Randomised Buffer Management with Bounded Delay against Adaptive Adversary

We give a new analysis of the RMix algorithm by Chin et al. for the Buffer Management with Bounded Delay problem (or online scheduling of unit jobs to maximise weighted throughput). Unlike the original proof of e/(e-1)-competitiveness, the new one holds even in adaptive-online adversary model. In fact, the proof works also for a slightly more general problem studied by Bie{'n}kowski et al.

💡 Research Summary

The paper revisits the classic Buffer Management with Bounded Delay problem, also known as online scheduling of unit‑length jobs to maximize weighted throughput. In this setting, jobs arrive over time, each job carries a weight and a deadline (the bounded delay), and a scheduler can transmit at most one job per time slot. The objective is to maximize the total weight of transmitted jobs.

The algorithm under study is RMix, originally introduced by Chin, Halldórsson, and Koutsoupias. RMix is a randomized online algorithm that, in each round, mixes two actions: it either transmits the heaviest pending job or, with a certain probability, transmits a job chosen according to a weight‑biased random rule. The original analysis proved that RMix achieves an expected competitive ratio of e/(e‑1) against a static online adversary – an adversary that fixes the entire input sequence in advance and then reveals it online.

The main contribution of this work is a new proof that the same competitive ratio holds against a much stronger adaptive online adversary. An adaptive adversary observes the algorithm’s past random choices and can generate future jobs (including their weights and deadlines) in response. This model captures realistic network environments where traffic patterns may be shaped dynamically by an opponent.

To obtain the stronger guarantee, the authors introduce a carefully crafted potential function Φ that captures both the current buffer state and the expected future contribution of jobs that may still arrive. The analysis proceeds in two complementary steps:

1. Forward‑expectation bound – For any round t, the expected weight of the job transmitted by RMix is at least (e‑1)/e times the weight that an optimal offline algorithm could obtain in that same round, even when the optimal algorithm is allowed to know the adversary’s current move. This step relies on the random mixing of RMix: the probability of picking the heaviest job is calibrated so that, in expectation, the algorithm “covers” a constant fraction of the optimal gain.

2. Potential‑drop guarantee – The potential function Φ is shown to decrease by at least a factor of (e‑1)/e in expectation after each round, regardless of how the adaptive adversary reacts. Intuitively, the decrease reflects the fact that the algorithm either removes a heavy job from the buffer or, when it transmits a lighter job, the random choice still reduces the expected future contribution of the remaining jobs.

Summing the forward‑expectation bound over all rounds and telescoping the potential‑drop inequality yields the classic e/(e‑1) competitive ratio for the total expected throughput.

The paper also demonstrates that the same analytical framework extends to a more general model studied by Bieńkowski et al., where job weights may change over time or where additional constraints (e.g., multiple parallel servers) are present. By augmenting the potential function with extra terms that capture weight dynamics, the authors prove that RMix’s competitive ratio remains unchanged as long as weight variations are bounded.

Beyond the theoretical contribution, the authors discuss practical implications. In real‑time networking, routers often employ buffer‑management policies that must operate under adversarial traffic (e.g., denial‑of‑service attacks). The new proof shows that a simple randomized rule like RMix is robust not only to worst‑case static traffic but also to traffic that adapts to the router’s past decisions. This robustness can be leveraged in the design of lightweight, provably near‑optimal scheduling modules for high‑speed switches, cloud‑based job dispatchers, and streaming platforms where latency constraints translate into bounded‑delay requirements.

In summary, the paper provides a clean, self‑contained proof that the RMix algorithm retains its e/(e‑1) expected competitive ratio against adaptive online adversaries and in a broader class of dynamic weight settings. The result bridges a gap between classic competitive analysis and more realistic adversarial models, reinforcing the relevance of randomized buffer‑management strategies in modern, security‑aware networked systems.