Folk theorem (game theory)

From Wikipedia, the free encyclopedia

In game theory, folk theorems are a class of theorems describing an abundance of Nash equilibrium payoff profiles in repeated games (Friedman 1971).[1] The original Folk Theorem concerned the payoffs of all the Nash equilibria of an infinitely repeated game. This result was called the Folk Theorem because it was widely known among game theorists in the 1950s, even though no one had published it. Friedman's (1971) Theorem concerns the payoffs of certain subgame-perfect Nash equilibria (SPE) of an infinitely repeated game, and so strengthens the original Folk Theorem by using a stronger equilibrium concept: subgame-perfect Nash equilibria rather than Nash equilibria.[2]

The Folk Theorem suggests that if the players are patient enough and far-sighted (i.e. if the discount factor ), then repeated interaction can result in virtually any average payoff in an SPE equilibrium.[3] "Virtually any" is here technically defined as "feasible" and "individually rational".

Setup and definitions[edit]

We start with a basic game, also known as the stage game, which is an n-player game. In this game, each player has finitely many actions to choose from, and they make their choices simultaneously and without knowledge of the other player's choices. The collective choices of the players leads to a payoff profile, i.e. to a payoff for each of the players. The mapping from collective choices to payoff profiles is known to the players, and each player aims to maximize their payoff. If the collective choice is denoted by x, the payoff that player i receives, also known as player i's utility, will be denoted by .

We then consider a repetition of this stage game, finitely or infinitely many times. In each repetition, each player chooses one of their stage game options, and when making that choice, they may take into account the choices of the other players in the prior iterations. In this repeated game, a strategy for one of the players is a deterministic rule that specifies the player's choice in each iteration of the stage game, based on all other player's choices in the prior iterations. A choice of strategy for each of the players is a strategy profile, and it leads to a payout profile for the repeated game. There are a number of different ways such a strategy profile can be translated into a payout profile, outlined below.

Any Nash equilibrium payoff profile of a repeated game must satisfy two properties:

  1. Individual rationality: the payoff must weakly dominate the minmax payoff profile of the constituent stage game. That is, the equilibrium payoff of each player must be at least as large as the minmax payoff of that player. This is because a player achieving less than their minmax payoff always has incentive to deviate by simply playing their minmax strategy at every history.
  2. Feasibility: the payoff must be a convex combination of possible payoff profiles of the stage game. This is because the payoff in a repeated game is just a weighted average of payoffs in the basic games.

Folk theorems are partially converse claims: they say that, under certain conditions (which are different in each folk theorem), every payoff profile that is both individually rational and feasible can be realized as a Nash equilibrium payoff profile of the repeated game.

There are various folk theorems; some relate to finitely-repeated games while others relate to infinitely-repeated games.[4]

Infinitely-repeated games without discounting[edit]

In the undiscounted model, the players are patient. They don't differentiate between utilities in different time periods. Hence, their utility in the repeated game is represented by the sum of utilities in the basic games.

When the game is infinite, a common model for the utility in the infinitely-repeated game is the limit inferior of mean utility: If the game results in a path of outcomes , where denotes the collective choices of the players at iteration t (t=0,1,2,...), player i's utility is defined as

where is the basic-game utility function of player i.

An infinitely-repeated game without discounting is often called a "supergame".

The folk theorem in this case is very simple and contains no pre-conditions: every individually rational and feasible payoff profile in the basic game is a Nash equilibrium payoff profile in the repeated game.

The proof employs what is called a grim[5] or grim trigger[6] strategy. All players start by playing the prescribed action and continue to do so until someone deviates. If player i deviates, all other players switch to picking the action which minmaxes player i forever after. The one-stage gain from deviation contributes 0 to the total utility of player i. The utility of a deviating player cannot be higher than his minmax payoff. Hence all players stay on the intended path and this is indeed a Nash equilibrium.

Subgame perfection[edit]

The above Nash equilibrium is not always subgame perfect. If punishment is costly for the punishers, the threat of punishment is not credible.

A subgame perfect equilibrium requires a slightly more complicated strategy.[5][7]: 146–149  The punishment should not last forever; it should last only a finite time which is sufficient to wipe out the gains from deviation. After that, the other players should return to the equilibrium path.

The limit-of-means criterion ensures that any finite-time punishment has no effect on the final outcome. Hence, limited-time punishment is a subgame-perfect equilibrium.

  • Coalition subgame-perfect equilibria:[8] An equilibrium is called a coalition Nash equilibrium if no coalition can gain from deviating. It is called a coalition subgame-perfect equilibrium if no coalition can gain from deviating after any history.[9] With the limit-of-means criterion, a payoff profile is attainable in coalition-Nash-equilibrium or in coalition-subgame-perfect-equilibrium, if-and-only-if it is Pareto efficient and weakly-coalition-individually-rational.[10]

Overtaking[edit]

Some authors claim that the limit-of-means criterion is unrealistic, because it implies that utilities in any finite time-span contribute 0 to the total utility. However, if the utilities in any finite time-span contribute a positive value, and the value is undiscounted, then it is impossible to attribute a finite numeric utility to an infinite outcome sequence. A possible solution to this problem is that, instead of defining a numeric utility for each infinite outcome sequence, we just define the preference relation between two infinite sequences. We say that agent (strictly) prefers the sequence of outcomes over the sequence , if:[6][7]: 139 [8]

For example, consider the sequences and . According to the limit-of-means criterion, they provide the same utility to player i, but according to the overtaking criterion, is better than for player i. See overtaking criterion for more information.

The folk theorems with the overtaking criterion are slightly weaker than with the limit-of-means criterion. Only outcomes that are strictly individually rational, can be attained in Nash equilibrium. This is because, if an agent deviates, he gains in the short run, and this gain can be wiped out only if the punishment gives the deviator strictly less utility than the agreement path. The following folk theorems are known for the overtaking criterion:

  • Strict stationary equilibria:[6] A Nash equilibrium is called strict if each player strictly prefers the infinite sequence of outcomes attained in equilibrium, over any other sequence he can deviate to. A Nash equilibrium is called stationary if the outcome is the same in each time-period. An outcome is attainable in strict-stationary-equilibrium if-and-only-if for every player the outcome is strictly better than the player's minimax outcome.[11]
  • Strict stationary subgame-perfect equilibria:[6] An outcome is attainable in strict-stationary-subgame-perfect-equilibrium, if for every player the outcome is strictly better than the player's minimax outcome (note that this is not an "if-and-only-if" result). To achieve subgame-perfect equilibrium with the overtaking criterion, it is required to punish not only the player that deviates from the agreement path, but also every player that does not cooperate in punishing the deviant.[7]: 149–150 
    • The "stationary equilibrium" concept can be generalized to a "periodic equilibrium", in which a finite number of outcomes is repeated periodically, and the payoff in a period is the arithmetic mean of the payoffs in the outcomes. That mean payoff should be strictly above the minimax payoff.[6]
  • Strict stationary coalition equilibria:[8] With the overtaking criterion, if an outcome is attainable in coalition-Nash-equilibrium, then it is Pareto efficient and weakly-coalition-individually-rational. On the other hand, if it is Pareto efficient and strongly-coalition-individually-rational[12] it can be attained in strict-stationary-coalition-equilibrium.

Infinitely-repeated games with discounting[edit]

Assume that the payoff of a player in an infinitely repeated game is given by the average discounted criterion with discount factor 0 < δ < 1:

The discount factor indicates how patient the players are. The factor is introduced so that the payoff remain bounded when .

The folk theorem in this case requires that the payoff profile in the repeated game strictly dominates the minmax payoff profile (i.e., each player receives strictly more than the minmax payoff).

Let a be a strategy profile of the stage game with payoff profile u which strictly dominates the minmax payoff profile. One can define a Nash equilibrium of the game with u as resulting payoff profile as follows:

1. All players start by playing a and continue to play a if no deviation occurs.
2. If any one player, say player i, deviated, play the strategy profile m which minmaxes i forever after.
3. Ignore multilateral deviations.

If player i gets ε more than his minmax payoff each stage by following 1, then the potential loss from punishment is

If δ is close to 1, this outweighs any finite one-stage gain, making the strategy a Nash equilibrium.

An alternative statement of this folk theorem[4] allows the equilibrium payoff profile u to be any individually rational feasible payoff profile; it only requires there exist an individually rational feasible payoff profile that strictly dominates the minmax payoff profile. Then, the folk theorem guarantees that it is possible to approach u in equilibrium to any desired precision (for every ε there exists a Nash equilibrium where the payoff profile is a distance ε away from u).

Subgame perfection[edit]

Attaining a subgame perfect equilibrium in discounted games is more difficult than in undiscounted games. The cost of punishment does not vanish (as with the limit-of-means criterion). It is not always possible to punish the non-punishers endlessly (as with the overtaking criterion) since the discount factor makes punishments far away in the future irrelevant for the present. Hence, a different approach is needed: the punishers should be rewarded.

This requires an additional assumption, that the set of feasible payoff profiles is full dimensional and the min-max profile lies in its interior. The strategy is as follows.

1. All players start by playing a and continue to play a if no deviation occurs.
2. If any one player, say player i, deviated, play the strategy profile m which minmaxes i for N periods. (Choose N and δ large enough so that no player has incentive to deviate from phase 1.)
3. If no players deviated from phase 2, all player ji gets rewarded ε above j's min-max forever after, while player i continues receiving his min-max. (Full-dimensionality and the interior assumption is needed here.)
4. If player j deviated from phase 2, all players restart phase 2 with j as target.
5. Ignore multilateral deviations.

Player ji now has no incentive to deviate from the punishment phase 2. This proves the subgame perfect folk theorem.

Finitely-repeated games without discount[edit]

Assume that the payoff of player i in a game that is repeated T times is given by a simple arithmetic mean:

A folk theorem for this case has the following additional requirement:[4]

In the basic game, for every player i, there is a Nash-equilibrium that is strictly better, for i, than his minmax payoff.

This requirement is stronger than the requirement for discounted infinite games, which is in turn stronger than the requirement for undiscounted infinite games.

This requirement is needed because of the last step. In the last step, the only stable outcome is a Nash-equilibrium in the basic game. Suppose a player i gains nothing from the Nash equilibrium (since it gives him only his minmax payoff). Then, there is no way to punish that player.

On the other hand, if for every player there is a basic equilibrium which is strictly better than minmax, a repeated-game equilibrium can be constructed in two phases:

  1. In the first phase, the players alternate strategies in the required frequencies to approximate the desired payoff profile.
  2. In the last phase, the players play the preferred equilibrium of each of the players in turn.

In the last phase, no player deviates since the actions are already a basic-game equilibrium. If an agent deviates in the first phase, he can be punished by minmaxing him in the last phase. If the game is sufficiently long, the effect of the last phase is negligible, so the equilibrium payoff approaches the desired profile.

Applications[edit]

Folk theorems can be applied to a diverse number of fields. For example:

  • Anthropology: in a community where all behavior is well known, and where members of the community know that they will continue to have to deal with each other, then any pattern of behavior (traditions, taboos, etc.) may be sustained by social norms so long as the individuals of the community are better off remaining in the community than they would be leaving the community (the minimax condition).
  • International politics: agreements between countries cannot be effectively enforced. They are kept, however, because relations between countries are long-term and countries can use "minimax strategies" against each other. This possibility often depends on the discount factor of the relevant countries. If a country is very impatient (pays little attention to future outcomes), then it may be difficult to punish it (or punish it in a credible way).[5]

On the other hand, MIT economist Franklin Fisher has noted that the folk theorem is not a positive theory.[13] In considering, for instance, oligopoly behavior, the folk theorem does not tell the economist what firms will do, but rather that cost and demand functions are not sufficient for a general theory of oligopoly, and the economists must include the context within which oligopolies operate in their theory.[13]

In 2007, Borgs et al. proved that, despite the folk theorem, in the general case computing the Nash equilibria for repeated games is not easier than computing the Nash equilibria for one-shot finite games, a problem which lies in the PPAD complexity class.[14] The practical consequence of this is that no efficient (polynomial-time) algorithm is known that computes the strategies required by folk theorems in the general case.

Summary of folk theorems[edit]

The following table compares various folk theorems in several aspects:

  • Horizon – whether the stage game is repeated finitely or infinitely many times.
  • Utilities – how the utility of a player in the repeated game is determined from the player's utilities in the stage game iterations.
  • Conditions on G (the stage game) – whether there are any technical conditions that should hold in the one-shot game in order for the theorem to work.
  • Conditions on x (the target payoff vector of the repeated game) – whether the theorem works for any individually rational and feasible payoff vector, or only on a subset of these vectors.
  • Equilibrium type – if all conditions are met, what kind of equilibrium is guaranteed by the theorem – Nash or Subgame-perfect?
  • Punishment type – what kind of punishment strategy is used to deter players from deviating?
Published by Horizon Utilities Conditions on G Conditions on x Guarantee Equilibrium type Punishment type
Benoit& Krishna[15] Finite () Arithmetic mean For every player there is an equilibrium payoff strictly better than minimax. None For all there is such that, if , for every there is equilibrium with payoff -close to . Nash
Aumann& Shapley[5] Infinite Limit of means None None Payoff exactly . Nash Grim
Aumann& Shapley[5] and Rubinstein[8][16] Infinite Limit of means None None Payoff exactly . Subgame-perfect Limited-time punishment.[7]: 146–149 
Rubinstein[6] Infinite Overtaking None Strictly above minimax. Single outcome or a periodic sequence. Subgame-perfect Punishing non-punishers.[7]: 149–150 
Rubinstein[8] Infinite Limit of means None Pareto-efficient and weakly-coalition-individually-rational[10] None Coalition-subgame-perfect
Rubinstein[8] Infinite Overtaking None Pareto-efficient and strongly-coalition-individually-rational[12] None Coalition-Nash
Fudenberg& Maskin[17] Infinite Sum with discount Correlated mixed strategies are allowed. Strictly above minimax. When is sufficiently near 1, there is an equilibrium with payoff exactly . Nash Grim
Fudenberg& Maskin[17] Infinite Sum with discount Only pure strategies are allowed. Strictly above minimax. For all there is such that, if , for every there is an equilibrium with payoff -close to . Nash Grim punishment.
Friedman (1971, 1977) Infinite Sum with discount Correlated mixed strategies are allowed. Strictly above a Nash-equilibrium in G. When is sufficiently near 1, there is equilibrium with payoff exactly . Subgame-perfect Grim punishment using the Nash-equilibrium.
Fudenberg& Maskin[17] Infinite Sum with discount Two players Strictly above minimax. For all there is such that, if , there is equilibrium with payoff exactly . Subgame-perfect Limited-time punishment.
Fudenberg& Maskin[17] Infinite Sum with discount The IR feasible space is full-dimensional.[18] Strictly above minimax. For all there is such that, if , there is equilibrium with payoff exactly . Subgame-perfect Rewarding the punishers.[7]: 150–153 

Folk theorems in other settings[edit]

In allusion to the folk theorems for repeated games, some authors have used the term "folk theorem" to refer to results on the set of possible equilibria or equilibrium payoffs in other settings, especially if the results are similar in what equilibrium payoffs they allow. For instance, Tennenholtz proves a "folk theorem" for program equilibrium. Many other folk theorems have been proved in settings with commitment.[19][20][21][22]

Notes[edit]

  1. ^ In mathematics, the term folk theorem refers generally to any theorem that is believed and discussed, but has not been published. Roger Myerson has recommended the more descriptive term "general feasibility theorem" for the game theory theorems discussed here. See Myerson, Roger B. Game Theory, Analysis of conflict, Cambridge, Harvard University Press (1991)
  2. ^ R. Gibbons (1992). A Primer in Game Theory. Harvester Wheatsheaf. p. 89. ISBN 0-7450-1160-8.
  3. ^ Jonathan Levin (2002). "Bargaining and Repeated Games" (PDF).
  4. ^ a b c Michael Maschler; Eilon Solan; Shmuel Zamir (2013). Game Theory. Cambridge University Press. pp. 176–180. ISBN 978-1-107-00548-8.
  5. ^ a b c d e Aumann, Robert J.; Shapley, Lloyd S. (1994). "Long-Term Competition—A Game-Theoretic Analysis" (PDF). Essays in Game Theory. p. 1. doi:10.1007/978-1-4612-2648-2_1. ISBN 978-1-4612-7621-0.
  6. ^ a b c d e f Rubinstein, Ariel (1979). "Equilibrium in supergames with the overtaking criterion". Journal of Economic Theory. 21: 1–9. doi:10.1016/0022-0531(79)90002-4.
  7. ^ a b c d e f Osborne, Martin J.; Rubinstein, Ariel (12 July 1994). A Course in Game Theory. ISBN 0-262-15041-7. LCCN 94008308. OL 1084491M.
  8. ^ a b c d e f Rubinstein, A. (1980). "Strong perfect equilibrium in supergames". International Journal of Game Theory. 9: 1–12. doi:10.1007/BF01784792. S2CID 122098115.
  9. ^ The paper uses the term "strong equilibrium". Here, to prevent ambiguity, the term "coalition equilibrium" is used instead.
  10. ^ a b For every nonempty coalition , there is a strategy of the other players () such that for any strategy played by , the payoff when plays is not [strictly better for all members of ].
  11. ^ In the 1979 paper, Rubinstein claims that an outcome is attainable in strict-stationary-equilibrium if-and-only-if for every player, the outcome is EITHER strictly better than the player's minimax outcome OR the outcome is weakly better than any other outcome the player can unilaterally deviate to. It is not clear how the second option is attainable in a strict equilibrium. In the 1994 book, this claim does not appear.
  12. ^ a b for every nonempty coalition , there is a strategy of the other players () such that for any strategy played by , the payoff is strictly worse for at least one member of .
  13. ^ a b Fisher, Franklin M. Games Economists Play: A Noncooperative View The RAND Journal of Economics, Vol. 20, No. 1. (Spring, 1989), pp. 113–124, this particular discussion is on page 118
  14. ^ Christian Borgs; Jennifer Chayes; Nicole Immorlica; Adam Tauman Kalai; Vahab Mirrokni; Christos Papadimitriou (2007). "The Myth of the Folk Theorem" (PDF). pp. 365–372.
  15. ^ Benoit, Jean-Pierre; Krishna, Vijay (1985). "Finitely Repeated Games". Econometrica. 53 (4): 905. doi:10.2307/1912660. JSTOR 1912660.
  16. ^ Rubinstein, Ariel (1994). "Equilibrium in Supergames". Essays in Game Theory. pp. 17–27. doi:10.1007/978-1-4612-2648-2_2. ISBN 978-1-4612-7621-0.
  17. ^ a b c d Fudenberg, Drew; Maskin, Eric (1986). "The Folk Theorem in Repeated Games with Discounting or with Incomplete Information". Econometrica. 54 (3): 533. CiteSeerX 10.1.1.308.5775. doi:10.2307/1911307. JSTOR 1911307.
  18. ^ There is a collection of IR feasible outcomes , one per player, such that for every players , and .
  19. ^ Forges, Françoise (2013). "A folk theorem for Bayesian games with commitment". Games and Economic Behavior. 78: 64–71. doi:10.1016/j.geb.2012.11.004.
  20. ^ Oesterheld, Caspar; Treutlein, Johannes; Grosse, Roger; Conitzer, Vincent; Foerster, Jakob (2023). "Similarity-based Cooperative Equilibrium". Proceedings of the Neural Information Processing Systems (NeurIPS).
  21. ^ Kalai, Adam Tauman; Kalai, Ehud; Lehrer, Ehud; Samet, Dov (2010). "A commitment folk theorem". Games and Economic Behavior. 69 (1): 127–137. doi:10.1016/j.geb.2009.09.008.
  22. ^ Block, Juan I.; Levine, David K. (2017). "A folk theorem with codes of conduct and communication". Economic Theory Bulletin. 5 (1): 9–19. doi:10.1007/s40505-016-0107-y.

References[edit]