---
title: Buckingham–Symplectic Networks (BuSyNet)
url: https://www.emergentmind.com/topics/buckingham-symplectic-networks-busynet
type: topic
---

# Buckingham–Symplectic Networks (BuSyNet)

Searching arXiv for the BuSyNet paper and closely related Hamiltonian-learning work.
Buckingham–Symplectic Networks (BuSyNet) are a deep learning architecture for Hamiltonian discovery that combines two inductive biases usually treated separately: dimensional consistency and symplectic action–angle structure. In BuSyNet, a strictly symplectic encoder maps observed phase-space trajectories $(q,p)$ to latent action–angle variables $(I,\theta)$, and a Buckingham-$\pi$–inspired symbolic head learns a Hamiltonian $H(I,m)$ with units of energy by constraining monomial exponents through $D_{\text{in}}\Psi = D_{\text{out}}$ [2604.00576]. The method is formulated for integrable Hamiltonian systems, where canonical transformations to action–angle coordinates reduce the dynamics to $\dot I = 0$ and $\dot \theta = \partial H/\partial I$, and it is evaluated on the harmonic oscillator and the Kepler two-body problem in two and three dimensions, where it recovers concise closed-form Hamiltonians and yields lower long-horizon state error and smaller energy variance than NN, HNN, and SympNet baselines [2604.00576].

## 1. Concept and scope

BuSyNet is motivated by the observation that the Hamiltonian is not merely a scalar predictor but a physical quantity with units of energy and a generator of symplectic dynamics. The framework therefore treats dimensional homogeneity and canonical structure as coupled constraints rather than independent regularizers. The first constraint is that the Hamiltonian must carry units $M L^2 T^{-2}$; the second is that, for integrable systems, there exists a symplectic change of variables to action–angle coordinates $(I,\theta)$ in which trajectories become rigid rotations on invariant tori [2604.00576].

The architecture is described as learning a symplectic encoder $S_\phi$ from observed phase-space trajectories to latent action–angle variables and a symbolic Hamiltonian head that combines the learned actions with known physical parameters. This design is intended to identify Hamiltonians precisely rather than only up to an additive constant, because the training procedure uses precomputed action integrals from the data as anchors for the latent variables [2604.00576].

Within the cited work, BuSyNet targets integrable Hamiltonian systems with full phase-space observations and known units for the relevant variables and parameters. The harmonic oscillator and the Kepler problem serve as canonical examples because both admit action–angle representations and closed-form Hamiltonians in those coordinates [2604.00576].

## 2. Mathematical framework

BuSyNet is built on the standard Hamiltonian formulation. For $z=(q,p)\in\mathbb{R}^{2n}$ and scalar Hamiltonian $H(z)$, Hamilton’s equations are
$$
\dot q=\frac{\partial H}{\partial p},\qquad \dot p=-\frac{\partial H}{\partial q},
$$
or, equivalently,
$$
\dot z = J \nabla_z H(z),
$$
with
$$
J=\begin{bmatrix} 0 & I \\ -I & 0 \end{bmatrix}\in\mathbb{R}^{2n\times 2n},
$$
where $J^T=-J=J^{-1}$ [2604.00576].

The relevant geometric structure is symplecticity. A differentiable map $S:\mathbb{R}^{2n}\to\mathbb{R}^{2n}$ is symplectic if its Jacobian satisfies
$$
S'(z)^T J S'(z)=J
$$
for all $z$. Canonical transformations preserve the symplectic form, and compositions of symplectic maps remain symplectic [2604.00576]. BuSyNet uses this property to ensure that the learned encoder is not merely approximately structure-preserving but symplectic by construction.

The latent representation is based on action–angle variables. For integrable Hamiltonians there exists a canonical transformation to $(I,\theta)\in\mathbb{R}^n\times\mathbb{T}^n$ such that $H=H(I)$ and
$$
\dot \theta = \omega(I)=\frac{\partial H}{\partial I},\qquad \dot I = 0.
$$
In these coordinates, motion lies on invariant $n$-tori and becomes a rigid rotation with constant frequency vector $\omega(I)$ [2604.00576]. This latent-space form is central to BuSyNet: the learned dynamics are constrained so that actions remain constant and angles evolve linearly.

## 3. Dimensional analysis and symbolic Hamiltonian construction

The unit-aware component of BuSyNet is a Buckingham-$\pi$–inspired symbolic head based on dimensional matrices and monomial parameterization. Each input variable $v_i$ is assigned a base-units vector $\Omega(v_i)\in\mathbb{Z}^r$, and these are collected into
$$
D_{\text{in}}=[\Omega(v_1)\ \cdots\ \Omega(v_d)]\in\mathbb{Z}^{r\times d}.
$$
A general monomial
$$
\Pi=\prod_{i=1}^d v_i^{\alpha_i}
$$
has units vector $D_{\text{in}}\alpha$ [2604.00576]. Classical Buckingham analysis obtains dimensionless groups from the nullspace equation $D_{\text{in}}\alpha=0$; BuSyNet extends this idea to produce outputs with prescribed units.

In the symbolic head, for a mini-batch $V\in\mathbb{R}^{m\times d}$,
$$
\Pi = \exp(\log |V| \Psi),
$$
where $\Psi\in\mathbb{R}^{d\times d'}$ contains trainable exponents. The unit constraint is imposed softly by
$$
D_{\text{in}}\Psi \approx D_{\text{out}},
$$
with loss
$$
L_{\text{dim}}=\|D_{\text{in}}\Psi-D_{\text{out}}\|_F^2.
$$
For Hamiltonian discovery, $D_{\text{out}}$ is the energy units vector $M L^2 T^{-2}$ [2604.00576].

The 1D oscillator illustrates the constraint. A candidate monomial $\Pi=I^\alpha m^\beta k^\gamma$ must satisfy
$$
\alpha[I]+\beta[m]+\gamma[k]=[\text{energy}],
$$
that is,
$$
\alpha[M L^2 T^{-1}] + \beta[M] + \gamma[M T^{-2}] = [M L^2 T^{-2}].
$$
The symbolic head learns $(\alpha,\beta,\gamma)$ subject to the dimensional constraint [2604.00576]. This mechanism narrows the admissible hypothesis class to forms with correct physical units, which the source identifies as important for extrapolation across parameter scales and interpretability.

The symbolic discovery proceeds in two stages. First, the head learns a latent monomial of the form
$$
\hat H = \Big[\sum_j I_j\Big]^\alpha \prod_i m_i^{\beta_i},
$$
with exponents trained under the unit constraint. Second, after $\alpha$ is identified and rounded to $10^{-3}$, the model replaces the intermediate form with an action-only expression. If $\alpha>0$,
$$
\hat H(I)=\sum_{k\in K_\alpha} a_k I_1^{k_1}\cdots I_n^{k_n},\qquad \sum_j k_j=\alpha.
$$
If $\alpha<0$,
$$
\frac{1}{\hat H(I)}=\sum_{k\in K_{|\alpha|}} a_k I_1^{k_1}\cdots I_n^{k_n}.
$$
For one-degree-of-freedom systems such as simple harmonic motion, the BuckiNet monomial itself already matches the true form and is used directly [2604.00576].

## 4. Architecture and training procedure

The encoder in BuSyNet is a G-SympNet layer stack $S_\phi:(q,p)\mapsto(\hat I,\hat \theta)$ that is symplectic by construction. The building blocks are canonical shear maps:
$$
\text{B-block}: (q,p)\mapsto(q,p+a(q)),
$$
$$
\text{A-block}: (q,p)\mapsto(q+b(p),p),
$$
with learnable functions $a$ and $b$ implemented as small neural networks. Alternating compositions such as $\cdots B\circ A\circ B\circ A\cdots$ remain symplectic, and each block is exactly invertible:
$$
A^{-1}:(q',p')\mapsto(q'-b(p'),p'),\qquad B^{-1}:(q',p')\mapsto(q',p'-a(q')).
$$
The reported implementation uses a 4-layer G-SympNet with 32 neurons per hidden layer [2604.00576].

A distinctive component is the use of precomputed action integrals to anchor the latent actions. For each action,
$$
I_i=\frac{1}{2\pi}\oint p_i\,dq_i = \frac{1}{2\pi}\int_0^{T_i} p_i \frac{dq_i}{dt}\,dt,
$$
where the period $T_i$ is estimated by FFT and the integral is evaluated numerically with Simpson’s rule [2604.00576]. These numerically estimated actions are then matched to the learned $\hat I$ through a reconstruction term, which the source presents as crucial for identifying $H$ precisely.

Training uses three principal losses:
$$
L_I=\|I-\hat I\|_2^2,
$$
$$
L_{\dot\theta}=\left\|\frac{\partial \hat H}{\partial \hat I}-\frac{d\hat\theta}{dt}\right\|_2^2,
$$
$$
L_{\dot I}=\left\|\frac{\partial \hat H}{\partial \hat\theta}+\frac{d\hat I}{dt}\right\|_2^2.
$$
Since $\hat H$ depends only on $\hat I$, one has $\partial \hat H/\partial \hat\theta = 0$, so the ideal latent dynamics satisfy $\dot{\hat I}=0$ [2604.00576]. Time derivatives are computed by the chain rule,
$$
\frac{d\hat I}{dt}=\frac{\partial \hat I}{\partial q}\dot q + \frac{\partial \hat I}{\partial p}\dot p,\qquad
\frac{d\hat\theta}{dt}=\frac{\partial \hat\theta}{\partial q}\dot q + \frac{\partial \hat\theta}{\partial p}\dot p,
$$
using automatic differentiation through $S_\phi$, with $\dot q$ and $\dot p$ obtained by numerically differentiating the trajectories [2604.00576].

The pipeline is four-stage. First, trajectories $(q(t),p(t))$ and known physical parameters are collected, and $D_{\text{in}}$ and $D_{\text{out}}$ are formed. Second, actions are precomputed by FFT-based period estimation and Simpson’s-rule integration. Third, the symbolic head is pretrained on $(I,m)$ by minimizing $L_{\text{dim}}$ with L-BFGS for 100 epochs and rounding learned exponents to the nearest $10^{-3}$. Fourth, the G-SympNet encoder and Hamiltonian coefficients are optimized with Adam for 1000 epochs at learning rate $10^{-3}$ using
$$
L(\phi,\{a_k\})=\lambda_I L_I+\lambda_{\dot\theta}L_{\dot\theta}+\lambda_{\dot I}L_{\dot I},
$$
with $\lambda_I=\lambda_{\dot\theta}=\lambda_{\dot I}=1$ in all experiments [2604.00576].

At inference time, latent-space rollout proceeds by encoding $z_k=(q_k,p_k)$ to $(\hat I_k,\hat\theta_k)$, computing $\hat\omega = \partial \hat H/\partial \hat I$, updating the angles with an Euler step,
$$
\hat\theta_{k+1}=\hat\theta_k+\hat\omega\,\Delta t,
$$
keeping actions fixed,
$$
\hat I_{k+1}=\hat I_k,
$$
and decoding via $S_\phi^{-1}$ to obtain $(q_{k+1},p_{k+1})$ [2604.00576]. Because the symplectic blocks are exactly invertible, decoding is performed exactly by reversing the G-SympNet blocks.

## 5. Empirical results and recovered Hamiltonians

The reported experiments cover the harmonic oscillator and the Kepler two-body problem in 2D and 3D. For the harmonic oscillator,
$$
H(q,p)=\frac{1}{2}kq^2+\frac{p^2}{2m},
$$
with action–angle form
$$
H(I)=\omega I,\qquad \omega=\sqrt{\frac{k}{m}}.
$$
The setup uses $m=1$, $k=1$, $\Delta t=0.001$, and $t\in[0,20\pi]$, with 1000 training points from $t\in[0,7]$ and testing on the remainder [2604.00576].

For the Kepler problem, the Hamiltonian is given in spherical coordinates as
$$
H(q,p)=\frac{1}{2m}\left(p_r^2+\frac{p_\theta^2}{r^2}+\frac{p_\phi^2}{r^2\sin^2\theta}\right)-\frac{k}{r},
$$
with $k=GMm$. In action–angle variables the 3D Hamiltonian is
$$
H(I)=-\frac{m k^2}{2(I_r+I_\theta+I_\phi)^2},
$$
and the 2D version omits $I_\phi$ [2604.00576]. The 3D setup uses $m=1$, $k=1$, $\Delta t=0.001$, $t\in[0,20]$, and initial conditions $r(0)=0.25$, $\theta(0)=\pi/2$, $\phi(0)=0$, $p_r(0)=0$, $p_\theta(0)=0$, $p_\phi(0)=0.5$, corresponding to a circular orbit of radius $0.25$; training uses $t\in[0,10]$, testing uses $[0,20]$, and the period is $T=\pi/4$. The 2D setup uses $(r(0),\theta(0),p_r(0),p_\theta(0))=(0.25,0,0,0.5)$ with the same train-test split and $T=\pi/4$ [2604.00576].

The discovered symbolic Hamiltonians are reported as follows.

| System | Recovered Hamiltonian |
|---|---|
| Simple harmonic oscillator | $\hat H_{\text{SHM}}(I)=I^{1.0000}k^{0.5000}m^{-0.5000}=\sqrt{k/m}\,I$ |
| Kepler 2D | $\hat H_{\text{Kepler2D}}(I)=-1/[2.000\,(I_r+I_\theta)^{2.000}]$ |
| Kepler 3D | $\hat H_{\text{Kepler3D}}(I)=-1/[2.000\,(I_r+I_\theta+I_\phi)^{2.000}]$ |

These expressions are stated to be consistent with the analytic forms for the chosen parameter settings, and the exponents are reported to match analytic values to rounding precision $10^{-3}$ [2604.00576]. For the oscillator, the recovered exponents $0.5$ on $k$ and $-0.5$ on $m$ are explicitly identified as physically meaningful in the source.

The recovered frequencies are also listed. For SHM, $\hat\omega = 1.0 = \omega_{\text{true}}$ for $m=k=1$. For the circular-orbit Kepler cases, the reported values are $\hat\omega_r=0.0$ and $\hat\omega_\theta=8.0$ in 2D, and $\hat\omega_r=0.0$, $\hat\omega_\theta=0.0$, and $\hat\omega_\phi=8.0$ in 3D [2604.00576]. This is consistent with the latent-space interpretation in which constant actions and linearly advancing angles directly expose the frequencies $\omega=\partial \hat H/\partial I$.

## 6. Accuracy, stability, interpretability, and limitations

The reported quantitative evaluation emphasizes long-horizon accuracy and stability. For the harmonic oscillator with period $T=2\pi$, the mean squared error on $[0,2T]$ is reported as NN $1.73\times 10^{-4}$, HNN $2.63\times 10^{-4}$, SympNet $1.74\times 10^{-5}$, and BuSyNet $7.89\times 10^{-6}$; on $[2T,20T]$ it is NN $3.12\times 10^{-3}$, HNN $3.75\times 10^{-1}$, SympNet $3.95\times 10^{-4}$, and BuSyNet $2.77\times 10^{-4}$ [2604.00576]. The energy variance $\operatorname{Var}[\hat H(t)]$ is NN $3.17\times 10^{-3}$, HNN $3.10\times 10^{-3}$, SympNet $4.14\times 10^{-6}$, and BuSyNet $9.88\times 10^{-12}$, which the source characterizes as best by 5–6 orders of magnitude [2604.00576].

For Kepler 2D over 25 periods, the state MSE is NN $5.20\times 10^{-3}$, HNN $2.46\times 10^{+1}$, SympNet $1.07\times 10^{-4}$, and BuSyNet $5.25\times 10^{-5}$; the energy variance is NN $2.74\times 10^{-5}$, HNN $6.49\times 10^{+2}$, SympNet $9.92\times 10^{-5}$, and BuSyNet $1.17\times 10^{-11}$ [2604.00576]. For Kepler 3D over 25 periods, the MSE is NN $1.38\times 10^{0}$, HNN $4.53\times 10^{+1}$, SympNet $4.10\times 10^{-5}$, and BuSyNet $3.51\times 10^{-5}$; the energy variance is NN $9.05\times 10^{-6}$, HNN $8.30\times 10^{+3}$, SympNet $3.09\times 10^{-6}$, and BuSyNet $2.32\times 10^{-11}$ [2604.00576].

The source does not report explicit ablations, but it states that comparisons to HNN and SympNet act as implicit ablations. Relative to HNN, adding the action–angle latent space plus units is said to reduce drift and phase error substantially. Relative to SympNet, adding the unit-aware symbolic Hamiltonian yields more accurate long-term prediction and orders-of-magnitude smaller energy variance [2604.00576]. Symbolic sparsity penalties and alternative latent dimensions are explicitly noted as unexplored.

Interpretability in BuSyNet has two components. First, the symbolic Hamiltonian makes physical content explicit. The oscillator example,
$$
\hat H=\sqrt{\frac{k}{m}}\,I,
$$
recovers the textbook action–angle form $H=\omega I$ and corresponds to the standard Hamiltonian $H=p^2/(2m)+(k/2)q^2$ in canonical variables [2604.00576]. Second, the latent variables themselves are interpretable: actions are constant, angles advance linearly, and frequencies are directly available from $\omega=\partial \hat H/\partial I$ [2604.00576].

The limitations stated in the source are substantial and delineate the scope of the method. BuSyNet assumes integrability and global action–angle coordinates; non-integrable or chaotic systems are not targeted. It assumes full $(q,p)$ observations and known base-unit assignments for inputs and outputs. Parameters entering the Hamiltonian must be present among the inputs to be discoverable. Numerical derivatives and action-integral estimates may be noise-sensitive, so smoothing or robust differentiation may be necessary. Increasing degrees of freedom enlarges the multi-index set $K_\alpha$ and raises SympNet capacity requirements, although end-to-end training is described as feasible for low-to-moderate dimensions [2604.00576]. A plausible implication is that the method’s strongest regime is low-dimensional, cleanly observed, integrable mechanics with explicit physical parameters.

## 7. Relation to prior work and terminological clarifications

The BuSyNet paper situates itself against two strands of prior Hamiltonian learning. Hamiltonian Neural Networks learn a Hamiltonian and derive dynamics via Hamilton’s equations, but they do not enforce units and do not reduce dynamics to action–angle coordinates. Symplectic networks such as SympNet enforce symplecticity, but they are typically trained as time-steppers and do not combine a learned canonical transformation with a dimensionally constrained symbolic Hamiltonian [2604.00576]. Within that comparison, BuSyNet is presented as pairing a symplectic action–angle encoder with a dimensionally aware symbolic head that outputs a closed-form Hamiltonian in correct energy units.

A separate terminological clarification concerns possible confusion with an unrelated use of “Buckingham–Symplectic Networks.” The paper “A ‘network of networks’ (from history to algebra)” works with complex-valued flows and tensions, lattices in $\mathbb{C}$, Siegel space, and symplectic group actions, and it explicitly states that it does not discuss Buckingham $\Pi$ dimensional analysis [2301.11271]. Its “connection to Buckingham–Symplectic Networks” is framed as a path for integrating Buckingham $\Pi$ into a lattice-and-Siegel framework rather than as the BuSyNet architecture for Hamiltonian discovery [2301.11271]. This suggests that the term may be used in two distinct contexts: one in Hamiltonian system identification from trajectories [2604.00576], and another in an algebraic-geometric construction over network lattices and Siegel modular forms [2301.11271].

In the Hamiltonian-learning sense, BuSyNet denotes the architecture that integrates a symplectic encoder, precomputed action integrals, and a Buckingham-$\pi$–inspired symbolic Hamiltonian head. Its defining claim is that jointly enforcing symplectic structure and dimensional consistency yields stable, accurate, and interpretable Hamiltonian discovery for integrable systems [2604.00576].

Source: https://www.emergentmind.com/topics/buckingham-symplectic-networks-busynet