Robust Hands-Off Principle

• Robust Hands-Off Principle is a framework for sparse control that equates nonconvex L⁰ and convex L¹ optimization under uncertainty.
• It employs a nonsmooth robust Pontryagin Maximum Principle to enforce a bang–off–bang control pattern, ensuring extended zero-input intervals.
• The method reduces semi-infinite constraints to finite convex programs through scenario sampling, guaranteeing terminal feasibility across uncertainties.

The Robust Hands-Off Principle provides a rigorous equivalence between minimum-support control (L⁰-optimal control) and L¹-optimal control for constrained linear systems with parametric uncertainty, under specified structural conditions. This principle enables practitioners to tackle highly sparse control objectives, such as maximizing intervals of zero input (“hands-off” operation), by solving convex L¹ programs that have identical optimal solution sets to their nonconvex L⁰ counterparts. The framework is built on a nonsmooth variant of the robust Pontryagin Maximum Principle, with robust feasibility ensured for an uncountable, compact family of uncertainty scenarios. Theoretical and computational apparatus is detailed, including finite reduction to convex programs via sampling and a globally convergent solver (Ganguly et al., 12 Jan 2026).

1. Mathematical Formulation of Robust Hands-Off Control

Robust hands-off control considers uncertain, parametric LTI plants of the form: $$\dot x(t;\lambda) = A(\lambda)x(t;\lambda) + b(\lambda)u(t),\quad x(0;\lambda) = \bar x,$$ where $\lambda$ ranges over a compact convex set $\Pscr \subset \mathbb R^\nu$, with input constraint $u(t) \in [-1,1]$ for a.e.\ $t\in[0,T]$, and terminal constraints $x(T;\lambda) \in C$, where $C$ is convex.

The core objective is

$J_0(u) := \|u\|_{L^0} = \mu(\{t : u(t) \neq 0\}),$

i.e., the total measure (duration) over which the control is nonzero. The robust admissible set comprises all $u$ such that the terminal constraint is satisfied for every $\lambda \in \Pscr$: $\mathcal U_\mathrm{ad} = \left\{ u \in L^\infty([0,T],\mathbb R): u(t) \in [-1,1]\;\forall t,~ x_u(T;\lambda) \in C\;\forall \lambda \in \Pscr \right\}.$ The robust hands-off problem is therefore: $\inf_{u \in \mathcal U_\mathrm{ad}} J_0(u).$

2. The Robust Hands-Off Principle: $L^0$–$L^1$ Equivalence

Given the combinatorial, nonconvex nature of L⁰ minimization, the robust hands-off principle states that, under the normality of the Pontryagin extremal and an ensemble controllability condition, every solution of the convex L¹ program

$\inf_{u \in \mathcal U_\mathrm{ad}} \|u\|_{L^1}$

is also optimal for the robust hands-off (L⁰) criterion: $\operatorname{Opt}_0 = \operatorname{Opt}_1,$ where $\operatorname{Opt}_0$ and $\operatorname{Opt}_1$ denote the sets of L⁰ and L¹ minimizers in $\mathcal U_\mathrm{ad}$ respectively (Ganguly et al., 12 Jan 2026).

The robust Pontryagin principle reveals all minimizers have bang–off–bang structure, i.e., $u^\star(t)\in\{-1,0,1\}$ a.e., so $J_1(u^\star) = J_0(u^\star)$. The L¹ convex relaxation does not destroy sparsity in this setting, ensuring that solving with tractable convex methods yields the maximum hands-off schedule.

3. Nonsmooth Robust Pontryagin Maximum Principle

The robust maximum principle for the L¹-relaxed problem employs an augmented Mayer-state

$\dot x_{d+1}(t) = |u(t)|,\qquad x_{d+1}(0) = 0,$

alongside costates $(p(t;\lambda), q(t;\lambda))$ and multipliers. The averaged switching function, defined via the supporting Radon measure $m$,

$\sigma(t) = \int_\Pscr p(t;\lambda)^\top b(\lambda) m(d\lambda),$

governs optimality: $$u^\star(t) \in \arg\max_{v \in [-1,1]} \{\sigma(t)v - \eta|v|\}.$$ Normality ($\eta>0$) enforces strict bang–off–bang control, guaranteeing $u^\star$ has minimal support (Ganguly et al., 12 Jan 2026).

4. Semi-Infinite Program Reduction and Algorithm

The robust L⁰/L¹ problem over uncountably many scenarios admits a reduction to a finite convex program:

1. Parameterize $u(t)$ using a basis of $N$ piecewise-constant dictionary functions, $u(t) = \Theta \Psi(t)$.
2. For decision variable $\Theta \in \mathbb R^N$, impose the $N$ most critical constraints for $\lambda^1,\ldots,\lambda^N \in \Pscr$ via

$$\psi(x_\Theta(T;\lambda^i)) \le 0,\qquad i = 1,\ldots,N,$$

ensuring feasibility over $\Pscr$ (Ganguly et al., 12 Jan 2026). The value of the SIP is attained using only these critical scenarios. A global optimization layer (e.g., simulated annealing) identifies the maximizing tuple of scenarios.

1. The final algorithm alternates between convex inner solves and global maximization, converging to an exact, robust maximally hands-off solution.

5. Numerical Demonstration and Comparison

On a spring–mass–damper system with three uncertain parameters ($\lambda \in [-0.1,0.1]^3$), the robust hands-off algorithm yields a sparse control profile ($<$20% active time) while satisfying all terminal constraints for $10,000$ sampled uncertainty scenarios. Compared to standard “scenario” approximations, the sparse robust solution achieves lower support and stricter terminal feasibility (Ganguly et al., 12 Jan 2026).

Table: Comparison of methods

Method	Achieved $J_1 = J_0$	Percentage Time Active
Sampling Theorem SIP	89.01	19.8%
Scenario $N=500$	86.88	34.5%
Scenario $N=1000$	87.01	32.7%

Formulations not based on the rigorous sampling theorem either violate constraints or incur significantly higher support.

6. Theoretical and Practical Implications

The robust hands-off principle provides:

• Factual guarantee: L⁰ and L¹-optimal control coincide under stated conditions, so convex optimization can be used without sacrificing sparsity.
• Bang-off-bang structure: All minimizers are strictly sparse, yielding maximal zero-control intervals.
• Full robustness: Terminal constraints are satisfied for all $\lambda \in \Pscr$ by construction.
• Algorithmic tractability: The framework reduces semi-infinite constraints to finite convex programs via exact scenario sampling.
• Scalability: Supports large-scale uncertain systems, subject to the feasibility of the sampling theorem and computational global maximization.

7. Extensions and Limitations

Potential directions include extending to mixed parametric and disturbance uncertainty, leveraging faster global solvers for the scenario maximization, and application to energy-efficient and networked control. The principle relies on normality and controllability in the robust PMP; failure of these can invalidate equivalence and require alternative nonconvex techniques. Multiple-input channels and nonlinear systems are not addressed in the cited results (Ganguly et al., 12 Jan 2026).