MORE GENERAL TOPOLOGY

1. The Cantor-Bendixson Theorem.

This probably doesn’t deserve a § all to itself, but it will be handy to know it in what follows. Recall
that a set S in a metric (or topological) space X is dense-in-itself 1 if every neighborhood of each s0 ∈ S
contains points s ∈ S different from s0 . One says that S is perfect if it is closed and dense-in-itself. A
topological (or metric) space is separable if it has a countable subset whose closure is the whole space.
Recall that for metric spaces, being separable is equivalent to being second countable: there is a countable
family B of open sets such that every open set is the union of the elements of B that it contains. This
property is inherited by subsets of these spaces (in their relativized metric).
Let (X, d) be a separable metric space in what follows, and let ∅ =
6 S. A point x ∈ X is a condensation
point of S if every neighborhood of x contains uncountably many points of S. Evidently every condensation
point of S belongs to S, and moreover
Proposition: Every uncountable set S in a separable metric space (X, d) contains a condensation point
of itself.
Proof. The property of being separable is inherited by subspaces, so (S, d|(S × S)) has a countable
family B = {Bn }∞ n=1 of open sets such that every open set is the union of the elements of B that it contains.
If no point of S is a condensation point of S, then every s ∈ S has a neighborhood containing only
S∞ countably
many points of S and therefore belongs to a Bn for which Bn ∩ S is countable. But then S = n=1 (Bn ∩ S)
is a countable union of countable sets and thus is countable.
Proposition: The set S c of condensation points of S ⊆ X is perfect.
Proof. It’s obvious that S c is closed. To see that S c is dense in itself, let xS∈ S c , let  > 0 be given, and
∞
write the punctured ball B(x; ) \ {x} as a countable union: B(x; ) \ {x} = n=1 [B(x; /n) \ B(x; /(n +
S∞ The uncountable set [B(x; ) \ {x}] ∩ S is thus written as a countable union [B(x; ) \ {x}] ∩ S =
1))].
n=1 {[(B(x; /n) \ B(x; /(n + 1))] ∩ S}. At least one of the uniands must be uncountable and therefore
contain a condensation point sn of itself: but that point is a fortiori a condensation point of S, so we have
exhibited sn ∈ [B(x; ) \ {x}] ∩ S c .
Lemma: If (X, d) is a countable complete metric space, then it is the closure of its set of isolated points.
Proof.S The finite case is obvious. In the infinite case, enumerate X = {xn }∞ n=1 . By the Baire category
theorem, ∞ n=1 {x n } ◦
is dense, and {x n } ◦
is nonempty if and only if x n is an isolated point of X.
Theorem [G. Cantor-I. Bendixson]: Every complete separable metric space (X, d) is the union of the
perfect set of its condensation points and a countable set in which the isolated points form a dense subset.
(Either of these subsets may be empty.)
Proof. As we just saw, the set X c of all condensation points of X is perfect. Its complement X \ X c
cannot be uncountable, for then it would contain a condensation point of itself which would of course belong
to X c , a contradiction. X c is closed, so its complement is open and therefore complete-metrizable: the Baire
category theorem holds for X \ X c . The lemma just proved tells us that X \ X c is the union of its isolated
points, and because X \ X c is open, a point that is isolated in this set is isolated in the big space X.
Various extreme cases are possible in the setting of the Cantor-Bendixson theorem: either or both sets
may be empty, the set of isolated points may be closed, or its closure may be disjoint from X c , contain some
points of X c , or contain X c . These possibilities are easily realized in the Euclidean plane, and the reader
should construct examples.

2. [The] Cantor Set(s).

In an earlier set of notes we defined the Cantor set to be the compact metric space 2N , the metric-space
product of a countable family of copies of 2N each equipped with the discrete metric. To prevent confusion it
1
Originally German in sich dicht; sometimes these words were run together. An echo of Kant’s Ding an sich?

1 Math 501:01 Fall 1999

may be convenient to call this metric space the Cantor space rather than Cantor set; then we can reserve
the latter name for certain subsets of the real line.
The customary construction of the Cantor middle-third set in R1 is given by Wheeden & Zygmund,
pp. 34–35. We can abstract from the construction slightly without changing things except in small detail,
as follows. Let a closed interval I ⊆ R—that does not reduce to a single point—be given, along with a
sequence {rn }∞ n=1 of real numbers, 0 < rn < 1 for each n ∈ N. For each n ∈ N we will construct a family
Cn of 2n pairwise-disjoint closed subintervals of I, indexed by the integers 0, . . . , 2n − 1, as follows. For a
given n, write these integers in binary notation, so each is a sequence of n elements of 2. For n = 1: remove
the open middle r1 -th of I, and label the left-hand closed interval thus produced as I0 and the right-hand
closed interval as I1 . Given Cn , to produce Cn+1 : for each interval Ib1 ··· bn ∈ Cn , remove the open middle
rn -th of this interval, and label the left-hand closed interval
S thus produced
S as Ib1 ··· bn 0 and the right-hand
closed interval as Ib1 ··· bn 1 .S It is clear inductively that Cn+1 ⊂ Cn ⊂ · · · ⊂ I and that if m denotes
Lebesgue measure then m( Cn ) = m(I) · [(1 − rn ) · · · (1 − r1 )]. Consequently, by the “monotone continuity”
∞ h[
\ i ∞
Y
of Lebesgue measure, the set C = Cn has Lebesgue measure (1 − rj ) = lim (1 − rn ) · · · (1 − r1 ).
n→∞
n=1 j=1

By choosing the rn ’s inductively, we can make m(C) come out equal to any preassigned number r ∈ [0, 1):
the sequence of partial products can be made to converge down from (1 − r1 ) to r simply by taking any r1
with 1 − r1 > r and then at each step choosing rn+1 such that (1 − rn+1 ) · [(1 − rn ) · · · (1 − r1 )] is, say, twice
as close to r as the preceding partial product. For those who want to do things all at once: it is easy to see
∞
Y P
that the product (1 − rj ) “diverges to zero” if and only if rj = ∞ and converges otherwise. {On one
j=1
hand,Qthe relation 1 + x ≤ eP
x
, valid for all x ∈ R as the second-degree Taylor remainder term shows, implies
n n
that j=1 (1 − rj ) ≤ exp(− j=1 rj ), so divergence of the series implies that the partial products go to zero.
P∞
On the other hand, if the series converges then for sufficiently large N one has j=N+1 rj < 1, and so the
inductively clear relation

X
N+p ∞
X
(1 − rN+p ) · · · (1 − rN+1 ) ≥ 1 − rj > 1 − rj > 0
j=N+1 j=N+1

shows that the sequence of partial products—which is monotone-decreasing and positive and must therefore
converge—cannot converge to zero.} Wheeden & Zygmund’s Problem 5, p. 48 gives one set of choices; Euler’s
product
Y∞  
sin πx x2
= 1− 2
πx n=1
n

also gives ready-made choices of x ∈ (0, 1) which for rn = x2 /n2 yield any measure one wants.
All these Cantor sets are topologically the same,2 because for any such set C
Proposition: The mapping 2N → C defined (with some initial abuse of language) by
∞
\
(b1 , b2 , . . . , bn , . . . ) 7→ Ib1 b2 ··· bj
j=1

is 1-1 onto and continuous in both directions.

Proof. The abuse of language comes about because the r. h. s. of the defining relation is a subset of
C rather than a point. However, since diam(Ib1 b2 ··· bj ) < diam(I)/2n , the diameter of the intersection is
zero (so it contains at most one point), and by the Cantor intersection theorem (or the compactness of I)
2
On the line, this statement is pretty much geometrically obvious. A later characterization of the Cantor space will be somewhat
less obvious.

2 Math 501:01 Fall 1999

the intersection T∞ is nonempty. The image of (b1 , b2, . . . , bn , . . . ) is then defined, formally, to be the unique
element of j=1 Ib1 b2 ··· bj . That this function is well defined is clear. Moreover, since the image of any
point in 2N whose first n coördinates are a fixed n-tuple (b1 , . . . , bn ) lies in the set Ib1 , ···, bn with diameter
< diam(I)/2n and any two points of 2N whose distance in the product metric is < 1/2n must agree in
their first n coördinates, the function is continuous. The function is 1-1 because if (b1 , . . . , bn , . . . ) and
(c1 , . . . , cn , . . . ) are different points of 2N , then there will be a first coördinate in which they differ: if it
is the (n + 1)-st coördinate with (say) bn+1 = 0 and cn+1 = 1, then their images will lie in the respective
sets Ib1 ··· bn 0 and Ib1 ··· bn 1 , and since these sets are disjoint, the images are different. The continuity of the
inverse function then follows from the well-known theorem3 that a 1-1 continuous mapping of one compact
metric space onto another necessarily has a continuous inverse. (By examining the geometry of the situation,
it is not difficult to see directly that the inverse mapping is continuous.)
Corollary: All Cantor sets in R1 constructed as described above are homeomorphic; indeed, all are
homeomorphic to the Cantor space 2N .
It is amusing to observe that the mappings of the Cantor space described above are also order-preserving
if the Cantor space is ordered lexicographically (i.e., one element of 2N precedes another if at the first index
for which the coördinates differ, the one has a 0 and the other has a 1). It may offer further amusement to
observe that the explicit formula

X∞
πn (b)
b = (b1 , b2 , . . . , ) 7→ 2 ·
n=1
3n

maps 2N onto the Cantor middle third set, and that similarly

X∞
πn (b)
b = (b1 , b2 , . . . , ) 7→
n=1
2n

gives the Cantor-Lebesgue function of Wheeden & Zygmund, p. 35 when 2N is identified with the Cantor
middle third set and the function restricted to it (the “Tietze extension theorem” of their Problem 18, p. 14
[or §3 below] then gives the whole Cantor-Lebesgue function). This is their Problem 2, p. 47.
The Cantor space can be characterized up to homeomorphism by a simple topological description.4
Definition: A metric (or topological) space (X, d) is said to be zero-dimensional or totally discon-
nected if for every point x ∈ X and neighborhood U of x, there is a neighborhood V with x ∈ V ⊆ U such
that V is both open and closed.
Proposition: If (X, d) is a perfect, totally disconnected compact metric space, then there is a 1-1
mapping of 2N onto X that is continuous in both directions.
Lemma: If (X, d) is a perfect, totally disconnected compact metric space and ∅ 6= V ⊆ X is both open
and closed, then V = V1 ∪ V2 where the Vj ’s are disjoint and nonempty and each is open and closed.
Proof. The Cantor-Bendixson theorem implies that every point of X is a condensation point, so V
contains more than one point (indeed, uncountably many points). Let x1 and x2 be distinct points of V , let
U be a neighborhood of x1 that does not contain x2 , and take open-and-closed V1 with x ∈ V1 ⊆ U . Then
V2 = V \ V1 is the “V2 ” required for the lemma.
Proof of the proposition. Since X is compact and every ball contains an open-and-closed set containing
its center, one can cover X with finitely many open-and-closed sets W0 , . . . , Wk−1 , each of diameter < 1.
By disjointizing these sets in the usual way if necessary, we can assume that they are disjoint (the results
of disjointizing open-and-closed sets by replacing W1 by W1 \ W0 , etc., discarding any empty sets that
3
The reader should check this if (s)he has never seen it before.
4
I believe this result is originally due to F. Hausdorff; see his M engenlehre, 3rd rev. ed., Dover n.d., p. 160, where he gives no
reference for the result.

3 Math 501:01 Fall 1999

occur, will again be open-and-closed sets). If the cardinality of {W0 , . . . , Wk−1 } is not a power of 2, we
can use the lemma to divide one or more of the Wj ’s into two nonempty disjoint open-and-closed sets until
the cardinality of the collection is a power of 2, say 2n1 , and we can index these by binary numbers of
n1 digits, say {U (0 · · · 0), . . . , U (1 · · · 1)}. We now repeat this process on each of the U ’s, so that each
is written as a disjoint union of the same number 2n2 −n1 of nonempty open-and-closed sets of diameter
< 1/2, each indexed by the string of n1 binary digits that indexed U (b1 · · · bn1 ) followed by the string of
n2 − n1 binary digits that indexes it as one of the 2n2−n1 open-and-closed sets into which U (b1 · · · bn1 )
was divided. Proceeding in this fashion, we can assign to each element (b1 , b2 , . . . ) of 2N a decreasing
sequence U (b1 · · · bn1 ) ⊃ U (b1 · · · bn1 bn1 +1 · · · bn2 ) ⊃ · · · ⊃ U (b1 · · · bnj ) ⊃ · · · (the running index is “j”)
of nonempty open-and-closed subsets of X such that the diameter of the j-th set is < 1/j. As in the case of
the Cantor sets in R1 , we can define a mapping from 2N to X by sending
2N → X
∞
\
(b1 , b2 , . . . ) 7→ U (b1 · · · bnj ) .
j=1

Because the diameters of the U ’s decrease to zero and the U ’s are closed, each of those intersections is a
single point and the mapping is well defined. The mapping is onto because each x ∈ X is contained in a
uniquely determined U (b1 · · · bnj ) at each step of the construction, and the mapping is continuous because if
the first nj digits of U (b1 · · · bnj · · · ) are given, the position of the corresponding element of X is determined
to within distance 1/j. It follows that this mapping, as a 1-1 continuous mapping of one compact metric
space onto another, is a homeomorphism of 2N onto X as desired.

3. Other Compact Metric Spaces Viewed from 2N .

A small modification of the construction that characterized 2N at the end of the preceding § shows that
every compact metric space is the image of 2N under a continuous mapping.
Proposition: If (X, d) is a compact metric space, then there is a continuous mapping of 2N onto X.
Proof. Since X is compact, one can cover X with an indexed family of finitely many closed (but probably
not disjoint) sets F0 , . . . , Fk−1, each of diameter < 1. If the cardinality of the set of indices is not a power
of 2, then we can repeat one or more of these sets in the indexed family so that the number of indices is a
power of 2, say 2n1 , and we can then index by binary numbers of n1 digits, say {F (0 · · · 0), . . . , F (1 · · · 1)}.
We now repeat this process on each of the F ’s, so that each is written as a union of an indexed family
with the same number 2n2 −n1 of nonempty closed sets of diameter < 1/2, each indexed by the string of
n1 binary digits that indexed F (b1 · · · bn1 ) followed by the string of n2 − n1 binary digits that indexes
it as one of the 2n2 −n1 closed sets into which F (b1 · · · bn1 ) is divided (possibly with repetitions and/or
overlapping). Proceeding in this fashion, we can assign to each element (b1 , b2 , . . . ) of 2N a decreasing
sequence F (b1 · · · bn1 ) ⊃ F (b1 · · · bn1 · · · bn2 ) ⊃ · · · ⊃ F (b1 · · · bnj ) ⊃ · · · (the running index is “j”) of
nonempty closed subsets of X such that the diameter of the j-th set is < 1/j. As in previous situations, we
can define a mapping from 2N to X by sending
2N → X
∞
\
(b1 , b2 , . . . ) 7→ F (b1 · · · bnj ) .
j=1

Because the diameters of the F ’s decrease to zero and the F ’s are closed, each of those intersections is a
single point and the mapping is well defined. The mapping is onto because each x ∈ X is contained in at
least one F (b1 · · · bnj ) at each step of the construction, and the mapping is continuous because if the first
nj digits of F (b1 · · · bnj · · · ) are given, then the position of the corresponding element of X is determined
to within distance 1/j. In this situation one cannot assert that the mapping is 1-1,5 but it certainly sends
2N onto X as desired.
5
In general, of course, it cannot be 1-1, because it would be a homeomorphism! It is easy to construct concrete examples of this
mapping: for example, send each element of 2N to the element of the unit interval with the corresponding binary expansion. Each
dyadic rational except 0 and 1 will have exactly two preimages.

4 Math 501:01 Fall 1999

 The existence of these mappings has some entertaining consequences. For example, since every compact
metric space is an image of 2N , the category of compact metric spaces (and continuous mappings) is a
“small category” in the sense that there is a set (as opposed to a class) of objects such that every object is
isomorphic to one of them. 6 Similarly, if ϕ : 2N → X is a continuous mapping onto (X, d) then the mapping

f → f ◦ϕ
C(X, R) → C(2N , R)

is an isometric linear mapping of the Banach space C(2N , R) onto a (closed) vector subspace of C(2N , R).
As we shall observe in a moment, there is also an isometric linear mapping of C(2N , R) onto a closed vector
subspace of C([0, 1], R), so these two spaces of continuous functions have the property that each “contains a
copy” of every space C(X, R) where (X, d) is a compact metric space.7
Lemma: Let (X, k · kX ) be a normed vector space and let C ⊆ I be a closed subset of an interval
I ⊆ R. There is an isometric linear mapping sending the vector space of continuous functions f : C → X in
its “uniform norm” kfk = sup{kf(t)kX : t ∈ I} into the vector space of continuous functions f : I → X in
its “uniform norm.” If K is a convex set in X that contains the range of f : C → X, then the range of f as
extended to I is contained in K.
Proof.8 Suppose first that the endpoints of I belong to C. Then I \ C is a disjoint union of open
subintervals of I whose endpoints lie in C. Extend f to each subinterval (a, b) ⊆ I \C by linear interpolation,
i.e., by setting
b−t t−a
f(t) = f(a) + f(b) for a < t < b .
b−a b−a
This gives a well-defined extension of f to I. It is obvious that the extension depends linearly on f. Because
the coefficients (b − t)/(b − a) and (t − a)/(b − a) are nonnegative and their sum is 1, any convex K ⊆ X
that contains the image f[C] of f on C will contain the range of the extension. In particular, since the ball
{x : kxkX ≤ r} whose radius r = sup{kf(t)kX : t ∈ C} is convex, the range of the extended f is contained
in this ball and so the supremum of the norms of the f(t)’s is not increased. We still have to show that the
extended f is continuous at each point z ∈ I. Continuity (or one-sided continuity respectively) is clear if z
is a point of I \ C or an endpoint of a component open interval of I \ C respectively, since then the extended
f is just a first-degree polynomial in a neighborhood of z or a one-sided neighborhood of z, respectively. If
the interval [z, z + δ] contains points of C different from z for arbitrarily small δ > 0, then given  > 0 we
may find δ > 0 for which z + δ ∈ C and such that t ∈ C ∩ [z, z + δ] ⇒ kf(t) − f(z)kX < . But then for any
component interval (a, b) of I \ C contained in [z, z + δ] we have, for any t ∈ (a, b),9

b−t t−a
kf(t) − f(z)kX = f(a) + f(b) − f(z) = b − t [f(a) − f(z)] + t − a [f(b) − f(z)
b − a b−a b−a b−a
X X
b−t t−a b−t t−a
≤ kf(a) − f(z)kX + kf(b) − f(z)kX < + =
b−a b−a b−a b−a

so the extended f is continuous from the right at t = z, and similarly (under corresponding hypotheses) it
will be continuous from the left. If the left or right endpoint of C does not happen to be an endpoint of I,
we simply extend f to be constant on the left or right of C respectively, and continuity and the norm are
again easily seen to be preserved.

6
Cf. the finitely generated modules over a ring.
7
In the presence of the Hahn-Banach theorem, this construction can be used to show that each of C([0,1],R) and C(2N ,R) contains
an isometric copy of every separable normed space.
8
This is the proof of the Tietze extension theorem that we hoped for as a solution to Wheeden & Zygmund’s Problem 18, p. 14,
extended a bit.
9
Really all the following computation does is use the fact that the ball of radius  and center f(z) in X is convex.

5 Math 501:01 Fall 1999

 Proposition [“Space-Filling Curves”] Let K be a compact convex set in a Banach space. Then there
is a continuous function f : [0, 1] → K that sends the unit interval of R onto K.
Proof. If C is the usual Cantor middle-third set, then we know there is a continuous mapping f : C → K
of C onto K (equipped with the relativized metric). The lemma we just proved allows us to extend f to a
continuous mapping f : [0, 1] → K which is “even more onto than it was.”
Thus it is possible to “scribble very fast” and completely shade the closed unit ball in any Rn . By
contrast, one has the following proposition, whose proof is an exercise for the reader:
Proposition: Let f : [a, b] → Rn be a curve of finite length. Then if n > 1 the n-dimensional Lebesgue
measure of its carrier, m(f[[a, b]]) = 0.
For n ≥ 2, it is thus impossible to cover the unit ball in Rn with a curve of finite length.

6 Math 501:01 Fall 1999