Symplectic Phase Space GAN Overview

Updated 12 July 2026

The paper introduces SPS-GAN, which integrates a conditional GAN backbone with Hamiltonian latent dynamics to model multiple conservative systems without predefining latent configuration dimensions.
It employs a cyclic-coordinate regularizer to encourage sparse latent representations, enabling the discovery of symmetry and effective degrees of freedom.
Tested on synthetic systems like mass-spring oscillators and pendulums, SPS-GAN demonstrates competitive trajectory accuracy and low energy drift across varying system parameters.

Searching arXiv for the SPS-GAN paper and closely related symplectic/phase-space generative modeling work to ground the article in current arXiv records. Symplectic Phase Space GAN (SPS-GAN) is a conditional generative model for dynamical systems that is designed to generate physically plausible trajectories across multiple mechanical systems and parameter settings while inferring a compact latent configuration space rather than assuming that the system’s generalized coordinates are known in advance. Introduced in “Physically Plausible Multi-System Trajectory Generation and Symmetry Discovery” (Liu et al., 26 Sep 2025), SPS-GAN combines a conditional GAN backbone with Hamiltonian latent dynamics and a cyclic-coordinate regularizer, with the stated aims of trajectory prediction, video generation, and symmetry discovery. Its defining claim is that one conditional model can represent families of conservative systems, operate from low-dimensional state-space measurements or raw video frames, and uncover low-dimensional latent structure associated with effective degrees of freedom.

1. Problem setting and conceptual scope

SPS-GAN addresses three linked problems. First, it seeks to generate trajectories that are dynamically coherent and approximately consistent with Hamiltonian structure, including low energy drift and rollouts that resemble true solutions of conservative dynamics. Second, it is intended to model multiple systems and parameter settings within one model rather than one fixed system with fixed physical parameters. Third, it does not assume that the relevant configuration-space dimension is already known, and instead aims to discover the configuration-space structure from observations (Liu et al., 26 Sep 2025).

In the reported experiments, the model is trained on five synthetic systems: mass-spring oscillator, ideal pendulum, double pendulum, planar two-body, and planar three-body. Conditioning variables include system identity and physical parameters, denoted by $\xi$ . This conditional structure is central: the model is intended to learn a family of Hamiltonians parameterized by system identity and physical properties rather than a single fixed Hamiltonian.

The paper positions SPS-GAN against prior mechanics-inspired neural models such as Hamiltonian Neural Networks, symplectic recurrent models, Hamiltonian Generative Networks, and related physics-informed video models. Those earlier approaches are characterized as typically handling one dynamical system at a time, with fixed physical parameters, and often assuming that the configuration-space dimension is already known. SPS-GAN explicitly targets the harder regime in which the system varies and the latent coordinates are not given.

A further conceptual distinction is modality. The method is described as being able to discover configuration-space structure “from arbitrary measurement types,” with concrete instantiations for low-dimensional Cartesian trajectories and RGB videos. The stronger claim about arbitrary measurements is supported by the learned mapping into a latent phase/configuration space and by adversarial learning in observation space; however, the full experimental instantiations are SPS-GAN-traj and SPS-GAN-video rather than a broader catalog of modalities.

2. Architecture and latent phase-space construction

SPS-GAN is presented as a three-stage conditional GAN with adversarial supervision in observation space. The common backbone consists of a configuration-space map, a Hamiltonian latent recurrent motion model, an observation-space decoder or generator, and observation-space discriminators (Liu et al., 26 Sep 2025).

Variant	Observation generator	Discriminator design
SPS-GAN-traj	MLP $G_T$	Recurrent discriminator $D_T$
SPS-GAN-video	CNN $G_I$	Image discriminator $D_I$ and video discriminator $D_V$

The first stage is the configuration-space map

$f(\epsilon_m,\xi): \mathbb{R} \to \mathbb{R}^{2\times d_{\text{lat}}},$

which maps a random motion sample $\epsilon_m$ and conditioning variable $\xi$ to an initial latent phase-space state

$\mathbf z_0=(q_0,p_0).$

The latent state is interpreted as a canonical-like representation with $G_T$ 0 coordinate variables and $G_T$ 1 momentum variables. The paper describes this as discovering the latent configuration or phase-space structure rather than prescribing it.

The second stage is a Hamiltonian neural network recurrent module conditioned on $G_T$ 2, with Hamiltonian

$G_T$ 3

Latent rollouts are generated by integrating Hamilton’s equations: $G_T$ 4 The recurrent latent module is implemented with an HNN and integrated with a symplectic Leapfrog scheme, following Symplectic Recurrent Neural Network. In the background discussion the paper assumes a separable Hamiltonian,

$G_T$ 5

and gives Leapfrog updates as

$G_T$ 6

$G_T$ 7

$G_T$ 8

The notation appears swapped relative to standard Leapfrog conventions; that is how the paper writes it. The operational point is that SPS-GAN uses a symplectic integrator rather than naive Euler rollout in order to reduce energy drift over long rollouts.

The third stage appends a content latent $G_T$ 9 to each latent motion state,

$D_T$ 0

so that motion and content are separated. For trajectories, the decoder is an MLP

$D_T$ 1

with generated frame

$D_T$ 2

A system-dependent binary mask $D_T$ 3 zeros out inactive particles: $D_T$ 4 For video, the decoder is a CNN

$D_T$ 5

with generated frame

$D_T$ 6

A subtle architectural point is the absence of an explicit encoder from observations into latent phase space. The generation pipeline is noise $D_T$ 7 latent phase state $D_T$ 8 latent rollout $D_T$ 9 decoded observations. The latent structure is therefore learned indirectly through observation-space decoding and adversarial objectives rather than through an explicit autoencoding or reconstruction pathway.

3. Objectives, symplectic bias, and symmetry discovery

The mechanics prior begins with Hamilton’s equations,

$G_I$ 0

and the observation that if a coordinate $G_I$ 1 does not contribute to the Hamiltonian, then

$G_I$ 2

Such coordinates are cyclic or ignorable. SPS-GAN uses this relation as the basis for its symmetry-discovery mechanism (Liu et al., 26 Sep 2025).

For trajectory generation, the written adversarial objective is

$G_I$ 3

For video generation, the paper gives

$G_I$ 4

These objectives place adversarial matching in observation space rather than latent space. The conditioning on $G_I$ 5 is conceptually that of a conditional GAN, even though the explicit loss expressions in the paper omit $G_I$ 6 inside discriminator notation.

The key additional term is the cyclic-coordinate loss,

$G_I$ 7

Because cyclic coordinates satisfy $G_I$ 8, penalizing $G_I$ 9 encourages some latent coordinates to become ignorable. The authors interpret this as encouraging sparse latent usage, minimizing the learned phase-space dimension, and identifying effective degrees of freedom.

In this framework, “symmetry discovery” does not mean symbolic recovery of conservation laws or exact analytical canonical coordinates. Rather, it means discovering a latent representation in which some coordinates become cyclic, so that the learned motion manifold has lower effective dimensionality and exhibits interpretable reductions consistent with Hamiltonian evolution. It is therefore more precise to say that SPS-GAN discovers minimal latent structure consistent with Hamiltonian dynamics than that it proves recovery of unique canonical coordinates.

4. Training protocol and reported empirical evidence

The data are synthetic simulations of five systems: mass-spring oscillator, ideal pendulum, double pendulum, planar two-body, and planar three-body. All systems are simulated with RK45 and time step $D_I$ 0, producing trajectories of 30 frames each. The video data are generated using the reimplementation pipeline of HGN. SPS-GAN-traj is trained for 50,000 epochs with batch size 128 for most systems and 160 for three-body; SPS-GAN-video is trained for 50,000 epochs with batch size 16. Optimization uses Adam with

$D_I$ 1

The configuration-space map and HNN are MLPs with hidden size 100 and ReLU activation. For trajectories, $D_I$ 2 is an MLP with hidden size 512 and Softplus activation. For video, $D_I$ 3 is a CNN with 32 filters and 3 channels. The latent dimensions are

$D_I$ 4

and for trajectories the output size is

$D_I$ 5

(Liu et al., 26 Sep 2025).

For single-system trajectory modeling, SPS-GAN is compared with supervised HNN. The evaluation uses the first generated Cartesian coordinate $D_I$ 6 as the initial condition and measures rollout MSE against ground truth.

System	SPS-GAN MSE	HNN MSE
Mass-spring	$D_I$ 7	$D_I$ 8
Pendulum	$D_I$ 9	$D_V$ 0
Double pendulum	$D_V$ 1	$D_V$ 2
Two-body	$D_V$ 3	$D_V$ 4
Three-body	$D_V$ 5	$D_V$ 6

These values are the basis for the claim that SPS-GAN achieves “performance on par with supervised models designed for single systems,” while outperforming HNN on the double pendulum and three-body examples. The paper also reports energy drift within $D_V$ 7 of ground-truth conserved energy on the pendulum example.

For symmetry and configuration-space discovery, the latent dimension is fixed at $D_V$ 8, yet t-SNE visualizations of the learned motion manifold suggest lower intrinsic dimensionality: 1D for two-body, 2D for double pendulum, and 2D for the constrained three-body special case. The paper contrasts this with FastICA, which does not recover similarly meaningful structure.

For single-system video generation, SPS-GAN is compared with HGAN and HGN using Fréchet Video Distance over 2048 generated videos of length 16.

System	SPS-GAN FVD	HGAN / HGN FVD
Mass-spring	25.63	45.68 / 385.08
Pendulum	40.57	91.64 / 688.12
Double pendulum	24.12	73.21 / 331.94
Two-body	87.12	105.85 / 830.91
Three-body	89.08	1981.10 / 451.40

The reported gains are especially large on the more difficult systems. On the chaotic double pendulum, the paper highlights a reduction in FVD by more than an order of magnitude relative to HGN. For multiple systems with parameter variation, reported FVDs are 182.15 for varied color plus varied physics, 135.38 for constant color plus varied physics, and 135.28 for varied color plus constant physics. The authors interpret these results as evidence that SPS-GAN can disentangle motion from appearance through the separation of motion latent $D_V$ 9/Hamiltonian rollout and content latent $f(\epsilon_m,\xi): \mathbb{R} \to \mathbb{R}^{2\times d_{\text{lat}}},$ 0, and that it extends to unseen system configurations and parameters.

5. Relation to adjacent symplectic and phase-space generative modeling

SPS-GAN is most directly connected to “Hamiltonian GAN” (Allen-Blanchette, 2023), which already used a GAN-based video generation pipeline with a learned configuration-space map, Hamiltonian latent dynamics, and a cyclic-coordinate loss, but in a single-system setting. SPS-GAN extends that line by making the model explicitly conditional on system labels and physical parameters, and by targeting multi-system generation and parameter generalization.

The method also sits near several strands of symplectic and phase-space machine learning. “Deep Neural Networks with Symplectic Preservation Properties” (He et al., 2024) proposes exactly invertible symplectomorphisms built from q-shearing, p-shearing, and symplectic stretching blocks, offering a hard architectural route to phase-space maps that preserve canonical geometry. “Symplectic Methods in Deep Learning” (Maslovskaya et al., 2024) and “Reduced-order modeling of Hamiltonian dynamics based on symplectic neural networks” (Chen et al., 16 Aug 2025) develop exact-by-construction symplectic architectures for deep learning and reduced-order modeling, including higher-order explicit symplectic partitioned Runge–Kutta networks and HénonNet-based symplectic embeddings. “Neural Canonical Transformation with Symplectic Flows” (Li et al., 2019) treats canonical transformations themselves as generative models, mapping physical variables to a latent system of independent harmonic oscillators.

A different but adjacent direction is phase-space generative modeling without strict symplecticity. “Generative Modeling with Phase Stochastic Bridges” (Chen et al., 2023) uses phase-space state augmentation $f(\epsilon_m,\xi): \mathbb{R} \to \mathbb{R}^{2\times d_{\text{lat}}},$ 1 and stochastic bridge dynamics to improve few-step generation, but does not impose Hamiltonian or symplectic structure. “Generative models on phase space” (Bogorad et al., 2 Apr 2026) shows how diffusion and flow matching can be confined exactly to relativistic Lorentz-invariant phase space by parameterization, emphasizing manifold-exact constraint preservation rather than canonical symplecticity. “Learning Physical Systems: Symplectification via Gauge Fixing in Dirac Structures” (Papatheodorou et al., 23 Jun 2025) addresses constrained and dissipative systems by learning a lift into an augmented nondegenerate symplectic space. Together, these works indicate that SPS-GAN belongs to a broader effort to combine generative modeling with exact or approximate physical geometry, though its particular combination of conditional GAN training, Hamiltonian latent rollout, and cyclic-coordinate sparsification is distinctive.

6. Strengths, limitations, and technical interpretation

The principal strengths claimed for SPS-GAN are that it unifies multi-system conditional generation with mechanics-informed latent dynamics, does not require prior knowledge of the true configuration-space dimension, supports both trajectory generation and video generation, and works from low-dimensional Cartesian observations or raw image sequences. Empirically, it shows low energy drift, strong FVD relative to prior physics-informed video generators, and trajectory accuracy competitive with supervised single-system HNN baselines despite modeling broader families of systems (Liu et al., 26 Sep 2025).

The method also has clear limitations. The experiments are on synthetic conservative systems rather than noisy real-world data. System labels and physical parameters are provided as conditioning variables rather than inferred. For trajectories, a system-specific output mask is assumed known from the dataset. The evidence for symmetry discovery is empirical and visualization-based, with no identifiability theorem for the learned coordinates. The latent “configuration space” is interpretable in a reduced-dimensional sense, but the paper does not prove that the learned coordinates are unique, canonical, or directly physically meaningful in the strict analytical-mechanics sense. Dissipative, controlled, or contact-rich systems are not treated, and sensitivity to noise, partial observability, or out-of-distribution measurements is not analyzed.

A common misconception is to treat SPS-GAN as an autoencoding Hamiltonian model. The reported architecture does not include an explicit encoder from observations into latent phase space, nor reconstruction losses, KL regularization, supervised derivative losses, or explicit parameter inference objectives. It is more accurately described as a generative latent-structure model trained through adversarial matching in observation space, with Hamiltonian latent rollout and a cyclic-coordinate regularizer.

In technical significance, SPS-GAN marks a shift from learning the Hamiltonian of one known system to learning a conditional generative latent phase space over many systems while discovering compact structure in that space. This suggests a broader research program in which structured generative models are asked not only to reproduce observations, but also to recover latent mechanics and symmetry-relevant coordinates from the observation distribution itself.