REVIEW 3 major objections 5 minor 1 cited by
The Disparate Effects of Partial Information in Bayesian Strategic Learning
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read For Bayesian agents, utility disparity is bounded and non-monotone in transparency, minimized at intermediate noise; for naive agents it grows unboundedly with noise.
desk verdict Real contribution with a misprinted centerpiece formula; the qualitative fairness-transparency findings survive, but the paper needs a correction pass before the precise characterizations are citable. 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 engine is the Gaussian-Gaussian conjugate update: with prior $N(\omega, \gamma^2 I)$ and signal $S = \theta_* + \sigma Z$, the posterior mean is $\omega + \beta(S-\omega)$ with shrinkage factor $\beta(\sigma,\gamma) = \gamma^2/(\gamma^2+\sigma^2)$. Substituting this posterior into the quadratic-cost best response $\Delta x_g = A_g^{-1} \mathbb{E}[\theta \mid S]$ yields feature changes that are affine in $\beta$, and taking differences of expected utilities yields closed-form disparity equations: Eq. 10 for naive agents and Eq. 12 for Bayesian agents. All monotonicity, boundedness, and neutrality results are read off these quadratic-in-$\beta$ expressions.
What would settle it
Simulate a linear scoring game with two groups, fixed cost matrices $A_1 \prec A_2$, a chosen prior $N(\theta_0, \gamma^2 I)$ with $\gamma > \gamma_c$, and Gaussian signal noise; measure average utility disparity across a grid of $\sigma$ and check that $F_u(\sigma)$ attains its unique minimum at $\sigma_{\min} = \gamma/\sqrt{1-\gamma_c^2/\gamma^2}$ and, for parameters in the region of Lemma 4.12, crosses zero at two distinct points. Then repeat with t-distributed noise, or with agents who believe the noise level is $\sigma' \neq \sigma$, and check whether the disparity remains bounded and whether the interior minimum persists.
Extended reading notes
Core claim
For Bayesian agents with a common Gaussian prior $N(\theta_0, \gamma^2 I)$, the utility disparity $F_u(\sigma)$ between the advantaged and disadvantaged groups has the closed form $F_u(\sigma) = -\frac{\beta^2}{2}(\|k_{\theta_*} - k_{\theta_0}\|^2 + \sigma^2 m) + \beta(\|k_{\theta_*} - k_{\theta_0}\|^2 + k_{\theta_0}^\top k_{\theta_*} - \tfrac{1}{2}k_{\theta_0}^\top k_{\theta_0})$, where $\beta(\sigma,\gamma) = \gamma^2/(\gamma^2+\sigma^2)$, $k_{\theta_*}$ and $k_{\theta_0}$ are the true model and prior mean transformed by $\sqrt{A_1^{-1} - A_2^{-1}}$, and $m = \operatorname{tr}(A_1^{-1}) - \operatorname{tr}(A_2^{-1}) > 0$. From this expression the paper classifies all behaviors: when prior variance satisfies $\gamma \le \gamma_c = \sqrt{(2/m)\|k_{\theta_*} - k_{\theta_0}\|}$, $F_u$ decreases monotonically with noise $\sigma$; when $\gamma > \gamma_c$, $F_u$ is non-monotone with a unique global minimum at $\sigma_{\min} = \gamma/\sqrt{1 - \gamma_c^2/\gamma^2}$, so the most equality-friendly transparency is strictly partial. The same expression determines when disparity is Exploitation, Burden, or Neutrality, including a regime with two distinct neutrality points. For naive agents the analogous closed form is $F_u(\sigma) = \tfrac{1}{2}\theta_*^\top(A_1^{-1}-A_2^{-1})\theta_* - \tfrac{1}{2}\sigma^2 m$, which is unbounded below as $\sigma \to \infty$.
Load-bearing premise
All closed forms assume the agent's posterior is exactly Gaussian-Gaussian, with a normal prior, normal signal noise of variance $\sigma^2$, and agents knowing that variance $\sigma$ exactly; if the noise is misperceived or non-Gaussian, the derived boundedness, monotonicity, and neutrality characterizations have no reason to hold.
Editorial extensions
If this is right
- For Bayesian agents, disclosing more information does not monotonically reduce inequality: whenever prior variance is large enough, utility disparity has an interior minimum at a strictly partial transparency level.
- For naive agents, utility disparity is unbounded as signal noise grows, so withholding information makes the low-cost group over-invest in useless feature changes and ultimately suffer more than the high-cost group.
- Score disparity for naive agents is invariant to transparency, but its variance increases with noise, so transparency affects individual-level randomness even when average score improvements are unchanged.
- When groups hold unequal priors modeled as subspace projections, both score and utility disparity are bounded by a factor times the information-overlap proxy $r_{1,2}(\theta_*)$, and the bound vanishes at full information.
- A learner can use the closed forms to locate neutrality points and, where they exist, set transparency to equalize outcomes exactly across groups.
Reading between the lines
- If the closed forms are correct, transparency is effectively a tunable fairness instrument, and the learner's choice of signal noise resembles an information-design problem in which revealing nothing or everything are both generically suboptimal.
- Because every closed form rests on the signal noise $\sigma$ being known to agents, a natural testable extension is whether misestimation of $\sigma$ or non-Gaussian noise destroys the boundedness and non-monotonicity; the naive model's divergence suggests the bounds are fragile outside the Gaussian conjugacy assumption.
- The regime with two neutrality points implies that a small change in transparency near the Burden interval flips which group is favored, a qualitatively different policy hazard from monotone fairness curves.
- The same machinery could be applied to measure disparity in terms of realized feature changes rather than expected utilities, which would connect these bounds to per-agent welfare rather than group averages.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies a strategic classification model in which agents observe the learner's linear scoring rule θ* only through a noisy signal S = θ* + σZ. Two agent models are analyzed: naive agents who best-respond to S as if it were θ*, and Bayesian agents with a Gaussian prior N(ω_g, γ²I) who form the conjugate posterior. For two groups differing in feature-modification cost matrices A1 ≺ A2, the paper derives closed-form score and utility disparities Fs and Fu as functions of σ, and characterizes monotonicity, boundedness, and neutrality/burden regions. In the naive model utility disparity diverges as σ → ∞, while in the Bayesian model it is bounded and can be minimized at intermediate transparency; a threshold γ_c separates monotone from non-monotone behavior. The paper then extends the Bayesian analysis to groups with unequal prior means given by projections onto subspaces and bounds disparities in terms of the information-overlap proxy of Bechavod et al. (2022).
Significance. If the formulas are corrected, the paper provides a clean, fully derived account of how transparency affects fairness under cost asymmetry, with concrete qualitative predictions: unbounded disparity for naive agents, bounded and possibly non-monotone disparity for Bayesian agents, and intermediate-transparency minima. The model is self-contained, all derivations are shown, and no fitting or simulation is used; the neutrality conditions and σ_min are explicit and falsifiable within the Gaussian-linear setting. The main limitations (linear rules, Gaussian noise, known noise level) are acknowledged by the authors. However, the central displayed formula in Theorem 4.10 contains an algebraic error, and a few supporting bounds in Section 5 use the wrong matrix norm; these must be repaired before the quantitative claims can be accepted.
major comments (3)
- [Theorem 4.10 / Eq. (12) and Appendix B.2] The displayed utility disparity formula in Eq. (12) is not what follows from the paper's own derivation. Substituting ω1 = ω2 = θ0 into the condensed expression in Appendix B.2 (Eq. (17)) yields Fu(σ) = −β²/2(∥kθ*−kθ0∥² + σ²m) + β∥kθ*−kθ0∥² + (kθ0^T kθ* − ∥kθ0∥²/2), whereas Eq. (12) also multiplies the dot-product term kθ0^T kθ* − ∥kθ0∥²/2 by β. The printed formula therefore has limit 0 as σ → ∞, contradicting the limit kθ0^T kθ* − ∥kθ0∥²/2 stated in Lemma 4.11 and used in the subsequent case analysis. The proofs in Appendix B.2.1 and of Lemma 4.12 in fact use the corrected form, so the qualitative conclusions appear to survive, but Theorem 4.10 and all statements that quote its boundary values must be corrected.
- [Lemma 4.11 / Appendix B.2.1] The threshold γ_c is stated as γ_c := sqrt(2/m |kθ* − kθ0|), but the first-order analysis in Appendix B.2.1 shows that the sign of F'_u(σ) changes at 2(k2+k3−2k1) = mγ², i.e., at γ² = (2/m)∥kθ*−kθ0∥². As printed, the threshold has the wrong scaling with the norm of kθ*−kθ0. Consequently Eq. (13) for σ_min, the condition γ > γ_c, and the feasibility region in Lemma 4.12 inherit this error and must be re-expressed with the corrected γ_c.
- [Lemmas 5.12 and 5.17] Both bounds state the right-hand side with ∥A∥^{-1}, but the derivation with A1 = A2 = A bounds expressions of the form θ*^T A^{-1}(Π1−Π2)θ*, whose norm is controlled by ∥A^{-1}∥, not by ∥A∥^{-1}. For non-scalar A these quantities differ; for example, with A = diag(1,1/2), θ* = (0,1), Π1 = diag(0,1), and Π2 = 0, the left side of the Lemma 5.12 bound at σ → ∞ equals 2 while the printed right side equals 1. Replace ∥A∥^{-1} by ∥A^{-1}∥, or explicitly restrict these lemmas to scalar cost matrices where the two norms coincide.
minor comments (5)
- [Eq. (12) and Lemma 4.11] The notations `(kθ*−kθ0)^2` and `|kθ*−kθ0|` are used without defining whether they denote a squared norm or a norm; since the surrounding text uses ∥·∥ for norms, please use ∥kθ*−kθ0∥² consistently.
- [Section 4.2, after Eq. (11)] The sentence describing Eq. (11) writes both weighted terms as kθ0^T kθ*; the second term should be kθ*^T kθ* to match the displayed formula.
- [Proof of Lemma 4.12] The notation Fu(γ) is used to mean Fu evaluated at σ = γ; because γ is already a model parameter, the evaluation point should be written explicitly.
- [Section 5.1] The phrase 'σ− → ∞' appears to be a typo and should read 'σ → ∞'.
- [Figure 3] The caption states that each panel uses a chosen prior mean θ0 but does not list the values, which makes the numerical illustrations hard to reproduce; please include the θ0 values used in each panel.
Circularity Check
No significant circularity: the closed forms and qualitative classifications are derived from stated Gaussian assumptions, not fitted to the predicted quantities or imported from self-citations.
full rationale
The derivation chain is self-contained. Lemma 3.1 solves the naive agent's stated concave program by first-order conditions, and Lemma 3.2 computes the Bayesian best response from the paper's explicitly stated Gaussian prior and Gaussian signal using the standard conjugate-posterior formula, with the textbook citation Murphy (2012) as external support. The fairness metrics in Eqs. (5)-(6) are definitions, and Theorems 4.6 and 4.10 are obtained by substituting the Lemma 3.2 expression into those definitions and taking expectations under the stated signal model; no parameter is fitted to the quantity being predicted, and the neutrality points, bounds, gamma_c threshold, sigma_min, and the two-neutrality-point region in Lemma 4.11 are algebraic consequences of the same closed form rather than imposed inputs. Section 5 likewise begins from Assumption 5.1, which states the subspace priors, and derives Eqs. (15)-(17) from the best-response expression; the information-overlap proxy of Bechavod et al. (2022) is used as a definition, not as a proof of the paper's claims. The paper does cite prior work by a co-author (Bechavod et al. 2022 and Cohen et al. 2024), but these citations introduce modeling assumptions or definitions rather than supplying the load-bearing mathematical content, which is re-derived internally. The skeptic's noted algebraic inconsistency between printed Eq. (12) and the appendix's condensed Eq. (17) would be a correctness issue about the displayed formula, not a circularity, because the derivation does not assume its own conclusion. Accordingly, no circular step is present.
Assumptions & free parameters
assumptions (8)
- domain assumption Learner's scoring rule is linear: y = x^T theta*
- domain assumption Agents observe a noisy signal S = theta* + sigma Z with Z ~ N(0,I) and known sigma
- domain assumption Prior beliefs are Gaussian: pi_g ~ N(omega_g, gamma^2 I)
- domain assumption Cost of feature modification is quadratic Mahalanobis with positive definite matrix A_g
- domain assumption Group 1 is advantaged: A1 <= A2 in the Loewner order
- domain assumption Both groups share the same prior mean theta_0 in Section 4.2
- domain assumption In Section 5, prior mean is Pi_g theta* and A_g^{-1} commutes with Pi_g
- standard math Concavity guarantees unique best responses via first-order conditions
Cite this review
Pith. "Pith review of The Disparate Effects of Partial Information in Bayesian Strategic Learning." pith.science (2026). https://pith.science/paper/UJH6EKQK
@misc{pith2026250600627,
author = {Pith},
title = {Pith review of: The Disparate Effects of Partial Information in Bayesian Strategic Learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/UJH6EKQK}},
note = {Machine review of arXiv:2506.00627}
}
read the original abstract
We study how partial information about scoring rules affects fairness in strategic learning settings. In strategic learning, a learner deploys a scoring rule, and agents respond strategically by modifying their features -- at some cost -- to improve their outcomes. However, in our work, agents do not observe the scoring rule directly; instead, they receive a noisy signal of said rule. We consider two different agent models: (i) naive agents, who take the noisy signal at face value, and (ii) Bayesian agents, who update a prior belief based on the signal. Our goal is to understand how disparities in outcomes arise between groups that differ in their costs of feature modification, and how these disparities vary with the level of transparency of the learner's rule. For naive agents, we show that utility disparities can grow unboundedly with noise, and that the group with lower costs can, perhaps counter-intuitively, be disproportionately harmed under limited transparency. In contrast, for Bayesian agents, disparities remain bounded. We provide a full characterization of disparities across groups as a function of the level of transparency and show that they can vary non-monotonically with noise; in particular, disparities are often minimized at intermediate levels of transparency. Finally, we extend our analysis to settings where groups differ not only in cost, but also in prior beliefs, and study how this asymmetry influences fairness.
Figures
Forward citations
Cited by 1 Pith paper
-
Desirable Effort Fairness and Optimality Trade-offs in Strategic Learning
Constraining a strategic classifier to keep desirable-effort incentives fair between two groups costs the principal an explicit accuracy or welfare loss bounded by the fairness tolerance beta.
Reference graph
Works this paper leans on
-
[3]
The role of randomness and noise in strategic classification.arXiv preprint arXiv:2005.08377,
Mark Braverman and Sumegha Garg. The role of randomness and noise in strategic classification.arXiv preprint arXiv:2005.08377,
arXiv 2005
-
[7]
Valia Efthymiou, Chara Podimata, Diptangshu Sen, and Juba Ziani. Incentivizing desirable effort profiles in strategic classification: The role of causality and uncertainty.arXiv preprint arXiv:2502.06749,
-
[9]
The disparate equilibria of algorithmic decision making when individuals invest rationally
Lydia T Liu, Ashia Wilson, Nika Haghtalab, Adam Tauman Kalai, Christian Borgs, and Jennifer Chayes. The disparate equilibria of algorithmic decision making when individuals invest rationally. InProceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pages 381–391,
work page 2020
-
[12]
16 arXivTemplateA PREPRINT A Proofs for the equilibrium computation Proof of Lemma. 3.1. The optimization problem for a Naive agent given by Equation 3 is concave for any given realization ofZ. Therefore, we can solve it using first order conditions as below. ∇u(∆x;g) = (θ ∗ +σZ g)−A g∆x= 0 where ∆x=x ′ −x . The solution to the above equation is given by ...
work page 2012
-
[2012]
Chara Podimata. Incentive-aware machine learning; robustness, fairness, improvement & causality.arXiv preprint arXiv:2505.05211,
-
[2018]
The Double-Edged Sword of Behavioral Responses in Strategic Classification: Theory and User Studies
Raman Ebrahimi, Kristen Vaccaro, and Parinaz Naghizadeh. The double-edged sword of behavioral responses in strategic classification: Theory and user studies.arXiv preprint arXiv:2410.18066,
-
[2020]
Fairness-aware contextual dynamic pricing with strategic buyers.arXiv preprint arXiv:2501.15338,
Pangpang Liu and Will Wei Sun. Fairness-aware contextual dynamic pricing with strategic buyers.arXiv preprint arXiv:2501.15338,
-
[2021]
On classification of strategic agents who can both game and improve.arXiv preprint arXiv:2203.00124,
Saba Ahmadi, Hedyeh Beyhaghi, Avrim Blum, and Keziah Naggita. On classification of strategic agents who can both game and improve.arXiv preprint arXiv:2203.00124,
Show all 12 references
-
[2022]
Machine bias
Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. Machine bias. propublica, may 23, 2016,
2016
-
[2023]
Bayesian strategic classification
Lee Cohen, Saeed Sharifi-Malvajerdi, Kevin Stangl, Ali Vakilian, and Juba Ziani. Bayesian strategic classification. arXiv preprint arXiv:2402.08758,
-
[2024]
Strategic classification from revealed preferences
Jinshuo Dong, Aaron Roth, Zachary Schutzman, Bo Waggoner, and Zhiwei Steven Wu. Strategic classification from revealed preferences. InProceedings of the 2018 ACM Conference on Economics and Computation, pages 55–70,
2018
-
[2025]
Strategic classification
Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classification. InProceedings of the 2016 ACM conference on innovations in theoretical computer science, pages 111–122,
2016
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.