3.1. Step A: dispensing with phrase syntax

Steps towards a pure head syntax for clitics have occasionally been taken in the literature, but they yield only partially satisfactory insights into clitics. Manzini (2022), updating Manzini and Savoia (2005), cf. also Manzini and Savoia (2017), assumes that clitics are heads throughout the derivation. Their labels include N, D, φ for direct case clitics and Appl for other clitics (with different flavors: e.g. Dat, Loc(ative), Gen(itive), more on this below). Thus clitics are not exponents of specialized functional categories (Voice etc.), nor are they inserted under/incorporated into an independently defined sequence of heads (no cartography). Rather they themselves build the functional spine of the predicate/event through projection of their labels. In order to implement her program, Manzini has recourse to standard Set Merge phrase structure. As clitic labels project on the functional spine of the verb, they create φP constituents (accusative, nominative) and ApplP constituents (with various flavours, Gen, Loc/Instrumental, Dat), whose internal order (or range of internal orders) matches that of independently postulated functional projections of the predicate. For instance, the final picture proposed for French looks as in (9).

Manzini also assumes that there are two paths to the satisfaction of the Theta Criterion. Following Chomsky (2000:93), an argument DP must be first merged in a theta configuration with a predicative head, at least when the theme (internal argument) role is concerned. However, Higginbotham (1985) also argues that an argument slot (the so called R role) is open at the N head of an NP, which he assumes to be a predicate. The slot remains open as the NP is variously modified, and is ultimately satisfied by a head, the D head. Similarly, Manzini’s proposal is that a D clitic may satisfy the Theta Criterion, again at first merge, by closing some projection of a V predicate head, provided it has an open argument. Hence, the idea that clitics are heads throughout the derivation does not necessarily pose a problem for theta licensing, though it implies that the licensing mechanism of clitics differs from that of arguments. The discussion about theta licensing will be resumed in section 3.3.

The analysis in (9), nevertheless, covers only part of the research questions listed in Section 1. Above all, it does not provide any hint about how to tackle the question of the final position where clitics surface – which is generally not the v phase, but rather the C-I phase. To reach a better understanding of clitic syntax, in the following subsection we propose to move a step further, by suggesting that other differences between clitics and XPs obtain because clitics are not set merged (even as heads), but pair merged, and in fact form Pair Merge sequences (Chomsky 2020; 2021).

3.2. Step B: dispensing with Set Merge

In this section, we develop a Pair Merge (as opposed to Set Merge) syntax for clitics. Specifically, we apply Chomsky’s (2020) construct of Pair Merge sequence to model the clitic group, hence incorporation.

The operation of Pair Merge of α with β, notated <α, β>, is defined by Chomsky (2004:118–120), with the explicit intent of modelling adjunction of α to β. Since Pair Merge is not nearly as widely used as Set Merge, we preliminarily remind the reader of some of its basic properties. The element adjoined to, β, is first set merged; adjunction then applies to replace β by <α, β>. Applications of Set Merge and Pair Merge interleave in the derivation. Thus <ADJ, NP> can be embedded in a Set Merge structure [DET <ADJ, NP>]. Conversely, ADJ can have internal set merged structure, for instance a complement. At Spell-Out, <α, β> undergoes simplification (SIMPL) to {α, β} for purposes of interpretation at SEM and linearization at PHON.⁷

Epstein, Kitahara and Seely (2016) explicitly extend Pair Merge to the other classical instance of adjunction, i.e. head adjunction. Chomsky (2020: 56) models at least inheritance pairs, e.g. <C, I>, in terms of Pair Merge. Another important extension of Pair Merge is introduced by Chomsky (2020: 49–52) in the form of the notion of Pair Merge Sequence. The crucial step is acknowledging that the structure formed by the modifiers in (10) is “not just a set of paired elements, it’s a sequence of paired elements” as witnessed by respectively constructions like (11) (cf. McCawley 1998).

Formally, as illustrated in (12), “we’re forming a sequence which begins with some conjunction, and then contains elements, each of which is predicated of something. So we have a sequence of elements that looks like [13], with links L_i”. As for L, the suggestion is that “L for nominal sequences is just n, the categorizer of each of the coordinated phrases”, assuming that roots are themselves deprived of a category (Marantz 1997). In any event, independently of whether we want to adopt acategorial Roots or not, Links n, v and the like are independently individuated as phase heads (weak or strong).⁸

One interesting consequence follows from this treatment of modifiers concerning constraints on movement. Specifically, “if you have the phrase old man, you can’t extract man and leave old; you can’t extract old and leave man”, as already indicated by Chomsky (2004: 120) who explicitly excludes moving β in <α, β>, while leaving α in situ. Hence “you have the Coordinate Structure Constraint [CSC] because every term is inaccessible. You have the adjunct island constraint because you can’t pull the elements out of adjuncts”.⁹

In this article, we model the clitic cluster as a Pair Merge sequence, where each member of the sequence is formed by a clitic head and a link v (but see section 4 for a revision). We adopt the conclusion of Manzini and Savoia (2017), Manzini (2022), reviewed in section 2.2. above, that the sequence of clitics is comparable to the ordinary sequence of argument-related functional projections in the sentence, specifically φP projections for accusative and nominative arguments (following Agree with v and I) and (variously flavoured) ApplP projections for oblique arguments. Parametric choices determine which of the possible orders of φ and Appl clitics is chosen by the language, yielding in particular crosslinguistic variation in the ordering of Acc and Dat clitics, cf. (3).

Earlier proposals on clitic structure were dismissed above because they predict a direct correspondence between the order of phrasal arguments and the order of clitics which is in fact not observed (neither match nor mirror). Pair Merge per se does not predict the (partially language specific) order of clitic sequences. However, Set Merge (roughly, argument structure building) and Pair Merge (roughly, head adjunction), taken together, provide us with the means for differentiating the order of phrasal arguments and that of clitics by purely syntactic means (recall that order phenomena are syntactic, though they are ultimately manifested by linearization).

Consider then the simple Italian sentence in (13). The internal order of the string points to lo ‘it’ merging first with v, followed by merger of the Appl clitic gli ‘to him’. This order of merger is independently attested in French with 1/2P datives and could easily fit into a Set Merge schema.

The core proposal we advance here however is that clitics do not Set Merge with VP. On the contrary a Pair Merge sequence is created, beginning with the Pair Merge of the inheritance pair <V, v*>, with v* acting as the Link, and continuing with the elements that modify the verb, in the sense that they specify its transitivity (the accusative clitic) or ditransitivity (the Appl_DAT clitic). The resulting Pair Merge structure can be represented as in (14), following Chomsky’s (2020) notation.

In this simple example, the order of the sequence of clitics is similar to the order of v-related heads, assuming that Pair Merge reflects the order in which theta roles are discharged. Schematizing, the correspondence between a conventional phrase marker for clitics (qua XP arguments) and the Pair Merge syntax envisaged here can be depicted as in (15a). The phrase marker in (15a) takes up again and generalizes the schema in (9) for French. The sequence in (15b) begins with the inheritance pair (V, v*), on top of which the other pairs are built in order of discharge of theta-roles.

The rough correspondence diagram in (15a), provided here to favour construal of the sequence in (15b), should crucially not be read as implying equivalence. The whole point of the discussion that precedes is that, while convergence at SEM should be insured between phrasal and clitic arguments, the syntactic structures implicated are not in a one-to-one relation. Set merged DPs and pair merged clitics are not necessarily ordered (hence linearized) under invariant conditions crosslinguistically, cf. (3). Importantly, evidence from Romance shows that the ordering of Dat/Acc clitics is not uniform within the same language. Instead, the order of clitics is often sensitive to animacy, which is immaterial in determining the relative order (dominance and precedence) of phrasal arguments: for instance, 1/2P datives and 3P datives in the same function are split in French, cf. (16).

Phrasal datives normally split according to whether they are selected or unselected, as revealed by binding facts (recall (1)–(2) vs. (4)–(5) above). This distinction is obliterated in clitic syntax: all types of datives (indirect objects, secondary objects, experiencers, benefactives, caused subjects, datives of possession, etc., see (17)) are mapped to a single clitic form (e.g. it. gli or fr. lui), which is linearized in the clitic string regardless of dominance relations observed among the corresponding phrasal arguments.

We believe that Pair Merge sequences best capture the lack of dominance in clitic strings. Indeed, going back to the definition of Pair Merge by Chomsky (2004; 2020), it is evident from the example considered (adjectival modification), that though the sequence is internally ordered (it is, precisely, a sequence), only the sequence as a whole is ordered by dominance in the Set Merge tree. In other words, the Pair Merge sequence, however internally articulated, is a single syntactic object in the general phrase marker.¹⁰ Despite being obvious (in the sense that it holds by construction, not as a theorem), this result implies another important consequence, namely that the internal structure of the sequence is inaccessible to movement. Only the entire sequence is a syntactic object accessible to movement – a point already discussed when introducing the general notion of Pair Merge sequence in (12).¹¹ In essence, we take this important difference between Pair Merge and Set Merge to yield the incorporation property (see also section 4).

Against this background we can take up again the examples in (7), which illustrate idioms involving one or more clitics, as in the case of the It. verb cavar-se-la ‘to get by’ (lit. ‘to remove it_fem for oneself’). As already observed, the clitics have regular syntax (e.g. the accusative la agrees with the perfect participle, cf the relevant example repeated in (18)).

Idioms, whether involving lexical DPs or clitics, must be listed in the lexicon. Thus, it is impossible to replace the clitic pronoun by means of a full pronoun, i.e. a phrasal pronoun merged as an XP. This is because V-XP pronoun idioms and X-V idioms must perforce be separate listemes. The listing of Cl-V as an idiom renders previous derivational history at best redundant, since the very reason why a phrasal position of first merge (or a pro) is postulated, namely theta-marking, does not hold here, as already mentioned in the discussion of (7).

More importantly, as already mentioned, Pair Merge accounts for the fact that, in general, the clitic string cannot be split, and cannot be split from the verb in ad-verbal clitic languages, i.e. the languages in which clitics must be adjacent to verbs, including Italian and French. Kayne (2000: 48) illustrates the impossibility of splitting clitics from one another by the examples in (19), involving restructuring environments.¹² The only possible split is between subject and object clitics, for instance in French interrogatives (20a) or in languages like Borgomanerese (Tortora 2015; Manzini and Savoia 2017) where object clitics are enclitic (20b), cf. section 4.

Data like (19) require an account of clitics where object clitics are accessible only as a cluster, possibly including the verb. In a Set Merge syntax, the only open option is to postulate incorporation of the clitics to the verb, independently of the specific mixed head/phrasal syntax proposed (Kayne 1991; Sportiche 1996; Roberts 2010). The solution we suggest is to resort to the alternative mechanism of Pair Merge sequences.

The terms of the theoretical debate concerning head movement and incorporation are well-understood. In GB theory and in some minimalist models (Kayne’s LCA, cartography) head movement comes for free – namely it results from simply intersecting the notion of head with the operation of movement. Chomsky (2001) points out a number of difficulties for actually saying that head movement comes for free – since the properties of head movement are different from those of DP-raising or wh-movement. As an alternative, he suggests that head movement is a PF operation; perhaps m-merger (Matushansky 2006) may be thought of as an implementation of this idea for clitics. In this article, we suggest a different path, namely taking the conventional (and well motivated) conclusion that cliticization is a core syntax phenomenon – but modelling it not via a Set Merge grammar but via a Pair Merge grammar.

In the present perspective, Pair Merge units of the general form <Cl, v> model clitics as heads adjoined to a functional projection of the verb. Pair Merge sequences of the general form in (15b) model the intuition that clitic heads form a complex syntactic objects with other clitic heads (and with the verb) – i.e. ‘incorporation’. This set of assumptions comes at no cost, since Pair Merge is independently needed to model adjunction in instances of modification (Chomsky 2004; 2020), as well as of verbal head syntax, as explicitly proposed by Epstein, Kitahara and Seely (2016). Similarly, the notion of Sequence is independently needed to model coordinations (Chomsky 2021). Therefore what is really new in the present proposal is simply an extension to clitics, based on the conclusion that they have a pure head syntax.

The question how Pair Merge and Sequences mesh in with Set Merge is also considered by the literature quoted. In present terms the question is how these ideas apply to derivations in which phrasal syntax mixes with clitic syntax. This is illustrated in the derivation in (21). As the first step, the Link v(*) is set merged in the general tree, as in (21a); in this we are simply following Chomsky (2004). The next step is also independent of the present work – namely the postulation of a Pair Merge couple <V, v> modelling v-V inheritance in (21b) (Epstein et al. 2016; Chomsky 2020). Adjunction of the clitic Cl₁, like other adjunctions, is modelled by external Pair Merge, yielding the Pair Merge couple <Cl₁, v*>. The relation of two elements, here Cl₁ and V merged to the same Link is assumed to generate a Pair Merge sequence, as in (21c). Other clitics can be pair merged to the Link and added to the sequence, as in step (21d)=(15b).

Summing up, our proposal that clitics are structured by Pair Merge, rather than by Set Merge contrasts with classical approaches treating the head status of clitics as derived from a fundamentally phrasal syntax. In (15) and again in (21) we have addressed the question whether and how a treatment in terms of Pair Merge is feasible. Another key question of feasibility is how different syntaxes can converge on the same argument structure, i.e. the theta-assignment question, to which we return in section 3.3.

In this section, we also addressed the question whether proceeding our way yields any empirical advantages. We made three main empirical points. First, under the phrasal derivation of clitics, we do not really expect there to be Cl-V idioms. Either the Cl forms an idiom with V at first merge, but then we would hardly expect movement to apply – or conversely the idiom is defined on derived structure, but wouldn’t the phrasal configuration at first merge already have determined theta-assignment? A second, more general point concerns the different syntactic order of phrasal and clitic arguments, which translates into different linear orders at EXT. Under the phrasal derivation of clitics, ordering is standardly accounted for as an effect of the division of labor between syntax and morphology. We envisage the contrast between Set Merge and Pair Merge within the syntax as replacing the more traditional contrast between syntactic phrase structure and morphological clitics. In other words, the syntax itself accounts for the lack of isomorphisms between the two orders, and precedence asymmetries at EXT simply reflect syntactic facts. Third, sequences derive the incorporation properties of the clitic string by construction and without need for any further operation. In section 4.1–4.2, we further argue that the present approach is more restrictive when it comes to predicting possible placements of the clitic string, while under classical Set Merge and movement approaches a number of phenomena are predicted to occur (such as mirror orders) that are hardly ever observed. We devote section 5 to a general comparative discussion of the two models. Before that, however, we must complete our discussion of convergence at SEM.

3.3. Step C: nonuniformity of theta licensing

Let us go back to derivations involving XP arguments, for instance (22). Following merger of the head V with the internal argument, Appl is merged, supporting the benefactive argument; this elementary predicate structure is then merged with v*. We follow Epstein et al. (2016), Chomsky (2020: 56) in modelling head raising by Pair Merge. Pair Merge <V, v*> corresponds to inheritance from v* to V; this in turn triggers Agree with the internal argument and accusative case licensing. Short movement of the internal argument (according to Chomsky (2015)) may be required, as also indicated.

Summarizing the results of previous sections, mixed Pair Merge and Set Merge structures in (22) can be taken as the starting point for structures including both phrasal and clitic arguments as in (23). In (24) we simply adjust the structure in (22) to include not a single Pair Merge, but a Pair Merge sequence, including a clitic.

Now, comparison between (22) and (24) leads us to the conclusion that, in order to pursue a Pair Merge account of clitics, we must discard Uniformity as a principle governing the syntax/semantics interface. In other words, in the spirit of Culicover and Jackendoff (2005: 6), we must depart from the claim that “the same meaning always maps onto the same syntactic structure”. In the lexicalist model, where a certain set of theta-roles is imputed to each lexical predicate, the mapping to a certain syntactic structure is enforced by principles such as the Uniformity of Theta Assignment Hypothesis (UTAH) of Baker (1988). In more recent constructionist models (in the sense of Borer 2005), the build-up of argument structure corresponds to the build-up of the event. In practice, both are structured around functional projections reflecting underlying semantic categories (result, process, etc. cf. Ramchand 2008). Hence, Uniformity holds by hypothesis.

The present proposal is incompatible with Uniformity in that phrasal argument structures are created by Set Merge and ordered by dominance, whereas clitic clusters are sequences of heads pair merged with v. This asymmetry raises two questions, namely whether there is independent evidence in other areas of grammar that Uniformity may not hold – and conversely how to explain the fact that clitic and phrasal syntax (full pronouns) can converge on the same meaning.

As for the first question, it is an evident fact that the same semantics can have different externalisations, both crosslinguistically and within the same language. An example among many is passive (lato sensu, i.e. arbitrarization of the external argument), which in the Romance languages can have a periphrastic construal (be + perfect participle, as in English) or a construal by the se clitic. This kind of evidence can be circumvented by assuming that syntactic structures are underlyingly identical, though they are differently externalized by the lexicon and the morphology of the language. Uniformity therefore entails that a rich morphological component works to opacise syntax to some extent (on this point, see section 2). It further leads to highly abstract, tendentially universal functional structures, typified by cartography, which raise issues of learnability (or evolveability, Chomsky et al. 2019).

As for whether there is an alternative to Uniformity, we need look no further than the autonomy of syntax thesis (Adger 2018 for recent discussion), under which we expect “the grammar to be able to support semantic description” (Chomsky 1957), but we do not expect semantics to be pre-encoded in syntax: syntax restricts interpretation without determining it. This model “attributes the relative uniformity of the expression of argument structure to the principles that interpret syntactic structure semantically” (Wood and Marantz 2017) as opposed to imputing it to uniform syntactic structures.

Consider then the simple sentence in (22a) again, containing a ditransitive verb, with the structure in (22b), or (25) below. We may assume that sister of V is the configuration for theme interpretation and sister (Spec) of Appl is the configuration for goal interpretation of the argument.

Phrasal structures are largely beyond the scope of the present article, however (26) is offered here as an illustration of how syntactic structures are interpreted at the semantics interface. Following a neo-Davidsonian approach (Parsons 1995; cf. Larson 2014), (25) receives the representation in (26), where bring denotes a predicate of events to which the arguments are related by relations of Agent (Gianni), Theme (the book), Goal (Paolo).

We contend that clitic arguments have an analogous interpretation, in (28), although they are structured in Pair Merge sequences as shown in (27b). In this respect, recall that Chomsky (2020) envisages a CONJ (conjunction) operator as heading the sequences, whose members are interpreted as conjoined. As it turns out, therefore, a conjunction semantics for argument structure of the type independently proposed for phrasal syntax follows by construction from Pair Merge sequences of clitics.

The mapping of (25) into (26) and the mapping of (27) into (28) rely on different mechanisms. In the former, as we saw, theta relations are configurationally defined, e.g. Theme is a relation established between a V and its sister. In the latter, recall that clitics are labelled by their intrinsic properties; labels include: N/D/φ for direct case clitics and Appl for oblique clitics, with different flavors, Dat(ive), Gen(itive) etc.¹³ Clitic morphology then conveys information as to whether clitics are interpreted as direct arguments or obliques, whereas hierarchical configurations play no role. Taking the suggestion of Manzini (2022), we assume that the saturation of the open slots of the predicate takes place in each of the Pair Merge links by the relation established between the clitic (the referential, argumental content) and the v (the predicative open expression). As Higginbotham (1985) proposes for the saturation of the R role in DPs, as long as a certain role of V is not saturated, it gets passed on to v, for instance by the Pair Merge operation that insures inheritance and perhaps to higher heads as long as the phasal domain is open (see the discussion in section 4).¹⁴

Granting these assumptions, we can conclude that though phrasal arguments and clitics may have different syntactic structures, they converge on expressing the same argument structures at the syntax/semantics interface.

4.1. The merging point of clitics: crosslinguistic variation

In the most-studied Romance languages such as French, Italian, Spanish, etc. clitics always occur close/attached to verbal heads. However, if we approach Romance from a microcomparative point of view, it turns out that, first, clitics are not necessarily attached to the verb and, second, in any given language one or more phase heads are capable of hosting clitics. To illustrate these points, we first focus on languages exhibiting so-called interpolation, i.e. the occurrence of XPs of various kinds between the clitics and the verbal head. Notice that the term interpolation is a symptom of our biased view on clitics, which we expect to be always attached (or close) to verbs as they are in major modern languages such as French, Italian or Spanish. However, if we turn our attention to less-known Romance vernaculars, including medieval languages, or to other linguistic groups and families (Pescarini, to appear) we find out that no privileged relation holds between clitics and verbs.

Evidence of clitics merged in I comes from Italo-Romance and western Ibero-Romance dialects that show interpolation of aspectual adverbs that are merged at the I/v boundary (Pollock 1989). The occurrence of clitics to the left of this class of adverbs, but to the right of complementizers, negative markers and subject clitics (as shown in (29a-c), respectively) indicates that clitics can be merged to a I head even if the inflected verb does not merge to the same head.

Evidence of clitics in C comes from medieval Romance languages that display Germanic-like inversion in main clauses (i.e. AUX > Subject > Participle; for an overview, see Pescarini 2021: 183–210 and references therein). In contexts where inversion is ruled out (especially, but not exclusively in embedded clauses), elements such as negation, preverbal subjects, IP-scrambled XPs¹⁸ can occur between clitics (in C) and the inflected verb that remains in I, see (30).

Lastly, dialects in which clitics never climb to the inflected verb offer evidence of clitics that are merged at the edge of v. In Borgomanerese, for instance, object clitics are not merged with the inflected verb, which climbs to I as in the other Romance languages. Clitics remain low as witnessed by the interpolation of aspectual vP adverbs in (31) (Tortora 2002; 2015). vP-internal arguments and adjuncts in Borgomanerese do not interpolate between the verb and enclitics.

The first row of Table 1 summarises the data illustrated in (29)–(31), showing that clitics can be merged with any phase head, including C and I (the latter via inheritance). By the same token, it seems reasonable to extend the conclusion reached for languages with so-called interpolation to the other languages – those summarised in the second row of Table 1 – where clitics and verbs are always adjacent.

The second row of Table 1 is easily illustrated. Languages with adverbal clitics in C are medieval Romance languages which display verb-subject inversion and never exhibit interpolation phenomena in contexts of V-to-C movement. When the verb is inverted (in C or at the I/C border), clitics can either precede or follow it, depending on factors such as polarity or focus-fronting, but they can never be separated from the inflected verb. Proclisis is normally found in questions, where object clitics and V precede inverted subjects, which at this chronological stage are not clitics, as in (32).

Modern Romance languages like Italian or French, illustrated in the text, are characterized by the placement of the verb and of the clitic group in I. In (33) we provide an illustration of the third attested typology, where the v attachment of clitics is signalled by the fact that they form a cluster with the perfect participle, not with the auxiliary in I.

The generalization that the verb and/or the clitic group can surface in a limited set of positions (see also Roberts 2010), namely all (and only) the phase heads (v-V, or I, or C), is actually predicted under the present approach to clitics, if C-I and v-V are the only (sentential) phase heads and if not every category can serve as Link in a sequence but only phase heads can (cf. Chomsky 2020).²²

This prediction is further confirmed by data from the DP domain, where clitics can occur as long as they Pair Merge with a phase head. Even remaining within the Romance languages, there are robust examples of cliticization in DPs with kinship terms, which turn up with clitic possessors in many Center and South Italian varieties (Manzini and Savoia 2005, III: 660–684), cf. (34). The kinship term may be Determiner-less, evoking raising of N to D (Longobardi 1994). However structures D-N-enclitic are also found, especially in the plural, when the kinship terms may be quite high in the structure, but not quite high as D, as in (34b).

Based on the evidence in (34), we may conclude that clitics are pair merged with the Link n, also a phase head (Chomsky 2020). There they are preceded by N in the D position or slightly lower. In any event the predictions of the analysis are verified for DPs as well.²³

Now, if we look at Table 1, merger to v is not even the majority option. We therefore revise our original proposal: clitics are not (necessarily) first merged with v. This happens in languages such as Borgomanerese in (31), in which clitics never climb.²⁴ As for languages in which clitics are found in I or C (regardless of the position of finite verbs), we contend that clitics are pair merged with I/C and do not climb from one phase to another.

Because of obvious limitations of space, we cannot pursue all the clitic typologies listed in Table 1. We however revise our discussion as concerns the modern Romance languages, which are the empirical focus of the present discussion. We do so by first considering further valuable evidence supporting our approach, brought by enclisis/proclisis alternations, namely enclisis in the imperative (section 4.2) and enclisis of the subject clitic in French interrogatives (sections 4.3–4).

4.2. The problem of enclisis

The languages in which clitics are always adjacent to verbs still display partial independence of the verb and of the (object) clitic group in proclisis/enclisis alternations. The Italian and French declarative examples that we have used so far all have proclisis, cf. (35a). However, the Romance languages present various patterns of enclisis. Modern Italian, for instance, has enclisis of object clitics in imperatives and non-finite clauses, as in (35b–c). French also exhibits enclisis of object clitics in imperatives, and in addition enclisis of subjects clitics in main interrogatives (subject clitics will be thoroughly analysed in section 4.3).

A considerable tradition of studies starting with Kayne (1991) imputes enclisis/proclisis alternations to head movement. More specifically, Kayne’s (1994) antisymmetric model excludes right adjunction and multiple adjunction. As a consequence, proclitics cannot be merged with the functional head hosting the verb; rather, clitics are merged with a silent head—or with a series of silent heads—that precedes the one hosting the verb. Enclisis results when the verbal head moves further, incorporating the clitics or skipping the head to which clitics are attached (for other attempts to derive enclisis circumventing the Head Movement Constraint, see Lema and Rivero 1991). Our account of enclisis, which will be presented in Section 4.4, follows the same intuition, but we contend that Pair Merge offers a more elegant solution. To illustrate our point, some preliminary remarks on enclisis are in order.

The conditions ruling enclisis/proclisis alternations are not uniform across Romance and across syntactic contexts. The triggers of enclisis vary crosslinguistically: they include polarity, mood, information structure, finiteness, etc. Enclisis in imperatives and infinitives, see (35b) vs (35c), do not result from the same mechanism; in fact, languages such as French display enclisis in imperatives, but proclisis in infinitives. Analogously, it is highly implausible that one and the same mechanism may account for cases of enclisis of object clitics to past participles in Piedmontese varieties in (33) and enclisis to finite verbs in present-day western Ibero-Romance or in medieval Romance (see (36)), where enclisis is compatible with topicalized XPs and barred by focus-fronted XPs (so-called Tobler-Mussafia law).

If we look beyond the specific conditions ruling enclisis, it is worth noting that none of them is computed in the v phase. In particular, enclisis correlates with properties of I/C, while in no language enclitic/proclitic placement is affected by the inner aspect or the argument structure of verbs. Enclisis/proclisis alternations therefore suggest that clitics are not merged to v even in languages such as Italian or French in which object clitics are always adverbal. If clitics were merged in v, nothing in the higher phase would be allowed to tamper with a syntactic object built in the v phase.

Furthermore, Set Merge and head movement cannot predict elementary generalisations on the make-up of proclitic vs enclitic strings, for instance that the object clitic group does not normally mirror, specifically under proclisis/enclisis alternations. Indeed, head(-projection) movement (rollup movement etc.) is expected to yield a very high incidence of mirror patterns, of the same kind one finds in other clausal domains (see Cinque 1999). On the contrary, when two or more clitics co-occur, the orders in proclisis and in enclisis do not mirror each other, see (37), as one would expect under a head incorporation analysis in which the verb incorporates the lowest clitic first and then the highest.

One way to circumvent this objection is by assuming, as Kayne (1994) does, that enclitics do not mirror because two clitics are often merged to a single head (see also Terzi 1999, from which the following tree is taken with minor modifications). Verbs are furthermore allowed to skip the complex head X in (38) without violating the Head Movement Constraint.

By assuming (38), however, we implicitly assume that, in languages such as French or Italian, proclitics are never incorporated to the inflected verb, a solution that raises more questions than it can answer. First, what is the nature of the head hosting clitics? Second, if clitics and verbs are not merged together, why cann’t they be separated in languages such as Italian, Romanian, or Spanish? We know that the position of the inflected verb in the Romance languages is subject to subtle variation, see (39) (from Ledgeway & Cruschina 2016), but proclitics occur immediately before the inflected verb. If adjacency were a coincidence, as (38) entails, one would expect adverbs to occur between proclitics and verbs more frequently than they actually do.

By the same token, one would expect proclitics to be separated from lexical verbs more readily than from auxiliaries, which cross-linguistically tend to target higher I-positions (see, e.g. Pollock 1989). Since these predictions are not borne out, we can safely conclude that in the majority of present-day Romance languages object clitics are incorporated into finite verbs. Hence, clitics have a hybrid behavior: they must be somehow incorporated to the verb, but at the same time they enjoy a higher degree of freedom than bona fide affixes, cf. section 1.

In our view, the hybrid behavior of clitics; yielding syntactic configurations such as (38), is better captured by assuming a Pair Merge structure whereby clitic pronouns are first merged in I (more on this in Section 4.4).

Note that the few cases of mirror orders in enclisis have attracted relatively high interest in the literature (see also Ordóñez 2002). Yet thorough surveys of their occurrence, available for Romance (Manzini & Savoia 2017) show that considerable constraints are placed on them. Specifically, mirroring sequences, such as well known French (37) only involve accusative clitics and some (allomorph of the) Appl clitic. Sequences of two Appl (oblique clitics) never mirror, e.g. in French (41).²⁷ What is more, mirror is hardly ever obligatory, and more often than not, alternates with preservation of the proclitic order, cf. (40c) for French (Laenzlinger 1993).

This is not to say that movement theories of clitics could not be adjusted to cover the relevant subregularities.²⁸ The argument is that some expectations brought about by the model are not realized.

4.3. Subject clitics

A prediction made by the model of clitic placement based on Set Merge and head movement is that there is a uniform syntax for all classes of clitics. This is particularly relevant when one considers subject clitics. Subject clitics are always adverbal (only negation and object clitics can occur between the subject clitic and the verb), and differ quite systematically from object clitics with respect to placement. Subject clitics occur in enclisis in contexts of V in C, like main interrogatives, though object clitics are proclitic in the same context (as expected with a finite verb).

The comparison between subject and object clitics in sentences like (42) shows that subject and object clitics do not form a single sequence of pair merged elements. Subject clitics are probably merged with a different phase head in (42b) (as a matter of fact, though the ordering of object clitics is complex and subject to language-specific irregularities, subject proclitics always precede object proclitics).

Subject clitics are not connected to the v phase, but rather to the I phase. A reason for this is that they are means for the satisfaction of the EPP in languages like North Italian varieties which have this constraint at the C-I phase (thus, subject clitics can be expletive). We therefore assume in (43b) that subject clitics in declarative clauses are pair merged with Link I. This presupposes that I acquires the phase head property by inheritance from C (Chomsky 2008), see also §4.4.²⁹

The subject clitic can close any open argument slot by predication (with v^(*)P) and satisfy the EPP (in this respect, Italo- and Gallo-Romance varieties exhibit variation; see Pescarini 2022).³⁰ This analysis once again separates clitic from phrasal derivations. Specifically, in standardised accounts of phrasal syntax, XP arguments are first merged in thematic position and eventually undergo movement to satisfy the EPP (in languages where it holds). As for subject clitics, we assume that they satisfy the EPP at first merge and can receive an argument slot via predication. Crucially, if all arguments slots have been filled, the subject clitic is still wellformed namely it is what we call an expletive, for instance in French (44).

4.4. Understanding proclisis/enclisis alternations

We now focus on interrogative inversion of the subject clitic. In sentences like (42b), which we have used to illustrate the relative independence of subject and object clitics, both are adverbal (strictly adjacent to the verb). Furthermore, we note that in proclisis no lexical material can be interpolated between subject clitics and object clitics (with the sole exception of the negation which is itself a clitic). In classical cliticisation theories, incorporation of the subject clitic into the verb depends on verb movement to C, where the verb also takes along object clitics previously incorporated into it.

Now, so far we have made explicit recourse to external Pair Merge to model clitics. We have also assumed that inheritance pairs such as < V, v^(*) > are modelled again by Pair Merge. If following Chomsky (2015), Epstein, Kitahara & Seely (2016) we explicitly endorse the derivation of inheritance pairs by internal Pair Merge, we have an elementary model for syntactic head movement as well. To obtain verb movement and subject clitic inversion in (39b), we keep following this model and assuming that the structure of reference is as in (45), cf. Epstein et al. (2016: 4). The derivational steps listed in (46) provide a useful sketch of how Set Merge and Pair Merge are interwoven in the derivation including now internal Pair Merge.

Let us begin by considering French (42a) reproduced below as (47a). Suppose we execute V-to-I movement by internal Pair Merge of the <V, v*> sequence with I. We obtain the basic sentential skeleton in (47b).

Following the general schema laid out in Table 1, clitics in modern Romance languages including French are pair merged with I. (48) represents the execution of this idea for examples like (47a). We assume that everything in the building of the object clitic string goes through as detailed in section 3, except that the locus of external Pair Merge of clitics is I. As for the subject clitic, Pair Merge with I is as detailed in section 4.3.

To better understand interrogative inversion, we must take a step back to the issue of enclisis as seen in imperatives (section 4.2) and distinguish at least two kinds of movement to C: (i) V-to-C in imperatives, yielding enclisis of object clitics, see (35) above and section 4.2; (ii) V-to-C in questions, yielding enclisis of subject clitics (but proclisis of object clitics, see (42b)).

Imperatives such as (35b), repeated in (49) for the sake of convenience are essentially subjectless. As discussed by Portner, Pak & Zanuttini (2019), the inflection of the imperative does not agree with the subject (the EPP argument) but rather with an addressee argument (which Portner, Pak & Zanuttini represent as part of the left periphery; on speaker/hearer projections see Giorgi 2009 a.o.). Therefore, v does not merge with I (Rivero 1994; Roberts 2010: ch. 5) but (directly) with C thus crossing object clitics, which remain pair merged in I, see (50).

Note that, as anticipated in section 4.1, we have now provided examples of derivations (in a modern Romance language) where clitics are not merged with the v Link, but directly with the I phase head/Link. Though this partially replaces the conclusions of section 3, it is still based on the rejection of Uniformity in that section, and the discussion of how convergence can be insured at the SEM interface. We assume that in these respects the merger structure of (50) is an extension of that already reviewed. As long as the v phase is open (as it is, at C-I only the complement of v having been shipped to the interface), saturation of open slots of the predicate v*P (via I) remains possible (and obligated as soon as argument material is merged in the form of clitics).

Interrogatives follow a different derivation, as witnessed by the absence of enclisis of object clitics and the peripheral position of DP subjects. Clitic subjects are customarily assumed to be merged in I also in inverted structures, but in fact this assumption is not justified. Consider that interrogatives prevent phrasal DP subjects from being licensed in I (interrogatives trigger so-called stylistic inversion, i.e. they require subjects to occur postverbally, or of course they can be topicalised). If I cannot license DP subjects, why should it license subject clitics?

A possible solution to the puzzle of interrogative inversion is to suppose, as in (51a), that subject clitics are pair merged in C, when C has interrogative force. Then inversion of I (along with its pair-merged clitics), takes place as shown in (51b) via internal Pair Merge with C.

This yields the order of (42b), repeated below as (52a), via the structure in (52b).³¹

As already mentioned, providing Pair Merge derivations for all the various clitic typologies illustrated in Table 1 is beyond the scope of the present article. Nevertheless, all of the general mechanisms are now in place for the execution of the research programme that Table 1 configures. Crucially, these include mechanisms for merging clitics not only in different configurations with respect to their notional phrasal counterparts, but also ‘at a distance’, namely with any phase head available in the Work Space.