acceptodds
Under review as a conference paper at ICLR 2027

Hilbert Mixture Wasserstein Distance for Efficient Gaussian Mixture Comparison

Abstract

Mixture Wasserstein compares Gaussian mixtures by optimizing both Gaussian factor alignment and component matching for every pair, making repeated comparisons costly. We introduce Hilbert Mixture Wasserstein (HMW), a metric built from a reusable ordered representation of each mixture. We encode each Gaussian by its mean and Cholesky factor and sort these vectors in a deterministic Hilbert order. Component weights set the lengths of consecutive subintervals of a unit interval. Overlaps between the two mixtures' subintervals specify a sparse coupling, with costs evaluated from the full feature vectors. Comparison is linear in the prepared representations' combined size, and equally weighted mixtures with a common component count admit exact finite Euclidean embeddings. HMW upper-bounds mixture Wasserstein, and we decompose the squared-distance gap into the costs of fixing factor alignment and component matching. We give conditions for a small gap and show that changes in component order can keep HMW bounded away from zero as mixture Wasserstein tends to zero. Constructed examples distinguish covariance response from component matching, and timing experiments examine scaling with mixture size and dimension. We also examine representation reuse in image retrieval and ordered grouping in density reduction and 3D Gaussian Splatting scene compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian mixture models, optimal transport, Wasserstein distances, Hilbert space-filling curves, mixture reduction, 3D Gaussian Splatting.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vGzYOpenReview ↗
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

  1. Bonet et al. (2023). Sliced-Wasserstein on Symmetric Positive Definite Matrices for M/EEG Signals. ICML. https://arxiv.org/abs/2303.05798

  2. 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

  3. Nguyen and Mueller (2024). Summarizing Bayesian Nonparametric Mixture Posterior -- Sliced Optimal Transport Metrics for Gaussian Mixtures. https://arxiv.org/abs/2411.14674

  4. Piening and Beinert (2026). Slicing Wasserstein Over Wasserstein Via Functional Optimal Transport. ICLR. https://arxiv.org/abs/2509.22138

  5. Bonet, Vauthier, and Korba (2025). Flowing Datasets with Wasserstein over Wasserstein Gradient Flows. ICML. https://arxiv.org/abs/2506.07534

  6. 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
AuthorsOpenReview ↗

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\sigma I: Sscale=diag(1,1)/2S_{\mathrm{scale}} = \mathrm{diag}(1,1)/\sqrt{2} (uniform expansion) and Sshape=diag(1,−1)/2S_{\mathrm{shape}} = \mathrm{diag}(1,-1)/\sqrt{2} (expanding one axis while contracting the other). An equal-axis 1D projection cancels SshapeS_{\mathrm{shape}} to first order. Dividing squared distances by ϵ2\epsilon^2 isolates directional sensitivity from magnitude. After calibrating uniform expansion to BW, MSW/DSMW weight traceless shape deformations by only 2/(d+2)2/(d+2) of BW (1/21/2 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 nn distinct SPD atoms and positive weights, at a source with positive recursive cut margin γ>0\gamma > 0. Write D=MWBWD=\mathrm{MW}_{\mathrm{BW}} and U=HMWU=\mathrm{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 max⁡i∥ei∥∞<γ/2\max_i \Vert{}e_i\Vert{}_\infty < \gamma/2; and converging weights make the mass-interval contribution vanish. For fixed weights, 0≤U2−D2≤∑iai∥ei∥20 \le U^2 - D^2 \le \sum_i a_i \Vert e_i\Vert^2. All 200 controls satisfy the corresponding bound for their weight setting. Local convergence equivalence does not imply U/D→1U/D\to1.

• 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=0.006136D = 0.006136 vs. U=1.999991U = 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.

AuthorsOpenReview ↗

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.

MethodTrainTruckGardenPlayroom
Hilbert18.720±0.04321.530±0.01623.202±0.01027.639±0.063
XYZ Lex17.506±0.03719.221±0.01320.974±0.01223.612±0.043
GHAP18.588±0.01021.551±0.03523.181±0.00627.414±0.017
LightGaussian17.238±0.03419.558±0.00621.903±0.01223.651±0.012
PUP 3D-GS16.615±0.03018.844±0.01322.271±0.01325.785±0.038
MesonGS17.268±0.01619.566±0.01722.171±0.00724.096±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 σ=10−4\sigma=10^{-4}) to component (x,y)(x,y) 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.

DatasetHMWLexMomentsDSMW-4SMix-W-4
MNIST65.6722.5621.6753.1540.48
USPS79.8939.7851.6766.5958.67
Fashion-MNIST65.1138.5648.6763.4859.96
KMNIST53.4422.2221.0037.2227.67

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.

AuthorsOpenReview ↗

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.

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

Comparing Gaussian mixtures is expensive. This paper proposes HMW for efficient GM comparison by generating a reusable representation for GM using sliced Wasserstein without solving OT problem for every pair of Gaussians. The algorithm enables linear-complexity comparison between GMs, and experiments demonstrate significant speedup.

Critical strengths
  1. The motivation of this paper is sound, and HMW offers a practical solution to this important problem.

  2. The proposal: reusable representation avoids repeated OT computation and can be very effective, especially when comparing many mixtures.

  3. The paper also provides a clear mathematical formulation for analysis and makes the method intuitive.

  4. The application demonstration and achieved efficiency gains are good and make the proposed method more practically convincing.

Critical weaknesses
  1. The introduction is hard to read. The current version reads more like an encyclopedia of related fields.

  2. The novelty of the proposed method seems somewhat limited, especially when HCP already orders samples with Hilbert curves and evaluates transport cost.

  3. Although MW is upper-bounded by the HMW, how the 2 quantities are close needs further attention and characterization. The potential distortion induced by method speed-up can be large, especially with the relative squared excess number in Table 5.

  4. The paper emphasizes efficiency speedups but insufficiently discusses the tradeoff between efficiency and quality loss, especially in downstream tasks.

Questions for authors
  1. Why is Hilbert order preferable to other simpler orderings? For example, in Table 6, simple Lex seems to be better than HMW.

  2. Refer to my weakness point #4, a Pareto graph would be very helpful in visualizing the trade-offs between quality and speed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer WkKw (Part 1/2)

Thank you for highlighting the practical value of reusable representations and efficiency gains.

On W1–W2. Clearer narrative and contribution relative to HCP (Sec. 1, 3, 4, App. D).

• Projection-free construction. Your feedback helped us center the introduction on how each mixture can be prepared once and reused. Sections 1 and 3 now present the construction directly. HMW performs no slicing or directional averaging. Each mixture is sorted once in full Gaussian feature space, weights set interval lengths on (0,1](0,1], and interval overlaps define at most K+M−1K+M-1 transport edges (O((K+M)p)O((K+M)p) comparison for pp features).

• Scope relative to HCP. We credit HCP for the ordered step-function metric and upper bound (also valid for Lex). Our Gaussian contributions are the exactness/residual analysis of fixed covariance factors, input conditions for rank stability, and the quantitative single-order obstruction, separating factor-pricing and matching excess.

On W3. Characterizing closeness to MW and large distortions (Sec. 4.3, 5.1–5.2, App. A.6, B.2, C.1–C.3).

• What drives the error. Your question helped us make the source of distortion explicit. Fixed-order matching dominates the excess on the tested unaligned mixtures. Write D=MWBWD=\mathrm{MW}_{\mathrm{BW}}, U=HMWU=\mathrm{HMW} and let V2V^2 be the cost of the ordered plan evaluated with exact BW costs. The decomposition

U2−D2=(U2−V2)⏟pricing A+(V2−D2)⏟matching BU^2-D^2=\underbrace{(U^2-V^2)}_{\text{pricing }A}+\underbrace{(V^2-D^2)}_{\text{matching }B}

attributes 71–89% of squared excess on dense pairs to BB, and 100% on diagonal K=512,d=3K=512, d=3 pairs (RMS U/D=2.106U/D=2.106). Section 5.2 pairs the runtime results with their approximation error: 346.1% mean relative error on the 11 pairs with exact MW references.

• A checkable accuracy guarantee (Theorem 2). We bound absolute error when correspondence ranks remain stable. Let γ>0\gamma > 0 be the smallest coordinate gap across the source sorter's recursive median splits. Intuition. Perturbations yi=xi+(t,0)+eiy_i = x_i + (t,0) + e_i with max⁡i∥ei∥∞<γ/2\max_i \Vert{}e_i\Vert{}_\infty < \gamma/2 cannot cross any cut, preserving correspondence ranks and bounding 0≤U2−D2≤∑iai∥ei∥20 \le U^2 - D^2 \le \sum_i a_i \Vert{}e_i\Vert{}^2 for unchanged weights. Here tt is a common mean translation. Changing weights adds a mass-interval term. All 200 within-radius controls satisfy the bound. Removing tt preserves mean squared excess (1.42×10−71.42\times10^{-7}) while mean relative error rises from 0.0000230.000023\\% to 7.387.38\\% as DD shrinks. Theorem 3 gives the complementary global obstruction for a single canonical order.

AuthorsOpenReview ↗

Response to Reviewer WkKw (Part 2/2)

On Q1. Why Lex won on the submitted digit benchmark (Sec. 5.3, Tables 1, 8, App. B.4, C.4).

• What explains Lex's advantage. Your question prompted us to revisit the encoding and identify a shared-grid alignment we had overlooked. In the 8×88\times 8 encoding, every image shared identical (x,y)(x,y) block centers. Lex sorted by xx then yy, matching the same spatial block across images. In the submitted digit benchmark, label R@1 was 97.0% for Lex and 94.0% for HMW.

• Breaking exact grid ties. Adding tiny independent (x,y)(x,y) Gaussian noise (σ=10−4\sigma = 10^{-4}, 0.035% of block spacing) across four 16-component image datasets breaks Lex's shared xx-ties while leaving intensities, covariances, and weights intact. Each dataset has 500 gallery images and 300 queries. R@1 (%) averages three perturbations and, for sliced methods, three direction pools; DSMW-4/SMix-W-4 use four direction pairs.

DatasetHMWLexMomentsDSMW-4SMix-W-4
MNIST65.6722.5621.6753.1540.48
USPS79.8939.7851.6766.5958.67
Fashion-MNIST65.1138.5648.6763.4859.96
KMNIST53.4422.2221.0037.2227.67

HMW has the highest mean accuracy among these five methods on all four datasets, losing 3.22–6.00 points from unperturbed inputs vs. Lex's 31.11–58.11 point decrease. Applying shared noise across images leaves Lex distances unchanged (<6×10−15< 6\times 10^{-15}), isolating the correspondence effect behind Lex's drop.

On W4/Q2. Pareto graphs and quality–speed trade-offs (Figures 6–7; Tables 8–13).

• Perturbed grid images (Figure 7, Tables 9–10). Following your suggestion, Figure 7 pairs R@1 with online query time for the same five methods. HMW lies on the measured frontier for MNIST, USPS and KMNIST; both sliced baselines use four direction pairs. At this budget, sliced methods have lower query cost. On Fashion-MNIST, DSMW-4 is slightly ahead in both quantities: 65.00%/0.334 ms versus HMW's 64.67%/0.538 ms. These paired points use the first perturbation and direction pool; the table above averages multiple perturbations and pools. Online time includes query preparation, scanning the prepared gallery and ranking, and excludes image encoding and gallery/direction preparation. These controls measure coordinate sensitivity around a shared grid.

• 3DGS compression (Figure 6, Table 2, Sec. 6.2). Hilbert (10.75 s, 22.773 dB) and LightGaussian (5.84 s, 20.588 dB) form the measured frontier in the four-scene average. All six methods use 1% retention and common 5,000-step tuning over three seeds; PUP 3D-GS, LightGaussian and MesonGS contribute their pruning stages. Hilbert has lower compression time and higher mean PSNR than Lex (13.57 s, 20.328 dB), GHAP (31.44 s, 22.684 dB), PUP (22.52 s, 20.879 dB) and MesonGS (90.84 s, 20.776 dB). These A800 timings include importance scoring and exclude tuning. Table 2 retains the scene-level differences, including GHAP's higher PSNR on Truck.

• Density reduction (Sec. 6.1). Across all 60 density-reduction inputs (Table 11), Hilbert improves on Lex in 27/60 cases with five fixed-group iterations and 40/60 with three CTD cycles. The separated and curved examples show that grouping quality depends on geometry.

Dear reviewer. Thank you for posting the response. I particularly like the new Pareto graph in Figure 7, which demonstrates the strength of HMW, despite being somewhat expensive to compute, relatively speaking. Figure 6 is also a nice add-on for practical impact, although the results feel a bit cherry-picked (for the scene selection), but the winning margin is significant.

As a side note, some newly added text in the updated PDF feels really rushed and hard to read or understand. I encourage the authors to improve the readability in the future version, if possible.

Overall, I am leaning toward raising my rating a little.

Dear Reviewer WkKw,

Thank you for taking the time to read the revision. We’re glad the Pareto plots helped make the trade-off between quality and speed clearer. You were right that some of the new material felt rushed.

We shortened the paper, mainly by removing repeated explanations and consolidating tables in the appendix. The introduction now starts with a concrete image retrieval problem. Section 3 walks through how two short lists of Gaussian components are matched, and Section 4.3 explains why small movements preserve their order before asking the reader to follow the theorem.

The 3D Gaussian Splatting experiment in Section 6.2 is a small application study showing how Hilbert ordering can group nearby Gaussians for compression. Our original submission used Train and Truck from Tanks and Temples. We added Garden from Mip-NeRF 360 and Playroom from Deep Blending in the revision. Table 2 reports the six-method comparison on all four scenes from our retention study, with results shown separately for each scene.

We welcome any further suggestions you may have. Thank you again for helping us make the paper clearer and easier to read.

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

This paper introduces Hilbert Mixture Wasserstein (HMW), a metric for comparing finite measures of Gaussian components through independently prepared representations. Each Gaussian is encoded by its mean and Cholesky factor, and the components are placed in a deterministic Hilbert order. Component weights define intervals whose overlaps induce a sparse transport plan. The paper establishes metric properties, an upper bound on mixture Wasserstein, and a decomposition of the excess cost into factor-alignment and component-matching terms. It also studies local covariance responses, conditional approximation guarantees, and discontinuities inherent to deterministic component ordering. Experiments examine computational cost, representation reuse, density reduction, and a spatial grouping extension for 3D Gaussian Splatting compression.

Critical strengths
  1. The mathematical construction is clear, and the main arguments appear sound under the stated assumptions. The feature-valued step-function representation provides a clean explanation of metricity and reusable comparisons. Separating the excess over mixture Wasserstein into factor-alignment and coupling contributions is also useful: it distinguishes two different sources of approximation error that would otherwise be conflated. The local covariance analysis further clarifies how slicing and Cholesky pricing respond differently to shape perturbations.
  2. The paper treats the limitations of deterministic ordering seriously. Remark 11 gives an informative obstruction that applies beyond the particular Hilbert implementation, showing why a single comparison-independent ordering cannot be globally continuous in the mixture Wasserstein topology. I also appreciate the transparency of the empirical reporting, including substantial coupling errors, competitive alternative representations, and application settings where the proposed approach is not better. For a theory-oriented paper, the experiments provide useful illustrations without needing to establish large-scale application superiority.
Critical weaknesses
  1. The theoretical contribution relative to existing Hilbert curve projection methods needs to be established more clearly. HCP already uses ordering-induced couplings evaluated in the original space and studies metricity, Wasserstein upper bounds, and topological nonequivalence. Much of the basic HMW construction appears closely related to applying this approach to a Gaussian feature measure. The Gaussian factor analysis, excess-cost decomposition, and ordering-independent obstruction are potentially valuable additions, but the paper does not sufficiently distinguish inherited properties, general consequences of fixed representations, and genuinely new Gaussian-specific results. My reservation concerns the significance and positioning of the theoretical advance, rather than the correctness of the basic metric construction.
  2. The analysis does not yet provide a strong justification for the particular combination of Hilbert ordering and Cholesky features. The near-Monge guarantee is valid for any ordering satisfying its cost-matrix assumption, while the perturbation result requires a restrictive reference configuration and preserved orders. Neither establishes when Hilbert ordering itself produces a good coupling. Similarly, the paper identifies the coordinate dependence of Cholesky pricing and discusses a principal-root alternative, but does not adequately explain the trade-off motivating its final choice. The experiments reinforce this question: coupling error can dominate the approximation gap, and lexicographic ordering performs better in the reported retrieval experiment. I do not expect a global approximation bound, which the paper’s own obstruction rules out, or substantially larger experiments. However, a more informative conditional analysis or a few targeted comparisons would strengthen the methodological justification.
Questions for authors
  1. Could the authors explicitly identify which results follow from existing HCP theory after introducing Gaussian features, and which constitute new results? In particular, how does the ordering-independent obstruction in Remark 11 advance the previously established topological nonequivalence of HCP and Wasserstein distances?
  2. Are there sufficient conditions stated directly in terms of the input mixtures that explain when Hilbert ordering yields a small coupling excess? A sorting-margin condition, a non-collinear geometric subclass, or an inexpensive diagnostic would make the conditional approximation results more informative than assuming that the resulting cost matrix is near Monge.
  3. What motivates Cholesky features over principal-root features? Could a small controlled comparison separate their effects on component pricing from their effects on ordering, while reporting preparation cost and approximation error?
  4. Could the runtime presentation more prominently distinguish the complete canonical implementation from the simplified path used in Figure 3? Please also clarify whether the sliced runtime comparisons change materially when directions are sampled uniformly on the sphere, consistently with the theoretical definitions, rather than from normalized positive-coordinate vectors.
5 comments
AuthorsOpenReview ↗

Response to Reviewer ZQDt (Part 1/2)

Thank you for recognizing the clarity of the construction and the value of separating factor-alignment and matching errors.

On W1/Q1. Contribution beyond HCP (Sec. 3–4, App. D, Table 14).

[Current manuscript, Secs. 1, 3–4, App. A.6]

• Contribution beyond HCP. We explicitly credit HCP for ordered coupling, canonical metricity, and the feasible upper bound (which also hold for Lex). Let D=MWBWD=\mathrm{MW}_{\mathrm{BW}}, U=HMWU=\mathrm{HMW} and let V2V^2 be the cost of the same ordered plan evaluated with exact BW costs. Our Gaussian analysis uses

U2−D2=(U2−V2)⏟factor excess A+(V2−D2)⏟matching excess B,U^2 - D^2 = \underbrace{(U^2 - V^2)}_{\text{factor excess } A} + \underbrace{(V^2 - D^2)}_{\text{matching excess } B},

which separates covariance-factor exactness/error (AA) from correspondence stability (BB).

• Single-order obstruction (Theorem 3). Extending HCP's topological example, we prove a quantitative Gaussian obstruction for any deterministic, label-invariant, partner-independent order assigning one contiguous interval per atom. Continuously swapping two atoms on a circle of radius r0>0r_0 > 0 returns to the same mixture, forcing a canonical order to jump; adjacent mixtures then satisfy DN=2r0sin⁡(π/(2N))→0D_N = 2r_0 \sin(\pi/(2N)) \to 0 while UN=VN=2r0cos⁡(π/(2N))→2r0U_N = V_N = 2r_0 \cos(\pi/(2N)) \to 2r_0 even with exact pricing (A=0A = 0).

On W2/Q2. Input-based conditions for small coupling excess (Sec. 4.3, App. A.4, A.6, C.3).

[Current manuscript, Secs. 4.3, 5.1, Apps. A.3, A.5, B.3]

• Sorting-margin guarantee (Theorem 2). Let γ>0\gamma > 0 be the minimum coordinate gap across the source sorter's recursive median cuts. Intuition. If target features satisfy yi=xi+(t,0)+eiy_i = x_i + (t, 0) + e_i with max⁡i∥ei∥∞<γ/2\max_i \Vert{}e_i\Vert{}_\infty < \gamma/2, no atom can cross any cut boundary after removing the common mean translation tt, so correspondence ranks stay identical. Unchanged weights give 0≤U2−D2≤∑iai∥ei∥20 \le U^2 - D^2 \le \sum_i a_i \Vert{}e_i\Vert{}^2 (plus an explicit interval-shift term if weights change), checkable in O(np)O(np) time for nn atoms with pp features, given recorded gaps and a correspondence. All 200 within-radius controls preserve ranks and satisfy the bound. On the same 50 perturbations at each level, the percentages preserving every rank (Hilbert/Lex) were 100/74 at 0.1γ0.1\gamma, 100/50 at 0.49γ0.49\gamma, 84/34 at γ\gamma, 50/14 at 4γ4\gamma, 22/8 at 16γ16\gamma and 12/2 at 64γ64\gamma.

• Noncollinear exactness (App. A.4). [Current manuscript, App. A.3] For diagonal covariances, if successive feature increments in both returned orders lie within a 45∘45^\circ cone around a common axis, their inner products are nonnegative, making the BW cost matrix Monge and U=DU = D.

• Reference certificate (App. A.4). [Current manuscript, Apps. A.3, B.3] For general SPD inputs, weighted isotonic regression fits monotone means along a supplied common axis, with a shared reference covariance. RMS feature residuals eP,eQe_P,e_Q give U−D≤2(eP+eQ)U-D\le2(e_P+e_Q) by the triangle inequality. Given the reference parameters and prepared features, the calculation is linear in their size. It certifies at most 10% relative error in 47/540 structured cases; 150/540 meet that threshold. None of the 140 random pairs has error at most 10% or a finite relative certificate.

AuthorsOpenReview ↗

Response to Reviewer ZQDt (Part 2/2)

On W2/Q3. Cholesky versus Principal Roots and Hilbert versus Lex (Sec. 4.2, 5.1, 5.3, App. A.5, C.1, C.4, Tables 3–4, 8).

[Current manuscript, Secs. 4.2, 5.1, 5.3, Apps. A.4, B.2, B.5, Tables 1, 3–4, 6]

• Pricing geometry (Proposition 1). For two Gaussian atoms, Cholesky pricing is exact iff their covariances share the same correlation matrix, whereas principal roots are exact iff ΣΛ=ΛΣ\Sigma\Lambda=\Lambda\Sigma and satisfy the known bound cBW≤cR≤2cBWc_{\mathrm{BW}}\le c_{\mathrm{R}}\le2c_{\mathrm{BW}} (Dinh et al., 2021).

• Controlled ablation (70 dense pairs). Crossing Cholesky (CC)/roots (RR) with Hilbert (HH)/Lex (LL) shows that CC is a faster default while RR reduces pricing error. Excess is mean 100(U/D−1)100(U/D-1); preparation/full times are medians for both mixtures.

Features/orderExcess (%)A/D2A/D^2B/D2B/D^2Prep./full ms
C/H60.210.3091.3171.060/1.262
C/L76.920.3301.9170.917/1.130
R/H51.490.0191.3322.189/2.392
R/L68.140.0201.9172.044/2.263

• Separating pricing from ordering. Repricing the fixed C/HC/H plan with RR reduces mean (U2−D2)/D2(U^2-D^2)/D^2 by 0.28910.2891, while switching to RR's Hilbert order adds back 0.01460.0146. The gain therefore comes from pricing. On 70 diagonal pairs, both factors give identical distances at 1.196/1.556 ms for Cholesky/roots. With Cholesky features, Hilbert is more accurate on 135 of the 140 pairs, and Lex has the shorter median full-call time in all 14 settings.

• Why Lex won on digits. The 8×88\times 8 block encoding gave every image identical grid centers, granting Lex exact block-to-block alignment. Perturbing component (x,y)(x,y) means independently by σ=10−4\sigma=10^{-4} breaks first-coordinate ties: Lex R@1 drops by 31.11–58.11 points across four datasets versus 3.22–6.00 for HMW. Shared offsets preserve Lex distances, isolating the correspondence effect.

On Q4. Complete preparation and sphere-uniform slicing (Sec. 5.2, Figure 3, Table 5, App. B.1, B.7, C.2).

[Current manuscript, Sec. 5.2, Figure 3, Table 5, Apps. B.1, B.4]

• Updated Figure 3 and Table 5. Figure 3 now plots full and simplified HMW alongside sphere-uniform DSMW/MSW on 28 common pairs. Full HMW includes duplicate merging and canonical enumeration and takes 1.83–3.76$\times$ as long as the simplified path; distances agree within 2.04×10−142.04\times10^{-14}. Full calls include preparation, matching and pricing; timings use one core.

(K,d)(K,d)Full HMW (ms)Sphere DSMW (ms)Sphere MSW (ms)MW/full HMW ratio
(100,3)1.0831.785.2911.62
(1000,3)4.16011.28256.25286.63
(1000,5)5.65612.46267.46418.81

MSW uses 16 sphere-uniform directions and DSMW 16 independent spatial/circle pairs, across five seeds. Speed ratios are medians of paired ratios. Switching from positive-orthant to sphere-uniform directions changes paired runtime by −3.00-3.00\\% to +0.31+0.31\\% (DSMW) and +0.81+0.81\\% to +5.54+5.54\\% (MSW); sphere DSMW setting medians remain 1.10–2.99$\times$ slower than full HMW.

AuthorsOpenReview ↗

Updated Manuscript References and Follow-up

Dear Reviewer ZQDt,

We’ve updated our earlier responses with references to the latest manuscript. The new references appear in bold alongside the original ones. If you prefer to use the earlier version of the paper, you can find it in the Supplementary Material. The shorter manuscript retains the analysis and experiments addressing all four of your questions. We shortened it by consolidating tables that covered the same ground and rewriting passages that were hard to follow.

As the discussion period draws to a close, we’d be grateful to hear whether our responses and additional experiments have addressed your concerns and whether you would consider revisiting your assessment in light of these changes. Thank you again for your time and for the care you put into your review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal and Discussion

Dear ACs, SACs, and PCs,

We sincerely thank you for handling our paper, and we greatly appreciate the reviewers' constructive feedback and recognition of our work. Given the length of the discussion, we have summarized the key points below for your convenience.

  • Summary of Our Response

    We addressed the reviewers' questions point by point and made the following updates:

    1. Clarifying HMW's Gaussian-specific contributions, local accuracy guarantees, and relationship to prior work.
    2. Adding factor/order ablations and full-call runtime comparisons, including sphere-uniform sliced baselines.
    3. Adding retrieval robustness and Pareto analyses on four image datasets.
    4. Expanding 3DGS comparisons to six methods across four scenes under common retention and recovery settings.
    5. Improving geometric intuition and readability and streamlining repeated material.

    Overall, we believe these concerns have been effectively resolved and should no longer serve as grounds for rejection.

  • Summary of Reviewer Feedback

    We are pleased to highlight that following the rebuttal process, our scores have been raised to 3, 3, 2. The detailed feedback is as follows:

    1. Reviewer ZQDt recognized the clear mathematical construction and the value of separating factor and matching errors, with no substantive concerns about soundness. We clarified the Gaussian-specific theoretical contributions and added controlled ablations of covariance factors and component ordering. Following our response, the reviewer explicitly confirmed: “My main concerns are addressed and I would like to keep my positive opinion on this paper.”
    2. Reviewer WkKw acknowledged the practical value of reusable representations and focused on the quality–speed trade-off. We added retrieval Pareto analyses and expanded 3DGS comparisons. In the follow-up, the reviewer stated that the Pareto analysis “demonstrates the strength of HMW” and described the 3DGS comparison as “a nice add-on for practical impact.” We also streamlined the manuscript in response to the readability feedback.
    3. Reviewer vGzY found the Gaussian Splatting application “very interesting” and raised questions about geometric intuition, topology, and positioning against prior work. We explained the covariance geometry and the conditions for local agreement of MW and HMW convergence, and clarified the distinctions from the six suggested references.

    In summary, Reviewers ZQDt and WkKw have given positive assessments of our work. Following the rebuttal, ZQDt explicitly confirmed that their main concerns were addressed and maintained their positive assessment, while WkKw acknowledged the strengths demonstrated by the Pareto analysis and additional 3DGS comparison. We have also responded point by point to vGzY's questions through theoretical explanations, literature comparisons, and additional experiments, further clarifying HMW's technical contributions and practical value.

We respectfully hope that the positive assessments from Reviewers ZQDt and WkKw, together with the expanded theoretical and experimental evidence, will be taken into account in evaluating the contribution and soundness of our work. We are grateful for the time and effort you and the reviewers have invested in our submission, and thank you again for coordinating the review process and for your fair and careful assessment.

Best regards,

Authors (Submission ID: 38938)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.