The current status of the syntactic cycle
With the advent of the Minimalist Program, syntactic theory has undergone several fundamental changes that radically affect the former analyses of the cycle, including the empirical content of the principle. In this section, I review how the relevant changes affect the earlier analyses.
Consider the derivation of (1a) under minimalist assumptions. Since there are no phrase structure rules or X-bar schemas, the phrase structure of the example is built up from the interaction of the concatenation operation (Merge) with the lexicon (perhaps via a numeration). Given that the relevant portion of the derivation concerns only the movement of the wh-phrases, I will ignore the movement and feature checking of the subjects and objects. One stage of (5a)’s derivation will be (2).
(1)
(2)
The Q-feature is strong in English because it induces overt wh-movement. Therefore, under current analyses, the feature must be checked as soon as it is introduced into the derivation; otherwise, the derivation will cancel. More precisely, if the Q-feature is a feature of a complementizer C, then the derivation cancels unless it is checked within CP, a projection of C.4 Therefore, one of the wh-phrases in (2) must raise to check the Q feature to prevent the derivation from canceling. Since this movement takes place prior to Spell-Out, the whole wh-phrase must move to [Spec, CP], giving (3).
(3)
Notice that the movement of what does not produce an empty category. Rather, the trace of the moved phrase is just a copy of the phrase. The derivation continues to build on (3) by adding lexical material until it reaches the external complementizer.
(4)
At this point, in order to get the derivation to violate the SCC, the Q-feature of what, assuming it is + Interpretable and therefore cannot delete, should be available to raise to check the Q-feature in the matrix C. That this is ever possible seems unlikely. Generally, when a wh-phrase checks a strong Q-feature, it cannot move on. Thus, (5) is never possible even when the derivation involves movement of what through the sentential complement CP.
(5)
However, even if it were possible to raise what to the matrix [Spec, CP], the copy of what in the complement CP would block the movement of where (via substitution) to the internal CP. Furthermore, Last Resort would also block the movement of where since the Q-feature in the complement CP has already been checked and therefore there would be no motivation for moving where at all. This also holds for the possible adjunction of where to the complement CP.
There is of course another derivation of (5a), which violates the SCC but would not be blocked by Last Resort. Consider the derivation that has reached the stage given in (6).
(6)
Assume for the moment that the unchecked strong Q-feature in the complement CP has not caused the derivation to cancel, as it would under the analysis in Chomsky 1995. The derivation that violates the SCC involves movement of what directly to the matrix CP, followed by the movement of where to the complement CP. In this derivation, each movement checks a strong feature and therefore satisfies Last Resort. Whether the interclausal movement to the matrix CP can occur before the intraclausal movement to the complement CP depends on how the movement operation is construed. If the operation is Move, then the closest landing site is the complement CP and therefore the interclausal movement violates the Minimal Link Condition (MLC; see Chomsky 1995, 296, (82)). But if the operation is Attract, then the interclausal movement should be allowed because there is no closer wh-phrase to attract. The intraclausal movement is allowed under either interpretation of the movement operation. Now if we reverse the order of the two movements, the results are different. If where moves to the complement CP first, then the movement of what to the matrix CP will violate the MLC even on the interpretation of the movement operation as Attract since where in the complement CP is closer to the attractor Q-feature in the matrix CP than what in complement object position.
We have discovered here that under current analytical assumptions concerning the construction of phrase structure and the interpretation of movement as a feature-checking operation, the empirical motivation for the SCC as an independent principle is limited to a very particular interpretation of the theory. Thus, there is a derivation of (1a) that violates the SCC but would be tolerated by the current theory only if the movement operation is construed as Attract and not Move and, crucially, an unchecked strong feature does not cancel a derivation. Otherwise, the empirical effects of the SCC appear to be handled by an independently motivated principle (Last Resort) plus the definition of feature strength and therefore provide no motivation for postulating the SCC as an axiom of the theory.
Before we consider cyclicity in the literature on minimalism, it is worth asking whether the original motivation for the cycle reappears, given that generalized transformations are again responsible for generating generalized P-markers. The answer appears to be negative because the procedure for constructing phrase structure does not allow for embedding operations of the sort that existed in the earlier theory Embedding is not brought about by a substitution operation that replaces a dummy node ∆ with a sentential construct. Rather, the sentential complement becomes embedded in a VP by concatenating with a verb that constitutes the head of the phrase created. Therefore, there is no way that the matrix domain is constructed separately from the complement domain so that the matrix domain could undergo some transformational operation prior to the embedding of the complement domain. Therefore, the general procedure for constructing phrase structure itself precludes the possibility that the original formulation of the cycle was meant to prohibit.
Chomsky (1993) proposes a constraint on substitution operations that gives in effect a version of the strict cycle. Recall that the 1993 proposal includes two substitution operations: a binary operation GT (generalized transformation), which maps two P markers onto a single generalized P-marker, and a singulary operation Move α, which maps one P-marker onto another. Both operations work in the same fashion. Each operation targets a P-marker K, adds Ø (“a designated empty position”), and substitutes α for Ø. For Move α, α is a P-marker contained within the target K. The new version of the strict cycle requires that “substitution operations always extend their target” (p. 23).
(7)
Chomsky offers two empirical arguments for this condition. First, without (6) “we would lose the effects of those cases of the ECP [Empty Category Principle] that fall under Relativized Minimality” (p. 23). Chomsky cites examples relating to superraising, the Head Movement Constraint, and wh-islands. Consider the superraising example cited in (8) (= Chomsky’s (19a)).
(8)
Chomsky is assuming here that (8) could be a stage in a legitimate derivation, whereas under the analysis in Chomsky 1995 this derivation would presumably cancel because the strong D-feature of the complement I of seems has not been checked within its maximal projection (see below for further discussion). The derivation that violates the Extension Condition involves the insertion of it into the embedded IP after John has moved to the matrix [Spec, IP], extending the targeted embedded projection of I but not the whole P marker, yielding (9).
(9)
The strength of the argument for a cyclic principle based on Relativized Minimality violations like superraising depends on three points: (a) that such constructions could actually be generated in this way, (b) that this derivation is not prohibited by any other principle or analysis that is more general, and (c) that there are no other possible derivations of these constructions that are not ruled out by the Extension Condition. In the latter case, we might wonder whether some generalization is being missed.
Regarding (a), it is not obvious that a [Spec, IP] position in the complement of seems can be created after that IP has been embedded in the matrix VP by concatenating with the matrix verb. For one thing, such an operation raises nontrivial questions about derived constituent structure. Nonetheless, there is a derivation of the appropriate sort that will not run into such problems: namely, moving John successive-cyclically first to the [Spec, IP] of the complement of seems and then to the matrix [Spec, IP], then substituting it for the intermediate trace of John. Under the copy theory of movement operations, we might be inclined to disallow the substitution of it on the grounds that it is distinct from John and therefore the operation would violate the nondistinctness conditions on substitutions (cf. Chomsky 1965). The effect of the nondistinctness condition is to prohibit the deletion of information from a derivation. In this way, it is really just another way to express the recoverability condition on deletions. Suppose we recognize this redundancy between the two conditions and eliminate nondistinctness in favor of recoverability, which has a broader coverage. Now, does the deletion of a trace violate recoverability? Presumably, it does not in the case of PF since trace deletion there is general requirement (see Nunes 1995). In the case of LF, the trace of the foot of a chain may be necessary for interpretation, but apparently the intermediate traces of the chain in question are not and thus presumably could be deleted without violating recoverability. If so, then substitution of it for John should be allowed.
This suggests that the copy theory of movement does not block trace erasure via substitution. The analysis carries over to all the examples considered in Freidin 1978— for instance, (5)–(6). Interestingly, both (1a) and the superraising case (9) contain chains that violate Subjacency. However, if there is no level of S-Structure, as assumed in current work, then Subjacency interpreted as a condition on representations would have to apply at LF. Nonetheless, we probably do not want to rely on such an analysis because the notion of bounding category is essentially unmotivated under minimalist assumptions and therefore suspect.
Fortunately there is another, more fundamental way to eliminate these derivations without recourse to suspect notions like bounding category or appeals to a cyclic principle. Let us suppose that there are no substitution operations in UG. That is, we assume that the elementary operations of UG are limited to concatenation (adjunction) and deletion. In earlier theories, the substitution operation was required for lexical insertion and therefore it was reasonable to assume that it was generally available for movement operations as well. Now that lexical items are inserted into a derivation by the concatenation operation, there is no reason to assume the existence of movement by substitution.
If this analysis is on the right track, then it is at least questionable whether there is a countercyclic derivation of the superraising construction (14). Furthermore, the proposed analysis automatically addresses point (b) above—namely whether the countercyclic derivation is prohibited by some other principle or more general analysis. Putting aside the fact that the derivation violates Subjacency construed as a condition on chain representations at LF, it is prohibited by the elimination of substitution as an elementary transformational operation of UG. The only remaining question is whether concatenation can apply to a subphrase of a P-marker.
Let us turn now to point (c), concerning other possible derivations of these constructions that do not violate the Extension Condition. Given the separation of feature-moving operations and category-moving operations, it may be possible to derive the superraising violations without violating either the Extension Condition or the strong feature analysis. Consider, for example, the following derivation. First, the D-feature of John in (13) raises to check the strong D-feature of is. Next, it is merged with the resulting P-marker, checking the agreement and Case features of is and the Case of it. Notice that the insertion of it here does not violate the Extension Condition, nor does it involve leaving an unchecked strong feature, which would automatically cancel the derivation. At this point, the D-feature of John can raise to the matrix I, checking another strong D-feature. Suppose then that the nominal phrase John merges with the matrix projection of I, thereby generating the violation of superraising that we are attempting to avoid by invoking cyclicity in some form, but without violating cyclicity.
This derivation raises several questions. One crucial question concerns whether or not feature movement involves only the movement of a single feature or, as assumed in Chomsky 1995c,6 the set of formal features of a lexical item—FF(LI). If the former, then the cyclic merger of it could be involved in checking the agreement and Case features of is, as proposed. If the latter and if the pied-piped FFs enter into a checking relation with I, then the Case feature of it would not be checked, causing the derivation to crash at both LF and PF. Under this analysis, the superraising construction violates Full Interpretation, not some formulation of the SCC. Suppose, however, that even under the interpretation that Move F automatically involves the pied-piping of FF(LI), it is possible that feature checking of Case and agreement features involves the specifier-head relation. This ought to be possible because that is exactly how all the features of I will be checked in a simple expletive construction (e.g., (10)).
(10)
That is, it will merge with is likely that John will be here on time, creating a specifier head relation with is and checking not only its—Interpretable Case and agreement features, but also its strong D-feature. So even on the pied-piping analysis of feature movement, it may not be necessary for the “free rider” features, those that have been carried along, to enter into feature-checking relations when the feature that carries them does.
Another question that arises concerns the movement of John rather than it to the matrix [Spec, IP]. If it were a simple question of movement of categories, then movement of the nominal John over the nominal it would violate the MLC. However, under the Move F analysis, the situation is not so straightforward. Given that minimally only a feature (along with other associated FF(LI)) will move to check the strong D-feature of the matrix I, it is possible that the moved D-feature of John will move again, rather than the features of it in the complement [Spec, IP]. Notice that the two positions are equidistant, because they are both in the checking domain of I (as discussed above), and therefore cannot be distinguished so that the MLC will prefer one movement option over the other.
The final step in the derivation of (9) involves the movement of the nominal category John to the matrix [Spec, IP]. If category movement is motivated solely for the purpose of convergence at PF but not for feature checking, then John moves directly from the IP complement of certain, presumably attracted by its FFs that merged with seem. This movement violates the MLC only if it can be attracted to the matrix [Spec, IP] instead of John. But if all the FFs of it have been checked in the complement of seems, then there is no reason to move it at all. If this is correct, then the derivation of the superraising case converges.
The convergent derivation of the superraising construction involves at least one highly suspect property: namely, that the features of a finite verb (or I) can be checked by the features of different nominal expressions. This could be avoided if feature checking is obligatory. Thus, the Case feature of John will be checked by is and therefore could not be raised to the matrix clause to check the Case feature of seems. This would follow from the economy condition on fewest steps. Checking the D-feature of I by raising the FFs of a nominal expression and then checking the Case and by merger of an expletive counts as two steps whereas checking all these features using only the raising operation counts as one step. This interpretation would ensure that the expletive winds up in the matrix [Spec, IP] since this is the only position in which its Case feature will be checked.
Under this analysis, (9) could result only if the expletive checks all the relevant features of is. Then the movement of John from the complement of certain directly to the matrix clause would be blocked by the MLC if Move F applies to the D-feature or the of the nominal. However, if Move F applies to the Case feature, then the MLC would not block the movement because the Case feature of it, having been checked, is no longer accessible. Thus, (9) could still converge, even if the two-step feature-checking option discussed above is blocked. Alternatively, the MLC could be strengthened so that the FFs that get carried along and enter into a checking relation are also considered in the determination of minimal links. Then, even when the Case feature chain does not violate the MLC, the D- and chains will.
If the superraising construction (9) does actually converge, then we might expect that some principle of economy will prefer the derivation of (11) to that of (9).
(11)
Thus, at the point in the derivation shown in (17) where there is a choice between moving the features of John and merging it, Move F is preferred to Merge.
(12)
If this generalizes to all derivations, then singulary transformations will be preferred to generalized transformations. Presumably, the computational system for human language (CHL) tries to make maximal use of what is already in a P-marker before incorporating additional material (but cf. Chomsky 1995c, 347).
Like the superraising case we have considered in such detail, the Head Movement Constraint (HMC) case is assumed to involve the insertion of a lexical item within a P marker. An HMC violation can be generated from (13) if fix is moved to C and then can is inserted in VP or IP.
(13)
The Extension Condition blocks the insertion of the modal. Such potential derivations are dubious on quite general grounds. C normally attracts only the finite form of the verb. If so, then there will be no reason for a nonfinite verb form to move to C and hence no way to generate HMC violations. If this line of reasoning is viable, then perhaps the entire range of HMC problems is just the result of an overly general and imprecise analysis.
Unlike the superraising and HMC cases, the wh-island case does not involve an instance of lexical insertion that is supposed to violate the Extension Condition. One derivation from (14) violates the Extension Condition straightforwardly.
(14)
If Move α targets the matrix C′ creating Ø external to C′ and substituting how for Ø then (7) prevents Move α from targeting the complement C′ and moving what to the complement [Spec, CP]. This derivation is also ruled out because of the unchecked strong Q-feature in the complement CP at the point that how is moved to matrix CP. So far, there is no strong empirical argument for the Extension Condition based on these examples.
There is of course a derivation of the wh-island violation (15) from (14) that violates neither the Extension Condition nor the strong feature cancellation analysis: namely, what moves to the complement [Spec, CP] and then how moves to the matrix [Spec, CP].
(15)
The second movement constitutes a clear violation of the MLC if the Q-feature of what in the [Spec, CP] of the complement could be attracted to check the Q-feature of the matrix C.
Given these analyses, no Relativized Minimality violation provides strong evidence for a cyclic principle.
Another empirical argument for the Extension Condition that does not involve Relativized Minimality violations concerns raising to complement position. Chomsky (1993) notes that given the Extension Condition, a structure like [X′ X YP] cannot be mapped onto [X′ X YP ZP] where ZP is raised from YP or inserted by GT. The strength of this argument depends on two closely linked assumptions: (a) that this consequence is unique to the Extension Condition and (b) that substitution operations will construct these structures that the Extension Condition is needed to prohibit. The second assumption, on which the first relies, seems far from obvious. It presupposes a substitution operation so unstructured that it can insert a designated empty position virtually anywhere in a P marker. Furthermore, the operation that could map [X′ X YP] onto [X′ X YP ZP] would be both structure-building and structure-destroying in a way that substitution operations are not supposed to be. So it seems doubtful that this kind of mapping would arise naturally.
Chomsky (1995c, 328) cites another case related to overt cyclicity involving the interaction of NP- and wh-movements (also A- and Ā-movements).
(16)
The wh-movement violates the Condition on Extraction Domain (GED), but can be derived without violating that condition if wh-movement to [Spec, CP] precedes NP movement into [Spec, IP] countercyclically. Chomsky states that the countercyclic derivation is prohibited by the current analysis of feature strength. The strong feature of I unchecked by NP-movement will cause the derivation to cancel before wh-movement can apply. He also suggests two other ways to exclude the countercyclic derivation of (16). But before we examine those, it is worth taking a closer look at the strong feature analysis.
For (16), too, the separation of category movement from feature movement has the potential for undermining the strong feature cancellation analysis. Thus, it might be possible for the strong D-feature of I to be checked via feature movement and still move who from the VP position. This would entail that it is possible to create the [Spec, IP] position after the phrasal projection of I has been embedded in CP as the complement of C. The overt movement of the NP is not forced by the strong feature cancellation analysis unless we assume that the strong feature of I is checked as a result of that movement, contrary to the Move F analysis. Therefore, the countercyclic derivation of (16) appears to argue for retaining the Extension Condition.
The problem with the Extension Condition is that it should prohibit feature movement generally since such operations (including overt head movement) do not extend the targeted P-marker as required.
Chomsky (1995c) suggests two other ways of preventing the countercyclic derivation of (16). One involves some notion of economy. Although the NP-movement in both derivations is the same, the wh-movement in the countercyclic derivation is longer and therefore might be excluded by comparing length of steps in derivations. As Chomsky notes, this is not appealing because it requires “a ‘global’ notion of economy of the sort we have sought to avoid” (p. 328). This would require that both derivations converge even though the construction is clearly deviant; therefore, the admissible (by hypothesis) cyclic derivation would have to be blocked in some other way. Another alternative Chomsky considers is the provision that “α can be attracted to K only if it contains no trace” (p. 365, (200)), which he proposes as a strengthening of the proviso that only the head of a chain may be attracted. This rules out the NP-movement in the countercyclic derivation rather than the wh-movement. Under this analysis, we lose the generalization expressed by the CED that extractions out of subject phrases in finite clauses are generally prohibited whether the subject phrase is created by movement or merger.
So far, we have been discussing how to eliminate the countercyclic derivation of (16). However, since (16) is deviant, all derivations must be eliminated—either the derivation is prohibited or the representations derived violate some output condition at PF or LF. Although Chomsky mentions the CED in reference to the cyclic derivation, we cannot appeal specifically to this condition because its formulation involves the notion “government,” which has been excluded from discussion within the Minimalist Program on the grounds that it is illegitimate. Therefore, the cyclic derivation of (16) must be excluded in some other way.
To this end, let us take a closer look at the cyclic derivation under the copy theory of movement, which involves the three steps illustrated in (17).
(17)
In (17a,b), α is a trace of its copy β. However, in (17c), β now contains a trace of who whereas α does not. Suppose that this renders α distinct from β so that α cannot function as the trace of β. The derivation would crash at LF because the nominal a picture of could not be assigned a θ-role. In other words, (16) constitutes a θ-Criterion violation— specifically, Functional Relatedness (Freidin 1978), now subsumed under FL Note that if the cyclic derivation crashes, then we cannot resort to economy to rule out the countercyclic derivation.
Returning now to the countercyclic derivation of (16), we have a number of options for excluding it. Under the copy theory of traces, the relevant steps of the derivation proceed as in (18).
(18)
In (18b,c), who in α is the trace of who in the matrix [Spec, CP]. In (18c), who in β, which is presumably a copy of a, is also a trace of who in [Spec, CP]. By construction, β binds α and who in [Spec, CP] binds who in α. What is not clear is the relation between who in [Spec, CP] and who in β. If we suppose that who in [Spec, CP] binds who in β, then we might object to the chain (who, whoβ, whoα) on the grounds that the link between whoβ and whoα is ill formed because whoβ does not bind whoα. Alternatively, if we impose the derivational definition of c-command (Epstein 1991), then who in [Spec, CP] will not c-command β, hence who in β. Since by construction who in β is a trace, β contains an unbound trace in violation of the Proper Binding Condition.
Under the copy theory of movement, there may be yet another explanation for the deviance of (16) regarding trace deletion. Consider the standard assumption that nontrivial chains must reduce at PF to trivial chains; that is, all traces must be erased because nontrivial chains are illegitimate objects at PF, violating FL The general case involves a chain whose linear order reproduces the linear order of the copies in the P marker and in which each adjacent pair of copies is in an asymmetric c-command relation such that the head of the chain asymmetrically c-commands the trace adjacent to it, that trace asymmetrically c-commands the trace adjacent to it, and so on to the foot of the chain. Such chains generally reduce to trivial chains at PF.
Before we discuss how this situation does not obtain in the derivations (17) and (18), we need to consider whether a copied element is necessarily always a trace of its copy. In this regard, consider (19).
(19)
Presumably, there is only one chain to consider, the chain involving the NP books about language and its trace in [Spec, IP]. That is, there is no chain between books in the matrix IP and the copy in the complement IP, and similarly for about and language. From this perspective, what is odd about a derivation like (17) is that a copied element who is not a trace when the phrase containing it is moved, but then becomes a trace when it alone is moved later. One way to deal with this is, as suggested above, to treat the phrase containing the nontrace copy as distinct from the phrase containing the trace copy Presumably this will block the required θ-role assignment. Furthermore, α and β would not form a chain and hence there would be no reason to delete α at PF.
Alternatively, we might resolve the issue as follows. Suppose that once a head-trace relation is established between two copies in a derivation, all copies must be considered as part of the chain. Thus, in both (17) and (18), the chain (20) is what must undergo chain reduction via trace deletion.
(20)
However, (20) does not conform to the standard case because whoβ does not asymmetrically c-command whoα (cf. Takahashi 1994). The counter-cyclic derivation adds another failure of asymmetric c-command between the first two elements of the chain on the derivational c-command analysis. The first is of course sufficient to block trace deletion in both derivations. As a result, both derivations crash at PF because they both contain illegitimate nontrivial chains, in violation of FI. If this is a viable analysis, then it should be preferable to one that requires an additional stipulation, like the Extension Condition, to impose cyclic derivations.
So far, we have yet to identify a strong empirical argument for the strict cycle based on the Extension Condition. The empirical motivation for the Extension Condition based on Relativized Minimality violations seems to dissolve when we probe the details of the analysis. Chomsky (1995c) reassesses these cases, noting that they involve two situations: (a) countercyclic operations and (b) skipping an already filled position. As we have seen with the superraising and wh-island cases, the MLC alone blocks countercyclic operations. Chomsky claims that the countercyclic insertion of heads does not arise (viz., HMC violations) because heads are inserted only by pure merger, which satisfies the Extension Condition. The other situation, involving skipped positions, is generally blocked by the MLC. This leads Chomsky to conclude that there may be no need to impose the Extension Condition on overt category movement.
