{"id":"794e9864-aa3c-4140-8e0a-64e36e3cf62e","arxiv_id":"2608.13131","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Huber-Wasserstein barycenters, defined by Huberizing the ground cost of optimal transport, are robust to outlying distributions, have breakdown point near 1/2, and interpolate between Wasserstein means and medians.","lead":"This paper proposes a robust average for collections of probability distributions by applying the Huber loss to individual transport displacements inside optimal transport, rather than to the final Wasserstein distance. It proves the resulting Huber-Wasserstein barycenters exist, are consistent, resist contamination up to half the sample, and interpolate between Wasserstein means and medians.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.16 is internally contradictory: the stated IF_out = c√α conflicts with the proof's c/√α, and the localized-contamination claim depends on which value is right.","rationale":"The central existence and consistency results — Theorem 3.4, Proposition 3.5, Theorem 3.11, and the one-dimensional influence-function/asymptotic-normality results (Propositions 3.13 and 3.14) — appear internally sound, and I found no flaw that would overturn them. The reader's chosen weakest assumption, the α-Lipschitz condition in Theorem 2.12, is a genuine limitation on the quantitative map-stability claim, but it is not load-bearing for the central barycenter claim: Theorem 3.6's interpolation limits are proved from cost convergence, not from map stability, and no later result depends on the map bound. The more pressing issue is Proposition 3.16, where the statement and proof give different values for IF_out. Because that proposition is the quantitative evidence for the abstract's localized-contamination advantage, the contradiction must be resolved before the manuscript can be accepted as written. I therefore keep the reader's CONDITIONAL verdict: the correction is localized and the main theoretical framework survives, but the text currently contains a stated result that is not the result its own proof establishes.","tokens_in":52583,"tokens_out":31540,"duration_ms":273737,"concrete_test":"Independently re-derive the outside-Huber influence function in Proposition 3.16 by solving the first-order condition for q_ε near 0: (1-ε) q_ε + ε c (q_ε - Q_η)/||q_ε - Q_η||_{L2} = 0. Dividing by ε and taking ε→0 gives q_ε/ε = c Q_η/||Q_η||_{L2} = c/√α 1_{(1-α,1)}. If this algebra is confirmed, correct the proposition's stated IF_out from c√α to c/√α and retain the qualitative conclusion; if an independent derivation produces c√α, reconcile the proof accordingly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and Section 3.4.1 advertise that displacement-wise Huberization retains first-order information lost by distance-based Huberization under localized shape contamination. Proposition 3.16 is the quantitative support: it states IF_in = c 1_{(1-α,1)} and IF_out = c√α 1_{(1-α,1)} for the contaminant η_M,α = (1-α)δ_0 + αδ_M. The proof, however, computes q_ε/ε → c Q_η / ||Q_η||_{L2} = c/√α 1_{(1-α,1)}. Since ||Q_η||_{L2} = M√α, the unit direction is (1/√α) on the upper block, so the limiting pointwise influence is c/√α, not c√α. The two readings have opposite implications for α < 1: the stated c√α would make the outside estimator more robust than the inside one, undercutting the paper's claim; the proof's c/√α supports the claim. This is an internal inconsistency in a motivating result, and Proposition 3.16 cannot be used as written. It does not threaten the core existence, consistency, breakdown-point, or interpolation theorems, but it must be settled by a direct re-derivation before the localized-contamination message is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a robust barycenter for distribution-valued data by placing the Huber loss inside the optimal transport ground cost rather than outside as in previously studied metric-space Huber means. The main object is m_c(P) = argmin_{ν∈P1(Rd)} Γ_{c,P}(ν), where Γ_{c,P}(ν) integrates the centered Huber transport cost T_{ρ_c}(ν,μ) − c∫||x||dμ(x). The authors establish existence, convexity, boundedness, and weak closedness of the barycenter set; consistency under one- and two-stage sampling; quantitative convergence to Wasserstein medians as c↓0 and to quadratic Wasserstein barycenters as c↑∞; a characterization via averaged Kantorovich potentials; a finite-sample breakdown point essentially 1/2; and, in dimension one, quantile characterizations, pointwise influence functions, asymptotic normality, and comparative efficiency with distance-based Huberization. Numerical experiments on MNIST images, London Underground flow profiles, and synthetic score distributions illustrate the interpolation between Wasserstein means and medians and the advantage of displacement-wise clipping under localized contamination. The supplementary appendix contains full proofs of the main theorems, including a decomposition-based proof of existence of Huber optimal transport maps, and a Sinkhorn-type algorithm with convergence guarantees for the regularized barycenter problem.","tokens_in":52856,"tokens_out":6174,"duration_ms":59016,"significance":"If the results are correct, the paper makes a useful contribution to robust distribution-valued data analysis. The construction is natural: it preserves the convex geometry of transport-based barycenters while clipping large individual displacements, and it provides an explicit interpolation between Wasserstein means and W1 Wasserstein medians. The theoretical package is substantial and mostly self-contained: existence, consistency, breakdown, and one-dimensional influence-function results are proved in detail, and the computational appendix supplies a convergent fixed-grid algorithm with reproducible code. The finite-sample breakdown-point theorem and the explicit scalar influence-function formulas are concrete, falsifiable statements. The central theoretical claims do not rely on fitted quantities and are not circular: the cutoff c is a user-specified parameter in the theory, and the data-driven c-selection in the experiments does not feed back into the theorem statements.","major_comments":[{"comment":"The stated influence function for the outside-Huber barycenter, IF_out = c√α 1_{(1−α,1)}, is inconsistent with the proof in Appendix A.2. The proof derives q_ε/ε → c Q_η/||Q_η||_{L2} = c/√α 1_{(1−α,1)}, since ||Q_η||_{L2} = M√α. Thus the pointwise outside influence is c/√α, not c√α. This is not a cosmetic typo: the two readings have opposite implications for α<1. The printed c√α would make the outside estimator more robust than the inside one and undercut the paper's localized-contamination claim, whereas the proof's c/√α supports the advertised conclusion. Proposition 3.16 is the quantitative support for the abstract's claim that displacement-wise Huberization retains first-order information lost by distance-based Huberization, so the discrepancy must be resolved by a direct re-derivation before that message is accepted. The core existence, consistency, breakdown, and interpolation theorems are not affected, but Proposition 3.16 cannot be cited as written.","section":"Section 3.4.1, Proposition 3.16"},{"comment":"The quantitative map stability result assumes that the quadratic optimal transport map H0 is α-Lipschitz, equivalently that the conjugate of the quadratic potential is 1/α-strongly convex. The paper gives no example of a pair (μ,ν) satisfying this condition and no fallback bound when it fails. For rough or non-Lipschitz quadratic maps, the L2 convergence of Huber OT maps to H0 as c→∞ is therefore not established; only the cost-level convergence of Proposition 2.11 is unconditional. Since the abstract and Section 2 list 'stability as the Huber parameter varies' among the contributions, this missing support should be addressed, either by exhibiting a nontrivial class satisfying the condition or by providing a bound under weaker assumptions.","section":"Section 2.4, Theorem 2.12"}],"minor_comments":[{"comment":"The sentence 'If μ≪L^d is such that μ(∂supp(μ)) = 0 and supp(μ), the directional derivative is linear...' is incomplete; it should presumably read 'and supp(μ) is connected'.","section":"Section 3.2, after Lemma 3.8"},{"comment":"The Sinkhorn regularization is chosen as ε_c = η times the 0.90-quantile of positive entries of the Huber cost matrix, so ε depends on both c and the data. The statement that observed changes along the c-path are attributable only to the Huber loss would be more convincing with a sensitivity check over η.","section":"Section 4.1"},{"comment":"The paper alternates between 'Huber–Wasserstein barycenter', 'Huber–Wasserstein mean', and 'Huber center' for the same object. Consistent terminology would help readers track which estimator is being discussed, especially in Tables 1 and 2.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the Proposition 3.16 discrepancy appears to be an algebraic slip in the statement rather than a flaw in the underlying derivation, since the proof's value c/√α supports the paper's intended conclusion. The unverified Lipschitz assumption in Theorem 2.12 is the more substantive gap. Both should be resolved before acceptance; the rest of the paper is sound in scope and well suited to stat.ME."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The construction is genuinely new: Huber loss placed inside the ground cost rather than outside the Wasserstein distance. The core theory is largely sound and the paper deserves a serious referee, but there is one concrete internal contradiction in Proposition 3.16 that must be resolved.\n\nWhat the paper does well: it separates itself cleanly from Lee–Jung metric Huber means, ROBOT hard truncation, and Brizzi et al.’s strictly convex h-Wasserstein barycenters. Existence and consistency (Thms 3.4, 3.5), the essentially-1/2 breakdown point (Thm 3.11), and the interpolation limits as c→0 and c→∞ (Thm 3.6) are real contributions. The one-dimensional influence function and asymptotic-efficiency results are useful and clearly derived. Theorem 2.8, existence of an optimal map for the Huber cost, is a heavy piece of work; the proof sketch is plausible. The related-work discussion is honest and the citation pattern looks appropriate. No fitted quantities feed back into the theory: c is a user parameter, so circularity is not an issue.\n\nSoft spots, in proportion. The biggest is Proposition 3.16. The stated influence function for the outside-Huber center is c√α, while the proof derives c/√α. The proof’s value is the one consistent with the derivation: Qη divided by its L2 norm has amplitude 1/√α on the upper block. The stated c√α would, for α<1, actually undercut the paper’s claim that inside Huberization is more robust under localized contamination; the proof’s c/√α supports that claim. As written, the proposition cannot be used. It is a motivating example rather than a load-bearing theorem, but it is advertised in the abstract and in Section 3.4.1, so it needs a direct re-derivation and corrected statement.\n\nTheorem 2.12 assumes the quadratic OT map is α-Lipschitz without giving examples or a fallback bound; that limits the quantitative map-stability claim, though the qualitative convergence is fine. Minor: the numerical experiments rely on single seeds without repeated-seed error bars, and the data-driven choice of c is reasonable but not deeply analyzed. Nothing else struck me as unsound.\n\nBottom line: I would send this to a competent referee, not desk-reject. The core results are defensible and the construction is worth publishing. The referee should insist on the Proposition 3.16 fix and a caveat on Theorem 2.12’s condition.","headline":"New construction, solid core theory, but Proposition 3.16 has a formula/proof mismatch that needs fixing before the localized-contamination claim is trusted.","tokens_in":53366,"tokens_out":2676,"would_cite":false,"duration_ms":25428,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G35","62G30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a robust barycenter for distribution-valued data can be built by applying the Huber loss to individual transport displacements inside the optimal transport cost, with existence, consistency, breakdown point ~1/2, and…","keywords":["Huber loss","Wasserstein barycenter","optimal transport","robust statistics","distribution-valued data","breakdown point","influence function","quantile functions"],"falsifier":"Take $\\mu$ uniform on $[0,1]$ and $\\nu$ the uniform mixture of $[0,1]$ and $[2,3]$; the quadratic transport map is a step function, hence not $\\alpha$-Lipschitz. Compute the one-dimensional Huber-optimal maps $H_c$ for increasing $c$ from the quantile equation and check whether $\\|H_0-H_c\\|_{L^2(\\mu)}$ tends to zero at the rate predicted by Theorem 2.12. If it does not, the Lipschitz assumption is essential.","tokens_in":52396,"feed_emoji":"📊","tokens_out":9798,"duration_ms":83713,"temperature":0.7,"pith_summary":"This paper tries to establish that a robust center for a collection of probability distributions can be obtained by putting the Huber loss inside the optimal transport cost, one displacement at a time, rather than applying it to the Wasserstein distance after optimization. The proposed Huber–Wasserstein barycenter minimizes $\\Gamma_{c,P}(\\nu)=\\int(T_{\\rho_c}(\\nu,\\mu)-c\\int\\|x\\|\\,d\\mu)\\,dP(\\mu)$, where $T_{\\rho_c}$ is the optimal transport cost with ground cost $\\rho_c$. The paper proves this minimizer exists, is bounded and convex, is consistent under both one-stage and two-stage sampling, and has a finite-sample breakdown point essentially $1/2$. As $c$ varies it interpolates between the classical Wasserstein mean and the $L^1$-type Wasserstein median, and in one dimension its quantile function solves a scalar Huber location problem at every level, giving a bounded influence function and an explicit efficiency trade-off. A reader should care because this yields a robust alternative to Wasserstein barycenters that keeps the convex landscape of transport-based objectives.","feed_headline":"Huber cost inside optimal transport yields robust distribution centers","feed_subtitle":"New barycenter resists up to half contaminated distributions and interpolates between Wasserstein mean and median.","key_machinery":"The carrying object is the Huber ground cost $\\rho_c(t)=\\frac12 t^2$ for $|t|<c$ and $\\rho_c(t)=c|t|-\\frac{c^2}{2}$ for $|t|\\ge c$, used inside the optimal transport problem $T_{\\rho_c}(\\mu,\\nu)=\\inf_\\pi \\int \\rho_c(\\|x-y\\|)\\,d\\pi(x,y)$. The barycenter objective subtracts $c\\int\\|x\\|\\,d\\mu$ so that $\\Gamma_{c,P}$ is finite for every $P$ on $P_1(\\mathbb{R}^d)$ without extra moment conditions. Three properties do the work: the dual potentials are $c$-Lipschitz, which yields Lipschitz stability of the cost in $W_1$; the quadratic-region part of an optimal Huber plan is cyclically monotone for the squared cost while the linear-region part reduces to an $L^1$ optimal transport ray decomposition, giving existence of a Monge map $H_c$; and in $d=1$ the quantile representation makes the barycenter a pointwise scalar Huber location problem with score $\\psi_c$. The averaged-potential characterization of minimizers connects the population problem to the empirical one.","core_discovery":"On the paper's own terms, the central claim is that the functional $\\Gamma_{c,P}$ is a well-posed robust barycenter: its minimizer set $m_c(P)$ is nonempty, bounded, convex, and closed in $W_1$; empirical plug-in barycenters converge almost surely under one- and two-stage sampling; and the finite-sample breakdown point is essentially $1/2$. In dimension one the barycenter is characterized pointwise in quantile space by $\\int \\psi_c(Q_\\nu(u)-Q_\\mu(u))\\,dP(\\mu)=0$, where $\\psi_c(t)=\\max\\{-c,\\min\\{t,c\\}\\}$, so each quantile displacement is clipped separately. The paper also establishes that the Huber-optimal transport problem admits a deterministic transport map for absolutely continuous sources, that optimal Kantorovich potentials are unique up to constants and $c$-Lipschitz, that barycenter minimizers are characterized by averaged Huber potentials, and that $c\\to\\infty$ recovers the quadratic Wasserstein barycenter while $c\\to 0$ recovers the $W_1$-type Wasserstein median after normalization. The distinction from metric-space Huber means is explicit: Huberizing inside the transport cost preserves convexity and clips individual displacements rather than whole distributions.","pith_inferences":["Beyond the paper, the same piecewise-quadratic/linear proof strategy should produce Monge solutions for any ground cost that is quadratic near zero and affine outside a cutoff, so the existence theory likely extends to smooth truncations of the quadratic cost.","The cellwise-versus-casewise analogy suggests a practical rule not tested here: when contamination is localized in a known region of the distributions, small $c$ should be chosen to preserve clean regions, whereas when entire observations are suspect, the outside-Huber construction may be preferable.","In $d>1$, the formal map-level comparison indicates that a measure with an anomalous region producing large displacements from the center should still contribute full weight in regions where its transport displacement is small; a direct numerical test on localized spatial contamination would show whether the one-dimensional gain persists.","Because the breakdown point reaches essentially $1/2$ while the influence function stays bounded, the estimator occupies a useful middle ground between the non-robust Wasserstein mean and the more aggressive Wasserstein median, which may make it attractive for adversarially contaminated distributional datasets."],"forward_implications":["Varying $c$ traces a continuous path from the classical Wasserstein barycenter ($c\\to\\infty$) to the $W_1$-type Wasserstein median ($c\\to 0$), so the method supplies a one-parameter family of robust centers rather than a single estimator.","In one dimension the estimator can be computed exactly by solving a scalar Huber location problem at each quantile level, with bounded influence function and asymptotic normality, making the robustness–efficiency trade-off explicit.","The finite-sample breakdown point is essentially $1/2$: replacing fewer than half of the observed distributions cannot drive the empirical barycenter arbitrarily far, while replacing slightly more than half can.","One-stage and two-stage plug-in barycenters are consistent, so the robust center remains valid when each observed distribution is itself estimated from its own sample.","When contamination changes only part of a distribution, clipping each transport displacement preserves information in unaffected regions, while random-translation contamination makes the inside- and outside-Huber constructions coincide."],"supporting_citations":[{"why":"Introduced Wasserstein barycenters, the quadratic-objective object that this paper robustifies by replacing the squared cost with the Huber cost.","marker":"Agueh and Carlier (2011)"},{"why":"Supplied the general existence and consistency theory for Wasserstein barycenters against which the new population theory is benchmarked.","marker":"Le Gouic and Loubes (2017)"},{"why":"Defined the $W_1$-type Wasserstein medians that appear as the $c\\to 0$ limit of the proposed barycenters.","marker":"Carlier et al. (2024)"},{"why":"Proposed metric-space Huber means, the outside-Huber construction the paper contrasts with and shows to differ under localized contamination.","marker":"Lee and Jung (2026)"},{"why":"Provided the $L^1$ optimal transport ray decomposition used in the proof that Huber-optimal transport admits a deterministic Monge map.","marker":"Ambrosio and Pratelli (2003)"},{"why":"Gave uniqueness of Kantorovich potentials for non-smooth costs, used to prove uniqueness of the Huber dual potentials up to additive constants.","marker":"Staudt et al. (2025)"},{"why":"Established the classical scalar Huber location estimator whose influence function, efficiency formula, and $c$ calibration the one-dimensional analysis follows.","marker":"Huber (1964)"},{"why":"Provided the entropic barycenter iterations that the appendix adapts into a Huber–Sinkhorn algorithm for computing the proposed barycenters.","marker":"Cuturi and Doucet (2014)"},{"why":"Supplied the general Kantorovich duality and optimality criteria on which the Huber-cost transport duality theorem rests.","marker":"Villani (2009)"}],"fun_headline_variants":["Huber cost inside OT yields robust barycenters","Clipped displacements give robust distribution means","Huber-Wasserstein barycenters resist half contamination","New barycenter: mean-median interpolation with breakdown 1/2","Displacement-wise Huberization beats distance-based robustness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative claim that the Huber transport map approaches the quadratic transport map as $c$ grows large depends on the quadratic transport map being Lipschitz with a fixed constant; the paper gives no example where this holds and no alternative bound when it fails.","fun_headline_variants_meta":{"raw":{"variants":["Huber cost inside OT yields robust barycenters","Clipped displacements give robust distribution means","Huber-Wasserstein barycenters resist half contamination","New barycenter: mean-median interpolation with breakdown 1/2","Displacement-wise Huberization beats distance-based robustness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001259,"raw_usage":{"total_tokens":5196,"prompt_tokens":1022,"completion_tokens":4174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":4097}},"tokens_in":638,"tokens_out":4174,"duration_ms":26659,"temperature":1.0,"reasoning_tokens":4097,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:10:41.039866+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $\\mu$ uniform on $[0,1]$ and $\\nu$ the uniform mixture of $[0,1]$ and $[2,3]$; the quadratic transport map is a step function, hence not $\\alpha$-Lipschitz. Compute the one-dimensional Huber-optimal maps $H_c$ for increasing $c$ from the quantile equation and check whether $\\|H_0-H_c\\|_{L^2(\\mu)}$ tends to zero at the rate predicted by Theorem 2.12. If it does not, the Lipschitz assumption is essential.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defined the $W_1$-type Wasserstein medians that appear as the $c\\to 0$ limit of the proposed barycenters."},{"cited_title":"Optimal Transportation and Applications , series =","cited_arxiv_id":null,"evidence_quote":"Provided the $L^1$ optimal transport ray decomposition used in the proof that Huber-optimal transport admits a deterministic Monge map."},{"cited_title":"Proceedings of the 31st International Conference on Machine Learning , series =","cited_arxiv_id":null,"evidence_quote":"Provided the entropic barycenter iterations that the appendix adapts into a Huber–Sinkhorn algorithm for computing the proposed barycenters."}],"review_version":1}