Abstract
In a decision problem under uncertainty, a decision maker considers a set of alternative actions whose consequences depend on uncertain factors beyond his control. Following Luce and Raiffa (Games and decisions: introduction and critical survey. Wiley, New York, 1957), we adopt a natural representation of such a situation which takes as primitives a set of conceivable actions A, a set of states S and a consequence function \(\rho {:}\,A\times S\rightarrow C\). Each action induces a map from states to consequences, a Savage act, and each mixed action induces a map from states to probability distributions over consequences, an Anscombe–Aumann act. Under an axiom of consequentialism, preferences over pure or mixed actions yield corresponding preferences over the induced acts. This observation allows us to relate the Luce–Raiffa description of a decision problem to the most common framework of modern decision theory which directly takes as primitive a preference relation over the set of all Anscombe–Aumann acts. The key advantage of the latter framework is the possibility of applying powerful convex analysis techniques as in the seminal work of Schmeidler (Econometrica 57:571–587, 1989) and the vast literature that followed. This paper shows that we can maintain the mathematical convenience of the Anscombe–Aumann framework within a description of decision problems which is closer to many applications and experiments. We argue that our framework is more expressive as it allows us to be both explicit and parsimonious about the assumed richness of the set of conceivable actions, and to directly capture preference for randomization as an expression of uncertainty aversion.
Similar content being viewed by others
Notes
See Gilboa and Marinacci (2013) for a recent survey.
Referred to as horse lotteries by Anscombe and Aumann (1963) and Anscombe–Aumann acts in the subsequent literature.
Formally, the set of all finitely supported probability distributions on C.
This is the framework they describe on p. 276. They later replace consequences \(c_{ij}\) with their utilities \(u_{ij}\).
In Eq. (1), \(\rho \left( a_{i},s_{j}\right) =c_{ij}\) for each action \(a_{i}\in A\) and each state \(s_{j}\in S\).
In other words, a decision problem is characterized by the set \(\mathcal {A}\) of actions that are available to the decision maker in a given choice situation.
This makes the conceptual status of subjective beliefs in games problematic from a decision theoretic perspective (Mariotti 1995).
By contrast, normal-form games also include a profile \(\left( u_{i}\right) _{i\in I}\) of von Neumann–Morgenstern utility functions on C.
Marschak and Radner (1972, Ch. 1) call them essentially equivalent.
See Swinkels (1989) for a detailed distinction among the various reduced forms of a game.
This equivalence class can be seen geometrically as a hyperplane.
According to the Office of the Comptroller of the Currency (1999) a derivative is ‘a financial contract whose value is derived from the performance of assets, interest rates, currency exchange rates, or indexes.’
A transitive binary relation which is also reflexive is called preorder.
That is, for every \(f{:}\,S\rightarrow C\) there exists \(a\in A\) such that \(\rho _{a}=f\). Formally, Savage assumes \(\left( A,S,C,\rho \right) =\left( C^{S},S,C,\varrho \right) \) where \(\varrho \left( f,s\right) =f\left( s\right) \) is the evaluation pairing. In this decision framework, the number of bets is \(\left| C\right| +\left| C\right| \left( \left| C\right| -1\right) \left( 2^{\left| S\right| -1}-1\right) \) while the number of acts is \(\left| C\right| ^{\left| S\right| }\). With 10 states and 10 consequences there are forty-six thousand bets and ten billion acts.
\({\mathbb {R}}_{+}^{X}\) is the set of functions from X to \({\mathbb {R}}_{+}\). With the usual abuse of notation we denote a probability distribution on X and the probability measure it induces on the set of all parts of X by the same Greek letter.
One should distinguish between mixed decision rules and behavioral decision rules. But, since we are considering finite state spaces, the distinction is immaterial because a version of Kuhn (1953) equivalence theorem holds.
See again, Gilboa and Marinacci (2013).
Formally, the embedding is \(\left[ \alpha \right] \mapsto \rho _{\alpha }\), where \(\left[ \alpha \right] \) is the \(\approx \)-equivalence class of \(\alpha \).
For example, (13) below is reduced, and the mixed actions \(\alpha =\frac{1}{2}\delta _{a_{1}}+\frac{1}{2}\delta _{a_{2}}\) and \(\beta =\frac{1}{2} \delta _{b_{1}}+\frac{1}{2}\delta _{b_{2}}\) are clearly distinct (they have disjoint support), but \(\rho _{\alpha }=\rho _{\beta }\), so that they are realization equivalent.
Note, reduction guarantees that \(\epsilon \) is well defined. Without reduction one should consider realization equivalence classes of mixed actions with support in the set of all sure actions. This is clearly possible, but leads to a notational cost which is not justified by the conceptual gain since mixed consequentialism will be maintained (see Proposition 3).
In fact, \(\rho _{\epsilon \left( \gamma \right) } \equiv \gamma \) as shown in Lemma 1 of “Appendix 2”.
The same happens in the Anscombe–Aumann framework where one writes \(\gamma \) not only to denote \(\gamma \in \Delta \left( C\right) \), but also the constant act \(\gamma _{S}\equiv \gamma \) (and \(\gamma \succsim \zeta \) means \(\gamma _{S}\succsim \zeta _{S}\) since the primitive preferences \(\succsim \) are defined on \(\Delta \left( C\right) ^{S}\)). In this case, the role of \(\epsilon \) is played by the embedding \(\gamma \hookrightarrow \gamma _{S}\) of random consequences onto constant Anscombe–Aumann acts.
Nevertheless, we remark that this is not a requirement of the Anscombe–Aumann framework, but rather a preferential assumption; in fact, Proposition 2 holds without any assumption on preferences.
As Luce and Raiffa (1957, p. 279) discuss, the choice of \(-u\) as a payoff function for nature is best seen as purely formal.
Formally, \(\alpha \geqslant _{u}\beta \) and \(\beta \geqslant _{u}\alpha \) imply \(\alpha \sim \beta \).
See “Appendix 1” for a list of the most common axioms in the Anscombe–Aumann framework.
One should write \(V_{u}\) and \(v_{u}\) instead of V and v since these functions obviously depend on u. The subscripts are omitted since u is cardinally unique.
We say that \(\alpha \left( a\right) \mu \left( s\right) \) is hybrid because it is the product of a chance \(\alpha \left( a\right) \) and a belief \(\mu \left( s\right) \). Likewise, we use the terms payoff for \(u\left( \rho \left( a,s\right) \right) \), expected payoff for the (objective) average \({\sum \nolimits _{a\in A}}\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) \) of payoffs, and expected utility for the (subjective) average \({\sum \nolimits _{s\in S}\mu }\left( s\right) {\sum \nolimits _{a\in A}}\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) \) of expected payoffs.
Again one should write \(r_{u}\) instead of r and again the subscript is omitted because of the cardinal uniqueness of u.
A fair coin here is just a random device generating two outcomes with the same 1/2 chance. The original paper of Ellsberg models it as another urn that the decision maker knows to contain 50 white balls and 50 black balls.
If the color of the drawn ball is white (resp. black), then the probability of winning by choosing this gamble is the chance that the coin toss assigns to betting on white (resp. black), and this chance is 1 / 2 since the coin is fair.
Note that, under the assumptions of Theorem 3, \(a_{1}\sim a_{2}\) implies \(\mu \left( B\right) =\mu \left( W\right) =1/2\).
See Dominiak and Schnedler (2011) for a recent perspective on uncertainty attitudes and preference for randomization.
Indeed, expected utility preferences satisfy it with indifference \(\sim \) instead of weak preference \(\succsim \).
See Strzalecki (2011) for an axiomatization.
See Ghirardato et al. (2005).
Building on Gilboa et al. (2010), another way to obtain maxmin expected utility is by replacing ‘extreme caution’ in Theorem 2 with ‘default to certainty’ with respect to the unambiguous preference
$$\begin{aligned} \alpha \succsim ^{*}\beta \iff q\alpha +(1-q)\eta \succsim q\beta +(1-q)\eta \quad \forall q\in \left[ 0,1\right] \text { and }\forall \eta \in \Delta \left( A\right) . \end{aligned}$$See Ghirardato et al. (2004).
As suggested by a referee, even on a purely theoretical level, realization equivalence appears to be very convincing when one considers a framework consisting of unambiguous sets of states, actions, and consequences. If there is ambiguity or a degree of unawareness regarding these primitive sets, it is no longer clear whether realization equivalence can be maintained as a prerequisite of consequentialism (see e.g., Karni and Viero 2014).
Again, we thank a referee for this observation.
We say ‘interpreted’ since this timeline and the timed disclosure of information are unmodeled. In principle, one could think of the decision maker committing to f and receiving the outcome of the resulting process.
Again the timeline and the timed disclosure of information are unmodeled. In principle, one could think that the decision maker commits to \(\alpha \) and receives the outcome of the resulting process.
See Kuzmics (2012) on this issue.
Seo (2009) studies the consequences of weakening this assumption.
Specifically, \(\succsim _{C}=\left\{ \left( c,d\right) \in C\times C:a_{c}\succsim a_{d}\text { for some/all }a_{c},a_{d}\in A\text { such that }\rho _{a_{c}}\equiv c,\rho _{a_{d} }\equiv d\right\} \).
In fact, \(\left\{ \left( f,g\right) \mid \alpha \succsim \beta ,\forall \alpha ,\beta \in \Delta \left( A\right) :\rho _{\alpha }=f\text { and }\rho _{\beta }=g\right\} =\left\{ \left( f,g\right) \mid \exists \alpha ,\beta \in \Delta \left( A\right) :\rho _{\alpha }\right. \left. =f\text {, }\rho _{\beta }=g\text {, and }\alpha \succsim \beta \right\} \subseteq \mathcal {F}\times \mathcal {F}\) and the relation \(\succsim _{\digamma }\) they define on \(\mathcal {F}\) is such that, if \(\alpha ,\beta \in \Delta \left( A\right) \) then \(\alpha \succsim \beta \iff \rho _{\alpha }\succsim _{\digamma }\rho _{\beta }\).
Notice that V is well defined since whenever \(\alpha _{\ell }^{\prime },\alpha _{\ell } ^{\prime \prime }\in \Delta _{\ell }\left( A\right) \) are such that \(\epsilon \left( \alpha _{\ell }^{\prime }\right) \sim \alpha \) and \(\epsilon \left( \alpha _{\ell }^{\prime \prime }\right) \sim \alpha \), then \(\alpha _{\ell }^{\prime }\sim _{\Delta \left( C\right) }\alpha _{\ell }^{\prime \prime }\) and \({\sum \nolimits _{c\in C}}\alpha _{\ell }^{\prime }\left( c\right) u\left( c\right) ={\sum \nolimits _{c\in C}}\alpha _{\ell }^{\prime \prime }\left( c\right) u\left( c\right) \). Also observe that if \(\alpha =\epsilon \left( \gamma \right) \in \Delta _{\ell }\left( A\right) \), one can choose \(\alpha _{\ell }=\gamma \) and obtain \(V\left( \epsilon \left( \gamma \right) \right) ={\sum \nolimits _{c\in C}}\gamma \left( c\right) u\left( c\right) ={\mathbb {E}}_{\gamma }\left[ u\right] \).
References
Anscombe, F.J., Aumann, R.J.: A definition of subjective probability. Ann. Math. Stat. 34, 199–205 (1963)
Arrow, K.J.: Rational choice functions and orderings. Economica 26, 121–127 (1959)
Battigalli, P., Cerreia-Vioglio, S., Maccheroni, F., Marinacci, M.: Mixed Extensions of Decision Problems under Uncertainty, IGIER Working Paper 485. Università Bocconi (2013)
Berger, A.: Statistical Decision Theory and Bayesian Analysis, 2nd edn. Springer, New York (1985)
Cerreia-Vioglio, S.: Maxmin Expected Utility on a Subjective State Space: Convex Preferences under Risk, manuscript. Columbia University (2009)
Cerreia-Vioglio, S., Ghirardato, P., Maccheroni, F., Marinacci, M., Siniscalchi, M.: Rational preferences under ambiguity. Econ. Theory 48, 341–375 (2011a)
Cerreia-Vioglio, S., Maccheroni, F., Marinacci, M., Montrucchio, L.: Uncertainty averse preferences. J. Econ. Theory 146, 1275–1330 (2011b)
Cerny, A.: Mathematical Techniques in Finance: Tools for Incomplete Markets, 2nd edn. Princeton University Press, Princeton (2009)
Debreu, G.: Theory of Value. Wiley, New York (1959)
Dekel, E., Lipman, B.L., Rustichini, A.: Representing preferences with a unique subjective state space. Econometrica 69, 891–934 (2001)
Dominiak, A., Schnedler, W.: Attitudes toward uncertainty and randomization: an experimental study. Econ. Theory. 48, 289–312 (2011)
Eichberger, J., Grant, S., Kelsey, D.: Randomization and Dynamic Consistency. manuscript, University of Exeter (2013)
Ellsberg, D.: Risk, ambiguity, and the savage axioms. Q. J. Econ. 75, 643–669 (1961)
Epstein, L.G., Halevy, Y.: No Two Experiments are Identical. manuscript, University of British Columbia (2014)
Fishburn, P.C.: Utility Theory for Decision Making. Wiley, New York (1970)
Ghirardato, P., Maccheroni, F., Marinacci, M., Siniscalchi, M.: A subjective spin on roulette wheels. Econometrica 71, 1897–1908 (2003)
Ghirardato, P., Maccheroni, F., Marinacci, M.: Differentiating ambiguity and ambiguity attitude. J. Econ. Theory 118, 133–173 (2004)
Ghirardato, P., Maccheroni, F., Marinacci, M.: Certainty independence and the separation of utility and beliefs. J. Econ. Theory 120, 129–136 (2005)
Gilboa, I., Maccheroni, F., Marinacci, M., Schmeidler, D.: Objective and subjective rationality in a multiple prior model. Econometrica 78, 755–770 (2010)
Gilboa, I., Marinacci, M.: Ambiguity and the Bayesian Paradigm. In: Acemoglu, D., Arellano, M., Dekel, E. (eds.) Advances in Economics and Econometrics: Theory and Applications. Cambridge University Press, Cambridge (2013)
Gilboa, I., Schmeidler, D.: Maxmin expected utility with non-unique prior. J. Math. Econ. 18, 141–153 (1989)
Glazer, J., Rubinstein, A.: An extensive game as a guide for solving a normal game. J. Econ. Theory 70, 32–42 (1996)
Hansen, L.P., Sargent, T.J.: Robust control and model uncertainty. Am. Econ. Rev. 91, 60–66 (2001)
Hansen, L.P., Sargent, T.J.: Robustness. Princeton University Press, Princeton (2008)
Herstein, I.N., Milnor, J.: An axiomatic approach to measurable utility. Econometrica 21, 291–297 (1953)
Karni, E., Maccheroni, F., Marinacci, M.: Ambiguity and nonexpected utility. In: Young, P., Zamir, S. (eds.) Handbook of Game Theory IV, pp. 901–948. Elsevier, Amsterdam (2014)
Karni, E., Viero, M.L.: Awareness of Unawareness: A Theory of Decision Making in the Face of Ignorance, QED Working Paper 1322, Queen’s University (2014)
Klibanoff, P.: Uncertainty, Decision, and Normal Form Games, manuscript. Northwestern University (1992)
Kreps, D.M.: A representation theorem for preference for flexibility. Econometrica 47, 565–577 (1979)
Kreps, D.M.: Notes on the Theory of Choice. Westview, Boulder (1988)
Kuhn, H.W.: Extensive games and the problem of information. In: Kuhn, H.W., Tucker, A.W. (eds.) Contributions to the Theory of Games II, pp. 193–216. Princeton, Princeton University Press (1953)
Kuzmics, C.: An Alternative Subjective Expected Utility Representation Theorem, manuscript. Bielefeld University (2012)
Lehrer, E., Teper, R.: Extension rules or what would the sage do? Am. Econ. J. Microecon. 6, 5–22 (2014)
LeRoy, S.F., Werner, J.: Principles of Financial Economics. Cambridge University Press, Cambridge (2000)
Luce, R.D., Raiffa, H.: Games and Decisions: Introduction and Critical Survey. Wiley, New York (1957)
Maccheroni, F.: Maxmin under Risk. Econ. Theory 19, 823–831 (2002)
Maccheroni, F., Marinacci, M., Rustichini, A.: Ambiguity aversion, robustness, and the variational representation of preferences. Econometrica 74, 1447–1498 (2006)
Mariotti, M.: Is bayesian rationality compatible with strategic rationality? Econ. J. 105, 1099–1109 (1995)
Marschak, J., Radner, R.: Economic Theory of Teams. Yale University Press, New Haven (1972)
Milnor, J.: Games against nature. In: Thrall, R.M., Coombs, C.H., Davis, D.L. (eds.) Decision Processes, pp. 49–59. Wiley, New York (1954)
Office of the Comptroller of the Currency: Bank Derivatives Report First Quarter, report. U.S, Department of Treasury (1999)
Ok, E.A.: Real Analysis with Economic Applications. Princeton University Press, Princeton (2007)
Raiffa, H.: Risk, ambiguity, and the savage axioms: comment. Q. J. Econ. 75, 690–694 (1961)
Rubinstein, A.: Comments on the interpretation of game theory. Econometrica 59, 909–924 (1991)
Saito, K.: Preference for Flexibility and Preference for Randomization under Ambiguity, manuscript. Caltech (2013)
Sarin, R., Wakker, P.P.: A single-stage approach to Anscombe and Aumann’s expected utility. Rev. Econ. Stud. 64, 399–409 (1997)
Savage, L.J.: The Foundations of Statistics. Wiley, New York (1954)
Schmeidler, D.: Subjective probability and expected utility without additivity. Econometrica 57, 571–587 (1989)
Seo, K.: Ambiguity and second-order belief. Econometrica 77, 1575–1605 (2009)
Strzalecki, T.: Axiomatic foundations of multiplier preferences. Econometrica 79, 47–73 (2011)
Swinkels, J.: Subgames and the Reduced Normal Form, manuscript. Princeton University (1989)
Wakker, P.P.: Prospect Theory: For Risk and Ambiguity. Cambridge University Press, Cambridge (2010)
Wald, A.: Statistical Decision Functions. Wiley, New York (1950)
Author information
Authors and Affiliations
Corresponding author
Additional information
We thank Veronica Cappelli, Karin Hendrix Blissitt, Giacomo Lanzani, Sujoy Mukerji, Kelly Gail Strada, and two anonymous referees, and the participants of the NES 20th anniversary conference (Moscow, December 2012) and D-TEA and RUD workshops (Paris, May 2013) for useful comments and suggestions. Simone Cerreia-Vioglio acknowledges the financial support of the European Research Council (advanced grant BRSCDP-TEA), Fabio Maccheroni acknowledges the financial support of the Italian Ministry of Education, Universities, and Research (Grant PRIN 20103S5RN3_005), and Massimo Marinacci acknowledges the financial support of the AXA Research Fund.
Appendices
Appendix 1: Anscombe–Aumann axioms
Here we report the axioms that are usually stated in the Anscombe–Aumann framework. In this section \(\succsim _{\mathcal {F}}\) and \(\succsim _{\mathcal {F}}^{\#}\) are binary relations on a convex subset \(\mathcal {F}\) of \(\Delta \left( C\right) ^{S}\) that contains the set \(\mathcal {C}\simeq \Delta \left( C\right) \) of all constant acts.
Axiom AA.1
(Weak order) \(\succsim _{\mathcal {F}}\) is complete and transitive.
Axiom AA.2
(Risk independence) For all \(\gamma ,\zeta ,\xi \in \mathcal {C}\), \(\gamma \sim _{\mathcal {F}}\zeta \) implies \(\dfrac{1}{2}\gamma +\dfrac{1}{2} \xi \sim _{\mathcal {F}}\dfrac{1}{2}\zeta +\dfrac{1}{2}\xi \).
Axiom AA.3
(Monotonicity) For all \(f,g\in \mathcal {F}\), if \(f(s)\succsim _{\mathcal {F}}g(s)\) for all \(s\in S\), then \(f\succsim _{\mathcal {F}}g\).
Axiom AA.4
(Continuity) For all \(f,g,h\in \mathcal {F}\), \(\left\{ q\in \left[ 0,1\right] :qf+\left( 1-q\right) g\succsim _{\mathcal {F} }h\right\} \) and \(\left\{ q\in \left[ 0,1\right] :h\succsim _{\mathcal {F} }qf+\right. \) \(\left. \left( 1-q\right) g\right\} \) are closed sets.
Axiom AA.5
(Default to certainty) For all \(f\in \mathcal {F}\) and \(\gamma \in \mathcal {C}\), \(f\not \succsim _{\mathcal {F}}^{\#}\gamma \) implies \(\gamma \succ _{\mathcal {F}}f\).
Axiom AA.6
(Independence) For all \(f,g,h\in \mathcal {F}\), \(f\sim _{\mathcal {F}}g\) implies \(\dfrac{1}{2}f+\dfrac{1}{2}h\sim _{\mathcal {F}} \dfrac{1}{2}g+\dfrac{1}{2}h\).
Axiom AA.7
(Non-triviality) There exist \(f,g\in \mathcal {F}\) such that \(f\succ _{\mathcal {F}}g\).
Axiom AA.8
(Uncertainty aversion) For all \(f,g\in \mathcal {F}\) and \(q\in \left( 0,1\right) \), \(f\sim _{\mathcal {F}}g\) implies \(qf+\left( 1-q\right) g\succsim _{\mathcal {F}}f\).
Axiom AA.9
(C-Independence) For all \(f,g\in \mathcal {F}\), \(\gamma \in \mathcal {C}\), and \(q\in \left( 0,1\right] \),
Appendix 2: Proofs and related material
Throughout this appendix, if \(\gamma \in \Delta \left( C\right) \) and \(u\in {\mathbb {R}}^{C}\), we indifferently write
Proof of Proposition 1
-
1.
For each \(a\in A\), \(\rho _{a}=\rho _{a}\) so that \(a\approx a\) and by consequentialism \(a\sim _{A}a\). Therefore \(\succsim _{A}\) is reflexive and transitive, that is, a preorder. Conversely, let \(\left( A,S,C,\rho \right) \) be reduced and \(\succsim _{A}\) be a preorder. For every \(a,b\in A\), \(a\approx b\) implies (because of reduction) \(a=b\), and reflexivity of \(\succsim _{A}\) delivers \(a\sim _{A}b\). Therefore \(\succsim _{A}\) satisfies consequentialism.
-
2.
Let \(c,d\in C\) and \(a_{c},b_{c},a_{d},b_{d}\in A\), not necessarily distinct, be such that
$$\begin{aligned} \rho \left( a_{c},\cdot \right) =\rho \left( b_{c},\cdot \right) \equiv c\text { and }\rho \left( a_{d},\cdot \right) =\rho \left( b_{d},\cdot \right) \equiv d\text {.} \end{aligned}$$(23)These sure actions exist since all sure actions are conceivable. By consequentialism, \(\succsim _{A}\) is a preorder and (23) implies \(a_{c}\sim _{A}b_{c}\) and \(a_{d}\sim _{A}b_{d}\). Therefore, by transitivity of \(\succsim _{A}\),
$$\begin{aligned} a_{c}\succsim _{A}a_{d}\iff b_{c}\succsim _{A}b_{d} \end{aligned}$$and \(\succsim _{C}\) is well defined.Footnote 48 Reflexivity and transitivity of \(\succsim _{C}\) follow from reflexivity and transitivity of \(\succsim _{A}\).
-
3.
The proof is similar to the one of the previous point, hence it is left to the reader. \(\square \)
Proof of Proposition 2
-
1.
Fix \(s\in S\). For each \(c\in C\),
$$\begin{aligned} \rho _{\delta _{a}}\left( c\mid s\right) =\delta _{a}\left( \left\{ b\in A:\rho \left( b,s\right) =c\right\} \right) =\left\{ \begin{array}{ll} 1 &{} \quad c=\rho \left( a,s\right) \\ 0 &{} \quad \text {otherwise} \end{array} \right. =\delta _{\rho \left( a,s\right) }\left( c\right) \end{aligned}$$that is \(\rho _{\delta _{a}}\left( s\right) =\delta _{\rho \left( a,s\right) }\). Write x instead of \(\delta _{x}\) if either \(x\in A\) or \(x\in C\), then
$$\begin{aligned} \digamma \left( a\right) =\digamma \left( \delta _{a}\right) =\left[ \begin{array}{c} \rho _{\delta _{a}}\left( s_{1}\right) \\ \rho _{\delta _{a}}\left( s_{2}\right) \\ \vdots \\ \rho _{\delta _{a}}\left( s_{n}\right) \end{array} \right] =\left[ \begin{array}{c} \delta _{\rho \left( a,s_{1}\right) }\\ \delta _{\rho \left( a,s_{2}\right) }\\ \vdots \\ \delta _{\rho \left( a,s_{n}\right) } \end{array} \right] =\left[ \begin{array}{c} \rho \left( a,s_{1}\right) \\ \rho \left( a,s_{2}\right) \\ \vdots \\ \rho \left( a,s_{n}\right) \end{array} \right] =\rho _{a} \end{aligned}$$as desired.
-
2.
Fix \(s\in S\). For each \(c\in C\),
$$\begin{aligned} \rho _{q\alpha +\left( 1-q\right) \beta }\left( c\mid s\right)&=\left( q\alpha +\left( 1-q\right) \beta \right) \left( \left\{ a\in A:\rho \left( a,s\right) =c\right\} \right) \\&=q\alpha \left( \left\{ a\in A:\rho \left( a,s\right) =c\right\} \right) +\left( 1-q\right) \beta \left( \left\{ a\in A:\rho \left( a,s\right) =c\right\} \right) \\&=q\rho _{\alpha }\left( c\mid s\right) +\left( 1-q\right) \rho _{\beta }\left( c\mid s\right) \end{aligned}$$that is, \(\rho _{q\alpha +\left( 1-q\right) \beta }=q\rho _{\alpha }+\left( 1-q\right) \rho _{\beta }\).
-
3.
Let D be an arbitrary nonempty subset of C. Then \(\Delta \left( D\right) \) is convex in \({\mathbb {R}}^{D}\), and its extreme points are the point-masses \(\left\{ \delta _{d}\right\} _{d\in D}\). Therefore \(\Delta \left( D\right) ^{S}\) is convex too. Next we show that its extreme points are the vectors of point-masses. If \(g\in \Delta \left( D\right) ^{S}\) is not extreme there exist \(g^{\prime },g^{\prime \prime }\in \Delta \left( D\right) ^{S}\) with \(g^{\prime }\ne g^{\prime \prime }\), say \(0\le g^{\prime }\left( d\mid s\right) <g^{\prime \prime }\left( d\mid s\right) \le 1\), and \(q\in \left( 0,1\right) \) such that \(g=qg^{\prime }+\left( 1-q\right) g^{\prime \prime }\). Then \(g\left( d\mid s\right) =qg^{\prime }\left( d\mid s\right) +\left( 1-q\right) g^{\prime \prime }\left( d\mid s\right) \in \left( 0,1\right) \) and \(g\left( s\right) \) is not a point-mass. Conversely, if \(g\in \Delta \left( D\right) ^{S}\) is not a vector of point-masses, then \(g\left( s\right) \) is not a point-mass for some \(s\in S\). Therefore, there exist \(g^{\prime }\left( s\right) ,g^{\prime \prime }\left( s\right) \in \Delta \left( D\right) \) with \(g^{\prime }\left( s\right) \ne g^{\prime \prime }\left( s\right) \) and \(q\in \left( 0,1\right) \) such that \(g\left( s\right) =qg^{\prime }\left( s\right) +\left( 1-q\right) g^{\prime \prime }\left( s\right) \). But this implies that \(g=qg^{\prime }+\left( 1-q\right) g^{\prime \prime }\) where \(g^{\prime }\) (resp. \(g^{\prime \prime }\)) is obtained replacing the s-th element \(g\left( s\right) \) of the vector g with \(g^{\prime }\left( s\right) \) (resp. \(g^{\prime \prime }\left( s\right) \)). Thus g is not extreme.
In particular, the generic extreme point of \(\Delta \left( D\right) ^{S}\) is
If part. Assume \(\left\{ \rho _{a}\right\} _{a\in A}=C^{S}\). For each \(e\in C^{S}\), there exists \(a_{e}\in A\) such that \(\rho _{a_{e}}=e\). For each \(f\in \Delta \left( C\right) ^{S}\), set \(D= {\textstyle \bigcup \nolimits _{s\in S}} \mathrm {supp}f\left( s\right) \). Then \(\Delta \left( D\right) ^{S}\) is compact in \({\mathbb {R}}^{D}\) since D is finite. By the Krein–Milman Theorem, there exists \(\phi \in \Delta \left( D^{S}\right) \) such that
where the last equality follows from point 1 of this proposition. By (24), and since \(\digamma \) is affine (point 2 of this proposition),
where \( {\displaystyle \sum \limits _{e\in D^{S}}} \phi \left( e\right) \delta _{a_{e}}\in \Delta \left( A\right) \) since \(D^{S}\) is finite and \(\Delta \left( A\right) \) is convex. Therefore, \(\digamma \) is onto.
Only if part. Assume \(\digamma \left( \Delta \left( A\right) \right) =\Delta \left( C\right) ^{S}\). For each \(e\in C^{S}\), there exists \(\alpha \in \Delta \left( A\right) \) such that
Partition the finite support of \(\alpha \) into realization equivalence classes so that
with \(\rho _{a_{i}}=\rho _{a_{i}^{\prime }}\) for all \(a_{i},a_{i}^{\prime }\in A_{i}\) and all \(i=1,2,\ldots ,k\) and \(\rho _{a_{i}}\ne \rho _{a_{j}}\) if \(a_{i}\in A_{i}\), \(a_{j}\in A_{j}\) and \(i\ne j\). Moreover, for each \(i=1,2,\ldots ,k\) arbitrarily select \(b_{i}\in A_{i}\). By (25), for each \(s\in S\),
but, for each \(i=1,2,\ldots ,k\), and all \(a_{i}\in A_{i}\),
Therefore, setting \(\beta \left( b_{i}\right) ={\sum \limits _{a_{i}\in A_{i}} }\alpha \left( a_{i}\right) \) for all \(i=1,2,\ldots ,k\) and \(\beta \left( a\right) =0\) if \(a\in A{\setminus }\left\{ b_{1},\ldots ,b_{k}\right\} \), \(\beta \in \Delta \left( A\right) \) is such that
and
The summands \(\left[ \delta _{\rho \left( b_{i},s_{1}\right) }\ \ldots \text { }\delta _{\rho \left( b_{i},s_{n}\right) }\right] ^{\intercal }\) are distinct (extreme) points in \(\Delta \left( C\right) ^{S}\) and \(\left[ \delta _{e\left( s_{1}\right) }\ \ldots \text { }\delta _{e\left( s_{n}\right) }\right] ^{\intercal }\) is an extreme point of \(\Delta \left( C\right) ^{S}\). Then \(\beta =\delta _{b}\) for some \(b\in \mathrm {supp}\beta \subseteq A\), and (26) implies \(e=\rho _{b}\). The arbitrary choice of e allows us to conclude that \(C^{S}\subseteq \left\{ \rho _{a}\right\} _{a\in A}\) and the converse inclusion is trivial. \(\square \)
Proof of Corollary 1
If \(\beta \ne \delta _{a}\) there exists \(b\ne a\) such that \(\beta \left( b\right) \ne 0\). Since the framework is reduced \(\rho _{b}\ne \rho _{a}\) and there exists \(s\in S\) such that \(\rho \left( a,s\right) \ne \rho \left( b,s\right) \). Setting \(c=\rho \left( b,s\right) \), it follows
and hence \(\rho _{\beta }\ne \rho _{\delta _{a}}\) and \(\delta _{a}\not \approx \beta \). By contrapositive \(\delta _{a}\approx \beta \Longrightarrow \beta =\delta _{a}\), and the converse is trivial. \(\square \)
Lemma 1
Let \(\left( A,S,C,\rho \right) \) be a decision framework, \(\alpha \in \Delta \left( A\right) \), \(u\in {\mathbb {R}}^{C}\), and \(\mu \in \Delta \left( S\right) \). Then:
-
1.
\(\rho _{\alpha }\left( s\right) ={\sum \limits _{a\in A}}\alpha \left( a\right) \delta _{\rho \left( a,s\right) }\) for all \(s\in S\);
-
2.
\(u\left( \rho _{\alpha }\left( s\right) \right) ={\mathbb {E}} _{\rho _{\alpha }\left( s\right) }\left[ u\right] ={\sum \limits _{a\in A} }\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) \);
-
3.
\(\int _{S}u\left( \rho _{\alpha }\left( s\right) \right) d\mu \left( s\right) ={\mathbb {E}}_{\alpha \times \mu }\left[ u\circ \rho \right] \).
Moreover, if \(\left( A,S,C,\rho \right) \) is a Luce–Raiffa framework, then
is affine and bijective. In particular:
-
4.
for each \(\alpha ={\sum \limits _{c\in C}}\alpha \left( cS\right) \delta _{cS}\in \Delta _{\ell }\left( A\right) \) the only \(\gamma \in \Delta \left( C\right) \) such that \(\epsilon \left( \gamma \right) =\alpha \) is \(\epsilon ^{-1}\left( \alpha \right) ={\sum \limits _{c\in C}}\alpha \left( cS\right) \delta _{c}\);
-
5.
for each \(\gamma \in \Delta \left( C\right) \) and each \(s\in S\), \(\rho _{\epsilon \left( \gamma \right) }\left( s\right) =\gamma \).
Proof
-
1.
Follows from points 1 and 2 of Proposition 2. Specifically,
$$\begin{aligned} \rho _{\alpha }=\digamma \left( {\displaystyle \sum \limits _{a\in A}} \alpha \left( a\right) \delta _{a}\right) = {\displaystyle \sum \limits _{a\in A}} \alpha \left( a\right) \digamma \left( \delta _{a}\right) ={\sum \limits _{a\in A}}\alpha \left( a\right) \rho _{\delta _{a}} \end{aligned}$$and therefore for each \(s\in S\)
$$\begin{aligned} \rho _{\alpha }\left( s\right) ={\sum \limits _{a\in A}}\alpha \left( a\right) \rho _{\delta _{a}}\left( s\right) ={\sum \limits _{a\in A}}\alpha \left( a\right) \delta _{\rho \left( a,s\right) }. \end{aligned}$$ -
2.
By definition, for all \(s\in S\),
$$\begin{aligned}&u\left( \rho _{\alpha }\left( s\right) \right) ={\mathbb {E}}_{\rho _{\alpha }\left( s\right) }\left[ u\right] ={\mathbb {E}}_{{\sum \limits _{a\in A} }\alpha \left( a\right) \delta _{\rho \left( a,s\right) }}\left[ u\right] ={\sum \limits _{a\in A}}\alpha \left( a\right) {\mathbb {E}}_{\delta _{\rho \left( a,s\right) }}\left[ u\right] \\&\quad ={\sum \limits _{a\in A}}\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) . \end{aligned}$$ -
3.
By definition,
$$\begin{aligned} \int _{S}u\left( \rho _{\alpha }\left( s\right) \right) d\mu \left( s\right)= & {} {\sum \limits _{s\in S}\mu }\left( s\right) {\sum \limits _{a\in A}} \alpha \left( a\right) u\left( \rho \left( a,s\right) \right) ={\sum \limits _{a\in A}\sum \limits _{s\in S}}\alpha \left( a\right) {\mu }\left( s\right) u\left( \rho \left( a,s\right) \right) \\= & {} {\mathbb {E}} _{\alpha \times \mu }\left[ u\circ \rho \right] . \end{aligned}$$If \(\left( A,S,C,\rho \right) \) is a Luce–Raiffa framework, since all sure actions are conceivable for each \(c\in C\) there exists a sure action \(a_{c}\in A\) such that \(\rho \left( a_{c},\cdot \right) \equiv c\), reduction instead guarantees that such sure action is unique and we denote it by cS. In other words, the map \(c\mapsto cS\) is well defined. It is also easy to check that its range is the set of all sure actions and that the map is injective. Therefore \(c\hookrightarrow cS\) is an embedding of C onto the set of all sure actions. The fact that \(\epsilon \) is affine and bijective immediately follows.
-
4.
Clearly \(\gamma ={\sum \limits _{c\in C}}\alpha \left( cS\right) \delta _{c}\in \Delta \left( C\right) \) and \(\gamma \left( c\right) =\alpha \left( cS\right) \) for all \(c\in C\), by definition of \(\epsilon \) it follows \(\epsilon \left( \gamma \right) ={\sum \limits _{c\in C}\gamma }\left( c\right) \delta _{cS}=\alpha \).
-
5.
By definition \(\epsilon \left( \gamma \right) ={\sum \limits _{c\in \mathrm {supp}\gamma }\gamma }\left( c\right) \delta _{cS}\), therefore, for each \(s\in S\),
$$\begin{aligned} \rho _{\epsilon \left( \gamma \right) }\left( s\right)= & {} \rho _{{\sum \limits _{c\in \mathrm {supp}\gamma }\gamma }\left( c\right) \delta _{cS}}\left( s\right) ={\sum \limits _{c\in \mathrm {supp}\gamma }\gamma }\left( c\right) \rho _{\delta _{cS}}\left( s\right) \nonumber \\= & {} {\sum \limits _{c\in \mathrm {supp}\gamma }\gamma }\left( c\right) \delta _{\rho \left( cS,s\right) }={\sum \limits _{c\in \mathrm {supp}\gamma }\gamma }\left( c\right) \delta _{c}=\gamma . \end{aligned}$$(27)\(\square \)
Proof of Theorem 1
(i) \(\Longrightarrow \) (ii). Since \(\succsim \) satisfies Axiom A.1, then
is a rational choice correspondence and generates \(\succsim \). Since \(\succsim \) also satisfies Axiom A.2, then \(\succsim _{\Delta \left( C\right) }\) satisfies Axioms 1, 2, and 3 of Herstein and Milnor (1953) and there exists \(u\in {\mathbb {R}}^{C}\) such that, if \(\gamma ,\zeta \in \Delta \left( C\right) \), then
and, by definition of \(\succsim _{\Delta \left( C\right) }\) this means
Therefore, if \(\alpha ,\beta \in \Delta \left( A\right) \), then
Now Axiom A.3 implies
moreover, if \(\alpha ,\beta \in \Delta _{\ell }\left( A\right) \), then \(\alpha =\epsilon \left( \gamma \right) \) and \(\beta =\epsilon \left( \zeta \right) \) for some \(\gamma ,\zeta \in \Delta \left( C\right) \), and
that is \(\succsim \), \(\succcurlyeq _{S}\), and \(\geqslant _{u}\) coincide on \(\Delta _{\ell }\left( A\right) \) where
This immediately implies, for each \(\mathcal {A}\in \mathcal {\wp }\left( \Delta \left( A\right) \right) \) such that \(\mathcal {A}\subseteq \Delta _{\ell }\left( A\right) \),
For a generic \(\mathcal {A}\in \mathcal {\wp }\left( \Delta \left( A\right) \right) \), by (28),
(ii) \(\Longrightarrow \) (i). If \(\succsim \) is generated by a rational choice correspondence \(\Gamma :\mathcal {\wp }\left( \Delta \left( A\right) \right) \rightarrow \mathcal {\wp }\left( \Delta \left( A\right) \right) \), then \(\succsim \) satisfies Axiom A.1. By (ii) there exists also \(u\in {\mathbb {R}}^{C}\) such that:
Therefore, if \(\alpha ,\beta \in \Delta \left( A\right) \),
but \(\left\{ \epsilon \left( \rho _{\alpha }\left( s\right) \right) ,\epsilon \left( \rho _{\beta }\left( s\right) \right) \right\} \subseteq \Delta _{\ell }\left( A\right) \) for all \(s\in S\), that is
Moreover, given \(\eta ,\kappa \in \Delta \left( A\right) \)
and (29) becomes
but, \(\rho _{\epsilon \left( \gamma \right) }\equiv \gamma \) for all \(\gamma \in \Delta \left( C\right) \), therefore
and \(\succsim \) satisfies Axiom A.3.
Finally, if \(\alpha ,\beta \in \Delta _{\ell }\left( A\right) \), then \(\alpha =\epsilon \left( \gamma \right) \) and \(\beta =\epsilon \left( \zeta \right) \) for some \(\gamma ,\zeta \in \Delta \left( C\right) \), and
but \(\mathrm {supp}\alpha ,\mathrm {supp}\beta \subseteq \left\{ cS\right\} _{c\in C}\) and by (30)
and so \(\succsim \) satisfies Axioms 2 and 3 of Herstein and Milnor (1953) on \(\Delta _{\ell }\left( A\right) \), that is, Axiom A.2.\(\square \)
Recall that if \(\gamma \in \Delta \left( C\right) \) and \(u\in {\mathbb {R}}^{C}\) we indifferently write \({\mathbb {E}}_{{\gamma }}\left[ u\right] \) or \(u\left( \gamma \right) \); with a similar abuse of notation, if \(f\in \Delta \left( C\right) ^{S}\) we denote by \(u\left( f\right) \) the element of \({\mathbb {R}}^{S}\) defined by
Lemma 2
Let \(\left( A,S,C,\rho \right) \) be a Luce–Raiffa framework, \(\mathcal {F}=\digamma \left( \Delta \left( A\right) \right) \), \(u\in {\mathbb {R}}^{C}\), and \(\succsim \) be a preorder on \(\Delta \left( A\right) \). Then:
-
1.
\(\mathcal {F}\) is a convex subset of \(\Delta \left( C\right) ^{S}\) containing all constant Anscombe–Aumann acts;
-
2.
if \(\succsim \) satisfies Axiom A.3, then \(\succsim \) satisfies mixed consequentialism;
-
3.
if all bets are conceivable, \(\left\{ u\left( f\right) :f\in \mathcal {F}\right\} =\left( \mathrm {co}\left( u\left( C\right) \right) \right) ^{S}=u\left( \Delta \left( C\right) \right) ^{S}\).
Moreover, if \(\succsim \) satisfies mixed consequentialism, then \(\succsim _{\digamma }\) is a well-defined preorder on \(\mathcal {F}\) and:
-
4.
for \(N=1,\,3,\,4,\,6,\,7,\,9,\,\succsim _{\digamma }\) satisfies Axiom AA.N if and only if \(\succsim \) satisfies Axiom A.N;
-
5.
Axiom AA.2 for \(\succsim _{\digamma }\) is implied by Axiom A.2 for \(\succsim \) (and they are equivalent under Axiom A.4);
-
6.
Axiom AA.5 for \(\succsim _{\digamma }\) is equivalent to Axiom A.5 for \(\succsim \) when \(\succsim _{\digamma }^{\#}\) is defined by
$$\begin{aligned} f\succsim _{\digamma }^{\#}g\iff \ f\left( s\right) \succsim _{\digamma }g\left( s\right) \quad \forall s\in S\text {;} \end{aligned}$$ -
7.
Axiom AA.8 for \(\succsim _{\digamma }\) implies Axiom A.8 for \(\succsim (\)and they are equivalent under Axioms A.1 and A.4).
Proof
-
1.
Follows from the fact that \(\digamma \) is affine and the fact that \(\rho _{\epsilon \left( \gamma \right) }\equiv \gamma \) for all \(\gamma \in \Delta \left( C\right) \).
-
2.
If \(\alpha ,\beta \in \Delta \left( A\right) \) and \(\alpha \approx \beta \), then \(\rho _{\alpha }\left( s\right) =\rho _{\beta }\left( s\right) \) for all \(s\in S\); since \(\succsim \) is a preorder \(\epsilon \left( \rho _{\alpha }\left( s\right) \right) \succsim \epsilon \left( \rho _{\beta }\left( s\right) \right) \), that is, \(\rho _{\alpha }\left( s\right) \succsim _{\Delta \left( C\right) }\rho _{\beta }\left( s\right) \) for all \(s\in S\), and Axiom A.3 implies \(\alpha \succsim \beta \). By a symmetric argument \(\beta \succsim \alpha \).
-
3.
Obviously, \(\left\{ u\left( f\right) :f\in \mathcal {F}\right\} \subseteq \left( \mathrm {co}\left( u\left( C\right) \right) \right) ^{S}=u\left( \Delta \left( C\right) \right) ^{S}\). Set \(U=\mathrm {co} \left( u\left( C\right) \right) \) and assume \(u\left( C\right) \) is not a singleton, otherwise the result is trivial. For every vector \(\varkappa =\left[ x_{1}\ \ldots \text { }x_{n}\right] ^{\intercal }\in U^{S}\) there exist \(c\ne d\) in C and \(q_{1},q_{2},\ldots ,q_{n}\in \left[ 0,1\right] \ \)such that
$$\begin{aligned} \left[ \begin{array}{c} x_{1}\\ x_{2}\\ \vdots \\ x_{n} \end{array} \right] =\left[ \begin{array}{c} q_{1}u\left( c\right) +\left( 1-q_{1}\right) u\left( d\right) \\ q_{2}u\left( c\right) +\left( 1-q_{2}\right) u\left( d\right) \\ \vdots \\ q_{n}u\left( c\right) +\left( 1-q_{n}\right) u\left( d\right) \end{array} \right] =\left[ \begin{array}{c} u\left( q_{1}\delta _{c}+\left( 1-q_{1}\right) \delta _{d}\right) \\ u\left( q_{2}\delta _{c}+\left( 1-q_{2}\right) \delta _{d}\right) \\ \vdots \\ u\left( q_{n}\delta _{c}+\left( 1-q_{n}\right) \delta _{d}\right) \end{array} \right] . \end{aligned}$$Set \(D=\left\{ c,d\right\} \), set \(B=\left\{ a\in A:\rho _{a}\left( S\right) \subseteq \left\{ c,d\right\} \right\} \), and consider the Luce–Raiffa framework \(\left( B,S,D,\rho \right) \). Since all the bets are conceivable in \(\left( A,S,C,\rho \right) \), then \(\left\{ \rho _{b}\right\} _{b\in B}=D^{S}\). In particular, by point 3 of Proposition 2 for each \(f\in \Delta \left( D\right) ^{S}\), there exists \(\beta _{f}\in \Delta \left( B\right) \) such that
$$\begin{aligned} \beta _{f}\left( \left\{ b\in B:\rho \left( b,s\right) =c\right\} \right) =f\left( c|s\right) \quad \forall s\in S. \end{aligned}$$Choose \(f\in \Delta \left( D\right) ^{S}\) such that \(f\left( c\mid s_{i}\right) =q_{i}=1-f\left( d\mid s_{i}\right) \) for all \(i=1,\ldots ,n\), and set \(\alpha \left( a\right) =\beta _{f}\left( a\right) \) if \(a\in B\) and \(\alpha \left( a\right) =0\) if \(a\in A{\setminus } B\). Then, \(\alpha \in \Delta \left( A\right) \) and, for all \(i=1,\ldots ,n\),
$$\begin{aligned} \rho _{\alpha }\left( c\mid s_{i}\right)&=\beta _{f}\left( \left\{ b\in B:\rho \left( b,s_{i}\right) =c\right\} \right) =f\left( c\mid s_{i}\right) =q_{i}\\ \rho _{\alpha }\left( d\mid s_{i}\right)&=f\left( d\mid s_{i}\right) =1-q_{i} \end{aligned}$$so that \(\rho _{\alpha }\left( s_{i}\right) =q_{i}\delta _{c}+\left( 1-q_{i}\right) \delta _{d}\) and \(u\left( \rho _{\alpha }\left( s_{i}\right) \right) =q_{i}u\left( c\right) +\left( 1-q_{i}\right) u\left( d\right) \), that is, \(u\left( \rho _{\alpha }\right) =\varkappa \).
Assume \(\succsim \) satisfies mixed consequentialism. Let \(f,g\in \mathcal {F}\) and \(\alpha _{f},\beta _{f},\alpha _{g},\beta _{g}\in \Delta \left( A\right) \), not necessarily distinct, be such that
These mixed actions exist since \(\mathcal {F}=\digamma \left( \Delta \left( A\right) \right) \). By mixed consequentialism, (31) implies \(\alpha _{f}\sim \beta _{f}\) and \(\alpha _{g}\sim \beta _{g}\) Therefore, by transitivity of \(\succsim \),
and \(\succsim _{\digamma }\) is well defined.Footnote 49 Reflexivity and transitivity of \(\succsim _{\digamma }\) follow from reflexivity and transitivity of \(\succsim \).
The verification of points 4, 5, and 6 is routine.
7. Assume \(\succsim _{\digamma }\) satisfies Axiom AA.8. If \(\alpha ,\beta \in \Delta \left( A\right) \) and \(\alpha \sim \beta \), then \(\rho _{\alpha } \sim _{\digamma }\rho _{\beta }\). By Axiom AA.8, \(2^{-1}\rho _{\alpha } +2^{-1}\rho _{\beta }\succsim _{\digamma }\rho _{\alpha }\), therefore \(\rho _{2^{-1}\alpha +2^{-1}\beta }\succsim _{\digamma }\rho _{\alpha }\) and \(2^{-1} \alpha +2^{-1}\beta \succsim \alpha \), so that \(\succsim \) satisfies Axiom A.8.
Conversely, assume \(\succsim \) satisfies Axioms A.1 and A.4. Next we show that Axiom A.8 implies that for all \(\alpha ,\beta \in \Delta \left( A\right) \) such that \(\alpha \sim \beta \) and all \(q\in \left( 0,1\right) \) it holds \(q\alpha +\left( 1-q\right) \beta \succsim \alpha \). Per contra, assume there exist \(\alpha ,\beta \in \Delta \left( A\right) \) and \(p\in \left( 0,1\right) \) such that \(\alpha \sim \beta \) and \(p\alpha +\left( 1-p\right) \beta \prec \alpha \). Set
Clearly \(p\in T\). Moreover, by Axioms A.1 and A.4, T is open in \(\left[ 0,1\right] \) and hence there exists O open in \({\mathbb {R}}\) such that \(T=O\cap \left[ 0,1\right] \), but \(T\subseteq \left( 0,1\right) \), therefore \(T=O\cap \left( 0,1\right) \) is open in \({\mathbb {R}}\). Therefore there exists an open interval in T that contains p. The set
is a union of pairwise overlapping open intervals, and so it is an open interval itself: \(p\in I=\left( \bar{q},\bar{r}\right) \subseteq T\subseteq \left( 0,1\right) \). If \(\bar{q}\in T\), there would exist \(\varepsilon >0\) such that \(\left( \bar{q}-\varepsilon ,\bar{q}+\varepsilon \right) \subseteq T\), and then \(p\in \left( \bar{q}-\varepsilon ,\bar{r}\right) \subseteq T\), whence
a contradiction. Therefore \(\bar{q}\notin T\) and (analogously) \(\bar{r}\notin T\), that is,
Since \(\left( \bar{q},\bar{r}\right) \subseteq T\) is nonempty, eventually, \(\bar{q}+n^{-1}\) and \(\bar{r}-n^{-1}\) belong to T, and by Axiom A.4
That is, \(\bar{q}\alpha +\left( 1-\bar{q}\right) \beta \sim \bar{r} \alpha +\left( 1-\bar{r}\right) \beta \sim \alpha \), and Axiom A.8 implies
but this is a contradiction since \(2^{-1}\bar{q}+2^{-1}\bar{r}\in \left( \bar{q},\bar{r}\right) \subseteq T\).
Now let \(f,g\in \mathcal {F}\) be such that \(f\sim _{\digamma }g\) and arbitrarily choose \(q\in \left( 0,1\right) \). Let \(\alpha ,\beta \in \Delta \left( A\right) \) be such that \(f=\rho _{\alpha }\) and \(g=\rho _{\beta }\), then \(f\sim _{\digamma }g\) implies \(\alpha \sim \beta \), Axioms A.1, A.4, and A.8 imply \(q\alpha +\left( 1-q\right) \beta \succsim \alpha \), whence \(\rho _{q\alpha +\left( 1-q\right) \beta }\succsim _{\digamma }\rho _{\alpha }\) and \(qf+\left( 1-q\right) g=q\rho _{\alpha }+\left( 1-q\right) \rho _{\beta }\succsim _{\digamma }\rho _{\alpha }=f\). As wanted. \(\square \)
Proof of Proposition 3
It immediately follows from Lemma 2. \(\square \)
In order to obtain representation (16), due to Cerreia-Vioglio et al. (2011b), we need two final pieces of notation. First, given \(u\in {\mathbb {R}}^{C}\), we denote by \(U=\mathrm {co} \left( u\left( C\right) \right) \) the smallest interval containing \(u\left( C\right) \) and by \(\mathcal {G}\left( U,\Delta \left( S\right) \right) \) the set of functions \(G:U\times \Delta \left( S\right) \rightarrow \left( -\infty ,\infty \right] \) such that:
-
1.
G is quasiconvex,
-
2.
\(\inf _{\sigma \in \Delta \left( S\right) }G\left( x,\sigma \right) =x\) for all \(x\in U\),
-
3.
G is increasing in the first component,
-
4.
the function \(\varkappa \mapsto \inf _{\sigma \in \Delta \left( S\right) }G\left( \varkappa \cdot \sigma ,\sigma \right) \) is continuous on \(U^{S}\).
Second, for each \(l\in -U\) and each \(\sigma \in \Delta \left( S\right) \), we denote by \(\mathcal {B}\left( l,\sigma \right) \) the set of all mixed actions that have Bayes risk level l under \(\sigma \), that is,
and by \(v\left( l,\sigma \right) \) the indirect payoff \(v\left( \mathcal {B}\left( l,\sigma \right) \right) \) of decision problem \(\left( \mathcal {B}\left( l,\sigma \right) ,S,C,\rho \right) \), as defined in (8).
Theorem 5
Let \(\left( A,S,C,\rho \right) \) be a Marschak–Radner framework and \(\succsim \) a binary relation on \(\Delta \left( A\right) \). The following conditions are equivalent:
-
(i)
\(\succsim \) is a non-trivial and continuous rational preference that satisfies uncertainty aversion;
-
(ii)
there exist a nonconstant \(u\in {\mathbb {R}}^{C}\) and \(G\in \mathcal {G}\left( U,\Delta \left( S\right) \right) \rightarrow \left( -\infty ,\infty \right] \) such that, if \(\alpha ,\beta \in \Delta \left( A\right) \),
$$\begin{aligned} \alpha \succsim \beta \iff \inf _{\sigma \in \Delta \left( S\right) }G\left( {\mathbb {E}}_{\alpha \times \sigma }\left[ u\circ \rho \right] ,\sigma \right) \ge \inf _{\sigma \in \Delta \left( S\right) }G\left( {\mathbb {E}}_{\beta \times \sigma }\left[ u\circ \rho \right] ,\sigma \right) . \end{aligned}$$(32)
In this case, u is cardinally unique and, for each u, the minimal element of \(\mathcal {G}\left( U,\Delta \left( S\right) \right) \) satisfying (32) is the indirect payoff
Continuous and uncertainty averse rational preferences are thus represented by
and (16) is obtained by setting
and by choosing \(\Sigma \) as the projection on \(\Delta \left( S\right) \) of the domain of R. In particular, when the minimal \(G_{u}\) described by (33) is considered,
is the indirect payoff of the decision problem in which the only available mixed actions are those with the same Bayes risk as \(\alpha \) under \(\sigma \). This is worth mentioning since the indirect payoff \(v\left( l,\sigma \right) \) can be seen as a comparative index of uncertainty aversion (see Cerreia-Vioglio et al. 2011b).
Proofs of Theorems 2, 3, 4, and 5
Let \(\succsim \) be a rational preference. As shown in the first part of the proof of Theorem 1, there exists \(u\in {\mathbb {R}}^{C}\) such that:
-
if \(\gamma ,\zeta \in \Delta \left( C\right) \), then \(\gamma \succsim _{\Delta \left( C\right) }\zeta \iff {\sum \limits _{c\in C}\gamma }\left( c\right) u\left( c\right) \ge {\sum \limits _{c\in C}\zeta }\left( c\right) u\left( c\right) \);
-
if \(\alpha ,\beta \in \Delta \left( A\right) \), then \(\alpha \succcurlyeq _{S}\beta \iff \alpha \geqslant _{u}\beta \iff {\sum \limits _{a\in A} }\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) \ge {\sum \limits _{a\in A}}\beta \left( a\right) u\left( \rho \left( a,s\right) \right) \) for all \(s\in S\);
-
if \(\alpha ,\beta \in \Delta _{\ell }\left( A\right) \), then \(\alpha \succsim \beta \iff \alpha \succcurlyeq _{S}\beta \iff {\sum \limits _{c\in C}} \alpha \left( cS\right) u\left( c\right) \ge {\sum \limits _{c\in C}} \beta \left( cS\right) u\left( c\right) \).
Moreover, if \(\succsim \) is continuous, for every \(\alpha \in \Delta \left( A\right) \) there exists \(\alpha _{\ell }\in \Delta \left( C\right) \) such that \(\alpha \sim \alpha _{\ell }\), more precisely, \(\alpha \sim \epsilon \left( \alpha _{\ell }\right) \). In fact, choosing w and m in S so that
it follows \(\epsilon \left( \rho _{\alpha }\left( m\right) \right) \succcurlyeq _{S}\alpha \succcurlyeq _{S}\epsilon \left( \rho _{\alpha }\left( w\right) \right) \). Whence \(\beta =\epsilon \left( \rho _{\alpha }\left( m\right) \right) \) and \(\eta =\epsilon \left( \rho _{\alpha }\left( w\right) \right) \) belong to \(\Delta _{\ell }\left( A\right) \) and, by Axiom A.3, \(\beta \succsim \alpha \succsim \eta \). Therefore, the nonempty and closed (by Axiom A.4) sets
cover (by Axiom A.1) the connected set \(\left[ 0,1\right] \). In particular, they cannot be disjoint, and there is \(q_{\ell }\in \left[ 0,1\right] \) such that \(q_{\ell }\eta +\left( 1-q_{\ell }\right) \beta \sim \alpha \); convexity of \(\Delta _{\ell }\left( A\right) =\epsilon \left( \Delta \left( C\right) \right) \) implies that \(q_{\ell }\eta +\left( 1-q_{\ell }\right) \beta =\epsilon \left( \alpha _{\ell }\right) \) for some \(\alpha _{\ell }\in \Delta \left( C\right) \).
If \(\alpha ,\beta \in \Delta \left( A\right) \), then
that is, the functional
represents \(\succsim \) on \(\Delta \left( A\right) \).Footnote 50
For each \(\alpha \in \Delta \left( A\right) \), by Eq. (35) and Axiom A.3, \(\epsilon \left( \rho _{\alpha }\left( m\right) \right) \succsim \alpha \succsim \epsilon \left( \rho _{\alpha }\left( w\right) \right) \) so that
Up until this point, we only assumed that \(\succsim \) satisfies Axioms A.1–A.4.
Theorem 2. If, in addition, Axiom A.5 holds, then \(\epsilon \left( \alpha _{\ell }\right) \precsim \alpha \) implies \(\alpha \succcurlyeq _{S}\epsilon \left( \alpha _{\ell }\right) \), that is \(\rho _{\alpha }\left( w\right) \succsim _{\Delta \left( C\right) } \alpha _{\ell }\), and
To sum up, \(V\left( \alpha \right) =\min _{s\in S}{\sum \nolimits _{a\in A} }\alpha \left( a\right) u\left( \rho \left( a,s\right) \right) \) for all \(\alpha \in \Delta \left( A\right) \), which proves (10). The converse is routine, hence omitted.\(\square \)
If \(\succsim \) satisfies Axioms A.1–A.4, by Lemma 2, \(\succsim _{\digamma }\) is well defined and satisfies Axioms AA.1–AA.4 on \(\mathcal {F}=\digamma \left( \Delta \left( A\right) \right) \). For each \(f=\rho _{\alpha }\in \mathcal {F}\), \(\alpha \sim \epsilon \left( \alpha _{\ell }\right) \) implies \(f\sim _{\digamma } \rho _{\epsilon \left( \alpha _{\ell }\right) }=\left( \alpha _{\ell }\right) _{S}\). In the Anscombe–Aumann jargon, \(\alpha _{\ell }\) is a certainty equivalent \(\gamma _{f}\) of \(f=\rho _{\alpha }\), more precisely,
Moreover, if \(g=\rho _{\beta }\in \mathcal {F}\), then
In particular, for every \(\gamma ,\zeta \in \Delta \left( C\right) \),
Set \(U=\mathrm {co}\left( u\left( C\right) \right) =u\left( \Delta \left( C\right) \right) \) and \(u\left( \mathcal {F}\right) =\left\{ u\left( f\right) :f\in \mathcal {F}\right\} \subseteq U^{S}\subseteq {\mathbb {R}}^{S}\), and for each \(\varkappa \in u\left( \mathcal {F}\right) \) define
Using the techniques of Cerreia-Vioglio et al. (2011a, b) it can be shown that \(I:u\left( \mathcal {F}\right) \rightarrow {\mathbb {R}}\) is well defined, monotone and normalized (that is \(I\left( x1_{S}\right) =x\) for all \(x\in U\)). Moreover, for each \(\alpha \in \Delta \left( A\right) \), \(V\left( \alpha \right) ={\mathbb {E}} _{\alpha _{\ell }}\left[ u\right] =u\left( \alpha _{\ell }\right) =I\left( u\left( \rho _{\alpha }\right) \right) \) and if \(f=\rho _{\alpha } ,g=\rho _{\beta }\in \mathcal {F}\)
Theorems 3 and 4. If \(\succsim \) satisfies Axiom A.6, then \(\succsim _{\digamma }\) satisfies Axiom AA.6 by Lemma 2, which, together with (36), implies that I is affine. Theorem 3 follows by the finite dimensional versions of the Krein-Rutman Extension Theorem (see, e.g. Ok 2007, p. 496) and the Riesz Representation Theorem. The additional assumptions of Theorem 4, non-triviality of \(\succsim \) and conceivability of all bets, guarantee that u is nonconstant and \(\mu \) is unique.\(\square \)
Theorem 5. Follows from Theorem 3 of Cerreia-Vioglio et al. (2011a, b) and Lemma 2 again, which guarantees that \(\left\{ u\left( f\right) :f\in \mathcal {F}\right\} =U^{S}\) when all bets are conceivable and that \(\succsim _{\digamma }\) satisfies Axiom AA.8 when \(\succsim \) satisfies Axioms A.1–A.4 and A.8. \(\square \)
Rights and permissions
About this article
Cite this article
Battigalli, P., Cerreia-Vioglio, S., Maccheroni, F. et al. Mixed extensions of decision problems under uncertainty. Econ Theory 63, 827–866 (2017). https://doi.org/10.1007/s00199-016-0972-5
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00199-016-0972-5