{"id":"ae7e88c7-46d0-4566-be29-543a8f6440d8","arxiv_id":"2602.06730","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A Wasserstein robust performative prediction framework with decision-dependent ambiguity claims linear convergence for two retraining algorithms, but the central convergence proof is not supported by the stated assumptions.","lead":"This paper proposes two algorithms for retraining predictive models when the people being predicted adapt to the model, using robust worst-case data distributions, and claims both converge to a stable solution. The convergence proofs rest on an unproven strong-convexity condition and analyze a full-batch update that the implemented algorithms do not run.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unsupported strong-convexity assertion invalidates Theorems 1-3: Assumptions 2-4 do not imply γ>0; proof cites Assumption 3 incorrectly.","rationale":"The reader's weakest assumption identifies exactly the missing strong-convexity requirement. I confirm that Assumptions 2-4 are insufficient: Lipschitz smoothness is an upper bound, and concavity in ζ says nothing about θ-curvature. The proof of Theorem 1 explicitly misattributes strong convexity to Assumption 3. A concrete counterexample with l≡0 satisfies all stated assumptions yet yields γ<0 and a concave ϕ, so the contraction argument collapses. This is the most load-bearing concern because it invalidates the core proof even before considering the discrepancy between stochastic algorithms and full-batch analysis or the Lagrangian/DRO equivalence. The rejection verdict remains appropriate unless the authors add a genuine strong-convexity assumption and rework the proofs.","tokens_in":20623,"tokens_out":4040,"duration_ms":38658,"concrete_test":"Define l(θ,ζ) ≡ 0, λ(θ)=λ_min + (H_λ/2)∥θ-θ0∥² on a compact Θ with nonempty interior, and Ξ with at least two points so D_ξ>0. This satisfies Assumptions 2-4 with L_θθ=0; then γ=-H_λD_ξ²<0, and ϕ=-λ(θ)∥ξ-ζ∥² is concave in θ, not strongly convex. This counterexample shows the assertion preceding Theorem 1 is false, so the proof of Theorem 1 cannot be repaired without adding a strong-convexity assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III states 'By Assumptions 2 and 4, we establish that ϕ(θ,ξ,ζ) is γ-strongly convex in θ' with γ = L_θθ - H_λD_ξ². Assumption 2 is a Lipschitz-smoothness upper bound on ∇θl; it gives ∇²θθl ⪯ L_θθI, not ⪰ γI. Assumption 3 concerns concavity in ζ. No lower curvature bound on l(·,ζ) is assumed. The Appendix proof of Theorem 1 writes 'From Assumption 3, ϕ(x,ξ,ζ) is γ-strongly convex in x,' which is a misattribution. Since the contraction inequalities (27)-(29) and the division by γ in Theorems 1-3 require γ>0, the central convergence and suboptimality claims are unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers performative prediction with a Wasserstein distributionally robust objective. It defines a decision-dependent ambiguity radius ρ(θ), relaxes the constrained DRO objective (4) to a Lagrangian-penalized form (5), and uses strong duality (Lemma 1) to obtain a min-max surrogate f(θ,ξ)=sup_ζ l(θ,ζ)-λ(θ)||ξ-ζ||². Building on this surrogate, the authors define a robust performative stable point and propose two algorithms: repeated risk minimization (Algorithm 1) and repeated gradient descent (Algorithm 2). Under Assumptions 1–4, Theorems 1 and 2 claim linear convergence to a unique stable point (or to a neighborhood in the inexact case), and Theorem 3 bounds the suboptimality gap. The paper also presents experiments on a credit-scoring dataset. The main theorems are not supported by the stated assumptions: the required strong convexity of ϕ in θ is asserted but not derived, and the theorems analyze batch expectation updates rather than the single-sample updates in the algorithms.","tokens_in":20840,"tokens_out":13103,"duration_ms":128799,"significance":"If the claims were true, this would be a worthwhile extension of performative prediction to data-driven Wasserstein DRO, particularly the decision-dependent radius and the convergence guarantees for both RRM and RGD. The paper also gives an explicit suboptimality gap and supports the theory with simulations. However, the missing strong-convexity condition is load-bearing in every theorem, and the discrepancy between the stated algorithms and the analyzed updates is substantial. The contribution is therefore not established in the current form.","major_comments":[{"comment":"The paper asserts that ϕ(θ,ξ,ζ) is γ-strongly convex in θ with γ = L_θθ - H_λ D_ξ², citing Assumptions 2 and 4. Assumption 2 is a Lipschitz-smoothness (upper curvature) condition on l; it gives ∇²_θθ l ⪯ L_θθ I, not ⪰. Assumption 3 is concavity in ζ, and Assumption 4 is a Lipschitz/Hessian bound on λ. No lower bound on ∇²_θθ l or on γ is provided. The appendix proof of Theorem 1 explicitly states 'From Assumption 3, ϕ(x,ξ,ζ) is γ-strongly convex in x', which is a misattribution. Since γ appears in the denominators of κ_rm, C_rm, κ_gd, C_gd and in Theorem 3, and the contraction inequalities (27)–(29) require γ>0, the central convergence and suboptimality theorems are not established by the stated assumptions.","section":"Section III, after Assumption 4; Appendix, Proof of Theorem 1"},{"comment":"The theorems do not analyze the algorithms as written. Algorithm 1 updates θ_{t+1} using a single sample ξ_t via (10), whereas Theorem 1 is stated for the deterministic expectation update (8) and the proof defines G(θ)=argmin J_θ(x) and then takes expectations over ξθ. Similarly, Algorithm 2 uses the single-sample gradient (12)/(13), while Theorem 2's proof analyzes the expected gradient E_{ξθ}[∇J_θ(θ)] in (33)–(40). No stochastic approximation argument (e.g., variance bound, martingale convergence, or high-probability guarantee) is given. Thus the claimed convergence of Algorithm 1 and Algorithm 2 is unproven.","section":"Section III-A, Algorithm 1 and Theorem 1; Section III-B, Algorithm 2 and Theorem 2"},{"comment":"Theorem 3 depends on the same unsupported strong-convexity property: the proof starts 'Since the function DPR(θ_s,·) is γ-strongly convex'. This is not a consequence of Assumptions 1–4. Also, the θ_o in (7a) is the minimizer of the empirical robust objective DPR(θ,θ), not of the original performative risk (1); the bound (15) therefore does not quantify the gap to the 'global performative optimum' claimed in the abstract. The final theorem is thus not established.","section":"Section IV, Theorem 3"},{"comment":"The relation between the constrained DRO problem (4) and the Lagrangian relaxation (5) is not specified. For fixed θ, strong duality relates the constrained problem to min_{λ≥0} [λρ(θ)+E[f]], but (5) fixes λ(θ) a priori as a function of θ. The paper does not state how λ(θ) is chosen from ρ(θ) or prove that minimizing (5) is equivalent to minimizing (4). Consequently the 'robust performative stable point' defined in (7b) is a fixed point of the penalized objective; the connection to the ambiguity-set problem in (4) is not established.","section":"Section II, Eqs. (4)–(5)"}],"minor_comments":[{"comment":"The displayed expression for ∇_θϕ(θ,ξ,ζ)-∇_θϕ(θ,ξ',ζ) has a missing parenthesis and incorrectly writes the first term as ∇θl(θ,ζ)-∇θl(θ,ζ).","section":"Appendix, proof of Lemma 3, Eq. (17)"},{"comment":"The term '2η^T T2' should be '2η T2'; T2 is scalar, not a vector.","section":"Appendix, proof of Theorem 2, Eq. (34)"},{"comment":"The symbol γ is used both as the strong-convexity modulus in Section III and as the regularization coefficient in the simulations. This notation clash is confusing; use a different symbol for the regularization parameter.","section":"Section V, after Eq. (L(θ,ξ))"},{"comment":"The y-axis uses a scaling factor c=||θ^S_0||^{-1}_2 that is not defined in the main text or figure captions. The marker convention for 'numerically zero' also deserves an explicit statement.","section":"Section V, Figs. 2 and 4"},{"comment":"The ambiguity set uses the Wasserstein distance with squared Euclidean cost, while Assumption 1 and Lemma 2 concern W1. The proof of Theorem 1 uses the W1 Kantorovich-Rubinstein bound, so the metric mismatch should be stated explicitly to avoid ambiguity.","section":"Section II, Assumption 1 and Lemma 2"}],"recommendation":"reject","confidential_remarks":"The paper has a reasonable high-level idea and the numerical study is useful. However, the main convergence and suboptimality theorems rely on a strong-convexity assertion that is not justified by the stated assumptions, and the theorems analyze different updates from the algorithms they claim to cover. These are load-bearing issues rather than presentation problems, so I cannot recommend acceptance. A substantial revision with an explicit strong-convexity condition and a proper analysis of the single-sample algorithms could make the paper publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I think the reader's rejection is right: the main theorems don't hold as written. The paper targets a real gap—prior Wasserstein performative prediction work used constant radii and skipped RGD convergence—and the decision-dependent penalty lambda(theta) is a sensible way to shrink ambiguity as the model stabilizes. The credit-scoring simulations are reasonable and suggest the method works empirically. The local Lipschitz computations (Lemma 3, Proposition 1) are mostly standard and appear correct. The load-bearing flaw is the strong-convexity assertion. Section III says phi is gamma-strongly convex in theta with gamma = L_theta_theta - H_lambda D_xi^2, citing Assumptions 2 and 4. Assumption 2 is a Lipschitz upper bound on gradients, not a lower curvature bound; Assumption 3 is concavity in zeta. Nothing in the assumptions gives a positive lower bound on the Hessian of l with respect to theta. The Appendix proof of Theorem 1 even claims strong convexity 'from Assumption 3,' which is a misattribution—concavity in zeta says nothing about convexity in theta. Every contraction inequality and every division by gamma requires gamma > 0, so Theorems 1–3 are unproven. This is not a minor gap; it is the engine of the paper. Two secondary issues reinforce the problem. First, the algorithms sample a single xi_t per iteration, but the convergence analysis treats G as the expected-risk minimizer or expected-gradient step. The theorems do not apply to the actual stochastic iterates. Second, the Lagrangian surrogate with decision-dependent lambda is not shown to be equivalent to the Wasserstein ambiguity-set problem; Lemma 1 covers constant lambda, and the extension needs a proof. The paper does show engagement with the literature and the proof structure suggests the authors know the standard contraction playbook. The flaw may be fixable by adding an explicit strong-convexity assumption and analyzing the stochastic updates separately. But as it stands, the central claims do not follow, and the paper should not be cited for convergence guarantees. Who gets value from this? A researcher working on performative prediction or decision-dependent DRO who wants to repair the theory and push the idea further. It deserves referee time in the sense that a careful referee can identify the missing assumption and guide a revision; I would not accept it in its current form. I would tell the editor: send it to review, but expect substantial revision before any claims can be trusted.","headline":"The main convergence theorems rest on an unproven strong-convexity assumption, so the paper's central claims are not established as written, but the decision-dependent penalty idea is worth revisiting.","tokens_in":723,"tokens_out":681,"would_cite":false,"duration_ms":32025,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","90C15","90C47"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that two iterative retraining algorithms converge linearly to a unique robust performative stable point when the model is trained against a Wasserstein ambiguity set with model-dependent radius, and bounds how far that poin","keywords":["performative prediction","distributionally robust optimization","Wasserstein ambiguity set","Lagrangian relaxation","repeated risk minimization","repeated gradient descent","performative stability","linear convergence"],"falsifier":"Construct a loss that is smooth and concave in ζ but not strongly convex in θ, such as l(θ,ξ)=θ·ξ, with λ(θ)=constant > 0 and compact Ξ; then L_θθ=0 so γ=0 and κ_rm is undefined. If repeated risk minimization still converges linearly in this setting, the theorem's stated Assumptions 1–4 do not suffice; if it fails, the strong-convexity gap is confirmed. A numerical run of Algorithm 1 on this counterexample would settle it.","tokens_in":20480,"feed_emoji":"🎯","tokens_out":6414,"duration_ms":60446,"temperature":0.7,"pith_summary":"The paper's aim is to show that performative distribution shift—where deploying a model changes the data distribution—can be handled by robust retraining against a Wasserstein ambiguity set. It develops distributionally robust repeated risk minimization and gradient descent, and proves that both converge linearly to a unique robust performative stable point under smoothness, concavity, and sensitivity assumptions; with inexact inner optimization they converge to a bounded neighborhood. The stable point is the fixed point of a Lagrangian-relaxed min-max objective, and the paper bounds its gap to the true performative optimum. If the proof is correct, practitioners can retrain on finite data with explicit convergence rates rather than ad hoc iterations.","feed_headline":"Robust retraining converges linearly under performative shifts","feed_subtitle":"Wasserstein ambiguity sets turn strategic distribution shift into a tractable min-max problem with linear convergence.","key_machinery":"The central mechanism is the Lagrangian relaxation duality (Lemma 1): sup_P (E_{ξ∼P} l(θ,ξ) − λ(θ) W(P, P̂(θ))) = E_{ξ∼P̂(θ)} sup_ζ [l(θ,ζ) − λ(θ)‖ξ−ζ‖²], which converts an infinite-dimensional robust objective into a sample-level min-max problem over adversarial perturbations ζ*(θ,ξ). The proof's contraction arguments then hinge on the robust surrogate φ(θ,ξ,ζ) = l(θ,ζ) − λ(θ)‖ξ−ζ‖² being γ-strongly convex in θ and μ-strongly concave in ζ, together with ε-sensitivity of the distribution map P̂(θ) in Wasserstein-1 distance. The rate κ_rm is exactly the product of that sensitivity and the surrogate's Lipschitz constant, divided by the strong-convexity modulus γ.","core_discovery":"The central claim is that the intractable worst-case distribution objective sup_P E_{ξ∼P}[l(θ,ξ)] with W(P,P̂(θ)) ≤ ρ(θ) can be replaced by the sample-level min-max objective E_{ξ∼P̂(θ)} sup_ζ [l(θ,ζ) − λ(θ)‖ξ−ζ‖²], and that the resulting retraining maps are contractions. Theorem 1 shows ‖G(θ)−G(θ′)‖ ≤ κ_rm‖θ−θ′‖ + C_rm with κ_rm = ε L^f_{θξ}/γ, giving linear convergence to a unique stable point θ_s when κ_rm<1 and to a neighborhood when the inner maximizer is ε-approximate. Theorem 2 extends this to projected gradient descent with κ_gd and a step-size condition. Theorem 3 bounds the myopia gap: ‖θ_s−θ_o‖ ≤ 2εL^f_ξ/γ and the risk gap by 2(εL^f_ξ)²/γ. These are the results the paper is trying","pith_inferences":["A concrete testable extension is to let the ambiguity radius be learned rather than hand-set; since the theory only requires λ(θ) bounded and smooth with H_λ D_ξ² < L_θθ, any adaptive scheduling scheme that keeps λ inside that envelope should inherit the linear convergence.","The bound ‖θ_s − θ_o‖ ≤ 2εL^f_ξ / γ suggests that the myopia gap is irreducible in the feedback loop, so no retraining algorithm that updates on current data can reach the performative optimum; the robust stable point is the natural attainable target.","If the loss does not have the asserted strong convexity in θ, the contraction denominator γ may vanish; in that regime one would expect either slower, sublinear convergence or oscillation, so the simulations' high-sensitivity spikes are consistent with the theory barely holding.","The min-max surrogate structure connects to adversarial training: ζ*(θ,ξ) can be read as the worst-case feature perturbation, so the framework provides a principled robustness interpretation for strategic feature manipulation in tabular classification."],"forward_implications":["If the assumptions hold, repeated risk minimization over the empirical distribution converges to a unique robust performative stable point at a linear rate; with inexact inner maximization it converges to a bounded neighborhood.","The same guarantee holds for repeated projected gradient descent when the step size is below the derived threshold, making the method applicable to high-dimensional problems without exact optimization oracles.","The stable point is provably close to the global performative optimum: the parameter gap is O(εL/γ) and the risk gap is O((εL)²/γ), so the performance loss due to myopic retraining is controlled by how sensitive the distribution is to the model.","Allowing the ambiguity radius to depend on θ subsumes constant-radius Wasserstein DRO and gives a principled way to shrink robustness as the model stabilizes.","In the credit-scoring simulation, the distributionally robust model maintains higher defaulter detection than standard performative prediction and a static classifier across strategic sensitivity levels."],"fun_headline_variants":["Wasserstein robust retraining converges linearly under strategic shifts","Tractable min-max turns performative shifts into linear convergence","Wasserstein stability gives linear convergence to performative stable point","Contraction-based retraining converges linearly amid distributional shifts","Worst-case robustness yields linear convergence for retraining"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that φ(θ,ξ,ζ) is γ-strongly convex in θ with γ = L_θθ − H_λ D_ξ² > 0; the paper asserts this follows from Assumptions 2 and 4, but those assumptions bound Lipschitz constants from above and do not supply the lower curvature bound that strong convexity requires.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein robust retraining converges linearly under strategic shifts","Tractable min-max turns performative shifts into linear convergence","Wasserstein stability gives linear convergence to performative stable point","Contraction-based retraining converges linearly amid distributional shifts","Worst-case robustness yields linear convergence for retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000816,"raw_usage":{"total_tokens":3457,"prompt_tokens":835,"completion_tokens":2622,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":2555}},"tokens_in":579,"tokens_out":2622,"duration_ms":16927,"temperature":1.0,"reasoning_tokens":2555,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:48:04.565208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a loss that is smooth and concave in ζ but not strongly convex in θ, such as l(θ,ξ)=θ·ξ, with λ(θ)=constant > 0 and compact Ξ; then L_θθ=0 so γ=0 and κ_rm is undefined. If repeated risk minimization still converges linearly in this setting, the theorem's stated Assumptions 1–4 do not suffice; if it fails, the strong-convexity gap is confirmed. A numerical run of Algorithm 1 on this counterexample would settle it.","supporting_citations":[],"review_version":1}