You are on page 1of 35

Integral Utility Models for Outcomes over Continuous Time*

Charles M. Harvey
Dept. of Decision and Information Sciences, University of Houston

Lars Peter sterdal


Department of Economics, University of Copenhagen

February 2009

PRELIMINARY VERSION. REFERENCES INCOMPLETE.


PLEASE DO NOT QUOTE WITHOUT THE AUTHORS PERMISSION

Abstract
Outcomes that occur over time can often be described as real- or vector-
valued functions defined on a bounded or unbounded interval of time.
This paper presents representation theorems for preferences between such
functions. Conditions on preferences are shown to be satisfied if and only
if the preferences are represented by a utility function that is an integral of
a discounting function times a utility scale on outcomes at instants of time.

Key words: continuous time, discounting, intertemporal equity, integral utility function.

* This paper is based on results from unpublished working papers by Harvey (1998a,b).
His research was supported in part by U.S. NSF/EPA Grant No. GAD-R825825010. It
subsumes a previous draft titled, Preferences between continuous streams of events,
[Discussion paper 05-12 (2005), Dept. of Economics, University of Copenhagen].

Correspondence: Lars Peter sterdal, Dept. of Economics, University of Copenhagen,


Studiestraede 6, DK-1455 Copenhagen K, Denmark. Email: lars.p.osterdal@econ.ku.dk
1

1. Introduction

An analyst who is developing an evaluation of alternatives in a decision model must


judge whether to model the consequences of the alternatives as outcomes that occur at
discrete times, to be called outcome-sequences, or as outcomes that occur over continuous
time, to be called outcome-streams. Outcome-sequences are modeled as functions defined
on a discrete set of times and outcome-streams are modeled as functions defined on an
interval of time, to be called a planning period. The judgment as to which type of model to
use depends on the nature of the data, the nature of the consequences, and the proclivities
of the analyst. Each type seems more appropriate under some circumstances.
This paper is concerned with outcome-streams. The outcomes are numbers or vectors
x, and the outcome-streams are real- or vector-valued functions x(t) defined on times t
in the planning period P. At an instant t, the value x = x(t) of an outcome-stream x
is a rate or an amount, or it is a vector of rates and amounts. For example, a value x(t)
might be: the rate of usage of a natural resource, one or more rates of monetary benefits
and costs, an amount that measures a type of environmental quality, or multiple amounts
that describe the health of an individual.
Comparisons between outcome-streams will be modeled as a binary relation , that
is, as a set of statements x y (which also use the symbol ). We will call a relation
a preference relation, and we will regard a statement x y as meaning that x is at
least as preferred as y. However, a relation can be interpreted equally well as hedonic
comparisons of the outcome-streams.
Relations of strict preference (f) and indifference (~) are defined in terms of a relation
, namely: x f y provided that x y and not y x, and x ~ y provided that x y
and y x. A function U(x) is called a utility function for a preference relation on a
specified set provided that: U(x) U(y) if and only if x y for any x, y in the set.
By a utility model, we will mean a result concerning a set of alternatives (outcome-
streams or some other type) which shows that if a preference relation defined on the set
satisfies certain conditions then there exists a utility function having certain properties
2

that represents the preference relation. Some utility models also show that the conditions
on the preference relation are necessary, that is, the existence of a utility function with the
specified properties implies the specified conditions on the preference relation.
Utility models for outcome-sequences are well-known. They provide different degrees
of generality regarding preferences. On the one hand, there are models in which the utility
function is a linear function, U(x) = a(t) x(t), of the outcomes x(t); e.g., Williams and
Nassar (1965). On the other hand, there are models in which the utility function has an
additive form, U(x) = a(t) u(x(t)); e.g., Koopmans (1960, 1972), Diamond (1965), and
Harvey (1986, 1995). Linear models require that u(x) is the identity function; thus, they
exclude issues of temporal balance, decreasing marginal utility, and intertemporal equity.
Linear models also exclude multivariable outcomes.
The results in this paper are utility models for outcome-streams. They provide utility
functions that have the integral form, U(x) = P a(t) u(x(t)) dt, where a(t) and u(x) are
functions that have specified properties. By analogy with utility models for outcome-
sequences that contain an additive utility function, U(x) = a(t) u(x(t)), we will call the
function U(x) in these models an integral utility function. The function a(t) will be
called a discounting function; the function u(x) will be called an outcome scale; and the
entire model will be called an integral utility model.
An outcome scale represents preferences between the outcomes in an integral utility
model. These preferences are defined by the preference relation on outcome-streams.
And a discounting function represents tradeoffs between amounts of the outcome scale
at different times. For more detailed interpretations of these functions, see the working
paper, Harvey and sterdal (2005), on which this paper is based.
A utility model for outcome-streams has been developed in which the utility function
has the form, U(x) = P a(t) x(t) dt, corresponding to a linear utility function in a model for
outcome-sequences, but no utility model has been developed in which the utility function
has the form, U(x) = P a(t) u(x(t)) dt, corresponding to an additive utility function. Some
researchers seem to have a contrary impression, and for this reason we need to provide an
account of previous research. Such an account is given in the next section.
3

2. Overview of the paper and previous research

This paper consists of four sections. The first two (Sections 3, 4) define outcome-
streams and describe conditions on a preference relation on outcome-streams, and the
next two (Sections 5, 6) present integral utility models that are based on these conditions.
Section 5 presents two models for outcome-streams that are defined on a planning
period with a finite horizon, that is, P = [0, T]. In the first model, the outcome-streams
are functions or vectors of functions that are step functions and in the second model the
outcome-streams are functions or vectors of functions that are Riemann integrable.
Section 6 presents two models for outcome-streams defined on a planning period with
an infinite horizon, that is, P = [0, ). In the first model, an outcome is specified which
intuitively speaking has the role of a null outcome, and each outcome-stream has this null
outcome at any time beyond a finite horizon that depends on the outcome-stream. In the
second model, the preference relation is defined on a set of outcome-streams and a subset
of those to be compared by the utility function is defined in terms of the preference rela-
tion. This subset includes the finite-horizon outcome-streams and certain infinite-horizon
outcome-streams. The dependence of the subset on preferences allows different instances
of the model to have outcome-streams with different types of behavior at infinity.
Instances of a model can have widely different characteristics. A discounting function
a(t) may or may not be a negative-exponential function. In particular, it can be constant
or tend to zero slowly as time tends to infinity in order to assign a greater importance to
outcomes in the distant future than that assigned by a negative-exponential discounting
function. Or it can tend to infinity as time tends to zero to represent extreme discounting
of future outcomes, even those in the near future.
The outcome scale u(x) is not required to have any property other than continuity.
For example, u(x) is not required to be additive in the case of multivariable outcomes.
In all four models, the outcome-streams are vectors of one or more functions that are
Riemann integrable on bounded intervals. For two reasons, this class of outcome-streams
seems more natural than the larger class with Riemann replaced by Lebesgue. The first
4

reason is technical. A virtue of Lebesgue integration is that a class of Lebesgue integrable


functions is closed under certain limit processes. But that feature is not relevant here since
sequences of outcome-streams are not considered. A virtue of Riemann integration is that a
Riemann integrable function can be approximated both above and below by step functions.
That feature is very relevant here since we use it in Section 5 to extend the model for step
functions to a model for Riemann integrable functions. It is an open question whether
utility models similar to those here are possible for Lebesgue integrable functions.
The second reason is concerned with applications of the models. Riemann integrable
functions seem adequate for describing most phenomena. Functions that are Lebesgue
integrable but not Riemann integrable, e.g., the characteristic function for a Cantor set,
are used to construct counterexamples. Moreover, we are sceptical that anyone can judge
whether or not a condition on preferences between such functions is appropriate.
The results here are if and only if results; they establish that a preference relation in
a model satisfies specified conditions if and only if it is represented by an integral utility
function that has specified properties. In this sense, we do not assume extra technical
conditions such as solvability or differentiability.
Proofs of results are provided in the Appendix. This material is surprisingly long. But
we have not found any way to construct shorter proofs.
It remains to describe previous research. We know of only two discounting models in
the literature for outcomes over continuous time. First, there is the work of Grodal and
Mertens (Grodal and Mertens, 1968 and Grodal, 2003). The Grodal-Mertens model shows
that if a preference relation defined on a set G of outcome-streams satisfies certain
conditions then there exists a integral function, U(x) = P a(t) u(x(t)) dt, such that: x f y
implies U(x) > U(y) for any x, y in G.
Since is complete (Grodal, 2003, p. 153), this is equivalent to the property that:
U(x) U(y) implies x y for any x, y in G. However, their model does not show the
converse implication, that: x y implies U(x) U(y) for x, y in G, and thus it is not
a utility model. What one can conclude is that there exists a preference relation such
5

that the set of statements x y is a subset of the set of statements x y and U(x) is a
utility function for .

Second, there is the work of Weibull (1985). His model is a utility model. It shows the
existence of a linear utility function, U(x) = P a(t) x(t) dt. Thus, it corresponds to models
for outcome-sequences that have a linear utility function. Like those models, it excludes
multivariable outcomes and the issues of preferences mentioned above.
In summary, the key differences between the above two models and the models in this
paper are that the Grodal-Mertens model is not a utility model and the Weibull model is a
linear model. However, there are other differences that also warrant mentioning.
The planning period in the Grodal and Weibull models is a measure space, and the
outcome-streams are measurable functions. For purposes of comparison, we will assume
that the measure space is the interval P = [0, ) with Lebesgue measure. Then, in each
model the set of outcome-streams is a specified set of Lebesgue measurable functions,
and the preference relation is defined on the specified set. The Grodal model assumes that
any constant outcome-stream is in the set. The Weibull model assumes that any outcome-
stream in the set has a finite integral and that for any outcome-stream x in the set and
any multiple a > 0 the outcome-stream a x is also in the set. Some of these modeling
assumptions seem insufficiently restrictive in that they allow outcome-streams that are far
removed from those of interest in an application, and some of the assumptions seem
overly restrictive in that they exclude outcome-streams that may be needed.
The Grodal model requires that the discounting function a(t) has a finite integral.
This excludes a discounting function with a zero discount rate and some (but not all)
discounting functions with a discount rate that decreases to zero as time tends to infinity.
By contrast, the Weibull model requires that the discounting function is bounded. This
excludes those discounting functions that tend to infinity as time tends to zero.

3. Outcomes and outcome-streams


6

Outcomes will be described by N 1 real-valued variables xj whose values are in


intervals Xj , j = 1, , N. An outcome will be a vector x = (x1, , xN ) in the product
set X = X1 XN . The set X will be called an outcome set.
A planning period P will be a bounded interval [0, T ], 0 < T < , or the unbounded
interval [0, ). An outcome-stream will be a vector-valued function x = (x1, , xN )
whose domain is a planning period P and whose values are in an outcome set X.
Outcome-streams of any type will be denoted by letters near the end of the alphabet,
e.g., x, y, etc., and outcome-streams that are constant will be denoted by letters near the
beginning of the alphabet, e.g., a, b, etc. An outcome-stream a is to have the value a
for any time t, and so forth. For two outcome-streams x, y and a time interval , ,
(x, , y) will denote the outcome-stream with (x, , y)(t) = x(t) for t in ,
and (x, , y)(t) = y(t) otherwise. And for outcome-streams x, y, z and two disjoint
time intervals , and , , (x, , y, , z) will have a similar meaning.

Step outcome-streams. Suppose that a, b denotes an interval with the endpoints a, b.


Here, a and b may or may not be in a, b. A partition p of an interval [0, T ] will be
a set of subintervals, a0, a1, , am1, am, where 0 = a0 a1 am1 am = T
and the subintervals are disjoint with the union [0, T ].
A step outcome-stream based on a partition p will be an outcome-stream of the form,
x(t) = x(i) = (x1(i), , xN(i) ) for t in ai1, ai , i = 1, , m. The set of step outcome-
streams based on a partition p will be denoted by Sp , and the union of the sets Sp will
be denoted by ST .
A step outcome-stream x = (x1, , xN ) has associated step functions, xj(t) = xj(i)
for t in ai1, ai , i = 1, , m. We define: the distance between two outcomes x, y
as, | x y | = j | xj yj | , the distance between two step functions xj , yj as, | xj yj | =
i (ai ai1) | xj(i) yj(i) | , and the distance between two step outcome-streams x, y as,
| x y | = i (ai ai1) | (x(i) y(i) | . Then, | x y | = j | xj yj | .

Preferences between outcomes. We will define a preference relation on outcomes so


that these preferences coincide with the preferences between outcome-streams that have a
7

constant outcome during a non-point interval of time. To formalize the situation, suppose
that a preference relation has been defined on a set C of outcome-streams. Then, we
make the following assumption.

Assumption 1. There exists a time > 0 in the planning period P and an outcome
in the outcome set X such that any outcome-stream (a[0,], ), a in X, is in the set C .

Now, we define a preference relation X on the outcome set X by:

a X b if and only if (a[0,], ) (b[0,], ) .

In the next section, we introduce conditions on which imply that the relation X

is the same for any choice of > 0 and . In the model in Section 6 with the planning
period P = [0, ) in which the set C depends on the preference relation , we will
choose as the null outcome mentioned above and in Section 2.
We will assume that the preference relation X is non-trivial in the sense below. The
purpose of this assumption is to avoid an uninteresting special case.

Assumption 2. The set X contains outcomes that are not indifferent.

Definition 1. A pair (C, ) will be called an outcome-stream space and the related pair
(X, X) will be called an outcome space provided that Assumptions 1, 2 are satisfied.

The definitions and assumptions regarding outcomes in this section generalize those
in Harvey (1998a,b). In those papers, the outcomes are amounts of a single variable in a
nonpoint interval and greater amounts are preferred.

4. Conditions on preferences

This section presents conditions on preferences in an outcome-stream space (C, ).


The later sections present integral utility models in which these conditions both imply and
are implied by the existence of an integral utility function that has specified properties.

(A) concurs with X on C: For any x, y in C,


(a) If x(t) X y(t) almost everywhere (a.e.) for t in P, then x y.
8

(b) If x(t) X y(t) a.e. in P and x(t) fX y(t) on a non-point interval, then x f y.

(B) is transitive on C: For any x, y, z in C, if x y and y z, then x z.


is complete on C: For any x, y in C, either x y or y x.

(C) is continuous on C with respect to ST : For any x in C and any w in ST C,


(a) If w p x, then there exists a > 0 such that | z w | < implies that z p x for
any z in ST C.
(b) If w f x, then there exists a > 0 such that | z w | < implies that z f x for
any z in ST C.

Intuitively speaking, the conditions (A)(C) on an outcome-stream space (C, ) are


ordinal conditions, e.g., they do not imply the existence of a utility function for that
has any particular form. Here, we state several implications of (A)(C) for the outcome
space (X, X) associated with (C, ).
A preference relation X will be called continuous provided that for any outcomes
x f y there exists a > 0 such that for any outcome w: |w x| < implies w f y and
|w y| < implies w p x.

Lemma 1. If an outcome-stream space (C, ) satisfies conditions (A)(C), then:


(a) X is transitive, complete, and continuous.
(b) There exists a continuous utility function for X.

(c) The range of any continuous utility function for X is a non-point interval.
(d) X does not depend on the choice of and , that is, for any time > 0 and
any outcomes a, b and , if the outcome-streams below are in C, then:

(a[0,], ) (b[0,], ) if and only if (a[0,], ) (b[0,], ) .

By contrast with (A)(C), the conditions below are additive conditions. Together with
(A)(C), they will imply the existence of a utility function that has an integral form.
9

(D) is preferentially independent on C: Suppose that , is a bounded interval in


P and that the following outcome-streams are in C. Then, (x, , x) (x, , y)
implies that (z, , x) (z, , y).

Condition (D) states that if two outcome-streams are equal during an interval ,
(so that a comparison depends on outcomes at other times), then the common outcome-
stream in , can be changed to another common outcome-stream in , without
changing the comparison. Conditions analogous to (D) play an essential role in additive
utility models: e.g., Debreu (1960) and Gorman (1968). To see the analogy, imagine that
a planning period P = [0, T ] is partitioned into subintervals, ai1, ai, i = 1, , m,
such that ai1, ai = , for some i . Then, for the set of outcome-streams that are
constant on each of the m subintervals in the partition, condition (D) reduces to a
condition of preferential independence for a set of outcomes with m components.
Two pairs a, b and c, d of outcomes will be called tradeoffs pairs with respect to
two intervals , and , in P provided that the intervals are bounded and disjoint
and (a,, d,, ) ~ (b,, c,, ). An outcome will be called a tradeoffs
midvalue of a pair a, a of outcomes for an interval , provided that there exists an
interval , and a pair c, d of outcomes such that a, and c, d are tradeoffs pairs
and , a and c, d are tradeoffs pairs with respect to the intervals , and , .
Condition (E) below is a requirement on preferences between outcome-streams of the
form (a,, b,, ). Thus, we use it only for sets C that include any such outcome-
stream. A variety of analogous conditions for vectors and discrete-time consequences are
described in Fishburn (1970), Krantz et al. (1972, page 305), and Harvey (1986, 1995).

(E) is midvalue independent on C: For any bounded intervals , , , in P,


if an outcome pair a, a has tradeoffs midvalues both on , and on , , then
the outcome pair a, a has the same tradeoffs midvalues on , and , .

5. Models for a bounded planning period


10

This section presents two integral utility models for outcome-streams that are defined
on a bounded planning period P = [0, T ]. First, we present a model for the set ST step
outcome-streams on [0, T ]. Then, we extend this result to present a model for a set of
outcome-streams x = ( x1, , xN ) on [0, T] whose component functions x1, , xN
are Riemann integrable on [0, T ].

Theorem 1. An outcome-stream space (ST , ), T > 0, satisfies conditions (A)(E) if


and only if it has a utility function of the form,

U(x) = 0T a(t)u(x(t)) dt , x in ST . (1)

such that the Lebesgue integral (1) exists for any x in ST and:
(a) The function u(x) defined on the outcome set X is continuous, has a non-point
interval range, and is a utility function for the outcome space (X , X ).

(b) The function a(t) defined on the planning period P = [0, T ] is non-negative and
Lebesgue integrable.
(c) The function A(t) = 0t a(s) ds defined on [0, T ] is strictly increasing, absolutely
continuous, and has the value A(0) = 0.
(d) The function U(x) is continuous on ST , that is, for any x in ST and any > 0
there exists a > 0 such that | z x | < implies | U(z) U(x) | < for any z in ST .
Moreover, the function u(x) is unique up to a positive affine transformation, and the
function A(t) is unique up to a positive multiple.

The function a(t) can be interpreted as a discounting function, and the indefinite
integral A(t) is then a cumulative discounting function. The fact that a(t) is not required
to be Riemann integrable is of practical importance since a Riemann integrable function
is bounded. In particular, the model allows discounting functions that are unbounded near
the present, t = 0, and thus represent extreme discounting of all future times.
The properties (a)-(c) of the functions u(x), a(t), and A(t) do not imply the joint
continuity property (d) of the function U(x). For a counterexample, see Harvey (1998b).
11

Since a step outcome-stream x is in a set Sp , i.e., x is a step outcome-stream with


respect to some partition p: ai1, ai, i = 1, , m, the integral (1) reduces to a sum,

m
U(x) = i=1( A(ai ) A(ai1 ) )u(x(i)) , x in Sp . (1)

It follows that U(x) has the same value for any partition p with x in Sp .
While each sum for U(x) has a finite number of terms, there is no upper bound on the
number of terms. Hence, the model in Theorem 1 is not a Debreu additive utility model.
Our method of proof proceeds in a direction opposite to the above derivation of (1)
from (1). First, we construct an additive utility model with a utility function (1) for a set
Sp ; next we extend this result to construct a model for the union ST of the sets Sp ; and
then we show that the utility function U(x) can be written as a Lebesgue integral (1).
Next, we extend the model in Theorem 1 to a larger set of outcome-streams.

Definition 2. An outcome-stream x = ( x1, , xN ) defined on [0, T] will be called a


Riemann outcome-stream provided that each function xj , j = 1, , N, is Riemann
integrable on [0, T ] and has upper and lower bounds that are in the interval Xj .
The set of Riemann outcome-streams on P = [0, T ] will be denoted by RT .

An outcome-stream x = ( x1, , xN ) is a Riemann outcome-stream if and only if


for each j = 1, , N the function xj is continuous almost everywhere and has bounds
in Xj . Hence, as previously noted the set of Riemann outcome-streams seems sufficient
for describing most phenonema.

Theorem 2. An outcome-stream space (RT , ), T > 0, satisfies conditions (A), (B) on


the set RT , satisfies condition (C) on the pair of sets RT , ST , and satisfies conditions
(D), (E) on the set ST if and only if it has a utility function of the form,

U(x) = 0T a(t)u( x (t)) dt , x in RT , (2)

such that the Lebesgue integral (2) exists for any x in RT and the functions u(x), a(t),
A(t), and U(x) have the properties (a)(d) in Theorem 1.
12

Moreover, the function u(x) is unique up to a positive affine transformation, and the
function A(t) is unique up to a positive multiple.

The proof of Theorem 2 relies on the fact that a real-valued function (t) on [0, T ] is
Riemann integrable if and only if it is Darboux integrable, that is, there exists increasing
(n) (n)
and decreasing sequences of step functions s , s (n) , n = 1, 2, , such that s (t)
(n)
(t) s (n) (t) , t in [0, T ], and the distances | s (n) s | tend to zero as n tends to
infinity. We show that for any Riemann outcome-stream x there exist Darboux sequences
of step outcome-streams for the functions xj associated with x, and we use this squeeze
property of Riemann outcome-streams to extend the integral utility model in Theorem 1
for step outcome-streams to an integral utility model for Riemann outcome-streams.

6. Models for an unbounded planning period

This section presents two models for outcome-streams defined on the planning period
P = [0, ). First, we present a model for the set of Riemann outcome-streams that equal
a null outcome beyond a time which depends on the outcome-stream. Then, we extend
this result to present a model for a larger set of outcome-streams. In this model, the set of
outcome-streams that are comparable, that is, the set of outcome-streams on which the
preference relation is complete, is specified in terms of the preference relation.
The models in Harvey (1986, 1995) for outcome-sequencesand models in Wakker
(1993) for discrete probability distributionsalso consider a preference relation defined
on a specified set and assume completeness of the preference relation only on a subset
that depends on the relation on the given set. Harvey argues that this type of dependence
permits an arbitrary sequence of discount weights, and Wakker argues that it is the
crucial change in the axioms of Savage (1954) that permits an unbounded utility function.

Definition 3. An outcome-stream defined on the planning period P = [0, ) will be


called a finite-horizon outcome-stream provided that for some time T > 0 it has the form
( x[0, T], ) where x[0, T] is a Riemann outcome-stream on [0, T ]. Here, is the null
13

outcome identified in Assumption 1. The set of finite-horizon outcome-streams will be


denoted by R .

An outcome-stream x in a set RT , T > 0, will be identified with the outcome-stream


( x[0, T], ) in the set R . Thus, R is the union of the sets RT , T > 0. Moreover, x
will be identified with the outcome-stream ( x[0, T], (T, T] ) in a set RT , T > T, and
thus RT is a subset of RT for T > T.

Theorem 3. An outcome-stream space (R , ) satisfies conditions (A), (B) on each set


RT , T > 0, satisfies condition (C) on each pair RT , ST , T > 0, and satisfies conditions
(D), (E) on each set ST , T > 0, if and only if (R , ) has a utility function of the form,

U(x) = limT 0T a(t)u( x (t)) dt , x in R , (3)

such that the improper Lebesgue integral (3) exists for any x in R and:
(a) The function u(x) defined on the outcome set X is continuous, has a non-point
interval range, is a utility function for the outcome space (X , X ), and u() = 0.
(b) The function a(t) defined on the interval [0, ) is non-negative and is Lebesgue
integrable on each interval [0, T ], T > 0.
(c) The function A(t) = 0t a(s) ds defined on [0, ) is strictly increasing on [0, ),
is absolutely continuous on each interval [0, T ], T > 0, and has the value A(0) = 0.
(d) For each T > 0, the function U(x) is continuous on ST , that is, for any x in ST
and any > 0 there exists a > 0 such that | z x | < implies | U(z) U(x) | < for
any z in ST .
Moreover, each of the functions u(x) and A(t) is unique up to a positive multiple.

Theorem 3 can be extended to a model for a set of outcome-streams for which the
improper integral (3) converges. To do so, we define a set of outcome-streams on [0, ),
and in terms of a preference relation on this set we define a smaller set and construct a
model for the preference relation restricted to this smaller set.
14

Definition 4. An outcome-stream x on [0, ) will be called a Riemann outcome-stream


on [0, ) provided that for any horizon T > 0 the restriction of x to [0, T] is in RT .
The set of Riemann outcome-streams on [0, ) will be denoted by R .

Suppose that a preference relation is defined on a set R . We do not assume that


is complete on R; instead, we will define a subset of R in terms of and assume
that is complete on the subset. Roughly speaking, the subset is to contain the outcome-
streams in R that become arbitrarily unimportant in the sufficiently distant future.
To make this idea precise, consider tradeoffs between the immediate future period
[0, 1] and an unbounded future period (t, ), t 1. Then, for an outcome-stream x we
can compare changes between two outcomes in the period [0, 1] with changes between
x and the null outcome-stream in the period (t, ).

Definition 5. An outcome-stream x in R will be called comparable provided that for


any outcomes a pX b pX c there exists a horizon T 1 such that for any t T :

( a[0, 1], ) p ( b[0, 1], , x(t, ) ) p ( a[0, 1], ) for any t T .

The set of comparable outcome-streams will be denoted by Rc .

Lemma 2. Suppose that a pair (R , ) is an outcome-stream space. Then:


(a) Any finite-horizon outcome-stream is comparable.
(b) For any two outcome-streams x, y in R , if x is comparable and there exists a
horizon T > 0 such that x(t) = y(t) for all t > T, then y is comparable.

Below, we present an integral utility model for an outcome-stream space (Rc , ). In


this model, the improper integral (4) converges for any outcome-stream in the subset Rc
of R . One may ask whether, conversely, any outcome-stream in R such that the
improper integral (4) converges is in the subset Rc . This statement is true if and only if
satisfies an additional condition; see Harvey (1998b).

Theorem 4. An outcome-stream space (R , ) satisfies conditions (A), (B), and (D)


on the set Rc , satisfies condition (C) on each pair of sets RT , ST , T > 0, and satisfies
15

condition (E) on each set ST , T > 0, if and only if the outcome-stream space (Rc , )
has a utility function of the form,

U(x) = limT 0T a(t)u( x (t)) dt , x in Rc , (4)

such that the improper Lebesgue integral (4) exists for any x in Rc and the functions
u(x), a(t), A(t), and U(x) have the properties (a)(d) in Theorem 3.
Moreover, each of the functions u(x) and A(t) is unique up to a positive multiple.
Appendix: Proofs of Results

Proof of Lemma 1. To show part (a), consider a pair of outcomes a, b. Then, a X b


if and only if (a[0,], ) (b[0,], ) by definition. Thus, condition (B) implies that X

is transitive and complete. To show that X is continuous, suppose that a fX b and thus
(a[0,], ) f (b[0,], ). For an outcome-stream (c[0,], ) of the same form, we have,
for example, | (c[0,], ) (a[0,], ) | = | c a |. Thus, condition (C) on the continuity
of implies that X is continuous.
N
To show parts (b), (c) note that since X is a connected subset of , the properties
(a) of X imply by a result of Debreu (1954) that X has a continuous utility function.
The connectedness of X implies that the range any such function is an interval, and
Assumption 2 implies that the range is not a point.
Part (d) is implied by the following stronger result.

Lemma A1. If an outcome-stream space (C, ) satisfies conditions (A), (B), then is
independent in the sense that for any non-point interval , , any outcomes a, b , and
any outcome-stream x in C :

a X b if and only if (a,, x) (b,, x) .

Proof. Assume that a X b. Then, (a,, x)(t) X (b,, x)(t) for t in P, and thus
(a,, x) (b,, x) by condition (A,a). Next, assume that a X b is false. Then,
b X a by the completeness of X which implies (b,, x) (a,, x) by the above
argument which implies that (a,, x) f (b,, x) is false.
16

Next, we obtain three preliminary results for Theorem 1.

Lemma A2. If an outcome-stream space (ST , ) satisfies conditions (A), (B), then:
(i) For any non-point interval , in [0, T ] , there exist x, y in ST such that x(t)
= y(t) for t not in , but x f y. (In this sense, any non-point interval is essential.)
(ii) For any point interval , in [0, T ] and any x, y in ST : if x(t) = y(t) for t
not in , , then x ~ y. (In this sense, any point interval is inessential.)
Proof. To show part (i), note that since X is complete, Assumption 2 implies that there
exist outcomes a fX b. Then, condition (A,b) implies that (a,, ) f (b,, ).
For part (ii), suppose that x, y are in ST with x(t) = y(t), t not in , . Then,
x(t) = y(t) a.e. and thus x(t) ~X y(t) a.e. which implies that x ~ y by condition (A,a).

Lemma A3. If an outcome-stream space (ST , ) satisfies conditions (A)-(C), then for
any non-point, disjoint intervals , , , in [0, T ] and any outcomes a pX b,
+ +
there exist outcomes a , b such that: a fX a, b pX b, and:
+
(a ,, a, , ) p (a,, b, , ) p (b,, b , , ).

+
Proof. We show the existence of an outcome a as described. The argument for the

existence of b is similar.
0 +
Define A = { x X: x pX a }, A = { x X: x ~X a }, and A = { x X: x fX a }.
These sets are pairwise disjoint, and since X is complete their union is X. Moreover,
0 +
the sets A and A are nonempty. Since X is continuous by Lemma 1(a), the sets A
+ + 0 + +
and A are open. Thus, the closure of A is a subset of A A . But A is not closed
0 0 +
since X is connected, and thus there exist an a in A that is in the closure of A .
+ + + 0
Hence, there exists a sequence of outcomes an in A such that | an a | tends
+
to zero as n tends to infinity. It follows that for any > 0 there exists an outcome a
+ 0 + 0
such that, | (a ,, a, , ) (a ,, a, , ) | = ( ) | a a | < .
0 0
Since a ~X a pX b, condition (A) implies that (a ,, a, , ) ~ (a,,
a, , ) p (a,, b, , ). By condition (C), it follows that there is a > 0
17

+ 0 +
such that: | (a ,, a, , ) (a ,, a, , ) | < implies (a ,, a,
0
, ) p (a ,, b, , ) ~ (a,, b, , ).

Lemma A4. Suppose that an outcome-stream space (ST , ) satisfies conditions (A)(E).
Suppose, moreover, that p: ai1, ai , i = 1, , m, is a partition of [0, T ] with at least
three non-point intervals. Then, the subspace (Sp , ) has a utility function of the form,

m
Up(x) = i=1 ai, p u p (x(i)) , x in Sp , (A1)

such that:
(a) The function up(x) defined on X is continuous and has a non-point interval range.
Moreover, it is a utility function for the outcome space (X , X ).

(b) A coefficient ai,p is positive if the interval ai1, ai is non-point and is zero
otherwise.
Moreover, the function up(x) is unique up to a positive affine transformation and the
coefficients ai,p are unique up to a common positive multiple.

Proof. Outcome-streams in Sp have the form, x(t) = x(i), t in ai1, ai, i = 1, , m.


m
Thus, there is a one-to-one correspondence between Sp and the set X = X X of
(m)
outcome vectors x = ( x(1), , x(m) ). Therefore, the preference relation on Sp
m m m m
induces a preference relation on X , and the space (X , ) can be identified
(m) (m)
with the space (Sp , ). We will define the distance between two vectors x , y in
m
X as the distance | x y | between the corresponding outcome-streams x, y in Sp .
Assume that (ST , ) satisfies conditions (A)(E). The conditions (B)(D) imply that
m
the relation is transitive, complete, continuous, and preferentially independent.
m
Lemma A2 implies that a component set X in X is essential if it corresponds to a
non-point interval in the partition p and is inessential if it corresponds to a point interval
m
in p. Thus, X has at least three essential components. Lemma A1 implies that is
independent and induces the relation X on each essential component and induces a
relation on each inessential component such that any two outcomes as indifferent.
18

m m
By a result of Debreu (1960), it follows that the space (X , ) has an additive
m (m)
utility function, U (x ) = iE ai ui(x(i)), where E denotes the set of indices of the
essential components, ui(x) is a continuous utility function for the preference relation

X on the i-th component, and each coefficient ai is positive. Lemma 1 implies that
each function ui(x) has a non-point interval range. Moreover, the functions ui(x) are
unique up to a common positive linear transformation and the coefficients ai are unique
up to a positive multiple.
m
Condition (E) implies that satisfies the condition of equal tradeoffs midvalues
defined in Harvey (1986). By use of an argument there, the functions ui(x) can be chosen
as a common function, which we denote by u(x). Defining ai = 0 for each inessential
m m
component, it follows that the vector space (X , ) has a utility function of the form,
m (m) m
U (x ) = i=1 ai, p u p (x(i)) , where the function u(x) and the coefficients ai have the
properties (a), (b) and the stated uniqueness properties. In particular, each coefficient for
an inessential component must be zero since the function u(x) is not constant, and thus
the coefficients ai are unique up to a positive multiple.
m m
By the identification of (X , ) and (ST , ), it follows that the function Up(x)
in (A1) with up(x) = u(x) and ai,p = ai is a utility function for (ST , ) and that it has
the stated properties. Thus, the proof of Lemma A4 is complete.

Suppose that p: ai1, ai, i = 1, , m, and q: bj1, bj, j = 1, , n, denote two


partitions of an interval [0, T ]. Since the intersections ai1, ai bj1, bj are pairwise
disjoint, they form another partition of [0, T ]. We will denote this partition by pq and
call it the conjunction of p and q. The sets Sp and Sp are subsets of the set Spq since
an outcome-stream that is in Sp or Sq is constant on each interval ai1, ai bj1, bj.
A partition with at least three non-point intervals will be called proper. The conjunction
of a proper partition and any partition is proper. Thus, the set ST is the union of the sets
Sp such that p is a proper partition.
19

Proof of Theorem 1. To show the forward implications, we consider utility functions


Up(x) in Lemma A4 whose functions up(x) and coefficients ai,p have been normalized,
and we paste the functions Up(x) together to construct a utility function of the form (1).
Consider any proper partition p. By Lemma A4(b), there exists coefficients ai,p > 0,
(1) (1)
and by Lemma 1 there exist outcomes a fa . Assume that the function up(x) and
(1)
the coefficients ai,p in a utility function (A1) are normalized such that up(a ) = 1,
(1)
up(a ) = 1, and i ai,p = 1. By Lemma A4, the function up(x) and the coefficients
ai,p are unique. Hence, the utility function Up(x) with these components is unique.
Suppose that Up(x) = i ai,p up(x(i)), Uq(x) = j aj,q uq(x(j)), and Upq(x) = i j
aij,pq upq(x(i, j)) are utility functions for two proper partitions p, q and their conjunction
pq. Assume that these utility functions have the properties (a)(d) in Lemma A4 and are
normalized in the sense defined above.
Since the sets Sp and Sq are subsets of the set Spq , the function Upq(x) is a utility
function for Sp and for Sq . Moreover, Upq(x) = i (j aij,pq) upq(x(i)) for x in Sp ,
and Upq(x) = j (i aij,pq) upq(x(j)) for x in Sq .
Consider the functions up(x), etc. Both up(x) and upq(x) are normalized functions
for the partition p, and thus up(x) = upq(x) for any x in X. And by a similar argument,
uq(x) = upq(x) for any x in X. Thus, up(x) = uq(x), x in X. Hence, the function u(x)
= up(x) is independent of the proper partition p.
Next, consider the coefficients ai,p , etc. Both ai,p and j aij,pq are normalized
coefficients for the partition p, and thus ai,p = j aij,pq for any i = 1, m. And by a
similar argument, aj,q = i aij,pq for any j = 1, n.
If two intervals, ah1, ah and bk1, bk, in p and q have equal endpoints, that
is, ah1 = bk1 and ah = bk , then ah,p = ak,q . For suppose that ah1 = bk1 and ah
= bk . Then, the interval ah1, ah bk1, bk in the conjunction pq also has these
endpoints. Thus, the intervals ai1, ai bk1, bk, i h, and ah1, ah bj1, bj,
j k, are point intervals, and hence aik,pq = 0, i h and ahj,pq = 0, j k. It follows
that ah,p = j ahj,pq = ahk,pq and ak,q = i aik,pq = ahk,pq , and thus ah,p = ak,q .
20

Hence, a coefficient ai,p is a function, ai,p = (ai1, ai), of the endpoints ai1, ai
of the interval ai1, ai. But what type of function?
Suppose that p has adjacent intervals ah1, ah, ah, ah+1, and q has an interval
bk1, bk = ah1, ah ah, ah+1. Then, ah,p = j ahj,pq = ahk,pq , ah+1,p = j ahj,pq

= ah+1,k,pq , and ak,q = i aik,pq = ahk,pq + ah+1,k,pq . Hence, ak,q = ah,p + ah+1,p ,
and thus (bk1, bk) = (ah1, ah+1) = (ah1, ah) + (ah, ah+1).
The endpoints are not restricted, and thus (a, c) = (a, b) + (b, c) for any a b c
in [0, T ]. To solve this simple functional equation, define A(t) = (0, t), t in [0, T ].
Then, (b, c) = A(c) A(b) and A(0) = (0, 0) = A(0) A(0) = 0.
The normalized utility function Up(x) for a set Sp can now be written as:
m
Up(x) = i=1( A(ai ) A(ai1 ) )u(x(i)) , x in Sp ,

where the functions u(x) and A(t) are independent of the proper partition p.
Both Up(x) and Upq(x) are normalized utility functions for Sp , and thus Up(x) =
Upq(x) for x in Sp . And for a similar reason, Uq(x) = Upq(x) for x in Sq . If an
outcome-stream x is in the sets Sp and Sq for two proper partitions p and q, then
Up(x) = Uq(x). Thus, for any x in ST , the amount U(x) in (2) is well-defined as the
common amount Up(x) for any proper partition p such that x is in Sp .
Moreover, the function U(x) is a utility function for ST . For consider any x, y in
ST . Then, x is in Sp and y is in Sq for some proper partitions p and q. Hence, x
and y are both in Spq , and thus U(x) = Upq(x) and U(y) = Upq(y). Therefore, x y
if and only if Upq(x) Upq(y) if and only if U(x) U(y).
The normalizations, i ai,p = 1, imply that A(T) = 1. Hence, U(a) = u(a) for any
outcome a. Moreover, an amount U(x) is a weighted average of amounts u(x(i)). Thus,
the range of the function U(x) equals the non-point interval range of the function u(x).
(1) (1)
The normalizations, up(a ) = 1, up(a ) = 1, imply that the common range of the
functions u(x) and U(x) includes the interval [1, 1]. Thus, for any amount 1 r 1
(r) (r) (r)
there exists an outcome a such that u(a ) = r and U(a ) = r .
21

Next, we show that the functions u(x), A(t), and U(x) have the properties (a)(c).
Lemma A4 implies that u(x) has the properties in (a) since u(x) = up(x) for any proper
partition p. Lemma A4 also shows that A(t) is strictly increasing on [0, T ] since any
coefficient ai,p for a nonpoint interval is positive. Moreover, A(0) = 0 as shown above.
To show that A(t) is absolutely continuous on [0, T ] it suffices to show that for any
n n
0 < < 1 there is a > 0 such that i=1(ai ai1 ) < implies i=1( A(ai ) A(ai1 ) )
< for any n and any pairwise disjoint intervals (ai1 , ai), i = 1, , n, in [0, T ].
(1)
For intervals (ai1 , ai) as described, define a step outcome-stream z by z(t) = a
(0) (0)
if t is in the union of the intervals (ai1 , ai) and z(t) = a otherwise. Then, | z a |
n (1) (0) n
= i=1(ai ai1 ) | a a | and U(z) = i=1( A(ai ) A(ai1 ) ) .
() (0) ()
Consider an outcome-stream a , 0 < < 1. Then, a p a . Condition (C) implies
(0) ()
that there exists a > 0 such that | z a | < implies z p a for z in ST . Define
(1) (0) 1 n (0) ()
= | a a | . Then, i=1(ai ai1 ) < implies | z a | < implies z p a
() n
implies U(z) < U(a ) implies i=1( A(ai ) A(ai1 ) ) < .
To show property (d), consider an x in ST and an > 0. As the primary case,
+ +
assume that there exist x , x in ST with U(x ) < U(x) <U(x ). Since the function
() ()
U(x) has an interval range, U(x) < U(x ) < U(x) < U(x ) < U(x) + for some
( ) ()
x , x in ST . Now, by condition (C) there exist 1 , 2 > 0 such that | z x | < 1
() ()
implies x p z, and | z x | < 2 implies z p x . Define = min{1 , 2}. Then,
() ()
| z x | < implies x p z p x implies U(x) < U(z) < U(x) + .
As a second case, assume that U(z) U(x) for any z in ST . Then, there exists an
( ) ()
x in ST such that U(x) < U(x ) < U(x). Now, by condition (C) there exists a
( )
> 0 such that | z x | < implies x p z for z in ST . Thus, | z x | < implies

( )
U(x ) < U(z) implies U(x) < U(z). The arguments are similar for the remaining
case that U(z) U(x) for any z in ST .
For the converse part of the proof, assume that an outcome-stream space (ST , ) has
a utility function U(x) of the form (1) with properties (a)(d). Then, it is straightforward
to show that (ST , ) satisfies conditions (A), (B), (D), and (E).
22

To show that (ST , ) satisfies the continuity condition (C), consider any x, y in ST
with x p y and thus U(x) < U(y). Define = U(y) U(x). By property (d), there exists
a > 0 such that | z x | < implies | U(z) U(x) | < for z in ST . The inequality
| U(z) U(x) | < implies U(z) < U(y) which implies z p y. By a similar argument, there
is a > 0 such that | z y | < implies z f x. Hence, condition (C) is satisfied.
It remains to show the uniqueness properties of the functions u(x) and A(t). Suppose
that U1(x) = P a1(t) u1(x(t)) dt and U2(x) = P a2(t) u2(x(t)) dt are utility functions for
(ST , ) with the properties (a)(d). Then, for any proper partition p, U1(x) = i ( A1(ai)
A1(ai1) ) u1(x(i)) and U2(x) = i ( A2(ai) A2(ai1) ) u2(x(i)) are utility functions for
the subset Sp of ST .
Lemma A4 implies that u2(x) = p u1(x) + p , x in X, where p > 0. Since u1(x)
has a non-point range, p , p are independent of p, and thus u2(x) is a positive linear
transformation of u1(x). Lemma A4 also implies that A2(ai) A2(ai1) = p ( A1(ai)
A1(ai1) ), i = 1, , m, where p > 0. Hence, A2(T) = p A1(T) by addition. Since
A1(T) and A2(T) are positive, p is independent of p. Then, A2(a1) = A2(a1) A2(a0)
= ( A1(a1) A1(a0) ) = A1(a1) where > 0 denotes the common value of p . But a1
can be any time in [0, T ), and thus A2(t) is a positive multiple of A1(t).
Conversely, if U1(x) = P a1(t) u1(x(t)) dt is a utility function for ST , and U2(x) =
P a2(t) u2(x(t)) dt is a function such that u2(x) = u1(x) + , > 0, and A2(t) = A1(t),
> 0, then U2(x) = P ( a1(t)) ( u1(x(t)) + ) dt is also a utility function for ST .

A sequence p(n), n = 1, 2, , of partitions of an interval P = [0, T] will be called


dissecting provided that: (i) For n = 1, 2, , each subinterval in p(n+1) is a subset of a
(n) (n)
subinterval in p(n), and (ii) limn maxi | ai ai1 | = 0. Dissecting sequences
can be constructed, e.g., by a process of bisection starting with the interval [0, T].
For two outcomes x, y, suppose that [x, y] denotes the product set of outcomes z
such that xj zj yj for j = 1, , N. Any non-empty set [x, y] is a connected, compact
subset of the set X of outcomes.
23

Lemma A5. Suppose that p(n), n = 1, 2, , is a dissecting sequence of partitions of


an interval P = [0, T] , that u(x) is a continuous function on a set X of outcomes, and
that x is a Riemann outcome-stream on [0, T]. Then:
(n)
(a) For n = 1, 2, , there exist step outcome-streams x , x (n) in Sp(n) such that
for each j = 1, , N:
(1) (2)
(i) xj (t) xj (t) xj(t) xj ( 2 ) (t) xj (1) (t) , t in [0, T].
(n)
(ii) limn | xj (n) xj | = 0.
(b) The function u(x(t)) is Riemann integrable on [0, T], and it has upper and lower
bounds that are in the range of u(x). For n = 1, 2, , there exist step outcome-streams
(n) (n) (n) (n)
y , z in Sp(n) and hence step functions u(y ), u(z ) such that:
(n) (n) (n) (n)
(i) xj (t) yj (t) xj (n) (t) and xj (t) zj (t) xj (n) (t) , t in [0, T], for
j = 1, , N.
(n) (n)
(ii) u(y (t)) u(x(t)) u(z (t)), t in [0, T].
(n) (n)
(iii) u(y (t)) u(s(t)) u(z (t)), t in [0, T], for any step outcome-stream s in
(n)
Sp(n) such that xj (t) sj(t) xj (n) (t) , t in [0, T], j = 1, , N.
(1) (2) (2) (1)
(iv) u(y (t)) u(y (t)) u(x(t)) u(z (t)) u(z (t)), t in [0, T].

Proof. To show part (a), note that for each j = 1, , N the function xj(t) is Riemann
integrable and has lower and upper bounds that are in the interval Xj .
(n) (n) (n)
For each partition p(n), define xj (i) = inf{ xj(t) : ai1 t ai } and xj (n) (i) =
(n) (n) (n) (n) (n)
sup{ xj(t) : ai1 t ai }, j = 1, , N. And define x (i) = (x1 (i), , xN (i) )
(n)
and x (n) (i) = ( x1(n) (i) , , x N (n) (i) ). Then, x (i) and x (n) (i) are in X. Define the
(n) (n) (n)
step outcome-streams x , x (n) in Sp(n) by: xj (t) = xj (i) and xj (n) (t) = xj (n) (i) for
(n) (n) (n)
t in ai1 , ai . Then, xj and xj (n) are lower and upper step functions for xj(t).
Since xj(t) is Riemann integrable, and hence Darboux integrable, the condition (ii)
of dissecting implies (ii). Also, the condition (i) of dissecting implies (i).
(n) (n)
To show part (b), note that for n = 1, 2, and for t in ai1 , ai , the vector
(n) (n)
x(t) is in the product set [ x (i), x (n) (i) ] in X. Since [ x (i), x (n) (i) ] is connected
and compact, the continuous function u(x) maps it onto a compact interval in the range
24

(n) (n)
of u(x). We will denote this image, u([x (i), x (n) (i) ]), by [u (i), u (n) (i) ]. Then,
(n) (n) (n)
u (i) u(x(t)) u (n) (i) for t in ai1 , ai .
(n) (n)
Since the numbers u (i) and u (n) (i) are in the image u([x (i), x (n) (i) ]), we have
(n) (n) (n) (n) (n)
u (i) = u(y (i)) and u (n) (i) = u(z (i)) for some outcomes y (i) and z (i) in the
(n) (n) (n) (n) (n)
set [x (i), x (n) (i) ]. Then, xj (i) yj (i) xj (n) (i) and xj (i) zj (i) xj (n) (i) .
(n) (n) (n) (n) (n) (n)
Now, define y (t) = y (i), z (t) = z (i) for t in ai1 , ai . Then, the step
(n) (n)
outcome-streams y and z are in Sp(n), and they satisfy the inequalities (i). The
(n) (n)
definitions of y and z imply the inequalities (ii) and (iii). And the definitions of
(n) (n)
y ,z and the inequalities (i), (ii) imply the inequalities (iv).

Lemma A6. Assume the situation described in Lemma A5. Suppose, moreover, that a(t)
is a non-negative, Lebesgue integrable function on P = [0, T] whose indefinite integral
A(t) = 0t a(s) ds is strictly increasing. Then:
(n) (n)
(a) The function a(t) u(x(t)) and the associated functions a(t) u(y (t)), a(t) u(z (t)),
n = 1, 2, , are Lebesgue integrable P = [0, T]. Moreover:
(n) (n)
(i) P a(t) u(y (t)) dt P a(t) u(x(t)) dt P a(t) u(z (t)) dt .
(n) (n)
(ii) P a(t) u(y (t)) dt P a(t) u(s(t)) dt P a(t) u(y (t)) dt for any outcome-
(n)
stream s in Sp(n) such that xj (t) sj(t) xj (n) (t) , t in [0, T], j = 1, , N.
(n) (n)
(iii) limn P a(t) u(y (t)) dt = limn P a(t) u(z (t)) dt = P a(t) u(x(t)) dt .
(c) For any x, y in RT , if u(x(t)) < u(y(t)) on a set of positive measure, then for any
> 0 there exists a non-point interval , in [0, T] and outcomes a, b such that:
(i) u(x(t)) u(a) < u(b) u(y(t)) for t in , .
(ii) , a(t) u(x(t)) dt A u(a) < A u(b) , a(t) u(y(t)) dt.
(iii) A u(b) , a(t) u(x(t)) dt < and , a(t) u(y(t)) dt A u(a) < .
Here, A = A() A() > 0 and, e.g., , a(t) u(b) dt = A u(b).

(n) (n)
Proof. Since yj , zj , and xj , j = 1, , N , are Riemann integrable on [0, T] and
(n) (n)
u(x) is continuous on X , the functions u(y (t)), u(z (t)), and u(x(t)) are Riemann
integrable on [0, T]. But a(t) is Lebesgue integrable on [0, T] , and thus the functions
(n) (n)
a(t) u(y (t)), a(t) u(z (t)), and a(t) u(x(t)) are Lebesgue integrable on [0, T] .
25

To show part (i) of (a), note that the inequalities in Lemma A5(b,ii) imply those in
part (i) since a(t) 0, t in [0, T].
(n) (n)
To show part (ii) of (a), define j (t) = xj (n) (t) xj (t), t in [0, T]. The functions
(n) (1) (2)
j (t) are non-negative step functions. Lemma A5(a,i) implies that j (t) j (t)
(n)
for t in [0, T] . Thus the sequence j (t), n = 1, 2, , converges for t in [0, T] .
(n)
Define j(t) = limn j (t), t in [0, T] . By the Monotone Convergence Theorem,
(n) (n)
j(t) is Lebesgue integrable and P j(t) dt = limn P j (t) dt. But P j (t) dt =
(n) (n)
| xj (n) xj |, and thus limn P j (t) dt = 0 by Lemma A5(a,ii).
To show that j(t) = 0 a.e. on [0, T] , define E = { t [0, T] : j(t) > 0 } and Em =
{ t [0, T] : j(t) > 1/m }, m = 1, 2, . Then, E = m Em . If the measure (E) of the
union is positive, then the measure (Em) of Em is positive for some m = 1, 2, .
But (Em) > 0 implies that P j(t) dt (1/m) (Em) > 0 which is a contradiction.
(n) (n)
Since xj (t) xj (t) xj (n) (t) , t in [0, T], it follows that limn xj (n) (t) =
(n) (n)
limn xj (t) = xj(t) a.e. on [0, T]. Lemma A5(b,i) now implies that: limn yj (t)
(n)
= limn zj (t) = xj(t) a.e. on [0, T] for j = 1, , N .
(n) (n)
Thus, limn u(y (t)) = limn u(z (t)) = u(x(t)) a.e. since u(x) is continuous.
(n) (n)
Hence, limn a(t) u(y (t)) = limn a(t) u(z (t)) = a(t) u(x(t)) a.e. on [0, T].
(n)
Lemma A5(b,iii) states in particular that the sequence, u(z (t)), n = 1, 2, , is
(n)
weakly decreasing for t in [0, T] . Hence, the sequence, a(t) u(z (t)), n = 1, 2, , is
weakly decreasing for t in [0, T] . The Monotone Convergence Theorem now implies
(n)
that, limn P a(t) u(z (t)) dt = P a(t) u(x(t)) dt . And a similar argument can be used
(n)
to show that, limn P a(t) u(y (t)) dt = P a(t) u(x(t)) dt .
To show part (c), assume that u(x(t)) < u(y(t)) on a set of positive measure in [0, T] .
The function (t) = u(y(t)) u(y(t)) is continuous a.e., and thus there exists a time t0
such that (t) is continuous at t0 and (t0) > 0. Define = (t0)/2. Then, there exists
a non-point interval , such that (t) > for t in , . Define u1 = inf{u(x(t)) :
t in , }, u2 = sup{u(x(t)) : t in , }, u3 = inf{u(y(t)) : t in , }, and u4 =
sup{u(y(t)) : t in , }. Then, u1 u2 u3 u4 where u3 u2 . By an argument
26

used for Lemma A5(b), the numbers u1, , u4 are in the range of u(x). Thus, there
exist outcomes a1, , a4 in X such that u1 = u(a1), etc. Hence, part (i) is established.
To show part (ii), we construct a sufficiently small subinterval , of , . But
, a(t) u(a3) dt , a(t) u(x(t)) dt (A() A())(u(a3) u(a1)) for any subinterval
, , and a similar inequality is true for y. The function A(t) is strictly increasing and

continuous, and thus we can choose < such that the inequalities in (ii) are true.

Proof of Theorem 2. For the forward part of the proof, assume that an outcome-stream
space (RT , ) satisfies the stated conditions. Then, restricted to the set ST satisfies
the conditions in Theorem 1. Thus, there exist functions u(x), a(t), A(t), U(x) with the
properties (a)-(d) in Theorem 1 such that U(x) is a utility function for the space (ST , ).
Moreover, the function u(x) is unique up to a positive linear transformation, and the
function A(t) is unique up to a positive multiple.
By Lemma A6(a), the function a(t) u(x(t)) is Lebesgue integrable for any x in RT
and thus U(x) = P a(t) u(x(t)) dt is well-defined on RT . Our task is to show that U(x)
is a utility function for the space (RT , ). Roughly speaking, the key step is to show that
any Riemann outcome-stream x can be approximated by step outcome-streams that are
simultaneously near to x in preference and near to x in utility.

(i) For any x in RT and any > 0, there exists a w in ST such that w ~ x and
| U(w) U(x) | < .

Proof: By Lemmas A5(b,iii) and A6(a,ii), there exist step outcome-streams y, z in a set
Sp(n) such that:
(1) u(y(t)) u(x(t)) u(z(t)), t in [0, T ].
(2) | U(y) U(x) | < and | U(z) U(x) | < .
By Theorem 1, u(x) is a utility function for X, and thus the inequalities (1) imply
that y(t) X x(t) X z(t), t in [0, T ]. Hence, y x z by condition (A). If x ~ y or
x ~ z, then by the inequalities (2) we are through.
In the remaining case, y p x and x p z, and thus y p z (by the completeness and
transitivity of ). In this case, we will construct a w between y and z that satisfies (i).
27

Define x = z + (1 ) y for 0 1. Then, each outcome-stream x is in the


(n)
set Sp(n) , and xj (t) (x)j(t) xj (n) (t) , t in [0, T], j = 1, , N, by Lemma
A5(b,i). Hence, U(y) U(x) U(z) by Lemma A6(a,i), and thus | U(x) U(x) | <
by (2).
One can check that x x = ( ) ( z y ) for 0 , 1. Thus, | x x | =
| | | z y |. Also, y p z implies that | z y | > 0 since y, z are in the set Sp(n) .
Define L = { : x p x } and U = { : x f x }. Then, L and U are disjoint, 0
is in L, and 1 is in U. Moreover, the sets L and U are open relative to [0, 1]. For
consider, e.g., an in L. Then, x p x, and thus by condition (C), there is a > 0
1
such that | x x | < implies x p x. Hence, | | < ( | z y |) implies that
x is in L. Since [0, 1] is connected, it follows that there exists a in [0, 1] that is
not in L or U. Then, x ~ x by the completeness of .

(ii) U(x) < U(y) implies x p y for any x, y in RT .


U(x) = U(y) implies x ~ y for any x, y in RT .

Proof. Assume that U(x) < U(y). Define = (U(y) U(x)). By (i), there exist w, z
in ST with w ~ x, z ~ y and | U(w) U(x) | < , | U(z) U(y) | < . Hence, U(w) <
U(z), and thus w p z since U(x) is a utility function for (ST , ). Therefore, x p y by
the transitivity of .
Assume that U(x) = U(y). By property (i), there exist w, z in ST such that w ~ x
and z ~ y. Then, U(w) = U(x) and U(z) = U(y) since otherwise, e.g., U(w) = U(x)
which implies w p x by (ii). Thus, U(w) = U(z). Since U(x) is a utility function for
(ST , ), this equality implies that w ~ z. Therefore, x ~ y by the transitivity of .

Property (ii) implies that U(x) is a utility function for (RT , ). For x y implies
not x p y which implies U(x) U(y) by (ii), and U(x) U(y) implies U(x) > U(y) or
U(x) = U(y) which implies x f y or x ~ y by (ii) which implies x y.
For the converse part of the proof, assume that a function, U(x) = P a(t) u(x(t)) dt, is
well-defined and is a utility function for an outcome-stream space (RT , ), and that the
functions u(x), a(t), A(t), and U(x) satisfy the properties (a)(d) in Theorem 1. Then:
28

satisfies condition (B) on RT since U(x) is a utility function on RT , satisfies


conditions (D) and (E) on ST by Theorem 1, and satisfies condition (A) on RT by
using property (a) in Theorem 1 and Lemma A6(c).
To show that satisfies condition (C) on the sets RT , ST , consider any x in RT
and w in ST such that w p x. Then, U(w) < U(x). Define = U(x) U(w) > 0. By
property (c) in Theorem 1, there exists a > 0 such that | z w | < implies |U(z)
U(w)| < for any z in ST . But |U(z) U(w)| < implies U(z) < U(x) implies z p x.
The argument when w f x is similar, and thus condition (C) is satisfied.
To prove the uniqueness properties of u(x) and A(t), consider two functions U1(x) =
P a1(t) u1(x(t)) dt, U2(x) = P a2(t) u2(x(t)) dt, and the associated functions, A1(t), A2(t).
If U1(x) and U2(x) are utility functions for (RT , ) which satisfy properties (a)(d)
in Theorem 1, then by Theorem 1 u2(x) is a positive linear transformation of u1(x) and
A2(t) is a positive multiple of A1(t). The converse argument is straightforward.

Proof of Theorem 3. Since RT is a subset of RT for T > T, R is the union of the


sets RT , T H, for any horizon H < . For our purposes, we choose H = 1.
First, we show the forward implications. Lemma 1 implies that there exists an outcome
+
x fX or an outcome x pX . The arguments are the same in both cases, so it suffices
+
to assume that there is an outcome x fX .
The assumptions in Theorem 3 are those in Theorem 2 for any horizon T > 0. Thus,
for any T 1 there exist a utility function UT(x) = [0,T] aT(t) uT(x(t)) dt for (RT , ) as
described in Theorem 2. Moreover, we can assume that the functions uT(x), AT(t) are
+
normalized such that uT() = 0, uT(x ) = 1, AT(0) = 0, AT(1) = 1, and thus uT(x) and
AT(t) are unique. Then, for any T> T 1, UT(x) and UT (x) are normalized utility
functions for RT , and thus uT(x) = uT (x), x in X, and AT(t) = AT (t), 0 t T.
Hence, the following functions are well-defined: the function u(x) = uT(x), x in X,
and the function A(t), 0 t < , defined by A(t) = AT(t) for 0 t T. We define a
function a(t), 0 t < , by a(t) = A(t) if the derivative A(t) exists and a(t) = 0
otherwise. Hence, a(t) = A(t) = AT(t) = aT(t) a.e. for 0 t T, and thus UT(x) =
[0,T] aT(t) uT(x(t)) dt = [0,T] a(t) u(x(t)) dt for x in RT . Finally, we define a function
29

U(x), x in R, by U(x) = UT(x) for x in RT . Since u() = 0, this definition implies


that, U(x) = limT UT(x) = limT [0,T] a(t) u(x(t)) dt, for any x in R.
Theorem 2 implies that U(x) is a utility function for any set RT , T 1, and that the
functions u(x), A(t), a(t), U(x) have properties (a)(d). Moreover, U(x) is a utility
function for (R , ). For if x, y are in R , then x is in RT and y is in RT for
some T, T 1. Define H = max{T, T}. Then, x and y are in RH and thus can be
compared by the normalized function UH(x). Hence, x y if and only if UH(x)
UH(y) if and only if U(x) U(y) since U(x) = UH(x) and U(y) = UH(y).
To show the converse implications, assume that there exist functions u(x), A(t), a(t),
and U(x) as described in Theorem 3. Then, Theorem 2 implies that for any T 0 the
preference relation satisfies conditions (A)(E) with regard to the sets RT and ST .
To show that each function A(t), u(x) is unique up to a positive multiple, consider
two utility functions U1(x) = limT [0,T] a1(t) u1(x(t)) dt and U2(x) = limT [0,T] a2(t)
u2(x(t)) dt as described in Theorem 3. Then, u1() = u2() = 0 and A1(0) = A2(0) = 0.
Since U1(x), U2(x) are utility functions for any set RT , T > 0, Theorem 2 implies that
there exist constants T > 0, T , T > 0 such that u2(x) = Tu1(x) + T , x in X , and
A2(t) = TA1(t), 0 t T . But T = 0 since u1() = u2() = 0. For T T: T u1(x) =
+
u2(x) = Tu1(x), x in X and T A1(t) = A2(t) = TA1(t), 0 t T. But u(x ) 0 and
A(T) 0, and thus T = T and T = T . Hence, T and T are independent of T > 0.

Proof of Lemma 2. For part (a), consider an x in R . Then, there is a T 1 such that
x(t,) = (t,) for t T. But a pX b pX c implies (a[0,1], ) pX (b[0,1], ) pX (c[0,1], )
by the definition of X and Lemma 1(d), and thus x satisfies Definition 5.
For part (b), consider x, y in R such that x is in Rc and there is a time T > 0
with x(t) = y(t), t > T. Then, x(t,) = y(t,) , t T, and thus y satisfies Definition 5.

Lemma A7. Suppose that an outcome-stream space (R , ) satisfies the conditions in


Theorem 4. Then, for any outcome-stream x in Rc , any outcomes a pX b pX c, and
any non-point, bounded interval , , there exists a time T such that:

(a,, ) p (b,, , x(t,)) p (c,, ), t T


30

Proof. We will show that for any non-point, bounded, disjoint intervals , and , ,
if an outcome-stream in Rc satisfies the comparability condition with respect to , ,
then it satisfies it with respect to , . This result implies the lemma because in the case
that , and , are not disjoint we can introduce a third interval that is disjoint
from , and from , and then use the result twice.
We will consider only the left-hand strict preferences since the argument for the right-
hand strict preferences is similar.
Suppose that x is an outcome-stream in Rc , a pX b are outcomes, and , , ,
are non-point, bounded, disjoint intervals such that x satisfies the comparability condition

with respect to , . Then, in particular for any outcome b pX b there exists a T

such that: (i) (b ,, ) p (b,, , x(t,)) for t T. Lemma A3 implies that there

exists such an outcome b such that: (ii) (b,, a,, ) p (b ,, b,, ).

Since (b ,, ) and (b,, , x(t,)) are in R by Lemma 2, condition (D) and

(i) imply: (iii) (b ,, b,, ) p (b,, b,, , x(t,)), t max{T, , }. Then,
(ii) and (iii) imply (b,, a,, ) p (b,, b,, , x(t,)) by transitivity, and thus
condition (D) implies (a,, ) p (b,, , x(t,)), t max{T, , }.

Proof of Theorem 4. First, we show the forward implications. Since R is a subset of


Rc , the space (R , ) satisfies the conditions of Theorem 3, and thus there exists a
utility function U(x) = limT [0,T] a(t) u(x(t)) dt for R as described in Theorem 3.
To show that U(x) converges for any x in Rc , it suffices to show that for any > 0
there exists a time T > 0 such that | U((x[0,t], )) U((x[0,s], )) | < for s, t T.
For x in Rc and > 0, choose outcomes a pX b pX c with A(1)(u(c) u(a)) < .
There exists a T 1 such that (a[0,1], ) p (b[0,1], , x(s,)) and (b[0,1], , x(t,)) p
(c[0,1], ) for s, t T. Therefore, (a[0,1], x(1,s], ) p (b[0,1], x(1,)) and (b[0,1], x(1,))
p (c[0,1], x(1,t], ) by condition (D). Hence, (a[0,1], x(1,s], ) p (c[0,1], x(1,t], ), and thus
U(a[0,1], x(1,s], ) < U(c[0,1], x(1,t], ) since U(x) is a utility function for R . Hence,
A(1)u(a) + U((x[0,s], )) < A(1)u(c) + U((x[0,t], )), and thus U((x[0,s], )) U((x[0,t], ))
< A(1)(u(c) u(a)) < for s, t T. Then, also U((x[0,t], )) U((x[0,s], )) for s, t T.
31

Next, we show that U(x) is a utility function for Rc . First, suppose that an x in Rc
is upper extremal, i.e., x y for y in Rc. Then, U(x) U(y) for y in Rc. For assume
that U(y) > U(x). Then, U((y[0,T], )) U((x[0,T], )) for some T > 0, and thus u(y(t))
> u(x(t)) on a nonpoint interval , in [0, T] by Lemma A6(c). Thus, y(t) fX x(t)
for t in , . But (y[0,T], x) is in Rc by Lemma 2, and thus (y,, x) f x by
condition (A). The argument is similar for an outcome-stream x that is lower extremal.
+
Next, suppose that an x in Rc is non-extremal, i.e., there exist y , y in Rc with y
+
p x p y . The argument for (i) below is similar to that for (i) in Theorem 3.
(i) For any non-extremal outcome-stream x in Rc and any > 0, there exist a time
T > 0 and an outcome-stream w in ST such that w ~ x and |U(w) U(x)| < .
To prove (i), it suffices to show that for any > 0 there exists a T > 0 and outcome-
+ + +
streams w , w in ST such that w x w and |U(w ) U(x)| < , |U(w ) U(x)| <
. For we can then use the proof of (i) in Theorem 3 to obtain a step outcome-stream of
+
the form, w = w + (1) w , 0 1, that satisfies (i).
+
To show the existence of a step outcome-stream w that has the properties described
above in terms of a given > 0, we will construct a sequence of outcome-streams leading
+
from x to w . The argument to show the existence of w is similar.
+ +
By condition (A), x p y implies that x(t) pX y (t) on a set of positive measure in
+
[0, ). Thus, there exists a time T1 > 0 such that x(t) pX y (t) on a set E [0, T1] of
+
positive measure. Then, u(x(t)) < u(y (t)) on E. Lemma A6(c) implies that there exists
a non-point interval , in [0, T1] and outcomes a, b such that: (1) u(x(t)) u(a) <
u(b) for t in , , and (2) , a(t) u(b) dt , a(t) u(x(t)) dt < /3.
By condition (A), the inequalities (1) imply that x (a,, x). And by Lemma A7
they imply that there exists a T2 such that (a,, , x(t,)) p (b,, ), t T2,
that is, (a,, [0,t],, x(t,)) p (b,, ), t T2. Condition (D) then implies that
(a,, x) p (x[0,t],, b,, ), t T2. Thus, x p (x[0,t],, b,, ), t T2.
By adding U(x[0,t],) to both terms in (2), we have |U((x[0,t],, b,, ))
U((x[0,t], ))| < /3 for t T1.
32

Next, note that since the integral U(x) converges there exists a T3 > 0 such that
|U(x) U((x[0,t], ))| < /3, t T3.
Finally, define T = max{T1, T2, T3}. By property (i) in the proof of Theorem 3, there
+ + +
exists an outcome-stream w in ST such that w ~ (x[0,T],, b,, ) and |U(w )
U((x[0,T],, b,, ))| < /3.
+ +
To conclude, x p (x[0,T],, b,, ) ~ w implies x p w by transitivity. And
+
|U(w ) U(x)| < by adding the above three inequalities.
The argument for (ii) below is the same as that for (ii) in Theorem 3. Moreover, the
argument that (ii) suffices to show that U(x) is a utility function for the non-extremal
outcome-streams in (Rc , ) is the same as in the proof of Theorem 3.
(ii) U(x) < U(y) implies x p y for any non-extremal x, y in Rc .
U(x) = U(y) implies x ~ y for any non-extremal x, y in Rc .
To show the converse implications, assume that a space (R , ) as described has a
utility function U(x) as described on the subspace (Rc , ). Then, it is straightforward
to verify that satisfies conditions (A), (B), and (D) on the set Rc . By Theorem 3,
satisfies condition (E) on each set ST, T > 0, and one can use an argument similar to that
in Theorem 3 to show that satisfies condition (C) on each pair of sets Rc , ST , T > 0.
By Theorem 3, each of the functions A(t), u(x) is unique up to a positive multiple.
References

Debreu, G. (1954). Representation of a Preference Ordering by a Numerical Function. In


Decision Processes. R. Thrall, C.H. Coombs and R. Davies (eds.), Wiley, New York.
Debreu, G. (1960). Topological Methods in Cardinal Utility Theory. In Mathematical
Methods in the Social Sciences, K.J. Arrow, S. Karlin, and P. Suppes (eds.), Stanford
University Press, Stanford, California.
Diamond, P.A. (1965). The Evaluation of Infinite Utility Streams. Econometrica 33, 170-
177.
Fishburn, P.C. (1970). Utility Theory for Decision Making. Wiley, New York.
Gorman, W.M. (1968). The Structure of Utility Functions. Review of Economic Studies
35, 367-390.
33

Grodal, B. (2003). A General Integral Representation. In Independence, Additivity,


Uncertainty, K. Vind, Springer-Verlag, Berlin.
Grodal, B. and J.-F. Mertens (1968). Integral Representation of Utility Functions.
Technical report CORE 6823, CORE.
Harvey, C.M. (1986). Value Functions for Infinite-Period Planning. Management Science
32, 1123-1139.
Harvey, C.M. (1995). Proportional Discounting of Future Costs and Benefits.
Mathematics of Operations Research 20, 381-399.
Harvey, C.M. (1998a). Value-Functions for Continuous-Time Consequences. Working
paper. Department of Decision and Information Sciences, University of Houston.
Harvey, C.M. (1998b). Proofs of Results in Value-Functions for Continuous-Time
Consequences. Working paper. Department of Decision and Information Sciences,
University of Houston.
Harvey, C.M. and L.P. sterdal (2005). Preferences Between Continuous Streams of
Events. Working paper. Department of Economics, University of Copenhagen.
Koopmans, T.C. (1960). Stationary Ordinal Utility and Impatience. Econometrica 28,
287-309.
Koopmans, T.C. (1972). Representation of Preference Orderings Over Time. In Decision
and Organization, C.B. McGuire and R. Radner (eds.), North Holland, Amsterdam.
Koopmans, T.C., P.A. Diamond and R.E. Williamson (1964). Stationary Utility and Time
Perspective. Econometrica 32, 82-100.
Krantz, D.H., R.D. Luce, P. Suppes and A. Tversky (1972). Foundations of Measure-
ment, Vol 1. Academic Press, New York.
Savage, L.J. (1954). The Foundations of Statistics. Wiley, New York (2nd ed. 1972,
Dover, New York).
Wakker, P. (1993). Unbounded Utility for Savages Foundations of Statistics, and
Other Models. Mathematics of Operations Research 18, 446-485.
Weibull, J.W. (1985). Discounted-Value Representations of Temporal Preferences.
Mathematics of Operations Research 10, 244-250.
34

Williams, A.C. and J.I. Nassar (1966). Financial Measurement of Capital Investments.
Management Science 12, 851-864.

You might also like