Nonperturbative Guiding-Center Models

• Nonperturbative guiding-center models are reduced descriptions of charged particle motion using exact invariants that bypass the limitations of small-parameter series.
• They utilize techniques like action-angle formulations, canonical Hamiltonian structures, and data-driven surrogates to capture dynamics in complex magnetic geometries.
• The approach improves accuracy in regimes where standard asymptotic expansions fail, such as in energetic particle orbits in stellarators and tokamaks.

Searching arXiv for recent and foundational papers on nonperturbative guiding-center models and closely related formulations. Nonperturbative guiding-center models are reduced descriptions of charged-particle dynamics in magnetized plasmas that seek to eliminate the fast cyclotron timescale without relying exclusively on an asymptotic expansion in the small parameter $\epsilon=\rho/L$, where $\rho$ is a gyroradius and $L$ a characteristic field scale. In the perturbative guiding-center framework, the reduced dynamics is organized as a power series in $\epsilon$, with the magnetic moment treated as an adiabatic invariant. By contrast, nonperturbative formulations aim to represent the reduced dynamics through exact or non-asymptotic invariants, exact action-angle constructions in special geometries, canonical or constrained Hamiltonian structures valid beyond specific coordinate choices, or data-driven surrogates learned directly from full-orbit dynamics (Burby et al., 2024). The term therefore denotes not a single formalism but a family of approaches that share the objective of retaining guiding-center efficiency while extending validity to regimes where conventional asymptotics may degrade, such as strong gradients, complex toroidal geometry, or energetic-particle orbits in stellarators and tokamaks (Brizard, 2017).

1. Historical and conceptual foundations

Classical guiding-center theory is based on separating fast gyromotion from slow drift motion in strong magnetic fields, typically under the ordering $\epsilon=\rho/L\ll 1$. In this setting, one introduces a near-identity transformation from particle variables to guiding-center variables and constructs an adiabatic invariant identified with the magnetic moment. The standard Lie-transform framework yields a Hamiltonian reduced system with noncanonical Poisson brackets and a hierarchy of corrections ordered in $\epsilon$ (Hirvijoki et al., 2014).

A nonperturbative guiding-center model departs from this logic in one of several distinct senses. One sense is exactness with respect to the underlying invariant: the reduced model is built from a quantity that is an exact constant of motion for the full Lorentz dynamics in a special class of fields and whose asymptotic expansion reproduces Kruskal’s invariant series to all orders in the perturbative regime (Hollas et al., 13 Aug 2025). Another sense is non-asymptotic validity in $\epsilon$: the reduced description is judged against exact or numerically resolved orbit averages and shown to remain accurate even when $\epsilon$ is not asymptotically small (Brizard, 2017). A third sense is nonperturbativity in geometry rather than finite-Larmor-radius ordering: the magnetic field may have islands, broken flux surfaces, or stochastic regions, while the reduced guiding-center equations are still represented canonically and integrated symplectically (Albert et al., 19 Mar 2025).

A related but distinct strand treats the reduced model itself as exact within a learned invariant manifold. In that formulation, the key object is a non-perturbative adiabatic invariant $\mathcal{J}$, interpreted as the Hamiltonian of a hidden $U(1)$ symmetry of the full Lorentz dynamics. Once $\rho$0 is known, the reduced guiding-center dynamics on a Poincaré section becomes an exact Hamiltonian system on the orbit space of that symmetry (Burby et al., 2024). This suggests that the central issue is often not whether fast gyromotion exists, but whether it can be quotiented out through a genuine invariant beyond the radius of convergence of the usual asymptotic series.

2. Exact invariants and action-based formulations

The most rigorous nonperturbative constructions presently available arise in symmetric magnetic fields, where the Lorentz dynamics is Liouville integrable. In fields with a two-parameter continuous symmetry group, one can construct exact action-angle coordinates for the full orbit and identify one action as the gyro-action $\rho$1, which serves as the nonperturbative invariant $\rho$2 (Hollas et al., 13 Aug 2025). This invariant is exact for all $\rho$3 in the symmetric geometry, while its power-series expansion agrees with Kruskal’s adiabatic invariant series to all orders when $\rho$4 (Hollas et al., 13 Aug 2025).

The slab field $\rho$5 and screw-pinch fields provide explicit examples. In the slab case, the exact action $\rho$6 is obtained by integrating the Liouville one-form over the cyclotron cycle on an invariant torus; the resulting Hamiltonian vector field generated by $\rho$7 satisfies Kruskal’s defining properties of the roto-rate in the perturbative limit (Hollas et al., 13 Aug 2025). In the screw-pinch case, the existence of exact Noether invariants associated with $\rho$8- and $\rho$9-translations reduces the full dynamics to a one-dimensional radial problem with an exact radial action integral $L$0 (Brizard, 12 Jan 2026).

That radial action integral is especially important because it gives a direct nonperturbative representation of the magnetic moment. In the doubly symmetric screw-pinch, the exact invariant

$L$1

matches the perturbative guiding-center magnetic-moment expansion through first order in magnetic-field nonuniformity and thereby represents the magnetic moment as a nonperturbative integral invariant (Brizard, 12 Jan 2026). This does not abolish perturbation theory; rather, it identifies the perturbative magnetic moment as the asymptotic series of an exact action in an integrable geometry.

A closely related structural point is that these exact invariants support reduced equations that eliminate the fast cyclotron angle without loss of fidelity in the symmetric setting. In the screw pinch, the nonperturbative guiding-center equations defined on a Poincaré section $L$2 reproduce the exact slow dynamics at successive gyro-crossings, thereby giving exact reduced predictions while suppressing the explicit cyclotron timescale (Hollas et al., 13 Aug 2025). This establishes a benchmark for nonperturbative guiding-center theory: exactness is achievable when the symmetry structure is strong enough to support a genuine global gyro-action.

3. Beyond asymptotic small-$L$3 validity

A central motivation for nonperturbative models is the observation that the formal condition $L$4 is sufficient but not always necessary for accurate guiding-center behavior in orbit-averaged quantities. In the exactly solvable problem of a particle in a straight magnetic field with constant transverse gradient,

$L$5

the full transverse motion can be solved in terms of Jacobi elliptic functions for arbitrary $L$6, with $L$7 (Brizard, 2017). Exact orbit averages then provide a stringent nonperturbative reference.

Two exact quantities are particularly informative. The orbit-averaged displacement is

$L$8

while the exact drift velocity is

$L$9

with $\epsilon$0 and $\epsilon$1 complete elliptic integrals (Brizard, 2017). Standard guiding-center theory predicts $\epsilon$2 and $\epsilon$3, formulas derived under $\epsilon$4. Yet the comparison shows excellent agreement well beyond the formal asymptotic domain: at $\epsilon$5, $\epsilon$6 versus $\epsilon$7, and $\epsilon$8 versus $\epsilon$9; at $\epsilon=\rho/L\ll 1$0, the errors are about $\epsilon=\rho/L\ll 1$1 and $\epsilon=\rho/L\ll 1$2, respectively (Brizard, 2017).

This does not mean that the perturbative derivation becomes nonperturbative. Rather, it shows that the practical domain of validity of the reduced dynamics may be controlled less rigidly by the formal parameter $\epsilon=\rho/L\ll 1$3 than by the persistence of smooth, periodic, near-adiabatic orbit structure. A plausible implication is that nonperturbative guiding-center modeling should not be understood solely as replacement of perturbation theory, but also as systematic calibration of reduced equations against exact orbit averages in regimes where formal asymptotics are borderline but adiabatic organization remains robust (Brizard, 2017).

The opposite regime is illustrated by energetic particles in realistic stellarator fields. In that context, the perturbative series for the magnetic moment may cease to approximate any useful invariant even when phase-space portraits still exhibit many invariant circles. In a 2D model field

$\epsilon=\rho/L\ll 1$4

with $\epsilon=\rho/L\ll 1$5, the standard asymptotic magnetic-moment series improves conservation at $\epsilon=\rho/L\ll 1$6, but for $\epsilon=\rho/L\ll 1$7 and $\epsilon=\rho/L\ll 1$8 higher-order truncations do not improve conservation and the optimal truncation collapses (Burby et al., 2024). This is precisely the regime where a genuinely non-perturbative invariant becomes necessary rather than optional.

4. Data-driven non-perturbative guiding-center dynamics

A data-driven non-perturbative guiding-center model has been proposed for exactly this energetic-particle regime. The formulation assumes a hidden non-perturbative $\epsilon=\rho/L\ll 1$9 symmetry of the full Lorentz dynamics with generator $\epsilon$0 and Hamiltonian $\epsilon$1, where $\epsilon$2 is an exact conserved quantity satisfying $\epsilon$3 under the Lorentz-force Poisson bracket (Burby et al., 2024). The guiding-center phase space is then identified with a global Poincaré section $\epsilon$4 intersecting each $\epsilon$5 orbit once, and the reduced dynamics is the Hamiltonian footpoint flow on $\epsilon$6.

The general reduced Poisson bracket on $\epsilon$7 depends only on the ambient bracket and on $\epsilon$8. For the special case $\epsilon$9, the non-perturbative guiding-center equations reduce to

$\epsilon$0

with

$\epsilon$1

(Burby et al., 2024). In the perturbative regime, substituting the asymptotic magnetic-moment series into these exact reduced equations recovers a perturbative guiding-center model. Outside that regime, the approach instead learns $\epsilon$2 directly from full-orbit data.

The learning procedure is notable for being constrained and linear rather than black-box. The invariant is represented spectrally on $\epsilon$3 using $\epsilon$4 Fourier$\epsilon$5Fourier$\epsilon$6Chebyshev modes, and determined by minimizing a Rayleigh quotient that combines a dynamical residual $\epsilon$7, which enforces agreement with drift velocities reconstructed from full-orbit data, and an invariant residual $\epsilon$8, which enforces invariance under the Poincaré map (Burby et al., 2024). Because both residuals are quadratic in the coefficients, the problem reduces to a single generalized eigenvalue problem.

The learned $\epsilon$9 reproduces full-orbit invariant circles substantially better than perturbative truncations at $\epsilon$0 and $\epsilon$1. More importantly, the resulting Poincaré map $\epsilon$2 outperforms the best available perturbative map by orders of magnitude over most of the section; for $\epsilon$3, higher-order perturbative truncations even become singular in the denominator $\epsilon$4, while the learned model remains well behaved (Burby et al., 2024). In this framework, “nonperturbative” means that all finite-Larmor-radius effects are absorbed into a globally learned invariant, rather than approximated by an $\epsilon$5-series.

This approach remains field-specific and relies on a substantial measure of quasiperiodic trajectories, so it is not a universal closed-form theory. Still, it suggests a route by which nonperturbative guiding-center dynamics can be operationalized in realistic stellarator optimization, where $\epsilon$6-particle confinement depends on orbit topology not faithfully captured by perturbative invariants (Burby et al., 2024).

5. Hamiltonian, canonical, and geometric formulations

A major line of research treats nonperturbativity not as replacement of the reduced model, but as reformulation of its geometric structure in a way that is robust to general magnetic topology or global coordinate obstructions.

One canonical approach shows that guiding-center dynamics can be described as a constrained canonical Hamiltonian system in six-dimensional phase space, with Hamiltonian

$\epsilon$7

subject to two constraints enforcing $\epsilon$8 (Zhang et al., 2024). The physical flow lies on a four-dimensional symplectic submanifold of the canonical phase space, and the guiding center is interpreted as a pseudo-particle with intrinsic magnetic moment $\epsilon$9. The resulting velocity decomposes into parallel motion plus curvature, grad-$\mathcal{J}$0, and $\mathcal{J}$1 drift terms that arise from the constraint forces (Zhang et al., 2024). This construction is nonperturbative in the sense that the canonicalization itself does not rely on an expansion in $\mathcal{J}$2, although the underlying Littlejohn Lagrangian remains the starting point.

A related canonical symplectic-structure approach embeds the dynamics in an extended canonical phase space and reorganizes it so that gyrorotation, field-aligned motion, and cross-field drift each correspond to a canonical pair (Neishtadt et al., 2019). In that formulation, the fast gyrodegree of freedom becomes an exact harmonic oscillator with action $\mathcal{J}$3, and the magnetic moment is $\mathcal{J}$4 (Neishtadt et al., 2019). This does not remove the strong-field ordering, but it renders the leading invariant structurally exact and amenable to standard canonical adiabatic theory, which is closer in spirit to an action-based nonperturbative model than the usual noncanonical Lie-transform construction.

For general toroidal magnetic fields, including perturbed tokamak fields with magnetic islands and stochastic regions, canonical coordinates can be constructed by a purely spatial transformation plus a gauge transformation, without requiring nested flux surfaces or magnetic-flux coordinates (Albert et al., 19 Mar 2025). The guiding-center Lagrangian then takes canonical form

$\mathcal{J}$5

with the canonical momenta defined directly from transformed vector-potential components (Albert et al., 19 Mar 2025). This is nonperturbative in magnetic geometry: islands, stochastic layers, and lower single-null divertor topology are treated directly rather than as perturbations of an integrable flux-surface configuration.

The numerical consequence is that fully symplectic schemes, such as the implicit midpoint rule in canonical coordinates, preserve invariants and orbit topology far better than non-symplectic schemes at comparable computational cost. In axisymmetric tests, the toroidal canonical momentum $\mathcal{J}$6 is conserved to numerical precision and energy remains bounded with small oscillations; in perturbed tokamak fields, island chains and banana-tip precession are preserved rather than smeared by secular numerical drift (Albert et al., 19 Mar 2025). The article’s abstract explicitly identifies these developments as a step toward gyrokinetic models that conserve physical invariants (Albert et al., 19 Mar 2025).

There is also a coordinate-free geometric formulation in which gyro-symmetry is represented as an $\mathcal{J}$7-action on a seven-dimensional extended phase space endowed with a Lagrangian one-form (Yu, 2021). In that picture, the guiding-center manifold is the orbit space of Kruskal’s rings, gyro-averaging is defined as group averaging over the $\mathcal{J}$8-action, and the obstruction to a global gyro-angle is expressed through the non-exactness of a connection-like one-form $\mathcal{J}$9, rather than through local coordinate singularities (Yu, 2021). This suggests that a fully nonperturbative model should be formulated globally on the symmetry bundle rather than in any local gyrophase chart.

6. Non-Hamiltonian extensions and conservative structure

Not all physically important guiding-center problems are Hamiltonian. Radiation reaction in nonuniform magnetic fields provides a clear example. In a classical relativistic point-charge model with Landau–Lifshitz radiation reaction, the Lorentz force dominates and the radiation-reaction force is ordered as a slow dissipative perturbation satisfying $U(1)$0, while $U(1)$1 controls magnetic nonuniformity (Hirvijoki et al., 2014). The guiding-center transformation can then be extended to the dissipative force by applying a push-forward to the phase-space continuity equation and defining effective reduced force components through projection coefficients and gyrophase averaging (Hirvijoki et al., 2014).

The resulting reduced equations are

$U(1)$2

where $U(1)$3 is not Hamiltonian (Hirvijoki et al., 2014). To first order in magnetic nonuniformity, the paper derives explicit expressions for the guiding-center position, parallel-momentum, and magnetic-moment damping terms. For example,

$U(1)$4

so radiation reaction breaks the exact invariance of $U(1)$5 and the conservation of $U(1)$6 (Hirvijoki et al., 2014).

This model is still perturbative in $U(1)$7, but it is nonperturbative in the sense that the full deterministic radiation-reaction force is projected into reduced phase space without a small pitch-angle or linearized approximation. The distinction matters conceptually: a nonperturbative guiding-center model need not be Hamiltonian, provided the reduced force respects the geometry of the reduced phase space. The article itself states that the model is “Hamiltonian + fully consistent dissipative RR at first order in field nonuniformity” and “effectively non-Hamiltonian” once radiation reaction is included (Hirvijoki et al., 2014).

A separate issue concerns conservation properties under coordinate choices. In Boozer coordinates for toroidal confinement, the guiding-center Lagrangian contains a noncanonical term proportional to the covariant component $U(1)$8. Dropping this term yields a canonical form commonly used in orbit-following codes, but the full unabridged Lagrangian retains it and remains a noncanonical Hamiltonian system (Bierwage et al., 2022). The paper proves that with the $U(1)$9 term retained, the equations preserve phase-space volume and energy-like invariants for a fixed-frequency, single-$\rho$00 shear Alfvén perturbation; inconsistent omission of certain accompanying small terms can break either energy conservation alone or both energy and Liouville conservation (Bierwage et al., 2022). The broader implication is that nonperturbative reduced models must respect the conservative structure of the phase-space Lagrangian even when certain coefficients appear small.

7. Scope, limitations, and current directions

The various nonperturbative guiding-center models differ sharply in scope. Exact action-based models are currently restricted to highly symmetric magnetic fields such as slabs and screw pinches, where Noether invariants render the full orbit integrable (Hollas et al., 13 Aug 2025). Their main value lies in establishing rigorous benchmarks and clarifying what a nonperturbative invariant should mean. Canonical and geometric reformulations broaden the admissible magnetic geometry, but most still inherit the basic strong-magnetization reduction from conventional guiding-center theory and therefore are not nonperturbative in finite-Larmor-radius ordering (Zhang et al., 2024).

Data-driven models reach regimes inaccessible to perturbation theory in practice, especially for energetic particles in stellarators, but they are field-specific and depend on training data and on the existence of a substantial quasiperiodic region in phase space (Burby et al., 2024). Symplectic canonicalization for general toroidal fields is geometry-nonperturbative and numerically robust, yet it presently uses the standard leading-order guiding-center model and does not include higher-order finite-Larmor-radius or time-dependent gyrocenter effects (Albert et al., 19 Mar 2025). Coordinate-free geometric formulations clarify global topology and gyro-symmetry, but the explicit reduced equations presented are first-order in the inhomogeneity parameter $\rho$01 (Yu, 2021).

Several concrete extension paths recur across the literature. One is higher-order action matching: in the screw pinch, the equality between the radial action and the guiding-center magnetic moment has been verified through first order in nonuniformity, and extension to higher order is explicitly proposed (Brizard, 12 Jan 2026). Another is orbit averaging beyond gyro averaging, particularly for runaway-electron kinetics with radiation reaction (Hirvijoki et al., 2014). A third is embedding canonical reduced models into fully structure-preserving gyrokinetic solvers, especially in the presence of islands and stochastic magnetic fields (Albert et al., 19 Mar 2025). A fourth is learning families of nonperturbative invariants across configuration space rather than training separately for each field, which the data-driven work identifies as a future possibility (Burby et al., 2024).

A recurring misconception is that “nonperturbative guiding-center” implies a complete abandonment of asymptotic reasoning. The surveyed work suggests a more nuanced picture. Exact nonperturbative invariants are currently available only in special cases; elsewhere, perturbative structure still organizes the geometry, the symplectic form, or the learned ansatz. What changes is the target: instead of treating the asymptotic series itself as the reduced model, nonperturbative approaches seek a global invariant, a global canonical structure, or an exact reduced Poincaré dynamics of which the perturbative series is merely the local asymptotic shadow (Hollas et al., 13 Aug 2025).

In plasma applications, this distinction is consequential. Fusion-born $\rho$02-particles in stellarators can have gyroradii large enough that perturbative magnetic-moment expansions lose their optimal truncation and fail to approximate any useful invariant, while orbit structure remains organized by invariant circles (Burby et al., 2024). In toroidal devices with strong 3D perturbations, preserving the Hamiltonian structure over long times requires canonical coordinates and symplectic integration even if the reduced equations are still formally leading order (Albert et al., 19 Mar 2025). In radiating relativistic populations, the reduced model must incorporate deterministic dissipation without violating the reduced phase-space geometry (Hirvijoki et al., 2014). For these reasons, the nonperturbative guiding-center model has emerged less as a single theory than as a research program: to identify reduced variables, invariants, and geometric structures that remain faithful when asymptotic smallness is weakened but the multiscale character of magnetized motion endures.