On the role of distances in defining voting rules

A voting rule is an algorithm for determining the winner in an election, and there are several approaches that have been used to justify the proposed rules. One justification is to show that a rule satisfies a set of desirable axioms that uniquely identify it. Another is to show that the calculation that it performs is actually maximum likelihood estimation relative to a certain model of noise that affects voters (MLE approach). The third approach, which has been recently actively investigated, is the so-called distance rationalizability framework. In it, a voting rule is defined via a class of consensus elections (i.e., a class of elections that have a clear winner) and a distance function. A candidate c is a winner of an election E if c wins in one of the consensus elections that are closest to E relative to the given distance. In this paper, we show that essentially any voting rule is distance-rationalizable if we do not restrict the two ingredients of the rule: the consensus class and the distance. Thus distance rationalizability of a rule does not by itself guarantee that the voting rule has any desirable properties. However, we demonstrate that the distance used to rationalize a given rule may provide useful information about this rule’s behavior. Specifically, we identify a large class of distances, which we call votewise distances, and show that if a rule is rationalized via a distance from this class, many important properties of this rule can be easily expressed in terms of the underlying distance. This enables us to provide a new characterization of scoring rules and to establish a connection with the MLE framework. We also give bounds on the complexity of the winner determination problem for distance-rationalizable rules.