# Archimedean Choice Functions: an Axiomatic Foundation for Imprecise Decision Making

If uncertainty is modelled by a probability measure, decisions are typically made by choosing the option with the highest expected utility. If an imprecise probability model is used instead, this decision rule can be generalised in several ways. We here focus on two such generalisations that apply to sets of probability measures: E-admissibility and maximality. Both of them can be regarded as special instances of so-called choice functions, a very general mathematical framework for decision making. For each of these two decision rules, we provide a set of necessary and sufficient conditions on choice functions that uniquely characterises this rule, thereby providing an axiomatic foundation for imprecise decision making with sets of probabilities. A representation theorem for Archimedean choice functions in terms of coherent lower previsions lies at the basis of both results.

## Authors

• 11 publications
• ### An alternative approach to coherent choice functions

Choice functions constitute a simple, direct and very general mathematic...
02/28/2019 ∙ by Jasper De Bock, et al. ∙ 0

• ### A Desirability-Based Axiomatisation for Coherent Choice Functions

Choice functions constitute a simple, direct and very general mathematic...
06/04/2018 ∙ by Jasper De Bock, et al. ∙ 0

• ### On a notion of independence proposed by Teddy Seidenfeld

Teddy Seidenfeld has been arguing for quite a long time that binary pref...
02/20/2021 ∙ by Jasper De Bock, et al. ∙ 0

• ### The Choice Function Framework for Online Policy Improvement

There are notable examples of online search improving over hand-coded or...
10/01/2019 ∙ by Murugeswari Issakkimuthu, et al. ∙ 0

• ### Constructive Decision Theory

In most contemporary approaches to decision making, a decision problem i...
06/23/2009 ∙ by Lawrence Blume, et al. ∙ 0

• ### Probabilistic Prediction for Binary Treatment Choice: with focus on personalized medicine

This paper extends my research applying statistical decision theory to t...
10/02/2021 ∙ by Charles F. Manski, et al. ∙ 0

• ### Equivariant ZFA with Choice: a position paper

We propose Equivariant ZFA with Choice as a foundation for nominal techn...
03/23/2018 ∙ by Murdoch J. Gabbay, et al. ∙ 0

##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

Decision making under uncertainty is typically carried out by combining an uncertainty model with a decision rule. If uncertainty is modelled by a probability measure, the by far most popular such decision rule is maximising expected utility, where one chooses the option—or makes the decision—whose expected utility with respect to this probability measure is the highest.

Uncertainty can also be modelled in various other ways though. The theory of imprecise probabilities, for example, offers a wide range of extensions of probability theory that provide more flexible modelling possibilities, such as differentiating between stochastic uncertainty and model uncertainty. The most straightforward such extension is to consider a set of probability measures instead of a single one, but one can also use interval probabilities, coherent lower previsions, sets of desirable gambles, belief functions, to name only a few.

For all these different types of uncertainty models, various decision rules have been developed, making the total number of possible combinations rather daunting. Choosing which combination of uncertainty model and decision rule to use is therefore difficult and often dealt with in a pragmatic fashion, by using a combination that one is familiar with, that is convenient or that is computationaly advantageous.

To motivate the use of a specific combination in a more principled way, one can also consider its properties. That is, one can aim to select an uncertainty model and decision rule whose resulting decisions satisfy the properties that one finds desirable for the decision problem at hand. In some cases, imposing a given set of properties as axioms can even rule out all combinations but one, thereby providing an axiomatic foundation for the use of a specific type of uncertainty model and decision rule. The famous work of Savage [savage1972], for example, provides an axiomatic foundation for maximising expected utility with respect to a probability measure.

The main contributions of this paper are axiomatic foundations for three specific decision rules that make use of imprecise probability models [troffaes2007]. The first two decision rules, called E-admissibility and maximality, apply to a set of probability measures; they both reduce to maximising expected utility when this set contains only a single probability measure, but are otherwise fundamentally different. The third decision rule applies to sets of coherent lower previsions; it is more abstract then the other two, but includes both of them as special cases. This allows us to use our axiomatic foundation for the third rule as an intermediate step towards axiomatising E-admissibility and maximality.

To obtain our results, we make extensive use of choice functions [seidenfeld2010, 2017vancamp:phdthesis]: a unifying framework for studying decision making. These choice functions require no reference to an uncertainty model or a decision rule, but are simply concerned with the decisions themselves, making them an excellent tool for comparing different methods. We will be especially interested in Archimedean choice functions, because all of the decision schemes that we consider are of this particular type.

## 2 Choice Functions and Uncertainty Models

A choice function , quite simply, is a function that chooses. Specifically, for every finite set of options, it returns a subset of . We here consider the special case where options are gambles: bounded real functions on some fixed state space . We let be the set of all gambles on and we use to denote the set of all finite subsets of , including the empty set. A choice function is then a map from to such that, for all , .

If contains only a single option , this means that is chosen from . If consists of multiple options, several interpretations can be adopted. On the one hand, this can be taken to mean that each of the options in is chosen. On the other hand, can also be regarded as a set of options among which it is not feasible to choose, in the sense that they are incomparable based on the available information; in other words: the elements of are rejected, but those in are not necessarily ‘chosen’. While our mathematical results further on do not require a philosophical stance in this matter, it will become apparent from our examples and interpretation that we have the later approach in mind.

A very popular class of choice functions—while not necessarily always called as such—are those that correspond to maximising expected utility. The idea there is to consider a probability measure and to let be the element(s) of whose expected value—or utility—is highest. The probability measure in question is often taken to be countably additive, but we will not impose this restriction here, and impose only finite additivity. These finitely additive probability measures are uniquely characterised by their corresponding expectation operators, which are linear real functionals on that dominate the infinum operator. We follow de Finetti in denoting these expectation operators by as well, and in calling them linear previsions [finetti1970, walley1991].

###### Definition 1 (Linear prevision)

A linear prevision  on  is a real-valued map on  that satisfies

1. [label=.,ref=,leftmargin=*,start=1]

2. for all ; boundedness

3. for all real and ; homogeneity

We denote the set of all linear previsions on  by .

For any such linear prevision—or equivalently, any finitely additive probability measure—the choice function obtained by maximising expected utility is defined by

 CP(A)\coloneqq{u∈A:(∀v∈A∖{u}) P(u)≥P(v)}\leavevmode\nobreak\ \leavevmode\nobreak% \ for all A∈Q. (1)

It returns the options in that have the highest prevision—or expectation—.

However, there are also many situations in which it is not feasible to represent uncertainty by a single prevision or probability measure [walley1991, Section 1.4.4]. In those cases, imprecise probability models can be used instead. The most straightforward such imprecise approach is to consider a non-empty set of linear previsions—or probability measures—as uncertainty model, the elements of which are regarded as candidates for some ‘true’ but unknown precise model.

In that context, maximising expected utility can be generalised in several ways [troffaes2007], of which we here consider two. The first is called E-admissibility; it chooses those options that maximise expected utility with respect to at least one precise model in :

 CEP(A)\coloneqq {u∈A:(∃P∈P)(∀v∈A∖{u}) P(u)≥P(v)}\leavevmode% \nobreak\ \leavevmode\nobreak\ for all A∈Q. (2)

The second generalisation is called maximality and starts from a partial order on the elements of . In particular, for any two options , the option is deemed better than if its expectation is higher for every . Decision making with maximality then consists in choosing the options in that are undominated in this order, in the sense that no other option is better than :

 CMP(A)\coloneqq{u∈A:(∀v∈A∖{u})(∃P∈P) P(u)≥P(v)}\leavevmode\nobreak\ \leavevmode\nobreak\ for all A∈Q. (3)

If contains only a single linear prevision, it can easily be seen that maximality and E-admissibility coincide and both reduce to maximising expected utility. In all other cases, however, these two approaches are different; see for example Proposition 3 in Section 9 for a formal statement.

One of the main aims of this paper is to characterise each of these two types of choice functions in terms of their properties. That is, we are looking for necessary and sufficient conditions under which a general choice function is of the form or , without assuming a priory the existence of a set of linear previsions . Such conditions will be presented in Section 7 and 8, respectively.

A crucial intermediate step in obtaining these two results will consist in finding a similar characterisation for choice functions that correspond to (sets of) coherent lower previsions [walley1991], a generalisation of linear previsions that replaces additivity by the weaker property of superadditivity.

###### Definition 2 (Coherent lower prevision)

A coherent lower prevision  on  is a real-valued map on  that satisfies

1. [label=.,ref=,leftmargin=*,start=1]

2. for all ; boundedness

3. for all real and ; positive homogeneity

We denote the set of all coherent lower previsions on  by .

That linear previsions are a special case of coherent lower previsions follows trivially from their definitions. There is however also a more profound connection between both concepts: coherent lower previsions are minima of linear ones.

###### Theorem 2.1

[walley1991, Section 3.3.3.] A real-valued map on is a coherent lower prevision if and only if there is a non-empty set of linear previsions such that

 P––(u)=min{P(u):P∈P} for all u∈L.

Alternatively, coherent lower previsions can also be given a direct behavioural interpretation in terms of gambling, without any reference to probability measures or linear previsions [walley1991, williams1975].

Regardless of their interpretation, with any given non-empty set of these coherent lower previsions, we can associate a choice function in the following way:

 CP(A)\coloneqq{u∈A:(∃P––∈P)(∀v∈A∖{u}) P––(v−u)≤0} for all A∈Q. (4)

If the lower previsions in are all linear, this definition reduces to E-admissibility, as can be seen by comparing Equations (2) and (4). What is far less obvious though, is that maximality is also a special case of Equation (4); see Theorem 8.1 in Section 8 for a formal statement. In that case, the sets in Equations (4) and (3) may of course—and typically will—be different.

Because the choice functions that correspond to E-admissibility and maximality are both of the form , with a set of coherent lower previsions, any attempt at characterising the former will of course benefit from characterising the latter. A large part of this paper will therefore be devoted to the development of necessary and sufficient conditions for a general choice function to be of the form . In order to obtain such conditions, we will interpret choice functions in terms of (strict) desirability and establish a connection with so-called sets of desirable option sets. This interpretation will lead to a natural set of conditions that, as we will eventually see in Section 6, uniquely characterises choice functions of the form . We start with a brief introduction to desirability and sets of desirable option sets.

## 3 Coherent Sets of Desirable Option Sets

The basic notion on which our interpretation for choice functions will be based, and from which our axiomatisation will eventually be derived, is that of a desirable option: an option that is strictly preferred over the status quo [couso2011, quaeghebeur2015:statement, walley1991]. In our case, where options are gambles on and the status quo is the zero gamble, this means that the uncertain—and possibly negative—reward , whose actual value depends on the uncertain state , is strictly preferred over the zero reward. In other words: gambling according to is strictly preferred over not gambling at all.

We will impose the following three principles on desirable options, where we use ‘’ as a shorthand notation for ‘, and ’.111There is no consensus on which properties to impose on desirability; plenty of subtle variations have been considered over the years—often under different names—and continue to exist alongside each other; the main ideas and results are always very similar though [couso2011, quaeghebeur2015:statement, walley1991, williams1975]. 2 is for example often strengthened by requiring that is desirable as soon as and ; we here prefer 2 because it combines more easily with the notion of strict desirability that we will consider in Section 5. Simmilar comments apply to 23 and 2 further on. The first two principles follow readily from the meaning of desirability. The third one follows from an assumption that rewards are expressed in a linear utility scale.

1. [label=.,ref=,start=1,leftmargin=*,topsep=5pt,itemsep=2pt]

2. is not desirable;

3. if , then is desirable;

4. if are desirable and , then is desirable.

The notion of a desirable option gives rise to two different frameworks for modelling a subject’s uncertainty about the value . The first, which is well established, is that of sets of desirable options—or sets of desirable gambles. The idea there is to consider a set that consists of options that are deemed desirable by a subject. If such a set is compatible with the principles 13, it is called coherent.

###### Definition 3 (Coherence for sets of desirable options)

A set of desirable options is coherent if it satisfies the following axioms:

1. [label=.,ref=,leftmargin=*]

2. ;

3. if  , then ;

4. if and , then .

We denote the set of all coherent sets of desirable options by .

A more general framework, which will serve as our main workhorse in this paper, is that of sets of desirable option sets [debock2018]. The idea here is to consider a set of so-called desirable option sets , which are finite sets of options that, according to our subject, are deemed to contain at least one desirable option. To say that is a desirable option set, for example, means that or is desirable. Crucially, the framework of sets of desirable option sets allows a subject to make this statement without having to specify—or know—which of the two options or is desirable. As explained in earlier work [pmlr-v103-de-bock19b, Section 3], it follows from 13 that any set of desirable option sets should satisfy the following axioms. If it does, we call coherent.

###### Definition 4 (Coherence for sets of desirable option sets)

A set of desirable option sets is coherent if it satisfies the following axioms:

1. [label=.,ref=,leftmargin=*,start=0]

2. if then also , for all ;

3. ;

4. for all with ;

5. if and if, for all and , , then also

 {λu,vu+μu,vv:u∈A1,v∈A2}∈K;
6. if and , then also , for all .

We denote the set of all coherent sets of desirable option sets by .

One particular way of obtaining a set of desirable option sets, is to derive it from a set of desirable options , as follows:

 KD\coloneqq{A∈Q:A∩D≠∅}. (5)

One can easily verify that if is coherent, then will be as well [pmlr-v103-de-bock19b, Proposition 8]. In general, however, sets of desirable option sets are more expressive than sets of desirable options. The link between both is provided by Theorem 3.1, which shows that a set of desirable option sets can be equivalently represented by a set of sets of desirable options.

###### Theorem 3.1

[pmlr-v103-de-bock19b, Theorem 9] A set of desirable option sets is coherent if and only if there is some non-empty set of coherent sets of desirable options such that .

In practice, modelling a subject’s uncertainty does not require her to specify a full coherent set of desirable option sets though. Instead, it suffices for her to provide an assessment , consisting of option sets that she considers desirable. If such an assessment is consistent with coherence, meaning that there is at least one coherent set of desirable option sets that includes , then this assessment can always be extended to a unique smallest—most conservative—coherent set of desirable option sets, called the natural extension of . This natural extension is given by

 Ex(A)\coloneqq⋂{K∈¯¯¯¯¯K:A⊆K}=⋂{KD:D∈¯D,A⊆KD}, (6)

as follows readily from Theorem 3.1. If is not consistent with coherence, is an empty intersection, which, by convention, we set equal to .

## 4 Strongly Archimedean Sets of Desirable Option Sets

That sets of desirable option sets can be used to axiomatise choice functions of the type , with a set of coherent lower previsions, was already demonstrated in earlier work [pmlr-v103-de-bock19b] for the specific case where is closed with respect to pointwise convergence. A key step in that result consisted in strengthening the interpretation of , replacing desirability with the stronger notion of strict desirability [walley1991, Section 3.7.7]. We here repeat the reasoning that led to this result, before adapting it in Section 5 to get rid of the closure condition.

We call a desirable option strictly desirable if there is some real such that is desirable. As a simple consequence of this definition and 13, we find that

1. [label=.,ref=,start=1,leftmargin=*,topsep=5pt,itemsep=2pt]

2. is not strictly desirable;

3. if , then is strictly desirable;

4. if are strictly desirable and , then is strictly desirable;

5. if is strictly desirable, then is strictly desirable for some real .

By applying these principles to sets of desirable options, we arrive at the concept of a coherent set of strictly desirable options: a coherent set of desirable options that is compatible with 4. What is particularly interesting about such sets is that they are in one-to-one correspondence with coherent lower previsions [pmlr-v103-de-bock19b, walley1991], thereby allowing us to move from desirability to lower previsions as a first step towards choice functions of the form . The problem with coherent sets of strictly desirable options, however, is that they correspond to a single lower prevision , while we which to consider a set of them. To achieve this, we again consider sets of desirable option sets, but now suitably adapted to strict desirability.

So consider any set of desirable option sets and let us interpret it in terms of strict desirability. That belongs to then means that contains at least one strictly desirable option. Given this interpretation, what properties should satisfy? Since the principles 13 are identical to 13, should clearly be coherent, meaning that it should satisfy 15. Formalising the implications of 4 is more tricky though, as it can be done in several ways.

The first and most straighforward approach is to impose the following immediate translation of 4 to desirable option sets, where for all and real :

 A−ϵ\coloneqq{u−ϵ:u∈A}
###### Definition 5 (Strong Archimedeanity for sets of desirable option sets)

A set of desirable option sets is Strongly Archimedean if it is coherent and satisfies222In earlier work [pmlr-v103-de-bock19b], we have referred to this property as Archimedeanity. With hindsight, however, we now prefer to reserve this terminology for the property in Definition 6.

1. [label=.,ref=,leftmargin=*]

2. if , then also for some real .

The reasoning behind this axiom goes as follows. Since is taken to mean that there is at least one that is strictly desirable, it follows from 4 that there is some real such that is strictly desirable. This implies that contains at least one strictly desirable option. I therefore seems sensible to impose that .

To explain the implications of this axiom, and how it is related to lower previsions, we need a way to link the latter to sets of desirable option sets. The first step is to associate, with any coherent lower prevision , a set of desirable option sets

 KP––\coloneqq{A∈Q:(∃u∈A)P––(u)>0}, (7)

The coherence of this set can be easily verified [pmlr-v103-de-bock19b, Propositions 8 and 24]. More generally, with any non-empty set of coherent lower previsions, we associate a set of desirable option sets

 KP\coloneqq⋂{KP––:P––∈P}. (8)

Coherence is again easily verified; it follows directly from the coherence of and the fact that coherence is preserved under taking intersections. The final tool that we need to explain the implications of strong Archimedeanity, does the opposite; it starts with a coherent set of desirable option sets , and associates a set of coherent lower previsions with it, defined by

 P––(K)\coloneqq{P––∈P––:K⊆KP––}.

If is strongly Archimedean, then as the following result shows, serves as a mathematically equivalent representation for , from wich can be recovered through Equation (8). The representing set will then furthermore be closed with respect to the topology induced by pointwise convergence.

###### Theorem 4.1

[pmlr-v103-de-bock19b, Theorem 28 and Proposition 24] A set of desirable option sets is strongly Archimedean if and only if there is some non-empty closed set of coherent lower previsions such that . Closure is with respect to pointwise convergence, and the largest such set  is then .

If the representing coherent lower previsions in or were linear, this result would already brings us very close to decision rules based on sets of linear previsions—or sets of probability measures. As we will see further on in Section 7, this can be achieved by imposing an additional axiom called mixingness. Before we do so, however, we will do away with the closure condition in Theorem 4.1, as it is overly restrictive. Imagine for example that we are modelling a subject’s uncertainty about the outcome of a coin toss, and that she beliefs the coin to be unfair. In terms of probabilities, this would mean that her probability for heads is different from one half. Strong Archimeanity is not compatible with such an assessment, as the set of probability measures that satisfy this (strict) probability constraint is not closed. Our first main contribution will consist in resolving this issue, by suitably modifying the notion of strong Archimedeanity.

## 5 Archimedean Sets of Desirable Option Sets

At first sight, it may seem as if 1 is the only way in which 4 can be translated to option sets. There is however also a second, far more subtle approach.

The crucial insight on which this second approach is based is that our interpretation in terms of strict desirability does not require to be known; it only imposes the existence of such an . Consider a subject whose uncertainty is represented by a set of desirable option sets and let us adopt an interpretation in terms of strict desirability. This implies that the option sets are option sets that, according to her beliefs, contain at least one strictly desirable option . As a consequence of 4, this implies that she beliefs that there is some real such that is strictly desirable. Hence, she believes that there is some real such that contains at least one strictly desirable option. Strong Archimedeanity, at that point, concludes that . However, this is only justified if our subject knows . If she doesn’t know , but only believes that there is such an , then there is no single for which she believes that contains at least one strictly desirable option. Since the option sets in are options sets for which our subject believes that they contain at least one strictly desirable option, it follows that need not necessarily imply that for some . Strong Archimedeanity is therefore indeed, as its name suggests, a bit too strong for our purposes.

So if we can’t infer that , what is it then that we can infer from and 4? As explained above, the only thing that can be inferred is that for any , there is some such that contains at least one strictly desirable option. Let us denote this specific epsilon by . Crucially, we may not know—or rather, our subject may not know—the specific value of . Nevertheless, any inferences we can make without knowing the specific value of , can and should still be made. Our approach will therefore consist in finding out what inferences can be made for a specific choice of the , to do this for every such choice, and to then only consider those inferences that can be made regardless of the specific choice of .

To formalize this, we consider the set of all functions that associate a strictly positive real with every option set in . As a consequence of our interpretation, we know that there is at least one such that, for every , contains a strictly desirable option.

Let us now assume for a minute that our subject does know for which specific in this is the case. In order to be compatible with 13, the resulting assessment

 Kϵ\coloneqq{A−ϵ(A):A∈K} (9)

should then be consistent with coherence, meaning that there is at least one coherent set of desirable option sets that includes . Whenever this is the case, then as explained in Section 3, we can use coherence to extend the assessment to the unique smallest coherent set of desirable option sets that incudes it: the natural extension of . Based on the assessment and coherence, each of the option sets in

must necessarily contain a strictly desirable option. Hence, still assuming for the moment that our subject knows

, it follows that every option set in should belong to .

Our subject may not know though; all we can infer from 4 is that there must be at least one for which the above is true. Let us denote this specific—but possibly unkown— by . Then as argued above, for every option set in , it follows from our interpretation that should also belong to . Since we don’t know , however, we don’t know for which option sets this is the case. What we can do though, is to consider those option sets that belong to for every possible . For those option sets, regardless of whether we know or not, it trivially follows that , and therefore, that should belong to . Any coherent set of desirable option sets that satisfies this property, we will call Archimedean.

###### Definition 6 (Archimedeanity for sets of desirable option sets)

A set of desirable option sets is Archimedean if it is coherent and satisfies

1. [label=.,ref=,leftmargin=*]

2. for any , if for all , then also .

Note that Archimedeanity also rules out the possibility that is inconsistent for all , for this would imply that , hence contradicting 2.

By replacing strong Archimedeanity with Archimedeanity, the condition that the representing set must be closed can be removed from Theorem 4.1, and we obtain a representation in terms of general sets of lower previsions.

###### Theorem 5.1

A set of desirable option sets is Archimedean if and only if there is some non-empty set of coherent lower previsions such that . The largest such set  is then .

The significance of this result is that it relates two very different things: sets of desirable option sets and sets of coherent lower previsions. While this may not yet be obvious, this is a major step in characterising choice functions of the form . In fact, we are nearly there. The only thing left to do is to connect choice functions with sets of desirable option sets. As we will explain in the next section, this connection comes quite naturally once we interpret choice functions in terms of (strict) desirability.

## 6 Archimedean Choice Functions

In order to provide choice functions with an interpretation, we need to explain what it means for an option to be chosen from , or alternatively, what it means for to be rejected from , in the sense that . We here adopt the latter approach. In particular, if our subject states that , we take this to mean that she is convinced that there is at least one other option in that is better that , where ‘ is better than ’ is taken to mean that is strictly desirable, or equivalently, that our subject is willing to pay a positive price to exchange the uncertain reward for .

This interpretation has two implications for . First, since , it immediately implies that should be translation invariant, in the sense that

 u∈C(A)⇔0∈C(A−u) for all A∈Q and u∈A, (10)

with . Second, for all such that , it implies that should contain at least one strictly desirable gamble. Indeed, if , then according to our interpretation, there is some such that is strictly desirable. Hence, indeed contains a strictly desirable option. For any choice function , this leads us to consider the set of desirable option sets

 KC\coloneqq{A∈Q:0∉C(A∪{0})}. (11)

According to our interpretation, each of the option sets in contains at least one strictly desirable option. Following the discussion in Section 6, we will therefore require to be Archimedean. When a choice function satisfies both of the conditions that are implied by our interpretation, we call it Archimedean.

###### Definition 7

A choice function is Archimedean if is Archimedean and is translation invariant.

Instead of deriving a set of desirable options sets from a choice function , we can also do the converse. That is, with any set of desirable option sets , we can associate a choice function , defined by

 CK(A)\coloneqq{u∈A:A⊖u∉K} for all% A∈Q, (12)

where . Similarly to , the expression for is motivated by our interpretation. Indeed, for any option , the statement that means that contains a strictly desirable option, so there is some such that is strictly desirable. This is exactly our interpretation for .

If a set of desirable option sets is Archimedean, then will be Archimedean as well. In fact, as our next result shows, every Archimedean choice function is of the form , with an Archimedean set of desirable option sets.

###### Proposition 1

Let be a choice function. Then is Archimedean if and only if there is an Archimedean set of desirable option sets such that . This set is then necessarily unique and furthermore equal to .

At this point, the hard work in characterising choice functions of the form is done. Proposition 1 relates Archimedean choice functions to Archimedean sets of desirable option sets, while Theorem 5.1 relates Archimedean sets of desirable option sets to sets of coherent lower previsions . Combining both results, we find that a choice function is Archimedean if and only if it is of the form .

###### Theorem 6.1

A choice function is Archimedean if and only if there is a non-empty set of coherent lower previsions such that . Whenever this is the case, the largest such set is .

Starting from this result, we will now proceed to axiomatise maximality and E-admissibility, by combining Archimedeanity with additional axioms.

Archimedeanity implies that a choice function is representable by a set of coherent lower previsions , in the sense that is of the form . As can be seen by comparing Equations (4) and (2), this already brings us very close E-admissibility. Indeed, all that we need in order to obtain E-admissibility is for the lower previsions in to be linear. That is, we would like the role of to be taken up by

 P(K)\coloneqq{P∈P:K⊆KP}

instead. To achieve this, we impose a property called mixingness on the Archimedean set of desirable option sets that corresponds to . For any option set , this property considers the set

 posi(A)\coloneqq{n∑i=1λiui:n∈N,λi>0,ui∈A},

of all positive linear combinations of the elements in , and requires that if any of these positive linear combinations—any mixture—is strictly desirable, then itself should contain a strictly desirable option as well.

###### Definition 8 (Mixingness)

A set of desirable option sets is mixing if it satisfies

1. [label=.,ref=,leftmargin=*]

2. if and , then also , for all ;

A choice function is called mixing if is.

As the following result shows, mixingness achieves exactly what we needed: for any coherent set of desirable option sets , it guarantees that the coherent lower previsions in are in fact linear.

###### Proposition 2

Let be a coherent set of desirable option sets that is mixing. Then for any , we have that . Hence, .

By combining this result with Theorem 6.1, it follows readily that Archimedean choice functions that are mixing always correspond to E-admissibility. The following result formalizes this. It furthermore shows that the converse is true as well, and that the largest set of linear previsions for which this correspondence with E-admissibility holds is given by .

###### Theorem 7.1

A choice function is Archimedean and mixing if and only if there is a non-empty set of linear previsions such that . The largest such set  is then .

## 8 Axiomatising Maximality

Having axiomatised E-admissibility, we now proceed to do the same for maximality. The link with Archimedeanity is not that obvious here though, because there is no immediate connection between Equations (4) and (3). Rather than focus on how to relate these two equations, we therefore zoom in on the properties of maximality itself. One such property, which is often used to illustrate the difference with E-admissibility, is that a choice function that corresponds to maximality is completely determined by its restriction to so-called binary choices—that is, choices between two options.

###### Definition 9 (Binarity)

A choice function is binary if for all and :

 u∈C(A)⇔(∀v∈A∖{u}) u∈C({u,v})

Inspired by this observation, we impose binarity as an additional axiom, alongside Archimedeanity. As the following result shows, these two conditions are necessary and sufficient for a choice function

to be of the form , hence providing an axiomatisation for decision making with maximality.

###### Theorem 8.1

A choice function is Archimedean and binary if and only if there is a non-empty set of linear previsions such that . The largest such set  is then .

The formal proof of this result is rather technical, but the basic idea behind the sufficiency proof is nevertheless quite intuitive. First, for every , the binarity of implies that for every . For every such , since is Archimedean, Theorem 6.1 furthermore implies that there is a coherent lower prevision such that . Because of Theorem 2.1, this in turn implies that there is a linear prevision such that and therefore also . The challenging part consist in showing that and establishing necessity.

## 9 An Axiomatisation for Maximising Expectation Utility

As we have seen in Sections 7 and 8, mixingness and binarity have quite a different effect on Archimedean choice functions. The former implies that they correspond to E-admissibility, while the latter leads to maximality. What is intriguing though is that the set of linear previsions is twice the same. Indeed, as can be seen from Theorem 7.1 and 8.1, we may assume without loss of generality that this set is equal to . For a choice function that is mixing and binary, we therefore find that , with . As the following result shows, this is only possible if is a singleton.

###### Proposition 3

Let be a non-empty set of linear previsions. Then if and only if consists of a single linear prevision .

As a fairly immediate consequence, we obtain the following axiomatic characterisation of choice functions that correspond to maximising expected utility.

###### Theorem 9.1

A choice function is Archimedean, binary and mixing if and only if there is a linear prevision such that .

## 10 Conclusion and Future Work

The main conclusion of this work is that choice functions, when interpreted in terms of (strict) desirability, can provide an axiomatic basis for decision making with sets of probability models. In particular, we were able to derive necessary and sufficient conditions for a choice function to correspond to either E-admissibility or maximality. As a byproduct, we also obtained a characterisation for choice functions that correspond to maximising expected utility.

The key concept on which these results were based is that of an Archimedean choice function, where Archimedeanity is itself a combination of several conditions. The first of these conditions is translation invariance; this condition is fairly simple and allows for a reduction from choice functions to sets of desirable option sets. The resulting set of desirable option sets should then satisfy two more conditions: coherence and 1. Coherence is also fairly simple, because it follows directly from the principles of desirability. The condition 1, however, is more involved, making it perhaps the least intuitive component of Archimedeanity.

The abstract character of 1 is not intrinsic to the property itself though, but rather to the framework on which it is imposed. In fact, the basic principle 4 on which 1 is based is very simple: if is strictly desirable, then there must be some positive real such that is strictly desirable as well. The reason why this simplicity does not translate to 1 is because we restrict attention to option sets that are finite. Consider for example an assessment of the form . This means that is strictly desirable and therefore implies, due to 4, that the option set contains at least one strictly desirable option. Hence, we should simply impose that this set belongs to . This is not possible though because is infinite, while our framework of sets of desirable option sets only considers finite option sets.

This situation can be remedied, and the axiom of Archimedeanity can be simplified, by developing and adopting a framework of sets of desirable options that allows for infinite option sets, and connecting it to a theory of choice funtions that chooses from possibly infinite option sets. Explaining how this works is beyond the scope and size of the present contribution though; I intend to report on those results elsewhere.

### Acknowlegements.

This work was funded by the BOF starting grant 01N04819 and is part of a larger research line on choice functions of Gert de Cooman and I. Within this line of research, this contribution is one of two parallel papers on Archimedean choice functions, one by each of us. My paper—this one—deals with the case where options are bounded real-valued functions, and is concerned with axiomatising Archimedean choice functions; in fact, I call a choice function Archimean if it satisfies those axioms. The paper of Gert defines Archimedeanity directly in terms of sets of coherent lower previsions, without providing an axiomatisation, but considers the more general case where options are elements of an abstract vector space; he extends the concept of a coherent lower prevision to this more general context an establishes links with even convexity. We would like to combine our respective results in future work.

## Appendix 0.A Proofs and Intermediate Results

The defining properties 13 of a coherent lower prevision imply various additional properties as well [walley1991, Section 2.6.1]; we will need the following two in our proofs:

1. [label=.,ref=,leftmargin=*,start=4]

2. for all constant options ;

3. for all and .

With any coherent lower prevision , we can also associate a corresponding coherent upper prevision , defined by

 ¯¯¯¯P(u)\coloneqq−P––(−u)\leavevmode\nobreak\ for all u∈L.

Linear previsions are lower previsions for which and coincide.

###### Proposition 4

[walley1991, Section 2.3.6] A coherent lower prevision on is a linear prevision if and only for all .

###### Lemma 1

For any , is coherent.

###### Proof

1 and 2 hold because due to 1. 3 holds because for any with , due to 1. Since 5 holds trivially, it remains to prove 4. So consider any and, for all and , some . Since , there is some such that . Similarly, there is some such that . It therefore follows from 3 and 2 that

 P––(λu∗,v∗u∗+μu∗,v∗v∗)≥λu∗,v∗P––(u∗)+μu∗,v∗P––(v∗)>0,

implying that

 {λu,vu+μu,vv:u∈A1,v∈A2}∈KP––.

###### Proposition 5

Let be an Archimedean set of desirable option sets. Then for any option set such that , there is a coherent lower prevision such that .

###### Proof

Let be an Archimedean set of desirable option sets and consider any option set such that .

Since is Archimedean, we know from Definition 6 that there is some such that . Because of Equation (6), this in turn implies that there must be some coherent set of desirable options such that and . For this coherent set of desirable options, we now consider the corresponding coherent lower prevision , defined by

 P––(u)\coloneqqsup{μ∈R:u−μ∈D}% \leavevmode\nobreak\ for all u∈L. (13)

That this functional is indeed a coherent lower prevision is a standard result and follows easily from the coherence of ; see for example [walley1991, Section 2.3.3]. It remains to show that and .

First assume ex absurdo that . It then follows from Equation (7) that there is an option such that . Because of Equation (13), this implies that there is some real such that . Since and 2 imply that , it therefore follows from 3 that . Since , this allows us to infer from Equation (5) that , a contradiction. Hence, it follows that .

Finally, to show that , we consider any and prove that . Consider any . Since , it follows from Equation (9) that . Since , this implies that . It therefore follows from Equation (5) that there is some such that , or equivalently, some such that . Since , it therefore follows from Equation (13) that . Since , we infer from Equation (7) that . ∎

###### Proof

For the ‘only if’ part of the statement, we assume that is an Archimedean set of desirable option sets.

For any such that , we know from Proposition 5 that there is some such that . Since due to 2, this implies that is non-empty because it contains . Furthermore, since for every , with , we also know that for every , and therefore, that . Since it follows from the definition of that also , we find that . Since is non-empty, it follows that the ‘only if’ part of the statement holds for .

For the ‘if’ part of the statement, we consider any non-empty set of coherent lower previsions such that . Since , it then follows that for all , implying that is a subset of . Hence, the largest such set is indeed . It remains to show that for any such set , is Archimedean, which is what we now set out to do.

We know from Lemma 1 that is coherent for every . Since and is non-empty, and because coherence is clearly preserved under taking non-empty intersections, it follows that coherent. According to Definition 6, the Archimedeanity of therefore hinges on whether or not it satisfies 1. We will show that it indeed satisfies this property. To that end, consider any such that for all . We need to prove that .

Consider any and any . It then follows from Equation (7) that there is some such that . Let . It then follows from 2 that , which, since , implies that . Since this is true for every , we have found some such that . Hence, since is coherent because of Lemma 1, it follows from Equation (6) that . Since by assumption, for all , we therefore have in particular that . Since this is true for every , we conclude that . ∎

###### Proof

First assume that is an Archimedean choice function, meaning that is Archimedean and is translation invariant. Consider the Archimedean set of desirable option sets . For any and , we then find that

 u∈CK(A)⇔A⊖u∉K⇔A⊖u∉KC ⇔0∈C((A⊖u)∪{0}) ⇔0∈C(A−u)⇔u∈C(A),

using Equation (11) for the first equivalence, Equation (12) for the third one, and translation invariance—Equation (10)—for the last. Hence, .

Next, consider any Archimedean set of desirable option sets such that . First observe that is translation invariant because, for all and ,

 u∈CK(A)⇔A⊖u∉K⇔(A−u)⊖0∉K⇔0∈CK(A−u),

using Equation (12) for the first and third equivalence. Next, observe that for all

 A∈KC ⇔0∉C(A∪{0}) ⇔0∉CK(A∪{0})⇔(A∪{0})⊖0∈K⇔A∖{0}∈K⇔A∈K,

using Equation (11) for the first equivalence, Equation (12) for the third equivalence and 1 and 5 for the fifth one. This implies that , thereby implying that is unique and equal to . Since is Archimedean, this also implies that is Archimedean and therefore, since we already know that is translation invariant, that is Archimedean. ∎

###### Proof

First assume that is Archimedean. It then follows from Proposition 1 that for the Archimedean set of desirable option sets . Since is Archimedean, it furthermore follows from Theorem 5.1 that , with . Hence, we find that , which is equal to because

 CKP(A) ={u∈A:A⊖u∉KP}

Conversely, consider any non-empty set of coherent lower previsions such that . It then follows from the derivation above that and from Theorem 5.1 that is Archimedean. Proposition 1 therefore implies that is Archimedean.

It remains to show that . To that end, consider any and any . Since , we know that . Since , it therefore follows from Equation (4) that there is some such that . Since , this implies that there is some such that , meaning that . Since was arbitrary, this implies that , hence . Since was arbitrary, we find that is indeed a subset of . ∎

###### Proof

Consider any and assume ex absurdo that . It then follows from Proposition 4 that there is some such that , or equivalently, such that . Since we know from 1 and 3 that

 0=P––(0)=P––(u−u)≥P––(u)+P––(−u),

this implies that . Let and define and . Then on the one hand, it follows from 2 that

 P––(u1)=P––(u−P––(u)−\nicefracϵ2)=P––(u)−P––(u)−\nicefracϵ2=−\nicefracϵ2<0

and

 P––(u2)=P––(−u+P––(u)+ϵ)=P––(−u)+P––(u)+ϵ=0,

so and therefore, since , also . On the other hand, however, since , it follows from the coherence of —and 3 and 5 in particular—that . Since , this contradicts the mixingness of . ∎

###### Proof

First assume that is Archimedean and mixing. This implies that the set of desirable option sets is Archimedean (and hence also coherent) and mixing, and therefore, it follows from Proposition 2 that . Since is Archimedean, it furthermore follows from Theorem 6.1 that , with non-empty. For every , this implies that

 CP(A) ={u∈A:(∃P∈P)(∀v∈A∖{u}) P(v−u)≤0} ={u∈A:(∃P∈P)(∀v∈A∖{u}) P(u)≥P(v)}=CEP(A), (14)

where the crucial second equality holds because the elements of are linear previsions. Hence, .

Consider now any non-empty set of linear previsions such that . It then follows from Equation (14) that </