Network Coding Capacity: A Functional Dependence Bound

Explicit characterization and computation of the multi-source network coding capacity region (or even bounds) is long standing open problem. In fact, finding the capacity region requires determination of the set of all entropic vectors $\Gamma^{*}$, which is known to be an extremely hard problem. On the other hand, calculating the explicitly known linear programming bound is very hard in practice due to an exponential growth in complexity as a function of network size. We give a new, easily computable outer bound, based on characterization of all functional dependencies in networks. We also show that the proposed bound is tighter than some known bounds.

💡 Research Summary

The paper addresses the longstanding challenge of characterizing and computing the capacity region for multi‑source network coding. Existing formulations require the set of all entropic vectors Γ* which is intractable for more than three random variables, and the known linear‑programming (LP) bound, while theoretically exact, suffers from an exponential blow‑up in the number of variables and constraints as the network grows. To overcome these obstacles, the authors introduce the notions of pseudo‑variables and pseudo‑entropy functions, which extend the classical entropy framework without requiring an underlying probability distribution. A pseudo‑entropy function g is required only to satisfy the polymatroid axioms (non‑decreasing, submodular, and zero at the empty set); it may or may not correspond to any real random variables. This abstraction allows the authors to work within a finite‑dimensional polyhedral space while retaining the essential information‑theoretic properties needed for network coding analysis.

The central contribution is the Functional Dependence Graph (FDG). An FDG is a directed graph whose vertices correspond to pseudo‑variables and whose edges encode local functional dependencies: for each vertex i, the pseudo‑entropy conditioned on its parent set π(i) must be zero, i.e., g(X_i | π(i)) = 0. Unlike earlier definitions, this FDG permits cycles, does not distinguish source from non‑source variables, and can represent additional functional relationships beyond those explicitly drawn. The authors define a graphical procedure, denoted A → B, that determines whether a set A functionally determines a set B by iteratively deleting outgoing edges from A and then repeatedly removing vertices with no incoming edges. If after this process no vertices of B remain, then A → B. Lemma 1 (Grandparent Lemma) and Theorem 1 prove that this graph‑based test correctly captures global functional dependence: whenever A → B, the conditional pseudo‑entropy g(B | A) vanishes.

Building on this, the paper introduces the concept of irreducible sets—subsets of vertices that cannot be further reduced by any proper subset determining them. Maximal irreducible sets are those that cannot be enlarged without losing the irreducibility property. For acyclic FDGs, every maximal irreducible set has the same pseudo‑entropy, and they can be enumerated efficiently using a recursive algorithm (Algorithm 1). For cyclic FDGs, a modified definition and a second recursive algorithm (Algorithm 2) are provided. The authors illustrate the process on the classic butterfly network, enumerating all maximal irreducible sets of its FDG.

The main theoretical result, Theorem 2 (Functional Dependence Bound), states that for any network coding instance defined by the usual constraints (source independence, encoding functions, decoding requirements, and edge capacity limits) together with an FDG G on the pseudo‑variables, the sum of source entropies is bounded above by the minimum total capacity over all maximal irreducible sets that contain no source variables. Formally,
∑{s∈S} g(Y_s) ≤ min{B∈𝔅_M} ∑_{e∈B} C_e,
where 𝔅_M denotes the collection of maximal irreducible sets of G excluding source nodes. The proof follows directly from the definition of irreducibility (which forces the pseudo‑entropy of the set B to equal that of the sources) and the sub‑additivity of pseudo‑entropy together with the edge‑capacity constraints.

Crucially, this bound is provably tighter than the previously known Network Sharing bound and the Information Dominance bound. Moreover, because the bound reduces to a combinatorial search over maximal irreducible sets, its computational complexity is polynomial in the size of the FDG, avoiding the exponential blow‑up of the full LP formulation. The authors demonstrate the improvement on the butterfly network example, where the new bound matches the known capacity while the older bounds are looser.

In conclusion, the paper provides a novel, graph‑theoretic framework that translates the intricate functional relationships inherent in network coding into a tractable combinatorial problem. By leveraging pseudo‑entropy and functional dependence graphs, the authors deliver an easily computable outer bound that is both tighter than existing bounds and scalable to larger networks. This work opens a promising direction for future research on practical capacity estimation, code design, and the exploration of tighter inner bounds that may approach the presented functional dependence bound.