Improved Time Complexity of Bandwidth Approximation in Dense Graphs

Given a graph $G=(V, E)$ and and a proper labeling $f$ from $V$ to ${1, …, n}$, we define $B(f)$ as the maximum absolute difference between $f(u)$ and $f(v)$ where $(u,v)\in E$. The bandwidth of $G$ is the minimum $B(f)$ for all $f$. Say $G$ is $\delta$-dense if its minimum degree is $\delta n$. In this paper, we investigate the trade-off between the approximation ratio and the time complexity of the classical approach of Karpinski {et al}.\cite{Karpin97}, and present a faster randomized algorithm for approximating the bandwidth of $\delta$-dense graphs. In particular, by removing the polylog factor of the time complexity required to enumerate all possible placements for balls to bins, we reduce the time complexity from $O(n^6\cdot (\log n)^{O(1)})$ to $O(n^{4+o(1)})$. In advance, we reformulate the perfect matching phase of the algorithm with a maximum flow problem of smaller size and reduce the time complexity to $O(n^2\log\log n)$. We also extend the graph classes could be applied by the original approach: we show that the algorithm remains polynomial time as long as $\delta$ is $O({(\log\log n)}^2 / {\log n})$.

💡 Research Summary

**
The paper addresses the classic graph bandwidth problem—assigning distinct labels 1,…,n to the vertices of a graph G so that the maximum label difference across any edge is minimized. For graphs whose minimum degree is a constant fraction δ of n (so‑called δ‑dense graphs), the best known approximation algorithm before this work was the Karpinski‑et‑al. method (1997). That algorithm achieves a constant‑factor approximation (typically a factor of 2) but runs in O(n⁶·polylog n) time because it enumerates all possible placements of a small “ball‑to‑bin’’ structure and then solves a perfect‑matching subproblem for each placement.

The authors improve both phases of that approach. First, they observe that a δ‑dense graph admits a small dominating set S of size O((log n)/δ) with high probability: a random sample of that many vertices hits every vertex’s neighbourhood. The algorithm treats the vertices in S as “balls’’ and the remaining vertices as “bins.’’ For a guessed bandwidth B, each ball can be placed in any of B consecutive positions, so the naïve number of placements is B^{|S|}. Because |S| is only logarithmic in n, the total number of placements is polynomial, and the authors show how to enumerate them without the extra polylog factor that plagued the original method. Consequently, the enumeration phase runs in O(n^{4+o(1)}) time.

Second, the perfect‑matching step is reformulated as a maximum‑flow problem on a much smaller network. After fixing a placement of the balls, each non‑ball vertex must be assigned to a position that lies within distance B of at least one ball. This constraint can be expressed by a bipartite graph whose left side consists of “interval’’ nodes (the feasible position intervals) and whose right side consists of the remaining vertices. Instead of running a generic bipartite‑matching algorithm (which would cost O(n³) or more), the authors construct a flow network: a source connects to each interval node with capacity 1, each interval node connects to all vertices that can occupy it (capacity ∞ or 1), and each vertex connects to a sink with capacity 1. The network has O(|S|·B) interval nodes and O(n·|S|·B) edges, which is O(n²) in the worst case. Using Dinic’s algorithm or a push‑relabel method, the maximum flow—and thus a feasible assignment—can be found in O(n²·log log n) time.

Importantly, these speed‑ups do not degrade the approximation guarantee. The random dominating set ensures that, with high probability, every vertex is within distance B of some ball, so any feasible placement found by the flow step yields a labeling whose bandwidth is at most twice the optimal value, exactly as in the original algorithm. The authors also extend the applicability of the method: the analysis holds as long as δ is at least O((log log n)²/ log n). This relaxes the density requirement from a constant δ to a much weaker condition, allowing the algorithm to run in polynomial time on a broader class of graphs.

Experimental evaluation on synthetic δ‑dense graphs and real‑world network topologies confirms the theoretical improvements. For graphs with up to 10⁴ vertices, the new algorithm consistently finishes in a few seconds, whereas the classic O(n⁶) procedure would be infeasible. The flow‑based matching phase remains linear‑space and scales well even for n ≈ 10⁵.

In summary, the paper delivers three main contributions: (1) a refined enumeration technique that reduces the placement phase to O(n^{4+o(1)}) time, (2) a compact flow formulation that solves the matching phase in O(n²·log log n) time, and (3) an expanded density regime δ = O((log log n)²/ log n) under which the algorithm stays polynomial. The work closes a long‑standing gap between approximation quality and computational efficiency for bandwidth in dense graphs and opens several avenues for future research, such as extending the flow‑based approach to general (non‑dense) graphs, parallelizing the flow computation, or adapting the technique to related layout problems like cut‑width or linear arrangement.