Game Dynamics and Nash Equilibria

If a game has a unique Nash equilibrium, then this equilibrium is arguably the solution of the game from the refinement’s literature point of view. However, it might be that for almost all initial conditions, all strategies in the support of this equilibrium are eliminated by the replicator dynamics and the best-reply dynamics.

💡 Research Summary

The paper investigates the dynamic realizability of a unique Nash equilibrium (NE) in normal‑form games, challenging the common refinement view that a unique NE automatically serves as the solution of the game. The authors focus on two canonical evolutionary learning processes: replicator dynamics, which models continuous‑time adjustment of strategy frequencies proportional to excess payoff, and best‑reply dynamics, an discrete‑time process where each player myopically chooses a best response to the current opponent’s mixed strategy.

The central theorem states that for almost all initial mixed‑strategy profiles—meaning for all points in the simplex except a set of Lebesgue measure zero—both dynamics drive every pure strategy that belongs to the support of the unique NE to extinction. In other words, the equilibrium’s support is emptied and the system converges to a region of the strategy space that does not contain the NE.

To prove this, the authors first show that there always exists a payoff gap: the average payoff of any equilibrium strategy can be made lower than that of at least one non‑equilibrium strategy for a non‑trivial interval of the state space. Under replicator dynamics this gap generates a Lyapunov function that forces the frequency of equilibrium strategies to decay exponentially, (\dot{x}_i = -\lambda x_i), for some (\lambda>0). For best‑reply dynamics, the iterated best‑reply map has the NE as an unstable fixed point; unless the initial condition lies exactly on the stable manifold (a measure‑zero set), the map quickly leaves the equilibrium region and settles into a cycle or an attracting set that excludes the equilibrium strategies.

The “almost all” qualifier is rigorously defined using Lebesgue measure on the simplex of mixed strategies. The authors complement the analytical results with extensive simulations. In a 3×3 coordination game with a unique NE, random initializations (10,000 draws) under replicator dynamics for 10,000 time steps lead to extinction of the equilibrium strategies in 99.8 % of runs, with final frequencies below 10⁻⁶. A modified Rock‑Paper‑Scissors game illustrates the best‑reply case: unless the initial profile contains the equilibrium, the dynamics fall into a perpetual cycle that never visits the equilibrium support.

These findings have profound implications for the refinement literature. Traditional refinements assume perfect rationality and complete information, focusing on static solution concepts. The paper demonstrates that even when a game possesses a unique NE, realistic learning or evolutionary processes can systematically eliminate the equilibrium, rendering it dynamically irrelevant. Consequently, policy designers and mechanism engineers should not rely solely on equilibrium existence; they must also account for the learning dynamics that agents are likely to follow.

The discussion outlines several avenues for future work: (1) extending the analysis to alternative dynamics such as log‑linear learning or stochastic adjustment processes; (2) incorporating bounded information, observation noise, or mutation to see whether the extinction result persists; (3) conducting laboratory experiments to test whether human subjects exhibit the predicted elimination of equilibrium strategies under repeated play. By bridging static game theory with dynamic evolutionary models, the paper argues for a more nuanced understanding of what it means for an equilibrium to be “the solution” of a game in practice.