Generalized Friendship Paradoxes in Network Science

Generalized friendship paradoxes occur when, on average, our friends have more of some attribute than us. These paradoxes are relevant to many aspects of human interaction, notably in social science and epidemiology. Here, we derive new theoretical results concerning the inevitability of a paradox arising, using a linear algebra perspective. Following the seminal 1991 work of Scott L. Feld, we consider two distinct ways to measure and compare averages, which may be regarded as global and local. For global averaging, we show that a generalized friendship paradox holds for a large family of walk-based centralities, including Katz centrality and total subgraph communicability, and also for nonbacktracking eigenvector centrality. However, we also find counterexamples for centralities based on walks of even length. For local averaging we establish a paradox for nonbacktracking eigenvector centrality and we characterize the cases where the paradox holds with equality for the walk-based case. Defining loneliness as the reciprocal of the number of friends, we show that for this attribute the generalized and local friendship paradoxes always hold in reverse. In this sense, we are always more lonely, on average, than our friends. We also derive global and local averaging paradoxes for the case where the arithmetic mean is replaced by the geometric mean. As well as unifying and adding to the literature in this area, we highlight some open questions.

💡 Research Summary

The paper presents a comprehensive theoretical investigation of generalized friendship paradoxes (GFP) in undirected, unweighted, connected networks, extending the classic observation that “your friends have more friends than you do” to arbitrary node attributes and to two distinct averaging schemes: global averaging (averaging over all nodes) and local averaging (averaging over each node’s ego‑network and then over all nodes).

Background. The original friendship paradox was proved by Feld (1991) using the inequality dᵀd/(1ᵀd) ≥ 1ᵀd/n, where d is the degree vector. Eom and Jo later generalized this to any positive attribute vector x, obtaining dᵀx/(1ᵀd) ≥ 1ᵀx/n, which holds precisely when x is non‑negatively correlated with degree. Higham showed that the paradox also holds when x is the Perron–Frobenius eigenvector of the adjacency matrix (eigenvector centrality). Cantwell et al. introduced a local version, Δ_i = (1/d_i)∑_j a_ij d_j − d_i, and proved that the average of Δ_i over all nodes is non‑negative, with equality only for regular graphs.

Global averaging results. The authors prove a powerful theorem (Theorem 3.1) stating that for any matrix function f that is monotone increasing on the spectrum of A, the attribute x = f(A) 1 satisfies the global GFP inequality. The proof exploits the eigen‑decomposition A = ∑ λ_j v_j v_jᵀ and the weights w_j = (v_jᵀ1)², reducing the inequality to a sum of terms (λ_i − λ_j)(f(λ_i) − f(λ_j)) which are non‑negative by monotonicity. Strict monotonicity yields equality only for regular graphs. This theorem immediately covers walk‑based centralities (x = A^ℓ 1 for odd ℓ), Katz centrality (x = (I − αA)⁻¹ 1 with 0 < α < 1/ρ(A)), and total subgraph communicability (x = exp(A) 1). Counterexamples are provided for even‑ℓ walk centralities, showing that the paradox does not hold universally.

Non‑backtracking eigenvector centrality. The paper introduces the non‑backtracking (Hashimoto) matrix and the associated 2n‑by‑2n eigenvalue problem (15). The dominant eigenvalue λ > 1 yields a centrality vector v_a (normalized so that 1ᵀv_a = 1). Theorem 3.2 proves that the global GFP holds for this centrality. The proof links λ to the reciprocal of the radius of convergence R of the Ihara zeta function (λ = 1/R) and invokes the Tarras conjecture (proved by Saito) that 1/R ≥ (1ᵀd)/n for graphs without degree‑one vertices and not a cycle. Special cases (λ = 1) correspond to regular graphs or cycles, where equality holds.

Local averaging results. Building on prior work, the authors show that the non‑backtracking eigenvector centrality also satisfies the local GFP inequality for all graphs. For walk‑based centralities, they refine the equality conditions (Theorem 8.1): when ℓ is odd, equality occurs only for regular graphs; when ℓ is even, equality occurs for regular or biregular graphs. This clarifies the role of walk length in the local paradox.

Loneliness paradox. Defining loneliness as the reciprocal degree (1/d), the paper proves (Theorems 6.1 and 6.2) that both global and local GFP hold in reverse: on average, our friends are less lonely than we are. This follows directly from the negative correlation between degree and its reciprocal.

Geometric mean version. Replacing the arithmetic mean with the geometric mean, the authors establish (Theorems 7.1 and 7.2) that the paradox still holds for degree, both globally and locally. This demonstrates robustness of the phenomenon to the choice of averaging operator.

Counterexamples and limitations. Section 8.2 presents explicit graphs where even‑ℓ walk centralities violate the global GFP, disproving any claim of universal applicability. The paper also discusses the impact of degree‑one vertices and cycles on the non‑backtracking results.

Empirical validation. Section 9 reports experiments on real‑world social networks (e.g., Facebook, Twitter) and synthetic models (Erdős‑Rényi, Barabási‑Albert, stochastic block models). The numerical results confirm the theoretical predictions: most centralities obey the paradox, while the identified counterexamples behave as expected.

Conclusions and open questions. The work unifies and extends the literature on friendship paradoxes, providing a linear‑algebraic framework that covers a broad class of centralities, both global and local averaging, and even alternative means (geometric). Open problems include characterizing the exact conditions for even‑ℓ walk paradox failures, extending the theory to weighted or temporal networks, and exploiting the paradox for more efficient sampling, immunization, or information‑diffusion strategies.

Overall, the paper delivers a rigorous, unified treatment of generalized friendship paradoxes, deepening our understanding of sampling bias in networks and offering new theoretical tools for network scientists and applied researchers alike.