2.1. Sub-extraction from subjects

Let us first look at English. Sub-extraction from the canonical subject position (labeled SpecTP in this paper) is impossible in English, as shown by the examples in (4).

1. (4)

   1. (a) *[_CP [Which Marx brother]_i did [_IP she say [_CP [_IP [a biography of t_i] is going to be published/will appear this year]]]]?

   2. (b) *[_CP [Which Marx brother]_i did [_IP she say [_CP that [_IP [a biographer of t_i] interviewed her/worked for her]]]]? (Lasnik & Park Reference Lasnik and Park2003: 650)

This fact has commonly been derived from Huang's (Reference Huang1982) Condition on Extraction Domains, and more specifically, the Subject Condition (Chomsky Reference Chomsky, Anderson and Kiparsky1973), which says that extraction from the canonical subject position is prohibited.Footnote ⁴ This is further confirmed by the data in (5), where extraction is possible from a position below the canonical subject position.

1. (5)

   1. (a) [_CP [Which candidate]_i were [_IP there [posters of t_i] all over the town]]?

   2. (b) *[_CP [Which candidate]_i were [_IP [posters of t_i] all over the town]]? (Lasnik & Park Reference Lasnik and Park2003: 651)

A number of authors have argued that the canonical subject position is generally a freezing position (see Ormazabal, Uriagereka & Uribe-Etxebarria Reference Ormazabal, Uriagereka and Uribe-Etxebarria1994; Takahashi Reference Takahashi1994; Stepanov Reference Stepanov2001; Boeckx Reference Boeckx2003, Reference Boeckx2008c: Chapter 5; Lasnik & Park Reference Lasnik and Park2003; Rizzi Reference Rizzi, Lai-Shen Cheng and Corver2006; Rizzi & Shlonsky Reference Rizzi, Shlonsky, Gärtner and Sauerland2007).Footnote ⁵ There is much cross-linguistic evidence for this view: (6) shows this for Spanish, (7) for Norwegian, (8) for Icelandic, (9) for Danish, and (10) for Swedish.Footnote ⁶

1. (6) [_CP [De qué conferenciantes]_i te parece [_CP que … (Spanish) of what speakers cl.to.you seem.3.sg that

   1. (a) [_IP me van a impresionar [las propuestas t_i]]]]?cl.to.me go.3.sg to to.impress the proposals

   2. (b) *[_IP [las propuestas t_i] me van a impresionar]]]? the proposals cl.to.me go.3.sg to to.impress‘Which speakers does it seem to you that the proposals by will impress me?’ (Uriagereka Reference Uriagereka1988: 118)

1. (7) *[_CP Hvem_i tror du [_CP at [_IP [brev fra t_i] kommer i morgen]]]? (Norwegian) who think you that letters from come in tomorrow‘Who do you think that letters from come tomorrow?’

1. (8) *[_CP Hverjum_i telur Jens [_CP að [_IP [mynd af t_i] hangi á skrifstofu Willochs]]]? who believes Jens that photograph of hangs on office Willoch's (Icelandic)‘Who does Jens believe that a photograph of hangs in Willoch's office?’

1. (9) *[_CP [Hvilken forfatter]_i blev [_IP [bøgerne af t_i] hurtig udsolgt]]? (Danish) which author were books.def of soon sold.out‘Which author were books by soon sold out?’

1. (10) *[_CP [Vilken kung]_i hänger [_IP [många porträtt av t_i] på Gripsholm]]? (Swedish) which king hang many portraits of at Gripsholm‘Which king hang many portraits of at Gripsholm?’ (Engdahl Reference Engdahl, Engdahl and Ejerhed1982: 164)

The Spanish examples in (6) are especially revealing as the contrast shows that extraction is possible from SpecvP, (6a), but not from SpecTP, (6b), assuming that in Spanish post-verbal subjects stay in SpecvP (see e.g. Gallego Reference Gallego2007). This is not a peculiar feature of Spanish, however; we have already seen in (5) that this is allowed in English ((5a) is repeated in (11) for convenience):Footnote ⁷

1. (11) [_CP [Which candidate]_i were [_IP there [posters of t_i] all over the town]]?

So far I have illustrated that SpecTP is a freezing position. The question is why this is the case. Here I will rely on the theory put forward in Boeckx (Reference Boeckx2008c: Chapter 5). By way of introducing his theory, I will focus on (pure) extraction and then return to sub-extraction once we have an understanding of how extraction works.

Crucial to Boeckx' approach is the understanding of Case. He follows Pesetsky & Torrego (Reference Pesetsky, Torrego and Kenstowicz2001: 361), who define nominative Case as follows:

1. (12) Nominative Case

   Nominative case is uT on D.

uT stands for an uninterpretable [T(ense)]-feature, and ‘on D’ means that it is present on a nominal phrase (i.e., a DP). Based on this, Boeckx (Reference Boeckx2008c: 172) formulates the domain of nominative Case as in (13). That is, (13) is the position where nominative Case is licensed.

1. (13) [Fin° _[+T] [– T° _[−ϕ]]]

The structure in (13) incorporates Chomsky's (Reference Chomsky, Gärtner and Sauerland2007, Reference Chomsky, Freidin, Otero and Zubizaretta2008) claim that assignment of nominative Case is dependent on a finite C, and assumes Rizzi's (Reference Rizzi and Haegeman1997: 288) cartography of the CP:

1. (14) ForceP … (TopicP) … (FocusP) … FinP TP

The ϕ-feature bundle on the functional head in (13) is present to indicate the presence of a Probe–Goal relation between the functional head and a DP (Boeckx Reference Boeckx2008c: 172). Phi-features are unvalued (hence the ‘minus’ sign) and they get valued through subject–verb agreement. Fin° comes with a [+T]-feature from the numeration, just like T° does (not shown in (13)). Here Boeckx follows Chomsky's (Reference Chomsky, Gärtner and Sauerland2007, Reference Chomsky, Freidin, Otero and Zubizaretta2008) claim that Fin° is the relevant Case licenser and not T°.Footnote ⁸

An assumption significant from the present point of view, which will become clearer as we go along, is that an item is frozen when it has checked a feature in a special position.Footnote ⁹ This follows from what Boeckx (Reference Boeckx2003) calls the Principle of Unambiguous Chain (PUC; see also N. Richards Reference Richards2001, Rizzi Reference Rizzi, Lai-Shen Cheng and Corver2006). The PUC says that a chain can contain at most one ‘strong’ position/occurrence; that is, an item can only move to one strong position. A strong position/occurrence is defined in terms of checking of a feature associated with an EPP property (Boeckx Reference Boeckx2008c: 165). Put slightly differently, movement to the canonical subject position is movement to a strong position. Obviously, there might be other positions that bear an EPP property, and the idea is that each such position counts as a strong position.Footnote ¹⁰ If an item moves to such a position, it cannot move any further. Richards (Reference Richards2001) has suggested a quite similar idea. He first suggests that the EPP property of an element α gives a signal to PF that α should be pronounced in the relevant context, and then formulates a condition on what constitutes a legitimate chain at the interface. In short, Richards argues that PF must receive unambiguous instructions about which part of a chain to pronounce. The EPP property is by hypothesis one such signal that gives an unambiguous instruction to the interfaces. In summary, we see that for both Boeckx and Richards, there can only be one strong position/EPP-position per chain, and this is central to our understanding of freezing effects. However, it should be pointed out that it is not clear that the system has to be organized exactly like this: The claim about there being only one strong position per chain is obviously axiomatic.

The picture is slightly more complicated than we have seen so far, though. The present theory suggests that an element can only move to a single feature-checking site, that is, remerge once (Boeckx Reference Boeckx2008c: 167; see also Boeckx Reference Boeckx2003 & Bošković Reference Bošković2007; Ormazabal et al. Reference Ormazabal, Uriagereka and Uribe-Etxebarria1994 contains an earlier and quite similar proposal). How does one define this single feature-checking site? On the assumption that a chain is defined by two domains, one associated with the External first-Merge position, the other with the Internal Merge position, Boeckx suggests the following characterization of what he calls a ‘checking domain’ (the diagram in (15) is adapted from Boeckx Reference Boeckx2008c: 171):Footnote ¹¹

1. (15)

   

Here we have two functional heads, H₁° and H₂°, and there is a specifier, marked ‘–’, where an item can undergo checking for the feature F. The original position of the element E that will remerge is also included. Crucially, this element will have to move into ‘–’. Unless the element E is externally merged in the ‘–’ position in (15), movement to this position is required for E to establish a Probe–Goal relation with H₁°. If movement does not happen, a relativized minimality violation will occur: H₁, H₂ and E all share a feature and they stand in an asymmetric c-command relation. Boeckx assumes that H₁ is the Probe and that it has to probe E, and given the intervening H₂, we have a standard case of intervention or relativized minimality. The only way E can escape the minimality effect is to move above H₂, that is, to remerge in ‘–’; if no movement happens, the derivation crashes. Upon remerge, an unambiguous checking site, of which only one can be established per chain, has been established, as discussed above.

Although I have kept the discussion of checking domains fairly general, the reader will not fail to recognize that the configuration in (15) is similar to the canonical subject position. (16) shows a schema for the assignment of nominative Case, where Fin° is assumed to be the bearer of nominative Case in English (Boeckx Reference Boeckx2008c, Chomsky Reference Chomsky, Freidin, Otero and Zubizaretta2008, Lohndal Reference Lohndal2009).Footnote ¹²

1. (16)

   1. (a) [Fin°_{[+T, −ϕ]} [– [T°_{[+T, −ϕ]}]] … [E_{[−T, +ϕ]}]] before valuation

   2. (b) [Fin°_{[+T, −ϕ]} [[E_{[−T, +ϕ]}] [T°_{[+T, −ϕ]}]] … [E]] movement of E

   3. (c) [Fin°_{[+T, +ϕ]} [[E_{[+T, +ϕ]}] [T°_{[+T, +ϕ]}]] … [E]] valuation

Boeckx (Reference Boeckx2008c) assumes that Fin° values nominative Case on the relevant DP (E in the structures above) without saying explicitly how Fin° actually qualifies as a Probe. Here I will interpret this to mean that Fin° also needs to have unvalued ϕ-features in order to be a Probe (see also Fortuny Reference Fortuny2008: Chapter 2 for extensive discussion of the link between Fin° and T°).Footnote ¹³ This is supported by complementizer agreement facts of the sort we find in West Flemish (Haegeman Reference Haegeman1992); see also M. Richards (Reference Richards2007) and Chomsky (Reference Chomsky, Freidin, Otero and Zubizaretta2008) for a more abstract version of this idea that generalizes to all languages.Footnote ¹⁴ E_[−F] moves to SpecTP and then the unvalued Case feature on E is valued. Since the elements Fin°, T° and E share a [T]-feature, movement is required. If there is no movement, the ϕ-features on Fin° will not get valued since the unvalued ϕ-features on T° will create a defective intervention effect. This means that (16) is an instance of a checking domain as defined, which means that the subject is frozen in the canonical subject position.

Interestingly, this theory also predicts that there should be no freezing effects when direct objects get their case checked (in a Pesetsky & Torrego Reference Pesetsky, Torrego and Kenstowicz2001, Reference Pesetsky, Torrego, Guéron and Lecarme2004 way). Direct objects are externally first-merged in their checking configuration as a complement to V°, where v° bears the relevant Case feature, on standard assumptions. (17) shows the relevant parts of the structure, though I will return to this in greater detail in Section 3.1.Footnote ¹⁵

1. (17) [v°_[+T] [V° [DP_[−T]]]]

Although direct objects are externally merged, they of course enter into checking, but not in a configuration of the kind given in (15). The latter assumption is motivated by the fact that raising of objects do not seem to be required in overt syntax in English (Lasnik Reference Lasnik1999, Reference Lasnik2003). Within Boeckx' theory, direct objects are eligible to move further. Below we will see that this fits with the data. However, note that the present account encompasses a disjunction: DPs can check Case in two different ways – via External first-Merge as in the case of direct objects and via Internal Merge in the case of subjects moving to SpecTP. Below we will see that direct objects can in fact also be checked in a derived position, in which case a freezing effect occurs. At the moment, I have no good reason for why there is a disjunction when it comes to checking positions for direct objects, so it remains a stipulation, as in most other theories.

This line of thought can be extended to the sub-extraction cases discussed above. Boeckx (Reference Boeckx2008c: 196) says that ‘in the illicit cases, the point of origin is a complete chain … that cannot be extended further’. It is not clear exactly what this means since it is not clear that there actually is a chain inside the entire constituent. However, what I think Boeckx is trying to say is that when the entire constituent has been checked in a checking domain, none of its internal parts can be extracted. This requires that what happens to the big constituent (e.g. Case checking) also happens to the constituent's internal parts. There exist sophisticated theoretical ways of achieving this; see Boeckx (Reference Boeckx2008c) for one suggestion.Footnote ¹⁶ Basically he suggests that when sub-extraction fails, the point of origin for sub-extraction is a complete chain that cannot be extended further. That is, since the constituent has already moved into a checking domain, we would get an ambiguous chain if parts of the moved constituent moved further since these parts would eventually have to reach another checking domain. This accounts for why sub-extraction out of in-situ subjects is better than sub-extraction out of displaced subjects, as we have seen illustrated above (see e.g. example (5)).Footnote ¹⁷ Thus we see that Boeckx assumes that the accounts of extraction and sub-extraction are very similar. This is not a logical necessity in his framework, but it is an assumption that he adopts. However, when we look at various data below, we will see that there are clear differences between the two types of extraction, and that more theoretical power is required to prevent sub-extraction than to prevent extraction. In particular, I will argue that Case is required in understanding why sub-extraction is more restricted than extraction if we want to treat extraction and sub-extraction in similar ways, as Boeckx does. This will lead me to address the difference between extraction and sub-extraction more extensively, which I will do in Section 4. The upshot of the discussion will be that the theory in Boeckx (Reference Boeckx2008c) does not offer a satisfactory explanation of why sub-extraction is prohibited given that it makes use of at least three ways to rule out the bad cases.

2.2. Extraction of subjects

Sub-extraction from subjects was shown to relate to the assumption that the subject moves into a checking domain, i.e. an unambiguous feature-checking site. We expect the same to be true of subjects being extracted, but a complication enters the picture at this point. It is a well-established fact that languages are subject to variation when it comes to subject extraction, and this is commonly related to that-trace effects (see Lohndal Reference Lohndal2009 for discussion). In this section, I will briefly show how the theory in Boeckx (Reference Boeckx2008c) and Lohndal (Reference Lohndal2009) derives that-trace effects, though I will discuss some examples that motivate changing some of the assumptions that Boeckx and Lohndal made. Since the main purpose of this section is to introduce the details of the framework, I will only focus on English; see Lohndal (Reference Lohndal2009) for an extensive discussion of how to accommodate the variation we find across dialects and languages.

The standard case of that-trace effects from English is given in (18).Footnote ¹⁸

1. (18)

   1. (a) *What do you think that is in the file?

   2. (b) What do you think is in the file?Footnote ¹⁹

Boeckx (Reference Boeckx2008c: 178) suggests that movement is allowed out of the embedded finite clause in (18b) because [T]-feature on Fin° is deactivated. When Fin° no longer has a valued [T]-feature, it cannot value the Case feature of the subject DP. That is, we have two configurations, as shown in (19).

1. (19)

   1. (a) [Fin°_[+T] [– T°_[−ϕ]]] subject extraction disallowed

   2. (b) [Fin°_[T] [– T°_[−ϕ]]] subject extraction allowed

The Standard English that-trace effect illustrated in (18a) above exhibits the configuration in (19a). When that is missing, as in (18b), Fin° does not have a valued [T]-feature, as is shown in (19b). Thus the DP does not receive Case, and can freely move to the main clause, where its Case feature is valued. I adopt the standard assumption that the subject moves through SpecTP even if there is no checking in this position, as long as the subject still needs to check some feature (see Boeckx Reference Boeckx2007 & Bošković Reference Bošković2007 for a defense of the latter claim). However, Boeckx (Reference Boeckx2008c) does not tell us where the Case feature of what is valued. One answer is provided by Lohndal (Reference Lohndal2009), who argues that what receives Case from the matrix Fin°. Unfortunately, this solution predicts (20) to be grammatical, contrary to fact.

1. (20) *Who was it expected t to kiss Mary?

Since the theory in Lohndal (Reference Lohndal2009) cannot account for this case, in this paper I argue that who actually receives Case from the matrix verb expected, similarly to Kayne (Reference Kayne1980), who argued that a matrix verb can assign Case to an element in COMP. On this assumption, I return to cases like (18b), here using the example in (21a). The relevant parts of a derivation where there is no that-trace effect are shown in (21b–e), where I ignore any other feature checking except Case.

1. (21)

   1. (a) Who do you think left?

   2. (b) [_Fin°[+T] do [_TP [_vP you_[−T] [_v°[+T] think [_FinP[T] [_TP who_[−T] left]]]]]]

   3. (c) [_Fin°[+T] do [_TP you_[+T] [_vPyou [_v°[+T] think [_FinP who_[−T] [_FinP[T] who_[−T] [_TP left]]]]]]]

   4. (d) [_Fin°[+T] do [_TP you_[+T] [_vPyou [_v°[+T] think [_FinP who_[+T] [_FinP[T]who [_TP left]]]]]]]

   5. (e) [_FinP who_[+T] [_Fin° do [_TP you_[+T] [_vPyou [_v°[+T] think [_FinPwho_[+T] [_FinP[T]who [_TP left]]]]]]]]

The derivation stages (21b–c) show how the main-clause subject you gets Case whereas (21d–e) show how the moved embedded subject who gets Case. The embedded SpecTP is not a checking domain because the [T]-feature on Fin° is deactiviated, thus the relevant relativized minimality situation will not emerge.Footnote ²⁰ This means that who is allowed to move further, potentially triggered by an EPP-feature. Note that this gives us a way to analyze (20) as well: On the standard assumption that accusative Case is absorbed in passives, there will be no Case checker for who if it is merged as the subject. Thus the sentence is predicted to be unacceptable.

Regarding dialects of English that do not exhibit that-trace effects (Sobin Reference Sobin1987, Reference Sobin2002), these show that the only possible configuration is (19b). Subject extraction is possible regardless of whether the complementizer is present or not. I am assuming that who in (22a) gets Case in the same way as in (21) above (pace Lohndal Reference Lohndal2009):

1. (22)

   1. (a) Who do you think that left?

   2. (b) [_Fin°[+T] do [_TP[−ϕ] [_vP you_[−T] [_v°[+T] think [_{Fin° [T]} that [_TP who_[−T] left]]]]]]

   3. (c) [_Fin°[+T] do [_TP[−ϕ] you_[+T] [_vPyou [_v°[+T] think [_FinP who_[−T] [_Fin°[T] that [_TP who_[−T] [_TP left]]]]]]]]

   4. (d) [_Fin°[+T] do [_TP you_[+T] [_vPyou [_v°[+T] think [_FinP who_[+T] [_Fin°[T] that [_TPwho [_TP left]]]]]]]]

   5. (e) [_FinP who_[+T] [_Fin°[+T] do [_TP you_[+T] [_vPyou [_v°[+T] think [_FinPwho [_Fin°[T] that [_TPwho [_TP left]]]]]]]]]

As we see, the only difference between the English derivations is whether the [T]-feature on Fin° is activated or not.Footnote ²¹ This crucially determines whether subject extraction is possible.Footnote ²²

Before moving on, we need to pause to look at a different case, namely (23), brought to my attention by Howard Lasnik (p.c.).Footnote ²³

1. (23)

   1. (a) I know who to say [t solved the problem].

   2. (b) It is unclear who to say [t solved the problem].

In (23), there does not seem to be a Case assigner for who, which seems to be a problem for the theory I have adopted here. However, the solution I offered for the cases where who is allowed to move out of an embedded finite clause seems to work for (23) as well. That is, I argue that say assigns Case to who in SpecCP in the examples in (23).Footnote ²⁴ Specifically, I assume that say has a valued [T]-feature that can check the unvalued [T]-feature on who. Since the C head of the CP that say embeds has a deactiviated [T]-feature, we do not have a checking domain in the sense of (15). Hence who can move further. One may wonder why who cannot remain in this lower position (*I know to say who solved the problem); this is presumably due to the fact that the matrix predicate selects for a wh-complement.Footnote ²⁵

This account does not derive the unacceptability of (24a), though the latter seems to be ruled out on independent grounds given that (24b) is also bad.

1. (24)

   1. (a) *I know who to say to solve the problem.

   2. (b) *I know why to say to solve the problem.

Note, however, that this account does handle the following data accurately:

1. (25)

   1. (a) *I don't know who to tell Bill to solve the problem.

   2. (b) I don't know why to tell Bill to solve the problem.

   3. (c) I don't know who to tell Bill to have t solve the problem.

In (25c), who has its Case checked by know. (25b) should be good since it involves an adjunct that does not need to be checked for Case, and (25a) should be bad because who has not moved and hence there is no trace. Since there is no trace, there is no meaningful interpretation.

The claim that say can assign Case to who in (23) may be complicated by the following data:

1. (26) I know who to say very clearly [t solved the problem].

Here the adjunct very clearly intervenes such that there is no adjacency between the verb and who. However, we know that adjacency matters in other cases, as is shown in (27a–c).

1. (27)

   1. (a) I sincerely believe John.

   2. (b) *I believe sincerely John.

   3. (c) *I believe very sincerely [John to be smart].

   4. (d) Who do you believe very sincerely [t to be smart]?

As Chomsky & Lasnik (Reference Chomsky and Lasnik1977) pointed out, sentences like (27c) improve if the subject is an extracted wh-element, as in (27d), which lends support to the analysis above. The generalization seems to be that adjacency matters for licensing elements in situ and ECM cases, but not when a constituent is in the specifier of CP. I do not have a good answer to why adjacency does not matter in the latter case and have to set that aside for future research.

There is a lot of empirical evidence in favor of this approach to that-trace effects and Comp-trace effects more generally. For instance, Lohndal (Reference Lohndal2007b, Reference Lohndal2009) argues that this approach can derive the variation among the Scandinavian languages when it comes to that-trace effects, and that it can also deal with relative clauses and extraction out of these. For reasons of space, I have to refer the reader to Boeckx (Reference Boeckx2008c) and in particular Lohndal (Reference Lohndal2007b, Reference Lohndal2009) for details. In the next section, I discuss local subject questions.

2.3. Local subject questions

So far I have discussed extraction from embedded questions and briefly touched on sub-extraction from subjects. I will now argue that this approach can also account for local subject questions like (28).

1. (28) Who told the story?

There is a lot of discussion in the literature as to whether the subject in (28) is in SpecCP or in SpecTP (see, among others, Travis Reference Travis1984, Vikner Reference Vikner1995, Zwart Reference Zwart1997). Rizzi & Shlonsky (Reference Rizzi, Shlonsky, Gärtner and Sauerland2007) discuss this question thoroughly, and conclude that it does reside in SpecCP. They advance an argument in favor of this position based on languages with overt question morphemes, like Imbabura Quechua:

1. (29) pi-taj shamu-rkawho-q left-agr‘Who left?’)

Rizzi & Shlonsky (Reference Rizzi, Shlonsky, Gärtner and Sauerland2007) argue that there is a covert expletive (called there below) directly merged in SpecTP, making it possible for the wh-phrase to circumvent being frozen in SpecTP. The theory I have adopted so far is neutral on whether the subject is in its Case position or in an A-bar position (Boeckx Reference Boeckx2008c: 193). I will follow Rizzi & Shlonsky's proposal and assume that there is a silent there in SpecTP. According to Koopman (Reference Koopman and Boeckx2006), there is evidence for a silent there in other languages (see also Rizzi & Shlonsky Reference Rizzi, Shlonsky, Gärtner and Sauerland2007), so (28) and (29) may be other such cases. What would falsify my claim is a case where the sequence ‘who there V’ is bad, as it would show that there cannot be a silent there in a matrix clause with wh-movement across the board in English. Bearing this point in mind, I need to clarify how the wh-phrase will get Case in an example like (30a). The relevant parts of the derivation are given in (30b–e).

1. (30)

   1. (a) Who left?

   2. (b) [_Fin°[+T] [_TP[−ϕ]there_[−T] [_vP who_[−T] left]]]

   3. (c) [_Fin°[+T] [_TP[−ϕ]there_[+T] [_vP who_[−T] left]]]

   4. (d) [_FinP who_[−T] [_Fin°[+T] [_TP[−ϕ]there_[+T] [_vPwho left]]]]

   5. (e) [_FinP who_[+T] [_Fin°[+T] [_TP[−ϕ]there_[+T] [_vPwho left]]]]

I assume that there is no intervention effect here because Case checking of there happens prior to A-bar movement of who, which may also be required to prevent a Multiple Agree configuration if the arguments in Haegeman & Lohndal (Reference Haegeman and Lohndal2010) extend to the present case. As for the details of the derivation, I assume that the wh-phrase bears a [−T]-feature, commands Fin° and is in a possible Agree configuration. Fin° is therefore responsible for valuing the Case feature on both the expletive and the wh-phrase. This account is fully in line with the overall theory of the present paper, namely that extraction/further movement is only possible when an item does not enter into a checking domain. For local subject questions, the silent expletive ensures that the first checking domain that the wh-phrase is able to enter is in SpecFinP.

The reader may at this point wonder how I would analyze V2 languages like Norwegian, given that most theories assume that the subject is in the A-bar domain in declarative sentences like (31).

1. (31) Han fortalte historien.he told story.def‘He told the story.’

Here I will have to assume that the Case feature [+T] is on the Force head in declarative main clauses (irrespective of whether there is a that-trace effect or not; see the variation in the Scandinavian languages discussed in Lohndal Reference Lohndal2009), which will then trigger movement to the C-domain. V2 is arguably the cue responsible for this fact since V2 generally only holds in main clauses. In V2 languages, the verb has been argued to target a position higher than Fin° in the left periphery (Westergaard & Vangsnes 2005, Bentzen Reference Bentzen2007, Julien Reference Julien2007), and it is thus reasonable to assume that V2 activates an articulated C-area (it ensures that Fin° and Force° are split, cf. Rizzi Reference Rizzi and Haegeman1997). Notice that this split ensures that the subject has to move into the C-domain. This movement is required for Case reasons, and made possible since SpecTP is no longer a checking domain due to the lack of a Case feature. This account also fits with the established fact that most Norwegian dialects do not have embedded V2 (though see Bentzen Reference Bentzen2007 for complications and some dialectal variation), which means that we do not expect subjects in embedded clauses to move to the C-area since there is no V2 that may trigger this movement.

In this section we have seen how local subject questions can be analyzed. In the next section I turn my attention to objects and some extraction phenomena from a comparative perspective.

3.1. Direct objects

In Section 2 I argued that there is a special freezing position for subjects, namely SpecTP. According to Rizzi (Reference Rizzi, Lai-Shen Cheng and Corver2006: 124), there is no corresponding freezing position for objects. The goal of this section is to show how this can be implemented to account for the Norwegian data in (32)–(33).

1. (32)

   1. (a) Hva tror du Peter vet t?what think you Peter knows‘What do you think Peter knows?’)

   2. (b) Hva tror du Peter maler et bilde av t?Footnote ²⁶what think you Peter draws a painting of‘What do you think Peter draws a painting of?’

1. (33)

   1. (a) Hva gav du barna t for noe til bursdagen?what gave you children.def for something to birthday.def

   2. (b) Hva gav du barna t til bursdagen?what gave you children.def to birthday.def‘What did you give the children for their birthday?’

As these data show, there are no freezing effects related to direct objects. This follows straightforwardly from the theory advocated for here as direct objects do not check their case in a checking domain (see also Section 2) but normally get their Case checked in situ. The Case of the direct object will be assigned by v° as in (34). I am assuming that the direct object is merged as a complement to V° and, following Pesetsky & Torrego (Reference Pesetsky, Torrego, Guéron and Lecarme2004) and Boeckx (Reference Boeckx2008c), I take it that the abstract representation of accusative Case is a [+T]-feature, and that the feature is morphologically realized either as nominative or accusative depending on the environment in which it occurs.

1. (34) [v°_[+T] [V° [DP_[−T]]]]

The structure in (34) is an ambiguous checking site; the direct object DP is externally first-merged in its position (see also the discussion above related to (15)). Thus this checking site does not prevent the object from moving further since it is not a checking domain as defined above. Nor does it prevent sub-extraction from the object. As (27) and (28) show, this account gives us the correct facts.

Interestingly, there is a peculiar asymmetry in the data discussed so far. We have seen that sub-extraction from subjects is generally banned, but that both sub-extraction from an argument and subject extraction are possible in a post-verbal position, as in the Spanish and English examples in (6) and (4) above. A prediction seems to be that we do not expect the same to hold true of direct objects since they do not enter into a checking domain (I will discuss object shift in Section 3.3). However, this prediction fails. Consider Basque, where sub-extraction from direct objects is banned but extraction of the entire direct object is grammatical. This is illustrated in (35).

1. (35)

   1. (a) *Nori_i buruzko sortu zituzten aurreko asteko istiluek [zurrumurruak t_i]?who about.of create aux last week scandals rumors‘Who₁ have last week's scandals caused [rumors about t₁]?’ (Uriagereka Reference Uriagereka1998: 395)

   2. (b) Zer₁ egiten duzu zuk t₁ hemen?what do aux you.erg here‘What are you doing here?’ (Etxepare & de Urbina 2003: 464)

We see here that in Basque sub-extraction from direct objects is impossible, while as we have seen both Norwegian and English allow such sub-extraction. The question is how we are to account for this in terms of the framework we are developing. Following Laka & Uriagereka (Reference Laka and Uriagereka1987), Uriagereka (Reference Uriagereka1999a) and Boeckx (Reference Boeckx2003, Reference Boeckx2008a, Reference Boeckxb, Reference Boeckxc), I will argue that this extraction fact is related to properties of Basque agreement. Basque shows object ϕ-feature agreement with the verb, contrary to languages like English and Norwegian, and it is reasonable that this agreement makes it impossible to sub-extract from the object in Basque. This account is supported by a vast number of examples where the absence of agreement makes extraction possible (see in particular Boeckx Reference Boeckx2003, Reference Boeckx2008a, Reference Boeckxb, Reference Boeckxc; Baker & Collins Reference Baker and Collins2006; see also the analysis of that-trace effects above). Verb–direct-object agreement yields the same effect as moving to a checking domain does, namely that sub-extraction becomes impossible. Boeckx (Reference Boeckx2003: 104) argues that the target for sub-extraction would be ϕ-complete, in which case sub-extraction is not licit. For this to be true in the present case, it would mean that verb–direct-object agreement somehow makes the entire direct object ϕ-complete. It is not entirely clear how this can be done technically and I leave the implementation of this for future research. For our purposes, it suffices to note the strong correlation between agreement and lack of sub-extraction. Notice, though, that since the direct object itself has not entered a checking domain, movement of the entire object is still possible. Thus there are now two ways in which sub-extraction of a DP becomes impossible: either by entering a checking domain (the case of subjects), or if there is agreement (in ϕ-features) between a verb and the DP (the case of objects).

We now have an asymmetry between extraction and sub-extraction. Extraction is ruled out whenever a constituent enters a checking domain. Sub-extraction is either ruled out by a constituent entering a checking domain or by ϕ-feature agreement. This already suggests that sub-extraction is more restricted than extraction since the environment where full extraction is impossible is a subset of the environments where sub-extraction is impossible. This is something that the theory in Boeckx (Reference Boeckx2008c) cannot capture as it stands. I have argued that a disjunction is necessary in order to capture the data, which suggests that Boeckx' theory does not offer an explanation of why sub-extraction is prohibited even if it looks like we are roughly dealing with the same type of structure each time sub-extraction is bad (that is, extraction out of a complex phrase). In the next section I will argue that this conclusion is strengthened by the fact that in order to account for indirect objects in Norwegian, we need yet a third mechanism to rule out sub-extraction.

3.2. Indirect objects

The data that this section seeks to account for are provided in (36).

1. (36)

   1. (a) *Hva gav du [t for en fyr] pakker til bursdagen?what gave you for a guy gifts for birthday.def

   2. (b) Hvem gav du t pakker til bursdagen?who gave you gifts for birthday.def‘Who did you give gifts for their birthday?’

The example in (36a) shows that it is impossible to sub-extract from an indirect object, whereas (36b) shows that it is possible to move the entire indirect object. As we have seen, Basque shows the same contrast with direct objects, but this section argues that the two cases are to be accounted for differently.

We know that indirect objects are special. They are claimed to be inherently theta-marked (and thus linked to theta-selection, see Chomsky Reference Chomsky1986), and often get dative case across languages. My account of the observed contrasts found with indirect objects is that the inherent Case assigned to them makes them opaque, thus sub-extraction is impossible (cf. Řezáč Reference Řezáč, Harbour, Adger and Béjar2008 on Basque and other languages, though see Uriagereka Reference Uriagereka2008: 157–158 for complicating factors for Basque). This holds because the Probe would not be able to locate any element within the indirect object that it could extract. In other words, inherent Case acts like a ‘shield’, preventing any probing into its inner structure. This shield might plausibly be a PP shell, as suggested by Kayne (Reference Kayne1984), Bittner & Hale (Reference Bittner and Hale1996) and more recently Řezáč (Reference Řezáč, Harbour, Adger and Béjar2008), although admittedly this in itself does not guarantee non-probing; however, see Řezáč for a specific proposal using phase theory. Indirect objects are therefore moveable as units, but it is impossible to sub-extract from them. Arguably, the reason they are movable as units is that they do not enter a checking domain, thus they can freely move to the left periphery of the clause. However, in Section 4 we will see that this picture is cross-linguistically more complicated in interesting and non-trivial ways.

Before we move on to briefly discuss object shift and scrambling, we should pause to reflect on what it means that we have three different factors that prevent sub-extraction: entering a checking domain, triggering agreement or being assigned inherent Case. These three factors make it clear that we should be well-prepared to deal with the empirical generalization that sub-extraction is more restricted than full extraction, though it also means that we are either missing a generalization or that we cannot really explain sub-extraction. This paper will suggest that the latter is more likely the case. Intuitively, it also makes sense that we need more factors to rule out a quite diverse set of environments where sub-extraction is banned, compared to the rather homogeneous environments where extraction is banned (the latter being by and large confined to checking domains, if Boeckx Reference Boeckx2008c and the present paper are correct). This is non-trivial from a theoretical perspective since it can be taken to mean that the prospects for developing a theory of sub-extraction are scant.

3.3. Some remarks on object shift and scrambling

The theory of Boeckx (Reference Boeckx2008c) makes an interesting prediction, which I have indirectly touched on already: when we have scrambling or object shift, the internal parts of the moved constituent should be frozen or immovable if the moved constituent enters a checking domain (earlier versions of this prediction are the Frozen Structure Constraint (Ross 1967/Reference Ross1986: 173), the Freezing Principle (Wexler & Culicover 1980: 179) and the Revised Extraction Constraint (Diesing Reference Diesing1992: 128). This prediction follows because a shifted object would have its Case checked in a checking domain. I am assuming that in cases of object shift and scrambling, Case is not checked until the shifted position is reached because the absence of checking in the base position is what makes object shift and scrambling possible in the first place. An abstract derivation is shown in (37), where the starting point looks something like (37a).

1. (37)

   1. (a) [_Agr°[+T] [DP_subjectv°_[T] [V° DP_[−T]]]]

   2. (b) [_Agr°[+T] DP_[−T] [DP_subject [v°_[T] [V° DP]]]]

   3. (c) [_Agr°[+T] DP_[+T] [DP_subject [v°_[T] [V° DP]]]]

What happens here is that the v° is deactivated and the Case feature is on a higher head, (37b), say Agr°, to use a familiar label (although the label should not be taken to have any theoretical significance). The deactivation is of the familiar sort that we saw in cases of subject extraction from finite embedded clauses without a complementizer (recall, for example, the analysis of English sentences without that as in (18b)). The object has to move in this case because otherwise the defective [T]-feature on v° will create an intervention effect which will block Case assignment. We can test that the derivation in (37) gives the right empirical results by looking at data related to these phenomena. I will use Icelandic and German data for illustration.

Icelandic does not have a general ban on floating quantifiers, as shown in (38).

1. (38)

   1. (a) Hún þekkir ekki [öll börnin].she knows not all children.def‘She doesn't know all the children,’

   2. (b) ?Börnin_i þekkir hún ekki [öll t_i].children.def knows she not all (Halldór Á. Sigurðsson, p.c.)

It is, however, impossible to strand a quantifier in a shifted position:

1. (39) *Börnin_i þekkir hún [öll t_i] ekki.children.def knows she all not (Halldór Á. Sigurðsson, p.c.)

These data seem to confirm our prediction: Shifted objects do not allow sub-extraction. However, this claim is somewhat complicated by the standard treatment of quantifier floating as in Sportiche (Reference Sportiche1988), where it is assumed that quantifiers can be floated in shifted positions. The data from Icelandic suggest that one cannot trivially equate quantifier float and sub-extraction, but that in this language, the lack of quantifier float in shifted positions may follow from conditions on sub-extraction.

Scrambling data from German provide the same picture as Icelandic, though without the additional complication. (40)–(41) illustrate that scrambled objects do not allow sub-extraction (for some complications, see Fanselow Reference Fanselow2001).

1. (40)

   1. (a) Was hat Otto immer [t für Romane] gelesen?what has Otto always for novels read‘What kind of novels has Otto always read?’

   2. (b) *Was hat Otto [t für Romane] immer gelesen?what has Otto for novels always read‘What kind of novels has Otto always read?’ (Diesing Reference Diesing1992: 129)

1. (41)

   1. (a) Über wen hat der Fritz letztes Jahr [ein Buch t] geschrieben?about whom has the Fritz last year a book written‘Who did Fritz write a book about last year?’

   2. (b) *Über wen hat der Fritz [ein Buch t] letztes Jahr geschrieben?about whom has the Fritz a book last year written‘Who did Fritz write a book about last year?’ (Müller Reference Müller1998: 125)

(40a) and (41a) show that sub-extraction from the base position is entirely licit, whereas (40b) and (41b) show that sub-extraction from the scrambled position is banned. This is in line with our prediction.

English provides further support. Here I draw on the observation in Lasnik (Reference Lasnik, Davies and Dubinsky2001) that an object is only an island for extraction when it has raised overtly; the contrast between (42a) and (42b) illustrates that. (See discussion of whether or not English has object shift in a series of papers in Lasnik Reference Lasnik1999, Reference Lasnik2003.)

1. (42)

   1. (a) Who_i did Mary call up [friends of t_i]?

   2. (b) ?*Who_i did Mary call [friends of t_i] up? (Lasnik Reference Lasnik, Davies and Dubinsky2001: 111)

In (42b) we have exactly the configuration shown in (37) above. The Case feature is on what I have labeled the Agr head in (42b), whereas it is on the v head in (42a). Thus the data from object shift and scrambling can be accounted for by the present theory. In the next section, I will look at cross-linguistic differences regarding indirect objects and their freezing properties.