HénonNet: Neural Symplectic Maps

• HénonNet is a neural architecture that composes symplectic Hénon maps to learn discrete, flux-preserving transformations in Hamiltonian systems.
• It replaces standard affine layers with potential-driven symplectic layers, enabling rapid computation of neural Poincaré maps and accurate phase-space diagnostics.
• The design supports reduced-order modeling and time-adaptive integration, though challenges remain in faithfully capturing global phase-space topology.

Searching arXiv for papers on HénonNet and closely related Hénon-map neural architectures. HénonNet denotes a class of structure-preserving neural architectures built from compositions of Hénon maps in order to learn symplectic transformations exactly. It was introduced as a feed-forward architecture for fast neural Poincaré maps of toroidal magnetic-field line flow, where the central requirement is not merely predictive accuracy but exact preservation of the flux-preserving geometry of the underlying return map. In the subsequent literature, HénonNet has been used more broadly as a symplectic, invertible map-based network for Hamiltonian dynamics, phase-space diagnostics, reduced-order modeling, and time-adaptive learning on irregular time grids (Burby et al., 2020).

1. Foundational construction

In its original formulation, HénonNet replaces generic affine-nonlinear layers by elementary Hénon maps. For $n$-dimensional variables $x,y$, the basic map is

$$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$

where $V:\mathbb{R}^n\to\mathbb{R}$ is a scalar potential and $\eta\in\mathbb{R}^n$ is a shift. Each such map is symplectic. The architecture relies on a theorem, cited from Turaev, that every symplectic map can be approximated arbitrarily well on compact sets by compositions of Hénon maps; specifically, $H[V,\eta]^{4N}$ can approximate any symplectic mapping to arbitrary accuracy, and the factor $4$ is emphasized because the fourth iterate of a trivial Hénon map is the identity. A Hénon layer is therefore defined as the fourfold composition

$L[V[W],\eta]=H[V[W],\eta]^4,$

with $V[W]$ implemented as a scalar feed-forward neural network, and a full HénonNet is a composition of such layers (Burby et al., 2020).

A persistent feature across later formulations is that HénonNet is defined through exact symplectic building blocks rather than through unconstrained function approximation. Later papers use related, but not identical, coordinate conventions. One formulation writes the base Hénon map as

$$H(V,\eta) \begin{pmatrix} \mathbf{x}\ \mathbf{y} \end{pmatrix} = \begin{pmatrix} \mathbf{y}+\eta\ \mathbf{x}+\nabla V(\mathbf{y}) \end{pmatrix},$$

with a Hénon layer $x,y$0, while another uses

$x,y$1

again composed four times per layer. These variations preserve the same architectural principle: exact symplecticity by construction, explicit invertibility, and composition-based approximation of symplectic dynamics (Chen et al., 16 Aug 2025).

2. Original application: fast neural Poincaré maps

The initial HénonNet application concerns a toroidal domain $x,y$2 with poloidal cross section $x,y$3 and divergence-free magnetic field $x,y$4. A field line is the trajectory of

$x,y$5

and the Poincaré map $x,y$6 sends a point to its next return to $x,y$7. Conventional evaluation of $x,y$8 requires numerical field-line integration between section crossings, so if there are $x,y$9 timesteps between intersections, $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$0 returns cost roughly $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$1. Learning $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$2 directly reduces this to repeated evaluation of a learned map $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$3, costing $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$4 with $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$5 (Burby et al., 2020).

The key physical constraint is flux preservation. For any region $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$6,

$$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$7

In toroidal coordinates $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$8, with $$H[V,\eta] \begin{pmatrix} x\ y \end{pmatrix} = \begin{pmatrix} y+\eta\ -x+\nabla V(y) \end{pmatrix},$$9, the map is shown to preserve a time-dependent $V:\mathbb{R}^n\to\mathbb{R}$0-form

$V:\mathbb{R}^n\to\mathbb{R}$1

and after the coordinate change

$V:\mathbb{R}^n\to\mathbb{R}$2

this becomes the canonical form $V:\mathbb{R}^n\to\mathbb{R}$3. Hence the field-line Poincaré map is canonical symplectic: $V:\mathbb{R}^n\to\mathbb{R}$4 HénonNet was designed to reproduce this structure exactly: each Hénon map is symplectic, the composition of symplectic maps is symplectic, and therefore every Hénon layer and every HénonNet satisfies the symplectic or flux-preserving constraint exactly (Burby et al., 2020).

Training in the original study is supervised. Sample points $V:\mathbb{R}^n\to\mathbb{R}$5 are drawn in the domain, labels $V:\mathbb{R}^n\to\mathbb{R}$6 are produced by a conventional field-line-following algorithm, typically RK4, and the network minimizes

$V:\mathbb{R}^n\to\mathbb{R}$7

The reported implementation uses Adam optimization and fully connected scalar-potential networks with $V:\mathbb{R}^n\to\mathbb{R}$8 activations. Accuracy is assessed by held-out test loss, long recursive rollouts, Poincaré-plot comparison, pointwise trajectory comparison, and conservation of the Hamiltonian in test Hamiltonian systems (Burby et al., 2020).

The speed advantage reported in the original paper is summarized below.

System	HénonNet	Reference integrator
Pendulum flow map, 2000 iterations	2.6 s	vectorized RK: 7.8 s
Perturbed pendulum, 1000 iterations	1.3 s	vectorized RK: 12.1 s
RMP-like magnetic field, 1000 iterations	4.26 s	vectorized RK: 34.2 s

Beyond timing, the magnetic-field experiments emphasize representational capacity. In the resonant magnetic perturbation case, the learned HénonNet reproduces the inner regular region, the chaotic sea, and a large magnetic island chain. When trained for fewer epochs, it first learns coiled hyperbolic invariant manifolds that generate a sticky chaotic region at the intended island location. The paper interprets these filaments as corresponding to low fast Lyapunov indicator regions and as likely acting like stable or unstable manifolds of hyperbolic periodic points, thereby producing long residence times near the target island location (Burby et al., 2020).

3. Position within structure-preserving neural dynamics

HénonNet belongs to a broader class of physics-constrained models for Hamiltonian systems, but its inductive bias is distinct. A conventional feed-forward neural network may learn an arbitrary vector field; Hamiltonian neural networks instead output a scalar Hamiltonian $V:\mathbb{R}^n\to\mathbb{R}$9, from which the vector field is obtained through Hamilton’s equations,

$\eta\in\mathbb{R}^n$0

In the Hénon–Heiles benchmark, Hamiltonian neural networks were shown to respect energy conservation and phase-space volume preservation more faithfully than conventional neural networks, to recover the Lyapunov-spectrum symmetry $\eta\in\mathbb{R}^n$1, and to reproduce the order-to-chaos transition measured by the smaller alignment index $\eta\in\mathbb{R}^n$2 with the threshold $\eta\in\mathbb{R}^n$3 for chaotic motion (Choudhary et al., 2019).

HénonNet differs from that scalar-Hamiltonian strategy by learning a symplectic map directly. In the original comparison with SympNet, SympNet layers are described as trainable linear symplectic maps plus a fixed symplectic activation, whereas HénonNet layers use a trainable scalar potential inside a symplectic Hénon-map construction. The full network is therefore map-centric rather than vector-field-centric. Later work places HénonNet within an even broader hierarchy, noting that HénonNet layers can be viewed as special cases of the more general GHNN or G-SympNet framework, while still retaining the specific advantages of Hénon-map composition and exact inversion (Burby et al., 2020).

This architectural distinction matters operationally. HénonNet is naturally suited to learning discrete return maps, reduced flow maps, or latent one-step symplectic updates. Hamiltonian neural networks are naturally suited to learning a continuous-time generator. The two approaches therefore occupy different positions in the structure-preserving modeling landscape rather than standing in a simple dominance relation. A plausible implication is that HénonNet is most naturally deployed where the target object is itself a discrete symplectic transformation, such as a Poincaré map or learned integrator (Choudhary et al., 2019).

4. Invertibility, diagnostics, and limitations

Later work on phase-space integrity treats HénonNet as a physically constrained architecture alongside SympNet and Generalized Hamiltonian Neural Networks. In that setting, HénonNet is defined by Hénon maps built from a learnable scalar potential $\eta\in\mathbb{R}^n$4 and a constant bias term $\eta\in\mathbb{R}^n$5, with a HénonNet layer given by four successive applications of the same Hénon map. The architecture is exactly invertible, and the inverse of the full network is obtained by reversing the order of layer inverses. This exact invertibility is especially relevant for Lagrangian Descriptor computations, which require both forward and backward propagation (Hasmi et al., 1 Apr 2026).

The same paper argues that trajectory error and energy preservation do not suffice to evaluate learned Hamiltonian models. It introduces Lagrangian Descriptors

$\eta\in\mathbb{R}^n$6

and converts LD fields into weighted probability density functions

$\eta\in\mathbb{R}^n$7

which are then compared using

$\eta\in\mathbb{R}^n$8

The purpose is to assess preservation of phase-space topology, including stable and unstable manifolds, separatrices, homoclinic tangles, and fixed-point structure, rather than merely short-horizon pointwise prediction (Hasmi et al., 1 Apr 2026).

Under this diagnostic, HénonNet exhibits clear limitations. In the Duffing oscillator, it learns the existence of the homoclinic structure but distorts the orbit geometry, especially in the lower half-plane of momentum space; its LD-weighted PDF shows asymmetric lobes and incorrect curvature, and its KL divergence is the largest among the four methods considered. In the three-mode nonlinear Schrödinger equation, it performs poorly again, with large phase shifts, distorted density distributions, and displaced or missing fixed points inside the orbit. Reservoir Computing, despite lacking explicit physical constraints, reproduces the homoclinic structure with higher fidelity in that study (Hasmi et al., 1 Apr 2026).

These results directly address a recurrent misconception: exact symplecticity and exact invertibility do not by themselves guarantee faithful reconstruction of global phase-space topology. HénonNet preserves the canonical geometric constraint it is designed to preserve, but later diagnostics show that topology-sensitive structures such as separatrices and homoclinic boundaries can still be misplaced or distorted (Hasmi et al., 1 Apr 2026).

5. Reduced-order and time-adaptive extensions

A substantial extension of the HénonNet program appears in symplectic reduced-order modeling. There, HénonNets are used both as a symplectic encoder-decoder for latent-space discovery and as a symplectic latent flow map for dynamics learning. The base Hénon map is written as

$\eta\in\mathbb{R}^n$9

a Hénon layer is $H[V,\eta]^{4N}$0, and a full HenonNet is

$H[V,\eta]^{4N}$1

The inverse is explicit: $H[V,\eta]^{4N}$2 and the network inverse is obtained by reversing the layer order. The paper also describes HenonNets as a symplectic generalization of triangular normalizing flows and combines them, optionally, with linear G-reflector layers $H[V,\eta]^{4N}$3 (Chen et al., 16 Aug 2025).

In that reduced-order framework, the encoder is built from a full-space HénonNet followed by a truncation operator, the decoder uses the inverse HénonNet composed with the inverse G-reflector and canonical inclusion, and the latent flow map is itself another HénonNet. Training uses a unified end-to-end loss combining reconstruction, one-step prediction, and Hamiltonian preservation: $H[V,\eta]^{4N}$4 The reported values are $H[V,\eta]^{4N}$5 for the linear and parametric wave examples and $H[V,\eta]^{4N}$6 for the nonlinear Schrödinger example. Additional strategies include multi-step autoregressive training and small Gaussian noise injection; training data are generated with the symplectic Störmer–Verlet method, and optimization uses Adam (Chen et al., 16 Aug 2025).

Selected reconstruction errors from that study show the scale of improvement reported for nonlinear symplectic embeddings.

System	HénonNet	HénonNet + G-reflector
Linear wave	$H[V,\eta]^{4N}$7	$H[V,\eta]^{4N}$8
Parametric linear wave	$H[V,\eta]^{4N}$9	$4$0
Nonlinear Schrödinger	$4$1	$4$2

A separate line of development addresses irregularly sampled data. Time-adaptive HénonNets, or T-HénonNets, modify the layer to depend explicitly on the step size $4$3: $4$4 with four compositions per layer and finite compositions across layers. For non-autonomous Hamiltonians, NAT-HénonNets additionally allow explicit time dependence in $4$5 and advance the time variable inside each layer. The main universal approximation results are restricted to separable Hamiltonians $4$6 in the autonomous case and $4$7 in the non-autonomous case; the paper explicitly shows that non-separable Hamiltonians remain outside the representable class. Numerically, T-HénonNet succeeds on the pendulum, fails on the non-separable linear Hamiltonian $4$8, and NAT-HénonNet succeeds on the forced harmonic oscillator where time-blind T-HénonNet fails because identical phase-space points can occur at different times (Janik et al., 24 Sep 2025).

6. Relation to the broader Hénon literature

Although HénonNet is a neural architecture rather than a classical dynamical system, its name and design sit within a wider Hénon and Hénon-like literature characterized by rich symbolic, geometric, and statistical structure. For the non-Archimedean quadratic Hénon family

$4$9

one study classifies parameter regions into good-reduction, attractor-bearing, asymmetric, and horseshoe regimes. In region $L[V[W],\eta]=H[V[W],\eta]^4,$0, the two-sided filled Julia set is a trapped attractor in Milnor’s sense; for infinitely many maps over $L[V[W],\eta]=H[V[W],\eta]^4,$1, this attractor is uncountably infinite, has Haar measure zero, contains no periodic points, and supports an SRB-type measure that equidistributes forward orbits. In region $L[V[W],\eta]=H[V[W],\eta]^4,$2, assuming $L[V[W],\eta]=H[V[W],\eta]^4,$3 is a square, the dynamics on the filled Julia set is topologically conjugate to the two-sided shift on bisequences in two symbols (Allen et al., 2016).

In complex dynamics, every complex Hénon map has been shown to possess a measure of maximal entropy that is exponentially mixing of all orders for Hölder observables and satisfies the Central Limit Theorem. The paper states the order-$L[V[W],\eta]=H[V[W],\eta]^4,$4 mixing rate in terms of

$L[V[W],\eta]=H[V[W],\eta]^4,$5

and writes the CLT variance as

$L[V[W],\eta]=H[V[W],\eta]^4,$6

These results describe a strong probabilistic regularity in the underlying Hénon family (Bianchi et al., 2023).

Other recent work on Hénon-like maps develops a renormalization theory for non-perturbative dissipative systems $L[V[W],\eta]=H[V[W],\eta]^4,$7 with bounded combinatorics. It shows that infinitely regularly renormalizable Hénon-like maps converge super-exponentially to the same one-dimensional renormalization attractor as unimodal maps, and that every infinitely renormalizable map is regularly unicritical, with a unique orbit of tangencies between strong-stable and center manifolds (Yang, 2024). This suggests that the term “Hénon” in HénonNet carries more than nominal significance: the underlying map family is associated, in the mathematical literature, with exact symplecticity or related geometric structure together with attractors, horseshoes, symbolic dynamics, and renormalization phenomena. HénonNet operationalizes only part of that legacy—chiefly the symplectic map structure—while later evaluation work shows that preserving this part exactly does not automatically recover the full global organization of phase space (Hasmi et al., 1 Apr 2026).