Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Static equilibria and Price of Anarchy hide dynamical instability and can certify non-rationalizable or chaotic multi-agent learning.

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 · grok-4.5

2026-07-14 03:25 UTC pith:YHQVDUWE

load-bearing objection Strong constructive package on dynamical instability of PoA anchors and CCE/SCE; two headline paradoxes rest on contested model relaxations that the paper treats as definitional. the 3 major comments →

arxiv 2607.11752 v1 pith:YHQVDUWE submitted 2026-07-13 cs.GT cs.LGcs.MAmath.DSmath.OC

Paradoxes of Game Theoretic Equilibria and Price of Anarchy

Georgios Piliouras , Ian Gemp , Siqi Liu , Luke Marris This is my paper
classification cs.GT cs.LGcs.MAmath.DSmath.OC MSC 91A1091A1491A2668Q2537N40
keywords Price of AnarchyNash equilibriumCoarse Correlated Equilibriumno-regret learningcongestion gamesstrict saddlesLi-Yorke chaosswap regret
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 argues that decades of algorithmic game theory have reduced multi-agent learning to static solution concepts—Nash, Correlated, and Coarse Correlated Equilibria—and to black-box no-regret guarantees, then used those objects to certify system efficiency via the Price of Anarchy. That reduction discards the continuous geometry of learning vector fields. The worst-case pure Nash equilibria that make robust PoA bounds tight are topologically unstable strict saddles (sometimes global repellers supported on almost-everywhere dominated strategies). Interior mixed Nash carry only C0 information and cannot tell cooperative from adversarial incentives. Once cost functions are allowed any strictly positive, increasing affine form (including negative intercepts that still map loads to positive costs), PoA becomes unbounded even in two-player, two-strategy congestion games. Projecting continuous trajectories onto empirical play yields CCE and proximal refinements that can place all mass on strictly dominated strategies; optimal O(1/T) swap-regret can coexist with chaotic limit sets. In the non-atomic polynomial setting long regarded as stable, discrete-time learning turns the unique Wardrop equilibrium into Li-Yorke chaos whose time-averaged social cost is exponentially worse (factor 2^p) than the classical sub-linear static bound. The authors therefore call for efficiency metrics that track dynamically accessible attractors rather than algebraically feasible fixed points.

Core claim

Reducing multi-agent learning to static NE/CE/CCE and black-box regret systematically obscures dynamic disequilibrium: the pure NE that tighten robust PoA are topologically unstable strict saddles or global repellers, PoA for affine congestion is unbounded once all strictly positive affine costs are admitted, CCE and SCE/PCE can be supported entirely on strictly dominated strategies, optimal swap-regret can coexist with chaos, and discrete-time non-atomic polynomial congestion yields Li-Yorke chaos with time-averaged inefficiency 2^p.

What carries the argument

Topological analysis of learning vector fields and potential landscapes (strict saddles via analytic extension, zero-trace Hessians of mixed NE, Chetaev instability) together with explicit constructions that embed non-rationalizable or chaotic orbits inside classical equilibrium polytopes and discrete-time maps.

Load-bearing premise

The paper judges classical PoA after allowing any latency that is positive and non-decreasing only on integer loads (including negative intercepts) and after agents use ordinary responsive step sizes rather than rates globally fine-tuned by total population and degree.

What would settle it

Exhibit a natural discrete-time learning rule, with step sizes set only from locally observed delays and without oracle knowledge of N or p, that converges to the Wardrop equilibrium of a two-path monomial network of degree p while keeping time-averaged social cost within the classical Theta(p/ln p) factor of optimum for large p.

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

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 / 5 minor

Summary. The paper argues that reducing multi-agent learning to static NE/CE/CCE and black-box regret obscures dynamic disequilibrium. It claims: (i) worst-case pure NE that tighten robust PoA are strict saddles of the extended potential (Theorems 1–2), sometimes global repellers on almost-everywhere dominated strategies; (ii) interior NE are C0-insensitive and can lock agents into min-max safety payoffs even in cooperative games (Theorems 3–6); (iii) PoA for affine congestion is unbounded once all strictly positive, nondecreasing affine costs on the discrete domain x≥1 are allowed (Theorem 7), and average PoA can likewise be made unbounded by translations (Theorem 8); (iv) CCE and SCE/PCE can support strictly dominated strategies, and optimal O(1/T) swap-regret can coexist with chaotic limit sets (Theorems 9–11); (v) discrete-time learning in non-atomic polynomial congestion yields Li-Yorke chaos with time-averaged inefficiency 2^p versus static Θ(p/ln p) PoA (Theorem 12). Appendices A–F supply constructive games, Hessians, Chetaev arguments, LP certificates, bipartite mirroring, and period-2 attractor analysis.

Significance. If the results hold under the paper’s modeling choices, they give a coherent geometric critique of anchoring efficiency to worst-case equilibria and of treating time-averaged CE/CCE as a full proxy for rational multi-agent learning. Strengths include explicit constructive counterexamples, closed-form Hessians for the canonical 5/2 PoA games, Chetaev-based instability, primal-dual certificates for SCE/PCE, a bipartite-mirroring reduction that restores multilinearity while preserving a chaotic manifold, and an exact 2^p quantification of discrete-time inefficiency in non-atomic polynomial routing. These are concrete, checkable contributions that extend prior chaos-in-games work (linear to higher degree; MWU/FTRL to L2/PGD) and force a sharper distinction between algebraic feasibility and dynamical accessibility. The policy implication—prefer dynamically grounded metrics over pure worst-case PoA—is significant for AGT and multi-agent learning if the community accepts the model relaxations.

major comments (3)
  1. §4.2 and Theorem 7: Unbounded pure/mixed PoA is obtained only after allowing affine intercepts that may be negative while costs remain positive and nondecreasing on the discrete domain x≥1. Classical robust-PoA statements (Christodoulou–Koutsoupias; Roughgarden smoothness) require a_e,b_e≥0. The manuscript treats this as a natural data-driven relaxation, but the claim that “the PoA becomes unbounded” is then a statement about an enlarged model class, not an internal contradiction of the classical theorems. The paper should either (a) restate Theorem 7 as “PoA is unbounded for the class of everywhere-positive nondecreasing affine latencies on x≥1,” with a clear comparison to the non-negative-coefficient subclass, or (b) give a quantitative continuity/robustness result showing that classical bounds degrade under small negative intercepts that still keep costs positive on the active domain.
  2. Theorem 12 and Remark 4: The Li-Yorke chaos and exact 2^p time-averaged inefficiency hold when the effective step-size a≈ε N^p exceeds the local stability threshold 2^{p+1}/p, i.e., when agents do not globally fine-tune ε=O(N^{-p}). The paper correctly notes that such fine-tuning requires oracle knowledge of N and p and is at odds with decentralized responsive learning. Still, much of the classical no-regret routing literature normalizes rates by maximum possible cost precisely to obtain convergence. The manuscript should state the theorem as a dichotomy (stable under oracle-normalized rates; chaotic under responsive rates) rather than as a blanket failure of the non-atomic framework, and should quantify how large a must be relative to local absolute delays for the 2^p attractor to appear under standard decreasing schedules (e.g., 1/√t).
  3. Theorem 11 and §7: The chaotic CE with O(1/T) swap-regret is confined to a symmetric invariant manifold of measure zero under Lebesgue prior; the paper acknowledges transverse instability and that generic noise ejects trajectories to distinct boundary cycles. The claim that “optimal swap-regret does not preclude macroscopic turbulence” is therefore true on that manifold and for nearby initial conditions that land on different cycles, but the measure-theoretic weight of the chaotic set itself is zero. The manuscript should clarify what is being claimed for generic (absolutely continuous) initial conditions: unpredictability of the eventual support, not persistence of positive-Lyapunov chaos almost everywhere. Without that clarification, the result can be read as stronger than the geometry supports.
minor comments (5)
  1. Figure 1 packs many claims into a single hierarchy diagram; some arrows (e.g., from PCE/SCE to dominated strategies) would benefit from a short caption pointer to the corresponding theorem.
  2. Definition 7 (strict saddle via analytic extension) is clear, but a one-line reminder that the zero-trace law forces mixed critical points to be saddles when the restricted Hessian is nonzero would help readers who skip Appendix A.4.
  3. In §5–6, “strong CCE” and “extreme CCE” are used; a brief cross-reference to Anagnostides et al. (or the paper’s own definition) would avoid ambiguity.
  4. Appendix E.2 (transverse instability under noise) is important for interpreting Theorem 11; a short pointer in the main text of §7 would make the measure-zero caveat more visible.
  5. Typos and notation: “eqilibria” appears in several theorem titles; “A verage” in Theorem 8 proof; consistent use of Φ vs. potential would help.

Circularity Check

1 steps flagged

No definitional loops, fitted predictions, or load-bearing uniqueness imports; constructive counterexamples and dynamical classifications are self-contained, with self-citations only as background for extensions of prior chaos results.

specific steps
  1. self citation load bearing [§8 (Phase Transitions), Theorem 12 and surrounding text; also Related Work §9.3]
    "Recent important contributions have illuminated this meta-stable regime, demonstrating the emergence of Li-Yorke chaos and qualitative inefficiencies in non-atomic congestion games under Multiplicative Weights Update (MWU) and Follow-the-Regularized-Leader (FTRL) with steep regularizers [Bielawski et al., 2021, 2025, Chotibut et al., 2020]. Building upon this work, we provide a generalized and exact quantification of this inefficiency for polynomial latencies..."

    The citations include overlapping authors and supply the existence of Li-Yorke chaos for MWU/FTRL; the present paper then claims a new exponential 2^p degradation. The step is only mildly circular because the paper supplies independent proofs of the stability threshold, the global period-2 attractor, and the exact limit of the time-averaged social-cost ratio; the self-citations function as background rather than as an unverified uniqueness or definitional premise that forces the new quantitative claim.

full rationale

The paper consists of existence/counterexample theorems (Theorems 1–12) and geometric/dynamical classifications of equilibria, potentials, and discrete-time maps. All central claims are established by explicit constructions (e.g., double-cycle and multi-hop networks for strict saddles/repellers; two-agent affine games with negative intercepts for unbounded PoA; grim-trigger uncoupled dynamics for dominated CCE; bipartite mirroring of Peixe–Rodrigues for chaos+swap-regret; period-2 attractors of MWU/PGD maps for 2^p inefficiency) together with direct calculations of Hessians, Lyapunov functions, Chetaev functions, invariant manifolds, and limits of empirical social cost. No free parameters are fitted to data and then re-labeled as predictions. Self-citations (Bielawski–Chotibut–Falniowski–Misiurewicz–Piliouras line on Li-Yorke chaos under MWU/FTRL) appear only as background that the present work extends (higher-degree polynomials, L2/PGD, exact 2^p quantification); the new statements are proved independently in the appendices and do not reduce to the cited results by construction. Classical PoA statements are not redefined; the paper deliberately relaxes non-negativity of intercepts and global step-size fine-tuning and shows the consequences. Hence circularity is negligible.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 2 invented entities

The paper is pure theory: load-bearing content is standard game theory and dynamical systems plus modeling choices about which cost representations and learning rates count as legitimate. Counterexample parameters (w, ε, a, p) are free in the mathematical sense but not fitted to empirical datasets; they are chosen to exhibit unbounded ratios or bifurcations. No new physical entities are postulated.

free parameters (3)
  • Affine intercept sign / coefficient non-negativity relaxation
    Classical PoA theory requires a_e,b_e≥0; unboundedness (Thm 7) depends on allowing negative intercepts while keeping c(x)>0 and nondecreasing for integer x≥1. This is a modeling dial, not a data fit, but the central sensitivity claim turns on it.
  • Effective MWU load a ≈ ε N^p
    Stability threshold a ≶ 2^{p+1}/p and the 2^p attractor limit (Thm 12) depend on this composite step-size parameter; chaos claims assume a is not globally fine-tuned to the O(N^{-p}) sweet spot.
  • Counterexample scales w, ε, δ in 2×2 affine games
    Used to drive PoA or APoA to infinity or to set basins (Fig. 5, Thms 7–8); chosen by construction, not estimated from data.
axioms (6)
  • domain assumption Smoothness framework and classical robust PoA for affine/polynomial congestion games (Roughgarden; Christodoulou–Koutsoupias; Awerbuch et al.)
    Used as the target of critique; tightness constructions are taken as the objects whose dynamics are analyzed (§2, Thms 1–2).
  • standard math Exact potential / Rosenthal potential for congestion games; multilinear extension for mixed strategies
    Underpins Hessian/trace arguments and Lyapunov analysis (Thms 1–4).
  • domain assumption Gradient-like continuous dynamics for which the potential is a strict Lyapunov function outside equilibria; interior-regular Riemannian metrics for measure-zero basins
    Property (ii)–(iii) of Thm 1 and global repeller claims depend on this class (replicator, smooth Riemannian gradient flows).
  • standard math No-regret / swap-regret definitions and standard continuous-time replicator as fluid limit of MWU
    Background for CCE/CE projections and O(1/T) claims (§2, Thms 9–11).
  • standard math Li-Yorke chaos and one-dimensional map tools; Shilnikov-type strange attractors in the cited polymatrix replicator family (Peixe–Rodrigues)
    Imported to lift chaos into a multilinear 6-player game on a symmetric manifold (Thm 11).
  • ad hoc to paper Agents may use fixed responsive discrete step sizes not normalized by global max congestion O(N^p)
    Remark 4 frames oracle fine-tuning as unrealistic; the exponential inefficiency claim is about this regime versus the classical normalized regime that restores convergence.
invented entities (2)
  • MinMax-Centered / MinMax-Regret performance scores no independent evidence
    purpose: Additive individual-rationality baselines used to show (λ,μ)-approximate optimal states can violate safety (Defs 8–9, Thm 5).
    Definitional metrics for the paper’s critique, not new physical objects; independent evidence is not applicable beyond the constructed games.
  • Combinatorial shadow of correlated play no independent evidence
    purpose: Name for information loss when continuous trajectories are projected to empirical CE/CCE distributions.
    Rhetorical/analytic framing rather than a new ontological entity; no external falsifiable handle beyond the theorems.

pith-pipeline@v1.1.0-grok45 · 57996 in / 4227 out tokens · 47214 ms · 2026-07-14T03:25:25.111333+00:00 · methodology

0 comments
read the original abstract

For decades, static solution concepts (Nash, Correlated, and Coarse Correlated Equilibria) and the Price of Anarchy (PoA) have formed the bedrock of algorithmic game theory, with no-regret learning proving fast convergence to such game-theoretic equilibria. We show that reducing multi-agent learning to static equilibrium and black-box regret analysis obscures underlying dynamic disequilibrium and game theoretic bounds. First, interior Nash equilibria lack $C^1$ vector field information, meaning agents cannot distinguish aligned from strictly opposing incentives. Inheriting this geometry, the worst-case pure Nash equilibria dictating robust PoA bounds manifest as topologically unstable strict saddles, and in canonical congestion games, as global repellers supported on almost everywhere strictly dominated strategies. Anchoring efficiency guarantees to these unstable states causes algebraic sensitivity; we prove that accommodating all strictly positive affine costs renders the PoA unbounded. Furthermore, projecting learning trajectories onto the discrete simplex of correlated play systematically accommodates non-rationalizable behavior. Evaluating dynamics via Coarse Correlated Equilibria or proximal refinements fails to preclude strictly dominated strategies. Moreover, optimal $O(1/T)$ swap-regret minimization does not preclude macroscopic turbulence, manifesting as chaotic limit sets even in minimal games. Finally, we examine the non-atomic limit of congestion games. Though considered highly stable with tight sub-linear $\Theta(p/\ln p)$ PoA bounds (where $p$ is the polynomial degree), we prove that under discrete-time learning, the unique equilibrium destabilizes into Li-Yorke chaos and global attractors whose time-averaged inefficiency degrades exponentially as $2^p$. These results necessitate re-evaluating worst-case equilibrium frameworks for dynamically grounded metrics.

Figures

Figures reproduced from arXiv: 2607.11752 by Georgios Piliouras, Ian Gemp, Luke Marris, Siqi Liu.

Figure 1
Figure 1. Figure 1: Paradoxes of Game-Theoretic Solution Concepts. arXiv:2607.11752v1 [cs.GT] 13 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Topological Degeneracy of Worst-Case PoA (I). Replicator dynamics in the linear con￾gestion game producing the tight lower bound of 5/2 for pure PoA [Christodoulou and Koutsoupias, 2005a],[Roughgarden, 2015] (Theorem 5.6, Example 5.7). The worst-case pure NE corresponds to a state of maximum potential—the most dynamically unstable state in the system. The strategies supporting the worst case pure Nash … view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Algebraic Constraints vs. Physical Validity. A comparison between a restricted and an unrestricted affine function in robust Price of Anarchy analysis. While both cost functions map strictly positive integer loads to positive costs, the critical parameter dictating their inclusion in PoA bounds relies entirely on whether the extrapolated, physically inaccessible continuous line crosses the 𝑦-axis above or … view at source ↗
Figure 5
Figure 5. Figure 5: Unbounded Price of Anarchy in Affine Congestion Games. Two parametric families of affine congestion games where the PoA is strictly unbounded. Pure Nash equilibrium states are bolded. Left: For games with 𝑤 > 1, the pure PoA is unbounded when accommodating affine cost functions of arbitrarily large slopes. Right: The mixed PoA is unbounded even for families of affine congestion games with an arbitrarily sm… view at source ↗
Figure 6
Figure 6. Figure 6: The Combinatorial Shadow of Chaos. Replicator dynamics in normal-form games simultaneously exhibit macroscopic chaos, optimal 𝑂(1/𝑇 ) convergence to CE, and optimal state-average convergence to NE. We demonstrate that this assumption does not hold dynamically. Evaluating learning algorithms strictly through the discrete combinatorial shadow of swap regret projects away the continuous￾time dimensions necess… view at source ↗
Figure 7
Figure 7. Figure 7: Bounded Swap Regret amidst Turbulence. The instantaneous swap regret of all agents remains strictly bounded, ensuring the time-average of the replicator dynamics converges provably at an optimal 𝑂(1/𝑇 ) rate to a chaotic correlated equilibrium. The games in question are polymatrix games with 6 agents, each possessing exactly two strategies. The implications of this result are clarified by contrasting it wi… view at source ↗
Figure 8
Figure 8. Figure 8: The emergence of Li-Yorke chaos and the concurrent satisfaction of Gradient Equilibrium (GEQ) [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Exact Measure-Theoretic Basins of Attraction. The simulated phase portrait of the replicator vector field after the volume-preserving transformation. The blue curve represents the mathematically exact invariant stable manifold 𝑥 2 (1−𝑥) = 𝑦 2 (1−𝑦) defining the separatrix. Because the saddle point is positioned at (2/3, 2/3), it severely constricts the basin of the high-potential (worst-case) equilibrium a… view at source ↗
Figure 10
Figure 10. Figure 10: Restoring Multilinearity via Bipartite Mirroring. The base 3-agent polymatrix game (left) contains self-loops (𝑃𝑖𝑖), forcing an agent’s utility to depend quadratically on their own mixed strategy, rendering the game non-concave and violating standard regret frameworks. By cloning the population into sets 𝑋 (blue) and 𝑌 (red), self-loops are unfolded into bidirectional bipartite interactions between identi… view at source ↗
Figure 11
Figure 11. Figure 11: Structural Robustness of Chaos. The distribution of the maximum Lyapunov exponents across all 289 simulated games illustrates that the vast majority of perturbed systems (> 95%) cluster well above the positive chaotic threshold. E.3 Numerical Demonstration of Symmetry Breaking and Distinct Limit Cycles In Theorem 11, we established that continuous-time replicator dynamics can exhibit chaotic limit sets in… view at source ↗
Figure 12
Figure 12. Figure 12: Symmetry Breaking and Emergence of Distinct Limit Cycles. Numerical simulation of the 6-agent mirrored polymatrix game under asymmetric initializations. Ten trajectories are initialized near the interior equilibrium with a microscopic asymmetric perturbation (10−5 ) between the 𝑋 and 𝑌 populations. While the trajectories initially shadow the chaotic limit set, the positive Lyapunov exponents rapidly ampli… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

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

    math.DS 2026-07 conditional novelty 6.0

    Despite Li-Yorke chaos, multiplicative-weights learning in a two-strategy congestion game still has natural invariant measures that fix long-run averages of payoffs, social cost, and regret.

Reference graph

Works this paper leans on

8 extracted references · 1 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    InInternational Conference on Machine Learning

    Follow-the-regularized-leader routes to chaos in routing games. InInternational Conference on Machine Learning. PMLR, 925–935. 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). Jakub...

    doi:10.1145/3717823.3718307 2024

  2. [2]

    Optimistic mirror descent in saddle-point problems: Going the extra(-gradient) mile. InICLR. https://openreview. net/forum?id=Bkg8jjC9KQ Panayotis Mertikopoulos, Christos Papadimitriou, and Georgios Piliouras. 2018. Cycles in Adversarial Regularized Learning. InProceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms(New Orleans, L...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1145/2806883 2018

  3. [3]

    double-cycle

    Solving the system (𝐻−𝜆𝐼)𝑣= 0yields the corresponding exact eigenvector: 𝑣= ©­­­ « 1 1√ 2−1√ 2−1 ª®®® ¬ Because √ 2> 1, every component of this eigenvector is strictly positive. In both canonical topologies, the explicit eigenvectors dictating the unstable manifold 𝑊 𝑢 consist strictly of positive components. Therefore, the tangent space of the repelling ...

    2005

  4. [4]

    + (1−𝜂)𝑎(𝑥 2 2)=𝜂𝑎(𝑥 ∗

  5. [5]

    (𝑥1 +1) + (1−𝜂)𝑎(𝑥 ∗

  6. [6]

    dummy players

    (𝑥2 +1) Substituting the exact block sizes (𝑥1 =2, 𝑥 2 =1, 𝑥 ∗ 1 =1, 𝑥 ∗ 2 =1) yields: 𝜂𝑎(4) + (1−𝜂)𝑎(1)=𝜂𝑎(3) + (1−𝜂)𝑎(2) 3𝜂+1=𝜂+2=⇒2𝜂=1=⇒𝜂=1/2 The indifference constraint mathematically mandates that 𝜂= 1/2. (see also Example 5.7 in [Roughgarden, 2015]). Consequently, the affine cost functions for both cycles possess identical slopes (𝑎/2), ensuring uni...

    2015

  7. [7]

    self-loops,

    The expected utility for each agent is exactly1. We evaluate the CCE condition. The optimal unconditional deviation for agent 1 is to permanently play a strictly dominating strategy, e.g., {𝑎1, 𝑐1}. Against the marginal distribution of 𝜋, agent 2 plays {𝑐2} with probability1 /2, allowing agent 1 to exclusively secure {𝑎1, 𝑐1} for a payoff of1 +𝜖 . 50 With...

    2011

  8. [8]

    The system flow strictly alternates between the states𝜎𝑎 and1 −𝜎𝑎

    Because divergence is strictly impossible at any interior point and at the boundary 𝜎𝑎 → 1/2, the condition 𝑎→ ∞ leaves only one possibility, forcing the limitlim 𝑎→∞ 𝜎𝑎 =0. The system flow strictly alternates between the states𝜎𝑎 and1 −𝜎𝑎. Due to the symmetric identical latency functions, the absolute social cost evaluates to exactly the same value at bo...