You are on page 1of 2

182 III.

Mathematical Concepts

U1 , . . . , Un , you can find a finite collection of open sets


V1 , . . . , Vm with the following properties:

(i) the sets Vi also cover the whole of X;


(ii) every Vi is a subset of at least one Ui ;
(iii) no point is contained in more than d + 1 of the Vi .

If X is a metric space, then we can choose our Ui to have


small diameter, thereby forcing the Vi to be small. So
this definition is basically saying that it is possible to
cover X with open sets with no d + 2 of them overlap-
ping, and that these open sets can be as small as you
like.
Figure 1 How to cover with squares
We then define the topological dimension of X to
so that no four overlap.
be the smallest d such that X is at most d dimen-
sional. And again it can be shown that this definition
of disjoint closed sets A and B, you can find disjoint assigns the “correct” dimension to the familiar shapes
open sets U and V with A ⊂ U and B ⊂ V such that of elementary geometry.
the complement Y of U ∪ V (that is, everything in X A fourth intuitive idea leads to concepts known as
that does not belong to either U or V ) has dimension homological and cohomological dimension. Associated
at most d − 1. The set Y is the barrier—the main differ- with any suitable topological space X, such as a man-
ence is that we have now asked for it to be closed. The ifold, are sequences of groups known as homology
induction starts with the empty set, which has dimen- and cohomology groups [IV.6 §4]. Here we will dis-
sion −1. Brouwer’s definition is known as the inductive cuss homology groups, but a very similar discussion
dimension of a set. is possible for cohomology. Roughly speaking, the nth
Here is another basic idea that leads to a useful def- homology group tells you how many interestingly dif-
inition of dimension, proposed by lebesgue [VI.72]. ferent continuous maps there are from closed n-dimen-
Suppose you want to cover an open interval of real sional manifolds M to X. If X is a manifold of dimension
numbers (that is, an interval that does not contain its less than n, then it can be shown that the nth homology
endpoints) with shorter open intervals. Then you will be group is trivial: in a sense, there is not enough room in
forced to make the shorter ones overlap, but you can X to define any map that is interestingly different from
do it in such a way that no point is contained in more a constant map. On the other hand, the nth homology
than two of your intervals: just start each new interval group of the n-sphere itself is Z, which says that one
close to the end of the previous one. can classify the maps from the n-sphere to itself by
Now suppose that you want to cover an open square means of an integer parameter.
(that is, one that does not contain its boundary) with It is therefore tempting to say that a space is at least
smaller open squares. Again you will be forced to make n dimensional if there is room inside it for interest-
the smaller squares overlap, but this time the situation ing maps from n-dimensional manifolds. This thought
is slightly worse: some points will have to be contained leads to a whole class of definitions. The homological
in three squares. However, if you take squares arranged dimension of a structure X is defined to be the largest
like bricks, as in figure 1, and expand them slightly, n for which some substructure of X has a nontrivial
then you can do the covering in such a way that no nth homology group. (It is necessary to consider sub-
four squares overlap. In general, it seems that to cover structures, because homology groups can also be trivial
a typical d-dimensional set with small open sets, you when there is too much room: it then becomes easy to
need to have overlaps of d + 1 sets but you do not need deform a continuous map and show that it is equiva-
to have overlaps greater than this. lent to a constant map.) However, homology is a very
The precise definition that this leads to is surpris- general concept and there are many different homology
ingly general: it makes sense not just for subsets of Rn theories, so there are many different notions of homo-
but even for an arbitrary topological space [III.90]. logical dimension. Some of these are geometric, but
We say that a set X is at most d dimensional if, how- there are also homology theories for algebraic struc-
ever you cover X with a finite collection of open sets tures: for example, using suitable theories, one can
III.17. Dimension 183

define the homological dimension of algebraic struc- remove the middle thirds of these two closed intervals
tures such as rings [III.81 §1] or groups [I.3 §2.1]. This to produce a set X2 , so X2 is the union of the intervals
1 2 1 2 7 8
is a very good example of geometrical ideas having an [0, 9 ], [ 9 , 3 ], [ 3 , 9 ], and [ 9 , 1].
algebraic payoff. In general, Xn is a union of closed intervals, and Xn+1
Now let us turn to a fifth and final (for this article at is what you get by removing the middle thirds of each
least) intuitive idea about dimension, namely the way it of these intervals—so Xn+1 consists of twice as many
affects how we measure size. If you want to convey how intervals as Xn , but they are a third of the size. Once you
big a shape X is, then a good way of doing so is to give have produced the sequence X0 , X1 , X2 , . . . , you define
the length of X if X is one dimensional, the area if it is the Cantor set to be the intersection of all the Xi : that
two dimensional, and the volume if it is three dimen- is, all the real numbers that remain, no matter how
sional. Of course, this presupposes that you already far you go with the process of removing middle thirds
know what the dimension is, but, as we shall see, there of intervals. It is not hard to show that these are pre-
is a way of deciding which measure is the most appro- cisely the numbers whose ternary expansions consist
priate without determining the dimension in advance. just of 0s and 2s. (There are some numbers that have
Then the tables are turned: we can actually define the two different ternary expansions. For instance, 13 can
dimension to be the number that corresponds to the be written either as 0.1 or as 0.02222 . . . . In such cases
best measure. we take the recurring expansion rather than the ter-
To do this, we use the fact that length, area, and vol- minating one. So 13 belongs to the Cantor set.) Indeed,
ume scale in different ways when you expand a shape. when you remove middle thirds for the nth time, you
If you take a curve and expand it by a factor of 2 (in all are removing all numbers that have a 1 in the nth place
directions), then its length doubles. More generally, if
after the “decimal” (in fact, ternary) point.
you expand by a factor of C, then the length multiplies
The Cantor set has many interesting properties. For
by C. However, if you take a two-dimensional shape and
example, it is uncountable [III.11], but it also has mea-
expand it by C, then its area multiplies by C 2 . (Roughly
sure [III.55] zero. Briefly, the first of these assertions
speaking, this is because each little portion of the shape
follows from the fact that there is a different element
expands by C “in two directions” so you have to mul-
of the Cantor set for every subset A of the natural num-
tiply the area by C twice.) And the volume of a three-
bers (just take the ternary number 0.a1 a2 a3 . . . , where
dimensional shape multiplies by C 3 : for instance, the
ai = 2 whenever i ∈ A and ai = 0 otherwise), and there
volume of a sphere of radius 3 is twenty-seven times
are uncountably many subsets of the natural numbers.
the volume of a sphere of radius 1.
To justify the second, note that the total length of the
It may look as though we still have to decide in
intervals making up Xn is ( 23 )n (since one removes a
advance whether we will talk about length, area, or vol-
third of Xn−1 to produce Xn ). Since the Cantor set is
ume before we can even begin to think about how the
contained in every Xn , its measure must be smaller
measurement scales when we expand the shape. But
than ( 23 )n , whatever n is, which means that it must be
this is not the case. For instance, if we expand a square
zero. Thus, the Cantor set is very large in one respect
by a factor of 2, then we obtain a new square that can
be divided up into four congruent copies of the original and very small in another.
square. So, without having decided in advance that we A further property of the Cantor set is that it is self-
are talking about area, we can say that the size of the similar. The set X1 consists of two intervals, and if you
new square is four times that of the old square. look at just one of these intervals as the middle thirds
This observation has a remarkable consequence: are repeatedly removed, then what you see is just like
there are sets to which it is natural to assign a dimen- the construction of the whole Cantor set, but scaled
sion that is not an integer! Perhaps the simplest exam- down by a factor of 3. That is, the Cantor set consists
ple is a famous set first defined by cantor [VI.54] and of two copies of itself, each scaled down by a factor
now known as the Cantor set. This set is produced as of 3. From this we deduce the following statement: if
follows. You start with the closed interval [0, 1], and you expand the Cantor set by a factor of 3, then you can
call it X0 . Then you form a set X1 by removing the mid- divide the expanded set up into two congruent copies
dle third of X0 : that is, you remove all points between of the original, so it is “twice as big.”
1 2 1 2
3 and 3 , but leave 3 and 3 themselves. So X1 is the
What consequence should this have for the dimen-
1 2
union of the closed intervals [0, 3 ] and [ 3 , 1]. Next, you sion of the Cantor set? Well, if the dimension is d, then

You might also like