Pith. sign in

REVIEW 2 major objections 4 minor 106 references

The paper shows that a proximal residual computed from convex subproblems certifies first-order stationarity for online nonconvex optimization with difference-of-convex composite losses and constraints.

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 →

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

T0 review · deepseek-v4-flash

2026-08-01 12:24 UTC pith:WT5J4URP

load-bearing objection A genuine extension of local-regret to CDCS objectives with CDCS constraints, but the stationarity certificate depends on a Slater condition that is not verified in the examples; worth refereeing after revision. the 2 major comments →

arxiv 2607.19553 v1 pith:WT5J4URP submitted 2026-07-21 math.OC cs.LG

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

Jingwei Ji , Jong-Shi Pang , Renyuan Xu This is my paper
classification math.OC cs.LG MSC 90C2690C2590C30
keywords online optimizationdifference-of-convexnonconvex nonsmooth optimizationproximal residualstationarity measuretangent conelocal regretconstraint qualification
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.

The paper studies online learning where each round's loss is a difference-of-convex function composed with a smooth map, and where the feasible region is itself defined by a constraint of the same kind. Because generic nonconvex problems are intractable, the paper adopts a stationarity-based benchmark rather than global optimality. Its central claim is that a time-smoothed proximal-linear algorithm, whose inner updates are convex subproblems, has a natural residual that is a proper stationarity measure: whenever the residual's fixed-point condition holds at a point, that point satisfies a first-order stationarity condition for the original nonconvex constrained problem. The key enabling result is a tangent-cone characterization showing that a carefully cushioned linearization of the constraint set preserves the local geometry of the true feasible set. If correct, the paper delivers a local-regret bound, a bound on the number of inner convex solves, and an error bound that turns the residual into a computable certificate of approximate stationarity.

Core claim

On its own terms, the paper's discovery is that the proximal-residual mapping G_t^μ(x)=μ(S_t^μ(x)−x), defined as the difference between the current point and the minimizer of a convex proximal-linear surrogate, is a faithful first-order stationarity measure for the nonconvex constrained problem. Proposition 2 proves that the fixed-point condition x=S_t^μ(x;v) implies 0∈b∂Φ_t(x)+T_X(x)^*, which is Bouligand stationarity of the time-smoothed loss over the original feasible set. The mechanism is Theorem 1: under a CDCS Slater condition, the tangent cone of the convex surrogate set equals the tangent cone of the true feasible set, so moving between the surrogate and the original problem loses no

What carries the argument

The cushioned linearization set L̃(x): a convex set formed by linearizing the inner smooth map ζ at x inside the convex function φ, linearizing the entire composite concave part ψ∘ζ, and adding quadratic cushion terms ±½Lβ‖y−x‖² to both sides. Theorem 1 says that under the CDCS Slater condition, T_{L̃(x)}(x)=T_X(x)=C(x), where C(x) is the cone of directions d satisfying φ'(ζ(x);Jζ(x)d) ≤ ∇ψ(ζ(x))ᵀJζ(x)d. This tangent-cone equality is what allows the proximal-linear mapping S_t^μ(x), a convex problem over L̃(x), to certify stationarity of the original nonconvex problem. The proximal residual G_t^μ(x) is the scaled displacement produced by this mapping.

Load-bearing premise

At every boundary point of the feasible region where the constraint is active, there must exist a direction along which the convex part grows strictly slower than the linearized concave part; without such a Slater-type direction, the cushioned surrogate set may fail to reproduce the true tangent cone, and the stationarity certificate can break.

What would settle it

Build a two-constraint CDCS example where both constraints are active at a boundary point so condition (28) has no solution, run the algorithm, and check whether a fixed point x̄=S_t^μ(x̄;v) still satisfies 0∈b∂Φ_t(x̄)+T_X(x̄)^*. If a fixed point occurs without the stationarity inclusion, that example marks exactly where the Slater assumption is load-bearing.

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

If this is right

  • The local-regret bound Regret_T ≲ T(δ/w+(L_f+L_g)L_θ/w)² says the iterates are approximately stationary on average, with per-round deviation controlled by the stopping tolerance and the smoothing window.
  • Each inner iteration solves a convex optimization problem, so the method runs with a convex oracle even though the objective and constraints are nonconvex.
  • The bound on total inner iterations, O(Mμ²w²/(μ−(L_f+L_g)β_θ)δ⁻²(1+T/w)), keeps the computational overhead finite and polynomial in the precision requirement.
  • Feeding the same loss at every round recovers an offline method: suitable choices of window and tolerance yield an iterate whose expected squared residual is below any ε, i.e., a first-order stationary point.
  • Under additional regularity, Proposition 4's error bound makes the residual a quantitative certificate: a small residual implies the next iterate is close to the true stationarity set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Not in the paper: the tangent-cone characterization is likely reusable for offline optimization with CDCS constraints, where it would justify convex-surrogate algorithms without a separate stationarity analysis.
  • Not in the paper: when the feasible region is a max of several CDCS constraints, the single-constraint Slater condition must be checked for that max representation; verifying or relaxing it for multiple simultaneously active constraints would extend the framework to common constraint structures.
  • Not in the paper: replacing the uniform sliding window with adaptive or exponentially weighted smoothing should preserve the proof's logic, since the regret argument only needs the smoothed objectives to change by a bounded amount between rounds.
  • Not in the paper: the residual could serve as a practical stopping criterion in stochastic or bandit-like online settings where only partial loss information is available at decision time.

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 / 4 minor

Summary. The paper proposes an online optimization framework for nonconvex nonsmooth losses of CDCS form f_t(θ_t(x))−g_t(θ_t(x)) over a nonconvex feasible set X={x: φ(ζ(x))−ψ(ζ(x))≤0}, where φ,ψ are convex, ψ,g are differentiable, and θ,ζ are smooth. The core algorithmic idea is a time-smoothed proximal-linear method whose inner subproblems are convex over a "cushioned linearization" set L̃(·). The paper defines a proximal residual G_t^μ(x)=μ(S_t^μ(x)−x) and a local-regret measure Σ∥G_t^μ(x_t)∥². Its main theoretical claims are: (i) Theorem 1, a tangent-cone characterization T_{L̃(x̄)}=T_X(x̄)=C(x̄) under a CDCS Slater CQ; (ii) Proposition 2, stating that a fixed point of S_t^μ implies first-order stationarity 0∈∂̂Φ_t(x̄)+T_X(x̄)*; (iii) Theorem 2, giving a local-regret bound of order T(δ/w+(L_f+L_g)L_θ/w)² and an inner-iteration bound; and (iv) Proposition 4, an error bound relating ∥G_t^μ(x)∥ to dist(0,∂̂Φ_t(x⁺)+T_X(x⁺)*). The proofs of the regret and iteration bounds are largely self-contained and appear coherent.

Significance. If the stated results hold, this would be a meaningful extension of the local-regret framework to a broad structured nonconvex nonsmooth class with nonconvex constraints, while keeping all inner updates convex. The tangent-cone characterization in Theorem 1 is of independent interest for offline CDCS-constrained optimization. The main regret bound has the expected scaling and the iteration-complexity analysis is nontrivial. However, the significance is currently weakened by two substantial gaps: the stationarity certificate depends on a CDCS Slater condition that is not verified for the paper's own examples and is not automatic even for simple constraints; and Proposition 4 relies on smoothness assumptions on ∇g and ∇ψ that are not stated in the paper's assumptions. These gaps affect the central claim that the residual is a proper stationarity measure, and one of the paper's advertised contributions.

major comments (2)
  1. [§3.1, Definition 2 and Theorem 1, eq. (28)] The CDCS Slater CQ (28) is load-bearing for the equality T_{L̃(x̄)}=T_X(x̄) and hence for Proposition 2, but it is not automatic and is not verified for the examples. Counterexample: let X=[0,∞) be represented by φ(t)=|t|, ψ(t)=t, ζ(x)=x³/(1+x²). All favorable-condition assumptions hold. At x̄=0, Jζ(0)=0, so no d0 satisfies φ'(0;0)<0; Slater fails. Then L̃(0)={0}, so T_{L̃}(0)={0}, but T_X(0)=[0,∞). With Φ(x)=−x, S(0)=0 and G(0)=0, yet 0∉∂̂Φ(0)+T_X(0)*=(−∞,−1]. Thus the residual is not a proper stationarity measure without Slater. Moreover, Example 4 (the Huber ℓ0 budget constraint, eq. (17)) fails (28) at every active point: at such a point the zero coordinates are exactly zero and the large coordinates have zero derivative, so φ'(x;d)−∇ψ(x)ᵀd = Σ_{j:x_j=0}|d_j| ≥ 0 for all d. The abstract's unqualified claim that the residual is a proper stationarity measure is therefore too broad; the
  2. [§3.5, Proposition 4, eqs. (95) and (99)] The proof of Proposition 4 invokes Lipschitz constants β_g and β_ψ for ∇g_s and ∇ψ, respectively. Neither Assumption 1 nor the favorable condition (Definition 1) includes smoothness of ∇g or ∇ψ; only L-Lipschitzness and differentiability of g and ψ are assumed. In bounding ∥∇g_s(θ_s(x⁺))−∇g_s(θ_s(x))∥ and the analogous term for ∇ψ, the proof uses exactly these unstated constants. Consequently, the error bound (81) is not established under the assumptions stated in the proposition. The proposition should be restated with explicit smoothness assumptions on ∇g_s and ∇ψ (and then these should be verified for the examples, e.g., the Huber function does have a Lipschitz gradient), or the claim should be modified to avoid these terms.
minor comments (4)
  1. [§3.5, eq. (81)] The displayed bound (81) has malformed typesetting with stray '+ +' and missing parentheses, making it hard to parse; it should be corrected.
  2. [Abstract and §1.1] The phrase 'proper stationarity measure for the original problem' should be qualified by the CDCS Slater condition, since Proposition 2 explicitly depends on it.
  3. [Example 4, eqs. (14)–(17)] The constraint (15) is only sufficient, not equivalent, to the ℓ0 budget (14); the paper says 'it suffices,' but the subsequent discussion of feasible regions should not imply that (15) exactly characterizes the sparsity constraint.
  4. [§1.3] The reference 'Zhu and Li (2026)' cites arXiv:2607.08954 with a July 2026 date; please verify that this reference is complete and accurately described.

Circularity Check

0 steps flagged

No significant circularity: the stationarity certificate and regret bound are derived from stated assumptions, not from fitted values or load-bearing self-citations.

full rationale

The paper's central claim is that the proximal residual G_t^mu(x)=mu(S_t^mu(x)-x) is a proper stationarity measure. This is proved in Proposition 2 via Theorem 1, which characterizes T_X(bar x) as C(bar x) under the CDCS Slater condition. The proof of Theorem 1 is self-contained: it starts from the definitions of the cushioned linearization set L~, the candidate cone C(bar x), and the Bouligand tangent cone, and derives the equality by explicit inequalities and a constructed in-set sequence. No fitted parameter is renamed as a prediction, and no external benchmark is defined in terms of the algorithm's output. The local-regret bound in Theorem 2 follows directly from the algorithm's stopping rule (residual below delta/w) plus a Lipschitz drift bound between consecutive smoothed objectives; it is a proof of a property of the proposed update, not a fitted or self-referential prediction. The paper does cite works co-authored by one of the current authors (Cui and Pang 2021; Pang et al. 2017), but these citations are used for terminology, standard DC calculus, and context, not as the justification of the central tangent-cone or stationarity results. The CDCS Slater condition is a genuine assumption; whether it holds for all listed examples is a correctness/verification concern rather than a circularity. Overall, the derivation chain is not circular.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The central claim rests on the favorably structured CDCS class, a Slater-type constraint qualification, and standard variational-analysis machinery. Proposition 4 additionally requires smoothness and regularity conditions that are only partially stated. No new physical or mathematical entities are postulated; the cushioned linearization set is an algorithmic construction, not an invented entity.

free parameters (3)
  • μ (proximal coefficient) = μ > (L_f+L_g)β_θ (Theorem 2)
    Chosen by the user or via backtracking; the descent analysis and all bounds depend on it.
  • δ (inner-loop tolerance)
    Stopping threshold for the inner proximal-linear loop; appears in the local-regret and iteration-complexity bounds.
  • w (smoothing window)
    Window length for the time-smoothed objective; controls the local-regret rate O(T/w^2) and the number of inner iterations.
axioms (7)
  • domain assumption Favorable CDCS condition (Definition 1): f,g convex; f L_f-Lipschitz; g L_g-Lipschitz and differentiable; θ β_θ-smooth with bounded Jacobian; range bounded on the relevant set.
    Defines the problem class and supplies all constants used in the bounds; if g is not differentiable or θ is not smooth, the proximal-linear construction fails.
  • domain assumption CDCS Slater CQ at active boundary points (Definition 2, eq. (28)): ∃d0 with φ'(ζ(¯x);Jζ(¯x)d0)<∇ψ(ζ(¯x))^T Jζ(¯x)d0.
    Necessary for Theorem 1's tangent-cone equality and for Proposition 2's stationarity certificate; not verified for the examples when multiple constraints are active.
  • domain assumption ζ(¯x)∈int(dom φ) at active boundary points.
    Used in Theorem 1 to ensure φ is locally Lipschitz and Hadamard directionally differentiable at ζ(¯x).
  • domain assumption Uniform parameters over all rounds (Assumption 1).
    Allows the time-smoothed averages and regret/iteration bounds to use uniform constants L_f,L_g,β_θ,L_θ,M.
  • domain assumption For Proposition 4: existence of KKT multipliers, κ_f/κ_ϕ Lipschitz subgradient continuity (eqs. (78)-(80)), an η>0 with H_x(u_x)≤−η, and λ>0⇒x_+∈∂X.
    These 'suitable conditions' control the Lagrange multiplier and the gap between linearized and true subgradients; they are substantial and not implied by Assumption 1.
  • ad hoc to paper Unstated β_g and β_ψ smoothness of ∇g_s and ∇ψ in Proposition 4's proof.
    The proof bounds ||∇g_s(θ_s(x_+))−∇g_s(θ_s(x))||≤L_θ^2 β_g||d|| and ||∇ψ(ζ(x_+))−∇ψ(ζ(x))||≤L_ζ^2 β_ψ||d||, but β_g and β_ψ are not part of Definition 1, Assumption 1, or the proposition statement.
  • standard math Standard variational-analysis results: Rockafellar-Wets Thm 10.6, Mordukhovich Prop 1.107, Bonnans-Shapiro Prop 2.47, tangent/polar cone calculus.
    Used throughout the proofs without reproving; these are standard background results in the field.

pith-pipeline@v1.3.0-alltime-deepseek · 25527 in / 19086 out tokens · 158403 ms · 2026-08-01T12:24:06.055455+00:00 · methodology

0 comments
read the original abstract

We study online optimization for a broad class of structured non-convex non-smooth problems where each loss is a composition of a difference-of-convex function with a smooth mapping, and the feasible region is defined by constraint functions of the same kind. We propose a time-smoothed proximal linear algorithm and a local-regret measure based on a proximal residual mapping. We show that this residual is a proper stationarity measure for the original problem: its fixed-point condition implies first-order stationarity. Our analysis relies on a tangent-cone characterization for a feasible region described by composite difference-of-convex constraints, which is of independent interest and allows each update to be computed via a convex optimization oracle, despite the non-convexity of the problem. We establish a local-regret bound and a bound on the total number of inner convex subproblems. We also derive an error bound connecting the proximal residual to the distance to stationarity, providing a quantitative certificate of approximate stationarity.

discussion (0)

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

Reference graph

Works this paper leans on

106 extracted references · 4 linked inside Pith

  1. [1]

    Foundations and Trends

    Introduction to online convex optimization , author=. Foundations and Trends. 2016 , publisher=

    2016

  2. [2]

    , author=

    X-Armed Bandits. , author=. Journal of Machine Learning Research , volume=

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Nearly tight bounds for the continuum-armed bandit problem , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Proceedings of the fortieth annual ACM symposium on Theory of computing , pages=

    Multi-armed bandits in metric spaces , author=. Proceedings of the fortieth annual ACM symposium on Theory of computing , pages=

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    Online convex optimization with stochastic constraints , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Online Learning under Adversarial Nonlinear Constraints , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    International Conference on Machine Learning , pages=

    Adaptive algorithms for online convex optimization with long-term constraints , author=. International Conference on Machine Learning , pages=. 2016 , organization=

    2016

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Online convex optimization for cumulative constraints , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    IEEE Transactions on automatic control , volume=

    Online convex optimization with time-varying constraints and bandit feedback , author=. IEEE Transactions on automatic control , volume=. 2018 , publisher=

    2018

  10. [10]

    SIAM Journal on Optimization , volume=

    Convergence analysis of a proximal-like minimization algorithm using Bregman functions , author=. SIAM Journal on Optimization , volume=. 1993 , publisher=

    1993

  11. [11]

    International Conference on Machine Learning , pages=

    The hedge algorithm on a continuum , author=. International Conference on Machine Learning , pages=. 2015 , organization=

    2015

  12. [12]

    Algorithmic Learning Theory , pages=

    Online non-convex learning: Following the perturbed leader is optimal , author=. Algorithmic Learning Theory , pages=. 2020 , organization=

    2020

  13. [13]

    Joint european conference on machine learning and knowledge discovery in databases , pages=

    Online learning in adversarial lipschitz environments , author=. Joint european conference on machine learning and knowledge discovery in databases , pages=. 2010 , organization=

    2010

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Online non-convex learning in dynamic environments , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Transactions on Machine Learning Research , year=

    Dynamic Regret Analysis of Safe Distributed Online Optimization for Convex and Non-convex Problems , author=. Transactions on Machine Learning Research , year=

  16. [16]

    International Conference on Machine Learning , pages=

    Efficient regret minimization in non-convex games , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    2017

  17. [18]

    International Conference on Machine Learning , pages=

    Regret minimization in stochastic non-convex learning via a proximal-gradient approach , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  18. [19]

    Advances in neural information processing systems , volume=

    Dynamic local regret for non-convex online forecasting , author=. Advances in neural information processing systems , volume=

  19. [20]

    2025 , month = feb, note =

    Nadav Hallak , title =. 2025 , month = feb, note =

    2025

  20. [21]

    The Thirty Sixth Annual Conference on Learning Theory , pages=

    Online nonconvex optimization with limited instantaneous oracle feedback , author=. The Thirty Sixth Annual Conference on Learning Theory , pages=. 2023 , organization=

    2023

  21. [22]

    , author=

    Some NP-complete problems in quadratic and nonlinear programming. , author=. Mathematical programming , volume=

  22. [23]

    IEEE Transactions on signal processing , volume=

    Sparse reconstruction by separable approximation , author=. IEEE Transactions on signal processing , volume=. 2009 , publisher=

    2009

  23. [24]

    The Annals of Statistics , volume=

    NEARLY UNBIASED VARIABLE SELECTION UNDER MINIMAX CONCAVE PENALTY , author=. The Annals of Statistics , volume=

  24. [25]

    Mechanical Systems and Signal Processing , volume=

    Sparse reconstruction for blade tip timing signal using generalized minimax-concave penalty , author=. Mechanical Systems and Signal Processing , volume=. 2021 , publisher=

    2021

  25. [26]

    Linear algebra and its applications , volume=

    Computing the numerical radius , author=. Linear algebra and its applications , volume=. 1996 , publisher=

    1996

  26. [27]

    Digital signal processing , volume=

    Vector minimax concave penalty for sparse representation , author=. Digital signal processing , volume=. 2018 , publisher=

    2018

  27. [28]

    IMA Journal of Numerical Analysis , volume=

    An algorithm for computing the numerical radius , author=. IMA Journal of Numerical Analysis , volume=. 1997 , publisher=

    1997

  28. [29]

    European Journal of Operational Research , volume=

    DC approximation approaches for sparse optimization , author=. European Journal of Operational Research , volume=. 2015 , publisher=

    2015

  29. [30]

    Journal of Convex Analysis , volume=

    Filling the gap between lower-C1 and lower-C2 functions , author=. Journal of Convex Analysis , volume=

  30. [31]

    arXiv preprint arXiv:1605.00125 , year=

    An accelerated algorithm for minimizing convex compositions , author=. arXiv preprint arXiv:1605.00125 , year=

    Pith/arXiv arXiv

  31. [32]

    Mathematical Programming , volume=

    Efficiency of minimizing compositions of convex functions and smooth maps , author=. Mathematical Programming , volume=. 2019 , publisher=

    2019

  32. [33]

    2009 , publisher=

    Variational analysis , author=. 2009 , publisher=

    2009

  33. [34]

    Variational Analysis and Generalized Differentiation. I. Basic Theory, II. Applications. , author=. 2009 , publisher=

    2009

  34. [35]

    Pacific Journal of mathematics , volume=

    Minimization of functions having Lipschitz continuous first partial derivatives , author=. Pacific Journal of mathematics , volume=. 1966 , publisher=

    1966

  35. [36]

    SIAM journal on imaging sciences , volume=

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems , author=. SIAM journal on imaging sciences , volume=. 2009 , publisher=

    2009

  36. [37]

    Mathematical Programming , volume=

    Descent methods for composite nondifferentiable optimization problems , author=. Mathematical Programming , volume=. 1985 , publisher=

    1985

  37. [38]

    International Journal of Systems Science , volume=

    A generalized proximal point algorithm for certain non-convex minimization problems , author=. International Journal of Systems Science , volume=. 1981 , publisher=

    1981

  38. [39]

    Mathematical Programming , volume=

    Composite proximal bundle method , author=. Mathematical Programming , volume=. 2013 , publisher=

    2013

  39. [40]

    IEEE Transactions on Control of Network Systems , volume=

    Dynamic regret bounds for constrained online nonconvex optimization based on polyak--lojasiewicz regions , author=. IEEE Transactions on Control of Network Systems , volume=. 2022 , publisher=

    2022

  40. [41]

    Acta mathematica vietnamica , volume=

    Convex analysis approach to DC programming: theory, algorithms and applications , author=. Acta mathematica vietnamica , volume=

  41. [42]

    Generalized differentiability/duality and optimization for problems dealing with differences of convex functions , author=. Convexity and Duality in Optimization: Proceedings of the Symposium on Convexity and Duality in Optimization Held at the University of Groningen, The Netherlands June 22, 1984 , pages=. 1985 , organization=

    1984

  42. [43]

    Applied mathematics and computation , volume=

    On the difference of two maximal monotone operators: Regularization and algorithmic approaches , author=. Applied mathematics and computation , volume=. 2008 , publisher=

    2008

  43. [44]

    SIAM Journal on Optimization , volume=

    Convergence of new inertial proximal methods for DC programming , author=. SIAM Journal on Optimization , volume=. 2008 , publisher=

    2008

  44. [45]

    Nonlinear Analysis and Optimization , pages=

    Global minimization of a difference of two convex functions , author=. Nonlinear Analysis and Optimization , pages=. 2009 , publisher=

    2009

  45. [46]

    2021 , publisher=

    Modern nonconvex nondifferentiable optimization , author=. 2021 , publisher=

    2021

  46. [47]

    Mathematics of Operations Research , volume=

    Minimizing compositions of differences-of-convex functions with smooth mappings , author=. Mathematics of Operations Research , volume=. 2024 , publisher=

    2024

  47. [48]

    Operations Research , volume=

    Wasserstein Distributionally Robust Optimization and Variation Regularization , author=. Operations Research , volume=. 2024 , publisher=

    2024

  48. [49]

    Computing

    Tian, Lai and So, Anthony Man-Cho , booktitle=. Computing. 2022 , publisher=

    2022

  49. [51]

    Mathematics of Operations Research , volume=

    Computing B-stationary points of nonsmooth DC programs , author=. Mathematics of Operations Research , volume=. 2017 , publisher=

    2017

  50. [52]

    Mathematics of Operations Research , volume=

    Penalty and augmented Lagrangian methods for constrained DC programming , author=. Mathematics of Operations Research , volume=. 2022 , publisher=

    2022

  51. [53]

    SIAM Journal on Optimization , volume=

    Proximally guided stochastic subgradient method for nonsmooth, nonconvex problems , author=. SIAM Journal on Optimization , volume=. 2019 , publisher=

    2019

  52. [54]

    SIAM Journal on optimization , volume=

    Robust stochastic approximation approach to stochastic programming , author=. SIAM Journal on optimization , volume=. 2009 , publisher=

    2009

  53. [55]

    Mathematical Programming Computation , volume=

    A stochastic approximation method for approximating the efficient frontier of chance-constrained nonlinear programs , author=. Mathematical Programming Computation , volume=. 2021 , publisher=

    2021

  54. [56]

    Computational Optimization and Applications , volume=

    Algorithms for stochastic optimization with function or expectation constraints , author=. Computational Optimization and Applications , volume=. 2020 , publisher=

    2020

  55. [57]

    Mathematics of Operations Research , volume=

    Stochastic approximation proximal method of multipliers for convex stochastic programming , author=. Mathematics of Operations Research , volume=. 2023 , publisher=

    2023

  56. [58]

    arXiv preprint arXiv:2002.05826 , year=

    Statistical learning with conditional value at risk , author=. arXiv preprint arXiv:2002.05826 , year=

    Pith/arXiv arXiv 2002

  57. [59]

    2013 , publisher=

    Perturbation analysis of optimization problems , author=. 2013 , publisher=

    2013

  58. [60]

    Set-Valued and Variational Analysis , volume=

    Nonconvex and nonsmooth approaches for affine chance-constrained stochastic programs , author=. Set-Valued and Variational Analysis , volume=. 2022 , publisher=

    2022

  59. [61]

    Statistica Sinica , pages=

    Waveshrink with firm shrinkage , author=. Statistica Sinica , pages=. 1997 , publisher=

    1997

  60. [62]

    Journal of computational and graphical statistics , volume=

    Wavelet shrinkage denoising using the non-negative garrote , author=. Journal of computational and graphical statistics , volume=. 1998 , publisher=

    1998

  61. [63]

    Minimization of functions having lipschitz continuous first partial derivatives

    Larry Armijo. Minimization of functions having lipschitz continuous first partial derivatives. Pacific Journal of mathematics, 16 0 (1): 0 1--3, 1966

    1966

  62. [64]

    A local regret in nonconvex online learning

    Sergul Aydore, Lee Dicker, and Dean Foster. A local regret in nonconvex online learning. arXiv preprint arXiv:1811.05095, 2018

    Pith/arXiv arXiv 2018

  63. [65]

    Dynamic local regret for non-convex online forecasting

    Sergul Aydore, Tianhao Zhu, and Dean P Foster. Dynamic local regret for non-convex online forecasting. Advances in neural information processing systems, 32, 2019

    2019

  64. [66]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems

    Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2 0 (1): 0 183--202, 2009

    2009

  65. [67]

    Perturbation analysis of optimization problems

    J Fr \'e d \'e ric Bonnans and Alexander Shapiro. Perturbation analysis of optimization problems. Springer Science & Business Media, 2013

    2013

  66. [68]

    X-armed bandits

    S \'e bastien Bubeck, R \'e mi Munos, Gilles Stoltz, and Csaba Szepesv \'a ri. X-armed bandits. Journal of Machine Learning Research, 12 0 (5), 2011

    2011

  67. [69]

    Descent methods for composite nondifferentiable optimization problems

    James V Burke. Descent methods for composite nondifferentiable optimization problems. Mathematical Programming, 33: 0 260--279, 1985

    1985

  68. [70]

    Dynamic regret analysis of safe distributed online optimization for convex and non-convex problems

    Ting-Jui Chang, Sapana Chaudhary, Dileep Kalathil, and Shahin Shahrampour. Dynamic regret analysis of safe distributed online optimization for convex and non-convex problems. Transactions on Machine Learning Research, 2023

    2023

  69. [71]

    Modern nonconvex nondifferentiable optimization

    Ying Cui and Jong-Shi Pang. Modern nonconvex nondifferentiable optimization. SIAM, 2021

    2021

  70. [72]

    A generalized proximal point algorithm for certain non-convex minimization problems

    Masao Fukushima and Hisashi Mine. A generalized proximal point algorithm for certain non-convex minimization problems. International Journal of Systems Science, 12 0 (8): 0 989--1000, 1981

    1981

  71. [73]

    Wavelet shrinkage denoising using the non-negative garrote

    Hong-Ye Gao. Wavelet shrinkage denoising using the non-negative garrote. Journal of computational and graphical statistics, 7 0 (4): 0 469--488, 1998

    1998

  72. [74]

    Waveshrink with firm shrinkage

    Hong-Ye Gao and Andrew G Bruce. Waveshrink with firm shrinkage. Statistica Sinica, pages 855--874, 1997

    1997

  73. [75]

    Kleywegt

    Rui Gao, Xi Chen, and Anton J. Kleywegt. Wasserstein distributionally robust optimization and variation regularization. Operations Research, 72 0 (3): 0 1177--1191, 2024. doi:10.1287/opre.2022.2383

    arXiv 2024

  74. [76]

    Online nonconvex optimization with limited instantaneous oracle feedback

    Ziwei Guan, Yi Zhou, and Yingbin Liang. Online nonconvex optimization with limited instantaneous oracle feedback. In The Thirty Sixth Annual Conference on Learning Theory, pages 3328--3355. PMLR, 2023

    2023

  75. [77]

    A sound local regret methodology for online nonconvex composite optimization, February 2025

    Nadav Hallak. A sound local regret methodology for online nonconvex composite optimization, February 2025. URL https://optimization-online.org/2025/02/a-sound-local-regret-methodology-for-online-nonconvex-composite-optimization/. Optimization Online, published February 17, 2025

    2025

  76. [78]

    Regret minimization in stochastic non-convex learning via a proximal-gradient approach

    Nadav Hallak, Panayotis Mertikopoulos, and Volkan Cevher. Regret minimization in stochastic non-convex learning via a proximal-gradient approach. In International Conference on Machine Learning, pages 4008--4017. PMLR, 2021

    2021

  77. [79]

    Efficient regret minimization in non-convex games

    Elad Hazan, Karan Singh, and Cyril Zhang. Efficient regret minimization in non-convex games. In International Conference on Machine Learning, pages 1433--1441. PMLR, 2017

    2017

  78. [80]

    Introduction to online convex optimization

    Elad Hazan et al. Introduction to online convex optimization. Foundations and Trends in Optimization , 2 0 (3-4): 0 157--325, 2016

    2016

  79. [81]

    An algorithm for computing the numerical radius

    Chunyang He and GA Watson. An algorithm for computing the numerical radius. IMA Journal of Numerical Analysis, 17 0 (3): 0 329--342, 1997

    1997

  80. [82]

    Generalized differentiability/duality and optimization for problems dealing with differences of convex functions

    J-B Hiriart-Urruty. Generalized differentiability/duality and optimization for problems dealing with differences of convex functions. In Convexity and Duality in Optimization: Proceedings of the Symposium on Convexity and Duality in Optimization Held at the University of Groningen, The Netherlands June 22, 1984, pages 37--70. Springer, 1985

    1984

Showing first 80 references.