Advances in Pure Mathematics, 2013, 3, 405-408
http://dx.doi.org/10.4236/apm.2013.34058 Published Online July 2013 (http://www.scirp.org/journal/apm)
On the Decomposition of a Bounded Closed
Interval of the Real Line into
Closed Sets
Edgar A. Cohen
Formerly of the Information Sciences Branch, Naval Surface Warfare Center, White Oak, Silver Spring, USA
Email: edcohenfam@yahoo.com
Received March 7, 2013; revised April 13, 2013; accepted May 14, 2013
Copyright © 2013 Edgar A. Cohen. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ABSTRACT
It has been shown by Sierpinski that a compact, Hausdorff, connected topological space (otherwise known as a contin-
uum) cannot be decomposed into either a finite number of two or more disjoint, nonempty, closed sets or a countably
infinite family of such sets. In particular, for a closed interval of the real line endowed with the usual topology, we see
that we cannot partition it into a countably infinite number of disjoint, nonempty closed sets. On the positive side,
however, one can certainly express such an interval as a union of c disjoint closed sets, where c is the cardinality of the
real line. For example, a closed interval is surely the union of its points, each set consisting of a single point being
closed. Surprisingly enough, except for a set of Lebesgue measure 0, these closed sets can be chosen to be perfect sets,
i.e., closed sets every point of which is an accumulation point. They even turn out to be nowhere dense (containing no
intervals). Such nowhere dense, perfect sets are sometimes called Cantor sets.
Keywords: Cantor Sets; Cardinality; Disjoint Closed Sets; Dyadic Representation; Interlacing Dyadic Expansions;
Generators of Sets; Nowhere Dense
1. Introduction
In a paper published in 1918 (see [1]), Sierpinski showed
that a compact, Hausdorff, connected topological space
cannot be partitioned into either a finite number of two or
more pairwise disjoint closed sets or a countably infinite
number of such sets. It follows then that any bounded
closed interval of the real line, in particular, [0,1], cannot
be so decomposed. A fortiori neither can the open inter-
val (0,1) be a union of such sets even though (0,1) is ex-
pressible as a countable union of overlapping closed sets.
In contrast [0,1] can certainly be expressed as the union
of its points, each of which is closed in the usual topol-
ogy (see [2]). Furthermore, there is an exercise given by I.
P. Natanson (see [3]) to show that [0,1] can be partition-
ed into c pairwise disjoint perfect sets. We shall show
that, except for a set of Lebesgue measure 0, [0,1] is, in
fact, decomposable into c pairwise disjoint perfect sets.
These sets also turn out to be nowhere dense. As a corol-
lary to this major result, we can show that there exists a
countable union of pairwise disjoint nowhere dense per-
fect sets which is dense in [0,1] even though this union
cannot be the interval [0,1] itself.
2. Results
Let us first repeat the Sierpinski result (see [1]) in the
form of interest here, namely:
Theorem 1. The closed interval [0,1] (in the usual to-
pology) cannot be partitioned into either a finite (two or
more) or a countably infinite number of pairwise disjoint
closed sets.
Proof. Sierpinski shows that the result is correct for
any compact connected Hausdorff space. Since the usual
topology is a metric topology and the closed interval [0,1]
is connected and, by the Heine-Borel Theorem (see [4]),
is compact, our case is simply a special instance of Sier-
pinski’s general theorem. Thus, our theorem is valid.
Corollary. The open interval (0,1) cannot be partition-
ed into a countable number of pairwise disjoint closed
sets.
Proof. If there were such a partition, one could adjoin
to it the two-point set {0,1} and thus partition the closed
interval [0,1] into a countable number of disjoint closed
sets. That would contradict the result of Theorem 1. So
the corollary follows.
Of course one can express [0,1] as a countable union
C
opyright © 2013 SciRes. APM
E. A. COHEN
406
of overlapping closed sets, e.g.,


3
1,1 1nn0,1
n
(1)
Let us next prove our major result, which parallels a
problem given by I. P. Natanson (see [3]).
Theorem 2. The closed interval [0,1] can be decom-
posed into c nowhere dense perfect sets, all of which, ex-
cept for a countable number, are pairwise disjoint.
Proof. Let x0 be a real number in the closed interval
[0,1], and consider its dyadic representation
0123
0.xi
ii
12
k
kk
k
i
i
 (2)
(so that ik = 0 or 1 for every k). Now for any real number
y likewise in [0,1] with
12
0.yj
j
12
k
kk
k
j
j
 (3)
consider

01122
0. kk
xy ijij ij
 21 2
11
,
22
kk
kk
kk
ij




(4)
which is thus represented by interlacing two dyadic ex-
pansions.
So any x0 with a given dyadic expansion can be con-
strued as a generator of a set {x0(y)}. We shall examine
the nature of these sets with the understanding that x0 as
given by Equation (2) may take on two binary represen-
tations if it has a terminating dyadic expansion. In that
case it will be necessary to use both expansions in the
process of forming Equation (4). Unfortunately this dual
representation will lead to some difficulties with regard
to disjointness of the sets which we generate, but they
will not be critical to the success of our analysis.
We first show that any given set {x0(y)} is closed. To
see this suppose that x0 = 0, the zero generator. We want
to show that any z not belonging to the set X0 = {x0(y)}
cannot be an accumulation point of the set. First let us
observe that the largest element of the set X0 is

00 0.010101 01xy
 2
1
11
,
3
2n
n


1.
n
iz

20.
n
iz
1j
(5)
an interlacing of x0 = 0 with y0 = 1. Thus, if z > 1/3, its
distance from the set is z – 1/3 > 0. So, since 0 belongs to
the set (an interlacing of 0 with itself), one need only
concern himself with those z which happen to lie in the
open interval (0,1/3). Since z does not lie in the set of
interest, clearly at least one of the digits of the generator
for z is not 0. Realizing now that i1(z) = 0, since z < 1/3,
there must exist an n1 such that in(z) = 0 for 1 n < n1 ,but
1 There are now two cases to be considered: (1)
in(z) = 1, n n1 and (2) there exists an n2 > n1 such that
Let us consider Case 1. If it so happens that
jn(z) = 1 for n n1, then z will belong to X0, since z
would have a dyadic expansion equivalent to that of a
real number with a terminating expansion in which
11n
and both in and jn = 0 for n n1. Observe that
such a phenomenon can occur despite the fact that the
generator of z is different from that of x0. However, also
note that such a phenomenon as this just implies that the
point actually belongs to the set, and we are interested at
this point in a z which is not in the set. So, in Case 1,
when z does not belong to X0, there must exist an m1 n1
such that 10.
m
j
Thus we have the following situa-
tion:
11
12
0.000000111 10111,
mm
zjj

 
1,0, 1,.ijinm
(6)
where in = jn = 0 when 1 n < n1, in = jn = 1 when n1 n
< m1, 11 1mmn Next, if 0
x
X
and
x < z, then the largest such x is (L for last and B for be-
fore)
0.000001010101 ,
LB
x

1.
n
j
(7)
with the first 1 in slot Now, from Equation (6),
1
11
11 1
21
2122 222
21 11
11 111
,
3
22 2
22
m
r
ir rnm
i nrmrm
j
zp


 

 
(8)
where p 0 and, from Equation (7),

1
1
221
11 1
.
3
22
LB in
in
x

.
(9)
Then, from Equations (8) and (9), we have

11 1
11
11
11
222 222
222 2
22 22
2
11111
33
22 2
21 11
33
22
11 111
2.
33
22
2
LB nmn
nm
nn
mn
zx p
p
p
 


  
 

 


(10)
We also note that
1
11
222
11
111
0.
3
22
2
r
rrm
rm rm
j
p

 
  

0.000 0001000 000 ,
FA
x
(11)
Now, when x > z, the smallest x in X0 (F for first and A
for after) is

11.
n
j
(12)
with a 1 in position
So, using Equations (8), (11),
and (12), one has
11 1
1
111
22222 2
22
22 22
1111
3
22 2
11
32
11 11
.
3222
FA nn m
m
mmm
zp
p


x
 


 

 


(13)
Copyright © 2013 SciRes. APM
E. A. COHEN 407
Now, using Equations (10) and (13), it obtains that the
distance between z and X0 is greater than or equal to
1
2
12m

1
, which is the minimum of the lower bounds pro-
vided by Expressions (10) and (13). Therefore, in Case
(1), z is not a limit point of X0.
Let us next consider Case (2). Here 1
n
iz
and
there exists n2 > n1 such that 2
n Let us assume
that n1 and n2 are the first indices for which this is valid.
So z now has the form

iz0.
2 2
1
0
n n
jj


1111
12311 1
0.0 001
nnnn
zjjjjjij

(14)
Here xLB is the same as before, as given by Equations
(7) and (9), and, from Equation (14), we have
  
2
1
1
221 2
11
1
22
2
n
r
rn
rr
jz iz
z




12
1 21
1
,
2
r
r
r r
nrn
iz

(15)
so that, using Equations (9) and (15),
11
21 2221
111 1
33
22
LB nn
zx qq


1 1
21
1 11
,
3
2 2
n n 

(16)
where q 0. In a similar manner, using Equations (12)
and (15), we have
11
21 2
11
22
FA nn
zx ss
11
2 2
11
,
22
nn
  (17)
where s 0. It follows that z is at a distance of at least
1
21
11
32n
12 3
0. ,giii
123
0. 000tiii
from X0, and again z cannot be a limit point of
the set. Thus X0 is closed when x0 = 0. Now observe that
every set {x0(y)} is just a translation of the set generated
by 0. In fact, given a generator
(18)
that set is simply translated by the real number
(19)
It follows that all of our sets are closed, and, further-
more, that their union will be [0,1]. The latter claim
stems from the fact that every real number in the closed
interval must have a generator and from the fact that any
real number can be obtained by interlacing some genera-
tor with another binary expansion.
Next we show that each set {x0(y)}, whose elements
are given by Equation (4), is a perfect set, i.e., all points
are points of accumulation. This is easy to see, for let us
consider a particular point x0(y0). If we change the value
of jn and keep all of the other entries the same, we obtain
another point of the same set whose distance from x0(y0)
is 1/22n. The result then follows, since n may be taken
arbitrarily large. Of course, the same result could have
been procured by proving it for the 0 generator case and
then appealing to translation invariance, as we did for
showing that all of our generated sets are closed.
To see that each set is actually nowhere dense, we
need only show that property for the set generated by 0
and then appeal to translation invariance. When x0 = 0, so
that ir = 0 for every r, consider any point x1 belonging to
X0 and another point x2 of that set arbitrarily close to x1.
We may suppose that the two points first differ for some
jn, which is 0 for x1 and 1 for x2. Then, if one modifies x1
by changing any ir for r > n from 0 to 1, he obtains a
point x3 of the complement of X0 lying between x1 and x2 .
It follows that X0 is nowhere dense and, by translation
invariance, that all of our sets are nowhere dense. So our
sets are actually Cantor sets, or Cantor-like sets, as some
might say.
We next note that our defining process itself shows
that, in general, any two sets, say {x0(y)} and {x1(y)},
will be disjoint, since they are constructed using different
generators. Also, clearly each set contains c points, since
it is built by interleaving all real numbers in [0,1] with a
given generator. Furthermore, there will be c sets, since
there are c generators.
However, there are a certain number of situations in
which our sets need not be disjoint, and these cases arise
only because rational numbers with terminating expan-
sions have two dyadic representations. Let us observe the
nature of these Cantor sets. Let us consider, for example,
the generators




0
0121
1
0121
2
0121
3
0121
0121
0.1000 000 000
0.0111 111 111
0.00111 111 111
0.000111 111 111
0.000 000 000
n
n
n
n
n
xiii
xiii
xiii
xiii
xiii
 
 
 
 

()i
(20)
where 0
x
, 1 i < , has i zeroes before an infinite ter-
minal string of 1’s and 0

x
is a finite expansion (the 0
expansion included). Then the following situations occur:


0
0y


and
1
0y


0
i
overlap, and
x
x
x
y


over-
0y



0
i
for 1 i < . However,
x
x
y will laps
be disjoint from any set whose generator is not expressi-
ble as a terminating dyadic expansion. The collection of
all such sets is clearly a countable one, and all remaining
sets, c in number, are pairwise disjoint. Therefore, except
for a countable family of Cantor sets, each clearly of the
same Lebesgue measure (since they are just translates of
one another), the remaining Cantor sets (constituting a
disjoint family of c sets) form a decomposition of [0,1].
This proves the theorem.
It is interesting that almost all of the Cantor sets de-
vised here consist only of irrational numbers as would, of
course, naturally follow if there are going to be c of them.
One will observe that, if the generator is irrational, then
Copyright © 2013 SciRes. APM
E. A. COHEN
Copyright © 2013 SciRes. APM
408
any number produced through interlacing will again be
irrational. In fact rational numbers are only produced
when a rational generator is interlaced with another ra-
tional number.
One also has the following corollary (see [5]):
Corollary 1. There is a subset X of [0,1] which is a
union of a countable number of nowhere dense perfect
sets and which is dense in [0,1].
Proof. Let X be the union of all Cantor sets whose
generators are of the form given in Equations (20). Sup-
pose that z is any real number which is not an element of
X. Let us form its dyadic representation, which, of course,
is unique, since it does not belong to X. Consider a fun-
damental neighborhood, or ball, about z. Then it is clear
that some truncation of its expansion will reside both
within the ball and within the set X. Since X is a count-
able union of nowhere dense, perfect sets, this proves the
corollary.
With just a little extra effort, one can even assert that
such sets can be chosen to be pairwise disjoint. In order
to achieve that, we can use the sets whose generators are
rational but not realizable as terminating expansions. The
union of all such sets (countable in number) is also dense
in [0,1] as the reader will easily see. In addition one has
the following result:
Corollary 2. All of our Cantor sets are of Lebesgue
measure 0. Therefore, X, of Corollary 1, is also of meas-
ure zero.
Proof. Our set X and the remaining Cantor sets de-
vised in the proof of Theorem 2 constitute c pairwise
disjoint sets whose union is the closed interval [0,1].
Now we can show that at least one of these sets must be
of zero Lebesgue measure. Suppose that this is not the
case. Let An be the collection of those sets whose meas-
ures exceed or equal 1/n. Since all sets are disjoint from
one another, clearly there are no more than n of them in
An. Of course, there are a countable number of such sets,
each of which can contain no more than a finite number
of sets of the collection. However, then we would have a
countable collection of such sets covering [0,1], and we
know that we have c of these sets whose union is [0,1].
Therefore, there must exist a set of measure zero among
them. If this set happens to be X, then any subset of it
also has measure 0, and, in particular, this is true of any
one of the Cantor sets. Since all of the Cantor sets are
translates of one another, we would conclude that all of
them have zero measure, On the other hand, if one of the
other sets has measure zero, then again all Cantor sets are
translates of it, and all our sets are of zero measure. This
establishes the corollary.
Corollary 3. Any nowhere dense perfect set C can it-
self be decomposed into c pairwise disjoint nowhere dense
perfect sets.
Proof. Since the set C is assumed to be a Cantor set,
we can use a well-known result originally due to L. E. J.
Brouwer, namely, that all Cantor sets are homeomorphic
to one another. This means that we can invoke the clas-
sical Cantor ternary set (see [3]) as a prototype. Since
only 0’s and 2’s are necessary to describe that set, this
means that there is a unique encoding of each point of
our given Cantor set C. It follows that our generator
scheme can now be utilized (using 0’s and 2’s rather than
0’s and 1’s) to obtain c pairwise disjoint nowhere dense
perfect subsets of C. In other words, we can obtain first
of all a partition of the classical Cantor set into c pairwise
disjoint Cantor subsets each of measure 0, and their im-
ages under a homeomorphism will also be Cantor sets
disjoint from one another. Therefore, we have a partition
of our original Cantor set C into c pairwise disjoint Can-
tor subsets. However, under a homeomorphism, measure
need not be preserved, so that these subsets need not all
be of measure 0.
3. Conclusion
We have established two important facts concerning the
decomposition of a closed interval of the real line into
disjoint closed sets: 1) Such a partition cannot be ob-
tained using either a finite (more than 1) or a countably
infinite number of closed sets, and, a fortiori, neither can
an open interval be so decomposed. The latter fact does
not seem to be generally emphasized in the literature. On
the other hand, it is easy to show that an open interval is
an Fσ set (a countable union of closed sets) if we drop the
requirement that the sets be pairwise disjoint; 2) In con-
trast a partition into c pairwise disjoint closed sets (with c
the power of the continuum) is always possible and, in
fact, except for a set of Lebesgue measure 0, that can be
accomplished with perfect sets (closed sets without iso-
lated points). It turns out that, in the theory that we have
developed here, the sets involved are also nowhere dense,
i.e., have a vacuous interior. Perfect, nowhere dense sets
are sometimes called Cantor sets.
REFERENCES
[1] W. Sierpinski, “Un Theoreme sur les Continus,” Tohoku
Mathematical Journal, Vol. 13, 1918, pp. 300-305.
[2] G. F. Simmons, “Introduction to Topology and Modern
Analysis,” McGraw Hill Book Company, Inc., New York,
1963, p. 92.
[3] I. P. Natanson, “Theory of Functions of a Real Variable,”
Frederick Ungar Publishing Co., Inc., New York, 1961,
pp. 49-50.
[4] D. W. Hall and G. L. Spencer II, “Elementary Topology,”
John Wiley and Sons, Inc., New York and Chapman &
Hall, Limited, London, 1955, p. 44.
[5] M. E. Munroe, “Introduction to Measure and Integration,”
Addison-Wesley Publishing Co., Inc., Reading, 1953, p. 70.