Coupled Multicommodity Flow

• Coupled multicommodity flow is a class of models where multiple commodity flows interact through shared resources like edge capacities, node processing, and common dynamic queues.
• The topic spans static formulations with edge, node, and matrix couplings to dynamic models that incorporate shared queues and aggregate congestion costs, addressing a variety of network challenges.
• Algorithmic frameworks and flow–cut dualities provide both exact and approximation solutions, informing practical applications in traffic assignment, telecommunication design, and large-scale network optimization.

Searching arXiv for the cited papers and closely related work on coupled multicommodity flow. Coupled multicommodity flow denotes a family of flow models in which multiple commodities are linked by shared resources or shared state variables, so that routing one commodity changes the feasible or optimal routing of others. In the narrowest usage, the coupling is the standard shared-capacity condition on edges; in broader usages, it includes node processing capacities, node-submodular or polymatroid constraints, quadratic edge energies, aggregate convex congestion costs, common dynamic queues, or shared infrastructure variables. The term therefore spans several technically distinct literatures, including network optimization, approximation algorithms, dynamic traffic assignment, network design, and graph-theoretic enumeration (Charikar et al., 2018, Chekuri et al., 2011, Sering, 2020, Lazzaro et al., 6 Aug 2025).

1. Classical multicommodity flow and the meaning of coupling

In the classical multicommodity flow problem, each commodity is a divisible flow from its own source to its own sink, subject to commodity-wise conservation and edge-capacity constraints. A canonical undirected formulation requires

$$\sum_{i=1}^k |f_i(e)| \le u(e)\qquad \forall e\in E,$$

or, in concurrent-flow form,

$$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$

so that the commodities interact through shared edge congestion (Kelner et al., 2013).

One strand of the literature treats this shared-capacity interaction itself as the coupling mechanism. The logistics formulation of the multicommodity network flow problem, for example, uses commodity-wise conservation together with the arc constraint

$\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$

so that vehicle assignment and capacity on arc $(i,j)$ jointly constrain all commodity flows (Chence et al., 2024). In this sense, coupled multicommodity flow is the standard multicommodity setting viewed from the perspective of resource competition.

A second strand reserves the term for models with additional joint structure beyond ordinary edge capacities. In that usage, the coupling may arise because flow consumes resources on both edges and vertices, because incident edges at a node satisfy a submodular capacity law, because the objective depends on aggregate edge load, or because dynamic queueing and earliest-arrival behavior are shared across commodities (Charikar et al., 2018, Chekuri et al., 2011, Sering, 2020). This broader usage is the one most often associated with specialized theory and algorithms.

2. Static models with edge, node, and matrix coupling

A prominent vertex-coupled model is multicommodity flow with in-network processing. On a directed graph $G=(V,E)$ with edge capacities $B:E\to \mathbb{R}_{\ge 0}$, vertex processing capacities $C:V\to \mathbb{R}_{\ge 0}$, and demands

$D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$

each unit of commodity flow must receive one unit of processing somewhere along its route. The resulting feasibility conditions combine edge capacities, vertex processing capacities, and a processing-completion requirement. In the polynomial edge-based LP, $f_i(e)$ is total flow of commodity $i$ on edge $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$0, $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$1 is its unprocessed portion, and $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$2 is processing assigned at $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$3, with the processing conservation identity

$$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$4

This model uses both links and vertices as consumable resources, and its decomposition requires 2-walks rather than simple paths because processing may occur mid-route (Charikar et al., 2018).

A different node-coupled generalization is the polymatroidal network model. Here capacities are not assigned independently edge-by-edge. In a directed network, each node $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$5 has submodular functions $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$6 on incoming edges and $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$7 on outgoing edges, imposing

$$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$8

In the undirected case there is a single polymatroid $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$9 on $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$0. This framework subsumes edge-capacitated and node-capacitated models and was motivated in part by wireless networks, where incident transmissions interfere and capacities are intrinsically coupled at a node (Chekuri et al., 2011).

A third static generalization replaces scalar edge resistances by matrix-valued edge energies. In quadratically coupled flows, each edge $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$1 carries a positive definite matrix $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$2, and the multicommodity electrical-flow analogue solves

$\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$3

The off-diagonal terms of $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$4 explicitly couple commodities on the same edge, and the associated edge saturation is

$\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$5

This quadratic coupling is not merely a reformulation of the usual $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$6 capacity region; it is an inner primitive used to approximate that region algorithmically (Kelner et al., 2012).

3. Coupling through dynamics, costs, and shared infrastructure

Dynamic traffic models create coupling through common queues and commodity-specific shortest paths. In multi-commodity Nash flows over time, each arc $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$7 has transit time $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$8, capacity $\sum_{k \in K} X_{kij} * L_k \le N_{ij} * W,$9, queue length

$(i,j)$0

waiting time

$(i,j)$1

and exit time

$(i,j)$2

Commodities share the total queue, but route according to their own earliest-arrival functions. The model’s distinctive coupling object is the foreign flow

$(i,j)$3

which enters the multi-commodity thin-flow equations. Existence is proved via multi-commodity thin flows with resetting and an infinite-dimensional variational inequality (Sering, 2020).

Another major class couples commodities through aggregate convex link costs. In convex multicommodity flow on a directed graph $(i,j)$4, the compact splittable formulation uses arc-flow ratios $(i,j)$5, rejection variables $(i,j)$6, capacities $(i,j)$7, and convex increasing arc costs $(i,j)$8, with objective

$(i,j)$9

subject to commodity-wise conservation and

$G=(V,E)$0

The splittable and unsplittable variants differ by allowing fractional or single-path routing for each commodity, but in both cases the cost of commodity $G=(V,E)$1 on arc $G=(V,E)$2 depends on the total load of all commodities on that arc (Beraud-Sudreau et al., 3 Feb 2026).

A related optimal-transport formulation couples commodities through a common conductivity field. Each edge carries a vector flux

$G=(V,E)$3

but all commodities share the same conductivity $G=(V,E)$4. The evolution law

$G=(V,E)$5

depends on the $G=(V,E)$6-norm of the vector of commodity potential differences, and the steady-state scaling is

$G=(V,E)$7

The corresponding constrained optimization problem generalizes one-commodity transport network design and shows that loops can arise from distinguishing different flow types (Lonardi et al., 2020).

Formulation itself can also be used to compress coupling. In the destination-based all-pairs model, all traffic with the same destination is aggregated into one commodity, using a flow matrix $G=(V,E)$8 with conservation

$G=(V,E)$9

and shared edge capacities

$B:E\to \mathbb{R}_{\ge 0}$0

This reduces the number of scalar edge-flow variables from roughly $B:E\to \mathbb{R}_{\ge 0}$1 to $B:E\to \mathbb{R}_{\ge 0}$2, which is central to GPU-oriented large-scale computation (Zhang et al., 29 Jan 2025).

4. Algorithmic frameworks

For the divisible-flow processing model, the max-throughput problem is tractable. The edge-based LP is polynomial-size, equivalent to the walk-based formulation, and admits decomposition into routable 2-walks in

$B:E\to \mathbb{R}_{\ge 0}$3

The same work also gives a multiplicative-weights-update algorithm with running time

$B:E\to \mathbb{R}_{\ge 0}$4

for a $B:E\to \mathbb{R}_{\ge 0}$5-approximation (Charikar et al., 2018).

Quadratically coupled flows were introduced precisely to make high-quality approximation algorithmically accessible. For undirected graphs with $B:E\to \mathbb{R}_{\ge 0}$6 edges and $B:E\to \mathbb{R}_{\ge 0}$7 commodities, the coupled-energy framework yields $B:E\to \mathbb{R}_{\ge 0}$8-approximate algorithms for maximum concurrent flow and maximum weighted multicommodity flow in

$B:E\to \mathbb{R}_{\ge 0}$9

time. The key technical point is that the multicommodity linear systems are not Laplacians, so the algorithm constructs well-conditioned edge matrices and solves the resulting systems via Laplacian preconditioning and preconditioned Chebyshev iteration (Kelner et al., 2012).

Almost-linear-time approximation frameworks then pushed the graph-size dependence further. In undirected graphs, a non-Euclidean gradient-descent framework combined with oblivious routings and flow sparsifiers gives a $C:V\to \mathbb{R}_{\ge 0}$0-approximate maximum concurrent multicommodity flow algorithm with running time

$C:V\to \mathbb{R}_{\ge 0}$1

improving over the earlier $C:V\to \mathbb{R}_{\ge 0}$2 bound (Kelner et al., 2013). On expanders, a localized version of Sherman’s algorithm gives the first local $C:V\to \mathbb{R}_{\ge 0}$3-approximate algorithm with runtime

$C:V\to \mathbb{R}_{\ge 0}$4

in the abstract formulation, and with an expander corollary

$C:V\to \mathbb{R}_{\ge 0}$5

(Li et al., 18 Jan 2025).

Directed graphs require different tools because undirected oblivious-routing structure is unavailable. A recent framework reduces directed multicommodity flow to composite $C:V\to \mathbb{R}_{\ge 0}$6-regression, then to constrained box-simplex games and composite $C:V\to \mathbb{R}_{\ge 0}$7-regression. Its headline guarantee is an almost-linear-time algorithm with

$C:V\to \mathbb{R}_{\ge 0}$8

dependence, up to $C:V\to \mathbb{R}_{\ge 0}$9 factors, for multiplicative $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$0-approximation of concurrent, maximum, and weighted maximum directed multicommodity flow (Chen et al., 31 Mar 2025).

Large-scale implementation has also become a topic in its own right. The destination-aggregated PDHG method for all-pairs flow exploits sparse incidence-matrix operations, columnwise simplex projections, and separable proximal steps on GPUs. Numerical experiments report acceleration of state-of-the-art generic commercial solvers by $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$1 to $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$2, and the method scales to instances with up to a billion variables (Zhang et al., 29 Jan 2025).

5. Flow–cut theory and geometric structure

Coupled multicommodity flow has a distinct flow–cut theory when capacities are node-based or submodular. In polymatroidal networks, the cut cost is itself node-coupled: a cut edge set $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$3 must be assigned to endpoints, and in the undirected case the cut value is

$D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$4

The multicut and sparsest-cut relaxations are expressed using Lovász extensions of the node submodular functions, and the dual of the maximum-throughput flow LP is equivalent to the corresponding convex cut relaxation (Chekuri et al., 2011).

This dual framework yields poly-logarithmic guarantees. In undirected polymatroidal networks, the paper proves

$D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$5

and

$D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$6

For directed symmetric demands on $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$7 pairs, the minimum multicut is bounded by

$D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$8

where $D=\{(s_i,t_i,k_i)\}_{i=1}^{|D|},$9 is the optimum throughput flow value (Chekuri et al., 2011).

Planarity creates a more delicate picture. The classical Okamura–Seymour theorem gives exact equality between cut conditions and feasible concurrent flow in the edge-capacitated single-face setting, but this exact statement fails for node capacities. A planar counterexample has

$f_i(e)$0

Nevertheless, there exists a universal constant $f_i(e)$1 such that if all demands lie on one face, then

$f_i(e)$2

and the same constant-factor statement extends to undirected polymatroid networks (Lee et al., 2012).

A related planar generalization studies demands whose endpoints lie on the same face, though different demands may live on different faces. For separable face instances, the flow-cut gap is at most $f_i(e)$3, proved by approximating face demands by two laminar families with weights $f_i(e)$4 and $f_i(e)$5. For general face instances with $f_i(e)$6 terminals on a face, the paper obtains an $f_i(e)$7 bound (Kumar, 2020).

6. Complexity landscape, applications, and broader usages

The computational status of coupled multicommodity flow is highly model-dependent. Some formulations are polynomial-time solvable. For any fixed digraph $f_i(e)$8, the convex many-commodity transshipment problem with variable number of commodities is polynomial-time solvable via $f_i(e)$9-fold integer programming, even with edge costs of the form

$i$0

and a second polynomial-time regime holds for multicommodity transportation on $i$1 when the numbers of commodities and suppliers are fixed (0906.5106). The divisible in-network processing model is also tractable by exact LP (Charikar et al., 2018).

Other formulations are explicitly NP-hard. The logistics MCNF formulation with binary routing variables $i$2 and vehicle counts $i$3 is a mixed-integer program and is stated to be NP-hard (Chence et al., 2024). The unsplittable convex multicommodity flow problem is likewise NP-hard, motivating stronger branch-and-price relaxations such as $i$4 and $i$5 (Beraud-Sudreau et al., 3 Feb 2026). Even within the processing literature, hardness appears as soon as the problem changes from divisible routing to network design: Directed Min Middlebox Node Purchase is hard to approximate within $i$6, Directed Budgeted Middlebox Node Purchase is hard to approximate better than $i$7, and the undirected purchase variants inherit hardness from Vertex Cover and Max-Cut style constructions (Charikar et al., 2018).

Applications span several network domains. NFV and middlebox placement motivate node-processing models in which each server can host all functions and the optimization decides how to split node resources and route traffic (Charikar et al., 2018). Wireless information flow motivates polymatroidal capacities that couple incident edges at a node (Chekuri et al., 2011). Dynamic traffic assignment motivates multi-commodity Nash flows over time with shared deterministic queues (Sering, 2020). Logistics networks use coupled arc capacities and vehicle assignments (Chence et al., 2024). Telecommunication design uses convex link costs that increase with utilization (Beraud-Sudreau et al., 3 Feb 2026). Optimal transport on adaptive networks studies how shared conductivities shape topology and when loops become optimal (Lonardi et al., 2020).

The term also has broader, non-optimization usages. One line of work replaces hard capacities, or capacities plus conservation, by relaxed pseudo-flow formulations with congestion penalties. In the equilibrium pseudo-flow model, capacities are relaxed and the paper states that zero-equilibrium pseudo-flow corresponds to feasibility whereas nonzero-equilibrium pseudo-flow corresponds to infeasibility (Liu, 2019). In the stable pseudo-flow model, both capacities and conservation are relaxed, edge congestion and vertex heights define a local potential difference

$i$8

and zero-stable versus nonzero-stable pseudo-flow plays the analogous role (Liu, 2021).

In a different direction again, coupled multicommodity flow appears as a generalized nowhere-zero flow in graph and matroid enumeration. For an oriented graph $i$9, a coupled $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$00-multicommodity flow is a tuple $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$01 of edge-flow functions satisfying Kirchhoff’s law together with nested edgewise vanishing implications; for even $$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$02, for example,

$$B^T f_i=\alpha d_i,\qquad \left\|U^{-1}\sum_{i=1}^k |f_i|\right\|_\infty \le 1,$$03

and so on. The counting function is orientation-independent and is evaluated by the chain characteristic polynomial of the dual graphic matroid (Lazzaro et al., 6 Aug 2025).

Taken together, these lines of work show that coupled multicommodity flow is not a single optimization problem but a technical umbrella for multicommodity systems with nontrivial interdependence. The common theme is that commodities cannot be optimized independently: they are linked by shared capacities, shared processing, shared congestion, shared queues, shared conductivities, or shared combinatorial rules. The resulting theory combines exact LP formulations, convex and integer programming, spectral and MWU-based approximation algorithms, flow-cut duality, dynamic equilibrium, and graph-theoretic generalization (Charikar et al., 2018, Chekuri et al., 2011, Kelner et al., 2012, Chen et al., 31 Mar 2025).