- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
Hierarchical WoW distance compute distance between measures over measures. While they can be computed, this is slow. Previous works have proposed slicing procedure as a workaround. The paper proposes a new slicing route via embedding measures in a Hilbert space and to use projections in that space.
Critical strengths
I think that this is a very nice paper, which considers the efficient computation of hierarchical OT distances. These distances and their computation have recently attracted attention and the proposed distance provides a very interesting and new direction. This is a clear and well-written paper based on very well-rounded theory and comparison with other works. In particular, for proposed distance we again get a well-behaved sample complexity which is useful for a lot of applications.
Given my overall impression, I would vote for acceptance. However, I still have some remaining concerns (see below). While I currently lean toward acceptance, my final recommendation will depend on how convincingly the authors address these concerns in their rebuttal.
Critical weaknesses
The 8-Gaussians and checkerboard experiment (Figure 7) is insufficiently visualized and evaluated. Basically, each checkerboard square is one measure and each Gaussian its own measure, i.e., a hierarchy of measures. Thus, the visualization should use color-coded point clouds to make clear that measures not mix and in the plot it should track the WoW distance over time, not the Euclidean Wasserstein distance. I think that the current presentation and evalution can be misleading for a reader.
The bandwidth is chosen pair-dependent in the OTDD (Figures 2–3), point clouds (Figure 4) and patches (Figure 5) experiments, not merely experiment-dependent. This leads to two issues: There is a computational overhead, which affects runtime and should be reported for runtime comparisons. More importantly, we do not get metric properties for pair-dependent bandwidth. Bandwidth parameters are known to be highly important for performance, so I would expect to see experimental results for fixed bandwidth. I think that the experiment should report results with fixed bandwidth and separate ones for tuned bandwidths.
Generally, the benchmarking is very focused on comparing against DSW, but alternative "Busemann-slicing" methods (Bonet et al, '26) and "Wasserstein-over-Sliced-Wasserstein" (as used eg in Piening, Duong, Steidl, '26) are not included.
Questions for authors
Questions/Remarks
Q1 The "exact" OTDD implementation uses an entropic solver. Given the small entropic regularization used, the benchmarking is still fine in that respect, but could you clarify that in the SuppMat?
Q2 The Pearson correlations use squared distances for VSW and DSW, but unsquared distances for s-OTDD, apparently inherited from the github repo of Piening and Beinert (2026). Could you instead report correlations using consistent powers to ensure a fair comparison? Spearman correlations should not be affected by this, so it is only a minor correction for consistency.
Q3 Is the bandwidth tuning overhead included in runtime analysis?
Q4 a) Why does Figure 7 seem as if DSW has not been run to convergence? In constrast, it seems as MMDSW has been run to convergence even though it is more expensive. Did I miss some convergence criterion?
b) Could you add some ablation of parameter for the minimization in Figure 7? It is a nice and interesting new experiment, but neither MMDSW nor DSW has been tried here so i would expect some kind of parameter tuning here for fairness. In particualr, the baseline DSW implementation seems to rely on a fixed coarse integration grid of [0, 1] for computation. I would expect that any gradient descent training should use either an exact integration or a very fine grid. The code is missing for this experiment too and would have been nice here.
Q5 Could you please add references for the OT-NNA, TLB, STLB, AE methods? They are currently not referenced, which is not good scientific practice. In particular, note that the methods are developed from Memoli's Gromov-Wasserstein which is also not referenced.
Q6 Could you also include the Busemann-based slicing method of Bonet et al. (2026) for some experiments?
Q7 An interesting direction would be to investigate whether VSW can also be used to construct explicit transport plans between measures. For example, Generalized Wasserstein Flow Matching (https://arxiv.org/abs/2605.08424) considers couplings between probability measures over probability measures ("meta-plans"), using Wasserstein-over-Sliced-Wasserstein for point-cloud generation. At the level of individual measures, Expected Sliced Transport Plans (EST) (https://arxiv.org/abs/2410.12176) constructs transport plans by averaging lifted one-dimensional couplings. Could an analogous EST-type construction be developed from VSW, potentially yielding computationally efficient meta-plans for point-cloud generation? (Potentially out of scope, but I would be curious)
Q8 HIPM (Catalano and Lavenant, 2024) is discussed but not compared experimentally. Given its favorable statistical guarantees and related test-function formulation, could you add an experimental comparison? Their implementation is mainly targeted at 1D and would presumably not scale to high-dimensional setting, but it might still be nice to reconsider one of their experiments or provide some novel illustrative toy experiment.
Q9 Euclidean MMD can be sliced too. It might be good to have a conceptual comparison in the SuppMat.
Q10
We could also define a Wasserstein-over-MMD (WoMMD) distance, which might seem closer in spirit to the proposed construction. So, we would have various options of estimating distances: MMDSW, DSW, WoSlicedWasserstein, sOTDD (without labels), VSW, WoW, WoMMD. Presumably all of them have very different geometries, but I would expect similar topologies.
Both MMD and Wasserstein metricizes weak convergence, but the geometric difference between MMD and Wasserstein is nicely illustrated via barycenter problems: Wasserstein induces a Euclidean interpolation with varying support and Euclidean MMD induces mixtures with fixed support. For free support barycenter approximations, we can just do gradient descent on particle locations. For such simulations, we often see that sliced Wasserstein behaves more similarly to Wasserstein than MMD does.
Long story short: Could you run a gradient descent based free support approximation of the barycenter between two translated metameasures for some of the different distances (ideally with some shared convergence criterion, sufficient particles, suitable visualization of the hierarchical measure via color-coding)?
This would add some intuition on the different geometries.
I think that would be a valuable addition to the SuppMat and trace back to the origins of sliced Wasserstein, which was primarily introduced for barycenter-type problems.
7 comments
Response to Reviewer kDCN (Part 2)
W2: Pair-dependent bandwidth
Bandwidth computation
We thank the reviewer for raising this concern. We would like to first note that the bandwidth is set by the median heuristic, which is a single statistic of the input points, with no search and no access to any ground truth required. We first clarify what the original runtimes measure. For every method, the runtimes in Figures 3 and 5 and Tables 1 and 2 time the distance computation only, and we exclude the bandwidth. This is done to isolate the cost of each distance method. In the original submission, was set to the median of all pairwise distances among 2000 points sampled from the pooled inner measures, i.e. about distances. In Figure 3 this took 33–83 ms per pair of meta-measures. However, we show that the same median can be estimated from 2000 independent random pairs of points, which costs 2–5 ms (10–30× less), much cheaper than the original. We next verify that the two estimators result in the same bandwidths and the same results. We compare them on the same pairs of meta-measures and the same random slices, so that only changes. In Figure 2, the new is within 2.1% of the original one for every pair, and the correlations with OTDD are essentially unchanged (Table A).
Table A. Correlation of VSW with OTDD in Figure 2, with the original vs. new bandwidth estimator.
| Experiment | Downstream metric | Original → new |
|---|---|---|
| Fig. 2, correlation with OTDD | Pearson / Spearman, MNIST, | 0.904 / 0.924 → 0.900 / 0.928 (largest change over all settings: 0.024) |
All experiments in the revision use the new estimator. That is, Figure 2,4,5, and Table 2 numbers for VSW are updated. However, none of the main conclusions or trends are changed. We have also added the new estimator's description to Appendix H.
Runtimes including the bandwidth. We then show that the computation overhead of the new bandwidth is negligible, and we report runtimes with and without the bandwidth computation overhead. Table B reports Figure 3 on MNIST at four dataset sizes, for the smallest and largest slice budgets.
Table B. Figure 3, MNIST: VSW runtime per distance on CPU (ms, mean of 10 draws), by dataset size.
| VSW runtime (ms) | 5,000 | 20,000 | 40,000 | 60,000 |
|---|---|---|---|---|
| 1k slices, distance only | 6.0 | 16.1 | 37.1 | 59.6 |
| 1k slices, + bandwidth | 8.3 | 18.5 | 39.1 | 61.7 |
| 10k slices, distance only | 23.7 | 33.0 | 52.8 | 76.4 |
| 10k slices, + bandwidth | 26.0 | 35.4 | 54.8 | 78.5 |
The bandwidth step does not depend on the number of slices, and its relative cost shrinks as the datasets grow (+3–4% at 60k samples). Across all datasets, sizes and slice budgets of Figure 3, VSW including the new bandwidth step remains its original trend, faster than DSW and other baselines.
Table C. Table 1 and Figure 5: runtime per distance of VSW without and with the new bandwidth step, and of DSW.
| Experiment | VSW, distance only | VSW + bandwidth | DSW |
|---|---|---|---|
| Table 1, ShapeNet chairs (s) | 0.0125 | 0.0134 | 0.0709 |
| Fig. 5, image patches (ms) | 67 | 89 | 298 |
Table D. Table 2: Runtime per distance (ms). The bandwidth is computed once per dataset, so we spread its cost over all pairs of shapes.
| Method | 2D Shapes | Animals | FAUST-500 | FAUST-1000 | MNIST-2000 |
|---|---|---|---|---|---|
| VSW, distance only | 0.35 | 0.34 | 2.80 | 3.76 | 7.48 |
| VSW, with new bandwidth | 0.36 | 0.35 | 2.81 | 3.77 | 7.49 |
| SQW | 0.93 | 0.95 | 6.51 | 10.15 | 18.46 |
With the new estimator, the bandwidth adds 2–5 ms per pair in Figure 3, 0.9 ms in Table 1, 22 ms in Figure 5, and at most 0.01 ms per distance in Table 2. With the new bandwidth computation overhead included, none of the comparisons in the paper changes.
Fixed bandwidth experiments
We have now added an ablation with a fixed bandwidth to Appendix F.2, Figures 8–9. We use the point-cloud and patch experiments. For each sweep, we compute a single bandwidth by the median heuristic on the reference meta-measures only (the side of each comparison that is kept fixed along the sweep), and use it for every pair. Since never depends on the targets, VSW is a single metric along the whole sweep. The pairs and the random slices are the same for every bandwidth, so only the bandwidth changes. The fixed generally gives the same trends as the per-pair bandwidth in all sweeps.
Bandwidth sweep. We also report every bandwidth for separately. VSW recovers the trends of the main text for bandwidths within a factor of two of and degrades gradually outside this range. VSW with a much smaller bandwidth detects that the target has changed, but not by how much. A much larger one only sees coarse differences. Tuning within this range changes the results only slightly, so the median heuristic, which needs no tuning, is already a good choice.
Response to Reviewer kDCN (Part 3)
W3: Additional baselines
We thank the reviewer for their suggestion. We have now added Busemann-slicing baselines (Bonet et al, 2026) to the dataset distance experiment (Figure 2, 3). Additionally, we added Wasserstein-over-Sliced-Wasserstein (WoSW) to the point-cloud and patch experiments (Figure 4,5 and Table 1), and summarized results are provided below. WoSW is also added to the barycenter computation experiments (Appendix G.2) (For additional details, please refer to our response to Q10).
Busemann-slicing methods. We include SWB1DG, which uses one-dimensional Gaussian geodesic rays, and SWBG, which uses a Gaussian approximation of each class after a t-SNE embedding. We use the official public implementation without modification and with its default hyperparameters. As for every other method in Figure 2, samples are projected linearly from the flattened images.
Both Busemann distances are more accurate than s-OTDD, and generally more accurate than DSW at small and moderate budgets. For few slices, VSW remains the most accurate method or is on par with the best one, and for large budgets, VSW is competitive with the Busemann distances.
We also timed both methods under the same protocol as Figure 3. At every dataset size and budget, SWB1DG is slower than s-OTDD and orders of magnitude slower than VSW. SWBG is dominated by the t-SNE embedding, and becomes too slow beyond small datasets. We therefore stop calls that take longer than the slowest s-OTDD call, and mark such calls as timed out in the figure.
Overall, the Busemann distances can match VSW's accuracy only with many more slices and at a much higher cost per evaluation. VSW therefore still offers the best accuracy–runtime trade-off. The revised paper includes both methods in Figures 2 and 3, and the implementation details in the appendix.
WoSW baseline. We evaluate WoSW on exactly the same data as the other methods in Figure 4, Figure 5 and Table 1, with projections. VSW still remains the fastest method: 0.0125 s vs. 0.145 s for WoSW on the ShapeNet benchmark (Table 1, 11.6×), and 67 ms vs. 159 ms on the image patches (Figure 5, 2.4×), where VSW also follows exact WoW more closely. On the low-dimensional point clouds of Figure 4, all methods show the same trends.
Questions
Q1: "exact" OTDD clarification
We thank the reviewer for their suggestion. We have revised Appendix H (“Optimal Transport Dataset Distance”) to clarify this.
Q2: Consistent powers in Pearson correlations
We thank the reviewer for bringing this to our attention. We have now updated Figure 2 to use the squared s-OTDD distance.
Q3: Is the bandwidth tuning overhead included in runtime analysis?
Please refer to our response to W2.
Q4(a) and (b): Figure 7: DSW convergence and parameter ablation
Q4(a): We appreciate the opportunity to clarify this. For this experiment, all methods are trained for the same fixed budget of 100,000 steps, and there is no early stopping. The curves end at different points since the x-axis is wall-clock time. Since MMDSW is the slowest per step, its curve extends furthest. In the revision, each method is still tuned within this 100k-step budget (see (b)). To make sure that DSW's result is not a convergence issue, we also trained DSW with its selected settings for 300,000 steps (3 seeds). Its final does not change, and remains above both VSW and MMDSW. The gap is therefore not due to training length.
Q4(b): We agree with the reviewer’s comment, and we sweep the main hyperparameters of each method on both datasets:
- VSW: slices , random features , the length of the initial large-bandwidth phase (10, 20 or 30% of the steps), the final bandwidth, and the learning rate (, , ).
- DSW: GP length scale , quantile grid of 50 or 200 points, slices (outer × inner) , and the learning rate (, ).
- MMDSW: Riesz exponent , projections , and the learning rate (, ).
For every method, we select the configuration with the lowest final and rerun it with 3 seeds for Figure 7. Table 4 in Appendix H lists the swept and selected values. Overall, every method is tuned with the same step budget, selection criterion and number of seeds to ensure the comparison is fair. After tuning, VSW matches MMDSW and improves on DSW, while training 2–5× faster than MMDSW in wall-clock time. We will also publicly release the code for all experiments, including the Figure 7 experiments and hyperparameter sweeps, upon acceptance.
Response to Reviewer kDCN (Part 1)
We thank the reviewer for their positive assessment, and their valuable comments. We address the raised questions and concerns below, and we have revised the paper accordingly.
We agree that both the visualization and evaluation should better reflect the hierarchical structure of the data. In the original submission, we reported point-level W22 as a simple measure of the overall generated sample quality, which does not account for the distinction between inner distributions. We have therefore updated Figure 7 to color-code the generated point clouds according to their nearest target component. We have also replaced the point-level W22 with Wasserstein-over-Wasserstein (WoW22), which explicitly accounts for both the inner and outer distributions. To ensure a consistent evaluation, we compute WoW against the target meta-measure with exactly uniform weights across its eight components (each assigned weight 1/8), rather than using a randomly sampled collection of target clouds. This avoids fluctuations in the number of target clouds per component that could otherwise affect the evaluation. We have clarified this protocol in Section 4 and Appendix H.