A special case of long distance agreement in Marathi

This paper discusses an unexplored case of long distance agreement (LDA) in Marathi in which agreement takes place across an embedded subjunctive clause. In this construction, three distinct headsan embedded verb, a matrix verb and what appears to be a clause-introducing complementizer-like morphemeagree with the embedded object. The paper demonstrates that the subjunctive clause embedded in the LDA construction is a restructured clause lacking several functional projections such as TP, NegP and the entire CP layer. Further, it argues that the complementizer-like agreeing morpheme is a clause-linker attached lower than the CP-layer (à la den Dikken 2006). LDA in this construction is mediated by this clause-linker and the restructured nature of the clause it introduces. The analysis in this paper is based within the Minimalist framework of probe-goal mechanism of Agree.


Introduction
In this paper, I discuss one particular construction in Marathi involving long distance agreement across a subjunctive embedded clause. In this construction three distinct heads-an embedded verb, a matrix verb and a complementizer-like element-agree with the same embedded argument, i.e. the object of the embedded clause. Consider the following Marathi sentences: (1) a. mirɑ=lɑ ɑmbɑ k h ɑ-w-ɑ-s-ɑ wɑʈ-t ̪-o. Mira=dat mango.m.sg eat-sbjv-m.sg-such-m.sg feel-ipfv-3m.sg 'Mira feels like eating a mango.' b. mirɑ=lɑ kəiri k h ɑ-w-i-∫-i wɑʈ-t ̪-e. Mira=dat raw.mango.f.sg eat-sbjv-f.sg-such-f.sg feel-ipfv-3f.sg 'Mira feels like eating a raw mango.' c. mirɑ=lɑ ɑmbe k h ɑ-w-e-s-e wɑʈ-t ̪-ɑt ̪. Mira=dat mango.m.pl eat-sbjv-m.pl-such-m.pl feel-ipfv-3m.pl 'Mira feels like eating mangoes.' In these examples, the matrix verb 'feel' takes an embedded clause as its complement. The embedded verb 'eat' carries subjunctive/optative morphology -ɑw-and it agrees with its object-'mango', 'raw mango' and the plural object 'mangoes' in (1.a), (1.b) and (1.c) respectively. The matrix verb 'feel' also agrees with the embedded object, which is not theta-marked by it. Moreover, there is a -s-, glossed here as 'such', 1 attached to the embed-ded verb, which also agrees with the same embedded object. Agreement of the matrix verb and -s-with the embedded object is an instance of 'long distance agreement' (LDA hereafter), where these two elements agree with an argument not selected by them. The present paper offers an analysis of LDA in this particular construction.
The phenomenon of LDA has been reported and analysed in several languages belonging to various language families across the world, such as Basque (Preminger 2009), Hindi-Urdu (Mahajan 1989;2017;Davison 1991;Butt 1995;Bhatt 2005;Chandra 2007;Keine 2013), Kashmiri (Subbarao & Munshi 2000 as cited in Bhatt (2005)), Tsez (Polinsky & Potsdam 2001) among others. The phenomenon of LDA is of empirical as well as theoretical importance as it raises issues regarding locality of syntactic phenomena and dependency relationships. Through the literature accumulated on LDA constructions over the last few decades, several cross-linguistically common properties of this phenomenon have emerged ('universals of LDA' as dubbed by Bhatt & Keine 2017). However, no consensus on the theoretical analyses of these constructions has been reached. There have been several proposals made in the literature, some of which will be briefly discussed later in section 4. See Bhatt & Keine (2017) for a critical survey of several proposals. LDA is in fact an epiphenomenon which evades a uniform analysis of several LDA constructions even within the same language. One has to provide case-by-case analysis of individual LDA constructions and look for fundamental properties underlying these distinct cases. One such instance of LDA is the focus of this paper.
LDA across the subjunctive clause in Marathi differs from LDA reported in other languages in two aspects: one, the embedded clause across which LDA takes places in Marathi is neither a typical infinitival clause as found in the languages like Hindi-Urdu and Kashmiri, nor is it any finite clause as seen in Tsez. It is specifically a subjunctive clause. Second, something that appears like a complementizer is overtly present, albeit in a bound form, and it participates in agreement. At the outset, this complementizer-like -sappears to be a phonologically reduced form of the clause-final complementizer əsə (see Subbārāo 2012: 165 for a remark to this effect). 2 Agreeing complementizers have been reported in some West Germanic languages like Bavarian (Bayer 1984) and some Bantu languages (Diercks 2013), but not in Indo-Aryan languages. I argue that the agreeing form -s-in Marathi LDA under discussion here is not a standard complementizer lexicalizing the C 0 . It is a sort of clause-linker, attached lower than the CP layer in the structure and introduces a restructured or a small-clause complement of a matrix verb (cf. den Dikken 2006;Philip 2012). I call it 'a (clause-)linker'. To my knowledge, both these properties-LDA across a subjunctive clause and an agreeing clause-linker-have not been reported in any of the Indo-Aryan languages. Marathi, therefore, offers interesting empirical insights.
In this paper, I demonstrate that LDA across the subjunctive clause in Marathi as shown in (1) results due to the mediation of the agreeing clause-linker -s-and it also goes handin-hand with restructuring properties of the embedded clause. The paper is organized as follows. Section two discusses basic facts about verb-agreement in Marathi. Section three introduces the LDA construction and discusses its basic properties in comparison to a closely-related construction without LDA (referred to as 'non-LDA construction' now on).
In section four, a sketch of the analysis proposed in this paper is provided. More specifically, I argue that the embedded clause in the LDA construction involves restructuring in the sense that it lacks some functional projections and is thus smaller than a full CP structure. Further, the form -s-has been analysed as a low-attached clause-linker distinct from the complementizer əsə in C 0 . In section five, I provide evidence for the restructured nature of the subjunctive clause in the LDA context, by demonstrating that this clause lacks TP, NegP and arguably a CP-layer as well. In section six, the mechanism of Agree deriving the effects of LDA has been spelled-out. Section seven concludes with some discussion on open issues. 3

Basic patterns of verb-agreement in Marathi
The finite verb in Marathi agrees with the highest DP argument which carries no overt case. This means, it agrees with the unmarked nominative subject irrespective of the case of the object DP (2.a). If the subject carries ergative case (2.b) or dative case (2.c), the verb agrees with the case-less object. If both the subject and the object carry an overt case, the verb shows default agreement morphology, which is the same as 3n.sg features (2.d).
These facts about Marathi are well-documented (Pandharipande 1997;Dhongde & Wali 2009 among others) and I am presenting them here only for immediate reference.
The subjunctive clause in Marathi, which is at the centre of the discussion in this paper, occurs both as an embedded clause and a root clause. The subject of a subjunctive clause in Marathi carries ergative or nominative case, depending on the (in)transitivity of the verb and animacy or agentivity of the subject. 4 Consider the examples in (3). If the subject has ergative case, as in (3.a), then in conformation with the pattern discussed above, the morphologically unmarked object controls agreement of the subjunctive verb. If the subject has nominative case, it controls agreement, as in (3.b).
(3) a. mirɑ=ne pepər wɑts-ɑw-ɑ. Mira.f.sg=erg paper.m.sg read-sbjv-3m.sg 'Mira should read a paper.' 3 Often in my discussion in this paper, facts of Marathi LDA construction have been described in relation to the well-known facts of Hindi-Urdu LDA, since LDA is most extensively analysed in Hindi-Urdu than in any Indo-Aryan languages. However, the purpose of this paper is not to present or revisit Hindi-Urdu facts. 4 Ergativity typically manifests in perfective aspect (Dixon 1994;Deal 2012 among others). However, in at least some Indo-Aryan languages, ergative subjects also occur in certain modal constructions. See Bhatt et al. (2011: ex.11) for Hindi-Urdu. In Marathi it is well-documented that ergative subjects also occur in a subjunctive clause (Damle 1970;Pandharipande 1997;Wali 2006;Dhongde & Wali 2009 In (3.a), the transitive verb 'read' takes the ergative subject and the verb agrees with the unmarked object 'paper '. In (3.b), the subject of the intransitive verb 'go' is in nominative case and so, the subject 'Mira' controls agreement of the verb. However in (3.c), the animate subject of the intransitive verb takes ergative case. In this sentence, as there is no unmarked argument, the verb shows default agreement.
An interesting feature of verb agreement in Marathi is 'double agreement'. The 2sg subject in Marathi triggers a special morpheme -s on the finite verb, when the subject carries either nominative or (null) ergative case. This morpheme occurs irrespective of whether the verb agrees with the subject or the object or shows default agreement morphology. In (4.a), the verb agrees with the nominative subject and carries the -s morpheme, while in (4.b) the verb agrees with the object and it still carries the same morpheme. Such cases where the verb agrees with the object and also shows the 2sg features of the subject, exhibit what I call 'double agreement' (cf. Wali 2005). Davison mentions some examples of this kind to suggest that Marathi has 'two slots for agreement ' (1991: ex.29). The -s morpheme is glossed separately.
(6) a. t ̪um h i kəwit ̪ɑ t∫ h ɑn wɑts-l-i-t ̪. you.pl.erg poem.f.sg nicely read-pfv-f.sg-2pl.sbj 'You all read a poem nicely.' b. ɑm h i/ t ̪jɑn=ni kəwit ̪ɑ t∫ h ɑn wɑts-l-i-(*t ̪). we.erg/ they=erg poem.f.sg nicely read-pfv-f.sg-(*2pl.sbj) 'We/ they read a poem nicely.' The morphemes -s and -t ̪ also occur in a subjunctive clause. As usual, the morpheme -s expresses 2sg features of the subject. However, the morpheme -t ̪ on a subjunctive verb expresses 3pl features on the object. It does not express any features of the subject. The morpheme -t ̪ in the subjunctive clause occurs whenever the verb agrees with the unmarked 3pl object, except when the subject of the clause is 2sg (in which case -s occurs on the verb). See examples (7).
The two double or second agreement morphemes -s and -t ̪ are peripheral in the sense that they occur on the left-most verbal element in a finite verb-complex. Recall that in (4), the -s appears on the main verb which is the only verb in the clause. However in (8), when the tense auxiliary is present, this morpheme jumps on the tense auxiliary instead of occurring on the verb.
The subjunctive verb does not take any tense auxiliary. However, it is negated by a special prohibitive negation which occurs sentence-finally. Interestingly, the 'double agreement' morpheme -s (2sg.sbj) and the second agreement morpheme -t ̪ indicating 3pl object, occur on the negation instead on the verb which now occurs in a nonfinite form.  It is for this reason, I treat both -s and -t ̪ morphemes as peripheral and coming from the 'second agreement slot', even in the case of subjunctive verbs. It is important to note further that, the prohibitive nəje shows agreement only through the morphemes -s (2sg.sbj) and -t ̪ (3pl.obj). In all other cases, the prohibitive does not show any agreement. It is in this regard too, the second-slot agreement differs from standard agreement. We will see later that the presence/absence of these two morphemes on the subjunctive verb provides additional support for the restructured nature of the subjunctive clause (cf. section 5.5.3).

Properties of the LDA construction
The example of LDA across the embedded subjunctive clause in (1.a) is repeated here as (11.a) for immediate reference. The sentence in (11.b) is often judged by the speakers as a paraphrase of the LDA construction in (11.a). 5 (11) a. mirɑ=lɑ ɑmbɑ k h ɑ-w-ɑ-s-ɑ wɑʈ-t ̪-o. Mira=dat mango.m.sg eat-sbjv-m.sg-such-m.sg feel-ipfv-3m.sg 'Mira feels like eating a mango.' b. mirɑ=lɑ ɑmbɑ k h ɑ-w-ɑ əsə wɑʈ-t ̪-ə. Mira=dat mango.m.sg eat-sbjv-m.sg such feel-ipfv-3n.sg 'Mira feels like eating a mango.' 5 A note on English translation of the Marathi constructions: The LDA construction in (11.a) expresses desire, and as such it could be naturally translated as "Mira wants to eat a mango". The non-LDA counterpart in (11.b) is ambiguous. It expresses Mira's desire to eat a mango. In addition, it also expresses Mira's opinion that she should eat a mango, whether or not she wants to do so. It could be naturally translated as "Mira thinks that she should eat a mango". This opinion-reading is missing from the LDA construction. In this paper, I keep aside the opinion/judgement reading of the construction in (11.b) and will only refer to its desire reading. In its desire reading, it is a non-LDA counterpart of (11.a) and is therefore, translated in the same way as the LDA construction. The choice of 'feel' to translate the Marathi verb wɑʈ is guided by the core lexical meaning of this verb, which is 'to feel'. It is a highly polysemous verb which is used to express the subject's feelings, thoughts, beliefs and opinions, and takes a wide variety of complement types as its argument, including an adjectival small-clause to give an expression comparable to "John finds Mary stupid" or a full clause as in (11.b).
The matrix verb wɑʈ 'feel' takes a dative-subject and therefore it cannot agree with the subject. In (11.a), as mentioned in section 1, both the matrix verb wɑʈ 'feel' and the form -s-agree with the embedded object 'a mango'. In (11.b), the matrix verb and the complementizer əsə show 3n.sg features (seemingly the default agreement morphology). Further note that in both the sentences, the embedded verb k h ɑ 'eat' agrees with its own object 'a mango'. In other words, the embedded verb agrees with its object, irrespective of whether LDA takes place or not. Let us now discuss some of the properties of LDA across the subjunctive clause.

(Non-)Optionality of LDA
In a variety of Hindi-Urdu discussed by Bhatt (2005), LDA is shown to be optional and it is characterised by co-variation of agreement of both matrix and embedded verbs: either both the verbs agree with the embedded object or both fail to agree, exhibiting default agreement. LDA across the subjunctive clause in Marathi is however, not optional. There is also no dependency between agreement of the embedded verb and that of the matrix verb. As seen in (11), the embedded verb in the subjunctive clause agrees with its own argument locally, irrespective of whether or not LDA takes place. However, agreement of the matrix verb co-varies with the presence of the form -s-and its agreement. The form -s-obligatorily agrees with the embedded object. And in that case, the matrix verb must agree with the embedded object as well. In (12.a), -s-agrees with the embedded object, while the matrix verb does not. In (12.b), while the matrix verb agrees, -s-does not. In (12.c), neither -s-nor the matrix verb agrees with the embedded object. All of these strings are unacceptable. Sentences in (12.b-c) particularly demonstrate that the form -s-must participate in LDA.
(13) *mirɑ=lɑ kəiri k h ɑ-w-i ə∫-i wɑʈ-t ̪-e. Mira=dat raw.mango.f.sg eat-sbjv-f.sg such-f.sg feel-ipfv-f.sg Intended: 'Mira feels like eating a raw mango.' In other words, there is co-variational dependency between the occurrence of the form -s-and LDA. If -s-occurs, it agrees and the matrix verb also agrees with the embedded object. If the complementizer əsə occurs instead, it cannot agree and the matrix verb also cannot agree with the embedded object. LDA across the subjunctive clause is not optional.

Absence of an overt subject
Another property of LDA across the subjunctive clause is that it does not permit an overt, lexical subject. In (14.a), the embedded subject 'Seema' is explicitly present in the LDA context and the sentence is unacceptable. When there is no LDA, an overt embedded subject is acceptable (14.b). The 'absent' embedded argument in the LDA construction is always interpreted to be co-referential with the matrix subject. But, the LDA construction also does not permit an anaphoric subject ɑpəɳ 'self' which is co-referential with (i.e. bound by) the matrix subject, while the non-LDA construction permits it.
The subject can be dropped in the non-LDA construction and is interpreted as co-referential with the matrix subject. Therefore, I assume that the covert subject in the non-LDA construction is an instance of pro-drop of an anaphoric subject. On the other hand, I claim that the embedded subject position in the LDA construction is absent altogether (see section 4). 6 Moreover, I will demonstrate that the embedded clause has a PRO argument in the spec-vP of the embedded verb, obligatorily controlled by the matrix subject (cf. section 5).

Interim summary
Let us summarize the properties of the LDA construction: • The matrix verb must be wɑʈ 'feel'. • The embedded verb must have a subjunctive/optative mood morpheme -ɑw-.
• The subjunctive embedded verb further carries a morpheme -s-whose status is yet to be determined. • The embedded clause must not have an overt subject. The agent of the embedded verb is interpreted as co-referential with the matrix subject. • The embedded verb agrees locally with the embedded object if present. Else, it would show default agreement in accordance with the general agreement patterns of Marathi.
(i) a. (from Goph by Gauri Deshpande) t ̪i kəd̪ h i je-t ̪-e-s-ə dz h ɑ-l-ə-j. she when come-ipfv-f.sg-such-3n.sg happen-pfv-3n.sg-prs 'I am eager for her to come.' (lit. To me, it's like when she would come.) b. mə=lɑ rɑd̪ h ɑ=ne puɳjɑ=lɑ dzɑ-u nəje-s-ə wɑʈ-t ̪-ə. I=dat Radha=erg Pune=dat go-nfin proh-such-3n.sg feel-ipfv-3n.sg 'I feel that Radha should not go to Pune.' Further, the form -s-in these cases can be replaced by əsə without any further change. I believe that -s-in these examples is a genuine case of a purely phonological reduction of the full complementizer əsə, a result of the connected speech, and it can be handled by low level phonetic rules.
• The morpheme -s-must agree with the embedded object if present. Else, it would show default agreement. • The matrix subject carries dative case, hence the matrix verb cannot agree with its subject. Instead, it must agree with the embedded object if present. Else it shows default agreement.

Analysis of LDA across the subjunctive: a proposal
In the literature on LDA in a variety of different languages, several analytic alternatives have been proposed. As agreement is treated as a local syntactic phenomenon, agreement becomes interesting and theoretically important when it appears to violate locality. In the literature on the analysis of LDA, two possible strategies of achieving long distance agreement have recurrently emerged: (i) restructuring in sense of Wurmbrand (2001), where an embedded structure is smaller than a full clause, lacking some functional projections, and (ii) movement of the agreement-triggering phrase in the domain of a target head. Mahajan (1989;2017); Chandra (2007) and Koopman (2006) propose movement to forge a very local spec-head relationship between the agreeing head and the trigger DP, while Polinsky & Potsdam (2001) and Keine (2013) propose movement of an argument to the edge of a domain (i.e. a clause) from where it can be accessed by agreeing heads outside the domain. The former movement has been referred to by Keine (2013) as 'long movement' and the latter as 'short movement'. The long-movement of the embedded object as proposed by Mahajan and Chandra is driven by case. In these accounts, case and agreement are associated. However, the case-driven long A-movement cannot account for LDA across the subjunctive clause in Marathi.
I propose that the embedded object in Marathi receives accusative case in its own clause even in the LDA construction. Recall that agreement of the embedded subjunctive verb is not dependent on LDA. The embedded subjunctive verb agrees with its unmarked object in both LDA and non-LDA contexts. Further, I argue that both types of subjunctive clauses have a vP projection and an agent argument (a lexical DP or a pro in the non-LDA context and a PRO argument in the LDA context). It implies, according to Burzio's Generalization, that the embedded v 0 in both types of subjunctives has the potential to assign accusative case to the object. Therefore, there is no obvious reason to infer that the case-assignment for the internal argument may be different in the embedded subjunctive in the LDA and the non-LDA contexts. Arguably, in both LDA and non-LDA contexts, the v 0 in the embedded clause assigns accusative case to the object. Thus, the embedded object in the LDA context does not need to move into the matrix clause to secure case.
Further, Bhatt & Keine (2017) use idioms as an argument against movement analysis for Hindi-Urdu LDA. They argue that scrambling of the embedded object which is part of an idiom decreases the availability of the idiomatic reading in Hindi (cf. (15)). The fact that the idiomatic reading is not disrupted under LDA in Hindi shows that the embedded object must not have moved out of the embedded predicate (cf. (16) A parallel argument can be made for Marathi: the idiomatic reading becomes deviant by scrambling the object included in the idiom. But under LDA, the idiomatic reading is not lost.
'Radha is feeling like idling away the time today.' Thus, movement-either long or short-of the embedded object is not needed for LDA to take place. In particular, the availability of the idiomatic reading suggests the embedded object does not move out of the vP, ruling out short-distance (i.e. within the clause) movement as well. Davison (1991) has also argued against the movement-based analysis of LDA in Hindi. Davison's analysis involves percolation of phi-features up in the syntactic projections yielding LDA. Another analysis of LDA which does not argue for movement is proposed by Legate (2005) who claims that agreement across syntactic phases (therefore, in long-distance manner) takes place through the mediation of intermediate heads (head-to-head agreement). Her analysis involves cyclic application of local agreement relationships bringing about the surface effect of LDA. Both Davison's and Legate's proposals have an appeal for explaining LDA across the subjunctive clause in Marathi in which multiple adjacent heads participate in agreement. In this construction, the presence and mediation of -s-facilitates LDA. The proposal which I adopt in this paper indeed involves cyclic application of local agreements, but it also differs in certain ways from both Davison's and Legate's approaches. In accordance with Davison's approach, I propose that in this instance of LDA, agreement in each clause is triggered by an argument of a verb in that clause: the embedded verb agrees with the embedded object, and the matrix verb agrees with its own internal argument which is clausal in nature. However, I do not resort to feature percolation or an equivalent mechanism as widely as Davison does. I also do not propose head-to-head agreement as Legate does. See section 6 for details.
One of the reasons for not adopting feature-percolation and cyclic agreement analyses as they are, is that they do not extend to other instances of LDA in Marathi or even to the standard monoclausal agreement of the main verb and the auxiliary. Note that in (19), the embedded non-finite verb does not agree with its argument while the matrix verb agrees with it, leading to LDA. Also in (20), the aspectual participle form of the main verb does not agree with the subject, while the past tense auxiliary does.
səmir kəwit ̪ɑ wɑts-ət ̪ hot ̪-ɑ. Samir.nom poem.f.pl read-ipfv be.pst-3m.sg 'Sameer was reading poems.' Cases such as these, where intermediate heads show no agreement, require extra stipulations (such as abstract agreement) in the feature-percolation and cyclic agreement analyses. Instead, I propose in section 6 that several heads in a clause independently agree with the same argument.
In another alternative proposed by Butt (1995), cyclic mechanism of local agreement is employed in addition to 'complex predicate formation'. Complex predicate formation involves 'fusion of event structures or argument structures of two verbs resulting in a simplex grammatical structure with a single predicate' (consisting of multiple lexical items) and a single subject (Butt 1995: 2). Typically such a complex predicate consists of a cluster of verbs. The Marathi LDA construction does not seem to involve complex predicate formation, as the embedded verb carries mood and agreement morphology which suggests that at least some structural material intervenes between the two verbs. 9 I propose that a different kind of verb-cluster formation is indeed possible in the LDA construction being discussed in this paper, but such a cluster-formation is not necessary for LDA (cf. section 5.5.1 and references cited therein).
Note that many of the proposals discussed above include restructuring of the embedded clause in Hindi LDA in addition to other mechanisms of agreement (cf . Koopman 2006;Keine 2013;Mahajan 2017). On the other hand, Bhatt (2005) and Boeckx (2004) propose that restructuring of the embedded clause is enough to account for LDA in Hindi. Bhatt claims that the restructured clause in Hindi-Urdu LDA is a subjectless TP with an infinitival T 0 (Inf 0 in his terminology) taking a vP complement with no external argument. Boeckx, on the other hand, claims that the embedded material in the Hindi LDA construction is a bare VP (2004: 32). In LDA in Marathi being discussed here, the embedded subjunctive clause also exhibits restructuring properties which I discuss at length in section 5. However, I argue that while LDA and the restructuring properties of the subjunctive clause always co-occur, LDA itself is achieved by a series of local agreements in which the verb in each clause agrees with its own internal argument.
My proposal is as follows. The subjunctive clause that occurs embedded in the non-LDA construction is a full CP. On the other hand, the subjunctive clause embedded in the LDA construction, is restructured in the sense that it lacks some functional projections that a full clause has, specifically, at least a TP, a NegP and crucially the entire CP layer. This restructured clause is selected by a clause-introducing functional element--s--which I analyse as a clause-linker (à la den Dikken 2006; Philip 2012) located in some functional projection lower than CP projections in a clause. For the want of a better alternative, I call this projection LinkP (cf. den Dikken 2006). The structure of the LDA construction is given in (21).  (21) is lexically specified to select for a restructured subjunctive clause. Also that, this linker has unvalued phi-features which make it probe and locate the embedded object as a goal, leading to LDA. As this clause-linker necessarily agrees with the embedded object, I claim that LDA across the subjunctive in Marathi is mediated by the agreeing form -s-. Since both LDA and the selection of the restructured clause depend on -s-, the two phenomena necessarily co-occur. Moreover, I shortly demonstrate that the linker is distinct from the complementizer əsə, despite its appearance as a reduced version of the latter.
The complementizer əsə always occurs in a full CP clause. Further, it has fixed, lexically valued phi-features (3n.sg). It does not probe. The structure of the non-LDA construction in which əsə occurs, is given in (22). In both LDA and non-LDA constructions, the matrix clause probes fail to agree with the matrix subject as it carries dative case. However, they agree locally with the internal argument of the matrix verb-CP in case of the non-LDA construction and LinkP in case of the LDA construction. However, since LinkP is headed by a linker carrying phi-features of the embedded object, the matrix verb also lands up with the same features through agreement with the LinkP. In other words, the presence of an agreeing linker facilitates the effect of LDA, while the presence of a non-agreeing complementizer yields the effect of blocking LDA. The mechanism of agreement in LDA and non-LDA constructions is discussed in detail in section 6. In the next section, I discuss the status of -s-as a linker.

A note on the status of -s-
Due to striking form-similarity between this linker and the final complementizer əsə, as well as between the LDA and non-LDA constructions, the form -s-is often judged as a phonologically reduced form of əsə. The two forms --s-and əsə may also be historically related (cf. Damle 1970), but there is no diachronic investigation available yet to verify this hunch. The form -s-, which I refer to as a linker, has extremely limited distribution, unlike other genuine complementizers in Marathi. The form -s-does not come attached to any clause other than the restructured subjunctive clause being discussed here. However, it occurs in two other contexts. In one context, the form -s-can attach to adjectives to express meanings similar to 'sort-of' or '-ish' in English (as in small-ish). In this case, both the adjective and the form -s-agree with the noun modified by the adjective. See examples in (23). I call it 'adjectival -s-'.
(23) a. t∫ h oʈ-i-∫-i mulgi small-f.sg-such-f.sg girl 'a small-ish girl / a small sort of girl' b. t∫ h oʈ-ɑ-s-ɑ mulgɑ small-m.sg-such-m.sg boy 'a smallish boy / a small sort of boy' In another instance, this form occurs attached to two negative auxiliaries-nɑhi which is a present-tense negative form and also the default negation in Marathi, and nəko which is the negative imperative/desirative (expressing the meaning 'not want/need'). The -s-attaching to these negations also participates in agreement. These negatives appear to be converted to adjectives by the addition of -s-, especially nəko-s-form, as can be seen from (24.c). The clause-linker is distinct from both the adjectival -s-and the negative-attaching -s-. I take the -s-attached negations to be lexicalized, fossilized forms, not productively derived in syntax. The reason is that these are the only negations that come attached with -s-. Both the adjectival -s-and the clause-linker are more productive.
The linker is distinct from the adjectival -s-due to following reasons. One, the adjectival -s-derives adjectives out of adjectives. However, the embedded subjunctive clause in the LDA context does not have properties of an adjectival phrase or an adjectival small clause. Indeed the matrix verb 'feel' can take a variety of complements, including an adjectival small clause such as in (25). However, the subjunctive clause attached with -s-is different from the adjectival small clause.
(25) mirɑ=lɑ [ɑmbe məhɑgɖe] wɑʈ-t ̪-ɑt ̪. Mira=dat mango.m.pl costly.m.pl feel-ipfv-3pl 'Mira finds mangoes costly.' The sentence embedding an adjectival small clause has a very different meaning than the LDA construction being discussed here, as the translation in (25) shows. Moreover, the two constructions are structurally very different: the adjectival small clause is a typical verbless clause which leads to the raising of a DP in it into the higher clause for case (or has ECM). I propose that the restructured subjunctive clause is bigger in size than the adjectival small-clause. It involves a PRO argument in its vP (see section 5.2). The LDA construction therefore, involves control and is thus structurally different from (25). Further, note that the adjectival small clause can be questioned by 'how', while the subjunctive verb cannot be. The answer to the question in (26.a) could involve true adjectives but not the -s-attached subjunctive verb.
(27) a. t∫ h ɑn gɑ-u ∫ək-el əsɑ mulgɑ nice sing-nfin can-fut.3sg such.m.sg boy 'a boy who can sing nicely' b. *t∫ h ɑn gɑ-u ∫ək-el-s-ɑ mulgɑ nice sing-nfin can-fut.3sg-such.m.sg boy Intended: 'a boy who can sing nicely' The adjectival -s-is optional in the sense that an adjective does not need to have this morpheme in order to be used. The linker may also be omitted as seen in (28), however such instances are relatively rare. Further, unlike in case of adjectives, where omission of -s-leads to the loss of -ish-like meanings, the omission of the linker -s-as in (28), seems to make no semantic difference. In other words, the clause-linker -s-appears to be semantically vacuous. 10 According to Philip (2012), complementizers are also a kind of linkers. Accordingly, as both -s-and əsə are linkers, their similarities or historical association may be accounted for. Nonetheless, they are distinct linkers, with their own distinct selectional and other lexical properties. The full complementizer əsə selects any finite clause, while the linker -s-selects only a restructured subjunctive clause. The full complementizer can also occur with a wide variety of matrix predicates such as m h əɳ 'say', sɑŋ 'tell', pɑh 'see', wit∫ɑr 'ask', kəɭ 'come.to.know' etc., including wɑʈ 'feel'. The linker occurs with only 'feel' matrix verb. Additionally, the full complementizer is a free and non-agreeing form. I propose that it has fixed 3n.sg features of its own lexically valued and therefore it does not act as a probe for agreement (cf. section 6.1 for details). The linker, on the other hand, is a bound form and it has unvalued phi-features. It therefore acts as a probe and agrees on finding a suitable goal to value its features. The complementizer əsə occupies the C-head, like any other complementizer. The linker, on the other hand, occurs immediately on top of the restructured clause it introduces. I demonstrate that this clause is a MoodP and it lacks TP as well as NegP. Thus, the linker heads a functional projection-LinkP-which occurs right above the MoodP. The linker does not occur in the CP-layer of a clause.
Treating -s-as a low-attached clause-linker allows us to capture the co-occurrence of LDA, the linker and the other properties of this construction which suggest restructuring in a principled way: both LDA and restructuring are the properties of this linker (cf. section 5). Further, it also suggests the absence of a phase boundary, which allows the movement of the embedded material into the matrix clause (cf. sections 5.5 and 5.7).

Evidence of restructuring
I now present arguments to demonstrate that the subjunctive embedded clause in the LDA construction involves syntactic restructuring. In the literature on restructuring, various diagnostic tests have been proposed, such as long-passivization (Wurmbrand 2001), anaphor-binding, control and agreement (Butt 1995) and NPI-licensing (Bhatt 2005). Most of these tests are not useful here: For instance, the long-passive test does not apply to Marathi LDA since the matrix verb in the LDA context under consideration is a dative-subject predicate which does not passivize (cf. Rosen & Wali 1989). However, transparency of the NPI-licensing can serve as one diagnostic, since it can be shown to occur due to restructured nature of the embedded clause. I demonstrate the restructured nature of the embedded subjunctive clause through the absence of positions such as TP, NegP etc. A crucial evidence of restructuring also comes from the fact that the embedded verb in the LDA context fails to participate in 'double or second agreement' (recall section 2).

Presence of a vP and a MoodP
First, we need to ascertain which functional projections the subjunctive embedded clause does have. The embedded verb in the LDA construction is necessarily an agentive verb. The presence of semantic agentivity, ascertained by the embedded verb's compatibility with an agentive adverb such as 'deliberately' and 'to one's satisfaction' suggests the presence of the vP-layer.

Presence of the obligatorily controlled PRO
The embedded subjunctive clause in the LDA context not only has a vP, but also has a PRO argument in the Spec-vP position. PRO takes the Agent theta role from the embedded verb and is obligatorily controlled by the matrix subject. I use two diagnostics-Davison's Dative Case Restriction and sloppy reading-to demonstrate the presence of PRO.

Dative case restriction
According to Davison's Dative Case Restriction (Davison 2008), in obligatory control constructions, dative-subjects do not occur as PRO. Davison uses this generalization as a test to determine if a construction involves obligatory control. This generalization has been shown to apply to complement clauses in Marathi (see Rosen & Wali (1989: 15), as quoted in Davison (2008: 42)  The impossibility of embedding a dative-subject predicate in the LDA construction suggests that the embedded clause in LDA involves a PRO argument. More specifically, the co-referential embedded argument-the receiver of a job in (31.a)-must be obligatorily controlled PRO. Note that in the non-LDA construction, there is a strong preference for an 11 The irrealis, subjunctive mood is often located low, over the verbal projections, in contrast to epistemic moods which are located higher in the CP-layer (Cinque 1999;Portner 2011). 12 There is at least one example where seemingly the Dative Case Restriction appears violated: (i) a. t ̪jɑ=lɑ widze=tsɑ d̪ h əkkɑ bəs-l-ɑ. he=dat electricity=gen shock.m.sg sit-pfv-3m.sg 'He got an electric shock.'

b. t ̪o [PRO widze=tsɑ
d̪ h əkkɑ bəs-un] me-l-ɑ. he PRO electricity=gen shock sit-cnprt die-pfv-3m.sg 'He died by getting an electric shock.' Note however, the embedded clause in this case is an adverbial clause in which the verb is in a conjunctive participial form. I did not find exceptions to Davison's Dative Case Restriction in any complement clauses in Marathi. overt subject (an anaphoric subject in (31.b)). This indicates that the covert subject in the non-LDA construction is not PRO.

Sloppy reading under ellipsis
Another common diagnostic used for the presence of a PRO is sloppy reading under ellipsis (Landau 2004 among many others). Consider (32).
(32) a. mirɑ=lɑ ɑmbe k h ɑ-w-e-s-e wɑʈ-t ̪-ɑt ̪ ɑɳi Mira=dat mango.m.pl eat-sbjv-m.pl-such.m.pl feel-ipfv-3pl and rɑd̪ h ɑ=lɑ=sud̪ d̪ h ɑ. Radha=dat=too 'Mira feels like eating mangoes and Radha does too.' b. mirɑ=lɑ ɑmbe k h ɑ-w-e-t ̪ əsə wɑʈ-t ̪-ə ɑɳi Mira=dat mango.m.pl eat-sbjv-m.pl such feel-ipfv-3n.sg and rɑd̪ h ɑ=lɑ=sud̪ d̪ h ɑ. Radha=dat=too 'Mira feels like eating mangoes and Radha does too.' In both these sentences, the entire matrix verb-phrase along with the clause embedded in it is deleted by ellipsis. In (32.a), the only reading possible is sloppy: Mira wants Mira to eat mangoes, while Radha wants Radha to eat mangoes. On the other hand, (32.b) is ambiguous between a sloppy reading and a strict reading. The preferred reading is sloppy in which Mira and Radha each want themselves to eat mangoes. However, the strict reading that Radha wants Mira to eat mangoes is also available. Obligatorily controlled PRO typically blocks the strict reading as in the case of (32.a).
From both of these tests, we can infer that the embedded subjunctive clause in the LDA construction has an obligatorily controlled PRO argument. The embedded clause in the non-LDA construction has a lexical DP or a pro.

Absence of a TP
While the embedded subjunctive clause in the LDA construction has vP and MoodP, it lacks functional projections higher than MoodP. I claim that this clause particularly lacks a TP projection. Recall that the subjunctive clause in Marathi can also occur as a root clause. I propose that the subjunctive that occurs as a root clause and as a complement clause in the non-LDA context has TP, but there is no evidence for the presence of TP in the subjunctive under LDA.
Consider sentences in (33) that depict instances of a subjunctive as a root clause. In (33.a) and (33.b), the subjunctive is used to express suggestions or injunctions such as in Do-s and Don't-s instructions. It expresses a speaker's wish in (33.c). In (33.b), the overt subject is optionally present. It is common in Marathi to drop the overt subject in an injunctive statement, when the injunction is deemed applicable universally. here silence.f.sg maintain-sbjv-f.sg 'Everybody / One should maintain silence here.' c. mə=lɑ tsɑŋgli nokri miɭ-ɑw-i. I=dat good.f.sg job get-sbjv-f.sg 'I should get a good job (I wish/hope).' Thus, when the subjunctive occurs as a root clause, a lexical subject (or a pro) is present. The lexical subject or the pro is also present in the subjunctive clause in the non-LDA contexts (cf. (14.b)). Further, in the injunctive reading as in (33.a-b), there is an overall generic sense-all future instances of the events of 'coming to school' or 'maintaining silence' come under the scope of these injunctions. The wish/hope reading obviously locates the event of 'getting a job' in future times. In other words, these root subjunctive clauses have an independent, i.e. deictic tense, which locates the events in future times with respect to the utterance time. I assume therefore that the root subjunctive also contains a syntactic projection of the tense, namely a TP. Borrowing from Stowell (1982), let us say the tense of the root subjunctive clause is always 'irrealis' or 'unrealised future'. 13 The irrealis tense or futurity of the subjunctive verb can further be confirmed from the fact that a temporal adverb 'yesterday' is incompatible with the subjunctive verb (see (34)).
(35) a. mirɑ=ne kəwit ̪ɑ wɑts-ɑw-i əsə rɑd̪ h ɑ m h əɳɑ-l-i. Mira=erg poem.f.sg read-sbjv-f.sg such Radha say-pfv-3f.sg 'Radha said that Mira should read a poem.' b. mirɑ=ne kəwit ̪ɑ wɑts-ɑw-i əsə rɑd̪ h ɑ=ne ʈ h ərəw-l-ə. Mira=erg poem.f.sg read-sbjv-f.sg such Radha=erg decide-pfv-3n.sg 'Radha decided that Mira should read a poem.' Note that the embedded clauses in (35) are introduced by a complementizer əsə which introduces only finite clauses (Bayer 1984;Pandharipande 1997;Dhongde & Wali 2009). The presence of a nominative or an ergative lexical subject, 14 as well as the proposal for the presence of a TP headed by the irrealis tense, are compatible with the claim that these subjunctive clauses may be finite in nature and have a full clausal structure. They are in fact CPs. Now consider (36). We have a non-LDA construction in (36.a). Note that both the matrix clause and the subjunctive clause can be independently modified by distinct temporal adverbs. In contrast, when LDA takes place, the embedded subjunctive clause cannot have an independent temporal adverb (36.b).
(36) a. kɑl mirɑ=lɑ ud̪ jɑ gɑɖi g h jɑ-w-i əsə wɑʈ-ət ̪ yesterday Mira=dat tomorrow car.f.sg take-sbjv-f.sg such feel-ipfv hot ̪-ə. be.pst-3n.sg 'Yesterday Mira was feeling like buying a car tomorrow.' b. *kɑl mirɑ=lɑ ud̪ jɑ gɑɖi g h jɑ-w-i-∫-i wɑʈ-ət ̪ yesterday Mira=dat tomorrow car.f.sg take-sbjv-f.sg-such-f.sg feel-ipfv hot ̪-i. be.pst-3f.sg Intended: 'Yesterday Mira was feeling like buying a car tomorrow.' The two events of wanting and of buying in (36.b) must be concomitant. In other words, the embedded clause in the LDA context appears to not have a tense independent of the tense of the matrix clause, as it does not have access to the utterance time. It is not incompatible with the proposal that the clause lacks the whole TP. Therefore, I assume that the subjunctive embedded clause in the LDA context lacks TP.

Absence of a NegP
Yet another evidence of restructuring comes from the prohibition of negation in the embedded subjunctive clause under LDA. Marathi has several 'negative auxiliaries' which inherently have tense and mood information in them (Damle 1970;Pandharipande 1997;Wali 2005;Dhongde & Wali 2009). In a typical verbal complex in any Marathi clause, the main verb precedes the negation and the negation precedes the tense and agreement morphology (see (37)). Given the linear order "V -Aspect -Negation-Tense" in Marathi, I assume that the NegP is located lower than the TP, but possibly higher than the AspP and the MoodP.
(38) a. rɑd̪ h ɑ=ne kəwit ̪ɑ wɑts-ɑw-i. Radha=erg poem.f.sg read-sbjv-f.sg 'Mira should read a poem.' b. rɑd̪ h ɑ=ne kəwit ̪ɑ wɑts-u nəje. Radha=erg poem.f.sg read-nfin proh 'Mira should not read a poem.' In the non-LDA construction, the embedded clause is negated with the prohibitive (39.a) just like in case of a root clause. It is also possible to negate both the embedded subjunctive verb and the matrix verb independently (39.b).

The effects of restructuring
The LDA and non-LDA constructions contrast with respect to some properties such as the scope of negation, the quantifier-scope etc. This contrast further supports arguments in favour of restructuring of the subjunctive clause in the LDA context. That the subjunctive clause under LDA is transparent for the scope of negation as well as for the quantifierscope can be demonstrated to be due to its restructured nature.

Transparency of the scope of negation
Despite the unavailability of NegP inside the restructured subjunctive, sometimes a tensed negative auxiliary seemingly occurs inside such a clause and takes the scope over the matrix clause. In the sentence in (41.a), the tensed negation occurs in between the embedded object and the embedded verb. So, it seems to be located inside the embedded clause. This linear order is not available for the non-LDA construction (41.b).
(42) koɳɑ=lɑ=hi ɑmbe nɑhi k h ɑ-w-e-s-e anyone=dat=emph mango.m.pl neg.prs eat-sbjv-m.pl-such-m.pl wɑʈ-l-e. feel-pfv-m.pl 'No one felt like eating mangoes.' Thus, the embedded subjunctive appears to be transparent for the scope of negation under LDA. The contrast between the LDA and non-LDA construction such as in (41-42) is used by Bhatt (2005) as a diagnostic for restructuring in Hindi-Urdu.
However, on a closer look, it is clear that the tensed negation in (41.a) and (42) must actually be located in the matrix clause. First, recall that the subjunctive verb can only be negated by the prohibitive in Marathi. The tensed negation must therefore negate the matrix verb. Secondly, a truly embedded negation cannot license the matrix NPI. For instance, in a non-LDA construction, the prohibitive inside the embedded clause negates the subjunctive verb and it cannot license the matrix NPI subject (43).
(43) *koɳɑ=lɑ=hi ɑmbe k h ɑ-u nəje-t ̪ əsə wɑʈ-l-ə. anyone=dat=emph mango.m.pl eat-nfin proh-3pl.obj such feel-pfv-3n.sg Intended: 'No one felt like eating mangoes.' Thus, the apparent 'transparency of the scope of negation' in (42), in fact indicates that the negation lies in the matrix clause. The linear order and the scope properties in (41) and (42) respectively, could be the result of a rightward head-movement of the embedded verb into the matrix clause across the tensed negation. The possibility of this movement corresponds with the restructured nature of the embedded subjunctive. Such a movement of the embedded verb is not possible out of the subjunctive clause in the non-LDA construction, which is a full CP. See Homer & Bhatt (2020) for a similar account for Hindi-Urdu which in essence can be extended to the Marathi data presented here.
The embedded subjunctive in the LDA construction is restructured and lacks a CP. The linker -s-is a bound form, which allows the head-movement of the embedded verb in Link 0 , where the verb adjoins the linker. The embedded subjunctive verb can then move into the matrix clause. It forms a cluster with the matrix verb. Then the cluster of the two verbs can move to right-adjoin the tensed negation (or possibly move past the tensed negation into some higher functional head), yielding the word order "embedded objecttensed negation -embedded verb -matrix verb" as observed in (41).
In the non-LDA construction, the head-movement of the embedded subjunctive verb into the matrix clause is blocked by the presence of the complementizer əsə. Since this complementizer is a free form occupying C 0 , I assume that it does not permit head-movement into C 0 . As a result, the embedded subjunctive verb will have to skip C 0 en route to the matrix clause. This would be violation of the Head Movement Constraint (HMC). The embedded subjunctive verb therefore cannot move into the matrix clause. It cannot form a verb-cluster with the matrix verb and as a result, the order "embedded object -tensed negation -embedded verb -matrix verb" is impossible for the non-LDA construction.
Thus, the seemingly embedded position of the tensed negation, as well as its apparent transparency of the scope from that position are facilitated by the linker and the restructured nature of the clause it introduces. The details of these head-movements are beyond the scope of this paper and need not detain us here. It must be noted that these headmovements and subsequent verb-cluster formation are optional, and are not necessary for LDA in Marathi (see also Homer & Bhatt 2020 for Hindi).

Object extraction and quantifier-scope
A subtle difference in the quantifier scope has been observed to correlate with LDA and non-LDA in Hindi-Urdu (Bhatt 2005;Chandra 2007). Similar effects are observed in Marathi too: (44) a. mirɑ=lɑ səgɭjɑ kəwit ̪ɑ wɑts-ɑw-jɑ-∫-ɑ wɑʈ-t ̪-ɑt ̪. Mira=dat all.f.pl poem.f.pl read-sbjv-f.pl-such-f.pl feel-ipfv-3pl (1) 'Mira feels like reading all poems.' (feel > all) (2) 'All poems are such that Mira feels like reading them.' (all > feel) b. mirɑ=lɑ səgɭjɑ kəwit ̪ɑ wɑts-ɑw-jɑ-t ̪ əsə wɑʈ-t ̪-ə. Mira=dat all.f.pl poem.f.pl read-sbjv-f.pl such feel-ipfv-3n.sg (1) 'Mira feels like reading all poems.' (feel > all) (2) #'All poems are such that Mira feels like reading them. ' (all > feel) In (44.a), Mira desires that she should read all the poems, say in a book. This reading is also available in (44.b). This is the narrow scope of the quantifier 'all' in the embedded object DP. However, in (44.a), the quantifier can also scope over the matrix verb 'feel', rendering the reading that 'all the poems (say in a book) are so nice that Mira desires to read them'. Assuming that the matrix scope of the quantifier is obtained by the movement of a quantifier DP over the matrix verb, such a movement of the object seems possible only in the LDA construction, but not in the non-LDA construction. Chandra (2007) uses this effect to demonstrate the movement of the embedded object in LDA. I believe that the embedded object can move over the matrix verb in LDA construction because the embedded clause in it is restructured and is not a phase. On the other hand, the wide scope of the quantifier is not possible in the non-LDA construction because the movement of the embedded object is blocked by the CP phase-boundary or by the finiteness of the embedded subjunctive clause. The ambiguity of the scope in (44.a) also indicates that in Marathi, the movement of the embedded object is optional and is not necessary for LDA.
In general, a definite embedded object can also be easily extracted out of the subjunctive clause in the LDA context, but not in the non-LDA construction as shown in (45) These examples suggest that there is a general constraint on the extraction of the embedded object out of the subjunctive clause in the non-LDA construction. Such a ban does not apply to the restructured embedded clause in the LDA construction. The object-extraction and the wide scope of the quantifier thus correlate with the restructured nature of the embedded subjunctive.

Absence of the double-agreement morphemes
Recall the discussion on the double-agreement phenomenon in Marathi. It is seen that the double-agreement morphemes -s and -t occur on the leftmost-usually a tense-carryingverbal element in a finite clause (cf. section 2). If the subjunctive clause in the LDA construction lacks NegP, TP, and arguably the entire CP domain, it is also predicated that the embedded verb cannot participate in the 'double-agreement' or the 'second-slot agreement'. In contrast, no such restriction is predicted on the embedded subjunctive verb in the non-LDA construction. These predictions are indeed borne out, lending support to the structural contrast between the two subjunctive clauses. ues the features on the probing head. Finding the goal is subject to minimality, thus the phi-features of the goal closest to the probe are used to value the features on the probe. I assume, following Preminger (2009;2014), that failing to find a goal does not lead the derivation to be crashed, instead the default agreement morphology shows up on the probing head.
I also assume that several functional heads in a Marathi clause, such as Aspect heads, Mood 0 , T 0 etc. carry phi-probes on them. I assume this as different aspectual participles and tense auxiliaries in Marathi have different agreement properties. This implies that several functional heads probe independently, but minimality results in them all agreeing with the same goal. I further assume that, because a bare verb does not show any agreement morphology, v 0 carries no probe. This is contrary to what is prevalent in the current literature. However, nothing crucially depends on it in my proposal here. If v 0 probes, the union of v 0 with higher functional heads like Asp 0 and T 0 in syntax will result in exhibiting a specific morphological agreement pattern on the inflected verb. The actual mechanism of morphological manifestation is not crucial for my proposal and therefore I ignore it here. For simplicity I also defer discussing the mechanism involved in double or second agreement. Let us briefly see how agreement takes place in a transitive clause in Marathi.
(48) a. mirɑ ɑmbe k h ɑ-t ̪ hot ̪-i. Mira.nom mango.m.pl eat-ipfv be.pst-3f.sg 'Mira was eating mangoes.' b. mirɑ=ne ɑmbe k h ɑ-ll-e hot ̪-e. Mira=erg mango.m.pl eat-pfv-m.pl be.pst-m.pl 'Mira had eaten mangoes.' The structure of the sentences in (48)  In (49.a), the Asp 0 hosts imperfective aspectual morpheme. When followed by a tense auxiliary, the imperfective participial form of the verb shows optional agreement in Marathi. In particular, in the presence of the past tense auxiliary, it does not show agreement. While I leave an account of optionality to future investigation, it would suffice here to assume that the imperfective aspectual head does not carry a phi-probe. Accordingly, no agreement morphology shows up on the imperfective participial verb. The T 0 carries a phi-probe and searches its c-command domain for a goal. It locates the external argument of the vP, which carries uninterpretable nominative case feature and is accessible for agreement. Thus, we get the standard local subject-agreement. 16 In (49.b) on the other hand, the main verb carries perfective aspect which always exhibits agreement in Marathi. Given that Marathi is a split ergative language, the external argument of a transitive vP carries ergative case feature which makes it inaccessible for agreement. The phi-probe on Asp 0 therefore probes further to locate the internal argument, here the DP 'mangoes'. Notice that the vP constitutes a phase. It is possible that the internal argument of the vP moves to the edge of the vP to secure accusative case. It is therefore accessible for the agree probe on Asp 0 and T 0 .
The T 0 , in a similar manner, probes to find the internal argument 'mangoes'. In case the internal argument is inaccessible due to an overt case on it, both Asp 0 and T 0 would fail to find a suitable goal. In that case, the default agreement morphology would be manifested on both.

Agreement mechanism with no LDA
Agreement in the non-LDA construction is essentially the same as in any monoclausal construction such as (48)(49). In each CP-matrix as well as embedded-local agreement takes place. First, the embedded Mood 0 probes. But as the external argument pro in the vP carries ergative case feature due to the subjunctive mood (recall section 2), it is inaccessible for agreement. The Mood 0 probes further to locate the internal argument, in this case, 'mangoes', which has possibly moved to the edge of the vP (movement not shown in the tree). The DP 'mangoes' would value the phi-features on the Mood 0 . This is shown by the Agreeprobing (1). The T 0 selecting MoodP (subjunctive) is always non-overt in Marathi. Thus, one could either say that the 'unrealized-future' tense in Marathi trivially probes to locate the object as a goal or that it does not carry any phi-probe at all. I propose that the complementizer əsə has valued phi-features 3n.sg. These features project to the CP. Thus, the CP headed by əsə can potentially serve as a goal (see also Davison 1991).
Both the Asp 0 and the T 0 in the matrix clause act as probes. But as the external argument 'Mira' carries dative case feature, it is inaccessible for agreement. They both probe further. Since the matrix verb is a dative-subject predicate, I assume that the matrix VP lacks the vP layer and hence does not constitute a phase in syntax. As a result the internal clausal argument of the matrix verb, the CP, is accessible to Asp 0 and T 0 . They agree with the CP, as shown by the probes (2) and (3). Since the CP is a phase in syntax, embedded arguments are not accessible for the matrix probes on Asp 0 and T 0 . Further, minimality implies they agree with the CP and not any goal inside the CP. According to the this analysis, the 3n.sg features on the matrix verb are not the manifestation of default agreement, but the result of local agreement with the CP argument with 3n.sg features.
Alternatively, it can be assumed that the complementizer əsə has no phi-features at all. Thus, the CP cannot serve as a goal. It is however, a phase, preventing the matrix clause probes from accessing embedded arguments. In that case, the matrix Asp 0 and T 0 fail to find a goal and the default agreement morphology manifests on them. 17 While this proposal is quite feasible too, I choose to assume that the complementizer əsə has fixed phi-features. This assumption has potential to capture the association between this complementizer and other instances of əsə in the language which typically participate in agreement (cf. (27) in which əsə occurs in a relative clause). It also opens a possibility of exploring differences between əsə and other complementizers in Marathi.

Agreement mechanism leading to LDA
Now, let us look at how agreement takes place in long distance manner: The Agree mechanism in the LDA construction (3) I assume that the PRO argument is invisible for agreement (Davison 1991). Thus, the embedded Mood 0 cannot agree with the external argument PRO, and needs to probe further. It locates the internal argument (possibly moved to the edge of vP) as a goal. This is shown by probe (1). The linker -s-carries a phi-probe, but it too cannot find PRO as a goal. Just like Mood 0 , the linker finds the embedded object as a goal, as shown by probe (2). I propose that these features percolate to LinkP which can now serve as a goal for other probes.
Like in the case of the non-LDA construction, the matrix Asp 0 and T 0 cannot access the dative-subject of the matrix clause. They probe further as shown by probes (3) and (4) respectively. They locate the internal argument of the matrix verb-LinkP-as a goal. Since the LinkP is not a phase in syntax, they can also access the embedded object directly. But minimality implies that they agree with LinkP rather than any argument embedded inside it. Since the LinkP has the same features as the embedded object, the matrix probes Asp 0 and T 0 also have the same features. This yields the effect of LDA in which the matrix verb appears to agree with the embedded object. In other words, LDA is derived through a series of local agreements. This proposal is essentially in line with the cyclic-agree analyses of Davison (1991);Butt (1995) and Legate (2005), but with crucial differences from all of these approaches in the implementation of the mechanism of agreement.
Notice that there is another possible alternative proposal in which the matrix Asp 0 and T 0 probes directly agree with the embedded object. Since the LinkP is a restructured clause and does not have a phase-boundary, the embedded object is accessible to the matrix probes. The rare instances in Marathi (cf. (28)) in which LDA takes place in the absence of the overt linker -s-in fact supports this analysis. 18 However, these instances can be accounted in my proposal as well. Under my analysis, the absence of the linker and the restructured nature of the embedded clause together imply that the embedded object is the closest accessible goal to the matrix probes. In the alternative proposal of direct agreement of the matrix verb, the linker and the probes in the matrix clause agree independently. Therefore, restructuring of the embedded clause is enough to account for LDA. The agreeing property of the linker -s-turns out to be an interesting but incidental fact. Under my proposal, both restructuring of the embedded clause and the agreeing nature of the linker play an important role in bringing about the patterns of LDA across the subjunctive clause.

Conclusions and open questions
Like its closely related Indo-Aryan languages, Marathi exhibits several constructions involving LDA. However, LDA across the subjunctive embedded clause in Marathi is empirically quite peculiar. I have demonstrated in this paper that this instance of LDA results due to an agreeing clause-linker -s-and a restructured subjunctive clause which it introduces.
This particular instance of LDA has been derived through a series of local agreements. In both the LDA and non-LDA constructions, the matrix verb agrees with its internal argument-the full CP in case of the non-LDA construction and the LinkP in case of LDA. The crucial difference between the two arises from the distinction between clause-introducer elements--s-and əsə. The former agrees with the embedded object, leading to the effect of LDA, while the latter does not agree. I have demonstrated that the form -s-is distinct from the clause-final complementizer əsə, despite its form-similarity with the latter. The form -s-is treated in fact as a low-attached linker. The complementizer əsə, in contrast, is a high-attached linker, a typical complementizer sitting in C 0 . The complementizer əsə has its fixed 3n.sg features and it occurs in a full-fledged CP clause, while the linker -s-agrees and introduces a restructured subjunctive clause. The correlation of the occurrence of the linker -s-, its agreement, LDA and the restructuring properties of the embedded subjunctive clause it introduces, is accounted for in this paper.
The distribution of the clause-linker -s-in Marathi is however, extremely limited. In fact, it occurs in just the one construction-the desirative construction involving LDA being discussed in this paper. As such, the analysis of this construction involving a clause-linker is extremely specific and independent justification for the proposed analysis of -s-as a clause-linker is hard to come by. One could argue that it is a clause-introducing element and hence a standard complementizer in C 0 . I have demonstrated the distinction between the two clause-final complementizers in section 4. Nothing crucially depends on locating it in C head. In any case, there is compelling evidence that the projections between CP and MoodP are absent. It is simpler to assume therefore that the whole CP-layer is also absent. However, analysing -s-as a low-attached linker may open possibilities for treating verb morphology in other multi-verb constructions in Marathi. The non-finite verbs in multiverb constructions in Marathi, bear some morphology including aspectual markers, a case marker, a morpheme -u sometimes labelled as inceptive (cf. Dhongde & Wali 2009) etc. Many of these constructions too exhibit LDA. It would be worth investigating in future if these morphemes are also some kind of clause-linkers.