Bourbaki

2. Metric spaces and metrizable spaces

1. Metrics and metric spaces

Definition 1

A metric on a set* XX is a finite pseudometric dd on XX such that the relation d(x,y)=0d(x, y) = 0 implies x=yx = y. A metric space is a set XX endowed with the structure defined by a given metric on XX.

A metric space XX is always considered as carrying the uniformity and the topology defined by the given metric on XX.

Example 1

The Euclidean distance d(x,y)d(x, y) (Chapter VI, § 2, no. 1) is a metric on real nn-dimensional number space Rn\mathbf{R}^n; so are the functions

sup⁡1≤i≤n∣xi−yi∣and∑i=1n∣xi−yi∣. \sup_{1 \leq i \leq n} |x_i - y_i| \quad \text{and} \quad \sum_{i=1}^n |x_i - y_i|.

All these metrics are equivalent (§\S 1, no. 2).

Example 2

On any set XX the pseudometric dd, defined by the relations d(x,x)=0d(x, x) = 0 and d(x,y)=1d(x, y) = 1 if x≠yx \neq y, is a metric. The uniformity it defines on XX is the discrete uniformity.

We have a definition equivalent to Definition 1 if we say that a metric is a finite pseudometric such that the uniformity defined by this pseudometric is Hausdorff; a finite pseudometric which is equivalent to a metric is therefore a metric.

Uniform spaces defined by a single pseudometric (which we may assume to be finite) can be reduced to metric spaces when the pseudometric is not a metric. Let ff be such a pseudometric on a set XX, and let U\mathscr{U} be the uniformity defined by ff; U\mathscr{U} is not Hausdorff, and the intersection of the entourages of U\mathscr{U} is the subset of X×XX \times X defined by the equivalence relation f(x,y)=0f(x, y) = 0. Let RR denote this relation. If x≡x′(modR)x \equiv x' \pmod{R}, then by the triangle inequality we have

f(x,y)≤f(x,x′)+f(x′,y)=f(x′,y) f(x, y) \leq f(x, x') + f(x', y) = f(x', y)

and similarly f(x′,y)≤f(x,y)f(x', y) \leq f(x, y), so that f(x,y)=f(x′,y)f(x, y) = f(x', y); in other words, ff is a function compatible (in xx and yy) with the equivalence relation RR (Set Theory, R, § 5, no. 7). Let f‾\overline{f} be the function induced by ff on the quotient set; f‾\overline{f} is defined on (X/R)×(X/R)(X/R) \times (X/R), and if xx and yy are any two points of XX and if x˙\dot{x} and y˙\dot{y} denote the equivalence classes (mod RR) of xx and yy respectively, then we have f‾(x˙,y˙)=f(x,y)\overline{f}(\dot{x}, \dot{y}) = f(x, y). It follows immediately that f‾\overline{f} is a metric on X/RX/R; it is called the metric associated with the pseudometric ff; furthermore, the uniformity it defines on X/RX/R is precisely the Hausdorff uniformity associated with U\mathscr{U} by the definition of this uniformity (Chapter II, § 3, no. 8, Remark). Thus, by passing to a suitable quotient space, the uniform structure defined by a single pseudometric can be reduced to the structure of a metric space.

Proposition 1 of § 1, no. 3 determines the structure of the completion of a metric space:

Proposition 1

Let XX be a metric space and let dd be its metric. If X^\hat{X} is the completion of XX (with respect to the uniformity defined by dd), the function dd can be extended by continuity to X^×X^\hat{X} \times \hat{X}; the extended function d‾\overline{d} is a metric on X^\hat{X}, and the uniformity of X^\hat{X} coincides with that defined by the metric d‾\overline{d}.

Proposition 1 of § 1, no. 3 shows that d‾\overline{d} is a finite pseudometric on X^\hat{X}, and that the uniformity defined by d‾\overline{d} on X^\hat{X} is the uniformity obtained by completion; since this latter uniformity is Hausdorff, d‾\overline{d} is a metric. Whenever we consider the completion of a metric space XX as a metric space, it is always to be understood that the metric on X^\hat{X} is that obtained by extending the metric on XX by continuity.

2. Structure of a metric space

Let XX and X′X' be two metric spaces, dd the metric on XX, d′d' the metric on X′X'. In accordance with the general definitions (Set Theory, R, § 8, no. 5) a one-to-one mapping ff of XX onto X′X' is an isomorphism of the metric space structure of XX onto that of X′X' if

(1)(1)
d(x,y)=d′(f(x),f(y)) d(x, y) = d'(f(x), f(y))

for all x∈Xx \in X and all y∈Xy \in X.

Note that if ff is a mapping of XX onto X′X' which satisfies the identity (1), then ff must be bijective and therefore an isomorphism of XX onto X′X'; such an isomorphism is also called an isometry (or an isometric mapping) of XX onto X′X'.

An isometry of X onto X' is of course an isomorphism of the uniformity (resp. topology) of X onto the uniformity (resp. topology) of X'; the converses of these statements are false, as is shown by the existence of distinct equivalent metrics (§ 1, no. 2).

Let X be a metric space, d the metric on X. For each a > 0 let V_a denote the subset of X × X consisting of all pairs (x, y) such that d(x, y) < a, and let W_a denote the subset of X × X consisting of all pairs (x, y) such that d(x, y) ≤ a. As a runs through the set of all real numbers > 0 (or merely a sequence of numbers > 0 which tends to 0), the sets V_a (resp. W_a) form a fundamental system of open (resp. closed) entourages of the uniformity of X, because of the continuity of d (§ 1, no. 3). We have V‾a⊂Wa\overline{V}_a \subset W_a, but these two sets are not necessarily the same.

By analogy with the case of the Euclidean distance on Rn\mathbf{R}^n, the set Va(x)V_a(x) [resp. Wa(x)W_a(x)] is called the open (resp. closed) ball with centre x and radius a; it is an open (resp. closed) set in X. Again, the set of all y∈Xy \in X such that d(x,y)=ad(x, y) = a is called the sphere with centre x and radius a; it is a closed set. From what has been said, the open (resp. closed) balls with centre x and radius a form a fundamental system of neighbourhoods of x as a runs through the set of all real numbers > 0, or a sequence of numbers > 0 which tends to 0.

The reader should beware of assuming that balls and spheres in an arbitrary metric space enjoy the same properties as the Euclidean balls and spheres studied in Chapter VI, § 2. Thus the closure of an open ball need not be the closed ball of the same centre and radius; the frontier of a closed ball need not be the sphere of the same centre and radius; an open (or closed) ball need not be connected; and a sphere can be empty (cf. Exercise 4).

Let A and B be any two non-empty subsets of a metric space X. The number

d(A,B)=inf⁡x∈A,y∈Bd(x,y) d(A, B) = \inf_{x \in A, y \in B} d(x, y)
is called the distance between the sets A and B. In particular we denote by d(x,A)d(x, A) the distance between the set {x}\{x\} and the set A; this is called the distance from the point x to the set A. Thus
d(x,A)=inf⁡y∈Ad(x,y) d(x, A) = \inf_{y \in A} d(x, y)
whence
d(A,B)=inf⁡x∈Ad(x,B) d(A, B) = \inf_{x \in A} d(x, B)
(Chapter IV, § 5, no. 4, Proposition 9).

Remark

If d(x,A)=ad(x, A) = a, it can happen that there is no point of A whose distance from x is equal to a. However, this situation can never arise if A is compact, for then Weierstrass's theorem (Chapter IV, § 6, no. 1, Theorem 1) shows that there exists y∈Ay \in A such that d(x,A)=d(x,y)d(x, A) = d(x, y).

Proposition 2

The statements* d(x,A)=0d(x, A) = 0 and x∈A‾x \in \overline{A} are equivalent.

For d(x,A)=0d(x, A) = 0 expresses the fact that the ball Va(x)V_a(x) meets AA whatever the value of a>0a > 0; and this is equivalent to x∈A‾x \in \overline{A}.

Proposition 3

The function* d(x,A)d(x, A) is uniformly continuous on XX.

Let xx and yy be any two points of XX; then given any ε>0\varepsilon > 0 there exists z∈Az \in A such that d(y,z)≤d(y,A)+εd(y, z) \leq d(y, A) + \varepsilon, and therefore

d(x,z)≤d(x,y)+d(y,z)≤d(x,y)+d(y,A)+ε d(x, z) \leq d(x, y) + d(y, z) \leq d(x, y) + d(y, A) + \varepsilon
by the triangle inequality.

A fortiori d(x,A)≤d(x,y)+d(y,A)+εd(x, A) \leq d(x, y) + d(y, A) + \varepsilon, and since ε\varepsilon is arbitrary it follows that d(x,A)≤d(x,y)+d(y,A)d(x, A) \leq d(x, y) + d(y, A). Similarly we have

d(y,A)≤d(x,y)+d(x,A), d(y, A) \leq d(x, y) + d(x, A),
so that
∣d(x,A)−d(y,A)∣≤d(x,y), |d(x, A) - d(y, A)| \leq d(x, y),
whence the result follows.

Remark

We can have d(A,B)=0d(A, B) = 0 for two subsets A,BA, B of XX such that A‾∩B‾=∅\overline{A} \cap \overline{B} = \varnothing, provided that neither subset consists of a single point. For example, on the real line R\mathbf{R}, the set of integers >0> 0 and the set of points of the sequence (n+1/2n)n≥1(n + 1/2^n)_{n \geq 1} are two disjoint closed sets whose distance a part is zero.

However, if AA is compact and BB is closed, the relation d(A,B)=0d(A, B) = 0 implies A∩B≠∅A \cap B \neq \varnothing, for by virtue of the relation

d(A,B)=inf⁡x∈Ad(x,B) d(A, B) = \inf_{x \in A} d(x, B)
it follows from Proposition 3 and Weierstrass's theorem that there exists x0∈Ax_0 \in A such that d(x0,B)=d(A,B)=0d(x_0, B) = d(A, B) = 0 and hence (Proposition 2) x0∈Bx_0 \in B.

The diameter of a non-empty subset AA of XX is the number (finite or equal to +∞+\infty)

δ(A)=sup⁡x∈A,y∈Ad(x,y). \delta(A) = \sup_{x \in A, y \in A} d(x, y).

The notion of a "W_a-small set" (Chapter II, § 3, no. 1) is identical with that of a set of diameter ≤a\leq a. A non-empty set AA consists of a single point if and only if δ(A)=0\delta(A) = 0.

A subset AA of XX is bounded (with respect to the metric dd) if its diameter is finite; equivalently, if for each point x0∈Xx_0 \in X, AA is contained in a ball with centre x0x_0. Every subset of a bounded set is bounded, and the union of a finite family of bounded sets is a bounded set.

Note that a subset of XX can be bounded with respect to a metric dd but unbounded with respect to a metric equivalent to dd (cf. § 1, no. 2).

3. Oscillation of a function

Related to the notion of diameter is that of the oscillation of a function ff, defined on an arbitrary set XX and taking its values in a metric space X′X'; if AA is any non-empty subset of XX, the diameter δ(f(A))\delta(f(A)) is called the oscillation of ff in AA.

If moreover XX is a subset of a topological space YY, and if x∈X‾x \in \overline{X}, the number

ω(x;f)=inf⁡δ(f(V∩X)) \omega(x; f) = \inf \delta(f(V \cap X))
(as VV runs through the neighbourhood filter of xx in YY) is called the oscillation of ff at x∈X‾x \in \overline{X}.

Proposition 4

The oscillation ω(x;f)\omega(x; f) of an arbitrary function ff, defined on a subset XX of a topological space YY and taking its values in a metric space X′X', is upper semi-continuous on X‾\overline{X}.

Let aa be any point of X‾\overline{X}; then for each k>ω(a;f)k > \omega(a; f) there exists an open neighbourhood VV of aa such that δ(f(V∩X))≤k\delta(f(V \cap X)) \leq k; for each x∈V∩X‾x \in V \cap \overline{X}, VV is a neighbourhood of xx and therefore

ω(x;f)≤δ(f(V∩X))≤k, \omega(x; f) \leq \delta(f(V \cap X)) \leq k,
which shows that ω\omega is upper semi-continuous at the point aa.

In order that ω(x;f)=0\omega(x; f) = 0 at a point x∈X‾x \in \overline{X} it is necessary and sufficient that for each ε>0\varepsilon > 0 there should exist a neighbourhood VV of xx such that f(V∩X)f(V \cap X) is contained in a ball of radius ε\varepsilon; if x∈Xx \in X, this condition expresses the fact that ff is continuous at the point xx (with respect to XX); if x∈X‾∩CXx \in \overline{X} \cap \mathbf{C}X, the image under ff of the trace on XX of the neighbourhood filter of xx in YY is a Cauchy filter base on X′X'; in particular:

Proposition 5

Let ff be a function defined on a subset XX of a topological space YY, taking its values in a complete metric space X′X'. Then ff has a limit relative to XX at a point x∈X‾x \in \overline{X} if and only if the oscillation of ff at xx is zero.

4. Metrizable uniform spaces

Definition 2

A metric on a set XX is said to be compatible with a uniformity UU on XX if the uniformity defined by the metric coincides with UU. A uniformity on a set XX is said to be metrizable if there is a metric on XX compatible with this uniformity. A uniform space is said to be metrizable if its uniformity is metrizable.

Distinct metrics can be compatible with the same uniformity; they are then equivalent (§ 1, no. 2, Definition 2).

Theorem 1

A uniformity is metrizable if and only if it is Hausdorff and the filter of entourages of the uniformity has a countable base.

The condition is necessary, for (with the notation of no. 2) the entourages V1/n(n≥1)V_{1/n} (n \geq 1) form a base of the filter of entourages of the uniformity of a metric space.

The condition is sufficient, for, if it is satisfied, the uniformity under consideration is defined by a single pseudometric, by Proposition 2 of § 1, no. 4; since the uniformity is Hausdorff, this pseudometric is a metric.

Corollary 1

A Hausdorff uniformity defined by a countable family of pseudometrics is metrizable.

For if (fn)(f_n) is a sequence of pseudometrics defining such a structure, the filter of entourages is generated by the countable family of sets f‾n1([0,1/m])\overline{f}_n^1([0, 1/m]), where mm and nn each run through the set of integers >0> 0.

Corollary 2

Every countable product of metrizable uniform spaces is metrizable.

For such a space is Hausdorff and its uniformity has a countable fundamental system of entourages (Chapter II, § 2, no. 6).

5. Metrizable topological spaces

Definition 3

A metric on a set XX is said to be compatible with a topology T\mathscr{T} on XX if the topology defined by this metric coincides with T\mathscr{T}. A topological space is said to be metrizable if there exists a metric on XX compatible with the topology of XX.

Two metrics on a set XX which are both compatible with the same topology T\mathscr{T} can be inequivalent.

The subspace R+∗\mathbf{R}_+^* of R\mathbf{R} provides an example of this. Both the uniformity induced by the additive uniformity of R\mathbf{R} and the uniformity induced by the multiplicative uniformity of R+∗\mathbf{R}_+^* are metrizable and are compatible with the topology of R+∗\mathbf{R}_+^*; but they are not comparable.

We remark also that there can exist non-metrizable uniformities compatible with the topology of a metrizable topological space (Exercise 7).

We shall content ourselves here with necessary conditions for the metrizability of a topological space (for a necessary and sufficient condition, cf. § 4, Exercise 22). In the first place, a space cannot be metrizable unless it is completely regular (indeed we shall see, in § 4, no. 1, Proposition 2, that a metrizable space is necessarily "normal", which is a stronger condition). On the other hand, Theorem 1 shows that:

Proposition 6

Every point of a metrizable space has a countable fundamental system of neighbourhoods.

More generally:

Proposition 7

In a metrizable space, every closed set is the intersection of a countable family of open sets, and every open set is the union of a countable family of closed sets.

Let dd be a metric compatible with the topology of a metrizable space XX. If AA is a closed subset of XX, it is the intersection of the open sets V1/n(A)V_{1/n}(A) [the set of all x∈Xx \in X such that d(x,A)<1/nd(x, A) < 1/n; cf. Proposition 2]. The second part of the proposition follows by taking complements.

Remark 1

These necessary conditions are not sufficient (cf. Exercise 13).

Remark 2

There are spaces in which every point has a countable fundamental system of neighbourhoods but in which there exist closed sets which are not countable intersections of open sets (Exercise 15); such spaces are not metrizable.

Corollary 2 of Theorem 1, no. 4, shows that a countable product of metrizable topological spaces is metrizable. Also the sum XX (Chapter I, § 2, no. 4) of an arbitrary family (Xi)i∈I(X_i)_{i \in I} of metrizable spaces is metrizable. For if did_i is a metric compatible with the topology of XiX_i for each i∈Ii \in I, we may assume that did_i is bounded and that the diameter of XiX_i is ≤1\leq 1; we can then define a distance dd compatible with the topology of XX by putting d(x,y)=di(x,y)d(x, y) = d_i(x, y) if xx and yy both belong to the same XiX_i, and d(x,y)=1d(x, y) = 1 otherwise.

6. Use of countable sequences

Proposition 6 is the origin of the part played by countable sequences of points in the theory of metrizable spaces; for many problems, they can be used to advantage in place of filters. This is because the neighbourhood filters of points of a metrizable space (and therefore also convergent filters) are determined by convergent sequences of points of the space: for since the neighbourhood filter of a point has a countable base, it is the intersection of the elementary filters finer than itself (Chapter I, § 6, no. 8, Proposition 11), i.e., of the elementary filters associated with sequences which converge to the point in question.

On the other hand, the notion of a convergent sequence is not adapted to the study of topological spaces in which there are points whose neighbourhood filter has no countable base. In particular, Hausdorff non-discrete topological spaces can be constructed in which, at every point xx, the intersection of any countable family of neighbourhoods of xx is again a neighbourhood of xx(*); in such a space the only convergent sequences are those in which all the terms are equal from a certain index onwards.

As examples of the use of countable sequences we give the following propositions:

Proposition 8

In a metrizable space* XX, a point xx lies in the closure of a non-empty subset AA of XX if and only if there is a sequence of points of AA which converges to xx.

We know already, from Chapter I, § 7, no. 3, that the condition is sufficient. To see that it is necessary, consider a countable fundamental system (Vn)(V_n) of neighbourhoods of xx such that Vn+1⊂VnV_{n+1} \subset V_n for each nn. If xx lies in the closure of AA then each VnV_n meets AA, and if xnx_n lies in Vn∩AV_n \cap A, the sequence (xn)(x_n) converges to xx(**).

From Proposition 8 we deduce:

Proposition 9

A metric space* XX is complete if and only if every Cauchy sequence in XX is convergent.

Let X^\hat{X} be the completion of XX. If there is a point x∈X^x \in \hat{X} which does not belong to XX, then there is a sequence (xn)(x_n) of points of XX which converges to xx, and this is a non-convergent Cauchy sequence in XX.

Proposition 10

Let* XX be a metrizable space and let ff be a mapping of XX into a topological space X′X'. Then ff is continuous at a point x∈Xx \in X if and only if, whenever (xn)(x_n) is a sequence of points of XX which converges to xx, the sequence (f(xn))(f(x_n)) converges to f(x)f(x) in X′X'.

The condition is necessary, from Chapter I, § 7, no. 4, Proposition 9, Corollary 1. To show that it is sufficient, consider the filter B′\mathscr{B}' of neighbourhoods of f(a)f(a) in X′X'; the hypothesis implies that f‾−1(B′)\overline{f}^{-1}(\mathscr{B}') is coarser than every elementary filter associated with a sequence which converges to aa, that is to say every elementary filter which converges to aa; but

() See e.g., J. Dieudonné, Notes de Tératopologie (I), Revue scientifique (Revue rose), 1939, p. 39. (*) This proposition can still be valid in certain spaces in which at least one point does not have any countable fundamental system of neighbourhoods; for example, the space obtained by compactifying an uncountable discrete space by adjoining a point at infinity (Chapter I, § 9, no. 8, Theorem 4).

the intersection of these filters is the neighbourhood filter of aa (Chapter I, § 6, no. 8, Proposition 11). Hence the result.

Note that Propositions 8 and 10 are valid in any space XX in which every point has a countable fundamental system of neighbourhoods.

7. Semi-continuous functions on a metrizable space

Proposition 11

Let XX be a metrizable space and let ff be a lower semi-continuous function on XX which takes its values in a closed interval [a,b][a, b] of R‾\overline{\mathbf{R}}. Then ff is the upper envelope of an increasing sequence of continuous functions on XX which take their values in [a,b][a, b].

By transfer of structure we may assume that a=0a = 0 and b=1b = 1.

(i) Suppose first that f=φAf = \varphi_A, where AA is an open subset of XX. Then the function gng_n defined by gn(x)=ninf⁡(d(x,X−A),1/n)g_n(x) = n \inf (d(x, X - A), 1/n) is continuous and ≥0\geq 0 on XX; also gn(x)=f(x)g_n(x) = f(x) when x∈X−Ax \in X - A and when

d(x,X−A)≥1/n. d(x, X - A) \geq 1/n.
It follows immediately that f=sup⁡n(gn)f = \sup_n (g_n).

(ii) In the general case consider, for each integer n≥1n \geq 1, the finite decreasing sequence of open sets

Ak=f−1([kn,+∞[)(0≤k≤n−1); A_k = f^{-1}\left( \left[ \frac{k}{n}, +\infty \right[ \right) \quad (0 \leq k \leq n - 1);
the function gn=1n∑k=1n−1φAkg_n = \frac{1}{n} \sum_{k=1}^{n-1} \varphi_{A_k} is lower semi-continuous, and we have 0≤f(x)−gn(x)≤1/n0 \leq f(x) - g_n(x) \leq 1/n for all nn; hence ff is the upper envelope of the sequence (gn)(g_n). On the other hand, gng_n is a linear combination with positive coefficients of a finite number of characteristic functions of open sets and is therefore the upper envelope of a countable sequence (hmn)m≥0(h_{mn})_{m \geq 0} of continuous functions ≥0\geq 0, by (i); hence f=sup⁡m,nhmnf = \sup_{m,n} h_{mn}. If we put
fn=sup⁡p≤n,q≤nhpq, f_n = \sup_{p \leq n, q \leq n} h_{pq},
we see that the sequence (fn)(f_n) is an increasing sequence of continuous functions ≥0\geq 0, with ff as upper envelope, and which do not take the value 1, since gn≤n−1/ng_n \leq n - 1/n.

8. Metrizable spaces of countable type

Definition 4

A metrizable space is said to be of countable type (or separable) if its topology has a countable base.

Clearly every subspace of a metrizable space of countable type is again of countable type. The definition of the base of the topology of a product space (Chapter I, § 4, no. 1) and Corollary 2 to Theorem 1 of no. 4 show that the product of a countable family of metrizable spaces of countable type is a metrizable space of countable type. Again, the sum of a countable family of metrizable spaces of countable type is a metrizable space of countable type (no. 5).

Proposition 12

If X is a metrizable topological space, the following statements are equivalent*:

a) X is of countable type. b) X has a countable dense subset. c) *X is homeomorphic to a subspace of the cube IN\mathbf{I}^{\mathbf{N}}, where I\mathbf{I} is the interval [0,1][0, 1] in R\mathbf{R}.

From the preceding remarks it is clear that c) implies a); a) implies b), for if (Un)(U_n) is a countable base of the topology of X and ana_n is a point of UnU_n, the ana_n form a dense subset of X. Finally let us show that b) implies c). Let (an)(a_n) be a dense sequence of points of X, and for each x∈Xx \in X let φ(x)\varphi(x) be the point (d(x,an))n∈N(d(x, a_n))_{n \in \mathbf{N}} of IN\mathbf{I}^{\mathbf{N}} (dd being a metric compatible with the topology of X, with respect to which the diameter of X is ≤1\leq 1). We shall show that φ\varphi is a homeomorphism of X onto a subspace of IN\mathbf{I}^{\mathbf{N}}. Indeed, φ\varphi is continuous, because each of the functions x→d(x,an)x \to d(x, a_n) is continuous; and φ\varphi is injective, because each point of X is the limit of some subsequence of (an)(a_n) (Proposition 8). Let B be a ball with centre x0x_0 and radius rr in X, and let nn be an integer such that

d(x0,an)<13r. d(x_0, a_n) < \frac{1}{3} r.

The image under φ\varphi of the set W of points x∈Xx \in X such that

∣d(x0,an)−d(x,an)∣<13r |d(x_0, a_n) - d(x, a_n)| < \frac{1}{3} r

is by definition a neighbourhood of φ(x0)\varphi(x_0) in φ(X)\varphi(X). But for each x∈Wx \in W we have d(x,an)<d(x0,an)+13r<23rd(x, a_n) < d(x_0, a_n) + \frac{1}{3} r < \frac{2}{3} r, whence

d(x,x0)≤d(x0,an)+d(x,an)<r, d(x, x_0) \leq d(x_0, a_n) + d(x, a_n) < r,

which shows that W is a neighbourhood of x0x_0 contained in B, and therefore that φ\varphi is a homeomorphism of X onto φ(X)\varphi(X).

Note that for an arbitrary topological space X, property b) does not necessarily imply the existence of a countable base, even if X is compact and every point of X has a countable fundamental system of neighbourhoods (Exercise 13; cf. Chapter I, § 1, Exercise 7).

Proposition 13

Let X be a topological space which has a countable base (Un)(U_n); then for each open covering (Vi)i∈I(V_i)_{i \in I} of X there exists a countable subset J of I such that (Vi)i∈J(V_i)_{i \in J} is a covering of X.

Let H be the subset of N consisting of indices n such that U_n is contained in at least one of the V_i; the sequence (U_n)_{n \in H} is a covering of X, because every point x \in X belongs to some V_i, and since V_i is open, there is an index n such that x \in U_n \subset V_i. Hence there is a mapping ψ\psi of H into I such that U_n \subset V_{\psi(n)} for each n \in H; taking J = ψ(H)\psi(H), which is countable, the proposition is proved.

9. Compact metric spaces; compact metrizable spaces

The criterion for precompactness of a uniform space (Chapter II, § 4, no. 2, Theorem 3) gives rise to the following proposition for metric spaces:

Proposition 14

A metric space X is precompact if and only if, for each ε>0\varepsilon > 0, there is a finite covering of X by sets of diameter ≤ε\leq \varepsilon.

If we adjoin the hypothesis that X is complete we have a criterion for compactness of metric spaces.

From Proposition 14 we get a topological criterion for compactness, applicable to metrizable spaces:

Proposition 15

A metrizable topological space X is compact if and only if every infinite sequence of points of X has a cluster point in X.

Axiom (C) of Chapter I, § 9, no. 1 shows that the condition is necessary. To show sufficiency, let d be a metric compatible with the topology of X. We show first that the metric space X so defined is complete: every Cauchy sequence in X has a cluster point and is therefore convergent (Chapter II, § 3, no. 2, Proposition 5, Corollary 2); hence X is complete, by Proposition 9. Next we shall show that X is precompact; if this were not so, then by Proposition 14 there would exist a real number α>0\alpha > 0 such that X could not be covered by any finite number of subsets of X of diameter ≤α\leq \alpha. We could then define by induction on n an infinite sequence (xn)(x_n) of points of X by the condition d(xp,xn)>12αd(x_p, x_n) > \frac{1}{2} \alpha for all p<np < n; and such a sequence can have no cluster point, since every ball of radius <12α< \frac{1}{2} \alpha contains at most one point of the sequence.

Corollary

A subset A of a metrizable topological space X is relatively compact if and only if every infinite sequence of points of A has a cluster point in X.

Let d be a metric compatible with the topology of X. We shall show that the space A‾\overline{A} is compact, by applying the criterion of Proposition 15. Let (xn)(x_n) be a sequence of points of A‾\overline{A}; then for each index n there exists yn∈Ay_n \in A such that d(xn,yn)<1/nd(x_n, y_n) < 1/n; the sequence (yn)(y_n) has, by hypothesis, a cluster point a∈Xa \in X, and a is also a cluster point of the sequence xix_i, for if ymy_m lies in the ball with centre aa and radius 1/n1/n, for some m>nm > n, then xmx_m lies in the ball with centre aa and radius 2/n2/n.

It should be remarked that Proposition 15 is not a consequence of the existence of a countable fundamental system of neighbourhoods at each point of XX; there are examples of non-metrizable, non-compact spaces in which every point has a countable fundamental system of neighbourhoods and every infinite sequence of points has a cluster point (Exercise 15).

Proposition 16

A compact space XX is metrizable if and only if its topology has a countable base.

The condition is necessary. By Proposition 14, for each integer n≥1n \geq 1 there is a finite subset AnA_n of XX such that the distance of AnA_n from every point of XX is ≤1/n\leq 1/n; the countable set A=⋃nAnA = \bigcup_n A_n is therefore dense in XX, and the result follows from Proposition 12 of no. 8.

The condition is sufficient. Let (Un)(U_n) be a countable base of the topology of XX. Every neighbourhood of a point of the diagonal Δ\Delta of X×XX \times X therefore contains a neighbourhood of the form Un×UnU_n \times U_n; applying the Borel-Lebesgue axiom to the compact subset Δ\Delta of X×XX \times X, it follows that every neighbourhood of Δ\Delta contains a finite union of sets of the form Un∖UnU_n \setminus U_n, which is a neighbourhood of Δ\Delta. Hence the neighbourhoods of Δ\Delta which are finite unions of sets of the form Un×UnU_n \times U_n form a fundamental system of entourages of the uniformity of XX (Chapter II, § 4, no. 1, Theorem 1), and the result therefore follows from Theorem 1 of no. 4.

Corollary

Let XX be a locally compact space and let X′X' be the compact space obtained by adjoining a point at infinity ω\omega to XX (Chapter I, § 9, no. 8). Then the following statements are equivalent:

a) The topology of XX has a countable base. b) X′X' is metrizable. c) XX is metrizable and σ\sigma-compact.

a⟹ba \Longrightarrow b: Let UnU_n be a countable base of the topology of XX. Each neighbourhood of a point x∈Xx \in X contains a compact neighbourhood of xx, which in turn contains a neighbourhood of xx equal to some UnU_n. Hence the relatively compact UnU_n form a base of the topology of XX, and we may therefore suppose that all the UnU_n are relatively compact. XX is therefore a countable union of compact sets U‾n\overline{U}_n, i.e. it is σ\sigma-compact; this implies that in X′X' the point ω\omega has a countable fundamental system VnV_n of open neighbourhoods (Chapter I, § 9, no. 9, Proposition 15, Corollary 2). Hence each neighbourhood of a point y∈X′y \in X' contains either one of the UnU_n or one of the VnV_n, which is a neighbourhood of yy, and so the UnU_n and the VnV_n form a countable base of the topology of X′X'. Hence X′X' is metrizable, by Proposition 16.

b) ⟹c)\Longrightarrow c): If X′X' is metrizable, then ω\omega has a countable fundamental system of neighbourhoods, and therefore XX is σ\sigma-compact by Chapter I, § 9, no. 9, Proposition 15, Corollary 2.

c) ⟹a)\Longrightarrow a): By hypothesis, there is an increasing sequence (Vn)(V_n) of relatively compact open sets which cover XX and are such that V‾n⊂Vn+1\overline{V}_n \subset V_{n+1} (Chapter I, § 9, no. 9, Proposition 15). The subspace V‾n\overline{V}_n is compact and metrizable and therefore has a countable base (Proposition 16), and therefore so does VnV_n. Let (Umn)m≥1(U_{mn})_{m \geq 1} be a base of the topology of VnV_n. For each x∈Xx \in X and each neighbourhood WW of xx, there exists nn such that x∈Vnx \in V_n, hence there exists mm such that x∈Umn⊂Vn∩Wx \in U_{mn} \subset V_n \cap W. Hence the sets UmnU_{mn} (m≥1,n≥1m \geq 1, n \geq 1) form a base of the topology of XX.

10. Quotient spaces of metrizable spaces

If XX is a metrizable space and RR is an equivalence relation on XX, the quotient space X/RX/R is not necessarily metrizable (even if XX is locally compact \* and X/RX/R is normal *). However:

Proposition 17

Every Hausdorff quotient space of a compact metrizable space is compact and metrizable.

Equivalently, if ff is a continuous mapping of a compact metrizable space XX into a Hausdorff space X′X', then f(X)f(X) is a metrizable subspace of X′X' (Chapter I, § 9, no. 4, Theorem 2, Corollary 4).

Let XX be a compact metrizable space, and let RR be an equivalence relation on XX such that X/RX/R is Hausdorff. Then X/RX/R is compact (Chapter I, § 9, no. 4, Theorem 2), hence by Proposition 16 it is enough to show that the topology of X/RX/R has a countable base. To do this, we use the facts that RR is closed (Chapter I, § 10, no. 4, Proposition 8) and that the classes mod RR are compact. Let φ\varphi be the canonical mapping of XX onto X/RX/R, and let (Un)(U_n) be a countable base of the topology of XX. Let zz be any point of X/RX/R and let VV be a neighbourhood of zz in X/RX/R; then φ‾−1(V)\overline{\varphi}^{-1}(V) is a neighbourhood in XX of the compact set φ‾−1(z)\overline{\varphi}^{-1}(z). If xx is any point of φ‾−1(z)\overline{\varphi}^{-1}(z), there is a set UnU_n containing xx and contained in φ‾−1(V)\overline{\varphi}^{-1}(V), and therefore by the Borel-Lebesgue axiom there is a finite open covering (Unk)1≤k≤r(U_{n_k})_{1 \leq k \leq r} of φ‾−1(z)\overline{\varphi}^{-1}(z) such that, if WW denotes ⋃kUnk\bigcup_k U_{n_k}, WW is a neighbourhood of φ‾−1(z)\overline{\varphi}^{-1}(z) contained in φ‾−1(V)\overline{\varphi}^{-1}(V). Since RR is closed, it follows that φ(W)\varphi(W) is a neighbourhood of zz in X/RX/R, contained in VV (Chapter I, § 5, no. 4, Proposition 10). Let B\mathscr{B} denote the set of interiors of sets of the form φ(W)\varphi(W), where WW runs through the set F\mathscr{F} of all finite unions of sets of the form UnU_n; we have then shown that B\mathscr{B} is a base of the topology of X/RX/R, and since F\mathfrak{F} is countable, so is B\mathscr{B}.

Proposition 18

Let XX be a complete metric space, let RR be an open equivalence relation on XX such that X/RX/R is Hausdorff, and let φ:X→X/R\varphi : X \to X/R be the canonical mapping. Then if KK is any compact subset of X/RX/R, there is a compact subset K′K' of XX such that φ(K′)=K\varphi(K') = K.

Let B1\mathscr{B}_1 be the set of all open balls of radius 1/21/2 in XX. As BB runs through B1\mathscr{B}_1, the sets φ(B)\varphi(B) form an open covering of KK, and therefore there exists a finite number of points x1,…,xmx_1, \ldots, x_m of XX such that the images under φ\varphi of the open balls with radius 1/21/2 and centre xix_i (1≤i≤m1 \leq i \leq m) form an open covering of KK. Let H1={x1,…,xm}H_1 = \{ x_1, \ldots, x_m \} and suppose that we have defined a finite set HiH_i, for 1<i≤n1 < i \leq n, such that:

(i) Hi⊂Hi+1H_i \subset H_{i+1} and each point of Hi+1H_{i+1} is at a distance <1/2i< 1/2^i from HiH_i, for 1≤i≤n−11 \leq i \leq n - 1;

(ii) the images under φ\varphi of the open balls of radius 1/2i1/2^i, with centres at the points of HiH_i, form an open covering of KK, for 1≤i≤n1 \leq i \leq n.

Let Bn+1\mathscr{B}_{n+1} be the set of all open balls of radius 1/2n+11/2^{n+1} whose centre xx is such that d(x,Hn)<1/2nd(x, H_n) < 1/2^n (dd being the metric on XX). The properties of HnH_n show that the sets φ(B)\varphi(B), for B∈Bn+1B \in \mathscr{B}_{n+1}, form an open covering of KK; hence there is a finite set Ln+1⊂XL_{n+1} \subset X such that the images under φ\varphi of the open balls of radius 1/2n+11/2^{n+1} whose centre belongs to Ln+1L_{n+1} form an open covering of KK. Taking Hn+1=Hn∪Ln+1H_{n+1} = H_n \cup L_{n+1}, we see that we can define inductively an infinite sequence (Hn)(H_n) of finite subsets of XX with properties (i) and (ii) above. Let H=⋃nHnH = \bigcup_n H_n, and let us show that HH is precompact. For each p>0p > 0 and each point zn+p∈Hn+pz_{n+p} \in H_{n+p}, there exists a sequence of points zn+i∈Hn+iz_{n+i} \in H_{n+i} (0≤i≤p−10 \leq i \leq p - 1) such that

d(zn+i,zn+i+1)<1/2n+ifor 0≤i≤p−1; d(z_{n+i}, z_{n+i+1}) < 1/2^{n+i} \quad \text{for } 0 \leq i \leq p - 1;

it follows that d(zn,zn+p)≤∑i=0p−11/2n+i≤1/2n−1d(z_n, z_{n+p}) \leq \sum_{i=0}^{p-1} 1/2^{n+i} \leq 1/2^{n-1}, and consequently d(y,Hn)≤1/2n−1d(y, H_n) \leq 1/2^{n-1} for all y∈Hy \in H, which proves our assertion. Since XX is complete, H‾\overline{H} is compact, hence φ(H‾)\varphi(\overline{H}) is compact. Next, let us show that K⊂φ(H‾)K \subset \varphi(\overline{H}). If z∈Kz \in K, then by definition d(H‾n,φ‾(z))≤1/2nd(\overline{H}_n, \overline{\varphi}(z)) \leq 1/2^n for all nn, and therefore d(H‾,φ‾(z))=0d(\overline{H}, \overline{\varphi}(z)) = 0; but φ‾(z)\overline{\varphi}(z) is closed and H‾\overline{H} is compact, so that this implies H‾∩φ‾(z)≠∅\overline{H} \cap \overline{\varphi}(z) \neq \emptyset (no. 2, Remark following Proposition 3); hence the assertion. Thus if K′=H‾∩φ‾(K)K' = \overline{H} \cap \overline{\varphi}(K), then K′K' is closed in H‾\overline{H} and therefore compact, and from what has been proved we have φ(K′)=K\varphi(K') = K.

Q.E.D.

Exercises

See the exercises for § 2.

Transcribed from the printed volume. Not translated.