{"id":"42b8c5a6-d01f-4d37-996d-25384ec508ad","arxiv_id":"2509.01076","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A proposed inverse-image Wasserstein DRO for noisy data is shown to contain a false equivalence in its reformulation, invalidating the paper's main 'blessing in disguise' result.","lead":"This paper proposes a distributionally robust optimization method that builds uncertainty sets over a hidden 'clean' distribution by pulling back a Wasserstein ball around noisy observations. It claims noisy data can improve decision-making, but the central convex reformulation that drives this conclusion is mathematically incorrect.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4.1's componentwise coupling assumption is not WLOG; a concrete mean-preserving noise instance makes the reformulation infeasible while the original DRO is finite, so the dominance theorem is unsupported.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the 'sufficient way' in Lemma 4.1 is not an equivalence. My analysis confirms this with an explicit, simple mean-preserving noise instance where the original noisy-data DRO is feasible and finite, but the proposed convex reformulation is infeasible/unbounded. Since the reformulation is used to prove Theorem 4.2 (dominance) and the POA comparisons, the central 'blessing in disguise' claim is not established. The statistical guarantees in Section 3 (finite-sample and consistency) appear to rely on the ambiguity-set definition rather than the reformulation and may be sound, but they do not rescue the main optimization conclusions. I find no other concern that is more load-bearing; the WLOG failure is both specific and decisive. No ad hominem intended; the issue is purely mathematical.","tokens_in":26345,"tokens_out":17127,"duration_ms":195410,"concrete_test":"Solve the finite LP for the counterexample: X={0,1}, E uniform on {-1,1}, N=2, observed x*=(0,2), ε=0.5, U(w,x)=w x, w=1. Compute the value of the original problem (5) via brute-force enumeration of all F and couplings (it is finite, equal to 0.5) and compare with the Lemma 4.1 reformulation (8). If they differ (e.g., the reformulation is infeasible or unbounded), the claimed equivalence fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Lemma 4.1 replaces the marginal constraint (1/N)Σ_j G*_j = T_O(F) with the stronger componentwise condition G*_j = T_O(α_j) for some decomposition F = (1/N)Σ_j α_j. This is only a sufficient way to satisfy the marginal constraint, not an equivalence: feasible couplings may have G*_j outside the image of T_O. The reformulation then solves a restricted problem, and its value need not equal the original DRO. A concrete counterexample: X={0,1}, noise E uniform on {-1,1} (mean-preserving), N=2, observed samples 0 and 2, ε=0.5, U(w,x)=w x with w≥0. The original problem has feasible latent F=δ_1 with coupling G*_1=δ_0, G*_2=δ_2, giving transport cost 0; the true inner value is 0.5w. In the reformulation, c_1(0)=c_1(1)=1, c_2(0)=2, c_2(1)=1, so any α satisfies (c_1+c_2)/2 ≥ 1; the cost constraint is infeasible for ε=0.5, and the dual is unbounded. Thus Lemma 4.1 is false, and Theorem 4.2, which relies on the reformulated constraints, does not establish g*_noise(ε) ≥ g*(ε).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributionally robust optimization framework for decision-making under noisy data. The ambiguity set is defined as the inverse image, under a known noise kernel, of a Wasserstein ball centered at the noisy empirical distribution. The central claims are: (i) the resulting noisy-data DRO admits a tractable convex reformulation; (ii) under mean-preserving noise it is less conservative than a direct DRO that treats noisy observations as clean, i.e., g*_noise(ε) ≥ g*(ε); (iii) this implies a lower price of ambiguity; and (iv) finite-sample guarantees and asymptotic consistency hold. The paper also presents a fair-resource-allocation application. The main proof of the reformulation rests on a false 'without loss of generality' assumption about the structure of optimal transport couplings, and the dominance theorem is therefore unsupported.","tokens_in":26759,"tokens_out":7746,"duration_ms":100747,"significance":"The inverse-image construction is a natural and potentially useful idea, and the statistical results in Section 3 appear to be largely independent of the disputed reformulation. If the dominance theorem were correct, it would be a surprising and practically relevant result. However, the central mathematical claim is not established: the proof of Lemma 4.1 contains a load-bearing error, and the counterexample in this report shows that the reformulated program can be unbounded while the original problem is finite. The numerical study solves the restricted surrogate, not the original noisy-data DRO, so it does not provide empirical support for the paper's main conclusion. Because the key comparison results (Theorem 4.2, Lemma 4.2, and the 'blessing in disguise' interpretation) all rely on the invalid reformulation, the paper's central contribution is not currently supported.","major_comments":[{"comment":"The proof replaces the marginal constraint (1/N)Σ_j G⋆_j = T_O(F) with the componentwise condition G⋆_j = T_O(α_j), where F = (1/N)Σ_j α_j. The text calls this 'a sufficient way' but then treats it as an equivalence by reparameterizing the entire optimization over α_j. This is not without loss of generality: feasible couplings may have components G⋆_j outside the image of T_O. Concretely, take X=[0,1], E uniform on {-1,1} (mean-preserving), N=2, observed samples 0 and 2, ε=0.5, U(w,x)=wx with w≥0. The latent F=δ_1 with G⋆_1=δ_0 and G⋆_2=δ_2 satisfies the original coupling constraints with zero transport cost, so the original inner problem is feasible and finite. In the reformulation, c_1(x)=1 for all x∈[0,1] and c_2(x)=2−x, so the average cost is at least 1; the constraint (1/N)Σ_j∫c_j dα_j ≤ 0.5 is infeasible and the dual is unbounded. Lemma 4.1 is therefore false as stated.","section":"Appendix A.2, proof of Lemma 4.1"},{"comment":"The dominance result g*_noise(ε) ≥ g*(ε) is proved entirely through the reformulated program (8). Since Lemma 4.1 is invalid, the theorem does not compare the original noisy-data DRO problem (5) with the direct DRO. In fact, the reformulation is a restriction of the original problem: it imposes G⋆_j = T_O(α_j), so its inner infimum is no smaller than the original inner infimum. Thus even the inequality g_surrogate ≥ g_direct is in the wrong direction to imply g_original ≥ g_direct; the original problem could have a smaller value. The numerical experiments in §5.2.2 solve the restricted surrogate, not Problem 2.1, so they cannot confirm Theorem 4.2 for the original model. The price-of-ambiguity dominance in Lemma 4.2 inherits the same unsupported conclusion.","section":"Theorem 4.2 / Lemma 4.2"},{"comment":"The tractability claim for Problem 2.1, as well as the sensitivity result dg*_noise/dε = −λ*(ε), depends on the dual program derived in Lemma 4.1. Once Lemma 4.1 is rejected, the 'shadow price of robustness' interpretation is not justified for the original problem. The paper does not provide an alternative derivation of a convex reformulation for the original coupling problem, so the proposed framework lacks a verified tractable formulation.","section":"Section 4.1 / Theorem 4.1"}],"minor_comments":[{"comment":"There are several typos, including 'p aper' in the abstract, and in Figure 4 the labels 'P ice of F ai ness', 'F AIR', and 'Optimal Value unde Va ying ε' are corrupted. Please proofread the text and figure labels.","section":"Abstract / Figure 4"},{"comment":"The statement says 'for all N ≥ 1, n ≠ 2' and then uses max{n,2}; for n=2 the exponent is the same as for n<2, so the exclusion of n=2 is unexplained. Either justify it or remove it.","section":"Theorem 3.1"},{"comment":"The numerical section says the DRO_noise model applies Lemma 4.1, but since Lemma 4.1 is invalid, the reported numbers are for the restricted surrogate. If the authors propose a corrected reformulation, the experiments should be rerun for the original problem. Also, details such as the solver, seeds, and code availability are not provided, which limits reproducibility.","section":"Section 5.2"},{"comment":"The notation P⋆,∞ is used without a definition. It appears to denote the infinite product measure for the noisy sample path, but this should be stated explicitly.","section":"Theorem 3.3 / Appendix A.1"}],"recommendation":"reject","confidential_remarks":"The flaw in Lemma 4.1 is not a presentation issue and is not local: it invalidates the paper's central theoretical claim. The counterexample is simple and directly refutes the reformulation as stated. I do not see a repair within the manuscript's current scope; a corrected result would require either additional assumptions on the noise kernel T_O or a different reformulation of the original coupling problem, and the dominance theorem would likely need to be re-derived from scratch. The statistical section may be salvageable, but the main contribution is not."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the takeaway: the central reformulation in Lemma 4.1 does not hold, and the paper's main claims about noisy data being a blessing are unsupported. The stress-test counterexample is legitimate, and the reader's reject verdict is right.\n\nThe genuinely useful pieces are the inverse-image construction and the statistical results. Pulling a Wasserstein ball back through a known noise kernel to define ambiguity over the latent distribution is a natural extension of Farokhi's total-variation work, and the paper spells it out clearly. Section 3's finite-sample guarantee and consistency results go through without the disputed reformulation and look formally sound. The price-of-ambiguity comparison is a reasonable framing, even if the definition is straightforward.\n\nThe problem is load-bearing. In the proof of Lemma 4.1, the coupling Π = (1/N)Σ δ_{x*_j} ⊗ G*_j has marginal constraint (1/N)Σ G*_j = T_O(F). The paper writes that a 'sufficient way' to satisfy this is to set G*_j = T_O(α_j) for a decomposition F = (1/N)Σ α_j, then reparameterizes as if this were an equivalence. It is not. Feasible couplings can have components outside the image of T_O, and restricting to the componentwise form changes the problem. The counterexample works: X={0,1}, mean-preserving noise uniform on {-1,1}, observed samples 0 and 2, ε=0.5. The original problem is feasible with value 0.5w; the reformulated program is infeasible and the dual is unbounded. Lemma 4.1 is false, and Theorem 4.2, the numerical experiments, and the price-of-ambiguity conclusions all inherit the flaw. The words 'sufficient way' at that step are an admission that equivalence is not being proven.\n\nThe dominance result may still be true—this counterexample refutes the proof, not the proposition—but the paper's stated contribution, the tractable convex reformulation, is not valid as written.\n\nRecommendation: the paper deserves a serious referee because the core idea and the statistical section merit engagement, but I would not cite the reformulation or dominance results in current form. If the authors can either prove the componentwise coupling is WLOG under stated conditions or find a genuinely equivalent lifted reformulation, the paper could be worth revisiting. As it stands, reject.","headline":"The inverse-image Wasserstein DRO idea is good and the statistical section looks sound, but Lemma 4.1 is false and the paper's main claims rest on it.","tokens_in":27154,"tokens_out":6452,"would_cite":false,"duration_ms":70762,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C47","90C34","49Q22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that explicitly modeling the noise process in distributionally robust optimization yields a provably higher worst-case utility than treating noisy data as clean, under mean-preserving noise, via an inverse-image ambiguity","keywords":["distributionally robust optimization","Wasserstein ambiguity set","noisy data","inverse-image construction","noise operator","price of ambiguity","mean-preserving noise","fair resource allocation"],"falsifier":"Enumerate the original max-min problem (5) on a small discretized instance — two or three latent states, two noisy samples, a two-point mean-preserving noise kernel — and compare the true worst-case value with the value of the convex reformulation (8). If any gap appears, or if the optimal coupling's components cannot all be written as T_O(α_j) for latent measures α_j, the reformulation is inexact and the proof of Theorem 4.2 no longer goes through. The same enumeration, run with a biased noise kernel of bias δ, would also check Proposition 4.1's bound g*_noise,δ(ε) ≥ g*(ε) − λ*δ.","tokens_in":26275,"feed_emoji":"🎲","tokens_out":10743,"duration_ms":117635,"temperature":0.7,"pith_summary":"This paper argues that noisy data, far from being a nuisance that must be cleaned away, can improve the guarantee a decision maker extracts from a distributionally robust optimization (DRO) model. The authors build the ambiguity set over the latent clean distribution by taking the inverse image of a Wasserstein ball centered at the noisy empirical distribution through a known noise kernel. Their central result, Theorem 4.2, is that under mean-preserving noise this 'noisy-data DRO' is less conservative than a direct DRO that treats noisy observations as clean: it has a weakly larger feasible region, a provably higher optimal value g*_noise(ε) ≥ g*(ε), and a lower price of ambiguity. The paper also proves a tractable convex reformulation, finite-sample out-of-sample guarantees, and asymptotic consistency, and shows numerically on a fair resource allocation problem that increasing ambiguity steers allocations toward equitable solutions while preserving more aggregate utility than raising the fairness parameter directly. If these claims hold, managers with corrupted data can model the noise as a source of robustness rather than an obstacle, and get better worst-case decisions than colleagues who ignore it.","feed_headline":"Modeling noise provably improves worst-case decisions","feed_subtitle":"A pull-back Wasserstein DRO shows noisy data can raise guaranteed utility and cut the price of ambiguity.","key_machinery":"The load-bearing object is the inverse-image ambiguity set B_{ε,O}(F̂⋆) = {F ∈ M(X) : d(T_O(F), F̂⋆) ≤ ε}, where T_O(F) = ∫ O(·|x)dF(x) is the noise operator pushing a latent distribution through a known noise kernel. Lemma 4.1 converts the max-min problem over this set into a semi-infinite convex program whose per-observation cost kernel is c_j(x) = ∫_{X⋆}∥x⋆ − x̂⋆_j∥ dO(x⋆|x); the dominance theorem then applies Jensen's inequality to this kernel under the mean-preserving assumption, showing the noisy kernel dominates the direct one pointwise.","core_discovery":"The paper's central claim is that the standard practice of applying DRO directly to noisy observations throws away information: because every norm is convex, a mean-preserving noise kernel inflates the expected transport cost to each observed sample, E_{O(·|x)}[∥X⋆ − x̂⋆_j∥] ≥ ∥x − x̂⋆_j∥, so each constraint of the direct model is relaxed pointwise and the feasible decision set grows. Consequently, for every radius ε > 0 the noisy-data DRO's worst-case optimal value dominates the direct model's, g*_noise(ε) ≥ g*(ε), and its price of ambiguity is no larger. The construction works by defining the latent ambiguity set as the pre-image of a Wasserstein ball through the noise operator, a pull-bac","pith_inferences":["The Jensen-inequality mechanism is generic: any mean-preserving spread inside a norm-based transport cost relaxes the same pointwise constraints, so the 'blessing of noise' should extend to other Wasserstein-based robust problems such as regression with noisy features or portfolio selection with noisy returns, even though the paper only demonstrates it for fair allocation.","The size of the blessing should be measurable and monotone in noise dispersion: for two mean-preserving noise kernels with the same mean but different variance, the gap g*_noise(ε) − g*(ε) should be larger for the higher-variance kernel; this is a direct testable corollary the paper does not state.","If the reformulation's decomposition step turns out to be inexact, the practical effect is that the numerical optimal values and POA gaps reported in Section 5 would be optimistic bounds rather than exact values; the dominance direction is plausible but the magnitudes would need re-derivation.","The ambiguity–fairness parallel hints at a broader design principle: a platform that cannot impose fairness constraints directly could deliberately inject calibrated ambiguity or noise into the robust model to steer allocations toward equity, a recipe the example suggests but the paper does not claim."],"forward_implications":["Decision makers with mean-preserving noisy observations can guarantee weakly higher worst-case utility by modeling the noise explicitly than by applying standard DRO to the noisy data.","The price of ambiguity — the relative utility loss from insisting on robustness — is no larger under noisy-data DRO, making robustness cheaper when noise is present.","The noiseless Wasserstein DRO becomes a special case (Dirac noise kernel), so the framework is a strict generalization of a widely used data-driven method.","Finite-sample out-of-sample guarantees at confidence 1−β and almost-sure convergence to the true stochastic optimum make the pull-back construction statistically sound for deployment.","In fair allocation, raising the ambiguity radius ε moves allocations toward equitable solutions while preserving more aggregate utility than raising the fairness parameter α, giving a lever for equity that does not sacrifice as much efficiency."],"supporting_citations":[{"why":"Provides the Wasserstein DRO paradigm, the concentration inequality the paper's Theorem 3.1 follows, and the direct model that serves as the dominance baseline.","marker":"Mohajerin Esfahani and Kuhn (2018)"},{"why":"Supplies the inverse-image pull-back principle for building latent-space ambiguity sets through a noise kernel, here ported from total variation to Wasserstein balls.","marker":"Farokhi (2023)"},{"why":"Establishes that noisy empirical distributions converge to the corrupted distribution F⋆ rather than F, the gap Lemma 3.1 closes in the Wasserstein metric.","marker":"Van Parys (2024)"},{"why":"Provides Lévy's continuity theorem and Borel–Cantelli, used in the noise-to-latent distance transfer and the almost-sure consistency proof.","marker":"Durrett (2019)"},{"why":"The price-of-fairness cost metric adapted to define the price of ambiguity and its dominance comparison.","marker":"Bertsimas et al. (2011)"},{"why":"Defines the α-fairness utility used in the fair resource allocation experiments.","marker":"Altman et al. (2008)"},{"why":"Underpins the choice of Wasserstein distance for distributions with non-overlapping supports.","marker":"Villani (2008)"}],"fun_headline_variants":["Noisy data can improve worst-case guarantees","Robust decisions profit from data noise","Wasserstein DRO finds hidden value in noise","Noise reduces cost of ambiguity in robust choice","Harnessing noise: less conservatism, better robust value"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The proof of the convex reformulation assumes that any transport coupling between the noisy empirical distribution and the noise-perturbed latent distribution can be split, one piece per observation, with each piece expressible as the noise-image of some latent component measure. This is shown to be sufficient but not necessary; if an optimal coupling requires pieces the noise operator cannot produce, the reformulation — and the dominance theorem built on it — might overstate","fun_headline_variants_meta":{"raw":{"variants":["Noisy data can improve worst-case guarantees","Robust decisions profit from data noise","Wasserstein DRO finds hidden value in noise","Noise reduces cost of ambiguity in robust choice","Harnessing noise: less conservatism, better robust value"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1253,"prompt_tokens":753,"completion_tokens":500,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":497,"tokens_out":500,"duration_ms":7097,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:57:52.852517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Enumerate the original max-min problem (5) on a small discretized instance — two or three latent states, two noisy samples, a two-point mean-preserving noise kernel — and compare the true worst-case value with the value of the convex reformulation (8). If any gap appears, or if the optimal coupling's components cannot all be written as T_O(α_j) for latent measures α_j, the reformulation is inexact and the proof of Theorem 4.2 no longer goes through. The same enumeration, run with a biased noise kernel of bias δ, would also check Proposition 4.1's bound g*_noise,δ(ε) ≥ g*(ε) − λ*δ.","supporting_citations":[{"cited_title":"and Kuhn, D","cited_arxiv_id":null,"evidence_quote":"Provides the Wasserstein DRO paradigm, the concentration inequality the paper's Theorem 3.1 follows, and the direct model that serves as the dominance baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the inverse-image pull-back principle for building latent-space ambiguity sets through a noise kernel, here ported from total variation to Wasserstein balls."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that noisy empirical distributions converge to the corrupted distribution F⋆ rather than F, the gap Lemma 3.1 closes in the Wasserstein metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Lévy's continuity theorem and Borel–Cantelli, used in the noise-to-latent distance transfer and the almost-sure consistency proof."},{"cited_title":"F., and Trichakis, N","cited_arxiv_id":null,"evidence_quote":"The price-of-fairness cost metric adapted to define the price of ambiguity and its dominance comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the α-fairness utility used in the fair resource allocation experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Underpins the choice of Wasserstein distance for distributions with non-overlapping supports."}],"review_version":1}