Pith. sign in

REVIEW 4 major objections 5 minor 53 references

Laplacian-regularized optimal transport provably produces cluster-aware couplings, and when the transport cost is block-constant the coupling is exactly block-constant with bounded non-negative rank.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 21:06 UTC pith:4XRW2Y75

load-bearing objection Sound theory for a known objective; the a posteriori bound is honest but unverified in most experiments, and the empirics are too thin to carry the practical claims. the 4 major comments →

arxiv 2607.16178 v1 pith:4XRW2Y75 submitted 2026-07-17 stat.ML cs.LGcs.NAmath.NAstat.ME

Cluster-Aware Matching via Laplacian Optimal Transport

Gabriel Samberg , YoonHaeng Hur , Yuehaw Khoo , Nir Sharon This is my paper
classification stat.ML cs.LGcs.NAmath.NAstat.ME
keywords optimal transportLaplacian regularizationcluster-aware matchingpoint cloud alignmentsimultaneous clusteringnon-negative rankspectral clusteringgraph Laplacian
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that when two point clouds carry an intrinsic cluster structure, matching should be done region-to-region rather than point-to-point, and that the right way to achieve this is to regularize entropic optimal transport with quadratic Laplacian terms built from each cloud's similarity graph. It proves that this Laplacian-regularized transport (LapOT) yields an optimal coupling that is close to its block-averaged form whenever the transport cost is nearly block-constant, the regularization weights are large, or the underlying graph Laplacians have large spectral gaps. In the exact block-constant case, the coupling is exactly block-constant and has non-negative rank bounded by the number of clusters. On top of this, the paper introduces Refined Simultaneous Clustering (RSC), which uses the LapOT coupling to produce consistent partitions of both clouds, and demonstrates on 3D shapes and stock-market data that this yields more meaningful alignments than independent clustering.

Core claim

The central claim is that the optimal coupling minimizing the LapOT objective — cost plus quadratic Laplacian penalties plus entropy — is approximately block-constant with respect to the connected components (or low-frequency eigenspaces) of the two similarity graph Laplacians. Theorem 1 bounds the Frobenius deviation from the block average by ||P_{X,r}C−C||∞/(λ_x μ_{r+1}) and ||C P_{Y,s}−C||∞/(λ_y μ_{s+1}). Consequently, if the cost is exactly block-constant, the coupling is exactly block-constant with non-negative rank at most min(r,s) (Corollary 1). Proposition 2 extends this to connected graphs: the coupling is close to its low-frequency projection whenever the a posteriori gap F(π*)−τ_λ

What carries the argument

The central object is the LapOT objective, q(π)=⟨π,C⟩+λ_x⟨π,L_X π⟩+λ_y⟨π,πL_Y⟩−λH(π), where L_X and L_Y are unnormalized graph Laplacians of similarity graphs on the two point clouds. The quadratic Laplacian terms penalize couplings whose rows vary for similar points in X and whose columns vary for similar points in Y, effectively a graph-smoothness prior on the transport plan. The proofs exploit the eigen-structure of the Laplacians: projecting onto the zero-eigenspace (or low-frequency subspace) corresponds to block-averaging rows/columns, and the spectral gap of the (r+1)-th eigenvalue controls how much the cost and entropy terms can distort the coupling.

Load-bearing premise

The guarantees become meaningful only if a hyperparameter regime exists in which the Laplacian regularization is strong enough to dominate the cost's fine-scale variation but weak enough to preserve the true cluster-level matching signal.

What would settle it

Solve the LapOT optimization exactly for similarity graphs with known connected components and a cost matrix that is exactly block-constant with respect to those partitions; if the optimal coupling fails to be exactly block-constant with non-negative rank at most min(r,s), Corollary 1 is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • When the transport cost is already determined at the cluster level (block-constant), the LapOT optimal coupling is exactly block-constant and decomposes into at most min(r,s) non-negative rank-one factors, so cluster-level matching is lossless.
  • Strengthening the Laplacian weights λ_x and λ_y drives the coupling toward its block average; in the infinite-regularization limit, the coupling becomes the entropic OT solution of the block-averaged cost P_{X,r} C P_{Y,s}.
  • Even when the similarity graphs are connected, the coupling remains close to a low-frequency structure whenever an a posteriori gap is small, so RSC can extract approximate cluster structure without exact connected components.
  • RSC, by clustering the rows and columns of the LapOT coupling and then refining the similarity graphs, yields partitions of both clouds that are aligned with each other, overcoming the instability of independent clustering.
  • In experiments on 3D shape alignment, cluster-wise matching via RSC remains accurate under higher noise levels than global distance-profile matching, and on correlated stock data it produces sector-like clusters aligned across two national markets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • We infer that the a posteriori bound of Proposition 2 can be inverted into a principled hyperparameter-selection rule: solve LapOT, evaluate F(π*)−τ_λ(C), and increase λ_x, λ_y until the bound is tight; the authors explicitly defer a systematic study.
  • We conjecture a cluster-level transport geometry: for block-constant costs, the exactly block-constant coupling defines a coarsened transport map between clusters that could serve as a hierarchical or coarse-to-fine alignment primitive.
  • The same Laplacian smoothing can be applied to multi-cloud matching by adding one quadratic Laplacian term per cloud, yielding a joint cluster-aware transport problem for three or more point sets.
  • The successful sector-like clustering of correlated stock returns suggests LapOT as a model-free cross-market alignment tool, but a controlled comparison against known sector labels would be needed to confirm that interpretation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Laplacian Optimal Transport (LapOT), which augments entropic OT with quadratic graph-Laplacian regularizers to encourage the transport coupling to respect cluster structure encoded by similarity graphs. The main theoretical results are: Theorem 1, showing that when the graphs have connected components and the marginals are block-constant, the optimal coupling is close to its block-averaged version with explicit non-asymptotic bounds (Eq. 11); Corollary 1, giving exact block-constant/low-rank structure when the cost is block-constant; Proposition 1, characterizing the infinite-regularization limits; and Proposition 2, an a posteriori bound for arbitrary Laplacians in terms of the gap between the LapOT and entropic-OT objective values. The paper also proposes Refined Simultaneous Clustering (RSC), which uses the LapOT coupling to define switch matrices, refine the similarity graphs, and then run spectral clustering. Experiments on 3D shape data (CAPOD), synthetic rotation alignment, and stock-market data are presented to support the method.

Significance. If the claims are substantiated, the paper offers a principled and potentially useful way to inject cluster structure into optimal transport, with self-contained proofs and a low-rank extension. Theorem 1's proof is clean and the bounds are non-asymptotic. Proposition 2 is a falsifiable computational certificate that could be valuable in practice. However, the significance depends on showing that the theoretical bridge (Proposition 2) actually operates in the regimes used in the experiments, and on stronger empirical validation. The current empirical support is largely qualitative or lacks statistical rigor, so the practical importance is not yet fully established.

major comments (4)
  1. [Section 3 / Section 5] Proposition 2 is the only theoretical result that applies to the connected RBF graphs and degree-based marginals used in the CAPOD and alignment experiments, but its bound is never evaluated in those experiments. The right-hand side involves the gap F(π*)−τ_λ(C); the paper gives no evidence that this gap is small relative to λx μ_{ℓ+1} and λy μ_{h+1} for any hyperparameter setting used. Figure 7 evaluates the bound only for the stock-market example. Without such a check, the claim that Proposition 2 provides the ‘theoretical mechanism behind the use of LapOT inside RSC’ (end of Section 3) is unsupported.
  2. [Section 2.2 / Section 4 Discussion] The bound in Proposition 2 becomes small when λx, λy are large, but large Laplacian weights can also overwhelm the matching signal in C, producing a smooth but cluster-agnostic coupling. The paper does not characterize or verify the existence of a hyperparameter regime in which the certificate is non-vacuous and the matching accuracy is preserved. Section 4 explicitly defers hyperparameter selection to future work, which leaves the central practical claim — that LapOT produces couplings that are both cluster-aware and match-preserving — without a demonstrated operating point.
  3. [Section 5.1, Table 1] Table 1 reports only mean relative rotation errors over ten runs, with no error bars, confidence intervals, or significance tests. At the highest noise level (SNR=2.91), the reported means for Our Method (0.20242) and Global DPM (0.25521) are close; without variance information it is impossible to assess whether the claimed robustness advantage is meaningful. This is a load-bearing empirical claim for the alignment application.
  4. [Sections 4 and 5, Figures 1–3, 8–9] The clustering results are evaluated only visually. No quantitative metric (e.g., adjusted Rand index, normalized mutual information, or correspondence consistency between the two point clouds) is reported to support the claim that RSC produces 'consistent and meaningful clusters' relative to independent clustering. Without a quantitative comparison, the central application of the paper is not empirically substantiated.
minor comments (5)
  1. [Appendix B] The derivation in Appendix B claims W1(μ_i, ν_j) ≥ |β_i − β_j|, relying on the equality 1/n Σ_k Cov(R_i, P_k/P_market R_k)/Var(R_market) = Cov(R_i, R_market)/Var(R_market). As defined, R_market is the return of the market index, which is not a price-weighted average of the individual P_k/P_market R_k terms in general. Please clarify the definition or correct the derivation.
  2. [Figure 7] The y-axis range (0 to 0.04) is not enough to judge whether the theoretical upper bound is small in an absolute sense. Reporting the actual values of F(π*)−τ_λ(C), λx, λy, and the relevant spectral gaps would strengthen the a posteriori certificate.
  3. [Section 2.3 / Algorithm 1] The low-rank LapOT variant is presented in detail but evaluated only qualitatively in Figure 3(c), with no comparison to the full-rank version in terms of runtime, accuracy, or cluster quality. A small quantitative comparison would help justify the added complexity.
  4. [Section 4] The switch-matrix refinement in Step 4 can remove all edges within clusters if k' is small, potentially disconnecting the graph. The paper does not discuss the sensitivity of RSC to k' or the interaction between k' and the final number of clusters k.
  5. [Throughout] All experiments use the solver from the authors' prior work [21] to compute π*. Please clarify whether the code is publicly available and provide the exact solver settings (tolerances, iteration counts) used, to support reproducibility.

Circularity Check

0 steps flagged

No significant circularity: LapOT theory is self-contained; the a-posteriori gap and hyperparameter regime are applicability caveats, not circular inputs.

full rationale

Theorem 1 (Eq. 11) is a genuine perturbation bound: the proof substitutes the block-averaged coupling P_{X,r}π* into the LapOT objective, uses feasibility from block-constant marginals, the Laplacian null-space identity L_X P_{X,r}=0, entropy concavity, and a spectral-gap lower bound; it does not assume the conclusion. Corollary 1 and Proposition 1 follow by direct substitution and compactness arguments. Proposition 2 is derived from the same optimality/spectral principle and is explicitly labeled "an a posteriori certificate" after solving LapOT, so it is not a fitted parameter renamed as a prediction; Figure 7 verifies the inequality numerically rather than predicting an unseen quantity. The self-citation [21] supplies a convex solver for (6) and is not load-bearing for the cluster-awareness theorem. The skeptical concern that the RHS of Proposition 2 may not be small for connected RBF graphs and degree-based marginals, and the paper's own statement that it leaves "a more detailed study of the hyperparameter selection for RSC to future work," are genuine validation/applicability gaps, not circular derivation. No step of the claimed derivation reduces to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on user-supplied similarity graphs, a set of unstated regularization hyperparameters, and compatibility between the matching cost and the graph structure. No genuinely new physical or model entities are introduced; the switch matrices in RSC are algorithmic constructs.

free parameters (5)
  • λx, λy, λ (LapOT regularization weights)
    Control the trade-off between matching cost and cluster-smoothness; values are not reported in the paper and are chosen separately for each experiment.
  • k and k' (RSC cluster counts) = k=5, k'=3 in Figure 1/4 experiments
    Number of final clusters and coarse switch clusters; selected by user; no selection rule beyond the heuristic comment about k' relative to k.
  • RBF kernel bandwidth σ for K_X, K_Y
    Defines similarity graphs that encode the cluster structure; exact bandwidths are not given.
  • rank r in low-rank LapOT = r=10 in Figure 3(c)
    Non-negative rank constraint for the low-rank extension; user-chosen.
  • optimization tuning (α, γ, δ, T)
    Stability parameter, step size, tolerance, and iteration count in Algorithm 3; not specified for experiments.
axioms (5)
  • standard math Unnormalized graph Laplacian L = diag(K1)-K for symmetric nonnegative K is PSD and its zero eigenspace is spanned by connected-component indicators.
    Used in Theorem 1 and Proposition 1 proofs; cited to spectral graph theory (Von Luxburg 2007).
  • standard math Shannon entropy H(P) = -Σ p_ij(log p_ij - 1) is concave, so block-averaging rows or columns of a coupling does not decrease entropy.
    Key inequality in Theorem 1 proof, used to drop the entropy difference term.
  • domain assumption The solver of Hur & Liang [21] converges to the minimizer of (6) with the claimed iteration complexity.
    All LapOT experiments rely on this external algorithm; the cited work is by one of the present paper's co-authors and no independent implementation is provided.
  • ad hoc to paper RBF similarity graphs with hand-picked bandwidth capture the meaningful cluster structure of the point clouds.
    All experiments construct K_X, K_Y via RBF kernels; if the graphs do not reflect true clusters, the Laplacian terms enforce the wrong structure.
  • domain assumption The matching cost C is compatible with the cluster structure: in the idealized regime it is nearly block-constant, and in the connected-graph regime the gap F(π*)-τ_λ(C) is small enough.
    Theorem 1 requires P_{X,r}C≈C (or exact block-constancy); Proposition 2's bound is useful only when the value gap is small. The paper does not verify this for the distance-profile or beta costs beyond one stock-market computation.

pith-pipeline@v1.3.0-alltime-deepseek · 23280 in / 16711 out tokens · 135391 ms · 2026-08-01T21:06:17.020982+00:00 · methodology

0 comments
read the original abstract

In many applications of matching, the point clouds to be matched are not merely unstructured sets of points but rather samples from distributions with an intrinsic cluster structure. In such cases, as individual points are often interchangeable within a coherent region, finding a robust region-to-region alignment is more desirable than establishing a precise point-to-point correspondence. To this end, we propose a novel approach for cluster-aware matching based on Laplacian Optimal Transport (LapOT). The key idea is to regularize the optimal transport problem with quadratic Laplacian terms constructed from similarity graphs of the point clouds, which encourages the optimal coupling to respect the cluster structure of both point sets. We also introduce Refined Simultaneous Clustering (RSC), a method that leverages the cluster-aware coupling obtained from LapOT to produce consistent partitions across the point sets, which can overcome the limitations of independent clustering and yield more stable and interpretable results. We demonstrate the effectiveness of our approach through theoretical analysis and empirical experiments, showing that LapOT indeed produces cluster-aware matching that leads to more consistent and meaningful alignments between point clouds.

Figures

Figures reproduced from arXiv: 2607.16178 by Gabriel Samberg, Nir Sharon, YoonHaeng Hur, YueHaw Khoo.

Figure 1
Figure 1. Figure 1: Clustering of two 3D human shapes from the CAPOD dataset [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Clustering of two 3D human shapes according to the clustering structure encoded in the switch [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Other examples of RSC applied to two point clouds from the CAPOD dataset. The first row shows [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Point clouds and the result of RSC. (a) shows the point cloud of a human (left) and the same [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Normalized daily stock prices for the top 50 companies from the S&P 500 (left) and top 50 Japanese [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The solution of LapOT applied to the US-Japan stock market data. We permuted the rows and [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison between the theoretical upper bound and the projection error [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The cluster structure of the top 50 stocks from S&P 500 found by the RSC. We see the rough [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The cluster structure of the top 50 Japanese stocks found by the RSC. We see the rough partition [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 5 linked inside Pith

  1. [1]

    Gromov-Wasserstein alignment of word embedding spaces

    David Alvarez-Melis and Tommi S Jaakkola. Gromov-Wasserstein alignment of word embedding spaces. arXiv preprint arXiv:1809.00013, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    Laplacian eigenmaps and spectral techniques for embedding and clustering

    Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps and spectral techniques for embedding and clustering. InAdvances in Neural Information Processing Systems, 2001

    2001

  3. [3]

    A sober look at clustering stability

    Shai Ben-David, Ulrike Von Luxburg, and D´ avid P´ al. A sober look at clustering stability. InConference on Learning Theory, 2006

    2006

  4. [4]

    A method for registration of 3-D shapes.IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):239–256, 1992

    Paul J Besl and Neil D McKay. A method for registration of 3-D shapes.IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):239–256, 1992

    1992

  5. [5]

    MREC: A fast and versatile framework for aligning and matching point clouds with applications to single cell molecular data.arXiv preprint arXiv:2001.01666, 2020

    Andrew J Blumberg, Mathieu Carriere, Michael A Mandell, Raul Rabadan, and Soledad Villar. MREC: A fast and versatile framework for aligning and matching point clouds with applications to single cell molecular data.arXiv preprint arXiv:2001.01666, 2020

    Pith/arXiv arXiv 2001

  6. [6]

    McGraw Hill, 2011

    Richard A Brealey, Stewart C Myers, and Franklin Allen.Principles of Corporate Finance. McGraw Hill, 2011

    2011

  7. [7]

    Learning generative models across incomparable spaces

    Charlotte Bunne, David Alvarez-Melis, Andreas Krause, and Stefanie Jegelka. Learning generative models across incomparable spaces. InInternational Conference on Machine Learning, 2019

    2019

  8. [8]

    Diffusion maps.Applied and Computational Harmonic Analysis, 21(1):5–30, 2006

    Ronald R Coifman and St´ ephane Lafon. Diffusion maps.Applied and Computational Harmonic Analysis, 21(1):5–30, 2006

    2006

  9. [9]

    Optimal transport for domain adaptation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1853–1865, 2017

    Nicolas Courty, R´ emi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1853–1865, 2017. 16

    2017

  10. [10]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems, 2013

    2013

  11. [11]

    Efficient random graph matching via degree profiles.Probability Theory and Related Fields, 179(1):29–115, 2021

    Jian Ding, Zongming Ma, Yihong Wu, and Jiaming Xu. Efficient random graph matching via degree profiles.Probability Theory and Related Fields, 179(1):29–115, 2021

    2021

  12. [12]

    On the existence of Monge maps for the Gromov-Wasserstein problem.Foundations of Computational Mathematics, 25(2):463–510, 2025

    Th´ eo Dumont, Th´ eo Lacombe, and Fran¸ cois-Xavier Vialard. On the existence of Monge maps for the Gromov-Wasserstein problem.Foundations of Computational Mathematics, 25(2):463–510, 2025

    2025

  13. [13]

    An algorithm for restricted least squares regression.Journal of the American Statistical Association, 78(384):837–842, 1983

    Richard L Dykstra. An algorithm for restricted least squares regression.Journal of the American Statistical Association, 78(384):837–842, 1983

    1983

  14. [14]

    Regularized discrete optimal transport.SIAM Journal on Imaging Sciences, 7(3):1853–1882, 2014

    Sira Ferradans, Nicolas Papadakis, Gabriel Peyr´ e, and Jean-Fran¸ cois Aujol. Regularized discrete optimal transport.SIAM Journal on Imaging Sciences, 7(3):1853–1882, 2014

    2014

  15. [15]

    Statistical optimal transport via factored couplings

    Aden Forrow, Jan-Christian H¨ utter, Mor Nitzan, Philippe Rigollet, Geoffrey Schiebinger, and Jonathan Weed. Statistical optimal transport via factored couplings. InInternational Conference on Artificial Intelligence and Statistics, 2019

    2019

  16. [16]

    Unsupervised alignment of embeddings with Wasserstein procrustes

    Edouard Grave, Armand Joulin, and Quentin Berthet. Unsupervised alignment of embeddings with Wasserstein procrustes. InInternational Conference on Artificial Intelligence and Statistics, 2019

    2019

  17. [17]

    On the added value of bootstrap analysis for K-means clustering.Journal of Classification, 32(2):268–284, 2015

    Joeri Hofmans, Eva Ceulemans, Douglas Steinley, and Iven Van Mechelen. On the added value of bootstrap analysis for K-means clustering.Journal of Classification, 32(2):268–284, 2015

    2015

  18. [18]

    Statistical significance of clustering using soft thresholding.Journal of Computational and Graphical Statistics, 24(4):975–993, 2015

    Hanwen Huang, Yufeng Liu, Ming Yuan, and JS Marron. Statistical significance of clustering using soft thresholding.Journal of Computational and Graphical Statistics, 24(4):975–993, 2015

    2015

  19. [19]

    Reversible Gromov–Monge sampler for simulation-based inference.SIAM Journal on Mathematics of Data Science, 6(2):283–310, 2024

    YoonHaeng Hur, Wenxuan Guo, and Tengyuan Liang. Reversible Gromov–Monge sampler for simulation-based inference.SIAM Journal on Mathematics of Data Science, 6(2):283–310, 2024

    2024

  20. [20]

    Robust point matching with distance profiles.Journal of Machine Learning Research, 26(205):1–38, 2025

    YoonHaeng Hur and Yuehaw Khoo. Robust point matching with distance profiles.Journal of Machine Learning Research, 26(205):1–38, 2025

    2025

  21. [21]

    A convexified matching approach to imputation and individual- ized inference.arXiv preprint arXiv:2407.05372, 2024

    YoonHaeng Hur and Tengyuan Liang. A convexified matching approach to imputation and individual- ized inference.arXiv preprint arXiv:2407.05372, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Inference for clustering: Conformal sets for cluster labels.arXiv preprint arXiv:2604.03488, 2026

    YoonHaeng Hur, Anirban Nath, and Genevera Allen. Inference for clustering: Conformal sets for cluster labels.arXiv preprint arXiv:2604.03488, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    M. K. Kerr and G. A. Churchill. Bootstrapping cluster analysis: Assessing the reliability of conclusions from microarray experiments.Proceedings of the National Academy of Sciences, 98(16):8961–8965, 2001

    2001

  24. [24]

    Assignment problems and the location of economic activities.Econometrica, pages 53–76, 1957

    Tjalling C Koopmans and Martin Beckmann. Assignment problems and the location of economic activities.Econometrica, pages 53–76, 1957

    1957

  25. [25]

    Mechanisms to improve clustering uncertain data with UKmeans.Data & Knowledge Engineering, 116:1–18, 2018

    Chuen-Ming Liu, Zhi-Ping Niu, and Kuan-Ting Liao. Mechanisms to improve clustering uncertain data with UKmeans.Data & Knowledge Engineering, 116:1–18, 2018

    2018

  26. [26]

    A survey for the quadratic assignment problem.European Journal of Operational Research, 176(2):657–690, 2007

    Eliane Maria Loiola, Nair Maria Maia De Abreu, Paulo Oswaldo Boaventura-Netto, Peter Hahn, and Tania Querido. A survey for the quadratic assignment problem.European Journal of Operational Research, 176(2):657–690, 2007

    2007

  27. [27]

    Gromov-Wasserstein distances and the metric approach to object matching.Founda- tions of Computational Mathematics, 11(4):417–487, 2011

    Facundo M´ emoli. Gromov-Wasserstein distances and the metric approach to object matching.Founda- tions of Computational Mathematics, 11(4):417–487, 2011

    2011

  28. [28]

    Comparison results for Gromov-Wasserstein and Gromov-Monge distances.ESAIM: Control, Optimisation and Calculus of Variations, 30:78, 2024

    Facundo M´ emoli and Tom Needham. Comparison results for Gromov-Wasserstein and Gromov-Monge distances.ESAIM: Control, Optimisation and Calculus of Variations, 30:78, 2024

    2024

  29. [29]

    Point set registration: Coherent point drift.IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(12):2262–2275, 2010

    Andriy Myronenko and Xubo Song. Point set registration: Coherent point drift.IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(12):2262–2275, 2010. 17

    2010

  30. [30]

    Weighted conformal clustering.arXiv preprint arXiv:2606.00436, 2026

    Anirban Nath, YoonHaeng Hur, and Genevera I Allen. Weighted conformal clustering.arXiv preprint arXiv:2606.00436, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    The canonically posed 3d objects dataset

    Panagiotis Papadakis. The canonically posed 3d objects dataset. InEurographics Workshop on 3D Object Retrieval, pages 33–36, 2014

    2014

  32. [32]

    Computational optimal transport: With applications to data science

    Gabriel Peyr´ e and Marco Cuturi. Computational optimal transport: With applications to data science. Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019

    2019

  33. [33]

    Gromov-Wasserstein averaging of kernel and distance matrices

    Gabriel Peyr´ e, Marco Cuturi, and Justin Solomon. Gromov-Wasserstein averaging of kernel and distance matrices. InInternational Conference on Machine Learning, 2016

    2016

  34. [34]

    Stability k-means clustering

    Alexander Rakhlin and Andrea Caponnetto. Stability k-means clustering. InAdvances in Neural Information Processing Systems, 2006

    2006

  35. [35]

    Aryan Tajmir Riahi, Geoffrey Woollard, Fr´ ed´ eric Poitevin, Anne Condon, and Khanh Dao Duc. Alig- nOT: An optimal transport based algorithm for fast 3d alignment with applications to cryogenic elec- tron microscopy density maps.IEEE/ACM Transactions on Computational Biology and Bioinformatics, 20(6):3842–3850, 2023

    2023

  36. [36]

    Fast point feature histograms (FPFH) for 3d registration.IEEE International Conference on Robotics and Automation, pages 3212–3217, 2009

    Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (FPFH) for 3d registration.IEEE International Conference on Robotics and Automation, pages 3212–3217, 2009

    2009

  37. [37]

    Recent advances in shape correspondence.The Visual Computer, 36(8):1705–1721, 2020

    Yusuf Sahillio˘ glu. Recent advances in shape correspondence.The Visual Computer, 36(8):1705–1721, 2020

    2020

  38. [38]

    Low-rank optimal transport: Approximation, statistics and debiasing

    Meyer Scetbon and Marco Cuturi. Low-rank optimal transport: Approximation, statistics and debiasing. InAdvances in Neural Information Processing Systems, 2022

    2022

  39. [39]

    Low-rank Sinkhorn factorization

    Meyer Scetbon, Marco Cuturi, and Gabriel Peyr´ e. Low-rank Sinkhorn factorization. InInternational Conference on Machine Learning, 2021

    2021

  40. [40]

    Geoffrey Schiebinger, Jian Shu, Marcin Tabaka, Brian Cleary, Vidya Subramanian, Aryeh Solomon, Joshua Gould, Siyan Liu, Stacie Lin, Peter Berube, Lia Lee, Jenny Chen, Justin Brumbaugh, Philippe Rigollet, Konrad Hochedlinger, Rudolf Jaenisch, Aviv Regev, and Eric S. Lander. Optimal-transport analysis of single-cell gene expression identifies developmental ...

    2019

  41. [41]

    Generalized-ICP

    Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Generalized-ICP. InRobotics: Science and Systems. The MIT Press, 2010

    2010

  42. [42]

    Critical limitations of consensus clustering in class discovery.Scientific Reports, 4(1):6207, 2014

    Yasin Senbabao˘ glu, George Michailidis, and Jun Z Li. Critical limitations of consensus clustering in class discovery.Scientific Reports, 4(1):6207, 2014

    2014

  43. [43]

    Alignment of density maps in Wasserstein distance.Biological Imaging, 4:e5, 2024

    Amit Singer and Ruiyi Yang. Alignment of density maps in Wasserstein distance.Biological Imaging, 4:e5, 2024

    2024

  44. [44]

    Lack of evidence for the transitional cerebellar progenitor.Nature, 643(8071):E1–E8, 2025

    Kyle S Smith, Yiran Li, Parthiv Haldipur, Brian L Gudenas, Kathleen J Millen, Volker Hovestadt, and Paul A Northcott. Lack of evidence for the transitional cerebellar progenitor.Nature, 643(8071):E1–E8, 2025

    2025

  45. [45]

    Entropic metric alignment for corre- spondence problems.ACM Transactions on Graphics, 35(4):1–13, 2016

    Justin Solomon, Gabriel Peyr´ e, Vladimir G Kim, and Suvrit Sra. Entropic metric alignment for corre- spondence problems.ACM Transactions on Graphics, 35(4):1–13, 2016

    2016

  46. [46]

    A survey on shape corre- spondence.Computer Graphics Forum, 30(6):1681–1707, 2011

    Oliver Van Kaick, Hao Zhang, Ghassan Hamarneh, and Daniel Cohen-Or. A survey on shape corre- spondence.Computer Graphics Forum, 30(6):1681–1707, 2011

    2011

  47. [47]

    American Mathematical Society, 2003

    C´ edric Villani.Topics in Optimal Transportation. American Mathematical Society, 2003

    2003

  48. [48]

    A tutorial on spectral clustering.Statistics and Computing, 17(4):395–416, 2007

    Ulrike Von Luxburg. A tutorial on spectral clustering.Statistics and Computing, 17(4):395–416, 2007. 18

    2007

  49. [49]

    Clustering stability: An overview.Foundations and Trends®in Machine Learning, 2(3):235–274, 2010

    Ulrike Von Luxburg. Clustering stability: An overview.Foundations and Trends®in Machine Learning, 2(3):235–274, 2010

    2010

  50. [50]

    Bayesian cluster analysis: Point estimation and credible balls (with discussion).Bayesian Analysis, 13(2):559–626, 2018

    Sara Wade and Zoubin Ghahramani. Bayesian cluster analysis: Point estimation and credible balls (with discussion).Bayesian Analysis, 13(2):559–626, 2018

    2018

  51. [51]

    The use of entropy maximising models, in the theory of trip distribution, mode split and route split.Journal of Transport Economics and Policy, pages 108–126, 1969

    Alan Geoffrey Wilson. The use of entropy maximising models, in the theory of trip distribution, mode split and route split.Journal of Transport Economics and Policy, pages 108–126, 1969

    1969

  52. [52]

    Optimal transport for brain-image alignment: Unveiling redundancy and synergy in neural information processing

    Yang Xiao, Wang Lu, Jie Ji, Ruimeng Ye, Gen Li, Xiaolong Ma, and Bo Hui. Optimal transport for brain-image alignment: Unveiling redundancy and synergy in neural information processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20445–20455, 2025

    2025

  53. [53]

    Gromov-Wasserstein learning for graph matching and node embedding

    Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-Wasserstein learning for graph matching and node embedding. InInternational Conference on Machine Learning, 2019. A Optimization of Low-Rank LapOT This section discusses the optimization scheme for solving the low-rank version of the LapOT problem (10). As noticed in [39], the division ...

    2019