Chapter 9

Theory of Utility Representation

This chapter centers around the problem of order-embedding a given preordered set in
(R. _) or in (R
n
. _) (or in some other naturally ordered function space). In the lexicon
of individual decision theory, this problem is known as the ordinal utility representation
problem. As the bulk of work in this area is in fact done within decision theory, we
shall adopt this terminology here as well.
There are mainly two approaches toward the utility representation problem. The rst
one is order-theoretic, and aims at identifying those complete (or incomplete) preorders
that can be order-embeded in R (or R
n
) by means of purely order-theoretic conditions.
The second is topological. In that approach the ground set of the preorder is given
some topological structure, and the involved preorders are assumed to satisfy suitable
continuity conditions. In return, we ask to nd continuous (or at least semicontinuous)
order embeddings (utility functions). We shall outline the elements of both of these
approaches below.
1 Preliminaries
1.1 Order-Preserving Real Maps
The bulk of this chapter is concerned with those order-preserving functions that map
a preordered set into the loset (R. _). (As usual, we denote the latter loset simply as
R below.) To conform with the related literature, we shall rename such functions as
follows:
Denition. Let (A. %) be a preordered set. We say that a real function , on A is
%-increasing if
r % implies ,(r) _ ,()
for every r. A. This function is said to be strictly %-increasing, if it is %-increasing
and
r ~ implies ,(r) ,()
for every r. A.
1
Given a preordered set (A. %), and a function , : A R. the notion of being
%-increasing is none other than being order-preserving. However, the notion of being
strictly %-increasing is new. In particular, it is weaker than being an order-embedding.
While it is readily checked that if , is an order-embedding, then it must be strictly
%-increasing, the converse is false. For instance, the real map x r
1
+ r
2
on R
2
is
strictly _-increasing, but it is far from being an order-embedding. Thus, in general, we
have:
order-embedding == strictly %-increasing map == %-increasing map
for any real map dened on a preordered set (A. %). Nonetheless, it is plain that the
discrepancy between the former two monotonicity notions disappear when % is a linear
order. That is:
order-embedding == strictly <-increasing map == <-increasing map
for any real map dened on a loset (A. <).
1.2 Losets and Q
Order-Embedding Countable Losets in Q
The following is a simple, but very useful, result of order theory. It was proved in 1895 by
Georg Cantor during the course of his work on the development of transnite numbers.
In words, it says that every countable loset can be order-embedded in the set of rational
numbers between 0 and 1 (and hence in Q).
Proposition 1.2.1. Let (A. <) be a loset. If A is countable, then there exists a strictly
<-increasing real function from A into Q (0. 1).
Proof. The claim is trivial when A is nite, so we assume that A is countably innite.
Let us enumerate A and Q (0. 1) as
A = r
1
. r
2
. ... and Q (0. 1) = :
1
. :
2
. ....
We construct the function , from A into Q (0. 1) as follows. First, let ,(r
1
) := :
1
.
Second, if r
1
< r
2
(or r
2
< r
1
), set ,(r
2
) as the rst enumerated element in :
2
. ...
such that :
1
_ ,(r
2
) (,(r
2
) _ :
1
. resp.). Proceeding inductively, for any : = 2. 3. ....
set
,(r
m
) := the rst element of :
1
. ...,(r
1
). .... ,(r
m1
)
which has the same order relation (w.r.t. _) to the numbers ,(r
1
). .... ,(r
m1
)
as r
m
has to the elements r
1
. .... r
m1
(w.r.t. <).
Since there is a rational number strictly between any two distinct rational numbers, this
well-denes , : A Q (0. 1). Moreover, it follows readily from our construction that,
for every r and in A. we have ,(r) _ ,() i r < .
2
Order-Dense Losets
Let us now pose the following question: Which losets are order-isomorphic to Q? Cer-
tainly, not all losets are up to the task. First of all, it is obvious that a nite loset cannot
be order-isomorphic to Q (as there cannot be a bijection between these two sets). A
countably innite loset does not do the job necessarily either. For instance, the losets
(N. _) and (Q. _) are not order-isomorphic, as the rst one has a minimum element and
the latter does not. Neither would the absence of extremum elements solve the problem.
Indeed, (Z. _) and (Q. _) are not order-isomorphic. After all, (Q. _) has the property
that for any distinct elements j and in Q. there is an element : in Q with j :.
Obviously, (Z. _) does not possess this order-theoretic property, and hence it cannot be
order-isomorphic to (Q. _). Put succinctly, (Q. _) is order-dense while (Z. _) is not.
Denition. Let (A. %) be a preordered set and 2 _ A. If, for any r and in A such that
r ~ . there exists an element . of 2 such that r ~ . ~ . we say that 2 is %-dense in
A. If 2 = A here, we simply say that A is %-dense.
Note. Order-denseness is an order-theoretic property in the sense that a preordered set (A, %)
satises it i every preordered set that is order-isomorphic to (A, %) also satises it.
It turns out that the absence of extremum elements and order-denseness jointly
characterize those countable losets that are order-isomorphic to (Q. _). This is, again,
a result of Georg Cantor.
Proposition 1.2.2. Let (A. <) be a loset with no <-maximum and no <-minimum
elements. If A is countable and <-dense, then (A. <) is order-isomorphic to (Q. _).
This result is often paraphrased as saying that any two countable order-dense losets
with no endpoints are isomorphic. As the proof of Proposition 1.2.2 parallels that of
Proposition 1.2.1, we leave it as an exercise here.
Exercises
1.2.1. Prove Proposition 1.2.2.
1.3 The Utility Representation Problem
First Impressions
In Section 3.2 of Chapter 1, we have mentioned that a preference relation on a set A
of choice alternatives is dened as a preorder on A in economics, and it is assumed to
contain all the information that concerns how an agent compares any two alternatives
according to her tastes.
1
Unfortunately, this preorder is often not the most convenient
1
There are more sophisticated notions of preference relations that stem from the theory of boundedly
rational choice. We shall not deal with such preferences in this text, however.
3
way of summarizing this information. Indeed, maximizing a binary relation (while a
well-dened matter) is a much less friendlier exercise than maximizing a real function.
Thus, it would be quite useful if we knew how and when one can nd a real function
that attaches to an alternative r a (strictly) higher value than an alternative i r is
ranked (strictly) above by a given preference relation. In economics, such a function
is called the utility function of the individual who possesses this preference relation. A
fundamental question in the theory of individual choice is therefore the following: What
sort of preference relations can be described by means of a utility function?
We begin by formally dening what it means to describe a preference relation by
means of a utility function.
Denition. Let % be a preference relation on a nonempty set A. We say that a function
n : A R represents % provided that
r % if and only if n(r) _ n()
for every r and in A. If such a function exists, we say that % admits a utility
representation, and refer to n as a utility function for %.
Note. In order-theoretic terms, the phrase n is a utility function for the preference relation %,
means simply that n is an order-embedding from the preordered set (A, %) into the loset (R, _).
Consequently, having a utility function is an order-theoretic property (that is, it is preserved
under any order-isomorphism).
In this chapter, we shall mostly be concerned with identifying those preference re-
lations that admit a utility representation. Let us then begin by noting that a utility
function for a preference relation is by no means unique. Rather, it is unique up to
strictly increasing transformations. Indeed, if % is a preference relation on a nonempty
set A. and n : A R is a utility function for %, then, obviously, , n is also a utility
function for %, where , is any strictly increasing real function on n(A). Less obvious is
the fact that the converse of this is also true:
Proposition 1.3.1. Let (A. %) be a preordered set, and n and two real functions on
A. Then, n and are both utility functions for % if, and only if, there exists a strictly
increasing real function , on n(A) such that = , n.
We leave the proof of this result as an exercise.
Remark. The if part of Proposition 1.3.1 shows that we can always represent a
preference relation % by a bounded utility function, provided that this preference relation
admits a utility representation. Indeed, if n is a utility function for %. then so is the
map arctan n. because r arctan r is a strictly increasing map from R onto (

2
.

2
).
In fact, we can control the length of the range of an ordinal utility function anyway we
wish. After all, for any real numbers with c < /. the map
_
ba

_
arctan n +
a+b
2
4
is a utility function for % whose range is contained in the interval (c. /).
Utility Representation on Countable Sets
Let us now turn to the problem of existence of a utility representation. In fact, thanks
to Proposition 1.2.1, we already have in our disposal an important result in this regard.
The following is deduced easily from Proposition 1.2.1 by the method of passing to the
quotient.
Proposition 1.3.2. There is a utility function n : A (0. 1) for any complete preference
relation % on a nonempty countable set A.
Proof. Clearly, the symmetric part ~ of % is an equivalence relation, and the quotient
set A,

is a partition of A (Proposition 2.1 of Chapter 1). We now dene the linear

order < on A,

as [r]

< []

i r % . By Proposition 1.2.1, there exists a strictly

<-increasing function , from A,

into (0. 1). But then n : A (0. 1). dened by

n(r) := ,([r]

). is a utility function for %.

Examples of Preferences without a Utility Representation
Unfortunately, the countability requirement cannot be dismissed in Proposition 1.3.2.
We now present some illustrations of this fact.
2
Example 1.3.1. Consider the lexicographic order <
lex
on R
2
we have introduced in
Example 3.1.4 of Chapter 1. This linear order cannot be represented by a utility function.
For, suppose n : R
2
R represents <
lex
. Then, for any real number c, we have n(c. c +
1) n(c. c) so that 1(c) := (n(c. c). n(c. c + 1)) is a nonempty interval. Moreover,
1(c) 1(/) = O for any distinct numbers c and /. because
n(/. /) n(c. c + 1) whenever / c.
and
n(/. / + 1) < n(c. c) whenever / < c.
Therefore, the map c 1(c) is an injection from R into 1(c) : c R. But since
1(c) : c R is countable, this entails that R is countable, a contradiction.
Example 1.3.2. Let < be a well-ordering on R recall Section 6.2 of Chapter 5. For
any real number r. dene
o := r R :
#
r is uncountable.
Then, set A := R if o = O. and A :=
#
(the <-minimum element of o) if o is nonempty.
We wish to show that the restriction of < to A. which we shall denote again by < for
convenience, does not admit a utility representation.
2
See Beardon et al. (2002) for a systematic analysis of complete preferences that do not admit a
utility representation.
5
Let us rst make some observations on A. Obviously, this set is uncountable. On
the other hand, by denition of A.
#
r is countable for every r in A. Furthermore, as
< is transitive, we have
#
r _ A for every r A. Conversely, if is a real number that
does not belong to

#
r : r A. then ~ A. that is, A _
#
. which implies
that
#
is uncountable, thereby showing that cannot belong to A. Taking stock: A is
an uncountable set such that
#
r : r A is countable for each r A. and
A =
_

#
r : r A.
Now, to derive a contradiction, let us assume that there is a utility function n : A
R that represents <. Without loss of generality, we may assume that n is bounded. Let
c := sup n(R). Then, for every rational number : with c :. there exists a real number
r such that c n(r) :. and hence
. R : : n(.) _ . R : n(r) n(.) =
#
r.
(The last equality follows from the hypothesis that n represents <.) Thus, . R : :
n(.) is a countable set for every rational number : with c :. And yet, for every real
number r.
#
r = . R : n(r) n(.) _
_
>r2Q
. R : : n(.).
(because Q is _-dense in R). and hence
A =
_

#
r : r A _
_
>r2Q
. R : : n(.).
It follows that A is contained in the union of a countable collection of countable sets,
and hence it is countable, a contradiction.
Just in case these examples may strike you as too abstract, we give two further
illustrations of preference relations that cannot be represented by a utility function,
which arise in economic analysis. The following example is due to Basu and Mitra (2003),
and it concerns the representation of a (social) preference relation over intergenerational
income streams in a way that is both egalitarian and ecient.
Example 1.3.3. Let A := 0. 1
1
. the set of all sequences whose terms are either 0 or
1. Let % be a complete preference relation on A such that
(r
m
) ~ (
m
) for any distinct (r
m
). (
m
) A with r
m
_
m
for each :; and
(r
m
) ~ (
m
) for any (r
m
). (
m
) A such that there exist two positive integers i
and , with r
i
=
j
. r
j
=
i
and r
k
=
k
for every positive integer / other than i
and ,.
6
(Think of any one member of A as an intergenerational wealth stream of a society (in
which there are two possible levels of income, low and high), and % as a social welfare
ordering. Then, this preference relation favors eciency in the sense that it ranks a
stream above another if every generation receives at least as much wealth in the former
stream as they do in the latter, with at least one generation receiving strictly more. It
is also equitable in the sense that it treats any two distinct generations equally in terms
of social welfare.) It is not dicult to show that such a complete preference relation %
exists by using Szpilrajns Theorem. We wish to show that no such preference relation
admits a utility function.
Let :
1
. :
2
. ... be an enumeration of Q. For each real number c and positive integer
:. we dene

m
(c) :=
_
1. if c :
m
0. otherwise,
and
r
m
(c) :=
_
1. if : = mini :
i
(c) = 0

m
(c). otherwise.
Notice that, for any real number c and positive integer :. the rst (eciency) property
of % implies (r
m
(c)) ~ (
m
(c)). Thus, if there were a utility function n : 0. 1
1
R
that represents %. the interval
1(c) := (n(r
1
(c). r
2
(c). ...). n(
1
(c).
2
(c). ...))
would be nonempty for each c R. On the other hand, for any real numbers c and /
with / c. the set i :
i
(c) = 1 is a subset of i :
i
(/) = 1. and there are innitely
many integers i that belong to the latter set but not to the former. By the two properties
of %, therefore, we must have (
m
(/)) ~ (r
m
(c)) whenever / c. (Why?) It follows
that 1(c) 1(/) = O whenever c and / are distinct. As in Example 1.3.1, we nd that
the map c 1(c) is an injection from R into 1(c) : c R. which contradicts the
uncountability of R.
The following example is due to Dubra and Echenique (2003), and concerns the
representation of preferences over information structures.
Example 1.3.4. Consider the partition lattice (Par([0. 1]). <). where / < E i for every
in / there is a 1 in E with _ 1. (In the context of information theory, any
one member of Par([0. 1]) is thought of as modeling the information that an agent will
have about the values of a given [0. 1]-valued random variable, when the value of this
variable is realized.) By Szpilrajns Theorem, there exists a linear order % on Par([0. 1])
such that 1 implies 1 ~ . (This preference relation ranks ner partitions (more
information) higher than coarser ones (less information).) You are asked to prove below
that this linear order cannot be represented by a utility function.
Exercises
1.3.1. Prove Proposition 1.3.1.
7
1.3.2. Let A be a nonempty set and % a complete preference relation on A that admits a utility
representation. Let o be a nonempty subset of A, and suppose that n : o R is a utility function for
% (o o). Does there necessarily exist a utility function l : A R for % such that l[
S
= n?
1.3.3. Let A be a nonempty set and % a complete preference relation on A such that A,

is countable.
Show that % admits a utility representation.
1.3.4. Prove the assertion made in Example 1.3.4.
2 Representation through Order-Separability
The Use of Order-Dense Sets for Utility Representation
While its countability requirement cannot be relaxed entirely, we can replace this con-
dition by various other properties in Proposition 1.3.2. In particular, the existence of a
countable order-dense subset is all one needs for guaranteeing the existence of a utility
representation. This is a fairly practical observation, due to Georg Cantor, but we shall
provide only a sketch of proof for it here. A more general result will be established
shortly.
Proposition 2.1. Let A be a nonempty set, and % a complete preference relation on A.
If A contains a countable %-dense set, then % can be represented by a utility function.
Proof. If ~ = O. then it is enough to take n as any constant function, so we may assume
~ ,= O to concentrate on the nontrivial case. Assume that there is a countable %-dense
set 2 in A. By Proposition 1.3.2, there exists a function : 2 (0. 1) that represents
% (2 2). Now, we dene n : A [0. 1] by setting n(r) = 0 if r is a %-minimum in
A. by setting n(r) = 1 if r is a %-maximum in A. and nally, by letting
n(r) := sup(.) : r ~ . 2.
otherwise. As ~ ,= O. and 2 is %-dense in A, this map is well-dened. The rest of the
proof consists of verifying that n actually represents % on A. We leave this part as an
exercise.
A Characterization of Representability by Utility Functions
It is important to note that Proposition 2.1 does not generalize Proposition 1.3.2. (For
instance, the latter result applies to (N. _). while N does not contain a _-dense set.)
However, it is possible to weaken the order-separability requirement we used in Propo-
sition 2.1 to a condition that is both necessary and sucient for the representability of
%. The following is one way of doing this.
Denition. For any preordered set (A. %), we say that % is separable if there is a
countable subset 2 of A such that for every r and in A with r ~ . we have
r % . % for some . 2. (1)
8
Note. This separability notion was introduced by Debreu (1954) who credits Leonid Savage for
suggesting it. The if part of the characterization theorem below was also proved in that paper.
If % is a preference relation on a nonempty set A such that there is a countable %-
dense set in A. then % is separable. Thus, the if part of the result below generalizes
Proposition 2.1. Furthermore, this result shows that the notions of separability and
representability by a utility function are one and the same for complete preference
relations.
Theorem 2.2. Let A be a nonempty set, and % a complete preference relation on A.
Then, % admits a utility representation if, and only if, it is separable.
Proof. Assume that % is separable, so there is a countable subset 2 of A such that (1)
holds for every r and in A with r ~ . Let us enumerate 2 as .
1
. .
2
. .... and dene
`(r) := i N : r ~ .
i
and `(r) := i N : .
i
~ r
for any r A. By transitivity of %, if r % , then `(r) _ `() and `(r) _ `().
Moreover, by separability of A. if r ~ . at least one of these containments must be
strict. Consequently, the map n : A [1. 1], dened by
n(r) :=

i2M(x)
2
i

i2N(x)
2
i
.
is strictly %-increasing. Conclusion: % admits a utility representation.
Conversely, assume that % is a linear order, and that there is a utility function
n : A R for %. Let J stand for the collection of all open and bounded intervals with
rational endpoints that intersect n(A). That is,
J := (:
1
. :
2
) : :
1
. :
2
Q and :
2
n(.) :
1
for some . A.
For each 1 J. we use the Axiom of Choice to pick an element of A whose utility value
belongs to 1. This denes a map 1 : J A such that
:
2
n(1(:
1
. :
2
)) :
1
for every (:
1
. :
2
) J.
(We write 1(:
1
. :
2
) for 1((:
1
. :
2
)) here.) Clearly, J. and hence 1(J). is countable. Next,
we let 1 stand for the collection of all (r. ) AA such that r ~ and r ~ . ~ for
no . A. As % is (assumed) antisymmetric, the map (r. ) (n(). n(r)) is injective
on 1. which shows that 1 is countable. Finally, we dene
2 := 1(J) ' r A : either (.. r) 1 or (r. .) 1 for some . A.
As 1(J) and 1 are countable, so is 2. To complete the argument, take any r and in
A such that r ~ . If (r. ) belongs to 1. then r 2. and, obviously, setting . := r
yields r % . ~ . If (r. ) does not belong to 1. then r ~ . ~ for some . A (by
denition of 1), and hence, we can nd rational numbers :
1
and :
2
such that
n(r) :
2
n(.) :
1
n().
9
Then, (:
1
. :
2
) J. so setting . := 1(:
1
. :
2
) yields r ~ . ~ . Conclusion: % is separable.
We have now proved the only if part of Theorem 2.2 under the assumption that
% is a linear order on A. The general case is settled by the method of passing to the
quotient (as in Proposition 1.3.2).
Exercises
2.1. Use Theorem 2.2 to prove that the lexicographic order on R
2
does not admit a utility representation.
2.2. Use Theorem 2.2 to prove the assertion made in Example 1.3.4.
2.3. (Birkho) Let A be a nonempty set, and % a complete preference relation on A. Prove: % admits
a utility representation i there exists a countable subset 7 of A such that for every r and j in A7
with r ~ j, we have r ~ . ~ j for some . 7.
Hint. Adapt the argument we gave for Proposition 2.1 for the if part. For the only if part,
reason as in the corresponding part of Theorem 2.2 with T dened exactly as in there except for (r, j)
coming from A1(J) A1(J) instead of A A. To prove that T is countable, we now need to prove
that for any (r, j) T, there is no . A such that r ~ . ~ j.
2.5. (Jaray) Let A be a nonempty set, and % a complete preference relation on A. Prove: % admits
a utility representation i there exists a countable subset 7 of A such that for every r and j in A with
r ~ j, we have
r % . ~ n % j for some ., n 7.
2.6. Let ~ be an interval order on a nonempty set A (Exercise 3.1.9 of Chapter 1). Prove: There exist
two real functions ) and q on A such that
r ~ j i )(r) q(j)
for every r and j in A, i there exists a countable subset 7 of A such that for every r and j in A with
r ~ j, we have
r ~ j implies
#
. _
#
r and . ~ j
for every r and j in A.
3 Representation through Semicontinuity
Separability vs. Semicontinuity
For a complicated prerodered set (A. %), it may be quite dicult to check if %. i:
separable or not, so Theorem 2.2 may not be of immediate use in determining whether
% admits a utility representation or not. In most applications, however, A has some
additional topological structure, and % satises a suitable continuity condition (that
links the order structure of % to the topological structure of A). In such situations,
and when the topology of A is suciently well-behaved, we may approach to the utility
representation problem by means of a dierent approach.
As a major example of what one can achieve by this approach we shall prove below
that every semicontinuous complete preference relation recall Section 3.1 of Chapter
5 admits a utility representation, provided that it is dened on a separable metric
space, or more generally, on a topological space with countable basis. (See Section 1.2
of Appendix.) This was proved rst by Trout Rader in 1963.
10
Proposition 3.1. Let A be a topological space with a countable basis and % a com-
plete preference relation on A. If % is upper (or lower) semicontinuous, then it can be
represented by a utility function.
Proof. Let us assume rst that % is upper semicontinuous. We begin by noting that,
as A is a topological space with a countable basis, there exists a countable collection O
of open subsets of A such that l =

C O : C _ l for every open set l in A.
(See Section 1.2 of Appendix.) This fact allows us to adapt the argument we gave in
the proof of Proposition 2.2 to the present setting.
Enumerate O as C
1
. C
2
. .... and dene
`(r) := i N : C
i
_
#
r
for each r in A. Then, by transitivity of %, we have `(r) _ `() for every r and
in with r % . Furthermore, because % is complete and upper semicontinuous,
#
r is an
open subset of A. and hence,
#
r =
_
C
i
: i `(r) for every r A.
Thus: `(r) `() for every r and in A with r ~ . Consequently, the map
n : A [0. 1], dened by
n(r) :=

i2M(x)
2
i
.
is strictly %-increasing, and we are done.
Next, assume that % is lower semicontinuous. Then, applying what we have just
established to the preordered set (A. -). we nd an upper semicontinuous utility function
n for -. But then n is a lower semicontinuous utility function for %, and our proof is
complete.
Corollary 3.2. Every complete and upper (or lower) semicontinuous preference relation
on a topological space with countable basis is separable.
Proof. Apply Theorem 2.2 and Proposition 3.3.
Semicontinuous Utility Representation
One can in fact show that an upper semicontinuous and complete preference relation
% on a topological space A with a countable basis can be represented by an upper
semicontinuous utility function. This is particularly easy when % is antisymmetric, for
then limsup n. where n is as found in Proposition 3.1, can be shown to be such a utility
function for %. To settle the general case, however, one needs a nontrivial modication
of the method of passing to the quotient.
Raders Representation Theorem. Let A be a topological space with a countable basis
and % a complete preference relation on A. If % is upper (lower) semicontinuous, then
it can be represented by an upper (lower) semicontinuous utility function.
11
Proof. In view of the duality argument given in the last paragraph of Proposition 3.1, it
is enough to establish the assertion in the case where % is upper semicontinuous. In that
case, we begin by applying Proposition 3.1 to obtain a utility function n for %. Without
loss of generality, we assume that n is bounded. Next, we dene the quotient set A,

as in the proof of Proposition 1.3.2. We topologize this set by means of the quotient
topology, that is, we declare a subset C of A,

to be open i

[r]

: [r]

C is an
open subset of A. (See Section 1.9 of Appendix.)
Dene the map l : A,

R by
l([r]

) := n(r).
Since n represents %. it assigns the same value to all members of any given [r]

. There-
fore, l is well-dened. Let \ denote the limsup of l (relative to the quotient topology),
that is, dene \ : A,

R by
\ ([r]

) := inf
O2O(x)
supn(.) : [.]

C.
where O(r) is the collection of all open subsets of A,

that contain [r]

. As the limsup of
any bounded real function on a topological space is upper semicontinuous, \ is an upper
semicontinuous function on A,

. Thus, the map : A R, dened by (r) := \ ([r]

).
is upper semicontinuous.
3
For future reference, we also note that
(r) _ n(r) for every r A. (2)
which is an obvious consequence of the denitions of \ and .
We shall complete our proof by showing that represents %. To this end, take
any elements r and of A. If r ~ . then, clearly, (r) = \ ([r]

) = \ ([]

) = ().
Suppose, then, r ~ . We wish to show that (r) () by distinguishing between two
cases.
Case 1: r ~ . ~ for some . A. In this case, the set C := [.]

: . ~ . includes
[]

. and we have
n(.) _ supn(.) : [.]

C.
Furthermore, by upper semicontinuity and completeness of %. the set
#
. is open in A.
while
#
. =
_
[.]

: [.]

C.
Thus, by denition of the quotient topology, C is an open subset of A,

. that is,
C O(). It then follows from the denition of that n(.) _ \ ([]

) = (). By
(2), then,
(r) _ n(r) n(.) _ ()
since n represents %.
Case 2: r ~ . ~ for no . A. In this case,
n() = maxn(.) : [.]

C
3
After all, this map equals \ j

, where j

: A A,

is the quotient map (which is rendered

continuous by the quotient topology).
12
where C := [.]

: r ~ .. As, again, C is an open subset of A,

that includes []

.
that is, C O(). it follows from the denition of that n() _ (). By (2), then,
n() = (). and hence,
(r) _ n(r) n() = ()
since n represents %.
Curiously, we cannot furnish a continuous utility function from the assumptions of
Raders Theorem. As you are asked to show below, there need not exist a continuous
utility function for an upper semicontinuous preference relation even on a Euclidean
space.
Exercises
3.1. Give a direct proof for Corollary 3.2.
3.2. Give an example of an upper semicontinuous preference relation on R
2
that does not admit a
continuous utility representation.
3.3. (Wold) Let % be a continuous preference relation on R
n
+
. Assume that % is monotonic in the
sense that
x y implies x ~ y
for any x and y in R
n
+
. Prove that the map
x maxa _ 0 : x % a1,
where 1 is the :-vector of 1s, is a continuous utility function for %.
3.4. Prove or disprove: Every monotonic and upper semicontinuous complete preference relation % on
R
2
+
can be represented by a continuous utility function.
3.5. Let A be a nonempty set, and % a complete preference relation on A. Then, the collection of all
sets of the form
#
r,
"
r and
#
r
"
j, where r and j vary over A, constitutes a basis for a topology. Prove
that this topology is second countable i % admits a utility representation.
3.6. (A Generalization of Raders Representation Theorem) Let A be a nonempty set, and % a complete
preference relation on A. Assume that there exists a countable subset 7 of A such that for every r and
j in A with r ~ j, we have r % . ~ n % j for some ., n 7. (Recall Exercise 2.5.)
a. Show that % admits a utility function that is upper semicontinuous with respect to the topology
on A generated by using
#
. : . 7 as a subbasis.
b. Derive Raders Representation Theorem from the previous observation.
4 The Open Gap Lemma
One of the most fundamental theorems of utility theory tells us that asking for the
continuity of % in Raders Representation Theorem would guarantee the existence of a
continuous utility function for this preference relation. This may not appear surprising
at rst. After all, Raders Representation Theorem ensures that there are at least two
utility functions for %, one upper semicontinuous, and the other lower semicontinuous.
Deducing from this that there is, in fact, one continuous utility function for %is, however,
quite dicult. To be able to prove this, we need to look into the order structure of R
somewhat closely.
13
Gaps and Jarays Lemma
Denition. Let o be a subset of R. A gap of o is a _-maximal nondegenerate interval
1 that is disjoint from o and that has lower and upper bounds in o. If, in addition, 1 is
an open interval, then it is said to be an open gap of o.
The following is a useful improvement of Proposition 1.2.1. In words, it says that the
function , in that result can be chosen to have the property that every rational number
: in Q (0. 1) with inf ,(A) < : < sup ,(A). belongs either to the range of , or to an
open gap of the range of ,. This result was proved by Jean-Yves Jaray in 1975.
Jarays Lemma. Let (A. <) be a loset. If A is countable, then there exists a strictly
<-increasing function , : A Q (0. 1) such that, for every rational number : in the
open interval (inf ,(A). sup ,(A)). we have either : ,(A) or
,(r) : ,() and (,(). ,(r)) ,(A) = O
for some r. A.
Proof. The claim is trivial when A is nite, so we assume that A is countably innite.
Let us enumerate A and Q (0. 1) as
A = r
1
. r
2
. ... and Q (0. 1) = :
1
. :
2
. ....
and construct the function , from A into Q(0. 1) exactly as in the proof of Proposition
1.1. Note that, by its construction, , has the following property: For every r. A
with r ~ .
,
_
the rst enumerated element
of
#
r
"

_
=
the rst enumerated element
of (,(). ,(r)) Q.
(3)
Now take any rational number : in (inf ,(A). sup ,(A)). Dene
J := min, N : :
1
. .... :
j
contains : and elements j. ,(A) with j : .
Next, dene
j

:= maxt :
1
. .... :
J
,(A) : : t
and
j

:= mint
1
. ....
J
,(A) : t :.
The key observation here is:
Claim. If (j

. j

) ,(A) ,= O. then : is either the rst enumerated element of

(j

. j

),(A). or it is enumerated before the rst enumerated element of (j

. j

),(A).
Proof. Suppose (j

. j

) ,(A) ,= O. and the rst enumerated element of (j

. j

)
,(A). say . is enumerated before :. (In particular, ,= j.) But then, by denition of
J. the numbers j

. j

and all belong to :

1
. .... :
J
,(A). And yet, we have either
14
j

: j

or j

: j

. while these inequalities cannot hold due to the

choice of numbers j

and j

.
Now, suppose : does not belong to an open gap of ,(A). We wish to show that :
belongs to ,(A). Since : (j

. j

). the interval (j

. j

) cannot be a gap of ,(A). that

is, (j

. j

) ,(A) ,= O. Of course, we have j

= ,() and j

= ,(r) for some . r A

with r ~ . Since , is strictly %-increasing, and (,(). ,(r)) ,(A) ,= O. we must have
#
r
"
,= O. Let c be the rst enumerated element of this set. Then, by (3),
,(c) = the rst enumerated element in (,(). ,(r)) Q.
In particular, either : = ,(c) or ,(c) is enumerated before :. But, since
,(c) (,(). ,(r)) ,(A) = (j

. j

) ,(A).
the Claim above shows that the latter case is impossible. Thus, : = ,(c). that is, :
belongs to the range of ,.
The Open Gap Lemma
We now come to one of the crucial building blocks of ordinal utility theory, the so-called
open gap lemma. This result was rst stated by Gerard Debreu in 1954, but the proof
that Debreu gave had a aw. A complete, but rather involved, proof rst appeared
in Debreu (1964). Since then a number of relatively simple proofs are obtained. For
instance, Bowen (1968) proved this result by means of a measure-theoretic method, and
Beardon (1992) by means of a topological method. We follow here the order-theoretic
approach of Jaray (1975).
The Open Gap Lemma. Let A be a nonempty subset of [0. 1]. Then, there exists
a strictly increasing function , : A [0. 1] such that every gap of ,(A) is either a
singleton or an open interval.
Proof. Let us rst establish the following auxiliary fact:
Claim. There exists a countable subset of A such that, for every r. A.
r implies r _ / c _ for some c. / . (4)
Proof. Since [0. 1] is a separable metric space, so is A. that is, there is a countable
dense subset o of A. Let 1 be the set of all endpoints of the gaps of A that belong to
A. Clearly, 1 is a countable subset of A (because A may have only countably many
gaps). We dene
:= o ' 1.
Obviously, is a countable subset of A. Now, take any r. A such that r . If
(. r) A = O. then (. r) is a gap of A. and hence r. 1. Then (4) holds with c :=
and / := r. If, on the other hand, (. r) A ,= O. then, since (. r) A is open in A and
o is dense in A. there must exist a number / in o such that r / . If (/. ) A = O.
15
then (/. ) is a gap of A. and hence /. 1. Then (4) holds with c := . Otherwise, by
denseness of o. there is an c o such that / c . and we are done.
Given the set found in the Claim above, we next use Jarays Lemma (with A
being ) to nd a strictly increasing function , : Q (0. 1) such that, for every
rational number : in (inf ,(). sup ,()). we have either : ,() or
,(/) : ,(c) and (,(c). ,(/)) ,() = O
for some c. / . Finally, we dene , : A [0. 1] by
,(r) := sup,(c) : r _ c .
Clearly, , is an extension of ,, that is, ,[
A
= ,. In addition, , is strictly increasing.
Indeed, for any r. A with r . the Claim above ensures that there exist c. /
such that r _ / c _ . Then, since , is strictly increasing,
,(r) _ ,(/) ,(c) _ ,(c) for all c with _ c
so that ,(r) ,(c) _ ,().
It remains to prove that every gap of ,(A) is either a singleton or an open interval.
Suppose that 1 is a non-singleton gap of ,(A). Then, O = 1,(A) _ 1,() = 1,().
Moreover, since 1 is a nondegenerate interval, it contains a rational number :. and we
have
sup ,() _ ,(r) : ,() _ inf ,().
where ,(r) and ,() are, respectively, any upper and lower bounds for 1 in ,(A). (We
have ,(r) : ,() because : 1 and 1 ,(A) = O.) Therefore, by the choice of ,.
there must exist numbers c and / in such that
: (,(c). ,(/)) and (,(c). ,(/)) ,() = O.
Since , is strictly increasing, the latter equation implies that (c. /) and are disjoint,
so, by the Claim above, (c. /) A = O. Since , is a strictly increasing extension of ,.
then,
(,(c). ,(/)) ,(A) = (,(c). ,(/)) ,(A) = O.
It follows that (,(c). ,(/)) is a gap of ,(A). Since ,(c) and ,(/) belong to ,(A).
(,(c). ,(/)) is a _-maximal gap of ,(A). But then, since 1 and (,(c). ,(/)) intersect
(for : belongs to both of these sets), we must have 1 = (,(c). ,(/)). so 1 is an open gap
of ,(A).
5 The Debreu-Eilenberg Representation Theorems
Making Utility Functions Continuous
The Open Gap Lemma facilitates the derivation of the two most famous theorems of
ordinal utility theory. In particular, we are now in a position to improve any utility
16
representation theorem to a continuous utility representation theorem in the case of
continuous preference relations.
Proposition 5.1. Let A be any topological space and let : A [0. 1] represent a
preference relation % on A. If % is continuous, then it is representable by a continuous
utility function.
Proof. Assume that % is continuous, and apply the Open Gap Lemma to nd a strictly
increasing , : (A) [0. 1] such that every gap of ,((A)) is either a singleton or an
open interval. Dene n := , and observe that n represents %. We will now prove
that n is upper semicontinuous its lower semicontinuity is established similarly.
Take an arbitrary real number c and let
o

:= . A : n(.) _ c.
Our task is to show that o

is a closed subset of A. If c = n(r) for some r A. then

o

= r
"
(because n represents %), and we are done by the upper semicontinuity of %.
We then consider the case where c Rn(A). Clearly, if c _ inf n(A). then we have
o

= A. and if c _ sup n(A). then o

= O; so our claim is trivial in these cases. Assume

then that c belongs to the open interval (inf n(A). sup n(A)). and let 1 be a gap of n(A)
that contains c. (Clearly, 1 is the union of all intervals in Rn(A) that contain c.) By
denition of n. either 1 = c or 1 = (c

. c

) for some c

and c

in n(A) with c

.
In the latter case, we have
o

= . A : n(.) _ c

=
"
.
where A satises n() = c

. Thus o

is a closed set, thanks to the upper semicon-

tinuity of %. In the former case, on the other hand,
o

: c , n(A)
Since o

is closed for each , n(A), and the intersection of any collection of closed sets
is closed, we nd again that o

is a closed set.
Two Theorems on Continuous Utility Representation
The following is one of the most celebrated results of utility theory.
The Debreu Representation Theorem. Let A be a topological space with a countable
basis, and % a complete preference relation on A. If % is continuous, then it can be
represented by a continuous utility function.
Proof. Apply Propositions 3.1 and 5.1.
In applications, A often takes the form of a separable metric space. As such a space
is a topological space with a countable basis, the Debreu Representation Theorem deals
with such cases readily:
17
Corollary 5.2. Every continuous preference relation on a separable metric space admits
a continuous utility representation.
Samuel Eilenberg proved in 1941 that any continuous preference relation on a con-
nected and separable topological space admits a continuous utility representation. (This
result replaces the countable basis requirement of Debreus Theorem with connected-
ness.) The technique we used to prove the Debreu Representation Theorem can also be
used to prove this result.
The Eilenberg Representation Theorem. Let A be a connected and separable topo-
logical space. and % a complete preference relation on A. If % is continuous, then it can
be represented by a continuous utility function.
Proof. In view of Proposition 5.1, all we need to prove here is that % admits a utility
representation. Let 1 be a countable dense subset of A, and pick any r. A with
r ~ . (If there is no such r and in A. then any constant function would represent
%.) Notice that
#
r and
"
are open subsets of A such that
#
r'
"
= A (because % is
complete and continuous). Since A is connected, we must then have
#
r
"
,= O. But
then
#
r
"
is a nonempty open subset of A. and hence, since a dense set intersects
every nonempty open set, we must have 1
#
r
"
,= O. This proves that 1 is %-dense
in A. Applying Proposition 2.1, therefore, we may conclude that % admits a utility
representation.
Note. There are various generalizations of the Eilenberg Representation Theorem. Campin et.
al. (2007), for instance, have recently shown that it is enough to take A to be locally connected
(and separable) in this result.
4
As we have mentioned above, the underlying set A of alternatives is usually taken
to be a metric space, say, in economic applications. (For instance, when modeling
choice over commodity bundles, one takes A to be R
n
. and when modeling choice over
intertemporal income streams, A is often taken as a suitable normed linear space of
real sequences.) Within such contexts, Debreus theorem is more general than that of
Eilenberg, because it does not require the connectedness hypothesis.
Application. (Choice Correspondences) We briey revisit the theory of choice corre-
spondences introduced in Section 3.3 of Chapter 5. Let A be a metric space, and C a
choice correspondence on k(A). the collection of all nonempty compact subsets of A.
Recall that the Fundamental Theorem of Revealed Preference tells us that C can be
thought of as arising from the maximization of a continuous preference relation, that is,
there exists a continuous and complete preference relation % on A such that
C(o) = max(o. %) for every o k(A).
4
A topological space A is said to be locally connected if, for any point r in A and any open
neighborhood l of r in A, there is a connected set \ with r \ _ l.
18
provided that C satises the rationality properties of WARP and C. In this case, if we
know also that A is separable, we may invoke the Debreu Representation Theorem
or more precisely, Corollary 5.2 to conclude that there exists a continuous (utility)
function n : A R such that
C(o) = arg max n(o) for every o k(A). (5)
where, of course, arg max n(o) stands for the set of all r in o that maximizes n on
o. This gives us a deeper version of our earlier nding on the representation of choice
correspondences:
Fundamental Theorem of Revealed Preference 2. Let A be a separable metric space,
and C a function from k(A) into 2
X
. Then, C is a choice correspondence on k(A) that
satises WARP and C if, and only if, there exists a continuous real function n on A
such that (5) holds.
6 Multi-Utility Representation
Obviously, a preference relation that admits a utility representation must be complete.
But this does not mean that there is no useful notion of utility representation for possibly
incomplete preference relations. Indeed, we have already seen in Section 4.2 of Chapter
1 that every preference relation can be represented as the intersection of a collection of
complete preference relations. If every one of the preference relations in that collection
admits a utility representation, we arrive at a useful functional representation for our
original relation.
Denition. Let % be a preference relation on a nonempty set A. We say that a nonempty
set | of real functions on A represents % provided that
r % if and only if n(r) _ n() for each n |
for every r. A. If such a set | exists, we say that % admits a multi-utility rep-
resentation. If | is nite (countable), we say that there exists a nite (countable)
multi-utility representation for %. If | represents %. and every n in | is strictly
%-increasing, then we say that | properly represents %, and that % admits a proper
multi-utility representation.
The following example is prototypical.
Example 6.1. The standard partial order _ of R
n
admits a (nite) multi-utility represen-
tation. Recall that this partial order is dened as r _ i r
i
_
i
for each i = 1. .... :.
19
Then, where n
i
: R
n
R is dened by n
i
(r) := r
i
for each i = 1. .... :. it is obvi-
ous that n
1
. .... n
n
represents _. This partial order also admits a proper multi-utility
representation. For,
_
n
j
+ `
n

i=1
n
i
: , = 1. .... : and ` 0
_
properly represents _.
It is worth noting that the reasoning we gave in the context of the example above
readily yields the following general principle:
Proposition 6.1. Every preference relation that admits a nite multi-utility represen-
tation admits also a proper (but not necessarily nite) multi-utility representation.
Another elementary observation concerns the universality of the notion of multi-
utility representation.
Proposition 6.2. There exists a multi-utility representation for every preference relation.
Proof. Let A be a nonempty set and % a preference relation on A. For any subset o of
A. let us write 1
S
for the indicator function of o on A; that is,
1
S
(r) :=
_
1. if r o.
0. if r Ao.
It is easily checked that | := 1
x
" : r A represents %.
Proposition 6.2 shows that the notion of utility representation is strictly more de-
manding than that of multi-utility representation for complete preference relations.
Example 6.2. It is plain that if a preference relation admits a utility representation,
then it admits a multi-utility representation. The converse is false, even for complete
preference relations. After all, we have seen in Example 2.1 that the lexicographic order
on R
2
cannot be represented by a utility function. Yet, this linear order admits a multi-
utility representation by Proposition 6.2.
We conclude by noting that the utility functions found in Proposition 6.2 provide us
with an upper semicontinuous multi-utility representation.
20
Proposition 6.3. Let A be a topological space and % a preference relation on A. If %
is upper (or lower) semicontinuous, then it can be represented by a set of upper (lower
resp.) semicontinuous utility functions.
Proof. Let % be an upper semicontinuous preference relation on A. Then, | := 1
x
" :
r A represents %. Moreover, for each r A, the set r
"
is closed in A by upper semi-
continuity of %. and hence 1
x
" is upper semicontinuous. For the lower semicontinuous
part of the assertion, it is enough to replace | with 1
x
# : r A in this reasoning.
Exercises
6.1. Let < be a partial order on R such that r ~ j i r j + 1 for every real numbers r and j. Does
< admit a multi-utility representation? A proper multi-utility representation? A nite multi-utility
representation?
6.2. Let A be a topological space and % a near-complete and upper semicontinuous preference relation
on A. Prove that there exists a nite set | of upper semicontinuous real functions on A that represents
%.
(Hint. Examine the proof of Proposition 5.2.4 of Chapter 3.)
6.3. (Uniqueness of Multi-Utility Representations) Given any nonempty set A, and a nonempty subset
| of R
X
, dene the map
U
: A R
U
by

U
(r)(n) := n(r).
Prove: Two nonempty subsets | and 1 of R
X
represent the same preference relation on A if, and only
if, there exists a map ) :
U
(A) R
V
such that (i)
V
= )
U
; and (ii) for every c, ,
U
(A), we
have c , i )(c) )(,).
6.4. (Ok) Let (A, %) be a preordered set such that % is near-complete and there is a countable 7 _ A
such that for every r, j A with r ~ j, we have r ~ . on j for some . 1. Prove that % admits a
nite multi-utility representation.
7 Continuous Multi-Utility Representation
We now turn our attention to the problem of nding a set of continuous utility functions
for a preference relation dened on a topological space. This problem is signicantly
more complicated than its semicontinuous counterpart. In most cases it requires one
to solve rst a monotonic extension problem. Fortunately, there are some results in
the topological literature which are tailor made to the needs of this problem. We shall
start this section by examining one major such result, and return to the continuous
multi-utility representation problem subsequently.
7.1 Nachbins Extension Theorem
In particular, much is accomplished by the following extension theorem, which was
proved by Leopoldo Nachbin in 1950, is of fundamental importance for the theory of
ordered topological spaces.
Nachbins Extension Theorem. Let A be a compact Hausdor topological space, % a
closed-continous preorder on A. and 2 a closed subset of A. If , : 2 R is a continuous
21
and %-increasing function, then there is a continuous and %-increasing function 1 : A
R such that 1[
Z
= ,.
Proof. TBW.
7.2 Levins Theorem
Recall that our present goal is to identify those preference relations on a given topo-
logical space that admit multi-utility representations by means of a set of continuous
utility functions. It is easy to verify that any preference relation that admits such a
representation must be closed-continuous. We shall now show that, provided that the
prize space under consideration is suitably well-behaved, the converse of this observation
is also true.
Theorem 7.2.1. Let A be a compact Hausdor topological space. and % a preference
relation on A. If % is closed-continuous, then it can be represented by a set of continuous
utility functions.
Proof. Dene
:= (r. ) A A : r % is false.
If = O. then r ~ for all r. A. so any nonempty set of constant functions on
A represents %. We assume, then, ,= O. In turn, for any (r. ) . we dene the
function
x;y
: r. R as

x;y
(r) := 0 and
x;y
() := 1.
which is a continuous and %-increasing function on r. . By Nachbins Extension
Theorem, for each (r. ) there is a continuous and %-increasing function n
x;y
: A
R such that n
x;y
(r) := 0 and n
x;y
() := 1. Then, | := n
x;y
: (r. ) represents %.
Indeed, for any r and in A. if ~ r, then n() _ n(r) for all n | (since each n |
is %-increasing) while
n
x;y
() = 1 0 = n
x;y
(r).
If ~ r, then n(r) = n() for all n | (since each n | is %-increasing). Finally, if
neither r % nor % r hold, then
n
x;y
() = 1 0 = n
x;y
(r) and n
y;x
(r) = 1 0 = n
y;x
()
because we have (r. ) and (. r) , respectively.
Thanks to the compactness of A. it is possible to sharpen Theorem 7.2.1 as follows:
Corollary 7.2.2. Let A be a compact Hausdor topological space. and % a preference
relation on A. If % is closed-continuous, then it can be represented by a countable set
of continuous utility functions.
22
Proof. As usual, we let C(A) stand for the set of all continuous real functions on A. and
metrize this set by means of the sup-metric. Since A is compact, C(A) is a separable
metric space.
5
Now, by Theorem 7.2.1, there exists a nonempty subset | of the set
C(A) that represents %. As a metric subspace of a separable metric space is separable,
there is a countable subset 1 of | which is dense in | (relative to the sup-metric).
We claim that 1 represents %. To see this, take any elements r and of A. Obviously,
if r % . then (r) _ () for every 1 (because 1 _ |). Conversely, suppose that
(r) _ () for all 1. (6)
Since | represents %, if r % were false, we would have n() n(r) for some
n |. Then, by denseness of 1 in |. we could nd a sequence (
m
) in 1 such that
|
m
n|
1
0. This implies that
m
()
m
(r) for a positive integer : large enough,
which contradicts (6).
Another shortcoming of Theorem 7.1 is that this theorem does not tell us if we can
properly represent the preference relation under consideration. However, as we shall now
show next, there is an easy remedy for this. First, let us recall the following elementary
fact from real analysis.
The Weierstrass `-Test. Let A be a topological space. and (n
m
) a sequence of con-
tinuous functions on A. Suppose that there exist a real sequence (`
1
. `
2
. ...) such that
|n
i
|
1
_ `
i
for each i = 1. 2. ...
and
1

i=1
`
i
< .
Then,

1
n
i
(r) exists for each r A. and r

1
n
i
(r) is a continuous map on A.
Proof. Dene ,
m
:= n
1
+ + n
m
for each positive integer :. Take any 0. Since

1
`
i
< . there exists a positive integer :

such that
1

i=m

`
i
< .
Then,

i=m

n
i
(r)

_
1

i=m

[n
i
(r)[ _
1

i=m

`
i
<
for every integer : _ :

and r A. It follows that

[,
m
(r) ,
m
(r)[ =

i=m

n
i
(r)

< for every r A and : _ :

.
5
This is a fairly straightforward consequence of the Stone-Weierstrass Theorem. (In fact, one can
show that, for any Hausdor topological space A, C(A) is separable i A is compact.)
23
Conclusion: (,
m
) is a Cauchy sequence in the metric space C
b
(A) of continuous and
bounded functions on A. Since C
b
(A) is a complete metric space see Example 2.4.5
of Appendix we may thus conclude that lim,
m
C
b
(A).
We are now ready for:
Levins Theorem. Let A be a compact Hausdor topological space. and % a preference
relation on A. If % is closed-continuous, then there is a continuous and strictly %-
increasing real function on A.
Proof. Assume that % is closed-continuous. By Corollary 7.2, there is a countable set
1 of continuous real functions on A which represents %. If 1 is nite, then the proof is
completed by using the trick we used at the end of Example 6.1. We thus assume that
1 is countably innite. Let us enumerate 1 as
1
.
2
. .... and dene
n
i
:=
2
i

i
1 +|
i
|
1
. i = 1. 2. ...
Since 1 represents %, we have
r % i n
i
(r) _ n
i
() for each i = 1. 2. ... (7)
while
r ~ implies r % and n
i
(r) n
i
() for some i = 1. 2. ... (8)
Moreover, as each [n
i
[ is bounded above by 2
i
, the Weierstrass `-test ensures that
n : A R is well-dened by
n(r) :=
1

i=1
n
i
(r)
as a continuous function on A. In turn, (7) and (8) guarantee that n is strictly %-
increasing.
Combining Theorem 7.2.1 and Levins Theorem yields the sharpening of Theorem
7.2.1 that we are after:
Theorem 7.2.3. Let A be a compact Hausdor topological space. and % a preference
relation on A. If % is closed-continuous, then it can be properly represented by a set of
continuous utility functions.
Proof. By Theorem 7.2.1, there exists a set | of continuous functions on A which
represents %. By Levins Theorem, there exists a strictly %-increasing real function ,
on A. Then
n + `, : n | and ` 0
properly represents %.
24
Note. With a slightly more involved argument, one can show that A can be taken to be a locally
compact and separable metric space (such as an arbitrary closed subspace of a Euclidean space)
in Levins Theorem. Since Evren and Ok (2009) prove that Theorem 7.1 is valid when A is such
a space, it follows that one can take A in Theorem 7.3 to be a locally compact and separable
metric space.
The next step on the agenda should be to look for ways of improving this theorem
towards obtaining a suitable continuous nite multi-utility representation theorem. A
natural conjecture is that this can be done by strengthening the upper semicontinuity
hypothesis to continuity of %. Unfortunately, this conjecture is false, and, in fact,
there does not seem to be a natural way of deriving a continuous nite multi-utility
representation theorem, at least, not by using the standard methods of utility theory.
We conclude our presentation with an example that aims to illustrate just how elusive
this representation problem is.
Example 7.2.1. Dene < on [0. 2] as
/ < c i
either 0 _ c. / _ 1 and / _ c,
or 1 _ c. / _ 2 and c _ /,
or / = 2.
Then, < is a closed-continuous partial order on [0. 2], and hence, by Theorem 7.2.1, there
exists a continuous multi-utility representation for it. Moreover, < is near-complete. In
fact, n. represents <, where n and are real maps on [0. 2], dened as
n(t) :=
_
_
_
1 t, if 0 _ t _ 1,
0. if 1 < t < 2,
1. if t = 2
and (t) :=
_
0. if 0 _ t _ 1.
t 1. if 1 < t _ 2.
Notice that is continuous and n is upper semicontinous. Moreover, n is continuous
everywhere but at 2.
Surprisingly, there does not exist a nite set | _ C[0. 2] that represents <. Indeed, if
| was such a set, then, for every non-constant member n of | we would have n(2) n(1).
and hence, by using niteness of |. we could nd large enough : [0. 1) and t [1. 2)
such that n(t) n(:) for all non-constant n |. which would contradict | representing
<.
Exercises
7.1. Let < be a partial order on R such that r ~ j i r j + 1 for every real numbers r and j. Does
< admit a multi-utility representation? A proper multi-utility representation? A nite multi-utility
representation?
7.2. Let A := [0, 1]
2
and consider the real maps ) and q on A dened as
)(x) := r
1
r
2
and q(x) :=
_
)(x), if x ,= 0,
2, if x = 0.
Dene the preorder % on A by x % y i )(x) _ )(y) and q(x) _ q(y).
a. Show that % admits a semicontinuous multi-utility representation.
b. Show that % does not admit a continuous multi-utility representation.
25