A Note on Learnable Nash Equilibrium

Abstract: A Nash equilibrium is learnable if there exists a myopic adjustment dynamic for which it is asymptotically stable. In generic symmetric two-player games, a Nash equilibrium is learnable if and only if it has index +1.

• The paper demonstrates that Nash equilibria with index +1 are learnable via explicitly constructed myopic adjustment dynamics.
• It employs linear algebra and real Schur decomposition to link the equilibrium's index to its local asymptotic stability.
• The findings position learnability as a robust equilibrium refinement, effectively distinguishing between index +1 and index -1 equilibria.

Learnable Nash Equilibrium in Symmetric Games: Index and Local Stability

Introduction and Contextualization

This work systematically addresses the characterization of learnable Nash equilibria in symmetric two-player (single population) games, focusing on the index-theoretic structure of equilibria and its relationship with learnability as defined through myopic adjustment dynamics. Building on results from evolutionary game theory and index theory for Nash equilibria, the author revisits and completes a theorem originally stated (but not proven in full generality) by Hofbauer concerning the equivalence between learnability of Nash equilibria and the index +1 property in generic symmetric games. The paper provides a rigorous linear algebraic proof for the if-direction: that index +1 guarantees learnability via suitable myopic adjustment dynamics, filling a gap in the literature.

Formal Definitions and Theoretical Developments

The work precisely formulates myopic adjustment dynamics $f: \Delta \to T$, where $\Delta$ is the simplex of mixed strategies and $T$ is the corresponding tangent space. Such a dynamic must be Lipschitz continuous, pay-increasing, and vanish on Nash equilibria. An equilibrium $x$ is termed learnable if there exists such a dynamic rendering $x$ locally (asymptotically) stable.

The paper makes a critical distinction between evolutionary stability (which frequently fails to yield refinements in generic games) and learnability, which, while originating from evolutionary dynamics concepts, can yield much sharper equilibrium selection criteria in generic settings. It is shown that, in contrast with traditional refinements like perfect, proper, and stable equilibria, learnability eliminates approximately half the equilibria (those of index -1) due to index summation constraints. This is significant because the only generally valid refinement in generic games via traditional criteria does not discard any equilibria, whereas learnability provides discriminatory power.

Mathematical Structure and Proof Techniques

The core result establishes the equivalence between the existence of a myopic adjusting dynamic for which a Nash equilibrium is locally stable and the equilibrium having index +1. While the necessity direction (learnability implies index +1) is known [see Demichelis & Ritzberger, JET 2003], the sufficiency is not. The paper proves:

Theorem: For a generic symmetric two-player game, any Nash equilibrium of index +1 is learnable.

The argument first reduces the problem to equilibria with full support (by genericity/quasi-strictness), then applies a local linearization in the neighborhood of the equilibrium to relate the properties of the adjustment dynamic to the spectral properties of the linearization matrix. The proof is constructive:

• The tangent dynamics is recast from $x$ to $y$, the displacement from equilibrium.
• The payoff dynamics is mapped to $f(x)' A x = g(y)' B y$, where $B$ encodes the linearized payoff structure.
• The adjustment dynamic is constructed as $g(y) = M B y$ with a positive semidefinite $\Delta$0 so that the product $\Delta$1 is Hurwitz stable (all eigenvalues with negative real part), employing Schur decompositions and block diagonalization for the general case.
• In the $\Delta$2 case, this is made explicit via explicit matrix constructions using $\Delta$3, leveraging the structure of $\Delta$4 matrices.
• The general construction employs real Schur decomposition, grouping positive and negative eigenvalues, and assigns appropriate stabilizing multipliers for each block.

The proof thereby shows that for any Nash equilibrium of index +1, there exists an explicit matrix $\Delta$5 such that the dynamics satisfies the myopic adjustment property and produces local asymptotic stability.

Implications and Contrasts

The main implication is that, in generic symmetric games, local stability under some plausible class of payoff-increasing dynamics (not tied to a specific biological or reinforcement interpretation) robustly picks out equilibria with index +1. Since the index summation property is fundamental, this removes approximately half of the equilibria, contrasting sharply with perfect, proper, and other refinements, which are vacuous in generic cases.

Notably, the work situates learnability as a potentially more operative refinement for equilibrium selection in evolutionary and learning contexts than previously recognized. Moreover, it provides a pathway for future theoretical advances:

• The explicit construction of payoff-increasing local dynamics stabilizing a given equilibrium of specified index suggests further exploration of explicit learning rules operationalizing these stability properties.
• The approach generalizes to non-symmetric or more complex population structures, possibly extending to evolutionary stability under broader classes of adjustment dynamics.
• The work calls attention to the relationship between the geometric (index-theoretic) structure of equilibria and their dynamic realizability, hinting at deeper connections between dynamical systems theory and equilibrium analysis in economic and biological game theory.

Numerical and Empirical Observations

While the result is theoretical, it carries interpretive weight as it predicts (robust to perturbation) that in nearly all games, only half the equilibria are learnable; for instance, in the canonical coordination game, pure equilibria are learnable, while the mixed equilibrium is not. In higher dimensions, nearly all interior (fully mixed) equilibria of index +1 are rendered learnable by small, genericistic perturbations, while those of index -1 are not, reflecting a deeply rooted structural distinction.

Conclusion

This paper rigorously establishes the equivalence between the learnability of Nash equilibria via local myopic adjustment dynamics and the index +1 property in generic symmetric two-player games, providing constructive procedures for dynamics leading to local stability of such equilibria and thus operationalizing a powerful equilibrium refinement. The findings clarify the scope and implications of learnability as a selection criterion, underpinning its importance in both theoretical and applied settings, while opening new directions for the characterization of dynamically stable and learnable structures in game-theoretic models.

Reference: "A Note on Learnable Nash Equilibrium" (2606.22701)