Designing CostSharing Methods for Bayesian Games
 663 Downloads
Abstract
We study the design of costsharing protocols for two fundamental resource allocation problems, the Set Cover and the Steiner Tree Problem, under environments of incomplete information (Bayesian model). Our objective is to design protocols where the worstcase Bayesian Nash equilibria have low cost, i.e. the Bayesian Price of Anarchy (PoA) is minimized. Although budget balance is a very natural requirement, it puts considerable restrictions on the design space, resulting in high PoA. We propose an alternative, relaxed requirement called budget balance in the equilibrium (BBiE). We show an interesting connection between algorithms for Oblivious Stochastic optimization problems and costsharing design with low PoA. We exploit this connection for both problems and we enforce approximate solutions of the stochastic problem, as Bayesian Nash equilibria, with the same guarantees on the PoA. More interestingly, we show how to obtain the same bounds on the PoA, by using anonymous posted prices which are desirable because they are easy to implement and, as we show, induce dominant strategies for the players.
Keywords
Price of anarchy Bayesian games Network design Costsharing games1 Introduction
A costsharing game is an abstract setting that describes interactions of selfish players in environments where the cost of the produced solution needs to be shared among the participants. A costsharing protocol prescribes how the incurred cost is split among the users. This defines a game that is played by the participants, who try to select outcomes that incur low personal costs. Chen, Roughgarden and Valiant [23] initiated the design aspect, seeking for protocols that induce approximately efficient equilibria, with low Price of Anarchy (PoA) [55]. Similarly, we study the design of costsharing protocols, for two wellstudied and very general resource allocation problems with numerous applications, the Set Cover and the Steiner tree (multicast) problem.
Set Cover Game
In the (weighted) set cover problem, there is a universe of n elements, U = {1, …, n}, and a family of subsets of U, \(\mathscr {F}=\{F_{1},\ldots , F_{m}\}\), with weights/costs \(c_{F_{1}},\ldots , c_{F_{m}}\). A subset of elements, X ⊆ U, needs to be covered by the Fi′s so that the total cost is minimized. We are interested in a game theoretic version, where there are X players and U possible types; each player’s type associates her with a specific element of U and X corresponds to the set of players’ types. Multiple players may have the same type. A player’s action is to chose a subset from \(\mathscr {F}\) that covers her element, and pay some costshare for using it. A costsharing method prescribes how the subsets’ costs are split among players.
Multicast Game
In a multicast game, there is a rooted (connected) undirected graph G = (V, E, t), where each edge e carries a nonnegative weight c_{e} and t is a designated root. There are k players and V  = n possible types; each player’s type associates him with a specific vertex of V which needs to establish connectivity with t. The players’ strategies are all the paths that connect their terminal with t. A costsharing method defines the costshares of the players.
CostSharing under Uncertainty
There are two different possible sources of uncertainty that may need to be considered in the above scenarios. Firstly, the designer needs to specify the costsharing protocol, having only partial information about the players’ types. Moreover, the players themselves, when they select their actions, may have incomplete knowledge about the types of the other players. We approach the former by using a stochastic model similar to [29], and the latter, as a Bayesian game, introduced by [48], which is an elegant way of modelling selfishness in partialinformation settings. In a Bayesian game, players do not know the private types of the other players, but only have beliefs, expressed by probability distributions over the possible realizations of the types.
The order of events is as follows; first, the designer specifies the costsharing methods, using the product probability distribution over the players’ types, then the players interact in the induced Bayesian game, and end up in a Bayesian Nash Equilibrium. We are interested in the design of protocols, where all equilibria have low cost i.e., the (Bayesian) PoA of the induced game is low.
Budget Balance in the Equilibrium (BBiE)
One of the axioms that [23] required in their design space, that every costsharing protocol should satisfy, is budget balance i.e., that the players’ costshares cover exactly the cost of any solution. Although budget balance is a very natural requirement, it puts considerable restrictions on the design space. However, since we expect that the players will end up in a Nash equilibrium, it is not clear why one should be interested to impose budget balance in nonequilibrium states; the players are going to deviate from such states anyway. We propose an alternative, relaxed requirement that we call budget balance in the equilibrium (BBiE). A BBiE costsharing protocol satisfies budget balance in all equilibria; for any nonequilibrium profile we do not impose this requirement. This natural relaxation enlarges the design space but maintains the desired property of balancing the cost in the equilibrium. More importantly, this amplification of the design space, allows us to design protocols that dramatically outperform the best possible PoA bounds obtained by budgetbalanced protocols. Indeed, by restricting to budgetbalanced protocols, a lower bound of Ω(n) exists, for the complete information set cover game [23]; we extend this lower bound for the Bayesian setting. We further show a lower bound of \({\Omega }(\sqrt {n})\), for the multicast Bayesian game. We demonstrate that, by designing BBiE protocols, we can enforce better solutions, that dramatically improve the PoA. For the set cover game, we improve the PoA to O(n/log n) (or O(log n) if m = poly(n)). Regarding the multicast game, we improve the PoA to O(1).
Posted Prices
It is a very common practice, especially in large markets and double auctions, for sellers to use posted prices. More closely to costsharing games is the model proposed by Kelly [51] regarding bandwidth allocation. Kelly’s mechanism processes players’ willingness to pay and posts a price for the whole bandwidth. Then each player pays a price proportional to the bandwidth she uses. This can be seen as pricing an infinitesimal quantity of bandwidth and the players, acting as pricetakers, choose some number of quantities to buy. It turns out that it is in the best interest of the players to buy the whole bandwidth.
The use of posted prices, to serve as costsharing mechanism, is highly desirable, but not always possible to achieve; a price is posted for each resource and then the players behave as price takers, picking up the cheapest possible resources that satisfy their requirements. Such a mechanism is desirable because it is extremely easy to implement and also induces dominant strategies. We stress that our main results can be implemented by anonymous posted prices.
1.1 Results and Discussion
We study the design of costsharing protocols for two fundamental resource allocation problems, the Set Cover and the Steiner tree problem. We are interested in environments of incomplete information where both the designer and the players have partial information, described by prior probability distributions over types. Our objective is to design costsharing protocols that are BBiE and the worstcase equilibria have low cost, i.e. the Bayesian PoA is minimized.
We show an interesting connection between algorithms for ObliviousStochastic optimization problems and costsharing design with low PoA. We exploit this for both problems and we are able to enforce approximate solutions of the stochastic problem, as Bayesian Nash equilibria, with the same guarantees on the PoA. Although this connection is quite simple, it results in significant improvement on the PoA comparing to budgetbalanced protocols. More precisely, we map each player to a single specific strategy and charge very high costs for any alternative strategy. In this way, their mapped strategy becomes a (strongly) dominant strategy. For the set cover game, we enforce the oblivious solution given by [45]. They apriori map each player i to some subset \(F_{i} \in \mathscr {F}\); then, if i is sampled, F_{i} should be in the induced solution. For the multicast game, the algorithm of [42], for the online Steiner tree problem, provides an oblivious solution.
BudgetBalanced Protocols
BBiE Protocols
PoA of BBiE protocols
BBiE protocols/posted prices  

Set cover  Multicast  
Complete information  1  1 
Bayesian  \(O(n/\log n)\)  O(1) 
Posted Prices
(Section 5). For the Bayesian (and stochastic) settings, expost BBiE cannot be obtained by anonymous prices. We first discuss limitations of other concepts, such as BBiE with “high” probability or bounded possible excess and deficit. Then, we examine prices that result in exante BBiE. We present anonymous prices with the same upper bounds as the BBiE protocols, for the unweighted set cover and for the multicast games, respectively. We stress that oblivious solutions may not be sufficient to guarantee low PoA for anonymous posted prices, in contrast to the BBiE protocols. This is because it is not clear anymore how to enforce players to choose desirable strategies, since anonymous prices are available to anyone. The reason that they exist here is due to the specific properties of the oblivious solution.
Regarding the weighted set cover game, we are only able to provide semianonymous prices with the same bounds; by semianonymous we mean that the prices for each player do not depend on her identity, but only on her type. We leave the case of anonymous prices as an open question. We remark that in all cases, posted prices induce dominant strategies for the players. Finally, for the polytime determinable prices, we give tight lower bounds.
PriorIndependent Mechanisms
(Section 6). Clearly, the above BBiE protocols and posted prices depend on the prior distribution. Priorindependent mechanisms are also of high interest and in Section 6 we discuss their limitations.
Complete Information Setting
(Section 7). We further study the complete information setting (see Tables 1 and 2). By using either BBiE protocols or anonymous posted prices, we enforce the strategy profile of the optimum solution. Note that, while trying to bound the PoA, computational issues are not of primary concern. However, if we stick to protocols that can be determined in polynomial time, we can upper bound the PoA of set cover and multicast games by O(log k) and 1.39, respectively, where k is the number of players. Moreover, we argue that there are no anonymous prices, computed in polynomial time, for the set cover game, with PoA= o(log k).
1.2 Related Work
There is a vast amount of research in costsharing games and so, we only mention some of the most related. Moulin and Shenker [57] studied costsharing games under mechanism design context; they characterized the budgetbalanced and group strategyproof mechanisms and identify the one with minimum welfare loss. In similar context, other papers considered (group)strategy proof and efficient mechanisms and relaxed the budgetbalanced constraint; Devanur, Mihail and Vazirani [34] and Immorlica, Mahdian and Mirrokni [50] studied the set cover game under this context showing positive and negative bounds on the fraction of the cost that is covered.
Regarding network design games, there is a long line of works mainly focusing on fair cost allocation (Shapley costsharing mechanism), originated by [5]. Anshelevich et al. [5] showed a tight Θ(log k) bound on the PoS for directed networks, while for undirected networks several variants have been studied [11, 12, 22, 24, 35] but the exact value of PoS still remains an open problem. For multicast games, Li [56] proved an upper bound of O(log k/loglog k), while for broadcast games, Fiat et al. [39] proved an O(loglog k) upper bound which was improved to constant due to Bilò, Flammini and Moscardelli [13]. The PoA of some special equilibria has been also studied in [17, 21].
Chen, Roughgarden and Valiant [23] were the first to study the design aspects for this game, identifying the best protocol with respect to the PoA and PoS in various cases, followed by [62] for parallel links, [40, 43, 54] for weighted congestion games, [29, 47, 59] for network games, [41] for routing games and [52] for resource allocation. The Bayesian Price of anarchy was first studied in auctions by [26]; see also [58] for routing games, and [61] for the PoS of Shapley protocol in costsharing games.
Close in spirit to our work is the notion of Coordination Mechanisms [25] which provide a way to improve the PoA in cases of incomplete information. Similar to our context, the designer has to decide in advance gamespecific policies, without knowing the exact input. Such mechanisms have been used for scheduling problems under the objective of minimising the makespan [2, 7, 16, 49, 53] or minimising the sum of players’ costs [1, 9, 33], as well as for simple routing games [10, 28].
Posted prices have been used for pricing in large markets. Kelso and Crawford [4] and Gul and Stacchetti [46] proved the existence of prices, for gross substitute valuations, that clear the market efficiently. Pricing bundles for combinatorial Walrasian equilibria was introduced by Feldman, Gravin and Lucier [37], who showed that half of the social welfare can be achieved. In a followup work [38], they considered Bayesian combinatorial auctions and they could guarantee half of the optimum welfare, by using anonymous posted prices. Dynamic pricing schemes has been used by Cohen, Eden, Fiat and Jez [31] in several online settings to induce the same performance as the best online algorithm, and by CohenAddad, Eden, Feldman and Fiat [32] in matching markets in order to achieve the optimal social welfare, for any tie breaking rule. For maximizing the revenue with posted price mechanisms see [3, 6, 8, 14, 18–20].
We further discuss some related work to the underlying problems that we consider here, the set cover and the minimum Steiner tree problems. Both problems are very well studied and known to be in NPcomplete. The best known approximations are O(log(k)) [30] (by using a simple greedy algorithm) and 1.39 [15]; in fact, for the set cover problem, Feige [36] showed that no improvement by a constant factor is likely. Research has been done regarding the stochastic model, Grandoni et al. [45] showed a roughly O(log nm) tight bound for the set cover problem and Garg et al. [42] gave bounds on the approximation of the stochastic online Steiner tree problem. A slightly different distribution is the independent activations. Shmoys and Talwar [60] demonstrated randomized and deterministic algorithms with constant approximations for the universal TSP problem, and Christodoulou and Sgouritsa [29] studied the multicast game presenting an ordered protocol with constant PoA.
2 Model
CostSharing Protocol

Stability: The induced game has always a pure (Bayes) Nash equilibrium.

Separability: The cost shares of each resource r are completely determined by the set of players that choose it.

BBiE: In any pure (Bayes) Nash equilibrium profile, the cost shares of the players choosing r should cover exactly the cost of r.
For the rest of the paper, by k we denote the number of players and by n the number of different types of the players, i.e. in the set cover game, U = n, and in the multicast game, V  = n.
Information Models
We study several information models, from the point of view of the designer and of other players, regarding the knowledge of players’ type. A player’s type is some resource: in the set cover game, it is some element from U that needs to be covered, and in the multicast game, it is some vertex of G, on which the player’s terminal lies, and requires connectivity with the root t. The parameters of the game is known to both the protocol designer and the participants. To be more specific, the tuple \((U,\mathscr {F},c)\) in the set cover game and the underlying (weighted) graph in the multicast game are commonly known.

Complete Information: The types of the players are common knowledge, i.e. they are known to all players and to the designer.

Stochastic/A priori: The players’ types are drawn from some product distribution D defined over the type set (U for set cover and V for multicast). The actual types are unknown to the designer, who is only aware of D. However, the players decide their strategies by knowing other players’ types.

Bayesian: The players’ types are drawn from some product distribution D defined over the type sets. Both the designer and the players know only D. The players now decide their strategies by knowing only D and not the actual types. A natural assumption is that every player knows her own type.
We assume that the players’ types are distributed i.i.d. (D = π^{k}) and the type of each player is drawn independently from some probability distribution π : R → [0, 1], with \({\sum }_{r\in R} \pi (r) = 1\); R is either U in the set cover game or V in the multicast game. For simplicity we write π_{r} instead of π(r).
Price of Anarchy (PoA)
3 Lower Bounds for BudgetBalanced Protocols
In this section, we show the lower bounds of budgetbalanced protocols, for the Bayesian setting.
Theorem 1
The Bayesian or stochastic PoA of any budgetbalancedprotocol, for the unweighted set cover game, is Ω(n).
Proof
Theorem 2
The Bayesian PoA of any budgetbalanced protocol, for the multicast game, is \({\Omega }(\sqrt {n})\) .
Proof
4 BBiE Protocols
In this section we drop the requirement of budget balance and instead we consider a more general class of costsharing protocols \(\mathscr {C}\), where the requirement is to preserve the budget balance in the equilibrium. For the rest of the paper, by h we denote a very high value with respect to the parameters of the game. h should be larger than the total costshare of any player by using any budgetbalanced protocol. It is sufficient that \(h > \max _{j} c_{F_{j}}\) for the set cover game and \(h > {\sum }_{e\in E}c_{e}\) for the multicast game. To show our results we will use known oblivious algorithms of the corresponding optimization problems and we will enforce their solution by applying appropriate costsharing protocols (or posted prices in Section 5); e.g. choices, not consistent with this solution, are highly expensive.
In an optimization problem, an oblivious algorithm assigns an action for each input component, based on the prior distribution, and independently of the realization of all other input components. Take as an example the multicast game, where the actions of an input (source) corresponds to the paths connecting the source to the root. An oblivious solution, maps each vertex to some path that connects it to the root, and is used in any realization of the input that contains this source. We associate the types of the players to the input components of the problem, and then we would like to enforce the players to follow the action decided by the oblivious algorithm for their type.
Theorem 3
LetG be any costsharing game and π the underlying optimization resource allocationproblem. Given any oblivious algorithm of π with approximation ratioρ,there exists a costsharing protocol\({\Xi } \in \mathscr {C}\)forG with PoA= O(ρ).
Proof
Suppose that R_{i} is the set of the resources allocated by the oblivious algorithm to the input component that serves as the type of some player i. Even though it is not quite correct, we will say that R_{i} are the resources allocated to player i. Let S_{r} be the set of players to whom resource r is allocated.
Note that Ξ assigns equal shares restricted to S_{r} and a high value h for other players. In fact, instead of equal shares we could use any budgetbalanced protocol restricted to S_{r}, for instance any generalized weighted Shapley protocol (for definition see [44]).
Note that any player i using a resource r ∉ R_{i} should pay h. By the definition of h, this is strictly more than \({\sum }_{r^{\prime }\in R_{i}} c_{r^{\prime }}\), which is the maximum she may pay if she deviates to R_{i}. Therefore, the only Nash equilibria are for each player i to choose some subset of R_{i}. This results to a PoA which is at most the same with the approximation ratio of the optimization problem, so PoA = O(ρ). Moreover, by the construction of Ξ, BBiE holds. □
The following corollaries hold for both the Bayesian and the stochastic setting.
Set Cover Game
Grandoni et al. [45] studied the stochastic problem, and they showed two mapping algorithms for the oblivious set cover problem (one for the unweighted problem which is lengthoblivious and one for the weighted problem which is lengthaware), which are almost O(log mn)competitive.
Theorem 3 implies the following corollary by using the results of [45].
Corollary 1
In the unweighted and weighted set cover game, there existlengthoblivious protocol\({\Xi }_{1} \in \mathscr {C}\)andlengthaware protocol\({\Xi }_{2} \in \mathscr {C}\),respectively, both computed in polynomial time, and with PoA ofO(log n),ifm = poly(n),and\(O\left (\frac {\log m}{\log \log m  \log \log n}\right )\),ifm ≫ n.
Multicast Game
Garg et al. [42] showed a constant approximation on the online Steiner tree problem. The idea is the following: sample a set S from the distribution π^{k} over the vertices and construct a minimum Steiner tree (or a constant approximation). Then connect each other vertex with its nearest vertex from S via shortest path. That way we end up with a spanning tree T (standard derandomization techniques can apply [29, 60, 63]). T defines a single path from each vertex to the root and this is an oblivious strategy for each players’ type. By using Theorem 3 and any constant approximation of the minimum Steiner tree (the best known is by [15]), the following corollary holds.
Corollary 2
In the multicast game, there exists\({\Xi }\in \mathscr {C}\)with PoA= O(1).
5 Posted Prices
In this section, we show how to set anonymous or semianonymous prices for the resources. Expost BBiE cannot be obtained by using anonymous posted prices, as the following example illustrates. Instead, we require exante BBiE. Furthermore, the same example serves to demonstrate that other natural variations of budget balance cannot be very promising: a) BBiE with “high” probability, b) bounded possible excess and deficit. Example 1 indicates that any anonymous posted prices may result in BBiE with probability at most \(O(1/\sqrt {k})\) and that no posted prices can guarantee good bounds on possible excess and deficit, i.e. for any posted prices, there are cases where the total shares for some resource are either at least \(\sqrt {k}\) or at most \(1/\sqrt {k}\) of the resource’s cost.
Example 1
Consider k i.i.d. players whose type is the uniform distribution over two elements e_{1}, e_{2} in set cover or two vertices v_{1}, v_{2} in multicast. In the set cover game, there are only two subset of unit cost, \(\mathscr {F} = \{\{e_{1}\},\{e_{2}\}\}\). In the multicast game there are only two edges, (v_{1}, t) and (v_{2}, t), of unit cost. The question that arises in both cases is how to set a price on a resource r of unit cost, when each player may use it with probability 1/2. Let q be the price for resource r. If 1/q is not an integer in {1, …, k}, then budget balance appears with zero probability. So, suppose that 1/q = k^{′}∈{1, …, k}, then budget balance appears only when k^{′} players use resource r and this happens with probability, \(\text {Pr}[\# \; players = k^{\prime }] = \binom kk^{\prime } \left (\frac 12\right )^{k^{\prime }} \left (1\frac 12\right )^{kk^{\prime }} \leq \binom k {\lfloor k/2 \rfloor } 1/2^{k} < 1/{\sqrt {k}}. \) Furthermore, for any price q for resource r, if \(q \geq 1/\sqrt {k}\) then, in the case that all players use r, the total shares sum up to at least \(k\cdot 1/\sqrt {k} = \sqrt {k}\). On the other hand, if \(q < 1/\sqrt {k}\) then, in the case that only one player uses r, her share is at most \(1/\sqrt {k}\). This means that we cannot guarantee good bounds on any possible excess and deficit.
Set Cover Game
To determine anonymous prices for the unweighted set cover game, we first state Lemma 1 to be used in stability arguments.
Lemma 1
For anya > b > 0 and integerk ≥ 2,\(\frac {a}{1(1a)^{k}} > \frac {b}{1(1b)^{k}}\).
Proof
Theorem 4
In the unweighted set cover game, there exist lengthoblivious and anonymous prices (computed in polynomial time) with PoAO(log n),ifm = poly(n),and\(O\left (\frac {\log m}{\log \log m  \log \log n}\right )\),ifm ≫ n.
Proof
We first argue that there exists a unique BayesNash equilibrium, where each player i chooses the set picked earlier by Algorithm 3 and covers her. For that it is sufficient to show that for any two sets A and B, such that \({\sum }_{i\in A} \pi _{i} > {\sum }_{i\in B} \pi _{i}\), \({k_{A}^{1}} > {k_{B}^{1}}\); so, if some player is covered by both A and B, the price set for A should be less than the price set for B and the player prefers A that is picked by Algorithm 3 before B.
By using (1), we need to show that \( \frac {k{\sum }_{i\in A} \pi _{i}}{1\left (1{\sum }_{i\in A} \pi _{i}\right )^{k}} > \frac {k{\sum }_{i\in B} \pi _{i}}{1\left (1{\sum }_{i\in B} \pi _{i}\right )^{k}},\) which is true for k ≥ 2, due to Lemma 1 by setting \(a={\sum }_{i\in A} \pi _{i}\) and \(b={\sum }_{i\in B} \pi _{i}\). Note that for k = 1, there exists only one player in the game and this is a trivial case to solve.
Next notice that, given that a set F is chosen by some player, the expected number of players paying for it is \({k_{F}^{1}}\), resulting in exante BBiE. As for the PoA, Grandoni et al. [45] analyzed the performance of Algorithm 3, for the stochastic problem. They didn’t consider any prices, instead they mapped each player to the first set considered by the algorithm and they used the mapping in order to form a set cover. Their cover though coincides with the equilibrium solution and therefore their results immediately provide bounds on the PoA. □
Theorem 5
In theweighted set cover game, there existlengthaware andsemianonymous prices (computed in polynomial time) with PoAO(log n),ifm = poly(n),and\(O\left (\frac {\log nm}{\log \log m  \log \log n}\right )\),ifm ≫ n.
Proof
By using the mapping of Algorithm 2 (Algorithm 2 of [45]), let S(F) be the set of elements mapped to set F. For each set \(F \in \mathscr {F}\), set the price to be \(\frac {c_{F}}{k_{S(F)}^{1}}\), for any player whose type is in S(F), and let the price be h for any other player. Such way, in the BayesNash equilibrium, each player chooses the set, to which she is mapped by Algorithm 2. Grandoni et al. [45] showed that the expected cost of that mapping is O(log n), if m = poly(n), and \(O\left (\frac {\log m}{\log \log m  \log \log n}\right )\), if m ≫ n, away from the expected cost of the optimum solution. Those also serve as upper bounds on the PoA of the induced game. Finally note that those prices satisfy exante BBiE, for the same reasons as the unweighted case. □
We complement our results by providing tight lower bounds for polytime determinable prices.
Proposition 1
Form = poly(n),there maynot exist anonymous posted prices for theunweighted set cover, or semianonymous posted pricesfor the weighted set cover, computed in polytime, withPoA= o(log n),unlessNP ⊆ DTIME(n^{O(log log n)}).
Proof
For any set cover problem \({\Pi } = (U,\mathscr {F}, c)\), we consider the game G with U players where each one is associated with a different element. Consider the stochastic or Bayesian game, where k ≫ n, and k is sufficiently large such that the probability that each element is the type of some player converges to 1. Then, we apply the prices on the stochastic G. It is easy to see that we can compute a Nash equilibrium in polynomial time, O(nm); players choose, among the sets that covers them, some with minimum price. All the chosen sets define a set cover for U. If there exist posted prices computed in polynomial time with PoA= o(log nm) = o(log n), this would imply a polynomial time algorithm for the setcover problem with approximation ratio o(log n). However, by [36], no polynomial time algorithm for the set cover problem can approximate the optimal solution by o(log k), unless NP ⊆ DTIME(n^{O(log log n)}), which results in a contradiction. □
Proposition 2
Form ≫ n,there maynot exist anonymous prices for unweighted setcover, or semianonymous prices for weighted set cover, withPoA\(=o\left (\frac {\log m}{\log \log m  \log \log n}\right )\).
Proof
On the contrary, suppose that such prices exist. Then, they would determine a mapping from the elements to the sets, meaning each element is covered by a specific set. The expected cost of the sets that cover the sampled elements would be \(o\left (\frac {\log m}{\log \log m  \log \log n}\right )\) away from the expected cost of the optimum solution. This contradicts the lower bound given by [45] (Theorem 4.2). □
Multicast Game
We construct a spanning tree T in the same way as in Section 4 and we use it to set the posted prices (computed in polynomial time).
Theorem 6
In the multicast game, there exist anonymous posted prices withPoA= O(1).
Proof
For each edge e ∈ E(T), let V (e) be the set of vertices that are disconnected from the root t in T ∖{e}. We set the price for each e ∈ E(T) as \(c_{e}/k_{V(e)}^{1}\). For each e ∉ E(T), the price is set to h. In the equilibrium each player chooses the path that connects her terminal with t via T. The constant PoA follows by [42] and the approximation of [15]. The expected total prices for e ∈ E(T) is \(k_{V(e)}^{1} c_{e}/k_{V(e)}^{1} = c_{e}\), if e is used, and 0 otherwise, resulting in exante BBiE. □
6 PriorIndependent Mechanisms
The design of priorindependent mechanisms is a more difficult task, as the objective now is to identify a single mechanism that always has good performance, under any distributional assumption. In this section, we show limitations of priorindependent mechanisms even for the restricted class of i.i.d. prior distributions.
BBiE Protocols
Satisfying BBiE with priorindependent protocols highly restricts the class of costsharing protocols and seems hard for natural classes of distribution, e.g. i.i.d., to find expost BBiE protocols with low PoA. Regarding the weighted set cover game with i.i.d. distributions, we can construct a lower bound of \({\Omega }(\sqrt {n})\) for all priorindependent mechanisms, which are expost BBiE.
Theorem 7
In the weighted set cover game, any priorindependent, expost BBiE protocol \({\Xi } \in \mathscr {C}\) has PoA \(={\Omega }(\sqrt {n})\) .
Proof
Consider n players, n + 1 elements/types U = {0, 1, …, n} and the family of sets \(\mathscr {F}=\{F_{0}, F_{1}, {\ldots } F_{n}, F_{all}\}\), with F_{j} = {j}, \(c_{F_{j}}= 1\) for all j, and F_{all} = {1, …, n}, \(c_{F_{all}}=\sqrt {n}\). Note that 0 is covered only by F_{0}, serving as a dummy set.
Given a BBiE, priorindependent protocol Ξ, suppose that there exists some F_{j}, j ≠ 0, where Ξ is not budgetbalanced, i.e. there exists a set of players S, such that if only S chooses F_{j}, the sum of their costshares are different from 1. Consider the prior distribution D_{1} = π^{n} with π(0) = π(j) = 1/2 and π(j^{′}) = 0 for any j^{′}∉{0, j}. With positive probability, 1/2^{n}, all player of S have type j and all other players have type 0. If all players of S choose F_{j} in any pure BayesNash equilibrium, expost BBiE is violated. So, there exists at least one player from S such that, whenever her type is j, she chooses F_{all} (and this happens with probability 1/2) which results in PoA\(={\Omega }(\sqrt {n})\).
Suppose now that Ξ is budgetbalanced for any F_{j}, where j ≠ 0. Let I be the set of players such that whenever i ∈ I is the only player choosing F_{all}, Ξ doesn’t charge \(\sqrt {n}\) to i. Consider the prior distribution D_{2} = π^{n} with π(0) = 1/2 and π(j) = 1/2n for all other j. With positive probability, 1/(2^{n}n), player i’s type is some j ≠ 0 and all other players’ type is 0. If for any type j ≠ 0 player i chooses F_{all} in any BayesNash equilibrium, expost BBiE is violated. Therefore, for any player i ∈ I, whenever her type is j, she chooses F_{j}.
We claim that the strategy profile, where any player i with type t_{i} chooses \(F_{t_{i}}\) is a BayesNash equilibrium. For any player i ∈ I there is no other valid strategy. For each player i ∉ I, whenever t_{i} ≠ 0, player i always pays at most 1 (due to budget balance in \(F_{t_{i}}\)), whereas if she deviates to F_{all} she pays \(\sqrt {n}\).
Each element j ≠ 0 is a type of at least one player with probability \(1\left (1\frac 1{2n}\right )^{n}\geq 1\frac 2e\), giving an expected cost of Ω(n) in the equilibrium. The expected optimum is at most \(1+\sqrt {n}\) by using only F_{0} and F_{all} and so PoA\(={\Omega }(\sqrt {n})\). □
Posted Prices
Setting priorindependent posted prices cannot guarantee any BBiE, even exante. Consider the set cover game (similar example exists for the multicast game) with n players, n elements and two subsets of unit costs, one containing element 1 and the other containing the rest. Suppose now that we post a price q for the first subset. If \(q \leq 1/\sqrt {n}\), for the uniform prior distribution, the expected number of players with type 1, given that there exists at least one, is \(\frac {n\cdot 1/n}{1(11/n)^{n}} \leq \frac {e}{e1}\). The expected cost shares for the first set are \(O (1/\sqrt {n})\), meaning that its cost is undercovered by a factor of \({\Omega }(\sqrt {n})\). If \(q > 1/\sqrt {n}\), consider the prior D = π^{n}, where π(1) = 1 and π(j) = 0 for all j ≠ 1. All players choose the first set and their total shares are \(n \cdot 1/\sqrt {n} = \sqrt {n}\) which exceeds the set’s cost by a factor of \(\sqrt {n}\). So, there is no way to avoid an over/undercharge of a resource by a factor better than \({\Theta }(\sqrt {n})\).
7 Complete Information
In the complete information setting, the input now is known and therefore for any feasible solution we can consider oblivious strategies. If \(\mathscr {F}^{\prime }\subseteq \mathscr {F}\) is any feasible solution with cost \(c(\mathscr {F}^{\prime })\) and \(\mathscr {F}^{*}\subseteq \mathscr {F}\) is the optimum solution with cost \(c(\mathscr {F}^{*})\), then by Theorem 3 we can enforce the solution \(\mathscr {F}^{\prime }\) and get a PoA of \(c(\mathscr {F}^{\prime })/c(\mathscr {F}^{*})\). In the following we consider \(\mathscr {F}^{\prime }\) as either the optimum solution or an approximation. Note that, while trying to bound the PoA, computational issues are not of primary concern.
Set Cover Game
By considering \(\mathscr {F}^{\prime }\), as either the optimum solution or its O(log k)approximation of the greedy algorithm, where k is the number of players to be covered, and by using Theorem 3 we get the following corollary.
Corollary 3
For the (weighted) set cover game there exists a protocol\({\Xi } \in \mathscr {C}\),that can be defined inexponential time, with PoA= 1 and a protocol\({\Xi } \in \mathscr {C}\),that can be defined inpolynomial time, with PoA= O(log k).
Next we show that there exist posted prices that can be computed in polynomial time with PoA = O(log k). We show that, under the restriction of setting the prices in polynomial time, this bound is tight. Then we drop the constrain of defining the prices in polynomial time, and we define posted prices with PoA = 1.
Theorem 8
There exist posted prices satisfying BBiE, that can be set inpolynomial time, for the (weighted) set cover game, with PoA= O(log k).
Proof
There exists a unique Nash equilibrium, where each player i chooses the set picked earlier by Algorithm 4 and covers her, let it be F^{i}. This is because any other set F that covers i and picked after F^{i} should have at least the same value \(\frac {c(F)}{{\sum }_{e \in F\cap X} k_{e}}\) with \(\frac {c(F^{i})}{{\sum }_{e \in F^{i}\cap X} k_{e}}\) at the time that F^{i} was picked. After processing F^{i}, F ∩ X is diminished by at least 1, and so the price of F should be strictly higher that the price of F^{i}.
In order to show the bound on the PoA, we consider a reduction from the set cover game \(G=(U, X,\mathscr {F},c,\mathbf {t})\), where t are players’ types, to a set cover problem \({\Pi } = (U^{\prime },\mathscr {F}^{\prime },c^{\prime })\), such that the PoA equals the approximation ratio of π. U^{′} is derived by X, after replacing each of its elements, e, by k_{e} copies \(e_{1}, \ldots , e_{k_{e}}\). For each set \(F \in \mathscr {F}\), we construct a set \(F^{\prime } \in \mathscr {F}^{\prime }\) of the same cost \((c_{F}=c^{\prime }_{F^{\prime }})\), first by erasing all elements belonging to U ∖ X, and then by replacing each of its remaining elements, e, by k_{e} copies \(e_{1}, \ldots , e_{k_{e}}\). We assume that each player i ∈{1, …, k_{e}} of type e is associated with element e_{i} of the constructed set cover problem.
It is easy to see that the greedy algorithm on π chooses the sets in the same order with Algorithm 4. Therefore, the approximation ratio for π equals the PoA of G. Notice that U^{′} = k and since the approximation ratio of the greedy is O(log k), the bound on the PoA follows. Note further that the sum of the prices for each such set used in the Nash equilibrium equals the cost of the set, that results in BBiE as desired. □
Proposition 3
There maynot exist posted prices, that are computedin polynomial time, for the set cover game, with PoA= o(log k),unlessNP ⊆ DTIME(n^{O(log log n)}).
Proof
On the contrary, suppose there exist posted prices computed in polynomial time with PoA= o(log k). This would impply a polynomial time algorithm for the set cover problem with approximation ratio of o(log k), but this is a contradiction due to [36]. □
Theorem 9
There exist posted prices satisfying BBiE, computed inexponential time, for the (weighted) setcover game, with PoA= 1.
Proof
As in the proof of Theorem 8, let k_{e} be the number of players with type e ∈ U and X ⊆ U be the set of elements/players that need to be covered. Moreover, let \(\mathscr {F}^{*}\subseteq \mathscr {F}\) be the optimum solution, found in exponential time. To set the prices we run Algorithm 4 but for input \((X,\mathscr {F}^{*})\). We set the prices for the rest of the sets \(\mathscr {F}\setminus \mathscr {F}^{*}\) equal to h.
By using similar arguments as in the proof of Theorem 8, in the (unique) Nash equilibrium, each player chooses the set picked earlier by Algorithm 4 and covers her. The prices for each set used in the Nash equilibrium equal the cost of the set resulting in BBiE. The difference here is that the Nash equilibrium uses only the sets of \(\mathscr {F}^{*}\), resulting in PoA = 1. □
Multicast Game
Similarly with the set cover game, we can easily get the following corollary by using Theorem 3, for the multicast game. For the second part, we use the 1.39approximation algorithm of [15].
Corollary 4
For the multicast game there exists a costsharing protocol\({\Xi } \in \mathscr {C}\),that can be defined inexponential time, with PoA= 1 and a costsharing protocol\({\Xi } \in \mathscr {C}\),that can be defined inpolynomial time, with PoA≤ 1.39.
We next use posted prices and show that the PoA is constant for the case of multicast game. By using the 1.39approximation algorithm of [15], the PoA is constant even if we require the prices to be computed in polynomial time.
Theorem 10
For the multicast game, there exist postedprices, computed in polynomial time, with PoA≤ 1.39 and posted prices, computed in exponential time, with PoA= 1.
Proof
Let S ⊆ V be the set of players’ terminal and T be the solution that approximates the minimum Steiner tree on the requested vertices S ∪{t} derived by the 1.39approximation algorithm of [15]. If we drop the requirement of computing the prices in polynomial time, T is the minimum Steiner tree. For each edge e ∈ E(T), let k_{e} be the number of players that are disconnected from t in T ∖{e}. We set the price for e ∈ E(T) as c_{e}/k_{e}. For each e ∉ E(T), we set the price to be h. In the Nash equilibrium each player will choose the unique path that connects her source with t in T, since any other path has high cost of at least h. Obviously, the players cover exactly the cost of each used edge and the cost of the Nash equilibrium equals the cost of T. □
Footnotes
 1.
In expost budget balance we require budget balance in every realization of the game. If the expected excess and deficit are zero, the budget balance is called exante.
Notes
Acknowledgements
Part of this work was done while all authors were visiting the Economics and Computation program of the Simons Institute for the Theory of Computing. The work of G. Christodoulou was supported by EPSRC grants EP/M008118/1, EP/K01000X/1 and Royal Society Leverhulme Trust Senior Research Fellowship LT140046. The work of S. Leonardi was partially supported by Google Focused award “Algorithms for Largescale Data analysis” and EU FET project no. 317532 “Multiplex”.
References
 1.Abed, F., Correa, J.R., Huang, C.C.: Optimal coordination mechanisms for multijob scheduling games. In: Proceedings of the 22th Annual European Symposium on Algorithms  ESA 2014, September 8–10, 2014, pp. 13–24. Wroclaw (2014)Google Scholar
 2.Abed, F., Huang, C.C.: Preemptive coordination mechanisms for unrelated machines. In: Proceedings of the 20th Annual European Symposium on Algorithms  ESA 2012, September 10–12, 2012, pp. 12–23. Ljubljana (2012)Google Scholar
 3.Adamczyk, M., Borodin, A., Ferraioli, D., de Keijzer, B., Leonardi, S.: Sequential posted price mechanisms with correlated valuations. In: Proceedings of the 11th International Conference on Web and Internet Economics  WINE 2015, December 9–12, 2015, pp. 1–15. Amsterdam (2015)Google Scholar
 4.Crawford, V.P., Kelso, A.S.: Job matching, coalition formation, and gross substitutes. Econometrica 50(6), 1483–1504 (1982)MathSciNetCrossRefzbMATHGoogle Scholar
 5.Anshelevich, E., Dasgupta, A., Kleinberg, J.M., Tardos, É., Wexler, T., Roughgarden, T.: The price of stability for network design with fair cost allocation. SIAM J. Comput. 38(4), 1602–1623 (2008)MathSciNetCrossRefzbMATHGoogle Scholar
 6.Anshelevich, E., Kar, K., Sekar, S.: Envyfree pricing in large markets approximating revenue and welfare. In: Proceedings of the 42nd International Colloquium on Automata, Languages, and Programming  ICALP 2015, July 6–10, 2015, Part I, pp. 52–64. Kyoto (2015)Google Scholar
 7.Azar, Y., Jain, K., Vahab, M.: (Almost) optimal coordination mechanisms for unrelated machine scheduling. In: Proceedings of the Nineteenth Annual ACMSIAM Symposium on Discrete Algorithms, SODA ’08, pp 323–332. Society for Industrial and Applied Mathematics, Philadelphia (2008)Google Scholar
 8.Babaioff, M., Blumrosen, L., Dughmi, S., Singer, Y.: Posting prices with unknown distributions. In: Proceedings of the Innovations in Computer Science  ICS 2010, January 7–9, 2011, pp. 166–178. Tsinghua University, Beijing (2011)Google Scholar
 9.Bhattacharya, S., Im, S., Kulkarni, J., Munagala, K.: Coordination mechanisms from (almost) all scheduling policies. In: Innovations in Theoretical Computer Science, ITCS’14, January 1214 2014, pp. 121–134. Princeton (2014)Google Scholar
 10.Bhattacharya, S., Kulkarni, J., Mirrokni, V.S. : Coordination mechanisms for selfish routing over time on a tree. In: Proceedings of the 41st International Colloquium on Automata, Languages, and Programming  ICALP 2014, July 8–11 2014, Part I, pp. 186–197. Copenhagen (2014)Google Scholar
 11.Biló, V., Bove, R.: Bounds on the price of stability of undirected network design games with three players. Journal of Interconnection Networks 12(12), 1–17 (2011)CrossRefGoogle Scholar
 12.Biló, V., Caragiannis, I., Fanelli, A., Monaco, G.: Improved lower bounds on the price of stability of undirected network design games. Theory Comput. Syst. 52(4), 668–686 (2013)MathSciNetCrossRefzbMATHGoogle Scholar
 13.Biló, V., Flammini, M., Moscardelli, L.: The price of stability for undirected broadcast network design with fair cost allocation is constant. In: 54th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2013, 2629 October, 2013, pp. 638–647. Berkeley (2013)Google Scholar
 14.Blumrosen, L., Holenstein, T.: Posted prices vs. negotiations: an asymptotic analysis. In: Proceedings 9th ACM conference on Electronic Commerce (EC2008), June 812, 2008, pp. 49. Chicago (2008)Google Scholar
 15.Byrka, J., Grandoni, F., Rothvoß, T., Sanitá, L.: An improved lpbased approximation for steiner tree. In: Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010, 58 June 2010, pp. 583–592. Cambridge (2010)Google Scholar
 16.Caragiannis, I.: Efficient coordination mechanisms for unrelated machines scheduling. In: SODA, pp 815–824 (2009)Google Scholar
 17.Charikar, M., Karloff, H.J., Mathieu, C., Naor, J., Saks, M.E.: Online multicast with egalitarian cost sharing. In: SPAA, pp 70–76 (2008)Google Scholar
 18.Chawla, S., Hartline, J.D., Kleinberg, R.D. : Algorithmic pricing via virtual valuations. In: Proceedings 8th ACM conference on Electronic Commerce (EC2007), June 1115, 2007, pp. 243–251. San Diego (2007)Google Scholar
 19.Chawla, S., Hartline, J.D., Malec, D.L., Sivan, B.: Multiparameter mechanism design and sequential posted pricing. In: Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010, Cambridge, 58 June 2010, pp 311–320 (2010)Google Scholar
 20.Chawla, S., Malec, D.L., Sivan, B.: The power of randomness in bayesian optimal mechanism design. Games and Economic Behavior 91, 297–317 (2015)MathSciNetCrossRefzbMATHGoogle Scholar
 21.Chekuri, C., Chuzhoy, J., LewinEytan, L., Naor, J., Orda, A.: Noncooperative multicast and facility location games. IEEE J. Sel. Areas Commun. 25(6), 1193–1206 (2007)CrossRefGoogle Scholar
 22.Chen, H.L., Roughgarden, T.: Network design with weighted players. Theory Comput. Syst. 45(2), 302–324 (2009)MathSciNetCrossRefzbMATHGoogle Scholar
 23.Chen, H.L., Roughgarden, T., Valiant, G.: Designing network protocols for good equilibria. SIAM J. Comput. 39(5), 1799–1832 (2010)MathSciNetCrossRefzbMATHGoogle Scholar
 24.Christodoulou, G., Chung, C., Ligett, K., Pyrga, E., van Stee, R.: On the price of stability for undirected network design. In: Proceedings of the 7th International Workshop on Approximation and Online Algorithms  WAOA 2009, September 10–11, 2009, Revised Papers, pp. 86–97. Copenhagen (2009)Google Scholar
 25.Christodoulou, G., Koutsoupias, E., Nanavati, A.: Coordination mechanisms. Theor. Comput. Sci. 410(36), 3327–3336 (2009)MathSciNetCrossRefzbMATHGoogle Scholar
 26.Christodoulou, G., kovács, A., Schapira, M: Bayesian combinatorial auctions. J. ACM 63(2), 11:1–11:19 (2016)MathSciNetCrossRefzbMATHGoogle Scholar
 27.Christodoulou, G., Leonardi, S., Sgouritsa, A.: Designing costsharing methods for bayesian games. In: Proceedings of the 9th International Symposium on Algorithmic Game Theory  SAGT 2016, September 19–21, 2016, pp. 327–339. Liverpool (2016)Google Scholar
 28.Christodoulou, G., Mehlhorn, K., Pyrga, E.: Improving the price of anarchy for selfish routing via coordination mechanisms. Algorithmica 69, 619–640 (2014)MathSciNetCrossRefzbMATHGoogle Scholar
 29.Christodoulou, G., Sgouritsa, A.: Designing networks with good equilibria under uncertainty. In: Proceedings of the Twentyseventh Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2016, January 1012, 2016, pp. 72–89. Arlington (2016)Google Scholar
 30.Chvatal, V.: A greedy heuristic for the setcovering problem. Math. Oper. Res. 4(3), 233–235 (1979)MathSciNetCrossRefzbMATHGoogle Scholar
 31.Cohen, I.R., Eden, A., Fiat, A., Jez, L.: Pricing online decisions: beyond auctions. In: Proceedings of the Twentysixth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2015, January 46, 2015, pp. 73–91. San Diego (2015)Google Scholar
 32.CohenAddad, V., Eden, A., Feldman, M., Fiat, A.: The invisible hand of dynamic market pricing. CoRR, arXiv:1511.05646 (2015)
 33.Cole, R., Correa, J.R., Gkatzelis, V., Mirrokni, V.S., Olver, N.: Inner product spaces for minsum coordination mechanisms. In: Proceedings of the 43rd ACM Symposium on Theory of Computing, STOC, 68 June 2011, pp. 539–548. San Jose (2011)Google Scholar
 34.Devanur, N.R., Mihail, M., Vazirani, V.V.: Strategyproof costsharing mechanisms for set cover and facility location games. Decis. Support. Syst. 39(1), 11–22 (2005)CrossRefGoogle Scholar
 35.Disser, Y., Feldmann, A.E., Klimm, M., Mihalák, M.: Improving the h _{k}bound on the price of stability in undirected shapley network design games. In: Proceedings of the 8th International Conference on Algorithms and Complexity  CIAC 2013, May 22–24, 2013, pp. 158–169. Barcelona (2013)Google Scholar
 36.Feige, U.: A threshold of ln n for approximating set cover. J. ACM 45(4), 634–652 (1998)MathSciNetCrossRefzbMATHGoogle Scholar
 37.Feldman, M., Gravin, N., Lucier, B.: Combinatorial walrasian equilibrium. In: Symposium on Theory of Computing Conference, STOC’13, June 14, 2013, pp. 61–70. Palo Alto (2013)Google Scholar
 38.Feldman, M., Gravin, N., Lucier, B: Combinatorial auctions via posted prices. In: Proceedings of the Twentysixth annual ACMSIAM Symposium on Discrete Algorithms, SODA 2015, January 46, 2015, pp. 123–135. San Diego (2015)Google Scholar
 39.Fiat, A., Kaplan, H., Levy, M., Olonetsky, S., Shabo, R.: On the price of stability for designing undirected networks with fair cost allocations. In: Proceedings of the 33rd International Colloquium on Automata, Languages and Programming, ICALP 2006, July 10–14, 2006, Part I, pp. 608–618. Venice (2006)Google Scholar
 40.Gairing, M., Kollias, K., Kotsialou, G.: Tight bounds for costsharing in weighted congestion games. In: Proceedings of the 42nd International Colloquium on Automata, Languages, and Programming  ICALP 2015, July 6–10, 2015, Part II, pp. 626–637. Kyoto (2015)Google Scholar
 41.Gairing, M., Kollias, K., Kotsialou, G.: Costsharing in generalised selfish routing. In: Proceedings of the 10th International Conference on Algorithms and Complexity  CIAC 2017, May 24–26, 2017, pp. 272–284. Athens (2017)Google Scholar
 42.Garg, N., Gupta, A., Leonardi, S., Sankowski, P.: Stochastic analyses for online combinatorial optimization problems. In: Proceedings of the Nineteenth Annual ACMSIAM Symposium on Discrete Algorithms, SODA 2008, January 2022, 2008, pp. 942–951. San Francisco (2008)Google Scholar
 43.Gkatzelis, V., Kollias, K., Roughgarden, T.: Optimal costsharing in weighted congestion games. In: Proceedings of the 10th International Conference on Web and Internet Economics  WINE 2014, December 14–17, 2014, pp. 72–88. Beijing (2014)Google Scholar
 44.Gopalakrishnan, R., Marden, J.R., Wierman, A.: Potential games are necessary to ensure pure nash equilibria in cost sharing games. In: EC, pp 563–564 (2013)Google Scholar
 45.Grandoni, F., Gupta, A., Leonardi, S., Miettinen, P., Sankowski, P., Singh, M.: Set covering with our eyes closed. SIAM J. Comput. 42(3), 808–830 (2013)MathSciNetCrossRefzbMATHGoogle Scholar
 46.Gul, F., Stacchetti, E.: Walrasian equilibrium with gross substitutes. J. Econ. Theory 87(1), 95–124 (1999)MathSciNetCrossRefzbMATHGoogle Scholar
 47.Harks, T., Huber, A., Surek, M.: A characterization of undirected graphs admitting optimal cost shares. CoRR, arXiv:1704.01983 (2017)
 48.Harsanyi, J.C.: Games with incomplete information played by bayesian players, iiii. part ii. bayesian equilibrium points. Manag. Sci. 14(5), 320–334 (1968)CrossRefzbMATHGoogle Scholar
 49.Immorlica, N., Li, L., Mirrokni, V.S., Schulz, A.: Coordination mechanisms for selfish scheduling. In: First International Workshop on Internet and Network Economics (WINE), pp 55–69 (2005)Google Scholar
 50.Immorlica, N., Mahdian, M., Mirrokni, V.S. : Limitations of crossmonotonic costsharing schemes. ACM Trans. Algorithms 4(2), 24:1–24:25 (2008)MathSciNetCrossRefzbMATHGoogle Scholar
 51.Kelly, F.: Charging and rate control for elastic traffic. Eur. Trans. Telecomm. 8(1), 33–37 (1997)CrossRefGoogle Scholar
 52.Klimm, M., Schmand, D.: Sharing nonanonymous costs of multiple resources optimally. In: Proceedings of the 9th international conference on Algorithms and Complexity  CIAC 2015, May 20–22, 2015, pp. 274–287. Paris (2015)Google Scholar
 53.Kollias, K.: Nonpreemptive coordination mechanisms for identical machine scheduling games. In: SIROCCO, pp 197–208 (2008)Google Scholar
 54.Kollias, K., Roughgarden, T.: Restoring pure equilibria to weighted congestion games. ACM Trans. Econ. Comput. 3(4), 21:1–21:24 (2015)MathSciNetCrossRefzbMATHGoogle Scholar
 55.Koutsoupias, E., Papadimitriou, C.H.: Worstcase equilibria. In: STACS, pp 404–413 (1999)Google Scholar
 56.Li, J.: An o(log(n)/log(log(n))) upper bound on the price of stability for undirected shapley network design games. Inf. Process. Lett. 109(15), 876–878 (2009)CrossRefzbMATHGoogle Scholar
 57.Moulin, H., Shenker, S.: Strategyproof sharing of submodular costs:budget balance versus efficiency. Econ. Theory 18(3), 511–533 (2001)MathSciNetCrossRefzbMATHGoogle Scholar
 58.Roughgarden, T.: The price of anarchy in games of incomplete information. ACM Trans. Econ. Comput. 3(1), 6 (2015)MathSciNetCrossRefGoogle Scholar
 59.Roughgarden, T., Schrijvers, O.: Network costsharing without anonymity. ACM Trans. Econ. Comput. 4(2), 8:1–8:24 (2016)MathSciNetCrossRefzbMATHGoogle Scholar
 60.Shmoys, D.B., Talwar, K.: A constant approximation algorithm for the a prioritraveling salesman problem. In: Proceedings of the 13th International Conference on Integer Programming and Combinatorial Optimization, IPCO 2008, May 26–28, 2008, pp. 331–343. Bertinoro (2008)Google Scholar
 61.Syrgkanis, V.: Price of stability in games of incomplete information. CoRR, arXiv:abs/1503.03739 (2015)
 62.von Falkenhausen, P., Harks, T.: Optimal cost sharing for resource selection games. Math. Oper Res. 38(1), 184–208 (2013)MathSciNetCrossRefzbMATHGoogle Scholar
 63.Williamson, D.P., van Zuylen, A.: A simpler and better derandomization of an approximation algorithm for single source rentorbuy. Oper. Res. Lett. 35(6), 707–712 (2007)MathSciNetCrossRefzbMATHGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.