Hedonic Games: Coalition Formation & Stability

• Hedonic games are a preference-based coalition formation model where agents' satisfaction is solely determined by the makeup of their own coalition.
• The approach defines key stability concepts such as core, Nash, and individual stability, and uses compact representations to manage computational complexity.
• It employs algorithmic, empirical, and PAC-learning methods to assess coalition stability, welfare optimization, and typical-case performance in multi-agent systems.

A hedonic games approach is a formal, preference-oriented methodology for modeling coalition formation among agents whose satisfaction depends solely on the composition of their own coalition. The approach is grounded in the specification, analysis, and computation of coalition structures based on agents' preferences over possible groups, with a rich taxonomy of stability, efficiency, and optimality concepts parameterizing the coalition formation landscape. This article surveys core formal definitions, representation techniques, algorithmic and complexity results, empirical and statistical methodologies, and contemporary applications that collectively constitute the hedonic games approach.

1. Formal Framework and Stability Concepts

Let $N = \{1,2,\dots,n\}$ denote a finite set of players (agents). A coalition is any nonempty $S \subseteq N$; a coalition structure (partition) $\pi = \{C_1, \dots, C_k\}$ is a set of disjoint coalitions whose union is $N$. In the hedonic games framework, each player $i$ has a preference relation $\succsim_i$ over the set $\mathcal{N}_i = \{S \subseteq N : i \in S\}$—that is, over all coalitions containing $i$.

The defining feature of a hedonic game is hedonicity: agent $i$'s satisfaction is fully determined by her own coalition, regardless of how the rest of $N \setminus \pi(i)$ is partitioned. Formally, a hedonic game is a pair $(N, \{\succsim_i\}_{i\in N})$.

Key stability concepts include:

• Core Stability: A partition $\pi$ is core-stable if there does not exist any blocking coalition $S \subseteq N$, $S \neq \emptyset$, such that $S \succ_i \pi(i)$ for all $i\in S$. The set of such partitions is called the core (Collins et al., 2020).
• Nash Stability: No individual player can profitably deviate to another (possibly empty) coalition (Aziz et al., 2011); formally, for all $i$ and $C\in\pi\cup\{\emptyset\}$, $C\cup\{i\} \not\succ_i \pi(i)$.
• Individual Stability (IS): No player can profitably deviate to another coalition if all members of the destination coalition weakly prefer the post-move grouping (Aziz et al., 2011).
• Contractual Nash Stability (CNS) and Contractual Individual Stability (CIS): Further restrict deviations to require non-harm to both departing and receiving coalitions (Gairing et al., 2015, Caskurlu et al., 2020).

The range of stability concepts allows the approach to model settings with varying rules for coalition entry/exit, and their combinations have well-defined implications for the existence and computation of stable partitions.

2. Representation and Structural Restriction

Since the number of possible coalitions is $2^n-1$ per agent, direct preference enumeration is intractable. The approach has developed several compact representation languages:

• Additively Separable Games: Each agent $i$ assigns a real value $v_i(j)$ to every other agent $j$, and values a coalition as $u_i(S) = \sum_{j\in S \setminus \{i\}} v_i(j)$ (Gairing et al., 2015).
• Fractional Hedonic Games (FHGs): $u_i(S) = \frac{1}{|S|} \sum_{j\in S} v_i(j)$, often with symmetric or unweighted $v_i(j)$ (Hanaka et al., 2023, Demeulemeester et al., 2024).
• Boolean Hedonic Games: Preferences specified via propositional (Boolean) goals over desired coalition properties (Aziz et al., 2015).
• Common Ranking Property (HGCRP): All agents share a coalition valuation function $U(S)$ (Caskurlu et al., 2022, Caskurlu et al., 2020).
• Subset-Additive and Subset-Neutral Games: Player preferences built from weights assigned to subgroups of their coalition, achieving full expressive power (Suksompong, 2018).
• Distance-Based (FEN) Games: Bipolar "friend/enemy" local ballots mapped to distance metrics (Hausdorff–Kendall-tau) over coalitions (Rey et al., 2022).
• Metric Hedonic Games: Agent types in metric space (e.g., on the line), with coalition costs defined by distances among members (Haye et al., 5 Feb 2026).

Structural restrictions (tree-like graphs, bounded coalition-size, bounded treewidth, block graphs, etc.) are crucial for algorithmic tractability (Hanaka et al., 2023, Igarashi et al., 2016).

3. Algorithmic Techniques and Complexity

The typical workflow for the hedonic games approach involves:

• Preference Representation using one of the above models (e.g., fixing $v_i(j)$ values or propositional goals).
• Stability and Welfare Computation: Checking for core, Nash, or individual stability, or maximizing utilitarian/egalitarian welfare. Key algorithmic strategies include:
  • Brute-force enumeration: For small $n$, enumerate all partitions (Bell number $B_n$) (Collins et al., 2020, Miles, 2017).
  • Dynamic programming on restricted graphs: For FHGs or ASHGs on bounded treewidth or block graphs, DP tracks partial coalition statistics (Hanaka et al., 2023).
  • Potential function methods: Many classes admit an exact or ordinal potential, ensuring that local improvement dynamics (best-response, swaps, or deviation-restricted moves) converge to stable outcomes (Gairing et al., 2015, Caskurlu et al., 2020).
  • Greedy and combinatorial algorithms: For HGCRP and neutrally anonymous games, greedy coalition selection by joint utility or coalition size suffices (Caskurlu et al., 2022, Suksompong, 2018).
  • Monte Carlo methods: Statistical approaches sample random hedonic instances and compute empirical distribution of core sizes or stable partitions (Collins et al., 2020, Bullinger et al., 2024).
• Complexity Landscape: The approach reveals rich computational distinctions:
  • Core nonemptiness and stability (NP- or $\Sigma_2^P$-complete in general), polynomial time for subclasses (e.g., neutrally anonymous, tree/graph-restricted, two-player coalition size) (Suksompong, 2018, Igarashi et al., 2016, Caskurlu et al., 2022).
  • Nash and IS existence is NP-hard for unstructured preferences, efficient in subclass (e.g., B-hedonic games, block graphs, anonymous games under single-peakedness, etc.) (Aziz et al., 2011, Hanaka et al., 2023, Bredereck et al., 2019).
  • PLS-completeness for various local-stability search problems, even under strong symmetry (Gairing et al., 2015, Igarashi et al., 2016).

4. Empirical, Probabilistic, and PAC-Learning Advances

Numerical and probabilistic methods expand the hedonic games approach to data-driven, "typical-case" and learning scenarios:

• Monte Carlo Analysis: Generate large samples of random (e.g., strict-preference) hedonic games to empirically estimate the distribution of core sizes and stability likelihood for increasing $n$ (up to $n=13$) (Collins et al., 2020). Key findings show that the fraction of random games with empty core increases with $n$, while the core-size distribution becomes increasingly right-skewed and can be well-fit by Weibull or Gamma distributions.
• $\varepsilon$-Fractional Core: Definition allows a small $\epsilon$ fraction of blocking coalitions, guaranteeing, via explicit construction, the existence and polynomial-time computability of approximately core-stable partitions for anonymous and simple fractional games (Fioravanti et al., 2023).
• Random Utility Models: High-probability results show, for instance, that in large random additively separable games, IS and CIS partitions exist w.h.p., but Nash stable partitions almost surely do not, pinpointing the mismatch between worst-case and typical-case theory (Bullinger et al., 2024).
• PAC Learning/Stabilizability: The approach leverages machine learning theory to quantify when agent preferences (or core-stable partitions) can be inferred and stabilized efficiently from samples. Efficient PAC learnability follows if the preference class has bounded pseudo-dimension, and stabilizability requires a sample-resistant core property. Also, explicit positive results for "bottom responsive" ($\mathcal{W}$-hedonic) games, and negative results for more expressive classes (Fioravanti et al., 2023).

5. Extensions: Welfare Maximization, Relaxations, and Applications

The hedonic games approach supports several advanced generalizations and applications:

• Welfare Maximization: For FHGs and related models, algorithmic work has addressed utilitarian and egalitarian coalition partition maximization using DP on tree-like structures and vertex-cover decompositions, with precise complexity boundaries (e.g., pseudo-polynomial for bounded treewidth) (Hanaka et al., 2023).
• Relaxed Core Notions: Unified frameworks (e.g., $\alpha$-hedonic games) relate small-blocking-coalition stability to approximate group-improvement stability across coalition sizes, yielding explicit bounds for classical classes and resolving open conjectures about core price of anarchy (Demeulemeester et al., 2024).
• Modeling Diversity and Metrics: Hedonic diversity games, metric games, and type-based models address coalition composition from diversity, expertise, or spatial proximity, and produce novel algorithmic, stability, and price-of-anarchy results (Bredereck et al., 2019, Caskurlu et al., 2020, Haye et al., 5 Feb 2026).
• Incentive Mechanisms and Federated Learning: The approach has shaped incentive-compatible, Nash-stable coalition formation protocols for decentralized learning and clustering, characterizing when potential game dynamics yield decentralized, stable clustering (Hasan, 2021).
• Simulators and Empirical Tools: Dedicated simulators exist to empirically validate theoretical results, test preference classes, and educate new researchers via hands-on coalition stability analysis (Miles, 2017).

6. Methodological Impact and Open Research Directions

The hedonic games approach systematically bridges the gap between social choice theory, combinatorial optimization, algorithm design, and empirical coalition analysis. Distinctive features include:

• Expressive modeling: Capable of capturing a vast range of real-world group formation phenomena: from roommate matching to network clustering, federated learning, political party formation, and more.
• Strong correspondence between preference structure and complexity: Neutrality, additivity, symmetry, or combinatorial restrictions often demarcate the boundary between tractable and intractable coalition-formation computation.
• Robustness under data-driven and learning paradigms: Probabilistic and PAC-oriented methods allow meaningful predictions about the typical behavior of coalition stability in large, data-driven populations.
• Unified existence, stability, and welfare guarantees: Classes with the common ranking property, monotone submodularity, or potential functions achieve simultaneous Nash/core/Pareto optimal partitions, with known efficiency trade-offs (Caskurlu et al., 2020, Caskurlu et al., 2022).
• Extensibility: The approach supports ongoing extensions to richer utility models (multi-layer, fractional, graphical), dynamic coalition formation, improved statistical/empirical validation, and integration with practical platforms via simulators and empirical workflow.

The field continues to address open challenges, such as tightening bounds for approximate stability, achieving PAC stabilization in expressive classes, and mitigating the computational impact of high-dimensional coalition form spaces in applications.

7. Summary Table: Representative Hedonic Game Variants

Model/Class	Preference Structure	Key Stability/Complexity Result
Additively Separable	$u_i(S) = \sum v_i(j)$	PLS-complete (NS/IS), polytime CIS (Gairing et al., 2015)
Fractional Hedonic	$u_i(S) = \frac{1}{|S|} \sum v_i(j)$	NP-hard for welfare opt.; core may be empty (Hanaka et al., 2023, Demeulemeester et al., 2024)
Boolean Hedonic	Propositional goal formula $\varphi_i$	Core exists, NS existence NP-complete (Aziz et al., 2015)
HGCRP (Common Ranking)	$u_i(S) = U(S)$ (joint utility)	Core+Indiv.+PO always exist; NP-hard for $|S|\leq 3$ (Caskurlu et al., 2022)
Subset-neutral	$u_i(S) = \sum w(T)$ over $T\subseteq S$	Always Nash/IS stable; core+IS polytime in anonymous (Suksompong, 2018)
Distance-based (FEN)	Metric from ordinal friend/enemy	Nash-stable existence NP-complete (Rey et al., 2022)
Metric Hedonic Games	Coalitional cost by metric on types	Swap-equilibrium always exists; PoA unbounded (Haye et al., 5 Feb 2026)
$\varepsilon$-fractional core	Exponential sample–fraction of blocks allowed	Existence/algorithms for simple/anonymous (Fioravanti et al., 2023)

This table is a non-exhaustive summary and many additional hybrid and specialized classes are actively studied.

---

References:

• (Collins et al., 2020): Generating Empirical Core Size Distributions of Hedonic Games using a Monte Carlo Method
• (Hanaka et al., 2023): Maximizing Utilitarian and Egalitarian Welfare of Fractional Hedonic Games on Tree-like Graphs
• (Suksompong, 2018): Individual and Group Stability in Neutral Restrictions of Hedonic Games
• (Gairing et al., 2015): Computing stable outcomes in symmetric additively-separable hedonic games
• (Caskurlu et al., 2022): On Hedonic Games with Common Ranking Property
• (Caskurlu et al., 2020): Hedonic Expertise Games
• (Fioravanti et al., 2023): $\varepsilon$-fractional Core Stability in Hedonic Games
• (Hasan, 2021): Incentive Mechanism Design for Federated Learning: Hedonic Game Approach
• (Haye et al., 5 Feb 2026): Metric Hedonic Games on the Line
• (Rey et al., 2022): FEN-Hedonic Games with Distance-Based Preferences
• (Bullinger et al., 2024): Stability in Random Hedonic Games
• (Fioravanti et al., 2023): PAC learning and stabilizing Hedonic Games: towards a unifying approach
• (Bredereck et al., 2019): Hedonic Diversity Games
• (Aziz et al., 2015): Boolean Hedonic Games
• (Aziz et al., 2011): Individual-based stability in hedonic games depending on the best or worst players
• (Miles, 2017): A Simulator for Hedonic Games
• (Demeulemeester et al., 2024): Quantifying Core Stability Relaxations in Hedonic Games