Case matching and syncretism in ATB-dependencies *

Syncretism has been reported to have the peculiar property of repairing violations of syntactic constraints, e.g. with agreement (Schütze 2003; Bhatt & Walkow 2013) and case matching (Citko 2005; van Craenenbroeck 2012). is paper puts forward the view that in one well-reported instance of syncretism repair of case-matching violations with ATB-movement, this repair follows directly from the nature of ATB movement. We pursue a novel movement-based analysis in which ATB movement involves the actual fusion of two syntactic objects, via intersection of feature sets. As well as deriving the one-to-many relation between llers and gaps in ATB, we show how the ‘repair’ eect of syncretism with case matching violations follows naturally under this approach.


. Introduction
is paper addresses a widely discussed instance of the 'repair e ect' of syncretism with violations of the case matching requirement in so-called Acrosse-Board (ATB) constructions such as ( ); see e.g. Ross ( ), Williams ( ), and de Vries (to appear) for an overview. ) and Citko ( , ), this case matching requirement can be circumvented if the extracted item is syncretic, i.e. has the same morphological form for the cases in question. Whereas the equivalent of 'what' in Polish has di erent forms in genitive and accusative ( ), 'who' is syncretic for genitive and accusative, and subsequently, ATB movement is possible despite the presence of a case mismatch ( ). Taken at face value, this 'repair by syncretism' seems to pose a challenge to a postsyntactic view of morphology such as Distributed Morphology (DM) (Halle & Marantz , Harley & Noyer , Embick & Noyer ), since it seems that a syntactic operation such as ATB movement can be licensed by morpho-phonological form. However, if syntax operates on abstract feature bundles with no morphological reality, as DM assumes, then it is di cult to reconcile this view with the observation that the case matching appears to be sensitive to the form of the elements in question. While some authors have attempted to maintain a DM view in the face of these facts by appealing to underspeci cation (e.g. Citko , Asarina ), we will show that none of these approaches is entirely satisfactory (see section . . ).
In this paper, we argue that the ameliorating e ect of syncretism on case matching violations in ATB dependencies can be made to follow naturally under the view that ATB movement and syncretism have a common denominator, namely intersection of feature sets. e central characteristic of ATB is a one-tomany relation between llers and gaps. From a derivational perspective, we capture the fact that two items seem to 'become one' by assuming that ATB movement involves movement of two items in parallel to an external workspace, where they are intersected to create a new item bearing the shared features of the movees. We argue that this view of ATB movement, and indeed the idea that movement in general (i.e. Internal Merge) must rst proceed via an external workspace, has some independent motivation. More importantly, under this view of ATB movement, the syncretism facts come for free if we assume that syncretic forms result from an underspeci ed exponent realizing a feature shared by both elements. For example, if a language has a syncretism between nominative and accusative, this can be captured by assuming that the syncretic form only realizes a feature such as [− ( )] that is present in both the speci cation of nominative and accusative. Under an intersection approach to ATB, the result of intersecting the feature sets of two DPs, each bearing nominative and accusative, would result in a new item bearing [− ], that is, the feature realized as the syncretic form. If two cases are non-syncretic, then their feature sets do not overlap. Intersection of case features thus results in the empty set, and therefore a crash in the derivation. In this way, the ameliorating e ect of syncretism on case mismatches follows independently from the mechanism of ATB movement and must not be independently stipulated. Since ATB movement involves intersection of feature sets, the only way for DPs bearing di erent cases to successfully undergo ATB movement is if they happen to have a case feature in common that is also realized by a syncretic exponent. e following paper is structured as follows. Section discusses the data surrounding 'repair by syncretism' in more detail and discusses the problems surrounding two previous DM-based approaches to this problem by Citko ( ) and Asarina ( ). Section provides the analysis of ATB based on intersection. In particular, section . discusses previous approaches to ATB movement, section . lays out a novel approach to ATB utilizing intersection of feature sets, section . shows how this approach can derive the syncretism facts in Polish and section . discusses some implications of the present approach for the analysis of Right Node Raising. Finally, section concludes the paper.

. Syncretism and syntax
ere are a number of cases in which syncretism has been reported to have the mysterious e ect of repairing violations of syntactic constraints. ere are a number of examples in which syncretism has an ameliorating e ect on what would otherwise be violations of strict constraints on agreement as well as case matching requirements. is section will discuss a few prominent examples from the literature, concluding with the focus of this paper: syncretism with case matching in ATB constructions.
. . Agreement in Icelandic DAT-NOM constructions e rst case of syncretism repair with agreement is in Icelandic DAT-NOM constructions. Icelandic is known to have quirky subjects in the dative, and these have been shown to behave like genuine subjects with regard to a number of diagnostics (Andrews , Zaenen et al. , Jónsson , Boeckx ). However, in Icelandic agreement targets the nominative DP rather than the dative subject ( ). Furthermore, if there is only a dative subject, as in passives of verbs assigning inherent dative, then agreement is default ( ) ( hjálpað. helped ' e girls were helped. ' (Sigurðsson ) However, there is a particular restriction on DAT-NOM constructions, namely that agreement with a non rd person nominative DP is ungrammatical (e.g. However, Bhatt & Walkow also show that medial conjuncts in coordinations of more than two DPs do not matter for this matching requirement. In general, medial conjuncts seem to be inaccessible as they cannot be targeted for agreement in South Slavic, for example (see Marušič et al. , Murphy & Puškar ).
In ( ), the rst conjunct is masculine, whereas the second is feminine and, as a result, this violates the matching requirement and no agreement form is possible. Furthermore, if conjuncts share the same gender but di er in number, the matching requirement is also violated and ungrammaticality results: Furthermore, as can be seen in the paradigm in ( ), there is a syncretism between masculine and feminine in the present singular, but not in the past singular forms. Bhatt & Walkow ( : f.) show that in 'Right-Node-Raising' constructions such as ( ), where the same matching requirements apply, mismatching gender on the subjects in each conjunct is only tolerated if the form is the syncretic present masculine singular. . . Case matching in free relatives As well as agreement, matching e ects for case have also been widely discussed in the literature. For example, free relatives in German have been shown to exhibit a case matching requirement with both the verb in the free relative and the host clause, see e.g. Gross & van Riemsdijk ( ), van Riemsdijk ( ) and Himmelreich (this volume). For example, in ( a) the wh-phrase in the free relative clause is unambiguously nominative. Since both the free relative and the host clause contain forms of the copula sein, which requires nominative, no problem arises. However, if the verb in the free relative assigns accusative as in ( b), then a mismatch arises between the verb scha en, which requires accusative, and the copula in the matrix clause, which requires a subject bearing nominative case. us, it seems that con icting case requirements imposed on elements in one-to-many relations such as free relatives (the same holds for ellipsis and ATB below) can be satis ed only if the cases in question are syncretic.
. . Case matching in sluicing A further example of 'repair by syncretism' with case matching comes from the domain of ellipsis. In sluicing, there is a general case matching requirement between the antecedent and the remnant. In an example such as ( ), somebody is the antecedent and who is the remnant.
( ) Somebody just arrived, but I don't know who ⟨just arrived⟩ In languages with somewhat richer case morphology, the case on the 'sluiced' wh-phrase has to match the antecedent. e classic example from German is given in ( ⟨ er he t schmeicheln atter will wants ⟩ 'He wants to atter somebody but we don't know who. ' (Ross ) e fact that there is a case matching requirement with the case assigned by the verb in the antecedent clause provides strong evidence for fully-edged elided syntactic structure in the ellipsis site (Merchant ). Furthermore, it is known that languages such as German which do not allow preposition stranding in ordinary wh-questions, also do not allow it in sluicing. However, van Craenenbroeck ( ) shows that, in certain cases, it is marginally possible An interesting account of these facts suggested by van Craenenbroeck ( ) rests on the assumption that sluices with omitted prepositions do not involve an isomorphic ellipsis site, but rather a cle such as 'who ⟨it is⟩' (see e.g. Szczegielniak , Rodrigues et al. , Nykiel , and van Craenenbroeck and Barros et al. for general discussion). Since the pivot of a cle has to bear nominative, there are con icting requirements imposed on the case of the sluice: On the one hand it has to match the case assigned by the antecedent, on the other, it has to be nominative in order to be compatible with the cle in the ellipsis site (but cf. Elliott & Murphy ). e reason why ( ) is possible, and ( ) is not, can be attributed to the fact that welche is syncretic for both the cases assigned in the antecedent clause (accusative) and in the cle in the ellipsis site (nominative), whereas the genitive form is not. e e ect of syncretism in licensing preposition omission can also be seen with was in ( ), which is syncretic for nominative and accusative as was already shown in the free relative examples ( ) and ( b).

. . Case matching in ATB movement
We now turn to case matching e ects with ATB movement, which will be the focus of the remainder of this paper. As was already brie y mentioned in section , in languages with rich case morphology, ATB constructions are characterized by an asymmetric dependency between one ller and two gaps. ere are various restrictions on what kind of gaps are possible in these constructions (see section . ), one of the more interesting ones being case matching. In languages with rich case morphology, the case assigned by the verb to each of the 'gaps' has to match. For example in Polish, the verbs widzieć 'see' and lubić 'like' both assign accusative and ATB movement is licensed ( ).
Janek However, since there is no syncretism between accusative and genitive, a mismatch between the two cases is ungrammatical: Janek In addition, Franks ( ) discusses case mismatches in relative clauses in Russian. In ( ), the relative pronoun kotoroj is syncretic for instrumental and dative, meaning that case matching is satis ed. A similar e ect is reported for German by te Velde ( ) (with an example that he attributes to van Oirsouw ) ( ) and Blümel ( ) ( ). e de nite determiner in German is not syncretic for nominative and accusative (dieser vs. diesen) and is therefore impossible in ATB con gurations. On the other hand, the form of the bare noun is invariant in all cases and therefore ( b) is reported to be grammatical by te Velde ( ).
( ) a. *Dieser this. A similar repair e ect in German is also discussed by Ott ( ) for cases of so-called 'split topicalization' as in ( ) (cf. Fanselow & Ćavar ). Whereas is does not seem to be the case for all speakers, as noted by te Velde himself. One plausible reason for this is that the example in ( b) violates the parallelism constraint on ATB movement proposed by Franks ( , ) stating that ATB movement must take place from somehow parallel structural positions (also see Kasai , Citko ). Here, the movement originates from an object position and a subject position, see section . . for further discussion. the word for 'women' is syncretic in dative and accusative (Frauen), 'men' is not (Männern vs. Männer). Accordingly, only the syncretic form is possible in split topicalization ( . . Interim summary

Käse
We have seen that a number of languages impose matching restrictions on items in certain constructions. In particular, there are case matching e ects that arise with 'sharing constructions' in which there is a one-to-many relation between llers and gaps. On an intuitive level, it seems that what look like bona de syntactic constraints are sensitive to the morpho-phonological form of linguistic objects. Taken at face value, the existence of 'repair by syncretism' would seem to be incompatible with postsyntactic 'late insertion' approaches to morphology, e.g. Distributed Morphology (Halle & Marantz , Harley & Noyer , Embick & Noyer , Nevins ). Proponents of this view assume that syntax operates on abstract feature bundles that do not contain any morpho-phonological information. Consequently, if matching violations can be overridden by paradigmatic identity of distinct cases, then this would seem to pose a serious challenge to this view. On the other hand, one could claim that the syncretism facts indicate that case matching should be a processing or PF constraint, rather than a syntactic one (cf. Smits , Vicente ). However, implementing a matching restriction in this module of the grammar would entail PF (or the parser) having access to syntax-speci c information about the case-assigning properties of individual verbs. is seems to be undesirable if we want to maintain a strictly modular view of grammar. As a result, we seem to be faced with the problem of 'domain leakage' , that is, whichever module of grammar case matching is implemented in, it will require access to information ordinarily reserved for a di erent module.
In what follows, we argue that this is not necessarily the case under the view that both the mechanism for ATB movement and the approach to syncretism share a common property; non-empty intersection of feature sets. In the following section, we propose a new approach to ATB that can explain the syncretism facts while still remaining compatible with a DM view of morphology.

. An intersection approach to ATB constructions
In this section, we present a new take on ATB constructions in which the one-tomany relation between llers and gaps is derived by an intersection operation that creates a single item from those originating in the gaps. It will be shown how this can directly derive the link between syncretism and ATB movement under the assumption that syncretism is derived by means of underspeci cation. First, section . discusses the main approaches to ATB in the literature and how these struggle to capture 'repair by syncretism' in a satisfactory way. Section . will lay out some of the core assumptions required for the analysis to follow. e following section . illustrates how an intersection-based approach to ATB can explain why case matching violations can only be repaired by syncretic forms and section . discusses some implications for Right Node Raising.

. . Previous approaches to ATB
A number of di erent theories of ATB movement have been proposed in the literature. Broadly speaking, they fall into one of two camps: ose that assume that there is 'extraction' from both conjuncts in parallel, what we might call 'symmetric approaches' , and those that assume that genuine extraction only takes place from one conjunct and the other gap is not related to movement ('asymmetric approaches'). Asymmetric approaches derive the second gap in an ATB structure either via a parasitic gap, sideward movement or ellipsis. Each of these approaches will be discussed in turn, considering the extent to which they can account for the syncretism facts. Subsequently, we will do the same for symmetric approaches which either assume genuine movement from both conjuncts or a multidominant structure.
. . . Parasitic gaps e rst kind of asymmetric approach to ATB assumes that extraction only takes place from the rst gap (e.g. Munn , , , Franks , Reich ), and the second gap contains a parasitic gap derived by empty operator movement (following the analysis of parasitic gaps in Chomsky ): Some motivation for this comes from the observation that certain reconstruction phenomena seem to behave asymmetrically, that is, they seem to only be able to reconstruct into the rst conjunct.
In terms of deriving syncretism, one could appeal to the fact that it has sometimes been argued that parasitic gaps also exhibit case matching e ects similar to what we nd in ATB. For example, Bayer ( ) argues that parasitic gap constructions in German exhibit case matching (Huybregts & van Riemsdijk , Kathol , Himmelreich this volume). In ( ), the parasitic gap is assigned dative by the verb anbieten 'o er' , whereas the real gap is assigned genitive by entsinnen 'remember' . ere seems to be the familiar case matching requirement ( ) that is alleviated by syncretism ( ). However, this is only true for some diagnostics (Principle A, Principle C and Weak Crossover). Other diagnostics such as Strong Crossover, variable binding, idiom reconstruction and scope reconstruction behave symmetrically (see Citko , Salzmann a,b for discussion). is seems to indicate that diagnostics that seem to behave asymmetrically are probably sensitive to e ects of linear proximity.
However, the idea that case matching in ATB is related to parasitic gaps is undermined by the fact that not all languages show case matching e ects with parasitic gaps, as also discussed by Himmelreich (this volume). Bondaruk ( , ) shows that Polish, the language with the most widely discussed examples of case matching in ATB, does in fact not seem to impose the same case matching requirement on parasitic gaps. In ( ), the form którą is unambiguously accusative and not syncretic for genitive. Nevertheless, a mismatch between the real gap and the parasitic gap is tolerated, in contrast to ATB constructions. If the explanation for case matching in ATB constructions came from the fact that ATB involves parasitic gaps, then this di erence in Polish would be entirely unexpected. Furthermore, there are a number of other more fundamental asymmetries across languages between ATB and parasitic gaps, in particular the much more restricted nature of parasitic gaps cross-linguistically (see Salzmann a for relevant discussion).

. . . Sideward movement
A closely-related approach involves the application of Nunes' ( , ) Sideward Movement operation to ATB (Hornstein & Nunes ). In this approach, the ller in the ATB con guration undergoes 'interarboreal' movement (i.e. between workspaces; cf. Bobaljik & Brown ). In the derivation of ATB, the moved item originates in the second clause of the conjunction, which is built in its own workspace ( a). It then undergoes sideward movement to the workspace in which the rst conjunct is built, where it is merged as the object of like ( b). At a later step, the vPs form a conjunct (now in the same workspace) ( c). Finally, the wh-phrase in the rst conjunct is extracted to SpecCP ( d).
is approach is therefore not entirely asymmetric since, in a sense, extraction does take place from both conjuncts, but crucially movement to SpecCP proceeds only from the rst conjunct and is therefore asymmetric. ). e case matching requirement could be straightforwardly captured by stipulating that only identical cases can be stacked, but it is unclear how syncretism could be invoked as a repair without opening the door to pre-syntactic morphology (also see Salzmann a: , fn. for discussion).

. . . Ellipsis
A di erent kind of asymmetric approach derives one of the ATB gaps via ellipsis (Ha , Salzmann a,b). In Ha's ( ) approach, it is the gap in the rst conjunct that is derived by ellipsis ( a), whereas Salzmann ( a,b) assumes that it is the second one ( b only Salzmann ( a) predicts asymmetric reconstruction in the rst conjunct). e ellipsis analysis, as all asymmetric approaches, faces the challenge that ATB has been argued to require a 'single identity reading' , which seems to implicate a movement gap in each conjunct (see e.g. Citko : , but cf. Munn , Salzmann a: , fn. ). However, instead, we will focus on the question of 'repair by syncretism' . Salzmann ( a: , fn. ) claims that 'once ellipsis is involved and if morphological mismatches are tolerated, one may expect case matches in ATB' . Indeed, one central characteristic of ellipsis is that it is known to tolerate form mismatches of various kinds (see e.g. Fiengo & May , Merchant ). While morphological mismatches under ellipsis provide a potentially interesting account of exceptions to a case matching requirement in ATB, it seems that an ellipsis-based account predicts that there should not be a case matching requirement at all. Consider example ( a), repeated below. Now, we turn to the symmetric approaches that assume that each of the ATB gaps is directly related to the ller. One particular approach that has Salzmann ( a: , fn. ) conjectures that the empirical situation surrounding 'repair by syncretism' might be more complicated, citing some inconsistency in Citko's reported judgements. Nevertheless, the syncretism repair facts for Polish seem to be relatively robust going back to Borsley ( ). Furthermore, the supposedly controversial case (an accusative/dative mismatch), which we discuss as example ( ), seems to conform to our expections in being ungrammatical. In general, if it is the phonological form, rather than features, that actually matters for mismatches under ellipsis, it seems that the ellipsis approach would be better o claiming that no case mismatches are tolerated under ellipsis and then only phonologically matching forms (identical or syncretic cases) would be correctly predicted to be possible in ATB. On the other hand, this would imply pre-syntactic morphology for Salzmann's ( a) syntactic implementation of ellipsis. For Ha ( ), the problem would be that RNR has been shown to feed ATB movement out of islands (Bachrach & Katzir : f.) and should therefore probably also be situated in the syntax. gained much traction in recent years is the multidominance approach to ATB (Citko , , Gračanin-Yüksek , , Bachrach & Katzir ). is approach assumes that the ller is related to each gap, however this is not derived by movement. Instead, a multidominant view of syntax is adopted in which an element can be in more than one position simultaneously. In an ATB construction, the wh-phrase is associated with both gaps and its derived position in SpecCP, however it is only pronounced in one of these positions ( ).
what is approach has the direct advantage that it can derive 'single identity readings' of ATB, that is, it is only possible to give a single individual answer, rather than a pair-list answer, to an ATB question: For other arguments in favour of a multidominance approach to ATB, see Citko ( , ). However, a problematic data point that is not o en discussed in conjunction with the multidominance approach is the fact that, in some languages, ATB movement can have resumptive pronouns in the gaps. For example in Akan (Niger Congo: Ghana),Ā-movement of animate DPs triggers obligatory resumption, also in ATB wh-questions (Saah , Korsah & Murphy )  ese data are problematic for multidominance accounts of ATB since, as is clear in ( ), they assume that the wh-phrase is syntactically present in both of the gaps. Whereas the multidominance account straightforwardly derives the fact that ATB movement leaves gaps, it does not seem to be possible to account for resumptive pronouns if the ller is also structurally present in its base positions.
Turning now to 'repair by syncretism' , Citko ( : .) explicitly addresses the question of how her multidominance approach can derive the fact that syncretism can repair case matching violations. Citko puts forward an explanation based on underspeci cation couched in the framework of Martin Salzmann (p.c.) suggests that this might not necessarily be fatal for a 'big DP' approach to resumption, in which the DP starts out in the same phrase as the resumption pronoun and is extracted (e.g. [DP DP [ D' D resumptive ]]) (e.g. Boeckx ). If the ATB-moved item multiply dominated the speci er of both 'big DPs' , then this might work. However, if one no longer has a movement approach, in which the resumptive pronoun is stranded, then it is unclear what the status of the 'big DP' is in such an analysis. A perennial problem is that these complex elements never occur overtly, so it is unclear what their motivation would be in a multidominance approach.
Distributed Morphology. She assumes that 'the lexicon contains a single whform, underspeci ed in such a way that it is compatible with both genitive and accusative' (Citko : ). Consider again example ( ), repeated below, where syncretic forms license a mismatch in case. Citko ( : ) then states that 'the lexicon contains a single form that is compatible with both accusative and genitive case feature by virtue of underspeci cation' (kogo) and this can be inserted into the terminal. e ungrammaticality of case mismatches in the inanimate wh-series where there is no syncretism ( ) (repeated below) is explained by the assumption that 'there is no single lexical item that can be inserted into this slot without a feature clash, [. . . ]  ] speci cations, that is, non-ATB environments where the wh-phrase is assigned only one case, following the Subset Principle (see ( ) below). e second and third options would incorrectly restrict the distribution of kogo to either genitive or accusative contexts respectively, but do not capture the fact that the forms are syncretic. Furthermore, regarding the illicit case mismatches without syncretism in ( ), Citko attributes the ungrammaticality to the fact that 'there is no single lexical item that can be inserted into this slot without a feature clash' ( : ). However, this is not a standard approach in DM, where Vocabulary Insertion relies on underspeci cation and the Subset Principle to regulate competition between exponents ( ).
( ) Subset Principle (Halle ; our emphasis) e phonological exponent of a Vocabulary Item is inserted into a morpheme in the terminal string if the item matches all or a subset of the grammatical features speci ed in the terminal morpheme. Insertion does not take place if the Vocabulary Item contains features not present in the morpheme. Where several Vocabulary Items meet the conditions for insertion, the item matching the greatest number of features speci ed in the terminal morpheme must be chosen.
us, if we have a terminal corresponding to an inanimate wh-phrase assigned both genitive and accusative, it is not true that we have a feature clash. Instead, the Subset Principle predicts that we should be able to insert either exponent since both ful l the Subset Principle and are equally speci c ( ).
One would be forced to have multiple entries for kogo, which would reduce the syncretism here to accidental homophony, see Asarina ( Alternatively, one could impose a ban against con icting features on a terminal itself which could trigger a repair that deletes both con icting features. is, however, seems implausible since deletion of only one of the con icting features would be su cient here. omas ( ) actually pursues this alternative strategy. She proposes a rule of Case Uni cation de ned in (i).  For reasons that are still poorly understood (but see section . . ), this particular kind of extraction can circumvent the Coordinate Structure Constraint, stating that extraction from a single conjunct is not possible (Ross , Grosu ). Furthermore, it is unclear how moving two items can result in a single ller (cf. Blümel , Weisser : ). is has typically been handled by construction speci c rules (Ross , Williams ), however this is something that the analysis to follow will explain. Since this approach is also symmetric, it shares with multidominance analyses the virtue of being able to explain single identity readings in ATB constructions.
Regarding the question of 'repair by syncretism' , current parallel extraction approaches have to more or less stipulate the case matching requirement in one way or another (e.g. Dyła : ). For example, Blümel ( ) simply states the matching requirement as part of a 'condition on chain formation' ( b):

( )
A movement chain must a. comprise non-distinct members (i.e. they must be featurally identical) b. be headed by a syntactic object which receives an exponent compatible with all lower chain members. (Blümel : ) Since chain formation is most plausibly syntactic in nature, Blümel's approach is clearly incompatible with a postsyntactic view of morphology (despite his claims to the contrary; Blümel : ). e new approach that we pursue is also a symmetric extraction approach. However, the crucial di erence to existing symmetric extraction approaches is that it can explain both the one-to-many relation between llers and gaps and the syncretism repair property.

. . eoretical assumptions
In the following, we propose an account of ATB dependencies that utilizes an intersection of the ATB-moved elements to create a single ller. In order to derive this, we will introduce new assumptions, or make some already existing ideas more explicit. e motivation for each of these assumptions will be discussed in turn.

. . . Movement via an external workspace
e existence of complex speci ers necessitates more than one workspace in a syntactic derivation. In ( ), the complex subject the man with the hat undergoes External Merge with v ′ as its speci er, however, this complex DP must have been built somewhere other than the current workspace, i.e. from another numeration, see e. Workspace is existence of an additional workspace has been exploited by Nunes ( , ) who assumes that it is possible for elements to undergo 'sideward' movement to another workspace of the local tree. Furthermore, there has been an e ort to dispense with a separate operation for movement, and instead view movement as a kind of Merge (e.g. Chomsky , Starke ). In particular, movement is assumed to be a variant of External Merge, with the di erence being whether the target of the operation is included in the same workspace (Collins & Stabler : ). Both operations have in common that they obey Chomsky's ( ) Extension Condition, stating that Merge must apply at the root node (i.e. extend the tree). We claim that one natural way to capture this is to actually decompose Internal Merge into two steps: Sideward Movement & External Merge. In the Copy of eory of Movement, it seems that Internal Merge is already o en (tacitly) assumed to consist of two steps: the rst step creates a copy, and the second step involves External Merge of this copy at the root (this is made explicit in Putnam , Stroik and Stroik & Putnam : ). One question that is not o en explicitly addressed is where exactly this moving copy is generated and stored. It seems desirable to assume that External Merge always accesses an item in a separate workspace. us, we assume that all instances of Internal Merge proceed in a two-step fashion as in ( ): 'sideward' movement to an external workspace (creating a copy) followed by External Merge at the root node.

. Parallel movement
Another assumption we make is that ATB involves parallel movement, that is, simultaneous movement from two distinct positions to a single landing site. It has been long noted that ATB must take place from 'parallel positions' (Williams , Franks , , Kasai , Citko ). For example, ATB extraction from a subject and object position is not possible: ( ) *I know a man who [Bill saw t ] and [t likes Mary]. (Williams : ) Furthermore, Franks ( , ) discusses ungrammatical examples of ATB movement from Russian, in which the case matching requirement is met, but the extraction is from di erent structural positions, and therefore illict. Given the Strict Cycle Condition (SSC) (Chomsky ), we can derive the parallelism requirement by assuming that ATB involves simultaneous movement to an external workspace as in ( ).
Note that the idea of parallel movement chains already exists in the literature, but in fact in the opposite sense. Chomsky ( ) proposes that it is possible for a wh-phrase to move in parallel to SpecTP and SpecCP (also see Bošković ). is idea is motivated by the fact that traces of wh-movement in Icelandic trigger defective intervention (i) in a way that A-movement does not (ii) (Holmberg & Hróardóttir ). e assumption is that the wh-phrase moves in parallel to SpecTP and SpecCP and its initial trace therefore counts as an A/A-trace and can trigger intervention. Crucially, this approach to parallel chains has one element moving to two positions simultaneously, rather than two elements undergoing movement to a single position. For reasons that will be made clear in the following section, parallel movement to an external workspace results in a single item, which is subsequently remerged into the structure. Assuming that the SSC holds across conjuncts (i.e. they are built in parallel), extraction from the same position (i.e. subject or object) will be possible, whereas extraction from di erent positions will result in a SSC violation.
Parallel movement would therefore seem to be restricted to coordinate structures. One possible reason for this could be that it is a Last Resort option to circumvent the Coordinate Structure Constraint (CSC, Ross ) that militates against extraction from a single conjunct. If we take a representational In ( ), movement takes place relatively late, however it is conceivable that this sideward movement takes place at a much earlier point of the derivation (e.g. vP) for reasons of phasehood and/or cyclicity.
While ATB movement is overwhelmingly found in coordinate structures, Vicente ( , to appear) claims that there are case of ATB extraction outside of coordination. He o ers some examples, whose status seems unclear. e following example from Munn ( ) is illustrative.
(i) Who did you send pictures of to ?
First, it is unclear whether these are parasitic gaps (but Vicente argues against this position due to the apparent availability of sloppy readings) and second, it is unclear to which extent these kinds of examples re ect a productive ATB strategy outside of coordinate structures. e examples in question do not seem particularly well-formed to us and overwhelmingly favour strict, single identity readings.

view of the CSC as in ( ) (see Mayr & Schmitt
: , Weisser : f., but cf. Kato ), then no extraction can take place from a single conjunct at any point of the derivation. Crucially, by moving in parallel we avoid both of the con gurations banned by ( ): Furthermore, this general approach can help to make sense of an interesting restriction on ATB movement reported by Kasai ( ), Citko ( , ). In multiple wh-fronting languages such as Polish, it is not possible to combine ATB extraction and multiple wh-fronting: ( ) a. *Kogo who.
[ TP Jan Jan 'What did Jan buy for whom and Peter send to whom?' (Citko : ) As a result, even in languages with the option of multiple wh-fronting, extraction from a coordinate structure must involve parallel movement. How exactly this parallel movement results in a single ller is discussed in the following section.

. . . Feature set intersection
In the previous section, we established our assumption that ATB movement proceeds in parallel to an external workspace, however, how does ATB extraction result in a single ller if two elements are moved simultaneously? We suggest that parallel movement to an external workspace results in set intersection of the feature sets of the moving elements. Recall that parallel sideward movement is viewed as a Last Resort solution to circumvent the CSC and is therefore not the norm. We assume that the external workspace has a restriction that it can hold a single moving item. As a result of this restriction, something must happen if two items move in parallel. It seems we have two options: (i) intersection of feature sets, (ii) uni cation of feature sets. e latter option would run into the same problems shown for Citko's DM approach in section . . , since the wh-phrase would bear both case values in a uni cation approach (also see footnote ). us, we assume that if more than a single item is moved (via the external workspace), intersection of the feature sets of these items must take place. In a simple example of ATB movement in ( ), both wh-phrases (with matching feature sets) are intersected in the external workspace, resulting in a single wh-phrase bearing the same features as the two moved items. is single element then re-enters the structure at the the landing site for ATB movement.
( ) CP In this way, we can derive the asymmetric relation between llers and gaps that is a hallmark of ATB dependencies. Furthermore, if the feature sets of the items do not intersect for a particular feature, for example animacy features with who ([+anim]) and what ([−anim]), then the value of that feature will be empty and thus result in a crash (given Full Interpretation; Chomsky ). Example ( ) is a somewhat trivial case, in which both of the intersected items have exactly the same features. However, as we show in the next section, this intersection operation has interesting, welcome consequences when cases do not match, but are syncretic.
One might also wonder how it is possible to intersect complex wh-phrases such as which book. One option is that the wh-determiner and NP are intersected separately, but this may not even be necessary if we view complex DPs from a set-theoretic perspective (e.g. Chomsky ). For example, a complex wh-phrase is a set containing two elements: a set containing the features of the determiner and another set containing the features of the NP (i).  Syncretism can then be captured by assuming that syncretic forms are underspeci ed and realize a feature that is present in both contexts. In other words, syncretic forms must have at least one feature in common (the one that the syncretic form realizes), i.e. their contexts' feature sets must overlap. For example, one can see in ( ) that animate wh-phrases in the genitive and the accusative share the feature [+gov]. us, the exponent kogo can be underspeci ed for only [+gov, +anim] and will therefore be inserted in both accusative and genitive animate contexts. We assume the following Vocabulary Items for Polish wh-phrases: Note that the features we use are arbitrary. However, for convenience, we adopt the conventional labels ±subject, ±governed and ±oblique without attributing these any semantic relevance. ey could easily be replaced by ±α, ±β, ±γ. An alternative approach is to use postsyntactic rules (e.g. impoverishment rules) to derive syncretism (see Himmelreich this volume), however, we nd this approach somewhat more ad hoc and are unsure of how it is compatible with 'repair' e ects of syncretism. Although ATB is independently assumed to involve intersection in order to derive the one-to-many relation we observe between llers and gaps, we also see that this will derive the case matching data, in particular, repair by syncretism in the following way: If we try to ATB-move two wh-phrases with mismatching cases, those with an overlapping feature will result in a successful intersection and -provided that there is a suitably underspeci ed VI -will be realized by that VI. We will show this in detail in the following sections.
. . . ATB with matching cases (no syncretism) An example of an ATB dependency with matching cases is given in ( ) where both verbs assign accusative case to an animate wh-phrase.
Janek As we saw in previous sections, both wh-phrases move in parallel via the external workspace. Given the assumption this workspace can only hold one item, both items undergo feature set intersection. Since in this case both items have exactly the same case features the newly formed item is identical to each of the two moving items, that is, it bears a fully speci ed accusative case. is new item is then merged from the external workspace into SpecCP.
At Spell-Out, only one of the four wh-vocabulary items from the inanimate series ( ), namely co, is speci ed for a subset of the wh-phrase's morphosyntactic features and therefore available for insertion. All other VIs are speci ed for at least one feature-value that is not part of the terminal. Hence, the accusative marker co is inserted as expected in accordance with the Subset Principle.  Workspace In other words, the newly formed wh-phrase is unvalued for case. Since there is no other case-assigner in the structure who at this point has not already assigned its case the item remains case-less until spell-out. A DP that does not have case, however, is in con ict with the Case Filter (or whatever ensures that DPs have case, e.g. Full Interpretation; Chomsky ). e derivation therefore crashes at the interface to PF.
. . . ATB with mismatching cases (with syncretism) e interesting case now concerns ATB movement with mismatching cases that happen to be realized by the same (syncretic)  , Gračanin-Yüksek ). Of these approaches, the general consensus seems to be that the evidence against an ATB movement account of RNR is pretty damning (e.g. Abels , Bachrach & Katzir , Larson , Barros & Vicente ). For example, RNR has been shown to display insensitivity to other processes that ordinarily constrain rightward movement, e.g. the Right Roof Constraint (Ross ). Consequently, the present debate focuses on whether the ellipsis or multidominance approach is correct, or even both (Barros & Vicente ). However, there is a potentially new argument in favour of movement, based on what has been shown here. If the current approach is correct that 'repair by syncretism' e ects that arise in 'sharing constructions' such as ATB whmovement cannot be adequately captured by ellipsis or multidominance (cf. sections . . and . . ), then nding such e ects with RNR would constitute an argument in favour of a movement-based approach. Asarina ( ) reports exactly this kind of data for RNR in Russian ( ).
However, see Sabbagh ( ) for some scope data that seem to support a movement-based account. Furthermore, Bachrach & Katzir ( ) make the interesting observation that RNR can feed, i.e. license, ATB movement out of islands. Although they have a rather involved account of this fact based on multidominance and 'delayed Spellout' , the most natural explanation of these facts would be that RNR is syntactic movement that can therefore interact with other kinds of extraction.
Asarina's ( ) approach to dealing with con icting case values on a single terminal di ers from Citko's ( ). When an element with a given feature matrix is assigned a second, di erent value for the already valued case feature, the whole feature matrix is duplicated to accommodate that value. e element then has two feature matrices that di er only in the value for the case feature. As long as both matrices can be spelled out by the same morphological rule (i.e. one that does not make reference to the distinct feature and is thus underspeci ed), the result is grammatical.
However, even though the rule should actually be able to spell out both feature matrices only one exponent exists on the surface. In e ect, this ties insertion of a VI into a terminal's feature matrix to a potential insertion of the same VI in the other feature matrix on the terminal even though that second insertion never actually happens. Roughly paraphrased: A VI may be inserted into a terminal with two feature matrices as long as it remains unclear which of the two it actually realizes. us, this leads back to the additional ban against a feature clash for vocabulary insertion: A VI may only be inserted into a terminal if it is not in con ict with any features on that terminal (even if they are in a di erent feature matrix). Another potential In ( ), 'plate' is not syncretic in the nominative and accusative cases (tarelka vs. tarelku) and therefore the mismatch is ungrammatical. However with 'saucer' , the form bljudce is syncretic for both cases and this licenses a mismatch.
is kind argument has not yet featured particularly in the debate on RNR, however it seems that the relevant ability of the theories at hand to capture 'repair by syncretism' e ects should have a bearing on the question of its correct treatment.

. Conclusion
In this paper, we have shown how one can derive the fact that case matching requirements imposed on ATB constructions by a number of languages can be circumvented by syncretism. On the face of it, these data seem to be incompatible with a postsyntactic view of morphology since morphological form seems to play a role in licensing ATB movement. Whereas existing approaches are forced to simply state 'repair by syncretism' as a fact, or their proposal turns out to be awed, we have shown that an intersection-based approach to ATB can derive the syncretism facts in an elegant way. In this approach, intersection is the operation independently required to derive the one-to-many 'sharing' characteristic of ATB, since it is otherwise puzzling why movement of two items results in a single ller. Once intersection is established as the core mechanism for deriving ATB, the syncretism facts follow naturally (given an underspeci cation approach to syncretism), rather than having to be problem is that if syntactic objects are understood as being just bundles of features duplicating an element's feature matrix is the same as duplicating the actual element itself. stated additionally. In particular, we have shown how the present approach can derive the classic facts of syncretism repair with case matching violations in Polish. Of course, any extension to other languages with similar e ects, such as German or Russian, may entail a di erent case decomposition (as the syncretism will most likely di er), however the basic mechanism will remain the same.
Furthermore, we have argued that only a movement-based account with intersection can adequately derive the ameliorating e ect of syncretism. e most (if not only) worked-out approach in multidominance theories, Citko ( ), is beset with a number of technical problems that become apparent once one tries to implement the proposal in a more explicit way. We have argued that this has potentially interesting consequences for the debate about the correct analysis of RNR. Multidominance is o en invoked as the most likely explanation for this phenomenon, however the existence of syncretism repair facts in these constructions (e.g. in Russian) would seem to undermine this, given the present inadequacy of multidominance approaches in deriving this. Of course, one could enrich these theories with further operations (e.g. intersection for con icting feature sets), but this lacks the elegance of an approach such as the present one, in which the motivation for intersection is independent of syncretism.