Universal Reduced-Operator Method and High-Order Global Curvature Bounds

Nikita Doikov; Yurii Nesterov

REVIEW 2 major objections 5 minor 28 references

Universal Reduced-Operator Method and High-Order Global Curvature Bounds

T0 review · 2 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A global curvature bound for nonlinear operators allows a parameter-free high-order method for composite variational inequalities whose worst-case complexity is set by the curvature and the domain's diameter.

desk verdict A serious, sound paper that extends the GCB framework to nonlinear operators and gives a universal high-order method for composite VIs; the one real gap is that Assumption 2 needs a stronger regularity condition at zero, which is easy to fix. read the letter →

arxiv 2511.07341 v2 pith:UWUP2XSP submitted 2025-11-10 math.OC

Nikita Doikov , Yurii Nesterov This is my paper

classification math.OC MSC 90C2590C4768Q25

keywords NonlinearOperatorsRiemannianGeometryGeodesicsVariationalInequalitiesGlobalCurvatureBoundComplexityHigh-OrderMethodsUniversal

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 introduces a global curvature bound (GCB) for arbitrary nonlinear operators between metric spaces and proves that, whenever the target space has nonpositive curvature, the bound grows at least quadratically under rescaling. This single scaling inequality is enough to construct a Universal Reduced-Operator Method of any order p ≥ 1 for composite variational inequalities—a class that covers convex minimization and min-max problems. The method is parameter-free: it never needs to know the operator's smoothness class, and its worst-case iteration count is expressed only through the inverse of the smoothed GCB and the domain's diameter. Under a standard power-law smoothness condition on the p-th derivative, these counts match the best known rates for such problems. The paper establishes this with a new reduced-operator step whose subproblem can be made monotone and efficiently solvable.

What carries the argument

The load-bearing object is the Global Curvature Bound κ_A(r), measuring how far A(γ(t)) sits from a geodesic in the image space, maximized over geodesics of length ≤ r. Theorem 1's inequality κ_A(βr) ≥ β² κ_A(r) is what gives all later bounds their power. Integral smoothing turns GCBs of high derivatives into σ̂_q(r), convex functions that bound Taylor approximation errors. The Reduced Operator step subtracts the Taylor model and a double regularization — quadratic plus power ‖y−x‖^(p+2) — leaving a residual whose norm and progress are controlled by Lemma 3; choosing the regularization through condition (4.4) gives a universal backtracking that adapts to the unseen σ̂_{p-1}.

What would settle it

On a positively curved surface such as a sphere with geodesic distance, take A as the identity map and compute κ_A(r) and κ_A(r/2) for small r; if κ_A(r/2) < κ_A(r)/4, Theorem 1's key inequality fails in a concrete case, invalidating the downstream bounds. For the method itself, on a bounded monotone problem with known diameter, simulate p = 1 and compare the required iterations to the bound (4.16); exceeding it would contradict Lemma 3.

Watch

Extended reading notes

Core claim

On its own terms, the central claim is that the global curvature bound κ_A(r) — the worst-case deviation of an operator A from geodesic interpolation over pairs at distance ≤ r — obeys κ_A(βr) ≥ β² κ_A(r) when the target space is nonpositively curved. Using integral smoothing of the GCBs of derivatives D^qV, the paper bounds the Taylor error of V and of its derivative by convex functions σ̂_q(r). With these, the Universal Reduced-Operator Method is shown, in Theorem 2, to terminate after at most (4/5)^(2/(p+1)) [R_0/c_p · max{1/σ̂_{p-1}^{-1}(δ/5), (5M_0/(2δ))^(1/(p+1))}]² iterations to reach δ-accuracy in the reduced-operator norm, and an analogous bound for ε-accuracy in the merit-function

Load-bearing premise

The target metric space must be nonpositively curved — otherwise the quadratic-growth inequality κ_A(βr) ≥ β² κ_A(r) can fail, and the paper's complexity bounds rest on that inequality.

Editorial extensions

If this is right

For p = 1 the method uses only the Jacobian of V and a convex subproblem, yielding a parameter-free extra-gradient-type algorithm for composite variational inequalities, including min-max problems.
On problem classes where D^p V has a power-law smoothness bound, the iteration counts reduce to the optimal rates known for monotone variational inequalities, but the method never receives those parameters.
The method offers both a stopping rule based on the norm of the reduced operator and a certified accuracy bound on the merit function, needing only the domain diameter as external information.
When V is the gradient of a convex function, the construction gives a universal high-order tensor method for composite minimization whose complexity is governed by curvature of derivatives rather than a prescribed smoothness class.

Reading between the lines

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

Editorial: because the norm-based guarantee in Theorem 2 uses only the weak-solution condition (4.15), the method may be applicable to non-monotone operators that admit a weak solution; the merit-function guarantee is the part that requires monotonicity.
Editorial: the GCB formalism should apply to operators built from sums of components with different power-law smoothness, which in general fit no single smoothness class; this is a testable regime where a fixed-parameter method would be suboptimal.
Editorial: Lemma 4 turns the subproblem monotone for p ≥ 2 under a condition numerically equivalent to the progress condition, so a practical high-order implementation hinges on an inner solver that respects condition (4.7); constructing such a solver is the paper's stated open direction.

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.

Desk Editor's Note

read the letter

Short version: this is a serious and mostly sound paper. The genuinely new pieces are Theorem 1, the quadratic-growth result for GCB of nonlinear operators between metric spaces, and the high-order integral smoothing machinery in Lemmas 1–2. The universal method (4.14) with complexity expressed directly through the inverse GCB is a real advance over prior reduced-gradient and GCB-for-functions work in [20,22]. I checked the main chain Theorem 1 → Lemma 3 → Theorem 2. The algebra in Lemma 3 is correct, and the telescoping in Theorem 2 is standard. Under Hölder smoothness the rates match [20], which is expected; the contribution is that the method needs no problem-class parameters and the analysis is driven by a single curvature object.

Now the soft spots, in order of severity. First, the stress-test note is half right. Assumption 2 only requires strict monotonicity of σ̂_{p-1}. Strict monotonicity does not rule out a right-limit c > 0 at zero. If that happens, then for δ < 5c condition (4.4) is unsatisfiable, the M-doubling loop in Step 1 never stops, and Theorem 2 is vacuous for that operator. The specific example in the note does not actually produce such a jump—for a differentiable V with bounded derivative, κ_V(r) is O(r)—but the structural concern is legitimate. The fix is trivial: add an explicit condition that σ̂_{p-1}(r) → 0 as r → 0, or simply require that (4.4) holds for the given δ. Second, the abstract's "fastest universal rate" claim is not supported by a lower-bound theorem in the paper. What is proven is a rate guarantee, and Corollary 3 recovers known optimal Hölder rates. Softening that wording is advisable. Third, for p ≥ 2 the subproblem (4.2) has no concrete solution algorithm; Section 5 gives a monotonicity condition but leaves implementation open. That is a real limitation but does not undercut the upper bounds. Fourth, the Busemann/non-positive-curvature assumption on the target space is essential for Theorem 1, so the advertised generality over arbitrary metric spaces is narrower than it sounds. The main examples all satisfy it.

Bottom line: the technical core is solid, the new GCB-for-operators concept is worth taking seriously, and the paper deserves a serious referee. A minor revision should tighten Assumption 2 and temper the optimality claim. I would cite it and would happily bring it to a reading group.

Referee Report

2 major / 5 minor

Summary. The paper introduces a Global Curvature Bound (GCB) for nonlinear operators between geodesic metric spaces and uses it to design a universal reduced-operator method of arbitrary order p≥1 for composite variational inequalities. The main result, Theorem 2, gives complexity bounds in terms of the inverse smoothed GCB σ̂_{p−1}^{-1}, and Corollary 3 recovers the known optimal rates under Hölder continuity of the p-th derivative. The framework generalizes the GCB idea from [22] and extends it to higher-order reduced-gradient methods.

Significance. If the technical assumptions are tightened, this is a valuable contribution. The abstract GCB concept and the quadratic-growth property in Theorem 1 are elegant and genuinely parameter-free: the method's complexity is expressed through an intrinsic curvature object rather than through Lipschitz/Hölder constants. The paper also gives a clean recovery of known optimal rates in Corollary 3, and the p=1 case is practically relevant because the subproblem is then a standard monotone VI. The proofs are detailed and mostly consistent. The main concern is a formal gap in Assumption 2 that affects the validity of the main theorem as stated.

major comments (2)

[Assumption 2 / Theorem 2, Eqs. (4.16)–(4.17)] Assumption 2 only requires σ̂_{p−1} to be strictly monotone. This does not imply σ̂_{p−1}(r)→0 as r→0, nor that σ̂_{p−1}^{-1}(δ/5) is finite. If σ̂_{p−1} has a jump at 0 (e.g., σ̂(0)=0 and σ̂(r)=c+r for r>0), then for δ<5c condition (4.4) has no solution, Lemma 3 cannot be invoked, the M-doubling loop in Step 1 of (4.14) may run forever, and (4.16) contains an undefined 1/σ̂^{-1}(δ/5). This is not just cosmetic: for p=1, σ̂_0=κ_V and no integral smoothing is available. Please add an explicit condition that σ̂_{p−1}(r)→0 as r→0+, or equivalently that (4.4) has a finite solution M, or prove that this follows from the GCB definition for differentiable monotone operators. Note that the concrete example V(x)=x+∫_0^x sin(1/t)dt has bounded derivative and hence κ_V(r)=O(r), so it does not by itself produce the jump; the issue is that the formal assumption is overbroad.
[Section 5, Eq. (5.1) and Algorithm (4.14)] For p≥2, subproblem (4.2) is guaranteed monotone only under the stronger condition (5.1), while the progress condition (4.4) used in Lemma 3 is weaker. Step 1 of the algorithm stops doubling M as soon as the progress condition or the small-norm condition holds, so it may stop at an M that does not satisfy (5.1). The statement 'by Lemma 3 each iteration is well defined' does not address solvability or uniqueness of (4.2) for such M, nor the computational task of finding a solution to a possibly non-monotone VI during the search phase. For p=1 this is harmless, because that subproblem is always strongly monotone. For p≥2, the main theorem is stated for an idealized algorithm whose inner steps are not fully specified. Either restrict the formal complexity claim to p=1, or modify the algorithm so that M is increased until both (4.4) and (5.1) hold, or fold the inexact-solver discussion from

minor comments (5)

[Section 4 heading] Typo: 'Variational Ineqiality' should be 'Variational Inequality'.
[Remark 6] Typo: 'monent' should be 'moment'.
[Assumption 2] The sentence introducing σ^{-1}_{p−1} should use the hat consistently: σ̂^{-1}_{p−1}.
[Theorem 1 proof] The reduction for β≤1/2 via β=β̄^k is terse; a short sentence specifying that β̄ is chosen in (1/2,1) and k is an integer would improve readability.
[References] Reference [22] is cited as '2025' with no venue; if it is an unpublished manuscript, this should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main derivation is self-contained and the inverse-GCB bounds are genuine complexity bounds, not fitted predictions.

full rationale

The paper's derivation chain is not circular. The Global Curvature Bound is defined independently as a geodesic-deviation measure (Eqs. (2.4)-(2.5)), and the key quadratic-growth inequality (Theorem 1, Eq. (2.7)) is proved under Assumption 1 rather than imported from the authors' earlier work. The smoothed high-order curvature characteristics sigma-hat_q are constructed from the operator's derivatives via Eq. (3.4), and Lemma 1 derives the Taylor-error bounds from that definition. The Universal Reduced-Operator Method (4.14) does not receive sigma-hat as an input: Step 1 adaptively doubles M until either the progress certificate (4.7) or the small-residual condition holds. Theorem 2's bound in terms of the inverse sigma-hat_{p-1}^{-1} is therefore a conventional worst-case complexity bound expressed through a problem-dependent curvature characteristic, not a prediction derived from a fitted quantity. The references to the authors' prior papers [20,22] provide context and the original GCB idea, but the load-bearing lemmas are proven in this paper. The statement in Corollary 3 that the rates 'recover the bounds' from [20] is a consistency check, not an assumption of the theorem. The limitations noted in the manuscript—possible difficulty satisfying condition (4.4) for pathological strictly monotone sigma-hat, and the open implementation question for p>=2 in Section 5—are correctness/completeness concerns, not circularity. No step in the derivation reduces to its own input by construction.

Assumptions & free parameters 1 free parameters · 7 assumptions · 1 invented entities

The central claim rests on the geometric non-positive-curvature assumption, on differentiability of V up to order p, on strict monotonicity of the smoothed GCB, and on monotonicity of V for the ε-bound. These are stated in the paper. The only hand-chosen algorithmic parameter is M0, whose effect on the complexity appears explicitly. No physical entities or fitted constants are introduced.

free parameters (1)

M0 (initial regularization multiplier) = user-specified; Remark 3 suggests M0 = 2ε/5 (1/c_p)^{p+1} D^p
Appears in the complexity bounds (4.16)/(4.17) through the max term (5M0/(2δ))^{1/(p+1)}; a poor choice degrades the bound. It is chosen by hand, not fitted to data.

assumptions (7)

domain assumption Target metric space N satisfies Assumption 1: distance between any two geodesics is convex (Busemann / non-positive curvature)
Used in the proof of Theorem 1 (Eq. (2.11)) and inherited by all smoothed-GCB inequalities (Lemma 2, Corollaries 1–2).
domain assumption σ̂_{p-1} is well-defined and strictly monotone (Assumption 2)
Needed to define σ̂_{p-1}^{-1} in Theorem 2 and to obtain the strict contradiction in Lemma 3 via (4.4)–(4.10).
domain assumption V is p-times differentiable on Q with well-defined curvatures κ_{D^{p-1}V}
The Taylor bounds (Lemma 1) and subproblem (4.2) require derivatives up to order p.
domain assumption V is monotone (4.13) for the ε-accuracy bound (4.17)
Used to prove Δ_k ≥ μ(bar x_k) and to relate the reduced-operator norm to the merit function.
domain assumption There exists a weak solution x⋆ satisfying (4.15)
Needed in Theorem 2 to conclude ⟨Vψ(x_i), x_i−x⋆⟩ ≥ 0 in the telescoping sum (4.18).
domain assumption Domain Q is bounded with known D = diam Q for certificate construction
Section 4 states that an auxiliary constraint ∥x∥ ≤ D can be added; the certificates Δ_k require maximizing over Q.
standard math Standard Taylor theorem with integral remainder, triangle inequality, and norm convexity
Used throughout Sections 3–4 without further justification.

invented entities (1)

Global Curvature Bound κ_A(r) for nonlinear operators between metric spaces
purpose: Quantifies how far A is from being geodesic-affine; used to state and prove universal complexity bounds without a priori smoothness assumptions.
New mathematical object defined in (2.4)–(2.5); its properties are proved in the paper, but it has no falsifiable handle outside the framework.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Universal Reduced-Operator Method and High-Order Global Curvature Bounds." pith.science (2026). https://pith.science/paper/UWUP2XSP

@misc{pith2026251107341,
  author       = {Pith},
  title        = {Pith review of: Universal Reduced-Operator Method and High-Order Global Curvature Bounds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWUP2XSP}},
  note         = {Machine review of arXiv:2511.07341}
}

abstract

In this paper, we develop a new concept of Global Curvature Bound (GCB) for an arbitrary nonlinear operator between abstract metric spaces. We use this notion to characterize the global complexity of high-order algorithms solving composite variational problems, which include convex minimization and min-max problems. We develop the new universal Reduced-Operator Method, which automatically achieves the fastest universal rate within our class, while our analysis does not need any specific assumptions about smoothness of the target nonlinear operator. Every step of our universal method of order $p \geq 1$ requires access to the $p$-th order derivative of the operator and the solution of a strictly monotone doubly regularized subproblem. For $p = 1$, this corresponds to computing the standard Jacobian matrix of the operator and solving a simple monotone subproblem, which can be handled using different methods of Convex Optimization. All our results are consequences of the new theorem on the quadratic growth of GCB for general nonlinear operators.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 3 linked inside Pith

[20]

High-order reduced-gradient methods for composite variational inequali- ties.arXiv preprint arXiv:2311.15154, 2023

Yurii Nesterov. High-order reduced-gradient methods for composite variational inequali- ties.arXiv preprint arXiv:2311.15154, 2023

arXiv 2023
[22]

Universal complexity bounds for universal gradient methods of convex optimization

Yurii Nesterov. Universal complexity bounds for universal gradient methods of convex optimization. 2025

2025
[1]

Optimal methods for higher-order smooth monotone variational inequalities.arXiv preprint arXiv:2205.06167, 2022

Deeksha Adil, Brian Bullins, Arun Jambulapati, and Sushant Sachdeva. Optimal methods for higher-order smooth monotone variational inequalities.arXiv preprint arXiv:2205.06167, 2022

arXiv 2022
[2]

Walter de Gruyter GmbH & Co KG, 2014

Miroslav Bac´ ak.Convex analysis and optimization in Hadamard spaces, volume 22. Walter de Gruyter GmbH & Co KG, 2014

2014
[3]

Princeton university press, 2009

Rajendra Bhatia.Positive definite matrices. Princeton university press, 2009

2009
[4]

Worst-case evaluation complexity for unconstrained nonlinear optimiza- tion using high-order regularized models.Mathematical Programming, 163(1-2):359–368, 2017

Ernesto G Birgin, JL Gardenghi, Jos´ e Mario Mart ´ ınez, Sandra Augusta Santos, and Philippe L Toint. Worst-case evaluation complexity for unconstrained nonlinear optimiza- tion using high-order regularized models.Mathematical Programming, 163(1-2):359–368, 2017

2017
[5]

Higher-order methods for convex-concave min-max optimization and monotone variational inequalities.SIAM Journal on Optimization, 32(3):2208–2229, 2022

Brian Bullins and Kevin A Lai. Higher-order methods for convex-concave min-max optimization and monotone variational inequalities.SIAM Journal on Optimization, 32(3):2208–2229, 2022

2022
[6]

Adaptive cubic regularisation methods for unconstrained optimization

Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Adaptive cubic regularisation methods for unconstrained optimization. Part I: motivation, convergence and numerical results.Mathematical Programming, 127(2):245–295, 2011

2011

Show all 28 references

[7]

Sharp worst-case evalua- tion complexity bounds for arbitrary-order nonconvex optimization with inexpensive con- straints.SIAM Journal on Optimization, 30(1):513–541, 2020

Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Sharp worst-case evalua- tion complexity bounds for arbitrary-order nonconvex optimization with inexpensive con- straints.SIAM Journal on Optimization, 30(1):513–541, 2020

2020
[8]

Minimizing quasi-self-concordant functions by gradient regularization of Newton method.Mathematical Programming, pages 1–39, 2025

Nikita Doikov. Minimizing quasi-self-concordant functions by gradient regularization of Newton method.Mathematical Programming, pages 1–39, 2025

2025
[9]

Super-universal regularized Newton method.SIAM Journal on Optimization, 34(1):27–56, 2024

Nikita Doikov, Konstantin Mishchenko, and Yurii Nesterov. Super-universal regularized Newton method.SIAM Journal on Optimization, 34(1):27–56, 2024

2024
[10]

Regularized newton methods for minimizing functions with H¨ older continuous hessians.SIAM Journal on Optimization, 27(1):478– 506, 2017

Geovani Nunes Grapiglia and Yurii Nesterov. Regularized newton methods for minimizing functions with H¨ older continuous hessians.SIAM Journal on Optimization, 27(1):478– 506, 2017

2017
[11]

Tensor methods for minimizing convex functions with H¨ older continuous higher-order derivatives.SIAM Journal on Optimiza- tion, 30(4):2750–2779, 2020

Geovani Nunes Grapiglia and Yurii Nesterov. Tensor methods for minimizing convex functions with H¨ older continuous higher-order derivatives.SIAM Journal on Optimiza- tion, 30(4):2750–2779, 2020

2020
[12]

Generalized optimistic methods for convex-concave saddle point problems.SIAM Journal on Optimization, 35(3):2066–2097, 2025

Ruichen Jiang and Aryan Mokhtari. Generalized optimistic methods for convex-concave saddle point problems.SIAM Journal on Optimization, 35(3):2066–2097, 2025
[13]

Extragradient method for finding saddle points and other prob- lems.Matekon, 13(4):35–49, 1977

Galina M Korpelevich. Extragradient method for finding saddle points and other prob- lems.Matekon, 13(4):35–49, 1977. 23

1977
[14]

Perseus: A simple and optimal high-order method for variational inequalities.Mathematical Programming, 209(1):609–650, 2025

Tianyi Lin and Michael I Jordan. Perseus: A simple and optimal high-order method for variational inequalities.Mathematical Programming, 209(1):609–650, 2025

2025
[15]

Arkadi Nemirovski. Prox-method with rate of convergenceO(1/t) for variational inequal- ities with lipschitz continuous monotone operators and smooth convex-concave saddle point problems.SIAM Journal on Optimization, 15(1):229–251, 2004

2004
[16]

Universal gradient methods for convex optimization problems.Mathe- matical Programming, 152(1):381–404, 2015

Yurii Nesterov. Universal gradient methods for convex optimization problems.Mathe- matical Programming, 152(1):381–404, 2015

2015
[17]

Springer, Cham, 2018

Yurii Nesterov.Lectures on convex optimization, volume 137. Springer, Cham, 2018

2018
[18]

Implementable tensor methods in unconstrained convex optimization

Yurii Nesterov. Implementable tensor methods in unconstrained convex optimization. Mathematical Programming, pages 1–27, 2019

2019
[19]

Superfast second-order methods for unconstrained convex optimization

Yurii Nesterov. Superfast second-order methods for unconstrained convex optimization. Journal of Optimization Theory and Applications, 191(1):1–30, 2021

2021
[21]

Primal subgradient methods with predefined step sizes.Journal of Op- timization Theory and Applications, 203(3):2083–2115, 2024

Yurii Nesterov. Primal subgradient methods with predefined step sizes.Journal of Op- timization Theory and Applications, 203(3):2083–2115, 2024
[23]

SIAM, Philadelphia, 1994

Yurii Nesterov and Arkadi Nemirovski.Interior-point polynomial algorithms in convex programming. SIAM, Philadelphia, 1994

1994
[24]

Cubic regularization of Newton’s method and its global performance.Mathematical Programming, 108(1):177–205, 2006

Yurii Nesterov and Boris Polyak. Cubic regularization of Newton’s method and its global performance.Mathematical Programming, 108(1):177–205, 2006

2006
[25]

On the Riemannian geometry defined by self- concordant barriers and interior-point methods.Foundations of Computational Math- ematics, 2(4):333–361, 2002

Yurii Nesterov and Michael Todd. On the Riemannian geometry defined by self- concordant barriers and interior-point methods.Foundations of Computational Math- ematics, 2(4):333–361, 2002

2002
[26]

Springer, 2006

Peter Petersen.Riemannian geometry. Springer, 2006

2006
[27]

Gradient-normalized smoothness for optimization with approximate Hessians.arXiv preprint arXiv:2506.13710, 2025

Andrei Semenov, Martin Jaggi, and Nikita Doikov. Gradient-normalized smoothness for optimization with approximate Hessians.arXiv preprint arXiv:2506.13710, 2025

2025 arXiv
[28]

Inequalities between the two kinds of eigenvalues of a linear transforma- tion.Proceedings of the national academy of sciences, 35(7):408–411, 1949

Hermann Weyl. Inequalities between the two kinds of eigenvalues of a linear transforma- tion.Proceedings of the national academy of sciences, 35(7):408–411, 1949. 24

1949

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [20]

High-order reduced-gradient methods for composite variational inequali- ties.arXiv preprint arXiv:2311.15154, 2023

Yurii Nesterov. High-order reduced-gradient methods for composite variational inequali- ties.arXiv preprint arXiv:2311.15154, 2023

arXiv 2023

[2] [22]

Universal complexity bounds for universal gradient methods of convex optimization

Yurii Nesterov. Universal complexity bounds for universal gradient methods of convex optimization. 2025

2025

[3] [1]

Optimal methods for higher-order smooth monotone variational inequalities.arXiv preprint arXiv:2205.06167, 2022

Deeksha Adil, Brian Bullins, Arun Jambulapati, and Sushant Sachdeva. Optimal methods for higher-order smooth monotone variational inequalities.arXiv preprint arXiv:2205.06167, 2022

arXiv 2022

[4] [2]

Walter de Gruyter GmbH & Co KG, 2014

Miroslav Bac´ ak.Convex analysis and optimization in Hadamard spaces, volume 22. Walter de Gruyter GmbH & Co KG, 2014

2014

[5] [3]

Princeton university press, 2009

Rajendra Bhatia.Positive definite matrices. Princeton university press, 2009

2009

[6] [4]

Worst-case evaluation complexity for unconstrained nonlinear optimiza- tion using high-order regularized models.Mathematical Programming, 163(1-2):359–368, 2017

Ernesto G Birgin, JL Gardenghi, Jos´ e Mario Mart ´ ınez, Sandra Augusta Santos, and Philippe L Toint. Worst-case evaluation complexity for unconstrained nonlinear optimiza- tion using high-order regularized models.Mathematical Programming, 163(1-2):359–368, 2017

2017

[7] [5]

Higher-order methods for convex-concave min-max optimization and monotone variational inequalities.SIAM Journal on Optimization, 32(3):2208–2229, 2022

Brian Bullins and Kevin A Lai. Higher-order methods for convex-concave min-max optimization and monotone variational inequalities.SIAM Journal on Optimization, 32(3):2208–2229, 2022

2022

[8] [6]

Adaptive cubic regularisation methods for unconstrained optimization

Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Adaptive cubic regularisation methods for unconstrained optimization. Part I: motivation, convergence and numerical results.Mathematical Programming, 127(2):245–295, 2011

2011