Adjunct control in German, Norwegian, and English

This paper presents an overview of adjunct control in German, Norwegian, and English, comprising adverbial infinitives, adverbial present and past participle constructions, as well as adverbial small clauses headed by the particle als in German, som in Norwegian, and as in English. We show that the height of the adjunction site (and thus, following scope-based adjunct theories, the underlying semantics of the adjuncts) determines the control possibilities. Based on a large set of data, we argue that event- and process-modifying adjuncts, i.e. adjuncts adjoined in the verbal domain (at the vP- or VP-level, respectively), display obligatory control (OC) properties, whereas sentence and speech act adverbials, which are adjoined at the TP- and CP-level, respectively, rather involve non-obligatory control (NOC). To capture these data theoretically, we propose a structural account that assumes that OC relations are syntactically licensed under upward Agree with PRO being the referentially defective probe that needs to be referentially identified in the course of the derivation by the controller, i.e. the goal. If the adjunct is adjoined in the verbal domain, such an Agree relation can be successfully established and OC is derived. If adjunction occurs in higher adjunction sites, i.e. outside of the verbal domain, feature valuation under Agree fails; as a last resort strategy, the control relation is then licensed on the basis of pragmatic factors, which yields NOC.


Introduction
In the control debates of the recent years (cf. Hornstein 1999;Landau 2000Landau , 2013Landau , 2015among others), adjunct control has only played an ancillary role. There are at least three reasons for this. Firstly, adjunct control comprises a very heterogeneous set of examples, empirically speaking. Secondly, as Landau (2013Landau ( , 2015Landau ( , 2017 points out, adjunct control cannot be categorized in unison as obligatory or non-obligatory control (OC vs. NOC). Instead, the distinction between OC and NOC cuts through the set of examples involving adjunct control, an observation that adds to its heterogeneous character. Thirdly, NOC as such has typically played a minor part in theories of control.
The aim of this paper is two-fold. In the first part, we focus on the empirical situation and provide insight not only into English, but also German and Norwegian data involving adjunct control, something which is much needed since the available literature is to a large extent based on English data only. To this end, we investigate adjunct control into (i) various adverbial infinitives (cf. Faarlund et al. 1997;Landau 2013;Høyem 2015, (ii) adverbial present and past participle constructions (cf. Kortmann 1991;Fabricius-Hansen and Haug 2012;Brodahl 2016Brodahl , 2020Brodahl and Høyem 2018;Høyem and Brodahl 2019), and (iii) adverbial small clauses headed by the particle als in German, som in Norwegian, as in English (cf. Emonds 1985;Eide 1996;Flaate 2007;Eide and Flaate 2017). Apart from providing comparable data, we show that there is, after all, a common underlying pattern. We argue that it depends on the underlying syntactic structure whether we get OC with the subject or the object as controller, or NOC. To be precise, it depends on the height of the adjunction site (which, in turn, hinges on the semantics of the adjunct): it turns out that adjunction in the vP-domain results in obligatory subject control, adjunction in the VP-domain results in obligatory object control, and a higher adjunction site yields NOC (see also Høyem and Brodahl 2019).
In the second part of the paper, we focus on the theoretical side and show how the data can be captured in accordance with phase theory, drawing on ideas by Fischer (2018) and Fischer and Høyem (2021).
The paper is structured as follows: in Sect. 2, empirical evidence is presentedwhat kind of adverbials can we observe, where do they adjoin, and which type of control do they display? The control analysis advocated in this paper will be outlined in Sect. 3. First, some underlying technical issues concerning the analysis are discussed, before the theory is then tested on the different adjuncts and control possibilities from Sect. 2. Section 4 concludes the paper.

Empirical evidence
First, we concentrate on empirical evidence from German, Norwegian, and English, which includes adverbial infinitives, adverbial present and past participle constructions, as well as adverbial small clauses headed by the particle als in German, som in Norwegian, and as in English. 1 These adjuncts display a wide range of adverbial readings and appear in most adverbial positions. Although these three languages differ typologically, with English and Norwegian being SVO languages and German an SOV language, the ordering restrictions among these adverbial adjuncts adhere to the same syntactico-semantic adverbial hierarchy.

Background assumptions
In line with scope-based adjunct theories, like Frey andPittner (1998, 1999), Pittner (1999), Haider (2000), Ernst (2002Ernst ( , 2014, we assume that adverbial adjuncts are located in different syntactic domains, which correlate with their semantic scope. According to the scope-based approach, adverbials modifying the speech act or proposition (i.e., speech act, frame, and sentence adverbials) are attached high in the clause, as CP-and TP-adjuncts, whereas event-modifying (causal, temporal, instrumental, etc.) and process-modifying (manner) adverbials are adjoined lower down in the tree structure, as vP-and VP-adjuncts, respectively.
(1) Ernst (2014, 115) CP TP vP VP Frey and Pittner (1999) Frame Proposition Event Process Ernst (2002) Speech-Act Proposition Event Specified Event Interestingly, adverbial small clauses headed by als/som/as, adverbial infinitives as well as adverbial present and past participle constructions all appear to adhere to this syntactico-semantic hierarchy (cf. Eide 1996;Flaate 2007;Høyem 2015Brodahl 2016Brodahl , 2020Eide and Flaate 2017;Brodahl and Høyem 2018;Høyem and Brodahl 2019). These adjuncts are syntactically and semantically underspecified in the sense that they are not inherently temporal, causal or manner, for example, but are interpreted as such in these adjunct positions (see also Businger 2011 and Høyem 2019 for a similar conclusion regarding absolute small clauses headed by mit ('with')). 2 Regarding the interpretation of PRO in these adjuncts, we will show below that the low adverbial adjuncts typically display the OC properties described by Landau (2013) outlined in (2).
(2) OC properties: a. The controller must be an argument of the adjunct's matrix clause. b. Long-distance and arbitrary control are ruled out. 3 c. OC PRO only allows a sloppy reading under ellipsis. d. OC allows a non-human PRO.
This means that the control status of these adjuncts is affected by their syntactic position; that is, adjuncts in the c-command domain of T (= vP-and VP-adjuncts) generally display OC properties, and adjuncts adjoined above T (= TP-and CPadjuncts) display NOC properties. Drawing on Frey andPittner (1998, 1999) and Ernst (2014), we assume that (only) speech act, frame, and sentence adverbials have scope over T (anchoring the event/proposition temporally) and are thus attached to a projection of TP or CP. Event-modifying adjuncts, for instance temporal, causal, and instrumental adverbials, are adjoined to a projection of vP (semantically corresponding to the whole or parts of the event), whereas process-modifying adjuncts, like manner adverbials, are adjoined to the VP; see also the illustration in (3).
Regarding the different control possibilities in the OC domain, i.e. vP/VP-internally, there is syntactic evidence that subject-controlled adjuncts adjoin higher in the tree structure than object-controlled adjuncts. The first piece of syntactic evidence for different adjunction sites comes from pre-and post-verbal word ordering restrictions, in the literature known as mirror effects (cf. Barbiers 1995, 102-113;Å farli 1997, 99;Pittner 1999, 304-310; see also Bowers 1993, 605-612). As illustrated in the German data in (4), subject-controlled adjuncts precede objectcontrolled adjuncts in preverbal position: (4)  (Høyem 2018, 376) We take this as evidence that object-controlled adjuncts are closer to the verb and therefore are adjoined lower than subject-controlled adjuncts, implying that subjectcontrolled adjuncts are vP-adjuncts, whereas accusative and dative object-controlled adjuncts are VP-adjuncts; cf. also Nissenbaum (2005) 4 for the same observation in subject-controlled rationale clauses and object-controlled purpose clauses, see (6) and (7) (Nissenbaum 2005, 4) That subject-controlled adjuncts are structurally higher than object-controlled adjuncts is further corroborated by binding effects in the form of (lack of) Principle C effects in ditransitive structures. 5 As far as the underlying structure for ditransitives involving a dative argument is concerned, we assume that the latter is introduced by Appl°(following Anagnostopoulou 1999;Pylkännen 2002;McFadden 2004McFadden , 2006; see also Sect. 3.3.1 below); this means that the dative argument is situated in a phrase between VP and vP.
Let us now look at the data in (8) and (10). These sentences involve adverbial small clauses headed by als. Example (8) involves control by the accusative object (ein Auto 'a car'). In (8a), the scenario is as follows: the dative argument (Peter) is an R-expression which is co-indexed with the pronoun sein ('his') inside the adverbial. This constellation is unproblematic, since the dative argument is not c-commanded by anything inside the adjunct; so the sentence is expected to be grammatical (which it is). However, if we use a pronoun as dative argument (ihm 'him') and a co-indexed R-expression in the adverbial, the sentence becomes ungrammatical (see (8b)). This suggests that a Principle C effect arises, i.e. the dative argument c-commands into the adjunct, which is exactly what we expect if this adverbial is adjoined at the VP-layer (see tree (9) Frey andPittner 1998, 1999;Pittner 1999;Frey 2003;Haegeman 2012;Frey and Truckenbrodt 2015;Solstad 2016). However, it is not trivial to find suitable structures if control into these adverbials is involved: what we are typically interested in when we apply these tests is whether there is a binding relation between a certain DP in the matrix clause and a DP inside the adjunct. So we have to look for structures in which the potential binder outside the adjunct is not co-indexed with PRO (otherwise, we would end up considering only a binding relation within the adjunct). Therefore, we use examples which involve control by the subject or accusative object and use the dative object as potential binder in the binding test. Interestingly, we get a different result if we consider subject-controlled adjuncts. In (10), we also use the dative argument Peter/ihm ('him')) as a potential binder for the DP seine ('his')/ Peters ('Peter's') inside the adverbial. 7 In this case, however, no Principle C effect arises if the co-indexed R-expression occurs inside the adjunct, which suggests that the dative argument does not c-command the adjunct in the case of subject control.  As far as speech act adverbials are concerned, scope-based adjunct theories suggest that they are high adjuncts which occur in the CP-domain. As a result, we would not expect a Principle C effect to arise when the adjunct contains an R-expression which is co-referent with the subject of the matrix clause. As the example in (12) shows, this prediction is indeed borne out. So the data in this section provide further evidence for the assumptions (i) that subject-controlled adjuncts occur higher in the structure than object-controlled adjuncts; i.e., we deal with vP-and VP-adjuncts, respectively, (ii) and that speech act adverbials are adjoined even higher in the structure, namely in the CP-domain.
As the following data show, PRO in these adjuncts displays all OC properties described by Landau (2013) and summarized in (2) above.

OC property no. 1
According to Landau's first OC property, the controller must be an argument of the adjunct's matrix clause (see (2a)). As illustrated below, subject control is attested in adverbial infinitives (see (13)), in adverbial present and past participle constructions (see (14)), and in adverbial small clauses headed by als/som/as (see (15) [PRO i/*j som student] fikk Jon i alltid rabatt på fly. as student obtained Jon always discount on flights 'As a student, Jon always obtained a discount on flights.' (Eide 1998, 53) c. English [PRO i/*j As a blonde], Mary i might look like Jane. (Fabricius-Hansen and Haug 2012, 36) Although subject control is a very frequent control relation in these types of adjuncts, also (accusative or dative) objects in the adjunct's matrix clause may function as an antecedent for OC PRO, shown in (16) [PRO i/*j Having undergone the German academic education], the English university system impressed him i a great deal. (Kortmann 1991, 8) 2.2.2 OC property no. 2 As outlined in (2b), the second OC property described by Landau states that OC PRO is never arbitrary or long-distance, which is exactly what we find in these adjuncts. In the German example (17a), the matrix clause does not contain a potential local controller for PRO; as a result, the sentence is ungrammatical, which shows that arbitrary control is not an option either. The same seems to be the case in the equivalent Norwegian and English examples in (17b, c). Here, PRO can have neither an arbitrary reading (as PRO arb ) nor a specific reading (as PRO i ), even though this would be required pragmatically. In German and Norwegian, one would have to use the finite counterpart of these adjuncts with a lexical subject (see (18a, b)); in English, a gerund with a lexical subject can be used to express the intended interpretation, as in (18c) To sum up, (17) and (19) show that the non-availability of a suitable controller in the matrix clause leads to ungrammaticality; resort to arbitrary control or control by an otherwise discourse-salient referent is excluded. Below, we add some more examples to illustrate explicitly that LD controllers in higher embedding clauses are also ruled out and that these observations are not restricted to adverbial infinitives headed by ohne/uten/without (see (17), (19)) but generally hold for adverbial infinitives (see (20a), (21a), (22a)), adverbial participle constructions (see (20b), (21b), (22b)), and small clause adjuncts headed by als/som/as (see (20c), (21c), (22c)).
(20) German a. Peter i behauptete, dass sie j vorbeigingen, [PRO *i/*arb/j ohne Peter claimed that they by.passed without etwas zu bemerken]. anything to notice 'Peter claimed that they had passed by without noticing anything.' (Haider 2015, 4)  To conclude, the data in this section have revealed that only OC by a syntactically local controller (i.e., a controller in the adjunct's matrix clause) is possible and that both long-distance and arbitrary control are ruled out.

OC property no. 3
Landau's third OC property states that in VP-ellipsis, OC PRO can only get a sloppy reading, never a strict reading (see (2c)). As shown in the data below, this is the case in all three languages, here exemplified by adverbial infinitives, where a strict reading should be possible for pragmatic reasons, but only a sloppy reading seems to be available. In all these examples, we see that PRO can only be controlled by an antecedent located in the second conjunct and not by an antecedent in the first conjunct. If we compare this to equivalent finite adjuncts with a lexical subject (see (25)), we can observe that they allow both a sloppy and a strict reading. undisturbed 'He took the children to school so that he could work undisturbed, and so did the grandparents sometimes.'

c. English
The professori took the essays home so that hei could grade them, and so did his wifej take the essays home so that hei/shej could grade them.
In the following example, a strict reading should be available for pragmatic reasons (since this event could imply that the crew went down with the ship). However, in this case only a sloppy reading seems to be possible, with the rather strange interpretation that the crew sank after taking in large amounts of water. in large amounts water 'The ship sank after taking in large amounts of water, and so did the crew.' To sum up, we can conclude that also the third OC property holds in adjunct control at the vP/VP-level.

OC property no. 4
According to the fourth OC property described by Landau (2013), the controller of OC PRO is not restricted to [+human], but can also be [-human] (see (2d)). This is clearly attested for adverbial infinitival, participial, and small clause adjuncts in German, Norwegian, and English; see (27) The examples above do not only show that the local antecedent can be non-human; they also reveal that it need not be the subject, but can alternatively be an object in the adjunct's matrix clause (see (27b), (28b), (29), (32), (35)).
To sum up, the tests in the preceding subsections have confirmed that adjunct control into various types of adjuncts adjoining in the verbal domain qualifies as OC.

NOC in TP/CP-adjuncts
Speech act and sentence adverbial adjuncts, on the other hand, appear to be NOC adjuncts, since (i) the controller can (see (36a)), but does not have to be an argument in the adjunct's matrix clause; (ii) PRO can have an arbitrary reference or refer to the speaker; and (iii) PRO is always [+human], 10 as shown in (36)-(38). This stands in contrast to Landau's (2013) OC criteria summarized above in (2) and thus 10 Note that adverbials adjoined at the CP-layer are speech act adverbials and thus speaker-oriented. As a result, they are automatically only compatible with a [+human] PRO; so this property follows automatically from the semantics of the involved adverbials. The same holds for adverbials adjoined at the TP-layer, since they typically involve predicates that obligatorily involve a person; see (i) and (ii) as an illustration.
(i) Salat i ist, [PRO speaker/*i ohne zu übertreiben], eine der gesündesten Beilagen. salad is without to exaggerate one the.GEN healthiest side.dishes 'Without exaggeration, salad is one of the healthiest side dishes.' (ii) Dieser Sturm i ist [PRO speaker/*i meteorologisch betrachtet] einer this storm is meteorologically considered.PRF.PTCP one der stärksten seit Jahren. the.GEN strongest for years 'Meteorologically speaking, this has been one of the severest storms for years.' characterizes NOC. Speech act and sentence adverbial readings are found in adverbial infinitives and adverbial participle constructions headed by a present or past participle in all three languages, but, to our knowledge, not in adverbial small clauses headed by the particle als/as/som (cf. Emonds 1985;Eide 1996;Flaate 2007 (Faarlund, et al. 1997, 811) c. adverbial past participle construction Han var [PRO speaker kort sagt] for dårleg. he was briefly said too bad 'He was, to put it briefly, too bad.' (Faarlund et al. 1997 To sum up, we have now reached the following conclusions: as the literature on scope-based adjunct theories has convincingly shown (see, among others, Frey andPittner 1998, 1999;Pittner 1999;Haider 2000;Ernst 2002Ernst , 2014, the adjunction height of adverbials depends on their underlying semantics. Considering speech act and sentence adverbials, we can therefore, on the one hand, say that they are generally adjoined at the CP/TP-level; on the other hand, the data in this section have shown that control into these adverbials qualifies as NOC (which means that PRO's interpretation hinges on pragmatic factors and can include arbitrary PRO or a discourse-salient referent like the speaker). Hence, we can generally conclude that control into adjuncts in the TP/CP-domain involves NOC. By contrast, as outlined in the scope-based adjunct theories, event-and processmodifying adverbials are adjoined in the verbal domain, i.e. at the vP/VP-level. As shown by the data in Sect. 2.2, we argued that control into event-modifying adverbials typically involves subject control, whereas control into processmodifying adverbials involves object control, and standard OC diagnostics have shown that both instances of control qualify as OC. In fact, that subject control involves a higher adjunction site than object control has further been corroborated by the binding effects discussed in Sect. 2.1.
In the remainder of this paper, we will now turn to a potential technical implementation of these observations.

Theoretical Approach
To illustrate how our analysis works, we will restrict ourselves to data from German for reasons of space; however, we assume that the analysis can be applied to Norwegian and English in the same way.
As shown in the first part of the paper, various instances of adjunct control (i.e. those involving adjunction in the verbal domain) involve OC. In Sects. 3.1-3.3, we will concentrate on these instances of control before we then turn to non-obligatory adjunct control in Sect. 3.4, which has been shown to be involved in adjunction in the TP/CP-domain.

Licensing of OC under Agree
In line with Fischer (2018) and Fischer and Høyem (2021), we argue that OC is licensed under upward Agree and propose a phase-based theory of adjunct control; i.e. we adopt a local-derivational view according to which the accessible domain at a given point in the derivation is restricted by the Phase Impenetrability Condition (PIC) (see (39)). As a result, syntactic licensing must occur locally within the respective accessible domain. We assume the following standard definitions:

(39) Phase Impenetrability Condition (PIC)
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. (Chomsky 2000, 108) (40) CPs and vPs are phases.
Our basic assumptions are the following (see also Fischer 2018;Fischer and Høyem 2021): PRO is a referentially defective empty argument which must be referentially identified. This means that PRO's interpretation is determined in the course of the syntactic derivation by an Agree relation with an accessible binder.
To encode this idea technically, we follow Wurmbrand (2017) and suggest that there is a distinction between formal uφ-features and semantic iφ-features. While formal uφ-features "feed (only) into PF and carry the values realized in morphology [...] semantic iφ-features [...] feed (only) into LF and carry the values interpreted in semantics" (Wurmbrand 2017, 32). We argue that the relevant feature on PRO that must be licensed in control relations is an unvalued semantic iφ-feature; 11 so the fact that this feature is unvalued in the beginning reflects PRO's referential defectiveness. That is, PRO enters the derivation with the feature specification {D, iφ:_}, and to get licensed, PRO probes upwards to find a goal in the accessible domain that values its unvalued semantic iφ-feature. 12 PRO's concrete interpretation can be determined once this feature has been valued under Agree; i.e. Agree syntactically links PRO to its controller (i.e., the goal), which means with respect to PRO's interpretation that, whatever the interpretation of the controller is, this will be the interpretation of PRO. 13 Formally, we adopt the following definitions, which combine assumptions by Wurmbrand (2011) (in terms of upward probing) and Pesetsky and Torrego (2007), Bošković (2009 et seq.), Wurmbrand (2011), among others, as far as the assumption is concerned that Agree is valuation-driven.
(41) Agree a. α agrees with γ iff (i) γ c-commands α, (ii) γ is the closest goal, and (iii) α and γ are both in the accessible domain (as defined in (39)). b. If α agrees with γ, feature [F:_ ] on α is valued by feature [F: val] on γ. 11 Note that Wurmbrand (2017) points out that anaphoric Ø N in ellipsis (being phonetically zero) only bears an iφ-feature (and no uφ-feature); it therefore seems natural to extend this classification to other covert nominal elements like PRO as well. In fact, our central point is that PRO is referentially defective and that its interpretation hinges on the syntactic licensing by the controller, which, as a result, determines PRO's interpretation. Technically, this could also be encoded in terms of β-features as proposed in Fischer (2004Fischer ( , 2006 as far as (anaphoric) binding is concerned and Fischer (2018) as far as control is concerned. 12 See also, for instance, Baker (2008), Schäfer (2008) In the subsequent sections, we will outline in detail how this basic mechanism can concretely be applied to the data we have discussed in the previous sections, i.e. how it can account for adjunct control in the languages under discussion.

Basic observations
One of our insights from Sect. 2 was that adjunction in the vP-domain results in obligatory subject control. As an example, consider, again (42) (adapted from (20a)).
(42) Sie i gingen vorbei, [ CP PRO i ohne etwas zu bemerken]. they passed by without anything to notice 'They passed by without noticing anything.' (Haider 2015, 4) In tree (43), we illustrate the point in the derivation when the vP-phase of the matrix clause is built; it contains the vP-adjunct as well as the subject in its base position, Specv.
(43) 14 Let us first have a closer look at the adjunct itself. Inside the adjunct, PRO is basegenerated in Specv, where it is theta-marked. Bearing an unvalued semantic iφfeature, it probes upwards to find a suitable goal, but without success. As a result, the derivation would be doomed to fail immediately if PRO did not move to the edge of the phase to remain accessible and thereby at least retain the possibility of 14 Concerning the subsequent trees, be aware of the following aspects: The trees display OV structures as is common for German. (ii) As default, the adjunct is represented as CP, but nothing hinges on this (see also footnote 6).
Generally, the inaccessible domain in trees like (43)  having its feature valued in the next phase. In our example, this means that PRO moves successive-cyclically via SpecT to SpecC inside the adjunct (see tree (43)). 15 This line of reasoning follows a number of proposals in the literature which suggest that the need to keep constituents accessible in order to enable feature checking later in the derivation triggers movement to the phase edge. 16 This way an early crash of the derivation can be avoided in phase-based derivations.
However, the position in which PRO occurs in (43), i.e., SpecC inside the adjunct, is the highest position to which it can move since adjuncts are islands for extraction (see Huang 1982). 17 At this point in the derivation, PRO (being at the edge of the CP-phase) is still accessible, whereas material below C inside the adjunct is not; however, as soon as the derivation proceeds and vP merges with T, all material inside the adjunct becomes inaccessible; 18 this is illustrated in tree (44). So tree (43) illustrates the last point in the derivation when PRO is still accessible.

Obligatory subject control into adjuncts
Let us now derive the control relation in examples like (42), which involves obligatory subject control. Our starting point is the scenario in tree (43).
In order to account for binding effects in adjuncts, Reinhart (1976) has already proposed that the subject DP in (43) c-commands the adjunct in this configuration. 19 In the following, we adopt Bruening's (2014) simplified and generalized version of this c-command definition, which is cited in (45). 20 (45) Node A c-commands node B iff the first branching node dominating A does not exclude B. (Bruening 2014, 356) (46) α excludes β iff no segment of α dominates β. (Chomsky 1986, 9) If we apply these definitions to tree (43), we get the following results: the first branching node dominating the subject is vP; since a segment of vP dominates the vP-adjunct, the vP does not exclude the adjunct-following (45), the subject therefore c-commands the adjunct and thus PRO inside it.
To sum up, we are faced with the following situation: PRO has an unvalued semantic iφ-feature and is looking for a suitable goal; the subject DP (which has a valued semantic iφ-feature) c-commands PRO, is the closest potential goal, and PRO is still in the accessible domain (see tree (43)). As a result, Agree between PRO and the subject DP can be established (following definition (41)). This yields the desired result: PRO's semantic iφ-feature can be valued by the subject DP, which means that we get obligatory subject control into the vP-adjunct. This is illustrated in tree (47). 19 The original c-command definition Reinhart (1976) proposes in this context is as follows: Node A c(onstituent)-commands node B iff the first branching node α 1 dominating A either dominates B or is immediately dominated by a node α 2 which dominates B, and α 2 is of the same category type as α 1 . (Reinhart 1976, 148) On the basis of this definition, Reinhart (1976) can rule out sentences like the following:  (47b)) 20 Note that Bruening himself argues for the notion of phase-command (instead of c-command).
Note in addition that, although accessibility is a precondition for both Agree and movement, this is not yet a sufficient condition for movement (in contrast to Agree; cf. (41a-iii)); hence, Agree into an adjunct can be possible while extraction out of it is illicit (see Fischer 2018 for a more detailed discussion). In short, this is the case because movement hinges on further restrictions (accessibility only being a necessary condition). Since we are only concerned with Agree relations and not with movement operations in this paper, we will not discuss any further what exactly blocks movement out of islands (i.e., which sufficient precondition is not fulfilled in these cases). 21 At this point, we just want to highlight the fact that Agree and movement involve different sufficient conditions and that an Agree relation between the subject and PRO can successfully be established in (47). 21 Of course, what movement might require on top of accessibility is a relevant question which has already been addressed elsewhere in the literature. For instance, Müller's (2010Müller's ( , 2011 account of CED effects makes the following proposal: intermediate movement steps are triggered by edge features (EFs) which have to be inserted on the (intermediate) target phase's head (see also Chomsky's 2000Chomsky's , 2001Chomsky's , 2008 for an earlier version of the Edge Feature Condition). Following Müller (2010Müller ( , 2011, EF insertion is only possible if the respective phase head is still active, i.e. if it still bears some other probe or structurebuilding feature. So this might be the additional requirement on movement we are looking for: EF insertion in the target phase must be possible, otherwise movement into this phase is blocked, completely independent of the question of accessibility. So in (47) this would mean that PRO cannot leave the adjunct, simply because EF insertion on v is blocked (since the phase head is already inactive at this point of the derivation because the vP is already complete).
In fact, in Müller's original account, adjuncts are treated as being specifiers of special functional projections (in line with Alexiadou 1997;Cinque 1999;see Müller 2010, 46); our treatment of adjuncts might involve a much simpler reason: in trees like (47), movement out of the adjunct to a Specv position is simply not possible because this would violate the Extension Condition (Chomsky 1995) since the tree would not be extended at the root (and direct movement out of the vP phase is excluded by the PIC).

Adjunction in the VP-domain
Let us now turn to adjunction in the VP-domain, which results in obligatory object control. To illustrate this once again, consider example (48) (repeated from (16b) and (16a)

Obligatory object control into adjuncts involving DP DAT
First, we want to consider sentences in which the controlling object bears dative Case. As pointed out before, we follow Anagnostopoulou (1999), Pylkännen (2002), McFadden (2004McFadden ( , 2006, , among others, and assume that Appl°i ntroduces the dative argument; the underlying structure thus looks as indicated in tree (49).
Since VP is not a phase, PRO is still accessible when the object DP enters the derivation. The DP c-commands PRO, has a valued semantic iφ-feature, and is thus the closest goal for PRO. As a result, an Agree relation can be established, which yields the desired result: we get obligatory object control into the VP-adjunct (again, following definition (41)). This is illustrated in detail in tree (50).

Obligatory object control into adjuncts involving DP ACC
If the controlling object bears accusative Case, the underlying scenario looks as indicated in tree (51).
Following the c-command definition in (45), the object DP c-commands PRO since VP, the first branching node dominating the object DP, does not exclude PRO since the latter is dominated by a VP segment. 22 Again, PRO enters the derivation with an unvalued semantic iφ-feature whereas the object DP bears a valued semantic iφfeature. Since the object DP c-commands PRO and both are accessible at this point in the derivation, an Agree relation can be established and valuation can take place; i.e. OC is derived (see (52)).

On the syntactic difference between OC and NOC
Let us now turn to those cases that involve higher adjunction sites, i.e. adjunction in the TP/CP-domain. As shown in Sect. 2.3, these instances of control all involve NOC, so we concluded that the empirical picture is as follows: control into adjuncts at the vP/VP-level classifies as OC, whereas adjunct control outside the verbal domain classifies as NOC. How does this distinction come about? We assume that it is the question of how PRO is ultimately licensed which determines whether we end up with OC or NOC; i.e. the underlying syntactic structure is the same in the two control types. 23 If we do not get OC outside the verbal domain, this suggests that, at this point in the derivation, DPs do not qualify as potential goals for PRO anymore, which then implies that the valuation of PRO's unvalued semantic iφ-feature under Agree fails. So we are left with the following two questions: (i) what happens to PRO if such an Agree relation cannot be established, and (ii) what is it that renders DPs in the TP/CP-domain useless for an Agree relation with PRO?
Concerning the first question, we follow Preminger (2014) in assuming that failed agreement does not necessarily lead to ungrammaticality; instead, we assume that the semantic interpretation we end up with if PRO's semantic iφ-feature remains unvalued in narrow syntax is determined by pragmatic factors (like logophoricity; see also Landau 2015, among others). As a result, we either get arbitrary control or control by a salient entity in the discourse (like the speaker). We will illustrate this with concrete examples in Sect. 3.4.2.
As far as the second question is concerned, the following observation can be made: DPs that would be high enough in the syntactic structure to be able to c-command PRO in these high adjuncts all share the common property that they have already been Case-marked before they undergo movement to this position. 24 So we argue that it is this property which excludes them from being potential goals for PRO. By contrast, we have the following scenario in the case of OC: accusative objects license PRO inside the VP and subjects license PRO inside the vP, i.e. before the respective Case-marking heads even enter the derivation. In the case of dative objects, we have assumed that their base position is SpecAppl; so they can first license PRO inside the VP-adjunct and then receive Case from Appl°. 25 That Case and A-processes are somehow related is not a new idea and has been implemented in different ways in the literature, starting with Chomsky (1981). What has typically been in the center of discussion in the past is the question of how or whether Case-marking affects A-movement. The central correlation proposed in the 1980s was that DP-movement is driven by the need to satisfy the Case Filter. McGinnis (1998) suggested the principle of Case Identification, which implies that "[o]nce the Case feature has been checked and deleted, it cannot identify a phrase for pied-piping, so Move is blocked" (McGinnis 1998, 36). So here as well, Case checking renders a DP unsuitable for certain subsequent processes. With the advent of Agree, it was then commonly assumed that Case valuation and φ-feature valuation go hand in hand. This is most obvious in Chomsky (2000), where [uCase]features are checked via Reverse Agree as a side effect of φ-Agree. But even in proposals which do not assume such a direct relation (like Pesetsky and Torrego 2007), the respective features typically appear jointly on the same constituents and establish Agree relations at the same point in the derivation. 26 The assumption that Case also influences other processes has culminated in the postulation of the Activity Condition (see Chomsky 2000Chomsky , 2001, according to which "DPs whose case feature is valued become inactive and thereby unable to undergo subsequent A-processes" (Keine 2018, 2, based on Chomsky 2000Chomsky 2001, 6). In the meantime, the Activity Condition as such has been critically reviewed at various places (see, for instance , Nevins 2004;Preminger 2014;Keine 2018), and in particular its role in constraining A-movement has been questioned. It may well be true that the Activity Condition as such does not hold; however, we still think that the central property that DPs outside the verbal domain share is that they enter this domain already Case-marked. Hence, we suggest that this is what excludes them from establishing an Agree relation with PRO, which would license OC. As a result, control into TP/CP-adjuncts can only involve NOC. 27 In the next subsection, we will show how our examples involving NOC are derived under these assumptions. In the trees below, we call DPs whose Case features have already been valued A-inactive and thus suggest that A-inactive DPs cannot license PRO syntactically, i.e., establish an OC relation via upward Agree.

Non-obligatory control into TP/CP-adjuncts
As we have seen in Sect. 2.3, NOC into TP-adjuncts can involve arbitrary PRO (PRO arb ) or control by a salient entity in the discourse, like the speaker; since CPadjuncts involve speech act adverbials, they are typically speaker-oriented and thus only involve PRO speaker . Consider again the examples in (53) (repeated from (36d)), (54a) (repeated from (36b)) and (54b) (repeated from (36c) without to exaggerate widely and broadly the beste Billiard-Spieler. best billiard-player 'He is, without exaggeration, the best billiard player ever.' (Pittner 1999, 338) b. CP-adjunct: [PRO speaker Um ehrlich zu sein,] ich habe nie viel von for honest to be I have never much of Kriminalromanen gehalten. crime.novels held 'To be honest, I have always disliked crime novels.' (Pittner 1999, 357) Let us now see how control into TP-adjuncts is derived. When the TP is completed, the syntactic structure looks as indicated in tree (55). Independent of whether there is a DP in SpecT or not (which might not be the case in German, as pointed out in footnote 15), such a DP would already be Case-marked and thus be A-inactive when the adjunct including PRO is merged into the derivation. As a result, PRO cannot establish an Agree relation with a DP in SpecT.
Since TP is not a phase, PRO remains accessible when the derivation proceeds. However, all DPs that could enter the derivation in the CP-domain are also Casemarked DPs, which means that they could not establish an Agree relation with PRO either (see tree 56). As a result, PRO's semantic iφ-feature remains unvalued in narrow syntax. 29 As outlined in the previous section (and in line with Preminger 2014), we do not assume that this results in ungrammaticality; instead, we suggest that this gives rise to a different licensing strategy based on pragmatic factors. 30 If there is no discourse-salient referent, we end up with a default interpretation. This is what happens in (53), which means that PRO is ultimately interpreted as PRO arb (see tree (56)). 29 Note that the situation would not change if the CP in (56) introduced an embedded clause; since CP is a phase, PRO would become inaccessible as soon as the root node in (56) would merge with a potential matrix V. So DPs in the matrix clause would not be accessible goals for a PRO inside an embedded TP-or CP-adjunct either. 30 That NOC results as a last resort strategy if the conditions for OC are not met has already been proposed earlier in the literature; for some recent implementations see, for instance, Fischer (2018) and McFadden and Sundaresan (2018). However, in contrast to these two proposals, we suggest that arbitrary PRO and speaker-oriented PRO are derived in the same way via the pragmatic licensing strategy (i.e., without a syntactic Agree operation). By contrast, Fischer (2018) and McFadden and Sundaresan (2018) assume that, in the latter case, there is an Agree relation between PRO and the attitude (or perspective) holder. (54), the attitude of some attitude holder (the speaker in this case) is reflected; i.e. we deal with attitudinal contexts. Generally speaking, the attitude holder corresponds to a salient entity in the discourse which could be the speaker (as in (54)) or a non-local antecedent (as, for instance, in long distance control). If the discourse provides such a salient referent, pragmatic licensing predicts that this is PRO's interpretation. In examples like (54), PRO is therefore ultimately predicted to be interpreted as PRO speaker . 31 As outlined in footnote 30, we assume that this control relation is not syntactically licensed. 32 However, following Speas (2004), Sigurðsson (2004Sigurðsson ( , 2014, Sundaresan and Pearson (2014), Landau (2015), Fischer and Pitteroff (2016), 33 among others, logophoric anchoring as such might of course be encoded in syntax in the following way: in attitudinal contexts, a logophoric center is projected in the left periphery, which introduces the attitude holder also syntactically. Since PRO speaker behaves like a logophor, 34 licensing of PRO speaker then boils down to standard logophoric licensing: the attitude holder is the antecedent a logophor needs to be licensed (cf. Zribi-Hertz 1989).

By contrast, in examples like
Hence, the underlying tree structure for sentences like (54a) would look as indicated in (57). Note that we remain agnostic as to the precise labeling of the heads in the extended CP; in tree (57), we therefore simply use the labels C α and C γ (for instance, following Sato and Kishida 2009, C γ could be called Point-of-View 31 In other examples, it might then be more neutral to replace the term PRO speaker with PRO attitude holder . 32 As pointed out by McFadden and Sundaresan (2018, 500), "flexibility of reference is characteristic of NOC"; if the interpretation of NOC PRO is not determined in narrow syntax but rather by pragmatic factors, this is not unexpected. 33 Fischer and Pitteroff have argued for this view in their talk 'Psych verbs and control into subject clauses in German' at GGS 42, University of Leipzig. 34 In fact, NOC PRO has long been shown to behave like a logophor; see, for instance, Kuno (1975), Landau (2013Landau ( , 2015. Projection (POVP)). Note, moreover, that the complete extended CP (CP α plus CP γ ) forms the next phase.
Finally, we briefly turn to CP-adjunctions (as illustrated in examples like (54b)). In fact, the analysis does not really differ from the one outlined above for sentences like (54a) (involving a TP-adjunct). DPs in the CP-domain cannot function as a goal for PRO anymore since they are already A-inactive at this point in the derivation. Thus, PRO's semantic iφ-feature cannot be valued in narrow syntax and the pragmatic licensing strategy applies instead. Since CP-adjuncts involve speech act adverbials, they typically occur in attitudinal contexts; this means we have an attitude holder which ultimately determines the interpretation of PRO; see tree (58). In sentence (54b), we can thus conclude that it involves PRO speaker .

Empirical situation
Considering data from German, Norwegian, and English, we have come to the conclusion that these three languages seem to behave alike with respect to adjunct control. In particular, we have made the following observation: although adjunct control comprises many different constructions and adverbials and the distinction between OC and NOC cuts through the set of examples, it all seems to boil down to the clear-cut distinction summarized in (59): (59) a. Adjunction in the vP-domain results in obligatory subject control.
b. Adjunction in the VP-domain results in obligatory object control. c. A higher adjunction site yields NOC.
Crucially, the adjunction site itself is determined independently by the underlying semantics of the respective adverbial adjuncts. As has been shown in the literature on scope-based adjunct theories (see, among others, Frey andPittner 1998, 1999;Pittner 1999;Haider 2000;Ernst 2002Ernst , 2014, process-modifying adverbials adjoin to VPs, event-modifying adverbials adjoin to vPs, sentence adverbials adjoin to TPs, and speech act adverbials adjoin to CPs. So apart from providing an empirical overview involving adverbials of all types from the three languages under discussion, the first part of this paper brings together these insights from the literature on adverbials and control theory and can thus link different types of control with different adjunction heights.

Theoretical implementation
To implement these findings technically, we have proposed a phase-based approach that works as follows: OC is licensed under upward Agree with PRO (the probe) looking for an accessible goal (ultimately the controller) to value its semantic iφfeature (following Wurmbrand's 2017 terminology) and thereby establish a binding relation which referentially identifies PRO.
For vP-adjuncts, the subject DP turned out to be the closest available goal and is thus predicted to be the obligatory controller (see tree (47)); hence, observation (59a) is correctly derived. Similarly, in the case of VP-adjuncts, the object DP turned out to be the closest available goal; hence, obligatory object control is straightforwardly predicted (see tree (50) for dative objects and tree (52) for accusative objects) and observation (59b) is thus derived.
If the adverbial is adjoined outside the verbal domain, i.e. if we consider control into TP-or CP-adjuncts, we generally face the following situation: DPs occurring in these domains have already become A-inactive at this point in the derivation, which means that they cannot function as a goal for PRO anymore even if they have moved into the accessible domain. Hence, PRO's unvalued semantic iφ-feature remains unvalued in narrow syntax, which triggers a last resort licensing strategy based on pragmatic factors. As a result, the semantic interpretation of PRO is determined as follows: if there is a salient referent in the discourse (like the attitude holder, which is often the speaker itself), this will be PRO's interpretation (see tree (57) and (58)); otherwise, we ultimately end up with arbitrary PRO (see tree (56)). So, generally speaking, NOC occurs if the structural requirements for OC cannot be met (see also, among others, Fischer 2018; Mc Fadden and Sundaresan 2018). As this happens in the case of TP-and CP-adjunction, we correctly predict observation (59c).
Since we consider the different underlying adjunction sites to be the pivotal point, our analysis as such does not predict differences for the three languages under investigation; therefore, the lack of variation that we have observed when comparing German, Norwegian, and English adjunct control is, after all, expected.