Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Normalized Radon Cumulative Distribution Transforms for Invariance and Robustness in Optimal Transport Based Image Classification

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Normalized Radon cumulative distribution transforms provably keep affinely warped image classes linearly separable, even when the warps are corrupted by bounded non-affine noise.

desk verdict A technically sound but narrow extension of the authors' mNR-CDT; the aNR-CDT guarantee covers only near-isotropic affine maps, and the numerics go beyond that regime without saying so. read the letter →

arxiv 2506.08761 v1 pith:3U6TJVYM submitted 2025-06-10 math.NA cs.CVcs.ITcs.NAmath.IT

Matthias Beckmann , Robert Beinert , Jonas Bresch This is my paper
classification math.NAcs.CVcs.ITcs.NAmath.IT MSC 94A0865D1868T10
keywords RadoncumulativedistributiontransformslicedWassersteindistancelinearseparabilityaffineinvarianceimageclassificationsmalldataregimeoptimaltransportrobustnesstoimpulsivenoise
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proves that two normalized versions of the Radon cumulative distribution transform turn image classes generated by affine transformations of a template into linearly separable sets in a fixed feature space, and that this separability survives small non-affine perturbations. The max-normalized transform works for arbitrary affine maps on compactly supported images but tolerates only small Wasserstein-infinity deformations. The newly introduced mean-normalized transform tolerates Wasserstein-2 perturbations, including impulsive noise, at the price of restricting the admissible affine maps to mildly anisotropic ones. If correct, these features give certified linear separability guarantees for small-data image classification.

What carries the argument

The unifying object is the normalized Radon-CDT N[μ](t,θ) = (bRθ[μ](t) − mean(bRθ[μ])) / std(bRθ[μ]), where bRθ[μ] is the one-dimensional cumulative distribution transform of the Radon projection in direction θ. The max-normalized version takes the pointwise maximum over θ, which collapses any affine class to a single template feature; the mean-normalized version averages over θ, turning an affine class into a small ball whose radius is controlled by the anisotropy ratio (σmax(A) − σmin(A))/σmin(A). Linear separability follows by bounding the perturbed features inside balls around the template features and applying Hahn-Banach separation of those balls.

What would settle it

Take two templates with distinct aNR-CDT features and generate finite class samples using affine maps whose anisotropy ratio exceeds the bound in Theorem 4.15, then test linear separability in aNR-CDT space; if a linear classifier separates all such samples, the theorem's restriction is not necessary, and if it fails, the restriction is essential. A second check: use a template nearly supported on a line segment so c0 is tiny, add Wasserstein-2 perturbations slightly above c0/2, and observe whether the aNR-CDT features of the two classes begin to overlap.

Watch

Extended reading notes

Core claim

The paper establishes two certified separability theorems. Theorem 4.9 shows that for compactly supported template measures with distinct max-normalized features, classes generated by arbitrary affine transformations plus Wasserstein-infinity perturbations of size below an explicit threshold remain linearly separable in mNR-CDT space. Theorem 4.15 shows that for templates in P*_2($R^{2}$) with distinct mean-normalized features, classes generated by affine transformations whose anisotropy ratio is bounded by a constant times the feature distance, plus Wasserstein-2 perturbations below an explicit threshold, remain linearly separable in aNR-CDT space. The bounds are explicit in terms of the minimal directional standard deviation of the templates, the diameters, and the feature distance, so a practitioner can check when the guarantee applies.

Load-bearing premise

The mean-normalized separability guarantee holds only when the affine distortion is mildly anisotropic: the ratio of largest to smallest singular values must stay below a threshold set by the feature distance between the templates, so strongly anisotropic scaling or shearing falls outside the certified regime.

Editorial extensions

If this is right

  • mNR-CDT provides a certified linear classifier for compactly supported image classes under arbitrary affine transformations, stable under small local deformations measured in Wasserstein-infinity distance.
  • aNR-CDT extends the certified regime to non-compactly supported measures and to Wasserstein-2 perturbations, which include impulsive noise that affects the whole support.
  • The explicit perturbation thresholds, ε < min{cμ, cν}/2, tell practitioners which template pairs and noise levels are guaranteed to remain separable.
  • The aNR-CDT robustness bound scales as 4ε/c0, so templates with well-spread directional support (large minimal standard deviation) enjoy stronger noise tolerance.
  • Both features require only quantile functions, means, standard deviations, and max/mean operations over angles, making them feasible for small-data regimes where deep classifiers lack training samples.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural practical extension is to switch feature representations based on the estimated anisotropy of the affine distortion: aNR-CDT for mildly anisotropic warps with noise, mNR-CDT for strongly anisotropic or shear-dominated warps.
  • The theory predicts that aNR-CDT separability degrades as the template approaches a one-dimensional support (c0 → 0); a benchmark on nearly-line-shaped templates would directly test this prediction.
  • The paper's own experiments suggest the mNR-CDT bounds are conservative, so refined sensitivity estimates could widen the certified perturbation radii without changing the feature definitions.
  • Because the aNR-CDT tolerates Wasserstein-2 perturbations, it could be used as a preprocessing step for optimal-transport-based registration or as a feature for clustering, where affine invariance and noise robustness are both desirable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies two normalizations of the Radon cumulative distribution transform (R-CDT) for image classification in the small-data regime. It first recalls the max-normalized R-CDT (mNR-CDT) and proves a stability estimate under Wasserstein-infinity perturbations of the template measures, extending the earlier affine-invariance separability result from the authors' conference paper. It then introduces a mean-normalized R-CDT (aNR-CDT), defined by averaging the normalized R-CDT over directions, and proves linear-separability guarantees for classes generated by affine transformations and Wasserstein-2 perturbations, under an explicit admissibility condition on the affine maps. Numerical experiments on synthetic symbols, Chinese characters, and LinMNIST report high classification accuracies for the new features compared with R-CDT and Euclidean baselines.

Significance. If the main theorems hold, the paper makes a useful contribution: it provides quantitative, parameter-free robustness bounds for two R-CDT variants, and it introduces a new feature whose stability is measured in Wasserstein-2 rather than Wasserstein-infinity, which is more realistic for deformations and impulsive noise. The proofs are detailed, with explicit constants, and the code is publicly available. The main caveat is that the aNR-CDT separability guarantee is restricted to a narrow admissibility set of near-isotropic affine maps, and the numerical experiments do not verify that they operate inside that certified regime. This limits the strength of the paper's claim that the experiments support the theoretical results.

major comments (3)
  1. [Theorem 4.15; Section 5.1.1] The admissibility set X in Theorem 4.15 restricts affine maps to those with (σmax(A)−σmin(A))/σmin(A) ≤ c ∥N_a[μ0]−N_a[ν0]∥_ρ / max{∥N[μ0]∥_{ρ×u_{S^1}}, ∥N[ν0]∥_{ρ×u_{S^1}}} with c<1/2. Since ∥N_a[μ0]−N_a[ν0]∥_ρ ≤ 2 max{∥N[μ0]∥, ∥N[ν0]∥} by Jensen's inequality, the admissible anisotropy ratio is always less than 2, and it can be arbitrarily small when the templates are close in aNR-CDT distance. The epsilon threshold in the same theorem is proportional to this distance, so it shrinks correspondingly. The experiments in Table 2 use scaling in [0.5,1.25] and shearing in [−45°,45°], which plausibly produce anisotropy ratios well above the certified bound; the paper does not report whether any of the Table 3 or Figure 7 configurations satisfy the theorem's epsilon condition. Thus the numerical evidence does not actually exercise the regime in which Theorem 4.15 provides a guarantee, and the claim that the experiments support the aNR-CDT robustness theorem is overstated. The authors should explicitly state this limitation and, ideally, add experiments inside the certified regime or a quantitative check of the theorem's assumptions.
  2. [Section 4.1, definition of N_m; Theorem 4.4] The mNR-CDT is defined through a pointwise maximum over the continuum S^1, and the text asserts N_m[μ]∈L∞_ρ by reference to [3, Proposition 6] without proof. Theorem 4.4 is likewise imported from [3, Theorem 1]. Since Theorem 4.9 and Corollary 4.8 rely on N_m as an L∞_ρ function, the manuscript should either prove measurability of the maximum (or give a precise statement with a proof) and state the exact assumptions under which Theorem 4.4 holds. Without measurability, the norm ∥N_m∥_{L∞_ρ} is not defined, and the ball-separation argument in Theorem 4.9 has no rigorous object. The paper should be self-contained on this point, especially because the mNR-CDT is one of the two main features studied.
  3. [Theorem 3.4] The theorem imposes no growth condition on the increasing bijections h∈H. For arbitrary h, h#R_{θ0}[μ0] need not lie in P2(R), and h∘b_{Rθ0}[μ0] need not belong to L^2_ρ(R). The proof asserts bF_{θ0}⊂L^2_ρ(R) from μ0,ν0∈P2, which is not justified. For the convexity argument, one needs ∫|h(b_{p0}(t))|^2 dρ(t)<∞ for all h∈H and for convex combinations. As stated, the theorem's conclusion—linear separability in L^2_ρ—is not well-defined for such h. This can be repaired by restricting H to functions with controlled growth (e.g., affine maps as in Example 3.6) or by requiring all class members to lie in P2(R2).
minor comments (6)
  1. [Abstract and Introduction] The word 'guaranties' should be 'guarantees' in several places.
  2. [Proposition 4.7] The displayed bound would be clearer with parentheses: (diam(μ0)+2ϵ)/(c0(c0−2ϵ)) · 4ϵ.
  3. [Section 5.1.1] The sentence 'In this and all experiment regarding the academic datasets' should read 'In this and all experiments regarding the academic datasets'.
  4. [Section 4.2] The name 'mean-normalized R-CDT' is initially confusing because the normalization step subtracts the mean; clarify that the 'mean' in the name refers to averaging over directions.
  5. [Table 2] The column headers repeat '∥·∥∞ ∥·∥2' without making clear which parameter range each block corresponds to; the layout should be made self-explanatory.
  6. [Figure 9] The caption says true labels are vertical and classified labels horizontal, which is opposite to the usual convention; please verify the orientation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the new aNR-CDT and perturbation results are proved from definitions, and the self-citations to [3] are transparent building blocks, not load-bearing redefinitions.

full rationale

The paper's central new claims are Theorems 4.9 and 4.15, and both are derived in-line. Proposition 4.5 proves a W-infinity stability bound for the R-CDT from the definition of the quantile function and the contraction property of the restricted Radon transform (Proposition 3.3). Lemma 4.6, Proposition 4.7, and Corollary 4.8 build on this to bound the mNR-CDT perturbation without fitting any parameter to the data; Theorem 4.9 then separates the two balls when the perturbation radius is smaller than half the template distance. The aNR-CDT part is equally self-contained: Lemma 4.10 establishes square-integrability, Theorem 4.11 derives a bound on the aNR-CDT deviation in terms of the anisotropy ratio via the Jacobian of the induced map on the circle, and Theorem 4.15 combines this with the W2-stability bound in Proposition 4.14. The admissibility sets X in Theorems 4.11 and 4.15 are explicit quantitative sufficiency conditions, not hidden assumptions that redefine the conclusion; the theorems do not claim separability outside X. The paper cites the authors' earlier work [3] for the mNR-CDT invariance theorem and for the measure-valued Radon transform duality, but these are external, peer-reviewed results with stated assumptions, and the new perturbation estimates do not reduce to them: Theorem 4.9's proof uses only the ball-radius bound, not the unperturbed collapse, and Theorem 4.15 is proved from first principles. The numerical experiments using scaling in [0.5,1.25] and shearing in [-45,45] degrees may operate outside the certified aNR-CDT regime, but that is a limitation of the theorem's applicability, not a circular derivation. The acknowledged looseness of the mNR-CDT bounds is likewise a quality-of-estimate issue, not circularity. No equation or fitted parameter is identified that makes the output equal to the input by construction.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The theoretical machinery rests on standard optimal transport results, a non-degeneracy condition on templates, and an ad hoc anisotropy restriction for the mean-normalized variant. No physical or ontological entities are introduced. The only free parameters are theorem thresholds c and c', which are chosen by hand and not fitted to data.

free parameters (2)
  • c
    Hand-chosen threshold in (0,1/2) defining the admissible anisotropy class X for aNR-CDT in Theorems 4.11 and 4.15. It is not fitted to data but is a design parameter of the theorem.
  • c_prime
    Auxiliary threshold in (c,1/2) introduced in Theorem 4.15 to set the perturbation radius epsilon. It is a proof device rather than a method parameter.
assumptions (6)
  • standard math Wasserstein distances attain their infima and the one-dimensional CDT equals the optimal transport map from the reference measure rho.
    Invoked in Section 2 and Section 3 to define W_p, W_infinity, and the CDT identifications.
  • standard math A fixed atomless reference measure rho with unit mass exists, e.g., uniform on [0,1].
    Used throughout Section 3 and Section 4 to define the CDT and the normalized transforms.
  • domain assumption Templates have dim(supp(mu)) > 1, so directional standard deviations are uniformly positive.
    Assumed through P*_2(R^2) and P*_c(R^2); Lemma 4.1 proves the needed positivity for such measures.
  • domain assumption For mNR-CDT, templates have compact support and perturbations are W_infinity-close.
    P*_c(R^2) is used in Theorem 4.4 and Propositions 4.7 and 4.9 to ensure N_m values lie in L_infinity_rho.
  • ad hoc to paper Admissible affine transformations for aNR-CDT are restricted to the set X with bounded anisotropy ratio.
    The class X in Theorems 4.11 and 4.15 is imposed by the proof to control the angular reparametrization error; it is not derived from the measurement setup.
  • domain assumption The templates are distinguishable in feature space: N_m[mu0] != N_m[nu0] or N_a[mu0] != N_a[nu0].
    Non-degeneracy condition stated in Theorems 4.4, 4.9, 4.11, and 4.15; without it linear separation is impossible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Normalized Radon Cumulative Distribution Transforms for Invariance and Robustness in Optimal Transport Based Image Classification." pith.science (2026). https://pith.science/paper/3U6TJVYM

@misc{pith2026250608761,
  author       = {Pith},
  title        = {Pith review of: Normalized Radon Cumulative Distribution Transforms for Invariance and Robustness in Optimal Transport Based Image Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3U6TJVYM}},
  note         = {Machine review of arXiv:2506.08761}
}
read the original abstract

The Radon cumulative distribution transform (R-CDT), is an easy-to-compute feature extractor that facilitates image classification tasks especially in the small data regime. It is closely related to the sliced Wasserstein distance and provably guaranties the linear separability of image classes that emerge from translations or scalings. In many real-world applications, like the recognition of watermarks in filigranology, however, the data is subject to general affine transformations originating from the measurement process. To overcome this issue, we recently introduced the so-called max-normalized R-CDT that only requires elementary operations and guaranties the separability under arbitrary affine transformations. The aim of this paper is to continue our study of the max-normalized R-CDT especially with respect to its robustness against non-affine image deformations. Our sensitivity analysis shows that its separability properties are stable provided the Wasserstein-infinity distance between the samples can be controlled. Since the Wasserstein-infinity distance only allows small local image deformations, we moreover introduce a mean-normalized version of the R-CDT. In this case, robustness relates to the Wasserstein-2 distance and also covers image deformations caused by impulsive noise for instance. Our theoretical results are supported by numerical experiments showing the effectiveness of our novel feature extractors as well as their robustness against local non-affine deformations and impulsive noise.

Figures

Figures reproduced from arXiv: 2506.08761 by the authors.

Figure 1
Figure 1. Templates for academic datasets with domain [− √1 2 , √1 2 ] 2 represented by 256×256 pixels. and the choice of ϵ > 0 guarantees that ∥Na[µ] − Na[µ0]∥ρ < c′ ∥Na[µ0] − Na[ν0]∥ρ=: r0. Consequently, NaF ⊆ Br0 (Na[µ0]) ⊂ L 2 ρ (R) and, analogously, NaG ⊆ Br0 (Na[ν0]) ⊂ L 2 ρ (R). Since c ′ ∈ (c, 1 2 ), Br0 (Na[µ0]) and Br0 (Na[ν0]) are linearly separable in L2 ρ (R). This implies the linear separability of NaF0 and NaG0… view at source ↗
Figure 2
Figure 2. mNR-CDT (left) and aNR￾CDT (right) of affine classes 5 and 12. The classes are generated using horizon￾tal/vertical scaling by a factor in [0.5, 1.25], rotation by an angle in [0◦ , 360◦ ], shearing of horizontal/vertical axis by an angle in [−45◦ , 45◦ ], and horizontal/vertical shifting by a pixel number in [−20, 20] [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Instances of corrupted data regarding non-affine deformations and impulsive/salt noise considered in the robustness analysis. The accompanying vector fields illustrate the optimal Wasserstein-2 transport between the corrupted datum ρdis and the true template ρorg [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Scala of exemplarily rendered salt noise of strength ℓ ∈ {1, . . . , 20} used in the experiments. The salt noise illustrated in [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]
Figure 7
Figure 7. Figure 7: Phase transition of NT classification accuracies for 3-class academic datasets with 10 samples per class and different salt noise (vertical: component numbers, horizontal: noise strengths, cf [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Phase transition of NT classification accuracies for 3-class academic datasets with 10 samples per class, non-affine deformations with frequencies in [0, 2] and variable amplitude ranges (horizontal), as well as salt noise of strength 9 according to [PITH_FULL_IMAGE:f…
Figure 9
Figure 9. Figure 9: Confusion maps for aNR-CDT in the 3. and 4. experiment in [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 11
Figure 11. Figure 11: Two samples of each class (zero to nine) of the LinMNIST dataset, generated by randomly rotating, anisotropically scaling and shifting a random MNIST sample of the corresponding class. Here, rotation angles are in [0◦ , 360◦ ], scalings in [0.75, 1.0] and shifts in [−…
Figure 12
Figure 12. Figure 12: Confusion maps for the best mNR-CDT and aNR-CDT results per number of training samples. curacy of 65% when using k = 11 and 50 training samples. In contrast to this, the Euclidean and R-CDT representation perform on the level of random guessing. Inspecting the confu￾s…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Unsupervised Ground Metric Learning

    math.OC 2025-07 conditional novelty 7.0 of 10

    New convergence proofs for stochastic fixed-point iterations in unsupervised ground metric learning, with extensions to Mahalanobis and graph Laplacian distances.

Reference graph

Works this paper leans on

30 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Aldroubi, R

    A. Aldroubi, R. Diaz Martin, I. V. Medri, G. K. Rohde, and S. Thareja , The signed cumulative distribution transform for 1-d signal analysis and classification , Foundations of Data Science, 4 (2022), pp. 137--163, https://doi.org/10.3934/fods.2022001

    work page doi:10.3934/fods.2022001 2022

  2. [2]

    Aldroubi, S

    A. Aldroubi, S. Li, and G. K. Rohde , Partitioning signal classes using transport transforms for data analysis and machine learning , Sampling Theory, Signal Processing, and Data Analysis, 19 (2021), p. 6, https://doi.org/10.1007/s43670-021-00009-z

    work page doi:10.1007/s43670-021-00009-z 2021

  3. [3]

    Beckmann, R

    M. Beckmann, R. Beinert, and J. Bresch , Max-normalized radon cumulative distribution transform for limited data classification , in Scale Space and Variational Methods in Computer Vision (SSVM), vol. 15667 of Lecture Notes in Computer Science, Springer, 2025, pp. 241--254, https://doi.org/10.1007/978-3-031-92366-1_19

    work page doi:10.1007/978-3-031-92366-1_19 2025

  4. [4]

    Beckmann and N

    M. Beckmann and N. Heilenk \"o tter , Equivariant neural networks for indirect measurements , SIAM Journal on Mathematics of Data Science, 6 (2024), pp. 579--601, https://doi.org/10.1137/23M1582862

    work page doi:10.1137/23m1582862 2024

  5. [5]

    Beier, R

    F. Beier, R. Beinert, and G. Steidl , On a linear G romov– W asserstein distance , IEEE Transactions on Image Processing, 31 (2022), pp. 7292--7305, https://doi.org/10.1109/TIP.2022.3221286

    work page arXiv 2022

  6. [6]

    Beinert, C

    R. Beinert, C. Heiss, and G. Steidl , On assignment problems related to G romov– W asserstein distances on the real line , SIAM Journal on Imaging Sciences, 16 (2023), pp. 1028--1032, https://doi.org/10.1137/22M1497808

    work page doi:10.1137/22m1497808 2023

  7. [7]

    Bliem , Handwritten chinese character hanzi datasets , 2022, https://www.kaggle.com/datasets/pascalbliem/handwritten-chinese-character-hanzi-datasets

    P. Bliem , Handwritten chinese character hanzi datasets , 2022, https://www.kaggle.com/datasets/pascalbliem/handwritten-chinese-character-hanzi-datasets. Accessed: March 26, 2025

    work page 2022

  8. [8]

    Bonneel, J

    N. Bonneel, J. Rabin, G. Peyr \'e , and H. Pfister , Sliced and R adon W asserstein barycenters of measures , Journal of Mathematical Imaging and Vision, 51 (2015), pp. 22--45, https://doi.org/10.1007/s10851-014-0506-3

    doi:10.1007/s10851-014-0506-3 2015

Show all 30 references
  1. [9]

    Cloninger, K

    A. Cloninger, K. Hamm, V. Khurana, and C. Moosm \" u ller , Linearized W asserstein dimensionality reduction with approximation guarantees , Applied and Computational Harmonic Analysis, 74 (2025), p. 101718, https://doi.org/10.1016/j.acha.2024.101718

    2025

  2. [10]

    Deng , The MNIST database of handwritten digit images for machine learning research , IEEE Signal Processing Magazine, 29 (2012), pp

    L. Deng , The MNIST database of handwritten digit images for machine learning research , IEEE Signal Processing Magazine, 29 (2012), pp. 141--142, https://doi.org/10.1109/MSP.2012.2211477

    2012

  3. [11]

    D \' i az Mart \' i n, I

    R. D \' i az Mart \' i n, I. V. Medri, and G. K. Rohde , Data representation with optimal transport , 2024, https://doi.org/10.48550/arXiv.2406.15503. arXiv:2406.15503

    2024 doi arXiv

  4. [12]

    C. R. Givens and R. M. Shortt , A class of W asserstein metrics for probability distributions , Michigan Mathematical Journal, 31 (1984), pp. 231--240, https://doi.org/10.1307/mmj/1029003026

    1984

  5. [13]

    Hauser, M

    D. Hauser, M. Beckmann, G. Koliander, and H. S. Stiehl , On image processing and pattern recognition for thermograms of watermarks in manuscripts -- a first proof-of-concept , in International Conference on Document Analysis and Recognition (ICDAR), 2024, pp. 91--107, https://...

    2024 doi

  6. [14]

    Helgason , The Radon Transform , Birkh \"a user, 2 ed., 1999

    S. Helgason , The Radon Transform , Birkh \"a user, 2 ed., 1999

    1999

  7. [15]

    Kolouri, S

    S. Kolouri, S. R. Park, and G. K. Rohde , The R adon cumulative distribution transform and its application to image classification , IEEE Transactions on Image Processing, 25 (2016), pp. 920--934, https://doi.org/10.1109/TIP.2015.2509419

    2016

  8. [16]

    Kolouri, S

    S. Kolouri, S. R. Park, M. Thorpe, D. Slepcev, and G. K. Rohde , Optimal mass transport , IEEE Signal Processing Magazine, 34 (2017), pp. 43--59, https://doi.org/10.1109/MSP.2017.2695801

    2017

  9. [17]

    X. Liu, R. D \' i az Mart \' i n, Y. Bai, A. Shahbazi, M. Thorpe, A. Aldroubi, and S. Kolouri , Expected sliced transport plans , 2024, https://doi.org/10.48550/arXiv.2410.12176. arXiv:2410.12176

    2024 doi arXiv

  10. [18]

    Moosm \"u ller and A

    C. Moosm \"u ller and A. Cloninger , Linear optimal transport embedding: provable W asserstein classification for certain rigid transformations and perturbations , Information and Inference: A Journal of the IMA, 12 (2023), pp. 363--389, https://doi.org/10.1093/imaiai/iaac023

    2023 doi

  11. [19]

    Natterer , The Mathematics of Computerized Tomography , SIAM, Philadelphia, 2001, https://doi.org/10.1137/1.9780898719284

    F. Natterer , The Mathematics of Computerized Tomography , SIAM, Philadelphia, 2001, https://doi.org/10.1137/1.9780898719284

    2001 doi

  12. [20]

    Park and D

    S. Park and D. Slep c ev , Geometry and analytic properties of the sliced W asserstein space , Journal of Functional Analysis, 289 (2025), p. 110975, https://doi.org/10.1016/j.jfa.2025.110975

    2025

  13. [21]

    S. R. Park, S. Kolouri, S. Kundu, and G. K. Rohde , The cumulative distribution transform and linear pattern classification , Applied and Computational Harmonic Analysis, 45 (2018), pp. 616--641, https://doi.org/10.1016/j.acha.2017.02.002

    2018 doi

  14. [22]

    Piening and R

    M. Piening and R. Beinert , Slicing the G aussian mixture W asserstein distance , 2025, https://doi.org/10.48550/arXiv.2504.08544. arXiv:2504.08544

    2025 doi arXiv

  15. [23]

    Quellmalz, R

    M. Quellmalz, R. Beinert, and G. Steidl , Sliced optimal transport on the sphere , Inverse Problems, 39 (2023), p. 105005, https://doi.org/10.1088/1361-6420/acf156

    2023 doi

  16. [24]

    Quellmalz, L

    M. Quellmalz, L. Buecher, and G. Steidl , Parallelly sliced optimal transport on spheres and on the rotation group , Journal of Mathematical Imaging and Vision, 66 (2024), pp. 951--976, https://doi.org/10.1007/s10851-024-01206-w

    2024 doi

  17. [25]

    U ber die B estimmung von F unktionen durch ihre I ntegralwerte l \

    J. Radon , \"U ber die B estimmung von F unktionen durch ihre I ntegralwerte l \"a ngs gewisser M annigfaltigkeiten , in Berichte \"u ber die Verhandlungen der K\"oniglich-S \"a chsischen Gesellschaft der Wissenschaften zu Leipzig, vol. 69 of Mathematisch-Physische Klasse, Kön...

    1917

  18. [26]

    A. G. Ramm and A. I. Katsevich , The Radon Transform and Local Tomography , CRC Press, 1996, https://doi.org/10.1201/9781003069331

    1996 doi

  19. [27]

    Shifat-E-Rabbi, X

    M. Shifat-E-Rabbi, X. Yin, A. H. M. Rubaiyat, S. Li, S. Kolouri, A. Aldroubi, J. M. Nichols, and G. K. Rohde , Radon cumulative distribution transform subspace modeling for image classification , Journal of Mathematical Imaging and Vision, 63 (2021), pp. 1185--1203, https://do...

    2021 doi

  20. [28]

    Shifat-E-Rabbi, Y

    M. Shifat-E-Rabbi, Y. Zhuang, S. Li, A. H. M. Rubaiyat, X. Yin, and G. K. Rohde , Invariance encoding in sliced- W asserstein space for image classification with limited training data , Pattern Recognition, 137 (2023), p. 109268, https://doi.org/10.1016/j.patcog.2022.109268

    2023

  21. [29]

    Villani , Topics in Optimal Transportation , American Mathematical Society, 2003, https://doi.org/10.1090/gsm/058

    C. Villani , Topics in Optimal Transportation , American Mathematical Society, 2003, https://doi.org/10.1090/gsm/058

    2003 doi

  22. [30]

    Zhuang, S

    Y. Zhuang, S. Li, M. Shifat-E-Rabbi, X. Yin, A. H. M. Rubaiyat, and G. K. Rohde , Local sliced W asserstein feature sets for illumination invariant face recognition , Pattern Recognition, 162 (2025), p. 111381, https://doi.org/10.1016/j.patcog.2025.111381

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.