acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Sliced-Wasserstein on Symmetric Positive Definite Matrices

Abstract

Comparing distributions of symmetric positive definite (SPD) matrices is a recurring operation in covariance-based learning, but high-dimensional projections can make repeated sliced-Wasserstein evaluations costly. We present SPD hierarchical sliced-Wasserstein (SPDHSW), which computes a small bank of projections in symmetric log-space and mixes their coordinates into many one-dimensional comparisons. Resampling the projection family at each update avoids confining optimization to a fixed subspace. Under Frobenius-uniform sampling, SPDHSW preserves the squared SPDSW population objective. Because the mixed projections share one bank, they are not independent, which prior analysis of hierarchical slicing overlooks; we characterize this dependence through an exact variance decomposition and harmonic analysis, and extend the results to the gradients used in optimization. This establishes separate precision criteria for discrepancy values and training gradients. Experiments on synthetic distributions, hand-action covariances, and EEG adaptation demonstrate improved alignment at a fixed matrix-projection count per update and reduced projection-dominated loop cost relative to large-budget flat slicing under shared optimization settings. Together, the construction and analysis provide a principled approach to allocating projection computation in learning systems that repeatedly compare SPD-valued distributions.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hierarchical Sliced Wasserstein, Sliced Wasserstein Distance, Symmetric Positive Definite Matrices, Riemannian Geometry, 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 3 reviews, above 27% of the others.
Reviewer_UvXzOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SPD Hierarchical Sliced-Wasserstein (SPDHSW) for comparing distributions of symmetric positive definite (SPD) matrices. The core of the method involves mapping SPD matrices to the space of symmetric matrices via the matrix logarithm, then first computing kk bottleneck projections and linearly combining them with LL random mixing vectors to generate multiple one-dimensional Wasserstein comparisons. The goal is to reduce the projection cost of the standard SPDSW from O(Lnm)O(Lnm) to O(knm+Lkn)O(knm+Lkn); k≪Lk \ll L is assumed. This idea is a combination of SPDSW (Bonnet et al., 2023) and HSW (Nguyen et al., 2023b).

Theoretically, this paper presents the following results. First, when using a direction sampler that is fully O(m)O(m)-invariant in Frobenius-orthonormal coordinates, the effective mixed direction can be decomposed as B=RUB = RU, where UU is a Frobenius-uniform direction and independent of R=∥B∥FR=\|B\|_F. Consequently, the population objective of SPDHSW is E[Rp]\mathbb{E}[R^p] times that of SPDSW for general pp, and when p=2p=2, the two squared objectives coincide exactly.

Second, considering that mixed slices sharing the same bottleneck bank are not independent of each other, we decompose the estimator variance into between-bank and within-bank terms: Vb(k)+Vw(k)L.V_b(k)+\frac{V_w(k)}{L}. Therefore, even when LL is increased, a variance floor denoted by Vb(k)V_b(k) remains. The paper further analyzes the contribution of the degree-two components via spherical-harmonic decomposition.

Third, by extending the same variance decomposition to the empirical log-coordinate particle gradient, we show that the accuracy of the scalar discrepancy and the accuracy of the optimization gradient can differ.

Experiments include synthetic SPD distributions, the FPHA hand-action covariance, several EEG domain-adaptation datasets, and SPDNet regularization. The results show that in high-dimensional settings where the matrix logarithm is precomputed or log-coordinates are optimized directly, the projection step and optimization loop can be significantly accelerated. On the other hand, downstream classification accuracy is generally similar to that of SPDSW, and alignment improvements are relatively small when the budget and learning rate for each method are chosen independently.

Critical strengths
  1. It clearly analyzes dependencies caused by shared bottlenecks. The strongest contribution of this paper lies in the statistical analysis of the hierarchical estimator rather than in the new SPD projection itself. In the single-bank estimator of the existing HSW, the various final slices share the same bottleneck directions and are therefore not independent Monte Carlo samples. This paper explicitly calculates this covariance and shows that the variance is of the form Vb(k)+Vw(k)/LV_b(k)+V_w(k)/L. This clearly explains why the L−1/2L^{-1/2} rate cannot be applied directly and fills an important gap in the previous analysis of HSW (Nguyen et al., 2023b).

  2. It clearly distinguishes between the population objective and the finite estimator. While the population objectives of SPDHSW and SPDSW are the same for p=2p=2, it is useful to distinguish that the variance and computational cost of their finite estimators differ. In particular, the paper concisely explains the population equality through the angular-radial factorization of the effective direction and shows that using normalized directions yields the same population scale even for general pp.

  3. It analyzes gradient variance separately, rather than scalar variance. In actual training, the stochastic gradient is more important than the discrepancy value itself. The paper points out that, since directional optimal coupling varies depending on the direction, the scalar cost and the gradient can exhibit different directional variability. Accordingly, it presents a specific criterion showing that hierarchical mixing does little to aid scalar estimation but can reduce gradient noise.

  4. The experiments cover a broad range and report various constraints with relative transparency. The paper covers projection-stage timing, total optimization timing, comparisons with equal projection counts, independently selected hyperparameters, fixed-direction controls, and diagnostics of scalar and gradient variance. In particular, it clearly reports that end-to-end speedup is virtually nonexistent in affine-invariant updates, where the matrix logarithm dominates the total runtime. The confirmation experiment using held-out high-gamma subjects is another strength of the experimental design.

  5. The computational benefits are clear in high-dimensional log-coordinate settings. When the matrix logarithm is cached or the optimization variable is already S=log⁡XS=\log X, the effect of reducing LL high-dimensional matrix projections to kk is substantial. In particular, the time reduction in projection-dominated loops is convincingly demonstrated in large EEG and FPHA settings where m=d(d+1)/2m=d(d+1)/2.

Critical weaknesses
  1. The methodological originality is limited. The algorithm essentially applies Nguyen et al.’s HSW bottleneck mixing to the SPDSW log-space construction proposed by Bonet et al. Since the log-Euclidean SPD space is isometric to Rm\mathbb{R}^m via Frobenius-orthonormal coordinate transformation, the proposed estimator can be viewed as Euclidean HSW performed in a symmetric log-space. The population constant-factor result and the shared-bank variance analysis are also results that apply directly to Euclidean HSW rather than being inherently specific to SPD. Therefore, the most important novelty of the paper should be more accurately positioned as the analysis and refinement of the existing HSW estimator rather than a new SPD distance.

  2. The gradient variance analysis is limited to diagnostics at a single fixed state. Theorem 3 analyzes the trace variance of the log-coordinate particle gradient at a fixed empirical state. However, it does not show how Vbg(k)V_b^g(k) and Vwg(k)V_w^g(k) change over the entire optimization trajectory, how the directional structure of the gradient covariance changes, or how the actual convergence rate or the range of stable learning rates improves. The descent bound in the Appendix B.9 is also a standard conditional inequality that assumes smoothness, and it does not prove that this condition holds throughout the entire trajectory of the actual sliced-Wasserstein objective. Furthermore, since the analysis focuses on the log-coordinate gradient, the conclusions regarding the affine-invariant Riemannian gradient or the general neural network parameter gradient are limited.

  3. Some of the strong experimental results may be influenced by hyperparameter comparisons. In FPHA, the result showing that SPDHSW (k=50,L=700)(k=50, L=700) achieves alignment similar to that of flat (L=700)(L=700) in a much shorter time is impressive. However, SPDHSW uses the same learning rate as flat (L=50L=50), and the objective of flat actually deteriorates. This can be explained not only by estimator quality but also by differences in the gradient-noise scale and optimal learning rate resulting from the projection budget. An equal-time FPHA comparison, in which the learning rate is selected independently for each method and budget, is necessary.

  4. From the perspective of scalar estimation, hierarchy is not always efficient. With the same number of final slices, flat SPDSW exhibits a lower scalar RMSE than SPDHSW, and in a total-work comparison targeting a 5% relative-RMSE, flat sampling is more efficient in many settings (d=22,63d=22,63). This implies that the advantages of the hierarchy are not applicable to general SPDSW estimation but are limited to specific optimization environments where high-dimensional projection costs dominate and the gradient floor is sufficiently low.

  5. The presentation of notation and key definitions needs improvement. In Theorem 2, VbV_b and VwV_w are first defined, but their notation is then changed to Vb(k)V_b(k) and Vw(k)V_w(k) without further explanation. These quantities actually depend not only on kk but also on fixed values of μ\mu and ν\nu, yet this dependency is omitted. Furthermore, hS(u)h_S(u) in Theorem 3 is introduced in the main text only as the “directional squared cost,” while the precise definition of optimal coupling does not appear until Appendix B.7. It is preferable to explicitly state in the main text the definitions necessary for understanding key theorems. Also, the sentences are broken up abruptly, making them difficult to read.

  6. Claims regarding computational efficiency should more clearly define their scope of application. In affine-invariant updates, the matrix logarithm is a cost shared by both SPDSW and SPDHSW; it is not a problem specific to SPDHSW. However, this also means that the proposed method does not generally accelerate the entire SPD optimization process but only accelerates the post-log projection stage. Therefore, the main computational claim should more prominently highlight the condition of "cached or directly optimized log coordinates."

Questions for authors
  1. The proposed SPDHSW estimator appears mathematically equivalent to applying Euclidean HSW in Frobenius-orthonormal log coordinates. What exactly do the authors consider to be the genuinely SPD-specific contribution? Which parts of Theorems 1–3 do not apply to general Euclidean HSW?

  2. Table 5 estimates Vbg(k)V_b^g(k) and Vwg(k)V_w^g(k) only at the initial FPHA state, before any alignment update. Do the normalized gradient floor and the advantage of hierarchical over flat persist along the optimization trajectory, particularly near convergence where the gradient magnitude becomes small relative to the stochastic gradient variance?

  3. Can you provide an equal wall-clock comparison in the FPHA results, where the learning rate is selected independently for each projection budget and method? In the current shared-rate comparison, it is difficult to distinguish whether the deterioration in the flat case (L=50L=50) is due to an underfitting estimator or a learning-rate mismatch.

  4. Can the gradient analysis be extended to affine-invariant Riemannian gradients or neural network parameter gradients? At a minimum, please explain how a deterministic Jacobian or metric conversion affects the variance decomposition under fixed conditions.

  5. Is it possible to compare independent flat directions, hierarchical mixing, orthogonal directions, or other variance-reduction samplers while maintaining the same computational cost? Based on the current comparison alone, it is difficult to determine whether the benefit is specific to the hierarchy or is a more general effect of structured-direction sampling.

  6. In the independently tuned High-Gamma setting, the improvement in alignment is small, and the classification accuracy is nearly identical. Please explain how the authors interpret the practical effect size of these results and in which real-world application scenarios the reduction in the alignment loop leads to a meaningful speedup for the entire pipeline.

Reviewer_7iQYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is about comparing distributions of SPD matrices. The authors approach this using sliced Wasserstein distance. To make the method efficient, they use a hierarchical (adaptive) technique, which comes from Nguyen 2023b. The contributions claimed by the paper are theoretical properties of this construction and then experimental results that (unsurprisingly) outperform flat estimators.

The theoretical contributions include showing that their method (SPDHSW) has an equivalence to SPDSW in terms of population distance, and then a careful analysis of the estimator's variance. Experiments use FPHA and EEG datasets and overall appear to validate the performance of the method.

In general, the paper appears to apply Nguyen 2023b to the family of distributions of SPD matrices, provides evaluations of this application, and provides a thorough theoretical characterization of the method in this setting.

Critical strengths

Hierarchical sliced Wasserstein (Nguyen 2023b) is an interesting adaptive technique that evidently gives good performance in this use case. The analysis of HSW itself (regardless of the SPD context) seems to go beyond Nguyen 2023b in several substantial ways. For example, as alluded to on L66 of the paper (and then detailed later on and in the appendix), the authors overcome a limitation of the MC error bound in Nguyen 2023b that treated slices as independent. So I believe that observations like this make this paper useful about HSW itself, not just in the SPD context. I also thought it was interesting that the authors estimated the precision of gradients in their FPHA experiments, not just the precision of the loss values. I'm not sure how generalizable of an insight that really is but it was interesting to see in those experiments.

Critical weaknesses

This section of the review asks for only one or two critical weaknesses. I move my other comments to the Questions for Authors section below.

  1. I am concerned about the novelty of this paper and the magnitude of the contribution. The paper seems to primarily be a direct application of Nguyen 2023b to a certain family of distributions (albeit then a careful characterization of the resulting behavior). Although there are a few insights about HSW to be gained by reading the paper, as mentioned above, I do not think they are compelling enough to really make this paper substantially new over Nguyen 2023b, at least at the level of an ICLR submission. In fact, Theorem 1 shows that SPDHSW is equal up to a constant to SPDSW - while I value that as part of the theory in this paper, it also works against the authors in that this isn't really that new over SPDSW then.

  2. The learning rates used by the baseline methods in the comparisons/experiments are not fully justified. For instance, for the tests whose results are shown in Fig. 2, could flat SW use a larger learning rate? What about on the FPHA trials? It looks like when you independently tune the learning rates as in Fig 3, the methods actually perform very simililarly. So while I believe that SPDHSW is efficient, the gap between SPDHSW and baselines is not fully convincing from what is shown. The authors would need to add exposition around choice of learning rates and add additional experiments. But from what I've seen, I fear that these additional experiments would show that SPDHSW actually has only marginal improvements over a well-tuned baseline method, so it might further reduce the magnitude of the contribution here.

Questions for authors
  1. The authors use fp64 on RTX A6000 cards. I know that on these GPUs, fp64 performance is super slow, compared to single-precision. So I would expect if the authors ran using single-precision for their FPHA experiments or other experiments, the speedup of SPDHSW might diminish significantly. Can the authors try this out for different experiments/results in the paper and see? I'd also want to see experiments with more fair learning rates, per comment #2 above.

  2. In the BNCI2014 test, e.g. Fig 4, SPDSW/SPDHSW gets to around 59%. Bonet 2023 achieve around 62% (see their Table 2, SPDSW column). Is the difference due to using a different sampler, or perhaps different learning rates (which learning rate(s) did you use here)?

  3. Although perhaps more of a comment than a question, I noted that many of the paper's conclusions come from results and tables in the appendix versus the main body - and that those results in the appendix often use different GPUs, update geometries, and number of seeds. Is there anything the authors could do to enhance consistency and clarity of the presentation so the paper could be easier to digest in this (albeit constrained) conference format?

Reviewer_9tufOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper applies hierarchical slicing to SPD distributions in symmetric log-space. It computes a small bank of matrix projections and linearly mixes them into many one-dimensional comparisons. It establishes a population relationship with SPDSW, analyzes shared-bank dependence and gradient variance, and evaluates estimation, distribution alignment, and downstream classification.

Critical strengths
  • Hierarchical slicing replaces many expensive slicing projections with cheaper linear combinations, reducing runtime when projection costs dominate.
Critical weaknesses
  • The methodological novelty is limited. The method combines existing SPD slicing [a] with Euclidean hierarchical slicing [b]. The central challenge is defining slices on the manifold, which has already been addressed for SPD matrices in [a] and more broadly for Cartan-Hadamard manifolds in [c]. Once scalar projections are available, existing Euclidean hierarchical mixing can be applied directly. The combination therefore introduces no new ingredient beyond these existing components.

Minor:

  • The speedup depends on whether matrix logarithms must be recomputed at each iteration. In Table 1, AIM updates recompute them at every step and achieve only a 1.01-fold speedup, whereas LEM updates compute them once before the loop and achieve a 5.92-fold speedup.
  • Large-scale dataset experiments are missing, leaving it unclear whether a limited bank of base directions preserves performance at scale.

References

[a] Sliced-Wasserstein on Symmetric Positive Definite Matrices for M/EEG Signals

[b] Hierarchical Sliced Wasserstein Distance

[c] Sliced-Wasserstein Distances and Flows on Cartan-Hadamard Manifolds

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.