{
  "id": "repeated-game",
  "title": "Repeated Game",
  "lang": "en",
  "source_lang": "en",
  "summary": "A repeated game models recurring strategic interaction in which future rewards, punishments, and information can change present incentives.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Supergame",
    "Iterated game"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/repeated-game/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/repeated-game/",
      "json": "https://aiwiki.page/zh/repeated-game.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A repeated game is a model in game theory in which a fixed group of players plays a given strategic interaction, called the stage game, more than once. Players can condition their current choices on information about previous play. Consequently, an action affects not only the current payoff but also how other players behave in the future. This connection between present conduct and future consequences distinguishes a repeated game from a collection of unrelated one-shot encounters.\n\nRepetition can sustain cooperation even when each player has an immediate incentive to act opportunistically. It can also sustain inefficient outcomes, costly punishment, and unequal distributions of gains. The theory therefore does not assert that long relationships necessarily produce cooperation. Instead, it identifies the conditions under which particular patterns of behavior are consistent with individual incentives.\n\nRepeated games are central to the analysis of continuing relationships among firms, workers and employers, trading partners, and governments. Their mathematical structure depends especially on the duration of interaction, the importance assigned to future payoffs, what players observe, and what they know about one another. These features determine whether promises and threats are credible, which outcomes can be sustained, and how experience changes behavior."
    },
    {
      "heading": "Formal structure and payoff criteria",
      "text": "A stage game can be written as \\(G=(N,(A_i)_{i\\in N},(u_i)_{i\\in N})\\), where \\(N\\) is the set of players, \\(A_i\\) is player \\(i\\)'s action set, and \\(u_i\\) assigns a payoff to each action profile \\(a=(a_1,\\ldots,a_n)\\). In period \\(t\\), the players choose actions \\(a^t\\), receive stage payoffs \\(u_i(a^t)\\), and obtain information about what occurred. The same stage game is then played again. In the standard model, actions do not alter the underlying action sets or payoff functions, although they may alter beliefs and future strategies.\n\nUnder perfect monitoring, the public history before period \\(t\\) is\n\n\\[h^t=(a^0,a^1,\\ldots,a^{t-1}).\\]\n\nA strategy specifies behavior after every possible history, including histories that would not occur if everyone followed the proposed strategy. Formally, a behavioral strategy maps each information history to a probability distribution over available actions. Such randomization is distinct from choosing an entire contingent plan at the outset, although familiar equivalence results relate these descriptions when players have perfect recall.\n\nFor a finite horizon of \\(T\\) periods, payoffs are commonly represented by a sum or average of stage payoffs. For an infinite horizon, a standard specification is normalized discounted utility:\n\n\\[U_i=(1-\\delta_i)\\, \\mathbb{E}\\!\\left[\\sum_{t=0}^{\\infty}\\delta_i^t u_i(a^t)\\right], \\qquad 0\\leq\\delta_i<1.\\]\n\nThe expectation accommodates randomized actions and signals. Player \\(i\\)'s discount factor \\(\\delta_i\\) measures the weight placed on future payoffs. Normalization ensures that receiving a constant stage payoff \\(c\\) forever has value \\(c\\), facilitating comparisons between repeated-game payoffs and stage-game payoff vectors.\n\nDiscounting can represent impatience, the opportunity cost of delayed income, or an exogenous risk that the relationship ends. If an undiscounted relationship independently continues after each period with probability \\(p\\), its expected payoff sum has weights \\(p^t\\). With both ordinary discounting and independent continuation risk, the effective discount factor is their product. A randomly terminated game can therefore have incentives resembling those of an infinite discounted game, even though actual relationships end.\n\nAnother criterion uses long-run average payoffs, such as the lower limit of expected time averages. Discounted and average-payoff formulations are related but not interchangeable without additional assumptions. A repeated game also differs from a stochastic game, where actions can change a payoff-relevant state. In a standard repeated game, history matters strategically; in a stochastic game, past actions can additionally change the physical or economic environment."
    },
    {
      "heading": "Finite repetition and backward induction",
      "text": "When the final period is known, reasoning often begins at the end. In that period, no future within the model remains to reward or punish current behavior. Players must therefore choose a stage-game Nash equilibrium in every final-period subgame. This observation underlies backward induction: solve the last decision problem, substitute its implications into the preceding one, and continue toward the beginning.\n\nThe clearest example is the finitely repeated prisoner’s dilemma. Each player chooses cooperation or defection. Defection strictly dominates cooperation in the stage game, so mutual defection is its unique Nash equilibrium. In the last round, both players defect. Since this last-round outcome cannot be changed by earlier behavior, the same reasoning applies to the penultimate round. Repeating the argument gives defection in every round of the unique subgame-perfect equilibrium.\n\nThe equilibrium refinement matters. A subgame-perfect equilibrium requires strategies to form a Nash equilibrium after every history at which a subgame begins, not merely along the intended path. With a finite horizon, perfect monitoring, and a unique stage-game Nash equilibrium, backward induction requires that equilibrium to be played after every history. Uniqueness here refers to the full stage-game equilibrium set, including mixed equilibria, rather than merely the set of pure equilibria.\n\nFinite repetition has richer possibilities when the stage game has several equilibria. Suppose one equilibrium gives a player a high payoff and another gives a low payoff. Earlier conduct can determine which equilibrium is selected later. Since both continuation outcomes are themselves equilibria, the reward and punishment can be credible even in the last round. Earlier actions that are not stage-game equilibria may then be sustainable because they influence the selection of subsequent equilibria.\n\nA known terminal date must also be distinguished from an uncertain one. A relationship with a constant independent probability of continuation has no publicly known final round from which backward induction can start. Likewise, uncertainty about an opponent’s objectives can change finite-horizon incentives. A player may cooperate to influence beliefs about its future conduct, despite knowing that the interaction eventually ends.\n\nThe finite-horizon result is therefore a conditional theorem, not a general claim about repeated human interaction. Its force depends jointly on the terminal structure, common knowledge of the game, equilibrium multiplicity, and the refinement being applied. Laboratory observations of early cooperation followed by increased defection near the end are relevant to these assumptions, but they do not invalidate the mathematical argument within its specified model."
    },
    {
      "heading": "Cooperation, retaliation, and incentive constraints",
      "text": "In an indefinitely repeated prisoner’s dilemma, future consequences can offset a current temptation to defect. Let the stage payoffs satisfy\n\n\\[T>R>P>S,\\]\n\nwhere \\(T\\) is the payoff from defecting against cooperation, \\(R\\) from mutual cooperation, \\(P\\) from mutual defection, and \\(S\\) from cooperating against defection. A common additional assumption, \\(2R>T+S\\), ensures that mutual cooperation produces a larger joint payoff than alternating between the two asymmetric outcomes.\n\nA grim-trigger strategy begins by cooperating and continues to cooperate as long as no defection has occurred. Following any defection, it prescribes defection forever. Against another grim-trigger player, continued cooperation yields normalized payoff \\(R\\). A unilateral defection yields\n\n\\[(1-\\delta)T+\\delta P,\\]\n\nbecause the deviator receives \\(T\\) immediately and then \\(P\\) forever. Cooperation is incentive compatible exactly when\n\n\\[R\\geq(1-\\delta)T+\\delta P,\\]\n\nor\n\n\\[\\delta\\geq\\frac{T-R}{T-P}.\\]\n\nFor \\(T=5\\), \\(R=3\\), \\(P=1\\), and \\(S=0\\), the threshold is \\(\\delta\\geq 1/2\\).\n\nThis calculation captures the basic mechanism of repeated-game cooperation. A higher temptation payoff makes cooperation harder to sustain, while a lower punishment payoff makes it easier. Greater patience also helps because it increases the weight of the continuation loss. After punishment begins, mutual defection is a stage-game equilibrium, making the prescribed punishment sequentially credible. Thus the argument establishes more than a favorable comparison along the cooperative path.\n\nOther strategies use shorter or more forgiving responses. A finite-punishment strategy may prescribe defection for a fixed number of rounds and then return to cooperation. Its incentive constraints must account for behavior both during cooperation and during punishment. Restarting or extending punishment after a fresh deviation can matter: an apparently straightforward punishment scheme may fail if a player gains by deviating while already being punished.\n\nTit for tat cooperates initially and then copies the opponent’s previous action. It became prominent through Robert Axelrod’s computer tournaments, but tournament success does not imply that it is a subgame-perfect equilibrium for every payoff specification and discount factor. Following an asymmetric outcome, two tit-for-tat players can alternate cooperation and defection indefinitely. Incentives at these off-path histories require separate examination.\n\nErrors further distinguish these strategies. Grim trigger can turn a single mistaken action into permanent breakdown, whereas forgiveness can restore cooperation. Forgiveness also reduces the expected cost of deliberate deviation. Repeated-game design consequently involves a trade-off between deterrence and resilience, rather than a universal ranking of harsh and lenient strategies."
    },
    {
      "heading": "Equilibrium payoffs and folk theorems",
      "text": "The folk theorem is a family of results characterizing the large set of payoffs that sufficiently patient players can sustain in repeated interaction. Its name reflects the circulation of versions of the result before definitive published treatments. There is no single theorem valid under every horizon, monitoring structure, or equilibrium concept. Each formulation specifies its own conditions on patience, feasibility, punishment, and information.\n\nFeasibility provides one restriction. If stage-game payoff vectors are \\(u(a)\\), the usual feasible set is\n\n\\[F=\\operatorname{co}\\{u(a):a\\in A\\},\\]\n\nwhere \\(\\operatorname{co}\\) denotes the convex hull. Convex combinations represent time-sharing or randomization among action profiles. Not every feasible payoff is an equilibrium payoff: a proposed allocation must also give each player enough to prevent profitable unilateral deviations.\n\nA second restriction is individual rationality. A commonly used benchmark is the player’s minmax payoff,\n\n\\[\\underline v_i =\\min_{\\alpha_{-i}}\\max_{a_i} \\mathbb{E}_{\\alpha_{-i}}[u_i(a_i,a_{-i})].\\]\n\nIt measures the best payoff player \\(i\\) can secure when the other players choose behavior to hold that payoff down. The admissible randomization and coordination among opponents must be specified; different conventions can affect the benchmark. An equilibrium cannot generally hold a player below what that player can guarantee by an appropriate response.\n\nA standard discounted, perfect-monitoring folk theorem says that, under suitable dimensionality conditions, every feasible payoff vector strictly above all players’ minmax levels can be supported as a subgame-perfect equilibrium when a common discount factor is sufficiently close to one. Some versions impose different qualifications or use different equilibrium concepts. Strict inequalities matter because boundary payoffs may leave no room to compensate players for carrying out punishments.\n\nThe central construction combines a target path with contingent punishments. A deviation triggers a continuation that reduces the deviator’s future payoff enough to offset its immediate gain. The difficult part is making every participant willing to carry out that continuation. Credible punishments may require rewards for punishers or further history-dependent phases; they need not consist simply of repeating a stage-game equilibrium.\n\nFolk theorems explain both the power and the limitations of repetition. Patience can support efficient cooperation, but it can also support unequal or inefficient allocations. The theory often predicts a set of possible equilibria rather than a unique outcome. Selecting among them requires further assumptions about institutions, communication, bargaining, learning, or strategic simplicity. Moreover, the discount-factor threshold depends on the target payoff and the game: “sufficiently patient” is not a universal numerical condition."
    },
    {
      "heading": "Monitoring, signals, and hidden actions",
      "text": "Perfect monitoring means that players observe the actions taken in previous periods. This assumption makes deviations identifiable and allows everyone to condition on the same action history. Many economic relationships instead involve hidden actions. A firm may observe market prices but not competitors’ production decisions; an employer may observe output but not a worker’s effort. Repeated-game analysis then depends on how actions generate observable signals.\n\nUnder imperfect public monitoring, all players observe the same signal, but it does not reveal actions exactly. A low market price, for example, might result either from a rival expanding output or from unexpectedly weak demand. A public strategy conditions on the common signal history. A perfect public equilibrium is an equilibrium in public strategies that remains an equilibrium after every public history.\n\nSuppose the public signal \\(y\\) has conditional distribution \\(p(y\\mid a)\\). If \\(w_i(y)\\) denotes player \\(i\\)'s continuation payoff after signal \\(y\\), then a pure prescribed action profile \\(a\\) must satisfy\n\n\\[(1-\\delta)u_i(a) +\\delta\\sum_y p(y\\mid a)w_i(y) \\geq (1-\\delta)u_i(a_i',a_{-i}) +\\delta\\sum_y p(y\\mid a_i',a_{-i})w_i(y)\\]\n\nfor every unilateral alternative \\(a_i'\\). Current incentives are therefore provided by the way a deviation changes the distribution of future rewards and punishments.\n\nThis formulation exposes the informational problem. A deviation that leaves the signal distribution unchanged cannot be deterred through signal-contingent continuation payoffs alone if it produces an immediate gain. When deviations do affect signals, punishment may sometimes follow unfavorable signals even though nobody actually deviated. The equilibrium must distinguish statistical evidence from certainty, and the cost of false alarms can reduce attainable payoffs.\n\nEdward Green and Robert Porter’s 1984 model of oligopoly illustrated how temporary price wars can arise under imperfect monitoring. A sufficiently unfavorable public market outcome can trigger a punishment phase even when all firms have followed the equilibrium strategy. Such episodes are not necessarily evidence that coordination has permanently collapsed; they can be part of the mechanism that sustains it.\n\nPrivate monitoring is more difficult because players receive different signals and may disagree about what happened. They may also be uncertain about what others believe. Public-history methods no longer transfer directly, and beliefs over private histories can become strategically important. Communication can help construct common information, but reports themselves must be incentive compatible. Monitoring quality thus affects not merely the precision of prediction but the structure of the equilibrium problem."
    },
    {
      "heading": "Incomplete information and reputation",
      "text": "Repetition with incomplete information introduces uncertainty about players’ payoff functions, available actions, or behavioral characteristics. A player may be an ordinary payoff maximizer, a type with an unusually strong preference for cooperation, or a commitment type constrained to follow a particular rule. Such interactions combine repeated play with the informational structure of a Bayesian game. Observed behavior affects both immediate payoffs and beliefs about the opponent.\n\nBeliefs are updated using Bayes’ theorem whenever the observed event has positive probability under the model. If a player’s initial probability of type \\(\\theta\\) is \\(\\mu(\\theta)\\), an observed action changes that assessment according to the likelihood of the action under each possible type. A strategic player anticipates this process. It may therefore choose an action with a relatively low current payoff because that action changes how opponents respond later.\n\nDavid Kreps, Paul Milgrom, John Roberts, and Robert Wilson’s 1982 analysis of the finitely repeated prisoner’s dilemma showed how incomplete information can support substantial cooperation in a finite relationship. A small probability that an opponent follows a cooperative contingent rule can give strategic players an incentive to preserve a cooperative reputation. This differs from the complete-information backward-induction argument because behavior conveys information about types.\n\nReputation is particularly important when a long-lived player meets a sequence of short-lived opponents. Each short-lived opponent may participate only once, but may observe the long-lived player’s past actions. The long-lived player can then have an incentive to imitate a commitment type whose apparent behavior induces favorable responses. Drew Fudenberg and David Levine’s 1989 reputation analysis established important conditions under which a patient long-run player can obtain a commitment-related payoff guarantee.\n\nInformation revelation can also be costly. In repeated zero-sum games with incomplete information, an informed player may conceal information to preserve future advantages or reveal it selectively to improve present decisions. Robert Aumann and Michael Maschler developed a major theory of this trade-off. Unlike mutually beneficial cooperation, these models concern strategic control of information in environments with directly opposed interests.\n\nReputation effects depend on observability, prior beliefs, the available commitment types, and the behavior of opponents. They do not imply that a favorable reputation can always be acquired or maintained. A finite-horizon reputation model can still display end effects, while imperfect observation can weaken the connection between conduct and inferred type. The mechanism is specifically the influence of current actions on future beliefs and responses, not simply the existence of a long relationship."
    },
    {
      "heading": "Recursive methods, computation, and learning",
      "text": "A repeated-game strategy may depend on an arbitrarily long history, making direct enumeration impractical. Recursive methods instead describe current play together with the continuation payoffs promised after each observable outcome. Under perfect monitoring and pure current actions, a candidate payoff vector \\(v\\) can be decomposed as\n\n\\[v_i=(1-\\delta)u_i(a)+\\delta w_i(a).\\]\n\nFor each alternative action \\(a_i'\\), incentive compatibility requires\n\n\\[v_i\\geq (1-\\delta)u_i(a_i',a_{-i}) +\\delta w_i(a_i',a_{-i}).\\]\n\nEvery continuation payoff must itself be supportable by an equilibrium.\n\nThis approach converts an infinite tree of histories into a problem about sets of attainable continuation values. A set is self-generating when its members can be decomposed into current payoffs and incentive-compatible continuations lying in the same set. Dilip Abreu, David Pearce, and Ennio Stacchetti developed influential recursive methods for repeated games, particularly in their 1990 work on discounted games with imperfect monitoring.\n\nThe recursive structure resembles a Bellman equation, but the object is often a set of equilibrium payoff vectors rather than a single agent’s optimal value function. Continuation values must satisfy several players’ incentives simultaneously. Numerical approaches may approximate payoff sets, solve supporting optimization problems, or iterate operators on candidate sets. Under suitable finite representations, some of these calculations reduce to linear programming, although general equilibrium computation need not be simple.\n\nAnother representation uses finite automata. A strategy has a finite collection of internal states, an action rule for each state, and a transition rule based on observed play. Grim trigger needs only a cooperative state and a punishment state. Tit for tat requires memory of the opponent’s latest action. Restricting the number of states gives a concrete model of limited strategic complexity, but can exclude equilibria requiring long or elaborate histories.\n\nLearning addresses a different question: how behavior develops when players do not initially know or compute an equilibrium. Procedures include best-response learning, fictitious play, regret minimization, and reinforcement learning. They use different information and generate different convergence properties. A learning rule may approach a correlated set of outcomes or achieve low regret without converging to a subgame-perfect equilibrium.\n\nThis distinction matters in multiagent machine learning. An algorithm that performs well against fixed opponents may behave poorly when all opponents adapt. Likewise, cooperation between trained agents may be fragile under changes in observation, discounting, or opponent behavior. Computational success, equilibrium verification, and empirical stability are separate properties and require separate tests."
    },
    {
      "heading": "Economic and institutional applications",
      "text": "Repeated games explain how behavior can be disciplined without a fully enforceable contract. If an opportunistic action is observable and jeopardizes sufficiently valuable future benefits, parties may refrain from it even when no court can impose the relevant penalty. This mechanism appears in supply relationships, professional services, employment, and repeated exchange. The essential enforcement resource is the loss of future relationship surplus.\n\nIn a relational employment arrangement, an employer may promise a discretionary bonus for good performance, while a worker supplies effort that is difficult to specify contractually. The employer’s promise is credible only if withholding the bonus would damage a sufficiently valuable future relationship. The worker’s effort must likewise be worthwhile given expected compensation and the consequences of poor performance. Both sides’ incentive constraints matter; patience alone does not ensure that every promised arrangement is self-enforcing.\n\nOligopoly provides a contrasting application because sustained cooperation among firms can harm buyers. A firm may profit immediately by undercutting a high-price arrangement, but lose future profits if competitors respond aggressively. In a simple model with cooperation profit \\(\\pi_C\\), deviation profit \\(\\pi_D\\), and perpetual punishment profit \\(\\pi_P\\), the incentive condition is\n\n\\[\\frac{\\pi_C}{1-\\delta} \\geq \\pi_D+\\frac{\\delta\\pi_P}{1-\\delta}.\\]\n\nThe same intertemporal logic that supports productive cooperation can therefore support restrictive market conduct. The equation is an analytical benchmark, not evidence that any particular market is coordinated.\n\nRepeated interaction also helps analyze agreements between governments when external enforcement is limited. Trade concessions may be supported by the prospect of reciprocal benefits and responses to noncompliance. Yet actual institutions involve changing membership, asymmetric power, domestic constraints, and uncertainty about policy effects. A repeated-game model isolates incentive mechanisms; it does not by itself establish the causes of a historical agreement or dispute.\n\nParticipation conditions are crucial in all these settings. If a player can leave the relationship, its outside option constrains promised continuation payoffs. Easy replacement of partners may weaken bilateral punishment, while transferable reputations can extend discipline beyond a single pair. Networks and public records can support community enforcement, but they also create questions about inaccurate reports and incentives to communicate honestly.\n\nEmpirical work must distinguish repeated-game mechanisms from alternatives such as enforceable contracts, social preferences, or technological complementarities. The same cooperative outcome can arise for different reasons. Evidence about monitoring, expected duration, responses to deviations, and partner replacement is therefore more informative than the mere observation that cooperation occurs repeatedly."
    },
    {
      "heading": "Experimental evidence and behavioral complications",
      "text": "Experiments on repeated games examine how actual choices respond to the horizon, payoff structure, information, and opportunities for retaliation. Early experimental work, including the repeated prisoner’s-dilemma experiment conducted by Merrill Flood and Melvin Dresher in 1950, revealed behavior more varied than simple one-shot equilibrium reasoning suggested. Subsequent laboratory research separated fixed-length interactions from indefinitely repeated interactions implemented through random continuation.\n\nA robust distinction in this literature is between cooperation being an equilibrium possibility and cooperation being reliably selected. Increasing the continuation probability can make cooperation incentive compatible, but participants must still coordinate on compatible strategies. They may be unsure whether an opponent will cooperate, whether an apparent defection was accidental, or whether punishment will be temporary. Strategic uncertainty can therefore inhibit cooperation even when mutual cooperation is sustainable.\n\nPedro Dal Bó’s 2005 experimental work systematically examined cooperation under different continuation probabilities. Dal Bó and Guillaume Fréchette’s later research investigated how cooperation develops with experience in indefinitely repeated games. This research treats learning and equilibrium selection as substantive questions rather than assuming that participants immediately identify and adopt the most efficient equilibrium.\n\nComputer tournaments provide a different kind of evidence. Axelrod’s tournaments, reported in 1980 and discussed in his 1984 book The Evolution of Cooperation, compared submitted strategies in repeated prisoner’s-dilemma encounters. Tit for tat, submitted by Anatol Rapoport, performed prominently. Its simplicity, responsiveness, and willingness to resume cooperation helped stimulate research across the social and biological sciences. Tournament rankings, however, depend on the pool of entrants, scoring rule, horizon, and error structure.\n\nBehavioral motives complicate interpretation. Participants may care about fairness, reciprocity, or the opponent’s payoff, rather than only their own monetary earnings. If these preferences are stable, they can be incorporated into a game’s utility functions. If behavior changes through experience or framing, a fixed-payoff model may be inadequate. Observed cooperation is therefore not automatically evidence for either self-interested repeated-game incentives or non-self-interested preferences alone.\n\nExperiments also connect repeated games with evolutionary game theory, where the frequencies of strategies change through reproduction or imitation rather than deliberate equilibrium calculation. A strategy can perform well in an evolving population without satisfying every sequential incentive condition of a rational-player model. Evolutionary stability, tournament performance, and subgame perfection answer different questions about strategic behavior."
    },
    {
      "heading": "Historical development and continuing research",
      "text": "The foundations of repeated-game theory emerged from twentieth-century work on strategic interaction, including John von Neumann and Oskar Morgenstern’s Theory of Games and Economic Behavior in 1944 and John Nash’s equilibrium results of 1950–1951. Nash equilibrium supplied a general solution concept, but applying it to repeated interaction exposed the importance of contingent behavior and threats. Robert Aumann’s 1959 work on infinitely repeated games was an early landmark in the systematic study of long-run strategic outcomes.\n\nReinhard Selten introduced subgame perfection in 1965, sharpening the distinction between a threatened punishment and a punishment that players would actually carry out. James Friedman’s 1971 analysis of noncooperative equilibria in supergames helped formalize how repeated interaction can sustain outcomes beyond stage-game equilibria. These developments made credibility a central requirement rather than treating future retaliation as an automatic consequence of deviation.\n\nThe 1980s brought major advances in both information and equilibrium payoff characterization. Green and Porter analyzed imperfect monitoring in 1984. Fudenberg and Eric Maskin’s 1986 folk-theorem work established influential conditions for sustaining feasible, individually rational payoffs in discounted repeated games. Research on reputation demonstrated how small uncertainties about player types could alter predictions, including in finitely repeated settings.\n\nLater work extended these approaches to richer monitoring structures. Abreu, Pearce, and Stacchetti developed recursive equilibrium analysis, while Fudenberg, Levine, and Maskin’s 1994 work established influential folk-theorem results under imperfect public information. Private monitoring proved especially demanding because the history relevant to incentives is not shared. Researchers have studied communication, statistical distinguishability of deviations, and conditions under which private observations still permit efficient outcomes.\n\nSeveral continuing research questions concern robustness. An equilibrium may rely on exact knowledge of payoffs, highly accurate observations, or a punishment system that all participants would prefer to abandon once reached. Renegotiation therefore creates a tension: a continuation can be credible against unilateral deviation yet vulnerable to a mutually agreed replacement. Different concepts of renegotiation-proofness impose different restrictions, so there is no single automatic refinement.\n\nOther work examines restricted memory, anonymous matching, network interaction, changing populations, and learning algorithms. These variations alter who can identify a deviation, who can respond, and how future opportunities depend on current conduct. They require explicit models of information transmission and strategic capacity, extending the original question from whether repetition can support cooperation to which forms of intertemporal discipline remain feasible under realistic informational and computational constraints."
    }
  ],
  "links": [
    "game-theory",
    "probability-distribution",
    "nash-equilibrium",
    "backward-induction",
    "prisoners-dilemma",
    "subgame-perfect-equilibrium",
    "convex-hull",
    "bayesian-game",
    "bayes-theorem",
    "zero-sum-game",
    "bellman-equation",
    "value-function",
    "linear-programming",
    "reinforcement-learning",
    "machine-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}