Critical Nodes In Directed Networks

Critical nodes or “middlemen” have an essential place in both social and economic networks when considering the flow of information and trade. This paper extends the concept of critical nodes to directed networks. We identify strong and weak middlemen. Node contestability is introduced as a form of competition in networks; a duality between uncontested intermediaries and middlemen is established. The brokerage power of middlemen is formally expressed and a general algorithm is constructed to measure the brokerage power of each node from the networks adjacency matrix. Augmentations of the brokerage power measure are discussed to encapsulate relevant centrality measures. We use these concepts to identify and measure middlemen in two empirical socio-economic networks, the elite marriage network of Renaissance Florence and Krackhardt’s advice network.

💡 Research Summary

The paper extends the notion of critical nodes, traditionally studied in undirected graphs, to directed networks where the direction of information or trade flows matters. A node h is defined as an (i, j)‑middleman if there exists at least one directed walk from i to j and h lies on every such walk. A node that fulfills this condition for any pair of distinct nodes is called a middleman. The authors distinguish two subclasses: strong middlemen, whose removal splits the network into two or more weakly‑connected components, and weak middlemen, whose removal leaves the overall network weakly connected but still destroys at least one directed i→j path.

To capture the competitive environment, the concept of contestability is introduced. A node is contested if alternative directed walks exist that can replace the flow it mediates; otherwise it is uncontested. The central theoretical result is a duality: a node is a middleman if and only if it is uncontested. This formalizes the economic intuition that middlemen can extract rents precisely when no alternative routes are available.

Building on this, the authors propose a quantitative “middle‑man power” measure. For each ordered pair (i, j) they compute the set M_{ij}(D) of all middlemen and count how many pairs are brokered by a given node. The measure assigns zero to non‑middlemen and a positive value proportional to the number of directed pairs whose communication depends exclusively on the node. An algorithmic implementation uses the adjacency matrix A: by summing powers A^k for k = 1 to n the algorithm obtains all reachable walks, then checks node‑wise inclusion in each (i, j) walk set. The procedure runs in polynomial time and can be applied to large‑scale directed graphs.

Empirically, the framework is applied to two classic socio‑economic networks. In the Florentine marriage network (Padgett & Ansell, 1993) the middle‑man power scores correctly identify the Medici and other historically dominant families as the most powerful brokers, confirming the method’s historical validity. In Krackhardt’s advice network (1987) the node identified as having the highest middle‑man power coincides with the “advice hub” originally highlighted by Krackhardt, while traditional centrality metrics such as betweenness and Bonacich centrality either over‑ or under‑estimate its influence.

A systematic comparison shows that betweenness centrality captures the frequency with which a node lies on shortest paths but ignores directionality and the existence of alternative longer routes; Bonacich centrality reflects the centrality of neighbours but does not directly measure exclusive control over directed communication. The proposed middle‑man power metric fills this gap by explicitly quantifying exclusive brokerage in directed settings.

The paper concludes that recognizing strong and weak middlemen, understanding their contestability, and measuring their brokerage power provide a richer analytical toolkit for economists, sociologists, and network scientists. The matrix‑based algorithm makes the approach computationally tractable, opening avenues for policy analysis (e.g., identifying critical infrastructure nodes), organizational design (e.g., spotting key information brokers), and further theoretical work on network robustness and power dynamics.