A Tight Bound for Shortest Augmenting Paths on Trees

The shortest augmenting path technique is one of the fundamental ideas used in maximum matching and maximum flow algorithms. Since being introduced by Edmonds and Karp in 1972, it has been widely applied in many different settings. Surprisingly, despite this extensive usage, it is still not well understood even in the simplest case: online bipartite matching problem on trees. In this problem a bipartite tree $T=(W \uplus B, E)$ is being revealed online, i.e., in each round one vertex from $B$ with its incident edges arrives. It was conjectured by Chaudhuri et. al. [K. Chaudhuri, C. Daskalakis, R. D. Kleinberg, and H. Lin. Online bipartite perfect matching with augmentations. In INFOCOM 2009] that the total length of all shortest augmenting paths found is $O(n \log n)$. In this paper, we prove a tight $O(n \log n)$ upper bound for the total length of shortest augmenting paths for trees improving over $O(n \log^2 n)$ bound [B. Bosek, D. Leniowski, P. Sankowski, and A. Zych. Shortest augmenting paths for online matchings on trees. In WAOA 2015].

💡 Research Summary

The paper studies the online bipartite matching problem restricted to trees, where a bipartite tree T = (W ⊎ B, E) is revealed vertex‑by‑vertex: at each round a black vertex bₜ∈B together with all its incident edges arrives, while the white vertices W are known in advance. The goal is to maintain a maximum matching after each arrival, using the classic Shortest Augmenting Path (SAP) algorithm, which always augments along the currently shortest augmenting path. The main question is how many edges are traversed in total by all augmenting paths over the entire sequence of arrivals.

Historically, Edmonds‑Karp introduced the SAP technique for flow, and Chaudhuri et al. (2009) conjectured that even in the worst arrival order the total length of augmenting paths should be O(n log n) for general bipartite graphs. For trees, Bosek et al. (2015) proved an O(n log² n) bound, leaving a gap to the conjectured optimal O(n log n). This work closes that gap by establishing a tight O(n log n) upper bound for trees.

The authors’ approach departs from directly analyzing the SAP paths. Instead they define a two‑player “mini‑max game” on a rooted tree. For any rooted tree T whose vertices are colored black (B) or white (W), each vertex v receives a “mini‑max value” recursively:

• For a black vertex b, mini‑max_T(b) = 1 + min_{w∈children(b)} mini‑max_T(w) (or ∞ if b has no children).
• For a white vertex w, mini‑max_T(w) = 1 + max_{b∈children(w)} mini‑max_T(b) (or 0 if w is a leaf).

The value represents the length of the worst‑case shortest augmenting path that an adversary could force when the algorithm always chooses the shortest possible edge. Two distance functions are defined for each vertex at time t:

• distₜ(v) = mini‑max_T(v), the primary distance to a free white leaf.
• sec‑distₜ(v) = mini‑max of the tree obtained after removing the child that determines the primary distance, i.e., the second‑best distance.

A crucial observation (Observation 2) is that both distₜ and sec‑distₜ are monotone non‑decreasing in t; adding a new black vertex can never decrease any mini‑max value. Lemma 3 shows that for any matching M, the actual shortest augmenting path ρₜ from the newly arrived black vertex bₜ to a free white vertex satisfies |ρₜ| ≤ distₜ(bₜ). Hence the mini‑max distance upper‑bounds the SAP path length.

The paper then connects the mini‑max distances to Hall’s condition. Lemma 4 proves that distₜ(b) = ∞ iff b violates Hall’s condition at time t. Vertices with infinite secondary distance are called “dead”. A dead black vertex is one for which sec‑distₜ(b) = ∞; a dead white vertex is one with distₜ(w) = ∞. The authors define the alive set Aₜ (vertices not dead at time t) and the dead set Dₜ, and study how vertices transition from alive to dead as the process evolves.

Several structural lemmas describe the behavior of dead vertices:

• Observation 6: a black leaf is dead immediately upon arrival; a black vertex becomes dead when it has at most one alive neighbor; a white vertex becomes dead when it has at least one dead neighbor.
• Lemma 7 shows that the mini‑max path starting from any alive vertex stays entirely within the alive region; it never passes through a vertex that died in the previous round.
• Lemma 8 guarantees that if a newly arrived black vertex does not break Hall’s condition, it must have at least one neighbor that was alive in the previous round.

These facts imply that each round either introduces no new dead vertices (if the new black vertex has at least two alive neighbors) or kills a limited, well‑structured set of vertices. By carefully charging the increase of distₜ values to the depth of the tree and using the monotonicity property, the authors bound the sum Σₜ distₜ(bₜ) by O(n log n). Since each SAP path length is bounded by the corresponding distₜ, the total length of all augmenting paths is also O(n log n).

The result matches the conjectured optimal bound for trees and improves the previous O(n log² n) bound. The paper’s contribution is twofold: it provides a clean combinatorial game‑theoretic framework that captures the worst‑case behavior of SAP on trees, and it introduces the dead/alive vertex analysis that isolates the portions of the tree where Hall’s condition is tight. These techniques may be extensible to general bipartite graphs, offering a promising route toward proving the full Chaudhuri conjecture.

In summary, the authors prove that for any online arrival order of black vertices on a bipartite tree, the Shortest Augmenting Path algorithm performs at most Θ(n log n) total edge traversals, establishing a tight bound and advancing our understanding of augmenting‑path based matching algorithms in dynamic settings.