Pith. sign in

REVIEW 3 major objections 3 minor 38 references

Establishing Boundary KKT Convergence of Mirror Descent through Reparameterization

T0 review · 3 major / 3 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read This paper proves that nonconvex mirror descent with a Legendre kernel converges to a boundary KKT point whenever the metric-flattening reparameterization admits a definable continuous boundary extension.

desk verdict Solid paper with a genuinely useful uniform Lagrangian-gradient bound, but the final KKT conclusion leans on an unproved earlier result of the authors that needs to be made self-contained. read the letter →

arxiv 2608.07248 v2 pith:EAVJ2WZP submitted 2026-08-07 math.OC cs.LG

classification math.OCcs.LG MSC 90C2690C3049J52
keywords mirrordescentboundaryKKTstationarityLegendrekernelmetric-flatteningreparameterizationKurdyka–ŁojasiewiczinequalityconformalcircuitdecompositiondefinablefunctionsBregmandivergence
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

Mirror descent is a workhorse for optimization over simplices and polytopes, but when the solution lies on the boundary of the feasible set the Legendre kernel's gradient blows up and the usual stationarity argument breaks. This paper closes that gap for smooth nonconvex problems with linear constraints by changing variables: it reparameterizes each coordinate so that the Hessian metric becomes the identity, and proves that if the inverse of this map extends continuously and definably to the boundary, then the reparameterized iterates have finite length and converge. The original iterates then converge to a point satisfying the Karush–Kuhn–Tucker conditions. This matters because a recent counterexample showed mirror descent can otherwise accumulate at non-KKT boundary points, so the boundary extension condition is what separates guaranteed stationarity from pathological winding.

What carries the argument

The engine is the metric-flattening reparameterization $S_i(x_i)=\int_{\bar x_i}^{x_i}\sqrt{h_i''(u)}\,du$, chosen so that the Hessian metric of the separable Legendre kernel becomes the identity, $D(S^{-1})^T\nabla^2\phi(S^{-1})D(S^{-1})=I$, and therefore cannot degenerate at the boundary. Its inverse must extend continuously and definably to the closure $N=\mathrm{cl}\,S(X^\circ)$; under that extension the reparameterized objective $E(z)=f(S^{-1}(z))$ is definable and satisfies the Kurdyka–Łojasiewicz property, so mirror descent becomes a Euclidean-type descent on $E$. The other essential piece is the conformal circuit decomposition of the mirror-step displacement, which yields a uniform bound on the Lagrangian gradient $\|\nabla f(x)+A^T\lambda\|_\infty\le\Gamma$ independent of distance to the boundary; this uniform bound lets the standard KL finite-length argument run all the way to the boundary.

What would settle it

Find a definable $C^1$ objective, a compact linearly constrained set, and one of the three allowed kernels (Shannon entropy, Fermi–Dirac entropy, or a power kernel with $1\le p<2$) such that mirror descent with $\bar\alpha L<1$ and $\sum\alpha_k=\infty$ accumulates at a boundary point violating the KKT inequalities; that would refute Theorem 4.1. A simpler probe is to run the same experiment with the logarithmic-barrier kernel $h''(t)=t^{-2}$, which fails the boundary-extension condition; if non-KKT accumulation appears there, the excluded case is genuinely different rather than a technicality.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is Theorem 4.1: under a finite curvature bound on the kernel ($\kappa<\infty$), a definable boundary extension of the coordinatewise metric-flattening map $S$, and an objective that is $C^1$, relatively smooth, and definable, mirror descent with stepsizes $0<\alpha_k\le\bar\alpha$, $\bar\alpha L<1$, and $\sum\alpha_k=\infty$ yields reparameterized iterates $z_k=S(x_k)$ with $\sum_k\|z_{k+1}-z_k\|<\infty$. Consequently $z_k\to z^\star$, the original iterates converge to $x^\star=S^{-1}(z^\star)$, and $x^\star$ is a KKT point of the linearly constrained problem. The proof applies the Kurdyka–Łojasiewicz inequality to the reparameterized objective after establishing a boundary-uniform bound on the Lagrangian gradient via a conformal circuit decomposition.

Load-bearing premise

The argument rests on the inverse of the metric-flattening map extending continuously and definably to the boundary of the reparameterized feasible set, a condition that excludes common kernels such as the logarithmic barrier; unless that extension exists the finite-length argument cannot start, and the KKT conclusion additionally inherits a cited proposition from the authors' earlier preprint that is not proved in this manuscript.

Editorial extensions

If this is right

  • For Shannon entropy, Fermi–Dirac entropy, and power kernels on a compact polyhedron, Theorem 4.1 turns mirror descent with diminishing nonsummable stepsizes into a guaranteed boundary-KKT method even when the limit lies on a face.
  • The reparameterized sequence has finite length, so the original iterates form a Cauchy sequence and converge to a single point rather than merely clustering.
  • Coordinatewise rates follow from the KL exponent: coordinates whose limit is on the boundary converge faster (squared rate for Shannon and Fermi–Dirac, rate $q_i=2/(2-p_i)$ for power kernels) than interior coordinates.
  • The same reparameterization proves that mirror-flow trajectories have finite length in reparameterized variables and converge to a KKT point.
  • The stepsize condition $\sum\alpha_k=\infty$ with $\alpha_k\le\bar\alpha<1/L$ permits stepsizes tending to zero, a flexibility not available in some related iterate-convergence frameworks.

Reading between the lines

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

  • The definable boundary-extension condition is essentially an integrability statement: $\int\sqrt{h_i''}$ must be finite so that $S_i$ reaches the boundary, which is why the logarithmic barrier (the $p=2$ power kernel, where $h''(t)=t^{-2}$) is excluded; one testable extension is to replace definability by tameness or semialgebraic regularity and see whether finite length survives.
  • The conformal-circuit bound suggests a transferable recipe for other Bregman-type methods (proximal point, proximal gradient, ADMM): stationarity at boundary limits can be certified through the elementary-vector structure of the constraint set rather than by taming the kernel's gradient blow-up.
  • A practical monitoring rule follows from the coordinatewise rates: if one tracks $S(x_k)$, residuals in boundary coordinates are expected to vanish faster than interior residuals, giving a principled stopping criterion.
  • Because the framework uses ordinary KL on the reparameterized objective, any definable objective with a known KL exponent yields explicit rates; it would be natural to test whether the rates are sharp by constructing objectives whose reparameterized KL exponent reaches the edge of the allowed range.
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 / 3 minor

Summary. The paper studies mirror descent for smooth nonconvex problems with linear constraints and separable Legendre kernels, aiming at convergence to boundary KKT points. It introduces a coordinatewise metric-flattening reparameterization S, assumes a definable boundary extension, and proves finite length of the reparameterized sequence S(x_k), convergence of x_k, and stationarity 0 ∈ ∂E(z*) for the reparameterized objective E. The main tool is a KL finite-length argument fed by a uniform, boundary-independent bound on the Lagrangian gradient obtained through a conformal circuit decomposition. The paper then concludes KKT stationarity of the original limit by invoking [16, Proposition 5.3], rather than by a self-contained proof. Applications are given for Shannon entropy, Fermi–Dirac entropy, and power kernels, together with coordinatewise convergence rates.

Significance. If the KKT step is made fully supported, this is a meaningful advance: it gives a broad positive result on boundary KKT convergence for nonconvex mirror descent, in contrast to the counterexample in [18], and the circuit-based uniform bound on the Lagrangian gradient (Lemma 4.8 and Proposition 4.11) is an elegant and apparently correct technical contribution. The metric-flattening perspective and the explicit coordinatewise rate examples are also useful. The main caveat is that the central KKT conclusion currently rests on an unpublished external proposition, so the significance is conditional on that step being verified or proved.

major comments (3)
  1. [Section 4.2.3, proof of Theorem 4.1] The KKT conclusion of Theorem 4.1 is not derived from the paper's own estimates. The proof establishes finite length of S(x_k), convergence to z* ∈ N, and 0 ∈ ∂E(z*), and then the final sentence invokes [16, Proposition 5.3]. This is load-bearing: equation (18) and the identity in (19) are only valid for z ∈ N∘, and at a boundary point the limiting subdifferential of E can contain extra normal directions coming from the box constraints on z. Therefore 0 ∈ ∂E(z*) does not by itself imply existence of λ* satisfying (2). The manuscript must either prove the needed statement from [16, Proposition 5.3] or give a direct argument for the KKT condition.
  2. [Section 1.1, statement on benign behavior under sequence convergence] The paper asserts without proof that [16, Proposition 5.3] implies every convergent mirror descent sequence with positive nonsummable stepsizes has a KKT limit. Since [16] is an unpublished preprint and the proposition is not stated here, its hypotheses are not checked. In particular, Theorem 4.1 explicitly allows α_k → 0, and the paper's own comparison with [10] emphasizes that its stepsize condition is more flexible than a lower-bound condition; whether [16, Proposition 5.3] covers the diminishing-stepsize case must be verified or proved.
  3. [Appendix B, proof of Theorem 4.3] The proof of Theorem 4.3 also relies on [16, Proposition 3.1] for global existence and uniqueness of the mirror flow in X∘. This is another external result from the authors' earlier preprint that is not proved in this manuscript. If the flow theorem is meant to be a contribution of the paper, the relevant existence statement should be stated as a hypothesis or proved in an appendix.
minor comments (3)
  1. [Section 3, Definition 3.1] The definition of a definable boundary extension includes compactness of N, but the text immediately says compactness is not essential and may be replaced by boundedness of the generated sequence. This should be reconciled, for example by stating the compactness as a standing assumption in the main theorems rather than as part of the definition.
  2. [Section 5.3 vs Proposition 3.3] Proposition 3.3(iii) states power kernels for p ∈ (1,2), while Section 5.3 uses p_i ∈ [1,2). The case p = 1 is already covered by the Shannon entropy example, but the ranges should be aligned or the p = 1 case explicitly referenced so that the reader is not left to infer the extension.
  3. [Section 4.2.3, Lemma 4.12] In the proof of Lemma 4.12, the phrase 'Since each limiting subdifferential is closed in finite dimensions' would be clearer as 'Since each ∂E(z_{k_j}) is closed in finite dimensions'; the current wording is slightly confusing. This is a presentation issue only.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity: the reparameterization and KL argument are self-contained; the final KKT step relies on an external proposition from the authors' own prior work, which is a verification concern but not a circular reduction.

full rationale

The paper's derivation chain is not circular. The reparameterized objective E is defined as f composed with S^{-1} on N, where S is the coordinatewise metric-flattening map and N is the closure of S(X°); these objects are defined from the kernel, the feasible set, and f, not from the desired KKT conclusion. The finite-length argument in the proof of Theorem 4.1 combines sufficient decrease (Lemma 2.6 and Proposition 4.9), a boundary-uniform relative-error bound (Proposition 4.11, obtained from the conformal-circuit Lemma 4.8 and the curvature estimate Lemma 4.10), and the standard KL finite-length Lemma 4.12. None of these steps fits a parameter to the conclusion or defines a quantity in terms of the theorem it is meant to prove. The definable-boundary-extension condition (Definition 3.1) is an explicit structural assumption, and Proposition 3.3 verifies it for the example kernels rather than smuggling it in through a citation. The only load-bearing external input is the use of the authors' own earlier results: [16, Proposition 5.3] supplies the final implication from convergence of a mirror descent sequence to KKT stationarity of its limit, and [16, Proposition 3.1] supplies global existence of the mirror flow in Theorem 4.3. These propositions are quoted as general statements with their own stated hypotheses (positive nonsummable stepsizes and convergence of the sequence), not as redefinitions of the current theorem. Under the reviewing rules, an external citation with independent content is real evidence and does not by itself constitute circularity. The gap here is a verification and completeness risk: the manuscript does not reproduce or verify the hypotheses of [16, Proposition 5.3], so if that proposition does not cover diminishing stepsizes, the KKT conclusion of Theorem 4.1 would be unsupported. But unsupported external support is not the same as a definitional or tautological derivation. For that reason, the appropriate finding is no significant circularity, with a small score reflecting the author-overlap burden on a central step.

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

No free parameters are fitted to data. The central claim rests on standard mathematical tools (KL, circuit decomposition, o-minimality) plus stated structural assumptions on the objective, kernel, and feasible set. The reparameterization map S is a mathematical construction, not a new entity.

assumptions (6)
  • standard math Proper lower-semicontinuous definable functions satisfy the Kurdyka-Lojasiewicz property.
    Used in Section 4.1 to assert E is KL on N and to apply Lemma 4.12 for finite length.
  • standard math Conformal circuit decomposition: every nonzero w in ker A splits into A-circuits cℓ with cℓ⊑w.
    Invoked in Lemma 4.8 to derive the uniform bound on the Lagrangian gradient; cited from Rockafellar [33] and Müller-Regensburger [28].
  • domain assumption Every convergent mirror descent sequence with positive nonsummable stepsizes has a KKT limit ([16, Proposition 5.3]).
    This is the final step of Theorem 4.1, converting sequence convergence to KKT stationarity. It is a result from the authors' own earlier preprint and is not proved in this manuscript.
  • domain assumption The mirror flow (5) has a unique global solution in X◦ for every x(0)∈X◦ ([16, Proposition 3.1]).
    Used in Theorem 4.3 to start the trajectory analysis; cited from the authors' prior work.
  • domain assumption f is L-smooth relative to φ on X◦ and f|X is definable in the common o-minimal structure.
    Assumed in Theorem 4.1 to obtain sufficient decrease (Lemma 2.6) and to make E definable and KL.
  • domain assumption The curvature bound κ<∞ holds and S admits a definable boundary extension.
    These are the joint structural conditions of Theorem 4.1, verified for the concrete kernels in Section 5 but not automatic in general.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Establishing Boundary KKT Convergence of Mirror Descent through Reparameterization." pith.science (2026). https://pith.science/paper/EAVJ2WZP

@misc{pith2026260807248,
  author       = {Pith},
  title        = {Pith review of: Establishing Boundary KKT Convergence of Mirror Descent through Reparameterization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EAVJ2WZP}},
  note         = {Machine review of arXiv:2608.07248}
}
read the original abstract

Sequence convergence to a boundary Karush--Kuhn--Tucker (KKT) point has long remained unclear for nonconvex mirror descent with Legendre kernels. The difficulty arises from the blow-up of the gradient of the Legendre kernel at the boundary. Recent work~\cite{dingtoh2026nonkkt} shows that mirror descent can accumulate at non-KKT boundary points despite decreasing objective values, precluding a convergence guarantee to KKT points in general. Despite this negative result, mirror descent remains effective in many real applications. Motivated by this contrast, we address the boundary difficulty directly and establish KKT convergence of mirror descent for a broad class of structured nonconvex problems. We analyze mirror descent in reparameterized variables, where the Hessian metric is flattened and remains nondegenerate as the boundary is approached. Under extension and definability conditions jointly coupling the objective, the Legendre kernel, and the feasible region, the reparameterized sequence has finite length and converges, thereby recovering convergence to a KKT point of the original sequence. Our general framework applies to some concrete instances: Shannon entropy, Fermi--Dirac entropy, and power kernels on polyhedron.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 4 canonical work pages

  1. [10]

    A Unified Framework for Iterate Convergence of Bregman Proximal Methods

    He Chen, Jiaming Fan, and Anthony Man-Cho So. A unified framework for iterate convergence of Bregman proximal methods.arXiv preprint arXiv:2608.05536, 2026

  2. [18]

    Non-KKT accumulation in entropic mirror descent, 2026

    Kuangyu Ding and Kim-Chuan Toh. Non-KKT accumulation in entropic mirror descent, 2026. arXiv:2608.01658

  3. [16]

    On exploration of an interior mirror descent flow for stochastic nonconvex constrained problem.arXiv preprint arXiv:2507.15264v3, 2025

    Kuangyu Ding and Kim-Chuan Toh. On exploration of an interior mirror descent flow for stochastic nonconvex constrained problem.arXiv preprint arXiv:2507.15264v3, 2025

  4. [1]

    Hessian Riemannian gradient flows in convex programming

    Felipe Alvarez, J´ erˆ ome Bolte, and Olivier Brahic. Hessian Riemannian gradient flows in convex programming. SIAM Journal on Control and Optimization, 43(2):477–501, 2004

  5. [2]

    Singular Riemannian barrier methods and gradient-projection dynamical systems for constrained optimization.Optimization, 53(5–6):435–454, 2004

    Hedy Attouch, J´ erˆ ome Bolte, Patrick Redont, and Marc Teboulle. Singular Riemannian barrier methods and gradient-projection dynamical systems for constrained optimization.Optimization, 53(5–6):435–454, 2004

  6. [3]

    Hedy Attouch, J´ erˆ ome Bolte, and Benar Fux Svaiter. Convergence of descent methods for semi-algebraic and tame problems: Proximal algorithms, forward–backward splitting, and regularized Gauss–Seidel methods.Mathematical Programming, 137:91–129, 2013

  7. [4]

    Bauschke, J´ erˆ ome Bolte, and Marc Teboulle

    Heinz H. Bauschke, J´ erˆ ome Bolte, and Marc Teboulle. A descent lemma beyond Lipschitz gradient continuity: First-order methods revisited and applications.Mathematics of Operations Research, 42(2):330–348, 2017

  8. [5]

    J´ erˆ ome Bolte, Aris Daniilidis, and Adrian S. Lewis. The Lojasiewicz inequality for nonsmooth subanalytic functions with applications to subgradient dynamical systems.SIAM Journal on Optimization, 17(4):1205–1223, 2007

Show all 38 references
  1. [6]

    Lewis, and Masahiro Shiota

    J´ erˆ ome Bolte, Aris Daniilidis, Adrian S. Lewis, and Masahiro Shiota. Clarke subgradients of stratifiable functions. SIAM Journal on Optimization, 18(2):556–572, 2007

  2. [7]

    Curiosities and counterexamples in smooth convex optimization.Mathematical Programming, 195(1–2):553–603, 2022

    J´ erˆ ome Bolte and Edouard Pauwels. Curiosities and counterexamples in smooth convex optimization.Mathematical Programming, 195(1–2):553–603, 2022

  3. [8]

    First order methods beyond convexity and Lipschitz gradient continuity with applications to quadratic inverse problems.SIAM Journal on Optimization, 28(3):2131–2151, 2018

    J´ erˆ ome Bolte, Shoham Sabach, Marc Teboulle, and Yakov Vaisbourd. First order methods beyond convexity and Lipschitz gradient continuity with applications to quadratic inverse problems.SIAM Journal on Optimization, 28(3):2131–2151, 2018

  4. [9]

    Bomze, Panayotis Mertikopoulos, Werner Schachinger, and Mathias Staudigl

    Immanuel M. Bomze, Panayotis Mertikopoulos, Werner Schachinger, and Mathias Staudigl. Hessian barrier algorithms for linearly constrained optimization problems.SIAM Journal on Optimization, 29(3):2100–2127, 2019

  5. [11]

    Spurious stationarity and hardness results for Bregman proximal- type algorithms.arXiv preprint arXiv:2404.08073, 2026

    He Chen, Jiajin Li, and Anthony Man-Cho So. Spurious stationarity and hardness results for Bregman proximal- type algorithms.arXiv preprint arXiv:2404.08073, 2026

  6. [12]

    On the iterate convergence of Bregman projected gradient method.arXiv preprint arXiv:2608.05035, 2026

    He Chen and Anthony Man-Cho So. On the iterate convergence of Bregman projected gradient method.arXiv preprint arXiv:2608.05035, 2026

  7. [13]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems 26, 2013

  8. [14]

    Dang and Guanghui Lan

    Cong D. Dang and Guanghui Lan. On the convergence properties of non-Euclidean extragradient methods for variational inequalities with generalized monotone operators.Computational Optimization and Applications, 60(2):277–310, 2015

  9. [15]

    Nonconvex stochastic Bregman proximal gradient method with application to deep learning.Journal of Machine Learning Research, 26(39):1–44, 2025

    Kuangyu Ding, Jingyang Li, and Kim-Chuan Toh. Nonconvex stochastic Bregman proximal gradient method with application to deep learning.Journal of Machine Learning Research, 26(39):1–44, 2025

  10. [17]

    Stochastic Bregman subgradient methods for nonsmooth nonconvex optimization problems.Journal of Optimization Theory and Applications, 206(3):67, 2025

    Kuangyu Ding and Kim-Chuan Toh. Stochastic Bregman subgradient methods for nonsmooth nonconvex optimization problems.Journal of Optimization Theory and Applications, 206(3):67, 2025

  11. [19]

    Doan, Subhonmesh Bose, D

    Thinh T. Doan, Subhonmesh Bose, D. Hoa Nguyen, and Carolyn L. Beck. Convergence of the iterates in mirror descent methods.IEEE Control Systems Letters, 3(1):114–119, 2019

  12. [20]

    A bregman ADMM for Bethe variational problem.arXiv preprint arXiv:2502.04613, 2025

    Yuehaw Khoo, Tianyun Tang, and Kim-Chuan Toh. A bregman ADMM for Bethe variational problem.arXiv preprint arXiv:2502.04613, 2025

  13. [21]

    On gradients of functions definable in O-minimal structures.Annales de l’Institut Fourier, 48(3):769–783, 1998

    Krzysztof Kurdyka. On gradients of functions definable in O-minimal structures.Annales de l’Institut Fourier, 48(3):769–783, 1998

  14. [22]

    A convergent single-loop algorithm for relaxation of Gromov–Wasserstein in graph data

    Jiajin Li, Jianheng Tang, Lemin Kong, Huikang Liu, Jia Li, Anthony Man-Cho So, and Jose Blanchet. A convergent single-loop algorithm for relaxation of Gromov–Wasserstein in graph data. InInternational Conference on Learning Representations, 2023. 22

  15. [23]

    Convergence of the exponentiated gradient method with Armijo line search

    Yen-Huan Li and Volkan Cevher. Convergence of the exponentiated gradient method with Armijo line search. Journal of Optimization Theory and Applications, 181(2):588–607, 2019

  16. [24]

    Lee, and Sanjeev Arora

    Zhiyuan Li, Tianhao Wang, Jason D. Lee, and Sanjeev Arora. Implicit bias of gradient descent on reparametrized models: On equivalence to mirror descent. InAdvances in Neural Information Processing Systems 35, pages 34626–34640, 2022

  17. [25]

    Freund, and Yurii Nesterov

    Haihao Lu, Robert M. Freund, and Yurii Nesterov. Relatively smooth convex optimization by first-order methods, and applications.SIAM Journal on Optimization, 28(1):333–354, 2018

  18. [26]

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

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

  19. [27]

    Global convergence of model function based Bregman proximal minimization algorithms.Journal of Global Optimization, 83(4):753–781, 2022

    Mahesh Chandra Mukkamala, Jalal Fadili, and Peter Ochs. Global convergence of model function based Bregman proximal minimization algorithms.Journal of Global Optimization, 83(4):753–781, 2022

  20. [28]

    Elementary vectors and conformal sums in polyhedral geometry and their relevance for metabolic pathway analysis.Frontiers in Genetics, 7:90, 2016

    Stefan M¨ uller and Georg Regensburger. Elementary vectors and conformal sums in polyhedral geometry and their relevance for metabolic pathway analysis.Frontiers in Genetics, 7:90, 2016

  21. [29]

    Computational optimal transport.Foundations and Trends in Machine Learning, 11(5–6):355–607, 2019

    Gabriel Peyr´ e and Marco Cuturi. Computational optimal transport.Foundations and Trends in Machine Learning, 11(5–6):355–607, 2019

  22. [30]

    Gromov–Wasserstein averaging of kernel and distance matrices

    Gabriel Peyr´ e, Marco Cuturi, and Justin Solomon. Gromov–Wasserstein averaging of kernel and distance matrices. InProceedings of the 33rd International Conference on Machine Learning, volume 48 ofProceedings of Machine Learning Research, pages 2664–2672, 2016

  23. [31]

    Shuffling the stochastic mirror descent via dual lipschitz continuity and kernel conditioning.arXiv preprint arXiv:2603.16042, 2026

    Junwen Qiu, Leilei Mei, and Junyu Zhang. Shuffling the stochastic mirror descent via dual lipschitz continuity and kernel conditioning.arXiv preprint arXiv:2603.16042, 2026

  24. [32]

    Entropic Gromov–Wasserstein distances: Stability and algorithms

    Gabriel Rioux, Ziv Goldfeld, and Kengo Kato. Entropic Gromov–Wasserstein distances: Stability and algorithms. Journal of Machine Learning Research, 25(363):1–52, 2024

  25. [33]

    Tyrrell Rockafellar

    R. Tyrrell Rockafellar. The elementary vectors of a subspace of RN . In R. C. Bose and T. A. Dowling, editors, Combinatorial Mathematics and Its Applications, pages 104–127. University of North Carolina Press, Chapel Hill, NC, 1969

  26. [34]

    Tyrrell Rockafellar.Convex Analysis, volume 28 ofPrinceton Mathematical Series

    R. Tyrrell Rockafellar.Convex Analysis, volume 28 ofPrinceton Mathematical Series. Princeton University Press, Princeton, NJ, 1970

  27. [35]

    Tyrrell Rockafellar and Roger J.-B

    R. Tyrrell Rockafellar and Roger J.-B. Wets.Variational Analysis. Springer, Berlin, 1998

  28. [36]

    Linear-time Gromov–Wasserstein distances using low-rank couplings and costs

    Meyer Scetbon, Gabriel Peyr´ e, and Marco Cuturi. Linear-time Gromov–Wasserstein distances using low-rank couplings and costs. InProceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 19347–19365, 2022

  29. [37]

    On the convergence rate of stochastic mirror descent for nonsmooth nonconvex optimization

    Siqi Zhang and Niao He. On the convergence rate of stochastic mirror descent for nonsmooth nonconvex optimization. arXiv:1806.04781, 2018

  30. [38]

    Boyd, and Peter W

    Zhengyuan Zhou, Panayotis Mertikopoulos, Nicholas Bambos, Stephen P. Boyd, and Peter W. Glynn. On the convergence of mirror descent beyond stochastic convex programming.SIAM Journal on Optimization, 30(1):687–716, 2020. 23

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.