{"id":"5a94f622-e053-498c-aebc-ad71a351797c","arxiv_id":"2608.02487","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Iterative c-rectified flow converges to optimal transport under regularity assumptions, and score-based plug-in estimation yields near-optimal transport-map rates.","lead":"Ordinary rectified flow can stop at a suboptimal transport coupling in Gaussian settings, but this paper proves that a cost-aware variant, c-rectified flow, converges to the optimal transport coupling under regularity conditions. It also derives minimax score-estimation rates and uses them to build a near-optimal statistical estimator of the optimal transport map.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's convergence depends on an unproved uniform C^{2+α,1+α} bound and uniform integrability for the iterated potentials; the paper never shows the c-rectified update propagates these from the initial coupling.","rationale":"The reader identified the uniform Hölder/compactness and uniform-integrability assumptions as the weakest load-bearing premise, and I agree. These assumptions are not consequences of the c-rectified update; they are external regularity hypotheses. The proof of Theorem 1 cannot proceed without them: the diagonal Arzelà–Ascoli argument requires the C^{2+α,1+α} bounds, and Lemma 5 requires the uniform integrability to interchange limits and expectations. No propagation theorem is stated for these bounds from the initial coupling to the iterates. The Gaussian section gives the only explicit verification, where the potentials are quadratic and the bounds are easy. For the general qualitative claim, the paper leaves an unverified condition at the center of the argument. This does not invalidate the theorem as a conditional statement, but it means the advertised 'always converges' conclusion rests on a strong, non-obvious assumption. The reader's CONDITIONAL verdict already reflects this, so my stress-test does not change the recommendation. The proposed numerical test would directly probe whether these hypotheses can fail for a natural smooth non-Gaussian example, thereby settling whether the concern is merely technical or a genuine restriction on the algorithm's domain of convergence.","tokens_in":56799,"tokens_out":20758,"duration_ms":198445,"concrete_test":"Implement a high-accuracy population-level simulation of Algorithm 1 in d=2 for a smooth non-Gaussian pair (e.g., P a Gaussian mixture and Q its pushforward by a known non-optimal map). For K=1,...,50, compute the potentials f^{(k)} by solving the projection onto gradient fields (e.g., via a spectral/Helmholtz solver on a grid), and monitor M_K := sup_{k≤K} ||f^{(k)}||_{C^{2+α,1+α}(B(0,R))} and U_K := sup_{M>0} sup_k E∫ g(Z_t^{(k)})^2 1_{g^2>M} dt. If M_K or U_K diverges with K, Theorem 1's hypotheses are not automatic; if they stay bounded, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is in Theorem 1: convergence to the OT coupling is made to depend on a uniform-in-k C^{2+α,1+α} bound for the minimizers f^{Z(k)} and on uniform integrability of ∫ g(Z_t^k)^2 dt. These are not derived from the initial coupling or from the c-rectified update; they are assumed. The proof uses the Hölder bound to run a diagonal Arzelà–Ascoli compactness argument (Appendix D) and uses the UI hypothesis to pass E[V_k]→E[V_*] via Lemma 5. Without both, the subsequential limit f* may fail to exist or L_{Z*}(f*)=0 may not follow, so the conclusion that the limit coupling is optimal collapses. The paper only verifies these conditions in the Gaussian/linear case (Section E), where potentials are quadratic; for general P,Q it provides no propagation mechanism. Thus the central qualitative claim 'iterative c-rectified flow always converges to optimal transport' is conditional on an opaque regularity condition that could fail for natural initial couplings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the c-rectified flow, a cost-aware variant of rectified flow that projects the learned velocity onto a gradient class while preserving endpoint marginals. Its main claims are: (i) ordinary rectified flow can stabilize at a non-optimal Gaussian coupling when the source and target covariance matrices do not commute; (ii) under compactness and uniform-integrability hypotheses, iterative c-rectified flow converges weakly to the optimal transport coupling and the transport cost converges to the optimal value; (iii) under projection-stability assumptions the iteration contracts the excess cost geometrically, both for quadratic and for strongly convex displacement costs; and (iv) a new score-estimation minimax theory yields plug-in estimators of optimal transport maps at rates that are claimed to be optimal for d>=3 and nearly parametric for d=1,2. The paper contains detailed proofs of the Gaussian calculations, explicit matrix formulas for the contractions, and matching upper and lower bounds for score estimation in the Gaussian-ratio model.","tokens_in":57036,"tokens_out":10724,"duration_ms":125804,"significance":"If the main structural claims are fully established, the paper would be a valuable contribution: it identifies a clear failure mode of vanilla rectified flow, proposes a principled cost-aware correction, and provides quantitative and statistical guarantees that are rare in this area. The Gaussian counterexample (Proposition 2) is clean and checkable, the Gaussian projection calculations in Section E are useful, and the score-estimation lower bounds in Appendix H are substantial. However, the central qualitative convergence theorem is conditional on regularity and uniform-integrability hypotheses whose propagation under the c-rectified update is not proved, and the quantitative contraction theorems require uniform projection stability that is only verified locally. The statistical optimality claim likewise relies on oracle operations rather than on a finite-step analysis of the iterative algorithm. The significance of the paper is therefore real but currently more prospective than established.","major_comments":[{"comment":"The main qualitative claim is made conditional on two hypotheses: a uniform C^{2+alpha,1+alpha} bound for the minimizers f^{Z(k)} on every ball, and uniform integrability of the family of integrals of g(Z_t^k)^2. These hypotheses are assumed for every k, but no lemma or proposition shows that they are propagated by the c-rectified update from the initial coupling. The proof uses them to run the diagonal Arzela-Ascoli argument and to pass E[V_k] to E[V_*] via Lemma 5; without them the subsequential limit f* and the identity L_{Z*}(f*)=0 are not justified. The only verification is the Gaussian/linear case in Section E, where the potentials are quadratic. The abstract's formulation that iterative c-rectified flow 'always converges' is therefore too strong as stated. This is the load-bearing step for the qualitative convergence claim and needs either a propagation theorem for natural classes","section":"Theorem 1 and Appendix D"},{"comment":"The exponential convergence results require Assumption 3 (resp. Assumption 5) to hold for every iterate T_k with the same constants. This is not derived from the one-step contraction theorem; it is an additional uniform structural assumption. Proposition 4 proves Assumption 3 only along smooth P-measure-preserving perturbations of the identity near T*, not that the iterates remain in the neighborhood where such a bound holds. Remark 1 acknowledges that entry into the admissible class is 'expected to occur' but no proof is given. Consequently, the advertised quantitative convergence guarantees are not established for arbitrary initial couplings; they are conditional on a condition that is verified only locally.","section":"Theorems 4-5 and 10-11; Assumptions 3 and 5"},{"comment":"The statistical estimator in Algorithm 3 requires, in Step 4, the exact optimal transport map between the projected estimated marginals, and the parenthetical 'equivalently the limiting map obtained by Algorithm 1' invokes the iterative c-rectified flow. However, Theorems 1 and 9 provide only asymptotic convergence and only under assumptions that are not verified for the plug-in marginals; no finite-K error bound is given. Thus the stated risk bound O(rho_n^2) in Proposition 5 is an oracle plug-in bound, not a guarantee for an estimator obtained by running finitely many steps of Algorithm 1. To support the claim that iterative c-rectified flow yields a rate-optimal estimator, the paper needs either a finite-iteration bias analysis or an explicit statement that the statistical theorem concerns the exact OT map and only the asymptotic guarantee of the iterative approximation.","section":"Section 4.2 and Algorithm 3"},{"comment":"The text describes the resulting estimator as 'rate-optimal' for optimal transport map estimation, but no matching lower bound for the map risk is proved in the manuscript. Theorem 6 gives minimax lower bounds only for the score-estimation problem; Theorem 7 gives upper bounds for the Wasserstein marginal risk; Proposition 5 transfers that upper bound to the map risk. To justify the word 'rate-optimal', the authors need either to prove a lower bound for the map estimation problem under the same class, or to cite precisely and verify the hypotheses of an existing minimax lower bound for OT maps (for instance, the smooth-Brenier-map lower bound of Hutter-Rigollet or Manole et al.) and show that the implied map smoothness is inherited from F_{alpha,d} and C^lambda_Lambda. As it stands, the optimality claim is an assertion rather than a proved consequence.","section":"Section 4.2, Proposition 5"}],"minor_comments":[{"comment":"The symbol P is used both for a probability measure (e.g. P(A), P(X)) and for the set of probability measures P(A). This overload is confusing in several places and should be disambiguated.","section":"Notation, Section 1.4"},{"comment":"The phrase 'Denote (Z*_0,Z*_1) as the unique optimal coupling' introduces uniqueness as an assumption, but it is not listed among the theorem hypotheses. Since uniqueness is needed to pass from subsequential convergence to full weak convergence, it should be stated explicitly as an assumption.","section":"Theorem 1"},{"comment":"Typo: 'a initial coupling' should be 'an initial coupling'. There are similar minor grammar issues elsewhere in Section 3.3.","section":"Theorem 5"},{"comment":"The notation '1 ∨ log_+(nt^{alpha+d/2})^{d/2}' is ambiguous: the exponent d/2 appears to apply to the whole logarithm expression in the first displayed branch, but the later derivations treat it as (log(...))^{d/2}. Clarify the convention.","section":"Equation (10)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth taking seriously. The genuinely new pieces are the qualitative convergence of iterative c-rectified flow to the optimal transport coupling, the explicit Gaussian contraction estimates, and the multivariate score minimax rates with matching lower bounds. The Gaussian section is the cleanest part: the commutativity condition for vanilla rectified flow is derived carefully, and the spectral criterion behind Theorem 2 is checkable. The proof of Theorem 1 via diagonal Arzelà–Ascoli is plausible, and the uniform-integrability hypothesis does exactly the work it is assigned. The score-rate machinery in Section 4 is substantial: the three-regime lower bound construction is nontrivial and the projected reverse sampler is a real algorithmic statement, not a placeholder.\n\nThe soft spots are where the reader put them, and I agree with most of that assessment. The stress-test criticism lands: Theorem 1 assumes, rather than establishes, the uniform C^{2+α,1+α} bound and the uniform integrability of the iterated potentials. There is no propagation argument showing the c-rectified update preserves these from the initial coupling; the paper only verifies them in the Gaussian/linear case where the potentials are quadratic. So the strong reading of “iterative c-rectified flow always converges to optimal transport” is conditional on an opaque regularity condition. The paper is not hiding this—the assumptions are stated explicitly—but the abstract and introduction let the qualifier drift out of view, and Remark 1 is where the unresolved gap lives.\n\nThe statistical claim is also slightly oversold. The rate-optimal estimator in Algorithm 3 is the exact optimal transport map between projected estimated marginals, not a finite-iteration sample-based c-rectified flow. Calling the infinite-iteration limit “equivalent” is true in the population sense but is not what the practitioner runs. That mismatch should be fixed in the writing. The general contraction theorems (4–5, 10–11) rest on projection stability, which is only verified locally near T*; that is a real limitation but a proportionate one, since the qualitative convergence result does not depend on it.\n\nCitation-wise, the paper properly credits the earlier Gaussian commutativity observations in Hertrich et al. and Mena et al., and the reliance on Liu (2022) for the fixed-point characterization is an import, not a circular step. No invented entities, one honest free parameter in the local-polynomial bandwidth.\n\nWho gets value: anyone working on flow-based transport or statistical estimation of optimal transport maps. It deserves a serious referee, not a desk reject. I would ask for a statement of Theorem 1 that is unambiguous about the regularity assumptions, a propagation lemma or concrete families where the uniform bounds hold, and a reworded statistical claim that separates the exact plug-in estimator from the iterative algorithm.","headline":"Real theoretical contributions in the qualitative convergence theorem and score minimax rates, but the headline 'always converges' is conditional on regularity assumptions the paper never derives, and the statistical claim trades on an exact plug-in rather than the iterative algorithm.","tokens_in":57556,"tokens_out":2006,"would_cite":true,"duration_ms":27617,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","62G05","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that iterating c-rectified flow—a cost-aware variant that projects velocities onto gradient fields while preserving marginals—converges to the optimal transport coupling under compactness and uniform-integrability assumptio","keywords":["optimal transport","rectified flow","c-rectified flow","statistical estimation","convergence rates","score estimation","minimax rates","Wasserstein distance"],"falsifier":"Run the population c-rectified iteration on centered Gaussian marginals N(0, Σ1), N(0, Σ2) with Σ1 and Σ2 chosen so that Σ1Σ2 ≠ Σ2Σ1 (say, Σ1=diag(4,1), Σ2 a rotation of it). The paper predicts the excess cost E||Z1^k−Z0^k||²−W₂²(P,Q) decays exponentially. If it stays bounded away from zero after many iterations, Theorem 1's convergence claim fails.","tokens_in":56636,"feed_emoji":"🎯","tokens_out":8875,"duration_ms":73829,"temperature":0.7,"pith_summary":"This paper tries to establish that a cost-aware variant of rectified flow, called c-rectified flow, is a principled algorithm for optimal transport. Ordinary rectified flow learns a velocity field by regression and can be iterated, but the paper shows in a Gaussian example that it can stabilize at a deterministic coupling that is not the optimal transport map unless the source and target covariance matrices commute. The proposed variant projects the learned velocity onto a gradient class compatible with the transport cost while preserving endpoint marginals; under compactness and uniform-integrability assumptions, the paper proves that iterating this projection drives the transport cost to the optimal value and the couplings converge weakly to the unique optimal coupling. Under additional projection-stability conditions, the convergence is quantitative: one step removes a fixed fraction of the excess cost, and iteration gives exponential decay, for both quadratic and strongly convex displacement costs. On the statistical side, the paper develops minimax-optimal score estimators for a Gaussian-ratio model and shows that plugging them into iterated c-rectified flow yields a rate-optimal estimator of the optimal transport map for d≥3 and a nearly parametric rate for d=1,2.","feed_headline":"Cost-aware projection makes iterated flow reach optimal transport","feed_subtitle":"Vanilla rectified flow can stall at suboptimal couplings; the cost-aware variant provably converges to the optimal map.","key_machinery":"Central object: the c-rectified flow update. From a coupling (X0,X1), take the linear interpolation X_t=(1−t)X0+tX1 and minimize L_{X,c}(f)=∫₀¹ E[m_c(Ẋ_t,∇f_t(X_t))]dt, where m_c is the Bregman divergence of the convex cost c; the new process follows dZ_t=∇c*(∇f_t(Z_t))dt. In the quadratic case, this is orthogonal projection of the conditional velocity onto gradient fields, i.e. removal of the divergence-free part. The key identity: a fixed point of this iteration has L=0 exactly for c-optimal couplings, so any subsequential limit with L=0 is the unique optimal coupling. Contraction rests on projection stability—a fixed fraction of T−T* must lie in the divergence-free subspace—which in the G","core_discovery":"c-rectified flow is claimed to be the right algorithmic object: unlike ordinary rectified flow, which in Gaussian cases can stabilize at a suboptimal coupling unless covariance matrices commute, the cost-aware gradient projection makes iteration converge to the optimal transport coupling. Under uniform Hölder and uniform-integrability assumptions, the transport cost along iterates converges to W₂²(P,Q) and the couplings converge weakly to the unique Brenier coupling. Under a projection-stability condition, one step contracts the excess cost by a fixed factor, giving exponential decay, for quadratic and general strongly convex displacement costs. The paper also supplies minimax score-estimati","pith_inferences":["Editorial inference: the qualitative convergence theorem likely extends to any setting with a unique optimal coupling and uniform compactness of the potentials; non-Gaussian marginals with log-concave densities are natural candidates, though the paper does not prove this.","Editorial inference: the projection-stability condition suggests a practical stopping rule—monitor the size of the divergence-free component of the velocity; when it vanishes, the iterates have reached the optimal fixed point.","Editorial inference: the Gaussian counterexample implies that any implementation of vanilla reflow on noncommuting covariance data silently converges to a suboptimal map; a cost-aware projection is therefore not an optional refinement but a necessary correction.","Editorial inference: since the plug-in analysis only uses Wasserstein control of the marginals, replacing the score-based marginal estimator with any other rate-optimal marginal estimator would transfer the same optimal transport map rate to c-rectified flow."],"forward_implications":["Iterating c-rectified flow is a principled population-level method for optimal transport: the transport cost converges to the optimal value and the coupling converges weakly to the unique optimal coupling.","Vanilla rectified flow, without cost-aware gradient projection, can stabilize at a non-optimal deterministic Gaussian coupling; practitioners should not expect reflow alone to recover optimal transport in general.","Under projection stability, the excess transport cost decays exponentially, so the number of iterations needed for a given accuracy can be bounded explicitly.","Combined with the paper's score-based marginal estimators, c-rectified flow yields a plug-in optimal transport map estimator with squared L2 risk at the minimax rate for d≥3 and nearly parametric in d=1,2.","The general-cost analogue shows the same convergence and contraction for strongly convex displacement costs, so the results are not tied to squared Euclidean distance."],"fun_headline_variants":["Cost-aware rectified flow provably reaches optimal transport","Iterated c-rectified flow: guaranteed optimal coupling","Cost projection fixes rectified flow's OT convergence","c-rectified flow converges to optimal transport map","Rectified flow upgrade: cost-aware projection ensures OT"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the assumption that, at every iteration, the smoothness of the rectified potentials and the growth of their gradients stay uniformly controlled, with the gradient integrability along the interpolated paths uniform over all steps; if those bounds fail, the proof cannot extract a limiting potential or show the limiting coupling is optimal.","fun_headline_variants_meta":{"raw":{"variants":["Cost-aware rectified flow provably reaches optimal transport","Iterated c-rectified flow: guaranteed optimal coupling","Cost projection fixes rectified flow's OT convergence","c-rectified flow converges to optimal transport map","Rectified flow upgrade: cost-aware projection ensures OT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000533,"raw_usage":{"total_tokens":2401,"prompt_tokens":745,"completion_tokens":1656,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":489,"tokens_out":1656,"duration_ms":410969,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T06:20:29.778942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the population c-rectified iteration on centered Gaussian marginals N(0, Σ1), N(0, Σ2) with Σ1 and Σ2 chosen so that Σ1Σ2 ≠ Σ2Σ1 (say, Σ1=diag(4,1), Σ2 a rotation of it). The paper predicts the excess cost E||Z1^k−Z0^k||²−W₂²(P,Q) decays exponentially. If it stays bounded away from zero after many iterations, Theorem 1's convergence claim fails.","supporting_citations":[],"review_version":1}