REVIEW 3 major objections 4 minor 1 cited by
The Triangle Friendship Paradox
T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Triangle counts break the friendship paradox: the average bias can be negative.
desk verdict The negative-bias message holds up, but the configuration-model section is the softest part—nonstandard triangle convention plus an unverifiable finite-n formula. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the triangle friendship-bias $\Delta_T[n]$ of (1.1)--(1.2), the average over vertices of the difference between the average triangle count of neighbours and the vertex's own triangle count. The argument uses a covariance representation, $\Delta_x[n]=\mathrm{Cov}(x_U,\kappa_U)$ with $\kappa_i=\sum_j A_{ij}/d_j$, to convert the sign of the bias into a correlation question; local weak convergence to pass from finite sparse graphs to rooted limiting graphs; and the graphon functional $\chi_T=\int_0^1 dx\,[D(x)^{-1}\int_0^1 dy\,\kappa(x,y)T(y)-T(x)]$ for dense random graphs. The rank-1 graphon case and a specific two-block graphon are the examples that settle when the paradox holds or fails in the dense regime.
What would settle it
Compute $\mathbb{E}[\Delta_T[n]]$ by exhaustive enumeration for a small sparse binomial random graph, say $n=10$ or $n=20$, and compare with the closed formula (1.5); any disagreement would refute Theorem 1.10(a). Alternatively, simulate the two-block graphon with $\alpha=0$, $\beta=1/4$, $\gamma=1/2$ and $p=10/33$ for large $n$ and check whether $n^{-2}\Delta_T[n]$ is negative as the paper predicts.
Extended reading notes
Core claim
The central claim is that the average triangle friendship-bias, defined by $\Delta_T[n]=\frac1n\sum_i\bigl[\frac1{d_i}\sum_j A_{ij}t_j-t_i\bigr]$, can be negative, unlike the degree and wedge versions of the friendship paradox. The paper proves that for partially completed star-graphs the bias is non-negative with an explicit closed form; for the sparse binomial random graph $G(n,\lambda/n)$ it proves $\lim_{n\to\infty} n\,\mathbb{E}[\Delta_T[n]]=\lambda^2-\lambda+(\lambda-\tfrac12\lambda^3)e^{-\lambda}$, which is strictly positive; for sparse configuration models it proves the limit $\zeta(c_1,c_2,c_3)=\frac{(c_2-c_1)^2}{2c_1^4}[(c_3-c_1c_2)-3(c_2-c_1^2)]\ge 0$, with equality only in the regular case; and for dense rank-1 graphons it proves $\chi_T\ge 0$, while giving a two-block graphon with $\chi_T<0$. In both sparse models the weak law of large numbers fails, because triangle-free realisations occur with positive probability in the limit.
Load-bearing premise
For the configuration-model theorems, the paper counts triangles only from triples of distinct vertices and drops the self-loop term in the bias formula, a convention that would change the finite-$n$ expectation, the limiting value, and the nonnegativity conclusion if the standard multi-graph counting including self-loops were used.
Editorial extensions
If this is right
- For locally tree-like sparse graphs the average triangle friendship-bias tends to zero, since triangle counts vanish in the local limit; nonzero bias requires visible triangle structure.
- In sparse binomial random graphs the expected bias is positive for large $n$ and grows like a strictly increasing function of the edge density, but with positive limiting probability the bias is exactly zero because the graph is triangle-free.
- In sparse configuration models the bias is asymptotically non-negative for all degree distributions satisfying the moment condition, with equality only for regular graphs.
- For dense random graphs, the bias scales like $n^2$ and its sign is controlled by the graphon functional $\chi_T$; rank-1 weighted graphs always satisfy the paradox, while some two-block stochastic block models violate it.
- Gluing two partially completed star-graphs together at a non-centre vertex preserves the non-negativity of the bias, but gluing four of them at a common vertex can produce a negative bias.
Reading between the lines
- The covariance interpretation suggests a general design principle: for any attribute, the average friendship-bias has the sign of the correlation between that attribute and the neighbour-averaging weight $\kappa_i$, so triangle counts fail precisely when high-triangle vertices are not the vertices with large $\kappa_i$.
- The negative two-block example implies that real networks with strong community structure could exhibit an anti-friendship paradox for triangles, where vertices inside dense communities have more triangles than their neighbours on average; this is a testable prediction for empirical network data.
- The limiting formulas could be turned into a cheap statistical test: from one large graph, estimate $\chi_T$ by averaging local triangle-ratio quantities, and use its sign as a diagnostic for whether the triangle friendship paradox holds in that network.
- The paper leaves heavy-tailed degree distributions, such as preferential attachment graphs, open; since triangles there concentrate in a small core, one might expect the bias to be positive, but that is an extrapolation beyond what the paper proves.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies the generalized friendship paradox for the attribute "number of triangles at a vertex". It defines the average triangle friendship-bias ΔT[n] in (1.1)-(1.2) and shows that, in contrast to degree- and wedge-based versions, this quantity is not always nonnegative. The negative claim is supported by a deterministic counterexample (Figure 2) and by two-block graphons (Theorem 1.13(b)). The paper also proves nonnegativity for partially completed star-graphs and their gluings (Theorems 1.6-1.7), derives an exact expectation and a limit ζ(λ) for sparse Erdős-Rényi graphs (Theorem 1.10), and gives a finite-n formula plus a nonnegative limit ζ(c1,c2,c3) for sparse configuration models under a stated triangle-counting convention (Theorem 1.11). Dense graphon limits are treated in Theorems 1.12 and 1.13.
Significance. The qualitative finding — that triangle-based generalized friendship paradox can fail — is a clean and useful addition to the friendship-paradox literature, and the deterministic counterexample is simple and convincing. The partially completed star-graph analysis and the sparse Erdős-Rényi computation are explicit and largely checkable from displayed sums, and the paper is honest about which computations are omitted or heuristic. The configuration-model theorem, however, is the least secure: its finite-n formula is supplied on request rather than proved, and its convention for counting triangles is nonstandard. If the configuration-model computation is provided and the convention is made precise in the theorem statement, the paper would be a solid contribution; in its current form the positive sparse-configuration-model claim is not independently verifiable.
major comments (3)
- [Section 3.3, Theorem 1.11(a)] The finite-n formula is not derived in the manuscript. After the displayed sums below (3.13), the proof states that the formula follows by a 'tedious but straightforward computation (available from the authors on request)'. Since the limiting formula in Theorem 1.11(b) is obtained by taking the limit of this unstated expression, the nonnegativity claim for sparse configuration models cannot be checked from the paper. Please include the full algebra in an appendix or state Theorem 1.11 as conditional on a supplied computation.
- [Paragraph before Theorem 1.11 and Eq. (3.9)] The configuration-model result is proved for a nonstandard triangle-counting convention. The text defines t_i with i≠j≠k, excluding repeated-vertex configurations, while retaining the A_ii t_i term in (1.1) because A_ii may be nonzero. Under the usual counting of triangles in a multigraph, which includes loops and repeated-vertex configurations, the finite-n expectation, the limit ζ(c1,c2,c3), and the nonnegativity conclusion are not established and will in general change. The theorem statement should either be explicitly conditional on the convention or should be reworked for the standard counting convention.
- [Section 3.1, proof of Theorem 1.9] The proof claims that ΔT[n](G_n) = E[t_{U_n}(G_n) | G_n], which is false: the left-hand side is the average of the triangle friendship biases, not the average triangle count. Moreover, local convergence in probability of rooted graphs would make the average of a bounded local functional converge in probability to the deterministic expectation under the limiting rooted law, not to the random variable ΔT as stated. The theorem and its proof should be corrected; Theorems 1.10 and 1.11 do not rely on this proof.
minor comments (4)
- [Section 1.3.2, after Theorem 1.10] The stated asymptotic ζ(λ) ∼ 1/2 λ^4 as λ ↓ 0 is inconsistent with (1.6); expanding (1.6) gives ζ(λ) = (1/3)λ^4 + O(λ^5).
- [Remark 4.1] The numerical example with p=0.4, α=0.1, β=0.3, γ=0.8 does not reproduce χT = −8.24 from (4.3); evaluating (4.3) for these parameters gives a value of order −10^−4. Please correct the value or clarify the computation.
- [Section 1.3.3, Dense Configuration Model] The sentence 'The formula in Theorem 1.10(a) is valid for any choice of degrees as long as n is finite' appears to refer to Theorem 1.11(a), since Theorem 1.10(a) is the sparse Erdős-Rényi formula, not a configuration-model formula.
- [Definition 1.5 and Figure 4] The notation in Figure 4, G(2,2,{2,5}), is inconsistent with the caption text 'two bands of adjacent triangles of width 2 and 4'; the second band should be labelled 4 rather than 5 if the notation matches the definition.
Circularity Check
No significant circularity: all central results are derived from the stated definitions and model assumptions, with no fitted parameters or load-bearing self-citations.
full rationale
The paper's Triangle Friendship Bias is defined in (1.1)-(1.2) as an explicit functional of the adjacency matrix and triangle counts, and each theorem is a direct computation or limit of that functional under a clearly stated graph model. The deterministic results (Theorems 1.6 and 1.7) are proved by summing the vertex-level biases according to vertex types. The sparse Erdős-Rényi result (Theorem 1.10) is an independent combinatorial expectation calculation leading to the closed-form ζ(λ). The configuration-model result (Theorem 1.11) is also a combinatorial computation under an explicitly stated convention about how triangles are counted; this convention is a modeling choice, not a parameter fitted to the conclusion. The dense graphon limit (Theorem 1.12) follows from the a.s. cut-metric convergence of the empirical graphon, and Theorem 1.13 is a direct substitution into (1.9)-(1.10). No fitted quantity is renamed as a prediction, and no load-bearing argument reduces to a self-citation. The only self-citations are background references (e.g., [7]) and are not used to force a result. The noted gaps, such as the finite-n configuration-model algebra being 'available from the authors upon request' and the stated normalization issues in some dense examples, are matters of completeness and correctness of presentation, not circularity. The central negative result is supported by explicit deterministic and two-block graphon counterexamples that do not depend on any fitted input.
Assumptions & free parameters
assumptions (5)
- standard math Local weak convergence framework and notation from van der Hofstad's book are used to define convergence and to justify Theorem 1.9.
- standard math Empirical graphons of G(n, κ) converge to κ in the cut-metric almost surely, and dominated convergence transfers this to the TFB limit.
- standard math The number of triangles in sparse ERRG and sparse CM converges to a Poisson distribution in the limit.
- ad hoc to paper For configuration-model results, triangle counts exclude self-loops and A_ii contributions; the TFB uses (1.1)-(1.2) with this convention.
- domain assumption In the dense graphon theorems, κ is continuous and its degree density D(x) is strictly positive for every x.
Cite this review
Pith. "Pith review of The Triangle Friendship Paradox." pith.science (2026). https://pith.science/paper/RU3ZT7QJ
@misc{pith2026250702627,
author = {Pith},
title = {Pith review of: The Triangle Friendship Paradox},
year = {2026},
howpublished = {\url{https://pith.science/paper/RU3ZT7QJ}},
note = {Machine review of arXiv:2507.02627}
}
read the original abstract
We consider the generalised friendship paradox, focussing on the number of triangles at a vertex as the relevant attribute. We show that, contrary to the setting where the attribute is the number of edges at a vertex or the number of wedges at a vertex, the average friendship-bias of the number of triangles at a vertex is not always non-negative. We identify classes of finite deterministic graphs for which the bias is non-negative, and provide examples of finite deterministic graphs for which it is not. For certain classes of sparse and dense random graphs, we compute the scaling of the bias in the limit as the number of vertices tends to infinity.
Figures
Figures from the paper (6 more)
Forward citations
Cited by 1 Pith paper
-
The Generalized Friendship Paradox for Eigenvectors
For inhomogeneous Erdős–Rényi graphs generated by a continuous graphon, the empirical distribution of eigenvector friendship bias converges to the law of (ρ/K(U)−1)φ(U), where ρ and φ are the principal eigenvalue and ...
Reference graph
Works this paper leans on
- [10]
-
[1]
S.L. Feld. Why your friends have more friends than you do. Am. J. Sociol. , 96(6):1464– 1477, 1991. 28
work page 1991
-
[2]
N.A. Christakis and J.H. Fowler. Social network sensors for early detection of contagious outbreaks. PloS one , 5(9):e12948, 2010
work page 2010
-
[3]
B. Nettasinghe and V. Krishnamurthy. “what do your friends think?”: Efficient polling methods for networks using friendship paradox. IEEE Transactions on Knowledge and Data Engineering, 33(3):1291–1305, 2021
work page 2021
-
[4]
G.T. Cantwell, A. Kirkley, and M.E.J. Newman. The friendship paradox in real and model networks. J. Complex Netw. , 9(2):Paper No. cnab011, 2021
work page 2021
-
[5]
S. Pal, F. Yu, Y. Novick, A. Swami, and A. Bar-Noy. A study on the friendship paradox– quantitative analysis and relationship with assortative mixing. Appl. Netw. Sci. , 4(1):1– 26, 2019
work page 2019
-
[6]
Y. Cao and S.M. Ross. The friendship paradox. Math. Sci. , 41(1):61–64, 2016
work page 2016
-
[7]
R.S. Hazra, F. den Hollander, and A. Parvaneh. The friendship paradox for sparse random graphs. Probab. Theory Relat. Fields, 2025. https://doi.org/10.1007/s00440-025-01365-w
Show all 16 references
-
[8]
Eom and H.-H
Y.-H. Eom and H.-H. Jo. Generalized friendship paradox in complex networks: The case of scientific collaboration. Scientific Reports, 4(1):4603, 2014
2014
-
[9]
Hodas, F
N. Hodas, F. Kooti, and K. Lerman. Friendship paradox redux: Your friends are more interesting than you. Proceedings of the International AAAI Conference on Web and Social Media, 7(1):225–233, 2021
2021
-
[11]
van der Hofstad
R. van der Hofstad. Random Graphs and Complex Networks , volume 2 of Cambridge Series in Statistical and Probabilistic Mathematics . Cambridge University Press, 2024
2024
-
[12]
van der Hofstad
R. van der Hofstad. Random Graphs and Complex Networks , volume 1 of Cambridge Series in Statistical and Probabilistic Mathematics . Cambridge University Press, 2017
2017
-
[13]
Lov´ asz
L. Lov´ asz. Large Networks and Graph Limits , volume 60 of American Mathematical Society Colloquium Publications. American Mathematical Society, Providence, RI, 2012
2012
-
[14]
Dhara and S
S. Dhara and S. Sen. Large deviation for uniform graphs with given degrees. Ann. Appl. Probab., 32(3):2327–2353, 2022
2022
-
[15]
Markering
M. Markering. The large deviation principle for inhomogeneous Erd˝ os-R´ enyi random graphs. J. Theoret. Probab., 36(4):711–727, 2023
2023
-
[16]
Athreya and A
S. Athreya and A. R¨ ollin. Dense graph limits under respondent-driven sampling. Ann. Appl. Probab., 26(4):2193–2210, 2016. 29
2016
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.