[Paper] Broken-symmetry shape discrimination on a driven Duffing ring

Source: arXiv - 2605.07475v1

Overview

Kaspar Anton Schindler investigates how a simple ring of coupled oscillators—think of a circular array of tiny mechanical or electronic resonators—can perform two fundamental computing tasks: bundling (storing multiple independent pieces of information) and binding (combining those pieces into a higher‑level representation). By driving the ring either linearly or in a nonlinear Duffing regime, the study shows how the system can extract shape information from time‑varying signals, even in noisy environments.

Key Contributions

• Unified framework for bundling & binding on a minimal closed substrate (a cycle graph of N nodes).
• Linear regime analysis: demonstrates that the ring’s eigenmodes act like a windowed FFT, achieving competitive feature extraction at high SNR and modest gains for transient, low‑SNR signals.
• Nonlinear Duffing regime: reveals a cubic mode‑mixing “shape‑discrimination” operation governed by a sparse integer‑wavenumber selection rule derived from the ring’s U(1) symmetry.
• Single‑number observable ( \phi_0 ): defines a compact metric that captures the bound representation’s response to input shape, with exact (\pi)-periodicity and a broken time‑reversal symmetry that preserves information.
• Robustness to noise: numerical experiments show (\phi_0) remains distinguishable from its symmetric baseline down to 0 dB input SNR.
• Open‑ended roadmap: outlines how the approach could be extended to richer physical substrates, more complex drives, and real‑world biological signals.

Methodology

1. Physical substrate – a ring of N identical oscillators coupled to nearest neighbours, mathematically represented as a cycle graph.
2. Master equation – a single differential equation governing each node’s displacement, with two regimes:
   • Linear: only the restoring force term is active.
   • Duffing: adds a cubic nonlinearity, turning the system into a driven Duffing oscillator.
3. Input encoding – temporal signals are injected uniformly across the ring (bundling). The shape of the signal is parameterised by a single angle (\theta) that controls its waveform symmetry.
4. Feature extraction – the system’s response is projected onto its eigenmodes (Fourier basis). In the Duffing regime, cubic interactions mix modes according to a selection rule that only allows certain integer wavenumber combinations, creating shape‑dependent harmonics.
5. Observable definition – the authors derive (\phi_0) by integrating the phase‑locked component of the response over one drive period, yielding a scalar that varies with input shape but is invariant to many irrelevant transformations.
6. Evaluation – synthetic test signals (sinusoids, chirps, transients) are run through the model with additive band‑limited Gaussian noise. Performance is compared against a standard windowed FFT baseline.

Results & Findings

• Linear regime matches the windowed FFT baseline at high SNR (≥ 20 dB) and modestly outperforms it for short‑duration transients at lower SNR (≈ 10 dB).
• Duffing regime produces new harmonic content that the linear system cannot generate, enabling discrimination of signal shapes that are indistinguishable in the linear spectrum.
• (\phi_0) behavior: exhibits perfect (\pi)-periodicity in the shape parameter (\theta); its trajectory across the quotient domain (i.e., after factoring out the symmetry) uniquely encodes the combined effect of binding and dissipation.
• Noise resilience: even with 0 dB input SNR, the mean (\phi_0) over many random seeds stays well above the symmetric‑attractor value, confirming that the observable retains shape information under severe noise.
• Symmetry analysis: time‑reversal symmetry is broken by the system’s intrinsic damping, which is essential for (\phi_0) to be a non‑degenerate descriptor.

Practical Implications

• Edge‑AI hardware: The ring architecture is a candidate for ultra‑low‑power analog processors that can perform feature extraction directly in the physical domain, reducing the need for digital FFTs.
• Neuromorphic sensing: The bundling‑binding paradigm mirrors how biological neural circuits store and combine sensory inputs; implementing it in hardware could lead to more brain‑inspired sensors.
• Robust signal classification: Because (\phi_0) survives heavy noise, devices built on this principle could be used in harsh environments (e.g., industrial monitoring, underwater acoustics) where traditional digital preprocessing struggles.
• Compact representation: A single scalar summarizing shape information simplifies downstream machine‑learning pipelines, enabling lightweight classifiers on microcontrollers.
• Design guidelines: The selection rule for mode mixing provides a clear recipe for engineering the coupling and nonlinearity to target specific harmonic signatures, facilitating custom analog filter banks.

Limitations & Future Work

• Synthetic only: Experiments are limited to numerically generated signals; real‑world data (speech, ECG, vibration) may introduce complexities not captured here.
• Single‑parameter shape model: The study varies only one shape angle; richer shape spaces (e.g., multi‑modal or non‑periodic patterns) remain unexplored.
• Scalability: The analysis focuses on modest ring sizes; performance and stability for large‑scale networks need investigation.
• Hardware realization: Translating the Duffing dynamics into practical MEMS, photonic, or electronic circuits will require careful handling of nonlinearity, damping, and fabrication tolerances.
• Extension to other substrates: The authors suggest exploring lattices, hierarchical graphs, or biologically inspired media, which could broaden the applicability of the bundling‑binding framework.

Authors

• Kaspar Anton Schindler

Paper Information

• arXiv ID: 2605.07475v1
• Categories: cs.NE, cs.ET, eess.SP
• Published: May 8, 2026
• PDF: Download PDF