Well-behaved [and Smooth] Aggregative Games Richard Cornes [ANU] Roger Hartley [Manchester Univ] What is an Aggregative Game? An aggregative game is one in which each player’s payoff function has the form i x i xi , j 1 x j i xi , X . n Examples include: Cournot oligopoly/oligopsony Pure public good provision Open access resources [Tragedy of the commons] Contest theory Diamond’s coconut search model Tragedy of the anticommons Aggregative Games: Individual Behavior A player’s marginal payoff function can be written as i xi , X i xi , X i xi , X xi X Ignoring corner solutions – which are easily accommodated the FOC associated with player i’s best response requires that i xˆi , X 0 Her replacement function [more generally, correspondence] is xˆi ri X A Well-known Example: Cournot Inverse demand function: P X a bX • • Cost Functions: Ci(xi) = cixi vi xi , X a bX xi ci xi i xˆi , X P X xˆi P' X ci 0 • In explicit form, taking account of nonnegativity constraints: a ci bX xˆi ri X max ,0, b defined for ri X X . Linear Cournot Geometry I qi qˆi ri Q qˆi bi Qi Qi n q j 1, j i j Q Aggregative Games: Equilibrium • A Nash noncooperative equilibrium of an n-player game is conventionally described by the n equations provided by the n players’ best response functions. This is problematic. • Using replacement functions, a Nash equilibrium is described by one equation in one unknown – the aggregative replacement function - regardless of how many heterogeneous players there are: n R X r j X X . j 1 • An alternative useful description is obtained by dividing both sides by the quantity X: n rj X n s j X 1 X j 1 j 1 Well-behaved Games • We are interested in games that are wellbehaved – that is, games that possess a unique equilibrium in pure strategies. • Specifically, we will present conditions that – are sufficient for a game to be well-behaved, – are easy to check, – have a straightforward geometric intuition, – are less restrictive than any with which we are familiar Familiar [and Restrictive] Conditions • Suppose we assume i xi , X C 1: 0, 0 xi X xi and i xi , X C2 : 0, 0 xi X X These conditions imply that xˆi i xi , X X 0, X i xi , X xi which implies a downward-sloping replacement function for player i. Hence the aggregate replacement function is downward-sloping wherever it is strictly positive. This establishes uniqueness. Existence is easily established from continuity considerations. Digression into best response functions i xi , X 0 beforeand afteran exogenousshock i xi , X i xi , X xi dxi X dxi dX i 0 i xi , X dxi X i xi , X i xi , X dX i xi X dxi 1 0 dX i These are the familiar bounds on the slopes of best response functions. It’s obvious why they work when n=2. Less obvious, though, when n>2! Linear Cournot Geometry II ri(X) r1(X)+r2(X)+r3(X) EN r1(X)+r2(X)+r3(X)=X r3(X) r2(X) r1(X) 45o X The Familiar Conditions and Cournot In Cournot with possibly nonlinear demand and cost functions, vi xi , X P X xi Ci xi i xi , X P X Ci ' xi xi P' X so that i xi , X C 1: 0 P' X Ci ' ' xi 0 xi and i xi , X C 2: 0 P' X xi P' ' X 0 X C1 and C2 are stated by Hahn [RES 1962] as sufficient for stability in oligopoly. They are stated by Corchon more generally in terms of the marginal payoff function as sufficient for existence and uniqueness in general aggregative games. Some Awkward Examples • Consider Cournot again, but with a constant elasticity demand function. This does not satisfy the familiar condition. But as we will show, it may be perfectly well-behaved. • The same is true of even the simplest Tullock contest • ... and of other examples, such as the static open access resource model, and other cost- and surplussharing games. Our Central Aim We now present conditions, sufficient for a game to be ‘well-behaved’, that are •Less restrictive than Hahn’s conditions •Naturally satisfied by all the awkward examples •Capable of a natural behavioral interpretation •As general as one can expect to find Our Sufficient Condition for ‘Good Behavior’ • Our two conditions are: A1: If xi , X satisfies0 xi X and i xi , X 0 , then i xi , X 0 xi A2 : If xi , X satisfies0 xi X and i xi , X 0 , then i xi , X i xi , X xi X 0 xi X The meaning of the new condition • The familiar (Hahn/Corchon) conditions require the replacement function, ri(X), to be everywhere downward-sloping. • Perhaps more familiar, they impose upper and lower bounds on the negatively sloped best response functions. • Our (new) conditions do not require monotonicity of the replacement function, but of the share function, ri X si X X • We call a player with a monotonic share function ‘regular’. More Geometry The following situation is not allowed by the familiar condition, but is allowed by us: xi xi=ri(X) ri(X*) si(X*) X X* The Share Function: A Simple Example Cournot with unit elasticity: 4 firms, p = 1/X, constant but idiosyncratic unit costs: (c1, c2, c3, c4) = (1, 2, 4, 5) vi xi , X p X xi ci xi 1 xi ci xi X 1 1 FOC : i xi , X ci 2 xˆi 0 X X xˆi ri ( X ) X ci X 2 xˆi si ( X ) 1 ci X X More correctly: si(X) = max{1 – ciX, 0} Our Conditions and Cournot Suppose (i) Inversedemandsat isfies p ' X 0 X 0, Xp X 0 as X 0 and p X 0 asX . (ii) For each i I , we haveci ' xi 0 if xi 0 and ci ' ' xi maxp ' X ,2 p ' X Xp' ' X if 0 xi X . Then our conditions A1 and A2 hold. Our conditions permit some degree of diminishing costs. For example, if demand is linear, diminishing costs do not threaten existence of unique equilibrium as long as ci ‘’ > b. Our Conditions in Other Models • Simple Tullock contests violate the familiar conditions. But they satisfy our less restrictive ones. • Our approach can analyse more complicated Tullock contests, allowing for risk aversion, more general contest success functions, and heterogeneity, and can identify conditions in these more general settings that are sufficient for ‘good behaviour’. •The simple over-fishing problem [tragedy of the commons], too, does not satisfy the familiar conditions, but does satisfy ours under standard assumptions. •Other simple cost and output sharing games have an aggregative structure, and under standard assumptions can be shown to be well-behaved. Beyond Well-behaved Games • Our sufficient conditions for uniqueness should not be regarded as sacred – many games violate them, and may have multiple equilibria. • Our methods – using replacement and share functions – can [and should] be used to analyse any game that has suitably aggregative structure. • The important point to take from this paper is the idea of conditioning player i’s behaviour on a common aggregate, rather than on the sum of all other players’ choices, excluding that of player i herself. Badly Behaved Games: An Example Tullock contest, identical risk averse players: Suppose that u w w 0.45w2 , w1 Then preferred share values are defined implicitly by 1 0.55 0.9Y 0 1 0.9 1 Y Y s(Y) 1 1 0.55 0.566 Y Badly Behaved Games: Example (cont.) • If each of 10 players has the utility function of our example, then there are multiple equilibria: – at the symmetric equilibrium, xi = 0.05633 for all i=1,…,10 – There are asymmetric equilibria in which three players each have x = 0.05633, the remaining seven each have x = 0, – There are asymmetric equilibria in which four players each have x = 0.14, the remaining seven each have x = 0, – And so on Badly Behaved Games: Another Example si(X) r 1 r Tullock contest: Identical risk neutral players pi xir r x j 1 j n Where r > 1 X Badly Behaved Games: Yet another example Diamond’s search model: If you really want to rule out many equilibria in Diamond’s model, you can do so by (i) making search costs rise very steeply with the level of search, and also (ii) Assuming that, even if no-one else searches, player i will find an edible coconut lying on the ground with prob > 0 if he searches. Relaxing our Conditions: I • We do not claim that our conditions are necessary and sufficient. • But we do claim that, if they are relaxed, existence of a unique equilibrium may be jeopardized unless a little extra structure is imposed on the game. Relaxing our Conditions: II • Suppose there are nt identical players of type t with a nonmonotonic share function, • However, suppose that for share values less than 1/nt, their share functions are regular. • Then, the aggregation of their share functions ensures that the aggregate share function is monotonic over the relevant range.