Network Design against the Bullwhip Effect in Complex Supply Chains

1. Introduction

• Bullwhip oscillations can raise costs and lead to stockouts or wastage, while supply networks couple multiple firms through heterogeneous transport delays.The introduction also notes that prior network-performance design approaches assume consensus dynamics unlike supply networks’ per-node feedback and heterogeneous delays.Source: §1 · p. 1
• The paper models heterogeneous-delay supply networks as directed acyclic graphs and quantifies bullwhip using the network transfer function’s H∞ norm.The model uses linear ordering policies and a network structure given by a directed acyclic graph.Source: §1 · p. 2
• Demand-to-order responses decompose into sums of serial supply chains along paths connecting nodes.This path-based view accommodates heterogeneous inbound delays, unlike analyses that assume one shared delay across every route.Source: §1 · p. 2
• Reallocating flows and adjusting route delays are proposed as ways to mitigate amplification, beyond changing ordering policies alone.The introduction identifies network structure and delays as design choices that prior policy-focused work left fixed.Source: §1 · pp. 1–2
• The paper develops approximate bullwhip-metric gradients and illustrates network optimization through rerouting and delay adjustment on a 14-node example.The nearest network-design work cited uses a specific demand model and assigns each node a single lead time despite multiple suppliers.Source: §1 · p. 2

2. Problem Setting

• The network is a directed acyclic graph whose edges represent supply routes with positive transport delays; sources have no suppliers and place no orders.Goods flow from supplier to customer, while orders flow in the opposite direction; exogenous demand may enter at any ordering node.Source: §2 · p. 2, Figure 1 · p. 2
• Each node’s demand combines weighted orders from its customers with exogenous demand, and its route weights specify how it splits orders among suppliers.The route weights represent fractions of a node’s orders and must satisfy the network’s weight constraints.Source: §2 · p. 2
• Route-specific deliveries arrive after their respective delays and sum into inventory, while the supply line tracks ordered goods not yet arrived.The ordering policy feeds back credited stock, defined as inventory position plus a fraction of the supply line.Source: §2 · p. 2, Figure 2 · p. 3
• The order-up-to policy applies proportional feedback to stock error with local gain K_i, while β_i determines how much in-transit supply is credited.At β_i = 0 the node ignores in-transit stock; at β_i = 1 it fully accounts for that stock, and β_i < 1 amplifies at any delay.Source: §2 · pp. 2–3
• The analysis assumes a directed acyclic network and linear ordering policies; its linear-deviation results apply to perturbations small enough to avoid negative orders or stock and capacity shortages.Discrete-time analogs replace integrators and delays with their z-transform counterparts, but integer-delay optimization becomes mixed-integer.Source: §2 · p. 3

3.1. Nodal dynamics

• A node’s demand responses share poles, while route weights and delays enter through W_i(s).The nodal dynamics reduce to a single feedback loop.Source: §3.1 · p. 3

3.2. Network dynamics

• In the DAG, demand-to-order responses sum products of nodal responses over paths, and the network is stable exactly when every node is stable.Network poles are exactly the poles of its component nodes.Source: §3.2 · pp. 4–5

3.3. Stability margins

• For a single route, crossover frequency is K, gain margin is π/(2Kτ), and phase margin is π/2−Kτ; increasing gain speeds response but erodes margins.At Kτ=π/2, the phase margin vanishes and peak amplification diverges.Source: §3.3 · p. 5

3.4. Margins with multiple suppliers

• Route responses interfere destructively: |W_i|≤1 at every frequency, with equality only when all routes are in phase.Source: §3.4 · pp. 5–6
• Splitting orders across routes with different delays improves stability margins by attenuating loop magnitude and phase lag; peak amplification falls together with bandwidth along the Pareto front.Optimal policies mix routes of different delays at every margin level.Source: §3.4 · pp. 5–6

4. Supply Chain Network Design

• The design minimizes the network-wide bullwhip metric over flow weights, lead-time reductions, and gains, subject to capacity, delay-floor, disk-margin, and spending constraints.The disk-margin equality selects the largest feasible gain at each node, so gains follow from the weights and delays.Source: §4 · p. 6

4.1. Gradients

• One backward pass computes gradients at each frequency with cost linear in edges; shortening Munich–Rotterdam lowers the peak by −2.1 dB per time unit, while Frankfurt–Madrid raises it by 0.6.Source: §4.1 · p. 7, Figure 7 · p. 7

4.2. Baselines and solution

• At b = 6, cost-minimal routing spends nothing on shortening, while the naive baseline spends its budget shortening long delays; the designs share constraints and disk margin.The comparison solves the design problem with sequential quadratic programming and a smoothed peak objective, then checks peaks on a finer frequency grid.Source: §4.2 · pp. 7–8

4.3. Results

• Routing alone removes 15.3 of the 20.3 dB gained at b = 6; the H∞ design amplifies less than either baseline at every budget.Source: Figure 8 · p. 8, §4.3 · p. 8
• The H∞ design uses 2.2 routes per node, keeps most routes below capacity, and mixes lead times while maintaining the cost-minimal design’s ωg at every budget.At b = 6, 46% of its flow is on routes longer than the median, versus 80% for cost-minimal routing and 28% for the naive design.Source: Figure 9 · p. 8, §4.3 · p. 8
• Without gradient guidance, shortening long routes raises responsiveness but does not reduce amplification, because it makes delays more uniform and removes attenuation from delay differences.Source: §4.3 · p. 8
• After a unit demand step, the H∞ design holds the most amplified supplier’s inventory near its setpoint, settling more slowly than the naive design.The cost-minimal response oscillates throughout the horizon, while the naive response swings several fold.Source: §4.3 · p. 8
• The H∞ design has higher algebraic connectivity than either baseline at every budget, indicating improved resilience to a lost route without a connectivity constraint.Source: §4.3 · pp. 8–9

• The paper models supply networks as DAGs and proves that demand-to-order transfer functions decompose into sums over network paths.Source: §5 · p. 9
• A 14-city example illustrates a systematic method for mitigating network amplification by reallocating flows and minimizing the network H∞ norm.Source: §5 · p. 9
• Future work will extend the results to more general ordering policies, stochastic lead times, capacity limits, and networks with cycles.Source: §5 · p. 9

• The proof uses induction on the topological index, with the base case at j = ℓ given by the single path (ℓ).Source: Appendix A · p. 9
• For a later node j, the proof assumes the path expansion holds for each earlier node k before applying the network relations.Source: Appendix A · p. 9
• Appending j to paths reaching its customers constructs the complete set of paths to j, so the double sum has one term per path.Source: Appendix A · p. 9

• At node i, the chain rule propagates objective sensitivity through its orders and into supplier demands; source nodes have zero demand derivative.Source: Appendix B · pp. 9–10
• Writing ψi = ∂Jρ/∂oi gives ∂Jρ/∂di = Tiψi, which is substituted into the derivative with respect to node orders.Source: Appendix B · p. 9
• At each frequency, the gradient combines weight and delay derivatives of Wi with a routing term, then sums across frequencies and takes real parts.Source: Appendix B · pp. 9–10