The Centipede game is a two-player extensive-form game in game theory, introduced by Robert W. Rosenthal in 1981, in which players alternate turns with perfect information, choosing at each decision node whether to "take" (end the game and claim the majority of an accumulating pot of payoffs) or "pass" (allow the pot to grow larger for the next player).[1] The game's structure resembles the elongated body of a centipede, with each segment representing a sequential decision point, and payoffs increase multiplicatively along the path until a terminal node is reached.[1]Backward induction reveals a unique subgame perfect Nash equilibrium in which the first player takes immediately on their initial turn, yielding the lowest possible payoffs for both players, as rational anticipation of defection at every subsequent node unravels cooperation from the end.[1] However, experimental studies consistently demonstrate that human players deviate from this prediction, often passing multiple times to extend the game and achieve higher joint payoffs, suggesting bounded rationality, fairness concerns, or learning effects that challenge standard rational choice assumptions.[2] In the seminal experiments by Richard D. McKelvey and Thomas R. Palfrey in 1992, using 4- and 6-move versions, subjects terminated the game after an average of about 2.4 and 3.5 passes, respectively, far exceeding the equilibrium path.[2]The Centipede game has become a cornerstone for investigating behavioral anomalies in strategic interactions, influencing research on topics such as altruism, reciprocity, and evolutionary game theory, while variants extend it to multi-player settings or imperfect information to model real-world dilemmas like arms races or environmental cooperation.[3] Its paradox underscores the limitations of complete rationality in finite games, prompting models like quantal response equilibrium to reconcile theory with observed play.
Introduction and History
Overview
The centipede game is a two-player, finite, extensive-form game in game theory designed to explore the tension between short-term individual gain and long-term mutual cooperation. In this game, players alternate turns deciding whether to claim a portion of a gradually increasing pot of resources or to pass the opportunity to the opponent, thereby allowing the pot to grow further. The structure creates a dilemma where continuing the game benefits both players if cooperation persists, but each decision point tempts defection for an immediate larger share at the other's expense.[1]The game highlights a core paradox in rational choice theory: while backward induction—a standard analytical tool—predicts that fully rational players will defect immediately to secure a modest gain, experimental observations consistently show participants engaging in cooperative passes far beyond what theory anticipates. This discrepancy challenges assumptions about human rationality and has made the centipede game a staple for studying bounded rationality, fairness, and social preferences in sequential decision-making.[1]Introduced by Robert W. Rosenthal in 1981, the centipede game draws its name from the game's tree-like structure, resembling the elongated body of a centipede, with each "leg" representing a potential decision point.[1]To illustrate the core dilemma, consider a simplified two-move version: Player 1 faces an initial pot and can choose to take a small share, leaving even less for Player 2, or pass to double the pot; if passed, Player 2 then decides whether to take the majority, ending the game with Player 1 receiving the remainder, or theoretically pass further—but since it ends here, rational play suggests Player 2 would take all available. This setup underscores how foresight of the opponent's incentives unravels cooperation from the end backward.[1]
Origins and Development
The centipede game was first introduced by Robert W. Rosenthal in his 1981 paper titled "Games of Perfect Information, Predatory Pricing, and the Chain-Store Paradox," published in the Journal of Economic Theory.[1] Rosenthal developed the game as a finite extensive-form model with perfect information to illustrate paradoxes arising from backward induction in sequential decision-making scenarios.[1]The primary motivation for creating the centipede game stemmed from challenges in industrial organization theory, particularly the chain-store paradox—originally posed by Reinhard Selten in 1978—and issues surrounding predatory pricing strategies under complete information.[1][4] In this context, Rosenthal used the game to highlight how rational players, anticipating future defections, would unravel cooperative opportunities from the end backward, leading to outcomes that seem counterintuitive despite perfect rationality assumptions.[1] The chain-store paradox involves an incumbent firm facing potential entrants in multiple markets, where backward induction predicts aggressive pricing in every period to deter entry, even though limited repetitions should allow for accommodation.[4]Following its introduction, the centipede game saw extensions in the 1990s that refined its structure for broader applications, such as analyzing simultaneous decision-making in normal form representations, though these did not fundamentally alter the core theoretical predictions of backward induction.[5] No major theoretical shifts emerged post-1980s, as the game's emphasis on unraveling under rationality remained central to discussions in non-cooperative game theory.[6]The centipede game significantly influenced the development of experimental game theory, prompting the first laboratory tests in the early 1990s to probe deviations from theoretical predictions. Pioneering experiments by Richard D. McKelvey and Thomas R. Palfrey in 1992 demonstrated that human subjects often continued play beyond what backward induction prescribes, sparking ongoing research into bounded rationality and behavioral anomalies.[2] Research on the centipede game has continued into the 2010s and beyond, examining topics such as preferences-dependent learning (Gamba and Regner 2019) and experienced payoff dynamics in cooperation (Sandholm, Izquierdo, and Izquierdo 2019).[7][8]
Game Mechanics
Rules of Play
The Centipede game involves two players, typically labeled Player 1 and Player 2, who alternate turns making decisions at a series of sequential nodes, with the total number of nodes fixed in advance and often even to balance the turns (for example, up to 100 nodes in extended versions). Player 1 initiates play at the first node, and the players continue alternating until the game concludes.At each node, the current player must choose between two actions: "take" (also called "stop" or "out"), which terminates the game and allocates the current payoffs immediately, or "pass" (also called "continue" or "in"), which augments the available payoff pot and transfers control to the opponent at the subsequent node. The initial pot is small, such that if Player 1 takes at the outset, they receive a modest positive payoff while Player 2 receives nothing. Each pass increases the pot, with payoffs growing along the path—in the original formulation linearly, though often multiplicatively (e.g., doubling) in experimental versions—enhancing the potential returns for both players should the game proceed further.[9][2]The game concludes either when a player selects "take," awarding the current player the larger portion of the pot and the opponent a smaller share (often 0 initially), or automatically upon reaching the final node, where the last player (Player 2 in even-length games) takes the pot, securing the highest possible payoff for themselves and a substantial but lesser amount for Player 1. In the standard formulation, the payoff structure is asymmetric, with continuation generally benefiting the second mover more in prolonged play, though some variants introduce symmetry by adjusting allocations. For illustration, a common short version with four nodes features initial payoffs of 1 for Player 1 and 0 for Player 2 if taken early, escalating to 4 for Player 2 and 2 for Player 1 at the end if all passes occur.[10]
Formal Definition
The centipede game is formalized as a two-player extensive-form game with perfect information, consisting of a finite sequence of alternating decision nodes labeled as "legs," where the total number of legs n is even. Player 1 moves first at leg 1 (an odd-numbered leg), followed by Player 2 at leg 2 (even-numbered), and so on up to leg n, with each player aware of all prior moves due to the perfect information structure. The game tree is a single path branching at each decision node, where the player to move chooses either to "take" (terminate the game and claim the larger share of the current pot) or "pass" (continue to the next leg, growing the pot).[9]At leg t, if the player takes, the payoffs are (a_t, b_t) for (Player 1, Player 2), where a_t + b_t equals the pot size at that leg, and the taking player receives the larger share to incentivize defection: specifically, a_t > b_t if t is odd (Player 1 takes) and b_t > a_t if t is even (Player 2 takes). In the original formulation, the pot size increases linearly with each pass; in many later versions, it grows multiplicatively by a factor greater than 1 (commonly 2). At the final leg n, the player takes the pot, with payoffs (a_n, b_n) where b_n > a_n. The strategy space consists of pure strategies for each player, defined as functions specifying an action (take or pass) at every possible decision node reachable by that player. For Player 1, this includes up to n/2 nodes (odd legs); for Player 2, up to n/2 nodes (even legs). Since information sets are singletons under perfect information, strategies fully specify behavior contingent on the history of passes.[9][2]
Theoretical Analysis
Backward Induction
Backward induction is a solution concept for finite games of perfect information, where strategies are determined by starting at the terminal nodes of the game tree and iteratively selecting the optimal action at each preceding decision node, assuming rational play in all subgames. In the Centipede game, this method assumes common knowledge of rationality among players, perfect information about the game structure and payoffs, and a payoff monotonicity condition where the total pot grows with each pass (typically by a factor r > 1), but at every node, the current player's payoff from taking exceeds their expected payoff from passing under optimal subsequent play.[1][11]To apply backward induction to the Centipede game, consider a simplified four-leg version where the pot starts at 1 unit, doubles with each pass, and the taker claims most of it (e.g., payoffs are structured such that taking at leg k gives the current player more than continuing). At the final leg (leg 4), the player to move has no option to pass further and will take the entire pot, securing their maximum possible payoff at that point—say, 8 units for Player 2 if the pot has grown to 8.[12][11]Moving backward to leg 3, the player (Player 1) anticipates that passing would lead to Player 2 taking at leg 4, yielding Player 1 only a small share (e.g., 0 units after growth). Since taking at leg 3 provides a higher payoff (e.g., 4 units for Player 1 from a pot of 4), rational play dictates taking immediately. This logic unravels further: at leg 2, Player 2 knows Player 1 will take at leg 3, so passing at leg 2 yields less than taking (e.g., 2 vs. 0), prompting defection; similarly, at leg 1, Player 1 foresees the chain of defections and takes the initial pot (1 unit), rather than passing to a path leading to 0.[1][13]The result is a unique subgame perfect Nash equilibrium in which Player 1 takes on the first move, terminating the game immediately and yielding payoffs of (1, 0) under the standard structure, despite the potential for higher joint payoffs if cooperation persisted. This equilibrium holds for any finite-length Centipede game due to the inductive unraveling process.[1][11]
Leg
Player to Move
Backward Induction Choice
Rationale
4
Player 2
Take (0, 8)
Terminal node; no pass option
3
Player 1
Take (4, 0)
Pass leads to (0, 8)
2
Player 2
Take (0, 2)
Pass leads to (4, 0) [self gets 0]
1
Player 1
Take (1, 0)
Pass leads to (0, 2) [self gets 0]
This table illustrates the unraveling for a basic four-leg Centipede with doubling pot and 100% take share, confirming the equilibrium path.[12]
Nash Equilibrium
In extensive-form games like the centipede game, a Nash equilibrium is a strategy profile in which no player can increase their payoff by unilaterally changing their strategy, assuming the other players' strategies remain fixed. This concept ensures stability in the overall game tree, but it does not necessarily require optimality in every possible subgame that might arise off the equilibrium path.[14]A refinement known as the subgame perfect Nash equilibrium (SPNE) addresses this by requiring that the strategy profile induces a Nash equilibrium in every subgame of the original game. In the finite centipede game, the unique SPNE is the strategy profile where both players always choose to take (defect) at every decision node they reach, leading to immediate defection by the first player and payoffs of (1, 0). This outcome is derived using backward induction, which eliminates non-optimal choices starting from the game's end.[15][16]Although the SPNE is unique, the centipede game admits other Nash equilibria that are not subgame perfect. These equilibria rely on non-credible threats or promises off the equilibrium path, such as a player committing to pass (continue) in a subgame where taking would yield a higher payoff if reached. However, all pure-strategy Nash equilibria in the finite game result in defection on the first move, as any profile supporting later play on the equilibrium path would allow profitable unilateral deviation by the first player. In infinite-horizon variants of the centipede game, additional equilibria supporting continued play can exist without such credibility issues, but the standard finite form does not permit them without non-credible elements.[15][16][13]The SPNE's robustness is further confirmed by the trembling-hand perfection refinement, which ensures that the equilibrium remains optimal even if players make small errors (trembles) in implementing their strategies with arbitrarily small probability. This property eliminates Nash equilibria supported only by knife-edge behaviors and underscores the SPNE as the strongest solution concept for the centipede game.[17]
Experimental Evidence
Key Findings
Laboratory experiments on the centipede game have consistently revealed substantial deviations from the theoretical prediction of immediate defection at the first decision node. In the seminal early study by McKelvey and Palfrey (1992), players exhibited high continuation rates in the initial stages, passing at the first node 93% of the time in 4-move games and 99% in 6-move games, with average game lengths reaching approximately 60% and 67% of the maximum, respectively.Subsequent research, including meta-analyses of experiments from 1992 to 2016, confirmed that continuation rates typically decrease progressively over successive legs, with games ending on average between 25% and 67% of their maximum length depending on design parameters such as payoff structure.[18] These patterns indicate persistent partial cooperation, though defection becomes more frequent as the end approaches, rarely reaching the final nodes in longer variants. Recent replications as of 2025 continue to show consistent patterns of partial cooperation.[19]Influencing factors include game length and stakes: shorter games tend to terminate earlier relative to their maximum, while higher monetary incentives lead to quicker defection compared to lower-stake conditions.[18] Although direct comparisons between incentivized and hypothetical play are limited, incentivized experiments generally show slightly lower continuation rates than non-binding hypothetical scenarios, suggesting greater realism amplifies strategic caution.Specific studies reinforce these trends; for instance, McKelvey and Palfrey (1992) observed elevated passing in early legs across variants, while cross-cultural experiments, such as a study comparing U.S. and Indian participants, reveal cultural differences, with Indian players defecting earlier on average (average game length of 1.63 rounds vs. 2.66 for U.S. players) despite some variations in decision timing.[20]Data from experiments spanning the 1990s through the 2020s remain remarkably consistent, with no significant shifts in cooperation levels post-2000, even as behavioral economics expanded, underscoring the robustness of observed anomalies.[18]
Explanations for Behavior
One prominent explanation for deviations from the subgame perfect equilibrium in the centipede game posits that players exhibit social preferences, such as altruism, fairness concerns, or inequality aversion, which motivate continuation moves to achieve mutual benefits rather than immediate self-interested taking.[21] In particular, models of inequity aversion, like the Fehr-Schmidt framework, suggest that players derive disutility from advantageous inequality, leading them to pass the pot to avoid leaving the opponent with a smaller payoff and to promote cooperative outcomes.[22] This preference-based approach accounts for observed continuation rates in experiments that exceed theoretical predictions of zero, as players weigh relational equity alongside monetary gains.[23]Bounded rationality models provide another key explanation, where players engage in limited depth of strategic thinking, assuming opponents have bounded foresight rather than perfect rationality.[24] Level-k thinking, an extension of cognitive hierarchy models, posits that players best respond to iteratively more sophisticated opponent behaviors, starting from a level-0 player who randomizes or assumes naive play; empirical analyses of centipede experiments show that level-1 or level-2 thinkers often continue early, explaining persistent deviations from backward induction. This framework, building on earlier work in guessing games, fits data where players stop later than equilibrium but earlier than fully cooperative play.[25]Learning and experience also influence behavior, with repeated exposure to the game reducing continuation rates over time as players adapt toward equilibrium play, though initial naivety leads to higher cooperation in early rounds.[3] Quantitative learning models, such as reinforcement or fictitious play, demonstrate that players update strategies based on observed outcomes, gradually converging closer to taking despite starting with optimistic beliefs about opponent cooperation.[3] However, full equilibrium learning remains incomplete even after multiple plays, reflecting persistent cognitive or preferential biases.[26]Additional factors include framing effects, where the presentation of the game—such as decomposed payoff structures emphasizing give-and-take—encourages more continuations by highlighting cooperative dynamics over competitive ones.[21] Stake size further modulates behavior, with smaller incentives promoting riskier continuations due to lower perceived costs of defection, while higher stakes align actions more closely with equilibrium predictions by amplifying self-interest.[27] These elements, intertwined with social preferences, underscore deviations driven by context beyond pure strategic calculation.[28]Critiques of these explanations highlight that no single model fully dominates, as pure rationality fails to capture data, while isolated preference or bounded rationality accounts overlook heterogeneous responses.[24] Hybrid approaches, such as quantal response equilibrium, better integrate noisy decision-making with strategic reasoning, allowing probabilistic deviations that match experimental patterns across player types.
Applications and Significance
In Game Theory
The centipede game serves as a key illustration of the chain-store paradox, highlighting the counterintuitive implications of backward induction in finite games of perfect information. In the chain-store paradox, originally formulated by Selten (1978), a monopolist faces sequential entry threats from potential competitors, and backward induction predicts acquiescence to entry at every stage despite incentives for deterrence early on, leading to a non-aggressive outcome that seems irrational. Similarly, Rosenthal (1981) introduced the centipede game to demonstrate how backward induction unravels potential cooperation in finite extensive-form games: starting from the terminal node where the final player takes the larger share, rational play propagates backward, compelling the first player to terminate immediately, yielding the subgame-perfect Nash equilibrium where no continuation occurs. This exposes limits of backward induction under strict rationality, as the predicted outcome forfeits mutually beneficial payoffs, prompting questions about the realism of assuming common knowledge of rationality in finite horizons.Theoretical extensions of the centipede game address these limitations by relaxing finiteness or perfection of information. In infinite-horizon versions, where the game potentially continues indefinitely with discounting, the unique subgame-perfect Nash equilibrium in pure strategies still prescribes immediate termination, but mixed-strategy equilibria allow for cooperation, with players randomizing continuation probabilities that sustain play for all but finitely many periods if payoffs are unbounded. Such models invoke the folk theorem for infinite-horizon repeated games, enabling a range of equilibria including cooperative outcomes sustainable via grim-trigger strategies, provided the discount factor is sufficiently high. Variants with noisy or imperfect information, such as observation errors or incomplete monitoring, further permit reputation-building and non-unraveling equilibria; for instance, small probabilities of "irrational" types can support continuation play akin to reputation effects in the chain-store setting.[29]The centipede game has fueled theoretical debates on rationality assumptions in game theory, challenging the standard backward induction outcome and influencing equilibrium refinements. Critics argue that common knowledge of rationality, required for backward induction, is implausibly strong in finite games, as players may doubt others' rationality at unreached nodes, leading to alternative rationalizable outcomes beyond the subgame-perfect equilibrium. Aumann (1995) defends backward induction by showing it follows from common knowledge of rationality in perfect-information games, yet this has spurred epistemic analyses questioning the assumption's applicability. These debates contributed to refinements like sequential equilibrium (Kreps and Wilson, 1982), which incorporates beliefs about off-path behavior and trembling-hand perfection to resolve indeterminacies in sequential games, providing a framework where limited rationality or noise sustains non-trivial play.[30]Since the 1990s, the centipede game has become a standard pedagogical tool in game theory education for teaching extensive-form games and backward induction. It appears prominently in influential textbooks, such as Osborne and Rubinstein's A Course in Game Theory (1994), where it exemplifies the tension between individual rationality and collective gain, and Fudenberg and Tirole's Game Theory (1991), which uses it to introduce subgame perfection and solution concepts. Its simplicity—alternating moves with escalating stakes—makes it ideal for classroom demonstrations of unraveling arguments without complex mathematics, fostering discussions on rationality and equilibrium selection in introductory and advanced courses.Recent theoretical work, such as a 2023 study on epistemic foundations using computational simulations and agent-based models, reaffirms the unraveling under strict rationality while exploring robustness. These models, assuming perfect rationality and common knowledge, consistently reproduce the backward induction outcome, with play terminating at the first node across thousands of iterations, underscoring the logical inevitability in finite settings. The computations also test extensions, showing that even minor deviations—like epsilon-rationality or finite precision—can delay but not prevent unraveling, influencing ongoing refinements in bounded rationality frameworks.[31]
In Behavioral Economics
In behavioral economics, the centipede game provides critical insights into the tension between cooperation and defection in sequential decision-making, revealing how players often prioritize long-term mutual gains over immediate self-interest, thereby informing models of trust and reciprocity. Experimental evidence consistently shows that participants cooperate far more than predicted by backward induction, with low rates, such as 1% to 15% of games ending on the first move in early experiments as rational choice theory expects, and cooperation rates increasing with game length and symmetric payoffs.[2][18] This behavior underscores reciprocity as a driver of sustained interaction, where players infer trustworthiness from prior passes, challenging purely selfish utility maximization and supporting hybrid models that incorporate other-regarding preferences.[18]The game's structure mirrors real-world applications in negotiations, auctions, and environmental agreements, where short-term defection temptations undermine collective outcomes. For auctions, the centipede game is isomorphic to Dutch auctions, where institutional formats (e.g., clock vs. tree representations) and move structures (sequential vs. simultaneous) influence bidding deviations from equilibrium, with clock formats leading to earlier "takes" analogous to overbidding.[32] In environmental contexts, the game illustrates how reciprocal commitments stabilize cooperation in agreements like treaties, where reassurance mechanisms prevent defection despite rising stakes, as seen in historical nonproliferation pacts that parallel climate accords requiring ongoing trust.[33]Behavioral interventions, such as communication, reputation-building, and commitment devices, significantly extend cooperative play in centipede experiments, often tested in field-like settings to mimic social dynamics. Social value orientations and reputation cues increase passing rates by enhancing perceived reciprocity, while direct communication or repeated interactions can raise cooperation, though group decisions sometimes accelerate defection due to reduced prosociality.[18]Commitment-enhancing tools, like purchasable bonuses or insurance against defection, further promote trust-based reciprocity, with mean exit points rising from 4.74 to 6.96 nodes when tools are available, and cross-cultural variations—higher in Japan (6.81) than Europe (4.69)—highlighting cultural influences on assurance-seeking.[34]The centipede game's empirical deviations contribute to broader critiques of rational choice theory, demonstrating bounded rationality and prosocial motivations over strict self-interest, while influencing integrations with frameworks like level-k thinking to explain persistent cooperation.[2][18] In the 2020s, it informs AI decision modeling in multi-agent systems, where reinforcement learning algorithms simulate human-like reciprocity to optimize sequential interactions, aiding designs for trustless environments like decentralized networks. A 2025 study on strategy method effects in centipede games found behavioral distortions in choiceelicitation, further highlighting methodological considerations in experimental designs.[35]