Académique Documents
Professionnel Documents
Culture Documents
Abstract We show that the Serial Cost Sharing method for heterogeneous goods [J. Econ. Theory 87 (1999) 275], and a large number of other cost sharing mechanisms, have the same strong strategic properties as the Serial Cost Sharing method for homogeneous goods [J. Econ. Theory 64 (1992) 178; Econometrica 60 (1992) 1009], including uniqueness of the Nash equilibrium for all utility proles and cost functions, dominance solvability, solvability in overwhelmed actions, and robustness to coalitional deviations. We describe several applications to cost / surplus sharing and the Internet. 2002 Elsevier Science B.V. All rights reserved.
Keywords: Cost sharing; Serial cost sharing; Ordinal cost sharing; O-solvable JEL classication: C71; H4
1. Introduction Cost sharing problems arise in many settings. The most common example is the shared production of resources in which a number of agents must share the costs of production for various goods (see, e.g., Young, 1985). Other important versions of these problems arise in the sharing of prots from cooperative ventures and even the shared use of network resources on the Internet (Friedman, 1999; Shenker, 1995). One important question in these settings is whether the chosen cost sharing method induce games with nice strategic properties. Restated, a key question is whether there will be incentives for the agents to manipulate the game. One extremely interesting result in this subject was presented by Moulin and Shenker
* Tel.: 1 1-607-255-9130; fax: 1 1-607-244-9129. E-mail address: friedman@orie.cornell.edu (E.J. Friedman). 0165-4896 / 02 / $ see front matter 2002 Elsevier Science B.V. All rights reserved. PII: S0165-4896( 02 )00012-4
146
(1992a,b) and Shenker (1995) who introduced a cost sharing mechanisms with extremely strong strategic properties for homogeneous goods. Dubbed the Serial Cost Sharing method 1 this mechanism leads to games in which the Nash equilibrium is unique, robust to coalitional deviations and the only rationalizable strategy prole. In addition, this Nash equilibrium is the unique outcome of adaptive learning (Milgrom and Roberts, 1991) and reasonable learning in asynchronous low information environments (Friedman and Shenker, 1998). This makes the Serial Cost Sharing method extremely attractive since one can be condent that the Nash equilibrium of the induced game is its natural outcome. For computer networks, this implies that fair queuing, which is a network service protocol that (approximately) implements the Serial Cost Sharing method leads to more stable service than standard rst-inrst-out protocols, which correspond to the averagecost mechanism (Friedman and Shenker, 1998). However, in many important applications goods are not homogeneous. In this setting it is natural to ask whether there exists an extension of the Serial Cost Sharing method which maintains its strong strategic properties. One natural generalization of the Serial Cost Sharing method to heterogeneous goods was introduced in Friedman and Moulin (1999) based on axiomatic (non-strategic) considerations. In this paper we show that this generalized Serial Cost Sharing method maintains the same strategic properties as the original Serial Cost Sharing method. However, we also show that this mechanism is not the only such mechanism with these properties. While the original Serial Cost Sharing method was essentially the only symmetric mechanism with these properties (Moulin and Shenker, 1992a,b), in the heterogeneous setting there is a large class of mechanisms which also share these strong strategic properties. Three particularly interesting examples of these are the Moulin Shenker method, the Axial Serial Path method and the Axial Serial Cost Sharing method, all of which are ordinally invariant and have many other interesting properties as described by Sprumont (1998).
2. The serial cost method and generalizations We consider a typical cost sharing problem: there are n agents, each of whom has a demand for good i, qi [ Q i 5 [0, q max ], where q 5 ( q1 , . . . , qn ), i [ N 5 h1,2, . . . ,n j and q max . 0. (Note that we can have q max 5 `.) The cost of serving these demands is C( q), which must be divided among the agents; the cost share for agent i is given by x i ( q ; C ). Note that we will identify agents with their goods; thus qi is both the demand of agent i and the demand for good i. Moulin and Shenker (1992a,b) considered the case where the goods are homogeneous i.e., the cost function can be written C( u q u ), where u q u 5 o n i 5 1 qi . They assume that C( ? ) is continuously differentiable and nondecreasing. In this setting, the Serial Cost Sharing
147
C( q ) ]]]]]. ( n 1 1 2 k )(n 2 k ) k 51
i 21
Although the Serial Cost Sharing method is well dened for any such cost function, in order for it to have strong strategic properties it is necessary that the cost function have increasing marginal costs. As in their paper, we will assume strictly increasing marginal costs to simplify the exposition.2 This is quite natural in many network contexts for which the Serial Cost Sharing method (or fair share) is particularly applicable. Friedman and Moulin (1999) extended the Serial Cost Sharing method to the case where the goods may be heterogeneous, i.e., C( q) can be an arbitrary nondecreasing and continuously differentiable function of its n variables. They describe the cost sharing mechanism given by x ( q ; C ) 5 i C(g SC (t ; q)) dg SC (*) i (t ; q)
0 SC i
as the natural generalization of the Serial Cost Sharing method, where i C( p) is the partial derivative of C with respect to pi evaluated at p, g SC is the path from 0 to q given by g SC (t ; q) 5 (tu) q, for t $ 0, where ( p q) i 5 min[ pi , qi ] is the componentwise minimum of the two vectors and u is the unit vector (1,1, . . . ,1). Note that in comparison to the AumannShapley method which integrates the partial derivative of the cost function along the line connecting 0 and q, the Serial Cost Sharing method integrates as far as possible along the diagonal and then once some agents have been completely served along diagonals in the subspace of the remaining agents. As we will see in Section 3, this heterogeneous version of the Serial Cost Sharing method has the same strong strategic properties as the homogeneous version when the marginal cost is strictly increasing in all variables. Intuitively this arises in the same way for heterogeneous goods as in the homogeneous case, since the cost shares of agents do not depend on demands that are larger than their own demand. This can be seen directly from (*) since the part of the path relevant to an agent is independent of higher demands. However, these strategic properties of the Heterogeneous Serial Cost Sharing method are not unique to the Serial Cost Sharing method, as they were in the homogeneous case (under symmetry and several other restrictions). In particular, consider any cost sharing mechanism which can be written as (*) where g SC is replaced by any continuous, non-decreasing path which can be written as follows. Let f (t ; C ) be any function which is nondecreasing and continuous in t (for xed C ) with f (0; C ) 5 0 and lim t ` (t ; C ) . q max for any C, then let the path be dened by g SC (t ; q) 5 f (t ; C ) q. We call these methods xed path methods and note that they are analogous to the paths studied
As they discuss, strictness can be relaxed by assuming that agents preferences are strictly convex.
148
by Moulin (1999a) in the discrete version of this problem and in both the continuous and discrete versions for a restricted class of homogeneous cost functions in Moulin (1999b). For functions, f, which are independent of C this includes many asymmetric methods, but the only symmetric one is the Serial Cost Sharing method. However, when f depends nontrivially on C there are many symmetric methods, since for any C, which is not symmetric for any group of agents the notion of symmetry does not constrain the choice of a path at all. Clearly, such methods include many methods which could be ruled out by locality assumptions (i.e., x( q ; C ) should not depend on properties of C outside of the domain [0, q ]); however, there still exist symmetric methods which are local. Some simple, but perhaps uninteresting methods include fi (t ; C ) 5 ai (C )t where ai (C ) is any function of the derivatives of C at 0. More complex and perhaps more interesting are several ordinally invariant methods.
149
method is a xed path method, the Axial Serial Cost Sharing method is not. Nonetheless, it is an easy corollary of our analysis, and was rst shown in Alcalde and Silva (1999), that the Axial Serial Cost Sharing method has the same strong strategic properties as the xed path methods. Theorem 1. The Axial Serial Cost Sharing method is not a xed path method. Proof. In Appendix A.
3. Cost sharing games We now consider the strategic properties of cost sharing games when xed path methods are used to allocate the costs. Following Moulin and Shenker (1992a,b) we assume that player i s utility is given by a concave utility function Ui ( qi , x i ) which is increasing in qi and decreasing in x i . Thus, given a set of players, N, a set of utility functions, Ui , a cost sharing method, x( ? ; ? ) and a cost function C this denes a noncooperative game in strategic form. We now present several examples.
3.1. Examples
An early example consists of a group of agents, who are identical Poisson sources of data sharing a network link.3 The cost function is the expected queue size which is given by C( q) 5 1 /( m 2 u q u ) where m is the capacity of the link and qi is the data transmission rate by agent i. In this case the precursor to the Serial Cost Sharing method was denoted fair share by Shenker (1990) and was based on a protocol known as fair queuing (see, e.g., Demers et al., 1990). When there are different types of trafc on the link the cost function is no longer homogeneous in the qi s. In this case, the cost function is heterogeneous in the goods and once again we can apply any xed path method to allocate the costs (queue lengths). An immediate generalization (in the case of Poisson sources) arises when there is a complex network with agents located at different nodes. In this case, if the Serial Cost Sharing method (or any other xed path method) is applied at each node, then the aggregate cost sharing corresponds to the application of the Heterogeneous Serial Cost Sharing method to the aggregate cost function. This model is analyzed in Shenker (1990). There are a wide variety of queuing problems which are mathematically equivalent to this problem, such as the ordering of web server or database requests and the allocation of dial in modems.4 They also arise in a variety of service facilities problems, see e.g., Mendelson and Whang (1990), Stidham (1992) and Friedman and Landsberg (1993).
3 4
See Friedman and Shenker (1998) for an extensive bibliography of such models. These are discussed in detail in Friedman (1997).
150
In all of these queuing models, the marginal cost is typically increasing in all variables, due to stochastic congestion effects, and thus our results are applicable. A different class of examples arise in surplus sharing problems where qi is the effort of agent i and C( q) is the group prot, when agents abilities are heterogeneous. In this case a players utility is decreasing in qi and increasing in x i , but when C( q) is concave and efforts are substitutes, our analysis is applicable.
4. Strategic properties We are interested in a variety of strategic properties which were analyzed by Moulin and Shenker (1992a,b). These are: uniqueness of the Nash equilibrium, strong equilibria, uniqueness of the set of rationalizable outcomes and convergence of adaptive learners. In addition, we are interested in the convergence of reasonable learners in network settings (Friedman and Shenker, 1998). We will show all of these results for xed path methods by showing that the game is O -solvable (Friedman and Shenker, 1998). These other properties all follow immediately as they are implied by O -solvability.5 Denition 1. (i) A strategy qi is overwhelmed by q i9 with respect to set S # Q 2 i if maxhUi ( qi , x i ( q ; C ))u q2 i [ S2 i j , minhUi ( q i9 , x i ( q i9 , q2 i ; C ))u q2 i [ S2 i j. When S 5 (S1 , S2 , . . . ,Sn ) let Oi (S ) denote the set of strategies which are not overwhelmed with respect to S2 i and let O(S ) 5 (O1 (S ), . . . ,On (S )). (ii) A game is O -solvable if the set O ` 5 lim n ` O n ( Q ) consists of a single strategy for each player, where O n is the functional composition of O with itself n times. We now present our main theorem. Theorem 2. Assume that the marginal cost ( i C( q)) is strictly increasing in all variables, x i ( ? ; ? ) is a xed path method and that preferences, Ui ( qi , x i ) are increasing in qi , decreasing in x i and concave. Then the induced game is O-solvable. The proof of this theorem is in Appendix A. However, it is straightforward to show that this theorem implies all the other strategic properties of interest and more. Corollary 1. Assume that the marginal cost is strictly increasing in all variables,
5
Moulin and Shenker (1992a,b) proved these results by iteratively eliminating dominated strategies, while we iteratively eliminate overwhelmed strategies a much stronger requirement, but surprisingly still possible in this problem using much of the same technical machinery.
151
x i ( ? ; ? ) is a xed path method and that preferences, Ui ( qi , x i ) are increasing in qi , decreasing in x i and concave. Then the induced game : (i) has a unique Nash equilibrium. (ii) has a unique rationalizable strategy. (iii) play will converge to the Nash equilibrium under adaptive learning (Milgrom and Roberts, 1991). (iv) the unique Nash equilibrium is strong. (v) play will converge to the Nash equilibrium under reasonable learning in network settings (Friedman and Shenker, 1998). Proof. Parts (i), (ii) and (iii) follow immediately from the fact that any overwhelmed strategy is also dominated and thus O -solvability implies dominance solvability which is well known to imply (i) and (ii) and was shown by Milgrom and Roberts (1991) to imply (iii). Parts (iv) and (v) were proven in Friedman and Shenker (1998). h Since the Heterogeneous Serial Cost Sharing method, the MoulinShenker Ordinal method and the Axial Serial Path methods are all xed path methods, this proves that they all have the same strategic properties as the homogeneous version of the Serial Cost Sharing method. Interestingly, even though it is not a xed path method, the Axial Serial Cost Sharing method has been shown to be solvable by the iterated deletion of dominated strategies by. In fact, it is straightforward to show using Lemma 1 in Appendix A, that it generates O -solvable games and also has very strong strategic properties, demonstrating that our characterization is not tight, i.e., there are methods which are not xed path methods which satisfy Theorem 2. Lastly, we note that Shenker (1992) has studied the properties of abstract serial mechanisms in general production economies. The relationship between his abstract analysis and our results merits further study.
152
Acknowledgements I would like to thank Herve Moulin, Scott Shenker and two anonymous referees for many helpful comments and Caltech for its hospitality during part of this work. This material is based upon work supported by the National Science Foundation under Grant No. ANI-9730162. www:http: / / orie.cornell.edu / friedman
Appendix A. A proof of theorems A.1. Theorem 1 Let C( q) 5 q1 1 q2 1 2q 2 1 q 2 . Then for any q, ( q, C ) is axially normalized. Thus, x 1 (b, b ) 5 C(b, b ) / 2 for any b . 0. A direct computation shows that the path cannot be entirely on the diagonal. Thus, there exists some point (a, b ) on the path such that a b. Assume that 0 , a , b. (The other case follows the same analysis.) Then from the formula for path generated methods, if b 2 e . a then x 1 (b 2 e, b ; C ) 5 C(b, b ) / 2 2 [C(b, b ) 2 C(b 2 e, b )]. However, by the denition of the Axial Serial Cost Sharing method x 1 (b 2 e, b ; C ) 5 C(b 2 e, b 2 e ) / 2. This implies that C(b, b ) / 2 2 [C(b, b ) 2 C(b 2 e, b )] 5 C(b 2 e, b 2 e ) / 2. Inserting the value for C shows that this is equivalent to the requirement that 2 b 2 * e 2 b 2 * e 2 1 e 3 5 0 which cannot be satised for some sufciently small e . 0, leading to a contradiction. Thus, the Axial Serial Cost Sharing method cannot be a xed path method. A.2. Theorem 2 We rst present some properties of xed path methods. Lemma 1. Assume that marginal cost is strictly increasing in all variables and that x i ( ? ; ? ) is a xed path method. Dene li ( qi ) 5 min [t ufi (t ) $ qi ]. Then : (a) x i ( q ; C ) is strictly increasing and strictly convex in qi (b) x i ( q ; C ) is nondecreasing in q j for all j 5 i. 9 ; C ). (c) For all q and q 9 j such that both l( q j ), j( q 9 j ) $ li ( qi ) then x i ( q ; C ) 5 x i ( q2 j , q j Proof. Using this notation, we can write the cost sharing formula as xi ( q; C ) 5
E E
l i ( qi )
i C(f (t ) q) d(fi (t ) qi ).
xi ( q; C ) 5
i C(f (t ) q 9 ) d(fi (t ) q 9 i ).
153
Thus
Di ( q) 5
l i ( q i 1d )
li( qi )
i C(f (t ) q 9 ) d(fi (t ) q 9 i ).
Since i C( p) . 0 for p . 0, Di ( q) . 0, x i ( q ; C ) is strictly increasing in qi . Also, since i C( p) is strictly increasing in p, Di ( q) is strictly increasing in qi and thus x i ( q ; C ) is strictly convex in qi , completing the proof of (a). (ii) For any d . 0 and j i let hi ( q) 5 x i ( q 9; C ) 2 x i ( q ; C ), where q 9 5 ( q2 j , q j 1 d ). Since the paths g (t ; q) and g (t ; q) agree for t , l j ( q j ), if l j ( q j ) $ li ( qi ), then hi ( q) 5 0, proving (c). Assume that l j ( q j ) , li ( qi ) then
hi ( q) 5
li( qi )
l j( q j )
Proving (b), from the monotonicity of C( ? ). h The following lemma is a slight modication of Theorem 10 in Friedman and Shenker (1998). Lemma 2. Consider a game kG, Q l, where G is the payoff function and Q i is a compact subset of R 1 . Assume there exist a set of nondecreasing functions li : Q i R 1 such that the following hold : j , q2 j ) for any q j $ q j , i j. (i) Gi ( q) $ Gi ( q j such that both l j ( q j ), l j ( q j ) $ li ( qi ), (ii) Gi ( q j , q2 j ) 5 Gi ( q j , q2 j ) for any q j , q ;i j. (iii) for each q2 i there exists an element BR i ( q2 i ) such that x i BR i ( q2 i ) Gi (BR i ( q2 i ), q2 i ) . Gi (x i , q2 i ). j , q2 ij ) for any q j such that l j ( q j ) $ li (BR i ( q2 i )). (iv) BR i ( q2 i ) 5 BR i ( q Then the game is O-solvable. The proof of this lemma is a straightforward modication of the proof (of Theorem 10) given in Friedman and Shenker (1998), which is for the case in which li ( qi ) 5 qi . We now show that games arising from generalized path methods satisfy these conditions: (i) This follows from (b) since x i ( q ; C ) is nondecreasing in q j . (ii) This follows directly from (c). (iii) The uniqueness of the best reply follows from the strict convexity of x i ( q ; C ) w.r.t. qi and the convexity of preferences. (iv) The best reply occurs at the unique point of tangency between the indifference
154
curves and the opportunity set, in the ( qi , x i ) plane. Neither of these is (locally) affected by such changes in q j by (c). Thus, the game induced by a xed path method satises the assumptions of the lemma and must be O-solvable, completing the proof. h
References
Alcalde, J., Silva, J., 1999. A procedure for sharing recycling costs. mimeo, University of Alicante. Demers, A., Keshav, S., Shenker, S., 1990. Analysis and simulation of a fair queueing algorithm. Journal of Internetworking 1 (1), 326. Friedman, E.J., 1997. Learnability in non-atomic externality games, with applications to computer networks (www.orie.cornell.edu / | friedman). Friedman, E.J., 1999. Strong monotonicity in surplus sharing. mimeo. Friedman, E.J., Landsberg, A.S., 1993. Short run dynamics of multi-class queues. Operations Research Letters 14, 221229. Friedman, E.J., Moulin, H., 1999. Three methods to share joint costs. Journal of Economic Theory 87 (2), 275312. Friedman, E.J., Shenker, S., 1998. Learning and implementation in the Internet, mimeo, available from www.orie.cornell.edu / | friedman. Mendelson, H., Whang, S., 1990. Optimal incentive-compatible priority pricing for the m / m / 1 queue. Operations Research 38 (5), 870883. Milgrom, P., Roberts, J., 1991. Adaptive and sophisticated learning in repeated normal form games. Games and Economic Behavior 3, 82100. Moulin, H., 1999a. Incremental cost sharing: Characterization by strategyproofness. Social Choice and Welfare 16 (2), 279320. Moulin, H., 1999b. Rationing a commodity along xed paths. Journal of Economic Theory 84, 4172. Moulin, H., Shenker, S., 1992a. Average cost pricing versus serial cost sharing: An axiomatic comparison. Journal of Economic Theory 64 (1), 178201. Moulin, H., Shenker, S., 1992b. Serial cost sharing. Econometrica 60, 10091037. Shenker, S., 1990. Efcient network allocations with selsh users. In: King, P.J.B., Mitrani, I., Pooley, R.J. (Eds.), Performance 90. North-Holland, New York, pp. 279285. Shenker, S., 1992. On the strategyproof and smooth allocation of private goods in production economies, mimeo. Shenker, S., 1995. Making greed work in networks: a game-theoretic analysis of switch service disciplines. IEEE /ACM Transactions on Networking 3, 819831. Sprumont, Y., 1998. Ordinal cost sharing. Journal of Economic Theory 81 (1), 126162. Stidham, S., 1992. Pricing and capacity decisions for a service facility: stability and multiple local optima. Management Science 38 (8), 11211139. Young, H.P., 1985. In: Cost Allocation: Methods, Principles and Applications. Elsevier, New York.