Several seemingly different types of representations are grouped together under this heading because they have an additive character and can be analyzed by similar mathematical methods. A general theory of additive measurement is presented in Fishburn (1992b), where it is applied to a variety of contexts, including positive extensive measurement, additive utility measurement for multiattribute alternatives, difference measurement for strength of preference comparisons, threshold measurement, expected utility, and comparative probability. The paper includes a condition for X of arbitrary cardinality that is necessary and sufficient for the existence of an additive representation. I will describe its approach for threshold measurement in Section 4. The present subsection considers only comparative probability and multiattribute utility to illustrate the additive theme. I include comparative probability under the preference rubric because its relation is often defined from preference comparisons. Suppose x and y are uncertain events. Let be the gamble that pays $100 if x obtains and $0 otherwise, and similarly for . Then the approach promoted by de Finetti (1937) and Savage (1954) defines if is preferred to .
We formulate X for the present discussion as a family of subsets of a universal set . For comparative probability, is a set of states, members of X are events, and X usually includes the empty event and universal event . The event set X may or may not be closed under operations like union, intersection, and complementation, and its members can have very different cardinalities.
The set of items to be compared by preference in multiattribute utility theory is a subset A of a Cartesian product set with . Each is a nonempty set, and we assume without loss of generality that the are mutually disjoint and every appears in at least one n-tuple in A. The universal set is defined as , and
so every member of X is an n-element subset of .
Suppose is finite and on X is a weak order. The basic additive representation uses for
It is common in the multiattribute case to denote the restriction of u on by so, when , . Then, when (4) holds, it remains valid when the origin of each is translated by adding a constant to all values. For comparative probability, we assume and that the union of and is monotonic, so . Then we can take and when (4) holds, so u becomes a probability distribution on . A necessary and sufficient condition for (4) referred to as cancellation, independence, or additivity, was identified first by Kraft, Pratt and Seidenberg (1959):
CANCELLATION: For every pair and of finite sequences of members of X such that
it is false that for and for some j.
Necessity of Cancellation for (4) follows from the fact that (5) implies
Hence if (4) holds and if Cancellation is violated by for all j and for some j, summation over j on the right side of (4) followed by cancellation of identical terms leaves the contradiction that 0 > 0. Sufficiency of Cancellation for (4) follows from solution theory for finite systems of linear inequalities by way of a solution-existence theorem known by various names, including the separating hyperplane lemma, the theorem of the alternative, Farkas's lemma, and Motzkin's lemma: see, for example, Scott (1964), Fishburn (1970, 1996a), or Krantz et al. (1971). Essentially the same separation lemma applies when is only assumed acyclic or a partial order, with slight modifications in Cancellation. For example, if (4) is to hold when is replaced by , we replace the last line of Cancellation by ``it is false that for all j.''
When (4) holds for finite under weak order, u is not generally unique in any simple sense. Special conditions that are not necessary for (4) but which yield simple uniqueness forms, such as absolute uniqueness for subjective probabilities, are described in Fishburn and Roberts (1989) and Fishburn (1989, 1994). Additional discussions of Cancellation for finite appears in the next section.
Theories of additive measurement for infinite usually assume nicely structured domains, such as for additive utility or for comparative probability. Most also use existence axioms that simplify cancellation conditions, promote representational uniqueness, and facilitate the derivation or assessment of u. Examples for (4) with in the multiattribute case of both the algebraic and topological varieties are detailed in Fishburn (1970), Krantz et al. (1971) and Wakker (1989). Their cancellation conditions use only m = 2 and m = 3 in Cancellation, and their functions as defined after (4) are unique up to similar positive affine transformations.
When is infinite for the comparative probability case, the weak order representation (4) is usually replaced by (1) in conjunction with , and
so that u is a finitely additive probability measure on X. Savage's (1954) elegant axiomatization for this representation assumes , weak order, , for all x, the m = 2 part of Cancellation which says that
and an Archimedean axiom involving finite partitions of . The representing measure is unique and satisfies the following divisibility property: if then for every there is an such that . Proof are given in Fishburn (1970, 1988) as well as Savage (1954). Survey material on related axiomatizations of comparative probability appears in Section 6 in Fishburn (1994).