \{ (i, V) \in I \times \mathcal{V}_x \ : \ x_i \in V \} is a directed set if we define (i,U)(j,V)    (ij and VU)(i, U) \le (j, V) \iff (i \ge j \text{ and } V \supseteq U). Let π:DI\pi : D \to I be the projection to II. It is order-preserving by definition and cofinal since xx is cluster. Finally, it it follows that xf(i,V)xx_{f(i, V)} \to x.

(3rd     \implies) Let {Vi}iI\{V_i\}_{i \in I} be an open cover. Define F={AI : A is finite}F = \{ A \subseteq I \ : \ A \text{ is finite}\}, this is a directed set. Contrapositively, if it has no finite subcover, take a choice function x:AXx : A \to X where xAiAVix_A \notin \bigcup_{i \in A} V_i. The net {xA}\{x_A\} cannot have a cluster point: If ViV_i is a neighborhood of xx, xAVix_A \notin V_i for any A{i}A \supseteq \{i\}. \quad\square

Definition. A point pp is said to be a cluster point of a sequence (xn)(x_n) if for any given nbd VV of pp, there are infinitely many nn of which xnVx_n \in V.

Note that every sequence is itself, a net, as the natural numbers are a directed set. The cluster point notion of nets is the same as above when applied to sequences. Similarly, a subsequence is defined by a function f:NNf : \N \to \N which is strictly increasing (as a consequence of being order-preserving and cofinal). A sequence converging to a point is the same notion to convergence of nets when applied to sequences. In a way, sequences are a special case of nets, which has a lot more structure but is also a lot more restricting. In general compact sets, they bring non-equivalent notions.

Lemma 1. If a sequence has a converging subsequence, then it has a cluster point.

Proof. Let xf(n)px_{f(n)} \to p. For any VV nbd of pp, there exists n0n_0 such that xf(n)Vx_{f(n)} \in V for nn0n \ge n_0, and there are infinitely many such nn. So pp is a cluster point. \quad\square

Proposition 5. \: CC     \iff Every sequence has a cluster point.

Proof. (    \implies) Suppose some sequence (xn)(x_n) has no cluster point. So every pp has a nbd VV such that V{xn}V \cap \{x_n\} is finite. Let Vn\mathcal{V}_n denote all open sets VV such that V{xn}{x1,,xn}V \cap \{x_n\} \subseteq \{x_1, \dots, x_n\}. If Vn=VnV_n = \bigcup \mathcal{V}_n, then {Vn}\{V_n\} is a countable cover of XX. It can’t have a finite cover, since xn+1Vnx_{n+1} \notin V_n. (    \impliedby) Suppose some countable cover {Vn}\{V_n\} has no finite subcover. Then choose x1V1x_1 \in V_1 and xn+1Xk=1nVkx_{n+1} \in X \setminus \bigcup_{k=1}^n V_k. No pp could be a cluster point of (xn)(x_n), since pVnp \in V_n for some nn and xmVnx_m \notin V_n for any m>nm > n. \quad\square

Definition. Given AXA \subseteq X, its sequential closure scl(A)\text{scl}(A) is the set of all pXp \in X of which there exists some sequence (xn)(x_n) in AA such that xnpx_n \to p. A set is sequentially closed if scl(A)A\text{scl}(A) \subseteq A.

Lemma 2. scl(A)A\text{scl}(A) \subseteq \cl{A} for any set AA. Every closed set is sequentially closed.

Proof. If (xn)A(x_n) \subseteq A with xnpx_n \to p, then by definition, for any nbd VV of pp there is an n0n_0 such that xnVx_n \in V for nn0n \ge n_0. So V{xn}V \cap \{x_n\} \ne \emptyset and therefore, pAp \in \cl{A}. The second fact follows from the fact that a set is closed if and only if AA\cl{A} \subseteq A. \quad\square

Definition. A space is sequential (Seq) if every sequentially closed is closed. A space is sequentially compact (SC) if every sequence has a converging subsequence.

Proposition 6. \: SC     \implies CC

Proof. Immediate from Lemma 1 and Proposition 5. \quad\square

This proves being countably compact is weaker than being compact or sequentially compact. However, both of the latter are incompatible, in the sense that there are topological spaces which are compact and not sequentially compact (using the axiom of choice, consider the product topology on [0,1][0,1][0,1]^{[0,1]}), and sequentially compact spaces which are not compact (The order topology on ω1\omega_1, the first uncountable ordinal). Conversely, there’s something beautiful about being able to say that a space is sequentially compact if it is sequential and compact. This is true, but the condition is not necessary, we only require countable compactness.

Proposition 7. \: (CC and Seq)     \implies SC

Proof. See Theorem 223 here. \quad\square