Pith. sign in

REVIEW 2 major objections 4 minor 2 cited by

The paper claims that Nesterov acceleration carries over to strongly convex-strongly concave bilinear saddle point problems, giving an optimal first-order method and an exponentially convergent continuous-time system.

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

2026-08-04 20:54 UTC pith:3MQALFMB

load-bearing objection The main theorem is false as stated (the constant in Theorem 1 is wrong and a simple example violates it), but the rate analysis is likely repairable and the paper deserves a serious referee. the 2 major comments →

arxiv 2509.08258 v1 pith:3MQALFMB submitted 2025-09-10 math.OC

Nesterov acceleration for strongly convex-strongly concave bilinear saddle point problems: discrete and continuous-time approaches

Xin He , Ya-Ping Fang This is my paper
classification math.OC MSC 90C2549M2734D0537N40
keywords optimal first-order algorithmNesterov accelerationbilinear saddle point problemstrongly convex-strongly concavecontinuous-time primal-dual dynamiclinear convergenceLyapunov analysis
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 aims to show that Nesterov acceleration for strongly convex optimization extends to bilinear saddle point problems min_x max_y F(x)+-G(y). It proposes a discrete primal-dual gradient method whose primal-dual gap and iterate error both contract at rate 1-min{sqrt(mu_F/L_F), sqrt(mu_G/L_G)} per iteration, the same rate an accelerated method would give if the two problems were decoupled. It also builds a constant-damping continuous-time system whose gap, trajectory error, and velocity decay like e^{-min{sqrt(mu_F), sqrt(mu_G)}t}, recovering the classical Nesterov ODE when A=0. If correct, the work supplies a first-order method and an ODE that match the strong-convexity benchmark, with an explicit discretization linking the two.

Core claim

The central claim is that for problem (1), with F mu_F-strongly convex and L_F-smooth, G mu_G-strongly convex and L_G-smooth, Algorithm 1 converges linearly with contraction factor 1-min{sqrt(mu_F r), sqrt(mu_G s)} for both the primal-dual gap L(x_k,y*)-L(x*,y_k) and the squared iterate error, whenever r<=1/L_F and s<=1/L_G; with r=1/L_F, s=1/L_G this becomes the advertised rate (8). The proof builds a Lyapunov sequence E_k = gap + (1/2r)||u_k||^2 + (1/2s)||v_k||^2, with u_k and v_k momentum-adjusted errors, and shows E_{k+1} <= (1-theta)E_k. The continuous-time system (20), with constant damping 2 sqrt(mu_F), 2 sqrt(mu_G) and look-ahead coupling through y+gamma y_dot and x+gamma x_dot, has

What carries the argument

The load-bearing parameter is theta = min{sqrt(mu_F r), sqrt(mu_G s)}, the common momentum coefficient. The interpolation step (x_bar_k, y_bar_k) = (x_k, y_k) + ((1-theta)/(1+theta))[(x_k, y_k) - (x_{k-1}, y_{k-1})] combines momentum with both curvatures, while the gradient evaluations use future-step extrapolation terms (1/theta)(x_{k+1}-x_k) and (1/theta)(y_{k+1}-y_k); in Algorithm 1 this is realized by solving a linear system with matrix I + (rs/theta^2) A^T A. The Lyapunov energy E_k telescopes at factor (1-theta). In continuous time the analogue is E(t) = gamma^2 (L(x(t),y*)-L(x*,y(t))) + 1/2 ||u(t)||^2 + 1/2 ||v(t)||^2 with u = x-x* + gamma x_dot, v = y-y* + gamma y_dot, gamma = max{1/

Load-bearing premise

The load-bearing premise is that the rate (1-min{sqrt(mu_F/L_F), sqrt(mu_G/L_G)})^k is truly optimal for the coupled problem; the paper infers this from the decoupled case A=0 and from matching NAG-SC, without proving a lower bound that rules out a slower rate when the coupling is present.

What would settle it

Take the quadratic instance F(x)=(mu_F/2)||x||^2, G(y)=(mu_G/2)||y||^2, and A=cI with c much larger than L_F and L_G; run Algorithm 1 with r=1/L_F, s=1/L_G. If the measured per-iteration contraction of the primal-dual gap stays away from 1-min{sqrt(mu_F/L_F), sqrt(mu_G/L_G)} or degrades as c grows, the claimed A-independent optimal rate is false.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • For every smooth strongly convex-strongly concave bilinear saddle problem satisfying the paper's assumptions, Algorithm 1 contracts the primal-dual gap and the squared iterate error by the per-iteration factor 1-min{sqrt(mu_F/L_F), sqrt(mu_G/L_G)}.
  • When A=0 the algorithm reduces to NAG-SC applied separately to F and G, so its rate matches the known optimal rate for decoupled strongly convex minimization.
  • The continuous-time system has a unique global solution even when gradients are only locally Lipschitz, and converges linearly in the primal-dual gap, trajectory error, and velocity, with a finite weighted velocity integral.
  • A time rescaling and explicit discretization of the continuous system recover Algorithm 1, so the ODE analysis functions as the continuous limit of the discrete method rather than a separate construction.
  • The advertised contraction factor is independent of the coupling matrix A, so asymptotically the bilinear term does not slow the rate beyond the strong-convexity parameters.

Where Pith is reading between the lines

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

  • The paper labels the rate optimal on the strength of the decoupled case A=0, but it proves no lower bound for the coupled problem; if a matching lower bound were later found to depend on ||A||, the advertised A-independent rate would be unattainable in the coupled regime.
  • Because the contraction factor ignores A, the method's behavior on ill-conditioned or large-coupling instances is an open question; testing A with norm far exceeding L_F and L_G would separate the theory from the numerics.
  • The momentum coefficient theta = min{sqrt(mu_F r), sqrt(mu_G s)} suggests an adaptive variant that updates the momentum as estimates of mu_F and mu_G improve; the Lyapunov proof would then need a perturbed-energy argument.
  • The continuous system's look-ahead terms y + gamma y_dot and x + gamma x_dot resemble implicit or symplectic discretizations, so a natural test is whether Algorithm 1 is exactly a symplectic Euler discretization of (20), which would explain the sharp rate.

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 studies strongly convex-strongly concave bilinear saddle point problems (1). It proposes a Nesterov-type first-order primal-dual method (Algorithm 1) and claims the rate (1 - min{sqrt(mu_F/L_F), sqrt(mu_G/L_G)})^k for both the primal-dual gap and the iterate gap, and calls this rate optimal. It also introduces an inertial primal-dual ODE (20), proves existence and uniqueness of global solutions and linear convergence at rate O(e^{-min{sqrt(mu_F), sqrt(mu_G)}t}), and reports numerical experiments. The main analysis is Lyapunov-based in both discrete and continuous time.

Significance. The claimed A-independent accelerated linear rate is interesting and, if valid, would be a genuine improvement over existing primal-dual methods. The Lyapunov derivations are self-contained, explicit, and do not fit parameters to data, which is a strength. The continuous-time result extends classical accelerated ODEs to the strongly convex-strongly concave saddle setting. However, Theorem 1 as stated is false because the constant C is not a valid upper bound; the rate appears repairable by using the energy E_1 as the constant. The optimality claim is plausible via the A=0 lower bound, but it is not formalized. These issues are substantive and require a major revision, though the core rate argument seems salvageable.

major comments (2)
  1. [Section 2, Theorem 1] The constant C in Theorem 1(i) is not a valid upper bound. The proof establishes E_{k+1} <= (1-theta) E_k with E_k = gap_k + (1/(2r))||u_k||^2 + (1/(2s))||v_k||^2 and u_k = theta(x_k - x*) + (1-theta)(x_k - x_{k-1}). Since Algorithm 1 starts from x_1 = x_0 and y_1 = y_0, one has u_1 = theta(x_1 - x*) and v_1 = theta(y_1 - y*), so E_1 = gap_1 + theta^2/(2r)||x_1 - x*||^2 + theta^2/(2s)||y_1 - y*||^2. The stated C = gap_1 + (1/(2r))||x_1||^2 + (1/(2s))||y_1||^2 does not dominate E_1. Concretely, take F(x)=0.5(x-1)^2, G(y)=0.5y^2, L_F=L_G=4, mu_F=mu_G=1, r=s=1/4, x_1=x_0=0, y_1=0, A=0. Then theta=1/2, x*=1, y*=0, gap_1=0.5, C=0.5, and the algorithm gives x_2=0.25, y_2=0, so gap_2=0.28125 > C(1-theta)=0.25. The asymptotic rate is repairable by replacing C with E_1, but Theorem 1 as stated is false. Please restate the theorem and adjust the constant in part (ii) accordingly.
  2. [Section 1, Eq. (8) and Remark 1] The 'optimal' designation is load-bearing but is not backed by a formal lower-bound statement. The informal argument reduces to the decoupled case A=0, where NAG-SC is optimal. This is the right idea in principle, because a lower bound on a subclass is a valid lower bound for the whole class, but the paper should state this as a precise minimax lower-bound proposition, with constants, and explain the transfer from function-value lower bounds to iterate-gap lower bounds. As written, the optimality claim in the title and abstract is a promise rather than a proved theorem.
minor comments (4)
  1. [Abstract and Section 1] The word 'iterative' appears where 'iterate' is meant: 'for both the primal-dual gap and the iterative' should be 'and the iterate gap'.
  2. [Lemma 2 proof] After Eq. (14), the text says 'the last equality follows from r <= 1/L_F', but this is an inequality: a nonnegative term (1/r - L_F/2)||x_{k+1}-xbar_k||^2 is dropped. Please change 'equality' to 'inequality'.
  3. [Section 3, system (34) and discretization] The displayed discretized system contains an index/notation error: the second equation has A(x_{k+1} + gamma/√h (x_k - x_k)), which is zero in the second term, while the resulting iteration (35) uses A(x_k + 1/theta (x_{k+1}-x_k)). Please correct the display. Also, the claim that the rescaled system (34) has the same convergence rate as (20) is asserted with 'one can show' but no proof is given; since this is used to motivate the link to Algorithm 1, the derivation should either be completed or explicitly labeled as heuristic.
  4. [§1 and Remark 1] The optimality argument via A=0 is plausible but informal. Even if a formal lower-bound proposition is not added, the authors should at least state explicitly that the lower bound is for the whole class because the decoupled subclass is contained in it.

Circularity Check

0 steps flagged

No significant circularity; the proofs are self-contained Lyapunov analyses and self-citations are not load-bearing.

full rationale

The paper's central results are derived rather than assumed. Algorithm 1 is constructed from Nesterov's NAG-SC template, and its convergence rate for both the primal-dual gap and the iterate gap is obtained by a fully explicit Lyapunov argument (Theorem 1 and Lemma 2): the energy E_k is defined from the algorithm's momentum terms, the contraction E_{k+1} ≤ (1-θ)E_k is proven algebraically, and the final rate follows from θ = min{√(μF r), √(μG s)}. No parameter is fitted to data, and no target quantity is re-used as an input. The continuous-time system (20) is analyzed with its own Lyapunov function, and the O(e^{-min{√μF,√μG}t}) rate follows directly from the dissipation inequality. The claim that rate (8) is optimal is justified by the decoupled A=0 case where the problem reduces to two independent NAG-SC problems; this is a valid minimax lower-bound argument (assuming the known optimality of NAG-SC for strongly convex smooth minimization), not a circular invocation. Self-citations ([11]–[16]) appear only in the introduction, motivation, and as examples of related techniques; they do not enter the proofs or provide the core premises. There are separate correctness/rigor concerns: the constant C in Theorem 1(i) appears to be misstated (the proof actually controls E_1 involving ||x1-x*||^2 and ||y1-y*||^2 rather than the printed ||x1||^2 and ||y1||^2), and the paper does not provide an explicit lower bound for the coupled setting when claiming optimality. These are mathematical or expository issues, not circularity. Hence the circularity score is low, reflecting only the presence of minor non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

No free parameters are fitted; the algorithm uses step sizes determined by smoothness and strong convexity constants. The 'optimal rate' claim rests on an unproven lower bound, which is an ad hoc assumption.

axioms (3)
  • domain assumption F is mu_F-strongly convex and L_F-smooth; G is mu_G-strongly convex and L_G-smooth (Assumption 1)
    Used throughout the discrete-time analysis; standard for these problems.
  • ad hoc to paper The convergence rate (8) is the information-theoretic lower bound for first-order primal-dual methods on problem (1)
    The paper uses 'optimal' without proof; the rate equals the NAG-SC rate in the decoupled case, but a lower bound for the coupled case is not provided. This is load-bearing for the optimality claim.
  • domain assumption Gradients of F and G are locally Lipschitz continuous (Assumption 2)
    Used for the continuous-time existence theory; weaker than global smoothness.

pith-pipeline@v1.3.0-alltime-deepseek · 18451 in / 12469 out tokens · 108956 ms · 2026-08-04T20:54:19.413185+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Nesterov acceleration for strongly convex-strongly concave bilinear saddle point problems: discrete and continuous-time approaches." pith.science (2026). https://pith.science/paper/3MQALFMB

@misc{pith2026250908258,
  author       = {Pith},
  title        = {Pith review of: Nesterov acceleration for strongly convex-strongly concave bilinear saddle point problems: discrete and continuous-time approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3MQALFMB}},
  note         = {Machine review of arXiv:2509.08258}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In this paper, we study a bilinear saddle point problem of the form $\min_{x}\max_{y} F(x) + \langle Ax, y \rangle - G(y)$, where $F$ and $G$ are $\mu_F$- and $\mu_G$-strongly convex functions, respectively. By incorporating Nesterov acceleration for strongly convex optimization, we first propose an optimal first-order discrete primal-dual gradient algorithm. We show that it achieves the optimal convergence rate $\mathcal{O}\left(\left(1 - \min\left\{\sqrt{\frac{\mu_F}{L_F}}, \sqrt{\frac{\mu_G}{L_G}}\right\}\right)^k\right)$ for both the primal-dual gap and the iterative, where $L_F$ and $L_G$ denote the smoothness constants of $F$ and $G$, respectively. We further develop a continuous-time accelerated primal-dual dynamical system with constant damping. Using the Lyapunov analysis method, we establish the existence and uniqueness of a global solution, as well as the linear convergence rate $\mathcal{O}(e^{-\min\{\sqrt{\mu_F},\sqrt{\mu_G}\}t})$. Notably, when $A = 0$, our methods recover the classical Nesterov accelerated methods for strongly convex unconstrained problems in both discrete and continuous-time. Numerical experiments are presented to support the theoretical convergence rates.

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. Inertial Primal Dual Dynamics with Hessian-driven Damping for Saddle Point Problems

    math.OC 2026-07 conditional novelty 6.0

    New inertial primal-dual ODEs with Hessian damping achieve O(1/t²) convex rates and O(1/t^{α−1}) strongly-convex rates without knowing the strong convexity moduli.

  2. Fast primal-dual methods for convex-concave bilinear saddle point problems: continuous-time dynamics and discrete algorithms

    math.OC 2026-06 unverdicted novelty 5.0

    Proves convergence to saddle points and o(1/t²) gap rates for continuous-time dynamics with α/t damping (α≥3) and for a structure-preserving discretization under a t_k sequence condition with ρ≤1.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages · cited by 2 Pith papers · 2 internal anchors

  1. [1]

    Attouch, H., Bot ¸, R.I., Csetnek, E.R.: Fast optimization via inertia l dynamics with closed-loop damping. J. Eur. Math. Soc. 25(5), 1985–2056 (2022)

    work page 1985

  2. [2]

    Bot ¸, R.I., Csetnek, E.R.: Second order forward-backward dyn amical systems for monotone inclusion problems. SIAM J. Control Optim. 54(3), 1423–1443 (2016)

    work page 2016

  3. [3]

    Borodich, E., Kormakov, G., Kovalev, D., Beznosikov, A., Gasnikov , A.: Near-optimal algorithm with complexity separation for strongly convex-strongly concave composite saddle point problems. Int. Conf. Comput. Optim. 83 (2024)

    work page 2024

  4. [4]

    Bot ¸, R.I., Nguyen, D.K.: Improved convergence rates and traj ectory convergence for primal-dual dynamical systems with vanishing damping. J. Differentia l Equations 303, 369–406 (2021)

    work page 2021

  5. [5]

    I., Csetnek, E

    Bot ¸, R. I., Csetnek, E. R., Nguyen, D. K.: Fast augmented Lagr angian method in the convex regime with convergence guarantees for the iterates. Ma th. Program. 200(1), 147-197 (2023)

    work page 2023

  6. [6]

    Cherukuri, A., Gharesifard, B., Cort´ es, J.: Saddle-point dynam ics: conditions for asymptotic stability of saddle points. SIAM J. Control Optim. 55(1), 486–511 (2017)

    work page 2017

  7. [7]

    B., Sidford, A., Tian, K.: Relative Lipschitzness in extrag radient methods and a direct recipe for acceleration

    Cohen, M. B., Sidford, A., Tian, K.: Relative Lipschitzness in extrag radient methods and a direct recipe for acceleration. arXiv:2011.06572 (2020)

    Pith/arXiv arXiv 2011

  8. [8]

    Ding, K.W., Fliege, J., Vuong, P.T.: Fast convergence of the primal-d ual dynamical system and corresponding algorithms for a nonsmooth bilinearly cou pled saddle point problem. Comput. Optim. Appl. 90, 151–192 (2025). 20

    work page 2025

  9. [9]

    S., Chen, J., Li, L., Xiao, L., Zhou, D.: Stochastic variance re duction methods for policy evaluation

    Du, S. S., Chen, J., Li, L., Xiao, L., Zhou, D.: Stochastic variance re duction methods for policy evaluation. In: Int. Conf. Mach. Learn., pp. 1049-1058 . PMLR (2017)

    work page 2017

  10. [10]

    Optimal Extragradient-Based Bilinearly-Coupled Saddle-Point Optimization

    Du, S. S., Gidel, G., Jordan, M. I., Li, C. J.: Optimal extragradient -based bilinearly- coupled saddle-point optimization. arXiv:2206.08573 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  11. [11]

    He, X., Hu, R., Fang, Y.P.: A second order primal-dual dynamical s ystem for a convex– concave bilinear saddle point problem. Appl. Math. Optim. 89(2), Art. no. 30 (2024)

    work page 2024

  12. [12]

    Automatica 146, Art

    He, X., Hu, R., Fang, Y.P.: Fast primal-dual algorithm via dynamical system for a linearly constrained convex optimization problem. Automatica 146, Art. no. 110547 (2022)

    work page 2022

  13. [13]

    Automatica, 183, Art

    He, X., He, D., Fang, Y.P.: Accelerated primal-dual methods for s trongly convex objective functions in continuous and discrete time. Automatica, 183, Art. no.112579 (2026)

    work page 2026

  14. [14]

    He, X., Hu, R., Fang, Y.P.: Convergence rates of inertial primal-d ual dynamical meth- ods for separable convex optimization problems. SIAM J. Control O ptim. 59(5), 3278–3301 (2021)

    work page 2021

  15. [15]

    Accelerated quadratic penalty dynamic appro aches with applications to distributed optimization

    He X, Guo L, He D. Accelerated quadratic penalty dynamic appro aches with applications to distributed optimization. Neural Networks 184, 107032 (2025)

    work page 2025

  16. [16]

    Second-order primal

    He, X., Hu, R., Fang, Y.P.: “Second-order primal” + “first-order dual” dynamical sys- tems with time scaling for linear equality constrained convex optimizat ion problems. IEEE Trans. Automat. Control 67(8), 4377–4383 (2022)

    work page 2022

  17. [17]

    Kovalev, D., Gasnikov, A., Richt´ arik, P.: Accelerated primal-dua l gradient method for smooth and convex-concave saddle-point problems with bilinear coupling. Adv. Neural Inf. Process. Syst. 35, 21725-21737 (2022)

    work page 2022

  18. [18]

    arXiv:2409.18285 (2024 )

    Li, W., Zeng, X., Pavel, L.: Primal-dual accelerated mirror-desce nt method for constrained bilinear saddle-point problems. arXiv:2409.18285 (2024 )

    Pith/arXiv arXiv 2024

  19. [19]

    Nature Singapore: Springer (2020)

    Lin, Z., Li, H., Fang, C.: Accelerated optimization for machine learn ing. Nature Singapore: Springer (2020)

    work page 2020

  20. [20]

    IEEE Trans

    Liu, Q., Wang, J.: A projection neural network for constrained quadratic minimax optimization. IEEE Trans. Neural Netw. Learn. Syst. 26(11), 2891-2900 (2015)

    work page 2015

  21. [21]

    A continuous perspective on the inertial corrected primal-dual proximal splitting

    Luo, H.: A continuous perspective on the inertial corrected pr imal-dual proximal splitting. arXiv:2405.14098 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Luo, H.: A universal accelerated primal-dual method for conve x optimization prob- lems. J. Optim. Theory Appl. 201(1), 280-312 (2024). 21

    work page 2024

  23. [23]

    Luo, H., Zhang, Z.: A unified differential equation solver approac h for separable convex optimization: Splitting, acceleration and nonergodic rate. Math. Co mp. 94, 3009-3041 (2025)

    work page 2025

  24. [24]

    Luo, H., Chen, L.: From differential equation solvers to accelera ted first-order methods for convex optimization. Math. Program. 195(1), 735–781 (2022)

    work page 2022

  25. [25]

    Springer, Berlin (2004)

    Nesterov, Y.: Introductory Lectures on Convex Optimization : A Basic Course. Springer, Berlin (2004)

    work page 2004

  26. [26]

    USSR Comput

    Polyak, B.T.: Some methods of speeding up the convergence of it eration methods. USSR Comput. Math. Math. Phys. 4(5), 1–17 (1964)

    work page 1964

  27. [27]

    Cambridge Univ

    Shalev-Shwartz, S., Ben-David, S.: Understanding machine lear ning: From theory to algorithms. Cambridge Univ. Press (2014)

    work page 2014

  28. [28]

    S., Jordan, M

    Shi, B., Du, S. S., Jordan, M. I., Su, W. J.: Understanding the acc eleration phe- nomenon via high-resolution differential equations. Math. Program . 195, 79-148 (2022)

    work page 2022

  29. [29]

    Shi, B., Du, S.S., Su, W., Jordan, M.I.: Acceleration via symplectic dis cretization of high-resolution differential equations. Adv. Neural Inf. Process . Syst. 32 (2019)

    work page 2019

  30. [30]

    Su, W., Boyd, S., Cand´ es, E.J.: A differential equation for modelin g Nesterov’s accel- erated gradient method: theory and insights. J. Mach. Learn. Re s. 17(153), 1–43 (2016)

    work page 2016

  31. [31]

    arXiv:2409.05301 (2 024)

    Sun, X., He, L., Long, X.J.: Tikhonov regularized inertial primal-du al dynamics for convex-concave bilinear saddle point problems. arXiv:2409.05301 (2 024)

    Pith/arXiv arXiv

  32. [32]

    American Math- ematical Society (2012)

    Teschl, G.: Ordinary differential equations and dynamical syste ms. American Math- ematical Society (2012)

    work page 2012

  33. [33]

    K., He, N., Oh, S.: Lifted primal-dual method f or bilinearly coupled smooth minimax optimization

    Thekumparampil, K. K., He, N., Oh, S.: Lifted primal-dual method f or bilinearly coupled smooth minimax optimization. In: AISTATS, pp. 4281-4308. PMLR (2022)

    work page 2022

  34. [34]

    Wang, Y., Li, J.: Improved algorithms for convex-concave minima x optimization. Adv. Neural Inf. Process. Syst. 33, 4800-4810 (2020)

    work page 2020

  35. [35]

    Wibisono, A., Wilson, A.C., Jordan, M.I.: A variational perspective o n accelerated methods in optimization. Proc. Natl. Acad. Sci. 113(46), E7351–E7358 (2016)

    work page 2016

  36. [36]

    Wilson, A.C., Recht, B., Jordan, M.I.: A Lyapunov analysis of accele rated methods in optimization. J. Mach. Learn. Res. 22(113), 1–34 (2021)

    work page 2021

  37. [37]

    Xie, G., Luo, L., Lian, Y., Zhang, Z.: Lower complexity bounds for fi nite-sum convex- concave minimax optimization problems. In: Int. Conf. Mach. Learn ., pp. 10504-10513. PMLR (2020). 22

    work page 2020

  38. [38]

    IF AC-Papers OnLine 53(2), 7362–7367 (2020)

    Zeng, X., Lei, J., Chen, J.: Accelerated first-order continuous -time algorithm for solving convex-concave bilinear saddle point problem. IF AC-Papers OnLine 53(2), 7362–7367 (2020)

    work page 2020

  39. [39]

    IEEE Trans

    Zeng, X., Lei, J., Chen, J.: Dynamical primal-dual accelerated me thod with appli- cations to network optimization. IEEE Trans. Automat. Control 68(3), 1760–1767 (2023)

    work page 2023

  40. [40]

    B.: Near-optimal lo cal convergence of alternating gradient descent-ascent for minimax optimization

    Zhang, G., Wang, Y., Lessard, L., Grosse, R. B.: Near-optimal lo cal convergence of alternating gradient descent-ascent for minimax optimization. I n: AISTATS, pp. 7659-7679. PMLR (2022)

    work page 2022

  41. [41]

    Zhao, Y., Liao, X., He, X., Zhou, M., Li, C.: Accelerated primal-dual mirror dynamics for centralized and distributed constrained convex optimization pr oblems. J. Mach. Learn. Res. 24(343), 1–59 (2023). 23

    work page 2023