Kuhn's Theorem (Game Theory)

Kuhn's theorem is a foundational result about games that unfold over time. It says that in a finite game where a player has perfect recall, meaning the player never forgets anything they previously knew or did, two different ways of randomizing are equally powerful. The player can either pick a complete plan of action at random before the game starts (a mixed strategy), or randomize separately at each decision point as it comes up (a behavior strategy). For every mixed strategy there is a behavior strategy that produces exactly the same probabilities over how the game can end, whatever the other players do, and the reverse holds too.

The theorem is named after the American mathematician Harold W. Kuhn (1925–2014), who proved it in his 1953 paper "Extensive Games and the Problem of Information." The paper appeared in Contributions to the Theory of Games, Volume II, edited by Kuhn and A. W. Tucker as number 28 in Princeton University Press's Annals of Mathematics Studies series, on pages 193 to 216.

Kuhn's theorem is rarely stated in the headline of a modern paper, but it is quietly used almost everywhere sequential games are analyzed. It is the reason economists can describe a poker player, a bargainer or a firm as choosing "what to do at each point" rather than choosing among an astronomically large list of complete plans. It is also central to modern algorithms that solve large games of imperfect information, including those used to solve versions of poker. And its failure when perfect recall is absent has produced some of the most debated puzzles in decision theory.

The idea in plain terms

Consider a player who will face several decisions during a game. There are two natural ways that player could introduce randomness.

Randomize once, at the start. Before play begins, write down every complete plan of action, one that says what to do at every decision point that could possibly arise. Then pick one plan by spinning a wheel, and follow it. This is a mixed strategy.

Randomize as you go. Each time a decision comes up, spin a separate wheel just for that decision. The wheels for different decisions are independent. This is a behavior strategy.

The first approach is more general in one sense. Choosing a whole plan at random can create correlation between decisions: "with probability one-half I'll do A now and C later; with probability one-half I'll do B now and D later." Independent wheels at each decision cannot obviously reproduce that kind of linkage.

Kuhn's insight was that if the player remembers their own past moves and what they knew, the correlation is unnecessary. When the later decision arrives, the player already knows whether they did A or B earlier, so they can simply use a different wheel depending on what they did. The memory does the work that correlation would otherwise do. The two approaches then generate the same distribution over outcomes.

If the player forgets, this breaks down. A forgetful player at the later decision does not know which earlier choice they made, so a single wheel at that point cannot depend on it. Some mixed strategies then have no behavior equivalent. In some unusual games with very severe forgetfulness, the reverse can also fail: behavior strategies can achieve things no mixed strategy can.

Historical background

Extensive games before Kuhn

John von Neumann and Oskar Morgenstern's Theory of Games and Economic Behavior (1944) described games that unfold over time, with moves, chance events and information. But their formulation was cumbersome, and much of their analysis converted such games into the normal form, a table listing each player's complete strategies and the resulting payoffs. That conversion is general but loses the structure of the game and can be enormous.

Kuhn's 1950 and 1953 papers

Harold Kuhn reformulated sequential games in terms of game trees. His short paper "Extensive Games," published in the Proceedings of the National Academy of Sciences in 1950 (volume 36, issue 10, pages 570 to 576), announced the approach. The 1953 paper developed it fully. In Kuhn's formulation:

• The game is a tree. Each non-terminal node belongs either to one of the players or to chance.
• Each player's nodes are grouped into information sets. A player who reaches a node in some information set knows only that they are somewhere in that set, not which node exactly.
• At every node in an information set, the player has the same set of available actions.
• Payoffs are attached to the terminal nodes.

This tree-and-information-set description is still the standard definition of an extensive-form game. In the 1953 paper Kuhn also defined perfect recall, introduced behavior strategies in this setting, proved the equivalence theorem that now carries his name, and showed that every finite game of perfect information (where every information set is a single node, as in chess) has an equilibrium in pure strategies, generalizing earlier work by Ernst Zermelo on chess.

Kuhn's theorem is often described as having a converse: perfect recall is not just sufficient for every mixed strategy to have a behavior equivalent, it is also, in Kuhn's characterization, the condition under which this holds in general. Textbook treatments present the theorem in both directions with this qualification.

Kuhn himself

Harold William Kuhn spent most of his career at Princeton University. Besides his game-theory work, he is known for the Karush–Kuhn–Tucker conditions in nonlinear optimization (from a 1951 paper with Albert W. Tucker) and for the Hungarian method for the assignment problem (1955). In 1950 he published a small poker model, "A simplified two-person poker," in the first volume of Contributions to the Theory of Games; the three-card game now called Kuhn poker remains a standard test case in game-solving research. He also edited the anthology Classics in Game Theory.

Extension to infinite games

Kuhn's theorem concerns finite games. Robert Aumann extended it to infinite extensive games in "Mixed and behavior strategies in infinite extensive games," in Advances in Game Theory, edited by M. Dresher, L. S. Shapley and A. W. Tucker (Annals of Mathematics Studies, volume 52, Princeton University Press, 1964, pages 627 to 650). Infinite games here include those with a continuum of possible actions at a decision point and those with infinitely many moves. Making the definitions precise in that setting requires measure theory, since "picking a plan at random" now means picking from an infinite-dimensional space.

The definitions

Pure strategies

A pure strategy for a player is a complete contingency plan: it specifies an action at every information set belonging to that player, including information sets that the plan itself prevents from being reached. If a player has k information sets and m_1, m_2, ..., m_k actions at them, the number of pure strategies is the product m_1 × m_2 × ... × m_k.

Mixed strategies

A mixed strategy is a probability distribution over pure strategies. The player randomizes once, before the game begins.

Behavior strategies

A behavior strategy assigns, to each of the player's information sets, a probability distribution over the actions available there. The randomizations at different information sets are independent. A behavior strategy is "local": it says how to act at each point given only what the player knows there.

Perfect recall

A player has perfect recall if, informally, they never forget anything they once knew, including their own past actions. Formally: for any two nodes in the same information set of that player, the sequence of that player's own earlier information sets and the actions taken at them, on the path from the root to each node, is identical.

Equivalently, if two nodes are in the same information set, the player cannot have reached them by taking different actions earlier or by passing through different earlier information sets of their own. If the player had done something different, or had known something different, they would remember it and therefore be able to tell the two nodes apart.

A game has perfect recall if every player has perfect recall. Perfect recall does not mean the player knows everything: they may still be ignorant of chance moves and other players' moves. It means only that their own knowledge never decreases.

Realization equivalence

Two strategies of the same player are realization equivalent (Kuhn used the term "equivalent") if, for every combination of strategies the other players might use, they induce the same probability distribution over the terminal nodes of the tree. Since payoffs depend only on terminal nodes, realization-equivalent strategies give every player the same expected payoff against every opponent behavior. They are interchangeable for all purposes of analysis.

Statement of the theorem

Kuhn's theorem. In a finite extensive-form game, suppose player i has perfect recall. Then for every mixed strategy of player i there is a behavior strategy of player i that is realization equivalent to it.

The other direction, from behavior strategies to mixed strategies, holds in a wide class of games without any recall assumption: given a behavior strategy, you can construct a mixed strategy by assigning each pure strategy the product of the probabilities the behavior strategy places on that pure strategy's actions. The resulting mixed strategy is realization equivalent to the behavior strategy as long as no path through the tree passes through the same information set twice. That condition fails only in games with "absentmindedness," discussed below.

So in finite games with perfect recall, mixed and behavior strategies are interchangeable. Several consequences follow directly:

• Nash equilibria in behavior strategies exist. John Nash proved that every finite game has a Nash equilibrium in mixed strategies. Under perfect recall, each player's equilibrium mixed strategy can be replaced by an equivalent behavior strategy, and the resulting profile is still an equilibrium.
• Behavior strategies are without loss of generality. Any analysis of a perfect-recall game can be done entirely in behavior strategies.
• Payoffs are preserved. Since outcome distributions match, so do expected payoffs, for every player.

The theorem does not say the equivalent behavior strategy is unique, and it does not say each behavior strategy corresponds to a unique mixed strategy. In fact, many different mixed strategies typically map to the same behavior strategy.

How the construction works

The proof is constructive. Given a mixed strategy σ for a perfect-recall player, the equivalent behavior strategy β is defined as a conditional probability.

For each information set h of the player:

1. Look at the pure strategies that are consistent with reaching h: those whose earlier choices are the ones on the path to h. Perfect recall ensures this set is well defined, because every node in h is preceded by the same earlier choices of this player.
2. Among those, look at the ones that choose action a at h.
3. Set β(a at h) = [total weight σ places on pure strategies that reach h and choose a] / [total weight σ places on pure strategies that reach h].

If the denominator is zero, so that the mixed strategy never reaches h, the behavior at h can be chosen arbitrarily, because it never affects the outcome.

Why the probabilities match

Consider any terminal node. To reach it, the player must take a specific sequence of actions at a specific sequence of their own information sets, say a₁ at h₁, then a₂ at h₂, and so on up to a_n at h_n. (Perfect recall ensures this sequence is the same however the opponents and chance behave on the way.) Under the mixed strategy, the probability that the player's part of this path is followed equals the total weight on pure strategies that take all of those actions.

Under the behavior strategy, the probability is the product β(a₁ at h₁) × β(a₂ at h₂) × ... × β(a_n at h_n). Each factor is a ratio: the weight on pure strategies consistent with the path up through a_t, divided by the weight on those consistent with the path up to h_t. Because of perfect recall, "consistent with the path up to h_{t+1}" is exactly the same set as "consistent with the path through a_t at h_t." So each numerator cancels the next denominator, and the product telescopes to the weight on pure strategies consistent with the whole sequence. The opponents' and chance's contributions to the probability are the same under both strategies, so the total probability of the terminal node is the same.

Perfect recall is used at exactly one point: to make sure that "the player's own earlier choices on the path to h" are the same for every node in h. Without that, the conditional probability in step 3 would mix together pure strategies that reached h in different ways, and the telescoping argument would fail.

Worked examples

All numbers in these examples are hypothetical and chosen for clarity.

Example 1: Converting a mixed strategy (perfect recall)

Player 1 moves twice. First she chooses Left (L) or Right (R). Then, after an intervening move by Player 2 that does not matter here, she moves again. She remembers her first move, so her second decision is in one of two information sets: h_L if she chose Left, with actions a or b, and h_R if she chose Right, with actions c or d.

She has 2 × 2 × 2 = 8 pure strategies, each specifying her first move, her choice at h_L and her choice at h_R. Suppose her mixed strategy puts weight on four of them:

Notice the correlation built into this mixed strategy: it never plays a with d or b with c.

Constructing the behavior strategy.

• First move: the weight on strategies choosing L is 0.2 + 0.1 = 0.3. So β(L) = 0.3 and β(R) = 0.7.
• At h_L: the strategies that reach h_L are those choosing L, with total weight 0.3. Of these, the ones choosing a have weight 0.2. So β(a at h_L) = 0.2/0.3 = 2/3, and β(b at h_L) = 1/3.
• At h_R: the strategies that reach h_R are those choosing R, with total weight 0.7. Of these, the ones choosing c have weight 0.3. So β(c at h_R) = 0.3/0.7 = 3/7, and β(d at h_R) = 4/7.

Checking the outcome probabilities. Player 1's part of the path to each terminal node:

They match exactly. The correlation in the mixed strategy (a with c, b with d) never mattered: whichever first move she made, only one of h_L or h_R is reached, so how she would have acted at the other one is irrelevant to the outcome.

Going back the other way. If we turn this behavior strategy into a mixed strategy by multiplying probabilities, we get a different mixed strategy. For example, (L, a, c) would receive 0.3 × 2/3 × 3/7 ≈ 0.086 rather than 0.2. That mixed strategy is realization equivalent to the original one even though it assigns different weights to pure strategies. This illustrates that many mixed strategies collapse to the same behavior strategy.

Example 2: Forgetting your own move

Change the game so that Player 1 forgets her first move. Her second decision is now a single information set where she cannot tell whether she chose L or R, with actions l and r. Meanwhile Player 2, who cannot see anything, simultaneously guesses Player 1's first move.

Payoffs (zero-sum, to Player 1): Player 1 gets 1 if her second move "matches" her first (L then l, or R then r) and Player 2's guess is wrong. Otherwise she gets 0.

With a mixed strategy, Player 1 can pick (L, l) with probability 1/2 and (R, r) with probability 1/2. She always matches. Whatever Player 2 guesses, it is wrong half the time. Player 1 guarantees an expected payoff of 1/2.

With a behavior strategy, she chooses L with some probability p at the first move, and l with some probability q at the second, independently, since the second decision cannot depend on a move she has forgotten. If Player 2 guesses L with probability g, Player 1's expected payoff is:

p × q × (1 − g) + (1 − p) × (1 − q) × g

Player 2 will guess whichever first move is more likely to be matched. The best Player 1 can guarantee is the smaller of p × q and (1 − p) × (1 − q). That smaller number can never exceed the square root of their product, √[p(1 − p) q(1 − q)], which is at most √(1/4 × 1/4) = 1/4. And p = q = 1/2 achieves exactly 1/4.

So the best behavior strategy guarantees 1/4, while the mixed strategy guarantees 1/2. Without perfect recall, the mixed strategy cannot be replicated by a behavior strategy, and the difference has real payoff consequences. The mixed strategy's correlation between first and second moves is valuable, and forgetting makes that correlation impossible to produce locally.

Example 3: The absent-minded driver

The most famous example of imperfect recall comes from Michele Piccione and Ariel Rubinstein's paper "On the Interpretation of Decision Problems with Imperfect Recall," published in Games and Economic Behavior in 1997 (volume 20, pages 3 to 24).

A driver on a highway wants to get home. There are two exits. Taking the first exit leads to a bad area (payoff 0). Taking the second exit leads home (payoff 4). Missing both exits and continuing to the end of the road gives payoff 1. The twist: the driver cannot tell the two exits apart, and cannot remember whether they have already passed one. Both exits are in the same information set. This is called absentmindedness: a single path through the tree passes through the same information set twice.

Pure strategies. The driver has only one information set, so there are only two pure strategies: "always exit" (gets 0, since they exit at the first exit) and "always continue" (gets 1). Any mixed strategy is a lottery between these and yields at most 1.

Behavior strategies. Suppose the driver continues with probability p each time they reach an exit. Then:

• Exit at the first exit with probability 1 − p, payoff 0.
• Continue at the first, exit at the second, with probability p(1 − p), payoff 4.
• Continue at both, with probability p², payoff 1.

Expected payoff = 4p(1 − p) + p² = 4p − 3p².

Setting the derivative 4 − 6p to zero gives p = 2/3. The expected payoff is 4 × 2/3 − 3 × 4/9 = 8/3 − 4/3 = 4/3.

Here behavior strategies do better than mixed strategies, the reverse of Example 2. Randomizing afresh at each exit lets the driver sometimes continue past the first exit and then turn off at the second, which no randomization over fixed plans can achieve.

The example became famous less for the Kuhn-theorem point than for a puzzle Piccione and Rubinstein raised: the plan that looks optimal at the start (continue with probability 2/3) seems not to look optimal when the driver is actually at an exit and reasons about where they might be. A special issue of Games and Economic Behavior carried responses, including one by Robert Aumann, Sergiu Hart and Motty Perry arguing that while the planning-stage and action-stage considerations differ, there is no genuine paradox or inconsistency. The debate continues in the philosophy of decision theory.

Example 4: Why the reduction matters, by counting

Consider a player who faces k binary decisions, each at its own information set. The player has 2^k pure strategies. A mixed strategy is a probability distribution over all of them, which takes 2^k − 1 numbers to specify. A behavior strategy needs just one number per information set, k numbers in total.

The gap grows exponentially. In real games such as card games or multi-stage negotiations, players have huge numbers of information sets, and working with mixed strategies over complete plans quickly becomes impossible. Kuhn's theorem guarantees that, under perfect recall, nothing is lost by working with the much smaller behavior-strategy description.

(Many pure strategies in the full list differ only in what they do at information sets they themselves rule out, so a "reduced" normal form can be smaller, but it still typically grows exponentially in the size of the tree.)

Example 5: Counting strategies in Kuhn poker

Kuhn poker shows the reduction on a real, if tiny, game. The deck has three cards: Jack, Queen and King. Each of two players antes one chip and is dealt one card. Player 1 can check or bet one chip. If Player 1 checks, Player 2 can check (leading to a showdown) or bet; if Player 2 bets, Player 1 can call or fold. If Player 1 bets at the start, Player 2 can call or fold.

Player 1's information sets. Player 1 knows her own card and the betting so far. She has a first decision for each of her three possible cards, and a second decision (call or fold after checking and facing a bet) for each card. That is 6 information sets, each with 2 actions, so 2^6 = 64 pure strategies.

Player 2's information sets. Player 2 knows his card and whether Player 1 checked or bet. For each of his three cards he has one decision after a bet (call or fold) and one after a check (bet or check). Again 6 binary information sets and 64 pure strategies.

A mixed strategy for either player is a distribution over 64 plans, requiring 63 numbers. A behavior strategy requires 6: for each card and situation, a single probability of betting or calling. Both players have perfect recall (each remembers their card and every action taken), so by Kuhn's theorem the six-number description loses nothing.

Kuhn solved this game in 1950. A frequently cited feature of his solution is that the first player has a whole family of equilibrium strategies that differ in how often they bluff with the worst card, and that the game is slightly unfavorable to the first player, whose expected value is −1/18 of a chip per hand. These features are easiest to state in behavior-strategy terms, for example as "bet with the Jack with some probability α, and with the King with probability 3α."

Perfect recall: which cases count

Because so much rests on perfect recall, it helps to see which situations satisfy it.

The team case deserves a comment. In bridge, for example, the two partners are sometimes modeled as a single player with a shared objective. But each partner cannot see the other's hand, so the combined "player" forgets, from one partner's turn to the other's, information it had a moment ago. Games of this kind lack perfect recall, and Kuhn's equivalence does not apply to the team as a whole. This is one reason team games are studied as a separate topic in computational game theory.

Why the theorem matters

Equilibrium analysis in sequential games

Most equilibrium concepts for extensive games are defined in terms of behavior strategies. Reinhard Selten's subgame-perfect equilibrium requires players to play optimally at every subgame, which is naturally expressed as a choice at each decision point. David Kreps and Robert Wilson's sequential equilibrium (Econometrica, 1982) pairs behavior strategies with beliefs at each information set. Perfect Bayesian equilibrium does the same. These concepts describe what a player does at each information set, which is exactly what a behavior strategy specifies. Kuhn's theorem reassures us that, in perfect-recall games, restricting attention to behavior strategies does not exclude any outcome a mixed strategy could achieve.

Modeling

When economists write down a model of bargaining, signaling, auctions with several rounds, or repeated interaction, they almost always describe players' strategies as rules for what to do at each point given what they know. That modeling habit is justified by Kuhn's theorem whenever perfect recall is assumed, which it nearly always is.

Computation: the sequence form

Solving games in the normal form is hopeless for large games because the number of pure strategies explodes. The sequence form, introduced in work by I. V. Romanovskii (1962) and developed by Bernhard von Stengel in "Efficient Computation of Behavior Strategies" (Games and Economic Behavior, 1996, volume 14, pages 220 to 246), represents a perfect-recall player's strategy by the probabilities of sequences of that player's own actions. Its size grows linearly, not exponentially, with the size of the game tree. For two-player zero-sum games with perfect recall, equilibria can then be found by linear programming in time polynomial in the size of the tree. Daphne Koller, Nimrod Megiddo and von Stengel's companion paper "Efficient Computation of Equilibria for Extensive Two-Person Games" (Games and Economic Behavior, 1996, pages 247 to 259) extended this approach.

The sequence form works precisely because of Kuhn's theorem: under perfect recall, a "realization plan" over sequences corresponds to a behavior strategy, and every mixed strategy is realization equivalent to one.

Computation: counterfactual regret minimization

Martin Zinkevich, Michael Johanson, Michael Bowling and Carmelo Piccione's paper "Regret Minimization in Games with Incomplete Information" (NIPS 2007) introduced counterfactual regret minimization (CFR). CFR maintains a separate regret-minimizing learner at each information set and updates local action probabilities, which is a behavior strategy by construction. The authors showed that minimizing these local, counterfactual regrets minimizes overall regret, so self-play converges toward a Nash equilibrium in two-player zero-sum games. They demonstrated the method on abstractions of limit Texas Hold'em with as many as 10^12 states.

Descendants of CFR were used by Bowling, Neil Burch, Michael Johanson and Oskari Tammelin, whose 2015 paper in Science reported that heads-up limit hold'em poker had been essentially solved. The whole approach depends on representing strategies in behavior form, and its guarantees depend on perfect recall, which is where Kuhn's theorem comes in.

Imperfect recall as a design choice

Researchers sometimes deliberately use game abstractions with imperfect recall, letting an artificial player "forget" some details to save memory and computation. Doing so gives up Kuhn's guarantee, so behavior strategies may no longer capture everything a mixed strategy could do, and the theory of equilibrium computation becomes more delicate. This is an active area of research in computational game theory.

Mixed and behavior strategies compared

Common misconceptions

"Perfect recall means perfect information." No. Perfect information means every player always knows the entire history of the game, including other players' moves and chance moves. Perfect recall means only that a player never forgets their own past information and actions. Poker has perfect recall but imperfect information: players remember everything they saw and did but cannot see opponents' cards.

"Kuhn's theorem says mixed and behavior strategies are the same thing." They are different objects. A mixed strategy is a distribution over complete plans, and a behavior strategy is a collection of local distributions. The theorem says they are realization equivalent under perfect recall, meaning they produce the same distribution over outcomes against every opponent strategy.

"The equivalent behavior strategy is unique." At information sets that the mixed strategy reaches with positive probability, the conditional probabilities are pinned down. At information sets the mixed strategy never reaches, the behavior strategy can be anything. And many different mixed strategies map to the same behavior strategy.

"Kuhn's theorem requires perfect recall for everyone." The statement is player by player. If player i has perfect recall, every mixed strategy of player i has an equivalent behavior strategy, regardless of whether other players have perfect recall.

"Imperfect recall always favors mixed strategies." Example 2 shows mixed strategies doing better under ordinary forgetfulness. Example 3, the absent-minded driver, shows behavior strategies doing better when a player can pass through the same information set twice. The direction depends on the kind of imperfect recall.

"Kuhn's theorem only matters for theorists." It underpins the representations used in practical solvers for large imperfect-information games. Algorithms like CFR work in behavior strategies, and their guarantees rely on perfect recall.

"The theorem was about infinite games." Kuhn's 1953 result is for finite games. The infinite-game extension is due to Robert Aumann (1964).

Frequently asked questions

What exactly is "equivalent" in Kuhn's theorem? Two strategies of the same player are equivalent if, against every possible strategy of the other players, they produce the same probability distribution over terminal nodes. Equal expected payoffs follow automatically.

Is there a simple test for perfect recall? Check every information set of the player. For any two nodes in the same set, trace the paths from the root. If the player's own sequence of earlier information sets and actions is the same on both paths, the condition holds at that set. It must hold at every set.

Why does perfect recall make correlation unnecessary? Because a player who remembers what they did can condition later choices on it. Any correlation a mixed strategy builds between an earlier and a later decision can be reproduced by using different local randomizations depending on the remembered earlier decision, as in Example 1.

Does every finite game have an equilibrium in behavior strategies? Every finite game with perfect recall does, by combining Nash's existence theorem for mixed strategies with Kuhn's theorem. Without perfect recall, equilibrium in behavior strategies can fail to exist in some games, which is one reason imperfect-recall games are harder to analyze.

What did Kuhn's 1953 paper contain besides this theorem? It gave the now-standard tree-and-information-set definition of extensive games, formally defined perfect recall, and proved that every finite game of perfect information has an equilibrium in pure strategies.

Who extended the theorem to infinite games? Robert Aumann, in "Mixed and behavior strategies in infinite extensive games," published in Advances in Game Theory (Annals of Mathematics Studies 52) in 1964.

What is Kuhn poker? A simplified poker game with a three-card deck that Kuhn analyzed in 1950. It is small enough to solve by hand but has bluffing and imperfect information, which makes it a common teaching and testing example in game theory and AI.

What is the absent-minded driver paradox about? A driver who cannot tell two exits apart and cannot remember passing one. It shows that with absentmindedness, behavior strategies can beat mixed strategies (expected payoff 4/3 versus 1), and it raised a debated question about whether optimal plans remain optimal when the decision is actually made.

How does this relate to reinforcement learning? Methods that learn a policy assigning action probabilities to each information state are learning behavior strategies. In multi-agent settings with imperfect information, Kuhn's theorem is part of the reason that restricting attention to such policies is not a limitation when agents have perfect recall.

Key terms

Summary

Kuhn's theorem, proved by Harold W. Kuhn in 1953, says that in finite extensive-form games a player with perfect recall loses nothing by randomizing locally at each decision rather than choosing a complete plan at random. The proof constructs the behavior strategy as a set of conditional probabilities and uses perfect recall to make the probabilities telescope. The theorem justifies the standard way of describing strategies in sequential games, underpins equilibrium refinements like subgame perfection and sequential equilibrium, and makes compact game-solving methods such as the sequence form and counterfactual regret minimization possible. When perfect recall fails the equivalence can break in either direction, as the forgetful-player example and the absent-minded driver show, and those failures have become important test cases in both game theory and the philosophy of decision-making.

Sources

• Kuhn's theorem (Wikipedia)
• Kuhn, H. W. (1953), "Extensive Games and the Problem of Information," Contributions to the Theory of Games II, pp. 193–216
• Kuhn, H. W. (1950), "Extensive Games," Proceedings of the National Academy of Sciences 36(10)
• Piccione, M. and Rubinstein, A. (1997), "On the Interpretation of Decision Problems with Imperfect Recall," Games and Economic Behavior 20
• Aumann, Hart and Perry, "The Absent-Minded Driver" (abstract page)
• von Stengel, B. (1996), "Efficient Computation of Behavior Strategies," Games and Economic Behavior 14
• Koller, Megiddo and von Stengel (1996), "Efficient Computation of Equilibria for Extensive Two-Person Games," Games and Economic Behavior 14
• Zinkevich, Johanson, Bowling and Piccione (2007), "Regret Minimization in Games with Incomplete Information," NeurIPS
• Bowling, Burch, Johanson and Tammelin (2015), "Heads-up limit hold'em poker is solved," Science 347