Three-Player Kuhn Poker: Equilibrium Analysis

• Three-player Kuhn poker is a three-player imperfect-information game with a four-card deck that serves as a benchmark for equilibrium analysis and strategic experimentation.
• Research reveals an infinite family of equilibria and rich dynamics in bluffing, calling, and betting, highlighting the complexity even in minimal games.
• Studies employ analytic, computational, and dynamic models to investigate equilibrium multiplicity, opponent exploitation, and efficient Nash equilibrium computation.

Searching arXiv for recent and foundational papers on three-player Kuhn poker and related equilibrium computation. {"query":"three-player Kuhn poker equilibrium multiplayer imperfect-information arXiv", "max_results": 10} arXiv search results:

1. (Ganzfried, 24 Jun 2026) — "Variable Bound Tightening for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games"
2. (Ganzfried, 30 Sep 2025) — "Quadratic Programming Approach for Nash Equilibrium Computation in Multiplayer Imperfect-Information Games"
3. (Ganzfried et al., 2022) — "Opponent Modeling in Multiplayer Imperfect-Information Games"
4. (Billingham, 2018) — "Equilibrium solutions of three player Kuhn poker with N>3 cards: A new numerical method using regularization and arc-length continuation"
5. (Billingham, 2017) — "Full street simplified three player Kuhn poker"
6. (Billingham, 2017) — "Simplified three player Kuhn poker" Three-player Kuhn poker is a three-player imperfect-information poker game played with a four-card deck $D=\{J,Q,K,A\}$, in which one distinct card is dealt to each player, each player antes 1 chip, and the hand is resolved after a single betting round with actions $\{K,B,C,F\}$ subject to the usual legality restrictions (Ganzfried et al., 2022). Within the literature it functions as a compact setting for analyzing multiplayer Nash equilibrium, equilibrium multiplicity, repeated-play dynamics, exact sequence-form computation, and opponent exploitation. Research on the game and its variants shows that even this small extensive-form model can exhibit an infinite family of equilibria, multiple coexisting analytic solution branches, oscillatory adaptive dynamics, and substantial separation between equilibrium play and data-driven exploitation (Ganzfried et al., 2022, Billingham, 2017, Billingham, 2018).

1. Formal extensive-form structure

In the standard formulation, chance deals one distinct card to each of the three players uniformly at random. Each player $i\in\{1,2,3\}$ privately observes only their own card $c_i\in D$, while the other cards remain hidden. For each player $i$, each private card $c\in D$, and each history prefix $h$ at which it is $i$’s turn, there is an information set $I_{i,c,h}$; all histories in the same information set are indistinguishable to that player (Ganzfried et al., 2022).

The legal actions depend on whether a bet is already on the table. If no bet has yet been made, the player chooses between check $(K)$ and bet $\{K,B,C,F\}$0. If a bet is on the table, the player chooses between call $\{K,B,C,F\}$1 and fold $\{K,B,C,F\}$2. The game has a single betting round, and histories are sequences over $\{K,B,C,F\}$3 with the usual restrictions. One exposition labels four aggressive-action situations: first move; after one check; after bet and one fold/call by player 2; and after bet and call by player 2 (Ganzfried et al., 2022).

A terminal history ends when either two players have folded or when all remaining players have called or checked out. If $\{K,B,C,F\}$4 denotes the total pot and $\{K,B,C,F\}$5 the highest card among players who have not folded, then player $\{K,B,C,F\}$6’s utility is

$\{K,B,C,F\}$7

with the equivalent zero-sum normalization $\{K,B,C,F\}$8 (Ganzfried et al., 2022).

Later computational work reformulates the same game in sequence form. Each player’s strategy is represented by a realization-plan vector satisfying flow-conservation constraints,

$\{K,B,C,F\}$9

where the first row of each constraint matrix enforces that the empty sequence has probability 1 (Ganzfried, 24 Jun 2026).

2. Nash equilibrium and equilibrium multiplicity

A mixed-strategy profile $i\in\{1,2,3\}$0 is a Nash equilibrium if, for each player $i\in\{1,2,3\}$1,

$i\in\{1,2,3\}$2

For three-player Kuhn poker, this condition is already nontrivial despite the game’s small size (Ganzfried et al., 2022).

The standard four-card game admits an infinite family of equilibria parameterized by 21 real parameters, as noted in the literature summarized by Ganzfried and collaborators. In experimental work on opponent modeling, three “robust” representatives from this family were selected: $i\in\{1,2,3\}$3, a lower-bound equilibrium; $i\in\{1,2,3\}$4, an upper-bound equilibrium; and $i\in\{1,2,3\}$5, a midpoint (max-entropy) equilibrium (Ganzfried et al., 2022). A common misconception is that this game has a unique canonical equilibrium. The equilibrium family described above directly contradicts that view.

Generalized analyses with $i\in\{1,2,3\}$6 cards clarify why the $i\in\{1,2,3\}$7 case is special. In the $i\in\{1,2,3\}$8 deck, each player’s equilibrium strategy breaks cleanly into bluffing cards, calling cards, value-betting cards, and occasional sandbagging with the best card. This structure yields at most three coexisting equilibria for generic pot size $i\in\{1,2,3\}$9, with exactly a continuum of equilibria at four special pot sizes, including $c_i\in D$0. By contrast, for $c_i\in D$1 the bluffing, calling, and value-betting ranges split into more complicated intervals, and the equilibrium-solution curve becomes much richer (Billingham, 2018).

The same generalized study reports that $c_i\in D$2 shows up to five distinct equilibria in some pot ranges, $c_i\in D$3 can have up to 24 distinct coexisting equilibria near $c_i\in D$4, and $c_i\in D$5 shows up to 15 branches for $c_i\in D$6 up to $c_i\in D$7. When $c_i\in D$8, a new equilibrium feature appears: a mid-value card can be bet in a way that is a value bet against one opponent and a bluff against the other, a phenomenon absent in two-player and $c_i\in D$9 games (Billingham, 2018). This suggests that three-player Kuhn poker is not merely a minimal extension of the two-player game; the multiplayer interaction changes the equilibrium geometry itself.

3. Analytic simplified variants

One line of work studies “one-third street Kuhn poker,” in which Players 1 and 2 are forced to check, after which Player 3 may check or bet one unit; if Player 3 bets, Player 1 may fold or call, and then Player 2 may overcall or fold. Here the total pot is $i$0, with each player anteing $i$1, and the strategic frequencies are $i$2 (Billingham, 2017).

For the full one-third street game, the equilibrium classification depends sharply on $i$3. For $i$4, the unique equilibrium is trivial: no one bluffs or calls. For $i$5, the equilibrium satisfies $i$6, $i$7, and $i$8. For $i$9, where $c\in D$0, the interior solution satisfies

$c\in D$1

For $c\in D$2, there are two distinct one-dimensional solution sets, labeled Solution A and Solution B (Billingham, 2017).

The same paper introduces a simplified three-parameter subgame, SKP, obtained by imposing $c\in D$3, $c\in D$4, $c\in D$5, and $c\in D$6. The remaining frequencies are $c\in D$7, $c\in D$8, and $c\in D$9. In this model there is a unique equilibrium for $h$0,

$h$1

while for $h$2 there are exactly three distinct equilibria: $h$3 These three vectors satisfy the three best-response equations simultaneously and lie in $h$4 only for $h$5 (Billingham, 2017).

A separate “full street simplified” model imposes a different restriction: the weakest card, $h$6, is “dead,” meaning that any player holding $h$7 must always check and always fold. This reduces the number of nontrivial betting frequencies from 23 to 11, namely

$h$8

with $h$9. In this game there are three ranges of pot sizes for which three distinct coexisting equilibrium solutions occur: $i$0; $i$1, where $i$2 and $i$3; and $i$4, where $i$5 (Billingham, 2017).

4. Dynamic adjustment and repeated play

Three-player Kuhn poker has also been studied as a model of learning dynamics rather than only static equilibrium. In SKP, if one sets $i$6, $i$7, and $i$8, a continuous-time learning model with logistic saturation yields

$i$9

$I_{i,c,h}$0

$I_{i,c,h}$1

All three SKP equilibria are fixed points of this ODE system, but linearization shows that $I_{i,c,h}$2 and $I_{i,c,h}$3 are nonlinear centers with nested periodic orbits in invariant planes and a transverse stable direction, while $I_{i,c,h}$4 is a saddle with a two-dimensional stable manifold separating the two basins of attraction (Billingham, 2017).

The same analysis concludes that generic orbits converge not to any equilibrium but to one of two families of limit cycles, described as “bistable oscillators,” oscillating around $I_{i,c,h}$5 or $I_{i,c,h}$6. For $I_{i,c,h}$7 only the $I_{i,c,h}$8 cycle is stable; for $I_{i,c,h}$9 both coexist. In a discrete repeated-play model with memory lengths $(K)$0 and adjustment speeds $(K)$1, small $(K)$2 and moderate $(K)$3 produce trajectories that track the ODE cycles near $(K)$4, whereas larger $(K)$5 or larger $(K)$6 produce larger, noisier relaxation-oscillation cycles (Billingham, 2017).

The full-street simplified model produces an analogous message at larger state dimension. Its adaptive ODE system is

$(K)$7

where $(K)$8 ranges over the 11 strategic frequencies. Every static equilibrium is a fixed point, but none of the equilibrium solutions is asymptotically stable: some are unstable, while others are non-hyperbolic with center manifolds on which nearby orbits oscillate. Depending on $(K)$9, numerical integration yields periodic motion, motion close to periodic, or long chaotic transients. At the same time, the long-run profit slopes $\{K,B,C,F\}$00 closely match the static equilibrium payoff of the nearby equilibrium being shadowed (Billingham, 2017). A plausible implication is that, in this domain, equilibrium analysis predicts long-run profitability more robustly than it predicts pointwise convergence of strategies.

5. Exact equilibrium computation

Recent work formulates exact Nash equilibrium computation in three-player Kuhn poker as a nonlinear complementarity problem and then as a nonconvex quadratically constrained feasibility program. For each player $\{K,B,C,F\}$01, the KKT conditions for best response consist of stationarity, primal feasibility, dual feasibility, and complementary slackness. Stacking the three players’ systems yields an NLCP in the realization plans, multipliers, and slack variables, with nonlinearity arising from bilinear payoff terms such as $\{K,B,C,F\}$02, $\{K,B,C,F\}$03, and $\{K,B,C,F\}$04, together with complementarity terms such as $\{K,B,C,F\}$05 (Ganzfried, 24 Jun 2026).

To obtain a QCP, auxiliary bilinear variables are introduced: $\{K,B,C,F\}$06 and similarly for the complementarity products. This makes the stationarity equations linear in the expanded variable set, while preserving the quadratic equalities. Because all nonlinearities are bilinear, the formulation is amenable to spatial branch-and-bound with McCormick envelopes (Ganzfried, 24 Jun 2026).

A 2025 implementation on the reduced game after removal of dominated actions reports the following dominated actions as safely prunable: calling a bet with Jack; folding to a bet with Ace; calling a bet with Queen after a bet and a call; and checking with Ace after two consecutive checks. After pruning, the game tree shrinks from 288 decision nodes to 252, and total nodes go from 601 to 415, while the number of information sets per player remains 16. The resulting QCP has 72 linear rows, 249 columns, and 198 explicit bilinear constraints, and the global solve time reported for Gurobi is 2.47 seconds, with verified $\{K,B,C,F\}$07-Nash error approximately $\{K,B,C,F\}$08 (Ganzfried, 30 Sep 2025).

A 2026 refinement derives finite bounds on slack and multiplier variables. In three-player Kuhn poker, terminal payoffs lie in $\{K,B,C,F\}$09, which yields

$\{K,B,C,F\}$10

Injecting these bounds produces tighter McCormick envelopes. Reported solve times on an Intel i7-1065G7 are $\{K,B,C,F\}$11 hours with no bounds, $\{K,B,C,F\}$12 s with slack bounds only, $\{K,B,C,F\}$13 s with multiplier bounds only, and $\{K,B,C,F\}$14 s with both bounds. The same study states that the slack-bounds-only configuration yields both the fastest solve and the smallest branch-and-bound tree (Ganzfried, 24 Jun 2026). This suggests that, in multiplayer extensive-form equilibrium computation, bounding complementarity slack variables can matter more than tightening dual multipliers.

6. Opponent modeling and exploitative play

Three-player Kuhn poker also serves as a testbed for opponent modeling in repeated multiplayer imperfect-information games. In the Bayesian best-response approach of Ganzfried, Wang, and Chiswick, offline sampling begins from a prior mean $\{K,B,C,F\}$15 for each opponent $\{K,B,C,F\}$16 and each position $\{K,B,C,F\}$17; the midpoint equilibrium $\{K,B,C,F\}$18 is used for all opponent priors. Action probabilities are clipped into $\{K,B,C,F\}$19, $\{K,B,C,F\}$20 independent Dirichlet samples are drawn with parameters $\{K,B,C,F\}$21, and the initial prior weight over sample pairs is uniform, $\{K,B,C,F\}$22 (Ganzfried et al., 2022).

Online play uses a switch-over schedule. For $\{K,B,C,F\}$23, the agent plays the lower-bound equilibrium $\{K,B,C,F\}$24; afterward it computes posterior mean opponent models $\{K,B,C,F\}$25, forms a best response $\{K,B,C,F\}$26, and plays that strategy. After each hand, the posterior over opponent sample indices is updated by Bayes’ rule using the likelihood of the observed public actions and any private cards revealed at showdown (Ganzfried et al., 2022).

The evaluation includes ten “class-project” agents $\{K,B,C,F\}$27 and the three exact-equilibrium agents $\{K,B,C,F\}$28. For each unordered triple of agents, all 6 seating permutations are played; each match consists of 3,000 hands per seating, and the match is duplicated with the same random deals for variance reduction. Performance is measured by average win rate in millichips per hand,

$\{K,B,C,F\}$29

with reported standard errors approximately $\{K,B,C,F\}$30–$\{K,B,C,F\}$31 millichips per hand (Ganzfried et al., 2022).

The reported win-rate ordering is led by the Bayesian best-response agent, MBBR, at $\{K,B,C,F\}$32, followed by $\{K,B,C,F\}$33 at $\{K,B,C,F\}$34, $\{K,B,C,F\}$35 at $\{K,B,C,F\}$36, and then the class agents down to $\{K,B,C,F\}$37 at $\{K,B,C,F\}$38. The paper states that MBBR “strictly outperform[s] all exact Nash equilibria (best of which earned $\{K,B,C,F\}$39) and all class-project bots.” It also reports that replacing the equilibrium-based prior with a naive $\{K,B,C,F\}$40 uninformed prior causes MBBR to drop near the bottom of the field (Ganzfried et al., 2022). Within this experimental setting, equilibrium serves as a safe prior and early-game policy, while posterior exploitation captures additional value against weaker or systematically biased opponents.