acceptodds
Under review as a conference paper at ICLR 2027

Relative Wasserstein Angle and the problem of -Nearest Gaussian Distribution

Abstract

Understanding the distributional structure of high-dimensional datasets has become an important topic, yet direct visual characterization is difficult. In this work, we develop a geometric framework for characterizing the distributional structure of empirical datasets by quantifying their deviation from the Gaussian family under the geometry induced by optimal transport theory. Building on the cone structure of the relative translation invariant quadratic Wasserstein space, we define two geometric quantities—the relative Wasserstein angle and the orthogonal projection distance—and show that they are well-defined because of the flat geometry of the filling cone between distributional rays. This formulation recasts the problem of measuring deviation from the Gaussian family as an orthogonal projection problem onto the Gaussian cone and reveals that the commonly used moment-matching Gaussian is, in general, not the -nearest Gaussian to a non-Gaussian distribution. In one dimension, we derive closed-form expressions for the proposed quantities and extend closed-form expressions to several other location–scale families, including uniform, Laplace, and logistic distributions. In higher dimensions, we develop a numerical approximation method for the proposed quantities based on empirical optimal transport and covariance-shape optimization. Our experimental results show the empirical convergence and stability of the proposed methods and reveal that the angle provides a robust and consistent measure of distributional non-Gaussianity. Moreover, these results provide empirical support for its potential use as an indicator of distributional heterogeneity.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
optimal transport, Wasserstein geometry, probability, and geometric ML

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RsNWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the problem of the closest Gaussian to a given measure in Wasserstein geometry.

Critical strengths

I think that the writing is very clear and nice to follow. The goal is clear and interesting. The ideas are nicely illustrated.

Overall, I think that this is nice and well-rounded paper, focussing on relatively narrow, but interesting problem.

Critical weaknesses

The experimental validation is weak, but this is okay from my point of view. This is a theoretical contribution and the experiments sufficiently validate everything.

The scope is very narrow, but I still think that it is an interesting problem.

Appendix H relies on numerical optimization to establish a claim about the globally nearest Gaussian, which does not constitute a rigorous proof. Monte Carlo optimization does not rigorously establish that the globally W₂-nearest Gaussian has different eigenvectors and eigenvalues from the moment-matching Gaussian.

Busemann connection: The RW₂ projection length coincides with the negative Busemann function of Bonet et al. (2026) for radial Wasserstein rays. The novelty of the proposed angle and projection is therefore unclear. ("Busemann Functions in the Wasserstein Space: Existence, Closed-Forms, and Applications to Slicing", C Bonet, E Cazelles, L Drumetz, N Courty)

Questions for authors

Could one extend the closed-form projection results for general location–scale families to derive explicit Busemann function formulas, as in Bonet et al. (2026, see above)? Could these formulas also be used to construct new Busemann-based slicing procedures? Being able to derive slicing algorithms would make the insights of the paper more interesting for practical applications.

Minor remarks

  • Gap in Theorem 3.2 proof (Appendix B.4): Uniqueness of the optimal transport map does not immediately establish uniqueness of the proposed affine coordinates (a,b)(a,b). Injectivity requires an additional argument.

  • Mismatch of assumptions: Theorem 3.2 assumes absolute continuity of at least one distribution, but the framework is applied to empirical distributions for which this condition need not hold.

  • Unjustified optimization reduction (Section 3.4): Imposing tr⁡(Σ)=1\operatorname{tr}(\Sigma)=1 changes the optimization objective. Equivalence of the optimal covariance shapes after optimizing scale must be explicitly established.

  • Missing convergence guarantees (Algorithm 2): The numerical procedure provides no guarantee of convergence to a globally W₂-nearest Gaussian, despite the terminology used.

Reviewer_c26xOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies optimal Gaussian approximation under the quadratic Wasserstein distance and the quantification of non-Gaussianity for empirical distributions. The authors introduce a relative Wasserstein angle and a projection distance in the space of distributions modulo translations, interpret the nearest-Gaussian problem as projection onto a Gaussian cone, and show that the moment-matching Gaussian is generally not optimal. The paper provides closed-form expressions in one dimension, extensions to other location–scale families, and a covariance optimization algorithm for higher dimensions. Experiments on synthetic data and four image datasets show improvements over the moment-matching baseline and explore the relationship between the proposed angle and distributional heterogeneity.

Critical strengths
  1. The authors proposed a clear problem with an intuitive geometric interpretation. Separating distributional scale from shape helps explain optimal Gaussian approximation and non-Gaussianity, and clarifies the distinction between moment matching and minimizing Wasserstein distance.
  2. The paper combines computable one-dimensional formulas, a high-dimensional approximation algorithm, and experiments on multiple datasets, demonstrating the preliminary feasibility of the framework.
Critical weaknesses
  1. The novelty is not sufficiently distinguished from prior work. Pages 6–7 acknowledge that the one-dimensional angle is a monotone transformation of the existing statistic of del Barrio et al., and therefore introduces no additional ranking or testing information. More importantly, Takatsu and Yokota’s work on the cone structure of Wasserstein spaces already establishes the cone structure and gives an angle formula for rays, but is not cited. The authors should identify precisely what is new in the angle definition and the filling-cone theorem. This distinction directly affects the assessment of the theoretical contribution.
  2. The accuracy and practical advantages of the high-dimensional solver are not sufficiently established. Stabilization of the iteration curves does not establish proximity to an optimal solution. The experiments lack exact or high-accuracy reference solutions, comparisons with alternative solvers, and calibration of finite-sample error. The actual setting uses 32 dimensions after PCA, with 512 data samples and 512 Gaussian samples. Although the angles decrease relative to moment matching, runtime for the full-distribution experiments increases from approximately 0.04 seconds to 3 seconds, leaving the computational benefits unclear.
  3. The application claims extend beyond the available evidence. Non-Gaussianity is not equivalent to heterogeneity, and trends along selected sequences of class mixtures do not establish a general relationship. For example, splitting a standard Gaussian by sign produces two non-Gaussian conditional distributions, while recombining them with equal weights recovers a Gaussian. Likewise, a lower coefficient of variation only indicates lower relative variability under the tested conditions; it does not establish robustness to outliers or contamination. These limitations weaken the claimed practical contribution of the metric.
Questions for authors
  1. Relative to Takatsu and Yokota’s results on cone structure and angles, and existing one-dimensional Wasserstein normality statistics, which propositions provide conclusions that do not follow directly from prior work? In particular, please clarify the novelty of Theorem 3.2.
  2. Could the algorithm be validated against the one-dimensional closed-form solution, the separable distributions in Appendix F, or high-accuracy low-dimensional solutions? Could experiments on known Gaussian distributions, varying sample size and dimension, distinguish optimization error, sampling error, and genuine non-Gaussianity?
  3. How is “heterogeneity” defined? Can it be assessed independently of non-Gaussianity, including cases where combining classes decreases the angle? If a general relationship cannot be established, would the authors narrow the corresponding claims?
Reviewer_LMUnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses Wasserstein angles and projections to quantify non-Gaussianity, derives one-dimensional formulas, and proposes a multivariate covariance-shape optimizer tested on Gaussian mixtures and image representations.

Critical strengths
  1. The geometric account clearly distinguishes moment matching from Wasserstein projection and explains the scope of pairwise flatness.
  2. The paper provides exact solutions in tractable cases and reproducible code. Two MNIST spot checks with larger Gaussian references retain an improvement over moment matching.
Critical weaknesses
  1. The multivariate solver can move away from a known optimum even without sampling error, and finite Gaussian sampling introduces additional bias. Its approximation accuracy therefore needs stronger support.
  2. The geometric novelty remains unclear relative to existing Wasserstein cone geometry [1] and Gaussian goodness-of-fit statistics [2].
  3. The experiments lack systematic validation against known exact solutions. Covariance-shape and angle errors should be measured across dimensions, step sizes, and Gaussian sample sizes.

[1] Cone Structure of L²-Wasserstein Spaces.

[2] Tests of Goodness of Fit Based on the L2-Wasserstein Distance.

Questions for authors
  1. How does the solver's error against exact solutions vary with dimension, step size, and Gaussian sample size?
  2. What is new in the cone/angle construction beyond prior Wasserstein geometry and normality statistics?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.