acceptodds
Under review as a conference paper at ICLR 2027

Stability of the Monge Map in Semi-Dual Optimal Transport

Abstract

This paper shows that the semi-dual formulation of the optimal transport problem has a degenerate saddle-point structure, and that its numerical solution is equivalent to solving a constrained optimization problem. We derive necessary and sufficient conditions for the convergence of Monge maps without requiring optimality of the dual potential. This analysis helps explain why, in practice, numerical algorithms often require more iterations to update the transport map than the potential.

Primary area:
optimization
Keywords:
neural optimal transport, Monge map, degenerate saddle point, constrained optimization, semi-dual optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8DSaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript correctly identifies a structural degeneracy in the semi-dual formulation and derives a corresponding stability estimate under strong regularity assumptions. However, the novelty is limited, the assumptions exclude most realistic data regimes, the experiments remain confined to synthetic settings where the theory is most favorable, and the critique of prior work is overstated. The overall contribution is incremental and does not meet the bar for acceptance at ICLR.

Critical strengths
  1. The paper cleanly isolates the flatness of the semi-dual objective at an optimal map and supplies an explicit L2L^2 stability estimate that does not require the dual potential itself to be optimal. This is a useful clarification for the neural-OT literature.

  2. The experimental illustrations (especially the timescale sweeps) make the practical consequence of the degeneracy concrete: one can obtain a good transport map while the potential remains far from optimal. This matches common empirical practice and is presented transparently.

Critical weaknesses
  1. Limited novelty and overstated contributions.The core observation—that the semi-dual objective F(ψ,t)F(\psi, t) becomes independent of ψ\psi once t=T⋆t = T^\star (i.e., the dual term vanishes when the marginal constraint is satisfied)—is a direct and essentially immediate consequence of the definition of the Lagrangian / semi-dual functional. The same flatness is already visible in classical duality arguments and in the analysis of (2,2)-WGANs / NOT. Framing this as a previously unrecognized “degenerate saddle-point structure” and claiming that existing convergence analyses are invalidated because they assume optimality of ψ\psi overstates the technical advance. The quantitative estimate (Theorem 4) is a reasonable combination of standard strong-convexity / Brenier regularity arguments with a Kantorovich–Rubinstein residual, but it does not rise to a major new theoretical insight.

  2. Extremely restrictive assumptions that undermine practical relevance.Theorem 4 requires densities in C0,αC^{0,\alpha} that are bounded away from zero, domains that are C2,αC^{2,\alpha} and uniformly convex, and a uniformly Lipschitz family of potentials. These conditions are far stronger than those typically satisfied by the distributions arising in generative modeling (images, tabular data, etc.). The paper itself notes that many real datasets live near lower-dimensional manifolds where Monge maps need not exist (Table 1), yet the main stability result simply assumes the classical Brenier setting. Consequently the claimed explanation of “why more map updates are needed” does not rigorously apply to the regimes in which neural OT is actually used.

  3. Experiments are too synthetic and insufficient to support the practical claims.The main figures (1–2) use only 1-D / 2-D Gaussians. The high-dimensional experiments (Appendix E) rely on an ICNN-generated continuous benchmark whose ground-truth map is itself produced by a strongly convex network—the most favorable setting possible for the theory. There are no experiments on real image data, no comparison against competitive baselines (e.g., recent neural OT or flow-matching methods), and no demonstration that the recommended timescale ratio Kηt/ηψK\eta_t/\eta_\psi yields measurable gains on downstream generative tasks. The observed phenomenon “map converges while potential diverges” is expected once the constraint is nearly satisfied; showing it on toy Gaussians does not constitute strong empirical support.

  4. Technical gaps and imprecise claims about prior work.Several statements about earlier papers (Fan et al., Rout et al., Tarasov et al., Korotin et al., Choi et al.) are presented as exposing fundamental flaws, yet the critique often reduces to “they assumed optimality of the potential.” Because the objective is flat once the map is optimal, those analyses remain valid in the regime they study. In addition, the continuous-time two-timescale system (22) is written down without a rigorous justification that the discrete alternating scheme converges to it, and the constant CC in the stability estimate is left completely opaque. The claim that the numerical procedure is “equivalent to a constrained optimization problem” is only heuristic; the Lagrangian is non-convex and the interchange of inf⁡\inf and sup⁡\sup is not justified under the paper’s own hypotheses.

  5. Insufficient significance for a top-tier venue. The paper essentially rediscovers a structural property of the semi-dual formulation, derives a stability bound under classical (and strong) regularity assumptions, and illustrates on synthetic data that the map can converge without the potential. While the observation is correct and the writing is generally clear, the incremental value relative to the already substantial literature on neural OT, two-timescale GAN dynamics, and stability of Brenier maps is modest. The work does not open a new algorithmic avenue, does not relax existing assumptions in a useful way, and does not deliver compelling empirical evidence that the community should change practice.

Questions for authors

please refer to the weaknesses above.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 8DSa

1. Novelty and prior work. The flatness itself is elementary and follows from the definition; we do not claim otherwise. Our contribution is making its consequences explicit by rewriting F(ψ,t)=∫c(x,t(x))dμ+∫ψ(y)d(ν−t♯μ)F(\psi,t)=\int c(x,t(x))d\mu + \int\psi(y) d(\nu-t_\sharp\mu) in terms of measures. Two kinds of hypotheses appear in earlier work. Some arguments use optimality of the potential (Rout et al. 2021, Thm 4.3). Since FF is independent of ψ\psi at T⋆T^\star, nothing in the objective drives ψ\psi toward ψ⋆\psi^\star, and Figs 1-2 show ψ\psi staying away from ψ⋆\psi^\star while the map converges, so these hypotheses are not enforced by training. Others restrict the potential class (Fan et al. 2021: cc-concave; Tarasov et al. 2026: λ\lambda-strongly convex). Convexity can be enforced by architecture, but at the cost of expressiveness. Our estimate needs only a uniform Lipschitz bound, so it covers unconstrained networks. Separately, Korotin et al. 2023, Lemma 1 is not about flatness: most likely authors conflated the existence of an infimum with a minimizer in the Monge problem (we provide Table 1 for convenience). Choi at al. 2025 oversee the flatness (Cor. 3.3). That's why we find it useful to formulate as a separate theorem.

On whether (17) follows from earlier works. We would ask the reviewer to indicate how the estimate would follow from results that assume optimality, cc-concavity or λ\lambda-strong convexity of the potential, because the implication goes in the opposite direction. On a compact domain, cc-concave potentials are Lipschitz, with a constant controlled by diam(Y)diam(Y), and potentials of the form ∣y∣2/2−v|y|²/2 - v with vv λ\lambda-strongly convex and Lipschitz also lie in the uniformly Lipschitz class. So those hypotheses are special cases of ours, while a Lipschitz ψ\psi need not be cc-concave, strongly convex or near-optimal. The proof has two parts. Step 2 (App. A) uses λ\lambda-strong convexity of the optimal potential v⋆v^\star, which comes from the regularity theory (Caffarelli) and is standard. Steps 3-4 are new: they remove ψ\psi from the estimate adding the term dKR(t♯μ,ν)d_{KR}(t_\sharp\mu,\nu). If the reviewer can show that this follows directly from a cited theorem, we will gladly shorten the proof and credit that result.

2. Assumptions. The hypotheses of Thm 4 are the standard setting for regularity of Brenier maps, and they give strong convexity of the dual potential. To our knowledge, estimate (17) is new even in this setting. We agree they exclude manifold-supported data and regard relaxing them as the main open problem. The explanation of why more map steps help (Sec. 4.2) is an interpretation in the regime where the theory applies, not a convergence guarantee for alternating optimization.

3. Experiments. The contribution is theoretical, and the experiments illustrate it rather than prove it. They require a known T⋆T^\star and ψ⋆\psi^\star. We acknowledge that 192-D benchmark benchmark is favorable to the theory by design, and real data has no ground-truth map. The solvers tested (OTP, MongeMap, MaxCorr, OTM) are representative semi-dual NOT formulations (Table 2). Flow matching is related but is a different problem, without the min-max structure we analyze. We do not claim downstream gains from a particular timescale ratio, and we updated abstract in the revision. If the reviewer has specific baselines in mind, we would be glad to add them.

4. Technical points. (a) Interchange of sup and inf. By Thm 3(iii), sup⁡ψF=+∞\sup_\psi F = +\infty for infeasible t, so inf-sup reduces to the Monge problem. This works for continuous ψ\psi, but not for Lipschitz ψ\psi. We corrected the wording in Sec. 5.5, which now states equivalence for specific Lipschitz constant. (b) The constant. From App. A, Step 4, C=(2/λ)max⁡(1,M)C=(2/\lambda)\max(1,M), with M=max⁡(2L′,L′⋅diamY)M = \max(2L′, L′·diam Y). Here λ\lambda is the strong-convexity constant of the dual potential depending on the domains and densities, and L′L′ is a common Lipschitz bound. (c) System (22) is the formal continuous limit of the alternating scheme (first variations of FF in tt and ψ\psi). We do not prove convergence of the discrete scheme to it, we use it in Discussion as a heuristic.

5. Significance. We see the contribution as the quantitative map-stability estimate that needs no optimal potential, and as a practical diagnostic: map convergence should be verified through dKR(t♯μ,ν)d_{KR}(t_\sharp\mu, \nu) in addition to the objective. Makkuva et al. 2020, Remark 3.4 reports that relaxing convexity constraints performs empirically better. That the map can converge while the potential does not may seem expected once the constraint is nearly satisfied, but it had not been quantified: Thm 4 bounds the map error by the objective gap and marginal mismatch for any ψ\psi in the Lipschitz class, and Remark 2 gives the converse.

6. Ethics flag. We could not find an ethics concern described in the review. Could the reviewer clarify what prompted it?

Reviewer_gBdGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper, the authors contribute further theoretical evidence to justify the convergence of Monge maps in the semi-dual formulation of the optimal transport problem. They do so by describing a new minimax formulation of the semi-dual objective and providing a bound on the error of any feasible solution to the semi-dual optimal transport problem which does not depend on the dual potential. Then, the authors experimentally justify that the convergence of the transport plan to the optimal Monge map is independent of the convergence of the dual potential.

Critical strengths

I found the primary strength of the paper to be the numerical experiments which show how the transport plan in the semi-continuous formulation can be stable even when the dual potentials are still unstable. In particular, the presentation of the phenomenon where convergence of the transport plan and dual potentials occur at different times was interesting to read. The authors also did a pretty good job of formulating the semi-dual optimal transport problem. I found their exposition pleasant to read, though I would just be careful of minor grammatical errors.

Critical weaknesses

My main issue with the paper is that the theoretical contributions which the authors claim to be novel seem to already be well known in the literature. I worry this issue may outweigh the experimental studies that the authors conduct. First of all, I still do not see a difference between the formulations of the semi-dual optimal transport problem given on lines 183 (existing version) and 200 (their formulation). If there is no fundamental difference, then shouldn't Theorem 3 be an immediate consequence of existing work? Moreover, in the statement of Theorem 4, it is unspecified what the value of CC is, but from the proof given in the appendix it appears that CC is proportional to the diameter of the set X \cup Y. If this is true, then shouldn't the L2L_2 norm between a plan and the optimal plan trivially be bounded by the difference in their optimal transport objectives divided by the diameter?

I would also point the authors to a vast literature on semi-discrete optimal transport, where optimization of the dual objective precisely relies upon this separation of the cost and mass which was claimed to be novel in Theorem 3. In particular, the gradient of the dual objective is proportional to the amount of mass contained in the Laguerre cell which is transported to a discrete mass. For a starting point on some references on semi-discrete OT, I would consider:

  • "Minkowski-type theorems and least-squares clustering" by Aurenhammer et al.
  • "Notions of optimal transport theory and how to implement them on a computer" by Levy and Schwindt
  • "Free boundaries in optimal transport and Monge-Ampere obstacle problems" by Caffarelli and McCann

While these references are restricted to the semi-discrete setting, many structural properties of the solutions are maintained. In particular, the paper by Levy and Schwindt seems useful to compare against for their description of the derivative of the dual objective solely as a function of the mass contained within the Laguerre cells.

I will put an asterisk on the criticisms I have about their theoretical contributions. While I think the theorems the authors prove are not novel, the authors extend these theorems to experimentally study their implications in practice for training NOT models. I think that further experimental studies still have value but are outshined by the emphasis on the theoretical contributions from the authors that I have found to be weak. Technically speaking, the theoretical contributions that the authors state are sound and may not have been written in the way the authors write them before.

Questions for authors
  • How does the left-hand side of line 200 differ from the semi-dual formulation described on line 183? The formulation described on line 200 seems to be a major contribution of the authors, and I currently do not see it as a contribution at all given the two formulations seem almost identical, except the authors move the position of the integral of the pushforward measure.

  • What is the constant C on line 260 of the statement of Theorem 4?

  • How does Theorem 4, the other major claimed theoretical contribution of the paper, proves any sort of convergence? There are no assumptions in Theorem 4 that state tkt_k converges to T∗T^* as k→∞k \to \infty. All that Theorem 4 states is that the L2L_2 norm between tkt_k and T∗T^* can be split into an error term in the semi-dual objective function and an error term in the mass transported. This instead seems to be a more precise formulation of the contribution claimed on line 45.

2 comments
AuthorsOpenReview ↗

Response to Reviewer gBdG

We thank the reviewer for the careful reading and the references.

Formulation (lines 183 and 200). The two forms are the same functional: ∫ψdν−∫ψ(t(x))dμ=∫ψ d(ν−t♯μ)\int\psi d\nu-\int\psi(t(x)) d\mu=\int\psi\,d(\nu-t_\sharp\mu). Nevertheless, the works we cite either treat the saddle point as strict (Choi et al. [2025] (Cor. 3.3), or state the objective in a form of Eq. (9) that does not make the degeneracy visible (Rout et al. [2021]; Tarasov et al. [2026], etc); to our knowledge, the constrained-optimization reformulation that exposes it explicitly does not appear elsewhere in the neural OT literature. We consider Thm 4, not the rewriting of FF, as the main contribution. The semi-discrete references are relevant: there the gradient of the dual, νj−μ(Lj(ψ))\nu_j-\mu(L_j(\psi)), is ν−t♯μ\nu-t_\sharp\mu evaluated at the induced map t=tψt=t_\psi, whose preimages are the Laguerre cells. We agree this identity is classical and have added the first two references and a discussion (lines 50--63). In semi-discrete OT the map tψt_\psi is computed exactly from ψ\psi, so the gradient vanishes only at the optimal ψ\psi and the objective is not flat. In neural OT, tt is a separate network approximating tψt_\psi with finitely many steps, so ν−t♯μ→0\nu-t_\sharp\mu\to0 as t→T⋆t\to T^\star for every ψ\psi. This is the degeneracy, and it is what Thm 4 addresses. The third reference is relevant for the unbalanced OT, we didn't include any references on this topic because of the page limit. We really appreciate for brining it to our attention, it will be useful for the extension of our work.

The constant. C=(2/λ)max⁡(1,M)C=(2/\lambda)\max(1,M) with M=max⁡(2L′,L′ diam Y)M=\max(2L',L'\,\mathrm{diam}\,Y), where λ\lambda is the strong-convexity constant of the optimal dual potential ϕ⋆\phi^\star (regularity theory, depending on the domains and density bounds) and L′L' is a common Lipschitz bound. It is now stated explicitly in Thm 4. CC scales as 1/λ1/\lambda, not as the diameter.

Is the estimate trivial? If we understand the question correctly, it asks whether ∥t−T⋆∥2≤∣F(ψ,t)−F(ψ,T⋆)∣/diam\|t-T^\star\|^2\le|F(\psi,t)-F(\psi,T^\star)|/\mathrm{diam}. This cannot hold. Take ψ=0\psi=0 and t(x)=2x−T⋆(x)t(x)=2x-T^\star(x) (e.g., on a unit 2D disk). Then F(0,t)=12∫∣x−t(x)∣2dμ=F(0,T⋆)F(0,t)=\tfrac12\int|x-t(x)|^2d\mu=F(0,T^\star), so the difference is 00, while ∥t−T⋆∥2=4∫∣x−T⋆∣2dμ>0\|t-T^\star\|^2=4\int|x-T^\star|^2d\mu>0 if μ≠ν\mu\ne\nu. This tt is infeasible, so a term controlling t♯μt_\sharp\mu is needed. What our estimate adds is the term dKR(t♯μ,ν)d_{KR}(t_\sharp\mu,\nu), which comes from bounding ∫ψd(ν−t♯μ)≤L′dKR(t♯μ,ν)\int\psi d(\nu-t_\sharp\mu)\le L' d_{KR}(t_\sharp\mu,\nu), so no optimality of ψ\psi is required. For feasible tt, the distance is controlled by the gap through quadratic growth of the objective at T⋆T^\star, which comes from the λ\lambda-strong convexity of the dual potential. The Lipschitz continuity of the cost gives only the opposite inequality.

Convergence. We agree that Thm 4 does not prove convergence of a particular algorithm, and it assumes none. Since it is an estimate and not an inequality in one direction only, the accurate statement is two-sided: under the assumptions of Thm 4, the reverse estimate also holds, so tk→T⋆ in L2(μ)⟺F(tk,ψk)→W22(μ,ν)  and  dKR(tk♯μ,ν)→0.t_k\to T^\star\ \text{in } L^2(\mu)\quad\Longleftrightarrow\quad F(t_k,\psi_k)\to W_2^2(\mu,\nu)\ \text{ and }\ d_{KR}(t_{k\sharp}\mu,\nu)\to0. This is a criterion for convergence of the map, and a more precise form of the estimate in the introduction (line 45). We revised abstract accordingly.

In the revision, we also add in the last paragraph of Sec. 5.5 a more clear statement when changing the order of sup and inf results in the same minimizer T⋆T^{\star}. It's possible, when the optimization set for ψ\psi contains ψ⋆\psi^{\star}.

I greatly appreciate the response of the authors. The discussion regarding whether or not Theorem 4 is trivial was useful to read. However, I still find that the result follows mostly intuitively: measure the difference in the current map tt and the optimal map T∗T^* through the KR distance of the images of the maps. I maintain my vote towards rejection. Below are the primary concerns I still have with the authors' response.

Problem Formulation. Yes, I agree that neural OT does not necessarily solve semi-discrete OT. I also agree that in the semi-discrete OT algorithms, the gradient of the dual is indeed νj−μ(Lj(ψ))\nu_j - \mu(L_j(\psi)), and this is equivalent to the authors' formulation of ν−t♯μ\nu - t_{\sharp} \mu. It is true that the dual is optimized when this gradient is equal to zero. However, similar to your neural OT model, most semi-discrete OT solvers do not ever attain an exact solution, and therefore one can only argue ν−t♯μ→0\nu - t_{\sharp} \mu \to 0 in the semi-discrete setting as well.

The constant CC. I am confused by this discussion. It is claimed that C=(1/λ)max⁡(1,2L′,L′,diam,Y)C = (1/\lambda) \max(1, 2L', L', \text{diam}, Y) does not depend on the diameter?

Convergence. While I agree that the statement should intuitively be two-sided, I still do not understand how line 45 implies that.

Reviewer_sKfLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the semi-dual objective used in neural optimal transport. It shows that the objective is independent of the dual potential whenever the map pushes the source distribution exactly onto the target. For quadratic cost, it bounds map error using the objective gap and distribution mismatch, under strong regularity assumptions and a uniform Lipschitz bound on the potentials. The potential itself need not be optimal. A constrained optimization view is also used to interpret map and potential update schedules. Gaussian experiments and a 192D synthetic benchmark with a known optimal map illustrate the difference between map and potential accuracy across existing solvers.

A partial rerun of the supplied code recovered some qualitative trends, but did not reproduce all reported results.

Critical strengths
  1. Theorem 4 provides a useful map-level stability bound without requiring an optimal learned potential. Under the stated regularity and uniform Lipschitz assumptions, it controls map error through objective discrepancy and marginal mismatch. This makes the distinction between learning a good map and learning a good potential precise.

  2. The synthetic benchmark has a known nonlinear optimal map, which is a good choice for this question. It allows direct assessment of map and potential errors rather than relying on distribution matching alone. Some rerun results, particularly for OTP, also support the observation that an accurate map can coexist with substantial potential error.

Critical weaknesses
  1. The potential-error references in the supplied code do not consistently match the training objectives. MongeMap uses a scaled quadratic cost but is evaluated against the untransformed convex conjugate. OTM scales the correlation reward without matching the reference potential to that scaling. MaxCorr also includes a quadratic regularizer despite being described as unregularized. These discrepancies affect the central empirical comparison between map and potential accuracy. The exact settings for Figures 1–2 could not be located in the package, and partial reruns did not recover all reported results. A different code version could explain some differences, but the connection between the submitted implementation and the figures remains unresolved.

  2. The explanation of update schedules goes beyond what the stability result establishes. Theorem 4 is a conditional error bound, not a convergence guarantee for alternating neural optimization. Flatness in the potential at a feasible map does not imply that an arbitrary potential leaves the map update stationary. The experiments also use a fixed number of outer iterations, so increasing K increases the number of map updates and the computation. Figure 2 shows little improvement from increasing K. Together, these leave the claimed benefit of faster map updates insufficiently supported.

Questions for authors
  1. [Evaluation clarification] What are the correct potential references for each solver after accounting for cost scaling, reward scaling, and regularization? Do the conclusions about map–potential separation remain after correcting the MongeMap and OTM references? Please also clarify the quadratic regularizer in MaxCorr and its description as unregularized.

  2. [Reproducibility clarification] Which scripts, resolved configurations, code version, and seed logs generated Figures 1–6? In a CPU rerun of the submitted appendix pipeline with K=1 and seed=0, OTP reached map RMSE 0.1556 at a learning-rate ratio of 1, but 25.8670 at a ratio of 0.02, compared with roughly 1.4 in Figure 3. The dependency versions satisfied the declared ranges. Could you clarify this discrepancy and provide the missing settings for Figures 1–2, including whether the reported results use best or final checkpoints?

  3. [Support for the training claim] What supports the two-timescale explanation beyond the conditional stability bound? In particular, is a common Lipschitz bound enforced throughout training, and what evidence separates the effect of increasing K from the additional computation?

1 comment
AuthorsOpenReview ↗

Response to Reviewer sKfL

Q1

The reviewer is correct, and this was our error. The public implementations [2, 4] of Fan et al. [1] and Rout et al. [3] normalize the cost or reward per dimension. When porting them, we hardcoded this as 1/641/64, the dimension of our earlier benchmark, and did not update it for the 192-D benchmark. The potential-error evaluation then compared the MongeMap [1], OTM [3] and MaxCorr [5] potentials against the unit-scale conjugate Φ∗\Phi^{\ast}, which is the correct reference only for the unit-scale, unregularized objective of Makkuva et al. [5]. Rescaling the cost leaves the optimal map unchanged but rescales the optimal potential, so the potential panels of Figures 4 and 6 mostly measured this mismatch. OTP used its correct reference, 12∥y∥2−Φ∗\tfrac12\lVert y\rVert^2-\Phi^{\ast}, so Figure 3 is unaffected. MaxCorr's "unregularized" description came from a prior version of the baseline. It includes a quadratic regularizer (λ=0.05(\lambda = 0.05), which we now state.

We now derive each solver's reference from the stationarity of its own map step. For example, MongeMap uses c∥y∥2−2cΦ∗c\lVert y\rVert^2-2c \Phi^{\ast} with c=1/64c = 1/64. Each reference is unit-tested against a closed-form problem. We also found that the OTM gradient-optimality penalty, inherited from the released OTM code [4], targeted E(x)\mathbb{E}(x) rather than the scaled sE(x)s \mathbb{E}(x). We fixed this and retrained OTM. Its map RMSE changes by −7%-7\% to +3%+3\%, improving in 19 of 25 settings.

Map metrics do not depend on the reference, and all 250 retrained MongeMap and MaxCorr runs reproduce their original map metrics exactly. We now report relative potential error, Var⁡(f−f∗)/Var⁡(f∗)\operatorname{Var}(f-f^{\ast})/\operatorname{Var}(f^{\ast}), together with its gradient analogue, where 1 means no better than a constant. With the corrected references, every solver's relative potential-gradient error exceeds its relative map error wherever the map is accurate (L2-UVP below 30%). The ratio is 1.6–3.2× for OTM and 2.8–7.8× for MongeMap. For OTP and MaxCorr it reaches 29× and 33× at small KK, falling to 2.5–4.6× at K≥10K \ge 10. Relative errors in the potential values are much smaller (mostly below 0.2), so the separation is in the gradient. In the revised PDF, panels (c) and (d) of Figures 3–6 show these relative errors, Figure 6 is regenerated from the retrained OTM runs, and Appendices D.1 and E state each solver's scaling and reference. We have revised the text to say this.

Q2

The leftmost point in Figure 3 is at ηψ/ηT=0.05\eta_{\psi}/\eta_{T} = 0.05, not 0.020.02. The list in Appendix D.1 that includes 0.020.02 is from an earlier set of experiments, and we have corrected it. At ratio 0.020.02, the large error the reviewer observed continues the trend in Figure 3. At ratio 1, the reviewer's 0.15560.1556 matches our 5-seed mean of ≈0.17\approx 0.17.

Each point is the final validation metric after 4096 iterations, reported as mean ±\pm std over 5 seeds. The per-run results.json uses the best checkpoint, so its values differ from the figures. The updated supplementary material contains the code, the exact sweep and post-evaluation commands, an environment lockfile, and, for every run in Figures 3–6, the resolved configuration, run specification, full metrics log and results. This is sufficient to rerun every experiment and regenerate every figure. The checkpoints (about 8 GB) are too large to attach, so the lockfile instead records the SHA-256 of each one.

Q3

We agree that Thm 4 is a conditional error bound, not a convergence guarantee for alternating optimization; the abstract and Sec. 4.2 now reworded. Flatness does not make the map update stationary for an arbitrary potential. At a feasible map, ∂ψF=0\partial_\psi F=0, so the objective gives ψ\psi no signal toward ψ⋆\psi^\star. In the map direction, ∂tF(ψ,t)\partial_t F(\psi,t) vanishes only if ∇ψ=∇ψ⋆\nabla\psi=\nabla\psi^\star. This asymmetry is the basis of our interpretation, now clarified in Sec. 4.2 and 5.5 with additional min/max interchange proof forcing sup⁡ψF(ψ,t)\sup_{\psi}F(\psi,t) too diverge.

We did not enforce a common Lipschitz bound. The solvers in Figs. 3-6 are the released implementations without clipping, so the hypothesis of Thm 4 is not verified there, and the experiments illustrate the theorem rather than test it. A sample maximum of ∣∇ψ∣|\nabla\psi| over validation points is only a lower estimate. It stays moderate for K>1K>1 (computed from the logs, we do not report it in the paper).

Larger KK also means more map updates, and Fig. 2 shows little improvement; we removed it to meet the page limit, and the appendix contains the parameter sweeps. We claim neither a downstream benefit nor a guarantee for a particular KK.

The code for Fig. 1 is in the supplement. We run the NOT algorithm, and once the map is close to optimal we add noise to the potential and continue training. The figure shows TT, ∇ψ\nabla\psi and F−W22F-W_2^2 for K=1K=1 at three ratios ηψ/ηT\eta_\psi/\eta_T; the protocol is explained after the figure.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.