On gradient descent-ascent flows in metric spaces

Noboru Isobe; Sho Shimoyama

REVIEW 2 major objections 4 minor 1 cited by

On gradient descent-ascent flows in metric spaces

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper establishes well-posedness—existence, uniqueness, and continuous dependence on initial data—of gradient descent-ascent flows on the product of two complete metric spaces, formulated through evolution variational inequalities…

desk verdict The EVI formulation of GDA and the Wasserstein application are worth attention, but the minimax theorem underpinning the resolvent is false as stated, so the abstract well-posedness theorem needs major repair. read the letter →

arxiv 2506.20258 v1 pith:IGC7UAAB submitted 2025-06-25 math.FA math.APmath.MGmath.OC

Noboru Isobe , Sho Shimoyama This is my paper · ORCID

classification math.FAmath.APmath.MGmath.OC MSC 49Q2247H0549J5290C4758E3046T0560B10

keywords gradientdescent-ascentflowsevolutionvariationalinequalitiesmetricspacesWassersteinsaddlepointproblemsminimaxtheoremminimizing-maximizingmovementconvex-concavefunctionals

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 establishes a metric-space theory for gradient descent-ascent (GDA) flows, the dynamical systems used to find saddle points of bivariate functionals in optimization, game theory, and robust control. Previously, such flows were only well understood in linear spaces, where maximal monotone operator theory applies; this paper extends them to the product of two complete metric spaces by formulating the flow as a solution of a system of evolution variational inequalities (EVIs). The central achievement is a well-posedness theorem: under convex-concavity and regularity assumptions, a unique GDA flow exists for every initial point in the effective domain and depends continuously on that point. The theory is then specialized to Wasserstein spaces over Hilbert spaces, yielding the first global existence and exponential saddle-point convergence results for Wasserstein GDA flows, thereby addressing an open problem in the mean-field two-player zero-sum game literature.

What carries the argument

The central object is the resolvent operator $J_\tau(z) := \arg\,\mathrm{minimax}_{z'\in X\times Y} \Phi_\tau(z;z')$, where $\Phi_\tau(x,y;x',y') = \phi(x',y') + \frac{1}{2\tau}\big(d_X^2(x',x) - d_Y^2(y',y)\big)$ is the squared-distance regularization of the bivariate functional $\phi$. Under Assumption 4.1, which requires this regularized functional to be $(\tau^{-1}+\lambda)$-convex-concave along suitable curves for all small $\tau$, the resolvent is single-valued, and the discrete minimizing-maximizing movement $J_{t/n}^n[w_0]$ converges to the unique EVIs solution. The proof relies on a metric-space minimax theorem (proved without compactness assumptions) that gives the saddle point defining $J_\tau$, and on a modified local slope $|\partial\phi|$, which measures the rate of decrease of the duality gap rather than of $\phi$ itself; this slope drives the error estimates and the exponential decay bounds.

What would settle it

Find two complete metric spaces and a functional satisfying the decomposition and boundary-value assumptions but not the joint convexity-concavity condition, for which two different EVIs solutions with the same initial point exist, or for which the minimizing-maximizing iterates fail to converge to a limit satisfying the integral EVIs.

Watch

Extended reading notes

Core claim

The paper proves that GDA flows can be defined and solved in a purely metric setting, without any linear or differentiable structure, by replacing the classical differential inclusions with a pair of evolution variational inequalities. Its main theorem states that for any $\lambda \in \mathbb{R}$, under an additive decomposition of the functional, a boundary-value convention on the effective domain, and a joint convexity-concavity condition on the regularized functional, every admissible initial point admits a unique global solution $S_t[w_0]$ of the EVIs system, with the semigroup property and continuity in the initial datum. This well-posedness is obtained through a minimizing-maximizing movement scheme: a resolvent operator $J_\tau$ maps the current point to the unique saddle point of the squared-distance-regularized functional, and the discrete-time iterates $J_{t/n}^n[w_0]$ converge to the continuous flow with an explicit order-1/2 error estimate. The same machinery yields structural properties, including $\lambda$-contraction, regularization, and an exponential decay bound for the duality gap along the flow. In the Wasserstein setting, the paper shows that the EVIs formulation is equivalent to the classically defined Wasserstein GDA flow, and that for strongly convex-concave functionals the flow converges exponentially to the unique saddle point.

Load-bearing premise

The entire construction depends on the joint convexity-concavity of the regularized functional along suitable curves for every small step size; if that condition fails, the resolvent may not be single-valued and the discrete scheme loses its definition.

Editorial extensions

If this is right

Wasserstein GDA flows now have a rigorous existence, uniqueness, and stability theory, so asymptotic questions can be posed with a well-defined dynamical system in hand.
For strongly convex-concave functionals on Wasserstein spaces over Hilbert spaces, the flow converges exponentially to the unique saddle point with an explicit rate $e^{-2\lambda t}$, an answer to the open convergence problem raised in [42] under the stated structural assumptions.
The quantitative order-1/2 error estimate for the discrete scheme provides the first convergence-rate guarantee for metric-space GDA algorithms, bridged to the continuum flow.
The framework automatically covers entropy-regularized mean-field zero-sum games, for which the relative-entropy terms give the required strong convexity-concavity, yielding existence and duality-gap decay without any linear structure.
When $\phi$ has no cross term (the case $\varphi=0$), the EVIs system decouples into two independent metric gradient flows, recovering the classical theory as a special case.

Reading between the lines

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

The joint convexity-concavity condition in Assumption 4.1 can be read as a geometric requirement on the product metric itself: well-posedness of GDA flows is tied to the squared distance being sufficiently curved in opposing directions, much as uniform convexity of the squared distance underlies purely minimizing gradient flows.
One natural stress test is whether the same well-posedness theory holds on Wasserstein spaces over Riemannian manifolds; the paper's assumption verification uses generalized geodesics, which exist in Hilbert bases, so extending that verification to curved base manifolds would be a significant testable step.
Since the modified slope tracks the duality gap rather than the energy, the exponential decay result suggests that practical discretizations of GDA on nonlinear spaces could monitor the duality gap instead of the functional value; the error estimate in the paper gives a quantitative justification for such monitoring.

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

Here's the short version. The paper gives a genuinely new formulation of GDA as a system of EVIs and a plausible route to Wasserstein well-posedness, but the abstract saddle-point theorem on which the resolvent depends (Theorem 1.7) is false as stated. That is not a cosmetic gap; the proof of Jτ single-valuedness uses it directly, so the general-metric-space version of Theorem 1.4 is not established.

The system of EVIs is a natural extension of the AGS/Muratori-Savaré framework, and the paper earns credit for the λ-contraction property, the quantitative error estimate for the minimizing-maximizing movement scheme, the modified slope, and the exponential duality-gap decay. The equivalence between EVIs and Wasserstein GDA (Theorem 5.7) is a concrete step toward the sort of convergence questions raised by Wang-Chizat.

The soft spots are serious. Theorem 1.7's proof says that since φY is strongly concave and upper semicontinuous, it has a maximizer. In a complete metric space that implication is simply false. Take X={0}, Y=ℓ², φ(0,y)=-ε||y||²-e^{⟨y,e₁⟩}. It is closed, satisfies Assumption 2.1, is 2ε-convex-concave along straight-line geodesics, but sup_y φ=0 and no point attains it. No saddle point exists. Since Remark 4.4 defines Jτ as the saddle point of Φτ via Theorem 1.7, the discrete scheme is not defined on arbitrary complete metric spaces. The Wasserstein case may survive, because the squared-distance and entropy terms add coercivity, but the paper does not prove it, and the 'general metric spaces' claim is overstated.

The reader's concerns about Assumption 4.1 are also fair: it is a strong joint convexity-concavity condition, verified only for Wasserstein spaces and CAT(0), so the scope is narrower than the abstract promises. These are the main issues; the rest of the analysis looks plausible and the citation pattern is appropriate.

Verdict: this deserves a serious referee, but it needs a genuine revision of Section 2 and the resolvent construction, not copyediting. I would not cite the abstract theorem in its current form.

Referee Report

2 major / 4 minor

Summary. The paper develops a metric-space theory of gradient descent-ascent (GDA) flows. It formulates a system of evolution variational inequalities (EVIs) for a bivariate functional on a product of complete metric spaces, constructs a discrete minimizing-maximizing movement scheme via a saddle-point resolvent, and proves well-posedness (Theorem 1.4), contraction properties, regularity estimates, and exponential decay of the Nikaidô–Isoda duality gap along the flow. These abstract results are then applied to Wasserstein spaces over Hilbert spaces, yielding existence and exponential convergence of Wasserstein GDA flows for strongly convex-concave functionals (Theorem 5.7, Corollary 5.8). The central technical tool is a claimed minimax theorem without compactness (Theorem 1.7).

Significance. If the main results were correct, the paper would provide a broad and timely extension of GDA theory beyond Banach spaces, addressing an open problem raised by Wang and Chizat [42] about Wasserstein GDA convergence. The manuscript contains substantial original machinery: a resolvent scheme with explicit quantitative error estimates, a modified slope adapted to the duality gap, and a self-contained treatment of the Wasserstein application. It does not rely on fitted parameters or numerical predictions, and the constructions are deterministic. However, the abstract well-posedness theorem depends crucially on a minimax existence result that is false as stated; since that result underpins the definition of the resolvent, the central claim is not established.

major comments (2)

[Section 2 (proof of Theorem 1.7)] The assertion 'Since φY is strongly-concave and upper semicontinuous, there exists the maximizer y∗∈DomφY' is false in a general complete metric space: strong concavity and upper semicontinuity do not imply attainment unless some coercivity or compactness condition holds. A concrete counterexample is X={0}, Y=ℓ², and φ(0,y)=−ε‖y‖²−e^{⟨y,e1⟩} with ε>0. Assumption 2.1 holds, φ is closed, and φ is 2ε-convex-concave along straight-line geodesics, so the hypotheses of Theorem 1.7 are satisfied. Yet sup_y φ(0,y)=0 (take y=−te1, t→∞) while φ(0,y)<0 for every y; hence no maximizer and no saddle point exist. Thus Theorem 1.7 is not merely missing a detail: it is false as stated.
[Section 4.1 (Definition 4.3 and Remark 4.4)] The resolvent Jτ is defined as the arg minimax of Φτ, and Remark 4.4 asserts that Jτ is single-valued by virtue of Theorem 1.7. Since Theorem 1.7 is false, the minimizing–maximizing movement scheme (4.3) is not well-defined for arbitrary complete metric spaces. Consequently Theorem 1.4, which is invoked in Section 5 to obtain the Wasserstein well-posedness results (Theorem 5.7 and Corollary 5.8), lacks the required foundation in its stated generality. A repair would require adding a coercivity or attainment hypothesis to Assumption 4.1, or verifying such a hypothesis separately in the Wasserstein setting; the paper currently does neither.

minor comments (4)

[Section 4.1 (Lemma 4.7)] The statement 'for each τ∈(0,λ−)' should read 'for each τ∈(0,1/λ−)', since λ−=max{−λ,0} has dimension of inverse time.
[Section 5.3 (Proposition 5.9)] In the first displayed inequality of the proof after (5.8), the second integral should be ∫ℓ(x′,y′) d(µ′⊗ν′)(x′,y′), not ∫ℓ(x,y) d(µ⊗ν)(x,y).
[Section 1 (Eq. (1.1))] The notation 'X×Y ∋ (x′,y′)↦→−Φτ' appears to contain a typographical artifact; the arrow should be simply '↦'.
[Section 3 (Eq. (3.1))] The convention that |∂ϕ|(x,y)=0 at isolated points of Domϕ is stated without justification; unless the lim sup is taken over sequences approaching from outside the singleton, the definition is ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the flow existence and convergence theorems are proved from explicit convex-concavity assumptions via a resolvent scheme, with no fitted parameters or self-citations.

full rationale

Circularity analysis: this paper is a pure mathematical construction. It contains no fitted parameters, no empirical predictions, and no data subsets, so the fitted-input-called-prediction pattern cannot arise. The central derivation is: define the EVI system (Definition 1.1); impose the joint regularity assumption that the regularized functional Phi_tau is (tau^-1+lambda)-convex-concave along curves (Assumption 4.1); define the resolvent J_tau as the saddle point of Phi_tau (Definition 4.3); invoke Theorem 1.7 to make J_tau single-valued; prove continuity of J_tau (Lemma 4.7), discrete EVI inequalities (Lemma 4.8), slope estimates (Lemma 4.10), and Cauchy estimates (Lemma 4.12); and finally pass to the continuous limit (Proposition 4.13 and Theorem 1.4). None of these steps presupposes the existence or uniqueness of the flow being proved. Assumption 4.1 is a hypothesis that makes the resolvent well-defined; it does not contain the conclusion of Theorem 1.4. The flow's EVI inequalities are obtained as the limit of the resolvent scheme, not assumed as an input. The Wasserstein results (Theorem 5.7 and Corollary 5.8) are derived from the abstract theorems rather than built into their statements. The modified slope |partial phi| and the Nikaido-Isoda duality gap are internal definitions used to quantify the inequalities; the exponential bound (3.3) is proved from those definitions and the EVI flow, not assumed. The reference list contains no work by Isobe or Shimoyama, so there are no self-citations and no uniqueness theorem imported from the authors' prior work. The only substantive concern in the manuscript is mathematical correctness, not circularity: the proof of Theorem 1.7 asserts that strong concavity plus upper semicontinuity guarantees the existence of a maximizer on a complete metric space, and the reviewer's counterexample challenges that assertion. Even if that proof is invalid, the issue is a potentially false intermediate lemma, not a derivation that reduces to its own inputs. Accordingly, no circular step is identifiable, and the paper's claimed derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

The paper introduces no fitted parameters; the only numbers chosen by hand are the regularity and convexity parameters (lambda, tau), which are assumptions, not fitted. The main burden lies in the axioms listed above, especially Assumption 4.1, which is strong and tailored to the proof technique. No physical entities are introduced; the modified slope is a mathematical definition.

assumptions (4)

domain assumption Assumption 2.1: phi is extended by +infinity on (DomX phi)^c x DomY phi and by -infinity on DomX phi x (DomY phi)^c.
Used throughout to restrict saddle points to the effective domain and to make the inf/sup interchange in Remark 2.7 valid.
domain assumption Assumption 1.3: phi(x,y) = phi(x,y) + psiX(x) + psiY(y) with psiX proper lsc, psiY proper usc, and phi real-valued continuous or locally Lipschitz.
Basis for the integral characterization of EVIs (Proposition 3.2) and for extending the flow from Dom(|∂phi|) to Dom phi in Theorem 1.4.
ad hoc to paper Assumption 4.1: for any z,z0,z1 in Dom phi there exist curves gamma and sigma such that Phi_tau is (tau^{-1}+lambda)-convex-concave along gamma x sigma for all tau in (0,1/lambda_-).
Core geometric hypothesis; makes the resolvent J_tau single-valued (Theorem 1.7) and yields the slope estimates and contraction property in Lemmas 4.10 and 4.12.
standard math Standard results from Ambrosio-Gigli-Savare [1] (EVI theory, Wasserstein subdifferential calculus, generalized geodesics) and Muratori-Savare [34] (integral characterization of EVI) are used as black boxes.
The proofs rely on these for existence of gradient flows, tangent vectors, and integration by parts in Wasserstein spaces; no independent verification is provided in this paper.

invented entities (1)

Modified local slope |∂phi| defined in (3.1)
purpose: Quantifies the rate at which the Nikaidô-Isoda duality gap decreases along the flow; used in the error estimate of Proposition 4.13 and the exponential bound in Theorem 3.3.
A new mathematical definition, not an observable entity; it is an analytical tool internal to the proofs and has no external falsifiable handle.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On gradient descent-ascent flows in metric spaces." pith.science (2026). https://pith.science/paper/IGC7UAAB

@misc{pith2026250620258,
  author       = {Pith},
  title        = {Pith review of: On gradient descent-ascent flows in metric spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGC7UAAB}},
  note         = {Machine review of arXiv:2506.20258}
}

abstract

Gradient descent-ascent (GDA) flows play a central role in finding saddle points of bivariate functionals, with applications in optimization, game theory, and robust control. While they are well-understood in Hilbert and Banach spaces via maximal monotone operator theory, their extension to general metric spaces, particularly Wasserstein spaces, has remained largely unexplored. In this paper, we develop a mathematical theory of GDA flows on the product of two complete metric spaces, formulating them as solutions to a system of evolution variational inequalities (EVIs) driven by a proper, closed functional $\phi$. Under mild convex-concave and regularity assumptions on $\phi$, we prove the existence, uniqueness, and stability of the flows via a novel minimizing-maximizing movement scheme and a minimax theorem on metric spaces. We establish a $\lambda$-contraction property, derive a quantitative error estimate for the discrete scheme, and demonstrate regularization effects analogous to classical gradient flows. Moreover, we obtain an exponential decay bound for the Nikaid\^o--Isoda duality gap along the flow. Focusing on Wasserstein spaces over Hilbert spaces, we show the global existence in time and the exponential convergence of the Wasserstein GDA flow to the unique saddle point for strongly convex-concave functionals. Our framework unifies and extends existing analyses, offering a metric-geometric perspective on GDA dynamics in nonlinear and non-smooth settings.

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

Local exponential stability of mean-field Langevin descent-ascent and associated particle system
cs.LG 2026-02 conditional novelty 7.0 of 10

Starting close to the mixed Nash equilibrium, mean-field Langevin descent-ascent converges exponentially fast in Wasserstein distance.

Reference graph

Works this paper leans on

45 extracted references · 38 canonical work pages · cited by 1 Pith paper

[42]

Open Problem: Convergence of Single-Timescale Mean-Field Langevin Descent-Ascent for Two-Player Zero-Sum Games

G. Wang and L. Chizat. “Open Problem: Convergence of Single-Timescale Mean-Field Langevin Descent-Ascent for Two-Player Zero-Sum Games.”Proceedings of Thirty Seventh Conference on Learning Theory. 247. Proceedings of Machine Learning Research. 2024, pp. 5345–5350. url: https://proceedings.mlr.press/v247/wang24c.html (cit. on pp. 2, 5, 6)

work page 2024
[1]

Gradient Flows in Metric Spaces and in the Space of Probability Measures

L. Ambrosio, N. Gigli, and G. Savaré. “Gradient Flows in Metric Spaces and in the Space of Probability Measures.” 2nd ed. Lectures in Mathematics ETH Zürich. Basel: Birkhäuser Verlag, 2008 (cit. on pp. 2–6, 9, 10, 12, 14, 16, 17, 19, 20, 25–29)

work page 2008
[2]

A Gradient Method for Approximating Saddle Points and Constrained Maxima

K. J. Arrow and L. Hurwicz. “A Gradient Method for Approximating Saddle Points and Constrained Maxima.”Traces and Emergence of Nonlinear Programming. Basel: Springer Basel, 2014, pp. 45–60.url: https://doi.org/10.1007/978-3-0348-0439-4_2 (cit. on pp. 2, 5). REFERENCES 31

work page doi:10.1007/978-3-0348-0439-4_2 2014
[3]

Studies in Linear and Non-Linear Programming

K. J. Arrow, L. Hurwicz, and H. Uzawa. “Studies in Linear and Non-Linear Programming.” II. Stanford Mathematical Studies in the Social Sciences. With contributions by H. B. Chenery, S. M. Johnson, S. Karlin, T. Marschak, R. M. Solow. Stanford University Press, Stanford, CA, 1958, pp. vii+229 (cit. on pp. 2, 5)

work page 1958
[4]

Maximal Monotone Operators Associated with Saddle Functions

H. Asakawa. “Maximal Monotone Operators Associated with Saddle Functions.” TRU Mathematics 22.2 1986, pp. 47–71 (cit. on p. 5)

work page 1986
[5]

Continuous Time Approximation of Nash Equilib- ria in Monotone Games

R. Awi, R. Hynd, and H. Mawi. “Continuous Time Approximation of Nash Equilib- ria in Monotone Games.” Communications in Contemporary Mathematics 26.05 2024, p. 2350021. eprint: https : / / doi . org / 10 . 1142 / S0219199723500219. url: https : //doi.org/10.1142/S0219199723500219 (cit. on p. 5)

work page doi:10.1142/s0219199723500219 2024
[6]

Convex Analysis and Optimization in Hadamard Spaces

M. Bacak. “Convex Analysis and Optimization in Hadamard Spaces.” Berlin, München, Boston: De Gruyter, 2014. url: https : / / doi . org / 10 . 1515 / 9783110361629(cit. on p. 17)

work page 2014
[7]

Uncertainty Quantification of the 4th Kind; Optimal Posterior Accuracy- UncertaintyTradeoffwiththeMinimumEnclosingBall

H. H. Bajgiran, P. Batlle, H. Owhadi, M. Samir, C. Scovel, M. Shirdel, M. Stanley, and P. Tavallali. “Uncertainty Quantification of the 4th Kind; Optimal Posterior Accuracy- UncertaintyTradeoffwiththeMinimumEnclosingBall.” Journal of Computational Physics 471 2022, p. 111608. url: https://www.sciencedirect.com/science/article/pii/ S0021999122006714 (cit. on p. 6)

work page 2022

Show all 45 references

[8]

Hypercontractivité De Semi-Groupes De Diffusion

D. Bakry and M. Émery. “Hypercontractivité De Semi-Groupes De Diffusion.”Comptes Rendus des Séances de l’Académie des Sciences. Série I. Mathématique 299.15 1984, pp. 775–778 (cit. on p. 6)

1984
[9]

Numerical Solution of Saddle Point Problems

M. Benzi, G. H. Golub, and J. Liesen. “Numerical Solution of Saddle Point Problems.” Acta Numerica14 2005, pp. 1–137 (cit. on p. 5)

2005
[10]

A Hybrid Variational Principle for the Keller–segel System in R2

Blanchet, Adrien, Carrillo, José Antonio, Kinderlehrer, David, Kowalczyk, Michał, Lau- rençot, Philippe, and Lisini, Stefano. “A Hybrid Variational Principle for the Keller–segel System in R2.” ESAIM: M2AN 49.6 2015, pp. 1553–1576.url: https://doi.org/10. 1051/m2an/2015021 (ci...

2015
[11]

Opérateurs Maximaux Monotones Et Semi-Groupes De Contractions Dans Les Espaces De Hilbert

H. Brézis. “Opérateurs Maximaux Monotones Et Semi-Groupes De Contractions Dans Les Espaces De Hilbert.”No. 5. North-Holland Mathematics Studies. Notas de Matemática, No. 50. [Mathematical Notes]. North-Holland Publishing Co., Amsterdam-London; Amer- ican Elsevier Publishing Co...

1973
[12]

Functional Analysis, Sobolev Spaces and Partial Differential Equations

H. Brézis. “Functional Analysis, Sobolev Spaces and Partial Differential Equations.” New York, NY: Springer New York, 2011.url: https : / / doi . org / 10 . 1007 / 978 - 0 - 387 - 70914-7_3 (cit. on p. 15)

2011
[13]

Convergence of the Min-Max Langevin Dynamics and Algorithm for Zero-Sum Games

Y. Cai, S. Mitra, X. Wang, and A. Wibisono. “Convergence of the Min-Max Langevin Dynamics and Algorithm for Zero-Sum Games.” 2024. arXiv:2412.20471 [cs.GT]. url: https://arxiv.org/abs/2412.20471 (cit. on p. 6)

2024 arXiv
[14]

Asymptotic Convergence of Constrained Primal- Dual Dynamics

A. Cherukuri, E. Mallada, and J. Cortés. “Asymptotic Convergence of Constrained Primal- Dual Dynamics.” Systems & Control Letters87 2016, pp. 10–15. url: https : / / www . sciencedirect.com/science/article/pii/S0167691115002078 (cit. on p. 5)

2016
[15]

Introduction to Gradient Flows in Metric Spaces

P. Clément. “Introduction to Gradient Flows in Metric Spaces.” 2009. url: https : / / math.leidenuniv.nl/reports/files/2009-09.pdf (cit. on p. 23)

2009
[16]

Introduction to Gradient Flows in Metric Spaces (II)

P. Clément. “Introduction to Gradient Flows in Metric Spaces (II).” 2010.url: https: //igk.math.uni-bielefeld.de/sites/default/files/study_materials/notes2.pdf (cit. on pp. 14, 15). 32 REFERENCES

2010
[17]

Game on Random Environment, Mean-Field Langevin System, and Neural Networks

G. Conforti, A. Kazeykina, and Z. Ren. “Game on Random Environment, Mean-Field Langevin System, and Neural Networks.”Mathematics of Operations Research48.1 2023, pp. 78–99. eprint:https://doi.org/10.1287/moor.2022.1252. url: https://doi.org/ 10.1287/moor.2022.1252 (cit. on p. 6)

2023
[18]

Strategic Distribution Shift of Interacting Agents via Coupled Gradient Flows

L. Conger, F. Hoffmann, E. Mazumdar, and L. Ratliff. “Strategic Distribution Shift of Interacting Agents via Coupled Gradient Flows.”Advances in Neural Information Pro- cessing Systems. 36. 2023, pp. 45971–46006. url: https : / / proceedings . neurips . cc / paper _ files / pa...

2023
[19]

Coupled Wasserstein Gradient Flows for Min-Max and Cooperative Games

L. Conger, F. Hoffmann, E. Mazumdar, and L. J. Ratliff. “Coupled Wasserstein Gradient Flows for Min-Max and Cooperative Games.” 2024. arXiv:2411.07403v1 [math.AP]. url: https://arxiv.org/abs/2411.07403 (cit. on p. 6)

2024 arXiv
[20]

Generation of Semi-Groups of Nonlinear Transforma- tions on General Banach Spaces

M. G. Crandall and T. M. Liggett. “Generation of Semi-Groups of Nonlinear Transforma- tions on General Banach Spaces.”American Journal of Mathematics93 1971, pp. 265–298. url: https://doi.org/10.2307/2373376 (cit. on pp. 5, 23)

1971 doi
[21]

Problemi Di Evoluzione in Spazi Metrici E Curve Di Massima Pendenza

E. De Giorgi, A. Marino, and M. Tosques. “Problemi Di Evoluzione in Spazi Metrici E Curve Di Massima Pendenza.” Italian.Atti della Accademia Nazionale dei Lincei. Classe di Scienze Fisiche, Matematiche e Naturali. Rendiconti68.3 1980, pp. 180–187.url: http: //eudml.org/doc/288...

1980
[22]

A Mean-Field Analysis of Two-Player Zero-Sum Games

C. Domingo-Enrich, S. Jelassi, A. Mensch, G. Rotskoff, and J. Bruna. “A Mean-Field Analysis of Two-Player Zero-Sum Games.”Advances in Neural Information Processing Systems.33. Curran Associates, Inc., 2020, pp. 20215–20226.url: https://proceedings. neurips.cc/paper_files/paper...

2020
[23]

Analysis of Kernel Mirror Prox for Measure Optimiza- tion

P. Dvurechensky and J.-J. Zhu. “Analysis of Kernel Mirror Prox for Measure Optimiza- tion.” Proceedings of The 27th International Conference on Artificial Intelligence and Sta- tistics.238. Proceedings of Machine Learning Research. 2024, pp. 2350–2358.url: https: //proceedings...

2024
[24]

Convex Analysis and Variational Problems

I. Ekeland and R. Témam. “Convex Analysis and Variational Problems.” Society for In- dustrial and Applied Mathematics, 1999 (cit. on pp. 7, 10)

1999
[25]

Gradient Flows of Time-Dependent Func- tionals in Metric Spaces and Applications to PDEs

L. C. F. Ferreira and J. C. Valencia-Guevara. “Gradient Flows of Time-Dependent Func- tionals in Metric Spaces and Applications to PDEs.”Monatshefte für Mathematik185.2 2018, pp. 231–268.url: https://doi.org/10.1007/s00605-017-1037-y (cit. on p. 5)

2018 doi
[26]

On the Subdifferential of a Saddle Function

J.-P. Gossez. “On the Subdifferential of a Saddle Function.”Journal of Functional Analysis 11.2 1972, pp. 220–230.url: https://www.sciencedirect.com/science/article/pii/ 0022123672900924 (cit. on pp. 5, 7)

1972
[27]

Finding Mixed Nash Equilibria of Generative Adver- sarial Networks

Y.-P. Hsieh, C. Liu, and V. Cevher. “Finding Mixed Nash Equilibria of Generative Adver- sarial Networks.”Proceedings of the 36th International Conference on Machine Learning
[28]

Evolution of Mixed Strategies in Monotone Games

R. Hynd. “Evolution of Mixed Strategies in Monotone Games.”SIAM Journal on Opti- mization 33.4 2023, pp. 2750–2771. eprint:https://doi.org/10.1137/22M1486066. url: https://doi.org/10.1137/22M1486066 (cit. on p. 5)

2023 doi
[29]

The Variational Formulation of the Fokker– Planck Equation

R. Jordan, D. Kinderlehrer, and F. Otto. “The Variational Formulation of the Fokker– Planck Equation.”SIAM Journal on Mathematical Analysis29.1 1998, pp. 1–17. eprint: REFERENCES 33 https://doi.org/10.1137/S0036141096303359 . url: https://doi.org/10.1137/ S0036141096303359 (ci...

1998 doi
[30]

Symmetric Mean-Field Langevin DynamicsforDistributionalMinimaxProblems

J. Kim, K. Yamamoto, K. Oko, Z. Yang, and T. Suzuki. “Symmetric Mean-Field Langevin DynamicsforDistributionalMinimaxProblems.” The Twelfth International Conference on Learning Representations. 2024. url: https://openreview.net/forum?id=YItWKZci78 (cit. on p. 6)

2024
[31]

Gradient Flow for the Boltzmann Entropy and Cheeger’s Energy on Time- Dependent Metric Measure Spaces

E. Kopfer. “Gradient Flow for the Boltzmann Entropy and Cheeger’s Energy on Time- Dependent Metric Measure Spaces.”Calculus of Variations and Partial Differential Equa- tions 57.1 2017, p. 20.url: https://doi.org/10.1007/s00526- 017- 1287- 5 (cit. on p. 5)

2017 doi
[32]

An O(sr)-Resolution ODE Framework for Understanding Discrete-Time Algo- rithms and Applications to the Linear Convergence of Minimax Problems

H. Lu. “An O(sr)-Resolution ODE Framework for Understanding Discrete-Time Algo- rithms and Applications to the Linear Convergence of Minimax Problems.”Mathematical Programming 194.1-2 2022, pp. 1061–1112.url: https://doi.org/10.1007/s10107- 021-01669-4 (cit. on p. 5)

2022 doi
[33]

The Variational Formulation of the Fully Parabolic Keller–Segel System with Degenerate Diffusion

Y. Mimura. “The Variational Formulation of the Fully Parabolic Keller–Segel System with Degenerate Diffusion.”Journal of Differential Equations263.2 2017, pp. 1477–1521.url: https://www.sciencedirect.com/science/article/pii/S0022039617301559 (cit. on p. 5)

2017
[34]

GradientFlowsandEvolutionVariationalInequalitiesinMet- ric Spaces. I: Structural Properties

M.MuratoriandG.Savaré.“GradientFlowsandEvolutionVariationalInequalitiesinMet- ric Spaces. I: Structural Properties.”Journal of Functional Analysis278.4 2020, p. 108347. url: https://www.sciencedirect.com/science/article/pii/S0022123619303416 (cit. on pp. 4, 9, 12–15)

2020
[35]

On the Arrow–Hurwicz Differential System for Linearly Constrained Convex Minimization

S. K. Niederländer. “On the Arrow–Hurwicz Differential System for Linearly Constrained Convex Minimization.” Optimization 73.7 2024, pp. 2313–2345. eprint: https : / / doi . org/10.1080/02331934.2023.2215799 . url: https://doi.org/10.1080/02331934. 2023.2215799 (cit. on p. 5)

2024
[36]

Note on Non-Cooperative Convex Games

H. Nikaidô and K. Isoda. “Note on Non-Cooperative Convex Games.” Pacific Journal of Mathematics 5 1955, pp. 807–815. url: http://projecteuclid.org/euclid.pjm/ 1171984836 (cit. on p. 4)

1955
[37]

Generalized Barycenters and Variance Maximization on Metric Spaces

B. Pass. “Generalized Barycenters and Variance Maximization on Metric Spaces.”Journal of Fixed Point Theory and Applications25.1 2023, p. 5.url: https://doi.org/10.1007/ s11784-022-01015-x (cit. on p. 6)

2023
[38]

Monotone Operators Associated with Saddle-Functions and Minimax Problems

R. T. Rockafellar. “Monotone Operators Associated with Saddle-Functions and Minimax Problems.” Nonlinear Functional Analysis. XVIII, Part 1. Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI, 1970, pp. 241–250 (cit. on pp. 4, 5)

1970
[39]

Monotone Operators and the Proximal Point Algorithm

R. T. Rockafellar. “Monotone Operators and the Proximal Point Algorithm.”SIAM Jour- nal on Control and Optimization14.5 1976, pp. 877–898. eprint:https://doi.org/10. 1137/0314056. url: https://doi.org/10.1137/0314056 (cit. on p. 5)

1976 doi
[40]

Saddle-Points and Convex Analysis

R. T. Rockafellar. “Saddle-Points and Convex Analysis.”Differential Games and Related Topics (Proc. Internat. Summer School, Varenna, 1970). North-Holland Publishing Co., Amsterdam-London, American Elsevier Publishing Co., Inc., New York, 1971, pp. 109–127 (cit. on pp. 2, 5, 6, 9, 10)

1970
[41]

Convex Functionals of Probability Measures and Nonlinear Diffusions on Manifolds

K.-T. Sturm. “Convex Functionals of Probability Measures and Nonlinear Diffusions on Manifolds.” Journal de Mathématiques Pures et Appliquées84.2 2005, pp. 149–168.url: https://www.sciencedirect.com/science/article/pii/S0021782404001485 (cit. on p. 6). 34 REFERENCES

2005
[43]

Saddle Flow Dynamics: Observable Certificates and Separable Regularization

P. You and E. Mallada. “Saddle Flow Dynamics: Observable Certificates and Separable Regularization.” 2021 American Control Conference (ACC). 2021, pp. 4817–4823 (cit. on p. 5)

2021
[44]

Sion’s Minimax Theorem in Geodesic Metric Spaces and a Riemannian Extragradient Algorithm

P. Zhang, J. Zhang, and S. Sra. “Sion’s Minimax Theorem in Geodesic Metric Spaces and a Riemannian Extragradient Algorithm.”SIAM Journal on Optimization33.4 2023, pp. 2885–2908. eprint:https://doi.org/10.1137/22M1505475. url: https://doi.org/ 10.1137/22M1505475 (cit. on pp. 6,...

2023 doi
[97]

2810–2819.url: https://proceedings.mlr.press/v97/hsieh19b.html (cit

2019, pp. 2810–2819.url: https://proceedings.mlr.press/v97/hsieh19b.html (cit. on p. 6)

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [42]

Open Problem: Convergence of Single-Timescale Mean-Field Langevin Descent-Ascent for Two-Player Zero-Sum Games

G. Wang and L. Chizat. “Open Problem: Convergence of Single-Timescale Mean-Field Langevin Descent-Ascent for Two-Player Zero-Sum Games.”Proceedings of Thirty Seventh Conference on Learning Theory. 247. Proceedings of Machine Learning Research. 2024, pp. 5345–5350. url: https://proceedings.mlr.press/v247/wang24c.html (cit. on pp. 2, 5, 6)

work page 2024

[2] [1]

Gradient Flows in Metric Spaces and in the Space of Probability Measures

L. Ambrosio, N. Gigli, and G. Savaré. “Gradient Flows in Metric Spaces and in the Space of Probability Measures.” 2nd ed. Lectures in Mathematics ETH Zürich. Basel: Birkhäuser Verlag, 2008 (cit. on pp. 2–6, 9, 10, 12, 14, 16, 17, 19, 20, 25–29)

work page 2008

[3] [2]

A Gradient Method for Approximating Saddle Points and Constrained Maxima

K. J. Arrow and L. Hurwicz. “A Gradient Method for Approximating Saddle Points and Constrained Maxima.”Traces and Emergence of Nonlinear Programming. Basel: Springer Basel, 2014, pp. 45–60.url: https://doi.org/10.1007/978-3-0348-0439-4_2 (cit. on pp. 2, 5). REFERENCES 31

work page doi:10.1007/978-3-0348-0439-4_2 2014

[4] [3]

Studies in Linear and Non-Linear Programming

K. J. Arrow, L. Hurwicz, and H. Uzawa. “Studies in Linear and Non-Linear Programming.” II. Stanford Mathematical Studies in the Social Sciences. With contributions by H. B. Chenery, S. M. Johnson, S. Karlin, T. Marschak, R. M. Solow. Stanford University Press, Stanford, CA, 1958, pp. vii+229 (cit. on pp. 2, 5)

work page 1958

[5] [4]

Maximal Monotone Operators Associated with Saddle Functions

H. Asakawa. “Maximal Monotone Operators Associated with Saddle Functions.” TRU Mathematics 22.2 1986, pp. 47–71 (cit. on p. 5)

work page 1986

[6] [5]

Continuous Time Approximation of Nash Equilib- ria in Monotone Games

R. Awi, R. Hynd, and H. Mawi. “Continuous Time Approximation of Nash Equilib- ria in Monotone Games.” Communications in Contemporary Mathematics 26.05 2024, p. 2350021. eprint: https : / / doi . org / 10 . 1142 / S0219199723500219. url: https : //doi.org/10.1142/S0219199723500219 (cit. on p. 5)

work page doi:10.1142/s0219199723500219 2024

[7] [6]

Convex Analysis and Optimization in Hadamard Spaces

M. Bacak. “Convex Analysis and Optimization in Hadamard Spaces.” Berlin, München, Boston: De Gruyter, 2014. url: https : / / doi . org / 10 . 1515 / 9783110361629(cit. on p. 17)

work page 2014

[8] [7]

Uncertainty Quantification of the 4th Kind; Optimal Posterior Accuracy- UncertaintyTradeoffwiththeMinimumEnclosingBall

H. H. Bajgiran, P. Batlle, H. Owhadi, M. Samir, C. Scovel, M. Shirdel, M. Stanley, and P. Tavallali. “Uncertainty Quantification of the 4th Kind; Optimal Posterior Accuracy- UncertaintyTradeoffwiththeMinimumEnclosingBall.” Journal of Computational Physics 471 2022, p. 111608. url: https://www.sciencedirect.com/science/article/pii/ S0021999122006714 (cit. on p. 6)

work page 2022