Joint Replenishment Problem

• Joint Replenishment Problem (JRP) is a supply chain challenge that coordinates orders across multiple items by incurring joint fixed costs and item-dependent costs.
• It balances economies of scale from consolidated orders against various temporal penalties such as holding, waiting, and backlog costs.
• Recent approaches employ deterministic models, LP rounding, EPTAS, and online algorithms to tackle its computational complexity and enhance efficiency.

Joint Replenishment Problem (JRP) is a fundamental inventory-management and supply-chain optimization problem in which multiple commodities share a common replenishment mechanism, so each order incurs a joint fixed cost together with item-dependent fixed costs for the commodities included in that order. Its central trade-off is between economies of scale from coordinated replenishment and temporal service penalties, which may take the form of waiting costs, holding costs, backlog costs, or hard deadlines. Across the literature, the term covers continuous-time periodic models, finite-horizon deterministic lot-sizing models, deadline and waiting-cost variants, online and non-clairvoyant formulations, and extensions integrating scheduling, stochastic demand, fairness constraints, and routing-style structure (Segev, 2024, Nagarajan et al., 2015, Ezra et al., 2024).

1. Core formulation and cost structure

In the classical two-level supply-chain view, goods move from a supplier to a shared warehouse and then to multiple retailers. A shipment to the warehouse incurs a fixed joint cost, and each retailer or item that joins the shipment incurs an additional fixed participation cost. In the deadline formulation JRP-D, each demand is a triple $(\rho,r,d)$ with retailer $\rho$, release time $r$, and deadline $d$; an order $(t,R)$ satisfies that demand when $\rho\in R$ and $t\in[r,d]$. More generally, JRP with waiting costs attaches to each order $\pi=(\rho,a,h)$ a nondecreasing waiting-cost function $h(t)$ for $t\ge a$, and the objective is the sum of shipping and waiting costs (Bienkowski et al., 2012, Bienkowski et al., 2013).

In the classical continuous-time deterministic periodic model, each item $\rho$0 has ordering cost $\rho$1, holding coefficient $\rho$2, and replenishment interval $\rho$3. The single-item EOQ building block is

$\rho$4

and the $\rho$5-item JRP objective is

$\rho$6

Here $\rho$7 is the long-run average joint-ordering cost induced by the union of all replenishment epochs, and the model assumes deterministic stationary demand, no shortages, and long-run average cost minimization (Segev, 2024, Segev, 23 Jun 2025).

Finite-horizon deterministic JRP is often expressed with a set function $\rho$8 for ordering any subset $\rho$9 of items in a period. In the classical additive case,

$r$0

whereas submodular JRP allows a more general monotone submodular setup-cost function. With demand points $r$1 and no backlogging or lost sales, one chooses order periods and assignments of demands to earlier orders to minimize ordering plus holding cost (Nagarajan et al., 2015).

2. Deterministic offline optimization and approximation

For finite-horizon deterministic JRP with submodular ordering costs and polynomial holding costs, an LP-rounding framework based on shadow intervals, interval stretching, and grouping by interval width yields an $r$2-approximation. In the linear holding-cost case this becomes the first sub-logarithmic approximation guarantee for submodular JRP, and in the fixed-lifetime perishable special case the same framework gives a 2-approximation (Nagarajan et al., 2015).

For JRP-D, the standard LP relaxation has a refined approximation theory. A randomized polynomial-time algorithm achieves approximation ratio $r$3, improving the previous $r$4 bound. The same paper establishes integrality-gap lower bounds of $r$5 and a stronger computer-assisted $r$6, together with an upper bound of $r$7. In the special case where all demand periods have equal length, it gives a $r$8-approximation, a lower bound of $r$9, and APX-hardness (Bienkowski et al., 2012).

A more recent offline branch studies rejection and fairness. In colorful or fairness-constrained JRP, each rejected demand consumes budget in one or more protected-feature dimensions, and the goal is to control rejections separately across those dimensions. For constant $d$0, the number of features, the first constant-factor approximation algorithms are a deterministic $d$1-approximation for general holding costs and a $d$2-approximation in the deadline case. These guarantees also improve prior bounds for the outlier-only special cases (Suriyanarayana et al., 2023).

For continuous-time deterministic JRP, recent work substantially improves the classical power-of-2 regime. In the variable-base model, an EPTAS achieves a $d$3-approximation in time

$d$4

while in the fixed-base model one obtains

$d$5

The same line shows that optimal evenly-spaced policies approximate JRP within factor $d$6, strictly improving on the longstanding power-of-2 factor $d$7 (Segev, 2024). A complementary development gives an EPTAS for near-optimal dynamic policies in continuous and discrete time: every continuous-time infinite-horizon instance can be reduced to a corresponding discrete-time $d$8-period instance with multiplicative loss at most $d$9, and the discrete-time problem admits a $(t,R)$0-approximation in time

$(t,R)$1

(Segev, 23 Jun 2025).

3. Complexity landscape

The computational complexity of periodic deterministic JRP has been a long-standing theme. For the stationary-demand, integer-cycle, infinite-horizon periodic JRP (PJRP), the problem is strongly NP-hard. The same reduction framework also shows the finite-horizon version is NP-hard, though not strongly NP-hard by that argument because the constructed horizon is not polynomially bounded (Cohen et al., 2015).

The continuous counterpart is also hard. The Continuous Periodic Joint Replenishment Problem (CPJRP), in the continuous-time, infinite-horizon, stationary-demand setting, is strongly NP-hard. This closes the complexity gap left open after hardness had been established for the discrete periodic version (Tuisov et al., 2020).

Hardness already appears at extremely small item counts. For two-item discrete-time JRP, both the aperiodic version with objective

$(t,R)$2

and the periodic version with objective

$(t,R)$3

are at least as hard as integer factorization. Under the assumption that the Riemann Hypothesis is correct, the periodic two-item variant is NP-hard under randomized reductions (Schulz et al., 2020).

Approximation hardness is also well developed. JRP is APX-hard, and the natural LP for JRP has integrality gap at least $(t,R)$4; both statements hold even for JRP-D (Bienkowski et al., 2013). In addition, the equal-length-demand-period case of JRP-D is APX-hard (Bienkowski et al., 2012).

4. Online, delay, and non-clairvoyant formulations

Online JRP has several distinct objective models. For arbitrary waiting-cost functions, the offline approximation ratio $(t,R)$5 breaks the earlier $(t,R)$6 barrier, while on the online side the competitive ratio for JRP-L, the linear-waiting-cost model, has lower bound $(t,R)$7. For online JRP-D, the optimal competitive ratio is 2 (Bienkowski et al., 2013).

A major development is the online model with both holding and backlog costs. In that setting, each request has an arrival time and a soft deadline; serving before deadline incurs holding cost, serving after deadline incurs backlog cost, and every service incurs a fixed joint service cost plus fixed item-dependent costs. The first constant-competitive results for this general model are a 3-competitive deterministic polynomial-time algorithm for the single-item case and a 30-competitive deterministic polynomial-time algorithm for the multi-item case, obtained by a greedy algorithm analyzed via dual fitting (Moseley et al., 2024).

Subsequent work removes the earlier uniformity assumptions. For arbitrary request-dependent monotone holding-cost and backlog-cost functions, one paper gives a 4-competitive deterministic algorithm for the single-item case and a 16-competitive deterministic algorithm for the general multi-item case. Its key conceptual move is to replace static deadline order by dynamic priority via virtual deadlines, so the served set need not be a deadline prefix (Azar et al., 21 Jul 2025). A later primal-dual wavefront algorithm improves these guarantees to 5-competitive for general JRP and $(t,R)$8 for the single-item lot-sizing special case, again for arbitrary monotone demand-specific holding and delay costs (Shmoys et al., 12 Feb 2026).

Another online direction studies delayed-service JRP under non-clairvoyance. In the non-clairvoyant subadditive model, requests have type $(t,R)$9, arrival time $\rho\in R$0, and nondecreasing delay-cost function $\rho\in R$1, while serving a set of types incurs a monotone subadditive cost $\rho\in R$2 with $\rho\in R$3. A deterministic $\rho\in R$4-competitive algorithm is obtained by approximating $\rho\in R$5 by a disjoint service function

$\rho\in R$6

and then reducing to independent TCP Acknowledgement instances. The same modular framework gives tight $\rho\in R$7-competitive algorithms for Multi-Level Aggregation and Weighted Symmetric Subadditive JRP (Ezra et al., 2024).

Prediction augmentation has also entered the online theory. In nonclairvoyant JRP-D with deadline predictions, the error is measured by the number $\rho\in R$8 of instantaneous item inversions, and a deterministic algorithm achieves competitive ratio

$\rho\in R$9

This yields the standard learning-augmented properties of consistency and robustness: $t\in[r,d]$0 with inversion-free predictions, and never worse than the best known $t\in[r,d]$1 nonclairvoyant guarantee (Dinitz et al., 20 Nov 2025).

5. Scheduling integration and stochastic control

JRP has also been integrated with single-machine scheduling. In one model, jobs have release dates, processing times, and required subsets of resource types; a job can start only after every required resource has been replenished in the interval from its release to its start. The resulting objectives include

$t\in[r,d]$2

This integrated problem is NP-hard in several restricted cases, but it is polynomially solvable for some structured regimes, including constant numbers of resources with unit or equal processing times, and one-resource $t\in[r,d]$3 variants. The paper also gives 2-competitive online algorithms for one-resource, unit-processing-time versions with $t\in[r,d]$4 and $t\in[r,d]$5 objectives (Györgyi et al., 2021).

A closely related one-item model couples replenishment with unit-time single-machine scheduling and minimizes replenishment cost plus maximum flow time. With release dates assumed integral and distinct, each replenishment costs $t\in[r,d]$6, and a job may start at time $t\in[r,d]$7 only if there exists a replenishment time $t\in[r,d]$8. A deterministic threshold policy is 2-competitive on arbitrary input, asymptotically $t\in[r,d]$9-competitive on $\pi=(\rho,a,h)$0-bounded input, and supported by lower bounds of $\pi=(\rho,a,h)$1, $\pi=(\rho,a,h)$2, and $\pi=(\rho,a,h)$3 under different assumptions (Györgyi et al., 2022).

Stochastic JRP appears in periodic-review, finite-horizon form under uncertain demand. A periodic-review stochastic JRP under Bookbinder and Tan’s static-dynamic uncertainty control policy fixes replenishment timing at the start of the horizon and dynamically orders up to prescribed levels in replenishment periods. A mixed-integer linear approximation based on first-order loss functions extends the earlier single-item $\pi=(\rho,a,h)$4 model to multi-item joint replenishment, and is used both as a heuristic policy class and as an approximation to the optimal $\pi=(\rho,a,h)$5 policy. The framework addresses nonstationary stochastic demand and includes a shortest-path reformulation for large stationary benchmark instances (Xiang et al., 2019).

6. Methodological themes and current frontiers

Several algorithmic ideas recur across the JRP literature. Deterministic finite-horizon models rely heavily on time-indexed LPs and structural rounding, including shadow intervals, extended shadow intervals, and width grouping for submodular JRP, as well as randomized LP rounding for JRP-D through the tally game and Wald’s Lemma (Nagarajan et al., 2015, Bienkowski et al., 2012). Fairness-constrained and outlier variants strengthen the natural LP by guessing a bounded amount of side information and then combine iterative rounding with pipage-style rounding (Suriyanarayana et al., 2023).

Online and non-clairvoyant work uses a different toolkit. Dual fitting underlies the first constant-competitive holding/backlog algorithm, primal-dual wavefront methods support the later 5-competitive arbitrary-cost result, and virtual deadlines provide a dynamic priority rule when actual deadlines are no longer sufficient (Moseley et al., 2024, Shmoys et al., 12 Feb 2026). For non-clairvoyant subadditive JRP, the main abstraction is a structural approximation of the service-cost function by disjoint functions, followed by a reduction to independent TCP Acknowledgement instances (Ezra et al., 2024).

Current limitations remain model-specific. For arbitrary subadditive non-clairvoyant JRP, the best general ratio still depends on the $\pi=(\rho,a,h)$6-stretch universal set cover bound, and the USC-based reduction is exponential-time in the fully general case (Ezra et al., 2024). In continuous-time deterministic JRP, the variable-base model has an EPTAS, but the fixed-base model is still described in terms of a $\pi=(\rho,a,h)$7 guarantee rather than an approximation scheme (Segev, 2024). For dynamic policies, the recent EPTAS does not resolve the exact hardness of optimization over all dynamic policies, and no lower bounds are known for dynamic JRP complexity in that line of work (Segev, 23 Jun 2025). For fairness-constrained JRP, constant-factor approximation requires the number of features $\pi=(\rho,a,h)$8 to be constant (Suriyanarayana et al., 2023). This suggests that the field’s frontier is no longer a single approximation ratio, but a collection of structural questions about policy classes, service-cost geometry, side constraints, and information models.