Low congestion online routing and an improved mistake bound for online prediction of graph labeling

In this paper, we show a connection between a certain online low-congestion routing problem and an online prediction of graph labeling. More specifically, we prove that if there exists a routing scheme that guarantees a congestion of $\alpha$ on any edge, there exists an online prediction algorithm with mistake bound $\alpha$ times the cut size, which is the size of the cut induced by the label partitioning of graph vertices. With previous known bound of $O(\log n)$ for $\alpha$ for the routing problem on trees with $n$ vertices, we obtain an improved prediction algorithm for graphs with high effective resistance. In contrast to previous approaches that move the graph problem into problems in vector space using graph Laplacian and rely on the analysis of the perceptron algorithm, our proof are purely combinatorial. Further more, our approach directly generalizes to the case where labels are not binary.

💡 Research Summary

The paper establishes a direct connection between two seemingly unrelated online problems: low‑congestion routing in a graph and online prediction of vertex labels on the same graph. In the prediction setting, a connected graph G = (V,E) is given together with an unknown labeling ℓ: V → L (the paper focuses on binary labels but the analysis works for any finite label set). In each round an adversary queries a vertex v_i; the algorithm outputs a prediction y_i, then learns the true label ℓ(v_i). The goal is to minimize the total number of mistakes, ignoring the first query.

The authors observe that if there exists an online routing algorithm A that, for any sequence of requests r_1,…,r_k, routes each request r_i (i>1) to some previously seen request r_j (j<i) using a path P_i, and guarantees that every edge e is used by at most α paths (i.e., the maximum congestion is α), then one can construct a prediction algorithm P_A with a mistake bound of α·|cut(ℓ)|, where cut(ℓ) = { (u,v)∈E | ℓ(u)≠ℓ(v) } and |cut(ℓ)| is the size of the cut induced by the true labeling. The construction is straightforward: when asked to label v_i, the algorithm runs the routing procedure A to connect v_i to a previously labeled vertex v_j, then predicts ℓ(v_j). If the prediction is wrong, ℓ(v_i)≠ℓ(v_j), which implies that the routing path P_i must contain at least one edge from cut(ℓ). By charging the mistake to any such edge, and noting that each edge can be charged at most α times (by the congestion guarantee), the total number of mistakes cannot exceed α·|cut(ℓ)|. This argument does not rely on the binary nature of the labels, so it extends to multi‑class labeling without any modification.

To obtain a concrete bound, the paper leverages known results on online routing in trees. Aw erbuch and Azar (1995) and later Chalermsook & Fakcharoenphol (2005) showed that for any tree with n vertices there exists an online routing algorithm that guarantees a congestion of O(log n) on every edge. By taking a spanning tree of the original graph G and applying the same algorithm, the same O(log n) congestion bound holds for general graphs. Combining this with the reduction above yields the main result: there exists an efficient online prediction algorithm that makes at most O(log |V|)·|cut(ℓ)| mistakes. For a path (line) graph this bound is optimal; an adversary can force Ω(log n) mistakes, matching the algorithm’s guarantee.

The authors compare their bound with several prior works. Herbster and Pontil (2007) obtained a bound of 4·|cut(ℓ)|·R_G + 2, where R_G is the maximum effective resistance between any pair of vertices. Since R_G can be as large as Θ(n) for a line graph, their bound becomes Θ(n·|cut|), much weaker than the O(log n)·|cut| bound presented here when R_G = Ω(log n). Herbster (2008) introduced a cluster‑based bound N(G,ρ) + 4·|cut(ℓ)|·ρ + 1, which is advantageous only when the graph decomposes into a small number of low‑diameter clusters; for high‑diameter graphs (e.g., lines) the bound again degrades to Ω(n). Pelckmans and Suykens (2008) gave a combinatorial algorithm with a guarantee that only counts edges between pairs of mistaken vertices; this guarantee can be vacuous (e.g., a line graph with a single cut edge) and does not bound the total number of mistakes.

The paper concludes with several open questions. The O(log n) congestion bound, while tight for trees, may be overly pessimistic for dense or well‑connected graphs where many short edge‑disjoint paths exist, potentially allowing much lower congestion. The authors suggest exploring connections between congestion and effective resistance, and investigating whether more recent online routing techniques (e.g., Rӓcke’s hierarchical decompositions, Harrelson et al.’s polynomial‑time decompositions) can be transplanted to improve the mistake bound. Finally, they note that their analysis cannot beat the trivial lower bound of |cut(ℓ)|, and achieving a bound that adapts to the distribution of cut edges (rather than counting each edge uniformly) remains an intriguing direction. Overall, the work provides a clean combinatorial bridge between online routing and online graph labeling, offering a simple, label‑agnostic algorithm with provably better guarantees than many earlier spectral‑based approaches.