Hamiltonian Dynamics with Control

Updated 16 May 2026

Hamiltonian dynamics with control is a framework that augments classical Hamiltonian systems with control inputs to shape energy landscapes and achieve robust, stable trajectories.
It leverages energy shaping and damping injection methods to modify the system’s energetic structure and ensure convergence and performance under disturbances.
Structure-preserving learning approaches, including physics-informed neural networks, enable efficient and interpretable control for diverse applications from robotics to quantum systems.

Hamiltonian dynamics with control describes the synthesis, analysis, and implementation of feedback and open-loop control policies for physical and engineered systems whose evolution is governed fundamentally or modelled efficiently by a Hamiltonian structure. The field spans classical, quantum, and distributed systems, and is characterized by approaches that explicitly exploit the geometric and energetic structure of the phase space to design trajectories, guarantees of stability, and robustness to disturbances or model uncertainty.

1. Mathematical Structure of Controlled Hamiltonian Systems

Hamiltonian dynamics arises from the specification of a total energy function $H(q,p)$ defined over a phase space, usually the cotangent bundle $T^*Q$ of a configuration manifold $Q$ , where $q$ are generalized coordinates and $p$ are conjugate momenta. In the uncontrolled setting, dynamics are generated by Hamilton’s equations: $\dot{q} = \frac{\partial H}{\partial p}, \qquad \dot{p} = -\frac{\partial H}{\partial q}.$ To incorporate control, additional terms are introduced to inject or extract energy or modify the evolution. The standard, control-augmented form is: $\begin{bmatrix} \dot{q}\ \dot{p} \end{bmatrix} = \begin{bmatrix} \frac{\partial H}{\partial p} \[4pt] -\frac{\partial H}{\partial q} \end{bmatrix} + \begin{bmatrix} 0 \ g(q)u \end{bmatrix},$ where $u$ are the control inputs and $g(q)$ is a map encoding how inputs couple into the momentum dynamics. This approach generalizes naturally to port-Hamiltonian systems, which embed energy transfer, interconnection, and dissipation: $\dot{x} = [J(x) - R(x)] \nabla_x H(x) + G(x)u,$ where $T^*Q$ 0 is a skew-symmetric interconnection structure, $T^*Q$ 1 is a dissipation matrix, and $T^*Q$ 2 is the input map (Duong et al., 2021, Zhong et al., 2020).

Hamiltonian control on manifolds and Lie groups such as $T^*Q$ 3 and $T^*Q$ 4 appears in robotics and rigid body dynamics, with phase space coordinates describing pose and velocity, and generalized momenta linked via configuration-dependent mass-inertia matrices (Duong et al., 2024, Duong et al., 2021).

In quantum systems and infinite-dimensional settings (PDEs), the Hamiltonian takes operator-valued or functional forms, with controls appearing either additively (e.g., $T^*Q$ 5 in quantum systems) or via boundary ports in distributed port-Hamiltonian systems (Bookatz et al., 2013, Beckers et al., 5 Apr 2026).

2. Control Synthesis: Energy Shaping, Damping Injection, and Adaptive Methods

Control design for Hamiltonian systems is based on passivity and energy principles. The primary methodology is energy-shaping and damping injection (IDA-PBC), which modifies the system’s energetic landscape and injects artificial dissipation for stabilization and trajectory tracking.

Energy shaping constructs a “desired” Hamiltonian $T^*Q$ 6 that has its minimum at the target equilibrium or follows a reference trajectory, augmenting the open-loop energy by artificial potentials and kinetic terms:

$T^*Q$ 7

where $T^*Q$ 8 is designed for regulation or tracking (Duong et al., 2021, Duong et al., 2024).

Damping injection introduces a feedback term proportional to generalized velocities, ensuring asymptotic convergence:

$T^*Q$ 9

where $Q$ 0 is a damping gain (Duong et al., 2021).

Total control law becomes $Q$ 1, with $Q$ 2 determined by matching the target port-Hamiltonian structure (Duong et al., 2021).

For systems with unknown or time-varying disturbances, adaptive Hamiltonian control leverages learned disturbance features. Offline, neural ODE architectures are trained to extract a basis for disturbances from data; online, adaptive laws compensate using error-driven updates of disturbance weights, all rigorously embedded within the Hamiltonian framework (Duong et al., 2021).

Stability proofs are based on Lyapunov arguments: the shaped Hamiltonian serves as a candidate, and the closed-loop energy is non-increasing, with strict decrease enforced through damping. Asymptotic and exponential stability of the regulated or tracked set is achieved under mild regularity and matching conditions (Duong et al., 2021, Duong et al., 2021, Duong et al., 2024).

3. Structure-Preserving Learning of Hamiltonian Dynamics

Data-driven approaches utilize physics-informed neural network architectures to enforce Hamiltonian structure in learned models:

Symplectic ODE-nets and port-Hamiltonian neural ODEs parameterize the mass matrix, potential, input maps, and dissipation using neural networks, with physical constraints (e.g., positive-definiteness via Cholesky factors) (Zhong et al., 2019, Zhong et al., 2020, Duong et al., 2021).
Manifold constraints (e.g., phase space on $Q$ 3, $Q$ 4) are enforced explicitly, guaranteeing that learned trajectories remain on the correct configuration manifold and conserve energy in the absence of input or dissipation (Duong et al., 2021, Duong et al., 2024).
Observation-space or point-cloud learning methods handle cases where states are inaccessible, directly minimizing losses in the raw observation domain while embedding Hamiltonian structure for efficiency and generalization (Altawaitan et al., 2023).
Learning dissipation is tractable, allowing for the explicit modeling and control of friction or drag as neural-network components in the dissipation matrix (Zhong et al., 2020, Altawaitan et al., 2023).

These architectures are statistically and computationally efficient, show improved generalization over black-box models (especially in long-horizon rollouts), and yield interpretable components (mass, energy, input gains, damping) for model-based controller synthesis (Duong et al., 2021, Zhong et al., 2019, Zhong et al., 2020).

4. Applications: Robotics, Quantum Systems, and Distributed Control

Robotics and Mechanical Systems:

Rigid-body and underactuated platforms (pendulums, quadrotors, hexarotors, mobile robots) are controlled using learned Hamiltonian models on $Q$ 5 or $Q$ 6. Energy shaping plus damping injection enables provably stable tracking and regulation, with real-time implementation verified in simulation and hardware (Duong et al., 2021, Li et al., 2022, Altawaitan et al., 2023).
Robust and safe navigation leverages Hamiltonian neural ODEs plus a virtual reference governor, which adaptively slows path progress to ensure safety with respect to obstacle constraints and model uncertainty, without online optimization (Li et al., 2022).

Quantum Control and Simulation:

Hamiltonian engineering in quantum devices employs Lie-algebraic and group-theoretic constructs to determine reachable subalgebras, to design control sets that guarantee full operator controllability, or confine evolution to symmetry-protected subspaces (Liang et al., 5 Mar 2026).
Bounded-strength Hamiltonian simulation (Eulerian simulation framework) enables the programmable simulation of complex many-body models (Kitaev, Heisenberg, etc.) under bounded control amplitudes, both for closed and open quantum systems, by concatenating group-structured control cycles (Bookatz et al., 2013).
Optimal control for quantum parameter estimation applies Krotov-type methods to design time-dependent pulses that maximize the quantum Fisher information (QFI), yielding orders-of-magnitude gains in estimation precision and robustness in open quantum systems (Qin et al., 2022).
Controllability of quadratic Hamiltonians with arbitrary linear drives is fully characterized via Kalman rank conditions and yields explicit analytical pulse families for quantum optics, circuit QED, and wave packet transport (Johnsson et al., 2023).

Distributed Systems:

Distributed port-Hamiltonian systems (dPHS) model PDEs with energy-based interconnection structures. Data-driven approaches use Gaussian processes to learn unknown Hamiltonian densities, coupled to boundary-interconnected finite-dimensional controllers, with probabilistic stability guarantees under model uncertainty (Beckers et al., 5 Apr 2026).

5. Analysis of Robustness, Uncertainty, and Stability

Hamiltonian control frameworks provide explicit tools for robustness and stability analysis:

Lyapunov-based analysis: The shaped Hamiltonian serves as a Lyapunov function, and closed-loop dissipation bounds energy drift and enforces convergence, even under bounded noise or unmodelled disturbances (Duong et al., 2021, Ackermann et al., 2024).
Passivity-based design: Stability in both deterministic and stochastic settings is established via energy dissipation inequalities and passive interconnection arguments, with unique invariant measures and convergence to the predefined set (Ackermann et al., 2024, Li et al., 2022, Beckers et al., 5 Apr 2026).
Model uncertainty quantification: Data-driven dPHS models incorporate posterior uncertainty from learned Hamiltonians to determine probabilistic stability regions—a direct extension of the classical passivity concept in the context of statistical learning (Beckers et al., 5 Apr 2026).

6. Algorithmic and Computational Approaches

Hamiltonian optimal control exploits the analytical structure of the Hamiltonian function:

Hamiltonian-based optimal control algorithms operate in the space of relaxed controls, descending via pointwise Hamiltonian minimization and Armijo-type step-size selection; projection schemes recover implementable, ordinary controls (Hale et al., 2016).
Gradient-based unitary design (Van Loan formalism) computes gradients of nested Dyson/Magnus series, framing effective Hamiltonian optimization as a bilinear control problem, amenable to automatic differentiation and global optimization (1904.02702).
Neural ODE and adjoint methods facilitate gradient computation for high-dimensional models and arbitrary observation spaces, supporting end-to-end controller learning and system identification (Duong et al., 2021, Altawaitan et al., 2023, Zhong et al., 2019).

7. Theoretical and Practical Implications

Energetic consistency: By embedding the structure of Hamiltonian dynamics in models and controllers, physical laws—conservation, passivity, invariance—are enforced by construction, enabling reliable extrapolation outside the training domain and robust performance in the presence of unmodeled effects (Duong et al., 2021, Zhong et al., 2019).
Sample and data efficiency: Physics-informed and structure-preserving strategies require fewer experimental trajectories, avoid energy drift and manifold violations, and produce interpretable surrogate models for further analytical exploration (Duong et al., 2021).
Versatility and expressiveness: Hamiltonian control encompasses classical, quantum, finite-, and infinite-dimensional systems, with methods translating across settings via underlying geometric and energetic principles (Liang et al., 5 Mar 2026, Beckers et al., 5 Apr 2026).
Integration with machine learning: Structured deep learning methods grounded in Hamiltonian mechanics yield models directly suitable for feedback synthesis (energy shaping, damping), while supporting provable guarantees in safety-critical robotics and high-precision quantum devices (Duong et al., 2021, Altawaitan et al., 2023, Li et al., 2022, Duong et al., 2024).

In summary, Hamiltonian dynamics with control provides a unified theoretical, algorithmic, and computational foundation for modeling, identification, and control of diverse physical systems. It leverages geometric structure to achieve robust, efficient, and interpretable performance in robotic, quantum, and distributed domains (Duong et al., 2021, Duong et al., 2024, Zhong et al., 2019, Altawaitan et al., 2023, Beckers et al., 5 Apr 2026, Li et al., 2022, Bookatz et al., 2013, Liang et al., 5 Mar 2026, Johnsson et al., 2023, Hale et al., 2016).