Heterogeneous Framework of Kim & Kim

• The paper establishes a rigorous framework for evolutionary game dynamics in heterogeneous populations by capturing full type–strategy correlations and ensuring unique equilibrium.
• It employs Poisson revision protocols and Lipschitz continuity conditions to guarantee the well-posedness and convergence of the dynamic system.
• The model generalizes classical replicator and best-response dynamics by incorporating heterogeneous payoff functions and flexible strategy revision processes.

The heterogeneous framework of Kim and Kim provides a mathematically rigorous foundation for evolutionary game dynamics in populations characterized by a continuum of types, allowing for heterogeneity both in agents’ payoff functions and in the revision protocols dictating behavioral switching. The construction captures the entire joint distribution of strategies and types, permitting full retention of type–strategy correlations and generalizing foundational results on equilibrium existence, stationarity, and Lyapunov stability from homogeneous to heterogeneous settings (Zusai, 2018).

1. Model Structure and Primitives

The population is formalized as the unit interval $\Omega = [0,1]$ equipped with Lebesgue measure $m_\Omega$. Each agent $\omega \in \Omega$ is assigned a type $\theta(\omega) \in \Theta$, where $\Theta$ is a complete, separable metric space with probability measure $\mu_\Theta \in \mathcal{P}(\Theta)$. Agents choose strategies from a finite set $\mathcal{A} = \{1,\ldots,S\}$, resulting in a measurable strategy–type profile $(s(\cdot), \theta(\cdot))$.

The fundamental object is the joint distribution $X = (X_s)_{s \in \mathcal{A}}$ on $\Theta \times \mathcal{A}$, determined by

$$X_s(B) = m_\Omega(\{\omega: s(\omega) = s \text{ and } \theta(\omega) \in B\}),\qquad B \text{ Borel in } \Theta,$$

which satisfies $\sum_s X_s(B) = \mu_\Theta(B)$. This encodes the entire mass of type–strategy combinations in the population.

The disintegration of $X$ yields a conditional strategy profile $x(\theta) \in \Delta^S$, the simplex of mixed strategies on $\mathcal{A}$, such that $X = \int_\Theta x(\theta) d\mu_\Theta(\theta)$ and $x_s(\theta)$ is the probability a type-$\theta$ agent plays $s$.

Payoff functions are specified by a population game map $$F: \mathcal{P}(\Theta \times \mathcal{A}) \to C(\Theta, \mathbb{R}^S)$$, assigning, for each $\theta \in \Theta$ and $s \in \mathcal{A}$, the vector $F_s[X](\theta)$—the payoff to a type-$\theta$ agent playing $s$ in the population profile $X$.

Bayesian–Nash equilibrium for this setting is a joint profile $X^*$ such that for $\mu_\Theta$-almost every $\theta$, the conditional strategy $x^*(\theta)$ places full probability on the best-response set to $F[X^*](\theta)$:

$$x^*(\theta) \in \Delta(\mathrm{BR}(F[X^*](\theta))),\qquad \text{a.e. } \theta,$$

with $\mathrm{BR}(\pi) = \{s: \pi_s = \max_j \pi_j\}$.

Revision protocols (agents’ strategy revision processes) are modeled as Poisson processes with instantaneous switching rates $\rho_{ss'}(\pi) \geq 0$, uniform across types except in extensions. Two principal classes are considered: (i) $L$-continuous (Lipschitz in $\pi$), and (ii) exact-optimization (switching only to best responses).

2. Evolutionary Dynamics Formulation

The core mean-field dynamic describes, for each $\theta$, the evolution of the conditional strategy profile $x(\theta)$: \begin{align*} \dot{x}s(\theta) &= \sum_r x_r(\theta) \rho{r s}(FX) - x_s(\theta) \sum_r \rho_{s r}(FX), \quad \mu_\Theta\text{-a.e.}\; \theta,\ \dot{x}(\theta) &= v(x(\theta), FX), \end{align*} where $v$ is the drift of the conditional strategy profile.

The dynamic on the joint population measure is

$$\dot{X}_s(B) = \int_B \dot{x}_s(\theta) d\mu_\Theta(\theta),\quad B \in \text{Borel}(\Theta),$$

which is compressed into the vector field notation $\dot{X} = V(X)$, mapping measures to measures.

This general form encompasses myriad population dynamics. For $|\Theta|=1$ or payoff independence from $\theta$, it reduces to standard homogeneous replicator, logit, Smith, or best-response dynamics.

3. Regularity and Well-Posedness

Existence and uniqueness of solution trajectories are ensured by regularity conditions enabling the application of the Picard–Lindelöf theorem in the Banach space of signed measures:

• (A1) $F$ is Lipschitz in $X$ (with respect to total variation norm $\|\cdot\|_{TV}$), uniformly in $\theta$.
• (A2) $\rho_{ss'}(\pi)$ is uniformly bounded in $\pi$.
• (A3) In the exact-optimization protocol class, the measure of types at which the best-response set changes discontinuously depends in a Lipschitz way on $X$.

Under these assumptions, $V$ is globally Lipschitz on a convex subset of signed measures: the system thus admits a unique, global solution $X(t)$ from every initial condition (Zusai, 2018).

4. Equilibrium Characterization and Stationarity

Distributional equilibrium under this framework exists by Glicksberg’s generalization of Kakutani’s fixed-point theorem. Specifically, if $F$ is continuous and (essentially) bounded, the correspondence

$$X \mapsto \arg\max_Y \int_\Theta F[X](\theta) \cdot y(\theta) d\mu_\Theta$$

is upper-hemicontinuous, convex-valued, and defined on the compact convex set $\mathcal{P}(\Theta\times\mathcal{A})$.

Stationarity of the dynamic aligns with the Bayesian–Nash equilibrium: $V(X) = 0$ if and only if $x(\theta)$ almost surely places all probability on $\mathrm{BR}(F[X](\theta))$:

$$V(X) = 0 \iff X \text{ is a Bayesian–Nash equilibrium}.$$

Thus, the invariant points of the mean-field dynamic correspond exactly to the generalized equilibrium notions.

5. Stability and Potential Games

A heterogeneous potential game admits a potential function $$f : \mathcal{P}(\Theta \times \mathcal{A}) \to \mathbb{R}$$ that is Fréchet-differentiable and weakly continuous, satisfying:

$$f(X+\Delta) = f(X) + \langle F[X], \Delta \rangle + o(\|\Delta\|_{TV}).$$

A dynamic possesses the positive correlation (PC) property if

$\pi \cdot v(\pi, x) \geq 0,$

with equality only when $v(\pi, x) = 0$.

The potential $f$ serves as a Lyapunov function for the dynamic:

$$\frac{d}{dt} f(X(t)) = \int_\Theta F[X](\theta) \cdot \dot{x}(t,\theta) d\mu_\Theta(\theta) = \int_\Theta \pi(\theta) \cdot v(\pi(\theta), x(\theta)) d\mu_\Theta \geq 0,$$

with strict inequality away from equilibrium. Cheung’s result (LaSalle’s theorem for the weak topology) implies:

• The set of equilibria is globally attracting.
• Any strict local maximizer of $f$ is asymptotically stable.
• Any isolated asymptotically stable equilibrium $X^*$ is a local strict maximizer of $f$.

This generalizes classical Lyapunov results for potential games to heterogeneous settings.

6. Representative Proof Structures

Several technical arguments underlie the framework:

• Lipschitz continuity of the dynamic: The drift $V$ is an integral of $\rho \circ F$ against $x$; under (A1) and Lipschitz $\rho$, small total-variation deviations in $X$ lead to uniformly small drifts in the measure space. For exact-optimization protocols, (A3) ensures the magnitude of best-response switching is controlled.
• Stationarity equivalence: The definition of equilibrium (maximum weight on best responses) exactly matches the zero drift condition in the dynamic. Thus, stationarity and equilibrium coincide.
• Lyapunov stability in potential games: The Fréchet-differentiability of $f$ yields a first-order expansion, and positive correlation ensures strict increase along trajectories unless at equilibrium, confirming Lyapunov stability via weak convergence arguments.

7. Examples and Relation to Homogeneous Dynamics

The framework abstracts and generalizes a broad range of population games and evolutionary processes:

• Additively separable aggregate games (ASAG): For

$$F_s[X](\theta) = F^0_s(\bar{x}) + \theta_s,\quad \bar{x}_s = \int_\Theta x_s(\theta) d\mu_\Theta,$$

if $F^0$ is a classical potential game, the extended game possesses potential

$$f(X) = f^0(\bar{x}) + \int_\Theta \theta \cdot x(\theta) d\mu_\Theta.$$

• Random matching (incomplete information): Agents, matched by $\mu_\Theta$, play payoff $u_{s,s'}(\theta,\theta')$, inducing

$$F_s[X](\theta) = \int_\Theta \sum_{s'} u_{s,s'}(\theta, \theta') X_{s'}(d\theta').$$

If $u$ is a two-player potential, the game admits a heterogeneous potential.

• Structured populations: For a continuum of sub-populations, each $\theta$ plays a two-population sub-game $F^0(\cdot, \cdot)$, with weights $g(\theta, \theta')$. Symmetry in $g$ implies existence of a potential via a double integral.
• Classical dynamics: Replicator, best-response (BNN), and imitation dynamics arise as special cases through the appropriate choice of $\rho$ in the main equation. When $\Theta$ is a singleton, the model collapses to standard homogeneous population dynamics on the simplex $\Delta^S$.

A plausible implication is that the framework enables rigorous analysis of large-scale strategic interaction in settings where population-level heterogeneity cannot be ignored, subsuming and extending classical results for homogeneous models (Zusai, 2018).