Pith. sign in

REVIEW 2 major objections 5 minor 64 references

A Unified Framework for Iterate Convergence of Bregman Proximal Methods

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

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

Pith's one-line read This paper proves that bounded Bregman proximal iterates converge to stationary points for nonconvex composite objectives, under an extended scaled Kurdyka–Łojasiewicz property and mild domain regularity.

desk verdict Substantial framework and a valuable mirror-flow result, but the main BPGM theorem has a proof gap in the H2 verification that needs fixing before the headline claim holds. read the letter →

arxiv 2608.05536 v1 pith:C4WR43ZS submitted 2026-08-06 math.OC

He Chen , Jiaming Fan , Anthony Man-Cho So This is my paper · ORCID
classification math.OC MSC 90C2690C2549J52
keywords BregmanproximalmethodsiterateconvergencescaledKurdyka–Łojasiewiczpropertynonconvexoptimizationmirrorflowsubanalyticfunctionso-minimalstructuresseparablekernels
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

The paper sets out to close a gap that has been open since the 1990s: proving that the actual iterates of Bregman proximal methods converge for nonconvex composite problems, not just their function values. It claims that convergence is governed by an extended scaled Kurdyka–Łojasiewicz property under the kernel, and that this property holds for every continuous subanalytic objective whenever the kernel's domain is closed. Under a regularity condition on the feasible domain, both the Bregman proximal point method and the Bregman proximal gradient method are shown to generate bounded sequences that converge to a stationary point with finite length. The same parameterization machinery yields the first trajectory convergence result for the continuous-time mirror flow without convexity or isolation assumptions.

What carries the argument

Two linked devices carry the argument. The first is the extended scaled Kurdyka–Łojasiewicz property: at each $x^*\in\operatorname{dom}(\partial H)$ there is a desingularizing function $\zeta$ with $\zeta'(H(x)-H(x^*))\,\operatorname{dist}(0,\nabla^2 h(x)^{-1/2}\partial H(x))\ge 1$ near $x^*$, paired with scaled sufficient decrease and scaled relative error inequalities whose common scaling matrix is the kernel Hessian at the new iterate, $\nabla^2 h(x_{k+1})^{1/2}$. The second is the kernel-dependent parameterization function $\psi(x)=\int_{x_0}^{x}\sqrt{\varphi''(s)}\,ds$ with inverse $\phi=\psi^{-1}$, which pulls the Bregman geometry back to Euclidean coordinates; it is the tool that proves Proposition 4 (subanalytic functions are SKŁ when the kernel domain is closed) and that turns mirror flow into the Euclidean gradient flow of $F\circ\phi$. Assumption 2—polyhedral domains, or closed convex domains with a normal-cone qualification—supplies the uniform bound on kernel-gradient gaps that makes (H1) and (H2) hold for BPPM and BPGM.

What would settle it

Find a bounded BPGM sequence for a continuous subanalytic objective with a closed-domain separable kernel, obeying the paper's step-size bounds, whose limit set contains two distinct points; the finite-length conclusion of Theorem 3 rules this out. A concrete place to look is the paper's own Example 4, where the Shannon kernel on $D=B((1,1),1)$ with $F(x)=-x_2+\delta_D(x)$ makes $\|\nabla h(x_{k+1})-\nabla h(x_k)\|_2\to+\infty$, showing why that problem sits outside Assumption 2.

Watch

Extended reading notes

Core claim

The central claim is that Bregman proximal methods inherit the full convergence theory of Euclidean proximal methods once the objective satisfies a scaled variant of the Kurdyka–Łojasiewicz property with respect to the kernel. Theorem 1 says any bounded sequence obeying scaled sufficient decrease (H1), scaled relative error (H2), and limiting stationarity (H3) converges to a stationary point with finite length. Theorems 2 and 3 verify (H1)–(H3) for bounded BPPM and BPGM sequences with step sizes $0<\alpha\le\alpha_k\le\bar\alpha$ (and $\bar\alpha<1/L$ for BPGM), giving what the authors state as the first iterate convergence guarantees for these methods over general closed-domain separable kernels and continuous subanalytic composite objectives. Proposition 4 supplies the key hypothesis: under closed domains, every continuous subanalytic function is SKŁ under the kernel. For the mirror flow, the paper transforms the trajectory through the kernel's parameterization function into a Euclidean gradient flow of $F\circ\phi$ and proves convergence to a stationary point for o-minimal definable objectives, without convexity or isolation assumptions.

Load-bearing premise

The load-bearing premise is that the feasible domain is well-posed relative to the kernel—polyhedral, or satisfying a normal-cone qualification—so that kernel-gradient gaps between successive iterates stay bounded; Example 4 shows that without this the scaled sufficient decrease and relative error inequalities can fail.

Editorial extensions

If this is right

  • Every bounded BPPM sequence satisfying the paper's assumptions converges to a stationary point with finite path length, with no convexity required.
  • Every bounded BPGM sequence converges under the same conditions when the step sizes stay below $1/L$ in the relative-smoothness sense.
  • For closed-domain separable kernels, every continuous subanalytic composite objective satisfies the required SKŁ property, so the convergence theorems apply to a very broad class of nonconvex problems.
  • The mirror flow result removes the convexity and isolation assumptions used in earlier trajectory convergence results for continuous-time Bregman dynamics.
  • The framework generalizes the classical Euclidean Kurdyka–Łojasiewicz convergence framework, adding the limiting-stationarity condition (H3) to rule out spurious stationary points.

Reading between the lines

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

  • Editorial extension: the change of variables $x=\phi(y)$ suggests that the SKŁ exponent of $F$ under $h$ should match the Euclidean KŁ exponent of $F\circ\phi$, which would convert these qualitative convergence results into convergence-rate estimates.
  • Editorial extension: the open-versus-closed kernel domain split appears to be the true boundary of the phenomenon; the Burg-entropy examples indicate that open-domain kernels may need a boundary-aware definition of stationarity rather than a sharper desingularizing function.
  • Editorial extension: the uniform kernel-gradient gap in Assumption 2 points toward a natural route for stochastic or inexact Bregman variants—keep that gap controlled in expectation, and the same three conditions should yield iterate convergence.
  • The paper itself flags that removing the closed-domain restriction, dropping the domain regularity condition, and obtaining non-asymptotic rates remain open; these are stated limitations rather than results.
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 develops a unified convergence framework for Bregman proximal methods (BPMs). It introduces an extended scaled Kurdyka-\L{}ojasiewicz (SK\L{}) property for separable kernels, proves this property for all continuous subanalytic functions when the kernel domain is closed, and verifies the three framework conditions (H1)--(H3) for the Bregman proximal point method and the Bregman proximal gradient method under regularity assumptions on the domain. It then uses kernel parameterization functions to transform the mirror flow into a Euclidean subgradient flow and obtains trajectory convergence for definable objectives with closed kernel domains. The stated headline results are Theorems 2 and 3 for iterate convergence and Theorem 4 for mirror-flow trajectory convergence.

Significance. If the technical issues below are resolved, this would be a substantial contribution to the convergence theory of Bregman proximal methods. The parameterization-function construction is a genuine new tool that connects the Bregman geometry to Euclidean K\L{} analysis, and the paper gives detailed proofs rather than fitting parameters to a preordained conclusion. The claimed scope -- general closed-domain separable kernels, composite objectives, and both discrete and continuous time -- is significantly broader than prior work, and the paper is honest about the limitations imposed by Assumption 2 and by open-domain kernels (Examples 2, 4, and 5). The contribution is therefore of clear interest to the optimization community, provided Theorem 3 can be repaired.

major comments (2)
  1. [Section 4.6.2 (Theorem 3)] The verification of (H2) for BPGM applies Proposition 7 (with Remark 1) to f, but Proposition 7 requires both Lh+f and Lh-f to be convex. Theorem 3 assumes only that Lh+F and Lh-F are convex, where F=f+g with g convex. From this one obtains Lh-f = (Lh-F)+g convex, but Lh+f = (Lh+F)-g need not be convex. For example, with h(x)=x^2/2, L=1, f(x)=-0.9 x^2, and g(x)=x^2, the theorem's assumptions hold while Lh+f = -0.4 x^2 is concave. Thus the displayed bound after the triangle inequality in Section 4.6.2 is not justified as written. A separate argument using Lh+F and Lh-F rather than Lh+f and Lh-f is needed, or the assumptions of Theorem 3 must be strengthened. Since Theorem 3 is one of the paper's main iterate-convergence results, this gap is load-bearing.
  2. [Theorem 1, Eq. (10)] The proof asserts a uniform strong-convexity bound \nabla^2 h(x) \succeq \sigma I on the whole ball B(x^*,d). If x^* lies on the boundary of dom(h), this ball may intersect the boundary, where \nabla^2 h is not defined. The proof only needs the bound for the iterates x_t, which lie in int(dom(h)), so Eq. (10) should be stated on B(x^*,d) \cap int(dom(h)) or the ball should be chosen inside int(dom(h)). This is a local rigor fix, but it should be made in the final version.
minor comments (5)
  1. [Section 5.3 (proof of Theorem 4)] The text refers to 'Proposition 5 (ii)' for the boundedness of y(t) and to 'Proposition 5 (iii)' for definability of the restriction of phi, but Proposition 5 has only items (i) and (ii). The correct references are Proposition 10 (ii) and Proposition 10 (i).
  2. [Lemma 1] In the proof of Lemma 1, the index sets I and J are defined identically as {i : \bar z_i \in int(dom(varphi))}; J should be the set of boundary indices. As printed, the proof is unreadable.
  3. [Section 4.4.2] In Case 2 of the proof of Proposition 8, the expression 'varphi_i(x_{k_l}) - varphi_i(x_{k_l+1})' should be 'varphi'(x_{k_l,i}) - varphi'(x_{k_l+1,i})' in order to justify the conclusion p^*_I = 0.
  4. [Definition 4 and Example 3] The symbol phi is used both for the univariate kernel component and for the newly introduced parameterization function. This clash becomes confusing in Example 3, where phi(x)=x log x and phi(y)=y^2 appear in the same paragraph. Rename one of the two functions.
  5. [Section 4.6.3] The verification of (H3) for BPGM is dismissed with 'the arguments are nearly the same' and omitted. Since Theorem 3 is a central result, the proof should either include the modified limit argument or at least spell out exactly which quantities are replaced and why the boundedness of {\nabla f(x_k)} suffices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: main theorems are proven from external KŁ results and the paper's own verifications; self-citations are not load-bearing.

full rationale

The paper's main derivation chain is not circular. Theorem 1 is a self-contained descent/relative-error/KŁ argument: given (H1)–(H3) and the extended SKŁ property, the proof bounds the total path length and uses (H3) only to upgrade convergence to stationarity; no conclusion is assumed in the hypotheses. The extension of SKŁ to closed-domain separable kernels (Proposition 4) is proved by constructing parameterization functions (Definition 4), composing the objective with a semialgebraic map G, and invoking the standard Łojasiewicz inequality for subanalytic functions [14, Theorem 3.1]; the growth estimates for φ'' (Proposition 1) and φ (Proposition 5) are proved in the appendix from Assumption 1(i) and external KŁ tools, not from the target theorem. The verifications of (H1), (H2), and (H3) for BPPM (Theorem 2) and BPGM (Theorem 3) use only the algorithm's optimality conditions, Assumption 2, and the paper's own Propositions 6–8; the only places where [25] and [26] are invoked are motivational (the SKŁ name and the spurious-stationarity phenomenon), and the formal proofs do not rely on those preprints as black boxes. The continuous-time result similarly uses the known mirror-descent/reparameterization equivalence (Proposition 9, cf. [43]) and external KŁ/definability theorems [14, 15], while proving the needed definability of ψ and φ (Proposition 10) and stationarity of limits (Lemma 5) inside the paper. There are no fitted parameters, no quantity is defined in terms of the target output, and no prediction is forced by construction. The skeptical concern about Theorem 3's H2 verification — that Proposition 7 is applied to f using assumptions made only on F = f + g — is a potential correctness gap in the proof, not a circular reduction; it would require repairing the argument, but it does not make the theorem's conclusion an input. Accordingly, the circularity score is 0.

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

No free parameters are fitted; the paper is a pure theory contribution. The listed axioms are the standard o-minimal/subanalytic infrastructure and the domain assumptions on the kernel, the objective, and the domain set. No new physical or mathematical entities are postulated beyond the parameterization functions, which are constructed from the kernel.

assumptions (7)
  • standard math O-minimal structures S(Ran) and S(Ran,exp) exist and definable sets have the stated closure properties
    Used in Section 2 and Proposition 10 to establish definability of parameterization functions.
  • standard math KŁ inequality holds for subanalytic functions (Bolte et al. [14]) and for o-minimal definable functions ([15])
    Central to Proposition 4 and Theorem 4.
  • standard math Nonsmooth subdifferential chain rule and sum rules from Rockafellar-Wets [50]
    Used in Proposition 4 proof and Lemma 3.
  • domain assumption Assumption 1: kernel is separable with 1/φ'' globally subanalytic and continuous on cl(dom(φ)), f∈C¹, g convex locally Lipschitz
    Defines the class of kernels and objectives covered.
  • domain assumption Assumption 2: dom(F) polyhedral or satisfies the normal-cone qualification N_F(x)∩span{e_j: x_j∈bd(dom(φ))}={0}
    Needed for Proposition 8 to bound the kernel gradient gap; Example 4 shows failure without it.
  • domain assumption Boundedness of the BPM sequence or mirror trajectory
    Assumed in Theorems 1-4; not proved.
  • domain assumption Definability of F in S(Ran,exp) for the mirror flow result
    Theorem 4 requires F definable in S(Ran,exp).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Unified Framework for Iterate Convergence of Bregman Proximal Methods." pith.science (2026). https://pith.science/paper/C4WR43ZS

@misc{pith2026260805536,
  author       = {Pith},
  title        = {Pith review of: A Unified Framework for Iterate Convergence of Bregman Proximal Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4WR43ZS}},
  note         = {Machine review of arXiv:2608.05536}
}
read the original abstract

Iterate convergence of Bregman proximal methods (BPMs) has long remained open, especially for nonconvex objectives. Recently, \citet{chen2026skl} made progress by establishing iterate convergence for a BPM via the so-called scaled Kurdyka-\L{}ojasiewicz (SK\L{}) property, but only for the Shannon entropy kernel and linearly constrained problems. In this paper, we develop a unified iterate convergence framework that applies to a broad group of kernels and composite objective functions. Our approach extends the analytical tools in \cite{chen2026skl}, in particular the SK\L{} property, which plays a central role in ensuring convergence of the generated sequences. By introducing kernel-dependent parameterization functions, we show that the extended SK\L{} property holds for all continuous subanalytic functions, particularly when the kernel has a closed domain. We then verify that the assumptions of the framework are satisfied by standard BPMs under mild regularity conditions, thereby establishing their iterate convergence for a wide range of objective functions. Furthermore, based on the parameterization functions, we show that the continuous-time BPM (mirror flow) converges to a stationary point for o-minimal definable objective functions, yielding the first trajectory convergence result for mirror flow without imposing convexity assumptions on the objective function or isolation assumptions on stationary points. Taken together, these discrete- and continuous-time convergence results provide a unified trajectory convergence theory for BPMs.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 44 canonical work pages

  1. [1]

    Convergence of the iterates of descent methods for analytic cost functions.SIAM Journal on Optimization, 16(2): 531–547, 2005

    Pierre-Antoine Absil, Robert Mahony, and Ben Andrews. Convergence of the iterates of descent methods for analytic cost functions.SIAM Journal on Optimization, 16(2): 531–547, 2005

    work page 2005

  2. [2]

    Hessian Riemannian gradient flows in convex programming.SIAM Journal on Control and Optimization, 43(2):477–501, 2004

    Felipe Alvarez, Jérôme Bolte, and Olivier Brahic. Hessian Riemannian gradient flows in convex programming.SIAM Journal on Control and Optimization, 43(2):477–501, 2004. 49

    work page 2004

  3. [3]

    Hessian Riemannian gradient flows in convex programming.SIAM Journal on Control and Optimization, 43:477–501, 2018

    Felipe Alvarez, Jérôme Bolte, and Olivier Brahic. Hessian Riemannian gradient flows in convex programming.SIAM Journal on Control and Optimization, 43:477–501, 2018

    work page 2018

  4. [4]

    Regularized Lotka-Volterra dynamical system as continuous proximal-like method in optimization.Journal of Optimization Theory and Applications, 121:541–570, 2004

    Hédy Attouch and Marc Teboulle. Regularized Lotka-Volterra dynamical system as continuous proximal-like method in optimization.Journal of Optimization Theory and Applications, 121:541–570, 2004

    work page 2004

  5. [5]

    Hedy Attouch, Jérôme 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, Series A, 137(1): 91–129, 2013

    2013

  6. [6]

    The rate of convergence of Bregman proximal methods: Local geometry versus regularity versus sharpness.SIAM Journal on Optimization, 34(3):2440–2471, 2024

    Waïss Azizian, Franck Iutzeler, Jérôme Malick, and Panayotis Mertikopoulos. The rate of convergence of Bregman proximal methods: Local geometry versus regularity versus sharpness.SIAM Journal on Optimization, 34(3):2440–2471, 2024

    2024

  7. [7]

    Fast composite optimization and statistical recovery in federated learning

    Yajie Bao, Michael Crawshaw, Shan Luo, and Mingrui Liu. Fast composite optimization and statistical recovery in federated learning. InProceedings of the 39th International Conference on Machine Learning, pages 1508–1536. PMLR, 2022

    work page 2022

  8. [8]

    Bauschke, Jérôme Bolte, and Marc Teboulle

    Heinz H. Bauschke, Jérôme 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

    work page 2017

Show all 64 references
  1. [9]

    Heinz H Bauschke, Jérôme Bolte, Jiawei Chen, Marc Teboulle, and Xianfu Wang. On linear convergence of non-Euclidean gradient methods without strong convexity and Lipschitz gradient continuity.Journal of Optimization Theory and Applications, 182(3): 1068–1087, 2019

    2019

  2. [10]

    MOS-SIAM Series on Optimization

    Amir Beck.First-Order Methods in Optimization. MOS-SIAM Series on Optimization. Society for Industrial and Applied Mathematics, Philadelphia, Pennsylvania, 2017

    2017

  3. [11]

    Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175, 2003

    Amir Beck and Marc Teboulle. Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175, 2003

    2003

  4. [12]

    Barrier operators and associated gradient-like dy- namical systems for constrained minimization problems.SIAM Journal on Control and Optimization, 42:1266–1292, 2003

    Jérôme Bolte and Marc Teboulle. Barrier operators and associated gradient-like dy- namical systems for constrained minimization problems.SIAM Journal on Control and Optimization, 42:1266–1292, 2003

    2003

  5. [13]

    Barrier operators and associated gradient-like dy- namical systems for constrained minimization problems.SIAM journal on control and optimization, 42(4):1266–1292, 2003

    Jérôme Bolte and Marc Teboulle. Barrier operators and associated gradient-like dy- namical systems for constrained minimization problems.SIAM journal on control and optimization, 42(4):1266–1292, 2003

    2003

  6. [14]

    The Łojasiewicz inequality for nons- mooth subanalytic functions with applications to subgradient dynamical systems.SIAM Journal on Optimization, 17(4):1205–1223, 2007

    Jérôme Bolte, Aris Daniilidis, and Adrian Lewis. The Łojasiewicz inequality for nons- mooth subanalytic functions with applications to subgradient dynamical systems.SIAM Journal on Optimization, 17(4):1205–1223, 2007. 50

    2007

  7. [15]

    Clarke subgradients of stratifiable functions.SIAM Journal on Optimization, 18(2):556–572, 2007

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

    2007

  8. [16]

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

    Jérôme 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

    2018

  9. [17]

    Hessian barrier algorithms for linearly constrained optimization problems

    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

    2019

  10. [18]

    Lev M Bregman. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming.USSR Computational Mathematics and Mathematical Physics, 7(3):200–217, 1967

    1967

  11. [19]

    Sparse and stable markowitz portfolios.Proceedings of the National Academy of Sciences, 106(30):12267–12272, 2009

    Joshua Brodie, Ingrid Daubechies, Christine De Mol, Domenico Giannone, and Ignace Loris. Sparse and stable markowitz portfolios.Proceedings of the National Academy of Sciences, 106(30):12267–12272, 2009

    2009

  12. [20]

    Theory of convex optimization for machine learning.arXiv preprint arXiv:1405.4980, 15:2, 2014

    Sébastien Bubeck. Theory of convex optimization for machine learning.arXiv preprint arXiv:1405.4980, 15:2, 2014

    2014 arXiv

  13. [21]

    The developments of proximal point algorithms.Journal of the Operations Research Society of China, 10(2):197–239, 2022

    Xing-Ju Cai, Ke Guo, Fan Jiang, Kai Wang, Zhong-Ming Wu, and De-Ren Han. The developments of proximal point algorithms.Journal of the Operations Research Society of China, 10(2):197–239, 2022

    2022

  14. [22]

    Robust uncertainty princi- ples: Exact signal reconstruction from highly incomplete frequency information.IEEE Transactions on Information Theory, 52(2):489–509, 2006

    Emmanuel J Candès, Justin Romberg, and Terence Tao. Robust uncertainty princi- ples: Exact signal reconstruction from highly incomplete frequency information.IEEE Transactions on Information Theory, 52(2):489–509, 2006

    2006

  15. [23]

    Proximal minimization algorithm with D- functions.Journal of Optimization Theory and Applications, 73(3):451–464, 1992

    Yair Censor and Stavros Andrea Zenios. Proximal minimization algorithm with D- functions.Journal of Optimization Theory and Applications, 73(3):451–464, 1992

    1992

  16. [24]

    Convergence analysis of a proximal-like minimization algorithm using Bregman functions.SIAM Journal on Optimization, 3(3):538–543, 1993

    Gong Chen and Marc Teboulle. Convergence analysis of a proximal-like minimization algorithm using Bregman functions.SIAM Journal on Optimization, 3(3):538–543, 1993

    1993

  17. [25]

    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

    2026 arXiv

  18. [26]

    Spurious stationarity and hardness results for Bregman proximal-type algorithms.arXiv preprint arXiv:2404.08073, version 3, 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, version 3, 2026

    2026 arXiv

  19. [27]

    On the linear convergence of Bregman proximal gradient methods with applications to Kullback–Leibler regression.arXiv preprint arXiv:2607.05539, 2026

    Jonathan Chirinos-Rodríguez, Christian Daniele, Cédric Févotte, and Emmanuel Soubies. On the linear convergence of Bregman proximal gradient methods with applications to Kullback–Leibler regression.arXiv preprint arXiv:2607.05539, 2026. 51

    2026 arXiv

  20. [28]

    Norm preserving extension of convex Lipschitz functions

    S Cobzas and C Mustata. Norm preserving extension of convex Lipschitz functions. Journal of Approxixation Theory, 24(3):236–244, 1978

    1978

  21. [29]

    Institut de Recherche Mathé- matiques de Rennes, 1999

    Michel Coste.An Introduction to O-minimal Geometry. Institut de Recherche Mathé- matiques de Rennes, 1999

    1999

  22. [30]

    Ingrid Daubechies, Michel Defrise, and Christine De Mol. An iterative thresholding algorithm for linear inverse problems with a sparsity constraint.Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 57(11):1413–1...

    2004

  23. [31]

    Stochastic subgra- dient method converges on tame functions.Foundations of Computational Mathematics, 20(1):119–154, 2020

    Damek Davis, Dmitriy Drusvyatskiy, Sham Kakade, and Jason D Lee. Stochastic subgra- dient method converges on tame functions.Foundations of Computational Mathematics, 20(1):119–154, 2020

    2020

  24. [32]

    A gen- eralized approach to portfolio optimization: Improving performance by constraining portfolio norms.Management Science, 55(5):798–812, 2009

    Victor DeMiguel, Lorenzo Garlappi, Francisco J Nogales, and Raman Uppal. A gen- eralized approach to portfolio optimization: Improving performance by constraining portfolio norms.Management Science, 55(5):798–812, 2009

    2009

  25. [33]

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

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

    2025 arXiv

  26. [34]

    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

    2025

  27. [35]

    Non-KKT accumulation in entropic mirror descent

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

    2026 arXiv

  28. [36]

    Compressed sensing.IEEE Transactions on Information Theory, 52 (4):1289–1306, 2006

    David L Donoho. Compressed sensing.IEEE Transactions on Information Theory, 52 (4):1289–1306, 2006

    2006

  29. [37]

    Chapman and Hall/CRC, 2025

    Lawrence C Evans.Measure Theory and Fine Properties of Functions. Chapman and Hall/CRC, 2025

    2025

  30. [38]

    Springer, New York, 2003

    Francisco Facchinei and Jong-Shi Pang.Finite-Dimensional Variational Inequalities and Complementarity Problems. Springer, New York, 2003

    2003

  31. [39]

    Central paths, generalized proximal point methods, and cauchy trajectories in Riemannian manifolds.SIAM Journal on Control and Optimization, 37(2):566–588, 1999

    Alfredo N Iusem, BF Svaiter, and João Xavier da Cruz Neto. Central paths, generalized proximal point methods, and cauchy trajectories in Riemannian manifolds.SIAM Journal on Control and Optimization, 37(2):566–588, 1999

    1999

  32. [40]

    Non-Convex Optimization for Machine Learning

    Prateek Jain and Purushottam Kar. Non-Convex Optimization for Machine Learning. Foundations and Trends®in Machine Learning, 10(3–4):142–336, 2017

    2017

  33. [41]

    Accelerated mirror descent in continuous and discrete time.Advances in Neural Information Processing Systems, 28, 2015

    Walid Krichene, Alexandre Bayen, and Peter L Bartlett. Accelerated mirror descent in continuous and discrete time.Advances in Neural Information Processing Systems, 28, 2015. 52

    2015

  34. [42]

    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. InProceedings of the 11th International Conference on Learning Repre- sentations (ICLR 2023), 2023

    2023

  35. [43]

    Implicit bias of gradient descent on reparametrized models: On equivalence to mirror descent.Advances in Neural Information Processing Systems, 35:34626–34640, 2022

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

    2022

  36. [44]

    Relatively smooth convex optimization by first-order methods, and applications.SIAM Journal on Optimization, 28(1):333–354, 2018

    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

    2018

  37. [45]

    Pearson, 1999

    Arthur Mattuck.Introduction to Analysis. Pearson, 1999

    1999

  38. [46]

    Problem complexity and method efficiency in optimization

    Arkadij Semenovic Nemirovskij and David Borisovich Yudin. Problem complexity and method efficiency in optimization. 1983

    1983

  39. [47]

    Feature selection,L1 vs

    Andrew Y Ng. Feature selection,L1 vs. L2 regularization, and rotational invariance. In Proceedings of the 21st International Conference on Machine Learning, page 78, 2004

    2004

  40. [48]

    On preparation theorems forran,exp-definable functions.Journal of Logic and Analysis, 15, 2023

    Andre Opris. On preparation theorems forran,exp-definable functions.Journal of Logic and Analysis, 15, 2023

    2023

  41. [49]

    On the sequential convergence of Lloyd’s algorithms.Mathematics of Operations Research, 2025

    Léo Portales, Elsa Cazelles, and Edouard Pauwels. On the sequential convergence of Lloyd’s algorithms.Mathematics of Operations Research, 2025

    2025

  42. [50]

    Springer Science & Business Media, Berlin, Heidelberg, 2009

    R Tyrrell Rockafellar and Roger J-B Wets.Variational Analysis, volume 317. Springer Science & Business Media, Berlin, Heidelberg, 2009

    2009

  43. [51]

    Princeton University Press, 2015

    Ralph Tyrell Rockafellar.Convex Analysis. Princeton University Press, 2015

    2015

  44. [52]

    Generalized self-concordant functions: A recipe for newton-type methods.Mathematical Programming, 178:145 – 213, 2019

    Tianxiao Sun and Quoc Tran-Dinh. Generalized self-concordant functions: A recipe for newton-type methods.Mathematical Programming, 178:145 – 213, 2019

    2019

  45. [53]

    Approximate Bregman proximal gradient algo- rithm for relatively smooth nonconvex optimization.Computational Optimization and Applications, 90(1):227–256, 2025

    Shota Takahashi and Akiko Takeda. Approximate Bregman proximal gradient algo- rithm for relatively smooth nonconvex optimization.Computational Optimization and Applications, 90(1):227–256, 2025

    2025

  46. [54]

    New Bregman proximal type algorithms for solving DC optimization problems.Computational Optimization and Applications, 83(3):893–931, 2022

    Shota Takahashi, Mituhiro Fukuda, and Mirai Tanaka. New Bregman proximal type algorithms for solving DC optimization problems.Computational Optimization and Applications, 83(3):893–931, 2022

    2022

  47. [55]

    A simplified view of first order methods for optimization.Mathematical Programming, Series B, 170(1):67–96, 2018

    Marc Teboulle. A simplified view of first order methods for optimization.Mathematical Programming, Series B, 170(1):67–96, 2018

    2018

  48. [56]

    A generalization of the Tarski-Seidenberg theorem, and some nondefinability results.American Mathematical Society, 15(2), 1986

    Lou Van den Dries. A generalization of the Tarski-Seidenberg theorem, and some nondefinability results.American Mathematical Society, 15(2), 1986. 53

    1986

  49. [57]

    Cambridge University Press, 1998

    Lou Van den Dries.Tame Topology and O-minimal Structures, volume 248. Cambridge University Press, 1998

    1998

  50. [58]

    Geometric categories and o-minimal structures

    Lou Van Den Dries and Chris Miller. Geometric categories and o-minimal structures. Duke Mathematical Journal, 84(2):497–540, 1996

    1996

  51. [59]

    Linear convergence of a proximal alternating minimization method with extrapolation forℓ1-norm principal component analysis.SIAM Journal on Optimization, 33(2):684–712, 2023

    Peng Wang, Huikang Liu, and Anthony Man-Cho So. Linear convergence of a proximal alternating minimization method with extrapolation forℓ1-norm principal component analysis.SIAM Journal on Optimization, 33(2):684–712, 2023

    2023

  52. [60]

    Inertial proximal gradient methods with Bregman regularization for a class of nonconvex optimization problems

    Zhongming Wu, Chongshou Li, Min Li, and Andrew Lim. Inertial proximal gradient methods with Bregman regularization for a class of nonconvex optimization problems. Journal of Global Optimization, 79:617–644, 2021

    2021

  53. [61]

    Bregman proximal point algorithm revisited: A new inexact version and its inertial variant.SIAM Journal on Optimization, 32(3):1523–1554, 2022

    Lei Yang and Kim-Chuan Toh. Bregman proximal point algorithm revisited: A new inexact version and its inertial variant.SIAM Journal on Optimization, 32(3):1523–1554, 2022

    2022

  54. [62]

    Inexact Bregman proximal gradient method and its inertial variant with absolute and partial relative stopping criteria.Mathematics of Operations Research, 2025

    Lei Yang and Kim-Chuan Toh. Inexact Bregman proximal gradient method and its inertial variant with absolute and partial relative stopping criteria.Mathematics of Operations Research, 2025

    2025

  55. [63]

    Proximal-like incremental aggregated gradient method with linear convergence under Bregman distance growth conditions

    Hui Zhang, Yu-Hong Dai, Lei Guo, and Wei Peng. Proximal-like incremental aggregated gradient method with linear convergence under Bregman distance growth conditions. Mathematics of Operations Research, 46(1):61–81, 2021

    2021

  56. [64]

    Level-set subdifferential error bounds and linear convergence of Bregman proximal gradient method.Journal of Optimization Theory and Applications, 189(3):889–918, 2021

    Daoli Zhu, Sien Deng, Minghua Li, and Lei Zhao. Level-set subdifferential error bounds and linear convergence of Bregman proximal gradient method.Journal of Optimization Theory and Applications, 189(3):889–918, 2021. 54

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.