Variables Must be Limited to a Single Feature

In this paper, I show that applying variables in the unconstrainted way that Halle (1962) first proposed causes the representation of certain phonotactic patterns to be simplified. I go on to show that this simplification prohibits an otherwise standard MaxEnt model from being biased in ways that reflect human behavior. As a solution, I propose limiting variables to occurring in the same feature across different segments. This not only prevents the pathological predictions for learnability, but it also makes variables represent more modern theories of assimilation and dissimilation like autosegmental spreading (Goldsmith, 1976) and surface correspondence constraints (Rose & Walker, 2004).