Pith. sign in

REVIEW 2 major objections 5 minor 29 references

Convergence Rates for Distribution Matching with Sliced Optimal Transport

T0 review · 2 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The slice-matching scheme provably converges to an isotropic Gaussian target with expected loss O(k^{1−2α}).

desk verdict The main theorem is not proven as written: Prop. D.2 treats a covariance-dependent extremal direction as if it were Haar-uniform, and Lemma D.3 is plainly false; the framework is still worth engaging. read the letter →

arxiv 2602.10691 v2 pith:MNQQ5K7W submitted 2026-02-11 stat.ML cs.LG

Gauthier Thurin (ENS-PSL) , Claire Boyer (LMO , IUF , CELESTE) , Kimia Nadjahi (ENS-PSL) This is my paper · ORCID
classification stat.MLcs.LG MSC 49Q2265K10
keywords Sliced-Wassersteindistancedistributionmatchingslice-matchingschemePolyak–ŁojasiewiczinequalitystochasticgradientdescentGaussianmeasuresnon-asymptoticconvergencerates
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 derives quantitative non-asymptotic convergence rates for the slice-matching scheme, an iterative method that matches distributions by repeatedly applying one-dimensional optimal transport maps along a random orthonormal basis. The central result is that when both the source and the target are Gaussian and the target is isotropic, the expected Sliced-Wasserstein objective decays as C/k^{2α−1} for step sizes γ_k = (k+1)^{−α} with 1/2 < α < 1. The proof works by showing that the Sliced-Wasserstein objective satisfies a Polyak–Łojasiewicz inequality along the trajectory, with random constants controlled through the smallest eigenvalue of the current covariance. A key ingredient is the choice of an orthonormal basis at each step: it matches first and second moments and keeps inverse eigenvalues bounded in expectation. The rate matters because it turns an empirically successful but theoretically poorly understood heuristic into a provably convergent stochastic optimization method.

What carries the argument

The load-bearing objects are the slice-matching map T_{σ,P}(x) = Σ_i t_{θ_i}(θ_iᵀx)θ_i acting along the columns of a Haar-random orthonormal basis P, and the covariance recursion Σ_{k+1} = ((1−γ_k)I + γ_k P D_k Pᵀ) Σ_k ((1−γ_k)I + γ_k P D_k Pᵀ)ᵀ with D_k = diag(√(θ_iᵀΛθ_i / θ_iᵀΣ_kθ_i)). The proof combines three ingredients: the gradient-variance decomposition 2F = ‖∇F‖²_σ + E‖T̄ − T_P‖²_σ; a Polyak–Łojasiewicz inequality F(σ) ≤ d²(d+2) / λ_min(Σ)² · ‖∇F‖²_σ along the iterates; and a recursion bounding λ_min(Σ_{k+1}) from below by λ_min(Σ_k) times a factor depending on one direction's ratio τ_θ. Orthonormal-basis sampling is what makes the drift term E[θᵀΣ_kθ − 1] ≤ 0, giving uniform-in-k mo

What would settle it

Run the explicit covariance recursion (18) for a non-isotropic Gaussian target with a fixed α and track E[λ_min(Σ_k)^{−p}] over many random seeds; if this quantity diverges with k for any p ≥ 1, then condition (21) fails and Theorem 4.1 cannot be extended to that target. Equivalently, for the isotropic case, a mismatch between the empirical slope of log E[F(σ_k)] and 1 − 2α over k up to about 10^4 would signal that the claimed rate, or its dimension constant, does not hold as stated.

Watch

Extended reading notes

Core claim

Writing the slice-matching update as a stochastic gradient step in Wasserstein space, the authors prove Theorem 4.1: if σ = N(0, Σ) and μ = N(0, Id), then for γ_k = (k+1)^{−α}, 1/2 < α < 1, there is C > 0 with E[F(σ_k)] ≤ C / k^{2α−1}, where F(σ) = (d/2) SW²(σ, μ). The argument decomposes the objective into squared gradient plus variance, obtains a Polyak–Łojasiewicz inequality whose constant is proportional to 1/λ_min(Σ_k)², and then controls inverse moments through an eigenvalue recursion derived from the explicit covariance update. The crucial step is that isotropy of the target makes the conditional drift E[θᵀΣ_kθ − 1 | A_k] = Tr(Σ_k)/d − 1 ≤ 0, so λ_min(Σ_k)^{−p} has bounded expectation

Load-bearing premise

The quantitative rate rests on the target being the isotropic Gaussian Λ = Id; for Λ ≠ Id the paper only provides an unverified sufficient condition (Eq. 21), so the rate is not established beyond the isotropic case.

Editorial extensions

If this is right

  • For an isotropic Gaussian target and a Gaussian source, with step sizes 1/(k+1)^α and 1/2 < α < 1, the expected Sliced-Wasserstein loss satisfies E[F(σ_k)] ≤ C/k^{2α−1}; convergence to the target is guaranteed at a polynomial rate.
  • The constants grow polynomially (roughly d²(d+2)) in the ambient dimension, so the bound predicts slower convergence in higher dimensions, which the reported experiments reproduce.
  • Sampling an orthonormal basis of directions at each step, rather than a single direction, is what allows inverse-eigenvalue moments to be controlled from the first iteration; the experiments show that this choice also suppresses fluctuations in the extreme eigenvalues.
  • Even without the Polyak–Łojasiewicz inequality, the stochastic-optimization viewpoint gives a convergence-to-critical-point guarantee at rate K^{α−1} for weighted average gradient norms, complementing earlier asymptotic results.
  • For co-diagonalizable Gaussian covariances with eigenvalues in [m, M], the new comparison W2² ≤ d(d+2)(M/m) SW2² replaces the usual exponential dimension dependence with a polynomial one.

Reading between the lines

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

  • If the claimed rate extends to general Gaussian targets, the sufficient condition (21) must hold; a natural test is to run the explicit covariance recursion with a non-isotropic Λ and check whether E[λ_min(Σ_k)^{−p}] stays bounded for all p. The paper's own numerical experiments hint that it does, but no proof is given.
  • The theorem is silent on the empirically best regime α ≤ 1/2, including α = 0 (the original constant-step algorithm). A plausible reading, which the authors flag as open, is that the polynomial bound captures only the tail of the convergence and that small-α behavior is governed by a different, perhaps dimension-dependent constant.
  • The event-based handling of random PL constants may transfer to other distribution-matching schemes (for instance, ones with diffusive regularization) as long as one can control the analogous random constants in expectation; if a similar spectral drift inequality holds, the same rate template would apply.
  • The single-direction variant's lack of inverse-moment control suggests that orthogonality is not merely an implementation detail: without it, worst-case drift can accumulate in low-eigenvalue directions, so one would expect strictly worse worst-case rates. This is an inference from the paper's analysis, not a proved statement.
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

2 major / 5 minor

Summary. The paper studies the slice-matching scheme (IDT with random orthonormal bases) for distribution matching in Wasserstein space. It derives a decomposition of the Sliced-Wasserstein gradient into variance and norm terms, establishes static Łojasiewicz/PL-type inequalities for bounded-density measures and for Gaussian measures with co-diagonalizable covariances, and proves a polynomial-dimension comparison W_2^2 ≤ C_{d,m,M} SW_2^2 in that Gaussian case. The main advertised result is Theorem 4.1: for a centered Gaussian source and an isotropic Gaussian target, with step sizes γ_k=(k+1)^{-α}, 1/2<α<1, the expected SW loss decays as O(k^{-(2α-1)}). The proof strategy is to verify a random PL condition (Assumption A) by controlling inverse moments of λ_min(Σ_k) along the covariance recursion, then applying a Chung-type lemma.

Significance. If the proof were valid, the result would be a useful quantitative non-asymptotic convergence rate for a practical sliced-OT algorithm, with explicit dependence on dimension and step-size, and the W2–SW2 comparison for Gaussians is of independent interest. The paper is also transparent about the isotropic-target restriction and provides reproducible numerical experiments. However, the moment control of λ_min is the linchpin of the proof, and the present argument contains a selection-bias error and a false lemma. Since the advertised rate is not established as written, the contribution is conditional on a correct spectral-control proof.

major comments (2)
  1. [Proposition D.2 (Eq. (51))] The conditional-expectation step is invalid because the direction θ_i is selected, not independent. In (45), θ_i is the column of P_{k+1} minimizing θ^TΛθ/θ^TΣ_kθ; when Λ=I it maximizes θ^TΣ_kθ among the d columns. After (51), the proof replaces E[θ_i^TΣ_kθ_i|A_k] by E_θ[θ^TΣ_kθ]=Tr(Σ_k)/d, which holds only for a fixed Haar-uniform θ independent of Σ_k. For d=2, Σ_k=diag(1.8,0.1), E[max_j θ_j^TΣ_kθ_j]≈1.49>Tr(Σ_k)/d=0.95, and can exceed 1. Thus E[(λ_min(Σ_{k+1}))^{-1}|A_k]≤(λ_min(Σ_k))^{-1} does not follow; neither do (49)–(50). Since these moment bounds are the sole verification of Assumption A for the isotropic Gaussian case, Theorem 4.1 is not proved.
  2. [Lemma D.3] The second inequality in Lemma D.3 is false as stated. For A=0.5I and d=2, the left-hand side at p=2 is E[(θ^TAθ−1)^2]=0.25, while the right-hand side is 1+∑_{r=1}^2 C(2,r)λ_max(A)^{r−1}(−1)^r = 1−2+0.5 = −0.5, a negative upper bound. The lemma is used in the proof of Proposition D.2 to bound E[(θ_i^TΣ_kθ_i−1)^p]; the statement needs correction even though a crude finite bound for fixed A is trivial.
minor comments (5)
  1. [Appendix E.1, Eq. (54)] The constant d^2(d+2)/λ_min(Σ_k)^2 does not appear to follow directly from Proposition 4.4 with the stated bounds; the derived constant seems to involve an extra factor of order d. This does not affect the rate, but the calculation should be checked.
  2. [Proposition D.1] Please state explicitly that θ_i and θ_j are the extremal columns arising in the Ostrowski bound. The current wording 'there exist directions' obscures the selection dependence that is central to the flaw in Proposition D.2.
  3. [Proposition 4.3] The inequality F(σ) ≤ 2C_ν(M/m)∥∇W2F(σ)∥_σ is a Łojasiewicz inequality with exponent 1/2, not a PL inequality of the form used in Assumption A. Calling it 'PL-like' may confuse the two notions.
  4. [Throughout] Minor typos and reference issues: 'weigths' in Proposition 3.3; Li and Moosmüller (2023) should give the arXiv identifier; the Sisouk et al. entry lacks full publication data.
  5. [Abstract and Conclusion] The abstract says convergence for 'Gaussian distributions', but Theorem 4.1 is specifically for an isotropic target. The conclusion is transparent about this, but the abstract should be aligned to avoid overclaiming.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the rate follows from re-proved PL/moment inequalities with explicit constants; the main caveats are an acknowledged isotropic-only target and a possible independence gap in the spectral proof, which is a correctness issue rather than circularity.

full rationale

The central claim (Theorem 4.1) is not fit-to-data and does not reduce to its inputs by construction. The proof chain is: (i) Theorem 4.2 converts a random-PL condition (Assumption A) into E[F(σ_k)] ≲ k^{-(2α-1)}; (ii) Proposition 4.4 establishes a static PL inequality for Gaussians, adapted from the non-self citation Chewi et al. (2020); (iii) Proposition B.8 re-proves the moment bound M2(σ_k) ≤ M2(µ) from the moment-matching identity; (iv) for the isotropic target Λ = I_d, Proposition D.2 supplies the inverse-eigenvalue moment bounds E[λ_min(Σ_k)^{-p}] < ∞. The constants in the final rate are explicit in d, m, M, and the step-size schedule; no parameter is fitted to loss data and then renamed a prediction. The only author self-citations (Nadjahi et al. 2019, 2020) appear in the literature review and are not load-bearing for the convergence theorem. The paper itself acknowledges the main limitation: Section 4.4(c) states 'We are able to verify (21) in the isotropic target case Λ = Id', and the Conclusion states 'our explicit rate requires an isotropic Gaussian target'; the non-isotropic case is therefore an admitted open scope condition, not a disguised circular step. There is, however, a potential proof gap in Proposition D.2: the conditional expectation E[θ_i^T Σ_k θ_i | A_k] = Tr(Σ_k)/d - 1 treats the direction θ_i from Proposition D.1 as Haar-uniform and independent of Σ_k, while θ_i is tied to the extremal eigen-direction of Σ_k (for Λ = I_d, effectively the direction maximizing θ^TΣ_kθ). If that equality fails, the moment bounds (49)-(50) do not follow as written. This is a substantive mathematical correctness concern, but it is not an equivalence-by-construction of the theorem and its inputs, so under the circularity rubric it is weighed as correctness risk and does not raise the circularity score.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

No fitted numerical constants appear in the derivation; the only hand-chosen quantity is the step-size schedule α, which is an algorithmic hyperparameter. The proof is conditional on Gaussian/elliptic structure, isotropic target, Haar-random bases, and standard analytic inequalities.

free parameters (1)
  • step-size exponent α and implicit first step = α in (1/2, 1); γ0 = 1 (implicit in proof)
    User-chosen hyperparameter, not fitted to data; the theorem quantifies over all α in the stated range and the proof appears to require a first step with γ0 = 1.
assumptions (7)
  • domain assumption 1D OT maps between Gaussian marginals are linear with slope sqrt(θ^TΛθ / θ^TΣθ)
    Used in Section 4.4 to derive the covariance recursion (18) and the eigenvalue estimates of Proposition D.1.
  • domain assumption Iterates remain Gaussian (or elliptically contoured) when source and target are Gaussian
    Relies on linearity of slice-matching maps; reduces the whole convergence analysis to spectral control of Σ_k.
  • domain assumption Target covariance is isotropic, Λ = Id, for the main theorem
    Proposition D.2 and Theorem 4.1 require Tr(Σ_k) ≤ Tr(Id) = d and the simplified ratio θ^TΛθ / θ^TΣθ = 1/(θ^TΣθ); the general-covariance case only has unverified sufficient condition (21).
  • domain assumption Random orthonormal bases P_k are i.i.d. Haar and independent of the past sigma-field
    Used in conditional expectation computations in Proposition D.2 and Lemma D.3, and for the unbiased stochastic gradient interpretation.
  • domain assumption Gradient and smoothness calculus for the Sliced-Wasserstein objective
    Invoked in Proposition B.1, Lemma B.5, and Proposition 4.3; the paper notes compact-support caveats and treats Gaussians through the explicit gradient formula.
  • domain assumption Poincaré inequality for the reference measure ν and uniform density bounds mν ≤ σ ≤ Mν
    Used only in the static PL inequality of Proposition 4.3; not needed for the main Gaussian theorem.
  • standard math Standard analytical tools: Robbins-Siegmund theorem, Chung's lemma, Ostrowski's theorem, von Neumann trace inequality
    Standard results used in the appendices; none are ad hoc to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Convergence Rates for Distribution Matching with Sliced Optimal Transport." pith.science (2026). https://pith.science/paper/MNQQ5K7W

@misc{pith2026260210691,
  author       = {Pith},
  title        = {Pith review of: Convergence Rates for Distribution Matching with Sliced Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MNQQ5K7W}},
  note         = {Machine review of arXiv:2602.10691}
}
read the original abstract

We study the slice-matching scheme, an efficient iterative method for distribution matching based on sliced optimal transport. We investigate convergence to the target distribution and derive quantitative non-asymptotic rates. To this end, we establish Lojasiewicz-type inequalities for the Sliced-Wasserstein objective. A key challenge is to control along the trajectory the constants in these inequalities. We show that this becomes tractable for Gaussian distributions. Specifically, eigenvalues are controlled when matching along random orthonormal bases at each iteration. We complement our theory with numerical experiments and illustrate the predicted dependence on dimension and step-size, as well as the stabilizing effect of orthonormal-basis sampling.

Figures

Figures reproduced from arXiv: 2602.10691 by the authors.

Figure 1
Figure 1. Evolution of SW2 2 (σk, µ) when σ = N (0, Σ) and µ = N (0, Id) 0 100 200 300 400 500 0.5 1.0 1.5 2.0 5 10 20 50 75 100 (a) α = 0 0 100 200 300 400 500 0.5 1.0 1.5 2.0 5 10 20 50 75 100 (b) α = 0.1 0 100 200 300 400 500 0.5 1.0 1.5 2.0 5 10 20 50 75 100 (c) α = 0.51 0 100 200 300 400 500 0.5 1.0 1.5 2.0 5 10 20 50 75 100 (d) α = 0.9 [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Minimum and maximum eigenvalues of Σk when σ = N (0, Σ) and µ = N (0, Id) Convergence and impact of (d, α) [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Comparison of sampling strategies: single direction [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Evolution of SW2 2 (σk, µ) for discrete source and target distributions. The source and target samples are distributed from Gaussian mixtures. 5.2 Beyond the Gaussian-to-Gaussian Setting [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Evolution of SW2 2 (σk, µ) when σ = N (0, Σ) and µ = N (0,Λ) F.2 Discrete source and target We also complement [PITH_FULL_IMAGE:figures/full_fig_p038_5.png]
Figure 6
Figure 6. Figure 6: Evolution of SW2 2 (σk, µ) for discrete source and target distributions. The source is sampled from a mixture of Gaussians, and the target is sampled from N (0, Id) Target Source (a) Setting 0 200 400 0.00 0.25 0.50 0.75 1.00 1.25 1.50 5 10 20 50 75 100 (b) α = 0 0 200…
Figure 7
Figure 7. Figure 7: Evolution of SW2 2 (σk, µ) for discrete source and target distributions. The source is sampled from a mixture of Gaussians, and the target is sampled from N (0,Λ) 0 200 400 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 5 10 20 50 75 100 (a) α = 0 0 200 400 0.0 0.2 0.4 0.6 0.8 5 10 2…
Figure 8
Figure 8. Figure 8: Evolution of SW2 2 (σk, µ) when σ = N (0, Σ) and µ = N (0, Id), with slice-matching maps along a single direction θk+1 instead of an orthonormal basis Pk+1 F.3 A single direction for the slice-matching scheme [PITH_FULL_IMAGE:figures/full_fig_p039_8.png]
Figure 9
Figure 9. Figure 9: Minimum and maximum eigenvalues of the estimated covariances [PITH_FULL_IMAGE:figures/full_fig_p040_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 canonical work pages

  1. [1]

    (31) Besides, for any σ, µ∈ P2(Rd), for any T ∈ L2(σ), SW 2 2 (T♯σ, µ) ≤ SW 2 2 (σ, µ) + 2⟨∇W2F (σ), T− Id⟩σ + 1 d ∥T − Id∥2 σ

    and σt = ((1 − t)T0 + tT1)♯σ, SW 2 2 (σt, µ) ≥ (1 − t)SW 2 2 (σ0, µ) + tSW 2 2 (σ1, µ) − t(1 − t) 1 d ∥T0 − T1∥2 σ, (30) and ⟨∇W2F (σ0) ◦ T0 − ∇W2F (σ1) ◦ T1, T0 − T1⟩σ ≤ 1 d ∥T0 − T1∥2 σ. (31) Besides, for any σ, µ∈ P2(Rd), for any T ∈ L2(σ), SW 2 2 (T♯σ, µ) ≤ SW 2 2 (σ, µ) + 2⟨∇W2F (σ), T− Id⟩σ + 1 d ∥T − Id∥2 σ. (32) Lemma B.5 is simply a rewriting of ...

    2025

  2. [2]

    Remark C.3 (Extension to elliptically contoured distributions)

    The final result follows from Tr(Γ)2 ≥ 0 and Tr(Γ2) = dX i=1 (σ2 i − λ2 i )2 = dX i=1 (σi − λi)2(σi + λi)2 ≥ 4m dX i=1 (σi − λi)2 ≥ 4mW 2 2 (µΣ, µΛ) . Remark C.3 (Extension to elliptically contoured distributions) . Proposition C.2 can be readily extended to the class of elliptically contoured distributions whose positive definite parameters are co-diagon...

    2021

  3. [4]

    (2024, Lemma

    deals with a similar recursion, and Jiang et al. (2024, Lemma

    2024

  4. [6]

    Properties of wasserstein gradient flows for the sliced-wasserstein distance

    Christophe Vauthier, Quentin M´erigot, and Anna Korba. Properties of wasserstein gradient flows for the sliced-wasserstein distance. arXiv preprint arXiv:2502.06525,

    arXiv

  5. [8]

    URL https://doi.org/10.1137/24M1656414

    doi: 10.1137/24M1656414. URL https://doi.org/10.1137/24M1656414. Eloi Tanguy, R´emi Flamary, and Julie Delon. Reconstructing discrete measures from projections. consequences on the empirical sliced wasserstein distance. Comptes Rendus. Math´ematique, 362(G10):1121–1129,

    work page doi:10.1137/24m1656414

  6. [9]

    that they verify the generalized parallelogram rule W 2 2 (σt, σ) = (1 − t)W 2 2 (σ0, σ) + tW 2 2 (σ1, σ) − t(1 − t)W 2 2 (σ0, σ1). (26) This can be easily verified by expanding the square in W 2 2 (σt, σ) via (25) and using the tricks t2 = t − t(1 − t) and (1 − t)2 = (1 − t) − t(1 − t), as in Kloeckner (2010, Proposition 4.1). Next, we turn to differenti...

    2010

  7. [10]

    Handbook of convergence theorems for (stochastic) gradient methods

    Guillaume Garrigos and Robert M Gower. Handbook of convergence theorems for (stochastic) gradient methods. arXiv preprint arXiv:2301.11235,

    arXiv

  8. [11]

    Antoine Godichon-Baggioni

    URL https://arxiv.org/abs/2406.05637. Antoine Godichon-Baggioni. Lp and almost sure rates of convergence of averaged stochastic gradient algorithms: locally strongly convex objective. ESAIM: Probability and Statistics, 23: 841–873,

    arXiv

Show all 29 references
  1. [12]

    URL https://doi.org/ 10.1214/20-AOS1987

    doi: 10.1214/20-AOS1987. URL https://doi.org/ 10.1214/20-AOS1987. L´ena¨ıc Chizat, Maria Colombo, and Xavier Fern ´andez-Real. Convergence of drift-diffusion pdes arising as wasserstein gradient flows of convex functions. arXiv preprint arXiv:2507.12385,

    arXiv

  2. [13]

    Luis A Caffarelli

    URL https://arxiv.org/abs/ 2510.16465. Luis A Caffarelli. The regularity of mappings with a convex potential. Journal of the American Mathematical Society, 5(1):99–104,

  3. [15]

    On the fenchel duality between strong convexity and lipschitz continuous gradient

    Xingyu Zhou. On the fenchel duality between strong convexity and lipschitz continuous gradient. arXiv preprint arXiv:1803.06573,

    arXiv

  4. [16]

    For further details, we refer the interesting reader to the classical references Ambrosio and Savar´e (2007); Santambrogio (2015)

    19 A Reminders on Wasserstein space This appendix gathers existing results useful for optimization over the space of probability distri- butions. For further details, we refer the interesting reader to the classical references Ambrosio and Savar´e (2007); Santambrogio (2015). ...

    2007

  5. [17]

    Interestingly enough, W 2 2 (·, σ) is strictly convex along (23) as soon as σ is absolutely continuous (Santambrogio, 2015, Proposition 7.19)

    and it is a building block for gradient flows in the Wasserstein space (Ambrosio and Savar ´e, 2007). Interestingly enough, W 2 2 (·, σ) is strictly convex along (23) as soon as σ is absolutely continuous (Santambrogio, 2015, Proposition 7.19). It is not hard to see that the s...

    2007

  6. [19]

    Thus, it only remains to verify that (61) fulfills the correct requirements

    generalizes this in several ways. Thus, it only remains to verify that (61) fulfills the correct requirements. To stick to the notations of Jiang et al. (2024, Lemma 19), we introduce yk = E[F (σk)], ak = 1/(k + γ)2−2α, ℓ1 = ℓ2 = 1, ℓ3 = C, τ = 2α/(2 − 2α), so that (61) rewrit...

    2024

  7. [20]

    Besides, one can find in Vauthier et al

    applied to Fσ yield ⟨∇Gσ(T0) − ∇Gσ(T1), T0 − T1⟩σ ≥ 0, which directly implies (31). Besides, one can find in Vauthier et al. (2025, Appendix B.6), namely the equations (140) and (155), that, for any σ, µ∈ P2(Rd): (a) for ξ0, ξ1 ∈ L2(σ), ξt = (1 − t)(Id + ξ0) + t(Id + ξ1) and σ...

    2025

  8. [21]

    By Proposition B.8, F (σk) ≤ 2M2(µ)

    yields KX k=0 γkE∥∇W2F (σk)∥2 σk ≤ F (σ0) + KX k=0 γ2 kEF (σk). By Proposition B.8, F (σk) ≤ 2M2(µ). The final result follows from dividing both sides of the above inequality by PK k=1 γk. C Proofs of Section 4: Łojasiewicz inequalities C.1 Proof of Proposition 4.3: a PL-like ...

    2024

  9. [22]

    Combining (38) and (39), SW 2 2 (σ, µ) ≤ W2(σ, µ) Z 1 0 ∥∇Ψ∥ρtdt

    yields Z Ψdσ − Z Ψdµ ≤ W2(σ, µ) Z 1 0 ∥∇Ψ∥ρtdt, (39) where ρt = ((1 − t)Id + tT µ σ )♯σ is the Wasserstein geodesic betweenσ and µ. Combining (38) and (39), SW 2 2 (σ, µ) ≤ W2(σ, µ) Z 1 0 ∥∇Ψ∥ρtdt. Taking the square and applying Jensen’s inequality, SW 4 2 (σ, µ) ≤ W 2 2 (σ, µ...

    2020

  10. [24]

    (45) In particular, Σk+1 is symmetric positive definite. Proof. The distribution σk+1 corresponds to the random vector (1 − γk)X + γkTPk+1(X), (46) where X ∼ N(0, Σk). Also, by definition, TPk+1(X) = dX ℓ=1 θℓtθℓ(X ⊤θℓ) = dX ℓ=1 τθℓθℓθ⊤ ℓ X = Pk+1DkP ⊤ k+1X, where Dk = diag(τθ...

    2021

  11. [25]

    36 Hence, gp/2 k ≤ 1/(k + 1)2α as soon as one chooses p ≥ 4α/(1 − α)

    and gp/2 k = 1 /(k + 1)(1−α)p/2. 36 Hence, gp/2 k ≤ 1/(k + 1)2α as soon as one chooses p ≥ 4α/(1 − α). In this case, (57) becomes, for some C = 2M2(µ)(√cp + 1), E[F (σk+1)] ≤ E[F (σk)] 1 − 1 k + 1 + C (k + 1)2α . The desired rate follows directly from Chung’s Lemma (Chung, 195...

    1954

  12. [26]

    Expected PL-like inequality

    or Bercu and Bigot (2021, Lemma A.3). Expected PL-like inequality. We now turn to show the second result. In the same way than in (55), Assumption A with τ = 2 would imply instead E[F (σk+1)|Ak] ≤ F (σk)1Gk − γkgkF (σk)21Gk + F (σk)1Gc k + F (σk)γ2 k, (58) where, in the last t...

    2021

  13. [27]

    Moulines and Bach (2011, Theorem

    2011

  14. [2011]

    Absolute continuity of wasserstein barycenters on manifolds with a lower ricci curva- ture bound

    Jianyu Ma. Absolute continuity of wasserstein barycenters on manifolds with a lower ricci curva- ture bound. arXiv preprint arXiv:2310.13832,

  15. [2016]

    doi: 10.1007/978-3-319-11259-6 23-1

    ISBN 978-3-319-11259-6. doi: 10.1007/978-3-319-11259-6 23-1. URL https://doi.org/10. 1007/978-3-319-11259-6_23-1 . Louis Grenioux, Alain Oliviero Durmus, Eric Moulines, and Marylou Gabri ´e. On sampling with approximate transport maps. In Andreas Krause, Emma Brunskill, Kyungh...

    doi

  16. [2018]

    Optimization with first order algorithms

    Charles Dossal, Samuel Hurault, and Nicolas Papadakis. Optimization with first order algorithms. arXiv preprint arXiv:2410.19506,

    arXiv

  17. [2021]

    Statistical optimal transport

    14 Sinho Chewi, Jonathan Niles-Weed, and Philippe Rigollet. Statistical optimal transport. arXiv preprint arXiv:2407.18163, 3,

    arXiv

  18. [2022]

    URL https://doi.org/10.1214/22-EJS2001

    doi: 10.1214/22-EJS2001. URL https://doi.org/10.1214/22-EJS2001. Eloi Tanguy. Convergence of SGD for training neural networks with sliced wasserstein losses. Transactions on Machine Learning Research,

    doi

  19. [2023]

    press/v202/grenioux23a.html

    URL https://proceedings.mlr. press/v202/grenioux23a.html. Sven Wang and Youssef Marzouk. On minimax density estimation via measure transport. arXiv preprint arXiv:2207.10231,

    arXiv

  20. [2024]

    Sangmin Park and Dejan Slep ˇcev

    URL https://arxiv.org/abs/2311.15874. Sangmin Park and Dejan Slep ˇcev. Geometry and analytic properties of the sliced wasserstein space. Journal of Functional Analysis, 289(7):110975,

    arXiv

  21. [2025]

    Measure transfer via stochastic slicing and matching.arXiv preprint arXiv:2307.05705,

    Shiying Li and Caroline Moosmueller. Measure transfer via stochastic slicing and matching.arXiv preprint arXiv:2307.05705,

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.