{
  "id": "subgame-perfect-equilibrium",
  "title": "Subgame-perfect Equilibrium",
  "lang": "en",
  "source_lang": "en",
  "summary": "A subgame-perfect equilibrium is a strategy profile that forms a Nash equilibrium in every subgame, excluding threats that would not be optimal if their decision points were reached.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Subgame-perfect Nash equilibrium",
    "subgame perfect equilibrium",
    "subgame perfection",
    "SPE",
    "SPNE"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/subgame-perfect-equilibrium/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/subgame-perfect-equilibrium/",
      "json": "https://aiwiki.page/zh/subgame-perfect-equilibrium.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A subgame-perfect equilibrium is a solution concept in game theory for games with sequential decisions. It is a strategy profile whose restriction to every subgame is a Nash equilibrium of that subgame. Consequently, prescribed behavior must remain mutually optimal not only along the predicted course of play but also in subgames that the equilibrium strategies never reach. This requirement rules out many equilibria supported by noncredible threats: actions that a player threatens to take but would not actually find optimal when the relevant decision arises."
    },
    {
      "heading": "Strategies, information, and subgames",
      "text": "Subgame perfection is defined for an extensive-form game, which represents the order of decisions, the available actions, what players know when choosing, and the payoffs associated with outcomes. A strategy is a complete contingent plan: it specifies what a player would do at every decision point or information set assigned to that player, including situations that the strategy itself makes unreachable.\n\nAn information set groups decision nodes that a player cannot distinguish when choosing an action. A subgame is a part of the original game that can be treated as a self-contained game. In a standard extensive-form representation, it must:\n\n- Begin at a node that is the sole member of its information set.\n- Include all successors of that node.\n- Contain either all or none of the nodes of each information set it intersects.\n\nThe final condition prevents a proposed subgame from cutting across an information set and thereby giving a player information that the original game does not provide.\n\nThe whole game counts as a subgame. A subgame beginning strictly after the initial node is called a proper subgame.\n\nIn a game of perfect information, every information set contains one node, so every decision node can begin a subgame. In games with imperfect information, many decision nodes cannot do so. This difference is central to both the usefulness and the limitations of subgame perfection."
    },
    {
      "heading": "Formal definition",
      "text": "Let \\(G\\) be an extensive-form game with players \\(i=1,\\ldots,n\\), and let\n\n\\[\\sigma=(\\sigma_1,\\ldots,\\sigma_n)\\]\n\nbe a strategy profile. For any subgame \\(H\\), write \\(\\sigma|_H\\) for the restriction of that profile to decisions within \\(H\\).\n\nThe profile \\(\\sigma\\) is a subgame-perfect equilibrium if, for every subgame \\(H\\), every player \\(i\\), and every admissible alternative continuation strategy \\(\\tau_i\\),\n\n\\[U_i^H(\\sigma_i|_H,\\sigma_{-i}|_H) \\geq U_i^H(\\tau_i,\\sigma_{-i}|_H).\\]\n\nHere, \\(U_i^H\\) is player \\(i\\)'s payoff in the subgame, using expected utility when play involves randomization or chance, and \\(\\sigma_{-i}\\) denotes the strategies of all players other than \\(i\\).\n\nThe inequality is evaluated by treating \\(H\\) as a game in its own right. It therefore applies even when the probability of reaching \\(H\\) under \\(\\sigma\\) is zero.\n\nBecause the whole game is a subgame, every subgame-perfect equilibrium is a Nash equilibrium. The converse is generally false. A Nash equilibrium tests whether a unilateral change of strategy is profitable from the beginning of the game; subgame perfection additionally tests incentives within every subgame."
    },
    {
      "heading": "Noncredible threats: an entry example",
      "text": "Consider a market-entry game involving an entrant and an incumbent.\n\n- The entrant chooses Out or Enter.\n- If the entrant stays out, payoffs are \\((0,2)\\).\n- If entry occurs, the incumbent chooses Accommodate or Fight.\n- Accommodation yields \\((1,1)\\), while fighting yields \\((-1,-1)\\).\n\nPayoffs are listed as entrant first, incumbent second.\n\nThe strategy profile Out, Fight is a Nash equilibrium. Given the incumbent's specified response, entry would lower the entrant's payoff from \\(0\\) to \\(-1\\). Given that the entrant stays out, changing the incumbent's response has no effect on the outcome or its payoff.\n\nHowever, this equilibrium is not subgame-perfect. In the subgame after entry, the incumbent receives \\(1\\) from accommodation and \\(-1\\) from fighting. Its threatened response is therefore not optimal when the decision actually arises.\n\nThe unique subgame-perfect equilibrium is Enter, Accommodate, yielding \\((1,1)\\).\n\nThe example illustrates why a strategy must specify off-path behavior. Although fighting does not occur in the first proposed equilibrium, its prescription is what supposedly deters entry. Subgame perfection tests whether that prescription is credible."
    },
    {
      "heading": "Backward induction",
      "text": "In finite games of perfect information, subgame-perfect equilibria can be found by backward induction:\n\n- At each final decision node, select an action maximizing the deciding player's payoff.\n- Replace that decision problem with its resulting continuation payoff.\n- Repeat the process at earlier decision nodes until reaching the initial node.\n\nThe resulting strategy profile specifies optimal behavior in every subgame, not merely an optimal sequence of actions along one path.\n\nEvery finite extensive-form game of perfect information has a subgame-perfect equilibrium in pure strategies. Indifference can produce multiple equilibria: different optimal choices at later nodes may also change incentives earlier in the game. Thus, backward induction is a method of finding equilibria, not a general guarantee of uniqueness.\n\nFor finite games with imperfect information, solving a subgame may require finding a Nash equilibrium of a simultaneous or otherwise strategically interdependent continuation game. Mixed strategies may be necessary. Every finite extensive-form game with perfect recall has a subgame-perfect equilibrium when randomization is allowed.\n\nBackward induction should also be distinguished from identifying an equilibrium outcome. Different strategy profiles can generate the same realized outcome while prescribing different actions elsewhere. Subgame perfection is a property of the complete profile."
    },
    {
      "heading": "The one-shot deviation principle",
      "text": "A useful method for verifying subgame perfection is the one-shot deviation principle. Instead of comparing a candidate strategy with every possible alternative continuation plan, it can be sufficient to check deviations at a single decision point, followed by a return to the prescribed strategy.\n\nThis principle holds for finite-horizon games with perfect information. It also holds in standard infinitely repeated games with observed actions, bounded stage payoffs, and geometric discounting. More general formulations require suitable assumptions, such as perfect recall and continuity at infinity: differences sufficiently far in the future must have an arbitrarily small effect on current payoffs.\n\nThe principle is particularly valuable when the set of complete strategies is enormous or infinite. Its applicability is not automatic for every infinite game; the information structure and payoff assumptions matter."
    },
    {
      "heading": "Repeated games",
      "text": "In a repeated game, players interact through successive plays of a stage game. When completed stages are publicly observed, a continuation after any publicly observed history defines a subgame. Subgame perfection therefore requires prescribed rewards and punishments to be credible after every such history.\n\nFinite repetition\n\nIf a stage game has a unique Nash equilibrium and is repeated a known finite number of times, with observed actions and payoffs given by an additive aggregation of stage payoffs, backward induction implies that this equilibrium is played after every history in every subgame-perfect equilibrium.\n\nThe final stage must use the unique stage-game equilibrium. Since its continuation payoff cannot vary across earlier histories, the same reasoning applies to the preceding stage and then recursively to all earlier stages.\n\nIf the stage game has multiple Nash equilibria, history-dependent selection among continuation equilibria can create additional incentives. The unique-equilibrium conclusion therefore cannot simply be extended to that case.\n\nInfinite repetition\n\nAn indefinite horizon can support behavior that would not be an equilibrium of the stage game.\n\nConsider an infinitely repeated prisoner's dilemma with stage payoffs satisfying\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.\n\nUnder a grim-trigger strategy, each player cooperates until a defection occurs and defects forever afterward. Let \\(0<\\delta<1\\) be the common discount factor. Continuing cooperation gives\n\n\\[\\frac{R}{1-\\delta},\\]\n\nwhereas defecting immediately against a cooperating opponent and then receiving the punishment payoff gives\n\n\\[T+\\frac{\\delta P}{1-\\delta}.\\]\n\nCooperation is incentive-compatible when\n\n\\[\\frac{R}{1-\\delta} \\geq T+\\frac{\\delta P}{1-\\delta},\\]\n\nor equivalently,\n\n\\[\\delta\\geq\\frac{T-R}{T-P}.\\]\n\nThe punishment is itself credible because mutual defection is a stage-game Nash equilibrium. Under these assumptions, the condition supports a subgame-perfect equilibrium. A severe punishment alone is insufficient: players must also be willing to carry it out once the punishment phase begins."
    },
    {
      "heading": "Imperfect information and stronger refinements",
      "text": "Subgame perfection is less restrictive when a game has few proper subgames. If the whole game is its only subgame, subgame-perfect equilibrium coincides with Nash equilibrium. In such a game, the concept may leave implausible behavior at unreached information sets unrestricted.\n\nStronger equilibrium concepts explicitly address choices and beliefs at information sets:\n\n- Perfect Bayesian equilibrium combines strategies with beliefs and imposes sequential rationality and belief updating. Its precise requirements vary among formulations.\n- Sequential equilibrium requires sequential rationality together with a consistency condition linking beliefs to limits of completely mixed strategy profiles. In finite games with perfect recall, every sequential equilibrium is subgame-perfect, but not every subgame-perfect equilibrium is sequential.\n- Trembling-hand perfect equilibrium examines robustness to small probabilities of mistakes. It is a distinct refinement, and its application to normal-form or extensive-form representations must be specified.\n\nSubgame perfection by itself does not provide a full account of beliefs after unexpected actions. It imposes equilibrium behavior wherever a genuine subgame begins, rather than at every information set individually."
    },
    {
      "heading": "Origins, applications, and limitations",
      "text": "Reinhard Selten introduced subgame perfection in 1965 as a refinement of Nash equilibrium for dynamic games. It became a central tool for analyzing credibility in economics, including market entry, bargaining, sequential competition, contracting, and repeated interaction.\n\nThe concept clarifies the difference between an announced future action and a credible continuation strategy. A threat may fail to be credible because carrying it out would harm the player making it. Changing the game's structure—for example, by introducing an enforceable commitment—can change which actions are optimal and therefore which equilibria are subgame-perfect.\n\nSeveral limitations affect its interpretation:\n\n- Multiplicity: the requirement may leave many equilibria and does not necessarily select a unique prediction.\n- Dependence on information: what players observe determines which continuations qualify as subgames.\n- Dependence on the model: commitment opportunities, available actions, timing, and payoffs can substantially change the equilibrium set.\n- Individual rather than collective stability: it rules out profitable unilateral deviations within subgames, not coordinated deviations by groups.\n- No efficiency guarantee: a subgame-perfect equilibrium need not maximize total payoff or benefit every player.\n- Computational difficulty: backward induction is straightforward for small trees, but large game trees and difficult continuation games can make equilibrium computation demanding.\n\nSubgame perfection is a consistency requirement for strategic plans under specified assumptions about rational choice and the game structure. It does not assert that observed participants invariably follow those plans."
    }
  ],
  "links": [
    "game-theory",
    "nash-equilibrium",
    "extensive-form-game",
    "perfect-information",
    "expected-value",
    "backward-induction",
    "mixed-strategy",
    "repeated-game",
    "prisoners-dilemma",
    "economics"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}