A theory of Conscientiousness∗

S. Cerreia-Vioglioa , Fabio Maccheronia , Massimo Marinaccia , Aldo Rustichinib

a
Department of Decision Sciences, Università Bocconi
b
Department of Economics, University of Minnesota

December 30, 2011

Abstract

We provide an axiomatic foundation for a personality trait which has important implications
for economic behavior, Conscientiousness, and two aspects of that factor, the inhibitive and the
proactive. We refer to these two aspects here with the names, probably more intuitive for
economists, of control and motivation. The first aspect is commonly associated in analysis of
individual behavior with the ability to override impulses and distractions when pursuing a goal.
The second is usually associated with the inclination to set ambitious goals.
Our setup and analysis closely follow those of standard decision theoretic analysis. In our
model an individual is characterized by a preference order over acts, which are maps from states
to lotteries over prizes. In the framework of Drèze, we allow the possibility that the individual
can affect the probability of the state which is realized, at some cost. The differences in this cost
of control make formal the differences in conscientiousness among individuals: a higher cost of
control over the probability corresponds to a lower degree of inhibitive side of conscientiousness.
The utility in each state deriving from the realization of an outcome is state dependent. An
important part of the research reported here is an axiomatic foundation of preferences with
moral hazard and state dependent preferences, first treated in Drèze. This utility evaluated
by the individual in reference to a subjective benchmark, or aspiration level. An extreme and
simple example is given by an individual who sets an aspiration level, which is a point in a
partially ordered space, and derives a positive utility when his outcome is larger than the set
level, and does not when the outcome is lower. The level of the aspiration set by the individual
corresponds to his motivation, which corresponds to the proactive side of Conscientiousness.

Very preliminary version !

∗
The authors gratefully acknowledge the financial support of the European Research Council (Advanced Grant,
BRSCDP-TEA) and of the National Science Foundation (grant SES 0924896).

1
1 Introduction
Personality theory and decision theory Personality theory and decision theory are two distinct
theories of human behavior. Although very different in historical origin and conceptual structure,
both try to identify stable characteristics of individuals, able to predict patterns of behavior of that
individual in different environments and circumstances. Personality theory wants to identify traits
of an individual, defined as relatively stable patterns of affect, behavior and cognition in a wide
variety of circumstances. Factors are, as in the statistical process of determining them, a limited
number of latent variables are able to explain a large number of possible patterns. If one accepts
this definition, then decision theory can be considered a theory of personality with a restricted
domain, because it also tried to identify and provide the conceptual structure for stable patterns
of behavior in choices involving dated stream of rewards under uncertainty. If one takes it as a
search for factors, one may define Decision Theory as the hypothesis that a large set of possible
different patterns of choice on the domain of such rewards can be explained by two simple factors,
the attitude to risk and that to delay.
So far these two theories have been largely independent. An integration of the two theories
seems necessary because several findings in the past years have shown that the predictive power of
classical decision theory is significantly increased once we introduce some measurement of personalty
traits of individuals, for example the Big Five factors. A first step in this integration is setting the
analysis of personality traits on the same formal level of analysis as classical decision theory, and
the axiomatic method seems the natural approach. The purpose of this paper is to provide such a
foundation for one of the factors, Conscientiousness, and its two aspects, motivation and control.
Why Conscientiousness? There are several reasons to out this trait at the forefront of the
analysis. First, Conscientiousness has been found to be important in explaining successful outcomes
for several economic variables. In the influential meta-analysis papers of Barrick and Mount ([2])
and Barrick et. al ([3]), Conscientiousness was found to be a good predictor of job performance. In
the survey paper of Roberts et al. ([24]), several longitudinal studies are examined, reporting the
comparative effect of IQ, Socioeconomic Status (SES) and personality traits at a later time. The
dependent variables considered are mortality, divorce, educational and occupational attainment.
For each, Conscientiousness has important and significant effects, comparable to those of IQ and
SES. The effects on Mortality are larger than those of other traits. For career success Judge et
al. ([16]) find that Conscientiousness has the largest standardized coefficient among the Big Five
factors. In the study [25], Personality traits and economic preferences compared side by side, and
the correlation with economic and social outcomes are are estimated for each. A higher score in
Conscientiousness is correlated in larger ability to hold a job, to reduce number of accidents in
drivers, and with Credit Score.
Second, Conscientiousness is a very specific human trait, in that it is based on the prominence
of long-term goals in directing purposeful behavior which is typically absent in other species. For

2
example in a review of 19 studies on 12 different mammalian and non-mammalian species Gosling
and John ([14]) reported factors similar to Extraversion, Agreeableness and Neuroticism in most
species, with Openness less represented than these are other three. A factor similar to Conscien-
tiousness has been reported only in chimpanzees, in the study by King and Figueredo [19]. They
asked 53 raters to rate 100 chimpanzees at 12 zoological parks. Raters had to choose a score on a 7
points scale for 43 adjectives, which were first briefly described. For example the definition of cau-
tious stated that Subject often seems attentive to possible harm or danger from its actions. Subject
avoids risky behaviors. Factor analysis of the ratings show the existence of six latent factors, five
of them close to the Big Five factors in human personality literature, including Conscientiousness.
The sixth factor was related to Dominance.
Third, Conscientiousness is among the Big Five personality traits the one that is more remote
form the conceptual structure of economic analysis, in spite of its importance for the analysis of
economic behavior. It is harder to model within standard economic and decision theoretic tra-
dition. This is not the case for the other four traits. For example the two traits of Neuroticism
and Extraversion may be mapped into the different attitudes to gains and losses as conceptual-
ized in Prospect Theory ([17]. Similarly, Intelligence can be modeled as a productive skill, and
Openness/Intellect can be modeled in a dynamic model under uncertainty as the ability to process
information. And Agreeableness as the inclination of an individual to consider into his utility the
material payoff of others.
Conscientiousness and its facets Conscientiousness “appears to reflect the tendency to
maintain motivational stability within the individual, to make plans and carry them out in an
organized and industrious manner” ([11]. In their systematization of the NEO Personality Inventory
(NEO-PI) Costa McCrae and Dye ([8]) conceptualize Conscientiousness as “having both proactive
and inhibitive aspects.” Since a good understanding of these two sides is crucial for the rest of
our model, and the conclusions, we will elaborate. Each of the Big Five factors is characterized by
six facets. In the case Conscientiousness, the six facets proposed in [8] are Order, Dutifulness, Self
Discipline, Achievement Striving, Competence, Deliberation. The first three pertain to the aspect
that Costa et alii call the inhibitive side of Conscientiousness, seen “in moral scrupulousness and
cautiousness” . The other three pertain to the proactive side, “seen most clearly in the need for
achievement and commitment to work.”
The two aspects have been proposed several times in further analysis of the Conscientiousness
factor. Hough ([18] and Mount and Barrick ([22] propose achievement and dependability as sub-
domains. Similarly, Roberts et al. [23] and (with slightly different terms) De Young et al. ([10])
suggest industriousness and orderliness. These two sides, that we are going to call here with the
terms of motivation and control, are the focus of our analysis.
Conscientiousness and Moral Hazard We approach the modeling of Conscientiousness from
a very familiar direction in economic analysis: moral hazard. The amount of care taken in avoiding
accidents is a simple and effective index of conscientiousness. So a model where individuals can

3
affect the probability of outcomes is a very natural choice. The setup we adopt is a classical decision
theoretic setup, that of Anscombe-Aumann ([1]): Individuals have preferences over acts, which are
maps from states to probability distributions over prizes. As in Drèze ([13]) the probability overs
states is at least partially under control of the individual, who can affect it at some utility cost. More
formally, let S be a finite state space, and Z a finite set of sure outcomes. Denote by X = ∆ (Z)
the set of all probability measures on Z, called random outcomes. The set F = X S of all maps
from S to X, is the set of acts; the lottery at state s induced by a generic act is denoted by f (·, s).
A binary relation % on F representing the agent’s choices, called preference.
The general representation we want to characterize is:
 
X
V (f ) = max  p(s)U (z − φ(s))f (z, s) − c(p) (1)
p∈∆(S)
S,Z

where we interpret φ as the function setting the aspiration level at state s and c is the cost to
control the occurrence of a state. In the representation (1) the utility of the outcome z in state s
is measured with reference to a state dependent benchmark represented by φ(s).
We will be in particular interested in the case in which the benchmark is an aspiration level,
so that a positive utility is reached when and only when the aspiration level is reached, and the
level of utility is otherwise independent of the outcome. For each f ∈ F and each s ∈ S, the image
f (·, s) of s under f is a probability measure on denoted by

f (A |s ) ∀A ⊆ Z.

Let for every state s the aspiration level be denoted by As . The evaluation of an act in this case is:
" #
X
V (f ) = max p(s)f (As |s ) − c(p) (2)
p∈∆(S)
s∈S

Outline In the rest of the paper, we will first define precisely in section 2 the environment we
consider. The form of our general representation 1 clearly suggests that an important intermediate
step in the process is the model of moral hazard with state dependent utilities and moral hazard.
This is the the main technical task of the paper, which we develop in section 3. Proofs are gathered
in section 5.

2 Setup and Basic definitions

The setup we have described earlier is called a genuine Anscombe-Aumann framework. It is often
convenient to consider the generalization of this framework obtained by replacing ∆ (Z) with an
arbitrary convex set X of some arbitrary vector space, we refer to the obtained framework as a (gen-
eralized) Anscombe-Aumann framework. The latter more general framework is used in most formal
statements and proofs, while we will focus our attention on the former in terms of interpretation.

4
 As usual the symbols  and ∼ denote respectively the asymmetric and symmetric parts of %.
If f, g ∈ F and s ∈ S, then fs g ∈ F is the act yielding f (s) in state s and g (s0 ) for all s0 6= s.
Note that clearly gs g = g. If f, g ∈ F and α ∈ [0, 1], f αg is a shorthand for αf + (1 − α) g. The
constant act xS ∈ F yielding x ∈ X in every state is simply denoted by x.
For each s ∈ S, %s is the binary relation defined on F by

f %s g ⇐⇒ fs h % gs h ∀h ∈ F (3)

it is called conditional preference;

2.1 Moral hazard

The variational state dependent preferences we are after are characterized by a functional V of the
form " #
X
V (f ) = max p(s)U (f (·, s), s) − c(p) (4)
p∈∆(S)
S

Let us consider the two elements separately. The utility function U : Z × S → R is a real valued
function, extended to X × S by linearity, that is
X
u (x, s) = x (z) u (z, s) ∀ (x, s) ∈ X × S (5)
z∈Z

The cost function c : ∆ (S) → [0, ∞] is a lower semi continuous, convex, and grounded
(minp∈∆(S) c (p) = 0) function.
This representation can be considered as the extension of the moral hazard model of Drèze [13]
to the case of a general cost function instead of the cost function which is the indicator of a closed
and convex set of probabilities. In fact, for c (p) = δC (p), (4) takes the form:

X
V (f ) = max p(s)U (f (·, s), s) (6)
p∈C
S

where C is a closed and convex subset of ∆(S).

2.2 Aspiration

For the model of aspiration, we assume in what follows that Z is a subset of a partially ordered
linear space W . An important special case is the one in which Z is a Riesz space and

u (z, s) = v (z − φ (s))

where v : Z → R is an increasing function and φ : S → Z represents the the agent’s (state

dependent) goal. Actually, we aim at the very special case

v = 1W +

5
where Z + is the positive cone of Z. In this case,
(
1 z ≥ φ (s)
u (z, s) = 1Z + (z − φ (s)) = = 1φ(s)+W + (z) (7)
0 otherwise

if f (s) is a degenerate lottery then

(
1 f (s) ≥ φ (s)
u (f (s) , s) =
0 otherwise

else
X X
u (f (s) , s) = f (z |s ) u (z, s) = f (z |s )
z∈Z z∈Z:z≥φ(s)

= f ({z ∈ Z : z ≥ φ (s)} |s ) = f φ (s) + Z + |s

= Pr {surpassing the goal in state s induced by the choice of f } .

If p ∈ ∆ (S) is the model that generates the occurrence of states,

Z X
f φ (s) + W + |s p (s)


u(f (s), s)dp(s) =
S s∈S

= Pr {surpassing the goal by the choice of f under p} .

A natural extension of (7) that does not require a Riesz space structure, but just a lattice
structure, is (
1 z ∈ As
u (z, s) = 1As (z) =
0 otherwise
where As is the aspiration level (set) in state S, that is any proper subset of Z, that is any set A
such that
a ∈ A, z ∈ A, and z ≥ a implies z ∈ A.

Since f (s) is, in general a lottery then

X X
u (f (s) , s) = f (z |s ) u (z, s) = f (z |s ) = f (As |s )
z∈Z z∈As

= Pr {achieving aspirations in state s induced by the choice of f } .

If p ∈ ∆ (S) is the model that generates the occurrence of states,

Z X
u(f (s), s)dp(s) = f (As |s ) p (s)
S s∈S

= Pr {achieving aspirations by the choice of f under p} .

Several different representations are possible. For example:

6
 1. The agent has no control over the probability: this s the expected utility model, where p = p0
is given
Z
V (f ) = f (As |s ) dp0 (s)
S

= Pr {achieving aspirations by the choice of f }

2. The agent has partial (resp. full) control over the probability: he can chose p, but he is con-
strained to the elements of C ⊂ ∆ (S) (resp. C = ∆ (S))
Z
V (f ) = max f (As |s ) dp(s)
p∈C S

= max Pr {achieving aspirations by the choice of f under p}

p∈C

3. The agent has costly control over the probability: he can chose p at a cost c (p)
Z 
V (f ) = max f (As |s ) dp(s) − c(p) . (8)
p∈∆(S) S

2.3 Strategy of proof

Let V be defined for all as in (4) for all f ∈ F. For each s ∈ S denote by us the s-section of u and
by Us the s-section of the extension of u defined in (5), that is

Us : X → R
P P
x 7→ u (x, s) = x (z) u (z, s) = x (z) us (z)
z∈Z z∈Z

for all x ∈ X. All functions Us : X → R, for s ∈ S, are affine, and so is their cartesian product
U : X S → RS defined by

U : XS → RS
f = [fs ]s∈S 7→ U (f ) = [Us (fs )]s∈S

for all f ∈ X S = F, that associates to every f = (fs )s∈S ∈ X S the vector f u = (us (f (s)))s∈S ∈ RS .
It is easy to show that (·)u is affine and takes values in Πs∈S us (X) = Φ.
Notice that Φ is a convex subset of RS and the map I : Φ → R defined by

I(ϕ) = max (hϕ, pi − c(p))

p∈∆(S)

is a convex niveloid (see the section 7 below), and we can then set the cost function to be:

c (p) = sup (hϕ, pi − I (ϕ)) = I ∗ (p) ∀p ∈ ∆ (S) .

ϕ∈Φ

7
3 Moral Hazard and State Dependent Utility

3.1 Axioms

In the sequel, we make use of the following properties of %.

Axiom A. 1 (Weak Order) The binary relation % is complete and transitive.

Axiom A. 2 (Continuity) For each f, g, h ∈ F, the sets {α ∈ [0, 1] : αf + (1 − α) g % h} and

{α ∈ [0, 1] : h % αf + (1 − α) g} are closed.

Axiom A. 3 (Conditional Preference) For each s ∈ S, if f, g, h ∈ F, then

fs g % g =⇒ fs h % gs h.

Axiom A. 4 (Conditional Independence) For each s ∈ S, if f, g, h ∈ F, then

1 1 1 1
f ∼s g =⇒ f + h ∼s g + h.
2 2 2 2

Axiom A. 5 (Conditional Nontriviality) For each s ∈ S, %s is nontrivial.

Axiom A. 6 (Value of Information) If f, g ∈ F and α ∈ [0, 1], then

f ∼ g =⇒ f % αf + (1 − α) g.

Axiom A. 7 (Best-worst Independence) If f, g, b, w ∈ F, α ∈ [0, 1], and

αf + (1 − α) (bλw) % αg + (1 − α) (bλw)

holds for some λ ∈ [0, 1], then it holds for all λ ∈ [0, 1], provided b %s h %s w for all s ∈ S and all
h ∈ F.

3.2 Representation Theorems

Let K be any non-empty subset of RS . A function I : K → R is:

1. normalized if I (t1S ) = t for all t ∈ R such that t1S ∈ X;

2. strictly increasing if I (ϕ) > I (ψ) for all ϕ, ψ ∈ K such that ϕ ≥ ψ and ϕ 6= ψ;

3. a niveloid if I (ϕ) − I (ψ) ≤ maxs∈S (ϕ (s) − ψ (s)) for all ϕ, ψ ∈ K.1

1
See Dolecki and Greco [11] and our [20].

8
Theorem 1 In a genuine Anscombe-Aumann framework, a binary relation % on F satisfies A.1-
A.5 if and only if there exist U : X × S → [0, 1], affine and onto with respect to the first component,
and a normalized, strictly increasing, and continuous function I : [0, 1]S → [0, 1] such that

V (f ) = I (U (f (·) , ·)) ∀f ∈ F

represents %.
Moreover, in this case,

1. f ∼ bV (f ) w for all f ∈ F, provided b %s h %s w for all s ∈ S and h ∈ F.

2. % satisfies A.6 if and only if I is quasiconvex.

3. % satisfies A.7 if and only if I is a niveloid.

Both U and I are unique.

Using the general results on Niveloids presented in the section 7 below, we can now conclude
the general representation theorem for variational preferences with state dependent utility. This
theorem is an extension of the characterization given by Drèze of state dependent preferences with
moral hazard, where the moral hazard had the form of the choice out of an element of a convex set
admissible probabilities.

Theorem 2 In a genuine Anscombe-Aumann framework, a binary relation % on F satisfies A.1-

A.7 if and only if there exist U : X × S → [0, 1], affine and onto with respect to the first component,
and a cost function c such that
X
V (f ) = max [ p(s)U (f (·, s), s) − c(p)]
p∈∆(S)
s

represents %. Moreover, in this case, the U function is unique and there is a unique maximal cost
function c.

4 Aspiration and Control

We now introduce the axiom that gives to the state dependent preferences the

Axiom A. 8 (Aspiration Sets) For each s ∈ S there is a set As ⊆ Z such that, for every pair
0
z, z ∈ Z:
0 0
1. z  z if z ∈ As , z ∈
/ As ,
0 0 0
2. z ∼ z if z, z ∈ As , or z, z ∈
/ As

The next theorem summarizes the results we have and provides the final representation including
both aspiration and control:

9
Theorem 3 A binary relation % on F satisfies A.1- A.8 and if and only if there exists a cost
function c such that
X
V (f ) = max [ p(s)1As (z)f (z, s) − c(p)] (9)
p∈∆(S)
s∈S,z∈Z

represents %.

5 Proofs
Denote by f, g, h generic elements of F, by x, y generic elements of X, by α, λ, µ generic elements
of [0, 1], by ϕ, ψ generic elements of [0, 1]S .
The constant element xS ∈ X S yielding x ∈ X in every state is simply denoted by x. Anal-
ogously, 1S and 0S in RS are sometimes denoted by 1 and 0, respectively. While 1s ∈ RS is the
indicator of state s, for each s ∈ S.

Lemma 1 If % satisfies A.1-A.3, then, for each s ∈ S, the binary relation %s satisfies A.1-A.2,
and

f %s g ⇐⇒ fs g % g ⇐⇒ ∃h ∈ F : fs h % gs h
f %s g ⇐⇒ f (s) %s g (s)
f s g ⇐⇒ fs g  g ⇐⇒ ∃h ∈ F : fs h  gs h ⇐⇒ fs h  gs h ∀h ∈ F
f s g ⇐⇒ f (s) s g (s) .

Moreover, if f, g ∈ F and f %s g for all s ∈ S then f % g, and f  g if additionally f s̄ g for

some s̄ ∈ S.

Proof. Arbitrarily choose s ∈ S.

If f %s g, then, by definition (3), fs g % gs g = g, and so fs h % gs h for some h ∈ F. Conversely,
if
fs h̄ % gs h̄ for some h̄ ∈ F (10)

set ḡ = gs h̄; then fs h̄ = fs ḡ, and (10) amounts to fs ḡ % ḡ, by A.3 this means fs h % ḡs h = gs h for
all h ∈ F. This proves the chain of characterizations of %s .
Notice that
fs h = f (s)s h and gs h = g (s)s h

for all f, g, h ∈ F. Then, if f, g ∈ F,

f %s g ⇐⇒ fs h % gs h ∀h ∈ F ⇐⇒ f (s)s h % g (s)s h ∀h ∈ F ⇐⇒ f (s) %s g (s) .

For each f, g ∈ F, choose h̄ ∈ F, by completeness of %, either fs h̄ % gs h̄ or gs h̄ % fs h̄, that is,

either f %s g or g %s f , and %s is complete. Transitivity of %s immediately follows by definition
(3) and transitivity of %. That is, %s satisfies A.1.

10
 As a consequence of completeness of %s ,

f s g ⇐⇒ g 6%s f
⇐⇒ q (gs h % fs h ∀h ∈ F)
⇐⇒ ∃h ∈ F : gs h 6% fs h
⇐⇒ ∃h ∈ F : fs h  gs h

moreover, using the equivalent definitions of %s ,

g 6%s f ⇐⇒ q (∃h ∈ F : gs h % fs h)
⇐⇒ gs h 6% fs h ∀h ∈ F
⇐⇒ fs h  gs h ∀h ∈ F.

Summing up,
f s g ⇐⇒ ∃h ∈ F : fs h  gs h ⇐⇒ fs h  gs h ∀h ∈ F

and so fs g  g is also equivalent to f s g. This proves the chain of characterizations of s (which,

by definition, is the asymmetric part of %s ).
The final equivalence for s follows from

f s g ⇐⇒ g 6%s f ⇐⇒ g (s) 6%s f (s) ⇐⇒ f (s) s g (s) .

For each f, g, h ∈ F, choose h̄ ∈ F, the set


{α ∈ [0, 1] : αf + (1 − α) g %s h} = α ∈ [0, 1] : (αf + (1 − α) g)s h̄ % hs h̄

= α ∈ [0, 1] : αfs h̄ + (1 − α) gs h̄ % hs h̄

is closed and so is {α ∈ [0, 1] : h %s αf + (1 − α) g}, that is, %s satisfies A.2.

Finally, assume f, g ∈ F and f %s g for all s ∈ S. Set S = {s1 , s2 , ..., sn } and recall that
fsi h % gsi h for all h ∈ F and i = 1, 2, ..., n, that is,
   
h (s1 ) h (s1 )
 ..   .. 

 .  
  . 

   
 h (si−1 )   h (si−1 ) 
   
 f (si )  %  g (si )  .
   
   
 h (s )   h (si+1 ) 
 i+1   
 .
..
  .. 
.
   
   
h (sn ) h (sn )

11
Moreover, if f sj g for some j = 1, 2, ..., n, then fsj h  gsj h for all h ∈ F. Therefore
         
f (s1 ) g (s1 ) g (s1 ) g (s1 ) g (s1 )
         
 f (s2 )   f (s2 )   g (s2 )
   g (s2 )   g (s2 ) 
       
 f (s3 )   f (s3 )   f (s3 )  g (s3 )   g (s3 ) 
         

% %  % ... %  %
.. .. .. .. ..
 
         
 .   .   .   .   . 
         
 f (s )   f (sn−1 )   f (sn−1 )
    g (s )   g (sn−1 ) 
 n−1     n−1   
f (sn ) f (sn ) f (sn ) f (sn ) g (sn )

and transitivity delivers f % g. Moreover, if f sj g for some j = 1, 2, ..., n, then

   
g (s1 ) g (s1 )
 ..   .. 

 .  
  . 

   
 g (sj−1 )   g (sj−1 ) 
   
 f (sj )    g (sj ) 
   
   
 f (s )   f (sj+1 ) 
 j+1   
 ..   .. 
. .
   
   
f (sn ) f (sn )

and f  g. 

Lemma 2 If % satisfies A.1-A.4, then for each s ∈ S, there exists an affine function Us : X → R
such that
f %s g ⇐⇒ Us (f (s)) ≥ Us (g (s)) .

Moreover,

1. if X = ∆ (Z) for some finite set Z, then % satisfies A.9;

2. if A.9 is satisfied, then Us (X) is a compact interval for all s ∈ S.

Proof. The result follows from Hernstein and Milnor [15] and Lemma 1. 

Lemma 3 If Us : X → R is affine for all s ∈ S, then the function

v : F → RS

defined for each s ∈ S by

vs (f ) = Us (f (s))

is affine.

Proof. Trivial. 

12
Axiom A. 9 (Conditional Boundedness) For each s ∈ S, %s admits a maximum and a mini-
mum element.

Proposition 1 A binary relation % on F satisfies A.1-A.5 and A.9 if and only if there exist
U : X × S → [0, 1], affine and onto with respect to the first component, and a normalized, strictly
increasing, and continuous function I : [0, 1]S → [0, 1] such that

V (f ) = I (U (f (·) , ·)) ∀f ∈ F (11)

represents %.
Moreover, in this case,

1. f ∼ bV (f ) w for all f ∈ F, provided b, w ∈ F are such that b %s h %s w for all s ∈ S and

h ∈ F.

2. % satisfies A.6 if and only if I is quasiconvex.

3. % satisfies A.7 if and only if I is a niveloid.

4. Us : X → [0, 1] represents %s on X for all s ∈ S.

Both U and I are unique.

Proof of Proposition 1 and Theorem 1. We first assume % on F satisfies A.1-A.5 and either
X = ∆ (Z) for some finite set Z or A.9 is satisfied.
By Lemma 2, for each s ∈ S, there exists an affine function Us : X → R such that

f %s g ⇐⇒ Us (f (s)) ≥ Us (g (s)) . (12)

For each s ∈ S, Us (X) is a compact interval [rs , ts ] ⊆ R and A.5 implies that rs < ts . The
assumption rs = 0 and ts = 1 causes no loss in generality and makes Us the only affine function for
which (12) holds.
In particular, the function v : F → [0, 1]S defined in Lemma 3 is affine and such that

f %s g ⇐⇒ vs (f ) ≥ vs (g) . (13)

The proof now proceeds through several steps.

Step 1. v is onto and b, w ∈ F are such that

b %s h %s w ∀s ∈ S, h ∈ F

if and only if v (b) = 1S and v (w) = 0S .

Proof of the Step. Let ϕ ∈ [0, 1]S , for each s ∈ S, there exists xs ∈ X such that Us (xs ) = ϕs
because Us : X → [0, 1] is onto. Set

f (s) = xs ∀s ∈ S

13
to obtain vs (f ) = Us (f (s)) = ϕs for all s ∈ S, that is v (f ) = ϕ. Therefore v is onto.
In particular, there exist b, w ∈ F such that

v (b) = 1S and v (w) = 0S . (14)

If b, w ∈ F satisfy (14), for each s ∈ S and each h ∈ F,

vs (b) = 1 ≥ vs (h) ≥ 0 = vs (w)

therefore, by (13)
b %s h %s w. (15)

Conversely, if b̄, w̄ ∈ F satisfy b̄ %s h %s w̄ for all s ∈ S and h ∈ F, choose b, w ∈ F that satisfy

(14). It follows that b̄ %s b, w %s w̄ and, by (15), b %s b̄, w̄ %s w, so that b̄ ∼s b and w̄ ∼s w for all
s ∈ S. By (13)


vs b̄ = vs (b) = 1 and vs (w̄) = vs (w) = 0 ∀s ∈ S


that is, v b̄ = 1S and v (w̄) = 0S . 

In the next five steps denote by b and w any two elements of F such that b %s h %s w for all
s ∈ S and h ∈ F.2 Affinity of v delivers

v (bλw) = λv (b) + (1 − λ) v (w) = λ1S ∀λ ∈ [0, 1] . (16)

Step 2. For each f ∈ F there exists a unique V (f ) ∈ [0, 1] such that f ∼ bV (f ) w and it does not
depend on the choice of b and w.3

Proof of the Step. Lemma 1 implies that b % f % w for each f ∈ F. A.1 and A.2 imply the
existence of λ such that f ∼ bλw. If f ∼ bµw for some µ > λ, by (16) and (13),

bµw s bλw ∀s ∈ S,

a contradiction, since Lemma 1 would then imply bµw  bλw. 

Step 3. V : F → [0, 1] represents % and is mixture continuous.

Proof of the Step. By (16) and (13), bλw % aµw if and only if λ ≥ µ.4 This implies that

f % g ⇐⇒ bV (f ) w % bV (g) w ⇐⇒ V (f ) ≥ V (g) ,
2
Such elements exist since v is onto.
3
In particular, V (bλw) = λ for all λ ∈ [0, 1].
4
If λ ≥ µ, then vs (bλw) = λ ≥ µ = vs (bµw) for all s ∈ S, whence bλw %s bµw for all s ∈ S, and Lemma 1
delivers bλw % bµw. The converse is shown at the end of the previous step, that is, λ 6≥ µ implies bλw 6% bµw.

14
that is, V represents %. As for mixture continuity, it is sufficient to observe that for each τ ∈ R
and f, g ∈ F, the set

 [0, 1] τ >1


{α ∈ [0, 1] : V (αf + (1 − α) g) ≤ τ } = {α ∈ [0, 1] : bτ w % αf + (1 − α) g} 0≤τ ≤1


 ∅ τ <0

is closed, and so is {α ∈ [0, 1] : V (αf + (1 − α) g) ≥ τ }. 

Step 4. If v (f ) ≥ v (g) (resp. v (f ) > v (g)), then f % g (resp. f  g).

Proof of the Step. Finally, v (f ) ≥ v (g), by definition means

vs (f ) ≥ vs (g) ∀s ∈ S

by (13),
f %s g ∀s ∈ S

which implies f % g by Lemma 1. 

Step 5. The correspondence I from [0, 1]S to [0, 1], defined for each ϕ ∈ [0, 1]S by

I : ϕ 7→ V (f ) if f ∈ F and v (f ) = ϕ

is a normalized and strictly increasing function I : [0, 1]S → [0, 1] such that

I (v (f )) = V (f ) ∀f ∈ F.

Proof of the Step. Since v is onto, for each ϕ ∈ [0, 1]S , there exists f ∈ F such that v (f ) = ϕ (the
correspondence has nonempty values in [0, 1]). Moreover, if f, g ∈ F are such that v (f ) = ϕ = v (g),
then f ∼ g by Step 4, and V (f ) = V (g). Therefore, the correspondence I is a function since it
has singleton values in [0, 1].
In general, for each ϕ ∈ [0, 1]S , set

bϕw (s) = ϕ (s) b (s) + (1 − ϕ (s)) w (s)

then

vs (bϕw) = Us (ϕ (s) b (s) + (1 − ϕ (s)) w (s)) = ϕ (s) Us (b (s)) + (1 − ϕ (s)) Us (w (s))

= ϕ (s) vs (b) + (1 − ϕ (s)) vs (w (s)) = ϕ (s)

that is
ϕ = v (bϕw)

and hence
I (ϕ) = V (bϕw) .

15
 In particular, for all λ ∈ [0, 1], I (λ1S ) = V (bλw) = λ and I is normalized. Finally, ϕ > ψ
imply v (f ) > v (g) if f, g ∈ F are such that v (f ) = ϕ and v (g) = ψ, by Step 4, f  g; therefore
V (f ) > V (g) and I (ϕ) > I (ψ). 

Step 6. I is continuous.

Proof of the Step. Consider ϕ, ψ ∈ [0, 1]S and τ ∈ R. Let f, g ∈ F be such that v (f ) = ϕ and
v (g) = ψ. Then,

{α ∈ [0, 1] : I (αϕ + (1 − α) ψ) ≤ τ } = {α ∈ [0, 1] : I (αv (f ) + (1 − α) v (g)) ≤ τ }

= {α ∈ [0, 1] : I (v (αf + (1 − α) g)) ≤ τ }
= {α ∈ [0, 1] : V (αf + (1 − α) g) ≤ τ } .

is closed since V is mixture continuous, and so is {α ∈ [0, 1] : I (αϕ + (1 − α) ψ) ≥ τ }. Since I is

increasing, by Proposition 43 of Cerreia-Vioglio, Maccheroni, Marinacci, and Montrucchio [6], it
follows that I is continuous. 

Steps 1-6 show that the axioms are sufficient for the representation, with U (x, s) = Us (x) for
all (x, s) ∈ X × S, and the representation satisfies Point 1 of the statement.

Conversely, assume that % is represented as in (11). For each s ∈ S, set Us (x) = U (x, s) for all
x ∈ X ×S. Since Us is affine for all s ∈ S, so is v defined as in Lemma 3, moreover, V (f ) = I (v (f ))
for all f ∈ F.

Step 7. % satisfies A.1 and A.2.

Proof of the Step. A.1 follows from existence of the representation V . Moreover, if f, g, h ∈ F then

{α ∈ [0, 1] : αf + (1 − α) g % h} = {α ∈ [0, 1] : I (v (αf + (1 − α) g)) ≥ I (v (h))}

= {α ∈ [0, 1] : I (αv (f ) + (1 − α) v (g)) ≥ I (v (h))} .

Since I is continuous, it follows that {α ∈ [0, 1] : αf + (1 − α) g % h} is closed (see, e.g., [6, Propo-
sition 43, page 1306]), and so is {α ∈ [0, 1] : h % αf + (1 − α) g}, that is, % satisfies A.2. 

Step 8. v is onto and such that

f %s g ⇐⇒ vs (f ) ≥ vs (g) . (17)

In particular, b, w ∈ F are such that

b %s h %s w ∀s ∈ S, h ∈ F (18)

if and only if v (b) = 1S and v (w) = 0S .

Proof of the Step. See Step 1 for the proof of the fact that v is onto. In particular, there exist
b, w ∈ F such that
v (b) = 1S and v (w) = 0S . (19)

16
Let s ∈ S. If f %s g, then fs w % gs w then, by (11), I (vs (f ) 1s ) = I (v (fs w)) ≥ I (v (gs w)) =
I (vs (g) 1s ). Since I is strictly increasing vs (f ) < vs (g) would imply I (vs (f ) 1s ) < I (vs (g) 1s ),
therefore vs (f ) ≥ vs (g). Conversely, vs (f ) ≥ vs (g) implies v (fs h) = vs (f ) 1s + v (h) − vs (h) 1s ≥
vs (g) 1s + v (h) − vs (h) 1s = v (gs h) and monotonicity of I implies I (v (fs h)) ≥ I (v (gs h)) for all
h ∈ F, and f %s g. This proves (17).
If b, w ∈ F satisfy (19), for each s ∈ S and each h ∈ F,

vs (b) = 1 ≥ vs (h) ≥ 0 = vs (w)

therefore, by (17)
b %s h %s w. (20)

Conversely, if b̄, w̄ ∈ F satisfy b̄ %s h %s w̄ for all s ∈ S and h ∈ F, choose b, w ∈ F that satisfy

(19). It follows that b̄ %s b, w %s w̄ and, by (15), b %s b̄, w̄ %s w, so that b̄ ∼s b and w̄ ∼s w for all
s ∈ S. By (17)


vs b̄ = vs (b) = 1 and vs (w̄) = vs (w) = 0 ∀s ∈ S


that is, v b̄ = 1S and v (w̄) = 0S . 

Denote by b and w any two elements of F v (b) = 1S and v (w) = 0S . Such elements exist since
v is onto and (18) guarantees that % satisfies A.9. On the other hand (17) delivers b s w for all
s ∈ S, and % satisfies A.5 Affinity of v delivers

v (bλw) = λv (b) + (1 − λ) v (w) = λ1S ∀λ ∈ [0, 1] . (21)

Also notice that, for each s ∈ S,

x %s y ⇐⇒ vs (x) ≥ vs (y) ⇐⇒ Us (x) ≥ Us (y) .

Therefore Us : X → [0, 1] represents %s on X, which together with maxx∈X Us (x) = 1 = 1 +

minx∈X Us (x) makes Us unique.

Step 9. % satisfies A.4.

Proof of the Step. Since v is affine and so is the projection on the s-th component,
1 1 1 1
f ∼s g ⇒ vs (f ) = vs (g) ⇒ vs (f ) + vs (h) = vs (g) + vs (h)
  2  2  2 2
1 1 1 1 1 1 1 1
⇒ vs f + h = vs g + h ⇒ f + h ∼s g + h.
2 2 2 2 2 2 2 2

The arbitrary chioce of s ∈ S and f, g, h ∈ F allows to deduce that % satisfies A.4. 

Step 10. % satisfies A.3.

Proof of the Step. Consider s ∈ S and f, g, h ∈ F, and assume fs g % g. Then v (fs g) = vs (f ) 1s +

v (g) − vs (g) 1s = v (g) + (vs (f ) − vs (g)) 1s . By strict monotonicity vs (f ) − vs (g) ≥ 0 (otherwise
I (v (fs g)) < I (v (g)) contradicting fs g % g). Therefore, v (fs h) = vs (f ) 1s + v (h) − vs (h) 1s ≥

17
vs (g) 1s + v (h) − vs (h) 1s = v (gs h) and monotonicity of I implies I (v (fs h)) ≥ I (v (gs h)) for all
h ∈ F, that is fs h % gs h. 

Steps 8-10 show that the axioms are necessary for the representation. Normalization of I and
(21) deliver
V (bV (f ) w) = I (v (bV (f ) w)) = I (V (f ) 1S ) = V (f ) .

Therefore, the representation satisfies Point 1 of the statement.

Let W : X ×S → [0, 1] be another function – affine and onto with respect to the first component
– and J : [0, 1]S → [0, 1] be another function – normalized, strictly increasing, and continuous –
such that the value function defined by

V 0 (f ) = J (W (f (·) , ·)) ∀f ∈ F

represents %.
The arguments used in the proof of necessity of the axioms for the representation show that
Ws : X → [0, 1] represents %s on X for each s ∈ S, which together with maxx∈X Ws (x) = 1 =
1 + minx∈X Ws (x) makes Ws unique. Thus W = U .
Moreover, for each ϕ ∈ [0, 1]S , let f ∈ F be such that ϕ = v (f ) = U (f (·) , ·) = W (f (·) , ·). By
Point 1 of the statement, and taking b, w ∈ F such that b %s h %s w for all s ∈ S,

f ∼ bV 0 (f ) w

then, by normalization of I and (21),

I (ϕ) = I (v (f )) = I v bV 0 (f ) w = V 0 (f ) = J (W (f (·) , ·)) = J (ϕ) .

That is J = I.
Uniqueness is established.

Finally, assume representation (11) holds and let v be defined as in Lemma 3.

Proof of Point 2. Assume % satisfies A.6. Consider ϕ, ψ ∈ [0, 1]|S| such that I (ϕ) = I (ψ) and
α ∈ [0, 1]. Let f, g ∈ F be such that ϕ = v (f ) and ψ = v (g) (which exist by Step 8). Then
I (v (f )) = I (v (g)), f ∼ g, and f % αf + (1 − α) g by A.6. Therefore,

I (ϕ) = V (f ) ≥ V (αf + (1 − α) g) = I (v (αf + (1 − α) g)) = I (αϕ + (1 − α) ψ) .

Since ϕ, ψ and α were arbitrarily chosen,

I (ϕ) ≥ I (αϕ + (1 − α) ψ) ∀α ∈ [0, 1] and ∀ϕ, ψ ∈ [0, 1]|S| : I (ϕ) = I (ψ) .

Since I is continuous, it follows that I is quasiconvex.

18
 Conversely, assume I is quasiconvex. Consider f, g ∈ F such that f ∼ g and α ∈ [0, 1]. Then
I (v (f )) = I (v (g)) and

V (αf + (1 − α) g) = I (v (αf + (1 − α) g)) = I (αv (f ) + (1 − α) v (g)) ≤ I (v (f )) = V (f )

where the inequality is a consequence of quasiconvexity and implies f % αf + (1 − α) g. Since f, g

and α were arbitrarily chosen, it follows that % satisfies A.6 

Proof of Point 3. Assume % satisfies A.7. Since I is monotonic and normalized, the characterization
results of [20], guarantee that, in order to show that I is a niveloid, it sufficies to prove that
 
1 1
I(αϕ + (1 − α) µ) = I(αϕ + (1 − α) ) + (1 − α) µ − (22)
2 2

for each ϕ ∈ [0, 1]|S| and α, µ ∈ [0, 1]. Consider f ∈ F such that v (f ) = ϕ and recall v (bµw) = µ,
by (21).
Notice that
1 1 1
α1S + (1 − α)
≥ αϕ + (1 − α) ≥ α0S + (1 − α)
2 2 2
1


then, by continuity of t 7→ I αt1S + (1 − α) 2 from [0, 1] → R and monotonicity of I, there is
λ ∈ [0, 1] such that    
1 1
I αλ1S + (1 − α) = I αϕ + (1 − α) . (23)
2 2
Normalization of I delivers
 
1 1
I αϕ + (1 − α) = αλ + (1 − α) . (24)
2 2

By (23) and since v αf + (1 − α) b 12 w = αϕ + (1 − α) 12 , it follows that

 
1 1 1
αf + (1 − α) b w ∼ b αλ + (1 − α) w = αbλw + (1 − α) b w
2 2 2

By A.7, this implies that

αf + (1 − α) bµw ∼ α (bλw) + (1 − α) bµw = b (αλ + (1 − α) µ) w

for all µ ∈ [0, 1], that is

I (αϕ + (1 − α) µ) = αλ + (1 − α) µ. (25)

Subtracting (24) from (25) delivers (22).

Conversely, assume I is a niveloid. If f, g ∈ F, α ∈ [0, 1], and

αf + (1 − α) (bλw) % αg + (1 − α) (bλw)

holds for some λ ∈ [0, 1]. Then,

I (αv (f ) + (1 − α) λ) ≥ I (αv (g) + (1 − α) λ)

19
since I is a niveloid this amounts to I (αv (f )) + (1 − α) λ ≥ I (αv (g)) + (1 − α) λ that is

I (αv (f )) ≥ I (αv (g))

which in turn delivers I (αv (f )) + (1 − α) µ ≥ I (αv (g)) + (1 − α) µ for all µ ∈ [0, 1], and the
properties of niveloids deliver

I (αv (f ) + (1 − α) µ) ≥ I (αv (g) + (1 − α) µ)

or
αf + (1 − α) (bµw) % αg + (1 − α) (bµw)

for all µ ∈ [0, 1]. That is % satisfies A.7. 

6 Conclusions
We have provided an axiomatic foundation of the Big Five factor Conscientiousness. The theory
provides condition under which two characteristics, the aspiration level and the cost of control,
can be uniquely identified from the choice behavior of the individual. These two characteristic
are reasonably accurate correspondent of the two aspects of Conscientiousness that have been
repeatedly identified in the literature, and labeled in the NEO-PI inventory proactive and inhibitive
side of the factor ([8]).
We consider as main contribution of the paper the successful extension of the axiomatic method
to Personality Theory. The model we have given here of the two aspects of conscientiousness is
very stylized, but it has the advantage of giving a clean correspondence between two parameters
that are in principle identifiable from behavior with general description of individual attitude that
have been so far based on factor analysis of survey questions.
This step is important in two respects. First, the analysis of the consequences of individual per-
sonality traits on behavior and performance is fast developing, and just as a sound axiomatization
of concepts such as risk aversion was needed to put the analysis of choice under uncertainty on a
sound quantitative basis, so we hope models like ours will provide a similar basis for the empiri-
cal and experimental analysis of individual behavior. Second, genetic and neuroeconomic analysis
of individuals differences is also progressing, and needs a precise definition of the phenotype (in
genetic analysis) and of the characteristics of the decision process that may be summarized by a
simple and operational measurement. Our model may provide the conceptual structure for such
measurement.

20
7 Niveloids
Let (S, Σ) be a measurable space. Denote by B0 (Σ) the set of all real-valued Σ-measurable simple
functions and by B (Σ) its supnorm closure (in the space of all real-valued bounded functions on
S). B0 (Σ, K) (resp. B (Σ, K)) is the set of all functions in B0 (Σ) (resp. B (Σ)) taking values in
the interval K ⊆ R.
When endowed with the supnorm, B0 (Σ) is a normed vector space and B (Σ) is a Banach space.
The norm dual of B0 (Σ) (resp. B (Σ)) is the space ba (Σ) of all bounded and finitely additive set
functions µ : Σ → R endowed with the total variation norm, the duality being
Z
hϕ, µi = ϕdµ

for all ϕ ∈ B0 (Σ) (resp. B (Σ)) and all µ ∈ ba (Σ) (see, e.g., [?, p. 258]). As it is well known,
on ∆ (Σ) the σ (ba (Σ) , B0 (Σ))-topology coincides with the σ (ba (Σ) , B (Σ))-topology, and they go
under the name of weak* topology; moreover a subset of ∆σ (Σ) is weakly* compact iff it is weakly
compact (i.e. compact in the weak topology of the Banach space ba (Σ)).
For ϕ, ψ ∈ B (Σ) we write ϕ ≥ ψ (resp. ϕ > ψ) if ϕ (s) ≥ ψ (s) (resp. ϕ (s) > ψ (s)) for all
s ∈ S.
Let Φ be any nonempty collection of elements of B (Σ), and Φc the constant functions in Φ.5
We call Φ a tube if Φ = Φ + R.6
Given a functional I : Φ → R, we say that I is:

(i) normalized if I (k) = k for all k ∈ Φc ;

(ii) monotonic if ϕ ≥ ψ implies I(ϕ) ≥ I(ψ) for all ϕ, ψ ∈ Φ;

(iii) vertically invariant if I(ϕ + c) = I(ϕ) + c for all ϕ ∈ Φ and c ∈ R such that ϕ + c ∈ Φ;

(iv) a niveloid if I (ϕ) − I (ψ) ≤ sup (ϕ − ψ) for all ϕ, ψ ∈ Φ.7

Remark 1 Notice that I is a niveloid iff I (ψ) − I (ϕ) ≥ − sup (ϕ − ψ) = inf (ψ − ϕ) for all
ϕ, ψ ∈ Φ iff inf (ψ − ϕ) ≤ I (ψ) − I (ϕ) ≤ sup (ψ − ϕ) for all ψ, ϕ ∈ Φ. Clearly a niveloid is
Lipschitz continuous of rank 1 in the supnorm (sup (ϕ − ψ) ≤ kϕ − ψk).

Remark 2 For all I : Φ → R, define I¯ : −Φ → R by I¯ (ϕ) = −I (−ϕ). It is easy to check that:

I¯ = I; I is normalized iff I¯ is normalized; I is monotonic iff I¯ is monotonic; I is vertically

invariant iff I¯ is vertically invariant; I is a niveloid iff I¯ is a niveloid.

5
As usual, we write k both for the real number k and for the constant function k1S ∈ B0 (Σ).
6
Clearly, if Φ is not a tube, then Φ + R is the smallest tube containing Φ.
7
Dolecki and Greco [?] call niveloid a monotonic and vertically invariant functional T : [−∞, ∞]S → [−∞, ∞].
Their Corollary 1.3 and our Lemma 7 explain why we chose to abuse this term.

21
7.1 Vertically invariant functionals

Next Lemma provides a useful condition for vertical invariance.

Lemma 4 Let Φ be a convex subset of B0 (Σ) (or B (Σ)) with 0 ∈ Φ and I : Φ → R be a functional
that satisfies
I(αϕ + (1 − α) k) = I(αϕ) + (1 − α) k (26)

for all ϕ ∈ Φ, k ∈ Φc , and α ∈ (0, 1). Then I is vertically invariant provided one of the following
conditions holds:

• Φ is open,

• I is continuous and 0 ∈ int (Φ),

• Φ = B0 (Σ, K) for some interval K ⊆ R such that 0 ∈ int (K).

Proof. If c = 0 then I(ϕ + c) = I(ϕ) + c for all ϕ ∈ Φ. It is sufficient to prove that I (ϕ + c) =

I (ϕ) + c for all ϕ ∈ Φ and c > 0 such that ϕ + c ∈ Φ.8
Let ϕ, ϕ + c ∈ Φ and c > 0.

Step 1. If ϕ, ϕ + c ∈ int (Φ), then I(ϕ + c) = I(ϕ) + c.

There exists α ∈ (0, 1) such that ϕ/α, (ϕ + c) /α ∈ int (Φ). Hence (ϕ + t) /α ∈ int (Φ) for each
t ∈ [0, c]. In fact, there exists γ ∈ [0, 1] such that t = γc and
ϕ+t ϕ + γc ϕ+c ϕ
= =γ + (1 − γ) ∈ int (Φ) .
α α α α
8
If c < 0, set ψ = ϕ + c, and d = −c. This yields ψ, ψ + d ∈ Φ and d > 0, then I (ψ + d) = I (ψ) + d, that is
I (ϕ) = I (ϕ + c) − c.

22
 c/n
Choose n ≥ 2 such that 1−α ∈ Φc .9 Then
 c c
I (ϕ + c) = I ϕ + + ... +
 n n 
c (n − 1) c
=I ϕ+ +
n n
  c(n−1)
 
ϕ+ n
= I α   + (1 − α) c/n 
α 1−α
   
ϕ + c(n−1)
= I α 
n
 + (1 − α) c/n
α 1−α
 
c (n − 1) c
=I ϕ+ +
n n
=···
 c  c (n − 1)
=I ϕ+ +
n n
c c (n − 1)
= I (ϕ) + +
n n
= I (ϕ) + c,

as wanted. 

Step 1 proves the lemma if Φ is open. If I is continuous and 0 ∈ int (Φ), since ϕ, ϕ + c ∈ Φ, then
1 − n1 ϕ and 1 − n1 (ϕ + c) ∈ int (Φ) for all n ≥ 1. But, by Step 1, I is vertically invariant on

int (Φ) and hence

       
1 1 1
I (ϕ + c) = lim I 1− (ϕ + c) = lim I 1− ϕ+ 1− c
n→∞ n n→∞ n n
    
1 1
= lim I 1− ϕ + 1− c = I (ϕ) + c.
n→∞ n n

It remains to prove the last case when K = [a, b) or (a, b] or [a, b] with −∞ ≤ a < 0 < b ≤ ∞.10

Step 2. Assume K contains b (and hence b < ∞) and a < ϕ < ϕ + c ≤ b, then I(ϕ + c) = I(ϕ) + c.
 c
b− b− c
Choose n ≥ 2 such that b − nc > 0 and ϕ/ b n > a. Set α = b n ∈ (0, 1). Notice that for all
h i
t ∈ nc , c(n−1)
n

c c (n − 1) c c
ϕ<ϕ+ ≤ϕ+t≤ϕ+ = ϕ + c − ≤ b − < b.
n n n n
Divide all the terms by α to obtain
 
 c c (n − 1)
a < ϕ/α < ϕ + /α ≤ (ϕ + t) /α ≤ ϕ + /α ≤ b
n n
9
This is possible since in any case 0 ∈ int (Φ).
10
If K = (a, b), then B0 (Σ, K) is open in B0 (Σ).

23
 h i
c c(n−1) c/n
and hence (ϕ + t) /α ∈ B0 (Σ, K) for each t ∈ n, n . Moreover 1−α = b ∈ K, and
 
c (n − 1) c
I (ϕ + c) = I ϕ + +
n n
    
c(n−1)
ϕ+ n
= I α   + (1 − α) c/n 
α 1−α
   
ϕ + c(n−1)
= I α 
n
 + (1 − α) c/n
α 1−α
 
c (n − 1) c
=I ϕ+ + .
n n
 
But a < ϕ < ϕ + c(n−1)
n < b implies ϕ, ϕ + c(n−1)
n ∈ int (Φ), and Step 1 guarantees I ϕ+ c(n−1)
n =
c(n−1)
I (ϕ) + n whence I (ϕ + c) = I (ϕ) + c. 

Step 2 concludes the proof if K = (a, b].

Step 3. Assume K contains a (and hence a > −∞) and a ≤ ϕ < ϕ + c < b, then I(ϕ + c) = I(ϕ) + c.

Consider −K and notice that −b < −ϕ − c < −ϕ ≤ −a, then ψ = −ϕ − c ∈ B0 (Σ, −K), c > 0
and ψ + c = −ϕ ∈ B0 (Σ, −K). Moreover, it is immediate to show that I¯ satisfies (26), by Step 2

I (ϕ + c) = −I¯ (−ϕ − c) = −I¯ (ψ)

= − I¯ (ψ + c) − c = −I¯ (−ϕ) + c

= I (ϕ) + c,

as wanted. 

Step 3 concludes the proof if K = [a, b).

c
If K = [a, b], then −∞ < a < b < ∞. If ϕ, ϕ + c ∈ B0 (Σ, K), then a ≤ ϕ < ϕ + 2 < b and
c
a<ϕ+ 2 < ϕ + c ≤ b, thus applying Step 2 and Step 3 we obtain
 c c
I (ϕ + c) = I ϕ + + = I (ϕ) + c.
2 2


Lemma 5 Let Φ be a convex subset of B0 (Σ) (or B (Σ)) and I : Φ → R a vertically invariant
functional that satisfies
I (αψ + (1 − α) ϕ) ≥ I (ϕ) (27)

for all ϕ, ψ ∈ Φ such that I (ψ) = I (ϕ) and α ∈ (0, 1). Then Φ is concave provided one of the
following conditions holds:

• I is continuous and int (Φ) is not empty,

24
 • Φ is a tube.

Proof. Assume I is continuous and int (Φ) is not empty. Let ϕ0 ∈ int (Φ), there exist ε > 0 such
that

N (ϕ0 , ε) = {ψ ∈ B0 (Σ) : kϕ0 − ψk ≤ ε}

= {ψ ∈ B0 (Σ) : ϕ0 − ε ≤ ψ ≤ ϕ0 + ε}

is contained in int (Φ). Moreover - by continuity - there exists ρ ∈ 0, 13 such that kϕ − ϕ0 k ≤ ρε

implies |I (ϕ) − I (ϕ0 )| ≤ 3ε . Then if ϕ, ψ ∈ N (ϕ0 , ρε), we have

2
|I (ϕ) − I (ψ)| ≤ |I (ϕ) − I (ϕ0 )| + |I (ϕ0 ) − I (ψ)| ≤ ε
3
and − 23 ε ≤ I (ϕ) − I (ψ) ≤ 2 2
3 ε. Setting t = I (ϕ) − I (ψ), we get − 3 ε ≤ t ≤ 2
3 ε. Notice that
− 13 ε ≤ −ρε ≤ ψ − ϕ0 ≤ ρε ≤ 1 1 1
3 ε and ϕ0 − 3 ε ≤ ψ ≤ ϕ0 + 3 ε. Summing up,

ϕ0 − ε ≤ ψ + t ≤ ϕ0 + ε

and ψ + t ∈ int (Φ). Since ψ ∈ int (Φ) too, then I (ψ + t) = I (ψ) + t = I (ϕ), so that

I (α (ψ + t) + (1 − α) ϕ) ≥ I (ϕ) . (28)

Hence,

I (ϕ) ≤ I (α (ψ + t) + (1 − α) ϕ) = I (αψ + (1 − α) ϕ + αt)

= I (αψ + (1 − α) ϕ) + αt
= I (αψ + (1 − α) ϕ) + α (I (ϕ) − I (ψ))

and
I (αψ + (1 − α) ϕ) ≥ αI (ψ) + (1 − α) I (ϕ) . (29)

We conclude that I is concave in N (ϕ0 , ρε). As the choice of N (ϕ0 , ε) was arbitrary, we conclude
that I is locally concave on int (Φ). A standard result from convex analysis yields concavity on
int (Φ). Finally, the continuity of I implies its concavity on the whole Φ. This proves the first
case. To prove the second, for all ϕ, ψ ∈ Φ and α ∈ (0, 1), set t = I (ϕ) − I (ψ). Since Φ is a tube,
ψ + t ∈ Φ, and I (ψ + t) = I (ψ) + t = I (ϕ). Repeat the argument leading from (28) to (29). 

Lemma 6 Let Φ be a nonempty subset of B (Σ) and I : Φ → R be a vertically invariant functional.

Then there exists a unique vertically invariant functional I˜ : Φ + R → R extending I to the tube
Φ + R generated by Φ. Moreover, if Φ is convex and I is concave, then (Φ + R is convex and) I˜ is
concave.

25
Proof. If there exists a vertically invariant functional I˜ : Φ + R → R extending I on Φ + R, then
for all ϕ + d ∈ Φ + R with ϕ ∈ Φ and d ∈ R, it satisfies

I˜ (ϕ + d) = I˜ (ϕ) + d = I (ϕ) + d. (30)

In particular it is unique. Next we show that Eq. (30) defines a vertically invariant functional (that
obviously extends I). If ϕ, ψ ∈ Φ, d, c ∈ R, and ϕ + d = ψ + c, then ϕ = ψ + c − d. In particular,
ψ ∈ Φ and c − d ∈ R are such that ψ + (c − d) ∈ Φ, and

I (ϕ) + d = I (ψ + c − d) + d
= I (ψ) + c − d + d = I (ψ) + c.

This proves that I˜ is well defined. If ϕ + d ∈ Φ + R (with ϕ ∈ Φ and d ∈ R) and c ∈ R, then

I˜ ((ϕ + d) + c) = I˜ (ϕ + d + c) = I (ϕ) + d + c = I˜ (ϕ + d) + c,

that is, I˜ is vertically invariant.

Assume I is concave. Let ϕ + d, ψ + t ∈ Φ + R with ϕ, ψ ∈ Φ and d, t ∈ R and α ∈ (0, 1). If
I˜ (ϕ + d) = I˜ (ψ + t) = c, then I (ϕ) = c − d and I (ψ) = c − t. Therefore I (αϕ + (1 − α) ψ) ≥
αI (ϕ) + (1 − α) I (ψ) = α (c − d) + (1 − α) (c − t), that is

I˜ (α (ϕ + d) + (1 − α) (ψ + t)) = I˜ ((αϕ + (1 − α) ψ) + αd + (1 − α) t)
= I (αϕ + (1 − α) ψ) + αd + (1 − α) t ≥ c.

By Lemma 5, since Φ + R is a tube, this means that I˜ is concave. 

7.2 Extensions of niveloids

In this section we obtain some novel results on the extension of niveloids (the first results on this
subject appear in [?]).

Lemma 7 Let Φ be a nonempty subset of B (Σ) and I : Φ → R. The following statements are
equivalent:

(i) I is vertically invariant and its unique vertically invariant extension I˜ to Φ + R is monotonic.

(ii) I is a niveloid.

In particular, if Φ is a tube, then I : Φ → R is a niveloid iff it is vertically invariant and

monotonic (see [?, Cor 1.3]).

Proof. Let I be vertically invariant and I˜ be monotonic. For all ϕ, ψ ∈ Φ, ϕ ≤ ψ + sup (ϕ − ψ),
but ϕ, ψ + sup (ϕ − ψ) ∈ Φ + R, then I˜ (ϕ) ≤ I˜ (ψ + sup (ϕ − ψ)) that is

I (ϕ) ≤ I (ψ) + sup (ϕ − ψ) .

26
Conversely, if I is a niveloid, for all ϕ ∈ Φ and c ∈ R such that ϕ + c ∈ Φ

c = inf ((ϕ + c) − ϕ) ≤ I (ϕ + c) − I (ϕ) ≤ sup ((ϕ + c) − ϕ) = c

that is I (ϕ + c) = I (ϕ) + c, and I is vertically invariant. Moreover, if ϕ, ψ ∈ Φ and d, t ∈ R are

such that ψ + t ≥ ϕ + d, then I (ψ) − I (ϕ) ≥ inf (ψ − ϕ) implies

I˜ (ψ + t) − I˜ (ϕ + d) = I (ψ) − I (ϕ) + t − d
≥ inf (ψ − ϕ) + t − d
= inf ((ψ + t) − (ϕ + d))
≥ 0,

that is I˜ is monotonic. 

Lemma 8 A vertically invariant and monotonic functional I : B0 (Σ, K) → R is a niveloid.

Proof. In view of Lemma 7, we just have to show that I˜ is monotonic. Let ϕ, ψ ∈ B0 (Σ, K)
and d, t, c ∈ R be such that ψ + t ≥ ϕ + d. We want to show that I (ψ) + t ≥ I (ϕ) + d, i.e., that
ψ + c ≥ ϕ implies I (ψ) + c ≥ I (ϕ).
Assume sup K = b < ∞ is not attained. If c < b − sup ψ, then ϕ ≤ ψ + c ≤ sup ψ + c < b, then
ψ + c ∈ B0 (Σ, K) and I (ϕ) ≤ I (ψ + c) = I (ψ) + c. Else c ≥ b − sup ψ ≥ 0 and there exists ε > 0
such that ϕ < b − ε < b. A fortiori c > (b − ε) − sup ψ. There are two subcases:

• c > (b − ε) − inf ψ, then I (ψ) + c ≥ I (inf ψ) + c = I˜ (0) + inf ψ + c > I˜ (0) + inf ψ + (b − ε) −
inf ψ ≥ I˜ (0) + b − ε = I (b − ε) ≥ I (ϕ) .

• c ≤ (b − ε) − inf ψ (that is inf ψ ≤ (b − ε) − c < sup ψ), then ψ + c ≥ ϕ implies (ψ + c) ∧

(b − ε) ≥ ϕ, but (ψ + c) ∧ (b − ε) ∈ B0 (Σ, K) and (ψ + c) ∧ (b − ε) = min {ψ + c, b − ε} − c +
c = min {ψ, b − ε − c} + c = (ψ ∧ (b − ε − c)) + c. Notice that also ψ ∧ (b − ε − c) ∈ B0 (Σ, K)
since (b − ε − c) ∈ [inf ψ, sup ψ) ⊆ K. Therefore

I (ψ) + c ≥ I (ψ ∧ (b − ε − c)) + c
= I ((ψ ∧ (b − ε − c)) + c)
= I ((ψ + c) ∧ (b − ε))
≥ I (ϕ) ,

as desired.

Assume that sup K = b < ∞ is attained. If c ≤ b − sup ψ, then ϕ ≤ ψ + c ≤ sup ψ + c ≤ b, then

ψ + c ∈ B0 (Σ, K) and I (ϕ) ≤ I (ψ + c) = I (ψ) + c. Else c > b − sup ψ ≥ 0 while ϕ ≤ b. There are
two subcases:

27
 • c ≥ b − inf ψ, then I (ψ) + c ≥ I (inf ψ) + c = I˜ (0) + inf ψ + c ≥ I˜ (0) + inf ψ + b − inf ψ =
I (b) ≥ I (ϕ) .

• c < b − inf ψ (that is inf ψ < b − c < sup ψ), then ψ + c ≥ ϕ implies (ψ + c) ∧ b ≥ ϕ,
but (ψ + c) ∧ b ∈ B0 (Σ, K) and (ψ + c) ∧ b = min {ψ + c, b} − c + c = min {ψ, b − c} + c =
(ψ ∧ (b − c)) + c. Notice that also ψ ∧ (b − c) ∈ B0 (Σ, K) since (b − c) ∈ (inf ψ, sup ψ) ⊆ K.
Therefore

I (ψ) + c ≥ I (ψ ∧ (b − c)) + c
= I ((ψ ∧ (b − c)) + c)
= I ((ψ + c) ∧ b)
≥ I (ϕ) ,

as desired.

Finally, if sup K = ∞, and ϕ ≤ ψ + c, then ψ + c ∈ B0 (Σ, K) and I (ϕ) ≤ I (ψ + c) = I (ψ) + c.



Lemma 9 Let I : Φ → R be a niveloid on a nonempty subset Φ of B (Σ), and set

n o
L = ϕ ∈ Φ + R : I˜ (ϕ) ≥ 0 + B Σ, R+ .

The functional defined on B (Σ) by

Iˆ (ϕ) = sup {c ∈ R : ϕ − c ∈ L} ∀ϕ ∈ B (Σ)

is the minimum niveloid on B (Σ) that extends I. Moreover, if Φ is convex and I is concave, then
Iˆ is concave.

Before entering the proof’s details, notice that if I is a niveloid on a tube Φ, then for all ϕ ∈ Φ,
I (ϕ) = sup {c ∈ R : c ≤ I (ϕ)} = sup {c ∈ R : ϕ − c ∈ {I ≥ 0}}, where {I ≥ 0} = {ϕ ∈ Φ : I (ϕ) ≥ 0}
(see also [?, p.160]).
n o
Proof. If ϕ ∈ Φ + R, and ϕ ∈ L, then ϕ ≥ ψ for some ψ ∈ I˜ ≥ 0 , whence I˜ (ϕ) ≥ I˜ (ψ) ≥ 0,
n o n o
that is ϕ ∈ I˜ ≥ 0 . This proves that ϕ ∈ Φ + R belongs to L iff it belongs to I˜ ≥ 0 . As a
consequence, for all ϕ ∈ Φ + R, we have
n n oo
I˜ (ϕ) = sup c ∈ R : ϕ − c ∈ I˜ ≥ 0

= sup {c ∈ R : ϕ − c ∈ L}
= Iˆ (ϕ) .

Then Iˆ : B (Σ) → [−∞, ∞] extends I,

˜ a fortiori I.
Notice that:

28
 • If ϕ ∈ L and ψ ≥ ϕ, then ψ ∈ L + B (Σ, R+ ) = L.

• If ϕ ∈
/ L and ψ ≤ ϕ, then ψ ∈
/ L.

• If ψ0 ∈ Φ, then ψ0 + d ∈ L iff I˜ (ψ0 + d) ≥ 0 iff d ≥ −I (ψ0 ). In particular, ψ0 − I (ψ0 ) ∈ L

and ψ0 − I (ψ0 ) − 1 ∈
/ L.

Let ψ0 ∈ Φ. For all ϕ ∈ B (Σ), ϕ − (inf ϕ − sup ψ0 + I (ψ0 )) ≥ ψ0 − I (ψ0 ) ∈ L, hence

6 ∅; therefore Iˆ (ϕ) > −∞. For all ϕ ∈ B (Σ) and all c ≥ sup ϕ − inf ψ0 +
{c ∈ R : ϕ − c ∈ L} =
I (ψ0 ) + 1, ϕ − c ≤ ϕ − (sup ϕ − inf ψ0 + I (ψ0 ) + 1) ≤ ψ0 − I (ψ0 ) − 1 ∈
/ L implies ϕ − c ∈
/ L, and
ˆ ˆ
{c ∈ R : ϕ − c ∈ L} is bounded above; therefore I (ϕ) < ∞. We conclude that I : B (Σ) → R.
If ψ ≥ ϕ and ϕ−c ∈ L, then ψ−c ≥ ϕ−c implies ψ−c ∈ L. It follows that {c ∈ R : ϕ − c ∈ L} ⊆
{c ∈ R : ψ − c ∈ L} and Iˆ (ψ) ≥ Iˆ (ϕ), i.e., Iˆ is monotonic.
Let d ∈ R and ϕ ∈ B (Σ), ϕ − c ∈ L iff (ϕ + d) − (c + d) ∈ L, that is

{c ∈ R : ϕ − c ∈ L} + d = {t ∈ R : (ϕ + d) − t ∈ L}

and

Iˆ (ϕ + d) = sup {t ∈ R : (ϕ + d) − t ∈ L}
= sup ({c ∈ R : ϕ − c ∈ L} + d)
= Iˆ (ϕ) + d.

That is Iˆ is a niveloid.
n o
Notice that Iˆ ≥ 0 = L. In fact, if Iˆ (ϕ) ≥ 0, then for all ε > 0 we have Iˆ (ϕ + ε) > 0, i.e.

sup {c ∈ R : ϕ + ε − c ∈ L} > 0,

therefore there exists c > 0 such that ϕ+ε−c ∈ L. This implies ϕ+ε ∈ L (since ϕ+ε ≥ ϕ+ε−c ∈ L).
n o
Since this is true for all ε, it follows ϕ = limn ϕ + n1 ∈ L, and we conclude Iˆ ≥ 0 ⊆ L.


n o
Conversely, for all ϕ ∈ L, ϕ − 0 ∈ L guarantees Iˆ (ϕ) ≥ 0; the continuity of Iˆ implies L ⊆ Iˆ ≥ 0 .
Let Jˆ be a niveloid on B (Σ) that extends I, then Jˆ coincides with I˜ on Φ + R. For all
n o
ψ ∈ L there exists ϕ ∈ I˜ ≥ 0 such that ψ ≥ ϕ, therefore Jˆ (ψ) ≥ Jˆ (ϕ) = I˜ (ϕ) ≥ 0. Then
n o n o
Iˆ ≥ 0 = L ⊆ Jˆ ≥ 0 , and this implies that for all ϕ ∈ B (Σ)
n n oo
Iˆ (ϕ) = sup c ∈ R : ϕ − c ∈ Iˆ ≥ 0
n n oo
≤ sup c ∈ R : ϕ − c ∈ Jˆ ≥ 0 = Jˆ (ϕ) .

This shows that Iˆ is the minimum niveloid on B (Σ) that extends I.

n o
Assume Φ is convex and I is concave, then I˜ is concave and I˜ ≥ 0 is convex. So L =
n o n o
I˜ ≥ 0 + B (Σ, R+ ) and Iˆ ≥ 0 = L are convex. This implies that Iˆ is concave. In fact, for all

29
 n o
ϕ, ψ ∈ B (Σ) such that Iˆ (ϕ) = Iˆ (ψ) = c, and α ∈ (0, 1), since ϕ − c, ψ − c ∈ Iˆ ≥ 0 , then

Iˆ (αϕ + (1 − α) ψ) − Iˆ (ϕ) = Iˆ (αϕ + (1 − α) ψ − c)

= Iˆ (α (ϕ − c) + (1 − α) (ψ − c)) ≥ 0,

and Lemma 5 guarantees concavity. 

Inspection of the proof shows that for a non-empty subset Φ of B0 (Σ) setting
n o
L0 = ϕ ∈ Φ : I˜ (ϕ) ≥ 0 + B0 Σ, R+

we could obtain the minimum niveloid extending I to B0 (Σ) .

7.3 Fenchel conjugates of concave niveloids

Remark 3 If I : B (Σ) → R is a concave niveloid, direct application of the Fenchel Moreau

Theorem (see, e.g., [?, p. 42]) guarantees

I (ϕ) = min (hϕ, µi − I ∗ (µ))

µ∈ba(Σ)

where I ∗ (µ) = inf ψ∈B(Σ) (hψ, µi − I (ψ)) is the Fenchel conjugate of I. If µ is not positive, there
exists ϕ ≥ 0 such that hϕ, µi < 0, then hαϕ, µi − I (αϕ) ≤ α hϕ, µi − I (0) for all α ≥ 0, whence
I ∗ (µ) = −∞. If µ (S) 6= 1, choose ψ ∈ B (Σ), then hψ + c, µi − I (ψ + c) = hψ, µi − I (ψ) +
c (µ (S) − 1) for all c ∈ R, and so I ∗ (µ) = −∞. That is,

I (ϕ) = min (hϕ, µi − I ∗ (µ)) .

µ∈∆(Σ)

In this section Φ is a (non-empty) convex subset of B (Σ) and I : Φ → R is a concave niveloid.

We set
∂π I (ϕ) = {p ∈ ∆ (Σ) : I (ψ) − I (ϕ) ≤ hψ − ϕ, pi for each ψ ∈ Φ} .

Notice that
n o
∂π I (ϕ) = p ∈ ∆ (Σ) : I˜ (ψ) − I˜ (ϕ) ≤ hψ − ϕ, pi for each ψ ∈ Φ + R .

Lemma 10 Let I : Φ → R be a concave niveloid. Then, ∂π I (ϕ) 6= ∅ for all ϕ ∈ Φ.

Proof. By Lemma 9, there exists a concave niveloid Iˆ on B (Σ) such that Iˆ|Φ = I. Let ϕ ∈ Φ.
Being a niveloid, Iˆ is Lipschitz continuous, then, its standard superdifferential at ϕ
n o
∂ Iˆ (ϕ) = µ ∈ ba (Σ) : Iˆ (ψ) − Iˆ (ϕ) ≤ hψ − ϕ, µi for each ψ ∈ B (Σ)

is nonempty (see, e.g., [?, p. 6-7]).

For all c ∈ R and µ ∈ ∂ Iˆ (ϕ) we have

Iˆ (ϕ) + c = Iˆ (ϕ + c) ≤ Iˆ (ϕ) + hϕ + c − ϕ, µi = Iˆ (ϕ) + cµ (S) ,

30
and so c ≤ cµ (S). This implies µ (S) = 1.
For all ψ ≥ 0 and µ ∈ ∂ Iˆ (ϕ) we have

hψ, µi = hϕ + ψ, µi − hϕ, µi ≥ Iˆ (ϕ + ψ) − Iˆ (ϕ) ≥ 0,

this implies µ ∈ ba+ (Σ).

Therefore, ∂ Iˆ (ϕ) ⊆ ∂π I (ϕ) and we conclude that ∂π I (ϕ) 6= ∅. 

Lemma 11 Let Φ be a convex subset of B (Σ) such that Φc 6= ∅, and I : Φ → R be a concave and
normalized niveloid. Then:

(i) For each ϕ ∈ Φ,

I (ϕ) = min (hϕ, pi − I ? (p)) = S min (hϕ, pi − I ? (p)) (31)

p∈∆(Σ) p∈ ψ∈Φ ∂π I(ψ)

where I ? : ∆ (Σ) → [−∞, 0] is given by

I ? (p) = inf (hψ, pi − I (ψ)) ∀p ∈ ∆ (Σ) .

ψ∈Φ

(ii) I ? is the maximal functional R : ∆ (Σ) → [−∞, 0] such that

I (ϕ) = inf (hϕ, pi − R (p)) ∀ϕ ∈ Φ. (32)

p∈∆(Σ)

(iii) I ? coincides with the Fenchel conjugate Iˆ∗ of Iˆ on ∆ (Σ) and

Iˆ (ϕ) = min (hϕ, pi − I ? (p)) ∀ϕ ∈ B (Σ) . (33)

p∈∆(Σ)

(iv) If (32) holds, and Ψ ⊆ Φ is such that sups∈S ψ (s) − inf s∈S ψ (s) < c for all ψ ∈ Ψ, then

I (ψ) = inf (hϕ, pi − R (p)) ∀ψ ∈ Ψ. (34)

{p∈∆(Σ):R(p)≥−c}

Proof. Notice that I ? (p) ≤ 0 for all p ∈ ∆ (Σ). For, if we take a constant k ∈ Φc we have
hk, pi = I (k) = k.
By definition of I ? , for all ϕ ∈ Φ and p ∈ ∆ (Σ)

I (ϕ) ≤ hϕ, pi − I ? (p) ; (35)

moreover,

p ∈ ∂π I (ϕ) ⇔ I (ϕ) ≥ I (ψ) − hψ, pi + hϕ, pi ∀ψ ∈ Φ

⇔ I (ϕ) ≥ sup (I (ψ) − hψ, pi) + hϕ, pi
ψ∈Φ

⇔ I (ϕ) ≥ hϕ, pi − inf (hψ, pi − I (ψ))

ψ∈Φ

⇔ I (ϕ) ≥ hϕ, pi − I ? (p)

⇔ I (ϕ) = hϕ, pi − I ? (p) .

31
Therefore, for all ϕ ∈ Φ

I (ϕ) = min (hϕ, pi − I ? (p)) ≥ S inf (hϕ, pi − I ? (p))

p∈∂π I(ϕ) p∈ ψ∈Φ ∂π I(ψ)

≥ inf (hϕ, pi − I ? (p)) ≥ I (ϕ) ,

p∈∆(Σ)

which implies (31). This proves (i). For later use, notice that if P is a subset of ∆ (Σ) such that
∂π I (ϕ) ∩ P 6= ∅ for all ϕ ∈ Φ, then the above argument yields

I (ϕ) = min (hϕ, pi − I ? (p)) . (36)

p∈P

Let R : ∆ (Σ) → [−∞, 0] be such that I (ϕ) = inf p∈∆(Σ) (hϕ, pi − R (p)) for all ϕ ∈ Φ. Then,

R (p) ≤ hϕ, pi − I (ϕ) ∀p ∈ ∆ (Σ) , ϕ ∈ Φ,

and hence
R (p) ≤ inf (hϕ, pi − I (ϕ)) = I ? (p) ∀p ∈ ∆ (Σ) .
ϕ∈Φ

This proves (ii).

For all ϕ ∈ B (Σ), set Jˆ (ϕ) = inf p∈∆(Σ) (hϕ, pi − I ? (p)), Jˆ is a normalized and concave niveloid
on B (Σ) that extends I. By (ii) applied to J, ˆ we obtain

Jˆ∗ (p) = Jˆ? (p) ≥ I ? (p) = inf (hϕ, pi − I (ϕ))

ϕ∈Φ
 
≥ inf ˆ
hϕ, pi − I (ϕ) (this is Iˆ∗ (p) )
ϕ∈B(Σ)
 
(since Iˆ ≤ J)
ˆ ≥ inf hϕ, pi − Jˆ (ϕ) = Jˆ∗ (p)
ϕ∈B(Σ)

that is Jˆ∗ (p) = I ? (p) = Iˆ∗ (p) for all p ∈ ∆ (Σ). Apply (i) (or Remark 3) to Iˆ to obtain
 
Iˆ (ϕ) = min hϕ, pi − Iˆ∗ (p) = min (hϕ, pi − I ? (p)) ∀ϕ ∈ B (Σ) .
p∈∆(Σ) p∈∆(Σ)

This completes the proof of (iii).

Finally, as to (iv), the monotonicity of Iˆ implies that inf s∈S ψ (s) = Iˆ (inf s∈S ψ (s)) ≤ I (ψ) , hψ, pi ≤
sups∈S ψ (s) = Iˆ (sups∈S ψ (s)) for all p ∈ ∆ (Σ) and all ψ ∈ Ψ. For each ψ ∈ Ψ, there exists ε > 0
such that sups∈S ψ (s) − inf s∈S ψ (s) + ε < c. For all p ∈ ∆ (Σ) such that R (p) < −c, we have

R (p) < − sup ψ (s) + inf ψ (s) − ε, and

s∈S s∈S

sup ψ (s) + ε < inf ψ (s) − R (p) , i.e.

s∈S s∈S

I (ψ) + ε ≤ sup ψ (s) + ε < inf ψ (s) − R (p) ≤ hϕ, pi − R (p) .

s∈S s∈S

On the other hand,

I (ψ) = inf (hψ, pi − R (p))

p∈∆(Σ)
 
= min inf (hψ, pi − R (p)) , inf (hψ, pi − R (p))
p∈{R<−c} p∈{R≥−c}

32
which concludes the proof, since inf p∈{R<−c} (hϕ, pi − R (p)) ≥ I (ψ) + ε. 

Remark 4 Inspection of the proof shows that: (i) ∂π I (ϕ) = arg minp∈∆(Σ) (hϕ, pi − I ? (p)). (ii)
If k ∈ Φc , then ∂π I (k) = {I ? = 0} = arg maxp∈∆(Σ) I ? (p). (iii) I ? is concave and weakly* upper
semicontinuous.

Corollary 1 Let Φ be a convex subset of B0 (Σ) (resp. B (Σ)) such that Φc 6= ∅ and Φ+R = B0 (Σ)
(resp. Φ + R = B (Σ)),11 and I : Φ → R be a concave and normalized niveloid. Then, I ? is the
Fenchel conjugate of the unique niveloid I˜ extending I to B0 (Σ) (resp. B (Σ)). In this case I ? is
the unique concave and weakly* upper semicontinuous function R : ∆ (Σ) → [−∞, 0] such that

I (ϕ) = inf (hϕ, pi − R (p)) ∀ϕ ∈ Φ.

p∈∆(Σ)

Proof. The equality

I ? (p) = inf (hψ, pi − I (ψ)) = inf (hψ, pi + c − I (ψ) − c)

ψ∈Φ ψ∈Φ
c∈R
   
= inf hψ + c, pi − I˜ (ψ + c) = inf hψ, pi − I˜ (ψ)
ψ∈Φ ψ∈B0 (Σ)
c∈R

yields the first part of the statement. Let R : ∆ (Σ) → [−∞, 0] be a concave and weakly* upper
semicontinuous functional such that I (ϕ) = inf p∈∆(Σ) (hϕ, pi − R (p)) for all ϕ ∈ Φ. Then, I˜ (ϕ) =
inf p∈∆(Σ) (hϕ, pi − R (p)) for all ϕ ∈ B0 (Σ). By the Fenchel-Moreau Theorem

R (p) = R∗∗ (p) = inf (hϕ, pi − R∗ (ϕ))

ϕ∈B0 (Σ)
  
= inf hϕ, pi − inf (hϕ, pi − R (p))
ϕ∈B0 (Σ) p∈∆(Σ)
 
= inf hϕ, pi − I˜ (ϕ) = I˜∗ (p) = I ? (p) ,
ϕ∈B0 (Σ)

as desired. 

11
E.g. Φ = B0 (Σ, K) with K an unbounded interval.

33
References
[1] Anscombe, F.J. & Aumann, R., (01963), A Definition of Subjective Probability, Annals of
mathematical Statistics, 34, 199–205.

[2] Barrick, M. R., & Mount, M. K. (1991). The big five personality dimensions and job perfor-
mance: A meta-analysis. Personnel Psychology, 44, 1–26

[3] Barrick, M.R., Mount, M.K., Judge, T.A. (2001). Personality and performance at the beginning
of the new millennium: What do we know and where do we go next? nternational Journal of
Selection and Assessment 9, 9-30

[4] Castagnoli, E., (1990), Qualche riflessione sull’Utilità Attesa, Ratio mathematica, 51–59.

[5] E. Castagnoli, LiCalzi, M., (2006), Benchmarking real-valued acts, Games and Economic Be-
havior, 236-253

[6] S. Cerreia-Vioglio, F. Maccheroni, M. Marinacci, and L. Montrucchio, Uncertainty averse pref-

erences, Journal of Economic Theory, 146, 1275–1330, 2011.

[7] Costa, P. T., Jr., & McCrae, R. R. (1992). Four ways five factors are basic. Personality and
Individual Differences, 13, 653–665.

[8] Costa, Paul T., Mccrae, Robert R. & Dye, David A., (1991), Facet scales for agreeableness
and conscientiousness: A revision of the NEO personality inventory, Personality and Individual
Differences, 12, 9, 887–898.

[9] DeYoung, C. G. & Gray, J. R. (2009). Personality Neuroscience: Explaining Individual Dif-
ferences in Affect, Behavior, and Cognition. In P. J. Corr & G. Matthews (Eds.), Cambridge
handbook of personality, 323–346, New York: Cambridge University Press

[10] DeYoung, C. G., Quilty, L. C., & Peterson, J. B. (2007). Between facets and domains: 10
Aspects of the Big Five, Journal of Personality and Social Psychology, 93, 880–896

[11] Dolecki, S. and Greco, G., Niveloids, Topological Methods in Nonlinear Analysis, 5, 1–22, 1995.

[12] Depue, R. A. & Collins, P. F. (1999). Neurobiology of the structure of personality: Dopamine,
facilitation of incentive motivation, and extraversion. Behavioral and Brain Sciences, 22, 491-
569.

[13] Drèze, J. H., (1987). Decision Theory with Moral hazard and state-dependent preferences,
Chapter 2 in Essays on Economic Decisions under Uncertainty, Cambridge University Press,
Cambridge, UK

[14] Gosling, S. D. & John, O. P. (1999). Personality dimensions in nonhuman animals: A

crossspecies review. Current Directions in Psychological Science, 8, 69–75.

34
[15] I. N Herstein and J. Milnor, An axiomatic approach to measurable utility, Econometrica, 21,
291–297, 1953.

[16] Judge, T. A., Higgins, C., Thoresen, C. J., & Barrick, M. R. (1999). The Big Five personality
traits, general mental ability, and career success across the life span. Personnel Psychology, 52,
621-652

[17] Kahneman, D. & Tversky, A. (1979) Prospect Theory: An Analysis of Decision under Risk,
Econometrica, 47, 263-291.

[18] Leatta M. Hough, (1992), The ’Big Five’ Personality Variables–Construct Confusion: Descrip-
tion Versus Prediction Human Performance, 5, 139–155.

[19] King, J.E., & Figueredo, A.J. (1997). The five-factor model plus dominance in chimpanzee
personality. Journal of Research in Personality, 31, 271–271.

[20] Maccheroni, F., Marinacci, M, & Rustichini, A. (2006), Niveloids and their extensions, mimeo

[21] P. K. Monteiro, Some results on the existence of utility functions on path connected spaces,
Journal of Mathematical Economics, 16, 147–156, 1987.

[22] Mount M.K., Barrick, M.R., (1995), The Big Five personality dimansions: Implications for
reaserach and practice in human resources management, Reserach in Personnel and Human
Respurces Management, 13, 153–200.

[23] Roberts, B. W., Chernyshenko, O. S., Stark, S., & Goldberg, L. R. (2005). The structure of
Conscientiousness: An empirical investigation based on seven major personality questionnaires.
Personnel Psychology, 58, 103-139.

[24] Roberts, B. W., Kuncel, N., Shiner, R., N., Caspi, A., & Goldberg, L. R. (2007). The power of
personality: The comparative validity of personality traits, socio-economic status, and cognitive
ability for predicting important life outcomes. Perspectives in Psychological Science, 2, 313–345

[25] Rustichini, A., DeYoung, C. G., Anderson, J. and Burks, S. (2011), Toward the Integration of
Personality Theory and Decision Theory in the Explanation of Economic and Health Behavior,
University of Minnesota, Mimeo

35