{"id":"756b4454-7c12-4a42-91b7-e258b98f4719","arxiv_id":"1908.10962","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A principled minimax training procedure over input convex neural networks learns the optimal quadratic-cost transport map as the gradient of a convex potential.","lead":"This paper introduces a minimax optimization framework that learns the optimal transport map between two distributions by training two input convex neural networks. The method aims to find the unique 2-Wasserstein optimal map, which is useful for deep generative modeling and other transport tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3's saddle-point attainment is not proven, and the practical ICNN optimization does not satisfy the theorem's convexity or strong-convexity hypotheses.","rationale":"The reader's weakest_assumption identifies precisely the gap between the infinite-dimensional regularity assumptions and the practical ICNN optimization. My reading of the paper supports this: Theorem 3.3 is stated with a proof sketch that only proves the inequality Vf(g) <= EQ[f*(Y)] for all g and equality at g = f*, but the attainment of the infimum over g and the supremum over f is asserted without proof. The appendix does not fill this gap; it only proves differentiability a.e. for the pointwise inequality. The claim that the learned map is 'optimal independent of initialization' therefore depends on an unproven saddle-point attainment result. The practical algorithm also relaxes the convexity of g (Remark 3.4 and the penalty term (8)) and does not enforce strong convexity that Theorem 3.6 requires. These are correct and concrete concerns that do not constitute an attack on the authors' integrity. The paper's contributions are still plausible and the experiments are suggestive, but the theory-practice gap is real, so the verdict should remain CONDITIONAL rather than full ACCEPT. I do not find a more load-bearing concern than this one. The honest recommendation is to require either a rigorous attainment proof in a revised version or a numerical verification of the saddle point structure. My concrete test is designed to settle whether the practical algorithm actually achieves the theoretical saddle point.","tokens_in":16334,"tokens_out":1652,"duration_ms":14442,"concrete_test":"Implement Algorithm 1 exactly as specified, but with the convexity regularization R(theta_g) removed (lambda = 0) and all weights of g strictly non-negative, for the checkerboard example in Section 4.1. Compare the learned transport map to the analytic or Barycentric-OT reference over 10 random initializations of both f and g. Also compute the gradient of the objective with respect to f at the converged solution to check whether the inner minimization over g has actually reached a global optimum for the current f. If the map varies across initializations or the inner optimization is not at its global minimum, the saddle-point attainment assumption underlying the 'initialization independence' claim is not satisfied by the practical algorithm.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that solving the minimax objective (5) recovers the exact optimal transport map rests on Theorem 3.3 asserting existence of an optimal pair (f0, g0) in infinite-dimensional convex function spaces. The proof sketch shows supg Vf(g) = EQ[f*(Y)] via a pointwise inequality and differentiability a.e., but it does not establish that the supremum over g in CVX(Q) is attained by some g0 in the function space, nor that the outer supremum over f is attained. Without attainment, the saddle point may be only an inf-sup value with no representable (f0, g0), so the claim that Algorithm 1's output is the optimal transport map is not justified. The appendix does not fill this gap; it only proves the a.e. differentiability used in the pointwise inequality. Additionally, the practical algorithm replaces CVX(P) and CVX(Q) by ICNN parameterizations and relaxes convexity of g with a soft penalty (8), while Theorem 3.3 requires exact convexity for the identity g = f*. The stability bound Theorem 3.6 assumes f is alpha-strongly convex, but the ICNN parameterization does not enforce strong convexity, only convexity, and the experiments do not verify the strong convexity constant alpha. This is not merely an outside-consensus disagreement; it is a gap between the theorem's hypotheses and the actual optimization problem being solved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new approach for learning optimal transport maps under the squared-Euclidean (W2) cost from samples. The authors introduce a minimax formulation over pairs of convex functions, derived from the Kantorovich dual via the convexification trick, and show in Theorem 3.3 that the solution of this minimax problem yields the optimal transport map as the gradient of one of the convex potentials. They parameterize the convex functions with input convex neural networks (ICNNs) and propose Algorithm 1, a stochastic minimax training procedure with a soft-convexity penalty for the second network. Numerical experiments on 2D synthetic distributions, high-dimensional Gaussians, Gaussian mixtures, and MNIST latents demonstrate the method's ability to recover optimal maps, including discontinuous maps, and its relative robustness to initialization compared to WGAN-type baselines.","tokens_in":16712,"tokens_out":12762,"duration_ms":109014,"significance":"The paper makes a useful methodological contribution by providing a clear, theory-grounded objective for learning W2-optimal transport maps, and demonstrates empirically that representing the map as the gradient of an ICNN allows sharp separation of disconnected target supports—an advantage over continuous network maps. The theoretical results, if fully established, would provide consistency and a stability estimate linking optimization gaps to map error. The experimental appendix is detailed and supports reproducibility. However, as presented, the proof of Theorem 3.3 omits the existence argument for the outer maximizer, and the stability theorem's proof contains an error in the constant; additionally, the bridge from the infinite-dimensional theory to the ICNN/penalized algorithm is not quantified. These gaps are fixable but require careful revision.","major_comments":[{"comment":"The proof sketch and Appendix A establish that for a fixed f, the inner supremum over g in CVX(Q) is attained by g=f^*, but they do not establish the existence of an f0 attaining the outer supremum in CVX(P) with f0^∗ ∈ L1(Q). The statement that '(f0,g0) achieves the infimum and supremum' is invoked without proof; this is a standard existence result in optimal transport, but the paper should either supply the argument or cite the precise theorem (e.g., Villani 2003, Theorem 2.9 or 2.12) in the proof. Without this, the saddle-point characterization and the claim that ∇g0 is the optimal transport map are not fully justified.","section":"Section 3 (Theorem 3.3) and Appendix A"},{"comment":"The proof uses the triangle inequality on the L2-norm after deriving ||∇g−∇f^*||^2_{L2(Q)} ≤ (2/α)ϵ1 and ||∇f^*−∇g0||^2_{L2(Q)} ≤ (2/α)ϵ2. For any two square-integrable functions u and v, ||u+v||^2 ≤ 2||u||^2 + 2||v||^2, so the triangle inequality yields ||∇g−∇g0||^2_{L2(Q)} ≤ (4/α)(ϵ1+ϵ2), not (2/α)(ϵ1+ϵ2) as stated. The constant in Theorem 3.6 thus appears to be incorrect under the proof as written; the authors should either correct the theorem to a constant of 4/α or find an alternative argument to retain the stated bound.","section":"Section 3.2 (Theorem 3.6) and Appendix B"},{"comment":"The theoretical guarantees in Theorems 3.3 and 3.6 are stated for exact optimization over the infinite-dimensional spaces CVX(P) and CVX(Q). The practical algorithm, however, optimizes over ICNN parameterizations and relaxes the convexity constraint on g by replacing it with the soft penalty R(θg) in Eq. (8). The paper does not provide a consistency result or an approximation bound that accounts for these parametric and penalization errors. Consequently, the abstract's claim that the learned mapping is 'optimal independent of how we initialize the neural networks' is not directly supported by the proven theorems; it is an empirical observation that appears consistent with the experiments but requires either a gap analysis or a more careful statement of the claim.","section":"Section 3.1 and Algorithm 1"},{"comment":"The stability bound in Theorem 3.6 assumes f is α-strongly convex, but the ICNN parameterization in Section 3.1 only enforces convexity (with nonnegative weights), and no strong-convexity regularization or verification is described in the experiments of Section 4. As stated, the theorem is vacuous if no learned potential satisfies the assumption; the authors should either discuss how α can be enforced or estimated for ICNNs or explicitly label the result as conditional on an unverified hypothesis.","section":"Section 3.2 (Theorem 3.6)"}],"minor_comments":[{"comment":"The expression 'inf_{f∈CVX(P)} E_P[f(X)] + E_Q[f^*(Y)]' should be parenthesized to make clear that the infimum applies to the sum of the two expectations, not only to the first term.","section":"Section 3, Eq. (4)"},{"comment":"The caption states that the source and target are 728-dimensional Gaussians, whereas the text in Section 4.4 specifies d=784; this is a typo.","section":"Table 1 caption"},{"comment":"The reported metric ||µT(Q)-µ||2 in the first row of Table 1 is scale-dependent; the relative error in the second row is more informative, and the first row could be complemented with a normalized version or a discussion of this dependence.","section":"Section 4.4, Table 1"},{"comment":"The step labeled 'Projection' is a thresholding of weights to nonnegative values; using a more descriptive name such as 'Weight clipping' would avoid confusion with projection onto a feasibility set.","section":"Algorithm 1"},{"comment":"In the proof of Theorem 3.3, the differentiability of f^* Q-a.e. is asserted to follow by arguments similar to those for g, but the argument is not spelled out; adding a sentence that f^* is convex and finite Q-a.e. (by f^*∈L1(Q)) would close this gap.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the main contribution is interesting. The identified issues are proof-detail and presentation problems rather than fundamental flaws; I expect a revision that addresses them would be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Makkuva et al. (arXiv:1908.10962). The paper's real contribution is a new minimax formulation for the 2-Wasserstein OT map: instead of computing the convex conjugate f* directly, you train a second convex network g to serve as f*, and the transport map is ∇g. That's new relative to Taghvaei & Jalali, which uses an ICNN for the potential but needs inner conjugate solves. The derivation sits on standard OT results, and the paper credits its predecessors clearly.\n\nWhat the paper does well: it turns a known but impractical semi-dual into something trainable, with a clean minimax objective. The experiments are suggestive: checkerboard and eight-Gaussian examples show the expected qualitative behavior, and the d=784 Gaussian-to-Gaussian example gives a quantitative check against a known map. The point about discontinuous maps is interesting; gradients of ReLU-type convex networks can have jumps, which standard continuous generators cannot represent.\n\nThe soft spots are real. Theorem 3.3's proof establishes that for a fixed f, the infimum over convex g is attained at g=f*, but the existence of a saddle point (f0,g0) in infinite-dimensional function space is not proven. The appendix covers only the fixed-f step; attainment of the outer supremum is inherited from the semi-dual but not shown. For an algorithmic paper that's probably an acceptable gap, but the headline claim 'solving (5) yields the exact OT map' is population-level and the attainment question is part of it. Theorem 3.6 assumes f is α-strongly convex; the ICNN parameterization does not enforce that, and the experiments never verify it. The practical algorithm also relaxes convexity of g with a soft penalty, so the exact identity g=f* is not enforced. These gaps mean the theory as stated doesn't cover the actual optimization problem, though the empirical method can still be sound.\n\nThe 'initialization independence' claim is a minor overclaim. Uniqueness of the OT map when Q has a density is solid, but finite-sample, finite-capacity training can still land in different local optima. The paper shows two trials for W1-LP and W2GAN and only one or a few for its own method; that's not a strong demonstration. No code is released either, which hurts reproducibility.\n\nOverall, this is a solid paper with a genuinely useful algorithmic idea and honest related-work discussion. The theory has gaps, but they are traceable and not deceptive. I'd send it to referees, and I'd cite it if I were writing about neural OT.","headline":"A genuinely new minimax formulation for learning the 2-Wasserstein map that is worth engaging, but the theory doesn't cover the practical algorithm and the initialization-independence claim is demonstrated only qualitatively.","tokens_in":17158,"tokens_out":3682,"would_cite":true,"duration_ms":35545,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","68T07","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A minimax game over convex functions recovers the exact optimal transport map.","keywords":["optimal transport","Wasserstein-2 distance","input convex neural networks","minimax optimization","convex conjugate","deep generative models","Brenier map","discontinuous transport maps"],"falsifier":"Take $Q = \\mathcal{N}(0,I_d)$ and $P = \\mathcal{N}(\\mu,I_d)$, where the exact optimal map is $T^*(y) = y + \\mu$, train the proposed algorithm from many different initializations, and compare each learned $\\nabla g$ with $T^*$ in $L^2(Q)$ while tracking the converged minimax value against the known squared Wasserstein distance $\\|\\mu\\|^2/2$; if the maps vary with initialization or the objective converges to a different value as network width grows, the saddle-point-to-map claim is disproved.","tokens_in":16100,"feed_emoji":"🎯","tokens_out":7234,"duration_ms":75090,"temperature":0.7,"pith_summary":"This paper claims that the optimal transport map between two distributions under quadratic cost can be learned by solving a minimax problem over pairs of convex functions, with no regularization of the Kantorovich constraint. It derives an identity in which the squared 2-Wasserstein distance equals a supremum over one convex function of an infimum over another, and proves that at the saddle point the gradient of the second function is the exact transport map. If this is right, it gives a principled training rule for deep generative models: the generator is the gradient of a convex network, converges to the same map regardless of initialization, and can represent discontinuous mappings that match targets with separated supports. Numerical experiments on synthetic and high-dimensional data support the claim.","feed_headline":"Optimal transport map falls out of a convex minimax game","feed_subtitle":"Two convex networks trained against each other give the cheapest mass movement, even when the target is disconnected.","key_machinery":"The carrying object is the functional $V_{P,Q}(f,g) = -\\mathbb{E}_P[f(X)] - \\mathbb{E}_Q[\\langle Y, \\nabla g(Y)\\rangle - f(\\nabla g(Y))]$ over pairs of convex functions, together with the identity $\\langle y, \\nabla g(y)\\rangle - f(\\nabla g(y)) \\le f^*(y)$, equality holding when $g = f^*$. This converts the hard Kantorovich feasibility constraint into an unconstrained saddle problem: the inner minimization selects $g$ as the convex conjugate of $f$, and the outer maximization selects the Brenier potential whose gradient is the optimal map. Input convex neural networks, networks whose output is convex in the input, supply the trainable parameterization of the convex functions, and a universality result for these networks is used to justify replacing the full convex class by the neural class.","core_discovery":"The central discovery is Theorem 3.3: for distributions $P$ and $Q$ with finite second moments, with $Q$ having a density, the squared 2-Wasserstein distance equals $\\sup_{f \\in \\mathrm{CVX}(P), f^* \\in L^1(Q)} \\inf_{g \\in \\mathrm{CVX}(Q)} V_{P,Q}(f,g) + C_{P,Q}$, where $V_{P,Q}(f,g) = -\\mathbb{E}_P[f(X)] - \\mathbb{E}_Q[\\langle Y, \\nabla g(Y)\\rangle - f(\\nabla g(Y))]$, and there is an optimal pair $(f_0,g_0)$ with $\\nabla g_0$ equal to the optimal transport map from $Q$ to $P$. The proof rests on the inequality $\\langle y, \\nabla g(y)\\rangle - f(\\nabla g(y)) \\le f^*(y)$, with equality at $g = f^*$, together with the Knott-Smith criterion and Brenier's theorem. The paper further claims that the proposed algorithm, which parameterizes $f$ and $g$ by input convex neural networks and alternates gradient updates, learns this optimal map independent of initialization, and that using gradients of convex networks permits discontinuous transport maps with sharp boundaries.","pith_inferences":["The stability bound requires $f$ to be $\\alpha$-strongly convex, a property the ICNN parameterization does not enforce; adding a small quadratic regularizer to $f$ would make the bound applicable, a modification the paper does not explore.","The minimax identity is representation-agnostic: any sufficiently expressive convex function class could replace ICNNs, and the same consistency would hold up to an approximation error that could be bounded by combining the universality result with Theorem 3.6.","The finite-sample behavior is left open; deriving sample complexity for the empirical minimax objective is a natural extension, and the paper's stability result only partially addresses it.","The dependence on $Q$ having a density suggests the method will need modification when the source distribution has atoms, since the differentiability-almost-everywhere argument in the proof would break."],"forward_implications":["An exact solution of the minimax gives the exact optimal transport map, not merely an arbitrary feasible transport.","Deep generative models trained this way inherit uniqueness: the same generator emerges from different initializations, so the model is not one of many arbitrary push-forwards.","Because the gradient of a convex ReLU-based network can have kinks, the learned map can jump across gaps, producing sharp boundaries for disconnected target supports.","Near-optimal solutions satisfy an explicit $L^2$ error bound proportional to the objective gaps divided by the strong-convexity parameter $\\alpha$, linking optimization error to transport error.","The framework turns optimal transport theory into a practical training signal for generative modeling, with robustness and discontinuity as concrete advantages over adversarial training."],"supporting_citations":[{"why":"Supplies the Knott-Smith criterion and Brenier's theorem that connect dual minimizers to the optimal transport map, the foundation of the formulation.","marker":"(Villani, 2003)"},{"why":"Introduces input convex neural networks, the parameterization used for the two convex functions $f$ and $g$.","marker":"(Amos et al., 2016)"},{"why":"Proposes the semi-dual ICNN approach to approximate the 2-Wasserstein distance, which the present minimax formulation extends and makes scalable.","marker":"(Taghvaei & Jalali, 2019)"},{"why":"Shows that input convex neural networks are dense in the class of convex functions, justifying the use of ICNNs as the approximating function class.","marker":"(Chen et al., 2018)"},{"why":"Provides the stability result for optimal transport maps used to prove Theorem 3.6, bounding the error of near-optimal maps.","marker":"(Hütter & Rigollet, 2019)"}],"fun_headline_variants":["Minimax convex nets learn optimal transport maps","Optimal transport without init bias via convex minimax","Discontinuous optimal transport from convex network gradients","Cheapest mass movement via convex neural game","Convex minimax yields optimal transport map"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the assumption that the infinite-dimensional minimax over all convex functions is attained by some pair $(f_0,g_0)$ and that $Q$ has a density; in the practical version, the stability bound additionally assumes $f$ is $\\alpha$-strongly convex, which the ICNN parameterization does not guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Minimax convex nets learn optimal transport maps","Optimal transport without init bias via convex minimax","Discontinuous optimal transport from convex network gradients","Cheapest mass movement via convex neural game","Convex minimax yields optimal transport map"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000577,"raw_usage":{"total_tokens":2717,"prompt_tokens":933,"completion_tokens":1784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1715}},"tokens_in":549,"tokens_out":1784,"duration_ms":13399,"temperature":1.0,"reasoning_tokens":1715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:27:37.372951+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $Q = \\mathcal{N}(0,I_d)$ and $P = \\mathcal{N}(\\mu,I_d)$, where the exact optimal map is $T^*(y) = y + \\mu$, train the proposed algorithm from many different initializations, and compare each learned $\\nabla g$ with $T^*$ in $L^2(Q)$ while tracking the converged minimax value against the known squared Wasserstein distance $\\|\\mu\\|^2/2$; if the maps vary with initialization or the objective converges to a different value as network width grows, the saddle-point-to-map claim is disproved.","supporting_citations":[],"review_version":1}