On Radon Measure
On Radon Measure
On Radon Measure
12/3/2016
1 Introduction
In this paper we define and discuss Radon measures, which are measures on certain topological spaces
which interact nicely with the underlying topology. Radon measures allow for approximation by open
sets or compact sets. They also allow for a unique representation of linear functionals as integrals so that
measure and integration can be handled with the tools of functional analysis [9]. We will also address
Radon spaces, on which every finite measure is Radon, and metrics on a space of Radon measures. In
particular, the Wasserstein distance defines a metric on Radon probability measures which has close ties
to transport optimization and other applied problems.
In general, this paper summarizes and follows the treatment of the subject in Folland’s Real Analysis [2],
and will reference other sources when used. Many propositions and theorems will be supplied without
proof, and the reader is referred to [2] for the details.
A topological space X is called locally compact if every point in X has a compact neighborhood. X
is called a Hausdorff space if every pair of distinct points in X have disjoint neighborhoods. A useful
consequence of this is that compact sets are closed in Hausdorff spaces [8]. For the rest of our discussion
of Radon measures, X will refer to a locally compact Hausdorff (LCH) space, and we will define measures
on the σ-algebra B, the Borel sets of X.
µ is called regular if it is both inner and outer regular on all Borel sets.
Essentially, these regularity conditions allow the measure of a set to be approximated by the measures of
open sets from the outside or compact (and hence closed) sets from the inside. Using these definitions,
we are ready to define Radon measure on an LCH space X.
Definition 2.2 (Radon measure). A Radon measure is a Borel measure that is finite on compact sets,
outer regular on all Borel sets, and inner regular on open sets.
Note that some authors define a Radon measure µ on the Borel σ-algebra of any Hausdorff space to be
any Borel measure that is inner regular on open sets and locally finite, meaning that for every point
x ∈ X there is an open neighborhood of x with finite measure [9, 7]. On an LCH space these definitions
1
are equivalent and some of the following results, including our formulation of the Riesz Representation
theorem, are dependent on the locally compact condition, so we use the LCH space definition given here.
Radon measures on LCH spaces have several useful properties. Recall that a measure µ is σ-finite if the
space X can be written as a countable union of finite-measure sets. Similarly, for the corollary below,
we say a set or a space X is σ-compact if it is a countable union of compact sets. These results state
that when working with σ-finite sets or measures we gain regularity.
Proposition 2.1. A Radon measure is inner regular on all of its σ-finite sets
Corollary 2.1.1.
The proof of 2.1 involves using the inner and outer regularity properties of the Radon measure to ap-
proximate the measure of σ-finite sets of finite measure, and for sets of infinite measure to apply that
technique to a sequence of sets, each of which has finite measure. From this, 2.1.1 follows directly from
the definitions of σ-finite and σ-compact.
Proposition 2.2. Let (X, B, µ) be a σ-finite Radon measure space, and let E ∈ B. Then
Proof. (Sketch) Consider a partition of E into disjoint subsets of finite measure, and approximate each
subset with an open set to within a factor of ε. Let U be the countable union of each of those open
sets such that µ(U \ E) < ε/2. Similarly construct an open set V which is the countable union of open
approximations of a partition of E c with µ(V \ E c ) < ε/2. Then let F = V c , and find that
2.) follows from 1.) in a straightforward way, from taking the limit over the countable unions so that in
the difference µ(B \ A) → 0.
Finally, on a different note, we present a useful approximation result for Lp spaces under Radon measures.
This last result is shown by first restricting our attention to sets E ∈ B of finite measure (since if
µ(E) = ∞, then χE 6∈ Lp (µ)). We then show that in the Lp norm we can approximate the characteristic
function χE with functions in Cc (X). This is done by applying 2.2 to obtain a compact K ⊂ E and an
open U ⊃ E such that µ(U \ K) < ε. Using Urysohn’s Lemma (3.1 below) we can find an f ∈ Cc (X)
such that χK ≤ f ≤ χU , employ our ε-bound on the difference, and thus approximate any Lp simple
function. These are in turn dense in Lp and so we have the proposition.
In order to prove a powerful representation result relating linear functionals and Radon measures, we
will first need to establish an important topological lemma and some definitions.
Theorem 3.1 (Urysohn’s Lemma). If X is a locally compact Hausdorff space and K ⊂ U ⊂ X, where
K is compact and U is open, there exists a continuous f : X → [0, 1] such that f = 1 on K and f = 0
outside a compact subset of U .
2
Urysohn’s lemma is very helpful in a situation which crops up often in these regularity proofs; we have
a compact K contained in an open U , and we are trying to approximate a characteristic function with
a continuous function. Urysohn’s lemma says that such a continuous function exists, approximating the
characteristic functions of K and U up to U \ K.
When considering any real or complex-valued function f on the domain of X, we will define the support
of f , denoted supp(f ), as the closure of the set {x ∈ X : f (x) 6= 0}. Going forward, we shall be concerned
chiefly with those functions which have compact support, meaning that supp(f ) is a compact set. We
then define Cc (X) as the space of all continuous functions f : X → C with compact support. Note that
since Radon measures are finite on compact sets, the support of any f ∈ Cc (X) has finite measure under
any Radon measure and the image of each f ∈ Cc (X) is bounded.
Let I be a real-valued linear functional on Cc (X), that is, for f ∈ Cc (X), I : f (x) → R. Then we say I is
positive if I(f ) ≥ 0 whenever f ≥ 0. In the following proposition, we formally state that positive linear
functionals are bounded on compact sets. In it, we use the uniform norm of f , which is equivalent on
Cc (X) to the infinity norm of f , defined as kf ku := sup{|f (x)| : x ∈ X}.
Proposition 3.2. If I is a positive linear functional on Cc (X), for each compact set K ⊂ X there exists
a constant CK such that |I(f )| ≤ CK kf ku for all f ∈ Cc (X) such that supp(f ) ⊂ K.
Now let µ be a Borel measure R on X, let B be the Borel σ-algebra R on X, and let E ∈ B. When µ is
finite on compact Rsets, then f dµ is finite for all f ∈ Cc (X) and f dµ ≥ 0 whenever f ≥ 0. Hence, the
mapping I : f 7→ f is a positive linear functional. In particular, since Radon measures are Borel and
finite on compact sets, the integral with respect to any Radon measure is a positive linear functional on
Cc (X). The Riesz Representation Theorem asserts that this is in fact the only kind of positive linear
functional on Cc (X): any positive linear functional can be expressed as the integral with respect to a
Radon measure.
In order to provide some intuition for how we might find such a measure µ relatingR a positive linear
functional I to the integral with respect to µ, observe that for any measure µ, µ(E) = χE dµ. Since we
R match, we might consider setting µ(E) = I(χE ) for every E ∈ B, in which case
want I and the integral to
we would have I(χE ) = χE dµ for every E ∈ B, which seems like a good start. However, characteristic
functions are neither continuous nor have compact support in general, and since I is a linear functional
on Cc (X), we cannot do the above exactly.
Instead, given an open set U , we will consider the supremum of I(f ) for 0 ≤ f ≤ χU . We will also need
to require that supp(f ) ⊂ U . We will write f ≺ U in the case that 0 ≤ f ≤ χU and supp(f ) ⊂ U . This
brings us to the full statement of the Riesz Representation Theorem.
Theorem 3.3 (The Riesz Representation Theorem). RIf I is a positive linear functional on Cc (X), there
is a unique Radon measure µ on X such that I(f ) = f dµ for all f ∈ Cc (X). Moreover, µ satisfies
µ(U ) = sup{I(f ) : f ∈ Cc (X) and f ≺ U } for all open U ⊂ X, (1)
µ(K) = inf{I(f ) : f ∈ Cc (X) and f ≥ χK } for all compact K ⊆ X. (2)
Proof. (Sketch) Proving the uniqueness of the measure µ given by this theorem is relatively straight-
forward. Using Urysohn’s lemma and the inner regularity of µ on open sets, we first demonstrate that
3.3(1) is satisfied. It is then apparent that µ is determined by I on open sets, and because of outer
regularity, is determined by I on Borel sets, and thus is unique.
Proving existence is much more involved, but still straightforward. Briefly, (i) an outer measure is
constructed on X, (ii) open sets are shown to be outer-measurable and a Borel measure is constructed
from the outer measure, which (iii) is shown to satisfy 3.3(2) above. It is demonstrated R that µ is
inner regular and hence Radon using that property. Finally, (iv) it is shown that I(f ) = f dµ for all
f ∈ Cc (X).
For (i), we define µ(U ) = sup{I(f ) : f ∈ Cc (X) and f ≺ U } for open U ⊂ X. Then, for arbitrary
E ⊂ X, we let µ∗ (E) = inf{µ(U ) : E ⊂ U, U open}. In order to show that µ∗ is an outer measure, it
3
S P
suffices to show that if (Ui ) is a sequence of open setsSwith Ui = U , then µ(U ) ≤ µ(Ui ), from which
it follows that µ∗ (E) = inf { µ(Ui ) : Ui open , E ⊂ Ui }, which defines an outer measure.
P
The proof of (ii) teases out the definitions of µ on open sets and µ∗ on arbitrary sets to show that open
sets are µ∗ -measurable. We prove (iii) from (ii), since it follows from Carathéodory’s extension theorem
that all Borel sets are µ∗ -measurable and that µ = µ∗ |BX is a measure. This measure µ satisfies 3.3(1)
and is outer regular by definition. From (iii), it follows that µ is inner regular.
R
To prove (iv), it suffices to show that I(f ) = f dµ for all f ∈ Cc (X, [0, 1]), since Cc (X) is the linear
span of Cc (X, [0, 1]). This is done by constructing a sequence (fi ) of N functions which sum to f , and
using the characteristic function of the support of each fi , 3.3(2), and outer regularity toRdemonstrate
that |I(f ) − f dµ| ≤ N1 µ(supp(f )). Since µ(supp(f )) is finite and N is arbitrary, I(f ) = f dµ.
R
4 Radon space
We will now consider the class of spaces where every finite measure is Radon [6].
Definition 4.1. A Hausdorff space X is called a Radon space if every finite Borel measure on X is a
Radon measure.
In Folland, it is shown that every second countable space is a Radon space [2]. In fact, the theorem given
is slightly stronger:
Theorem 4.1. Let X be a locally compact Hausdorff space in which every open set is σ-compact (which
is the case, for example, if X is second countable). Then every Borel measure on X which is finite on
compact sets is regular and hence Radon.
Proof. (Sketch) Consider a Borel measure µR which is finite on compact sets. Then each f ∈ Cc (X) has
finite integral with respect to µ, so I(f ) = f dµ is a positive linear functional on Cc (X). We can then
apply the Riesz Representation Theorem 3.3 to obtain a Radon measure ν associated with I. First we
show that µ(U S) = ν(U ) for all open U . Using the σ-compactness of U , choose a sequence (Kn ) of compact
K such that Ki = U . Using Urysohn’s Lemma 3.1, we can construct aR sequence (fnR) of functions with
compact support, increasing pointwise to χU . Then we get µ(U ) = lim fn dµ = lim fn dν = ν(U ), the
last inequality following from the monotone convergence theorem.
Next we show regularity. Let E be a Borel set, ε > 0. Then by 2.2, there are closed F and open V such
that F ⊂ E ⊂ V , with ν(V \ F ) < ε. Since V \ F is open, we get that µ(V \ F ) < ε. By subadditivity
and monotonicity of measures, this gives us µ(V ) ≤ µ(E) + ε, so µ is outer regular. Similarly, we have
µ(F ) ≥ µ(E) − ε. Since F is closed andSE σ-compact, F is σ-compact, so we can choose a increasing
sequence (Ki ) of compact Ki ⊂ F with Ki = F , in which case µ(Ki ) converges to µ(F ), so µ is inner
regular.
Thus every finite measure on a second countable space is Radon. Since every separable metric space is
second countable, every separable metric space is a Radon space.
We say that a function f vanishes at infinity if for any ε > 0, {x ∈ X : |f (x) ≥ ε} is compact.
We will now turn our attention towards C0 (X), the spaces of continuous functions which vanish at infin-
ity. In any locally compact Hausdorff space, C0 (X) is the uniform closure of Cc (X). In this section, we
will give a complete description of C0 (X)∗ , the bounded linear functionals on C0 (X). The first important
fact is that real linear functionals on C0 (X) have a “Jordan decomposition.”
4
Theorem 5.1. If I ∈ C0 (X, R)∗ , there is are positive linear functionals I + , I − ∈ C0 (X, R)∗ such that
I = I + − I −.
A signed Radon measure is a signed Borel measure whose positive and negative variations are Radon
measures. We will denote the space of signed Radon measures on X by M (X) and for µ ∈ M (X) define
kµk = |µ|(X). It can be shown that M (X) is a complete normed vector space.
Proof. (Sketch) To see that there is a bijection from M (X) to C0 (X)∗ , we first note that every bounded
linear functional on Cc (X) extends continuously to C0 (X), since C0 (X) is the uniform closure of Cc (X).
SoR each linear
R functional is given by integration against some Radon measure. On the other hand,
| f dµ| ≤ |f |d|µ| ≤ kf kµ kµk, so every integral against a Radon measure is a bounded linear functional.
In order to show that the norms are equal, first note that the above inequality gives that kIµ k ≤ kµk.
dµ
To show the other direction, we approximate the Radon-Nikodym derivative d|µ| by some f ∈ Cc (X)
R
such that kµk ≤ | f dµ| + ε, from which it follows that kµk ≤ kIµ k.
The Riesz Representation Theorem gives a complete characterization of the space of bounded linear
functionals on C0 (X) - it is exactly the space of signed Radon measures.
Let M+ (X) be the space of finite positive Radon measures. M+ (X) ⊂ M (X) is clearly closed under
addition and scalar multiplication for scalars c ≥ 0 and contains µ ≡ 0, and thus is a pointed convex
cone. We now define a distance on this space.
Definition 6.1 (Radon distance). The Radon distance between measures µ, ν ∈ M+ (X) is
Z
ρ(µ, ν) := sup f (x)d(µ − ν) : continuous real-valued f : X → [−1, 1] (3)
X
Since all elements of M (X) are finite measures, we are guaranteed this supremum exists in R. It can be
shown that ρ is a metric on M+ (X) and that M+ (X) is a complete metric space under ρ [9].
The Radon metric may seem to be a natural choice for applications on M+ (X), but in fact it is often
too restrictive. For example, let us consider the oft-used space of probability measures in M+ (X),
defined:
P(X) := {µ ∈ M+ (X) : µ(X) = 1}.
Under the Radon metric, P(X) is not sequentially compact, meaning there is no guarantee that every
sequence has a convergent subsequence [9]. For this and other reasons, other more useful metrics have
been defined on P(X), chief among them, the Wasserstein metric.
First, we need to define Π(µ, ν) as the set of all couplings in P(X × X): pairings of measures on (X, B)
which preserve the marginal distributions of µ and ν respectively on X. That is,
π(A × X) = µ(A)
Π(µ, ν) := π ∈ P(X × X) : , for each A ∈ B
π(X × A) = ν(A)
For a Radon space X imbued with a metric d, let us define Pp (X) as the collection of probability mea-
R moment, for p ∈ [1, ∞). This means that for any µ ∈ Pp (X), around any fixed
sures with a finite p-th
central point x0 ∈ X, X d(x, x0 )p dµ(x) < ∞.
5
Definition 6.2 (p-th Wasserstein distance). For µ, ν ∈ Pp (X) the p-th Wasserstein distance between
µ and ν is defined:
Z 1/p
Wp (µ, ν) := inf d(x, y)p dπ(x, y) (4)
π∈Π(µ,ν) X×X
or equivalently:
inf kdkLp (π)
π∈Π(µ,ν)
It can be shown that if X is a compact metric space, i.e. if X is complete and totally bounded under
metric d, then Wp is a metric on Pp (X), sometimes called the Kantorovich metric [1, 9, 5].
The support of a measure µ is the largest closed subset of X for which every open neighborhood
of any point in the set has strictly positive measure. When µ and ν have bounded support, which is
certainly the case for X compact, the W1 metric has a dual formulation which is closely related to the
Radon metric [1, 4]
The formulations (5) and (3) of the Wasserstein-1 and Radon metrics are quite similar. In fact, it is
easily shown (see [3] for proof) that, if C is the constant bound for d on X compact, then
From this, we infer that convergence in the Radon metric implies convergence in the Wasserstein-1 metric
[10], although the converse is not necessarily true.
In the field of tranportation, the Wasserstein metric is closely tied to both the Monge and Kantorovich
transport problems, which are concerned with optimally moving mass between two probability distri-
butions on the same space. In the non-linear Monge problem, we seek a map between two probability
spaces which will minimize the integral of a cost function on the domain space. By contrast, the linear
Kantorovich problem seeks a probability coupling π which minimizes the integral of a cost function on the
product space. This eventually lead to the Wasserstein metric (thus sometimes called the Kantorovich-
Rubinstein metric) and the dual formulation (6.1) [1]. W1 is also referred to as the “earth mover’s
distance”, at least in part because Monge’s original problem was constructed in the context of optimally
transporting construction soil, where the probability distributions represented the conserved volume of
earth, and the cost function represented distance moved.
Many other applications of the Wasserstein distance exist, including but certainly not limited to: [4]
References
[1] Vladimir I. Bogachev and Aleksandr V. Kolesnikov. The Monge-Kantorovich problem: achieve-
ments, connections, and perspectives. Russian Mathematical Surveys, 67(5):785–890, November
2012.
[2] Gerald B. Folland. Real Analysis: Modern Techniques and Their Applications, chapter 7, pages
211–233. John Wiley & Sons, 2nd edition, May 1999.
6
[3] Alison L. Gibbs and Francis Edward Su. On choosing and bounding probability metrics. Interna-
tional Statistical Review, 70(3):419–435, 2002.
[4] Kazumasa Kuwada. Probability analysis on Wasserstein distance. In Probabilistic symposium - Re-
search meeting, December 19-22, 2011. www.math.tohoku.ac.jp/~aida/workshop/H23-Daisympo/
kuwada.pdf.
[5] Filippo Santambrogio. The Wasserstein Distances. Course notes, URL: https://www.math.u-psud.
fr/~filippo/Wp.pdf, March 20, 2011.
[6] Laurent Schwartz. Radon Measures on Arbitrary Topological Spaces and Cylindrical Measures. Ox-
ford University Press, 1973.