Discontinuity-Sensitive Optimal Control

• Discontinuity-sensitive optimal control is a class of methods that explicitly models and manages abrupt changes and nonsmooth behaviors in control systems.
• It employs advanced analytical frameworks—including anisotropic measures, piecewise differentiability, and persistent homology—to accurately capture switching surfaces and mode transitions.
• These techniques improve system reliability and computational efficiency in diverse applications such as robotics, obstacle avoidance, and financial engineering.

Discontinuity-sensitive optimal control denotes a class of methodologies and analytical frameworks designed to explicitly identify, represent, and efficiently handle discontinuities and nonsmooth structures inherent to the value function, optimal control policy, or system trajectories in optimal control problems. Such discontinuities typically manifest as switching surfaces, mode transitions, impulsive controls, state or co-state jumps, or abrupt changes in the parameter–solution mapping, often driven by system dynamics, constraints, multi-modality, obstacles, or nonconvexities. Ignoring these features degrades solution quality, reliability, and computational tractability. Discontinuity-sensitive approaches, therefore, combine mathematical analysis, control-theoretic insight, and algorithmic advances to both quantify and computationally resolve the effects of discontinuities in deterministic, stochastic, and hybrid control settings.

1. Sources and Manifestations of Discontinuity in Optimal Control

Discontinuities naturally arise in a broad range of optimal control formulations:

• Switching Surfaces and Bang–Bang Controls: Certain Hamilton–Jacobi–Bellman (HJB) structures or minimum-time problems produce piecewise constant (“bang–bang”) optimal controls, inducing discontinuities in both control input and co-state (2208.00065, Oguri, 2023).
• Multi-Modal and Homotopy Classes: Nonconvexity, obstacle-avoidance, and system/environmental symmetry foster multiple distinct minimizers (homotopy), resulting in discontinuous parameter–solution maps (Merkt et al., 2020, Tang et al., 2018). For instance, the cart-pole swing-up or quadrotor-obstacle navigation can have solution branches featuring abrupt transitions.
• Impulsive and Chattering Phenomena: Minimizing sequences can feature oscillations or converge to impulsive controls, driving the system through Dirac-type inputs and causing state-jumps (Henrion et al., 2018, Fusco et al., 2024).
• Flux Discontinuities and Hybrid Dynamics: Discontinuous fluxes in conservation laws, hybrid systems with mode switching, and contact mechanics yield kinks or jumps in the value function and policy (Adimurthi et al., 2016, Banjanin et al., 2017).
• Discontinuous Terminal or Path Costs: Piecewise constant or indicator-type Mayer costs and discontinuous payoff functions translate immediately into value-function discontinuities (Bouchet et al., 2021, Esfahani et al., 2012, Hamadène et al., 2019).
• Stochastic Switching and càdlàg Paths: In stochastic optimal control, discontinuities arise from càdlàg cost processes and through barrier-type obstacles in reflected BSDEs and QVI frameworks (Hamadène et al., 2019).

The principal consequences of such phenomena include the failure of classical differentiability assumptions, multi-valuedness of optimal solutions, and ill-conditioning for gradient-based algorithms. Value functions may be only piecewise differentiable (PC^r), locally Lipschitz, or merely upper/lower-semicontinuous.

2. Analytical Foundations: Mathematical Representations and Theory

A variety of analytical frameworks have been constructed to precisely represent and analyze discontinuities:

• Anisotropic Parametrized Measures: Functional-analytic extensions of Young, DiPerna–Majda, and anisotropic parametrized measures compactify control/state trajectories, interpret oscillation/concentration/jump phenomena, and allow relaxation to measure-valued solutions. The relaxed cost functional decomposes into Lebesgue (smooth) plus atomic (jump) contributions, and can be solved via convexified LP/SDP formulations (Henrion et al., 2018).
• Piecewise Differentiability and Clarke Subgradients: For contact-rich or hybrid systems, value and policy functions exhibit only PC^r-regularity. The Clarke subdifferential framework is used to define generalized gradients and guide subgradient or bundle-based optimization (Banjanin et al., 2017).
• Persistence-based Topological Analysis: Tools from algebraic topology, specifically persistent homology, extract physically meaningful discontinuity and multi-modality structure from large trajectory datasets, separating solution branches via Betti numbers (holes in the space) and supporting clustering for learning (Merkt et al., 2020).
• Advanced PDE Theory: Weak viscosity solutions, semicontinuous envelopes, and quasi-variational inequalities accommodate discontinuities of the value function or solution in both deterministic and stochastic HJB settings. For discontinuous payoff functions, uniqueness and composition must be interpreted in terms of lower- or upper-semicontinuous envelopes (Esfahani et al., 2012, Hamadène et al., 2019).
• Hybrid Maximum Principles and Impulsive Analysis: For impulsive systems, especially with delays, the necessary conditions require discontinuity-sensitive adjoint equations, Weierstrass conditions over attached controls, and admit non-uniqueness even in scalar-valued measure settings (Fusco et al., 2024). In unbounded control, infimum gaps may appear or vanish based on normality of extremals (Motta et al., 2018).

3. Discontinuity-Sensitive Numerical Solution Methods

Capturing and efficiently solving for discontinuous solutions requires numerical schemes and optimization methods that depart fundamentally from standard smooth approaches:

• Hybrid and Mode-Tracking Trajectory Optimization: Approaches explicitly enumerate or encode mode-sequences, freeze time around state jumps, or use Filippov inclusions to shift discontinuities into the right-hand side of the ODE, mitigating combinatorial explosion in mode events (Nurkanović et al., 2020).
• Discontinuity-Bounding and Adaptive Discretization: Flexible mesh methods treat mesh locations as variables, allowing nodes to cluster at discontinuities. Edge-detection and jump-function approximations automate targeted $h$-refinement, placing collocation points precisely at switching times and preserving approximation accuracy in the presence of chattering (Miller et al., 2020, Nita et al., 2022, Abadia-Doyle et al., 2024).
• Derivative-Free and Hybrid Optimization: Problems with piecewise constant or indicator-type costs—where gradients vanish or are undefined—are handled using outer-loop derivative-free optimizers (e.g., MADS/Nomad) to shepherd singular state variables, while retaining smooth optimizers for the remaining trajectory (Bouchet et al., 2021).
• Smoothing and Regularization: Smooth indirect methods replace discontinuous activation, multiplier, or costate functions with sharp $C^1$ approximations (e.g., sigmoids, mollifiers, Gaussian kernels) that preserve safety and convergence while rendering the underlying TPBVPs numerically accessible (Oguri, 2023).
• Mixture-of-Experts and Regime Clustering: For learning parameter–solution mappings, solutions are clustered according to geometric or topological features; within each cluster, a separate regressor is trained, and hard- or soft-gating identifies the correct expert at inference (Merkt et al., 2020, Tang et al., 2018).

4. Learning and Warm-Start Strategies Adapted to Discontinuity

Classic “universal” regressors and function approximators (e.g., MLPs mapping problem parameters to trajectory seeds) fail in discontinuous or multi-modal regimes, since smoothing across discontinuities produces invalid averages and low-quality warm-starts. Discontinuity-sensitive learning instantiates:

• Topology-Guided Clustering: Persistent homology identifies discontinuous branches (homotopy classes) by extracting long-lived $H_1$ cycles in trajectory datasets. Each branch is assigned a cluster, and clustering drives the training of a mixture-of-experts (Merkt et al., 2020).
• Mixture-of-Experts with Decoupled Training: Regressor/classifier MoE architectures, trained with hard gating and domain-aware features, sharply outperform single-model approaches at capturing abrupt transitions and tracking performance, especially in high-stakes multi-modal tasks like obstacle-rich navigation (Tang et al., 2018).
• Warm-Start Reliability Metrics: Discontinuity-sensitive warm-start prediction achieves nearly 100% convergence rates in DDP-based shooting methods (e.g., MoE-topology ≈99.8% vs. vanilla MLP ≈17.2% in quadrotor maze), while exhibiting dramatic reductions in required iterations and tracking cost (Merkt et al., 2020).
• Regularization via Network Architecture: In bang–bang cases, steep but continuous (e.g., tanh) activations in actor-critic structures regularize discontinuous control maps, accurately approximating switching boundaries without chattering and preserving convergence (2208.00065).

5. Applications, Algorithmic Tradeoffs, and Impact

Discontinuity-sensitive optimal control is essential in domains where system or problem non-smoothness is fundamentally unavoidable:

• Contact-Rich Robotics: Legged locomotion, manipulation, and hybrid systems require piecewise or subgradient-based optimization, mode-tracking, and PC^1 sensitivity analysis (Banjanin et al., 2017).
• Obstacle Avoidance and Motion Planning: Minimum-time and clearance functions in kinodynamic environments contain propagating discontinuity manifolds; precise identification enables navigation algorithms to avoid sharp cost increases that can destabilize planners (Armstrong et al., 2022).
• PDE-Constrained Control and High-Dimensional Dynamics: Discontinuous Galerkin, hp-mesh, and time-dG methods ensure sharp layer resolution and stability in convection-dominated or reaction-diffusion PDEs, as seen in ecological or fluids applications (Karatzas, 2024, Akman et al., 2013).
• Stochastic Switching and Financial Engineering: Control under càdlàg cost processes, quasi-variational inequalities, and reflected BSDEs represent discontinuous stochastic phenomena such as regime-switching costs in finance and energy (Hamadène et al., 2019).
• Impulsive and Time-Delay Systems: Measure-theoretic and extended-process approaches characterize necessary conditions and non-uniqueness in settings with both impulsivity and feedback, such as networked or delay-coupled control systems (Fusco et al., 2024).

Compared to smooth-only frameworks, discontinuity-sensitive methods deliver substantial gains in solution quality, robustness, and solver efficiency, especially in tasks characterized by multi-modality, switching, or physical nonsmoothness.

6. Limitations, Scalability, and Open Directions

Several challenges persist in discontinuity-sensitive optimal control:

• Scalability: Persistence-based and measure-theoretic methods can be computationally intense (e.g., persistent homology filtrations scale as $O((N(T-1))^2)$), though recent parallelization and sparsification mitigate this for large datasets (Merkt et al., 2020).
• Cluster Design and Expert Selection: MoE methods rely on appropriate clustering, which may require domain insight or the development of automatic geometric/topological feature extraction. Over-segmentation risks data sparsity, while under-segmentation fails to resolve discontinuities (Tang et al., 2018).
• Theoretical Gaps: Proof of global convergence for flexible mesh, jump detection, and hybrid optimization methods under repeated discontinuity bracketing remains open, as does rigorous analysis of hp–$h$/$p$ tradeoffs (Nita et al., 2022, Miller et al., 2020).
• Combinatorial Explosion in Hybrid and High-Dimensional Settings: Fully enumerative hybrid schemes rapidly become intractable as the number of modes/scenarios scales, despite time-freezing and Filippov-type representations (Nurkanović et al., 2020).
• Extensions to Path-Dependent and Non-Markovian Problems: Path-dependent stochastic control and high-index DAE systems with state or control-triggered discontinuities remain challenging (Hamadène et al., 2019, Nita et al., 2022).

This suggests that future development of discontinuity-sensitive optimal control will likely focus on efficient high-dimensional algorithms, automated identification of discontinuity structure, and rigorous convergence guarantees for hybrid learning-numerics frameworks.