Many-to-Many Matching: Techniques & Insights

• Many-to-many matching strategies are a paradigm that assigns multiple partners under quota and capacity constraints with rigorously defined mathematical models.
• Algorithmic techniques include Hungarian-based reductions, dynamic programming, geometric approaches, and optimal transport frameworks to meet diverse application needs.
• These methods are applied in resource allocation, market design, wireless caching, and reviewer assignment, highlighting their practical impact in complex systems.

Many-to-many matching strategies encompass a broad array of mathematical, algorithmic, and economic models for associating members of two (or more) sets to potentially multiple partners, subject to constraints, objective functions, and preferences. These strategies play foundational roles in combinatorial optimization, stable matching theory, computational geometry, resource allocation, market design, and machine learning. Recent research investigates matching with capacities, demands, externalities, geometric constraints, stability, Pareto-optimality, and strategic considerations, yielding both efficient algorithms and deep complexity-theoretic insights.

1. Formal Models, Constraints, and Solution Concepts

The many-to-many matching paradigm generalizes classical one-to-one and many-to-one matching by permitting both sides to be matched to multiple partners, typically up to prescribed capacities or quotas. Formally, given two finite sets $A$ and $B$, a many-to-many matching $M \subseteq A \times B$ must satisfy lower and upper bounds ($\alpha_i, \alpha'_i$ for $a_i \in A$ and $\beta_j, \beta'_j$ for $b_j \in B$) on the number of matches per element (Rajabi-Alni et al., 2022, Rajabi-Alni, 2013, Rajabi-Alni et al., 2012). Each element may be subject to demand (minimum partners required) and capacity (maximum allowed matches).

Preferences over partners can be strict, weak, or include ties; extensions compare bundles lexicographically or allow for externalities. Solution concepts include minimum-cost matchings (for geometric or graph-theoretic settings) (Bandyapadhyay et al., 2021, Bandyapadhyay et al., 2024), stable matchings (blocking-pair-free under agent preferences) (Hamidouche et al., 2014, Bonifacio et al., 2021, Glitzner et al., 16 May 2025, Biró et al., 2022), Pareto-optimal matchings (Cechlárová et al., 2015), and notions of popularity or criticality under lower quotas (Nasre et al., 2022, Kavitha et al., 2024).

2. Algorithmic Techniques for Many-to-Many Matching

Algorithmic approaches for many-to-many matching reflect the underlying mathematical structure and application context.

• Hungarian-based reductions: Demands and capacity constraints are enforced by constructing expanded bipartite graphs with node splitting and dummy vertices, reducing the problem to an assignment of artificial elements and then applying the Hungarian method (Rajabi-Alni et al., 2022, Rajabi-Alni, 2013). Complexity can be $O(n^4)$ or better in special cases, and feasibility boils down to the existence of perfect matchings respecting all quotas.
• Dynamic programming (1D geometric matching): When input sets lie on the real line, DP recurrences capture the evolution of feasible matchings block-wise; greedy pairings between blocks guarantee non-crossing solutions and $O(n^2)$ or even linear time under further restrictions (Rajabi-Alni et al., 2012, Rajabi-Alni et al., 2019, Rajabi-Alni et al., 2017, Rajabi-Alni et al., 2019). These algorithms exploit the one-dimensional structure and capacity separation properties.
• Geometric matching in higher dimensions: Matching in $\mathbb{R}^2$ and beyond uses reductions to edge covers, perfect matchings, and advanced data structures (e.g., dynamic bichromatic closest pair, Voronoi diagrams), yielding exact $O(n^2 \cdot \operatorname{polylog} n)$ and $(1+\varepsilon)$-approximate $O(n^{3/2} \cdot \operatorname{polylog} n)$ algorithms (Bandyapadhyay et al., 2021, Bandyapadhyay et al., 2024).
• Relaxation-based graph matching: For graph-alignment tasks, cluster-to-cluster assignments are formulated via quartic objective functions, relaxed to continuous domains, and solved using conditional gradient (Frank–Wolfe style) methods; final matchings may require clustering or greedy integer projection (Zaslavskiy et al., 2010).
• Deferred-acceptance and trading-cycle algorithms: Stability-oriented approaches extend Gale–Shapley and TTC schemes to many-to-many, often requiring preferences to be substitutable or lexicographic; they yield stable, strong-core, or Pareto-optimal matchings under suitable conditions (Hamidouche et al., 2014, Biró et al., 2022, Bonifacio et al., 2021, Mordig et al., 2021, Glitzner et al., 16 May 2025). Some advanced mechanisms (GSDT) handle ties via flow networks and are capable of generating all POMs (Cechlárová et al., 2015).
• Optimal transport frameworks: Many-to-many matching is reformulated as a budget-constrained OT problem with sparsity-inducing $q$-entropy regularization, solved via penalty-decomposition algorithms (Liu et al., 31 Mar 2025). This approach enables explicit control over matching cardinalities, prevents degeneration to the one-to-one case, and generalizes to soft and hard budget constraints.

3. Stability, Core, Pareto, and Popularity Criteria

Stable many-to-many matchings require that no pair (or, in strong core cases, no coalition) can mutually improve their situation by deviating, under agents' preferences (Hamidouche et al., 2014, Bonifacio et al., 2021, Biró et al., 2022, Glitzner et al., 16 May 2025). Substitutability and the law of aggregate demand ensure lattice structure and tractability for the set of stable matchings (Bonifacio et al., 2021). Pairwise and coalition stability are distinct; the strong core may be empty even under lexicographic preferences, and its non-emptiness is NP-hard to ascertain (Biró et al., 2022).

Pareto-optimality requires that no alternative matching strictly improves some agents without worsening others. Mechanisms such as GSDT and men-proposing DA generalize serial dictatorship and guarantee Pareto-optimality in polynomial time if orderings are fixed accordingly (Cechlárová et al., 2015). Verification and generation of all POMs is coNP-complete.

Popularity extends the voting-based notion from one-to-one to many-to-many: a matching is popular if, in pairwise comparisons over all feasible matchings, it never loses a majority of votes, generalizing to perfect and critical matchings under lower quotas (Kavitha et al., 2024, Nasre et al., 2022). Polynomial-time algorithms exist for finding (min-cost) popular perfect matchings in capacitated bipartite settings by reduction to colorful stable-matchings, and similar methods apply in critical and lower-quota problems.

4. Complexity and Computational Hardness

Many-to-many matching problems display a wide spectrum of computational complexity:

Problem type	Hardness	Efficient algorithms
MMDC with lower/upper quotas (general)	Assignment problem (P)	Hungarian-based ($O(n^4)$–$O(n^6)$)
Geometric matching ($\mathbb{R}^d$)	NP-hard	$O(n^2)$, $O(n^{3/2} \operatorname{polylog} n)$, $O_\varepsilon(n\log n)$ approximate
Stable core (two-sided, lexicographic)	NP-hard to detect non-emptiness	TTC yields near-feasible matching
Pareto-optimality checking	coNP-complete	DA-based generates max-size POM
Popular/perfect critical matchings	Polynomial	Multi-level DA, LP dual approaches
MM with additive externalities	coNP-complete/NP-hard (neutral, pessimistic)	Polynomial (optimistic, FPT)

NP-hardness emerges in strong-core existence, Pareto verification, and general market stability; efficient algorithms are possible in restricted geometries, under substitutable preferences, or using FPT techniques when the number of “controversial” edges is small (Brânzei et al., 2012, Biró et al., 2022, Glitzner et al., 16 May 2025, Zaslavskiy et al., 2010, Nasre et al., 2022).

5. Applications and Empirical Results

Many-to-many matching strategies underlie contemporary resource allocation, assignment, and clustering systems:

• Wireless caching: Pairwise-stable cache placement in small-cell networks markedly increases satisfaction ratios and reduces latency compared to random allocation (Hamidouche et al., 2014).
• UAV relay selection: Distributed potential-matching approaches maximize global throughput and satisfaction, nearly attaining optimality under reasonable message/iteration complexity (Liu et al., 2017).
• Assignment and reviewer allocation: PolyGS extends Gale–Shapley to quotas on both sides, supporting multi-phase programs such as ELLIS 2020 with high match quality and proposer-optimality (Mordig et al., 2021).
• Geometric data mining: Efficient geometric matching and graph alignment methods underpin clustering, recognition, and cross-domain linkage tasks, exploiting structure for combinatorial speed-ups (Bandyapadhyay et al., 2021, Bandyapadhyay et al., 2024, Zaslavskiy et al., 2010).
• Market and mechanism design: Deferred-acceptance and TTC variants accommodate strategic and fairness constraints in multi-partner markets, handling ties, quotas, and externalities (Cechlárová et al., 2015, Biró et al., 2022, Brânzei et al., 2012).

Empirical analyses reveal subtle phenomena: mid-range capacities may decrease the likelihood of solvability in non-bipartite fixtures (Glitzner et al., 16 May 2025); near-feasible matchings often require few capacity adjustments. The Rural Hospitals Theorem generalizes to many-to-many scenarios: all stable matchings allocate the same total degree to each agent, matching market invariance (Glitzner et al., 16 May 2025, Nasre et al., 2022).

6. Extensions, Limitations, and Strategic Considerations

Contemporary research investigates robustness to ties, externalities, multiple quotas, groupwise constraints, and strategic manipulation. In many-to-many with ties, every POM can be generated via GSDT, but truthfulness is generally unattainable except under contiguous applicant orderings; manipulation risks are inherent (Cechlárová et al., 2015). Additive externalities in matching games produce varied stability definitions, with tractable optimistic solutions and generically hard neutral/pessimistic cases (Brânzei et al., 2012).

Continuous relaxations and optimal transport models offer high flexibility but guarantee only local optimality for non-convex objectives; penalty methods and $q$-entropy regularization enable capacity control without artificial degeneracy to one-to-one assignments (Liu et al., 31 Mar 2025).

7. Comparative Table of Algorithmic Paradigms

Setting	Algorithmic Approach	Guarantees/Complexity
MMDC (assignment, arbitrary cost)	Node-splitting/Hungarian	$O(n^4)$–$O(n^6)$, exact
1D geometric with capacities/demands	DP blockwise scan, greedy pairing	$O(n^2)$–$O(n)$, optimal
Planar geometric matching	Reduction, geometric data structs	$O(n^2\operatorname{polylog}n)$ exact, $O_\varepsilon(n\log n)$ approx
Pareto-optimal (many-to-many)	DA/TTC/GSDT algorithms	Polynomial (for near-feasibility), coNP-complete verification
Popular/critical (lower quotas)	Multi-level DA, LP duality	Polynomial, size/rural-hospital invariance
Stable non-bipartite (Stable Fixtures)	GSP/half-matching partition, ILP	$O(n^4)$ for GSP, $O_{ILP}(n)$ for optimization
Additive externalities (matching games)	Optimistic stability, FPT	FPT in “controversial” edges, polynomial membership check

In conclusion, many-to-many matching strategies form a rich toolkit for combinatorial, economic, and computational problems where multi-partner allocations are essential. They offer precise control over assignment structures, accommodate complex preferences, and support precise notions of efficiency, stability, and popularity—all under rigorous mathematical and algorithmic frameworks.