Normally Hyperbolic Invariant Tori

• Normally Hyperbolic Invariant Tori are tori whose tangent dynamics is dominated by transverse contraction and expansion, ensuring robust persistence under small perturbations.
• They serve as organizing centers in dynamical systems, emerging in models like Hénon-like maps and periodic forcing, and play a key role in bifurcation analyses (e.g., secondary Hopf bifurcation).
• Numerical and analytical methods such as ergodic averaging and computer-assisted proofs are used to verify invariant splittings, detect bundle collisions, and quantify hyperbolicity breakdown.

Searching arXiv for the papers on arXiv and closely related work on normally hyperbolic invariant tori and manifolds. Normally hyperbolic invariant tori are invariant tori whose tangent dynamics is dominated by transverse contraction and expansion, so that the torus fits the general framework of a normally hyperbolic invariant manifold. In the standard formulation used for invariant manifold theory, one has an invariant splitting

$T_MQ = TM \oplus E^+ \oplus E^-,$

together with growth and decay estimates in the tangent, unstable, and stable bundles; for compact tori this is the usual NHIM setting, while in concrete applications the same structure appears as attracting quasiperiodic invariant circles, bifurcating $2$-tori in extended phase space, or torus leaves inside larger normally hyperbolic objects such as cylinders and laminations (Eldering, 2012, Linroth, 2019, Pereira et al., 14 Jul 2025).

1. Definition and conceptual framework

The general NHIM definition used in the literature places an invariant torus within a splitting

$T_MQ = TM \oplus E^+ \oplus E^-,$

with continuous globally bounded projections and $D\Phi^t$-invariance, together with estimates

$$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$

$$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$

The stronger $r$-normal hyperbolicity condition is

$\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$

A compact invariant torus $M=\mathbb T^n$ is a special case of a compact NHIM, and if it is normally attracting then the standard persistence theorem applies directly under small perturbation (Eldering, 2012).

In the map setting, an attracting quasiperiodic invariant circle provides the one-dimensional instance of a normally hyperbolic invariant torus: a $2$0-torus $2$1 embedded in $2$2, with one tangent direction carrying quasiperiodic dynamics and a two-dimensional normal bundle carrying contraction. For such circles the tangent Lyapunov exponent is $2$3, and normal hyperbolicity means, informally, that the normal contraction dominates the tangent dynamics and that the invariant splitting exists regularly (Linroth, 2019).

This framework is distinct from several adjacent notions. A compact invariant torus may be normally hyperbolic without being intrinsically hyperbolic, while an Anosov torus is defined instead by the induced action on $2$4, equivalently by the condition that $2$5 is isotopic to a hyperbolic automorphism. The classification of Anosov tori in irreducible $2$6-manifolds therefore concerns intrinsic torus dynamics rather than domination in the ambient phase space (Hertz et al., 2010).

A second essential distinction is between NHIM persistence and KAM persistence. NHIM theory assumes exponential contraction or expansion transverse to the torus and yields robustness under $2$7-small perturbations, whereas KAM theory treats normally elliptic or fully center-like tori and depends on frequency vectors, small divisors, and nondegeneracy assumptions. In this sense, normally hyperbolic invariant tori belong to a different persistence mechanism from KAM tori, even though both are toroidal invariant sets (Eldering, 2012).

2. Canonical realizations and bifurcation mechanisms

One concrete realization is the attracting invariant circle in a three-dimensional Hénon-like map

$2$8

with $2$9. The invariant circle $T_MQ = TM \oplus E^+ \oplus E^-,$0 is parameterized by $T_MQ = TM \oplus E^+ \oplus E^-,$1 satisfying

$T_MQ = TM \oplus E^+ \oplus E^-,$2

where $T_MQ = TM \oplus E^+ \oplus E^-,$3 is an orientation-preserving circle map with rotation number

$T_MQ = TM \oplus E^+ \oplus E^-,$4

Along $T_MQ = TM \oplus E^+ \oplus E^-,$5 one assumes a splitting

$T_MQ = TM \oplus E^+ \oplus E^-,$6

where $T_MQ = TM \oplus E^+ \oplus E^-,$7 is the tangent bundle and $T_MQ = TM \oplus E^+ \oplus E^-,$8 is the contracting normal bundle, with

$T_MQ = TM \oplus E^+ \oplus E^-,$9

This is the simplest nontrivial toroidal NHIM setting represented in the cited literature (Linroth, 2019).

A second mechanism is bifurcation from a periodic orbit. In a $D\Phi^t$0 one-parameter family

$D\Phi^t$1

a secondary Hopf bifurcation of the flow is encoded by a Neimark–Sacker bifurcation of the Poincaré map. Under the hypotheses

$D\Phi^t$2

$D\Phi^t$3

$D\Phi^t$4

$D\Phi^t$5

and nondegeneracy of the first Lyapunov coefficient $D\Phi^t$6, the classical secondary Hopf theorem produces a unique invariant torus on the side

$D\Phi^t$7

The main theorem of the 2025 paper strengthens this by proving that the bifurcating torus is normally hyperbolic; it has exactly $D\Phi^t$8 stable directions in the subcritical case and $D\Phi^t$9 stable directions in the supercritical case (Pereira et al., 14 Jul 2025).

Periodically forced planar systems fit the same scheme after passage to extended phase space. For

$$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$0

with $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$1-periodic time dependence, the autonomous reformulation on $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$2 is

$$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$3

Here the extended phase space is three-dimensional, so the invariant torus has codimension one, and the averaging theorem strengthened in (Pereira et al., 14 Jul 2025) shows that the torus produced by the secondary Hopf mechanism is normally hyperbolic and repelling if $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$4, or attracting if $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$5 (Pereira et al., 14 Jul 2025).

A third setting is more indirect but geometrically instructive. In a periodically driven rank-$$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$6-saddle reaction model, the primary normally hyperbolic object is a time-dependent NHIM rather than a torus. However, when the restricted dynamics on that NHIM is examined by a stroboscopic Poincaré map, a period-one fixed point corresponding to a transition-state trajectory is surrounded by near-integrable invariant curves. In the autonomous extended formulation these invariant curves are naturally interpreted as $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$7-tori lying within the NHIM, so the paper provides numerical evidence for torus-like invariant sets on a normally hyperbolic manifold even though it does not claim that the NHIM itself is a torus (Tschöpe et al., 2020).

3. Persistence, regularity, and generalized settings

Persistence theory for normally hyperbolic invariant manifolds extends directly to tori but with important qualifications. For noncompact NHIMs in bounded geometry, the main theorem assumes a complete Riemannian manifold of bounded geometry, a $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$8 vector field, and a connected complete submanifold $$\forall t\in\mathbb R,\ (m,x)\in TM:\quad \|D\Phi^t(m)x\|\le C_M e^{\rho_M |t|}\|x\|,$$9 that is $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$0-normally hyperbolic with empty unstable bundle,

$$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$1

Then, for every sufficiently small $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$2, there exists $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$3 such that if

$$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$4

there is a unique invariant submanifold $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$5 in the $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$6-neighborhood of $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$7, diffeomorphic to $$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$8, with

$$\forall t\le 0,\ (m,x)\in E^+:\quad \|D\Phi^t(m)x\|\le C_+ e^{\rho_+ t}\|x\|,$$9

For a normally attracting torus $$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$0, this yields direct persistence. The same work emphasizes that regularity is limited by the spectral gap: even if the system is $$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$1, the persisted NHIM may fail to be $$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$2 when the gap only supports $$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$3 regularity (Eldering, 2012).

The same theme appears in the Banach-space theory of partially normally hyperbolic invariant manifolds. There, the center bundle may strictly contain the tangent bundle,

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$4

and the main trichotomy theorem constructs local center-stable, center-unstable, and center manifolds

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$5

for generalized dynamical systems modeled by correspondences. In the torus case, a whiskered torus is defined by an embedding $$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$6 satisfying

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$7

together with a splitting

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$8

If

$$\forall t\ge 0,\ (m,x)\in E^-:\quad \|D\Phi^t(m)x\|\le C_- e^{\rho_- t}\|x\|.$$9

and $r$0, then the torus is normally hyperbolic. The theory then provides $r$1, $r$2, $r$3, and strong stable and unstable laminations $r$4, $r$5, even for ill-posed PDEs modeled by correspondences (Chen, 2019).

A weaker persistence result is available when post-perturbation rate conditions fail. In that topological setting, one starts from a compact NHIM $r$6 for a diffeomorphism $r$7 with splitting

$r$8

and estimates

$r$9

The perturbed map need only satisfy a covering relation

$\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$0

in a bundle neighborhood $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$1. The result is then persistence to an invariant set that projects onto the whole base $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$2, but that set need not be a manifold. For a normally hyperbolic torus, this means that one can retain an invariant torus-shaped set near the original torus even after losing the usual rate conditions, but not necessarily a persisted torus in the classical Fenichel sense (Capinski et al., 2018).

4. Breakdown and loss of hyperbolicity

The numerical study of three-dimensional Hénon-like maps shows that persistence boundaries for quasiperiodic normally hyperbolic invariant circles need not be detected by Lyapunov exponents alone. Along contour segments in parameter space where the attracting invariant circle has constant irrational rotation number, especially the golden mean, the minimum angle between the tangent bundle and the slow contracting bundle tends to zero near the endpoint of the contour, while the Lyapunov exponents remain separated: one stays at $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$3, and the other two remain negative and apparently distinct (Linroth, 2019).

In the Oseledets splitting

$\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$4

the relevant exponents on the attractor satisfy

$\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$5

with $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$6 tangent and $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$7 normal. The principal observation is that the failure of hyperbolicity occurs through the geometry of the splitting: the tangent bundle $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$8 and the slow contracting bundle $\rho_- < -r\rho_M \le 0 \le r\rho_M < \rho_+.$9 collide and lose smoothness, even though the exponents do not approach each other. The paper states this as “the collision and loss of smoothness of two of the invariant Lyapunov bundles while the Lyapunov exponents all remain distinct” (Linroth, 2019).

This phenomenon is a major correction to a common simplification. Spectral separation alone is not always sufficient to detect the boundary of persistence. A torus can lose effective normal hyperbolicity through degeneration of the invariant splitting while transverse exponents remain bounded away from the tangential exponent. The paper therefore identifies bundle geometry, minimum angles between tangent and normal bundles, and regularity of the bundle fields as primary diagnostics in computations (Linroth, 2019).

The same source also highlights a second misconception. The attracting invariant circle itself may remain visually smooth at macroscopic scale while the conjugacy to rigid rotation loses regularity and the Lyapunov bundles become nonsmooth. Breakdown is therefore not necessarily the sudden geometric destruction of the torus; it can instead be a collapse of the regular invariant splitting on a thin set, with the attractor still appearing globally circle-like (Linroth, 2019).

5. Detection, computation, and numerical diagnostics

Several complementary computational paradigms appear in the literature. One is orbit-based ergodic averaging. For attracting quasiperiodic invariant circles, the Weighted Birkhoff method replaces standard Birkhoff sums by

$M=\mathbb T^n$0

with a $M=\mathbb T^n$1 bump $M=\mathbb T^n$2. For Diophantine rotation number and sufficiently regular observables,

$M=\mathbb T^n$3

This makes it possible to compute rotation numbers, Lyapunov exponents, and Lyapunov bundles efficiently enough to detect the bundle-collision mechanism described above (Linroth, 2019).

A second approach is direct stabilization of dynamics on a normally hyperbolic manifold. In the periodically driven reaction model, the NHIM is represented by a neural network regression map

$M=\mathbb T^n$4

with $M=\mathbb T^n$5 input neurons, hidden layers of sizes $M=\mathbb T^n$6, and $M=\mathbb T^n$7 output neurons. The network is trained on $M=\mathbb T^n$8 points for $M=\mathbb T^n$9 epochs with Adam at learning rate $2$00. During time stepping, one keeps $2$01 and corrects the drifted reaction variables back to the learned NHIM values after every step. This does not learn the reduced vector field directly; it learns the embedding of the NHIM and uses it as a projection mechanism. The resulting stabilized trajectories allow Poincaré surfaces of section on the NHIM, identification of the transition-state periodic orbit as a fixed point, and observation of invariant curves interpreted as sections of $2$02-tori in the extended system (Tschöpe et al., 2020).

A third approach is computer-assisted topological verification. For maps, one can prove existence of a NHIM from a numerical guess without assuming a perturbative regime. The method uses local boxes, covering relations

$2$03

cone inequalities based on quadratic forms

$2$04

and forward/backward bounds computed from interval Jacobians. The main theorem then yields continuous monomorphisms

$2$05

with $2$06 invariant. Applied to a driven logistic map, this proves that numerical evidence for a chaotic attractor is false and that the attractor is a normally hyperbolic invariant curve (Capinski et al., 2011).

A fourth computational paradigm appears in celestial mechanics. In the restricted planar elliptic three-body problem at the $2$07 Kirkwood gap, the analysis begins with a normally hyperbolic invariant cylinder

$2$08

foliated by $2$09-dimensional invariant tori. From its homoclinic channels one constructs a separatrix map and then, via isolating blocks, a normally hyperbolic invariant lamination

$2$10

The induced dynamics on this NHIL is a partially hyperbolic skew-shift of the form

$2$11

This shows how torus leaves inside a NHIM can generate larger symbolic hyperbolic structures rather than remaining the sole organizing objects (Guardia et al., 20 Mar 2026).

6. Applications, extensions, and broader geometric landscape

Normally hyperbolic invariant tori now appear in several broad application domains. In polynomial vector fields, a three-dimensional analogue of Hilbert’s $2$12th problem replaces planar limit cycles by isolated invariant tori. For

$2$13

the counting functions

$2$14

$2$15

measure isolated invariant tori and normally hyperbolic invariant tori, respectively. An explicit lifting mechanism turns a planar system with $2$16 hyperbolic limit cycles into a $2$17-dimensional system with at least $2$18 normally hyperbolic invariant tori, using

$2$19

This yields

$2$20

and, via a Christopher–Lloyd replication argument,

$2$21

A later paper proves the strict monotonicity statement

$2$22

whenever $2$23 is finite, using torus bifurcation near Hopf–Zero equilibria (Novaes et al., 2022, Arakaki et al., 18 Apr 2025).

Near resonances in nearly integrable Hamiltonian systems, however, the principal normally hyperbolic object is often not a torus but a cylinder. At strong double resonance in a $2$24-degree-of-freedom Hamiltonian, one obtains a $2$25-dimensional normally hyperbolic invariant cylinder crossing the resonance, even though the reduced slow dynamics on $2$26 is generically chaotic. In a related three-degree-of-freedom setting, large NHICs cover the resonant path except inside neighborhoods of size

$2$27

around finitely many strong double resonant points. These cylinders are built from families of periodic orbits rather than quasiperiodic torus leaves, and they organize diffusion across resonance (Kaloshin et al., 2012, Cheng, 2015).

The torus viewpoint also extends to partially hyperbolic fibrations. For accessible center-bunched fibred partially hyperbolic systems with $2$28-dimensional center leaves, the general trichotomy states that one has either distinct center Lyapunov exponents, or an invariant continuous line field or pair of line fields tangent to the center leaves, or a continuous conformal structure on the center leaves invariant under both the dynamics and stable and unstable holonomies. In the torus case, a neighborhood of an elliptic affine extension

$2$29

supports a rigidity theorem: under the nonhyperbolic alternatives, the dynamics is topologically conjugate to an $2$30-affine extension of the Anosov base (2207.13236).

Finally, several papers clarify that normal hyperbolicity is a robust organizing principle even when no torus is present. In asymptotically Kerr spacetimes, the trapped set is a codimension-two normally hyperbolic invariant manifold in the characteristic set, with codimension-one stable and unstable manifolds, and the analysis deliberately avoids relying on complete integrability or torus foliations. This reinforces a final structural point: normally hyperbolic invariant tori are one important special case of a broader NHIM framework, but current research repeatedly shows that cylinders, laminations, and trapped sets may be the more stable global objects, with tori appearing as leaves, bifurcating submanifolds, or local sections of the larger normally hyperbolic geometry (Hintz, 2018).