{
  "id": "bayesian-game",
  "title": "Bayesian Game",
  "lang": "en",
  "source_lang": "en",
  "summary": "A Bayesian game models strategic interaction in which players have incomplete information and choose actions using beliefs about privately known characteristics.",
  "category": "mathematics",
  "subcategory": "applied",
  "type": "concept",
  "aliases": [
    "Game of incomplete information",
    "Incomplete-information game"
  ],
  "generated_at": "2026-11-06",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/bayesian-game/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/bayesian-game/",
      "json": "https://aiwiki.page/zh/bayesian-game.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A Bayesian game is a model in game theory in which players lack complete information about strategically relevant features of the situation, such as other players’ preferences, costs, valuations, or available resources. Private information is represented by a player’s type, while beliefs about unknown types are represented by probability distributions. A strategy specifies what a player would do for each possible type, and equilibrium requires each player’s choices to maximize expected payoff given their information and the other players’ strategies."
    },
    {
      "heading": "Incomplete information and types",
      "text": "In a complete-information game, the structure of the game—including the players’ payoff functions—is common knowledge. Players may nevertheless be uncertain about which actions others will choose. A Bayesian game adds uncertainty about characteristics that help determine those choices or their consequences.\n\nExamples include:\n\n- A bidder knows its own valuation of an object but not its competitors’ valuations.\n- A firm knows its production cost but not a rival’s cost.\n- A seller has information about product quality that a buyer does not possess.\n- A negotiator knows its own willingness to accept an agreement but not the other party’s.\n\nA type summarizes a player’s private information relevant to the model. It may include a valuation, a cost, a private signal, or information about another player’s information. A type need not describe a psychological personality or a fixed behavioral disposition.\n\nInformation can also concern an external state rather than a player’s own characteristics. For example, several firms might receive different private signals about future demand. Such situations can be represented by adding a state variable and specifying how signals and payoffs depend on that state.\n\nIncomplete versus imperfect information\n\nIncomplete information and imperfect information describe different features of a game.\n\nIncomplete information concerns uncertainty about the game’s payoff-relevant characteristics. Imperfect information concerns a player’s inability to observe some earlier moves when choosing an action.\n\nA simultaneous-move game can have complete information even though players do not observe one another’s current choices. Conversely, players may observe all previous actions yet remain uncertain about a participant’s privately known cost. Dynamic Bayesian games often involve both kinds of uncertainty."
    },
    {
      "heading": "Formal model",
      "text": "A standard finite Bayesian game can be written as\n\n\\[G=\\left(N,(A_i)_{i\\in N},(T_i)_{i\\in N},p,(u_i)_{i\\in N}\\right),\\]\n\nwhere:\n\n- \\(N=\\{1,\\ldots,n\\}\\) is the set of players.\n- \\(A_i\\) is player \\(i\\)’s action set.\n- \\(T_i\\) is player \\(i\\)’s type set.\n- \\(p\\) is a joint probability distribution on the type profiles \\(T=\\prod_{i\\in N}T_i\\).\n- \\(u_i:A\\times T\\to\\mathbb{R}\\) is player \\(i\\)’s payoff function, where \\(A=\\prod_{i\\in N}A_i\\).\n\nThe notation \\(t_{-i}\\) denotes the types of all players except \\(i\\), and \\(a_{-i}\\) similarly denotes their actions.\n\nIn the standard interpretation, a type profile \\(t\\) is drawn according to \\(p\\). Each player \\(i\\) observes their own type \\(t_i\\), but not necessarily anyone else’s. The players then choose actions.\n\nThe distribution \\(p\\) is often called the common prior. A shared prior does not imply that players have identical information or identical beliefs after observing their types. Instead, their beliefs are obtained by conditioning the same prior on different private observations. For a type with positive marginal probability,\n\n\\[p(t_{-i}\\mid t_i) = \\frac{p(t_i,t_{-i})} {\\sum_{\\hat t_{-i}}p(t_i,\\hat t_{-i})}.\\]\n\nThis is an application of conditional probability. Types may be correlated: learning one’s own type can provide information about others’ types. Assuming independence is an additional modeling restriction, not part of the definition of a Bayesian game.\n\nMore general models allow player-specific beliefs rather than a common prior, or use continuous type and action spaces. With continuous types, beliefs are expressed through conditional distributions, and sums are replaced by integrals where appropriate."
    },
    {
      "heading": "Strategies and expected payoffs",
      "text": "A pure strategy is a type-contingent action rule,\n\n\\[s_i:T_i\\to A_i.\\]\n\nThus, a strategy specifies an action for every possible type, not merely for the type that happens to be realized.\n\nA mixed strategy can be represented as a probability distribution over pure type-contingent strategies. For a static Bayesian game, randomized behavior is also conveniently described by a rule\n\n\\[\\sigma_i:T_i\\to\\Delta(A_i),\\]\n\nwhere \\(\\Delta(A_i)\\) is the set of probability distributions over actions. This specifies an action distribution for each type.\n\nFor a pure strategy profile \\(s\\), the expected payoff to type \\(t_i\\) from choosing action \\(a_i\\) is\n\n\\[U_i(a_i,s_{-i}\\mid t_i) = \\sum_{t_{-i}} p(t_{-i}\\mid t_i) u_i\\!\\left(a_i,s_{-i}(t_{-i});t_i,t_{-i}\\right).\\]\n\nThis expected value accounts for uncertainty about other players’ types and, through their strategies, the actions those types will select. With randomized strategies, the expectation also includes the players’ randomization.\n\nThree perspectives are useful:\n\n- Ex ante: before a player learns their type.\n- Interim: after learning their own type but before learning others’ private information.\n- Ex post: conditional on the realized type profile.\n\nBayesian equilibrium is ordinarily formulated using interim incentives."
    },
    {
      "heading": "Bayesian Nash equilibrium",
      "text": "A Bayesian Nash equilibrium is a strategy profile in which each player’s strategy is optimal for every own type with positive probability, given the other players’ strategies and the player’s conditional beliefs.\n\nFor pure strategies, \\(s^*\\) is an equilibrium if\n\n\\[U_i(s_i^*(t_i),s_{-i}^*\\mid t_i) \\geq U_i(a_i,s_{-i}^*\\mid t_i)\\]\n\nfor every player \\(i\\), every relevant type \\(t_i\\), and every alternative action \\(a_i\\in A_i\\).\n\nThe condition combines two ideas:\n\n- Each type chooses an action that maximizes its conditional expected payoff.\n- Beliefs about other players’ actions are consistent with those players’ equilibrium strategies.\n\nA player need not choose the action that would be best if all private information were revealed. The action must instead be best given what that player actually knows.\n\nBayesian Nash equilibrium extends Nash equilibrium to incomplete-information settings. If all type sets are singletons, the definition reduces to ordinary Nash equilibrium.\n\nExistence and limitations\n\nEvery finite Bayesian game has an equilibrium in randomized strategies. One way to establish this is to construct a finite normal-form game whose pure strategies are the type-contingent plans and apply the mixed-strategy existence theorem for finite games.\n\nPure-strategy equilibrium need not exist. For infinite action or type spaces, existence requires additional assumptions, such as appropriate continuity, compactness, and measurability conditions.\n\nEquilibrium also need not be unique. Different equilibria can produce different actions, allocations, or payoffs under the same information structure. The existence of an equilibrium therefore does not by itself yield a unique prediction."
    },
    {
      "heading": "Harsanyi’s transformation",
      "text": "The modern framework was developed by John C. Harsanyi in work published in 1967–1968. Its central device, commonly called the Harsanyi transformation, represents uncertainty about player characteristics as an initial move by a fictitious player, “Nature.”\n\nThe transformed game proceeds as follows:\n\n- Nature selects a type profile according to the specified distribution.\n- Each player receives the information associated with their type.\n- Players choose actions, simultaneously or sequentially.\n- Payoffs depend on the actions and realized types.\n\nThis converts a game of incomplete information into a game with an explicit chance move and a fully specified structure. The resulting game generally has imperfect information because players do not observe Nature’s entire choice.\n\nNature need not represent literal random assignment. The construction can model uncertainty about an already existing fact, such as a firm’s production cost.\n\nAn important role of types is to organize beliefs about beliefs: a player may be uncertain about what another player knows, what that player believes, and so on. Simple applied models usually specify a common prior and signal structure rather than explicitly listing these higher-order beliefs."
    },
    {
      "heading": "Example: a first-price sealed-bid auction",
      "text": "Consider two risk-neutral bidders competing for one object. Each bidder privately observes a valuation \\(v_i\\), independently drawn from the uniform distribution on \\([0,1]\\). Both distributions and the auction rules are common knowledge.\n\nEach bidder submits a nonnegative bid. The higher bidder wins and pays their own bid; the loser pays nothing. A winner with valuation \\(v_i\\) and bid \\(b_i\\) receives payoff \\(v_i-b_i\\).\n\nA symmetric Bayesian Nash equilibrium is\n\n\\[b(v)=\\frac{v}{2}.\\]\n\nTo see why, suppose the opponent follows this rule. For a bid \\(b\\in[0,1/2]\\), the bidder wins when the opponent’s valuation is below \\(2b\\), an event with probability \\(2b\\). Expected payoff is therefore\n\n\\[(v-b)\\,2b.\\]\n\nDifferentiating with respect to \\(b\\) gives\n\n\\[2v-4b,\\]\n\nwhich is zero at \\(b=v/2\\). The objective is concave, and bids above \\(1/2\\) cannot improve on the best bid in the stated interval. Ties occur with probability zero under the continuous valuation distribution.\n\nThis example illustrates why a strategy is a function of type: each possible valuation determines a different bid. It also shows that equilibrium behavior need not reveal private information truthfully through the action itself.\n\nWith \\(n\\geq2\\) bidders under the same independent uniform-value and risk-neutrality assumptions, the symmetric equilibrium becomes\n\n\\[b(v)=\\frac{n-1}{n}v.\\]\n\nThese formulas depend on the assumptions. Correlated valuations, risk aversion, reserve prices, or uncertainty about a common underlying value can change the equilibrium."
    },
    {
      "heading": "Dynamic Bayesian games and belief updating",
      "text": "In a dynamic Bayesian game, players may observe actions that reveal information about private types. Beliefs then evolve through Bayes’ theorem.\n\nFor example, a worker’s educational choice may convey information to an employer about privately known productivity. Such interactions are studied in signaling games. Different types may choose different observable actions, producing a separating equilibrium, or choose the same action, producing a pooling equilibrium.\n\nBayesian Nash equilibrium alone can permit behavior that is not optimal after some histories, particularly histories that occur with zero probability under equilibrium play. Dynamic analysis therefore often uses refinements such as perfect Bayesian equilibrium and sequential equilibrium.\n\nThese concepts combine contingent strategies with beliefs at decision points and impose requirements of sequential optimality and belief consistency. They are not interchangeable: definitions of perfect Bayesian equilibrium vary, whereas sequential equilibrium imposes a specific consistency requirement using limits of completely mixed strategies."
    },
    {
      "heading": "Applications",
      "text": "Bayesian games provide a framework for analyzing strategic behavior when information is dispersed.\n\nAuctions and procurement. Bidders have private valuations or costs. Models distinguish private-value settings from settings where bidders’ information concerns a common underlying value.\n\nContracts and regulation. A principal may design terms for an agent whose costs or characteristics are privately known. These models study how information constraints affect feasible contracts.\n\nIndustrial organization. Firms may compete while holding private information about costs, capacity, demand, or intended investment.\n\nBargaining. Negotiators may have private reservation values or information about the benefits of agreement. Offers and delays can themselves convey information.\n\nMechanism design. Mechanism design reverses the usual question: rather than predicting behavior under fixed rules, it studies rules intended to implement particular outcomes. Bayesian incentive compatibility requires truthful reporting to maximize each type’s expected payoff when other participants report truthfully.\n\nComputational strategic interaction. Bayesian-game models are used in security allocation and multiagent systems where participants possess private information about objectives, resources, or capabilities."
    },
    {
      "heading": "Modeling assumptions and computational challenges",
      "text": "The informational structure is central to a Bayesian game, not a secondary detail. Changing what players observe, how their types are correlated, or what they believe about one another can change equilibrium outcomes even when the physical actions remain unchanged.\n\nThe common-prior assumption provides a tractable way to connect private information and beliefs. It excludes some forms of prior disagreement, however, and may be difficult to justify in particular applications. Likewise, representing uncertainty by a known probability distribution can be restrictive when relevant probabilities are themselves poorly understood.\n\nComputation can be demanding. If player \\(i\\) has \\(|T_i|\\) types and \\(|A_i|\\) actions available at each type, the number of pure type-contingent strategies is\n\n\\[|A_i|^{|T_i|}.\\]\n\nConsequently, directly constructing the normal-form game can be impractical even for moderately sized models. Computational approaches may instead use type-based formulations, exploit special structure, discretize continuous spaces, or seek approximate equilibria.\n\nFinally, equilibrium predictions depend on assumptions about payoff maximization, beliefs, and strategic consistency. A Bayesian game specifies a disciplined model of decisions under private information; it does not establish that actual participants possess the assumed beliefs or select a particular equilibrium."
    }
  ],
  "links": [
    "game-theory",
    "joint-probability-distribution",
    "conditional-probability",
    "statistical-independence",
    "mixed-strategy",
    "expected-value",
    "nash-equilibrium",
    "uniform-distribution",
    "bayes-theorem",
    "mechanism-design"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}