Coordinate-Independent GSPT

• ci-GSPT is a coordinate-independent framework that characterizes singular perturbations by encoding slow–fast dynamics through geometric invariants such as critical manifolds and fast fibre bundles.
• It employs the parametrization method to compute slow manifolds, fast fibre bundles, and reduced dynamics simultaneously via iterative homological equations.
• The approach integrates techniques like Floquet averaging and blow-up analysis to extend classical Fenichel theory and capture complex multi-timescale behaviors in applications.

Searching arXiv for coordinate-independent geometric singular perturbation theory and closely related parametrization/averaging work. Coordinate-independent geometric singular perturbation theory (ci-GSPT) is a framework for analyzing singularly perturbed ordinary differential equations with multiple time scales without privileging a priori a specific separation into slow and fast coordinate components. Instead, slow–fast structure is encoded geometrically through critical manifolds, tangent bundles, fast fibre bundles, invariant splittings, and reduced vector fields defined intrinsically on manifolds. In this formulation, coordinate changes act by pushforward and pullback, while normal hyperbolicity, reduced flows, and persistence statements remain geometric objects. Recent developments have combined this viewpoint with the parametrization method, higher-order homological equations, and Floquet-theoretic averaging on manifolds of periodic orbits, thereby extending ci-GSPT from equilibria on critical manifolds to slow drift along normally hyperbolic manifolds foliated by periodic orbits (Jelbart et al., 2021, Lizarraga et al., 2020, Rink et al., 8 Sep 2025).

1. Geometric formulation and basic objects

In contrast to the classical “standard form” of Fenichel theory, where a system is written in coordinates

$$x' = \varepsilon g(x,y,\varepsilon), \qquad y' = h(x,y,\varepsilon),$$

ci-GSPT treats systems on manifolds in a coordinate-free manner. A basic formulation is

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$

where $M$ is a manifold of states, $f:M\to\mathbb{R}$ defines the critical manifold $S=\{z\in M:f(z)=0\}$, $N:M\to TM$ is a nonzero vector field spanning the fast direction, and $G$ collects the lower-order processes or fluxes. In a more general perturbative setting one considers

$$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$

with $S_0=\{y:f_0(y)=0\}$ a $k$-dimensional critical manifold (Jelbart et al., 2021, Lapuz et al., 5 Aug 2025).

The defining geometric feature is the splitting of the restricted tangent bundle into slow and fast directions. At points of a normally hyperbolic critical manifold,

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$0

so the tangent bundle $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$1 carries the reduced drift, while $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$2 carries the fast linear fibre directions. The corresponding coordinate-independent reduced problem is obtained by oblique projection along fast fibres: $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$3 with

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$4

This construction is invariant under smooth changes of variables: if $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$5 is a diffeomorphism, then $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$6 and $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$7 transform by pushforward and pullback, preserving $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$8, normal hyperbolicity, and reduced flows (Lapuz et al., 5 Aug 2025, Jelbart et al., 2021).

Normal hyperbolicity is measured through the fast linearization transverse to the critical manifold. In the scalar-fast-direction setting of dimension two, the nontrivial eigenvalue along $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$9 is

$M$0

while in higher dimensions one checks the nontrivial eigenvalues of $M$1 along $M$2. Attracting, repelling, and degenerate subsets are then determined by the sign or spectral location of these transverse eigenvalues (Jelbart et al., 2021, Lapuz et al., 5 Aug 2025).

2. Parametrization method and intrinsic homological equations

A major recent development is the use of the parametrization method to compute slow manifolds, fast fibre bundles, and the reduced dynamics simultaneously. In this approach one seeks an embedding $M$3 of the perturbed slow manifold and a reduced vector field $M$4 satisfying the conjugacy or invariance equation

$M$5

Formal expansions

$M$6

lead to a hierarchy of homological equations whose solvability is controlled by the intrinsic tangent/normal splitting (Lizarraga et al., 2020, Lapuz et al., 5 Aug 2025).

The first nontrivial reduced flow is

$M$7

and, order by order, the parametrization method yields explicit updates

$M$8

where $M$9 is built from lower-order terms and $f:M\to\mathbb{R}$0. In the equivalent formulation of the 2020 parametrization paper, one decomposes

$f:M\to\mathbb{R}$1

with $f:M\to\mathbb{R}$2 a gauge fixing the reparametrization freedom and $f:M\to\mathbb{R}$3 determined uniquely by solvability. The same method extends to the variational equation and computes the perturbed fast fibre bundle through an extended conjugacy equation for $f:M\to\mathbb{R}$4 (Lizarraga et al., 2020).

This intrinsic formulation is explicitly reparametrization-invariant. If $f:M\to\mathbb{R}$5 is a smooth reparametrization, then

$f:M\to\mathbb{R}$6

The geometry therefore resides in the invariant manifold and bundle structure rather than in a particular chart. A further consequence is the appearance of hidden timescales: the second-order term

$f:M\to\mathbb{R}$7

can generate infra-slow drift even when $f:M\to\mathbb{R}$8. This is the basis of the top-down, nested-timescale constructions developed for systems with three or more timescales (Lizarraga et al., 2020).

3. Periodic-orbit manifolds and coordinate-independent averaging

A substantial extension of ci-GSPT concerns slow–fast systems whose singular limit contains a manifold of periodic orbits rather than a manifold of equilibria. In that setting one considers

$f:M\to\mathbb{R}$9

and assumes that the layer problem $S=\{z\in M:f(z)=0\}$0 possesses a $S=\{z\in M:f(z)=0\}$1-parameter family of periodic orbits forming an embedded $S=\{z\in M:f(z)=0\}$2-dimensional manifold $S=\{z\in M:f(z)=0\}$3. With an embedding

$S=\{z\in M:f(z)=0\}$4

the unperturbed flow on $S=\{z\in M:f(z)=0\}$5 is $S=\{z\in M:f(z)=0\}$6, and Floquet theory yields exactly $S=\{z\in M:f(z)=0\}$7 zero exponents in tangent directions and $S=\{z\in M:f(z)=0\}$8 exponents with nonzero real part, uniformly bounded away from zero (Rink et al., 8 Sep 2025).

The coordinate-independent invariance equation is now posed for the embedding $S=\{z\in M:f(z)=0\}$9 of the perturbed invariant manifold and its reduced flow $N:M\to TM$0: $N:M\to TM$1 With expansions

$N:M\to TM$2

one obtains homological equations

$N:M\to TM$3

The Fredholm alternative is enforced באמצעות the adjoint variational equation

$N:M\to TM$4

and the averaged drift is recovered by projecting the homological equation onto $N:M\to TM$5. In maximally reduced gauge, the first-order drift is determined by

$N:M\to TM$6

Because the construction is derived from the invariance equation, the resulting reduced flow is invariant under smooth reparametrizations of phase and slow variables (Rink et al., 8 Sep 2025).

The main theorem in this setting gives a coordinate-independent Pontryagin–Rodygin theorem consistent with Fenichel persistence. For any finite order $N:M\to TM$7, there exist smooth series

$N:M\to TM$8

$N:M\to TM$9

such that the truncated pair satisfies

$G$0

with $G$1 independent of $G$2, and the reduced flow is maximally reduced, meaning independent of $G$3. On a section $G$4, true trajectories on the persistent NHIM $G$5 remain $G$6-close to the periodic fibre over the averaged trajectory for times $G$7. This recovers the classical averaging timescale while placing it within a geometrically consistent NHIM construction (Rink et al., 8 Sep 2025).

4. Multiple timescales, switching, and singular geometry

ci-GSPT is particularly effective when time-scale separation is region-dependent or hidden in higher-order terms. The parametrization framework formalizes an $G$8-timescale system through nested embeddings $G$9, reduced vector fields $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$0, and normally hyperbolic critical manifolds

$$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$1

with conjugacy relations

$$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$2

The leading infra-slow flow on $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$3 is produced by a double projection of $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$4, so three or more timescales can emerge even when the original system appears two-timescale (Lizarraga et al., 2020).

The calcium-dynamics application provides a process-oriented realization of this idea. There the system is cast in the ci-GSPT form $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$5, with five small parameters identified and related to a single perturbation parameter by a polynomial scaling law. The resulting singular perturbation problem has a time-scale separation that depends on the region of state space. In regime $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$6, $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$7, the critical manifolds are $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$8 and $$\frac{dy}{dt}=\mathcal F(y,\varepsilon)=\sum_{i=0}^j \varepsilon^i F_i(y), \qquad F_0(y)=N_0(y)f_0(y),$$9, with $S_0=\{y:f_0(y)=0\}$0 normally hyperbolic attracting for $S_0=\{y:f_0(y)=0\}$1 and $S_0=\{y:f_0(y)=0\}$2 everywhere degenerate. In regime $S_0=\{y:f_0(y)=0\}$3, $S_0=\{y:f_0(y)=0\}$4, rescaling yields a folded critical manifold $S_0=\{y:f_0(y)=0\}$5 with a regular fold $S_0=\{y:f_0(y)=0\}$6. The loss of normal hyperbolicity along $S_0=\{y:f_0(y)=0\}$7 and at the contact point $S_0=\{y:f_0(y)=0\}$8 is resolved by cylindrical and spherical blow-up, allowing the construction of a singular orbit $S_0=\{y:f_0(y)=0\}$9 and a global Poincaré map (Jelbart et al., 2021).

Under Assumption 1 of that paper, the main theorem states that for $k$0 the system has a unique stable relaxation cycle $k$1 satisfying

$k$2

with Floquet exponent bounded above by $k$3 for some $k$4. The period satisfies

$k$5

This demonstrates that ci-GSPT is not restricted to globally fixed slow/fast coordinates and that blow-up becomes necessary when normal hyperbolicity is lost (Jelbart et al., 2021).

5. Representative applications and model classes

The coordinate-independent viewpoint has been used across non-standard fast–slow systems, conductance-based neuronal models, and chemical reaction networks.

In the fast–slow Bazykin–Berezovskaya predator–prey model with Allee effect, the system is written as

$k$6

with critical manifold $k$7. Here neither $k$8 nor $k$9 is globally fast or slow; instead the fast dynamics is governed by $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$00 and the slow drift by $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$01. Normal hyperbolicity fails at $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$02, and the transition across that point is handled by a coordinate-independent entry–exit relation. The fast subsystem has the first integral

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$03

yielding a fast impact map $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$04, while the slow exit map $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$05 is determined by an explicit integral identity. Their composition organizes extinction, bistability, heteroclinic cycles, stable limit cycles, and singular Hopf behavior, with heteroclinic threshold $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$06 and Hopf threshold $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$07 (Borsotti et al., 14 Apr 2025).

In the multiple-timescale Hodgkin–Huxley equations, Fenichel theory is used first to reduce the original four-dimensional system to a global three-dimensional system on an attracting slow manifold $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$08, and then ci-GSPT is used again on the reduced problem. The critical manifold

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$09

has attracting and repelling sheets $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$10 and $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$11 separated by a fold set $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$12. In the $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$13-slow and $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$14-slow regimes one obtains further one-dimensional slow manifolds $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$15 and $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$16, folded singularities $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$17 and $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$18, and return maps whose displacement functions determine double-epoch MMOs, single-epoch MMOs, and, in the $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$19-slow regime, relaxation oscillation. The connected/aligned/remote classification of folded singularities is explicitly geometric and does not depend on a preferred chart (Kaklamanos et al., 2022).

In chemical reaction networks, ci-GSPT has been proposed as an alternative to selecting among sQSSA, tQSSA, and rQSSA. For Michaelis–Menten kinetics, the method provides a unique reduced model on any normally hyperbolic, attracting component of the critical manifold, and the 2025 study systematically explores parameter configurations across three orders of magnitude. It reports distinct model reductions for 14 relevant parameter configurations of the irreversible reaction scheme and 25 for the reversible reaction scheme, while also classifying fast fibre orientations into Class S, Class R, and Class T. For the Kim–Forger model, the same framework yields a new reduction without the need of a coordinate transformation, with critical manifold $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$20 and an explicit three-dimensional reduced system (Lapuz et al., 5 Aug 2025).

These applications show a common pattern: ci-GSPT replaces a priori variable selection by geometric structures determined by the vector field itself. This suggests that its main comparative advantage lies in regimes where standard-form reductions are ambiguous, region-dependent, or inaccessible without additional transformations.

6. Long-time validity, limitations, and common misconceptions

Beyond local-in-time reduction, ci-GSPT also supports convergence results on unbounded slow-time intervals. A coordinate-independent version of Hoppensteadt’s convergence theorem considers a family of vector fields

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$21

on a manifold $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$22, with critical manifold

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$23

Under a compact positively invariant set $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$24, a global parameterization of $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$25, normal hyperbolicity of the fast directions, and a Lyapunov function $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$26 for the reduced flow $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$27, the theorem yields uniform convergence on all closed subsets of $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$28 in slow time: $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$29 In the one-dimensional slow-manifold case, the Lyapunov function can be taken explicitly as

$$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$30

This result strengthens the usual compact-time Tikhonov–Fenichel reduction by supplying long-time shadowing after the initial layer (Lax et al., 2016).

The framework nevertheless has clear structural limitations. Across the cited works, the recurring assumptions are smooth dependence on parameters and charts, full rank of the embedding, and a uniform spectral gap separating neutral and hyperbolic directions. The periodic-orbit averaging construction requires smooth $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$31, $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$32, smooth Floquet projectors, and exclusion of singular points where periodic orbits collapse into equilibria. The CRN reductions are valid on normally hyperbolic attracting subsets. Near loss of normal hyperbolicity, such as folds, SNPO, or $$z' = H(z,\varepsilon) = N(z) f(z) + \varepsilon G(z,\varepsilon), \qquad z\in M,\quad 0<\varepsilon\ll 1,$$33 near SNIC, the basic Fenichel-based construction breaks down and blow-up or related techniques are required (Rink et al., 8 Sep 2025, Lapuz et al., 5 Aug 2025, Jelbart et al., 2021).

Several misconceptions are therefore excluded by the literature. ci-GSPT is not a replacement for Fenichel theory; rather, it is a coordinate-independent formulation that uses invariant manifolds, bundle splittings, and projectors to make Fenichel-type persistence constructive. It is not restricted to two timescales, since hidden timescales can emerge from higher-order homological terms and nested reductions. It is also not merely a notational reformulation of standard slow–fast form, because its central objects are the critical manifold, fast fibre bundle, and intrinsic reduced flow, all defined without committing to a global split into slow and fast coordinates (Lizarraga et al., 2020, Rink et al., 8 Sep 2025, Lapuz et al., 5 Aug 2025).

Taken together, these developments position ci-GSPT as a geometric synthesis of Tikhonov–Fenichel reduction, invariant manifold theory, parametrization methods, averaging, and blow-up analysis. Within its stated hypotheses, it yields reduced models that are intrinsic, systematically improvable to higher order, and explicitly tied to the geometry of the persistent invariant set.