Matroid-Constrained Fair Division

• Matroid-constrained fair division is an allocation process that requires each agent’s bundle to form an independent set or basis, ensuring combinatorial feasibility.
• The study demonstrates that equitable partitioning via exchange graphs provides exact fairness guarantees—EF1 for tri-valued and MMS for bi-valued valuations—within polynomial time.
• The work unifies combinatorial optimization with fairness metrics, resolving longstanding conjectures and opening new research avenues in multi-set and general valuation settings.

Matroid-constrained fair division studies the allocation of indivisible items among agents subject to combinatorial feasibility requirements modeled by matroids—a fundamental abstraction capturing independence, partitioning, and exchangeability constraints. This paradigm extends classical fair division by enforcing that each agent's bundle must be an independent set, or typically a basis, of a specified matroid. Matroid-constrained fair division encompasses classic resource allocation scenarios such as the assignment of shifts, referees, or assets under group, quota, or combinatorial restrictions, and it stands at the confluence of combinatorial optimization, algorithmic fairness, and matroid theory.

1. Matroidal Foundations and Formal Model

A matroid $M = (E, \mathcal{I})$ is defined by a finite ground set $E$ and a family of independent sets $\mathcal{I} \subseteq 2^E$ satisfying:

• Nonempty: $\emptyset \in \mathcal{I}$.
• Hereditary: $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$.
• Augmentation: $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$ with $I \cup \{e\} \in \mathcal{I}$.

Every maximal independent set (a basis) has cardinality $r(M)$, the rank. Examples include uniform matroids (cardinality constraints), partition matroids (group quotas), and graphic matroids (spanning trees).

In the fair division setting, let $N = [n]$ be the set of agents and $E$ the set of $E$0 indivisible items. Each agent $E$1 is assigned a valuation $E$2, and feasibility is imposed by a matroid constraint. An allocation $E$3 is feasible if $E$4 for each $E$5, usually $E$6 must be a basis, with $E$7. The core questions are: under what conditions do classical and algorithmic fairness notions—envy-freeness up to one item (EF1), maximin share (MMS), Pareto optimality—admit existence and efficient computation in this matroidal regime?

2. Existence Theorems: Equitable Partition and MMS

The principal structural result is the equitability theorem: if $E$8 can be partitioned into $E$9 bases, then for any subset $\mathcal{I} \subseteq 2^E$0 there is a partition $\mathcal{I} \subseteq 2^E$1 (each $\mathcal{I} \subseteq 2^E$2 a basis) such that $\mathcal{I} \subseteq 2^E$3 are as balanced as possible, i.e.,

$\mathcal{I} \subseteq 2^E$4

This result, proven via iterative exchange operations in the exchange graph of pairs of bases, settles the Matroid Equitability Conjecture and extends beyond graphic, regular, and strongly base-orderable cases (Akrami et al., 16 Jul 2025).

For applications to MMS, for any bi-valued additive valuation $\mathcal{I} \subseteq 2^E$5, the ground set of unit-valued goods can be partitioned equitably into $\mathcal{I} \subseteq 2^E$6 bases (one per agent), giving each $\mathcal{I} \subseteq 2^E$7 or $\mathcal{I} \subseteq 2^E$8, thereby matching the exact maximin share (MMS) guarantee for all agents. The allocation can be constructed globally using a matching argument on the bipartite graph between bases and agents, recursively establishing existence in polynomial time.

3. Exchange Graphs and Algorithmic Frameworks

Key to constructive proofs is the directed exchange graph $\mathcal{I} \subseteq 2^E$9 for pairs of bases $\emptyset \in \mathcal{I}$0. Vertices correspond to items in $\emptyset \in \mathcal{I}$1; symmetric directed cycles encode exchangeable subsets $\emptyset \in \mathcal{I}$2 such that $\emptyset \in \mathcal{I}$3 and $\emptyset \in \mathcal{I}$4 are also bases. When the $\emptyset \in \mathcal{I}$5-counts in two bases are imbalanced, one locates an appropriate exchangeable set and swaps, incrementally narrowing discrepancies.

The general iterative procedure (Algorithm A) involves repeatedly identifying bases whose intersection with the target set $\emptyset \in \mathcal{I}$6 differ by more than $\emptyset \in \mathcal{I}$7, then augmenting the smaller by one using cycles in $\emptyset \in \mathcal{I}$8. This process converges in polynomial time. For more complex objectives (e.g., two disjoint sets $\emptyset \in \mathcal{I}$9), lex-minimization of weighted $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$0-potentials guides the selection and application of exchanges (Algorithm B).

MMS allocations for all agents are obtained by stitching together individual equitable splits using matching and Hall's theorem, with recursive reduction upon Hall violators (Akrami et al., 16 Jul 2025).

4. Fair Division Notions under Matroid Constraints

4.1 Envy-Freeness up to One Item (EF1)

An allocation $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$1 is EF1 if, for all $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$2 with $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$3, there exists $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$4 such that $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$5. Under matroid constraints, the above equitability result ensures that for identical, tri-valued additive valuations, the allocation into bases via equitable splitting of the sets of value-$$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$6 and value-$$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$7 goods guarantees EF1. Any higher envy would allow a swap argument restoring EF1; thus, no agent envies another by more than one item, and a polynomial-time construction exists (Akrami et al., 16 Jul 2025).

4.2 Maximin Share (MMS)

Given $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$8 the set of feasible base partitions, agent $$J \in \mathcal{I}, I \subseteq J \implies I \in \mathcal{I}$$9's MMS is $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$0. For bi-valued additive valuations, equitable splitting achieves the exact MMS for each agent because the bundles' intersection sizes match the optimal uniform split. The global allocation is then found by matching and recursion.

5. Equitable Splitting for Multiple Subsets

For two disjoint sets $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$1, it is generally impossible to balance both intersections perfectly across all $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$2 bases; however, it is always possible to achieve $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$3 within $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$4, and $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$5 within $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$6, for all bases $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$7—a result that cannot be improved in general matroids (as shown by the $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$8-graphic matroid). The potential-minimization proof ensures that no better tradeoff is possible, and the algorithmic implementation remains polynomial-time.

Scenario	Balance possible in $$I, J \in \mathcal{I},\ |J| > |I| \implies \exists e \in J \setminus I$$9	Balance possible in $I \cup \{e\} \in \mathcal{I}$0	Reference
$I \cup \{e\} \in \mathcal{I}$1 only	$I \cup \{e\} \in \mathcal{I}$2	N/A	(Akrami et al., 16 Jul 2025)
$I \cup \{e\} \in \mathcal{I}$3	$I \cup \{e\} \in \mathcal{I}$4	$I \cup \{e\} \in \mathcal{I}$5	(Akrami et al., 16 Jul 2025)

For higher-order equitability ($I \cup \{e\} \in \mathcal{I}$6 disjoint sets), the optimal gap is open; the conjecture is that $I \cup \{e\} \in \mathcal{I}$7-balancing suffices for all $I \cup \{e\} \in \mathcal{I}$8.

6. Complexity and Algorithmic Aspects

Each exchange operates by searching for short cycles in the exchange graph, an $I \cup \{e\} \in \mathcal{I}$9 task, with the total number of required exchanges bounded by the maximum imbalance for the relevant subset(s). For the full allocation process, combining equitable partition with matching to agents, the complexity is $r(M)$0 or better, as each recursive step strictly decreases the problem size.

These results hold for arbitrary matroids, subsuming all previously known classes (graphic, partition, regular, strongly base-orderable). Under the stated conditions, both EF1 (identical tri-valued valuations) and MMS (bi-valued valuations) allocations are obtained exactly and efficiently. Extension to general identical valuations for EF1 or multivalued MMS remains open.

7. Impact, Extensions, and Open Directions

The matroid equitability theorem provides unified combinatorial and algorithmic foundations for balanced splitting in matroids, settling several longstanding conjectures and directly enabling exact fairness guarantees under broad constraint regimes (Akrami et al., 16 Jul 2025). This advance has immediate implications for resource allocation under combinatorial constraints, as it lifts the frontier of provable fairness and tractability to the entire class of matroids.

Limitations regard the restrictions on valuation types: EF1 is established only for identical tri-valued valuations, and MMS for bi-valued. Extending these results to arbitrary identical additive valuations (for EF1) or multi-valued/heterogeneous valuations (for MMS) is a central open question.

Further research directions include:

• Generalization to multiple ($r(M)$1) subsets for equitable partitioning.
• Characterization of exchange sequences and their minimal lengths, with connections to Rota's basis conjecture and basis-sorting problems in matroid theory.
• Algorithmic and structural exploration beyond the union of bases setting, potentially involving nonbasis constraints or broader independence system generalizations.

Matroid-constrained fair division thus stands as a foundational synthesis, enabling fair, efficient, and combinatorially robust allocation in the presence of rich real-world feasibility constraints.