Locality , control , and non-adjoined islands

The goal of this paper is twofold: empirically, it is shown that obligatory control (OC) into islands is not restricted to control into certain adjuncts, but can also involve non-adjoined islands. This poses a serious problem for the movement theory of control (MTC), whose analysis of OC into adjuncts crucially relies on the fact that adjunction is involved. Second, the paper seeks to explore to what extent control theory is compatible with phase theory based on a strict version of the Phase Impenetrability Condition (PIC). In order to reconcile these locality considerations with the observed control patterns in the context of islands, the paper assumes a moderately local relationship between controller and controllee. The basic idea of the proposed theory is that the controllee starts out as an empty argument which needs to be referentially identified under Agree. To this end, it moves from phase edge to phase edge (in accordance with the PIC) until it can be licensed by the controller. In contrast to the MTC, the target position of controllee movement is not the controller position itself; thus, control into islands (including non-adjoined islands) can be derived more easily, since the control relation can already be established when the controller is at the edge of the highest phase inside the island and the controller is merged in the next higher phase. Hence, the theory is compatible with phase theory and can in particular account for the observed control patterns involving adjoined and non-adjoined islands.


Introduction
Since the development of the movement theory of control (MTC), which argues that the controllee in obligatory control structures is but a residue of A-movement (see Hornstein 1999 et seq.), a lively debate has been going on concerning the adequate handling of control (for alternative approaches, cf., for instance, Landau 2000;2015;Culicover & Jackendoff 2001;2006).In this paper, I want to focus on the question of how local control dependencies actually are in view of the fact that, on the one hand, controller and controllee might be separated by an island, but on the other hand, phase theory demands a certain degree of locality.So the question arises as to how these two aspects can be unified in a theory of control.
In fact, one major conceptual goal of this paper is thus part of a bigger enterprise, namely to explore the consequences of taking the phase model seriously.The phase model adopts a local-derivational view of syntax in which the accessible domain is restricted by the Phase Impenetrability Condition (PIC; see Chomsky 2000 et seq.). 1 As a consequence, all syntactic operations are expected to be local, and apparent non-local dependencies must be reanalyzed in a way that allows a reconciliation with phase theory.So we need to take a look at all sorts of non-local syntactic phenomena and check whether they can be remodeled accordingly.One well-known example is to split up non-local movement relations into smaller steps by means of successive-cyclic movement, and there are many more examples in the literature where it has been proposed that apparent non-local relations should be analyzed in terms of local dependencies.For instance, with respect to long distance agreement, Legate (2005) argues that it involves cyclic agreement, and Polinsky & Potsdam's (2001) analysis of long distance agreement also aims to achieve an underlying local agreement configuration by means of covert movement to the edge of the embedded clause.Furthermore, Camacho's (2010) analysis of switch reference is also based on cyclic agreement operations, and early examples from binding theory (which do not yet involve the phase model, of course) include Pica's (1987) or Cole, Hermon & Sung's (1990) treatment of simple reflexives in terms of LF head-movement to derive long distance binding in a local way.Fischer's (2004;2006) theory of reflexivization then adopts the phase model and proposes a PIC-compatible analysis of anaphoric and pronominal binding; and this paper aims to investigate control relations from a PIC-oriented perspective.Why should control be subject to the PIC to begin with?As will be shown in section 2, control is sensitive to certain syntactic island effects, which means that it is governed by syntactic locality.Since it would not be parsimonious to assume that there is a syntactic relation that obeys some syntactic locality constraints and yet is not subject to the PIC, we are led to assume that control is indeed restricted by the PIC.
In a non-movement theory of control, controller and controllee typically occur in different clauses throughout the derivation, and their relationship is therefore non-local.This means that it is not compatible with the PIC without further assumptions. 2The MTC, by contrast, expresses the most local configuration one can think of, since controller and controllee are related by movement and thus represented by copies of the same DP; however, as the discussion in section 4.2 will show, in its current form, the MTC is not really compatible with phase theory either (cf.also Drummond & Hornstein 2014).Moreover, if we take seriously the notion that islands impose restrictions on syntactic movement, the idea that movement takes place all the way up to the controller's position (as suggested by the MTC) might be too radical given that obligatory control is possible into (certain types of) islands.In fact, proponents of the MTC have proposed some strategies to handle control into adjuncts in particular, but, crucially, these cannot be extended to non-adjoined islands.However, as section 2 reveals, control into islands of this type exists as well.An alternative is clearly needed.This paper will therefore explore a hybrid theory of control (HTC) which will combine aspects of both the MTC and non-movement approaches to control.It will assume that there is movement (to model control in terms of a local relationship and thereby make it compatible with phase theory), but not all the way up to the controller's position (to keep up the idea of strict islandhood).The basic idea is that the controllee is merged as (i) Phase Impenetrability Condition (PIC) (Chomsky 2000: 108) The domain of a head X of a phase XP is not accessible to operations outside XP; only X and its edge are accessible to such operations.
(ii) The domain of a head corresponds to its c-command domain.
(iii) The edge of a head X is the residue outside X′; it comprises specifiers and elements adjoined to XP.
(iv) CPs and vPs are phases. 2This might not be the case to the same extent in all control scenarios since not all clause boundaries necessarily coincide with phase boundaries (cf., for instance, restructuring; see Wurmbrand 2001 a.o.); however, it is true for many control configurations.
an empty, referentially defective argument (EA) which probes upwards to find a suitable goal (=the controller) that licenses EA under Agree.As the search domain of EA is restricted by the PIC, EA has to move from phase edge to phase edge until it can be licensed; however, in contrast to the predictions of the MTC, it can stop moving when it is at the edge of the phase preceding the phase in which the controller enters the derivation.
Hence, licensing of control and the availability of movement are dissociated -this means that even if EA is inside an island, it can be licensed by a controller as long as EA is at the edge of the island and thus still accessible when the controller enters the derivation. 3So we can conclude that the HTC takes control dependencies to be local in a moderate way, meaning that while the HTC follows phase theory, it at the same time accepts that control dependencies might be disrupted by island boundaries.
Although the HTC contrasts with both the MTC and non-movement theories of control, the focus of this paper is on comparison with the MTC.Apart from the fact that PRO-based theories comprise a less homogeneous class of analyses, one basic insight of the paper is that control into non-adjoined islands is problematic for the MTC in particular.
The paper is organized as follows: section 2 starts with a brief overview of control into islands and presents the data involving obligatory control into non-adjoined islands.Section 3 reviews how the MTC accounts for OC into adjuncts and discusses why this strategy cannot be applied successfully in the case at hand.Moreover, the section reconsiders the Icelandic adjunct OC data involving extraposition discussed by Wood (2012;2014), which are also problematic for the MTC, and a similar point is made for German.Section 4 then returns to the locality issue; it focuses on the compatibility of phase theory and previous analyses of control, before section 5 then introduces an alternative account, the hybrid theory of control (HTC).First, the underlying mechanism will be discussed, before concrete examples and further consequences of the proposal are illuminated.Section 6 then returns to control into islands and shows how this is captured under the HTC, including control into non-adjoined islands.Crucially, we will see that, while this is barred by the MTC, control relations of this sort can easily be derived by the HTC.In section 7, nonobligatory control is discussed, and section 8 contains some concluding remarks on the distribution of the underlying empty argument.Section 9 concludes with a brief summary.

The general picture
If we consider control into islands, the picture is bipartite.On the one hand, there are well-known examples of obligatory control into certain adjuncts that have been discussed in the literature before.These comprise sentences like (1) and (2) (the latter being an Icelandic example involving extraposition). 41) Hornstein (1999: 88) John 1 heard Mary 2 [without PRO 1/*2 entering the room].
(2) Icelandic (Wood 2012: 323) Þeir 1 ákváðu (það) [að PRO 1 heimsaekja Ólaf].they.m.nom decided it.accto visit Olaf.acc 'They decided to visit Olaf.'In view of these example, the question might arise of whether control is sensitive to syntactic islands at all.However, crucially, obligatory control relations cannot be established across all kinds of syntactic islands.Two cases in point are subject islands (see (3)) and speech act or sentence adverbial adjuncts (see (4), ( 5)), which involve non-obligatory control. 5The grammaticality of sentence (3) shows that the matrix subject does not obligatorily control PRO; if this were the case, Binding Principle B would be violated (due to the coindexed pronoun in the infinitival clause) and the sentence would be expected to be ungrammatical.Similarly, in (4) and (5), PRO is not controlled by a local, c-commanding antecedent; instead, we observe instances of arbitrary or speaker PRO.
(3) This shows that control is not completely insensitive to islands, which means that it is subject to syntactic locality restrictions.Turning to the MTC, which derives OC via movement, it is a priori unexpected that we observe OC into islands at all, since extraction out of islands is generally blocked.In order to derive cases like (1), Hornstein (1999 et seq.) therefore proposes an analysis based on sideward movement, which does not involve movement out of the adjunct.As Wood (2012) already points out, this analysis cannot readily be extended to the extraposition data in (2) (see also section 3.2).Crucially, however, these data are not the only problematic island data for the MTC, because there is a third type of construction that displays OC into islands which the MTC cannot account for: OC into non-adjoined islands.

OC into non-adjoined islands
In this section, the focus will be on data from German.The general scenario we want to look at involves infinitival clauses located inside an island in the verb's complement position.Typically, these infinitival clauses can be extraposed in German (and sometimes have to be), but, crucially, extraposition is by no means always obligatory, as the broad range of examples shows. 6,7The crucial point is that we do find grammatical non-extraposed exam-ples (cf.also Kiss 2005: fn.6), and these have to be derived.Note, moreover, that, due to the underlying OV-structure in German, it is easy to see whether extraposition has taken place or not because of the different resulting surface word order: if the infinitival CP is extraposed, the participle (being located in V) occurs between the sentential pronoun or the embedding DP and the right-adjoined CP; otherwise, it occurs sentence-finally.In the following examples, the latter can be observed (i.e. the participle is the last word), which means that they do not involve extraposition and clearly are of the type control into a non-adjoined island.(6) involves a sentential proform plus related CP, and (7)/(8) are standard complex NP constraint configurations.The corresponding (b)-and (c)-examples confirm the islandhood of the underlying constituent γP, since both wh-movement and topicalization are blocked.Hornstein 1999;Landau 2000;2013;Sichel 2010).First, OC PRO requires a local, c-commanding antecedent; i.e., the controller must be an argument of the embedding clause and long-distance or arbitrary control are ruled out.This is illustrated in ( 9) and ( 10), in which non-c-commanding or non-local controllers are illicit.Moreover,in (11), Binding Theory helps to make this point even clearer.Note that the German verb blamieren can be used in a reflexive or a transitive context: sich blamieren ('to make a fool of oneself') vs. jemanden blamieren ('to embarrass someone').The fact that, in (11-a), it is impossible to use a pronoun that is bound by the matrix subject in the infinitival clause suggests that Principle B is violated, which means that PRO must be controlled by Hans (i.e., we have OC).8In (11-b), on the other hand, the pronoun is replaced with the simple anaphor sich in the infinitival clause.Since this is grammatical, we can conclude that sich is locally bound (following Principle A), which suggests that PRO is coreferent with the matrix subject Hans and can thus license the anaphor.9

OC into islands and the MTC
Recall that, at first sight, OC into adjuncts (as in (14-a)) seems to be problematic for the MTC, since its central assumption is that control involves movement of the controllee to the controller's position; however, extraction out of adjuncts is typically barred (see (14-b)).The MTC's answer to this dilemma is sideward movement.The underlying idea is that the controller DP (John in (14-a)) is moved out of the adjunct into the matrix clause while adjunct and matrix clause are still unconnected -that is, according to the MTC, control into adjuncts relies on an interarboreal operation before adjunction takes place.
The concrete derivation of (14-a) thus proceeds as follows: first As shown above, the MTC can account for data like these by invoking sideward movement: before XP is merged with a projection of α, the DP occupying the controllee position inside XP can move sideways to the controller position and thereby establish the control relation. 12his means that sideward movement precedes adjunction (and is therefore licit).
The scenario which concerns adjuncts created by movement is illustrated in (18).
(18) scenario 2: XP adjunct created by movement (internal merge) 10 See Huang (1982), who unifies the Subject Condition (according to which extraction out of subjects is not possible) and the Adjunct Condition (which claims that extraction out of adjuncts is not possible) under the Condition on Extraction Domain (CED).
11 The labeling in the trees follows Bare Phrase Structure, according to which the labels are defined on the basis of their relative positions.The different projections can thus be defined as follows: α is a maximal projection iff α does not project; α is a minimal projection iff α is directly selected from the numeration; α is an intermediate projection iff α is neither a maximal nor a minimal projection.As a result, the mother node of the controller is α ′ , not αP.
12 Landau (2003) and others have questioned whether sideward movement should be allowed or not (see Landau's 2003 objections against it), but this debate will be ignored here.What is crucial for us is that the MTC has a means to derive the scenario illustrated in (17).9 11 As shown above, the MTC can account for data like these by invoking sideward movement: before XP is merged with a projection of α, the DP occupying the controllee position inside XP can move sideways to the controller position and thereby establish the control relation. 12This means that sideward movement precedes adjunction (and is therefore licit).
The scenario which concerns adjuncts created by movement is illustrated in (18).
(18) scenario 2: XP adjunct created by movement (internal merge) The crucial difference between (17) and ( 18) is that in the latter case concatenation of XP with the main derivation takes place much earlier, namely inside γP.Hence, sideward movement is blocked: at the point in the derivation when XP (and with it the controllee inside it) is still unconnected to the rest (i.e.before γP is completed), the target position of the required movement (= Specα, the controller position) is not yet part of the derivation, given a bottomup derivation.In addition, movement out of XP at the point represented in (18) (i.e. after adjunction has taken place) yields a classical CED effect.So neither ordering, first sideward movement and then adjunction to αP, nor first adjunction to αP and then movement from the controllee to the controller position, can derive (18).There is, however, a third possibility to allow for movement from the controllee to the controller position in (18): if movement takes place after XP is merged into the derivation but before it is adjoined to αP (in (18), this would correspond to movement out of γP to Specα, before XP moves to the adjoined position).Note, however, that this option is only available if γP itself is not an island for extraction (nor is any other YP an island that is dominated by γP and dominates (the base position of) XP).
This leads to a third scenario involving control into islands: the one illustrated in (19).It does not involve control into an adjunct but rather control into a non-adjoined island.For the MTC, such a structure cannot exist (recall that we are talking about obligatory control): it is impossible to extract the controllee out of γP before the latter is merged into the structure and becomes an island (i.e.sideward movement, in analogy to control into adjuncts, is not an option), because the landing site is not yet available at this point; but once it is concatenated, extraction out of γP is blocked because of γP's island status; see (19).( 19) scenario 3: control into non-adjoined islands αP The crucial difference between (17) and ( 18) is that in the latter case concatenation of XP with the main derivation takes place much earlier, namely inside γP.Hence, sideward movement is blocked: at the point in the derivation when XP (and with it the controllee inside it) is still unconnected to the rest (i.e.before γP is completed), the target position of the required movement (=Specα, the controller position) is not yet part of the derivation, given a bottom-up derivation.In addition, movement out of XP at the point represented in (18) (i.e. after adjunction has taken place) yields a classical CED effect.So neither ordering, first sideward movement and then adjunction to αP, nor first adjunction to αP and then movement from the controllee to the controller position, can derive (18).There is, however, a third possibility to allow for movement from the controllee to the controller position in (18): if movement takes place after XP is merged into the derivation but before it is adjoined to αP (in (18), this would correspond to movement out of γP to Specα, before XP moves to the adjoined position).Note, however, that this option is only available if γP itself is not an island for extraction (nor is any other YP an island that is dominated by γP and dominates (the base position of) XP).
This leads to a third scenario involving control into islands: the one illustrated in (19).It does not involve control into an adjunct but rather control into a non-adjoined island.For the MTC, such a structure cannot exist (recall that we are talking about obligatory control): it is impossible to extract the controllee out of γP before the latter is merged into the structure and becomes an island (i.e.sideward movement, in analogy to control into adjuncts, is not an option), because the landing site is not yet available at this point; but once it is concatenated, extraction out of γP is blocked because of γP's island status; see ( 19). ( 19) scenario 3: control into non-adjoined islands controller position in ( 18): if movement takes place after XP is merged into the derivation but before it is adjoined to αP (in ( 18), this would correspond to movement out of γP to Specα, before XP moves to the adjoined position).Note, however, that this option is only available if γP itself is not an island for extraction (nor is any other YP an island that is dominated by γP and dominates (the base position of) XP).This leads to a third scenario involving control into islands: the one illustrated in (19).It does not involve control into an adjunct but rather control into a non-adjoined island.For the MTC, such a structure cannot exist (recall that we are talking about obligatory control): it is impossible to extract the controllee out of γP before the latter is merged into the structure and becomes an island (i.e.sideward movement, in analogy to control into adjuncts, is not an option), because the landing site is not yet available at this point; but once it is concatenated, extraction out of γP is blocked because of γP's island status; see ( 19). ( 19) scenario 3: control into non-adjoined islands To sum up, the MTC makes the following predictions: (i) OC into an extraposed adjunct (=scenario 2) can only be derived if the underlying constituent γP is not an island; (ii) OC into non-adjoined islands (=scenario 3) is predicted to be impossible.However, control scenarios of the latter type do exist -this has been shown explicitly in section 2. And this means that they fall outside the purview of what the MTC can handle.

Extraposition in Icelandic and German
But as mentioned above, even control into extraposed clauses is problematic for the MTC if the underlying constituent from which extraposition takes place is an island.This has already been reported for Icelandic in Wood (2012;2014), and in this section, I will moreover have a brief look at object extraposition in German.
The set of data from Icelandic considered here is taken from Wood (2012), who made this observation: if we look at sentences with an extraposed infinitival clause, an asymmetry arises between movement and control.While obligatory control across the sentential pronoun það is grammatical (see ( 20)), this pronoun blocks movement of all sorts; i.e., both standard A′-and A-movement across það is impossible, as illustrated in (21) ( involving topicalization) and ( 22) (involving raising).( 20 Let us now take a look at similar data from the literature on German.We can observe that, in German, we also find sentential pronouns of the það-type; moreover, as has been noted before (cf., for instance, Webelhuth 1992;Müller 1995), they also occur optionally and block CP topicalization; see ( 23). ( 23) German (Webelhuth 1992: 101) a. Ich bereue (es), dass Maria wegfährt.I regret it that Maria goes away 'I regret that Maria is going away.' b.Dass Maria wegfährt, bereue ich (*es).
that Maria goes away regret I it 'I regret that Maria is going away.'While (23) involves only finite complement clauses, ( 24)-( 26) show that the pattern can also be extended to non-finite complement clauses and topicalization involving extraction out of the embedded CP: as in Icelandic, the latter is illicit (see (24-b)-(26-b)), whereas control across the intervening pronoun is not blocked (see (24-a)-(26-a)).In fact, the observed intervention effect does not only occur with topicalization; other instances of A′-movement are equally affected, as example ( 27) involving wh-movement shows. 14The same holds for A-movement: as in Icelandic (see ( 22)), the sentential pronoun is ruled out in raising constructions (cf. the ambiguous verb beginnen ('begin'), which occurs in a raising construction in ( 28) and in a control construction in ( 29)).( 27 To sum up, it has been shown that these extraposition data from German behave like Icelandic in neither allowing A-nor A′-movement out of the extraposed infinitive.This means that there is no point in the derivation at which extraction is possible, which suggests that the underlying constituent from which extraposition takes place must be an island for leftward movement. 15However, control into the extraposed infinitive is common, hence, these extraposition data also pose a problem to the MTC.

Locality and control
After having considered the empirical data that motivate the development of an alternative, let us now briefly return to the second issue of this paper: control and its compatibility with phase theory.In general, the typical control scheme looks as illustrated in (30): the controller is part of the matrix clause and the controllee functions as the subject of an infinitival complement clause.
( 15 That rightward and leftward movement behave differently is well-known; while rightward movement seems to be less sensitive to island restrictions, it is, on the other hand, strictly clause-bounded (cf. the Right Roof Constraint), which is not true for leftward movement.I have nothing to say about why this is the case, but see, for instance, Müller (1995), who proposes an account of this asymmetry in terms of improper movement.
From a phase-theoretic point of view, this is problematic, because it implies that controller and controllee are separated from each other by a phase boundary (the embedded CP) and thus occur in different phases (recall the definitions from footnote 1).Since the controllee does not occur at the edge of the lower CP but rather in SpecT, the canonical subject position, it is no longer accessible when the controller enters the derivation in the next higher phase (see the illustration in (31), in which crossed out material represents those parts of the derivation that have already become inaccessible).So we can conclude that, following the standard view, control involves a dependency that is not readily compatible with phase theory. (

Phase theory and non-movement theories of control
Although focus here is on the comparison with the MTC, let us briefly take a look at traditional PRO-based theories of control.Here, the locality problem is relatively obvious.Without further assumptions, the distance between controller and controllee (=PRO) is too large to be compatible with phase theory, see (31).
Of course, considerations like these were not an issue when the first PRO-based theories were proposed in the 1980s.However, since the development of phase theory, little attention has been devoted to its compatibility with control theory.In fact, two PRO-based theories which have adopted the phase model at an underlying level are those by Landau (2000;2015) (although locality considerations also do not play a central role there).The theory proposed in Landau (2000 et seq.) involves an Agree relation between a functional head in the matrix clause and PRO in the embedded SpecT position, for which he originally proposed a relaxation of the PIC in order to make it work (this was in connection with his analysis of exhaustive control); see Landau (2000: 69;2004: 843).In Landau (2015), he adopted an alternative view: he commented on the Agree model and suggested that "OC complements […] are weak phases" (Landau 2015: 12).As a consequence, PRO is still accessible inside the infinitival complement clause even if it is not at its edge.
However, both views involve a relativization of the underlying locality restrictions (which means that the representational residue is enlarged, i.e. the amount of representation that must be kept in the workspace in between the spellout of two phases; see also footnote 1).So what this paper sets out to do is answer the following question: is it possible to model control in a strictly phase-based theory in which extra assumptions that ease locality restrictions are not needed?This is the conceptual motivation for sketching this alternative model, although Landau (2015) lists some more shortcomings of the Agreebased model from Landau (2000). 16n his two-tiered theory of control, Landau (2015) proposes movement of PRO to the edge of the embedded clause (=SpecFin in his model), which is thus compatible with the PIC (although this is not the driving force behind the assumed movement).The underlying idea is that exhaustive control (EC) is derived via a predication relation: PRO, which is assumed to be a minimal pronoun with the features {D, uφ}, moves to SpecFin (=the edge of the infinitival clause), since the predicative head Fin is looking for a nominal operator (i.e., [uD] on Fin attracts PRO).As a result, PRO-movement yields an open predicate, which is applied to the controlling DP and thereby saturated.
So in contrast to Landau (2000 et seq.),Landau (2015) also assumes movement of PRO to the phase edge, an assumption I will also argue for below.One major difference between his model and the basic assumptions of the approach developed here concerns the different impact of partial control (PC). 17While Landau takes the EC-PC split to be the source of a distinct syntactic treatment, I follow Pearson (2013;2016), Pitteroff et al. (2017a), a.o., in assuming that partial control readings are construed in the semantics and do not give rise to a different implementation in the syntactic component. 18 On the other hand, the central data that lead to the postulation of the hybrid theory of control (OC into islands) is not discussed in Landau (2015); so the foci of the two approaches differ, and similarities in the technical implementation have developed independently. 19A more thorough discussion of PC would go beyond the scope of this paper, but the interested reader is referred to Pitteroff et al. (2017a) as regards arguments for a semantic treatment of PC.The focus here will be control into islands and the locality of control dependencies.Therefore we now turn to the MTC and its compatibility with phase theory.

Phase theory and the MTC
Since the basic assumption of the MTC is that controller and controllee are related via movement, this puts the locality question in a different light.If the MTC is right, the control relation does not have to be established once the controller is merged into the derivation -instead, the only thing that must be guaranteed is that movement is an available option.In other words, the potential problem is not that the distance between controller and controllee might be too large to ultimately license control; that control involves a non-local dependency instead implies that movement has to be split into available movement steps, and the potential danger is that the designated controller position might not be accessible via movement.To stick with example (30-a) (John tries to win), consider the structures in (32).(32-a) displays the basic idea that the embedded subject is a copy of John, left behind by movement.Since the underlying idea of phase theory is that material that has been rendered inaccessible cannot be moved anymore and that only material in the previous phase head or phase edge is in the accessible domain (see footnote 1), the 17 If controller and controllee are referentially identical, the result is exhaustive control; see (i-a).In examples like (i-b), by contrast, the controller is just a proper subpart of the set of people denoted by the controllee -this is partial control.

Exhaustive control (EC)
John 1 tries PRO 1 to win. b.
Partial control (PC) (Landau 2000: 5) The chair 1 preferred PRO 1+ to gather at 6.Note that, in the recent literature, the terms PC/EC predicate have typically been replaced by the terms attitudinal vs. non-attitudinal predicate, which also suggests that this split among predicates is not exclusively responsible for a PC/EC reading, but has other impacts as well.In any case, the fact that I follow a semantic treatment of PC does not mean that I dismiss the observed split between attitudinal vs. non-attitudinal predicates; see also section 5.4 and 7.2. 19Note that an early version of the HTC has already been presented in 2011 at the University of Tübingen.A movement-based approach to control is therefore compatible in principle with phase theory as long as it is assumed that the non-local movement relation between controller and controllee is broken up into smaller movement steps which proceed from phase edge to phase edge.So far, so good; up to this point the MTC seems to be ideally suited to a local derivational approach to control.However, as The MTC accounts for this contrast in grammaticality as follows: (32-a) can be derived via sideward movement, as the adjunction site (adjunction to vP) is above the target of sideward movement (=Specv).When deriving (32-b), however, the target position of wh-movement is SpecC, i.e. the adjunction site (=adjunction to vP) is below this position.Therefore, the only available order of operations is (i) concatenating adjunct and main clause and then (ii) extracting who out of the adjunct.This follows automatically from the Extension Condition (see Chomsky 1993;1995), according to which syntactic operations have to extend the root.However, this implies at the same time that (32-b) violates the CED and is therefore ungrammatical.
Note that in order to derive the ungrammaticality of (32-b), it is absolutely essential that wh-movement targets a position above the adjunction site.As regards the site of adjunction, Drummond & Hornstein (2014) point out that "the relevant class of adjuncts must adjoin below C" (Drummond & Hornstein 2014: 451); so if wh-movement directly targets SpecC, the ungrammaticality of (32-b) follows.However, any intermediate landing site for wh-phrases below SpecC would undermine the finding of (32-b)'s ungrammaticality.Therefore, the assumption that wh-phrases move successive-cyclically via the phase edge Specv (as required by the PIC) cannot be maintained by the MTC: if the wh-phrase stopped in Specv (a position below the adjunction site), sideward movement could apply before adjunction takes place, and the ungrammatical sentence in (32-b) could ultimately be derived.Hence, Drummond & Hornstein's movement-based account of control into adjuncts is, after all, incompatible with a strict interpretation of phase theory.So if we want to explore the possibility of modeling control relations in a local way restricted by the PIC, we have to come up with an alternative.

Basic assumptions
In this section, we want to take a look at such an alternative approach, the so-called hybrid theory of control.It illustrates what a theory of control could look like that is compatible with both a strict interpretation of phase theory and all three control scenarios discussed in section 3.1.In this section, the basic underlying assumptions of the HTC will be introduced, followed by the data considered before.Basically, the idea is the following: the controllee has to move closer to the controller position to be able to establish the control relation in a local configuration (in accordance with the PIC); however, the controllee is not forced to move out of islands to license control into islands -just being at their edge suffices.
Technically, this is implemented as follows: the controllee is merged in the derivation as an empty argument (= EA) which is referentially defective.This is encoded in syntax in terms of the feature specification {D, β:_}.The β-feature can be viewed as a syntactically reified binding index feature, and that EA carries an unvalued β-feature indicates that EA needs to be referentially identified.This is achieved under Agree (which is assumed to involve upward probing) with another element bearing a valued β-feature.At the C-I interface, Agree involving β-feature checking is interpreted as binding. 20Overt DPs bear valued β-features, which means that they typically function as goal for EA and end up as binders or controllers of EA. 21Formally, we can adopt a version of Wurmbrand's definition of (Reverse) Agree (cf.Wurmbrand 2011: 3):22 (34) Agree: A feature [F:_ ] on α is valued by a feature [F: val] on γ iff (i) γ c-commands α, (ii) γ is the closest goal, and (iii) α is accessible to γ.
The derivation of obligatory control then proceeds as follows: the D-feature allows EA to be merged into an argument position; from here it probes upwards to find a goal/licensor (as to upward probing, cf. also Schäfer 2008;Hicks 2009;Zeijlstra 2012;Wurmbrand 2011;2013;Bjorkman & Zeijlstra 2014).If there is no potential antecedent present in the phase containing EA (as is the case in OC due to the non-local dependency), the need to establish an Agree configuration forces it to move to the phase edge, from which it probes further (see section 5.2 for details). 23When a DP is merged, EA finds a goal and can be licensed under Agree; i.e., the β-feature of EA is valued, which means that EA is interpreted as being bound by the controller. 24 Comparing the HTC to its predecessors, we can conclude that, as in the MTC, the controllee has to move to be licensed, the licensing conditions are not control-specific (i.e., no independent control module is needed), and non-obligatory control might involve last resort if no syntactic licensor can be found (see section 7 for details).As in PRO-based Fischer: Locality, control, and non-adjoined islands Art.82, page 17 of 40 theories, however, it is assumed that the controllee is an independent argument receiving its own θ-role (i.e., the Theta Criterion is not dispensed with).In addition, as in Landau (2000;2004), Agree is a basic licensing mechanism of control.The hybrid nature of the approach thus follows from the fact that the licensing of control involves first movement and then Agree.25

Successive-cyclic movement of EA
Before we turn to concrete examples that demonstrate how the HTC works in practice, let us briefly address the following questions: how is movement of EA triggered, why does it stop at the edge of islands, and is it ensured that it stops in time to prevent overgeneration?
The question of what triggers successive-cyclic movement is presumably as old as the idea that movement stops in intermediate positions.One possible implementation has been proposed by Chomsky (2000;2001;2008) in terms of EPP or edge features, an approach that has been taken up and advanced by Müller (2010;2011).Sticking to the underlying assumption that all instances of movement are feature-driven, one way to trigger intermediate movement steps is the insertion of so-called edge features on the head of the target phase.The insertion of these features is regulated by the so-called Edge Feature Condition.26(35) Edge Feature Condition -Chomsky's version (Müller 2010: 37, based on Chomsky 2000: 109;2001: 34;2008: 149) The head X of phase XP may be assigned an edge feature after the phase XP is otherwise complete, but only if the assignment has an effect on outcome.
As Müller (2010: 37) explains, "[g]iven [( 35)], phase heads can be assigned additional (i.e., noninherent) edge features in the course of the derivation "if the assignment has an effect on outcome" -that is, if it serves to implement intermediate movement steps required by the PIC".In the case at hand, this means the following: if the β-feature of EA cannot be valued within the current phase, the only way to maintain the possibility of valuing it later in the derivation involves movement of EA to the phase edge to ensure that it remains accessible.Hence, edge feature insertion takes place and triggers movement of EA to the edge of the current phase. 27 Successive-cyclic movement of EA either stops if a potential goal enters the derivation or if EA occurs at the edge of an island and any further step involved movement out of this island.Let us have a closer look at the second scenario.What is crucial for the HTC is that the availability of Agree and movement are dissociated since extraction out of islands must be prohibited, whereas licensing of control into (certain) islands under Agree must be possible.Therefore it is important to keep in mind that although accessibility (i.e.being at the very least at the phase edge of the previous phase) is a precondition for both Agree and movement, this is not yet a sufficient condition for the latter.In the case of islands, movement within the island (i.e. in particular to the edge of the highest phase contained in it) is not restricted; it is movement beyond which is forbidden.
In fact, for the HTC it is not really important what exactly this extra requirement for movement is which finally blocks extraction out of islands, but, to be concrete, let us follow Müller's (2010;2011) proposal.What he suggests is that the Edge Feature Condition be slightly modified in such a way that edge feature insertion is only possible if the phase head in the targeted phase is still active, which means that it must still have some structure-building or probe feature on it. 28In the context of islands, however, the typical configuration is this: the current phase head has already been rendered inactive when we try to extract something out of the island, because the insertion of islands is typically the last operation that takes place in a phase, and, as a result, all features on the phase head have already been used up.Thus, no edge feature can be assigned to the phase head anymore, which means that nothing can move out of the island to the current phase edge.
But what about direct movement out of the island into a higher phase (i.e.without intermediate stop at the current phase's edge)?This is also ruled out since material inside the island is no longer accessible after the completion of the current phase -so we would end up with a violation of the PIC.Hence, movement out of islands is predicted to be ungrammatical (even if the element we try to extract is located at the island's edge), while Agree into an island is possible (of course only as long as its edge is still accessible); we will return to the latter scenario in section 6.3. 29 Since successive-cyclic movement of EA is apparently only stopped if the latter is trapped in an island or Agree can be established, the question arises of whether the theory overgenerates in the following sense: if there is no island involved, we expect EA to be able to move successive-cyclically from phase edge to phase edge until it finally finds a goal.In fact, examples involving long EA-movement are rare, because the standard scenario is this: the embedded clause containing EA is an internal argument of the control verb (unless the clause containing EA is a subject clause, which is an island and thus stops EA-movement anyway); 30 this implies that not only does the next higher clause host the control verb, but also the latter's external (and potentially second internal) argument, i.e.DPs which stop EA-movement because they can function as goal.The only structure one can try to come up with is a combination of control and raising (since raising predicates do not have an external argument that could function as goal for EA), see (36).(36) John 28 That is, edge feature insertion must apply before XP is complete, in contrast to the original definition in (35) (see Müller 2010: 37). 29That subjects and adjuncts are islands for extraction is an old observation (see, for instance, Huang 1982), and one question that has been discussed ever since is what these two types of constituents have in common that distinguishes them from complements.One difference that can be observed is that the insertion of a subject as well as that of vP adjuncts happens at the very end of constructing the vP phase.In Müller's approach, last-merged specifiers in a phase turn out to be islands for extraction, and this is exactly how subjects and adjuncts can be characterized: external merge of the subject in Specv is typically the last operation (in the relevant sense, see below) that takes place in the vP phase (only vP adjuncts might additionally be inserted afterwards).As far as adjuncts are concerned, I deviate from Müller's theory in that I do not assume that they are last-merged specifiers of further functional projections (as proposed, for instance, in Cinque 1999).Alternatively, I assume that adjuncts are not subcategorized by any head, and therefore the insertion of a vP adjunct does not involve a subcategorization feature on v; instead, I assume that vP adjuncts can be inserted freely as a last operation in the vP phase.As a result, vP adjuncts can be inserted when the subject has already been merged into the derivation and the phase head v is already bare of any features.Insertion of the subject thus takes place prior to vP adjunction and is therefore, in fact, the penultimate operation in the phase; however, it is the operation that renders the phase head inactive and thereby predicts extraction out of the subject to be impossible. 30See section 7 as regards examples involving subject clauses.(36) is an example in which EA is base-generated in the most deeply embedded clause, raises (via SpecT) into the medial clause, moves to the phase edge SpecC, and finally finds a goal in the matrix clause.Hence, the HTC predicts that the sentence should be grammatical, which it is.Note, moreover, that (36) illustrates again the locality problem of the standard PRO-based theories alluded to in the beginning -if we do not assume EA-movement to SpecC, the controllee is no longer accessible when the controller John enters the derivation. 31,32 However, what about examples like ( 37) or ( 38), which are ungrammatical?Since no island restricts movement, it should be possible for EA to move to SpecC, from where it could probe upwards and agree with John in the matrix clause.(37)  (intended: 'John hopes that he seems to be smart.')What distinguishes the grammatical example of long EA-movement in (36) from the ungrammatical ones in ( 37) and ( 38) is the occurrence of an expletive in the latter case.Why should an expletive block licensing of EA by the matrix subject?A potential explanation would be to assume that the expletive is, at first sight, a potential goal for EA.This would imply that expletives bear valued β-features and can thus agree with EA, which would prevent EA from moving beyond the expletive (note that it was suggested above that overt nominals generally bear valued β-features).But recall what β-features are about: checking relations involving β-features are interpreted as binding; i.e., a val-31 However, note also that movement of EA through the intermediate SpecT positions is required independently by the EPP, which shows once more that it is hardly possible to come up with an example that involves long EA-movement independently. 32In fact, one reviewer suggests that the information from the controllee might alternatively be transmitted via every phase head on the path to the controller, or via null categories in the specifier of every phase.The question (s)he raises is whether this could do away with the locality problem.The latter scenario (with null categories in every phase edge on the path) seems to me similar to the assumption that EA itself moves to these specifiers; however, it would mean that we would have to posit additional empty elements in these positions.If the phase heads on the path to the controller are taken as mediating elements, we would have to come up with some motivation and technical implementation for that; in any case, the outcome would not be different on the assumption that this kind of feature transmission is restricted by the PIC just like movement (for instance, we would expect it not to be blocked in the case of non-adjoined islands, but in the case of subject islands, where EA would no longer be in the accessible domain when the next phase head entered the derivation; see (i)). (i) successive-cyclically from phase edge to phase edge until it finally finds a goal.In fact, examples involving long EA-movement are rare, because the standard scenario is this: the embedded clause containing EA is an internal argument of the control verb (unless the clause containing EA is a subject clause, which is an island and thus stops EA-movement anyway); 30 this implies that not only does the next higher clause host the control verb, but also the latter's external (and potentially second internal) argument(s), i.e.DPs which stop EA-movement because they can function as goal.The only structure one can try to come up with is a combination of control and raising (since raising predicates do not have an external argument that could function as goal for EA), see (36).
( 36) (36) is an example in which EA is base-generated in the most deeply embedded clause, raises (via SpecT) into the medial clause, moves to the phase edge SpecC, and finally finds a goal in the matrix clause.Hence, the HTC predicts that the sentence should be grammatical, which it is.Note, moreover, that (36) illustrates again the locality problem of the standard PRObased theories alluded to in the beginning -if we do not assume EA-movement to SpecC, the controllee is no longer accessible when the controller John enters the derivation. 31,32 However, what about examples like ( 37) or ( 38), which are ungrammatical?Since no island restricts movement, it should be possible for EA to move to SpecC, from where it could probe 30 See section 7 as regards examples involving subject clauses. 31However, note also that movement of EA through the intermediate SpecT positions is required independently by the EPP, which shows once more that it is hardly possible to come up with an example that involves long EA-movement independently.
32 In fact, one reviewer suggests that the information from the controllee might alternatively be transmitted via every phase head on the path to the controller, or via null categories in the specifier of every phase.The question (s)he raises is whether this could do away with the locality problem.The latter scenario (with null categories in every phase edge on the path) seems to me similar to the assumption that EA itself moves to these specifiers; however, it would mean that we would have to posit additional empty elements in these positions.If the phase heads on the path to the controller are taken as mediating elements, we would have to come up with some motivation and technical implementation for that; in any case, the outcome would not be different on the assumption that this kind of feature transmission is restricted by the PIC just like movement (for instance, we would expect it not to be blocked in the case of non-adjoined islands, but in the case of subject islands, where EA would no longer be in the accessible domain when the next phase head entered the derivation; see (i)).A third option might be to assume massive mediation by every node on the path from controllee to controller.This would imply that control relations involve uniform rather than punctuated movement paths in the sense of Abels ( 2003), which would mean that all nodes along the path are affected.As a result, EA itself would not need to move to the phase edge to remain accessible.However, one would also expect that EA's features inside a subject clause could be passed on via CP subj.and vP out of the vP phase.But this would predict that we get OC in the case of subject clauses, contrary to the facts (see section 7.2).But, of course, alternative technical implementations are certainly possible.

20
A third option might be to assume massive mediation by every node on the path from controllee to controller.This would imply that control relations involve uniform rather than punctuated movement paths in the sense of Abels ( 2003), which would mean that all nodes along the path are affected.As a result, EA itself would not need to move to the phase edge to remain accessible.However, one would also expect that EA's features inside a subject clause could be passed on via CP subj.and vP out of the vP phase.But this would predict that we get OC in the case of subject clauses, contrary to the facts (see section 7.2).But, of course, alternative technical implementations are certainly possible.
ued β-feature of an R-expression, for instance, can be viewed as encoding its reference.
Talking about expletives, they lack this semantics; so if they bear a β-feature, its value must be negatively defined, which could be represented as Ø-value. 33As a consequence, the Agree relation between the expletive and EA can syntactically take place, but it cannot be interpreted at the C-I interface and the derivation crashes.Moreover, I think that sentences like (37) (which involves the expletive there) face an additional problem since licensing of the existential construction also fails.Broadly speaking, EA would have to function as associate NP in the underlying expletive-associate relation, but since EA is defective at various levels, it cannot accomplish this task.In terms of Hazout's (2004;2008) theory of existential constructions, this can be explained as follows.According to Hazout, the relation between there and its associate is an underlying subject-predicate relation that gives rise to an agreement relation.Crucially, this agreement relation comes about as a result of percolating up the φ-features of the predicate nominal. 34However, EA is referentially defective; thus, it does not only bear an unvalued β-feature, but also unvalued φ-features, which renders it useless for the licensing of the existential construction.To sum up, as regards sentences involving expletive there, we can conclude that licensing fails in two directions: both there and EA try to get licensed by the other, but both attempts fail.

Subject control and the HTC
Although empirically, the main focus of the present paper is on control into (non-)adjoined islands, any control theory should be able to derive standard subject control configurations.Let us therefore briefly go back to our initial example (30-a), repeated in (39-a), to see how subject control is derived under the HTC.As for EA, it is assumed that this referentially defective argument is part of the lexicon, and inserting it into the numeration is in principle optional.However, if it is not inserted in (39-b), the derivation will crash later on because of a violation of the Theta Criterion.Hence, only the numeration in (39-b) can derive (39-a). 35  (39) a. John tries to win. b.Underlying numeration: Num = {{John, tries}, {EA, to, win}} The derivation then proceeds as follows.In Specv, EA is inserted as the external argument of win and is assigned the latter's external θ-role. 36Then it moves to the embedded SpecT position to check the EPP on T, 37 and finally to the edge of the embedded CP in order to remain accessible, as it still needs to value its β-feature; so the last step simply takes place in order to prevent the derivation from crashing (see ( 40)).
( Now the matrix clause is derived.After merging the matrix verb try, the matrix subject John enters the derivation in Specv and is assigned the external θ-role by the matrix predicate.Note that, due to its movement to the edge of CP, EA is still accessible when John is merged into the structure (John is then in Specv of the matrix clause and EA in SpecC, the edge of the preceding phase; see ( 41)).
(41) [ vP John [θ,β: val]  The derivation of ( 42) proceeds as follows: first, the embedded clause is built (which is identical to the derivation of the embedded clause in subject control structures).In Specv, EA is inserted as the external argument of surrender and is assigned the latter's external θ-role.Then it moves to the embedded SpecT position to check the EPP-feature on T, and finally to the edge of the embedded CP in order to remain accessible, as it still needs to value its β-feature; see ( 43).
( What is unexpected at first sight is that EA apparently does not choose Mary, the matrix object, as controller, although it is merged into the derivation prior to the matrix subject. However, this can be explained if we have a closer look at the type of verb that is used in the matrix clause.Typically, control verbs used in this context are attitude verbs, i.e. verbs that are used to "report on a mental state or a communicative act of some individual" (Pearson 2015: 1). 39Attitudinal contexts involve an attitude holder, "the bearer of the attitude or the agent of the reported speech act" (Pearson 2015: 1), and it has been proposed in the literature that this attitude holder is represented in the syntax. 40This means that in sentences like (45), a logophoric center is projected in the left periphery of the embedded clause, which introduces the attitude holder in a specifier position; as a result, the underlying structure is as shown in ( 46). ( 46) Crucially, this structure reveals that the closest binder for EA is not an argument in the matrix clause, but rather the attitude holder inside the embedded CP (which bears a valued β-feature since it is referentially specified).For (45), this means that EA's β-feature can be checked under Agree at the point in the derivation shown in (47). 4147) [ CP attitude holde [β:val] Cº [ TP EA [β:val] to Whether the attitude holder is the referent of the matrix subject is determined by semantic/pragmatic factors.I will defer a more thorough discussion to future research since a full-fledged analysis of these data is beyond the scope of this paper.

Control into islands and the HTC
Let us now return to the focus of this paper: the three different scenarios from section 3 involving control into adjuncts and non-adjoined islands.The following three subsections address each of these three scenarios and their analysis under the HTC. 39One test that can be applied to distinguish attitude from non-attitude verbs refers to the observation that attitude complements are tensed, whereas complements of non-attitude verbs are not (cf.Landau 2015).Hence, (i-a) (which contains the attitude verb promise) is felicitous, whereas (i-b) and (i-c) are not.
Yesterday, John promised Mary to call Anna tomorrow.b.
#Yesterday, John tried to call Anna tomorrow.c.
#Yesterday, John forced Bill to surrender tomorrow.

OC into adjuncts and the HTC
We start with control into adjuncts created by external merge.This is exemplified by example (48) (repeated from (16-a)).
Again, we are free to choose between numeration (49-a) and (49-b); however, in (49-a), the derivation will crash because it will inevitably violate the Theta Criterion, since there is no external argument for enter.
(49) a. Num 1 = {{John, heard, Mary}, {without, entering, the, room}} b.Num 2 = {{John, heard, Mary}, {EA, without, entering, the, room}} The adjunct is then derived as follows: EA is inserted in Specv, where it gets its θ-role from enter.Since its β-feature is still unvalued, it starts moving, first to SpecT, where it checks the EPP-feature on T, and then to SpecC, the edge of the phase and the edge of the adjunct.Now the adjunct is merged into the derivation, which is illustrated in (52). 43  (52) Now the adjunct is merged into the derivation, which is illustrated in (52). 43

John heard Mary
Here we have the following configuration: both John and the adjunct are at the edge of the vP phase, meaning that they c-command each other in the sense of the category-based definition of c-command by Kayne (1994), see (53).(54) Chomsky (1986: 7;9) a. X excludes Y if no segment of X dominates Y. b.X is dominated by Y only if it is dominated by every segment of Y.
For ( 52), this has the following effect: since only one segment of v ′ dominates John, the category v ′ (which all in all consists of two segments) does not dominate the latter.The first category dominating John is therefore vP, which also dominates the adjunct (consequently John ccommands the adjunct).Hence, John also c-commands EA at the adjunct's edge, which is still accessible at this point of the derivation. 45As a result, John can function as licensor of EA - Here we have the following configuration: both John and the adjunct are at the edge of the vP phase, meaning that they c-command each other in the sense of the category-based definition of c-command by Kayne (1994), see (53). 44 42 Details concerning the structure of gerunds will not be covered here. 43Recall that, following Bare Phrase Structure, John is dominated by v′ in (52); see also footnote 11. 44 In fact, category-based versions of c-command have often been proposed when licensing mechanisms under c-command involving adjoined structures have been investigated.It has been empirically important, for instance, in May's (1985) derivation of scopal relations after quantifier raising, or Kayne's (1994) approach to linearization based on the Linear Correspondence Axiom.Also from a theoretical point of view, category-based definitions have often been adopted in the literature when adjoined structures/multi-segment categories have been scrutinized; cf.Chomsky (1995: 338-340), or Sheehan (2013).I agree with Sheehan (2013: 15) in that "while it is true that category-based definitions of c-command appear complex when described verbally, they are more simple to represent graphically" -so ( 53) is not a complication of the notion of c-command but rather helps to clarify relationships in multi-segment structures.
(53) Category-based definition of c-command (Kayne 1994: 16; 18) X c-commands Y iff X and Y are categories and X excludes Y and every category dominating X dominates Y.
(54) Chomsky (1986: 7;9) a. X excludes Y if no segment of X dominates Y. b.X is dominated by Y only if it is dominated by every segment of Y.
For ( 52), this has the following effect: since only one segment of v′ dominates John, the category v′ (which all in all consists of two segments) does not dominate the latter.The first category dominating John is therefore vP, which also dominates the adjunct (consequently John c-commands the adjunct).Hence, John also EA at the adjunct's edge, which is still accessible at this point of the derivation. 45As a result, John can function as licensor of EA -it can value EA's β-feature and thereby establish the control relation. 46 45 Note that this is the last point in the derivation when EA is still accessible; when T merges with vP, EA is rendered inaccessible, which explains why EA cannot move out of the adjunct (=island) directly into a higher phase.On the other hand, intermediate movement to the edge of vP is not possible either, following Müller's (2010;2011) Edge Feature Condition, since v has already become inactive in (52), and therefore edge feature insertion is blocked, which would have to trigger this intermediate movement step (recall section 5.2). 46Note that the object, by contrast, is not in a position where it could license EA; i.e., obligatory object control into adjuncts is ruled out.This does not imply, however, that the object cannot bind variables inside the adjunct -LF movement to Specv can derive these readings; cf., for instance, John read every book 1 without reviewing it 1 (Hornstein 1999: 88).(However, at this point in the derivation, EA has already picked the subject as a goal.)Moreover, one reviewer has raised an interesting question: what about Object Shift in languages like Scandinavian?If the object raises to Specv, can it control EA inside a vP adjunct?In Norwegian, this is not the case (see (i)): as in English, we only get subject control even if the object has undergone Object Shift.

(i)
Norwegian (Inghild Flaate Høyem, p.c.) Jon 1 hørte henne 2 aldri/ikke [uten EA 1/*2 å gå inn i rommet].John 1 heard her 2 never/not without to go into room.the'John did not hear her/has never heard her without entering the room.'So how can it be excluded that the object is a potential goal for EA? First, it is far from clear whether Specv is the final landing site of Object Shift; following, for instance, Bošković (2012), Object Shift ultimately targets a position higher than vP.On this assumption, Specv is just an intermediate landing site (the object has to move via Specv, a phase edge, in order to satisfy the PIC).Following Müller's assumptions on successive-cyclic movement and edge feature insertion, this implies that the object targets a lower Specv position than the subject (see Müller 2010: 45).Although both, the object and the subject, now c-command EA inside the adjunct, the subject is the closer goal (since the path between probe and goal is shorter as fewer nodes intervene), and this is why we only get subject control; see (ii).
(ii) 6.2 Non-adjoined islands and the HTC Let us now turn to control into a non-adjoined island (i.e.scenario 3 from section 3, which could not be derived by the MTC since sideward movement cannot circumvent such an island).As an example, consider (55) (repeated from (8-a)).( 55 EA is inserted as the external argument of the embedded predicate, and since its β-feature is unvalued, it moves to the edge of CP.This suffices for the licensing of EA, since in the next phase (= vP) the controller is merged into the derivation (in Specv) and can license EA at the edge of the previous phase under Agree (see ( 56 So how can it be excluded that the object is a potential goal for EA? First, it is far from clear whether Specv is the final landing site of Object Shift; following, for instance, Bošković (2012), Object Shift ultimately targets a position higher than vP.On this assumption, Specv is just an intermediate landing site (the object has to move via Specv, a phase edge, in order to satisfy the PIC).Following Müller's assumptions on successive-cyclic movement and edge feature insertion, this implies that the object targets a lower Specv position than the subject (see Müller 2010: 45).Although both, the object and the subject, now c-command EA inside the adjunct, the subject is the closer goal (since the path between probe and goal is shorter as fewer nodes intervene), and this is why we only get subject control; see (ii).

Non-adjoined islands and the HTC
Let us now turn to control into a non-adjoined island (i.e.scenario 3 from section 3, which could not be derived by the MTC since sideward movement cannot circumvent such an island).As an example, consider (55) (repeated from (8-a)).
( EA is inserted as the external argument of the embedded predicate, and since its β-feature is unvalued, it moves to the edge CP.This suffices for the licensing of EA, since in the next phase (=vP) the controller is merged into the derivation (in Specv) and can license EA at the edge of the previous phase under Agree (see ( 56 But what about more deeply embedded infinitival clauses, i.e. scenarios in which more phases intervene between EA and the matrix subject?Consider the pair of sentences in (60).The verb inside the infinitival clause is chosen in such a way that it can be used reflexively (sich weiterbilden ('to educate oneself further')) or in a transitive way (jemanden weiterbilden ('to educate sb.further')).In this way, the underlying control relationships become more obvious, since the reflexive sich is only licensed if it is bound by EA (Binding Principle A) and a pronoun coindexed with EA is necessarily ruled out (Binding Principle B).
Note that the intervening DP das Angebot is not a potential licensor of EA since it does not c-command EA.Moreover, it can be excluded that there is an empty controller inside the island in form of a covert PP in the complement position of the noun: although Angebot ('offer') might take overt complements of this type (see ( 58)), this is not the case for other nouns that can occur inside such islands (see ( 59 But what about more deeply embedded infinitival clauses, i.e. scenarios in which more phases intervene between EA and the matrix subject?Consider the pair of sentences in (60).The verb inside the infinitival clause is chosen in such a way that it can be used reflexively (sich weiterbilden ('to educate oneself further')) or in a transitive way (jemanden weiterbilden ('to educate sb.further')).In this way, the underlying control relationships become more obvious, since the reflexive sich is only licensed if it is bound by EA (Binding Principle A) and a pronoun coindexed with EA is necessarily ruled out (Binding Principle B).
Let us first have a look at (60-a).It has the same structure as the examples discussed so far in this subsection: EA is inside a complex DP that functions as a direct object (das Angebot, sich/ihn weiterzubilden ('the offer to educate himself/him further')).In such a configuration, we can observe obligatory subject control.This is confirmed by the Principle B effect in (60-a), which indicates that EA must be bound by the matrix subject, as predicted by the HTC.In (60-b), the infinitival clause is more deeply embedded since the complex DP in which it is located is part of a relative clause that modifies the matrix object.As a result, two phase boundaries intervene between the subject trace (t 1 ) and EA.Following the HTC, we thus expect that subject control by Peter should be blocked.This prediction is indeed borne out as the binding effects in (60-b) show.The fact that the sentence is grammatical if the pronoun ihn is coindexed with the matrix subject Peter indicates that EA must bear another index; otherwise, the configuration would give rise to a Principle B violation because the pronoun would be bound in its binding domain.
To sum up, (60-a) shows that obligatory control holds if EA and the controller are only separated by one phase boundary; if more phase boundaries intervene, as in (60-b), obligatory control by the matrix subject is blocked.49And this is exactly what the HTC predicts.

Extrapositon and the HTC
Finally, let us take a look at the second scenario introduced in section 3, namely control into adjuncts created by movement (see (61), repeated from (18)).I suggest that this is also the underlying structure for control configurations involving extraposition.
(61) scenario 2: XP adjunct created by movement (internal merge) 6.3 Extrapositon and the HTC Finally, let us take a look at the second scenario introduced in section 3, namely control into adjuncts created by movement (see ( 61), repeated from ( 18)).I suggest that this is also the underlying structure for control configurations involving extraposition.
(61) scenario 2: XP adjunct created by movement (internal merge) In the literature, different ways to account for extraposition have been proposed.I will follow the movement-based approach (cf., among others, Bierwisch 1963;Reinhart 1980;Baltin 1982;Büring & Hartmann 1995;1997;Müller 1995;1997), which means that in examples like ( 62) and ( 63), the surface position of the extraposed CP is taken to be the result of rightward movement. 50 49 Note, however, that we get obligatory control by die Frau ('the woman'), which is expected, since there is only one phase boundary between the coindexed relative pronoun and EA.
50 Alternatively, two other strategies have been suggested in the literature: (i) the base-generation approach,

28
In the literature, different ways to account for extraposition have been proposed.I will follow the movement-based approach (cf., among others, Bierwisch 1963;Reinhart 1980;Baltin 1982;Büring & Hartmann 1995;1997;Müller 1995;1997) As already mentioned in section 2.2, extraposition of the infinitival clause is sometimes obligatory (as in ( 62)) and sometimes optional (as in ( 63)). 5150 Alternatively, two other strategies have been suggested in the literature: (i) the base-generation approach, according to which the extraposed XP is considered to be base-generated in its surface position (cf., for instance, Koster 1978;Culicover & Rochemont 1990;Webelhuth 1992;Haider 1997), and (ii) the PF approach, according to which extraposition is not a syntactic phenomenon; i.e., syntactically, the extraposed XP never occurs in the extraposed position.Instead, it is simply spelled out there (cf., for instance, Truckenbrodt 1995;Göbbel 2007).Moreover, combinations of the above-mentioned strategies have also been proposed, the underlying assumption being that there are different types of extraposition which should be analyzed differently.A common distinction that has been argued to be relevant is the argumentadjunct distinction (cf., for instance, Fox & Nissenbaum 1999;Kiss 2005;Inaba 2007;Hunter & Frank 2014).
In fact, the base-generation approach does not seem to be an option for data like (62), since it has been argued in the literature that the sentential pronoun and the embedded CP underlyingly form a constituent (see Thráinsson 1979;Wood 2014); if CP were base-generated in the extraposed position, it would not even be adjacent to the sentential pronoun at any point in the derivation.Note, however, that the base generation approach as such would be fully compatible with the HTC; in this case, control into the extraposed CP would boil down to control into an adjunct created by external merge, which has been discussed in detail in section 6.1. 51Recall that in the latter case, different additional factors have an impact on the naturalness of the sentence.
In (63) (which is basically the extraposed variant of (55) from section 6.2),I have therefore added the phrase mit den größeren Jungs ('with the older boys') to make it sound more natural.Note that the nonextraposed variant is equally grammatical: OC involving the attitude holder as controller; but since the attitude holder is not realized overtly, it gives the impression that there is no local, obligatory controller around (see also Fischer & Flaate Høyem 2017).In any case, this basically means that licensing of EA in examples like ( 70)-( 72) boils down to standard logophoric licensing (see also Zribi-Hertz 1989), which is not surprising since NOC PRO has long been shown to behave like a logophor (see, for instance, Kuno 1975;Landau 2013;2015). 59o if we compare the derivation of arbitrary control in (69) with that of example (71), which also involves a subject clause containing EA, the difference is simply that the latter involves an attitude holder which can license EA under Agree (see ( 73)).( 73) that licensing of EA in examples like (70)-( 72) boils down to standard logophoric licensing (see also Zribi-Hertz 1989), which is not surprising since NOC PRO has long been shown to behave like a logophor (see, for instance, Kuno 1975;Landau 2013;2015). 59 So if we compare the derivation of arbitrary control in (69) with that of example (71), which also involves a subject clause containing EA, the difference is simply that the latter involves an attitude holder which can license EA under Agree (see ( 73)).( 73) CP EA 1 firing the football coach T vP has t subj.clauseturned off ... Note, moreover, that theories that take NOC to be logophoric in nature make the following prediction: if a sentence is changed in such a way that the attitude holder changes, it is predicted that EA must be interpreted differently as well.This is borne out in the following example, which contrasts with (70) (Amy sagte, dass es Spaß machte, mit Dan zu tanzen/'Amy said that dancing with Dan had been fun').In (74), it is no longer Amy's point of view that is reported on; hence, EA cannot refer to Amy anymore.
( 8 Concluding remarks on distribution and realization of EA Crucially, the main goal of this paper was to develop a theory of control which is compatible with a certain conceptual perspective (cf.its PIC-oriented nature) capturing a specific set of empirical observations (particularly control into (non-)adjoined islands).Following Chomsky (1981), this means that we have to answer the following question concerning PRO/EA: "how is its reference determined?" (Chomsky 1981: 74).This is what falls under control theory.And this is what this paper focuses on.
Still, there are related issues which concern the context in which EA occurs.60This paper 59 As for discourse factors licensing logophoricity in general, see Kuno (1987);Fischer (2015); and others.As far as the relation between logophors and NOC is concerned, see also Sundaresan (2012); Nishigauchi (2014); Charnavel (2015).
60 Chomsky (1981: 74) formulates two further questions related to PRO: "where may it appear" and "where must it appear"?He concludes that they do not fall under control theory; instead, "the first question falls under general principles of the theories of government and binding, the second under the projection principle and Case theory" (which summarizes the answers GB-theory has provided).

34
Note, moreover, that theories that take NOC to be logophoric in nature make the following prediction: if a sentence is changed in such a way that the attitude holder changes, it is predicted that EA must be interpreted differently as well.This is borne out in the following example, which contrasts with (70) (Amy sagte, dass es Spaß machte, mit Dan zu tanzen/'Amy said that dancing with Dan had been fun').In (74), it is no longer Amy's point of view that is reported on; hence, EA cannot refer to Amy anymore.

(74)
German Amy 1 wurde erzählt, dass es Spaß machte, [EA *1/2 mit Dan zu tanzen].Amy was told that it fun made with Dan to dance 'Amy was told that dancing with Dan had been fun.'

Concluding remarks on distribution and realization of EA
Crucially, the main goal of this paper was to develop a theory of control which is compatible with a certain conceptual perspective (cf.its PIC-oriented nature) capturing a specific set of empirical observations (particularly control into (non-)adjoined islands).Following Chomsky (1981), this means that we have to answer the following question concerning PRO/EA: "how is its reference determined?" (Chomsky 1981: 74).This is what falls under control theory.And this is what this paper focuses on.
Still, there are related issues which concern the context in which EA occurs. 60This paper will not provide an ultimate answer to the question of where EA must or must not occur; however, this section will sketch some basic ideas.
Returning to the underlying assumption that EA can be inserted freely into the numeration, the following question arises.What ensures that EA surfaces in the subject position of infinitivals and not in another argument position?In particular, why is OC into finite clauses, as illustrated in (75), ruled out? (75) *John 1 said that EA 1 likes pizza.
In fact, what goes wrong in examples like these is not necessarily the licensing of EA; instead, the problem rather seems to be that finite T cannot be properly licensed if EA is the subject.It is standardly assumed that T bears φ-features that must be valued under Agree with the subject DP (see, for instance, Chomsky 2000;2001). 61However, in SpecT, EA itself is not yet licensed, which means that it is not yet referentially identified and can therefore not function as a goal for φ-Agree.But what is different in the case of non-finite T, where we typically find EA?Assume that finite T differs from non-finite T insofar as the latter does not bear φ-features.If this is the case, non-finite T does not depend on a subject that can value φ-features, and therefore referentially defective EA can occur in the subject position of non-finite T without causing any damage.
However, what if EA did not only correspond to PRO (hence also the more neutral term EA)? Could it also be the origin of other (empty) categories?After all, structures like (76) might not be completely out.The remaining section is devoted to some very tentative ideas concerning these questions.Of course, languages like English or German do not allow these structures.On closer inspection, though, it is not at all that clear why we would want to rule out (76) completely.What (76-a) actually displays is a sentence with a non-overt argument in the subject position of a finite clause.This is exactly what is found in pro-drop languages, and although it has been standard to consider PRO and pro two distinct empty categories since the 1980s (see Chomsky 1982;Rizzi 1986), the idea that PRO and pro might have the same origin is not new (see Chomsky 1981;Borer 1989;Huang 1989;Manzini 2009;Duguine 2015;McFadden & Sundaresan 2016).As an illustration, let us briefly consider the Italian example in (77).
(77) Italian Canto.sing.1sg'I sing.' The standard analysis (following Rizzi 1986) is to assume that the syntactic derivation involves pro as external argument (for a different view, see, for instance, Borer 1986;Alexiadou & Anagnostopoulou 1998).According to the underlying assumptions of the HTC, EA can be inserted into the numeration, and so it is easy to see that EA could take over the role of what is standardly called pro.Since without EA the Theta Criterion would be violated, a successful derivation must include EA.Under the assumption that PRO and pro are underlyingly the same element, 62 what distinguishes EA in (77) from EA in OC would then be not an inherent property of EA itself, but would have to be derived from other structural distinctions.Of course, much more needs to be said about the concrete differences that turn EA into a PRO-or pro-like element, but this must be left for future research.
Turning briefly to (76-b), the situation is the following: the object position is occupied by a non-overt argument which ends up being bound by the subject John.Again, this might be a scenario we do not want to abandon completely.Although they are typically phonologically realized, this configuration is reminiscent of that of anaphors.So, in the end, EA might even end up being the source of anaphors, independent restrictions forcing us to spell out EA phonologically in this context (cf.also Hornstein 2001, who extends his movement approach to anaphors as well).However, a more elaborate analysis in this direction lies definitely beyond the scope of this paper.
In any case, the central insight is that EA need not be control-specific (unlike PRO), and therefore it could well be the case that it also appears in other constructions.So EA might be considered the source of OC PRO and NOC PRO as well as of pro (and arguably even of anaphors).Thus, these elements might not be inherently different in the lexicon, but could simply emerge because of differences in the syntactic environment.But these are issues left for future research.

Conclusion
This paper set out to develop a theory of control that (i) is compatible with phase theory and (ii) can straightforwardly account for control into adjoined and non-adjoined islands, two aspects which have proved to be problematic for the MTC.
It has been assumed that the lexicon hosts an empty argument (EA), which is referentially defective and therefore bears an unvalued β-feature (which basically corresponds to a syntactically reified binding index feature).In control structures, EA is part of the derivation since otherwise the Theta Criterion would be violated.In the course of the syntactic derivation, EA probes upwards to find a goal which can value its β-feature under Agree -hence, EA moves from phase edge to phase edge until a potential goal is merged into the next higher phase and licenses EA by valuing EA's β-feature, which is interpreted as binding at the C-I interface.Typically, the licensor is the controlling DP.In NOC structures, two different scenarios can arise: in attitudinal contexts, the attitude holder (which is syntactically represented in the left periphery) is accessible to EA, agrees with it and thus determines EA's interpretation.In non-attitudinal contexts, there is no controller in the accessible domain and arbitrary control thus arises as the result of failure to agree.
Since EA is not control-specific, the theory can presumably also be extended to include pro (and arguably even anaphors).Depending on underlying structural distinctions, these forms might simply emerge as different realizations of EA.
If the HTC is on the right track, the answer to the locality question is this: control is more local than traditional PRO-based theories would have us believe, but less local than suggested by the MTC.The conceptual advantage of the HTC is that it allows us to take the PIC seriously; thus, the HTC can be considered to be part of a bigger program which aims at reanalyzing all kinds of syntactic phenomena in a local-derivational way.At the same time, it allows us to take islandhood seriously -after all, control into islands does not involve extraction out of them, but only movement to the edge of the highest phase within them.As a c onsequence, under the HTC it does not matter whether an island is adjoined or non-adjoined.theory of control, NOC = non-obligatory control, OC = obligatory control, PC = partial control, PIC = Phase Impenetrability Condition (i) the controlled clauses in (6-a)-(8-a) involve OC (see (9)-(13)); (ii) the infinitival clauses are embedded inside islands because they block extraction (see (6-b)/(6-c)-(8-b)/(8-c)); (iii) and, as the word order clearly shows, these examples do not involve extraposition but are genuine examples of the type OC into a non-adjoined island.
laughed at Mary [without <John> falling over].b. *Who did John laugh at Bill before Mary spoke to <who>]?
movement indicated in (32-a) is only compatible with the PIC if it proceeds via the edge of the CP phase, as indicated in (32-b).(32) a. John tries [ CP <John> to win].b. [ vP John tries [ CP <John> [ TP <John> to [ vP <John> win]]].
Drummond & Hornstein (2014) have pointed out, successive-cyclic movement from phase edge to phase edge undermines the MTC-based account of why control into adjuncts is possible in examples like (33-a), whereas wh-extraction out of adjuncts is not (see (33-b)).(33) Drummond & Hornstein (2014: 450) a. John laughed at Mary [without <John> falling over].(=(14-a)) b. *Who did John laugh at Bill [before Mary spoke to <who>]?(=(14-b)) 44 definition of c-command(Kayne 1994: 16;  18) X c-commands Y iff X and Y are categories and X excludes Y and every category dominating X dominates Y.
Mathis accepted gladly the offer to join the hockey game.'Note that these structures are clear instances of obligatory control.This is confirmed in the following by applying several OC diagnostics from the literature (see, for instance,

Mathis accepted gladly the offer to join the hockey game, and Lasse accepted gladly the offer to join the hockey game.' Finally, it has been observed that OC PRO need not be human (see Landau 2013 a.o.).
What is also well-known is that OC PRO only allows a sloppy interpretation under ellipsis; this is illustrated in (12).The meaning of (12-a) is represented in (12-b), and as the indexation shows, a sloppy reading of PRO is obligatory -hence, this must be obligatory control.
13 (24) German a. Er 1 hat (es) bereut/bedauert, PRO 1 Maria verletzt zu haben. he has it regretted Maria hurt to have 'He regretted having hurt Maria.' b. a. Er 1 bittet dich (darum), PRO 1 die Unterlagen morgen mitzubringen.he asks you for it the documents tomorrow with.to.bring 'He is asking you to bring the documents tomorrow.'b.Bierwisch (1963: 135) Die Unterlagen 2 bittet er 1 dich (*darum), PRO 1 t 2 morgen mitzubringen.the documents asks he you for it tomorrow with.to.bring 'He is asking you to bring the documents tomorrow.''Peter regretted having hurt Maria, and Hans did too.' b.Peter 1 hat (es) bereut [PRO 1 Maria verletzt zu haben], und Hans 2 hat (es) bereut Peter has it regretted Maria hurt to have and Hans has it regretted [PRO *1/2 Maria verletzt zu haben].Maria hurt to have → only sloppy reading available (=OC property) 30) a. John 1 tries [ CP [ TP PRO 1 to win]].b. [ matr.clause controller … [ emb.clause controllee … ]] Pitteroff et al. 2017a;;et seq.; 2015), whether we end up with PC or EC depends on the matrix predicate, which were thus called PC vs. EC predicates inLandau (2000). Byst, Boeckx, Hornstein & Nunes (2010),Sheehan (2014),Pitteroff et al. (2017a; b)and others have argued that the type of embedded predicate also plays a role; see the German example below, which allows a PC reading and involves an EC matrix predicate plus an embedded predicate which licenses a comitative (grammaticality judgements are based on an experimental investigation of PC in German; seePitteroff et al. 2017a; b).
1 hopes [ CP EA 1 [ TP t EA to seem [ TP t EA to [ vP t EA be smart]]]] *John 1 hopes [ CP EA 1 for [ TP there to seem [ TP t EA to [ vP t EA be smart]]] (intended: 'John hopes to seem to be smart.')(38) *John 1 hopes [ CP EA 1 that it seems [ TP t EA to [ vP t EA be smart]]].

5.4 Object control, promise-verbs, and the HTC So
tries [*θ*] [ VP t tries [ CP EA[β:val][ TP t′ EA to [ vP t EA win t win ]]]]] So EA's β-feature can finally be valued by the matrix subject under Agree, and EA is interpreted as being bound by John. 38ar, we have only considered examples in which the control verb takes two arguments, the infinitival clause and an external argument; the controller was therefore always the external argument, i.e. the subject of the matrix clause.In this section, we will briefly turn to control verbs that select in addition a second internal argument.Let us first turn to standard object control constructions and see how examples like (42) can be derived.
Bresnan 1982;Farkas 1988;he embedded CP and θPetter 1998;Stiebels 2007;Polinsky 2011;Landau 2013)n SpecV and is assigned the second internal θ-role of force.Since Bill and EA are now both accessible and the former c-commands EA, the control relation can be established -Bill can value EA's β-feature under Agree.As a result, we get object control; see (44).Bill[θ,β:val]forced [*θ*] [ CP EA [β:val] [ TP t′ EA to [ vP t EA surrender [ VP t surrender ]]]]]I do not have much to add to the discussion of promise-verbs or control shift (cf., for instance,Bresnan 1982;Farkas 1988; Sag & Pollard 1991;Petter 1998;Stiebels 2007;Polinsky 2011;Landau 2013); but for the sake of completeness, let us briefly have a look at examples like (45).
vP vP John heard [ VP t heard Mary]] Mathis accepted gladly the offer to join the hockey game.' )). 47(56) [ vP Mathis [β:val ] das Angebot, [ CP EA [β:val] [Hockey mitzuspielen]], angenommen] Mathis accepted gladly the offer to join the hockey game.' )).47Note that the intervening DP das Angebot is not a potential licensor of EA since it does not c-command EA.Moreover, it can be excluded that there is an empty controller inside the island in form of a covert PP in the complement position of the noun: although Angebot ('offer') might take overt complements of this type (see (58)), this is not the case for other nouns that can occur inside such islands (see (59)). )).

*1/2 sich *1/2 / ihn 1/*2
48 Peter accepted gladly the offer to undergo further training.'b.Peter 1 hat [ vP t 1 die Frau 2 , [ CP die 2 das Angebot, [ CP EA , which means that in examples like (62) and (63), the surface position of the extraposed CP is taken to be the result of rightward movement.50Mathisaccepted gladly the offer to play hockey with the older boys.'