acceptodds
Under review as a conference paper at ICLR 2027

Riemannian Sliced-Wasserstein Distances on Correlation Matrices

Abstract

Covariance matrices are widely used in Riemannian learning, but they change with marginal signal amplitudes even when the underlying dependence structure remains similar. This paper studies sliced-Wasserstein regularization on full-rank correlation matrices. We consider two geometric branches. The first uses four zero-curvature pullback metrics ECM, LECM, OLM, and LSM, for which the geodesic and Busemann coordinates differ only by sign and CorSW is equivalent to Euclidean sliced-Wasserstein in the corresponding coordinate space. The second branch, PHCML, lifts normalized Cholesky coordinates to the Lorentz model and gives two distinct discrepancies based on Busemann and geodesic projections. We use these discrepancies either alone or together with covariance-based SPD sliced-Wasserstein regularization. Experiments across multiple datasets and network architectures show that CorSW consistently improves downstream performance. For the four flat metrics, caching the embeddings reduces subsequent projections to inner-product operations per sample

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
representative learning, matrix learning, optimal transport, data mining, sliced Wasserstein distance

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_dyEhOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes correlation sliced-Wasserstein (CorSW) on full-rank correlation matrices. It instantiates four pullback-flat geometries (ECM, LECM, OLM, and LSM), constructs Log and Busemann slices on the PHCM (Poly-hyperbolic-Cholesky metrics) product of hyperbolic factors, and introduces a joint Lorentz representation, PHCML-HCat. It distinguishes its closed-form Product-Geodesic surrogate from exact nearest-product-geodesic projection and establishes route-specific metric or pseudometric properties. Validation includes controlled distribution transport, SPD classification, EEG transfer, graph learning, covariance-correlation fusion, and native loss-module timing. The results indicate task-dependent advantages of correlation alignment and complementary behavior of the product and HCat representations.

Critical strengths
  • Removing marginal channel scale is well motivated when amplitude is a nuisance. This paper compares correlation-only and covariance-correlation alignment under shared backbones and seven seeds across several task families. Its RawCorr and Corr-SPD controls help distinguish correlation normalization from the subsequent geometry choice.
  • This paper explicitly distinguishes the original PHCM product geometry from the constant curvature PHCML pullback metric. It also identifies Product Geodesic as a surrogate, derives the exact nearest-geodesic optimization problem, and explains a sufficient condition for coincidence.
Critical weaknesses
  • The significance of the correlation-specific extension needs a sharper case beyond established constructions. The pullback-flat reduction and projection identity follow the framework of Bonet et al. (2025), particularly Proposition 7 and Lemma 25. The Product-Busemann weighted-sum identity is already Proposition 16 of that work, and Lorentz direct concatenation is an existing operation in Bdeir et al. (2024). The authors acknoledge these papers, so the most distinctive components appear to be the PHCML metric choice, the product-geodesic analysis, and the empirical comparison. The paper would be more persuasive if it established what practical insight the new metric provides relative to its strong flat variants.
  • The general-parameter theory contains inconsistencies. In Corollary 3.2, Eq. (7) uses X=log⁡CX=\log C in the OLM inner product, whereas the OLM coordinate is H=off⁡(log⁡C)H=\operatorname{off}(\log C). Dropping the diagonal is harmless for the canonical Frobenius pairing with a hollow direction, but not for the general (α,β,γ)(\alpha,\beta,\gamma) pairing in Eq. (56). For example, take C=diag⁡((1,1/2;1/2,1),I2),(α,β,γ)=(1,0,1),v=E12+E216.C=\operatorname{diag} ( (1,1/2; 1/2,1),I_2), (\alpha,\beta,\gamma)=(1,0,1), v=\frac{E_{12}+E_{21}}{\sqrt{6}}. Here EijE_{ij} denotes a matrix unit and vv is a unit direction for the stated metric. The correct coordinate −⟨H,v⟩(1,0,1)-\langle H,v\rangle_{(1,0,1)} is approximately −1.35-1.35, whereas substituting X=log⁡CX=\log C into Eq. (7) gives −1.11-1.11.
  • I'm not familiar with this area, so I would like to take into account the comments of other reviewers.
Questions for authors

Same as weakness.

Reviewer_VKKkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposed Sliced-Wasserstein (SW) distances for full-rank correlation matrices, termed CorSW. For four flat metrics (ECM, LECM, OLM, LSM), CorSW reduces to Euclidean SW unto an isometry. On the product-hyperbolic Cholesky metric (PHCM), it introduces Product-Log/Busemann/Geodesic coordinates. PHCML joins the rows on one hyperboloid via Lorentz concatenation, with closed-form Log, Busemann and geodesic slices. Experiments are on gradient flow, SPDNet/TSMNet/SPDGCN regularization (nine benchmarks) and EEG alignment.

Critical strengths
  1. Nice motivated focus on correlation geometry. Comparing correlation matrices removes marginal-scale variation and targets dependence structure, making the approach relevant when channel amplitudes vary across subjects or sessions (which in practice happens).
  2. A Unified framework. CorSW brings four flat correlation geometries, product-hyperbolic geometry, and a joint Lorentz representation into a common distribution-comparison framework. This is I think the biggest selling pitch.
  3. Substantive theoretical analysis. The paper derives slicing coordinates, establishes metric properties under stated conditions, and carefully distinguishes its Product-Geodesic construction from nearest-geodesic projection.
  4. Diverse empirical validation. Nine benchmarks spanning SPD classification, EEG transfer, and graph learning demonstrate complementary strengths across the proposed variants. Also the authors showed favorable computational costs in high dimensions.
Critical weaknesses
  1. Novelty is largely incremental. a) The flat case is Euclidean SW after a known isometry, using the framework of Bonet et al. (2025). b) The curved cases instantiate that framework on the PHCM geometry of Chen et al. (2026). c) Hu et al. (2026) already applied OLM/LSM sliced-Wasserstein to EEG.

  2. “Significantly outperform” is a max-over-variants claim. With about 12 variants × 2 modes per setting against one baseline, “at least one variant above both references” is expected under noise. Also there is no significant testing, so very hard to see the "significantly" outperform part. Several margins are tiny: LECM on FPHA is +0.03 over no alignment, and Cora differences are within noise. SPDSW is worse than no alignment on FPHA (61.29 vs 68.99), which suggests the baseline may be poorly tuned.

  3. The motivation (amplitude invariance) isn’t isolated. There is no controlled experiment where marginal variance is perturbed to show SPDSW sensitivity versus CorSW invariance. Gains appear on Radar, HDM05 and graph tasks where amplitude shift is not the issue. Correlation-only EEG alignment often hurts (Table 22: ECM 56.15 vs 77.87 without alignment on BCIC), so the covariance term carries the result, what happens if you remove this term-an ablation maybe?

  4. Some numbers mismatch a) In Table 2, PHCML-B on BCIC-IV-2a is identical to No-align (77.87±3.92). Is the alignment term inactive? b) No-align on BNCI is 81.50 in Table 2 but 78.47 in Table 22.

  5. Protocol asymmetries. An input-scale sweep ({0.25, 0.5, 0.75}, selected on validation) is described only for PHCML-Busemann. Please clarify whether other variants and baselines received equal tuning (hopefully that's the case to have an apple-to-apple comparison, but not clearly mentioned).

  6. Cholesky-based geometries are not permutation-invariant. ECM, LECM, PHCM and PHCML depend on channel ordering, which is arbitrary for EEG, whereas OLM/LSM are invariant. No invariance test is reported.

  7. Cost claims are inconsistent. Table 9 lists O(d²) for CorSW and O(d³) for SPDSW, but the Cholesky/matrix-log maps are O(d³). The “complementary behavior” narrative is post hoc. Different winners per column are expected under noise. There is no guidance on which variant to use.

Questions for authors
  1. Which test supports “significantly”, and how is multiplicity handled? Would a single pre-specified variant still win?
  2. Why does the No-align BNCI numbers differ between tables?
  3. How sensitive are Cholesky-based variants to channel permutation? Something like this I pondered myself, so I have personal interest in this and I think it's an important question to ask.
  4. Can you show a comparison experiment, and compare to Hu et al. (2026)?
  5. Can you compute/report end-to-end costs including the Cholesky/log maps?
Reviewer_Rx5vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper develops correlation-sliced Wasserstein (CorSW) objectives for comparing the distributions of full-rank correlation matrices. The flat branch uses four existing correlation geometries, ECM, LECM, OLM, and LSM, and reduces to Euclidean sliced-Wasserstein after the corresponding isometries. The curved branches either retain the product-hyperbolic Cholesky representation or combine its factors into a single Lorentz hyperboloid through HCat. The paper provides explicit slicing coordinates, distinguishes its Product-Geodesic surrogate from exact nearest-geodesic projection, and establishes metricity for several routes while qualifying the Busemann guarantees through injectivity conditions.

The objectives are evaluated as distribution regularizers and domain-alignment losses across SPD classification, EEG transfer, and graph learning. Additional experiments examine controlled transport, covariance–correlation fusion, correlation-normalized baselines, and loss-evaluation runtime. The results indicate useful but task-dependent benefits.

Critical strengths
  1. A relevant representation-level motivation with a practical alignment formulation. Correlation normalization removes positive channel-wise amplitude scaling while retaining inter-channel correlations. This is a meaningful choice when marginal power differences are nuisance variation rather than task-relevant information. The paper also allows covariance and correlation alignment to be combined, rather than assuming that amplitude information should always be discarded. The ability to introduce these objectives without replacing the existing backbone makes the approach practically relevant.

  2. Explicit constructions supported by broad experimental coverage. The paper gives implementable formulas for several correlation-specific sliced objectives and carefully distinguishes flat, product-hyperbolic, and joint-Lorentz representations. The Klein-coordinate argument for PHCML-Geodesic metricity is particularly clear. Experimentally, the coverage across SPD classification, EEG transfer, and graph learning is valuable, and the RawCorr and Corr-SPD controls help assess whether the results reflect more than just correlation normalization. The runtime study also separates repeated loss-step cost from direction construction and includes the native coordinate maps in the forward pass. Together, these elements make the paper a useful methodological contribution.

Critical weaknesses
  1. Robustness to the regularization needed for empirical correlations is not established. Appendix E.1 specifies a fixed diagonal jitter of 10−610^{-6} before Cholesky or matrix-logarithm operations, but does not evaluate sensitivity to this choice. For a centered sample correlation matrix constructed from nn observations of dd variables with nonzero sample variances, rank⁡(C^)≤min⁡(d,n−1)\operatorname{rank}(\widehat C)\leq \min(d,n-1); consequently, it is necessarily singular when d≥nd\geq n. Even above this threshold, small eigenvalues can make logarithmic coordinates and normalized Cholesky coordinates sensitive to perturbations.

The full-rank restriction is legitimate, and I am not requesting a new theory on rank-deficient strata. The concern is whether the regularization used to obtain admissible empirical inputs materially changes the alignment objective, gradients, or downstream results. Appendix E.2 demonstrates valid transport trajectories near the boundary on Cor+(3)\mathrm{Cor}^{+}(3), but does not test estimation from short observation windows or vary regularization strength. Without such a study, the practical robustness of the proposed objectives remains uncertain in a regime directly relevant to their intended use.

  1. The practical benefit of cross-factor aggregation is not isolated. Section 4.3 and Appendix H.8 correctly show that aggregating factor coordinates before Wasserstein matching can distinguish distributions with identical factor marginals. However, this establishes a representational distinction, not that the mechanism improves downstream performance. The experimental comparisons between PHCM-Product and PHCML-HCat do not isolate this issue, because both aggregate information across factors and also differ in geometry.

A matched comparison between joint Product-Busemann slicing and a sum of factor-wise Busemann losses, using the same backbone, factor representation, training protocol, and comparable projection/computation budget, is needed to substantiate the practical importance of this mechanism. This is consequential because cross-factor aggregation is highlighted as a central contribution, while the basic weighted-sum product-Busemann formula is inherited from the general product-manifold construction of Bonet et al. (2025, Proposition 16). I do not expect a curved geometry to outperform every flat alternative; rather, I would like evidence identifying when this particular design choice provides an advantage.

Questions for authors
  1. Regularization and finite-sample robustness. At what stage is the 10−610^{-6} jitter applied: to the covariance before correlation normalization, to the correlation matrix followed by renormalization, or elsewhere? Please specify any additional shrinkage or eigenvalue flooring. How do downstream accuracy and numerical stability change when varying both the number of observations per matrix relative to its dimension and the regularization strength? A focused study reporting accuracy, minimum eigenvalues or condition numbers, and nonfinite losses/gradients or factorization failures would directly affect my assessment. Here, the observation count refers to the samples used to estimate each matrix, not the number of matrices in a minibatch.

  2. Direct evidence for cross-factor aggregation. Can you compare joint Product-Busemann slicing with the factor-wise objective in Eq. (127), holding the factor coordinates, backbone, optimization protocol, and tuning effort fixed? Please account for the different projection and sorting costs rather than matching only the nominal number of directions. Does the joint objective improve downstream performance in a representative setting, and does it respond as expected in a controlled setting in which factor marginals are held fixed while their dependence changes? This would establish whether the mathematical distinction highlighted in Section 4.3 explains a practical advantage.

  3. Statistical reporting. What statistical test, pairing unit, and significance threshold support the claims that all highlighted CorSW results significantly outperform SPDSW? Please also reconcile the statement in Section 5.1 and Appendix E.1 that methods share seven seeds with Table 16, whose caption reports five seeds for inter-subject PHCM-Product results. Were the relevant comparisons evaluated on matching seeds and held-out subjects? These details matter when interpreting relatively small improvements and comparing the product and HCat branches.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.