Robust Stable Matchings: Theory & Methods

• Robust stable matchings are solutions to two-sided matching problems that remain stable despite perturbations, uncertainties, or incomplete preference information.
• The survey covers formal models such as δ-robust, ε-stable, and (a, b)-supermatches, each quantifying stability resilience under specific adversarial or noisy conditions.
• Algorithmic frameworks range from efficient polynomial-time methods for single-agent perturbations to NP-complete scenarios, highlighting trade-offs between robustness and computational feasibility.

A robust stable matching is a solution to the stable marriage (or more generally, two-sided matching) problem that remains stable in the face of specified changes, errors, or uncertainties in the agents’ preference profiles. The study of robust stable matchings addresses fundamental questions about the resilience of stable matching solutions under various forms of perturbations, partial or unreliable information, and dynamic environments. Several formal models of robustness have been introduced, each targeting distinct operational goals or adversarial models. This article surveys the primary concepts, algorithmic frameworks, structural results, and complexity boundaries that define the landscape of robust stable matchings.

1. Formal Models of Robust Stability

The literature recognizes several notions of robustness, differentiated by the kind and scope of perturbations:

• Instance Robustness: Given two preference profiles $A$ and $B$ (for the same set of agents), a matching is robust if it is stable in both—i.e., $M\in\mathcal{M}_A\cap\mathcal{M}_B$. Robustness may be defined under a class of perturbations: for example, when $B$ is obtained from $A$ by permuting a single agent’s list, by a single adjacent swap, or by changing up to $k$ agents’ lists arbitrarily (Gangam et al., 12 Jan 2026).
• $\delta$-Robust Stability: Given a profile $P$ and swap distance $\delta$, $M$ is $\delta$-robust if it is stable in all profiles within total swap distance at most $\delta$: $M\in\mathcal{SM}(P')$ for every $P'$ with $\tau(P,P')\leq\delta$ (Chen et al., 2019). This imposes stability against all small perturbations to the input.
• $\epsilon$-Stable Matching: In models with cardinal utilities, a matching is $\epsilon$-stable if no unmatched pair both gain at least $\epsilon$ by switching, i.e., both partners' utilities increase by at least $\epsilon$ by leaving their assignment for each other. This models robustness to agents who only deviate if the improvement is substantial (“switching costs”) (Pittel et al., 18 Jul 2025).
• (a, b)-Supermatches: A stable matching $M$ is an $(a,b)$-supermatch if, whenever any $a$ non-fixed pairs “break up,” it is possible to repair $M$ to another stable matching by changing at most $b$ additional pairs (Genc et al., 2017, Genc et al., 2017). This quantifies robustness to forced breakups or departures.
• Input-Error and Probabilistic Robustness: One may consider probabilistic models where input errors (e.g., upward shifts in a random agent’s list chosen according to some distribution) occur. The objective is to maximize the probability that a matching remains stable after a random error (Mai et al., 2018).
• Robustness under Partial Information: When agents specify incomplete or partial preference information, robust matchings minimize the worst-case number of blocking pairs across all completions of the input (Menon et al., 2018).

2. Lattice Structure and Algebraic Characterizations

Stable matchings of a fixed instance form a distributive lattice under the dominance order (worker/firm or man/woman dominance). Robust subfamilies frequently inherit and interact with this structure:

• Sublattice Properties: When two instances $A$, $B$ differ by a single agent’s full-list permutation, or a single adjacent swap, the set $\mathcal{M}_A\cap\mathcal{M}_B$ forms a distributive sublattice of both underlying lattices $\mathcal{L}_A, \mathcal{L}_B$ (Gangam et al., 2018, Gangam et al., 12 Jan 2026). The join- and meet-semisublattice properties hold for the complements.
• Rotation Poset and Birkhoff Compression: Every stable-matching lattice is isomorphic to the lattice of down-sets of a finite poset of “rotations.” Robust sublattices correspond to compressions of the rotation poset: partitioning rotations into “meta-rotations” preserving precedence constraints (Gangam et al., 2018). This is operationalized in the bouquet construction, which represents robust matchings as closed sets in a compressed poset.
• Succinct Representations: The compressed poset representation allows efficient enumeration and optimization over the robust sublattice (Gangam et al., 2018, Mai et al., 2018). Every robust matching corresponds to a closed set in the compressed poset.
• Loss of Lattice Structure: If changes affect two or more agents per side (type $(p, q)$ with $p+q \ge 4$), the intersection $\mathcal{M}_A\cap\mathcal{M}_B$ may not be a sublattice—structural properties, compressibility, and polyhedral integrality can fail (Gangam et al., 12 Jan 2026).

3. Algorithmic Frameworks and Complexity

Algorithmic results depend sharply on the scope of perturbation:

Perturbation Regime	Polynomial-Time?	Sublattice?	Integral Polytope?	Reference
None (classical)	Yes	Yes	Yes	Gale–Shapley
Single-agent full swap	Yes	Yes	Yes	(Gangam et al., 2018, Gangam et al., 12 Jan 2026)
Up to $k=O(1)$ agents	XP ($O(n^k)$)	No	No	(Gangam et al., 12 Jan 2026)
General ($k=\Omega(n)$)	NP-complete	No	No	(Gangam et al., 12 Jan 2026, Genc et al., 2017)

• Bouquet Algorithm: For single-agent or “single-side” perturbations, the bouquet-finding algorithm constructs the sublattice of robust matchings in $O(n^2 \cdot T_{mem})$ time by identifying “flowers” and “tails” in the rotation poset (Gangam et al., 2018).
• XP-Time Algorithms: For $k$ changed agents, iterating over candidate partial matchings for these agents and applying DA yields $O(n^{k+2})$-time algorithms (Gangam et al., 12 Jan 2026).
• (a, b)-Supermatch Decision: It is NP-complete to decide existence of general $(a, b)$-supermatches, even for $(1,1)$ in restricted instances. Polynomial-time verification is possible for a given candidate matching when $a=1$ (Genc et al., 2017, Genc et al., 2017).
• Probabilistic Robustness/Single-Shift: For input errors limited to a polynomial-sized class (e.g., upward shifts in one agent’s list), a combinatorial max-flow/closed-set LP yields robust matchings and a compact sublattice representation (Mai et al., 2018).
• $\delta$-Robust (Swap Distance): Testing existence and optimizing for $\delta$-robust matchings is in $O(n^4)$ for strict lists, but NP-hard with ties or under near-stability objectives (Chen et al., 2019). Optimizing for egalitarian cost over robust solutions is tractable via totally unimodular LP.

4. Quantitative Robustness: Measures and Thresholds

• Worst-Case vs. Average-Case: The minimal number of swaps that destroy the stability of a given matching (worst-case robustness) is typically $1$ for men-optimal matchings; average-case measures (fraction of random $\ell$-swap perturbations that preserve stability) decay more slowly, with 50% thresholds at $1–2$ swaps per list in synthetic instances (Boehmer et al., 2024).
• Correlation with Blocking-Pair Proximity: Robustness correlates with blocking-pair proximity; matchings minimizing summed rank of partner assignments (“summed-rank-minimizing”) exhibit greater average-case robustness than standard extremal stable matchings (Boehmer et al., 2024).
• $\epsilon$-Stable Threshold Phenomena: In random cardinal-utility models, the number of $\epsilon$-stable matchings exhibits a sharp phase transition: for $\epsilon=O(n^{-1}\log n)$, the expected count grows only polynomially; for $\epsilon\gg n^{-1}\log n$ it becomes super-polynomial, leading to a combinatorial explosion of approximately robust solutions (Pittel et al., 18 Jul 2025).
• Partial Information and Super-Blocking Pairs: Under incomplete preference orders, robustness is measured via minimax of super-blocking pairs, tight approximations are characterized, and inapproximability bounds are established beyond special cases (Menon et al., 2018).

5. Perturbation-Robustness, Dynamic and Distributed Models

• Agent Departures: Perturbation-robust stable matchings explicitly incorporate the probability of agents leaving the market. The objective balances expected social cost and regret (distance to the optimal re-matching post-departure), optimized via a rotation-DAG/flow framework (Jacobovic, 2016).
• Distributed and Local Robustness: In distributed and local-information settings, almost-stable matchings (with $O(\epsilon n)$ blocking pairs) can be attained in $O(1)$ synchronous rounds and exhibit robustness to local perturbations, in contrast to the global fragility of perfect stability (0812.4893).
• Byzantine Robustness: In adversarial distributed environments, the solvability of stable matching under up to $t_L$ and $t_R$ Byzantine agents is characterized precisely as a function of the connectivity, authentication, and cryptographic protocols, with tight thresholds and protocol constructions for feasible regimes (Constantinescu et al., 9 Feb 2025).

6. Applications, Implications, and Extensions

Robust stable matchings are central for the design of platforms and protocols in environments where inputs are subject to noise, agents’ preferences are uncertain, information is incomplete, or the platform must guarantee resilience to departures and local failures.

• Market Design Implications: In large random markets or DA-based mechanisms, robust equilibrium results show that even widespread non-truthful reporting yields outcomes asymptotically equivalent to full truthfulness, provided stability is enforced (Artemov et al., 2022).
• Robustness vs. Social-Optimality Trade-offs: There are structural tensions between robustness (especially for large $\delta$ or strong adversarial models) and objectives such as minimizing social cost, maximizing the number of matched agents, or achieving egalitarian outcomes (Chen et al., 2019).
• Correlated Utility Models: In settings with correlated public and private utilities, all stable matchings for high-rated agents yield similar values, and robust matching protocols can use short lists ($O(\log n)$), with most agents having unique stable partners (Agarwal et al., 2022).
• Polyhedral Descriptions: LP relaxations of the robust stable matching problem are integral in single-side perturbation regimes but lose integrality if two or more agents per side may change, demonstrating a geometric phase transition in tractability (Gangam et al., 12 Jan 2026).

7. Open Questions and Future Directions

• Full characterization and efficient enumeration of robust matching lattices under multi-agent perturbations and for incomplete, many-to-one, or more general matching models.
• Tight approximability and parameterized complexity in robust and near-stable regimes, especially for values of $\delta$ scaling with $n$ (Chen et al., 2019).
• Integration of robust stable matching principles into practical platforms facing noisy, incomplete, or adversarially manipulated inputs.
• Deeper study of the relationship between blocking-pair structure, lattice geometry, and robust solution counts in real-world and synthetic preference models.

Robust stable matchings are a fertile interface of algorithmic game theory, combinatorics, and operations research, with ongoing developments in fine-grained structural theory, computational complexity, and market design applications (Gangam et al., 12 Jan 2026, Gangam et al., 2018, Boehmer et al., 2024).