For and against the identity hypothesis
It seems clear that the constrained model of Lexical Phonology assumed here will be unable to generate all surface differences between related dialects from a common underlying inventory and set of representations, or at least that such composite analyses will be strongly dispreferred for reasons of learnability and coherence with external evidence. In terms of the abstractness of the synchronic system proposed for individual varieties, this has obvious advantages. However, before accepting this conclusion unconditionally, we should ask what independent evidence there might be for or against panlectal phonology (see also McMahon 1992).

Variation studies were certainly not central to the early generative enterprise, where dialects of a single language were derived from the same set of underlying representations in phonology (Thomas 1967, Newton 1972, Brown 1972) or deep structures in syntax (Klima 1964); surface divergences followed from differences in the form, ordering and inventory of rules. The only controversy involved the character of the basal, underlying forms themselves, which were sometimes argued to be drawn uniquely from one dialect (Brown 1972), and sometimes to be neutral between dialects (Thomas 1967). In short, although distinct languages were permitted to differ at the underlying level, related dialects were not. Early synchronic generative linguistics thus adopted the diachronic methodology of internal and comparative reconstruction, whose practitioners aim to reduce variation to earlier invariance (Hock 1986).

Every assumption in the previous paragraph can be, and has been, challenged. First, the derivation of dialect differences from a single set of underlying forms follows from the assumption that grammars should be maximally simple and economical, and that differences in the rule system `cost' less than differences in the underlying representations. Evaluation according to simplicity, however, is based solely on internal evidence from distribution and alternation, and often conflicts with external evidence involving language change, dialect variation, speech errors and speaker judgements, for instance. The issue of simplicity relates directly to the claim that `long-term memory constraints prompt speakers to limit storage to idiosyncratic information and to maximize the computing of predictable information' (Harris and Lindsey 1995: 48) - a view which, as Harris and Lindsey continue, `has never been seriously defended in the psycholinguistic literature'. Even more interestingly for present purposes, Harris and Lindsey raise this point in a critique of underspecification theory, which is similarly predicated on the alleged need for maximal simplicity at the underlying level. As we shall see, underspecification is open to many of the same objections as the identity hypothesis; we should not therefore be tempted to retain these assumptions of simplicity purely because they allow underspecification. Indeed, Goldsmith (1995b: 17) argues that underspecification is by no means the only, or indeed the most obvious solution even if simplicity is seen as a general desideratum.

The identity hypothesis also logically includes the problematic premise that synchronic dialect differences result from changes in a language which was formerly without variation. This attitude is sometimes made explicit, as in Newton's (1972: 1) description of the dialects of Modern Greek as `the outcome of historical changes acting on an originally uniform language'. Of course, no known extant or attested language is without variation (as dialect atlases of e.g. Middle English show), and not even reconstructed languages like Proto-Indo-European are entirely homogeneous (pace Pulgram 1959, 1961); as Hock (1986: 569) observes, `isoglosses for ... different changes intersect in such a criss-crossing fashion as to suggest a single, dialectally highly diversified proto language'.

We might shrug this objection off as wilful misinterpretation of professional shorthand (on the principle that everybody knows we don't mean the predicted invariance really existed, even if some early generativists seem to get a bit carried away sometimes). It is less easy to evade the fact that defining related dialects as sharing the same underlying forms, but different languages as differing at the underlying level, prevents us from seeing dialect and language variation as the continuum that geographical and sociolinguistic investigation has shown it to be. Even the traditional family tree model of historical linguistics is based on the assumption that dialects may diverge across time and become distinct languages; but this pattern is obscured if related dialects cannot differ underlyingly, while related languages characteristically do. It follows also that the status of the basal forms of generative dialectology is unclear, especially if they are neutral between dialects. Brown (1972), in a study of Lumasaaba which involves the derivation of southern forms from northern ones, produces two highly significant disclaimers. First, she notes that `it is not suggested here that the model of Common Luma saaba phonology outlined here bears any relation to the process of language acquisition or production for any speaker of any dialect of Lumasaaba' (1972: 147). A little later, she adds that `I do not suggest that the southern dialects derive historically from any existing northern dialects, nor that the presentation [here] provides a reasonable frame work for a synchronic description of any one of the southern dialects. My intention is simply to demonstrate that the dialects can be shown to be related to each other by a small number of quite general rules' (1972: 171). The power of SGP means that this aim can easily be achieved; however, the validity of a basal level which is avowedly synchronically, diachronically and psycholinguistically inadequate or even irrelevant must surely be called into question.

Instead, LP encourages a view whereby even different speakers may have different underlying representations and rule systems, a case made convincingly by Giegerich (in press). This in turn allows potential incorporation of insights from sociolinguistics (Labov 1972, Milroy and Milroy 1985, Milroy 1992), where cumulative innovations by individual speakers are recognized as the key to understanding language variation and change. Without a way of according theoretical status to such innovations, and of modelling the shift from individual to dialect to language variation, we lose these valuable connections. Since most of the historical work in SGP preceded these sociolinguistic insights, it is perhaps not surprising that SGP lacked a coherent diachronic side. But there is no excuse now.

There seems no compelling reason to retain the identity hypothesis. Chambers and Trudgill (1980: 50), although accepting this conclusion in principle, worry that `unless differences in lexical entries are constrained in some way, it does mean that it would in theory be possible ... to incorporate totally unrelated varieties such as English and Chinese into the same system'. Of course, if different dialects are to become different languages across time, there should be a continuum between dialect and language variation, and distantly related, though not unrelated, languages may therefore show residual similarities in their grammars. The loss of a linguistic definition of dialect may also be a minor problem, since language and dialect may more fruitfully be regarded as sociopolitical rather than purely linguistic notions. Nonetheless, one important goal for future research might be an assessment of how much variation is compatible with subsumption under a common underlying system. Although I suspect that finding a general answer to that question is a forlorn hope, the development of constrained models of phonology like the one presented here may be a step in the right direction in supplying a limit to the variation which can be included in any specific case.

Nonetheless, before concluding that underlying dialectal identity is unwarranted, we should consider inter-dialectal communication. In SGP, the identity hypothesis automatically accounted for comprehension between speakers of varieties of the same language. Adaptive accommodations of non-standard towards standard forms simply involved manipulations of low-level rules: as Harris (1985: 341) documents, non-standard speakers were assumed to invoke `footstep-following' (the adoption of a rule from the target standard variety) or `step-retracing' (the loss or suppression of a rule usually implemented in the non-standard dialect but not in the target). If underlying unity is essential to allow for cross-dialectal communication and adaptive change, then a Lexical Phonology which cannot incorporate common underlying forms and derive all necessary surface differences by rule must, after all, be inadequate.

However, varieties of English may differ to an extent irreconcilable with inclusion in a common underlying system; in these cases, adaptive changes cannot be analyzed simply as manipulations of rules. Harris (1984) provides a particularly compelling syntactic case of this kind from Hiberno-English, which incorporates a four-way present-tense distinction of simple he goes, progressive he's going, iterative perfective he does go, and iterative imperfective he be's going. Harris argues that these are not borrowings from Irish, but retentions from Early Modern English, some shared with other varieties of English. He also contends that the degree of divergence from the standard is incompatible with derivation by rule from a shared source. Instead, he extends the familiar picture of the creole continuum, where `shifting between basilectal and acrolectal poles proceeds via the radical restructuring of underlying representations, not merely through the manipulation of low-level rules' (1984: 314), to the interface between standard and non-creole vernacular constructions.

Furthermore, `If we are attempting to establish a theory of language which claims to explain how native speakers understand each other, we must also investigate how it is they often misunderstand each other' (Lodge 1984a: 15). That is, communicative breakdowns do occur among speakers of different, but related varieties: the SGP assumption of common underlying forms should presumably rule out this possibility (although this is not made explicit in the SGP literature, given that communicative breakdown is a matter of performance). Harris (1984) reports mismatches and misunderstandings between even superficially similar syntactic constructions; and Lodge (1984a: 15) raises the interesting question of how speakers can understand another variety without necessarily being able to produce it, when both abilities should follow from a common underlying system. For instance, northern English speakers lacking the [ʊ]~[Λ] distinction can understand these vowels in RP, but may not be able to mimic them. Along similar lines, Lodge (1984b) presented eighty-eight undergraduates with spoken forms like [peʔɹəƗ] petrol, [mɑs] mass, [bΛƗ] bull and [stɹe:] straw, asking them whether they used these forms themselves; had heard them but would not use them; guessed that there might be English speakers who did use them; or felt they were not possible English forms. Quite a number of his informants, including some from the areas where the focus forms are attested, judged them as non-English; one Belfast student felt not even a foreigner would use [bΛƗ], a Belfast form. Lodge argues that `if forms are not accepted as being English by native speakers, then this is an indication that a panlectal approach to phonology ... is inappropriate ...' (1984b: 21).

Harris concludes that `in general it is fair to say that cross-dialectal understanding succeeds in spite of structural differences rather than because of complete structural identity' (1985: 346). To understand related varieties, speakers will, when necessary, invoke ad hoc, idiosyncratic comprehension or pattern-matching strategies. As for adaptive change, we might invoke `shifts in the selection of alternative lexical representations rather than the manipulation of synchronic process rules' (Harris 1985: 341). In other words, altering output to conform to some target standard variety involves lexeme-by-lexeme phonemic redistribution. This assumption shows affinities with Andersen's (1973) important hypotheses on abductive change, whereby learners who have restructured their underlying inventories may nonetheless innovate one-off pattern matching rules to forestall correction by older speakers retaining under lying and hence surface forms appropriate to an earlier stage of the language. These intermediate speakers may be crucial to the progress of a change, since their restructured underliers, though conflicting with their own corrected production, may make them more likely to accept novel pronunciations by the next generation.

It seems, then, that we are justified in renouncing the identity hypothesis, and in favoring phonological models which, unlike SGP, do not or, even better, cannot derive all surface dialect forms from a single under lying level. It follows that a language, in Lexical Phonological terms, must be seen as a collection of related varieties, but with no underlying identity or unity. As Lass (1987: 4) puts it:
To say that `Scots is a dialect of English' does not imply the (real) existence of an `English' of which it's a dialect. Rather that `English' is the name given to a cluster of (relatively) mutually comprehensible speech forms (the dialects) that share more features with each other than they do with any other conventionally named dialect clusters (`Dutch', `German', etc.).

If we are not tied to a notion of language as common underlying system, then we can also account for the gradual divergence of dialects becoming the gradual divergence of languages; on this analysis, dialect and language variation are only quantitatively, not qualitatively distinct. Of course, core systems (like the one Lass (1987: 5) calls `a semi-fictitious idealized ``core'' English') can be useful expository shorthand; I used just such a composite system for Scots dialects and SSE. But such core systems cover a multitude of real dialect-specific and indeed speaker-specific systems, and the extent to which they are themselves `real' in any sense will depend on what is allowed in the phonological model we are using.