Improved Online Algorithms for Inventory Management Problems with Holding and Delay Costs: Riding the Wave Makes Things Simpler, Stronger, & More General

The Joint Replenishment Problem (JRP) is a classical inventory management problem, that aims to model the trade-off between coordinating orders for multiple commodities (and their cost) with holding costs incurred by meeting demand in advance. Moseley, Niaparast and Ravi introduced a natural online generalization of the JRP in which inventory corresponding to demands may be replenished late, for a delay cost, or early, for a holding cost. They established that when the holding and delay costs are monotone and uniform across demands, there is a 30-competitive algorithm that employs a greedy strategy and a dual-fitting based analysis. We develop a 5-competitive algorithm that handles arbitrary monotone demand-specific holding and delay cost functions, thus simultaneously improving upon the competitive ratio and relaxing the uniformity assumption. Our primal-dual algorithm is in the spirit of the work Buchbinder, Kimbrel, Levi, Makarychev, and Sviridenko, which maintains a wavefront dual solution to decide when to place an order and which items to order. The main twist is in deciding which requests to serve early. In contrast to the work of Moseley et al., which ranks early requests in ascending order of desired service time and serves them until their total holding cost matches the ordering cost incurred for that item, we extend to the non-uniform case by instead ranking in ascending order of when the delay cost of a demand would reach its current holding cost. An important special case of the JRP is the single-item lot-sizing problem. Here, Moseley et al. gave a 3-competitive algorithm when the holding and delay costs are uniform across demands. We provide a new algorithm for which the competitive ratio is $ϕ+1 \approx 2.681$, where $ϕ$ is the golden ratio, which again holds for arbitrary monotone holding-delay costs.

💡 Research Summary

The paper studies the Joint Replenishment Problem (JRP) and its special case, the single‑item lot‑sizing problem, in an online setting where each demand may have its own monotone holding‑and‑delay cost function. Prior work by Moseley, Niaparast and Ravi gave a 30‑competitive algorithm but required that all demands share the same cost functions. This paper removes the uniformity assumption and presents a 5‑competitive online algorithm for the general JRP, as well as a (ϕ + 1) ≈ 2.681‑competitive algorithm for the single‑item case, improving upon the previous 3‑competitive bound.

The authors first formalize the JRP: there are N item types, a fixed ordering cost K₀ for any order, and a per‑item cost Kᵢ for each type i included in an order. Each demand d = (i, t) has a desired service time t and a holding‑delay cost function H_{i,t}(s) that is non‑increasing for s ≤ t (holding cost) and non‑decreasing for s > t (delay cost), with H_{i,t}(t)=0. The goal is to serve every demand while minimizing the sum of ordering costs and incurred holding‑delay costs.

The core technical contribution is a primal‑dual algorithm that maintains a “wavefront” value τ that continuously moves from time 1 to T. For every active demand the dual variable bₜ is increased at a rate equal to the derivative of its holding‑delay function at τ, and for each pair (t, s) the dual variable z_{t,s} is increased at the same rate whenever bₜ ≥ H_{t}(s). When a constraint of the form z_{t,s} ≤ K (or z_{t,i,s} ≤ Kᵢ) becomes tight, the corresponding demand is made inactive, its dual variables are frozen, and an actual replenishment order is placed at the current τ. This process guarantees a feasible dual solution whose objective value is at least 1/5 of the optimal offline cost.

A crucial novelty lies in deciding which future‑time demands to serve early. Instead of sorting demands by their desired service times (as in the earlier 30‑competitive algorithm), the authors compute for each demand the “virtual deadline” – the earliest τ at which its delay cost would equal its current holding cost. Demands are then ranked by increasing virtual deadline, and those whose virtual deadline arrives first are added to the current order. This ensures that an early service is only performed when it is justified by the cost trade‑off, eliminating the pathological cases that caused the previous algorithm’s competitive ratio to blow up under non‑uniform costs.

For the single‑item lot‑sizing problem, the same wavefront framework applies, but the analysis simplifies because there is only one item type. The algorithm again raises bₜ and the associated z_{t,s} until a constraint z_{t,s} ≤ K becomes tight, at which point an order is placed. By carefully bounding the total cost incurred relative to the dual objective, the authors prove a competitive ratio of ϕ + 1, where ϕ ≈ 1.618 is the golden ratio. This improves the earlier 3‑competitive bound (which required uniform costs) and approaches the known lower bound of 2 for the delay‑only version of the problem.

The competitive analysis proceeds in two steps. First, the dual construction shows that the sum of bₜ values (the dual objective) is at least 1/5 of the optimal offline cost for the general JRP (or at least 1/(ϕ+1) for the single‑item case). Second, the algorithm’s actual incurred cost (ordering plus holding‑delay) is shown to be at most 5 times (or ϕ+1 times) the dual objective, because each order corresponds to a tight dual constraint and the wavefront never over‑pays for any demand. Consequently, the overall competitive ratios follow directly.

Beyond the immediate results, the paper highlights that the wavefront primal‑dual technique, combined with the virtual‑deadline ranking, yields a conceptually simpler analysis than the earlier dual‑fitting approach. This simplicity suggests that the method could extend to more complex inventory settings such as multi‑level aggregation, network design with delay, or set‑cover‑type aggregation problems where only monotonicity (not uniformity) is required.

In summary, the authors deliver:

1. A 5‑competitive online algorithm for the JRP with arbitrary monotone, demand‑specific holding‑and‑delay costs.
2. A (ϕ + 1)‑competitive algorithm for the single‑item lot‑sizing problem under the same general cost model.
3. A clean primal‑dual wavefront framework that decides early service based on virtual deadlines, eliminating the need for uniform cost assumptions.
4. Insightful analysis that bridges the gap between the best known upper bounds (30 and 3) and the theoretical lower bounds (2 for delay‑only), moving the field significantly forward.

These contributions advance the theory of online inventory management and provide practical algorithmic tools for systems where cost functions vary across items and time.