3.1 Is it control?

PRO is always a subject. If we are dealing with control, we predict that the possessor can only be interpreted as the IRC subject and never as any internal argument of the IRC (unless that internal argument is a derived subject). This prediction is borne out. Consider the following:

1. (17)

   1. (a) This is the book (for John) to read to Mary.

   2. (b) This is John's book to read to Mary.

   3. (c) *This is Mary's book (for John) to read to.

2. (18)

   1. (a) That is the general (for the Emperor) to give a slave to.

   2. (b) ^?That is the Emperor's general to give a slave to.

   3. (c) *That is the slave's general (for the Emperor) to give to.

3. (19)

   1. (a) This is the patient (for the new surgeon) to operate on.

   2. (b) This is the new surgeon's patient to operate on.

   3. (c) #This is the patient's new surgeon to operate on.

4. (20)

   1. (a) This is the man to fix the sink.

   2. (b) *This is the sink's man to fix.

The (a) examples are the baseline IRCs without prenominal possessors. (17b), (18b) and (19b) show that the prenominal possessor can easily be interpreted as the IRC subject. In contrast, (17c) shows that the prenominal possessor cannot be interpreted as the IRC indirect object. Similarly, (18c) shows that the prenominal possessor cannot be interpreted as the IRC direct object (see also (20b)), and (19c) is odd precisely because world knowledge tells us that surgeons operate on patients and not vice versa. It thus looks as if we are dealing with a control relation, i.e. the prenominal possessor is capable of controlling the IRC subject PRO.

A clear consequence is that this control interpretation should be impossible in subject IRCs, as in (20b), since the IRC subject position (standardly taken to be an A′-variable) is linked with the RC head, not the prenominal possessor attached to the RC head. The same reasoning applies to passivised IRCs, which involve derived subjects.

1. (21) This is John's book to be read.

Even if one considers the subject gap of a subject IRC to be PRO rather than an A′-variable (see Bhatt Reference Bhatt1999), PRO will be interpreted as the RC head book and not as the prenominal possessor John. The prenominal possessor cannot be interpreted as the implicit external argument of read either, consistent with a control analysis. Consequently, subject IRCs are excluded from the present article. Subsequent references to IRCs thus refer exclusively to non-subject IRCs unless otherwise stated.

So far, I have simply been assuming that there is a PRO subject in IRCs. That PRO is present in IRCs is plausible for wh-IRCs, i.e. IRCs introduced by an overt wh-relative pronoun (with obligatory preposition pied-piping), since the presence of the wh-relative pronoun is standardly taken to indicate the presence of at least some portion of the C-domain (see Douglas Reference Douglas2016 for further discussion) and hence the presence of a subject in SpecTP. It is not immediately obvious whether Ø-IRCs, i.e. IRCs introduced by neither a wh-relative pronoun nor the complementiser for, should have a PRO subject. Douglas (Reference Douglas2016: chapter 3) proposes that Ø-IRCs may lack a C-domain altogether, but does not make any claims about whether Ø-IRCs have a full T-domain or not. However, the relation between the prenominal possessor and IRC subject does not seem to be affected by the presence or absence of a wh-relative pronoun. I take this to indicate that both wh-IRCs and Ø-IRCs have a PRO subject. Note that in standard English, PRO and subject traces are not permitted with the complementiser for. For-IRCs, i.e. IRCs introduced by the complementiser for, are thus incompatible with the phenomenon at issue.Footnote ³

Further evidence for the presence of PRO in both wh-IRCs and Ø-IRCs comes from the fact that anaphors are permitted in IRCs.

1. (22)

   1. (a) These are the sweets on which PRO_i to gorge yourselves_i.

   2. (b) These are the sweets PRO_i to share with each other_i.

Successful binding of anaphors in IRCs suggests that the IRC subject is a PRO, rather than an implicit argument, since implicit arguments cannot bind (see Wurmbrand Reference Wurmbrand2001).

To summarise, focusing exclusively on non-subject IRCs, the data show that a prenominal possessor can be interpreted as the IRC subject but not as any other IRC-internal argument. I thus conclude that the relation between the prenominal possessor and the IRC subject is one of control.

3.2 OC or not?

Since Williams (Reference Williams1980), a distinction has been made between Obligatory Control (henceforth, OC) and Non-Obligatory Control (henceforth, NOC) (see Landau Reference Landau2000, Reference Landau2003) for a thorough overview of approaches to control and for extensive discussion of various OC/NOC diagnostics that have been proposed).

For concreteness and because they are widely recognised and adopted, I will use the OC/NOC diagnostics given in Landau (Reference Landau2000: 31):

1. (23) The OC/NOC Categories

   1. (a) Arbitrary Control is impossible in OC, possible in NOC.

   2. (b) Long-distance Control is impossible in OC, possible in NOC.

   3. (c) Strict reading of PRO is impossible in OC, possible in NOC.

   4. (d) De re reading of PRO is impossible in OC (only de se), possible in NOC.

These diagnostics are illustrated below for canonical control cases. The examples in (24)–(27) are (slightly adapted) from Landau (Reference Landau2000: 34–6).

(24) shows that PRO can be interpreted as arbitrary in NOC but not in OC contexts, and (25) shows that long-distance control is possible in NOC but not in OC contexts.

1. (24)

   

1. (25)

   

(26) shows that, under ellipsis, the strict reading of PRO is possible in NOC but not in OC contexts (elided material is demarcated by angled brackets).

1. (26) Strict reading of PRO

   1. (a) John tried [PRO_John to leave early], and Bill did <try [PRO_Bill/*John to leave early]> too.(OC)

   2. (b) John thinks that [PRO_John feeding himself] will be difficult, and Bill does <think that [PRO_John/Bill feeding himself] will be difficult> too.(NOC)

Finally, (27) shows that the de re reading of PRO is possible in NOC but not in OC contexts (where only the de se interpretation is available).

1. (27)

   

(27a, b) are true in the context given, but (27c) is false. This shows that PRO in (27c), the OC example, must be interpreted de se (and cannot be interpreted de re), whilst in (27b), the NOC example, PRO can be interpreted de re.

I now turn to our IRC examples, applying Landau's diagnostics where applicable to see whether the control relation between the prenominal possessor and IRC subject is one of OC or NOC. As will be seen, the results are intriguingly conflicting.Footnote ⁴

Turning first to the de re/de se diagnostic, this is impossible to test with IRCs. The de re/de se diagnostic requires the presence of an attitude predicate between the controller and PRO. In our examples, the prenominal possessor is directly attached to the RC head which is directly modified by the IRC. It is thus impossible to construct examples with an attitude predicate between the prenominal possessor and the IRC subject PRO.

Now consider strict/sloppy interpretations. As shown in (28), the strict reading of PRO is impossible in IRCs, i.e. PRO can only be interpreted sloppily under ellipsis.

1. (28) This is John's book to read and that is Mary's.

   1. (a) This is John's book to read and that is Mary's <book PRO_Mary to read>.

   2. (b) *This is John's book to read and that is Mary's <book PRO_John to read>.

This holds even in a context where, for example, various people (including Mary) are choosing books for John to read. To express such an interpretation, an overt subject with for is required, as in (29) with the structure in (30). In such cases, the overt subject is in the antecedent of the ellipsis.

1. (29) This is John_i’s book for him_i to read and that is Mary's.

2. (30) This is John_i’s book for him_i to read and that is Mary's <book for him_i to read>.

The evidence from (28) thus strongly suggests that the relation between the prenominal possessor and the IRC PRO subject is OC.

Now consider long-distance control. As (31) shows, long-distance control is impossible in IRCs.

1. (31)

   1. (a) This is John's book PRO_John to read (to himself).

   2. (b) Mary said this is John's book PRO_John/*Mary to read (to himself/*herself).

   3. (c) Mary and Julie said this is John's book PRO_{John/*Mary and Julie} to read (to himself/*each other).

This holds even in a context where, for example, John has chosen a book for Mary (or Mary and Julie) to read. Again, to express such an interpretation, an overt subject with for must be used instead.

1. (32)

   1. (a) Mary_i said this is John's book for her_i to read (to herself_i).

   2. (b) [Mary and Julie]_i said this is John's book for them_i to read (to each other_i).

This diagnostic thus also suggests that we are dealing with OC. If the reference of PRO were free, the long-distance restriction would be unexpected.Footnote ⁵

The diagnostics so far suggest that prenominal possessor control into IRCs is an instance of OC, not NOC. Before moving on to the final diagnostic, let us consider a few predictions that are made if we are dealing with OC.

First, if OC is at stake, we would expect c-command to be relevant, i.e. if the prenominal possessor controls the IRC subject, we would expect that it must c-command the IRC subject.Footnote ⁶ This expectation is borne out. Compare the following:

1. (33) (a) This is John_i’s book PRO_John to read (to himself_i/*him_i).

   

2. (b) This is John's girlfriend_i’s book PRO_{John's girlfriend} to read (to herself_i/*her_i).

   

3. (c) *This is John_i’s girlfriend's book PRO_John to read (to himself_i).

   

The prenominal possessor DP (John in (33a) and John's girlfriend in (33b)) c-commands PRO and can thus control it, as made clear by the satisfaction of Condition A and the violation of Condition B. However, if the would-be controller is a prenominal possessor embedded inside another prenominal possessor, as John is in (33c), this DP fails to c-command PRO and so control cannot be established and the anaphor himself cannot be bound, thus violating Condition A.

Note that there is no obvious semantic reason why John cannot control PRO in (33c), as this is interpretation is perfectly possible with an overt infinitival subject co-referent with John.

1. (34) This is John_i’s girlfriend's book for John_i/him_i to read.

The reason thus seems to be purely syntactic.

The importance of c-command can also be seen using postnominal possessors as well, though the argument is somewhat more involved. Consider (35):

1. (35)

   1. (a) This is John's book to read.

   2. This is John's book PRO_John to read.

   3. (b) This is the book of John's to read.

   4. *This is the book of John's PRO_John to read.

PRO can be controlled by John if John is a prenominal possessor, as in (35a), but not if John is a postnominal possessor, as in (35b).

Independent evidence from Condition C effects confirms that prenominal possessors attached to the RC head c-command the IRC subject, whilst postnominal possessors do not. The contrast in (36) shows that a prenominal possessor c-commands the IRC subject. Binding conditions are satisfied in (36b), but Condition C is violated in (36a).

1. (36)

   1. (a) *This is his_i book for John_i to read.

   2. (b) This is John_i’s book for him_i to read.

Evidence involving postnominal possessors is a bit more complicated. First, note that a postnominal possessor version of (36a) does not violate Condition C.

1. (37) This is that book of his_i for John_i to read.

This is compatible with the idea that the postnominal possessor does not c-command the IRC subject. However, there is a potential confound here. RC heads can generally optionally reconstruct into the IRC in English (see Douglas Reference Douglas2016). Therefore, (37) may be grammatical because the RC head and everything embedded in it (including the postnominal possessor) reconstructs into the IRC to a position c-commanded by the IRC subject. This confound is not present with prenominal possessors, since prenominal possessors (like the external determiner) do not reconstruct into RCs generally in English, i.e. they are external to the constituent containing the RC head and IRC. This is confirmed by the ungrammaticality of (36a); if the prenominal possessor were able to reconstruct with the RC head to an IRC-internal position c-commanded by the IRC subject, a Condition C violation would be avoided and we would expect the example to be grammatical, contrary to fact. Returning to the issue of postnominal possessors, we thus need an example where the RC head is forced not to reconstruct. Consider the following:

1. (38) Context: John has written a book about Mary's adventure, but he is not happy with it and is reluctant to publish it. Mary, however, wants people to know about her adventure and must persuade John to publish it. One can say of the book:

2. This is the book of his_i about Mary_j’s adventure for her_j to persuade John_i to publish.

Here, the RC head is book of his about Mary's adventure. If it were to reconstruct into the IRC, we would get a Condition C violation since her would then c-command Mary. Assuming that lack of reconstruction of Mary implies lack of reconstruction of the entire RC head and all its modifiers,Footnote ⁷ this means that his is not interpreted in the IRC in (38). Now, his can be co-referential with John in (38), suggesting that postnominal possessors do not c-command into the IRC. In contrast, if his is a prenominal possessor attached to the RC head, it cannot be co-referential with John.

1. (39) *This is his_i book about Mary_j’s adventure for her_j to persuade John_i to publish.

Therefore, controlling for potential reconstruction effects, evidence from Condition C supports the claim that prenominal possessors attached to the RC head c-command the IRC and IRC subject, whilst postnominal possessors do not.

This contrast between prenominal and postnominal possessors falls out reasonably straightforwardly from most analyses of RCs. Consider the following schematic structures (assuming that the RC head does not reconstruct into the IRC):

1. (40) (a) John's book to read

   

2. (b) the book of John's to read

   

On reasonably standard assumptions, the external determiner attached to the RC head has no (I)RC-internal representation, supported by the fact that it never reconstructs (see, e.g., Kayne Reference Kayne1994; Bianchi Reference Bianchi1999; Aoun & Li Reference Aoun and Li2003; Salzmann Reference Salzmann2006, among many others). Furthermore, the external determiner c-commands the (I)RC (either because it takes the (I)RC as its complement or because it scopes over both the RC head and the (I)RC; the latter being illustrated here). Now, on the standard assumption that prenominal possessors (except those in compounds such as men's shoes or children's book) are in SpecDP (see Abney Reference Abney1987; Alexiadou, Haegeman & Stavrou Reference Alexiadou, Haegeman and Stavrou2007), prenominal possessors are higher than the external determiner and would thus c-command the (I)RC and everything in it, including the IRC subject, as in (40a). Postnominal possessors, however, are contained inside a PP attached to the RC head, as in (40b). They are thus unable to c-command out of the PP and out of the RC head, and hence cannot c-command into the IRC or control the IRC subject.

These facts also support the claim made above that the reference of the IRC subject PRO is not free. If it were, we might expect accidental co-reference to be possible between PRO and a possessor regardless of the possessor's structural position. These facts thus support the idea that the control relation between the prenominal possessor and the IRC subject is one of OC.

Assuming that partial control is a species of OC (Landau Reference Landau2000, Reference Landau2008, Reference Landau2013), we would also predict partial control readings to be possible in IRC contexts. As the following examples show, this prediction is borne out.

1. (41) Context: There are tours around the set of the Fifty Shades of Grey film. The tour guide is pointing out the various rooms and what Christian Grey uses each room for (censored version!).

   1. (a) This is his room to meet in / This is his room in which to meet.

   2. (b) This is his room to kiss in / This is his room in which to kiss.

   3. (c) This is his room to hug in / This is his room in which to hug.

   4. This is his_i room PRO_i+ to meet/kiss/hug in.

   5. This is his_i room in which PRO_i+ to meet/kiss/hug.

All of these examples exhibit partial control, i.e. the referent of PRO properly includes the controller (indicated by the index i+). Furthermore, verbs like hug (and for many speakers kiss as well) do not take comitative arguments in English. This shows that the partial control interpretation is not dependent on a covert comitative argument (pace Boeckx, Hornstein & Nunes Reference Boeckx, Hornstein and Nunes2010) at least in English (see also Sheehan Reference Sheehan, Lahousse and Marzo2014; Landau Reference Landau2016).

Partial control can also be seen in the following example:

1. (42)

   1. A: You had to be at school at 7am?!

   2. B: That was the headmaster's time to gather at! It certainly wouldn't have been mine!

The presence of partial control in IRCs thus supports the results from the structural diagnostics above, namely that we are dealing with OC.Footnote ⁸

The final OC/NOC diagnostic concerns arbitrary control. Given the evidence for OC that we have seen above, we would expect that PRO cannot be interpreted as arbitrary since PRO_arb is impossible in OC contexts by definition. However, it is perfectly possible to have a prenominal possessor attached to the RC head whilst simultaneously interpreting the IRC subject as arbitrary PRO. Consider our by-now familiar example:

1. (43) This is John's book to read.

We already know that this example has an interpretation where the IRC subject is interpreted as John. However, it also has an interpretation where PRO is arbitrary, for example, John may have recommended a book for others to read.Footnote ⁹ This is represented in (44).

1. (44) This is John's book PRO_arb to read.

On this interpretation it is possible to have a pronoun in the IRC in non-subject position that is co-referential with John without triggering a Condition B violation.

1. (45) This is John_i’s book PRO_arb to read to him_i.

When John does control PRO, a Condition B violation is present, as in (46a). To express such an interpretation, an anaphor satisfying Condition A must be used instead, as in (46b).

1. (46)

   1. (a) *This is John_i’s book PRO_John to read to him_i.

   2. (b) This is John_i’s book PRO_John to read to himself_i.

Inducing Condition B violations is also considered a strong test for partial control. Consider the following non-IRC OC context:

1. (47)

   1. (a) John_i wants PRO_i+ to gather at 6pm.

   2. (b) *John_i wants PRO_i+ to gather at 6pm without him_i.

In an OC context, John is included in the referent of PRO; co-indexing John and him thus leads to a Condition B violation. Now, let us consider IRCs. If PRO_arb interpretations are available, we would expect that it would be possible for him to be co-indexed with John on certain interpretations. This prediction is borne out. Consider (48):

1. (48) This is John_i’s room in which to gather at 6pm without him_i.

According to my consultants, in (48) either John is considered to be one of those gathering at 6pm, in which case my consultants report a Condition B violation, or John is not considered as such (for example, he may simply have suggested a room for other people to gather in), in which case my consultants report no Condition B violation. I thus take this as further evidence that IRCs permit both partial control and PRO_arb interpretations.

We saw above that the strict reading of PRO is impossible under ellipsis, as in (28) repeated below.

1. (49) This is John's book to read and that is Mary's.

   1. (a) This is John's book to read and that is Mary's <book PRO_Mary to read>.

   2. (b) *This is John's book to read and that is Mary's <book PRO_John to read>.

A PRO_arb interpretation can also be licensed in the ellipsis site, but only if such an interpretation is present in the antecedent, as in (50a, b). Furthermore, a PRO_arb interpretation in the antecedent cannot license a controlled-PRO interpretation in the ellipsis site, as in (50c).

1. (50)

   1. (a) This is John's book PRO_arb to read and that is Mary's <book PRO_arb to read>.

   2. (b) *This is John's book PRO_John to read and that is Mary's <book PRO_arb to read>.

   3. (c) *This is John's book PRO_arb to read and that is Mary's <book PRO_Mary to read>.

It is important to note that PRO is interpreted either as co-referential with the prenominal possessor or as arbitrary. As we observed above in relation to long-distance control and strict/sloppy interpretations under ellipsis, the reference of PRO in IRCs is not free. The relevant examples from above are repeated below as (51a, b).

1. (51)

   1. (a) *This is John's book to read and that is Mary's <book PRO_John to read>.

   2. (b) Mary said this is John's book PRO_John/*Mary to read.

If the reference of PRO were free, we would not expect these restrictions since the relevant interpretations would presumably be available through accidental co-reference. Therefore, we conclude that, in IRCs, PRO can be interpreted either as being controlled by the prenominal possessor or as arbitrary.

To summarise, the structural diagnostics very strongly suggest that we are dealing with OC. However, as far as interpretation is concerned, control is evidently not obligatory. In the next section, I will discuss an approach to this paradoxical state of affairs.

4.1 Previous accounts of control by a possessor

Control into IRCs cannot be assimilated to previous accounts of control by a possessor. Control by a possessor can be seen in so-called logophoric extension contexts and control within nominalisations. Let us first consider logophoric extension contexts. Landau (Reference Landau2000: 109ff.) notes that a controller seems not to be a direct argument of the matrix predicate in a well-defined set of cases. The following examples are taken from Landau (Reference Landau2000: 109–10):

1. (52)

   1. (a) It would help Bill_i’s development [PRO_i to behave himself in public].

   2. (b) PRO_i finishing his work on time is important to John_i’s development.

   3. (c) PRO_i finishing his work on time is important to John's friends_i.

2. (53)

   1. (a) *It would help Bill_i’s friends [PRO_i to behave himself in public].

   2. (b) It would help Bill's confidence [PRO to plan his itinerary in advance].

   3. (c) *It would help Bill's car [PRO to plan his itinerary in advance].

   4. (d) [PRO causing an uproar] is important for John's career.

Landau notes that the class of nouns that can contain the controller (as a possessor) is quite small and denotes abstract notions reflecting the individuality of the controller via actions, characters traits or social attributes (Landau Reference Landau2000: 110). This class contains nouns like career, status, confidence, performance, development, image, reputation, behaviour, etc. When a prenominal possessor denoting an individual, X, is attached to one of these nouns, Landau calls the result the logophoric extension of X. Landau (Reference Landau2000: 111) suggests that the class of logophoric extensions could be assimilated to the class of inalienably possessed nouns: these nouns do not introduce new individuals to the discourse, but rather highlight some aspect of the individual denoted by the possessor. Consequently, Landau suggests that such nouns do not block the index of the possessor (or, alternatively, such nouns inherit the index of their possessors) and so, in a way, the possessor can be considered an argument of the matrix predicate. However, our IRC examples are not amenable to a similar analysis since the prenominal possessor can control the IRC subject regardless of whether the RC head belongs to the class of logophoric extensions or not.

Now let us consider control within nominalisations. This is potentially more relevant since, in our examples, the possessor controls PRO within the DP projected by the external determiner. Hornstein (Reference Hornstein and Hendrick2003), cited in Landau (Reference Landau2013: 215), notes that possessors can be related to their head nouns in a number of different ways. In some cases, it looks as if we can choose between OC and NOC, which is particularly interesting from our perspective. Consider (54) (Landau Reference Landau2013: 215):

1. (54) [John_i’s plan [PRO_j≠i to bury him_i in the pit]] just won't work.

(54) can have an interpretation where PRO is disjoint in reference from John (ensured by the Condition B effect that would otherwise arise), showing that PRO is not obligatorily controlled by the possessor. However, Landau points out a potential confound. This interpretation relies on John not being interpreted as the thematic agent of plan, but in such cases, plan has a result reading rather than an eventive/process reading. Consequently, in such cases, plan does not take genuine arguments. Landau concludes that there is OC within DPs in derived nominals on their event readings (parallel to clauses) but not on their result readings.

However, in our examples, we cannot make recourse to an event vs result distinction to account for the controlled-PRO vs PRO_arb alternation that we observed above because the RC head is clearly not necessarily eventive and need not be a nominalisation of a canonical control predicate at all, as (55) shows.

1. (55) This is John_i’s book to read to him_i.

This example also forces disjoint reference between the prenominal possessor and PRO, i.e. in this example PRO cannot be interpreted as John and must be interpreted as arbitrary. However, unlike in (54), this is not plausibly related to any result vs eventive ambiguity relating to the RC head book. I thus conclude that previous accounts of control by a possessor cannot be extended to control into IRCs by a prenominal possessor.

4.2 Against a movement account

Control into IRCs would be problematic for any purely movement-based account (e.g. Hornstein Reference Hornstein1999; Manzini & Roussou Reference Manzini and Roussou2000; Boeckx, Hornstein & Nunes Reference Boeckx, Hornstein and Nunes2010) which attempts to say that the prenominal possessor has moved out of the IRC to its surface position.

First, movement out of IRCs with wh-relative pronouns is prohibited.

1. (56)

   1. (a) This is John's book in which to write a message.

   2. (b) *What (kind of message) is this John's book in which to write?

   3. (c) *the message that this is John's book in which to write

It therefore seems unlikely that the IRC subject has moved out of an IRC with a wh-relative pronoun to become the prenominal possessor. Furthermore, recall that partial control into IRCs with wh-relative pronouns is permitted. As pointed out by Landau (Reference Landau2003), partial control is very difficult to capture on a movement account of control. Furthermore, as pointed out above, whilst exhaustive control and a covert comitative might give the illusion of partial control (Boeckx, Hornstein & Nunes Reference Boeckx, Hornstein and Nunes2010), this analysis has been argued to be incorrect for English at least (Sheehan Reference Sheehan, Lahousse and Marzo2014; Landau Reference Landau2016).

The situation with IRCs without a wh-relative pronoun is potentially more debatable. In some cases, extraction from within the IRC is permitted, but in other cases it is not. In general, it seems that if the resulting configuration involves a nested dependency, the result is fine (at least for some speakers), but if it involves a crossing dependency, the result is degraded or ungrammatical.Footnote ¹⁰ Consider the following examples:

1. (57)

   1. (a) This is John's book to give to that student.

   2. (b) Which student is this John's book to give to?

   3. (c) To which student is this John's book to give?

   4. (d) the student who/that this is John's book to give to

   5. (e) the student to whom this is John's book to give

2. (58)

   1. (a) This is John's message to write in that book.

   2. (b) Which book is this John's message to write in?

   3. (c) the book that this is John's message to write in

In (57) and (58), the direct object of the infinitival clause has been relativised. As the data show, extracting a prepositional object from the IRC is fine. Configurationally, in such examples the IRC relativisation chain is nested within the chain formed by movement of the prepositional object, i.e. there is a nested dependency.

If, however, the prepositional object of the infinitival clause is relativised, and one tries to extract the direct object from the IRC, as in (59) and (60), the result is degraded or ungrammatical.

1. (59)

   1. (a) This is John's student to give that book to.

   2. (b) ^?/*Which book is this John's student to give to?

   3. (c) ^?/*the book that this is John's student to give to

2. (60)

   1. (a) This is John's book to write that message in.

   2. (b) ^?/*Which message is this John's book to write in?

   3. (c) ^?/*the message that this is John's book to write in

Configurationally, the chain formed by movement of the direct object crosses over the IRC relativisation chain, i.e. there is a crossing dependency.

Importantly, if one were to extract the IRC subject from the IRC, it would always cross the IRC relativisation chain and hence would always result in a crossing dependency. We would thus expect the result to be degraded or ungrammatical, contrary to fact.

I thus conclude that the control relation between a prenominal possessor and the IRC subject is not established by movement.

4.3 An Agree-based account

I will assume that the control relation is established by the operation Agree.Footnote ¹¹ Agree directly between the controller and PRO yields exhaustive control, whilst Agree between the controller and PRO that is mediated by the C-domain of the infinitival clause yields partial control (Landau Reference Landau2004, Reference Landau2008, Reference Landau2013, Reference Landau2015).

1. (61)

   

Since control into IRCs permits partial control, I will assume that the Agree relation between the prenominal possessor and PRO is mediated by a C-domain head in the IRC.Footnote ¹²

Following McFadden & Sundaresan (Reference McFadden, Sundaresan, Hammerly and Prickett2016), I will assume that OC PRO, NOC PRO, PRO_arb and definite pro are manifestations of the same underlying minimal pronoun (UPro). As a minimal pronoun, UPro is maximally underspecified, i.e. contains only unvalued features, when it enters the derivation. Assuming its unvalued features must be valued at Spellout, UPro must enter an Agree relation. According to McFadden & Sundaresan, OC PRO, NOC PRO and pro are all consequences of UPro entering into an Agree relation: OC PRO results when UPro Agrees with a matrix controller, NOC PRO results when UPro Agrees with a perspective-holder in the left-periphery of the infinitival clause, and pro results when UPro Agrees with a null topic in the left-periphery. As for PRO_arb, McFadden & Sundaresan propose that, if UPro fails to Agree with an antecedent, the derivation does not crash (see Preminger Reference Preminger2014) and UPro receives a default PRO_arb interpretation. In other words, PRO_arb is the elsewhere case and only arises where Agree fails to be established.

I have claimed above that PRO in IRCs is either controlled or interpreted as PRO_arb, i.e. NOC PRO and pro are unavailable in English. The absence of pro is perhaps unsurprising for English given that it is a non-pro-drop language. Furthermore, I have independently argued in previous work that English IRCs (and English RCs more generally) do not permit topics in their C-domains (Douglas Reference Douglas2016), i.e. UPro cannot Agree with a null topic in English and hence pro will not arise in IRCs. As for the absence of NOC PRO, we would have to say that IRCs do not contain a perspective holder in the left-periphery. Generally, if there is a perspective holder, we would expect only mind-possessing subjects to be possible, since only minds can hold perspectives. If, however, a perspective holder is absent, both mind-possessing and mind-lacking subjects should be possible. As (62) shows, it appears to be possible to have a mind-lacking controller.

1. (62) Context: A landscape gardener is designing an elaborate rock garden with several waterfalls and rivulets. The gardener takes the client to the proposed site of the rock garden and explains where various things will be, including where the water will be flowing. The gardener can indicate the proposed path of the water by saying:

   1. (a) This will be the water's path to flow down.

   2. (b) This will be the water's path down which to flow.

   3. (c) This will be the water's path to follow.

Water is not mind-possessing yet it can still serve as the controller of the IRC PRO, thus suggesting that there is no perspective holder in an IRC and hence accounting for the absence of NOC PRO.Footnote ¹³

We can now turn to the analysis of controlled-PRO and PRO_arb. Let us first consider IRCs where the prenominal possessor does control PRO, i.e. when a (mediated) Agree relation is established between the prenominal possessor and UPro. For this to happen, the prenominal possessor must (i) c-command UPro, and (ii) be local to UPro. That UPro is c-commanded by the prenominal possessor was demonstrated above, so I will focus here on the issue of locality, which I will assume is defined in phasal terms. Consider again the following structure (note that the IRC is now illustrated as having at least part of a C-domain given the framework adopted above):

1. (63) John's book to read

2. 

For John to be able to control PRO, there can be no phase boundary between them (where phase boundary means the maximal projection of a phase head). Infinitival clauses in canonical control contexts are generally considered not to be phasal because they lack all or part of the C-domain (see, e.g., Landau Reference Landau2015). More specifically relating to IRCs, recall from (57) and (58) that extraction out of IRCs is possible provided that nested dependencies are involved. As pointed out by an anonymous reviewer, assuming that the ‘phase’ edge is occupied by the relative operator, the fact that extraction is still permitted suggests that IRCs cannot be phases. In addition, in previous work I have independently argued that IRCs either lack the C-domain or only contain a truncated C-domain (Douglas Reference Douglas2016), so may not be phasal. One piece of evidence in favour of this comes from Distinctness effects. Richards (Reference Richards2010) proposes that the reason wh-relative pronouns must pied-pipe a preposition in IRCs, as shown in (64), but not in finite RCs, as shown in (65), is due to Distinctness.

1. (64)

   1. (a) The book in which to write is there.

   2. (b) *The book which to write in is there.

2. (65)

   1. (a) The book in which you should write is there.

   2. (b) The book which you should write in is there.

Distinctness says that two categories of the same type cannot be in the same Spellout domain because they will be unlinearisable. Adapting the details of Richards’ exact analysis slightly (see Douglas Reference Douglas2016 for discussion), Richards proposes that the RC head and wh-relative pronoun are in the same Spellout domain in IRC contexts and so cannot both be DPs. To avoid this problem, the wh-relative pronoun pied-pipes a preposition, i.e. it is a PP. Because P is phasal in English, the DP wh-relative pronoun and the DP RC head are no longer in the same Spellout domain, and the structure can be successfully linearised. In contrast, wh-relative pronouns in finite RC contexts do not have to pied-pipe a preposition. Richards proposes that this is because finite RCs have a phase boundary. Consequently, the RC head and wh-relative pronoun will always be in separate Spellout domains in finite RC contexts. Exact details aside, the important claim is that there is no phase boundary at the left edge of an IRC.Footnote ¹⁴

English D is commonly assumed to be a phase head (see, e.g., Bošković Reference Bošković2014) and the prenominal possessor is standardly taken to occupy SpecDP in English. Elements in the phase edge may form Agree relations with elements in the phase head's complement. Furthermore, recall that restrictive RCs (including IRCs) modify NPs, which are considered to be non-phasal in English (Bošković Reference Bošković2014). Consequently, we would not necessarily expect there to be a phase boundary between the prenominal possessor and PRO (though see below). These considerations lead us to expect that the prenominal possessor and PRO are local in the phase-theoretic sense.

I will thus assume that UPro Agrees with a C head and that this C head Agrees with the prenominal possessor in SpecDP.Footnote ¹⁵ Because there is no phase boundary between SpecDP and UPro, UPro can be valued by the prenominal possessor via the C head. I will set aside the interesting question of which C head this is and what its properties are for future research. Landau (Reference Landau2015), for example, draws a very tight connection between this C-position and logophoricity. However, if my characterisation of the IRC data is correct (recall (62)), this C head should not be equated with the logophoric centre (see also McFadden & Sundaresan Reference McFadden, Sundaresan, Hammerly and Prickett2016).

Let us now consider IRCs where PRO is interpreted as PRO_arb. According to the theory of control being pursued here, this means that we are dealing with a case where UPro fails to be valued by anything via Agree. There are thus two analytic possibilities that could in principle account for why the prenominal possessor fails to value UPro: (i) the c-command condition is not met, or (ii) the locality condition is not met.

We can rule out the c-command condition option by showing that the prenominal possessor c-commands UPro even in cases where UPro is interpreted as PRO_arb. This can be shown using Condition C effects. We have already seen that prenominal possessors generally c-command the RC head and the RC, as shown by the Condition C violation in (36a), repeated below.

1. (66) *This is his_i book for John_i to read.

If it were possible for the prenominal possessor to appear in a position that did not c-command the IRC subject (except for those possessors appearing in compounds, as mentioned above), we would expect it to be possible to avoid the Condition C violation. However, this does not seem to be possible, suggesting that the c-command relations between the prenominal possessor and the IRC subject do not change.

Indeed, we can get a Condition C violation in the IRC whilst simultaneously getting a PRO_arb interpretation. Recall (55), repeated as (67a). Swapping the pronoun and R-expression results in a Condition C violation, as in (67b), showing that the pronominal prenominal possessor c-commands into the IRC.

1. (67)

   1. (a) This is John_i’s book PRO_arb to read to him_i.

   2. (b) *This is his_i book PRO_arb to read to John_i.

The same point can be made for (68) (thanks to an anonymous reviewer for suggesting the example from which this is adapted).

1. (68)

   1. (a) This is Mary_i’s image to compare oneself to using her_i recognition software.

   2. (b) *This is her_i image to compare oneself to using Mary_i’s recognition software.

In (68), use of oneself in the IRC is designed to make the PRO_arb interpretation of the IRC subject more salient. The adjunct using Mary's/her recognition software modifies the IRC verb phrase, i.e. it is the means by which the comparison is carried out, and thus it is c-commanded by anything that c-commands the IRC as a whole. (68b) shows that her and Mary cannot be co-indexed, i.e. a Condition C violation, thus showing that her (the prenominal possessor) c-commands into the IRC without necessarily simultaneously controlling the IRC subject (PRO_arb). (68a) shows that switching the pronominal and R-expression results in a grammatical example.

These data show that the c-command relations between the IRC PRO subject and the prenominal possessor attached to the RC head do not change depending on whether PRO is controlled or PRO_arb. Consequently, a violation of the c-command condition cannot be the reason for the failure of the prenominal possessor to value UPro in PRO_arb cases.

By the process of elimination, this leaves the locality condition. The existence of PRO_arb interpretations suggests that there must be a phase boundary between the prenominal possessor and UPro. There are in principle two ways of implementing this idea: (i) the prenominal possessor is merged higher than SpecDP in PRO_arb contexts, i.e. it is external to the DP phase, or (ii) there is a DP-internal phase boundary between the prenominal possessor and UPro (this phase boundary would have to be optional to account for the availability of the controlled-PRO interpretation).

Option (i) is reminiscent of possessor raising constructions in which the possessor is syntactically external to the DP with which it is semantically associated. This could be implemented either as a type of raising operation or as a type of control (see Deal Reference Deal2013 for an overview). However, syntactically speaking, in cases of possessor raising, the possessor and possessum generally constitute separate constituents and separate arguments of the verb. If this were so in our IRC examples, we might expect it to be possible to extract one without the other. However, this is not possible.

1. (69) *Which book to read did you say that this was John's?

2. (cf. You said that this was John's crime novel PRO_arb to read)

This suggests that we are not dealing with anything resembling a possessor raising construction.

This leaves option (ii); there is an optional phase boundary between the prenominal possessor and UPro. This would predict that there is an intermediate landing site for elements undergoing successive cyclic movement. We saw above that wh-movement from IRCs is permitted provided there is a nested dependency between the relativised element and the wh-extracted element (at least for some speakers). We should therefore, in principle, be able to test whether an element can be licensed only in that intermediate position. Following Fox (Reference Fox2000), I will attempt to isolate this position by using a combination of variable binding and Condition C diagnostics. Consider the following on the interpretation where Mary is the subject of write:

1. (70) This is every fan_i’s message for Mary_j to write on the picture that he_i took of her_j.

We can force a Condition C violation by switching Mary and her (again, on the interpretation where her is the subject of write).

1. (71) *This is every fan_i’s message for her_j to write on the picture that he_i took of Mary_j.

For every fan to variable bind he, the former must c-command the latter, and to satisfy Condition C, her must not c-command Mary. Now consider what happens when the constituent containing picture and its relative clause is questioned. (70) thus becomes (72).

1. (72) On which picture that he_i took of her_j is this every fan_i’s message for Mary_j to write?

(72) permits simultaneous variable binding of he and satisfaction of Condition C. For this to happen, the questioned constituent is interpreted in a position c-commanded by every fan (via reconstruction). Now consider what happens when Mary and her are switched, i.e. (71) becomes (73).

1. (73) On which picture that he_i took of Mary_j is this every fan_i’s message for her_j to write?

Unlike (71), which contains a Condition C violation, (73) permits simultaneous variable binding of he and satisfaction of Condition C. For this to be possible, the questioned constituent must be reconstructing to a position where it is c-commanded by every fan but not c-commanded by her. The availability of this interpretation in (73) indicates the presence of an intermediate landing site somewhere between the IRC subject (her) and the prenominal possessor (every fan). Assuming this is evidence for a phase boundary, when this phase boundary is present, a UPro subject will be spelled out before the prenominal possessor in SpecDP has a chance to value UPro's features. By assumption, this does not cause the derivation to crash, but it does result in UPro's features being valued by default, yielding PRO_arb.

Ideally, one would like to show that this intermediate phase edge is absent in cases where the prenominal possessor controls the IRC subject. However, the sort of test employed above will not work as the control relation itself forces the prenominal possessor and IRC subject to be co-referential (or partially so in the case of partial control). An intermediate landing site between the prenominal possessor and IRC subject would thus be indistinguishable from an intermediate landing site in clause-medial position in the IRC (at least in terms of the interpretations it licences).

Therefore, given the tools and diagnostics currently available, I conclude that the proposal of there being an optional phase boundary between the prenominal possessor and IRC subject is at least not inconsistent with the data. However, further research would be required to make any firmer conclusions. I thus tentatively conclude that there is an optional phase boundary between the prenominal possessor and UPro. When the phase boundary is absent, the prenominal possessor values UPro's features (mediated by a C head) resulting in the controlled-PRO interpretation; when the phase boundary is present, UPro is spelled out before the prenominal possessor can value UPro's features resulting in the PRO_arb interpretation.