Markovian Sliced Wasserstein Distances: Beyond Independent Projections

Sliced Wasserstein (SW) distance suffers from redundant projections due to independent uniform random projecting directions. To partially overcome the issue, max K sliced Wasserstein (Max-K-SW) distance ($K\geq 1$), seeks the best discriminative orthogonal projecting directions. Despite being able to reduce the number of projections, the metricity of Max-K-SW cannot be guaranteed in practice due to the non-optimality of the optimization. Moreover, the orthogonality constraint is also computationally expensive and might not be effective. To address the problem, we introduce a new family of SW distances, named Markovian sliced Wasserstein (MSW) distance, which imposes a first-order Markov structure on projecting directions. We discuss various members of MSW by specifying the Markov structure including the prior distribution, the transition distribution, and the burning and thinning technique. Moreover, we investigate the theoretical properties of MSW including topological properties (metricity, weak convergence, and connection to other distances), statistical properties (sample complexity, and Monte Carlo estimation error), and computational properties (computational complexity and memory complexity). Finally, we compare MSW distances with previous SW variants in various applications such as gradient flows, color transfer, and deep generative modeling to demonstrate the favorable performance of MSW.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer YkFo7/10 · confidence 4/52023-06-19

Summary

The authors focus on the study of sliced Wasserstein distances. It consists in projecting measures onto Euclidean lines, such that OT between these projected measures has low complexity because it amounts to sort samples. The projections are unit vectors of the d-dimensional sphere of R^d, and in practice one samples a set of projections uniformly samples over the unit sphere. However, this might be sub-efficient since some projections might be redundant by sampling close directions, or uniformative by projecting in directions where measures are similar. To solve this issue, the authors propose to sample projections non-uniformly. In particular they consider sampling processus satisfying a first-order Markov property. They propose a sliced OT variant called 'Markovian Sliced Wasserstein'. They prove metric, weak* convergence and statistical properties. They propose two families of Markov processes, focusing on orthogonality or data-driven. They propose a 'burning and thinning' variant to reduce the number of projections and improve complexity. They run experiments of gradient flows, color transfer and generative learning to illustrate the properties of their losses.

Strengths

- I believe that the directions tackled by the authors is interesting, in that they propose ways to change the sampling process of projections, with the hope that it improves the properties of the discretized sliced divergences in practice. - The contributions are thorough, with both methodological, theoretical and numerical content on their proposed sliced OT variant.

Weaknesses

Major issues: - To my mind, the use of a Markov chain does not seem to bring a significant improvement. Concerning Table 1, after running the gradient flow, all W2 distances are of order of $10^{-5}$, whatever the choosen loss. The running times are also equivalent up to a factor $2$. In table 2, the FID and IS scores are not significantly different when we take into account the variance of scores (which is nice to have included it). My take-home message of these experiments is the negative result (as an open question to prove rigorously) that there exists no sampling process which yields a significantly better error rate of the Monte-Carlo approximation. However, I acknowledge that exploring this question of sampling is an interesting, and that even absence of significant improvement is a beneficial insight to the ML community. - Section (3.3) on 'burning and thinning' processes is not sufficiently motivated. First, is it a standard approach in MCMC sampling ? If so, could the authors provide references in the main body about this ? Second, what are the main advantages of doing this ? There should be a key property highlighting the benefits in the main body, otherwise it looks like a variant with no main benefit. Except the reduced computational complexity, nothing else is discussed. If the sampling error rates scale the same as before, I am not motivated by using a more elaborate variant, and would personnally prefer to use a simple yet straightforward algorithm. Numerical experiment also do not support a significant advantage of using these techniques. - I have an issue with the proof of Proposition 2. In the proof there is a term $\log (8 / \delta)$ which somehow disappears when the 'universal constant' C is defined. However, this term seem not to converge to zero when the number of samples goes to infinity. Could the authors provide more details on this proof ? With all due respect to the authors, I am not completely sure of the rigor of the proof yielding this statistical rate. Minor remarks: - Line 51: Do we only use one projection to optimize Max-SW via gradient ascent ? It seems to be a terrible idea, probe to fall into local maxima. Also why not solving sliced Wasserstein for a set of random projections and then keep the projection yielding the largest score ? - Line 62: Do we really need in practice to have a loss which is a distance ? In practice training a model by minimizing a loss which is only definite (e.g. Kullback-Leibler) seems to be suffice. - Line 222: From what I understand, L is the number of Monte-Carlo chains, and T is the number of projections for a given MC chain. Thus the number of total projections is $T \times L$. - Line 251: Maybe a notation 'Proj' and not 'Prod' would be more relevant to denote a projection operation. - Line 263: How do you tune $\kappa$ in the Input-Aware sampling ? Do we need to scale it w.r.t. loss, data, or something else ? Some insights would be welcome here. - In general figures have legends which are too small to read without zooming significantly. Could the authors increase their size ? - Line 343: You could comment that this remark on having a larger T is better by connecting it to Proposition 3, since the error scales proportionnaly to $1 / T \sqrt{L}$. Theory indeed show it is more efficient to take T large rather than L.

Questions

See my questions which are detailed above.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

They adressed the societal impact of their work.

Reviewer 72x26/10 · confidence 4/52023-07-03

Summary

This paper introduces Markovian Sliced-Wasserstein distances, which imposes a Markovian structure on the projecting directions. The author impose different distributions on the transition kernel.

Strengths

- The paper is well written and easy to follow. The idea is well-presented. - Each important theoretical quantity is studied (metricity, sample complexity, MC approximation error, computational complexity). - Different choices of transition kernels are proposed and studied.

Weaknesses

- I think experiments are not that convincing for people to adopt this method. For instance, Figure 3 graphs and Table 2 results are not enough to show that this method is better than the other ones (see standard deviations for instance). - I think this is the type of paper that iterates in the study of SW projections, with a standard theoretical study, good experiments, well-written, but that may not be groundbreaking.

Questions

- The experiments in generative modeling are not very convincing, would you have the time to do some other experiments? I think it would benefit the paper by giving more impact to the paper. - One of the best method for smart SW projections is Distributional SW. Why did the authors did not compare their method with it? - I am waiting to see other reviewers questions, to see if it raises other questions on my side. I am willing to discuss and modify my score in accordance if the authors address my points.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

No need to discuss societal impact here.

Reviewer yZut7/10 · confidence 5/52023-07-03

Summary

This paper presents a novel variant of the sliced Wasserstein (SW) distance, addressing the challenges of redundancy and loss of metricity in other SW distance variants, such as Max-K-SW. The proposed method introduces the use of Markov chain-based projection directions selection to overcome these issues. The theoretical contributions of the paper include proving that the new variant of SW distance is a metric, establishing its ability to induce weak convergence. Additionally, sample complexity bounds are derived, demonstrating that the proposed distance, called MSW, does not suffer from the curse of dimensionality. The authors also analyze the Monte Carlo approximation error for computing the distance. Empirical evaluations are conducted to validate the performance of MSW compared to other SW distance variants. The results demonstrate the superiority of MSW in various tasks, showcasing its effectiveness and advantages in practical applications.

Strengths

- The paper presents a novel and intriguing idea of utilizing a Markov chain to select projection directions. This approach offers a fresh perspective on addressing the challenges of redundancy and loss of metricity in other variants of the sliced Wasserstein (SW) distance. The novelty and innovation of this idea contribute to the strength of the paper. - The authors showcase a commendable strength in their thorough theoretical analysis, covering multiple aspects such as topology and statistics. This comprehensive analysis ensures that the proposed method is well-founded and supported by solid theoretical underpinnings. - The experimental findings presented in the paper are convincing and demonstrate the superior performance of the newly introduced method compared to other variants of the SW distance. However, I have some questions regarding these experiments and see below.

Weaknesses

One potential weakness of the paper lies in the claim that Max-K-SW suffers from a loss of metricity when not obtaining the optimal projections. While this issue is acknowledged, it would be beneficial for the authors to emphasize how their proposed method, even with the use of Monte Carlo approximation, mitigates or eliminates this problem.

Questions

- It would be valuable if the authors could provide further explanation regarding why the newly introduced method outperforms other variants of the sliced Wasserstein (SW) distance empirically. Considering Proposition 1, which suggests that Max-SW is a better proxy for the Wasserstein distance, is it due to the errors introduced in computing the optimal projection direction? Additional insights into the specific factors that contribute to the superior performance would enhance the understanding of the proposed method's advantages. typos - line 593: "We samples" should be "We sample" - line 703: in the limit I assume that you mean $\mu_{\phi(k)}$ instead of $\mu_k$. - line 704: in the limit, $(\mu_k,\mu)$ shouldn't be in subscript. In general the subsequence notation is weird in this proof. For example, $\phi$ appears twice in the proof with different meanings. Please fix this. - line 727: do you mean $\Theta$ is the compact support of $\mu$? - line 735: it's better to write something like $\mathcal{B}_\theta$ to make the set depending on $\theta$ explicitly.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Reviewer AyBm6/10 · confidence 3/52023-07-04

Summary

The authors propose a variant of the sliced Wasserstein distance that samples the projection directions from a Markov chain instead of uniformly from the hypersphere. They propose two Markov chain designs: the first chooses the next projection vector to be orthogonal to the previous one and the second is deterministic and basically performs projected gradient ascent on the projected Wasserstein distance.

Strengths

- The authors analyze their proposed method theoretically in various ways - Multiple experiments compare the proposed method against baselines

Weaknesses

1. In high dimensions, two random vectors will be almost orthogonal to each other with high probability. Therefore, oMSW should be basically equivalent to SW which is confirmed by Table 2 and Table 4 in the appendix (compare SW(L=45) to oMSW(L=3,T=15), so that they use the same number of samples; both methods take roughly the same time and lead to almost the same Wasserstein-2 distance) 1. Table 2: The majority of results overlap so much (i.e. many are within one standard deviation of each other) that no conclusions can be drawn. In particular, putting any of the numbers in bold is misleading. 1. oMSW is only evaluated in Table 2 (which does not really show anything, see previous point) 1. Burned-thinned MSW is introduced as a performance enhancement but (1) its impact only evaluated in the appendix (2) its impact on performance is difficult to understand from Table 3 and it looks like there is no real benefit to it 1. Paragraph "Stationarity of $\sigma_T(\theta_T)" is has neither positive nor negative results and is vague overall

Questions

1. How do you choose the parameters of the vMF distribution? Why would there be any preference regarding the first projection direction? 1. Why would the stationary distribution of the $\sigma_T(\theta_T)$ for uniform initialization and orthogonal chain not be the uniform distribution? 1. If orthogonality of directions is a desirable property, shouldn't oMSW always be worse than K-SW for $T = K$?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors did not discuss any limitations of their method.

Area Chair gcWm2023-08-11

Discussions

Dear reviewers and authors, Thank you very much for your work on this submission and its evaluation. Now that the authors have responded to the reviews, I **strongly encourage** the reviewers to acknowledge the review, to look at other reviews and rebuttals for this submission, and to adjust their scores if needed. Thanks to those that have already done so. Authors have the possibility to reply if further questions are needed, until the 16th. Thank you very much to all, Area Chair

Reviewer 72x22023-08-13

Answer to rebuttal

I would like to thank the authors for their rebuttal. As the experiments with DSW are included in the paper, I am increasing my score by 1. Overall I think that this is a good paper, well written with a good set of experiments. I am still not 100% convinced in the adoption of the method in practice, but I think that the direction of dependent projections is worth exploring.

Authorsrebuttal2023-08-13

Response to Reviewer

Thank you for your response and for increasing the score to 6. We deeply appreciate the comment "the direction of dependent projections is worth exploring.". Overall, the main goal of MSW is to improve the performance of SW in applications. Should the reviewer have any specific concerns within practical contexts, we would be delighted to engage in a more detailed discussion about the adoption of MSW. Best regards,

Reviewer AyBm2023-08-13

Thank you for the clarifications. I have raised my score accordingly.

Authorsrebuttal2023-08-13

Response to Reviewer

Thank you for raising the score to 6. Please feel free to ask if the reviewer still has questions. We are happy to make further discussion. Best regards,

Reviewer YkFo2023-08-15

Response

I thank the authors for their rebuttal. The proof seems to the best of my understanding rigorous. I also join the opinion of Reviewer 72x2 that the performance improvement of sampling dependent projections is entirely convincing, but proposing such idea might inspire new directions of research for the community. For these reasons I decide to increase my score.

Authorsrebuttal2023-08-15

Response to Reviewer

Thank you for raising the score to 7. We really appreciate your questions and constructive feedback which helps us to improve the paper. Please feel free to ask if you still have other questions. Best regards, Authors,

Reviewer yZut2023-08-16

Thank you for addressing my questions. I would like to clarify a little of my question. I was referring to Proposition 1 (i) in the manuscript which is talking about some inequalities related different metrics instead of sampling complexity mentioned in proposition 2. \ From this proposition, it seems that Max-SW is theoretically closer to the Wasserstein distance, suggesting it might be a more accurate theoretical proxy. While I acknowledge the authors' arguments regarding the computational challenges of Max-SW, I'm curious if there are any practical experiments that demonstrate that the inequality MSW<Max-SW does not hold consistently in practice. Showcasing such practical limitations would further substantiate the superiority of the proposed method over Max-SW.

Authorsrebuttal2023-08-16

Response to Reviewer

Thank you for a very interesting suggestion. This is the type of experiment that we did not have in our paper. We have calculated the values of iMSW and Max-SW between two point clouds as in the rebuttal setting. We use the same setting as in the rebuttal, and we get the maximal value of Max-SW is 12.7 and the maximal value of iMSW is 23.2. Therefore, it seems that iMSW reaches a "better" projecting direction than the Max-SW by using two Markov chains. Nevertheless, despite the fact that Max-SW seems to be a better proxy for Wasserstein distance in theory, Max-SW might not be robust in the setting of noisy and contaminated distribution, which might happen in large datasets e.g., CIFAR10, CelebA. In this setup, the finding of the max projecting direction could be affected significantly by the noise. Overall, we will add this report and the discussion to our revision. We are happy to discuss this further if the reviewer is still not satisfied with the answer. Also, please feel free to ask if the reviewer still has questions. Best regards,

Reviewer yZut2023-08-19

I'm satisfied with the answer and I would like to maintain my supportive score for the paper.

Authorsrebuttal2023-08-19

Response to Reviewer

We would like to thank the reviewer for the supportive score. We will include all the discussion in the revision. Best regards, Authors,

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC