In traditional control engineering, optimization and physical dynamics are treated as separate worlds. If you want to optimize a power grid or a chemical plant, the standard "timescale separation" rule dictates that your optimizer must be much slower than the system it controls. You wait for the plant to reach a steady state, take a measurement, run an optimization algorithm to convergence, and then apply the new setpoint.
This "wait-and-see" approach is safe but sluggish. Florian Dorfler and Mattia Bianchi have recently shown that this separation is not a mathematical necessity, but a self-imposed limitation. By treating optimization algorithms as dynamical systems rather than static solvers, researchers are proving that we can run optimization and physical control at the exact same speed.
Optimization as a Dynamical System
The core intuition behind this shift is to stop viewing an optimization algorithm (like Gradient Descent) as a black box that produces an answer. Instead, view it as a differential equation—a dynamical system where the "state" is the current guess of the optimal solution.
When you connect this "optimizer-system" to a "physical-system" (the plant), you get two interconnected dynamical systems. As Jaap Eising describes in On analysis of open optimization algorithms, these are "open" systems that constantly exchange information. The plant provides real-time measurements, and the optimizer provides a continuous stream of control updates.
The Small-Gain Stability Condition
If both systems are moving at the same time, how do we prevent them from oscillating out of control? The answer lies in the Small-Gain Theorem.
In Stability and Robustness of Distributed Suboptimal Model Predictive Control, Giuseppe Belgioioso and colleagues explain that stability is guaranteed if the "gain" (the sensitivity) of the loop is less than one. Specifically, if the optimizer converges toward the optimum faster than the plant's state changes can "disturb" the optimization landscape, the whole closed-loop system remains stable.
This allows for Suboptimal MPC, where you don't wait for the optimizer to find the perfect answer. You just take one step of the algorithm, apply it, and move on. Yujia Yang and Chris Manzie have applied this to multi-agent systems with limited communication, showing that even with "noisy" or incomplete optimization steps, the physical system can still be steered safely.
Breaking the Timescale Barrier
The most significant breakthrough in this area is the removal of the requirement that the optimizer must be "fast enough." Florian Dorfler and Mattia Bianchi developed a "max-type" Lyapunov function in A Stability Condition for Online Feedback Optimization without Timescale Separation that proves stability even when the controller and the plant operate on identical timescales.
The intuition is a parameter dominance condition: as long as the internal "self-stabilizing" forces of the plant and the optimizer are stronger than the "interference" they cause each other, the system will converge. This allows for incredibly responsive control. For example, Marta Zagorowska and Alisa Rupenyan demonstrated that by ignoring timescale separation, they could improve the tracking performance of centrifugal compressors by up to 87% (Tuning of Online Feedback Optimization for setpoint tracking in centrifugal compressors).
Model-Free and Robust Optimization
This framework is also remarkably robust to model errors. Because the optimizer is constantly receiving feedback from the actual physical plant, it doesn't need a perfect internal model. Zhiyu He and Saverio Bolognani showed in Model-Free Nonlinear Feedback Optimization that you can drive complex nonlinear systems to their optimal operating points using only zeroth-order (measurement-based) information.
By treating the optimizer as a living part of the system's dynamics, we move from "calculating the best move" to "evolving toward the best state," enabling real-time optimization in environments that were previously too fast or too complex to handle.