Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Natural invariant measures impose statistical order on chaotic learning dynamics

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

2026-08-01 06:40 UTC pith:WFHKBIMB

load-bearing objection Sound core with a genuinely useful framing; the 'full-spectrum' catalogue is numerically illustrated, not proven, and the abstract oversells it. the 3 major comments →

arxiv 2607.21805 v1 pith:WFHKBIMB submitted 2026-07-23 math.DS cs.LGecon.TH

Natural Invariant Measures for Chaotic Game Dynamics: Finding Order in Chaos

Jakub Bielawski , Thiparat Chotibut , Fryderyk Falniowski , Micha{\l} Misiurewicz , Georgios Piliouras This is my paper
classification math.DS cs.LGecon.TH MSC 37E0537A0537D4591A26
keywords multiplicative weights updatenatural invariant measureschaotic dynamicscongestion gamesergodic theoryinterval mapsNash equilibriumsocial cost
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 claims that chaotic multiplicative-weights learning in a two-strategy congestion game has a rigorous statistical description through natural invariant measures. It proves that even when trajectories never converge to Nash equilibrium, long-run time averages of play equal the Nash fraction b. Any invariant measure that does not charge the boundary states gives the same average for affine observables, and bounds for convex and concave ones. Consequently, social cost and regret have well-defined long-run averages despite chaos. It also shows that this learning algorithm realizes the full range of one-dimensional dynamical behavior, including unique or multiple absolutely continuous invariant measures, periodic attractors, and coexisting chaotic and stable regimes.

Core claim

For the two-strategy congestion game under multiplicative weights update, the dynamics reduces to the map f_MW(x)=x/(x+(1-x)exp(a(x-b))). Through a logarithmic change of variables it becomes the family F(y)=y+b-1/(exp(-ay)+1). The paper proves that for every initial state x in (0,1), the time average of play converges to the Nash fraction b even when Li-Yorke chaos prevents pointwise convergence. Consequently, any invariant probability measure that assigns zero mass to the endpoints has affine-observable average equal to the equilibrium value, while convex and concave observables satisfy Jensen-type inequalities. When a natural invariant measure exists, these space averages equal long-run ti

What carries the argument

The central object is the multiplicative-weights map f_MW and its logarithmic conjugate F(y)=y+b-1/(exp(-ay)+1). The conjugacy turns the strategy-fraction dynamics into a bimodal interval map with negative Schwarzian derivative and two nondegenerate critical points when a>4. Natural invariant measures, defined as invariant probability measures obtained as limits of time-averages from absolutely continuous reference measures, act as the statistical attractors. For negative-Schwarzian interval maps, the support of an absolutely continuous invariant measure must be a cycle of intervals, and each attracting periodic orbit or such measure is fed by one of the two critical points. This dichotomy o

Load-bearing premise

The paper leans on an unproved dichotomy that for almost all parameter values the dynamics is either an attracting periodic orbit or supports an absolutely continuous invariant measure, imported from the quadratic family; if a third statistically relevant behavior appears on a positive-measure parameter set, the full-spectrum conclusions no longer follow.

What would settle it

In the parameter rectangle a in [27.5,40], b in [0.25,0.5], simulate long orbits seeded at the two critical points and construct the empirical invariant measure. If a positive-area subregion shows an attractor that is neither an attracting periodic orbit nor a measure with a density with respect to Lebesgue measure, for instance a histogram with persistent fractal support that is not a cycle of intervals, then the two-type catalogue fails.

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

If this is right

  • Long-run average play in the chaotic regime equals the Nash equilibrium fraction b, so equilibrium predictions survive in a time-averaged sense even when trajectories never settle.
  • Time-averaged social cost is well defined and is bounded below by the social cost at the Nash equilibrium; similarly, time-averaged regret is bounded below by its equilibrium value.
  • For rational b=k/n with sufficiently large learning rate a, Lebesgue-almost every trajectory is attracted to a periodic orbit of period n, and all continuous observables exhibit the same long-run time average as that orbit.
  • The paper's catalogue shows that MWU in a two-strategy congestion game can realize unique or multiple absolutely continuous invariant measures, as well as coexisting chaotic and periodic attractors, matching the known range of interval-map behavior.
  • If the natural-measure framework holds, numerical histograms from long simulations are not artifacts; they approximate the true statistical law of the chaotic learning process.

Where Pith is reading between the lines

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

  • A testable extension: treat the two critical points as probes of invariant statistics. For parameters where both lie in the same invariant cycle of intervals, the empirical histogram should be independent of which critical point seeds the orbit, while disjoint cycles should produce distinct coexisting statistical laws.
  • The averaging theorems depend mainly on the fact that time averages converge to the Nash fraction; any learning rule whose reduced one-dimensional dynamics satisfies the same hypotheses would inherit the same form of statistical predictability.
  • If the two-type dichotomy holds for this family, bifurcation scans become practical statistical phase maps: they show which critical point feeds which natural measure, and therefore which learning rates make long-run welfare predictable.
  • An economic reading the paper leaves implicit: chaotic learning does not necessarily undermine price-of-anarchy reasoning, because the relevant averages remain tied to equilibrium benchmarks even when individual trajectories are unpredictable.

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

3 major / 6 minor

Summary. The paper studies the long-run statistical behavior of the Multiplicative Weights Update (MWU) algorithm in a two-strategy congestion game, modeled by the one-dimensional family of maps f_MW (the (MW) family). The theoretical core (Section 4) proves general results relating time-averages of continuous observables to invariant measures: Theorem 4.1 constructs an invariant measure on the omega-limit set from any convergent subsequence of time-averages; Theorem 4.2 shows that if the time-average trajectory tends to a unique interior fixed point p, then every invariant measure gives O(p) for affine observables and satisfies the Jensen inequality for convex/concave observables; Proposition 4.3 shows that an attracting periodic orbit controls time-averaged observations for almost every initial condition; Proposition 4.4 constrains the support of an absolutely continuous ergodic invariant measure to a cycle of intervals. These results are then applied to f_MW. Corollary 5.2 states that for any invariant measure not charging the boundary, affine observables integrate to their Nash value and convex/concave observables satisfy the corresponding inequality; Corollary 5.4 describes the periodic-orbit case; Corollary 5.5 gives the cycle-of-intervals support. The paper then presents a catalogue of possible behaviors (attracting periodic orbits, absolutely continuous invariant measures, and coexisting cycles) with numerical experiments, and discusses social cost and regret. The advertised '

Significance. If fully established, the framework would be valuable: it gives a rigorous way to compute long-run averages of economic observables in a setting where MWU dynamics are chaotic, and the convex-observable inequalities are a clean addition. The theoretical results in Section 4 are transparent and correct under the natural reading, and Corollaries 5.2 and 5.4-5.5 follow from the cited prior work. The paper also makes a useful connection between game dynamics and one-dimensional ergodic theory, and its numerical experiments illustrate the proposed phenomena. However, the headline claim that the (MW) family exhibits the 'full spectrum' of one-dimensional dynamical behaviors is not proven. The dichotomy into attracting periodic orbits versus absolutely continuous invariant measures is imported from a suspicion rather than established for this family, and the existence of parameters for several catalogue cases rests on visual monotonicity claims and heuristic bifurcation arguments. The proven contribution is therefore the ergodic-theoretic framework and the conditional consequences, while the comprehensive classification remains conjectural.

major comments (3)
  1. [Section 5.2] The classification into exactly two long-run behaviors (attracting periodic orbit or absolutely continuous invariant measure) is central to the paper's 'full spectrum' claim, but it is not proven for the (MW) family. The paper states: 'There are reasons [13,15] to suspect that similar properties hold for other "natural" families of maps with negative Schwarzian derivative and nondegenerate critical points, so we will concentrate on those two types of behavior.' This is an explicit assumption, not a theorem. If the (MW) family admits other statistically relevant invariant measures (e.g., wild attractors or non-ergodic measures with positive Lebesgue measure basins), then the subsequent catalogue and the abstract's 'comprehensive statistical characterization' do not follow. The authors must either prove a Lyubich-type dichotomy for the (MW) family or substantially weaken the claims to cond
  2. [Section 5.3.2 / Lemma 5.6] The only proposed proof of existence of an absolutely continuous invariant measure (case 1.b.i) is not rigorous. The application of Lemma 5.6 requires verifying that the map G=(F^3(y_l)-y_f, F^2(y_r)-y_l) maps the rectangle boundary homeomorphically onto a Jordan curve with the origin in the bounded component. The verification is based on the visual assertion 'all functions whose graphs we see in the figure are monotone' and on quadrant observations from the plots. The paper itself acknowledges the initial evidence is 'very week [sic] numerical evidence.' Without a rigorous computer-assisted verification (e.g., interval arithmetic) of the boundary behavior, the existence of parameters (a,b) with the required critical orbit condition is not established. This is load-bearing because it is the sole existence argument for a unique a.c.i.m. in the (MW) family.
  3. [Sections 5.3.3 and 5.4.2] Cases 1.b.ii and 2.c are asserted on the basis of heuristic arguments rather than proofs. Section 5.3.3 says that as a varies 'it should go through the usual unimodal bifurcations [21,22], so there are values of a for which there is an absolutely continuous invariant measure,' and Section 5.4.2 claims that 'for the suitably chosen preperiodic combinatorial patterns' there is a parameter value with two coexisting a.c.i.m.s. These are plausible numerical observations, not mathematically verified statements. Consequently, the claim in Section 5.2 that the game dynamics 'demonstrates all of these behaviors' is not supported by the evidence presented. The paper should either provide rigorous existence proofs (e.g., via the same critical-orbit conditions and validated numerics) or explicitly mark these cases as conjectures supported by simulations.
minor comments (6)
  1. [Section 5.3.2] Typo: 'very week numerical evidence' should be 'very weak numerical evidence.'
  2. [Section 2] The Schwarzian derivative formula is written as Sf= f'''/f' − 3/2 (f''/f')^2; add parentheses to avoid ambiguity: Sf = f'''/f' − (3/2)(f''/f')^2.
  3. [Theorem 4.2] The assumption that (3) holds 'for all x except for a set of measure zero' should explicitly say 'with respect to μ' in the theorem statement. The footnote clarifies the intention, but stating it in the theorem avoids confusion, especially since the result is later applied to different invariant measures.
  4. [Section 5.5.2] The statement that the limit of the time-average regret is equal to the space average of (x−b)^2 with respect to an invariant probability measure is only valid for subsequential limits unless convergence of the time-average is established. The phrase 'provided this limit exists' appears, but the sentence as written is stronger than Theorem 4.1 guarantees. Also, the conclusion that regret is 'lowest in the Nash equilibrium play' reduces to ∫(x−b)^2 dμ ≥ 0, which is a triviality; consider clarifying the intended content.
  5. [Figure 3 captions] The caption 'The (a,b)-plane' should be more descriptive, e.g., 'The rectangle in the (a,b)-plane used for the boundary check in Section 5.3.2.'
  6. [Section 5.4.2] The phrase 'there are a lot (countably many) values of b' is informal. If a precise statement is intended, it should be justified or rephrased as 'infinitely many' or 'a countable set' with a reference.

Circularity Check

0 steps flagged

No constructional circularity; the statistical conclusions rest on prior published theorems, and the two-type catalogue is explicitly a suspicion rather than a definitional shortcut.

full rationale

I find no circularity in the claimed derivation chain. Corollary 5.2 is obtained by applying the general Theorem 4.2, with the needed time-average convergence supplied by the prior published result [20]; although [20] shares authors with the present paper, it is a separate, externally checkable theorem and is not a fitted input. In fact, the identity ∫x dμ = b follows directly from invariance of the conjugate map F(y)=y+b−1/(e^{-ay}+1), so the central equality is not an artifact of the paper's own construction. Corollary 5.4 likewise imports Lemma 5.1 from the independent published paper [11], and Proposition 4.4 uses standard results [13,15]. The Section 5.2 restriction to only attracting periodic orbits or absolutely continuous invariant measures is explicitly introduced as a 'suspicion' ('There are reasons [13,15] to suspect that similar properties hold for other "natural" families ... so we will concentrate on those two types of behavior'), so the full-spectrum catalogue is conditional on an unproven dichotomy, which is a correctness/completeness limitation rather than circularity. The existence argument for an absolutely continuous invariant measure in Section 5.3.2 relies on visual monotonicity and the paper's own 'very week numerical evidence'; again, that is a rigor gap, not a reduction to the conclusion. No fitted parameters are relabeled as predictions, and no equation is shown to be equivalent to its own input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 9 axioms · 0 invented entities

The paper introduces no fitted parameters and no new physical/mathematical entities; the natural invariant measure is a standard ergodic-theory object. The ledger is dominated by imports: the measurable-observable bounds rest on the authors' own [20] and the period-n attractors on their own [11]; the classification's completeness rests on a disclosed but unproven Lyubich-dichotomy extension; and the unique-a.c.i.m. existence rests on Misiurewicz's theorem applied to a bimodal map without explicit verification of the theorem's hypotheses. These are legitimate as published results, but they are assumptions relative to this paper's own derivations.

axioms (9)
  • standard math Birkhoff ergodic theorem, Krylov-Bogolyubov existence of invariant measures, Krein-Milman and Banach-Alaoglu theorems
    Used without proof throughout Section 4 (proofs of Theorems 4.1 and 4.2).
  • standard math Blank-Bunimovich definition of natural measure, plus the reduction to checking one measure ([39])
    Section 3; the paper's central object is imported from [12], with [39] justifying the single-measure check for ergodic natural measures.
  • standard math Negative-Schwarzian one-dimensional theory: attracting periodic orbits attract a critical point [22]; Blokh-Lyubich decomposition [13]; Bruin-Lopez Proposition 24 [15] (wild attractors have Lebesgue measure zero)
    Load-bearing for Proposition 4.4 and for the structural claims of the catalogue in Section 5.2.
  • domain assumption Lyubich dichotomy (regular or stochastic) for the quadratic family extends to the (MW) family: for Lebesgue-a.e. parameter there is either an attracting periodic orbit or an absolutely continuous invariant measure
    Section 5.2: 'There are reasons [13,15] to suspect...' - disclosed as a suspicion, not a theorem. Underpins the completeness of the behaviour catalogue and the statistical-predictability narrative.
  • domain assumption Misiurewicz condition ([38]) implies existence of an absolutely continuous invariant measure for the (MW) map
    Section 5.3.2. The theorem is cited for the preperiodic-to-repelling-fixed-point configuration; whether [38] covers the bimodal (two-critical-point) case with both critical points non-recurrent is not verified in the text.
  • domain assumption The time-average of MWU play converges to b for every x in (0,1) ([20])
    Imported from the authors' prior work; the load-bearing input for Corollary 5.2 and the Figure 9 claim that the time average of x is b.
  • domain assumption For b=k/n with coprime k,n, the (MW) map has an attracting period-n orbit attracting Lebesgue-a.e. initial points for sufficiently large a ([11], Lemma 5.1)
    Imported from the authors' prior work; load-bearing for Corollary 5.4 (time-averages converge to the periodic-orbit average).
  • domain assumption Regret formula: lim R_T/T = N times the limit of the time-average of (x-b)^2 ([19])
    Imported from the authors' prior NeurIPS paper; connects the convex observable results to regret in Section 5.5.2.
  • standard math Jordan-Schoenflies theorem (in the proof of Lemma 5.6)
    Used to normalize the boundary map in the covering argument that proves the zero-crossing existence for the unique a.c.i.m. case.

pith-pipeline@v1.3.0-alltime-deepseek · 28081 in / 37848 out tokens · 368692 ms · 2026-08-01T06:40:46.952670+00:00 · methodology

0 comments
read the original abstract

We study the long-term behavior of the Multiplicative Weights Update (MWU) algorithm in game settings where learning dynamics frequently fail to converge to Nash equilibria and instead exhibit Li-Yorke chaos. While such chaos precludes the prediction of specific long-term strategy profiles, it does not imply a lack of statistical structure. We demonstrate that natural invariant measures - a fundamental concept from ergodic theory - provide the rigorous framework necessary to find order within this chaos. Focusing on a two-strategy congestion game, we prove that these measures allow for a comprehensive statistical characterization of the dynamics. Crucially, we show that this framework extends beyond simple strategy frequencies to \emph{general observables}, enabling the precise calculation of long-term time averages for broad classes of economic metrics - including payoffs, social cost, and regret - despite chaos. Our results reveal that this simple learning algorithm captures the full spectrum of behaviors found in one-dimensional dynamical systems, from unique or multiple absolutely continuous invariant measures to complex periodic attractors as well as coexisting chaotic and stable (periodic) behaviors. By bridging game theory and dynamical systems, we show that statistical predictability is attainable even in the absence of pointwise convergence.

Figures

Figures reproduced from arXiv: 2607.21805 by Fryderyk Falniowski, Georgios Piliouras, Jakub Bielawski, Micha{\l} Misiurewicz, Thiparat Chotibut.

Figure 1
Figure 1. Figure 1: Various periodic orbits for various parameter values. The horizontal axis is [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: 𝐹 2 (𝑦𝑟) = 𝑦𝑙 and 𝐹 3 (𝑦𝑙) = 𝑦𝑓 . However, this is a very week numerical evidence, since small changes in parameters can change the picture in an unpredictable way. To get a stronger evidence, consider the rectangle in the (𝑎, 𝑏)-plane with sides at 𝑎 = 27.5, 𝑎 = 40, 𝑏 = 0.25, 𝑏 = 0.5 (see Figures 3a–3d). For the four sides of the rectangle, we draw the graphs of 𝐹 3 (𝑦𝑙) − 𝑦𝑓 (black) and 𝐹 2 (𝑦𝑟) − 𝑦𝑙 (gr… view at source ↗
Figure 3
Figure 3. Figure 3: The (𝑎, 𝑏)-plane [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Here 𝑏 = 0.38 and 𝑎 (on the horizontal axis) varies from 26.5 to 27. The positions of critical points are marked in red and yellow. 6 (see also Figures 6a and 6c). For 𝑏 = 0.3868, in the period 4 cycle there is an attracting periodic orbit of period 4 (Figures 6b and 6d) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Bifurcation diagram for 𝑎 = 26 and 𝑏 (on the horizontal axis) varies from 0.386 to 0.39. The attracting sets in the cycles of intervals are in black (period 6 cycle) and blue (period 4 cycle). The positions of critical points are marked in red. 5.4.2 Two measures. Let us find case 2.c. Figures 7 and 8 show bifurcation diagrams for 𝑎 = 27 and 𝑎 = 27.01 respectively, as 𝑏 varies from 0.386 to 0.388. In those… view at source ↗
Figure 6
Figure 6. Figure 6: Two disjoint cycles of intervals for 𝑎 = 26, 𝑏 = 0.3868. of 𝑎 between 27 and 27.01 for which both critical points follow those patterns. For this value of 𝑎 there are two absolutely continuous invariant measures with disjoint supports (in the red and black areas). 5.5 Example observables in economics Here we present two economic/machine learning notions for which there exist observables that can be applied… view at source ↗
Figure 7
Figure 7. Figure 7: Here 𝑎 = 27 and 𝑏 (on the horizontal axis) varies from 0.386 to 0.388 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Here 𝑎 = 27.01 and 𝑏 (on the horizontal axis) varies from 0.386 to 0.388. Despite the fact that for sufficiently large 𝑎 the fraction 𝑥𝑛 may not converge to the equilibrium, the natural invariant measure discussed earlier ensures that the time-averaged cost still converges to a well-defined value. Intuitively, while the system may enter limit cycles or chaotic orbits for large 𝑎 (and hence never settle exa… view at source ↗
Figure 9
Figure 9. Figure 9: An illustration of the cost function, cobweb diagrams, and trajectories for the fraction of agents using [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

57 extracted references · 2 canonical work pages

  1. [1]

    Ioannis Anagnostides, Ian Gemp, Georgios Piliouras, and Kelly Spendlove. 2026. Chaos in Autobidding Auctions. In Proceedings of the 27th ACM Conference on Economics and Computation (EC ’26)

    2026

  2. [2]

    Gabriel P Andrade, Rafael Frongillo, and Georgios Piliouras. 2021. Learning in matrix games can be arbitrarily complex. InConference on Learning Theory. PMLR, 159–185

    2021

  3. [3]

    Andrade, Rafael Frongillo, and Georgios Piliouras

    Gabriel P. Andrade, Rafael Frongillo, and Georgios Piliouras. 2023. No-Regret Learning in Games is Turing Complete. InProceedings of the 24th ACM Conference on Economics and Computation. Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras19

    2023

  4. [4]

    Sanjeev Arora, Elad Hazan, and Satyen Kale. 2012. The Multiplicative Weights Update Method: a Meta-Algorithm and Applications.Theory of Computing8, 1 (2012), 121–164

    2012

  5. [5]

    Brian Arthur

    W. Brian Arthur. 1994. Inductive reasoning and bounded rationality.The American economic review84, 2 (1994), 406–411

    1994

  6. [6]

    Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. 1995. Gambling in a rigged casino: The adversarial multi-armed bandit problem. InProceedings of IEEE 36th annual foundations of computer science. IEEE, 322–331

    1995

  7. [7]

    James P Bailey and Georgios Piliouras. 2018. Multiplicative weights update in zero-sum games. InProceedings of the 2018 ACM Conference on Economics and Computation. 321–338

    2018

  8. [8]

    Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Grzegorz Kosiorowski, Michał Misiurewicz, and Georgios Piliouras. 2021. Follow-the-regularized-leader routes to chaos in routing games. InInternational Conference on Machine Learning. PMLR, 925–935

    2021

  9. [9]

    Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras. 2024. Memory loss can prevent chaos in games dynamics.Chaos: An Interdisciplinary Journal of Nonlinear Science34, 1 (2024)

    2024

  10. [10]

    Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras. 2025. Hetero- geneity, reinforcement learning, and chaos in population games.Proceedings of the National Academy of Sciences122, 25 (2025), e2319929121

    2025

  11. [11]

    Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras. 2025. Interval maps mimicking circle rotations.Communications in Nonlinear Science and Numerical Simulation(2025), 108963

    2025

  12. [12]

    Michael Blank and Leonid Bunimovich. 2003. Multicomponent dynamical systems: SRB measures and phase transitions. Nonlinearity16, 1 (2003), 387–401

    2003

  13. [13]

    Alexander Blokh and Mikhail Lyubich. 1990. On decomposition of one-dimensional dynamical systems into ergodic components. The case of negative Schwarzian.Leningr. Math. J1 (1990), 137–155

    1990

  14. [14]

    George W Brown. 1951. Iterative solution of games by fictitious play.Activity analysis of production and allocation13, 1 (1951), 374–376

    1951

  15. [15]

    Henk Bruin and Víctor Jiménez López. 2010. On the Lebesgue measure of Li-Yorke pairs for interval maps.Communi- cations in Mathematical Physics299, 2 (2010), 523–560

    2010

  16. [16]

    Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolo, and Sergio Pastorello. 2020. Artificial intelligence, algorithmic pricing, and collusion.American Economic Review110, 10 (2020), 3267–3297

    2020

  17. [17]

    Yun Kuen Cheung and Georgios Piliouras. 2020. Chaos, extremism and optimism: Volume analysis of learning in games.Advances in Neural Information Processing Systems33 (2020), 9039–9049

    2020

  18. [18]

    Yun Kuen Cheung, Georgios Piliouras, and Yixin Tao. 2022. The evolution of uncertainty of learning in games. In International Conference on Learning Representations

    2022

  19. [19]

    Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras. 2020. The route to chaos in routing games: When is Price of Anarchy too optimistic?Advances in Neural Information Processing Systems33 (2020), 766–777

    2020

  20. [20]

    Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras. 2021. Family of chaotic maps from game theory.Dynamical Systems36, 1 (2021), 48–63

    2021

  21. [21]

    Pierre Collet and J-P Eckmann. 1983. Positive Liapunov exponents and absolute continuity for maps of the interval. Ergodic Theory and Dynamical Systems3, 1 (1983), 13–46

    1983

  22. [22]

    2012.One-dimensional dynamics

    Welington De Melo and Sebastian Van Strien. 2012.One-dimensional dynamics. Vol. 25. Springer Science & Business Media

    2012

  23. [23]

    2010.Computational topology: an introduction

    Herbert Edelsbrunner and John Harer. 2010.Computational topology: an introduction. American Mathematical Soc

    2010

  24. [24]

    2011.Ergodic Theory with a view towards Number Theory

    Manfred Einsiedler and Thomas Ward. 2011.Ergodic Theory with a view towards Number Theory. Springer London

    2011

  25. [25]

    1996.Chaotic regimes in a dynamical system of the type many predators-one prey

    Timo Eirola, Alexandr V Osipov, and Gunnar Söderbacka. 1996.Chaotic regimes in a dynamical system of the type many predators-one prey. Helsinki University of Technology, Institute of Mathematics

    1996

  26. [26]

    Fryderyk Falniowski and Panayotis Mertikopoulos. 2025. On the discrete-time origins of the replicator dynamics: From convergence to instability and chaos.International Journal of Game Theory54, 1 (2025), 7

    2025

  27. [27]

    J Doyne Farmer and Duncan Foley. 2009. The economy needs agent-based modelling.Nature460, 7256 (2009), 685–686

    2009

  28. [28]

    Paweł Góra and Abraham Boyarsky. 1988. Why computers like Lebesgue measure.Computers & Mathematics with Applications16, 4 (1988), 321–329

    1988

  29. [29]

    Timo Klein. 2021. Autonomous algorithmic collusion: Q-learning under sequential pricing.The RAND Journal of Economics52, 3 (2021), 538–558

    2021

  30. [30]

    Sergey Kryzhevich, Viktor Avrutin, and Gunnar Söderbacka. 2021. Bistability in a one-dimensional model of a two-predators-one-prey population dynamics system.Lobachevskii Journal of Mathematics42, 14 (2021), 3486–3496

    2021

  31. [31]

    Stefanos Leonardos, Barnabé Monnot, Daniël Reijsbergen, Efstratios Skoulakis, and Georgios Piliouras. 2021. Dynamical analysis of the eip-1559 ethereum fee market. InProceedings of the 3rd ACM Conference on Advances in Financial Technologies. 114–126. Jakub Bielawski, Thiparat Chotibut, Fryderyk Falniowski, Michał Misiurewicz, and Georgios Piliouras20

    2021

  32. [32]

    Stefanos Leonardos, Daniël Reijsbergen, Barnabé Monnot, and Georgios Piliouras. 2023. Optimality despite chaos in fee markets. InInternational Conference on Financial Cryptography and Data Security. Springer, 346–362

    2023

  33. [33]

    David Levy. 1994. Chaos theory and strategy: Theory, application, and managerial implications.Strategic management journal15, S2 (1994), 167–178

    1994

  34. [34]

    Tien-Yien Li and James A. Yorke. 1975. Period Three Implies Chaos.The American Mathematical Monthly82, 10 (1975), 985–992

    1975

  35. [35]

    Wayne Lin, Georgios Piliouras, Ryann Sim, and Antonios Varvitsiotis. 2024. No-Regret Learning and Equilibrium Computation in Quantum Games.Quantum8 (2024), 1569

    2024

  36. [36]

    Mikhail Lyubich. 2002. Almost every real quadratic map is either regular or stochastic.Annals of Mathematics, Second Series156, 1 (2002), 1–78. https://doi.org/10.2307/3597183

    work page doi:10.2307/3597183 2002

  37. [37]

    Konstantinos Metaxas and Themistoklis P Sapsis. 2026. Risk-Sensitive Learning in Population Games under Extreme Events: Bifurcations and Chaotic Dynamics.arXiv preprint arXiv:2606.29967(2026)

    Pith/arXiv arXiv 2026

  38. [38]

    Michał Misiurewicz. 1981. Absolutely continuous measures for certain maps of an interval.Publications Mathématiques de l’Institut des Hautes Études Scientifiques53, 1 (1981), 17–51

    1981

  39. [39]

    Michał Misiurewicz. 2005. Ergodic natural measures. InAlgebraic and Topological Dynamics, S. Kolyada, Y. Manin, and T. Ward (Eds.). Contemporary Mathematics, Vol. 385. American Mathematical Society, Providence, RI, 1–6

    2005

  40. [40]

    W Garrett Mitchener and Martin A Nowak. 2004. Chaos and language.Proceedings of the Royal Society of London. Series B: Biological Sciences271, 1540 (2004), 701–704

    2004

  41. [41]

    Archan Mukhopadhyay and Sagar Chakraborty. 2020. Deciphering chaos in evolutionary games.Chaos: An Interdisci- plinary Journal of Nonlinear Science30, 12 (2020)

    2020

  42. [42]

    Sai Ganesh Nagarajan, David Balduzzi, and Georgios Piliouras. 2020. From chaos to order: Symmetry and conservation laws in game dynamics. InInternational Conference on Machine Learning. PMLR, 7186–7196

    2020

  43. [43]

    Vazirani

    Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V. Vazirani. 2007.Algorithmic Game Theory. Cambridge University Press, New York, NY, USA

    2007

  44. [44]

    Martin A Nowak and Robert M May. 1992. Evolutionary games and spatial chaos.nature359, 6398 (1992), 826–829

    1992

  45. [45]

    Gerasimos Palaiopanos, Ioannis Panageas, and Georgios Piliouras. 2017. Multiplicative weights update with constant step-size in congestion games: Convergence, limit cycles and chaos. InAdvances in Neural Information Processing Systems. 5872–5882

    2017

  46. [46]

    Telmo Peixe and Alexandre Rodrigues. 2022. Persistent strange attractors in 3D polymatrix replicators.Physica D: Nonlinear Phenomena438 (2022), 133346

    2022

  47. [47]

    Georgios Piliouras, Ian Gemp, Siqi Liu, and Luke Marris. 2026. Paradoxes of Game Theoretic Equilibria and Price of Anarchy.arXiv preprint arXiv:2607.11752(2026)

    Pith/arXiv arXiv 2026

  48. [48]

    Georgios Piliouras and Fang-Yi Yu. 2023. Multi-agent performative prediction: From global stability and optimality to chaos. InProceedings of the 24th ACM Conference on Economics and Computation. 1047–1074

    2023

  49. [49]

    Rosenthal

    Robert W. Rosenthal. 1973. A class of games possessing pure-strategy Nash equilibria.International Journal of Game Theory2, 1 (1973), 65–67

    1973

  50. [50]

    James B. T. Sanders, J. Doyne Farmer, and Tobias Galla. 2018. The prevalence of chaotic dynamics in games with many players.Scientific Reports8 (2018)

    2018

  51. [51]

    Sandholm

    William H. Sandholm. 2010.Population Games and Evolutionary Dynamics. MIT Press

    2010

  52. [52]

    Doyne Farmer

    Yuzuru Sato, Eizo Akiyama, and J. Doyne Farmer. 2002. Chaos in learning a simple two-person game.Pro- ceedings of the National Academy of Sciences99, 7 (2002), 4748–4751. https://doi.org/10.1073/pnas.032086299 arXiv:http://www.pnas.org/content/99/7/4748.full.pdf+html

    work page doi:10.1073/pnas.032086299 2002

  53. [53]

    2012.Online learning and online convex optimization

    Shai Shalev-Shwartz. 2012.Online learning and online convex optimization. Vol. 4. Now Publishers, Inc. 107–194 pages

    2012

  54. [54]

    Brian Skyrms. 1992. Chaos and the explanatory significance of equilibrium: Strange attractors in evolutionary game dynamics. InPSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association, Vol. 1992. Philosophy of Science Association, 374–394

    1992

  55. [55]

    Peter D Taylor and Leo B Jonker. 1978. Evolutionary stable strategies with two types of player.Mathematical biosciences 40, 1-2 (1978), 145–156

    1978

  56. [56]

    Christopher JCH Watkins and Peter Dayan. 1992. Q-learning.Machine learning8 (1992), 279–292

    1992

  57. [57]

    Lai-Sang Young. 2002. What are SRB measures, and which dynamical systems have them?Journal of Statistical Physics 108 (2002), 733–754

    2002