Pith. sign in

REVIEW 4 major objections 3 minor 19 references

Perceived Fairness in Networks

T0 review · 4 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A decision rule can satisfy demographic parity population-wide and still appear systematically discriminatory when individuals judge fairness by comparing their own outcome with the average outcome of their network neighbors.

desk verdict Useful formal framework for perceived fairness in networks, but the central theorem that DP rules yield a linear perception gap is not proven—the paper's own expansion cancels under DP and the rescue mechanism is asserted, not derived. read the letter →

arxiv 2510.12028 v2 pith:BL7KCCRF submitted 2025-10-14 econ.TH cs.GT

Arthur Charpentier This is my paper · ORCID
classification econ.THcs.GT MSC 91D3005C80
keywords perceivedfairnessdemographicparityhomophilystochasticblockmodelfriendshipparadoxnetworktopologyalgorithmicassortativemixing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that fairness as experienced by people is a local, network-dependent quantity rather than a global statistic. It defines perceived fairness as the fraction of individuals in a group whose own outcome is at least the average outcome of their neighbors, and studies how this quantity behaves under homophily. In a two-group network with assortative mixing, the paper claims that even a rule satisfying demographic parity produces a perceived fairness gap between groups that grows linearly with the homophily index. It also proves that perceived fairness converges to objective fairness as comparison neighborhoods expand, that degree-outcome correlation creates exposure bias, and that clustering dampens the dispersion of perceived fairness. If the main theorem is right, standard fairness audits that check only population-level parity can certify a rule while a perception of discrimination persists in segregated networks.

What carries the argument

The central object is the fairness perception indicator F^(d)(i;h) = 1{E_i^(d)[h] <= h(i)}, which records whether individual i's own acceptance probability is at least the average acceptance probability in its d-neighborhood. The argument is carried by the contrast between two exposure averages: the node average and the edge-weighted average, whose difference equals Cov(d,h) / E[d], the friendship-paradox identity. Homophily enters through a two-block stochastic block model parameterized by rho = (p_in - p_out)/(p_in + p_out), and a first-order expansion of the group-level perception probability in rho is what converts topology into a perceived fairness gap. The degree-exposure mechanism des

What would settle it

Simulate a two-block network with equal group averages and outcomes independent of degree, then measure the perceived fairness gap as the homophily index increases; if the gap does not grow linearly, the asserted degree-exposure mechanism fails. Alternatively, compute the exact first-order coefficient in rho under demographic parity — the paper leaves it as an unspecified kappa(pi_A, pi_B) — and check whether it is nonzero.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 3.1: in a homophilous two-block stochastic block model, the expected depth-1 perceived fairness gap E[Delta_1(h)] between groups equals c(pi_A, pi_B) rho Gamma(h) + o(rho), so it is generically non-zero and grows linearly with homophily even when the decision rule h satisfies demographic parity. The paper interprets this as evidence that perceived discrimination is not a sign of a biased decision rule but an emergent property of local observation: people benchmark themselves against neighbors, and homophily makes those benchmarks differ systematically by group. A corollary (Proposition 3.2) bounds the gap by the network's modularity, and Proposition 4.4 s

Load-bearing premise

The central result relies on an unproved degree-exposure lemma: with equal average outcomes between groups, neighbor sampling biased by degree must still create a perceived gap that grows with segregation — if that mechanism fails, parity-fair rules show no such gap.

Editorial extensions

If this is right

  • Fairness audits that ignore network structure can approve a rule that every group in a segregated network perceives as systematically unfair.
  • The perceived fairness gap grows with the homophily index, so more segregated interaction networks deepen perceived discrimination without any change in the decision rule.
  • Deepening the comparison neighborhood reverses the effect: on a connected graph, perceived fairness converges to demographic parity as the neighborhood radius grows.
  • If outcomes are positively correlated with degree, the average person sees neighbors with better outcomes than the population average, a downward bias in perceived fairness.
  • Degree-equalizing or clustering-reducing rewires weaken the dispersion of perceived fairness, so network design can shape fairness experience without changing allocations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • This suggests a direct empirical test: in lending, insurance, or workplace data, self-reported perceived discrimination should track neighborhood composition and degree, not just global outcome gaps.
  • A policy consequence the paper leaves implicit is that changing what people can compare — for instance, by publishing global statistics or adding cross-group ties — could reduce perceived unfairness without altering the allocation rule.
  • The linear-response technique should extend to directed, weighted, or temporally evolving networks; the degree-covariance term would become a measure of how strongly outcome correlates with neighbor-sampling probability.
  • The convergence theorem implies a measurement caveat: in large sparse networks the diameter is large, so at practical observation depths the perceived gap can remain substantial even though it eventually vanishes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes a network-based model of 'perceived fairness,' in which individuals compare their own outcome to the average outcome of their network neighbors. The main advertised claim is that even a decision rule satisfying demographic parity (DP) can generate a nonzero and, under homophily, linearly growing perceived fairness gap. The paper proves that as neighborhoods grow, perceived fairness converges to objective fairness (Prop. 3.1/4.1); that degree–outcome correlation creates exposure bias (Prop. 4.2); and, as central results, that homophily linearly amplifies the perceived gap under DP (Thm. 3.1 and Prop. 4.3). Additional propositions claim that modularity bounds the gap and that clustering reduces dispersion.

Significance. If rigorously established, the central result would be valuable: it would show that standard demographic-parity audits are insufficient in networked environments and that network topology must be part of fairness assessment. The paper's setting is well motivated, and the definitional framework (local comparison operators, exposure bias) is a useful contribution. However, the central theorem's proof is invalid for the DP case, and the advertised linear amplification under DP is not actually derived; it is asserted in a remark. The paper also relies on unproven or externally deferred results for other propositions. Because the main claim is the paper's raison d'être, the significance of the current manuscript is limited until a correct proof is supplied.

major comments (4)
  1. [§3.6, Theorem 3.1] The proof's Step 3 derives E[Δ1] = ψσ(0)(μA−μB)[1−4ρπAπB] + o(ρ). Under DP, μA=μB, so this expression is o(ρ) — there is no linear-in-ρ term. Remark 3.1 concedes this and points to a degree-exposure mechanism in Remark 3.2, but that mechanism is asserted without derivation. The 'transcribing' step that converts the coefficient into c(πA,πB)ρΓ(h) is algebraically inconsistent: with μA=μB the displayed expression has no linear term. Moreover, Eqs. (8)–(10) use unweighted group means, implicitly assuming h is degree-independent within groups, even though Γ(h) includes degree-weighted neighbor exposures. The theorem's central claim is therefore not established.
  2. [§4.2, Proposition 4.3] This proposition repeats the same linear-response expansion as Theorem 3.1. It again uses unweighted group means and does not handle degree-dependent h, despite the statement of Γ(h) involving degree-weighted exposures. No new argument is provided to overcome the DP cancellation identified in Remark 3.1. The claimed result that homophily linearly amplifies the perceived gap under DP is thus unsupported.
  3. [§3.7–3.8, Propositions 3.2 and 3.3] These results are used in the paper's summary of 'takeaways' but are not proven. Proposition 3.2 is given only as a proof idea ('expand the group difference using B...'). Proposition 3.3's proof is deferred to Charpentier and Ratz (2025), a same-author arXiv preprint, without reproducing the argument. A theoretical result whose proof is entirely in a different, not-yet-published source is not an established result in this manuscript.
  4. [§5, Simulation section] The simulation's data-generating process violates DP: H_i = 0.7 H_group + 0.3 H_degree + ε, with H_group ~ Beta(4,2) for group A and Beta(2,4) for group B. These Beta distributions have means 2/3 and 1/3, so E[H|S] is substantially different across groups. The figure legend states the global fairness gap 'remains nearly constant,' not that it is zero. The simulation therefore cannot confirm the DP-specific claim of Theorem 3.1.
minor comments (3)
  1. [§3.4, Eq. (3)] The perception indicator uses a strict/non-strict inequality; the smooth-approximation argument later assumes a continuous density at 0. The exposition would benefit from a precise convention on the tie case, though this does not affect the main criticism.
  2. [§5, Figure 1] No error bars, confidence intervals, or number of simulation replications are reported. Since the claim is about a linear trend, some measure of variability is needed.
  3. [Throughout] The terms 'objective fairness,' 'global fairness,' and 'demographic parity' are sometimes used interchangeably; the paper should consistently distinguish the decision rule property (DP) from the population-level perception limit in Prop. 3.1.

Circularity Check

2 steps flagged · score 4.0 of 10

No fitted-input circularity, but the central DP theorem's derived linear term vanishes under its own assumption and the nonzero gap is deferred to an unproved degree-exposure remark; a supporting proposition also relies on a same-author citation.

  1. other [Theorem 3.1 proof, Steps 1–3 and Remark 3.2 (Section 3.6)]
    "If DP holds for H and H∼Bernoulli(h), then μA = μB, so the mean shift in (10) cancels and the leading linear term above vanishes. ... E[Δ1(h)] = 0 + ρ·κ(πA, πB)·Cov(h, d) + o(ρ), for some κ(πA, πB) > 0 (derivable by the same expansion with degree weights)."

    In the proof, Eq. (12) gives m_s = (π_s' − 2ρπ_sπ_s')(μ_s − μ_s') + o(ρ), and Step 3 gives E[Δ1(h)] = ψ_σ(0)(m_A − m_B) + o(ρ) = ψ_σ(0)(μ_A − μ_B)[1 − 4ρπ_Aπ_B] + o(ρ). Since the theorem assumes DP, μ_A = μ_B, so the only derived first-order term is zero by construction. The nonzero linear slope advertised in (7) is not obtained from the expansion; it is assigned to a degree-exposure term that is merely asserted as 'derivable by the same expansion' without a derivation. Thus the central prediction rests on an unproved additional mechanism rather than following from the paper's own equations.

  2. self citation load bearing [Section 3.8, Proposition 3.3 proof sketch]
    "thresholding preserves a weak reduction in dispersion (Charpentier and Ratz, 2025)."

    The key variance-reduction inequality for Proposition 3.3 is not proved in this paper; the decisive step is deferred to a paper sharing the present author. The proposition is presented as a new analytical result, but its load-bearing thresholding claim is imported from the authors' own prior work without an independent derivation in the text.

full rationale

The paper does not fit parameters and then rename them as predictions; the simulation is illustrative and the formal framework is self-contained. The more subtle issue is Theorem 3.1: the proof's linear-response expansion has a first-order term proportional to μ_A − μ_B, which DP forces to zero. The nonzero perceived gap under DP is then attributed in Remark 3.2 to a degree-exposure effect, but the promised 'same expansion with degree weights' is not shown. This is an omitted-proof/correctness problem that borders on circularity because the advertised conclusion is inserted as an unproved assertion rather than derived from the model's equations. Separately, Proposition 3.3 delegates its key inequality to a same-author citation. These are partial concerns: the perception formalism, friendship-paradox identity, and convergence result have independent mathematical content, and there is no classic fitted-input or definitional circularity. Score 4 reflects the central theorem's unproved rescue mechanism plus the load-bearing self-citation, without treating the paper as entirely reducible to its inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central model postulates a specific binary perception operator and assumes demographic parity as the objective standard; the main theorem's proof additionally depends on an unverified smooth-CDF limit interchange. Simulation parameters are hand-picked and no code is supplied. No new physical or ontological entities are introduced; the fairness-perception operator is a mathematical definition accounted for under axioms.

free parameters (4)
  • alpha (α) = 0.7
    Simulation outcome model Hi = α H_group + (1−α) H_degree + ε; fixed by hand with no sensitivity analysis (Section 5).
  • Beta shape parameters for H_group = Beta(4,2) for group A, Beta(2,4) for group B
    Chosen by hand; means differ (2/3 vs 1/3), so the simulated 'global fairness' baseline is not actually held at zero (Section 5).
  • noise variance = 0.05^2
    Chosen by hand to add individual heterogeneity in the simulation (Section 5).
  • smoothing scale σ in Ψ_σ = σ↓0 (limit)
    Proof device in Theorem 3.1; the coefficient ψ_σ(0) is retained as an unspecified finite constant and the interchange of limits is not justified.
assumptions (6)
  • domain assumption Perceived fairness is F^(d)(i;h) = 1{E_i^(d)[h] ≤ h(i)}
    Eq. (3), Section 3.2. The entire analysis is about this specific binary local comparison; no empirical evidence is given that human fairness perception equals this operator.
  • domain assumption Objective fairness = demographic parity: P[H=1|S=A] = P[H=1|S=B]
    Eq. (1), Section 3.1. If another objective criterion were used (e.g., equalized odds), the results would need re-derivation.
  • domain assumption Two-block SBM with pin > pout, homophily index ρ = (pin−pout)/(pin+pout)
    Section 3.5, Eq. (6). The linear-response formula is derived only in this generative network model.
  • standard math Neighbor group labels are conditionally i.i.d. given the ego's group in the SBM
    Used to derive the neighbor-label probability θ_s in Theorem 3.1 proof, Eq. (8).
  • ad hoc to paper Smooth CDF approximation Ψ_σ → indicator and σ↓0 preserves first-order terms
    Theorem 3.1 proof, Step 1 and footnote 1. The limit interchange is not justified, and under DP it fails to produce a linear term (Remark 3.1).
  • domain assumption Connectedness and non-degenerate h for convergence results
    Propositions 3.1 and 4.1 require a connected graph and positive true/false positive rates for the stated convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perceived Fairness in Networks." pith.science (2026). https://pith.science/paper/BL7KCCRF

@misc{pith2026251012028,
  author       = {Pith},
  title        = {Pith review of: Perceived Fairness in Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BL7KCCRF}},
  note         = {Machine review of arXiv:2510.12028}
}
read the original abstract

The usual definitions of algorithmic fairness focus on population-level statistics, such as demographic parity or equal opportunity. However, in many social or economic contexts, fairness is not perceived globally, but locally, through an individual's peer network and comparisons. We propose a theoretical model of perceived fairness networks, in which each individual's sense of discrimination depends on the local topology of interactions. We show that even if a decision rule satisfies standard criteria of fairness, perceived discrimination can persist or even increase in the presence of homophily or assortative mixing. We propose a formalism for the concept of fairness perception, linking network structure, local observation, and social perception. Analytical and simulation results highlight how network topology affects the divergence between objective fairness and perceived fairness, with implications for algorithmic governance and applications in finance and collaborative insurance.

Figures

Figures reproduced from arXiv: 2510.12028 by the authors.

Figure 1
Figure 1. Relationship between network homophily and perceived fairness gap. [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 1 linked inside Pith

  1. [1]

    , author Hardt, M

    author Barocas, S. , author Hardt, M. , author Narayanan, A. , year 2017 . title Fairness in machine learning . journal Nips tutorial volume 1 , pages 2017

    2017

  2. [2]

    , author Matthews, K.A

    author Brown, C. , author Matthews, K.A. , author Bromberger, J.T. , author Chang, Y. , year 2006 . title The relation between perceived unfair treatment and blood pressure in a racially/ethnically diverse sample of women . journal American Journal of Epidemiology volume 164 , pages 257--262

    2006

  3. [3]

    , author Kirkley, A

    author Cantwell, G.T. , author Kirkley, A. , author Newman, M.E.J. , year 2021 . title The friendship paradox in real and model networks . journal Journal of Complex Networks volume 9 , pages cnab011

    2021

  4. [4]

    , author Ratz, P

    author Charpentier, A. , author Ratz, P. , year 2025 . title Linear risk sharing on networks . journal arXiv volume 2509.21411

    arXiv 2025

  5. [5]

    , author Gonz \'a lez-Sanz, A

    author De Lara, L. , author Gonz \'a lez-Sanz, A. , author Asher, N. , author Risser, L. , author Loubes, J.M. , year 2024 . title Transport-based counterfactual models . journal Journal of Machine Learning Research volume 25 , pages 1--59

    2024

  6. [6]

    , author Hardt, M

    author Dwork, C. , author Hardt, M. , author Pitassi, T. , author Reingold, O. , author Zemel, R. , year 2012 . title Fairness through awareness , in: booktitle Proceedings of the 3rd innovations in theoretical computer science conference , pp. pages 214--226

    2012

  7. [7]

    , author Charpentier, A

    author Fernandes Machado , A. , author Charpentier, A. , author Gallic, E. , year 2025 . title Sequential conditional transport on probabilistic graphs for interpretable counterfactual fairness , in: booktitle Proceedings of the AAAI Conference on Artificial Intelligence , pp. pages 19358--19366

    2025

  8. [8]

    , author McDaniel, M

    author Gonzalez, D. , author McDaniel, M. , author Kenney, G.M. , author Skopec, L. , year 2021 . title Perceptions of unfair treatment or judgment due to race or ethnicity in five settings . journal Washington, DC: Urban Institute

    2021

Show all 19 references
  1. [9]

    , author Price, E

    author Hardt, M. , author Price, E. , author Srebro, N. , year 2016 . title Equality of opportunity in supervised learning . journal Advances in neural information processing systems volume 29 , pages 3315--3323

    2016

  2. [10]

    , author Loftus, J

    author Kusner, M.J. , author Loftus, J. , author Russell, C. , author Silva, R. , year 2017 . title Counterfactual fairness , in: booktitle Advances in Neural Information Processing Systems , pp. pages 4066--4076

    2017

  3. [11]

    , author Smith-Lovin, L

    author McPherson, M. , author Smith-Lovin, L. , author Cook, J.M. , year 2001 . title Birds of a feather: Homophily in social networks . journal Annual review of sociology volume 27 , pages 415--444

    2001

  4. [12]

    , year 2003

    author Newman, M.E. , year 2003 . title Mixing patterns in networks . journal Physical review E volume 67 , pages 026126

    2003

  5. [13]

    , author Richman, L.S

    author Pascoe, E.A. , author Richman, L.S. , year 2009 . title Perceived discrimination and health: a meta-analytic review. journal Psychological Bulletin volume 135 , pages 531

    2009

  6. [14]

    , author Peixoto, T.P

    author Peel, L. , author Peixoto, T.P. , author De Domenico, M. , year 2022 . title Statistical inference links data and theory in network science . journal Nature Communications volume 13 , pages 6794

    2022

  7. [15]

    , author Branscombe, N.R

    author Schmitt, M.T. , author Branscombe, N.R. , author Postmes, T. , author Garcia, A. , year 2014 . title The consequences of perceived discrimination for psychological well-being: a meta-analytic review. journal Psychological bulletin volume 140 , pages 921

    2014

  8. [16]

    , author Cheek Jr, N.H

    author Shrum, W. , author Cheek Jr, N.H. , author MacD, S. , year 1988 . title Friendship in school: Gender and racial homophily . journal Sociology of Education , pages 227--239

    1988

  9. [17]

    , author Percus, A.G

    author Wu, X.Z. , author Percus, A.G. , author Lerman, K. , year 2017 . title Neighbor-neighbor correlations explain measurement bias in networks . journal Scientific Reports volume 7 , pages 5576

    2017

  10. [18]

    , year 1990

    author Yamaguchi, K. , year 1990 . title Homophily and social distance in the choice of multiple friends an analysis based on conditionally symmetric log-bilinear association models . journal Journal of the American Statistical Association volume 85 , pages 356--366

    1990

  11. [19]

    , author Liu, T

    author Zhou, Z. , author Liu, T. , author Bai, R. , author Gao, J. , author Kocaoglu, M. , author Inouye, D.I. , year 2024 . title Counterfactual fairness by combining factual and counterfactual predictions . journal arXiv volume 2408.03425

    2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.