Hybrid Optimal Control Methods

• Hybrid optimal control is a framework for systems that combine smooth, continuous state evolution with discrete event-driven transitions and resets.
• It employs methodologies like the hybrid maximum principle, dynamic programming, and advanced numerical methods to handle nonconvexity and combinatorial complexity.
• Applications span power electronics, automotive energy management, quantum systems, and PDE control, addressing challenges in scalability and robustness.

Hybrid optimal control concerns the analysis, synthesis, and optimization of controlled dynamical systems that combine both continuous-time evolution and discrete (event-driven or state-triggered) transitions, including phenomena such as mode-switching, state resets, and resets that may involve changes in the dimension of the state space. Such systems are central to fields ranging from power electronics and automotive systems to process control and quantum systems. The optimal control of hybrid systems requires novel mathematical models, tailored necessary conditions for optimality, and specialized algorithms capable of handling both the continuous and discrete aspects of the dynamics, as well as the combinatorial complexity introduced by logic-driven mode transitions and nontrivial reset maps.

1. Mathematical Formulations of Hybrid Systems and Hybrid Optimal Control

Hybrid systems are formalized on smooth, finite-dimensional manifolds $M$, equipped with a collection of continuous dynamics, event (guard) manifolds $S \subset M$ governing the occurrence of discrete transitions, and reset maps $\Delta : S \to M$ that can effect state jumps, potentially of lower dimension than $S$. Controls $u(t)$ are typically piecewise-continuous functions taking values in a closed subset $U \subset \mathbb{R}^m$, and the continuous state $x$ evolves according to vector fields $f: M \times U \to TM$ which are $C^\infty$ in both arguments except on the guard set, where discrete resets may occur. The canonical hybrid control system is

$$\dot{x} = f(x, u), \quad x \notin S; \qquad x^+ = \Delta(x^-), \quad x^- \in S.$$

The objective is the minimization of a hybrid cost functional of Bolza type: $$J(u; t_0, x_0) = \int_{t_0}^T \ell(x(t), u(t))\, dt + \varphi(x(T)),$$ with appropriate terminal and transversality conditions. Admissible hybrid trajectories undergo a finite number of discrete transitions (reset times $t_1 < \cdots < t_N$), with no Zeno behavior, and satisfy regularity requirements on both the control and the flows between resets (Clark et al., 2024).

2. Necessary Conditions: Hybrid Maximum Principle and Hamiltonian Variation

For hybrid systems, necessary conditions for optimality generalize Pontryagin’s Maximum Principle (PMP) and take the form of the Hybrid Maximum (or Minimum) Principle. The continuous-time arcs adhere to canonical Hamiltonian equations: $$\dot{x} = \frac{\partial H}{\partial p}, \quad \dot{p} = -\frac{\partial H}{\partial x}, \qquad H(x, p) = \min_{u \in U} \langle p, f(x, u)\rangle + \ell(x, u).$$ At each discrete event, the so-called Hamiltonian jump conditions govern the evolution of the costate $p$:

• General reset (diffeomorphism):

$$p^+ \circ D\Delta(x^-) - p^- \in \operatorname{Ann}(T_{x^-}S), \qquad H(x^+, p^+) = H(x^-, p^-).$$

• Submersive resets (dimension drop): The jump condition is generally overdetermined. Feasibility requires the pre-reset $p^-$ annihilates the directions in $\ker D\Delta(x^-)$, and the set of valid post-reset $p^+$ forms an affine space determined by right-pseudo-inverses of $D\Delta$ and energy-matching (Hamiltonian) constraints (Clark et al., 2024, Pakniyat et al., 2017). For submersive resets, this set is non-singleton, and uniqueness is restored by propagating the admissible set forward and intersecting again at the next event.

Boundary conditions include terminal transversality (alignment of costate with the gradient of the terminal cost along admissible directions) (Clark et al., 2024, Pakniyat et al., 2016). In the presence of switching costs or mode-dependent costs, jump conditions for the adjoint and the cost functional include explicit gradients of the cost with respect to switching variables (Pakniyat et al., 2017, Pakniyat et al., 2016).

3. Structural and Geometric Features of Hybrid Resets

When the reset map $\Delta$ is a submersion onto its image with constant rank $r < \dim S$, its differential $D\Delta$ necessarily has a nontrivial kernel. This leads to situations where, at a reset, the co-state jump condition is under- or over-determined: unless $p^-$ annihilates the kernel, no solution exists; otherwise, there are infinitely many. The space of post-reset costates $p^+$ is an affine manifold of codimension $r$ inside the cotangent space of the new mode (Clark et al., 2024).

A selection principle emerges from the hybrid geometry: propagate the entire family of admissible $p^+$ under the continuous Hamiltonian flow until the next reset, where consistency must be imposed again. Generically, the intersection of the propagated set with the admissibility manifold is finite, selecting the true optimal $p^+$. This mechanism is required, for instance, in systems with energy loss at impact or changes in hidden or "gauge" variables (Clark et al., 2024).

4. Algorithmic Approaches and Numerical Methods

Multiple algorithmic paradigms address the intrinsic combinatorial and nonsmooth structure of hybrid optimal control:

• Discretize-then-optimize with direct transcription: Discrete-time approximations (e.g., explicit Euler, direct collocation) encode the system as a mixed-integer nonlinear program (MINLP) with state, control, and mode-selection variables (Nikitina et al., 2024, Westenbroek et al., 2015). The resulting problem is attacked using augmented Lagrangian frameworks and SMIL techniques, yielding critical points that are first-order stationary (not necessarily globally optimal).
• Relaxed control and measure-based convexification: Convex relaxations based on occupation measures, sum-of-squares hierarchies, and moment-SDP approximations enable the computation of globally valid lower bounds for highly nonconvex objectives (prominent in power electronics modulation) (Miller et al., 8 Dec 2025).
• Complementarity constraint reformulations: Mode selection and switching logic are imposed via mathematical programs with complementarity constraints (MPCC), and solved via moving finite elements, smoothing regularizations, and globalization via NLP/MILP hybridization (Kazi et al., 5 Mar 2025, Hempel et al., 2015). Under special geometric conditions (e.g., block-PSD, rank-one structure), full Karush-Kuhn-Tucker (KKT) regularity and stationarity can be achieved.
• Dynamic programming and hybrid Bellman principles: In low-dimensional, discrete-time cases, dynamic programming approaches using grids on the hybrid state and control spaces remain practical, especially for embedded implementation in automotive systems (Miretti et al., 2023).
• Two-stage (hybrid) optimization in quantum systems: In high-dimensional systems, a hybrid policy combining global, low-dimensional, gradient-free search (simplex/Nelder-Mead) with local, high-dimensional, gradient-based refinement (e.g. Krotov's method or GRAPE) achieves order-of-magnitude acceleration and maintains solution simplicity (Goerz et al., 2015, Egger et al., 2014).

5. Case Studies and Domain-Specific Applications

Hybrid optimal control methodology has been successfully applied across a diverse set of engineering and scientific domains:

• Impact systems and submersive resets: In a bouncing ball with internal variable, enforcing the refined jump rules yields a closed-form, globally optimal two-bounce strategy, confirmed by cost evaluation over competing solutions (Clark et al., 2024).
• Quantum control: In holonomic phase gate synthesis for transmons, hybrid (simplex + Krotov) optimization achieves high-fidelity gates, rapid convergence, and simple, experimentally robust controls (Goerz et al., 2015, Egger et al., 2014).
• Power electronics and OPP: Synthesis of optimal pulse patterns for multilevel converters via hybrid optimal control and measure relaxations enables near-global optimality with tractable conic programming relaxations; this approach scales effectively with system size (Miller et al., 8 Dec 2025).
• Automotive energy management: Supervisory control of hybrid electric vehicles uses hybrid optimal principles via NMPC, direct collocation, and DP, leading to efficient power-split strategies and physically interpretable rule-extraction for real-time controllers (Uthaichana et al., 2018, Miretti et al., 2023).
• Hybrid PDE systems: Hybrid optimal control over switching semilinear parabolic PDEs yields nontrivial regularity results and extends Hamiltonian constancy properties across infinite-dimensional transitions (Court et al., 2016).

6. Theoretical Unification and Equivalence with Dynamic Programming

The Hybrid Minimum Principle and Hybrid Dynamic Programming are unified under broad regularity conditions: the adjoint process from the necessary conditions and the gradient of the value function in the dynamic programming formulation satisfy identical ODEs and boundary/jump conditions along optimal trajectories (Pakniyat et al., 2016). In the linear-quadratic regime, this equivalence can be made explicit, with coupled Riccati equations and feedback formulas, both in finite and infinite dimensions (Pakniyat et al., 2016, Court et al., 2016).

7. Open Problems and Future Directions

Challenges remain in the development of scalable, globally optimal solvers for hybrid optimal control in the presence of high-dimensional combinatorial structure and non-convexity. Current research targets:

• Exploiting geometric mechanics for symmetry reduction (e.g., Lie-Poisson reduction) in hybrid Lie group systems, which introduces extra state dimensions but may lead to computationally tractable reduced-order models (Clark et al., 2024).
• Developing theoretically sound globalization and convergence guarantees for hybrid MPCCs and mixed-integer augmented Lagrangian algorithms operating under realistic constraint qualifications (Nikitina et al., 2024, Kazi et al., 5 Mar 2025).
• Efficient high-order methods for hybrid systems with sliding modes, Zeno behavior, or nontrivial reset geometry (including non-injective resets) (Pytlak et al., 2021).
• Enhanced integration with stochastic filtering and partial observation, allowing for tractable synthesis and feedback in hybrid stochastic environments (Lv et al., 2023).

Hybrid optimal control remains an area of rapid development at the interface of geometry, nonsmooth analysis, combinatorial optimization, and high-dimensional computation, drawing on theory and application from diverse scientific and engineering fields (Clark et al., 2024, Miller et al., 8 Dec 2025, Pakniyat et al., 2017, Hempel et al., 2015, Pakniyat et al., 2016, Clark et al., 2024, Court et al., 2016).