acceptodds
Under review as a conference paper at ICLR 2027

Estimation Rates for Optimal Transport Maps to the Barycenter in Multidimensional Spaces

Abstract

The optimal transport barycenter problem seeks transport maps that align multiple distributions with their barycenter, a central object in applications such as image interpolation and fair regression. Although several algorithms estimate these maps in continuous multidimensional spaces, their statistical estimation rates remain largely unknown. We establish, to our knowledge, the first convergence rate for optimal transport map estimation to the barycenter under the quadratic cost in multidimensional Euclidean spaces. Crucially, the rate is governed solely by the approximation error of the estimator's class, possibly non-convexly parametrized such as deep neural networks, and requires no regularity of the underlying transport maps. The resulting rate interpolates between the fast rate under the Poincar\'e inequality and the slow rate in general, two extremes previously established only in isolation for transport map estimation between two distributions.

Primary area:
learning theory
Keywords:
optimal transport, optimal transport barycenter, transport map estimation, convergence rates, nonparametric statistics, Poincaré inequality

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fx4nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the estimation of the optimal transport maps in the Wasserstein barycenter problem. The authors design an objective function based on a known potential-minimization characterization of such maps [1]. They introduce an interpolation regime between the unconstrained case and the Poincaré inequality-constrained case, and then provide a stability bound and convergence rates.

[1] Korotin et al. Continuous Wasserstein-2 Barycenter Estimation without Minimax Optimization. ICLR2021

Critical strengths

This paper is trying to address a real gap for barycenter estimation on the continuous multi-dimensional measures and aims for a rather general set of measures.

Critical weaknesses
  1. This paper seems rather incremental to me compared to Fukuchi (2025). The main change is the scope from 1D to multi-dimensional measures, but the extra mechanism needed for the extension is standard. The estimator, the potential construction and the stability bound carry over from the 1D case with little change, and the truncation used to handle unbounded supports is also a standard technique. (See Fukuchi (2025) Remark 4.1, Proposition 5.6)

  2. The authors motivate the estimation of barycenter maps through applications such as image interpolation and fair regression, but never show such applications. It is also unclear how the optimization is carried out. This is not a deal breaker, but it makes the paper appear incomplete, especially since the theoretical contribution is incremental.

  3. The paper claims its estimator accommodates the deep neural network parametrizations widely used in practice. However, Def. 2 requires bijective gradient maps that are monotone, L-Lipschitz, and exactly congruent everywhere, and standard architectures do not satisfy this. For instance, Korotin et al. (2021) enforce congruence only approximately, through a regularizing term.

Questions for authors
  1. What is the motivation for the intermediate regime α∈(0,1)\alpha\in (0, 1)? α=0,1\alpha=0, 1 are both well-motivated. For measures that satisfies the Poincaré-inequality, intermediate α\alpha gives no improvement. The usual Poincaré inequality with constant CC implies the order-α\alpha inequality with CαC^\alpha for every α∈[0,1]\alpha\in[0, 1]. For measures with disconnected support, especially heavy-tailed measures, the tail term ΞR\Xi_R decays slowly. Can you give me some examples of measures in such a regime with the estimation rates?
  2. How do you address the case where ∣∣f∣∣∞=∞||f||_\infty=\infty for α∈(0,1)\alpha\in (0, 1)?
3 comments

We thank the reviewer for the careful reading and the insightful comments. We address the concerns and questions one by one.

W1. Novelty relative to Fukuchi (2025)

We would like to highlight that our analysis goes significantly beyond Fukuchi (2025) in three respects. First, Fukuchi (2025) assumes a bounded domain in R\mathbb{R}, whereas our analysis allows unbounded and heavy-tailed measures through the clipping argument and the tail term ΞR\Xi_R. Second, Proposition 5.6 of Fukuchi (2025) requires the true maps to be bi-Lipschitz, whereas Theorem 3 requires regularity only of the candidate maps. Hence, Theorem 2 imposes no regularity on the true maps, which affect the rate only through the approximation error AA. Third, we bring the Poincaré inequality of order α\alpha to the analysis of transport map estimation, which is a significant technical contribution. It bridges the fast-rate analysis under the usual Poincaré inequality and the slow-rate analysis without it, which have been developed separately, and it extends the analysis to heavy-tailed measures that fail the usual Poincaré inequality, as demonstrated in Section 4.1.

W2. Applications and optimization

Several algorithms estimate the transport maps to the barycenter in multidimensional spaces, but their convergence rates have been unknown, and our paper fills this gap. In particular, our estimator is the empirical counterpart of the potential minimization of Korotin et al. (2021), whose neural-network implementation is evaluated empirically in their paper. Hence, our result also offers insight into the statistical performance of this practical approach, although their implementation enforces congruency only approximately. Regarding the optimization, the minimality of the estimator is used only to make the second term of the decomposition in Section 5.2 non-positive. Consequently, for any approximate minimizer with optimization error η\eta, the same analysis gives the bound in Theorem 2 with AA replaced by A+ηA + \eta, which cleanly separates the statistical and optimization errors.

W3. Neural networks and exact congruency

We agree that standard architectures do not satisfy Definition 2 as they are. Our claim is that classes satisfying Definition 2 can be built from neural networks. Section 4.1 provides such a class from sparse ReQU networks, for which Prop. 1 verifies all the requirements of Definition 2, including exact congruency everywhere without a regularizing term.

Q1. Motivation for α∈(0,1)\alpha \in (0,1)

We agree that α=1\alpha = 1 is preferable when the usual Poincaré inequality holds. The intermediate regime targets measures that do not satisfy it, such as polynomially tailed ones, since the usual Poincaré inequality implies sub-exponential tails. For example, the density proportional to (1+∥x∥2)−(p+κ)/2(1+\Vert x\Vert^2)^{-(p+\kappa)/2} satisfies the inequality of order α=κ/(2κ+p+2)\alpha = \kappa/(2\kappa+p+2). For sufficiently large κ\kappa, Prop. 2 gives the resulting rate, which accounts for the polynomial decay of ΞR\Xi_R.

Q2. The case ∥f∥∞=∞\Vert f\Vert_\infty = \infty

In this case, the right-hand side of Definition 1 is infinite for α<1\alpha < 1 unless ∇f=0\nabla f = 0, and the inequality imposes no constraint.

I thank the authors for the response. My concern about the intermediate regime of α has been addressed. To clarify my first point, I understand the differences between the result itself and Fukuchi (2025). However, my take is that although this is a first rate for the problem, it is obtained largely by combining existing ingredients: the stability bound that avoids Lipschitz assumptions on the true maps, which comes from Korotin et al. (2021), and standard truncation to handle unbounded supports. I do give credit for bringing the order-α Poincaré inequality into this analysis. I also still do not see how the result applies in practice. The ReQu class in Section 4.1 is a special construction for which no empirical evidence is given. Without experiments or a computable estimator, it is unclear how the guarantee can be useful. For these reasons, I will keep my score.

Edit: The reason behind my Q2 is that Def1 is vacuous for unbounded ff when α<1\alpha < 1, while for measures with unbounded support the potential differences that the analysis needs to control are unbounded. As I understand it, this is handled correctly, since Def. 1 is applied only after truncation. For heavy-tailed measures, which are the motivation for α<1\alpha < 1, the rate in Prop. 2 is much slower than in the light-tailed case. This is a minor point, and I have no remaining concern here.

We thank the reviewer for the follow-up and the clarification.

Combination of existing ingredients

We respectfully but strongly disagree, as explained in the general response. In particular, the stability bound of Korotin et al. (2021) provides no convergence rate, and the truncation used in existing analyses cannot be applied as it is, because it breaks the congruency and the local Lipschitz continuity required by the generalized Poincaré inequality. In addition, the revised paper now establishes a new stability bound (Theorem 3), which requires congruency only on a ball.

Practical usefulness

Theoretical analyses of transport map estimation, such as Divol et al. (2025), also establish statistical guarantees without experiments, and their value lies in clarifying which rates are achievable and under which conditions. Our paper provides such a guarantee for the barycenter problem, for which no convergence rate in multidimensional spaces was available. The guarantee is not tied to the ReQU class: Theorem 2 applies to any class satisfying Definition 2, and to approximate minimizers with AA replaced by A+ηA + \eta, where η\eta is the optimization error.

Reviewer_WtNiOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper provides a theoretical statistical analysis of estimating optimal transport maps to the barycenter in multidimensional Euclidean spaces. While optimal transport barycenters are widely used for tasks like image interpolation and fair regression, the convergence rates for estimating the maps to these barycenters were previously unknown, except in 1D or discrete settings. The authors establish the first convergence rate for this problem under quadratic cost, showing that the rate is governed by the approximation error of the estimator's function class and the geometry of the underlying measures. Crucially, their result allows for non-convex parameterizations (like deep neural networks), requires no regularity on the true transport maps, and smoothly interpolates between the fast rate (under the Poincaré inequality) and the slow rate (in general).

Critical strengths
  1. It provides the first rigorous convergence rate for transport map estimation to the barycenter in multidimensional spaces, filling a significant gap in the statistical learning theory of optimal transport. 2). The convergence bound depends only on the approximation error of the estimator's class. This means it applies to highly expressive, non-convex parametrizations like deep neural networks, unlike previous theoretical works that required strict convexity or regularity.
Critical weaknesses
  1. The analysis relies on certain assumptions about the underlying probability measures, such as finite second moments and specific behavior on small sets (vanishing on small sets), which might not hold for highly irregular or heavy-tailed real-world data.
  2. The paper is not easy-to-follow
  3. The paper is purely theoretical and focuses on the statistical error of an estimator, it does not propose a new, faster computational algorithm for actually finding the barycenter or the transport maps in practice
Questions for authors
  1. Can the generalized Poincaré inequality (and the resulting interpolation between fast and slow rates) be empirically estimated or verified for real-world datasets before training the transport maps?
  2. How does the estimation error behave if the true optimal transport maps are highly discontinuous (e.g., when the target distributions have disconnected supports), given that the bound does not require regularity on the maps?
  3. How does the proposed convergence rate explicitly depend on the dimension of the Euclidean space, and does it suffer from the curse of dimensionality for standard neural network classes?
3 comments

We thank the reviewer for the careful reading and the insightful comments. We address the concerns and questions one by one.

W1. Assumptions on the measures

Finite second moments are needed for the barycenter with respect to the optimal transport cost to be well defined, and vanishing on small sets is the standard condition under which the transport maps to the barycenter exist and are unique (Agueh & Carlier, 2011). Beyond these two conditions, our results require no regularity of the transport maps and cover heavy-tailed measures. Indeed, Prop. 2 gives explicit rates for polynomially tailed densities.

W3. No new algorithm

We would like to highlight that our main contribution is an estimator of the transport maps to the barycenter with a provable convergence rate, which has been unknown in multidimensional spaces. Providing a faster computational algorithm is beyond our scope. We believe that this guarantee is valuable in its own right, because our estimator is the empirical counterpart of the potential minimization that Korotin et al. (2021) implement with neural networks. Hence, our result also offers insight into the statistical performance of this practical approach. Moreover, our analysis extends to approximate minimizers, with the optimization error added to the approximation error AA.

Q1. Estimating the order α\alpha

The estimator does not use α\alpha, which appears only in the analysis. Hence, α\alpha need not be estimated, although the choice of (R,L,D)(R, L, D) still matters and can be made by cross-validation.

Q2. Discontinuous maps and disconnected supports

Theorem 2 requires no regularity of the true maps, including continuity; they enter only through the approximation error AA. Discontinuous maps make AA larger for Lipschitz classes, which is traded off against the factor 2L2L. Our result also applies to measures with disconnected supports. In this case, the Poincaré inequality of order α>0\alpha > 0 fails, and Theorem 2 with α=0\alpha = 0 gives the slow rate.

Q3. Dependence on the dimension

The dimension enters through DD and AA. For Hölder potentials of smoothness γ\gamma and the ReQU class, the rate is n~−2(γ−1)/(p+2(γ−1))+ϵ\tilde n^{-2(\gamma-1)/(p+2(\gamma-1))+\epsilon} for any ϵ>0\epsilon > 0 under light tails (Prop. 2). It suffers from the curse of dimensionality, as is typical in nonparametric estimation. Overcoming the curse of dimensionality is beyond the scope of this paper.

We would like to kindly follow up on our response. We have uploaded a revised paper, which also reflects the follow-up discussion with the other reviewers summarized in the general response. We would appreciate it if you could let us know whether our response has addressed your concerns, including the assumptions on the measures, the estimation of the order α, and the dependence on the dimension. We would be happy to clarify any remaining points.

Reviewer_WtNiOpenReview ↗

Answer

Thank you, you address my question, but additional ablation experiments about dependence on dimensionality are still necessary more in the next revision.

Reviewer_4BCzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies estimation of optimal transport maps from several distributions to their unknown Wasserstein barycenter in Rp\mathbb{R}^p. It minimizes an empirical potential over a class of tuples of continuous bijective gradient maps satisfying congruency, monotonicity, and Lipschitz constraints and combines stability with an empirical-process analysis. For zero potential approximation error, the claimed statistical term is (Dlog⁡β(n~)n~)12−α\left(\frac{D\log^{\beta}(\tilde{n})}{\tilde{n}}\right)^{\frac{1}{2-\alpha}}, suppressing dependence on L,RL,R and distributional constants. The claimed rate ranges from slow to fast as the order α\alpha of the Poincaré inequality varies from 0 to 1. The analysis avoids direct smoothness assumptions on the true maps, subject to the stated measure, class, and tail conditions. The paper is theoretical.

Critical strengths

A meaningful statistical contribution if the main theorem is established.

Joint estimation of transport maps to an unknown multidimensional barycenter is not directly covered by deterministic barycentric-map stability [1] or two-marginal oracle analyses [2]. An oracle bound that accounts for approximation error and shows how the rate depends on the order of the Poincaré inequality would be a substantive contribution.

Critical weaknesses
  1. The main statistical guarantee requires additional entropy control and proof corrections.
    Theorem 2 allows (R,L,D)(R,L,D), and hence the estimator class, to depend on the sample size. However, Definition 2(d) provides only an asymptotic covering entropy bound for each fixed class. Appendix I (lines 1308–1311) uses this condition at a sample-size-dependent resolution without controlling how the scale at which the bound becomes effective depends on the class. An appropriate uniform finite-scale entropy condition may repair this step, but such a condition and its dependence on (R,L,D)(R,L,D) are not established in the current theorem. Consequently, the proof does not establish Theorem 2 for growing classes under the stated assumptions. Appendix H has independent proof issues. In the A=0A=0 argument (pp. 23–24), the proof assumes u>tu > t, using uu as shorthand for the expected potential error and tt for the proposed bound, but derives only t≤ut\le u and calls this a contradiction. The optimization at lines 1187–1194 drops a factor of two: writing the two terms as δQ2+mδ\delta Q^{2}+\frac{m}{\delta} with Q,m>0Q,m > 0, inf⁡δ>0(δQ2+mδ)=2Qm\inf_{\delta > 0}\left(\delta Q^{2}+\frac{m}{\delta}\right)=2Q\sqrt{m}. Moreover, Theorem 2 covers A=0A=0 or A=ω(sn~)A=\omega(s_{\tilde{n}}), leaving out 0<A=O(sn~)0 < A=O(s_{\tilde{n}}), including the balance used in Implication 1; sn~s_{\tilde{n}} denotes the displayed statistical term. A corrected oracle-bound argument and recalculated constants may resolve these issues, but the current proof does not justify the main statistical guarantee.
  2. The claimed neural-network applicability remains insufficiently demonstrated.
    The abstract and contribution statement emphasize neural-network parameterizations, but allowing a nonconvex parameterization and citing network covering-number results do not establish a class satisfying all of Definition 2 while achieving useful approximation error. Congruency in Definition 2(a) must hold jointly with continuity, bijectivity, the gradient structure, monotonicity, and Lipschitz constraints. Separately, Theorem 3 warrants a direct comparison with [1, Theorem 4.1, Eqs. (9)–(10)], which already bounds the squared map error by the potential gap times twice the candidate potentials' smoothness constant, under exact congruency. The measure assumptions differ, and that stability result does not subsume the present statistical analysis of transport maps to the barycenter. References

[1] Alexander Korotin, Lingxiao Li, Justin Solomon, and Evgeny Burnaev. Continuous Wasserstein-2 Barycenter Estimation without Minimax Optimization. ICLR, 2021.

[2] Shubo Li, Yizhe Ding, Lingzhou Xue, and Runze Li. Stability and Oracle Inequalities for Optimal Transport Maps between General Distributions. NeurIPS, 2025.

Questions for authors
  1. Can the authors justify the finite-scale entropy bound in Appendix I when the function class depends on the sample size? If an additional uniform entropy assumption is needed, can the authors state its dependence on (R,L,D)(R,L,D) and establish whether Theorem 2 and the approximation–estimation trade-off in Implication 1 remain valid under the revised assumptions?

  2. Can the authors repair the contradiction argument and constant calculations in Appendix H and establish an oracle bound covering the missing regime 0<A=O(sn~)0 < A=O(s_{\tilde{n}}), where sn~s_{\tilde{n}} denotes the statistical term in Theorem 2? Can the argument also cover α=0\alpha=0 and justify Implication 1?

Additional question

1.Can the authors provide a concrete neural-network class satisfying Definition 2, with controlled entropy and an approximation error that decreases as the class grows, or qualify the claimed neural-network applicability? Can the authors also clarify how the input-measure assumptions in Theorem 3 differ from those in [1, Theorem 4.1]?

3 comments

We thank the reviewer for the careful reading, including the proofs, and the insightful comments. We address the concerns and questions one by one.

W1 and Q1. Entropy control for growing classes

We thank the reviewer for pointing out this critical issue. We changed Definition 2(d) to require ln⁡N(ϵ,TR,L,D,∥⋅∥Lˉ∞(BR))≤Dln⁡β(1/ϵ)\ln N(\epsilon, \mathcal T_{R,L,D}, \Vert\cdot\Vert_{\bar{L}^\infty(B_R)}) \le D\ln^\beta(1/\epsilon) for all ϵ∈(0,1/(eDLR)]\epsilon \in (0, 1/(eDLR)], instead of an asymptotic lim sup⁡\limsup. The range of ϵ\epsilon makes the dependence on (R,L,D)(R,L,D) explicit. Theorem 2 now assumes DL2R2=o(n~)DL^2R^2 = o(\tilde n), which ensures that the resolution used in the proof lies in this range. Under these revised assumptions, Theorem 2 and Implication 1 remain valid, and the ReQU class in Section 4.1 satisfies the condition with β=1\beta = 1.

W1 and Q2. The proof of the case A=0A = 0 in Appendix H

The reviewer is right that the last step only gave a lower bound consistent with the assumption u>tu > t. We now bound uα/2<utα/2−1u^{\alpha/2} < u t^{\alpha/2-1} before rearranging, which yields the contradiction under the same condition on Bn~B_{\tilde n}, including the case α=0\alpha = 0. Since this argument requires the statistical term to be small, the first case of Theorem 2 now states explicitly that it is o(1)o(1).

W1 and Q2. Missing factor of two

We thank the reviewer for pointing out this issue. We carefully checked all the proofs and corrected the constants; the rates are unchanged.

W1 and Q2. The balance used in Implication 1

We weakened the condition of the second case of the main theorem: it now suffices that A≤an~A \le a_{\tilde n} for some an~=O(1)a_{\tilde n} = O(1) with an~=ω(sn~)a_{\tilde n} = \omega(s_{\tilde n}). This covers 0<A=O(sn~)0 < A = O(s_{\tilde n}), including the balanced choice of DD in Implication 1.

W2 and AQ1. A concrete neural-network class

We provide a concrete neural-network class built from sparse ReQU networks in Section 4.1. Congruency holds exactly by construction, and Prop. 1 verifies all conditions of Definition 2. Prop. 2 shows that the approximation error decreases as the network grows and gives the resulting convergence rates for light- and heavy-tailed measures.

W2 and AQ1. Comparison with Korotin et al. (2021, Thm. 4.1)

We agree that Theorem 3 is closely related to Korotin et al. (2021, Eqs. (9)–(10)). Under exact congruency, their bound coincides with Theorem 3, with their smoothness constant BB in place of LL. The input-measure assumptions differ: Korotin et al. assume absolutely continuous measures with finite second moments, one of which has a bounded density, whereas Theorem 3 requires only finite second moments and vanishing on small sets. In the revised version, we attribute Theorem 3 to Korotin et al. (2021), as it follows from a slight modification of their proof of Theorem 4.1.

Thank you for the detailed response and the revised manuscript. Some of my concerns have been resolved. The revised entropy condition makes the dependence on (R,L,D)(R,L,D) explicit, and the argument in Appendix H is now correct.

The ReQU construction answers my question about a concrete class satisfying Definition 2. However, the condition μ0>1−λ(K)\mu_0>1-\lambda^{(K)} in Proposition 2 appears restrictive for large KK. I suggest acknowledging this limitation and qualifying the claim that the estimator accommodates neural networks "widely used in practice."

Regarding novelty, the estimator and overall analysis closely follow Fukuchi (2025), while the stability bound is due to Korotin et al. (2021). Definition 1 also appears closely related to the weak Poincaré inequality [3, 4], which is not cited. Could the authors clarify this connection and better distinguish their technical contribution from existing results?

I appreciate the improvements but maintain my rating of 2 due to the remaining concerns about theoretical novelty and practical applicability.

[3] Thomas M. Liggett. L2L_2 Rates of Convergence for Attractive Reversible Nearest Particle Systems: The Critical Case. The Annals of Probability, 19(3):935–959, 1991.

[4] Michael Röckner and Feng-Yu Wang. Weak Poincaré Inequalities and L2L^2-Convergence Rates of Markov Semigroups. Journal of Functional Analysis, 185:564–603, 2001.

We thank the reviewer for the follow-up and the constructive suggestions, which we have incorporated into the uploaded revised paper.

Novelty and the weak Poincaré inequality

We respectfully but strongly disagree that our analysis closely follows existing results, as explained in the general response, which also discusses the difference from Fukuchi (2025). We thank the reviewer for pointing out the weak Poincaré inequality. For α∈(0,1)\alpha \in (0,1), Definition 1 is equivalent, up to constants, to the weak Poincaré inequality of Röckner and Wang (2001) with a rate function proportional to r−(1−α)/αr^{-(1-\alpha)/\alpha}, and the revised paper states this connection after Definition 1.

Practical applicability and the condition μ0>1−λ(K)\mu_0 > 1 - \lambda^{(K)}

We thank the reviewer for this suggestion. In the revised paper, Proposition 2 holds for every μ0>0\mu_0 > 0, as summarized in the general response. Hence, the ReQU construction in Section 4.1 applies regardless of the number of distributions KK, and the rates in Theorem 2 are unchanged. We also qualified the claim on neural networks as suggested: the contributions now state that the estimator admits classes built from deep neural networks, as demonstrated in Section 4.1.

Comments on the paper

2 comments

We thank all the reviewers for their careful reviews and insightful comments. We would like to highlight that our main contribution is to establish, to our knowledge, the first convergence rate for estimating transport maps to the barycenter in multidimensional Euclidean spaces. Existing results are limited to one-dimensional or discrete probability measures. Our convergence rate is characterized by the Poincaré inequality of order α\alpha. Bringing this notion to the analysis of optimal transport map estimation is a key contribution of this paper. Existing analyses of transport map estimation, even between two distributions, rely either on the usual Poincaré inequality, which yields the fast rate, or on no such inequality, which yields the slow rate, and these two lines of analysis have been developed separately. The Poincaré inequality of order α\alpha bridges them within a single analysis and extends it to measures that fail the usual Poincaré inequality, such as heavy-tailed ones. We expect this tool to be of independent interest beyond the barycenter problem.

We made the following revisions to address the reviewers' concerns. The convergence rates in the main results are unchanged.

  1. Finite-scale entropy condition (4BCz). Definition 2(d) now requires ln⁡N(ϵ,TR,L,D,∥⋅∥Lˉ∞(BR))≤Dln⁡β(1/ϵ)\ln N(\epsilon, \mathcal T_{R,L,D}, \Vert\cdot\Vert_{\bar{L}^\infty(B_R)}) \le D\ln^\beta(1/\epsilon) for all ϵ∈(0,1/(eDLR)]\epsilon \in (0, 1/(eDLR)], instead of an asymptotic lim sup⁡\limsup.
  2. A concrete neural-network class (4BCz, fx4n, WtNi). New Section 4.1 constructs an (R,L,D)(R,L,D)-class from sparse ReQU networks that satisfies every condition of Definition 2. It also derives explicit rates for Hölder potentials under light-tailed and polynomially tailed measures (Prop. 2).
  3. Weaker assumption in the second case of the main theorem (4BCz). It now suffices that the approximation error is bounded above by some sequence an~a_{\tilde n}.
  4. Corrected proof of Theorem 4 (4BCz). We fixed the contradiction argument for A=0A = 0, which now also covers α=0\alpha = 0, and recomputed the constants.

We thank Reviewers 4BCz and fx4n for the follow-up discussion.

We respectfully but strongly disagree that our result is just a combination of the existing ingredients. First, the stability results for two distributions, such as Divol et al. (2025) and Li et al. (2025), cannot be applied to our problem, because an estimator based on their techniques requires samples from the barycenter, which are unavailable. Second, the stability result of Korotin et al. (2021) can be applied to our problem, but they do not provide a convergence rate, and deriving a convergence rate from their stability bound requires nontrivial effort. Third, Fukuchi (2025) establishes a convergence rate for the potential minimization, but only for one-dimensional bounded distributions under the standard Poincaré inequality. Our result covers multidimensional unbounded distributions under the generalized Poincaré inequality (Definition 1), including heavy-tailed ones. This extension cannot be obtained just by applying the truncation used in existing analyses, such as Divol et al. (2025) and Li et al. (2025), which restrict the potentials to a ball. The unknown barycenter is eliminated through congruency, a pointwise identity that couples all the potentials on the barycenter side, and truncating each input measure to a ball breaks this identity because the truncated regions are mapped to different regions of the barycenter. Moreover, the generalized Poincaré inequality requires a finite supremum norm and local Lipschitz continuity, and the restriction to a ball destroys the latter. We therefore keep the objective untruncated, control the region outside the ball through the coupling of the input measures in Theorem 3, and clip the potentials in the concentration step, which keeps them bounded and locally Lipschitz.

We have revised the paper based on the follow-up comments and uploaded it. The rates in Theorem 2 are unchanged.

  1. (Proposition 2 for every μ0>0\mu_0 > 0 and a new stability bound (4BCz, fx4n).) By slightly modifying the main analysis, we show that Proposition 2 no longer requires μ0>1−λ(K)\mu_0 > 1 - \lambda^{(K)}. Specifically, Definition 2 now requires congruency only on the ball B2LRB_{2LR}, in exchange for μ\mu-strong monotonicity, which is not restrictive because μ>0\mu > 0 can be arbitrarily small. The analysis under the new Definition 2 yields a new stability bound (Theorem 3) with an additional tail term, which is negligible under the assumptions of Theorem 2.
  2. (Weak Poincaré inequality (4BCz).) After Definition 1, we clarify its connection to the weak Poincaré inequality of Röckner and Wang (2001). Although Definition 1 is a special case of their inequality, bringing this concept into the analysis of transport map estimation is a non-trivial significant contribution.
  3. (Clarified contributions (4BCz, fx4n).) We clarified the contributions discussed above in the paper. Section 2 now explains that the stability results for two distributions require samples from the target measure, which are unavailable for the barycenter. The overview of the analysis in Section 5 now explains why the truncation used in existing analyses is not applicable and that we instead clip the potentials.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.