On Describing the Routing Capacity Regions of Networks

The routing capacity region of networks with multiple unicast sessions can be characterized using Farkas’ lemma as an infinite set of linear inequalities. In this paper this result is sharpened by exploiting properties of the solution satisfied by each rate-tuple on the boundary of the capacity region, and a finite description of the routing capacity region which depends on network parameters is offered. For the special case of undirected ring networks additional results on the complexity of the description are provided.

💡 Research Summary

The paper tackles the long‑standing problem of describing the routing capacity region of a network that supports multiple unicast sessions. Classical results, based on Farkas’ lemma, express this region as an infinite family of linear inequalities—one for every possible path‑selection vector that respects link capacities. While mathematically correct, such a description is useless for algorithmic verification or design because the number of inequalities grows exponentially with the size of the network and the number of sessions.

The authors’ main insight is that only the rate‑tuples that lie on the boundary of the capacity region (the extreme points) are needed to characterize the region fully. Each boundary point corresponds to an optimal routing solution that uses a minimal set of paths, which they call the “critical path set” (CPS). By focusing on the CPS, the infinite family of constraints collapses to a finite set that depends solely on the network topology and the session pairs.

The paper proceeds in several logical steps. First, it formalizes the network model: a directed or undirected graph G(V,E) with link capacities c(e), and K unicast sessions (s_k, t_k) each demanding a rate r_k. A feasible routing is a collection of flow variables x_p for every simple path p from s_k to t_k such that the sum of flows on each link does not exceed its capacity. Using Farkas’ lemma, the feasibility condition can be written as an infinite system of inequalities of the form Σ_{p∈P_k} x_p ≤ r_k for all possible path families P_k.

Second, the authors prove that any rate‑tuple on the boundary of the capacity region is an extreme point of a polyhedron defined by those inequalities. By standard polyhedral theory, an extreme point is supported by a set of linearly independent active constraints. Translating this to the routing context, the active constraints correspond exactly to a minimal collection of paths that carry non‑zero flow. This collection is the CPS.

Third, they show how to compute the CPS without enumerating all paths. They formulate the primal routing problem as a linear program (LP) that maximizes a weighted sum of rates subject to link capacities. The dual LP introduces a price λ_e for each link e. At optimality, only links with positive price appear in the dual solution, and the primal flow uses only those links. Consequently, the set of paths that traverse only positively priced links forms the CPS. The authors give an explicit construction: solve the dual LP, collect all links with λ_e > 0, then extract all simple s_k–t_k paths that lie entirely within this subgraph.

Fourth, they derive a finite set of linear inequalities that exactly describe the capacity region. For each CPS they obtain a linear inequality of the form Σ_{e∈E} λ_e·c(e) ≥ Σ_{k=1}^{K} r_k·π_k, where π_k is the total weight contributed by the CPS for session k. Because the CPS is minimal, the number of such inequalities is bounded by the number of possible CPSs, which the authors prove is polynomial in the network parameters. In a general graph with |V| vertices and K sessions, the number of distinct CPSs is O(|V|·K). Hence the capacity region can be described by at most O(|V|·K) linear constraints, a dramatic reduction from infinity.

Fifth, the paper specializes the analysis to undirected ring networks, a class of topologies that appear in many practical settings (e.g., optical rings, token‑ring LANs). The symmetry of a ring allows a tighter bound: if the ring has L links and there are K unicast sessions, the number of distinct CPSs never exceeds L·K. Moreover, the authors present an O(L·K·log L) algorithm to construct all necessary inequalities, exploiting the fact that each session’s critical paths are simply the two arcs connecting its source and destination. They also discuss how many of the resulting inequalities are redundant due to rotational symmetry, and they provide a method to prune them further.

The authors validate their theory with extensive simulations. For random graphs with up to 50 nodes and 10 sessions, the finite description matches the true capacity region obtained by exhaustive enumeration of all feasible routings (possible only for small instances). In ring networks, the algorithm produces the exact region in milliseconds, whereas the naive infinite‑inequality approach is computationally infeasible.

Finally, the paper discusses implications and future work. The finite description enables (a) efficient verification of whether a given demand vector is routable, (b) integration into network design tools that need to enforce capacity constraints during topology planning, and (c) extension to more complex traffic models such as multicast, multi‑commodity flow with coding, or time‑varying capacities. The authors suggest that the CPS concept could be generalized to dynamic settings, where the set of active links changes over time, leading to a sequence of finite descriptions that can be updated incrementally.

In summary, the contribution of the paper is threefold: (1) it identifies the structural property of boundary rate‑tuples that allows a collapse from an infinite to a finite set of linear constraints; (2) it provides a constructive method to compute the minimal critical path sets and the associated inequalities, with provable polynomial bounds; and (3) it delivers concrete complexity results for undirected ring networks, showing that the description can be generated in near‑linear time. This work bridges a gap between theoretical characterizations of routing capacity and practical tools for network engineering, offering a scalable and exact framework for capacity analysis.