CONNECTED SUBSTITUTES AND INVERTIBILITY OF DEMAND BY Steven Berry, Amit Gandhi, and Philip Haile COWLES FOUNDATION PAPER NO. 1391 COWLES FOUNDATION FOR RESEARCH IN ECONOMICS YALE UNIVERSITY Box 208281 New Haven, Connecticut 06520-8281 2013 http://cowles.econ.yale.edu/ Econometrica,Vol.81,No.5(September,2013),2087–2111 CONNECTEDSUBSTITUTESANDINVERTIBILITYOFDEMAND BYSTEVENBERRY,AMITGANDHI,ANDPHILIPHAILE1 Weconsidertheinvertibility(injectivity)ofanonparametricnonseparabledemand system.Invertibilityofdemandisimportantinseveralcontexts,includingidentification ofdemand,estimationofdemand,testingofrevealedpreference,andeconomicthe- oryexploitingexistenceofaninversedemandfunctionor(inanexchangeeconomy) uniqueness of Walrasian equilibrium prices. We introduce the notion of “connected substitutes” and show thatthis structure is sufficientfor invertibility. Theconnected substitutesconditionsrequireweaksubstitutionbetweenallgoodsandsufficientstrict substitutiontonecessitatetreatingtheminasingledemandsystem.Theconnectedsub- stitutesconditionshavetransparenteconomicinterpretation,areeasilychecked,and aresatisfiedinmanystandardmodels.Theyneedonlyholdundersometransformation ofdemandandcanaccommodatemanymodelsinwhichgoodsarecomplements.They allowonetoshowinvertibilitywithoutstrictgrosssubstitutes,functionalformrestric- tions,smoothnessassumptions,orstrongdomainrestrictions.Whentherestrictionto weaksubstitutesismaintained,oursufficientconditionsarealso“nearlynecessary”for evenlocalinvertibility. KEYWORDS:Univalence,injectivity,globalinverse,weaksubstitutes,complements. 1. INTRODUCTION WECONSIDERTHEINVERTIBILITY(INJECTIVITY)ofanonparametricnonsep- arabledemandsystem.Invertibilityofdemandisimportantinseveraltheoret- icalandappliedcontexts,includingidentificationofdemand,estimationofde- mandsystems,testingofrevealedpreference,andeconomictheoryexploiting existenceofaninversedemandfunctionor(inanexchangeeconomy)unique- ness of Walrasian equilibrium prices. We introduce the notion of “connected substitutes”andshowthatthisstructureissufficientforinvertibility. Weconsiderageneralsettinginwhichdemandforgoods1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J ischarac- terizedby (cid:2) (cid:3) (1) σ(x)= σ (x)(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)σ (x) :X ⊆RJ →RJ(cid:2) 1 J where x=(x (cid:2)(cid:3)(cid:3)(cid:3)(cid:2)x ) is a vector of demand shifters. All other arguments of 1 J the demand system are held fixed. This setup nests many special cases of in- terest. Points σ(x) might represent vectors of market shares, quantities de- manded, choice probabilities, or expenditure shares. The demand shifters x 1ThispapercombinesandexpandsonselectedmaterialfirstexploredinGandhi(2008)and BerryandHaile(2009).WehavebenefittedfromhelpfulcommentsfromJean-MarcRobin,Don Brown,thereferees,andparticipantsinseminarsatLSE,UCL,Wisconsin,Yale,the2011UCL workshopon“ConsumerBehaviorandWelfareMeasurement,”andthe2011“Econometricsof Demand”conferenceatMIT.AdamKaporprovidedcapableresearchassistance.Financialsup- portfromtheNationalScienceFoundationisgratefullyacknowledged. ©2013TheEconometricSociety DOI:10.3982/ECTA10135 2088 S.BERRY,A.GANDHI,ANDP.HAILE might be prices, qualities, unobserved characteristics of the goods, or latent preferenceshocks.SeveralexamplesinSection2illustrate. The connected substitutes structure involves two conditions. First, goods must be “weak substitutes” in the sense that, all else equal, an increase in x j (e.g.,afallinj’sprice)weaklylowersdemandforallothergoods.Second,we require “connected strict substitution”—roughly, sufficient strict substitution between goods to require treating them in one demand system. These con- ditions have transparent economic interpretation and are easily confirmed in many standard models. They need only hold under some transformation of thedemandsystemandcanaccommodatemanysettingswithcomplementary goods. The connected substitutes conditions allow us to show invertibility without thefunctionalformrestrictions,smoothnessassumptions,orstrongdomainre- strictionsreliedonpreviously.Wealsoprovideapartialnecessityresultbycon- sidering the special case of differentiable demand. There we show that when the weak substitutes condition is maintained, connected strict substitution is necessary for nonsingularity of the Jacobian matrix. Thus, given weak substi- tutes,connectedsubstitutesissufficientforglobalinvertibilityand“nearlynec- essary” for even local invertibility. A corollary to our two theorems is a new globalinversefunctiontheorem. Important to our approach is explicit treatment of a “good 0” whose “de- mand”isdefinedbytheidentity (cid:4)J (2) σ (x)=1− σ (x)(cid:3) 0 j j=1 Formally,(2)issimplyadefinitionofthefunctionσ (x).Itsinterpretationwill 0 vary with the application. When demand is expressed in shares (e.g., choice probabilitiesormarketshares),good 0 mightbea“real”good—forexample, anumeraire good,an“outside good,”ora goodrelative to which utilities are normalized.Theidentity(2)willthenfollowfromthefactthatsharessumto1. Inotherapplications,good 0 willbeapurelyartificialnotion introducedonly as a technical device (see the examples below). This can be useful even when anoutsidegoodisalsomodeled(seeAppendixC). Itisclearfrom(2)that(1)characterizesthefulldemandsystemevenwhen good 0 is a real good. Nonetheless, explicitly accounting for the demand for good0inthiscasesimplifiesimpositionoftheconnectedsubstitutesstructure on all goods. When good 0 is an artificial good, including it in the connected substitutesconditionsprovesusefulaswell.Aswillbeclearbelow,itstrength- ens the weak substitutes requirement in a natural way while weakening the requirementofconnectedstrictsubstitution. AlsoimportanttoourapproachisapotentialdistinctionbetweenthesetX andthesubsetofthisdomainonwhichinjectivityofσ isinquestion.Ademand system generally will not be injective at points mapping to zero demand for CONNECTEDSUBSTITUTES 2089 somegoodj.Forexample,raisinggoodj’sprice(loweringx )atsuchapoint j typically will not change any good’s demand. So when considering conditions ensuringinjectivity,itisnaturaltorestrictattentiontotheset (cid:5) (cid:6) X˜ = x∈X :σ (x)>0∀j>0 j or even to a strict subset of X˜—for example, considering only values present in a given data set, or only positive prices even though σ is defined on all of RJ (e.g., multinomial logit or probit). Allowing such possibilities, we consider injectivityofσ onanyset X∗⊆X (typicallyX∗⊆X˜).Ratherthanstartingfromtherestrictionofσ toX∗,how- ever, it proves helpful to allow X∗ (cid:7)=X. For example, one can impose useful regularity conditions on X with little or no loss (we will assume it is a Carte- sianproduct).Incontrast,assumptionsontheshapeortopologicalproperties of X∗ implicitly restrict either the function σ or the subset of X˜ on which in- jectivitycanbedemonstrated.Avoidingsuchrestrictionsisonesignificantway inwhichwebreakfromthepriorliterature. There is a large literature on the injectivity of real functions, most of it de- velopingconditionsensuringthatalocallyinvertiblefunctionisgloballyinvert- ible.ThisliteraturegoesbackatleasttoHadamard(1906a,1906b).Although wecannotattemptafullreviewhere,themonographofParthasarathy(1983) provides an extensive treatment, and references to more recent work can be found in, for example, Parthasarathy and Ravindran (2003) and Gowda and Ravindran(2000).Localinvertibilityisitselfanopenquestioninmanyimpor- tantdemandmodels,whereusefulconditionslikelocalstrictgrosssubstitutes orlocalstrictdiagonaldominancefailbecauseeachgoodsubstitutesonlywith “nearby”goodsintheproductspace(severalexamplesbelowillustrate).Even when local invertibility is given, sufficient conditions for global invertibility in this literature have proven either inadequate for our purpose (ruling out im- portant models of demand) or problematic in the sense that transparent eco- nomicassumptionsdeliveringthese conditions havebeenoverlyrestrictive or evendifficulttoidentify. Acentralresultinthisliteratureistheglobal“univalence”theoremofGale and Nikaido (1965). Gale and Nikaido considered a differentiable real func- tionσ withnonsingularJacobianonX∗=X.2 Theyshowedthatσ isglobally injectiveifX∗ isarectangle(aproductofintervals)andtheJacobianisevery- where a P-matrix (all principal minors are strictly positive). While this result isoftenreliedontoensureinvertibilityofdemand,itsrequirementsareoften 2GaleandNikaido(1965)madenodistinctionbetweenX andX∗.Whenthesediffer,Gale andNikaidoimplicitlyfocusedontherestrictionofσtoX∗. 2090 S.BERRY,A.GANDHI,ANDP.HAILE problematic.Differentiabilityisessential,butfailsinsomeimportantmodels. Examplesincludethosewithdemanddefinedonadiscretedomain(e.g.,agrid ofprices),randomutilitymodelswithdiscretedistributions,orfinitemixtures ofverticalmodels.Givendifferentiability,theP-matrixconditioncanbediffi- cult to interpret, verify, or derive from widely applicable primitive conditions (seetheexamplesbelow).Finally,thepremiseofnonsingularJacobianonrect- angularX∗ isoftenasignificantlimitation.Forexample,ifxrepresentsprices and σ (x)=0, an increase in j’s price has no effect on any good’s demand, j implying a singular Jacobian. However, the set X˜ (which avoids σ (x)=0) is j often not a rectangle. In a market with vertically differentiated goods (e.g., MussaandRosen(1978)),alowerqualitygoodhasnodemandunlessitsprice is strictly below that of all higher quality goods. So if −x is the price vector, X˜ generally will not be a rectangle. Other examples include linear demand models, models of spatial differentiation (e.g., Salop (1979)), the “pure char- acteristics” model of Berry and Pakes (2007), and the Lancasterian model in AppendixC.WhenX˜ isnotarectangle,theGale–Nikaidoresultcandemon- strateinvertibilityonlyonarectangularstrictsubsetofX˜. Theliteratureonglobalinvertibilityhasnotbeenfocusedoninvertibilityof demand, and we are unaware of any result that avoids the broad limitations of the Gale–Nikaido result when applied to demand systems. This includes results based on Hadamard’s classic theorem.3 All require some combination of smoothness conditions, restrictions on the domain of interest, and restric- tions on the function (or its Jacobian) that are violated by important exam- plesand/oraredifficulttomotivatewithnaturaleconomicassumptions.4 The connected substitutes conditions avoid these limitations. They have clear in- terpretation and are easily checked based on qualitative features of the de- mand system. They hold in a wide range of models studied in practice and imply injectivity without any smoothness requirement or restriction on the setX∗. The plan of the paper is as follows. Section 2 provides several examples that motivate our interest, tie our general formulation to more familiar spe- 3See,forexample,Palais(1959)orthereferencesinRadulescuandRadulescu(1980). 4Afewresults,startingwithMas-Colell(1979),useadditionalsmoothnessconditionstoallow thedomaintobeanyfulldimensioncompactconvexpolyhedron.However,inapplicationsto demand,thenaturaldomainofinterestX˜ canbeopen,unbounded,and/ornonconvex.Exam- plesincludestandardmodelsofverticalorhorizontaldifferentiationorthe“purecharacteris- ticsmodel”ofBerryandPakes(2007).AndwhiletheboundariesofX˜ (iftheyexist)areoften planeswhenutilitiesarelinearinx,thisisnotageneralfeature.AlesscitedresultinGaleand Nikaido(1965)allowsarbitraryconvexdomainbutstrengthenstheJacobianconditiontorequire positive quasidefiniteness. Positive quasidefiniteness (even weak quasidefiniteness, explored in extensions)isviolatedonX˜ inprominentdemandmodels,includingthatofBerry,Levinsohn, andPakes(1995)whenxis(minus)thepricevector. CONNECTEDSUBSTITUTES 2091 cialcases,andprovideconnectionstorelatedwork.Wecompletethesetupin Section 3 and present the connected substitutes conditions in Section 4. We giveourmainresultinSection5.Section6presentsasecondtheoremthatun- derliesourpartialexplorationofnecessity,enablesustoprovidetightlinksto the classic results of Hadamard and of Gale and Nikaido, and has additional implicationsofimportancetotheeconometricsofdifferentiatedproductsmar- kets. 2. EXAMPLES EstimationofDiscreteChoiceDemandModels A large empirical literature uses random utility discrete choice models to study demand for differentiated products, building on pioneering work of McFadden (1974, 1981), Bresnahan (1981, 1987), and others. Conditional in- directutilitiesarenormalizedrelativetothatofgood0,oftenanoutsidegood representingpurchaseofgoodsnotexplicitlyunderstudy.Muchoftherecent literature follows Berry (1994) in modeling price endogeneity through a vec- tor of product-specific unobservables x, with each x shifting tastes for good j j monotonically. Holding observables fixed, σ(x) gives the vector of choice probabilities(ormarketshares).Becauseeachσ isanonlinearfunctionofthe j entirevectorofunobservablesx,invertibilityisnontrivial.However,itisessen- tial to standard estimation approaches, including those of Berry, Levinsohn, and Pakes (1995), Berry and Pakes (2007), and Dube, Fox, and Su (2012).5 Berry(1994)providedsufficientconditionsforinvertibilitythatincludelinear- ity of utilities, differentiability of σ (x), and strict gross substitutes.6 We re- j laxallthreeconditions,openingthepossibilityofdevelopingestimatorsbased on inverse demand functions for new extensions of standard models, includ- ing semiparametric or nonparametric models (e.g., Gandhi and Nevo (2011), Souza-Rodrigues(2011)). 5TheBerry,Levinsohn,andPakes(1995)estimationalgorithmalsoexploitsthefactthat,in themodelstheyconsidered,σissurjectiveatallparametervalues.Giveninjectivity,thisensures thatevenatwrong(i.e.,trial)parametervalues,theobservedchoiceprobabilitiescanbeinverted. Thispropertyisnotnecessaryforallestimationmethodsorforotherpurposesmotivatinginterest intheinverse.However,Gandhi(2010)providedsufficientconditionsforanonparametricmodel andalsodiscussedasolutionalgorithm. 6Although Berry (1994) assumed strict gross substitutes, his proof only required that each insidegoodstrictlysubstitutetotheoutsidegood.HotzandMiller(1993)providedaninvertibility theoremforasimilarclassofmodels,althoughtheyprovidedacompleteproofonlyforlocal,not global, invertibility. Berry and Pakes (2007) stated an invertibility result for a discrete choice model relaxing some assumptions in Berry (1994), while still assuming the linearity of utility inx.Theirproofisincomplete,althoughaddingthesecondofourtwoconnectedsubstitutes j conditionswouldcorrectthisdeficiency. 2092 S.BERRY,A.GANDHI,ANDP.HAILE NonparametricIdentificationofDemand Separate from practical estimation issues, there has been growing interest in the question of whether demand models in the spirit of Berry, Levinsohn, andPakes(1995)areidentifiedwithoutthestrongfunctionalformanddistri- butional assumptions typically used in applications. Berry and Haile (2010a, 2010b) have recently provided affirmative answers for nonparametric mod- els in which x is a vector of unobservables reflecting latent tastes in a market and/orunobservedcharacteristicsofthegoodsinamarket.Conditioningonall observables,oneobtainschoiceprobabilitiesoftheform(1).Theinvertibility resultbelowprovidesanessentiallemmaformanyofBerryandHaile’sidenti- ficationresults,manyofwhichextendimmediatelytoanydemandsystemsat- isfyingtheconnectedsubstitutesconditions.RelatedworkincludesChiappori andKomunjer(2009),whichimposedfunctionalformandsupportconditions ensuringthatinvertibilityfollowsfromtheresultsofGaleandNikaido(1965) orHadamard(Palais(1959)). InvertingforPreferenceShocksinContinuousDemandSystems Beckert and Blundell (2008) recently considered a model in which utility fromabundleofconsumptionquantities q=(q (cid:2)(cid:3)(cid:3)(cid:3)(cid:2)q )isgivenbyastrictly 0 J increasing C2 function u(q(cid:2)x), with x∈RJ denoting latent demand shocks. The price of good 0 is normalized to 1. Given total expenditure m and prices p=(p (cid:2)(cid:3)(cid:3)(cid:3)(cid:2)p ) for the remaining goods, qu(cid:7)antities demanded are given by 1 J q =h (p(cid:2)m(cid:2)x) j =1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J, with q =m− p q . Beckert and Blundell j j 0 j>0 j j (2008) considered invertibility of this demand system in the latent vector x, pointing out that this is a necessary step toward identification of demand or testingofstochasticrevealedpreferencerestrictions(e.g.,BlockandMarschak (1960), McFadden and Richter (1971, 1990), Falmagne (1978), McFadden (2004)).Theyprovidedseveralinvertibilityresults.Onerequiresmarginalrates of substitution between good 0 and goods j>0 to be multiplicatively separa- ble in x, with an invertible matrix of coefficients. Alternatively, they provided conditions(onfunctionalformand/orderivativematricesofmarginalratesof substitution)implyingtheGale–NikaidoJacobianrequirement.Weprovidean alternativetosuchrestrictions. One way to translate their model to ours is through expenditure shares. To dothis,fixpandm,letσ (x)=p h(p(cid:2)m(cid:2)x)/mforj>0.Expenditureshares j j sumto1,implyingtheidentity(2).Othertransformationsarealsopossible(see Example 1 below). And although Beckert and Blundell represented all goods intheeconomybyj=0(cid:2)1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J,acommonalternativeistoconsiderdemand foramorelimitedsetofgoods—forexample,thoseinaparticularproductcat- egory.Inthatcase,therewillnolongerbeagoodwhosedemandisdetermined from the others’ through the budget constraint, and it will be natural to have a demand shock x for every good j. This situation is also easily accommo- j dated. Holding prices and all other demand shifters fixed, let σ (x) now give j CONNECTEDSUBSTITUTES 2093 thequantityofgoodj demandedforj=1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J.Tocompletethemappingto ourmodel,let(2)definetheobjectσ (x).Ahintattherolethisartificialgood 0 0playsbelowcanbeseenbyobservingthatarisein σ (x) representsafallin 0 thedemandforgoodsj>0asawhole. ExistenceofInverseDemand,UniquenessofWalrasianEquilibrium Let −x be the price of good j. Conditional on all other demand shifters, j letσ (x)givethequantitydemandedofgoodj.Aneedforinvertibledemand j arises in several contexts. In an exchange economy, invertibility of aggregate WalrasiandemandisequivalenttouniquenessofWalrasianequilibriumprices. Our connected substitutes assumptions relax the strict gross substitutes prop- erty that is a standard sufficient condition for uniqueness. In a partial equi- librium setting, invertibility of aggregate Marshallian demand is required for competition in quantities to be well defined. The result of Gale and Nikaido (1965) has often been employed to show uniqueness. Cheng (1985) provided economically interpretable sufficient conditions, showing that the Gale and Nikaido(1965)Jacobianconditionholdsunderthedominantdiagonalcondi- tionofMcKenzie(1960)andarestrictiontostrictgrosssubstitutes.Inaddition to the limitations of requiring differentiability and, especially, a rectangular domain, the requirement of strict gross substitutes (here and in several other resultscitedabove)rulesoutmanystandardmodelsofdifferentiatedproducts, where substitution is only “local,” that is, between goods that are adjacent in the product space (see, e.g., Figure 1 and Appendix A below). Our invertibil- ityresultavoidstheselimitations.Herewewouldagainusetheidentity(2)to introduceanartificialgood0asatechnicaldevice. 3. MODEL Weconsiderademandsystemcharacterizedby(1).Recallthat x∈X ⊆RJ is a vector of demand shifters and that all other determinants of demand are held fixed.7 Although we refer to σ as “demand” (and to σ (x) as “demand j forgoodj”),σ maybeanytransformationofthedemandsystem,forexample, σ(x)=g◦f(x),wheref(x)givesquantitiesdemandedandg:f(X)→RJ.In thiscase, σ isinjective onlyif f is.Ourconnectedsubstitutesassumptionson σ are postponed to the following section; however, one should think of x as j amonotonicshifterofdemandforgoodj.Intheexamplesabove, x iseither j (minus) the price of good j, the unobserved quality of good j, or a shock to tasteforgoodj.Inalloftheseexamples,monotonicityisastandardproperty. LetJ ={0(cid:2)1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J}.Recallthatweseekinjectivityofσ onX∗⊆X,andthat wehavedefinedtheauxiliaryfunctionσ (x)in(2). 0 7Whengood0isarealgoodrelativetowhichpricesorutilitiesarenormalized,thisincludes allcharacteristicsofthisgood.Forexample,wedonotruleoutthepossibilitythatgood0hasa pricex ,butareholdingitfixed(e.g.,at1). 0 2094 S.BERRY,A.GANDHI,ANDP.HAILE ASSUMPTION1: X isaCartesianproduct. This assumption can be relaxed (see Berry, Gandhi, and Haile (2011) for details) but appears to be innocuous in most applications. We contrast this withGaleandNikaido’sassumptionthatX∗ isarectangle.Thereisasuperfi- cialsimilarity,sincearectangleisaspecialcaseofaCartesianproduct.Buta fundamental distinction is that we place no restriction on the set X∗ (see the discussionintheIntroduction).Further,Assumption1playstheroleofareg- ularityconditionhere,whereasrectangularityofX∗ isintegraltotheproofof GaleandNikaido’sresult(seealsoMoréandRheinboldt(1973))andlimitsits applicability. 4. CONNECTEDSUBSTITUTES Ourmainrequirementforinvertibilityisapairofconditionscharacterizing connected substitutes. The first is that the goods are weak substitutes in x in thesensethatwhenx increases(e.g.,j’spricefalls),demandforgoodsk(cid:7)=j j doesnotincrease. ASSUMPTION 2—Weak Substitutes: σj(x) is weakly decreasing in xk for all j∈{0(cid:2)1(cid:2)(cid:3)(cid:3)(cid:3)(cid:2)J},k∈/{0(cid:2)j}. We make three comments on this restriction. First, in a discrete choice model, it is implied by the standard assumptions that x is excluded from the j conditional indirect utilities of goods k(cid:7)=j and that the conditional indirect utilityofgoodj isincreasinginx . j Second, although Assumption 2 appears to rule out complements, it does not. In the case of indivisible goods, demand for complements can be rep- resented as arising from a discrete choice model in which every bundle is a distinctchoice(e.g.,Gentzkow(2004)).Asalreadynoted,theweaksubstitutes condition is mild in a discrete choice demand system. In the case of divisible goods, the fact that σ may be any transformation of the demand system en- ables Assumption 2 to admit some models of complements, including some witharbitrarilystrongcomplementarity.Example1belowillustrates. Finally, consider the relation of this assumption to a requirement of strict gross substitutes. If good zero is a real good, then our weak substitutes con- dition is weaker, corresponding to the usual notion of weak gross substitutes. When good 0 is an artificial good, Assumption 2 strengthens the weak gross substitutesconditioninanaturalway:takingthecasewherexis(minus)price, all else equal, a fall in the price of some good j >0 cannot cause the total demand(overallgoods)tofall. To state the second condition characterizing connected substitutes, we first defineadirectionalnotionof(strict)substitution. CONNECTEDSUBSTITUTES 2095 DEFINITION1: Goodjsubstitutestogoodkatxifσk(x)isstrictlydecreasing inx . j Consideradeclineinx withallelseheldfixed.ByAssumption2,thisweakly j raises σ (x) for all k(cid:7)=j. The goods to which j substitutes are those whose k demands σ (x) strictly rise. When x is (minus) the price of good j, this is k j a standard notion. Definition 1 merely extends this notion to other demand shifters that may play the role of x. Although this is a directional notion, in most examples it is symmetric; that is, j substitutes to k if and only if (iff) k substitutes to j.8 An exception is substitution to good 0: since any demand shiftersforgood0areheldfixed,Definition1doesnotdefinesubstitutionfrom good0toothergoods.9 Itwillbeusefultorepresentsubstitutionamongthegoodswiththedirected graphofamatrixΣ(x)whoseelementsare (cid:8) 1{goodj substitutestogoodkatx}(cid:2) j>0(cid:2) Σ = j+1(cid:2)k+1 0(cid:2) j=0(cid:3) ThedirectedgraphofΣ(x)hasnodes(vertices)representingeachgoodanda directed edge from node k to node (cid:6) whenever good k substitutes to good (cid:6) atx. ASSUMPTION 3—Connected Strict Substitution: For all x∈X∗, the directed graphofΣ(x)has,fromeverynodek(cid:7)=0,adirectedpathtonode0. Figure 1 illustrates the directed graphs of Σ(x) at generic x∈X˜ for some standard models of differentiated products, letting −x be the price vector and assuming (as usual) that each conditional indirect utility is strictly de- creasing in price. The connected substitutes conditions hold for X∗ ⊆X˜ in all of these models. As panel (e) illustrates, they hold even when J is com- posedofindependentgoodsandeitheranoutsidegoodoranartificialgood0. Eachoftheseexampleshasanextensiontomodelsofdiscrete/continuousde- mand(e.g.,NovshekandSonnenschein(1979),Hanemann(1984),Dubinand McFaden (1984)), models of multiple discrete choice (e.g., Hendel (1999), Dube(2004)),andmodelsofdifferentiatedproductsdemand(e.g.,Deneckere and Rothschild (1992), Perloff and Salop (1985)) that provide a foundation forrepresentativeconsumermodelsofmonopolisticcompetition(e.g.,Spence 8See,forexample,Berry,Gandhi,andHaile(2011).Weemphasizethatthisreferstosymmetry ofthebinarynotionofsubstitutiondefinedabove,nottosymmetryofanymagnitudes. 9Ifgood0isarealgooddesignatedtonormalizeutilitiesorprices,onecanimagineexpanding xtoincludex anddefiningsubstitutionfromgood0toothergoodspriortothenormalization 0 thatfixesx .IfAssumption3,below,holdsundertheoriginaldesignationofgood0,itwillhold 0 foralldesignationsofgood0aslongassubstitution(usingtheexpandedvectorx=(x (cid:2)(cid:3)(cid:3)(cid:3)x )) 0 J issymmetric(inthesensediscussedabove).
Description: