1.1 Problematic properties

The interesting aspects of TCs and MPs we will focus on here stem from two main properties they both have:

In particular, we will argue that clause size plays a crucial role in allowing for the specific dependency displayed by TCs/MPs and in determining its constraints, by the selection of a Voice head with particular properties. In other words, (3a) is a consequence of (3b).

Let us now see why the dependency in TCs/MPs is so problematic. First, it is bounded, as shown by the ungrammaticality of (4)–(5), where the argument targeted by the dependency is in a doubly embedded clause.

This fact about (most) Romance TCs has been known since Kayne (1975); Aissen & Perlmutter (1976); Radford (1977); Rizzi (1982), but has not been addressed for MPs (though see Napoli 1976; Giurgea & Soare 2010 on infinitival relatives, which we believe are not exactly the same configuration as MPs - see sec. 2.2). The ungrammaticality of (4)–(5) is clear evidence that these constructions employ a cross-clausal A-dependency (like raising, control, and restructuring).³ Crucially, however, in TCs and MPs the argument targeted by the dependency with the matrix subject position must be the internal argument (IA) of a transitive verb, whereas in raising, control, and restructuring configurations the matrix subject corresponds to the external argument (EA) if the embedded verb is transitive (or unergative), or to the IA if the embedded verb is unaccusative. The relevant contrasts are shown in Catalan (6)–(8) (Italian behaves in the same way) and summarised in Table 1:⁴

The pattern in control, raising, and restructuring configuration is well-known: these dependencies are local and target the highest argument of the embedded verb, regardless of whether it is transitive or not. This captures the fact that if the embedded verb is transitive it must be passivised in order for its IA to be targeted by the dependency (as in 8a–c). Surprisingly, TCs and MPs have the opposite pattern: only IAs of transitive verbs can be extracted, despite the presence of an implicit EA (ImpEA) in the verb’s argument structure and without overt passivisation. The contrast is summarised in (9):

The schema in (9c) assumes that the relevant restructuring verbs are functional heads, following Wurmbrand (2001) and Cinque (2006), among others. The comparison in (9) does not include configurations that are sometimes labelled clause union (Rizzi 1982; Sheehan 2016; Pineda & Sheehan 2022) or lexical restructuring (Wurmbrand 2001), which have been claimed to be different from (functional) restructuring as they involve a matrix lexical verb selecting a reduced structure (so, there would be two lexical domains in the clause). This is the case, for instance, of Romance causative and perception verb constructions (Sheehan 2016). In sec. 5, we will further discuss the differences between these notions and we will argue that TCs and MPs are in fact quite similar to this latter group, rather than to (functional) restructuring verbs.

Previous literature both in the generative framework (e.g. Montalbetti et al. 1982; Giurgea & Soare 2010; Bosque & Gallego 2011) and beyond (e.g. Aguila-Multner & Crysmann 2022) has argued that TCs/MPs in Romance are indeed syntactically passive, but cannot provide a satisfactory account of why the embedded infinitive does not show participial passive morphology nor, at least in Catalan and Italian, by-phrases (cf. Zwart 2012 on Dutch). Other works have recognised that TCs show some properties resembling passives but account for the movement of the IA to the matrix in different ways: TCs involve a nominalisation according to Authier & Reed (2009) or a bare VP according to Canac Marquis (1996; cf. Wurmbrand 1994 on German). These accounts, however, fail to explain why unaccusative verbs are banned from TCs.

1.2 The present work

In light of the problems presented above, this work aims to explain the unusual pattern TCs and MPs show in (9) by focusing on the size of their verbal complement. More specifically, we will first introduce some fine-grained syntactic tests which are able to detect functional structure in the embedded clause as well as some empirical caveats which need to be made explicit when dealing with TCs and MPs (in sec. 2). The clause size tests will show that the embedded complement of TCs/MPs does not have any functional structure above Voice in both Catalan and Italian: we outline the results in sec. 3, and propose an analysis based on the idea that such a small complement clause must have a defective Voice head (rather than active or passive Voice), which is unable to assign accusative but does not license participial passive morphology nor by-phrases. As further support to our proposal, we compare the canonical TCs and MPs (which involve an infinitive without any particular marking, as in 1–2) to alternative constructions which are only productive in colloquial Catalan and involve a resumptive clitic on the infinitive:

In sec. 4, we show that when a resumptive clitic is present (signalling that regular active Voice is used), the embedded clause has much more functional structure than in the bare (i.e. clitic-less) constructions and is structurally similar to a purpose/result clause. We also highlight that this alternation between two possible sizes and, relatedly, different types of Voice is not unique to Catalan but can also be found in other Romance varieties (especially Italo-Romance dialects). Consequently, (in sec. 5) we spell out the main generalisation stemming from the comparison: defective Voice is only compatible with an extremely reduced size of the embedded clause; if the embedded clause is bigger (i.e. an IP/CP), regular active/passive Voice is found. We then provide a way of explaining this correlation, relying on the selectional properties of the various Voice heads.

2.1 Testing clause size

To obtain a clear picture of the size of the embedded clauses under study, we resorted to the many tests found in the (cartographic) literature on the clausal functional structure (Rizzi 1976; 2000; Cinque 1999; 2006; Wurmbrand 2001; 2024; Ledgeway & Lombardi 2005; Groothuis 2019; Villalba 2022). Our choice to primarily rely on the cartographic framework for the selection of the tests is not arbitrary. On the one hand, it offers a robust and well-tested methodological toolkit, particularly suited to determine the different blocks of sentence structure. On the other hand, the detailed maps drawn in the last 30 years for many areas and languages have proven a very valuable playground for testing our hypothesis on syntactic variation, language acquisition, and language pathologies (see discussions in Shlonsky 2010; Rizzi 2013; Cinque & Rizzi 2015; Rizzi & Cinque 2016). Building on this tradition, we selected the variables specified in Table 2 below to assess what functional projections are present in the embedded clauses of TCs/MPs. Nevertheless, many of the tests we employ do not necessarily need strict cartographic assumptions and have been widely used to measure clause size and functional structure outside of the cartographic framework: this is the case at least for auxiliaries, modal verbs, negation, and clitics. This implies that the results and the main theoretical insights of the present paper are no less valid (although they might need a different technical formalisation) if we do without the richly articulated functional spine proposed by cartographers. Furthermore, it could alternatively be assumed that the clauses rejecting inflectional material have impoverished or defective CP/IP layers (Paradís 2019), rather than missing, or that some operation of structure removal (Pesetsky 2023; Müller 2025) has been activated. Changing the assumptions in such a way does not pose any significant problems for our claims as far as we can tell, provided that configurations like TCs/MPs can still be distinguished from restructuring verbs (see the discussion in sec. 5.4).

Consistently with our chosen diagnostics, we assume the simplified structure in (12) for a full clause, where the various areas can be further decomposed as specified in Table 2. The labels for the various functional projections which make up the macro-areas are adapted from Cinque (1999; 2006); Rizzi (2000).

The order of the individual projections within each of the three macro-areas (CP, High IP, Low IP) does not really matter for our purposes. In other words, we are more interested in these three subdivisions than in the individual functional projections, so the fine-grained tests can be seen as a means of assessing the presence of CP, High IP and Low IP via multiple diagnostics. Our results will also show that the projections making up each area have a generally consistent pattern (e.g. either all or none of the aspectual verbs in the Low IP are grammatical).

Variables T1–T2 (fronting focus and high adverbs) are associated with a rich left periphery, so they are expected to signal the presence of a (complete) CP and the highest portion of the split IP. Variables T3–T7 (perfect/progressive auxiliary, clausal negation, non-object clitics, modal and high restructuring verbs) have been associated with the higher inflectional spine of the clause (above Voice), so they signal the presence of a full IP area. In other words, we assume with Cinque that all of this material realises functional projections. Finally, variable T8 has been argued to be linked to an Asp projection below Voice by Cinque (2006) as these verbs allow long passives. We follow this analysis for expository purposes. Alternatively, Wurmbrand (2001); Amadas (2002) argue that restructuring verbs allowing long passives are lexical: this idea is still compatible with how we use the test, since lexical verbs are of course predicted to be grammatical even in very small structures.

The novel data about the two patterns have been collected through various questionnaires that asked native speakers for grammaticality judgments. We used both written and in person questionnaires; in both cases, the speakers could give a judgement (and possibly additional comments) about the examples in their own words. In total, 6 native Catalan speakers (mostly from the Barcelona area) and 24 native Italian speakers (from various regions) were consulted. In both cases, the informants were predominantly linguists or individuals with university-level education. The Catalan speakers were asked to judge both pattern 1 and 2 so that they could compare the two constructions. We note more problematic cases next to the relevant examples. See also sec. 4.1 for more information about the productivity of the resumptive variants (i.e. pattern 2).

2.2 A note on related constructions

In this work, we do not consider a construction that may be somehow related to the ones studied: infinitival relatives (13). At the behest of a reviewer, who signals a close relationship between these relatives and MP, here we explain why.

The fist reason to exclude this construction is its distribution: while the infinitival in TCs/MPs is a complement, infinitival relatives are adjunct clauses/optional modifiers of the NP. The second reason is that the similarities with TCs/MPs are much less obvious when we look at them in some detail. For instance, Catalan infinitival relatives are perfectly fine with unaccusative verbs (14), unlike TCs/MPs (7).

Obviously, this is unexpected if we assume a common structure for all constructions.

A third reason for not treating MPs as a subspecies of infinitival relatives concerns the availability of subjects not allowed as heads of infinitival relatives. As a rule, any subject pronoun or proper name in the MP will be predicted to be impossible altogether as a modal infinitive relative:

This is obviously expected if we are dealing with two different constructions, and only the latter involves a modifier structure: neither pronouns nor proper names admit restrictive modification. Crucially, this restriction does not affect MPs, suggesting that an analysis deriving MPs from infinitival relatives with a raising DP, while not impossible to imagine, must face empirical and theoretical challenges.

Finally, MP and infinitival relatives differ with respect to low modals: while the former do not admit a low modal like potere ‘can’ (1, see also sec. 3), the latter admit them quite easily, as already pointed out by Napoli (1976); Villalba (2022).

All in all, while we do not deny that a connection may exist between TCs/MPs and infinitival relatives as suggested by the reviewer, the differences just mentioned justify our choice to leave them out of our study at this stage. This move also ensures that we are not introducing additional variables and complications that may be unrelated to our hypothesis and distort the analysis of TCs/MPs.

3.1 Data

The focus variable was consistently judged by all speakers across both languages and constructions: focus fronting was impossible in TCs/MPs in Catalan and Italian.

The same results were found for high adverbs, which were judged impossible.

A similar pattern was found with perfect (22)–(23) and progressive auxiliaries (24)–(25):

When we considered clausal negation, the judgments became sometimes less negative, but sentences were mostly judged ungrammatical:⁵

As already observed by Rizzi (2000) for Italian, the presence of non-object clitic pronouns on the infinitive is very marginal:⁶

When we considered high restructuring verbs like Cat./It. voler(e) ‘want’, Cat. mirar ‘hope’ and deixar de ‘stop’, and It. soler(e) ‘use to’ and smettere di ‘stop’, results were also quite bad, with no differences between TC and MP:

These verbs all classify as ‘high restructuring verbs’ since they generally disallow long passives (33) (showing that they realise functional heads above Voice, Wurmbrand 2001; Amadas 2002; Cinque 2006), as reported by Paradís (2019: 388) for Catalan and by Rizzi (1976); Cinque (2006: 66ff) for Italian.⁷

Modal verbs, which generally disallow long passives too,⁸ display the same situation as high restructuring verbs, since they were judged as very unnatural:

Up to this point, all tests were negative, suggesting that TCs/MPs lack a full sentence structure in the embedded clause. When we considered lower landmarks, judgments reversed. For example, low aspectual restructuring verbs were judged possible (vs high restructuring ones, above).

Low restructuring verbs (38) are lower than Voice, as shown by the possibility of long passives (Rizzi 1976; Amadas 2002; Cinque 2006):

Therefore, only restructuring verbs which allow long passives are possible in TCs/MPs. Thus, only material merged below Voice seems to be grammatical in Catalan and Italian TCs/MPs, where the infinitive does not have any marking.

3.2 Discussion and analysis

In light of these results, we can immediately notice that there are no significant differences between Catalan and Italian, despite using distinct introductory elements (de and per in Catalan, and da in Italian). Therefore, we propose that the complement clause has the same syntax across the two constructions and the two languages. More precisely, the syntactic tests reveal that the infinitive does not project any functional structure above Voice: the embeddded clause must then be just a VoiceP (39).

An immediate consequence of this observation is that the introductory elements of TCs/MPs (namely It. da and Cat. de/per) are not C heads, at least if we assume a rigid functional hierarchy in the clausal structure, which entails that if I is missing, C must be missing too.⁹ This situation is similar to Romance restructuring/clause-union configurations, which also involve preposition-like introductory elements not corresponding to a C layer (e.g. It. provare a ‘try’, Cat. mirar de ‘hope’): according to Cinque (2006); Ledgeway (2016); Groothuis (2019), these elements can simply emerge from the combination of a predicate with its complement, and are just there for selectional/idiosyncratic reasons; for this reason, clauses introduced by these elements can be of different sizes, not necessarily CPs. More specifically, Cinque (2006: 45) states that these elements can be analysed as Ps, which are needed in order for an infinitive to be selected by certain verbs. Alternatively, Wurmbrand (2001: 109–110) concludes that since the German infinitive introducer zu, often found in restructuring/clause-union configurations, “bears no semantic content and does not seem to fulfil any syntactic function, it is essentially ‘invisible”’ and its distribution depends on an arbitrary selectional property of the matrix predicate. An additional alternative could be analysing the prepositional element as part of the main predicate: an analysis along these lines is proposed by Haider (2024: 14), according to whom these particles “can be regarded virtually like a particle of a particle verb.” For concreteness, we will follow Wurmbrand and not represent the introductory elements in the structure –but our proposal is compatible with all the above-mentioned technical solutions.

We can now turn our attention to the Voice properties of TCs/MPs. Voice in TCs/MPs does not fully resemble either active or passive Voice: on the one hand, it cannot be active because the IA is promoted via an A-dependency (cf. the discussion in sec. 1); on the other hand, it cannot be passive, as passive morphology and by-phrases are disallowed (40)–(41).¹⁰

This means that proposals like Giurgea & Soare (2010) cannot capture TCs/MPs in Catalan and Italian well, although we do share the idea that TCs/MPs are similar to passives in many respects: Giurgea and Soare report that some speakers accept indefinite or inanimate by-phrases in Romanian/French infinitival relatives (which look similar to TCs/MPs) so they claim that these constructions must involve a syntactically passive infinitive which exceptionally does not show passive morphology. Since both passive morphology and by-phrases (even when they are indefinite or inanimate)¹¹ are ungrammatical in Catalan and Italian TCs/MPs, we need to find a different solution to account for why neither of these properties normally associated to passives is found on TCs/MPs.

Furthermore, we can assess the status of the implicit EA of the embedded verb. If the implicit EA were a PRO or pro, it could be modified by a depictive like enfadats (‘angry’) (Landau 2010). This is shown by the impersonal construction in (42a–b), which contrasts with a participial passive (42c), where the implicit EA has a weaker syntactic status (if any):

The same holds in Italian:

TCs and MPs interestingly pattern like passives, which shows their implicit EA is not PRO or pro:

Once again, this suggests that the infinitive does not have active Voice, which would require a PRO/pro (i.e. a DP) as EA.¹²

Let us now spell out our assumptions on Voice alternations for the languages at issue. First, we outline in (45)–(46) the syntactic properties of active and passive Voice (cf. Legate 2021 for discussion on the main properties of canonical passives):

The way these properties should be formalised has been discussed extensively in the literature. In particular, it is not clear whether one or multiple heads in the Voice domain are responsible for the differences above, and what the division of labour might be (see Wood & Tyler 2023; Kallulli & Roberts 2025 for discussion). This is important because cross-linguistically there are many instances of passive-like constructions which show interesting contrasts with canonical passives, in particular when they lack one or more of the properties outlined in (46) (Legate 2021). Participial passives in Italian (47a) and Catalan (47b) show all the properties in (46).

On the other hand, in this work we do not consider se-passives, as they show different syntactic properties from participial passives (see e.g. Dobrovie-Sorin 2017 for a review).

For the purposes of this paper, we simply assume that the various passive properties exhibited by participial passives in contrast with the properties of active transitive constructions all rely on the head Voice, depending on the following features (mostly following Wurmbrand 2016; Wurmbrand & Shimamura 2017; Bryant et al. 2023 and references therein, see also Schäfer 2008 for the [AG] feature):

The differences between active and passive Voice can then be encoded as follows:

To explain the unusual behaviour of TCs/MPs we propose that Voice in their verbal complement is defective in the following way: it cannot assign accusative (which is why the IA can be promoted) and it encodes a passive-like implicit EA that cannot be merged as a DP or as a by-phrase (which is why only transitive verbs are allowed, but there is never an explicit or a PRO/pro agent for the verb), but it can only have default active-looking morphology. Under our assumptions, this amounts to only having the [AG] feature, but neither [PART-PASS] nor [ACC] - making this Voice head defective, as in (50).

To recapitulate, in the active (49a), Voice assigns accusative (a property that we encode as an [ACC] feature on Voice) and introduces an agent in the semantics (encoded as an [AG] feature), which in this case corresponds to a DP in Spec, VoiceP (i.e. the EA), in compliance with Burzio’s generalisation. In the passive (49b), Voice does not assign accusative and does not introduce a DP in its Specifier (again complying with Burzio’s generalisation), but it still encodes an agent (which can be overt as a by-phrase) and therefore has the [AG] feature, as well as a [PART-PASS] feature activating passive morphology and licensing the optional by-phrase. Finally, when Voice is defective (50), it still encodes an agent as [AG] without being able to assign accusative, but it cannot realise such an agent overtly as a by-phrase and does not require passive morphology.¹³

The idea that reduced complement clauses involve a somewhat defective Voice (or v) head has been put forth several times in previous literature, with independent motivations. For instance, Giurgea & Soare (2010) and Pitteroff (2015) propose that a Voice head with passive syntax but no passive morphology is involved in Romance infinitival relatives and TCs, and German let-middles, respectively; similarly, Manzini (2022) argues that Romance causatives select a defective vP with a passive-like (“ergative”) alignment and regular infinitive morphology (cf. Folli & Harley 2007); finally, Wurmbrand (2016); Wurmbrand & Shimamura (2017); Bryant et al. (2023) claim that in many languages reduced complement clauses (like the ones found in German long passives) have an underspecified Voice head possibly surfacing with default morphology which cannot assign accusative nor merge an overt EA (cf. Bosque & Gallego 2011 for a similar suggestion on Romance long passives and similar constructions). The formalisation and the label associated with this type of Voice varies depending on the author and on the details of the specific constructions at issue.

For our purposes, the simple implementation in (49)–(50) will suffice, even if there are other possible ways of representing causative semantics, the implicit agent, and participial passive morphology. A possible alternative could rely on two layers in the Voice domain (as in e.g. Collins 2005; 2024; Sigurdsson 2011; Merchant 2013; Bruening 2013; Alexiadou et al. 2015): in a passive, the lower layer (v) is responsible for agentive semantics, while the higher layer (Voice) introduces participial morphology and licenses by-phrases, as shown in (51a). The pattern we observed in Italian/Catalan TCs/MPs can then be explained by positing that the higher layer is missing in those constructions, consistently with the clause size tests in the previous subsection.¹⁴

The idea expressed by the formalisations in (51b) and (50) is the same: in TCs/MPs, a component of the canonical passive syntax is missing, whether the missing component is a feature on Voice or a layer in the Voice/v domain. At the same time, unaccusative verbs are always excluded (under either set of assumptions) because they are incompatible with the mandatory presence of a head introducing an agent in the semantics (encoded by [AG]), i.e. the same reason why they cannot be passivised. We see the variant in (49)–(50) as more straightforward for the purposes of this paper so we will use it in the remaining part of the paper, but the proposed account is compatible with the formalisation in (51a–b), too.

It is now easy to see how defective Voice can derive the unusual argument realisation pattern found in TCs/MPs: as (52c) and (53c)¹⁵ show, defective Voice ensures that an implicit agent is present in the interpretation (and therefore the verb has to be transitive) without being a syntactic intervener for A-movement of the IA –like in the passive. However, the agent cannot surface as a by-phrase and there cannot be participial passive morphology (sec. 1).

For MPs, we assume that the main verb realises a V head with a modal meaning. That the verb is not functional/just an auxiliary is independently confirmed by the impossibility of clitic climbing in MPs (54), contrasting with the participial passive auxiliary (55).

A reviewer challenges this assumption by arguing that a normal copula is used in (54) due to the fact that a similar infinitival clause can appear as an infinitival relative in adnominal position, and claims that the copula in MPs is functional:

We have already discussed why we think that infinitival relatives (like 56) have syntactic differences with TCs/MPs in sec. 2.1, which is why we do not propose an automatic extension of our analysis to those configurations. The exclusion of infinitival relatives entails that the problem raised by the reviewer would go away on its own, as the parallelism does not hold. Besides those arguments, even if we accepted that the standard copula were used in MPs, then we could assume that it is nevertheless a (semi-)lexical unaccusative verb (v or V), following a part of the literature (e.g. Wurmbrand 1994: 98 on German MPs, Gallego & Uriagereka 2016 on Spanish estar, see also Arche et al. 2019 for a review). It would be harder to explain, however, what introduces the deontic/possibility interpretation in this case. Alternatively, we could assume that the modal meaning is actually conveyed by a null adjective in MPs (cf. Kayne 2014 on the English be to periphrasis), which would be a lexical category, while the verb can be treated as the normal copula (whether lexical or functional). This option is particularly tempting for Italian, as this assumption would make MPs a straightforward subcase of TCs with a null adjective, and could also work for Catalan (even if it uses a different preposition).¹⁶ What matters most for the purposes of this paper is that, on a par with TCs, the infinitival complement of MPs is not selected by a functional category (see discussion in sec. 5).

If we accept the proposal in (52)–(53), the issue is then how we can constrain the distribution of defective Voice: we will show that it is indeed the case (at least in Catalan and Italian) that defective Voice is only available in extremely reduced clauses, namely VoiceP complements, whereas active/passive Voice requires at least an IP. We will explain this as a constraint on selection in sec. 5.

4.1 Data

Focus fronting is very marginal or outright excluded in both constructions, just as in their bare counterparts:

The same results were found for high adverbs, which where judged impossible, just as happened with bare TCs and MPS:

When we considered perfect and progressive auxiliaries, the sentences with a resumptive TC were judged grammatical, in sharp contrast with the non-resumptive options.

On the other hand, resumptive MPs are not acceptable with such auxiliaries: we think this is due to an independent incompatibility between the inherent modal/aspectual meaning of the estar per periphrasis and the aspectual value of the auxiliaries.

As González & Martín Gómez (2019) argue for Spanish, estar periphrases semantically encode a precondition for a future event to happen, which necessarily gives the construction a prospective meaning. If this is the case in (68), we can expect that perfect and progressive auxiliaries will be forbidden independently of the size of the clause involved. In other words, the ungrammaticality of these examples is not due to the lack of a full IP structure, but a side effect of the incompatibility of estar periphrases with perfect/progressive aspect.

When we considered clausal negation, the judgments were much more positive than with clitic-less variants in this case too:

Similarly, the presence of non-object clitics on the infinitive was judged fully grammatical if a resumptive object clitic is also present:

When we considered high restructuring verbs like voler ‘want’, deixar de ‘stop’, or mirar de ‘hope’, results were ameliorated by the clitic for both resumptive TCs and MPs, in contrast with their bare counterparts.

This pattern was reproduced with modal verbs, which were possible (although sometimes still not fully natural, probably due to the presence of two modal elements) in both constructions with the resumptive clitic:

When we considered lower landmarks of clause structure, judgments were also fine –just like they were in the non-resumptive options. This is the case of low restructuring verbs:

Thus, in this test we found no difference between bare and clitic resumptive patterns.

4.2 Discussion and analysis

In light of these results, we can immediately notice that the complement clause of resumptive TCs and MPs is structurally bigger than that of their clitic-less counterparts. While bare TCs and MPs were argued to have a VoiceP complement, resumptive TCs and MPs involve at least a full IP and possibly FinP, just lacking the higher left periphery (like most non-finite clauses; e.g. Villalba 2019; 2022 on infinitival relatives). The crucial tests involved high restructuring verbs, modals, clausal negation, and aspectual auxiliaries: they were judged grammatical for resumptive TCs and MPs, whereas they were clearly bad for their non-resumptive counterparts (sec. 3). Obviously, all tests involving a lower part of the structure were equally good in both patterns. Therefore, we claim that the embedded clause of resumptive TCs and MPs has the functional projections in (74).

The only discordant note was the impossibility of perfect and progressive auxiliaries in Catalan MPs (68), which was a side effect of the aspectual restrictions imposed by the verb estar.

As for the Voice properties of resumptive TCs and MPs, these are much less unusual than in bare TCs and MPs: active Voice is clearly involved, as signalled by the object clitic. Furthermore, the implicit EA of the infinitive is projected as a (null) DP in this case (as expected with active Voice), while it is not in the bare variants. This is evidenced by the acceptability of depictives:

Further evidence that the EA is fully projected comes from its alternation with postverbal lexical subjects, which in Catalan are only available in a few types of infinitives (Rigau 1995; Mensching 1999; Sitaridou 2002), including result (76a) and subject clauses (76b), which have been consequently analysed as merging a pro when there is no lexical subject (see in particular Rigau 1995).¹⁹ In other words, if a null implicit EA can alternate with postverbal subjects in an infinitive, then it is pro.

This option is indeed available in both resumptive TCs and MPs, as shown in (77a-b):

This is a strong argument that these constructions just involve regular active Voice in the embedded clause.

Therefore, the analysis we propose for the embedded clause of resumptive TCs/MPs is the following:²⁰

We do not make specific technical assumptions on how the resumptive dependency between the object clitic and the matrix subject is established, as it does not matter for our claims. A reviewer correctly points out that if we assume that the matrix subject DP is base-generated in (78), then the adjective would need to assign it a θ-role, differently from the non-resumptive variant. We do not think this would be a problem. In fact, something very similar has been proposed by Remberger (2024) for Sardinian fàchere a… when it means ‘it is possible to…’, which alternates between a raising variant (non-thematic subject) and a resumptive variant (thematic subject).

Furthermore, in our investigation, we found inter-speaker variation as to whether, when the resumptive clitic is present, the dependency can cross a clause boundary. Only some informants accepted (79).

To the extent that (79) is acceptable (in stark contrast with the non-resumptive variant 4a), it resembles English TCs, which are also unbounded, and suggests an analysis of the resumptive pronoun as bound via an A’-dependency (Chomsky 1977; 1981; Hicks 2009).

In conclusion, resumptive TCs and MPs are derived as in (80) and (81) respectively:²¹

5.1 The empirical generalisation

The results of the tests about clause size (T1–T8) and the status of the implicit EA (T9–T10) are summarised in Table 3. These highlight important differences between the two variants of TCs/MPs under both aspects.

Among the most striking data in Table 3 are the minimal contrasts depending on the presence of the clitic when IP material above Voice and below Fin is included. A resumptive clitic makes such sentences grammatical. Similarly, only TCs/MPs with a resumptive object clitic pass the tests associated with the syntactic presence of a null EA (as a PRO or, more probably, a pro). This ameliorating effect highlights the connection between the presence of an object clitic, active Voice (Voice_{[AG, ACC]}), and a rich functional structure. If the clitic is absent, TCs/MPs involve a much smaller verbal complement (up to VoiceP) and what we labelled defective Voice (Voice_[AG]). The two proposed structures for the embedded clause of bare TCs/MPs (in Catalan and Italian) and resumptive TCs/MPs (colloquial Catalan only) are repeated below, respectively:

Similar results were found by Russo Cardona (2023) in certain Italo-Romance dialects, which present two main types of TCs differing in size and Voice properties. For instance, consider the following pair from Neapolitan (83). As shown by the optionality of the clitic in (83a), both the plain and the resumptive TC are possible in this dialect. However, when functional material (in the IP projections above Voice), like a dative clitic, is added (83b), only the resumptive variant (which relies on standard active Voice) is possible. This is exactly the same pattern we found in Catalan.²²

As a consequence, the empirical generalisation about the interaction of clause size and Voice in TCs/MPs emerging from this state of affairs is (84):

In other words, active Voice is only found in embedded clauses with a rich functional structure, whereas defective Voice is found in smaller clauses projecting only up to Voice. Thus, the amount of functional structure on top of Voice correlates with what kind of Voice can be merged, on the basis of the TC/MP data.²³

Finally, there is also another interesting correlation emerging from our data:

This is very much in line with a recent claim by Satik (2024), who independently proposes that the smaller a clause is, the more defective its subject has to be. Under our account, (85) will be easily derivable from the constraints on the different types of Voice, which we outline below.

5.2 Selectional constraints on Voice

How do we explain the link between Voice properties and clause size? Why can’t defective Voice appear in bigger clauses, and why can’t active Voice be found in smaller ones, in the context of TCs/MPs? Even if it is intuitive that defective Voice might be associated with a poorer functional spine, nothing in our system so far derives this distribution, and nothing prevents active (or passive) Voice from being selected in VoiceP complements.

Similarly, various previous works relying on some similar kind of defective Voice do not specify under what syntactic conditions this type of Voice can be merged and why it cannot be dominated by a rich functional structure, and/or why there is no competition with regular active Voice. For instance, Pitteroff (2015) claims that German let-middles have a canonical passive Voice head in the infinitival clause, which lacks a functional head above Voice hosting passive morphology, so that the infinitive shows up with default morphology (86).

Although Pitteroff (2015: 48) correctly notes that this kind of Voice is always found in reduced complement clauses (cf. Wurmbrand 2001 et seq.) which lack the projections that would be realised by passive morphology, it remains unclear under this account why active Voice is not possible in these reduced complement clauses. Furthermore, Pitteroff reports that by-phrases are possible in let-middles, which supports his analysis of the construction as a genuine passive (cf. the discussion in sec. 3.3 on Giurgea & Soare’s 2010 proposal on Romanian/French infinitival relatives). This is an empirical difference with the constructions at issue here, so we cannot simply extend this type of analysis to TCs/MPs.

Other works (sometimes implicitly) assume that their equivalent to what we call defective Voice has to be in a clause that is small enough. For example, Manzini (2022) proposes that Romance causatives (e.g. 87 in Italian) involve a defective vP with “ergative alignment” (so the IA is the highest argument, and the EA is demoted or oblique from the start) like in a passive (Manzini 2017).

However, the distribution of the ergative v head remains somewhat unclear, even more so if it is the same in both causatives and passives, which have different morphological marking, and different possibilities regarding the overt realisation of the EA as a PP (see Burzio 1986; Guasti 1993 for details). So, these existing accounts, though they may work well for the individual constructions they are concerned with, do not provide a satisfactory explanation for the pattern we have identified in the present paper.

Furthermore, an explanation that is exclusively based on clause size is also bound to fail. In other words, (84) is not accurate enough, once we take other configurations into account: it cannot be the case that each type of Voice (defective vs active/passive) maps to a specific size of the embedded clause. For instance, as discussed in sec. 1, modal restructuring verbs like Cat. poder/It. potere ‘can’ take a VoiceP complement, yet they allow passive morphology on the lower verb (88a–b) - unlike TCs/MPs - which implies they do not have defective Voice even if the embedded verb does not have its own functional structure.

Thus, restructuring configurations like (88) are a counterexample to the claim that a VoiceP complement is always associated with defective Voice, which would be the most immediate solution of the puzzle in (84).

A suggestion towards a more refined explanation for the distribution of the proposed types of Voice can be found in the existing literature. According to Wurmbrand (2016: 269–271) (see also Wurmbrand & Shimamura 2017; Bryant et al. 2023 for more recent implementations), an underspecified head is found in what Wurmbrand calls “Voice restructuring” configurations cross-linguistically: under her account, a special, underspecified Voice head must be selected directly by the matrix verb, so that its distribution is constrained by a lexical specification present only on certain verbs; so, it follows that there cannot be an intervening T layer between the selecting predicate and the special Voice/v head, as lexical selection must be unmediated.

Expanding on this idea, we submit that the lexically-specified selectional constraints in (89) constrain the distribution of the different types of Voice (at least in the languages we are considering, but see Russo Cardona 2023 for the formulation of this idea about Italo-Romance dialects).²⁴

The idea that active/passive Voice can only be selected by a functional head (i.e. 89a), rather than directly by a lexical predicate predicts that in embedded clauses real active Voice is found only if the embedded clause is at least as big as an IP. This straightforwardly captures the fact that active Voice is only observed in resumptive TCs/MPs,²⁵ which have a FinP embedded clause, rather than in bare TCs/MPs, where Voice is not dominated by an I head. Furthermore, restructuring configurations as in (88) now become easy to derive: since the matrix verb is an I head in such cases, as we have assumed throughout the present paper (following Picallo 1990; Wurmbrand 2001; Cinque 2006), the VoiceP projected by the lower verb is selected by an I head; therefore, it can be active or passive, but not defective.

Defective Voice, on the other hand, can only be directly selected by specific lexical predicates (89b). Adjectives allowing the TC, and the verbs estar (per) in Catalan and essere (da) in Italian, qualify as such predicates, meaning that their lexical entry specifies that they can select a (defective) VoiceP complement. Of course this does not mean that they can only select a VoiceP, but just that they have this possibility - unlike other lexical predicates. We think this explanation is very desirable, as it leaves a lot of space for nano-variation (in the sense of Roberts 2019): languages, and possibly even individual grammars, are likely to differ a lot in which items can take a VoiceP complement. Thus, it is easy to explain, for instance, the contrast in (90) in Italian simply as an arbitrary lexical difference between the two adjectives.

Furthermore, the lack of a resumptive variant for TCs/MPs in standard Italian can similarly be accounted for by simply appealing to a lexically encoded, idiosyncratic difference: Italian tough-adjectives cannot generally select/be modified by a FinP embedded clause headed by da.

5.3 Predictions

The predictions of (89) naturally extend beyond TCs/MPs. First, all control and raising complement clauses are predicted to disallow defective Voice, and to allow active or passive Voice, as they always have at least an IP layer above Voice.²⁶ This explains the facts shown in sec. 1 about these types of complements: here, regular passivisation of the embedded verb is needed to establish the relevant dependency between the IA and the matrix subject position.

A more interesting case, which will not cover exhaustively, is that of perception verbs. These constructions in Catalan and Italian seem to confirm the generalisation we put forth, and thus obey the constraints in (89). For instance, Sheehan & Cyrino (2024) argue that in English and Brazilian Portuguese, perception verbs can take an active/passive VoiceP complement, as shown in (91a–b), adapted from Sheehan & Cyrino (2024):

According to our claims, the same cannot be possible in Catalan and Italian, since perception verbs are lexical predicates selecting a VoiceP clause which allows the active/passive alternation. Our analysis instead predicts that in Catalan and Italian perception verbs should either take a VoiceP complement with defective Voice, or a bigger complement (at least an IP) with active/passive Voice. In fact, both options seem to be possible in these languages (as also claimed by Ciutescu 2013, among many others). In both Catalan and Italian, perception verbs can involve a highly reduced clause (no negation, no clitics), with a demoted/oblique agent²⁷ (92) but no passive morphology: this is the defective VoiceP complement option (cf. faire-à causatives).