3.1 Core assumptions

My core proposal is based on five assumptions, which I will elaborate on here.

Assumption no. 1 echoes an old observation that goes back to Frege: propositional material does not take scope over force. That is, a negative question is not a negation of a question, and an NI is not a negation of a command (e.g., Han 2000, Han 2001, Zeijlstra 2006, and Zeijlstra 2013):

This is a significant fact about natural language and underlies many analyses of NIs. Han 2000: 46, for example, observes that “NIs are unavailable in some languages because they instantiate a syntactic configuration in which negation would take scope over the imperative operator in C.”

Assumption no. 2 is that the perfective NI in (12a) violates formal properties of a syntactic mechanism; that is, it is not a result of a purely semantic/interpretative constraint. Negative perfective commands can easily be expressed with structures like (12b).

This is in line with the large body of literature that has argued that the ban on NIs is essentially syntactic; examples include Schmerling 1982, Zanuttini 1991, Platzack & Rosengren 1994, Rivero 1994, Zanuttini 1994a, Zanuttini 1994b, Rivero & Terzi 1995, Han 1998, Platzack & Rosengren 1998, Han 2000, and Han 2001.

Assumption no. 3 is that BCMS imperatives involve a complex probe, which consists of two unvalued features: (a) i-mood[__]: an interpretable but unvalued imperative force feature, which needs to be valued via Agree by the uninterpretable but valued version of the same feature, u-mood[imp] (e.g., Pesetsky & Torrego 2007), and (b) u-addressee[__]: an uninterpretable, unvalued ϕ-probe, which needs to agree with a second person argument. I need to clarify some aspects of this proposal.

The first feature, i-mood[__], is an example of the attribute: value feature system in which imperative is one of the values that the attribute mood may have (e.g., Adger & Svenonius 2011). This feature interacts with other probes on other functional heads and does not search for nominal arguments or assign case. The interpretable–uninterpretable contrast is used as in Pesetsky & Torrego 2007; I will spell it out in a moment.

The second feature, u-addressee[__], is different: it can only be valued by a particular element, namely a second person subject. This is in spirit like Béjar & Rezac 2009’s feature system but is formulated differently. Béjar & Rezac’s approach has articulated probes, which can be specified to search for a certain type of argument. Thus, the more highly articulated the probe is the more highly specified a DP must be to match all of the probe’s features; for example, Béjar & Rezac recognize a flat probe, a partially articulated probe, and a fully articulated probe. Consequently, a fully articulated probe can be satisfied only by a certain, very specific type of argument (e.g., first or second person only, depending on the language). The feature that I proposed above (u-addressee[____]) does the same job in the sense that it restricts the type of argument that can satisfy it, but it is formally more similar to what is found in Nevins 2007, which uses a version of Multiple Agree to explain different varieties of Person–Case Constraint (PCC) effects.² In Nevins’s system the search can be relativized to arguments with a certain type of marked feature (e.g., participant), which again puts a limit on what kind of arguments can satisfy the search. On his approach, for instance, “the strong PCC results from a Probe that is specifically looking for contrastive values of [Author]” (Nevins 2007: 296). Similarly, the feature I propose, u-addressee[__], looks for a second person argument.

But I also assume that this feature is uninterpretable (in addition to being unvalued) following the standard Agree approach to case agreement in the subject position (e.g., Chomsky 2000 and Chomsky 2001). Namely, to be deleted before LF (as an uninterpretable feature) u-addressee[__] must be valued by an argument in its domain that has interpretable ϕ-features but that lacks case and is, therefore, visible for search. After Agree takes place, u-addressee[__] is valued (which I will represent as u-addressee[val] throughout the article) and deleted before LF, while the subject is assigned nominative.

The one complication that I add to this standard story is that in the case of imperatives the ϕ-probe in T is limited to search only for second person arguments. In other words, only second person pronouns may satisfy this probe and be assigned case (see below for more details). The fact that third person referential nominals and quantifiers cannot be subjects of imperatives in BCMS is then directly explained under my proposals: these expressions, not being specified for a second person feature, cannot value u-addressee[__], which in turn cannot be deleted before LF. At the same time, the nominal in the subject position cannot be assigned case, so the derivation crashes.

As we already saw in section 1, a simple fact about BCMS is that the verb shows special imperative morphology, but crucially this is not where we interpret the imperative force. An effective way of getting around this problem is to assume that the verb moves all the way to C to license imperative force (e.g., Rivero 1994 and Rivero & Terzi 1995 for Romance). But this can be hardly maintained for BCMS, which shows little evidence of verb movement to C. In fact, examples like (13) indicate that the verb stays very low; here the negated imperative verb is preceded by an adverb nikako ‘by no means’ and the second position clitic mu ‘him,’ and even another type of adverb like sutra ‘tomorrow’ (or slučajno ‘accidentally’) may be added between the clitic and negation.

On most accounts both adverbs and second position clitics are located in the left periphery of the clause (C domain). Many accounts in fact take second position clitics to be located in C (e.g., Franks & Progovac 1994, Wilder & Ćavar 1994, Progovac 1996, and Tomić 1996). This strongly suggests that the verb in (13) does not move to C. Consequently, it cannot be the case that it acquires imperative morphology by moving to C. But then we have the following conundrum: we interpret (13) as a (negative) command, but the imperative force cannot be interpreted compositionally where we see the imperative morphology. Force is interpreted above the propositional material, and the imperative verb in (13) is embedded deep within it.

Pesetsky & Torrego 2007 observes a similar situation in English tense agreement: although the main verb carries past and present tense agreement, it does not acquire it by moving to T (unlike in French; e.g., Emonds 1978 and Pollock 1989). Thus, although tense is interpreted in T, tense morphology is on the verb in a lower position. We see the same type of problem in (13). I will adopt Pesetsky & Torrego’s feature sharing approach, which assumes the four way feature typology given in (14). In this system, in addition to uninterpretable unvalued (u-F[__]) and interpretable valued (i-F[val]) features (e.g., Chomsky 2000 and Chomsky 2001), there are interpretable unvalued (i-F[__]) and uninterpretable valued (u-F[val]) features.

An example of uninterpretable valued features could be grammatical gender on inanimate nouns; although an inanimate noun may have a grammatical gender value (e.g., feminine), this is not interpreted (see Bošković 2009 and Wurmbrand 2012, among others, for more details). An example of interpretable unvalued features would be tense in English. Pesetsky & Torrego 2007 posits an i-T[__], as in (15b), which from T probes the uninterpretable but valued version of the same feature located on walked.

Pesetsky & Torrego also assume Brody 1997’s Thesis of Radical Interpretability, according to which each feature must receive a semantic interpretation in some syntactic location. Thus, the structure in (15b) will be uninterpretable, unless i-T[__] is valued. In this spirit, I propose that the BCMS imperative probe includes i-mood[__]: an interpretable but unvalued imperative feature, which needs to be valued via Agree by the uninterpretable but valued version of the same feature u-mood[imp] located in a lower position.

But the BCMS imperative probe also includes u-addressee[__]: an uninterpretable, unvalued ϕ-probe, which needs to agree with a second person argument. The imperative verb can have three forms, as shown in (16)—second person singular, second person plural, and first person plural.

Significantly, (16c) can only be interpreted as inclusive, which shows that this probe is searching for a [+addressee] argument. The question then is, what is the nature of vocative expressions, which very often accompany imperatives? Unlike in English, there is a special vocative ending in BCMS for many nominals, which makes identifying the vocative expression straightforward. For instance, in (17), Milane is the vocative form of the name Milan, which seemingly functions as the second person subject.

But as (17) shows, the second person subject is optional and can be made overt. The vocative expression is then really licensed by the presence of the addressee, regardless of the clause type or the syntactic position of the addressee argument. In (18a) Milane marks the addressee to which the question is posed, and in (18b) it marks the addressee in the object position of a declarative sentence. Note also that the vocative expression can appear in a variety of positions, as (18c) shows.

It seems safe to conclude that vocatives are not subjects of imperatives but simply mark the presence of an addressee argument in the subject position in imperatives (for discussion of vocatives in English as non-subjects see Downing 1969 and Davies 1986, among others).³ This predicts that any non-second person nominative subject will be ungrammatical as the subject of an imperative in BCMS, which is correct. The sentence in (19a) is ungrammatical with the nominative form Milan as the subject, and (19b) is ungrammatical with the nominative form svako ‘everyone’ as the subject.

Notice that the quantifier svi ‘all’ can be located in what appears to be the subject position, as in (20). The crucial difference, however, is that svi is an adjectival modifier (it agrees with the noun/pronoun in gender and case; e.g., svih vas is ‘all of you’ in genitive) and the underlying second person nominative subject can always be unearthed.

Quantifiers like svako ‘everyone,’ on the other hand, are incompatible with the nominative pronoun vi and must combine with a PP od vas ‘from you,’ in which the pronoun is genitive. This situation is exactly predicted by my analysis—imperatives fail every time one can show that there is no underlying second person nominative pronoun in the subject position.

A separate question is why quantifiers like everyone can function as subjects of imperatives in English. Zanuttini 2008 shows that quantificational subjects are possible in English imperatives but that their interpretation is restricted to ranging over the set of addressees, unlike in questions and declaratives. One empirical contrast that supports this is that quantificational subjects can bind a second person pronoun only in imperatives (e.g., Zanuttini 2008: 190–191):

Zanuttini also observes that something like (21a) is possible only if everyone can be followed by of you, indicating that the quantificational subject of imperatives includes a second person pronoun; this is very similar to my analysis of (20).

The question then is why (20b) is ungrammatical, unlike English (21a), even if PP od vas ‘from you’ is added after svako. In (21a), the hidden second person pronoun is identified via binding/co-indexation; there is no specific verbal form in English reserved only for imperatives, as in BCMS (see (16)). This difference in morphological paradigms between English and BCMS is envisioned here as arising from the difference in the types of probes these languages have in imperatives. As discussed above, I suggest that BCMS has a probe that can be licensed only by an addressee argument, so consequently only an addressee argument can have nominative case and be the subject of an imperative, which is all clearly indicated in the BCMS verbal morphology. Since English lacks this morphology, the English imperative ϕ-probe is not specified to look for an addressee argument per se and thus can assign nominative to arguments other than second person (like everyone in (21a)). Thus, the fact that English imperative subjects still involve a second person pronoun might not be due to specifications built into the ϕ-probe but follow from some other syntactic properties of imperatives in English.⁴

Assumption no. 4 is that at least some agreement is feature sharing (Frampton & Gutmann 2006, Pesetsky & Torrego 2007, Kratzer 2009, Abels 2012, Ackema & Neelman 2013, and Keine to appear, among others):

An advantage of this model that is important for the present purposes is that it enables so-called parasitic agreement. For example, A and B may agree with each other, producing a single instance of some feature F, which is now shared between them and is still unvalued. But if A subsequently agrees with some third element C that bears a valued version of the same feature F, this will lead to feature sharing between A, B, and C, which will now all be valued for F. This advantage will become clearer in the next subsections, where I present derivations of structures.

Finally, assumption no. 5 is that perfective verbs in BCMS are located in the complement of the external argument–introducing head v whereas imperfective verbs are located higher than v (e.g., Svenonius 2004 and Travis 2010:

In section 3.6, I will provide empirical evidence for this claim based on reflexive passivization, nominalizations, and idioms.

3.2 Positive imperatives: derivations

Consider first declarative sentences in BCMS with periphrastic verbal forms. The future tense consists of the auxiliary verb će ‘will’ followed by either an infinitive, as in (24a), or, as in (24b), the so-called da clause (e.g., Browne 1987, Progovac 1993, Stjepanović 2004, Tomić 2004, Veselinović 2019, and Todorović & Wurmbrand 2020), which has present tense inflection but which is in terms of tense interpretation dependent on some higher element. The simple past example in (24c) consists of the auxiliary verb je ‘is’ and the participle form.

For all three examples I will assume that the auxiliary verb is in T. This will also become relevant in the discussion of analytic NIs.

The grammatical positive imperative examples from (5a) are repeated here.

Since the imperative probe includes u-addr[__], which is a ϕ-feature and a property of the T domain, it probes from T. I will also assume that the uninterpretable version of i-mood[__] is located either in Asp_imfv or Asp_pfv. Thus, these features are properties of functional heads: the interpretable, unvalued i-mood[__] can be located in C or T, whereas u-mood[imp] is either in Asp_imfv or in Asp_pfv. I will assume in addition that imperfective verbs move to Asp_imfv and perfective verbs to Asp_pfv. Consider then the derivation in (26), where the uninterpretable version is in Asp_pfv.

First, u-addr[__] in T probes the addressee subject and gets valued (26a). Then i-mood[____] probes u-mood[imp] in Asp_pfv and gets valued (26b). As a result u-addr[__] in Asp_pfv is also valued, due to (parasitic) feature sharing. Finally, u-addr[val] in Asp_pfv is spelled out, but the presence of u-mood[imp] conditions its morphological form, so that the second person inflection in imperatives is different from that in declaratives and questions.

In (27), u-mood[imp] is in Asp_imfv. First, u-addr[__] in Asp_imfv probes the subject and gets valued. Then i-mood[__] in T probes u-mood[imp] in Asp_imfv and gets valued. u-addr[__] in T is also valued, due to feature sharing, and u-addr[val] in Asp_imfv is spelled out.

3.3 NIs: derivations

The negative forms of the declarative sentences in (24) are as follows.

Given the distribution of negation in these examples I propose that NegP projects above TP (e.g., Zanuttini 1991):

The presence of NegP does not affect the Agree system in constructions like (28), but it does in the case of imperatives. Since the imperative force cannot be in the scope of negation in the logical forms of sentences, I propose that in NIs, the imperative probe is located in C (as a type of a last resort; see section 3.5). This will now create locality issues, given the Phase Impenetrability Condition (Chomsky 2001: 13):

According to (30), the complement of the v phase head—VP in (31)—is inaccessible for operations from the higher phase head C.

The BCMS NIs from (5b) are repeated in (32), and the derivation of the grammatical imperfective NI is given in (33).

In (33), u-mood[imp] is in Asp_imfv, and u-addr[__] in Asp_imfv probes the addressee subject first and gets valued. i-mood[__] in C then probes u-mood[imp] in Asp_imfv and gets valued. u-addr[__] in C is also valued, due to feature sharing, and u-mood[imp] in Asp_imfv is spelled out.

In (34), however, u-mood[imp] is in Asp_pfv, which is spelled out when C is merged. Thus, the ban on perfective NIs in BCMS is essentially a violation of syntactic locality: neither u-addr[__] in Asp_pfv nor i-mood[__] in C can be valued, since they belong to different spellout domains. The complex probe in C first probes the subject, and u-addr[__] gets valued. The problem is, however, that i-mood[__] cannot be valued by u-mood[imp], due to the Phase Impenetrability Condition. In addition, u-addr[__] in Asp_pfv cannot be valued.

Thus, on the present account the ban on perfective NIs in BCMS falls out straightforwardly.

3.4 Exception I: analytic NIs

In this section I show how analytic NIs, repeated here from (6), support the present proposal.

I assume that the da clause in (35b) is a MoodP (Veselinović 2019) or a vP (e.g., Todorović & Wurmbrand 2020). Diachronically, ne+moj goes back to the negated imperative of the modal verb moći ‘can, be able’ (e.g., Szucsich 2010), but synchronically the moj part is “bleached” of any semantic content. Moj seems to be functioning as a “dummy” verb that supplies the uninterpretable but valued feature u-mood[imp], which i-mood[__] on C can successfully probe in a syntactically licit configuration. The BCMS innovation, relative to the rest of Slavic, is thus that u-mood[imp] can be located in T, in addition to perfective or imperfective Asp.

Notice that a very similar construction exists in the northern dialects of Italy, as discussed in detail by Zanuttini 1991 and Kayne 1992. First, Italian in general can license a matrix infinitive in the function of an NI; this is surprising given that infinitives can otherwise only be licensed in embedded contexts. Kayne observes that such infinitives require the presence of negation, specifically the negative marker non, not some general type of negation:

Kayne proposes that the negative marker is not licensed by the infinitive but rather by an empty modal that in turn licenses the infinitive. According to this view, in (36) we have an instance of clitic climbing, where the clitic is not adjoined to the infinitive but to the phonetically unrealized modal. Very nice support for this proposal comes from the northern Italian dialects that have an overtly realized verbal form specific to the negative infinitival in the imperative construction, namely sta, which corresponds to the BCMS moj. Like in BCMS, the infinitive is preceded by sta in the NI, something that is not possible in positive contexts:

Zanuttini concludes from this that “sta in these languages is the lexical realization of tense, and this is why an occurrence of the preverbal negative marker (no) is possible in this type of imperative” (Zanuttini 1991: 81). Thus, the core assumptions behind my analysis of BCMS analytic NIs align well with Kayne and Zanuttini’s proposals regarding (36) and (37). That is, on my analysis, since the locality problem that the use of moj is meant to resolve is created by the presence of NegP, it also follows that moj cannot be used to license positive imperatives, in which the locality problem in question never arises. This is indeed correct:

My derivation for analytic NIs, then, is as in (39), with u-mood[imp] in T. u-addr[__] in T probes the addressee subject and gets valued. i-mood[__] in C probes u-mood[imp] in T and gets valued. u-addr[__] in C is also valued, due to feature sharing. u-addr[val] in T is spelled out as moj. Finally, the main verb takes either the infinitive (uninflected) form or the default present tense form (da clause). Crucially, the main verb does not participate in Agree with i-mood[__].

3.5 Exception II: lack of voluntary agency

Another fact that any account of the ban on perfective NIs should be able to explain is why perfective NIs become possible with verbs that lack voluntary agency:

Other verbs that allow perfective NIs in BCMS include iznenaditi se ‘get surprised,’ razočarati se ‘get disappointed,’ obrukati se ‘get embarrassed,’ and začuditi se ‘get surprised.’

In cases like (40), where the perfective form is used, the presupposition is that the subject is not in control of the action and that they might inadvertently perform an undesirable action. A particularly nice contrast is provided by the verb propustiti ‘let through,’ which is a perfective verb, ambiguous between agentive and non-agentive readings. The agentive use in (41a) is not compatible with the NI form (context: addressing a bouncer protecting a door), while the non-agentive use in (41b) is.

I take this obligatory lack of agency of the verbs in (40) and (41b) to be related to the “deficiency” of their corresponding vPs. On phase-based approaches, v introduces the external argument/agent (e.g., Hale & Keyser 1993, Chomsky 1995, and Kratzer 1996) and defines a phase domain (Chomsky 2000 and Chomsky 2001). These constructions either do not have a vP or their vP is deficient in the sense that it does not define a phase domain or trigger cyclic transfer (i.e., it is not a “strong phase”; Chomsky 2001). The point of variation among different Slavic languages then comes down to the question of which verbs under which circumstances may be associated with this type of “non-agentive,” “deficient” v head.⁵

The derivation of a structure like (40) is given in (42), in which the uninterpretable version of the imperative feature is in Asp_pfv. First u-addr[__] in C probes the addressee subject and gets valued. i-mood[__] then probes u-mood[imp] in Asp_pfv and gets valued. u-addr[__] in Asp_pfv is also valued, due to feature sharing. Finally u-addr[val] in Asp_pfv is spelled out.

An issue that could be raised at this point about my overall proposal is how it is possible for the complex imperative probe to be located in two different functional heads, namely C and T. If imperative is a clause type, then its features belong more naturally in the C domain; but if it has a probe that agrees with the subject in ϕ-features, it more naturally belongs to the T domain. One possibility is to assume that the BCMS lexicon contains both T and C specified for the same features and that some formal bias favors the use of T over C whenever possible.

But I think my proposal is most directly compatible with the feature inheritance system of Chomsky 2008, where it is proposed that ϕ-features originate in phase heads and are inherited by their complements (for discussion and different implementations of this idea see Richards 2007, Saito 2011, Richards 2012, Aldridge 2018, and Branigan 2020, among others). In other words, ϕ-features and tense features associated with the inflectional system are not inherent properties of T; instead, they belong to the phase head C. This analysis is in part motivated by the fact that T, for example, carries ϕ-features only in the presence of C. The proposal directly accounts for an old observation that a raising/ECM infinitival T, which is not selected by C, also lacks tense and ϕ-features. At the same time, it allows a uniform characterization of phase heads (C, v*) as the locus of uninterpretable features.

The picture that we arrive at regarding BCMS is that ϕ-probes originate in C in both declaratives and imperatives. The ϕ-probe is always inherited by T in declaratives, even when they are negated. Since the declarative ϕ-probe does not encode any force, it can remain in the scope of negation (located between C and T) without creating problems at LF. Or in other words, force in declaratives (and interrogatives) is always encoded in C, as it is not part of their agreement ϕ-probes. Consequently, force always scopes over the propositional material and negation in declaratives and questions.

The imperative ϕ-probe is also inherited by T in positive contexts, but since this probe encodes force, it cannot stay in the scope of negation.⁶ So, as a last resort, the imperative ϕ-probe can probe directly from C, which creates locality issues, as already discussed, since the complement of v is not accessible for operations from C. Thus, the ban on perfective NIs in BCMS can be explained.

One might wonder at this point how this analysis of BCMS NIs fits with other languages with some type of ban on NIs. For example, Rivero 1994 and Rivero & Terzi 1995 argue that the verb moves all the way to C in Romance, attracted by a strong imperative feature, whereas the verb in Slavic stays lower. We could think of this as a type of parameter: (a) in languages like Romance, the imperative probe can probe from C, triggering verb movement to C, and (b) in languages like BCMS, the imperative probe is inherited by T (when possible), and there is no verb movement, only agreement.

This would be similar to Carstens 2005’s agreement parameter whereby there are languages that have just simple downward agreement with a nominal without subsequent movement of the nominal and languages like Bantu in which the nominal moves after entering Agree with the probe (i.e., if F agrees with XP, then XP must check the EPP feature on F). Perhaps the same kind of parameter exists for verb movement, which is behind the contrast between Romance and BCMS.

As already mentioned (see footnote 3), my proposal about a special imperative probe is in many ways similar to Zanuttini 2008’s Jussive Phrase, which is argued to exist in all clauses conventionally associated with directive force and which abstracts over the subject and prevents it from entering a predicate relation. In my analysis, T specified for u-addr[__] can be present only in imperatives, which is similar to the idea of Jussive Phrase.

Bennis 2006 proposes that Dutch imperatives have a special feature called [2] in C, which can be assigned to the subject through agreement. Jensen 2003 also argues that imperatives have a special TP that includes a tense feature (anchored to speech time) and a person feature. All these analyses have in common the idea that second person features are uniquely encoded in the syntactic representation of imperatives, which I agree with. Thus, my complex imperative probe includes an uninterpretable feature in search of an addressee argument, which is not present in declaratives or questions—this is unique to imperatives. But I assume that this imperative probe is not an inherent property of T or some other designated functional head. Both declarative and imperative ϕ-features originate in C and are inherited by T. When negation is projected between C and T, feature inheritance by T fails only with the imperative probe since, unlike the declarative probe, it also encodes force and consequently cannot take scope under negation. One could say that the interaction between negation and the imperative probe in BCMS in fact reveals that feature inheritance takes place (which is otherwise not very easy to detect empirically).

Finally, the C–T feature inheritance system could be perhaps extended to AspPs, on the assumption that the higher, imperfective Asp is also a phase head. Then the uninterpretable but valued imperative features might originate in the higher AspP and be inherited by the lower AspP (assuming that vP would not block it). For proposals that AspP is a phase see Aelbrecht 2010, Harwood 2013, Aelbrecht & Harwood 2015, Harwood 2015, Fenger 2020, Kim 2024, and Kwong 2024, among others.⁷

3.6 On the structural position of perfective and imperfective verbs

One of the central assumptions of this article is that imperfective verbs in BCMS are higher than perfectives and that the v phase head is located between them (e.g., Travis 2010). In this section I will provide empirical support for this claim, which is based on reflexive passivizations, nominalizations, and idioms.

3.7 Interim summary

I have proposed that the ban on perfective NIs in BCMS is essentially due to a violation of syntactic locality. The imperative probe needs to enter Agree with the main verb, and it can probe from T or C, depending on whether NegP is projected between CP and TP. If NegP is projected, the imperative probe is located in C, since it cannot be in the scope of negation. But given the Phase Impenetrability Condition, nothing in the complement of v will be accessible for probing from C; by assumption, the complement of v includes perfective but not imperfective verbs. Consequently, only imperfective NIs are possible. In section 3.6 I presented empirical arguments (based on reflexive passivization, nominalizations, and idioms) for the claim that AspP_imfv is indeed located above vP, in contrast to AspP_pfv. I also, in section 3.4, discussed analytic NIs and showed that they are not constrained by aspect because the imperative inflection that the imperative probe has to value via Agree is not on the main verb but on a dummy verb close to the imperative probe. Analytic NIs are, however, constructions with many peculiar properties, which deserve special attention. I discuss analytic NIs in more detail in the next section.

4.1 Negative concord item licensing

A very useful test in this context is licensing of negative concord items, since it is sensitive to structure: it can occur within a single clause (or across a restructuring verb) but never across a finite CP (e.g., Progovac 1993, Progovac 1994, and Stjepanović 2004). In (58) negation cannot license the negative concord item in the subject position of ‘eat’; instead, the “long distance” negative polarity item iko can be used.

In the analytic NIs discussed so far, however—see (6)/(35), repeated in (59)—the negative concord item is licensed in the object position.

This contrast indicates that ne-moj in (58) combines with a CP, while in (59a) and (59b) it does not.

Consider also the examples in (60). First, in (60a), the da clause can have its own subject, different from the addressee, which can be overt and which triggers non-second person agreement on pojesti ‘eat.’ In this case, however, negation in ne-moj cannot license the negative concord item in the object position, unless there is another negation right before the main verb, as in (60b). But this then affects the interpretation, since we have a case of double negation now.

However, negative concord item licensing can (unexpectedly) become unavailable even in sentences like (59b), where moj and the main verb both have second person agreement. First, note that (61) can be felicitously used as a warning not to eat anything over there (e.g., since the food will be poisoned).

However, in a minimally different structure in (62a), which only has an extra adjunct nekim slučajem ‘by some accident’ attached between ne-moj and da, the negative concord item cannot be licensed. The negative concord item is possible only if another negation is added right before the main verb, as in (62b). However, (62b) is now a warning not to be/stay hungry, since it has double negation.

4.2 Analysis

The key observation regarding (62) is that ‘by some accident’ is compatible with the meaning in which the addressee is not aware that a certain state of affairs can occur and the speaker’s warning to the addressee is to avoid that state of affairs. In (61), on the other hand, the addressee is in control of their actions, and the command is to not perform a certain action. I propose that (58) underlyingly looks, in fact, like (63) and has two hidden verbs dozvoliti ‘to allow’ and desiti se ‘to happen.’

Note that it is always possible to pronounce these verbs. Also, neither dozvoliti ‘allow’ nor desiti se ‘to happen’ are restructuring verbs, like želeti ‘want’ in (64). Although desiti se combines with da clauses, it is not a restructuring verb even when it combines with a very simple da clause, as in (65); this indicates that even here the da clause is a CP.

The presence of hidden dozvoliti and desiti se is not only directly compatible with the subtle meaning of (62, 63), but it also explains the lack of negative concord item licensing, since desiti se always introduces a CP. This also resolves the problems of (57), which is really like (63): the addressee gets its thematic role from dozvoliti, which in turn take [desiti se ‘to happen’ + CP] as its complement.¹²

That this process of omitting verbs is general in BCMS is independently supported by Veselinović 2019, which shows convincingly that this process underlies the root–epistemic ambiguity in this language. The epistemic reading is biclausal, as in (66a), where mora ‘must’ combines with the matrix verb biti ‘to be,’ which takes a full CP complement. But biti ‘be’ is often dropped, which makes these structures superficially similar to true, monoclausal root modals, like the one in (66b). The structures that Veselinović proposes for these constructions are in (67).

BCMS da clauses can have different sizes (Browne 1987, Progovac 1993, Progovac 1994, Stjepanović 2004, Tomić 2004, Veselinović 2019, and Todorović & Wurmbrand 2020, among others). For instance, Todorović & Wurmbrand 2020 proposes that they come in three sizes: (i) CPs—propositional complements of ‘claim,’ ‘believe,’ or ‘think,’ which have their own tense properties independent of the matrix tense, (ii) TPs—future irrealis complements of ‘decide’ or ‘plan,’ whose event must occur in the future with respect to the matrix event, and (iii) vPs—tenseless complements of ‘try,’ ‘begin,’ or ‘manage.’ The sentence in (66a) superficially looks monoclausal, with mora in some higher modal projection combining with a TP or vP size da clause. But as Veselinović 2019 shows, all tests (e.g., negative concord item licensing) indicate that there is a CP boundary between mora and the da clause. Once we add to that the fact that biti can always be pronounced in (66a) and that biti selects for the CP size da clause, it becomes clear that (66a) must have a structure like (67a).

This type of gestalt effect, I argue, also happens in structures like (62a). The da clause after ne-moj here is really a CP, which is selected by the underlyingly present, covert verb desiti se ‘to happen.’