Bourbaki

4. Normal spaces

1. Definition of normal spaces

Axiom (OIVO_{IV}) for uniformizable spaces (§1§ 1, no. 5) can be stated in the following form: given any closed set AA and any point x∈CAx \in \mathbf{C}A, there is a continuous mapping of XX into [0,1][0, 1] which is equal to 0 at xx and is equal to 1 at every point of AA; this property can again be expressed by saying that in a uniformizable space we can separate a point and a closed set (not containing the point) by a continuous real-valued function.

We shall now study spaces in which it is possible in the same way to separate two disjoint closed sets by a continuous real-valued function:

Definition 1

A topological space XX is said to be normal if it is Hausdorff and satisfies the following axiom:

(Ov)(O_v) If AA and BB are any two disjoint closed subsets of XX, there exists a continuous mapping of XX into [0,1][0, 1] which is equal to 00 at every point of AA and to 11 at every point of BB.

Clearly every normal space is completely regular; but there are completely regular spaces which are not normal (see Exercises 9, 10, 13, 26 and § 5, Exercises 15 and 16).

The statement of Axiom (Ov)(O_v), like that of (OIV)(O_{IV}), involves the real line R\mathbf{R} as an auxiliary set. But there is a condition equivalent to (Ov)(O_v) which involves no auxiliary set:

Theorem 1 (Urysohn)

Axiom (Ov)(O_v) is equivalent to the following:

(Ov′)(O'_v) If AA and BB are any two disjoint closed subsets of XX, then there exist two disjoint open sets U,VU, V such that A⊂UA \subset U and B⊂VB \subset V.

It follows immediately that (Ov)(O_v) implies (Ov′)(O'_v), for if ff is a continuous mapping of XX into [0,1][0, 1] which is equal to 00 on AA and to 11 on BB, then the open sets f‾−1([0,1/2[)\overline{f}^{-1}([0, 1/2[) and f‾−1(]1/2,1])\overline{f}^{-1}(]1/2, 1]) contain AA and BB respectively and do not intersect.

To prove the converse, notice first that (Ov′)(O'_v) is equivalent to the following axiom:

(Ov′′)(O''_v) Given any closed set AA and any open neighbourhood VV of AA, there exists an open neighbourhood WW of AA such that W‾⊂V\overline{W} \subset V.

If there is a continuous mapping f:X→[−1,+1]f : X \to [-1, +1] which is equal to −1-1 on AA and to +1+1 on BB, and if we put U(t)=f‾−1([−1,t[)U(t) = \overline{f}^{-1}([-1, t[) for each t∈[0,1]t \in [0, 1], then we have defined a family of open sets in XX, indexed by [0,1][0, 1], such that (i) A⊂U(0)A \subset U(0), (ii) B⊂CU(1)B \subset \mathbf{C}U(1) and (iii) for each pair of real numbers t,t′t, t' such that 0≤t<t′≤10 \leq t < t' \leq 1 we have

(1)(1)
U‾(t)⊂U(t′); \overline{U}(t) \subset U(t');

for U(t)U(t) is contained in the closed set f‾−1([−1,t])\overline{f}^{-1}([-1, t]). Conversely, suppose that we have defined a family (U(t))(U(t)) of open sets (0≤t≤1)(0 \leq t \leq 1) with these three properties (i), (ii) and (iii). For each x∈Xx \in X, put g(x)=1g(x) = 1 if x∈CU(1)x \in \mathbf{C}U(1), and if x∈U(1)x \in U(1) let g(x)g(x) be the greatest lower bound of the values of tt such that x∈U(t)x \in U(t). Clearly 0≤g(x)≤10 \leq g(x) \leq 1 for each x∈Xx \in X, g(x)=0g(x) = 0 on AA, g(x)=1g(x) = 1 on BB; also gg is continuous on XX, for if we put g(x)=ag(x) = a, we have ∣g(y)−g(x)∣≤ε|g(y) - g(x)| \leq \varepsilon for all y∈U(a+ε)∩CU‾(a−ε)y \in U(a + \varepsilon) \cap \overline{\mathbf{C}U}(a - \varepsilon), which is a neighbourhood of xx by (i) [with the conventions that U(a+ε)=XU(a + \varepsilon) = X if a+ε>1a + \varepsilon > 1, and U(a−ε)=∅U(a - \varepsilon) = \varnothing if a−ε<0a - \varepsilon < 0].

Hence Theorem 1 will be proved if we can define a family (U(t))(U(t)) of open sets satisfying conditions (i), (ii) and (iii) above; to do this we use Axiom (OV′)(\mathrm{O}'_V). Take U(1)=CBU(1) = \mathbf{C}B; since A⊂U(1)A \subset U(1) there exists an open set U(o)U(o) such that A⊂U(o)A \subset U(o) and U(o)⊂U(1)U(o) \subset U(1) by (OV′)(\mathrm{O}'_V). Suppose then that for each dyadic number k/2nk/2^n (k=0,1,…,2nk = 0, 1, \ldots, 2^n) we have defined an open set U(k/2n)U(k/2^n), these sets being such that U‾(k/2n)⊂U((k+1)/2n)\overline{U}(k/2^n) \subset U((k+1)/2^n) for 0≤k≤2n−10 \leq k \leq 2^n - 1. For each dyadic number

(2k+1)/2n+1(0≤k≤2n−1) (2k + 1)/2^{n+1} \quad (0 \leq k \leq 2^n - 1)
there exists by (OV′)(\mathrm{O}'_V) an open set U((2k+1)/2n+1)U((2k + 1)/2^{n+1}) such that
U‾(k/2n)⊂U((2k+1)/2n+1) \overline{U}(k/2^n) \subset U((2k + 1)/2^{n+1})
and
U‾((2k+1)/2n+1)⊂U((k+1)/2n). \overline{U}((2k + 1)/2^{n+1}) \subset U((k + 1)/2^n).
Hence for each dyadic number rr such that 0≤r≤10 \leq r \leq 1 we can define an open set U(r)U(r), such that A⊂U(o)A \subset U(o), B⊂CU(1)B \subset \mathbf{C}U(1), and
(2)(2)
U‾(r)⊂U(r′) \overline{U}(r) \subset U(r')
for each pair of dyadic numbers r,r′r, r' such that 0≤r<r′≤10 \leq r < r' \leq 1.

Now define, for each real number t∈[0,1]t \in [0, 1],

U(t)=⋃r≤tU(r)(r dyadic); U(t) = \bigcup_{r \leq t} U(r) \quad (r \text{ dyadic});
by (2), this definition agrees with the preceding one for tt dyadic; also, if 0≤t<t′≤10 \leq t < t' \leq 1, then there exist two dyadic numbers r,r′r, r' such that t≤r<r′≤t′t \leq r < r' \leq t'; by (2) we have U(t)⊂U(r)U(t) \subset U(r), hence
U‾(t)⊂U‾(r)⊂U(r′)⊂U(t′); \overline{U}(t) \subset \overline{U}(r) \subset U(r') \subset U(t');
this proves (1) and therefore completes the proof.

Theorem 1 will enable us to show that two important categories of topological spaces are normal. In the first place:

Proposition 1

A compact space is normal.

A compact space satisfies axiom (OV′)(\mathrm{O}'_V), by Proposition 2 of Chapter I, § 9, no. 2.

As to locally compact spaces, every point of such a space has a compact neighbourhood, which is a normal subspace; but there are examples of locally compact spaces which are not normal (cf. Exercises 9 and 26, and § 5, Exercise 15).

Proposition 2

A metrizable space is normal.

Let XX be a metrizable space and let dd be a metric compatible with the topology of XX. Let A,BA, B be two disjoint closed subsets of XX; since the functions d(x,A)d(x, A) and d(x,B)d(x, B) are continuous, the set UU (resp. VV) of points xx such that d(x,A)<d(x,B)d(x, A) < d(x, B) [resp. d(x,B)<d(x,A)d(x, B) < d(x, A)] is open; clearly A⊂UA \subset U and B⊂VB \subset V and U∩V=∅U \cap V = \varnothing, hence Axiom (OV′)(O'_V) is satisfied.

Remark 1

Proposition 2 gives another necessary condition for metrizability; but this condition, even in conjunction with all the necessary conditions given in § 2, does not give a set of sufficient conditions for metrizability (cf. Exercise 6 and § 5, Exercise 10).

Remark 2

There are examples of normal spaces which are neither metrizable nor locally compact (see § 5, Exercise 16).

By (OV′)(O'_V), every closed subset of a normal space is a normal subspace; but this is not always the case for an arbitrary subset of a normal space.

For example, a completely regular space which is not normal is homeomorphic to a subspace of a compact space (§ 1, no. 5, Proposition 3), and the latter is normal.

Finally we record that the product of two normal spaces is not necessarily normal (see Exercise 9 and § 5, Exercise 16).

2. Extension of a continuous real-valued function

Let XX and YY be two topological spaces and let A≠XA \neq X be a closed subset of XX. If ff is a continuous mapping of AA into YY, it is not always possible to extend ff to a continuous mapping of the whole of XX into YY. When Y=R‾Y = \overline{\mathbf{R}}, the possibility of such an extension is determined by the following theorem:

Theorem 2 (Urysohn)

Axiom* (OV)(O_V) is equivalent to the following property: (OV′′)(O''_V) Given any closed subset AA of XX and any continuous real-valued function ff (finite or not) defined on AA, there exists an extension gg of ff to the whole space XX, which is a continuous mapping of XX into R‾\overline{\mathbf{R}}.

It is easy to see that (OV′′)(O''_V) implies (OV)(O_V); for if BB and CC are two disjoint closed subsets of XX, then the function which is equal to 0 on BB and equal to ii on CC is defined and continuous on the closed set B∪CB \cup C, hence by (OV′′)(O''_{V}) has a continuous extension ff to XX. If g=inf⁡(f+,1)g = \inf (f^{+}, 1), then gg is continuous on XX, takes its values in [0,1][0, 1] and is equal to oo on BB and to ii on CC.

Let us show conversely that (OV)(O_{V}) implies (OV′′)(O''_{V}). Since R‾\overline{\mathbf{R}} and the interval [−i,+i][ -i, +i ] are homeomorphic, we need consider only the case where the continuous mapping f:A→R‾f : A \to \overline{\mathbf{R}} takes its values in [−i,+i][ -i, +i ]. We shall construct an extension gg of ff to XX by forming a sequence (gn)(g_{n}) of continuous functions on XX, such that the sequence (gn(x))(g_{n}(x)) converges for all x∈Xx \in X to a point of the interval [−i,+i][ -i, +i ]; this limit will, by definition, be the value of gg at xx, and it will follow from the choice of the gng_{n} that the function gg satisfies the required conditions.

The definition of the gng_{n} rests on the following lemma:

Lemma 1

Let uu be a continuous mapping of AA into [−i,+i][ -i, +i ]; then there is a continuous mapping vv of XX into [−i/3,+i/3][ -i/3, +i/3 ], such that ∣u(x)−v(x)∣≤2/3|u(x) - v(x)| \leq 2/3 for all x∈Ax \in A.

Let HH be the set of all x∈Ax \in A such that −i≤u(x)≤−i/3-i \leq u(x) \leq -i/3, and let KK be the set of all x∈Ax \in A such that i/3≤u(x)≤ii/3 \leq u(x) \leq i; HH and KK are closed in AA, and therefore in XX, and do not intersect; hence by (OV)(O_{V}) there is a continuous mapping vv of XX into [−i/3,+i/3][ -i/3, +i/3 ] which is equal to −i/3-i/3 on HH and to +i/3+i/3 on KK. The mapping vv satisfies the conditions of the lemma.

We now define the functions gng_{n} by induction. Applying the lemma with u=fu = f, we define g0g_{0} to be a continuous mapping of XX into [−i/3,+i/3][ -i/3, +i/3 ] such that ∣f(x)−g0(x)∣≤2/3|f(x) - g_{0}(x)| \leq 2/3 for all x∈Ax \in A. Suppose now that a continuous mapping gng_{n} of XX into the interval

[−i+(23)n+1,i−(23)n+1] [-i + (\frac{2}{3})^{n+1}, i - (\frac{2}{3})^{n+1}]

has been defined, such that ∣f(x)−gn(x)∣≤(23)n+1|f(x) - g_{n}(x)| \leq (\frac{2}{3})^{n+1} for all x∈Ax \in A. Applying the lemma to the function u(x)=(23)n+1(f(x)−gn(x))u(x) = (\frac{2}{3})^{n+1} (f(x) - g_{n}(x)), we see that there exists a continuous mapping hn+1h_{n+1} of XX into the interval

[−2n+13n+2,2n+13n+2] \left[ -\frac{2^{n+1}}{3^{n+2}}, \frac{2^{n+1}}{3^{n+2}} \right]

such that

∣f(x)−gn(x)−hn+1(x)∣≤(2/3)n+2 |f(x) - g_{n}(x) - h_{n+1}(x)| \leq (2/3)^{n+2}

for all x∈Ax \in A; the induction is completed by taking gn+1=gn+hn+1g_{n+1} = g_{n} + h_{n+1}, since this function satisfies the inequality ∣gn+1(x)∣≤1−(2/3)n+2|g_{n+1}(x)| \leq 1 - (2/3)^{n+2} for all x∈Xx \in X, by virtue of the definition of hn+1h_{n+1}.

From this definition it follows that, for m≥pm \geq p and n≥pn \geq p, we have

∣gm(x)−gn(x)∣≤2p+13p+2∑k=0∞(2/3)k=(2/3)p+1 |g_{m}(x) - g_{n}(x)| \leq \frac{2^{p+1}}{3^{p+2}} \sum_{k=0}^{\infty} (2/3)^{k} = (2/3)^{p+1}

at each point x∈Xx \in X; hence the sequence (gn(x))(g_n(x)) is a Cauchy sequence for each x∈Xx \in X, and therefore converges to a point g(x)g(x) of the interval [−1,+1][ -1, +1 ]; and since f(x)−gn(x)f(x) - g_n(x) tends to 0 for all x∈Ax \in A as n→∞n \to \infty, gg is an extension of ff to XX. It remains therefore only to show that gg is continuous on XX.

Now let xx be any point of XX; then, given any ε>0\varepsilon > 0, there exists an integer n0n_0 such that ∣gm(y)−gn(y)∣≤ε|g_m(y) - g_n(y)| \leq \varepsilon for all y∈Xy \in X and all m≥n0m \geq n_0 and all n≥n0n \geq n_0; hence, letting mm tend to +∞+\infty, we have

∣g(y)−gn(y)∣≤ε. |g(y) - g_n(y)| \leq \varepsilon.

Let VV be a neighbourhood of xx such that ∣gn(y)−gn(x)∣≤ε|g_n(y) - g_n(x)| \leq \varepsilon for all y∈Vy \in V; then, for each y∈Vy \in V we shall have

∣g(y)−g(x)∣≤∣g(y)−gn(y)∣+∣gn(y)−gn(x)∣+∣g(x)−gn(x)∣≤3ε, |g(y) - g(x)| \leq |g(y) - g_n(y)| + |g_n(y) - g_n(x)| + |g(x) - g_n(x)| \leq 3\varepsilon,

which shows that gg is continuous at xx, and completes the proof of Theorem 2. (The last part of the proof uses, in a particular case, the idea of uniform convergence, which we shall study in general in Chapter X, no. 1.)

Corollary

If ff is a finite continuous real-valued function defined on AA, then there exists a finite continuous real-valued function gg defined on XX, which extends ff.

First consider the case in which f(x)≥0f(x) \geq 0 for all x∈Ax \in A; then there is a continuous extension g1g_1 of ff to XX, taking its values in [0,+∞][0, +\infty]. If we put B=g1‾(+∞)B = \overline{g_1}(+\infty), then BB is closed and by hypothesis does not meet AA; the function hh which is equal to ff on AA and to 0 on BB is therefore a continuous function on the closed set A∪BA \cup B. Let g2g_2 be a continuous extension of hh to XX, again taking its values in [0,+∞][0, +\infty]; the function g=inf⁡(g1,g2)g = \inf(g_1, g_2) is then a continuous extension of ff to XX, whose values are ≥0\geq 0 and finite at every point of XX.

To pass to the general case, it is enough to remark that, if ff is finite and continuous on AA, then so are f+f^+ and f−f^-; extending f+f^+ and f−f^- to XX by finite continuous functions g1g_1 and g2g_2 respectively, we see that the function g1−g2g_1 - g_2 is finite and continuous on XX and extends ff.

Remark

If XX is a normal space and if AA is a closed subset of XX, there exists also a continuous extension to XX of every continuous mapping ff of AA into a cube K1K^1 (§ 1, no. 5); for we have then f=(fi)i∈I,fif = (f_i)_{i \in I}, f_i being a continuous mapping of AA into the compact interval KK of R\mathbf{R}; since there exists a continuous mapping gi:X→Kg_i : X \to K which extends fif_i, the mapping g=(gi)g = (g_i) is a continuous extension of ff to XX.

3. Locally finite open coverings of a closed set in a normal space

Theorem 3

Let (Ai)i∈I(A_i)_{i \in I} be a locally finite open covering of a closed set YY in a normal space XX. Then there is an open covering (Bi)i∈I(B_i)_{i \in I} of YY such that B‾i⊂Ai\overline{B}_i \subset A_i for each i∈Ii \in I.

Well-order the index set II (Set Theory, Chapter III, § 2, no. 3, Theorem 1). We shall define a family (Bi)i∈I(B_i)_{i \in I} of open sets in XX, by transfinite induction, such that (i) B‾i⊂Ai\overline{B}_i \subset A_i for each i∈Ii \in I; (ii) for each i∈Ii \in I, the family formed by the BλB_\lambda such that λ≤i\lambda \leq i and by the AλA_\lambda such that λ>i\lambda > i is an open covering of YY. Suppose that we have defined the BiB_i for i<γi < \gamma so that (i) and (ii) are satisfied for all i<γi < \gamma, and let us show that we can define BγB_\gamma in such a way that (i) and (ii) are also satisfied for i=γi = \gamma. Let us first show that the BiB_i for which i<γi < \gamma and the AiA_i for which i≥γi \geq \gamma form a covering of YY. By hypothesis, for each x∈Yx \in Y there is only a finite number of indices λ∈I\lambda \in I such that x∈Aλx \in A_\lambda, say λ1<λ2<⋯<λn\lambda_1 < \lambda_2 < \cdots < \lambda_n; let λh\lambda_h be the greatest of the λi\lambda_i such that λi<γ\lambda_i < \gamma; if h<nh < n we have x∈Aλnx \in A_{\lambda_n} and λn≥γ\lambda_n \geq \gamma, and if h=nh = n the inductive hypothesis shows that xx belongs to some BλB_\lambda such that λ≤λn<γ\lambda \leq \lambda_n < \gamma, and our assertion follows.

Now put C=(∁Y)∪(⋃i<γBi)∪(⋃i>γAi)C = (\complement Y) \cup \left( \bigcup_{i < \gamma} B_i \right) \cup \left( \bigcup_{i > \gamma} A_i \right); CC is open, and from what has been said we have ∁Aγ⊂C\complement A_\gamma \subset C; by virtue of Axiom (OV′′)(O''_V) for normal spaces, there is therefore an open set VV such that ∁Aγ⊂V‾⊂V⊂C\complement A_\gamma \subset \overline{V} \subset V \subset C. If we put Bγ=∁V‾B_\gamma = \complement \overline{V}, we have B‾γ⊂∁V⊂Aγ\overline{B}_\gamma \subset \complement V \subset A_\gamma and Bγ∪C=XB_\gamma \cup C = X, so that the BiB_i such that i≤γi \leq \gamma and the AiA_i such that i>γi > \gamma cover YY.

Remark

Note that we have used only the fact that the covering (Ai)(A_i) is point-finite, i.e., that every point of XX belongs to only a finite number of sets AiA_i.

Definition 2

Let XX be a topological space and let ff be a real-valued function defined on XX. The support of ff, denoted by Supp⁡(f)\operatorname{Supp}(f), is the smallest closed set SS in XX such that f(x)=0f(x) = 0 for all x∉Sx \notin S.

In other words, Supp⁡(f)\operatorname{Supp}(f) is the closure in XX of the set of all x∈Xx \in X such that f(x)≠0f(x) \neq 0; or again, it is the set of all x∈Xx \in X such that every neighbourhood of xx contains a point yy for which f(y)≠0f(y) \neq 0.

Let (fi)i∈I(f_i)_{i \in I} be a family of finite real-valued functions on XX whose supports form a locally finite family; then the sum ∑i∈Ifi(x)\sum_{i \in I} f_i(x) is defined for each x∈Xx \in X (since it contains only a finite number of non-zero terms). The finite real-valued function x→∑i∈Ifi(x)x \to \sum_{i \in I} f_i(x) is called the sum of the family (fi)(f_i), and is denoted by ∑i∈Ifi\sum_{i \in I} f_i. If each of the fif_i is continuous, then so is f=∑i∈Ifif = \sum_{i \in I} f_i; for if xx is any point of XX, there is a neighbourhood VV of xx which meets only a finite number of supports of the fif_i, and hence there is a finite subset HH of II such that f(y)=∑i∈Hfi(y)f(y) = \sum_{i \in H} f_i(y) for all y∈Vy \in V.

Definition 3

Given a family* (Ai)i∈I(A_i)_{i \in I} of subsets of a topological space XX, a family (fi)i∈I(f_i)_{i \in I} of real-valued functions defined on XX is said to be subordinate to the family (Ai)i∈I(A_i)_{i \in I} if we have Supp⁡(fi)⊂Ai\operatorname{Supp}(f_i) \subset A_i for each index i∈Ii \in I.

A continuous partition of unity on XX is any family (fi)i∈I(f_i)_{i \in I} of real-valued continuous functions ≥0\geq 0 on XX whose supports form a locally finite family and which are such that ∑i∈Ifi(x)=1\sum_{i \in I} f_i(x) = 1 for all x∈Xx \in X.

Proposition 3

Given any locally finite open covering* (Ai)i∈I(A_i)_{i \in I} of a normal space XX, there exists a continuous partition of unity (fi)i∈I(f_i)_{i \in I} on XX, subordinate to the covering (Ai)i∈I(A_i)_{i \in I}.

By Theorem 3 there exists an open covering (Bi)i∈I(B_i)_{i \in I} of XX such that B‾i⊂Ai\overline{B}_i \subset A_i for each i∈Ii \in I, and it is clear that the covering (Bi)(B_i) is locally finite. By axiom (OV′′)(O''_V), for each i∈Ii \in I there exists an open set CiC_i such that B‾i⊂Ci⊂C‾i⊂Ai\overline{B}_i \subset C_i \subset \overline{C}_i \subset A_i. By axiom (OV)(O_V), for each i∈Ii \in I there exists a continuous mapping gig_i of XX into [0,1][0,1], such that gi(x)=1g_i(x) = 1 on B‾i\overline{B}_i and such that the support of g‾i\overline{g}_i is contained in C‾i\overline{C}_i, and therefore contained in AiA_i. Since (Bi)(B_i) is a covering of XX, we have ∑i∈Igi(x)>0\sum_{i \in I} g_i(x) > 0 for each x∈Xx \in X; if we put

fi(x)=gi(x)∑i∈Igi(x) f_i(x) = \frac{g_i(x)}{\sum_{i \in I} g_i(x)}
for all i∈Ii \in I and all x∈Xx \in X, then the fif_i form a continuous partition of unity subordinate to the covering (Ai)(A_i).

Corollary

Given any locally finite open covering* (Ai)(A_i) of a closed set FF in a normal space XX, there exists a family (fi)(f_i) of continuous real-valued functions ≥0\geq 0 on XX, subordinate to the covering (Ai)i∈I(A_i)_{i \in I} and such that ∑i∈Ifi(x)=1\sum_{i \in I} f_i(x) = 1 for all x∈Fx \in F and ∑i∈Ifi(x)≤1\sum_{i \in I} f_i(x) \leq 1 for all x∈Xx \in X.

The family of sets consisting of CFCF and the AiA_i is a locally finite open covering of XX. There is therefore a continuous partition of unity subordinate to this covering, consisting of a family (fi)i∈I(f_i)_{i \in I} such that Supp⁡(fi)⊂Ai\operatorname{Supp}(f_i) \subset A_i for each i∈Ii \in I, and a function gg whose support is contained in the complement of FF. The family (fi)(f_i) clearly satisfies the required conditions.

4. Paracompact spaces

We recall (Chapter I, § 9, no. 10) that a topological space XX is said to be paracompact if it is Hausdorff and if every open covering of XX has a locally finite open refinement.

Proposition 4

Every paracompact space is normal.

This is a consequence of the following lemma:

Lemma 2

Let A,BA, B be two disjoint closed subsets of a paracompact space XX. If for each x∈Ax \in A there is an open neighbourhood VxV_x of xx and an open neighbourhood WxW_x of BB which do not intersect, then there exists an open neighbourhood TT of AA and an open neighbourhood UU of BB which do not intersect.

Assuming the truth of this lemma for the moment, we can apply it to the case where BB consists of a single point, because XX is Hausdorff, and it shows then that XX is regular. We can then apply Lemma 2 again to any two disjoint closed subsets of XX, and this shows that Axiom (OV′)(O'_V) is satisfied.

To prove the lemma, consider the open covering of XX consisting of CACA and the VxV_x, where x∈Ax \in A; let (Ti)i∈I(T_i)_{i \in I} be a locally finite open refinement of this covering. Then, by definition, if A∩Ti≠∅A \cap T_i \neq \emptyset there exists xi∈Ax_i \in A such that Ti⊂VxiT_i \subset V_{x_i}. Let TT be the open set which is the union of the TiT_i which meet AA, and let us show that there is an open neighbourhood UU of BB which does not meet TT. For each y∈By \in B there is an open neighbourhood SyS_y of yy which meets only a finite number of sets TiT_i; let JJ be the finite subset of II consisting of those indices ii such that TiT_i meets both SyS_y and AA; if we put Uy=Sy∩⋂i∈JWxiU_y = S_y \cap \bigcap_{i \in J} W_{x_i}, then UyU_y is an open neighbourhood of yy which meets none of the TiT_i which meet AA, and hence Uy∩T=∅U_y \cap T = \emptyset. Let U=⋃y∈BUyU = \bigcup_{y \in B} U_y; then U′U' is an open neighbourhood of BB which does not meet TT, and the lemma is proved.

There exist normal spaces which are not paracompact (Exercise 19).

Corollary 1

Given any open covering (Ai)i∈I(A_i)_{i \in I} of a paracompact space XX, there exists a continuous partition of unity (fi)i∈I(f_i)_{i \in I} on xx, subordinate to the covering (Ai)(A_i).

Let (Uλ)λ∈L(U_\lambda)_{\lambda \in L} be a locally finite open covering of XX which refines the covering (Ai)i∈I(A_i)_{i \in I}; then there is a mapping φ:L→I\varphi : L \to I such that Uλ⊂Aφ(λ)U_\lambda \subset A_{\varphi(\lambda)} for each λ∈L\lambda \in L. By Propositions 3 and 4, there exists a continuous partition of unity (gλ)λ∈L(g_\lambda)_{\lambda \in L} subordinate to (Uλ)(U_\lambda). For each i∈Ii \in I, put

fi=∑φ(λ)=igλ; f_i = \sum_{\varphi(\lambda) = i} g_\lambda;

this sum is defined and continuous since the supports of the gλg_\lambda form a locally finite covering; moreover, the union BiB_i of the supports of the gλg_\lambda such that φ(λ)=i\varphi(\lambda) = i is closed, by Proposition 4 of Chapter I, § 1, no. 6, and is contained in AiA_i. Since we have fi(x)=0f_i(x) = 0 whenever x∈CBix \in \mathbf{C}B_i, the support of fif_i is contained in BiB_i, and therefore in AiA_i. On the other hand, the family BiB_i is locally finite, because for each x∈Xx \in X there is a neighbourhood VV of xx and a finite subset HH of II such that V∩Uλ=∅V \cap U_\lambda = \emptyset for all λ∉H\lambda \notin H, and it follows therefore that V∩Bi=∅V \cap B_i = \emptyset for all i∉φ(H)i \notin \varphi(H). Finally, for each x∈Xx \in X we have

1=∑λ∈Lgλ(x)=∑i∈I(∑φ(λ)=igλ(x))=∑i∈Ifi(x), 1 = \sum_{\lambda \in L} g_\lambda(x) = \sum_{i \in I} \left( \sum_{\varphi(\lambda) = i} g_\lambda(x) \right) = \sum_{i \in I} f_i(x),

and the proof is complete.

Corollary 2

If F is a closed subset of a paracompact space X, then every neighbourhood of F in X contains a closed (and therefore paracompact) neighbourhood of F.

By Proposition 16 of Chapter I, § 9, no. 10, any closed subspace of X is paracompact; the corollary therefore follows from Proposition 4 and Axiom (Ov′)(O_v').

5. Paracompactness of metrizable spaces

The following theorem sharpens the result of Proposition 2 of no. 1:

Theorem 4

Every metrizable space is paracompact.

The theorem is a consequence of the following four lemmas.

Lemma 3

Let R=(Uα)α∈A\mathscr{R} = (U_\alpha)_{\alpha \in A} be an open covering of a metrizable space X. Then there is a sequence (Gn)(\mathscr{G}_n) of locally finite families of open subsets of X, such that S=⋃nSn\mathscr{S} = \bigcup_n \mathscr{S}_n is an open covering of X which refines R\mathscr{R}.

Let dd be a metric on X compatible with its topology. For each α∈A\alpha \in A and each integer nn, let FnαF_{n\alpha} denote the set of all x∈Uαx \in U_\alpha such that d(x,X−Uα)≥2−nd(x, X - U_\alpha) \geq 2^{-n}. Since X−UαX - U_\alpha is closed, we have Uα=⋃nFnαU_\alpha = \bigcup_n F_{n\alpha}.

Well-order the set AA; for each α∈A\alpha \in A and each integer nn, let GnαG_{n\alpha} be the set of all x∈Fnαx \in F_{n\alpha} such that x∉Fn+1,βx \notin F_{n+1,\beta} for all β<α\beta < \alpha, and let VnαV_{n\alpha} be the set of all y∈Xy \in X such that d(y,Gnα)>2−n−3d(y, G_{n\alpha}) > 2^{-n-3}. VnαV_{n\alpha} is clearly an open set; on the other hand, Vnα⊂UαV_{n\alpha} \subset U_\alpha, because for each y∈Vnαy \in V_{n\alpha} there exists x∈Gnαx \in G_{n\alpha} such that d(x,y)≤2−n−1d(x, y) \leq 2^{-n-1}, and since x∈Fnαx \in F_{n\alpha}, we have

d(y,X−Uα)≥d(x,X−Uα)−d(x,y)≥2−n−1, d(y, X - U_\alpha) \geq d(x, X - U_\alpha) - d(x, y) \geq 2^{-n-1},

so that y∈Uαy \in U_\alpha. Furthermore, for each x∈Xx \in X let α\alpha be the smallest index in AA such that x∈Uαx \in U_\alpha; then there is an integer nn such that x∈Fnαx \in F_{n\alpha}, and the definition of α\alpha implies that x∈Gnαx \in G_{n\alpha}, so that x∈Vnαx \in V_{n\alpha}. This shows that if we put Sn=(Vnα)α∈A\mathfrak{S}_n = (V_{n\alpha})_{\alpha \in A}, then S=⋃nSn\mathfrak{S} = \bigcup_n \mathfrak{S}_n is an open covering of XX which refines R\mathscr{R}; thus it remains to be shown that each of the families Sn\mathfrak{S}_n is locally finite. To this end we shall first show that d(Gnα,Gnβ)≥2−n−1d(G_{n\alpha}, G_{n\beta}) \geq 2^{-n-1} if α≠β\alpha \neq \beta. Suppose that β<α\beta < \alpha; then if x∈Gnαx \in G_{n\alpha} and y∈Fnβy \in F_{n\beta} we have x∉Fn+1,βx \notin F_{n+1,\beta} by definition, hence d(x,X−Uβ)<2−n−1d(x, X - U_\beta) < 2^{-n-1} and d(y,X−Uβ)≥2−nd(y, X - U_\beta) \geq 2^{-n}, and therefore d(x,y)≥2−n−1d(x, y) \geq 2^{-n-1}; since Gnβ⊂FnβG_{n\beta} \subset F_{n\beta}, the assertion follows. From this it follows immediately, using the definition of VnαV_{n\alpha} and VnβV_{n\beta}, that d(Vnα,Vnβ)≥2−n−2d(V_{n\alpha}, V_{n\beta}) \geq 2^{-n-2}. From this last inequality we deduce that, for each z∈Xz \in X, the open ball with centre zz and radius 2−n−32^{-n-3} meets at most one set of the family Sn\mathfrak{S}_n; hence Sn\mathfrak{S}_n is a locally finite family, and the proof is complete.

Lemma 4

Let* (Sn)(\mathfrak{S}_n) be a sequence of locally finite families of open sets in a topological space XX, such that

S=⋃nSn \mathfrak{S} = \bigcup_n \mathfrak{S}_n
is a covering of XX. Then there exists a locally finite (but not necessarily open) covering B\mathscr{B} of XX which refines S\mathfrak{S}.

Let EnE_n be the open set in XX which is the union of all the sets of Sn\mathfrak{S}_n; let UnU_n denote ⋃k=1nEk\bigcup_{k=1}^n E_k and let AnA_n denote Un−Un−1U_n - U_{n-1} (with U0=∅U_0 = \varnothing). Consider the set B\mathscr{B} of subsets V∩AnV \cap A_n, where V∈SnV \in \mathfrak{S}_n and nn is any integer; we shall show that B\mathscr{B} satisfies the conditions of the lemma. For each x∈Xx \in X there is an integer nn such that x∈Anx \in A_n, since the AnA_n form a partition of XX; thus x∈Enx \in E_n and there exists V∈SnV \in \mathfrak{S}_n such that x∈Vx \in V; so x∈V∩Anx \in V \cap A_n, and we have proved that B\mathscr{B} is a covering of XX. Clearly, this covering refines S\mathfrak{S}. On the other hand, for each x∈Xx \in X there exists an integer nn such that x∈Unx \in U_n; since UnU_n is open and the Sm\mathfrak{S}_m are locally finite families, there exists a neighbourhood WmW_m of xx, for each mm, contained in UnU_n, which meets only a finite number of sets of Sm\mathfrak{S}_m; hence the neighbourhood W=⋂m=1nWmW = \bigcap_{m=1}^n W_m of xx meets only a finite number of sets of B\mathscr{B}, because W∩Ap=∅W \cap A_p = \varnothing for p>np > n. Hence B\mathscr{B} is locally finite, and Lemma 4 is proved.

Lemma 5

Let* XX be a regular space such that, for each open covering R\mathscr{R} of XX, there exists a (not necessarily open) locally finite covering U\mathscr{U} of XX which refines R\mathscr{R}. Then for each open covering R\mathscr{R} of XX there exists a locally finite closed covering F\mathfrak{F} of XX which refines R\mathscr{R}.

Let R\mathfrak{R} be any open covering of XX. For each x∈Xx \in X there is an open set U∈RU \in \mathfrak{R} which contains xx, and therefore (since XX is regular) an open neighbourhood VxV_x of xx such that V‾x⊂U\overline{V}_x \subset U. The family B\mathscr{B} formed by the VxV_x is an open covering of XX, hence by hypothesis there is a locally finite covering B′\mathscr{B}' which is finer than B\mathscr{B}. Let F\mathscr{F} be the family of closures of the sets of B\mathscr{B}. Since the covering B′\mathscr{B}' formed by the V‾x\overline{V}_x is finer than R\mathfrak{R}, and since F\mathscr{F} is finer than B′\mathscr{B}', it follows that F\mathscr{F} is a closed covering of XX which refines R\mathfrak{R}. But also F\mathscr{F} is locally finite, because if an open set does not meet a set B∈BB \in \mathscr{B}, then it does not meet its closure B‾\overline{B} either.

Lemma 6

Let XX be a Hausdorff space such that, given any open covering R\mathfrak{R} of XX, there exists a locally finite closed covering F\mathscr{F} of XX which refines R\mathfrak{R}. Then XX is paracompact.

Let R\mathfrak{R} be any open covering of XX. We have to show that there is a locally finite open covering of XX which refines R\mathfrak{R}. Let A\mathscr{A} be a locally finite covering (closed or not) of XX which refines R\mathfrak{R}; for each x∈Xx \in X, let WxW_x be an open neighbourhood of xx which meets only a finite number of sets of A\mathscr{A}. The family W\mathscr{W} of sets WxW_x is an open covering of XX. Let F\mathscr{F} be a locally finite closed covering of XX which refines W\mathscr{W}. For each A∈AA \in \mathscr{A}, let UAU_A be a set of R\mathfrak{R} which contains AA, and let CAC_A be the union of the sets F∈FF \in \mathscr{F} such that A∩F=∅A \cap F = \varnothing. Since F\mathscr{F} is locally finite, CAC_A is closed in XX (Chapter I, § 1, no. 6, Proposition 4) and therefore A′=UA∩(X−CA)A' = U_A \cap (X - C_A) is open. Since we have A∩CA=∅A \cap C_A = \varnothing and A⊂UAA \subset U_A, it follows that A⊂A′A \subset A', and the family A′\mathscr{A}' of sets A′A', as AA runs through A\mathscr{A}, is an open covering of XX; moreover, since A′⊂UA∈RA' \subset U_A \in \mathfrak{R}, A′\mathscr{A}' refines R\mathfrak{R}. It remains to show that A′\mathscr{A}' is locally finite. For each x∈Xx \in X there is a neighbourhood TT of xx which meets only a finite number of sets of F\mathscr{F}, say F1,…,FnF_1, \ldots, F_n. Since each FiF_i is contained in a set of the form WyiW_{y_i}, by definition FiF_i meets only a finite number of sets of A\mathscr{A}; let AijA_{ij} (1≤j≤si1 \leq j \leq s_i) be these sets. If AA is a set of A\mathscr{A} other than one of the AijA_{ij} (1≤i≤n, 1≤j≤si1 \leq i \leq n,\ 1 \leq j \leq s_i), it follows from the definitions that A′A' meets none of the FiF_i, and therefore does not meet T⊂⋃i=1nFiT \subset \bigcup_{i=1}^n F_i. This completes the proof of Lemma 6 and of Theorem 4.

Exercises

See the exercises for § 4.

Transcribed from the printed volume. Not translated.