Symplectic Convolutional Neural Network

• Symplectic CNN is a deep learning architecture that enforces the canonical symplectic form in convolutional layers, preserving phase-space volume in Hamiltonian systems.
• It employs a matrix formulation of convolution and symplectic pooling techniques to embed geometric constraints, ensuring invertibility and robust gradient flow.
• Applications include structure-preserving reduced-order modeling for Hamiltonian PDEs like the wave, nonlinear Schrödinger, and sine-Gordon equations with improved accuracy over standard methods.

Searching arXiv for the specified papers and closely related work on symplectic neural networks and symplectic CNNs. A symplectic convolutional neural network (CNN) is a convolutional architecture whose layers are parameterized so that the resulting map is symplectic, meaning that its Jacobian preserves the canonical symplectic form associated with Hamiltonian phase space. In the current literature, the most explicit realization is a symplectic convolutional autoencoder, denoted SympCAE, which combines symplectic neural networks, proper symplectic decomposition, and tensor techniques for structure-preserving dimensionality reduction of Hamiltonian systems; its stated applications are the wave equation, the nonlinear Schrödinger equation, and the sine-Gordon equation (Yıldız et al., 27 Aug 2025).

1. Emergence from symplectic deep learning

Symplectic CNNs emerged from a broader program in geometric deep learning that seeks architectures with exact structure preservation rather than approximate enforcement through regularization. Several closely related constructions established the underlying ingredients before a convolutional realization was given. One line of work proposed deep neural networks whose outputs are invertible symplectomorphisms and whose symplecticity is enforced structurally, with no need for penalty terms in the loss (He et al., 2024). Another line constructed symplectic networks from explicit higher-order numerical methods, emphasizing non-vanishing gradients and universal approximation after feature augmentation (Maslovskaya et al., 2024). A further framework developed SympNets as compositions of symplectic flow maps, proving universality in the space of Hamiltonian diffeomorphisms and exact representation results for linear symplectic systems (Tapley, 2024). Earlier work on neural canonical transformation formulated canonical maps as symplectic flows and explicitly discussed the extension of such ideas to convolutional architectures (Li et al., 2019).

Within this trajectory, the symplectic CNN is not merely an invertible CNN. Its defining objective is to make the convolutional action itself compatible with Hamiltonian geometry. This suggests that the convolutional setting is being treated as a special case of a more general principle: build each learnable block as a symplectic map and obtain a globally symplectic network by composition.

2. Symplectic structure and the convolutional setting

The mathematical condition used in the symplectic CNN literature is the standard symplectic Jacobian identity. For a nonlinear map $\psi : \mathbb{R}^{2k} \to \mathbb{R}^{2n}$, symplecticity is expressed as

$$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$

with

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$

In the linear case, the condition reduces to

$A^\top J_{2n} A = J_{2k}.$

These formulas are used to define when a convolutional layer, a pooling layer, or an encoder-decoder block is symplectic (Yıldız et al., 27 Aug 2025).

The same literature emphasizes that symplecticity is stronger than ordinary invertibility and stronger than generic volume-preserving behavior. Real NVP, which motivates some symplectic constructions, provides explicit invertibility but is not generally volume-preserving, let alone symplectic (He et al., 2024). Conversely, canonical transformations in symplectic-flow models satisfy the symplectic condition

$M J M^\top = J,$

and therefore preserve Hamiltonian evolution and phase-space volume (Li et al., 2019). A common misconception is to treat any ODE-inspired or invertible architecture as automatically symplectic. The cited works explicitly distinguish these notions.

For CNNs, the key technical move is to rewrite convolution in matrix form. The symplectic CNN paper states that any convolution operation can be written in an equivalent matrix-vector form,

$y = T x,$

where $T$ is a block matrix assembling channel-wise convolutional weights; in one dimension this gives a block Toeplitz representation, and in two dimensions a block-block Toeplitz matrix with block Toeplitz substructure (Yıldız et al., 27 Aug 2025). This equivalence allows the symplectic constraint to be imposed on the convolution layer through matrix parameterization.

3. Architectural construction of the symplectic CNN

The concrete architecture presented in the literature is the symplectic convolutional autoencoder, SympCAE. Its encoder consists of symplectic convolutional and activation layers followed by symplectic pooling and PSD-like projection layers; its decoder uses transposed PSD, unpooling, and symplectic deconvolutions with activation. The paper states that the entire pipeline is symplectic by design (Yıldız et al., 27 Aug 2025).

The defining components can be summarized as follows:

Component	Mathematical form	Role
Convolution layer	$y = T x$	Matrix realization of convolution
Symplectic condition	$A^\top J_{2n} A = J_{2k}$	Enforces linear symplecticity
Symplectic pooling	$P(x)=\begin{bmatrix}\Phi(x_1)&0\0&\Phi(x_1)\end{bmatrix}$	Symplectic downsampling for 2-channel input
Symplectic unpooling	$$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$0	Decoder inverse of pooling
PSD-like layer	$$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$1, with $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$2	Symplectic bottleneck projection

For two input and two output channels, corresponding to $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$3 and $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$4 in Hamiltonian systems, one admissible symplectic linear map is an upper-triangular block matrix of the form

$$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$5

where $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$6 is symmetric and Toeplitz-structured for convolution. For channel-upsizing, the paper gives a symplectic “copy” layer

$$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$7

with $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$8 for symplecticity (Yıldız et al., 27 Aug 2025). More generally, the convolutional layers are built from block matrices with identity and symmetric Toeplitz blocks, together with channel-number constraints that maintain symplecticity.

A distinct contribution is the symplectic pooling layer. Given a standard max-pooling operation with selection or binary mask matrix $$(D \psi_x)^\top J_{2n} D \psi_x = J_{2k}, \qquad \forall x \in \mathbb{R}^{2k},$$9, the symplectic pooling map for a two-channel input $$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$0 is defined as

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$1

or with $$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$2 in the lower variant. The same source states that $$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$3, and therefore the combined pooling map is symplectic (Yıldız et al., 27 Aug 2025). This addresses a point that had remained largely open in earlier discussions: pooling is incorporated without abandoning the structure-preserving requirement.

4. Relation to SympNets, symplectomorphisms, and geometric integrators

The symplectic CNN is best understood as a convolutional specialization of the broader SympNet program. In one influential construction, a symplectic neural network is any finite composition of three explicit symplectic building blocks: q-shearing,

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$4

p-shearing,

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$5

and stretching,

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$6

Each block is explicitly invertible, and symplecticity is guaranteed by construction, so no extra penalty or loss term is needed to enforce it (He et al., 2024).

A second formulation derives symplectic networks from geometric integration. There, each layer is a full step of an explicit symplectic partitioned Runge-Kutta method, with coefficients satisfying

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$7

for symplecticity. These networks are called SPRK Nets and are presented as higher-order generalizations of a symplectic-Euler Hamiltonian network (Maslovskaya et al., 2024). A third formulation represents a SympNet as a composition of exact symplectic flow maps generated by basis Hamiltonians, with universality in the class of Hamiltonian diffeomorphisms and exact representation of linear symplectic transformations by P-SympNets (Tapley, 2024).

The symplectic CNN paper states that its symplectic convolutional layers are parameterized using a scheme inspired by LA-SympNets and that the network alternates symplectic linear modules with symplectic activation modules (Yıldız et al., 27 Aug 2025). This indicates that the convolutional model does not introduce a separate notion of symplecticity; rather, it transfers established symplectic-network design principles into Toeplitz-structured, channel-coupled CNN layers.

5. Theoretical properties

The central theoretical property is exact symplecticity by construction. Because the set of symplectic maps is closed under composition, a network assembled from symplectic convolution, activation, pooling, and projection blocks remains symplectic globally (Yıldız et al., 27 Aug 2025).

The broader symplectic-network literature associates this design with several additional guarantees. Symplectic networks were shown to have a non-vanishing gradient property, and one formulation states that for a fully symplectic layer the gradient norm across layers satisfies

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$8

so backpropagation through arbitrary depth does not exponentially contract nor blow up gradients (Maslovskaya et al., 2024). A closely related statement appears in the SympNet framework, where products of symplectic Jacobians have lower-bounded norm, again implying that gradients do not vanish as depth increases (Tapley, 2024).

Another recurrent property is explicit invertibility. SymplectoNet layers admit inverses by sign reversal of the generating scalar functions, for example

$$J_{2n} = \begin{bmatrix} 0 & I_n \ -I_n & 0 \end{bmatrix}.$$9

and canonical-transformation models use invertible coordinate maps together with momentum updates derived from Jacobians (He et al., 2024, Li et al., 2019). In the CNN setting, invertibility is expressed operationally through decoder design, transposed PSD, and symplectic unpooling (Yıldız et al., 27 Aug 2025).

The literature also distinguishes symplecticity from mere numerical sophistication. High-order methods can improve CNN accuracy, but the cited review explicitly states that such methods do not generally guarantee non-vanishing gradients or preservation of geometric structure (Maslovskaya et al., 2024). This is a substantive distinction rather than a terminological one.

6. Applications, empirical results, and scope

The reported applications of the symplectic CNN are all Hamiltonian PDEs. For the wave equation,

$A^\top J_{2n} A = J_{2k}.$0

with canonical variables $A^\top J_{2n} A = J_{2k}.$1 and $A^\top J_{2n} A = J_{2k}.$2, the discretized evolution is written as

$A^\top J_{2n} A = J_{2k}.$3

which the paper describes as a composition of symplectic maps mimicked by the SympCAE structure (Yıldız et al., 27 Aug 2025). For the nonlinear Schrödinger equation,

$A^\top J_{2n} A = J_{2k}.$4

with $A^\top J_{2n} A = J_{2k}.$5, and for the two-dimensional sine-Gordon equation,

$A^\top J_{2n} A = J_{2k}.$6

the same architecture is applied after structure-preserving discretization (Yıldız et al., 27 Aug 2025).

The numerical comparisons reported for SympCAE versus proper symplectic decomposition are specific and favorable at low latent dimension:

Problem	Latent dimension	SympCAE	PSD
Wave equation reconstruction error	$A^\top J_{2n} A = J_{2k}.$7	$A^\top J_{2n} A = J_{2k}.$8	$A^\top J_{2n} A = J_{2k}.$9
NLS reconstruction error	$M J M^\top = J,$0	$M J M^\top = J,$1	$M J M^\top = J,$2
Sine-Gordon reconstruction error	$M J M^\top = J,$3	$M J M^\top = J,$4	$M J M^\top = J,$5

The same source states that, after encoding, a SympNet is trained on latent trajectories and yields accurate long-term prediction for the wave equation, and that latent ODE integration with SympNet plus decoder reconstructs the NLS solution accurately (Yıldız et al., 27 Aug 2025). The abstract summarizes the empirical conclusion succinctly: the numerical results indicate that the symplectic CNN outperforms the linear symplectic autoencoder obtained via proper symplectic decomposition.

The current scope of the literature remains specific. One 2024 study on symplectic methods in deep learning explicitly states that it does not present implementation of SPRK Nets as convolutional neural networks directly, although the techniques are compatible with CNNs in principle (Maslovskaya et al., 2024). Earlier work on neural canonical transformations also frames convolutional extensions as a natural direction rather than as a completed architecture (Li et al., 2019). Accordingly, the most concrete notion of a “symplectic CNN” in the cited record is the SympCAE family for Hamiltonian model reduction and latent dynamics. A plausible implication is that future variants may broaden this class beyond autoencoding and reduced-order modeling, but that expansion lies beyond the claims currently documented in the cited papers.