Federal Reserve Bank of Minneapolis Research Department Staff Report 91 March 1984 TIME CONSISTENCY OF OPTIMAL PLANS: AN ELEMENTARY PRINER Michael J. Stutzer Federal Reserve Bank of Minneapolis ABSTRACT, ‘Mme consistent optimal plans are defined within the context of a simple, discrete time optim control framework. Three possible sources of inconsistency are identified and discussed with refer- ence to the literature. ‘The views expressed herein are solely those of the author and do not necessarily represent the views of the Federal Reserve Bank of Minneapolis or the Federal Reserve System.Introduction A planning problem consists of the following elements: 1. A designated agent, termed the primordial planner, who chooses a (possibly infinite) planning horizon, the initial time of which is chosen, without losses of generality, to be t = 0. ‘The final time of the horizon is denoted T, with T = © for an infinite horizon. 2. A set of present and future sequences (or continua) of decisions, termed plans. A subset of these plans is determined by the primordial planner to be feasible, given his perceptions of ‘the problem's constraints. 3+ A system linking the feasible plans to outcomes they 4. Performance criteria used ty the primordial planner to rank the desirability of feasible plan outcomes. The plan that is ranked highest (if one exists) is termed an optimal or, equiva- lently, a primordial plan. As the optimal plan is implemented, other planners my want to change it. They may have fundamental differences with the primordial planner about one or more of the planning elenents. For example, new instruments may become available that change the constraints determining the feasible plans. Unforeseeable nev informtion may become known about the system. Or, society my demand that the planner fundamentally change the performance criteria initially adopted. In any of these events, it seems appropriate that the “optimal” plan be changed to reflect the new elements.-2- Tt is possible, though, that future planners will be in agreement vith the primordial planner about the last three plen— ning elements. Because they evaluate the plans at later starting dates, hovever, their planning horizons are not the sane as those of the primordial planner. Future planners vill only concern themselves with the impacts of their proposed plane on their own present and future, ignoring their om pasts. ‘This myopic (from the viewpoint of the primordial planner) behavior my lead future Planners to deviate from the primordial plan, even if they all agree with the primordial planner on planning elenente 2-4. From @ normative point of view, if one views the primordial planner's choice of horizon as the appropriate choice, then this myopic be- havior 4s undesirable. Can it somehow be avoided? If the primordial planner cannot force future planners to implement the optimal plan, he can only hope that they will, of their ow free will, choose to implement it. A planning problem having the property that ite optim plan will be implenented by all future planners, each of whom is free to change it but will not do so, is termed a planning problem with @ consistent optinal plan, Even if a planning problem does not have consistent optimal plans, some future planner at time t < T may freely choose to implement thet part of the optimal plan under his control. A planning problem whose optimal plan has this property is termed a problem with a t-consistent optimal plan. As we will see, only certain problems with very special structures have consistent or t-consistent optim plans. Strotz (1956) and Burness (1976) have shown that these structures also-3- require that, in any problem that involves discounting the future, ‘the discount function mst be of a very special form. In economic planning problems, Kydland and Prescott (1980), Calvo (1978), and Fischer (1980) have shown that these special structures require expectations to be adaptive rather than rational or, in some other sense, determined ty future policy decisions. Finally, ve vill also see that problems with what are usually termed nonseparable Performance criteria do not possess consistent optimal plans. I will present these results in a unified vay through the use of discrete-time optimal control model. The model is ‘broad enough to encompass all of the special structures above, yet simple enough to be understood ty people unfamiliar with dynamic optimization techniques. In the process, though, the reader will Decome acquainted with the so-called Bellman Principle of Opti- mlity, which is a necessary condition characterizing optim) Plens in planning problems with certain special structures. We will see that, when discounting 1s not used, the problems whose optimal plans are characterized ty Bellmn's principle are pre- cisely those problems whose optimal plans are consistent. When @iscounting 1s used, however, this 4s no longer the case, and a aifferent test for consistency is required. In problens whose optimal plans are not consistent, the primordial planner mst either find some way to force future planners to implement her plan, or adopt some other strategy. If the former course of action is impossible, the primordial planner may choose to propose some other plan that will be freely im plemented by future planners. Such a plan is termed consistent,and in general there my be many consistent, albeit suboptim1, plans. In problems not involving discounting, plans that are consistent are precisely those derived via the use of Bellman's principle, regardless of whether or not it characterizes the optim plan. As mentioned above, these consistent plans are suboptimal unless the problem bas e structure in which Bellman's principle does characterize the optim plan. ‘The Optimal Planning Problen ‘The most general optimal planning problem considered here can always be placed in the form: wo Oy sD yee Ky Poking ot) Bete Key = fy (My Pps Xp given, t = Ose, Dy © OL (Ky)e For each t, the primordial planner chooses a vector Dy from a choice set Cy(X,) to solve (1), producing an optimal or primordial ‘Blan (D§,+++,DK)+ The vector-valued state equations f, determine the evolution of the state vector X%- Note that ve permit the possibility that future decisions may affect current states. ‘This simple setup is broad enough to exhibit all known causes of incon— sistency while avoiding complications concerning the existence end computation of optimal plans inherent in both infinite horizon and stochastic planning problems. An illustrative example follows.Example: Capital Budgeting Consider the following problem faced ty a firm's primor- @ial planner, who chooses a plan to ellocate a fixed capital budget K among T competing projects. Project t starts in period t and, for simplicity, is assumed to lest one period. A dollar Anvested in the project starting (and ending) in period t earns a dividend of By. An investment of D, dollars in period t 4s then assumed to yield dividends of B,D,. In addition, retained earn- ings in period t, denoted X,, are taxed in that period at the rate ‘The undiscounted return to the firm in period t ia then: (2) Uz (Ky sDg) = DEB — PyXes t = Oseee se At time t = 0, the firm's primordial planner correctly perceives a future interest rate series (1,,+++,ip)s In deriving her plan, she discounts the return in period t ty the factor 1/(1+i,) + (tig) + se. + (1#4,). Defining the 4: (3) r(tjt) = 1/(14t gy) (1H _yp) (0,0) = 1, ‘the primordial planner's maximand is then: r (4) J w(t,0)0, (4, 504) To complete the specification of the primordial plan— ner's problem, we derive the state equation for retained earnings Xge Starting at t = 0, the planner invests Dp from the total-6- capital fund K, leaving before-tax retained earnings of K - Doe After-tax retained earnings of X; = (1-Py) (K-Dy) are then left for future investment purposes. In period t = 1, investment of Dy leaves before-tax retained earnings of X - Dj, and after-tax earnings of Xp = (1-P,) (Xy-D)) are available for future invest- ment purposes. Continuing in this mnner, we see that: 4 (5) Xgaq = GPL) ODE) Ff, 0G DL). t= Oseee es Hy = Ke The primordial planner's problem is then: . 4 (6) Dp ghy PEs elHyrPg) # UysDgrene KysDakgyar#) Bete Key = (KI Fl%psPy)s 2 oTs Xo = Ks Dy © Oy (X,) = [0,x4], where fy is given ty (5). Classification of Optimal Planning Problems There are several subclasses of problem (1) that prove useful in the following. One subclass, termed nonanticipatory, constitutes those problems in which the current state vector depends only on current and rast decision vectors, ise., m Keer = Fe(KesDorree sDg)s 6 = Opeee st By redefining the current state X, to include the past decision vectors Do,+++,Dz_1, the state equations of e nonanticipatory problem can alvays be written in the form:(8) Kear = feXpeDy)s t= 0 he most general nonanticipatory problem is illustrated in Figure Le Another important subclass of problems are the monotoni~ cally separable problems, in which the objective function U is separable and monotone in the following sense: (9) U(X) Dos MP Xn) [Uy (Dg) Uy (Dy eves sth Cy sDy) sUingg ggg) = 6, [Ug (3D) ot (Uy Oy 5g) +2 Un Xp sDp) Unga (Spa)? ana for any fixed value of Up, a; is @ monotonically non~ decreasing function of ap. A common type of monotonically separable problem is the additive separable problen: . Y= YD) + Unag Mpeg)» (9a) (Xp, * ° t=0 ete oXmeDipe Kanga as diagramed in Figure 2; and the discounted additive separable problem: 2 (90) v= J wle,ou (xy Dg) + (42,0000), +0 where r(t,0) is a real valued discount function, giving the pri- mordial planner's discount factor for decisions implemented at time te-8- Definition of Consistent Plans Roughly speaking, a consistent plan is a sequence of deciatons (Dgy+++ Dy) having the property that, for each time t, D, is the first decision in a plan that maximizes the "objective Pay above maximization treats the state evolution Xj ,+++,K, para~ have been determined. At each t, the remaining" after Dj, metrically; and it calculates the impacts of Don the future states X.yzsreKp and utilities Uzs+eeUp More precisely, we adopt two form] definitions within our planning framevork--one for undiscounted problems and the other for discounted ones. For undiscounted problems, a plan (Dy,+e+4D,) 48 con= sistent if and only if for each 0< + <7, 5, is the first com ponent in a solution of the following problem: 2D U(X sD g2Xq2Dy 3+ a Pe Kee Qo) yom Poy Pg) Pua 7 Bote Kean = ty llysDyseeesDy_y DyseeeaDgls t= ty with Xp given and (K,,+++5%p) ae the state evolution remting from (Dy, saDp) and the state equations f. This property can be weakened by requiring only that 5, be the first component of a solution vector to (10) for some particular > Os This veaker Property is termed t-consistencye The plan is consistent if and only if it is t-consistent for all t= 1,..., For discounted problems with the performance index (9b), the formation (10) will not do because (10) assumes that the maximand of a planner at time t isr * (22) LT r(t,0)U, 0%, DE) + F(T2,0)U ps (pyg )s tt which implies that the planner at t @iscounta his ovn current and future utilities as heavily as did the primordial planner t per- jods earlier. A more realistic assumption treats any future planners and the primordial planner symmetrically by assuming that the planner at t discounts future utilities Upyy,+++,Up in the same way that the primordial planner discounts her own future utilities Uj,.+.,Upg+ For example, if it is assumed that the prinordial planner does not discount her current utility Up, then we also mst assume that the planner at t does not discount his own current utility Uy. Denoting the discount factor for U, of a sDp) 4s dis~ planer at time t ty r(tyt), we say a plan (Dos counted consistent 1¢ and only if for each t,0< 16%, 5, te the first component in a solution vector to the folloving problem: 2 Pp & v(t, t)U, (Dy) + F(T TU yyy (ngy) (12) Fy (Xy Dg rere Dey Dy veer sDp)s & = 7, ry (Xp Dg aT Bete Xoo If, for some particular t > 0, D, is the first component in a solution vector to (9b), we say that the plan is iscounted consistent. Of course, the optinal plen (D§,+++,DR) in either an undiscounted or @ discounted problem may or may not be consistent or discounted consistent. In the following sections, I explore ‘the relationship between consistency and optimlity.~10- The Relationship betveen Consistency and Optimality for Undiscounted Problems Bellvan's Prineiple of Optimality Richard Bellman formilated his optimization technique in a 1957 work entitled Dynamic Programming, The beeis for the technique, called the Principle of Optimality, ie a necessary condition characterizing optim1 plans for certain types of plan— ning problens. According to Bellman's description: ‘A optimal policy has the property that whatever the initial state and initial de- cision are, the remining decisions mst constitute en optimal policy with regard to the state resulting from the firet decision. (p. 83) In our jargon, the applicability of Bellmn's principle implies ‘that the optim plan (D¥,+++,DK) in an undiscounted problem is Inconsistent because (Df,+..,DE) solves (10) for t = 1; its first component, Dt, is thus the first decision in the optimal plan. If Bellman's principle applies at each t > 0, considering the initial state to be Xtand the initial decision to be Dt_,, then the optim plan is consistent. As shown ty Mitten (1964), and restated in Nenhauser (1966), Beliman's principle applies at each t for nonanticipatory, monotonically separable problems, which proves that they have consistent optiml plans. ‘The application of Bellman's principle in the important special case of nonanticipatory, additive separ- ability is demonstrated below:-u- + Up(%psDq) (13) Uy (yaDy) + ave + UE Oy .DA) + * Une naa) Bete Xen = fe(XpaDy)s Mp elven, t= OyeeesTy Dy © Cy (x) and is diagrammed in Figure 2. To apply Bellman's principle, we first note that the nonanticipatory nature of the state equation allows us to compute Xy4z Fecuretvely as follows: (a4) X = £9(Kq.Po) Xp = f1(X5Dz) = £4(£9(%q5P9) sD, 3 = £ppsBp) = ply (t9(%s09) Dy) Da) Kear = Pe (Ke oDg) = ty (P41 (fol ++(fo(XpsDp) Dz) sDa) ++) ,Dy) Thus, any future state depends solely on decisions mde prior to ‘that time and to the initial state Y>, which is clear from a glance at Figure 2. By substituting the above relations into the additively separable objective function U, we see that the problem becomes: (as) ik Dp Uy (Xp sPq) + Uy (fq (Xp sPq)D4) + Ug (£4 (9% 0g)D,)sPQ) + ore + Upp aTy) + Unga Chaya)»~12- with Xo given and Dy chosen from the eet Cy(Xq), t = 7 It is the special structure of (13) that permits the application of Bellman's principle. Noting that D, does not occur tn Up, we can solve (13) ty: (a6) max[y(%5Dq) + y mex 5 Uy CRD.) + eve + Upp aDp) + = pp, 0 Moen na)!» Bets Dy © Colo) Dg © Oy Ky)» Xo given XE = 5%.) Kear = f4(QeDy) t= 1, This to precisely vhat Helinan's principle saya-—that the rexain= ing decisions (Df,--+,DA) mst constitute an optiml plan with regard to the state [xt = f,(X5,D§)] resulting from the first decision Dg. Also, thie ability to decompose the optintzation problen shows that the optinal plan solving (10) is 1-consis- tent. A planner at time 1 who reconsiders the optim plan Df,+++,D computed ty the primordial planner would solve (10) for 1 = 1, taking Xf = f(%p,0§) as given. Pictorially, the planner at time 1 ignores the part of Figure 2 to the left of % in cal- culating bis plan. This is the same problem (and thus produces the same solution) as the imer mximtzation in (16) computed ty ‘the primordial planner. Th fact, at each t > 0, Bellmn's prin ciple still applies to (10), vhose solution is thus t-conaistent for all t, iss, consistent. To see this, note from (16) that its inner maximization can be similarly decomposed into two parts:-13- 2 max fu, (x, D+ Spy na P an UO 2Dg) + Tagg (Xp) r = max[U,(X,,D,) + (u,0g,0,) + 2 04D.) + sea) # oe ny ltB) + Fler Ugg Mpg) IMs By © 6%) Bete Dy © 0,(%4) De © Cy (%) % given YS = f(x, .D,) Xyer = Py (%poDp) te2,000 57 because th does not depend on the future decisions Dy,+++,Dy. ‘Tms, Bellman's principle applies again, and the optim plan is also 2-consistent. Continuing in this mnner, we see that the original T + 1 variable optimization problem (13) can be de- composed into T +1 single-variable optimization problems, yield~ ing the dynamic programming decomposition (8) pep, Y= BrloU%o-20) + pena) teat max [Up (Xp Dp) + Ungy (Xnyy) Jee] », T where Xyua = fy(ysDy)s t = Oseees? and Dy © Cy(Xp)» Therefore, Bellman's principle applies at each t > 03 and the optimal plan in the undiscounted, additive separable, nonanticipatory problem is consistent. A planner at any time t > 0 thus ignores the part of Figure 2 to the left of X, in computing his plan. As ws men— tioned earlier, Bellmn's principle will aleo apply at each t in the more general class of undiscounted, monotonically separable, nonanticipatory problems maximizing (9) subject to (8). Optim1 plans for these problens are then also consistent. To sumarize, Fellmn's principle applies to undiscounted, monotonically sep--1be arable, nonanticipatory problems. In any undiscounted problem, the applicability of Bellnan's principle at each t implies that ‘the optim plan is consistent. Discounted Consistency for Discounted Problems ‘The discounted, additive separable, nonanticipatory problen of maximizing (9b) subject to (8) can also be solved ty applying Bellman's principle at each t. ‘This 1s easily done by redefining Uz to be r(t,0)U,(X,,D4), which brings (9b) into the form of (13)+ One then applies the proof just given. This fact, though, has no bearing on whether or not the optimal plan is discounted consistent. Although the primordial plan satisfies (10) for each t, @iscounted consistency requires the satisfaction of (12) for each t. In a continuous time set— ting, Burness, generalizing a result of Strotz, found necessary and sufficient conditions that mst be satisfied ty r(t,t) to ensure discounted consistency of the optiml plan. His proof involves the use of the calculus of variations. Our discrete-time formiation allows us to derive the discrete-time analog of his result using simple calculus. Theoren: A necessary and sufficient condition for the discounted consistency of an optimal plan maximizing (9b) subject to (8) is that, (9) EL A MEE), tor ants, t,t, t's As a simple corollary, if the discount function is given by (3), ‘then the optimal plan is discounted consistent.-15- Proof: We first prove necessity. The final decision Dk of the optimal plan is found by solving (20) max H(T,0) Uy (XH5Dq) + r(THL OV as (Fp (XESDy)) . where X& is the state at time T of the primordial plan. Assuming an interior mximum, solving (20) yields: au, Way af, ¥(7,0) aye (xe,0H) + (42,0) tt. Wp (xp,DB) = 0 or ay, n <2 (x8 ,pe) (41,0) _ Dyes (21) r(t,0) ~~ By BF Ty Bay PH) A planner at time 1 would solve (12), with his final decision Dy computed ty solving: (22) max rT) KqsDg) + (TL Tg (fe EpsDg) >, ” where X, is the state at time T and is teken paranetrically. Solving (22) yields: ne ay Tatty x(etyt) _ a (23) r(T,t) ayy Ep rae Fe, HP php Coneistency of the optimal plen requires the sequence of future planners at each time t to choose the same decision that the primordial planner did. This means that Dy = D¥, Dy = Bevo By Rom Roms ‘the primordial plan requires X, = X% and D, = DE, so thst the D = Dk, which, through the state equations, implies that > %yyq = hye In particular, consistency of right-hand sides of (21) and (23) are equal. Thus, we see that-16- r(T#1,0) _ r(T+1,1) (en) SGROY "ECT + = Pees tees, that the ratio L725) 15 independent of te Next, consider the primordial planner's choice Dt_,+ This is formed ty solving: (25) max [r(T-1,0)Uy 5 (X89 Dy) + Poor sax [e(1,0) tgp OR Py g)s0p) 2 + (42,0) pg (p(t. g (RR g Pye g) Dg) IIT whose solution satisfies the condition that the partial derivative of (25) with respect to Dp_y equals zero, or Un (26) r(-1,0) se (XK, DA) + m2 Wy a¢, nf x(1,0) ye gE (om, DH) + 9fp Oy MET ET Wy fp Pfr r(T+1,0) SE ge pp (XK, DH) = 0+ yf Pg Dividing by r(T,0) and rearranging leaves: x(n-1,0) 2Uy fa (en 1,0) yy” By yy x(o41,0) Unga fn PF p.- + ae yt =. ¥(T,0) BF, FF ay where all derivatives are evaluated along the optimal plan.-1T- ‘The planner at any time t who reconsiders this choice of Dyz Would solve: (28) max [r(T-1,7)Up_ 5 (XK 9 Dy_y) Poa + marl Mpg OF. 2 (TLD 9 (fg Hg Dig) Dg) HT + a?) Consistency requires its solution to be the same as that of (25), 50 a modified condition (27) results, with 0 replaced by t: r(toa,t) 2Upa (29) = aap a x(teiye) Umer 2% p PF pa x(T1,t) he we tare gust coon, EE) te sacozentent ot xr(?-1,1) Therefore, (29) implies that “HT, z) mst also be independent of Ty Aneluding, ty (27), r= 0. But then, of course, EMLAT) mt also be independent of t Considering the primordial and future planners’ problens for finding Dp, and imposing consistency, we similarly find thos Fey and Bebe are also independent of t+ Continuing the process for Dpz,+++4Do, it is tedious, tut simple, to find ‘that it 1s necessary that a discounted consistent, optimal plan in @ nonanticipatory, additive separable problem satisfy (19). The proof of the sufficiency of (19) for the discounted consistency of the optimal plan is straightforvard and will be omitted here.-18- The Relevance of Strotz-Burness The problem both Strotz and Burness use to discuss this result is one where the primordial planner, who elso plays the role of all future planners, is planning her own future consump- tion stream. If the planner's preference ordering over future consumption streans is not representable ty a discounted separable utility (9b) with r(t,t) of the necessary form (19), then the primordial plan is not consistent. Strotz views the lack of consistency as damaging to a normative theory of behavior, which prescribes that the plan that should be implemented is the pri- mordial plan. From this viewpoint, the consumer who does not implement her primordial plan is myopic: An individual, who because he does not dis- count all future pleasures at a constant rate of interest finds himself continuously repud- jating his past plans, my learn to distrust his future behavior, and my do something about it. (Strotz 1956, 173). Strotz proposes that such en individual may choose either of tvo courses: At time zero, the individual chooses to Precommit her future decisions irrevocably; or the individual decides to successively recalculate the optimal plan. The latter course would produce @ plan in which, for each t, the chosen D, is found by solving (12) for t. By definition, this procedure always generates discounted consistent plen, which is not the prinor— dial plan unless r(t,t) is of the neceseary form (20). Strotz's alarm is unjustified in the type of problem he @iscussed where there is only a single decision mker, the primor— dial planners Unless the primordial planner has changed her mind-19- about the appropriate horizon--or about the appropriate perfor- mance criterion, state equations, and constreints operative over that horizon--she never would reconsider the primordial plan. Its consistency (or lack of) is then a moot point. If the primordial planner did change her mind about one or more of these planning elements, then she would not necessarily consider the primordial plan to be optimal anymore. ‘The fact that it is also inconsistent in this event does not imply that the primordiel planner is myopic. Rather, the economist who posits thet such e decision maker will implement the primordial plan has made an incorrect ascumption. Many economic modelers of a single decision mker's Dehavior have assumed that the primordial plan will be imple- mented. These modelers assume that the single decision mker (eeg-, a firm or a consumer) maximizes some objective (e.g-, profits or utility) over some fixed horizon, usually chosen to be infinite. Examples include the dynamic competitive firm models treated by Sargent (1979). Among theories of the firm, the most widely accepted behavioral assumption is that firms maximize iscounted profits, using the discount function (3) of our capital budgeting example. Among theories of the consumer, there is less agreement on the admissible forms of the consumer's intertemporal utility function. However, the assumption commonly used in the recent rationel expectations models of Sargent and others is that 2 consumer's intertemporal preferences are representable ty a dis~ counted, additively separable, time-invariant utility function:=20- (30) Gptue,) 4 True, +20 nore + = gly ss a "constant rate of tine preference.” The corollary to the umese theoren implies tat non anticipatory models of the firm with a discount function (3) will be consistent. Thus, even if the mnagement planner changes the firm's decision horizon and reconsiders the primordial plan, the primordial plan will still be implemented. Nonanticipatory models of the coneumer with a dtsccunted, additively separable utility of ‘the form (30) will also be consistent (as noted ty Strotz) because ‘the constant-rate discount function used in them ie: tet tar 1 =to = ay Taya) t-t Gines (31) (t,t) = 3 and ts thus of the sane formas (3), with 4g = 1, for all t of course, even if a consumer's utility function is additively sep- arable, it 19 posstble that the consumer's discount function will not be of the form (19). In this event, an economist who admits the possibility that the consumer will change hie horizon and reconsider his primordial plan mst then also admit that the primordial plan is inconsistent and will not be implemented. Strotz's claim that inconsistency of primordial plans leads to undesirable myopic behavior has more validity in models with future planners other than the primordial planner. ‘their decision horizons differ from that of the primordial planners In these models, even if all future planners agree with the primor—-21- diel planner's choice of performance criterion, and with her assessment of the state equations and constraints, they still my not implement the primordial plan unless r(t,t) satisfies (19). If one believes that the primordial plenner's choice of horizon is, in some sense, the correct choice, then the inconsistency certainly seems to cause undesirable myopic behavior. Optimal grovth modele that maximize social welfare over long or infinite horizons mist admit the existence of future planners. ‘Inconsistency of their optiml plans is likely because ‘they are often proposed for use in planning environments where it is impossible to precommit the behavior of future planners and because there is seldom any good reason to believe that discount- ing will be of the form (19). Several theorists assume that the social welfare function is of the form (30), in which case the op- tim plan is consistent. But the price that these models pay is the same paid by models of individual agents thet assume this form. These theorists mist rule out the possibility that the planner's preference orderings may not be representable ty social welfare functions of that form Xoopmans (1960) has found axioms characterizing preference ordering that are representable by (30). No one has determined whether these axioms are plausible in positive models or desirable in normative models. Anticipatory Behavior as @ Cause of Inconsistency Anticipatory behavior as a pervasive cause of inconsis- tency has been stressed by Kydland and Prescott (1977, 1980), Fischer, and the introductory article in Incas and Sargent (1981).~22- The concept can be fully illustrated ty the simple, additive separable, anticipatory problem illustrated in Figure 3. The planner's problem is to: (32) max Up(Xq Dp) + 0, (X, ,D,) + 1(x,) Dos. or Bate Xy = feol(XqsDqDy), Xp = £y(Xz,D,)« Substituting the state equations in the objective function, prob- Jem (32) 4s to find DX and DE solving (33) mx U6 50g) + Uy (f(s sD, IsD,) + Unt C.D,» 07 The solution of (32) will, in general, be inconsistent. The problem is undiscounted, so consistency requires that (33) be decomposed by the Bellman principle: (34) max [Up (Xp Do )#maxlU, (x#,D,) + Up (2, OD, )) 11, D D, 0 1 vhere (X$,x$) is the state trajectory determined ty f, fy, and the optimal plan DB, Df. However, (3h) does not follow from (33), so (33) is not I-consistent. ‘This is because e future planner's choice of D, at ‘tame period 1, taking the state Xf as given, will solve the inner maximization in (34): (35) aa 10, Oxf) + 0 (4,0) ))1- a The result is @ decision D, that, unlike the optinal plan's D&, takes no account of the indirect impact Dy has on U, through X, = fo(XqsD9sPy)+ Thus B, # DK, so the optimal plan is inconsis~ tent. This is obvious from Figure 3, where we see that a future~23- planner at time 1 who accepts X¥ as given will, of course, ignore the effect D, has on %y. ‘The future plenner worries only about the effects Dy has on Uy, Xo, and Dox Kydland and Prescott argue that the assumption of per- fect foresight or rational expectations in merceconomic planning models introduces such anticipatory phenomena into them, which cause the primordial plans to be inconsistent. In our paradigm, @ macroeconomic planning model has the government choosing policy variables 0p,+++,Dp to maximize some performance criterion U subject to the behavioral rules of firms and consumers, as given in the form of state equations. These rules take the general form: e e (36) yyy = Py lKysDooeee Dy sDeyg reer sDE)y t= Opens y Watch means that the relevant behavior of firms and consuners in the next perfod 1s sone function of past behavior end government decisions as well as their expectation of future government deci- tons D2, se+esD8+ the assumption of perfect foresight in the determintetitc model above 1s that: . 3 for each t = - (37) Peay = Peay? J = leressTots for each t = 0. T-1, which creates the most general possible planning problem (1) considered here. Unlike the previous example, the state X, in gach period t depends on future decisions. Then, not only is the optimal plan inconsistent, it 1s not even t-consistent for any te Furthermore, this negative result still holds when the assumption of perfect foresight is weakened. Yor example, con-~2h- sider the myopic perfect foresight case illustrated in Figure h. There, the behavioral state at time t depends not on all future decisions, but only on that decision mde in period t + 1: (38) Kear = Fe % Pe Pear) = £0 2D Pear)» ‘The same thing that happens in Figure 3 happens in each pertod, so that an optiml plan for an economy with myopic perfect foresight is neither consistent nor t-consistent for any t > 0. In fact, any systemtic relationship between expected and actual future decisions my cause inconsistency. In more sophisticated stochastic models where future expectations are randon variables, the rational expectations Typotheste thet D2, equals the mean of Diyy will cause inconsistency for the same reasons As long as there is sone systematic connection between current, subjective probabilities of future decisions and the objective probabilities of those decisions, inconsistency my cceurs Kydland and Prescott (1977) give some microeconomic examples of anticipatory phenomena, vhich fit into the simple two period model of Figure 3. For example, they mention the case of optim government patent policies. There, it is optim to offer a mumber of patente (Df) initially to induce an optimal level of inventive activity (xt). But in period 1, after these inventions are created, it seems best to remove all the patents--leaving no (B, = 0) patents, rather than leaving DY = D& patents--o that the inventions can be produced ty competitive mrkets rather then by monopolies. Inventors who expected this to happen (i.e, Dy =~25- D, = 0), however, would not have undertaken the optiml level of inventive activity (Xt). So, the optimal level of patents that should remain is Dt #5, = 0, and the optimal plan is inconate- tent. Kydland and Prescott (1977, 1980), as well as Fischer, @iscuss the inconsistency of optim taxation plens in worlds where firms have future determined expectations. However, they offer radically different remedies for this problem The former advocate Strotz's first course, ie., thet future decisions be precommitted ty the primordial planner. They propose to do this by having the primordial planner substitute simple policy rules— wherein future planners are bound to use predetermined rules to make decisions only nominally under their control--for the discre- ‘ion thet future planners would otherwise have in making deci~ sions. Although these rules are clearly suboptimal from the primordial planner's point of view, Kydland and Prescott's simla- tions show that these rules appear to outperform the suboptimal solution thet would othervise be implemented at the discretion of future planners. However, as Fischer pointed out, there will generally still be an incentive for future planners to break these suboptimal (from their viewpoint) rules. Of course, one might argue that the primordial planner could adopt constitutional or other binding constraints to force the future planners to follow these policy rules. But if that were feasible, why couldn't the primordial planner adopt constraints to prevent the optimal plan itself from ever being changed?~26- Rather than advocate the substitution of simple rules for discretion, Fischer contends that the use of discretion my be warranted on occasion. The patent example cited earlier can be used to illustrate Fischer's reasoning. If the future planners repeatedly revoke patents granted earlier to investors, then eventually all future inventors would expect this to happen and would cease inventing. The repeated revoking of patents in a rational expectations environment is thus undesirable. However, the occasional revocation of patents that would have produced vast monopoly profits will probably not reduce total inventive activity very mch, end it will produce mch social benefit through lowered monopoly profits. Fischer raises the possibility that randomize- tion of decisions my be of use in devising such e strategy. In the patent example, the government might randomly cancel patents, using some preannounced and expected probability distribution to do soe The resulting outcome of this stochastic plan my be etter than that of a simple, deterministic rule, particularly if the inventors most willing to bear the risk are also the most likely to produce valuable inventions. Once again, though, why wouldn't future planners have an incentive to change a previously adopted randomization rule? All of the problems created by time consistency due to anticipatory phenomena are the result of a fundamental paradox. ‘This paradox stems from the logical inconsistency of the following three assumptions common to analysts' models of optim planning with anticipatory phenomena:-2T- (A1) Assume that agents correctly anticipate future planning actions (decisions, policy rules, randomi- tation rules, etc.) that are relevant to their om welfare. (a2) Assume that future planners can observe the agents! past anticipations of current and future plenning actions. (43) Asoume thet future planners have the discretion to implement actions that vere not anticipated in the past ty agents. Time inconsistency implies that planners endowed with the povers of (A2) and (A3) will actually implement actions that were not anticipated by past agents, thus violating the "rational expecta~ tions" assumption (A1). Alternatively, if an analyst insists that (a1) and (a2) mst be valid, then (A3) mst be violated; i.e., the analyst cannot admit the possibility that future planners have the discretion to change the primordial plan. In this case, future planners have no "free will" to act if agents can correctly antic- ipate these actions beforehand. They are preordained to follow prior anticipations that, because of time inconsistency, vould not be realized if planners did have "free will." Because of this, it is difficult to understand the logic of estimating the effects of olicy regime changes" in models (e.g-, Iucas and Sargent 1961) that purport to have rational expectations operating indefinitely ‘both before and after regime changes, tut which assume that the planners have free will to change the regime to one unanticipated earlier.~28- Nonseparability as a Cause of Inconsistency Even in problems without anticipatory elements, a non- separable performance eriterion can lead to inconsistency of the optinal plan. To see this, consider the simple, undiscounted two- period problem below: (39) mx U(Xy sPo oXq sD, ) 0°? Sete % = fo(%qsDq)s Xo given Dy © Cy(X,)» Consistency requires that the Bellnan principle apply to ‘tthe optimal solution, ee, that for any fixed Xo: (40) ome U(X Dp 2% 2D) = mex mx U(X so sq (Xp 2D) D, ) 0a 0 Unfortunately, it is well kmown that the Joint maximization of a function of two variables cannot always be decomposed into the two one-variable problems that consistency requires it to be. As mentioned earlier, Mitten (1964) found simple sufficiency condi- tions (9) on the form of U thet guarantee that Bellman's principle applies. The additive separable case (5) satisfies these condi- tions, as does the mltiplicative form . (41) U(Xq sDysXq 9D y see sXopeDep) = oo Uy (X,»D,)> when the U's are constructed so thet U, > 0 for all Dy © Cy(Xp)« Tt is not at all clear that optimal grovth problems will, or should, have additively separable utility functions. If the utility function is not even monotonically separable, then the primordial plan will, in general, be inconsistent, even though discounting is not used.-29- Conclusions We have seen that any of the following elements will ead to inconsistency of primordial plans: 1. Discounting with a discount function not satisfying (19). 2. Anticipatory elements, such es expectations that are, in some sense, determined ty future policy decisions. 3+ Nonseparable performance criteria. T have argued that discounting my not be of the appro- priate form in optimal growth problems. Anticipatory elemente seem likely in any economle problem where controlled agents have oth an incentive to forecast future decisions and the skills to forecast in some vay that is systemtically related to the actual future decisions. Wonseparable performance criteria my occur in problems involving intertemporal preferences over future consump- tion streams, unless one is confident that the restrictive axioms needed to justify additive separable utilities characterize actual or desirable behavior.(Hy) + Catty)hg ra] (ty) fa dq he woTqoug AuozedyoTqueUoN oTqesedeg eATITPPY Z ouneTA (ee x 0% ) g Cd] 2) f | Cax)?s Fa, Ir Gos4a) seth gig wetqoug AvoqedrorqueuoN | ounsT.gCos%q) 211909 (M5) ata ee Fen (tata + (Paty) Orem t i [ [ Thao) F EH a4) Fs FR aetoa)s FE (048) %s FE I | | Te TT a | 144, ‘a = g(a) squatsosog s00ss0g oTdomy y ounsta (%)?n (atx) a ('a80%)%s fF ty T Ye worgoug Auoqedrorquy eTqesedos earaTPPY £ cundrg=30- References Bellman, R. 1957+ _Dynamfe programming. Princeton, Me Princeton University Press. Burness, Ee S. 1976. A note on consistent naive intertemporal decision making and an application to the case of uncertein lifetime. Review of Economic Studies 43: 5h7-l9. Calvo, G- 1978+ On the time consistency of optimal policy in a monetary economy. Heonometrica 46: 1411-28. Fischer, 8. 1980. Dynamic inconsistency, cooperation and the benevolent dissembling government. Journal of Economie Dynem- ies and Control 2: 93-107. Koopmans, T. 1960. Stationary ordinal utility and impatience, Econometrica 28: 287-309. Kydland F., and Prescott, F. 1977. Rules rather than @iscretion: the inconsistency of optim plans. Journal of Political Economy 85: 473-91. + 1980. Dynamic optim taxation, rational expectations, and optimal control. Journal of Economic Dynamics and Control 2: 19-91. Incas, Re, and Sargent, T. 1981. Rational expectations end econometric practice. Minneapolis: University of Minnesota Press. Mitten, Le 1964. Composition principles for aynthesis of optinal miltistage process. Operations Research 12: 610-19+ Nemhauser, G. 1966. Introduction to dynamic programming. New York: John Wiley.-31- Sargent, T. 1979. Macroeconomic theory. New York: Academie Press. Strotz, R. 1956. Myopia and inconsistency in dynamic utility maximization. Review of Feonomic Studies 23: 165-80.