- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper investigates the problem of estimating an efficient optimal-transport-based distance between Gaussian mixtures. It builds on previously introduced sliced estimators and proposes a alternative estimator based linearization (via embedding Gaussians into a Hilbert space). The main application is the reduction of Gaussian components in Gaussian splatting, i.e., data compression for 3D objects.
Critical strengths
The Gaussian splatting application is very interesting.
The linearization can lead to significant accelerations.
Critical weaknesses
Writing I would say that the writing fails to motivate the construction. As particular example, I would point out Subsection 2.2. Sentences like "The first term changes scale equally in every direction, while the traceless term changes shape. To compare local sensitivity, divide each squared distance" have interesting ideas and insights underlying them, but seem very vague to me. It would have been nice to provide more intuition.
I think that tools such as the ICLR-provided "Paper Assistant Tool" could have been leveraged more to improve writing.
Relevance While I think that the Gaussian Splat experiment is quite interesting, it is a rather small benchmarking over a small dataset with only one refernce method and the differences over the reference method are very small. Also the proposed ordering does not seem to be an optimal choice for all settings, see stronger lexicographic baseline ("Lex") in Table 6 outperforming the proposed distance.
While the retrieval benchmarking provides a nice framework for benchmarking, it is very far from practical applications.
Topological Differences to MW Unlike for DSMW, convergence in MW distance does not induce convergence in the proposed distances. I think that it is nicely illustrated in Appendix A.7, but I think that this is not what we would want from such a distance.
Generally, the (Bures)Wasserstein space is a nonlinear space. While linearized OT has been useful in a lot of applications, linearization quickly become very coarse approximation of the original topology without additional assumption.
Novelty
A lot of the ideas have been proposed been before. In particular, the Mixture Wasserstein Distance of (Delon&Desolneux) can be seen as a special case of the Wasserstein-over-Wasserstein distance restricted to metameasures restricted to Gaussian support (see eg Bonet et al 5., Piening et all 4. and 6.). I would refer to E.1 in Ref. 5 below for an (incomplete) overview.
In addition the DSMW used as comparison in the paper, Bonet et al. propose sliced estimators based on Busemann functions (2. below) and Nguyen (3.) constructs very similar sliced estimates to the DSMW based on earlier work by Bonet (1.) Beyond Gaussian, Piening&Beinert use a Hilbert space embedding for more general measures in 4 - this is a different approach than the proposed one, but a distinction should be drawn.
Moreover, I would note that partial linearization of such hierarchical distances has also been proposed in 6 coming from a different direction.
Lastly, the main metric theorem holds for any canonical deterministic ordering, not specifically Hilbert ordering. (Example: lexicographic baseline)
Relevant references
-
Bonet et al. (2023). Sliced-Wasserstein on Symmetric Positive Definite Matrices for M/EEG Signals. ICML. https://arxiv.org/abs/2303.05798
-
Bonet et al. (2026). Busemann Functions in the Wasserstein Space: Existence, Closed-Forms, and Applications to Slicing. AISTATS. https://proceedings.mlr.press/v300/bonet26a.html
-
Nguyen and Mueller (2024). Summarizing Bayesian Nonparametric Mixture Posterior -- Sliced Optimal Transport Metrics for Gaussian Mixtures. https://arxiv.org/abs/2411.14674
-
Piening and Beinert (2026). Slicing Wasserstein Over Wasserstein Via Functional Optimal Transport. ICLR. https://arxiv.org/abs/2509.22138
-
Bonet, Vauthier, and Korba (2025). Flowing Datasets with Wasserstein over Wasserstein Gradient Flows. ICML. https://arxiv.org/abs/2506.07534
-
Piening, Duong, and Steidl (2026). Generalized Wasserstein Flow Matching: Transport Plans, Everywhere, All at Once. https://arxiv.org/abs/2605.08424
Questions for authors
Can you establish a very clear use case for the proposed distance for 3D applications? How does your method perform if compared here to your lexicographic baseline?
6 comments
Response to Reviewer vGzY (Part 2/2)
On Relevance / Q1. Four-scene 3DGS compression with six methods (Sec. 6.2, Table 2, Figure 6, App. B.6, C.6).
• Practical result. At 1% retention with common recovery, Hilbert improves on Lex by 1.214–4.027 dB and on the highest-PSNR pruning baseline in each scene by 0.931–1.964 dB. GHAP remains close and has the highest PSNR on Truck.
• Matched comparison. Table 2 compares Hilbert, XYZ Lex, GHAP and the pruning stages of PUP 3D-GS, LightGaussian and MesonGS on Train, Truck, Garden and Playroom. All six start from identical scene checkpoints and share cameras, resolution and 5,000 tuning steps over three seeds. The pruning methods use their published importance scores in one pass followed by common recovery; attributes remain in full precision. Target counts are 10,265/20,686/58,347/25,461; GHAP retains 4/7/5/2 fewer primitives.
• Why grouping helps. Our spatial extension orders Gaussian centers, then applies GHAP's within-block reduction. Hilbert and Lex use identical group sizes and refinement, differing only in center order. Hilbert forms spatially compact groups before refining representatives. On Playroom at 1%, Lex's initial within-group center RMS dispersion is 16.64 times Hilbert's. The following held-out PSNR values are mean ± SD over three tuning seeds.
| Method | Train | Truck | Garden | Playroom |
|---|---|---|---|---|
| Hilbert | 18.720±0.043 | 21.530±0.016 | 23.202±0.010 | 27.639±0.063 |
| XYZ Lex | 17.506±0.037 | 19.221±0.013 | 20.974±0.012 | 23.612±0.043 |
| GHAP | 18.588±0.010 | 21.551±0.035 | 23.181±0.006 | 27.414±0.017 |
| LightGaussian | 17.238±0.034 | 19.558±0.006 | 21.903±0.012 | 23.651±0.012 |
| PUP 3D-GS | 16.615±0.030 | 18.844±0.013 | 22.271±0.013 | 25.785±0.038 |
| MesonGS | 17.268±0.016 | 19.566±0.017 | 22.171±0.007 | 24.096±0.015 |
• Speed–quality trade-off. Hilbert and LightGaussian form the measured frontier in the four-scene average. Figure 6 uses the same six methods and four scenes as Table 2. Hilbert averages 10.75 s / 22.773 dB, versus LightGaussian's 5.84 s / 20.588 dB and GHAP's 31.44 s / 22.684 dB. Times include pruning-score computation, are measured separately on the same A800, and exclude shared tuning. Tables 12–13 report the separate Hilbert/Lex/GHAP retention sweep on RTX 3090 GPUs.
On Relevance. Full-distance retrieval and why Lex won on aligned digits (Sec. 5.3, Figure 7, Tables 8–10, App. B.4, C.4).
• What the Lex comparison revealed (Table 1, Figure 7). Our original preprocessing fixed Gaussian spatial means at shared block centers, giving Lex exact block correspondence. We overlooked this advantage when interpreting the submitted UCI digit result (Lex 97.0% R@1; HMW 94.0%). New controls on MNIST, USPS, Fashion-MNIST and KMNIST add independent Gaussian noise (standard deviation ) to component means, keeping intensity means, covariances and weights fixed. Relative to each dataset's unperturbed inputs, Lex R@1 drops by 31.11–58.11 percentage points, versus 3.22–6.00 for HMW. HMW has the highest mean R@1 among the five methods below, including four-direction DSMW and SMix-W. Scores average three perturbations and, for sliced methods, three direction pools; each dataset has 500 gallery images and 300 queries.
| Dataset | HMW | Lex | Moments | DSMW-4 | SMix-W-4 |
|---|---|---|---|---|---|
| MNIST | 65.67 | 22.56 | 21.67 | 53.15 | 40.48 |
| USPS | 79.89 | 39.78 | 51.67 | 66.59 | 58.67 |
| Fashion-MNIST | 65.11 | 38.56 | 48.67 | 63.48 | 59.96 |
| KMNIST | 53.44 | 22.22 | 21.00 | 37.22 | 27.67 |
The matched five-method accuracy–time comparison places HMW on the measured frontier for MNIST, USPS and KMNIST (Figure 7, Table 10). Both sliced methods use four direction pairs and have lower query cost; DSMW-4 is also slightly more accurate on Fashion-MNIST in the timed configuration. Those points use the first perturbation and direction pool, separately from the multi-seed means above. The result measures coordinate sensitivity around a shared grid. Section 6.2 evaluates the spatial grouping extension in 3DGS compression.
On the Ethics-Review Flag
Could you clarify the concern behind the ethics flag so we can address it directly? This work studies Gaussian mixture transport on public benchmarks. The AI Use Statement (p. 11) describes the assistance received and our responsibility for the claims, proofs, experiments and citations.
Response to Reviewer vGzY (Part 1/2)
Thank you for highlighting the 3DGS application and for the helpful references.
On Writing. Geometric intuition for scale, shape, and normalization (Sec. 2.2, Figure 2).
Your request for intuition led us to illustrate how projection can cancel opposing changes in shape. Section 2.2 considers two unit-norm perturbations around σI: Sscale=diag(1,1)/2 (uniform expansion) and Sshape=diag(1,−1)/2 (expanding one axis while contracting the other). An equal-axis 1D projection cancels Sshape to first order. Dividing squared distances by ϵ2 isolates directional sensitivity from magnitude. After calibrating uniform expansion to BW, MSW/DSMW weight traceless shape deformations by only 2/(d+2) of BW (1/2 in 2D), illustrating the directional sensitivity behind Figure 2. Factor pricing is exact; panel (a) also uses exact matching.
On Topology. Assumptions preserving MW convergence (Sec. 4.3, App. A.6–A.7, C.3).
• Local topological equivalence (Theorem 2). Your concern helped us clarify when MW and HMW convergence agree. Theorem 2 establishes this locally on the class of mixtures with n distinct SPD atoms and positive weights, at a source with positive recursive cut margin γ>0. Write D=MWBW and U=HMW. Intuition. Separated source atoms force an eventual 1-to-1 correspondence under small MW; positive cut gaps prevent rank swaps once feature residuals after a common mean translation satisfy maxi∥ei∥∞<γ/2; and converging weights make the mass-interval contribution vanish. For fixed weights, 0≤U2−D2≤∑iai∥ei∥2. All 200 controls satisfy the corresponding bound for their weight setting. Local convergence equivalence does not imply U/D→1.
• Global boundary jump (Theorem 3). Continuously exchanging two atoms forces a nonvanishing jump for any deterministic, label-invariant, partner-independent order assigning one contiguous interval per atom (e.g., D=0.006136 vs. U=1.999991 with identical covariances).
On Novelty. Positioning against the 6 suggested references (Sec. 1, App. D, Table 14).
Your references helped us clarify HMW's independently prepared step-function representation of each mixture, built from full Gaussian features without projections. Section 1 and Appendix D discuss Gaussian MW as a Gaussian-supported WoW metric (Delon & Desolneux, 2020; Bonet et al., 2025); geometric SPD and Busemann slicing (Bonet et al., 2023, 2026); and Nguyen & Mueller's (2024 preprint, cited as 2025 v5) SMix-W, benchmarked alongside DSMW. SMix-W uses log standard deviations in stage two, versus DSMW's standard deviations. Piening & Beinert (ICLR 2026) embed 1D-projected components as quantile functions for functional slicing. Unlike reference-based lazy plans (Piening et al., 2026), HMW orders each mixture independently following HCP (Li et al., 2024), adding Gaussian factor-exactness characterizations, margin-based rank stability, and the quantitative single-order obstruction.