Dynamic Matroids: Base Packing and Covering

In this paper, we consider dynamic matroids, where elements can be inserted to or deleted from the ground set over time. The independent sets change to reflect the current ground set. As matroids are central to the study of many combinatorial optimization problems, it is a natural next step to also consider them in a dynamic setting. The study of dynamic matroids has the potential to generalize several dynamic graph problems, including, but not limited to, arboricity and maximum bipartite matching. We contribute by providing efficient algorithms for some fundamental matroid questions. In particular, we study the most basic question of maintaining a base dynamically, providing an essential building block for future algorithms. We further utilize this result and consider the elementary problems of base packing and base covering. We provide a deterministic algorithm that maintains a $(1\pm \varepsilon)$-approximation of the base packing number $Φ$ in $O(Φ\cdot \text{poly}(\log n, \varepsilon^{-1}))$ queries per update. Similarly, we provide a deterministic algorithm that maintains a $(1\pm \varepsilon)$-approximation of the base covering number $β$ in $O(β\cdot \text{poly}(\log n, \varepsilon^{-1}))$ queries per update. Moreover, we give an algorithm that maintains a $(1\pm \varepsilon)$-approximation of the base covering number $β$ in $O(\text{poly}(\log n, \varepsilon^{-1}))$ queries per update against an oblivious adversary. These results are obtained by exploring the relationship between base collections, a generalization of tree-packings, and base packing and covering respectively. We provide structural theorems to formalize these connections, and show how they lead to simple dynamic algorithms.

💡 Research Summary

The paper introduces a new computational model called dynamic matroids, in which the ground set of elements evolves over time through insertions and deletions, and the family of independent sets adapts automatically to these changes. This model generalizes static matroids and captures a wide range of combinatorial structures, including graphic, linear, and partition matroids. The authors distinguish between two adversarial settings—oblivious (updates are fixed in advance) and adaptive (updates may depend on the algorithm’s state)—and design algorithms that are either deterministic (robust against adaptive adversaries) or randomized (robust against oblivious adversaries).

The first technical contribution is a deterministic algorithm for maintaining a minimum‑weight base of a dynamic matroid. Building on a rank‑oracle model (the algorithm can query the rank of any subset), the method updates the current minimum‑weight base with only O(log n) rank queries per insertion or deletion. This improves on prior work (BMNT23) that handled deletions only and required ˜O(rk(E)) queries per update. The new approach leverages a “light exchange” technique reminiscent of dynamic MST algorithms (Frederickson, Eppstein‑Galil‑Italiano‑Nissenzweig) and avoids any dependence on the matroid’s rank beyond logarithmic factors.

The core of the paper addresses two fundamental matroid parameters in a dynamic setting:

1. Base Packing Number (Φ) – defined as the fractional packing number Φ(M) = min_{A⊂E, rk(A)<rk(E)} |A| / (rk(E)−rk(A)). Its integer part ⌊Φ⌋ equals the maximum number of pairwise disjoint bases that can be packed. The authors prove a structural theorem (Theorem 1.2) linking Φ to a newly defined base collection, a generalization of tree‑packings. Using this connection, they devise a deterministic algorithm that maintains a (1 ± ε)‑approximation of Φ with O(Φ·ε⁻⁴·log³ n) rank queries per update (worst‑case) or O(Φ·ε⁻⁴·log³ n) amortized queries. The algorithm’s cost scales only with Φ (or an upper bound Φ_max) and not with the ground‑set size n, yielding an exponential improvement over static methods that need ˜O(n·Φ/ε²) queries after each change.

2. Base Covering Number (β) – defined as the fractional covering number β(M) = max_{A⊂E, A≠∅} |A| / rk(A). Its ceiling ⌈β⌉ is the smallest number of bases needed to cover all elements. Theorem 1.3 establishes a dual structural relationship between β and base collections. Deterministically, the authors achieve a (1 ± ε)‑approximation of β with O(β·ε⁻⁴·log³ n) rank queries per update. More strikingly, they present a randomized algorithm (Theorem 1.4) that, against an oblivious adversary, maintains a (1 ± ε)‑approximation of β using only O(poly(log n, ε⁻¹)) rank queries per update, independent of β’s magnitude. This algorithm succeeds with high probability and relies on a sampling‑based refinement of the greedy base collection.

Both packing and covering algorithms hinge on a greedy base collection data structure. The structure stores a set of bases that are locally optimal with respect to weight or size. When an element is inserted or deleted, only the bases that directly involve that element need to be examined and possibly replaced, which can be verified via a constant number of rank queries. The greedy update rule guarantees that the collection remains within a (1 ± ε) factor of the optimal fractional solution, thanks to the structural theorems that bound the loss incurred by local adjustments.

The paper also discusses specializations to the graphic matroid, where bases correspond to spanning trees. In this case, the dynamic base packing problem coincides with dynamic tree‑packing (related to min‑cut), and the dynamic base covering problem corresponds to arboricity. Existing dynamic algorithms for arboricity (e.g., De Vos & Christiansen 2025) achieve poly(log n) update time for exact values; the authors’ results match these bounds for approximate values while extending to arbitrary matroids. They note that exact dynamic packing/covering for general matroids would likely require breakthroughs beyond current techniques, making their (1 ± ε) approximations the realistic target.

Beyond graphics, the authors argue that the framework applies to linear matroids (column independence in matrices) and partition matroids, where base packing and covering have interpretations in matrix factorization and resource allocation. The rank‑oracle model is natural for these settings, as rank queries can be answered via Gaussian elimination or specialized data structures.

In summary, the paper makes three major contributions:

• A deterministic O(log n)‑query algorithm for maintaining a minimum‑weight base in fully dynamic matroids.
• Deterministic (and, for covering, randomized) algorithms that maintain (1 ± ε)‑approximations of the fractional base packing number Φ and covering number β with query complexities that depend only on Φ, β (or polylogarithmic factors), not on the ground‑set size.
• A unifying structural theory based on base collections, which generalizes tree‑packing concepts and enables simple, local greedy updates.

These results open a new line of research in dynamic combinatorial optimization, providing tools that could be leveraged for real‑time network design, streaming linear algebra, and other applications where the underlying independence structure evolves continuously.