Bounding the Optimal Performance of Online Randomized Primal-Dual Methods

The online randomized primal-dual method has widespread applications in online algorithm design and analysis. A key challenge is identifying an appropriate function space, $F$, in which we search for an optimal updating function $f \in F$ that yields the best possible lower bound on the competitiveness of a given algorithm. The choice of $F$ must balance two competing objectives: on one hand, it should impose sufficient simplifying conditions on $f$ to facilitate worst-case analysis and establish a valid lower bound; on the other hand, it should remain general enough to offer a broad selection of candidate functions. The tradeoff is that any additional constraints on $f$ that can facilitate competitive analysis may also lead to a suboptimal choice, weakening the resulting lower bound. To address this challenge, we propose an auxiliary-LP-based framework capable of effectively approximating the best possible competitiveness achievable when applying the randomized primal-dual method to different function spaces. Specifically, we examine the framework introduced by Huang and Zhang (SICOMP 2024), which analyzes Stochastic Balance for vertex-weighted online matching with stochastic rewards. Our approach yields both lower and upper bounds on the best possible competitiveness attainable using the randomized primal-dual method for different choices of $F$. Notably, we establish that Stochastic Balance achieves a competitiveness of at least $0.5796$ for the problem (under equal vanishing probabilities), improving upon the previous bound of $0.576$ by Huang and Zhang (SICOMP 2024). Meanwhile, our analysis yields an upper bound of $0.5810$ for a function space strictly larger than that considered in Huang and Zhang (SICOMP 2024).

💡 Research Summary

The paper tackles a fundamental issue in the design of online algorithms that rely on the randomized primal‑dual (RPD) framework: how to choose the update function f that splits the primal gain between dual variables. While the RPD method has yielded optimal 1 − 1/e competitive ratios for classic problems such as online matching, Adwords, and ranking, the choice of the function space F from which f is drawn has traditionally been driven by analytical convenience rather than optimality. Imposing strong structural constraints (e.g., fixing f(z) to a constant for large z, requiring piecewise linearity, etc.) simplifies the inner minimization that defines the competitive ratio but may also unnecessarily limit the achievable bound.

To address this trade‑off, the authors introduce an auxiliary‑LP (AUG‑LP) based framework that can both approximate the optimal value of the RPD program for a given F and provide provable upper bounds on the best possible competitive ratio achievable within F. The framework is applied to the vertex‑weighted online matching with stochastic rewards (OM‑SR) problem, specifically to the Stochastic Balance (ST‑BA) algorithm introduced by Huang and Zhang (SICOMP 2024). In this setting, each offline vertex u has an exponential threshold Θ_u, and each arriving online vertex v must be assigned to an available neighbor. The algorithm updates dual variables according to a single‑variable gain‑sharing function f(ℓ_u), where ℓ_u is the current load on u.

The core of the analysis consists of two linear programs:

1. AUG‑LP – For a fixed f, the adversary’s optimal strategy in the inner minimization (the worst‑case expectation of dual constraints) is encoded as a linear program. By discretizing the domain of f and expressing expectations as linear combinations of discretized values, the originally infinite‑dimensional problem becomes a finite LP that can be solved numerically.

2. AUG‑UB‑LP – To bound the optimal value over an entire function space F, the authors approximate the graph of f by a polygonal chain. The slopes and intercepts of this chain become decision variables, and the constraints enforce monotonicity, range