2.1 The OC Signature

Before any meaningful investigation of a grammatical construction can be carried out, one must delineate the class of phenomena that fall under this construction. Such “working definitions” are heuristic in nature. As knowledge accumulates, and as more languages enter the data pool, the working definitions change to reflect our better understanding of what is essential to the grammatical construction of interest and what is merely coincidental to it. In linguistic research, in fact, it is expected that later incarnations of the definitions will be sparser and more abstract than earlier ones. Typically, notions like linear order, overtness, and so on, initially taken to be essential, are subsequently discarded in view of cross-linguistic variability, and commonalities are gradually discerned at deeper levels.Footnote ²

This has also been the fate of the notion of OC, first introduced in Williams (Reference Williams1980), which has undergone considerable revisions ever since.Footnote ³ As a starting point for current research, I will adopt the OC signature of Landau (Reference Landau2013), slightly adjusted to reflect some recent findings in the field.

1. (4)

   The OC signature
   In a construction [… Xi … [S Yi …] …], where Y is the subject of clause S, if:
   a.	X must be (a) codependent(s) of S, AND
   b.	Y (or part of it) must be interpreted as a bound variable
   then this is an OC construction (X = controller, Y = controllee).

In (4), the controller and controllee are neutrally labeled X and Y, rather than DP and PRO, respectively. This is meant to allow implicit controllers, which do not surface as overt DPs, and also to allow controllees that surface as overt pronouns rather than as a null category (see Section 3.1). A “dependent” of S is either an argument or an adjunct of S; thus, (4a) subsumes both complement OC, where the controller and S are co-arguments, and adjunct OC (on the latter, see Section 5.2). Condition (4a) does not require the controller to be unique, thus allowing both control shift and split control, as long as the two potential controllers are codependents (in effect, co-arguments) of S. Finally, the parentheses in (4b) are meant to allow partial control (PC; see Section 3.1).

To illustrate with familiar examples, consider the following contrast.

1. 1. a.

      Venessai was impressed that Mikej had tried [PROj/*i to regain	OC
      his/*her composure].

   2. b.

      Venessai was impressed that Mikej had understood that	NOC
      [PROi/j regaining his/her composure] was essential to the discussion.

Suppose that the examples in (5) are not drawn from English but from some understudied language, and your goal as a linguist is to identify their nature with respect to control. Based on informant judgments, you observe that in (5a), Venessa and the bracketed clause are not codependents, and control fails; whereas in (5b), both Venessa and Mike are not codependents of the bracketed clause, and control goes through with either of them. This indicates that (5a) abides by criterion (4a) and hence stands a good chance of instantiating OC, while (5b) violates this criterion and so must be NOC. To strengthen this conclusion, you can apply criterion (4b).

1. 1. a. Only Mike tried to regain his composure.

   2. b. Only Mike understood that regaining his composure was essentialto the discussion.

Imagine that Venessa and Bob also tried to bring it about that Mike regain his composure. In this situation, (6a) is still true, because PRO must be understood as a bound variable, and Mike is the only individual of whom the property λx.try(x,regain-composure(x)) holds. Hence, the complement of try passes the second criterion of OC (4b) too. Imagine next that Venessa and Bob also understood that Mike’s regaining his composure was essential to the discussion. On this scenario, (6b) has a false reading; namely, the reading that ascribes to Mike the property λx. understood(x,essential-to-discussion(regaining-composure(Mike)). The sentence is false because the PRO subject of the gerund regaining his composure is understood referentially (as Mike) and not as a bound variable; not only did Mike reach the relevant understanding concerning Mike, but so did Venessa and Bob. This indicates that the subject gerund in (6b) falls under NOC also with respect to criterion (4b).

The criteria in (4) are particularly useful in distinguishing OC from pronominal anaphora. Pronouns, including null pronouns, often take intrasentential antecedents, but the relation between the antecedent and the pronoun is looser than the one required by the OC signature. Nevertheless, there is one construction that is famously similar to OC, and which the criteria in (4) are too coarse to distinguish from OC: That is Raising. We now turn to the OC-Raising distinction.

2.2 OC versus Raising

In Raising constructions, a nominal that is thematically an argument of the embedded predicate appears in the matrix clause, due to A-movement. The Raising predicate itself assigns no θ-role to the raised nominal; that is why nonthematic arguments (expletives and idiom chunks) can appear in Raising but not in control constructions. An expletive cannot get any θ-role, being nonreferential. An idiom chunk must be interpreted within the idiomatic phrase. If separated from the idiomatic phrase by movement (= Raising), the idiom chunk is still interpretable in its base copy. In control, however, there is no copy of the controller inside the complement clause; hence, the idiom chunk is forced to be interpreted as an argument of the control predicate, leading to the loss of the idiomatic reading and often to anomalous interpretations.

Examples (7a) and (7b) show the compatibility of Raising-to-Subject predicates with expletives and idiom chunks, while (7c) and (7d) show the incompatibility of subject control predicates with these elements. Similarly (8a,b) versus (8c,d) show the parallel contrast between Raising-to-Object and object control predicates (see (2a) and (2b) for the distinction between subject and object control).Footnote ⁴

1. 1. a. It seems that John is happy.

   2. b. My leg appeared to have been pulled.

   3. c. * It hoped that John is happy.

   4. d. * My leg attempted to be pulled.

1. 1. a. We expect it to be snowing all weekend.

   2. b. The police prevented tabs from being kept on their informer.

   3. c. * We convinced it to be snowing all weekend.

   4. d. * The police dissuaded tabs from being kept on their informer.

As noted, these differences reflect different syntactic derivations, corresponding to movement in the case of Raising as opposed to referential antecedence in the case of control. The raised nominal starts out in the embedded Spec,TP and raises to the matrix Spec,TP (in Raising-to-Subject) or the matrix Spec,VP (in Raising-to-Object); see (9a) and (9b). The controlled nominal, namely PRO, occupies the embedded Spec,TP position; see (9c) and (9d).

1. 1. a.

      Billi appeared [TP Billi to feel better].

      Raising

   2. b.

      They expectedk [VP Billi tk [TP Billi to remain silent]].

      Raising

   3. c.

      Billi hoped [CP [TP PROi to feel better]].

      Control

   4. d.

      They convinced Bill [CP [TP PROi to remain silent]].

      Control

Accordingly, Raising and OC display a wide array of empirical contrasts beyond the familiar ones in (7) and (8).Footnote ⁵ These contrasts can be grouped into three categories. (i) Low thematic source: The raised nominal receives its θ-role in the lower clause, whereas the controller nominal receives its θ-role in the matrix clause. (ii) “Reconstruction”: A raised nominal can take scope, or be bound, in its embedded copy position, whereas a controller nominal is necessarily interpreted in the matrix clause. (iii) Derivational history: The raised nominal displays properties of a copy in a movement chain, whereas the controller nominal does not.

Starting with (i), this is the source of the contrast in the tolerance to nonthematic arguments in (7) and (8). It is also the reason why lexical (“quirky”) case assigned in the lower clause is preserved on a raised nominal but not on a controller. Quirky case can be seen as a morphological reflection of θ-assignment, hence its preservation in Raising indicates the low thematic source of the raised nominal. A minimal pair in Icelandic, taken from Sigurðsson (Reference Sigurðsson2008), is shown in (10) (see also Thráinsson Reference Thráinsson1979, Bobaljik and Landau Reference Bobaljik and Landau2009). The embedded passive verb “helped” assigns quirky DAT to its subject. This case lodges on PRO in the control complement (10a) and is not passed on to the controller DP (which is NOM). In contrast, the quirky DAT case moves along with the raised nominal in Raising (10b).

1. 1. a.

      Mennirnir/*Mönnunum	vonast	til	[að PRO	verða	hjálpað].	Control
      men.the.NOM/*DAT	hope	for	to	be	helped
      “The men hope to be both helped.”

   2. b.

      Mönnunum/*Mennirnir	virðist [mönnunum hafa verið hjálpað].		Raising
      men.the.DAT/*NOM	seem	have been helped
      “The men seem to have both been helped.”

Consider (ii), “reconstruction” contrasts.Footnote ⁶ The raised nominal can be interpreted in its original merge position, but a controller must be interpreted in the matrix clause because it has no embedded copy (PRO is not a copy). This contrast shows up in many ways; we illustrate it here with scope reconstruction (11) and with Weak Crossover effects (12).Footnote ⁷

1. 1. a.

      Seven contestants are likely to lose on the next round.	Raising
      7con.≫ likely, likely≫7con.

   2. b.

      Seven contestants are afraid to lose on the next round.	Control
      7con.≫ afraid, *afraid>7con.

1. 1. a.

      Hisi medical condition worries every patienti.

   2. b.

      Hisi medical condition seems to worry every patienti.

      Raising

   3. c.

      * Hisi doctor tried to worry every patienti.

      Control

Such contrasts in reconstruction are consistently found in other languages as well (see Anagnostopoulou and Alexiadou Reference Anagnostopoulou, Alexiadou, Bird, Carnie, Haugen and Norquest1999 on Greek, Wurmbrand Reference Wurmbrand, Bird, Carnie, Haugen and Norquest1999 on German, and Baykov and Rudnev Reference Baykov and Rudnev2020 on Russian).

Scope reconstruction in Raising allows the raised nominal to be interpreted de dicto, but a controller must be read de re. Under the de dicto reading, an indefinite nominal introduces an entity in the context of thought, hence there is no commitment to its existence in the actual world. Under the de re reading, an indefinite nominal introduces an entity in the context of utterance, hence it is assumed to exist in the actual world. This is reflected in the following contrast.

1. 1. a.

      Smith expected a unicorn to drink the apple juice.

      Raising

   2. b.

      # Smith tempted a unicorn to drink the apple juice.

      Control

Parallel contrasts in reconstruction emerge when the matrix QP interacts with a negative complementizer (Baltin Reference Baltin, Irwin and Maldonado2009).

Finally, we turn to (iii), derivational history. Certain grammatical conditions are sensitive to whether a given syntactic position is derived (by movement) or not. For example, Rizzi’s (Reference Rizzi and Borer1986) Chain Condition blocks A-movement across a coindexed reflexive clitic (a corollary of the Minimal Link Condition; Chomsky Reference Chomsky1995), as in *[DP_{i …} si_{i …} t_i]. This correctly predicts that subject control across a (dative) reflexive clitic will be possible (14b) but Raising to subject will not (14a) (copies are marked as traces for expository purposes only). Note that (14b) is active and not passive, the “be” auxiliary occurring due to reflexivization.

1. 1. a.

      *I	due		candidatii	sii	risultavano	[ti poter	vincere].
      the	two		candidates	to.each-other	appeared	to.be.able	to.win
      (“The two candidates appeared to each other to be able to win.”)

   2. b.

      I	due	concorrentii	sii	sono	promessi	[di PROi	essere	leali].
      the	two	competitors	to.each-other	were	promised	DI	to.be	loyal
      “The two competitors promised to each other to be loyal.”

Another derivational constraint involves Freezing – the ban on extraction from derived positions (Wexler and Culicover Reference Wexler and Culicover1980). It is an old observation that the postverbal nominal in expect-type constructions blocks subextraction (15a,b) but that in persuade-type construction does not (15c,d) (Chomsky Reference Chomsky, Anderson and Kiparsky1973, Postal Reference Postal1974, Runner Reference Runner2006).

1. 1. a.

      * Whoi did you expectk [VP [stories about ti]j tk [tj to terrify John]]?

      (cf. Whoi did you hear [stories about ti])?

   2. b.

      * Whoi did you findk [VP [pictures of ti]j tk [tj to be offensive]]?

      (cf. Whoi did you find [pictures of ti]?)

   3. c.

      Whoi did you persuadek [VP [friends of ti]j tk [PROj to join us]]?

   4. d.

      Whoi did you pushk [VP [friends of ti]j tk [PROj to reconsider their position]]?

While Chomsky (Reference Chomsky, Anderson and Kiparsky1973), assuming no Raising to Object, traced these facts to Subjacency (more currently, the Subject Condition), on the Raising analysis they follow from the Freezing Principle.

It is worth mentioning some Raising-Control contrasts that were thought to be solid but research has proven to be unreliable. One property that was traditionally associated with Raising is clause-“defectiveness.” Within the GB theory of the 1980s, this idea was cashed out in terms of clause size: Raising clauses were analyzed as bare TPs, in contrast to control clauses, analyzed as ordinary CPs. However, this idea is dubious in both directions. First, restructuring gives rise to control complements smaller than CP in many languages (Wurmbrand Reference Wurmbrand, Bui and Özyıldız2015). Conversely, “hyper-raising” out of finite CPs is also attested in several languages (Wurmbrand Reference Wurmbrand, Ronai, Stigliano and Sun2019). While clause size often correlates with the Raising-Control contrast language-internally, it is not a criterion of universal validity.

A second property concerns potential overtness of the dependent position – the controllee or the Raising “trace.” It is already known that overt Raising is unnecessary in null subject languages; the matrix subject position may remain null (or contain a null expletive), and the nominal remains unraised in its embedded base position (see Halpert Reference Halpert2019 for some theoretical implications). We also know that controllees may surface as overt pronouns (see Section 3.1). Finally, the phenomenon of backward control suggests that they can also surface as full lexical DPs (Polinsky and Potsdam Reference Polinsky and Potsdam2002, Fukuda Reference Fukuda2008, Potsdam Reference Potsdam2009). Thus, how the control or Raising dependency is spelled out is not a reliable distinction between the two constructions. In contrast, the properties demonstrated in (8)–(15) provide a solid battery of contrasts that have been repeatedly confirmed in the literature throughout the years.

3.1 Syntactic Theories

The earliest account of OC in generative grammar was a syntactic one – The Equi-NP Deletion rule of Rosenbaum (Reference Rosenbaum1967, Reference Rosenbaum, Jacobs and Rosenbaum1970). Both the relata of the rule – two identical NPs – and the rule itself were strictly syntax-internal. Moreover, the rule was subject to a syntactic locality principle – the Minimal Distance Principle (MDP) – which required the controller to be the closest NP to the controlled (deleted) NP. The MDP was revived in a PRO-based analysis by Larson (Reference Larson1991) (see also Martin Reference Martin1996, Manzini and Roussou Reference Manzini and Roussou2000). Within the Movement Theory of Control, the MDP was adopted as descriptively correct, its effects reduced to the Minimal Link Condition, specifically, the ban on A-movement across an intervening A-position (Hornstein Reference Hornstein1999 and subsequent work).

The main ingredients of Rosenbaum’s proposal have been rejected by subsequent research. This early work in the 1970s revealed that the controlled position manifests interpretive properties typical of pronouns rather than of lexical NPs (e.g., bound variable readings, split control); more fundamentally, the deletion analysis did not explain the fundamental effect of obligatory control, for it said nothing about why deletion is sometimes impossible (John believed *(Mary) to be smart), sometimes possible (John wanted (Mary) to win), and sometimes mandatory (John managed (*Mary) to win). Likewise, the MDP faced serious counterexamples from the outset, much beyond the familiar promise-example, which furthermore bore the unmistakable signature of lexical sensitivity, much to the detriment of a purely syntactic constraint. These matters have been discussed extensively elsewhere, so I do not elaborate on them here.Footnote ⁸ Interestingly, the classical Equi-NP analysis is reincarnated in Chomsky’s recent “Form Copy” theory of control (Chomsky 2021, Chomsky et al. 2023, Chomsky 2024), which is predicated on the idea that lexical NPs can be rendered copies not just by movement, leading to deletion of the lower copy (see Landau to appear for a detailed critique of that proposal).

Nonetheless, the lingering heritage of the Equi-NP Deletion account was its most fundamental and yet nontrivial aspect: Superficially subjectless clauses do contain a subject at some abstract grammatical level. The nature of the abstractness was much debated in the following years, but current consensus retains the original idea that (barring reduced/restructuring complements) OC is indeed a dependence between two syntactic relata. The syntactic reality of PRO is revealed in a variety of interactions with syntactic processes, such as binding, agreement, and case concord (see Landau Reference Landau2013: chapter 3).

Perhaps the most straightforward evidence for the syntactic reality of the controlled position is the fact that in many situations, across many languages, it is realized by an explicit nominal. One class of cases involves controlled overt pronouns. In several Niger-Congo languages, irrealis complements of volitional and implicative verbs display OC with an obligatory pronominal subject, as in the Gã example (16a) (Allotey Reference Allotey2021). In other languages, typically (but not only) in Romance, PRO alternates with an overt pronoun, but the alternation is semantically significant: The controlled overt pronoun is associated with exhaustive or contrastive focus, often accompanied by some focus-sensitive particle. Example (16b) is from Hungarian (Szabolcsi Reference Szabolcsi, Irwin and Maldonado2009).

1. 1. a.

      Gbekebiii	lɛ	nye	[(ni)	*(ameii/*j)	he	shia].
      Children	DET	manged	COMP	3.PL	buy.INF	home
      “The children managed to buy a home.”

   2. b.

      Nem	felejtettem	el	[én	is	aláírni	a	levelet].
      Not	forgot.1SG	PFX	I	too	to.sign	the	letter.ACC
      “I didn’t forget to bring it about that I too sign the letter.”

Controlled pronouns in Chirag Dargwa also require a focus particle, but in this language, the embedded subject may even surface as a conjunction, with the controlled pronoun (or long-distance reflexive) occurring as one conjunct, a rare case of overt PC (Ganenkov Reference Ganenkov2023); on PC, see Section 3.3.

Furthermore, in some languages, the controlled position can surface as a full DP, a phenomenon labeled backward control (Polinsky and Potsdam Reference Polinsky and Potsdam2002, Fukuda Reference Fukuda2008, Potsdam Reference Potsdam2009): The embedded subject position hosts a lexical DP and the matrix controller position is null, yet reveals its syntactic reality by a number of tests. The pair in (17) illustrates an alternation between standard (forward) and backward control in Malagasy (LNK = linker, CT = circumstantial topic [voice], TT = theme topic [voice]), taken from Potsdam Reference Potsdam2009: 765). Abstracting away from the extracted embedded object, in (17a) the controller occurs as the matrix object and surfaces with accusative case; in (17b), the controller occurs as the embedded subject and surfaces with nominative case.

1. 1. a.

      Forward control
      trano-n’	iza	no	naneren’	i Mery	ahy	[hofafana]?
      house-LNK	who	FOC	force.CT	Mary	1SG. ACC	sweep.TT

   2. b.

      Backward control
      trano-n’	iza	no	naneren’ i Mery	[hofafa-	ko]?
      house-LNK	who	FOC	force.CT Mary	sweep.TT	1SG.NOM
      “Whose house did Mary force me to sweep?”

Pronominal control and backward control make two points eminently clear. First, OC cannot be fully reduced to a relation between a matrix predicate/argument and a subjectless clause; there must be a way to establish OC with a syntactic controllee. Second, PRO and pronouns do not exhaust the spellout possibilities for controlled DPs. Backward control is most readily explained by the Movement Theory of Control (MTC) as an instance of “low spellout” in an A-chain, and it is still the strongest argument in its favor.Footnote ⁹ It should be noted, though, that backward control is quite tricky to demonstrate; several early proposals invoking it have been retracted or reanalyzed without it (Kwon et al. Reference Kwon, Monahan, Polinsky, Hornstein and Polinsky2010, Yoshimoto Reference Yoshimoto2013, Alexiadou and Anagnostopoulou Reference Alexiadou, Anagnostopoulou, Mucha, Hartmann and Trawiński2021, Pietraszko 2021). The general validity of the MTC as a viable theory of control has been subject to much criticism from different angles. If backward control will ultimately be reanalyzed in one of the ways alluded to above or otherwise, the MTC will lose its main empirical motivation.Footnote ¹⁰

Within minimalism, the other main syntactic approach to OC is the Agree-based account.Footnote ¹¹ On this approach, PRO enters an Agree relation with a matrix element (directly or indirectly via the embedded Agr) and consequently values its ϕ-features; on some versions, it also values an indexical feature. This agreement chain is translated at LF as variable binding. Initially, this approach was designed to capture noncanonical phenomena like finite and partial control (on PC, see Section 3.1); however, subsequent work has shown that these phenomena can also be captured under alternative conceptions (e.g., Pearson Reference Pearson2016, Vinka Reference Vinka, Saami, Dalmi, Tsedryk and Cegłowski2022).

Finite control has been documented in a range of languages (all the Balkan languages, Persian, Kannada, Korean, Japanese, Arabic, Amharic, South Saami), challenging the classical view that linked control to nonfiniteness (e.g., via assumptions about the case-deficiency of nonfinite domains). Two examples are given in the following. Example (18a) shows finite control in Amharic, where the prospective aspectual marker lɨ- introduces an irrealis complement; the embedded verb is imperfective and fully inflected (Leung and Halefom Reference Leung and Halefom2017: 13). Example (18b) shows finite control in Aromanian, where the complementizer ta introduces a complement hosting the subjunctive particle s(i) and an inflected verb (Manzini and Savoia Reference Manzini and Savoia2018: 239).

1. 1. a.

      käbbädä l-i-bärr-Ø					märrät’ä-Ø.	(Amharic)
      Kebede CM-3S.MS-fly.IMP-3SG.M prefer.PERF−3SG.M
      “Kebede preferred to fly.”

   2. b.

      am	uʁitə	ta	s	u	vɛd.	(Aromanian)
      have.1SG	started	that	PRT	it	see.1SG
      “I began to see it.”

The most immediate outcome of the Agree-based theory is that PRO must formally agree with the controller, a fact that is, surprisingly, not easy to capture under purely semantic conceptions of OC; see (63) and extensive discussion in Landau (Reference Landau2016b, Reference Landau, Patel-Grosz, Grosz and Zobel2018).

There is, however, one empirical effect that falls out naturally from the Agree-based analysis of Landau (Reference Landau2000, Reference Landau2004) but not under any alternative; it involves the so-called “Visser’s generalization,” which restricts OC by implicit passive agents (Landau Reference Landau2000: 169–179). As van Urk (Reference van Urk2013) shows, the true generalization is about the interaction of implicit control and agreement.

1. (19) Revised Visser’s generalization (RVG)

   Implicit subjects cannot control if T agrees with a referential DP.

In other words, implicit subjects of impersonal passives are free of the RVG. The contrast is illustrated in (20) for Norwegian; van Urk shows that it also holds in German, Dutch, Swedish, Icelandic, and English (the latter allowing less opportunity to observe the RVG due to its unproductive usage of impersonal passives).

1. 1. a.

      *Jeg	ble	lovet	å	gi	meg	gaver.
      I	was	promised	COMP	give.INF	me.ACC	gifts
      (Lit.) “I was promised to give me gifts.”

   2. b.

      Det	ble	lovet	å	gi	meg	gaver.
      there	was	promised	COMP	give.INF	me.ACC	gifts
      (Lit.) “It was promised to give me gifts.”

Note that RVG is not about morphological agreement; the forms ble “was” in (20a) and (20b) are identical. Nor is it about A-movement; van Urk cites sentences parallel to (20a), except that the matrix object remains in situ (in German and Icelandic), and they are equally ungrammatical. Thus, it is the formation of a syntactic Agree relation between a referential DP and T that somehow obstructs implicit OC.

van Urk reasons that RVG reflects the key role of the operation Agree is mediating OC. Following Landau (Reference Landau2000, 2004) (and unlike later executions, such as Fischer Reference Fischer2018 and McFadden and Sundaresan Reference McFadden and Sundaresan2018), he assumes that a matrix functional head mediates OC. Thus, subject control is established between the matrix subject and PRO because both Agree with T, resulting in feature sharing; object control works the same, via light v. Thus, if T is unavailable to Agree with the controller, OC fails.

van Urk further assumes that the implicit subject of passive is represented in [Spec,vP] as a syntactic D head denoting an existential quantifier, call it D_∃. In impersonal passives (20b), T agrees both with this D_∃ and with PRO, successfully resulting in implicit subject control. In personal passives (20a), however, T agrees with the matrix goal argument, and thus control by the implicit subject is rendered impossible.Footnote ¹² The reason why T must Agree with the overt goal DP and not with D_∃ is that only the former requires case-licensing, and case-licensing is parasitic on ϕ-Agree.

Note that the RVG only specifies when implicit arguments cannot control; it does not specify when they can. It therefore makes no prediction that OC will succeed in all impersonal passives. Indeed, impersonal passives of implicative verbs resist OC in many languages, for reasons independent of agreement (see the discussion of (64b)).

Currently, the RVG is only explainable on the Agree-based theory of OC and not on alternative theories, specifically on the version of that theory that depends on the mediating role of T in subject control. Before closing, let me point out two empirical challenges to the RVG – one that turns out to be spurious, the other one real. Consider first implicit subject control into adjuncts, which is not blocked by agreement with a matrix argument.

1. 1. a. A shed was built to store the tools in.

   2. b. The game was played wearing no shoes.

   3. c. The president was elected without considering his competence.

These examples are problematic to the RVG only if they instantiate OC by Agree. However, implicit agent control into adjuncts is an instance of NOC, as shown in Landau (Reference Landau, Halpert, Kotek and van Urk2017, Reference Landau2021a). This is relatively easy to see in (21a), where the storer need not be the builder, but the point is more general and can be demonstrated for other types of adjuncts, using and manipulating the kind of contextual information to which NOC is sensitive, as shown in the works cited.

The other counterexample to RVG is real; it involves implicit subject control into interrogative complements as in (22a).

1. 1. a. Mary was asked where to throw the trash.

   2. b. Ben_i knew that Mary was asked by Sue_j where [PRO_*i/j to throw the trash].

   3. c. *Mary was asked by Ben_i when to call him_i.[cf. Mary was asked by Ben_i when people should call him_i]

   4. d. *Mary was asked how to improve myself.

Examples (22b)–(22d) serve to demonstrate that interrogative complements under passive verbs instantiate OC and not NOC, thus blocking long-distance, arbitrary, and deictic control, respectively.Footnote ¹³ Given this, (22a) is a genuine exception to the RVG and the contrast between it and (20a) remains a puzzle.

Like other syntactic approaches to OC, the Agree-based account is not free of problems. A fundamental issue, which receives no satisfactory answer, is the distribution of PRO. In the Agree model, PRO is specified [–R], which in turn can only be checked by defective T heads – lacking either semantic tense or morphological agreement. This captures a surface correlation (roughly, the aversion of indicative clauses to PRO) but is little more than a formal redescription of the problem. A second problem concerns the treatment of oblique controllers, which are quite common (e.g., It is incumbent [upon them] to cooperate with the police). Oblique arguments are almost universally inert for external agreement, yet they must be visible to the alleged OC-creating Agree operation – even when they cannot trigger agreement on any functional head.

In sum: OC cannot be fully explained by strictly syntactic theories, since they leave too many issues unanswered – from the lexical aspects of controller choice to atypical locality or agreement conditions (unmatched in standard instances of Move and Agree). Nonetheless, the thorough studies of OC within syntactic frameworks have uncovered a wealth of cross-linguistic data to be reckoned with in any comprehensive theory of OC. Importantly, these data establish beyond reasonable doubt that OC does implicate a syntactic dependency of some sort, no matter what other types of grammatical machinery it incorporates.

3.2 Lexical-Semantic Theories

Starting from the early 1970s, an alternative to the syntactic account of OC has developed, in which the crucial information used to establish the control relation is lexical and not syntactic. In this research tradition, control rules make reference to thematic roles, thematic hierarchies, semantic classes, and other semantic notions derivable from the lexical content of the control predicate (Jackendoff Reference Jackendoff1974, Chierchia Reference Chierchia1984, Nishigauchi Reference Nishigauchi1984, Xu Reference Xu1986, Sag and Pollard Reference Sag and Pollard1991, Růžička Reference Růžička1999, Rooryck Reference Rooryck2000, 2007, Jackendoff and Culicover Reference Jackendoff and Culicover2003).

Basic observations about the choice of controller motivate a lexical ingredient in any adequate account of OC.

1. 1. a. Diane_i promised/pledged to Mark_j [PRO_i/*j to give a hand].

   2. b. Diane_i persuaded/encouraged Mark_j [PRO_*i/j to give a hand].

1. 1. a. The promise that was given by Diane_i to Mark_j [PRO_i/*j to give a hand].

   2. b. Mark_j was encouraged [PRO_j to give a hand].

1. 1. a. Diane_i asked/begged/said to/shouted to Mark_j [PRO_*i/j to give a hand].

   2. b. Diane_i asked/begged [PRO_i/*j to give a hand].

   3. c. Diane_i said/shouted [PRO_*i/j to give a hand].

Verbs with similar meanings tend to select the same argument as a controller. Thus, verbs of commitment select the source of the commitment as a controller (23a), while verbs of influence select the target of influence as the controller (23b). These roles are identified thematically and not by any syntactic position, as seen in (24). That distance (or the MDP) cannot fully address these issues is further shown in (25): While some verbs switch from object control to subject control when their object is dropped (or implicit) (25b), others resist this shift and retain the interpretation of control by the (implicit) object (25c). Once again, the relevant verb classes are semantically defined.

Throughout the years, it has proven extremely difficult to pinpoint the semantic components that determine the choice of controller. The main problem with lexicalist accounts is that the analytic categories are either vague or just approximate, with too many cases listed as “exceptions.” For example, Chierchia’s (Reference Chierchia1984) thematic hierarchy Theme > Source > Goal > … correctly predicts subject (= source) control with promise, but wrongly predicts it with other communication verbs like tell, order, and require, which must be labeled as “marked.” Sag and Pollard (Reference Sag and Pollard1991) lump all object control verbs under the category influence, even though the object controller of communication verbs is not influenced (e.g., observe the test What Mary did to John was force/*tell him to clean up his room).

Jackendoff and Culicover (Reference Jackendoff and Culicover2003) classify control predicates by underlying “conceptual predicates” (e.g., INTEND, OBLIGATED, REQUEST, etc.), arguing that OC is established at Conceptual Structure (CS) between a designated argument of the conceptual predicate and the embedded subject. Appeal to hidden CS predicates is problematic insofar as speakers have no direct intuitions about the meanings of these predicates; speakers can evaluate what intend means but not what the hypothetical INTEND does. Such explanations thus risk circularity (see Boeckx et al. Reference Boeckx, Hornstein and Nunes2010b: 230–237 and Landau Reference Landau2013: 135–136 for detailed critiques).Footnote ¹⁴

A fundamental shortcoming of all purely lexical theories of OC is their inability to identify the controllee position. That position can only be defined syntactically, as the embedded subject, which may bear any thematic role whatsoever (see Landau Reference Landau2010: 363, fn. 4).

1. (26)

   a.	John planned [PRO to work harder].	PRO = agent
   b.	John planned [PRO to be a TV host].	PRO = stative bearer of property
   c.	John avoided [PRO receiving mail].	PRO = goal
   d.	John remembered [PRO fearing ghosts].	PRO = experiencer
   e.	John tried [PRO to be elected].	PRO = patient

Now, lexicalist theories might analyze some of these cases with hidden coercion, restoring an invariant, embedded thematic role as the target of control; but that would still beg the question of why that hypothetical role must be realized as the syntactic subject of the complement. The problem is compounded in those lexicalist theories that deny the existence of a syntactic null subject and must therefore reintroduce subjecthood into the lexicon.Footnote ¹⁵

Coercion is commonly invoked to explain control shift from subject to object or vice versa, as in be-allowed-to complements. Thus, a canonical subject control verb like promise shifts to object control and a canonical object control verb like ask shifts to subject control.

1. 1. a. She promised him_i [PRO_i to be allowed to take a picture of himself].

   2. b. He_i asked her [PRO_i to be allowed to take a picture of himself].

The phenomenon of control shift has been documented early on (Rosenbaum Reference Rosenbaum1967: 92, fn. 13) and received numerous treatments; see Landau (2013: 136–148) for discussion and appraisal of the different proposals. One evident implication of control shift is that the lexical semantics of the control predicate alone is not sufficient to determine the controller. Thus, control shift is facilitated by deagentivized and in particular by modalized complements, where PRO is construed as the recipient of permission. Unfortunately, thematic concepts like “causative” of “beneficiary,” which figure in some of the lexicalist accounts, are too coarse to pick out this specific flavor of control shift complements (see Uegaki Reference Uegaki2011 for an attempt to formally model control shift).

More fundamentally, because both the matrix and the embedded predicate contribute to the acceptability of control shift, lexicalist accounts are led to posit cross-clausal lexical dependencies. This runs counter to the basic idea that lexical relations are strictly local (i.e., spanning a single argument structure). Of course, one can abandon this idea, as is done within HPSG and the Parallel Architecture of Jackendoff, but that would rob the debate about whether OC is to be handled in the syntax or in the lexicon of much of its interest; a lexicon rich enough to express cross-clausal dependencies is for all practical purposes already “syntactic.”

A more pressing concern is that control shift appears to be sensitive not only to strictly lexical information but also to the pragmatics of the reported event. Thus, in many languages, an explicit modal is not necessary in the complement, as long as modality is somehow understood from the surrounding context. A helpful hint comes from authority relations between the two participants. Consider sentence (28a) in Hebrew. Out of the blue, it is ambiguous. Whether the object control interpretation (28b) or the subject control one is intended (28c) is entirely up to the context. If we know that the male has authority over the female, we select the former, and if we know that it is the female who has authority over the male, we select the latter interpretation.

1. 1. a.

      hui	bikeš	mimenaj	[PROi/j	le’hiša’er].
      he	asked	from.her		to.stay

   2. b.

      He requested from her that she would stay.

   3. c.

      He asked for her permission to stay.

The understanding that pragmatics has an important role in choosing the controller in OC – even if syntax has a decisive role in delimiting the domain in which it must occur, in accordance with the OC signature – has first made its entrance to control studies in Farkas (Reference Farkas1988). Farkas introduced the semantic notion Responsibility: RESP(i,s) holds between an initiator i and a situation s just in case i intentionally brings about s. In the unmarked case, like (23a) and (23b), the controller of an infinitive describing situation s is just the matrix argument that is the initiator of s. In control shift situations, such as (27), the controller is the individual whose actions are determined by the initiator.

While RESP in Farkas’ analysis, unfortunately, is an intuitive rather than a formalized concept, it transcends the limits of strictly lexicalist accounts by explicitly incorporating pragmatic considerations. The concept of “initiator” is broader than “agent,” and applies to secondary agents of passives (e.g., King was deliberately arrested) as well as to NOC by extra-sentential antecedents (e.g., The shop window has a big sale sign in it in order to attract customers). The RESP relation is not reducible to thematic roles, and has a “global” character in that it is sensitive both to the matrix and the embedded eventuality. In these respects, it can be thought of a precursor of the ESA theories to be discussed in the next subsection.

Last to be mentioned within the lexicalist camp are recent attempts to model the semantics of OC using the formal apparatus of attitude reports (Stephenson Reference Stephenson2010, Pearson Reference Pearson2013, Reference Pearson2016, Reference Pearson2018). These accounts are lexicalist insofar as they derive the OC dependency from the lexical denotation of the control predicate; however, that should not imply that they deny the existence of PRO (normally they acknowledge it) or the importance of syntax in feeding compositional semantics with the right structures for interpretation. Commonly in these accounts, a λ-abstractor is inserted at the edge of the complement, binding PRO and yielding a property denotation. This property serves as the first argument of the control verb, which, being an attitude predicate, quantifies over structured worlds/contexts, for example, over tuples of <individual,world,time>. It is the lexical meaning of this predicate that asserts the control implication; namely, that in each of these contexts, the complement property holds of the doxastic counterpart of the attitude holder (e.g., subject control with want) or of its addressee (e.g., object control with tell). The notion of “doxastic counterpart” encodes self-identification, accounting for the obligatory de se reading of PRO in attitude complements.

The biggest advantage of formal semantic analyses of OC is their explicitness; one can track exactly how the meaning of an OC construction is composed from the meaning of its parts. However, these analyses leave a few major questions unanswered.

First, the syntactic relation between the controller and PRO is lost, since PRO is locally bound by an operator, which bears no syntactic relation to the controller. This leaves unexplained the basic fact of agreement – PRO must agree with the controller, even in cases where the agreement is merely formal; see (63), Landau (Reference Landau2016b), and especially Landau (Reference Landau, Patel-Grosz, Grosz and Zobel2018). Second, the formal semantic accounts do not really address the hard question of controller choice: What makes certain verbs cluster together and differently from other verbs with respect to subject or object control? Specifying in the lexical entry of each control verb the “designated argument,” whose doxastic counterpart is identified with the individual coordinate of the world-time-individual triplet quantified over by the predicate, misses obvious generalizations that cut across OC verbs.Footnote ¹⁶ Furthermore, we have seen that the relevant generalizations are sensitive to variable pragmatic information, which cannot be specified in lexical entries. Finally, the basic assumption that OC complements uniformly denote properties is at odds with a number of distinctions that naturally fall out of a systematic cut between property complements and propositional complements (see Section 5 on the dual analysis of control).

In sum, lexicalist accounts of OC have made important contributions to our understanding that OC is not a “blind” syntactic dependency between two positions like, for example, A-movement, which is oblivious to the semantic roles of the related positions. The semantic roles, as well as the overall construal of causal and deontic relations that tie together the matrix and the embedded eventualities, all conspire to determine which matrix argument is selected as the controller. Lexical entries cannot encode every aspect of these construals, but they certainly restrict the range of construals available in control constructions and, correspondingly, the choice of controller.

3.3 Embedded Speech Act Theories

While lexicalist theories focus their attention on the argument structure of the control verb and try to formulate systematic correlations between thematic roles and controller choice, ESA theories highlight the type of speech act – by which we include both actual speech acts and mental “acts” such as decisions, plans, and so on – expressed by the OC complement. The leading idea behind this approach is that the choice of controller naturally falls out from the proper identification of this speech act, using a couple of systematic “bridging” principles. These bridging principles make crucial reference to indexical information tagged on the participants in the speech event.

One can discern two strands of research converging on ESA theories. The first strand was launched in Postal (Reference Postal1970), specifically in the appendix to that study, and was gradually developed, syntactically and semantically, in Kuno (Reference Kuno1972), Bianchi (Reference Bianchi, Guéron and Tasmowski2003), Schlenker (Reference Schlenker2003, Reference Schlenker, von Heusinger, Maienborn and Portner2011), Anand and Nevins (Reference Anand, Nevins and Young2004), Anand (Reference Anand2006), Baker (Reference Baker2008), Stephenson (Reference Stephenson2010), Landau (Reference Landau2015, Reference Landau, Patel-Grosz, Grosz and Zobel2018), and Stegovec (Reference Stegovec2019). The second strand has been developed within studies of OC in East Asian languages, mostly Japanese and Korean. Curiously, these two strands have evolved quite independently of each other, although they reach very similar conclusions. I will start by describing Postal’s original idea and then proceed directly to the literature on East Asian languages, which provides striking support for it.

Postal (Reference Postal1970) made three key observations regarding the relation between OC and ESAs. First, infinitival OC complements often have a finite counterpart with a characteristic modal. Second, the understood controller of the silent subject of the infinitive (later called PRO) corresponds to the understood antecedent of the subject pronoun in the modal finite counterpart; see the pairs (a)–(b) in (29)–(32). Third, there is a systematic correlation between the choice of matrix subject or matrix object antecedent of the pronominal subject in the Indirect Discourse (ID) complements in the (b) examples and a first or second pronoun in their Direct Discourse (DD) counterparts in the (c) examples.

1. 1. a. Harry told Betty_i PRO_i to marry him.

   2. b. Harry told Betty_i that she_i should marry him.

   3. c. (You) marry me, Harry told Betty.

1. 1. a. Harry asked Betty_i PRO_i to marry him.

   2. b. Harry asked Betty_i if she_i would marry him.

   3. c. Will you (please) marry me, Harry asked Betty.

1. 1. a. Harry_i promised Betty PRO_i to leave.

   2. b. Harry_i promised Betty that he_i would leave.

   3. c. I will leave, Harry promised Mary.

1. 1. a. Harry_i asked Betty when PRO_i to leave.

   2. b. Harry_i asked Betty when he_i should leave.

   3. c. When should I leave, Harry asked Betty.

The emerging generalization, which Postal stated, was this.

1. (33) Generalization: If the DD subject is second person, the ID subject (PRO) isobject-controlled; if the DD subject is first person, the ID subject is subject-controlled.

Note that Postal’s reference to modality would more accurately be described today as mood, that is, illocutionary force. It is the parallelism between the force of the complements in the (a)–(b) pairs that is tied to the choice of antecedent: For example, the subject of the content of a promise is identified with the promiser (Speaker), while the subject of the content of a directive is identified with the one being directed (Addressee). Deontic modality here is a concomitant feature of speech acts imposing commitments on either the speaker or the addressee.

As to the most interesting question of why controller choice would correlate with the person feature of a DD counterpart, Postal had the following to say (p. 496): “If Fact C [generalization (33), IL], which represents a correlation between properties of ‘corresponding’ Direct and Indirect Discourse sentences, is to provide an explanation for the operation of EQUI in Indirect Discourse sentences, it must be the case that underlying structures of such Indirect Discourse sentences manifest properties somehow linked to the surface properties of the relevant Direct Discourse sentences.”

In other words, underlying PRO there must be some “ancestor” with properties “somehow linked” to the indexical pronouns that figure in the DD paraphrases. While current theory no longer relates ID and DD (or finite and nonfinite clauses) transformationally, Postal’s insight is conceptually independent of such derivations. At its core, it amounts to the appealing idea that in those complements that paraphrase ESA, OC PRO is interpreted as if it were an indexical pronoun. This insight has proven very fruitful in subsequent work on OC. In particular, it finds interesting support in the grammatical expression of OC in East Asian languages.Footnote ¹⁷

The Japanese/Korean scholarship on OC as ESA capitalizes on two observations: (i) In matrix clauses, certain mood particles restrict their subject to first person, second person, or inclusive plural first person (“I+you”); (ii) in embedded clauses, the very same mood particles restrict the choice of controller to the matrix speaker/author, addressee, or their combination.

Starting with (i), a first person subject is selected by the promissive (PRM), intentive (INT), and optative (OPT) mood markers; second person is selected by the imperative (IMP) mood marker; and inclusive first plural person subject is selected by the exhortative (EXH) mood marker (Japanese data from Matsuda Reference Matsuda, Mucha, Hartmann and Trawiński2021: 147–148).Footnote ¹⁸

1. 1. a.

      Watasi/Watasitati/??Anata/??Anatatati/??Kare/??Karera}-wa hatiji-ni kaer-u.
      I/We/??You/??You.PL/??He/??They-TOP	eight-at	go.home-PRM
      “I’ll/We’ll go home at eight.”

   2. b.

      *Watasi/*Watasitati/Omae/Omaetati/*Kare/*Karera-wa hayaku ronbun kak-e!
      *I/*We/You/You.PL/*He/*They-TOP	soon paper write-IMP
      “(You) write that paper soon!”

   3. c.

      Watasitati/??Anatatati/*Karera-wa shukudai si-yoo.
      We/??You.PL/*They-TOP	homework do-EXH
      “Why don’t we do our homework.” / “Let’s do our homework.”

To my knowledge, it was Pak et al. (Reference Pak, Portner and Zanuttini2008) and Madigan (Reference Madigan2008a, Reference Madigan2008b) who first proposed to connect between observations (i) and (ii) and derive the choice of controller from the choice of the mood marker on the embedded clause, in other words, from the type of speech act embedded under the control verb, which is exactly Postal’s (Reference Postal1970) original insight. This idea has been put to use, in various forms, in Lee (Reference Lee2009), Seo and Hoe (Reference Seo and Hoe2015), Sisovics (Reference Sisovics2018), Matsuda (Reference Matsuda2019, Reference Matsuda, Mucha, Hartmann and Trawiński2021), and Liao and Wang (Reference Liao and Wang2022).Footnote ¹⁹

As Lee and Madigan observed, there must be some fit between the matrix verb and the embedded mood markers. A verb like order cannot embed a promise and a verb like promise cannot embed a directive. Nonetheless, many verbs are somewhat flexible and allow more than one choice. It is here that the effect of the mood marker is most striking. The examples in (35) show how the choice of the embedded mood marker determines the control options under the Korean verb mal “say” (Madigan Reference Madigan2008b: 167, 174).

1. 1. a.

      Inho1-ka	Hwun2-eykey pro1/2/3	swuyeng-ha-n-ta-ko	mal-ha-yess-ta.
      Inho-NOM	Hwun-DAT	swim-do-IND-DC-C	tell-do-PST-DC
      “Inho1 told Hwun2 that he1/she2/someone3 is swimming.”

   2. b.

      Inho1-ka	Hwun2-eykey PRO*1/2/*3	swuyeng-ha-keyss-ta-ko
      Inho-NOM	Hwun-DAT	swim-VOL-DC-C
      mal-ha-yess-ta.
      tell-do-PST-DC
      “Inho1 said to Hwun2 that he1/*she2/*someone would swim.”

   3. c.

      Inho1-ka	Hwun2-eykey PRO*1/ 2/*3	swuyeng-ha-la-ko	mal-ha-yess-ta.
      Inho-NOM	Hwun-DAT	swim-do-IMP-C	tell-do-PST-DC
      “Inho told Hwun to swim.”

   4. d.

      Inho1-ka	Hwun2-eykey PRO1 +2/*1/*2/*3	swuyeng-ha-ca-ko	mal-ha-yess-ta.
      Inho-NOM	Hwun-DAT	swim-do-EXH-C	tell-do-PST-DC
      “Inho1 said to Hwun2 that they1+2/*3 should go swimming.”

The indicative present tense marker (n)un-ta does not induce control (35a); the volitional marker keyss, which is speaker-oriented in matrix clauses, induces subject control in complements (35b); the imperative marker la, which is addressee-oriented in matrix clauses, induces object control in complements (35c); and the exhortative marker ca, which is oriented to the speaker + addressee in matrix clauses, induces split control in complements (35d).

Consider now how this causal relation between the mood marker and controller choice is cashed out in the relevant studies. In Pak et al. Reference Pak, Portner and Zanuttini2008, promises, directives, and so on are expressed by jussive clauses, which essentially contribute properties to the “To-Do List” of the speaker/addressee. Mood markers are functional heads projecting a JussP, establishing agreement with the local DP (= subject) and endowing it with interpretable person features, interpreted as referential presuppositions. They also introduce a λ-binder to bind the subject variable and produce the property added to the To-Do List (Zanuttini et al. Reference Zanuttini, Pak and Portner2012).

In embedded contexts, the speaker/addressee participant is shifted to the reported speech (or thought) event. If c* is the utterance context and c’ is the reported context, then Speaker(c’) and Addressee(c’) will pick out the subject and object of the matrix verb, respectively. This is achieved by context shift, in analogy to treatments of indexical shift, where the attitude verb is taken to quantify over contexts, for example, tuples of <world, time, speaker, addressee> (Schlenker Reference Schlenker2003). Pak et al. (Reference Pak, Portner and Zanuttini2008) and Madigan (Reference Madigan2008a) make direct reference to [person] as the key feature identifying the matrix controller, whereas Madigan (Reference Madigan2008b) uses thematic roles like Agent to do so. However, both assume that PRO is, in essence, a shifted indexical.

The ESA theory of OC is appealing in a number of ways and seems particularly suitable for East Asian languages. It is predicated on a very natural intuition – namely, that (attitude) OC verbs are just an instance of verbs introducing ID; that this discourse typically involves deontic commitments to bring about a certain state of affairs; and that the identity of the party entrusted with bringing it about is systematically correlated with, and predictable from, the type of thought or speech act involved (intention, promise, directive, exhortation, etc.). Finally, it offers a natural account of the de se reading characteristic of PRO in attitude complements; it is essentially reduced to the obligatory de se reading associated with first person indexicals.Footnote ²⁰

One general puzzle about OC that may receive a principled explanation within the ESA theory is the existence of PC (Wilkinson Reference Wilkinson1971, Landau Reference Landau2000, Reference Landau2004, Reference Landau2016b, Sheehan Reference Sheehan2012, Reference Sheehan, Lahousse and Marzo2014, Reference Sheehan, Gonçalves and Santos2018a,Reference Sheehan, Cognola and Casalicchiob, White and Grano Reference White, Grano, Urtzi Etxeberria, Irurtzun and Leferman2014, Pearson Reference Pearson2016, Grano Reference Grano2017a, Pitterofff et al. 2017, Authier and Reed Reference Authier and Reed2018, Reference Authier, Reed and Vogel2020, Pitteroff and Sheehan Reference Pitteroff, Sheehan, Hucklebridge and Nelson2018). In genuine PC, the referential relation between the controller and the controllee is a subset relation, rather than identity (the latter is the case only in “fake PC”; see Pitterofff et al. 2017, Pitteroff and Sheehan 2018). The “residue,” uncontrolled part is picked up from salient antecedents (in discourse or in the conversational context; a PC construal is notated as index “i+” on PRO).

1. (36) We thought that …

   1. a. The chair_i preferred [PRO_i+ to gather at 6].

   2. b. Bill_i regretted [PRO_i+ meeting without a concrete agenda].

   3. c. Mary_i wondered [whether PRO_i+ to apply together for the grant].

   4. d. It was humiliating to the chair_i [PRO_i+ to disperse so abruptly].

PC is attested in complements of attitude predicates; thus, it is not available, for example, in implicative complements (37a). Furthermore, the plurality of PRO under PC is of a rather abstract nature, resisting morphological plural marking as well as distributive readings (37b)–(37c) (Landau Reference Landau2016a,Reference Landaub, Authier and Reed Reference Authier and Reed2018, Reference Authier, Reed and Vogel2020), at least in most languages.

1. 1. a.

      * (We thought that) Johni managed [PROi+ to gather at 6].

   2. b.

      * (We thought that) the chairi preferred [PROi+ to work as partners].

   3. c.

      (We thought that) Helen wanted [PROi+ to (*each) answer a different question].

There is a great deal more to say about PC; recent research has documented restricted deviations from the generalizations exemplified in (37) in various languages, which space limitations prevent us from discussing (see Landau Reference Landau2024). Nevertheless, these core properties hold across many languages and constructions to warrant a principled explanation.

While PC can be modeled in many different ways – via syntactic agreement, lexical entailments, or pragmatic implicatures – many of these ways remain descriptive. A true insight appeared in Matsuda (Reference Matsuda2019, Reference Matsuda, Mucha, Hartmann and Trawiński2021), where a link between PC and the associative semantics of indexical pronouns was proposed. Matsuda’s account proceeds in two steps. First, it is noted that the standard semantics of [person] is associative; thus, we means “a group including the speaker” and you_PL means “a group including the addressee” (rather than “speakers” and “addressees”); See Noyer 1992, Cysouw 2003, Bobaljik 2008, and Wechsler 2019. Indeed, this is how indexical features are interpreted within the presuppositional approach to ϕ-features (Heim Reference Heim, Harbour, Adger and Béjar2008, Kratzer Reference Kratzer2009):

1. 1. a. $\text{[AUTHOR]}$ ^g,c = λx_e:x includes the speaker/thinker in c.x.

   2. b. $\text{[ADDRESSEE]}$ ^g,c = λx_e:x includes the addressee in c.x.

Second, following the ESA theory, OC PRO is essentially an indexical pronoun – first person in subject control and second person in object control. Thus, the associative semantics is available by default to OC PRO in attitude contexts (reported speech or thought) and requires no special amendments.

Nonetheless, as pointed out in Landau (Reference Landau2015: 35−37), a full reduction of OC PRO to shifted indexicals still faces significant difficulties. First, the distribution of shifted indexicals is very different from that of OC PRO, being restricted to a handful of verbs (often no more than three), in contexts of indicative (uncontrolled) complements, often distinguishing between first and second person pronouns, and so on, not to mention that OC is by far more common than indexical shift. Perhaps the most immediate challenge to this theory is the fact that unlike shifted indexicals, OC PRO conceals its indexical character and inherits its morphological [person] value from the controller (as indicated on agreeing elements).

1. 1. a. John planned [PRO_[pers:3] to promote himself/*myself].

   2. b. John planned: “I will promote myself.”

As Landau (Reference Landau2015: 37) observes,

Indeed, recognizing the agreement problem in the analysis of Korean OC, Seo and Hoe (2015) go as far as proposing that despite appearances, OC complements are not embedded jussive clauses but rather embedded subjunctive clauses whose heads do not encode indexical information.

The morphological features of PRO present a genuine problem for ESA theories of OC, which persists in more sophisticated implementations, like Matsuda (Reference Matsuda2019, Reference Matsuda, Mucha, Hartmann and Trawiński2021). As the authors before her, Matsuda assumes that OC complements are evaluated against a shifted context. However, in her account, PRO is not a minimal pronoun, but rather intrinsically valued for participant features – [+SP,–AD] for speaker, [–SP,+AD] for addressee, and [+SP,+AD] for inclusive speaker. These are crucially semantic features, anchored to a given context, as distinct from morphological person features. Matsuda proposes that when anchored to the utterance context, these features yield an overt pronoun with first or second person morphology. However, when anchored to the shifted context, they yield a controlled null third person subject – that is, PRO.

Distributional and typological issues that face the reduction of PRO to shifted indexicals face this proposal too. In addition, because it rejects the minimal pronoun analysis, Matsuda’s account is committed to a three-way lexical distinction between three types of PRO: a first person PRO for speaker control, a second person PRO for addressee control, and a person-less PRO for nonattitude control.

While the anchoring of the pronoun’s interpretation to context is semantically natural, the anchoring of its spellout to context leaves open the question of why null PRO cannot occur in embedded object positions. In Matsuda (Reference Matsuda2019: 109), this is attributed to agreement with T, governed by locality. However, Matsuda (Reference Matsuda, Mucha, Hartmann and Trawiński2021) discards this idea, presumably to avoid positing a [tense] feature on PRO. This overgenerates pronouns specified, for example, [+SP_c,–AD_c] (with c the reported context) even in embedded object positions. Finally, the possibility of first/second person PRO becomes a problem, especially when they are misaligned with the [SP]/[AD] coordinates of the utterance context.

1. 1. a. You decided [PRO to harm yourself].

      Controller: [–SP_c*,+AD _c*]; PRO: [+SP_c’,–AD _c’]

   2. b. You urged me [PRO to harm myself].

      Controller: [+SP_c*,–AD _c*]; PRO: [–SP_c’,+AD _c’]

Here, it seems that PRO morphologically agrees with the controller. However, because her spellout rules cannot generate first/second person on PRO, Matsuda is led to suggest that PRO is always third person, even in (40a) and (40b), and that the reflexives agree directly with the controller. This, however, runs counter to standard conceptions of agreement and binding, and (as Matsuda Reference Matsuda2019: 192 admits) cannot explain why the same agreement facts hold when the reflexive is read de se, namely, on its natural reading, where it is semantically bound by PRO.Footnote ²¹

Overall, then, ESA theories of OC offer an insightful treatment of the de se interpretation, of the occurrence of mood markers in OC complements, of controller choice, and of the existence of PC. However, they still grapple with the morphological agreement properties of OC PRO.

5.1 The Two-Tiered Theory: Complement Control

The starting point of the Two-Tiered Theory of Control (henceforth, TTC) is the recognition that Williams’ fundamental insight was correct: Some understood subjects are interpreted by predication while others require reference to some “logophoric” representation of antecedents. Let us consider OC by predication first.

That predication must be involved in some cases of OC is uncontroversial; reduced complements in restructuring environments, presumably, do not project a subject position, hence their external argument is shared with that of the matrix predicate via complex predicate formation (Wurmbrand Reference Wurmbrand, Zwart and Abraham2002, Reference Wurmbrand2003). What Landau (Reference Landau2015) proposed, adopting earlier analyses by Chomsky (Reference Chomsky1980), Chierchia (Reference Chierchia, Bartsch, van Benthem and van Emde Boas1990), and Clark (Reference Clark1990), is that clausal complements can also function as predicates if their PRO subject is treated as an operator. Specifically, a Fin head attracts PRO to its specifier, and the resulting chain is interpreted as λ-abstraction over the subject position. This is possible because PRO is a minimal pronoun devoid of any inherent denotation; hence, it cannot saturate a predicate, only form one. FinP is the projection where finiteness is encoded (Rizzi Reference Rizzi and Haegeman1997); while propositional, full clauses project a CP layer above FinP (see (58)), predicative clauses lack that layer.

1. (56) Predicative clause: [_FinP PRO_i Fin [_TP PRO_i …]]

What type of predicates instantiate OC this way, namely, by taking a predicative complement? These are the same predicates that do not license PC (see Section 3.1). Their common semantic property is that they do not introduce attitudes.Footnote ²⁴

1. (57) Predicative control: nonattitude predicates

   1. a. Implicativesdare, manage, make sure, bother, remember, get, see fit, condescend, try, avoid, forget, fail, refrain, decline, neglect, force, compel.

   2. b. Aspectualbegin, start, continue, finish, stop, resume.

   3. c. Modalhave, need, may, should, is able, must.

   4. d. Evaluative (adjectives)rude, silly, smart, kind, (im)polite, bold, modest, cruel, cowardly, crazy.

From the semantic point of view, this is a heterogeneous class. All but the modal predicates carry some actuality entailment. Thus, positive implicative verbs (remember) and evaluative predicates (rude) entail their complement, while negative implicative verbs (forget) entail its negation. Aspectual verbs (continue) entail that the complement holds to some (possibly incomplete) degree. While some of the verbs select experiencers, the truth or falsity of the complement is evaluated not against the mental attitude of that experiencer but rather against the actual world; this is what defines these verbs as nonattitude. On Landau’s (Reference Landau2015) analysis, a direct predication relation is established between the controller and the nonfinite complement. The resulting state of affairs is asserted to hold either in the actual world or (in the case of modal complements) in a set of possible worlds conforming to some norm. It is a defining feature of predicative control that the bearer of the property denoted by the complement is the referent of the controller in the actual world. Things are different in logophoric control, to which we now turn.

If predicative clauses constitute the first tier of OC, logophoric clauses are constructed as a second tier above them: A “perspectival” C head takes the FinP predicate of (56) as a complement, and projects a null pro as a specifier. This pro is associated (via a lexical presupposition on C) with a coordinate of the reported context – either the AUTHOR or ADDRESSEE of the matrix clause.Footnote ²⁵ The OC dependency is broken into two links: variable binding between the controller and pro, and predication between pro and FinP. Because PRO mediates this predication relation (being the λ-operator that binds the embedded subject position), it shares the features that pro received from the controller, and the FinP property is understood as holding of the AUTHOR or ADDRESSEE of the matrix clause. Note that the OC complement (= CP) denotes a proposition, differently from the OC complement (= FinP) of predicative control verbs, which denotes a property. This fundamental distinction has important repercussions, as we will see in the following.

1. (58) Logophoric clause: [_CP pro C_+log [_FinP PRO_i Fin [_TP PRO_i …]]]

The predicates instantiating OC by taking a logophoric complement are the same predicates that license PC. Their common semantic property is that they do introduce attitudes.

1. (59) Logophoric control: attitude predicates

   1. a. Factivesglad, sad, regret, like, dislike, hate, loath, surprised, shocked, sorry.

   2. b. Propositionalbelieve, think, suppose, imagine, say, claim, assert, affirm, declare, deny.

   3. c. Desiderativeswant, prefer, yearn, arrange, hope, afraid, refuse, agree, plan, aspire, offer, decide, mean, intend, resolve, strive, demand, promise, choose, eager, ready.

   4. d. Interrogativeswonder, ask, find out, interrogate, inquire, contemplate, deliberate, guess, grasp, understand, know, unclear.

In these OC complements, one observes the obligatory de se reading (when AUTHOR-controlled) or de te reading (when ADDRESSEE-controlled) of PRO; see Morgan (Reference Morgan, Campbell, James and Alice1970), Chierchia (Reference Chierchia, Bartsch, van Benthem and van Emde Boas1990), Percus and Sauerland (Reference Percus, Sauerland and Weisberger2003), Schlenker (Reference Schlenker2003), von Stechow (Reference von Stechow, Kadowaki and Kawahara2003), Anand (Reference Anand2006), Pearson (Reference Pearson2016, Reference Pearson2018), Hintzen and Martin (Reference Hintzen and Martin2021), and Pearson and Roeper (Reference Pearson and Roeper2022). These readings are consistently typical of OC under attitude predicates (but not of OC elsewhere). What they amount to is the observation that an attitude OC complement contributes information not about the matrix subject (AUTHOR) or object (ADDRESSEE), but rather about “the image” these participants have in the eyes of the attitude holder; technically, the “doxastic counterparts” of the matrix participants.

The specific character of de se/de te readings emerges in situation of misidentification, either of the self or of the addressee.

1. (60) Obligatory de seContext: John watches a clip of himself, caught on CCTV camera, making suspicious moves near some house at night. He doesn’t recognize himself but comes to believe that the person he watches is a burglar.

   1. a. John_i hoped that he_i would be caught. de re: true

   2. b. John_i hoped [PRO_i to be caught]. de se: false

1. (61) Obligatory de teContext: Betty is about to take Johnny to the movies. Earlier, Johnny messed up the living room, but Betty thinks it was Billy who did it. She tells Johnny: “Whoever messed up the living room should tidy it up!”

   1. a. Betty told Johnny_i that [he_i should tidy up the living room]. de re: true

   2. b. Betty told Johnny_i [PRO_i to tidy up the living room]. de te: false

Although de se readings in attitude OC contexts were thought to be unshakable, recent work suggests that at least with some OC verbs they are pragmatically defeasible. Pearson and Roeper (Reference Pearson and Roeper2022) observe that in contexts of ignorance disclaimers, such as unwittingly/unintentionally/unknowingly, or under the scope of in effect, the de se entailment is excluded. They describe a context in which Mary is the judge of a baking contest and, blindfolded, tastes her own cookie without recognizing it, and then declares that whoever baked that cookie deserves the prize. The following statements are all judged true in this scenario, despite the fact that Mary made (or intended) no first person statement of the kind “I deserve the prize.”

1. 1. a. Mary_i unwittingly/unintentionally/unknowingly claimed[PRO_i to deserve the prize].

   2. b. Without realizing it/In effect, Mary_i claimed [PRO_i to deserve the prize].

   3. c. Mary didn’t realize it, but she_i claimed [PRO_i to deserve the prize].

Taking in effect as their leading cue, Pearson and Roeper write:

Notably, Pearson and Roeper do not describe the readings in (62) as de re; these readings just share a certain property with de re (namely, foregrounding the speaker’s perspective). It is thus still possible to maintain that de se is definitional for PRO in attitude OC. Furthermore, the “excluded entailment” effect is lexically restricted; while present with claim, it is unavailable with want (see Pearson and Roeper Reference Pearson and Roeper2022 for details of the explanation).

Proceeding with the implications of de se construals, Landau (Reference Landau, Patel-Grosz, Grosz and Zobel2018) pointed out there is a tension between the leading accounts of how PRO comes to be obligatorily construed de se (or de te) and the leading accounts of how it comes to agree with the controller. In the former category, we have semantic models that take OC complements to denote properties or centered worlds (Chierchia Reference Chierchia1984, Reference Chierchia, Bartsch, van Benthem and van Emde Boas1990, von Stechow Reference von Stechow, Kadowaki and Kawahara2003, Stephenson Reference Stephenson2010, Pearson Reference Pearson2013). In these models, PRO is bound by a λ-operator located at the edge of the complement. This operator guarantees that the OC complement be of the right semantic type required by the attitude verb. Crucially, however, the operator bears no syntactic relation to the controller; they are only related in the semantics, and indirectly so. Thus, the robust fact of agreement in OC is left unexplained, assuming, as is standard, that agreement piggybacks some syntactic relation.

It is worth noting that the prospects of reducing agreement to semantic matching between the controller and PRO are rather grim. This can be seen with imposter or hybrid nouns, evincing a split between their formal and semantic features. The imposter noun the present authors governs either third person (formal) or first person (semantic) agreement (63a) (adapted from Collins and Postal Reference Collins and Postal2012: 19), and the German hybrid noun Mädchen governs either neuter gender (formal) or feminine gender (semantic) agreement (63b) (Wurmbrand, pers. comm.). The features of PRO are detectable on embedded, agreeing reflexives and pronouns, which must be locally bound by PRO.

1. 1. a. The present authors_i plan [PRO_i to devote themselves/ourselvesto ecology].

   2. b. Das Mädchen hat versprochen, [PRO_i sein/ihr Bestes zu geben].the girl has promised, its/her best to give“The girl promised to do her best.”

Without entering the intricate morphosyntax of hybrid nouns, suffice it to say that the mere availability of formal, nonsemantic agreement in attitude OC complements is an insurmountable obstacle for any attempt to reduce agreement to semantic matching. For semantic matching must make reference to the denotational values of the observed ϕ-features of PRO. However, these values can be either mismatched with the reference of the controller, or just uninterpretable. Thus, agreement in OC, including attitude OC, is an irreducible syntactic phenomenon. This phenomenon finds a natural explanation in the TTC insofar as it invokes variable binding and insofar as variable binding is a standard vehicle of agreement (Heim Reference Heim, Harbour, Adger and Béjar2008, Kratzer Reference Kratzer2009). By contrast, within the popular property-based view of OC, the explanation of agreement is “somewhat stipulative on every account” (Schlenker Reference Schlenker, von Heusinger, Maienborn and Portner2011: 1575).

Looking beyond agreement and de se, the TTC makes a range of predictions that have been largely confirmed. The first set of predictions follows from the highly strict nature of predication, which resists different types of “noncanonical” OC. Thus, control shift and split control are excluded with predicative complements but allowed (in principle) in logophoric ones. This is because a predicate in a given structure can only apply to a unique argument, whereas the intermediary logophoric pro in (58) can be anchored to either AUTHOR or ADDRESSEE coordinates, or possibly to their sum (Madigan Reference Madigan2008b). Likewise, “partial” readings are not obtainable under direct predication but can be modeled via an intermediate pronoun (see, e.g., Matsuda Reference Matsuda2019). For this reason, PC is typical of attitude complements.Footnote ²⁶

The possibility of implicit control into predicative complements is less clear. Landau (Reference Landau2015) argued (following Landau Reference Landau2010) that predicates must be saturated by overt arguments, and therefore controllers may not be implicit in aspectual or implicative constructions, as opposed to desiderative ones. In (64), this is illustrated with subject control verbs in Hebrew, where the controller becomes implicit upon passivization.

1. 1. a. Logophoric controlhuxlat / tuxnan / huvtax le’hitkadem ba-proyekt.was.decided / was.planned / was.promised to.move.forward in.the.project ‘It was decided/planned/promised to move forward with the project.’

   2. b. Predicative control*hufsak / nusa / niškax le’hitkadem ba-proyekt.was.stopped / was.tried / was.forgotten to.move.forward in.the-project‘It was stopped/tried/forgotten to move forward with the project.’

The same split is attested with object control verbs in Hebrew: Object drop is possible only with desiderative verbs and not with implicative verbs.

1. 1. a.

      Logophoric object control
      ha-menahel		civa / pakad / asar / laxac			(alay)
      the-manager		ordered / commanded / prohibited / pressured (on.me)
      lešatef		pe’ula	ba-misrad.
      to.cooperate		action	in.the-office
      “The manager ordered / commanded / forbade / pressured (me)
      to cooperate in the office.”

   2. b.

      Predicative object control
      Gil	kafa / hikša / hekel / hišpia				*(alay)
      Gil	compelled/made.it.difficult/made.it.easy/influenced *(on.me)
      le’hitpater	etmol.
      to.quit	yesterday
      “Gil compelled / made it difficult for / made it easy for / influenced *(me) to quit yesterday.”

Against this evidence from Landau (Reference Landau2015) (who cites parallel data in Polish and Russian), Pitteroff and Schäfer (Reference Pitteroff and Schäfer2019) point out that examples parallel to (64b), such as the Dutch example (66), are possible in a number of Germanic languages, although there is “huge variation” in judgments.

1. (66)

   Er	werd	vergeten/verzuimd	(om)	als	collectief	te	spelen,
   there	was	forgotten/missed	C	as	collective	to	play,
   juist	wat	normaliter	de	sterke	kracht	is	van	het	team.
   just	what	normally	the	strong	power	is	of	this	team.
   “People forgot/failed to play as a collective, which usually is the strength of this team.”

It is notable that counterexamples to (65b) have not been reported.Footnote ²⁷ One can tentatively conclude that cross-linguistic variation in the control capacity of implicit arguments is only attested with implicit agents of passive. Recent work indeed suggests that the binary distinction between Active and Passive Voice is an oversimplification; Voice heads come in different degrees of “strength” and featural specification (Legate Reference Legate2021, Sigurðsson and Wood Reference Sigurðsson and Wood2021). It is thus quite possible that these Voice heads, or the external arguments they project, would correspondingly vary in their visibility to grammatical dependencies such as control, predication, and binding. Fleshing out the details of such a typology is a task for future research.

Let us mention two further consequences of the TTC, one concerning the size of the control complement and the other concerning its semantic type. Comparing (56) and (58), we observe that predicative clauses are smaller than logophoric ones, lacking at least the CP layer (and even more layers in restructuring complements). This is to be expected insofar as the former express modality, manipulation, phasal status, and achievement, while the latter express desires, fears, epistemic states, and speech acts. Typological research has shown that clausal complements of the latter type are larger than those of the former type, the difference keyed to the occurrence of designated heads in the left periphery of the clause, encoding the different types of attitudes (Lohninger and Wurmbrand Reference Lohninger, Wurmbrand, Benz, Frey, Krifka, McFadden and Żygis2024).

Concretely, to the extent that such functional heads are spelled out, we expect the left-periphery of logophoric complements to be more richly specified than that of predicative complements. Looking at complementizers, this tendency is consistent, although often concealed due to the common syncretism between functional heads in the left periphery (usually at most one is pronounced). In some languages, infinitival complementizers distribute across verb classes with hardly any systematic restrictions (i.e., a/de/di prepositional complementizers in Romance, að in Icelandic); in others, attitude complements are introduced by overt complementizers and nonattitude complements are not. Notably, no language exhibits the opposite pattern (overt complementizers in nonattitude complements and no complementizer in attitude complements).

As an illustration, consider the distribution of the Polish complementizer żeby (Bondaruk Reference Bondaruk2004, Citko Reference Citko2012). It is obviative with subjunctive complements and with some desiderative infinitival complements. With object control verbs and a few subject control verbs, it has no effect on control, but with many subject control verbs its presence licenses NC. What is relatively clear is that predicative complements (modal, aspectual, and implicative) reject żeby; it only occurs in logophoric complements. For example, żeby may or must occur with desiderative and epistemic subject control complements (67b) and (67c) but not with implicative complements (67a).

1. 1. (a.)

      Jani	zdołał	[(*żeby) PROi/*j	śpiewać].
      Jan	managed	C	sing.INF
      “Jan managed to sing.”

   2. b.

      Jani	wolał	[PROi/*j	śpiewać] /	Jani	wolał	[żeby pro*i/j	śpiewać].
      Jan	preferred		sing.INF	Jan	preferred	C	sing.INF
      “Jan preferred to sing. / Jan preferred for others to sing.”

   3. c.

      Jan marzył	[*(żeby) PROi/*j	śpiewać].
      Jan dreamed	C	sing.INF
      “Jan dreamed to sing.”

A parallel contrast exists between implicative and desiderative object control complements.

1. 1. a.

      Mareki	dał	mi	[(*żeby) PROi/*j	poprowadzić	swój	samochód].
      Mark	let	me	C	drive.INF	his	car
      “Mark let me drive his car.”

   2. b.

      Jan	kazał	Piotrowii	[(żeby) PROi/*j	nie biegać	po ulicy].
      John	told	Peter	C	not run.INF	on street
      “John told Peter not to run on the street.”

A similar distribution is displayed by the jussive modals yao “must/will/want” and its negative counterpart bie “NEG.IMP” and bu-yao “NEG-should/NEG.IMP” in Mandarin Chinese (Liao and Wang Reference Liao and Wang2022). These are banned from predicative OC complements (69) (which must use standard negation instead) and occur only in logophoric ones (70) (note that (70b) exhibits PC as well). Liao and Wang analyze them semantically as the imperative operator of Stegovec (Reference Stegovec2019) and identify them with Landau’s C^OC, the complementizer of logophoric OC complements.

1. 1. a.

      Zhangsanii	kaishi	[*bie/bu PROi	zai-nali	tiaowu].
      Zhangsan	start	NEG.IMP/NEG	in-there	dance
      “Zhangsan started to not dance in that place.”

   2. b.

      Zhangsani	neng(gou)	[*bie/bu PROi	zheme	zuo].
      Zhangsan	can	NEG.IMP/NEG	so	do
      “Zhangsan is able to not do so.”

1. 1. a.

      Zhangsani	dasuan	[bie/*bu PROi	tai-zao	jiehun].
      Zhangsan	intend	NEG.IMP/NEG	too-early	marry
      “Zhangsan intended not to get married too young.”

   2. b.

      Zhangsan	quan	Lisii [bie/*bu PROi+	zai-liu-dian	jianmian].
      Zhangsan	persuade	Lisi NEG.IMP/NEG	at-six-o’clock	meet
      “Zhangsan persuaded Lisi not to meet at six.”

In fact, it is not required that the Fin head of predicative complements be null, as long as it is distinct from the C head of logophoric complements. A language where both heads are overt and distinct may well be Moro (Kordofanian; South Sudan), as documented in Jenks and Rose (Reference Jenks, Rose, Kandybowicz and Torrence2017). The complementizer nə́- only occurs in implicative complements (71a), whereas the complementizer t̪á only occurs in logophoric complements (71b).Footnote ²⁸

1. 1. a.

      kúk:u g-ɘndɘtʃɘn-ú	(n)-áŋə́ -↓lə́və́tʃ-a	ŋál:o(-ŋ)
      Kuku CLg-(RTC-)try-PFV	C−3SG.INF-hide-INF	Ngalo-ACC
      “Kuku tried to hide Nghalo.”

   2. b.

      é-g-a-mwandǝð-ó	kúk:u-ŋ	t̪á	ɜ́ŋ-ɘ́-↓búg-í	ís:íɘ.
      1SG-CLg-RTC-ask-PFV	Kuku-ACC	C	3SG.INF-give-INF	CLj.gun
      “I asked Kuku to shoot the gun.”

What is even more striking is that nə́- also introduces nonsubject relative clauses in Moro. Given the predicative analysis of OC in (71a), it is a pleasant convergence to find the same complementizer introducing both predicative OC complements and predicative modifier clauses.

A final cross-linguistic pattern that naturally falls into place under the TTC involves the tolerance of OC complements to lexical subjects. While attitude and nonattitude complements exhibit the same OC signature when their subject is null (except for the de se/de te property), they differ dramatically in their tolerance to a lexical subject: Only attitude complements allow it.

1. (72) Complements of attitude predicates allow lexical subjects; complements of nonattitude predicates disallow lexical subjects.

Generaization (72) is very broad. It crucially abstracts away from the syntactic realization of the complement – finite or not, nominalized or not, and so on. It was first stated by Grano (Reference Grano2015: 19) in terms of the EC-PC contrast, but we have seen that this contrast reflects the deeper cut between attitude and nonattitude complements.

In some languages, no special grammatical device is needed to license a lexical subject in an infinitival complement (thus turning it from OC to NC); this is the situation in Malayalam (Mohanan Reference Mohanan1982) and in Tamil (Sundaresan and McFadden Reference Sundaresan and McFadden2009), the latter illustrated in (73). Note that a null subject is obligatorily controlled under both the implicative try and the desiderative want, but a lexical subject is tolerated only under want.

1. 1. a.

      ramani [PROi/*j/	(*Anand)	saadatt-ai	saappiɖ-a]	paa-tt-aan.
      Raman.NOM	Anand	rice-ACC	eat.INF	try-PST−3 M.SG
      “Raman tried (*Anand) to eat rice.”

   2. b.

      champa-vukkui	[PROi/*j/	Sudha oru samosa-vai saappiɖ-a]	veɳɖ-um.
      Champa-DAT		Sudha a samosa-ACC eat-INF	want−3 N.SG
      “Champa wants (Sudha) to eat a samosa.”

The contrast appears to be replicated in English, but this is only because want is an ECM verb, which can license an embedded subject, as opposed to, for example, decide. What is quite clear in the Dravidian languages in that the embedded subject is licensed regardless of the particular matrix verb; all attitude complements, finite or nonfinite, accept lexical subjects. Even more striking evidence comes from Irish, a language notorious for its free licensing of (accusative) lexical subjects in infinitives (McCloskey Reference McCloskey1980a, Reference McCloskey, Goldberg, Mackaye and Michael1985, McCloskey and Sells Reference McCloskey and Sells1988, Bondaruk Reference Bondaruk2006). Yet McCloskey (Reference McCloskey1980b) observed that the nonfinite complements of some predicates reject a lexical subject. Although he has not attempted any generalization, the examples he provided, given in (74a) and (74b), contain modal and implicative verbs, namely, nonattitude predicates (Irish infinitives are formed from Verbal Nouns [VN], usually following the particle a). Attitude predicates, on the other hand, allow either OC or NC ((74c) and (74d) are from Bondaruk Reference Bondaruk2006, and (74e) is from McCloskey Reference McCloskey, Goldberg, Mackaye and Michael1985).

1. (74) Nonattitude infinitival complements in Irish: ✓PRO, *lexical subject

   1. a.

      Ní thiocfadh proi libh [PROi/	*Nollaig	imeacht	chomh luath sin.
      NEG could.2SG	Noel	leave.VN	so.early
      “You couldn’t (*Noel) leave so early.”

   2. b.

      Rinne	séi	iarracht [PROi/*na		daoine teach	a	thógáil].
      made	he	attempt	those	people house	PRT	build.VN
      “He tries (*those people) to build houses.”
      Attitude infinitival complements in Irish: ✓PRO, ✓lexical subject

   3. c.

      Ba	mhaith	liom proi [PROi	imeacht	/ é	a	meacht].
      COP	good	with.1SG	go.VN	him	PRT	go.VN
      “I would like (him) to go.”

   4. d.

      Tá	mé	sásta [PROi	a	bheith	anseo].
      am	I	glad	PRT be.	VN	here
      “I’m glad to be here.”

   5. e.

      Bheinn	sásta	[iad	a	bheith	i láthair].
      I.would.be	glad	them	PRT	be.VN	present
      “I would be glad for them to be present.”

To see the effect of (72) in English, observe that for-infinitivals are excluded in nonattitude complements (Grano Reference Grano2015: 19). The marginal acceptability of lexical subjects in the complements of certain implicative verbs, as in (75b), is due to lexical coercion, by which the core meaning of the verb is modified or extended (see Jackendoff and Culicover Reference Jackendoff and Culicover2003, Grano Reference Grano2017b). However, language-internal factors, partly arbitrary in nature, somewhat obscure the picture of English infinitivals. For example, not all desiderative verbs take a for-infinitive (75c), and wh-infinitivals, although expressing attitudes, consistently reject lexical subjects (75d) (Chomsky and Lasnik Reference Chomsky and Lasnik1977).

1. 1. a. *John began/had [for Bill to solve the problem].

   2. b. #John tried/managed [for Bill to solve the problem].

   3. c. *Mary persuaded Fred [for the kids to buy ice cream].

   4. d. *She wondered [where (* (for) people) to go].

All these “irregularities” of infinitives disappear under finite complementation, where lexical subjects are uniformly licensed. The problem is that nonattitude predicates also reject finite complements, so there does not seem to be any clean testing ground in English where the empirical consequences of (72) can be fully observed across the attitude–nonattitude divide.

In fact, there is one testing ground that overcomes these difficulties – gerunds. English gerunds have no trouble licensing an internal subject. Therefore, when placed in complement position, the (non)availability of a lexical subject inside the gerund purely reflects general principles. As it happens, it precisely reflects (72).

Pires (Reference Pires2007) pointed out that gerundive complements fall into two classes: One class, which he described as [+tense], accepts either PRO or a lexical subject, whereas the other class, [–tense], only accepts a PRO subject. Following Landau’s (Reference Landau2015) restatement of the [±tense] distinction in terms of [±attitude], this split is aligned with (72).

1. (76) Nonattitude gerundive complements: ✓PRO, *lexical subject

   1. a. Philip_i tried/avoided [PRO_i/*Jane driving in the freeway].

   2. b. Jane_i started/resumed [PRO_i/*Frank talking to us].

   Attitude gerundive complements: ✓PRO, ✓lexical subject

   1. c. Sue_i favored/insisted on [PRO_i/Anna moving to Chicago].

   2. d. They_i imagined/suggested [PRO_i/Paul joining the trip].

English has another variant of gerundive complementation, where the gerund is introduced by a preposition. These P-gerund constructions, studied in Landau (Reference Landau2021b), also fall into two broad categories: a causative-implicative class and a nonimplicative class. Each class consists of several subclasses, depending on the preposition involved. What is crucial to note is that all and only the complements of the implicative (nonattitude) verbs reject a lexical subject.

1. • Nonattitude P-gerund complements: ✓PRO, *lexical subject

   1. a. She fooled us_i into [PRO_i/*Bob) thinking she was sick].

   2. b. They confined Beth_i to [PRO_i/*her son’s eating dog food].

   3. c. John talked Sue’s partner_i out of [PRO_i/*her accepting a bribe].

   4. d. John restrained Sue_i from [PRO_i/*her candidate’s makinga long statement].

      Attitude P-gerund complements: ✓PRO, ✓lexical subject

   5. e. He_i accused me of [PRO_i/Frank being suspended].

   6. f. Bill credits Jane_i with [PRO_i/David finding the courage to resist].

   7. g. I’d caution you_i against [PRO_i/anyone taking this too far for now].

As Landau (Reference Landau2021b) shows, this split is correlated with a cluster of other contrasts, all predictable from the different ways in which control is established in predicative constructions as opposed to logophoric constructions.

Generalization (72) falls out naturally from the TTC’s commitment to the underlying duality of OC complementation – (56) versus (58). In predicative control, the complement denotes a predicate; specifically, the derived predicate FinP. In logophoric control, the complement denotes a proposition; specifically, the proposition obtained from applying the predicative FinP to the logophoric pro in [Spec,CP]. Thus, control complements do come in two semantic types, contra the uniformist property-theories stemming from Chierchia Reference Chierchia1984, but in line with the fundamental duality envisioned in Wurmbrand (Reference Wurmbrand, Zwart and Abraham2002). Because nonattitude verbs select properties, their complements cannot host a lexical subject, which would saturate the property and yield a type mismatch. Because attitude verbs select propositions, their complements may surface with a lexical subject. Importantly, either (58) (with PRO) or a clause with a lexical subject may satisfy the selectional requirement for a proposition. While the latter option is not consistently available in all nonfinite environments (due to poorly understood distributional restrictions), it is robust enough in those environments that allow a “fair” testing.

In the next section, we will see how the property–proposition divide extends further to account for the fundamental arrangement of nonfinite adjuncts with respect to control.

5.2 The Two-Tiered Theory: Adjunct Control

Throughout much of the history of work on control, adjuncts have received considerably less attention than complements. Some major theories of control have been developed without considering adjuncts at all (Sag and Pollard Reference Sag and Pollard1991, Landau Reference Landau2000, Jackendoff and Culicover Reference Jackendoff and Culicover2003); others mention adjuncts summarily, illustrating one or two types at the most, only to assimilate them to standard OC (Manzini Reference Manzini1983, Mohanan Reference Mohanan1983, Clark Reference Clark1990, Hornstein Reference Hornstein1999, Reference Hornstein and Hendrick2003, Pires Reference Pires2007, Fischer Reference Fischer2018, McFadden and Sundaresan Reference McFadden and Sundaresan2018). This situation has fortunately changed in recent years, with concentrated studies of adjunct control that pay much closer attention to the different types of adjuncts, their modificational flavor and how it relates to the OC/NOC classification (Green Reference Green2018, Reference Green2019, Landau Reference Landau2021a, Fischer and Flaate Høyem Reference Fischer and Flaate Høyem2022; see Williams Reference Williams, Larson, Iatridou, Lahiri and Higginbotham1992 for an important precursor). In this section, I lay out the major results of this recent work and the research questions it raises.

It was Williams (Reference Williams, Larson, Iatridou, Lahiri and Higginbotham1992) who first pointed out that adjuncts display a dual behavior: Sometimes they pattern with OC and sometimes with NOC. Following earlier observations in Chomsky (Reference Chomsky1981: 324–327), Williams noted that PRO in NOC must be [+human], a feature he attributed to its logophoric nature. Obligatory control, in contrast, operates by predication, accepting inanimate subjects and even weather-it. By way of illustration, consider the following examples.

1. 1. a.

      Around here, it can’t snow [before PROi raining].

      OC

   2. b.

      There won’t be any progress [without PRO insisting	NOC
      on guidance from the outside].

Williams assumed that OC and NOC adjuncts are attached at different heights; specifically, that NOC adjuncts are higher than the subject. This idea is part of what Landau (Reference Landau2021b: 93) calls “the classical view.” On this view, OC and NOC in adjuncts are in complementary distribution, because (i) OC requires c-command by the controller, NOC does not, and (ii) OC is mandatory whenever possible (NOC is “last resort”). It follows from the classical view that NOC adjuncts are not c-commanded by the subject.Footnote ²⁹

However, Green (Reference Green2018) and Landau (Reference Landau2021b) marshal a series of arguments refuting the classical view. To begin with, OC and NOC are not mutually exclusive. Although normally in competition, the nature of this competition is pragmatic rather than grammatical, hence it is not a rigid constraint. Careful testing reveals that OC and NOC can even coexist in the same sentence (Green Reference Green2018: 40). Thus, “last resort” and “elsewhere case” are the wrong concepts to describe the interplay between OC and NOC in adjuncts.

1. (79) The pool_i was the perfect temperature [after PRO_i/arb being in the hot sun all day].

Second, NOC adjuncts can demonstrably attach below the subject. As Landau (Reference Landau2021b: 95–96) shows, they can scope below negation and be elided as part of VP-ellipsis. Note that the rationale clause in (80a) involves speaker-control and the temporal clause in (80b) involves arbitrary control, two types of NOC.

1. 1. a. The door is not open in order to greet anyone, I just needed some fresh air.

   2. b. In the summer, the night sky is frequently an unforgettable spectaclewhen camping in the desert, but in the winter it rarely is.

Third, even lower adjuncts, located inside the VP, may display NOC. Consider Object Purpose Clauses, whose embedded object is bound by (an operator bound by) the matrix theme, and whose PRO subject may display NOC. The following examples demonstrate that this NOC reading persists even though the adjunct is necessarily VP-internal (hence, resists stranding under VP-ellipsis).

1. 1. a. The sterile bandages have been placed in the small backpack[PRO_arb to use in case of serious injury],and the plasters have been ___ too.

   2. b. * The sterile bandages have been placed in the small backpack[PRO_arb to use in case of serious injury],but the plasters have been ___ [PRO_arb to use for bruises only].

Conversely, TP-adjuncts may display OC despite claims that this position is reserved for NOC (e.g., Fischer and Flaate Høyem Reference Fischer and Flaate Høyem2022). In (82), the matrix subject can be neither a logophoric antecedent (being inanimate) nor a topic (being a negative quantifier). Correspondingly, it does not qualify for NOC. The observed control relation into the initial adjunct, therefore, must be OC.

1. (82) After falling into this acid, nothing survives.

See Landau (Reference Landau2021b: 94–98) for further, extensive evidence against the classical view.Footnote ³⁰

Instead of positing unmotivated correlations between an adjunct’s position and the OC–NOC distinction, we should look for a theory that ties the structural position of adjuncts to something else. Landau (Reference Landau2021a) argues that this “something else” is compositionality: OC adjuncts and NOC adjuncts are of different semantic types, and this difference explains which syntactic nodes they may combine with (to yield an interpretable adjunction structure) versus which they cannot (due to a type mismatch). Crucially, the semantic-type distinction is the same fundamental distinction already assumed in the TTC for complement OC; namely, the distinction between predicative clauses and propositional clauses.

While complement clauses split into these two types directly, as in (56)/(58), adjunct clauses split indirectly, via a mediating P head. This P head may be overt (before, without, in order, etc.) or null, as it is in rationale clauses without in order, stimulus clauses (e.g., He smiled to see my response), and so on.

1. 1. a. Predicative adjunct: [_PP P_Pred [_FinP PRO_i Fin [_TP PRO_i …]]]

   2. b. Propositional adjunct: [_PP P_Prop [_CP pro C_+log [_FinP PRO_i Fin [_TP PRO_i …]]]]

The way these two semantic types map to control types is asymmetric. Propositional adjuncts can be realized either as NC (with a lexical subject inside the adjunct) or NOC. Predicative adjuncts can be realized either as strict OC adjuncts, namely, adjuncts that never alternate with NOC or NC, or as the OC variant of alternating OC/NOC adjuncts.

The understanding that controlled adjuncts split in this asymmetric fashion is fairly recent. Most earlier work simply assumed that adjuncts display OC or NOC. However, Green (Reference Green2018, Reference Green2019) and Landau (Reference Landau2021a) established a sharp dichotomy between adjuncts that never alternate with NOC and those that do.

1. (84) Controlled adjuncts in English

   1. a. Strict OC adjuncts: Goal, Result, Stimulus, Subject Purpose Clause (SPC)

   2. b. Alternating OC/NOC adjuncts: Object Purpose Clause (OPC), Rationale,

      Temporal, Absolutive, Justification, Telic.

For a full, systematic empirical description of all these adjuncts, see Landau (Reference Landau2021a). Here, we will just present a few illustrative contrasts. A result clause modifies an inchoative event by elaborating on its result (85a). As a strict OC adjunct, its controller must be the matrix subject and no other contextually available antecedent (85b), although this would be possible with a temporal adjunct (85c) on its NOC variant, as often happens when the local subject is inanimate.

1. 1. a. The door_i opened again [PRO_i to reveal a strangely decorated room].

   2. b. *The door_i opened again [PRO to try to close it_i].

   3. c. The door_i opened again [after PRO trying to close it_i]. (Green Reference Green2019: 16)

Because strict OC adjuncts are necessarily predicative, they require the presence of an overt controller (see (64b)/(65b)). Alternating OC/NOC adjuncts, on the other hand, may access an implicit controller via the propositional NOC route. We thus predict a contrast in tolerance to implicit control. The following examples compare an SPC (strict OC) and an OPC (alternating OC/NOC), confirming the prediction: Only the latter allows the controller to be dropped in the presence of the adjunct (note that the relevant objects are optional, in principle).

1. 1. a.

      We’re now hiring (people).

   2. b.

      We’re now hiring *(peoplei) [PROi to manage the marketing for us]. SPC

   3. c.

      They provided (mei) a connector cable [PROi to charge my	OPC
      device with].

The cut between the two kinds of adjuncts ultimately reflects the distinction between the basic semantic types – property versus proposition. One of Landau’s (Reference Landau2021a) main point is that this semantic type is independently detectable by the existence of a “propositional variant” for the controlled adjunct. A propositional variant for a controlled adjunct is an adjunct introduced by the same preposition, contributing the same modificational relation, except that it hosts a lexical subject. While some adjuncts allow an alternation between PRO and DP (… without Bill/PRO saying much), others restrict the propositional variant to finite clauses (… while PRO watching TV / while Bill watched TV). However, insofar as such a variant exists, it indicates that the adjunct head may either be P_Pred or P_Prop, and the adjunct is classified as alternating OC/NOC. When no such variant exists, the head of the adjunct is uniquely P_Pred, and the adjunct is classified as strict OC. The propositional variant criterion states that adjuncts without a propositional variant are strictly predicative and require strict OC, while adjuncts with a propositional variant are either predicative (OC) or propositional (NOC).

For example, a stimulus clause may not host a lexical subject but a Justification Clause may, (87a) versus (88a). Correspondingly, the former rejects non-c-commanding control (a sign of NOC), whereas the latter accepts it, (87b) versus (88b).

1. 1. a. Bill wept [(*for his wife) to hear the tragic story].

   2. b. [Her_i kids]_j wept [PRO_j/*i to hear the tragic story].

1. 1. a. Our life was blessed [for her being so much a part of it].

   2. b. Her_i kids were punished [for PRO_i letting them ruin the place].

This state of affairs is summarized in table (89).

1. (89) S-selection in nonfinite adjuncts

It is striking that Strict NOC adjuncts are unattested. The best candidate for such adjuncts are speech act oriented (SAO) adjuncts (Quirk et al. Reference Quirk, Greenbaum, Leach and Svartvik1985: 1068–1073, Kortmann Reference Kortmann1991, Meinunger Reference Meinunger2006, Lyngfelt Reference Lyngfelt2009, Duffley Reference Duffley2014: 99–102), where PRO is controlled by the speaker in declarative contexts and by the addressee in interrogative contexts.

1. 1. a. [PRO judging from my/*your experience], John would be better offwithout Mary.

   2. b. [PRO judging from your/*my experience], would John be better offwithout Mary?

However, Landau (Reference Landau2021a) points out that SAO adjuncts are subject to severe restrictions that do not generally apply to NOC adjuncts. First, while both types of adjuncts undergo perspectival shift when embedded, an SAO adjunct is uniquely anchored to either AUTHOR or ADDRESSEE, not both.

1. 1. a. NOC adjunct (PRO = reported AUTHOR/ADDRESSEE)

      John_i told Mary_j that [PRO_i/j having such experience],

      this job would be a piece of cake.

   2. b. SAO adjunct (PRO = reported AUTHOR)

      John_i told Mary_j that [PRO_i/*j judging from experience],

      such offers were very rare.

Tellingly, SAO adjuncts reject lexical subjects, unlike all other NOC adjuncts.

1. 1. a. NOC adjunct

      [His income now being secure], John is better off without Mary.

   2. b. SAO adjunct

      * [For me to be absolutely frank], John would be better off without Mary.

Landau proposes that SAO adjuncts are in fact OC adjuncts, predicated of a null nominal in the left periphery of the clause, representing the AUTHOR or ADDRESSEE.

Given this picture, table (89) represents a genuine asymmetry that calls for explanation. By default, all controlled adjuncts map to predicates; a subset of them maps to propositions, as an additional option. What is the source of the predicative default? Landau claims that it is rooted in Economy of Projection (EoP): All else being equal, a more minimal structure is favored over a less minimal one (Chomsky Reference Chomsky and Freidin1991, Safir Reference Safir1993, Grimshaw Reference Grimshaw, Lust, Suñer and Whitman1994, Bošković Reference Bošković1996, Speas Reference Speas, Ackema, Brandt, Schoorlemmer and Weerman2006). Looking at the structures of the two types of adjuncts in (83a) and (83b), it is clear that the predicative adjunct is more minimal than the propositional one, FinP being properly contained in CP. Predicative adjuncts account for all instances of local control; hence, they are the first choice of the child in parsing input conveying this interpretation. As positive evidence for a propositional variant, in the shape of adjuncts with lexical subjects, is gradually accumulated, the child would entertain an additional, propositional variant, along with the suitable head (P_Prop) (see Landau Reference Landau2021a: 165–171 for developmental evidence). This, in turn, would grant these adjuncts the option of NOC.

Note that EoP only regulates the choice between different derivations of the same interpretation. Yet OC and NOC often compete for distinct interpretations. An oft-made observation is that a local human controller is strongly favored over any alternative construal, for example, long-distance or arbitrary control (Kortmann Reference Kortmann1991, Kawasaki Reference Kawasaki1993, Lyngfelt Reference Lyngfelt2000). This has led many researchers to misclassify adjunct control as OC, as in the Italian example (93a) (Sundaresan Reference Sundaresan2014). Yet if the matrix subject is inanimate, the same type of adjunct, in the same position, supports extra-sentential control (93b) (Landau Reference Landau2021b: 106).

1. 1. a.

      [PROi/*j	detestando	il	pesce],	Giannii	compró	solo	carne.
      	detest.GER	the	fish	Gianni	bought	only	meat
      “[PROi/*j detesting the fish], Giannii bought only meat.”

   2. b.

      [PRO	detestando	gli	altri],	la	vita	diventa	difficile.
      	hate.GER	the	others	the	life	gets	difficult
      “Life gets difficult when one hates the others.”

This preference is strong but not absolute (attesting to its pragmatic source). Occasionally, a local human controller can be skipped, as in the following examples from Español-Echevarría (Reference Español-Echevarría, Billerey and Lillehaugen2000: 101) and Green (Reference Green2018: 36).

1. 1. a. Bill_i will introduce the ambassador to the president [in order PRO to give him_i the opportunity to observe their reactions].

   2. b. Strangely, the candidates talked avidly when we_i asked them where they were from, but they hesitated [after PRO_i asking them about their work].

Why is a local human controller so strongly favored, though? The answer is straightforward. The OC derivation is guided by the locality of predication, so it clearly picks out the local subject. The NOC derivation is guided by salience along two dimensions – logophoricity and topicality (see Section 4). A human antecedent is already ranked high on these two scales, as humans are the canonical perspective holders and also make better sentence topics (on the animacy hierarchy, see Comrie Reference Comrie1981, DuBois Reference DuBois1987, Song Reference Song2001, Swierskia Reference Swierskia2004, Kuno Reference Kuno, Horn and Ward2006). Thus, in the common scenario of a local human subject, both the OC derivation and the NOC derivation converge on the same reading. Indeed, we might suppose that EoP will favor the OC derivation under these circumstances. Any nonlocal control interpretation would have to be supported by exceptional salience of the remote antecedent, surpassing that of the local human subject.

This indeed is quite rare, leading to the false impression that the OC reading somehow “blocks” the NOC reading. Strictly speaking, however, this never happens. Only when the readings coincide do the two derivations compete directly (and EoP adjudicates in favor of OC). Otherwise, the choice between the reading delivered by OC and the one (or several ones) delivered by NOC is a matter of degree, a complex interaction of pragmatic salience, processing, linear order, and so on. All too often the NOC reading will be ranked so low as to be practically inaccessible. Yet as far as the grammar is concerned, it is generated alongside the OC reading.