When designing control systems for real-world applications—such as steering a spacecraft or managing a wind turbine's power grid—engineers must account for uncertainty. Whether this uncertainty stems from unmodeled dynamics, fluctuating environments, or incomplete statistical data, the goal remains the same: guarantee safety and performance under the worst-case scenario.
Mathematically, this objective is framed as a minimax optimization problem:
$$\min_{\text{Controller}} \max_{\text{Uncertainty}} \text{Cost}(\text{Controller}, \text{Uncertainty})$$
While conceptually elegant, this formulation is notoriously intractable. The inner maximization often searches over an infinite-dimensional space of disturbances or probability distributions, while the outer minimization operates over non-convex policy spaces.
A rapidly growing body of research solves this dilemma through convex reformulation. By leveraging duality theory, alternative system parameterizations, and matrix inequalities, researchers transform these seemingly impossible minimax problems into tractable convex optimization problems—specifically Semidefinite Programs (SDPs)—that can be solved globally and efficiently in polynomial time.
Flattening the Min-Max via Duality
The primary obstacle in robust control is the inner maximization. If we must guarantee performance against an infinite set of potential noise distributions (known as an ambiguity set), evaluating the worst-case scenario directly is computationally prohibitive.
To bypass this, researchers utilize Distributionally Robust Optimization (DRO) combined with optimization duality. The core intuition is that maximizing a linear objective over a constrained set of probability distributions is dual to a deterministic minimization problem with a finite number of constraints.
Intractable Min-Max Problem
min_u max_P E_P[Cost] (Over infinite-dimensional distributions)
│
▼ [Strong Duality / Conic Representation]
Tractable Convex Program
min_u,λ Subject to Linear Matrix Inequalities (LMIs)
For instance, Peter Coppens et al. demonstrated this in Data-driven distributionally robust LQR with multiplicative noise. Instead of assuming exact knowledge of noise statistics, they optimize against an entire family of distributions. Through duality, they prove that this infinite-dimensional minimax problem collapses into a standard, finite-dimensional SDP.
This approach becomes even more powerful when dealing with real-time constraints. Peter Coppens extended this concept to Model Predictive Control (MPC) by expressing risk measures as conic constraints in Data-Driven Distributionally Robust MPC for Constrained Stochastic Systems. By mapping risk directly to conic geometry, the author ensures that the controller can update its uncertainty model dynamically from data while maintaining strict safety guarantees.
Taming Wasserstein Ambiguity Sets
When data is scarce, we often define the ambiguity set as a "ball" around an empirical distribution using the Wasserstein metric. To make this tractable, researchers use mathematical approximations to map these statistical distances into geometric constraints.
In Distributionally Robust Density Control with Wasserstein Ambiguity Sets, Joshua Pilipovsky tackles the problem of steering an entire state distribution under severe, non-Gaussian disturbances. By using the Gelbrich distance as an outer approximation of the Wasserstein metric, the author flattens the distribution-level maximization into a clean set of matrix inequalities, enabling real-time robust density control. Similarly, Jean-Sébastien Brouillon et al. leverage Wasserstein ambiguity sets in Distributionally Robust Infinite-Horizon Control: From a Pool of Samples to the Design of Dependable Controllers, proving that quadratic control objectives over these sets can be solved via finite-dimensional SDPs.
Changing Coordinates: System-Level and Covariance Parameterizations
Even if the inner maximization over uncertainty is resolved, the outer optimization over the controller parameters often remains non-convex. This is because the system's states and control inputs interact multiplicatively over time, creating non-linear dependencies.
To restore convexity, researchers change the "coordinates" of the optimization problem. Instead of optimizing the controller gains directly, they optimize over the closed-loop trajectories or system responses themselves.
System Level Synthesis (SLS)
Under the System Level Synthesis framework, the decision variables become the transfer functions that map disturbances to states and inputs. Because the system dynamics constrain these transfer functions linearly, the resulting optimization space is perfectly convex.
Traditional Control Design (Non-Convex)
Optimize: Controller Gain (K)
State Equation: x_t+1 = (A + B*K)x_t + w_t <-- Multiplicative interaction
System Level Synthesis (Convex)
Optimize: Closed-loop responses (Φ_x, Φ_u)
Constraint: [I -A, -B] [Φ_x; Φ_u] = I <-- Linear constraint
This parameterization has unlocked solutions to complex performance metrics like regret minimization, where a controller seeks to minimize its suboptimality relative to an ideal, omniscient controller that can see the future. In A System Level Approach to Regret Optimal Control, Alexandre Didier et al. used SLS to transform the non-convex quadratic constraints of dynamic regret into a convex SDP, allowing engineers to enforce real-world state and input constraints on regret-optimal systems.
This coordinate shift is equally potent for state estimation. Jean-Sébastien Brouillon and Giancarlo Ferrari‐Trecate applied SLS to observer design in Minimal regret state estimation of time-varying systems, turning the tracking of time-varying systems into an efficiently solvable convex program.
Covariance Steering
Another powerful parameterization focuses on steering the state covariance matrix directly. In Computationally Efficient Covariance Steering for Systems Subject to Parametric Disturbances and Chance Constraints, Jacob W. Knaup demonstrates that by treating the state covariance as the primary optimization variable, complex probabilistic "chance constraints" (e.g., ensuring a robot has a 99% probability of avoiding an obstacle) can be mapped into tractable convex constraints. This allows for direct control over system uncertainty without relying on expensive Monte Carlo simulations.
Taming Dynamic Uncertainties via Matrix Inequalities
In many practical scenarios, uncertainty cannot be modeled simply as additive noise. Instead, it manifests as dynamic unmodeled behavior—such as structural flexibility in an aircraft wing or delayed actuator responses.
To handle these structured, dynamic uncertainties, researchers use Integral Quadratic Constraints (IQCs). IQCs bound the energy of the uncertainty's input and output signals using quadratic forms.
The mathematical bridge that brings IQCs into the realm of convex optimization is the S-procedure. The S-procedure is a theorem that provides a clean condition under which a set of quadratic inequalities implies another quadratic inequality. It allows an "all-quantified" statement—such as "for all possible dynamic uncertainties, the system remains stable"—to be replaced by a single Linear Matrix Inequality (LMI).
In Output-feedback model predictive control under dynamic uncertainties using integral quadratic constraints, Lukas Schwenkel et al. leveraged this mechanism to design robust output-feedback controllers. By embedding IQCs into an MPC framework and applying the S-procedure, the authors convexified the robust estimation and control loops simultaneously. This ensures that the controller and estimator parameters can be computed reliably at every time step, guaranteeing recursive feasibility and stability despite unmodeled system dynamics.
Broad Applications of Convex Reformulation
The strategy of convexification is not isolated to a single control paradigm; it serves as a universal translator across diverse engineering domains:
- Adaptive Control & Learning: In Learning Linear-Quadratic Regulators Efficiently with only √T Regret, Alon Cohen et al. solved a long-standing open question in adaptive control by reformulating the LQR exploration problem as an SDP, achieving the theoretical optimal regret bound efficiently.
- Sensor Fusion: W. Chen et al. introduced a moment-constrained marginal distributionally robust Kalman filter in A Marginal Distributionally Robust Kalman Filter for Sensor Fusion, converting a minimax sensor-coordination problem into a stable, convex optimization problem to handle unknown sensor correlations.
- Aerospace & Energy Systems: From the fuel-optimal space rendezvous algorithms developed by Denis Arzelier et al. in Linearized Impulsive Fixed-Time Fuel-Optimal Space rendezvous: A New Numerical Approach, to the wind turbine power smoothing designed by Tobias Gybel Hovgaard and John Bagterp Jørgensen in Model predictive control for wind power gradients, the recipe remains consistent.
By identifying the underlying algebraic structure of the system and applying duality or reparameterization, non-convex physical constraints are systematically mapped into equivalent convex forms.
Conclusion
Convex reformulation fundamentally changes our approach to robust estimation and control. Instead of abandoning mathematical rigor for heuristics when faced with non-convexity, this research direction proves that many worst-case optimization problems possess hidden convex structures. By exposing these structures through duality, System Level Synthesis, and matrix inequalities, the control community continues to turn theoretically intractable safety problems into practical, deployable algorithms.
Key Foundational Work
Linearized Impulsive Fixed-Time Fuel-Optimal Space rendezvous: A New Numerical Approach (2016) — contributed 2 concept elements to this direction.
Model predictive control for wind power gradients (2014) — contributed 1 concept elements to this direction.