4.

Radon-Nikodym Theorems
In this section we discuss a very important property which has many important
applications.
Definition. Let X be a non-empty set, and let A be a -algebra on X. Given
two measures and on A, we say that has the Radon-Nikodym property relative
to , if there exists a measurable function f : X [0, ], such that
Z
(1)
(A) =
f d, A A.
A

Here we use the convention which defines the integral in the right hand side by
 R
Z
f A d if f A L1+ (X, A, )
X
f d =

if f A 6 L1+ (X, A, )
A
In this case, we say that f is a density for relative to .
The Radon-Nikodym property has an equivalent useful formulation.
Proposition 4.1 (Change of Variables). Let X be a non-empty set, and let
A be a -algebra on X, let and be measures on A, and let f : X [0, ] be a
measurable function.
A. The following are equivalent
(i) has the Radon-Nikodym property relative to , and f is a density for
relative to ;
(ii) for every measurable function h : X [0, ], one has the equality1
Z
Z
(2)
h d =
hf d.
X

B. If and f are as above, and K is either R or C, then the equality (2)

also holds for those measurable functions h : X K with h L1K (X, A, ) and
hf L1K (X, A, ).
Proof. A. (i) (ii). Assume property (i) holds, which means that we have
(1). Fix a measurable function h : X [0, ], and use Theorem III.3.2, to find a
sequence (hn )
n=1 A-ElemR (X), with
(a) 0 h1 h2 h;
(b) limn hn (x) = h(x), x X.
Of course, we also have
(a0 ) 0 h1 f h2 f hf ;
(b0 ) limn hn (x)f (x) = h(x)f (x), x X.
Using the Monotone Convergence Theorem, we then get the equalities
Z
Z
Z
Z
(3)
h d = lim
hn d and
hf d = lim
hn f d
X

1 For the product hf we use the conventions 0 = 0 = 0, and t = t = ,

t (0, ].
300

4. Radon-Nikodym Theorems

301

Pp
Notice that, if we fix n and we write hn = k=1 k Ak , for some A1 , . . . , Ap A,
and 1 > > p > 0, then
Z
Z
p
p Z
X
X
hn d =
k (Ak ) =
k Ak f d =
hn f d,
X

k=1

k=1

so using (3), we immediately get (2).

The implication (ii) (i) is trivial, using functions of the form h = A , A A.
B. Suppose has the Radon-Nikodym property relative to , and f is a density
for relative to , and let h : X K be a measurable function with h L1K (X, A, )
and hf L1K (X, A, ). In the complex case, using the inequalities |Re h| |h| and
|Im h| |h|, it is clear that both functions Re h and Im h belong to L1 (X, A, ),
and also the products (Re h)f and (Im h)f belong to L1 (X, A, ). This shows that
it suffices to prove (2) under the additional hypothesis that h is real-valued. In this
case we consider the functions h , defined by
h+ = max{h, 0} and h = max{h, 0}.
Since we have 0 h |h|, it follows that h L1+ (X, A, ), as well as h f
L1+ (X, A, ). In particular, we get the equalities
Z
Z
Z
Z
Z
Z
+

+
(4)
h d =
h d
h d and
hf d =
h f d
h f d.
X

Since h 0, we can use property A.(ii) above, and we have

Z
Z

h d =
h f d,
X

and then the desired equality (2) immediately follows from (4).

One important issue is the uniqueness of the density. For this purpose, it will
be helpful to introduce the following.
Definition. Let T be one of the spaces [, ] or C, and let r be some
relation on T (in our case r will be either =, or , or , on [, ]).
Given a measurable space (X, A, ), and two measurable functions f1 , f2 : X T ,
f1 r f2 , -l.a.e.
if the set


A = x X : f1 (x) r f2 (x)
belongs to A, and it has locally -null complement in X, i.e. [X r A] F ) = 0,
for every set F A with (F ) < . (If r is one of the relations listed above, the
set A automatically belongs to A, so all intersections [X r A] F , F A, also
belong to A.) The abreviation -l.a.e. stands for -locally-almost everywhere.
Remark that one has the implication
f1 r f2 , -a.e. f1 r f2 , -l.a.e.
Remark that, when is -finite, then the other implication also holds:
f1 r f2 , -l.a.e. f1 r f2 , -a.e.
With this terminology, one has the following uniqueness result.

302

CHAPTER IV: INTEGRATION THEORY

Proposition 4.2. Suppose A is a algebra on some non-empty set X, and

and are measures on A, such that has the Radon-Nikodym property relative to
. If f, g : X [0, ] are densities for relative to , then
f = g, -l.a.e.
In particular, if is -finite, then
f = g, -a.e.


Proof. Consider the set B = x X : f (x) 6= g(x) , which belongs to A.
We need to prove that B is locally -null, i.e. one has (B F ) = 0, for all F A
with (F ) < . Fix F A with (F ) < , and let us write B F = D E,
where




D = x B F : f (x) < g(x) and E = x B F : f (x) > g(x) .
If we define, for each integer n 1, the sets




Dn = x B F : f (x) + n1 g(x) and En = x B F : f (x) g(x) + n1 ,
then it is clear that
BF =DE =

(Dn En ),

n=1

so in order to prove that (B F ) = 0, it suffices to show that (Dn ) = (En ) = 0,

n 1.
Fix n 1. It is obvious that f (x) < , x Dn , so if we define the sequence
(Dnk )
k=1 A, by


Dnk = x Dn : f (x) k , k 1,
S
we have the equality Dn = k=1 Dnk , so in order to prove that (Dn ) = 0, it suffices
to show that (Dnk ) = 0, k 1. On the one hand, since f (x) k, k 1, using
the inclusion Dnk F , we get
Z
Z
(Dnk ) =
f d
k Dnk d = k(Dnk ) k(F ) < .
k
Dn

On the other hand, since g(x) f (x) + n1 , x Dnk , we get

Z
Z
(Dnk ) =
g d
(f Dnk + n1 Dnk ) d =
k
Dn

Z
=
X

Z
f Dnk d +

1
k
n Dn

d = (Dnk ) + n1 (Dnk ).

(Dnk )

Since
< , the above inequality forces (Dnk ) = 0.
The fact that (En ) = 0, n 1, is proven the exact same way.

In general, the uniqueness of the density does not hold -a.e., as it is seen from
the following.

Example 4.1. Take X to be some non-empty set, put A = {, X , and define
the measure on A, by () and (X) = . It is clear that has the RadonNikodym property realtive to itself, but as sensities one can choose for instance the
constant functions f = 1 and g = 2. Clearly, the equality f = g, -a.e. is not true.

4. Radon-Nikodym Theorems

303

Remark 4.1. The local almost uniqueness result, given in Proposition 4.2,
holds under slightly weaker assumptions. Namely, if (X, A, ) is a measure space,
and if f, g : X [0, ] are measurable functions for which we have the equality
Z
Z
f d =
g d,
A

for all A A with (A) < , then we still have the equality f = g, -l.a.e.
This


follows actually from Proposition 4.2, applied to functions of the form f A and g A .
Let us introduce the following.
Notations. For a measure space (X, A, ) we define
A0 = {N A : (N ) = 0};
Afin = {F A : (F ) < };
A0,loc = {A A : (A F ) = 0, F Afin }.
With these notations, we have the inclusions
A0 = A0,loc Afin A0,loc A,
and A0 and A0,loc are in fact -rings.
Comment. The locally-almost everywhere terminology is actually designed
to hide some pathologies under the rug. For instance, if (X, A, ) is a degenerate
measure space , i.e. (A) {0, }, A A, then anything happens locally
almost-everywhere, which means that we have the equality A0,loc = A.
At the other end, there is a particular type of measure spaces on which, even in
the absence of -finiteness, the notions of locally-almost everywhere and almost
everywhere coincide, i.e. we have the equality A0,loc = A0 . Such spaces are
described by the following.
Definition. A measure space (X, A, ) is said to be nowhere degenerate, or
with finite subset property, if
(f) for every set A A with (A) > 0, there exists some set F A, with
F A, and 0 < (F ) < .
With this terminology, one has the following result.
Proposition 4.3. For a measure space (X, A, ), the following are equivalent:
(i) A0,loc = A0 ;
(ii) (X, A, ) has the finite subset property.
Proof. (i) (ii). Assume A0,loc = A0 , and let us prove that (X, A, ) has
the finite subset property. We argue by contradiction, so let us assume there exists
some set A A, with (A) = , such that (B) {0, }, for every B A, with
B A. In particular, if we start with some arbitrary F Afin , using the fact
that (A F ) (F ) < , we see that we must have (A F ) = 0. This proves
precisely that A A0,loc . By assumption, it follows that A A0 , i.e. (A) = 0,
which is impossible.
(ii) (i). Assume that (X, A, ) has the finite subset property, and let us
prove the equality (i). Since one inclusion is always true, all we need to prove is
the inclusion A0,loc A0 , which equivalent to the inclusion A0,loc Afin . Start
with some set A A0,loc , but assume (A) = . On the one hand, using the finite

304

CHAPTER IV: INTEGRATION THEORY

subset property, there exists some set F A with F A and (F ) > 0. On the
other hand, since A A0,loc , we have (F ) = 0, which is impossible.

Example 4.2. Take X be an uncountable set, let A = P(X), and let be the
counting measure, i.e.

Card A if A is finite
(A) =

if A is infinite
Then (X, P(X), ) has the finite subset property, but is not -finite.
When we restrict to integrable functions, the two notions -l.a.e, and -a.e.
coincide. More precisely, we have the following.
Proposition 4.4. Let (X, A, ) be a measure space, let K be one of the fields
R or C, and let p [1, ). For a function f LpK (X, A, ), the following are
equivalent:
(i) f = 0, -l.a.e.
(ii) f = 0, -a.e.
Proof. Of course, we only need to prove the implication (i) (ii). Assume
f = 0, -l.a.e. Using the function g = |f |p , we can assume that p = 1 and f (x) 0,
x SX. Consider then the set N = {x X : f (x) > 0}, and write it as a union

N = n=1 Nn , where
Nn = {x X : f (x)

1
n },

n 1.

Of course, all we need is the fact that (Nn ) = 0, n 1. Fix n 1. On the one
hand, the assumption on f , it follows that Nn A0,loc . On the other hand, the
inequality n1 Nn f , forces the elementary function n1 Nn to be -integrable, i.e.
(Nn ) < . Consequently we have
N A0,loc Afin = A0 . 
Comment. In what follows we will discuss several results, which all have as
conclusion the fact that one measure has the Radon-Nikodym property with respect
to another one. All such results will be called Radon-Nikodym Theorems.
The first result is in fact quite general, in the sense that it works for finite
signed or complex measures.
Theorem 4.1 (Easy Radon-Nikodym Theorem). Let (X, A, ) be a finite
measure space, let K denote one of the fields R or C, and let C > 0 be some
constant. Suppose is a K-valued measure on A, such that
|(A)| C(A), A A.
Then there exists some function f L1K (X, A, ), such that
Z
(5)
(A) =
f d, A A.
A

Moreover:
(i) Any function f L1K (X, A, ), satisfying (5) has the property |f | C, a.e. If is an honest measure, then one also has the inequality |f | 0,
-a.e.
(ii) A function satisfying (5) is essentially unique, in the sense that, whenever
f1 , f2 L1K (X, A, ) satisfy (5), it follows that f1 = f2 , -a.e.

4. Radon-Nikodym Theorems

305

R
Proof. The ideea is to somehow make sense of X h d, for suitable measurable
functions
h, and to examine the properties of such a number relative to the integral
R
h
d.
The
second integral is of course defined, for instance for h L1K (X, A, ),
X
but the first integral is not, because is not an honest measure. The proof will
be carried on in several steps.
Step 1: There exist four honest finite measures k , k = 1, 2, 3, 4, and numbers k , k = 1, 2, 3, 4, such that = 1 1 + 2 2 + 3 3 + 4 4 , and
k C, k = 1, 2, 3, 4.

(6)

In the case K = R we use the Hahn-Jordan decomposition = + . We also

know that ||, the variation measure of . In this case we take 1 = 1,
1 = + , 2 = 1, 2 = , and we set 3 = 4 = 0, 3 = 4 = 0.
In the case K = C, we write = + i, with and finite signed measures,
and we use the Hahn-Jordan decompositions = + and = + . We
also know that the variation measures of and satisfy || || and || ||, so
we also have || and ||. In this case we can then take 1 = 1, 1 = + ,
2 = 1, 2 = , 3 = i, 3 = + , 4 = i, 4 = .
Notice that in either case we have
k ||, k = 1, 2, 3, 4.
By Remark III.8.5 it follows that we have || C, so we immediately get the
inequalities (6).
Step 2: For any measurable function h : X [0, ], one has the inequality
Z
Z
(7)
h dk C
h d, k = 1, 2, 3, 4.
X

To prove this, we choose a sequence of elementary functions (hn )

n=1 A-ElemR (X),
with
0 h1 h2 . . . (everywhere),
limn hn (x) = h(x), x X,
so that by the General Monotone Convergence Theorem, we get the equalities
Z
Z
Z
Z
h d = lim
hn d and
h dk = lim
hn dk , k = 1, 2, 3, 4.
X

This means that, in order to prove (7), it suffices to prove it under the extra
assumption that h is elementary. In this case, we have
h = 1 B1 + + p Bp ,
with 1 , . . . , p 0 and B1 , . . . , Bp A. The inequality is then immediate, from
(6) since we have
Z
Z
p
p
X
X
h dk =
j k (Bj ) C
(Bj ) = C
h d.
X

j=1

j=1

As a consequence of Step 2, we get the fact that, for every k = 1, 2, 3, 4, one

has the inclusions
L1K (X, A, ) L1K (X, A, k ) and NK (X, A, ) NK (X, A, k ).
Taking quotients, this gives rise to correctly defined linear maps
(8)

k : L1K (X, A, ) 3 h 7 h L1K (X, A, k ), k = 1, 2, 3, 4.

306

CHAPTER IV: INTEGRATION THEORY

(Here we use the abusive notation that identifies an element in L1 with a function
in L1 , which is defined almost uniquely.) Moreover, one has the inequality
Z
Z
|h| dk C
|h| d, h L1K (X, A, ), k = 1, 2, 3, 4,
X

in other words, the linear maps (8) are all continuous. For every k = 1, 2, 3, 4, let
k denote the integration map
Z
1
k : LK (X, A, k ) 3 h 7
h dk K.
X

We know (see Remark 3.5) that the k s are continuous. In particular, the compositions k = k k : L1K (X, A, ) K, which are defined by
Z
k : L1K (X, A, ) 3 h 7
h dk , k = 1, 2, 3, 4,
X

are linear and continuous.

We now use Proposition 3.3 which states that one has an inclusion
: L2K (X, A, ) , L1K (X, A, ),

(9)

which is in fact a linear continuous map. So if we consider the compositions k =

k , which are defined by
Z
k : L1K (X, A, ) 3 h 7
h dk , k = 1, 2, 3, 4,
X

then these compositions are linear and continuous. Apply then Riesz Theorem (in
the form given in Remark 3.4), to find functions f1 , f2 , f3 , f4 L2K (X, A, ), such
that
k (h) = hfk , hi, h L2K (X, A, ), k = 1, 2, 3, 4.
In particular, using functions of the form h = A , A A (which all belong to
L2K (X, A, ), due to the finiteness of ), we get
Z
Z
k (A) =
A dk =
fk A d, A A, k = 1, 2, 3, 4.
X

Finally, if we define the function f = 1 f1 + 2 f2 + 3 f3 + 4 f4 L2K (X, A, ),

then the above equalities immediately give the equality (5).
At this point we only know that f belongs to L2K (X, A, ). Using the inclusion
(9), it turns out that f indeed belongs to L1K (X, A, ).
Let us prove now the additional properties (i) and (ii).
To prove the first assertion in (i), we start off by fixing some function f
L1K (X, A, ), which satisfies (5), and we define the set

A = {x X : |f (x)| > C ,
for which we must prove that (A) = 0. Since f is measurable, it follows that A
belongs to A. Consider the rational unit sphere SQ1 in K, defined as

{1, 1}
if K = R
(10)
SQ1 =
{e2it : t Q} if K = C
The point is that SQ1 is dense in the unit sphere S 1 in K:
S 1 = { K : || = 1},

4. Radon-Nikodym Theorems

so we immediately have the equality A =

307

A , where

A = {x X : Re[f (x)] > C .
1
SQ

Since SQ1 is countable, in order to prove that (A) = 0, it then suffices to show that
(A ) = 0, SQ1 . Fix then SQ1 , and consider the K-valued measure = .
It is clear that we still have
(11)

|(A)| = |(A)| C(A), A A,

as well as the equality

Z
(12)

f d, A A.

(A) =
A

For each integer n 1, let us define the set

An = {x X : Re[f (x)] C + n1 ,
S
so that we obviously have the equality A = n=1 An . In particular, in order to
prove (A ) = 0, it suffices to prove that (An ) = 0, n 1. Fix for the moment
n 1. Using (12), it follows that
Z
 Z
Z
Re (An ) = Re
f d =
Re[f ] d =
Re[f ] An d.
An

An

1
,
n ) An

Since we have Re[f ] An (C +

the above inequality can be continued
with
Z
Re (An )
(C + n1 ) An d = (C + n1 )(An ).
X

Of course, this will give

|(An )| Re (An ) (C + n1 )(An ).
Note now that, using (11), this will finally give
C(An ) (C + n1 )(An ),
which clearly forces (An ) = 0.
Having proven that |f | C, -a.e., let us turn our attention now to the uniqueness property (ii). Suppose f1 , f2 L1K (X, A, ) are such that
Z
Z
(A) =
f1 d =
f2 d, A A.
A

Consider then the difference f = f1 f2 and the trivial measure 0 = 0. Obviously

we have
|0 (A)| n1 (A), A A,
for every integer n 1, as well as
Z
0 (A) =
f d, A A.
A

By the first assertion in (i), it follows that

1
, -a.e.,
n
for every n 1. So if we take the sets (Nn )
n=1 A defined by
|f1 f2 | = |f |

Nn = {x X : |f1 (x) f2 (x)| >

1
n },

308

CHAPTER IV: INTEGRATION THEORY

S
then (Nn ) = 0, n 1. Of course, if we put N = n=1 Nn , then on the one hand
we have (N ) = 0, and on the other hand, we have
f1 (x) f2 (x) = 0, x X r N,
which means that we indeed have f1 = f2 , -a.e.
Finally, let us prove the second assertion in (i), which starts with the assumption
that is an honest measure. Let f L1K (X, A, ) satisfy (5). By the uniqueness
property (ii), it follows immediately that
f = Re f, -a.e.,
so we can assume that f is already real-valued. Consider the honest measure
= C , and notice that the function g : X R defined by
g(x) = C f (x), x X,
clearly has the property
Z
(A) =

g d, A A.

Since we obviously have

0 (A) C(A), A A,
by the first assertion of (i), applied to the measure and the function g, it follows
that |g| C, -a.e. In other words, we have now a combined inequality:


max |f |, |C f | C, -a.e.
Of course, since f is real valued, this forces f 0, -a.e.

In what follows we are going to offer various generalizations of Theorem 4.1.

There are several directions in which Theorem 4.1 can be generalized. The main
direction, which we present here, will aim at weakening the condition || C.
The following result explains that in fact the case of K-valued measures can be
always reduced to the case of honest finite ones.
Proposition 4.5 (Polar Decomposition). Let A be a -algebra on a non-empty
set X, let K be one of the fields R or C, and let be a K-valued measure on A. Let
|| denote the variation measure of . There exists some function f L1K (X, A, ||),
such that
Z
(13)
(A) =
f d||, A A.
A

Moreover
(i) Any function f L1K (X, A, ||), satisfying (13) has the property |f | = 1,
||-a.e.
(ii) A function satisfying (13) is essentially unique, in the sense that, whenever f1 , f2 L1K (X, A, ||) satisfy (13), it follows that f1 = f2 , ||-a.e.
Proof. We know that
|(A)| ||(A), A A.
So if we apply Theorem 4.1 for the finite measure = || and C = 1, we immediately
get the existence of f L1K (X, A, ||), satisfying (13). Again by Theorem 4.1, the

4. Radon-Nikodym Theorems

309

uniqueness property (ii) is automatic, and we also have |f | 1, ||-a.e. To prove

the fact that we have in fact the equality |f | = 1, ||-a.e., we define the set
A = {x X : |f (x)| < 1},
which belongs to A, and we prove that ||(A) = 0. If we define the sequence of sets
(An )
n=1 A, by
An = {x X : |f (x)| 1 n1 }, n 1,
S
then we clearly have A = n=1 An , so all we have to show is the fact that ||(An ) =
0, n 1. Fix n 1. For every B A, with B An , we have
|f (x)| 1 n1 , x B,
so using (13) we get
Z
 Z
Z


|(B)| = 
f d||
|f | d||
(1 n1 ) d|| = (1 n1 )||(B).
B
B
B
S
Now if we take an arbitrary pairwise disjoint sequence (Bk )
k=1 A, with
k=1 Bk =
An , then the above estimate will give

X
X
|(Bk )| (1 n1 )
||(Bk ) = (1 n1 )||(An ).
k=1

k=1

Taking supremum in the left hand side, and using the definition of the variation
measure, the above estimate will finally give
||(An ) (1 n1 )||(An ),
which clearly forces ||(An ) = 0.

Remark 4.2. The case K = R can be slighly generalized, to include the case
of infinite signed measures. If is a signed measure on A and if we consider the
Hahn-Jordan set decomposition (X + , X ), then the density f is simply the function

1
if x X +
f (x) =
1 if x X
The equality (13) will then hold only for those sets A A with ||(A) < .
Since || is allowed to be infinite, as explained in Example 4.1, the only version of
uniqueness property (ii) will hold with ||-l.a.e in place of ||-a.e Likewise, the
absolute value property (i) will have to be replaced with |f | = 1, ||-l.a.e
Comment. Up to this point, it seems that the hypotheses from Theorem 4.1
are essential, particularly the dominance condition || C. It is worth discussing
this property in a bit more detail, especially having in mind that we plan to weaken
it as much as possible.
Notation. Suppose A is a -algebra on some non-empty set X, and suppose
and are honest (not necessarily finite) measures on A. We shall write
b ,
if there exists some constant C > 0, such that
(A) C(A), A A.
A few steps in the proof of Theorem 4.1 hold even without the finiteness assumption, as indicated by the follwing.

310

CHAPTER IV: INTEGRATION THEORY

Exercise 1*. Suppose A is a algebra on some non-empty set X, and suppose

and are honest measures on A. Prove the following.
(i) If b , then one has the inclusions
NK (X, A, ) NK (X, A, ) and LpK (X, A, ) LpK (X, A, ), p [1, ).
Consequently (see the proof of Theorem 4.1) one has linear maps
LpK (X, A, ) 3 h 7 h LpK (X, A, ), p [1, ).
Show that these linear maps are continuous.
(ii) Conversely, assuming one has the inclusion
LpK0 (X, A, ) LpK0 (X, A, ),
for some p0 [1, ), prove that b .
Hint: To prove (ii) show first one has the inclusion L1+ (X, A, ) L1+ (X, A, ). Then show that
the quantity
Z

Z
C = sup
h d : h L1+ (X, A, ),
h d 1
X

1
is finite. If C = , there exists some sequence (hn )
n=1 L+ (X, A, ), with
Z
Z
hn d 1 and
h d 4n , n 1.
X

Consider then the series

A A.

1
n=1 2n hn ,

and get a contradiction. Finally prove that (A) C(A),

It is the moment now to introduce the following relation, which is a highly

non-trivial weakening of the relation b.
Definition. Let A is a -algebra on some non-empty set X, and suppose
and are honest (not necessarily finite) measures on A. We say that is
absolutely continuous with respect to , if for every A A, one has the implication
(14)

(A) = 0 = (A) = 0.

In this case we are going to use the notation

 .
It is obvious that one always has the implication
b  .
Remarks 4.3. Let (X, A, ) be a measure space. A. If is an honest measure
on A, which has the Radon-Nikodym property relative to , then  . This is
pretty obvious, since if we pick f : X [0, ] to be a density for realtive to ,
then for every A A with (A) = 0, we have f A = 0, -a.e., so we get
Z
Z
(A) =
f d =
f A d = 0.
A

B. For an honest measure on A, the relation  is equivalent to the

inclusion
NK (X, A, ) NK (X, A, ).
By Exercise 1, this already suggests that the relation  is much weaker than b
(see Exercise 2 below).
C. If is either a signed or a complex measure on A, then the following are
equivalent:

4. Radon-Nikodym Theorems

311

(i) the variation measure || is absolutely continuous with respect to ;

(ii) for every A A, one has the implication (14)
The implication (i) (ii) is trivial, since one has
|(A)| ||(A), A A.
The implication (ii) (i) is also clear, since if we start with some A A with
(A) = 0, then we get |(B)| = 0, for all B A with B A, and then arguing
exactly as in the proof of Proposition 4.3, we get ||(A) = 0.
Convention. Using Remark 4.2.A, we extend the definition of absolute continuity, and the notation  to include the case when is either a signed
measure, or a complex measure on A. In other words, the notation  means
that ||  .
The following techincal result is key for the second Radon-Nikodym Theorem.
Lemma 4.1. Let (X, A, ) be a finite measure space, and let be an honest
measure on A, with  . Then there exists a sequence (n )
n=1 , of honest
measures on A, such that
(i) n b , n 1; in particular the measures n , n 1 are all finite;
(ii) 1 2 . . . ;
(iii) limn n (A) = (A), A A.
Proof. Let us define
n = (n) , n 1.
Recall (see III.8, the Lattice Property; it is essential here that one of the measures,
namely n, is finite) that by construction n has the following properties:
(a) n n and n ;
(b) whenever is a measure with n and , it follows that n .
Property (a) above already gives condition (i). It will be helpful to notice that
property (a) also gives the inequality
(15)

n , n 1.

The monotonicity condition is now trivial, since by (b) the inequalities n1

(n 1) n and n1 , imply n1 (n) = n .
To derive property (iii), it will be helpful to recall the actual definition of the
operation . Fix for the moment n 1. One first considers the signed measure

n = n , and its Hahn-Jordan decomposition n = +

n . In our case, we
+

get n n and n . With these notations the measures n are defined by

n = n +
n , n 1. If we fix, for each n 1, a Hahn-Jordan set decomposition
(Xn+ , Xn ) for X relative to n , then we have
n (A) = (A Xn+ ) + n(A Xn ), A A, n 1.
S
T
+

Consider then the sets X

= n=1 Xn+ and X
= n=1 . It is clear that X
A,

+
and X = X r X .
Fix now a set A A, and let us prove the equality (iii). On the one hand, the

obvious inclusions Xn X
, combined with (16), give the inequalities
(16)

(17)

n (A) (A Xn+ ) + n(A X

), n 1.

312

CHAPTER IV: INTEGRATION THEORY

On the other hand, since n+1 = + n , n 1, using Lemma III.8.2, we get the
relations
X1+ X2+ . . . .

(Recall that the notation D E stands for (D r E) = 0.) Since  , we also

have the relations

A X1+ A X2+ . . . ,

so using Proposition III.4.3, one gets the equality

+
(A X
) = lim (A Xn+ ).
n

Combining this with the inequalities (15) and (17) then yields the inequality


+

(18) (A) lim sup n (A) lim inf n (A) (A X

) + lim n(A X
) .
n

There are two posibilities here.

Case I : (A X
) > 0.
In this case, the estimate (18) forces
(A) = lim sup n (A) = lim inf n (A) = .
n

Case II : (A X
) = 0.

In this case, using absolute continuity, we get (A X

) = 0, and the equality
+

A = (A X ) (A X ) yields
+
(A) = (A X
).

Then (18) forces

lim sup n (A) = lim inf n (A) = (A).
n

In either case, the concluison is the same: limn n (A) = (A).

After the above preparation, we are now in position to prove the following.
Theorem 4.2 (Radon-Nikodym Theorem: the finite case). Let (X, A, ) be a
finite measure space.
A. If is an honest measure on A, with  , then there exists a measurable
function f : X [0, ], such that
Z
(19)
(A) =
f d, A A.
A

Moreover, such a function is essentially unique, in the sense that, whenever f1 , f2 :

X [0, ] are measurable functions, that satisfy (19), it follows that f1 = f2 ,
-a.e.
B. Let K be either R or C. If is a K-valued measure on A, with  , then
there exists a function f L1K (X, A, ), such that
Z
(20)
(A) =
f d, A A.
A

Moreover:
(i) A function f L1K (X, A, ) satisfying (20) is essentially unique, in the
sense that, whenever f1 , f2 L1K (X, A, ) satisfy (20), it follows that
f1 = f2 , -a.e.

4. Radon-Nikodym Theorems

313

(ii) If f L1K (X, A, ) is any function satisfying (20), then the variation
measure || of is given by
Z
||(A) =
|f | d, A A.
A

Proof. A. Use Lemma 4.1 to find a sequence (n )

n=1 of honest measures
on A, such that
n b , n 1; in particular the measures n , n 1 are all finite;
1 2 . . . ;
limn n (A) = (A), A A.
For each n 1, we apply the Easy Radon-Nikodym Theorem 4.1, to find some
measurable function fn : X R, such that
Z
n (A) =
fn d, A A.
A

Claim: The sequence (fn )

n=1 satisfies
0 fn fn+1 , -a.e., n 1.
Fix n 1. On the one hand, since the n s are honest finite measures, and
n b , by part (i) of Theorem 4.1, it follows that fn 0, -a.e. On other hand,
since n+1 n is also an honest finite measure with n+1 n b , and with
density fn+1 fn , again by part (i) of Theorem 4.1, it follows that fn+1 fn 0,
-a.e.
Having proven the above Claim, let us define the function f : X [0, ], by


f (x) = lim inf max{fn (x), 0} x X.
n

It is obvious that f is measurable. By the Claim, we have in fact the equality

f = -a.e.- lim fn .
n

Since we also have

f A = -a.e.- lim fn A , A A,
n

using the Claim and the Monotone Convergence Theorem, we get

Z
Z
Z
Z
f d =
f A d = lim
fn A d = lim
fn d =
A

= lim n (A) = (A), A A.

n

Having shown that f satisfies (19), let us observe that the uniqueness property
stated in part A is a consequence of Proposition 4.2.
B. Let be a K-valued. In particular, the variation measure || is finite, so by
the Polar Decomposition (Proposition 4.3) there exists some measurable function
h : X K, such that
Z
(21)
(A) =
h d||, A A,
A

and such that |h| = 1, ||-a.e. Replacing h with the measurable function h0 : X
K, defined by

h(x) if |h(x)| = 1
0
h (x) =
1
if |h(x)| =
6 1

314

CHAPTER IV: INTEGRATION THEORY

we can assume that in fact we have

|h(x)| = 1, x X.
Apply then part A, to the measure ||, which is again absolutely continuous with
respect to , to find some measurable function g : X [0, ], such that
Z
||(A) =
g d, A A.
A

Remark that, since

Z
g d = ||(X) < ,
X

it follows that g L1+ (X, A, ). Fix for the moment some set A A. On the one
hand, since
|h A | 1,

(22)

and || is finite, it follows that h A L1K (X, A, ||). On the other hand, since
g L1+ (X, A, ), using (22) we get the fact that h A g L1K (X, A, ). Using the
Change of Variable formula (Proposition 4.1) we then get the equality
Z
Z
h A d|| =
h A g d,
X

which by (21) reads:

Z
(A) =

hg d.
A

Now the function f0 = hg (which has |f0 | = g) belongs to L1K (X, A, ), and clearly
satisfies (20).
To prove the uniqueness property (i), we start with two functions f1 , f2
L1K (X, A, ) which satisfy
Z
Z
f1 d =
f2 d = (A), A A.
A

If we define the function = f1 f2 L1K (X, A, ), then we clearly have

Z
Z
d =
0 d = (A), A A,
A

where is the zero measure. Since , using Theorem 4.1 it follows that = 0,
-a.e.
To prove (ii) we start with some f L1K (X, A, ) that satisfies (20), and we
use the uniqueness property (i) to get the equality f = f0 , -a.e., where f0 is the
function constructed above. In particular, using the construction of f0 , the fact
that |f0 | = g, and the fact that g is a density for || relative to , we get
Z
Z
Z
|f | d =
|f0 | d =
g d = ||(A), A A. 
A

At this point we would like to go further, beyond the finite case. The following
generalization of Theorem 4.2 is pretty straightforward.

4. Radon-Nikodym Theorems

315

Corollary 4.1 (Radon-Nikodym Theorem: the -finite case). Let (X, A, )

be a -finite measure space.
A. If is an honest measure on A, with  , then there exists a measurable
function f : X [0, ], such that
Z
(23)
(A) =
f d, A A.
A

Moreover, such a function is essentially unique, in the sense that, whenever f1 , f2 :

X [0, ] are measurable functions, that satisfy (19), it follows that f1 = f2 ,
-a.e.
B. Let K be either R or C. If is a K-valued measure on A, with  , then
there exists a function f L1K (X, A, ), such that
Z
(24)
(A) =
f d, A A.
A

Moreover:
(i) A function f L1K (X, A, ) satisfying (20) is essentially unique, in the
sense that, whenever f1 , f2 L1K (X, A, ) satisfy (24), it follows that
f1 = f2 , -a.e.
(ii) If f L1K (X, A, ) is any function satisfying (24), then the variation
measure || of is given by
Z
||(A) =
|f | d, A A.
A

S Proof. Since is -finite, there exists a sequence (An )n=1 Afin , with
An1 ), n 2. Then
n=1 An = X. Put X1 = A1 and Xn = An r (A1 S

(Xn )

A
is
pairwise
disjoint,
and
we
still
have
n=1
n=1 Xn = X. The Corolfin
lary follows
then
immediately
from
Theorem
4.2,
applied
to the measure spaces




(Xn , AXn , Xn ) and the measures Xn and Xn respectively. What is used here
is the fact that, if K denotes one of the sets [0, ], R or C, then for a function

f : X K the fact that f is measurable, is equivalent to the fact that f X is
n
measurable for each n 1. Moreover, given two functions
f1 , f2 : X K, the con
dition f1 = f2 , -a.e. is equivalent to the fact that f1 Xn = f2 Xn , -a.e., n 1.
1
Finally, forf : X K(= R, C), the
 condition f LK (X, A, ), is equivalent to the
1



fact that f X LK (Xn , A X , X ), n 1, and
n

Z
X
n=1

Xn





f Xn d Xn < . 

Comment. The -finite case of the Radon-Nikodym Theorem, given above, is

in fact a particular case of a more general version (Theorem 4.3 below). In order
to formulate this, we need a concept which has already appeared earlier in III.5.
Recall that a measure space (X, A, ) is said to be decomposable, if there exists a
pairwise disjoint subcollection F Afin , such that
S
(i) F F F = X;
(ii) for a set A X, the condition A A is equiavelnt to the condition
A F A, F F;

316

CHAPTER IV: INTEGRATION THEORY

(iii) one has the equality

(A) =

(A F ), Afin .

F F

Such a collection F is then called a decomposition of (X, A, ). Condition (ii)

is referred to as the patching property, because it characterizes measurability as
follows.
(p) Given a measurable space (Y, B), a function
f : (X, A) (Y, B) is mea
surable, if and only if all restrictions F F : (F, AF ) (Y, B), F F, are
measurable.
Theorem 4.3 (Radon-Nikodym Theorem: the decomposable case). Let (X, A, )
be a decomposable measure space. Let A-fin be the collection of all --finite sets
in A, that is,

[


A-fin = A A : there exists (An )

A
,
with
A
=
An .
n=1
fin
n=1

A. If is an honest measure on A, with  , then there exists a measurable

function f : X [0, ], such that
Z
(25)
(A) =
f d, A A-fin .
A

Moreover, such a function is locally essentially unique, in the sense that, whenever
f1 , f2 : X [0, ] are measurable functions, that satisfy (25), it follows that
f1 = f2 , -l.a.e.
B. Let K be either R or C. If is a K-valued measure on A, with  , then
there exists a function f L1K (X, A, ), such that
Z
(26)
(A) =
f d, A A-fin .
A

Moreover:
(i) A function f L1K (X, A, ) satisfying (26) is essentially unique, in the
sense that, whenever f1 , f2 L1K (X, A, ) satisfy (26), it follows that
f1 = f2 , -a.e.
(ii) If f L1K (X, A, ) is any function satisfying (26), then the variation
measure ||, of , satisfies
Z
||(A) =
|f | d, A A-fin .
A

Proof. Fix F to be a decomposition for (X, A, ).



A. For every F F, we apply Theorem 4.2 to the measure space (F, Af , F )

and the measure F , to find some measurable function fF : F [0, ], such that
Z

(A) =
fF d, A AF .
A

Using the patching

property, there exists a measurable function f : X [0, ],

such that f F = fF , F F. The key feature we ar going to prove is a particular
case of (25).
R
Claim 1: (A) = A f d, A Afin .

4. Radon-Nikodym Theorems

317

Fix A Afin . On the one hand, we know that

X
(A) =
(A F ).
F F

Since the sum is finite, it follows that the subcollection



F(A) = F F : (A F ) > 0
S
is at most countable. We then form the set A = F F(A [A F ], which is clearly a
subset of A. The difference D = A r A has again (D) < , so its measure is also
given as
X
(D) =
(D F ).
F F

Notice however that we have (D F ) = 0, F. (If F F(A), we already have

D F = , whereas if F F r F(A), we have D F A F , with (A F ) = 0.)
Using then the above equality, we get (D) = 0. By abosulte continuity we also
get (D) = 0. Using the equality A = A Dn , and -additivity (it is essential here
that F(A) is countable), it follows that
X
=
(A) = (A)
(A F ).
F F(A)

Using the hypothesis, we then get

(27)

X Z

(A) =

f d.

AF

F F(A)

Now if we list F(A) = {Fk }

k=1 , and if we take a partial sum, we have
Z
Z
Z
n
X
f d =
f d =
f Gn d,
k=1

AFk

Gn

where
Gn =

p
[

[A Fk ], n 1.

k=1

It is clear that we have

f G1 f G2 . . . ,
limn (f Gn )(x) = (f A )(x), x X,
so using the Monotone Convergence Theorem, it follows that
Z
Z
Z
lim
f Gn d =
f A d =
f d.
n

Using (27) we then get

Z
(A) = lim

Z
f Gn d =

f d.

= 0, it follows that
On the other hand, since (A r A)
Z
Z
f d =
f d,
A

so the preceding equality immediately gives the desired equality

Z
(A) =
f d.
A

318

CHAPTER IV: INTEGRATION THEORY

At this point let us remark that the local almost uniqueness of f already follows
from Remark 4.1.

Let us prove now the equality (25). Start

Swith some set A A-fin , and choose

a sequence (An )n=1 Afin , such that A = n=1 An . Define the sequence (Bn )
n=1
by
Bn = A1 An , n 1,
S
so that we still have Bn Afin , n 1, as well as A = n=1 Bn , but moreover we
have B1 B2 . . . . For each n 1, using Claim 1, we have the equality
Z
(Bn ) =
f d.
Bn

Using these equalities, combined with

0 f B1 f B 2 . . . ,
limn (f Bn )(x) = (f B )(x), x X,
the Monotone Convergence Theorem, combined with continuity yields
Z
Z
Z
Z
f d =
f B d = lim
f Bn d = lim
d = lim (Bn ) = (A).
B

Bn

B. We start off by choosing a measurable function h : X K, with |h| = 1,

such that
Z
(A) =
h d||, A A.
A

Using part A, there exists some measurable function g0 : X [0, ], such that
Z
(28)
||(A) =
g0 d, A A-fin .
A

At this point, g0 may not be integrable, but we have the freedom to perturb it (l.a.e.) to try to make it integrable. This is done as follows. Consider the collection


F0 = F F : ||(F ) > 0 .
Since S
|| is finite, it follows that F0 is at most countable. Define then the set
X0 = F F0 F A-fin . Since X0 is --finite, every set A A with A X0 , is
--finite, so we have
Z

||(A) =
g0 d, A AX .
A

Applying the -finite

version
of the Radon-Nikodym Theorem to the -finite mea

sure space (X0 , AX0 , X0 ) and the finite measure X0 , it follows that the density



g0 X0 belongs to L1+ (X0 , AX0 , X0 ), which means that the function g = g0 X0
belongs to L1+ (X, A, ). With this choice of g, let us prove now that the equality
(28) still holds, with g in place of g0 . Exactly as in the proof of part A, it suffices
to prove only the equality
Z
(29)
||(A) =
g d, A Afin .
A

Claim 2: ||(A) = ||(A X0 ), A A-fin .

4. Radon-Nikodym Theorems

319

Since (use the fact that || is finite) the equality is equivalent to

||(A r X0 ) = 0, A A-fin ,
it suffices to prove it only for A Afin . If A Afin , using the properties of the
decomposition F, we have
X
X
X
||(A) =
||(A F ) =
||(A F ) +
||(A F ) =
F F

= ||

F F0

||(A F ) =

F FrF0

F F0

= ||(A X0 ) +

F FrF0

[A F ] +
X

||(A F ).

F F 0 (A)

Notice now that, for F F r F0 , we have ||(F ) = 0, which gives ||(A F ) = 0,

so the Claim follows immediately from the above computation.
Having proven the above Claim, let us prove now (29). Fix A Afin . The
desired equality is now immediate from Claim 2, combined with (28):
Z
Z
||(A) = ||(A X0 ) =
g0 d =
g0 AX0 d =
AX0
X
Z
Z
Z
=
g0 X0 A d =
g A d =
g .
X

L1+ (X, A, ),

it follows that
Define now the function f0 = hg. Since |f0 | = g
f0 L1K (X, A, ). Let us prove that f0 satisfies the equality (26). Start with some
A A-fin . On the one hand, using Claim 2, we have
|(A r X0 )| ||(A r X0 ) = 0,
so we get (A) = (A
 the -finite version of the Radon-Nikodym
 X0 ). Using
Theorem for (X0 , AX0 , X0 ) and X0 , we then have
Z
Z
(A) = (A X0 ) =
hg0 d =
hg0 AX0 d =
AX0
X
Z
Z
Z
Z
=
hg0 X0 A d =
hg A d =
hg d =
f0 d.
X

We now prove the uniqueness property (i) of f (-a.e.!). Assume f L1K (X, A, )
is another function, such that
Z
(A) =
f d, A A-fin .
A

Claim 3: f = f0 , -l.a.e.
What we need to show here is the fact that
f B = f0 B , -a.e., B Afin .
But this follows immediately from the uniqueness
from part B of Theorem 4.2,


 ,  ) and the measure  , which has
applied to the finite
measure
space
(B,
A
B
B
B

both f B and f0 B as densities.
Using Claim 3, we now have f f0 L1K (X, A, ), with f f0 = 0, -l.a.e.,
so we can apply Proposition 4.4, which forces f f0 = 0, -a.e., so we indeed get
f = f0 , -a.e.

320

CHAPTER IV: INTEGRATION THEORY

Property (ii) is obvious, since by (i), any function f L1K (X, A, ), that satisfies
(26), automatically satisfies |f | = |f0 | = g, -a.e.

Comment. One should be aware of the (severe) limitations of Theorem 4.3,
notably the fact that the equalities (25) and (26) hold only for A A-fin . For
example, if one considers the measure space (X, P(X), ), with X uncountable,
and defined by

if A is uncountable
(A) =
0
if A is countable
This measure
space is decomposable, with a decomposition consisting of singletons:

F = {x} : x X . For a measure on P(X), the condition  means
precisely that (A) = 0 for all countable subsets A X. In this case the equality
(25) says practically nothing, since it is restricted solely to countable sets A X,
when both sides are zero.
In this example, it is also instructive to analyze the case when is finite (see part
B in Theorem 4.3). If we follow the proofS
of the Theorem, we see
 that at some point

we have constructed a certain set X0 = F F0 , where F0 = F F : (F ) > 0 .
In our situation however it turns out that X0 = . This example brings up a very
interesting question, which turns out to sit at the very foundation of set theory.
Question: Does there exists an uncountable set X, and a finite measure
on P(X), such that (X) > 0, but (A) = 0, for every countable subset
A X?
(The above vanishing condition is of course equivalent to the fact that ({x}) = 0,
x X.) It turns out that, not only that the answer of this question is unkown, but
in fact several mathematicians are seriously thinking of proposing it as an axiom
to be added to the current system of axioms used in set theory!
The limitations of Theorem 4.3 also force limitations in the Change of Variables
property (see Proposition 4.1), which in this case has the following statement.
Proposition 4.6 (Local Change of Variables). Let (X, A, ) be a measure
space, and let be a measure on A, and let f : X [0, ] be a measurable
function.
A. The following are equivalent:
(i) one has
Z
(A) =
A

f d, A A-fin ;

(ii) for every measurable function h : X [0, ], with the property that the
set Eh = {x X : h(x) 6= 0} belongs to A-fin , one has the equality
Z
Z
(30)
h d =
hf d.
X

B. If and f are as above, and K is either R or C, then the equality (30)

also holds for those measurable functions h : X K with Eh A-fin , for which
h L1K (X, A, ) and hf L1K (X, A, ).
Proof. A. (i) (ii). Assume (i) holds. Start with some measurable function
h : X [0, ], such that the set Eh = {x X : h(x) 6= 0} belongs to A-fin . The
equality
 (30) is then immediate fromProposition 4.1, appliedto the measure space
(Eh , AE , E ), and the measure E , which has density f E .
h

4. Radon-Nikodym Theorems

321

(ii) (i). Assume (ii) holds. If we start with some A A-fin , then obviously
the measurable function h = A will have Eh = A, so by (ii) we immediately get
Z
Z
Z
(A) =
A d =
A f d =
f d.
X

B. Assume now and f satisfy the equivalent conditions (i) and (ii). Suppose
h : X K is measurable, with Eh A-fin , such that h L1K (X, A, ) and
hf L1K (X, A, ). Then the equality
4.1,
 (30)
 follows again from Proposition

applied to the measure space (Eh , AE , E ), and the measure E , which has
h
h
h


density f E .
h