Coalitional stochastic stability in games, networks and markets
Ryoji Sawa∗ Department of Economics, University of Wisconsin-Madison
November 24, 2011
Abstract This paper examines a dynamic process of unilateral and joint deviations of agents and the resulting stochastic evolution of social conventions in a class of interactions that includes normal form games, network formation games, and simple exchange economies. Over time agents unilater- ally and jointly revise their strategies based on the improvements that the new strategy profile offers them. In addition to the optimization process, there are persistent random shocks on agents utility that potentially lead to switching to suboptimal strategies. Under a logit specification of choice prob- abilities, we characterize the set of states that will be observed in the long-run as noise vanishes. We apply these results to examples including certain potential games and network formation games, as well as to the formation of free trade agreements across countries.
Keywords: Logit-response dynamics; Coalitions; Stochastic stability; Network formation.
JEL Classification Numbers: C72, C73.
∗Address: 1180 Observatory Drive, Madison, WI 53706-1393, United States., telephone: 1-608-262-0200, e-mail: [email protected]. The author is grateful to William Sandholm, Marzena Rostek and Marek Weretka for their advice and sugges- tions. The author also thanks seminar participants at University of Wisconsin-Madison for their comments and suggestions.
1 1 Introduction
Our economic and social life is often conducted within a group of agents, such as people, firms or countries. For example, firms may form an R & D alliances and found a joint research venture rather than independently conducting R & D. When forming an alliance, two or more firms make a joint decision, or simultaneously make decisions while taking into account other firms’ decisions. In these settings, it is sometimes more appropriate to use a stronger solution concept than Nash equilibrium, one that accounts for joint deviations by groups of agents. In this paper, we introduce an evolutionary equilibrium selection approach for solution concepts stronger than Nash equilibrium, solution concepts that account for coalitional deviations. The importance of stronger solution concepts has been emphasized in recent years due to a grow- ing interest in social and economic networks. This is because Nash equilibrium has relatively weak predictive power in games on networks, particularly network formation games. Examples of such games are co-author networks (e.g. Jackson and Wolinsky (1996)), R & D networks among firms (e.g. Goyal and Moraga-Gonzalez´ (2001)), and trade agreement networks among countries (e.g. Furu- sawa and Konishi (2007)). When link formation between two agents requires both agents’ consent, Nash equilibrium fails to capture the fact that it may be beneficial for two agents to form a link. For instance, it is always a Nash equilibrium for no player to consent to any link, resulting in the empty network. It is thus natural in this context to use a stronger solution concept that is robust against deviations by pairs of agents as in the notion of pairwise stable equilibrium. Once joint deviations are admissible, it is reasonable to consider some settings which are not described by games because some unilateral deviations may be infeasible. For instance, consider a Gale-Shapley marriage problem. Suppose that an individual’s strategy is to choose whom to marry, and that a man and a woman will be married if and only if both agree to do so. This implies that
the set of feasible strategy profiles is restricted: letting si denote individual i’s strategy, man m can choose sm = w if and only if woman w chooses sw = m. We call these generalizations of normal-form games interactions. Even when we apply stronger solution concepts to interactions, we may still face the problem of multiple equilibria. For example, in the Gale-Shapley marriage problem, any stable matching is a pairwise stable equilibrium. The average number of stable matchings grows more than proportion- ally as the population increases, making it difficult to predict which will emerge. To address this issue and single out one equilibrium, we apply the concept of stochastic stability. In the stochastic stability approach, we embed a static interaction in a dynamic process in which players revise their strategies based on the improvement that the new strategies offer them relative to the current strategies. We perform a stability test by adding stochastic noise to agents’ strategy choices in the above dynamic and studying the long-run outcomes as noise vanishes. This approach has been well developed to select an equilibrium among the set of Nash equilibria.1 However, despite the importance of decentralized but possibly joint decisions in many social and economic settings, the development of frameworks for analyzing stochastic stability in this setting is in its infancy. A pioneering paper by Jackson and Watts (2002) studied pairwise stochastically stable networks that are immune against random shocks in two players’ decisions. They restricted their
1See Chapter 12 of Sandholm (2010) for an overview.
2 attention to network formation games. The present paper follows a similar line to Jackson and Watts (2002), but extends the range of applications in three directions; (i) applying the stochastic stability approach to a broader class of environments, (ii) extending pairwise stochastic stability to stochastic stability with respect to an arbitrary set of feasible coalitions, and (iii) employing the logit-response dynamic, under which more serious mistakes are less likely to occur. Point (i) allows us to study stochastically stability in a broader class of environments. We define an interaction denoted by I = N, (Si, ui)i∈N , S with player set N, strategy sets Si, payoff functions ui and the set of feasible strategy profiles S ⊆ S = ∏i∈N Si. The restriction on strategy profiles allows us to consider settings where some unilateral deviations are prohibited or infeasible. For example, in an exchange economy, a consumer cannot unilaterally change her allocation, since doing so requires another consumer who is willing to exchange with her. Point (ii) allows us to apply our model in environments where more than two agents may form a coalition. Let R be a family of sets of players, representing the collection of feasible coalitions. We define an R-stable equilibrium of interaction I to be a feasible strategy profile which is immune against deviations by any coalition in R. Suppose for instance that S = S, so that I is a normal form game. Then if R is the family of singletons, R-stable equilibrium corresponds to Nash equilibrium; if R is the family of singletons and pairs, it corresponds to pairwise stable equilibrium; and if R is the family of all nonempty sets of players, it corresponds to strong equilibrium. If not all strategy profiles are feasible, then neither are all unilateral deviations, so some choice of R is necessary to define a meaningful notion of equilibrium. Point (iii) allows us to apply stochastic stability in settings where agents, while sometimes mak- ing mistakes, assign smaller probabilities to actions which deliver smaller payoffs. As an example, consider the probability that people send an email to unintended recipients. People will be more careful, and so make such a mistake with lower probability, when their email includes confidential information, e.g. a strategic move of their corporation. Our interest is to characterize strategy profiles that are stochastically stable against not only indi- vidual deviations but also coalitional deviations. We introduce a dynamic process in which players randomly form coalitions and revise their strategies in the following way. First, a new strategy profile for them is randomly proposed, and assessed by each agent. In the unperturbed updating process, an agent agrees to the new strategy profile if it yields her higher payoff than the original strategy profile. The new strategy profiles is accepted if all agents in the coalition agree. To this unperturbed dynamic process, we add stochastic noise that leads agents to agree to a new strategy profile accord- ing to the logit choice rule. Thus, coalitional deviations sometimes occur even if not all members of the coalition benefit, and this occurs with a probability that declines in the total cost to coalition members. As a consequence, the stochastic process visits every strategy profile repeatedly, and pre- dictions can be made concerning the relative amounts of time that the process spends at each. We consider the behavior of this system as the level of noise in agents’ choice becomes small, defining the stochastically stable strategy profiles to be those which are observed with positive frequency in the long run as noise vanishes. Our paper characterizes the stochastically stable strategy profiles of the logit-response dynamic when coalitional deviations are possible. The intuition for which strategy profiles are stochasti-
3 cally stable comes from the changes in payoffs by switching strategies. Under the logit choice rule, if agents face larger decreases in payoffs by switching to new strategies, it will be less likely that they agree to such revisions. Thus, the decrease in payoffs by a revision represents its unlikeliness. Roughly, the stochastically stable strategy profiles are those that minimize the sum of the payoff de- creases by being switched from other strategy profiles. Those strategy profiles are harder to get away from and easier to get back to in the stochastic dynamic. We apply our main result (Theorem 1) to simple exchange economy and show that the stochasti- cally stable allocations are those maximizing the sum of players’ utilities. Note that these allocations are Pareto efficient, but not uniquely so. We generalize our finding by characterizing a class of inter- actions that exhibit coalitional potential functions, showing that the stochastically stable outcomes are those maximizing the coalitional potential. We use this construction to derive a similar result for the cost-sharing among agents in a fixed network, where every two agents forming a link are required to share the cost of maintaining the link. We show that, in a fixed star network, the greater portion of a link’s cost is sponsored by the peripheral agent as the number of peripheral agents in- creases. We also apply our main result to a network formation game in which countries form free trade agreements (FTAs). FTA formations have been analysed as network formation games by recent stud- ies, such as Goyal and Joshi (2006) and Furusawa and Konishi (2007). Goyal and Joshi (2006) showed that there are two sets of pairwise stable equilibria; one set represents global free trade, while the other represents divided trading blocs. An unanswered question is whether the decentralized FTA formations will likely lead to global free trade. We answer this question in affirmative; our model shows that a complete network is stochastically stable when the number of countries is a realistic number. Related literature As written above, pairwise stable equilibrium has been often used in the literature on networks. The papers most closely related to this one are Jackson and Wolinsky (1996), Watts (2001) and Jackson and Watts (2002). The notion of pairwise stability is introduced in Jackson and Wolinsky (1996). Watts (2001) and Jackson and Watts (2002) analyze the formation of networks in a dynamic framework. Watts restricted attention to an unperturbed dynamic and discussed whether an efficient network would emerge. Watts found that an efficient network appears less likely as the number of agents increases. Jackson and Watts (2002) adopted the stochastic stability approach of Kandori et al. (1993) and Young (1993) in order to select among pairwise stable equilibria. This approach, known as the best response with mutations, has agents choose every suboptimal strategy with same probability. Although this approach is the most popular one to examine stochastic stability, it does not allow for the possibility that costly mistakes may be less common than minor mistakes. One model that accounts for this possibility is the logit choice rule of Blume (1993) which we adopt here. Under the logit rule, players choose an action according to the logit function, which assigns larger probability to those actions which offer larger payoffs, and so takes into account the magnitudes of payoffs when considering agents’ mistakes. Other closely related papers in the stochastic stability literature are Alos-Ferrer´ and Netzer (2010), Kandori et al. (2008) and Newton (2011). Alos-Ferrer´ and Netzer (2010) characterize the stochasti- cally stable outcomes of the logit response dynamic in finite normal form games. Like our model,
4 their model handles deviations by more than one agent; but unlike in our model, these agents do not account for the new choices of simultaneously revising agents. Thus, Alos-Ferrer´ and Netzer (2010) examined stochastic stability of Nash equilibrium by introducing simultaneous unilateral de- viations, while my model examines stochastic stability of R-stable equilibrium by introducing joint deviations. Kandori et al. (2008) employed a similar logit dynamic to study a simple exchange econ- omy showing that stochastically stable allocations are those maximizing the sum of players’ utility functions. As written above, we define a class of interactions that exhibit coalitional potentials, and show that the economies considered in Kandori et al. (2008) fall into this category. Newton (2011) modifies the stochastic stability approach of Kandori et al. (1993) and Young (1993) by introducing vanishingly small probabilities of coalitional deviations, and uses this approach to examine the ro- bustness of Nash equilibrium. In contrast, we assume that forming a coalition is a likely event and focus on stochastic stability of stronger solution concepts than Nash equilibrium. The paper is organized as follows. Section 2 describes the model. Section 3 introduces the dy- namic process and offers the characterization of stochastically stable states. Applications of our framework are laid out in Section 4. Section 5 concludes.
2 The Model
2.1 Interactions Let G = N, (Si, ui)i∈N denote a normal-form game with player set N = {1, . . . , n}, finite strat- egy sets Si and payoff functions ui. Let S = ∏i∈N Si denote the set of pure strategy profiles. The payoff functions are defined as ui : S → R. We define an interaction to be a collection I = N, (Si, ui)i∈N , S where S ⊆ S is the set of feasi- ble strategy profiles for the interaction. In interaction I, the payoff functions need only be defined on
S , i.e. ui : S → R. Since S may be a strict subset of S, interactions include not only normal-form games, but also more general multi-agent choice environments.
For a given strategy si ∈ Si of player i, let S−i(si) = {s−i ∈ ∏j6=i Sj : (si, s−i) ∈ S } denote the set of pure strategy profiles of i’s opponents that are feasible in combination with si. For a coalition J ⊆ N and strategy profile s ∈ S , write sJ and s−J to represent a strategy profile of agents in J and J not in J respectively. Let S (s−J) = {sJ ∈ ∏i∈J Si : (sJ, s−J) ∈ S } denote the set of pure strategy −J profiles of players in J that are feasible in combination with s−J. Similarly, let S (sJ) = {s−J ∈
∏i∈/J Si : (sJ, s−J) ∈ S }.
2.2 Feasible Coalitions and R-Stable Equilibria
We allow agents to jointly revise their strategies by forming a coalition. Which coalitions are feasible may vary among applications. For example, it would be appropriate that a set of feasible coalitions is a set of single agents and pairs in network formation games, as in pairwise stable equi- librium. However, that set may not be appropriate in some exchange economies in which more than two agents may form a coalition.2 To meet the variety of needs across applications, we introduce a
2For example, there has been three-way exchanges in kidney exchange, i.e. three patients exchange their kidney-donors. See Roth et al. (2007).
5 flexible solution concept. Let R ⊆ P (N) \{φ} denote a class of feasible coalitions, where P (N) denotes the power set of N and φ denotes empty set. The following solution concept describes a strategy profile such that for any coalitional deviation by J ∈ R, there is at least one agent in J who will not be better off. In the strict formation, there is at least one agent in J who will be worse off.
0 Definition 1. A strategy profile s ∈ S is an R-stable equilibrium in (I, R) if for all J ∈ R and all sJ ∈ J S (s−J), 0 ui (s) ≥ ui sJ, s−J for some i ∈ J.
0 J A strategy profile s ∈ S is a strict R-stable equilibrium in (I, R) if for all J ∈ R and all sJ ∈ S (s−J),
0 ui (s) > ui sJ, s−J for some i ∈ J.
Note that R-stable equilibrium is equivalent to Nash equilibrium when R = N is the set of single agents. R-stable equilibrium is equivalent to a version of pairwise stable equilibrium when R = {J ⊆ N : |J| ≤ 2}.3 R-stable equilibrium is equivalent to strong equilibrium (Aumann (1959)) when R = P (N) \{φ}.
2.3 Examples of Interactions
To illustrate the sorts of environments represented by interactions, we introduce three examples. Example 1 (Normal-form games). When S = S, interaction I = N, (Si, ui)i∈N , S corresponds to normal-form game G = N, (Si, ui)i∈N . Note that when R = N in Example 1, (I, R) corresponds to a noncooperative normal form game. The next example shows a network formation game where both agents need to consent to form a link.
Example 2 (Network formation games). A strategy of agent i in interaction I is her choice of a vector n−1 of contributions si ∈ Si = {0, 1} . Let sij be an entry of si and representing agent i’s contribution to forming link ij. Link ij is formed if and only if sij = sji = 1. Payoffs derive from two sources: the collaborative work among connected agents and the costs of maintaining links. The feasible strategy profile space is given by S = S. A network g is a set of links formed by agents. Let s ∈ S be a strategy profile and g (s) the network generated by s, that is, (ij) ∈ g(s) if and only if sij = sji = 1. The payoffs of agent i are given by
ui (si, s−i) = ϕ (g (s) , si) (1) where ϕ is a function that maps a network and the agent’s strategy to her payoffs.
In Example 2, it is natural to imagine that a pair of agents will form a link if it is beneficial for both agents. Then, a reasonable solution concept is R-stable equilibrium with R = {J ⊆ N : |J| ≤ 2}. The next example shows an interaction which is not a normal form game, i.e. S ⊂ S.
3A pairwise stable equilibrium requires that there is no pair of players such that one player be strictly better off and the other be weakly better off. For variations, see Section 5 of Jackson and Wolinsky (1996).
6 Example 3 (Simple exchange economies). We define an interaction I to represent a housing market as in Shapley and Scarf (1974). Let H denote the finite set of (indivisible) houses with |H| = n. Let the strategy space for agent i be ( )
n h Si = si ∈ {0, 1} ∑ si = 1 h∈H
h Strategy si can be interpreted as a house allocation for agent i. An entry si = 1 in si ∈ Si represents agent i owning house h. Define the set of feasible strategy profile as ( )
h S = s ∈ S ∑ si = 1 ∀h ∈ H . i∈N
In words, S is the set of allocations in which each owner has exactly one house. Note that under S , no single agent can unilaterally change her strategy or her allocation. Thus, to study behavior in this interaction, we need to specify which groups of owners may jointly change their strategies by swapping their houses.
3 Dynamics and Stochastically Stable States
In this section, we apply the stochastic stability approach to interaction I with families of feasible coalitions. This is briefly summarized as follows. In this approach, we embed a static interaction in a dynamic process in which agents randomly form coalitions and revise their strategies based on improvements in their payoffs in the presence of stochastic payoff shocks. We examine the limiting probability distribution over strategy profiles as the level of stochastic shocks approaches zero.
3.1 The Unperturbed Dynamic
We describe an unperturbed dynamic, i.e. a dynamic with no stochastic payoff shocks. The state of the system st in period t is a vector of the strategies played by each agent and is defined by
t t t s = s1,..., sn ∈ S
t where si denote the strategy played by agent i in period t. The dynamic interaction proceeds as follows. In each period t, a set of agents, denoted by J, is randomly selected from R to revise their 0 J t strategy. For agents in coalition J, one feasible strategy profile sJ ∈ S (s−J) is proposed at random. 0 An agent will agree to sJ if she will be strictly better off by switching. If an agent is indifferent between the current strategy and the proposed strategy, she will agree with probability α ∈ (0, 1). 0 0 Otherwise, she disagrees. If all agents in J agree with playing sJ, agents will switch to sJ for the next round. If at least one agent disagrees, then no agent will alter his current strategy. The dynamic above determines a Markov chain with state space S . Note that we use state s ∈ S and strategy profile s ∈ S interchangeably since these two sets coincide. Now, we discuss the 0 transition probabilities of the Markov chain for two arbitrary strategy profiles s, s ∈ S . Let qJ
7 denote the probability that exactly players in coalition J receive revision opportunities.4 We assume
that qJ > 0 for all J ∈ R. In words, coalition J has positive probability to revise in the dynamic if J 0 is feasible in interaction I. qs0 (J, s) denotes the probability that s is chosen as a new strategy profile when agents form coalition J in state s. We sometimes call R the family of revising sets in the sense that each element of R is the set of agents simultaneously revising their strategy profile. For any two 0 0 strategy profiles s, s ∈ S , let Rs,s0 = J ∈ R sk = sk ∀k ∈/ J denote the set of coalitions potentially leading from s to s0. The transition probabilities for (s, s0) of the unperturbed dynamic is given by " # 0 0 Ps,s0 = ∑ qJ qs0 (J, s) ∏Ψ uj (s) , uj sJ, s−J , (2) J∈Rs,s0 j∈J
where 0 x > y Ψ (x, y) = α x = y 1 x < y.
0 We say that state s is absorbing in the unperturbed dynamic if Ps,s = 1, i.e. the process does not exit s. We also say that a nonempty set of states RC is a recurrent class in the unperturbed dynamic if there is zero probability that the unperturbed dynamic can cause the process to exit RC and a positive probability of moving from any state in RC to any other state in RC in a finite number of time periods. Note that a singleton recurrent class must consist of an absorbing state. The following proposition relates an absorbing state with a strict R-stable equilbrium.
Proposition 1. State s is absorbing in the unperturbed dynamic if and only if s is a strict R-stable equilbrium in (I, R).
Proof. The ’if’ part, recall that if s is a strict R-stable equilbrium, then there exist at least one agent who will be strictly worse off for any coalitional deviation. Then, equation (2) implies that the tran- sition probabilities from s to any other states are zero. For the ’only if’ part, suppose not. There exists absorbing state s that is not a strict R-stable equilbrium. Then, there exists J ∈ R such that all agents in J will be weakly better off by some joint deviations. Equation (2) implies that state s is not absorbing, a contradiction.
The above proposition tells us that the unperturbed process starting from a state that is a strict R-stable equilibrium does not exit the state. Also note that the unperturbed dynamic starting from any state must lead to a strict R-stable equilibrium or a recurrent class consisting of multiple states (where these states are repeatedly visited). This is because the unperturbed process cannot exit a recurrent class if it starts there. This suggests that the unperturbed dynamic does not provide unique predictions. The reason is that it does not distinguish any strict R-stable equilibrium even if these equilibria may differ in the robustness against some exogenous events, e.g. errors made by agents. The approach introduced below can be thought of a robustness check for equilibria. In what follows, we add stochastic shocks to the unperturbed dynamic and examine how resilient these equilibria and cycles are to stochastic shocks.
4 Probability qJ is independent of time, but it may depend on the current state. All the results will hold without change if qJ > 0 for one state implies that qJ > 0 for any state.
8 3.2 The Coalitional Logit Dynamic
We now formally describe a stochastic dynamic behavior of interaction I. We focus on a logit- response dynamic of Blume (1993). As an example of the logit rule, suppose that the current strategy 0 profile is given by s and the set revising agents is given by J. Also suppose that sJ is proposed as 0 their new strategy profile. The probability that all agents in coalition J agree with strategy profile sJ is given by
h −1 0 i exp η ui sJ, s−J 0 Pr members in J agree with sJ s = h i . (3) ∏ −1 0 −1 i∈J exp η ui sJ, s−J + exp [η ui (s)]
where parameter η ∈ (0, ∞) denotes the noise level of the logit-response function and we assume that η is small. Note that agent i takes into account other agents’ new strategies in the LHS of equation (3). The logit response dynamic is a Markov chain on the state space S with stationary transition probabilities. The probability for transition (s, s0) is given by h i exp η−1 u s0 , s η ∑j∈J j J −J P = q q 0 (J, s) . (4) s,s0 ∑ J s h h i i ∈ −1 0 −1 J Rs,s0 ∏j∈J exp η uj sJ, s−J + exp η uj (s)
Note that the unperturbed dynamic is given by the limiting dynamic as η approaches zero.5
3.3 Limiting Stationary Distributions and Stochastic Stability
Let πη denote the stationary distribution of the Markov chain with noise level η. Also let πη(s) denote the probability that πη places on state s. πη(s) represents the fraction of time in which state s is observed over a long time horizon. It is also the probability that state s will be observed at any given time t, provided that t is sufficiently large. Thus, the agents’ behavior is nicely summarized by πη in the long-run. We say that state s is R-stochastically stable if the limiting stationary distribution places positive probability on s.
η Definition 2. State s is R-stochastically stable if limη→0 π (s) > 0. To determine which states will be observed often in the long run, we now introduce several definitions in order to compute the unlikeliness of transitions. Given a state s, define a s-tree to be a directed graph T such that there exists a unique path from any state s0 ∈ S to s. The following definition is due to Alos-Ferrer´ and Netzer (2010). 0 Definition 3. A revision s-tree is a pair (T, ρ) where (i) T is an s-tree, (ii) (s, s ) ∈ T only if Rs,s0 6= φ, and 0 0 (iii) ρ : T → R is such that ρ (s, s ) ∈ Rs,s0 for all (s, s ) ∈ T. Let T (s) denote the set of revision s-trees. The waste of a revision tree (T, ρ) ∈ T (s) is defined as n 0 o W (T, ρ) = ∑ ∑ max uj (s) − uj sJ, s−J , 0 . (5) (s,s0)∈T j∈J
5More precisely, the unperturbed dynamic with α = 1/2 is given by the limiting dynamic as η approaches zero. However, observe that the set of recurrent classes in the unperturbed dynamic does not differ for α ∈ (0, 1). Our analysis will not differ for all α ∈ (0, 1).
9 where J = ρ (s, s0) for (s, s0) ∈ T. Note that the main difference in the waste from Alos-Ferrer´ and 0 Netzer (2010) is that each agent j ∈ J assesses a coalitional deviation sJ instead of an individual de- viation. Roughly, the waste is the sum of the decreases in agents’ payoffs along the tree. A stochastic potential of s-tree is defined as W (s) = min W (T, ρ) . (T,ρ)∈T (s) We cannot directly apply the stochastic potential function defined in Alos-Ferrer´ and Netzer (2010), although the analysis is carried out in a similar way proposed in their paper. Agents does not consider the effect of strategy simultaneously chosen by other agents in Alos-Ferrer´ and Net- zer (2010). By contrast, revising agents take into account their revising actions in this paper. As η approaches zero, the stationary distribution converges to a unique limiting stationary distribution. Our main result is the following theorem which offers the characterization of R-stochastically stable states.
Theorem 1. A state is R-stochastically stable if and only if it minimizes W (s) among all states.
Its proof and most of profs in this paper are relegated to Appendix.
3.4 Conditions for Selection among R-stable Equilibria
Predictions of the stochastic stability approach are made concerning the relative amounts of time that the stochastic dynamic process spends in each state. Since the stochastic dynamic process gravi- tates to recurrent classes of the unperturbed dynamic, the limiting stationary distribution may place positive probability on non-singleton recurrent classes. Since a recurrent class does not necessarily correspond to an R-stable equilibrium, the dynamic of our model may not always select R-stable equilibrium In this section, we characterize the class of games in which our model selects one or some R-stable equilibria given the set of feasible coalitions R. Our model selects a subset of R-stable equilibria if there does not exist any non-singleton recurrent class. We characterize a sufficient condition in terms of payoffs under which a subset of R-stable equilibria is selected. This condition corresponds to the acyclic condition of Young (1993) when S = S and R = N. We define an weakly improving path as follows, and define a cycle of better responses.
Definition 4. An weakly improving path from a strategy profile s1 to another strategy profile sK is a sequence of strategy profiles {s1, s2,..., sK}, such that, for all k = 1, . . . , K − 1,
(i) Rsk,sk+1 is not empty, and
k+1 k (ii) ∃J ∈ Rsk,sk+1 such that ui(s ) ≥ ui(s ) ∀i ∈ J.
An weakly improving path {s1,..., sK} is a sequence of strategy profiles that will be observed with positive probability in an unperturbed dynamic starting from s1. A set of strategy profiles C with |C| ≥ 2 form a cycle if there exists an weakly improving path connecting a strategy profile s to another strategy profile s0 for all s ∈ C and s0 ∈ C. A cycle C is closed if there is no weakly improving path from any s ∈ C that leads to some s0 ∈ C. We call (I, R) R-acyclic if (I, R) has no closed cycle.
10 It is clear that a non-singleton recurrent class corresponds to a closed cycle. If the interaction is R- acyclic, then the process has only singleton recurrent classes (strict R-stable equilibria), and selects a subset of them. It is formally shown by the following proposition.
Proposition 2. Suppose that (I, R) is R-acyclic. Then, any stochastically stable state is a strict R-stable equilibrium.
3.5 A Radius-Coradius theorem
Therem 1 provides an algorithm to determine R-stochastically stable states. However, it might be a burden to compute a stochastic potential of each state. In the stochastic stability literature, several methods have been developed to reduce these computational burdens. One of the most powerful methods to determine stochastically stable states is the Radius-Coradius theorem of Ellison (2000). In this section, we prove an analogous result to the Radius-Coradius theorem in Ellison (2000). In Section 4.4, we show that our result is useful to determine R-stochastically stable states.
A directed graph d (s1, sk) on S is a path if d (s1, sk) is a finite, repetition-free sequence of transi- tions {(s1, s2) , (s2, s3) ,..., (sk−1, sk)} such that si ∈ S for all i = 1, . . . , k. A path d (s1, sk) is feasible if Rsisi+1 is not empty for all i = 1, . . . , k − 1. Define a revision path as a pair (d (s1, sk) , ρ) such that 0 ρ (si, si+1) ∈ Rsi,si+1 for all (si, si+1) ∈ d (s1, sk). Let D (s, s ) be the set of all revision paths with initial point s and terminal point s0. Let the waste W (d(s, s0), ρ) be the waste of (d(s, s0), ρ) as a revision tree. The basin of attraction of a state s, B (s) ⊆ S , is the set of all states s0 such that there exists a path (d, ρ) ∈ D (s0, s) with W (d, ρ) = 0. The radius of state s is defined as
0 0 R (s) = min W (d, ρ) s ∈/ B (s) , (d, ρ) ∈ D s, s . (6)
Also the coradius of state s is defined as
00 0 00 CR (s) = max min W (d, ρ) s ∈ B (s) , (d, ρ) ∈ D s , s . (7) s0∈/B(s)
Let U denote an arbitrary set of states. Similarly to B (s), B (U) ⊆ S is the set of all states s0 such that there exists a path (d, ρ) ∈ D (s0, s) with W (d, ρ) = 0 for some s ∈ U. The radius of U and the coradius of U are defined as
0 0 R (U) = min min W (d, ρ) s ∈/ B (U) , (d, ρ) ∈ D s, s , (8) s∈U 00 0 00 CR (U) = max min W (d, ρ) s ∈ B (U) , (d, ρ) ∈ D s , s . (9) s0∈/B(U)
R (U) is the minimum waste for the state to move away from the basin of U. CR (U) is the maximum waste for the state to move into the basin of U.
Theorem 2 (Radius-Coradius). Let U denote the set of absorbing states of the unperturbed dynamic. Sup- pose a subset of absorbing states U1 ⊂ U such that R (U1) > CR (U1). Then, the limiting stationary distribution places probability one on U1.
11 In the remainder of this paper, we study applications of our model.
4 Applications
4.1 Interactions with Coalitional Potentials
In this section, we introduce a class of interactions which exhibit a function named coalitional potential. Then, we show that the stochastically stable strategy profiles are those maximizing coali- tional potential in this class of interactions. This result provides a much simpler way to determine the stochastically stable strategy profiles relative to Theorem 1. Let I = N, (Si, ui)i∈N , S be an interaction and R the set of feasible coalitions. We define a coalitional potential and a coalitional potential interaction as follows.
0 J Definition 5. A function P : S → R is a coalitional potential for (I, R) if for all J ∈ R, sJ, sJ ∈ S , −J s−J ∈ S 0 0 ∑ uj sJ, s−J − uj sJ, s−J = P sJ, s−J − P sJ, s−J . j∈J Definition 6. (I, R) is called a potential interaction if it admits a coalitional potential.
Note that a coalitional potential function is a potential function in the sense of Monderer and Shapley (1996) when S = S and R = N. In the following sections 4.2 and 4.3, we show that this class of coalitional potential interactions includes interesting examples. The following lemma relates the maximizer of P with R-stable equilibrium of I. And the subsequent proposition shows that our model always selects R-stable equilibrium that maximizes P.
Lemma 1. Suppose that (I, R) is a coalitional potential interaction with potential P. A strategy profile s∗ ∈ S is an R-stable equilibrium if s∗ maximizes P.
J Proof. For any J ∈ R and sJ ∈ S , observe that
∗ ∗ ∗ ∗ ∑ ui sJ, s−J − ui (s ) = P sJ, s−J − P (s ) ≤ 0. i∈J
∗ ∗ This implies that if there exists agent i ∈ J such that ui sJ, s−J > ui (s ), then there must exist at ∗ ∗ ∗ least one agent j ∈ J such that uj sJ, s−J < uj (s ). According to Definition 1, s is an R-stable equilibrium
Note that the converse of the lemma is not necessarily true. There may exist an R-stable equi- librium that does not maximize P. A similar fact is well known for potential games. For coalitional potentials, see the example below.
Example 4. Suppose that interaction I with N = {1, 2}, Si = {A, B} and S = {(A, A), (B, B)} and that R = {(1, 2)}. Utility functions are given by u1(A, A) = 5, u1(B, B) = 3, u2(A, A) = 1, and u2(B, B) = 4. The interpretation of this interaction is that two people choose which restaurant A or B to go together. They have different preferences over restaurants. Observe that the coalitional potential of (I, R) is given by u1 + u2. Both strategy profiles are R-stable equilibria, but (A, A) does not maximize the potential.
12 The following proposition, together with the lemma above, shows that our model always selects a particular subset of R-stable equilibria.
Proposition 3. Suppose that (I, R) is a coalitional potential interaction with potential P. State s∗ is R- stochastically stable if and only if s∗ maximizes P.
Proposition 3 is a generalization of results of Blume (1997) which considered potential games, i.e. S = S and R = N.
4.2 An application in simple exchange economy
In this section, we consider an simple exchange economy of Kandori et al. (2008). We use this economy as an example of an interaction that exhibits a coalitional potential. In an exchange econ- omy, coalitions are naturally formed, since two or more agents are involved with an exchange of goods. Following Kandori et al. (2008), we suppose a house swapping market introduced by Shap- ley and Scarf (1974). Let H denote the finite set of houses with |H| = n. Let the strategy space for agent i be ( )
n h Si = si ∈ {0, 1} ∑ si = 1 h∈H
h An entry si = 1 in si ∈ Si represents that agent i owns house h. Strategy si can be interpreted as a house allocation for agent i. Let ( )
h S = s ∈ S ∑ si = 1 ∀h ∈ H . i∈N
In words, S is the set of states in which every agent has exactly one house. Thus, strategy profile space S can be interpreted as the space of feasible allocations. The utility of agent i is given by ( ) = h = ∈ R ui s ui h si 1 for s S . Let revising set be such that the perturbed dynamic visits every allocation, or state, with positive probability. The dynamic interaction proceeds as follows. At the beginning of each period, a coalition J ∈ R is chosen. They choose one possible allocation 0 J sJ ∈ S s−J where s denotes the current state, or the current allocation, and
J 0 S s−J = s ∈ S : si = si ∀i ∈/ J . (10)
0 0 The probability that s is chosen is exogenously given by qs0 (s, J) such that qs0 (s, J) = 0 if s ∈/ J 0 S s−J . Agents in J agree with s with the probability given by (4). Otherwise, the trade does not occur in a given period. We have the following characterization of R-stochastically stable allocations.
Proposition 4. The limiting stationary distribution places probability one on the set of allocations that maxi- mize the sum of the agents’ utility functions.
A version of the result has been proved by Kandori et al. (2008). However, the above proposition provides slightly more general result in the sense that it does not require the following restrictions of Kandori et al. (2008); (i) probability qJ must be independent of the current strategy profile, and
(ii) probability qs0 (J, s) must be symmetric in the candidate strategy distribution, that is, qs0 (J, s) = 0 qs(J, s ) must hold.
13 Restriction (i) rules out settings where coalition formations depend on the strategy profile. For instance, suppose a listing service, e.g. claigslist.org, that provides a list of potential traders. When people use a listing service to find their counter-part, the coalition formations may depend on the current allocation. This is because a typical listing service provides a list sorted by categories, e.g. houses sorted by location. It implies that the probability people find their counter-part on such listing services would depend on the current allocation. One implication of restriction (ii) is that there cannot exist any favorable choice of strategy pro- files. To see this, observe that
0 ∑ qs0 (J, s) = 1 ⇔ ∑ qs(J, s ) = 1. 0 J 0 J s ∈S (s−J ) s ∈S (s−J )
J 1 2 3 1 2 3 For the intuition, suppose J = 1, 2, 3, S (s−J) = {s , s , s }, and that s Pareto domitnes s and s . Then, it is natural to imagine that people pick up s1 with higher probability when the allocation 2 3 2 3 is s or s , i.e. qs1 (J, s) > 0.5 for s ∈ {s , s }. However, restriction (ii) rules out these probability distributions.
4.2.1 Assigning a utility function to preference order
Ben-Shoham et al. (2004) considered a house allocation problem as a triple N, H, {i}i∈N where i is a complete, transitive preference order over H. Let hi (s) denote the house allocated for agent i in allocation s. In their model, the probability that agent i and j trade is given by εni+nj where
nx = h ∈ H : hx (s) x h x hy (s) for (x, y) ∈ {(i, j) , (j, i)} . (11)
The envy level of agent i in allocation s is defined as ei (s) = |{h ∈ H : h x hx (s)}|, that is the num- ber of houses agent i prefers to his current house. Ben-Shoham et al. (2004) showed that the minimum ni+nj envy allocation is selected when the transition probability is given by ε with nx defined in (11). The present model can yield the exactly same result by assigning the following utility to each house allocation:
ui (s) = |{h ∈ H : hi (s) x h}| .
It is obvious that ui (s) correctly represents the agent i’s preference. Following Ben-Shoham et al. (2004), let R = {J ⊂ N : |J| = 2}. Proposition 4 tells us that states maximizing the sum of utilities are stochastically stable. Observe that ( ) ( ) ∗ ∗ 0 ∗ ∗ 0 s : ui (s ) = max ui s = s : [H − ui (s )] = min H − ui s ∑ 0∈ ∑ ∑ 0∈ ∑ i∈N s S i∈N i∈N s S i∈N ( ) ∗ ∗ 0 = s : ei (s ) = min ei s . ∑ 0∈ ∑ i∈N s S i∈N
Thus, our model provides the same prediction as Ben-Shoham et al. (2004) by assigning a linear utility to the preference order.
14 4.2.2 Exchange economy with K goods
It is straightforward to extend the analysis to an exchange economy with more than one good. Suppose an economy consisting of consumer set N = {1, . . . , n} and K goods. Consumer i’s con- 1 K K 6 k sumption bundle is denoted by si = {si ,..., si } ∈ Si = Z+. An entry si ≥ 0 represents agent i’s holdings of good k. The total amount of each good k, or the total endowment of k, is denoted by ek ≥ 0 for k = 1, . . . , K. Let the space of feasible allocations S be ( )
k S = s ∈ S ∑ si = ek ∀k = 1, . . . , K . i∈N
We assume that the intrinsic utility of agent i is given by ui (s) = ui (si). Again, let the class of feasible coalitions R be such that the perturbed dynamic visits every allocation with positive probability. The dynamic interaction proceeds similarly. At the beginning of each period, a coalition J ∈ R is chosen. 0 J J Consumers in J choose one allocation sJ ∈ S s−J where S s−J is given by (10). They will agree to the exchange and switch to allocation s0 if every consumer yields higher utility subject to the random shocks. Define the social surplus of this economy as SS(s) = ∑i∈N ui(s). According to a similar discussion to the house-exchange economy, it is obvious that SS(s) is a coalitional potential of this interaction. Then, the following proposition characterizes R-stochastically stable allocations. Since the proof is straightforward, it is stated without proof.
Proposition 5. Suppose an exchange economy consisting of n consumers and K goods. The limiting station- ary distribution places probability one on the set of allocations that maximize the social surplus.
The proposition holds even if utility function is decreasing over some goods. However, note that the endowment constraints must hold for all goods. An interpretation of such a good is that it might be a ’task’ that must be carried out by somebody. The following example illustrates the case.
Example 5 (Divide the household chores). Two individuals 1 and 2 share their monthly (30 days) household duties, say cooking and dish washing denoted by c and d respectively. Assume that each x x duty must be done everyday so that ∑i=1,2 si = 30 holds for x ∈ {c, d}. Also assume that si ∈ Z for i ∈ {1, 2} and x ∈ {c, d}. Suppose their utility functions are given by
c d u1 = α1 log (30 − s1) + (1 − α1) log (30 − s1) c d u2 = α2 log (30 − s2) + (1 − α2) log (30 − s2) where 0 < αi < 1 for i ∈ {1, 2}. αi is interpreted as the relative dislike of cooking to dish washing for individual i. At the beginning of each month, they negotiate to swap their duties. Assuming that
α1 = 2/3 and α2 = 1/3, the R-stochastic stable allocation is given by
c 30α2 d 30α1 s1 = = 10, s1 = = 20. α1 + α2 α1 + α2
6Each good is divisible (but not perfectly).
15 4.3 Cost sharing in a fixed network: distance-based payoffs
In this section, we study cost-sharing among agents in a fixed network, denoted by g, with the distance-based payoffs. Network g is a set of links and we write (ij) ∈ g if there is a link between
agents i and j in g. The set of neighbors of agent i denoted by gi is the set of agent with which agent i has a link. We say that i1 and ik are connected if there exists a set of distinct agents {i1, i2,..., ik} such that {i1i2,..., ik−1ik} ⊂ g. Let sij denote agent i’s contribution to link (ij) ∈ g. For all (ij) ∈ g, a strategy space of agent i with regards to link (ij) is such that Sij = {0, γ, 2γ,..., c − γ, c} where γ > 0. Note that no element in Sij is = × negative, i.e. there is no transfer between agents. Agent i’s strategy space is defined as Si j∈gi Sij. Since links are fixed, a strategy profile s = (s1,..., sn) is such that sij + sji = c > 0 for any (ij) ∈ g. Since a change in cost-sharing must involve two agents, we define R = {J ⊂ N : |J| = 2}. Let S denote the set of feasible strategy profiles, that is
S = s ∈ ×i∈NSi : sij + sji = c ∀ (ij) ∈ g .
Let d (i, j) denote the distance between i and j in terms of the number of links in the shortest path between them. Let d (i, j) = ∞ if i and j are not connected. A utility function is distance-based if ! d(i,j) ui (si, s−i) = ϕ ∑ δ − ∑ sij j6=i j∈gi
for some δ < 1 and an increasing function ϕ : R → R. The interpretation of the utility function is that agent i joined some collaborative work if she has an access to or is being accessed by another agent. The following lemma is immediate, thus provided without proof.
Lemma 2. Any state s is a singleton absorbing state of the unperturbed dynamic.
The lemma implies that there is no cycle in the unperturbed dynamic.
Proposition 6. A state s is R-stochastically stable if and only if s maximizes the sum of the agents’ utilities.
Proof. Note that a change in strategies sij and sji affects only agents i and j’s utility functions. Then, the cost sharing interaction admits a coalitional potential P such that for all s ∈ S
P (s) = ∑ ui (s) . i∈N
The claim follows from Proposition 3.
The characterization of stochastically stable states depends on the form of utility function. The following corollary is immediate.
Corollary 1. If ϕ is linear, all states are R-stochastically stable. If ϕ is concave, any R-stochastically stable state s satisfies the following properties: for all (ij) ∈ g
Yi (s) − Yj (s) ≤ γ if sij > 0 and sji > 0,
Yj (s) − Yi (s) < γ if sij = c and sji = 0,
16 where d(i,j) Yi (s) = ∑ δ − ∑ sij. j6=i j∈gi
Proof. It is obvious if ϕ is linear. For the concave case, suppose s ∈ S such that Yi (s) − Yj (s) > γ 0 0 and sij, sji > 0 for some i, j. Let sij = sij + γ and sji = sji − γ. The concavity implies that