REVIEW 5 major objections 5 minor 2 cited by
On exploration of an interior mirror descent flow for stochastic nonconvex constrained problem
T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read One continuous flow explains why mirror descent and Hessian barrier methods can stop at spurious stationary points, and provides conditions that rule those stops out.
desk verdict The continuous-flow unification and stable-set reading of spurious points are solid and useful, but the paper's own headline application to constrained mirror descent rests on an unverified condition, so the discrete claims are not yet established. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the interior Riemannian subgradient flow (7), with $\nabla^2\varphi$ inducing a Riemannian metric on $C$ and $P_{T_xM}$ the metric projection onto the tangent space of $M$. Its equivalent dual form (8), $\frac{d}{dt}\nabla\varphi(x(t)) \in -\partial f(x(t)) + N_{x(t)}M$, is what makes the mirror descent connection explicit: discretizing (8) with a Bregman step gives the mirror descent update, while discretizing (7) with a retraction gives the Hessian barrier method. The load-bearing distinction is between the flow's stable set $S$ and the optimization problem's stationary set $\Omega$; the argument uses $f$ itself as a Lyapunov function, a separating-hyperplane construction on the active constraints to produce a repelling direction at points of $S \setminus \Omega$, and stochastic approximation theory to transfer trajectory statements to discrete iterates.
What would settle it
Run the flow (7) for a bounded $C^2$ function $f$ with affine constraints $Ax = b$ and the entropy kernel, starting near a spurious point $x^* \in S \setminus \Omega$ that satisfies $s(x^*) + x^* \ge 0$; if the trajectory accumulates at $x^*$, or a mirror descent run with step size $\eta_k = 1/k$ accumulates there, then Theorem 4.1(ii) and Proposition 3.7 would be contradicted. A concrete experimental signature would be $x_i(t)$ failing to grow exponentially while the trajectory remains indefinitely in a neighborhood of $x^*$, contrary to the Gronwall-based escape bound.
Extended reading notes
Core claim
On its own terms, the paper establishes that the differential inclusion (7), equivalently written as $\frac{d}{dt}\nabla\varphi(x(t)) \in -\partial f(x(t)) + N_{x(t)}M$, is the continuous object underlying both the Hessian barrier method and mirror descent: the former is a retraction-based discretization of the projected Riemannian subgradient, and the latter is a Bregman discretization of the dual form of the same flow. The stable set $S = \{x : 0 \in P_{T_xM}\nabla^2\varphi(x)^{-1}\partial f(x)\}$ can be strictly larger than the true stationary set $\Omega = \{x : 0 \in \partial f(x) + N_M(x) + N_C(x)\}$; the difference $S \setminus \Omega$ is precisely where the spurious stationary points of mirror descent live, and the paper identifies them with fixed points of the extended Bregman update map. It shows that every trajectory of the flow exits a neighborhood of any point in $S \setminus \Omega$ in finite time, so a convergent trajectory must land in $\Omega$. If the complementarity condition $s(x) + x \ge 0$ holds on the $\omega$-limit set, or if $S \cap \partial C$ has only isolated points, then the escape dynamics preclude subsequential convergence to spurious points. Absent such regularity, generically perturbing the objective to $f + \langle \nabla\varphi, v\rangle$ and the constraint to $c + u$ makes the stable set have no cluster points by a Morse--Sard argument, so the flow and its discrete approximations converge to stationary points of the perturbed problem, which are nearly stationary for the original problem.
Load-bearing premise
The load-bearing premise is that each discrete method tracks the continuous flow accurately: the retraction error in Assumption 5.1.6, and the analogous condition (18) for mirror descent, must vanish uniformly as the step size goes to zero, and the paper itself notes that no retraction can satisfy this in a one-dimensional interval example and that verifying (18) beyond an entropy kernel is difficult.
Editorial extensions
If this is right
- Any trajectory of (7) that converges must converge to a true stationary point in $\Omega$, because spurious points in $S \setminus \Omega$ are escaped in finite time.
- When the complementarity condition $s(x) + x \ge 0$ holds on the $\omega$-limit set, or when $S \cap \partial C$ consists of isolated points, the flow and the associated discrete interior-point methods subsequentially converge to $\Omega$ instead of spurious points.
- Without those conditions, perturbing $f$ to $f + \langle \nabla\varphi, v\rangle$ and $c$ to $c + u$ makes the stable set isolated for almost all small $(u,v)$, so the trajectory reaches a stationary point of the perturbed problem, which is approximately stationary for the original one.
- The proposed discretizations are interior point methods whose iterates remain in $M \cap C$, and their convergence inherits the continuous flow's conclusions whenever the retraction error scaled by step size vanishes uniformly.
- The spurious stationary points of mirror descent correspond exactly to fixed points of the extended Bregman update map, so the paper transfers that phenomenon from a property of one algorithm to a property of the unifying flow.
Reading between the lines
- This stable-set criterion suggests a screening test for any new Bregman or barrier scheme: compute $S \setminus \Omega$ for the kernel and constraint geometry; if it is nonempty, spurious limits or slow escape should be expected unless a retraction condition like Assumption 5.1.6 or condition (18) holds.
- The escape mechanism is quantitative: the Gronwall argument makes a coordinate grow exponentially while the trajectory remains near a spurious point, so Proposition 3.7 could be turned into a numerical diagnostic that measures exponential repulsion to certify whether a near-stationary iterate is spurious.
- If the paper's transfer is correct, the honest failure mode of mirror descent is geometric rather than due to step sizes: any kernel whose Hessian inverse degenerates at the boundary enlarges the stable set, and comparing kernels by the size of $S \setminus \Omega$ could become a principled kernel-design criterion.
- The perturbation strategy is effectively a random smoothing of the manifold, and for structured cones such as the positive semidefinite cone it preserves membership of the slack variable in the dual cone, hinting at practical regularized interior-point solvers for conic programs.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies a nonsmooth, nonconvex optimization problem over the intersection of the closure of an open convex set C and a smooth manifold M. It introduces the differential inclusion (7), ẋ(t) ∈ -P_{T_{x(t)}M} ∇²φ(x(t))^{-1} ∂f(x(t)), and interprets this flow as a unified continuous-time model for both the Hessian-barrier method and mirror descent. The central claim is that the 'spurious stationary points' observed in both algorithms are exactly the points in the stable set S = {x : 0 ∈ P_{T_xM}∇²φ(x)^{-1}∂f(x)} that are not true stationary points Ω = {x : 0 ∈ ∂f(x) + N_M(x) + N_C(x)}. The paper develops finite-time escape results near S \ Ω, provides two conditions (complementarity and isolation) under which trajectories avoid S \ Ω, proposes a random perturbation strategy, and derives two discrete interior-point algorithms whose convergence is analyzed through stochastic approximation. The authors are explicit that several steps are heuristic or incomplete, including the verification of condition (18) for constrained mirror descent and the nonlinear extension of Theorem 4.1.
Significance. If the main claims were fully established, the paper would supply a valuable unifying explanation for a known pathological behavior: the spurious-stationary-point phenomenon of mirror descent and Hessian-barrier methods would be traced to the stable set of a single continuous flow. The conceptual link between the extended mirror-descent mapping of [19] and the stable set S is attractive, and the paper is commendably honest about its limitations. However, the load-bearing bridge from the continuous flow to constrained mirror descent is verified only in the unconstrained case, and several central proofs—most notably Theorem 4.1(i), Proposition 3.4, and the nonlinear extension in §4.1—contain gaps or are self-flagged as problematic. The significance is therefore conditional: the paper provides a promising framework and several correct-looking continuous-flow arguments, but not yet a fully supported explanation of the constrained mirror-descent phenomenon that motivates the title and abstract.
major comments (5)
- [§5.2, Eq. (18), Example 5.1] The paper's explanation of the spurious-stationary-point phenomenon in constrained mirror descent rests entirely on condition (18), which would make the interpolated mirror-descent process a perturbed solution of the differential inclusion (7). The authors state that verifying (18) 'is generally difficult' and verify it only in the unconstrained case M = Rⁿ, where the Aᵀy term is absent. The constrained case is exactly the regime of the phenomenon being explained: near the boundary of C the preconditioner ∇²φ(x)^{-1} degenerates for typical barriers such as entropy, and the dual variable y couples the escape direction to the equality constraint. Without (18), the stochastic-approximation theorem of Benaim et al. is not applicable, and none of the escape, avoidance, or stationarity conclusions of Sections 3 and 4 are known to hold for the mirror-descent iterates. The claimed unification with the constrained mirror-descent spurious-point results of [19] is therefore unsupported as written.
- [§4.1, Theorem 4.1(i)] The proof of Theorem 4.1(i) asserts that when s(x) + x < 0 at a spurious point, there is a neighborhood in which ẋ > δ > 0. For the entropy kernel, H(x)^{-1} = diag(x) and the flow components satisfy ẋ_i = -x_i s_i(x), so as x_i → 0 the velocity tends to 0 even when s_i(x) ≤ -δ. The uniform linear lower bound is false. The proof's case analysis also does not exclude the possibility that the trajectory exits a neighborhood and later re-enters it; the text says 'we can argue in case 1' but case 1 does not establish that re-entry is impossible. The stronger 'never subsequentially converge' claim therefore does not follow from the given argument. A Gronwall-type log-coordinate argument as in Proposition 3.7 may repair the claim, but as written the proof is invalid.
- [§4.1, nonlinear extension] The paragraph extending Theorem 4.1 to nonlinear constraints C = {x : g(x) ≤ 0} contains the sentence 'the complementarity condition s(x) + x ≥ 0 ... has some issues', followed by an assertion that the repelling mechanism works 'analogous to the linear case' without a proof. Since the theorem statement explicitly claims the nonlinear extension, the text itself flags a missing proof. This is not a minor presentation issue: the nonlinear constraint setting is one of the stated contributions of the paper. The extension must either be proved in full or the theorem must be restricted to the linear case and the abstract and introduction adjusted accordingly.
- [§3.1, Proposition 3.4] The proof of the inclusion NC(x) ⊂ Null((∇²φ(x))^{-1}) is not valid as written. The proof defines a set S of limits ∇²φ(x_k)e_k → d with e_k → 0 and shows that v ∈ span(∂∞φ(x)) can be represented as such a limit, but it never derives that H(x)^{-1}v = 0; the set S is not identified with the nullspace of H(x)^{-1}. In addition, the equality span(∂∞φ(x)) = NC(x) is asserted rather than proved for a general open convex C. Proposition 3.4 underlies the stable-set characterization used in Proposition 3.5, Proposition 3.6, and Theorem 4.1, so this gap is load-bearing rather than cosmetic.
- [§4.2, Lemma 4.1 and Theorem 4.3] The random-perturbation argument is not fully justified. The proof of Lemma 4.1 applies Morse-Sard theory to the map F(x, λ), but the required differentiability is not established: φ is only assumed C², so ∇²φ(x)^{-1}∇f(x) need not be C¹, and C¹ Morse-Sard is false in general for maps Rᵈ → Rᵈ. Even if local uniqueness of solutions follows from the implicit function theorem, that only shows the stable set S(u,v) is discrete; it does not by itself rule out accumulation of isolated points within a bounded region, and no boundedness of S(u,v) is shown. Since the perturbation strategy is the paper's answer for the generic case, this argument needs a careful repair or a precise set of additional hypotheses.
minor comments (5)
- [Abstract and throughout] There are numerous typos and grammatical errors, including 'unifily', 'stationsrt point', 'supercoecive', 'Gronwalls inequality', and 'separate theorem' for 'separation theorem'. The abstract also says 'strict complementarity conditions' while the actual sufficient condition in Theorem 4.1 is the non-strict inequality s(x) + x ≥ 0 on the ω-limit set.
- [§2.3, Proposition 2.1 proof] The proof of Proposition 2.1 contains a sign error: the KKT condition '0 = H(z)^{-1}(d(z) + Aᵀy)' should presumably be '0 = H(z)^{-1}(d(z) - Aᵀy)', and the expansion of the projection term omits a term. This is fixable but should be corrected for clarity.
- [§3.1, Proposition 3.3] The equivalence between (7) and (8) is central to the mirror-descent interpretation, but its proof is omitted with 'similar to the proof of Proposition 2.1'. Given that Proposition 2.1 itself has sign issues, the authors should provide a complete proof of Proposition 3.3 or a precise reference.
- [§5.1, Eq. (16)] Equation (16) has a typo: the displayed chain ends with 'xk = −ηk(vk + ξ̃k)' but should be 'xk+1 = xk − ηk(vk + ξ̃k)'. This makes the algebra in that paragraph confusing.
- [§5.2, heuristic algorithm] The proposed 'heuristic Riemannian mirror descent scheme' (R-Breg) is presented with only a brief discussion and is not analyzed beyond the unverified condition (18). If the paper keeps this algorithm, it should either be clearly labeled as a proposal open for future work or provided with some form of justification; otherwise the reader may expect a convergence theorem that is not supplied.
Circularity Check
No significant circularity: the central stable-set characterization is proved against independent external objects, and the unverified condition (18) is a gap, not a circular step.
full rationale
No significant circularity. The central claim — that spurious stationary points of the Hessian barrier method and mirror descent scheme correspond to the stable set S = {x : 0 in P_{T_xM} grad^2 phi(x)^(-1) partial f(x)} \ Omega — is established by theorems checked against independent external objects: the extended mirror-descent mapping T^eta of [19] in Proposition 3.6, and the Hessian-Riemannian flow of [1] in Section 2.3. The equivalence between the differential inclusion (7) and the Bregman formulation (8) is proved directly in Proposition 3.2, and the stable-set characterization in Proposition 3.4 rests on stated Legendre/barrier hypotheses rather than on the conclusions it is used to explain. The only notable self-citation, [25] in Section 5.2, asserts that unconstrained Hessian barrier and mirror descent share the same differential inclusion; the paper immediately re-derives the connection via Taylor expansion, so the citation is not load-bearing. The paper itself flags the main limitation in Section 5.2: condition (18), which transfers all flow conclusions to constrained mirror descent, is said to be 'generally difficult' to verify, and Example 5.1 verifies it only in the unconstrained case M = R^n. This is a substantive correctness gap — the discrete-to-continuous bridge for constrained mirror descent is an unverified hypothesis, not an input that forces the claimed conclusion by definition. Likewise, the phrase 'has some issues' in the nonlinear extension of Theorem 4.1 signals an omitted or incomplete argument, but again this is a missing proof, not a circular reduction. The paper is self-contained against external benchmarks and does not fit parameters to data, rename known consequences as predictions, or import a uniqueness theorem from the authors' own prior work.
Assumptions & free parameters
assumptions (9)
- domain assumption LICQ holds at every point of the manifold M = {x : c(x)=0}.
- domain assumption The Riemannian projection operator P_x ∇²φ(x)^{-1} extends continuously to the closure M ∩ C.
- domain assumption φ is a C² Legendre function with locally Lipschitz ∇²φ, and f is path-differentiable; the boundedness condition of Assumption 3.1 holds.
- domain assumption Separable kernel structure of Assumption 3.2: each φ_i is a Legendre function over R_+ with φ_i''(x_i)=x_i^{-γ} ψ_i(x_i).
- ad hoc to paper For Proposition 3.4, sup_{x∈C∩ρB} ‖∇²φ(x)^{-1}∇φ(x)‖ < ∞ and span(∂∞φ(x)) ⊃ Null((∇²φ(x))^{-1}).
- domain assumption Stochastic approximation conditions of Assumptions 2.1 and 5.1: bounded iterates, stepsizes with sum ∞ and η_k=o(1/log k), martingale difference noise, closed graph, and the asymptotic averaging condition.
- ad hoc to paper Assumption 5.1.6: uniform vanishing of the retraction error η^{-1} ‖R_x(-η(...)) - (x - η(...))‖ → 0.
- domain assumption Assumption 5.2: f is lower bounded and the set of critical values has empty interior (weak Sard condition).
- standard math Morse-Sard theorem and implicit function theorem are applied in Lemma 4.1 and Theorem 4.3 without explicit smoothness and compactness hypotheses.
Cite this review
Pith. "Pith review of On exploration of an interior mirror descent flow for stochastic nonconvex constrained problem." pith.science (2026). https://pith.science/paper/LXFTPTHI
@misc{pith2026250715264,
author = {Pith},
title = {Pith review of: On exploration of an interior mirror descent flow for stochastic nonconvex constrained problem},
year = {2026},
howpublished = {\url{https://pith.science/paper/LXFTPTHI}},
note = {Machine review of arXiv:2507.15264}
}
read the original abstract
We study a nonsmooth nonconvex optimization problem defined over nonconvex constraints, where the feasible set is given by the intersection of the closure of an open set and a smooth manifold. By endowing the open set with a Riemannian metric induced by a barrier function, we obtain a Riemannian subgradient flow formulated as a differential inclusion, which remains strictly within the interior of the feasible set. This continuous dynamical system unifies two classes of iterative optimization methods, namely the Hessian barrier method and mirror descent scheme, by revealing that these methods can be interpreted as discrete approximations of the continuous flow. We explore the long-term behavior of the trajectories generated by this dynamical system and show that the existing deficient convergence properties of the Hessian barrier and mirror descent scheme can be unifily and more insightfully interpreted through these of the continuous trajectory. For instance, the notorious spurious stationary points \cite{chen2024spurious} observed in Hessian barrier method and mirror descent scheme are interpreted as stable equilibria of the dynamical system that do not correspond to real stationary points of the original optimization problem. We provide two sufficient condition such that these spurious stationary points can be avoided if the strict complementarity conditions holds. In the absence of these regularity condition, we propose a random perturbation strategy that ensures the trajectory converges (subsequentially) to an approximate stationary point. Building on these insights, we introduce two iterative Riemannian subgradient methods, form of interior point methods, that generalizes the existing Hessian barrier method and mirror descent scheme for solving nonsmooth nonconvex optimization problems.
Figures
Forward citations
Cited by 2 Pith papers
-
A Unified Framework for Iterate Convergence of Bregman Proximal Methods
A unified framework using scaled Kurdyka-Lojasiewicz inequalities shows that Bregman proximal point and gradient methods, and mirror flow, converge for closed-domain separable kernels and subanalytic or definable objectives.
-
A Support-Set Algorithm for Optimization Problems with Nonnegative and Orthogonal Constraints
A support-set algorithm for nonnegative orthogonal optimization provably converges to first-order stationary points in O(epsilon^-2) iterations, with closed-form subproblem solutions.
Reference graph
Works this paper leans on
-
[19]
H. Chen, J. Li, and A. M.-C. So. Spurious stationarity and hardness results for mirror descent. arXiv preprint arXiv:2404.08073 , 2024
arXiv 2024
-
[1]
F. Alvarez, J. Bolte, and O. Brahic. Hessian Riemannian gradient flows in convex program- ming. SIAM J. Control and Optimization , 43(2):477–501, 2004
work page 2004
-
[2]
H. Attouch, J. Bolte, P. Redont, and M. Teboulle. Singular riemannian barrier methods and gradient-projection dynamical systems for constrained optimization. Optimization, 53(5- 6):435–454, 2004
work page 2004
-
[3]
J.-P. Aubin and A. Cellina. Differential inclusions: set-valued maps and viability theory , volume 264. Springer Science & Business Media, 2012
work page 2012
-
[4]
H. H. Bauschke, J. Bolte, and M. Teboulle. A descent lemma beyond Lipschitz gradient con- tinuity: first-order methods revisited and applications. Mathematics of Operations Research , 42(2):330–348, 2017
work page 2017
-
[5]
M. Bena ¨ ım, J. Hofbauer, and S. Sorin. Stochastic approximations and differential inclusions. SIAM J. Control and Optimization , 44(1):328–348, 2005
work page 2005
-
[6]
M. Bena ¨ ım, J. Hofbauer, and S. Sorin. Stochastic approximations and differential inclusions, part ii: Applications. Mathematics of Operations Research, 31(4):673–695, 2006
work page 2006
-
[7]
J.-D. Benamou, G. Carlier, M. Cuturi, L. Nenna, and G. Peyr´ e. Iterative Bregman projections for regularized transportation problems. SIAM J. Scientific Computing , 37(2):A1111–A1138, 2015
work page 2015
Show all 69 references
-
[8]
D. P. Bertsekas. Nonlinear programming. J. of the Operational Research Society , 48(3):334– 334, 1997
1997
-
[9]
Bolte, A
J. Bolte, A. Daniilidis, A. Lewis, and M. Shiota. Clarke subgradients of stratifiable functions. SIAM J. Optimization , 18(2):556–572, 2007
2007
-
[10]
Bolte and E
J. Bolte and E. Pauwels. Conservative set valued fields, automatic differentiation, stochastic gradient methods and deep learning. Mathematical Programming, 188:19–51, 2021
2021
-
[11]
Bolte, S
J. Bolte, S. Sabach, M. Teboulle, and Y. Vaisbourd. First order methods beyond convexity and lipschitz gradient continuity with applications to quadratic inverse problems. SIAM J. Optimization, 28(3):2131–2151, 2018
2018
-
[12]
Bolte and M
J. Bolte and M. Teboulle. Barrier operators and associated gradient-like dynamical systems for constrained minimization problems. SIAM J. Control and Optimization , 42(4):1266–1292, 2003
2003
-
[13]
I. M. Bomze, P. Mertikopoulos, W. Schachinger, and M. Staudigl. Hessian barrier algorithms for linearly constrained optimization problems. SIAM Journal on Optimization , 29(3):2100– 2127, 2019
2019
-
[14]
V. S. Borkar. Stochastic approximation: a dynamical systems viewpoint , volume 48. Springer, 2009. 30
2009
-
[15]
L. 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
-
[16]
R. H. Byrd, M. E. Hribar, and J. Nocedal. An interior point algorithm for large-scale nonlinear programming. SIAM Journal on Optimization , 9(4):877–900, 1999
1999
-
[17]
Cambier and P.-A
L. Cambier and P.-A. Absil. Robust low-rank matrix completion by riemannian optimization. SIAM Journal on Scientific Computing , 38(5):S440–S460, 2016
2016
-
[18]
Castera, J
C. Castera, J. Bolte, C. F´ evotte, and E. Pauwels. An inertial newton algorithm for deep learning. J. of Machine Learning Research , 22(1):5977–6007, 2021
2021
-
[20]
H. T. Chu, L. Liang, K.-C. Toh, and L. Yang. An efficient implementable inexact entropic prox- imal point algorithm for a class of linear programming problems. Computational Optimization and Applications, 85(1):107–146, 2023
2023
-
[21]
F. H. Clarke. Optimization and nonsmooth analysis . SIAM, 1990
1990
-
[22]
Davis, D
D. Davis, D. Drusvyatskiy, and L. Jiang. Subgradient methods near active manifolds: saddle point avoidance, local convergence, and asymptotic normality. arXiv preprint arXiv:2108.11832, page 170, 2021
2021 arXiv
-
[23]
Davis, D
D. Davis, D. Drusvyatskiy, S. Kakade, and J. D. Lee. Stochastic subgradient method converges on tame functions. Foundations of Computational Mathematics , 20(1):119–154, 2020
2020
-
[24]
K. Ding, J. Li, and K.-C. Toh. Nonconvex Stochastic Bregman Proximal Gradient Method with Application to Deep Learning. arXiv preprint arXiv:2306.14522 , 2023
2023 arXiv
-
[25]
Ding and K.-C
K. Ding and K.-C. Toh. Stochastic bregman subgradient methods for nonsmooth nonconvex optimization problems. arXiv preprint arXiv:2404.17386 , 2024
2024 arXiv
-
[26]
K. Ding, N. Xiao, and K.-C. Toh. Adam-family methods with decoupled weight decay in deep learning. arXiv preprint arXiv:2310.08858 , 2023
2023 arXiv
-
[27]
Ding and S
L. Ding and S. J. Wright. On squared-variable formulations. arXiv preprint arXiv:2310.01784, 2023
2023 arXiv
-
[28]
Dragomir, A
R.-A. Dragomir, A. dAspremont, and J. Bolte. Quartic first-order methods for low-rank minimization. J. of Optimization Theory and Applications , 189:341–363, 2021
2021
-
[29]
R. A. Dragomir, M. Even, and H. Hendrikx. Fast stochastic Bregman gradient methods: Sharp analysis and variance reduction. In International Conference on Machine Learning , pages 2815–2825. PMLR, 2021
2021
-
[30]
Dragomir, A
R.-A. Dragomir, A. B. Taylor, A. dAspremont, and J. Bolte. Optimal complexity and certifi- cation of Bregman first-order methods. Mathematical Programming, pages 1–43, 2021. 31
2021
-
[31]
J. C. Duchi and F. Ruan. Stochastic methods for composite and weakly convex optimization problems. SIAM J. Optimization , 28(4):3229–3259, 2018
2018
-
[32]
Dvurechensky and M
P. Dvurechensky and M. Staudigl. Hessian barrier algorithms for non-convex conic optimiza- tion. Mathematical Programming, 209(1):171–229, 2025
2025
-
[33]
A. V. Fiacco and G. P. McCormick. Nonlinear Programming: Sequential Unconstrained Min- imization Techniques. Wiley, 1968
1968
-
[34]
Frank and P
M. Frank and P. Wolfe. An algorithm for quadratic programming. Naval Research Logistics Quarterly, 3(1-2):95–110, 1956
1956
-
[35]
Hsieh, M
Y.-P. Hsieh, M. R. Karimi Jaghargh, A. Krause, and P. Mertikopoulos. Riemannian stochastic optimization methods avoid strict saddle points. Advances in Neural Information Processing Systems, 36:29580–29601, 2023
2023
-
[36]
M. Jaggi. Revisiting frank-wolfe: Projection-free sparse convex optimization. In ICML, pages 427–435, 2013
2013
-
[37]
Jiang, X
B. Jiang, X. Meng, Z. Wen, and X. Chen. An exact penalty approach for optimization with nonnegative orthogonality constraints. Mathematical Programming, 198(1):855–897, 2023
2023
-
[38]
Josz and L
C. Josz and L. Lai. Global stability of first-order methods for coercive tame functions. Math- ematical Programming, 207(1):551–576, 2024
2024
-
[39]
Karmarkar
N. Karmarkar. A new polynomial-time algorithm for linear programming. Combinatorica, 4(4):373–395, 1984
1984
-
[40]
Y. Khoo, T. Tang, and K.-C. Toh. A bregman admm for bethe variational problem. arXiv preprint arXiv:2502.04613, 2025
2025
-
[41]
C. Kolb, C. L. M¨ uller, B. Bischl, and D. R¨ ugamer. Smoothing the edges: A general framework for smooth optimization in sparse regularization using hadamard overparametrization. CoRR, 2023
2023
-
[42]
T. Le. Nonsmooth nonconvex stochastic heavy ball. arXiv preprint arXiv:2304.13328 , 2023
2023 arXiv
-
[43]
J. D. Lee, I. Panageas, G. Piliouras, M. Simchowitz, M. I. Jordan, and B. Recht. First-order methods almost always avoid strict saddle points. Mathematical programming, 176:311–337, 2019
2019
-
[44]
J. D. Lee, M. Simchowitz, M. I. Jordan, and B. Recht. Gradient descent only converges to minimizers. In Conference on learning theory , pages 1246–1257. PMLR, 2016
2016
-
[45]
Levin, J
E. Levin, J. Kileel, and N. Boumal. The effect of smooth parametrizations on nonconvex optimization landscapes. Mathematical Programming, 209(1):63–111, 2025
2025
-
[46]
Z. Li, T. Wang, J. D. Lee, and S. Arora. Implicit bias of gradient descent on reparametrized models: On equivalence to mirror descent. Advances in Neural Information Processing Systems, 35:34626–34640, 2022. 32
2022
-
[47]
H. Lu, R. M. Freund, and Y. Nesterov. Relatively smooth convex optimization by first-order methods, and applications. SIAM J. Optimization , 28(1):333–354, 2018
2018
-
[48]
M. C. Mukkamala, F. Westerkamp, E. Laude, D. Cremers, and P. Ochs. Bregman proximal framework for deep linear neural networks. arXiv preprint arXiv:1910.03638 , 2019
1910 arXiv
-
[49]
A. S. Nemirovskij and D. B. Yudin. Problem complexity and method efficiency in optimization. Wiley-Interscience, 1983
1983
-
[50]
Nesterov et al
Y. Nesterov et al. Lectures on convex optimization , volume 137. Springer, 2018
2018
-
[51]
Nesterov and A
Y. Nesterov and A. Nemirovskii. Interior-Point Polynomial Algorithms in Convex Program- ming. SIAM, 1994
1994
-
[52]
Panageas and G
I. Panageas and G. Piliouras. Gradient descent only converges to minimizers: Non-isolated critical points and invariant regions. arXiv preprint arXiv:1605.00405 , 2016
2016 arXiv
-
[53]
Panageas, G
I. Panageas, G. Piliouras, and X. Wang. First-order methods almost always avoid saddle points: The case of vanishing step-sizes. Advances in Neural Information Processing Systems , 32, 2019
2019
-
[54]
M. J. D. Powell. A method for nonlinear constraints in minimization problems. In R. Fletcher, editor, Optimization, pages 283–298. Academic Press, 1969
1969
-
[55]
R. T. Rockafellar. A dual approach to solving nonlinear programming problems by uncon- strained optimization. Math. Programming, 5:354–373, 1973
1973
-
[56]
R. T. Rockafellar. Convex analysis , volume 11. Princeton University Press, 1997
1997
-
[57]
M. Shub. Global stability of dynamical systems . Springer Science & Business Media, 2013
2013
-
[58]
Tang and K.-C
T. Tang and K.-C. Toh. Optimization over convex polyhedra via hadamard parametrizations. Mathematical Programming, pages 1–41, 2024
2024
-
[59]
Tseng, I
P. Tseng, I. M. Bomze, and W. Schachinger. A first-order interior-point method for linearly constrained smooth optimization. Mathematical Programming, 127:399–424, 2011
2011
-
[60]
Vandereycken
B. Vandereycken. Low-rank matrix completion by riemannian optimization. SIAM Journal on Optimization , 23(2):1214–1236, 2013
2013
-
[61]
S. J. Wright. Primal-Dual Interior-Point Methods . SIAM, 1997
1997
-
[62]
N. Xiao, K. Ding, X. Hu, and K.-C. Toh. Developing lagrangian-based methods for nonsmooth nonconvex optimization. arXiv preprint arXiv:2404.09438 , 2024
2024 arXiv
-
[63]
N. Xiao, X. Hu, X. Liu, and K.-C. Toh. Adam-family methods for nonsmooth optimization with convergence guarantees. arXiv preprint arXiv:2305.03938 , 2023
2023 arXiv
-
[64]
N. Xiao, X. Hu, and K.-C. Toh. Convergence guarantees for stochastic subgradient methods in nonsmooth nonconvex optimization. arXiv preprint arXiv:2307.10053 , 2023. 33
2023 arXiv
-
[65]
N. Xiao, T. Tang, S. Wang, and K.-C. Toh. An exact penalty approach for equality constrained optimization over a convex set. arXiv preprint arXiv:2505.02495 , 2025
2025 arXiv
-
[66]
Yang and K.-C
L. Yang and K.-C. Toh. Bregman proximal point algorithm revisited: A new inexact version and its inertial variant. SIAM J. Optimization , 32(3):1523–1554, 2022
2022
-
[67]
Zass and A
R. Zass and A. Shashua. Nonnegative sparse pca. Advances in neural information processing systems, 19, 2006
2006
-
[68]
J. Zhang. Stochastic bregman proximal gradient method revisited: Kernel conditioning and painless variance reduction. arXiv preprint arXiv:2401.03155 , 2024
2024
-
[69]
H. Zou, T. Hastie, and R. Tibshirani. Sparse principal component analysis. Journal of com- putational and graphical statistics , 15(2):265–286, 2006. 34
2006
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.