Isostable Framework for Nonlinear Dynamics
- The isostable coordinate framework is a nonlinear dynamics description that uses slowest decaying Koopman eigenfunctions or Floquet modes to organize state-space evolution.
- It defines invariant slow manifolds and underpins reduced-order models, enabling data-driven inference and effective handling of forcing in fixed-point and oscillatory systems.
- Advanced numerical strategies, including predictor–corrector methods and asymptotic expansions, address backward integration instability and stiffness in these high-order reductions.
Searching arXiv for the specified papers to ground the article in current metadata and citations. arxiv_search.query({"12search_query12 OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12"," OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12,"12sort_by12 arxiv_search.query({"12search_query12 and Computation of Slow Manifolds Using the Isostable Coordinate System\" OR 12ti:\12 Inference of High-Accuracy Isostable-Based Dynamical Models in Response to External Inputs\" OR 12ti:\12 for stochastic oscillators\" OR 12ti:\12 into oscillator network dynamics using a phase-isostable framework\"","12max_results12 OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12,"12sort_by12 The isostable coordinate framework is a coordinate description of nonlinear dynamics in which state-space evolution is organized by asymptotic decay rates associated with Koopman eigenfunctions or, for periodic orbits, Floquet modes. For systems with a stable fixed point, the framework focuses on the slowest decaying principal Koopman eigenmodes and defines scalar coordinates whose level sets are isostables; for stochastic oscillators and deterministic limit cycles, closely related constructions produce amplitude-like coordinates complementary to asymptotic phase. In the formulations developed for fixed-point attractors, external-input reductions, stochastic oscillators, and oscillator networks, isostable coordinates provide a common language for invariant geometry, reduced-order modeling, and bifurcation analysis (&&&12search_query12&&&, &&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&, &&&12max_results12&&&, &&&12sort_by12&&&).
12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12. Koopman-theoretic definition and local structure
For a smooth autonomous ODE
PRESERVED_PLACEHOLDER_12search_query12^
with hyperbolic fixed point PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, the Koopman operator acts on observables PRESERVED_PLACEHOLDER_12max_results12^ by
PRESERVED_PLACEHOLDER_12sort_by12^
where PRESERVED_PLACEHOLDER_12relevance12^ is the flow. Although the state-space dynamics are nonlinear, PRESERVED_PLACEHOLDER_12sort_order12^ is linear but generally infinite-dimensional. A Koopman eigenfunction PRESERVED_PLACEHOLDER_12descending12^ and eigenvalue PRESERVED_PLACEHOLDER_12search_query12^ satisfy
PRESERVED_PLACEHOLDER_12ti:\12^
When the Jacobian PRESERVED_PLACEHOLDER_12 OR ti:\12^ has eigenvalues ordered by
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12^
the associated PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ are the principal eigenfunctions, and the real part or modulus of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12^ yields an isostable coordinate measuring distance along the PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12-th mode (&&&12search_query12&&&).
A constructive limit definition is obtained by fixing left and right eigenvectors PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12^ of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12, normalized so that PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12, and setting
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12^
Its level sets PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12^ are the isostables associated with decay rate PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12 OR ti:\12. Differentiation along trajectories gives
PRESERVED_PLACEHOLDER_12max_results12search_query12^
so each principal isostable decays or grows exponentially at rate PRESERVED_PLACEHOLDER_12max_results12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ (&&&12search_query12&&&).
A closely related normalization appears in isostable reduction for forced systems. There, Koopman eigenfunctions PRESERVED_PLACEHOLDER_12max_results12max_results12^ satisfy the spectral PDE
PRESERVED_PLACEHOLDER_12max_results12sort_by12^
with normalization
PRESERVED_PLACEHOLDER_12max_results12relevance12^
and the real-valued isostable coordinate PRESERVED_PLACEHOLDER_12max_results12sort_order12^ is defined by filtering out all but the slowest decaying Koopman mode through a limit of the same form (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&). Near the fixed point, these definitions recover the linear eigendirections because PRESERVED_PLACEHOLDER_12max_results12descending12^ locally (&&&12search_query12&&&).
12max_results12. Slow manifolds as zero sets of fast isostables
When there is a pronounced spectral gap at index PRESERVED_PLACEHOLDER_12max_results12search_query12,
PRESERVED_PLACEHOLDER_12max_results12ti:\12^
the framework defines a PRESERVED_PLACEHOLDER_12max_results12 OR ti:\12-dimensional slow manifold by setting all faster isostable amplitudes to zero:
PRESERVED_PLACEHOLDER_12sort_by12search_query12^
This identifies the slow manifold as the set on which the fastest decaying principal Koopman coordinates vanish (&&&12search_query12&&&).
The definition immediately implies invariance. On PRESERVED_PLACEHOLDER_12sort_by12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, the coordinates PRESERVED_PLACEHOLDER_12sort_by12max_results12^ are identically zero, so their derivatives vanish under PRESERVED_PLACEHOLDER_12sort_by12sort_by12, and the flow remains on PRESERVED_PLACEHOLDER_12sort_by12relevance12. Near PRESERVED_PLACEHOLDER_12sort_by12sort_order12, the manifold reduces to PRESERVED_PLACEHOLDER_12sort_by12descending12, the span of the slow eigenvectors. Away from the linear regime, PRESERVED_PLACEHOLDER_12sort_by12search_query12^ is a nonlinear manifold of codimension PRESERVED_PLACEHOLDER_12sort_by12ti:\12, locally transverse to the fast eigendirections (&&&12search_query12&&&).
This formulation is not merely a local tangent-space statement. The motivating point of the construction is that the condition PRESERVED_PLACEHOLDER_12sort_by12 OR ti:\12^ remains meaningful beyond the linear neighborhood of the attractor, so the same coordinate language used for spectral decomposition also gives a geometric definition of the slow manifold in the nonlinear regime (&&&12search_query12&&&). This suggests a direct bridge between Koopman spectral objects and model order reduction: the reduced dynamics are obtained by retaining the slowest coordinates and eliminating fast amplitudes by an invariant constraint.
12sort_by12. Backward-time computation and numerical strategies
Direct backward integration of the original ODE from a point on PRESERVED_PLACEHOLDER_12relevance12search_query12^ is numerically unstable once the trajectory leaves a small neighborhood of PRESERVED_PLACEHOLDER_12relevance12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, because any small error in a fast isostable PRESERVED_PLACEHOLDER_12relevance12max_results12^ blows up like PRESERVED_PLACEHOLDER_12relevance12sort_by12. To avoid this, the framework rewrites the backward dynamics in isostable coordinates. With backward-time variable PRESERVED_PLACEHOLDER_12relevance12relevance12, one has
PRESERVED_PLACEHOLDER_12relevance12sort_order12^
where PRESERVED_PLACEHOLDER_12relevance12descending12. Stacking PRESERVED_PLACEHOLDER_12relevance12search_query12^ into a matrix and enforcing PRESERVED_PLACEHOLDER_12relevance12ti:\12^ on PRESERVED_PLACEHOLDER_12relevance12 OR ti:\12^ yields
PRESERVED_PLACEHOLDER_12sort_order12search_query12^
If PRESERVED_PLACEHOLDER_12sort_order12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ and PRESERVED_PLACEHOLDER_12sort_order12max_results12^ are known along PRESERVED_PLACEHOLDER_12sort_order12sort_by12, this evolution marches backward without exciting fast modes (&&&12search_query12&&&).
For the slow coordinates PRESERVED_PLACEHOLDER_12sort_order12relevance12, the isostable gradients satisfy the adjoint variational equation
PRESERVED_PLACEHOLDER_12sort_order12sort_order12^
with initial condition PRESERVED_PLACEHOLDER_12sort_order12descending12. Because these are slow modes, backward integration remains accurate over substantial intervals. The fast gradients PRESERVED_PLACEHOLDER_12sort_order12search_query12, however, are not computed directly. Instead, the computation uses their duals PRESERVED_PLACEHOLDER_12sort_order12ti:\12, which solve
PRESERVED_PLACEHOLDER_12sort_order12 OR ti:\12^
with bi-orthonormality PRESERVED_PLACEHOLDER_12descending12search_query12. Since the backward formula only needs the subspace spanned by PRESERVED_PLACEHOLDER_12descending12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, it is enough to approximate the orthogonal complement of PRESERVED_PLACEHOLDER_12descending12max_results12^ (&&&12search_query12&&&).
Two strategies are developed. The asymptotic-expansion approach expands PRESERVED_PLACEHOLDER_12descending12sort_by12^ and PRESERVED_PLACEHOLDER_12descending12relevance12^ in Taylor series in the isostables near PRESERVED_PLACEHOLDER_12descending12sort_order12, for example
PRESERVED_PLACEHOLDER_12descending12descending12^
and then matches with PRESERVED_PLACEHOLDER_12descending12search_query12^ to solve for PRESERVED_PLACEHOLDER_12descending12ti:\12^ to arbitrary order. The method works well in low dimension or mild nonlinearity but becomes unwieldy beyond PRESERVED_PLACEHOLDER_12descending12 OR ti:\12–PRESERVED_PLACEHOLDER_12search_query12search_query12-th order. The predictor–corrector approach instead uses the approximation PRESERVED_PLACEHOLDER_12search_query12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, integrates backward using the slow gradients, then corrects the resulting point by solving for a displacement PRESERVED_PLACEHOLDER_12search_query12max_results12^ in the fast eigendirections so that the exact isostable-gradient condition is restored (&&&12search_query12&&&).
The reported numerical issues are threefold. Timescale separation makes direct backward integration exponentially unstable in fast modes; the reformulated backward equation avoids this by keeping PRESERVED_PLACEHOLDER_12search_query12sort_by12. Stiffness arises in the adjoint equation when PRESERVED_PLACEHOLDER_12search_query12relevance12^ is very negative, motivating a stiff integrator or small steps. Non-uniqueness requires choosing PRESERVED_PLACEHOLDER_12search_query12sort_order12^ so that there are no resonances PRESERVED_PLACEHOLDER_12search_query12descending12^ for slow indices; this guarantees convergence of the asymptotic expansions (&&&12search_query12&&&).
12relevance12. Reduced-order models with forcing and data-driven inference
For systems with a stable fixed point driven by an input PRESERVED_PLACEHOLDER_12search_query12search_query12^ entering through PRESERVED_PLACEHOLDER_12search_query12ti:\12, an PRESERVED_PLACEHOLDER_12search_query12 OR ti:\12-mode isostable reduction keeps only the slowest coordinates PRESERVED_PLACEHOLDER_12ti:\12search_query12^ and represents the reduced dynamics and output as
PRESERVED_PLACEHOLDER_12ti:\12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^
Here PRESERVED_PLACEHOLDER_12ti:\12max_results12^ and PRESERVED_PLACEHOLDER_12ti:\12sort_by12^ are expanded in multivariate Taylor series in PRESERVED_PLACEHOLDER_12ti:\12relevance12, and truncating at total degree PRESERVED_PLACEHOLDER_12ti:\12sort_order12^ yields a PRESERVED_PLACEHOLDER_12ti:\12descending12-th-order accurate reduced model (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&).
When the underlying equations are unknown, the coefficients can be inferred from steady-state responses to sinusoidal probing. With rank-one input
PRESERVED_PLACEHOLDER_12ti:\12search_query12^
the steady-state output admits an PRESERVED_PLACEHOLDER_12ti:\12ti:\12-expansion and Fourier series
PRESERVED_PLACEHOLDER_12ti:\12 OR ti:\12^
The key structural fact is that the PRESERVED_PLACEHOLDER_12 OR ti:\12search_query12-th harmonic first appears at PRESERVED_PLACEHOLDER_12 OR ti:\12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, and its amplitude is a linear combination of the PRESERVED_PLACEHOLDER_12 OR ti:\12max_results12-th-order Taylor coefficients in PRESERVED_PLACEHOLDER_12 OR ti:\12sort_by12^ and PRESERVED_PLACEHOLDER_12 OR ti:\12relevance12. At first order,
PRESERVED_PLACEHOLDER_12 OR ti:\12sort_order12^
with PRESERVED_PLACEHOLDER_12 OR ti:\12descending12^ collecting measured first harmonics, PRESERVED_PLACEHOLDER_12 OR ti:\12search_query12^ known analytically in terms of PRESERVED_PLACEHOLDER_12 OR ti:\12ti:\12, and PRESERVED_PLACEHOLDER_12 OR ti:\12 OR ti:\12^ containing the unknown coefficient products. Pseudoinverse solution gives
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12search_query12^
Second and higher orders have the same structure, with remainder terms depending only on lower-order quantities. The eigenvalues PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ can also be refined by Newton iteration on the first-order relation (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&).
The algorithmic summary consists of a preliminary data-driven estimate of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12max_results12^ from delay-embedded outputs and POD, symbolic assembly of response matrices PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12sort_by12, harmonic extraction at probe frequencies PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12relevance12, pseudoinverse recovery of the coefficient vectors PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12sort_order12, and optional Newton refinement before assembling the final reduced model (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&).
The examples emphasize accuracy under finite forcing amplitudes. In a two-dimensional test model with PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12descending12^ and PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12search_query12, one isostable suffices, the third-order reduction matches the full dynamics perfectly when PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12ti:\12^ and PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12 OR ti:\12^ are known, and data-driven inference from noisy output recovers PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12^ together with coefficients up to cubic order. For a population of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ synaptically coupled Morris–Lecar-type neurons, two complex-conjugate isostables suffice, and the second-order inferred model outperforms the linear one by an order of magnitude in mean absolute error under composite transient inputs. For the one-dimensional Burgers’ equation projected onto five POD modes, a second-order three-isostable model reduces PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12-error by two orders of magnitude compared to the first-order model under composite multi-frequency boundary inputs (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&).
A related input-driven reduction appears in the Goodwin oscillator example used for slow-manifold computation. After adding an external input PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12^ to the PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12-equation and restricting to the manifold PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12^ with PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12, the reduced model becomes
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12^
where PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12^ is tabulated along the computed manifold. In that example, the reduced one-complex-dimensional ODE reproduces steady-state periodic and period-doubling behavior far beyond the linear approximation; the full and isostable-reduced models both show a period-doubling bifurcation at input amplitude PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12 OR ti:\12–PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12search_query12, whereas a naive linearization fails to capture the bifurcation, and the maximum amplitude of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ versus PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12max_results12^ matches to within a few percent (&&&12search_query12&&&).
12sort_order12. Stochastic and phase–amplitude extensions
For stochastic oscillators governed by the Itô SDE
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12sort_by12^
with diffusion tensor PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12relevance12, the relevant operator is the backward Kolmogorov, or stochastic Koopman, operator
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12sort_order12^
Under standard ellipticity and suitable boundary conditions, PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12descending12^ and PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12search_query12^ admit discrete spectra and biorthogonal eigenbases. A system is termed robustly oscillatory when the nonzero eigenvalue of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12ti:\12^ with maximal real part is a complex conjugate pair PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12max_results12 OR ti:\12, with PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12search_query12, PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, and all other eigenvalues satisfy PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12max_results12. The asymptotic phase eigenfunction is then PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12sort_by12, and the asymptotic phase is
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12relevance12^
If there is also a unique real eigenvalue PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12sort_order12^ with maximal real part among real eigenvalues, the corresponding real eigenfunction PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12descending12^ is interpreted as the stochastic isostable coordinate (&&&12max_results12&&&).
Normalization is fixed by biorthogonality, and PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12search_query12^ is typically shifted so that its zero-level set defines an effective limit cycle
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12ti:\12^
In the phase–amplitude variables
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_by12 OR ti:\12^
Itô’s formula yields
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12search_query12^
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^
Thus, in the mean sense,
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12max_results12^
The noisy linear focus provides a closed-form example: for the planar Ornstein–Uhlenbeck process with
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12sort_by12^
one obtains
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12relevance12^
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12sort_order12^
so the effective limit cycle is the circle PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12descending12^ (&&&12max_results12&&&).
For deterministic limit cycles, the analogous construction uses Floquet theory rather than Koopman eigenfunctions at a fixed point. If the uncoupled ODE admits a stable PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12search_query12-periodic orbit PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12ti:\12^ with dominant nontrivial Floquet exponent PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12relevance12 OR ti:\12, the asymptotic phase PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12search_query12^ is defined by PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12, and the slowest-decaying transverse coordinate is an isostable PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12max_results12^ satisfying
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12sort_by12^
Near the cycle, the phase–isostable normal form under weak perturbation PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12relevance12^ is
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12sort_order12^
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12descending12^
where PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12search_query12^ is the iPRC and PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12ti:\12^ is the iIRC (&&&12sort_by12&&&). The phase and isostable coordinates therefore separate tangential timing from the slowest transverse amplitude decay.
12descending12. Network formulations, comparative scope, and limitations
For PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12sort_order12 OR ti:\12^ identical coupled oscillators
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12search_query12^
retaining the phase PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ and a single slow isostable PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12max_results12^ at each node gives, after Taylor expansion and first-order averaging in PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12sort_by12,
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12relevance12^
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12sort_order12^
The six PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12descending12-periodic coupling functions PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12search_query12^ are defined by averaging PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12ti:\12^ over one period and depend on PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12descending12 OR ti:\12, the coupling Jacobians, and the Floquet eigenfunction PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12search_query12^ (&&&12sort_by12&&&).
The reduced network equations support explicit existence and stability conditions for phase-locked states. For a PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ phase-locked solution PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12max_results12^ with constant PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12sort_by12, the locked amplitudes satisfy PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12relevance12^ and the collective frequency shift satisfies
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12sort_order12^
Linearization produces a PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12descending12^ Jacobian in block form PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12search_query12, and stability requires all eigenvalues except the rotational zero mode to have negative real part. In the globally coupled case, synchrony reduces to a PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12ti:\12^ matrix PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12search_query12 OR ti:\12, so stability requires
PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12search_query12^
Taking PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12^ recovers the classical phase-only criterion PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12max_results12^ (&&&12sort_by12&&&).
The mean-field complex Ginzburg–Landau equation provides an analytic benchmark. There, the phase-isostable reduction reproduces the synchrony boundary exactly, yields compact analytic expressions for the splay and antisynchrony boundaries, and tracks the true bifurcation loci qualitatively over a wide parameter range, whereas second- and third-order phase reductions agree only locally near PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12sort_by12. It also predicts bistability regions between synchrony and splay in agreement with the full system. In globally coupled Morris–Lecar networks, the PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12relevance12-dimensional phase-isostable equations predict loss and restabilization of synchrony, a narrow window of stable antisynchrony, off-invariant-manifold phase-locked states, quasiperiodic solutions born in Hopf bifurcations, and cluster states for PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12sort_order12, while the first-order phase model predicts only trivial stable antisynchrony and unstable synchrony (&&&12sort_by12&&&).
Across the fixed-point, forced, stochastic, and network settings, several limitations are explicit. Higher-order Taylor inference suffers from combinatorial growth in the number of coefficients and from noise sensitivity because fitting divides by PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12descending12; careful choice of PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12search_query12^ can mitigate this. The data-driven fixed-point framework requires a stable fixed point and forcing experiments that remain inside its basin of attraction. In slow-manifold computation, stiffness and fast-mode instability constrain backward marching, and asymptotic expansions require a nonresonance condition on slow indices. In network reduction, higher-order pure phase reductions become cumbersome for large PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12ti:\12, whereas the phase-isostable model remains compact by retaining one extra coordinate per node and six coupling functions PRESERVED_PLACEHOLDER_12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12ti:\12 OR ti:\12^ (&&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&, &&&12search_query12&&&, &&&12sort_by12&&&).
Taken together, these formulations define the isostable coordinate framework as a family of reductions in which the dominant decay structure of a nonlinear system is represented explicitly. At fixed points, the coordinates identify invariant slow manifolds by the condition that fast isostables vanish; under forcing, they support high-order reduced models and purely data-driven inference; in stochastic oscillators, they supply an amplitude variable dual to asymptotic phase; and in oscillator networks, they extend phase reduction by tracking slow transverse deviations from the attracting cycle (&&&12search_query12&&&, &&&12id:(Wilson, 18 Jul 2025) OR id:(Wilson, 2021) OR id:(Pérez-Cervera et al., 2021) OR id:(Nicks et al., 2023)12&&&, &&&12max_results12&&&, &&&12sort_by12&&&).