1. Introduction
This paper investigates a previously unexplored split in Greek that involves possessor constructions embedded under Ps. In Greek possessor constructions, a genitive possessor, pjanu ‘whose,’ can be separated from a possessum, to podhilato ‘the bike’ (Horrocks and Stavrou Reference Horrocks and Stavrou1987, Mathieu and Sitaridou Reference Mathieu, Sitaridou, Batllori, Hernanz, Picallo and Roca2005, Alexiadou et al. Reference Alexiadou, Haegeman and Stavrou2007, Angelopoulos Reference Angelopoulos2019, Angelopoulos and Michelioudakis Reference Angelopoulos and Michelioudakis2024 i.a.):
I present an asymmetry split within PPs. Although a possessor can never be separated from a possessum embedded under a P (2a), some PPs allow what appears to be the movement of a non-constituent in (2b). In this pattern (hereafter, PP-split), the possessor and the P co-occur in the left periphery to the exclusion of the possessum, which is left stranded postverbally. An important finding of this paper is that PP-splitting is not allowed with all PPs. Instead, its availability depends on the specific interpretation of the PP. For instance, while PPs expressing subject matter allow for splitting, as we saw in (2b), PPs with an evidential interpretation block both standard possessor movement and the formation of a PP-split (3).
Similar splits are found across different languages, particularly within Slavic languages.Footnote 1 These splits have given rise to different analyses, with important implications for our understanding of syntactic structure and the division of labor between syntax and phonology. Syntactic accounts of these splits have been proposed in Franks and Progovac (Reference Franks, Progovac, Fowler, Cooper and Ludwig1994); Kayne (Reference Kayne2002); Bašić (Reference Bašić2008, Reference Bašić and Subotić2009); Abels (Reference Abels2003, Reference Abels2012); and Talić (Reference Talić2019). In contrast, proponents of approaches employing mechanisms like distributed deletion that selectively delete copies at PF, include Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002); Bošković (Reference Bošković, Szajbel-Keck, Burns and Kavitskaya2015); Fanselow and Féry (Reference Fanselow and Féry2006); and Bondarenko and Davis (Reference Bondarenko and Davis2023), among others. Crucially, the syntactic approach finds support in the observation that the interpretation of the PP directly affects its potential for splitting, as it aligns with established findings that PPs occupy different syntactic positions depending on their interpretation (Cinque Reference Cinque2006, Alexiadou and Anagnostopoulou Reference Alexiadou and Anagnostopoulou2007, Schweikert Reference Schweikert2005 i.a.). Distributed deletion, however, faces difficulties with PP-splits, requiring principles beyond those of a purely syntactic analysis to account for their distribution. Moreover, when applied to Greek, it predicts ungrammatical structures. Consequently, a syntactic analysis of Greek PP-splits provides evidence against distributed deletion as a mechanism. To formalize this, this paper argues that there are principles which effectively prohibit distributed deletion in grammar, as summarized in (4). This has far-reaching consequences, calling for a re-examination of other splits in other languages (see Goncharov Reference Goncharov2015). This paper focuses on Greek PP-splits shown in (2b) to demonstrate the issues of distributed deletion and motivate (4).
This paper proceeds as follows. Section 2 demonstrates that PP splits in Greek are allowed only with certain Ps, depending on their interpretation. This section reveals the striking behavior of a particular P: it allows or blocks PP-splits based on its interpretation. Building on these empirical observations, Section 3 proposes an analysis of the cases in which a possessor is separated from the possessum in Greek, extending this analysis to account for both the availability and unavailability of PP-splits. Section 4 discusses the numerous challenges faced by distributed deletion accounts to explain this complex pattern. By demonstrating the inadequacy of this approach, this section strengthens the argument for the constraint proposed in (4). It also discusses the principles behind the prohibition of distributed deletion as a mechanism of natural languages. Section 5 concludes.
2. Possessor movement and PP-splits in Greek
I present a study of possessor movement and PP-splits in Greek. I examine a variety of PPs expressing evidential, temporal, benefactive, source, agent, causer, manner, instrument, locative (either denoting a static location or direction), comitative and subject matter (in the sense of Pesetsky Reference Pesetsky1996). These PPs range in structural complexity, from simple Ps (kata ‘according to’) to complex Ps (simfona me ‘according to’) allowing us to investigate whether structural complexity plays any role in the formation of PP-splits. As we shall see, it does not.
All of the Ps illustrated in the table above can be combined with a possessum that takes accusative case and a dependent possessor that takes genitive case. As shown in (2a), a possessor originating within a PP cannot occur alone in the left periphery of the clause, stranding P and the possessum. This restriction applies to all Ps listed in Table 1, as illustrated in the second row.Footnote 2 To further illustrate, similar to the restriction observed in (2a), possessors originating in a benefactive (5), also cannot appear alone clause initially.
Possessor movement and PP-splitting across PPs
Conversely, PP-splitting (2b) is permitted with certain types of PPs but not with others. Splitting is allowed with agent, causer, benefactive, instrument, source, locative, comitative, subject matter and manner PPs (see 7 for an example with agent PPs). However, evidential and temporal PPs do not permit splitting (see 6 and 8, repeated from 3, respectively, for examples with temporal and evidential PPs).Footnote 3 Beyond the simple Ps discussed so far, such as ja in (2b), there are Ps like simfona me, which differ from simple Ps in that they are morphologically complex. Like plain Ps, complex Ps can express different interpretations. Interestingly, the availability of PP-splitting for such a P depends on its interpretation. While simfona me blocks splitting when used with an evidential meaning (8), it allows splitting when expressing manner (9). This contrast shows that a P’s capacity to license splitting is determined by its interpretation rather than its structural complexity.Footnote 4
Consequently, we have detected two distinct groups of PPs, shown in the pairs in (8) and (9). The group of PPs in (8) is different because it prevents both possessor movement and PP-splitting, as opposed to the group in (9), which only prevents the first. The different behavior of these PPs, and in particular, this of simfona me in (8) and (9) suggests that PP-splitting depends on the interpretation of the PP. Having observed the varying behavior of PPs with different interpretations with respect to PP-splitting, I will now demonstrate how this behavior can be captured under a syntactic analysis.
3. A syntactic analysis of splits
I begin with the analysis of plain cases of possessor extraction and then show how the account extends to PP-splits.
3.1. A syntactic analysis of possessor extraction
To understand why PPs behave differently in PP splitting, we first need to analyze how possessors are separated from a simple DP possessum argument, as shown in (10a), repeated from (1).Footnote 5 This analysis is defended in length in Angelopoulos (Reference Angelopoulos2019) and Angelopoulos and Michelioudakis (Reference Angelopoulos and Michelioudakis2024). Here, I discuss key points of their analysis, and extend it to PPs.
To start with, this analysis is based on the observation that possessors in Greek can appear before the noun, as in (10b). In this position, the possessor and the accusative possessum form a constituent, evidenced by their ability to undergo both short and long-distance movement to the left periphery of the clause, as shown in (10c) with short movement.
This prenominal position, which I label FrontP, is projected in the left periphery of the DP. Similar positions have been identified in various languages, including Gungbe (Aboh Reference Aboh2004), Italian (Cardinaletti Reference Cardinaletti, Alexiadou and Wilder1998, Giusti Reference Giusti and Frascarelli2006), Polish (Siewierska and Uhlírová Reference Siewierska, Uhlírová and Siewierska1998), Turkish (Rijkhoff Reference Rijkhoff and Siewierska1997), and Bangla (Syed Reference Syed and Steindl2015), among many others. The exact function of FrontP, whether it should be analyzed as FocusP (Ntelitheos Reference Ntelitheos2004) or something else (Szendrői Reference Szendrői2010), is not crucial for the current analysis; what matters is that it exists.Footnote
6 The existence of this position accounts for the fact that the accusative possessum can be moved on its own, as in (11a). In this case, the possessor moves to Spec, FrontP, as shown in (11b), which allows the possessum to form part of a DP constituent that can subsequently move. In contrast, if prenominal possessors were located in Spec, DP (Horrocks and Stavrou Reference Horrocks and Stavrou1987), as in (11c), the possessum should not be able to move on its own. This is because intermediate projections like
$ {\mathrm{D}}^{\prime } $
, shown in (11c), are generally not movable.Footnote
7
, Footnote
8
In cases where a possessor is separated from the possessum, as in (10a), I assume that the possessor can never be moved out of the DP to Spec, CP. This restriction stems from the DP acting as a bounding node that blocks extraction out of it to the left periphery (Angelopoulos and Michelioudakis Reference Angelopoulos and Michelioudakis2024). To be more precise, following Aissen and Polian (Reference Aissen and Polian2025, (82)) and Angelopoulos and Michelioudakis (Reference Angelopoulos and Michelioudakis2025, Reference Angelopoulos and Michelioudakis2026), I encode the opacity of the DP-domain as a case of ‘selective opacity’ (Keine Reference Keine2019). In particular, I assume that any phrase in the extended projection of N0, which also includes D and Front, i.e., Front
$ > $
D
$ > $
N, is opaque to a wh-/focus/topic/ or relative probe on a c-commanding C0. In Keine’s terms, any node in the extended projection of N0 is a ‘horizon’ for a wh, relative, topic or focus probe on C0, rendering elements dominated by that node invisible to the probe. In Keine’s notation:
The only licit movement step of the possessor is that shown in (13), where it moves to Spec, FrontP. This is allowed because it constitutes movement within the extended projection of N, and is not motivated by C.Footnote 9
Subsequently, the DP remnant that contains the accusative possessum, the bike, is optionally allowed to undergo movement into a middle field position between the TP and the thematic position of external arguments, that is, the vP. This movement step of the possessum DP resembles regular accusative objects in Modern Greek. These objects can also move to this middle-field position, resulting in VOS word order. In this order, the object moves to a middle-field position – let us call it YP, as in (14) – lower than the verb in T but higher than the external argument which arguably stays in Spec, vP when it appears postverbally (Alexiadou and Anagnostopoulou Reference Alexiadou and Anagnostopoulou2001, Anagnostopoulou Reference Anagnostopoulou2003, Alexiadou Reference Alexiadou2006, Roussou and Tsimpli Reference Roussou and I-M2006, Georgiafentis Reference Georgiafentis2001, Reference Georgiafentis2003 i.a.).
This movement step is compatible with our assumptions about nominal opacity in (12). Selective opacity is probe-specific, being defined with respect to a particular probe. In the present case, the relevant probe is the wh/relative/topic/focus feature on C0. Opacity with respect to this probe does not entail opacity with respect to other features. Movement of the DP to the middle field is therefore expected to be possible under selective opacity, because, as will be discussed in the next section, it is an A-movement step triggered by a different probe, rather than by C bearing a wh-/focus/topic or relative feature.
The derivation below illustrates the three movement steps that underlie cases where the possessor occupies the left periphery. First, the possessor moves to Spec, FrontP. Next, the possessum DP moves to the middle-field, landing in Spec, YP, where objects are moved in VOS orders. Finally, the possessor undergoes further movement to the left periphery within the FrontP remnant, which gives rise to the surface order in (10a).
In this view, the derivations underlying (10a) and (10c) share a key similarity: in both cases, the entire FrontP moves to the left periphery of the clause. The crucial difference lies in whether the accusative possessum moves out of FrontP before this movement. In (10a), the possessum evacuates FrontP, resulting in the possessor being separated from the possessum. Conversely, if the possessum remains within FrontP, it moves together with the possessor to the left periphery. To explain these movements, I adopt the standard assumption that all syntactic movement is feature-driven (Chomsky Reference Chomsky, Hale and Keyser1993).Footnote 10 In (10c), the possessor carries two features: a ‘Front’ feature,Footnote 11 that triggers movement to Spec, FrontP and a ‘Q’ feature, triggering movement to Spec, CP.Footnote 12 Similarly, the possessum must also carry a feature to license its movement to the middle-field. Notably, prenominal possessors in Greek obligatorily have a distinct emphatic interpretation (Ntelitheos Reference Ntelitheos2004), often lacking in unfronted possessors (see 18a vs 18a for a fronted and an unfronted possessor). This suggests that movement to Spec, FrontP (i.e., the prenominal position) is motivated by interpretive reasons. Therefore, I analyze this movement as criterial movement (see discussion in Angelopoulos and Michelioudakis Reference Angelopoulos and Michelioudakis2024 for evidence that Spec, FrontP is a criterial position), akin to movement to Spec, TopicP or Spec, FocusP. Consequently, in (13), the possessor, having undergone criterial movement to Spec, FrontP, is frozen in place, as a result of criterial freezing (see Rizzi Reference Rizzi, Cheng and Corver2006, Reference Rizzi and Panagiotidis2010), and cannot move on its own. To satisfy its Q-feature, it must pied-pipe the entire FrontP to Spec,FocusP.
3.2. Middle-field movement
In this section, I present arguments for the view that plain possessor movement always involves the movement of the accusative possessum to the middle field, and, crucially, that this movement constitutes an A-movement step. Moreover, the target position for these arguments is the same as that of objects in the VOS order. These assumptions are supported by new data from ditransitive constructions, particularly those containing a dative argument. In this applicative frame, the default order is dative preceding accusative (16a), though the reverse order is also possible (16b). A striking fact emerges in (16c) and (16d): possessor sub-extraction is systematically blocked from an accusative argument in ditransitive constructions, irrespective of whether the accusative appears before (16c) or after the dative argument (16d).
If possessor splitting proceeded in a successive cyclic fashion through Spec, DP, as proposed by Horrocks and Stavrou (Reference Horrocks and Stavrou1987), both (16c) and (16d) should be grammatical, contrary to fact. In (16c), the possessor should be able to move through Spec, DP of the accusative argument and then past the dative argument, given that A-bar movement is generally possible across an A-position, much like A-bar movement across a subject. Likewise, in (16d), the possessor should be extractable via Spec, DP in the pre-dative position, unless no movement out of the accusative is possible in this derived position – an expectation consistent with the so-called freezing constraint.
The proposed analysis accounts for these facts more robustly: (16c) is ungrammatical because the accusative object remains in its base position, that is, it has not moved to the middle-field, as evidenced by its surface order after the dative. (16d) is also ungrammatical, despite the accusative object undergoing A-movement to the middle field, because A-movement across a dative object is independently known to be impossible (Anagnostopoulou Reference Anagnostopoulou2003).Footnote 13 Crucially, Anagnostopoulou (Reference Anagnostopoulou2003) has shown that cliticization or Clitic Doubling of the dative argument obviates its intervention, thereby permitting A-movement of the accusative object past it. This suggests that cliticization of the dative should also facilitate possessor extraction from the accusative, as movement to the middle field should now be possible. Indeed, this prediction is borne out (17).Footnote 14
Now a question is whether the possessum undergoes the same type of movement and to the same position as objects when they appear in the VOS order. There are different analyses for the VOS order in Greek. Some of them assume movement to the middle field (Alexiadou and Anagnostopoulou Reference Alexiadou and Anagnostopoulou2001, Roussou and Tsimpli Reference Roussou and I-M2006), whereas others take it that the object moves alongside the verb (Georgiafentis Reference Georgiafentis2001). In support of the former, Alexiadou and Anagnostopoulou (Reference Alexiadou and Anagnostopoulou2001) offers a compelling explanation for obligatory object movement in VOS constructions. They argue that in VOS orders, the subject occupies Spec, vP. Consequently, the object cannot remain in its base position because the VP cannot contain two DPs with unvalued case features at spell-out, requiring its movement. Since this movement step is due to case, it must also be an instance of A-movement, and, as we saw in (16), the accusative possessum also undergoes A-movement. Since both undergo the same type of movement step, it makes sense that objects in the VOS order as well as accusative possessums in plain possessor extraction move to the same position. Nevertheless, the motivation for this movement step seems to be different in the two cases. In particular, as we are going to see in the next section, the accusative possessum undergoes the same middle-field movement step in PP-splits as well. Nevertheless, the possessum in this case originates in a case position – the P-complement position – so its movement to the middle-field cannot be for case, in contrast to the movement of accusative objects in the plain VOS cases. This in turn suggests that middle-field A-movement in Greek is not uniformly motivated by case reasons. Instead, it seems plausible that Greek also has scrambling, which is similarly not always case-related. While the precise feature motivating scrambling remains an open question, with some works proposing a dedicated scrambling feature (Müller Reference Müller, Wilder, Gärtner and Bierwisch1997, Reference Müller1998 i.a.), I will assume that the same feature underlies the movement observed in the derivation underlying PP-splits.
3.3. A syntactic analysis of PP-splits
Let us consider now how we can extend this analysis to analyze the patterns found with PPs. We begin with a baseline sentence in (18a). As shown in (18b), the possessor can appear before the noun even in those cases where the noun is embedded under a P. However, (18c) illustrates that the possessor cannot occur in its own clause initially. The PP in (18a) also allows for PP-splitting, as demonstrated in (18d), in which case P and the possessor are moved to the left periphery, and appear separated from the possessum, which in turn appears postverbally.
Following Pesetsky (Reference Pesetsky1996), I take PPs expressing subject matter like the one in (18) to be externally merged in the V-complement position. FrontP is also available in this environment that allows for a prenominal possessor (18b). However, since the possessor undergoes criterial freezing in Spec, FrontP, it can never be extracted out of a PP, thus accounting for (18c) and the second row of Table 1, which illustrates the same fact for all PPs. Likewise, the possessor cannot be moved in a single movement step out of the DP where it originates because the DP is an opaque domain for extraction. On the other hand, after movement of the possessor to Spec, FrontP, where it satisfies its Front feature, the DP remnant comprising the possessum, tin epitihia ‘the success,’ is allowed to undergo movement into the middle-field. This movement step into the middle-field may involve an additional movement step through the PP specifier (not shown in 19). This is required if, following Abels (Reference Abels2003), Ps in languages that do not allow P-stranding like Greek are phase heads.Footnote 15
The possessor also has a Q-feature that it needs to satisfy after moving to Spec, FrontP. Nevertheless, because the possessor has undergone criterial freezing in Spec, FrontP and therefore cannot move out of this position, its Q-feature can only be satisfied by pied-piping the PP remnant to the clause’s left periphery, as illustrated in the structure below. Crucially, movement of FrontP alone, stranding P, is impossible because P is a phase in Greek, so this requires FrontP to move from the complement of P to its specifier, a step ruled out by Antilocality (Abels Reference Abels2003). This correctly predicts why the P must move with the possessor to the left periphery.
To summarize, the analysis in (20) proposes that the ability of a PP to undergo PP-splitting is determined by its merge height relative to a middle-field position, YP in (19). Specifically, PP-splitting is permitted when a PP originates lower than YP, allowing A-movement of the accusative possessum to that position. This prediction is supported by the fact that benefactive, agent, causer and subject matter PPs are argumental PPs; thus they are E-merged within the vP, lower than YP. Manner PPs, similar to manner adverbials, are standardly assumed to be licensed in the vP/VP area (Alexiadou Reference Alexiadou1997, Cinque Reference Cinque1999 i.a.). Likewise, Michelioudakis and Angelopoulos (Reference Michelioudakis and Angelopoulos2019) have shown that instrument and comitative PPs are licensed by a verbal projection responsible for an eventive interpretation. Lastly, directional PPs are typically assumed to merge as arguments. In contrast, PPs merged above the middle field – high PPs – are predicted to resist splitting. Unlike low PPs (merged below YP), high PPs attach in the C-domain above TP, where the only possible probe is C. The accusative possessum thus cannot evacuate the DP where it originates, as movement out of the DP cannot be triggered by a C-probe (see nominal opacity in 12). PP-splitting is therefore blocked, since it requires a movement step of the accusative from inside the DP where it originates – a step impossible with high PPs. I propose that this is the case for evidential and temporal PPs, which I argue are merged above TP. This hypothesis aligns with existing research demonstrating that PPs occupy different syntactic positions depending on their interpretation (Schweikert Reference Schweikert2005, Alexiadou and Anagnostopoulou Reference Alexiadou and Anagnostopoulou2007, Cinque Reference Cinque2006). In this literature, a standard assumption is that evidential and temporal PPs are merged higher than benefactive, source, agent, and other PPs allowing PP-splitting; (21) illustrates where the different PPs are merged relative to YP is provided below:
The two central components of the analysis, namely, that P-splits involve a multi-step derivation and that PPs are merged hierarchically, are supported from the behavior of different types of genitives and from Condition C effects. We begin with the genitives. The analysis predicts that only genitives that can undergo the initial DP-internal fronting to Spec, FrontP should be able to form a P-split. This prediction is borne out. Consider non-possessive, appositive genitives like ton Vrixelon in (22a). Unlike true possessors, these genitives cannot be moved to a pre-nominal position, as shown by the ungrammaticality of (22b) (Angelopoulos and Michelioudakis Reference Angelopoulos and Michelioudakis2024).Footnote 16
Because this initial DP-internal fronting step to Spec, FrontP is a necessary precondition for creating the remnant DP, the analysis correctly predicts that such constructions cannot form a P-split. The ungrammaticality of (23) confirms this.
In short, the inability of appositive genitives to undergo DP-internal fronting directly correlates with their inability to participate in P-splits, strongly supporting the claim that the former is a prerequisite for the latter.
The idea that PPs are merged in distinct syntactic heights finds support in Condition C data, which I present in the contrast below between evidential and comitative PPs. This contrast is replicated between evidential and temporal PPs, and the rest of the PPs that allow PP-splitting. Specifically, both evidential and comitative PPs can occur clause initially. However, whereas evidential PPs usually prefer this position, comitative PPs only do so when containing a focused XP, e.g., secretary in (24b). This contrast already suggests that evidential PPs are E-merged higher than comitative PPs. In (24a) and (24b), both PPs contain a proper name. It can be coindexed with the subject of the clause, shown as pro, when within an evidential PP, but not when in a comitative PP. This aligns with the assumption that evidential PPs are merged in the left periphery; specifically, evidential PPs are base generated in the left periphery, above TP where pro is interpreted (Angelopoulos and Sportiche Reference Angelopoulos and Sportiche2021). Because of this, the proper name within the evidential PP and pro do not c-command each other, thus, the two can be freely coindexed, as shown in (24a). In contrast, comitative PPs originate below the middle-field, and thus, lower than Spec, TP. Their presence in the left periphery, as in (24b), results from movement from their base position. Hence, a lower PP copy containing the proper name exists under Spec, TP, where pro is interpreted. Due to Condition C, the proper name and pro cannot be coindexed in (24b) since pro c-commands the proper name in the lower PP copy.Footnote 17 , Footnote 18
4. *Distributed deletion
Several analyses assume distributed deletion to account for nominal splits like those examined here (Fanselow and Ćavar Reference Fanselow, Ćavar and Alexiadou2002; Pereltsvaig Reference Pereltsvaig2008; and Bondarenko and Davis Reference Bondarenko and Davis2023 i.a.). However, a challenge for distributed deletion accounts is determining the conditions under which it applies as well as why it applies the way it does.
Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) propose that splits arise only in the context of operator movement, specifically connecting split NPs to information structure. This approach proposes that if a constituent XP, composed of a head X and, e.g., an NP-complement, carries distinct features (e.g., [Topic] on X and [Focus] on NP), then the entire XP moves as a single unit to distinct feature-checking positions (e.g., Spec, TopP and Spec, FocP), as illustrated in (25). Subsequently, X and NP are spelled out in the position where they have their feature checked. Consequently, in (25), the NP is spelled out in the lower XP copy, while X is spelled out in the high XP copy, yielding a split.Footnote 19
While applying (25) to derive an NP-split seems straightforward, deriving a PP-split requires additional principles. For example, Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) examine the Croatian split in (26a). Deriving this split requires two movements of the PP, na kakvo stablo ‘what kind of tree,’ from its base-generated position, as illustrated in (26b). Stablo ‘tree’ checks a feature in the intermediate position, while nkakvo ‘kind of’ checks a feature in the left periphery. Stablo is spelled out in the intermediate copy where it satisfies its feature, and kakvo is spelled out in the left periphery where it satisfies its own feature. To account for the obligatory adjacency of the P with kakvo, Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) posit an additional principle and introduce a corresponding feature: a P must possess a strong P-feature that triggers its movement. Moreover, the P must be phonologically adjacent to its selected category (see Goncharov Reference Goncharov2015 for criticism).
The conditions for licensing a split proposed by Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) are also met in the Greek PP-splits we analyze. Specifically, similar to the XP structure in (25), the PP comprises two parts, the possessum and the possessee, each carrying a formal feature. Therefore, we would expect PP-splits under this view to be derived from distributed deletion. To illustrate how this would work, consider the PP-split in (18d), repeated below:
Under a distributed deletion approach, three copies of the PP would be generated: one in its base position, the V-complement position, one in the middle field, where the accusative possessum checks a feature also involved in the movement of object in VOS, and a third in the left periphery, where the possessor checks its Q-feature. Distributed deletion would then apply, spelling out the possessor and possessum in the positions where they check their features, resulting in the split.
A question that arises in light of (28) is why the P is spelled out high together with the possessor. Our proposed account offers a straightforward explanation for this phenomenon (see discussion preceding 20). However, under Fanselow and Ćavar’s (Reference Fanselow, Ćavar and Alexiadou2002) approach, we would expect the P to be spelled out next to its selected complement, the success, in the intermediate copy. Therefore, this observation presents initial evidence that a distributed deletion analysis yields incorrect predictions. Now, consider the ungrammatical sentence in (18c), repeated below for convenience:
The ungrammaticality of this example follows from the proposed account because possessors undergoing criterial freezing inside the PP cannot be extracted out of it (see discussion preceding 20). In contrast, a distributed deletion account incorrectly predicts it to be grammatical. Under such an account, the possessor, whose, would be spelled out in the left periphery, where its Q-feature is checked, as shown in (30). The accusative possessum, success, would be spelled out in the middle field, checking the feature associated with object movement in VOS orders. The P would also be pronounced in the middle field, adhering to Fanselow and Ćavar’s principle that Ps must be spelled out next to the element it selects.
We next turn to the limitations of a more recent alternative proposed by Bondarenko and Davis (Reference Bondarenko and Davis2023), who formalize distributed deletion as follows:
This PF rule faces several theoretical problems. First, it remains unclear why it applies only to NPs and PPs, what prevents it from targeting CPs or VPs as well? How can a PF rule see syntactic categories and apply to a subset of them? The proposal offers no explanation. Second, the semantics it presupposes are problematic: why should the rule be limited to contrastive focus and topicalization, but not extend to plain focus?
Even setting these theoretical issues aside, the rule makes incorrect empirical predictions. In the case of PPs, since Ps do not bear features like Topic or Focus, the same problem arises as in the earlier account: what determines the realization of the preposition? Furthermore, consider appositive genitives, which, as previously discussed, cannot be fronted prenominally (i.e., to Spec, FrontP), as shown in (32a) and (32b) (repeated from 22a and 22b). These genitives also fail to form P-splits, as shown in (32c) (repeated from 23). This presents an important problem for the PF rule above, as it should be able to apply and thus derive a P-split, just as with other PPs.
The PP in (32c) contains a DP with a contrastive-focus interpretation, ton Vrixelon. If P-splits were derived via the same PF rule, we would expect this PP to form a P-split as well, just like other PPs (see Table 1): the P and the contrastive focused DP should be able to be spelled out in the left periphery stranding the possessum postverbally, However, this prediction is incorrect. Consequently, the main shortcoming of the PF-based account is that it fails to capture the observed correlation between the availability of prenominal genitives and the possibility of P-splitting – a correlation that the present proposal successfully explains.
Other accounts of distributed deletion face similar issues: they fail to generalize beyond the languages they were designed for, or rely on additional stipulative mechanisms. This is, for instance, the case for Iquito, where a distributed deletion analysis has been proposed (Murphy and Wilson Reference Murphy and Wilson2025). However, this account is highly language-specific and cannot be extended to Greek or other languages. Moreover, it requires an additional diacritic feature to operate. We will not discuss this proposal further, as it has limited relevance for cross-linguistic comparison.
An anonymous reviewer notes that the lack of cross-linguistic generalizability of distributed deletion is not necessarily problematic, since languages are expected to differ in how their syntactic structures are externalized, which might allow distributed deletion to operate differently across languages.Footnote 20 This is in principle correct. Under a single mechanism of copy deletion, for instance, cross-linguistic differences in PF conditions may determine whether the higher copy, the lower copy, or both copies are spelled out. A similar scenario could also hold for distributed deletion if the same operation (e.g., the version proposed by Fanselow and Ćavar Reference Fanselow, Ćavar and Alexiadou2002) applied universally, with languages differing only, for example, in whether the relevant feature that creates the condition for distributed deletion is Topic in one language and Focus in another.
However, this is not the situation we find with distributed deletion. In existing proposals, split constituents in different languages are analyzed as the result of entirely different distributed deletion mechanisms, none of which can be straightforwardly extended to other languages. As discussed above, some accounts assume that the XP carrying the movement-triggering feature must be spelled out, while others assume the opposite, that all XPs except the one bearing the movement-triggering feature are pronounced, and still others introduce additional machinery, such as diacritic features. Some of them tie the application of distributed deletion to information structure properties whereas others do not. The limitation of distributed deletion, therefore, is that no single mechanism of it can be shown to underlie split constituents cross-linguistically.
One could alternatively assume that languages do not differ in how distributed deletion operates, but rather in whether they allow the operation at all. On this view, Greek may simply lack distributed deletion, while other languages have it. However, this scenario is not compatible with how distributed deletion has been proposed in the literature. Under current proposals – especially those linking the operation to Topic or Focus – the prediction is that distributed deletion should in principle be available in all languages, since such information-structural properties appear to be universally available. Moreover, nothing in the proposed conditions on distributed deletion suggests a principled way in which the operation could be active in one language but entirely absent in another. Consequently, if we examine a language that satisfies the conditions under which distributed deletion is predicted to apply and show that its split constructions cannot be derived through this operation – as argued here for Greek – this constitutes strong evidence against distributed deletion as a grammatical mechanism more generally.
To sum up, the proposed analysis fares better than a distributed deletion alternative because it only works with what is independently needed; a general PF-mechanism of copy deletion that applies to every instance of internal merge.Footnote 21 No additional PF-mechanisms that only derive splits are needed, nor are additional assumptions required to account for where, for example, a P is spelled out, nor additional spurious features, such as a strong P-feature that is important in Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) to motivate movement of the P. Furthermore, the principles that apply to distributed deletion lack theoretical grounding or are very specific to certain languages or constructions, which in turn make a cross-linguistic analysis not possible. For instance, Fanselow and Ćavar’s PF principle governing P spell-out faces this challenge: it is not clear why a PF principle should be sensitive to P’s selectional requirements. Also, as we saw, this principle does not hold in Greek. Furthermore, as discussed in Murphy and Wilson (Reference Murphy and Wilson2025), Fanselow and Ćavar’s distributed deletion (or other versions of distributed deletion) account cannot be extended to Iquito splits. This has as a consequence that a new, language-specific distributed deletion mechanism with novel constraints is needed. This approach raises a significant concern: the absence of a principled account of distributed deletion allows arbitrary modifications in its application and the postulation of novel PF-principles, thereby rendering the theory effectively unfalsifiable. Specifically, the present proposal predicts that, within the DP domain, the availability of both a prenominal position and a middle-field position permits the formation of a split. When such a DP is embedded under a PP, the possibility of P-splitting depends on the height at which P is merged – either below or above the middle-field position – as well as on the phase status of P. The above data and considerations therefore provide strong evidence against distributed deletion in the formation of splits. Combined with similar conclusions in Goncharov (Reference Goncharov2012), who proposes a syntactic alternative to the splits initially posited by Fanselow and Ćavar (Reference Fanselow, Ćavar and Alexiadou2002) as evidence for distributed deletion, these concerns suggest that distributed deletion is not attested in natural language.
Before we close, a question to address is why exactly distributed deletion is not allowed in natural languages. It is reasonable to assume that constituency plays an important role here. Syntax respects constituency and the picture that arises from deletion phenomena, such as ellipsis, is that constituency is respected there too, as only constituents can be deleted (Fox Reference Fox2000 and Merchant Reference Merchant2001, among others). Consequently, distributed deletion seems to be ruled out in natural languages simply because it is a mechanism that disregards constituency, permitting the deletion of non-constituents.
5. Conclusion
This paper investigated a split of Greek in which a possessor within a PP is separated, along with P, from the possessum. New data revealed that this split is sensitive to the PP’s semantic interpretation. PPs are merged at distinct syntactic heights in accordance with their interpretation. Thus, the availability of PP-splitting is linked to the PP’s merge height relative to an independently motivated middle-field position. However, a distributed deletion account proves inadequate. It not only requires supplementary principles to explain PP-splits but also yields incorrect predictions. Similar criticisms have been leveled against distributed deletion accounts of splitting phenomena in other languages (Goncharov Reference Goncharov2012, Reference Goncharov2015), thus refuting the existence of distributed deletion.
Acknowledgments
The author also wishes to thank Artemis Alexiadou, Elena Anagnostopoulou, Richard Kayne, Hilda Koopman, Dimitris Michelioudakis, Léa Nash, Anna Roussou, Christina Sevdali, Ur Shlonsky, Dominique Sportiche, Vassilios Spyropoulos, Gary Thoms, the audience of the Caulage summer school and three anonymous reviewers for their invaluable feedback and comments on earlier versions of this paper.
Funding Statement
This study was funded by the Marie Skłodowska-Curie fellowship ‘Form-Meaning Mappings in Language’ (Project ID: 101147196), which the author gratefully acknowledges.
Open access
