Hamiltonian-Informed Flows

• Hamiltonian-informed flows are dynamical systems that integrate symplectic geometry, conservation laws, and variational principles to model complex distributions.
• They are implemented in neural architectures using symplectic integration schemes like Leapfrog/Verlet, ensuring volume preservation and backward-error control.
• Applications span kinetic PDEs, generative modeling, Bayesian inference, and quantum simulation, offering efficient and geometrically faithful simulation methods.

Hamiltonian-informed flows are a class of dynamical systems, methodologies, and deep learning architectures leveraging Hamiltonian structure to model, simulate, or sample from complex distributions and partial differential equations. Core to this approach is the embedding of phase-space symplectic geometry, conservation laws, and variational principles, often on manifolds of probability densities or function spaces, to induce structure-preserving, volume-conserving (or stochastic) flows for applications in kinetic PDEs, generative modeling, Bayesian inference, optimal transport, and quantum simulation.

1. Geometric and Variational Foundations

Hamiltonian-informed flows generalize the classical Hamiltonian formalism to spaces of probability distributions, function spaces, and neural representations. The prototypical example is the $L^2$-Wasserstein manifold $P_2(\mathbb R^d)$ of probability densities, equipped with the Otto-Wasserstein metric. The tangent space at a density $\rho$ is identified with zero-mean perturbations, $$T_\rho P_2(\mathbb R^d) = \{\xi(x) : \int\xi(x)dx=0\}$$, each written as $\xi = -\nabla\cdot(\rho\nabla\phi)$ for a potential $\phi$. The associated inner product $$\langle \xi, \eta\rangle_\rho = \int\langle\nabla\phi,\nabla\psi\rangle\rho dx$$ defines a Riemannian metric yielding geodesics equivalent to 2-Wasserstein optimal transport flows (Cui et al., 2021, Chow et al., 2019).

The variational principle underlying these flows combines kinetic and potential energy on the manifold: $$\mathcal{L}(\rho, v) = \int \tfrac12|v(x)|^2\rho(x) dx - \mathcal V(\rho),$$ constrained by the continuity equation $\partial_t\rho + \nabla\cdot(\rho v)=0$. Introduction of noise at the action level (Wong–Zakai corrections) leads to stochastic Euler–Lagrange and Hamiltonian equations for density evolution—a key structural feature distinct from ad hoc SDE perturbations (Cui et al., 2021).

2. Hamiltonian-Informed Neural Architectures

Hamiltonian-informed flows are operationalized in a variety of deep learning settings, especially as volume-preserving neural normalizing flows, symplectic surrogates for time integration, and operator learning for PDEs:

• Neural Hamiltonian Flows (NHF)/Symplectic Neural Flows: These networks parameterize exact or approximate symplectic maps, sometimes via compositions of analytically symplectic layers or time-dependent Hamiltonians. For example, SympFlow builds flows as compositions of $q$-only and $P_2(\mathbb R^d)$0-only Hamiltonian layers, guaranteeing symplecticity and backward-error control (Canizares et al., 2024). Leapfrog and Verlet integration schemes are often embedded directly in the architecture for exact volume conservation (Souveton et al., 7 May 2025, Rezende et al., 2019).
• Taylor-Expansion and Remainder Surrogates: Flow-map surrogates combine truncated Taylor expansions of the solution operator with a neural remainder term, increasing accuracy over long time-intervals while preserving the geometric structure dictated by chosen integrators (Fang et al., 29 Oct 2025).
• Hamiltonian Velocity Predictors and Score Matching: Recent approaches define phase-space ODEs with parameterized force fields, learning via novel Hamiltonian score matching objectives that leverage trajectories of the induced Hamiltonian system for ultimate generative modeling and score estimation (Holderrieth et al., 2024).

3. Hamiltonian Flows in Probability and Density Manifolds

Extending classical Hamiltonian mechanics to the space of probability densities enables powerful new descriptions of sampling, fluid/kinetic equations, and generative models:

• Wasserstein Hamiltonian Flows: Euler–Lagrange variations in the Wasserstein density manifold yield Hamiltonian PDEs for $P_2(\mathbb R^d)$1 pairs, covering the Liouville, Vlasov, Schrödinger, and Schrödinger bridge equations (Chow et al., 2019). Noise introduced at the variational level leads to stochastic Hamiltonian PDEs on probability space, supporting applications such as MCMC, mean-field games, and stochastic control (Cui et al., 2021).
• Hamiltonian Normalizing and Generative Flows: Hamiltonian flows implemented by neural networks (e.g., PDE-NHF (Souveton et al., 7 May 2025)) or velocity predictors (e.g., HGF (Holderrieth et al., 2024)) target distributional evolution by enforcing exact symplecticity, invertibility, and prescribed conservation properties.
• Equivariant and Symmetry-Preserving Flows: Imposing symmetries via constraints on the Hamiltonian (e.g., by enforcing commute relations with Lie-algebra generators) yields flows that are equivariant to group actions, enhancing data efficiency and generalization, and connecting directly to the logic of disentangled representation learning (Rezende et al., 2019).

4. Applications in Physics, Machine Learning, and Inference

Hamiltonian-informed flows are deployed in domains requiring geometric fidelity, computational efficiency, and interpretability:

Domain	Problem Type	Role of Hamiltonian-Informed Flow
Kinetic/Plasma	Vlasov–Poisson PDEs	Surrogate solvers, fast sampling (Souveton et al., 7 May 2025)
Quantum Systems	Schrödinger, NLS, reacting flows	Structure-preserving simulation, quantum computing (Lu et al., 2023, Cui et al., 2021)
Generative Models	Diffusion, flow matching	Unified modeling via phase-space ODEs (Holderrieth et al., 2024, Layden et al., 9 Oct 2025)
Bayesian Inference	Posterior approximation, coreset construction	Volume-preserving posterior flows, evidence bounds (Chen et al., 2022)
Optimal Transport	Trajectory optimization, Kantorovich duality	HJB-based OT with Hamiltonian structure, non-smooth costs (Buzun et al., 23 Jul 2025)

Notable results include accurate surrogate operators for high-dimensional kinetic problems, principled generative models rivaling state-of-the-art diffusion/flow-matching baselines, exact posterior recovery in compressed Bayesian coresets, and Hamilton–Jacobi-driven OT updates in the presence of obstacles (Souveton et al., 7 May 2025, Buzun et al., 23 Jul 2025, Holderrieth et al., 2024, Chen et al., 2022).

5. Theoretical Guarantees and Advantages

Hamiltonian-informed flows enforce and leverage structural properties that are otherwise lost in black-box architectures:

• Symplecticity and Volume Preservation: Discrete updates are by construction symplectic (e.g., via Leapfrog/Verlet integration or analytic symplectic layers), guaranteeing volume preservation and bounded energy drift, critical for stable long-time evolution (Canizares et al., 2024, Souveton et al., 7 May 2025, Fang et al., 29 Oct 2025, Rezende et al., 2019).
• Backward-Error Analysis: Approximated flows converge to "shadow" Hamiltonians, with quantitative energy conservation bounds that grow only linearly or sublinearly with time, depending on the architecture (Canizares et al., 2024).
• Geometric Generalization: Embedding symmetries (translation, permutation, rotation) provides exact equivariance guarantees, improved generalization, and direct links to representation learning (Rezende et al., 2019, Souveton et al., 7 May 2025).
• Sampling Efficiency: Volume preservation eliminates the need for Jacobian determinant calculations and mitigates rare-event drift. Phase-space rotation (e.g., oscillator Hamiltonians) improves conditioning and numerical stability in generative sampling (Holderrieth et al., 2024).

6. Limitations, Extensions, and Future Directions

Despite their advantages, current Hamiltonian-informed flows face several challenges:

• Scalability to High Dimensions: The computational burden of gradient and Hessian evaluations is significant in large-$P_2(\mathbb R^d)$2 settings unless specialized architectures (e.g., Deep Sets for invariant potentials) or custom kernels are employed (Souveton et al., 7 May 2025, Canizares et al., 2024, Rezende et al., 2019).
• Training Complexity: Accurate enforcement of symplecticity, equivariance, or score-matching requires sophisticated loss functions (e.g., residuals to integrators, Poisson bracket penalties), second-order derivatives, and careful regularization (Fang et al., 29 Oct 2025, Rezende et al., 2019).
• Non-Smooth Costs: Applications to non-differentiable OT or obstacle-laden domains necessitate additional regularization (e.g., angular-acceleration penalties) for stable learning (Buzun et al., 23 Jul 2025).
• Extensions: Directions include port-Hamiltonian flows for open systems, adaptation to high-dimensional field theories, quantum simulation (via explicit continuity Hamiltonians), and parareal or local-step operator learning for stiff, multi-scale systems (Fang et al., 29 Oct 2025, Layden et al., 9 Oct 2025, Lu et al., 2023).

Ongoing research continues to extend Hamiltonian-informed flows to broader problem classes, integrating further geometric and physical structure to unlock advances in simulation, inference, and generative modeling.