Perfect Information¶
In an extensive-form game, every decision maker can identify the exact history at each turn, so every decision information set is a singleton.
Core Idea¶
A game of perfect information is an extensive-form game in which a player about to move knows which decision node has been reached. Formally, every decision information set is a singleton. Thus the mover can distinguish every relevant prior path through the tree, including any earlier chance outcome that could change the current node. This is a condition on observation of history, not a claim that a player is omniscient, predicts an opponent's future move, or knows an unknown parameter outside the modeled game.[1]
For a finite perfect-information game with specified payoffs and sequential decisions, choosing optimal actions from terminal nodes backward yields at least one pure-strategy subgame-perfect equilibrium. This theorem is conditional on the modeled tree and incentives; a perfect-information label does not make a huge game computationally easy or certify that real agents reason backward. The formal game-tree sense is this entry's identity. A market's colloquial “perfect information” about prices or product quality is not the same singleton-information-set condition and should not be silently folded into it.[1][2]
Structural Signature¶
Sig role-phrases: histories as tree nodes; decision maker; information-set partition; singleton test; contingent strategy; finite backward-induction consequence.
- Game tree and histories: each node represents a distinct action history; terminal nodes have outcomes or payoffs in the specified model.[1]
- Mover: a designated player chooses among available actions at a decision node. The question is what that player can distinguish then, not what a later observer knows.
- Information set: nodes the mover cannot tell apart are grouped; actions available must agree within the set.[1]
- Singleton condition: perfect information means each such set has one node. Linking two decision nodes because an earlier choice is hidden breaks the condition.[1]
- Contingent plan: a strategy can specify an action after each distinct observed history, including histories not reached by the planned path.[1]
- Solution consequence: for a finite tree, backward induction selects optimal continuations at every subgame and gives a pure-strategy subgame-perfect equilibrium; this is a theorem, not the definition.[1]
Condensed: observed history at every decision node → singleton information sets; finite payoff games then admit backward-induction solutions.
What It Is Not¶
- Not complete information: complete information concerns the modeled players, strategies and payoffs/types. Perfect information concerns the mover's distinction among histories. A simultaneous game may have common-known payoffs but a non-singleton information set for the player represented as moving second.[1][3]
- Not perfect recall alone: remembering one's own earlier choices is weaker than observing all relevant earlier moves.
- Not foreknowledge: the opponent's future action is not already known.
- Not a unique equilibrium theorem: ties can give multiple backward-induction continuations. The finite-game claim is existence of a pure-strategy subgame-perfect equilibrium.[1]
- Not market transparency by definition: knowing all current prices or quality information is a different economic idealization, not necessarily a sequential game with singleton decision sets.
- Not unbounded determinacy: any win/lose determinacy claim must specify a finite tree or additional hypotheses for an infinite game.
Scope of Application¶
The precise setting is a sequential extensive-form representation. Every history that ends at a player's decision is identifiable to that player. Chance can be in the tree, but a chance result that affects the mover's node cannot be hidden from that mover while the representation is still perfect information. Whether actors know the game's payoff function is a separate modeling question; ordinary backward-induction examples specify it.[1][3]
MIT's entry-deterrence tree has an entrant who chooses In or Out and an incumbent who, if entry occurs, observes In before choosing Accommodate or Fight. There is one decision node for the incumbent after In, so the information test passes. Backward induction gives Accommodate after entry and In initially in the lecture's payoffs. A strategic-form Nash equilibrium sustained by a threat to Fight after entry is not subgame-perfect. The singleton condition made the continuation an independent subgame, but the payoff comparison—not information alone—selects the equilibrium.[1]
The same lecture models a two-stage matching-pennies variant where player 2 can observe player 1's H or T before choosing. Player 2 then has two distinct decision nodes, not one blurred set; their strategy specifies what to do after both possible first moves. When the lecture instead joins these nodes into one information set to represent simultaneous choice, the strategic situation becomes imperfect-information even though the payoff table is known. That one changed observation relation is a clean identity boundary.[1]
Clarity¶
Ask exactly which player knows which prior branch at the moment of choosing. Publicly visible history suffices only if every decision-making player can distinguish the node relevant to their move. “Full information” is imprecise: it might mean known payoffs, observed history, or a rich signal about state. Draw the information-set partition. A singleton at each decision node proves perfect information; a non-singleton set disproves it. The assumption is about a representation, so two models of an institution can differ when one hides a move or signal.[1][3]
The finite backward-induction result does not extend without care to infinite trees or games with simultaneous hidden actions. Nor does a mathematically defined equilibrium imply agents can compute it for chess-sized state spaces. The claimed output depends on payoff and rationality assumptions distinct from the informational classification.[1][2]
The singleton definition itself has no built-in moral or engineering cost. Backward induction is a conditional implication of a finite modeled game, while computational size and behavioral realism are limits on using that implication rather than additional structural tensions.[2]
Manages Complexity¶
Singleton information sets permit the analyst to reason at each history separately: the continuation after one branch is a proper subgame, so terminal choices can be evaluated and rolled backward. This replaces simultaneous reasoning over indistinguishable nodes by conditional reasoning over visible histories. It does not eliminate game-tree growth. The entry game is tiny enough for backward induction to be transparent; merely labeling chess perfect-information does not give a practical complete solution. Joining the matching-pennies nodes to hide player 1's choice removes a continuation distinction on which the simple backward-induction demonstration depended.[1][3]
Abstract Reasoning¶
Represent an alleged case as a tree, annotate its mover at each node, and mark which nodes each mover can distinguish. If all decision information sets are singletons, the formal classification follows regardless of whether payoffs are favorable, players are altruistic, or the tree is large. If an earlier action is unobserved, link the compatible nodes; the classification changes even if the actual action sequence and terminal payoff labels are unchanged.[1][3]
For a finite perfect-information payoff game, start at terminal predecessor nodes, choose best available actions (allowing ties), replace their continuation subtrees by the selected payoffs, and proceed toward the root. The resulting contingent strategy must include decisions in off-path subgames as well. A bare predicted first move without those contingencies is not the full subgame-perfect result.[1]
Diagnostic: At every turn, is the mover's information set a singleton, and are any claimed equilibrium consequences scoped to a finite specified payoff game?
Knowledge Transfer¶
The live Information prime names a broad portable concern with distinctions available to an agent. This entry specializes it to a precise partition of histories at decision nodes. The move from observation to contingent planning travels to decision trees and sequential mechanism models, but the singleton rule itself does not describe general market transparency or intelligence. Replacing a hidden move by an observable one changes the model's subgames and sometimes its credible commitments; replacing unknown payoffs with known ones addresses a different axis.[1][3]
Examples¶
Entry deterrence: observed entry and a credible continuation¶
In Ozdaglar's MIT lecture tree, Out gives (1,2), while In leads the incumbent to choose Accommodate, giving (2,1), or Fight, giving (0,0), with payoff order entrant, incumbent. The incumbent observes entry and prefers its payoff 1 under Accommodate over 0 under Fight. The entrant therefore chooses In, receiving 2 rather than 1. A threat of Fight can support a normal-form Nash equilibrium (Out,Fight), but it fails subgame perfection because Fight is not optimal in the reached-after-In subgame. This is a source's stylized model, not an empirical claim about actual incumbents.[1]
Mapped back: tree = entrant's In/Out followed by incumbent's A/F after In; decision set = the observed In node alone; contingent choice = A; finite backward induction = (In,A); boundary = payoffs, not mere visibility, determine that choice.
Sequential matching pennies versus simultaneous choice¶
Ozdaglar's lecture explicitly contrasts two-stage observed matching pennies with a simultaneous-choice extensive-form representation. In the first, player 2 observes whether player 1 chose H or T and can condition their second move on that history; their two decision nodes are separate singleton sets. In the simultaneous version, the second mover's two nodes are joined because the first move is unobserved. The two forms can have the same terminal payoff table yet different information and strategy possibilities. One cannot import the pure backward-induction consequence from the observed tree into the joined-node game.[1]
Mapped back: history nodes = after H and after T; information partition = two singletons when observed, one two-node set when hidden; changed role = observation, not terminal payoff labels; boundary = hidden first move makes the extensive form imperfect-information.
A board position is not a payoff oracle¶
A chess position has publicly visible legal-move history and no hidden cards in the usual game model; its next mover can identify the current node. That makes it a familiar Classification example, but the mere classification is not a source-attested practical chess-solving algorithm. The finite-tree theorem concerns existence under formalized terminal outcomes; computational feasibility is a distinct question. This example is illustrative rather than a claim that the cited lectures solved chess.[1]
Mapped back: history = public position and prior moves; information set = current decision node; theorem boundary = formal existence does not imply feasible full search.
Structural Tensions¶
Observability versus strategic concealment. Revealing an earlier move lets a later player condition a response on the actual branch and makes the corresponding decision nodes distinct. Concealing it can protect a first mover from a tailored response but prevents the second mover from exploiting branch-specific information. In Ozdaglar's observed versus simultaneous matching-pennies representations, the terminal payoff table can stay fixed while this informational choice changes possible contingent strategies. Diagnostic: Which actor benefits from revealing or hiding the first move in this payoff structure?[1]
Structural–Framed Character¶
The core classification lies toward the structural end: a game tree plus its information-set partition decides it without asking whether outcomes are desirable. Its evaluative weight enters when a modeler asks whether observability, credibility, or the predicted equilibrium is desirable for participants. Human practice matters twice: institutions determine which actions are actually visible, and analysts decide which histories, signals and payoffs belong in the formal tree. The terminology arose in game theory's extensive-form analysis rather than as a promise of everyday omniscience; an economic market's “perfect information” can travel under the same words while requiring a different test. Importing singleton sets into market price disclosure without a sequential decision model is invalid, while recognizing different observed-history games under one formal condition is justified. Conversely, complete knowledge of a payoff matrix does not make hidden simultaneous choices observable. Its character: a structural condition on what each mover can distinguish in a sequential game, with conditional solution consequences and context-dependent strategic value.[1][3]
Structural Core vs. Domain Accent¶
The portable skeleton is that an agent's possible histories are partitioned by what it can distinguish; the broad Information prime can own that abstraction. The domain-bound mechanism is a game tree, player-assigned decision nodes, information sets and subgame/backward-induction reasoning under finite payoffs. The named entry fails the prime bar because “all information is available” outside a sequential extensive-form representation does not entail singleton decision sets, and general information structures may have no strategic player or off-path contingent action. The market usage is an adjacent vocabulary sense, not a reason to promote this game-theoretic condition to a universal prime.
Instantiates / Related Primes¶
Unparented root: no verified live formal extensive-game genus supplies a necessary parent for the singleton-information-set condition. Game is a played activity, not every analytical model; Game Form maps actions to outcomes without payoffs and need not encode this extensive history partition. Information, Determinacy and Perfect Competition remain neighbors, not forced parents. This condition is not omniscience, and backward induction retains finite sequential payoff hypotheses.
Neighborhood in Abstraction Space¶
Perfect Information sits in a moderately populated region (56th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Game-Theoretic Models & Paradoxes (37 abstractions)
Nearest neighbors
- Cognitive Hierarchy Theory — 0.86
- Bayesian Nash Equilibrium — 0.85
- Subgame Perfect Equilibrium — 0.85
- Game complexity — 0.85
- Centipede Game — 0.85
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Complete information concerns knowledge of game structure/payoffs or types, not observed moves. Perfect recall concerns remembering one's own information/action history, not seeing all others' past choices. Subgame perfection is a solution refinement, not the information condition. Perfect competition/market information has different institutional assumptions and no automatic singleton-set equivalence.[1][3]
References¶
[1] Asu Ozdaglar, MIT 6.254 Lecture 12, “Extensive Form Games”, pp. 4, 9–13, 19–22: entry game, observed/hidden matching pennies, information sets and finite pure-strategy SPE. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x
[2] Alexander Wolitzky, MIT 14.126 Lecture 2, “Equilibrium Refinements”, pp. 1–4: backward induction's rationality/behavioral assumptions. registry ↩a ↩b ↩c
[3] MIT 17.810 Lecture 4, “Extensive Form Games with Complete Information”, pp. 25–27 and 37–38: complete/perfect distinction and finite backward induction. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h