The problem

All languages have distributional regularities: patterns which restrict what sounds can appear where, including nowhere, as determined by local syntagmatic factors independent of any particular morphemic alternations. Early Generative Phonology tended to slight the study of distributional relations in favour of morphophonemics, perhaps because word-relatedness phonology was thought to be more productive of theoretical depth, reliably leading the analyst beyond the merely observable. But over the last few decades it has become clear that much morphophonemics can be understood as accommodation to phonotactic requirements, e.g., Kisseberth (1970), Sommerstein (1974), Kiparsky (1980), Goldsmith (1993), etc. A German-like voice-neutralising alternation system resolves rapidly when the phonotactics of obstruent voicing is recognised. And even as celebrated a problem in abstractness and opacity as Yawelmani Yokuts vocalic phonology turns on a surface-visible asymmetry in height-contrasts between long and short vowels.

Distributions require nontrivial learning: the data do not explicitly indicate the nature, or even the presence, of distributional regularities, and every distributional statement goes beyond what can be observed as fact, the ‘positive evidence’. From seeing X in this or that environment the learner must somehow conclude ‘X can only appear under these conditions and never anywhere else’ – when such a conclusion is warranted.

A familiar learning hazard is immediately encountered. Multiple grammars can be consistent with the same data, grammars which are empirically distinct in that they make different predictions about other forms not represented in the data.