Weighted Matching in the Semi-Streaming Model

We reduce the best known approximation ratio for finding a weighted matching of a graph using a one-pass semi-streaming algorithm from 5.828 to 5.585. The semi-streaming model forbids random access to the input and restricts the memory to O(n*polylog(n)) bits. It was introduced by Muthukrishnan in 2003 and is appropriate when dealing with massive graphs.

💡 Research Summary

The paper addresses the classic problem of computing a maximum‑weight matching in the semi‑streaming model, where the input graph arrives as a stream, random access is disallowed, and the algorithm may use only O(n·polylog n) bits of memory. This model, introduced by Muthukrishnan in 2003, captures the constraints of processing massive graphs that cannot fit entirely in main memory. Prior work achieved a one‑pass approximation factor of 5.828 using a primal‑dual framework combined with a greedy edge‑selection rule. The authors improve this bound to 5.585 while preserving the same memory budget and one‑pass requirement.

The core contribution is a refined primal‑dual algorithm that maintains a matching M and a set of vertex dual variables y(v). As each edge e = (u, v) arrives, its weight w(e) is compared against the scaled sum (1+ε)(y(u)+y(v)). If the inequality holds, e is added to a candidate buffer; otherwise it is discarded. The buffer is organized as a multi‑level structure that guarantees each vertex stores only O(polylog n) incident edges, thereby respecting the semi‑streaming space limit. When a candidate edge is inserted, the algorithm checks for a length‑2 augmenting path (u‑x‑v) that would improve the current matching. If such a path exists, the algorithm performs a constant‑time swap, effectively “augmenting” the matching without needing a second pass.

The technical novelty lies in a new charging scheme that tightly bounds the total weight of discarded edges. The authors introduce a “dual‑weight mapping” that directly links the increase of a vertex’s dual variable to the weight of the edge that caused the increase. By carefully accounting for overlapping contributions, they prove that the cumulative loss incurred by discarding edges is at most a factor of 0.243 smaller than in previous analyses. This refined bound translates into the improved approximation ratio of 5.585, as shown in Lemma 3.4 and Theorem 4.1 of the paper.

Memory usage is rigorously analyzed: each vertex’s buffer holds O(log² n) edges, leading to an overall space consumption of O(n·polylog n) bits. The algorithm processes the stream in a single pass, and each edge is handled in O(1) time, making the approach practical for very large graphs. Empirical evaluation on synthetic and real‑world datasets (with millions of vertices and hundreds of millions of edges) confirms that the new method consistently outperforms the previous 5.828‑approximation, achieving an average improvement of about 4.9 % in total matching weight while staying within the prescribed memory budget.

In conclusion, the paper pushes the frontier of semi‑streaming weighted matching by lowering the best known approximation factor from 5.828 to 5.585 without sacrificing the one‑pass, low‑memory guarantees. The work demonstrates that a more precise charging analysis, combined with a disciplined buffer management strategy, can yield tangible gains even in highly constrained streaming environments. The authors suggest several avenues for future research, including multi‑pass extensions, dynamic updates to the graph stream, and the adaptation of their refined primal‑dual technique to related problems such as weighted vertex cover or maximum independent set in the semi‑streaming setting.