In the last twenty years, there has been an explosion of research in decision theory. Various decision-making procedures have been proposed as descriptive or normative alternatives to expected utility theory, such as info-gap theory [Ben-Haim, 2006], Choquet expected utility [Asano and Kojima, 2015, Chateauneuf et al., 2003], and qualitative binary possibilistic utility [Giang and Shenoy, 2005, Weng, 2006]. Another approach is to refine expected utility theory by allowing utility functions and probability measures to contain infinitesimal or infinite elements. This refinement permits coherent modeling of Pascal’s Wager [Herzberg, 2011], and allows uncountably many pairwise disjoint events to have nonzero probability, circumnavigating some of the troubles of standard expected utility theory [Hammond, 1999]. It has also been defended on game-theoretic grounds [Hammond, 1994].

One of the most popular methods for modeling infinitesimals is Abraham Robinson’s nonstandard analysis, which enriches the field of real numbers with infinitely small and infinitely large numbers in a way that sensibly preserves the field’s underlying structure [Goldblatt, 1998, Robinson, 1996]. This extension of the reals is termed the hyperreal numbers and denoted by

We axiomatize a decision-theoretic model for agents who generally agree with Savage’s postulates but are unwilling to gamble extremely valuable goods against relatively unimportant ones at even the slimmest of odds. Our theorem clarifies the postulates a person would need to accept in order to embrace nonstandard expected utility; this highlights the gap between Bayesians and decision theorists in the school of Ellsburg [Ellsberg, 2016].

A variety of perspectives stress the advantages of working with nonstandard probability theory rather than its more granular standard counterpart. It is natural to ask what advantages the nonstandard approach has for other decision theories. Many of these theories have been axiomatized in a Savage-like manner [Sarin and Wakker, 1992, Weng, 2013]), and these axiomatizations may be adapted to a nonstandard setting by the transfer principle, theorem 3.7. We leave this question for future researchers to investigate.

We begin our paper with two motivating examples. We proceed with a brief overview of relevant constructions and theorems from nonstandard analysis. We restate Savage’s Theorem for reference, and close by stating and proving our own theorem.

Suppose a friend of yours offers you the following bet. You will roll a fair six-sided die. If the die lands on its edge, you must pay your friend $1000. Otherwise, nothing happens. You clearly should reject this bet, because it is impossible for playing to make you better off than refusing. You are at risk for losing $1000, and there is no possible reward for taking the bet. On the other hand, the probability of a fair six-sided die landing on one of its edges is 0, and so an adherent of expected utility theory would be indifferent about taking this bet. Expected utility theory has been hailed as a normative theory for decision-making [Bernoulli, 1954, Fishburn, 1970, Jallais et al., 2008, Neumann and Morgenstern, 1953, Savage, 1972], so this is a serious quandary.

This paradox occurs because in standard probability theory, an event may be completely possible and yet have probability 0. Under such circumstances, even monumental awards or penalties will be ignored in an expected utility calculation. In order to resolve this difficulty, it is natural to refine expected utility theory so that it can correctly interact with events of negligible probability, or with outcomes of superlative significance. Once we formalize such a refinement, we may ask what postulates govern this refined expected utility theory. Our theorem 5.1 definitively answers this question for nonstandard expected utility theory.

Suppose now that you are an economist considering whether it is normatively correct to obey Savage’s seven postulates for rational decision-making (see theorem 4.5). You agree that if you had time to consider every conceivable option, you would have weakly ordered preferences (S1), and your preference between two decisions should depend only on those cases where their outcomes differ (S2). You agree that your preferences between outcomes should be state-independent (S3), and that whether you prefer to bet on one event than another should not depend on exactly what prize you would earn (S4). You don’t mind excluding trivial decision-making (S5), and you agree that if you prefer every possible outcome of one decision to another one considered holistically, you should prefer the former decision to the latter(S7).

However, you are uncomfortable agreeing that you would be willing to bet your life against a penny, at any odds. In general, you are unwilling, when faced with two decisions, to finitely partition your state space so that what happens on any given partition will not reverse your preferences (S6). But you would be willing to risk catastrophe to claim a small reward if the odds in your favor were literally infinite. You wonder what decision-theoretic systems could correctly model your preferences. This curiosity might be academic, it might be economic, or it might be personal. In any case, theorem 5.1 answers this question as well.

We survey some elementary constructions and theorems from nonstandard analysis. No proofs are provided; our exposition is taken directly from chapters 2 through 4 of [Goldblatt, 1998], so curious readers are encouraged to study further there.

Definition 3.1. A nonprincipal ultrafilter on

1. For all 2. If A

3. ultrafilter)

4. For all 5. For all there exists

The existence of nonprincipal ultrafilters is guaranteed by Zorn’s lemma.

Observation 3.2. Every nonprinicipal filter is cofinite: if has finite cardinality, then

Definition 3.3. Fix U a nonprincipal ultrafilter. For S an arbitrary set, let where (

Observation 3.4. The relation is an equivalence relation on . We identify S with its image in under the diagonal embedding

For ) be a representative of ) be a representative of t. We extend a relation Similarly, we extend by letting be the equivalence class containing (). Similar extensions can be made of binary operations, Cartesian products, fibered products, and other set-theoretic constructions. All are well-defined.

is a totally ordered field which contains infinitesimal elements. In other words, we can find for every natural number n. Note that for every natural number contains infinite elements as well.

Note that a function may be hyperbounded without being bounded.

) be a representative for

is at most countable, then . On the other hand, let element containing (1, we see that is finite, so for all

One of the most important theorems in nonstandard analysis is the transfer principle, which asserts that S and its nonstandard extension have essentially the same structure.

Theorem 3.7 (Transfer Principle). Let R be a relational structure, the mathematical language of R, comprising the relation and function symbols of R together with logical connectives, existential quantifiers, and parentheses. A defined is true if and only if

Roughly what this means is that any statement about a mathematical object S is true if and only if its nonstandard analogue is true. Here is obtained from by extending the objects, relations, and functions in

As the Transfer Principle suggests, it is possible to develop nonstandard analysis on purely model-theoretic grounds, without any reference to ultrafilters. However, we will not pursue this train of thought here (but see [Robinson, 1996]).

Leonard Savage proved that any decision-maker who complied with seven plausible rationality axioms behaved as if he followed expected utility theory. We reproduce his axioms and theorem here. Readers interested in a more thorough treatment are directed to [Fishburn, 1970] and [Savage, 1972].

First, we provide some definitions. The state space S is the collection of all possible states of the world. We assume the collection S is exhaustive and mutually exclusive, so the world is identified with exactly one element of S. We take X to be space of possible outcomes, or results of the actor’s choice. Let denote the space of conceivable decisions that the actor could make. In practice, an actor will normally only be able to choose from a small subset of D, but Savage considered it reasonable to assume a normatively rational actor would have the capacity to compare any two hypothetical decisions. This is not a restrictive assumption, as we may enrich the decision space with additional options without forcing any changes to the actor’s original preferences. We also assume the actor’s menu of choices does not alter the state of the world, and neither does the particular decision makes.

The actor has preferences between possible decisions, described by the binary relation “is preferred to g.” From this primitive binary relation, we may define to hold if , we also say if the constant decision returning x is preferred to the constant decision returning y. At the outset, we make no assumptions about the structure of this preference relation.

We employ the following definitions.

Suppose that you prefer outcome x to outcome y, and you believe that event A is more likely than event B. Obtaining x if event A occurs and y otherwise is clearly preferable to obtaining x if event B occurs and y otherwise.

if whenever We read is more likely than

Now suppose you are certain event A occurs. Then when choosing between decisions, what would happen if A did not occur is irrelevant.

, we write (if for all read (is preferred to

If an event A is impossible, then what would happen if it occurred is always irrelevant.

Then there is a unique finitely additive probability measure such that for all

Proof. Theorem 14.1 of [Fishburn, 1970].

Theorem 5.1. With all the notation of Savage’s Theorem (4.5), let satisfy axioms S1, S2, S3, S4, S5, S7, and the following modification of S6: S6 then there exists a countable partition that for Then there is a unique countably additive nonstandard probability measure such that for all

Proof. Apply the transfer principle to Savage’s Theorem. For axioms S1-S5 and S7, this relabels becomes “For any , there exists ∗and a partition such that for all .” But for any given , there are at most countably many . Refining our partition if necessary, we may take countable partition. Then there exists a bijection , so we may relabel

Observe also that as S and X have no assumed structure, neither do . Then we can write . Applying the transfer principle to the assertion that for all and pairwise disjoint measurable sets in see that for disjoint measurable sets in Thus is countably additive.

satisfies conditions S1-S5, S6 , S7 above. Then we may take to be bounded.

positive affine transformation of which is bounded.

[Asano and Kojima, 2015] Asano, T. and Kojima, H. (2015). An axiomatization of Choquet expected utility with cominimum independence. Theory and Decision, 78(1):117–139.

[Ben-Haim, 2006] Ben-Haim, Y. (2006). Info-Gap Decision Theory: Decisions Under Severe Uncertainty. Academic Press.

[Bernoulli, 1954] Bernoulli, D. (1954). Exposition of a new theory on the measurement of risk. Econometrica, 22:23–36.

[Chateauneuf et al., 2003] Chateauneuf, A., Eichberger, J., and Grant, S. (2003). A simple axiomatization and constructive representation proof for Choquet expected utility. Econom. Theory, 22(4):907–915.

[Ellsberg, 2016] Ellsberg, D. (2016). Risk, Ambiguity and Decision. Routledge.

[Fishburn, 1970] Fishburn, P. C. (1970). Utility theory for decision making. Technical report, DTIC Docu- ment.

[Giang and Shenoy, 2005] Giang, P. H. and Shenoy, P. P. (2005). Two axiomatic approaches to decision making using possibility theory. European J. Oper. Res., 162(2):450–467.

[Goldblatt, 1998] Goldblatt, R. (1998). Lectures on the hyperreals, volume 188 of Graduate Texts in Mathematics. Springer-Verlag, New York. An introduction to nonstandard analysis.

[Hammond, 1994] Hammond, P. J. (1994). Elementary non-Archimedean representations of probability for decision theory and games. In Patrick Suppes: scientific philosopher, Vol. 1, volume 233 of Synthese Lib., pages 25–61. Kluwer Acad. Publ., Dordrecht.

[Hammond, 1999] Hammond, P. J. (1999). Non-Archimedean subjective probabilities in decision theory and games. Math. Social Sci., 38(2):139–156.

[Herzberg, 2011] Herzberg, F. (2011). Hyperreal expected utilities and Pascal’s wager. Logique et Anal. (N.S.), 54(213):69–108.

[Jallais et al., 2008] Jallais, S., Pradier, P.-C., and Teira, D. (2008). Facts, norms and expected utility func- tions. History of the Human Sciences, 21(2):45–62.

[Neumann and Morgenstern, 1953] Neumann, J. and Morgenstern, O. (1953). Theory of games and economic behavior. Princeton Univ. Press, Princeton, NJ, 3. ed. edition.

[Robinson, 1996] Robinson, A. (1996). Non-standard analysis. Princeton Landmarks in Mathematics. Princeton University Press, Princeton, NJ. Reprint of the second (1974) edition, With a foreword by Wilhelmus A. J. Luxemburg.

[Sarin and Wakker, 1992] Sarin, R. and Wakker, P. (1992). A simple axiomatization of nonadditive expected utility. Econometrica, 60(6):1255–1272.

[Savage, 1972] Savage, L. J. (1972). The foundations of statistics. Dover Publications, Inc., New York, revised edition.

[Weng, 2006] Weng, P. (2006). An axiomatic approach to qualitative decision theory with binary possibilistic utility. In Proceedings of the 2006 Conference on ECAI 2006: 17th European Conference on Artificial Intelligence August 29 – September 1, 2006, Riva Del Garda, Italy, pages 467–471, Amsterdam, The Netherlands, The Netherlands. IOS Press.

[Weng, 2013] Weng, P. (2013). Axiomatic Foundations of Generalized Qualitative Utility. In Multi-