Superiority, Subjacency, and Economy
In his introduction to volume four of Noam Chomsky: Critical Assessments, Carlos Otero writes:
In true scientific inquiry (in fact, in any kind of rational inquiry including the natural sciences), productive people try to identify and come to understand major factors and see what can be explained in terms of them. They anticipate that there will always be a periphery of unexplained phenomena (a range of nuances and minor effects that require auxiliary assumptions) which should be very sharply separated. (1994:17–18)
Generative grammar provides an ideal case study, as Otero has documented. This inquiry has from the outset been a search for guiding principles, the dominant structures, and the major consequences.
In the history of generative grammar over the past four decades, among the major factors investigated have been rule systems and systems of grammatical principles and related parameters. One extremely successful line of research has been the elimination of language-particular and construction-specific transformations in favor of three optimally general and presumably universal transformational rules: Substitute α for β, Adjoin α to β, and Delete α (the former two referred to under the designation “Move α” or all three under the designation “Affect α”). This was brought about by the gradual development (1970 to 1979) of a set of general grammatical principles involving phrase structure, bounding, binding, Case, government, and predicate/argument structure, many of which function as conditions on representations, rather than conditions on the application of rules (i.e., on derivations), raising a question about the role of derivations in grammar. If most grammatical principles function as conditions on representations, then perhaps all conditions on derivations could be replaced by conditions on representations so that derivations would become essentially epiphenomenal. One result in this direction was the demonstration in Freidin (1978) that the empirical effects of the Strict Cycle Condition of Chomsky (1973) could be derived from independently motivated conditions on representations. Such derivations highlighted the interconnection of the proposed conditions, beginning with Chomsky’s demonstration of how many of Ross’s island constraints (Ross 1967/19864)—and in particular the Complex NP Constraint—followed from the Subjacency Condition (Chomsky 1973, 1997b—henceforth Subjacency). This led to the useful research strategy of trying to derive the empirical effects of one condition from another (or others) that appear to be more basic or have a broader empirical range. The guiding principle in this research is to eliminate redundancy (i.e., overlapping conditions) under the assumption that the language faculty is nonredundant. When a construction is prohibited by more than one condition, it is assumed that this construction is more strongly deviant than those that are prohibited by only one of the conditions. This is an empirical hypothesis, though not as yet particularly well motivated.
The Superiority Condition (Chomsky 1973) provides an interesting case study of the connections between grammatical principles and the effort to subsume certain ones under others. The original formulation of the Superiority Condition (henceforth Superiority) in Chomsky (1973) applies exclusively to leftward movement:
A category Z is superior to Y when Z asymmetrically c-commands Y. The standard examples involve wh-movement in a sentence containing multiple wh-phrases where one phrase is superior to the other, as in (2a–b):
Given a rule of wh-movement that can move either the subject who or the object what into [Spec,CP], Superiority prohibits the movement of what because it is asymmetrically c-commanded by who. Under this formulation Superiority is a condition on the application of wh-movement and is thus a condition on derivations.
One widely discussed attempt to derive the empirical effects of Superiority involves the Empty Category Principle (ECP) of Chomsky (1981), under the crucial assumption that the ECP applies to LF representations. Under this proposal (2a) and (2b) would have the LF representations along the lines of (3a) and (3b) respectively:
Note that this analysis assumes that unmoved wh-phrases in (2) are simply adjoined at LF to the wh-phrase in [Spec,CP]. It follows from the adjunction operation that the head of the phrase to which a wh-phrase is adjoined remains the head of the entire adjunction structure. Thus, in (3a) who fails to antecedent govern its trace because it does not c command it—in contrast to (3b) where NP1 c-commands and hence antecedent governs the trace in subject position. In both (3a–b) the trace in object position satisfies the ECP because it is lexically governed by the verb. On this analysis Superiority a condition of derivations, is reduced to the ECP, a condition on LF representations. This analysis crucially depends on a disjunctive formulation of the ECP where either lexical head government or antecedent government suffices (as in Lasnik and Saito 1984, 1992).
While this reduction may work for Superiority violations involving a superior subject in a finite clause, it fails for all other superior positions. Consider (4a–b) with the representations (5a–b):
In (4a) the complement subject will undergo exceptional Case marking by the matrix verb. Under the standard GB analysis the infinitival subject position is lexically governed by the matrix verb. Thus antecedent government of the trace in complement subject position by the wh-phrase in [Spec, CP] is not required. Under the Spec/head agreement analysis of Case assignment (Chomsky and Lasnik 1993, Chomsky 1993), who(m) would move to [Spec, AgrO-P] in the matrix clause where it would antecedent govern its trace in complement subject position. (5a) therefore involves a Superiority, not an ECP, violation. In (5b) the superior wh-phrase is in object position where it is lexically governed. Thus (5b) is also a Superiority violation, but not an ECP violation.
Another configurational analysis of superiority effects is given in Pesetsky (1987) involving the relation between paths of moved wh-phrases (or in more contemporary terms, between wh-chains). Specifically, Pesetsky proposes a Nested Dependency Condition as formulated in (6):
In (7a) the chain (who1, t1) is nested inside the chain (what book2 t2), whereas in (7b), the chains intersect because what book lies inside the chain created by the movement of who while its trace is outside that chain.
To handle Superiority violations within a single clause (e.g., (2a)), Pesetsky adopts the analysis of Jaeggli (1981), where the wh-phrase that remains in-situ at S-structure is adjoined to the clause that dominates the moved wh-phrase—that is, adjunction to CP in our analysis or to S′ in Jaeggli’s. This analysis corresponds to the standard raising analysis of quantifiers (May 1977) in which scope ambiguities are resolved in LF representations—specifically, the quantifier with wider scope will c-command the quantifier with narrower scope at LF. This seems inappropriate since multiple wh-phrase constructions do not manifest such scopal ambiguities. The answer to a question like (2b) is simply a paired list of people and objects. It involves a mapping from readers to things read. In contrast the LF representation allowed by the NDC (i.e., where what takes wide scope over who) corresponds to an interpretation where the mapping is from things read to readers. Yet a list of what was read by whom (e.g., “Ulysses was read by Bill and Mary, and Middlemarch was read by Jane, Fred, and Sally,…”) is a thoroughly unnatural response to a question like (2b).
The NDC generalizes to certain Subjacency violations. Thus (8) violates both the NDC and Subjacency:
The chain (who2, t2) intersects the chain (what1, t1) and the link of the latter chain crosses two bounding categories (2 IPs). Nonetheless, (8) cannot be attributed to a crossed binding constraint like the NDC because of examples like (9), which are equally unacceptable but where there is no occurrence of crossed binding, as discussed in Freidin and Lasnik (1981).
Presumably (9) is only a Subjacency violation. It is not a Superiority violation because at the point in the derivation where what moves to the internal [Spec, CP], it asymmetrically c-commands who (on virtually every analysis of double object constructions that has been proposed in the literature). In contrast the movement of who to the internal [Spec, CP] in (8) violates Superiority because the wh phrase what is superior. Such paired examples suggest that a geometric approach to movement constraints like the NDC (or the Path Containment Condition of Pesetsky 1982b) is not viable.
At this point we might raise the question of whether the movement in (9) of who to the external [Spec, CP] over what in the internal [Spec, CP] constitutes a Superiority violation. Undoubtedly, the internal [Spec, CP] c-commands the object of the PP in the embedded IP. Nonetheless, the wh-phrase in [Spec, CP] could not move to the external [Spec, CP] because then the selectional requirements of wonder would not be satisfied. If we change the matrix verb to remember, which does not require a [+WH] CP complement, then it is possible to derive (10):
The question is whether the movement of what to the matrix [Spec, CP] is driven by Superiority as well as Subjacency which is violated in (9) by the movement of who to the external [Spec, CP]. In other words, is Superiority limited to wh-phrases in grammatical function positions? This seems like a reasonable way to sharpen the formulation of Superiority16 Note that those Romance languages in which a relative pronoun can be extracted out of a wh-island provide some evidence that Superiority does not apply between a superior wh-phrase in [Spec, CP] and a wh-phrase in a grammatical function position. If it did, then those constructions should be blocked by Superiority.
Another connection between Superiority and Subjacency arises in constructions such as (11), which can be associated with two distinct S-structures (12a–b):
The chain (what1, t1) in (12a) violates Subjacency and Superiority (since who, which asymmetrically c-commands the complement object from the internal [Spec, CP], could also move to the external [Spec, CP]). In contrast, derivation of the chain link between the two traces in (12b) violates only Superiority. Suppose that the language faculty assigns examples like (11) a unique mental representation at each level of representation. Within the recent “minimalist” framework (Chomsky 1993, 1995c) in which the levels of representation are restricted to the interface levels PF and LF, the problem we have identified cannot be discussed in terms of assigning two distinct S-structure representations. Rather, the computation of (11) “crashes”—that is, it fails to converge (meaning “yield a legitimate structural description”) at one or both interface levels. Under a standard GB analysis (11) is computationally ambiguous. Computational ambiguity of this sort seems rather inefficient for language design, so perhaps it can be excluded. Let us conjecture that the only kind of computational ambiguity we will find in languages is that in which the two computations are convergent, as in the standard case of structural ambiguity (e.g., a review of a book by two linguists). Given this conjecture, the language faculty assigns a unique computation to (11) and therefore (11) crashes for a particular reason. If this is correct, then certain aspects of derivations are not epiphenomenal, as some earlier research was interpreted as suggesting.
From this perspective the problem with (11) is to determine why the computation crashes. Does the failure to converge involve PF or LF (or both)? The answer will depend on how and where Superiority and Subjacency apply to derivations or the representations they create, which in turn depends on other factors concerning the derivation of (11). Putting aside the issue of at what level or levels nonconvergence occurs, we have two distinct possibilities: the computation of (11) crashes because of Superiority or Subjacency. The choice between these two possibilities can be resolved if it follows from the theory of grammar that either the wh-phrase in subject position must move into [Spec, CP] prior to the movement of the wh-phrase in object position (in which case we get a Subjacency violation) or the wh-phrase in subject position does not move to a [Spec, CP] until after the wh-phase in object position moves to the external [Spec, CP]. Under the latter scenario Subjacency is not violated, but Superiority is. Of course, in the absence of a theory that chooses one analysis over the other, we might try to identify some data whose analysis requires one option and excludes the other.
The standard versions of GB do not distinguish the two computations of (11) given in (12). Thus, (11) is excluded either because it violates two conditions or only one. Under the assumption that a construction that violates two general principles will be perceived as more deviant than one that violates only one of the two, this analysis of (11) cannot stand. One way of computing (11) must be wrong.
Fortunately, there is a way to tease apart the Superiority versus Subjacency analyses empirically. Consider the well-known fact that superiority effects disappear when bare interrogative pronouns are replaced by which-phrases, as illustrated in the contrast between (13a) and (13b):
This fact gives us a tool for testing the sort of violation that occurs in (11). If the corresponding construction with which-phrases instead of bare interrogative pronouns is less deviant, then we know that (11) involves a Superiority violation.
Given that (14) is significantly better than (11)—some speakers (myself included) find it essentially normal, (11) appears to be a Superiority violation, not a Subjacency violation.
Having identified some empirical evidence for a unique analysis of (11), we now need to show how the analysis follows from the theory of grammar. I am assuming that examples like (13b) and (14) are not necessarily part of the primary language data of every child who acquires a grammar of English. Even if they were, the analysis of constructions like (11) does not follow automatically.
Within the current minimalist framework as sketched in Chomsky (1993) there are two economy of derivation principles that could be used to predict that a wh-phrase in subject position that is not moved out of its clause remains in-situ in the derivation to PF19 (henceforth “overt syntax”): Procrastinate and Greed. Procrastinate is based on the notion that LF movement is somehow preferable to overt movement (see Chomsky 1993:30). Thus, if nothing requires the movement of a subject wh-phrase in overt syntax, it will not move in the overt syntax. The derivation to LF does require the covert movement of a wh-phrase subject to [Spec, CP] in order to create a proper quantifier/variable structure. Otherwise, the derivation crashes at LF because the result violates Full Interpretation (FI)—a quantifier that does not bind a variable cannot be interpreted. Greed, a somewhat related principle, is characterized as selfserving Last Resort, where the Last Resort principle licenses a step in a derivation “only if it is necessary for convergence” (Chomsky 1993:32). “Derivations are driven by the narrow mechanical requirement of feature checking only, not by a ‘search for intelligibility’ or the like” (Chomsky 1993:33). Under Chomsky’s version of Greed, an element moves only to check its own features. As will be discussed below, because this requirement appears to be too strong for wh-movement, a modified version of Greed will be proposed.
Given that quantifier/variable structures need not occur in overt syntax (as in the case of non-wh-quantifiers), it is reasonable to assume that the construction of these structures can be postponed to the derivation to LF. The application of Procrastinate to (11) therefore suggests that (12b) should be a viable PF representation. However, Greed pulls harder in the other direction.
Following the proposal for feature checking in Chomsky (1993), let us assume that languages have strong or weak features; that strong features are visible at PF whereas weak features are invisible; that features that are checked “disappear”; and that strong features that remain unchecked at PF cause a derivation to crash. Let us further assume that [+WH] is a strong feature in English. Therefore, it must be checked in overt syntax, or the derivation will crash at PF. For reasons that will become clear below, I am going to assume that this wh-feature is the one attached to C and that it is checked via Spec/head agreement when a wh-phrase moves into the relevant [Spec, CP]. Thus, if C contains [+WH], this feature must be checked in the overt syntax, or the derivation will crash at PF.
Consider a verb such as expect that does not allow indirect question complements. The sentence (15a) could be represented as either (15b) or (15c):
In (15b) the wh-phrase has failed to move to the external [Spec, CP], hence the strong wh-feature in the external C is not checked, and the derivation crashes at PF because it is visible. At LF the wh-phrase can still move to the external [Spec, CP] to check the wh feature (too late in this instance); however, this will create the requisite quantifier variable structure needed to interpret the wh-phrase at LF. So the derivation converges at LF. Greed applies both to LF, where it governs the movement of an object or infinitival complement subject NP to [Spec, AgrO-P], and to PF, where it determines the movement of the subject of a finite clause from [Spec, VP] to [Spec, AgrS-P] (= [Spec, IP]). In (15c) there is no strong feature to be checked, thus no reason for the wh-phrase to move at all in the overt syntax or the covert syntax, given Greed. Thus, the derivation of (15c) crashes at LF because the wh-phrase cannot be properly interpreted—a violation of FI.
If we substitute the verb wonder, which obligatorily selects a [+ WH] C, for the matrix verb in (15a), the derivation crashes at PF because this strong feature is not checked. Greed would allow the wh-phrase to move to the internal [Spec, CP] at LF, hence the derivation would converge at LF. If that movement did not occur, then the derivation crashes at LF too for the familiar reason. If the matrix verb in (15a) is changed to forget, which optionally selects a [+ WH] C, then either the matrix C or the complement C or both could contain [+WH]. If both do and there is only one wh-phrase, then one will not be checked, and the derivation crashes at PF. If there are two wh-phrases in the construction, then each wh-feature can be checked and the derivation converges at PF (as in (14)). If neither matrix nor complement C contains the wh-feature, then Greed causes the derivation to crash at LF as an FI violation.
Now consider the case in which there are multiple wh-phrases and only one [+ WH] C—e.g., who read what? If feature checking applied to a feature of the wh-phrase, then such questions should crash at PF, contrary to the fact. For this reason we have been assuming that a wh-phrase checks the wh-feature in C and not conversely. As for the LF representation of multiple wh-questions in which one wh-phrase remains in-situ, let us assume that the moved wh-phrase provides an adjunction site to which the wh-phrase in situ can move at LF to create the appropriate quantifier/variable structure.
This analysis requires that we modify the analysis of Greed for wh-movement. The wh-phrase is not moving into [Spec, CP] to check its own [+ WH] feature, but rather the one contained in C. Let us assume that the purpose of Greed is to maximize convergent derivations (i.e., profits). If a particular movement does not lead to a convergent derivation, then Greed blocks it.
Apparently the wh-feature in C is independent of whatever feature accounts for the formation of yes/no questions. Thus (16a) crashes either at PF because there is a [+ WH] feature in the external C that is not checked or at LF because the wh-phrase cannot move to create a quantifier/variable structure.
In contrast the wh-phrase in (16b) converges at PF because the [+WH] feature in the matrix [Spec, CP] is checked by the wh-phrase and at LF for the familiar reason. (16c) does not involve a [+WH] C, as indicated by the contrast between (16a) and (16c).
With this analysis in mind we can return to the computation of (11) as a violation of Superiority rather than Subjacency This concerns the way feature checking is carried out when a wh-phrase occurs in [Spec, IP]. To see what this involves, let us consider the verbs expect, forget, and wonder, which differ in their ability to take indirect question complements. As the paradigm in (17) illustrates, expect cannot take an indirect question complement, forget may (but does not have to), and wonder must.
(17a) crashes at LF for the same reasons that (15a) does, depending on which analysis of the external C we choose. In (17b) who winds up in the matrix [Spec, CP] to have its [+ WH] feature checked by [+ WH] C. The chain it forms involves the complement [Spec, CP]—in which case Subjacency is satisfied. (17c) and (17e) seem to be identical in the relevant respects. In both complements the wh-phrase subject raises to [Spec, CP] to check the [+ WH] feature. (17f) crashes at PF because the [+ WH] feature of the complement C has not been checked. In contrast the feature of the complement C in (17d) is [− WH] and therefore does not require overt material in [Spec, CP].
Notice that we have been assuming that the wh-phrase in [Spec, IP] raises to [Spec, CP] so that checking of the wh-feature can occur. There is another possibility—namely, that checking occurs without movement to [Spec, CP]. If this is feasible, then Greed will rule out the string-vacuous movement of the wh-phrase in the overt syntax. The movement of the wh-phrase from [Spec, IP] to [Spec, CP] would occur in the covert syntax, the derivation to LF, as predicted by Procrastinate.
Such derivations would require that the [+ WH] feature of complement C be checked by the wh-phrase in complement subject position. While this analysis might appear feasible for (17c) and (17e), it would create a problem for constructions such as (18) where the matrix verb selects a [− WH] C:
If the selectional feature on C for forget and wonder can be transferred to I, then presumably the same thing should happen with expect, thereby creating a Spec/head agreement violation. Since (18) is perfectly acceptable, its computation should not violate any constraint on movement or selectional restriction. In this way constructions such as (18) argue against the possibility that the [+ WH] feature on C can be checked by a wh phrase in [Spec, IP]. (As we will see below, an even stronger empirical argument can be given against this analysis.)
Given Greed, the computation of (18a) cannot involve successive cyclic Move α through the internal [Spec, CP] because movement of the wh-phrase to this position is not motivated by feature checking. Also, a two-step derivation would violate the principle of Least Effort (Chomsky 1991), provided there is a way of accounting for this long distance movement in a single step. Following Chomsky (1993), let us now assume that the basic movement operation is not Move a, but rather Form Chain. Form Chain constructs (18b) in one step by moving which books to the external [Spec, CP] and at the same time constructing the chain consisting of the trace in the internal [Spec, CP] and the trace in complement object position. The chain is well formed since its links are minimal (i.e., satisfy Subjacency etc.). Given this formulation, it follows that Subjacency, now a condition on chain links, must be construed as a condition on representations. In this way (18) provides empirical evidence for Form Chain under a theory that includes Greed and a principle of Least Effort.
In (18) the wh-phrase in the complement subject position could not be raised to the internal [Spec, CP] position, thereby creating a putative Subjacency violation. However, a corresponding example with forget (e.g., (14) above) is susceptible to this analysis. Assuming the Form Chain analysis, the derivation of (14) with respect to wh-movement can be computed as two steps in (19) or one step in (20):
In (20) the one-step consists of moving the complement object to the external [Spec, CP] and forming a chain, which satisfies Subjacency. The two-step derivation in (19) can be given in two ways, depending on the order of the operations. If the complement subject which students is moved first to the complement [Spec, CP], then the application of Form Chain to the complement object which books will create a chain that violates Subjacency. The alternative derivation is countercyclic: which books moves long distance to the matrix [Spec, CP], forming a chain that satisfies Subjacency, and then which students moves into the complement [Spec, CP]. A principle of Least Effort selects (20) over (19) ceteris paribus. Given the status of (14), it seems clear that the proper computation for constructions such as (14) is (20) and not (19). As noted earlier, Subjacency violations do not exhibit varying effects depending on whether a which-phrase or a bare interrogative is involved. Furthermore, can, like expect, select a [− WH] C as discussed above.
If the complement C in (20) does not contain [+ WH], Greed prevents the complement subject from moving to complement [Spec, CP] at LF to create a quantifier/variable structure. Instead, it must move to the external [Spec, CP], adjoining to the wh-phrase there to create a complex quantifier that is appropriate for the pair-list reading of such constructions. However, when both wh-phrases in a multiple wh-question have moved to a [Spec, CP] in the overt syntax, the pair-list reading is blocked, as illustrated in (21):
In (21b) the question asks for a pair-list answer, whereas in (21a) it does not—instead, it asks for the identity of student x such that the professor told x which book to read. The difference in interpretation is directly related to the difference in the syntactic structures of (21a) and (21b). In (21a) both wh-phrases have moved in the overt syntax to [Spec, CP] positions; therefore, at LF they form independent quantifier/variable structures. In contrast only one wh-phrase has moved to [Spec, CP] in (21b), the other remaining in situ. However, at LF, the in-situ wh-phrase (which student) must move to [Spec, CP] in order to form a quantifier/variable structure. The only c-commanding [Spec, CP] position is already filled with a wh-phrase, therefore the in-situ wh-phrase adjoins to the wh phrase in [Spec, CP], forming an absorption structure that gives the pair-list reading. Notice also that the nonpair-list reading of (14) may be computationally more complex than that of (21 a) because it involves two intersecting quantifier/variable chains, while those in (21 a) do not intersect.
Although this analysis accounts for expect and forget (e.g., (18) and (14)), the corresponding construction with wonder, which obligatorily selects [+ WH] C, unlike forget, raises some further problems:
The status of (22) is on a par with (14), though perhaps slightly degraded because of the computational complexity noted above. Significantly, (22) does not have a pair-list reading like (14), but only the one asking “for which books x did John wonder which students had borrowed x.” The fact that (18) can only have a pair-list reading and (22) cannot would follow if the complement subject of (22) moves to [Spec, CP] in the overt syntax. In this way (22) provides evidence against the Vacuous Movement Hypothesis (cf. Chomsky 1986b).
As with the corresponding example with forget, (22) has two potential analyses, given in (23):
It would seem that (23b), the blocked Superiority violation derivation, and not (23a), the Subjacency violation, provides the appropriate computation. And yet the selectional feature of wonder requires (23a) at PF. If (23a) must feed the PF computation of (22), the countercyclic derivation seems the only way to avoid a Subjacency violation.
We have been assuming that feature checking of wh-features applies at PF for good reason. Consider what results if we adopt the alternative assumption—namely, that wh features are checked at LF. (22) is no longer a problem because the wh-phrase in [Spec, IP] could raise to the complement [Spec, CP] at LF to check the wh-feature in C. However, this mechanism would also check the same feature in (24a), so that the explanation for the deviance of such constructions is lost.
If forget in (24b) selects [+ WH], checking features at LF will not account for such constructions—in contrast to feature checking at PF, where the wh-phrase remains in situ. Thus, checking of wh-features occurs at PF, not LF.
It is worth noting that this account of (22) provides evidence for the derivational (as opposed to a representational) approach to grammar:
Under a derivational approach, computation typically involves simple steps expressible in terms of natural relations and properties, with the context that makes them natural wiped out by later steps of the derivation and not visible in the representations to which it converges. (Chomsky 1995b)
The second step in the derivation of (23a) wipes out the context that allowed the first step.
Ironically, the proposed derivation for (22) violates the Strict Cycle Condition (SCC) as formulated in Chomsky (1973). This would be a serious problem if the SCC were an axiom of the theory, but not if the empirical effects of the SCC followed from other considerations (see Freidin 1978, Kitahara 1995). Given the Form Chain analysis of wh movement outlined above, it should be clear that Subjacency alone no longer accounts for any deviant wh-island construction. Consider the derivation of (25):
As in (22), the movement of each wh-phrase in (25) is motivated by Greed. If which books moves to the internal [Spec, CP] first, then the chain formed by a second movement of which students to the external [Spec, CP] will violate Subjacency. Presumably, this derivation converges at PF but crashes at LF. If, however, the two applications for Form Chain are reversed, then neither chain formed will violate Subjacency. The reason that (25) is deviant, but (22) is not, must follow from some other condition—in this case the EC P. The chain (which students1, t1) in (25) violates the ECP at LF because the trace is not properly governed. In contrast the trace in the LF chain (which books2, t2) in (22) must be properly governed. In the latter case it does not seem as if antecedent government will hold since both (22) and (25) appear to be relativized minimality violations of the same sort. This suggests that we may need to retain head government to explain the difference between (22) and (25), which involve another subject/object asymmetry.
Chomsky (1993) proposes that the notion “shortest link,” expressible in terms of the operation Form Chain, might be used to incorporate parts of Subjacency and the ECP under the intuitive formulation of a Shortest Movement Condition expressed in (26):
Chomsky suggests that the phenomena of Superiority and Relativized Minimality (including superraising, the Head Movement Constraint, and [Spec,CP] islands (including wh-islands)) should fall out from such economy considerations, though an explicit account is not provided. (See Kitahara 1993 for a more detailed analysis of Superiority effects.)
To see how the Shortest Movement Condition (henceforth Shortest Movement) subsumes the effects of Superiority, consider how it would account for the simplest case, exemplified by (2). A question immediately arises about what part of the two derivations is relevant to the computation of this condition. Every derivation will involve an overt part, the derivation to PF, and a covert part, the derivation to LF. In the overt part, (2a–b) would involve (27a–b) respectively—pretending that the subject does not originate inside VP to keep the representations simple.
The link between what and its trace in object position is obviously longer than the one between who and its trace. However, in the LF representations (3a–b) (repeated here for convenience), the links are the same for each derivation since both phrases have moved to the same two positions:
This shows that Shortest Movement applies in the overt syntax, not to LF. It also shows that Shortest Movement will not account for the difference between (22) and (25) since each example involves the movement of both wh-phrases to the same two positions.
The same situation obtains in the more complicated case of (28) where the two wh phrases start out in different clauses:
Taking (28a–b) to represent a pair of two convergent derivations in the overt syntax (PF), the link between who1 and its trace is, again obviously, shorter that the link between who2 and its trace. Thus, the difference in acceptability between (28a) and (28b) follows from Shortest Movement. Given Greed, the wh-phrase in complement object position cannot move to the [Spec, CP] of the complement because persuade does not select a [+ WH] C. (28b) is only a Superiority violation under the Form Chain analysis given that a trace of the moved wh-phrase will occur in the complement [Spec, CP], whereas under a Move α analysis (assuming Greed) it would be both a Superiority and a Subjacency violation.
As with the analysis of (11), these two analyses can be distinguished by replacing the bare interrogative pronouns in (28) with which-phrases:
(29b) is perfectly normal. However, the derivation of (29b) is identical to that of (28b), which is blocked by Shortest Movement. This contrast argues against Shortest Movement as formulated in (26). The condition is simply too general to make the fine distinctions apparently needed.
Having dispensed with Shortest Movement, we can return to the two analyses of (28b) under Greed. With Form Chain (28b) is only a Superiority violation. Given this we would expect (29b) to be normal, as it is. Under the Move α analysis (29b) would still be a Subjacency violation since (28b) is both a Superiority and Subjacency violation. Thus (29b) shows that (28b) is only a Superiority violation, thereby providing further empirical evidence for the Form Chain analysis of movement. That is, Move α appears to be incompatible with Greed.
Notice that there are other constructions whose derivations involve no apparent difference between the Move α and Form Chain analyses and about which Shortest Movement also makes the wrong prediction. The derivation of (30a-b) involves one NP movement from complement subject to matrix subject and one wh-movement from matrix clause to [Spec, CP]:
(30b) should be blocked by (30a) under Shortest Movement. As might be expected, the same phenomena occur with passive constructions, which also involve one NP movement:
As illustrated, in (32) and (33), these constructions can be recast as indirect question complements of forget and wonder:
Moreover, all of these except (33a) can be turned into direct questions with only a slight degradation in acceptability, as in (34) and (35):
(35a) violates the selectional property of wonder, hence the derivation crashes at PF because the [+ WH] feature in C of the CP complement of wonder is not checked. In (35b) that feature is checked by the movement of the complement subject which students into the complement [Spec, CP]. Notice that there are two possible derivations here using Form Chain. Given that there is a [+ WH] feature in C of the CP complement of wonder, which books could move to the [Spec, CP] of that complement, forming a chain, and then move to the matrix [Spec, CP], forming a second chain. Alternatively, the wh-phrase could move directly to the matrix [Spec, CP], forming a single chain. Economy considerations along the lines of Least Effort (see Chomsky 1991) would preclude the first option. The wh-feature attached to the CP complement of wonder will be checked in PF by the overt movement of which students from [Spec, IP] to [Spec, CP]. Note that this derivation provides another more complicated example that necessarily violates the Strict Cycle Condition.
So far nothing precludes the possibility of a cyclic derivation of constructions like (22) and (35b), where the long-distance movement would invariably result in a chain that violates Subjacency. This is not a problem, however, as long as the countercyclic derivation is available. In general when the theory allows a convergent and a nonconvergent derivation for the same string, the convergent derivation usually masks the nonconvergent one. Thus, we can ignore the cyclic derivation that gives a bad result, though ideally the cyclic derivation would be blocked by some principle of grammar.
Under the countercyclic Form Chain analysis of wh-movement discussed above, no Subjacency violation could occur in a construction with a single embedded clausal complement. Thus the standard cases of wh-island violations involving the movement of two wh-phrase complements of V, as in (36), can no longer be explained as Subjacency violations.
On the cyclic derivation the examples in (36) violate Subjacency in the usual way, but not on the countercyclic derivation. Therefore, something aside from Subjacency must prohibit the countercyclic movement inside the complement CP. Just prior to this movement Form Chain creates the following structures:
Since a trace is an empty category analogue of its antecedent, t1 in (37a) is an NP, whereas t2 in (37b) is a PP. The countercyclic movement of the wh-PP in (37 a) or the wh-NP in (37b) should be ruled out by the nondistinctness on substitutions because the traces are categorially distinct from the phrases that replace them. In this way both possible derivations of (36) have a bad outcome.
At this point we might wonder whether Subjacency plays any role at all regarding the movement of multiple wh-phrases in a sentence. That it does play a crucial role can be demonstrated with examples containing at least two successive sentential embeddings:
With respect to wh-movement the derivation of (38) involves two steps: which students moves from the subject of read to [Spec, CP] of the complement of remember, and which books moves from the object of read to the [Spec, CP] of the root clause. This can be accomplished either cyclically or countercyclically. In the cyclic derivation the movement of which books is a straightforward Subjacency violation. The application of Form Chain in this instance creates a chain, one of whose links will cross two IPs because the intervening [Spec, CP] is occupied by which students. The countercyclic derivation is somewhat more interesting. Form Chain moves which books to the root CP and creates a chain with traces in all the intervening [Spec, CP] positions. This chain observes Subjacency. The next step in the derivation requires Form Chain to move which students to the [Spec, CP] of the complement of remember. If the resulting structure is as given in (39) (where t2 is the trace of which books), then the chain formed by which students and its trace constitutes another Subjacency violation.
(39) results only if Form Chain cannot replace a trace with another trace. In contrast Form Chain can substitute a wh-phrase for a trace (for feature checking, as required by Greed). Note that even the cyclic derivation can be interpreted in this way. Thus Subjacency violations for such constructions follow as a consequence from the restriction on chain formation by the operation Form Chain.
The analysis of wh-movement presented here has been motivated by principles of economy, primarily Greed with assistance from Least Effort. Taking Greed to be perhaps the major motivating force for substitution operations leads to the choice of Form Chain over Move α as the proper formulation for the movement transformation. As a result successive cyclic movement is revealed as an illusion, its effects subsumed under Subjacency (applied to the chain links created by Form Chain) and a rather natural prohibition against replacing one trace with another via Form Chain. The analysis of Superiority phenomena demonstrates that Form Chain is not subject to a Shortest Movement constraint (as proposed in Chomsky 1993) and therefore plays no role in motivating Form Chain over Move α. Furthermore, the analysis of wh-movement discussed in this paper suggests that some derivations may have to be countercyclic, not strictly cyclic. Therefore derivations may be countercyclic or cyclic, with improper consequences ruled out by other factors.
Much research in generative grammar over the past four decades has quite properly attempted to identify and explore large-scale generalizations about syntactic processes, including the A-over-A Principle, the wh-Island Condition, the Strict Cycle Condition, Subjacency, Koster’s Locality Principle (Koster 1978b), Connectedness (Kayne 1983), Global Harmony (Koster 1986), and currently the Shortest Movement Condition. It seems to me that the history of the field has shown that these large-scale syntactic generalizations tend to break down at various points and have to be restricted or replaced by more fine-grained principles (e.g., Greed, which essentially concerns morphological properties). It is therefore not surprising to find that a closer look at superiority phenomena and wh-movement under minimalist assumptions leads to the conclusion that successive cyclic movement is illusory, that derivations can be countercyclic and hence the empirical effects of the Strict Cycle Condition must be derivable from other principles of grammar, that the Shortest Movement Condition does not appear to be viable, and that the concept “wh-island” turns out to be completely spurious.
In the middle of Samuel Beckett’s play Endgame a character exclaims (“With fervour”), “Ah the old questions, the old answers, there’s nothing like them!” This is a profoundly irrational attitude, as anyone actively engaged in rational inquiry (including the natural sciences, as Carlos Otero pointedly notes in the quote at the beginning of this paper) knows. And as any generative grammarian who has been following the field over the past few decades knows all too well, it is demonstrably false. The great power of rational inquiry is that it can lead us to see the world in a new and hopefully clearer light. It liberates us from the tyranny of habit and the domination of concepts whose authority over us is without real justification. The more accurate our understanding of the world, the better our chances of finding real solutions to the problems we study. One hopes that this uniquely human ability can be applied to other aspects of human activity to improve the way we live as individuals and in society.
