1 Introduction

A fundamental question for control theory is the distribution question: what principles regulate the relative distribution of controlled arguments (also known as PRO)1 as compared against non-controlled arguments (which include full NPs, ordinary pronouns, and pro)? At least two cross-linguistically valid dimensions to PRO’s distribution appear to be purely syntactic in nature: with very few possible exceptions, PRO can appear only in subject position, as illustrated in (1), and only in a nonfinite clause, as illustrated in (2). (See e.g., Landau 2013: Chapter 4 for an overview).

(1) a.   John wants [PRO to see Bill].
  b. *John wants [Bill to see PRO].
(2) a.   John claims [PRO to be tall].
  b. *John claims [that PRO is tall].

The conclusion that the contrast in (1) is purely syntactic in nature is based on the premise that there is no semantic factor on which the ungrammaticality of (1b) could be blamed; supporting evidence comes from the fact that if the embedded clause in (1b) is passivized so that PRO ends up in subject position, the result is grammatical, as in (3). On the intended reading, (1b) and (3) are truth-conditionally equivalent, yet only the latter is grammatical.

(3) John wants [PRO to be seen by Bill].

Similarly, the conclusion that the contrast in (2) is purely syntactic is based on the premise that there is no (relevant) semantic difference between finite and nonfinite clauses on which the contrast in (2) could be blamed; supporting evidence comes from the fact that on the intended reading, (2a) and (2b) are truth-conditionally equivalent, yet only the former is grammatical.

But when these two dimensions are factored out, there are residual distributional puzzles in PRO’s behavior whose source is more obscure. This paper takes as its focus one such puzzle, namely the observation that the verb try readily accepts control complements (4a) but rejects non-control complements (for many speakers of English)2 (4b); cf. want which comfortably accepts both control (5a) and non-control complements (5b) equally without fuss.

(4) a.    John tried [PRO to open the door].
  b. %John tried [for Bill to open the door].
(5) a.    John wanted [PRO to open the door].
  b.    John wanted [(for) Bill to open the door].

What is the source of the asymmetry between (4) and (5)? At least two analytical options are a priori conceivable. On one view — call it the idiosyncrasy approach — the asymmetry between (4) and (5) reflects a purely arbitrary, lexically idiosyncratic accident of (some dialects of) English: want syntactically selects for both control and non-control complements, whereas try syntactically selects for control complements only. On this view, the observed split is no deeper or more interesting than the fact that the verb depend just so happens to select for the preposition on whereas the verb wait just so happens to select for both on and for.

On another view — call it the semantic approach — the asymmetry between (4) and (5) reflects a principled semantic difference between want and try: want has a denotation whereby control and non-control complements both yield sensible truth conditions whereas try has a denotation whereby non-control complements yield structures that are either outright uninterpretable or otherwise semantically anomalous (in the relevant dialects).3

What factors are at stake in deciding between these two approaches? Two are critical. The first is the cross-linguistic picture. Broadly speaking, the idiosyncrasy approach predicts no systematic cross-linguistic agreement in how the meaning of an embedding predicate relates to its complementation options, whereas the semantic approach does predict such agreement. The second is independent semantic evidence: if try rejects non-control complements on semantic grounds, then such an analysis ought to be consistent with — if not predictable from — our most sophisticated understanding of the semantics of try-sentences.

Turning first to the cross-linguistic factor, what we in fact see is neither fully regular nor fully irregular; instead, we see a kind of systematically constrained irregularity. As pointed out by Grano (2015b) — from whom Table 1 below is borrowed — we see cross-linguistic agreement for want in accepting non-control complements, and cross-linguistic agreement for aspectual verbs like begin in rejecting non-control complements (as long as the syntax of the language is respected — in many cases this means using a subjunctive clause in testing for the possibility of a non-control complement).4 This is exactly what we would expect if there is a semantic component to PRO’s distribution but unexpected if the relative distribution of control and non-control complements is lexically idiosyncratic. But for try, we see instability, both across dialects of English, as well as across languages, with some including French and Mandarin rejecting non-control complements but others like Greek, Hebrew and Spanish allowing them.5 This is of course expected on the idiosyncrasy approach, but what would still remain unexplained on the idiosyncrasy approach is why the observed cross-linguistic and cross-dialectal instability is confined to try to the exclusion of want and begin. In light of this tension, a major goal of this paper is to pursue a compromise whereby try’s behavior has a principled semantic basis in that non-control try-sentences are semantically anomalous, but this underlying regularity gets obscured by variation in the availability of a coercion mechanism that repairs the anomaly. In its general shape, this proposal bears a close resemblance to some of the ideas expressed about control by Perlmutter (1968); Jackendoff & Culicover (2003); Grano (2012, 2015a).

English French Mandarin Greek Hebrew Spanish

want
try % * *
begin * * * * * *

Table 1

Cross-linguistic availability of overt embedded subjects (Grano 2015b).

Turning next to the semantic factor, the only two formally explicit treatments of the semantics of try-sentences that I am aware of are Sharvit (2003) and Grano (2011), both of whose insights will be weaved into the analysis pursued below. Sharvit (2003: 422) is explicitly neutral about whether try’s syntactic behavior “has a deeper semantic reason” or not; it is not something that follows from the semantics she proposes. Grano (2011) attempts an explicit connection between try’s syntactic behavior and its semantics, and ends up pursuing an interface account whereby (following Sharvit) try has an aspectual semantics, and (building on Cinque 1999, 2004) morphemes with aspectual semantics are realized as inflectional-layer functional heads whose syntax precludes overt embedded subjects. A goal of this paper is to try to make the connection between meaning and syntactic behavior even tighter, in a way that does not rely on mapping principles that relate aspectual morphemes to particular syntactic positions.

The organization of the rest of the paper is as follows. Section 2 lays out novel cross-linguistic data that form part of the empirical basis of the account to be developed, with a focus on a puzzling interaction between control and temporal orientation that obtains for try-complements in Spanish and Hebrew. Section 3 develops a semantics for try as used in control sentences. Section 4 turns to non-control try-sentences and argues that they have an additional causative component not present in their control counterparts. Section 5 argues that this causative component is the result of coercion. Section 6 discusses how the core analysis might be scaled up to account for the broader range of observed cross-linguistic variation in try’s behavior. Finally, Section 7 concludes.

2 Some novel data

2.1 The core puzzle

Wurmbrand (2014) argues that English infinitival complements come in three temporo-aspectual classes. Future infinitives are diagnosed based on their compatibility with time frame adverbials that are future-oriented with respect to the matrix reference time, as in (6a–b) (cf. Stowell 1982). Simultaneous propositional infinitives are diagnosed based on their incompatibility with episodically interpreted, aspectually zero-marked event descriptions, as in (7a–b). Finally, simultaneous nonpropositional infinitives are diagnosed based on their incompatibility with time frame adverbials that are future-oriented with respect to the matrix reference time, as in (8).6

(6) Future infinitives
  a. (Today,) John wanted [PRO to leave tomorrow].
  b. (Today,) John expected [Bill to leave tomorrow].
(7) Simultaneous propositional infinitives
  a. John claimed [PRO to {be leaving/have left/#leave} at 3].
  b. John believed [Bill to {be leaving/have left/#leave} at 3].
(8) Simultaneous nonpropositional infinitives
  a. John tried [PRO to leave (#tomorrow)].
  b. John managed [PRO to leave (#tomorrow)].
  c. John began [PRO to leave (#tomorrow)].

Turning to the cross-linguistic situation, we observe here that in Spanish and Hebrew, want and try both behave like their respective English counterparts in that want enables future orientation (9a)/(10a) whereas try does not (9b)/(10b).

    1. (9)
    1. Spanish
    1.  
    1. a.
    1. Hoy
    2. today
    1. Juan
    2. Juan
    1. quería
    2. wanted.3SG
    1. [PRO
    2.  
    1. abrir
    2. open.INF
    1. la
    2. the
    1. puerta
    2. door
    1. (mañana)].
    2. tomorrow
    1. ‘Today, Juan wanted to open the door (tomorrow).’
    1.  
    1. b.
    1. Hoy
    2. today
    1. Juan
    2. Juan
    1. ha
    2. have.3SG
    1. intentado
    2. tried
    1. [PRO
    2.  
    1. abrir
    2. open.INF
    1. la
    2. the
    1. puerta
    2. door
    1. (#mañana)].
    2. tomorrow
    1. ‘Today, Juan tried to open the door (tomorrow).’
    1. (10)
    1. Hebrew
    1.  
    1. a.
    1. Ha-yom
    2. the-day
    1. Dani
    2. Dani
    1. raca
    2. wanted.3MS
    1. [PRO
    2.  
    1. liftoax
    2. open.INF
    1. et
    2. ACC
    1. ha-delet
    2. the-door
    1. (maxar)].
    2. tomorrow
    1. ‘Today, Dani wanted to open the door (tomorrow).’
    1.  
    1. b.
    1. Ha-yom
    2. the-day
    1. Dani
    2. Dani
    1. nisa
    2. tried.3MS
    1. [PRO
    2.  
    1. liftoax
    2. open.INF
    1. et
    2. ACC
    1. ha-delet
    2. the-door
    1. (#maxar)].
    2. tomorrow
    1. ‘Today, Dani tried to open the door (tomorrow).’

Unlike in (standard) English, however, Spanish and Hebrew try also have the option of taking a non-control complement, as in (11)–(12). (11), for example, could be used to report a scenario in which Juan did something with the intention that his action would result in Pedro opening the door (e.g., he might have simply asked him to do so, or tried reasoning with him, or employed subtle psychological manipulation).

    1. (11)
    1. Spanish
    1. Juan
    2. Juan
    1. ha
    2. have.3SG
    1. intentado
    2. tried
    1. [que
    2. that
    1. Pedro
    2. Pedro
    1. abriese
    2. opened.SBJV.3SG
    1. la
    2. the
    1. puerta].
    2. door
    1. Lit.: ‘Juan tried for Pedro to open the door.’
    1. (12)
    1. Hebrew
    1. Dani
    2. Dani
    1. nisa
    2. tried.3MS
    1. [she-Moshe
    2. that-Moshe
    1. yiftax
    2. open.3MS.FUT
    1. et
    2. ACC
    1. ha-delet].
    2. the-door
    1. Lit.: ‘Dani tried for Moshe to open the door.’

Against that backdrop, the novel and crucial observation is that when try is used with a non-control complement, future orientation now becomes possible in both Spanish and (for at least some speakers) Hebrew, as witnessed in (13) and (14) respectively. (13), for example, could be uttered in a scenario where Juan did something today (making a request, or whatever) with the intention that this action would result in Pedro opening the door tomorrow.7

    1. (13)
    1. Spanish
    1. Hoy
    2. today
    1. Juan
    2. Juan
    1. ha
    2. have.3SG
    1. intentado
    2. tried
    1. [que
    2. that
    1. Pedro
    2. Pedro
    1. abriese
    2. opened.SBJV.3SG
    1. la
    2. the
    1. puerta
    2. door
    1. mañana].
    2. tomorrow
    1. Lit.: ‘Today, Juan tried for Pedro to open the door tomorrow.’
    1. (14)
    1. Hebrew
    1. %Ha-yom
    2.   the-day
    1. Dani
    2. Dani
    1. nisa
    2. tried.3MS
    1. [she-Moshe
    2. that-Moshe
    1. yiftax
    2. open.3MS.FUT
    1. et
    2. ACC
    1. ha-delet
    2. the-door
    1. maxar].
    2. tomorrow
    1. Lit.: ‘Today, Dani tried for Moshe to open the door tomorrow.’

The nature of the puzzle should be clear: if we hypothesize that future orientation is unacceptable in control try-sentences because control infinitives cannot be future-oriented, then we make the wrong predictions about control want-sentences, which do enable future orientation. But if we hypothesize that future orientation is unacceptable in control try-sentences because try disallows future-oriented complements, then we make the wrong predictions about non-control try-sentences, which also do enable future orientation. Rather, it must be something about the interaction of try with control that conspires to preclude future orientation and similarly something about the interaction of try with the absence of control that conspires to enable future orientation.

In this light, the narrow goal of this paper is to propose an account of the pattern of data in (9)–(14). In a nutshell, the proposed account has three ingredients:

(15) a. Control try can only combine with simultaneous infinitives because try has an aspectual semantics that precludes future orientation (cf. Sharvit 2003; Grano 2011) (Section 3).
  b. Non-control try involves an additional causative layer of meaning that renders it compatible with future orientation (Section 4).
  c. The source of the additional causative layer for non-control try is coercion (cf. Perlmutter 1968; Jackendoff & Culicover 2003; Grano 2012, 2015a) (Section 5).

In what follows, I unpack each of these components in turn. First, though, I document two other cross-linguistically attested patterns of data regarding try-sentences that the proposed analysis will ultimately have to be responsible for as well and to which we will ultimately return in Section 6.

2.2 Two other patterns

Although the data from the previous subsection constitute the core explanandum of this paper, evidently not all languages can be categorized as English-like or as Spanish/Hebrew-like in how try behaves, and so it will be important to ensure that the proposed account leaves room for the other attested patterns as well. In particular, I am aware of two other patterns.

First, as seen in (16), Greek behaves like Spanish and Hebrew in that want enables future oriented complements (16a) and try enables (future-oriented) non-control complements (16c). Unlike Spanish and Hebrew (and for that matter English), though, Greek also enables future orientation with control try, as witnessed in (16b) which is taken from Roussou (2009: 1826).8

    1. (16)
    1. Greek
    1.  
    1. a.
    1. O
    2. the
    1. Kostas
    2. Kostas
    1. ithele
    2. wanted.3SG
    1. simera
    2. today
    1. [na
    2. SBJV
    1. fiji
    2. leave.3SG
    1. PRO
    2.  
    1. avrio].
    2. tomorrow
    1. ‘Today, Kostas wanted to leave tomorrow.’
    1.  
    1. b.
    1. O
    2. the
    1. Kostas
    2. Kostas
    1. prospathise
    2. tried.3SG
    1. simera
    2. today
    1. [na
    2. SBJV
    1. fiji
    2. leave.3SG
    1. PRO
    2.  
    1. avrio].
    2. tomorrow
    1. Lit.: ‘Today, Kostas tried to leave tomorrow.’
    1.  
    1. c.
    1. O
    2. the
    1. Kostas
    2. Kostas
    1. prospathise
    2. tried.3SG
    1. simera
    2. today
    1. [na
    2. SBJV
    1. fiji
    2. leave
    1. i
    2. the
    1. Maria
    2. Maria
    1. avrio].
    2. tomorrow
    1. Lit.: ‘Today, Kostas tried for Maria to leave tomorrow.’

Second, preliminary evidence suggests that (Brazilian) Portuguese may instantiate yet another pattern. According to my consultant, Portuguese want unsurprisingly enables future orientation, as in (17a). Portuguese try accepts control complements and (at least marginally) non-control complements, but does not enable future orientation with either, as seen in (17b–c) (cf. also Note 7 above on a potential point of variation among Hebrew speakers that would make one variety of Hebrew look just like Portuguese with respect to this pattern of data).9

    1. (17)
    1. Portuguese
    1.  
    1. a.
    1. Hoje
    2. today
    1. o
    2. the
    1. João
    2. Joao
    1. queria
    2. wanted.3SG
    1. [PRO
    2.  
    1. abrir
    2. open.INF
    1. a
    2. the
    1. porta
    2. door
    1. (amanhã)].
    2. tomorrow
    1. ‘Today, Joao wanted to open the door (tomorrow).’
    1.  
    1. b.
    1. Hoje
    2. today
    1. o
    2. the
    1. João
    2. Joao
    1. tentou
    2. tried.3SG
    1. [PRO
    2.  
    1. abrir
    2. open.INF
    1. o
    2. the
    1. porta
    2. door
    1. (#amanhã)].
    2. tomorrow
    1. ‘Today, Joao tried to open the door (tomorrow).’
    1.  
    1. c.
    1. ?Hoje
    2.   today
    1. o
    2. the
    1. João
    2. Joao
    1. tentou
    2. tried.3SG
    1. [que
    2. that
    1. o
    2. the
    1. Pedro
    2. Pedro
    1. abrisse
    2. opened.SBJV.3SG
    1. a
    2. the
    1. porta
    2. door
    1. (#amanhã)].
    2. tomorrow
    1. Lit.: ‘Today, Joao tried for Pedro to open the door (tomorrow).’

According to my informant, to express the intended meaning of (17c), it is necessary to insert an overt causative predicate, as in (18).

    1. (18)
    1. Portuguese
    1. Hoje
    2. today
    1. o
    2. the
    1. João
    2. Joao
    1. tentou
    2. tried.3SG
    1. [PRO
    2.  
    1. fazer
    2. make
    1. o
    2. the
    1. Pedro
    2. Pedro
    1. abrir
    2. open.INF
    1. a
    2. the
    1. porta
    2. door
    1. amanhã].
    2. tomorrow
    1. ‘Today, Joao tried to make Pedro open the door tomorrow.’

Where does this leave us? Taking a step back, if the behavior of try with respect to the availability of control vs. non-control complements and the availability of simultaneous vs. future-oriented complements were independent of one another, then we would expect that in a given language, try with a control complement could be either simultaneous, future-oriented, or ungrammatical; and independently of that, try with a non-control complement could similarly be either simultaneous, future-oriented, or ungrammatical. Ignoring the hypothetical case where try is ungrammatical with both control and non-control complements alike, this yields a total of eight possible language types as schematized in Table 2.10 Of these eight types, only four are attested: first we have languages like Brazilian Portuguese and Greek in which complements to try always have the same temporal profile regardless of whether they involve control or not (types a and e in the table). Then we have Spanish and Hebrew in which control complements are simultaneous but non-control complements are future-oriented (type b). Finally, we have languages in which control complements are simultaneous and non-control complements are ungrammatical. Such languages include (standard) English — as well as French (Rochette 1988), German (Wurmbrand 2001), and Mandarin (Grano 2015a) (type c), based on the reports of this verb’s behavior in the cited references. The other four logically possible scenarios (types d, f, g, and h) are as far as I know unattested.

CONTROL NON-CONTROL LANGUAGE

a. S S Brazilian Portuguese
b. S F Spanish, Hebrew
c. S U (standard) English, French, Mandarin, German
d. F S (unattested)
e. F F Greek
f. F U (unattested)
g. U S (unattested)
h. U F (unattested)

Table 2

The cross-linguistic behavior of try (key: S = simultaneous; F = future-oriented; U = ungrammatical).

After arriving at an account of types b and c languages in Sections 3–5 below, I will return in Section 6 below to a discussion of how the core account might accommodate type a and e languages as well without erroneously letting in the unattested types.

3 A semantics for try as used in control sentences

Any adequate semantics for try as used in sentences like (19) is responsible for at least two facts.

(19) John tried to open the door.

First, as emphasized by Sharvit (2003), in order for (19) to be true, it must be the case that John did something; i.e., there must be some event whose agent is John. That this is so is evidenced by the observation that it sounds contradictory to pair a try-sentence with a denial that the relevant individual took action, as in (20); cf. the corresponding want- and intend-sentences in (21) which are purely attitudinal and hence not similarly contradictory. We’ll call this the action component of try.

(20) #John tried to open the door, but he didn’t do anything about it.
(21)   John wanted/intended to open the door, but he didn’t do anything about it.

Second, as also pointed out by Sharvit (2003) and further discussed by Grano (2011), (19) also conveys that John had a particular kind of mental attitude — namely an intention — to open the door. This is evidenced by the contradictory status of (22); cf. the corresponding want-sentence in (22) which is not similarly contradictory. We’ll call this the intention component of try.11

(22) #John tried to open the door, but he had no intention of opening the door.
(23)   John wanted to open the door, but he had no intention of opening the door.

Putting together the action component with the intention component leads us to (24) as a first approximation of the semantics for the try-sentence in (19). I assume here a Hintikkan semantics for intention reports whereby INT is a function from individuals x and worlds w to the set of worlds compatible with x’s intentions in w. This function is borrowed from Stephenson (2010: 423–424); for discussion and alternative options, see Grano (2017). Also, because it is orthogonal, we abstract away here from the need to account for the obligatory de se interpretation of attitude reports expressed by control sentences; for options broadly consistent with the other theoretical choices made in this paper, see Stephenson (2010); Pearson (2016).

(24) Proposed denotation for John tried to open the door (version 1 of 2):
  e[Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[open(e′) ∧ Ag(e′, j) ∧ Th(e′, d) in w′]]
  ‘There is some event e whose agent is John and all worlds compatible with John’s intentions in w are worlds in which there is some event e′ which is an opening event whose agent is John and whose theme is the door.’

As it stands, though, (24) is inadequate, because although it correctly conveys that John acted and that John intended to open the door, it does not specify that John acted with the intention of opening the door; i.e., his intention was that the action in question result in the fulfillment of the intention (cf. Searle’s 1983 concept “intention in action”). Hence we incorrectly predict that (19) would be true in a scenario where John’s action had nothing to do with the named intention.

In order to rectify this shortcoming, we need the truth conditions to specify some sort of link between the action John took and the outcome he intended. There are various ways one could go with this; here we will basically adopt the approach taken by Sharvit (2003), whose insight was that the relationship that a try-sentence establishes between action and intended outcome is very reminiscent of Dowty’s (1977) imperfective paradox. In particular, a progressive sentence like (25a) asserts the existence of some event that stands in a part-whole relationship with the event asserted to exist by the sentence’s non-progressive counterpart in (25b).

(25) a. John was opening the door.
  b. John opened the door.

Here we’ll implement this insight somewhat differently from how Sharvit does by simply adding to the content of John’s intention the clause that the action John took is an initial stage of the outcome he intended, as in (26) (with bolding to show what’s been added). In other words, in order for John’s intention to be satisfied, it has to be the case that the outcome he intended (successfully opening the door) had as an initial stage the action he took. The “initial-stage-of” relation used here (⊂init) is borrowed from Piñango and Deo’s (2016) work on aspectual verbs.

(26) Proposed denotation for John tried to open the door (version 2 of 2):
  e[Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[einite′ ∧ open(e′) ∧ Ag(e′, j) ∧ Th(e′, d) in w′]]
  ‘There is some event e whose agent is John and all worlds compatible with John’s intentions in w are worlds in which e is an initial stage of some event e′ which is an opening event whose agent is John and whose theme is the door.’

Treating try-sentences as naming the initial stage in an action raises a question about how try differs from begin when begin is used as a control predicate. Here I follow Grano (2011) who proposes, building on relevant work in action theory (see e.g., O’Shaughnessy 1973), that volitional events have as their initial stage a mental action that precedes and causes externally observable actions. Whereas try merely requires that the event is underway at least to the mental action stage, begin requires that the action has been successfully externalized. In other words, we need to read “⊂init” in (26) as relating an event to its mental action stage, and would need to define another relation for begin that relates an event to some portion of its externalized onset. Following Grano (2011), support for this difference comes from the observation that if an agent is (perhaps unknowingly) paralyzed, that agent can try to raise his arm (i.e., execute a mental action intended to cause an arm raising) but cannot begin to raise his arm, as seen in (27).12

(27) CONTEXT: John is paralyzed from the neck down.
  a.   John tried to raise his arm.
  b. #John began to raise his arm.

Armed with this semantics, we are now in a position to say something about the unacceptability of try with future-oriented infinitives. By way of background, consider the anomalous sentence in (28).

(28) #Today John opened the door tomorrow.

We can make sense of this anomaly by assigning (28) a denotation like (29), so that it comes out contradictory: the event’s runtime is asserted to be included both in the day of the utterance time (i.e., “today”) and in the day after the utterance time (i.e., “tomorrow”). Because “today” and “tomorrow” do not overlap temporally, this is a logical contradiction. (In the formula, t* denotes the utterance time, DAY-OF is a function from time intervals to the relevant 24-hour period that contains them, and DAY-AFTER is a function from time intervals to the relevant 24-hour period that follows them).

(29) e[open (e) ∧ Ag(e, j) ∧ Th(e, d) ∧ τ(e) ⊆ DAY-OF(t*) ∧ τ(e) ⊆ DAY-AFTER(t*)]
  ‘There is some event e such that e is an opening event whose agent is John and whose theme is the door and the runtime of e is included in the day of the utterance time and the runtime of e is included in the day after the utterance time.’

Returning to try, the result now is that a future-oriented infinitive like in (30) will under the current proposal yield the semantics in (31).

(30) #Today John tried to open the door tomorrow.
(31) e[τ(e) ⊆ DAY-OF(t*) ∧ Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[e ⊂inite′ ∧ open(e′) ∧ Ag(e′, j) ∧ Th(e′, d) ∧ τ(e′) ⊆ DAY-AFTER(t*) in w′]]
  ‘There is some event e whose runtime is included in the day of the utterance time and whose agent is John and all worlds compatible with John’s intentions in w are worlds in which e is an initial stage of an opening event whose agent is John and whose theme is the door and whose runtime is included in the day after the utterance time.’

The trying event e is asserted to be contained in “today” and the intended outcome e′ is asserted to be contained (in John’s intention alternatives) in “tomorrow”. But since e stands in a part-whole relationship with e′, it is impossible for the former to be contained in “today” and the latter in “tomorrow”. This impossibility can hence be understood as the source of the perceived unacceptability of the sentence. This is a good result: the temporal orientation of try does not have to be stipulated but rather is an emergent consequence of try’s aspectual semantics (cf. Wurmbrand 2014: Note 29 for a similar point).

4 A causative component for non-control try

Consider now the semantics of try-sentences that involve non-control complements (in those varieties of English and those languages that permit such structures). The null hypothesis would be that a sentence like (32) has a semantics just like its control counterpart except that the thematic relation borne by the subject position of the lower clause is not bound by the matrix subject but rather has its own referentially independent subject, in this case Bill, as in (33).

(32) John tried [for Bill to open the door].
(33) Proposed denotation for John tried for Bill to open the door (version 1 of 2):
  e[Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[einite′open(e′) ∧ Ag(e′, b) ∧ Th(e′, d) in w′]]
  ‘There is some event e whose agent is John and all worlds compatible with John’s intentions in w are worlds in which e is an initial stage of an event of Bill opening the door.’

But if this is right, then we erroneously predict that future-oriented infinitives should be unacceptable with non-control try-sentences: (33) has all the same temporo-aspectual properties as its control counterpart and hence future-oriented infinitives should give rise to the same anomalous outcome as discussed in the previous section. Yet as we saw in Section 2, non-control try-sentences enable future-oriented interpretations in Spanish and Hebrew.

Suppose then that non-control try-sentences have a bit more semantic structure than their control counterparts; in particular, suppose that they assert not that the agent of the trying intended for the trying to be an initial stage in the intended outcome but rather that the agent of the trying intended for the trying to be the initial stage in an event that stands in a causal relationship with the intended outcome, as in (34).

(34) Proposed denotation for John tried for Bill to open the door (version 2 of 2):
  e[Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[einite′ ∧ ∃e″[CAUSE(e′, e″) ∧ open(e″) ∧ Ag(e″, b) ∧ Th(e″, d) in w′]]]
  ‘There is some event e whose agent is John and all worlds compatible with John’s intentions in w are worlds in which e is an initial stage of an event e′ such that e′ causes an event of Bill opening the door.’

The representation in (34) carries out the intuition that in a non-control try-sentence, the subject of try is understood as playing a causal role with respect to the eventuality associated with the complement of try. And if this is right, then (35) has a denotation like (36).

(35) Today John tried [for Bill to open the door tomorrow].
(36) e[τ(e) ⊆ DAY-OF(t*) ∧ Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[einite′ ∧ ∃e″[CAUSE(e′, e″) ∧ open(e″) ∧ Ag(e″, b) ∧ Th(e″, d) ∧ τ(e′) ⊆ DAY-AFTER (t*)] in w′]]
  ‘There is some event e whose runtime is included in the day of the utterance time and whose agent is John and all worlds compatible with John’s intentions in w are worlds in which e is an initial stage of an event e′ such that e′ causes an event of Bill opening the door whose runtime is included in the day after the utterance time.’

Is there any semantic anomaly or contradiction in (36)? It depends on what kinds of constraints there are on the temporal relationship borne by events that stand in the hypothesized CAUSE-relation. As long as these events need not overlap temporally — that is, as long as the causing event is allowed to precede and not overlap with the caused event — then we predict that (35) is acceptable (in those languages like Spanish and Hebrew that allow non-control try to begin with).

As it turns out, causative expressions vary as to whether they allow for temporal distance between the causing event and the caused event. For example, as pointed out by Bjorkman & Cowper (2013), causative make admits future-oriented complements, as in (37a), whereas causative have does not, as in (37b).

(37) Bjorkman & Cowper (2013: 2)
  a.   They made the team throw the game on Monday by threatening them on Sunday night.
  b. #They had the team throw the game on Monday by threatening them on Sunday night.

On some theories, this distinction correlates with the distinction between direct and indirect causation; see e.g., Martin & Schäfer (2014) for discussion. For our purposes, it suffices simply to note the existence of causative expressions like make, which serve as an existence proof that our hypothesized CAUSE predicate is in principle compatible with future orientation.

5 Causation by coercion

We have solved the puzzle by deus ex machina: in non-control try-sentences, a causative predicate disrupts the inclusion relation between the trying event and the intended outcome, enabling future orientation. But can this causative predicate be independently motivated?

The strategy we’ll take here is to independently motivate the causative predicate as a coercion mechanism used to satisfy a selectional restriction try has to the effect that its complement must have an unsaturated agentive thematic relation. To start with, consider the fact that try sounds odd in combination with outcomes that are not ordinarily construable as being amenable to deliberate control (cf. Farkas 1988; Jackendoff & Culicover 2003; Grano 2017). (38), for example, sounds odd because it seems to presuppose that John potentially has control over — or at least believes himself to potentially have control over — the timing of his sexual development.

(38) #John tried to go through puberty.

The status of this meaning component as a presupposition is supported by the observation that questioning, negation, and conditionalization of the sentence preserves the perceived oddity, as seen in (39).

(39) a. #Did John try to go through puberty?
  b. #John did not try to go through puberty.
  c. #If John tries to go through puberty, I’ll be surprised.

This suggests that it is sensible to model this meaning component as a selectional restriction on try (though see Grano 2017 for a more extensive discussion of the nuances involved in adopting this approach). Working backwards from our semantics for try-sentences from Section 3, repeated in (40), the semantics that we end up with for try itself is as indicated in (41). The assumption here is that the complement to try denotes a function of type ⟨e, ⟨Ɛ, st⟩⟩, i.e., a function from individuals to functions from eventualities to propositions.

(40) Proposed denotation for John tried to open the door (version 2 of 2):
  e[Ag(e, j) ∧ ∀w′ ∈ INTj,w: ∃e′[einite′ ∧ open(e′) ∧ Ag(e′, j) ∧ Th(e′, d) in w′]]
(41) TRY(P)(x)(e)(w) = 1 iff…
  Ag(e, x) ∧ ∀w′ ∈ INTx,w: ∃e′[einite′ ∧ P(x)(e′)(w′)]

To encode the suggested selectional restriction, we can augment the denotation as in (42). With this revision in place, a try-sentence will have a defined denotation only if try’s first argument P is such that any individual and any eventuality it applies to results in the entailment that the relevant individual stands in the agent relation with the relevant eventuality (for any arbitrary evaluation world). In those cases where this selectional restriction is met, the resulting denotation is the same as what we had before.

(42) TRY(P)(x)(e)(w) is defined only if…
  ye′∀w′[P(y)(e′)(w′) → Ag(e′, y) in w′]
  where defined, TRY(P)(x)(e)(w) = 1 iff…
  Ag(e, x) ∧ ∀w′ ∈ INTx,w: ∃e′[e <inite′ ∧ P(x)(e′)(w′)]

So if P is instantiated by go through puberty, the resulting sentence has an undefined semantics unless go through puberty can be construed as entailing an agent relation for its individual argument.

Now consider again a non-control sentence like (43) as considered in a dialect of English in which it is grammatical.

(43) John tried for Bill to leave.

The first question that needs to be dealt with has to do with semantic type of for-to clauses. In a great many environments, control clauses are syntactically interchangeable with for-to clauses. These environments include subject position (44), infinitival relatives (45), purpose clauses (46), and complements to desiderative predicates (47).

(44) [{PRO/For John} to go to the store] would be a great idea.
(45) This would be a great book [{PRO/for us} to read].
(46) I opened a window [{PRO/for them} to get some fresh air].
(47) a. I am dying [{PRO/for John} to get this job].
  b. I long [{PRO/for John} to get this job].
  c. I was hoping [{PRO/for John} to come over].
  d. I want very much [{PRO/for John} to help those in need].
  e. I was so happy [{PRO/for John} to get the job].

The distributional similarity between control clauses and for-to clauses renders plausible the hypothesis that they are also type-theoretically similar, and so I will pursue the idea — contra Dowty (1985) but in line with more recent work in the literature on the semantics of control such as Stephenson (2010); Pearson (2013) — that control and non-control clauses are type-theoretically identical. More specifically, I will assume following Pearson (2013) that they are both property-denoting. See Pearson (2013) for an extensive defense of the view that clauses denote properties. What this will mean for a for-to clause like (48) is that it instantiates vacuous lambda binding: the clause denotes a function whose input is an individual argument x, but that argument does not figure anywhere in the output of the function.

(48) ⟦for Bill to leave⟧ = λxλeλw.leave(e) ∧ Ag(e, b) in w

Putting together the semantics for try in (42) with the semantics for for-to infinitives in (48), the important observation now is that although there will be no type-theoretic difficulty in combining try with a for-to clause, there will be a presupposition failure: it is not the case that any arbitrary individual and eventuality that the function in (48) applies to will be such that they stand in the agent relation; this will hold only in the special case where that individual is Bill. This gives us an account for why try resists non-control complements in many languages including standard English, Mandarin, and French: such sentences suffer a presupposition failure.

As for those languages and those dialects of English where try accepts non-control complements, the suggestion is that in these languages, a coercion mechanism is available to repair the selectional restriction violation (cf. Perlmutter 1968; Jackendoff & Culicover 2003; Grano 2012, 2015a who propose essentially this, but without formalizing it in the way done here).13 One initially plausible way in which this coercion mechanism could work would be as defined in (49). This coercion operator combines with an arbitrary P and adds an agent relation linked to the unsaturated individual argument, which would yield the result in (50) for the case at hand.

(49) ⟦OpC⟧ = λPe, ⟨Ɛ, st⟩⟩λxλeλw.P(x)(e)(w) ∧ Ag(e, x)
(50) ⟦OpC⟧(λxλeλw.leave(e) ∧ Ag(e, b)) =
  λxλeλw.leave(e) ∧ Ag(e, b) ∧ Ag(e, x)

But such an operator is not of very good general use because it will frequently run afoul of thematic uniqueness (see e.g., Parsons 1990), i.e., the requirement that for any given thematic relation and any given eventuality, there can be at most one individual that stands in that thematic relation with that eventuality. The solution then is for the coercion operator to introduce its own eventuality that stands in a causation relation with the eventuality introduced by P and that can have its own agent relation with no risk of violating thematic uniqueness. This gives us (51), which will yield the result in (52) for the case at hand.

(51) ⟦Op′C⟧ = λPe, ⟨Ɛ, st⟩⟩λxλeλw.∃eCAUSE(e, e′) ∧ Ag(e, x) ∧ P(x)(e′)(w)
(52) ⟦Op′C⟧(λxλeλw.leave(e) ∧ Ag(e, b)) =
  λxλeλw.CAUSE(e, e′) ∧ Ag(e, x) ∧ leave(e′) ∧ Ag(e′, b)

The causative predicate is thereby motivated for non-control try-sentences, and the puzzle we started with is thereby solved.

Before moving on, one question that should be addressed is whether positing coercion leads to an overgeneration problem when we look at other verbs besides try. While a full cross-linguistic investigation of how different verb types behave with respect to control is well beyond the scope of this paper, one cross-linguistically stable trend identified by Grano (2015b) is that aspectual verbs robustly resist non-control complements even in languages like Greek, Hebrew and Spanish that allow non-control complements to try (see Table 1 in Section 1 above). Similarly, I am not aware of and would not expect to find any variety of English that accepts sentences like (53). The question then is why, on the coercion account, languages that allow coercion for try do not similarly allow for coercion with ‘begin’ so that a sentence like (53) would be acceptable and receive an interpretation like (54).

(53) *John began [for Bill to open the door].
(54)   John began [PRO to get Bill to open the door].

In response to this I will simply point to Grano (2016) for a relevant proposal. In a nutshell, what Grano proposes is that what goes wrong in (53) is not the lack of control per se but rather the presence of complementizer for, which is syntactically required to license the overt subject. Grano proposes that for introduces a modal semantics (roughly, priority modality in the sense of Portner 2007), and this modality conflicts with the semantics of begin which is arguably modal but not priority-oriented (i.e., not bouletic, teological, or deontic). By contrast, predicates like want and try do have a priority-oriented modal semantics and so are in principle semantically compatible with complementizer for. Scaling this up to the cross-linguistic picture, Grano speculates that the subjunctive and other morphology needed to syntactically license an overt subject in other languages has a similar modal semantics that conflicts with the semantics of aspectual verbs. If this is on the right track, then we have an account of why (53) should be ruled out even in languages that have coercion: even with coercion in place, there is a semantic conflict between the meaning of the embedding verb and the semantics of the morphosyntactic machinery needed to introduce the subject.

6 The two other patterns revisited

In the preceding three sections, I arrived at an account of the behavior of try in (standard) English-like languages vs. Spanish/Hebrew-like languages. If these two language types exhausted the cross-linguistic behavior of try, then the account would lend itself to the overall picture in (55).

(55) The cross-linguistic grammar of try (version 1 of 2)
  a. Universally, try has an aspectual semantics that precludes future orientation.
  b. Universally, try has a selectional restriction that precludes non-control complements.
  c. In some languages but not others, a coercion mechanism repairs violations to the selectional restriction and thereby enables non-control complements.
  d. In all those languages that have the relevant coercion mechanism, this mechanism enables future orientation.

Evidently, though, this account is too strong, because as we saw in Section 2.2, not all languages fall into the two patterns that have been the focus of this paper: instead, Greek and Brazilian Portuguese each exemplify their own distinct pattern. In other words, in the terms of Table 2 of Section 2.2, (55) predicts that we should only be able to find type b and type c languages to the exclusion of the six other hypothetical types. How can we loosen (55) so as to let in type a languages (Brazilian Portuguese) and type e languages (Greek) without loosening things up so much that we erroneously let in the four unattested types (types d, f, g, and h)?

Returning first to Portuguese, what we saw is that try disallows future orientation regardless of whether it takes a control complement or a non-control complement. This state of affairs is actually relatively easily to accommodate within the overall account: we already observed at the end of Section 4 that causative predicates vary as to whether they enable future orientation; within English, for example, causative make allows it but causative have does not. Hence we can make sense of the Portuguese facts by appealing to the hypothesis that among those languages that employ coercion to enable non-control complements to try, some have a causative semantics that enables future orientation (a make-like causative) and others have a cuasative semantics that do not enable future orientation (a have-like causative). In other words, we can revise the overall account to that in (56), with italics to show what’s been changed.

(56) The cross-linguistic grammar of try (version 2 of 2)
  a. Universally, try has an aspectual semantics that precludes future orientation.
  b. Universally, try has a selectional restriction that precludes non-control complements.
  c. In some languages but not others, a coercion mechanism repairs violations to the selectional restriction and thereby enables non-control complements.
  d. In some of those languages that have the relevant coercion mechanism, this mechanism enables future orientation.

Returning now to Greek, what we saw in this language is that try allows for future orientation with both control and non-control complements alike. Accommodating this pattern is not as straightfoward: if try universally names an initial stage of the action denoted by its complement, then a ban on future orientation would seem to be an inescapable result for the same reason that we would not expect to find a language in which the translation equivalent of Yesterday John left tomorrow could be a felicitous utterance. But two hypotheses come to mind. First, it could be that the Greek word that we translate as try (i.e., prospatho) does not have the kind of aspectual semantics pursued by Sharvit (2003) and Grano (2011) for English try. In particular, it could have some other kind of semantics that renders it consistent with future orientation. Consider for example the verbs plan and decide in English. Like try, the verbs plan and decide have an action component and an intention component, but they allow for a gap of time between the planning/deciding and the intended outcome, as witnessed by examples like (57).

(57) a. (Today,) John planned [PRO to open the door tomorrow].
  b. (Today,) John decided [PRO to open the door tomorrow].

It is conceivable that Greek prospatho behaves likewise or is perhaps polysemous between a ‘try’ reading and a ‘plan’/‘decide’ reading.

A second possibility is that Greek try is like its cross-linguistic counterparts but that there is something about the grammar of Greek that enables a causative component even in sentences where it is not forced by the presence of a non-control complement. It is tempting to relate this hypothesis to the more general fact about Greek that it altogether lacks nonfinite complementation: verbs are always morphologically marked for subject agreement and for tense and aspect, possibly signaling a richer structure of the sort that could host a silent morpheme that enables future orientation.14 But finite complementation per se cannot be a sufficient condition for future orientation because even within Greek, future orientation is banned in complements to aspectual verbs, as seen in (58).

    1. (58)
    1. Greek (Roussou 2009: 1826)
    1. O
    2. the
    1. Kostas
    2. Kostas
    1. arxise
    2. began.3SG
    1. simera
    2. today
    1. [na
    2. SBJV
    1. odhiji
    2. drive.3SG
    1. PRO
    2.  
    1. (#avrio)].
    2. tomorrow
    1. ‘Kostas began today to drive (tomorrow).’

While further research will be needed to decide which (if either) of these hypotheses is correct for Greek try, either hypothesis is broadly compatible with this paper’s core proposals.

7 Conclusion

The central conclusion of this paper is that superficial cross-linguistic variation in try’s syntactic behavior conceals an underlying regularity: try semantically demands a control complement because only a control complement satisfies the presupposition that try’s complement bear an unsaturated agent relation, and apparent counterexamples are handled via a coercion mechanism that is independently detectable in that it enables future orientation in some languages. In its general shape this proposal is closely allied with ideas expressed by Perlmutter (1968); Jackendoff & Culicover (2003); Grano (2012, 2015a), though I have gone beyond these earlier works by recruiting novel cross-linguistic data and also by formalizing the proposal in a way that is informed by work on the formal semantics of try-sentences.

In closing, a note is in order on the broader implications of this study for the division of labor between syntax and semantics in regulating the distribution of control. At the outset of the paper I identified two purely syntactic dimensions to PRO’s distribution: PRO can only occur in subject position and PRO can only appear as an argument of a nonfinite clause. Are these the only purely syntactic dimensions to PRO’s distribution? Can all distributional differences that turn on the choice of the embedding verb be understood in terms of semantic principles? While this would be a welcome result, I am pessimistic about its prospects, given the (English-internal) contrast in (59)–(60) as well as the (cross-linguistic) contrast in (60)–(61). Why is it that claim allows control complements but believe does not? Although claim and believe are not semantically identical, they are similar enough in meaning that it is hard to see how a semantic account of (59)–(60) would be possible. Instead, it looks like lexical idiosyncrasy. Similarly, why is it that in Italian unlike English, ‘believe’ allows control complements? Again, it is hard to see how English believe and Italian credere could differ semantically in just such a way as to produce the contrast in (60)–(61). And so far as I know, (61) has no special interpretive properties that we could use to build an argument that Italian ‘believe’ affords some coercion mechanism that enables control complements and that is not available to its English counterpart.

(59)   John claims [PRO to be tall].
(60) *John believes [PRO to be tall].

    1. (61)
    1. Italian
    1. Gianni
    2. Gianni
    1. crede
    2. believe.3SG
    1. [di
    2. of
    1. PRO
    2.  
    1. essere
    2. be.INF
    1. alto].
    2. tall
    1. ‘Gianni believes himself to be tall.’

The suggested conclusion then is that the distribution of control as a function of embedding verb choice is sometimes a matter of semantics and sometimes a matter of syntax. A comprehensive picture requires attention to both of these components of grammar.