Myerson Value Approach in Cooperative Games

• Myerson Value Approach is a framework that applies the Shapley value to graph-restricted games, ensuring fair allocation by incorporating connectivity and feasibility constraints.
• It extends to weighted and egalitarian variants, addressing player heterogeneity and solidarity, with applications in auctions, network games, and multi-agent systems.
• Algorithmic strategies such as dynamic programming and Monte Carlo sampling make it computationally tractable for complex and large-scale networks.

The Myerson value approach is a foundational framework in cooperative game theory and mechanism design, centered on allocating payoffs or optimizing outcomes in environments where network or graph constraints, or informational structure, play an essential role. It extends the Shapley value principle of fair attribution or optimal allocation by incorporating connectivity, feasibility, and probabilistic formation of coalitions, with rigorous axiomatic and algorithmic underpinnings. The concept has been generalized from classic communication games to network games with probabilistic formation, multi-agent systems explanation, graph signal attribution, weighted capacities, auction theory, and beyond.

1. Formal Definition and Key Properties

The classical Myerson value is defined for graph-restricted transferable-utility games, which model scenarios where cooperation among agents is governed by an explicit communication or network structure. Given a finite set of players $N$ and a characteristic function $v:2^N\to\mathbb{R}$ with $v(\emptyset)=0$, the Myerson value incorporates a network $g$ (or more generally a probability distribution over networks) that restricts feasible coalitions.

• For deterministic networks, the restricted game $\omega_{(v,g)}$ maps $S\subseteq N$ to $v(g|S)$: the total value generated by the subnetwork induced on $S$.
• The Myerson value $Y_i^m(v,g)$ for player $i$ is the Shapley value of this graph-restricted game:

$$Y_i^m(v,g) = \sum_{S\subseteq N\setminus\{i\}} \frac{|S|! (n - |S| - 1)!}{n!} [v(g|S \cup \{i\}) - v(g|S)]$$

• In the context of variable network games (where networks form stochastically per distribution $\pi$), the expected Myerson value $\Psi^m(v,\pi)$ is the $\pi$-expectation of $Y^m$:

$$\Psi^m_i(v,\pi) = \sum_{g \in \mathcal{G}^N} \pi(g) Y_i^m(v,g)$$

This is equivalent to applying the Shapley value to the expected restricted game $\omega_{(v,\pi)}(S) = \sum_g \pi(g) v(g|S)$ (Chakrabarti et al., 2021).

• The Myerson value satisfies several axioms:
  • Component balance: For any component $h$ of the extent of $\pi$, $\sum_{i \in N(h)} \Psi^m_i(v,\pi)$ equals expected wealth in $h$.
  • Equal bargaining power: The impact of removing a link is distributed equally across its endpoints.
  • Balanced contributions: The effect of removing a player is symmetric across all pairs.

Uniqueness of the Myerson value among allocation rules is guaranteed under these axioms for component-additive network games (Chakrabarti et al., 2021).

2. Generalizations: Weighted and Egalitarian Myerson Values

To accommodate heterogeneity in player capacity or solidarity preferences, several generalizations have been developed.

Weighted Myerson Value:

• Each player has a weight $w_i > 0$, quantifying link-formation capacity or importance.
• For every subnetwork $g'$ containing $i$, the weighted value apportions the Harsanyi dividend $A_{g'}(v)$ in the ratio $w_i / \sum_{j \in N(g')} w_j$:

$$\phi^w_i(g, v) = \sum_{g' \subseteq g,\, i \in N(g')} \frac{w_i}{\sum_{j \in N(g')} w_j} A_{g'}(v)$$

• Characterized by component balance plus weighted bargaining power; specializes to the standard Myerson value when $w_i$ are uniform (Kakoty et al., 2024).

Component-wise Egalitarian Myerson Value:

• Introduces a solidarity parameter $\lambda \in [0,1]$:

$$\Phi^{CEM}_i(v, g; \lambda) = \lambda \Phi^M_i(v, g) + (1-\lambda) \frac{v(C(i,g))}{|C(i,g)|}$$

• Interpolates between marginalist Myerson and pure egalitarian equal-split within components.
• Characterized via three different axiom sets, all adhering to component balance, additivity, and monotonicity criteria. Admits a noncooperative SPNE implementation via a bidding game (Borkotokey et al., 2022).

3. Applications in Network Games and Cooperative Multi-Agent Systems

The Myerson value methodology underpins diverse applications where cooperative benefits must be fairly and efficiently attributed among structurally constrained agents:

• Network partitioning and community detection: The Nash-stable partition concept leverages Myerson values to assign players to coalitions by comparing their payoffs across clusterings. This results in “natural” communities aligned with link structure and resolution parameters (Avrachenkov et al., 2017).
• Attribute and policy explainability in multi-agent systems: Exploiting a hierarchical knowledge graph encoding agent–attribute–policy dependencies, the Myerson decomposition allows efficient calculation of the global and local contributions. This yields dramatic reduction in computational cost (up to 66x faster than direct Shapley) while preserving fairness guarantees (Angelotti et al., 2022).
• Monte Carlo estimation for large networks: Given the intractable enumeration involved, unbiased approximation of the Myerson value uses hybrid schemes combining exact calculation over small/large coalitions and Monte Carlo sampling over intermediates. Hybrid estimators outperform pure permutation or connected-coalition sampling in both error and efficiency (Tarkowski et al., 2019).

4. Mechanism Design: Auctions and Virtual Values

The Myerson value approach is central to modern auction theory, where it underpins optimal and approximately optimal mechanism design:

• Single-parameter auctions: Myerson’s original result is that the revenue-maximizing auction allocates the item to maximize ironed virtual value $\phi(v)$ and charges a critical payment, with virtual value given (for single-dimensional continuous types) by $\phi(v) = v - (1 - F(v))/f(v)$ (Alaei et al., 2012, Giannakopoulos et al., 2024).
• General convex constraints: The framework extends via strong LP duality and KKT theory to multi-parameter and constrained settings, yielding generalized virtual values and associated payment identities. Under total unimodularity in feasibility constraints, the optimal auction can always be taken deterministic (Giannakopoulos et al., 2024).
• Multi-dimensional and non-linear preferences: Revenue-linearity is both necessary and sufficient for marginal-revenue (virtual value) maximization to yield full optimality. Where revenue-linearity fails, approximation guarantees can still be obtained (e.g., 4-approximation in unit-demand settings) (Alaei et al., 2012).

5. Extensions: Structure-Aware Attribution and Higher-Order Interactions

The framework has evolved to model not just player-wise contributions, but also structure-aware and interaction-level attributions:

• Myerson–Taylor interaction index: For graph-structured models, the Myerson–Taylor index generalizes the Shapley–Taylor interaction index by incorporating connectivity constraints, allowing attribution not just to nodes but to motifs and higher-order input interactions. The index is uniquely characterized by linearity, restricted null player, component efficiency, coalitional fairness, and interaction distribution axioms (Bui et al., 2024).
• Explanation of GNNs and black-box models: Application of the Myerson–Taylor interaction index to GNN explainability via the MAGE explainer enables extraction of motifs (connected subgraphs) most responsible for a model’s prediction, offering both rigor and alignment with human-interpretable patterns (Bui et al., 2024).

6. Algorithmic and Computational Strategies

Computing Myerson-type values faces severe combinatorial barriers:

• Exact computation for $n$ agents involves $O(2^n)$ coalitions, with further overhead for connectivity tests.
• Dynamic programming and memoization: By decomposing the game over the network’s connected components or via a hierarchical decomposition reflecting known agent dependencies, substantial computational savings can be realized (e.g., exponential reduction in the effective number of coalitions) (Angelotti et al., 2022).
• Monte Carlo hybrid methods: Exact computation for small/large coalitions and sampling for intermediate sizes provides optimal trade-offs; permutation-based estimators are unbiased with scalable variance bounds (Tarkowski et al., 2019).

7. Conceptual Significance and Theoretical Insights

The Myerson value approach unifies several fundamental themes in economic, network, and algorithmic theory:

• Fair division under structural constraints: It extends Shapley’s principle of average marginal contribution to environments where cooperation is mediated by explicit connectivity, stochastic network formation, or weighted bargaining capacities (Chakrabarti et al., 2021, Kakoty et al., 2024).
• Revenue optimization via marginal analysis: In mechanism design, it clarifies when revenue-optimal mechanisms can be interpreted as virtual welfare maximizers and underpins key revenue-equivalence and duality results (Alaei et al., 2012, Giannakopoulos et al., 2024, Zuo, 2017).
• Axiomatic uniqueness: Rigorous axiomatizations show that component balance, bargaining power symmetry, and (weighted) fairness are both necessary and sufficient to determine the value formula, robust to extensions such as weighting and solidarity parameters (Chakrabarti et al., 2021, Kakoty et al., 2024, Borkotokey et al., 2022).
• Algorithmic tractability: Decomposition, dynamic programming, and hybrid sampling are central to rendering these normative solutions computationally practical for large-scale networks and complex systems (Angelotti et al., 2022, Tarkowski et al., 2019).
• Interpretability and explainability: The Myerson framework is increasingly vital for attributing responsibility in structured machine learning models and for fair explanation of outcomes in AI systems (Angelotti et al., 2022, Bui et al., 2024).

The Myerson value approach thus provides a mathematically rigorous, axiomatized, and computationally principled methodology for fair division, optimal decision-making, and explainable allocation in networked multi-agent environments.