A major goal of any theory of word order is to identify patterns in the variation across grammars and to propose principles that describe, predict and explain these patterns. A "burning issue" in this area, as in others, involves the (in)adequacy of current cross-linguistic principles when numerous language types are considered: there are systematic and puzzling exceptions to general principles such as the head ordering parameter; and there are patterns, e.g. hierarchically arranged data sets rather than binary options, that are not currently being predicted. With regard to explanation, there has been little interest in the "why" question in formal grammar: principles of ordering have been stipulated in what has been essentially an explanatory vacuum. Why, for example, should there be a head ordering parameter? Why are there heads at all? Why are some categories adjacent and others not? Why are some categories asymmetrically ordered whereas others are symmetrical?

In this paper I outline an approach to word order variation that provides answers to such questions and that builds on the following premise: patterns in cross-linguistic grammatical rules are highly correlated with patterns of preference in the performance of languages possessing several structures of a given type. The preferred structures of performance, in this context preferred word orders, are those that we find conventionalized in languages with fewer variants. This general theory, called the "Performance-Grammar Correspondence Hypothesis", gives us a basis for predicting and explaining cross-linguistic patterns in ordering, including the exceptions to stipulations and many hitherto unpredicted data sets.

The correspondence between performance and grammars is illustrated with respect to just one aspect of word order: adjacency. I first examine variation data from English and Japanese in which speakers have a choice between the adjacency or non-adjacency of certain categories to their heads. There are systematic preferences in performance, and an efficiency principle of "Minimize Domains" is proposed. I then show that this same principle can be found in the fixed conventions of grammars in languages with fewer options. This provides an explanation, derived from language use and processing, for why ordering universals and tendencies are the way they are and it gives us a principled and non-stipulative basis for formulating our grammatical axioms in one way rather than another and for making better empirical predictions.