Squared Effort in Control Systems

• Squared effort is a quadratic cost structure that penalizes the magnitude of control inputs, serving as a regularizer for control trajectories in optimal and stochastic control settings.
• It underpins analytical tractability by yielding linear feedback laws in deterministic problems and explicit solutions in settings involving stochastic dynamics and partial differential equations.
• The quadratic form facilitates efficient numerical optimization, enabling high-order discretization methods and robust implementations in constrained and mean field control applications.

The squared effort, also commonly termed L₂ acceleration or control cost, is a quadratic functional penalizing the magnitude of a control input over time. This cost structure is foundational in optimal control theory, stochastic control, and contract theory, serving both as a regularizer of control trajectories and as an economic model of effort costs. The archetypal objective is the minimization of functionals of the form $\int_0^T \|u(t)\|^2 dt$ (or extensions thereof), subject to system dynamics and possibly further constraints. The ubiquity of the quadratic form arises from its analytical tractability—yielding linear feedback laws in linear settings, semilinear or fully nonlinear partial differential equations (PDEs) in more general frameworks, and admitting explicit solutions or rigorous existence and uniqueness results under broad circumstances.

1. Mathematical Formulation and Canonical Problems

The classical L₂ effort cost is defined via the objective functional

$J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$

where $u(t)$ is a control trajectory. This cost arises naturally in deterministic, stochastic, and mean field settings. In the context of controlled diffusions, the state evolves as

$dX_t = u_t\,dt + \sigma\,dW_t,$

with a running cost $c(u_t) = \|u_t\|^2$ (Chen et al., 11 Mar 2025). In deterministic systems, problems such as the double integrator (second-order dynamics) over $[0,T]$ with $\ddot{x}(t) = u(t)$ and $J(u) = \int_0^T u(t)^2 dt$ are canonical (Lewis, 2010, Achdou, 6 Oct 2025).

Squaring the control (the Euclidean norm in vector cases) ensures strict convexity, analytic tractability, and penalizes large, oscillatory, or energy-expensive control signals.

2. Solution Structures: Pontryagin Principle, HJB, and Explicit Formulas

The minimization of the L₂ cost under given system dynamics is addressed by the Pontryagin Maximum Principle (PMP) or, in stochastic settings, the Hamilton–Jacobi–Bellman (HJB) equation. For instance, in the double-integrator setting over a fixed interval, the Hamiltonian is

$H = \lambda_1 \dot{x} + \lambda_2 u - u^2,$

where $u^*(t) = \lambda_2(t)/2$ is optimal away from constraints. For the data $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$0, $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$1, with zero initial and terminal velocities, the optimal unconstrained control is affine in time: $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$2 If $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$3 exceeds admissible limits, bang-singular-bang structures arise (Lewis, 2010).

In stochastic principal–agent models, the agent's value function $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$4 solves the semilinear PDE

$J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$5

with terminal condition $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$6, where $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$7 is a concave utility (Chen et al., 11 Mar 2025). For control-affine problems with quadratic cost, the Schrödinger equation with a nonlinearity quadratic in the gradient arises systematically.

For control on acceleration (second derivative), the value function $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$8 satisfies, in the interior: $J[u(\cdot)] = \int_0^T \|u(t)\|^2 dt,$9 Closed-form solutions have been derived for $u(t)$0, with state constraint multipliers enforcing boundaries (Achdou, 6 Oct 2025).

3. Analytical Properties and Existence Theorems

The quadratic cost structure enables rigorous analysis of existence, uniqueness, and regularity of optimal solutions. In the principal–agent context with quadratic effort, the existence of an optimal contract follows from the solvability of a corresponding fully nonlinear degenerate parabolic PDE: $u(t)$1 $u(t)$2, with appropriate terminal and growth conditions (Chen et al., 11 Mar 2025). Under sufficient smoothness and boundedness assumptions, $u(t)$3 exists uniquely with monotonicity $u(t)$4 uniformly.

In deterministic boundary-value problems, explicit closed forms for $u(t)$5, $u(t)$6, and the minimal cost function $u(t)$7 exist for the unconstrained case. When state or control constraints are present, the cost function remains strictly monotonic, and the optimal control displays switching behavior with explicitly computable switching times (Lewis, 2010, Achdou, 6 Oct 2025).

For mean field and state-constrained problems, the value function may be piecewise defined, and singular arcs (constant or affine controls) play a central role in connecting coasts and brakes (Achdou, 6 Oct 2025).

4. Numerical and Algorithmic Implementations

The discretization and numerical optimization of L₂ costs are optimized by exploiting the quadratic structure. In long-horizon nonlinear optimal control, the Gauss–Newton Runge–Kutta (GNRK) discretization integrates the L₂ cost with high-order accuracy and efficient Hessian computation: $u(t)$8 is discretized over $u(t)$9 intervals as

$dX_t = u_t\,dt + \sigma\,dW_t,$0

where $dX_t = u_t\,dt + \sigma\,dW_t,$1 is piecewise constant. The Gauss–Newton approximation yields a block-diagonal, positive semidefinite Hessian, facilitating sequential quadratic programming (SQP) and preserving solver sparsity. The method leverages the least-squares structure for efficient gradient and Hessian evaluations without requiring full second-order derivatives (Frey et al., 2023).

In the regime of expensive control (very small $dX_t = u_t\,dt + \sigma\,dW_t,$2 in $dX_t = u_t\,dt + \sigma\,dW_t,$3), perturbative expansion in $dX_t = u_t\,dt + \sigma\,dW_t,$4 allows the recasting of non-convex optimization into a hierarchy of convex semidefinite programs using sum-of-squares (SOS) relaxations (Huang et al., 2016).

5. Applications and Structural Consequences

Quadratic effort costs are central to a range of applications:

• Principal–agent contract theory: Drift-controlled diffusions with quadratic cost in the agent's effort yield explicit analytic solutions and guarantee the existence of a unique, optimal contract when coupled with well-posedness of the associated nonlinear PDE (Chen et al., 11 Mar 2025).
• Feedback control of fluid flows: The “expensive control” methodology has been validated in reduced-order models and direct numerical simulation of flows (vortex shedding control) by regularizing feedback laws via L₂ costs, with explicit controller forms obtained through SOS programming (Huang et al., 2016).
• State-constrained mean field games: Squared acceleration cost yields tractable closed-form and piecewise-smooth solutions, accommodating state constraints and providing insight into transition regimes between coasting and braking (Achdou, 6 Oct 2025).
• Classical trajectory control: Problems of moving point masses with fixed or free terminal time, quadratic acceleration penalties, and control bounds yield affine, bang–singular–bang, or pure singular solutions (Lewis, 2010).

A selection of mathematical regimes and solution forms is shown below:

Regime	Control Form	Minimal Cost
Unconstrained, fixed $dX_t = u_t\,dt + \sigma\,dW_t,$5	$dX_t = u_t\,dt + \sigma\,dW_t,$6	$dX_t = u_t\,dt + \sigma\,dW_t,$7
Control bounded, $dX_t = u_t\,dt + \sigma\,dW_t,$8	Singular arc (affine control)	$dX_t = u_t\,dt + \sigma\,dW_t,$9
Control bounded, $c(u_t) = \|u_t\|^2$0	Bang-singular-bang ($c(u_t) = \|u_t\|^2$1 segments + singular)	Piecewise, strictly decreasing
State-constrained acceleration	Piecewise: coast, smooth brake, or singular brake	Piecewise formula by region

6. Limitations, Degeneracy, and Structural Monotonicity

A notable feature of the L₂ effort cost is that, in free horizon problems, the minimal cost decreases without bound as $c(u_t) = \|u_t\|^2$2 (e.g., $c(u_t) = \|u_t\|^2$3), hence no optimizer exists for the free-time case—the infimum is not attained (Lewis, 2010). For state-constrained or bounded control cases, monotonicity and explicit switching times can be established, but singularities at constraint boundaries or regime transitions must be treated explicitly (Achdou, 6 Oct 2025).

In stochastic contract design, degenerate directions in the nonlinear PDE (e.g., no term in $c(u_t) = \|u_t\|^2$4) are a result of the quadratic structure and must be handled via precise monotonicity and regularity arguments to ensure well-posedness (Chen et al., 11 Mar 2025).

7. Connections and Extensions

The L₂ cost structure generalizes directly to quadratic forms $c(u_t) = \|u_t\|^2$5, with $c(u_t) = \|u_t\|^2$6, accommodating weighted penalties in multi-input settings (Frey et al., 2023). Elliptical “moderation incentives” and higher-order control norms provide alternative penalties that smoothly interpolate between submaximal control and maximum-effort bang arcs (Lewis, 2010). The quadratic cost's analytical tractability makes it the foundational case for both extensions to nonlinear costs (e.g., $c(u_t) = \|u_t\|^2$7, $c(u_t) = \|u_t\|^2$8) and for the development of efficient direct and indirect numerical methods, as in GNRK or SOS/SDP frameworks (Frey et al., 2023, Huang et al., 2016).

Quadratic effort costs remain the canonical tool for regularizing and analyzing the complexity, smoothness, and implementability of optimal controls across deterministic, stochastic, and game-theoretic optimization settings.