Homoclinic Intersections and the Macroscopic Observability of Arnold Tongues in the Forced-Dissipative Duffing System

Abstract: Dissipative chaos often exhibits abrupt transitions to periodic windows or vanishing states, but these transitions occur far below macroscopic theoretical boundaries such as the Melnikov threshold. In this study, we re-evaluate the transversal intersections (microscopic) of invariant manifolds from a deterministic and entropic perspective for the ``Arnold tongues'' shown by synchronization to forced inputs in the parameter space. We applied an algorithm that digitally determines manifold intersections as binary values ($1.0$ or $0.0$) without numerical interpolation. As a result of scanning the $Ω- F$ parameter plane at a resolution of $5000 \times 5000$ ($25$ million points) with the damping coefficient fixed at $k = 0.2$, it became possible to globally capture the relationship between the macroscopic phase-locked regions shown by the Arnold tongues and the microscopic manifold intersections (chaotic regions). Furthermore, from a one-dimensional cross-section whose computational accuracy was verified, we confirmed a dynamical case where the region with homoclinic intersections (topological entropy $h_T > 0$) is a necessary condition for the region where chaos manifests (Kolmogorov-Sinai entropy $h_{KS} > 0$), and simultaneously confirmed the existence of a region where the intersection of the primary saddle solution does not serve as a necessary condition.

• The paper introduces a GPU-based, deterministic binarization algorithm to classify transversal homoclinic intersections in a 5000×5000 scan of the Ω−F parameter plane.
• The resulting geometric map reveals two regimes: one where homoclinic intersections exist but the attractor is periodic, and another where chaos persists without primary-saddle intersections.
• The findings refine the classical Melnikov boundary, suggesting that multiple coexisting geometric structures can govern macroscopic uncertainty, challenging the single-saddle homoclinic analysis.

Overview

This paper by Hikihara examines the relationship between microscopic geometric structures in phase space—specifically transversal homoclinic intersections of invariant manifolds—and macroscopic observability of phase locking (Arnold tongues) in the forced-dissipative double-well Duffing oscillator. The central methodological contribution is a deterministic, non-interpolated binarization algorithm, executed on GPU hardware, that classifies each point of a $5000 \times 5000$ ($25$ million-point) scan of the $\Omega - F$ parameter plane at fixed damping $k = 0.2$ according to whether transversal intersections of the stable and unstable manifolds of the primary Poincaré saddle exist. The resulting overlay of this binary geometric flag on the steady-state variance (a proxy for Arnold tongue structure) reveals two regimes that jointly refine the classical picture: intervals where homoclinic intersections persist yet the observed attractor is periodic, and intervals where chaos persists despite the absence of intersections of the primary saddle.

Background and motivation

The Melnikov criterion [Melnikov1963] supplies an analytic threshold for tangency between stable and unstable manifolds and is conventionally treated as a necessary condition for chaos. The paper's premise is that dissipative chaotic attractors are frequently drawn into periodic windows well below the Melnikov boundary, so that Lyapunov-based or coarse-grained statistical indicators cannot explain the geometric mechanism of these early transitions. Prior work—Ueda's identification of Duffing chaos [Ueda1991], Wiggins' global transport theory [Wiggins1992], Gekle et al.'s extraction of multidimensional topology from periodic orbits [Gekle2006], and Gallas' mapping of periodic-window organization [Gallas1993]—established the geometric program, but computational limitations precluded high-resolution global manifold tracking. The present work argues that modern GPU-parallelized parameter-space mapping [Hegedus2020, Rybin2026] makes it feasible to close this gap, connecting Arnold tongue structure—which has recently been invoked for entrainment analysis in biological systems [Sanchez2022]—directly to non-interpolated manifold intersection data.

System and binarization methodology

The system studied is

$\ddot{x} + k\dot{x} - x + x^3 = F\cos\Omega t,$

with $k = 0.2$ fixed, a value reported previously to sustain complex homoclinic intersections. The workflow consists of five stages:

1. Tensorized mesh: the entire $\Omega - F$ plane is batched as a tensor on GPU VRAM to avoid host–device transfer bottlenecks.
2. Batch Newton–Raphson continuation tracks the primary saddle fixed point $\mathbf{z}^*$ of the Poincaré map $\mathcal{P}$ across all parameter points, seeding from previous converged solutions.
3. Eigenvalue analysis of a numerically constructed local Jacobian (central differences with $h = 10^{-5}$) yields the initial slopes of the unstable ($25$0) and stable ($25$1) branches.
4. Manifold evolution: 2 million initial conditions per branch are integrated forward, with adaptive density concentrated near the saddle via a power-law weighting of exponent $25$2, mitigating artifacts from intense topological stretching without any linear interpolation.
5. Deterministic sign-change detection: a signed distance function between the evolved unstable-manifold endpoints and the linearized stable manifold is evaluated; any sign reversal within the search window sets flag $25$3, otherwise (including saddle disappearance) flag $25$4.

A convergence study on the cross-section $25$5 shows that the binarized map is pixel-level identical under time resolutions of 128, 256, and 512 RK4 steps per forcing period, supporting the claim that the results are free of truncation-error artifacts. This claim is plausible given the binary nature of the output but rests on a single cross-section; convergence over the full plane is not separately demonstrated.

Global structure of the parameter plane

The two-dimensional overlay exhibits three notable features. First, the boundaries of the homoclinic-existence region converge toward a resonance cusp near $25$6, with interlocking filaments tracking fractally along the edges of strongly phase-locked (high-variance) windows. Second, above the Melnikov lower limit, the manifold-intersection regions coexist with—but largely avoid—the interiors of the Arnold tongues, indicating that microscopic geometry constrains phase-locking boundaries over a vast region. Third, below the Melnikov line, islands of homoclinic existence appear "over" synchronization regions, demonstrating directly that sub-threshold intersections do not necessarily disrupt macroscopic phase locking. This constitutes a quantitative refinement of the classical interpretation of the Melnikov threshold as the operative chaos boundary.

Necessary-condition structure along a one-dimensional section

The cross-section at $25$7 yields the paper's most substantive dynamical findings, which fall into two contrasting categories:

Interval	Homoclinic flag	Steady-state variance	Interpretation
$25$8	$25$9	$\Omega - F$0	Non-attracting chaotic saddle
$\Omega - F$1	$\Omega - F$2	$\Omega - F$3	Chaos without primary-saddle intersections

Intersections present, chaos absent: In $\Omega - F$4, transversal intersections guarantee $\Omega - F$5 (an embedded horseshoe), yet the physical measure collapses onto a period-one locked state, giving $\Omega - F$6. This quantitatively confirms that homoclinicity is necessary but not sufficient for observable chaos: the gap between topological entropy (total possibility) and Kolmogorov–Sinai entropy (physical observability) is realized concretely as a non-attracting chaotic saddle whose basin measure vanishes in the steady state.

Chaos present, primary-saddle intersections absent: In $\Omega - F$7, the tracked primary saddle has strictly no homoclinic intersections, yet the attractor sustains finite non-periodic fluctuations. The authors attribute this to higher-period saddles or topological structures born from early period-doubling bifurcations, implying that multiple coexisting geometric constituents—not solely the primary saddle—can govern macroscopic uncertainty. This is the paper's strongest and potentially contentious claim: it identifies a "new root of chaos" not captured by single-saddle homoclinic analysis, though the responsible secondary structure is inferred rather than explicitly computed.

Phase-space verification (Fig. 3) supports both cases: at $\Omega - F$8 the manifolds form an unambiguous tangle, while at $\Omega - F$9 the unstable leading edge visibly bypasses the outer loop of the stable manifold, with the saddle displaced to $k = 0.2$0.

Limitations and open questions

Several caveats bear on the strength of the conclusions. The analysis is restricted to a single damping value ($k = 0.2$1) and to intersections involving the primary saddle only; the conjectured higher-order saddles responsible for chaos in $k = 0.2$2 are not themselves tracked, so the alternative geometric mechanism remains hypothetical within the data presented. The binarization uses a linear approximation of the stable manifold near the saddle combined with endpoint distances after long integration—an approximation whose sensitivity to search-window choice is not characterized. Convergence was verified only along one cross-section, leaving full-plane robustness assumed rather than shown. Finally, the identification of the variance-vanishing interval as a chaotic saddle is consistent with theory but is not independently confirmed by, e.g., transient-lifetime scaling or saddle-repeller computation.

Conclusion

By combining GPU-tensorized continuation, non-interpolated manifold construction, and deterministic sign-change classification, the paper produces a 25-million-point geometric map that overlays homoclinic existence directly onto Arnold tongue structure in the forced Duffing system. Its principal results are twofold: a demonstrated regime where $k = 0.2$3 coexists with $k = 0.2$4 (necessary-but-not-sufficient intersections sustaining only a chaotic saddle), and a regime where chaos persists despite absent primary-saddle intersections, implicating coexisting higher-order geometric structures. The work reopens, with modern computational means, the deterministic geometric program for explaining why macroscopic transitions depart from Melnikov-type thresholds, while leaving open the explicit identification of the non-primary saddles that carry the chaotic dynamics in the anomalous region.