(Linear programming isthe name used for problems in wich te g maximize or minimize a linear function subject to linear inequality As the opening quotation from Lovasz indicates, linear mathematical technique of immense Practical importance. Since 1940s, there have been important theoretical and computational its methods are now used every day in many parts of the Because of its extensive use in economic decision proble should have some basic knowledge of this theory. However, even beyond its practical applications.(in particular, the dualit Programming is a basis for understanding more complicated op with even more interesting economic applications.) In principle, any linear programming problem (oft can be solved, if a solution exists. The simplex me G. B. Danwig is a very efficient numerical finite number of operations. This method fact, for economists, it is probably more of LP than the details of the simplex met LP problem, it is natural to use one of Programs to find the solution.1 Sec. 19.1 / Preliminaries 723 oA preliminaries ea : general linear programming problem with only two decision variables involves maximizing or minimizing a linear function t= 1X1 +¢2%2 (criterion function) subject 10 constraints ay 1x) +aI2%2 < by 2)x) +0222 < be "Sell (inequality constraints) GmiX1 + Om2%2 S bm usually, we also impose explicit nonnegativity constraints on x; and x2: x1 20, %220 — (nonnegativity constraints) Note that the direction of the inequalities in the inequality constraints is a conven- sion in the sense that any inequality of the form ax; + bxz > ¢ is equivalent to the inequality —2*1 — bx, 5 -¢. A Graphical Approach to Simple LP Problems LP problems with only two decision variables can be solved by a simple geometric method. x Example 19-1 2 . A baker has 150 kilograms of flour, 22 kilos of sugar, and 27.5 kilos of butter with which to make two types of cake. ‘Suppose that making one dozen A cakes requires 3 kilos of flour, 1 kilo of sugar, and 1 kilo of butter, whereas making one dozen B cakes requires 6 kilos of flour, 0.5 kilo of sugar, and 1 kilo of butter. Suppose that the profit from one dozen A cakes is 20 and from one dozen B cakes is 30. ‘How many dozen A cakes (x1) and how many dozen B cakes (x2) will maximize the baker's profit? Solution ‘An ourpat of x dozen A cakes and x; dozen B cakes would need 3x; + 6x2 kilos of flour. Because there are only 150 kilos of flour, the inequality 3x, + 6x2 < 150 (flour constraint) w) must hold. Similarily, for sugar, x1 +0.5x2 522 (sugar constraint) Q)724 Chapter 19 / Linear Programming and for butter, xy+2x2 527.5 (butter Of course, x1 = 0 and x2 > 0. The profit ‘A cakes and x2 dozen B cakes is 20x) + 30x In short, the problem is to 3x, + 6x2 max z = 20x) + 30x2 S- x + 0.522 xp big aa, This problem will now be solved graphically. ‘The output pair (x;,%2) is called fe lem [5] if all the five constraints are which is 3x1 + 672 < 150. If we use all inputs. Figure 19.1 shows the straight lines the sugar border, and the butter border. Ta o the nonnegative quadrant, the set of feasible p {feasible (or admissible) region, is the shaded set S is a so-called convex polyhedron, and | C, and D are called extreme points of the FIGURE 19.1 Flour Border 32) +612=150 ‘Sugar Border x1+0.522=22 204 Bunter Border x1+22=27.5Sec. 19.1 / Preliminaries 725 FIGURE 192 In principle, the baker can find the point in the feasible region that maximizes profit by calculating 20x; + 30xz at each point of S and picking the highest value. In practice, this is impossible because there are infinitely many feasible points. Let us argue this way instead. Can the baker obtain a profit of 600? If so, the straight line 20x; + 30x2 = 600 must have points in common with S. This line is represented in Fig. 19.2 by dashed line Ly. It does have points in common with 5. (One of them is (x1, x2) = (0, 20), where no A cakes are produced, but 20 dozen B cakes are, and the profit is 20-0 + 30-20 = 600.) Can the baker do better? Yes. For instance, the straight line 20x; + 30x2 = 601 also has points in common with S and the profit is 601. In fact, the straight lines 20x; +30x2=c (cis a constant) (4 are all parallel to 20x; + 30x; = 600. As c increases, the line given by [6] moves out farther and farther to the northeast. It is clear that the straight line [6) that has the highest value of c and still has a point in common with S is dashed line L2 in the figure. It touches set S at point B. Note that B is at the intersection of the flour border and the butter border. Its coordinates, therefore, satisfy the two equations: 3x) +6x=150 and xy +2 = 27.5 Solving these two simultaneous equations yields x; = 5 and x, = 22.5. So the baker maximizes profit by baking 5 dozen A cakes and 22.5 dozen B cakes. This uses all the available flour and butter, but 22—5—0.5-22.5 = 5.75 kilos of sugar are left over. The profit earned is 20x; + 30x2 = 775. Our next example is a minimization problem.726 Chapter 19 / Linear Programming Example 19.2 A firm is producing two goods, A and B. It has two ‘ produce the two goods in the following quantities (per hour). tha ing Factory 1 Factory 2 Good A 10 20 Good B 25 25 ‘The firm receives an order for 300 units of A and 500 units of B. Ths operating the two factories are 10,000 and 8000 per hour. Formulate thea programming problem of minimizing the total cost of meeting wees Solution Let u; and up be the number of hours that the two . operate to produce the order. Then 10uy+20u2 units of good A are Factories and 25uy + 25u2 units of good B. Because 300 units of A and 500 uni B are required, IS of 10u; + 20u2 > 300 25uy + 25u2 > 500 wu In addition, of course, u; > 0 and uz > 0. The total costs of operating the two factories for u, and uz hours, respectively, are 10,000; +8000u2. The problem is, therefore, 10u; + 20u2 > 300 min 10,000 u; +8000u2 subject to Ast F peun-> 500) 9 tem ‘The feasible set S is shown in Fig. 19.3. Because the inequalities in (1] are of the > type, the feasible set lies to the northeast. ‘We show some of FIGURE 19.3 a 7. 25uy + 25u2 = 500Sec. 19.1 / Preliminaries 727 the evel curves 10,000u; + 8000u = c, marked L =o, }» Lz, and L;. three correspond to the values 100,000; 160,000; a 240,000 of pa level c. As ¢ increases, the level curve moves farther and farther to the northeast. The solution to the minimization problem is clearly the level curve that touches the feasible set S at point A with coordinates (0,20). Hence, the optimal solution IS to operate factory 2 for 20 hours and not to use factory 1 at all. The minimum cost of producing the order is 160,000. ‘The graphical method of solving linear programming problems works well when there are only two decision variables. It is possible in principle to extend the method to the case with three decision variables. Then the feasible set is a convex polyhedron in 3-space, and the level surfaces of the criterion function are planes in 3-space. However, it is not easy to visualize the solution in such cases. For more than three decision variables, no geometric method is available. (By using duality theory, however, one can solve LP problems geometrically when either the number of unknowns or the number of constraints is less than or equal to 3; see Section 19.5.) Both the previous examples had optimal solutions. If the feasible region is unbounded, however, a (finite) optimal solution might not exist, as is the case in Problem 2. The General LP Problem ‘The general LP problem is that of maximizing or minimizing 2=01%1 +--+ Ca, (criterion, or objective, function) 19.1] with cj, +, 8 given constants, subject to m constraints a bx inequality constraints) {19.23 = = S bn where elements aij and by are given constants. Usually, we assume explicitly that x) 20,....%,20 — (nonnegativity constraints) 19.3) There is no essential difference between a minimization problem and a maxi- nization problem, because the optimal solution (x, ..., x3) that minimizes [19.1] subject to [19.2] and [193] also maximizes —z, and minz = —max(—z). An nvector (x1, ..,1,) that satisfies (19,2) and [19.3] is called feasible.728 — Chapter 19 / Linear Programming boundary are called faces and the points O, P, Q. ‘extreme points. In n-space, a convex polyhedron If an LP problem has a solution, then it must point. The simplex method is a procedure that : treme point to another in such a way that the value of decreases, until we come to a point such that by n is impossible to increase the value of the criterion fi the optimal solution. Problems 1. Use the graphical method to solve the following 3x + 2m 56 n+4n 54 uy + 3a 21 2u; + Suz 2 2 —2 + 3m = % max 2x; + 5x2 st. < a. max 3x) +422 sit b. min 104; +27u2 sit { @. max 81, +92 e. mx -2) +)Sec. 19.2 / Introduction to Duality Theory 729 2. a. Is there a solution to the following problem? ut on max x) +22 S.t. SF (ems 420,220 p. Is there a solution if the criterion function is z = —x) — x3? 3, Set A consists of all (x1, x2) satisfying 2+ <2 x)20,m20 a) +2r <8 Solve the following problems with A as the feasible set: a. max X2 b. max x) © max 3x) + 2x2 d. min 2x) — 2x2 e max 2x) +4r) f. min — 3x) — 2% 4. A firm produces two types of television sets, an inexpensive type (A) and an expensive type (B). The firm eams a profit of 700 from each TV of type A, and 1000 for each TV of type B. There are three stages of the production process. Stage I requires 3 hours of labor on each set of type A and 5 hours of labor on each set of type B. The total available number of hours is 3900. Stage II requires 1 hour of labor on each set of type A and 3 hours on each set of type B. The total labor they have is 2100 hours. At stage III, 2 hours of labor are needed for both types, and 2200 hours of labor are available. How many TV sets of each type should the firm produce to maximize its profit? 5, Replace the criterion function in Example 19.1 values of r will the maximum profit still be at 1 by 20x) + tap. For what =5 and x2 = 22.5? 19.2 Introduction to Duality Theory nn problem involving scarce resources, an economist will often ask: What happens to the optimal solution if the availability of the resources changes? For linear programs, answers to questions of this type are intimately related to the so-called duality theory of LP. As a point of departure, let us again consider the baker's problem in Example 19.1. Example 19.3 Suppose the baker were fo receive (free of charge) one extra Kilo of flour. How much would this extra kilo add to his maximum profit? How much would an extra kilo of sugar contribute to profit? Or an extra kilo of butter? Solution If the baker receives one extra kilo of flour, his flour border would become 3x1 + 6x2 = 151. It is clear from Fig. 19.2 that the feasible Confronted with an optimizatio730 Chapter 19 / Linear Programming set § will expand slightly and point B will the butter border. The new optimal point B’ will ines 3x) + 6x2 = 151 and x) + x = 2755. x) = 14/3 and x: = 137/6, Then the criterion 20(14/3) + 30(137/6) = 2335/3 = 775 + 10; 7 If the baker receives an extra kilo of sugar, but the optimal point is still at B. Recall that at problem, the baker had 5.75 Kilos of unused sugz An extra kilo of butter would give a new opti tion of the lines 3x; +6x2 = 150 and x; +2 = 28.5. gives x) = 7 and x2 = 21.5 with 20x; + 30x2 = 7 ‘These results can be summarized as foll a. An extra kilo of flour would increase the b. An extra kilo of sugar would increase t cc. An exta kilo of buner would increase: Se interesting properties. Suppose (x1,x2) is a feasible pair in in the | (1), (2), and 3] in Example 19.1 are satisfied. 0, and {3} by 10. Because the multipliers are all preserved. That is, (10)Gx + 6x2) < 107) ] O(n +0.5x) <0-22 1 (x1 + x2) < 10 s 10x; + 20x2 + 10x; + 10x2 which reduces to Thus, using the “magic” numbers uj, Proved that if (x), x2) is any feasible pair, then be less than or equal to 775. Because x, = 5 an 775, we have in this way proved algebraically- = ‘Sec. 19.2 / Introduction to Duality Theory 731 ‘The pattern revealed in this example tums up in all linear ing ‘oblems. In fact, the numbers uj, u3, and u$ are solutions to a Eee eckien called the dual of the original problem. The Dual Problem Consider once again the baker’s problem, which is ; 3x, + 6 < 150 max 20x; +30x2 subjectto ¢ x) + 05% < 22 m220 [1] a+ om S215 Suppose the baker gets tired of running the business. (After all, there must be many complaints about the rather plain cakes.) Somebody else wants to take over ind buy all the baker's ingredients. The baker intends to charge a price for each Kilo of flour, uz for each kilo of sugar, and u3 for each kilo of butter. Because one dozen A cakes requires 3 kilos of flour and 1 kilo each of sugar and bumer, the baker will charge 3uy-+us-+us for the ingredients needed to produce ‘a dozen A cakes. The baker originally had a profit of 20 for each dozen A cakes, and he wants to eam at least as much from these ingredients if he quis. Hence, the baker insists that the prices (1, uz, us) must satisfy 3u, +u2 +u3 > 20 table to use the ingredients himself to produce A Otherwise, it would be more profi before for the ingredients cakes. If the baker also wants to eam at Jeast as much as needed to produce a dozen B cakes, the requirement is 6uy + 0.Su2 + U3 2 30 presumably, the entrant wants to buy the baker's resources as inexpensively as posible. Te total cost of 150 kilos of flour, 22 kilos of sugar, and 27.5 kilos of bower ig 150uy +224 +27.5us. In order to pay as litle as possible while having the baker accept the offer, the entrant should suggest prices u; > 0, 42 2 0, and us > 0, which solve eS min 150u, + 22u2 +27.5u3 subject to Hee eee z wus 2 2 with u, > 0, u2 2 0, and u3 2 0. An interesting question is this: If the baker lets the entrant take over the business and solve problem (2], will the baker earn as rauch as before? It tums out that the answer is yes. The solution to (2) is uj = 10/3, and u3 = 0, and u3 = 10, and the amount the baker gets for selling the resources is 150uj +22u5 +27.5u5 = 715, which is precisely the maximum value of the criterion function in problem [1].732 Chapter 19 / Linear Programming ‘The entrant pays for each ingredient exactly the marging which was calculated previously. In particular, the Price the baker has more than he can use optimally, = Problem [2] is called the dual of problem (1). The be closely related. Let us explain in general how to cx problem. The General Case Consider the general LP problem ax + max ¢1x1 +~-- + Cn%n subject to 4 -- GmiX1 + a with nonnegativity constraints x; > 0,...,%, 20. ‘The dual of [19.4] is the LP problem min buy +---+bmUm Subject to { sense with nonnegativity constraints u; > 0, ..., Um = oF is constructed using exactly the same coefficients biss-+) Bm as in [19.4]. ; In problem [19.4], which we now refer to as variables x),..._ and m constraints (disregarding the nonn the primal is a maximization problem, the dual is a problems, all variables are nonnegative. The m const are of the “ess than or equal to” type, whereas the 7 of the “greater than or equal to” type. The in either problem are the right-hand side elements problem. Finally, the two matrices formed by the co constraints in the primal and dual problems are they take the form ay a2 42) G22... ; A= | ee v | and Al Gm Gm2 =. eh Check carefully that problem [2] is the dual of problem | | pe crete DS eae ae2 Sec. 19.3 / The Duality Theorem 733 Matrix Formulation Let vs introduce the following column vectors (matrices): -0) <0 0-0) = ‘Then the primal can be written as follows (with A and A’ given by [19.6])- max ¢c'x subjectto Ax0 [19.8] ‘And the dual can be written as min b’n subject to Alu>c, u>0. Itis more convenient, however, to write the dual in a slightly different way. Transposing tara > ¢ using the rules in (12.44) of Section 12.9 yields u'A > ¢, and moreover piu = wb, So the dual can be written as ‘ min ub subject to wWA>c, u20 19.9] Problems 1. Consider Problem 1(a) in Section 19.1. "a. Replace the constraint 3x; + 2xz < 6 by 3x +2x2 <7. Find the new optimal solution and compute the increase uj in the criterion function. b. Replace the constraint x; + 4x2 < 4 by x, +4r2 <5. Find the new optimal solution and compute the increase w3 in the criterion function. ce By the same argument as in Example 19.3, prove that if (x1, x2) is feasible in the original problem, then the criterion function can never be larger than 36/5. 2. Write down the dual to Problems 1(a) ‘and (b) in Section 19.1. 3. Write down the dual to Problem 1(d) in Section 19.1. 19.3 The Duality Theorem This section presents the main results relating the two solutions to an LP problem and its dual, We begin by considering the baker’s problem yet again. Example 19.4 Consider problems (1) and (2) in Section 19.2. Suppose that (x1, x2) is an arbitrary feasible par in [1], which means that x > 0, x2 2 0, and the three inequalities in (I] are satisfied. Let (us, ug, us) be an arbitrary feasible Triple in (2). Multiply the < inequalities in [1] by the nonnegative numbers tu, up, and up, respectively, and then add the inequalities. The result is the734 Chapter 19 / Linear Programming a new inequality (Bx; + 6x2)ie1 + (a1 + 0.5x2)u2 + (1 + x2)u5 < 150uy +29, @+275,, iy Rearranging the terms on the left-hand side yields (Guy + ua + u5)x1 + (6) + 0.5u2 + u3)x2 < 150m) + 224, +27. 242754. iy by the This gives Guy + uz + u3)x1 + (6uy + 0.5u2 + u3)x2 > 20r + 30x, In a corresponding way, we multiply the > inequalities in (2 negative numbers x; and x2, respectively, and add the results (+s) From [+] and [++] together, it follows that 150u; + 222 +27.5us > 20x) + 30x, ir for all feasible (x1, x2) in problem [1] and for all feasible (1,13.43)j problem [2]. We conclude that in this example, the criterion function : the dual problem is always greater than or equal to the criterion fine. tion of the primal problem, whatever feasible (x)..x2) and (uy, 3,13) are chosen. The inequality [+] is valid for the feasible pair (x), x2) = (5,22) in particular. For each feasible triple (wu), u2.u3), we therefore obtain 150u; + 22u2 + 27.5u3 > 20-5 +30-22.5 =775 It follows that if we can find a feasible triple (uj, u3, u3) for prob- em [2] such that 150uj + 22u$ + 27.5u3 = 775, then (uj, u5, 13) must solve problem (2), because no lower value of the criterion function is ob- tainable. In Section 19.2, we saw that for (uj. u3, 43) = (10/3, 0, 10), the criterion function in the dual did have the value 775. Hence, (10/3, 0,10) solves the dual problem. The first general result we prove is the following: Theorem 19.1 If (x1, ....%n) is feasible in the primal problem [19.4] and (u1,..., um) is feasible in the dual problem [19.5], then [19.10] Dyuy +++ + Buti = C1X) +++ + CpXq So the dual criterion function has a value that is always at least as large as that of the primal.Sec. 19.3 / The Duality Theorem 735 Proof Multiply the m inequalities in (19.4) by the nonnegative numbers -1++Mm» then add. Also, multiply the n inequalities in (19.5) by the non- negative numbers x1,...,n, then add. These two operations yield the two inequalities (ayixr #0 + inka) Ho + GmiX1 +--+ Gan Xn)lm Siu) +-<- + bm (ayy Hoe + mim) x1 Ho + Gin) +--+ + Gmntim)in > 1X1 +--+ Cnn By rearranging the terms on the left-hand side of each inequality, we see that each is equal to the double sum S77, 07. ajjuixj. Then (19.10) follows immediately. From Theorem 19.1 another interesting result is obtained: ‘Theorem 19.2 Suppose that (xf, ...,x¢) and (uj,...,u3,) are feasible in problems (19-4] and [19.5], respectively, and that caxt beset Catt = buf +21 bly, 19.1) ren (ef, ---+3) solves problem [19.4] and (uj,..-,4%) solves problem (19.5). Let (x), ....%,) be an arbitrary feasible n-vector for problem (19.4). Proof Um = uz, as well as [19.11], we obtain Using [19.10] with wy = uf. -..- c1xy +20 Gata Sbyuy +27 + mtn Sexy tt en ka This proves that (x7, cm) solves [19.4]. ‘Suppose that (u ". Um) is feasible for problem [19.5]. Then [19.10] and (19.11) together imply that dyer +--+ + bmtim Dept + ente = by +--+ Pmt This proves that: (uj, ..-. 44%) solves (19.5). Theorem 19.2 shows that if we are able to find feasible solutions for prob- lems (19.4) and [19.5] that give the same value to the criterion function in each of the two problems, then these feasible solutions are, in fact, optimal solutions.736 Chapter 19 / Linear Programming Theorem 19.3 (The Duality Theorem) Jem [194] has a (finite) optimal solution. Then the dual problem [19.5 also has a (finite) opumal solution, and the corresponding values of the a terion functions are equal. If the primal has no bounded optimum, i dual has no feasible solution. The most important result in duality theory is the followin, 8: Suppose the primal. prob. then the The proofs of Theorems 19.1 and 19.2 were very simple. It is much more es to prove the first statement in Theorem 19.3 concerning the existence of to the dual, and so we shall not attempt to provide a proof here. The Solution ment in Theorem 19.3, however, follows readily from inequality (19.10), e% (u),.-+.un) is any feasible solution to the dual problem, then iis, +. cane is a finite number greater than or equal to any number cix +--+ cay. «nut (t1..-+5%s) is feasible in the primal. This puts an upper bound on the rou” values of ¢)x1 +--+ Gn%n- possible Note: Itis a useful exercise to formulate and prove Theorems 19.1 and 19.2 usin matrix algebra. Let us do so for Theorem 19.1. Suppose x is feasible in (19.8) and w is feasible in [19.9]. Then f u’b > u'(Ax) = (W/A)x > cx You should note carefully how these inequalities comespond to those we established in the earlier proof of Theorem 19.1. Problems 1. Consider the problem 20 21 nt 4x + 5y max 2x +7y subject to 1s 5 IA 1A a. Solve it by a geometric argument. b. Write down the duab and solve it by a geometric argument ¢. Are the values of the criterion functions equal? (If not, then act to Theorem 19.3, you have made a mistake.) 2. Write down the dual to the problem in Example 19.2 and solve it Check that the optimal values of the criterion functions are equal. ee 3. A firm produces small and medium television sets. The profit is ee each small and 500 for each medium television set. Each televisiol t to be processed in three different divisions. Each small television req respectively 2, 1, and 1 hour in divisions 1, 2, and 3. The comesP™® | numbers for the medium television sets are 1, 4, and 2. Suppose diviss cording “iSec. 19.4 / A General Economic Interpretation 737 and 2 both have a capacity of at most 16 hours per day, and division 3 has a capacity of at most 11 hours per day. Let x; and x2 denote the number of smal] and medium television sets that are produced per day. ‘a. Show that in order to maximize profits per day, one must solve the following problem: 2 + m < 16 max 400x; + 500x2 subjectto ¢ x; + 4x2 < 16 x, 20,220 x) + 2n < 11 b. Solve this problem graphically. . If the firm could increase its capacity by 1 hour a day in just one of the three divisions,-which should be the first to have its capacity increased? 194A General Economic Interpretation This section gives a general economic interpretation of the LP problem [19.4] and its dual [19.5]. ‘Think of a firm that produces one or several outputs using m differ- ent resources as inputs. There are n different activities (or processes) involved in the production process. A typical activity is characterized by the fact that running it a unit level requires a certain amount of each resource. If ai; is the number of units of resource é that are needed to run activity j at unit level, the vector with components G1 j, 22j----+@mj EXPFESSeS the m different total resource requirements for running activity j at unit level. If we run the activities at levels x1, ---,%n, the total resource requirement can be expressed as the column vector ‘laaoat) If the available resources are by, ..-, bm, then the feasible activity levels are those aS that satisfy the m constraints in [19.4]. The nonnegativity constraints reflect the fact that we cannot run the activities at negative levels. Each activity brings a certain “reward.” Let cj be a measure of the reward (or value) created by running activity j at unit level. The total reward obtained by Tunning the activities at levels x), ...,%n is then ¢)x1-+-*-+Cn%n- The problem of the firm is therefore to solve the following LP problem: Find those levels for the n activities that maximize the total reward, subject 10 the given resource constraints. ‘The problemi in Example 19:1’can be interpreted in this way. There are two activities of making each of the two types of cake, and there are three resources— flour, sugar, and butter. Let us tum to the dual problem [19.5]. In order to be in business, the firm has to use some resources. Each resource, therefore, has a value or price. Let uj738 Chapter 19 / Linear Programming be the price associated with one unit of resource j. Rather than th; a market price for resource j, we should think of it as (somehow) think of tr relative contribution that one unit of resource j makes to the total ing 3 These prices are not real prices, so they are often called shadow pris ic Tesulp Because d1j.02/,---+@nj are the numbers of units of each of, es, needed to run activity j at unit level, ay/uy + apju2 + -- (shadow) cost of running activity j at unit level. Because of running activity j at unit level, the m res +H On tip i + Gmjum is 0/3 te evade Cj — (Qyjuy + 42jU2 +--+ + Anjum) can be regarded as the (shadow) profit of running activity j at uni that the jth constraint in the dual problem {19.5 says that RE Gado, Nee running activity j at unit level is < 0. Profit from The criterion function Z = buy +--+ Bnim in the dual LP problem measures the (shadow) value of the initial stock of all the resources. The dual problem is, therefore, the following: Among all choices of nonnegative shadow prices uy, ..-.Um such that the profit of running each activity at unit level is < 0 Jind one that minimizes the (shadow) value of the initial resources. sn) The Optimal Dual Variables as Shadow Prices Consider again the primal problem yx) + -+> + Ginkn S By max c1x1+++*+Cn%n Subject to { --- : mix) + + + Gnntn S bm with x) > , > 0. What happens to the optimal value of the criterion function if the numbers b;,...,y change? If the changes Abj.-..,Abm are positive, then the feasible set increases and the new optimal value of the criterion function cannot be smaller. (Usually it increases.) The following analysis also applies when some or all the changes Ab; Abm are negative. How big is the change in the optimal value?, Suppose (xj, (gf + Ax)....,x3+ Axq) are optimal solutions to the primal when side is respectively (by, ..., Bm) and (bj + bj, ---,bm+Abm). If Abis-++9° axe all sufficiently small, the duals of the two problems often have the same optimal solution u*,....u%,. Then, according to Theorem 19.3, one has Cuxf tee cnx = byuy +--+ + bmi, cuCay + Ay) 4 --- teg lag + Ate) = (61 + Abi uy +--+ Om + Sdn dln Hence, by subtraction, CyAxy Hoo + CpAty = UAB, +++ + UZ DmSec. 19.5 / Complementary Slackness 739 Here the eta Bea is change we obtain in the criterion function in (+) when by, -+-2 Om are changed by Abj,..., Abm, respecti thi: i ! a ean spectively. Denoting this change in uphby +--+ + us Abn (19.12) je: The assumption underlying [19.12] is that the changes in the b’s d e optimal dual variables to change. Se ae ae No! \‘cause the problems 1. Consider Problem 1 in Section 19.3, 4x + Sy < 20 max 2x +7y subjecrto {52 S21 x2>0,y20 We found that the optimal solution of this problem was x" = 0, and y" = 3. with 2x* + 7y* = 21. The optimal solution of the dual was uj = 0 and u = 1. Suppose we change 20 to 20.1 and 21 to 20.8. What is the corresponding change in the criterion funetion? 49.5 Complementary Slackness Consider again the baker's problem [1] in Section 19.2 and its dual (2]. The solution to [i] was x} = 5 and x3 = 22.5, with the first and the third inequalities both ‘atisfied with equality. The solution to the dual was uj = 10/3, u3 =O, and u3 = 10, with both inequalities in the dual satisfied with equality. Thus, in this example, i 2 the first and second inequalities ee 2 Oe (eum ae setsied with equality a om the first and third inequalities =O, {2 “hr ceimal are saished'with equatiry = ecause the shadow prices of flour and butter are posi- the available flour and butter are used. ‘We do not zero—it is not a scarce resource. We interpret [2] this way: Be tive, in the optimal solution, all use all the available sugar, so its shadow price is These results hold more generally. Indeed, first, consider the problem auxitanx Sh max c)x, +¢2x2 subject to 4 @2x1+anrt2 = bz 20,20 GB) aX) +432%2 0. Suppose (xj. [4]. Then x3) solves (3 ax} +.a2x3 Sb (a) aaxj+anx3 Shr (0) is 3 a rut + ¢ a3)x; +a32%3 <3 4 Multiply the three inequalities in [5](a) by the three and uj, respectively. Then add the results. This yields 1 ope Multiply the two inequalities in [5](b) by xj and This gives . (aru + a2yud + a31u3)xq + (ar2uj + a22u3 + asat Looking closely at the left-hand sides of the two in that they are equal. Moreover, by the duality theorem right-hand sides of (6) and [7] are equal. Hence, be replaced by equalities. Thus, replacing < by because uf, us, and uf are all > 0, the three terms in their sum would be <0. Because their sum is 0, a (ajixt + aj2x5 — by)uj =O We conclude that aixt +52 Using the fact that > in [7] can be replaced by same way as before, we also get ayju} + aru} + asus > cj The results in (9] and (10) are called the comple: The arguments used to show that these conditions areSec. 195 / Complementary Siackness 741 forward way to the general case. Furthermore, the same complementary slackness conditions are also sufficient for optimality. Here is a general statement and proof: 19.4 (Complementary Slackness) __ Suppose that the primal 19.4] of Section 19.2 has an. opumal solution x* = (Xf, ++ Any ‘gs the dual [19.5] has an optimal solution u* = (uj, ..., “7,). Then Beoot n, and j = versely, if 3] and [19. “Mm, Ayu +--+ Oily 2 Cj (= Gj if x7 > 0) [19.13] ayxt to-++ Gynt, S bj (= by if uj > 0) 19.14] x¢ and u* have all their components nonnegative and satisfy 14), then x* and u* solve the primal problem [19.4] and the {195}. respectively. Proof Suppose x° solves [19.4] and u* solves [19.5]. Then. in particular (see (19.8) and [19.9). Ax*c 103) Multiplying the first inequality in {1 on the left by (u*)’ > 0 and the second inequality on the right by x* > 0 yields (atYAxt s(a")'b and (u")Ax > cx [2] ‘According to Theorem 19.3, (u*)’b = c’x*. So both inequalities in (2] must be equalities. They can be written as (oy(ax"—by)=0 and ((a"'A— ex" = 0 GB) But these two equations are equivalent to Dogar +--+ ajexn — 4) =0 4) jut and n Denny +--+ amin — ex7 = 0 6) t= For j = 1.....m one has both us > O and aj + +--+ GjnXq — bj <0. SO cach term ia the sum [4] is <0. ‘The sum of all the terms is 0, so there can be no negative term, because there is no positive term to cancel it. Hence, eachee 742 Chapter 19 / Linear Programming \y Note: Using the general economic interpretations we ga tions [19.13] and [19.14] can be interpreted as follows: If the shadow price of resource j is positive ( of resource j must be used in any optimum LP Problems If the solution to either the primal or the dual) plementary slackness conditions can help find the determining which constraints are slack. Let us look at Example 19.5 term in the sum (4] must be 0. Therefore. ott Nw arlajxy +++ aynry — 8) =0 Now [19.14] follows immediately. Property (19.13] is pro by noting how (5] implies that 3 LN flout + Femah—ed =O Gat Suppose on the other hand that x* and u* have all nonnegative and satisfy [19.13] and [19.14]. It follows im and (7) are satisfied. So summing over j and i. respecti 5]. From these equ: it follows that 07 bj equal. te bay yd Consider the problem max 3x1 + 4x2 + 6x5 subject to {4 + with x1 > 0, x2 2 0, and x3 > 0. rhe Sova by a geometric argument. Then use complementat Solution The dual problem is Buy + up min 2u; +u2 subject of uy + 2u uy + 6u2Sec. 19.5 / Complementary Stackness 743 » \ { Using the geometric solution technique shown in Example 19.2, we find the ( solution j= 2/5, and ug = 9/5. Then 3uj + u3 = 3. uj + 2u3 = 4, and ult 6u5 > 6. —— —— ee Y What do we know about the solution (xj, x3,.x3) to [1]? According to [19.14], because uj > 0 and u§ > 0, both inequalities in (1] are satisfied with equality. So 3xptagtagy=2 0 and xp +27 + 6x3 =1 GB) —_—__ pees Now x$ cannot be > 0, otherwise [19.13] would imply u; +6u5 = 6. Hence, x} = 0. Letting x3 = 0 in (3] and solving for xj and x3 gives y=3/, =, =O This is the solution to problem [1]. Note that the optimal values of the criterion functions in the two problems are indeed equal: 2uy +u3 = 13/5 \ and 3x1 + 4x} + 6x5 = 13/5, just as they should be according to the duality a theorem. : i >) The Kuhn—Tucker Theorem Applied to Linear Programs Consider the general linear programming problem aK yx) $+ +Ginkn SO ( max ¢)X)+7°°+Cnkn SL { a x) 20, .%,%1 20 Omi X1 + °+° + Omnkn S Om 0 ‘This problem is obviously a special case of the general nonlinear programming problem max f (x1, +260%8) St { 2h Bn (X1 which was studied in Section 18.9. Note that the functions f and gy in (1] are all linear, so are concave as well as convex. Thus, the Kuhn-Tucker conditions {18.42} and (18.43) are sufficient for optimality of a vector x = (x1, £2.~--.%») satistying the constraints specified in [1]. Let us see what form these conditions take in the linear case. If we let 4; = uy for j = 1..--.m, conditions [18.42] and [18.43] become Go —G@my toe + Onin) $0 (H0ifx > 0) GH.) BI .+m) (4) up 20 (SOif ary +--+ aja, < bj)Bn + 744 Chapter 19 / Linear Programming \ \ No aany to ante £8 and \) = a11x1 — When combined with the requirement that x satisfy the constraints ce these conditions are precisely the complementary slackness conditio bem rem 19.4. ne in a Duality When Some Constraints are Equalities Suppose that one of the constraints in the primal problem is the equality Qj1X) +--+ + Ginkn = Bj t] ‘Then we can replace it by the two inequalities —Ainkn Sb) fag) in order to put the problem into standard form. Constraint [+] thus gives rise two dual variables u and u”. In the matrix describing the dual constraints, the Se columns associated with uj and uj’ are equal except for opposite signs everywhere. Therefore, we can replace the two variables ui and u/’ with u; =u! — ul) but then there is no restriction on the sign of u;. We see that if the ith constrain in the primal is an equality, then the ith dual variable has an unrestricted sign. This is consistent with the economic interpretation we have given. If we are forced to use all of resource i, then it is not surprising that the resource may command a negative shadow price—it may be something that is harmful in excess. For instance, if the baker of Example 19.1 was forced to include all the stock of sugar in the cakes, the best point in Fig. 19.2 would be B, not C. Some profit would be lost. From the symmetry between the primal and the dual, we-sealize now that if riables in the primal-has.an unrestricted sign..then.the corresponding one of the constraint in the dual is an equality. _—-—_— or ee Problems 1. Consider Problem 1 of Section 19.3. The optimal solution of the primal was: x* =0, and y* = 3, whereas uj = 0, was the optimal solution of the dual. Verify that [19.13] and [19.14] are satisfied in this case. 2. a. Solve the following problem geometrically: nt+o2 eel nt we 5 20, 220 min y: +2y2 subjectto y 9 PS 5 nz yi — 2y2 = =20 b. Write down the dual problem and solve it. 5 sraint yi +692 2 | . What happens to the optimal dual variables if the cons is changed to y1 + 6y2 > 15.1? ——Sec. 19.5 / Complementary Slackness 745 3. A firm produces two commodities A and B. The firm has three factories that jointly produce both commodities in the amounts per hour given in the following table: Factory 1 Factory 2 Factory 3 Commodity A 10 20 20 Commodity B 20 10 20 The firm receives an order for 300 units of A and 500 units of B. The cost per hour of running factories 1, 2, and 3 are respectively 10,000: 8000; and 11,000. a. Let yj, yp, and y3, respectively, denote the number of hours for which the three factories are used. Write down the linear programming problem of minimizing the costs of fulfilling the order. b. Show that the dual of the problem in pant (a) is 10x; + 20x. < 10.000 max 300x; +500x2 s.t.¢ 20x; + 10x. < 8000 x 20.220 20x; + 20x2 < 11.000 Solve this problem and then find the solution of the problem in part (a). c. By how much will the minimum cost of production increase if the cost per hour in factory 1 increases by 100? Harder Problems 4. The following problem has arisen in production theory: glx! ef? +--+ gaY 0. Here &/, Vj. V2, ¥!..... =" are fixed constants. (Superscripts i = 1. ..., N refer to production units; V and Vs are available quantities of two resources; each ¥/ is a capacity constraint on the output of unit j.) Write down the dual to prob- Jem [1], with q, and q2 denoting the dual prices associated with the first two constraints in (1), whereas r', r?, ..., r" are the dual variables associated with the last N constraints. Let the values of the variables that solve the746 Chapter 19 / Linear Programming two problems be those indicated with a hat, ‘optimum, then } 0 = Hate tr = gigy + de tr! > 1 ext = Gy > OR + HEP R™ HEPRN < Vi => GY Batten F>0 => # 5. Consider the LP problem xy + x2 max 3x;+2x2 subject to 4 2x) + x2 — 3 IA IA IA, x) + 2x. — 2x3 ‘a. Suppose x3 is a fixed number. Solve the b. Solve the problem for any fixed value of value of 3x; + 2x2 becomes a function of x3. maximize it. ca ¢. Do the results in part (b) say anything about 1 problem in which x3 can also be chosen?