Evolutionary Hypergame Dynamics

• Evolutionary hypergame dynamics is a framework where agents have access to different strategy subsets, resulting in asymmetric interactions and emergent behavior.
• It features distinct phases including cyclic dominance, absorption by a single strategy, and uncertainty, driven by local spatial clustering and finite-size effects.
• The framework extends to include introspection, fluctuating payoffs, and higher-order interactions, offering insights into strategic revival and cooperative shifts.

Searching arXiv for the cited papers to ground the article in current arXiv records. Evolutionary hypergame dynamics denotes evolutionary game dynamics in which players do not all share the same strategy set. Rather than assuming that every individual has full knowledge about and full access to the complete set of available strategies, the framework admits heterogeneous “strategy awareness/availability,” so that different agents can access different subsets of the strategic environment. In the canonical lattice formulation, this heterogeneity produces dynamical behavior that is qualitatively different from standard evolutionary games on networks, including single-strategy absorption phases, a cyclic competition (“rock-paper-scissors”) type of phase, an uncertain phase in which the dominant strategy adopted by the population is unpredictable, and the phenomenon of strategy revival (Jiang et al., 2018). Subsequent formulations broaden the concept by studying the learning and evolution of the strategy set itself, as well as higher-order interaction structures on hypergraphs, while related work on recurrence and fluctuating payoffs clarifies how cyclicity, metastability, and environmental variability reshape long-run evolutionary outcomes (Zhang et al., 29 Sep 2025, Civilini et al., 2021, Boone et al., 2019, Stollmeier et al., 2018).

1. Conceptual basis and scope

A common assumption employed in most previous works on evolutionary game dynamics is that every individual player has full knowledge about and full access to the complete set of available strategies. Hypergames relax that assumption by allowing players not to have an identical strategy set, thereby reflecting diversity in knowledge, experience, and background among individuals (Jiang et al., 2018). In this sense, the interaction may still be governed by a common underlying payoff structure, but each player’s perceived and feasible strategy space differs. The resulting asymmetry is not a payoff asymmetry in the ordinary sense; it is an asymmetry in access to options (Zhang et al., 29 Sep 2025).

Within the supplied literature, two related but distinct usages appear. In the lattice model of “Evolutionary hypergame dynamics,” the term refers to a population in which each agent can adopt only two out of three available strategies, and the heterogeneity of these two-strategy sets is the key ingredient that changes the evolutionary logic (Jiang et al., 2018). In “Evolutionary hypergame dynamics: Introspection reasoning and social learning,” the unit of selection can be the strategy set itself, so that strategy sets evolve on a slow timescale while strategies within a fixed set are revised on a fast timescale by introspection dynamics (Zhang et al., 29 Sep 2025). A broader interpretation also appears in the hypergraph model of group choice dilemmas, which is described as “best seen as an instance of evolutionary hypergame dynamics” because individuals are embedded in overlapping higher-order groups of varying sizes rather than only pairwise interactions (Civilini et al., 2021).

This suggests that evolutionary hypergame dynamics is not a single model but a family of models unified by heterogeneous strategic opportunity. The supplied papers consistently emphasize that such heterogeneity can generate asymmetric strategic interactions, alter effective dominance relations, and produce macroscopic outcomes that are absent from standard evolutionary game dynamics (Jiang et al., 2018, Zhang et al., 29 Sep 2025).

2. Canonical lattice formulation

The prototypical model employs a generalized prisoner’s dilemma game with three available strategies: cooperation $C$, defection $D$, and loneliness or non-participation $L$. The payoff matrix is

$$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$

where $b>1$ is the temptation to defect and $\sigma$ is the payoff of loneliness. In the simulations, $b=1.02$, $\sigma=0.25$, and the noise is $\kappa=0.1$ (Jiang et al., 2018).

The defining hypergame assumption is that each agent can adopt only two of the three strategies, with probabilities that sum to $1$. There are exactly three player types: $D$0 where the rows correspond to $D$1, and $D$2 is the key control or bifurcation parameter (Jiang et al., 2018). Thus $D$3 mixes $D$4 and $D$5, $D$6 mixes $D$7 and $D$8, and $D$9 mixes $L$0 and $L$1.

The population lives on a square lattice with periodic boundary conditions, and each agent interacts only with its nearest neighbors. For node $L$2, the total payoff is

$L$3

where $L$4 is the pairwise payoff from interaction between nodes $L$5 and $L$6. Strategy update occurs via a Fermi imitation rule,

$L$7

This is standard in evolutionary game theory, but here it operates on the mixed-strategy hypergame types $L$8 rather than on a common strategy set (Jiang et al., 2018).

The principal observable is the equilibrium frequency of each type: $L$9 with indicator $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$0 if agent $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$1 is of type $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$2, and $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$3 otherwise (Jiang et al., 2018).

3. Phase structure, cyclic dominance, and uncertainty

As $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$4 is varied, the lattice hypergame exhibits several qualitatively distinct phases. Near the extremes $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$5 or $$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$6, the mixed strategies become close to pure strategies, and the system behaves more like a conventional evolutionary game on a lattice. In this regime, one strategy dominates, the other two disappear early, and the final state is effectively a two-type coexistence near the boundary, or absorption by a single type depending on the side of the parameter range. The left and right endpoints are related by the limits

$$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$7

which explains the symmetry between the small-$$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$8 and large-$$M= \left(\begin{array}{ccc} & C & D & L\ C & 1 & 0 & 0\ D & b & 0 & 0\ L & \sigma & \sigma & \sigma \end{array}\right),$$9 behavior (Jiang et al., 2018).

For intermediate $b>1$0, roughly

$b>1$1

all three types coexist, and the interaction pattern becomes cyclic, like rock-paper-scissors. The key pairwise dominance relations are

$b>1$2

The paper emphasizes that this cyclic relation is not imposed externally; it emerges from the payoff structure and local spatial clustering. Spatially, the lattice organizes into large domains of different strategy types with active interfaces, and strategy changes happen mainly on boundaries between clusters, producing a self-organized cyclic competition (Jiang et al., 2018).

For approximately

$b>1$3

the system enters the uncertain phase. In this regime, the system is still dominated by competition among all three types for a long time, but eventually, because of fluctuation-driven extinction, the final state becomes one of two possible single-strategy absorptions: either $b>1$4-absorption or $b>1$5-absorption. The dominant final strategy is unpredictable from a single realization. The stated mechanism is that the system spends a long transient time in an RPS-like state, but one of the weak strategies may become so rare that finite-size fluctuations eliminate it. Once one type goes extinct, the cyclic balance breaks and the remaining interactions drive the system into one of the absorbing states. The transition boundary between the cyclic region and the uncertain phase shifts with lattice size, and the final outcome depends on which rare strategy dies first (Jiang et al., 2018).

For roughly

$b>1$6

the system becomes dominated by $b>1$7, and the final state is essentially all $b>1$8. Near $b>1$9, the system eventually transitions toward $\sigma$0-dominance again, consistent with the symmetry relation between the ends of the parameter interval (Jiang et al., 2018).

A related but distinct cyclicity appears in the replicator-dynamics literature. In one-player zero-sum games under replicator dynamics, interior orbits are Poincaré recurrent if and only if there exists an interior Nash equilibrium, and in dimension $\sigma$1, if a one-player zero-sum game has an interior Nash equilibrium, then every interior point belongs to a periodic orbit (Boone et al., 2019). This does not describe the lattice hypergame directly, but it clarifies that cyclic or recurrent behavior in evolutionary systems can arise from structurally different mechanisms: in zero-sum replicator dynamics through volume preservation, recurrence, and low-dimensional geometry; in the lattice hypergame through local clustering, heterogeneous strategy access, and interface invasion (Jiang et al., 2018, Boone et al., 2019).

4. Strategy revival and extinction ordering

One of the striking findings of the lattice model is strategy revival. In the stated sense, strategy revival means that a strategy’s frequency becomes extremely small, even nearly zero, yet later it reappears and eventually dominates the entire population (Jiang et al., 2018). The supplied example concerns region 3, where $\sigma$2 can become almost extinct, but later take over the lattice.

The mechanism is explicitly spatial. The population first forms clusters. A weak $\sigma$3 cluster may collapse. This can isolate a tiny $\sigma$4 cluster inside a sea of $\sigma$5. Since $\sigma$6 has a local advantage over $\sigma$7, that tiny remnant can expand. Once $\sigma$8 is gone, $\sigma$9 can eliminate all remaining $b=1.02$0 and dominate. Conversely, if $b=1.02$1 disappears first, then $b=1.02$2 wins. Revival is therefore tied to the order of extinction among the weak types (Jiang et al., 2018).

The article’s explanation distinguishes near-extinction from permanent extinction. In ordinary evolutionary games with a common strategy set, extinction usually means permanent loss. Here, because strategy availability is heterogeneous and the local dominance relations are cyclic, a near-extinct type can persist in a spatial refuge and later exploit changes in neighborhood composition. The paper also notes that this revival persists even on small-world rewired versions of the lattice, not just the pure square lattice (Jiang et al., 2018).

A useful comparison comes from the zero-sum replicator setting. There, if no interior Nash equilibrium exists, then all interior initial conditions converge to the boundary, and strategies not in the support of any equilibrium vanish in the limit of all orbits (Boone et al., 2019). In that framework, collapse to the boundary is a theorem-level asymptotic classification. In the lattice hypergame, by contrast, long transients, finite-size extinction events, and stochastic revival effects govern the uncertain phase. This suggests that the fate of a rare strategy in evolutionary hypergame dynamics cannot be inferred solely from static support conditions; local refuges and extinction ordering matter (Jiang et al., 2018, Boone et al., 2019).

5. Analytical treatments: pair approximation, mean-field closure, and long-run behavior

To analyze the lattice hypergame, the paper uses pair approximation focusing on interactions at cluster boundaries. The boundary between two strategy clusters is modeled on an 8-node motif with 2 focal nodes in the center and 6 neighbors around them. The assumption is that strategic changes occur mostly at cluster interfaces, and boundaries are usually between only two strategy types. There are six regular boundary configurations and three irregular ones, making nine cases total. For two strategies $b=1.02$3 and $b=1.02$4, the focal payoffs in each configuration are denoted $b=1.02$5 and $b=1.02$6, $b=1.02$7, and the average payoffs are

$b=1.02$8

This yields the approximate pairwise replacement probability

$b=1.02$9

It is interpreted as an approximation to the actual strategy transformation probability on the lattice (Jiang et al., 2018).

For $\sigma=0.25$0, the pair approximation finds

$\sigma=0.25$1

which reproduces the replacement direction $\sigma=0.25$2 beats $\sigma=0.25$3, $\sigma=0.25$4 beats $\sigma=0.25$5, and $\sigma=0.25$6 beats $\sigma=0.25$7, exactly the RPS cycle. For $\sigma=0.25$8, the approximation shows a reversal: $\sigma=0.25$9, so $\kappa=0.1$0 beats $\kappa=0.1$1, while $\kappa=0.1$2, so $\kappa=0.1$3 also eliminates $\kappa=0.1$4. Thus $\kappa=0.1$5 becomes the dominant absorbing strategy (Jiang et al., 2018).

The paper also introduces a mean-field theory for the global frequencies $\kappa=0.1$6: $\kappa=0.1$7

$\kappa=0.1$8

with

$\kappa=0.1$9

The stated interpretation is that growth of each type comes from pairwise replacement against its competitors, the terms $1$0 represent encounter frequencies under mean-field mixing, and the replacement probabilities come from the pair approximation (Jiang et al., 2018).

The authors state that the mean-field theory matches simulation well in regions 1, 2, and 4, and reproduces the main phase boundaries where the pairwise invasion probabilities cross $1$1. It does not fully capture region 3, because that region is governed by very long transient times, finite-size extinction events, and stochastic revival effects. Region 3 is therefore beyond simple mean-field closure (Jiang et al., 2018).

For continuous-time replicator dynamics in zero-sum games, a different analytical structure applies. If an interior Nash equilibrium exists, then the sum of KL divergences to that equilibrium is constant along trajectories, which implies boundedness away from the boundary; combined with volume preservation, Poincaré recurrence follows (Boone et al., 2019). This is not the method used in the lattice hypergame, but it provides a mathematically sharp contrast between conservative recurrent dynamics and fluctuation-driven metastable coexistence.

6. Extensions: introspection, fluctuating payoffs, and higher-order interaction structure

The 2025 paper extends the hypergame framework by combining heterogeneous access to strategy sets with introspection reasoning and social learning. For the three-strategy system $1$2, it considers both fixed-size two-strategy sets

$1$3

and the full collection of all nonempty subsets

$1$4

At the fast timescale, players revise strategies by introspection dynamics with switching probability

$1$5

where $1$6 is the introspection strength. At the slow timescale, the objects that evolve are the strategy sets themselves. In the well-mixed case this is modeled by replicator dynamics, while in the spatial version players live on a $1$7 square lattice with periodic boundary conditions and adopt a neighbor’s strategy set with probability

$1$8

The reported phases include loner dominance, coexistence of multiple strategy sets, and cooperation-loner dominance by $1$9. The key conceptual result is that heightened rationality significantly promotes cooperative behaviors, and the paper’s summary states that stronger introspection favors cooperation (Zhang et al., 29 Sep 2025).

A central analytical example is $D$00 against itself. The paper derives a stationary distribution over the states $D$01, $D$02, $D$03, and $D$04, and the important qualitative result is that $D$05 increases with $D$06, while as $D$07, $D$08. Thus, with strong introspection, two $D$09 players are increasingly likely to settle into mutual cooperation rather than drifting to loner behavior. Against pure defectors $D$10, the expected defector payoff against $D$11 decreases monotonically with $D$12, and the paper derives a threshold in $D$13 beyond which $D$14 beats $D$15 (Zhang et al., 29 Sep 2025).

Another extension concerns fluctuating payoffs. The paper on “Unfair and Anomalous Evolutionary Dynamics from Fluctuating Payoffs” shows that the usual practice of replacing a time-varying game by its arithmetic average can be fundamentally wrong because payoff noise is multiplicative, so long-run outcomes are governed by geometric rather than arithmetic averaging. The effective stationary condition under fluctuating payoffs is

$D$16

equivalently $D$17, and for small fluctuations

$D$18

The paper explicitly describes the resulting behavior as “hypergame-like behavior”: the effective strategic environment experienced by the population differs qualitatively from the nominal static game. It further states that temporal variability reshapes strategic success, coexistence can become unfair, equilibria can shift, split, or disappear, and selection can reverse direction (Stollmeier et al., 2018).

Higher-order interaction structure supplies a further extension. In the hypergraph model of group choice dilemmas, the population is represented by a hypergraph with an incidence matrix $D$19, hyperdegree $D$20, and group size $D$21. The imitation probability under weak selection is

$D$22

and the group-level coarse graining yields a replicator-like equation

$D$23

When the co-membership distribution satisfies $D$24 with $D$25, numerical quasi-stationary simulations agree with the mean-field prediction and the system converges to the Nash equilibrium $D$26. When $D$27 with $D$28, the second moment diverges, hubs become highly influential, and the simulations show a phase transition from the absorbing all-safe state to a nontrivial QS state that then rapidly moves toward widespread risky adoption. The model is described as a mechanism for irrational herding and radical behavior in social groups (Civilini et al., 2021).

Taken together, these extensions indicate that evolutionary hypergame dynamics is a flexible framework for studying how heterogeneous access to strategies, endogenous learning within restricted strategy sets, fluctuating environments, and higher-order social structure alter evolutionary outcomes. A plausible implication is that the central hypergame effect is not confined to one topology or one update rule; it persists whenever heterogeneity changes the feasible strategic environment experienced by agents (Jiang et al., 2018, Zhang et al., 29 Sep 2025, Stollmeier et al., 2018, Civilini et al., 2021).