Chapter 2.

Length measure on R
The idea here is to develop the idea of the length of a subset of R in such a way that we can
manage to dene the Lebesgue integral in more or less the way outlined in Chapter 0.
Ideally we should be able to talk about the length (total length) of any subset of the line, but
this turns out to be unworkable. Well, at least lots of things that seem obvious cant work unless
we restrict ourselves to subsets of R that are not too terrible. Well explain later in precise terms
which kinds of sets are going to be allowed.
The approach we take is to start with lengths of intervals, in fact not even of all possible
kinds of intervals, and then build up to more complicated sets. If we wanted to cater for double
integrals and triple integrals (and more generally integrals of functions of n real variables) we
should start with areas of rectangles in R
2
, or volumes of boxes in R
3
, or n-dimensional volume
of n-dimensional boxes in R
n
, and build up from there. We will not actually do these more
general cases, but the approach we use on R can be adapted fairly easily to work on R
n
.
2.1 Algebras of subsets of R
As we said, we cant manage with lengths of all subsets of R, that is we will not be able to assign
a satisfactory length to each set in the power set P(R), the collection of all subsets. We do insist
that we have our length making sense for a (Boolean) algebra of subsets.
Denition 2.1.1. A collection A of subsets of R is called an algebra of subsets of R if
(a) A
(b) E A E
c
A
(c) E
1
, E
2
A E
1
E
2
A
(In words, A contains the empty set, is closed under taking complements and under unions (of
two members).)
Lemma 2.1.2. If A is an algebra of subsets of R, then
(i) R A
(ii) E
1
, E
2
A E
1
E
2
A
(iii) E
1
, E
2
, . . . , E
k
A

k
j=1
E
j
A and

k
j=1
E
j
A
Proof. (i) R =
c
A by Denition 2.1.1 (a) and (b).
(ii) E
1
, E
2
A E
1
E
2
= (E
c
1
E
c
2
)
c
(by De Morgans laws) E
1
E
2
A (because
from Denition 2.1.1 (b) we have E
c
1
, E
c
2
A, thus E
c
1
E
c
2
A by Denition 2.1.1 (c),
and so (E
c
1
E
c
2
)
c
A by Denition 2.1.1 (b) again).
2 201112 Mathematics MA2224
(iii) This follows by induction on k since

k+1
j=1
E
j
=
_

k
j=1
E
j
_
E
k
, and similarly for inter-
sections.
Denition 2.1.3. We dene the interval algebra I to be the collection of all nite unions of
intervals of one of the types
(a, b], (, b], (a, ), (, )
We include the empty set in I (as the union of the empty collection of such intervals!).
(So the right hand end point is included in these intervals when it is nite and the left-hand
end point is not included.)
Lemma 2.1.4. I is an algebra (of subsets of R).
Proof. This is almost obvious, but a formal check is perhaps a good idea. If we take the union
of two members of I, say of two nite unions E
1
=

n
j=1
I
j
and E
2
=

m
k=1
I

k
(where the I
j
and I

k
are intervals of the allowed types), then
E
1
E
2
=
n
_
j=1
I
j

m
_
k=1
I

k
is again a nite union of intervals and so E
1
E
2
I.
Considering the intersection E
1
E
2
(instead of the union) we have
E
1
E
2
=
n
_
j=1
(I
j
E
2
) =
n
_
j=1
m
_
k=1
(I
j
I

k
)
and to see that E
1
E
2
I notice that each I
j
I

j
is either empty or another interval of the
same type. By induction, we have then that nite intersections of sets in I will be in I.
The complement of one of the intervals we are allowing is either another of these intervals or
the union of two such. For instance
(a, b]
c
= (, a] (b, ), and (, b]
c
= (b, ).
By De Morgans laws then, if E I with E =

n
j=1
I
j
, we have
E
c
=
n

j=1
I
c
j
and we have just noted that each I
c
j
I. So E
c
I (using what we just showed about nite
intersections).
Length measure 3
Remark 2.1.5. One thing to note is that if two intervals of the types above overlap or share an
endpoint, then their union is another interval of the same kind.
So we can write every set in I uniquely as one of the following
, R,
(a
1
, b
1
] (a
2
, b
2
] (a
n
, b
n
],
(, b
1
] (a
2
, b
2
] (a
n
, b
n
],
(a
1
, b
1
] (a
2
, b
2
] (a
n
, ), or
(, b
1
] (a
2
, b
2
] (a
n
, ),
where
a
1
< b
1
< a
2
< b
2
< a
n
< b
n
.
Remark 2.1.6. We want to dene a length function on subsets of R, certainly including the sets
in I for a start. But some of the intervals above have innite length and so we are led to allow
as a possible length.
That means we introduce an extended system of nonnegative numbers by adding an extra
symbol to the usual [0, ). We regard as bigger than every ordinary x [0, ), that is we
extend the denition of x
1
< x
2
for x
1
, x
2
0 to have x < for all x [0, ). Then we have
[0, ] including .
We will also want to allow most arithmetic operations to be done where is allowed. So we
will extend the denition of addition and multiplication of positive numbers so that
x += + x = += (0 x < ),
x= x = = (0 < x < ),
0= 0 = 0
and sometimes we will have subtraction (mostly where we end up with a positive result) but we
will not dene .
So we will have x = for 0 x < .
We could go further and also introduce (smaller than all x R) and then get an extended
real line [, ], but we will not really need that.
Note that the rule 0 = 0 = 0 is not always good to allow, for instance in the case of
limits, but it will be appropriate in the ways that we will encounter it for integration.
Denition 2.1.7. We dene a length function (or measure) m: I [0, ] by taking
m((a, b]) = b a, m((, b]) = m((a, )) = m((, )) =
and dening m(E) for general E I as the sum of the lengths of the intervals in the unique
representation mentioned above in Remark 2.1.5.
4 201112 Mathematics MA2224
Terminology 2.1.8. We say that a function m: A [0, ] is nitely additive if whenever
E
1
, E
2
, . . . , E
n
A are disjoint (that is E
j
E
k
= for j = k), then
m(E
1
E
2
E
n
) = m(E
1
) + m(E
2
) + + m(E
n
).
By induction, to check nite additivity we only need to check it for n = 2.
We say that m is nitely subadditive if whenever E
1
, E
2
, . . . , E
n
A, then
m(E
1
E
2
E
n
) m(E
1
) + m(E
2
) + + m(E
n
).
Again the case n = 2 implies the general case.
Lemma 2.1.9. The length function m we have dened on I has the properties
(i) E
1
, E
2
I, E
1
E
2
= m(E
1
E
2
) = m(E
1
) m(E
2
) (which implies nite
additivity);
(ii) E
1
, E
2
I, E
1
E
2
m(E
1
) m(E
2
) (monotonicity);
(iii) m is nitely subadditive.
Proof. Most of this is quite easy, almost obvious perhaps, but maybe not so easy to organise into
a proper proof.
(i) It is sufcient to establish this in the case when E
1
E
2
is an interval. The reason is this.
We know E
1
E
2
I and so we can write it in the form E
1
E
2
=

n
j=1
I
j
where
the I
j
are intervals of the allowed types and share no endpoints that is we are taking
E
1
E
2
represented in the form considered in Remark 2.1.5. Since E
1
E
1
E
2
, it must
be that when we write E
1
in the form we mentioned in Remark 2.1.5, then the intervals
used for E
1
must all be contained in one of the intervals I
j
. So it is clear then that m(E
1
) =

n
j=1
m(E
1
I
j
) and similarly m(E
2
) =

n
j=1
m(E
2
I
j
). Now I
j
= (E
1
I
j
)(E
2
I
j
)
(since I
j
E
1
E
2
) and that justies the simplication.
Assuming then that E
1
E
2
= I is a single interval, the intervals that make up E
1
and E
2
must t together to ll up I, with the intervals for E
1
alternating with those for E
2
(because
E
1
E
2
= ). It is then pretty obvious that the sum of the lengths of the intervals for E
1
together with those of E
2
must be m(I). So m(I) = m(E
1
) + m(E
2
), or m(E
1
E
2
) =
m(E
1
) + m(E
2
).
As noted before, this implies nite additivity.
(ii) If E
1
E
2
, then E
2
= E
1
(E
2
\ E
1
), a disjoint union, and so (by (i))
m(E
2
) = m(E
1
) + m(E
2
\ E
1
) m(E
1
).
(iii) For E
1
, E
2
I, we get E
1
E
2
= E
1
(E
2
\ E
1
), a disjoint union and so
m(E
1
E
2
) = m(E
1
) + m(E
2
\ E
1
) m(E
1
) + m(E
2
)
using (i) and (ii). As noted before, this implies nite subadditivity.
Length measure 5
Remark 2.1.10. Looking back at the example with y = sin(1/x) in 0.3 we already indicated
that we need to be able to deal with more than just lengths of nite unions of intervals. We need
to be able at least to deal with sets that are made up of countably many intervals, and the obvious
way to dene their total length is to add up all the lengths.
This brings us to a more general kind of additivity, called countable additivity.
Before we get to that, notice that there is a nice thing about series

n=1
t
n
of terms t
n

[0, ]. If there is any term where t
n
0
= , then

N
n=1
t
n
= whenever N n
0
and so it
seems perfectly logical to say that

n=1
t
n
= . On the other hand, if t
n
< always then
the sequence of partial sums s
N
=

N
n=1
t
n
is a monotone increasing sequence. If the sequence
of partial sums is bounded above (by a nite number), then lim
N
s
N
= sup
N1
s
N
and this
number is

n=1
t
n
(in the ordinary sense). If the sequence of partial sums is not bounded above
by a nite quantity, we dene

n=1
t
n
= .
So every series

n=1
t
n
of terms t
n
in [0, ] has a sum.
Another good thing is that the sum remains the same if we change the order of the terms.
That follows from the fact that the sum is the supremum (in [0, ]) of the partial sums.
Terminology 2.1.11. If A is an algebra and m: A [0, ], then we say that m is countably
additive if whenever E
1
, E
2
, E
3
, . . . is an innite sequence of disjoint sets in A such that the
union

n=1
E
n
A, then
m
_

_
n=1
E
n
_
=

n=1
m(E
n
).
What we want is to get to a situation where we consider only algebras Athat are closed under
taking countable unions, but the algebra I does not have that property. We will need one more
result about the length function we have dened on I.
Proposition 2.1.12. The length function on I is countably additive.
Proof. As in the proof of nite additivity, it is enough to consider the case where an interval
I I is a countable union I =

n=1
E
n
of disjoint E
n
I. Furthermore, we can write
each E
n
=

jn
j=1
I
n,j
as nite disjoint union of intervals of one of the allowed types (where
m(E
n
) =

jn
j=1
m(I
n,j
)) and so we can reduce to the case where I =

n=1
I
n
is a disjoint union
of intervals I
n
. [To spell it out more, the idea is that E = E
1
E
2
= I
1,1
I
1,2
I
1,n
1

I
2,1
I
2,n
2
I
3,1
and we number these intervals as I
1
, I
2
, . . ..]
Since

N
n=1
I
n
I for N nite, we can apply nite additivity and monotonicity to get
N

n=1
m(I
n
) = m
_
N
_
n=1
I
n
_
m(I)
for each N. Letting N gives half of what we want to show

n=1
m(I
n
) m(I).
6 201112 Mathematics MA2224
It remains to prove holds also.
If any of the intervals I
n
is of innite length, then it is clear that I I
n
must be innite also
and so m(I) =

n=1
I
n
in this case.
So we assume each I
n
is a nite interval (a
n
, b
n
]. Fix > 0 arbitrarily small and choose a
nite closed interval [a
0
, b
0
] I. If I = (a, b] is a nite interval, take a
0
= a + and b
0
= b
(where we only consider smaller that ba). If I is innite take [a
0
, b
0
] a nite interval contained
it of any large length.
Now (a
0
, b
0
] [a
0
, b
0
]

n=1
(a
n
, b
n
]

n=1
(a
n
, b
n
+ /2
n
) and we have an open cover
{(a
n
, b
n
+ /2
n
) : n N} of the compact interval [a
0
, b
0
]. By the Heine Borel theorem, there
must be a nite subcover
[a
0
, b
0
] (a
n
1
, b
n
1
+ /2
n
1
) (a
n
2
, b
n
2
+ /2
n
2
) (a
n
k
, b
n
k
+ /2
n
k
)
and then we have
(a
0
, b
0
]
k
_
j=1
(a
n
j
, b
n
j
+ /2
n
j
].
By subadditivity of m
b
0
a
0

k

j=1
(b
n
j
a
n
j
+ /2
n
j
)

n=1
(b
n
a
n
+ /2
n
)
=

n=1
(b
n
a
n
) +

n=1
/2
n
=

n=1
(b
n
a
n
) +
=

n=1
m(I
n
) +
In the case where m(I) = b a < then we get
b
0
a
0
= b (a + )

n=1
m(I
n
) + m(I) = b a

n=1
m(I
n
) + 2.
But this is true no matter how small > 0 is and so we have m(I)

n=1
m(I
n
) if m(I) < .
If m(I) = we can choose b
0
a
0
as big as we like (with = 1 say) to get

n=1
m(I
n
) =
= m(I).
Length measure 7
Corollary 2.1.13. The length function on I is countably subadditive in the sense that when-
ever E
1
, E
2
, E
3
, . . . is an innite sequence of sets in I such that the union

n=1
E
n
I, then
m
_

_
n=1
E
j
_

n=1
m(E
n
).
Proof. We can deduce this from countable additivity (Proposition 2.1.12) and monotonicty as
follows. Let E =

n=1
E
n
where E
n
I for each n and we are supposing that E I also.
We generate a sequence of disjoint sets F
1
, F
2
, . . . I such that F
n
E
n
,

n
j=1
F
j
=

n
j=1
E
j
for each n and

n=1
F
n
= E. To do that let F
1
= E
1
and inductively dene
F
n+1
= E
n
\
_
n
_
j=1
E
j
_
= E
n

_
n
_
j=1
E
j
_
c
I
for n 1.
The properties claimed for (F
n
)

n=1
are rather easy to check (by induction) and the by count-
able addiivity of m we have
m(E) = m
_

_
n=1
F
n
_
=

n=1
m(F
n
)

n=1
m(E
n
)
(because F
n
E
n
m(F
n
) m(F
n
) using monotonicty of m)
Notation 2.1.14. If E R and x
0
R we dene
x
0
+ E = {x
0
+ x : x E}
(and refer to x
0
+ E as the translate of E by x
0
).
Proposition 2.1.15 (Translation invariance). For E I and x
0
R we have x
0
+ E I and
m(x
0
+ E) = m(E).
Proof. Exercise.
2.2 Outer measure
Denition 2.2.1 (Outer measure). We now dene the outer measure of an arbitrary subset S R
to be
m

(S) = inf
_

n=1
m(E
n
) : E
1
, E
2
, . . . I with S

_
n=1
E
n
_
(with the understanding that if

n=1
m(E
n
) = always, then m

(S) = ).
8 201112 Mathematics MA2224
Remark 2.2.2. Outer measure is dened on the algebra P(R) of all subsets of R, but it fails to
be even nitely additive. We will identify a smaller algebra, the algebra of Lebesgue measurable
subsets, on which m

is countably additive. In the meantime, we will give some properties that

m

does have.
It should be clear soon that m

is dened to be as large as it can be subject to two requirements

(a) countably subadditive and (b) agrees with m on I.
In the arguments that follow, you may like to notice any time we delve into what I is.
We will be making use of rather general facts the fact that I is an algebra, countable (sub)
additivity of m, monotonicity of m, m() = 0 and translation invariance of m. But we wont be
delving into the details of m or I.
Since I is an algebra, for any S R we can take E
1
= R, E
n
= for n = 2, 3 . . . and
so there is always at least one possible choice for the sequence (E
n
)

n=1
used in the denition of
m

(S).
Proposition 2.2.3. (i) m

() = 0
(ii) S
1
S
2
R m

(S
1
) m

(S
2
) (monotonicity)
(iii) If S
1
, S
2
, . . . R, then
m

_
n=1
S
n
_

n=1
m

(S
n
)
(this is called countable subadditivity)
(iv) If E I, then m

(E) = m(E).
(v) For S R and x
0
R, m

(x
0
+ S) = m

(S) (translation invariance of outer measure).

Proof. (i) We can take E
n
= I for each n 1 and then

n=1
E
n
while

n=1
m(E
n
) =
0. So m

() 0. Hence m

() = 0.
(ii) If m

(S
2
) = then certainly the inequality is true. For the case m

(E
2
) < , if
E
1
, E
2
, . . . I with S
2

n=1
E
n
, then certainly also S
1

n=1
E
n
and m

(S
1
)

n=1
m(E
n
). But this sum can be arbitrarily close to m

(S
2
) and so m

(S
1
) m

(S
2
).
(iii) If

n=1
m

(S
n
) = , then the inequality is true and so we consider the case

n=1
m

(S
n
) <
. Fix > 0 and for each n, choose E
n,1
, E
n,2
, . . . I with
S
n

_
j=1
E
n,j
and

j=1
m(E
n,j
) m

(S
n
) +

2
n
.
Then, taking S =

n=1
S
n
we have
S

_
n=1

_
j=1
E
n,j
Length measure 9
(a countable number of sets in I that can be arranged as a single sequence) and so
m

(S)

n=1

j=1
m(E
n,j
)

n=1
_
m

(S
n
) +

2
n
_
=

n=1
m

(S
n
) + .
Since we can take > 0 arbitrarily small we have the result.
(iv) For E I we can take E
1
= E and E
n
= for n = 2, 3, . . .. That gives E
1
, E
2
, . . . I
with E

n=1
E
n
and shows that m

(E)

n=1
m(E
n
) = m(E) + 0 = m(E).
On the other hand if we have any E
1
, E
2
, . . . I with E

n=1
E
n
, then we can say that
E = E
_

_
n=1
E
n
_
=

_
n=1
E E
n
I
and apply countable subadditivity of m (Corollary 2.1.13) to deduce
m(E)

n=1
m(E E
n
)

n=1
m(E
n
).
In this way we conclude that m(E) m

(E). Combining with the reverse inequality

obtained rst we get equality, as required.
(v) We show m

(x
0
+ S) m

(S) rst. Of course if m

(S) = , this is true. If m

(S) <
, for any > 0 we must be able to nd E
1
, E
2
, . . . I so that S

n=1
E
n
and

n=1
m(E
n
) m

(S) +. Then x
0
+S

n=1
(x
0
+E
n
), where we know x
0
+E
n
I
with m(x
0
+ E
n
) = m(E
n
) (by translation invariance of m), and so we have
m

(x
0
+ S)

n=1
m(x
0
+ E
n
) =

n=1
m(E
n
) m

(S) + .
As > 0 can be arbitrary, we conclude m

(x
0
+ S) m

(S).
The reverse inequality follows by applying the same fact to x
0
and

S = x
0
+ S to get
m

((x
0
) +

S) m

S), which means m

(S) m

(x
0
+ S). So we have equality.
Corollary 2.2.4. m

is nitely subadditive. That is, if S

1
, S
2
, . . . , S
n
R, then
m

_
n
_
j=1
S
j
_

j=1
m

(S
j
).
Proof. Exercise.
10 201112 Mathematics MA2224
2.3 Lebesgue measurable sets
Although we have not justied the claim that m

fails to be nitely additive, we use it as a way

to distinguish the well-behaved sets.
Denition 2.3.1. We say that a subset F R is Lebesgue measurable (or measurable with
respect to the outer measure m

) if for every subset S R,

m

(S) = m

(S F) + m

(S F
c
).
We denote the collection of Lebesgue measurable sets by L.
The criterion used to dene L is called the Carath eodory criterion. Our aim now is to show
that this class L is a well-behaved algebra, in fact a -algebra and that the restriction of m

to
L is also well-behaved. We need some denitions so we can express things concisely.
Notice that the Carath eodory criterion says that F and F
c
divides every subset S R addi-
tively for m

. Here is a small simplication of the criterion.

Lemma 2.3.2. A set F R is in L if it satises
m

(S) m

(S F) + m

(S F
c
).
for each S R.
Proof. By subadditivity of m

, we always know
m

(S) m

(S F) + m

(S F
c
).
and combining with the assumption in the lemma we get equality,
Denition 2.3.3. An algebra A(of subsets of R) is called a -algebra (sigma-algebra) if when-
ever E
1
, E
2
, . . . A, then

n=1
E
n
A. [In words: A is closed under the operation of taking
countable unions.]
Denition 2.3.4. If is a -algebra (of subsets of R) and : [0, ] is a function, then we
call a measure on if it satises
(a) () = 0
(b) is countably additive, that is whenever E
1
, E
2
, . . . are disjoint, then (

n=1
E
n
) =

n=1
(E
n
).
Theorem 2.3.5 (Carath eodory). L is a -algebra, I L and the restriction of outer measure
m

to L denes a measure on L.
Proof. Step 1: L is an algebra.
(i) L
For any S R, m

(S ) +m

(S
c
) = m

() +m

(S) = 0 +m

(S) = m

(S).
Length measure 11
(ii) F L F
c
L
For any S R, m

(S F
c
) + m

(S (F
c
)
c
) = m

(S F
c
) + m

(S F) =
m

(S F) + m

(S F
c
) = m

(S) by measurability of F.
(iii) F
1
, F
2
L F
1
F
2
L.
We rst show F
1
F
2
L (for F
1
, F
2
L). Take S R. It may help to look at
this Venn diagram to follow the argument.
F2 F1
S1
S2
S3
S4
S
Notice that
m

(S) = m

(S F
1
) + m

(S F
c
1
)
= m

(S F
1
F
2
) + m

(S F
1
F
c
2
) + m

(S F
c
1
)
(using rst F
1
L, then F
2
L). But also
m

(S (F
1
F
2
)
c
) = m

(S (F
1
F
2
)
c
F
1
) + m

(S (F
1
F
2
)
c
F
c
1
)
(F
1
and F
c
1
split S (F
1
F
2
)
c
additively)
= m

(S (F
c
1
F
c
2
) F
1
) + m

(S (F
c
1
F
c
2
) F
c
1
)
= m

(S F
c
2
F
1
) + m

(S F
c
1
)
= m

(S F
1
F
c
2
) + m

(S F
c
1
)
and so we get
m

(S) = m

(S F
1
F
2
) + m

(S (F
1
F
2
)
c
).
This shows F
1
F
2
L
As F
1
F
2
= (F
c
1
F
c
2
)
c
, we get F
1
F
2
L because F
c
1
, F
c
2
L F
c
1
F
c
2

L (F
c
1
F
c
2
)
c
L, or F
1
F
2
L.
Step 2: L is a -algebra and m

is countably additive on L.
12 201112 Mathematics MA2224
If F
1
, F
2
, . . . L we aim to show that F =

n=1
F
n
L. It is enough to do that
when F
1
, F
2
, . . . are disjoint, since we can replace F
1
, F
2
, . . . by

F
1
= F
1
,

F
2
= F
2
F
c
1
,
. . . ,

F
n
= F
n

n1
j=1
F
j
_
c
, thus getting a disjoint sequence in L with the same union
F =

n=1

F
n
.
So we assume that F
1
, F
2
, . . . are disjoint, F =

n=1
F
n
.
Take any S R. We claim that m

(S) m

(S F) +m

(S F
c
) always (which shows
F L via Lemma 2.3.2).
Take G
n
=

n
j=1
F
n
(which we know is in L as L is an algebra) and notice that
m

(S) m

(S G
n
) + m

(S G
c
n
).
Since F
n
L we have
m

(S G
n
) = m

(S G
n
F
n
) + m

(S G
n
F
c
n
) = m

(S F
n
) + m

(S G
n1
)
(for n > 1). Since F
1
= G
1
we then can see by induction on n that
m

(S G
n
) =
n

j=1
m

(S F
j
),
and so
m

(S)
n

j=1
m

(S F
j
) + m

(S G
c
n
).
However G
n
F implies G
c
n
F
c
and so (by monotonicity of m

) we get
m

(S)
n

j=1
m

(S F
j
) + m

(S F
c
).
Let n to get
m

(S)

n=1
m

(S F
n
) + m

(S F
c
). (1)
From subadditivity of m

, this inequality implies

m

(S) m

_
n=1
S F
n
_
+ m

(S F
c
) = m

(S F) + m

(S F
c
).
So F L is veried.
Using (1) for S = F we get
m

(F) = m

_
n=1
F
n
_

n=1
m

(F
n
)
and subadditivity of m

gives the opposite inequality. That shows countable additivity of

m

on L.
Length measure 13
Step 3: I L.
Fix E I. Take any S R. We claim that m

(S) m

(S E) + m

(S E
c
) always
(which shows E L via Lemma 2.3.2). If m

(S) = there is nothing to do.

For any > 0 (xed) and then there must be E
1
, E
2
, . . . I with S

n=1
E
n
and

n=1
m(E
n
) < m

(S) + .
We then have S E

n=1
E
n
E, and E
n
E I. So
m

(S E)

n=1
m(E
n
E).
Similarly, as E
c
I,
m

(S E
c
)

n=1
m(E
n
E
c
).
But m(E
n
) = m(E
n
E) + m(E
n
E
c
) (nite additivity of m) and so
m

(S E) + m

(S E
c
)

n=1
m(E
n
E) +

n=1
m(E
n
E
c
)
=

n=1
(m(E
n
E) + m(E
n
E
c
))
=

n=1
m(E
n
)
m

(S) + .
As this holds for each > 0 we have m

(SE)+m

(SE
c
) m

(S) and that completes

the argument.
Remark 2.3.6. It may be a good idea now to recap a little on where we have reached.
We began with lengths of certain nite intervals (a, b] (with a < b), where we assigned the
length m((a, b]) = b a, together with certain innite intervals to which we assigned innite
length. Just to make things more convenient we worked with nite unions of intervals of these
types so that we would have an algebra of sets to work with.
For lengths of sets in that algebra, the interval algebra I, we showed that certain nice prop-
erties worked.
We then dened outer measure for arbitrary subsets of R, basically by making m

(E) be the
biggest thing it could be if we wanted it to agree with m(E) when E is in the interval algebra I
and to be countably subadditive.
Finally we identied a -algebra L containing I and on which m

is well-behaved in the
sense of being countably additive.
We have pretty much what we need to start looking at integration, but rst we take stock a bit
of this -algebra L.
14 201112 Mathematics MA2224
Proposition 2.3.7. L is translation invariant. That is if F L and x
0
R, then x
0
+F L.
Proof. Exercise. (It is not that hard because we already know m

(x
0
+ E) = m

(E) for every

E R. See Proposition 2.2.3 (v).)
Proposition 2.3.8. If C is a set of subsets of R (so that means C P(R) is a subset of the power
set of R, the collection of all subsets of R), then C is a -algebra if and only if it satises
(i) C
(ii) E C E
c
C
(iii) E
1
, E
2
, . . . C

n=1
E
n
C.
Proof. We dened a -algebra to be an algebra with the additional property of being closed
under taking countable unions. So every -algebra C has these three properties.
For the converse, suppose C has these three properties. We are just missing the property of
an algebra that
E
1
, E
2
C E
1
E
2
C.
However, if E
1
, E
2
C we can make an innite sequence by dening E
n
= for n 3. Now
E
1
, E
2
, . . . C and so we have

n=1
E
n
C. But

n=1
E
n
= E
1
E
2
because the other sets in
the sequence are all empty. So we do have E
1
E
2
C.
Proposition 2.3.9. For any set S P(R) of subsets of R, there is a smallest -algebra (of
subsets of R) that contains S.
We call it the -algebra generated by S.
Proof. This is sort of easy, but in an abstract way. There is certainly one possible -algebra
containing S, that is P(R).
What we do is look at all possible -algebra, the set
S = { : P(R) a -algebra and S }.
Then we take their intersection

S
=

and argue that

S
is still a -algebra. In fact that
S
S and is the contained in every S
by the way we dened it.
We check
(i)
S
because for each S.
(ii) E
S
E
c

S
because E
S
E for each S, and so E
c
for each S. Thus
E
c

S
.
Length measure 15
(iii) E
1
, E
2
, . . .
S

n=1
E
n

S
If E
n

S
for n = 1, 2, . . ., then we have E
1
, E
2
, . . . for each S. So

n=1
E
n

for each S, and hence

n=1
E
n

S
.
Finally S
S
holds because S for each S.
Lemma 2.3.10. If is a -algebra (of subsets of R) and E
1
, E
2
, . . . , then

n=1
E
n
.
(So -algebras are closed under countable intersections as well as under countable unions.)
Proof. This is because De Morgans laws allow us to turn intersections unto unions via taking
complements:

n=1
E
n
=
_

_
n=1
E
c
n
_
c
We have E
c
n
for all n, hence

n=1
E
c
n
, hence (

n=1
E
c
n
)
c
.
Theorem 2.3.11. The following -algebras of subsets of R are all the same -algebra (and
usually called the Borel -algebra on R)
1.
1
= the -algebra generated by I (the interval algebra)
2.
2
= the -algebra generated by the collection of all nite open intervals (a, b) (with
a < b)
3.
3
= the -algebra generated by the collection of all nite closed intervals [a, b] (with
a b)
4.
4
= the -algebra generated by the collection of all open subsets of R
5.
5
= the -algebra generated by the collection of all closed subsets of R
Proof. It helps to notice rst that one point sets are included in each of these -algebras.
Starting with the interval algebra I and any x
0
R we have {x
0
} =

n=1
_
x
0

1
n
, x
0

and
so {x
0
} is in any -algebra that contains I. Thus {x
0
}
1
.
For open intervals, we can say instead that {x
0
} =

n=1
_
x
0

1
n
, x
0
+
1
n
_
. So we have
{x
0
}
2
.
That also covers
4
(open intervals are open sets).
Since single point sets are closed, we have no bother with {x
0
}
5
(for each x
0
R).
Since single point sets are closed intervals also, because [a, a] = {a}, the fact that {x
0
}
3
is
also easy.
So we have that any -algebra that contains I (in particular
1
) must contain nite open
intervals (a, b) = (a, b] \ {b} = (a, b]){b})
c
. It follows that
1

2
.

2

3
since [a, b] = (a, b) {a} {b} and
3

2
also because (a, b) = [a, b] ({a}
{b})
c
. So
2
=
3
.
To show
2

4
we show that every open subset U R is a countable union of open
intervals. For each x U, there is > 0 with (x, x+) U. (That is what it means for U to
16 201112 Mathematics MA2224
be an open set.) Now there are rational numbers q
1
, q
2
with x < q
1
< x < q
2
< x+ and that
means x (q
1
, q
2
) U. It follows that U is the union of all open intervals (q
1
, q
2
) with rational
endpoints q
1
< q
2
that are contained in U. But there are fewer such intervals than there are pairs
(q
1
, q
2
) Q Q = Q
2
. Since Q
2
is a countable set, the collection of these rational intervals
contained in U is countable. So U is a countable union of open intervals, hence U
2
. This
shows
2

4
.
To show
4

1
we need only show
4
I. We have (, ), (a, )
4
because
they are open. Also (, b] = (, b) {b} and (a, b] = (a, b) {b} are
4
. So I
4
because sets in I are nite unions of these kinds of intervals.
From
1

2
=
3

4

1
, we have
1
=
2
=
3
=
4
. Finally
4
=
5
because
closed sets are exactly the complements of open sets.
Corollary 2.3.12. The Borel -algebra (of subsets of R) is the -algebra generated by the inter-
vals (, b] with b R.
Proof. We have
(a, b] = (, b] ((, a])
c
.
Also (, b]
c
= (b, ) and
c
= R = (, ). So any -algebra that contains the intervals
(, b] must also contain I.
The reverse inclusion, that the -algebra generated by I must contain the -algebra gener-
ated by the intervals (, b] is easy because each (, b] I.
Remark 2.3.13. We have not proved that there any sets not included in L. This is true, though
the proof relies on the axiom of choice, something that cannot be proved based on the usual
axioms of mathematical logic.
The idea for the proof is to dene an equivalence relation on R by x y x y Q.
Then choose one x [0, 1] belonging to each equivalence class and let E be the set of those x.
Now
[0, 1]
_
qQ[1,1]
(q + E) [1, 2]
We know m

(q + E) = m

(E) for each q. If E was in L, then so would q + E for each q, and

countable additivity of m

on L would give a contradiction from

1 = m

([0, 1]) m

_
_
_
qQ[1,1]
(q + E)
_
_
=

qQ[1,1]
m

(q + E) m

([1, 2]) = 3
If m

(E) > 0, the sum would be , while if m

(E) = 0, the sum would be 0. Neither possibility

is between 1 and 3. So E / L.
Remark 2.3.14. We have also not proved that there are fewer Borel sets than there are Lebesgue
measurable sets, but that is also true (and does not need the axiom of choice it needs just a bit
more about how to distinguish different sizes of uncountable sets than we have looked into).
It is not too hard to show that for each E L, there is a set B is in the Borel -algebra with
m

(EB) = 0 (where EB = (E \ B) (B \ E) is the symmetric difference). I think we can

Length measure 17
omit that proof. On tutorial sheet 5 we showed the opposite: that if B is in the Borel -algebra
and E R has m

(EB) = 0, then E L.
So, in a way, the difference between the Borel -algebra and L comes from the fact that L
includes all sets E with m

(E) = 0.
R. Timoney February 16, 2012