Cedric Boeckx, Norbert Hornstein, Jairo Nunes Control as Movement 2010

http://www.cambridge.org/9780521195454

This page intentionally left blank

C O N T RO L A S M OV E M E N T

The movement theory of control (MTC) makes one major claim: that controlrelations in sentences like John wants to leave are grammatically medi-ated by movement. This goes against the traditional view that such sentencesinvolve not movement, but binding, and analogizes control to raising, albeitwith one important distinction: whereas the target of movement in controlstructures is a theta position, in raising it is a non-theta position; however, thegrammatical procedures underlying the two constructions are the same. Thisbook presents the main arguments for MTC and shows it to have many theo-retical advantages, the biggest being that it reduces the kinds of grammaticaloperations that the grammar allows, an important advantage in a minimalistsetting. It also addresses the main arguments against MTC, using examplesfrom control shift, adjunct control, and the control structure of promise,showing MTC to be conceptually, theoretically, and empirically superior toother approaches.

cedric boeckx is Research Professor at the Catalan Institute for AdvancedStudies (ICREA), and a member of the Center for Theoretical Linguistics atthe Universitat Autonoma de Barcelona.

norbert hornstein is Professor in the Department of Linguistics at theUniversity of Maryland, College Park.

jairo nunes is Professor in the Department of Linguistics at the Universi-dade de Sao Paulo, Brazil.

In this series

81 roger lass : Historical linguistics and language change82 john m. anderson : A notional theory of syntactic categories83 bernd heine : Possession: cognitive sources, forces and grammaticalization84 nomi erteschik-shir : The dynamics of focus structure85 john coleman : Phonological representations: their names, forms and powers86 christina y. bethin : Slavic prosody: language change and phonological theory87 barbara dancygier : Conditionals and prediction88 claire lefebvre : Creole genesis and the acquisition of grammar: the case of Haitian

creole89 heinz giegerich : Lexical strata in English90 keren rice : Morpheme order and semantic scope91 april m cmahon : Lexical phonology and the history of English92 matthew y. chen : Tone Sandhi: patterns across Chinese dialects93 gregory t. stump : Inflectional morphology: a theory of paradigm structure94 joan bybee : Phonology and language use95 laurie bauer : Morphological productivity96 thomas ernst : The syntax of adjuncts97 elizabeth closs traugott and richard b. dasher : Regularity in semantic

change98 maya hickmann : Childrens discourse: person, space and time across languages99 diane blakemore : Relevance and linguistic meaning: the semantics and pragmatics of

discourse markers100 ian roberts and anna roussou : Syntactic change: a minimalist approach to

grammaticalization101 donka minkova : Alliteration and sound change in early English102 mark c. baker : Lexical categories: verbs, nouns and adjectives103 carlota s. smith : Modes of discourse: the local structure of texts104 rochelle lieber : Morphology and lexical semantics105 holger diessel : The acquisition of complex sentences106 sharon inkelas and cheryl zoll : Reduplication: doubling in morphology107 susan edwards : Fluent aphasia108 barbara dancygier and eve sweetser : Mental spaces in grammar: conditional

constructions109 hew baerman , dunstan brown, and greville g. corbett : The

syntaxmorphology interface: a study of syncretism110 marcus tomalin : Linguistics and the formal sciences: the origins of generative

grammar111 samuel d. epstein and t. daniel seely : Derivations in minimalism112 paul de lacy : Markedness: reduction and preservation in phonology113 yehuda n. falk : Subjects and their properties114 p. h. matthews : Syntactic relations: a critical survey115 mark c. baker : The syntax of agreement and concord116 gillian catriona ramchand : Verb meaning and the lexicon: a first phase syntax117 pieter muysken : Functional categories118 juan uriagereka : Syntactic anchors: on semantic structuring119 d. robert ladd : Intonational phonology, second edition120 leonard h. babby : The syntax of argument structure121 b. elan dresher : The contrastive hierarchy in phonology122 david adger , daniel harbour, and laurel j. watkins : Mirrors and

microparameters: phrase structure beyond free word order123 niina ning zhang : Coordination in syntax124 neil smith : Acquiring phonology125 nina topintzi : Onsets: suprasegmental and prosodic behaviour126 cedric boeckx , norbert hornstein, and jairo nunes : Control as movement

Earlier issues not listed are also available

CAMBRIDGE STUDIES IN LINGUISTICS

General editors: p. austin, j . bresnan, b. comrie,s . crain, w. dressler, c. j . ewen, r. lass ,d. lightfoot, k. rice, i . roberts,s . romaine, n. v. smith

Control as Movement

CONTROL AS MOVEMENT

C E D R I C B O E C K XICREA/Universitat Autonoma de Barcelona

NORBERT HORNS T E I NUniversity of Maryland, College Park

JAIRO NUNE SUniversidade de Sao Paulo, Brazil

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,

So Paulo, Delhi, Dubai, Tokyo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK

First published in print format

ISBN-13 978-0-521-19545-4

ISBN-13 978-0-511-78955-7

Cedric Boeckx, Norbert Hornstein, and Jairo Nunes 2010

2010

Information on this title: www.cambridge.org/9780521195454

This publication is in copyright. Subject to statutory exception and to the

provision of relevant collective licensing agreements, no reproduction of any part

may take place without the written permission of Cambridge University Press.

Cambridge University Press has no responsibility for the persistence or accuracy

of urls for external or third-party internet websites referred to in this publication,

and does not guarantee that any content on such websites is, or will remain,

accurate or appropriate.

Published in the United States of America by Cambridge University Press, New York

www.cambridge.org

eBook (NetLibrary)

Hardback
http://www.cambridge.orghttp://www.cambridge.org/9780521195454

Contents

Acknowledgments page x

1 Introduction 1

2 Some historical background 52.1 Introduction 52.2 What any theory of control should account for 52.3 Control in the standard-theory framework 62.4 Control in GB 92.5 Non-movement approaches to control within minimalism 16

2.5.1 The null-case approach 162.5.2 The Agree approach 20

2.6 Conclusion 35

3 Basic properties of the movement theory of control 363.1 Introduction 363.2 Departing from the null hypothesis: historical, architectural, and

empirical reasons 373.3 Back to the future: elimination of DS and the revival of the

null hypothesis 433.4 Controlled PROs as A-movement traces 46

3.4.1 Configurational properties 473.4.2 Interpretive properties 493.4.3 Phonetic properties and grammatical status 52

3.5 Conclusion 56

4 Empirical advantages 594.1 Introduction 594.2 Morphological invisibility 594.3 Interclausal agreement 604.4 Finite control 63

4.4.1 Finite control and hyper-raising 70

vii

viii Contents

4.4.2 Finite control, islands, and intervention effects 754.4.3 Summary 79

4.5 The movement theory of control under the copy theory of movement 794.5.1 Adjunct control and sideward movement 834.5.2 The movement theory of control and morphological

restrictions on copies 984.5.3 Backward control 1024.5.4 Phonetic realization of multiple copies and copy control 115

4.6 Conclusion 123

5 Empirical challenges and solutions 1255.1 Introduction 1255.2 Passives, obligatory control, and Vissers generalization 125

5.2.1 Relativizing A-movement 1275.2.2 Impersonal passives 1325.2.3 Finite control vs. hyper-raising 136

5.3 Nominals and control 1415.3.1 Finite control into noun-complement clauses in Brazilian

Portuguese 1425.3.2 Raising into nominals in Hebrew 1475.3.3 The contrast between raising nominals and control nominals

in English 1495.4 Obligatory control and morphological case 152

5.4.1 Quirky case and the contrast between raising and controlin Icelandic 152

5.4.2 Apparent case-marked PROs 1605.5 The minimal-distance principle, control shift, and the logic

of minimality 1695.5.1 Control with promise-type verbs 1715.5.2 Control shift 1765.5.3 Summary 181

5.6 Partial and split control 1825.6.1 Partial control 1835.6.2 Split control 190

5.7 Conclusion 194

6 On non-obligatory control 1956.1 Introduction 1956.2 Obligatory vs. non-obligatory control and economy computations 1966.3 Some problems 2026.4 A proposal 2046.5 Conclusion 209

Contents ix

7 Some notes on semantic approaches to control 2107.1 Introduction 2107.2 General problems with selectional approaches to obligatory control 2107.3 Simpler syntax 216

7.3.1 Some putative problems for the movement theory of control 2177.3.2 Challenges for simpler syntax 226

7.4 Conclusion 237

8 The movement theory of control and the minimalistprogram 238

8.1 Introduction 2388.2 Movement within minimalism and the movement theory of control 2398.3 The movement theory of control and the minimalist architecture

of UG 2418.4 Inclusiveness, bare phrase structure, and the movement theory

of control 2458.5 Conclusion 248

References 250Index 261

Acknowledgments

Previous versions of (part of) the material discussed here have been presentedat the following universities: Connecticut, Harvard, Leiden, Lisbon, Maryland,New York, Rutgers, Sao Paulo, Stony Brook, Tilburg, and Utrecht; and at thefollowing meetings: ANPOLL 2003, XVIII Colloquium on Generative Gram-mar, Edges in Syntax, EVELIN 2004, GLOW XXX, Going Romance 2007,LSA 2005, Romania Nova II, Ways of Structure Building, and V Workshop onFormal Linguistics at USP. We would like to thank these audiences for com-ments and suggestions. Special thanks to Zeljko Boskovic, Hans Broekhuis,Lisa Cheng, Marcelo Ferreira, Michael Gagnon, Terje Lohndal, Carme Picallo,and Johan Rooryck.

We would also like to acknowledge the support received from theGeneralitat de Catalunya (grant 2009SGR1079; first author), NSF (grantNSD.BCS.0722648; second author), and CNPq and FAPESP (grants302262/20083 and 2006/009652; third author).

x

1 Introduction

In the following pages we develop an extended argument for a proposal whoseconceptual simplicity and empirical success will, we trust, be evident to allreaders. The proposal says that (obligatory) control is movement, more specif-ically, A-movement. We propose that the phenomena that have been usedto motivate a special and separate control construction are best explained ifcontrol is treated as an A-movement dependency, on a par with other phe-nomena that have been traditionally treated in terms of A-movement such aspassive, raising, and (local) scrambling. Put another way, we claim that main-taining the constructional specificity of control (in whatever form, be it interms of the PRO theorem [e.g., Chomsky 1981], null case [e.g., Chomskyand Lasnik 1993; Martin 1996; and Boskovic 1997], or ad hoc anaphorictense-agreement dependencies [e.g., Landau 1999, 2000, 2004]) significantlyhampers our understanding of the phenomenon as it leads to explanations thatare roughly as complex as the phenomenon itself.

Despite virtues that we believe are transparent (see e.g., Hornstein 1999,2001), the movement theory of control (hereafter, MTC) has proven to be quitecontroversial.1 We believe that there are several reasons for this. The first one ishistorical. Differentiating raising from control in terms of movement has beena fixed point within generative grammar from the earliest accounts within thestandard theory to current versions of minimalism (see Davies and Dubinsky2004). Under this long-held view, which became crystallized in GB with theformulation of the (construction-specific) control module (Chomsky 1981), ifraising involves movement, control cannot. It is thus not surprising that theMTC has been welcomed with considerable skepticism, as its basic proposal isexactly to analyze control in terms of (A-)movement. However, such historicalbias should not deter us from a fair evaluation of the conceptual properties andempirical coverage of the MTC.

1 See e.g., Landau (2000, 2003); Culicover and Jackendoff (2001, 2005); Kiss (2005); and vanCraenenbroeck, Rooryck, and van den Wyngaerd (2005) for a useful sample.

1

2 Introduction

The second reason behind the controversy is also related to the long interestcontrol has enjoyed within the generative tradition. Over the years, controlphenomena have been richly described. Consequently, any new approach willlikely fail, at least initially, to adequately handle some of the relevant data.Moreover, if the novel approach is conceptually tighter than the more descrip-tive accounts that it aims to replace (as we believe to be the case with theMTC), some features of the phenomenon heretofore assumed to be central maynot be accommodated at all. This should occasion no surprise, as it reflectsthe well-known tension between description and explanation. Odd as it mayseem, failure to cover a data point may be a mark of progress if those that arecovered follow in a more principled fashion. The virtues of a proposal can beseriously misevaluated unless one keeps score of both what facts are coveredand how facts are explained. A weak theory can often be easily extended toaccommodate yet another data point, and this is not a virtue. Correspondingly,a tight theory may miss some facts and this is not necessarily a vice, par-ticularly if the account is comparatively recent and the full implications of itsresources have not yet been fully developed. We believe that many have beentoo impressed by these apparent problems without considering how the MTCmight be developed to handle them. In fact, we believe that the MTC actuallyfaces few empirical difficulties (and none of principle), whereas the currentalternatives both face very serious empirical hurdles (e.g., backward control)and often empirically succeed by stipulating what should be explained (e.g.,the distribution of PRO through null case). One aim of what follows is to makethis case in detail.

Finally, it is fair to say that the resistance to MTC is in part due to theinadequacies and limitations of previous versions of the MTC (including ourown work), which we have tried to overcome here. Addressing the vigorouscritiques of MTC here and in previous work (Hornstein 2003; Boeckx andHornstein 2003, 2004, 2006a; Nunes 2007; Boeckx, Hornstein, and Nunes inpress) has allowed us to rectify some errors, clarify the proposal, and sharpen thearguments. This stimulating intellectual exercise has led us to better appreciatethe consequences of the MTC and has in fact convinced us that it covers evenmore empirical ground than we at first thought, as we will argue in the followingchapters.

For all these reasons, we thought that a detailed defense of MTC requireda monograph. But before we launch our defense of MTC, a few notes are inorder.

First, we cannot emphasize enough that MTC does not equate control withraising. Since the MTC was first proposed, it has been regularly objected

Introduction 3

that the MTC cannot be right because of features that control has, but raisingdoes not, and vice versa. However, control is raising only in the descriptivesense that control is an instance of A-movement, but it is not raising quaconstruction. In other words, all the MTC is saying is that, like the derivation ofraising, passive, or local scrambling constructions, the derivation of obligatory-control constructions also involves A-movement. The different properties ofconstructions involving wh-movement and topicalization, for instance, do notargue against analyzing them in terms of A-movement. Similarly, we urgethe reader not to dismiss our proposal simply because (unanalyzed) controlraising asymmetries exist. Although raising often proves useful in illustratingproperties of A-movement that carry over to control, it is a ladder that ought tobe kicked away as theory advances. In the chapters that follow, we in fact arguethat controlraising asymmetries generally reduce to independent factors something we take to be an indication that the MTC is on the right track.

Second, the MTC is actually not a radically new idea. It goes back as faras Bowers (1973), who already proposed that raising and control should bebasically generated in the same way. However, as the proposal conflicted withcore principles of almost every model of UG from Aspects to GB, it did not findfertile soil to blossom for a long time. This scenario drastically changed whenthe minimalist program came into the picture. Chomskys (1993) proposal thatD-structure should be eliminated provided a very natural conceptual niche forthe MTC within the generative enterprise as it removed the major theoreticalobstacle that prevented movement to -positions. In a system with D-structure,movement to -positions is a non-issue, for movement can only take place once-assignment is taken care of. By contrast, in a system without D-structure,where movement and -assignment intersperse, movement to -positions arisesat least as a logical possibility. Thus, whether or not it is a sound option hasto be determined on the basis of the other architectural features of the system,as well as its empirical coverage. We hope to show that the MTC fits snuglywith some leading minimalist conceptions and thus constitutes an interestingargument in its favor.

Third, as minimalism aspires to explain why UG properties are the waythey are, we are interested in developing a theory of control that deducesthe properties of control configurations from more basic postulates, ratherthan merely listing the possible controllers, controllees, control predicates, andcontrol complements coded as features of individual lexical items.

Finally, although our specific implementation of MTC is the one that hasbeen extended to the broadest range of data thus far, it is certainly not the onlyone possible. ONeil (1995), Manzini and Roussou (2000), Kayne (2002), and

4 Introduction

Bowers (2006) share the spirit but not the details of our analysis. For reasonsof space, we will not be able to do proper justice to these works and the readeris invited to evaluate each different implementation in its own right.

Let us close this introductory chapter by providing an overview of the sub-sequent chapters. Chapter 2 offers a brief overview of how control is handledin the standard-theory framework, in GB, and in non-movement approacheswithin minimalism. Chapter 3 lays out the broad features of our version of theMTC. Chapter 4 discusses some of the empirical advantages that the MTC has.Chapter 5 addresses many of the empirical challenges that have been consid-ered to be fatal to the MTC and proposes solutions compatible with the MTC.Chapter 6 presents our take on how non-obligatory control is to be analyzed.Chapter 7 discusses the extent to which the MTC is based on more solid concep-tual and empirical grounds than semantic/selectional approaches to obligatorycontrol. Finally, Chapter 8 concludes the monograph.

2 Some historical background

2.1 Introduction

Up to very recently, there had been a more or less uncontroversial view thatcontrol phenomena should be analyzed in terms of special grammatical primi-tives (e.g., PRO) and construction-specific interpretive systems (e.g., the controlmodule). In this chapter, we examine how this conception of control was instan-tiated in the standard-theory framework (section 2.3), in GB (section 2.4), andin non-movement analyses within the minimalist program (section 2.5), brieflyoutlining what we take to be the virtues and problems of each approach.1 Thisdiscussion will provide the general background for us to discuss the core prop-erties of (our version of) the MTC in Chapter 3 and evaluate its adequacy inthe face of the general desiderata for grammatical downsizing explored in theminimalist program.

2.2 What any theory of control should account for

A theoretically sound approach to control one that goes beyond the merelisting of the properties involved in control must meet (at least) the followingfour requirements.

First, it must specify the kinds of control structures that are made availableby UG and explain how and why they differ. Assuming, for instance, thatobligatory control (OC) and non-obligatory control (NOC) are different, theirdifferences should be reduced to more basic properties of the system.

Second, it must correctly describe the configurational properties of control,accounting for the positions that the controller and the controllee can occupy.In addition, it should provide an account as to why the controller and thecontrollee are so configured. Assuming, for instance, that the controllee can

1 For much more detailed discussion, we urge the reader to consult Davies and Dubinskys (2004)excellent history of generative treatments of raising and control.

5


only appear in a subset of possible positions (e.g., ungoverned subjects), whyare controllees so restricted?

Third, it must account for the interpretation of the controllee, explaininghow the antecedent of the controllee is determined and specifying what kind ofanaphoric relation obtains between the controllee and its antecedent (in bothOC and NOC constructions) and why these relations obtain and not others.For instance, assuming that controllers must locally bind controllees in OCconstructions, why is the control relation so restricted in these cases?

Fourth, it must specify the nature of the controllee: what is its place amongthe inventory of null expressions provided by UG? Is it a formative special tocontrol constructions or is it something that is independently attested?

In the next sections, we briefly review how these concerns have beenaddressed from the standard-theory model to the minimalist program.

2.3 Control in the standard-theory framework

Within the framework of the standard theory, control phenomena were codedin the obligatory transformation referred to as equi(valent) NP deletion (END),which for our current purposes can be described as follows:2

(1) X-NP-Y-[S {for/poss}-NP-Z]-WStructural description: 1 2 3 4 5 6 7 Structural change: 1 2 3 4 6 7Conditions: i. 2 = 5

ii. the minimal-distance principle is satisfied

Irrelevant details aside, END applies to the (a)-structures in (2)(5), for instance,and converts them in the corresponding (b)-sentences.

(2) a. John tried/wanted/hoped [for John to leave early] b. John tried/wanted/hoped to leave early

(3) a. John regrets/insisted on/prefers [poss John leaving early] b. John regrets/insisted on/prefers leaving early

(4) a. John persuaded/ordered/forced/asked/told Mary [for Mary to leaveearly]

b. John persuaded/ordered/forced/asked/told Mary to leave early

(5) a. John kissed Mary before/after/without [poss John asking if he could] b. John kissed Mary before/after/without asking if he could

2 Here we abstract away from issues that are orthogonal to our discussion such as the interactionbetween END and the rule of complementizer deletion, which has the effect of deleting the termnumbered 4 in (1). See Rosenbaum (1967, 1970) for discussion.

2.3 Control in the standard-theory framework 7

According to this approach, there is nothing of special interest in the natureof the controllee. It is a regular NP in the underlying structure and the factthat the corresponding surface position is phonetically null follows from thekind of transformation END is. It is a deletion transformation that removesthe targeted NP, leaving nothing at surface structure. To put it differently, thesuperficial phonetic difference between controller and controllee results notfrom intrinsic lexical properties of the controllee, but from properties of thecomputation itself, i.e., that END is a deletion operation.

As far as the configurational properties of control are concerned, END explic-itly specifies that the controllee (the target of deletion) must occur in the subjectposition of infinitival clauses (for-clauses) and gerunds (poss-clauses), and thatthe controller must be the closest NP (in compliance with the minimal-distanceprinciple). Thus, according to the minimal-distance principle, sentences suchas (4b) must be derived from the structures in (4a) and not from the one in(6) below, which would incorrectly allow the understood subject of the embed-ded clause to be interpreted as being coreferential with the matrix subject. Asopposed to what we find in (4a), the antecedent of the controllee in (6) is notthe closest NP around. As for adjunct control in sentences such as (5), theminimal-distance principle is satisfied under the assumption that the embeddedclause is adjoined to the matrix clause and, as such, it is structurally closer tothe subject than it is to the object.3

(6) John persuaded/ordered/forced/asked/told Mary [for John to leave early]

Finally, the interpretation properties of control are enforced by condition(i), which requires that controller and controllee be identical, which wasunderstood in terms of coreference.

This general approach was refined within the standard theory as more com-plex control structures were considered, but its axiomatic (i.e., stipulative)nature remained. The configurational and interpretive properties of controlwere analyzed as irreducible features of the END transformation itself. Thisby no means diminishes the value of these earlier approaches to control. Iden-tifying the different properties of control phenomena with such formal rigor

3 END as stated is not entirely adequate empirically. Given (1) above, the structure in (ia), forexample, should allow for control by Mary in (ib):

(i) a. John persuaded a friend of Mary [for Mary to leave]b. John persuaded a friend of Mary to leave

It should be clear how requiring that some sort of command relation hold between the antecedentNP and the deleted one will help screen out cases like (i), where the wrong NP is chosen.


was unquestionably an achievement, with large consequences for theorizingbeyond control structures, and it paved the way for subsequent reanalyses inGB and in the minimalist program.

Before we leave this brief review, two points are worth mentioning which willbe relevant to the discussion of these later reanalyses, including the MTC. Thefirst one regards an empirical problem that the standard-theory approach facedin relation to the way it handled the interpretive properties of control. As we sawabove, the controller and the controllee were taken to be lexically identical andthe semantic relation between them was understood as coreference. Problemsarise when the controller is not a referential NP, as exemplified by the contrastbetween (7) and (8).

(7) a. [John wants [John to win]] b. John wants to win

(8) a. [Everyone wants [everyone to win]] b. Everyone wants to win

Whereas (7a) might be taken to roughly represent the meaning of (7b),(8a) in no way represents the interpretation of (8b), which should rather beparaphrased as Everyone wants himself to win. This suggests that, instead ofan NP identical to (i.e., coreferential with) its controller in underlying structure,what we actually need is a kind of bound anaphor or an expression that canbe so interpreted.4 The obvious question then is how to obtain this boundinterpretation.

The second point worth mentioning concerns the identification of anothertype of control. Relatively early on, END was distinguished from a relatedoperation dubbed super-equi (SEND). This operation also deletes a subjectof a non-finite clause but, in contrast to END, it operates across unboundedstretches of sentential material, as illustrated in (9).5

4 If there is an anaphoric relation in control structures, then END is unlikely to be a chopping(gap-leaving) rule. Rather, it is more like the rules of reflexivization or pronominalization,which were operations governed by command relations. The problem is that control structuresdo not appear to leave lexical residues like the other construal operations. They appear to requirea phonetic gap. Seen from a contemporary perspective, the problem of how to characterize therules that lead to control structures (are they chopping rules or construal rules?) highlightsthe tension that we will see constantly recurring: how best to account for both the distribution ofthe controllee and its interpretation.

5 See Grinder (1970). Data such as (9) are not the sorts of cases Grinder discussed, but they fallunder the SEND rubric.

2.4 Control in GB 9

(9) a. [S1 John said [S2 that Mary believes [S3 that [S4 John washing himself]would make a good impression on possible employers]]]

b. John said that Mary believes that washing himself would make a goodimpression on possible employers

Note that (9) violates the minimal-distance principle, as Mary intervenesbetween the target of deletion (John in S4) and its antecedent (John in S1).Moreover, in contrast to standard END configurations, the controllee is notwithin a clausal complement (or adjunct) of a higher predicate. In (9), forinstance, the controllee is within the sentential subject of S3. The followingquestion then arises: what is the relation between END and SEND? Or to putthe question somewhat differently: why should UG have two rules that have thesame effect (deletion of an identical NP), but apply to different configurations?6

In the next sections we examine some answers to these two issues that wereoffered within GB and the minimalist program.

2.4 Control in GB

Building on earlier work in the extended standard theory (EST), the GBapproach to control is considerably more ambitious and empirically more suc-cessful than the standard-theory model.

Within GB, the controllee is a PRO, a base-generated NP containing nolexical material ([NP ]). This conception of the controllee as a base-generatednon-lexical formative arises as a natural consequence of the GB assumptionsregarding the base component. The GB theory of the base includes both phrase-structure rules, like the ones in (10), and lexical-insertion operations, like theones in (11).

(10) a. S NP INFL VPb. VP V NPc. NP N

(11) a. N John/he/it/Billb. V kiss/see/admirec. INFL past/to

These two types of rules operate in tandem to generate structures such as (12)below. However, they can also be used to generate structures like (13), wherethe subject of the clause has been generated by the phrase-structure component

6 Grinder (1970) actually collapsed END and SEND. However, later approaches identified manysubstantial differences between the constructions underlying END and SEND that are bettercaptured if two kinds of control are recognized, as we shall see below.


but has not been filled by lexical insertion. In short, a theory of the base factoredinto a set of phrase-structure rules and lexical-insertion operations has room foran element like PRO: it is what one gets when one generates an NP structurebut does not subject it to lexical insertion.

(12) [S [NP John] past [VP see [NP Bill]]]

(13) [S [NP ] to [VP see [NP Bill]]]

This way of understanding PRO has an interesting consequence for theconstructions that were captured by END in the standard theory. If one assumesthat categories without lexical content are uninterpretable unless provided withcontent (by being linked with an antecedent, for example) and, furthermore,that the principle of full interpretation does not tolerate contentless structures,then the requirement that PRO must have an antecedent follows naturally.7

We wish to stress this point as it is important for some of the discussion thatfollows. If one treats PRO as a lexical element, it is hard to explain why PROmust be phonetically null and why it requires an antecedent. Of course, it ispossible to stipulate that these two features are inherent properties of a specificlexical item (PRO), but this cannot explain why PRO is necessarily anaphoricand null. Moreover, so conceived, PRO is a rather unusual lexical element as ithas no positive properties. It has no phonetic matrix and its only semantic featureis the requirement that it must be coindexed with a grammatical antecedent.8

This point is worth emphasizing. PRO, on this view, is not simply a semanticallydependent expression that needs to be interpreted with respect to some salientelement in the discourse (e.g., like the other in John ate one of the bagels.Harry ate the other.). Rather, PRO is specified as needing an antecedent in aparticular structural configuration. However, this is a very odd lexical feature asit is only definable in configurational (i.e., grammatical) terms. In other words,invoking such features in the construction of lexical items (be it PRO or anyother item) is just a way of simulating a grammatical requirement via lexicalstipulation.9

The GB approach offers a sounder alternative as it treats PROs properties asthe result of interacting grammatical principles. This feature of the GB analysis

7 See Chomsky (1980: 8): If Coindex does not apply and the embedded clause contains PRO,then we end up with a free variable in LF; an improper representation, not a sentence but anopen sentence.

8 This point is similar to Chomskys (1995) argument against considering Agr as a lexical category.Given that its only features are uninterpretable, a preferable approach, all things being equal, isto take these features as belonging to related true lexical categories.

9 For a discussion of reflexives and bound pronouns in light of this discussion, see Hornstein(2001, 2007).

2.4 Control in GB 11

of control is clearly a desirable one for any theory to have. Any adequatetheory of control should eschew lexically stipulating PROs basic propertiesand specify how grammatical principles interact so that the desired propertiesof PRO emerge. We consider some possibilities below (see also Chapter 3).

Being a grammatical, non-lexical formative specified as [NP ], PRO is infact quite similar to NP-traces (standard traces of A-movement) in GB.10 Whatdistinguishes them is neither their internal structures nor their interpretation,but how they are introduced in the derivation and how they get their indices.11

PRO is inserted at D-structure, but is only coindexed later in the derivation.In contrast, NP-traces receive their indices as they are created in a movementoperation (they must be coindexed with the NP that moves). However, afterPROs get their indices (at S-structure or LF), they become completely indistin-guishable from NP-traces. Notice that, once we take PRO and NP-traces to beindistinguishable at some point(s) in the derivation, we are already very closeto the MTC. We return to this point in Chapters 3 and 4.

The GB account of the distribution of PRO is similarly ambitious. Ratherthan simply stipulate that it appears in the subject position of non-finite clauses,GB strove to derive this fact from the binding theory. The proposal, known asthe PRO theorem, went as follows (see Chomsky 1981). PRO was taken tobe a pronominal anaphor and, as such, subject to both principles A and B.12

Principle A states that an anaphor must be bound in its domain; principle Bthat a pronoun must be free in its domain. Under the assumption that theseprinciples apply within the same domain, they end up imposing contradictoryrequirements on a pronominal anaphor, namely, that it should be both freeand bound in the same domain. The only way for such an expression to meetboth requirements is for it to vacuously satisfy them, i.e., by not meeting thenecessary conditions for these requirements to be enforced. Thus, PRO cannothave a binding domain. Given that the binding domain for an expression wasdefined (in one of its formulations) as the smallest clause within which it isgoverned, then PRO does not have a binding domain if it is ungoverned (if it hasno governor, for instance). Finally, if one takes an Infl head to be a governor if

10 See Chomsky (1977: 82): We may take PRO to be just a base-generated t(x) [trace of x], x avariable; i.e., as a base generated NPx, an NP without an index.

11 See Chomsky (1977: 82): trace and PRO are the same element; they differ only in the way theindex is assigned as a residue of a movement rule in one case, and by a rule of control in theother . . . Note also that PRO is a non-terminal.

12 Saying that PRO was a pronominal anaphor does not imply in the context of GB that it was alexical formative. Traces, for instance, were treated as anaphors (Chomsky 1981) despite theirclearly being non-lexical.


it is finite but not if it is non-finite (to and ing), one then derives the distributionof PRO: it can only appear in the subject position of non-finite clauses.

A side benefit of this reasoning is that it provides an account of why PROmust be phonetically null. Within GB, case theory requires that nominals withphonetic content bear case and case is taken to be assigned under government. IfPRO only appears in ungoverned positions, it cannot be case marked. Therefore,PRO cannot have phonetic content, for otherwise the case filter would beviolated. Once again, this makes PRO very similar to NP-traces. These toooccur in caseless positions and, not surprisingly, are phonetically null.

Notice also that, by taking PRO to be a non-lexical formative, the problemposed by quantified expressions in the standard-theory framework dissolves. Asentence like (8b), for instance, repeated below in (14a), will be associated witha structure along the lines of (14b), where PRO does not have quantificationproperties on its own, but is rather interpreted as a bound variable, as desired.

(14) a. Everyone wants to winb. [Everyone wants [PRO to win]]

At first sight, taking PRO to be a pronominal anaphor also seems to haveother welcome consequences as far as its interpretation is concerned. Oneindeed finds examples of its anaphoric behavior, as illustrated in (15), as wellas examples of its pronominal behavior, as illustrated in (16).13

(15) a. It was expected PRO to shave himselfb. John1 thinks that it was expected PRO1 to shave himselfc. John1s campaign expects PRO1 to shave himselfd. John1 expects PRO1 to win and Bill2 does too (and Bill expects himself to

win, not and Bill expects John to win)e. [The unfortunate]1 expects PRO1 to get a medalf. [Only Churchill]1 remembers PRO1 giving the Blood, Sweat, and Tears

speech

(16) a. It is illegal PRO to park hereb. John1 thinks that Mary said that PRO1 shaving himself is vitalc. John1s friends believe that PRO1 keeping himself under control is vital if

he is to succeedd. John1 thinks that PRO1 getting his resume in order is crucial and Bill does

too (Bill2 thinks that his1/2 getting his resume in order is crucial)e. [The unfortunate]1 believes that PRO1 getting a medal is unlikelyf. Only Churchill remembers that PRO giving the BST speech was

momentous

13 On the properties illustrated in (15) and (16) as well as further data and discussion, see e.g.,Fodor (1975), Williams (1980), Lebeaux (1985), and Higginbotham (1992).


In (15), PRO roughly behaves like a reflexive. In configurational terms,it requires an antecedent (cf. [15a]) which must be local (cf. [15b]) and c-command it (cf. [15c]). On the interpretation side, it only supports a sloppyinterpretation under VP ellipsis (cf. [15d]), a de se reading in sentences suchas (15e) (i.e., it is only felicitous if the unfortunate is conscious of who he isand expects himself to get a medal), and a bound reading when its antecedentis associated with only (that is, [15f] can be paraphrased as Only Churchillis such that he remembers himself giving the BST speech and not as OnlyChurchill remembers that Churchill gave the BST speech). By contrast, in(16) PRO behaves like a pronoun in every respect. Hence, it does not requirean antecedent (cf. [16a]) and, where there is an antecedent, the antecedentneed not be local (cf. [16b]) or c-command it (cf. [16c]). In addition, (16d)allows both strict and sloppy readings, (16e) permits both de se and non-de seinterpretations, and (16f) may be falsified by situations in which people otherthan Churchill recall the import of the BST speech.

Despite appearances, the data in (15) and (16) actually turn out to be quiteproblematic for the specification of PRO as a pronominal anaphor within GB.Notice that PRO displays either properties of reflexives or properties of pro-nouns. But in no case does it display properties of both pronouns and reflexives.Not by coincidence were data such as (15) and (16) handled by two differenttransformations (END and SEND, respectively) within the standard theory.Thus, it makes more sense to assume that PRO is ambiguous between a reflex-ive and a pronoun, than to assume it is a pronominal anaphor. However, thisambiguity thesis completely undermines the PRO theorem, as the theorem cru-cially assumes the existence of an element that is simultaneously a pronounand an anaphor. In turn, if the PRO theorem falls, we are left with no accountof the distribution of PRO.

The requirements of the PRO theorem have one further architectural con-sequence: in order to explain the distributional properties of PRO in terms ofthe PRO theorem, some other component of the grammar must be responsiblefor PROs specific interpretation in a given configuration. This accounts forthe addition of the control module in the GB framework. The control modulerecognizes two types of control: obligatory control (OC), illustrated in (15),and non-obligatory control (NOC), illustrated in (16). In the case of OC, thecontroller is lexically specified as an argument of the embedding control verband, in the case of non-local control, other (rarely specified but frequentlyadverted to) principles come into play.

Notice that this amounts to saying that, like in the earlier treatment in terms ofEND and SEND, OC and NOC are rather distinct types of relations. Importantly,it is tacitly assumed that the control module somehow obliterates the pronominal


specification of PRO in OC constructions, and, conversely, its anaphoric spec-ification in NOC constructions. The problem is not so much that the details ofhow this would be achieved are never spelled out, but that this tacit assumptioncasts suspicion over the initial specification of PRO as a pronominal anaphor.Why should UG provide PRO with such a specification only to see it blottedout later? After all, this does not happen with standard pronouns and anaphors:they must live with the pronominal or anaphoric specifications stated in theirbirth certificate.

One could reply that we should learn to live with the PRO module giventhe nice results we obtain with the PRO theorem concerning the distribution ofPRO. However, this apparent success does not survive closer scrutiny either.The first thing to be noted is that the PRO-theorem account of the distributionalproperties of PRO is intrinsically associated with a specific formulation ofbinding domains, one in which government is essentially the one and onlyrequirement to be satisfied. Recall that all that matters for PRO to vacuouslysatisfy both principles A and B is that it does not have a binding domain.To lack a governor is certainly one way for PRO to be deprived of a bindingdomain. But, if the correct definition of binding domain ends up including otherrequirements, there may be other ways for PRO to lack a domain. Take, forinstance, the definition in (17) (see Chomsky 1981).

(17) is a binding domain for iff is the minimal NP or S containing , inwhich is governed and has a subject accessible to .

This is not the place to review the various reasons for including the notion ofaccessible subject in the definition of binding domain within GB.14 The relevantpoint for our discussion is that, once accessible subjects become part of thedefinition of binding domains, PRO may also lack a domain if it does not havean accessible subject. This in turn undermines the account of the distribution ofPRO in terms of the PRO theorem, as government is no longer the only playeron the field.15 To put it broadly, if binding domains are to be formulated alongthe lines of (17), the account of the distribution of PRO exclusively in terms ofgovernment involves an independent axiom, rather than a theorem.

But the problem is actually worse than these remarks suggest. Recall that acrucial assumption in the PRO-theorem account is that a finite Infl is a governor,but a non-finite Infl is not. From an empirical point of view, this assumption

14 See Lasnik and Uriagereka (1988) for a good discussion of the notion of accessible subject andthe motivations for its inclusion in the definition of binding domain.

15 See Bouchard (1984) for discussion.


is challenged by languages like Brazilian Portuguese, which allow obligatorycontrol into indicative clauses, as we will see in detail in sections 2.5.2.2 and4.4 below. In order to make room for finite control in such languages, the PRO-theorem account would be forced to assume that their finite Infls are optionalgovernors. However, it is not at all obvious how this assumption can be formallyencoded in the system. Given that government is a structural relation, being agovernor cannot be listed as a lexical property for the reasons discussed above.Such a lexical specification would be comparable to saying, for instance, that agiven lexical item is lexically specified as being unable to c-command.16 Thisjust does not make sense. What is required is a structural reason for preventinga non-finite Infl from governing its Spec. But, regardless of the definition ofgovernment one assumes, if a finite Infl can govern its Spec, so should a non-finite Infl, as the two structural configurations are identical. Again, to assumethe opposite would be parallel to saying that, although the configurational Spec-head relation is exactly the same in both cases, a finite Infl head m-commandsits Spec, but a non-finite Infl does not. In short, when details are considered,the distributional properties of PRO do not follow theorematically and it is noteven obvious how to convert the PRO theorem into an axiom, as it is unnaturalto encode structural properties as lexical features or formulate different notionsof government for different lexical items.

Even the apparent benefit of the account of PROs lack of case has an undesir-able consequence. An A-chain in GB must be headed by a case-marked positionunless it is headed by PRO.17 This statement is transparently troublesome. If A-chains are independently subject to a case-licensing requirement (say, Aouns[1979] visibility condition, which requires that -roles be associated with casein order to be visible at LF), why should A-chains headed by PRO be exemptedfrom such a requirement? Notice in particular that, given that chains headedby pro and null operators also required case licensing, PROs lack of phoneticcontent could not be the reason for this exception.

To sum up: despite its laudable ambitions and its improvement over thestandard-theory approach to control, the GB approach has significant empiricaland theoretical problems. On the plus side, treating PRO as a grammaticalformative circumvents the previous problem related to control involving quan-tified expressions, and accounts for why PRO is phonetically null and why (atleast in the case of OC) it needs a grammatical antecedent. On the down side,the account of the distribution of PRO turns out on closer consideration to be

16 See Hornstein, Nunes, and Grohmann (2005) for a discussion of this point.17 See Chomsky (1981: 334ff.) for discussion.


less a theorem than an axiomatic stipulation. Moreover, the assumption thatPRO is a pronominal anaphor leads to empirical problems as the system cannotpredict when PRO behaves like an anaphor and when it behaves like a pronoun.A separate control module must then be added to the theory to specify theinterpretive properties of PRO. Moreover, the construction-specific flavor ofthis new addition to the model is at odds with the general goal of the principles-and-parameters theory of deducing properties of rules and constructions fromthe interaction of more basic features. It is no wonder that the control modulealways felt like an appendix to the model and never occupied a bright spotamong GBs theoretical achievements. The GB take on control was thereforeripe for a minimalist reanalysis.

2.5 Non-movement approaches to control within minimalism

2.5.1 The null-case approachGiven that the addition of the construction-specific control module in GB wasprompted by the problematic assumption that PRO is a pronominal anaphor,one would in principle expect that the abandonment of this assumption shouldalso lead to the abandonment of the control module. However, history and logicare known to frequently go their separate ways. The first minimalist reanalysisof control, outlined in Chomsky and Lasnik (1993), gave up on the accountof PRO in terms of its alleged pronominal-anaphoric nature, but basically leftintact the assumption that the interpretation of PRO required a special modulein the system. Let us consider how the distributional properties of PRO arehandled on this account.

Take the contrast between (18a) and (18b), for instance.

(18) a. John hoped [PRO1 to be elected t1]b. John hoped [PRO1 to appear to t1 [that Bill was innocent]]

From the perspective of GB, PRO cannot occupy a governed position as itwould then meet the requirements for binding theory to apply and would endup violating principle A or principle B. Hence, PRO cannot remain in theobject position of the embedded verb in (18a) or the preposition to in (18b).However, once it moves to the subject position of the infinitival clause, which,by assumption, is ungoverned, it should circumvent the binding violation inboth (18a) and (18b). The ungrammaticality of (18b) is therefore unaccountedfor within GB. Notice that the contrast in (18) mimics the contrast in the ECMconstructions in (19), which can be straightforwardly captured if one assumes

2.5 Non-movement approaches to control 17

that a given expression cannot move from a case-marked position to anothercase-marked position.

(19) a. We expected [John1 to be hired t1]b. We never expected [John1 to appear to t1 [that the job was easy]]

Based on the parallelism between pairs like (18) and (19), Chomsky andLasnik (1993) propose a case-based account of the distributional properties ofPRO under which A-chains headed by PRO are not exceptional as far as caselicensing is concerned, as they were in GB (see section 2.4).18 The gist of theirproposal is that PRO must be licensed by a special kind of case, dubbed nullcase, which is checked by some non-finite Infl heads. Under the assumptionthat the infinitival to in (18) checks null case, movement of PRO is licit in(18a) as it proceeds from a caseless (passives generally do not check case)to a case-checking position, but not in (18b), where it proceeds from a case-checking to another case-checking position. This proposal also extends to thestandard cases regarding the distribution of PRO. Thus, under this view, PROcannot appear in the subject position of finite clauses or in the object positionof a transitive verb, as respectively illustrated in (20), because these are notpositions in which null case is checked.

(20) a. John hoped (that) PRO could eat a bagelb. Bill saw PRO

Notice that some extra assumption must be made in order to capture thestandard contrast between (21) and (22) below, for instance. In other words,the null-case approach must somehow ensure that the infinitival to of controlconstructions can license PRO, but not the infinitival to of ECM or raisingconstructions. The obvious question is how to independently distinguish theto that can check null case in (21) from its siblings in (22), which cannot. Onething is certain. One cannot simply say that these heads are lexically ambiguousin terms of their specification for case checking; otherwise, structures corre-sponding to (22) should be grammatical with the case-checking version of to.

(21) John hopes [PRO to graduate soon]

(22) a. I believe [PRO to be nice]b. It seems [PRO to be nice]

18 The idea of accounting for the distribution of PRO in terms of case finds its origins in Bouchards(1984) proposal that PRO cannot appear in a case-marked position.


Martin (1996, 2001) is the most fully worked out version of the null-caseapproach to the distribution of PRO, which attempts to couch the distinctionbetween (21) and (22) on more solid grounds. Building on Stowells (1982)proposal that control infinitives are tensed whereas ECM and raising infinitivalsare tenseless, Martin proposes that only tensed infinitivals check null case.19

Tying null case to tense has the virtue of rendering it more natural and lessstipulative. Under this perspective, null case would be very similar to nomina-tive case, as both would be checked by a tensed Infl, differing only in termsof their morphological realization. Unfortunately, the proposed independentdiagnostics for distinguishing tensed from tenseless infinitivals fail to yieldthe expected divide between control predicates, on the one hand, and ECMand raising predicates, on the other, as convincingly shown by Wurmbrand(2005).20

Take the contrasts between the infinitival complements of the control verbdecide and the ECM verb believe in (23)(26) (from Wurmbrand 2005), forexample.

(23) a. At 6, Leo decided to sing in the shower right thenb. At 6, Leo believed Bill to sing in the shower right then

(24) a. Leo decided yesterday to leave tomorrowb. John believes/believed Mary to be pregnant

(25) a. Leo decided [[to leave] [which was/is true]]b. Leo believes [[John to be smart] [which is true]]

(26) a. Leo doesnt want John to sing in the shower, but he decided to, anywayb. Leo believes John to be honest and she believes Frank to, as well

The contrasts above are supposed to show that the control infinitival is tensedas it is compatible with eventive predicates (cf. [23a]), triggers a future reading(cf. [24a]), requires an irrealis interpretation (that is, the truth of the complementis left unspecified at the time of the utterance; cf. [25a]), and licenses VPellipsis (cf. [26a]). Conversely, the ECM/raising infinitival clauses are takento be tenseless as they are incompatible with eventive predicates (cf. [23b]),require a simultaneous interpretation with respect to the embedding clause(cf. [24b]),21 allow a realis interpretation (cf. [25b]), and do not license VPellipsis (cf. [26b]).

19 See also Boskovic (1997) for relevant discussion.20 For further discussion and arguments against null case and its ties to tense, see also Landau

(2000), Pires (2001, 2006), Baltin and Barrett (2002), and Hornstein (2003).21 See Hornstein (1990) for a discussion of this interpretation in the context of sequence-of-tense

constructions.


The above paradigm does indeed distinguish decide, a control verb, frombelieve, an ECM verb. The problem, as Wurmbrand (2005) shows, is that thecriteria do not generalize to other control and ECM/raising cases. For instance,the control verb claim does not license eventive predicates (cf. [27a] below)and allows a realis interpretation for its complement (cf. [27b]), whereas thecontrol verb manage does not trigger a future reading (cf. [27c]). In turn,the infinitival complement of the ECM verb expect is compatible with aneventive predicate (cf. [28a]), does not permit a realis interpretation (cf. [28b]),and allows a non-simultaneous interpretation (cf. [28c]).

(27) a. At 6, Leo claimed to sing in the shower right thenb. Leo claimed [[to be a king], which was true]c. John managed to bring his toys tomorrow

(28) a. The bridge is expected to collapse tomorrowb. The train is expected [[to arrive late tomorrow] [which is true]]c. The printer is expected to work again tomorrow

Wurmbrand (2005) also reviews the VP ellipsis data and observes that thedata that purport to demonstrate a distinction between control (where it isallowed) and raising (where it is prohibited) are subject to substantial speakervariation (when the contrast exists at all). Besides, the clear acceptability ofraising examples such as the ones in (29) indicates that the licensing of VPellipsis fails to cleanly distinguish control from raising.

(29) a. The tower started to fall down and the church began to as wellb. John expects the printer to break down whereas Peter expects the copier toc. They say that Mary doesnt know French but she seems to

The above arguments, which decouple tense properties from control infini-tivals, are seconded by the observation that PRO may also occur in gerundivesubject positions, despite the fact that gerunds are generally analyzed as nottensed (see Stowell 1982; Pires 2001, 2006). This is illustrated in (30) below,where the gerund licenses PRO but not the temporal adverb.

(30) a. John hated [PRO eating turnips (tomorrow)]b. John preferred [PRO eating turnips (tomorrow)]

The overall conclusion one reaches is that, whatever tense properties non-finite clauses have, they do not seem to be useful for distinguishing raising fromcontrol configurations. There are surely differences between raising and controlcomplements, but this varies across verbs and there is no apparent systematicway to distinguish the two classes using the tense diagnostics mentioned


above. Thus, although conceptually appealing, the attempt to analyze null caseas similar to nominative by associating it to a form of tense ends up failing.

This is really bad news. Once the distribution of PRO cannot be reduced toa [tense] feature of T, null case finds no independent motivation within thesystem and follows from nothing but the attested distribution of PRO. And thepicture is not very glamorous. In order to work, the null-case approach requiresthree stipulations: (i) PRO has no phonetic content; (ii) null case must beassigned to PRO; and (iii) only PRO can bear null case. These three stipulationstrack but do not explain the facts under discussion. In other words, despite itsexplanatory aspirations, it seems fair to say that the null-case approach amountsto stipulating that PRO appears where it does and that it has the phoneticproperties it has.

What of PROs interpretive properties? Here there is some good news. Withthe PRO theorem abandoned, PRO can be treated as ambiguous, a null reflexivein some contexts (OC cases) and a null pronoun in others (NOC cases). It isthen possible to reduce the interpretive properties of PRO to the interpretiveproperties of pronouns and reflexives. For example, that OC PRO requires alocal, c-commanding antecedent follows from its being subject to principle Aof the binding theory (or whatever substitutes for principle A). The fact thatNOC PRO does not need an antecedent follows its being pronominal.

Given such a reduction, what remains to be determined is why OC and NOCPROs distribute as they do, i.e., why reflexive PRO appears in OC contextsand pronominal PRO in NOC contexts. One can, of course, stipulate thatcertain predicates select for OC and so for reflexive-like PROs, while others donot. However, it is not clear how this is to be implemented grammatically (seeChapters 6 and 7 below). First, it is not clear how selection of embedded subjectsby matrix verbs (so-called control predicates) is to be stated. If selection is ahead-to-head relation, then OC is not an obvious case of selection. Second,adjunct control seems to pattern like OC and, on the standard assumption thatpredicates can select complements but not adjuncts, then adjunct control isexpected to be NOC, contrary to fact. These are issues that we revisit in laterchapters. What is worth noting here is that simply reducing OC to somethinglike principle A and NOC to something like principle B does not by itself sufficeto account for the interpretive properties of OC and NOC configurations.

2.5.2 The Agree approachLet us now consider Landaus (1999, 2000, 2004) alternative approach to con-trol. Like the null-case approach reviewed in the previous section, Landautakes the existence of PRO for granted but, unlike proponents of the null-case


approach, he takes PRO to bear regular case like any other DP. In addition tothis take on case, three other aspects of Landaus approach stand out: (i) thespecial attention given to partial-control constructions; (ii) the dependenceof obligatory control on the postulation of certain features and feature specifi-cations; (iii) the interpretation of PRO mediated by (a version of) Chomskys(2000, 2001) Agree operation.

Let us examine each of these major aspects of Landaus system, leaving thediscussion of whether or not PRO bears regular case to section 5.4.2 below.22

2.5.2.1 The relevance of partial controlPartial control refers to control constructions where an embedded predicatemust take a (semantically) plural subject, but the antecedent of the controlleeis (semantically) singular, as illustrated in (31).

(31) The chair hoped [PRO to gather/meet at 6/to apply together for the grant]

In (31), the matrix subject is understood as a member of the set of peo-ple denoted by the embedded subject. Assuming that this interpretive factshows that controller and controllee are not identical, Landau takes partial-control constructions to be a strong argument for a PRO-based account ofcontrol. According to him, the mismatch in interpretation between PRO andits antecedent results from PRO being independently specified for the semanticfeature mereology, which characterizes group names (for instance, committeeis [+Mer], while chair is [Mer]), as illustrated in (32).

(32) The chair[Mer] hoped [PRO[+Mer] to gather/meet at 6/to apply together forthe grant]

It is a great merit of Landaus work to have shown that partial control is indeedan instance of obligatory control (the controllee requires a local c-commandingantecedent, triggers sloppy readings under ellipsis, and enforces de se readings,for instance), and to have provided a very detailed description of the types ofpredicates that allow partial control. Landau argues that a tensed infinitive suchas the complement of desiderative verbs licenses it, but an untensed infinitivesuch as the complement of implicative verbs does not, as illustrated by thecontrast between (31) and (33) (see section 2.5.2.2 below for details).

22 Here we will primarily focus on Landau (2004)s analysis of obligatory control, which he takesto replace his older treatment (Landau 1999, 2000). For discussion of the limitations of hisprevious treatment, see Hornstein (2003) and Landau (2007) for a rejoinder.


(33) The chair managed [PRO to gather/meet at 6/to apply together for the grant]

It is fair to say that, after Landaus work, partial control came to be partof the empirical basis that any approach to obligatory control must take intoconsideration. However, the amount of ad hoc machinery required to accountfor partial control in Landaus system, as we will see below in section 2.5.2.3,ends up undermining the initial appeal that a PRO-based theory appears to have.And there are empirical problems, as well. As observed by Hornstein (2003),it is not the case that any predicate that selects a plural subject licenses partialcontrol, as shown in (34).

(34) a. They sang alike/were mutually supportingb. John hoped/wants [PRO to sing alike/to be mutually supporting]

Notice that the matrix predicate of (34b) is of the type that licenses partialcontrol (cf. [31]). So (34b) shows that partial control must in part be deter-mined by properties of the embedded predicate. In fact, Hornstein suggeststhat what seems to distinguish the predicates that support partial control fromthe ones that do not is that the former can select a commitative PP, as shownin (35).

(35) a. The chair met/gathered/applied together for the grant with Billb. The chair sang alike/is mutually supporting with Bill

The data in (36)(37) further show that being compatible with a commitativePP is not sufficient for partial control to be licensed: the commitative must beselected.

(36) a. The chair met/gathered/applied together for the grant (with Bill)b. The chair left/went out (with Bill)c. The committee left/went out

(37) The chair preferred [PRO to leave/go out at 6](exhaustive control: OK; partial control: )

Example (36b) shows that, as opposed to what happens with meet/ gather/applytogether in (36a), the commitative associated with leave/go out is not selected.In turn, (36c) shows that a [+Mer] noun can be the subject of leave/go out.Now, given that in Landaus system PRO can always be intrinsically specified as[+Mer], one would expect that a sentence such as (37), whose matrix predicateis of the type that licenses partial control, should allow a partial-control readingwith a [+Mer] PRO. But this does not happen. Example (37) only has anexhaustive control reading.


The fact that the availability of partial control is contingent on there beinga predicate that selects a commitative complement suggests that, rather thaninvolving a plural subject, partial control may in fact involve the licensing ofa null commitative argument in a standard (exhaustive) obligatory-controlconstruction. That is, a sentence such as (38a) should actually be representedas in (38b) (still keeping PRO for purposes of discussion), where pro is a nullcommitative argument.

(38) a. The chair preferred to meet at 6b. [The chair]i hoped [PROi to[+tense] meet prok at 3]

Here is not the place for us to pursue the suggestion encapsulated in (38b) (seesection 5.6.1 below for discussion). What is relevant for our current purposes isto point out that, in Landaus system, the availability of partial control shouldbe quite free once the tense requirements on the infinitive are satisfied. It isindeed quite mysterious in his system why partial control should depend on thepotential licensing of commitative arguments within the infinitival clause. Andif partial control turns out to be more related to the licensing of null commitativearguments, whatever accounts for exhaustive control should also cover partialcontrol. In other words, if something along the lines of (38b) is on the righttrack, partial control does not intrinsically favor a PRO-based approach and weare back to the original question of what the best account of the null embeddedsubject of (38a) is (see section 5.6.1 below for a suggestion of how partialcontrol can be analyzed under the MTC).

2.5.2.2 [Tense] and [Agr] features and finite controlThe second major aspect of Landaus system is the specific typology of controlconfigurations involving both non-finite and finite clauses it establishes.Following a venerable tradition, Landau assumes that the local environment ofthe embedded subject must provide all the necessary information to determinewhether it must, can, or cannot be PRO. In particular, Landau takes the rel-evant local licensing features to be (semantic) [T(ense)] and (morphological)[Agr(eement)]. Where Landau departs from previous accounts is in the waythese features conspire to determine the nature of control, as shown in (39)(from Landau 2004: 840).23

23 EC and PC in (39) stand for exhaustive and partial control respectively. C(ontrol)-subjunctivesand F(ree)-subjunctives are distinct in that only the former necessarily require an obligatory-control interpretation of their subjects. For purposes of exposition, below we use I for the tensehead T in order to distinguish it from the tense feature [T].


(39) Obligatory control No control

EC-infinitive Balkan C-subjunctive

Hebrew3rd-personsubjunctive

PC-infinitive BalkanF-subjunctive

indicative

I0 [T, Agr] [T, +Agr] [+T, +Agr] [+T, Agr] [+T, +Agr] [+T, +Agr]C0 [T] [T] [+T, +Agr] [+T, (+Agr)] [+T, +Agr]

Consider the infinitives in (39), for instance. As mentioned in section 2.5.2.1,Landau has argued that the essential difference between an infinitival that allowspartial control and one that disallows it is its tense properties: an infinitivalI allows both exhaustive and partial control if specified as [+T], but onlyexhaustive control if specified as [T]. This difference is meant to capturethe fact that the infinitival clauses that allow partial control can be temporallyindependent from the matrix clause, as illustrated in (40) below. Given that thetense properties of I are predicted by the selecting predicate and that selectionis a local relation, the [T] features of I are accordingly replicated on C in (39).Thus, a verb like hope, for instance, selects a CP headed by C[+T], which inturn selects an IP headed by I[+T].

(40) a. Yesterday John hoped to travel tomorrowb. Yesterday John managed to travel tomorrow

As Landau observes, the basic intuition underlying the typology in (39) isthat obligatory-control configurations do not form a natural class; they are infact the complement subset of the natural class of non-controlled environments.Putting aside the case of Hebrew third-person subjunctives for the moment, thegeneralization is that if I is positively specified for both [T] and [Agr], it doesnot trigger obligatory control. On the other hand, a single negative specificationfor [T] or [Agr] ([+T, Agr] or [T, +Agr]) or a negative specification on both([T, Agr]) will necessarily lead to obligatory control. In sum, obligatorycontrol is the elsewhere case.

Given this feature distribution, it follows that indicative complements shouldnot display obligatory control. As Landau (2004: 849850) puts it, the onlygeneralization in this domain that appears to be universal is the incompat-ibility of indicative clauses with OC. Anything else is possible, under cer-tain circumstances. However, this generalization is falsified by referential(i.e., non-expletive, non-arbitrary) null subjects in (colloquial) Brazilian Por-tuguese. As extensively argued by Ferreira (2000, 2004, 2009) and Rodrigues(2002, 2004), null subjects in Brazilian Portuguese show all the diagnostics of


obligatory control. Take the Brazilian Portuguese sentences in (41)(45), forinstance.24

(41) Comprou um carro novoBought a car newShe/he bought a new car

(42) [[o Joao] disse que [o pai d[o Pedro]] acha que vaiThe Joao said that the father of-the Pedro thinks that goes

ser promovido]be promotedJoaoi said that [Pedrojs father]k thinks that hek/ i/ j/ l is going to bepromoted

(43) So o Joao acha que vai ganhar a corridaOnly the Joao thinks that goes win the raceOnly Joao is an x such that x thinks that x will win the raceNOT: Only Joao is an x such that x thinks that he, Joao, will win the race

(44) O Joao esta achando que vai ganhar a corrida e oThe Joao is thinking that goes win the race and the

Pedro tambem estaPedro too isJoao thinks that hes going to win the race and Pedro does, too (think thathe, Pedro, is going to win the race)

(45) O infeliz acha que devia receber uma medalhaThe unfortunate thinks that should receive a medalThe unfortunate thinks that he himself should receive a medal

Example (41) shows that null subjects in Brazilian Portuguese require anantecedent25 and (42), that the antecedent must be the closest c-commandingDP. As for interpretation matters, a null subject in Brazilian Portuguese isinterpreted as a bound variable when its antecedent is an only-DP (cf. [43]); itobligatorily triggers sloppy reading under ellipsis (cf. [44]); and it only admitsa de se reading in sentences such as (45). Importantly, in all the sentences of(41)(45), the null subject displays the diagnostics of obligatory control despitethe fact that it is within a standard indicative clause.

The existence of finite control into indicative complements in BrazilianPortuguese therefore presents prima facie problems for the typology proposed

24 See Ferreira (2000, 2004, 2009) and Rodrigues (2002, 2004) for additional tests.25 Referential null subjects in matrix clauses in Brazilian Portuguese can only be licensed as

instances of topic-drop (see Ferreira 2000, Modesto 2000, and Rodrigues 2004 for relevantdiscussion).


by Landau.26 Below we discuss the implications of this empirical fact withinLandaus Agree-based approach.

2.5.2.3 Determining the interpretation of obligatorily controlledPRO via Agree

In addition to the features [T] and [Agr] to be hosted by C and I, Landau (2004:841) also proposes that DPs must be featurally specified as to whether or notthey support independent reference ([R]): lexical DPs and pro are specifiedas [+R] and PRO as [R]. According to Landau (p. 841), [b]oth values on[R] are interpretable, when occurring on nominal phrases. However, the [R]feature makes PRO a potential goal for agreement, for this feature acts as aninstruction to coindex the -features of PRO with those of an antecedent; Agreeis a way of achieving that (p. 843). The feature [R] is also assigned to somefunctional categories, according to the rule in (46).

(46) R-assignment rule (Landau 2004: 842)For X0[T, Agr] {I0, C0, . . .}: [+R]/X0[__], if = = + [R]/elsewhere

Given these assumptions, let us consider the derivation of an exhaustive controlconstruction such as (47) in Landaus system, which is given in (48).

(47) John managed to fix the car

(48) Agree[DP I2 [ . . . tDP . . . [CP C[T] [IP PRO[R] I1[T, Agr, R] [tPRO . . . ]]]]]

Agree Agree Agree

Agreement between I1 and PRO in (48) deletes I1s [Agr] and [R] features.Agreement between C and I1 then deletes Cs [T] feature. Finally, afteragreeing with the matrix subject, I2 agrees with PRO, coindexing their featuresand licensing PROs [R] feature.27

26 See Rodrigues (2004) for arguments that Finnish may also allow obligatory control into finiteindicative clauses.

27 The only relevant difference between (48) and a typical obligatory-control subjunctive in Greeksuch as (i) in Landaus system is that in the latter, I1 has overt agreement morphology ([+Agr]),rather than abstract agreement ([Agr]), as represented in (ii).(i) Greek (Terzi 1997):

I Maria1 prospathise Pro1/2 na divasithe Maria tried.3SG PRT read.3SGMaria tried to read


In turn, a partial-control construction like (49) is to be derived along the linesof (50).

(49) The chair hoped to meet at 6(50) Agree

[DP I2[+T, +Agr, +R] [ . . . tDP . . . [CP C[+T, +Agr, +R] [IP PRO[R]Agree Agree

I1[+T, Agr, R] [tPRO . . . ]]]]]Agree

As before, agreement between I1 and PRO deletes I1s [Agr] and [R] fea-tures. Agreement between C and I1 now deletes Cs [+T] and [+Agr] features,but not its [+R] feature, for it mismatches the [R] feature of I1. C then checksits [+R] feature with I2. Notice that I2 agrees with C and not with PRO, whichraises the question of how PRO can license its [R] feature. According toLandau (p. 845), this feature gets licensed in virtue of I2 agreeing with C,which in turn is coindexed with PRO via I1. Furthermore, Landau assumes(p. 849) that if I2 and PRO do not agree directly, their [Mer] features need notmatch. If I2 is specified as [Mer] and PRO is inherently specified as [+Mer],a partial-control effect will arise.

Let us finally consider the last type of obligatory-control configuration listedin (39): Hebrew third-person subjunctives.28 Given the feature specification forHebrew subjunctives in (39), the derivation of a sentence such as (51) proceedsas in (52).

(51) Hebrew (Landau 2004)Gili hivtiax [se- eci yitnaheg yafe]Gil promised that will-behave.3SG.M wellGil promised to behave

(52) Agree[DP I2[+R] [ . . . tDP . . . [CP C[+T, +Agr, +R] [IP PRO[R] I1[+T, +Agr, +R] [tPRO . . . ]]]]]

Agree Agree Agree

(ii) Agree[DP I2 [ . . . tDP . . . [CP C[T] [IP PRO[R] I1[T, +Agr, R] [tPRO]]]]]

Agree Agree Agree

28 Landau (2004: 815, 846) attributes the lack of a derivation with an uncontrolled third-personpro in Hebrew subjunctives to the non-existence of referential third-person pro in the language.More specifically, he assumes Shlonskys (1997) proposal that third-person pros in Hebrew arenull Num heads and, because they are null, they cannot support a third-person feature hostedby a higher D-head.


Agreement between I1 and PRO in (52) checks the [+Agr] feature of I1, butnot its [+R] feature, which mismatches the [R] feature of PRO. Agreementbetween C and I1 then checks all of the features of C and the [+R] feature of I1.If the matrix I2 had a [R] feature, the remaining unchecked [R] feature ofPRO would be licensed by agreement with I2. However, in (52) I2 is specifiedas [+R]. The [R] feature of PRO must then be indirectly licensed in virtue ofthe agreement relations between I2 and C, between C and I1, and between I1and PRO.

As the reader can easily check, the feature specifications and computationsproposed above are such that they track, but do not explain, the distributionand interpretation of PRO. Landau (p. 842) in fact acknowledges that his R-assignment rule is an honest stipulation, which played the role of case inprevious models. Unfortunately, if the distribution and interpretation of PROis to rest on a stipulation, calling it honest does not make the analysis lessstipulative. In other words, it is subject to the same criticism made to thenull-case approach: the distribution and interpretation of PRO ends up beingstipulated under the guise of lexical features.

It is also worth pointing out that, under the label Agree, Landaus proposalactually groups different kinds of relations, which do not obviously form anatural class. Thus, in addition to the familiar valuation procedure involving a[interpretable] and a [+interpretable] feature of Chomsky (2001), the Agreeoperation assumed by Landau encompasses three other types of relations. First,it admits relations between two [interpretable] features such as the agreementbetween C and I1 with respect to [Agr] features in (50) or the agreement betweenI2 and C in (52) with respect to [R] features. According to Landau (p. 849),[t]he fact that C bears [+Agr] does not stop this feature from entering Agreewith [Agr] of I; recall that [+Agr] on C represents abstract [Agr] to beginwith (in most cases), thus [Agr] on both heads is semantically uninterpretableand phonologically null. That may be so, but the resort to features which aremotivated neither in LF nor in PF terms not only is completely at odds with coreminimalist assumptions, but also reinforces the impression that these featuresare only redescribing the facts to be explained.

The second type of relations encompassed by Landaus version of Agreeinclude coindexing relations such as the agreement between I2 and PRO in(48) to license PROs [R] feature (which was assumed to be a [+interpretable]feature, as mentioned above). Finally, it also includes composite-coindexingrelations such as the licensing of the [R] feature of PRO in (50) and (52),which involves the conjunction of three basic agreement relations: between I2and C, between C and I1, and between I1 and PRO. Even if we put aside the fact


that coindexing and feature valuation/deletion seem to be of different nature,it is not at all trivial to explain how the composition of the three agreementrelations mentioned should result in coindexing. Recall that, in (50) and (52),I2 and C agree with respect to [R] as do I1 and PRO, but C and I1 agree withrespect to [Agr]. In virtue of these two agreement relations, PRO agrees with I1through some kind of transitivity assumption. However, it is worth asking howtransitivity arises given that the agreement relations computed do not target thesame type of feature. Note that, if A is taller than B and B is fatter than C, thenone can conclude nothing regarding As height or weight as regards C. However,for the account above to work, we must assume that this logic is overturned whencertain feature sets are involved, which in turn brings the obvious minimalistquestion: why are these features endowed with their alleged properties? Thisshows that the proposed transitivity in the account of (51) does not follow asa point of logic, but is rather a stipulated feature in Landaus system. Thus,the proposed composite-coindexing relations should be subject to the sameskepticism we accord the Barriers approach to A-movement, which licensesA-traces by resorting to a chain coindexing mechanism combining Spec-headagreement with head-to-head government (see Chomsky 1986a, section 11).

The non-explanatory nature of the proposal is further highlighted whenLandaus account of (51) is examined in light of his take on the impossibility ofPRO in indicative clauses. As we saw in (39), the feature specification proposedfor indicatives involved the features [+T] and [+Agr] for I and no features forC. The reason for C not to be associated with [T] features is that the tense valueof I is completely independent from the matrix clause. Furthermore, sinceLandau (p. 840) assumes that the presence of [Agr] on C is parasitic on [+T], ifindicative C does not have [+T], it cannot have [Agr] either. Finally, if it is notspecified for both features, it cannot be associated with an [R] feature, accordingto the R-assignment rule in (46). That being so, Landau (p. 843) claims thatthe reason why PRO cannot be licensed in the indicative configuration in (53)below (Landaus [40b]) is that Agree fails due to a feature mismatch in theR value between I and PRO. Thus, indicative clauses with independent tenseuniversally do not display OC.

(53) [DP I2 [ . . . tDP . . . [CP C [IP I1[+T, +Agr, +R] [VP PRO[R] . . . ]]]]]Agree

Agree

This specific claim now introduces an additional aspect of composite-agreement relations: feature mismatch is taken to cause a derivational crashunder direct agreement, like the relation between I1[+R] and PRO[R] in (53),


but not under composite agreement, like the relation among I2[+R]-C[+R]-I1[R]-PRO[R] in (50). Putting aside the fact that no motivation was providedfor why these two instantiations of Agree should yield opposite results, itis important to point out that this stipulated aspect of composite agreementleads to overgeneration. Notice that the feature mismatch at the derivationalstep depicted in (53), that is, before PRO moves, cannot be the reason forthe derivation to crash. As we saw in the derivation proposed by Landau forHebrew subjunctive control in (52), mismatch in the values for [R] by itself isnot a problem if the features can be licensed later on in the derivation. Considerfor instance the structure in (54), which depicts the movement of PRO in (53).

(54) [DP I2 [ . . . tDP . . . [CP C [IP PRO[R] I1[+T, +Agr, +R] [VP tPRO . . . ]]]]]Agree Agree

If I2 in (54) is specified as [R], it will be able to agree with PRO, but the [+R]feature of I1 will remain unchecked, causing the derivation to crash. Suppose,by contrast, that I2 is specified as [+R]. As such, it should be able to agree withI1, as represented in (55), checking the [+R] feature of the latter.(55) Agree

[DP I2[+R] [ . . . tDP . . . [CP C [IP PRO[R] I1[+T, +Agr, +R] [VP tPRO . . . ]]]]]Agree Agree

What about the [R] feature of PRO? Recall from (50) and (52) that PROcan be indirectly licensed by a chain of agreement relations. In (52), forexample, its [R] feature is taken to be licensed in virtue of PROs havingagreed with I1, which had agreed with C, which in turn had agreed with I2.That being so, there should be no reason for PRO not to get licensed in (55) viaa composite-agreement relation. That is, its [R] feature should be licensedonce PRO has agreed with I1, which agrees with I2. Crucially, compositeagreement is assumed to be oblivious to feature mismatch. In other words,once the composite-agreement relations proposed by Landau are assumed,finite control into indicatives becomes freely available.

In fairness, Landau (2004: 846847) seems to assume that the [+R] featureof I1 cannot be checked by a probe higher than C: We still account for the factthat indicative complements in Hebrew do not display OC. In a configurationlike (40b) [= (53) above], as opposed to (43b) [= (52) above], the [+R] featureof I remains unchecked as no corresponding feature exists on the indicativeC. Notice, however, that C does not prevent a higher probe from agreeing withthe embedded subject in (48), for instance. Given that the embedded subjectand the embedded I are equidistant (see Chomsky 1995), it does not seemplausible to exclude the checking of the [+R] feature of I1 in (55) based on the


intervention of C. Notice also that C in (55) has no features that could block anagreement relation between a higher probe and I1.

A phase-based approach to this conundrum is of no help either. Landau(2004, footnote 26) claims that [a]lthough not at the edge of its phase, PRO isvisible to Agree from the outside since its (- and R-) features are interpretable(hence, never erased). Note, however, that in Landaus system the - and R-features of PRO are only valued after agreement. Thus, it is plausible to assumethat spell-out/transfer must be halted until PRO has its features valued and, ifthis is so, we are back to the technical question of why I1 in (55) cannot bechecked by the matrix probe if spell-out/transfer is on hold. Of course, one mayattempt to specify the inner workings of spell-out/transfer in such a way thatPRO becomes immune to spell-out/transfer at the relevant derivational step, butnot I1. But that would only add to an already loaded machinery, without actuallyshedding light on the discussion. Still, such an attempt would require furthercomplications. Recall that, under composite-agreement relations in Landaussystem (cf. [52] for Hebrew and [ii] in footnote 27 for Balkan languages), thehigher probe agrees with the embedded C, which is also coindexed with PROvia I (Landau 2004: 845). This in turn indicates that the embedded I muststill be available to the computation at the derivational step where PRO is tohave its R-feature licensed.

To wrap up: if composite relations must be assumed in order to accountfor Hebrew subjunctive control, control into indicative clauses becomes freelyallowed. Although this may be good news for languages such as BrazilianPortuguese, as discussed in section 2.5.2.2, it is certainly unwelcome for mostlanguages.

2.5.2.4 Simplifying Landaus calculus of controlLet us examine what the relevant property of Brazilian Portuguese indicativeclauses is that triggers obligatory control. Ferreira (2000, 2004, 2009) proposesthat finite Ts in Brazilian Portuguese are ambiguous in being associated witheither a complete or an incomplete set of -features and that obligatory controlis licensed in clauses with a -incomplete T.29 Nunes (2007, 2008a) reinterpretsFerreiras proposal in terms of the presence or absence of the feature [person] inT. He observes that the verbal-agreement paradigm of finite clauses in BrazilianPortuguese is such that the only inflection that overtly encodes both number and

29 Ferreira (2000, 2004, 2009) (as well as Rodrigues 2002, 2004) in fact analyzes null subjects inBrazilian Portuguese in terms of the MTC. Thus, in his system a -incomplete T does not valuethe case feature of the subject of its clause, which can then undergo A-movement to the matrixclause. We will leav

Cedric Boeckx, Norbert Hornstein, Jairo Nunes Control as Movement 2010

Documents

Transcript of Cedric Boeckx, Norbert Hornstein, Jairo Nunes Control as Movement 2010