When we run an iterative optimization algorithm like gradient descent, we update a state variable step by step: $x_{k+1} = x_k - \alpha \nabla f(x_k)$. To an optimization theorist, this is an algorithmic sequence designed to minimize a cost function. To a control engineer, however, this looks exactly like a discrete-time dynamical system where the algorithm's state updates act as a physical plant, and the gradient acts as a feedback controller.
This conceptual bridge reveals a profound duality. We can analyze optimization algorithms using robust control theory, and conversely, we can deploy optimization algorithms directly as real-time feedback controllers for physical systems.
Analyzing Algorithms via Robust Control
Proving that an optimization algorithm converges typically requires hand-crafting a "Lyapunov function" (or energy function) that decreases at every iteration. For complex or accelerated algorithms, finding this function is notoriously difficult and often relies on algebraic intuition.
By framing optimization as a closed-loop control system, we can automate this process. We split the algorithm into two blocks:
- A Linear Time-Invariant (LTI) System: This represents the fixed algorithmic structure (such as momentum buffers or step-size rules).
- A Nonlinear Feedback Block: This represents the gradient or subdifferential of the objective function.
+---------------------------------+
-----> | Linear Algorithm State | ----> State (x_k)
| (e.g., Momentum, Step Sizes) | |
+---------------------------------+ |
^ v
| +---------------+
+------------ | Gradient / |
| Function Oracle|
+---------------+
Because the objective function satisfies certain geometric properties (like $L$-smoothness or strong convexity), we can bound the behavior of the nonlinear gradient block using Integral Quadratic Constraints (IQCs)—a standard tool from robust control used to analyze unmodeled or complex nonlinearities.
As demonstrated by Robert Nishihara et al. in their analysis of the Alternating Direction Method of Multipliers (ADMM), this separation allows researchers to translate convergence proofs into Semidefinite Programs (SDPs). Instead of inventing a proof by hand, a researcher can solve a small numerical optimization problem to automatically certify an algorithm's linear convergence rate and select optimal parameters.
This framework has proven remarkably versatile across different optimization paradigms:
- Nonsmooth and Non-strongly Convex Problems: Ribeiro Alejandro et al. extended IQC analysis to non-strongly convex functions, constructing time-varying Lyapunov functions to systematically certify subexponential convergence rates.
- Multi-Operator Splitting: Han Wang et al. modeled the Three-Operator Splitting (TOS) algorithm as an LTI system with multiple nonlinear feedback loops, deriving automated convergence bounds via matrix inequalities.
- Multi-Player Games: Moving beyond single-objective minimization, Guodong Zhang et al. applied IQCs to analyze first-order methods in smooth, strongly monotone games (such as GANs), revealing why certain accelerated methods fail in stochastic game environments.
- Nonsmooth Composite Optimization: Dongsheng Ding et al. established tight, worst-case exponential convergence rates for proximal augmented Lagrangian dynamics by constructing quadratic Lyapunov functions through an interconnected control lens.
- Stochastic Variance Reduction: By leveraging dissipativity theory—a control-theoretic framework centered on energy conservation—Bin Hu et al. unified the analysis of stochastic algorithms like SVRG and Katyusha, converting variance-reduction guarantees into solvable SDPs.
Even high-order tensor methods benefit from this perspective. Tianyi Lin et al. showed that the complex algebraic conditions required for discrete accelerated tensor methods are actually discrete approximations of continuous-time closed-loop feedback laws that dynamically link step sizes to gradient magnitudes.
Online Feedback Optimization
If optimization algorithms can be modeled as controllers, can we use them to control physical systems in real time? This is the core idea behind Online Feedback Optimization (OFO).
In traditional Model Predictive Control (MPC), a computer uses a high-fidelity mathematical model of a physical system to simulate future outcomes and solve an optimization problem at every time step. This approach struggles if the system is highly nonlinear, or if the environment changes unexpectedly.
OFO flips this paradigm. Instead of running an optimization algorithm to completion inside a computer using a model, the optimization algorithm is run directly on the physical system, one iteration at a time.
+--------------------------------------------------------+
| Online Optimization Loop |
| |
| +--------------------+ Control Input (u_k) |
| | Gradient Update | ----------------------------+ |
| | (Algorithmic Step) | | |
| +--------------------+ v |
| ^ +------------+
| | | Physical |
| +------------------------------ | Plant |
| Measured Output (y_k) +------------+
+--------------------------------------------------------+
In this setup, the physical plant acts as its own computational oracle. The algorithm applies a single gradient step as a physical control input ($u_k$). The physical system naturally responds according to the laws of physics. Sensors measure the actual output ($y_k$), and this real-time measurement is fed back to compute the next gradient step.
This approach works beautifully because the physical system inherently embodies its own exact physics, bypassing the need for a perfect predictive model. Lukas Ortmann et al. provided the first real-world experimental validation of this concept by using OFO to regulate voltages in a power distribution grid, achieving optimal performance using only a rudimentary grid model and noisy sensors. In subsequent work, Lukas Ortmann et al. demonstrated that a constraint-based OFO controller could successfully coordinate distributed energy resources (like solar inverters and batteries) to mitigate sudden grid disturbances from electric vehicle charging.
To make this mathematically rigorous, Gianluca Bianchin et al. combined projected primal-dual gradient flows with singular perturbation theory. They proved that when an optimization algorithm acts directly as a feedback controller, the closed-loop system can track moving optimal trajectories while strictly respecting hardware and safety constraints. Furthermore, Saverio Bolognani et al. developed a completely model-free nonlinear OFO framework that estimates gradients directly from system outputs, ensuring stable closed-loop operation even when the underlying system dynamics are entirely unknown or time-varying.
Blurring the Line Between Solver and Plant
When an optimization algorithm is embedded into a physical loop, we can no longer treat the solver as an instantaneous black box. The solver's internal states and the plant's physical states evolve together as a single, interconnected dynamical system.
This interconnected view is essential when dealing with hardware limitations. For example, in high-speed applications, an optimization solver might not have enough time to run to convergence before the next control action is required. Anusha Srikanthan et al. analyzed this scenario in ADMM-based linear MPC. By modeling the joint evolution of the plant and the incomplete solver states, they proved that the system remains stable and satisfies constraints even when the solver is stopped early after a fixed, finite number of iterations, provided a warm-start strategy is maintained.
Similarly, Dominic Liao-McPherson et al. leveraged forward-backward splitting and constraint tightening within an iterative learning control framework. They demonstrated that injecting real-world error measurements directly into the gradient steps allows high-precision robots to safely adapt to non-repetitive noise and model mismatches from trial to trial.
Beyond Physical Hardware
The utility of treating optimization as a controlled dynamical system extends well beyond physical engineering into purely algorithmic domains:
- Meta-Optimization: Tuning hyperparameters like learning rates or momentum coefficients is notoriously difficult due to non-convex loss landscapes. Xinyi Chen et al. reframed meta-optimization as a disturbance-feedback control problem, treating past gradients as external disturbances to dynamically optimize the hyperparameters of first-order methods.
- Concept Drift: In non-stationary data environments, machine learning models face performance degradation as data distributions shift. Tianyu Pan et al. designed a Lyapunov-stable adaptive control framework that treats the training process as a controlled system, dynamically adjusting learning rates and fusion weights to keep prediction errors bounded during multimodal concept drift.
- Distributed Optimization: In large-scale machine learning paradigms like Federated Learning, algorithms must contend with gradient clipping, differential privacy, and communication compression. Mingyi Hong et al. unified these diverse distributed algorithms under a single stochastic multi-rate feedback control framework, allowing algorithmic modifications to be analyzed as standard control components with predictable impacts on stability.
Ultimately, whether the goal is to automate a convergence proof for a machine learning optimizer or to regulate a power grid in real time, treating optimization as a feedback control system replaces ad-hoc heuristics with systematic, control-theoretic guarantees.
Key Foundational Work
A General Analysis of the Convergence of ADMM (2015) — contributed 2 concept elements to this direction.
Analysis and Design of Optimization Algorithms via Integral Quadratic Constraints (2016) — contributed 1 concept elements to this direction.
Optimal control of fast and high-fidelity quantum gates with electron and nuclear spins of a nitrogen-vacancy center in diamond (2015) — contributed 1 concept elements to this direction.