Pith. sign in

REVIEW 2 major objections 6 minor 38 references

Stochastic Moving Anchor Algorithms and a Popov's Scheme with Moving Anchor

T0 review · 2 major / 6 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 Stochastic moving-anchor methods keep their accelerated convergence rate in smooth convex-concave minimax problems, provided the oracle noise decays quickly enough along the run.

desk verdict A legitimate stochastic extension of moving-anchor EAG with a real, fixable proof gap in the main theorem; the Popov half is honestly labeled heuristic. read the letter →

arxiv 2506.07290 v1 pith:M3YO26AF submitted 2025-06-08 math.OC

James Alcala , Yat Tin Chow , Mahesh Sunkula This is my paper
classification math.OC MSC 90C4790C2565K05
keywords movinganchorstochasticextragradientsaddlepointproblemsminimaxoptimizationsquaredgradientnormHalpernanchoringPopov'sschemeacceleratedconvergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether the accelerated convergence of moving-anchor extragradient methods survives when the gradient oracle becomes stochastic. It answers yes for smooth convex-concave saddle point (minimax) problems, provided the stochastic noise along the run decays as the fourth power of the iteration count. The main result, Theorem 4.9, bounds the squared norm of the saddle operator by a constant over the product $(k+1)(k+2)$, matching the order-optimal $O(1/k^2)$ rate known for the deterministic method, with an explicit extra term that accumulates the variance. This matters because large-scale machine-learning and game-theoretic problems rely on stochastic oracles for efficiency, and the moving anchor had previously only known accelerated guarantees in deterministic settings. The paper also proposes a moving-anchor version of Popov's scheme, which needs only one operator evaluation per iteration, with numerics indicating the same accelerated behavior.

What carries the argument

The carrying mechanism is the moving anchor: an auxiliary sequence $\bar{z}_k$ is updated as $\bar{z}_{k+1}=\bar{z}_k+\gamma_{k+1}G(z_{k+1})$, and every iterate is pulled toward the anchor by a $1/(k+2)$ drift, generalizing the fixed anchor $z_0$ used in earlier anchoring methods. The analysis runs through a stochastic Lyapunov functional $$V_k = A_k\|G(z_k)\|^2 + B_k\langle G(z_k), z_k-\bar{z}_k\rangle + c_k\|z^*-\bar{z}_k\|^2,$$ whose expected descent is controlled by variance terms rather than forced to be nonpositive. Lemma 4.4 bounds the expected drop from below by a negative multiple of the oracle variances at $z_k$ and $z_{k+1/2}$; under Condition 4.1 those variances are summable, so the supermartingale convergence theorem yields almost-sure convergence and the rate bound follows by unpacking $V_k$.

What would settle it

Run stochastic moving-anchor EAG-V on a finite-sum convex-concave problem in which one component gradient is nonzero at the saddle point; Condition 4.1 then fails because the relevant noise term cannot decay as $k^{-4}$ once the iterates approach the solution. Record $\|G(z_k)\|^2$: if it still decreases like $1/k^2$, the condition is unnecessary for the rate, while a plateau or slowdown would show the condition is doing the work the proof assigns it.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the stochastic moving-anchor EAG-V iterations (4.5)--(4.8), using independent unbiased gradient estimates and the variance-aware stepsize (4.8), converge almost surely in the Lyapunov sense and satisfy $$\|G(z_k)\|^2 \leq \frac{4\big((\alpha_0 $R^{2}$+c_0)\|z_0-z^*\|^2 + \mathrm{sum}(k-1)\big)}{\alpha_\infty (k+1)(k+2)}$$ whenever $c_\infty \geq 1/\alpha_\infty$. The rate is $O(1/k^2)$ on the squared gradient norm at the last iterate, matching the deterministic moving-anchor rate and preserving the constant-factor speedups observed numerically. The extra term $\mathrm{sum}(k-1)$ accumulates the oracle variances along the path and is provably summable under the paper's variance condition, so it does not change the order of the rate. This is, to the authors' knowledge, the first accelerated anchored stochastic rate in the moving-anchor setting.

Load-bearing premise

The load-bearing premise is that the stochastic oracle's noise shrinks like the fourth power of the iteration count at every visited point; the theorem's own $1/k^2$ conclusion does not imply that, and it fails for finite-sum oracles with noise that persists at the solution.

Editorial extensions

If this is right

  • If Theorem 4.9 is correct, stochastic moving-anchor EAG-V achieves a last-iterate squared-gradient-norm rate of $O(1/k^2)$ for smooth convex-concave saddle problems, matching the best known deterministic anchored rate despite using random gradient estimates.
  • The moving anchor retains its practical advantage over fixed anchoring in the stochastic regime: the numerical comparisons in Section 6 show the moving-anchor variants ahead by a constant factor, consistent with the deterministic findings.
  • The summable-variance machinery gives a template for stochastic analyses of other anchored methods: control the Lyapunov descent by the oracle variance and require that variance to decay along the path.
  • The proposed moving-anchor Popov's scheme, if its conjectured theory materializes, would give an accelerated method with a single operator evaluation per iteration, an efficiency improvement over the two-evaluation EAG family.
  • Condition 4.1 ties the rate to the oracle's behavior on the specific stochastic path; in practice the paper's experiments suggest that setting the condition-number factor to one improves speed substantially in well-conditioned regimes.

Reading between the lines

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

  • A natural next step the paper leaves implicit is to replace Condition 4.1 with a variance-reduction mechanism such as control variates that forces the noise to decay along the path, which would make the accelerated rate valid for finite-sum oracles with noise at the solution.
  • The observed sign flip in the numerical experiments—negative anchor step winning for EAG-V and positive anchor step winning for the earlier FEG work—suggests the preferred anchor direction is tied to the operator-evaluation structure, a hypothesis one could test by swapping the two update formulas.
  • Because Condition 4.1 is not implied by the theorem's own conclusion, a demanding reader can test the theory by measuring the noise term along actual runs; if it fails to decay like $k^{-4}$, the bound's variance sum is not covered and a different analysis would be needed.
  • The moving-anchor Popov schemes, with descent directions taken from either the last iterate or the extrapolator, hint at a general principle: anchoring acceleration may tolerate a wide family of anchor direction choices, which would generalize both fixed- and moving-anchor theories.
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 / 6 minor

Summary. The paper introduces stochastic moving-anchor variants of the EAG-V extragradient algorithm for smooth convex-concave saddle point problems. The central theoretical contribution is Theorem 4.9, which claims a rate ||G(z_k)||^2 = O(1/k^2) for the stochastic moving-anchor EAG-V iterations under an unbiased stochastic oracle and a variance-decay Condition 4.1. The proof proceeds through a stochastic Lyapunov analysis: Lemma 4.4 establishes a conditional-expectation drift bound, Lemma 4.7 shows summability of variance terms, and Theorem 4.8 gives almost-sure convergence of the Lyapunov functional. The paper also proposes two moving-anchor Popov schemes, for which no convergence theorem is proved, and presents numerical experiments comparing stochastic moving-anchor methods with fixed-anchor counterparts and the moving-anchor Popov schemes.

Significance. If the main theorem were valid, it would be one of the first stochastic accelerated anchored rates for saddle point problems, extending the deterministic moving-anchor framework of [1] to oracle noise. The paper is honest about the limitations of the Popov section, explicitly stating that its convergence theory is incomplete. The numerical experiments provide a useful initial comparison, and the Lyapunov analysis is a natural extension of prior anchored-method analyses. However, the central result as stated is not proven: the proof of Theorem 4.9 converts a conditional-expectation inequality into a pathwise inequality, and Condition 4.1 is substantially stronger than the theorem's own conclusion. The in-expectation version appears recoverable, but the current claim is overstated.

major comments (2)
  1. [§4, proof of Theorem 4.9] The first displayed chain of the proof writes V_k ≤ V_{k-1} + ... ≤ V_0 + sum(k−1), citing (4.50). But (4.50) is the conditional-expectation inequality E[V_{k+1} | \bar z_k, z_k] ≤ V_k + W_k, which does not imply the pathwise inequality V_k ≤ V_{k−1} + W_{k−1}. A nonnegative supermartingale can exceed its initial value with positive probability, so the chain leading to the deterministic-looking rate bound on ||G(z_k)||^2 is unjustified. The argument does support an in-expectation bound E[V_k] ≤ V_0 + Σ_{j<k} E[W_j] under the summability supplied by Lemma 4.7, hence E||G(z_k)||^2 = O(1/k^2). Theorem 4.9 and the abstract should be restated in expectation form, or an additional pathwise argument must be supplied.
  2. [§4, Condition 4.1] Condition 4.1 requires (C_G(z_k)−1)||G(z_k)||^2 ≤ C_1/(k+1)^4 and the analogous conditional expectation at half-steps. This is much stronger than the O(1/k^2) gradient-norm bound that Theorem 4.9 claims to prove. For a finite-sum oracle with sample-wise noise at the solution, C_G(z)−1 is bounded below by a positive constant, so Condition 4.1 would force ||G(z_k)||^2 = O(k^{−4}) along the stochastic path — a rate that the theorem does not deliver and that is not established by the Lyapunov analysis. Since Lemma 4.7's summability argument relies directly on Condition 4.1, the main theorem is conditional on an assumption that is not shown to hold for any concrete oracle class and is not implied by the theorem's own conclusion. The authors should either prove Condition 4.1 for a natural class of stochastic oracles or replace it with a weaker condition and adjust the claimed rate.
minor comments (6)
  1. [§3.2, Lemmas 3.1 and 4.2] Lemma 3.1's proof is deferred to Lemma 4.2, but Lemma 4.2 is stated with a different stepsize range and in the stochastic setting; moreover, the proof of Lemma 4.2 reuses the symbol η for the expression in (4.10), shadowing the η defined in the lemma statement, and the 'without loss of generality' reduction to R=1 and K_G=1 is not explained for K_G because the recurrence (3.5) does not involve K_G.
  2. [§6.1, numerical experiments] The theory requires K_G = N C_G in the stepsize definition (4.8), but the experiments set K_G = 1 for substantial numerical improvement; the paper should comment on this discrepancy between the theoretical requirement and the numerical practice.
  3. [§4, Theorem 4.9 statement] The theorem states a rate for ||G(z_k)||^2 without specifying the mode of convergence (almost sure, in expectation, or in probability); given the proof's reliance on conditional expectations, the authors should explicitly state which mode is claimed.
  4. [§4, equations (4.43)-(4.44)] The inequalities in (4.43) and (4.44) are correct but not fully explained; the roles of β_k < 1 and the positivity of (1−α_k^2R^2−β_k) should be stated explicitly so the reader can follow the sign of the dropped terms.
  5. [§4, equations (4.19)-(4.20)] The notation K_G(zk+1) appears in the denominator of (4.19) while the stepsize definition (4.8) uses a constant K_G; please clarify whether K_G is evaluated at specific points or is a global bound, and ensure the notation is consistent.
  6. [§6, figures] The figure captions do not identify which curve corresponds to which variant (positive versus negative γ_k, fixed versus moving anchor), and the experiments do not report multiple seeds or error bars, which would strengthen the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the stochastic rate follows from a self-contained Lyapunov argument, with prior-work citations providing context rather than load-bearing support.

full rationale

The central claim (Theorem 4.9) is not derived by assuming its own conclusion. The stochastic moving-anchor EAG-V algorithm is analyzed through Lemma 4.4, which establishes a conditional-expectation Lyapunov drift bound with explicit variance terms: E[V_k - V_{k+1} | \bar z_k, z_k] is bounded below by a negative combination of Var(z_k) and E[Var(z_{k+1/2}) | \bar z_k, z_k]. This is a genuine stochastic estimate, not a restatement of the convergence rate. Condition 4.1 controls the oracle's relative variance decay, and Lemma 4.7 uses it only to make the accumulated variance terms summable; the theorem's O(1/k^2) bound is then extracted from the lower bound V_k >= (alpha_infty/4)(k+1)(k+2)||G(z_k)||^2 via monotonicity and Young's inequality. The displayed rate is therefore not equivalent to any input assumption. The paper does cite the authors' own prior preprint [1] for the deterministic moving-anchor lemmas and the sequence recurrences, but the stochastic proof does not invoke the conclusion of that deterministic theorem; the self-citations are contextual, not load-bearing. A separate note: the proof of Theorem 4.9 applies the conditional inequality (4.50) pathwise, whereas (4.50) is an in-expectation statement; this is a correctness concern in the write-up, but it is not a circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central stochastic rate is supported by standard convex-concave smoothness and unbiased-oracle assumptions plus two bespoke additions: the deterministic moving-anchor machinery of [1] (same authors) and Condition 4.1, which assumes variance decays at k^{-4}. The constants alpha_0, c_0, delta_k, and K_G are free algorithmic choices; K_G is used as 1 in the numerics despite the theory defining it as N*C_G.

free parameters (4)
  • alpha_0 (initial stepsize) = in experiments: 0.9*(3/4)/(R*sqrt(K_G))
    Initial stepsize must lie in (0, eta); the whole sequence alpha_k and the rate constant depend on it. It is chosen by hand, not fitted to data.
  • c_0 (initial Lyapunov weight) = in experiments: 1.01*(4/3)*e^(pi^2/6)*R*sqrt(K_G)
    Initial Lyapunov weight, constrained by c_infty >= 1/alpha_infty; affects the constant in Theorem 4.9.
  • delta_k sequence = unspecified, chosen so sum log(1+delta_k) < infinity
    Defines c_k and gamma_k; no canonical choice is given, and c_infty depends on it.
  • K_G (noise-normalization parameter) = set to 1 in experiments; theory defines K_G=N*C_G
    Noise-normalization parameter in the stepsize and anchor update. The experiments deliberately use K_G=1 despite the theoretical definition, and the authors say this improved results.
assumptions (5)
  • domain assumption Deterministic moving-anchor Lyapunov lemmas from [1] (Lemma 3.2 and Theorem 3.4) are correct.
    The stochastic analysis in Section 4 imports the deterministic Lyapunov descent, the recurrences for A_k, B_k, alpha_k, and the O(1/k^2) deterministic rate directly from the same authors' previous preprint [1] without reproving them.
  • ad hoc to paper Condition 4.1: (C_G(z_k)-1)||G(z_k)||^2 <= C_1/(k+1)^4 and E[(C_G(z_{k+1/2})-1)||G(z_{k+1/2})||^2] <= C_2/(k+1)^4.
    This is the load-bearing noise-control assumption. It makes the drift W_k summable in Lemma 4.7, but is not implied by bounded variance or by the theorem's own O(1/k^2) gradient-norm conclusion.
  • domain assumption The oracle is unbiased and componentwise R_i-Lipschitz with R^2 = sum R_i^2.
    Standard stochastic oracle model stated in Section 4 conditions (1) and (3); used throughout the variance estimates in Lemma 4.4.
  • ad hoc to paper c_0 and {delta_k} can be chosen so that c_infty >= 1/alpha_infty while preserving the Lyapunov relations.
    Theorem 4.9 requires c_infty >= 1/alpha_infty; the paper asserts freedom in this choice but does not prove existence for general instances.
  • standard math Supermartingale convergence theorem [7] applies to the Lyapunov process.
    Used in Theorem 4.8 to conclude almost-sure convergence of V_k.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stochastic Moving Anchor Algorithms and a Popov's Scheme with Moving Anchor." pith.science (2026). https://pith.science/paper/M3YO26AF

@misc{pith2026250607290,
  author       = {Pith},
  title        = {Pith review of: Stochastic Moving Anchor Algorithms and a Popov's Scheme with Moving Anchor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3YO26AF}},
  note         = {Machine review of arXiv:2506.07290}
}
read the original abstract

Since their introduction, anchoring methods in extragradient-type saddlepoint problems have inspired a flurry of research due to their ability to provide order-optimal rates of accelerated convergence in very general problem settings. Such guarantees are especially important as researchers consider problems in artificial intelligence (AI) and machine learning (ML), where large problem sizes demand immense computational power. Much of the more recent works explore theoretical aspects of this new acceleration framework, connecting it to existing methods and order-optimal convergence rates from the literature. However, in practice introducing stochastic oracles allows for more computational efficiency given the size of many modern optimization problems. To this end, this work provides the moving anchor variants [1] of the original anchoring algorithms [36] with stochastic implementations and robust analyses to bridge the gap from deterministic to stochastic algorithm settings. In particular, we demonstrate that an accelerated convergence rate theory for stochastic oracles also exists for our moving anchor scheme, itself a generalization of the original fixed anchor algorithms, and provide numerical results that validate our theoretical findings. We also develop a tentative moving anchor Popov scheme based on the work in [33], with promising numerical results pointing towards an as-of-yet uncovered general convergence theory for such methods.

Figures

Figures reproduced from arXiv: 2506.07290 by the authors.

Figure 1
Figure 1. Comparison of the grad-norm squared of three stochastic EAG￾V variants of interest on a toy ‘almost bilinear’ problem. In [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the grad-norm squared of three stochastic EAG￾V anchoring variants on a nonlinear game. In [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the fixed anchor Popov scheme versus two moving anchor variants with a descending anchor. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [1]

    James K Alcala, Yat Tin Chow, and Mahesh Sunkula,Moving anchor extragradient methods for smooth structured minimax problems, arXiv preprint arXiv:2308.12359 (2023)

    arXiv 2023

  2. [20]

    Sucheol Lee and Donghwan Kim,Fast extra gradient methods for smooth structured nonconvex- nonconcave minimax problems, (2021)

    work page 2021

  3. [2]

    2863–2873

    Wa ¨ ıss Azizian, Ioannis Mitliagkas, Simon Lacoste-Julien, and Gauthier Gidel,A tight and unified analysis of gradient-based methods for a whole spectrum of differentiable games, International conference on artificial intelligence and statistics, PMLR, 2020, pp. 2863–2873

    work page 2020

  4. [3]

    2482–2508

    Raef Bassily, Crist´ obal Guzm´ an, and Michael Menart,Differentially private algorithms for the stochastic saddle point problem with optimal rates for the strong gap, The Thirty Sixth Annual Conference on Learning Theory, PMLR, 2023, pp. 2482–2508

    work page 2023

  5. [4]

    Yang Cai, Argyris Oikonomou, and Weiqiang Zheng,Tight last-iterate convergence of the extragradient and the optimistic gradient descent-ascent algorithm for constrained monotone variational inequalities, 2022

    work page 2022

  6. [5]

    Tatjana Chavdarova, Gauthier Gidel, Fran¸ cois Fleuret, and Simon Lacoste-Julien,Reducing noise in gan training with variance reduced extragradient, Advances in Neural Information Processing Systems 32(2019)

    work page 2019

  7. [6]

    4, 1779–1814

    Yunmei Chen, Guanghui Lan, and Yuyuan Ouyang,Optimal primal-dual methods for a class of saddle point problems, SIAM Journal on Optimization24(2014), no. 4, 1779–1814

    2014

  8. [7]

    2, 1221–1248

    Patrick L Combettes and Jean-Christophe Pesquet,Stochastic quasi-fej´ er block-coordinate fixed point iterations with random sweeping, SIAM Journal on Optimization25(2015), no. 2, 1221–1248

    work page 2015

Show all 38 references
  1. [8]

    Damek Davis and Wotao Yin,A three-operator splitting scheme and its optimization applications, Set- valued and variational analysis25(2017), 829–858

    2017

  2. [9]

    1428– 1451

    Jelena Diakonikolas,Halpern iteration for near-optimal and parameter-free monotone inclusion and strong solutions to variational inequalities, Conference on Learning Theory, PMLR, 2020, pp. 1428– 1451

    2020

  3. [10]

    1049– 1058

    Simon S Du, Jianshu Chen, Lihong Li, Lin Xiao, and Dengyong Zhou,Stochastic variance reduction methods for policy evaluation, International Conference on Machine Learning, PMLR, 2017, pp. 1049– 1058. 20

    2017

  4. [11]

    Noah Golowich, Sarath Pattathil, Constantinos Daskalakis, and Asuman Ozdaglar,Last iterate is slower than averaged iterate in smooth convex-concave saddle point problems, 2020

    2020

  5. [12]

    Tom´ as Gonz´ alez, Crist´ obal Guzm´ an, and Courtney Paquette,Mirror descent algorithms with nearly dimension-independent rates for differentially-private stochastic saddle-point problems, arXiv preprint arXiv:2403.02912 (2024)

    2024

  6. [13]

    Ian J Goodfellow, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, and Jean Pouget-Abadie,Generative adversarial nets, Advances in neural information processing systems27(2014), 2672–2680

    2014

  7. [14]

    7865–7901

    Eduard Gorbunov, Hugo Berard, Gauthier Gidel, and Nicolas Loizou,Stochastic extragradient: General analysis and improved rates, International Conference on Artificial Intelligence and Statistics, PMLR, 2022, pp. 7865–7901

    2022

  8. [15]

    11614–11641

    Eduard Gorbunov, Adrien Taylor, Samuel Horv´ ath, and Gauthier Gidel,Convergence of proximal point and extragradient-based methods beyond monotonicity: the case of negative comonotonicity, International Conference on Machine Learning, PMLR, 2023, pp. 11614–11641

    2023

  9. [16]

    6, 957–961

    Benjamin Halpern,Fixed points of nonexpanding maps, Bulletin of the American Mathematical Society 73(1967), no. 6, 957–961

    1967

  10. [17]

    Wallach, H

    Yu-Guan Hsieh, Franck Iutzeler, J´ erˆ ome Malick, and Panayotis Mertikopoulos,On the convergence of single-call stochastic extra-gradient methods, Advances in Neural Information Processing Systems (H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch´ e-Buc, E. Fox, and R. Ga...

    2019

  11. [18]

    Galina M Korpelevich,The extragradient method for finding saddle points and other problems, Matecon 12(1976), 747–756

    1976

  12. [19]

    Ryu,Accelerating value iteration with anchoring, Thirty-seventh Conference on Neural Information Processing Systems, 2023

    Jongmin Lee and Ernest K. Ryu,Accelerating value iteration with anchoring, Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  13. [21]

    Sucheol Lee and Donghwan Kim,Semi-anchored multi-step gradient descent ascent method for structured nonconvex-nonconcave composite minimax problems, arXiv preprint arXiv:2105.15042 (2021)

    2021 arXiv

  14. [22]

    9793–9826

    Chris Junchi Li, Yaodong Yu, Nicolas Loizou, Gauthier Gidel, Yi Ma, Nicolas Le Roux, and Michael Jor- dan,On the convergence of stochastic extragradient for bilinear games using restarted iteration averaging, International Conference on Artificial Intelligence and Statistics, ...

    2022

  15. [23]

    2, 405–418

    Felix Lieder,On the convergence rate of the halpern-iteration, Optimization letters15(2021), no. 2, 405–418

    2021

  16. [24]

    Mingrui Liu, Youssef Mroueh, Jerret Ross, Wei Zhang, Xiaodong Cui, Payel Das, and Tianbao Yang,To- wards better understanding of adaptive gradient algorithms in generative adversarial nets, arXiv preprint arXiv:1912.11940 (2019)

    2019 arXiv

  17. [25]

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu,Towards deep learning models resistant to adversarial attacks, arXiv preprint arXiv:1706.06083 (2017)

    2017 arXiv

  18. [26]

    Malitsky,Projected reflected gradient methods for monotone variational inequalities, SIAM Journal on Optimization25(2015), no

    Yu. Malitsky,Projected reflected gradient methods for monotone variational inequalities, SIAM Journal on Optimization25(2015), no. 1, 502–520

    2015

  19. [27]

    269, 1983, pp

    Yurii Nesterov,A method for unconstrained convex minimization problem with the rate of convergence o (1/kˆ 2), Doklady an ussr, vol. 269, 1983, pp. 543–547

    1983

  20. [28]

    Levon Nurbekyan, Siting Liu, and Yat Tin Chow,Monotone inclusion methods for a class of second-order non-potential mean-field games, 2024

    2024

  21. [29]

    Leonid Denisovich Popov,A modification of the arrow-hurwicz method for search of saddle points, Math- ematical notes of the Academy of Sciences of the USSR28(1980), 845–848

    1980

  22. [30]

    Ernest K Ryu, Kun Yuan, and Wotao Yin,Ode analysis of stochastic gradient methods with optimism and anchoring for minimax problems and gans, (2019)

    2019

  23. [31]

    Suh, Jisun Park, and Ernest K

    Jaewook J. Suh, Jisun Park, and Ernest K. Ryu,Continuous-time analysis of anchor acceleration, Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  24. [32]

    Quoc Tran-Dinh,The connection between nesterov’s accelerated methods and halpern fixed-point itera- tions, arXiv preprint arXiv:2203.04869 (2022)

    2022 arXiv

  25. [33]

    Quoc Tran-Dinh and Yang Luo,Halpern-type accelerated and splitting algorithms for monotone inclu- sions, arXiv preprint arXiv:2110.08150 (2021)

    2021 arXiv

  26. [34]

    3, 1943–1967

    Killian Wood and Emiliano Dall’Anese,Stochastic saddle point problems with decision-dependent distri- butions, SIAM Journal on Optimization33(2023), no. 3, 1943–1967

    2023

  27. [35]

    Suh, and Ernest K

    Taeho Yoon, Jaeyeon Kim, Jaewook J. Suh, and Ernest K. Ryu,Optimal acceleration for minimax and fixed-point problems is not unique, Proceedings of the 41st International Conference on Machine Learning (Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria O...

    2024

  28. [36]

    Taeho Yoon and Ernest K Ryu,Accelerated algorithms for smooth convex-concave minimax problems with o(1/k2)rate on squared gradient norm, Proceedings of the 38th International Conference on Machine Learning139(2021), 12098–12109

    2021

  29. [37]

    Ryu,Accelerated minimax algorithms flock together, SIAM Journal on Op- timization35(2025), no

    TaeHo Yoon and Ernest K. Ryu,Accelerated minimax algorithms flock together, SIAM Journal on Op- timization35(2025), no. 1, 180–209

    2025

  30. [38]

    Angela Yuan, Chris Junchi Li, Gauthier Gidel, Michael Jordan, Quanquan Gu, and Simon S Du,Optimal extragradient-based algorithms for stochastic variational inequalities with separable structure, Advances in Neural Information Processing Systems36(2023), 33338–33351. Department...

    2023

Pith tools

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