acceptodds
Under review as a conference paper at ICLR 2027

Sliced-Regularized Optimal Transport

Abstract

We propose a new regularized optimal transport (OT) formulation, termed sliced-regularized optimal transport (SROT). Unlike entropic OT (EOT), which regularizes the transport plan toward an independent coupling, SROT regularizes it toward a smoothened sliced OT (SOT) plan. To the best of our knowledge, SROT is the first approach to leverage a version of SOT plan as a reference to improve classical OT. We provide a formal definition of SROT, derive its dual formulation, and provide a post-Bayesian interpretation of SROT. We then develop a Sinkhorn-style algorithm for efficient computation, retaining the same scalability advantages as EOT. By incorporating a scalable SOT plan as a prior, SROT yields more accurate approximations of the optimal transport plan than EOT under the same level of regularization. Moreover, the resulting transport plan improves upon the reference SOT plan itself. We further introduce the corresponding OT divergence induced by SROT, named SROT divergence, and analyze its topological and computational properties. Finally, we validate our approach through experiments on synthetic datasets and color transfer tasks, demonstrating that SROT is better than both EOT and SOT in approximating exact OT. Additional experiments on gradient flows further highlight the advantages of SROT divergence.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Optimal Transport, Sliced Optimal Transport

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 2 reviews, above 27% of the others.
Reviewer_VbzGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper capitalizes on recent advances in sliced Optimal Transport plans and proposes a novel regularized optimal transport problem, whose regularization term is the KL divergence between the coupling and the sliced Optimal Transport plan. The computation integrates well into the Sinkhorn solver. Experiments show that it is closer to the exact Optimal Transport plan compared to the entropic Optimal Transport plan.

Critical strengths
  1. This is a natural extension of entropic optimal transport plan, and it sits well at the intersection of sliced OT and entropic OT.
  2. The methods are written very clearly and are easy to follow.
  3. The experiments show consistent improvement in OT plan fidelity over the entropic OT plan at the same regularization level.
Critical weaknesses
  1. Theorem 2 is not well justified; see Question 2.

  2. Experiments are small-scale, in which exact OT is cheap and fast to compute. This does not match the paper's underlying motivation that exact OT is a computational bottleneck. Experiments at larger sample sizes and in higher dimensions would be more convincing.

Minor:

  • Lines 264–265: the adjusted ground cost C′=C−εlog⁡PSOTC' = C - \varepsilon \log P^{\mathrm{SOT}} decreases, rather than increases, the cost of pairs with high probability under the SOT plan.
Questions for authors
  1. In Figure 2, the SOT plot for the bottom row seems to be wrong. I do not expect the uniform slicing distribution to give such a clean radial matching.
  2. In its current form, the proof of Theorem 2 does not establish the weak continuity or strict convexity of FεF_\varepsilon: πSOT\pi^{\mathrm{SOT}} depends on μ\mu through 1D matching and disintegration, which breaks the argument of Feydy et al. (2019) and is not weakly continuous.
  3. Could the authors state precisely in what sense the SROT divergence recovers "a version of" MMD as ε→∞\varepsilon \rightarrow \infty?

[1] Feydy et al. Interpolating between optimal transport and mmd using sinkhorn divergence. AISTATS 2019.

Reviewer_8RScOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Sliced-Regularized optimal transport (SROT), which is a regularized OT formulation that replaces independent coupling that is typically used in entropic optimal transport with a reference coupling constructed using sliced OT (SOT). The main idea is that the SOT coupling is an better approximation of the optimal transport plan and therefore represents a better regularization prior. The authors derive dual formulations and Sinkhorn-type optimization algorithms. They also introduce a SROT divergence and derive properties such as non-negativity, symmetry and and weak convergence.

Critical strengths

It is interesting that the regularization smearing effects seems to be less compared to standard entropic regularization.

The paper derived dual formulation and Sinkhorn-like computational methods.

The assumptions in 3.1. about, e.g., full support

Critical weaknesses

Limited theoretical guarantees on, e.g., approximation quality.

Theorem 2 assumes positive definiteness of a kernel corresponding to the SOT reference plan. However, this depends on the input distributions, and it is unclear whether the assumptions hold generally.

The simulations are quite limited and only considers small examples, and in particular the simulations considering the computational speed.

Questions for authors

In which cases can do one obtain a measure πSOT\pi^{SOT} which has support covering the product of the supports of μ\mu and ν\nu?

How expensive are are the preprocessing of πSOT\pi^{SOT} when the number of data points are large?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.