Modern control theory has moved far beyond standard linear-quadratic regulators (LQR). To handle real-world uncertainties, engineers use advanced frameworks like distributionally robust control, mixed $H_2/H_\infty$ optimization, and regret-optimal control. However, a fundamental mathematical paradox arises when solving these problems over an infinite horizon in the frequency domain: the mathematically exact optimal controllers are often non-rational.
As shown by Vikrant Malik et al. in Optimal Infinite-Horizon Mixed H2/ H∞ Control, these optimal control laws frequently contain irrational mathematical terms, such as square roots of polynomials.
Why Non-Rationality Prevents Physical Implementation
Why do irrational terms matter? Standard digital hardware—such as microcontrollers or embedded systems—operates on finite-dimensional state-space models. A standard controller updates its internal state and outputs control signals using a finite matrix equation:
$$x_{k+1} = Ax_k + Bu_k$$ $$y_k = Cx_k + Du_k$$
In the frequency domain, this structure restricts the controller to rational transfer functions—meaning they must be expressible as a ratio of polynomials (like a standard $z$-transform).
A non-rational transfer function, by contrast, corresponds to an infinite-dimensional state space. To execute such a control loop perfectly, physical hardware would need to store and process an infinite history of past states and disturbances. As Taylan Kargin et al. point out in Infinite-Horizon Distributionally Robust Regret-Optimal Control, this lack of a finite-order state-space representation makes direct implementation on real-time embedded hardware impossible.
The Power of Low-Order Rational Approximations
To bridge this gap, researchers do not try to implement the exact irrational controller. Instead, they look for a low-order rational transfer function that closely mimics the ideal controller's frequency response.
The core intuition for why this works traces back to functional analysis. In the context of $l^1$-optimal control, Andrew Eberhard et al. demonstrated in Slice convergence of parametrised sums of convex functions in non-reflexive spaces that physically realizable controllers with rational $z$-transforms can achieve performance levels arbitrarily close to the theoretical optimum. Because the underlying system behaviors are bounded-input bounded-output stable, the infinite-dimensional tail of the ideal controller contains diminishing amounts of energy.
Remarkably, you do not need a high-order polynomial to capture this energy. Work by Joudi Hajar et al. in The Distributionally Robust Infinite-Horizon LQR reveals that a simple second-order rational approximation is often sufficient to achieve nearly all the performance benefits of the complex, infinite-dimensional ideal model, bringing the expected control cost within one percent of the theoretical optimum.
Engineering the Approximation via Convex Optimization
How do we find the best rational approximation? Rather than relying on heuristic curve-fitting, the modern approach frames this as a structured optimization problem.
Luca Furieri et al. established in An Input-Output Parametrization of Stabilizing Controllers: amidst Youla and System Level Synthesis that infinite-dimensional frequency-domain problems can be systematically approximated as finite-dimensional programs without sacrificing closed-loop stability. By treating specific closed-loop transfer matrices as design parameters, they bypass complex pre-computations and open the door to convex optimization.
Building on this, the authors of Distributionally Robust Kalman Filtering over an Infinite-Horizon map non-rational controllers into standard finite-dimensional forms by solving a convex feasibility problem. This method minimizes the $H_\infty$ norm error between the ideal non-rational controller and the target rational polynomial ratio:
$$\min_{\text{Rational } \hat{K}} \| K_{\text{exact}} - \hat{K} \|_{\infty}$$
However, the choice of the approximation order requires care. As Joudi Hajar et al. caution, selecting an insufficient approximation degree can lead to severe performance degradation and extreme cost inflation under specific uncertainty bounds. When tuned correctly, as emphasized by Taylan Kargin et al. in Wasserstein distributionally robust regret-optimal control over infinite-horizon, this optimization pipeline yields high-quality, low-dimensional controllers that make advanced, theoretically ideal control laws practical for real-time embedded systems.
Key Foundational Work
Slice convergence of parametrised sums of convex functions in non-reflexive spaces (1999) — contributed 1 concept elements to this direction.
An Input-Output Parametrization of Stabilizing Controllers: amidst Youla and System Level Synthesis (2019) — contributed 1 concept elements to this direction.