Equant Geometry Recasts Nonlinear Oscillations as Clocks

Nonlinear oscillators often have a clear rhythm but an unclear phase. In high-dimensional biological systems, the trajectory can be distorted, asymmetric, or embedded in many state variables, so the usual geometric angle around a cycle is not a physically reliable clock. This paper proposes a data-driven answer: identify an observation point, inspired by Ptolemy's equant, from which the oscillator sweeps area as uniformly as possible and use that view to define phase.

The result is a learned dynamical clock. The authors train neural maps that send a limit cycle to the unit circle, recover an inverse map back to the original system, and compute phase response functions under perturbations. The contribution is partly conceptual and partly operational: it gives a procedure for extracting phase, response, and synchronisation laws directly from trajectory data rather than requiring a hand-designed phase coordinate.

Figure 1 lays out the analogy. Ptolemy's equant made planetary motion appear uniform from an offset point; here, the offset point becomes a learned geometric observer for rhythmic systems.

Core Contribution

The paper's central claim is that a broad class of oscillatory systems can be represented as uniform rotation through an equant-induced coordinate. Formally, the phase $\phi(x)$ is trained so that its derivative along the vector field is constant, $\dot{\phi}=\omega$, while an inverse map $\chi(\phi)$ reconstructs the limit cycle. The equant $x^*$ is then selected by minimizing normalized areal non-uniformity, a dimensionless measure of how evenly the viewed trajectory sweeps area over one period.

That is the main difference from ordinary phase reduction. Standard phase coordinates are mathematically valid but can be hard to interpret geometrically, especially when the limit cycle is irregular. The authors add a physical viewing condition: phase should correspond to the angle seen from an optimal observer. This makes the phase not just a coordinate, but an interpretable clock tied to the system's geometry.

Technical Approach

Second, an invertible neural network extends the learned relation between the original limit cycle and the unit disk to identify the equant. The architecture uses affine invertible mappings, permutation mappings, and dimension alignment so that a two-dimensional disk representation can be matched to systems embedded in higher-dimensional state spaces. This is where the Ptolemaic analogy becomes a computational object: the equant is not assumed from the plot, but optimized by minimizing areal non-uniformity.

Results and Analysis

The strongest empirical evidence is the breadth of systems tested rather than a single benchmark score. The authors report validation across 10 representative oscillatory systems, including Escherichia coli quorum sensing, FitzHugh-Nagumo neurons, Wilson-Cowan dynamics, Lotka-Volterra dynamics, mammalian circadian clocks, thalamic neurons, cell-cycle Cdk models, Morris-Lecar dynamics, Selkov dynamics, and laser dynamics. The evaluation uses three criteria: frequency accuracy, uniformity of the learned equant, and interpretability of the induced viewing map.

Figure 5 is the most compact evidence for the broad claim. It shows frequency accuracy and equant quality across the tested systems, and it uses FitzHugh-Nagumo dynamics to connect equant geometry to critical transitions. As the control parameter approaches a Hopf transition, the optimal equant non-uniformity $K^*$ decreases approximately linearly; extrapolating to $K^*=0$ predicts a critical value of 1.575, close to the true Hopf point 1.574.

The perturbation experiments are more specialized but useful. For weak cyclic perturbations of FitzHugh-Nagumo parameters, the learned phase dynamics predicts a geometric phase that closely tracks full numerical simulations across perturbation periods and strengths. For synchronisation, the paper evaluates five realistic networks with 20 to 1059 nodes and compares reduced phase predictions against original trajectories. The reduced model matches electrical and memristor-based coupling better than chemical synaptic coupling, because the latter has a non-vanishing term even at synchrony and distorts the oscillation shape.

Figure 6 illustrates this boundary. The method gives accurate reduced synchronisation behaviour for coupling forms close to the phase-reduction assumptions, but it is not a universal replacement for full dynamics when perturbations change the geometry of the cycle.

Limitations

The evidence supports the method as a general phase-extraction and analysis framework for simulated or modeled oscillators with identifiable limit cycles. It is less conclusive as a claim about arbitrary real-world data. The approach assumes dynamics remain near the limit cycle under perturbation, requires enough trajectory information to train the autoencoder and invertible map, and depends on the quality of the learned vector-field or model-based dynamics. The supplementary comparisons with neural ODE and adjoint methods are useful, but the paper remains mostly a computational and theoretical demonstration rather than a prospective test on noisy experimental measurement pipelines.