4 1 0 2 n a J 0 2 On Convergence in the Spatial AK Growth Models ] N G . by G. Aldashev, S. Aldashev, T. Carletti n i f - Report naXys-03-2014 19 01 2014 q [ 1 v 7 8 8 4 . 1 0 53.5 4 1 53 v: 52.5 Xi k(x,t) 52 51.5 r a 51 50.5 1 3 2 0.5 1 x 0 0 time Namur Center for Complex Systems University of Namur 8, rempart de la vierge, B5000 Namur (Belgium) http://www.naxys.be On Convergence in the Spatial AK Growth Models∗ Gani Aldashev† Serik Aldashev‡ Timoteo Carletti§ 19 01 2014 Abstract Recent research in economic theory attempts to study optimal economic growth and spatial location of economicactivityinaunifiedframework. Sofar,thekeyresultofthisliterature-asymptoticconvergence,even in the absence of decreasing returnsto capital - relies on specific assumptions about the objective of thesocial planner. We show that this result does not depend on such restrictive assumptions and obtains for a broader classofobjectivefunctions. Wealsogeneralizethisfinding,allowingforthetime-varyingtechnologyparameter, and providean explicit solution for thedynamicsof spatial distribution of thecapital stock. Keywords: Economic Growth; Convergence; Spatial Dynamics; Partial Differential Equations. JEL codes: C60, O40, O11, R11. ∗This research has been partially financed by the Belgian National Bank. The second author kindly acknowledges the support of theNamurCenterforComplexSystems(naXys)visitingproject. ThispaperpresentsresearchresultsoftheBelgianNetworkDYSCO (DynamicalSystems,Control,andOptimization),fundedbytheInteruniversityAttractionPolesProgramme,initiatedbytheBelgian State, SciencePolicyOffice. Thescientificresponsibilityrestswithitsauthor(s). †Corresponding author. Department of Economics and CRED, Universityof Namur. Mailingaddress: Department of Economics, 8RempartdelaVierge,5000Namur,Belgium. Email: [email protected]. Tel.: +3281724862, fax: +3281724840. ‡Abai Kazakh National University. Mailing address: Faculty of Mathematics and Physics, Tole Bi Street, 86, 050000 Almaty, Kazakhstan. Email: [email protected]. §NaXyS and Department of Mathematics, University of Namur. Mailingaddress: Rempart de laVierge8, 5000 Namur, Belgium. Email: [email protected]. 1 G. Aldashev, S. Aldashev, T. Carletti 2 1 Introduction Two major fields of research in theoretical economics - theory of economic growth and models of new economic geography - have been developing separately and independently until recently. The need to integrate them in a common framework, as well as the underlying fundamental mechanisms between endogenous growth and agglom- erationof economic activity - in particular,increasingreturns to scale - was forcefully stated by Lucas (1988)and, in more detail, by Krugman (1997). Over the last decade, economists have been trying to construct such unified analytical frameworks that allow to capture and describe the evolution of economic activity both in space and in time. Most of this recent literature has been summarized in a survey by Desmet and Rossi-Hansberg (2010), as well as in Chapters 18 and 19 of Acemoglu (2008). What emerges from these surveys is that while a commonly agreed framework has not yet emerged, research has been evolving using several different and highly promising approaches. One of these approaches is to build optimal growth models with spatial dimension allowing for continuous space and time, in which spatio-temporal evolution of capital stock is described by a partial differential equation, usually of the diffusion (parabolic) type. This framework, initially suggested by Isard and Liossatos (1979), has beenrigorouslydevelopedrecentlybyBrito(2004)andBoucekkineetal. (2009),forthespatialversionofRamsey- Cass-Koopmans optimal growth model (i.e. with decreasing returns to capital) and by Boucekkine et al. (2013) for an AK growth model. The majorresultinallofthese papersis the asymptotic disappearanceofspatialinequality,i.e. convergenceof thecapitalstockovertimetothesamelevelinalltheregions,despitetheinitiallyheterogeneousspatialdistribution ofcapital. ThisisnottoosurprisingforthespatialRamsey-Cass-Koopmansmodel,giventhedecreasingreturnsto capital, and thus an intuitive economic reasonfor the flow of capital from more capital-abundant to capital-scarce regions. However, the convergence result is much more surprising for the spatial AK endogenous growth model, given that the returns to capital are constant; it is well known that the non-spatial versions of the AK model exhibit non-convergence(Romer 1986,Lucas 1988, Rebelo 1991). Boucekkine et al. (2013) rely on a specific assumption on the objective of the benevolent social planner: she treats equally all the individuals, independently of their location (and thus initial endowment of capital). Given this, the social planner chooses to smooth the (detrended) consumption, both across space and time. This assumption could be difficult to justify according to certain ethical criteria, for example, giving more weight to individuals located inareaswith a lowerinitiallevel ofcapitalstock. Thus, a naturalquestionarises: How general is the convergence result in the spatial AK model? In particular, does this surprising result rely on the particular assumptions concerning the objective function of the benevolent social planner? This note provides an answer to this question and generalizes the findings of Boucekkine et al. (2013) in two importantways. First,wefindthattheconvergenceresultdoesnotrelyonrestrictiveassumptionsontheobjective G. Aldashev, S. Aldashev, T. Carletti 3 of the social planner. In particular, we show that the asymptotic convergence result holds for a program of the social planner with any objective function which gives rise to a continuous consumption function, provided that the presentdiscounted value ofthe flow ofmaximum consumption(onthe entire space)does not exceedthe initial capital stock at any point of the space. Second, we show that these results hold for the AK production function in which technology (A) evolves over time. This is an important generalization,because technologicalchange can potentially alter the spatial dynamics of capital stock, and thus a priori it is unclear whether the convergence result would hold in such settings. The rest of the paper is organized as follows. Section 2 presents the model. Section 3 develops our analytical results. Section 4 highlights directions for future research and concludes. Technical proofs are relegated to the Appendix. 2 Model In modelling the spatial economic growth process, we follow the approach initiated by Brito (2004) and further developed by Boucekkine et al. (2009) and Boucekkine et al. (2013). Spatial dimension is modelled as a circle of radius one, on which atomistic economic agents are assumed to be uniformly distributed. Thus, letting x to denote the geographic location of an agent, we have x ∈ T = [0,2π], where boundary points of T coincide. This circle is a stylized representation of different regions in a country (in case of a closed economy), or - allowing for perfect capital mobility - can represent the global economy. Time is assumed to be continuous and evolves from zero to infinity: t∈R+ =[0,∞). Our main object of interest is the spatial distribution of capital stock and its evolution over time. We denote the capitalaccumulatedatmomentt inpoint xwith k(x,t). The initialdistributionof capitalstock k(x,0) onthe circle is a known function k (x). 0 TheproductionfunctionisofAK-type(seeAcemoglu,2008,chapter11,fordetailedanalysisofthepropertiesof thenon-spatialgrowthmodelswiththeAKaggregateproductionfunction),i.e. thereturnstocapitalareconstant. Importantly,weallowfortechnologytochangeovertime (althoughthe technologicalfrontierinanygivenmoment isthe sameineverypointofthe circle). Thus, A(x,t)=A(t) denotesthe leveloftechnologyattime tinthe whole space. The instantaneous budget constraint of the agent located at point x at moment t is y(x,t)=c(x,t)+τ(x,t)+s(x,t), (1) and simply states that the production y(x,t) is divided between consumption c(x,t), trade balance τ(x,t) (given thateachregionis asmallopeneconomy),andsavings(x,t). Savingrepresentscapitalaccumulationintothe next instant of time: s(x,t)=kt(x,t). G. Aldashev, S. Aldashev, T. Carletti 4 As stated above, production uses the linear AK-technology: y(x,t)=A(t)k(x,t). Finally, concerning the trade balance, we assume that there is perfect capital mobility. In other words, consider a region(arc)Rofthecircle. Theconsumptioninexcessofunsavedoutputminusconsumptionof”domestic”output (i.e. of output produced within the region) comes from other regions, which is reflected by the trade balance of this region. However, in the balance of payments of the region R, this excess consumption has to be financed by capital outflows. Thus, the regionaltrade balance simply equals the symmetric of the inflow of capital from one of its border minus the outflow from the other border: τ(x,t)dx=−[kx(b,t)−kx(a,t)], (2) ZR where b and a are the boundaries of the regionR. Using the FundamentalTheorem ofCalculus, the trade balance can thus be written as τ(x,t)dx=− kxx(x,t)dx, ZR ZR which, for a length of the region R tending to zero, simply becomes τ(x,t)=−kxx(x,t). Thus, the instantaneous budget constraint (1) can be written as the following equation of motion of capital: A(t)k(x,t)=c(x,t)−kxx(x,t)+kt(x,t), (3) and this constraint must hold for any point x and moment t. Moreover, given that we represent the space as a circle, the values of the capital stock must coincide at the endpoints of the interval T=[0,2π], and the smooth-pasting condition must also hold, at any moment t: k(0,t)=k(2π,t) and kx(0,t)=kx(2π,t). (4) The problem of optimal growth in this economy is that of a social planner that maximizes a certain objective functionJ(k ,c(x,t))bychoosingthe consumptionfunctionc(x,t), subjecttothe instantaneousbudgetconstraint 0 (3), the boundary value conditions (4), and the initial value condition k(x,0) = k (x). Clearly, the value of the 0 capital stock k(x,t) must be non-negative everywhere and in any moment of time. Formally, this accounts to: Problem 1 Find a non-negative classical solution, namely a continuous function in the closed domain Ω¯, where Ω = T×R+, twice-continuously differentiable with respect to x in Ω, of the linear parabolic partial differential equation kt =kxx+A(t)k(x,t)−c(x,t) ∀(x,t)∈Ω, (5) G. Aldashev, S. Aldashev, T. Carletti 5 that satisfies the initial condition k(x,0)=k (x), ∀x∈T. (6) 0 The problem of the social planner is a highly complicated infinite-dimensional optimal control problem, where complicationsessentiallyarisebecauseoneoftheconstraintsisintheformofapartialdifferentialequation. Inakey contribution,Boucekkineetal. (2013)developananalyticalmethodologythatallowstoovercomethischallengeby adapting the dynamic programming methods to this infinite-dimensional problem. However, they need to impose a specific form on the objective function of the social planner, in order to obtain a characterizationof the optimal consumption function c(x,t). Instead, we attack this problem differently. We study the problem of finding a non-negative classical solution of the partial differential equation describing the equation of motion of capital stock, for a general (continuous) consumption function c(x,t). In doing so, we determine two different sufficient conditions on the consumption function that guarantee the uniqueness and non-negativity of the explicit solution of the PDE problem. The first condition leads to a space-invariant consumption function (and is thus equivalent to the one posited by the objective function of the socialplanner in Boucekkine et al. 2013).However,the second condition is more general, and has a different economic interpretation. Next, we show that the asymptotic properties of the solution are similar to the ones determined by Boucekkine et al. (2013); in particular, we prove the convergence of the capital stock in every point of the circle to the same level as t → ∞. Crucially, the second sufficient condition on the consumption function that we find is considerably weaker than those of Boucekkine et al. (2013). We thus show that the convergence result in the spatial AK model is not driven by a particular objective function of the social planner. 3 Analysis The aim of this section is twofold: (i) to provide solutions (Theorems 1 and 2) of Problem 1 under two different assumptions on the consumption function c(x,t), and (ii) to study the asymptotic behavior of such solutions (Propositions 6 and 7). Assumingthespatio-temporalconsumptionfunctiontobeasmoothconcavefunctionwithrespecttothespatial variable for any positive time moment, the following result holds: Theorem 1 Let Ω=T×R+. Assume the functions A:R+ →R+, k :T→R+∪{0} and c:T×R+ →R+∪{0} 0 are continuous in their respective domains. Assume also that cxx(x,t)≤0 ∀(x,t)∈T×R+, (7) and t k0(x)≥ e−R0sA(z)dzc(x,s)ds ∀(x,t)∈Ω. (8) Z 0 G. Aldashev, S. Aldashev, T. Carletti 6 Then the Problem 1 admits a unique non-negative classical solution. Note that given the coincidence of the endpoints of the interval T = [0,2π] and the smooth-pasting condition (4), the concavityassumption(7) leads to a consumption function that is invariantin space. Thus, it is equivalent to the one posited by the objective function of the social planner in Boucekkine et al. (2013). On the one hand, this is re-assuring,as (together with the convergence results in Section 3.2) it shows that the two approaches lead to the same conclusions. However, one may wonder whether the analysis extends to a less restrictive sufficiency condition. We show that this is indeed the case; in fact, the following result holds: Theorem 2 Let Ω=T×R+. Assume the functions A:R+ →R+, k :T→R+∪{0} and c:T×R+ →R+∪{0} 0 are continuous in their respective domains. Assume, moreover, that t k0(x)≥ e−R0sA(z)dzmaxc(x,s)ds ∀(x,t)∈Ω. (9) Z x∈T 0 Then the Problem 1 admits a unique non-negative classical solution. Theorem 2 only requires that the consumption function is continuous and that the present discounted value of the flow of maximum consumption (on the entire space) does not exceed the initial capital stock at any point of the space. Note that the discounting is done using a (time-varying) technology parameter. The economic interpretation of this condition is as follows. Note that (9) allows for spatial inequality in consumption; it just imposes the upper bound on the (present discounted value of the) highest values of this consumption. Moreover, given the discounting, it allows for an increasing spatial inequality in consumption over time. The upper bound imposed depends on the initial distribution of the capital stock, and in particular, the condition (9) is most stringent for the lowest initial capital stock on the circle. In other words, it is more difficult to satisfy this condition when, ceteris paribus, the initial spatial inequality in capital stock is higher.1 3.1 Proof of the main results Let’s start by observing that we can simplify Eq. (5) by removing the term A(t)k(x,t) using the following Lemma (whose proof is presented in the Appendix): Lemma 1 (Equivalent solutions) Let k(x,t) and h(x,t) be two positive functions defined in Ω and related to each other by h(x,t)=e−R0tA(s)dsk(x,t). (10) 1Theconditions(8)and(9)holdforallpositivet;hence,theyimplythattheintegralontherighthandsideconvergesatthelimit. Thisimplies lim e−R0tA(s)dsc(x,t) =0 ∀x∈T, t→∞h i and lim e−R0tA(s)dsmaxc(x,t) =0, t→∞(cid:20) x∈T (cid:21) whichconstrainstheconsumptionfunctiontogrowovertimeataslightlylowerratethanthetime-varyingtechnology parameter. G. Aldashev, S. Aldashev, T. Carletti 7 Then k(x,t) is a positive solution of (5) with initial condition (6) if and only if h(x,t) is a positive solution of ht =hxx−γ(x,t) ∀(x,t)∈Ω, (11) with the same initial condition (6), where γ(x,t)=e−R0tA(s)dsc(x,t). The proofs of Theorems 1 and 2 are constructed with following steps.2 We first find a formal solution, i.e. a Fourier series that, order by order, solves (5). Proposition 1 (Formal solution) Let us define, for any positive integer n, λn =n2 and, for (x,y,t)∈T2×R+, the Green’s function G(x,y,t)= e−λntcos[n(x−y)] . (12) nX≥0 Then the function h(x,t) given by π t π 1 1 h(x,t)= G(x,y,t)k (y)dy− ds G(x,y,t−s)γ(y,s)dy, (13) π Z−π 0 π Z0 Z−π is a formal solution of Eq. (11). Thesecondstepistoprovethatthesolutionprovidedbythepreviouspropositionisactuallyaclassicalsolution. Proposition 2 (Classical solution) Under the above assumptions the function G(x,y,t), respectively h(x,t), is continuous in T2 × [0,+∞), respectively T × [0,+∞), and twice differentiable in T2 × (0,+∞), respectively T×(0,+∞). The third steps is to prove the uniqueness of the classical solution: Proposition 3 (Uniqueness) The classical solution of the Eq. (11) with initial condition k(x,t) = k (x) is 0 unique. Considering the above-mentionedresults and using Lemma 1, we can conclude that k(x,t) = h(x,t)eR0tA(s)ds (14) π t π := eR0tA(s)ds(cid:20)π1 Z−πG(x,y,t)k0(y)dy− π1 Z0 dsZ−πG(x,y,t−s)γ(y,s)dy(cid:21) , is the unique classical solution of (5) with initial condition k (x). 0 The laststepforprovingTheorem1is to showthatthe solutiongivenby (14) isnon-negativeforall(x,t)∈Ω, that is the function describing the evolution of the stock of capital in space and time k(x,t) is everywhere and alwaysnon-negative. Forthis, weneedapreliminaryresult(Lemma2)whoseproofcanbe foundinthe Appendix. 2Forthesakeofclarity,alltheproofsexceptthoseconcerningthenon-negativity arerelegatedtotheAppendix. G. Aldashev, S. Aldashev, T. Carletti 8 Proposition 4 (Non-negativity) Leth(x,t)betheclassicalsolutionoftheEq.(11)inΩwithboundarycondition h(x,0)=k (x)≥0 for all x∈T and assume γ(x,t)≥0 for all (x,t)∈Ω. Under the hypotheses 0 γ (x,t)≤0 ∀(x,t)∈Ω, xx and t k (x)≥ γ(x,s)ds ∀(x,t)∈Ω, 0 Z 0 we have h(x,t)≥0 ∀(x,t)∈Ω. (15) Proof. Let T >0, ǫ>0 and let us define the auxiliary function t v(x,t)=h(x,t)+ǫt+ γ(x,s)ds, Z 0 for all (x,t)∈T×[0,T]. A straightforwardcomputation gives: t t vt−vxx =ht−hxx+ǫ+γ(x,t)− γxx(x,s)ds=ǫ− γxx(x,s)ds≥ǫ>0, Z Z 0 0 wherewe use the factthat ht−hxx =−γ(x,t) andthe assumptionγxx(x,s)≤0. We canthus apply the Lemma2 to v and conclude that it attains its minimum at some (a,τ)∈T×{0}. Hence, we have that for all (x,t)∈T×[0,T] t h(x,t)+ǫt+ γ(x,s)ds=v(x,t)≥v(x,0)=h(x,0)=k (x), 0 Z 0 thus t h(x,t)+ǫt≥k (x)− γ(x,s)ds≥0, 0 Z 0 where the rightmost inequality holds because of the assumption on k . We can finally pass to the limit ǫ→0 and 0 conclude that h(x,t)≥0 (x,t)∈T×[0,T]. The arbitrariness of T completes the proof. Because the function k(x,t) has the same sign that h(x,t), we conclude that k(x,t) is also non-negative in Ω and this concludes the proof of Theorem 1. The proofof Theorem 2 can be achievedin a very similar way; the only difference lies on the way we prove the non-negativity of the solution (14) under the assumptions of Theorem 2. Proposition 5 (Non-negativity) Leth(x,t)betheclassicalsolutionoftheEq.(11)inΩwithboundarycondition h(x,0)=k (x)≥0 for all x∈T and assume γ(x,t)≥0 for all (x,t)∈Ω. Under the hypothesis 0 t k (x)≥ maxγ(x,s)ds ∀(x,t)∈Ω, 0 Z x∈T 0 G. Aldashev, S. Aldashev, T. Carletti 9 we have h(x,t)≥0 ∀(x,t)∈Ω. (16) Proof. Let T >0, ǫ>0 and let us define the auxiliary function t v(x,t)=h(x,t)+ǫt+ maxγ(x,s)ds, Z x∈T 0 for all (x,t)∈T×[0,T]. A direct computation provides vt−vxx ≥ 0 and, thus, by Lemma 2, v attains its minimum at some (a,τ) ∈ T × {0}. Then, following an argument similar to the one used in the previous Proposition we conclude that t h(x,t)+ǫt≥k0(x)− 0 maxx∈Tγ(x,s)ds≥0. Passingto the limit ǫ→0, andusing the arbitrarinessofT, we get R the result. Onceagain,becausek(x,t)andh(x,t)differbyapositivefunction,wecanconcludethatk(x,t)isnon-negative in Ω. Ourresultsaboveshowthatifthe initialspatialinequalityincapitalis nottoo starkandthatthe consumption does not differ too muchacrossspace (so as to respect (9)), a givenconsumption function uniquely determines the spatialgrowthprocess. However,thenaturalquestionremainswhethersuchdynamics(always)leadstoconvergence of capital stock across space over time. We address this question below. 3.2 Asymptotic behavior and convergence The aim of this sub-section is to study the asymptotic behavior of the classical solution k(x,t) for large t. Our analysisrelies on the behaviorof the Green’s function G(x,y,t) for larget (that is fully characterizedby Lemma 3 in the Appendix). Let us first consider the case of time-independent technology, i.e. A(t) = A for all t. Our main result is the 0 following Proposition 6 Let (9) hold andk(x,t) betheclassical non-negativesolution of (5)with initial condition k(x,0)= k (x) and A(t)=A ∈R+. Then, assuming c(t) to be bounded, we have 0 0 e ∞ lim k(x,t)e−A0t =k − e−A0sc(s)ds (17) t→∞ 0 Z (cid:2) (cid:3) 0 e e uniformly in T, where π π 1 1 k = k (x)dx and c(t)= c(x,t)dx. 0 π Z−π 0 π Z−π e e Proof. From the explicit form for the classical non-negative solution k(x,t) in the case A(t)=A we get: 0 π t π 1 1 k(x,t)e−A0t = G(x,y,t)k (y)dy− ds G(x,y,t−s)e−A0sc(y,s)dy, π Z−π 0 π Z0 Z−π