Optimal procurement when suppliers are divided into groups and the buyer is constrained to buy from all the groups P. Samuel NJIKI Université de Montréal and CIREQ May 2008 Abstract We consider an agency willing to achieve a project. For example the project can be the building of equipments or a research project. The project is divided into smaller projects and contracts for the completion of these projects are granted through a procurement auction. The participants are agents coming from di¤erent groups. We assume that the agency is committed to allocate the contracts so that agents from all groups win. We study the problem of optimally choosing a procurement mechanism to complete the project. Sellers have private information on their cost and their payo¤s are linear. We derive a mechanism that minimizes the expected price paid by the buyer when he is committed to "all the groups". Under this optimal mechanism the buyer allocates the suprojects to a set of agents coming from all possible groups and having the smallest aggregate virtual cost (the sum of the ironed out virtual costs of agents in the set) among all sets of agents who may win the project. Keywords: Optimal mechanism, Procurement, (Aggregate) Virtual costs, Iron out, Compatible suppliers, Groups. JEL classi…cation: C72, D44, D82. 1 Introduction A buyer willing to procure some units of heterogeneous items will often seek to minimize the total price of the desired quantity. Beside this objective, a buyer facing potential suppliers from di¤erent groups may sometimes want to procure the items in such a way that all the groups are represented by the actual suppliers. For example, consider a research project commanded by a government agency. The agency may divide ps.njiki.njiki@umontreal.ca 1 the project into subprojects and award the contracts related to these subprojects through a procurement auction. Assume that participants are researchers a¢ liated to institutions in di¤erent provinces of the country. A government willing to encourage the research in all the provinces may design a procurement mechanism so that winners come from all the provinces. Another example is that of an international institution seeking to hire a given number of individuals among job candidates coming from di¤erent countries. If the institution is created and/or …nanced by these countries. Then it is natural for the institution to hire from all these countries. In this paper we derive a procurement mechanism that minimizes the expected price of the units when the buyer is constrained to buy the units from actual suppliers coming from all the groups he faces. We also derive a mechanism that minimizes the expected price of the units when the buyer wants to procure each item independently of the others. Then, assuming that the buyer can implement these mechanisms, we compare the expected prices of the units under the two mechanisms. The literature on optimal auctions has ‡ourished since the early 1980’s. Economists have designed various revenue maximizing (or cost minimizing, in the case of procurement auctions) mechanisms in the contexts of single unit auctions (ex. Myerson, 1981), multi-unit auctions (ex. Branco, 1996) and multi-object auctions (ex. Armstrong, 2000). To the best of our knowledge, the available literature does not consider design problems where potential buyers (suppliers in the case of procurement auctions) are not always compatible; that is while every potential supplier may win a priori, only some subsets of the potential buyers (or suppliers) may win simultaneously. This aspect of the problem is considered in the present paper using the approach laid in (Myerson, 1981). The notion of virtual value appears to be crucial in the literature on optimal auction. The equivalent notion in procurement auction is that of virtual costs. We …nd that virtual costs are also a key instrument in the optimal mechanism under the constraint of the groups. Virtual costs are summed over every potential winning subset of suppliers and the aggregate virtual costs are used to order these subsets; contracts are then awarded to suppliers of one of the maximal subsets. The model we adopt is the same as in (Armantier and Njiki, 2008). Instead of searching for optimal auctions (Armantier and Njiki, 2008) are interested in particular auctions. On the one hand they consider two speci…c procurement auctions where contracts are awarded so that all the groups are represented by the winners of the auction, and on the other hand they consider traditional procurement auctions where items are procured independently through second price auctions and …rst price 2 auctions. The expected prices of the units under Nash equilibria of the games induced by these auctions are then compared. They identi…ed a mechanism under constraint of the groups that yields a lower expected price than two independent second price auctions, under assumptions that imply costs correlation or asymmetry. Due to the assumption that costs are independent in our model, we …nd that the optimal independent mechanism yields a lower expected price than the optimal mechanism under the constraint of the groups. Nevertheless we provide conditions under which the two prices are equal. We use a simple model with a single buyer willing to buy two heterogeneous items (a unit of each), and four potential suppliers coming from two di¤erent groups. There each supplier can only supply one item and for each item there are two potential suppliers from di¤erent groups. Suppliers are supposed to have private information about their supply costs and these costs are independent. Our results however generalize when there is more than two groups or items. The abstract approach used here to solve the problem can be applied to more complex models. The rest of the paper is organized as follows. In section 2 we present the model and introduce some useful de…nitions. In section 3 we clarify what mechanisms are considered feasible and focus on direct mechanisms. The optimal mechanisms are derived in section 4. Examples and applications are considered in section 5. Then we conclude in section 6. Notation 1 (Note) Scalars and scalar functions are denoted by lowercase letters. Vectors and vector functions are denoted by boldface lowercase letters. u i denotes the vector u without the component of order i: u = (ui ; u i ). Finally conditional expectations are denoted by the uppercase of the same letter e.g.: X(ci ) = Ec i [x(ci ; c i )]. 2 The model We consider two suppliers (1 and 2) from a group l and two suppliers (3 and 4) from a group h. Suppliers 1 and 3 produce the same item and suppliers 2 and 4 produce another item. A buyer would like to buy a unit of each of these two items through a procurement auction. If supplier i 2 N = f1; 2; 3; 4g wins the procurement it will cost ci to provide one unit. Costs are independently distributed and supplier i’s cost is distributed according to the probability density function fi ; and the cumulative density function Fi with support i = [ci ; ci ]. Denote = ¯ i the Cartesian product of these supports. The costs’distributions i2N and the groups of the suppliers are common knowledge to the buyer and the suppliers. Every supplier observes privately his own cost but not the 3 other suppliers’ costs. There is no cooperation between the suppliers. Note that in this model suppliers win the procurement in pairs: the buyer can only buy from one of the following pairs f1; 2g ; f2; 3g ; f3; 4g or f4; 1g :1 There are many di¤erent mechanisms for procuring the two units. For example, the buyer can buy the units through two simultaneous …rst price auctions: suppliers are asked to bid the price they are willing to accept, and the buyer buys the …rst unit from the supplier with the lowest bid among suppliers 1 and 3 at a price equal to the lowest bid, and he buys the second unit from the supplier with the lowest bid among suppliers 2 and 4 at a price equal to this lowest bid: He can also buy each unit through a second price auction (for each unit the potential suppliers submit their bids, and the item is bought from the supplier with the lowest bid at a price equal to the second lowest bid). The buyer can even practice di¤erent kind of auctions on each item. In the mechanisms considered so far the two items are bought independently. The buyer’s decision concerning the supplier to whom he purchases the …rst item and the …rst item’s unit price is independent on his decision concerning the supplier to whom he purchases the second item and the second item’s unit price. In that sense, the procurement of one unit does not dependent on the procurement of the other unit. There exist mechanisms under which the procurements of the two units are not independent. For example, consider the following rule: "the buyer must buy the units from suppliers coming from all the groups". We shall refer to this rule as the rule (R). Under this rule the buyer can buy only from the pairs f1; 4g and f2; 3g:2 Thus if he buys the …rst item from supplier 1 he must buy the second unit from supplier 4: Consider for example the auction format where supplier i submits a bid bi and the units are bought from the pair f1; 4g if b1 + b4 < b2 + b3 ; and from the pair f2; 3g if b1 + b4 > b2 + b3 . In any case the winners are paid a price equal to their bid but the other suppliers receive nothing. This auction format is studied (among others) in (Armantier and Njiki, 2008) in order to model procurement at the European space agency. We introduce some basic de…nitions in the rest of this section. Let P = ff1; 2g ; f2; 3g ; f3; 4g ; f4; 1gg : We already mentioned that the pair of suppliers who win the procurement is necessarily one of the pairs in P. In a typical procurement mechanism game, the buyer …rst announces 1 The pair f1; 3g is excluded because 1 and 3 sell the same good and the buyer needs only a single unit of it. f2; 4g is excluded for the same reason. 2 The pair f1; 2g is excluded because 1 and 2 come from the same country though they sell di¤erent goods. f3; 4g is excluded for the same reason. 4 the procurement rules to the suppliers. Suppliers observe privately their own costs and place their bids. Finally the buyer collects the bids and buys the items according to the rules set before. Let i be the set of all possible bids for supplier i 2 N ; and let 4 Q = is determined by the type of information that i : The set i=1 the buyer requires to each supplier during the procurement process. So we may assume that the buyer knows the set of the information he requires. Therefore suppliers cannot bid out of that set. A strategy for supplier i is a function i : i ! i transforming i’s cost into a bid. An allocation rule is a function q = (qG )G2P : !R4+ such that: X for any bid vector b 2 ; qG (b) = 1: G2P In other words q(b) is a probability distribution on P and qG (b) is the probability that pair G wins. The individual allocation rule associated with the allocation rule q is a function x = (xi )i2N : !R4+ such that: for any supplier i 2 N and any bid vector b 2 ; xi (b) = X qG (b): G2P=i2G xi (b) is supplier i’s winning probability with a bid vector b.3 Note that this de…nition implies that x1 (b) + x3 (b) = 1 = x2 (b) + x4 (b): A payment rule is a function t = (ti )i2N : !R4 such that ti (b) is the amount of money paid to supplier i when the bid vector is b. Note that this payment can be negative, meaning that supplier i will have to make a transfer to the buyer rather than receive from him. A procurement mechanism is de…ned by a set of bids, an allocation rule and a payment rule. We use the notation ( ; q; t) to refer to a procurement mechanism with a set of bids , an allocation rule q and a payment rule t: It is important to note that we do not mention the individual allocation rule in the de…nition of a mechanism since it is uniquely determined by the allocation rule. On the contrary a given individual allocation rule might be associated with more than one allocation rule.4 What is usually referred to as the allocation rule in auction literature 3 For example x1 (b) = qf1;2g (b) + qf4;1g (b): Indeed let x = (xi )i2N : !R4+ be such that x1 (b)+x3 (b) = 1 = x2 (b)+x4 (b): Consider a function : ![0; 1] such that max(0; x3 + x4 1) min(x3 ; x4 ); exists because x3 ; x4 2 [0; 1] ) 0 max(0; x3 + x4 1) min(x3 ; x4 ) 1: Take for example qf3;4g = ; qf4;1g = x4 ; qf3;2g = x3 ; qf1;2g = 1 x3 x4 + ; 4 5 actually corresponds to our individual allocation rule. The reason for this di¤erence is that our de…nition is designed to grasp the possibility of an additional constraint on the compatibility of the suppliers. The rule (R) is an example of such constraints. In this case suppliers from the same group are not compatible in the sense that they cannot win simultaneously. A procurement mechanism ( ; q; t) is independent if there exists two 4 4 functions : 1 : 2 3 ! R and 4 ! R such that, for any b2 ; (i) (x1 (b); x3 (b); t1 (b); t3 (b)) = (b1 ; b3 ); (ii) (x2 (b); x4 (b); t2 (b); t4 (b)) = (b2 ; b4 ): To understand this de…nition, remember that suppliers 1 and 3 are selling the same item as well as 2 and 4; so the conditions (allocation and payment) under which one item is purchased depend solely on the message sent by the suppliers of that item and not the message sent by the suppliers of the other item. An allocation rule q follows the rule (R) if: for any b 2 ; qf1;4g (b) + qf2;3g (b) = 1: In that case if x is the associated individual allocation rule, for all b2 ; x1 (b) = x4 (b) = qf1;4g (b) and x2 (b) = x3 (b) = qf2;3g (b): A procurement mechanism ( ; q; t) follows the rule (R) if q follows the rule (R). Our concern in the next section is to de…ne the set of all possible mechanisms. 3 Direct mechanisms The set of bid vectors can be a complex object, depending on the information the buyer requires from the suppliers. This makes di¢ cult the problem of optimally choosing a mechanism. Direct mechanisms are a particular class of mechanisms where each supplier is asked to directly X It is easy to see that qG = 1 and for any G 2 P; qG 0: Thus q is an allocation rule. Moreover the individual allocation rule associated with q is precisely x; but the allocation rule q depends on the selection . 6 report his cost. Formally a direct mechanism is a mechanism where the set of bid vectors is : When ( ; q; t) is a direct mechanism we shall simply denote it (q; t): 3.1 The revelation principle A procurement mechanism induces a game of incomplete information between the suppliers, and the notion of direct mechanism has been de…ned in the broader context of games with incomplete information. In such games players observe privately an information considered as their types, they send a message and resources are allocated on the basis of the messages sent and prede…ned rules. The search for an optimal mechanism can be simpli…ed if one can restrict attention to direct mechanisms. The well known revelation principle allows us to make such a restriction without loss of generality. This principle states that: given a game of incomplete information and a Bayesian Nash equilibrium (BNE), there exist a direct mechanism (with the same outcomes as the …rst game) for which it is a BNE to report honestly the types. For the interested reader we provide (in appendix) a version of the proof of the revelation principle in the framework we have set earlier. Proposition 2 Revelation principle Given a mechanism ( ; q; t) and an equilibrium for that mechanism , there exists a direct mechanism (q; t) in which it is an equilibrium for each supplier to report honestly his cost and the outcomes are the same as in the equilibrium of the …rst mechanism. 3.2 Incentive compatible and individually rational direct mechanisms Suppliers need not report their true costs in a direct mechanism since this information is private; if the buyer cares about truth, he must choose a procurement mechanism that gives them incentives to do so. This condition imposes further restrictions on mechanisms that may be chosen: a procurement mechanism must be incentive compatible and individually rational. Before we de…ne these two concepts we need to introduce some more notations. Consider a direct mechanism (q; t) with an individual allocation rule x: Let Z Xi (mi ) = xi (mi ; c i )f i (c i )dc i 7 i (1) be the (interim) winning probability of supplier i if he reports the value mi given that the other suppliers report their true costs. And let Z Ti (mi ) = ti (mi ; c i )f i (c i )dc (2) i i be the (interim) expected payment received by supplier i if he reports the value mi given that the other suppliers report their costs honestly. The (interim) expected pro…t of supplier i when he reports mi (rather than ci ) and the other suppliers report their true costs is: i (mi ; ci ) = Ec i [ti (mi ; c i ) xi (mi ; c i )ci ] = Ti (mi ) Xi (mi )ci : (3) In particular: i (ci ; ci ) Ti (ci ) (4) Xi (ci )ci : is the expected pro…t when supplier i reports his true cost ci . If honesty (reporting the true cost) is an equilibrium then i (ci ; ci ) is supplier i’s pro…t at equilibrium: Xi (ci ) and Ti (ci ) are respectively the winning probability and the expected payment of supplier i at equilibrium. i (ci ; ci ) A mechanism is individually rational (IR) if: for any i 2 N and ci 2 i; i (ci ; ci ) 0: (5) This means that even the suppliers with the worst costs will make non negative pro…ts if they participate in the procurement honestly when all the other players do so. A mechanism is incentive compatible (IC) if: for all i 2 N and ci 2 i; i (ci ; ci ) = max i (mi ; ci ) = fTi (mi ) Xi (mi )ci g: mi 2 i (6) This means that reporting honestly his cost give a supplier the highest expected pro…t when the other suppliers report their true costs. In other words the mechanism (q; t) is incentive compatible if honesty is an interim Bayesian Nash Equilibrium (BNE) of the game induced by (q; t): The next proposition characterizes an IC mechanism by the winning probabilities and the expected payment functions. 8 Proposition 3 A mechanism (q; t) is IC if and only if: for all i 2 N ; the function Xi is decreasing (7) and, for all ci 2 i ; Ti (ci ) = Ti (ci ) Xi (ci )ci + Xi (ci )ci + Z ci Xi (ti )dti : (8) ci Thus, when honesty is a BNE, suppliers with the lowest costs have the highest interim winning probabilities. And these winning probabilities determine the expected payments up to a constant. equation (8) is well known in the literature on auction design as the revenue equivalence theorem. A proof is available in appendix. In the next section we suppose the buyer has the choice of the procurement mechanism. We look for the mechanism he will choose depending on whether he is looking for an independent mechanism or a mechanism that follows the rule (R). We also assume that the buyer is motivated by the minimization of the expected price of the two units. 4 Designing the optimal mechanisms In this section we …nd mechanisms that minimize the expected amount paid by the buyer (i.e. the expected price of the two units) among IC and IR direct mechanisms. We do this by restricting our search to mechanisms satisfying the rule (R) and then to independent mechanisms. Consider an incentive compatible direct mechanism (q; t) with an individual allocation rule x: Below is an expression of the expected price P (q; t) of the units under this mechanism. X X X X P (q; t) = E ti (c) = E[ti (c)] = EEc i [ti (ci ; c i )] = ETi (ci ) i2N i2N i2N i2N using (8), E[Ti (ci )] = Ti (ci ) + Z ci c ¯i Xi (ci )ci + Z Z ci Xi (ci )ci fi (ci )dci c ¯i ci Xi (ti )dti fi (ci )dci ci Fubini’s theorem implies, 9 Z ci c ¯i Z ci Xi (ti )dti fi (ci )dci = ci = Z Z ci c Z¯ ici ti fi (ci )dci Xi (ti )dti c ¯i Fi (ti )Xi (ti )dti : c ¯i Therefore, E[Ti (ci )] = Ti (ci ) Xi (ci )ci + Z Z ci [Xi (ci ) ci fi (ci ) + Fi (ci )Xi (ci )]dci c ¯i ci Fi (ci ) )Xi (ci ) fi (ci )dci fi (ci ) ci Z¯ F (ci ) )xi (c) f (c)dc (using (1)). = i (ci ; ci ) + (ci + i fi (ci ) = i (ci ; ci ) + (ci + It follows that, P (q; t) = X i (ci ; ci ) + Z (X ) i2N i2N Fi (m) fi (m) (9) Hi (ci )xi (c) f (c)dc where Hi (m) = m + for all i 2 N and m 2 i ; the function Hi is usually called the virtual cost of supplier i: We can also express the expected price in terms of the allocation rule rather than the individual allocation rule: 8 Z <X 9 = Hi (ci ) qG (c) f (c)dc P (q; t) = i (ci ; ci ) + : ; i2N i2N G2P=i2G 8 9 Z < = X X X = Hi (ci )qG (c) f (c)dc i (ci ; ci ) + : ; i2N i2N G2P=i2G ) Z (X X X = Hi (ci )qG (c) f (c)dc i (ci ; ci ) + X i2N = X X G2P i2G i (ci ; ci ) i2N + Z (X qG (c) G2P X i2G ) Hi (ci ) f (c)dc: We de…ne the aggregate virtual cost of the pair G as SG (c) P Hi (ci ): i2G And the price can be written as: P (q; t) = X i2N i (ci ; ci ) + Z (X G2P 10 ) qG (c)SG (c) f (c)dc: (10) For any i 2 N , consider the function Ki : [0; 1] ! R; Ki (zi ) = 1 (zi ) Hi (t)fi (t)dt: Note that Ki0 (Fi (ci )) = Hi (ci ): 0 ^ i : [0; 1] ! R be the convex hull5 of Ki : K ^ i is di¤erentiable Let K ^ ^ ^ 0 (Fi (ci )): almost surely. Let Hi : i ! R be such that Hi (ci ) = K i Then it is known that: ^ i (0) = Ki (0) and K ^ i (1) = Ki (1); (a) K ^ (b) Ki (zi ) Ki (zi ) for any zi 2 [0; 1] ; ^ i (zi ) < Ki (zi ) then K ^ i0 is constant in some neighborhood of (c) if K ^ zi ; hence Hi is constant in some neighborhood of Fi 1 (zi ). ^ i (ci ) is called the ironed out virtual cost of supplier i: S^G (c) PH ^ i (ci ) will be called the aggregate ironed out virtual cost. H R Fi i2G Let P^ (q; t) = X i (ci ; ci ) + i2N Z (X ) ^ i (ci )xi (c) f (c)dc: H i2N (11) P^ is obtained from the expression of P by replacing virtual costs by the ironed out virtual costs. Thus we can also write: ) Z (X X qG (c)S^G (c) f (c)dc: (12) P^ (q; t) = i (ci ; ci ) + G2P i2N The following lemma will also be useful. Lemma 4 for any IC mechanism (q; t): P^ (q; t) Proof. P^ (q; t) P (q; t) = Z (X ) ^ i (ci )xi (c) f (c)dc H i2N XZ n ^ i (ci ) = H i2N = XZ ci i2N ci i2N ci i2N c ¯i n ^ i (ci ) H P (q; t): Z (X i2N o Hi (ci ) xi (c)f (c)dc; ¯ X Z ci n ^ i Fi )0 (ci ) = (K o ^ i ), Ki0 (Fi (ci )) Xi (ci )f (ci )dci (by de…nition of Hi and H o (Ki Fi )0 (ci ) Xi (ci )dci . i.e. the greatest convex function g : [0; 1] ! R such that g(zi ) zi 2 [0; 1] : 11 Hi (ci )xi (c) f (c)dc; o Hi (ci ) Xi (ci )f (ci )dci (using (1)), ¯ X Z ci n ^ 0 (Fi (ci )) = K i 5 ) Ki (zi ); for any Integrating by part, Xn ^ i Fi )(ci ) P^ (q; t) P (q; t) = (K i2N Z ci ci P^ (q; t) X ¯n ^ i (1) P (q; t) = K i2N Z ci c ¯i P^ (q; t) n ^i (K P (q; t) = n ^ i Fi )(ci ) (K XZ i2N o ^ i Fi )(ci ) + (Ki Fi )(ci ) (K ¯ ¯ o Fi )(ci ) (Ki Fi )(ci ) dXi (ci ); o ^ i (0) + Ki (0) Ki (1) K ci c ¯i (Ki Fi )(ci ) n ^ i (Fi (ci )) K o (Ki Fi )(ci ) dXi (ci ); o Ki (Fi (ci )) dXi (ci ) (using (a)). Since the mechanism is IC, proposition 3 implies that dXi (ci ) 0. ^ Furthermore property (b) implies Ki (Fi (ci )) Ki (Fi (ci )) 0. It follows that P^ (q; t) P (q; t) 0: We now derive the optimal mechanism under the rule (R). 4.1 Optimal mechanism under the rule (R) Consider the allocation rule qR de…ned as follow: 8 R > < qf1;4g (m) = 1 if S^f1;4g (m) S^f2;3g (m); R for any m 2 ; qf2;3g (m) = 1 if S^f1;4g (m) > S^f2;3g (m); > : q (m) = 0 if G 2 ff1; 2g ; f3; 4gg: G R R Note that qf1;4g (m) + qf2;3g (m) = 1 for any m 2 . Under this allocation rule the buyer buys from the pair with the smallest aggregate ironed out virtual cost between the pairs f1; 4g and f2; 3g. Thus qR follows the rule (R). Let xR be the individual allocation rule associated with qR ; xR is such that, for any m 2 , R R R R R xR 1 (m) = x4 (m) = qf1;4g (m) and x2 (m) = x3 (m) = qf2;3g (m): (14) Finally consider the following payment rule: Z ci R R for any m 2 ; ti (m) = xi (m)mi + xR i (t; m i )dt: (15) mi R Note that we have in particular tR i (ci ; m i ) = xi (ci ; m i )ci for all m i 2 i . Then taking the expectation TiR (ci ) = XiR (ci )ci and …nally 12 (13) the expected pro…t of supplier i with costs ci when he bids honestly is R R XiR (ci )ci = 0: The consequence is that the mechai (ci ; ci ) = Ti (ci ) nism (qR ; tR ) always yields a worse expected pro…t than any individually rational mechanism (q ; t) for the supplier i with costs ci when he bids honestly: i (ci ; ci ) 0 = R i (ci ; ci ): And we may prove the following result. Proposition 5 The direct mechanism (qR ; tR ) minimizes the expected price of the two units among all direct mechanisms that are IC, IR and follow the rule (R). Moreover the minimum expected price is given by R = E[min(S^f1;4g (c); S^f2;3g (c))]: Proof. If (q; t) is IC and follows the rule (R) then for any c 2 ; qf1;4g (c)+ qf2;3g (c) = 1 and qG (c) = 0 if G 2 ff1; 2g ; f3; 4gg ; Then we have Z (X G2P ) Z n o qG (c)S^G (c) f (c)dc = qf1;4g (c)S^f1;4g (c) + qf2;3g (c)S^f2;3g (c) f (c)dc Z = Z Z qf1;4g (c) min(S^f1;4g (c); S^f2;3g (c)) f (c)dc +qf2;3g (c) min(S^f1;4g (c); S^f2;3g (c)) qf1;4g (c) + qf2;3g (c) min(S^f1;4g (c); S^f2;3g (c))f (c)dc R R qf1;4g (c) + qf2;3g (c) min(S^f1;4g (c); S^f2;3g (c))f (c)dc Z n o R R ^ ^ = qf1;4g (c)Sf1;4g (c) + qf2;3g (c)Sf2;3g (c) f (c)dc ) Z (X R = qG (c)S^G (c) f (c)dc: = G2P If (q; t) is also IR we just showed that P i2N i (ci ; ci ) 0= P R i (ci ; ci ): i2N We conclude that P^ (q; t) P^ (qR ; tR ) for any IC and IR mechanism (q; t) following the rule (R). Putting this with lemma 4 we obtain that, for any IC and IR mechanism (q; t) following the rule (R) P (q; t) P^ (qR ; tR ): It will be su¢ cient to show that P^ (qR ; tR ) = P (qR ; tR ) and that (qR ; tR ) is an IC and IR mechanism following the rule (R). R By de…nition tR xR i (m) i (m)mi and (integrating over m i ) Ti (mi ) XiR (mi )mi ; i.e. R 0: So (qR ; tR ) is IR. i (mi ) R R We show that (q ; t ) is IC using The two conditions R ci ofRproposition R R R 3. First taking the de…nition of t : ti (c) = xi (c)ci + ci xi (t; c i )dt; 13 integrating this equality over c i leads to TiR (ci ) = XiR (ci )ci + R ci and R R Xi (t; c i )dt and since R XiR (ci )ci = 0; TiR (ci ) = i (ci ; ci ) = Ti (ci ) ci R c TiR (ci ) XiR (ci )ci + XiR (ci )ci + cii XiR (t; c i )dt: We will show that XiR is decreasing when i = 1; the proof is similar for any i 2 N .6 ^ 1 is increasing we have H ^ 1 (s) H ^ 1 (t) Let s; t 2 1 : s < t; since H ^ 1 (s)+ H ^ 4 (c4 ) H ^ 1 (t)+ H ^ 4 (c4 ): Therefore the following implication and H is true: ^ 1 (t) + H ^ 4 (c4 ) H ^ 2 (c2 ) + H ^ 3 (c3 ) ) H ^ 1 (s) + H ^ 4 (c4 ) H ^ 2 (c2 ) + H ^ 3 (c3 ): H In other words: S^f1;4g (t; c 1 ) S^f2;3g (t; c 1 ) ) S^f1;4g (s; c 1 ) S^f2;3g (s; c 1 ); or R R qf1;4g (t; c 1 ) = 1 ) qf1;4g (s; c 1 ) = 1: R R So qf1;4g (s; c 1 ) qf1;4g (t; c 1 ); and …nally (using 14): xR 1 (s; c 1 ) R x1 (t; c 1 ): Since c 1 is arbitrary we may take the integral over c 1 and obtain X1R (s) X1R (t): Thus X1R is decreasing. Since (qR ; tR ) is IC then, by equation 13: o X Z ci n R R R R ^ ^ Ki (Fi (ci )) Ki (Fi (ci )) dXiR (ci ); P (q ; t ) P (q ; t ) = i2N c ¯i ^ i (Fi (ci )) Ki (Fi (ci )) 6= 0 then K ^ i (Fi (ci )) Ki (Fi (ci )) < 0 and If K ^ i (ci ) is constant in some neighborhood of ci (property (c)) and so, H XiR (ci ) is also constant7 (i.e. dXiR (ci ) = 0) in o some neighborhood R ci n ^ i (Fi (ci ) Ki (Fi (ci ) dXiR (ci ) = 0 and of ci : We then conclude c K i ¯ P^ (qR ; tR ) P (qR ; tR ) = 0: 6 The large inequalities in the implications below must be replaced by strict inequalities if i = 2 or i = 3: This is necessary because of the de…nition of qR where ties are solved in favor of the pair f1; 4g : 7 Without loss of generality suppose i = 1: ^ ^ ^ 2 (c2 ) + H ^ 3 (c3 ). By Recall that xR H 1 (c1 ; c 1 ) = 1R if and only if H1 (c1 ) + H4 (c4 ) R R de…nition we have X1 (c1 ) = 3 x1 (c1 ; c 1 )f 1 (c 1 )dc 1 . Since the integral is taken ^ i )i6=1 and on H ^ 1 (c1 ). Thus it is over c 1 we conclude X1R (c1 ) depends only on (H ^ constant if H1 (c1 ) is constant. 14 Finally the expected price is: Z n o X R R R R R ^ ^ ^ qf1;4g (c)Sf1;4g (c) + qf2;3g (c)Sf2;3g (c) f (c)dc = P (q ; t ) = i (ci ; ci ) + i2N Z n o R R ^ ^ = qf1;4g (c)Sf1;4g (c) + qf2;3g (c)Sf2;3g (c) f (c)dc (because i (ci ; ci ) = 0) Z n o R R (c) min(S^f1;4g (c); S^f2;3g (c)) f (c)dc (c) min(S^f1;4g (c); S^f2;3g (c)) + qf2;3g = qf1;4g h i R ^ ^ = E min(Sf1;4g (c); Sf2;3g (c)) . 4.2 Optimal independent mechanism For any m 2 let xI1 (m) = ^ 1 (m1 ) H ^ 3 (m3 ) 1 if H ^ 1 (m1 ) > H ^ 3 (m3 ) ; 0 if H (16) xI2 (m) = ^ 2 (m2 ) H ^ 4 (m4 ) 1 if H ^ 2 (m2 ) > H ^ 4 (m4 ) ; 0 if H (17) xI3 (m) = 1 xI1 (m) and xI4 (m) = 1 and tIi (m) = xIi (m)mi + Z xI2 (m); (18) ci xIi (t; m i )dt: (19) mi We showed in section 2 that there exist at least one allocation rule I )G2P such that the vector function xI = (xIi )i2N is the inq = (qG dividual allocation rule associated with qI :8 The mechanism (q I ; tI ) is independent since xI3 ; xI1 ; tI3 ; tI1 depend only on m1 or m3 ; and xI2 ; xI4 ; tI2 ; tI4 depend only on m2 or m4 : We see that in the sense of (Myerson 1981) (xI3 ; xI1 ; tI3 ; tI1 ) is a procurement mechanism for the …rst item. Likewise (xI2 ; xI4 ; tI2 ; tI4 ) is a procurement mechanism for the second item. By de…nition any independent mechanism is actually the sum of two independent mechanisms for each item. Moreover it is known that (xI3 ; xI1 ; tI3 ; tI1 ) I (resp. (xI2 ; xI4 ; tI2 ; tI4 )) is the optimal procurement mechanism of the …rst (resp. second) item. Because the expected price P (q; t) is an additively separable function of (x3 ; x1 ; t3 ; t1 ) and (x2 ; x4 ; t2 ; t4 ), (q I ; tI ) is optimal among independent mechanisms. Our next result is that this mechanism actually minimizes the expected price of the items among all IC and IR mechanisms, including those satisfying the rule (R). 8 We …x qI once for all. 15 Proposition 6 The direct mechanism (qI ; tI ) minimizes the expected price of the two units among all IC and IR mechanisms. Moreover ^ 1 (c1 ); H ^ 3 (c3 )) + the minimum expected price is given by I = E[min(H ^ 2 (c2 ); H ^ 4 (c4 ))]: min(H The proof is similar to the proof of proposition 5 and is found in the appendix. A consequence of this is that the expected price of the units when the buyer implements optimally the rule (R) is higher than the expected price when he implements independent mechanisms optimally. R i.e I . In the next proposition we characterize conditions under I which = R or I < R : ^ 1 (c1 ) H ^ 3 (c3 )][H ^ 2 (c2 ) H ^ 4 (c4 )] Proposition 7 I = R if and only if [H 0 for almost every c 2 : ^ 1 (c1 ) An equivalent result would be: I < R if and only if [H ^ 3 (c3 )][H ^ 2 (c2 ) H ^ 4 (c4 )] > 0 with positive measure: H ^ 1 (c1 ) H ^ 3 (c3 )] and [H ^ 2 (c2 ) H ^ 4 (c4 )] The last condition means that [H have the same sign. Assume suppliers are …rms and interpret the supply cost as their technology levels. In the optimal mechanisms derived the buyer is interested in …rms’virtual cost (that is his actual vision of the technologic di¤erence between the two) rather than their costs. And the ^ 1 (c1 ) H ^ 3 (c3 )][H ^ 2 (c2 ) H ^ 4 (c4 )] > 0 means that one country condition [H dominates the other in the buyer’s actual view. That this happens with positive probability is a su¢ cient and necessary condition for I < R . A proof is provided in appendix. We illustrate these results in the particular case of power distributions. 5 Example: power distributions We suppose Fi (ci ) = ci c i ¯ ci c i ¯ ai and ai 1 for all i 2 N and ci 2 probability density functions are given by fi (ci ) = ai ci c i ¯ ci c i ¯ ci c i ¯ ai 1 the virtual cost of supplier i is increasing in ci : Hi (ci ) = ci + ^ i. (1 + a1i )ci a1i ci . Since Hi is increasing we have Hi = H ¯ The condition for I = R writes ^ 1 (c1 ) [H ^ 3 (c3 )][H ^ 2 (c2 ) H ^ 1 (c1 ) The domain of H ^ 1 (c1 ) [H ¯ ^ 4 (c4 )] H 0 almost surely. ^ 3 (c3 ) is H ^ 3 (c3 ); H ^ 1 (c1 ) H 16 ^ 3 (c3 )] = [x; x]: H ¯ ¯ i. The , and Fi (ci ) fi (ci ) = ^ 2 (c2 ) Likewise the domain of H ^ 2 (c2 ) [H ¯ ^ 4 (c4 ) is H ^ 4 (c4 ); H ^ 2 (c2 ) H ^ 4 (c4 )] = [y; y]: H ¯ ¯ ^ i ’s are di¤erentiable and increasing, the set Because the H n o ^ 1 (c1 ) H ^ 3 (c3 )][H ^ 2 (c2 ) H ^ 4 (c4 )] 0 c 2 : [H is of full measure if and only if the set (x; y) 2 [x; x] [y; y] : xy 0 ¯ ¯ is also of full measure. The later is true if and only if [x 0 and y 0] or [x 0 and y 0] . ¯ ¯ [x ¯ ^ 1 (c1 ) H ^ 3 (c3 ) and H ^ 2 (c2 ) H ^ 4 (c4 )] 0] , [H ¯ ¯ 1 1 1 1 c] c and (1 + )c2 c (1 + )c3 a3 a3 ¯ 3 a2 a2 ¯ 2 ¯ 4 1 1 c3 + (c3 c3 ) and c2 + (c2 c2 ) c4 ]: ¯ ¯ ¯ a3 a2 0 and y , [c1 ¯ , [c1 ¯ Note that ci >ci and condition c1 c3 + a13 (c3 c3 ) means that supplier ¯ ¯ ¯ 1’s minimal unit cost c1 is su¢ ciently higher than his direct opponent’s ¯ maximal unit cost c3 . And the second condition means that supplier 4’s minimal unit cost c4 is su¢ ciently higher than his direct opponent’s ¯ maximal unit cost c2 . In other word group h = f3; 4g has a better technology for producing item 1 (via supplier 3) while the group l = f1; 2g has a better technology for producing item 2 (via supplier 2). Similarly the other condition [x 0 and y 0] implies that group l = ¯ f1; 2g has a better technology for producing item 1 (via supplier 1) while group h = f3; 4g has a better technology for producing item 2 (via supplier 4). None of the group has a better technology for producing both items. In the rest of the section we assume ci = 1 and ci = 0 for i 2 N . Let ¯ "i = (1 + a1i ). We provide simple expressions of the optimal mechanisms I and examine the evolution of R when the parameters ai increase. We know that I < R . Since Fi (ci ) = cai i , an increase in ai (a decrease in "i ) results in an increase in the probability of having low costs. Thus the parameter ai somehow describes the technology level of supplier i or at least what is believed about his technology level. The greater ai the better i’s technology. 5.1 Optimal mechanisms Using the expressions of the ironed out virtual costs the allocation rule qR is such that, for any m 2 : 17 R R "2 m 2 + "3 m 3 xR 1 (m) = x4 (m) = qf1;4g (m) = 1 if "1 m1 + "4 m4 R R R x2 (m) = x3 (m) = qf2;3g (m) = 1 if "1 m1 + "4 m4 > "2 m2 + "3 m3 : Given an arbitrary seller i 2 N we denote the other sellers i0 , j, and j 0 so that i and j (resp. i0 and j 0 ) sell the same good, and i and i0 (resp. j and j 0 ) belong to the same group.9 Thus we may show that supplier i 2 N receives payment: tR i (m) = min(ci ; "j mj +"j 0 mj 0 "i0 mi0 ) "i 0 if xR i (m) = 1 R if xi (m) = 0: So supplier i reports a unit cost of mi and the buyer uses a subjective correction factor "i in order to compare the bids of the suppliers. We shall call "actualized bid" the numbers "i mi . So the buyer compares actualized bids and allocates the contracts to the pair with the lowest (aggregate) actualized bid. Then if supplier i wins, he is paid an amount equal to his actualized bid plus the absolute value of the gap between aggregate actualized bids of the two pairs10 (all of this "disactualized"), provided that it does not exceed the maximal cost ci . Similarly we have xIi (m) = 1 if "i mi < "j mj 0 if "i mi > "j mj with the ties (i.e. situations where "i mi = "j mj ) solved in favor of suppliers 1 and 4. But this is irrelevant in the computation of the payment rule that can be written as: tIi (m) = min(ci ; 0 "j mj ) "i if xIi (m) = 1 if xIi (m) = 0: We see that in this case i wins if his actualized bid is lower than his opponent’s. Then he is paid an amount equal to his opponent’s actualized bid "disactualized" using supplier i’s correction factor, provided that it does not exceed the maximal cost ci . 9 We use the same notation throughout this section. for example, if i = 1 then i 0 = 2; j = 3 and j 0 = 4: 10 Indeed, "j mj + "j 0 mj 0 "i0 mi0 = "i mi + j("j mj + "j 0 mj 0 ) ("i0 mi0 + "i mi )j: If i = 1 for example, "3 m3 + "2 m2 "4 m4 = "1 m1 + j("3 m3 + "2 m2 ) ("1 m1 + "4 m4 )j: 18 5.2 Expected prices and changes in technology We know from the proof of proposition 7 in appendix that R I 1 ^ ^ ^ ^ ^ ^ ^ ^ = [jH 1 (c1 ) H3 (c3 )j+jH2 (c2 ) H4 (c4 )j j(H1 (c1 ) H3 (c3 )) (H2 (c2 ) H4 (c4 ))j]: 2 Substituting we obtain: R I 1 = E fj"1 c1 "3 c3 j + j"2 c2 "4 c4 j j("1 c1 "3 c3 ) ("2 c2 2 1 = E fmin[j"1 c1 "3 c3 j; j"2 c2 "4 c4 j]:1[("1 c1 "3 c3 )("2 c2 2 "4 c4 )jg "4 c 4 ) 0]g ; where 1(A) = 1 if assertion A is true, and 1(A) = 0 otherwise. We impose some more structure in the costs’distributions. Assume that "1 = "2 = "l > "h = "3 = "4 , that is, supplier from the same group have the same technology level but the group h has in general a higher technology than group l. We may think of the groups as countries. Consider an improvement of the technologies in both countries. This corresponds to an increase in ai , a decrease in "i and also an increase in "i ai = (1 + ai ): We make the additional assumption that this change in the parameters leaves constant the ratio ""hl . This implies that the technology level increases faster in the high-tech country than in the low-tech country. The following shows that such changes in technologies I result in an increase of R . i.e within this context an increase in the technology gaps between the two countries results in an increase in the di¤erence between the expected price when the rule (R) is optimally implemented and the expected price when items are optimally procured independently. Indeed the gap between the expected prices under the two rules depends on the parameters "l and "h as follows: R I "h "h "h "h 1 c3 j; jc2 c4 j]:1[(c1 c3 )(c2 c4 ) 0] = E "l min[jc1 2 "l "l "l "l Z "l min[jc1 ""hl c3 j; jc2 ""hl c4 j]: 1 = dc 1[(c1 ""hl c3 )(c2 ""hl c4 ) 0]al ca11 1 al ca22 1 ah ca33 1 ah ca44 1 2 Z "l al min[jc1 ""hl c3 j; jc2 ""hl c4 j]: 1 = dc; 1[(c1 ""hl c3 )(c2 ""hl c4 ) 0]ca11 1 al ca22 1 ah ca33 1 ah ca44 1 2 and the conclusion follows immediately. 19 6 Conclusion In this paper we have derived a procurement mechanism minimizing the expected price when the buyer is constrained to buy the units from suppliers coming from the many groups he faces. Under the optimal mechanism, virtual costs are summed over every potential winning subset of suppliers and the aggregate virtual costs are used to order these subsets; then contracts are awarded to suppliers of one of the maximal subsets. We have also derived an optimal mechanism when each item is procured independently of the others. Due to the assumption of costs independence the optimal independent mechanism consists of independent Myerson auctions on each item and yields a lower expected price than the optimal mechanism under the constraint of the groups. In the simple model with only two groups, we have found that if the probability that a group dominates the other in terms of virtual costs is positive then the two mechanisms are not equivalent. They are equivalent in the opposite case. Future directions of research may include relaxing the assumption of independent costs and allow in the model situations where some suppliers may supply more than one item. 7 References Armantier,O., Njiki, P.S. Fair return principle : e¢ ciency and cost effectiveness. Working paper, 2008. Armstrong, M. Optimal multi-object auctions. The Review of Economic Studies, 67(3):455-481, 2000. Branco, F. Multiple unit auctions of an indivisible good. Economic Theory, 8:77-101, 1996. Myerson, R. Optimal auction design. Mathematics of Operations Research, 6(1):58-73, 1981. A Appendix Proof of the revelation principle. Consider a mechanism ( ; q; t) and an equilibrium ; de…ne qG (c) = qG ( (c)) and ti (c) = ti ( (c)); And let x be the individual allocation mechanism associated with q: Suppose that under the direct mechanism (q; t) the other suppliers (than i) report their true costs, supplier i’s expected pro…t when he bids 20 mi 2 i rather than the true cost ci is: Ec i [ti (mi ; c i ) xi (mi ; c i )ci ]; And we have, ti (mi ; c i ) xi (mi ; c i )ci = ti ( (mi ; c i )) xi ( (mi ; c i ))ci = ti ( i (mi ); i (c i )) xi ( i (mi ); i (c i ))ci : Observe that Ec i [ti ( i (mi ); i (c i )) xi ( i (mi ); i (c i ))ci ] is supplier i’s expected pro…t when he bids i (mi ); and the other suppliers strategy is i under the mechanism ( ; q; t): Since is an equilibrium, Ec i [ti ( i (mi ); Ec i [ti ( i (ci ); = Ec i [ti (ci ; c i ) i (c i )) xi ( i (mi ); i (c i ))ci ] xi ( i (ci ); i (c i ))ci ] i (c i )) xi (ci ; c i )ci ]: Thus, Ec i [ti (mi ; c i ) xi (mi ; c i )ci ] Ec i [ti (ci ; c i ) xi (ci ; c i )ci ]; implying that supplier i’s best response is to report his true cost. Proof of proposition 3. If (q; t) is an IC mechanism then the functions i are convex. Indeed let 2 [0; 1]; and u; v 2 i : i( u + (1 )v) = max fTi (mi ) mi 2 i mi 2 i Xi (mi )( u + (1 )v)g = max[ fTi (mi ) Xi (mi )ug + (1 )fTi (mi ) maxfTi (mi ) Xi (mi )ug + (1 ) max fTi (mi ) = mi 2 i i (u) + (1 ) i (v) mi 2 Xi (mi )vg] Xi (mi )vg i i being convex is di¤erentiable almost everywhere in the interior of its domain and the derivative is increasing. The envelope theorem applied to condition (6) implies that: 0 i (ci ) = Xi (ci ): Thus Xi is increasing and therefore Xi is decreasing. (20) also implies that: Z ci Xi (ti )dti : i (ci ; ci ) = i (ci ; ci ) + ci 21 (20) (21) Replacing i (ci ; ci ) by Ti (ci ) Xi (ci )ci leads to (8). Conversely, from (8) we can have the expressions of Ti (ci ) and Ti (mi ): Substituting these expressions in what follows, we have: Z mi [Ti (ci ) Xi (ci )ci ] [Ti (mi ) Xi (mi )ci ] = Xi (ti )dti Xi (mi )(mi ci ); ci or equivalently, i (ci ; ci ) [Ti (mi ) Xi (mi )ci ] = Z mi Xi (ti )dti Xi (mi )(mi ci ): ci This last expression is positive if the function Xi is decreasing. Therefore condition (6) holds and the mechanism is IC. proof of proposition 6. If (q; t) is an IC mechanism then for any c 2 , Using x1 (c) + x3 (c) = 1 = x2 (c) + x4 (c), h i h i X ^ i (ci )xi (c) = H ^ 1 (c1 )x1 (c) + H ^ 3 (c3 )x3 (c) + H ^ 2 (c2 )x2 (c) + H ^ 4 (c4 )x4 (c) H i2N h i h i ^ 1 (c1 ); H ^ 3 (c3 )) + min(H ^ 2 (c2 ); H ^ 4 (c4 )) min(H h i h i I I I I ^ ^ ^ ^ = H1 (c1 )x1 (c) + H3 (c3 )x3 (c) + H2 (c2 )x2 (c) + H4 (c4 )x4 (c) X ^ i (ci )xI (c): H = i i2N Therefore P^ (q; t) P^ (qI ; tI ) and because of lemma 4 P (q; t) I I P^ (q ; t ): We know from equation (13) that if (qI ; tI ) is IC then: o X Z ci n I I I I ^ ^ P (q ; t ) P (q ; t ) = Ki (Fi (ci )) Ki (Fi (ci )) dXiI (ci ); i2N c ¯i ^ i (Fi (ci )) Ki (Fi (ci )) 6= 0 then K ^ i (Fi (ci )) Ki (Fi (ci )) < 0 If K ^ i (ci ) is constant in some neighborhood of ci (property (c)) and and H I so Xi (ci ) is also constant11n(i.e. dXiI (ci ) = 0) in some neighborhood o R ci I ^ i (Fi (ci )) Ki (Fi (ci )) dX (ci ) = 0 and of ci : We may conclude c K i ¯i 11 I Without loss of generality suppose i = 1. Recall that 1 ; c 1 ) = 1 if and only R x1 (c I I ^ 1 (c1 ) ^ 3 (c3 ). By de…nition we have X (c1 ) = if H H x (c 1 1 ; c 1 )f 1 (c 1 )dc 1 . 1 1 I ^ i )i6=1 and Since the integral is taken over c 1 we conclude X1 (c1 ) depends only on (H ^ 1 (c1 ). Thus it is constant if H ^ 1 (c1 ) is constant. on H 22 P^ (qI ; tI ) = P (qI ; tI ): Thus P (q; t) P (qI ; tI ) for any IC mechanism (q; t): We need to show that (qI ; tI ) is IR and actually IC. First, note that xIi (m)mi ; taking the by de…nition of the payment rule tI : tIi (m) I I 0; and Xi (mi )mi i.e. Ii (mi ) integral over m i ; we have Ti (mi ) I I (q ; t ) is IR. We show that (qI ; tI ) is IC using proposition 3; consider two suppliers i and j selling the same item. We will show that XiI is decreasing in the case i 2 f1; 2g : The proof is similar in the case where i 2 f3; 4g except that the large inequalities in the implication below must be replaced by strict inequalities. ^ i (t) ^ i (s) and the Let t; s 2 i and suppose t s: Therefore H H following assertion is true: ^ i (s) H ^ j (mj ) ) H ^ i (t) H ^ j (mj ); H i.e. xIi (s; m i ) = 1 ) xIi (t; m i ) = 1; or equivalently xIi (s; m i ) xIi (t; m i ): XiI (t); and XiI is deSince m i is arbitrary we also have XiI (s) creasing. Using the de…nition of the payment rule and taking the integral over m i ; we deduce: Z ci I I Ti (mi ) = Xi (mi )mi + XiI (t; c i )dt mi and I i (ci ) XiI (ci )ci = 0; Rc and …nally: TiI (mi ) = Ii (ci )+XiI (mi )mi + mii XiI (t; c i )dt. Using Proposition 3 we conclude (qI ; tI ) is IC. From the …rst lines of our proof we have I = P (qI ; tI ) = P^ (qI ; tI ) = ^ 1 (c1 ); H ^ 3 (c3 )) + min(H ^ 2 (c2 ); H ^ 4 (c4 ))]: E[min(H proof of proposition 7. R I = TiI (ci ) ^ 1 (c1 ) + H ^ 4 (c4 ); H ^ 2 (c2 ) + H ^ 3 (c3 ))] = E[min(H ^ 1 (c1 ); H ^ 3 (c3 )) + min(H ^ 2 (c2 ); H ^ 4 (c4 ))] E[min(H ^ 1 (c1 ) + H ^ 4 (c4 ); H ^ 2 (c2 ) + H ^ 3 (c3 )) = E[min(H ^ 1 (c1 ); H ^ 3 (c3 )) min(H ^ 2 (c2 ); H ^ 4 (c4 ))]; min(H 23 Using the equality min(a; b) = R I 1 ^ = [jH 1 (c1 ) 2 ^ 1 (c1 ) j(H a+b ja bj 2 we obtain: ^ 3 (c3 )j + jH ^ 2 (c2 ) H ^ 3 (c3 )) H ^ 2 (c2 ) (H ^ 4 (c4 )j H ^ 4 (c4 ))j]: H The triangular inequality implies that: ^ 1 (c1 ) H ^ 3 (c3 )j+jH ^ 2 (c2 ) H ^ 4 (c4 )j j(H ^ 1 (c1 ) H ^ 3 (c3 )) (H ^ 2 (c2 ) H ^ 4 (c4 ))j jH I ^ 1 (c1 ) H ^ 3 (c3 )j + jH ^ 2 (c2 ) Then we have R = 0 if and only if jH ^ 4 (c4 )j = j(H ^ 1 (c1 ) H ^ 3 (c3 )) (H ^ 2 (c2 ) H ^ 4 (c4 ))j almost surely. Using H the square of both sides of the equality we obtain the equivalent assertion ^ 1 (c1 ) H ^ 3 (c3 ))jj(H ^ 2 (c2 ) H ^ 4 (c4 ))j = 2(H ^ 1 (c1 ) H ^ 3 (c3 ))(H ^ 2 (c2 ) 2j(H ^ 4 (c4 )) almost surely. And this simply means (H ^ 1 (c1 ) H ^ 3 (c3 ))(H ^ 2 (c2 ) H ^ H4 (c4 )) 0 almost surely. 24 0