acceptodds
Under review as a conference paper at ICLR 2027

Learning with empirical measures: practical transformers with universality guarantees

Abstract

We propose the *Quantile Transformer*, a transformer-based architecture for learning with empirical measures. This architecture alleviates the computational burden of existing transformer layers for this task, which either require all-to-all particle interactions or rely on expensive optimal transport subprocedures. For measures on , we encode each empirical measure as a finite-dimensional token derived from order statistics and establish universal approximation guarantees for continuous maps on compact classes. For measures on , we extend both the construction and the universality result to compact classes of elliptical distributions using finitely many one-dimensional projections. Preliminary synthetic experiments validate our approach in terms of both learning performance and computational efficiency.

Primary area:
learning theory
Keywords:
measure-to-measure learning, distributional inputs, Wasserstein space, empirical measures, sequence-to-sequence learning, transformers, universal approximation, quantile representations, sliced Wasserstein distance, elliptical distributions

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FDwhOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Prior work in the evolution of point clouds has used a microscopic approach, where the evolution of each particle is tracked and predicted. The authors of the submission counsel for a macroscopic approach, where the quantiles of the initial and final distribution are monitored. The authors show that for simple problems, summarizing the population using such an empirical measure and compressing this into a finite-dimensional token leads to good performance.

Critical strengths

The strength of the method lies in its alleviation of the computational burden for problems where one has to predict how populations evolve under an unknown transformation. When the detailed trajectories of individual points are not required, this reduces to a measure-to-measure (M2M) regression problem. Vandergrift et al have detailed this approach. The authors advance this approach and in their macroscopic formalism the attention mechanism operates on measures instead of individual particles. This leads to savings in computational complexity. One has to note, that this reduction in complexity has been achieved by simplifying the problem to a macro-scale. In my opinion, Baptista et al have also considered this, but eschewed a quantile approach due to its problems in scaling to higher dimensions.

Critical weaknesses
  1. As the Quantile Transformer relies on quantiles, how do the authors deal with scaling to higher dimensional point clouds? There the authors are obligated to resort to simplifications such as Sliced Projections. In higher dimensions, the universal approximation guarantees established by the authors break down in my opinion.

  2. In my opinion, the method should not work in cases where the distribution has a disconnected support.

  3. The quantile / measure approach ignores any localized interactions between the particles.

  4. The approach is very sensitive to smaller sizes of samples. This limitation is exacerbated as the dimensionality of the space increases.

Questions for authors
  1. As the Quantile Transformer relies on quantiles, how do the authors deal with scaling to higher dimensional point clouds? There the authors are obligated to resort to simplifications such as Sliced Projections. In higher dimensions, the universal approximation guarantees established by the authors break down in my opinion. Please provide addiitonal proof for this.

  2. In my opinion, the method should not work in cases where the distribution has a disconnected support. How does the method extend to such cases.

  3. The quantile / measure approach ignores any localized interactions between the particles. What are the ramifications of this?

  4. The approach is very sensitive to smaller sizes of samples. This limitation is exacerbated as the dimensionality of the space increases. What is the minimum number of samples for R^d and how does this compare with the M2M approach of Vandergrift et al?

4 comments
AuthorsOpenReview ↗

Response – Part I: On the scope of our theoretical results

Dear reviewer FDwh,

Thank you for your time, comments, and clarification questions, which helped us improve our manuscript. We address concerns point by point below. In addition, we summarize manuscript changes and some common points raised by reviewers in a general response. Note that revised manuscript additions are in blue. References are in the final comment.

  1. Regarding the scope of our universal approximation result in higher dimensions.

    We address the reviewer's claim that

    There [in higher dimensions] the authors are obligated to resort to simplifications such as Sliced Projections. In higher dimensions, the universal approximation guarantees established by the authors break down in my opinion.

    We respectfully note there is a misunderstanding of our results on this point. We devote Sec. 4 entirely to describing a slicing-based architecture which is provably universal whenever the measures are elliptical, with PD covariance, and belong to P2(Rd)\mathcal P_2(\mathbb R^d) (see Prop. 4.2 and Thm. 4.1). We prove this result by relying on the finite Cramer-Wold type result of [1], which ensures that elliptical measures are identifiable from a finite number of carefully chosen slices such as Vd\mathcal V_d in Eq. (15). All claims are rigorously proven in the respective appendix sections.

    If, however, the reviewer refers to generic measures in Rd\mathbb R^d, then indeed, the finite set of slices we consider is not identifiable for such generic measures. We emphasize that we do not claim such a result. Indeed, in the Conclusions Sec. 6 we explicitly state that the general case remains open.

  2. Regarding connectedness of measure support.

    We address the reviewer's claim that

    In my opinion, the method should not work in cases where the distribution has a disconnected support.

    We respectfully point out that our theoretical results disagree with the reviewer's claim. In particular, Thms 3.1 and 4.1 make no assumption of the connectedness of the support, and include measures with disconnected support whenever they satisfy our assumptions, i.e., when the 1D measures belong to P2(R)\mathcal P_2(\mathbb R), and the higher-dimensional measures are elliptical, with PD covariance and belong to P2(Rd)\mathcal P_2(\mathbb R^d) for d≥2d\geq 2. For example, Thm. 3.1 accommodates all finitely supported measures with at least two distinct support points (which have disconnected support by definition) since they belong to P2(R)\mathcal P_2(\mathbb R). For the elliptical case in Rd,d≥2\mathbb R^d, d\geq 2, take e.g., X=RUX=RU, where U∼Unif(Sd−1)U\sim\mathrm{Unif}(\mathbb S^{d-1}), R⊥UR\perp U, and P(R=1)=P(R=2)=12\mathbb P(R=1)=\mathbb P(R=2)=\tfrac12. Then Law(X)\mathrm{Law}(X) is spherically symmetric, hence elliptical, has PD covariance, disconnected support Sd−1∪2Sd−1\mathbb S^{d-1}\cup 2\mathbb S^{d-1}, and belongs to P2(Rd)\mathcal P_2(\mathbb R^d). This type of measure thus satisfies all assumptions needed by Thm 4.1. In short, our method includes measures with disconnected support as long as they comply with the conditions of Thms. 3.1 and 4.1.

AuthorsOpenReview ↗

Response – Part II: On the absence of localized interactions and the minimum number of samples

  1. On possible ramifications of ignoring localized particle interactions.

    • From a formal approximation standpoint, there are no ramifications in the sense that, whenever measures evolve through local particle interactions and this evolution gives rise to a W2W_2-continuous map between the compact sets allowed by Thms. 3.1 and Thm 4.1, there exists a Quantile Transformer that can approximate it for a sufficiently large M with high probability.

    • From a practical standpoint, a possible ramification concerns learning efficiency, i.e., how many training sequences are needed to reach a certain W2W_2 test error level. Specifically, a particle transformer with neighborhood attention applies a shared rule directly to nearby particles, which can make such local dynamics easier to learn from finite data. In contrast, since the Quantile Transformer attends over measure tokens, it learns how local interactions affect the next distribution without explicitly updating each particle from its neighbors. It may therefore require more training sequences to attain the same next-measure W2W_2 error on unseen sequences. We include this discussion in App. A.2.

  2. On a universal, dd-dependent minimum MM.

    • We address the following reviewer question

      The approach is very sensitive to smaller sizes of samples. This limitation is exacerbated as the dimensionality of the space increases. What is the minimum number of samples for R^d and how does this compare with the M2M approach of Vandergrift et al?

      There is no universal minimum MM for Rd\mathbb R^d depending exclusively on dd. For a given error ε\varepsilon and failure probability δ\delta, Thm 4.1 gives only a sufficient threshold M0M_0 that also depends on the distribution class, the continuity modulus of φ~\widetilde\varphi in Sec. 2.1--Step 1, and the elliptical decoder stability. Likewise, Vandergrift et al. [3] give no minimum MM.

      There is, however, a dimensional limitation shared by both methods: for d>4d > 4, approximating a single distribution within W2≤εW_2\leq\varepsilon can require O(ε−d)\mathcal O(\varepsilon^{-d}) particles. Fournier and Guillin [2] show that for a fixed confidence, the worst-case W2W_2 error of an MM-particle empirical measure scales as M−1/dM^{-1/d}, even for laws supported on the unit ball. In particular, the uniform law on the unit ball is relevant since it belongs to our elliptical class: here, a matching lower bound of Ω(M−1/d)\Omega(M^{-1/d}) can be extracted from [2], point c) just below their Thm. 1. Put differently, even when the target law is known, the W2W_2 error cannot decrease past M−1/dM^{-1/d} for this example. Thus, at fixed ε\varepsilon, any method outputting an MM-particle empirical measure needs exponentially-in-dd many particles to approximate the target law. Our M0M_0 in Thm 4.1 may be larger, since it further depends on other problem-related quantities as mentioned above.

    • We add dimension-dependent degradation tests for the Quantile Transformer and M2M-W2W_2 of [3] on the Gaussianity-preserving McKean Vlasov dynamics (see revised Fig. 3, Sec. 5 and App. H.8). Specifically, we update our experiments with new results for d=6d=6 and M=32,500M=32, 500 (besides the prior d=3d=3, M=128M=128). Due to lack of space, the (d=3,M=500)(d=3, M=500) and (d=6,M=500)(d=6, M=500) results are left to the appendix section. Fig. 3 (e-f) depicts the average test W2W_2 normalized by d\sqrt{d} to account for dimensional scaling. Reducing MM from 128128 to 3232 increases mean normalized error by 24.624.6\\% for QT versus 44.544.5\\% for M2M-W2W_2 at d=3d=3, and by 17.017.0\\% versus 10.110.1\\% at d=6d=6. Increasing MM from 128128 to 500500 decreases QT's error by 12.512.5\\% but increases M2M-W2W_2's by 19.419.4\\% at d=3d=3. At d=6d=6, errors increase by 9.39.3\\% and 15.515.5\\%, respectively. Furthermore, at fixed MM, both methods’ normalized errors increased from d=3d=3 to d=6d=6. QT has a consistent advantage and a smaller variance across all settings. We hypothesize the latter is due to QT avoiding intermediate particle reconstruction (via token-based training), thus preventing particle-level error accumulation during recursive rollout. We also note the two methods' markedly different training times: across the 6 settings, M2M M2M-W2W_2 requires ≈16×\approx 16\times the mean training time of QT for 10510^5 updates on an NVIDIA L4, including periodic validation.

Our new experiments, discussions, and clarifications cover the reviewer's raised concerns in full. If the reviewer deems our response satisfactory, we kindly request reconsideration of the score. Otherwise, we are happy to answer any further questions.

[1] Fraiman et al. (2023), A Cramér–Wold theorem for elliptical distributions.

[2] Fournier & Guillin (2015), On the rate of convergence in Wasserstein distance of the empirical measure.

[3] Vandergrift et al. (2026), Measure-to-measure Regression with Transformers.

Reviewer_FDwhOpenReview ↗

Response to the authors

Thank you for the response. Whilst my questions were addressed in letter, I retain grave concerns about this approach requiring many, many additional samples as it completely ignores any localized interactions between the particles, which are part and parcel of any real world application. However, an intransigent stance would defeat the purpose of a discussion period between authors and reviewers. More importantly, I asked questions and the authors responded to most of them. I shall duly increase my score.

We thank the reviewer for engaging with our response and reconsidering their assessment. We wish to provide further clarification regarding the concern that QT can potentially require "many, many additional samples".

If this concern refers to the number of particles per measure, the exponential-in-dimension limitation applies to every method returning an MM-particle measure (as per our prior example of the uniform law on the unit ball). Furthermore, any method aiming to learn a map between measures is bound to pay for potential ill-conditioning of the target φ\varphi and its interplay with MM.

Separately, if this concerns the number of training sequences per problem instance, particle-based attention (which models localized interactions) could improve learning efficiency for some subset of dynamics. Our discussion identifies this as a hypothesis requiring empirical investigation. This does not remove the computational disadvantage providing the core motivation of our work: a particle-based transformer operating on TT-length sequences of MM-atom empirical measures with dense attention has attention cost O(T2M2d)\mathcal O(T^2M^2d) and requires expensive Optimal Transport computations for training. QT instead attends over TT measure tokens, with the O(T2Md2)\mathcal O(T^2Md^2) forward-pass cost stated in Sec. 4, and trains using cheap Euclidean token losses.

Our QT therefore offers a point of tradeoff, whose value has to be established per application. In particular, our experiment comparing QT with M2M-W2W_2 (Sec. 5, App. H.8) highlights QT's computational advantage in this case: QT achieves lower error and lower variance in all 66 settings when training on the same set of sequences and with the same sample rate MM. Thus, here, QT requires neither additional particles per measure nor additional training trajectories. We provide the full experiment statistics below, which clearly highlights the computational advantage of QT across several dimensions and sample rates.

ddMMM2M time (min)QT time (min)M2M/QT timeM2M peak memory (GiB)QT peak memory (GiB)
33277.8619.583.98×0.3700.266
3128117.3618.416.38×0.6900.409
3500773.2919.9638.74×2.0630.980
63280.5719.584.12×0.3700.388
6128125.0819.366.46×0.6940.903
6500855.1624.0135.62×2.0622.926
Mean—338.2220.1516.79×*——

*Ratio of mean times, computed before rounding. The arithmetic mean of the six individual time ratios is 15.88×. Memory denotes PyTorch's recorded peak allocated GPU memory. The experiments were run on an NVIDIA L4 GPU

Reviewer_Xr8ROpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes the Quantile Transformer, a transformer-based framework for learning sequence-to-sequence maps between probability measures from finite samples. The approximator takes the form D∘T∘ED\circ\mathcal{T}\circ E, where EE and DD are a fixed encoder/decoder pair and T\mathcal{T} is a learned transformer. Training minimizes squared Euclidean error in the encoded representation, exploiting quantile-based Wasserstein geometry and decoder stability to control approximation error in the W2W_2-space. The paper develops a comprehensive theoretical framework, including universal approximation guarantees for one-dimensional and multivariate elliptical measures, and evaluates the method through synthetic experiments.

Critical strengths
  • The theoretical framework is elegantly constructed and supported by substantial, largely sound analysis.
  • The computational implementation is efficient, and the synthetic experiments are informative and honestly reported.
Critical weaknesses
  • The restriction to 1D and multivariate elliptical distributions limits the method's applicability to general machine learning problems.
  • The numerical experiments use only synthetic data and do not examine robustness to nonelliptical multivariate targets, an important form of model misspecification.
  • The numerical results also do not demonstrate a consistent advantage over the simpler Moment Transformer baseline.
Questions for authors
  • Could the authors evaluate robustness to model misspecification, using synthetic experiments with multivariate target measures that are not elliptical?
  • Could the authors include experiments on real-world data? A simple example came to my mind is to apply a temporally structured transformations AtA_t to a real 2D image XX, then predict At+1XA_{t+1}X from A1X,…,AtXA_1X,\dotsc,A_tX, treating each image as a probability measure.
  • I include some minor suggestions here due to lack of suitable space; they do not substantially affect my assessment:
    • There are some formatting inconsistencies, such W2W_2 was typeset in upright sometimes, and a theorem reference on page 5 appears in a footnote rather than in the main text like the rest of the citations.
    • One of the citation (Beranger et al.) contains a strange string after the title: "B. beranger et al."
    • The related-work discussion would benefit from discussing connections and differences with work on predicting transport maps. Some suggested citations:
      • Supervised Training of Conditional Monge Maps (NeurIPS 2022)
      • Meta Optimal Transport (ICML 2023)
      • Embedding Empirical Distributions for Computing Optimal Transport Maps (ISIT 2025)
3 comments
AuthorsOpenReview ↗

Response – Part I: Additional experiments and fixing inconsistency issues

Dear Reviewer Xr8R,

Thank you for your time and constructive comments, which helped us improve our manuscript. We address concerns point by point below. In addition, we summarize manuscript changes and some common points raised by reviewers in a general response. Note that revised manuscript additions are in blue. References are in the final comment.

  1. On robustness to non-ellipticity. We add a new experiment using time-varying 3D Gaussian mixtures with constant mean and covariance, studied at two levels of ellipticity violation (see Sec. 5 and App. H.10). To isolate token prediction error from decoding error, we decoded both population-oracle tokens and QT predictions with an elliptical decoder and a mixture-aware decoder. With mixture-aware decoding, QT nearly matches the true-distribution oracle. With elliptical decoding, even the population-oracle tokens incur increasing error as the departure grows, pointing to decoder misspecification as the main source of excess error in this experiment. QT with elliptical decoding maintains a stable performance for mild-to-moderate ellipticity violations, but significantly degrades in strong cases (see Fig. 39 App. H.10).

  2. On real-world experiments. While we agree that real-world experiments are always a beneficial addition, we wish to emphasize that, as a learning theory submission, our paper's contribution is primarily theoretical. The goal of our experiments is to provide comprehensive preliminary validation to motivate further studies (empirical or theoretical). Indeed, we explicitly leave both the case of general measure classes and real-world experiments as future work in the concluding Sec. 6. We highlight the improved experimental Sec. 5 made possible by the increased page limit and reviewer feedback. The updated experiments are:

    • (not new) One-dimensional measure sequences: linear-evolution Gaussian (App. H.2), nonlinear evolution Gaussian (not new, App. H.3), nonlinear evolution Student-t (note the heavy-tailedness; not new, App. H.5), mixtures with evolving mode separation (not new, App. H.4).
    • (not new) High-dimensional measure sequences of a fixed elliptical class: linear-evolution Gaussian (App. H.6), nonlinear evolution Student-t (note the heavy-tailedness; App. H.7)
    • (new) High-dimensional non-elliptical measure sequences with varying degrees of non-ellipticity: the experiment described in the point just above
    • (new) High-dimensional measure sequences with varying ellipticity class: the Gaussian-to-elliptical sequences described in next comment (Sec. 5, App. H.9)
    • (updated with more settings) Comparison with the particle-based M2M-W2W_2 transformer of [1]: We expand the McKean–Vlasov comparison to cover d∈3,6d\in\\{3,6\\} and M∈32,128,500M\in\\{32,128,500\\}. QT achieves lower mean test W2/dW_2/\sqrt d in all settings, and has smaller variance. Furthermore, across settings, M2M-W2W_2 requires ≈16×\approx 16\times the mean training time of QT for 10510^5 updates, including periodic validation.
  3. Fixed inconsistency issues and added related work. We fixed the inconsistencies and the citation issues raised by the reviewer. Also, we added a brief discussion of the additional literature mentioned by the reviewer—see the blue section in Appendix A.1.

AuthorsOpenReview ↗

Response – Part II: On the comparison between the Quantile and Moment Transformers

  1. On the Quantile Transformer (QT) vs. Moment Transformer (MT) comparison, and when QT should be preferred despite its linear-in-MM token.

    We wish to emphasize that QT's close performance to that of MT is, in fact, a strong empirical result, due to the knowledge asymmetry between the two models.

    • First, Figs. 2 (a)-cc and (g) respectively depict 1D Gaussian and Student-t sequence experiments, where QT operates without any assumption on the distribution class besides inclusion in P2(R)\mathcal P_2(\mathbb R). In contrast, MT knows the distributions are respectively Gaussian and Student-t (this knowledge lies in the token format and the specialized Gaussian/Student-t decoders). Thus, the two models have matching accuracy (overlapping plot lines), despite an enormous information gap regarding the distribution family. This behaviour extends to bimodal mixtures with time-varying degrees of separation (App. H.4); and Gaussian sequences with non-linear mean and variance dynamics (App. H.3).

    • Second, Figs. 2 (d)-ff and (h) respectively depict 5D Gaussian and Student-t experiments, where QT assumes only generic ellipticity (a broad class). In contrast, MT knows the correct parametric class (Gaussian, and Student-t, respectively). Here again, QT matches MT's performance, despite the information asymmetry.

    • Finally, we propose a new experiment emphasizing the shortcoming of MT for sequences with varying radial law (i.e., changing elliptical families). Details are in the revised Sec. 5 and App. H.9. In short, we evolve an initial Gaussian μ1\mu_1 through repeated applications of a non-linear map ff, producing a sequence of non-Gaussian elliptical measures for t≥1t\geq1. Here, the radial law, and hence generator ψ\psi in Def. 4.2, changes at every tt. QT accommodates these changes by construction and remains competitive with population oracles, whereas MT's decoder becomes misspecified after t=1t=1 and performance degrades. Crucially, supplying MT with the true radial law per step would mean providing the very distributional dynamics it is supposed to learn.

    In summary, these experiments highlight a tradeoff: when the parametric class is known for every μt, t∈[T]\mu_t, \, t\in[T], MT is a more economical choice due to its smaller tokens. Otherwise, QT is preferable since it is maximally flexible in 1D (no class assumptions besides inclusion in P2(R)\mathcal P_2(\mathbb R)) and adaptive to the broad class of elliptical distributions in high dimensions.

Our new experiments, discussions, and clarifications cover the reviewer's raised concerns in full. If the reviewer deems our response satisfactory, we kindly request reconsideration of the score. Otherwise, we are happy to answer any further questions.

[1] Vandergrift et al. (2026), Measure-to-measure Regression with Transformers.

I thank the authors for the additional experiments. They clarify when the QT offers an advantage over the MT and how its decoder behaves when ellipticity is violated. These results address my main empirical concerns, and I shall raise my score.

Reviewer_KygSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Quantile Transformer, a transformer-based architecture for learning mappings between sequences of probability distributions represented by finite samples. It represents each empirical distribution as a single finite-dimensional token, rather than treating every sampled particle as an individual token. This reduces computational costs while retaining theoretical universal approximation guarantees.

Critical strengths
  1. The paper establishes universal approximation guarantees for continuous measure-to-measure mappings over compact classes of one-dimensional and multidimensional elliptical distributions.

  2. The proposed quantile-based representation allows a standard Transformer to process each empirical distribution as a single token, avoiding expensive particle-level attention and optimal transport computations during training.

Critical weaknesses

My main concern is the gap between theoretical analysis and experimental results.

  1. The theoretical guarantees require a sufficiently large number of particles M (Theorem 3.1, 4.1), while the proposed representation has a token dimension that grows linearly with M. Increasing M introduces substantial computational and memory costs. The tradeoff between approximation accuracy and computational efficiency is unknown.

  2. In Fig. 2(a–f), compared to the Gaussian Moment Transformer, the Quantile Transformer doesn't show a clear improvement in prediction accuracy. The paper emphasizes practicality, but Fig. 2 does not establish why a representation containing hundreds or thousands of quantile coordinates is preferable to a much smaller representation when both deliver similar prediction accuracy.

  3. Experiments with more complex distributional structures would help demonstrate the effectiveness of the proposed approach.

Questions for authors
  1. How does approximation error scale with the number of particles and the associated computational cost?
4 comments
AuthorsOpenReview ↗

Response – Part I: on token size and computation/approximation tradeoff

Dear reviewer KygS,

Thank you for your time, constructive comments, and clarification questions, which helped us improve our manuscript. We address concerns point by point below. In addition, we summarize manuscript changes and some common points raised by reviewers in a general response. Note that revised manuscript additions are in blue. References are in the final comment.

  1. On the linear-in-MM dependence of token dimension. This kind of dependence is typical of architectures aiming to approximate unknown measures via samples. Note, e.g., the well-established particle-based transformer of [1], where the number of "switches" (i.e., layers) in their continuous-depth architecture is linear in MM (see discussion in [1] Section 1.4.3, point 2). For us, this dependence moves to width, rather than depth. Furthermore, note that the class of functions covered in [1] (measure-to-measure) is strictly included in ours (sequence-of-measures to sequence-of-measures), whenever measures are elliptical and have PD covariance. We include this remark in App. A.2.

  2. On the computation vs. ε\varepsilon trade-off.

    • The formal trade-off between computation and accuracy ε\varepsilon is already quantified by Prop. 3.2, which establishes the (ε,δ)(\varepsilon, \delta)-dependence of a sufficient MM in Thm. 3.1 (i.e., M0M_0), in conjunction with the "Computational complexity" paragraphs of Sec. 3 and 4, which give the MM-dependent complexity of forward passes, encoding and decoding. Training complexity is inherited from standard Euclidean transformers (we train in token space), where the MM-dependence enters through token size.

      Now, we briefly explain how M1M_1 and M2M_2 in Prop. 3.2 relate to ε\varepsilon in Thm. 3.1 (Thm. 4.1 follows similarly, with an additional ε\varepsilon-dependence from the decoder). Thm. 3.1 starts from the good event defined in Eq. (14) of the main paper,\hspace{70mm} A_{M,\alpha}(\mathbf{\mu}) = \left\\{\max_{t\in[T]}W_2(\widehat\mu_{t,M},\mu_t)\leq\alpha\right\\}. We let Mgood(α,δ)M_{\mathrm{good}}(\alpha,\delta) be a threshold s.t. ∀M≥Mgood(α,δ)\forall M\geq M_{\mathrm{good}}(\alpha,\delta), $\,\,\sup_{\mathbf{\mu}\in \mathcal K^T} \mathbb P_{\mathbf{\mu}}\left(A_{M,\alpha}(\mathbf{\mu})^C\right) \leq\delta. $ Thus, sampling MgoodM_{\mathrm{good}} particles ensures the good event holds w.p. at least 1−δ1-\delta for every admissible input.

      Now, MgoodM_{\mathrm{good}} depends on ε\varepsilon through α\alpha. On the good event, QT's W2W_2 error is bounded by the sum of i) Tθ\mathcal T_\theta's Euclidean error, ii) the quantile discretization error, and iii) ωφ~(4α)\omega_{\widetilde\varphi}(4\alpha), where ωφ~\omega_{\widetilde\varphi} is the continuity modulus of target φ~\widetilde \varphi, in Sec. 2.1, Step 1.

      We choose i) to be <ε/3< \varepsilon/3 via Cor. 2.1. (indep. of MM, see Eq. 60). For ii) we need an Mproj(ε)M_{\mathrm{proj}}(\varepsilon) large enough s.t. the quantile discretization error is <ε/3<\varepsilon/3 (Eq.66). For iii) we need α\alpha s.t. ωφ~(4α)≤ϵ/3\omega_{\widetilde\varphi}(4\alpha)\leq\epsilon/3 (Eq. 67). Thus, the threshold M0M_0 in Thm. 3.1 is M0=max⁡Mgood(α,δ),Mproj(ϵ)M_0=\max\\{M_{\mathrm{good}}(\alpha,\delta),M_{\mathrm{proj}}(\epsilon)\\}. In Prop. 3.2, M1=Mgood(α,δ)M_1 = M_{\mathrm{good}}(\alpha,\delta) and M2=Mproj(ϵ)M_2 = M_{\mathrm{proj}}(\epsilon).

      Note that, for an arbitrary continuous φ\varphi, Mgood(α,δ)M_{\mathrm{good}}(\alpha,\delta) can be arbitrarily large since ωφ~\omega_{\widetilde\varphi} can decay arbitrarily slowly. However, under stringent assumptions like Lipschitzness (point cc of Prop. 3.2), this dependence can be made explicit. E.g., if quantiles are uniformly LL-Lipschitz and φ\varphi is HH-Lipschitz, then ωφ~(4α)≤4Hα\omega_{\widetilde\varphi}(4\alpha)\leq4H\alpha. Taking α=ϵ/(12H)\alpha=\epsilon/(12H) gives Mgood∈O(LH/ϵ)2ℓM_{\mathrm{good}}\in\mathcal O (LH/\epsilon)^2\ell and Mproj∈O(L/ϵ)M_{\mathrm{proj}} \in \mathcal O (L/\epsilon) for 1D. Thus O(ϵ−2)\mathcal O(\epsilon^{-2}) particles per measure suffice at fixed T,δ,L,HT,\delta,L,H.

    • We also highlight a practical tradeoff. The linear-in-MM token marks our choice to store all MM particle projections, for a high-resolution approximation of the quantile function. Using Eq. (10) in the manuscript, we can instead divide the quantile domain into B:=B(M)<MB := B(M) < M equal-probability bins and store the avg. quantile value per bin. Thus, we still use MM particles to estimate quantiles, but at a lower resolution, and hence a smaller token. Of course, we still require lim⁡M→∞B(M)=∞\lim_{M\to\infty} B(M) = \infty, since we approximate unknown measures from samples. At fixed MM however, this gives a practical tradeoff between quantile approximation accuracy and token size. The arguments in Thms. 3.1 and 4.1 extend here since binning error vanishes uniformly over the same compact sets as M→∞M\to\infty. Thus, token width can grow sublinearly in MM and retain the asymptotic guarantee. We include this remark in App. A.2.

AuthorsOpenReview ↗

Response – Part II: on Quantile vs. Moment Transformer comparison, and experiments with more challenging distributions

  1. On the Quantile Transformer (QT) vs. Moment Transformer (MT) comparison, and when QT should be preferred despite its linear-in-MM token. We wish to emphasize that QT's close performance to that of MT is, in fact, a strong empirical result, due to the knowledge asymmetry between the two models.

    • First, Figs. 2 (a)-cc and (g) respectively depict 1D Gaussian and Student-t sequence experiments, where QT operates without any assumption on the distribution class besides inclusion in P2(R)\mathcal P_2(\mathbb R). In contrast, MT knows the distributions are respectively Gaussian and Student-t (this knowledge lies in the token format and the specialized Gaussian/Student-t decoders). Thus, the two models have matching accuracy (overlapping plot lines), despite an enormous information gap regarding the distribution family. This behaviour extends to bimodal mixtures with time-varying degrees of separation (App. H.4); and Gaussian sequences with non-linear mean and variance dynamics (App. H.3).

    • Second, Figs. 2 (d)-ff and (h) respectively depict 5D Gaussian and Student-t experiments, where QT assumes only generic ellipticity (a broad class). In contrast, MT knows the correct parametric class (Gaussian, and Student-t, respectively). Here again, QT matches MT's performance, despite the information asymmetry.

    • Finally, we propose a new experiment emphasizing the shortcoming of MT for sequences with varying radial law (i.e., changing elliptical families). Details are in the revised Sec. 5 and App. H.9. In short, we evolve an initial Gaussian μ1\mu_1 through repeated applications of a non-linear map ff, producing a sequence of non-Gaussian elliptical measures for t≥1t\geq1. Here, the radial law, and hence generator ψ\psi in Def. 4.2, changes at every tt. QT accommodates these changes by construction and remains competitive with population oracles, whereas MT's decoder becomes misspecified after t=1t=1 and performance degrades. Crucially, supplying MT with the true radial law per step would mean providing the very distributional dynamics it is supposed to learn.

    In summary, these experiments highlight a tradeoff: when the parametric class is known for every μt, t∈[T]\mu_t, \, t\in[T], MT is a more economical choice due to its smaller tokens. Otherwise, QT is preferable since it is maximally flexible in 1D (no class assumptions besides P2(R)\mathcal P_2(\mathbb R)) and adaptive to the broad class of elliptical distributions in high dimensions.

  2. On including experiments with more complex distributional structures. We used the extra page allowance to expand the experimental section. Now, Sec. 5 discusses all experiments previously isolated in the appendix, and introduces new ones. In short, we now cover:

    • (not new) One-dimensional measure sequences: linear-evolution Gaussian (Sec. 5, App. H.2), nonlinear evolution Gaussian (App. H.3), nonlinear evolution Student-t (note the heavy-tailedness; Sec. 5, App. H.5), mixtures with evolving mode separation (App. H.4).
    • (not new) High-dimensional measure sequences of a fixed elliptical class: linear-evolution Gaussian (Sec. 5, App. H.6), nonlinear evolution Student-t (note the heavy-tailedness; Sec. 5, App. H.7)
    • (new) High-dimensional measure sequences with varying ellipticity class: the Gaussian-to-elliptical sequences described above (Sec. 5, App. H.9)
    • (new) High-dimensional non-elliptical measure sequences with varying degrees of non-ellipticity: we consider time-varying 3D Gaussian mixtures with constant mean (03\mathbf{0}_3) and covariance (I3\mathbf{I}_3) and different levels of ellipticity violation (mode separation) --- see Sec. 5, App. H.10. We isolate the token prediction errors from the decoding errors by using a mixture-aware decoder in addition to the elliptical one. We observe that QT is resilient to mild-to-moderate levels of non-ellipticity, and errors mostly stem from a misspecified decoder rather than an inaccurate quantile token representation.
    • (not new, but updated) Comparison with the particle-based M2M-W2W_2 transformer of [2]: We expand the McKean–Vlasov comparison to cover d∈3,6d\in\\{3,6\\} and M∈32,128,500M\in\\{32,128,500\\} (Sec. 5 and App. H.8). QT achieves lower mean test W2/dW_2/\sqrt d in all settings, and has smaller variance. Furthermore, across settings, M2M-W2W_2 requires ≈16×\approx 16\times the mean training time of QT for 10510^5 updates, including periodic validation.

Our new experiments, discussions, and clarifications cover the reviewer's raised concerns in full. If the reviewer deems our response satisfactory, we kindly request reconsideration of the score. Otherwise, we are happy to answer any further questions.

[1] Geshkovski et al. (2026), Measure-to-measure interpolation using transformers.

[2] Vandergrift et al. (2026), Measure-to-measure Regression with Transformers.

Thank you for the detailed response. It has addressed most of my concerns. Although I still have some reservations about the practical use of the proposed method, I acknowledge its theoretical contribution and will raise my score.

Thank you for taking the time to evaluate our response, for the positive feedback, and for helping us improve the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary and new revision

Dear Area Chair and Reviewers,

We thank the reviewers for their constructive feedback and their time and patience in reviewing our manuscript. Our manuscript significantly improved the experiments in Sec. 5, largely by using the additional page to propose new settings and discuss the implications of the existing ones in more detail (all changes are highlighted in blue). We summarize the major changes below.

  1. New experiments on elliptical sequences with a time-varying radial law (revised Sec. 5, App. H.9), which emphasize the advantage of our Quantile Transformer (QT) over the Moment Transformer (MT) baseline. Here, an initially Gaussian distribution evolves into a non-Gaussian elliptical sequence of subsequent measures. The distribution-agnostic QT remains competitive with distribution-aware population oracles, while MT’s inflexible Gaussian decoder causes its predictions to progressively degrade. Crucially, supplying MT with the true radial law per step would mean providing the very dynamics it is supposed to learn.
  2. New experiments showcasing the Quantile Transformer's (QT) robustness to violations of ellipticity (revised Sec. 5, App. H.10). Here, we consider 3D Gaussian mixtures with fixed mean/covariance at two levels of departure from ellipticity (mode separation). To isolate token prediction error from decoding error, we decoded both population-oracle tokens and QT predictions with an elliptical decoder and a mixture-aware decoder. With mixture-aware decoding, QT nearly matches the true-distribution oracle. With elliptical decoding, even the population-oracle tokens incur increasing error as the departure grows, pointing to decoder misspecification as the main source of excess error in this experiment. QT with elliptical decoding maintains a stable performance for mild-to-moderate ellipticity violations, but significantly degrades in strong cases (see Fig. 39 App. H.10)
  3. New experiments on dimension and particle count scaling relative to M2M-W2W_2 [1] (revised Sec. 5, App. H.8). We expand the McKean–Vlasov comparison to cover d∈3,6d\in\\{3,6\\} and M∈32,128,500M\in\\{32,128,500\\}. QT achieves lower mean test W2/dW_2/\sqrt d in all settings, and has smaller variance. Furthermore, across settings, M2M-W2W_2 requires ≈16×\approx 16\times the mean training time of QT for 10510^5 updates, including periodic validation.
  4. New discussion on our linear-in-MM token size, the absence of particle interactions, and further related work (App. A.1–A.2). First, we remark on the linear-in-MM dependence of other measure-based architectures, like the particle transformer of [2]. We further identify a practical tradeoff between token size and quantile approximation accuracy: averaging quantiles within B(M)<MB(M) < M bins allows tokens to grow sublinearly in MM as long as lim⁡M→∞B(M)=∞\lim_{M\to\infty}B(M) = \infty. We additionally hypothesize on how the absence of local particle interactions might affect learning efficiency compared to particle-based architectures. Finally, we improve coverage of related work.

In closing, we wish to briefly revisit several recurring points raised in this review round:

  1. Scope of our work. Theorems 3.1 and 4.1 establish universality on W2W_2-compact classes in P2(R)\mathcal P_2(\mathbb R) and on W2W_2-compact classes of multivariate elliptical laws with PD covariance. The general case lies beyond the scope of our present paper, and is explicitly left as future work in Sec. 6.
  2. The significance of matching performance between the Quantile and Moment Transformers (QT vs. MT). QT matching MT's performance on fixed-class benchmarks is, in fact, a strong empirical result, due to the knowledge asymmetry of the two models. MT’s decoder knows the parametric family. QT only assumes inclusion in P2(R)\mathcal P_2(\mathbb R) in 1D, and generic ellipticity with PD covariance additionally for higher dimensions. When the radial law changes over time (App. H.9), MT’s fixed decoder significantly underperforms. Here, the radial law (parametric family) gives the very dynamics MT is asked to learn.
  3. The class of functions covered by us vs. related literature. Related particle-based architectures [1,2] address the narrower setting of measure-to-measure prediction, where the learning target φ\varphi processes a single input measure. In contrast, our targets represent sequence-of-measures to sequence-of-measures maps. Whenever [2]'s assumptions align with the conditions of Thm. 4.1, our setting strictly includes theirs. In particular, for time-changing sequences, their setting covers only order-1 autoregressions, while ours can consider a TT-length history.

We hope the new revision clarifies the scope, advantages, and limitations of our work. We are happy to address any remaining questions.

Best wishes,

The authors

[1] Vandergrift et al. (2026), Measure-to-measure Regression with Transformers.

[2]Geshkovski et al. (2026), Measure-to-measure interpolation using transformers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.