7.

Measure Theory
You have already studied Lebesgue measure in Rn.
Here we consider a more abstract setup. This will al-
low us to develop a unified framework to deal with many
different situations, such as measures supported on sub-
manifolds or fractals, or infinite dimensional spaces, such
as Wiener measure used in the definition of Brownian
motion.

Definition 7.1. A measure space is a set X together

with a σ-algebra M, a collection of subsets of X, and
a measure µ : M → [0, ∞], such that
• M is required to contain the empty set and to
be closed under countable unions and complementation
(hence also closed under countable intersections).
• µ is required to be countably additive: if E1, E2, . . .
is a countable family of disjoint measurable sets, then
( ) ∑
µ ∪n En = µ(En).
n

1
 Example.
• Lebesgue measure on Rn, with the σ-algebra of Lebesgue
measurable sets.
• Counting measure on any set, where every set is
measurable.
• Surface measures.
• Hausdorff measures, which are ‘fractional dimen-
sional’ measures of subsets of Rn. We will introduce
these in due course.

How to construct measure spaces

Measures can be constructed from ‘exterior measures’.
Let X be any set. An exterior measure µ∗ is a
function from P(X) to [0, ∞] such that
(i) µ∗(∅) = 0;
(ii) If E1 ⊂ E2, then µ∗(E1) ≤ µ∗(E2);
(iii) if E1, E2, . . . is a countable family of sets, then
( ) ∑
µ∗ ∪n En ≤ µ∗(En).
n
 There is a general construction, due to Carathéodory,
of obtaining a measure from an exterior measure. We
say that E ⊂ X is measurable if, for every subset A ⊂
X, we have

µ∗(A) = µ∗(A ∩ E) + µ∗(A ∩ E c).

That is, E separates every set efficiently: we have equal-

ity in the equation above rather than ‘≤’ which holds
automatically. Heuristically, we can think of a non-
measurable set as being one that looks big ‘from the
outside’ but small ‘from the inside’, and the exterior
measure as measuring the size of the set as viewed from
the outside. But µ∗(A) − µ∗(A ∩ E c) in some sense
measures the size of A ∩ E from the inside, so if we get
equality, then this indicates that E should be measur-
able.

Theorem 7.2. Let µ∗ be an exterior measure on X.

Then the subset M of P(X) consisting of measurable
sets forms a σ-algebra, and µ∗ restricted to M is a
measure.

Proof: 1. It is clear that E ∈ M =⇒ E c ∈ M.

 2. To show that the union of two sets E1, E2 ∈ M is
in M, we use
µ∗(A) = µ∗(A ∩ E1) + µ∗(A ∩ E1c)
and then write each of these two terms as a sum as
follows

µ∗(A ∩ E1) = µ∗(A ∩ E1 ∩ E2) + µ∗(A ∩ E1 ∩ E2c)

µ∗(A ∩ E1c) = µ∗(A ∩ E1c ∩ E2) + µ∗(A ∩ E1c ∩ E2c).
Now using subadditivity of µ∗, we have
µ∗(A ∩ (E1 ∪ E2)) ≤ µ∗((A ∩ E1 ∩ E2))
+ µ∗((A ∩ E1 ∩ E2c)) + µ∗((A ∩ E1c ∩ E2))
(Draw a diagram!) Combining the two we get
µ∗(A ∩ (E1 ∪ E2)) + µ∗(A ∩ E1c ∩ E2c) ≤ µ∗(A)
or equivalently
µ∗(A) ≥ µ∗(A ∩ (E1 ∪ E2)) + µ∗(A ∩ (E1 ∪ E2)c),
since A ∩ E1c ∩ E2c = A ∩ (E1 ∪ E2)c. Equality in this
equation follows from subadditivity, which gives the ‘≤’
direction.
2.’ Additivity for two sets follows directly from the
definition of measurability: put A = E1 ∪ E2 and E =
E1, and obtain
µ∗(E1 ∪ E2) = µ∗(E1) + µ∗(E2).
One then uses induction to obtain finite additivity.
3. To show closure under countable unions, let Ei ∈
M. Without loss of generality, the Ei are disjoint. Let
Gn = E1 ∪ · · · ∪ En, and let G = ∪En. Then each
Gn ∈ M, and we have
µ∗(A) = µ∗(A ∩ Gn) + µ∗(A ∩ Gcn)
≥ µ∗(A ∩ Gn) + µ∗(A ∩ Gc)
and since A ∩ Gn = ∪nj=1A ∩ Ej so µ∗(A ∩ Gn) =
∑
µ∗(A ∩ Ej ), using finite additivity as proved earlier,
we now have
∑ n
µ∗(A ∩ Ej ) ≤ µ∗(A) − µ∗(A ∩ Gc).
j=1
Hence, taking the limit as n → ∞,
∑∞
µ∗(A ∩ Ej ) ≤ µ∗(A) − µ∗(A ∩ Gc).
j=1
∑∞
By countable subadditivity, µ∗(A ∩ G) ≤ j=1 µ∗ (A ∩
Ej ), and so
µ∗(A ∩ G) ≤ µ∗(A) − µ∗(A ∩ Gc)
The other inequality is automatic, so we have
µ∗(A ∩ G) = µ∗(A) − µ∗(A ∩ Gc).
Thus G ∈ M.
3’ Finally we need to show countable additivity. Since
∑∞
j=1 µ∗ (A ∩ Ej ) is pinched between two quantities we
now know are equal, we also have
∞
∑
µ∗(A ∩ Ej ) = µ∗(A ∩ G).
j=1
Finally putting A = G we get countable additivity of
µ∗ on M. □
This measure µ has the property of being complete: if
Z is a measurable set with measure zero, then every sub-
set of Z is measurable (with measure zero, of course).
Any measure that is not complete can be completed in
a natural way: see exercise 2 in the text.
 Measures on metric spaces
The theorem above is very nice, but how do you find
out which sets are measurable? Suppose that X is a
metric space. Then we would like at least all open sets
to be measurable. This is guaranteed by the following
condition that relates the metric and measure properties
of X. Before stating it we make the following

Definition 7.3. The Borel σ-algebra BX is the inter-

section of all σ-algebras containing all the open sets of
X. (Note that the intersection of any collection of σ-
algebras is itself a σ-algebra.) A measure defined on BX
is called a Borel measure on X.

Said another way, the Borel σ-algebra is the smallest

σ-algebra containing all the open sets of X.
Proposition 7.4. Let µ∗ be an outer measure on
the metric space X. Suppose that
(7.1) µ∗(A ∪ B) = µ∗(A) + µ∗(B)
whenever dist(A, B) > 0. Then all open sets in X
are measurable, and hence the induced measure µ is
a Borel measure.

• Such an exterior measure is called a metric exte-

rior measure.
• Recall that dist(A, B) is defined to be the infimum of
distances d(a, b) where a ∈ A and b ∈ B. For example
the distance in R between (0, 1) and (1, 2) is zero, even
though the sets are disjoint.
Proof: Let O be an open set. Define On to be the subset
of O of points distance > 1/n from the boundary of O.
We first claim that µ∗(O) = limn µ∗(On). (Note that
≥ is trivial here). Consider the ‘shells’
{ }
1 1
Sn = On \ On−1 = x ≥ d(x, ∂O) > ,
n−1 n
for n ≥ 2 with S1 = O1. Note that since O is open, we
have a formula
∪
O = ON ∪ Sn
n>N
for each N .
For even n, these sets are a positive distance apart,
by an easy triangle inequality argument. If x ∈ Sn and
y ∈ Sn+2 we have d(x, ∂O) ≤ d(x, y) + d(y, ∂O), and
1
so d(x, y) > n(n+1) . (Similarly for the odd n.)
We now split into two cases, depending on whether
µ∗(O) is finite or infinite. First suppose that µ∗(O) <
∞. Then, by condition (??), we have
∑
µ∗(S2n) = µ∗(∪nS2n) ≤ µ∗(O) < ∞.
n
Similarly for the odd shells:
∑
µ∗(S2n+1) ≤ µ∗(O) < ∞.
n
Hence the sum of the tails of these infinite series can
be made arbitrarily small. In particular, for any ϵ > 0
there exists N such that
∑
µ∗(Sn) < ϵ.
n>N
∪
Using subadditivity for O = ON ∪ n>N Sn
∑
µ∗(O) ≤ µ∗(ON ) + µ∗(Sn)
n>N
and thus
µ∗(ON ) ≥ µ∗(O) − ϵ,
proving the claim. If on the other hand, µ∗(O) = ∞,
then we have by subadditivity
∑
µ∗(O) ≤ µ∗(Sn)
n≥1

and so
∑
µ∗(Sn) = ∞,
n
and using the positive distance between the even and
odd slices as before, this implies that µ(On) → ∞.
We can apply the same argument to µ(A∩O), for any
set A, obtaining µ∗(A ∩ O) = limn µ∗(A ∩ On). Note
here that even though A ∩ O is not open, the identity
 ∪
O = ON ∪ Sn implies
n>N
∪
A ∩ O = A ∩ ON ∪ (A ∩ Sn),
n>N
which was all we needed.
Now, take any set A. Using condition (??) we see that
for any n
µ∗(A) ≥ µ∗(A ∩ On) + µ∗(A \ O).
Taking the limit as n → ∞, we obtain
µ∗(A) ≥ µ∗(A ∩ O) + µ∗(A \ O).
The inequality ≤ is immediate from subadditivity, so
we see that O is measurable. □
 We also give a result about approximation of measur-
able sets by open and closed sets.

Proposition 7.5. Let X be a metric space and µ

a Borel measure which is finite on all balls in X
of finite radius. Then for each Borel set E and
each ϵ > 0 there is a bigger open set O ⊃ E with
µ(O \ E) < ϵ, and a smaller closed set F ⊂ E such
that µ(E \ F ) < ϵ.

Let us refer to the two properties as outer regularity

and inner regularity.
Proof: We show that closed sets have these properties,
and the collection C of Borel sets having these properties
is a σ-algebra. The Borel σ-algebra is, by definition, a
sub-σ-algebra of C, proving our claim.
 For a closed set F , inner regularity is trivial. For outer
regularity, we write Fk = F ∩ B(x0, k) for each k ∈ N,
where x0 is an arbitrarily chosen point. Each Fk is
closed and has finite measure, since it is a subset of
B(x0, k). which by hypothesis has finite measure. For
each Fk we define a family of open neighbourhoods

Ok,l = {x ∈ X | d(x, Fk ) < 2−l },

which ‘shrink’ down to Fk .

Then, since Fk is closed, we have ∩l Ok,l = Fk . Ob-
serve that for any l we have the disjoint decomposition
∪
Ok,l = Fk ∪ Ok,i \ Ok,i+1
i≥l

(where every set here is Borel). Looking at this for

∪
l = 0, we see that there is some l(k) so µ( i≥l(k) Ok,i \
Ok,i+1) < ϵ2−k . Then we have

µ(Ok,l(k)) < µ(Fk ) + ϵ2−k .

Then define O = ∪k Ok,N (k), which is an open set, and

observe
∑ ∑
µ(O\F ) ≤ µ(Ok,N (k)\F ) ≤ µ(Ok,N (k)\Fk ) = ϵ.
k k
It is easy to check that C is closed under complemen-
tation, so it remains to check closedness under count-
able unions. Outer regularity is straightforward using
an ϵ2−k argument: Given E = ∪Ei, choose Oi with
µ(Oi \ Ei) < ϵ2−k . Then define O = ∪Oi, which is
automatically an open set. Then
µ(O \ E) = µ(∪iOi \ E)
≤ µ(∪Oi \ Ei)
∑
≤ µ(Oi \ Ei)
< ϵ.
 The same argument doesn’t work for inner regular-
ity, because an infinite union of closed sets need not be
closed. This argument does establish inner regularity
for finite unions, however.
This observation means that given a countable family
Ek in C, with E = ∪Ek , we may assume that it is in-
′
creasing. Now define Ek,n = Ek ∩ (B(x0, n) \ B(x0, n −
1)) and analogously En′ . Since µ(Ek,n
′
) is increasing (in
′
k) and ∪k Ek,n = En′ has finite measure, there exists
′
a k(n) such that µ(Ek(n),n ) > µ(En′ ) − ϵ2−n−1. We
′ ′
approximate Ek(n),n with a closed set Fn ⊂ Ek(n),n
to within ϵ2−n−1. Then the union of the Fn is closed
(since at most finitely many intersect in any given ball
B(x0, n), and F is closed iff F ∩ B(x0, n) is closed for
all n) and approximates E within ϵ. □
 Premeasures
We have seen how to construct a measure from an ex-
terior measure. An exterior measure may in turn be
constructed from a more basic object called a premea-
sure. This is defined on an algebra, rather than a σ-
algebra, of subsets of X, that is, a collection of subsets
of X containing the empty set and closed under finite
unions and complementation. Let A be an algebra of
subsets of X. Then a premeasure µ0 : A → [0, ∞] is a
function satisfying
(i) µ0(∅) = 0;
(ii) If E1, E2, . . . is a countable collection of disjoint
subsets of A, and ∪k Ek ∈ A, then
( ) ∑
µ0 ∪n En = µ0(En).
n
Given a premeasure µ0, we define for any set E ⊂ X
{∑
∞ }
µ∗(E) = inf µ0(Ej ) | E ⊂ ∪j Ej , and each Ej ∈ A .
n=1
(Observe the infimum is over a nonempty set, because
X ∈ A covers any set E.)
Theorem 7.6. The function µ∗ is an exterior mea-
sure, such that
(i) µ∗(E) = µ0(E) for all E ⊂ A;
(ii) Every A ⊂ A is measurable.
Consequently, µ0 extends to a measure on the σ-
algebra generated by A.

This, if you recall, was precisely how Lebesgue mea-

sure was defined.
Proof: It is straightforward to show that µ∗ is an ex-
terior measure. To show that µ∗(E) = µ0(E) for all
E ∈ A, we take a cover of E by elements Ej ∈ A. Then
define Ek′ = E ∩ (Ek \ ∪k−1
j=1 E j ). Now E = ∪E ′
k and the
sets Ek′ are disjoint and all in A. We then have µ0(E) =
∑ ′
∑
j µ0 (Ej ), and therefore, µ0 (E) ≤ j µ0 (Ej ), since
Ej′ ⊂ Ej . Taking the inf of the right hand side shows
that µ0(E) ≤ µ∗(E). The ≥ inequality is easy, by
covering E using {E}.
 To show that E ∈ A is measurable, take any set A ⊂
X, and any covering Ej of A. Using finite additivity of
µ0 on A, we have
µ0(Ej ) = µ0(Ej ∩ E) + µ0(Ej ∩ E c).
Summing in j, we find that
∑ ∑ ∑
µ0(Ej ) = µ0(Ej ∩ E) + µ0(Ej ∩ E c)
j
≥ µ∗(A ∩ E) + µ∗(A ∩ E c).
Since this is true for every cover of A, we find that
µ∗(A) ≥ µ∗(A ∩ E) + µ∗(A ∩ E c),
which shows that E is measurable since the ≤ inequality
is obvious. □
 For later use we note the following. Let Aσ be the
collection of countable unions of elements of A and let
Aσδ denote the collection of countable intersections of
elements of Aσ . (Of course, if A were actually a σ-
algebra, then we would have A = Aσ = Aσδ .) Then

Proposition 7.7. For any set A, and any ϵ > 0,

there exists an E1 ∈ Aσ such that A ⊂ E1 and
µ∗(E1) < µ∗(A) + ϵ, and there exists E2 ∈ Aσδ such
that A ⊂ E2 and µ∗(A) = µ∗(E2).

Exercise. What is the relation between the Borel

sets of R and the Lebesgue measurable sets of R?