Pith. sign in

REVIEW 2 major objections 8 minor 300 references

The paper introduces the distributional determinantal point process (dDPP), a repulsive point process whose atoms are probability distributions, and shows it yields well-separated, interpretable clusters of distribution-valued data.

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 06:32 UTC pith:IFG6SQWB

load-bearing objection A useful repulsive point-process prior for distribution-valued data with a real but fixable gap in the posterior-consistency proof. the 2 major comments →

arxiv 2607.21847 v1 pith:IFG6SQWB submitted 2026-07-23 stat.ME cs.LGstat.APstat.COstat.ML

Distributional Determinantal Point Process for Repulsive Clustering of Distributions

Khai Nguyen , Yang Ni , Elizabeth Juarez-Colunga , Peter Mueller This is my paper
classification stat.ME cs.LGstat.APstat.COstat.ML MSC 62F1560G5562H30
keywords distributional determinantal point processrepulsive priorsliced Wasserstein kernelgeneralized Bayesrandom partitionsposterior consistencydistributional clusteringempirical concentration
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's central aim is to make repulsive clustering work directly on probability distributions, not just on points. It defines the dDPP through an L-ensemble with a sliced-Wasserstein kernel, proves this is a well-defined point process on a compact space of distributions, and shows plug-in estimates from i.i.d. samples converge at a parametric rate. The dDPP is then used as a prior over atoms in a generalized Bayesian mixture model, with a proof that the posterior concentrates on the best mixing measure as data size grows. This matters because distribution-valued data (gene-expression profiles, seizure diaries) often need clusters that are separated and interpretable, and previous nonparametric Bayesian methods tend to over-fragment. On two real datasets, the dDPP yields fewer, better-separated clusters than a Dirichlet-process baseline with essentially no loss of likelihood.

Core claim

The central claim is that the distributional determinantal point process (dDPP), defined by the Janossy density j_k(ν_1,...,ν_k) ∝ det(L(ν_i,ν_j)) with L(ν,ν') = exp(-γ SW2^2(ν,ν')), is a valid repulsive point process on a compact set of probability distributions equipped with the sliced-Wasserstein topology. Because sliced-Wasserstein distance is Hilbertian, the kernel is positive definite, and the associated integral operator is trace-class, so the normalization constant is finite and the process is well-defined. In the discrete setting the paper derives dimension-free O(m^{-1/2}) concentration bounds for plug-in estimators of the L-ensemble, correlation kernel, and their determinants. In

What carries the argument

The load-bearing object is the sliced-Wasserstein kernel L(ν,ν') = exp(-γ SW2^2(ν,ν')), whose positive definiteness follows from the Hilbertian property SW2^2(ν,ν') = ||Ψ(ν)-Ψ(ν')||^2 with an explicit feature map Ψ into L^2(S^{d-1}×[0,1]). This kernel defines an L-ensemble whose Janossy density det(L(ν_i,ν_j)) penalizes configurations containing close atoms. The correlation kernel K = L(I+L)^{-1} and its Schur complement ∆K(θ*,θ_new) quantify how strongly a candidate atom is penalized for being near occupied atoms. Trace-class of the integral operator T_L (via spectral decomposition) ensures the Fredholm determinant is finite, making the point process well-defined.

Load-bearing premise

The posterior-concentration proof requires that every configuration of distinct distributions has a strictly positive definite sliced-Wasserstein kernel matrix; the paper proves positive definiteness of the kernel but not strictness on distinct points.

What would settle it

Compute the eigenvalues of the sliced-Wasserstein Gram matrix for a finite set of distinct distributions that are close in SW distance but not identical—for example, several Gaussians with different means and variances. If any eigenvalue is zero to numerical precision, the strict positive definiteness needed for the full-support argument in the proof of Theorem 3 fails. More decisively, find a configuration {ν_1,...,ν_k} of distinct distributions with det(L(ν_i,ν_j))=0; then the dDPP assigns zero probability to that configuration, contradicting the full-support assertion.

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

If this is right

  • Provides a principled repulsive prior for mixture models over distribution-valued data, addressing over-fragmentation in nonparametric Bayesian clustering.
  • Justifies replacing true distributions by empirical measures in practice: the concentration bounds guarantee the DPP probabilities are estimated at a parametric rate.
  • Guarantees statistical consistency: the posterior of the mixing weights converges to the population risk minimizer, so inferences are anchored as data accumulate.
  • Gives a practical MCMC scheme: the marginal sampler uses precomputed SW kernel matrices and rank-one Cholesky updates, with computational cost polynomial in the number of atoms.
  • On single-cell and epilepsy data, dDPP partitions are more parsimonious and better separated than DP partitions, with near-identical generalized likelihood.

Where Pith is reading between the lines

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

  • The proof of posterior consistency assumes strict positive definiteness of the SW kernel on distinct distributions, but the paper only proves positive definiteness; verifying strict positive definiteness (or finding a counterexample) would settle whether the full-support step in the proof is sound.
  • The same L-ensemble construction could be applied to other Hilbertian divergences, such as the energy distance, to obtain different repulsive structure on distribution spaces.
  • The finite discrete atom set Θ simplifies computation and theory; extending to a continuous base measure with slice sampling could allow inference over an infinite space of distributions, at the cost of revisiting the trace-class and consistency arguments.
  • The paper's own limitation note that generalized likelihood may deviate from fully model-based Bayesian inference suggests a natural next step: building a generative likelihood for distribution-valued atoms that preserves the SW repulsion.

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

2 major / 8 minor

Summary. The paper introduces the distributional determinantal point process (dDPP), a repulsive point process whose atoms are probability distributions, constructed via an L-ensemble with a sliced Wasserstein (SW) kernel. It proves well-definedness of the dDPP on a compact space of distributions (Theorem 1), derives O(m^{-1/2}) concentration bounds for plug-in estimators of the L and correlation kernels using empirical measures (Theorem 2 and Corollary 1), and proposes a generalized Bayesian mixture model with a dDPP prior over atoms and an SW-based generalized likelihood. A posterior consistency theorem for the mixing weights (Theorem 3) is stated and proved in the supplement. The paper also develops marginal MCMC inference, a decision-theoretic posterior summarization using sliced Wasserstein barycenters, and applies the method to single-cell and epilepsy data.

Significance. If the results are correct, this is a useful extension of DPP priors to distribution-valued data, with theoretical guarantees that go beyond existing repulsive mixture models. The concentration bounds are dimension-free, the computational strategy exploits the SW distance effectively, and the applications address relevant scientific problems. The posterior consistency theorem is a central contribution, as it provides a misspecified-Bayes justification for the proposed loss-based inference. However, the proof of Theorem 3 contains an unsupported strict positive definiteness claim that is load-bearing; the gap is likely repairable but must be fixed before the result can be accepted.

major comments (2)
  1. [Supplementary A.4, Eq. (75)] The full-support argument for the dDPP prior asserts that 'the SW kernel matrix is strictly positive definite on distinct points', so that det(L(ν_p,ν_q))>0 for every subconfiguration. Theorem 1 (Supplementary A.1) proves only positive definiteness (non-negativity of the quadratic form) via a power-series expansion; it does not rule out zero eigenvalues for distinct distributions. This strictness is essential: Eq. (82) requires Π(U*_ρ)>0 for every ρ, and if some distinct subconfiguration had a singular SW Gram matrix, the dDPP could assign zero probability to that subconfiguration, potentially zero mass to open subsets of Δ^{Q-1}. Thus the proof of Theorem 3 is incomplete. The gap is likely fixable by adding a lemma that the feature map Ψ(ν)=F^{-1}_{P_v♯ν}(·) is injective (by Cramér-Wold) and that the Gaussian RBF kernel exp(-γ||Ψ(ν)-Ψ(ν')||^2) on L^2(S^{d-1}×[0,1]) is strictly positive
  2. [Assumption 1, Section 3.1] The two sufficient conditions listed under Assumption 1 (uniform tightness and uniform integrability of second moments) imply relative compactness of Θ in W2, not compactness; closedness of Θ in the relevant topology is also required. Since the assumption itself states that Θ is compact, this is not fatal to Theorem 1, which proceeds under that assumption, but the phrasing 'This holds if...' is technically inaccurate. Please either add closedness as a third condition or explicitly state that the conditions are meant to establish relative compactness and that Θ is taken to be its own closure.
minor comments (8)
  1. [Section 1, notation] In the notation paragraph, 'A vector (π_1,...,π_k)∈Δ_k satisfies π_k≥0 for each k = 1,...,k' uses the same symbol k for both the dimension and the index. It should be something like 'π_j≥0 for each j=1,...,k'.
  2. [Section 6.1, first paragraph] Typo: 'we used the D2PP for inference' should be 'dDPP'.
  3. [Supplementary A.2, equations (32)-(36)] There are OCR-like typographical errors: 'ˆθo,(m)' appears instead of 'ˆν_i^{(m)}' in several equations, and the subscript 'i' is missing in 'P_v♯ˆθo,(m)' and similar expressions. These make the proof harder to follow.
  4. [Section 5.2, update assignments] The sentence 'For new atoms, we propose naux (is later set to 1) candidates...' is awkward; consider rewriting to 'we propose naux candidates, with naux later set to 1' or similar.
  5. [Section 4, Eq. (6)] The parameter w is called a 'bandwidth parameter'. In a generalized likelihood, w is better described as an inverse-temperature/scale parameter; 'bandwidth' is usually reserved for kernels such as γ in the SW kernel. Please adjust the terminology for clarity.
  6. [Section 5.1, discussion after Eq. (10)] The statement 'Since SW2(θ_new,θ_new)=0, we have ΔK∈[0,1]' is imprecise. The fact that ΔK is a Schur complement of a positive definite kernel matrix implies ΔK∈[0,1] because K(θ_new,θ_new)=1 and the subtracted term is nonnegative; the SW2(θ_new,θ_new)=0 observation alone is not the reason. Please clarify.
  7. [References] The references 'Peyré & Cuturi (2019)' and 'Peyré, G., Cuturi, M. et al. (2019)' are duplicated and should be merged into a single entry.
  8. [Section 7, Conclusion] The authors note two limitations: the generalized-Bayes nature and the finite discrete base measure. These are appropriately acknowledged, but the strict positive definiteness gap in Theorem 3 is not mentioned there; it should be addressed in the main text or supplement.

Circularity Check

0 steps flagged

No significant circularity: the dDPP construction, plug-in concentration bounds, and posterior-consistency theorem are self-contained; only a minor non-load-bearing self-citation and an unproven strict-positive-definiteness assertion in Theorem 3's proof.

full rationale

The dDPP is defined directly as a DPP with an SW kernel (Def. 1, Eq. 3); Theorem 1 is a standard positive-definiteness plus trace-class argument and does not assume its conclusion. Theorem 2 derives plug-in concentration from DKW and McDiarmid inequalities; the bounds are not fitted to the quantities they bound, and Corollary 1 is a determinant perturbation bound with explicit constants. Theorem 3 is a misspecified-Bayes consistency theorem: it defines M* as the minimizer of the population risk of the same SW generalized likelihood used in the model, and proves posterior concentration on M* via a uniform law of large numbers and prior full support. This is a self-contained mathematical statement, not an external validation of the model; targeting the same loss as the likelihood is not circular because the theorem does not claim the loss-minimizer equals some external truth. The only notable concern is Eq. (75) in the proof of Theorem 3, where the proof asserts that 'the SW kernel matrix is strictly positive definite on distinct points' to establish full support of the dDPP prior; Theorem 1 proves only positive definiteness, so this is an omitted proof (a correctness risk), not a circular reduction. The self-citation to Nguyen & Mueller (2026) in Sec. 5.3 underlies the posterior-summarization framework, but the hierarchical optimal-transport utility and the Bayes-rule estimator (11) are new; the citation is not load-bearing. Overall, no derivation step reduces to its own input by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The dDPP is a new stochastic model, not a new physical entity. The central claim depends on the hand-chosen hyperparameters gamma, w, and alpha, plus the compactness and exponential-moment assumptions that make the theorems go through. No particles, forces, or new conserved quantities are introduced.

free parameters (3)
  • gamma (SW kernel bandwidth) = 10^3, 10^4, 10^5 (single-cell); 10, 100, 1000 (HEP)
    Controls the prior repulsion strength in L(nu,nu') = exp(-gamma * SW2^2); chosen by hand over a grid, not estimated.
  • w (generalized likelihood bandwidth) = 5e5, 7e5, 1e6 (single-cell); 100, 200, 500 (HEP)
    Controls concentration of l(F|theta) = exp(-w * SW2^2); chosen relative to the median pairwise SW2^2 distance; grid reported.
  • alpha (Gamma mark shape) = Not stated in experiments
    Shape of s_h ~ Gamma(alpha,1) in model (7); affects cluster sizes and prior weight distribution; no value or sensitivity analysis is given in the main text.
axioms (6)
  • domain assumption Assumption 1: Theta is compact under SW2
    Required for Theorem 1 to apply Mercer's theorem and show trace-class; trivial in the finite-discrete setting used in applications, but excludes unbounded supports in the general formulation.
  • standard math Mercer's theorem on compact metric spaces
    Used to spectrally decompose L and to identify det(I + T_L) with prod(1 + lambda_i) in the proof of Theorem 1.
  • domain assumption SW2 metrizes weak convergence on P2(Rd)
    Invoked to establish continuity of the SW kernel and empirical convergence; cited to Nadjahi et al. (2020).
  • standard math DPP existence theory for L-ensembles with trace-class kernels
    Definition 1 and Theorem 1 rely on the standard result that a symmetric positive definite trace-class kernel defines a DPP; the paper does not verify Janossy consistency directly.
  • domain assumption Strict positive definiteness of the SW Gram matrix on distinct distributions
    Asserted in Supplementary A.4 for full support of the dDPP prior; not proved, load-bearing for Theorem 3.
  • domain assumption Exponential moment condition (8)
    Assumed for finiteness of the population risk and the proof of the uniform strong law in Theorem 3.

pith-pipeline@v1.3.0-alltime-deepseek · 37519 in / 18774 out tokens · 189654 ms · 2026-08-01T06:32:30.604814+00:00 · methodology

0 comments
read the original abstract

We introduce the distributional determinantal point process (dDPP) as a novel repulsive point process whose atoms are probability distributions rather than points in a real space. The dDPP is constructed via an L-ensemble with a sliced Wasserstein (SW) kernel between distributions. We show its validity as a well-defined point process. In the discrete setting, we derive concentration results for plug-in estimators of the L-ensemble, the correlation kernel, and their determinants given i.i.d. samples from the distributional atoms. Leveraging this framework, we propose a distribution-valued random partition model by way of a repulsive generalized Bayesian mixture model. The model places a dDPP prior over the atoms of the mixing measure and defines a generalized likelihood based on SW distance. To summarize posterior inference, we develop a decision-theoretic approach to report a point estimate of the mixing measure as a Bayes rule under a hierarchical optimal transport utility function. The latter is a natural choice given that the mixing measure is itself a distribution over distributions. We use the proposed framework for inference with single-cell gene expression data and human epilepsy data, producing interpretable and well-separated clusters that reflect meaningful structure in the data.

Figures

Figures reproduced from arXiv: 2607.21847 by Elizabeth Juarez-Colunga, Khai Nguyen, Peter Mueller, Yang Ni.

Figure 1
Figure 1. Figure 1: Single cell data. UMAP visualization based on point estimates [PITH_FULL_IMAGE:figures/full_fig_p023_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Singe cell data - PCA visualization of the atoms of the mixing measure. The [PITH_FULL_IMAGE:figures/full_fig_p024_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: HEP data. Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: HEP data. The atoms of the point estimate [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: HEP data. Visualization of all patients with corresponding clusters label. Similar [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p049_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: PCA visualization of some donors in clusters the dDPP model ( [PITH_FULL_IMAGE:figures/full_fig_p052_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: PCA visualization of some donors in clusters the dDPP model ( [PITH_FULL_IMAGE:figures/full_fig_p053_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: PCA visualization of some donors in clusters the dDPP model ( [PITH_FULL_IMAGE:figures/full_fig_p054_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: PCA visualization of some donors in clusters the dDPP model ( [PITH_FULL_IMAGE:figures/full_fig_p055_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p056_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p057_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p058_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p059_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Same as Figure [PITH_FULL_IMAGE:figures/full_fig_p060_15.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

300 extracted references · 21 linked inside Pith

  1. [1]

    Single-cell

    Yazar, Seyhan and Alquicira-Hernandez, Jose and Wing, Kristof and Senabouth, Anne and Gordon, M Grace and Andersen, Stacey and Lu, Qinyi and Rowson, Antonia and Taylor, Thomas RP and Clarke, Linda and others , journal=. Single-cell. 2022 , publisher=

    2022

  2. [2]

    Summarizing Nonparametric

    Nguyen, Khai and Mueller, Peter , journal=. Summarizing Nonparametric. 2026 , publisher=

    2026

  3. [3]

    Bayesian Analysis , volume=

    Robust probabilistic inference via a constrained transport metric , author=. Bayesian Analysis , volume=. 2025 , publisher=

    2025

  4. [4]

    Reversible jump

    Green, Peter J , journal=. Reversible jump. 1995 , publisher=

    1995

  5. [5]

    Journal of Computational and Graphical Statistics , volume=

    Search algorithms and loss functions for Bayesian clustering , author=. Journal of Computational and Graphical Statistics , volume=. 2022 , publisher=

    2022

  6. [6]

    2017 , publisher=

    Fundamentals of nonparametric Bayesian inference , author=. 2017 , publisher=

    2017

  7. [7]

    Journal of the Royal Statistical Society: Series B (Methodological) , volume=

    A general class of coefficients of divergence of one distribution from another , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1966 , publisher=

    1966

  8. [8]

    2021 , publisher=

    Introduction to linear regression analysis , author=. 2021 , publisher=

    2021

  9. [9]

    Scalar-on-image regression via the soft-thresholded

    Kang, Jian and Reich, Brian J and Staicu, Ana-Maria , journal=. Scalar-on-image regression via the soft-thresholded. 2018 , publisher=

    2018

  10. [10]

    Annual Review of Statistics and Its Application , volume=

    Functional regression , author=. Annual Review of Statistics and Its Application , volume=. 2015 , publisher=

    2015

  11. [11]

    Bernoulli , pages=

    Adaptive Bayesian density regression for high-dimensional data , author=. Bernoulli , pages=. 2016 , publisher=

    2016

  12. [12]

    Biometrika , volume=

    Distribution-on-distribution regression via optimal transport maps , author=. Biometrika , volume=. 2022 , publisher=

    2022

  13. [13]

    Journal of the American Statistical Association , volume=

    Wasserstein regression , author=. Journal of the American Statistical Association , volume=. 2023 , publisher=

    2023

  14. [14]

    Journal of the American Statistical Association , volume=

    Optimal sample size for multiple testing: the case of gene expression microarrays , author=. Journal of the American Statistical Association , volume=. 2004 , publisher=

    2004

  15. [15]

    Generalized

    Sturm, Karl-Theodor , journal=. Generalized. 2011 , publisher=

    2011

  16. [16]

    Journal of the American Statistical Association , volume=

    Practical Bayesian density estimation using mixtures of normals , author=. Journal of the American Statistical Association , volume=. 1997 , publisher=

    1997

  17. [17]

    Properties of discrete sliced

    Tanguy, Eloi and Flamary, R. Properties of discrete sliced. Mathematics of Computation , volume=

  18. [18]

    A Wasserstein-type distance in the space of

    Delon, Julie and Desolneux, Agnes , journal=. A Wasserstein-type distance in the space of. 2020 , publisher=

    2020

  19. [19]

    Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=

    A look at some data on the Old Faithful geyser , author=. Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=. 1990 , publisher=

    1990

  20. [20]

    Statistics and Computing , volume=

    Optimal Bayesian estimators for latent variable cluster models , author=. Statistics and Computing , volume=. 2018 , publisher=

    2018

  21. [21]

    Journal of Computational and Graphical Statistics , volume=

    Bayesian model-based clustering procedures , author=. Journal of Computational and Graphical Statistics , volume=. 2007 , publisher=

    2007

  22. [22]

    Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29--June 2, 2011, Revised Selected Papers 3 , pages=

    Wasserstein barycenter and its application to texture mixing , author=. Scale Space and Variational Methods in Computer Vision: Third International Conference, SSVM 2011, Ein-Gedi, Israel, May 29--June 2, 2011, Revised Selected Papers 3 , pages=. 2012 , organization=

    2011

  23. [23]

    Proceedings of the 26th annual international conference on machine learning , pages=

    Information theoretic measures for clusterings comparison: is a correction for chance necessary? , author=. Proceedings of the 26th annual international conference on machine learning , pages=

  24. [24]

    Journal of multivariate analysis , volume=

    Comparing clusterings—an information based distance , author=. Journal of multivariate analysis , volume=. 2007 , publisher=

    2007

  25. [25]

    Biometrika , volume=

    Bayesian cluster analysis , author=. Biometrika , volume=. 1978 , publisher=

    1978

  26. [26]

    Bayesian Analysis , volume=

    Bayesian Cluster Analysis: Point Estimation and Credible Balls (with Discussion) , author=. Bayesian Analysis , volume=

  27. [27]

    Learning with minibatch

    Fatras, Kilian and Zine, Younes and Flamary, R. Learning with minibatch. AISTATS 2020-23nd International Conference on Artificial Intelligence and Statistics , volume=

    2020

  28. [28]

    Journal of Machine Learning Research (JMLR) , Pages =

    On the efficiency of entropic regularized algorithms for optimal transport , Author =. Journal of Machine Learning Research (JMLR) , Pages =

  29. [29]

    Foundations and Trends in Machine Learning , volume=

    Convex Optimization: Algorithms and Complexity , author=. Foundations and Trends in Machine Learning , volume=. 2015 , publisher=

    2015

  30. [30]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Unbalanced minibatch Optimal Transport; applications to Domain Adaptation , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  31. [31]

    2009 , publisher=

    Optimal transport: old and new , author=. 2009 , publisher=

    2009

  32. [32]

    Mathematics of Computation , volume=

    Scaling algorithms for unbalanced optimal transport problems , author=. Mathematics of Computation , volume=

  33. [33]

    arXiv preprint arXiv:1910.12958 , year=

    Sinkhorn divergences for unbalanced optimal transport , author=. arXiv preprint arXiv:1910.12958 , year=

    Pith/arXiv arXiv 1910

  34. [34]

    Archive for rational mechanics and analysis , volume=

    The optimal partial transport problem , author=. Archive for rational mechanics and analysis , volume=. 2010 , publisher=

    2010

  35. [35]

    Advances in Neural Information Processing Systems 33 (NeurIPS 2020) , year=

    Partial Optimal Transport with Applications on Positive-Unlabeled Learning , author=. Advances in Neural Information Processing Systems 33 (NeurIPS 2020) , year=

    2020

  36. [36]

    Learning generative models with

    Genevay, Aude and Peyr. Learning generative models with. International Conference on Artificial Intelligence and Statistics , pages=. 2018 , organization=

    2018

  37. [37]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages=

    Deepjdot: Deep joint distribution optimal transport for unsupervised domain adaptation , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages=

  38. [38]

    Optimal Transport:

    Sommerfeld, Max and Schrieber, J. Optimal Transport:. Journal of Machine Learning Research , volume=

  39. [39]

    On Transportation of Mini-batches:

    Nguyen, Khai and Nguyen, Dang and Nguyen, Quoc and Pham, Tung and Bui, Hung and Phung, Dinh and Le, Trung and Ho, Nhat , journal=. On Transportation of Mini-batches:

  40. [40]

    Unbalanced optimal transport: Dynamic and

    Chizat, Lenaic and Peyr. Unbalanced optimal transport: Dynamic and. Journal of Functional Analysis , volume=. 2018 , publisher=

    2018

  41. [41]

    On parameter estimation with the

    Bernton, Espen and Jacob, Pierre E and Gerber, Mathieu and Robert, Christian P , journal=. On parameter estimation with the. 2019 , publisher=

    2019

  42. [42]

    The Annals of Statistics , Volume =

    Vector quantile regression: An optimal transport approach , Author =. The Annals of Statistics , Volume =

  43. [43]

    Journal of the American Statistical association , volume=

    Objective criteria for the evaluation of clustering methods , author=. Journal of the American Statistical association , volume=. 1971 , publisher=

    1971

  44. [44]

    Leluc, R. Sliced-. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  45. [45]

    Journal of the American statistical Association , volume=

    Gibbs sampling methods for stick-breaking priors , author=. Journal of the American statistical Association , volume=. 2001 , publisher=

    2001

  46. [46]

    Advances in neural information processing systems , volume=

    Bayesian deep learning and a probabilistic perspective of generalization , author=. Advances in neural information processing systems , volume=

  47. [47]

    Schwartz, Lorraine , journal=. On. 1965 , publisher=

    1965

  48. [48]

    Approximate

    Ishwaran, Hemant and James, Lancelot F , journal=. Approximate. 2002 , publisher=

    2002

  49. [49]

    2019 , publisher=

    Riemannian geometric statistics in medical image analysis , author=. 2019 , publisher=

    2019

  50. [50]

    Annals of Statistics , pages=

    Convergence rates of posterior distributions , author=. Annals of Statistics , pages=. 2000 , publisher=

    2000

  51. [51]

    International conference on learning representations , volume=

    Learning mixed-curvature representations in products of model spaces , author=. International conference on learning representations , volume=

  52. [52]

    Max-Sliced

    Deshpande, Ishan and Hu, Yuan-Ting and Sun, Ruoyu and Pyrros, Ayis and Siddiqui, Nasir and Koyejo, Sanmi and Zhao, Zhizhen and Forsyth, David and Schwing, Alexander G , booktitle=. Max-Sliced

  53. [53]

    Generative modeling using the sliced

    Deshpande, Ishan and Zhang, Ziyu and Schwing, Alexander G , booktitle=. Generative modeling using the sliced

  54. [54]

    Wu, Jiqing and Huang, Zhiwu and Acharya, Dinesh and Li, Wen and Thoma, Janine and Paudel, Danda Pani and Gool, Luc Van , booktitle=. Sliced

  55. [55]

    The Annals of Mathematical Statistics , volume=

    On the identifiability of finite mixtures , author=. The Annals of Mathematical Statistics , volume=. 1968 , publisher=

    1968

  56. [56]

    Subspace Robust

    Paty, Fran. Subspace Robust. International Conference on Machine Learning , pages=

  57. [57]

    Strong equivalence between metrics of

    Bayraktar, Erhan and Guo, Gaoyue , journal=. Strong equivalence between metrics of

  58. [58]

    Asymptotic guarantees for learning generative models with the sliced-

    Nadjahi, Kimia and Durmus, Alain and Simsekli, Umut and Badeau, Roland , booktitle=. Asymptotic guarantees for learning generative models with the sliced-

  59. [59]

    arXiv preprint arXiv:1506.03365 , Year =

    Yu, Fisher and Zhang, Yinda and Song, Shuran and Seff, Ari and Xiao, Jianxiong , Title =. arXiv preprint arXiv:1506.03365 , Year =

    Pith/arXiv arXiv

  60. [60]

    High-dimensional statistics: A non-asymptotic viewpoint , author=

  61. [61]

    Advances in Neural Information Processing Systems , year=

    Hierarchical Hybrid Sliced Wasserstein: A Scalable Metric for Heterogeneous Joint Distributions , author=. Advances in Neural Information Processing Systems , year=

  62. [62]

    International Conference on Machine Learning , pages=

    Arjovsky, Martin and Chintala, Soumith and Bottou, L. International Conference on Machine Learning , pages=

  63. [63]

    International Conference on Machine Learning , pages=

    On Efficient Optimal Transport: An Analysis of Greedy and Accelerated Mirror Descent Algorithms , author=. International Conference on Machine Learning , pages=

  64. [64]

    Tolstikhin, Ilya and Bousquet, Olivier and Gelly, Sylvain and Schoelkopf, Bernhard , booktitle=

  65. [65]

    Master's thesis, Department of Computer Science, University of Toronto , year=

    Learning multiple layers of features from tiny images , author=. Master's thesis, Department of Computer Science, University of Toronto , year=

  66. [66]

    Augmented Sliced

    Chen, Xiongjie and Yang, Yongxin and Li, Yunpeng , journal=. Augmented Sliced

  67. [67]

    Amortized Optimization

    Shu, Rui , booktitle=. Amortized Optimization

  68. [68]

    2020 , eprint=

    Geoopt: Riemannian Optimization in PyTorch , author=. 2020 , eprint=

    2020

  69. [69]

    International Conference on Artificial Intelligence and Statistics , pages=

    On projection robust optimal transport: Sample complexity and model misspecification , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2021 , organization=

    2021

  70. [70]

    Near-linear time approximation algorithms for optimal transport via

    Altschuler, Jason and Niles-Weed, Jonathan and Rigollet, Philippe , booktitle=. Near-linear time approximation algorithms for optimal transport via

  71. [71]

    Advances in Neural Information Processing Systems , pages=

    Joint distribution optimal transportation for domain adaptation , author=. Advances in Neural Information Processing Systems , pages=

  72. [72]

    Advances in Neural Information Processing Systems , pages=

    Sinkhorn distances: Lightspeed computation of optimal transport , author=. Advances in Neural Information Processing Systems , pages=

  73. [73]

    The American Mathematical Monthly , volume=

    Diagonal equivalence to matrices with prescribed row and column sums , author=. The American Mathematical Monthly , volume=. 1967 , publisher=

    1967

  74. [74]

    Sliced and

    Bonneel, Nicolas and Rabin, Julien and Peyr. Sliced and. Journal of Mathematical Imaging and Vision , volume=

  75. [75]

    2014 IEEE International Conference on Image Processing (ICIP) , pages=

    Adaptive color transfer with relaxed optimal transport , author=. 2014 IEEE International Conference on Image Processing (ICIP) , pages=. 2014 , organization=

    2014

  76. [76]

    Advances in Neural Information Processing Systems , pages=

    Subspace detours: Building transport plans that are optimal on subspace projections , author=. Advances in Neural Information Processing Systems , pages=

  77. [77]

    2010 IEEE International Conference on Image Processing , pages=

    Regularization of transportation maps for color and contrast transfer , author=. 2010 IEEE International Conference on Image Processing , pages=. 2010 , organization=

    2010

  78. [78]

    Advances in Neural Information Processing Systems , pages=

    Large-scale optimal transport map estimation using projection pursuit , author=. Advances in Neural Information Processing Systems , pages=

  79. [79]

    Journal of the American Statistical Association , volume=

    Sliced inverse regression for dimension reduction , author=. Journal of the American Statistical Association , volume=. 1991 , publisher=

    1991

  80. [80]

    Journal of the American Statistical Association , volume=

    On Directional Regression for Dimension Reduction , author=. Journal of the American Statistical Association , volume=. 2007 , publisher=

    2007

Showing first 80 references.