Abstract
Ancient Greek verbal morphology involved extensive allomorphy of lexical morphemes, most of which was phonologically and semantically arbitrary, lexically idiosyncratic, and functionally redundant. In the subsequent history of the language this allomorphy was reduced, partly through analogical levelling, where an allomorphic alternation is eliminated in favour of a single phonological expression of underlying meaning. This kind of reduction of arbitrary complexity is often observed in the development of morphological systems, which has inspired a common view of morphological change as being guided by universal preferences, which nudge morphological systems along paths which will lead them to a more optimal status. This paper uses some systematic empirical data from the history of Greek to put to the test two questions about analogical levelling and the role of optimisation'. Firstly, is levelling motivated by a universal preference for a one-to-one alignment of meaning and form in language? Secondly, is the direction of levelling determined by universal preferences for particular ways of marking morphosyntactic distinctions? I will argue that the answer to both questions is no: the changes in the Greek data I have examined are remarkably well predicted by language-specific, formal properties of paradigms, without the need to invoke universal preferences. These facts are best accommodated if speaker competence includes detailed probabilistic information about the predictive structure of paradigms, which has important implications for morphological theory, as well as historical linguistics.