{"id":"a3a8c0ad-7156-4f3a-8d56-b669e3cf647d","arxiv_id":"2501.10162","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A convex physics-informed neural network with a Hausdorff-distance boundary loss approximates optimal transport maps in two and three dimensions.","lead":"This paper introduces a neural-network solver that computes the cheapest way to move material from one region to another by solving the Monge-Ampere equation with a convex neural network. The method uses a Hausdorff-distance-based loss to enforce the transport boundary condition, and tests it on 2D and 3D examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never states how nonnegativity of ICNN weights is maintained during Adam/L-BFGS; squaring only at initialization does not keep the trained network convex, so the central OT-map claim is unverified as written.","rationale":"The central claim is that an ICNN potential's gradient approximates the optimal transport map. That claim requires the trained network to be convex. The only stated measure toward convexity is squaring the initial weights; no subsequent projection or reparameterization is described. Since Adam and L-BFGS are unconstrained optimizers, weights can leave the nonnegative orthant, so the 'input convex' property is not guaranteed after training. This is more specific and more immediately falsifiable than the general nonconvex-loss concern: it can be checked by inspecting the final weights. If the authors in fact use a reparameterization, they should state it; if not, the numerical results do not demonstrate an ICNN-based solution. The reader's concern about loss-landscape reliability is real but secondary; I therefore partially agree with the reader. I would keep the conditional verdict pending code and implementation clarification.","tokens_in":10953,"tokens_out":6013,"duration_ms":64907,"concrete_test":"Inspect the released code or trained weights for Section 4.2: check the final W(l) for negative entries after L-BFGS. If negatives exist, re-train with an explicit positivity-preserving parameterization (e.g., W = softplus(A)) and compare the reported L2 error and boundary loss; if the constrained run no longer reaches ~1e-3, the central claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines the ICNN via nonnegative weight matrices W(l), and Section 4.1 says the weights are initialized by squaring them to guarantee positivity. For the trained network to remain an ICNN, every optimizer step must preserve W(l) >= 0. The paper describes only unconstrained Adam followed by L-BFGS (Section 4.1, Table 1); it never states that weights are projected, clipped, or reparametrized (e.g., W = A^2 or W = softplus(A)) during training. With ordinary gradient updates, squared initialization does not keep weights nonnegative. If any W(l) entry becomes negative, u_NN is not convex, the gradient is not necessarily a monotone OT map, and the central claim that convexity is enforced by the architecture is unsupported. This is a concrete risk independent of the nonconvex-loss concern: even a globally converged loss could be attained by a non-convex network. The authors also report that clipping weights gives unsatisfactory results, which makes the actual constraint mechanism especially important to specify.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed neural network (PINN) method for the quadratic-cost optimal transport problem, writing the optimal map as the gradient of a convex potential u satisfying the generalized Monge-Ampère equation det(D^2u) = f/g(∇u) with the transport boundary condition ∇u(∂X) = ∂Y. The authors represent u by an input-convex neural network (ICNN), impose the PDE through a strong-form residual, and enforce the boundary condition through a discretized Hausdorff-distance loss. Numerical experiments cover disk-to-ellipse, ellipse-to-rotated-ellipse, Gaussian-to-uniform, Gaussian-to-Gaussian, bimodal Gaussian-to-uniform, and a 3D cube example; the Adam and L-BFGS optimizers are compared, and sensitivity to epochs, collocation points, and boundary-point ratio is studied. The claimed result is an accurate convex neural potential whose gradient approximates the optimal transport map, with reported L2 errors around 1e-3 in the exact-solution cases.","tokens_in":11236,"tokens_out":10966,"duration_ms":111364,"significance":"The combination of ICNNs with a PDE residual and a Hausdorff-type boundary loss is natural and potentially useful, and the paper includes honest sensitivity studies and an explicit exact-solution test in Section 4.3 that allows pointwise error evaluation. If the implementation issues identified below are resolved, the method could be a competitive tool for moderate-dimensional optimal transport problems. At present, however, the central claim that the architecture enforces convexity is not substantiated, because the positivity of the ICNN weights is only described at initialization. In addition, the main validation in Section 4.2 targets an incorrect exact solution, and the Gaussian experiments lack quantitative error reporting. The rotated-ellipse test gives credible pointwise errors around 1e-3, and the authors are transparent about run-to-run variability, but the missing details on convexity preservation and the Section 4.2 inconsistency are load-bearing and need to be fixed.","major_comments":[{"comment":"The ICNN in Eq. (3.1) is convex only when all W(l) entries are nonnegative. Section 4.1 says the weights are initialized by element-wise squaring, but the subsequent training uses unconstrained Adam followed by L-BFGS (Table 1), and no projection, clipping, or reparametrization is described. Ordinary gradient steps can make entries of W(l) negative, in which case the trained network is not an ICNN and ∇u_NN is not guaranteed to be the gradient of a convex potential. Please state explicitly how W(l) ≥ 0 is maintained throughout training, or show empirically that it is preserved; without this, the assertion that convexity is enforced by construction is unsupported.","section":"Section 3.2 and Section 4.1"},{"comment":"The exact map stated after the definition of the ellipse, ∇uex(x) = (2x1+7, (1/2)x2), is inconsistent with Y = { (y1-3.5)^2/2^2 + y2^2/0.5^2 < 1 }. For x1 ∈ [-1,1], 2x1+7 ranges over [5,9], i.e. an ellipse centered at 7, not at 3.5. The correct affine map for the stated Y is (2x1+3.5, 0.5x2). As written, Table 2 and Figures 2-4 report errors against a solution of a different problem; please correct either the definition of Y or the exact map and recompute the reported errors.","section":"Section 4.2"},{"comment":"The PDE residual is defined via the ratio f(x)/g(∇u(x)), with g(y) = (1/π)χ_Y(y) or g = χ_{[0,1]^2}. During training, ∇u_NN(x_i) can leave the support of g, making the denominator zero and the loss undefined at those collocation points. The manuscript does not describe any regularization, clamping, or reformulation that makes the residual finite in this situation. This is a genuine issue for compactly supported target densities and should be addressed, for example by using a smoothed density or by explicitly treating points outside the target support.","section":"Sections 3.3 and 4.2-4.7"},{"comment":"The Gaussian and bimodal experiments are only evaluated visually, through histograms and pointwise error plots; no error norms, convergence rates, or quantitative comparisons are given. Section 4.4 states that the error 'is still of a similar order', but no numerical value supports that claim. Since these tests are the main evidence for nonuniform densities and for the 3D extension, at least a quantitative error measure (e.g., against a reference solution or a histogram-distance metric) should be reported.","section":"Sections 4.4-4.7"}],"minor_comments":[{"comment":"The captions label both bottom panels as 'Bottom left' and 'Top right' for the second component and its error; the second error panel should be 'Bottom right'.","section":"Figure 4 and Figure 12 captions"},{"comment":"Calling the first term of EOT a control of 'injectivity' is not accurate; the term ∑ dist(∇u(x_i), {y_j})^2 only enforces that each image lies near the target boundary, not that the map is injective. Please rephrase or justify.","section":"Section 3.3, Eq. (3.4)"},{"comment":"The explicit optimal map for two centered ellipses, T(x)=M_Y R_θ M_X^{-1}x with the given formula for θ, is asserted without derivation or reference; a citation or a short argument would help the reader trust this exact solution.","section":"Section 4.3"},{"comment":"The conclusion that the model 'is not particularly affected by the random sampling' is difficult to reconcile with Table 2 (L-BFGS L2 error 0.00267 ± 0.00376) and with the large sample variation visible in Figure 6; please qualify this statement.","section":"Section 5"},{"comment":"The statement that the code is 'available on request' is not a reproducible-software statement; a persistent repository link would strengthen the paper.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the missing description of how ICNN weight nonnegativity is preserved during training; the paper's central claim depends on it. The Section 4.2 exact-solution inconsistency is likely a typo but must be fixed. The method is plausible and worth revising rather than rejecting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read of 2501.10162.\n\nThe paper has a genuinely useful idea: use an Input Convex Neural Network to solve the generalized Monge–Ampère equation with transport boundary conditions for quadratic-cost OT, and encode those boundary conditions as a discrete Hausdorff-distance loss. That specific combination does not appear in the cited Dirichlet-Monge–Ampère PINN work or the PICANN density-estimation work. The rotated-ellipse experiment, which has an explicit solution, shows pointwise errors around 1e-3 and the authors do a decent sensitivity analysis over random samples. The acknowledgement in Section 5 that the PINN error is 'rather difficult to control' is honest.\n\nBut there is a load-bearing gap. The ICNN architecture requires all W(l) to be nonnegative throughout training. The paper only says the weights are initialized by squaring them. It never says what happens during Adam/L-BFGS updates. Unconstrained gradient steps can push entries negative, and once that happens the network is no longer convex, so the gradient map is not necessarily the OT map. Clipping was tried and deemed unsatisfactory, which makes the missing mechanism even more pressing. I'd want to see either a genuine reparametrization (W = A^2 or softplus), a projection step, or at least a post-training check that the trained weights are nonnegative and the Hessian is PSD on the collocation points. Without that, the central claim of the paper is unverified.\n\nThere are smaller problems. The exact solution in Section 4.2 uses an offset of +7 in the first gradient component, but the ellipse is centered at x1 = 3.5, so for x1 in [−1,1] the gradient maps to [5,9] instead of [1.5,5.5]. That looks like a typo, but it undermines the reported validation numbers for that case. The Gaussian examples (4.4–4.7) provide only histograms and no quantitative error measures; the authors claim the error is 'of a similar order' without reporting values. And there are no baseline comparisons to standard OT solvers. The lack of public code is a minor annoyance; 'available on request' is fine but not ideal.\n\nOverall the idea is worth pursuing and the paper deserves a serious referee, but as written it needs a substantive revision, not just copy-editing. The authors must clarify how convexity is maintained during optimization and correct the exact-solution mistake. If they can do that, the method would be a solid contribution to the PINN-OT literature.","headline":"A promising PINN-OT idea whose central convexity claim is unsupported as written, because the paper never says how ICNN weights stay nonnegative during training.","tokens_in":11693,"tokens_out":3153,"would_cite":false,"duration_ms":29507,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35J96","49Q22","65N99","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Input convex neural networks trained on the Monge–Ampère equation can approximate the optimal transport map as the gradient of a convex potential, with L2 errors around 1e-3 on test problems with known solutions.","keywords":["optimal transport","Monge-Ampère equation","input convex neural networks","physics-informed neural networks","transport boundary conditions","Hausdorff distance","sensitivity analysis","logistics"],"falsifier":"Take a two-dimensional problem with a known nonlinear optimal map, run the protocol with a dense boundary discretization, and measure the one-sided Hausdorff distance from $\\nabla u_{NN}(\\partial X)$ to $\\partial Y$; the paper's claim implies this distance decreases toward zero as the number of collocation points grows, so a persistent positive floor while the interior residual is small would falsify the boundary-condition enforcement.","tokens_in":10775,"feed_emoji":"🚚","tokens_out":7961,"duration_ms":74458,"temperature":0.7,"pith_summary":"The paper proposes a physics-informed neural network method for the quadratic-cost optimal transport problem in two and three dimensions. It claims that an input convex neural network can approximate the convex potential whose gradient is the optimal transport map, by minimizing a loss that combines a generalized Monge–Ampère equation residual $\\det(D^2u)=f/g(\\nabla u)$ with a discrete Hausdorff-distance boundary term enforcing $\\nabla u(\\partial X)=\\partial Y$. On test problems with known solutions, the gradient of the trained network matches the exact transport map to errors around $10^{-3}$ in $L^2$ norm. If this holds, the method offers a mesh-free, architecture-enforced-convexity route to optimal transport maps, with the transport boundary condition handled without prescribing a boundary potential.","feed_headline":"Convex neural nets approximate optimal transport maps","feed_subtitle":"Convex neural nets solve the Monge-Ampère equation, with gradients accurate to about 1e-3.","key_machinery":"The central object is the input convex neural network (ICNN), a feedforward network whose hidden-to-hidden weight matrices are constrained to have nonnegative entries and whose activation is convex and nondecreasing, so the network output is convex in its input. The loss is the sum of the squared Monge–Ampère residual at interior collocation points and a boundary term $E_{OT}$ defined by the two-sided discrete Hausdorff distance between $\\nabla u_{NN}(\\partial X)$ and $\\partial Y$. The training protocol uses an identity-map initialization for $\\nabla u$, then Adam followed by L-BFGS, with softplus activations and weight-squaring to maintain positivity.","core_discovery":"The central claim is that the optimal transport map for quadratic cost between two absolutely continuous probability measures can be approximated by training an input convex neural network to solve the generalized Monge–Ampère equation $\\det(D^2u)=f/g(\\nabla u)$ with transport boundary condition $\\nabla u(\\partial X)=\\partial Y$. Because the ICNN architecture guarantees convexity of the approximate potential $u_{NN}$ by construction, the trained gradient $\\nabla u_{NN}$ is a legitimate candidate transport map. The paper demonstrates numerically on disk-to-ellipse, ellipse-to-rotated-ellipse, and Gaussian-to-uniform problems that this gradient converges to the exact map, and it enforces the transport boundary condition through a differentiable discrete Hausdorff-distance loss on the images of boundary collocation points, controlling both injectivity and surjectivity of the map.","pith_inferences":["Editorial inference: the identity-map initialization plus ICNN convexity could serve as a warm start for higher-dimensional or obstacle-constrained transport problems, where the authors expect training to scale poorly.","Editorial inference: because the Hausdorff boundary loss is defined on point sets, it could be extended to partial or semi-discrete boundary matchings without reformulating the PDE, a direction the paper does not pursue.","Editorial inference: a natural stress test is to compare the reported ~1e-3 gradient accuracy against a finite-difference or finite-element optimal transport solver on the non-analytic examples; the paper does not include such a comparison."],"forward_implications":["The trained gradient $\\nabla u_{NN}$ provides a direct approximation of the optimal transport map for quadratic cost, bypassing discrete linear-programming or iterative schemes.","The method handles non-constant source and target densities: Gaussian-to-Gaussian and bimodal-Gaussian-to-uniform examples show transported samples matching the target distribution, so it applies beyond uniform densities.","The Hausdorff-based boundary loss is generic: it requires only point clouds on $\\partial X$ and $\\partial Y$, so it can enforce transport boundary conditions for arbitrary target shapes, and simplifies to a Neumann condition for square-to-square maps.","The sensitivity experiments indicate that increasing the number of collocation points or training epochs decreases the gradient error, and that setting the number of boundary points equal to the number of collocation points is the best configuration."],"supporting_citations":[{"why":"Introduces input convex neural networks, the architecture whose nonnegative weights and convex nondecreasing activations guarantee convexity of the approximate solution.","marker":"[2]"},{"why":"Supplies the classical theorem that the quadratic-cost optimal transport map is the gradient of a convex potential satisfying the generalized Monge–Ampère equation.","marker":"[14]"},{"why":"Applies ICNNs to the Dirichlet Monge–Ampère problem and supplies the softplus activation, weight-squaring initialization, and evidence on scaling with dimension.","marker":"[21]"},{"why":"Introduces the physics-informed neural network residual minimization framework used for the PDE loss.","marker":"[26]"},{"why":"Documents the Adam-then-L-BFGS training protocol used to obtain stable convergence in the numerical experiments.","marker":"[27]"},{"why":"Uses convex neural networks for optimal transport density estimation and motivates the identity-map initialization for the gradient of the solution.","marker":"[29]"}],"fun_headline_variants":["Convex neural nets compute optimal transport maps","Convex PINNs crack the Monge-Ampère equation","ICNNs enforce convexity for accurate transport maps","Monge-Ampère solved by convex neural networks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that minimizing this nonconvex PINN loss from the identity-gradient initialization reliably lands on a network whose gradient approximates the exact optimal map, despite the authors' own observation that ICNNs are notoriously difficult to train and that run-to-run variability makes the error hard to control.","fun_headline_variants_meta":{"raw":{"variants":["Convex neural nets compute optimal transport maps","Convex PINNs crack the Monge-Ampère equation","ICNNs enforce convexity for accurate transport maps","Monge-Ampère solved by convex neural networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000342,"raw_usage":{"total_tokens":1817,"prompt_tokens":814,"completion_tokens":1003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":940}},"tokens_in":430,"tokens_out":1003,"duration_ms":10335,"temperature":1.0,"reasoning_tokens":940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:22:26.155484+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-dimensional problem with a known nonlinear optimal map, run the protocol with a dense boundary discretization, and measure the one-sided Hausdorff distance from $\\nabla u_{NN}(\\partial X)$ to $\\partial Y$; the paper's claim implies this distance decreases toward zero as the number of collocation points grows, so a persistent positive floor while the interior residual is small would falsify the boundary-condition enforcement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces input convex neural networks, the architecture whose nonnegative weights and convex nondecreasing activations guarantee convexity of the approximate solution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the classical theorem that the quadratic-cost optimal transport map is the gradient of a convex potential satisfying the generalized Monge–Ampère equation."},{"cited_title":"Nystr¨ om and M","cited_arxiv_id":null,"evidence_quote":"Applies ICNNs to the Dirichlet Monge–Ampère problem and supplies the softplus activation, weight-squaring initialization, and evidence on scaling with dimension."},{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Introduces the physics-informed neural network residual minimization framework used for the PDE loss."},{"cited_title":"Rathore, W","cited_arxiv_id":null,"evidence_quote":"Documents the Adam-then-L-BFGS training protocol used to obtain stable convergence in the numerical experiments."},{"cited_title":"Singh, M","cited_arxiv_id":null,"evidence_quote":"Uses convex neural networks for optimal transport density estimation and motivates the identity-map initialization for the gradient of the solution."}],"review_version":1}