Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Local regularity alone yields finite-time KKT rates for nonconvex composite constrained problems via a truncated prox-linear ALM.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.5

2026-07-13 05:31 UTC pith:ZR7VHPHD

load-bearing objection Solid first nonasymptotic ALM rates for nonsmooth nonconvex composite inequalities under only local multiplier regularity; the finite-time feasibility-to-KKT transfer is the real contribution. the 2 major comments →

arxiv 2607.08954 v1 pith:ZR7VHPHD submitted 2026-07-09 math.OC cs.LGstat.ML

Nonconvex Composite Functional Constraints via First-Order Augmented Lagrangian Methods under Local Regularity

classification math.OC cs.LGstat.ML MSC 90C2690C3090C4649J52
keywords nonconvex constrained optimizationconvex-composite structureaugmented Lagrangianprox-linear methodlocal conic regularityKKT residualdual error boundminimax reformulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper shows that first-order augmented Lagrangian methods can produce approximate KKT points for nonsmooth nonconvex problems whose objective and inequality constraints are convex outer functions of smooth maps, without global multiplier bounds. The algorithm truncates the dual variables to a compact set so that a nonsmooth minimax analysis applies, then uses a large enough penalty to force almost all iterates into a near-feasible region. On that region a local conic regularity condition bounds the multipliers and deactivates the artificial truncation, converting truncated stationarity into a genuine KKT certificate. With dual regularization the KKT residual decays as O(K^{-1/3}); without regularization, and under piecewise-linear outer structure, the rate improves to O(K^{-1/2}). A sympathetic reader cares because the argument separates feasibility recovery from multiplier control and thereby weakens the regularity assumptions that previous nonasymptotic ALM analyses required.

Core claim

For a sufficiently large penalty parameter, all but a controlled number of iterates of the smoothed prox-linear ALM enter a near-feasible region on which local conic regularity uniformly bounds the associated prox-linear multipliers, rendering the artificial dual truncation inactive; the resulting KKT residual of the original constrained problem is then O(K^{-1/3}) with dual regularization and O(K^{-1/2}) without it under piecewise-linear outer functions.

What carries the argument

The finite-time KKT-transfer mechanism: Lyapunov counting arguments first produce near-feasible, nearly stationary iterates; local conic regularity then bounds the prox-linear multipliers so that the artificial dual radius becomes inactive and truncated minimax stationarity becomes a genuine KKT residual.

Load-bearing premise

A uniform positive lower bound on how far the linearized constraint gradients stay from the origin on a neighborhood of near-feasible points; if that constant vanishes, the multiplier bound and the whole transfer argument fail.

What would settle it

Construct a convex-composite problem that satisfies every other hypothesis yet has no positive local conic-regularity constant on any near-feasible set, then check whether the algorithm’s dual iterates remain unbounded or the claimed KKT residual rates fail to hold.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper develops nonasymptotic KKT complexity guarantees for a smoothed prox-linear augmented Lagrangian method applied to nonsmooth nonconvex problems whose objective and inequality constraints are convex-composite (convex Lipschitz outer functions of smooth inner maps). To handle the lack of a priori multiplier bounds and the artificial dual truncation needed for minimax analysis, the authors introduce a compact dual set Y and prove a finite-time transfer mechanism: for large enough penalty, all but a controlled number of iterates enter a near-feasible region on which a local conic regularity condition (Assumption 2.2) uniformly bounds the prox-linear multipliers, rendering the dual truncation inactive. With dual regularization this yields an O(K^{-1/3}) KKT residual rate; without regularization, under piecewise-linear outer functions and local structural assumptions, a local dual error bound yields O(K^{-1/2}). The argument is organized as a clean chain of Lyapunov descent, dual error bounds, counting lemmas, and residual conversion culminating in Theorem 4.1.

Significance. If correct, the result is a genuine advance for nonasymptotic ALM analysis of nonsmooth nonconvex functional inequalities: multiplier control is obtained from purely local regularity on a recovered near-feasible set rather than from global error bounds, uniform CQs, or a priori multiplier bounds. The separation of feasibility recovery (penalty + Lyapunov counting, independent of any CQ), local multiplier boundedness, dual-truncation inactivity, and dual error bounds is conceptually clean and should be reusable. The local dual error bound for piecewise-linear composite dual maps (Proposition 4.3), obtained via a lifted epigraphical NLP and Robinson strong regularity, is of independent technical interest. Constants are tracked explicitly and the parameter-selection order is largely consistent. The work sits squarely in the current complexity literature on nonconvex constrained first-order methods and strengthens the case that local regularity can replace global multiplier-control assumptions once near-feasibility is algorithmically enforced.

major comments (2)
  1. [Theorem 4.1(i), Proposition 4.6, Eq. (4.4)] Theorem 4.1(i) and the surrounding parameter choices: when ry = Θ(K^{-1/3}), β = Θ(K^{-1/3}) and ξ = Θ(K^{-2/3}), one has 1/(βξ) = Θ(K), so the subtracted term 7(Φ0 - fmin)/(2βξ) in NK is Θ(K). For NK = Ω(K) (needed for the claimed O(K^{-1/3}) residual) the hidden constants inside the Θ notation must be chosen so that this term is at most, say, K/2, after which ρ is fixed large enough that the feasibility-counting term is also o(K). The paper uses Θ notation without spelling out this constant-selection step. Please add a short remark after Theorem 4.1 verifying that admissible constants exist and that they remain compatible with the upper bounds on β coming from ω1 (which itself blows up as ry o 0).
  2. [Remark 4.3, Definition 2.1, Eq. (4.5), Proposition 4.1] Parameter interdependence and selection order: L ho := L(1 + Ry + ρ Rx) depends on Ry, while the dual-radius lower bound (4.5) itself depends on ρ, √ξ and √rx. The algorithm requires all of ρ, Ry, rx > L ho, α, β to be fixed before iteration begins. Remark 4.3 correctly notes that Φ0 - fmin is independent of ρ and Ry, but the full cascade (ρ first, then Ry, then rx, then α/β, with ξ possibly K-dependent) is never collected in one place. A single explicit selection protocol (even if only asymptotic) would remove any doubt that the conditions of Propositions 4.1, 4.4 and 4.6 can be satisfied simultaneously.
minor comments (6)
  1. [Remark 3.1] Remark 3.1 acknowledges that the analysis assumes exact solutions of the strongly convex prox-linear subproblems. A one-sentence pointer to the relative-error / summable-error criteria under which the Lyapunov and counting arguments would survive (even without a full inexact proof) would help practitioners.
  2. [Assumption 4.1(iii), Remark 4.1] Assumption 4.1(iii) (primal interiority) is used only for the local dual error bound. Remark 4.1 already notes the alternative of including NX in the regularity condition; a brief forward reference from the statement of Assumption 4.1 to that remark would improve readability.
  3. [Algorithm 1, Introduction] The initialization requires a feasible x0. This is standard but should be flagged in the abstract or introduction as a standing hypothesis of Algorithm 1, since many competing primal-only methods do not need it.
  4. [Assumption 2.1(ii), Eq. (5.11)] Notation: the same symbol L is used both for the composite Lipschitz constant and (implicitly) as a bound on subgradient norms after (5.11). A short clarifying sentence would avoid confusion.
  5. [Proof of Proposition 4.5] In the proof of Proposition 4.5, the three cases for the inequality (5.8) are carefully checked; a one-line summary that the argument is componentwise and uses only the projection property onto the ℓ1-ball would help the reader navigate the case split.
  6. [References] Several references to concurrent or very recent arXiv preprints (e.g., [20], [37]) are natural; ensure final bibliographic data are updated at production time.

Circularity Check

0 steps flagged

No circularity: KKT rates follow from Lyapunov counting, penalty-driven near-feasibility, and local CQ applied only after iterates enter the near-feasible region; R_y is chosen after the multiplier bound, not by definition of the target residual.

full rationale

The paper is a self-contained nonasymptotic complexity analysis under explicitly stated assumptions (Ass. 2.1–2.2, and Ass. 4.1 only for the unregularized local dual error bound). The derivation chain separates cleanly: (i) basic Lyapunov decrease of Φ on the compact dual set Y (Prop. 4.1, standard nonconvex-concave minimax estimates); (ii) dual error bounds that absorb the sensitivity term (global from strong concavity when r_y>0; local from piecewise-linear structure + pointwise LICQ/strict complementarity when r_y=0); (iii) a pure penalty counting argument that forces all but O(1/(ρδ²)) iterates into a near-feasible region, independent of any CQ (Prop. 4.5); (iv) on those iterates Ass. 2.2 supplies a uniform bound on the prox-linear multipliers γ^k, which is used to choose R_y large enough that the artificial truncation of Y is inactive (Prop. 4.6); (v) residual conversion to an original-problem KKT certificate (Lemma 5.2 → Thm. 4.1). Remark 4.3 shows Φ_0−f_min is independent of ρ and R_y, so the parameter order is consistent rather than circular. Self-citations ([20],[21],[37]) supply standard minimax Lyapunov tools; they are not uniqueness theorems that force the rates, and the central KKT-transfer mechanism is developed in the paper. There are no fitted parameters, no data-driven “predictions,” and no renaming of known empirical patterns. Score 0 is therefore the correct outcome.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The paper is a pure complexity analysis. All free parameters are algorithmic step-sizes or penalty values chosen to satisfy explicit inequalities derived from problem constants; the axioms are standard convex-analysis facts plus the two domain-specific regularity conditions that define the problem class.

free parameters (3)
  • penalty ρ
    Must be larger than 2(Φ_0 - f_min)/δ^{2}; chosen after the Lyapunov gap is known, not fitted to data.
  • dual radius R_y
    Chosen strictly larger than the local multiplier bound produced by Assumption 2.2; algorithmic device, not a fitted constant.
  • dual regularization r_y
    Set to Θ(K^{-1/3}) for the slower rate or to zero for the faster rate; free algorithmic choice.
axioms (4)
  • domain assumption Assumption 2.1: outer functions convex Lipschitz, inner maps C^{2} with Lipschitz Jacobian, X compact convex.
    Defines the convex-composite problem class; standard in prox-linear literature.
  • domain assumption Assumption 2.2: local uniform conic multiplier regularity on the near-feasible set R_δ cq.
    Load-bearing local CQ that replaces global error bounds used in prior ALM analyses.
  • domain assumption Assumption 4.1: piecewise-linear outer functions, strict complementarity, pointwise active-plane LICQ, primal interiority (used only for r_y=0).
    Enables the local dual error bound that removes dual regularization bias.
  • standard math Standard convex subdifferential calculus, Danskin theorem, projection nonexpansiveness, Robinson strong regularity.
    Invoked throughout the dual-error-bound and residual-conversion proofs.
invented entities (1)
  • Auxiliary compact dual set Y = {y ≥ 0 : ||y||_1 ล R_y} independent evidence
    purpose: Makes the dual domain compact so that projected primal-dual estimates apply; later shown inactive at good iterates.
    Analytical device, not a physical entity; independent evidence is the explicit construction and the inactivity proof.

pith-pipeline@v1.1.0-grok45 · 35375 in / 2521 out tokens · 27601 ms · 2026-07-13T05:31:26.288382+00:00 · methodology

0 comments
read the original abstract

We study nonasymptotic convergence of primal-dual methods for a class of nonconvex constrained optimization problems with a convex-composite structure. In this class, both the objective and the functional inequality constraints are given by convex Lipschitz outer functions composed with smooth nonlinear inner mappings. The analysis is complicated by constraint violation in a nonconvex functional inequality system and by the lack of an a priori bound on the multipliers. To address these issues, we restrict the dual variable to an auxiliary compact set and analyze a smoothed prox-linear augmented Lagrangian method through a nonsmooth nonconvex-concave minimax reformulation. The main contribution is a finite-time mechanism for converting stationarity of the truncated minimax problem into a KKT certificate for the original constrained problem. We show that, for a sufficiently large penalty parameter, all but a controlled number of iterates enter a near-feasible region. On this region, a local conic regularity condition uniformly bounds the associated prox-linear multipliers and thereby makes the artificial dual truncation inactive at the selected iterates. Building on this mechanism, we establish explicit convergence rates for the proposed method in terms of the KKT residual. With dual regularization, a global dual error bound together with a bias-balancing argument gives an $O(K^{-1/3})$ rate. In the unregularized case, under additional local structural assumptions including piecewise linearity of the outer functions, a local dual error bound yields the sharper $O(K^{-1/2})$ rate.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Online Optimization of Difference-of-Convex Compositions with Smooth Mappings

    math.OC 2026-07 conditional novelty 6.0

    An online proximal-linear algorithm for difference-of-convex-composite objectives and constraints attains O(T/w^2) local regret, with a proximal residual that certifies first-order stationarity.

Reference graph

Works this paper leans on

37 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Complexity of single loop algorithms for nonlin- ear programming with stochastic objective and constraints

    Ahmet Alacaoglu and Stephen J Wright. Complexity of single loop algorithms for nonlin- ear programming with stochastic objective and constraints. InInternational Conference on Artificial Intelligence and Statistics, pages 4627–4635. PMLR, 2024. 22

  2. [2]

    A relaxed quasinormality condition and the boundedness of dual augmented La- grangian sequences.SIAM Journal on Optimization, 35(4):2474–2489, 2025

    Roberto Andreani, Gabriel Haeser, Maria Laura Schuverdt, and Leornardo Delarmelina Secchin. A relaxed quasinormality condition and the boundedness of dual augmented La- grangian sequences.SIAM Journal on Optimization, 35(4):2474–2489, 2025

  3. [3]

    A relaxed constant positive linear dependence constraint qualification and applications.Mathematical Programming, 135(1):255–273, 2012

    Roberto Andreani, Gabriel Haeser, Maria Laura Schuverdt, and Paulo JS Silva. A relaxed constant positive linear dependence constraint qualification and applications.Mathematical Programming, 135(1):255–273, 2012

  4. [4]

    Princeton University Press, 2009

    Aharon Ben-Tal, Laurent El Ghaoui, and Arkadi Nemirovski.Robust Optimization, volume 28. Princeton University Press, 2009

  5. [5]

    Brown, and Constantine Caramanis

    Dimitris Bertsimas, David B. Brown, and Constantine Caramanis. Theory and applications of robust optimization.SIAM review, 53(3):464–501, 2011

  6. [6]

    Nonconvex Lagrangian-based optimization: monitoring schemes and global convergence.Mathematics of Operations Research, 43(4):1210– 1232, 2018

    Jérôme Bolte, Shoham Sabach, and Marc Teboulle. Nonconvex Lagrangian-based optimization: monitoring schemes and global convergence.Mathematics of Operations Research, 43(4):1210– 1232, 2018

  7. [7]

    Frédéric Bonnans and Alexander Shapiro.Perturbation Analysis of Optimization Problems

    J. Frédéric Bonnans and Alexander Shapiro.Perturbation Analysis of Optimization Problems. Springer, 2000

  8. [8]

    Stochastic first-order methods for convex and nonconvex functional constrained optimization.Mathematical Programming, 197(1):215–279, 2023

    Digvijay Boob, Qi Deng, and Guanghui Lan. Stochastic first-order methods for convex and nonconvex functional constrained optimization.Mathematical Programming, 197(1):215–279, 2023

  9. [9]

    Distributed optimization and statistical learning via the alternating direction method of multipliers.Foun- dations and Trends in Machine Learning, 3(1):1–122, 2011

    Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating direction method of multipliers.Foun- dations and Trends in Machine Learning, 3(1):1–122, 2011

  10. [10]

    Burke and Abraham Engle

    James V. Burke and Abraham Engle. Strong metric (sub)regularity of Karush–Kuhn–Tucker mappings for piecewise linear-quadratic convex-composite optimization and the quadratic con- vergence of Newton’s method.Mathematics of Operations Research, 45(3):1164–1192, 2020

  11. [11]

    Damped proximal augmented Lagrangian method for weakly-convex problems with convex constraints.Mathematical Programming Computation, pages 1–50, 2026

    Hari Dahal, Wei Liu, and Yangyang Xu. Damped proximal augmented Lagrangian method for weakly-convex problems with convex constraints.Mathematical Programming Computation, pages 1–50, 2026

  12. [12]

    Stochastic model-based minimization of weakly con- vex functions.SIAM Journal on Optimization, 29(1):207–239, 2019

    Damek Davis and Dmitriy Drusvyatskiy. Stochastic model-based minimization of weakly con- vex functions.SIAM Journal on Optimization, 29(1):207–239, 2019

  13. [13]

    Dontchev and R

    Asen L. Dontchev and R. Tyrrell Rockafellar.Implicit Functions and Solution Mappings, volume 543. Springer, 2009

  14. [14]

    Efficiency of minimizing compositions of convex functions and smooth maps.Mathematical Programming, 178:503–558, 2019

    Dmitriy Drusvyatskiy and Courtney Paquette. Efficiency of minimizing compositions of convex functions and smooth maps.Mathematical Programming, 178:503–558, 2019

  15. [15]

    Chuan He, Zhaosong Lu, and Ting Kei Pong. A Newton-CG based augmented Lagrangian method for finding a second-order stationary point of nonconvex equality constrained opti- mization with complexity guarantees.SIAM Journal on Optimization, 33(3):1734–1766, 2023

  16. [16]

    Hestenes

    Magnus R. Hestenes. Multiplier and gradient methods.Journal of Optimization Theory and Applications, 4(5):303–320, 1969. 23

  17. [17]

    Inexact Moreau envelope Lagrangian method for non-convex constrained optimization under local error bound conditions on constraint func- tions.arXiv preprint arXiv:2502.19764, 2025

    Yankun Huang, Qihang Lin, and Yangyang Xu. Inexact Moreau envelope Lagrangian method for non-convex constrained optimization under local error bound conditions on constraint func- tions.arXiv preprint arXiv:2502.19764, 2025

  18. [18]

    First-order methods for nonsmooth nonconvex func- tional constrained optimization with or without slater points.SIAM Journal on Optimization, 35(2):1300–1329, 2025

    Zhichao Jia and Benjamin Grimmer. First-order methods for nonsmooth nonconvex func- tional constrained optimization with or without slater points.SIAM Journal on Optimization, 35(2):1300–1329, 2025

  19. [19]

    Lewis and Stephen J

    Adrian S. Lewis and Stephen J. Wright. A proximal method for composite minimization. Mathematical Programming, 158(1):501–546, 2016

  20. [20]

    Smoothing meets perturba- tion: Unified and tight analysis for nonconvex-concave minimax optimization.arXiv preprint arXiv:2602.14185, 2026

    Jiajin Li, Mahesh Nagarajan, Siyu Pan, and Nanxi Zhang. Smoothing meets perturba- tion: Unified and tight analysis for nonconvex-concave minimax optimization.arXiv preprint arXiv:2602.14185, 2026

  21. [21]

    Nonsmooth nonconvex–nonconcave minimax optimization: Primal–dual balancing and iteration complexity analysis.Mathematical Programming, 214(1-2):591–641, 2025

    Jiajin Li, Linglingzhi Zhu, and Anthony Man-Cho So. Nonsmooth nonconvex–nonconcave minimax optimization: Primal–dual balancing and iteration complexity analysis.Mathematical Programming, 214(1-2):591–641, 2025

  22. [22]

    Stochastic inexact aug- mented Lagrangian method for nonconvex expectation constrained optimization.Computa- tional Optimization and Applications, 87(1):117–147, 2024

    Zichong Li, Pin-Yu Chen, Sijia Liu, Songtao Lu, and Yangyang Xu. Stochastic inexact aug- mented Lagrangian method for nonconvex expectation constrained optimization.Computa- tional Optimization and Applications, 87(1):117–147, 2024

  23. [23]

    Complexity of an inexact proximal-point penalty method for constrained smooth non-convex optimization.Computational Optimization and Applications, 82(1):175–224, 2022

    Qihang Lin, Runchao Ma, and Yangyang Xu. Complexity of an inexact proximal-point penalty method for constrained smooth non-convex optimization.Computational Optimization and Applications, 82(1):175–224, 2022

  24. [24]

    A SPIDER-type stochastic subgradient method for expectation- constrained nonconvex nonsmooth optimization.SIAM Journal on Optimization, 36(2):1125– 1153, 2026

    Wei Liu and Yangyang Xu. A SPIDER-type stochastic subgradient method for expectation- constrained nonconvex nonsmooth optimization.SIAM Journal on Optimization, 36(2):1125– 1153, 2026

  25. [25]

    Quadratically regularized subgradient methods for weakly convex optimization with weakly convex constraints

    Runchao Ma, Qihang Lin, and Tianbao Yang. Quadratically regularized subgradient methods for weakly convex optimization with weakly convex constraints. InInternational Conference on Machine Learning, pages 6554–6564. PMLR, 2020

  26. [26]

    Michael J. D. Powell. A method for nonlinear constraints in minimization problems.Optimiza- tion, pages 283–298, 1969

  27. [27]

    Smoothed proximal Lagrangian method for nonlinear constrained programs.arXiv preprint arXiv:2408.15047, 2024

    Wenqiang Pu, Kaizhao Sun, and Jiawei Zhang. Smoothed proximal Lagrangian method for nonlinear constrained programs.arXiv preprint arXiv:2408.15047, 2024

  28. [28]

    Robinson

    Stephen M. Robinson. Strongly regular generalized equations.Mathematics of Operations Research, 5(1):43–62, 1980

  29. [29]

    Tyrrell Rockafellar

    R. Tyrrell Rockafellar. A dual approach to solving nonlinear programming problems by uncon- strained optimization.Mathematical Programming, 5(1):354–373, 1973

  30. [30]

    Tyrrell Rockafellar

    R. Tyrrell Rockafellar. Monotone operators and the proximal point algorithm.SIAM Journal on Control and Optimization, 14(5):877–898, 1976

  31. [31]

    Optimizationofconditionalvalue-at-risk.Journal of Risk, 2:21–42, 2000

    R.TyrrellRockafellarandStanislavUryasev. Optimizationofconditionalvalue-at-risk.Journal of Risk, 2:21–42, 2000. 24

  32. [32]

    Yue Xie and Stephen J. Wright. Complexity of proximal augmented Lagrangian for nonconvex optimization with nonlinear equality constraints.Journal of Scientific Computing, 86:1–30, 2021

  33. [33]

    First-order methods for constrained convex programming based on linearized augmented Lagrangian function.INFORMS Journal on Optimization, 3(1):89–117, 2021

    Yangyang Xu. First-order methods for constrained convex programming based on linearized augmented Lagrangian function.INFORMS Journal on Optimization, 3(1):89–117, 2021

  34. [34]

    Single-loop algorithms for stochastic nonconvex optimization with weakly convex constraints.Transactions on Machine Learning Research, 2026

    Ming Yang, Gang Li, Quanqi Hu, Qihang Lin, and Tianbao Yang. Single-loop algorithms for stochastic nonconvex optimization with weakly convex constraints.Transactions on Machine Learning Research, 2026

  35. [35]

    A single-loop smoothed gradient descent-ascent algorithm for nonconvex-concave min-max problems.Advances in Neural Infor- mation Processing Systems, 33:7377–7389, 2020

    Jiawei Zhang, Peijun Xiao, Ruoyu Sun, and Zhiquan Luo. A single-loop smoothed gradient descent-ascent algorithm for nonconvex-concave min-max problems.Advances in Neural Infor- mation Processing Systems, 33:7377–7389, 2020

  36. [36]

    A first-order primal-dual method for nonconvex constrained optimization based on the augmented Lagrangian.Mathematics of Operations Research, 49(1):125–150, 2024

    Daoli Zhu, Lei Zhao, and Shuzhong Zhang. A first-order primal-dual method for nonconvex constrained optimization based on the augmented Lagrangian.Mathematics of Operations Research, 49(1):125–150, 2024

  37. [37]

    Primal-dual meth- ods for nonsmooth nonconvex optimization with orthogonality constraints.arXiv preprint arXiv:2604.04130, 2026

    Linglingzhi Zhu, Wentao Ding, Shangyuan Liu, and Anthony Man-Cho So. Primal-dual meth- ods for nonsmooth nonconvex optimization with orthogonality constraints.arXiv preprint arXiv:2604.04130, 2026. A Useful Technical Lemmas To begin with, we introduce the weakly convex function which plays an important role in our following analysis. Definition A.1.The fu...