Neural Canonical Transformation (NCT)

• Neural Canonical Transformation (NCT) is a cross-disciplinary design pattern that uses neural networks to learn canonical representations that preserve key system symmetries.
• It applies to domains like Hamiltonian mechanics, many-body physics, 3D vision, and control by mapping complex structures into simpler latent forms such as harmonic oscillators.
• Its methodologies range from enforcing symplectic and unitary conditions to adaptive canonicalization, addressing challenges like non-uniqueness and optimization ambiguities.

Neural Canonical Transformation (NCT) denotes a family of methods in which a neural model learns a canonical transformation, canonical representation, or canonical coordinate system that simplifies a target problem while preserving the relevant structure of the original system. In classical Hamiltonian mechanics, NCT is a symplectic neural network that maps phase-space variables to latent canonical coordinates with an independent harmonic oscillator Hamiltonian (Li et al., 2019). In finite-temperature many-body physics and vibrational spectroscopy, it is a variational density-matrix or wavefunction construction based on unitary coordinate transforms, autoregressive occupation models, and normalizing flows (Xie et al., 2021, Xie et al., 2022, Wang et al., 28 Sep 2025). In geometric learning, 3D vision, and neural fields, the same pattern appears as learned canonicalization of group orbits, poses, coordinate frames, or operator bases, including pose-canonicalized neural radiance fields, learned canonicalization functions, adaptive canonicalization, canonical factors for factored neural fields, and learned phase-correlation bases (Agaram et al., 2022, Kaba et al., 2022, Lin et al., 29 Sep 2025, Yi et al., 2023, Reynolds, 16 Jun 2026). In nonlinear control, analogous neural canonical control structures learn state and input transformations into Brunovsky form (Ecker et al., 2023). Taken together, these usages suggest that NCT is best understood as a cross-disciplinary design pattern rather than a single architecture.

1. Canonical transformation as an organizing principle

Across the literature, the term “canonical” retains its classical meaning: a change of variables or representation that converts a difficult object into a simpler one without discarding the structure that matters. The learned object can be a symplectic map, a unitary basis change, a pose-canonicalized field, a group-orbit representative, or a state/input transformation for feedback linearization. What is shared is the attempt to make symmetry, dynamics, or geometry explicit rather than implicit.

Domain	Canonical object	Representative paper
Hamiltonian mechanics	$x=(p,q)\mapsto z=(P,Q)$ with latent independent harmonic oscillator Hamiltonian	(Li et al., 2019)
Interacting fermions	Variational density matrix over canonically transformed Slater determinants	(Xie et al., 2021, Xie et al., 2022)
Vibrational spectra	Normalizing-flow coordinate transform of a harmonic-oscillator basis	(Wang et al., 28 Sep 2025)
3D vision / neural fields	Arbitrary-pose NeRF to canonical field	(Agaram et al., 2022)
Geometric learning	Learned or adaptive canonicalization over group orbits	(Kaba et al., 2022, Lin et al., 29 Sep 2025, Reynolds, 16 Jun 2026)
Nonlinear control	State and input transforms to sampled-data Brunovsky form	(Ecker et al., 2023)

A common source of confusion is that NCT does not designate one mathematical formalism. In some papers the transformation is required to be symplectic, in others unitary, equivariant, or merely approximately invertible. In some settings the canonical object is a latent variable, whereas in others it is an explicit pose, a canonical coordinate field, or a canonicalized input to a downstream backbone. This diversity is not a contradiction; it reflects the fact that canonicalization is being used to solve different structural problems.

2. Symplectic and Hamiltonian formulations

The formulation that explicitly introduced the name “Neural Canonical Transformation” models canonical transformations of classical Hamiltonian systems with symplectic neural networks (Li et al., 2019). A system with $n$ degrees of freedom is written in phase-space coordinates $x=(p,q)\in\mathbb{R}^{2n}$, with Hamiltonian dynamics

$\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$

A canonical transformation $\mathcal{T}:x\mapsto z=(P,Q)$ has Jacobian $M$ satisfying

$MJM^T=J.$

This symplectic condition preserves phase-space volume and the Hamiltonian form of the dynamics.

In this framework, the neural network $\Phi_\theta:x\mapsto z$ is constrained to be symplectic and is trained so that the latent Hamiltonian becomes

$K(z)=\sum_{k=1}^n \frac{P_k^2+\omega_k^2Q_k^2}{2},$

a sum of independent harmonic oscillators with learnable frequencies $\omega_k>0$. The corresponding latent equilibrium density is Gaussian, with diagonal covariance determined by the $n$0. The learned latent variables are therefore canonical collective modes ordered by dynamical time scale.

The main experimental architecture is a neural point transformation. Coordinates are transformed by an invertible neural map

$n$1

implemented with Real NVP, and momenta are transformed by

$n$2

Because this construction is canonical by design, the symplectic property is enforced structurally rather than by a penalty. The paper also describes linear symplectic transformations via the matrix exponential of the symplectic Lie algebra and continuous symplectic flows generated by $n$3 (Li et al., 2019).

Two training regimes are given. When the Hamiltonian is known analytically, NCT minimizes the variational free energy

$n$4

which is a KL objective against the physical canonical ensemble. When only samples are available, NCT is trained by maximum likelihood on the phase-space density. The reported applications include a two-dimensional ring potential, a harmonic chain, alanine dipeptide, and MNIST, with the latent frequencies identifying slow collective modes and enabling latent interpolation and conceptual compression (Li et al., 2019).

3. Variational density matrices, interacting fermions, and vibrational spectra

A closely related line of work uses neural canonical transformations as variational finite-temperature ansätze for interacting fermions in the continuum (Xie et al., 2021). The density matrix is written as

$n$5

where $n$6 is a discrete probabilistic model and each basis state is obtained by a common unitary transformation $n$7 applied to a reference Slater determinant $n$8. Restricting to point transformations yields an explicit coordinate map $n$9 such that

$x=(p,q)\in\mathbb{R}^{2n}$0

At the level of densities, this is a continuous normalizing flow. Because Monte Carlo estimators depend on $x=(p,q)\in\mathbb{R}^{2n}$1, the method avoids the usual path-integral fermion sign problem in low-temperature thermodynamics (Xie et al., 2021).

The same variational philosophy was adapted to the two-dimensional spin-polarized electron gas to extract the quasiparticle effective mass $x=(p,q)\in\mathbb{R}^{2n}$2 from entropy rather than finite-size excitation energies (Xie et al., 2022). The density matrix is

$x=(p,q)\in\mathbb{R}^{2n}$3

where $x=(p,q)\in\mathbb{R}^{2n}$4 is modeled by a Causal Transformer over momentum occupations and $x=(p,q)\in\mathbb{R}^{2n}$5 is a canonically transformed Slater determinant in coordinate space. The entropy per particle is obtained directly from the autoregressive model,

$x=(p,q)\in\mathbb{R}^{2n}$6

and at low temperature the Fermi-liquid relation

$x=(p,q)\in\mathbb{R}^{2n}$7

is used. The calculation reports a suppression of the effective mass in the two-dimensional spin-polarized electron gas, more pronounced in the low-density strong-coupling region (Xie et al., 2022).

NCT was later extended to vibrational spectra and excited states of fluxional molecules, notably $x=(p,q)\in\mathbb{R}^{2n}$8, using atomic Cartesian coordinates rather than normal modes (Wang et al., 28 Sep 2025). Starting from a latent harmonic-oscillator basis $x=(p,q)\in\mathbb{R}^{2n}$9, a Real-NVP flow $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$0 defines physical-space wavefunctions

$\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$1

All states are optimized simultaneously by minimizing the sum of their variational energies. On $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$2, this formulation recovers the ground and excited states on a full-dimensional CCSD(T)/aug-cc-pVTZ potential energy surface, finds preferences for the three stationary points on the potential energy surface, and shows that the dynamics of low- and high-energy states are characterized by continuous hydrogen swapping (Wang et al., 28 Sep 2025).

4. Learned canonicalization in geometric learning and neural fields

In geometric learning, canonicalization is used to replace group-equivariant architectures with a learned map to a canonical representative of each orbit. A generic model has the form

$\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$3

where $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$4 outputs a group element, $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$5 is an arbitrary backbone, and $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$6 maps outputs back to the original frame (Kaba et al., 2022). If $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$7 is equivariant in the sense that $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$8, then $\dot{x}=\nabla_x H(x)\,J, \qquad J=\begin{pmatrix}0&I\-I&0\end{pmatrix}.$9 is $\mathcal{T}:x\mapsto z=(P,Q)$0-equivariant for any $\mathcal{T}:x\mapsto z=(P,Q)$1. The same work proves a universality theorem: if $\mathcal{T}:x\mapsto z=(P,Q)$2 is continuous and the backbone universally approximates continuous functions, then the composite model universally approximates continuous equivariant functions (Kaba et al., 2022).

Adaptive canonicalization generalizes this by letting the canonicalization depend on both the input and the current network. In prior maximization, the canonical representative is chosen to maximize a class-wise neural score over the transformation group,

$\mathcal{T}:x\mapsto z=(P,Q)$3

This construction yields continuous and symmetry-preserving models and a universal approximation theorem for continuous symmetry-preserving functions (Lin et al., 29 Sep 2025). The same principle was instantiated for spectral graph neural networks, where orthogonal ambiguities inside eigenspaces are resolved by adaptive canonicalization, and for point clouds with $\mathcal{T}:x\mapsto z=(P,Q)$4 symmetry, where adaptive canonicalization outperformed data augmentation, standard canonicalization, and equivariant architectures on the reported tasks (Lin et al., 29 Sep 2025).

Pose-canonicalized neural fields provide an explicitly 3D version of this idea. Canonical Field Network (CaFi-Net) takes a pre-trained NeRF of a single object instance in arbitrary unknown 3D pose and predicts a canonical field with consistent 3D pose across the entire category (Agaram et al., 2022). Translation and scale are handled by centering and normalizing the bounding box diagonal, leaving a rotation-canonicalization problem in $\mathcal{T}:x\mapsto z=(P,Q)$5. From sampled densities $\mathcal{T}:x\mapsto z=(P,Q)$6 and coordinates $\mathcal{T}:x\mapsto z=(P,Q)$7, the network predicts canonical coordinates $\mathcal{T}:x\mapsto z=(P,Q)$8 and a set of equivariant rotation candidates $\mathcal{T}:x\mapsto z=(P,Q)$9, with the best transform selected by

$M$0

Training is self-supervised through a canonicalization loss, an orthonormality loss, and a Siamese Chamfer loss across instances of the same category. On a dataset of 1300 NeRF models across 13 object categories, the method matches or exceeds point-cloud baselines on category-level and equivariance-consistency metrics (Agaram et al., 2022).

Canonicalization also appears inside the representation itself. “Canonical Factors for Hybrid Neural Fields” introduces TILTED, which replaces fixed axis-aligned projections in factored neural fields by learnable rotations $M$1, with $M$2 parameterized by unit quaternions and optimized by Riemannian ADAM (Yi et al., 2023). The motivation is an axis-aligned bias of factored feature volumes: in a two-dimensional model problem, a rank-1 factorization exactly represents an axis-aligned square, whereas a $M$3-rotated square requires rank growing with resolution unless the coordinate system is rotated back. The paper proves that jointly learning the transformation and the factorization succeeds with drastically improved efficiency, and reports improvements across image, signed distance, and radiance-field reconstruction tasks, including capabilities comparable to baselines that are 2x larger (Yi et al., 2023).

A more operator-theoretic version is Neural Phase Correlation, which learns a basis in which the inter-observation operator decomposes into $M$4 rotations (Reynolds, 16 Jun 2026). Instead of encoding each observation independently, it represents the transformation directly via paired learned filters and bilinear interactions. In image registration this yields local orthogonal operators acting by planar rotations on learned invariant subspaces; in the one-dimensional quantum harmonic oscillator it recovers the Hermite-function eigenstates and the quantized energy levels of the unknown Hamiltonian from time-evolved wavefunction pairs alone (Reynolds, 16 Jun 2026). This suggests a strong connection between NCT and learned spectral diagonalization.

5. Control-theoretic formulations

In nonlinear sampled-data control, neural canonical control structures implement feedback linearization by learning state and input transformations into Brunovsky form (Ecker et al., 2023). Starting from

$M$5

the method learns

$M$6

with inverse maps $M$7 and $M$8, such that the latent dynamics are the fixed canonical system

$M$9

The architecture is a neural auto-encoder with four one-hidden-layer feedforward networks: state encoder, state decoder, input encoder, and input decoder. The total loss is

$MJM^T=J.$0

combining state reconstruction, input reconstruction, one-step prediction in original coordinates, and one-step prediction in canonical coordinates (Ecker et al., 2023).

Once the canonical dynamics have been identified, a linear outer-loop controller is designed in the $MJM^T=J.$1 coordinates, and the control input is mapped back through

$MJM^T=J.$2

The paper demonstrates this on an academic exactly feedback-linearizable nonlinear system and on a single-mast stacker crane. A transfer-learning setting is also reported: the outer-loop controller is trained on a nominal system $MJM^T=J.$3, applied to a target system $MJM^T=J.$4 with parameters varying within $MJM^T=J.$5, and the neural transformations are then fine-tuned on new recordings from $MJM^T=J.$6, after which the closed-loop behavior becomes close to the nominal one (Ecker et al., 2023).

6. Limitations, ambiguities, and open directions

The surveyed literature identifies several recurring limitations. In symplectic-flow NCT, the latent harmonic Hamiltonian implies an integrable target model, so generic interacting systems can only be approximated as nearly integrable (Li et al., 2019). In finite-temperature fermionic applications, the ansatz remains variationally biased, optimization is nontrivial, and finite-size effects remain important, even when entropy is directly accessible (Xie et al., 2022). In the $MJM^T=J.$7 application, the Real-NVP flow is not permutation-equivariant for the hydrogens, so the computed eigenstates do not strictly obey the full permutation symmetry (Wang et al., 28 Sep 2025).

In learned canonicalization for equivariant learning, discontinuities and non-uniqueness are intrinsic when orbits have nontrivial stabilizers or when a global canonical representative does not exist smoothly (Kaba et al., 2022). Adaptive canonicalization addresses continuity at the function level, but its prior-maximization form requires solving class-wise optimization problems at runtime, which increases inference cost (Lin et al., 29 Sep 2025). In 3D vision, CaFi-Net requires 360° NeRFs, is category-specific, and faces symmetry ambiguity for objects such as monitors (Agaram et al., 2022). In neural fields, TILTED currently uses only rotations, and its advantages can be obscured by evaluation protocols built around axis-aligned synthetic scenes (Yi et al., 2023). In control, the latent Brunovsky dynamics are exact by construction, but the paper does not provide a rigorous proof that approximate neural transformations preserve closed-loop stability in the original coordinates (Ecker et al., 2023).

A broader implication is that canonicalization is powerful precisely because it concentrates structure, but that concentration creates sensitivity to ambiguity, topology, and optimization. Future directions stated across the literature include more expressive symplectic and invertible architectures, symmetry-aware canonicalization, partial-field and occlusion-robust canonicalization for neural fields, improved architectures for bijective control transforms, and broader treatment of noncompact or more complex symmetry groups (Li et al., 2019, Agaram et al., 2022, Ecker et al., 2023, Lin et al., 29 Sep 2025). A plausible synthesis is that the long-term development of NCT will depend less on a single canonical architecture than on better ways to encode the relevant conserved structure—symplectic, unitary, equivariant, or controllable—inside learned transformations.