Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Accelerating Diagonal Methods for Bilevel Optimization: Unified Convergence via Continuous-Time Dynamics

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proves that two diagonal bilevel algorithms reach little-o inner rates and weak convergence to a solution under Hölderian growth or the Attouch–Czarnecki condition.

desk verdict A genuinely strong unified Lyapunov analysis for accelerated bilevel diagonal methods, with a real but patchable formulation gap in the assumptions and one missing factor in a proof. read the letter →

arxiv 2505.14389 v1 pith:T5VGE7MU submitted 2025-05-20 math.OC

Radu Ioan Boţ , Enis Chenchene , Ernö Robert Csetnek , David Alexander Hulett This is my paper · ORCID
classification math.OC MSC 90C2591A6590C9937L0565K1065K0546N10
keywords bileveloptimizationTikhonovregularizationNesterovaccelerationproximal-gradientmethodLyapunovanalysisHölderiangrowthAttouch–Czarneckiconditionweakconvergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper analyzes two diagonal (Tikhonov-regularized) methods for bilevel problems of the form $\min\{H(x): x\in\operatorname{argmin} F\}$, where the outer objective $H$ is minimized over the minimizer set of a composite convex inner function $F$. The central claim is that, when the regularization parameter $\varepsilon_k=c/(k+\beta)^\delta$ is tuned to one of two critical regimes, both the plain proximal-gradient method and its Nesterov-accelerated variant achieve little-o decay of the inner residual ($o(k^{-1})$ and $o(k^{-2})$, respectively), drive the outer objective to its constrained minimum, and make the iterates converge weakly to a solution of the bilevel problem in an infinite-dimensional Hilbert space. The conditions needed are Hölderian growth of the inner function for $\delta\in(\eta/\rho^*, \eta)$, or the Attouch–Czarnecki summability condition at the critical value $\delta=\eta$, where $\eta=1$ for Algorithm 1 and $\eta=2$ for Algorithm 2. This matters because it upgrades earlier big-O bounds to little-o rates, extends the analysis to composite and infinite-dimensional settings, and makes explicit the trade-off between inner geometry and regularization decay.

What carries the argument

The load-bearing object is a discrete Lyapunov energy of the form $$E^\lambda_k := t_k^\eta(\Psi_{k-1}(x_k)-\Psi_{k-1}(x_*)) + V_k,$$ where $\Psi_k=F+\varepsilon_k H$ is the regularized inner objective, $t_k=\theta(k+\gamma)$, and $V_k$ is a kinetic term that is $\frac{\lambda}{2}\|x_k-x_*\|^2$ for Algorithm 1 and a momentum-based term for Algorithm 2. The key identity is the dissipativity estimate (15), which says that the energy can only decrease, up to a compensated term proportional to the outer residual; the sign and summability of the compensation are controlled by $\zeta_{k,\delta}$. Lemma A.7 is the engine that converts this one inequality into big-O and little-o rates, summability of the residuals, existence of energy limits, and, through Opial's lemma, weak convergence of the iterates. The continuous-time Lyapunov functions in Appendix B have the same structure, which is why the discrete and continuous analyses are essentially equivalent.

What would settle it

Run Algorithm 2 on a composite bilevel problem in which $f$ is quadratically growing, $\hat f$ is an indicator of a set, and $\operatorname{argmin} F$ is strictly contained in $\operatorname{argmin} f$; if $F(x_k)-\min F$ fails to be $o(k^{-2})$ or the iterates fail to converge weakly to a point in $\operatorname{argmin} F$, then the theorem as printed is false.

Watch

Extended reading notes

Core claim

The paper's discovery is that both the gradient and accelerated gradient diagonal schemes are governed by one dissipativity inequality, $$E_{k+1}-E_k+\zeta_{k,\delta}(H(x_k)-H(x_*))+g_k\leq -C_2 $k^{{\eta-1}}$(F(x_k)-F(x_*)),$$ with $\zeta_{k,\delta}$ of order $k^{\eta-1}\varepsilon_k$. From this single inequality, the auxiliary Lemma A.7 extracts three regimes: $\delta>\eta$ gives only $O(k^{-\eta})$ inner decay with no guarantee of convergence to a bilevel solution; $\delta=\eta$ together with the Attouch–Czarnecki condition gives $o(k^{-\eta})$ inner decay, convergence of the outer objective to its constrained minimum, and weak convergence of the whole sequence; and $\eta/\rho^*<\delta<\eta$ together with $\rho$-Hölderian growth gives the same little-o inner rate, an $o(k^{\delta-\eta})$ outer residual, and weak convergence. Theorems 2.3 and 3.3 state these results for $\eta=1$ and $\eta=2$, and Appendix B proves the analogous continuous-time statements for the first- and second-order regularization dynamics that guided the algorithms. The argument hinges on translating the geometry of $F$ into a summability condition on the conjugate function $(F-\min F)^*$ and then letting the Lyapunov inequality convert that summability into rates and convergence.

Load-bearing premise

The load-bearing premise is that the stated Hölderian or Attouch–Czarnecki condition holds for the full inner function $F$ and not only for its smooth component $f$, because the proofs need residuals measured with respect to $\operatorname{argmin} F$ while the printed assumptions are written with $\operatorname{argmin} f$.

Editorial extensions

If this is right

  • At $\delta=\eta$ with the Attouch–Czarnecki condition, Algorithm 1 satisfies $F(x_k)-\min F=o(k^{-1})$ and Algorithm 2 satisfies $F(x_k)-\min F=o(k^{-2})$; both sequences converge weakly to a solution of the bilevel problem and $H(x_k)$ tends to the constrained minimum.
  • In the Hölderian regime $\eta/\rho^*<\delta<\eta$, the same little-o inner rates hold and the outer residual decays as $o(k^{\delta-\eta})$, with weak convergence of the iterates.
  • The accelerated scheme with $\eta=2$ contains the recent FBi-PG method as a special case, so its convergence results improve those of that method rather than only matching them.
  • Every discrete result has a continuous-time counterpart for the gradient system (3) and the second-order system (4), so the rates transfer to trajectories of those dynamics.
  • When $\delta>\eta$, only big-O inner rates are guaranteed and there is no claim of convergence to a bilevel solution, showing that the critical regimes are precisely where acceleration pays off.

Reading between the lines

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

  • Beyond the paper, the same Lyapunov inequality suggests that the critical choice $\delta=\eta$ is not a technical artefact but the smallest regularization decay that still guarantees convergence to a bilevel solution, so practitioners balancing inner and outer accuracy should choose the critical value rather than larger $\delta$.
  • A natural extension is to feed the dissipativity inequality with inexact or stochastic gradient oracles; since the proof only uses $L$-smoothness and summability of $I_k^r$, the rates should survive with additional noise-summability assumptions.
  • One editorial reading, flagged in the manuscript itself by the mismatch between assumptions stated for $f$ and theorems applied to $F$, is that a corrected statement should write Assumptions 1 and 2 for the full inner function $F$; under that reading the composite case is coherent.
  • In finite-dimensional settings, the weak convergence results plus coercivity bounds likely upgrade to strong convergence or convergence of distances to the solution set, a direction the paper only touches in Section 4.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper studies diagonal (Tikhonov-type) methods for convex bilevel programs of the form min H(x) subject to x in argmin F, in real Hilbert spaces. It proposes two algorithms: a proximal-gradient diagonal method (Algorithm 1) and a Nesterov-accelerated variant (Algorithm 2), both designed as discretizations of the continuous-time systems (3) and (4). The main results, Theorems 2.3 and 3.3, establish O(k^{-1}) and O(k^{-2}) inner-function rates when the regularization parameter decays faster than the critical rate, and little-o rates of the same orders, together with weak convergence to a solution of the bilevel problem, under either a Hölderian growth condition on the inner function (Assumption 1) or the Attouch–Czarnecki condition (Assumption 2). The convergence analysis is unified through a single discrete Lyapunov energy and a general dissipativity lemma (Lemma A.7), and continuous-time counterparts are given in Appendix B. Numerical experiments on a quadratic benchmark and a logistic-regression bilevel problem illustrate the behavior of the algorithms.

Significance. If the technical gaps identified below are repaired, this is a useful and substantial contribution. The unified Lyapunov framework centered on Lemma A.7 is clean and appears to be the right vehicle for treating both first- and second-order diagonal schemes under one set of geometric conditions. The little-o rates improve the big-O rates of Merchav–Sabach–Teboulle [26] and extend the analysis to infinite-dimensional Hilbert spaces, which is a genuine step beyond the current literature. The paper also gives explicit credit to prior work and ships reproducible code for the numerical experiments, which is commendable. The central claims are derived from dissipativity estimates rather than assumed target rates, and I see no circularity in the convergence arguments. The main obstacles are formal correctness issues in the statement and use of Assumptions 1 and 2; these are fixable but currently affect load-bearing parts of the proofs.

major comments (2)
  1. [§1.3, Lemma A.1, Theorems 2.3 and 3.3]
  2. [Proofs of Theorems 2.3 and 3.3; Lemma A.7, condition (C2)]
minor comments (4)
  1. [Theorem 3.3 proof, Eq. (53)] The left-hand side of Eq. (53) appears to have a sign error: it should read H(x_{k_j}) − H(x*) rather than H(x_{k_j}) + H(x*), since q_k is defined with H(x_k) − H(x*).
  2. [Section 5.3, Algorithm 1 parameters] In the experimental setup, the text says "For Algorithm 1 we set ... s = 1.95/L_f", but Algorithm 1 uses a step size denoted by θ, not s; the parameter s is introduced for Algorithm 2. Please use consistent notation.
  3. [Lemma 3.1, Eq. (37)] The proof derives the dissipativity estimate with a factor t_k on the right-hand side and then replaces t_k by θk using t_k ≥ θk; this is correct, but the replacement should be stated explicitly at Eq. (37) to avoid the appearance of a new factor k without explanation.
  4. [Appendix B and general typography] There are several small typographical issues: "Furtheremore" in Theorem B.3, "W eakening" in the Section 4 heading, and a few garbled author-name marks in the references (e.g., [9]). These should be cleaned up in revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all convergence rates are derived from explicit Lyapunov dissipativity estimates under independently stated geometric conditions, with no fitted parameter renamed as a prediction and no load-bearing self-citation.

full rationale

The paper's central claims are the little-o rates and weak convergence for Algorithms 1 and 2. These are obtained from a single discrete Lyapunov energy (Eqs. (18) and (36)) and the dissipativity estimates in Lemmas 2.1 and 3.1, which reduce to inequality (15)/(80) in Lemma A.7. The geometric inputs, Assumptions 1 and 2 (Eqs. (8) and (9)), concern the inner function's growth or the summability of a conjugate expression involving the regularization parameters; they are stated independently of the algorithm's iterates and of the target rates. No parameter in the proof is fitted to the reported rates, and no rate is assumed to reach a target. The use of prior works is non-circular: the Attouch-Czarnecki condition is attributed to [4], the Lyapunov technique follows [3], and the comparison baseline [26] is an external method whose rates are improved, not presupposed. The paper's self-citations [7,8] appear only in motivational remarks about Tikhonov methods and are not used to justify the main theorems. The one substantive issue is a statement mismatch: Assumptions 1 and 2 in Section 1.3 are phrased for the smooth component f with argmin f, while Lemma A.1 and the proofs of Theorems 2.3 and 3.3 require the conditions on the full inner function F so that dist(x, argmin F) and (F - min F)^* appear. The theorems themselves condition on 'F satisfies Assumption 1/2', indicating the intended corrected reading; this is a presentation gap in the assumptions rather than a case of a result being equivalent to its inputs by construction. Under the intended reading the derivations are self-contained. Therefore the circularity score is 0.

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

The central results rest on geometric conditions on the inner function F (Holderian growth or Attouch-Czarnecki), the qualification condition, and the polynomial schedule for epsilon_k. No free parameters are fitted to data; the algorithmic hyperparameters are user choices inside the theorem's allowed ranges. No new postulated entities are introduced.

assumptions (6)
  • domain assumption Holderian growth on the inner function F (as used in the proofs, despite the f in the printed Assumption 1)
    Eq. (8) with dist(x,argmin F) and F(x)-min F; needed for Lemma A.1 and conditions (C1)-(C2).
  • domain assumption Attouch-Czarnecki summability condition on the inner function F (as used, despite the f in Assumption 2)
    Eq. (9) with (F-min F)* and sigma_argmin F; needed for the delta=eta critical case.
  • domain assumption Qualification condition Assumption 3
    Eq. (12), needed to produce p* in N_argmin F(x*) with -p* in dH(x*), used in (13).
  • domain assumption Polynomial regularization decay epsilon_k = c/(k+beta)^delta
    Eq. (5); all rates are expressed in terms of delta, and the analysis is restricted to this parameterization.
  • domain assumption Well-posedness of (1): argmin F nonempty, solution set nonempty, inf H > -infinity
    Stated before (1); required for the Lyapunov functions and limit values.
  • standard math Standard convex analysis facts (Fenchel-Moreau, prox calculus, Opial's lemma, Stolz-Cesaro)
    Used throughout the appendix proofs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerating Diagonal Methods for Bilevel Optimization: Unified Convergence via Continuous-Time Dynamics." pith.science (2026). https://pith.science/paper/T5VGE7MU

@misc{pith2026250514389,
  author       = {Pith},
  title        = {Pith review of: Accelerating Diagonal Methods for Bilevel Optimization: Unified Convergence via Continuous-Time Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T5VGE7MU}},
  note         = {Machine review of arXiv:2505.14389}
}
read the original abstract

We analyze fast diagonal methods for simple bilevel programs. Guided by the analysis of the corresponding continuous-time dynamics, we provide a unified convergence analysis under general geometric conditions, including H\"olderian growth and the Attouch-Czarnecki condition. Our results yield explicit convergence rates and guarantee weak convergence to a solution of the bilevel problem. In particular, we improve and extend recent results on accelerated schemes, offering novel insights into the trade-offs between geometry, regularization decay, and algorithmic design. Numerical experiments illustrate the advantages of more flexible methods and support our theoretical findings.

Figures

Figures reproduced from arXiv: 2505.14389 by the authors.

Figure 1
Figure 1. Geometric setting considered in this paper, where [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Result of the experiment in Subsection 5.2: Comparing inner and outer function residuals as well as distance to solution for the methods listed in Subsection 5.1. 101 102 103 Iteration number (k) 10−8 10−6 10−4 10−2 100 F(x k)−min F (a) Inner objective residual 101 102 103 Iteration number (k) 10−2 10−1 100 101 102 103 |H(x k)− H(x ∗)| 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 Value of δ (b) Outer objective residual [PIT… view at source ↗
Figure 3
Figure 3. Results of experiment in Section 5.2: Comparing residual decrease for Algorithm 2 as a function of the iteration number emphasizing the choice of δ ∈ (1, 2). Although the objective of this paper is not to compete with existing algorithms, we can observe that the improved flexibility of Algorithm 2 plays a crucial role in performance, yielding results that clearly surpass those of FBi-PG. In this example, specificall… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Results of experiment in Subsection 5.3. The dotted lines in Figure 4a represent the rates O(k −δ/2 ) and O(k −δ ), with δ = 1.9. the standard sigmoid function. Due to overparametrization, among the optimal classifiers, we seek the one that minimizes the ℓ 1 norm by so…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. The iterates of FISTA converge even under inexact computations and stochastic gradients

    math.OC 2025-11 conditional novelty 6.0 of 10

    Weak (and almost-sure weak) convergence of inexact and stochastically noised FISTA iterates in infinite-dimensional Hilbert spaces, with O(1/t_k^2)-type objective rates.

Reference graph

Works this paper leans on

39 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [26]

    Dynamic FISTA for Convex Composite Bi-Level Optimization

    R. Merchav, S. Sabach, and M. Teboulle. “A fast algorithm for convex composite bi-Level opti- mization”. In: arXiv preprint, arXiv:2407.21221 (2024)

    work page Pith review arXiv 2024

  2. [23]

    On the convergence of proximal gradient methods for convex simple bilevel optimization

    P. Latafat, A. Themelis, S. Villa, and P. Patrinos. “On the convergence of proximal gradient methods for convex simple bilevel optimization”. In: Journal of Optimization Theory and Appli- cations 204 (51) (2025). 31

    work page 2025

  3. [1]

    Asymptotic development by Γ-convergence

    G. Anzellotti and S. Baldo. “Asymptotic development by Γ-convergence”. In: Applied Mathe- matics and Optimization 27 (2) (1993), pp. 105–123

    work page 1993

  4. [2]

    Viscosity solutions of minimization problems

    H. Attouch. “Viscosity solutions of minimization problems”. In: SIAM Journal on Optimization 6 (3) (1996), pp. 769–806

    work page 1996

  5. [3]

    Convergence of iterates for first-order opti- mization algorithms with inertia and Hessian driven damping

    H. Attouch, Z. Chbani, J. Fadili, and H. Riahi. “Convergence of iterates for first-order opti- mization algorithms with inertia and Hessian driven damping”. In: Optimization 72 (5) (2023), pp. 1199–1238

    work page 2023

  6. [4]

    Asymptotic behavior of coupled dynamical systems with multiscale aspects

    H. Attouch and M.-O. Czarnecki. “Asymptotic behavior of coupled dynamical systems with multiscale aspects”. In: Journal of Differential Equations 248 (6) (2010), pp. 1315–1344

    work page 2010

  7. [5]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems

    A. Beck and M. Teboulle. “A fast iterative shrinkage-thresholding algorithm for linear inverse problems”. In: SIAM Journal on Imaging Sciences 2 (1) (2009), pp. 183–202

    work page 2009

  8. [6]

    Coresets via bilevel optimization for continual learning and streaming

    Z. Borsos, M. Mutny, and A. Krause. “Coresets via bilevel optimization for continual learning and streaming”. In: Advances in Neural Information Processing Systems . Ed. by H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin. Vol. 33. Curran Associates, Inc., 2020, pp. 14879– 14890. 30

    work page 2020

Show all 39 references
  1. [7]

    Generalized Fast Krasnoselskii–Mann method with preconditioners

    R. I. Bot ¸, E. Chenchene, and J. M. Fadili. “Generalized Fast Krasnoselskii–Mann method with preconditioners”. In: arXiv preprint, arXiv:2411.18574 (2024)

    2024

  2. [8]

    Extra-Gradient method with flexible anchoring: strong convergence and fast residual decay

    R. I. Bot ¸ and E. Chenchene. “Extra-Gradient method with flexible anchoring: strong convergence and fast residual decay”. In: arXiv preprint, arXiv:2410.14369 (2024)

    2024 arXiv

  3. [9]

    Fast optimistic gradient descent ascent (OGDA) method in continuous and discrete time

    R. I. Bot ,, E. R. Csetnek, and D.-K. Nguyen. “Fast optimistic gradient descent ascent (OGDA) method in continuous and discrete time”. In: Foundations of Computational Mathematics 25 (2025), pp. 163–222

    2025

  4. [10]

    Proximal point algorithm controlled by a slowly vanishing term: applications to hierarchical minimization

    A. Cabot. “Proximal point algorithm controlled by a slowly vanishing term: applications to hierarchical minimization”. In: SIAM Journal on Optimization 15 (2) (2005), pp. 555–572

    2005

  5. [11]

    An accelerated gradient method for con- vex smooth simple bilevel optimization

    J. Cao, R. Jiang, E. Y. Hamedani, and A. Mokhtari. “An accelerated gradient method for con- vex smooth simple bilevel optimization”. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2024

    2024

  6. [12]

    A hybrid proximal generalized conditional gradi- ent method and application to total variation parameter learning

    E. Chenchene, A. Hosseini, and K. Bredies. “A hybrid proximal generalized conditional gradi- ent method and application to total variation parameter learning”. In: 2023 European Control Conference (ECC). 2023, pp. 1–6

    2023

  7. [13]

    Asymptotic analysis of the exponential penalty trajectory in linear programming

    R. Cominetti and J. San Mart ´ ın. “Asymptotic analysis of the exponential penalty trajectory in linear programming”. In: Mathematical Programming 67 (1) (1994), pp. 169–187

    1994

  8. [14]

    Viscosity solutions of Hamilton–Jacobi equations

    M. G. Crandall and P.-L. Lions. “Viscosity solutions of Hamilton–Jacobi equations”. In: Trans- actions of the American Mathematical Society 277 (1) (1983), pp. 1–42

    1983

  9. [15]

    Sinkhorn distances: lightspeed computation of optimal transport

    M. Cuturi. “Sinkhorn distances: lightspeed computation of optimal transport”. In: Advances in Neural Information Processing Systems. Ed. by C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger. Vol. 26. Curran Associates, Inc., 2013

    2013

  10. [16]

    Bilevel Optimization

    S. Dempe and A. Zemkoho. “Bilevel Optimization”. In: Springer Optimization and its Applica- tions. Vol. 161. Springer, 2020

    2020

  11. [17]

    Methodology and first-order algorithms for solving nonsmooth and non- strongly convex bilevel optimization problems

    L. Doron and S. Shtern. “Methodology and first-order algorithms for solving nonsmooth and non- strongly convex bilevel optimization problems”. In: Mathematical Programming 201 (1) (2023), pp. 521–558

    2023

  12. [18]

    Ekeland and R

    I. Ekeland and R. Temam. Analyse Convexe et Probl` emes Variationnels. Etudes Math´ ematiques. Dunod, 1974

    1974

  13. [19]

    Exact regularization of convex programs

    M. P. Friedlander and P. Tseng. “Exact regularization of convex programs”. In: SIAM Journal on Optimization 18 (4) (2008), pp. 1326–1350

    2008

  14. [20]

    New proximal point algorithms for convex minimization

    O. G¨ uler. “New proximal point algorithms for convex minimization”. In: SIAM Journal on Op- timization 2 (4) (1992), pp. 649–664

    1992

  15. [21]

    A method with convergence rates for optimization prob- lems with variational inequality constraints

    H. D. Kaushik and F. Yousefian. “A method with convergence rates for optimization prob- lems with variational inequality constraints”. In: SIAM Journal on Optimization 31 (3) (2021), pp. 2171–2198

    2021

  16. [22]

    A bilevel optimization approach for parameter learning in variational models

    K. Kunisch and T. Pock. “A bilevel optimization approach for parameter learning in variational models”. In: SIAM Journal on Imaging Sciences 6 (2) (2013), pp. 938–983

    2013

  17. [24]

    On the convergence rate of the Halpern-iteration

    F. Lieder. “On the convergence rate of the Halpern-iteration”. In: Optimization Letters 15 (2) (2021), pp. 405–418

    2021

  18. [25]

    Convex bi-level optimization problems with nonsmooth outer ob- jective function

    R. Merchav and S. Sabach. “Convex bi-level optimization problems with nonsmooth outer ob- jective function”. In: SIAM Journal on Optimization 33 (4) (2023), pp. 3114–3142

    2023

  19. [27]

    A. S. Nemirovskij and D. B. Yudin. Problem Complexity and Method Efficiency in Optimization . A Wiley-Interscience publication. Wiley, 1983

    1983

  20. [28]

    A method for solving the convex programming problem with convergence rate O(1/k2)

    Y. Nesterov. “A method for solving the convex programming problem with convergence rate O(1/k2)”. In: Proceedings of the USSR Academy of Sciences 269 (1983), pp. 543–547

    1983

  21. [29]

    Weak convergence of the sequence of successive approximations for nonexpansive mappings

    Z. Opial. “Weak convergence of the sequence of successive approximations for nonexpansive mappings”. In: Bulletin of the American Mathematical Society 73 (1967), pp. 591–597

    1967

  22. [30]

    Scikit-learn: machine learning in Python

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. “Scikit-learn: machine learning in Python”. In:Journal of machine learning research 12 (Oct) (2011), pp. 2825–2830

    2011

  23. [31]

    Coupling the gradient method with a general exterior penalization scheme for convex minimization

    J. Peypouquet. “Coupling the gradient method with a general exterior penalization scheme for convex minimization”. In: Journal of Optimization Theory and Applications 153 (2012), pp. 123– 138

    2012

  24. [32]

    Meta-learning with implicit gradients

    A. Rajeswaran, C. Finn, S. M. Kakade, and S. Levine. “Meta-learning with implicit gradients”. In: Advances in Neural Information Processing Systems . Ed. by H. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alch´ e-Buc, E. Fox, and R. Garnett. Vol. 32. Curran Associates, Inc., 2019

    2019

  25. [33]

    A first order method for solving convex bilevel optimization problems

    S. Sabach and S. Shtern. “A first order method for solving convex bilevel optimization problems”. In: SIAM Journal on Optimization 27 (2) (2017), pp. 640–660

    2017

  26. [34]

    An explicit descent method for bilevel convex optimization

    M. Solodov. “An explicit descent method for bilevel convex optimization”. In: Journal of Convex Analysis 14 (2) (2007), p. 227

    2007

  27. [35]

    A bundle method for a class of bilevel nonsmooth convex minimization prob- lems

    M. V. Solodov. “A bundle method for a class of bilevel nonsmooth convex minimization prob- lems”. In: SIAM Journal on Optimization 18 (1) (2007), pp. 242–259

    2007

  28. [36]

    A differential equation for modeling Nesterov’s accelerated gradient method: theory and insights

    W. Su, S. Boyd, and E. J. Cand` es. “A differential equation for modeling Nesterov’s accelerated gradient method: theory and insights”. In:Journal of Machine Learning Research 17 (153) (2016), pp. 1–43

    2016

  29. [37]

    On the solution of incorrectly formulated problems and the regularization method

    A. N. Tikhonov. “On the solution of incorrectly formulated problems and the regularization method”. In: Doklady Akademii Nauk SSSR 151 (1963). In Russian, pp. 501–504

    1963

  30. [38]

    Wolberg, O

    W. Wolberg, O. Mangasarian, N. Street, and W. Street. Breast Cancer Wisconsin (Diagnostic). UCI Machine Learning Repository. doi.org/10.24432/C5DW2B. 1993

    1993 doi

  31. [39]

    Accelerated algorithms for smooth convex-concave minimax problems with O(1/k2) rate on squared gradient norm

    T. Yoon and E. K. Ryu. “Accelerated algorithms for smooth convex-concave minimax problems with O(1/k2) rate on squared gradient norm”. In: Proceedings of the 38th International Confer- ence on Machine Learning . Vol. 139. Proceedings of Machine Learning Research. PMLR, July 20...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.