Introduction
For decades, $H_\infty$ control has served as the gold standard for robust control, providing a rigorous framework for managing worst-case disturbances in dynamical systems. However, this framework has traditionally been tethered to quadratic cost functions (e.g., $x^T Q x$). While mathematically convenient—leading to the well-known Algebraic Riccati Equation (ARE)—quadratic costs are inherently rigid. They struggle to encode "hard" constraints, such as actuator saturation limits or safety envelopes, without resorting to ad-hoc workarounds or computationally expensive online optimization.
This paper breaks that long-standing bond by introducing a novel approach using Bregman divergences. The authors derive a closed-form, nonlinear $H_\infty$ controller capable of handling general strictly convex costs over an infinite horizon. This provides the flexibility of Robust Model Predictive Control (MPC) without the requirement of solving complex optimization problems at every time step.
The Problem: The Rigidity of Quadratics
In robust control, the primary goal is to minimize the impact of external disturbances on system performance. The classical $H_\infty$ solution relies on the "completion-of-squares" technique, which naturally favors quadratic penalties. While these penalties are easy to compute, they cannot represent physical realities like power limits or "keep-out" zones effectively.
Currently, engineers seeking to use non-quadratic costs typically turn to Robust MPC. However, MPC's reliance on short, finite horizons and its high computational demand at each time step make it difficult to implement in high-speed or resource-constrained environments. The authors aim to bridge this gap by providing a "set-it-and-forget-it" nonlinear controller that maintains the efficiency of classical methods while allowing for much richer cost structures.
Methodology: From Squares to Bregman Divergences
The core innovation of this research is the replacement of the classical squared norm with the Bregman divergence. A Bregman divergence measures the "gap" between a convex function and its linear approximation, offering a more generalized way to define distance and cost.
The Non-Quadratic Riccati Identity
By leveraging the Bregman three-point identity, the authors derive a new, generalized version of the Riccati equation. Unlike the standard matrix-based ARE, this is a functional identity involving Fenchel duals—a mathematical tool for representing convex functions in "slope space." This identity ensures that the system maintains a telescoping sum property, which is the underlying algebraic requirement for $H_\infty$ stability.
Offline Synthesis
To make this theoretical breakthrough practical, the paper outlines a "recipe" for controller synthesis (Theorems 3, 4, and 5). A designer can define specific cost structures—such as a desired state penalty—and solve a one-time convex feasibility program. Once this offline step is complete, the resulting controller is a simple, time-invariant nonlinear function that can be executed instantly in real-time.
Key Insights and Results
Natural Nonlinearity
One of the most compelling aspects of this approach is that the nonlinearity of the controller arises naturally from the cost function. For example, if a designer uses a cost function that approaches infinity as the state reaches a certain boundary, the resulting controller will automatically "push back" harder as the system nears that limit. This creates a built-in safety shield without the need for external logic or clipping.
Theoretical Generalization
This framework is not merely a new method but a true generalization of robust control theory. If standard quadratic costs are applied to the authors' formulas, the mathematics collapses back into the classical $H_\infty$ equations and the standard linear feedback law.
Performance in Simulation
The authors demonstrated the efficacy of their approach through several simulated scenarios:
- Input-Limited Control: The controller naturally saturated at hard limits while maintaining stability, whereas traditional controllers would require ad-hoc clipping that could compromise performance.
- Safety Envelopes: The system successfully maintained a "hard safety envelope" (e.g., $|x| \leq 0.2$), keeping the state within bounds even when subjected to various noise profiles, including White, Uniform, and Laplacian disturbances.
Limitations and Future Directions
While the paper marks a significant advancement, several challenges remain for future research:
- Actuation Constraints: The current practical synthesis assumes the system is fully or over-actuated. Applying this to under-actuated systems—common in robotics—remains a complex open problem.
- Storage Function Discovery: Finding the appropriate "storage function" $p(x)$ for an arbitrary set of costs is still non-trivial. The current method requires fixing certain parts of the cost structure to ensure feasibility.
- Output Feedback: The current work assumes "full information" (access to all states and disturbances). Extending the Bregman approach to an observer-based framework for systems with noisy sensors is a necessary next step for broader industrial application.
Summary updated