Location Specific Constraints in Matrix and Subordinate Clauses w Jane Grimshaw Rutgers University grimshaw@ruccs.rutgers.edu August 2006

Introduction

Matrix and subordinate clauses can be different. This paper has two goals. One is to set out a general perspective on the differences between matrix and subordinate clauses, based on constraint interaction (Prince and Smolensky 1993, 2004). The other is to solve an empirical problem with a long history: the fact that inversion of a subject and auxiliary does not occur in subordinate interrogatives in English. I argue that both goals can be at least partly satisfied in theory which posits constraint families whose members evaluate the same structural configurations in different locations. The structures that emerge as optimal can be different where a constraint evaluating a phrase in one location is ranked differently from the same constraint evaluating a phrase in a different location. The constraints on matrix and subordinate clauses are in essence the same, and they are the same as those governing the structure of all projections. The different patterns enforced by these constraints in various places within sentences results from ranking among location-neutral constraints and the two kinds of location specific constraints. The ranking governs their interaction, which in turn determines grammaticality. Two structural locations are targeted by the constraints: the highest projection of a matrix clause and the highest projection of a subordinate clause. As a result, the differences between matrix and subordinate clauses are predicted to lie at their tops, since these are the targets of the location-specific constraints (henceforth LS constraints). Projections which are not at the top of a matrix or at the top of a subordinate are subject only to the general, location-neutral version of the constraints.1 Complements to certain groups of verbs, which notoriously pattern with matrix clauses in some respects, are argued to be subject to both matrix-level and subordinate-level constraints. This proposal aims to answer the question of whether the differences between main and subordinate clauses can be understood as systematic. Most centrally, I will argue that HEAD LEFT (the constraint requiring left alignment of a head in its maximal projection) for subordinate domains (HD LFT /sub) is ranked higher in English than the general HD LFT constraint, which is separated in the ranking by several conflicting constraints. HD LFT /sub, in the cases explored here, is dominated only by OB SPEC /mtx (See Grimshaw in prep, the supplement to this paper, for further rankings.) The result is that mis-aligned heads have different consequences at the edge of a subordinate clause (where they violate HD LFT /sub) than they do elsewhere, because of the different ranking of HD LFT /sub and HD LFT itself. This unifies a variety of grammatical phenomena: the behavior of complementizers, adjunction and inversion at the edge of subordinate clauses. I argue that all three follow from the avoidance of HD LFT /sub violations in the highest projection of a subordinate clause.
Of course the properties of the highest projection could in principle have consequences for the lower, e.g. through head-to-head relationships, but these are not enforced by LS constraints, and we will not encounter examples here.

No direct evidence for the constraint H D L FT /mtx is found in the cases analyzed here, and I include it only because a general theory which does encompass all the motivated constraints seems to include this one.
This constraint is related to Pesetskys (1998) constraint LE(CP): The first pronounced word in CP is a function word related to the main verb of that CP (p 351). Both constraints are violated in a CP containing a specifier and a head. However, they do not agree in every case. Of particular importance here, LE(CP) is violated if a CP has no head, and it is violated if the (possibly extended) head of the CP is not at its left edge. Thus it is violated where either O B H D or H D L FT is violated (as a H D L FT alignment constraint with the quantification reversed would be.) As a result LE(CP) does not distinguish between the structure [W h __ IP] where the head is empty and [W h V IP ], where inversion has occurred, and the head V is not on the left: both violate the constraint. The difference between these candidates is crucial for the analysis of inversion patterns. LE(CP)has a different theoretical status too: it is not an alignment constraint, nor an instance of a general constraint like H D L FT.
LS constraint then it must also violate the general one. However, violation of the general constraint does not entail violation of an LS constraint.7 4. Complementizers in Matrix and Subordinate Clauses
English complementizers are impossible in matrix clauses, but allowed in subordinates. (Complements without that are analyzed in Section7.1.) The proposed line of explanation holds that the structure with no complementizer is preferred over the C-IP structure in matrix clauses, while the structure with a complementizer is preferred for subordinates; it is not, for example, a lexical property which determines the distribution of that. (8) a. b. c. *That the president resigned The president resigned They reported that the president resigned
Violations incurred by C-IP and IP clauses8 OB HD SPEC LFT OB SPEC w! HD LFT w IP w IP
[C [Spec I VP] ] [Spec I VP]
The grammaticality of C-IP (a C with an IP complement) for subordinates is not accounted for, since no constraint prefers the C-IP structure over the IP candidate, as (9) illustrates. Although C-IP is left-headed, and IP is not, the candidates tie on HD LFT , both having one violation in VP and one in IP. HD LFT therefore cannot be the decisive constraint. In fact, OB SPEC will be the only relevant constraint and it will always prefer the structure with no complementizer, even for subordinate clauses. However, the location of the HD LFT violations is different in the two cases. In the IP candidate, there is a violation of HD LFT in the top projection, since the specifier (ie. the subject) intervenes between I and the left edge of IP. In the CP candidate, HD LFT is satisfied in the top projection: nothing intervenes between the C and the left edge of CP, since CP has no specifier. The HD LFT violation in the top projection can be decisive in choosing between C-IP and IP in subordinate clauses if there is a LS HD LFT constraint governing subordinate projections. Since HD LFT is not decisive in choosing between the same candidates in matrix clauses, we must conclude that a violation at the left edge of a subordinate clause is not evaluated in the same way as a violation in another location. So the evidence for HD LFT /sub comes from the existence of fatal HD LFT violations where the general HD LFT constraint cannot be responsible.

For example: if the highest projections in two candidates each violate H D L FT and H D L FT /sub, a lower projection can violate H D L FT in one candidate and not in the other. The two candidates tie on H D L FT /sub, but not on H D L FT. If two candidates each have one H D L FT violation, but in different locations, the candidates then tie on H D L FT , but not on H D L FT /sub.
Throughout this paper I show violations incurred if every projection contains a specifier, a head and a complement, unless otherwise stated. Moreover, I dont show violations within VP since they are never significant in the comparisons at stake. I do indicate the location of VP-external head alignment violations.
In a matrix clause, the constraint HD LFT /sub (or any other LS constraint for subordinate projections) is vacuously satisfied, hence the highest ranked constraint on which the candidates differ is OB SPEC , which is fatally violated in the C-IP structure. This analysis is shown in (10). In a subordinate clause, HD LFT /sub is not vacuously satisfied: if the top projection contains a head that is not left-aligned with the maximal projection then HD LFT /sub is violated. Thus the IP candidate in (11) violates HD LFT /sub. Now we know that this constraint dominates OB SPEC , so that the C-IP candidate is chosen over the IP, as demonstrated in (11)9. (10) A complementizer in a matrix clause SPEC LFT a. b.L [C [Spec I VP] ] [ Spec I VP] OB HD HD LFT /sub OB SPEC w! HD LFT wIP wIP
A complementizer in a subordinate clause SPEC LFT OB HD HD LFT /sub OB SPEC * w! HD LFT w IP w IP

a. L b.

V [C [Spec I VP ]] V [ Spec I VP]
At the edge of a subordinate clause, therefore, and only there, a left-headed structure is optimal even in a grammar where OB SPEC dominates the general HD LFT constraint. (As it must in English, see (2)). HD LFT is ranked too low to force the choice of the C-IP structure. (12) HD LFT /sub >> OB SPEC >> HD LFT
The posited constraints govern structure: obligatory structural elements and alignment. They make no reference to the features or other properties of the particular elements which realize the structure.
It is possible to satisfy O B -S PEC in a subordinate clause by raising the subject into the highest specifier position, but this candidate violates H D L FT /sub, so it loses to candidate a. for the same reason as the IP candidate, namely the structure of its top projection.

However, positing HD LFT /sub, ranked above OB HD , which in turn ranks above HD LFT , prevents inversion just in a subordinate projection. (18) HD LFT /sub >> OB HD >> HD LFT
We now know, then, that HD LFT /sub dominates both OB HD and OB SPEC , which in turn dominate HD LFT. We can compare optima for the candidates in (19), all of which have a specifier, a complement, and either a head filled by movement or an empty head. (19) schematizes the cases, abstracting away from the nature (wh or negative) of the specifier in each case. (19) Inversion in a matrix, subordinate or internal projection SPEC LFT MATRIX PROJECTION: a. b. L Spec __ IP Spec V IP w! wIP wIP wCP HD LFT/sub OB HD OB SPEC HD LFT
SUBORDINATE PROJECTION: c. L d. V [Spec __ IP] V [Spec V IP] w! w wIP wIP wCP
INTERNAL PROJECTION: e. f.L V[ that [XP Spec __ IP] V [ that [XP Spec V IP] w! wIP wIP wXP
In Grimshaw 1997, I proposed that O B H D dominates S TAY , giving rise to inversion. That paper does not recognize the role that H D L FT inevitably plays in governing inversion. S TAY is an economy of movement constraint, now possibly eliminated. See Grimshaw 2006.
In the matrix projection (candidates a. and b.) HD LFT /sub is satisfied vacuously, and OB HD forcesinversion.12 In the internal projection XP (candidates e. and f.) HD LFT /sub is satisfied because the projection of that, which is the highest projection in the verbal extended projection, is left-headed. Inversion within the internal projection adds a HD LFT violation, but since HD LFT is dominated by OB HD , the inversion candidate is still the optimum13. Inversion at the top of the subordinate clause ( in candidate d.) is the one case which violates HD LFT /sub, since it is located in a subordinate projection. In this case the non-inverted candidate is optimal. With the three constraints: HD LFT , HD -LFT /sub, OB HD , there are six possible rankings. The two rankings with OB HD as the highest ranked constraint induce inversion in all three environments. The two rankings with both HD LFT constraints dominating OB HD force an empty head in all three environments. The remaining rankings split the HD LFT constraints: when just HD LFT dominates OB HD , an empty head will occur in all three environments, and when just HD LFT /sub dominates OB HD , a head must be present in every configuration except in subordinates. Polar questions also show inversion in main clauses and not in subordinates, and the solution proposed here generalizes immediately to them, provided that they have filled specifiers (containing a null operator presumably.) They thus satisfy OB SPEC , and inversion within them violates HD LFT /sub when in the relevant location. As for if and whether as heads, they will violate HD LFT /sub if they co-occur with a null operator in their specifiers. Under this analysis, faithfulness must force their presence (cf. the proposal in Bakovic and Keer 2001). A location specific member of a family of constraints, here one specific to subordinate projections, crucially ranked above a constraint which dominates the location free member of the same family, derives the failure of I to C at the top of subordinate clauses and hence in subordinate interrogatives. In this hypothesis, when a structure is eliminated it is eliminated once and for all, and this does not depend on stipulated lexical properties of any heads. All specifier-head-complement structures are eliminated for subordinate clauses, regardless of the nature of the head or the specifier: a complementizer is ruled out in a subordinate interrogative for the same reason as inversion. In contrast, analyses that eliminate inversion on the grounds that it interferes with selection (see 8.1) say nothing about why a complementizer cannot fill the head position, even if inversion is impossible. Of course all hypotheses about subordination must account for the fact that the head is a complementizer and not, for example, a raised auxiliary verb. The place to look is probably other constraints in the theory which assess the markedness of chains and of individual functional heads. Grimshaw 2006 is a preliminary work on this topic.

Later in the analysis, constraints that are specific to matrix projections will be motivated, and their ranking becomes crucial for deriving inversion in matrix projections. See Section 8 for the final ranking.
Green (1996), challenging the notion of a syntactic root vs. non-root distinction, cites (among several other kinds of example) instances of negative-induced subject auxiliary inversion in subordinate clauses, viewing these as evidence that inversion is not forbidden in subordinate clauses. The analysis given here precisely characterizes inversion as illegitimate not within subordinate clauses, but at the edge of a subordinate clause, in the projection which is subject to the H D L FT /sub constraint.
Another way to violate HD LFT /sub
The distribution of adjuncts also shows the effects of constraints relativized to syntactic domains. As established in McCloskey 1992, 2006, certain temporal adjuncts can occur at the left edge of a matrix clause, but not a subordinate clause. When next week precedes the complementizer in (20) it must be construed as modifying the higher clause. Adjunction is impossible at the left edge of a subordinate clause, a familiar location. (20) a. b. c. *They say next week that they will leave They say that next week they will leave Next week they will leave
To simplify matters slightly, I will compare only the subordinate clause structures illustrated in (20), which I will refer to as external and internal adjunction, finessing the issue of exactly where the internal adjunction should be located and why. If alignment constraints are satisfied only by alignment to the outer segment, i.e. the whole node, the adjunct in an adjunct-specifier-head sequence is perfectly aligned, but the head is worse aligned than if the adjunct is not present, and the same holds for the specifier. Adjunction to IP therefore adds a violation of HD LFT and SPEC LFT in the IP, as argued in Zepter 2000. SPEC LFT prefers the ungrammatical CP adjunction (when CP has no specifier), so it cannot explain why CP adjunction is ungrammatical. HD LFT does not select candidate b. either, since a. and b. tie on HD LFT. Neither SPEC LFT nor HD LFT selects the desired optimum. (21) Adjunction to a subordinate projection HD LFT/sub

a. b.L V [ CP Adjunct [ CP C [ IP Spec I VP ]] V [ CP C [ IP Adjunct [ IP Spec I VP]]

SPEC LFT

OB SPEC w

HD LFT

wIP wCP ww IP
However, the location of the HD LFT violations is different in the two structures; the ungrammatical candidate has a disruption of head structure in the subordinate projection; the grammatical candidate has the disruption inside an internal projection. The winning candidate satisfies HD LFT /sub, while the losing candidate violates it. Once we factor in the violation of HD LFT /sub, we see why adjunction to CP is ungrammatical and internal adjunction grammatical. Since SPEC LFT favors external adjunction and HD LFT /sub favors internal adjunction, the (previously undetermined) ranking between them will decide the optimum. (Since ranking is responsible for the choice, we expect to find cross-linguistic variation in external adjunction, a possibly dangerous prediction.) (22) HD LFT /sub >> SPEC LFT
If adjuncts are analyzed as occupying specifier positions along the lines of Cinque 1999, HD LFT /sub is still crucial. See Grimshaw in prep. for details.
Adjunction is similarly ungrammatical outside subordinate interrogatives (McCloskey 1992, 2006). (23) a. b. *She knows usually where he goes *We found out next week where they will go
This case is informative because the candidate with external adjunction does not violate HD LFT /sub, since the highest projection of the complement has no head. See Grimshaw in prep. for an analysis based on another LS constraint: SPEC LFT /sub. In essence, the solution for complementizers, inversion and adjunction is the same: the LS constraint dominates a conflicting constraint which itself dominates the general constraint. The ranking of the LS constraint HD LFT /sub above OB SPEC chooses a CP for subordinate propositions. The ranking of the same LS constraint above OB HD derives the distribution of inversion. Its ranking above SPEC LFT forces internal over external adjunction. 7. Complements, main and root clauses Certain apparent complements show properties which seem intermediate between main clauses and true complements. Examples include complements with no complementizers, those with structures such as V2in many Germanic languages, and those with topicalization. (The verbs which take such complements may be related to the verbs which allow extraction from their complements, the bridge verbs of Erteshik-Shir 1973.14 ) One thread that runs through the literature on the topic, and the one that I will follow up on here, is the idea that these are in some sense, not complements. For instance, they are clause types which have illocutionary force (Hooper and Thompson 1973), or have a paratactic relation to the higher clause (de Haan 2001), or have a different structure. Emonds (2004) proposes that certain clauses with root-like properties can be Discourse Projections, making them structurally parallel to true matrix clauses. Furthermore research on the semantics of interrogative and propositional complements supports the conclusion that, crudely speaking, the complements of wonder, ask, inquire, which can only be interrogative in form, are indeed questions, like matrix questions. Complements of verbs like know, find out, discover, on other hand denote sets of propositions. Similarly, complements of verbs like say, think, hope are really assertions, while complements of verbs like realize and deduce are not. I refer the reader to McCloskeys discussion of these points, which I have drawn on freely here. I will refer to the clauses at issue as subordinate roots, or s-roots. I will refer to true complements by the obvious nomenclature, and reserve matrix for the highest projection of a clausal structure. What is the grammar of an s-root? The answer I propose here is that the structure of s-root clauses is a consequence of the constraints responsible for the syntax of matrix clauses, the constraints responsible for the syntax of purely subordinate clauses, and the ranking among them. This is because, I propose, s-root clauses are subject to both LS constraints governing matrix projections and LS constraints on subordinate projections. This proposal instantiates a view in which apparent complements can resemble main clauses, and extends the above analysis of propositional and interrogative complements.

See for example, Iatridou and Kroch (1992); M ller and Sternefeld (1993), who propose that the verbs are the same, and de Haan (2001), Vikner (1995), who argue that they are not.

Propositional s-roots

Propositional complements can have not just the C-IP structure analyzed so far, but also a bare structure, as in They say they will leave next week. (Doherty 1997).(The hypothesis that these are not CPs is hardly uncontroversial. See Ogawa 2001, Kishimoto 2006 for recent defenses of the CP hypothesis.) Here I pursue another angle on the patterns displayed by such complements, which takes as central the fact that while apparently embedded, these clauses resemble matrix clauses in lacking a complementizer. How s-roots clauses compare with matrix and subordinates depends on how they are evaluated by the constraints. The four obvious possibilities are: they count as matrix clauses, as subordinate clauses, as neither, or, as both. If s-root clauses are evaluated as subordinate clauses, the C-IP structure will always win. If they are evaluated as matrix clauses the IP structure will always win. If they are evaluated by neither, there is no reason to expect any relationship between structural patterns of matrix and/or subordinate projections and patterns found in s-root clauses. The final option predicts that s-roots are not always identical to a matrix, or always identical to a subordinate, but their grammar is systematically related to both. Suppose that certain verbs (say, speak, tell, think, hope.) admit s-root clauses as (perhaps pseudo-) complements, and that these count as matrix for evaluation by the constraints. (The definition of matrix projection in (4) needs to be revised. One possibility would follow the de Haan (2001:22) and Emonds (2004:85) line and posit a node dominating both matrix clauses and roots.) For an s-root clause, the desired winner is the IP structure. The two candidates tie on HD LFT. We know that HD LFT /sub dominates OB SPEC , otherwise IP structures would always be chosen over CP structures in subordinate clauses, see (11). Yet choosing the IP optimum requires the opposite ranking. Again we have a ranking paradox which can be resolved by an LS constraint, in this case, a constraint which penalizes the OB SPEC violation in the matrix projection, and thus eliminates the C-IP structure. Constraints relativized to the matrix are relevant in the root but not in the subordinate, so the right result will follow if OB SPEC /mtx has priority over HD LFT /sub but HD LFT /sub takes precedence over OB SPEC , as before. The rankings in (24) select the IP candidate in (25). (24) OB SPEC /mtx >> HD LFT /sub [>> OB SPEC from before]

Apart from the introduction of the LS constraint and ranking in (24), the analysis is unchanged. HD LFT /sub dominates SPEC LFT , OB HD and OB SPEC (the ranking among which is unknown), and all three of these dominate HD LFT.
In earlier work I suggested (Grimshaw 1997: 411) that both options were possible because they both count as optima, having the same constraint violation profiles. However, this cannot be right. In terms of just the constraints of this paper, the CP contains a violation of O B S PEC which IP does not, and the IP violates H D L FT /sub, if it is subject to the subordinate LS constraints. Analyses based on faithfulness (and in some cases neutralization) can be found in Legendre et al 1995, Bakovic and Keer 2001. Pesetsky (1998) proposes a tiedconstraint solution.
The analysis given here does not (apparently) generalize to the optionality of that in a relative clause.
No complementizer in an s- root OB SPEC/mtx HD LFT/sub SPEC LFT OB HD OB SPEC w w HD LFT
V [ s-root that IP] V [s-root IP]
The proposal made in Bakovic and Keer 2001 is entirely compatible with that made here. They suggest that the complementizer, or a set of features encoding its properties, is freely included in the input. Thus there is an input with C and an input without C. The two inputs correspond to distinct outputs, except when a markedness constraint dominates faithfulness to the input C. Placed within the present framework, the analysis will be as follows: HD LFT /sub chooses the C candidate as optimal, OB SPEC /mtx chooses the bare candidate as optimal, provided that these markedness constraints dominate the relevant faithfulness constraints, exactly as in Bakovic and Keers analysis S-roots are informative in three ways. First, they establish the need for a location specific version of the constraint OB SPEC. Second, they establish that constraints can be specific to matrix projections, as well as subordinate projections. Third, they show that the ranking between matrix LS constraints and subordinate LS constraints can be determined directly (and not just indirectly, via the rankings of each with other constraints). They can conflict in s- roots, and their rankings can potentially be determined from the conflicts. (In contrast, in a matrix projection, all /sub constraints are vacuously satisfied. In a subordinate projection all /mtx constraints are vacuously satisfied. No direct conflicts between the two sets of constraints are therefore possible.) We now have direct motivation for two OB SPEC constraints, and hypothesize a subordinate version in accord with a theory which constructs LS constraints for both matrix and subordinate projections. The full consequences of positing all three constraints are discussed in the supplement, Grimshaw in prep. (26) The OB SPEC family A matrix projection has a specifier (OB SPEC /mtx) A subordinate projection has a specifier (OB SPEC /sub) A projection has a specifier (OB SPEC )

The candidate V [ that Adj IP] is, according to this analysis, ungrammatical if the clause is an s-root. It violates O B S PEC /mtx, as does candidate c. The source of the grammatical sentence must be the true subordinate complement, and not the s-root. As a subordinate projection, the highest phrase in this analysis satisfies O B S PEC /mtx vacuously.
In order to complete the picture, it is necessary to look at adjunction to interrogative s-roots. See the supplement, Grimshaw in prep.
is allowed. McCloskey (2006) cites examples like these from various sources:20 (30) a. b. c. I wondered was he illiterate I asked him from what source could the reprisals come I wonder what is he like at all
McCloskey shows that these are not reported matrix questions, with parenthetical question-taking predicates: What should we do, I wonder. Among other points, he cites sequence of tense, pronominal binding and island effects. Not all predicates allow the inversion: find out, discover, know. do not21. These examples are also McCloskeys: (31) a. b. Ive never found out would he really have come with me *The police discovered who had they beaten up.
In the terms of the present analysis of subordination, I propose that these are s-root interrogatives, which like s-root propositions are subject to the matrix and subordinate LS constraints. We have seen that standard English selects the same structures for matrix and s-root propositions, namely IPs, but selects different structures for matrix and s-root interrogatives. The Local English dialect (McCloskeys term) selects the same structures for matrix and s-root clauses, in both the case of interrogatives and the case of propositions. I assume there is no lexical difference between the varieties of English, i.e. that both have the same s-root taking verbs in their lexicons. The grammar must be the locus of the dialect difference; the rankings of the relevant matrix and subordinate constraints determine the outcome. The same constraints that group matrix and s-root together for propositions must group s-root and subordinate together for interrogatives in Standard English(again McCloskeys term), and s-root and matrix together for Local English. We know that HD -LFT /sub >> OB HD in LocalE English (this is what prevents inversion in subordinate interrogatives such as those in (31)). So inversion cannot be forced by the general OB HD constraint. It is forced, however, by a matrix LS constraint, OB HD /mtx, which dominates HD LFT /sub and HD LFT , both of which prefer the loser in (32). (32) Inversion in an interrogative s-root: Local English (LE)
O B S PEC /mtx O B H D /mtx H D L FT /sub SPEC LFT OB HD OB SPEC HD LFT
V [s-root wh __ IP V [s-root wh V IP
I refer the reader to McCloskey 2006 for references to other work on inversion in subordinates.
The higher context matters too: Green 1996 cites the difference (pointed out in her own earlier work) between She wants to know who did I appoint and *She already knows who did I appoint. See McCloskey 2006 for discussion of such examples.

LE: OB HD /mtx >> HD LFT /sub >> OB HD
As above, polar s-root interrogatives will match wh s-roots (in both dialects) if they contain an operator in specifier position. The correct prediction for Standard English is maintained if the new constraint, OB HD /mtx is dominated by HdLft/sub. (34) Inversion in an s- root interrogative: Standard English (SE)
O B S PEC /mtx H D L FT /sub O B H D /mtx SPEC LFT OBHD OB SPEC HD LFT
V [s-root wh __ IP V [s-root wh V IP w!
Both OB HD and OB HD /mtx prefer the loser, so they must be dominated by some version of HD LFT. Since we know that HD LFT itself is ranked below OB HD , it must be HD LFT /sub which is over-riding the effects of the obligatory head constraints. (35) SE: HD LFT /sub >> OB HD , OB HD /mtx
In sum, a HD LFT constraint must both dominate and be dominated by an OB HD constraint for SE. An OB HD constraint must both dominate and be dominated by a HD LFT constraint for LE. If the dominating constraint is recognized as pertaining to a different domain than the dominated constraint, there is no paradox, and the LE/SE constrast is characterized in terms of alternative rankings of universal constraints. Since the s-root is subject to the matrix and subordinate level constraints, the LS constraints can come into direct conflict. It is the ranking of a matrix constraint, OB HD /mtx, relative to a subordinate constraint, HD LFT /sub, which is critical in distinguishing the two dialects. The ranking of HD LFT /sub over the OB HD constraints which it conflicts with in subordinates, namely OB HD /sub and OB HD , has many consequences. It excludes inversion or complementizers in subordinates, requires C in subordinate propositions, disallows adjunction to subordinate propositions. Because this is maintained in both varieties of English, observable differences between them are few. The consequences of elevating OB HD /mtx over HD LFT /sub or vice versa will be very limited, in fact visible only in s-roots. 8. Constraints and Rankings
This analysis gives direct evidence for the LS constraints in the second column of (36), and indirect evidence for the LS constraints in the third column.
(36) General Constraints Motivated Elsewhere 22 HD LFT SPEC LFT OB HD OB SPEC OB HD /mtx OB SPEC /mtx Directly Motivated LS Constraints HD LFT /sub Inferred LS Constraints HD LFT /mtx SPEC LFT /sub OB HD /sub OB SPEC /sub SPEC LFT /mtx
The rankings determined in this paper, with the LS constraints bolded, are: (37) HD LFT /sub HD LFT /sub HD LFT /sub OB SPEC /mtx OB HD /mtx HD LFT /sub >> >> >> >> >> >> OB SPEC OB HD SPEC LFT HD LFT /sub HD LFT /sub OB HD /mtx >> >> >> >> >> >> HD LFT HD LFT HD LFT OB SPEC OB HD HD LFT (propositional complements) (interrogative complements) (adjunction) (s-root declaratives) (LE s-root interrogatives) (SE s-root interrogatives)

Main clauses, subordinate clauses and s-root clauses are all different, and it seems that no theory can maintain that they are exactly the same. The issue is how to confront the differences. One way is to posit principles which arbitrarily distinguish among these clause types. PURE EP (Grimshaw 1997) is an example. The present proposal deconstructs PURE EP , deriving the generalization which it stipulates from the interaction of simple constraints. This move approaches the goal of maintaining maximal systematicity; developing a theory of the differences between main and subordinate clauses, and not a list of the differences. The strategy of positing location specific constraints is to treat the conditions governing matrix and subordinate clauses as part of a system, namely the syntax of the language as a whole, and the syntax of all languages. The limitation of location specific effects to highest projections limits inter- and intra-linguistic variation between matrix, s-root and subordinate clauses to their highest projections (and of course any consequences which follow from properties of the highest projections by virtue of further constraint interaction). In brief, for each constraint-mandated effect, s-root clauses must be either the same as both matrix and subordinate,
or the same as one of the two. It is not possible for matrix and subordinate to pattern together leaving s-roots to behave differently. (See Grimshaw in prep., the supplement, for development of this point). Families of LS constraints predict that the structural properties of matrix and subordinate clauses cannot be arbitrarily different. They can differ only in properties governed by the constraint families. A matrix clause cannot differ from a subordinate unless some LS constraint or constraint interaction enforces the difference, which can only happen if the constraints involved are part of the system of general and LS constraints. The constraint system evaluates all other candidates, and determines optima everywhere in the language (and every other language). Likewise, any general obligatory element or alignment constraint now has LS counterparts, which in turn must evaluate matrix, subordinate and root projections. Properties of matrix clauses are inextricably intertwined with properties of phrases in general. Constraints on matrix clauses are inextricably intertwined with the general theory of constraints
References Adger, David, Ccile de Cat and George Tsoulas eds. 2004. Peripheries. Kluwer. Bakovic, Eric and Edward Keer. 2001. Optionality and ineffability. Rutgers Optimality Archive 384-0300. In Legendre, et al. eds. 97-112. Beckman, Jill. 1997. Positional faithfulness, positional neutralisation and Shona vowel harmony Phonology 14. 1-46. Chomsky, Noam. 1986. Barriers. MIT Press. Cinque, Guglielmo. 1999. Adverbs and Functional Heads. Oxford University Press. Doherty, Cathal. 1993. Clauses without that: the case for bare sentential complementation in English. Doctoral dissertation. University of California, Santa Cruz. Doherty, Cathal. 1997. Clauses without complementizers: finite IP complementation in English. The Linguistic Review 14. 197-220 Emonds, Joseph. 1976. A Transformational Approach to English Syntax. Academic Press. Emonds, Joseph. 2004. Unspecified categories as the key to root constructions. In Adger et al eds. 75-120. Erteshik-Shir, Nomi. 1973. On the nature of island constraints. Doctoral dissertation. MIT. Green, Georgia. 1996. Distinguishing main and subordinate clause: the ROOT of the problem. ms. University of Illinois. http://www.linguistics.uiuc.edu/g-green/ Grimshaw, Jane. 1997. Projection heads and optimality, Linguistic Inquiry 28. 373-422. Grimshaw, Jane. 1998. Constraints on constraints in optimality theoretic syntax. Ms. Rutgers University. Grimshaw, Jane. 2000. Optimal clitic positions and the lexicon in romance clitic systems. In Legendre, et al. eds. 205-240. Grimshaw, Jane. 1991. Extended projection. Ms. Brandeis University. Revised version in Grimshaw 2005. Grimshaw, Jane. 2001. Economy of structure in OT. Rutgers Optimality Archive 434-0601. Grimshaw, Jane. 2002. Economy of structure in OT. In Carpenter, Coetzee, & de Lacy, eds. Papers in Optimality Theory II. University of Massachusetts Occasional Papers 26. 81-120. Grimshaw, Jane. 2005. Words and Structure. Center for the Study of Language and Information. Grimshaw, Jane. 2006. Chains as unfaithful optima. In Bakovic Ito, and McCarthy eds. Wondering at the Natural Fecundity of Things: Essays in Honor of Alan Prince. http://repositories.cdlib.org/lrc/prince/ Grimshaw, Jane. In prep. A supplement to Location Specific Constraints in Matrix and Subordinate Clauses. Grimshaw, Jane and Vieri Samek-Lodovici. 1998. Optimal subjects and subject universals. In Is the Best Good Enough?, P. Barbosa et al. eds, MIT Press. 193-219. De Haan. Germen J. 2001. More is going on upstairs than downstairs. The Journal of Comparative Germanic

Linguistics 4. 3-38. Hooper, Joan and Sandra A. Thompson. 1973. On the applicability of root transformations. Linguistic Inquiry 4. 465-498. Kishimoto, Hideki. 2006. On the existence of null complementizers in syntax. Linguistic Inquiry 37. 339345. Legendre, Graldine. 1998. Second position clitics in a V2 language: conflict resolution in Macedonian. In J. Austin and A. Lawson, (eds.). Proceedings of the 1997 ESCOL Meeting. CLC Publications, Cornell University. 139-149. Legendre, Graldine, C. Wilson, P. Smolensky, K. Homer, and W. Raymond. 1995. Optimality and Wh-Extraction. In Papers in Optimality Theory, J. Beckman, L. Walsh Dickey, and S. Urbanczyk, eds. University of Massachusetts Occasional Papers 18. GLSA, UMass, Amherst, 607-636. Legendre, Graldine, Jane Grimshaw and and Sten Vikner eds. Optimality Theoretic Syntax, MIT Press, 2000. Lombardi, Linda. 1999. Positional faithfulness and voicing assimilation in optimality theory. Natural Language and Linguistic Theory 17. 267-302. McCarthy, John. & Alan. Prince. 1993. Generalized alignment. In G. Booij & J. van Marle (eds.), Yearbookof Morphology. 79-154. Dordrecht: Kluwer. McCloskey, James. 1992. Adjunction, selection and embedded verb second. ms. UCSC. McCloskey, James. 2006. Questions and questioning in a local English. In Cross-Linguistic Research in Syntax and Semantics: Negation, Tense and Clausal Architecture, Raffaella Zanuttini, Hctor Campos, Elena Herburger and Paul H. Portner eds.,Georgetown University Press. Publication expected 2006. Mller, Gereon and Wolfgang Sternefeld. 1993. Improper movement and unambiguous binding. Linguistic Inquiry 34. 461-507. Nelson, Nicole. 2003. Asymmetric Anchoring. Ph.D. Dissertation, Rutgers University. Ogawa, Yoshiki. 2001. A unified theory of verbal and nominal projections. Oxford University Press. Pesetsky, David. 1998. Some optimality principles of sentence pronunciation.Iin Is the Best Good Enough?, P. Barbosa et al. eds, MIT Press 337-383. 1998. Prince, Alan and Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generative grammar. RuCCS Technical Report #2. Rutgers University Center for Cognitive Science, Piscataway NJ. Prince, Alan and Paul Smolensky. 2004. Optimality Theory: Constraint interaction in generative grammar. Blackwells. Alan Prince 2002. Arguing optimality. In Carpenter, Coetzee, & de Lacy, eds. Papers in Optimality Theory II. 269-304. University of Massachusetts Occasional Papers 26. Rutgers Optimality Archive 562-1102 Rizzi, Luigi. 1991. Residual verb second and the Wh-Criterion, Technical Reports in Formal and Computational Linguistics 2, University of Geneva. Rizzi, Luigi, and Ian Roberts. 1989. Complex inversion in French. Probus 1. 1-30. Samek-Lodovici, Vieri. 1998. Opposite constraints: Left and right focus-alignment in Kanakuru. Lingua 104. 111-130. Samek-Lodovici, Vieri. 2005. Prosody-syntax interactions in the expression of focus. Natural Language and Linguistic Theory. 23. 687-755. Samek-Lodovici, Vieri and Alan Prince 2005. Fundamental properties of harmonic bounding. Rutgers Optimality Archive 785-1205. Vikner, Sten. 1995. Verb movement and expletive subjects in the germanic languages. Oxford University Press. Vikner, Sten. 1999. The outer layers of embedded clauses. Handout from the 3rd workshop on optimality

Theory. Universitt Stuttgart. Zepter, Alexandra 2000. Specifiers and adjuncts. Ms. Rutgers University. Rutgers Optimality Archive 413-0900 Zoll, Cheryl. 1998. Positional asymmetries and licensing. Rutgers Optimality Archive 282-0998. w Acknowledgements: Id like to thank many individuals who have provided important input into this project over the past several years: Eric Bakovic, Eva Engels, Gisbert Fanselow, Georgia Green, Edward Keer, Soowon Kim, Alan Prince, Roger Schwarzschild, Elizabeth Traugott, Tom Werner, and Alex Zepter. Vieri Samek-Lodovici and Sten Vikner gave me detailed comments on Grimshaw 1998, an ancestor of the present paper. The project also profited from discussion with Jim McCloskey, whose work is fundamental to this paper. Audiences at the following locations and events sharpened the proposal in many ways: the Rutgers Optimality Research Group (RORG); the syntax seminar at Rutgers University in 1997; the LSA Institute at the University of Illinois in 1999; Deutschen Gesellschaft fr Sprachwissenschaft; the Linguistics Association of Great Britain in 2000; Stanford University in 2000; University of Potsdam in 2001; the LSA Annual Meeting in 2001; and SUNY Stonybrook in 2002.