Discontinuous transition to loop formation in optimal supply networks

Supply networks in biology (leaf venation, vasculature) and technology (power grids, hydraulic systems) often balance a trade-off between construction/maintenance cost and resilience to failures or fluctuations. Real networks frequently exhibit loops that enhance topological resilience, yet loops are costly. Prior numerical work showed loops emerge when inputs fluctuate strongly or links can be damaged, but the precise mechanism for loop formation in minimal-dissipation networks remained unclear. This study investigates when and how the first loops form in optimal linear flow networks under cost constraints, focusing on two resilience settings—fluctuating sources/sinks and stochastic edge damage—and asks whether loop formation occurs continuously or discontinuously, and how to predict the location and conditions for the first loop.

The paper situates its work within research on optimal transport and supply networks across biology and engineering, including vascular and venation networks, power grids, hydraulic networks, and drainage basins. Previous studies established trade-offs between cost and resilience and observed looped vs. treelike structures, with loops arising under fluctuations or damage (e.g., Katifori et al. 2010; Corson 2010; Bohn & Magnasco 2007). Additional contexts include resilience in traffic and communication networks, structural resilience via percolation, and optimization approaches in hydraulic engineering. The authors highlight that while loop emergence with fluctuations or damage has been observed numerically, the theoretical mechanism (type of transition) and predictive criteria for first-loop formation in minimal-dissipation networks remained to be clarified.

• Network model: Linear flow network on a graph with node injections/withdrawals P_n, oriented edges e with capacities k_e and flows F_e. Flows satisfy Kirchhoff’s current law (sum of incident flows equals P_n) and are linear in nodal potential differences: F_e = k_e (θ_n − θ_m). Kirchhoff’s voltage law ensures zero loop sum of impedances times flows.
• Objective: Minimize total dissipation D = Σ_e F_e^2 / k_e subject to a global resource constraint Σ_e k_e = K (cost parameter γ enters the optimality conditions via the known scaling relation; Lagrange multiplier method yields a self-consistency: k_e ∝ ⟨F_e^2⟩^{1/γ}).
• Two resilience models:
  1. Fluctuating sink model: P_n are random (balanced in each realization). Minimize average dissipation ⟨D⟩ = Σ_e ⟨F_e^2⟩/k_e over realizations, with capacities fixed across realizations. Using Lagrange multipliers gives the self-consistency equation for k_e in terms of ⟨F_e^2⟩.
  2. Edge-damage model: Deterministic sources/sinks (single source supplies N−1 sinks). Partial damage on a single edge l modeled as k_e → (1−Δ_e) k_e with Δ_l = Δ ∈ (0,1), Δ_e=0 for e≠l. Minimize average dissipation over all singleton damage scenarios under the same resource constraint.
• Small-network analysis: Five-node example with an optional fifth edge closing a loop. Without the loop (tree), dissipation ⟨D_tree⟩ is computed explicitly in terms of sink statistics and γ. With the loop, a cycle-flow degree of freedom f is introduced and eliminated using KVL, yielding a dissipation expression depending only on capacities, enabling direct minimization.
• Generalization to arbitrary trees with one added edge (single-loop case): • In a tree, flows F_e depend only on topology and P via a tree matrix T; capacities do not affect flows in trees. • Adding a new edge l=(m,n) defines edge sets L and R along paths from source to m and n; the cycle flow f is solved from KVL. The loopy dissipation can be written as D_loopy = Σ_e∈T F_e^2/k_e + k^{-1} (Σ_{e∈R} F_e/k_e − Σ_{e∈L} F_e/k_e)^2. Averaging and rewriting introduce B_{m,n} and C_{m,n} terms depending on capacities along L∪R. • Optimization with inequality x ≥ 0 for the new-edge capacity uses Karush-Kuhn-Tucker (KKT) conditions. The KKT condition for the new edge gives either x=0 or a finite positive value depending on B_{m,n}, C_{m,n}, γ, and the Lagrange multiplier.
• Analytical results: • Theorem: For γ∈(0,1), the tree (x=0) is always a KKT point and is isolated—loopy minima cannot bifurcate continuously from the tree solution. • Approximation to predict first-loop location: assume capacities on L∪R change by a near-unity factor upon adding a small x; then loop benefit is governed by the squared potential drop across the prospective new edge B(k_e) ≈ (θ_m − θ_n)^2 computed on the optimal tree.
• Topological predictor for dissipation: Derive an expression for tree-network dissipation at local minima in terms of edge betweenness centrality N_2(e) with i.i.d. Gaussian sinks (using moments ⟨P_i^2⟩ and ⟨P_i P_j⟩). This yields D_tree ∝ Σ_e (N_2(e) σ^2 + N_2(e)^2 μ^2)^{1/2} and motivates using betweenness to rank local minima in loopy networks.
• Numerical procedures: For larger grids (triangular lattice potential edges), start from the optimal tree, add all non-tree edges with tiny capacity, renormalize to satisfy the resource constraint, and apply an iterative relaxation method (from Corson 2010) to converge to local minima. Track loop formation as parameters (γ, σ, Δ) vary and compare with pressure-drop predictions. Edge betweenness computed via NetworkX implementations.

• Discontinuous loop formation: In both the fluctuating sink model and the edge-damage model, the transition from a treelike to a loopy optimal network is discontinuous. The capacity of the new loop-forming edge jumps from zero to a finite value at the transition.
• Saddle-node bifurcation mechanism: New loopy minima arise via saddle-node bifurcations, not via continuous bifurcations from the tree solution. This holds when varying either the fluctuation strength or the cost parameter, and also in the damage model when varying damage fraction or cost.
• Tree remains a local optimum: For γ<1, the tree (no-loop) solution is always an isolated KKT point, i.e., a local minimum persists even after loopy minima emerge.
• Predicting loop location and threshold: The average squared potential (pressure) drop across candidate edges in the optimal tree accurately predicts where the first loops will form and correlates strongly with the critical cost parameter at which loops become beneficial.
• Beyond the first loop: The discontinuous character persists for subsequent loops in larger networks; each new loop appears with a finite nonzero capacity as costs decrease or fluctuations increase.
• Betweenness-dissipation link: An explicit relation ties minimal dissipation in tree networks to edge betweenness centrality. This provides a topological heuristic to rank local minima in loopy networks; strong correlation between the tree-based betweenness estimate and actual dissipations is observed, even with many loops.
• Practical implication: Simple topological/pressure-drop measures can predict loop formation in biological and engineered supply networks, informing design strategies balancing cost and resilience.

The study addresses the open question of how loops emerge in optimal supply networks under cost constraints and resilience considerations. By proving that loop formation occurs via saddle-node bifurcations, it explains the observed discontinuities in optimal capacities at the onset of redundancy. This clarifies the mechanism across distinct resilience models (fluctuation-driven and damage-driven), indicating generality. The identification of pressure drop as a predictor connects mechanical stress relief concepts to vascular development and provides a practical rule for where to add redundancy. The derived connection between edge betweenness and minimal dissipation links network topology to function, enabling heuristic ranking of candidate structures without fully solving the coupled flow-capacity problem. These findings bear on the design of resilient infrastructures (power and water networks) and the interpretation of biological network morphologies, complementing structural resilience insights from percolation theory with flow-based criteria.

The paper establishes that loops in dissipation-optimized supply networks emerge discontinuously via saddle-node bifurcations, with the tree solution remaining an isolated local optimum. It provides analytical and numerical evidence in both fluctuating-sink and edge-damage models, shows that discontinuity persists beyond the first loop, and introduces predictive metrics—average pressure drop for loop location/threshold and edge betweenness for estimating dissipation at minima. These results bridge topological measures and functional optimization, offering practical guidance for adding redundancy in biological and engineered networks. Future work could extend beyond linear flow assumptions, explore broader cost functions and constraints, analyze non-Gaussian or correlated fluctuations, and apply the predictive framework to empirical systems at scale.

• Linear flow assumption: The analysis relies on linear relationships between flows and potentials; applicability to nonlinear flow regimes is not addressed.
• Resource constraint form: Optimization focuses on dissipation minimization with a specific global capacity constraint; although similar transitions are noted for alternative formulations, results may depend on cost modeling details.
• Analytical scope: Closed-form solutions for optimal capacities in near-tree networks were not obtained for arbitrary topologies and γ; predictions for loop emergence use approximations (e.g., small changes in capacities along L and R).
• Validation domain: Numerical demonstrations primarily use planar triangular grids and specific parameter regimes; while a simple nonplanar case is shown in supplementary material, broad generalization to diverse real-world topologies is not exhaustively tested.