Optimization in Economic Theory Second Edition BY AVINASH K. DIXIT OXFORD UNIVERSITY PRESSOsi University Press, Walton Street, Oxford OX? 6DP Ovfand New York Toronto Delhi Hombay Cuesta Madras Karachi Peraing Java Singapore Hone Kong Toke ‘Nairobi Dares Salaam Cape Town Meliowne Auetland and asselted companies ‘Berlin Inada Osford is atrade mark of Ovfrd University Press Pubiishd nthe United Stes nv Oxford University Press Ne York Avinash K Dit 1990 ie pabhisod 1990 aperivak eprnte (90, 1991 (38) Allright reserved opr of this publication maybe reproduced, stordindreticul yr orton, ams formor bans meas ‘lecinmt. mec lune phon, recondng orohereise, witha! ‘he prior pcan of Oxford Unversity Pree Tsk wl jee the canton that sal tb way oftrade on viernive hele, 2-0, red oat otheregeeratated ‘is te pubte's pe onset inform of nding cove? cenier thant stich ii publsher al witht asia condi Inlaid biped nt he subsequent purchase? Iii Library Celine in Pablication Data Dis Avinash, K Oprimicaion economic theory 1 Esonomies Opinion Tide 28001515 1SBNO-19-877211-4 ISBN 0-19-877210-6 (PAK) and, ibrar of Congres Cataloging in Publication Data Data no avalable Typeset by A. K. Dir using TEX Pet and Bound Groat Brainy Baas ta, ‘Gudford and King's Lon PREFACE ‘Making optimal use of scarce resources, that is, maximizing subject to constraints, is the central theme of economies. But students of ‘economics axe often taught the mathematies of constrained maxi- ‘ization as a branch of mathesuaties, and its economic applications follow separately. An integrated treatment that relates the math fematies to the economics from the beginning has the potential for providing a quicker and deeper understanding. This book aims to give such an exposition. Temphasize economic intuition rather than mathematical rigor. Proofs of the mathematical theorems are structured to bring out points of economic interest and facilitate economic applications. The illustrative examples and exercises are also chosen for their economic interest and usefulness. ‘The first edition of this book was published in 1976. The een- tral aim of the book is still valid, but the subject hins changed great deal over the years. Therefare I have revised the text very substantially. A chapter on uncertainty, with some treatment of topies like Snance and asymmetric information, is now indispens- able. [have added such a chapter, and have also expanded the chapter on dynamic programming to treat uncertainty. ‘Most chapters have been thoroughly rewritten, and many new examples and exercises added. One innovation deserves special mention. When the first edition was written, the main mode of ex- position in elementary and intermediate microeoonomnies was geo- metric, based on the tangency between a budget line and an indif- ference curve, oF a cost line and a production isoquant. Nowadays this shibboleth of tangeney seems lese prevalent. Thevefore I have used a starting-point that is simpler and economically space intu- itive, namely the search for costless improvements through ‘arbi- ‘rage’ operations. This allows an integrated treatment of tangency and comer optima, and its intuition extends much mote readily to situations involving time, uncertainty ete. In the years since the frst edition of this book was published, the mathematical training of economics students his improved sub- stantially. [have taken advantage of this by going a little deeper into some topics, letting the pace pick up in the last three chap- ters, and sketching the proof of the central result of constrained maximization the Kuhn Tucker theres ~ in a mathematical ap- pendix. But the book remains aimed at the majority of economiesvi Preface students in the last two years of undergraduate studies o the first year of graduate work, not the small minority who plan to become ‘uathematical economists, ‘The main use of the book is as a supplementary text inn crocconomies courses at the intermediate and advanced levels, Tt can be used as the main text, but should then be supplemented by fother books or articles, inchiding those listed for fusther reading at the end of each chapter. Short courses on consumer and pro- ducer behavior ean be based on chapters 1-8. Chapters 9 snd 10 are independent of each othe. Therefore courses that include op- timization over time but not uncertainty can omit chapter 9 and parts of 11. Conversely, those that include uncertainty but not time can omit chapters 10 and 11 ‘The examples ae fully solved out as far as the main line of reasoning is concerned, but some details are omitted. ‘The exer- ‘isos contain further development of the theory in the text, as well ‘as applications. The examples and exercises are an integral part of learning from the book, and I urge readers to work carefully through them. T thank many readers who gave me useful suggestions for im proving on the first edition, Among them, Pete Kyle deserves special mention, Tasn very grateful to Richard Quandt for reading the entize manuscript of the second edition with great eare and pointing out several errors, [also thank Barry Nalebuff and Carl Shapiro for reading several chapters. Peter Kenen’s Lav says that there is always at least one more typo than yor think, and I retain responsibility for the errors that rexnain. Princeton AKD. December 1989 CONTENTS 1. Introduction 2, Lasgrange’s Method 3, Extensions and Generalizations 4, Shadow Prices 5, Maximum Value Punetions 6. Convex Sets and Their Separation 7. Concave Programming 8. Second-Order Conditions 9. Uncertainty 10. Time: The Maxinum Principle IL. Dynamic Progeaaming Appendix ~ Kuhn-Tucker Theorera Index 10 4 40 35 69 86 105 122 M5 162 asl 1871 Introduction Feonomics has been defined as the study of making the best use of scarce resources, that is, of maximization subject to constraints ‘The criterion being maximized, and the constraints imposed on the choiee, vary from one context to the next: houseliolds’ co sumption and labor supply, firms’ production, and governments? swolicies, But all constrained maximization problems have a com mon mathematical structure, which in turn generates @ common economic intuition for them. This book sims to outline the math ‘matics and develop the intuition ‘The standard model of consumer's choice provides a good point of departure. The basie concepts are a budget line and a set of indifference curves. ‘The points on the budget line represent all affordable combinations of two goods. The family of indifference ‘curves represents the objective, namely to reach as high a curve as possible, The optimum is where an indifference curve is tangential to the badget line, Figure 1.1 shows the failiar picture Tn this chapter I shall develop this model further, using, ver~ bal and geometric arguments, but with an eye toward the math- ematical sharpening and generalization that will occupy us in later chapters. Tt helps to give a little algebraic content to the various mag- nitudes in Figure 1.1. Write I for the consumer’s inoney income, x and pe foc the prices of the two goods, and zy aud 1 for their quantities. The budget line, where the expenditnze exactly equals ucome, can then be expressed by the equation may tree (2) ‘The consumer's preferences over the amounts 2 and 2 of these goods are sepresented by a numerical scale, called the utility function, This assigns to each bundle (21,72) of goods & nun ber U(rs,22), called its utility level. In any comparison among, alternative bundles, the preferred bundle is the one that receives the highest utility number, Along an indifference eusve, all points2 Optimization in Economic Theory indifference ‘optimum, choice budget Tine Fig. 1.1 ~ The consumer's optimum choice ‘must have the same utility number. Therefore such a curve has an equation U(x1,22) = constant a2) ‘The exposition of the theory now proceeds by calculating the slopes of the budget line and the indifference curve. For tangency between the two, the slopes imust be equal. I shall do this soon. But let ine begin with a much simpler and more intuitive approach, ‘The Arbitrage Argument ‘The idea is to have the consumer start with any trial allocation of bis budget, and contemplate a change. If this leads to a bundle of goods that he rates higher on his utility seale, then it is to be adopted as a new trial allocation, Onec a bundle is found that cannot be bettered in this way, it will be the optimum allocation, a will be the one actually consuined. Thus the impossibility of ‘an improvement will serve as the test of optimality ‘The change does not en merely a reallocation of some amount of money from the purchase any additional expenditures; it is Introduction. 3 ‘fone good to the other. Ifthe initial allocation is not optimal, this ‘an raise the consumer's utility. When the consumer has made the Ibest choice and reached his personal equilibrium, such opporta ities for doing better with no net increase in expenditure vanish. is has a close parallel in financial markets. Ontside a market ceili un, participants ean make ‘arbitrage’ profits ata zero net donthy, taking advantage of discrepancies in prices of the same asvet in different markets. In equilibrium there ae no such opportun- ities. In fact the very process of people seeking arbitrage profits Iwings about the equilivium. I shall exploit this intuitive parallel labeling this whole fine of reasoning the “abitrage arguinent’ au the resulting optimality condition the ‘no rbitrage condition” If goods are indivisible, these changes aust oceu in diserte sieps. However, itis oft @ good approximation to suppose that oods are pevfectly divisible. Then t fintesimal amounts, oF what economists call maeginnl adjustments ven for sceningly indivisible goods, sch as ears or other large asumer durables, there are dinuensions of qty ete. that allow continnons adjustment. In any event, seh ssarginal changes ave hiages can oceur it in- the subject of our analysis in thin bovk “The standard syusbol for “a sinall (manginal, infsitesimal) claange in the variable 2” dz, This i not to be thought of as the product of two variables d and 2, mt an entity dein its own fight. ‘The use of such infinitesimal magnus ean be justified Figorosly, but for the must part Tokall use them in a loose heuts {ie way. Where you doubt s statement involving infitesimals, or are unsure Thave used thens correctly, you shoul rowork the ag cat ssing proper calculus method, starting with a finite change Asrand then going tothe linit av Ae +0 First suppose tat inital location of tie bulge has postive amounts 35 and of both goody. Now eontenplat a sonal ari trage operation, or a marginal relocation ofa small but postive suinount of come df from goo 2 to good 1. In physical tees, this means buying dI/p, units wiore of good 1 and dI/pz units less of good 2. Let MU, and MU; denote the nneginal tities ofthe toro goods. This means that a stall change dry in the quantity of good I changes tity by MU dn, nts and snilanly MUs dea for good 2. When the quantities of hoth goods are changing, the two effects ean be added together, so the change in uility is MU; dry + MUzdry 3)4 Optimization in Economie Theory Iu ater chapters I shall express this more rigorously using partial derivatives and Taylor series, but the simple statement will sullice here. The important point is that any one marginal adjust: rnient is so small that any changes in the marginal utilities then selves during its course can be neglected. Of course if many sueh narginal adjustments are strung together, the marginal utilities will change gradually over this sequence, In particular, if 2 rises end/or x falls, MU2/MUy will fall; this is the principle of the diminishing marginal rate of substitution in consumption. But for the moment Lam speaking of just one marginal adjustment. ‘The effect of the arbitrage operation on utility is then easy to compute. The increase of dI/py in the quantity of good 1 raises utility by MU dI/p,, while the decrease of dI/py in the quow of good 2 lowers utility by MU; dT/p,. The net increase in utility is therefore (MU fp — MUz/p2) a. If this expression is positive, the consumer will eaery out Ys reallocation and try further reallocations in the same direction. If the initial consumption bundle is optimam, therefore, the expres: sion cannot be positive. This is a part of the nno-arbitrage’eriterion of optimality. Since dI was chosen positive, we can divide by it ‘and write the criterion as (MU, Jp — MU2/p2) <0. (aa) Now suppose the etiterion is uot met, that is, suppose the left-hand side expression in (1.4) is > 0. Therefore some switch of expenditure toward good 1 is desirable. How far should this process go? Recall that as 21 increases and 22 decreases, MU; will gradually fall relative to MU, Eventually a point will be reached sthere the expression in (1.4) is zero, and no facther move jn this direction ean raise the consumer's utility level Next consider a reallocation in the opposite dvection, This will switch the signs in all the above arguments. If the initial allocation is optinmim, we must have (MU; /p, ~ MUz/p2) > 0. (3) If this is false, that is, if the expression is <0, then the process of increasing x and decreasing 21 will be carried out until the ‘expression reaches zero, Introduction 8 We can combine the two criteria of optimality of the initial consumption bundle into one: ifthe allocation (11,72), with both ‘quantities positive, is optinmum, thes the marginal utilities at this ie amust satisfy MU; /p1 = MUa/ps (16) "his is the overall ‘no-arbitrage’ condition. ‘The economic interpre: tion is that at the optimum the consumer should be indifferent hetween allocating the marginal unit of money to the one good or he other ‘The Tangency Condition Using Calculus Phe more complex but more commonly used way to derive the same vunudition of optimality is based on the tangency of the budget line (1.1) and the indifference curve (1.2), Write the equation of the Indget line a5 = (Lp2) ~ 21 (ps/p2). ‘Vhwu we see at once that the slope of the line is (pa /ps) in numerical value, The slope of the indifference curvo is the marginal rate of substitution (MRS) in consumption, and it equals the ratio of the ‘lites (MU; MU), A heutistic derivation of this is a= follows, Ifa marginal iss of dy units of good 1 is just compensated hoy the marginal gain of dza units of good 2, then the marginal rate of substitution (RS) is the ratio dra/dzy. But the exact offset vf the gain can be written as an equation in utility units snanginal MU d2y = MU; des, MRS =dra/ley MU /MUs. (az) Incidentally, note the reversal of the subscripts 1 and 2 between Ute mumerator and the denominator of the two ratios, This is not ‘typographical error; that is how the ratios are velated, as you east see by cross-multiplying to get back to the equation just above, At the optimum, the slope of the indifference curve (the mar- inal rate of substitution in consumption) is equal to the slope of the budget line (the price ratio). Therefore MU,/MU2 = pals (8)6 Optimization in Beonomic Theory This is equivalent to the optimality criterion (1.6) we derived. before. However, the verbal argument underlying (1.6) has some advantages over the geometry of (1.8). Corner Solutions ‘The arbitrage argument is somewhat easier to adapt to the case where one of the goods is not bought at all. Since this occurs fat one end or the other of the budget line, such optima are often called comer solutions. Suppose all income is spent on good 2 in the initial allocation, so xy = 0 and r2 = I/p2. Now of the two directions sf improvement, namely trying to ineyease x, and to decrease it, onily the former is possible. Therefore only the criterion (1.4) corresponding to this test survives. Since x, cannot de decreased any further, the argument leading to (1.5) cannot be made, The economic interpretation of (1.4) holding for the consumption bundle (0, /p2) is equally simple: even the first little unit of income spent on good 1 does not bring enough utility benefit to mntch that from the last unit spent on good 2 Marginal Utility of Income ‘The second advantage of the azbitrage argument is even more im portant. Retum to the situation where both goods are initially ought, and the eriterion (1.6) for optimality. Now suppose our consuiner is given an extra amount d/ of ineome to spend. He could spend it all on good 1, buy (df/p1) more units oft and achieve an additional (MU; d1/p,) units of uility. Or he could spend it all fon good 2, when his utility would increase by (AfU, di /p2) units. But the two increments to utility are equal by (1.6). Therefore at the margin the allocation of the infinitesimal amownt df of extra income to good 1, of good 2, oF indeed any mixtute of the two, is @ ratter of indifference to the consumer. Then we cai call the util ity increment per unit of marginal addition to income simply the marginal wility of income, without hothering to specify how the :narginal addition to the income is spent. Write \ for this marginal utility of income. Then the dF units of extra income raise utility by AAI units, Equating this to the two other ways of writing the same uliity gain in terms of spending on each of the two goods, wwe Rind d MU, |p. = MU2/p2. 9) Introduction 7 We sce that the common valwe of the two sides in the opti- tality criterion (2.6) has a very useful cconoric interpretation — it is the marginal utility of income. 4 very similar interpretation is possible for the no-arbitrage conditions in all constrained opti ‘uicstion problems; [shall explain this in greater detail and snify it muder the concept of Lagrange multipliers or shadow prices Many Goods and Constraints he generalization of the nnalysis to cover eases where there are several goods, some of which may not be bought at all, is equally easy. Suppose there are n goods, with prices (p,,p2y.--Pa) and jquntities (21, 22,...2n)- For all goods bought in positive amounts tl the optimum, the ratio of marginal ntility to price must have mon value, which ean then be interpreted as the marginal nlilty of income A. For all goods not bought, the ratio of marginal ulilty to price snust be smaller than, or at best equal to, this vale, In symbols, for any good 4, ita >0, wed es tnt van {ep eet (1.10) The method can also be extended to allow several constraints. We needl a separate ) for each constraint, and it can be interpreted us the marginal utility of relaxing that constraint, T shall diseuss this in Chapter 3 Non-binding Constraints a consumer theory, but will prove very important lly, consider an extension that is not of great relevance in some other ap: plications. Picture a consumer with aa income 50 Inege that he is satiated, and fails to spend it all. ‘The budget equation (1.1) should be replaced by an inequality Posi t pam <1. We can bring this within the scope of the above theory simply by defining a new good 23, ‘unspent income’, which has price equal8 Optimization in Economie Theory to one and yields no utility. (Note that I am not talking of saving, which enables a non-satiated consumer to spend more on desired goods in the future, but of totally useless unspent income.) The Duddget equation becomes Pin + para tay = \dwehave MU} = 0, Since we are supposing that the consumer is choosing a positive amount of 2, (1.10) for i= $ gives A= 0. This makes intuitive sense: if the consumer does not even spend all the income he has, then the marginal wtility of an inerement to income should be zero. In tuen, we can use this in (1.10) corresponding to the other goods, aud obtain MU; =0 for i = 1 and 2, Thus these goods are consumed at a level that yields zero marginal utility, that is, to the point of satiation. We started on a simple and intuitive excursion into a con: sumer's choice problem, using nothing but the computation of ‘ar Ditrage’ gains and losses from marginal adjustments and building them into criteria of optimality. This has already brought 1s to a very important general way of characterizing optima subject to constraints, In faet (1.10) is nothing but a form of a basic result of the theory of optimization subject to constraints, namely the Kuhn-Tucker Theorem, The extension to the case of a satiated consumer is an instance of the general principle known as Comple: mentary Slackness. In the chapters that follow, I shall gradually develop the general theory in a more systematic way, making use of the ealeulus, to develop and to sharpen the intuitive ideas in troduced here Preliminary Reading To read this book with ease and profit, you should be farnil iar with the basic concepts of elementary or intermediate micro economics: transformation curves and indifference eurves, margi- nal calculations done verbally or geometrically, ete. WF you need, to acquire this background, any good introductory economics or intermediate microeconomics textbook will serve. T mention just ‘one of each kind WILLIAM J. BauMOL and ALAN S, BLINDER, Economics Principles and Poliey, Chicago: Harcourt Brace Jovanovich, 4th. eda., 1987. Introduction 9 HAL R. VARIAN, Intermediate Microcconomies, New York: Norton, 1987, The mathematics needed is at the college calculus level. You will noed knowledge of partial derivatives right away in Chapter 2. Vector and matrix notation, with simple sums and products of dese entities, appears in Chapter 3, Some concepts of convex- ity axe developed in the book ax needed. Definite matrices and saundtatie forms are required in Chapter 8, and some elementary integration in Chapters 9 to 11 "Those who weed to acquire some of these mathematical tech niques have the choice of proper mathematics books, mathematics Innoks designed for economists, and economics books which explain ‘the mathematics along the way. Here once again is one example of wach Kind. SERGE LANG, Calculus of Several Variables, Berlin: Springer: Verlag, 3rd edn., 1987, chs. 1-6, 13-16. ALPHA C, CHIANG, Fundamental Methods of Mathematical Beonomies, New York: McGraw-Hill, 3rd edn, 1984, ehs. 4, 5, 7, 8 ALAsDaih SMITH, A Mathematical Introduction to Economies, Oxford, UK: Blackwell, 1982.2 Lagrange’s Method Statement of the Problem Let us begin the development of the subject to constraints, using a setting very close to the consumer choice model of Chapter 1. Suppose the choice variables are 2, and 22. Ushall write them more compactly as a vector 2 arranged in a column, aS (2). Initially, I shall use vectors only to abbreviate lists of components: actual operations with vectors and matrices will appear gradually We will need notation that distinguishes between the general vector x and some particular value of such as the optimum, while remembering the family resemblance between the two: both are vectors of choice variables. I shall generally use the symbol F to denote the optimum value of the general variable x. ‘The -neral theory of optimization components of 2 will be written Z) ete ‘The function to be maximized, called the objective function, is F(x), The constraint is a general non-linear ene (2) 1) where Gis function and ca given constant. The model of Chapter 1 was a special case of this: F was the utility function, G was a linear function showing the expenditure, Ga)=pintprre, and ¢ was income, If it helps, you can now think of @ as a more gcneral now-linear expenditure or eost fanction, such as would arise if the consumer faced a quaattity diseount or premium price sched- ule, With this notation, the problem in this chapter is to find we value # that maximizes F(x) subject to G(x) = ¢ Lagrange’s Method u ‘The Arbitrage Argument As we did in Chapter 1, start with a trial point and an infinitesimal siyuge, Let the particular trial point be 2, and the infinitesimal oe a= (%) We want (+ dz) to satisfy the constraint, and see if it yields a higher value of the objective function, Since the change in is infinitesimal, we can approximate the changes in values in functions of x by the first-order linear terms in their Taylor series. Let subscripts on # function denote ity partial derivatives with respect to the indicated argument; for example F, is OF/Ox). We must remember that in general each partial derivative is itself a function of the whole vector x. Only in the special case where F is linear will the pactial derivatives be constants; only in the special ease where F is additively separable (function of 21 plus a function of zz) will F, be independent of irz and vice versa. Thus we shioald write the pattial derivatives as Inetions F(x) and F(z). When these are evaiusted at our initial tsial point £, their values will be written Fi(z) and F,(2). ‘The first-order Taylor approximation of the change in F(x) 1s a result of the infinitesimal move from the general point x to ride is F(z) = F(x + ds) — Fle) = Ala) de + Rlz)dra (22) Observe the similarity with the expression for the change in util: ity (1.3) of Chapter 1; the marginal utilities that were motivated there by economic intuition are simply the partial derivatives of the utility function with respect to the amounts of the two goods consumed There is a similar expression for the change in Gt) G(x) = Gale) dey + Ga(x) dea (23) in Chapter 1, the partial derivatives of @ were simply the prices If we now think of G asa more general non-linear outlay or expen= dituse fanction, the partial derivatives are the marginal prices of the respective commodities.2 Optimization in Beonomie Theory Now we can modify the arbitrage argument of Chapter 1 to apply to the new more general setting. Start at a point # where the constraint (2.1) olds, and consider a change dz such that (2-+dz) also satisfies the constraint, Then @G(2) = 0. Using (2.8) with the particular initial point, we have Gi(@) dry = ~G,(2) dro. Call the common value of these two sides de. Then our arbitrage consists of reallocating an amount de inthe value of the funetion G(x) away from zy and towed 1, Fit suppose G,(2) and G,(z) are both non-zero. Then dx =defG(z), and — dry = de/G(z) ‘The resulting change in the value of the objective function is foxnd by substituting into (2.2) as AP (z) = [Fi(@)/@\(8) — Fel 23/Gxl)) de (24) If the bracketed expression is non-zero, then F(2) ean be in: creased by choosing de to have the same sign as that of the brack- ‘ted expression, AS before, we can turn this around te find a no-arbitrage condition that holds when 2 is the optimum choice So long as neither 2 nor 2 has hit some natural boundary such, fas zero, then changes de of either sig are possible. If x is opti um, then uo such change should be capable of increasing F(z) Therefore the bracketed expression in (2.4) should be zero, or Fy(2)/G,(2) = Fx(2)/Cx(2). (25) This is the analog of the condition (1.6) of the previous chapter. Note the exact statement: if the optimum choice is 2, then it satisfies (2.5). I have not established any implication the other ‘way round, so there is no guarantee that a solution to (2.5) is the optimum. ‘This is the difference between necessary and suficient conditions, and I will discuss it in more detail later in this chapter If # lies at some natural boundary, for example if one of 21 and Fy is zero when both must be non-negative, then only one sided changes de are meaningful, and we get an inequality that Lagrange's Method 13 orrespouds to (14) and (1.5). [shall not consider this case in this chapter, but shall return to it in the next. As in Chapter 1, let us define A to be the cesamon value of ‘he two sides in (2.5). Then we ean write that equation as a set of {owe equations Fy) = Gz), 12. (28) Remember that the 3 of Chapter 1 could be interpreted as the ‘wouginal utility of income, In the same way, the A just introduced har out to be the rate at which the optimum value of F(2) re- svenuls to. change in c. T shall develop this interpretation and jcations in Chapter 4. ‘The main topic im the rest of this »ter involves writing (2.6) in a way that easily extends to more yeneral settings, and provides a method for finding +, But Grst a couple of necessary digressions, Constraint Qualification What happens if say G,(#) is zero? Now #1 can be changed slightly without affecting the constraint. If Fj (2) is not zero, itis desirable lo so, For example, if F(z) is pesitive, then F(x) can be in- creased by raising 2). This goes on until either Fy(2) drops to zero, ‘8 Gy(z) becomes non-zero. In the cousumption interpretation, a consumer will go on using more of a free good either until he is salinted, of until the marginal unit of the good is no longer free. ‘Therofore if G,(2) is zero and z is optinnum, then F,(2) must be 210, too. We can define the ratio of these two zeros as we please, imud there is no harin in defining it so that (2.5) is satisfied. What if Gi(z) and Gx(2) are both zero? This might mean Hint both goods ate free, and should be consumed to the point of satiation. But there is a more ominous possibility arising from the tunizks of algebra and calculus. Take the budget line (1.1) of the previous chapter, and write its equation as (pity + Pete — 1)? = 0. ‘his is an unnecessarily complicated way of writing (2.1), but the two are mathematically fully equivalent, and we should see if the change makes any difference, Let G(z) be the function on the left-hand side of this. Then Gil2)= 3p (ray + p22 —1),M4 Optimization in Boonomée Theory which is always zero when = satisfies the budget constraint, The same is true for G,(2r). Goods are not free at the margin, and yet thir quantities have zero effect on the constraint function, In such a case, our method runs into trouble, ‘The formal mathernatical theory eops out and simply refuses to deal with such eases by assuming a condition called a Constraint Qualification, In the present instance that simply amounts to ns: suming that at least one of G,(7) and Gz(2) is non-zero. If this is not true in a particular application, then conditions like (2.6) may be invalid there. Luckily, failure of the Constraint Qualifi- cation is rarely a problem in practice. In the rare cases where it arises, it can usually be circumvented by writing the algebraic form of the constraint differently, as with the budget constraint above, But students of the subject should not forget the problem alto- gether; it is a favorite source of trick questions in examinations. If something strange seems to be going wrong when you try the stan- dard methods, you should check to see if the problem violates the Constraint Qualification. But I shall omit further mention of this complication except in the formal statements and mathematical proofs The Tangency Argument ‘The second digression relates the arbitrage argument to the tan- gency condition more familiar from elementary economics texts Tn Chapter 1 we saw an alternative way to obtain the condition (1.6), based on the tangency of the budget line and an indifference curve, The same can be done for (2.5). Figure 2.1 shows the story Along the curve G(r) = ¢, we have dG(2) = 0, aud from (2.3) we can calculate the slope of the tangent to the curve at 2 as dg /dzy = ~G\(2)/Gn(2). 7) Note the reversal of the subseripts, exactly as in Chapter 1. Note also that if G,(z) = 0 the eurve is vertical; this is not a serious problem, If both Gi(2) and G,(z) are zero, the slope is not well defined, and the methoel may run into & problean as was explained just above. In most economic applications, @ is an increasing fanetion of both arguments. Then Gi(z) and Ga(2) are botl pos itive, drz/d2, along the curve is negative, and we have the usual downward sloping transformation frontier for the constraint. Lagrange's Method 6 agent Fig. 2.1 ~ The tangeney solution A contour of the objective function F, tnt is, a curve of equal values F(z), runs through the point x. The slope of the tangent tw the contour at this point is similarly calculated: dry dey = —Fi(2)/ Plz). (238) ‘Once again, in most economic applications, F is an increasing, func tion and the contour (indifference curve) is downward sloping, If is optimum, the two curves must be matually tangential, Unt is, have the same slope, at this point, quating the two expressions, we have Fla\/Falz) = Gil2V/G(2), (29) which is equivalent to (2.5). Necessary vs. Sufficient Conditions Recall what the above argument established: if # maximizes F(z) subject to G(x) = e, then (2.5) holds, In other words, the condition16 Optimization in Beonomic Theory (2.5) is a logical consequence of the optimality of 2. Therefore it is called a uccessary condition for optimality. To be more precise, since it involves the first-order derivatives of the functions F and G, itis called the first-order necessary condition, In searching for an optimum, the first-order necessary eondl tion helps to narrow down the search, The necessary conditions were established starting with an assumed known optiznum 2. But wwe turn the story around by treating the components 4 atd #2 1as unknowns, and the constraint (2.1) and the necessary condition (2.5) as two equations that will deterasine them, ‘Typically, there isa whole continuous range of values of 2 satisfying the constraint (2.1), But there are only a few values of %, and if we are lucky, just one, that also satisfies the condition (2.5) If we know from separate reasoning that our probles indeed hhas a solution, and wo find that there is a unique = satisfying the constraint and the first-order necessary condition, then it must be the solution we seek. If there are multiple solutions to (2.1) and (2.5) taken together, then all are candidates for optimality as far fas the present analysis is coucemed, and some other methiod must be used to find the correct solution. Even then, the first-order necessary condition (2.5) will have eut down quite drastically the nuaber of candidate points we need to examine, ‘The main reason that the first-order necessary condition does not always lead us to the right solution is that the sane first-order consition is also necessary for the problem of minimizing the saa function F(z)subject to the same constraint G(s) =e. Minimizing, F(2) is the same as maximizing —F(z). By the same reasoning 1s that leading to (2.8), we can find the slope of a contour of this funetion: dzz/dr, = -[-F(2))/[-Fa2)] ~Fi(a)/Fale) Equating the value of this at F to (2.7), we get (2.9) again, There is also the point that to obtain the condition, we asked if the value F(x) could be improved by making small changes in 2. If not, then 2 is better than the comparison points in a small neighborhood, or it yields a local peak of F(z). Now a function ean in general have several such local peaks, and several local troughs too, The same first-order necessary condition will be true at all these points. Only one will give a true or global maximum, Lagrange's Method 7 Finally, either ® maximum or a minimum implies (2.5) but vat vice versa. Therefore the condition might be satisfied at os th tieifhborhood. As a simple example, consider F{2) = 2° wher eisalar, We have F"(0) = 0, but « =O gives neither & maximum (2). ly distinguish such eases, any point satisfying the first-order sini is one of the stationary points. To locate it among these ‘inulcates, we need some other test. Such tests typically rely on the eurvatixe, oF the second-order derivatives, of the functions. In Figure 2.1 the curvatutes of the contours of F and G have been hoon correctly for # maximum, ‘Thas the curve Glz) = © gets ‘atten as zy decreases and zy increases along it; the economic in- texpretation is that the marginal rate of transformation of zy into 1 limiishes as more and more of such a transformation is earried is neither a maximum nor a minimum, even in a small hot n minimnan of ¥y conditions is called a stationary point, The true opti vit, Similarly, the contour of F shows a diminishing marginal rate uulstitution, lists involving curvatures or second-order derivatives are the subject of Chapters 6-8. These tests differ from the first-order eon- ‘lions of this chapter in another Way: af such a condition holds, then the point in question is a maximum, at least in comparison with neighboring points; the condition ensures optimality. There- 1w such @ condition is called a second-order sufficient condition, Lagrange’s Method Now let us express the first-order necessary condition (2.6) in a way that is easy to remember and use, This is called Lagrange's Metliod after its inventor, Note that we want to use the condition Iu solve for the optimum #. We introduced A as the common vale nif the two sides in (2.5), s0 i€ is just as much unknown as the vptinuim 2, That is, we have to determine it as an integral past of ‘hw solution. In the meantime, call it an undetermined Lagrange susltipliee, Define a new function, called the Lagrangian, L(2,) = Flr) + Ale~ G2) (220) Ni at (that L is also a function of ¢, and of any other parameters pear in the functional forms of F and G, Such arguments uf L-will be shown explicitly only when they are important in the sles18 Optimization in Economic Theory Denote the partial derivatives of L by L,=0L/02;, Ly =AL/0a ‘Then Lian Fiz) -AGjlz), Late.) G2), ‘The first-order necessary condition (2.6) becomes just Ly = 0 for j = 1 and 2, and the constraint (2.1) simply Ly = 0. Then we ‘can state the result of the whole argument so far into a simple statement Lagrange’s Theorem: Suppose 7 is a two-dimensional vector, ¢ is a scalar, and F and G functions taking scalar values. Define the function L as in (2.10). If x maximizes F(z) subject to Gl2) with no other constraints (such as non-negativity), and if G,(z) ¢ O for at least one j, then there is a value of A stich that Lyz,A)=0 for j=1,2 LA(2,)= en) Remember that the theorem provides necessaty conditions for optimality. In other words, it starts with a known optimum 2, and establishes that it must satisfy (2.11). But in practice, mich of the use of the theorem is in helping us narrow down the search for an initially unknown optimum. We regard (2.11) as three equations for the three unknowns #), #2, and A, The equations are generally non-linear and neither existence nor uniqueness of the solution is guaranteed. If the conditions have no solution, the reason may he either that the maximization problem itself has no solution, or that the Constraint Qualifieation fails and the first-order conditions are inapplicable. If the conditions have multiple solutions, we need the second-order conditions to arbitrate between the candidate sol tions, But in most of our applieations, the problems will be well posed enough that the first-order necessary conditions take ws to the unique soltion. I shall now develop some examples that use Lagrange's method, and offer some exercises for you to attempt similar solutions. After you have gained soune experience of prob: lems with two vasiables and one constraint, you will be ready for the extensions considered in the next chapter Lagrange’s Method 19 While the notation keeps the theoretical developments clear ly distinguishing the general point from the particular optimum, 1 becomes cumbersome in applications where we are searehing for ‘aw unknown optimum, Therefore we often drop the bar on x in ‘nulitions like (2.11) when using thesn in particular contexts, such ‘ws the examples below. Examples Heumple 2.1: Preferences that Ingly Constant Budget Shares Consider a consumer choosing between two goods and y, with prices p and q respectively. (The notation x), 2 ete iu the theoretical part because it generalizes more easily to several sls and constraints, but the x, y notation is simpler in examples With just two goods.) His income is I, so the budget constraint is was used Suppose the utility funetion is U ay) =a In(z) +B Inly), (212) where a, B axe positive constants and hy a sithuns, Write the Lagrangian jes natural og L(xy,d) =a In(z) +8 Infy) +A [1 pa ay Newall that dn()/di + waulitions (2.11) become 1s. Therefore the first-order ne AL/Oz= 0/2 ~rp=0, OL Jy = Bly —dg=0, aul OL/O\=1—p2~qu=0. Tu solve these, substitute for x, y from the first two into the hand, ‘This gives A= (a4 Sil, (213)20 Optimization in Economie Theory sand then al al weap YEG ew ‘These are the demand functions, namely the solutions for the opti- suwuin quantities in terms of the prices, income and the given para micters a, 8. ‘We can write them alternatively as a w_b ats" To atf pe ae (2.15) In other words, for the utity fanetion specified, the shares of in- come spent on the two goods are constants, This is @ convenient property, and one that is sometimes close enough to reality. In the initial exploration of theoretical models, this specification is often the crucial simplification that yields eoncrete results that suggest the directions for further analysis and testing. Therefore this fane tion isa favorite of economists Note that in (2.18) the marginal utility of income is inversely proportional to the income, This might seem a natural conseqitence of the intuitively appealing idea of diminishing marginal utility. But that is a treacherous concept; see Exercise 2.1 below. Bzomple £.2: Guns vs. Butter Consider an economy with 100 units of labor. It ean produce guns © or butter y. To produce + guns, it takes 2? units of labor; Tikewise y* units of labor are needed to produce y guns. Therefore the economy's resource constraint is ay? = 100. Geometsically, you can easly see that the production possibility frontier is a quarter-citce. The objective function to be maximized is Flz,y) = ar + by. where a, 6 are given positive constants ‘To solve this problem, form the Lagrangian U2,y,d) = a2 + by +9 (100-2? —y?), Tagrange's Method a he first-order conditions are OL/Oz = a~2dx ab/ay =6-2Ay =0, aLjar = 100 — Substitute from the first two into the third to get ays 100 = (a? + 8)/(4)*), A= (a? 467)'/2/20, The 10 a/fa? +88)", 10 d/(a? +B), (2.16) You can think of a, b as the weights or social values attached to the two goods, and then (2.16) gives the economy's optimal applies as functions of these weights. Ifboth weights are incscased. iw equal proportions, say doubled, then the optimurn quantities ‘sud y are unchanged. The supplies are homogeneous of degree veto in the values, so only the relative values matter. ‘The supply ‘ofeach good ineseases as its relative value increases. In later work, ‘expecially the chapter on comparative statics, we shall see bow renerally valid such properties are. xercises jercise 2.1: The Cobb-Douglas Utility Function {der the consumer's problem as ix Example 1, but with a dllfesent utility fanction 0 defined by Gaary’ Show that it yields the same constant-budget-share demand fane= tions (2.14) as above, (Hint to simplify the solution proces: eli nuute the Lagrange multiplier between the first-order conditions for the two goods, Tl gives a relation between x and y. Simplify2 Optimization in Beonoméc Theory this as far as possible, and then use it and the budget. constraint to solve for the quantities.) Note that the two utility functions are Hnked Ulz,y) =I B(2,y)}, or (zy) = explU(,y) This illustrates that changing the utili transformation does not affeet the consumer's optimum choice. If ‘observed deimand behavior is all that matters, then the forma of the utility function is indeterminate (and irelevant) to within such transformation, Any properties that depend om the choice of particular form are meaningless. One such is diminishing marginal utility of income, If we write the multiplier for this problem as A to distinguish it from the A of Example 2.1, then you should verify tnt y finetion by any iscreasing, ong ero op If (a+ 8) > 1, then J increases with income, Tn some circumstances, specific forms of the utility function play special roles, This happens when some assumptions about interpersonal comparability are made, or when fonctions that are additively separable across time-petiods or states of the world are used for representing preferences in situations involving time or uncertainty, But in all these cases, the primacy of the special functional forms arises from those other considerations, not from the underlying mechanism of individual choice. When these other considerations are absent, we are free to transform the tility function for computational convenience, Note that changing both a and @ in the same proportions leaves demand snaffected in Example 2.1 a8 well as in Exercise 21, A glance at ‘equations (2.14) or (2.17) shows that it is convenient to choose these proportions so that a+ = 1 Beercise 2.2: The Linear Bzpenditure System Return once again to the consumer of E bout let the utility fanction be modified to U, where ample 2.1, Bay) a In(z— x0) +8 In(y — yor Lagrange’s Method 23 where 2g and yy are given constants, and a +9 = 1. Show that we optimal expenditures on the two goods are linear functions of ieome and prices: pz =al+ Spm —aque, y= 81 Spro + aque. "This slight modification of the utility function brings with it a twush richer range of possible optimum choice. The budget shares nf the two goods can now vary systematically with income and mies, One good can be a necessity and the other a luxury (but hwitler good can be inferior since a and i must be positive to keep tHe iinginal wtilities positive), But the expenditures still have a sunple functional form. For these reasons, this specification was opilar in the early empirical work on consumer demand Barwise £.8: Production and Cost-Minimization Conpider a producer who rents machines IC at r per year and hires Inher L at sage w per year to produce output Q, where Q=VK+vi prose he wishes to produce a fixed quantity Q at minimum wal. Find his factor demand functions, Show that the Lagrange tuuihiplier is given by A= 2wr Q/(w +r). Siuygest an econoinic interpretation for A. Now let p denote the price of output, Suppose the producer aay the quantity of output, and seeks to maximize profi rw that kis optimum outpat supply i Q=plw+r)/(2wr) Holate this to your interpretation of 3. Further Reading Vor supplementary treatments of Lagrange’s method, see Varian (op. cit.}, appendices to chs. 5 and 20, and Smith (op. eit.) ch. 2 urets. 1-4), ch, 4 (sects. 1-3), “The development of the theory in this book is relatively intu- ive and heuristic, There are several textbooks that are mathe twusieally more rigorous; I mention just one MICHAEL D. INTRILIGATOR, Mathematical Optimization and ‘onomic Theory, Englewood Cliffs, NJ: Prentice-Hall, 1971; ch. 1 is hot Lagrange’s method,3 Extensions and Generalizations More Variables and Constraints If there aren choice variables (21,22)...2q), we simply Jet the veetor + have n components, Then (2.11) is extended to j 1,2).2-m, and we have (n+ 1) equations in the (w+ 1) unknowns, nately the n components of 2 and the number A. If there are sn constraints, write them as Gie)se, 112.2, where the functions are identified by superscripts to avoid con: {sion with partial derivatives, which are being denoted by sub- scripts. For the mament, continue to ignore any other restrictions such as non-negativity on the variables. We need m 0, Generalization of tae reasoning in Chapter 1 that led to the inequality condition (1.10) owe gives us positiv AG So Gia) so tlz) Extensions and Generalizations w Unee again Tomit the formal proof. More generally, some components of # may be positive and stlwrs zero, Then an equation like (3.2) should bold for the partial Ulerivative of the Lagrangian with reapect to every component that tw ponitive, and an inequality like the one just above with respect to very component that is 2er0, at the initial point. In other words, ‘ every j, we should have L(z)<0, #20, (33) with at least one of these holding as an equation. In exceptional ewes, both might hold as equations. But the logical possibility that Iwoth inequalities may be striet is ruled ont by te requirements of inaity "The requirement that at least one inequality in (3.7) should hob as an eqwation is sometimes stated more compactly as 3; L,(2)=0. rhe point is that the product can be zero only if at least one of the factors is zero. A pair of matched inequalities like (3.7), not both of which van be strict, is said to show complementary slackness. A single inuquality, say 2) 2 0, is Binding if it holds as an equation, that is, Ar, is at the extrome limit of its permitted range; the inequality is sic to be stack if 2; is positive, and so has some room to maneuver fore hitting an extreme. Each one of the pair of inequalitie (3.7) therefore complen nilck, the other must be binding We can collect all the component inequalities in (3.7) into vec lors, Here I shall use the following notation: x > 0 means that Fy > 0 for ever ponent inequalities is strict; 2 >> 0 means that all the component inequalities are striet, Then (3.7) becomes ints the slackness ia the other: if one is yy J; 2 > O means that at least one of these com LAz)<%, £20, — with complementary slackness, (3.7) it being understood that complementary daekness holds for each sommponent pair in the veetor inequalities (Once again, I summarize the result for future reference:28 Optimization in Beonomie Theory Lagrange’s Theorem with Non-Negative Variables: Suppose + isan n-dimensional vector, ean m-dimensional vector, F a function liking scalar values, G a function taking m-dimensional vector values, and m 0, and the constraint qualification rank G,(2) =m holds, then there is a value of A such that LA(Z,A) $0, #20, with complementary slackness, (3.7) ond 1a(z,A) =0. (36 First-order necessary conditions are supposed to narrow dow: four search for a solution. How does that work in this instance? We start without knowledge of which components of the solution are going to be positive and which ones zero, particular pattern, say #) > 0, 2 =0, 2 > 0, we get Li(z) = 0, 2; = 0, La(z) = 0, assumed pattern leads to a set of n equations from (3.7) may not have a solution at all, and even if they do, it may not satisfy the other inequality conditions required from the pattern, But iffa solution satisfying the requirements exists, then it becomes ‘candidate for the optimum choice, ‘There are 2" pattems of positive and zero components in the n-dimensional veetor 2, We can then repent th every one of these patterns and find other candidates for optimality Then our search narrows down to all these candidates. If we assume a , then from (3.7) In other words, ai These sane exercise for In some applications the search i relatively enay to perfor the simplex method for solving linens programming problems is es sentially a systematic algori But in general the search is too exhaustive and exhausting. If we bad to carry i cout every time, the prospects for walving constrnined optimization problems would be very poor. Luckily many problets of prac tical interest offer short cts or systems for searching among, tk patterns. In most basie contexts of economic theory, we ean mak attern of equntions and inequalities, im for sich a seareb Hood guesses about the Tike Extensions and Generalizations 29 peeved on that basis, and use second-order sufficient conditions us verify that the resulting solution is indeed the optimum, Incquality Constraints Now we can consider more general inequality constraints. ‘This is J considerable economic importance, because sully there is no ‘eanpnlsion to use all of avnilable income or some resouree, and we should determine in the process of solution whether it is optimal os ae it fly. Suppose the frst component in the constraint need only hold vm wn inequality Gz) Sey. We ean fit (ick as the sisinner’s problem of Chapter 1 is into the context discussed before using the same ntroduction of the “unspent income’ variable in the Let us define a new variable any = — GMa). (38) Iu terms ofthe enlarged setof variables, the consiraint has become r-exact equation, The new variable zn ie rest tegative, but we know how to handle that. Write Las the Lagrangian fr the new problem, to distinguish 41 fron the Z of the old one. Then 1d to be non Blais... tating tyes Am) = Fleiss. myAieoss Am) 495 [er = GY(21,.-.29) — tnt +0 A [er -Gia,.-2n) = Uezayeeety Aye Am) Ab Ea All the other first-order conditions are as before, but we have new one with respect to zus1 AL/Arag1=—M SO, Ze41 20, with complementary slackness, Recalt that G(x) = OL/0Ay0 Optimization in Economic Theory Extensions and Generalizations a Therefore the condition can be written Kukn-Tucker Theorem: Suppose « is an n-dimensional vector, 1 nv dimensional veetor, F a function taking scalar values, G a OL/O, 20, 20, (3.tuetion taking m-dimensional vector values. Define with complementary slackness: a form that is nicely symmetsi 1(2,A) = Fla) + Ale~ G(x) | (3.4) with the condition (3.6) for non-negative variables We can similarly allow all constraints to be inequalities. Ifwslwie A is an m-dimensional row vector. Suppose # maximizes do that, there is no reason in general to maintain the restrietion!'(r) subject to G(x) 0, and the constraint. qualifiea: mm 0, then (3.12) gives A= 1/g, and then (8.11) becomes p/q > 00, which is not true. Therefore cannot arise either, The economie reason is that the first small lunit of 2 has infinite marginal utility, so it cannot be optimal to Next consider x > 0 and y = 0. Then from the budget con: straint 2 = I/p, and from (3.11), = a/I. Using this in (3.12), we have 1 < ag/I, ot I< ag. This is n condition on the given parameters of the problem, and they may or may not satisfy it. If they do, the premises of the case are mutually consistent and we fe a candidate for optimality. ‘economies of this is simple Finally, if both 2 and y are positive, (3-11) and (3.12) give a/(p2) == 1/4, 80.2 = ag/p. Then the budget constraint gives y = 1/4 — a. This is logically consistent if 1 > ag. Once again this may or may not be satistied; if i is, we have a candidate for optimality Bztensions and Generalizations 33 This completes the discussion of the eases, Now let us infer swine useful ideas from the procedure and the results. ‘The first thing to note is that six of the eight possible patterns could be ruled out using economic sense; after some experience it is possible soning quite a lot in this way Next, observe that the space of all possible values of p, 4, nd Tis split into two exhaustive and mutually exclusive sets by tus ent down the amount of formal re tlw conditions on the parameters that makc each of the Inst two ‘ses internally consistent, One requires I< ag, while the other requires > aq. In many applications, a similar neat classification will energe, ‘Third, when I= ag, we have A= 1/q. Therefore andy =0; 1-Ag=0 woth inequalities in (3.12) hold as exact equations. When I dis: assed complementary slacknes: following (3.6), 1 mentioned this sibility, and called it an exceptional case, Now we see why: it rive at the special configuration of parameters where the solution 1s just at the point of switehing from one pattern of equations and ickness conditions to another inequalities in the complementazy pisttern Finally, let us restate the optimum choice rule I Taq, 2=ag/p and y=I/q—a To see the solution more clearly, carry out the thought ex eriment of starting at a very low level of income and raising it wrndually. At first, all of income is spent on After a point, the expenditure on z is kept coustant, and all wlditional income is spent on y. We can think of good 2 as an exemplar of a necessity: it has an absolute first claim on income, good x and none on Init once its needs are satisfied, all extra income can go toward ther goods Quasi-inear preferences are usefel when we want to isolate ine sector or industry and wish to avoid the feedback of sllects on the demand for its goods. This is often called “partial ‘wilibriom’; a better name would be ‘industry analysis’, The as siption that changes in income do not affect the demand for the4 Optimization in Economic Theory good in question is obviously not meant to be taken literally, bu often proves nn acceptable approximation or simplification for thy purpose, Example §.2: Technological Unemployment Suppose an economy has 300 units of labor and 450 units of land ‘These can be used in the production of wheat and beef. Bach uni of wheat requires 2 of labor and 1 of land; each unit of beef requir 1 of labor and 2 of land. A plan to produce + units of wheat and y units of beef feasible if its requizements of each factor of production are les than the available amount of that factor: Qety < 300, (au ety < 450. (315 Each of the inequalities allows all points on or below a stenight lin: The sot that is feasible given both constraints is the quadrilaters OABC of Figure 3.1. The north-east fronticr of the feasible set, « the production possibility frontier, is ABC. ‘Along AB, (3.14) holds as az equation, while along BC, (3.1 does. Only at B do both hold as equations, Everywhere else, thes js unemployment of one factor or the other You might be tempted to assume that it will be optimun + have full employment, and achieve production at B with 2 = 51 dy = 200. However, that is not necessarily so Suppose the society has an objective or social welfare functic defined over the quantities of the two goods of the simple form w have used before (ey) =a nz 4B Iny (6 where a and 8 are given positive constants, and (a+) — 1 We know from the intuition developed in the previous exa ple that non-negativity constraints on x and y are not going to Is binding, So let us leave them out from the start, Write the bas range multiplier for the constraint (3.14) as \ and that for (3.1 Form the Lagrangian 84 L(y Au) =a In 248 In y+A[800—22~ y) 4 p[450—2—2y Ertensions and Generalizations 5 ’ B =(50.200) Feasible set ‘ 150 0 - Fig. 3.1 ~ Production and unen he lirstorder conditions are (37) (3.18) WW) —22—y>0, AZO — with complementary slackness, (3.19) K-22 >0, #20 — with complementary slackness, (3.20) Betwoon (8.19) and (3.20) we have four possible patterns of ‘yuntions and inequalities. We should suspect that it is not guing. sensible to keep both factors less than fully employed, that is, ta huave A= 0 and =O. Let us check this out: using A= 0 = p uw (417) and (3.18) would give a = 0 = 4, and that is not so. 'herefore this case is ruled out, and we are left with three.2 Optimization in Economic Theory second line, \ is an m-dimensional row veetor, G_(#) is an n-by-n matrix, and dé is an n-dimensional column vector, The final re- sult is the product of the gow vector 3 and the column vector de of ‘equal disnensions m; therefore it is a sealar, In fact it is the inner product of the two vectors: Dde= By Xidew The result is important enough to be stated separately for seferenice: Interpretation of Lagrange Multipliers: If v is the maxims of F(z) subject to a vector of constraints G(x) = ¢, and \ is the row vector of multipliers for the constraints, then change de tat results from an infinitesimal change de is given by dv =) de. 42) It should be stressed that (4.2) gives only the first-order or linear approximation to the change in v if the change in e is more tina infinitesimol. For such changes, we can entry the Taylor ex- pansion to higher orders and find a closer approximation. This will De done, although for a different purpose, in Chapter 8. Shadow Prices To illustrate and explain (4.2), consider « planned economy for which @ production plan z is to be chosen to maximize a social welfare function F(x). The vector of the plan’s resource require ments is G(r), and the vector of the available amounts of these resourees is c. Suppose the problem has been solved, and the vec- tor of the Lagrange multipliers A is known. Now suppose some power outside the economy puts « small additional amount de, of the first resource (say labor) at its disposal. The optimization problem can be solved afresh with the new labor constraint to de: termine the new pattern of production. But we know the resultant increase in social welfare without having to do this calculation: it is simply Ay dey. We ean then say that the multiplier Ay is the marginal product of labor in this economy, measuped in units of its social welfare. This is clearly a vital piece of economic informa: tion, and that is why Lagrange’s method and his multipliers are so important in economies. Shadow Prices 43 If there is only one scarce input, then a paraphrase of the at- sgunent of Chapter 1 yields another very instructive way of looking AL this result. Suppose we use the additional labor input to raise the quantity of a particular good, say good j, leaving the outputs of al the other goods unchanged. Since we are assuming full em- ployment of labor in both situations, the increase dy in the output ff the chosen good must satisfy oye de, es G42). ‘The resultant increase in social welfare is By(@) dt, = [F@)/G3e)] de “The condition of optimality (2.5) says that the ratio in the square Trackers should be the same forall j. ‘Thercfore the effect of the Turginal increase in labor spply on socal welfare is independent Ut how the extra labor is used. ‘That is why we ean spenk una Iigiousy of the marginal product of Ibo. ‘Now suppose the additional Inbor can only be used at some cost. The muasimaum the economy is wiling to pay in terms ofits Sm socal welfare units is clearly Ay per marginal unit of cy. Any Ghats payinent leaves it with a postive net beneft from using the Tatra lnbor; for any lnager payment the cost exceeds the beueft tn this natural senso, the Lagrange mltipie is the demand price fhe planner places on labor services. A price expressed in nits of socitl welfare may seem strange, but a minor wodifieation brings Zr into familar light. Consider ome other resource, say lad, and tuber it 2. Now suppose the econonny is offered the serviees of tn extra dey of labor, but saked to give in return the serviers of tly of land. The net gain in social welfare foam this tnsseton is (isd Xz dey), Therefore the aoost lel the planner is willing to give up is (u/Ma)dey., Then itis equally natural to call the falio (y/a) the desaal price of a unit of labor measured in tite of land. You know from microeconomic theory that relative vies rather than absolute oes gowern market exchage;stslaty Ihe selative tangaitudes of the Lagrange mcliliers for diferent resources govern the planncr’s willingness to exchange one resonree for another. Ifa neighboring economy has a different tradeoff between the two resources on neeount of differences in their relative availability36 Optimization in Boonomic Theory First consider the case, 4 = 0 and j1 > 0, which I shall Iabel Case (i). Here (3.17) gives z = a/y and (3.18) gives y = B/2p. Since ye > 0, (3.20) then becomes WO=242y=(at Aa, or x= 1/450 ‘Then x= 450@ and y = 225 8. It remains to check out if the feasibility condition in (3.19) is We need 300 > 2+ y = 9000 +2958 = 900 ~ 675 8, or 8 > 8/9, The other eases can be checked out in the shall merely state the results: IFA > 0 and = 0 (Case (ii), we get 2 = 1504, y The case is internally consistent if 8 < 2/3. IFA > 0 and p> 0 (Case (ii), we get the full employment point x = 50 and y = 200, This case is internally consistent if 2/3< 8 < 8/9, ‘The solution gives several useful i range of parameters splits nicely into exhaustive and mutually ex clusive regions, in each of which just one of the cases yields a can didate for optimality. For low values of 8, the solution lies along, the line AB. Then there is a middle range where the solution stays at the point B. Finally, for high values of 3 the solution lies along the line BC. ‘The social indifference curves for the objective fanction (3.16) are like hyperbolas. ‘The higher is the weight attached to y (relative to the weight of 2), the more willing is society to sacrifice x for y, that is, the flatter are the hyperbolas. Therefore for low 1 Wwe get a tangeney of a social welfare contour and the production possibility set along the segment AB, for medium values we have ‘a comer solution at B, and for high values a tangency along BC. Next note that at any point along AB (except B), itis optimal to keep some land unemployed. ‘To see why, note that the goods have fixed coefficients of input requirements, and wheat requires atively more labor. If we wish to use the unemployed land, 9° must do s0 by producing less wheat and more beef. To try it the bother way round would increase the labor requizement, but labor is already fully employed. But this is a situation with a relatively same way, and | 3003, sits, Once again, the Extensions and Generalizations 37 lw i; beet is not highly valued relative to wheat, and the required netifice of wheat is not worth while If enough substitution in production were possible, then the Milfcalty would not arise and both factors conld be fully employed. The unemployment in this setting is a consequence of the rigid teehuology, not of any effective demand failures or coordination fires Finally, look again at the complementary slackness conditions (2.19) and (3.20). If one factor is not Sully employed at the opti- twin, then the Lagrange multipber for its constraint is zero. Ia Clunpter 1, the multiplier on the consumer's budget constraint was ke marginal utility of money. 1 this problem gives the effect on social welfare of having another wnvrginal anit of that factor. Then complementary slackness be: In the same way, each multiplier ws economically quite intuitive: if it is optimal not to employ ‘the available amount fully, then an increment must be worthless Inthe next chapter Tshall develop this idea in more detail Exercises Kerveise 8.1: Retioning Suppose a consumer bas the utility fanetion U(er,22,25) =a In(er) +42 In(z2) +03 (zs), (3.21) wluye the aj are positive constants suinming to one, The budget semstraint is pity + pata t Paes | wldition, the consumer faces a rationing consteaint: he is not allowed to buy more than k units of good 1 Solve the optimization problem, Under what condition on the ‘uriows parameters is the rationing constraint binding? Show that when the rationing constraint binds, the income Hust. the consumer would have liked to spend on good 1 but cannot shisa is now split between goods 2 and 3 in the proportions @, : a3 Would you expect rationing of b tun Initter and rice in this way? What is the property of the utility Innnetion (3.21) that produces the result, and hiow would you expect ‘hw bread-butter-rice ease to differ? fad purchases to affect demands38 Optimization in Economie Theory Exercise 9.8: Distribution Between Envious Consumers There is a fixed total ¥ of goods at the dieposal of society, The fre two consumers who envy each other. Ifvonsumer 1 gets ¥y an consumer 2 gets Yp, their utilities are UW=K-kYF, Wave ky?, where & is a positive constact Yo <¥, and maximize Uy + U2 ‘Show that if ¥ > 1/k, the resource constraint will be slack » the optimum. Interpret the result, The allocation auast satisly Yt Exercise 8.9: Investment Allocation A capital sum C is available for allocation among n investmew projects, If the non-negative amount 2; is allocated to. projes j for j = 1, 2, «.. n, the expected return from this portfolio « projects is Vlas 1 The allocation is to be chosen to maximize this. Find the first-order necessary conditions from the Kubn-Tucls Theorem, Define #=Y(as/85) Show that (2) If © > H, then a part of the total sum available is lef unused (Gi) Hay > (H-C)/K for all j, then every project will receiv some funding, (iii) If any project receives lower a than any project that gets some funding, ero funding, then it must have Further Reading A more rigorous treatment of the more g¢ ‘Tucker theorems is in Intriligator (op.cit.) ei. 4 Quasi-linear preferences and their applications are discusses ral Lagrange and Kuli in more detail by Extensions and Generalizations 39 HAL VARIAN, Microeconomic Analysis, New York: Norton, ‘wul ciu., 1984, pp, 278-83, ‘A good modem treatment of rationing is in Peren NEARY and KEVIN ROUERTS, “The theory of house lwald Ichavior under rationing’, European Economie Review, 13 (ISD), pp. 29-42. ‘They use methods that will not be developed ww this book until Chapter 8, but it will be worth while to return to this reference ther4 Shadow Prices Comparative Statics Lagrange’s method, and its extensions and generalizationsin Chap 43 all introduce an undeterinined multiplier for each eonstvaint ‘The values of these multiplices are found as part of the sol tion. The heuristic diseussion of the consumer choice problem int Chapter 1 offered an economic interpretation for its Lagrange mul: tiplicr: it was the marginal utility of income. In Chapters 2 and 3 T hinted that a similar interpretation holds rch more generally for constrained optimization problems, That is the focus of this chapter. A constrained optimization problem has several parameters as ddatn, In the maximization of F(x) subject to G(2) =, the para nieter ¢ i an obvious example. There are also other parameters that appear in the definitions of the functions F and G, for exani- ple the weights a, 8, and the prices, in the exantples and exercises of Chapter 2. Economists often need to know how the solution to the problem will change if these parameters take different values. In consumer theory, we discuss the income and Substitution effects bof price changes by comparing the optimum choices for different budget lines. In the theory of a firm's production and supply, its rinrginal cost is the difference between the easts of producing two Jiffetent levels of output when the firm chooses the least-cost input tix for each outpat level. The general method of comparing sol- utions for various parameter changes is ealled comparative staties and the importance of Lagrange multipliers lies in the fact that they provide the answer to a very i question. miportant comparative static Equality Constraints Let us begin in the simple setting of Chapter 2, with two choice variables (21,2), an objective function F(z}, and one equality constraint G(x) =e. Let 2 denote the optimum choice, and » = F(2) the highest attainable value, Now suppose ¢ increases by an, Shadow Prices a Uufinitesimal amount de, Let (# + di) be the new optimum choice, 4+ do the new optinura value. Note a slight difference between the usage here and that of Chapter 2. There the aim was to test 2 for optimality, and we did this ly considering arbitrary devintions dz from it. ‘This led us tw the first-order necessity conditions that held nt the optirmam 1 New the increment di is not arbitrary; it is the optimwm small ‘huge in the choice, arising in response to a small change in the meters. For these sanall changes, we cant use the first-order Taylor ap- oximations to the changes in the values of F and G. We have dy = FUR + dz) — F(2) Fila) day + Fa()de A[Gu(R) dn + Gx( 3) de, 2 [G(e +z) - G(a)| [let de) — In the derivation, the second and the fourth lines are the Taylor npproximations, the third line uses the first-order condition (2.6), nnd the fifth line uses the constraint (2.1). The result. can now be written, dofde (4a) Thus the multiplier is the rate of change of the maximum ible value of the objective function with respect to a change iw the parameter om the right-hand side of the constraint. Now we fmt see the marginal utility of iucome in Chapter 1 as a special nse of this more general result The case of several choice variables and many equation con- ntruints is no harder, In vector-matrix notation, the argument is 1m fact identical. Look at the first section of Chapter 3. Let the tight-hand side of the vector constraint change by de, and write dz for the resulting change in the optinnum veetor £. Then. dv (r+ de) — Fz) = Fe(2) dt Gx(2) dz = d [GUE + dz) — G(z)| = Ade. inse a moment to check the sizes of the various vectors and mat- tives being, sultiplied. For example, in the first expression of the4 Shadow Prices Comparative Statics Lagrange’s method, and its extensions and generalizations in Chap. ter 3, all introduce an undetermined multiplier for each constraint ‘The values of these multipliers are found as a past of the solu tion. The heuristic discussion of the consumer choice problem in Chapter 1 offered an economic interpretation for its Lagrange saul tiplier: it was the marginal utility of income. In Chapters 2 and 3 T hinted that a similar interpretation holds much more generally for constrained optimization problems. That is the focus of this chapter, A constrained optimization problem has several parameters ns data. In the maximization of F(x) subject to G(2) = ¢, the para meter ¢ is an obvious example, There ate also other parameters that appear in the definitions of the functions F and G, for exans ple the weights a, 9, and the prices, in the examples aud exercises ‘of Chapter 2. Economists often need to know how the solution to the problem will change if these parameters take different values In comsmer theory, we discuss the income and substitution effects of price changes by comparing the optimum choices for different Dudget lines. In the theory of a firm's production and supply, its marginal cost is the difference between the costs of producing two different levels of output when the firm chooses the Ieast-cost input mix for each output level. The general method of comparing, sol tutions for various parameter changes is called comparative statics, and the importance of Lagrange suultipliers lies in the fact that they provide the answer to a very important comparative static question, Equality Constraints Lot us begin in the simple setting of Chapter 2, with two choice variables (21,22), an objective function F(z), and one equality constraint G(z) = e. Lot 2 denote the optimun choice, and v F(z) the highest attainable value. Now suppose ¢ increases by an Shadow Prices 4 smal amount de. Let (2-+ dz) be the new optimum choice, md + do the new optimum value. Note a slight difference between the usage here and that of Chapter 2, There the aim was to test 2 for optimality, and we did this by considering arbitrary deviations dr from it. ‘This led us lw the first-order necessary conditions that held at the optimum +. Now the inesement di is not arbiteary; it is the optimus small change in the choice, arising in response to a small change in the ppavanneters. For these small changes, we ean use the first-order Taylor ap- proximations to the ehanges in the values of F and G. We have dy = F( + de) — F(z) = F(@dry + F(a) doa =A[Gile) der + Gol 2) d%2} =A [Gla +e) —G13)) =A [let dey —e] = Ade In the derivation, the second and the fourth lines are the Taylor ‘approximations, the third line uses the first-order condition (2.6), rund the fifth Fine uses the constraint (2.1). The result can now be written dofide =r (4a) ‘Thus the multiplier is the rate of change of the maximum tainable value of the objective function with respect to a change in the parameter on the right-hand side of the constraint. Now we con see the marginal tility of income in Chapter 1 as a special cease of tis more general result ‘The ease of several choice variables and many equation con- straints is no harder. In vector-matrix notation, the argument is in fact identical. Look at the first section of Chapter 3. Let the right-hand side of the vector constraint change by de, and write dz for the resulting change in the optimum vector £. Then dy = F(z + dz) - F(z) = Fy() dr Gz) dz = d (GE + dz) ~ G(z)] = Ade. Pause a moment to cheek the sizes of the various vectors and nat rives being multiplied. For example, in the fist expression of the42 Optimization in Economic Theory second line, \ is an m-dimensional row vector, Ge(#) is an m-by-n matrix, and di is an n-dimensional column vector. The final re sult is the produet of the row vector A and the column vector de of ‘equal dimensions m; therefore it is a sealar. In fact it is the inner product of the two vectors: Dde= 3s Avdey ‘The result js important enough to be stated separately for reference: Interpreietion of Lagrange Multipliers: If v is the maximum of F(z) subject to a vector of constraints G(r) = ¢, and \ is the row vector of multipliers for the constraints, then change dv that results from an infinitesimal change de is given by dy = Ade (42) 1k should be stressed that (4.2) gives only the fist-order oe linear approximation to the change in vif the change in ¢ is move thon infinitesimal. For such changes, we can earry the Taylor ex- pansion to higher orders and find a loser approximation. This wil be done, although for a different purpose, in Chapter & Shadow Prices To illustrate and explain (4.2), consider a planned economy for which a production plan 7 is to be chose to maximize a social welfare function F(z). The vector of the plan's resource require- ments is G(r), and the vector of the available amounts of these resources isc. Suppose the problem has been solved, and the vee- tor of the Lagrange multipliers 2 is known. Now suppose some power outside the cconomy puts a small additional amount dey of the first resource (say labor) at its disposal. The optimization problem can be solved afresh with the new labor constraint to de- termine the new pattern of production, But we know the resultant increase in social welfare without having to do this calculation: it is simply 4, dey. We can then say that the multiplier 3, is the I product of labor in this economy, measured in units of its social welfare. This is clearly a vital piece of econornie informa- tion, and that is why Lagrange’s method and his multipliers are 90 innportant in economies, Shodow Prices 43 If there is only one searce input, then a paraphrase of the ar- ‘xument of Chapter 1 yields another very instructive way of looking At this result. Suppose we use the additional labor input to raise the quantity of a particular good, say good j, leaving the outputs of all the other goods unchanged, Since we are assuring full ployment of labor in both situations, the increase dé; in the output. ‘of the chosen good must satisfy Gi(e)d2;=dey, or dB, = dey /G}(z). ‘The resultant increase in social welfare is Fy(z)dz, = [Fy(2)/GY(2)] des ‘The condition of optimality (2.5) says that the ratio in the square Inackets should be the same for all j. Therefore the effect of the marginal increase in labor supply on social welfare is independent of how the extra labor is used. That is why we can speak unas: figuously of the marginal product of labor. ‘Now suppose the additional labor can only be used at some cost. The maximum the economy is willing to pay in terms of its ‘own social welfare units is clearly Ay per marginal unit of ¢, Any smaller payment leaves it with a positive net benefit from using the ‘extra labor; for any larger payment the cost exeoeds the benefit In this natural sense, the Lagrange multiplier is the demand price the planner places on labor services. A price expressed in units of social welfare may seem strange, but a minor modification brings it into familiar light, Consider some other resource, say land, and ammber it 2. Now suppose the economy is offered the services of fun extra dey of labor, but asked to give in return the services of dep of land. The net gain in social welfare from: this transaction is (Oi dey — Ae des). Therefore the most land the planner is willing to give up is (Ai/Az) dey. Then it is equally natural to call the ratio (Ai/\2) the demand price of a unit of Jabor measured in ‘mits of land. You know from microeconomic theory that relative prices rather than absolute ones govern market exchange; similaely the relative magnitudes of the Lagrange multipliers for diferent resources govern the planner’s willingness to exchange one resource for another. Ifa neighboring economy has a different trade-off between the two resources on account of differenecs in their relative availability“4 Optimization in Beonomic Theory or technology, then theres & possiblity of matually advantageous trade in factor services between the two. (Even if lactor services cannot be traded, exchange of goods made using these factors can secure some or even al of this mutual gain, but details of that would take us too fat into the theory of intermational trade.) OF course the internal organization of the economy ised have sothing todo with prices, nd the Lagpange malplice fr te lar constraint need not equal the wage that ncially pai for exch Labor may simply be directed to various tasks in command economy. (There are serious conceptual and practical problems in so doing, as most Soviel-stple economies have now realized, but that agin is another story.) But the plan impictly places values on the rsonrers, and the planners nnetstning of the economy and ofits possible bottlenecks will be improved by Paying attention to the multipliers tht reflect these wales Now consider ax economy tliat does allonte resoutes ving tatkets. In equilibrium, the prices age such that the dean and supplies chosen by individuals solving tir own constrained maximization problems are equal inthe nggeegate. Now wippone an economist sets out to evaluate the performance of the economy ssing some given criterion. To get a comparison standard, he wil solve the planning problets of maximizing this ertrion function subject to constraints arising from the eeunomy'sremoece aval ability, technology, and information transion, ‘The solution will include a vector of Lagrange multiplies for the resource con Strait You may think there is little reson why the market should replicate this planned allocation, anv equally bite season why the Lagrange multipliers should have anything todo with the marker prices. But there are important cases where the optismm ean be replicated in the market, and the Lagrange smuliplicrs ave pro Portional to the market prices ofthe resources: the relative prices equal the corresponding ratios of multipliers fn such cases the économist is tempted to say thatthe economy is guided by an ne visible hand? to his planned optimum. Such a ease is worked out i detain Example 41. Terese on many special asituptions whose validity is often doubifa and most of modera econemie theory i ‘concerned with questions of what happens when thowe gssunpione ste not mel. But the ease has great importance asthe point of de. vture for all suc analysis ad as a practical matter neany people Shadow Prices 45 Inieve in the optimality of the market mechanism. Therefore it eserves careful study. To evoke the connection with prices, and yot maintain a com: rvptual distinetion from market priees, Lagrange multipliers are sften called shadow prices, inequality Constraints ‘An economic question now arises, We expect prices to be non: wgative, but so far we have seen to reason why’ the shadow prices (Lagrange multipliers) should be non-negative. In the planing pplication wsed in the above exposition, the multipliers measured tHe increase in social welfare resulting from incrcased avsilaility of seazee resources. Having mote of a resource Leaves all previons pprnbiction opportunities available and adds some new ones. ‘This shendd allow the planner to achieve at least as hgh a level of socal welfare, and in most instances a higher level. In the sane way, in Uwe general problem of constrained optimization, a relaxation of te constraint should be a desirable thing. Can the mathematies a this intuition? One dificulty is that in the genernl formulation of the prob- lov with equality constraints, an increase in the right-hand side of n constraint equation need not mean relaxation of the eon- iraint. Trivially, we could have written the constraint G'(2) = cy tw =G'(z) = —c, and an increase in the right-hand side of the new forin would be a decrease inthe quantity ¢, of resource i. Also, not nll of the constraints nocd be ones of resource availability. Por ex: tmp, we might want to maximize the amount of investment while casting a minimum acceptable provision of some consumer goods Now an increase in this stipalated minimum tightens the economic constraint, s0 a smaller amount of investment ean be squeezed out td the multiplier is negative. Here the multiplier is ike the slope ‘ofa transformation Sanetion (consumption iato investment). We should expect such a eurve to be downward-sloping, and should inerpret minus the slope as the shadow price ‘These examples show that if we want non-negative shadow tices, we mast be careful to write the constraints in such a way that an increase in the right-hand side does relax the restrictions ca the eheice. ‘There is another, more important, consideration, There may. he eases in which the marginal value of a resource tims negative46 Optimization in Economic Theory beyond a point. For example, too many workers may simply im terfere with one another's effort In such a ease, x further increase in the quantity of the resource will mean a lower maximum value of the objective function and a negative mnltiplie. But in sveh a situation, it would be better not to use the resource in such ex se 5.2 considered a situation where it was optimal to throw away some quantity of = food in the face of overwhelming envy effects. Mathematically, te equality constraint forces use ofthe entize amount available. If the constraint were an inequality, G¥(z) $ ci, then we woold have the freedom to Teave resources idle when this serves the goals of optimization In practice there may be some costs of leaving resources idle Unemployment of labor might be thought to be socially undesir- able, and somne capital, especially brains, ean rust when unused. In such situations we should include these costs in the objective function, Provided this has been done, theres no economic reason to deny ourselves the freedom of leaving some part of the resource endowment unused if this leads to a better outcome, ‘This diseussion has an exact parallel in matket prices, too. If some ‘goods’ are actually bade’, we expeet them to have negative prices. More generally, it is the assumption of free disposal that ‘eres non-negative prices ‘The Kuhn “Tucker Theorem stated in Chapter 3 gives the fist- order necessary conditions for maximization subject to inequality constraints. It immediately confirms this intuition. ‘The condi- tion (3.10) says that the vector of the Lagrange multipliers is non negative. It yields a further result of considerable importance. The vector inequality \ > O shows complementary slackness with Ly 20, which is just another way of writing G(z) < ¢. For every i, mt least one of the pair cessive amounts even if it is available, Exe GaSe N20, holds as an equation, If resource ¥ is not fully used, then its shadow price is zoro; a resource with a positive shadow price must be fally| sed. This supports and completes the interpretation of shadow piers as the marginal value products of the resources. If part of some resource is alerady idle, then any increment in it will also he left dle, ‘The maximum value of the objective function will not, Shadow Prices 7 ‘luge, and the shadow price will be zero. On the other hand,» Iwoitive shadow price means that » marginal increment jn resource tinilability can be put to good use. Then none of the amount ixinally available can have been left idle in the original plan Youu should return to the account of technologieal unemployment Hxample 3.2 and examine the results there inthis light. ‘There is just one tricky poiat to be taken eare of. Suppose e's sel that resource ¢is just on the point of becoming superfluous a: tive tuaggin. This amount is fally used, but any increment will be left amused. Conplementary slackness does not tell us whether the swltipler will be positive or zero at this point. In fact the answer In specific to each peoblem, and depends on whether the slope of he imaximum value o showa as a fenction of e, drops smoothly or nuddenly to zer0 at the borderline point. ia) i) Fig, 4.1 ~ Resource quantities and shadow prices Figure 4.1 shows both possibilities. In (a) the drop is smooth, ‘wud the multiplier at the point in question is zero. In (b) the drop insndden, and any value of A, between the slope of the curve to te left and its slope to the right (zero) will serve at the point48 Optimization in Economic Theory of transition. This happens in the context of linear programmin, (Example 7.1). ° * Examples UH errr ae ee ee ee eee Sur thle cinae the mamuan Ties be @ semua ieee ee ame gga nici he bak ie eee ee ee tammy ac iy el on te = Urea, cay ee) (43) Social welfare is a function of these utility levels w= Wasa ue) ‘The constraints are that for each good, its allocation to the indi viduals should add up to no more than the total amount available, When the utilities and social welfare are increasing functions, it is clear that no goods are going to be wasted, so we can express the constraints as equations fay tia tet for g=1,2,...@, (44) ie o Lagrange's Theorem. Let x, be the multiplier for the constraint on good g, and form the Lagrangian L=WO any. HG) Ute. +L l-¥ sal, tea)) where the arguments of [ and the ranges of summation are omitted, for brevity. ‘When deriving the first-order conditions, we must differentiate L with respect to every ey using the chain nile, This gives (W/O) (OU) Beg) — ny = 0. (45) Shadow Prices 49 All the partial derivatives are to be evaluated at the optimum as tial; [omit them for ease of notation. ‘The multipliers ry are also ubtnined as a part of the solution. New suppose the 1, are made the prices of the goods, Every eonsumer cis given a money income Ie, and allowed to choose his rronsumnption vector to maximize his utility (4.3) subject to the Wurget constraint mratmia bet rang = le (48) jx optimization will be characterized by the conditions BUSJOrey =e, an allg and c. As usual, Xe is the marginal utility of income for If me compare (4.5) and (4.7), we sce that they coincide pro- vised we set OWJdue=1fdy or (OW/Oue) o= 1, (48) furalle, This can be done by adjusting the money incomes Ie. The Iefe hand side of the second equation in (4.8) is simply the marainal tilect on social welfare of giving a unit of income: to consumer ¢} it Iw the marginal effect on e's own utility times the effect of a unit of his utility on social welfare. In other words, the distribution of income should be arranged nm that at the margin the social value of every consumer's income Feith same. Once this is done, they ean be left free to choose their rvtual consumption bundles. This is the ‘invisible hand” result for the distribution problem. "The argument comparing first-order conditions is not fully rg nus, but better proofs exist. The important thing is to revognize the erucial assumptions that lead to the result, Here the most im ntant is the dependence of every consuiner's utility only on his tun consumption quantities, [fone consumer's utility depends on tmther's consumption, this is called an ‘external effect’ or ‘exter- hlitg", Such effects can interfere with the simple decentralization nsumption through prices. In essence, we must charge each ‘eansner not just for the scarcity value of his consumption, but ‘hho for the harm his consumption causes to the wlility of others50 Optimization in Economic Theory (or pay him for the benefit he confers on others). Such prices can bbe person-specific, and the market implementation becomes mucl ‘more complicated. Brample 4.8: Duty-Free Purchases Let us tuen from the dreary image that central planning an come distribution always invoke, to the consumption decision of jetsetter. He can buy various brands of liquor at his home-tow store, or at the duty-free stores of the various airports he travel, through. ‘The duty-fee stores have cheaper prices, but the to tal quantity he eat buy Where is restricted by his home country’s customs regulations ‘There are n brands, Let p be the row vector of home-tow prices and q that of duty-free prices. The duty-free prices ars tuiformly lower: q 0 and yj =0, we have BUA; = Ap, (aaa) sand BU Bes < Ag; +p (4.15) Of these, (4.14) i familiar, Dut even it benefits from a reinter- pretation, The left-hand side is just the marginal utility of brand J. The right-hand side is the marginal opportunity cost of buying i€ at the home-town store: to do so takes p, of income which cane uot then be used for other purchases, and the utility value of this muh income is 25. ‘This in turn casts Hight on (4.15). Its right-hand side isthe marginal opportunity cost of baying « unit of brand j at a duty frce store. This requires g; of income having utility value Ag;. But ‘also uses up a unit of the duty-free allowance, which has the shadow price . The total opportunity cost of the purchase is the sam of these two components, If the brand is uot bought at the duty-free store it must be because the oporlunity cost of so doing exceeds the marginal utility from its consumption, Now the principle is clear: buy eaeh brand at the outlet with the lower opportunity cost. Note that Dgikw p/ Therefore our jet-setter should rank the brands by their absolute price differences in the two kinds of stores. The brands with the largest price differences are bought at the duty-free stores, and those with the smallest price differences, at the home-town store. ‘The meeting-point of the two is chosen so as to use up the duty free allowance. There may be at anost one brand that is bought at oth kinds of stores. Incidentally, if the duty-free allowance restricts the total value 4 of purchases instead of the quantity ey, tlten tie solution will be similar, but the brands will be ranked by their relative price differences instead of the absolute ones, I shall leave this case for te readers to work out as an exercise Shedow Prices Exercises Brercise 4.1: The Invisible Hand ~ Production Continue with the notation of Example 4.1, but now allow produc- tion of the goods. Let there be F factor inputs, available in fixed quantities Zy for f= 1,2,..., F. H2y of factor f is used in the production of good g, the output X, is given by the production function Xy = (215 229, 2) (4.16) |Add these constraints to the earlier problem. Verify that the first-order conditions of optimum distribution aze the same as be- fore, but new conditions for optimusa factor allocation sxe added. Interpret the Lagrange multipliers. Can production be dee ied, with one firm producing each good? Show that the sum of the incomes J, handed out to the consumers equals the value of aggregate output. Bzereive 4.8: The Invisible Hand ~ Factor Supplies Now let even the factor supplies Zy be a part of the optimization Suppose each consumer ¢ supplies zcy of factor f. These amounts affect his utility adversely; there is disutility from supplying fac- tors. Find the first-order conditions. Interpret the Lagrange pliers and diseuss the implementation of the optismum in a market Iramework Now you must distinguish two sourees of income for the con- sumers: their earnings from the factor services they supply, and the lump sums I. they get from the government. These lunp sums must now be varied (in a person-specifie way) te attain the condition (4.8), Show that the total of the luup sums liaded ont to the consumers equals the total profit in prodiction, that is, the value of output minus the payments to the factors, pulti- Brercise 4.9: Borrowing ond Lending Consider a consumer planning his consumption over two years. He will have income 1, during the first year and , during the second. In each year there are two goods to consume. Ip year 1, the prices are p, and 91, #nd the corresponding quantities £; ands y,. In year 2, we similarly have pp, g2 and 2, yp. The utility function is ty = an In(21)+ 41 In(yr) +2 ln(ze) + B2 In(y).4 Optimization in Beononn Thearg This is to be maximized subject to two nly! cont raints, one for wach yea. Solve this problem, and find the wielipers \y snl Ay for the two constraints. Examine how they dep! ss money income and other parameters of the problem How much more of year-2 income will yeamnuer require if he is to give up dl; of yeae-1 incom? tw utter words, what js the rate of retum needed to induce hinw fe save a Uittle? You would expect borrowing and lending institut wns ise in am econ ‘omy populated by stich consuaners, Whit serie: wlio will borrow and who will lend? Further reading An excellent di ffcient allocation of resources is in TIALLING C. KOOPMANS, Three Bonus wn the State of Boo nomic Science, New York: MeGraw-Hill, sion of the role of prive. in cbcentealizing an 5 Maximum Value Functions ‘The last chapter introduced the concept of comparative staties, and nsed it to interpret the Lagrange multipliers as the rates of change of the maximum attainable value of the objective fanetion with respect to the right-hand sides of the constraint equations, Many ‘other parameters enter the objective funetion and the constraint functions, and the maxismum attainable value of the objective fune- tion depends on them all, The anethod used before ean be adapted ‘0 understand the nature of this more general dependence. ‘That is the object of this chapter. As there, I shall begin with the ease e exact equalities, and consider inequality where all constraints 4 constraints later. Parameters in the Objective Function Consider first the case where the parameters affect the maxinand alone. A common example is a producer who chooses a mix of in outs to minimize the cost of producing a given target output, The oriees of the inputs are parameters that affect his objective fune- jon, But the constraint, which says that the chosen inputs should yield the desired output, involves only the production funetion and not the prices. As another example, consider a small country that shooses its production pattern to maximize the national product evaluated st world prices; now these prices are parameters thot enerally, suppose a vector 8 of para meters enters the objective function, so 2 is chosen to maximize (2,8) subject to the usual vector constraint G(2) = ¢. Lageange"s hold with a slight nnodifieation ~ we now recognize the dependence of F and the La. grougian on 8, Thus affect the maximand. More frst order necessary conditions (3.5) anu ( 1(2,,6) = F(2,8) + Mle~ Ge) and the optimum 2 satisfies the first-order conditions 0. L(#,A.8)=0, LA(?.2056 Optimization in Beonomic Theory White » for the maxinnum value once again. Suppose @ changes to (8+ d@). Correspondingly, let the optimum # change to (2 + dz), and the maximum value to (v + dv). Using first-order Taylor pproximations as in Chapter 4, we eau find an expression for dv, dy (e+ de,0 + d8) ~ F(z,8) = Fi(z,0) de + Fo(2.0)d0 Gu(a) de + Fol 2,0) 40 Ful, 8) a8. (a) In this caleulation, the passage from the secoud to the thied line uses Lagrange’s first-order condition, atl the passage to the last Tine uses the fact that the value of G stays equal Lo ein the course of the change in 8. Once again, for changes in @ that are large enough to make the first-order approximation invalid, we ean eatry the series expansion further tofind closer approximations to changesin v. But the result above hins grent interest because of its simplicity. It says that to find the first-order change in the maxiaw value of the objective function in response to changes in paranoters that do not affect, the constraints, we need not worry about the simultaneous change jn the optimum choice 2 itself. All we have to do is to ealeulate the partial elect of the parameter change, aul evalnite the expression ft the initial optinuam choice. The cost-minimization problem mentioned! above illustrates this well. Suppose x is the vector of inputs, G the production finetion, and ¢ the required output quantity. Let @ be the row vector of input prices. Then the producer minimizes @2, that is, maximizes F(z,@)=—@2 subject to Gl) = Write the resulting mas smam as (—»). Thon (5.1) gives d(—v) = d8 Fo(2,0) = 8. Note that d@ is a row vector, so we interpret Fy as n column voetor and write the inner product as shown. The result is tv Mezimum Value Functions 87 Now v is just the minimum cost of producing output ¢ when inymt prices aze @ When input prices change, the producer will ge his input. mix, using less of the inputs that, have become twlatively more expensive and more of the others. That is, he will ubstitute along. an isoquant of the production function. But tht 7 in (5.2) is the optimal choice for the original parameter vector not the one for @+ 6, or some average. In other words, the lirstorder change in the cost is just the change in the cost of the siginal optimum #, as if Bxed coefficients ruled. Another way of looking at this is useful. If we write v = 0 and differentiate, we have do =O de +d0 2, "The first term on the right-hand side is the value of the clang in the input mix, using the original priees. But at those prices, the priginal mix is chosen optimally, therefore the value of any change in it must be zero to the first order. This just leaves: the second orm, as in (5.2). ‘The Envelope Theorem ‘The algebra of the previous section is illustrated grometziealy i Figure 5.1. For a particular value of 8 say 6), suppese the nptinnn choice is #4, The two eurves represent two fnetions of @ One is F (21,0), where 2 is held fixed at 2! as @ varies, ‘The wth is the ‘optimum value function linking v and @, where 2 is alloweel to vary ‘optimally as @ varies. Formally, this funetion is defined by V() = max { Fl2,4) | G(x) =e } (53) which is read as “V(8) is the maximum over x of F(r,@) subject to G(x) =e. Next write the optimum choice 4 = X(8), then we have tself as a function V(0) = FX (6). 8), ‘The two functions V(0) and F(2",6) coincide at &, becanse 4" happens to be the optimal choice there. For other values of 4, unless 2 remains the optimal choiee, the eurve showing the ‘optimsam value function will be higher than that of F(s!,@). (In event, it cannot be lower.) Therefore the two curves should be tuntually tangential at @, and that is just what (5.1) expressesOptimization in Beonomic Theory i) ) oo Fig, 5.1 ~ ‘The Envelope Theorem Similarly, we could draw the graph of F(?,6), where 22 is the optimal choice at @. ‘This would touch the graph of the optiinal value function V(B) at @2. In faet we could draw a whole fanily of curves of F(x,0) for a whole range of fixed values of 2, ench optimal for some 6. Ne member of this family of eurves contd ever ross ubove the graph of (8), and each would be tangential to the ‘optimal value function at that value of @ where its + happened to 1be the optimal choice. In other words, the optimal value function is the upper envelope of the faunily of value functions, in each of which the choice variables are held fixed. That is why the fornnula. (5.1) is often referred to as the Envelope Theorem: In the cost-minimization application, for exannple, let a scalar parameter # denote the price of just one input. When the vector ‘of inpat quantities is held fixed, the cost of production is a linear fiction of @. ‘The minitnized cost as a function of 8 is the lower welope (not upper, because this is a problem of miniztization, not ‘wiaximization) of all these straight lines. Figure 5.2 shows this, The mains focus here is on a first-order or tangency prop: ‘erly: where the upper envelope meets one member of the farnily aie Maximum Value Functions 9, oc cost lines for fixed v2 2 minimum cost vo 6, Fig. 6.2 — The minimum cost finetion cof curves, the two are tangential. In the cost-minimization exam ple, let denote the quantity ofthe input whose pice # ix bring vivied. ‘The slope of ea line equals the fixed # along it. “The Tine touches the minim cost fumiction at that & where this 2 is snl, Therefore the ope of te mii cot faction a Y point is just the optimal value of x there. In other words, the int cont anetioncartie thin che fafermatien abn th optimus choices of inputs. This idea will be developed farther i Pxample 5.2. vrei aco onde oc curator property i al event. Fig te 5.1 shows cach F(z, 8) a8 8 concave enrve and V(B) as 2 convex curve. But more generally, the envelape aust be more convex thas ‘niy member of the family of which it isthe envelope. "Tus the cost function for any fixed input choice is linear in input prices, but the lower envelope (the minimum cost curve) is concave. ‘This prop. ‘sty will be stedied in more detail in Chapter & and it will lead to an itmpoctaat comparative statie result called the Le Chatelce Samuelson Principle.60 Optimization in Economic Theory Parameters Affecting All Functions Now suppose G as well as F involves @. The calculation proceeds as above, except that the change in G is no longer zero. Suppose the vector constraint is G(x,8) = c, where @ and e are distinet. Then G(2,8) de + Go(2,0) dO Using thi the previous chain of equations, we have do == Gye, 0) d0 + Fal, 0)d0 Lol#,4,8) db (4) ‘The difference between this and (5.1) has an intuitive explanation When @ affects the constraints, a change d8 has Uke direct effect of increasing the value of G by Ga(z,9)d8. This acts exactly like a ‘equal reduction in ¢. The interpretation of the Lagrange multiplier tells us that the equivalent reduetion in e reduces v by \Go(#,8) 48 ‘This is just the additional term in (5.4) when compared to (5.1). In the previous chapter, a similar comparative static analysix with respect to changes in the parameters c led to (4.2), which gaw us the important interpretation of the Lagrange multipliers. Ti more general formulation of this chapter ean subsume the carl cease. ‘To see this explicitly, define a larger vector of parameters, & which includes @ and cas subvectors, and write the constraints as G(z,8) = C(2,8) ‘The Lagrangian can now be written as E(2,A,8) = F(x,8)- G28), and (5.4) becomes av = Ee, ,8) a8 Separating the subveetors in 8, note that Gleb) AB = Gaz, 0) 0 — de ‘Therefore the expression for dv becomes dy = La(2,,0)d0 + Ade, RET Mazimnum Value Punctions a which includes (5.3) and (4.2) as special cases. Some Choice Variables Fixed When the parameters @ in the constrained optimization problem change, so do the optimum choice vector and the maximum value v. But we have seen that the first-order effect on v can be caleu lated by holding 2 fixed at the old optimum and finding the partial effect of @ on». If the parameters affect the constraints, we must remember to include the contribution of the equivalent reduction in the right-hand-side magnitudes ¢, but we still ignore the ehange This suggests a generalization, Is the effect on » th ‘when only some components of x adjust to their new opt ls while others rnust be kept fixed at the oxiginal levels? isons of this kind are commex in econonuies. ‘The ios! poo example is the distinction between the sont run ated He Tong run rome quantities that can be varied optintally in the long eum nnust be held fixed in the short run. nat lev This question can be tackled using the saune ealeuls method tis was used in deriving (5.1) and (5.4). But the geometry that provided the intuition for the envelope property «lor thw jb far tore simply. Let us begin by stating the question somewhat noe precisely. Partition the veetor z into two subvectors y saul =, Sulsuning the right-hand sides of the constraints into the psuranicier vector as explained above, write the problem as: maximize F(y, 2,0) subject to Gly,2,8) (55) There are two versions. In the long run, bot y aud = are ehoiee vectors, while in the short sun, = is held fixed aan] only allowed to vary. For the latter problem to be meaningful, the ynunlier of constznints must be less than the dimension of y Write the long-run optimum choices and the resulting value ts functions of 8, say y=), 28), In the short run, = should be treated as just another paras along with 8, and the optirmim choice y and the resulting value v are functions of (2,8), say ¥=¥(28), =V@). (56) ¥ (2,8). (7)62 Optimization in Beonomic Theory ‘The use of the same symbols ¥ and V to denote different func tions in the two eases should not eanse confusion since the distinet ‘arguments will be displayed as appropriate. The definition of optimization gives at once V()> V2.8) for all (2,0), with equality if 2 = Z(8), the optinnal choice, ‘Then, just as in Figure 5.1, the graph of V(8) is the upper envelope of the curves showing V(2,8) as functions of @ for the whole range of possible values of Ifthe functions ate differentiable, we ens: conehade that VI(B) = Vo(Z8),8), (58) where the right-hand side is the partial derivative of the short-run ‘optimnan value function V(=,8) taken holding the first argument 2 fixed, but evaluated at the point = = (0), Now the geometry alerts us to a poten problem that the caTeulus approach would have conecaled, namely that the functions nay not be differentiable. Even when the mnderlying objective and constraint functions F and @ are as sinooth as one might like, the optimum value funetion V can have sudden changes of slope, We saw an example of this in connection with the interpretation of Largange multipliers in the last scetion of Chapter 4, aud in Fig ure 4.1. When there are inequality constraints, or non-negativity constraints on the choice variables, these eau be binding for one range of parameter values, and slack clsewhere. The objective fimetion may respond differently to parameter changes depending ‘on the configuration of binding or sack coustrniats. At the point where there isa regime change, from binding to slack or vice versa oh of the maximum value function suay have a kink, Ta many applications we will not be concerned with such reginne ‘changes and will be able to use (5.8), but the possibility of its fhilure should be kept in mind, In seine contests such as linear programming, changes of slope necessarily arise as the parameter values move from one combination of tight and slack constraints to another Mosimum Value Functions 63 Examples Fxample §.1: Short-Rum and Long-Run Costs As an illustration of the Envelope Theorem on its home ground, ronsider the relation between short-run and long-run cost curves for the production function Qa(REy (39) where Q is output, J is capital fixed in the short run, and L is labor. Returns to seale are constant if a = 2, increasing if a <2, rand decreasing if a > 2 Let w be the wage rate and r the user cost of capital (He rental price of capital services if they are rented, or the sum of interest and depreciation costs if capital equipment is purchased). ‘The long-run cost function is C(wy7,Q)= ip (wh tek [KL =Q" } (5.0) Using Lagrange’s method, the cost-minimizing input choices are seen to be Ka(wQry, b= (rQ*fw)' (Ga) ‘Then (0, 7,Q) = 2lawry? Qe (5.2) oc Exercise 5.1 below for a snore gen In the short run, there is no frevdom of chive. Iontpnt Q is lo be produced using capital I, labor L = Q°/ um amid the cost function beconies he hired, Clr QW) = Qe pe ke (9.13) If there were a third input, say als, whose quantity ean be vatied in the short ran, tere would be a short-rm cost= ‘minimization problem to be solved. Ushll leave this: asa exercise ‘The long-run marginal cost is found by differentiating (5.12): Catw,7,Q) = awry"? Qre (5.14)ot Optimization in Economic Theory In the short run, (5.13) gives Cale. Q,K) = awQ™/K. (6.15) If the value of 1 happens to be the long-run optimum given by (5.11), then the short-run marginal cost (5.15) and the long-run marginal cost (5.14) coincide, as the Envetage Theorem reauires. Ezample 5.2: Consumer Demand ‘The most important new idea introduced this chapter was to regard the maxinzum value of the objective as a function of the parameters of the problem, Such functions contain a lot of eco: nomically useful information, which ean be used to simplify the treatment of optimizing behavior in many applications, This ex ample treats the ease of consumer demand theory based on utility ‘maximization, Consider a consumer who maximizes utility U(2) subject to the budget constraint pr = I, whore p is a row vector of prices, 2 ‘column vector of quantities, and J is money income. The para meters of the problem are p and I, and the resulting maximum utility is a function V(p,1). This is called the indirect utility fune- tion, to distinguish it from the direct utility funetion U(r) defined ‘over the quanti Some properties of V are evident. For example, changing all priees and incoine in the same proportion leaves the budget. con- straint unchanged, and thus does not aifect the optimal choice or the resulting utility. Therefore V is homogeneous of degree zero in (p,J). We will have oceasion to study some other properties later. The focus of interest here is the application of the Envelope ‘Theorem, or snore specifically, the formula (5.4). Note that the Lagrangian is Le dyp.T) = Ule) + AUT ~ pa) Therefore Vind ‘evaluated at the optimum. Similarly the column veetor of deri tives of V with respect to the prices is Lyle, dopsP (5.16) Vol(p, 1) = Ly(2,A,p,0 (17) Mazimum Value Functions 6 yin evaluated at the optimum, Of course the utility maximizing, tantites 2 comprise the (vector) demand function D(p,1). There fore we can divide (5.17) by (5.16) and write Dip.) = Ver. D/VilPs 2). (6.18) This is a useful and important result, If we are given the consumer's (direct) utility function and asked to find his demand functions, we have to carry out the whole constrained maxis tion solution, which is messy even in the simplest eases. On the other hand, if we are given his indirect utility function, we eau fine the demands by differentiation alone, Thus it is nich sinpler to marie means of inulireet lity functions, Particularly in general equilibrinnn wwdels where consumers are only one part of the story, th ‘of notation makes a great deal of diffe itances of this in the chapters to follow. ‘Next consider the mirror-image problem where the ix scen as minimizing the expenditure rquived for uttai target utility level. The resulting mininman exp tion of the price veetor p and the target utility level». Call it the: consumer's ezpenditure function, and waite it E(y.u). Keeping fixed and changing all prices in'the same proportion will simply. slunge the necessary expenditure by this proportion; therefore B ix homogencons of degree one in p for fixed w. Other pruyyrtivs of E-will be developed as necessary. One again, ons fens is out the velop property Write the Lagrangian for the minimization robles ur information about conssiners by cconony of ent nee, Wee will see sone litre is Tune: Lapp.) = pak ulu U2) Arguing as bofore, we have Ex(pyu) = (5.19) “This Lagrasige multiplier gives the inerense in expeaiture reyuied to achieve marginal increase in the utility level. ‘Therefore itis just the reciprocal of the marginal wtlty of mony X above. Next Ex(p.u)66 Optimization in Economic Theory Cost-tninimizing commodity choices for a given utility level are the Hicksian compensated demand functions C(p,u). The situation is as if, following any price change, the consumer is compensated by changing his snoney income just enough to leave him on the sans inulfference eurve as before. Thus we have shown that Clu Ex(p.w). (5.20) This expression is even simpler than that for the uncompensated demand functions D above. Finally, we can relate the indirect utility function and the expenditure function, and thereby the uncompensated and. cox, petsated demand functions. Suppose we begin with a utility lew u, and find = E(p,u). Then we assign this J as the money in come, and find the uliy-maxiizing choice. So long as all price are positive, as will be the case in most elementay econotie a plications, this gives back the utility level we started with, thats = V(p. 1). The cost-minimaizing ehoice of x is also the atility maximizing choice, that is, C(p,1) = D(p, 1) so long as u and 1 ae related as above. Take the jth component equation, and dif ferentiate it with respect to pg. Hold u fixed, but make I = Ep.) a fonction of py. By the chain evle Llp.) = Dip.) + Dito.) Eel.) But Exp) = CM py) = Dip.d), the demand for good k, Therefore Ch(p.u) = Dip.) + DY.) Dip.) (a This relation compensated demand functions is called the Slutsky-Hicks equa tion, Some readers may know it in a different notation: ip between the derivatives of compensated and wn Mazimum Value Functions 6 Broreive 5.1: The Cobb-Douglas Cost Function Consider a production funetion =4]]p a, (522) where y is output, the 2, are inputs, and A and the a, axe postive constants, Let w = (1;) be the veetor of input prices, and show that the minimum cost of producing a given entpiat level y i Cw,y) = Ayal? TT (sso, where 8 = Oj aj. 1A <1, enlenlate the corres profit function r(p,u) where p is the ontpast-pr wrong if 3 > 17 Brereise 5.2: The CBS Expenditure Punetion Suppose the direct utility function is Uley) = las? + sy any where x and y are the quantities of the two goods, atl 1, andl are given constants, with a, 8 positive aud p = 1. Slow Unit the expenditure function is of the form Ep.q.u) = lap + by] (5.25) where p, q are the prices of the goods, w is the tility level, saul a b, and r'are constants that can be expressed in tern: of o,f, ane a Find the compensated demand functions saul shui that the ratio of the cost-minimizing, quantities is aly = (of) (ale) ‘The elasticity of (2/y) with respect to (a/P) atn(2/u) Tin (a/p) the elasticity of substitution in proxrtion. Show that in this example, it is constant andl equal to (1 =»). What cond tion must be imposed on p to ensure a noxt-awgative elasticity of substitution, that is, r-<1?68 Optimization in Economie Theory Further Reading [An interesting ecount of the discovery of the Envelope Theorem Jacon VINER, ‘Cost curves aud supply curves’, as reprinted in George J. Stigler and Kenneth E. Boulding (eds.), Readengs in Price Theory, Homewood, IL: Irwin, 1952. Samuelson clarified the confusion that sursounded Vinee’s dis covery; his original exposition je still worth reading: PAUL A. SAMWASON, Foundations of Beonomic Analysis Cambridge, MA: Harvard University Press, 1047, pp. 34-6. For further discussion of the Envelope Theorem, see Vasian, Micrneconomie Analysis (op.it.), mathematical appendix, or DAVID KREPS, A Course in Microeconomic Theory, Prince ton, NJ: Princeton University Press, 1990, el. 7. For more detailed tseatments of the indirect utility function ‘and the expenditure function, see Varian (ibid. ch, 3) and Kreps (bid. ch. 2); for more on a firm’s cost and profit functions, Varian (ibid, ch. 1) and Kreps (ibid. ch. 7), 6 Convex Sets and Their Separation ‘The Separation Property Lagrange’s Theorem and the Kuhn Tucker Theoret ive 1s fies order necessary conditions for constrained maxisnization problems, The conditions are not in general sufficient to deteswuine the op: timum. The same necessary conditions would arise if wer wanted to minimize the same objective function subject to Une sane cont straints, When I discussed this point in Chapter 2, 1 mentioned that maxima and minima can be distinysished by examining the curvatures of the objective and consteaint functions, ‘Phat idea i developed in this chapter and Chapters 7 wil 8. Bogin with the problem of choosing the welor wr te aaxinize F(z) subject to a sealar constraint G2) <¢. Let sr driuote tle optimum choice, and now write # for the anaxinnuy value Figure 2.1 showed the familiar tangeucy; wow I shall interpret the solution in a new and useful way, That figure sbuwed tae contour where G(2) = ¢. Now we need to know the set of all points that fulfil the more general inequality constraint Gli) <6, oF the lower contour set of G for the value c. In Figure 6,1 it i Uh shaded area A below the contour G(z) = ¢. Similily, Figger showed only the contonr where the F(z Now we ace the set of all points where F(x) > @, or the upper contour art of F for 4, This is the shaded area B in Figure 6.1. ‘The fguire asannes F and G@ are increasing functions, This is true in sunny eronexnie applications, as when F is a welfare function anel Gs roquirement function of the outpat vector 2. But « similar figure ean be drawn for other eases ‘The two curves meet tangentially at 2, and the figure shows their common tengent line at this point, ‘The eurvatures are chosen, to ensure a maximum. In Chapter 2 the curvatures were said to reflect diminishing marginal rates of substitution and transforma tion; now T offer a somewhat different interpretation:0 Optimization in Eeonomic Theory Fig. 6.1 - Separation by the common tangent Note that the sets B and A lie one to each side of their com mon tangent, with only their common point # on thnt Tine. In other words, the common tangent separates the z-plane into two halves, each containing one of the sets, In higlier dimensions, the common tangent will be a hyperplane; it will likewise separate the two contour sets This separation is the crucial property that allows us to dis tinguish maxiraa from minima, and obtain sufficient conditions for the maximization problem. Now we must make explieit the hid den conditions on the functions F and G that ensure the right curvature, Convex Sets and Functions Ench of the contour sets in Figure 6.1 bulges outward. That is why each bends assay from the coramon tangent at Z, and cannot I back to meet the other set once again. This property of \walging outward is called convexity. A geometric test of convexity i that given any two points of the set, the whole line seginent joining them should lic in the set. Mlgebraically, a set $ of points Conver Sets and Their Separation a rndimensional space is called conver if, given any two points 2h 8) and 2° = (ah at, zh) in $ and any real imumber 6 in the closed interval [0,1}, the point (@ 2° + (1—6)2"), (Oxy +L ~A)z}, O25 +(1— Oded... 2h + (1-8) ) when written out explicitly in terms of the components is also in 8 Applied to the lower contour set of G, this means that if 2* amd 2° satisfy the constraint, so does @2* + (1—6).°. In eco: asic applications, constraints often reflect Emited availability of resources. Thus G(x) might be the amiennt of labor necessary to produce the vector 2, and c the ainount of labor avaiable. fi sch ‘eontext, the converity ofthe set Gl) min( Fle"), FUN), (G2) for all 2%, 2% and for all @ € [0,1)- Sucl a fusetion will be called squasi-concave,m Optimization in Economie Theory Fig. 6.1 ~ Separation by the common tangent Note that the sets B and A lic one to each side of their cont mon tangent, with only their common point # on thnt tine. Ia other words, the contmon tangent separates the x-plane into two hhalves, each containing one of the sets. In higher dimensions, the conmnon tangent, will be a hyperplane; it will likewise separate the ‘ona contour sets, separation is the erucial property that allows us to dis. tinguish maxima ftom minima, and obtsie.sulficient conditions for the maximization problem. Now we amust snake explicit the bid: den conditions on the functions F and G that ensure the right curvature, Convex Sets and Functions Each of the contour sets in Figure G1 bulges outward. ‘That is why each bends away from the common tangent at F, and eannot bend back to meet the other set once again. This property of Dulging outward is called convexity. A geometsic test of convex is tat given any two points of the sel, the whole ine segmers joining them should lic in the set, Algebraically, a set $ of points Conver Sets and Their Separation a n-dimensional space is called conves if, given any two points ata, ... 28) and 2) = (z},24, 28) in S and any real number # in the closed interval [01}, the point (82° + (1 ~#)2"), (G2$ 4 8)a4, Ox + (1 — Bah, BR + (1 B)a%) when written out explicitly in terms of the components, is abso in 8 Applied to the lower contour set of G, this means that if 2 id x? satisfy the constraint, so does 62" +(1—@):t*. In eco- applications, constraints often reflect insited availability of resources. Thus G2) might be the amount of labor necessary to yeoduce the vector 4, and the amount of labor avi context, the convevity ofthe set G2) < e aust that weighed twerage of two production pluses does wt seed siete Tabu ta one of the extremes. This rules out siguficwit econosies af scale or specialization, Sinmiladly, applied to the yppet content set of B ble. lu scl convexity means that « weighted average of xo ronstamption plans is at least as good as one of the extremes thi ssi a taste for diversity. Algebraically, the condition states that Ga") Scand Gla") Se imply GGr" (1 Met) ‘The most severe test of this arises when ane of the extreme sales equals c, therefore we ean state the condition alternatively as G(6r* + (1-024) < max (Gle") GU}, (63) for all 2%, 24 nnd for all @ € [0,1]. Au dled avdvinntage of this form is that it does not involve the particular number «Since in practice we will have to solve the unsimization problem fo general value of ¢, we will ueed to invoke the canelition for all 6, and a general statement like (6.1) A function G satisfying (6.1) is said to be quesi-ronpea. The parallel condition on F will be the best sy for des FO=" 4 (1-24) > min FL"), FUN), (02) for all 2, 2* and for all @ € [0,1]. Such a fmetion will be called spuasi-concave.2 Optimization in Economie Theory ‘The quasi in the above definitions serves to distinguish then from somewhat stronger properties that often arise in optinnization problems. In the usual economic interpretations, quasi-convexity ‘cortesponds to a diminishing marginal rate of transformation along the constraint curve, while convexity of the constraint funetion cor responds to diminshing returns to scale. Similarly, quasi-coneavity of the objective function means a diminishing marginal rate of sub- stitution along an indifference curve, while concavity is diminishing, rmargival utility, Formally, we define G to be convex if G(Ox* + (1 8)x*) < OG(x*) +(1-8)G(x"), (6.3) for all 2%, 26 and for all 8 € [0,1]. The right-hand side of (6.8) is obviously no larger than that of (6.1) 8G(24) 41-8) U2) <8 max( G(x), Gle*)) + (1 ~6) maxi G2"), G2") < max( Gl"), G(2")). Therefore if (6.8) holds, so does (6.1). In other words, a convex function is quasi-convex; convexity is the stronger property of the two, Similarly, we define P to be concave if F(Ox" + (1 ~)e') > OF(x*) + (1-8) F(2*)), (6.4) for all x4, 2 and for all @ € [0,1]. In words, the graph of the function lies on or above the chord joining any two points of it Figure 6.2 illustrates this for the ease a scalar. It is easy to verify that (6.4) implies (6.2): a concave function is also quasi-concave, Note that the inequalities in (6,3) and (6.4) are weak, ‘Thete- fore convex and concave functions are allowed to have straight-line segments in their graphs, where the chord coincides with the graph, Tn particular, a linear function is simultancously convex and con- cave. Later we will have occasion to strengthien the concepts of convexity and concavity. An alternative definition of a concave function is sometimes ‘useful. Consider the (n+ 1)-dimensional spare consisting of points Convex Sets and Their Separation 3 tangent Fa) (FLA) chord (Ak) Fig. 6.2 ~ Concave function like (2,») where 2 is an n-dimensional vector and v is m scalar, In this, define the set, F ={(z,v) |v S F(2)}, ‘Then F is a concave function if and only if F is a convex sets check. this using (6.4), Tn other words, n concave function traps x convex. set underneath its graph, This is easy to sce from Figui 62 Similar properties of convex fanetions are equally easy to stat so I shall leave them to the reader. Fizally, if the fimctions are differentiable, the properties of concavity, quisi-voncavity ele. eam bbe expressed in terms of first- nnd sero order derivatives; Tsall ddo this in Chapter 7 T must define two more concepts to be able to state the basic mathematical result I neai. A point 2° € Sis called an interior point if it is surrounded for some distance by points of the set, that is, if there is a number r > O such that all points x within distance of 2” are in S. Ia the plane, such points will forn a dise of radius rand centre 2°, Then, a point that is interior neither to $ nor to4 Optimization in Beonomic Theory tue rest of the space is called a Boundary point of S. ‘Thus 2? is a homudary point of $ if, for any r > 0, we ean find points in $ as well as points not in S within distance r of 2”. Iu Figure 6.1, for example, 7 is a boundary point of B and also a bonulary point of A. Any 2 for which F(z) > 9 is an interior point of B so long as F is continuous. Similarly any point 2 with G(2) < eis an interior point of A so long as G is continuous. The two sets have only the boundary point 7 in common. 1ud the common tangent separates them, Let the equation of this tangent be prepini ter where p is a row vector of coefficients, so pz is the inner product or the sum of products of corresponding components, of p and x Of course p # O, that is, at least one of py and pp must be non-zero, Since 7 lies on the separating line, we have pe pi B+ pao For all points x to one side of the line, pz is greater than 6, and for all those on the other side, itis less. If the sets had no points in common at all, there would be fa clear gap between them, and once again we could draw a line that separated them, although it need not be a tangent to cither set. But if the sets had interior points in common, then any line entirely above the set A would have to ent into the set B, and any line entirely below B would have to cut into A; separation would be impossible, Convexity of the sets is also important; Figure 6.3 shows two ‘cases, in cach of which one of the sets is not convex. The common tangent euts into the non-convex set, and separation fails, lized in the following theor hope most readers will ind the pictorial argument 0 shall not give a formal proof. I shall only state the result in the simplest form that suits my purpose, even though more general results of this kind are available. All these ideas are for Separation Theorem: If B and A are two convex sets, that lave no interior points in common, and at least one of the sets has 8 non-empty interior, then we ean find non-zero vector p and a Conver Sets and Their Separation % fa) ny Fig, 6.3 ~ Partial failue of decentralization umber B such that the hyperplane p2 = 6 separates the two sets, a 8. This is the basic revolt of this section, Optimization by Separation: Given a quasi-concave function F and @ quasi-convex function G, the point ¥ maximizes F(x) subject to the constraint G(2) < eif, and only if, there isa non-zero veetor such that (2) maximizes pz subjert to Ble) So, and (Gi) # minimizes px subject to F(2) > 8, The generalization to several constraints is straightforward The set B, of points for which G'(z) < ¢; is convex if G* is quasi convex. If this is so for all i, then the set B of points satisfying all the constraints, being the interseetion of the convex sets Aj, is itself convex; this is easy to verify using the definition of a convex set. Then the argument proceeds as above. Note the ‘if and only if” in the statement of the result. The stated conditions are both necessary and sufficient for optimality if 2 is optimal, the conditions will hold, and vice versa. In both respects, the result goes beyond the Lagrange or Kutin-Tucker conditions we met before. This is the first time we have met a sufficient condition, therefore it is of interest in itself. But the conditions are not ensy to verify in practical applications. After all, we have replaced one optimization problem by two, and have given no useful criterion for judging when either one of them solved. In the next two chapters we shall see sufficient conditions that are more useful in this regard, ‘The real benefit from splitting the maximization problem into two separate problems, in each of which the objective function is linear, comes from its economic interpretation. Tt raises the pos sibility of decentralizing optisnal resource allocations using prices, Conver Sets and Their Separation 7 ‘To give Ue simplest interpretation, suppose + is the production- ccun-consuimption vector, the constraints reflect limited resource availability, and the objective is the utility function, Now inter pret pas the row vector of prices of outputs. ‘Then part (i) ofthe above theorem says that the optimum # would be produced by an entrepreneur secking the maximum value of output, and (i) says that would also be demanded by a consumer trying to reach the target utility level & in the least-cost manner. If we assume away some technical complications that arise when there are free goods, this is equivalent to maximizing wility subject to the budget cone straint p2 vv. Then we can find a neighboring point that is still in B, and has a smaller value of pz than 2, To be formal, take positive, and define 2° by. B= ep) for Because # i an interior point of B, fore sinall enough, 2 isin B. And , pa! =pi—ep pet p3)—aF beer? 0. This we have found a point + € B with pr < h contradicting the separation property. This contradiction stems from the initial supposition that # was not sigue, ‘Therefore that oppesition must be wrong. Thus strict quasi-concavity of F im- plies the uniqueness of the maximizes Striet quasi-convexity of G womld do equally well. When there sce several constraints, we have to assume evtry component con niraint function G! to be strictly quasi-convex to ensure stone Convesity ofthe constraint set A Examples Example 6.1: Mlusiration of Separation Suppose there are two non-negative choice variables labeled 2 and y, and the functions F and @ aro given by +y F(a,y) y and Glau) The constraint is G(r, y) < 25. Figure 6.5 illustrates this. "The fensile sot Al consists of the aquartercirelo and all points below it; houndnses of the upper con- tour sets of F for various values » are shown as + fuily of rectan- ular hyperbolas. The optimum oceurs at (3.1) ~ (5/v2,5/ V3), and the maximized value of F(z,y) is @ = 123. ‘The upper contour set B corresponding to ® touches the fens- iM ot atthe optimum; they are separated by the common tangent82 Optimization in Economic Theory Fig, 6.5 ~ Ilustration of separation 44 y=5V2, The upper contour set of F for the larger value 1s hes no points in common with A, aud we ean draw 8 separating Tine ry 8 through the clear gap between them For a stale value than 8, s8y 10, the upper contour set of Frand the fasibl set tase inttior points in cominon (the lene-shaped area of tht intersection), and the two cannot be separated Ezample 6.2: Indirect Utility and Ezpenditure Functions A utility maximizing consumer's indirect utility funetion and ex penditure fanction were defined in Example 52. Here I shall ex famine their convexity properties. Begin with the expenditure function E(p,u) = min (pr | U(2) 2 u} (65) This is concave as a function of p for cach fixed w. To see this, tak any two price vectors p* and p', and any number @ in [0,1]. Let PF =p" + (1— Op. Then we need to show that (pu) 2 OE" u) + (10) E(w). Ca Convez Sets and Their Separation 83 Let xf achieve the expenditure minimization for p®, OF course x" must satisfy the constraint, so U(2") > u. The constraiut does nnot involve the price vectors, so 2° is also feasible when the price vector is p* or p, In each ease it cannot achieve a sinaller exper diture than the mainimum: phat > Rip) and ph2°> Elphu). (68) Multiply the first of these by 8 and the second by (18); this leaves the directions of the inequalities unchanged sinee both multipliers re non-negative, Adding the two, piat = (Op" + (1— B)ph) 2° > OE(p*,u) + (1 = 8) Ely) ‘The left-hand side is just E(p',u). This proves (6.7). ‘The economic intuition is that as the priee yeetor changes: ene could leave the quantity vector wicluniged. Then the expenditure would change linessly with the priee. To the extent that there ix substitution along the indifference curves, the quantity choice ‘can he adapted to the changing prices. This will change the ex: penditure slower than linearly, that is, the minimized expenditure will be a concave function of prices. “Another way of looking at this is to exataine the worst ease, when there is no substitution, in consumption. The indifference curves are L-shaped, and x is the optimal way of achieving the utility level u, irrespective of the prices, ‘The two inequalities in (6.8) hold as equations, and then (6.7) is an equation, too: expenditure is linear in. prices. ‘Next the indirect utility function, V(p.D) = max (Ula) Lee ST} (69) This is quasi-convex in (p, 1). The proof follows i sini (pF) and (p',2*) be any two price income wetors, and @ any amumber in (0,1). Let (PT) =O TE (= KT, amd suppose 2° solves the atility- maximization problem for (p° I). It satisfies the constraint, 20 por" it and phot >t4 Optimization in Economic Theory Multiplying the first by @, the second by (1 ~ @) and adding, » would get p' 2° > I*, contradicting the feasibility of 2° for (p°I° In whichever situation 2° is feasible, its utility U(e*) eanne cexecod the maximum utility achievable in that situation. Therefor: ‘at least one of Ul) < Vi) and U2) < Vip) rst be true, Then V(r 1) S max( Vip", I"), (PhD), which is just the statement of quazi-convexity of V. In other words, the lower contour sets of the indirect util ity function are convex, This has an unfortunate consequence When the xoverament chooses indivect taxes optimally, itis in ef fect choosing prices to maximize an indirect utility function. On result says that the objective function has the wrong curvatuy for a maximization problem, ‘Therefore sufficient conditions fo» ‘optimal tax problems are hard to verify Exercises Exercise 6.1: Commodities that Cause Disutility How is Figure 6.1 altered when (a) one of the choice vail is labor, which gives disutility to consumers and is an input 1 production, and (b) when one of the goods is pollution, sshich giv: disutility to consumers and is the by-product of an economically esirable good which is the other choice variable? Interpret tli associated prices in each of these contexts. Exercise 6.2: Convesity of a Firm's Profit Function A firm chooses vectors ¢ of inputs and y of outputs subject to » production possitility constraint G(2,y) <0, to maximize profi (qu ~p2), where q denotes the row vector of output prices and p that of input prices. Let II(q,p) be the maximized profit expressed as a function of the prices. Prove thnt TT is a convex function uf (ap) Convez Sets and Their Separation 85 Bzercise 6.8: Corner Solutions Consider an economy with Iabor endowment L. Tt can produce two goods 2, and rz. A unit of good j needs a fixed amount a, units of labor, so the production possibility constraint is ort tan 22 Fla) - Fla") (74) When x is one-dimensional, F,(x*) is just the slope of the tangent to Fl) at 2%, 90 the lethand side isthe vertical distance along ths tangent as we move from 2° to 2°. Then (7) soy that the change in the value of concave function is overestimated hy its tangent, that is the tangent lies above the curve. More geucrally, the lefchand side is just the Tear tern: in the Tylor snpproxt imation to the change in F(2) using 2* as the intial point this npproximation isan overestimate ofthe disage in Fe} Similarly, if G is adiferentiable convex fete, tea Gr(2*)(2' ~ 2) < Gla) ~ Gla"). (72) A particularly important class of opt sins eon eave objective fimetion and convex constraint functions the term concave programming is often used to describe the general problem of this kind, and it is the subject of the next section. Concave Programming Consider the maximiration of F() subject to a vector coustraint G(x) < ¢, where F is differentiable and concave, and exch com poneat constraint fuuction G' is differentiable aad convex. This i called the generat problem of concave programming. For con- ereteness and economic intevest, J shall use the terminclogy of the production problem, where « is the vector of outputs, F(z) is the revenue from the sale of the outpats, ea fixed veevor of input sap- plies, and G(x) is the vector of inputs needed to produce . But the mathematics is independent of this interpretation, The conditions of optimality for # partieular value of © are found by first considering the problew for a geval ¢, ‘Tlhen the optimum choice of 2, say 2, and che masinmon value «= oth become functions of ¢. Let X(c) denote the optinnnis function, and V(c} the maximum value fonction. ‘The first result is that V is a non-deereasing fimetion, This is bocatise an 2 that was feasible fora given value of e remains feasible when any component of ¢ increases, so the maximum value cannot decrease.88 Optimization in Economic Theory ‘The next result is that V is concave, Let ¢ and c! be any two input supply vectors, with F=X(0), #=X(C), =Vehy 8 = VEC) Since the optimum choices must be feasible, we have GZ) < ean Gat) <¢. Now let @ be any number in [0,l}. For V to be concave, i should be possible to achieve revenue at least as high as 8 V(e) 4 (16) ¥(c) when the input supply vector is Be + (1-8) eA natural eandidate for the output vector is 87 + (18) 2". The fs point to cheek is whether this is feasible. For each i, the convexity of Gt implies (2 +(1~ 8) 2") F(z) + (1 ~ 8) F(a) 2 Ov + (1B) ‘Thus we have found a feasible output veetor that yields revenue 1! least a high as the expression on the extreme right of this clusis of inequalities. The maximum value, V(8e+(1—8)¢'}, can be 1 smaller. This is the result we want to prove. ‘The economies behind this is that the convexity of G rule: ‘oat economies of scale or specialization in production, ensuring that a weighted average of outputs can be produced using the same weighted average of inputs. Then the concavity of F ensur that the resulting revenue is at least as high as the same weighted average of the separate revenues. ‘As V is a concave function, the set of peiuts on or below it graph is « convex set, This is an (rm + 1)-dirmensional vet, tle collection of all points (c,v) such that v < V(c). That is, revewx of at least » can be produced using the input vector e. Therefor it is natural to think of it as the set of production possibilities for revenue. Figure 7.1 shows this set as the shaded area A in the cass where c is a sealar. Since V is non-decreasing and concave, the Concave Programming 50 Fig, 7.1 ~The value function in coneave prograsnsning vt has a frontier that shows a positive but diminishing marginal produet of the input in producing revenue. Convex sets are meant to be separated from other convex sets, "To do this in the most useful way for the present purpose, choose a point (c*,0*) in A such that o* = V(¢*). This must be a boundary point, since for any r > 0, the point (c*,»*—r) isin A but (c*,o* + 1) is not. Now define Bas the set of all points (e,v) sueh’ that eS et and w > v", that is, revenue v cannot be attained with nputs ¢ except when ¢ = c* and v = 0°. Thnx the set B serves the same role as the corresponding set in Chapter 6, Cleasly B is n convex set with a non-empty interior, and A andl B have only hnoundaey points in common. ‘Therefore the Separation Thewre can be applied. For reasons that will become elear ina moment, 1 write the equation of the separating liyperplane as tw how ba unt Ae, where tis a scalar, \is an ra-dimensional row vector, saul the signs nv so chosen that {Sb forall (eA wae{S} frattees, 8)0 Optimization in Bconomic Theory ‘The first point to note is that ¢ and A must both be now negative. For exeinple, suppose 4 is negative. Now consider the point (c*,o* + 1}, which is clearly in B. We have ao" FU) -Ae = beh, which contradicts the separation property, Sins, for enck & 1, 2, ..-m, considering points (c* eu"), where e is a veeten with its ith component equal to 1 and all other components zero. wwe see that Ay cannot be negative Now comes a more subtle question: ean ¢ equal 2er0? Let as sve the consequences of that, For the equation of the hyperplane to be meaningful, the combined vector (1,2) must be son-zer0, Ife = 0, therefore, at least one component of A must be now zero, that is, positive. The equation of the separating lyyperplatw becomes —Ne = —e*, or Me ~ e*) = 0. For all (e,0) i A, se have “Ae S —e*, or Afe—e*) > 0, In the scalar constraint ese te separating line is vertical at c*, and the set A lies entirely to the right of it Figure 7.2 shows two ways in whidi thie can happen. In both cases, there are no feasible points to the left of e*; predustion is impossible if input supply falls short of this level, Tn some appli cations, this can happen because of indivisibilities. ‘The two eases Uiffer in the behavior of V(c) #8 ¢ approaches e* from the sight In case (a), the marginal revenue product of the resouree goes ts infinity, and only a vertical separating line will do. In ease (bh) the limit of the marginal revenue product stays finite, aud while vertical separating line exists, there are also many other separatine lines with finite slope, aud therefore positive ¢, This shows that the conditions soon to be found for ensuring a positive « axe only sufficient and not necessary. ‘The natural condision is to rule out indivisibility, If the set A has any points to the left of c*, then it cannot hive ait infinite slope at ct. For this, there must be an 2° such thst. G(s") < e and F(2°) is defined; then we can choose (G(z"), F(a") as the desired point in A. If there are several constraints, we noed thi corresponding vector inequality G(2°) < ¢*. This then is the constraint qualification for the concave programing problem, It is sometimes called the Slater condition, ‘To prove that the Slater condition implies a positive 1, suppow Unt the condition holds but « =. Then at least one component of Concave Programming a PEK (a) Fig. 7.2 ~ Failure of the constraint qualification A must be positive. Now every component of G(r?) — negative. Then is strictly eye ee AGE", =e) <0, Ieause at least one component product ou the right side is negative, and all are non-ponitive, But the point (Gtr), Fe") it in A, therefore by the separation property AG?) =e Fle) 1G fo M(G(22) — 67) 20. We have proved the sae expression te be both, conclude that the supposition ¢ = 0 umust be wrong. ‘The separation propetty (7.3) is wnaffictesd if we multiply 6, 1, and every component of \ by the snaue peitive immer, Once we ean be sure that «#0, wo ean choose this senie to make « In cconomnie terms, « and A constitute a systen of shalow prices, sgative and non-negative: the eoatsndietion forces us too Optimization in Economic Theory + toa th revenue and 2 for the inputs. Only relative prices mut tor for economic decisions, and in setting « tevenme to be the numéraire. This seems an obvious choice, axl | shuall adopt it heneeforth, But sometimes we might wish to do wrwise; for example F(z) might give the revenie in x force 1, we are choosin, rnieeney, and then ¢ would be an exehange rate to convert it inl vestie currency units. The important thing is to establish eon Intions under which the marginal revenue produet of inputs at the ptinuum is finite, ensuring that the proposed numeraire is not « free wood. Next observe that by the separation property, (e",") achiev Uwe auaeimum value of (v ~ Ac) among all points (¢,) € A. This has an important economic implication, If we interpret A as thy wector of shadow prices of the inputs, then (» — Le) is the profi ‘that accrues when a producer uses inputs ¢ to produce revents 1». Saee all points in A represent feasible production plans of this kind, the result says that a profit-maximizing producer will ick (¢7,0"), He need not be aware that in fact the availability o! iupnts is limited to ct, He may think himself free to choose a ight c*. Th hii the searcity. The interpretation is special, but the principle general and important: constrained choice can be converted int «, but ends up choosing the prices 2 bring home ts unconstrained choice if the proper searsity costs or shadow vals of the constraints are netted out of the criterion function. To tix economist, this is the most important feature of Lagrange's Meth! in eoneave prograimning, The shadow price interpretation of A can be confirmed som (e)) isin A. So's what more formally. For any e, the point (c, the separation property we have V(e) Aes V(et) Ae’ V() — Vie" S Mee), (ra ‘The linear function on the right overestimates changes in the vals of V. This looks very much like the concavity property (7.1), a! suggests that A should equal Ve(c*), the vector of partial derivative of V at c*, That would make A the vector of the marginal oven roilucts of the inputs at the optimum, or the vector of shadow for the constraints. But one difficulty remains: we ennin Concave Programming 93 bbe sure that V is differentiable. So far in this chapter we have n ven assumed F and G to be differentiable, but even when we do, V may fail to be. This ean happen for the same reason as we sass in Chapter 4 and Figure 4.1. Different inequality constraints may hold as exact equalities for different ranges of the parameters, and where the solution switches from one regime to ancther, the slope of V may change suddenly Exen when such discontinuities in the slope of V exist, a very natural generalization of the concept of diminishing returns holds, As the value of some component of ¢ increases, the corresponding, partial derivative of V may jump downward, hut not upward, This follows from the concavity of V. ‘The asterisks, having served their purpose of distingishing 1 particular point in the (c,») space for separation, may now be dropped. Let us consider a general point (¢, V(¢}) with its assoc nted multiplier veetor A. Compare this with a neighboring point where only the ith input is increased: (¢+he!, V(e-+ he*)) where hn is a positive scalar and et is a vector with its ith component equal to 1 and all others zero, ‘Then (7.4) becones V(c+ het)—V(c) < Ahe bhi Since h is positive, we ean divide by it to write [Viet het) — NSA, It is easy to show that by the concavity of V, the left-hauel side is a non-inereasing function of f, and therefore must have a limit 48h goes to zero from positive values. In a diagram showing ¢y and V(6), the expression on the left-hand side is simply the slope of a chord joining the point (c,V(c)) to an adjaevit point to the tight (because h > 0). Is kit is defined as the “vightWord” partial desivative of V with respect tothe ith coordinate of eal written V;*(e). Thus we have proved that Ve) ViNCe)om Optimization in Economie Theory a ua) Fig. 7.3 ~ Generalized marginal products ofa scalar o es, 2, has been ept inthe Figure 7.3 illustrates this for the ea So far the vector of choiee vari Dackground. Let us bring it in explieitly. Recall that & suaximizes F(z) subject to Glz) $e. Let \ be the vector of shadow prices of inputs found from the separating hyperplane as in Figure 7.1. The point (F(z), G(2)) is in A, so the separation property gives F(z) — AG(z) < V(e) = de (76) OF course F(z) = V(e), 90 Ale G(a)] <0. (a7) Now every component of the tow vector 4 is non-negative, and every component of the column vector [e ~ Gl2)] is sls non niegative. Therefore for every i, the produet Xs (cs — G'(2)] is now: negative, ‘The inner produet in (7.7) is the sum of these terms It can he non-positive only if each ter Therefore for every i, either A; = 0 oF ej ~ G'() = 0. ‘This is just the notion Concave Programming 95 of complementary slackness developed in Chapter 3, and it fits in well with the interpretation of A; as the shadow price of the Dh constraint aid any constraint with a positive shadows price must be binding Note an implication of complementary slackness: the inequalities in (7.6) and (7.7) must in fact hold as equations Finally, since (F(x), G(x) is in A for any 2, and since (7.6) holds as an equation because of complementary slackness, the sep: aration property gives ‘The shadow poiee of my slack constraint is 2esn, F(z} AG(#) < Flr) = Gir} for all s. ‘That is, # maximizes F[s) — AG|r) without any com straints, This is an alter ing choice variables, of how slides prices allow us to convert the ve statement, in terms of the underly original constrained revenuc-nvasiniiation problenn into an uncox strained profit-naximisation problem. All of the above reasoning can now be sununarized into the basic theorem of this section: Necessary conditions of Concave Progearnming: Suppose that F is a conease function and G is a veetor couvex function, and Lat there exists an 2° satisfying G(x") < e. Ife maximizes F(z) subject to Gl) <6, then there is a row veetor A such that (i) # maximizes Fle) — G(x) without any constraints, and Gi) 40, Gz) < ¢ with complementary slackness. of this requires F and @ to linve derivatives, But if the functions are differentiable, then we have the first-order necessary conditions for the maximization in (i), namely Fos) AG,() = 0. (79) In terms of the Lagrangian £(e,) of Chapter 3, (7.0) becomes: E,(2,2) = 0. This is just the condition (3.5) of Tay orem with Inequality Constraints, Heve I lid ot ianpos aay aos: negativity constraints on the choiee variables , but Hat was just to keep the algebra 9s simple as possible added on without cansing ange cw clilfiondies, seed ehow we get the corresponding condition (3.7) af the Khas Tucker Theorem. 1 leave it ser to verify this. auge's Th Suli conditions ean be an exercise for the96 Optinnication in Economic Theory There is one respect in w yond the general Lagrange or Kubn—Tucker conditions. The con Uitions of Chapter 3 merely set the first-order derivatives of the Lagrangian with respect to the choice variables equal to zero. Th was not sufficient to ensure maximization, and in general there ‘was 10 claim that # maximized the Lagrangian. When F is con cave and G is convex, part (3) of the above theorem on Concavt- Progaming is easily transformed into L(z,A) < L(2,) for all so @ does maximize the Lagrangian. The distinction does make ‘economic sense: we know that profit-maximization at given prices ‘can be problematie when there are economies of seale. Price must still equal marginal cost at the optimum, but profit need not be maximized, even in comparison with neighboring points. Tn the same way, our interpretation of Lagrange’s method ns convert ing constrained revenuie-maximization into unconstrained profit maximization must be confined to the case of coneave program, ming, ‘The previons paragraph was on the verge of saying that the first-order conditions are sufficient to yield a tive maxinzum in the ‘concave progeamining problem, That is indeed so, ‘The argument proceeds in two parts. First, suppose 3 satisfies the conditions (i) and (ii) ia the statement of the necessary conditions. Then, for any x, we have concave programming goes be FU) ~AG(a) > Fle) ~ AGI), or using complementary slackness, F(a) Ae Fla)— AG(2), Wa is feasible, Gl) < 6, and then F(a) 2 F(a) + Ale~ G2) 2 Fie) Next suppose # satisfies the first-order conditions (7.9). Sinee F is concave, G is convex, and A> 0, F— 2G is concave, Then (7.1) applied te this function gives [F(2) — 4G(e)] - IF2) — 9G=N) SE) — 4G (@)}( ~ 2). But the right hand side is zero by (7.9). Therefore Fl2)~AG(2) < F(z) - AG(z), Concave Programming 7 or # maximizes Fl) —AG(x) without any constraints, ‘This is summed up in the next theorem: Sufficient Conditions for Concave Programming: If # and are such that () # maximizes F(x) — AG(«) without any constyaints, and ii) A> 0, G2) Se with complementary slackness, then z maximizes F(x) subject to G(z) Se. If (FAG) is eoncave (for which in tur it suffices to have F concave ana G couwex), then (7.9) implies (i) above. Note that-no constraint qualifieation appears in tle suffieie conditions; it pertains to the val y of the necessary eotutitions Quasi-Concave Programming In the separation approach of Chapter 6, F wis awiely quack concave arid exch component constraint unetion in @ was quas convex. In this chapter the stronger assnyption of eensaity ane convexity has been made so far. lu fact the weaker astimptiens oF quasi-concavity and -convexity make litle difference tthe wees sty conditions. They yietd swfficiewt conditions ike the on for concave programing, but only in the presence of some hte: technical conditions that are quite comple to establish, ‘Therefore {shall discuss only a linited version of quas-conenve programming, suamely one where the objective funetion is qunsicconcave aud the constraint function js linear. Of course the sirror-ininge case of a Finear objective and a quasi-convex constraint ean be treated ia the same way. Therefore my analysis covers each of the decentsalized pair of decision problems into which the separation approach of Chapter 6 split the general quasi-concave programing proble First we must establish a property of quasi-concave functions tnt is similar to the ‘overestimation by the tangent” property of concave functions, Start with the definition: if Fis quasieoneave, then for any 2 and 2 and for any @ in (0,1), we are to have F((1~6)2 424) > min(( Fie"), F(2")), Suppose F(x) > F(x4). Then F(a" +6(2"—5")) > Fla")98 Optimization in Beonomic Theory Fix 24, 2°, and regard the left-hand side as a function of 8, say (8). Then the inequality is simply A(G) > AQ) for all @>0 (and <1), Therefore h'(0) > 0. But by the ebain rule, W() = F(2* + 8(2" —2*)) (24 — 2") Evaluating this at @ 0, we have Fi(a*) (28 = 24) 30. (70) This holds for all 2%, 2° such that F(x) = F(a), Note that F, is a row veetor, so the expression on the left-hand side is an inet product. Now consider the maximization of F(.r) subject to px 0 the constraint is binding, and this is the only ease 1 shall cousider, (By taking cubic transforms of F, it is possible to points where (7-11) halds with A= 0, bus that is not a ease of sufficient ec tary exposition.) The aim is to prove that if F is continuous and quasi-conesve, the conditions (7.11) are also sufficient, That is, i et spurious stations mic interest for an eleren they are satisfied by # and > 0, then # solves the qunsi-coucave programming problem, To prove this, consider any x such that F(x) > F(z) =. 1 shall prove that 2 is not feasible, that is, pr > b. Start by using (7.10) with 2 =F and 26 = 2. Then F(z) > F(z) implies F(z) (2 — 2) 20. Substitute (7.11) into this and divide by the positive number \ te get plea) >0, or pr > pe In other words, the upper contour set of Fz) for the vale # is contained in the hal of the z-space on oF above the constraint line Concave Programming 99 Fig. 7.4 ~ Quasi-concave objective and linear constraint Figure 7.4 illustrates this, Geometrically, the vector Fy(2) is normal (perpendiculae) to the contour of P(x) at # The veetor is normal to the constraint line p2 = B at any point on it, The usual condition of tangency between the two curves is equivalent to saying that their normal veetors be parallel. That is just what (7.11) expresses, with the constant of proportionality equal to 2 Since F js continuous and F(x) > F(#), in faet 2 is au interior of the upper contour set sefore it is a poi 1a intesion point of the set px > 5. In other words, it satisins par > b. So auy yielding value yields no greater value than F(x). This completes (hn wrenter than F(Z) is infeasible, or any feasible x pron Uniqueness ‘The above sufficient conditions for concave as well as yuasi-concave programing are weak in the sense that they establish that no ‘other feasible choice > ean do better thau #. ‘They do ot rule out the existence of other feasible choices that yield F(.r) = Flz). In other words, they do aot establish the uniqueness of the optimum.106 Optimization in Economic Theory As in Chapter 6, an additional condition, namely a strengthen of the concept of coneavity or quasi-coucavity, gives 1 For example, the definition of striet concavity requires all in terior points of the chord joining any two points on the graph of the function to lie strictly below the graph, ‘This rules out any straight line segments whore the graph and the chord can coincide Formally, call F strictly concave if, for distinct 2*, 2°, and for any B in (0,1) except the end-points 0 and 1, we have F(O2" + (1-82!) > O F(a") + (1B) F(x). (712) If the objective function # in the coneave programing, probless isstrietly concave, then the maximizes 2 is unique, The proof peo ceeds hy nssianing another equally good choice, say %, and showing that. (2 +2) can do even better. T shall eave the simple detail to the reader Examples Beample 7.1: Linear Programming An important special ease of concave progranuning is the theory of linear programming. Here the objective and constraint funetion: ae Tinea Fe) Br an Gtr where a is nn n-dimensional row vector and B an m-by-n matsix Now F(z) a, G(x) =B, Sign constraints x > Qare also imposed. When the coustraint fune tions are Finear, no constraint qualification is needed; interested! readers can see the reason for this fron the formal development of the Kati Tucker theory in the Appeattx, All the conditions of concave progratuning are fulfilled, and the appropriate Kubn-Tucker conditions (3,7) and (3.10) are nee essary as well as sufficient, There is n small new notational point In this problem we will have occasion to consider the Lagrangy sultipliers as variables. ‘Therefore their paxticular valves corre sponding to the optimum of the problem at hand will be indicates! by placing bars over the corresponding symbols, Concone Programming 301 ‘The Lagrangian is Lah artAle~ Br}, (7.13), sand the optimum 7, 4 satisfy the conditions a-ABE0, F>0, with complementary slackness, (7.14) €~BE20, 420, with complementary slackness. (7.18) Between them, (7.14) and (7.15) contain 2" combinations ‘of patterns of equations and inequalities. Not all of these are per tmissible. Generally, if k of the constraints in (7.15) hold with ‘quality, this puts k restrictions on the n-dimensional vector 2. To determine it, we should have (nk) more conditions from (7.14), $0 exactly this number of the non-negativity constysints should bind When this is the ease, the cossespondinig cows! Sey is also cor- rect, Each such (2,1) paie is called “basic solution’. ‘Phe space of parameters («,¢, B) splits into regions, i eacle of which one biesie olution obtains. At the boundaries where taro sich sepions meet, tere are additional non-basie solutions where bots inequities i some of the complementary slackness pairs hold as equstions ‘The transition from one such region to another einses 1 sud den change in the Lagrange multipliers, and therefore kinks in the ‘maximum value fanetion of the kind we saw in Figuye 4.1 Now consider a new linear progeammning problem: fad a sow vector y to minimize yc subject to the constraints y B > a, y > 0, where the vectors a, ¢, and the matrix B, are exactly as before. La ‘our maximization terminology, this ean be written as max (ye) subject to —yBS—ay>0. Except for an interchange of row vectors and column veetors, this is of the sane form as the above problent involving «. Thevefore we ean introduce a column vector je of multipliers, and defixe the Lagrangian My.) = ‘The optimum y and ji are defined by the Kuhn “Tucker conditions yer[-aty Ble (7.16) secossmny anil sutficient e+ BS, 920, with complementary slackwess, (7.17)102 Optimization in Economic Theory —a+9B>0, #20, with complementary slackness. (7.18) Now (7.17) is exactly the same as (7.15), and (7.18) is the same ss (Td), if we replace 7 by A, and ji by 2. In other words, the optiauin # aud } of the origival problem solve the new problem, With their oles interchanged: 2 is the optimal vector of the choiee vasinbles in the now problem, and 2 is the corresponding vector of the multiplies ‘The new problens is anid to be dual to the osiginal, whiel is then ealled che primal problem in the pair. ‘This eaptures au important «conomie relationship between prices and quantities in economies. The pritnal problem his the staudard interpretation Lot 2 be the vector of output quantities, and @ that of prices or unit values of the outputs. The matrix J coutains unit inpwl coeflcients, $0 B zis the vector of input requirements for producing 2 Finally, cis the vector of input sspplies When the optimum 2 is found, the corresponding 3 is: the vector of shiadow prices of the inputs. We just saw that anonx all the woetors 4 that satisfy the constraints AB > a and 4> 0, the X yields the minimum value of Ae. ‘Thus the shadow prices aninimize the cost of the inputs ¢, Note that the jth component of AB is 3 A, Baz, which is just the cost of the bundle of inputs needed to produce one init of ood j,caleulated using the shadow prices of the inputs. Thus the coustzaint is that the vector of sue, input costs is at least ax great ass the vector of the unit valles of outputs, In other words, the sade: prices of inpits ensue that 1 food can make a stsietly positive profit ~ a standard ‘conpetitive condition in economies. ‘Complementary slaekness in (7.14) ensures that if the unit cost of production of good j actually exceeds its price, that i, pro duction would entail a loss when calculated at che proper shadow prices of inputs, then the good will be produced in zero quantity Conversely, if good j is produced in postive quantity, then the unit cost exactly equals the price, that is, the profit is exactly zero, ‘This can be summarized by observing that complementary slackness in (7.14) implies or=\Ba, ‘and that in (7.15) gives Ne= ABs. Concave Programming 103 Combining the two, we have az=de, (Z19) ‘This says that the value of the optinwun output equals the cost of the factor supplies evalunted at the shadow prices, ‘The result ean be interpreted as the farniliar eireular flow of income, that is, iiational product equals national income, Finally, it is easy to check that if we take the dual problem steps of find ite dual, we return to the primal. In other words, duality is ‘reflexive’ ‘This is in essence the duality theory of Hinear programming cexeepl for one point, We took the optinuun # ns our starting-point, paying no attention to the existence of the sohution. This may he problematic, either because the cousteaiuts any be autwally inconsistent, of because they may define an unbounded feasible set and the objective finetion may’ tend te infinity as we pncved ‘out along this unbounded set. T shall leave the treatament of this issue to moze advaaiced texts ‘as our stacting-poiat and go through the nec Example 7.2: Failure of Profit-musineiz Por a scala x, consider the maximization of F(x) = G(x) = 4 <1. Since F is increasing, the optinnn occurs ale = 1 ‘The Kuln-Tucker conditions apply, and give 9 = ¢ But x = 1 docs not maximize F(z) — G(x) witht om straints. In fact ef ~ ez can be macde arbitrarily large by inerens ng x beyond 1. Lagrange’s method does not convert the origin ronstained maxiuization problem into an uaconstesined prolit uaximization problem. The difficulty is that F is not concave tion subject to Exercises Brercise 7.1: Minimization Develop the theory of minimization of a convex function alou lines parallel to those used in this chapter for maximization of a concave function, Brercise 7.2: Convesity of Maximum Value Function Let @ be a vector of parameters, and consider the problem of choos: ng x to maximize F(z,6) subject to Glz) Se. Let V(@) de104 ptimization in Economie Theory tie massimum value as a function of the parameters. Prove that if F is convex as funetion of 8 for each fixed 2, then V is convex, In Chapter 5 we saw geomctrienlly (Figure 5.2) that the min imum cost of producing a given quantity of output, regarded as a function of input prices, is concave. Derive that formally as a corollary of the above general result Beercise 7.3: More on Linear Programming Show that the optimal solution # of the linear-programaning prob: lem of Exam fre such that le 7.1, and the corresponding vector of rmultipliers 3. 12,3) < (#4) $ L2,) for all non-negative x and A. In other words, # maximizes the Lagrangian when 4 = A, and 3 minimizes the Lagrangian when 2 =, nother words, the graph of the Lagrangian in (2, ) space is shaped like « saddle, Therefore (2,4) is said to be a saddle-point of the Lagrangian. Let V(a,c} denote the mmaximnin value function of the linear programming problem, Show that V is convex in a for each fixed! ¢, and concave in ¢ for each fixed Further Reading ‘The best treatment of linear programming continues to be Ropenr DoweMAn, PAUL A. SAMUELSON, and ROBERT M SoLow, Linear Programming and Economie Analy w York MeGravw-Hall, 1958. On 4 the most valuable KENNETH ARROW and ALAIN ENTHOVEN, ‘Quasi-concave Programming’, Beanometrica, 29(4), 1981, pp. 779-800, iasi-coneave programing, the original article remain 8 Second-Order Conditions Local and Global Maxima The previous chapter developed sufficient couitions fov optimal ity, using properties like concavity and quasi-concavity, Those were defined globally, that is, over the full domain of definition of thie functions. For example, a function is called concave if the tangent at any point lies on or above the graph of the funetion, or if thy graph lies on or above the chord joining any two points of it, Cor respondingly, the conditions are sufficient for a global maxis: the F satisfying them does at least as well as any other feasible . Inasense the conditions are ideal; if they are met, we have no further worry that some distant point somewhere may do better tan the one we are looking at. But in many applications the fane tions do not have the desired property over their whole domain of definition, In this chapter the focus and constraint funetions in a small neighborhood of the proposed on the curvature of the objective sed in terms of the second optimum. The conditions are expr order derivatives of the functions at this point. Conditions ot such derivatives are sullicient for local optima; if they lol, the proposed point does better than all points in a sufficiently snuall neighborhood of it This is a useful property when global conditions ate not muck Moreover, it has a valuable by-prndutet, Ia Chapter 4 [introduced the term comparative statics for the gewexal nsctlod ofc of solutions in response to small eli win prassaucters, las f wwe have been concerned only with Ube comparative states of the ‘maximum value v, and not of the optinuuan chuice variables 1, Now we focus on the latter, Jt turns out that the second-order cond tions play an instrumental role in deters statie responses of the optimum choice varinbles. Therefore I shall develop the theory of second-or ning, the comparative der conditions and their applieation to comparative statics in parallel. I shall start with simple cases106 Optimization in Beonomic Theory where the maximization is not subject to auy constraints, and eo on to the moze complex theory of constrained maximization, Unconstrained Maximization First suppose a s alar variable 2 is being chosen to maximize F(z) Let # be n candidate for the optinium choice, aud expand F(x) it 18 Toylor series around 2 F(a) = Fe) + PG) (2-2) + FPM@)(2— 2 (8) ‘The first-order necessary condition for optimality is F'(z) = 0. Using it, we ean write (8.1) as F(z) ~ F(a) = 4 PMa)(e ~ 2F + (82) For x near enough to 2, the quadratic term will dominate the higher-order terms (concealed in the ...) in the Tay Therefore if #"(2) is positive, we will be able to find an 1 near enough to x for which F(z) > F(#). In other words, ¢ will no yield a maxinum of F(2) in a small ucighborhood, OF course it ‘ould not yield a maximum over the whole range of F citer. This argument gives us a second-order necessary condition for # to yield a niaximum, local oF global, of F(z), namely or expansion, BNE) <0. (83) If this derivative is negative, then the quadratic term in (8.2 will be negative, In a small enough interval around 2, we will hav F(2) < F(2), irrespective of the signs of higher order terms. Thus siven that (8.1) holds, PME) <0 (8.4) is second-order sufficient condition for # to give a local maximum of F(z). Note two points of difference between (8.3) and ($4). The former is a weak inequality while the latter is the corresponding stziet inequality. The former is 8 nev ssary condition for local or ut condition for local maxima only. Similar remarks apply to second-orde: conditions in more general contexts. Therefore I shall concentrate on the local suliciency role of second-order conditions, and leave it to the sslobal maxima, while the latter is a suific readers to formulate the corresponding necessary ones. Second-Order Conditions 107 A local maximum satisfying the second-order conditions is called a regular maximum. If the maximum is “rteglar’, Unt is, P"(z) = 0, we have to look at higher-order F(z) = 0 is a necessary condition, and F(z) < 0 is a sulicient condition. I shall leave aside such complications. Now suppose the problem involves a parameter @. ‘The fist= onder condition is crivatives. "Thon Flee) 0. (8.5) This implicitly defines £ a8 a function of @ said we need to know hhow the optisum choice will respond to elanges in @. Dillerentt ating (8.5) totally, we lave F(z, 0)d0 + Fool B)d 8, de /d8 = — Fale], (.0,0) (Ko At a regular maximum, the denominator oo Uv right hanel side bs negative. Then the sign of dé/d@ is the wu as that of Ue et partial derivative Fyg at the optimum, ‘This shovss ota lany the second-order condition helps 1s in assessing. the: qirabittive elferks of parameter changes on the optizan chic Asa simple economic illustration, consider a preit sa:rsiiit firm whose demand curve, and heace the rovenne ervey shill Let R(z,8) be the revenue as a function of output sane i sfufl yar eter 8. Arrange matters so that Ity is always positive: sn inva in @ shifts the demand and the revenue curves npicansl, \ cles Intion similar to that above will show tt at inctrine ins ill inerease the profit-maximizing output # if Ryg ix positive. tat is, if the increase in @ shifts the marginal revemie upwatel. New 13s perfectly possible that as @ increases, the average yevensie (Ihe te mand curve) shifts up but the marginal reverse shifts slows: what is needed is a twist that reduces the elasticity of dean, ‘Then a favorable shift of demand will eause output to fall, ‘This is whol tundeslies many fond paradoxes aud trick questions in vleumentary Let us turn to the case of maximization with a vector of choice variables, but still no constraints, Now the Taylor expmnsin ie F(a) = F(z) + Fe(3)(2-2)4+§ (2—2)" Feel@)(r—r) +. (87)108 Optimizotion in Economic Theory exe Fre isthe symmetsic square matrix ofthe second-order par tial derivatives Fya = 3°F/02,O24, the superscript T denotes th transpose operation to change the eokinn veetor into a ron vector and the second-order terms are the quadratic form (2-2)? Feel@)(2-2) = OY Fal@)(2)-#)) (2-7) (8.8) Before turning to the maximization problem, it is useful to note how (8.7) gives a new characterization of concavity. If F is concave, it les on or below its tangent, oF its value is less than or ‘equal to that given by the Taylor expansion up to first order. For this to be true when 2 is sufficiently close to 2 where the second order terms dominate the difference, we must have (= 2)" Fae(3)(2- 3) <0 forall = ‘A quacdeatic form y" My, where M is a symmetric matrix, is called negative definite if its value is negative for all y # 0. This corsesponds to the following condition on M principal minor of order & of M, that is, the submatsix formed by the elements common to any k rows and the k columms with the same numbers as the chosen rows. The determinant of such « principal minor should have the sign of (—1)§, negative iff is od and positive if kis even. Morcover, this should hold for any choice of the k rows and corresponding columns, and for every & from 1 to n, the dimension of the matrix. In fact it suffices to check this for the leading principal minors, namely those formed by taking the fSirst k rows and columns: Consider » ‘The form is called negative seme definite if its value is non-positive for all y; the conditions for this are the corresponding weak inequalities on the principal minors. With these definitions, we say that F is concave at 2 if Fre(2) is, the matrix of a negative semi definite quadratic form. Now consider the maximization problem. The first-order nec essary condition is F,(#) = 0, and then the second-order sufficient condition is that the quadratic form on the right-hand side of (8.7) is negative forall x # 7, that is, negative definite, The correspond ing necessary condition is that it be negative sexi-definite. The objective function need not be concave in the general sense defined Second-Order Conditions 109 in Chapter 6, but it must be concave at the point # in the sense just defined, All this once again helps in doing comparative staties. Sup. pose a vector of parameters @ enters the definition of F. ‘The first-order condition is F,(2,8) = 0. Differentiating, we have Fes(2,0)d2 + Fea( 7,0) d0 = 0, which looks exactly like the corresponding equation for the one variable case, exeept that now dr and d? are vectors, suid Fe and Fg are matrices. The solution for di is as =Fiel 2,8) Foal 2.8) 08, (89) The inverse of a negative definite matsix is also negative definite, ‘and the information about the sign of its minors ean Le combined with the information about Fe in specific problems tw find the signs of the changes in some choice variables as some parameters change. The results with many variables and parameters are thus not as simple or general as those before. But I shall offer oxe appliention in Example 54 Constrained Optimization [shall begin with the simplest case of two choice variables and one equation constraint, and then state hiow its results extend to more sseneral situations. Consider the maximization of F(.r1,c2) subject to Gl21,22) = ¢, where both F and @ are increasing functions Figures 2.1, 6.1, and 6.3 consideved various aspeets of this. The erueial point is the relative curvature of te lowo contours through the optinuns 2; the coatour of F should be of their ‘arguments snore convex than that of G. To express this nlgebrnically, we have aul find the second derivative of this function, Begin with the contour of F(e1,72) ‘The expression for the first-order derivative was obtained in conjunction with the tangency argument of Figure (28) is to think of 22 as a function of zy slong each eontout Eqvation drz/dr, = Fi(01.22)/ Falta, ra)110 Optimization in Economic Theory Now we must differentiate this again, remembering that x on th right-hand side is a function of 21. Therefore AFP) er drt dey 2 (Fis + Fiz dra/dzy) ~ Fi (Pas + Fan dvo/des) Fe 2 Fi 27 Fi Fo + Fi? Fo Pe The arguments (21,72) of all the derivatives of F are suppressed for brevity A similar expression can be derived for the second derivative slong the constraint curve, ‘The second-order sufficient conditiou for £ to be a local optimum is that dzz/diy? along the F coutows should be greater than that along the G contour, Using the first order necessary conditions Py) Gz) for j= 1,2 remembering that we are assuming the Fy and G, to be positive ‘and simplifying, the second-order condition becomes, Gy? (Fy — AGa1) 2G: Ga Fig 4 Gia) + G1? (Far — Gan) <0. (8.10) ‘This is mow newtly expressed in matrix notation: evaluated abe [Pu-AGu Fe-AGe —Gy det | Fay-AGay Fin AGn -Gz]>0. (811) L-@ -G. 0 The conditions for the general problem with n choice variables and m V8"), Expanding the two expeessions on the loft suid the sight around @ in Taylor series, we have V(2(0),8) + Vol 2(8),8)(6" ~ 8) + 5 Vool Z(@),8) (0! — 8)? + SVB) + Val) (6" — 0) + 5 Vae( 0) (0 ~ 0)? +na Optimization in Economic Theory ‘The first-order envelope property allows us to cancel the first two terms on the left-hand side with those on the sight-handl side. ‘This leaves (Vou 2(8),8) — ‘Taking 6" sufficiently close to 8, the quadratic tenn doniinates the 131 of the expansion (concealed in ...) on the left the inequality to hold in this situation, we mast a (8)) (OP +... 50 rand side. For Vaal Z(6),) < Voo(@), (813) is proves that the long-run maximum value function ‘8 convex as the short-ras one at the point where the ¢wo are tangent. For suitably ‘regular’ maxima, we have a strict inequality like (8,13). T shall not pursue this refinement, but will talee wp aa: Jmportant application of (8.13) in Example 8.2 it least Examples Example 8.1: Consumer Theory In Example 5.2 the consumer's expenditure function E(p, 1) was dofined as the minimum outlay requited to attain the utility level 1 at prices p. The compensated demand function C(p, u) wns the minimization problem, vector of quantities that solved this cos The envelope property implied that Clo.) = Eyton), the veetor of price-derivatives of the expenditure function. Tn Example 6.2 we showed that the expenditure fianetion was concave. Now we know the characterization of concavity in terns of second-order derivatives, and ean use this to obtain useful prop ‘erties of compensated demand functions Differentiating the above relationship, we have Cylpast) = Byplpse) (s.4) Since the matrix of second-order derivatives on the right-hand side is symmetric, we have the symmetry of substitution effects of price changes 80? fp = OC*/8p, = Eye Second-Order Conditions 15 Next, sinee E is concave, the matrix on the vight-hand side is negative semi-definite. In particular, its diagonal entries, being, L-by-1 minors, must be <0 herefore OC} Jp; <0. for all j (815) In other words, the own substitution effects of price changes are non-positive, The same result follows even more simply from the very con: cept of a maximum. Manipulation of the ‘revealed prefe ‘equalities showing that the optimum does at least as well as any’ ‘other feasible choice leads ns to the desired result. Suppose p* anid ph are two price vectors, and 2*, 2" the corresponding compensated omands, Both attain the same utility level u. Since 2* gives the smaller expenditure when prices are p" and 2* does the sane for vf, we have pz? 0, OexfOpe <0 for all j,k Brercise 8.2: More on Derived Demand Consider a competitive firm like that of Example 8.3, but withow the assumption of constant returns to scale, Suppose its total eos! fanction is C(w,y), so the marginal eost is Cy(w,). Assame tht the marginal cost curve is rising, or Cyy > 8. ‘The firm takes thr output price p and the input price vector w as given, and maxi izes profit. Find the set of equations that determin its inp demand vector x. Examine the role of te cross-partial deriva PC/Oydwx in determining the sign of Or,/Iiee. Intexpret_ you result Brereise 8.9: Minimization Develop second-order conditions for unconstrained and constrain! minimization problems by analogy with the snaximization proly lems of the text. You will need to define positive definite au! semi-definite quadratic forms, and the signs of the principal 1 nors of their matrices Further Reading For a mote detailed treatment of consumer and produ see Varian, Microeconomic Analysis, op. cit., chs. 1, 3, or JAMES M. Henberson and RICHARD E. QUAND, Micro economic Theory: A Mathematical Approach, 3rd edn., New Yor! MoGraw-Hill, 1080, chs. 2-5. +: theory Second-Order Conditions 11 ‘The use of second-order conditions in comparative statics was Irisieered by Sasnuclson; his classic treatise i still essential reading. PAUL A, SAMUELSON, Foundations of Economic Anclysis, Cambridge, MA: Harvard University Press, 1947. Further reflections on these issues, including a discussion of the LeChatelier- Samuelson Principle, is in his Nobel Prize lecture, PAUL A. SAMUELSON, ‘Maximum principles in analytical cco: hnoinies’, American Beonomic Review, 62(3), 1972, pp. 249-62. For an extension of this principle, with applications, see EUGENE SILBENRENG, “The LeChatelier Principle as a covol- Inzy to a generalized envelope theorem’, Journal of Economic The ory, 3(2), 1971, pp. 146-56. A useful reference that will belp you avoid some pitfalls in the use of optitnization theory is KNUT SYDSETER, ‘Tetter to the Editor on some frequently occurring errors in the eeouoinie literature pertaining to problems of maxinin and minizna’, Journal of Economic Theory, 9(4), 1974 pp. 464-66,9 Uncertainty Expected Utility 1 formal sense, the theory of optimization under uncertain does not roquite any new mathematical theory as suc "Phe cit variables lead to random outcomes with objective or subjeetiv probability distributions, The objective functions are apprepriats probability weighted aver aio subjoct to some constraints. The general theory develope in the fist eight chapters continues to apply. In fact the special slructute of problems of choice under wicertaiuty lends to speci results that were not availble fw the general matlienaties of cou ses (expected values). The choices a strained optinnization, Let us make these ideas a little more precise, Suppose thir after the decision at hand has been taken, the world could evolu in any of a number of different states of the world, or element ys. These ate called differs ry events, Suppose at first that the states ave discrete and finite in number, indexed by i = 1,2... 1 Write p, for their probabilities, objective or subjective as may ls proriate for the application being considered. ‘These are 1x niegilive and add to one, The economically relevant outeomes in the alternative situations will typ wealth, or profit accruing to the decision-maker. Denote the by Yi. Most of the time I shall take the ¥; to be scalars, bu in gencral they could be vectors, Then we can write a generm cally be the levels of incom objective function FY Yay Yona Pays Pm) The ehoiee o control variables x will affect some or all of the ¥ and the pj; the 2 variables may also be subject to some addition coustraints, ‘The maximization problem ean then be solved 1y familiar methods, Uncertainty 123 Under certain restrictions on the preferences, the objective fanetion ean be expressed in a special form, namely the mathe- suatical expectation (probability weighted average) of the values of utility finetion U af the outcomes in the different states: Som The function U is called the von Neumann-Morgeastera utility finetion (of income, wealth, or profit as the ens whole expression (9.1) is called the expected utility (02) ny be), and the “This formulation is very useful in its simplicity and its ability to capture some economically interesting aspects of belavior such ns sisk-aversion, Therefore it is used almost universally in appli cations. Al the work I diseuss is based on it. Since any focns is ‘ou the techniques of optimizstion, [ shall not state or disenss the conditions under which expected ritility sunximszation is valid, Int sefer the reader to the books by Arrow and Kreps cited at the of the chapter. Recent research has begun to nd clop mote general by Machina cited at the yet tractable alternatives; sce the sur ead of the daapter In most situations, oxe would nok exp + the decision-maker to be indifferent to the risk involved in getting oue of a whale range ‘of values of ¥. Suppose there are just two states, with distinct outcomes ¥; and Ya, and positive probabilities p and (1 — p) re spectively. Compare this with an actuarially oqaivalent alternative, where the mathematical expectation of ¥, naunely [p¥i-H(1~p) Yal, is reecived with certainty. A decision-maker who is risk- refer the sure sain. TI UipYs £1 —D)¥s) > PV) 4A VIVO). This just says that U is (strictly) concave iu the range [¥.¥3} More generally, strict concavity iimplies o(S>x) Sno (92) when the probabilities are positive nel the outcomes distinct. If U is twice differentiable, 0" < 0 correspon' to tisk-nversiast124 Optimization in Economic Theory Once again, the decision variables x affect some or all of thy outcomes and probabilities, setting up the basis for an optimization problem. A couple of quick illustrations will develop the economie intuition for this, First suppose ¥j < Yo, s0 the first state entails some loss at least relative to the second. A natural response would be nit of # ots you X if'state 1 occurs. If this insurance industry is perfectly competitive, and each firs can pool a large number of independen risks, thea insurance will be actuatially fair, Then pX = 2, 0 X = 2/p, Therefore Y; changes to (Yi ~ x + 2/p), and Yp t (V4 ~ 1); remeunber that the premium is paid in advance, that is in both states. The value of the objective funetion becomes purchase insurance, Suppose an advance premium payin PU ~2+2/p) + (1 plU% ==), ‘The first-order contition for x to be optimum is fond using thy chain rule UY, = 2+ 2/p)(Ifp- 1) —P)U¥2 — 2) = 0, UY ~ 2 + 2/p) =U ~ 2). IF" <0, this is also sufficient, and itmplies Yj — 2-4 2/p = Thus @ risk-averse person will buy actuarially fair insurance to th point where the outcomes in different states are equal. He will sure, or hedge, fully Next suppose the probability of the bad outcome 1 ean In reduced by incurring an expense 7 in advance. In specific context this might mean using a more relinble but more expensive product or exercising more eare in the risky activity when the act of bein) careful causes disutility. Now we make p a function of 2; this will be decreasing, and since p is bounded below, it will generally | xx. The objective function can be written as p)U(% — 2) +1 Ul ~ 2) Then (2) = = p'(2) [0% — 2) — UH — 2) — (02) U' = 4 s]U' = 2)}. (O28 Uncertainty 125 The first term on the right-hand side is the expected marginal ben: fit, of care or quality, being the product of the margival reduction. in the probability of the bad outcome, and the difference in util constitute the ities between the two outcomes, ‘The other tern tmarginal cost of care or quality. The optimum = is defined by the fisst-order condition Finally, suppose both insurance and care variables are avail ble. The insurance company cannot tell whether care was exer cised, it ean only observe the outcome. If actuarially fair insurance is available, the objective function is Visor tb a/ph by the same reasoning as before. Let the common value of these he Ya. Then, by analogy with (9.3), we have bela, = 2)- 0% 2-2 + 2/02) = {(2) U's =~ 2 + #/p2)) + IL 2)]U' ~ 2-2} U'(¥o) <0 when x iv chosen optimally. In words, the marginal benefit from cate vanishes when there is full insurance, while the marginal cost stays positive, Therefore the optimum of eare oveurs at the corer 20. Inother words, the availability of full inmsrinice destioys the incentive to exercise costly care. Thais is known as ‘nsoral lard in the insurance jargon. In practice, only partial insuesiee will he nnilable when moral hnzand is preset This is an example of the general problem of ayymanetsic in- formation: one side in nn economic transactions hiv: knowledge of sonne relevant variable like risk, effort, or quality that the other side lacks. Then the transaction cannot take place int a classical com. petitive market at arm's length; a coutsact or an incentive scheme Jias to be designed to get around the information asymmetry as for tus possible, A vast new atea of economic theory of such situations126 Optimization in Economic Theory has opened up in the Inst decade or so, T diseuss two simple prob lems of this Wind in Examples 9.1 and 2, and offer some refereue' for farther reading at tine end of the chapter. he rest of this chap ations of choice under sncertainty, particularly portfalio choie deals with more conventional situ ‘This is somewhat mnore conveniently treated in terms of contin number of states of the world. Thus we eplace the index by a random variable r defined over the range [r, 7], the probabilities p, by a density funtion f(0), and the expected utility expression (9.1) fous random vatiables rather than a finite EU) U(r) fr) (oa ‘The choice variable + will then become an aditional argument in the functious ¥ and f. Theinterpretation of risk-nversion paralle: (0.2), The mathematical expectation of Y is ayl= fl Yesiver Then U" < 0 implies UY) > EVO: (051 {his isa application of Jensen's Inequality. One Safe and One Risky Asset An investor has initial wealth IV, Investing 2 in the risky yields the total (principal plus interest) of 2(1 +r), where » i fa random varinble with ¢l pays zero interest; this ean be generalized but it makes the notatios little messier. Now the final (random) wealth is density funetion f(r). "The sage asset W = (Wy —2)+c(l +r) =Wo tar (06 The amount + must be in the range (0,1); ‘short sales’, aw borrowing at the riskless rate to invest in the risky assct, a not allowed. ‘The investor hias a sti Morgenstern atility function U, and chooses 2 to maximize ly concave von Neuman EIUGW)] UQWy ber) fled (or Uncertainty wm Let (2) denote this expression regarded as a function of 2. ‘Then sto [ev oHerenfende=0 Ii pnetienlar, eco = Vine fO rSle)ae= Ue) EU) Note that U"(IW9) is non-random, and therefore can be takes out- side the integration (sum (probabilistic average) of the rate of interest ou thie risky asset i positive, then 4'(0) is positive, and the optimum x carmel be zeve In other words, the risk-averse investor will buy at least some of fn actuarially good investment If > 0, then 4!(2) > 0 for all x, and itis optimal to put all of Wy in the risky asset. More typically, the investor will hold some ofeach asset, The first-order condition with respect to 2 is ition). If the mathematical expectation [vor rensoyte =. (98) If there is an -< Wo satisfying this, then the concavity of U sarantces that if is the global optima The next obvions step is to study the comparative staties of this choice. First suppose Wo changes. Recognize Wy as a para. nneter in 4, So the maximand is (2, Wo). Write its partial deriva tive with respect to 2 as dey and that with respect to Tp a8 de ‘Then the first-order condition is dy(2. Wa) = 0. Diftventinting this \otally as in Chapter 8. we find dz /dWy = ~bewls Wall bealr.Wo. cor is negative by the second-order condition, so the The desom sign of the comparative static derivative is the sane as the sign of the numerator. Now UMW + ar) fr}ae (99) 2,Wa)198 Optimization in Economic Theory Since an interior ¢ ean be optimal only if r < 0.< F, we cannot fis the sign of (9.9) without further work depend on the property of a measure of risk-aversion, Re that U"" <0 means risk-aversion; a useful quantitative measure « this turns out to be he answer turns ont AQW) = -U"UN YIU) (oan) To interpret this, we consider decreasing the variance of the di tribution of WV slightly leave the investor indifferent. This marginal rate of substitution turns out to be } ACW). A higher A(W) means the investor i willing to give up more of the mean retum to get a given sul decrease in vatiance. Therefore A(W) is called the investor's ah solute risk-aversion, We would expect a wealthier investor to In more tolerant of a given marginal tisk, that is A(IV) should be decreasing function, ‘Then the result is that if the nbsolute risk-aversion ACI) i 1 decreasing function of wealth, then dew is positive: that is, wealthier investor will hold more of the risky asset. and asking what decrease in its mean will To prove this, note that for r <0, we have =U"(Wy + en) /U"(Wo + 2r) > -U"(Wo)/U" 9) = A(Wo) Multiplying by —r, 7 U"(Wo + 2r)/U"(Wy + 2r) > ~ ACW) Pr U"(Wy+2r) > —A(Wo)rU'(Wo ter). (9.11) Forr > 0 U"(Wo + 27)/0" Wo + ar) < —U"(WG)/U" Wa) = AW). Multiplying by r, =r "(Wo + 27)/U"(Wo + 27) —A( Wo) rU(Wo + 20) once again, Uncertainty 129 We have shoven that the inequality (9.11) holds for all r, pos- itive or negative. Integrating it [rere ear sinner > ale) [Ort +20) Ferre ‘The right-hand side is zero by (2.8). This proves the result ‘The next natural questions are the effects on 2 of shifts in the density function f, particularly an increase in risk, and in the utiliyy function U, patticular! these topies would take far too much space, and I must leave them fan increase in risk-aversion, But to proper microeconomics texts and research articles cited at the tnd of the chapter. Portfolio Choice Here we allow any number of risky assets, The treatment of tis in financial economics usually assuines that the investor's objective can be expressed as a function of the mean M and the standard deviation $ of wealth. This is quite a restrictive assiamption. In the expected utility framework, it corresponds to one of two eve: cial cases: (1) If the von Neumann-Morgenstern utility function's quadratic, uw W-law?, where a > 0 is constant, then BIUCW)) = M — baa’ 5S?) (2) Ifench asset has a normally distributed return, ten wealth has anormal distribution, and the expectation of any vou Neumann: Morgenstern utility finnetion can be expressed int terns of its sean fund variance, A specific function of interest in this context is UE) = expat), with @ > 0, for which BIUCW)| = —exp(—a [Mt — 4S?) ‘Then expected utility is maximized when (M — ba S?) is. Using (9.10), we can sce that for this function the absolute risk-aversion is constant and equal to «130 Optimization in Beonomic Theory In wither of the two cases where mean-variance analysis is sy plicable, the indifference curves of expected utility can be shows in (M.S) space. By constructing the Gasible frontier between and $, we can study the portfolio-choiee problem dingeanaticall The initial wealth will be held fixed throughout this exere so normalize it to unity, Suppose there age n assets. Write tlw {otal returns (principal plus interest) 8a vector r= (P72... ta ‘These are random variables. Let their mathe form a vector = (1, H2 eal expectativs jon)s and let the (symmetrie, positis definite) variance-covariance matrix of the gross returns be ¥: (2,)). The portfolio is a veetor of proportions of wealth invests! in the various assets, 2 = (2, ¢2 ... 29) ‘The randoms Sal wealth is We ‘The mean and the vatianee of final wealth are respectively Yo m= 2" M ne s=yy eer Note that both M and $? nze functions of x To find the feasible trans we want to minimize the standard deviation for a given mean, th oemation frontier between M and minimize $= (27 Sx)", subject to 27 = M, 27e=1, ’ where ¢ is a vector of ones The sninimized $ is a convex fiction of M. It anay have « decreasing portion, but for large Mf it is increasing, tus present ing a trade-off between return and tisk. The case of ews asset suffices to Dring ont the main points, therefore I restric attention tech tion from previous chapters; I shal] outline i i to it here. But the general case illustrates the use of some niques of optiniz Pxample 9.3, Breertainty in With two assets, let x stand for x1, s0 22 =1~ 2, Then M = 12 + (1 ~ p2)25 (9.12) aad F-2Ke rt (ey + Ba) (9.3) where I have defined Ky so} -poron Ky =a} poo, mid p is the corrclation coefficient between the random returns 1n and ra, that is, the covariance divided by the product of thei standavd deviations. Order the assets so that jay > ja. Then the equations (9.12 nnd (9.13) define the frontier iu (A, 8) space in torus of a para meter. As x increases from 0 to 1, J nnd $ changes from 22 to 0}. Aloug, the way. invrvayes front pty to ge S$ dSfdr = Ke + (Wy + Ka) (9.4) So 0, dS/dz = (02 ~ por). and at 1, dS/dz =~ po. If ay > op, so there is a risk-return trade-off between the two completely specialized portfolios, then d8/dz is sure to be positive wat ¢ = 1, that is, there is bound to be a trade-off over some range of high retwrns. Even if o1 <2, that is, asset 1 dominates asset 2, there may be a diversification benefit from mixing thew ad therefore a trade-off if p is sufficiently less More generally, the mininnan-varianee pottfolio i K; af — porns > Rt Re ~ (P= omoa) + (o} — ev) ‘The conditions for this to ie in [0,1], thnt is, uot involve any short tt from the equation wales, are evi Differentiating (9.14) again, $ #Sfas? + (aS/de)? = C132 Optimization in Beonomic Theory (503) efficent frontier indifference —_—. Ms Fig. 9.1 Portfolio choice in mean-variance fraumework Simplifying, we find sin d?S/ds* = sgn (1 p*) of of > 0. ‘Therefore the frontier in (M.S) space is convex: Figure 9.1 sums up the results so far. Two extn sare worth pointing ont. If p = 1, the frontier is simply the straight linc joining (11501) to (jz,@2)- If p= 1, then the minimum variance is 2er0, and the frontier consists of two straight lives joining this riskless portfolio point to each of (14),01) an (2 Next suppose there is one riskless asset with a Sure gross ceturn bo, and risky assets are as before. Ifa fraetion 29 of wealth is held in the risklese asset, then M = topo +27 p= 26 p49 + (1 ~ 0) €" py where the vector £ gives the proportions of the various assets i the risky part of the portfolio, Likewise 2? Sa =(1— su)" TEE, Uncertainty 133 S= (1-40) (EQ). ‘Therefore the fonsible points are found by joining {¢.0) to cach of the points ou the fro produces the efficient frontier ABP, if short sales are not allowes fand the straight-line frontier ABC if they are. In this ease the risky posets arc only ever held in the proportions represented by the point B. Everyone holds a inix of the suite asset and the portfolio B, the nix. depending on the attitude to risk. This 1 becomes a mutual fund, ier previously obtained. In Figure 9:2, this proportions of th s 2 - i uw Portfolio clioice wit a rises asset Examples Beample 9.1: Managerial Incentives Bliciting the right amount of effort frou subordinates is a problem for capitalist owners and socialist plannev example. An owner has to hire a mange the project succeeds, it will produce vale V ke. Hove is a simple to rim a project. TE "The probability1st Optimization in Beonomic Theory of suc depends on the quality of the manager's work. Gives high quality, the project will suceeed with probability », but low quality will reduce this to 9. The basie salary needed to attract manager is w. But he has to exert himself more to achieve high quality, and will do so only if he is paid a premium. Bath the owner and the manager are risk-neutral, that is, each maximize the mathematical expectation of his monetary returns (minus th equivalent cost of effort in the case of the manager) First suppose the owner ean observe the quality of the man ager’s work, and compensate the effort dizeetly. His expected prolit from eliciting high-quality work is pV — (w+), while low-quality operation would get him qV —w. The interesting ease to consider is one where (a) high quality yields more profit than low quality that is, pV-(wte)>qV—w, of (p—g}V>e, (9.15) ‘and ()) the paofit with high quality is positive, that is, pV >wte (9.16) Now suppose the owner cannot observe the quality of effort If the owner offers the premium © to a mannger in return for 8° unverifiable promise to provide high-quality work, the manger ¢n cheat and deliver low quality instead. ‘This lowers tke probability of success of the project, but so Tong ax 1 > p> q > 0, the owner cannot infer the quality of the manager's effort by observing single success or failure, and therefore has no recourse against the enc manager's breach of c« Therefore the owner must base his payment scheme on the only thing he ean observe, namely suceess or failure. Suppose he pays the manager 2 if the project succeeds, and y if it fails. Gives this scheme, the manager will choose high-quality effort if this gives him greater benefit net of the cost of his efort, that is, if prt (1-p)y-e>ar4(l—ay. If the two sides are equal, the n wager is indifferent between high and low-quality effort. It is usual to suppose that so long as i! makes no difference to him, he acts like a nice gny and breaks the Uncertainty 135 tic favorably to the owner, that is, chooses high quality. ‘Therefore the inequality is weak rather than strict. It simplifies to (p-glz-a) 26 (9.17) Secondly, the manager will agree to work for the owner if pet+(1—p) ytp(r—u)2wte (9.18) ‘These then are the constraints that give the manager the right incentives, ‘The owner's expected profit is x=pV-[pe+(—p)y]=p¥—y-ple—y)- (9.19) He wants to maximize this subject to the two incentive constraints above. It is obvious that he should make y and (2 — y) as small as possible. Then the constraints will hold with equality, and r-y=elp-gh y=wte-eri(p—ah yew-egl(p—g) te wte(l~aip—a). (9.20) These have a simple interpretation: the manage ward for success or unm a 's compensation consists of the basic salary plus a penalty for failure With these values, the owner's expected profit is the same as when he could observe the manager's effort direetly the inability to observe effort has made no difference But there is a difficulty. Nothing gnarantees y > 0. ‘Thus the optimum sehenie may involve a fine that the wi owner if the project fails. In practice such a fine is ver, ager pays to the difficult to extract. If fines are ruled out, the problem must be solved with fn additional constraint y > 0. The solution is to go as far as136 Optimizetion in Beonomic Theory possible, namely set y= 0. Then x must satisfy the remaining t ‘sanstraints, so Fe(wteyp wud 22 elip—a) But when the y in the unconstrained solution is negative, w +-€ less than ep/(p~ 4), and the latter is the binding constraint. Se 2 =ef(p—4), and the owner's expected profit is PV —epi(p~ 4) By the conditions (9.25-16) assumed at the outset this sles thw pV ~(w+e), but still positive. ‘Thus the nee to pay output-bases incentives ineans a smaller expected profit, but tot by so mse to inake the whole enterprise unprofitable ‘This example is just the beginning of a large body of rece research on the sig of incentive contracts, The problein consid cred here is similar to that of moral hazard in insurnace, where the company could not observe the rsk-redicing care on te patt of Ui polieyhokler, We designed the best contznct for a single profes. Us fn ongoing relationship of this kind, the average of the anager suceesses overtime provides statistical information about his effort enabling a better contract that generates higher expected profit Similarly, if the planuer hos several similar managers who fac some common risk, then observation of thir relative peeforiat cau provide information about their relative efforts. Finally, i Ae example there was a liquidity constraint ~ the manager could not be fied — but no risk-aversion. Allowing the manager to bx sk-averse brings aldtional issues ofthe efficent alloention of rish ‘between the planner and the manager. Some references to the bu sgeoning literatane on such matters are provided atthe end of the chapter The next example concerns a diferent probles of information ‘Theplanner may not know the ianate quality of the manager. This too, has « porallel in insurance, where the policyholder bis bette infortuation about his own risk: class than does the company. This gives tise to ‘adverse selection’; an insurance policy is especialy tzactive to those who know their ovat risk to be figher than the duds implicit iu the terms of the policy, and who are therefor suudesirable estomers from the point of view of the company, It Uncertainty 137 wants to design a contract that will handle this problem. For variety, I shall construct the example in a different context Example 9.2: Cost-Plus Contracts Governinent expenditures for defense, health, and yes, even educa tion are often made on a cost-plus basis, ‘That is, the goverment reimburses the supplier’s eost plus a normal profit. ‘The govern tnent's puechasing agency usually does not know the true cost of production of these goods or services. If a supplier who can pro- hace the good at low cost pretends te have a higher cost, he gets a higher reimbursement, which he ean enjoy by paying high salaries to himself and other workers, spending lavishly on offices and such, facilities ete. If the government wishes to avoid the excessive pay ment, it must devise & scheme that climinates the temptation to Inisstate cost, Here is a very simple example ‘Suppose the truc average cost of produetion ean take just one of two values, ¢1 and ¢2, with ey < eg. Bach figure already includes horuial profit. The problem is that the higher figure may be real for pretense; the government cannot tell the differenes. In other words, the supplier can be either of two types, low cost or hi cost, and the purchasing ageney cannot tell whether it faces a irewsinely high-cost supplier, o a type-cy who is pretencling to be a type-ez and plimaing to enjoy the extra "The government enn decide Low many «nits it will buy, how much it will pay, plier. Suppose it announces that if the supplier claims to hes average cost ¢,, for i= 1 or 2, it will buy qi units, and pay a total san of Ry for them. What are the constraints on these choices? First, of the two cost figures, if that happens to be genuine, the supplier should be willing to participate, That is, his east sbuntd be covered Reg frie 12 (oan) Secondly, it should be optimal for the supplier to reveal his true cast, ‘These are called incentive compatibility euusteaiats, If the supplier's true average cost is ey, he should not wish to pretend it fs cy, and the other way round. Econonie intuition suggests that only the first of these will be binding (a tvuly high-cost firm will hot preterid to be low-cost), but a proper theotetieal treatment should prove that, and not assume it at the outset. Now if a firmhis Optimization in Beonoméc Theory with cost ¢1 pretends to have cost es, it sells q2 units instead of gy andl line revenue Ry instead of Ry, but its actual cost in its profit ealeulation stays at ¢. Therefore the constraint for a firm with fost cy to report it truthfully is Ry cig 2 Reco (9.22) Similaely for a Berm with true ost 2, the constenint is Roan? Ron (923) Suppose the government gets hen (g) from quantity 9 where Bis an increasing and strictly concave funtion. Suppose its estimate of the probability of the true east being ¢; 3 6, anc that of c: is @2 = 1-61. Then its expected benefit met of the payments to the supplier is, 4 [Blgx) ~ Ry] + 6 Bla2) ~ Re) (928) ‘The optimum scheme will maximize this subject to the partic pation and incentive compatibility constraints (9-21-23). For the snoment I shall ignore non-negativity constraints on the q, and Ry Form d Lagrangian 4 (Blas) ~ Ru] + 8 | Bar) ~ Ra] + a [Ri ~ e141) + a2 [Re ~ e2 a2) Ar [Ry ~ Ro — egy + er ga) + ala = Ry = en gn ben ash Most of the economically interesting results can be found with- out solving the whole problem, First we add the incentive com: patibility constraints together and simplify to write (eee) (an ~ 42) 20, (9.25) so if the supplier declares lower cost, he will sell at least as mh, and typically more. ‘This is a part of the incentive for the low cost-ty f = 0 and q > 0. ‘That will effectively climinate the incen- tive to pretend 10 have high cost, But this carries a risk: if the pe fir to respond truthfully. Indeed the solution may have supplier turns out to have genuinely high cost, the government lhaviaig committed itself to the seheme, will be unable to purchase Uncertainty 139 from him even though stich a traasaction might be er post desir fable, Therefore such a solution will arise only if either this risk is small, oF or its consequences ate sufficiently stall, that is, if 6 B'(0) is finite and simall. I shall leave the details to the reader, and rssume henceforth that q2 and gy are both postive. Now consider the participation constraints (0.21). The aita is to prove that both easmnot be slack. Begin by noting that if type 2 firm makes positive profil, so docs a type-1 firm, To see this, suppose Ry — ca 42 > 0, then from type T's incentive compatibility ‘constraint (9.22) we have Ry ern > (c2—c1)a2 > 0. Next note the first-order conditions for Ry and Ry + dy debi = Ay + Ae te =O. If firins of both types maké positive profits, both participation implies sn = constraints are slack, and complementary slackness O= pig, Then = — Ae = Or but both 8 and @ are positive, so this is Therefore the optinusm selvere must lave Roa Rag >o. "The positive pure profit is anathier part of the ineeut firm has for revealing its low cot truthfully. Its average revenne RiJan excecals its average cost cr, butt hy type 2' incentive cour patibility constraint, it eamot excerl e Of course the goveruient will mule My as small as possible ). while meeting type I's incentive conpmtibility eomstenint (9 Substituting for Rp in it, we have Risen (a —o)ae (927) With this, it is easy to verify that. (9.28) is antommatically satisfied, and it is a strict inequality so long as q: > q2 which is true al the optimum, ‘Thus we have proved the intuitive suggestion that theM0 Optimization in Feonomic Theory truly high-cost firm will low-cost firm will be Now the govern want to deflate its cost, but the truly the verge of wanting to inflate its cost, cat's objective function ean be rewritten as 45 { B(qy) — en 9: ~ (ex es) g2] + 82 Blas) ~ e202} The first-order conditions with respect to gy and gy are Bla) = (9.28) and BY ga) = €2 + (81/82) (co — e1) These have a role in the incentive scheme, too. OF the two. (9.28) is straightforward. If the supplier declares himself to be low-cost, the government buys from him a quantity that makes the marginal benefit equal to the marginal (equals average) cost But (9.29) is more complicated. If the supplier declares himself te De high-cost, that in itself would be a reason for buying less; tlhe equality of marginal benefit and cost would ituply B'(g2) = e2 and therefore gz < 41. In fact the sight-band side of (9.29) is greater than cso the government buys even less from a high-cost supplier ‘This again makes it less tempting for the low-cost supplicr to inflate cost, The policy dies ancan giving up a net marginal social gain when the supplier is really of type 2, but it is desirable to incur some such cost to achieve the offsetting gain on the incentive sid. Another way of expressing the idea is to note that the possibility Of the cost being higher offers a temptation for the low-cost fin to pretend a high eost; this is like an external diseconomy eased by a high-cost firm. The right-hand side of (9,29) adds the shadow cost of this externality to cz to get the true social cost of a type-2 firm. ‘Then the amount g2 bought from it equates the sarginel Denefit with this true social cost, (9.29) Bsample 9.9: The Mean-Variance Prontier Here I examine the convexity of the transformation frontier in port folio selection with n assets. ‘This fis out the technical details omitted from the text, and incidentally provides a nice illustra tion of the use of convexity and maxinnm-value fanctions. The argument proceeds Ws to intermediate results or lemmas. Lemma t: (2) (27 B.r)4 is a convex function of x, Uncertainty Mi aT De, then differentiating, and ‘Therefore O60 =E- 5 We show that the matrix der is positive semi-definite. For any vector 2, 2a ePErsTE2/¢ Be- (PEs /@TEs), OFT ber = ‘This expression is now-negative by the Cauchy-Schwartz inequal- ity Lemma 2: If ¢{) is convex, and min{ #2) [wT =MyeTs=1), then S is a convex function of M Proof: Let My, My be any two values of M, and let 23, 2? the respective minimizers. ‘Then for any @€ [0,1), wT (2) 4 (1 6)27) = OM + (1-8) A, P02! 41-824) So the average portfolio is feasible. Therefore S(O Mi, + (1~8) Ma) < 402" + (1 — 8") < 04x!) +1 — AAC") (Mh) + (1 8) SUB). This proves the desired convexity. Now I indicate how to obtain the explicit solution for the frontier S(M). It is easier to2 Optimization im Beonomie Theory mize $*. Let a, 8 denote the multipi a 4, 9 denote the multipliers on the respective coustraints. The first-order condition ig i ute, Solving for x and substituting in the conatrai 2 (WD) +B (ED Wy WIE e) +8 (CEe) ‘These two equations ean be solved for a and 8, which in turn gives cxpeeions for Hh pte potas» nadie aseteat Se terns of M. \ "s Exercises: Beeteite 9.1: Taxation of Risky Income Swoprne that in the mode of one safe nnd one tsky asset in the text, we intoduce taxation of intent conve on the sky eee (with deduction allowed for lose) nt the rater whine rn ‘Thus the final random wealth is : W=Wo + (-r)er, Show that the first-order condition for an interior optiniun x is [ e0ns0-nen ins Deduce that + change, the optiznn the optinin ¢ changes heping (1) constant. Thensor ifthe tn ate on esky toon netoases o does the amount of wealth held in the risky asset. Si iggest econo intuition for this seemingly pasadena! reat Exercise 9.2: Saving with Uncertainty A consumer lives for two periods. I periods. Income in period 1 is sure and equal to Yi, The income ¥p in period 2 ean he random, It saves § from his period-1 income, he gets total return (pri plus interest) of r$ in period 2, where + ean be randoin, His Uncertainty M3 ‘objective is to: maximize the expected present value of the «tities of consumption in the two pesind: BOY - 5) 45 BLU +r 9), where U’ > O and U" <0. Write down the first- and second-order conditions. Show that as Yj increases, $ also increases but at a smaller rate, that is, the marginal propeusity to save lies between O and 1 Next suppose that ¥ is sure but r is random, and examine the effect of an increase in Y>. Final random, and examine the effect of an increase in r, Further Reading A basic exposition of the economies of uncertainty is in Varian, Microcconomic Analysis, ch. 3, sects. 18-2). For a fuller treat ‘ment, including a discussion of the validity of expected-utility max imization, see Davio M. KWEPs, Notes on the Theory of Choiee, Boulder, CO: Westview Press, 1988, ‘The classic, and still highly recommended treatment of ex ected utility theory aud its applications is KeNnerit J. ARNOW, Bxsays in the Theory of Risk. Bearing, Amsterdam: North-Holland, 1971, chs, 1-3. A review of recent work on Criterion fonctions snore general than expeeted utility is MARK MACIINA, ‘Choice under uncertainty; Problenas solved and unsolved’, Journal of Economie Perspeetioze, I(1}, 1987, pp. 11-84, The general theory of the comp: ir effects of cexeases in tisk and in risk-avetvion ean be: Said iy Micuart Roruscimy and Josrrn St risk: (1) A defintion, and (2) Economie Consequences’, Journal of Bconomie Theory, 2, 1970, pp. 2% 43 and 3, 1971, py 66 84, and suppose 7 is sure bat ¥, ive tn 12, “Increasing evens nn risk Peren DIAMOND and Josten Sticitt7Z, risk aversion’, Journal of Beononie: Pheory, 8, WT4, pp. 337-60. The theory of portfolio selection is set forth iu every book on financial economies, At an clementary level, the interested seader ean consultMt Optimization in Economie Theary Ruchanp A. BREALEY and STEWART MvEnS, Principles of Corporate Finance, 2nd edn., New York: McGraw-Hill, 1984, A tore advanced treatment is in Cil-ru HUANG and RoERT H. LrtzeNMERGER, Foundations of Financial Beonomics, Amsterdam: North-Holland, 1988, Varian (ibid., ch. 8) gives a brief treatment of asynsmetric information. For more detailed expositions see : avi Krps, A Course in Micraeconomic Theory, op. cit chs. 16-18, 7 Porn on ot 10 Time: The Maximum Principle Statement of the Problem As in the case of uncertainty, the study of optimization over time requires no new general principles, The variables to be chosen will pertain to different dates, but we can always stack them together jn one large vector z, and the general problem remains one of max- imizing the value of a function F(x) subject to a veetor inequality constraint G(x) ver But in fact this constraint will always bind along an optiznal path, so [shall use the simpler equation form, In addition to the constraints that govern changes in stocks. tere may be constraints on dh ibe vasiables pertaining to any one date, such as Glyn ent) £0, 0.2) where G is yeetor funetion, An example would be a constrain! that requires consuimption not to exceed grass output. Constraints for stocks and Rows to be nose (10.2), Another special feature that often occurs in optimization ov. time is tlt the criterion funetion is additively separable: it ean be expressed as the stm of functions, each of which depends on te varinbles pertaining to only one date tive enn also be included inn Y Fluent 03} For example, a firm maximizing the discounted present value of its stroum of profits would naturally have such an objective, and tin Time: The Maximum Principle ur would enter the fusetion explicitly in the forin of discount factors: (1+ r)-t, whore r is the rate of interest, For a comsusner's choice lover time (the saving decision), it is often convenient to assume that the utility function is additively separable over time, ‘This is a restriction on preferences. Roughly, it requires that ‘the marginal sate of substitution between Inch and dinner is independent of the faniouut of breakfast”; aucat example due, [ believe, to Henry Wan, As additive separability over time is commonly analytical tractability Just one more detail remains to be specified. ‘The ti al stocks or state variables yp must be the result of some unsperified history: wwe simply take them as given. Similarly, when the optimization ends at a Gnite T, we must specify some terminal condition, and the simplest is a requirement of a fixed veetor ype, of stocks to be bequeathed to the future. ith the expected utility formulation in the ease of wucertainty, ¥ assumed because of its span of the optimization problem begins at t= 0. ‘The init ‘The Maximum Principle Now the choice variables are the yp for # = 1,2,... 7 aval t for # = 0,1,...T. These axe subject to the constraints (10.1) and (10.2), each holding for # = 0,1,...7. The objective function is sven by (1033). We ean define shadow prices and forin the Lagrangiaat as us Let Ay denote the reltipliees for the cemtraints (10.2); these lave: the usual interpretation of shadow prices of the constraguts on ae tivities at f, The shadow prices of (10:1) are ew, sud snore in. teresting. ‘They tell as the amount of the first-order inevesse in the objective fanction if the constraint on the relaxed, that is, if we are given a gift of a siall addition to the stock yea. Therefore they ate the shalow prices of the storks at (t+ 1), and I shall denote them by t+ White £ for the Lagrangian of the full intertemporal problem, Then weve i storks is £20 (Fone trees [set Qluize8 mon] = Glve zt) }- 104)Optimization in Beonomic Theory ous to list on the left-hand side. In this chapter I shal use the first-order necessary conditions established in Chapter 3 without ever explicitly comparing the optima with other feasible choiees. Therefore I shall not need bars to distinguish specific points from general ones, and shall make the notation simpler by ‘omitting then, [shall also assume that the appropriate constraint qualification is alseays met, The first-order conditions with respect to 2: for t= 0,1,...7 are simple’ ajax Fuyistet) + ran Quy ze) 1 Gales 205t (05) ‘Those with respect to yr ate made more complicated beease exc! 1 appears in two terms of the sum, For example, a appears it the finetions F, Q, and @ and as 2; in the term ¢ = 1, amd also 98 —m} yy in the term # = 0. We ean rearrange the expression so that each yp appears in only one term. portion of (10.4): fake only the relevaat YE (ye — yen) = (vo —mn)-+ ma (a — Ye) + oot Has (Ue — Yrs) = wom ti (em) beet ur (ra Fe) — Ure ra = ve (me — 1) + 0m — Ure Read «a0.6) Then (10.4) becomes LD? { Flan sut) + m1 Aun eet) Falmer =m) NGUn 208) } + F(yo, 20,0) + ™ Ql va, 20,0) + Yo 1 — Yrgs Fry: (10.7) ‘The terms left hanging in the last line pertain to yo aud yrs which ace not choice variables. The first-order conditions on yy fr hhe arguments of Care all the yr, #4, Ay and mr4,; they are too Time: The Mazimum Principle 49 T ase BE 21 = Fylyestet) + Ress Qyle s08) + mei — te Me Gul t) rege r= —[Fulvetet) +8141 Qylue eet) = MGy(u 20 t)]: (10.8) ‘These conditions can be weitten in a snore compact and eco: nomically illuminating way. Define a new function JZ, called the Hamiltonian, by (yz Rt) = Pluses) + Qu 2 (10.9) ‘Then (10.5) says that the controls 2, at t should be chosen to maximize H(ye,2,te41sths subject to the constraint Gye, 21.t) < 0. Write H*(yeytevav8) for the resulting maximum value Define the Lagrangian of this single-pesiod optimization prob. Jean E (not to be confused with the £ for the full problem over all periods) as 1 = Myst teet) ~MGvn 20) (10.10) ‘Then (10.8) is more simply written as rept = Lgl verze tet) In the static maximization problem, only te 2; are ehwiee var lables and the yr and ess are parameters, Therefore the Envelope Theorem applies, and we have mua — A= —Hglgus mines! (0.11) ‘The Envelope Theorem also gives Hz = Ly = Q, evaluated ft the optimum. ‘Therefore we can verite (10.1) in» form that is symmetric to (10.9) gern ue = Helvetia (10.12) ‘The results can be summed up as follows.150 Optio fiom in Economic Theory The Maximum Principle: The fust-order necessary conditions for the maximization of (10.3) subject to (10.1) and (10.2) are (for each t, 2¢ maximizes the Hamiltonian H (ans s1ste410t| subject to the single-period constraints G(y,,. 7+.) < 0, and {ii) the changes in yy and xy over time ate governed by the Alifference equations (10.11) sau (10.12), ‘This principle proves useful in solving such problems applications. But its grentest conceptual merit isin the interpretation of the maximization condition (j). It is clear thet we would not want to choose =; to maximize F(yp,24,¢)¢ we kos Uhat the choice oft alfectsye41 via (10.1), and therefore affects tle terms in the objective fanction at times (41 ete. In the prodhetion interpretation, for example, a big splurge of consumption tod’ would increase uilty today, but would mean a lower capital stock for the future, and therefore less consumption ard less utility the future. How can we eaptute all these future effects in a siniph way? By using the shadow price of the afected stock, of cour ‘The effet of zon yoga equals its effect on Qlyn set), an tls resulting change in the objective faction is found by mutiplyin this by the shadow price me41 of yagi. That is just what we add to F to get the Hamiltonian, ‘Thus the Hamiltonian offers a simpl seny of altcring the enc-petiod objective funetion Flys srt) to take fo acconnt the future consequences of the eliiee of the contrat att The condition (10.8), or equivalently (10.11) also has a useful ‘conoinie interpretation, A marginal unit of yy yields the marginal setutn Fy(t)—2y Gy(t) within the period ¢, paying proper attention to the shadow cost of the single-petiod constraint, ancl an extra Qs) the next period valued at rays. (Note that T have ted thy angument ¢ instead of the full (y, 21,0) for brevity.) These ean be thought of as a dividend. The change in price ta, ~ vb ike a capital gain, except that the prices are in present-value terme 50 "41 contains an extra discount factor that captures the usual interest or opportunity cost of enrrying yy for one period. When 38 optimum, the overall marginal return, or the sim of these component, shouldbe aro, That ix jst what (10.8) express whew written as (Fo(8) — A Gy(t)] + (0) O10) + [reg m=O. (10.13) Time: The Mazimum Principle 151 other words, the shadow prices take on values that do not permit ‘a pure or excess return from holding the stock; this is an intertem: poral no-arbitrage condition, ‘What if the terminal condition o stocks gry, had not been imposed? As these stocks contribute nothing to the objective fune- tion, the optimal policy should keep them as low as possible, ast ally zero. But in some cases it may be desireble to build up stocks first to provide output aud utility, and then, depreciation limits how fast they can be run down before the terminal date. If any ast be worthless, in other words, we positive stocks are left, they should have Yres 20, Trai 20, with complementary slackness. More generally, if there ism eonstenint yryy 2 Be we get tri, Frei 20, with complementary slackness. (10.14) Such a condition on terminal stocks and their shadow prices is often called a transversality condition Continuous-Time Model Up to now Ihave treated tiie as passing in a diserete succession of periods. This permits the development of the theory as a special case of the standard Lagrange-Kcun~Tucker theory, and easy eco nomic interpretation of the conditions. But when solving acta specific examples, it turns ont to he much nore convenient to think of fime as a continuous variable. There is no real theoretieal reason for preferring the one or the other. For mnesonie convenience, T shall write discrete time as a subscript, and continuous tine as function argument ia parentheses We ean think of continions tine ms the Fist elon wer take diserete periods of length At, and let this slirink to zevo. This requires some modifications inv (10-1 33). Flows are now rates per unit time, 50 the right-hand side of the sloek flow relation (10.1) must be multiplied by the length At of the period, The equation becomes Wet AL) = ult) = Q(uld,=10.0 AL Dividing by At and letting this go to zew gives the tine-derivative of the stock on the left-hand side. Te is emuventional to indicate this by a dot placed over the variable, Tins we have HE) = Qlu(t), =(0).8) (20.15)152 Optimization in Economic Theory Only @ notation in (10.2) to get I change from subseripts to arguments is needed Glult}. 208), € 6 (10.16) ‘The suan in (10.3) is more conmplicated. The total spa of tinw from 0 to T is split into T/A¢ little discrete periods. Indexing these periods by i, the sum can be written as rat So Fiat) id) de ‘The limit of this sum, when At goes to zero, is the integeal fF Fond cioa7y Aut incidental advantage is that Te4.a, converges to tr, and thus storks at time # de not go awkwardly together with shadow prices at {4-1 in the Hamiltonian. Defining H as in (10.9), 2¢ max izes H(y(t),2(0), m(t)-4)) subject to Glu(t),2(t},t) <0. Writing LH for the maximum value function, y(#) and x(t) satisfy the pait of differential equations HE) = Hult), 1.1), (10.18) and Ht ~Hylalt).x(6)0) (10.19) We could formally derive these conditions by first defining the Lagrangian £ of the full problem by analogy with (10.4) caf { Fuit.200.9 + x18) [QU = NH} Gult) 2(0).4) ft (20.20) 0.4) ute)]) ‘The analog of the rearrangement in (10.6) is integration by parts sep atoat= [a(n a{ete+ 0x10) — ME a(T). (021) Time: The Mazimum Principle 153 ‘Thea cof {Fu = NAG(U(t) 2(4),4) } A+ #(0) (0) ~ n(T) y(T).(20.22) A083 + xLE) Oly), 2(0).0) + CE) ACE Now we can think of the integcal just like n sum, and differentiate ‘with respect to 2(t) and y(t) © get the first-order conditions Frlult), (tht) +) QWlt)s 210),1) = MUG.y(t)2(),8) = (10.23) and Fylvlt}, 2000) + HO Qs(olt) 200,04 #0) A(#) Gy(y(t)e-(0)st) = 0. (10.24) (20.28) is the condition for 2(#) to maximize the Hamiltonian, and (20.24) parallels (10.13), the intertemporad atbitrage equation Bilt) NE)G Ct) 4 HL) Qylt) +H) = 0. (20.25) Of course inatters are not really that simple, Integrals enamot be differentiated in the ordinary sense with respect to variable at one instant of time, aud a sigorous theory of optimization with coutinuous tine is very complicated. But short-ruts: fike the one above do lead to usable results, This will saflice for ines» those demanding more cast purse ty vefere of the chapter, Practical applications of the Masinun Prineiple procevd by deriving the differential equations (10.18. 19}, and solving them subject to the appropriate conditions al f= and 7’ If time does not enter the Hamiltonian explicitly, the solutions ean be shown igeomctrically i the (y,) spares stich 1 pictorial representation is| called a phase diagram. Its use is best explained by illustrating it in the context of a specific problem of economie interest. ¥ shall develop such an applieation in Example 102.rey Optimization in Beonomie Theory Examples Example 10.1: Life-Cycle Saving Consider a worker with a known span of life T, over whiel be will carn wages at a constant rate 1, and receive interest at a constant rate ron accumulated savings, of pay the same rate ot auccunmated debts. ‘Thus when his stock of accunulated assets (debt if negative) equals k, his flow income is (w + rk). Writing + for his consumption flow, capital accumulation is governed by wtrk—e Note that k and ¢ are functions of f. The general point of evaluation js taken as understood; only special values are shown explicitly when needed. In technical language, f is the state variable, and ¢ the control variable, Suppose there ate no inheritanees or bequests, so the ent point conditions are KO) KT) =0. (10.26) Suppose there are no other constraints on choice. ‘The instanta neous utility function is In(e), and there is a utility discount rate ps0 the maximand is ale) ee To use the Maximum Principle, define the Hamiltonian Hao) eH +x (w rk 0). (0.27) ‘The condition for ¢ to maximize 1 is cher (10.28) Substituting in (10.27), the maximized Hamiltonian heeomes Ht = ~(In(r) + pt) "+m (wt rk) ‘The differential equations for k and = are = 0H" /Or =wtrk-* : (10.29) Time: The Mazimum Principle 158 and o re (2030) OH" /Ok = ‘The general solution of (10.30) is obvious: wet, (oat) where ro is a constant to be determined. Substituting this in (10.29), we have swtrk= mm Now ake yfdt = (kr ke" ‘which integrates to eo = MO) = w [Le fr = wot [1 eo. Since we know &(T), this equation fixes 9, and completes the solution ‘Some economically important facts can be found without know. ing the complete solution. Using (10.31) in (10.28) we ean write tro “This shows tha the worker's optimum consumption grows aver i Metineifr > p. Sice consumption nnd wages rast al eee He whole iene inthe sense of having aqua dncrited prs Calves, the implies 1 Jie later yeare Im ote words. He commer saves only en Phils pases ad athe las ars of is Hn vg Me opponite happens if rp. Some titutinal constants ine prevent bi om having aasrigg af fir ey and of soure the wok economy could nok Harem with all connec lvenpting to disave, But few ae separate Nes * eperely the simplest example of life-cycle saving, The tncory can be geneaied oinclade worecoupheated preferences, Tabor supply sd etneent choles taxation, uncertainty, Hid iy comtraints, and many move features of reality156 Optimization in Economie Theory Esamnple 10.2: Optionum Growth Thisis alo a problem of optimal saving, Pine economy as a whole. The change of perspective brings with it two new features. First, the rate of setumn to saving rant be ag Fosenous market rate of interest as it would be for an indivices) {it must be the endogenous marginal procct of capital, Secondly, W7e is no logical terminal date to the plan, I shall develop tke theory by formally letting T = co, a theory 1% and mention the attendant, but from the point of view (The usunl assumptions ae that Fis increasing snd ste proportional rate 6. If the consumption flow is c, tl hen the ri accumulation equation js cneiie bFU)~ bee (0039) Tie initia capital stock 6(0) i give straints, Ct ‘The utility of the flow of constanption ts 0% set ty of he sotion is O(c), increasing m There are no other con- ¥ discos ate is p, 80 the an ff 0. at An ol obvious potential dificult is i emerge ofthis eg That “hee ‘ds a sufficiently targe I shall leave out the details, el To apply the Maxinuin Prnepes define the Hea H=U(6) e+ [F(R ~ 84 e] The condition for ¢ to maxiniine H iy Ui) ee" (10.33) ‘The differential equation sntisfied by x is OH/O = —n( Fh) ~ 8 (v.34) We could solve (10.33) for cand substitute the r which would then ot dine join (10.84) in giving ys a pair of differential Time: The Mazimum Principle 1st equations for & and x. Actually it is more convenient to work, terms of xe" = 9 say, because the pair of differential eyuatinns for k and 6 docs not involve tine explicitly. Here T shall take a Afferent approuch, and work with k and ¢ Differentiation of (10.33) gives # = (UMC) é- pU"(c)] e-". Using (28.8) and simplifying, we find &_ PW) 048) 103s e ne) ‘ d where n(e) is the elasticity with whief the margial atility of eon. suinption deelines as consumption increases: le) =e UNV") Observe that Example 10.1 had a formally identicad structuse, with F'U&} constant and equal to r, 6 = 0, and n(c) constant aad eal tol Now we ean use (10.32) and (10.35) as the pair of differen tial equations in & and e. ‘Time does not enter explicitly, and we can show the solutions in a diagram; see Figure 10.1. Given say point (kc), we ean Sad the velocities (f,é) from the differential equations. ‘These can be shown by a small veetor arrow attache to (hye). If we do this for all points, we can join successive 3 rows together ane! find whole paths of motion it the (he) ys: Given an initial point, the differential equations detente Uwe sub. sequent change to proood alosy thesnltive ptt posing hye 2his point, Two such paths eainmot cress, beesnte the direction of motion is uniquely deternsined by te equations set starting: point ‘The easiest way to mnuderstand the plies elingtonn is to 2000 nize that cach of k and ¢ caw increas ar deere Uns die pte fs split into four regions racks corre north-east, south-east ete, Fry ¢ < F(R) 6k, whi is the region below the exe This curve las its peak when FUb)— 8 is a 1 defined by FY(E) = 6. Taming to (10.85), 6 see that in creases when F'(k) > p46; note thet n(e} > @ since UT is increasing, unin, to snoventent toward the (10.39), we se Cin & increases if ok158 Optimization in Beonomie Theory and strictly concave. But F is also increasing and strictly concave} therefore e increases when k < RY, where P'(k*) = p +6. Father, when p > 0, which we need for convergence, we have &* < kf Putting together all this information, we get the pattern of paths shown in the figure, Writing e* = F(A*) — 6 R*, we sce that there are exactly two fri that converge to (K".*), one ftom the lft atthe other from the right, Ail other paths diverge, and are asymptotic to, ot ‘even hit, one of the axes, . venience We are given k(0) but not o(0), so we must try out alternative possiilites for o(0) and see where they lead. If we choox e(0) such that the path starting at (K(0),0(0)) eonverges to (K*e"), all is well. If we choose any other o(0), the path diverges to one of Time: ‘The Mazin Principle 159 the axes, Along the /-axis the consumption goes to zero, Along the c-axis consumption grows for a while, but capital rans out and eventually capital, output, and therefore consumption must fall to zero. Neither possibility Jooks aetractive. This suggests that the right choice of o(0) is on the stable path directly above (0), Indeed, an appeal to sufficient conditions, which I shall discuss briefly in Chapter 11, shows that such a choice is indeed the right A general feature of the solution is apparent: c is higher when is higher. But the figure does not tell us whether ¢/ F(k) inereases with £, and in fact this ean go either Way for particular forms of F and U. Thus we cannot say in general that richer societies should ‘optimaity save a larger proportion of their income, Exercises, Broscive 10.4: Life-Cycle Saving Solve the problem of Example 10.1 with the instantaneous utility function changed to Uh=P-9, «>0. ‘Thus the marginal utility is U'(e) =e The earlier example is the special case where € = 1, as ean be verified by taking linits casing L'Hapital’s Rule, ‘Next suppose the consumer inherits assets ky and plans to leave a bequest of ky, How large ean ky be before the problem has no feasible solution? Beereive £0.2: Optimum Growth Interpret the variable 4 defined in Example 10.2, Show Maat k and @ satisfy the pair of differential eynations B= Fk) ak G0), and dy S[FU)— pa), where G is the function inverse to U". Draw the phase diagram, which should look just like Figure 10.1 but reflected upside down. Complete the solution160 Optimization in Economie Theory Exercise 10,8: Butry-Deterrence ‘The demand curve in nn an industry at time f is given by alt) =a —bplt), where a and b are positive constauts, 9(t) and g(t) are respectively the price and the quantity. There és one large firm that sets the price, and a fringe of sinal firms that accept this price and sel their entire outpit, New fringe firms enter if the lange firm charges 3 price greater than p*. Write 2(¢) for the output of the fringe firms. ‘The initial 2(0) is given; 2(#) satisfies the differential equation Ht) = KLM) ‘The large firm's sales are g()—2(#), and its average cost is constant ‘and equial to ¢, Therefore the discounted present valve of its profits f [plt) —e}[a— a(t) — bplt)] enPa, pis the rate of interest. Assume p* > ¢ Apply the Maxinmn Principle to this problem, taking + as the stnte ynrinble and p as the control varinble, Coustruct the phase diagram in (2,p) space, Finel te qualitative features of the optimom pricing poliey of the large fim, Obtain eonditions on the parameters of the problem ruler which the competing firms retain positive sales in the limit as t goes to oo. Further Reading A more rigorous treatment of the Maximum Principle and its ap- plications is in Intriligator (op.cit.), chs. 14 and 16. See also his dis. 11 and 12 for related ideas and techniques. For relatively simple expositions of the theory of opti S NOMERT M. SOLOW, Growth Theory: An Exposition, Oxford versity Press, 1970, ch. 5, and AVINASH. Dixtr, The Theory of Equilibrium Growth, Ox: ford University Press, 1976, eli. 5, 7 For details of the entey deterze Danius W. Gaskins, ‘Dynamic limit pricing: Optimal prie ing under threst ofentry’, Journal of Economie Theory, 3(3), 1971, -e problein (Exercise 10.3), Time: The Mazimum Principle 61 pp. 806-22, and the comment on it by NORMAN TRELAND, ibid, '5(2), 1972, pp. 303-5. . ‘Although time is the natural independent variable in applica- tions of the Maximum Principle, the mathematies is independent of this interpretation, Indeed, problems of the optimal layout of toads in cities have been handled using the same method, and thete {is interpreted as distance. An example is . TAVINASHE K.. DIxt?, ‘The optimum factory town’, Bell Journal of Economics and Management Science, Auturan 1973, pp. 637 511 Dynamic Programming ‘The Bellman Equation In Chapter 10 we studied optimization over time. Th (io) expreing increment sock or stat variably, and the additively separable form (10.3) of the objective funetion, Were spe- cial features that enabled is to express the first-order corslitions in “tefl apc fo snl Masia Pipl, Pye Ione emperaly ae when tne and wernt sent sether, a8 they ao often do in reality ‘The vectors of initial stocks yo and terminal stocks wry were given when we maximized " YE Pleseth (20.3) subject to the constraints mest ~ ve = Olde 208s (0.2) and Oye zet) <0, (20.2) for t= 0,1... T, Keep the terminal stock require nent fixed for sximuin value as a function of the initial stocks, say V(up)- ne stocks, say V(yp). The vector of deriva vey Va) wl bo the vector ofthe show pies of thee ii now. As in Chapter 5, we ean define the sesulting 1 The separability ofthe objective and the constraints allows us to generalize this greatly. lnstend of the statting tie 0, consider another particular time, say ¢ = 1. For the decisions starting at 1 the only thing that matters about the past is the vector tr of stocks that emerges fiom the past decisions, We can take that Dynarnic Programming 163 as parninetric, and start the whole problem afresh at r- In other words, we maximize a sum just like (10.3), but extending only from r to T, subject to constraints just Tike (10.1) and (10.2), buat hokling only for r, r-+1,.-- 7. Let Viye,7) be the maxisnam value function of this problema; the explicit argument 7 is necessary pecause the limit of the summation depends on it. The veetor of derivatives Vy(y-,7) is the marginal inczease in the inaximized stim ‘when we start with a small increment to the initial stocks yr at wthat is, the veetor of shuadow prices of the initial stocks for the Optimization problem that starts at 7. What happens when we embed the sub-problem starting at + into the full problem starting at 0? In Chapter 10 we could Jaterpret the Lagrange multiplier on the constsaint (10.1) for r as the shadow price vector #41 of stocks at (7-41). A slight relaxation Of this constraint aneant an exogenous increase in veya, and the inultiplier told us the resultant inerease in the objective function (10.3), At fist sight this differs from the veetor of derivatives Voluseast + 1) for the sub-problem starting at (7 + 1). In the fall problem, we know at time 0 that the stocks at (r + 1) are doing to increase a little. Then we can plan ahead and change the control variables at earlier dates. For example, if we know at time 0 that a windfall of wealth is going to occur at (7 +1}, ‘will constume more in anticipation as well as after the realization, But the Envelope Theorem comes to the rescue. For the sinall changes that are involved when we look at first-order derivatives, the direct effect on the objective function is all that counts; the induced effect of optinial readjustments in the choice vi fore we can indeed identify Ue derivatives Vy jabs ean bbe ignored. ‘The with the shadow prices w at all ties ‘Now pick any t, and consider the decision abo tye control variables 21 at tnt time, Consider the conequeners of aay partic ldlar chojee of =. It will lead to next period's stocks mryy aceneling, to (10.2). Thereafter it renains to solve the subproblent starting Bt (¢4 sand achieve the maxinnus value V(gigast + 1). Then the total value starting with ye at Cean be broken down into two terme: Flye tet) that acerues at once, and V(yrgist + 1) that vemrars thereafter, The eboice of ¢1 should maximize the sum of these two terms. In other words, Viaust) = max { Fluse + Vuent +f Ob)164 Optimization in Economic Theory ‘subject to the constraints (10.1) and (10.2) for just this one f ‘This eauation gives usa brute force way of solving the original Wamization problem. ‘The idea isto start at the end and precect fact times recursively, At time thece is no future only the fixed terminal stock requirement 9,4, . ‘Therefore Viv. T) = 1x Fle, zr, T) subject to UF QMYertT)= yess and — Slyry2e.T) <0. its Js in principle a strnightforwend static optimization problem, iia Seles the maximum value fonction V(ys-T). ‘That Len hag be used on the right-hand! side in (11.1) fort = T 1. Tyee {hother static problem, and yields the maximum value fini [Gren TV) And so on ail the way back to 0. In serge {his works only for the simplest problems. Analytical sash ce Pitts Kind ae possible only when the functions P, G, wat @ REE vey simple: forms. Numerical slutions ean be eonmated for somewhat hander problems, but if the state variables hen Wector of more than two di:ensions, even that quickly becomes Irmarateable. Luckily, the brute force method is only a backstop, {p.many economic applications, thete are better methode te heel the solution, or at feast obtain tiseful insights about i, ‘This method of optimization overtime asa succession of static Programming problems was pioneered by Richaed Bellinen, aml bamed Dynamic programming, The iden thne whatever the de- 22m Bt & the subsequent decisions should proceed optimally fog the suboroblen tating a (1) elma Bae a Supetimalty. The maximum value fiction V(yy,t) i called oe Delian value function, and eqnation (17.1) the Belin equation det us look at the maximization problem on the right haecl Gas Of he Bellman equation, Substituting for ves, from (10.1) We are to choose = to maximize Pye 268) 4+ Ven + Qe ses t)yt 41) subject to Cnet) <0. Dynamic: Programming 165 Letting Ar denote the row vector of the multipliers on the con: straints, the first-order conditions are Fil tnt) + Vo(diaa t+ 1) Qclse cat) Ae Guest) = 0. Recognizing the derivatives Vj a8 the shadow prices r, this becomes et) + tres Qelde zest) — Gkyes 206) = 6. Fl joe ate exactly the fit-oner conditions fry to maine the Munntenian H(t tet) dened in Chapter 10, sbyet tte Sipieperiod cosas 02) nether. Thos Danan og mingled to these ele fr ttting the coi viable athe Masi Prince / : tn act the Maximum Prinsple and Dynamic Progrsnning ave fly ejuvaten!allemative seta optimization ove fe You should wee whichever simpler fr tackling the pastedny prablm at hand. The Maxinnin Principles generally beter wen Che's continous an here ancestiny, Dynatic Poot tring inthe oppose ete of crete tne and netting, But thats nota bard an fast rule Inter in ths chopter I shall llutrote the us of Dynamic Programming in sme cconomie appctions, ‘To conde this tetlon Im i to tabla the ineremporl arbitrage equation (G0.18) inn diferent wag. When x is hosen pty, (01.1) es with equality tbat Vue t) = Flues te.t) 4+ V(veriyt + D). i peas on Differentiate this with respect to yy, noting that yeqs depend ye and using the Envelope Theorem on the right Jsun) sik, ‘Thete Voltet) = Fyuiyeet) + Volueynt #1) LF Oat sist)) = Gyn 2et) Using the shadow prices x, this becomes (10.13), Uncertainty Dynamic Programming is particularly well suited to optimization problems that combine tine and uncertainty. Suppese that the Process governing the evolution of stocks ye through time has a16 Optimization in Beonomie Theory random component. Given the stocks yy at the beginning of pe- riod t, and the controls =, dusing the period, we know only the probability density function of next period's stocks yy4s. Write this as olyrysiyinze)- ‘The arguments are separated out for no tational clarity. The first axgsunent yrs1 is the actual vector of random variables whose probability density Canction this is; the others are like parameters that can alter the fanetional form of the distribution. Asa sinple example, yess could be a vector normal distribution with a mean vector j2arud a vasinnee-eoxasiance ma- trix B both of which depend on (y1,21). AS an even more special case, je could equal ye-+ QCahy 8), the value of yey in the previous, discussion without uncertainty: Now the problein is to maxinize the mathematical expectation of (10.8), subject to (10.2) for allt, and (10.1) seplaced by the stochastic law of motion for yes1 described by the function d. Write V(yit) for the maximum value function of the subproblem starting at t. For the moment fix the choice of z. Consider what happens after the actual value of y.41 becomes known at the beginning of iod (14-1). The rest of the decisions will be made optimally, and ld V(yessst + 1). From oe perspective of period ¢, this is till fe random vatinble, and we are concemed abut its mathematical expectation, zivvent+nl= f Veveeit 1) luresitesse) dies, (11.2) where the integral is taken over the ran over which yegr is dis- Vevnt) = max { Fn zat) + EMinent +1}, (18) ‘The maximization on the tght-haa side of (11.8) is somewhat snore dificult than the corresponding certainty ease (11-1). The frstonder condition with respect to =, reinesdiferentintion of with respect to =r inside the integral, and the results of that ean be hard to characterize and understand. Dut in principle (11.3) allows us to staet at and solve the problem recursively backsned 00 just as before. In simple but seal models, the solution ean the completed analytially. At the end ofthe chapter, I develop two examples of Dynamic Progranuning wmder uncertainty applied to Dynamic Programming 167 ‘economic problems, and give references where readers ean prtrse the topie further. Continuous Time ‘e Maxinmim Principle could be formulated for diserete or com eae eee ante Doping Res iti Cha ter 10 the problem was formulated as the maximization of [ F(y(t),2(t),t) dt, (10.17) subject to the law of motion Hl) = QV. AOD. (v8) and the instantancons constraint Glut) 2(8).1) $0 (10.16) efine V(y(t);t) as the maximum value function of the sub-problem Dae VOCs at sal rel fe, upto the Control variables take the value 2(t). Then the contsibution to the Chjective function over this small interval will be F(y(t),2(#)s#) a ‘The stocks at (t+ a0) will be incremented by ult + at) — #(2) = Qlul@).2(8), 4) at, and optimal policies from then on will yield V(y(t + 0.1 4 db). Bellman’s Principle of Optimality gives Vive max { Flylt)o=(nt)de-V(y(eratyt bel) Js 114) 7 subject to (10.15) and (10.16). Expand the right Issud side in a Taylor series: (y(t + dt), t + at) = V(ult)at) + Volt st) tule + dd) — yl] + ValuCDot) dt = V (ult) £)-+ Vol (t)ot) Qalte (Ot) at + Awl 1) at Substituting in (11.4), we sce that V(y(t),#) eaneels from the two, Cides, and then the equation ean be divided through by dt. (A168 Optimization in Beonomic Theory more rigorous argument would use finite increments O¢ and then take limits.) This gives De ame { PlulO,2(.8) + YM D plu). 19.9 +Velvlthds 1s) subject to the instantaneous constraint (10.16). Since Vy{y(t)st) is the vector of shadow prices x(¢), the max ana is js he Halton a) (0 (0) of Chapter 10 and the result is the maximized Maniltonian H*(y(t), en (114) ean be written 7 ae Valu), €) + HCO, Vy(ylt)s Bot (16) Since H* is a known function, this is a partial differential this is a partial differential equation for the Bellman Value Function V. With suitably defined bound: ary conditions in particular appcation, it can be solved. Once again, anzlytical solutions are available only in very simple special cases. But mumerical solutions are becoming increasingly Viable as computing technology improves, ‘Transversality Conditions We have so far kept the terminal time T and the associated target, stock requitement yy, fixed, while allowing the initial time and stocks to vary in order to set up the Bellman Value Function. But the reverse approach leads to some useful insights, too. I shall do this in continuous time, leaving the corresponding diserete-titne expressions for readers to derive, Write 1(y,€) for the maximum intogral of F over [0,1], with a fixed initial stock vector yy and a variable requirement y at t, When a greater stock must be left at the end, the maximized integral is smaller, and the shadow price interpretation becomes x(f) = —W,(y(¢),¢).. Splitting the problem into an initial interval (0,¢ ~ dt) and a small final interval Ataf we can apply Relinans ripe of Optimality a8 above and obtain, WiU(O,0) — Hut), Wy, ¢ 0. au7) This alternative approach allows us to extend the theory to snore general end-point conditions, It is natural to accept a fixed Dynamic Programming 169 initial time and historically given initial stocks at that tis forms of terininal conditions other than a fixed date and stock are easily conceivable, For example we may wish to attain a given target stock in the smallest possible time, or may have a more general trade-off between the terminal time and stocks. Suppose the aim is to maximize the integral (10.17), choosing both T sud (7) subject to a general constraint TyPT) $0. (ars) ‘This can be tackled in two stages. First we caf regard 7 and y(T) ‘a fixed, solve the standard Dynamic Programming problem, and find the value fuaction 1V(y(T),T) defined above. Then we choose T and y(T) to maximize this fonetion subject to the constraint (118). This isa standard static optimization problem, with frst order conditions Wy(WT)T)=ELUTIT Way(TIT) = 6 Jl T)T). where € is the Lagrange multiplier. Using the shadow price and Hamiltonian notation, these become EMUT)-T), HUT) (P),T) = § IdlD)-T) (19) ‘These say that the vector (7,—H") should be parsilol to th vector (Jy, J+) when both are evaluated at the optinnnmn tein point, Site the latter vector is perpendicular to the const urface J(y,t) = 0, the conditions say that the former vector ix iso perpendicular to the saine surface. ‘Therefore the conditions (11.9) ace called the transversality conditions "As aun example, suppose we wish to reach a given tuKet, soy ¥*, 0 minimum time. Then W(T) = —T ix to he maxinsized Subject to yiT) = yr. Newt Jr is identically 1 (11.9) says thet at the optimally chosen cind-point te Hansltosian H* sould loo be zero, Silly, i 7 is xed ut y(T) is unconstrained, then 4 in identically zero, and the transversality couition 3 ®(T) = 0 ‘These serve a8 boundary conditions that help us pin down the solution to the Dynamic Programming problem x(T)vie, Optimization in Bconomic Theory Infinite Horizons Another extension of the intertemporal optimization problem is important in many economic contexts. Often there is no natural way to specify the terminal date for the decisions being optimized, In fact we can rarely fix a date in advance and claim that consider- ations beyond it can be totally disregarded. This may be a minor problem for an individual, but it becomes mote and more impor- tant as we consider wider and wider contexts of decision-making, such as an extended family, a firm, and the economy as a whole Keeping the time-horizon finite, we can recognize that the termi- nal stocks will provide utility Rows beyond the horizon, and thus wlireetly take the future into account. But this is an imperfect solution. We eannot specify the correct terminal stock target, or a valuation function like J above, without paying explicit at tion to the fyture beyond the horizon. But that means sol ng 2 problem exactly like the original one tat with a longer horizon. Of course, there is no logical stopping point to this argument, and at forces us to allow an infinite horizon. We run into some technical problems when we consider deci- sions over an infinite time-horizon, The most important is that the integral (or the suin in discrete time) may not converge. A typical ‘cxample of this occnrs in the optimum growth problem of Exam- ple 10:2. Suppose the production funetion and the utility funetion are both linear, Fi =3kh Ul=6 and the marginal product of capital # exceeds the utility discount rate p. Neglect depreciation, that is, set § = 0. Now consider di verting one unit of output from consumption to saving, and letting it compound up to time 7, and consuming the additional output then. The extra output at T is eT, aud the present value ofits consumption is 8-0)", which exceeds 1, the opportunity cost of the current consumption forgone. Thus the postponement of con sumption is always desirable, and longer and longer postponement can make the utility integral larger and larger. But the limit of such policies means no consumption at all, which is the worst of all policies. ‘The condition thot is sufficient, and often necessary, fo ruling out such pathologies says that the shadow value of the tersainal stocks should go to 20: im, mT) y(T) =0. (21.20) Dynamic Programming am ‘The rigorous prou ofthis statements beyond our seope ete, Bot Tote at «natural extensog ofthe transversality com (ohics just the complementary alikness condition) for af Jech pebblem with a nonnegative teen stock reqs an fe this new condition, inte horizon optimization is no dfrent fram the finite horizon ease. Bollnan's Principle of Otuality at one tells us why. Consider any finite sian sub Soy te itil wed ternal condions Bred by the lange pete forthe subprobem, the Maximum Principle Dynamic Freer cette ppl But the iil and terminal ines eee cubpreblem eould be aititesy, x0 the conditions ans fact hold for the entire range (0.20), ‘An application of the infinite-horizon transversality condition fe to ne atime growth problem of Example 10.2. The paths ths chvenge to the steady stale (Rc*) in the phse ding as ie fouditon; the divergent palin general do oot This Slea ae toueeet the divergent pan, and forgiven itil epitl HHO) alee Sel conumption level e{0) tie on converge path Examples Bromple 11.1: Search ‘This iw greatly simplified model of job search. Tt does not aim to be realistic; its purpose is to introduce you to the Dyaumie Programing approach, and to prepare you for richer modes ite at the end of the chapter. ; "There is a whole spectrum of jobs paying different wages i the economy. ‘The cumulative distribution function ~ the probs bility that a randomly selected job pays w ar less is Or), The corresponding density funetion is gw) = "(w). The individu Knows these probabibities, but not the details of where say par Geular job is to be found of the wage it pays he must engage iat search for that information. ‘To fine ont about one joby he must stay unemployed with zcs0 income ud search for one period. At the beginning, of the next period he can either accept the job just found or continue search. If ie rejects the job he cannot return to, it later, His objective is to mnxinsize the mathenatical expectation ff the discounted present value of his wages. Ifthe interest rate is 7, define the discount factor 6 = 1/(1 +r), soa dollar in ¢ years? iow is worth 6 dollars now,172 Optimization in Beonomic Theory Consider the person starting, the last possible period of his working life. Suppose the job he observed during his sesech of the previous period pays 2 > 0. It is clearly better to accept this and work one peried than to remain unemployed. Thus he will accept any positive x, and his value function is simply Vi(z) = x, where the subscript denotes the number of periods of working life left Next suppose the same searcher has two periods of working life left, and the wage he observed last is x. If he accepts it, he gets 2 this period and the next, that is, a discounted present value of z +x. Ife searches, he gets no income this period, but finds out another job prospect and the wage y it pays. The next period being the last of his working life, he will accept it. Therefore the ‘expected income next period is the mathematical expectation Bul= [ vena, and this must be multiplied by 6 to get the present value in this period. Of the two alternatives, the searcher will piek the better, 90 Va(2) = max ((14 6), 6 Ely]) oy ‘There is critical value zy] = 6Ey0+8), (a2) such that the searcher with two periods of working life left will ‘accept the most recent offer 2 if and only if is greater than = ‘Therefore 23 i called this veareer’s reservation wage. In the same way we could define 2f for a searcher in his tnt period, and as we saw above, xf = 0 More generally, with n periods left, accepting the Intest = gets (45484... he = 2 (1-H), Continued search gets @ new observation y and an optimal decision starting next period (s0 discounted by 6) with (n — 1) periods to g0. Bellnsan's Equation becomes Vo(z) = max (2(1- 8)/(1— 8), 6 E[Va-ilyi})- (11.13) | | Dynamie Programming 173 ‘This produces a reservation wage #3. An induction argument mucky shows that the sequence of reservation wages 2, is increas ae he iy searchers with longer working ives alta of aie iiote selective about shat jobs they will accept, We can also solve for the functions V4(2) wing (11.18) recursively. If we let go to infinity, Va(2) and Va~s(2) alike converge to 4 limiting function V(z), whieh satisfies V(2) = max ( 2/(1-6), 5 EIV(u))). Since the same function V appears on both sides, the solution of (21.14) invalves a citcular (or move properly, fixed-point) reason: ing. ‘Think of the right-hand side as an operator, or a “function fof a function’, that starts with a function V and generates a new function, ‘Then we look for a particular V that leads to itself in this sway. This view also provides a method of numerical computation. Start with any V and apply the operator to get ® new one, then ‘apply the same operator to that, and so on. This process converges to the solution so long as a solution exists, which it does in this ‘example so long as 6 < 1 Thuisvelge he solution should have the same qualitative fs ture as that for long but finite life-spans, namely there should be fa reservation wage, of a critical 2° such that wage offers above this are accepted and those below trigger continued search. I shall proceed on this assumption and see where it takes us, For 2 > 2", (11.14) becomes: qi.) Viz) For smaller 2, we have V(2)= 6 EV), say, which is independent of 2. Bauating the limits of these ter native expressions as goes to 2° from the tight aad the left, V(at) = 2 (18) = SEU (i1.is) /0~6), Now avinl= [7 vorere ie vesvacnins tw/Q — 8) owl dy = ver) acy + vo) f wstedawa Optimization in Economic Theory Therefore V(z*) [1 — 6 B(2*)) = [6/1 —8)] _ why) dy, £ vty) dy. Since we know the functions and 9, this equation ean be solved. for 2*, and then the funetion V is coustant up to 2* and equal to x/(1 ~8) thereafter. s*[1-60(2")} a1.) Example 11.2: Saving under Uncertainty ‘This is again a simplified ‘starter’ example, Consider a consumer ‘vith an infinitely long life, wealth W that earss 2 random total re turn (principal lis interest) of r per period, and no other income. Consumption of C in any one period gives him utility wey-erya~— quazy with € > 0, as in Exercise 10.1. His utility discount factor is 4, and the objective over time is to maximize the mathematical ‘expectation of the discounted present value of utility Starting period with woalth W, if he consumes C and saves (W —C), his random wealth at the start of the next period will be r(W’ ~ C). Writing V(W) for his Beinn Value Function, the Bellman Equation is VOW) = max (CJ(L~ 0) +6 BV(r —C))]). (118) ‘The neat way to solve this problem is to guess a solution of 1 particular form and then verify it. Since wealth is split into consumption over a number of perieds and each gives utility of the form (11.17), @ natural form to try is vi) =A wy 2, (aiagy sxhere A is a constant to be determined. Using this in (11.18), we have A ge (Co aay e Om AA pacger oy}. cs) Dynamic Programming 175 ‘This simplifies to C=Ww/a+B) where I have used the abbreviation (a2) n p= (sai) Subotiouting back (1.20) and siping, we nd Ale (1 = tl Bie ‘This determines A provided which become he comin fr existence af nation tthe prob ae ae gee Tees ee anrengnug of te abate aly ous bis cae seting buck (1020 me ee “ cpw =1- (291) aia De ee eee cen eee Te Fe a ele a a eager ee Seesea te ce astaang means rae i Ferri vais Asp mepikla Soe span Gt la() i tormoly dese ith atta devin a, Then standard formals for te sexs cae El] = (lr)! exp(—e( ~ 6)? /2) ‘To see the consequences, consider the case of «= 1, New ou inerease in Ejr], holding @ fixed, decreases the eonsauptivny seeallly ratio, while an inerease in a, with Elp] fixed, messes Han rat ‘The opposite results hold if € > 1.116 Optimization in Economie Theory > t= rot optimum Fig. 11.1 ~ The shortest distance Brample 11.3: The Shortest Distance This example has no economic content, but has the great mesit, that the answer is known at the outset, enabling us to focus on the techniques that much better. Also, it illustrates the point that although the independent variable ¢ in the theory bas a natural interpretation as time, any other variable such as space can serve the sane formal role and the theory continues to apply. Consider the path of minimum length between the points (0,0) ‘and (1,0) in the plane; see Figure 11.1. Call the horizontal coordin- ate f and the vertical coordinate y. Tt is clenr that any patl which loops or winds cannot be of minimum length, becasise we ean sim ply omit the loop or an S:shape to get a shorter one. We can therefore restriet discussion to the ease where y is a single-valued function of t. The distance between the adjacent points (t,y) and (t+ dty + dy) is [(at)? + (dy)?}2. Let dy/dt = =, the control variable. Then we are to maximize -f[ (+2(y)'? at, (i123) Dynamic Progromming wT sbretie ‘i(t) = 2(t), (11.24) i the Stamm Pence, define the Hainan Ha 42 re. (a125) “The first-order condition for + to maximize this is —a4ey ete which simplifies to /(l— ty? (11.26) ‘Then the maximized Hamiltonian is Ht = (1-28), (127) Now the two differential equations are YH On = x f(A — x2), (11.28) and —OH*/Oy (11.29) ‘Thus r is constant along the optimal path, and then so is 2. But the only constant 2 that will keep y(1) = y(O) is zevo. Thus the shortest path is the horizontal straight line joining (0,0) aud (1,0) We can apply Dynamic Programming and consider 1 sonne ‘what more general problem, namely finding the shortest ie join ing (00) to the line t= 1. First we consider the suyprobem joing (090) to the general point (7, y) with T > 0 but y nnconstenined. "The theory above eontinues to apply, and we sew Hat. the straight line y(t) = ty/T does the job, Now it remain to clivose y opti- mally. ‘The transversality condition for that is a{') = 0. Since 1 is constant along the optimal path, it unist he 2ovo everywhere ‘Then # must be zero, and therefore y(7) = 0 is optimal, In other words, the shortest path from a point te a line is the straight from the point and perpendicular to the given fine178 Optimization in Beonomic Theory Exercises Exercise 11.1: Search Consider a variant of the search problem of Example 11.1. Suppose that the searcher can return to old jobs that he had rejected. So ‘at any time, the state variable z is the best offer he had observed up to that point. If he rejects this, searees and observes y, then next period he will start with the better of z and y. Thus instead of (11.14) we have Vaz) mnax (x (1~6")/(1~ 6), 6 E[Vy_s(max(z,y))]). Show that there is a reservation wage 2%, and that the sequence of reservation wages is increasing in n. Write down the Bellman Equation as n goes to infinity, and characterize its r*. Note that ELV (max(x,y))] ff versaars [ven onray = V(2) 8(2) + L Viv) stu) dy Exercise 11.2: Intensity of Research Effort ‘The research and development needed for a new product consists of completion of a nutuber of stages. ‘Think of these as a continuum of length L. If2(t) stages are completed at tie é, and flow cost o(#) is incurred on the R.&eD. program at this time, then the completion process evolves according to the differential equation #= fI(e). When all L stages are completed, a reward R acerues. If this occurs at time T, and r > 0 is the rate of interest, the net present value of the program is Ret f" qyerta ‘The aim i to choose the compl (aver [0,7] to maximize this. ion time T and the profile of costs, Dynamic Programming 179 Define V(2) to be the value function for the coresponding problem that starts with x stages completed, Applying Bellman’s Principle over an initial suuall time interval df, we have V(x) max [edt + V(e + St) 7 *] Expanding the right-hand side in Taylor seri asin the derivation of (11.5), convert this into + V(2) = max [V"(2) fle) ~¢) (11.30) For the ease fc) =e, with 0 0, and the constraint qualifies ion holds, anaiely the submatrix of G,(z) formed by taking those rows i for which G'(#) = ej has the maximum possible rank. Then there is a value of J such that L(2,9) <0, #20, with complementary slacknew, (3.7) and Lj(Z,) 20, X20, with complementary slackues. (3.10) Before we begin the proof, we ean simplify the notation. ‘The <0fer j= 1 fn. Thus we can subsume them into te general ineqnality non-negativity constraints can be expressed as a constraints, whose number m is increased accordingly. We can also choose the labels on the component finsetions G* suc tht Constzaints 1,2... kage binding, nd 1, «ate slack, of Bo, feini2..k ow { (any Fla), (43) Note that we do not need any restrictions arising from the con- straints K-41, ....m; since the constraints hold as strict inequalities ‘at 2, by continuity they will go on holding for x stfteiently near Now write = +d where de is infinitesimal. Then we want toreplace (4.2) and (4.3) by the first-order Taylor approximations: Giz)dr <0 fori=1,2,...k, (44) and F,(2) de > 0, (4s) ‘This is valid provided the I-by-n matrix G,(2) formed by stacking the k row: G!(2) has rank £. Tf it has a smnaller rank, it has too large mull'space, that is, too many veetors di yield ero when multiplied by this matrix. ‘Then many more dr satisfy the linear approximation (4.4) than do z+ dr satisfy the true coustraints (4.2). AS a result, the first-order conditions ean fail even though is optizum, I shall illustrate this using an example with two variables and {wo constraints. Suppose the objective function is Flees) =n and the constraints are S22at or Gre) srry <0, Soa of Cum eetal ce The Kuhn-Tucker Theorem 183 Lig Fig. A.1 — Failure of the constraint qualification Figure A.1 shows the feasible set as the shaded area, and then it is clenr that (0.0) solves the constrained maximization problem, At this point both constraints bind (4 = 2), and the vectors of derivatives of the constraints are G4(0,0)=(0,-1) and G4(0,0) = (0,1). ‘The rows are linearly dependent, and the rank of the matrix of derivatives is 1, which is < &. ‘The linear approximation (44) becomes dey <0 tnd dey <0, thiis, dry =0, hich is the whole horizontal axis. Points te the right of the origin are not in fact a linear approxinnation to th feasible set, but they appear as feasible deviations in the linear approximation to the functions. Now we also see the sevise of the rexnack made earlier in the context of linear programmaing (Example 7.1, p. 100) that no con straint qualification was needed there, When the constraints are184 Optimization in Economic Theory Fig. A-2: The Kubn~Tucker Theorem already linear, there is no need to find linear approximations to them, so the issue does not arise Let us proceed with the assumption (the Consteaint Qualifica- tion) that the rank of G.(2) equals k. Figure A.2 shows the situa- tion. Tt is drawn so that all functions are inereasing in 2, but that is only to conform to the economic intuition; other configurations ineke no difference. The feasible region, and the upper contour set of F(z) for its optimal value, are both shown by hatch-marks on their boundaries. The vestor of derivatives of each of the funetions is also shown; each is normal (perpendicular) to the contour curve of the corresponding function. When # is optimum, the feasible region and the uppe jon and the upper con tour set of the objective function should not interseet. Assuming ‘The Kuhn- Tucker Theorem 185 validity of the constraint qualification, their linear approximations defined by the deviations in (44) and (A.5) should not intrsect tether. For this to be true, the vector of derivatives Fy(2) should Tie in the cone formed by the vectors Gi (z) of the derivatives of the constraint functions. ‘This is more easily seon frou tHe eqtivalent statement in the opposite direction (contrapositive) from the figs ture: f F,(2) were to lie outside this cone, the contout of F through, ir would cut into the feasible region, so a neighboring, feasible point with @ higher value could be found, and z could wot be optimal, Points in a Suite cone are non-negative linear combinations of the vectors that define the cone. Therefore there are non-negative rmumbers Ay, Xz, +++ Ak such that Fa) =Y) NG) (asy Ic is harmless to extend the range of summation to m by defining Desay v1 dm all equal to zero. The extended (4.6) is just the condition £,(2,4) = 0 to which (3.7) reduces when there are ne non-negativity constraints (remember they are subsumed into the inequality constraints and so do not have a separate role). ‘The complementary slackness condition (3.10) i fulfilled beeause of the the A, are defined for the slack constraints. This completes the sketel of the proof. Further Reading [A more detailed proof ean be found in E. MALINVAUD Lectures on Microeconomic Theory, Auster dam: North-Holland, 2nd edn., 1985, Appendix by J-C. Millevent Tputiligator (op. cit.) ch. treats the concave case in fuller detailIndex Bold-face page numbers represent the definition or the main occurrence of a term in the text. Italie page numbers indicate tan example or an exercise where the concept figures protiinently. Roman page numbers are for other significant occurences Adverse selection 180-40 Deceutralization 77 Arbitrage 3, 12 Demand functions 65.6 compensated 66, 14-16 Bellman Equation 164, 172, ents 14 Dynamic programming 164 Bellman Value Function 164, oie * 178-8 Entry-dotersence 160 Bellman’s Principle of Optima. Etvelope Theorem 57-8, 69 lity 164 second-order 50, 113 ary point 74, 88 ad utility 123, Badge fie | Bxpentiture function 82-9 Boundar First order conditions 16 nocessary 16, 18, 95 sufficient 97-8 gts Invisible hand 44, 48-9, 53 Global maxima 106 udiference curve 1 Indirect utility function 8.4 Inequality constraints 29-30, Kula Tucker Theoresn 32 Convex set 7 Toe) Comer solution 6, 27-8, 85 Lagrange multipliers 24, 42 Cost curves 69 Lagrange’s method 17-18, Cost function 63, 116 26188 Lagrangian 17 Le Chatelier-Samuelson Principle 115 Linear programming 100-4 Marginal cost 29 Marginal rate of substitution 5 of transformation 17 Marginal utiliry of income 6 Maximum principle 150 Maximum Value Function 57, 87, 103 Mean-variance frontier 130, 140-2 Moral hazard 1 139-6 Necessary conditions 16 No-arbitsage comdition 5, 151 Non-negativity constraints 26-8, 51-2 Numéraire 92 Objective function 10 Optimum growth 156-59 Phase diagram 153, 158 Persfolin choice 129: 140-2 Producer behavior 28, 68-4 84, 116-20 Production possibility frontier 20, $4, 85 Profit function 84 Index Quadratic form 108 ‘Quasi-concave funetion 71, 76 Quasi-concave programming 97-9 Quasi-convex function 71, 76, 4 Rationing 97-8 Regular maxima 107 Revealed preference 115 Risk-aversion 123, 129 Saving behavior £42, 154-5, 159, 14-5 Second-order conditions 17 necessary 106 sufficient 107, 111, 118-9 Search behavior 171-4, 178 Separation Theorem 4, 81,89 Shadow prices 42-5 Slater condition 90 Slutsky-Hicks equation 66 State variables 146 Stationary point 17 Substitution effect 115, ‘Transversality conditions 151, 168-9 Uncertainty 122, 165 Uniqueness 78-80, 99-100 Utility function 1, 19, 21-3