Online Graph Balancing and the Power of Two Choices

Abstract: In the classic online graph balancing problem, edges arrive sequentially and must be oriented immediately upon arrival, to minimize the maximum in-degree. For adversarial arrivals, the natural greedy algorithm is $O(\log n)$-competitive, and this bound is the best possible for any algorithm, even with randomization. We study this problem in the i.i.d. model where a base graph $G$ is known in advance and each arrival is an independent uniformly random edge of $G$. This model generalizes the standard power-of-two choices setting, corresponding to $G = K_n$, where the greedy algorithm achieves an $O(\log!\log n)$ guarantee. We ask whether a similar bound is possible for arbitrary base graphs. While the greedy algorithm is optimal for adversarial arrivals and also for i.i.d. arrivals from regular base graphs (such as $G = K_n$), we show that it can perform poorly in general: there exist mildly irregular graphs $G$ for which greedy is $\widetildeΩ(\log n)$-competitive under i.i.d. arrivals. In sharp contrast, our main result is an $O(\log!\log n)$-competitive online algorithm for every base graph $G$; this is optimal up to constant factors, since an $Ω(\log!\log n)$ lower bound already holds even for the complete graph $G = K_n$. The key new idea is a notion of log-skewness for graphs, which captures the irregular substructures in $G$ that force the offline optimum to be large. Moreover, we show that any base graph can be decomposed into ``skew-biregular'' pieces at only $O(\log!\log n)$ scales of log-skewness, and use this to design a decomposition-based variant of greedy that is $O(\log!\log n)$-competitive.

• The paper introduces log-skewness to quantify irregular subgraph structures that critically impact online load balancing under i.i.d. arrivals.
• It decomposes any graph into skew-biregular subgraphs, enabling a Threshold-Greedy algorithm to achieve an optimal O(log log n) competitive ratio.
• The work reveals that even mildly irregular graphs can significantly impair standard greedy methods, necessitating structural adaptations in online scenarios.

Online Graph Balancing and the Power of Two Choices: A Technical Overview

Problem Formulation and Context

This paper addresses the online graph balancing problem under stochastic arrivals, providing a sharp characterization of the achievable competitive ratio. The problem considers a known base graph $G$ on $n$ vertices; edges arrive sequentially, each sampled uniformly at random (i.i.d.\ model) from $E(G)$. Upon each arrival, the edge must be irrevocably oriented toward one of its endpoints, with the goal of minimizing the maximum in-degree (maximum vertex load) experienced at any node.

The adversarial version, where the sequence of arrivals is controlled by an adversary, admits a tight $O(\log n)$-competitive greedy algorithm, and this is best possible even with randomization. However, in practice, arrival sequences often follow a stochastic process, and the classical “power of two choices” literature (particularly in balls-into-bins) demonstrates significant improvements in this less pessimistic i.i.d.\ regime. When $G$ is the complete graph $K_n$, the greedy algorithm achieves an $O(\log\!\log n)$ bound, which is tight for randomized algorithms.

The question investigated is: For arbitrary base graphs $G$, can one achieve a similarly strong $O(\log\!\log n)$ competitive ratio in the i.i.d.\ setting?

Limitations of Greedy and Introduction of Log-Skewness

While greedy orientation is optimal for both adversarial arrivals and i.i.d.\ arrivals from (almost) regular base graphs (as shown in "Balanced Allocation on Graphs" [kenthapadi2006balanced]), this paper proves that greedy can perform suboptimally for even mildly irregular $G$. Specifically, there exist graphs where the greedy algorithm is $n$0-competitive under i.i.d.\ arrivals, while the offline optimum remains $n$1.

To address this, the authors introduce the log-skewness parameter $n$2, capturing irregular bipartite substructures that can inflate the offline optimum. Log-skewness, formally, measures the maximum imbalance (in a logarithmic sense) across bipartite subgraphs, normalized by minimum left degree and average degree. They demonstrate that $n$3, tightly characterizing the offline lower bound in terms of irregular subgraph structure.

Decomposition via Skew-Biregular Subgraphs

Building on the log-skewness concept, the authors analyze which graph classes admit $n$4-competitiveness via greedy-like strategies. While greedy fails in general, they identify the critical role of skew-biregular subgraphs—bipartite subgraphs in which the degree imbalance is precisely controlled so that each side’s vertex degrees scale proportionally. On these subgraphs, greedy achieves optimal $n$5 competitiveness.

The main structural result demonstrates that any base graph can be decomposed, in almost-linear time, into $n$6 edge-disjoint skew-biregular subgraphs. This decomposition, with layers corresponding to progressively greater log-skewness, reduces the general problem to a small number of well-understood instances. The decomposition is shown to be tight up to constants; the layer count cannot be improved, as evidenced by the lower bound constructions.

The Threshold-Greedy Algorithm and Analysis

The central online algorithm is Threshold-Greedy, which simultaneously handles all decomposition pieces without paying an additional (iterative) $n$7 penalty. For each base scale (decomposition class), left vertices are permitted to absorb a bounded number of “threshold” loads (i.e., initial arrivals among their incident edges), selected by class. All subsequent arrivals are handled greedily, but with the load computation excluding the threshold phases. The precise thresholds are carefully calibrated based on $n$8 and $n$9 at each class, ensuring the witness tree analysis remains subcritical.

The technical core of the analysis establishes that, with high probability, the greedy (non-threshold) portions of the sampled graph have all connected components bounded by $E(G)$0, leveraging both the decomposition properties and log-skewness bottleneck. This extends classical witness tree (branching process) arguments from regular to highly irregular, decomposed base graphs, demonstrating crucial independence and negative association properties under the i.i.d.\ model for the probability estimation.

The global result is that, for any $E(G)$1 and any number of arrivals $E(G)$2, Threshold-Greedy is $E(G)$3-competitive. This matches the lower bound up to multiplicative constants for $E(G)$4.

Lower Bound for Greedy

A significant structural lower bound is also proved: For certain mildly irregular graphs, even with all vertex degrees within the range $E(G)$5, the greedy algorithm (even with random tie-breaking) attains load $E(G)$6 while an offline optimum is still $E(G)$7. The construction is a layered bipartite graph with calibrated degree growth, confirming that irregularity cannot be ignored and that previously developed regularity-based techniques do not suffice.

Implications and Future Directions

The results of this paper demonstrate that the i.i.d.\ model is strictly more tractable than the adversarial order model when full base graph knowledge is available, but only when online algorithms can adapt to skew and underlying structure—standard greedy is insufficient. The introduction of log-skewness as a unifying parameter will guide future work in online load balancing and stochastic network orientations, especially in sparse, irregular, or partially specified settings.

Pragmatically, the decomposition and Threshold-Greedy framework can be viewed as a variant of policy design robust to hidden structure, and suggests possible extensions to related domains:

• Online load balancing on unrelated machines, extending beyond the 2-choice bipartite case to hypergraphs or $E(G)$8-choices, remains open.
• The impact of partial or noisy knowledge of the base graph, addressing learning-augmented algorithms where the underlying graph is not fully known or must be inferred online, is highlighted, especially given the contrast to the substantially weaker guarantees known for random-order arrivals (where only $E(G)$9 competitiveness is achievable).

Conclusion

This paper resolves the stochastic (i.i.d.) online graph balancing problem on arbitrary base graphs, determining the optimal competitive ratio up to constant factors. Threshold-Greedy, combined with a canonical decomposition into log-skewness scales, achieves $O(\log n)$0-competitiveness provided the base graph is known, even in the presence of significant irregularity. The work exposes structural limitations of greedy and provides a new fine-grained invariant (log-skewness), which likely has broader applications in stochastic and online combinatorial optimization.

Reference: "Online Graph Balancing and the Power of Two Choices" (2604.04159)