Buy-backs, Price Wars, and Collusion Enforcement∗ Jimmy Chan Wenzhang Zhang † June 2009 Abstract Recent studies of cartel operation show that in most cartels both sales and price data were private information, and side payments between firms were used in addition to price wars to enforce collusion. This paper proposes a simple mechanism for collusion enforcement that incorporates these features. The mechanism applies to both price and quantity competition, and it allows for multi-product firms. ∗ † We thank Joe Harrington for comments and suggestions. School of Economics, Shanghai University of Finance and Economics, Shanghai 200433, China. E-mail: [email protected] (Chan) and [email protected] (Zhang). 1 1 Introduction In enforcing a collusion agreement, a cartel must contend with the problem that member firms may cut price or over sell secretly (Stigler, 1964). In a seminal paper Green and Porter (1984) show that a cartel that could not monitor its members’ outputs may still deter cheating by using a price war as a collective punishment. In a recent case study Harrington (2006) shows that, instead of starting a price war, many cartels settled disputes through compensation. In the citric acid cartel, for example, firms agreed to a set of sales quotas, and firms that sold above quota were required to purchase product from those that sold below.1 In one case, Haarmann & Reimer were required to purchase 7000 tons of citric acid from ADM.2 While such compensation schemes should in principle discourage over-producing, they might not work in practice because firms could lie about their sales. Many cartels had no safeguard against under-reporting. While some used trade associations or other third parties to collect data, it is likely that a determined firm would find a way to fudge the data. For any third party enforcement that would make it impossible for a firm to lie would also likely to arouse the suspicion of the antitrust authority. Nevertheless, Harrington (2006) concludes that cartels appeared to take the self-reported data seriously. Suslow (2005) finds that cartels that used some type of self-imposed penalty schemes were more stable. In this paper we show how a cartel can enforce collusion with self-reported data with a scheme similar to the one described above. The scheme applies to both price and quantity competition, and it allows for multi-product firms. Under the scheme firms agree to a collusive agreement that maximizes cartel profit and a set of profit targets. They report their profits at the end of each period. The reported profit shortfalls—the difference between the profit targets and reported profits—determine the side-payments between firms and the probability that collusion breaks down and a price war erupts. A firm tends to pay less when it 1 Harrington (2006) reports that some form of compensation schemes were used in the car- tels in the market for choline chloride, citric acid, lysine, organic peroxides, sodium gluconate, sorbates, most vitamins, and zinc phosphate. Also see Levenstein and Suslow (2006). 2 Harrington (2006) 2 reports a lower profit, but the gain is offset by a higher probability of a price war. In reality, collusion may break down if one or more firms suspect some firms have been cheating about their sales. Taking into account of both its payment to other firms and the loss of profit in case of a price war, a firm’s expected “penalty” is independent of its own report and increasing in the reported profit shortfalls of the other firms. If the profit targets are set very high, then the penalty will fully internalize the externality a firm’s action imposes on the other firms. However, in practice the profit targets could not be set very high, because firms would refuse to pay a very high penalty. But if the targets are set, for example, near the expected profits, a firm which secretly cut price may get lucky and escape punishment when an exceptionally high industry demand, due to a positive demand shock, masks the effect of the price cut. We show that despite this “truncated distribution” problem, the firms could be motivated to comply with the collusion agreement under a weak condition that is satisfied either when the demand shock is multiplicative or additive, or when its density is log-concave. Penalties in our model must include both side payments and a price war. If only side payments are used, firms would under-report to raise the penalties of the other firms. On the other hand, if only a price war is used, firms would overreport to avoid starting a price war. This explains why compensation schemes were adopted in many cartels. Since a price war is used to ensure truth-telling, collusion will break down occasionally due to random demand shocks. By contrast, a cartel that monitors firms’ sales directly could enforce collusion with only side-payments (Harrington and Skrzypacz (2007)). This suggests that arrangements that allow firms to publicize sales information credibly may be anti-competitive. 1.1 Related Literature Our model is inspired by the recent works of Harrington and Skrzypacz, which are the first to analyze the role of side payments in collusion enforcement. Harrington and Skrzypacz (2007) considers an environment where prices are private but sales are public. They show that by requiring the firms to pay a per unit tax, a cartel 3 can induce the firms to follow the collusion agreement. Harrington and Skrzypacz (2008) extend their 2007 model to the case where both prices and sales are private.3 The key difference between that model and ours lies in the way the firms are induced to reveal their sales. Since in their model a firm is required to pay a tax when its own sales is too high, it would have incentives to underreport. They show that firms could be induced to tell the truth when the correlation between firms’ demands satisfies certain conditions. No such restriction is needed in our model since a firm’s report is used only to determine the penalties of the other firms. More broadly, our model is also related to the literature of repeated games with communication under private monitoring (Kandori and Matsushima, 1998, Compte, 1998, Fudenberg and Levine, 2007, Obara, 2008, and Zheng, 2008). In this literature collusion is usually supported by using a statistical test that the collusive action is the least likely to fail among all actions. In our model, that would suggest a two-sided test where a firm is punished when other firms’ profits are either too high or too low. On the contrary, the test we used is a one-sided test where a firm is punished only when other firms’ profits are too low. Hence, in our model a firm can reduce its expected penalty by raising its price, but the gain would be smaller than the loss in its own profit that a higher price would cause. Another difference between this paper and the repeated-game literature is that the latter tends to focus on the question of asymptotic efficiency as the discount factor goes to one. By contrast, the objective of the present paper is to come up with an enforcement scheme that is both versatile and consistent with actual cartel operation. 2 Model There are n firms, indexed by i = 1, 2, . . . , n. The horizon is infinite and time periods are indexed by t = 0, 1, . . . . In each period t, firms play the following stage game: (1) each firm i simultaneously chooses an action si,t from a set Si ; (2) 3 Aoyagi (2002) analyzes collusion in a repeated Betrand game in which the demand shocks are positively correlated. 4 Figure 1: Time line the stochastic profit πi,t of each firm i is realized; both the action si,t and profit πi,t are private information of firm i; (3) each firm i simultaneously reports a profit π̂i,t to the other firms; (4) based on the reported profits, each firm i simultaneously makes a side-payment τij,t to each firm j; (5) firms observe the outcome xt of a public randomization device; xt is uniformly distributed between 0 and 1.4 See Figure 1. The action si can be broadly interpreted. For example, it can be price, quantity, or quality. We assume that it is one-dimensional and the action set Si is an interval. In section 4.1 we also allow the action si to be multi-dimensional to capture the fact that most cartels involve firms that produce multiple goods or sell in multiple markets. When an action profile s ≡ (s1 , s2 , . . . , sn ) is played, the stochastic profit πi of firm i has a smooth distribution function Fi (·, s) with finite support [π i , π i ]. Conditional on any sequence of strategies, the profits are independent over time but possibly correlated across firms. We make two assumptions about the profit 4 A note on notation: The subscript “i, t” of a variable refers to firm i in period t. Similarly, the subscript “ij, t” refers to firms i and j in period t. The same variable with subscript t refer to the n-vector or n × n matrix of the variable. For example τt is n × n side-payment matrix in period t. The subscript −i refer to a n-vector minus the i-th component. We drop the subscript t when doing so does not cause confusion. 5 distributions. Assumption 1. For each firm i, Fi (πi , sj , s−j ) is decreasing in sj for each πi and each j. The assumption means that a higher action of firm j shifts the distribution of firm i’s profit to the right in the sense of first-order stochastic dominance. Assumption 2. For each pair of distinct firms i and j, each action profile s−j of the firms other than j, and each Kij < π i , the ratio Z Kij πi ∂Fi (πi , sj , s−j ) dπi ∂sj Z πi πi ∂Fi (πi , sj , s−j ) dπi ∂sj (1) is weakly decreasing in sj . Note that by integration by parts E [πi |s] = π i − R πi πi Fi (πi , sj , s−j ) dπi . Roughly speaking, Assumption 2 requires that the effect of si on the low end of the distribution of πj is larger when si is smaller. Assumptions 1 and 2 are satisfied for a large class of demand functions. Example: Bertrand Competition. Suppose the firms are competing in price. Each firm i has a constant marginal cost, ci , and must choose a price pi ∈ [ci , p]. Let p ≡ (p1 , . . . , pn ) be a price vector. Firm i’s stochastic demand is qi (p, εi1 , εi2 ) = h (εi1 qbi (p) + εi2 ) , where h is strictly increasing, εi1 and εi2 are random shocks, and qbi (p) is positive for all p, decreasing in pi and increasing in pj , j 6= i. The random demand shock, εik , k = 1, 2, has a smooth distribution on a finite support [εik , εik ] with εik ≥ 0. Lemma 1. Assumptions 1 and 2 are satisfied in the Bertrand example if either (1) h is linear or (2) εi1 is a constant and the density of εi2 is log-concave. An analogous result holds for Cournot competition. The proof of Lemma 1 and its Cournot counterpart is provided in the Appendix. Assumption 1 is satisfied because qi (p, εi1 , εi2 ) increases in pj for any εi1 and εi2 . The conditions in Lemma 6 1 therefore are needed only for Assumption 2. If h is linear, the effect a marginal change in pj on πj is independent of the realization of εi2 , and the ratio in (1) is equal to the probability that πj is less than Kij , an event which is less likely when pj is high. Intuitively, the effect of pj concentrates on the high end of the distribution of πi because πi is unlikely to be less than Kij when pj is high. The random shock may enter the demand function in a non-linear way. For example, if h is strictly concave, then the effect a marginal change in pj on πj will be smaller when εi2 is smaller. In this case, Assumption 2 will still hold if there is no multiplicative shock and the density of εi2 is log-concave. Many common distributions, including normal and logistic, are log-concave. Under either condition, there is no restriction on the functional form of qbi . Let s∗ = (s∗1 , s∗2 , . . . , s∗n ) be an action profile that maximizes the expected P industrial profit ni=1 E [πi |s]. We assume that the stage game has a pure-strategy N N Nash equilibrium sN = (sN 1 , s2 , . . . , sn ) such that n X ∗ E [πi |s ] > n X i=1 E πi |sN . i=1 Let mpub = π̂t , τt , xt denote the public information in period t, and mpri t i,t = si,t , πi,t firm i’s private information in that period. At the beginning of period t, pub pub all firms observe a public history hpub = (mpub t 0 , m1 , . . . , mt−1 ), and, in addition, pri pri pri each firm i observes a private history hpri i,t = (mi,0 , mi,1 , . . . , m i,t−1 ). Player i’s full pub pri 5 information at the beginning of period t is denoted by hi,t = ht , hi,t . An ac- tion strategy of firm i is a function ai that maps a history hi,t to an action ai (hi,t ) in Si . A reporting strategy of firm i is a function ri that maps hi,t , mpri to a report i,t ri (hi,t , mpri i,t ) in [π i , π i ]. A transfer strategy of firm i is a n-vector bi = (bi1 , . . . , bin ), pri where each bij maps hi,t , mpri i,t , π̂t to a nonnegative real number bij hi,t , mi,t , π̂t . A strategy σi of firm i is the collection of an action, a reporting, and a profile of transfer strategies (ai , ri , bi ). Given a strategy profile σ = (σ1 , σ2 , . . . , σn ), firm i’s 5 pri Define each of hpub 0 , hi,0 and hi,0 as the empty set ∅. 7 expected average discounted payoff in the repeated game is " ! # ∞ X X vi (σ) ≡ E (1 − δ) δ t πi,t − τij,t − τji,t σ , t=0 j6=i where E[ ·|σ] is the expectation with respect to the probability measure on histories induced by the strategy profile σ, and δ is the common discount factor. A strategy is a public strategy if its continuation strategy after a history hi,t depends only on the public history hpub t . A profile of public strategies is a perfect public equilibrium if the continuation strategy profile after each period constitutes a Nash equilibrium. 3 Trigger-Strategy Equilibrium We focus on equilibrium with a simple strategy structure. A truth-telling triggerstrategy profile ρ is defined by three components: a collusive action s, a probability function p : [π, π]n → [0, 1], and an n×n payment matrix β, where each component βij : [π, π]n → [0, ∞). Firms start off in the collusive state, in which they play an action profile s and report their profits truthfully. Based on the reported profits π̂t , each firm i pays βij (π̂t ) ≥ 0 to each firm j. If all firms make the side-payments, they switch to the non-collusive state in the next period with probability p(π̂t ). If any of them fails to pay, they switch to the non-collusive state with probability one. The non-collusive state is absorbing. Once the firms enter the non-collusive state, they play the stage-game Nash equilibrium forever; the reports in the noncollusive state are irrelevant and there are no transfers. To summarize, under a trigger-strategy profile ρ = (s, p, β), a public history hpub is in the collusive state if (i) t = 0, or (ii) hpub t t−1 is in the collusive state, and τij,t−1 = βij (π̂t−1 ) for all i, j, and xt−1 ≥ p(π̂t−1 ); otherwise, t is in the non-collusive state. Let C (s, p, β) denote the set of public histories in the collusive state. Firm 8 i’s strategy σiρ = (aρi , riρ , bρi ) under ρ is given as follows: ( si if hpub ∈ C (s, p, β) t ai (hi,t ) = ; sN otherwise i ri (hi,t , si,t , πi,t ) = πi,t ; ( bij (hi,t , si,t , πi,t , π̂t ) = βij (π̂t ) if hpub ∈ C (s, p, β) t 0 otherwise . Firm i’s expected payoff from σ ρ is (1 − δ) (E [πi |s] − E [βij (π) − βji (π) |s|]) + δE [p (π) |s] viN vi (σ ) = . 1 − δ (1 − E [p (π) |s]) ρ Define firm i’s “continuation payoff” conditional on π̂ in the collusive state— its payoff that comes after the realization of the current period profit, including the current period side-payments and the expected payoffs from the next period onward—as wiρ (π̂) ≡ vi (σ ρ ) − p(π̂)(vi (σ ρ ) − viN ) − δ −1 (1 − δ) X (βij (π̂) − βji (π̂)) , (2) j6=i where viN = E[πi |sN ] is the expected profit of firm i in the stage-game Nash equilibrium. The second term in (2) is the expected profit loss due to a price war, and the third term is firm i’s net payment to other firms. Together, they constitute a “penalty” for firm i. We will focus on trigger-strategy profiles with the following features. Definition 1. A trigger-strategy profile (s, p, β) is enforceable if for all i and π, wiρ (π̂) ≥ viN . A unenforceable trigger-strategy profile could never be an equilibrium as firm i would be better off not paying the side-payments it owes when wiρ (π̂) < viN . Definition 2. A trigger-strategy profile (s, p, β) involves no simultaneous payments if a firm never makes a strictly positive payment to one firm and receives one from another at the same time. That is, for all i, there is no π̂ where both P P β (π̂) and ij j j βji (π̂) are strictly positive. 9 Let Π (s) denote all profit profiles that has a positive density when some firm i chooses some s0i ∈ Si and all other firms j choose sj . The set Π (s) includes all profit profiles that might occur under any choice of firm i when other firms are choosing s−i . Definition 3. Let K and λ be n × n matrices with λij > 0 for all i and j. A trigger-strategy profile ρ implements a linear penalty scheme (K, λ) if for all i and all π̂ ∈ Π (s) , wiρ (π̂) = vi (σ ρ ) − δ −1 (1 − δ) X λij max (Kij − π̂j , 0) . (3) j6=i Each Kij is a target for firm j’s profit. Under (K, λ), firm i is required to pay a penalty linear in the profit shortfall of each firm j 6= i below Kij in the collusion state. Its average discounted profit in the collusion state is Z Kij X K,λ vi (s) ≡ E[πi |s] − λij Fj (πj , s) dπj . j6=i πj Proposition 1. Suppose Assumptions 1 and 2 hold. Then for any linear penalty scheme (K, λ) such that R K ∂F Rπ 1. λij π ij ∂sij dπj = π j ∗ j j s=s 2. (1 − δ) supπ̂∈Π(s∗ ) ∂Fj ∂si s=s∗ dπj , for each distinct i and j; and K,λ ∗ N λ max (K − π b , 0) ≤ δ v (s ) − v , for each ij j i i j6=i ij P i, then there is a perfect public equilibrium where the firms choose a trigger-strategy profile with collusive action s∗ that implements the penalty scheme (K, λ). Proposition 1 says that when conditions 1 and 2 are met, a cartel can use a linear penalty scheme to enforce a collusive action that maximizes the cartel profit. The right-hand side of the equation in condition 1 is the marginal effect of si on firm j’s profit. Since the penalty scheme captures only the effect of si on firm j’s profit below Kij , condition 1 requires that λij be scaled up so that the marginal effect of si on the penalty is the same as that on firm j’s profit. This ensures that it is optimal for each firm i to choose s∗i in the collusion state. The left-hand 10 side of the equation in condition 2 is an upper bound of firm i’s penalty under (K, λ). Condition 2 requires that this bound be less than the value the scheme can destroy through a price war. The condition ensures that the penalty can be implemented by a trigger-strategy profile and that the firms have the incentive to make the required side-payments. We prove Proposition 1 through two lemmas. Lemma 2. Under Assumptions 1 and 2, if (K, λ) satisfies condition 1 in Proposition 1, then " s∗i ∈ arg max E πi − si X j6=i # λij max (Kij − πj , 0) si , s∗−i . Proof of Lemma 2. It is useful to rewrite the objective function in (4) as " # X X ∗ ∗ E −λij max (Kij − πj , 0) − πj si , s−i . E πj si , s−i + j (4) (5) j6=i The first term in (5) is the expected cartel profit, and the second is the expected difference between firm i’s penalty associated with firm j’s profit shortfall and firm j’s profit. By definition, s∗i maximizes the first term of (5). Hence, Lemma 2 holds if s∗i also maximizes Hij si , s∗−i ; Kij , λij ∗ ≡ E −λij max (Kij − πj , 0) − πj si , s−i (6) for each j 6= i. Notice that if Kij > π j and λij = 1 for each j 6= i, then Hij si , s∗−i ; Kij , λij = P − j6=i Kij for all si . In this case, firm i’s penalty fully captures the externality of si on the profits of the other firms. However, as we shall see, profit targets are bounded from above by the continuation value of collusion and might not be set above π j . Suppose Kij is less than π j . Through integrating Hij by parts and differentiating it with respect to si , we have Z K Z πj ∂Hij si , s∗−i ; Kij , λij ∂Fj ∂Fj = −λij (πj , si , s−i ) dπj + (πj , si , s−i ) dπj . ∂si π j ∂si π j ∂si (7) 11 By condition 1, ∂Hij s∗i , s∗−i ; Kij , λij = 0. ∂si It follows from Assumption 2 that ∂Hij si , s∗−i ; Kij , λij = ∂si R πj ∗ Z Kij ∗ /∂s dπ ∂F π , s , s i j j j i −i ∂Fj πj , si , s−i λij − R πj ≥0 − Kij ∗ ∂s i ∂Fj πj , si , s−i /∂si dπj πj π j if and only if si ≤ s∗i . Intuitively, Assumption 2 ensures that firm i’s expected penalty decreases faster than the profit of firm j increases (as si increases) when si is less than s∗i , and slower when si is greater than s∗i . Lemma 3. If (K, λ) satisfies condition 2 in Proposition 1, then it can be implemented by an enforceable trigger-strategy profile that requires no simultaneous payments and in which s∗ is chosen in the collusive state. Proof of Lemma 3. We first choose p (b π ) so that the expected value to be destroyed through a price war is equal to the total penalty that is required. For each π b ∈ Π (s), set P i p (b π) ≡ (1 − δ) δ b, 0) j6=i λij max (Kij − π . P K,λ ∗ N (s ) − v v i i i −1 P (8) Since, by condition 2, the maximum penalty is less than the continuation value of collusion for each firm, the total penalty for all firms must be less than the total value. Hence, p (b π ) ∈ [0, 1] for all π b ∈ Π (s). To obtain the right penalty for each firm (i.e. (3)), set the net payment of each firm i to be X λij max (Kij − π̂j , 0) − δ (1 − δ)−1 p(π̂) viK,λ (s∗ ) − viN . βinet (π̂) ≡ (9) j6=i These net payments could be implemented by the side-payment scheme: net β net (π̂) Pmin(βj (π̂),0) if βinet (π̂) ≥ 0, net i k min(βk (π̂),0) βij (π̂) ≡ 0 otherwise. 12 (10) According to β, each firm i with a positive net payment pays each firm j with a negative payment an amount in proportion to firm j’s share of the total payment changing hand. It is straightforward to see that β involves no simultaneous payments. Finally, the enforceability of ρ = (s, p, β) follows directly from condition 2. Proof of Proposition 1. Since a firm’s report under (s∗ , p, β) does not affects its own penalty, it is always optimal for each firm to report truthfully. If firm i refuses to make the side-payments, its continuation will be viN . Since (s∗ , p, β) is enforceable, it is optimal for each firm i to comply with β. Finally, by (4), it is optimal for firm i to choose s∗i in the collusion state.6 Proposition 1 says that a cartel can use a linear penalty scheme to enforce a collusive action s∗ so long as the maximum penalty required for each firm is smaller than the continuation value of collusion. The following proposition shows that this condition will be met when the demand shock is small and the firms are sufficiently patient. Proposition 2. Suppose Assumptions 1 and 2 hold and there is some positive constants κ1 and κ2 such that κ1 < |∂Fj (πj , s) /∂si | ≤ κ for all s, πj , i, and j. Then there exist δ̄ < 1 and > 0 such that if Pr (|πi − E[πi |s]| ≥ | s) ≤ for each i and each s, then for all δ > δ̄ there is a trigger-strategy perfect public equilibrium where the firms choose s∗ in the collusion state. Moreover, firm i’s average discounted equilibrium payoff converges to E [πi |s∗ ] as tends to zero. 4 Extensions 4.1 Multiple markets and multiple instruments Suppose that there are M markets, indexed by m = 1, 2, . . . , M , and the action space Si of each firm i is a convex subset of RL , the Euclidean space of dimension 6 Strictly speaking, p and β are defined only for π b ∈ Π (s∗ ). Any profit report π b∈ / Π (s∗ ) is not reacheable by unilateral deviation. For completeness, we can set p (b π ) = 1 and βij (b π) = 0 for all π b∈ / Π (s∗ ). 13 L. Let πi,m denote the profit realization of firm i in market m, and si,l denote the l-th component of an action si in Si . We assume that, for each i and each m, there exists a function φi,m : ×nj=1 Sj → R such that, conditional on a action profile s = (sj,l )j≤n, l≤L , the cumulative distribution function of firm i’s profit πi,m has the form Fi,m (·, φi,m (s)) with support [π i,m , π i,m ]. Assumption 3. For each firm j, each market m and each level of profit realization πj,m , we have A. the probability Fj,m (πj,m , φj,m ) is decreasing in φj,m ; and B. the partial derivative ∂Fj,m (πj,m , φj,m )/∂φj,m is integrable with respect to πj,m , and, for each K ≤ π j,m , the ratio Z πj,m Z K ∂Fj,m ∂Fj,m (πj,m , φj,m ) dπj,m (πj,m , φj,m ) dπj,m π j,m ∂φj,m π j,m ∂φj,m is weakly decreasing in φj,m . Assumption 4. For each pair of distinct firms i and j and each market m, the function φj,m is differentiable and quasi-concave in si . Assumption 5. There exist positive constants κ1 and κ2 such that κ1 ≤ ∂Fj,m (πj,m , s) ≤ κ2 , ∂si,l for each s, each i and j, each l, and each πj,m . We now modify the penalty scheme in the benchmark model to encompass the situations of multiple markets and/or multiple instruments. We still let Π (s) denote all profit profiles that has a positive density when some firm i chooses some s0i ∈ Si and all other firms j choose sj . Definition 4. Let K = (Kij,m )1≤m≤M ;1≤i,j≤n;i6=j and λ = (λij,m )1≤m≤M ;1≤i,j≤n;i6=j be such that Kij,m ≤ π j,m and λij,m > 0 for all i, j and m. A trigger-strategy profile ρ implements a linear penalty scheme (K, λ) if for all i and all π̂ ∈ Π (s) , wiρ (π̂) = vi (σ ρ ) − δ −1 (1 − δ) M XX j6=i m=1 14 λij,m max (Kij,m − π̂j,m , 0) . (11) Each Kij,m denotes a target for firm j’s profit in the market m, and firm i is required to pay a penalty linear in the profit shortfall of firm j below Kij,m in the collusion state. Its average discounted profit in the collusion state is "M # XX Z Kij,m M X K,λ vi (s) ≡ E πi,m s − λij,m Fj,m (πj,m , s) dπj,m . m=1 π j,m j6=i m=1 As before, we let s∗ = (s∗i,l )1≤i≤n;1≤l≤L be an action profile maximizing the P P cartel profit ni=1 M m=1 E [πi,m |s], and assume that the stage game has a purestrategy Nash equilibrium sN = (sN i,l )1≤i≤n;1≤l≤L such that M X n X ∗ E [πi,m |s ] > m=1 i=1 Let viN = PM m=1 M X n X E πi,m |sN . m=1 i=1 E πi,m |sN denote firm i’s profit given by sN . Parallel to Propositions 1 and 2, we have the following two propositions. Proposition 3. Suppose Assumptions 3 and 4 hold. Then for any linear penalty scheme (K, λ) such that RK ∂Fj,m 1. λij,m π ij,m ∂φj,m j,m s=s∗ dπj,m = R πj,m π j,m ∂Fj,m ∂φj,m s=s∗ dπj,m , for each distinct i and j, and each m; and 2. (1 − δ) supπ̂∈Π(s∗ ) PM m=1 K,λ ∗ N min , , 0 ≤ δ v (s ) − v λ max K − π b ij,m i j,m i j6=i ij,m P for each i, then there is a perfect public equilibrium where the firms choose a trigger-strategy profile with collusive action s∗ that implements the penalty scheme (K, λ). Proposition 4. Suppose Assumptions 3, 4 and 5 hold. Then there exist δ̄ < 1 and > 0 such that if Pr (|πi,m − E[πi,m |s]| ≥ | s) ≤ for each i, each m and each s, then for all δ > δ̄ there is a trigger-strategy perfect public equilibrium where the firms choose s∗ in the collusion state. Moreover, firm i’s average discounted P ∗ equilibrium payoff converges to M m=1 E [πi,m |s ] as tends to zero. The proofs for these two propositions are identical to those for Propositions 1 and 2, except that we need to replace Lemma 2 by the following lemma, which 15 generalizes the one-dimensional case and assures that the firms have incentives to play the collusive action profile s∗ . Lemma 4. Under Assumptions 3 and 4, if (K, λ) satisfies condition 1 in Proposition 3, then " s∗i ∈ arg max E si X πi,m − m XX m j6=i # λij,m max (Kij,m − πj,m , 0) si , s∗−i . (12) Proof. See Appendix. 4.2 Implementing general action profiles One distinguishing feature of our enforcement mechanism, which implements that action profile s∗ , is to exploit the fact that s∗ maximizes the cartel profit in designing penalty. A natural question, then, is whether it can be extended to implement more general action profiles that are less efficient. This is important because, when the discount factor δ is not sufficiently close to one and the efficient profile s∗ is not implementable, some less efficient action profiles may still be implementable. In this subsection we address this question by first outlining in general what action profiles could be implemented and how our mechanism should be extended to accomplish this and then characterizing by an example the set of action profiles that are implementable for a fixed discount factor. Definition 5. An action profile s = (s1 , s2 , . . . , sn ) is supportable if, for each i, there exists a weight θi = (θ1i , θ2i , . . . , θni ) with θji ≥ 0 for each j 6= i, θii > 0, and Pn i j=1 θj = 1, such that E X n θji πj j=1 X n s ≥E θji πj j=1 0 si , s−i for each s0i . Remark 1. By setting θji = 0 if j 6= i and θii = 1, for each i and each j, each Nash equilibrium is supportable. By setting θ1 = θ2 = · · · = θn , each efficient action profile is supportable. 16 Suppose that (θ1 , θ2 , . . . , θn ) is a profile of weights that support an action profile s. We still set each firm i’s continuation payoff at end of each period to be X wi (π) = vi − λij max (Kij − πj , 0) (1 − δ)δ −1 . (13) j6=i Note that Z πi Z X 0 πi dFi (πi , si , s−i ) − λij πi Z j6=i πi = πi dFi (πi , s0i , s−i ) πi − + X Kij πj (θii )−1 θji Z j6=i X (θii )−1 θji j6=i Z πj (Kij − πj ) dFj (πj , s0i , s−i ) πj dFj (πj , s0i , s−i ) πj πj dFj (πj , s0i , s−i ) πj − X πj j6=i Z Kij λij (Kij − πj ) dFj (πj , s0i , s−i ) . πj Therefore, if we set the constants λij such that Z πj Z Kij ∂Fj (πj , s) ∂Fj (πj , s) i −1 i (θi ) θj dπj = λij dπj , 0 ∂si ∂s0i πj πj then similar arguments can be used to show that, given such reward functions wi , it is optimal for firms to follow the action profile s, and the results in Section ?? still hold. Example: Bertrand competition with linear demand and additive shock. For illustration we consider a special case of the Bertrand competition discussed in Section 2. Suppose that there are two firms and demand of each firm i takes the form qi = qbi (p1 , p2 ) + εi = 1 − pi + αpj + εi . where α ∈ (0, 1) and the additive shock εi is uniformly distributed on the interval [−1/2, 1/2]. For simplicity we set the constant marginal cost ci to be zero. We want to characterize the sets of price pairs (p1 , p2 ) that are supportable and, for each supportable price pair, the pairs of weights that support it. Suppose that θi = (θ1i , θ2i ), where i = 1, 2, satisfies θii > 0, θji ≥ 0, and θii + θji = 1. For p1 to maximize E[θ11 p1 q1 + θ21 p2 q2 ] = θ11 p1 (1 − p1 + αp2 ) + θ21 p2 (1 − p2 + αp1 ) , 17 we must have θ11 − 2θ11 p1 + αp2 = 0. Similarly, for p2 to maximize E[θ12 p1 q1 + θ22 p2 q2 ], we must have θ22 − 2θ22 p2 + αp1 = 0. Solving for θ11 and θ22 yields θ11 = αp2 2p1 − 1 θ22 = and αp1 . 2p2 − 1 The constraints 0 < θ11 ≤ 1 and 0 < θ22 ≤ 1 translate into 1 − 2p1 + αp2 ≤ 0 1 − 2p2 + αp1 ≤ 0. and (14) The nonnegativity constraints q1 ≥ 0 and q2 ≥ 0 require that 1 − p1 + αp2 ≥ 0, and 1 − p2 + αp1 ≥ 0. (15) Therefore the set of supportable pairs (p1 , p2 ) are those in the quadrangle bounded by the four inequalities in (14) and (15), and each such (p1 , p2 ) is supportable by the pair of weight pair αp2 αp2 1 θ = ,1 − 2p1 − 1 2p1 − 1 A 2 and θ = αp1 αp1 , 1− 2p2 − 1 2p2 − 1 . Illustration of Assumption 2 A.1 Proof of Lemma 1 Suppose that the random demand shock εi1 has a distribution Gi1 on a finite support [εi1 , εi1 ] with εi1 ≥ 0, and conditional on εi1 , the shock εi2 has a smooth distribution Gi2 (·|εi1 ) on a finite support [εi2 , εi2 ]. We use πi to denote the profit of firm i. Given the prices p and the noises εi1 and εi2 , the profit πi can be calculated to be πi = (pi − ci ) qi (p, εi1 , εi2 ). 18 Therefore πi has the distribution Z εi1 −1 Fi (πi , p) ≡ Gi2 h εi1 πi pi − ci − εi1 qbi (p) εi1 dGi1 (εi1 ) . That Fi satisfies Assumption 1 is clear as qbi is increasing in pj for each j 6= i. To see that Fi also satisfies Assumption 2, we fix a j different from i and differentiate Fi with respect to pj to get Z εi1 ∂Fi πi ∂b qi −1 =− gi2 h − εi1 qbi (p) εi1 εi1 dGi1 (εi1 ) , ∂qj pi − ci ∂pj εi1 where gi2 denotes the density function of εi2 . Now, integrating ∂Fi /∂qj over the truncated profits and changing the order of integrations, we obtain Z K ∂Fi dπi π i ∂qj Z K Z εi1 ∂b qi πi −1 − εi1 qbi (p) εi1 εi1 dGi1 (εi1 ) dπi =− gi2 h pi − ci ∂pj εi1 πi Z K Z ∂b qi εi1 πi −1 =− εi1 gi2 h − εi1 qbi (p) εi1 dπi dGi1 (εi1 ) . ∂pj εi1 pi − ci πi Next we proceed separately. First suppose that the function h is linear and h (εi1 qbi (p) + εi2 ) = k (εi1 qbi (p) + εi2 ) + l, where k and l are constants and k > 0. Then Z K ∂Fi dπi π i ∂qj Z Z K ∂b qi εi1 1 πi =− εi1 gi2 − l − εi1 qbi (p) εi1 dπi dGi1 (εi1 ) ∂pj εi1 k pi − ci πi Z εi1 1 K ∂b qi =− k(pi − ci ) εi1 Gi2 − l − εi1 qbi (p) εi1 dGi1 (εi1 ) , ∂pj k pi − ci εi1 and Z πi πi ∂Fi ∂b qi dπi = − k(pi − ci ) ∂qj ∂pj 19 Z εi1 εi1 dGi1 (εi1 ) . εi1 So we have the following expression for the ratio of the relative effects of a marginal change in pj , R K ∂Fi π ∂qj dπi πi ∂Fi ∂qj dπi R πii R εi1 = εi1 εi1 Gi2 1 k K pi −ci − l − εi1 qbi (p) εi1 dGi1 (εi1 ) R εi1 εi1 εi1 dGi1 (εi1 ) , which is decreasing in pj as qbi (p) is increasing in pj . Now suppose, instead, that εi1 is a constant and the density gi2 is log-concave. In this case we have R K ∂Fi π ∂qj πi ∂Fi ∂qj R πii πi −1 g h − ε q b (p) dπi i1 i pi −ci π i i2 . = Rπ i πi −1 dπi g h − ε q b (p) dπ i2 i1 i i pi −ci π RK dπi i To see that this ratio is weakly decreasing in pj , it suffices to show that RK πi −1 g h − ε q b (p) dπi i2 i1 i pi −ci πi R ≡ Rπ i i − εi1 qbi (p) dπi g h−1 piπ−c K i2 i is weakly decreasing in ph . We differentiate it with respect to pj to get R RK R πi 0 K 0 g dπ g dπ i2 i i dπ g i2 dR ∂b qi π i i π , = −εi1 · R Ki − RKπi i2 R πi dpj ∂pj gi2 dπi gi2 dπi gi2 dπi K πi K where, to simplify notation, we omit the argument of gi2 . Note that ∂b qi /∂pj ≤ 0 and εi1 ≥ 0, to have dR/dpj ≥ 0, it suffices that RK 0 R πi 0 g dπi g dπi π i i2 ≥ RKπi i2 . RK g dπ g dπ i2 i i2 i K π i By denoting γ1 (πi ) = R K πi gi2 gi2 dπi and γ2 (πi ) = R πi K gi2 gi2 dπi , we can rewrite the above inequality as Z K 0 Z πi 0 gi2 gi2 γ1 (πi ) dπi ≥ γ2 (πi ) dπi , K gi2 π i gi2 0 which follows from the fact that h−1 is increasing (since h is), and that gi2 /gi2 is decreasing (since gi2 is log-concave). 20 A.2 Cournot Competition Suppose now that the firms are competing in quantity. Each firm i decides its level of production qi ∈ [0, q] and has a cost function ci (·). Let q ≡ (q1 , . . . , qn ) be a profile of production decisions. Firm i’s stochastic inverse demand is pi (q, εi1 , εi2 ) = h (εi1 pbi (q) + εi2 ) , where h is strictly increasing, εi1 and εi2 are random shocks, and pbi (q) is positive for all q, decreasing in qj for each j. The random demand shock εi1 has a distribution Gi1 on a finite support [εi1 , εi1 ] with εi1 ≥ 0, and conditional on εi1 , the shock εi2 has a smooth distribution Gi2 (·|εi1 ) on a finite support [εi2 , εi2 ]. The profit of firm i, given q, εi1 , and εi2 , can be calculated to be πi = pi (q, εi1 , εi2 ) qi − ci (qi ) , and has the distribution Z εi1 πi + ci (qi ) −1 Gi2 h Fi (πi , q) ≡ − εi1 pbi (q) εi1 dGi1 (εi1 ) . qi εi1 The following result can be proven similarly. Lemma 5. Each distribution function Fi satisfies Assumptions 1 and 2 (with sj = −qj for each j) if either (1) h is linear or (2) εi1 is a constant and the density of εi2 is log-concave. B Proof of Proposition 2 By the assumption that κ1 < |∂Fj (πj , s∗ ) /∂si | < κ2 , we have πj ∂Fj (πj ,s∗ ) ∂si dπj πj ∂Fj (πj ,s∗ ) ∂si dπj R πj λij = R Kij ≤ π j − π j κ2 . Kij − π j κ1 Let be the smallest real number such that Pr (|πi − E[πi |s]| ≥ | s) ≤ , 21 for each i and each s. We set Kij = E[πj |s∗ ] + . Since Z Kij Z ∗ Kij −2 Fj (πj , s ) dπj = πj Z ∗ Kij Fj (πj , s ) dπj + Fj (πj , s∗ ) dπj Kij −2 πj ≤(Kij − 2 − π j ) + 2 ≤ (Kij − π j + 2), we have X Z Kij λij X π j − π j κ2 (Kij − π j + 2), Kij − π j κ1 j6=i Fj (πj , s∗ ) dπj ≤ πj j6=i which tends to zero as tends to zero. Therefore condition 2 holds for all δ sufficiently close to one and sufficiently close to zero. This also implies that the efficiency loss Z Kij X XX ∗ λij (E[πi |s ] − vi ) = Fj (πj , s∗ ) dπj i i πj j6=i tends to zero as tends to zero. C Proof of Lemma 4 Given the constants λij,m satisfying Z Kij,m λij,m π j,m ∂Fj,m (πj,m , φj,m (s∗ )) dπj,m = ∂φj,m Z π j,m π j,m ∂Fj,m (πj,m , φj,m (s∗ )) dπj,m , ∂φj,m (16) we need to show that, for each firm i, the collusive action s∗i maximizes XZ m π i,m πi,m dFi,m πi,m , φj,m si , s∗−i π i,m − XX j6=i m Z Kij,m λij,m (Kij,m − πj,m ) dFj,m πj,m , φj,m si , s∗−i π j,m 22 . Again, from the assumption that s∗ maximizes the total industrial profits, it follows that s∗i also maximizes the total industrial profits, given that the other firms are following s∗−i . Therefore, it suffices to show that, for each si , − XX j6=i Z XX j6=i Z Kij,m λij,m (Kij,m − πj,m ) dFj,m πj,m , φj,m si , s∗−i π j,m m j6=i (17) π j,m XXZ m (Kij,m − πj,m ) dFj,m (πj,m , φj,m (s∗ )) π j,m m + ≥ Kij,m λij,m ∗ πj,m dFj,m (πj,m , φj,m (s )) π j,m − XXZ j6=i m π j,m πj,m dFj,m πj,m , φj,m si , s∗−i . π j,m Using integration by parts, we can rewrite this as Z Kij,m XX λij,m Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m si , s∗−i dπj,m j6=i ≤ π j,m m XXZ j6=i m π j,m Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m si , s∗−i dπj,m . π j,m (18) Define sti = ts∗i + (1 − t)si and note that φj,m sti , s∗−i is quasi-concave in t. Consider the following two cases separately. Case 1. Suppose that φj,m si , s∗−i ≥ φj,m (s∗ ). Then, since φj,m sti , s∗−i is quasi-concave in t, either φj,m sti , s∗−i is decreasing in the interval [0, 1] (from φj,m si , s∗−i to φj,m (s∗ )), or it is first increasing and then decreasing. In the second case there must exist some t0 such that φj,m sti0 , s∗−i = φj,m si , s∗−i and φj,m sti , s∗−i is decreasing in the interval [t0 , 1]. Note that t0 = 0 also has these properties in the first case. Therefore in both cases we have φj,m (s∗ ) ≤ φj,m sti , s∗−i ≤ φj,m sti0 , s∗−i = φj,m si , s∗−i for each t ∈ [t0 , 1]. Then, for each t ∈ [t0 , 1], we have, by the assumption of 23 quasi-concavity, X ∂φj,m sti , s∗−i 1 (s∗i,l − si,l ) = − ∇φj,m sti , s∗−i · si − sti ≤ 0, ∂si,l t m (19) and, by Assumption 3, Kij,m Z λij,m π j,m ∂Fj,m πj,m , φj,m sti , s∗−i dπj,m ∂φj,m Z πj,m ∂Fj,m πj,m , φj,m sti , s∗−i dπj,m ≥ 0. (20) − ∂φj,m π j,m Hence, from φj,m sti0 , s∗−i = φj,m si , s∗−i , we have Z Kij,m λij,m Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m si , s∗−i dπj,m π j,m Z π j,m − Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m si , s∗−i dπj,m π j,m Z Kij,m =λij,m Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m sti0 , s∗−i dπj,m π j,m Z π j,m − Fj,m (πj,m , φj,m (s∗ )) − Fj,m πj,m , φj,m sti0 , s∗−i dπj,m . π j,m The right-hand side, by rewriting the integrand, is, Z Kij,m Z 1 ∂Fj,m πj,m , φj,m sti , s∗−i X ∂φj,m sti , s∗−i ∗ (si,l − si,l )dt dπj,m λij,m ∂φj,m ∂s i,l t0 π j,m m Z πj,m Z 1 ∂Fj,m πj,m , φj,m sti , s∗−i X ∂φj,m sti , s∗−i ∗ (si,l − si,l )dt dπj,m − ∂φj,m ∂si,l π j,m t0 m or, by Fubini’s theorem, is Z 1X Z Kij,m ∂φj,m sti , s∗−i ∗ ∂Fj,m πj,m , φj,m sti , s∗−i (si,l − si,l ) λij,m dπj,m ∂si,l ∂φj,m π j,m t0 m Z πj,m ∂Fj,m πj,m , φj,m sti , s∗−i dπj,m dt, − ∂φj,m π j,m which is less than or equal to zero by (19) and (20). This proves (17). 24 Case 2. Suppose that φj,m si , s∗−i ≤ φj,m (s∗ ). Then, since φj,m sti , s∗−i is quasi-concave in t, either φj,m sti , s∗−i is increasing in the interval [0, 1] (from φj,m si , s∗−i to φj,m (s∗ )), or it is first increasing and then decreasing. In the second case there must exist some t0 such that φj,m sti0 , s∗−i = φj,m si , s∗−i and φj,m sti , s∗−i is increasing in the interval [0, t0 ]. Note that t0 = 1 also has these properties in the first case. Therefore φj,m sti , s∗−i ≤ φj,m sti0 , s∗−i = φj,m (s∗ ) for each t ∈ [0, t0 ]. Then, for each t ∈ [0, t0 ], we have, by the assumption of quasi-concavity, X ∂φj,m sti , s∗−i 1 (s∗i,l − si,l ) = ∇φj,m sti , s∗−i · s∗i − sti ≥ 0, ∂si,l 1−t m and, by Assumption 3, Z Kij,m λij,m π j,m ∂Fj,m πj,m , φj,m sti , s∗−i dπj,m ∂φj,m Z πj,m ∂Fj,m πj,m , φj,m sti , s∗−i − dπj,m ≤ 0. ∂φj,m π j,m With these results (17) can be similarly proven as in Case 1. References Aoyagi, Masaki, 2002, Collusion in dynamic bertrand oligopoly with correlated private signals and communication, Journal of Economic Theory 102, 229 – 248. Compte, Olivier, 1998, Communication in repeated games with imperfect private monitoring, Econometrica 66, 597–626. Fudenberg, Drew, and David K. Levine, 2007, The Nash-threats folk theorem with communication and approximate common knowledge in two player games, Journal of Economic Theory 132, 461 – 473. Green, Edward J., and Robert H. Porter, 1984, Noncooperative collusion under imperfect price information, Econometrica 52, 87–100. 25 Harrington, Joseph E. Jr., 2006, How do cartels operate?, Foundations and Trends in Microeconomics 2, 1–105. , and Andrzej Skrzypacz, 2007, Collusion under monitoring of sales, The RAND Journal of Economics 38, 314–331. , 2008, Collusion with monitoring based on self-reported sales, Working Paper, Dept of Economics, Johns Hopkins University. Kandori, Michihiro, and Hitoshi Matsushima, 1998, Private observation, communication and collusion, Econometrica 66, 627–652. Levenstein, Margaret C., and Valerie Y. Suslow, 2006, What determines cartel success?, Journal of Economic Literature 44, 43–95. Obara, Ichiro, 2008, Folk theorem with communication, Journal of Economic Theory In Press, Corrected Proof, –. Stigler, George J., 1964, A theory of oligopoly, The Journal of Political Economy 72, 44–61. Suslow, Valerie Y., 2005, Cartel contract duration: Empirical evidence from interwar international cartels, Industrial and Corporate Change 14, 705 – 744. Zheng, Bingyong, 2008, Approximate efficiency in repeated games with correlated private signals, Games and Economic Behavior 63, 406 – 416. 26
© Copyright 2025 Paperzz