Online Feedback Optimization (OFO)

• Online Feedback Optimization (OFO) is a control paradigm that integrates optimization algorithms into the feedback loop to steer systems toward optimal steady states using real-time measurements.
• OFO replaces offline computed optimizers by embedding gradient-based or quadratic programming controllers that adjust inputs dynamically to track both static and time-varying objectives.
• OFO has been validated across domains—from power systems to social networks—demonstrating its capability to merge measurement-driven control with adaptive estimation for enhanced performance and stability.

Online Feedback Optimization (OFO) is a control paradigm in which an optimization algorithm is embedded directly in the feedback loop of a dynamical system so as to steer the plant to an optimal steady state, or to track the optimizer of a time-varying steady-state problem, by using real-time measurements rather than a full plant model. In the canonical formulation, the plant has dynamics

$\dot x = f(x,u), \qquad y = g(x),$

with a well-defined steady-state input–output map $y=h(u)$, and OFO targets the steady-state program

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$

Its central feature is that the optimizer is not run offline and then handed to a tracking controller; instead, the optimization dynamics themselves act as the controller, with measured outputs replacing unknown steady-state evaluations (Bianchi et al., 2024).

1. Formal problem statement and core mathematical structure

A standard OFO setup assumes that for each constant input $u$, the plant is globally asymptotically stable, admits a unique equilibrium $x=s(u)$, and induces a continuously differentiable steady-state output map $h(u)=g(s(u))$ with locally Lipschitz sensitivity $\nabla h(u)$. The steady-state objective is

$$\min_{u\in\mathbb{R}^m,\;y\in\mathbb{R}^p}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u),$$

which can be reduced to the unconstrained problem

$$\min_{u\in\mathbb{R}^m}\ \tilde\Phi(u), \qquad \tilde\Phi(u):=\Phi(u,h(u)).$$

The optimal steady state is the pair $(u^\star,y^\star)$ with $y=h(u)$0 solving the reduced problem, $y=h(u)$1, and $y=h(u)$2 (Bianchi et al., 2024).

The reduced gradient is obtained by the chain rule,

$y=h(u)$3

so a nominal open-loop gradient flow would be

$y=h(u)$4

OFO replaces the unavailable steady-state quantity $y=h(u)$5 by the measured output $y=h(u)$6 and closes the loop through

$y=h(u)$7

yielding the interconnected system

$y=h(u)$8

In this sense, OFO differs both from classical setpoint optimization, in which $y=h(u)$9 is computed offline, and from standard feedback control, which stabilizes references without explicitly solving a steady-state optimization problem (Bianchi et al., 2024).

The same principle appears in discrete-time settings. For time-varying AC-OPF tracking, a projected-gradient OFO law is written as

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$0

where $$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$1 is an input–output sensitivity matrix and $$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$2 is the measured grid output. This turns a one-step optimization iterate into a feedback controller that tracks the moving OPF solution in real time (Picallo et al., 2021).

2. Algorithmic realizations and controller architectures

The most basic OFO controllers are gradient or projected-gradient flows. When input constraints $$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$3 are present, smooth projected dynamics take forms such as

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$4

and nonsmooth variants based on tangent-cone projection are also used:

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$5

These formulations keep the optimizer in the loop while enforcing feasibility of the plant input (Bianchi et al., 2024).

A second major class uses projected quadratic programs. In real-time curative control for transmission grids, the OFO update is

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$6

where $$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$7 is the solution of a convex QP that keeps the future input and a linearized prediction of the future output inside admissible sets. The unconstrained descent direction is

$$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$8

and the QP computes the closest feasible direction under input and output constraints. This architecture is expressly designed for nonconvex physical tasks while preserving the interpretation of optimization as feedback (Ortmann et al., 2022).

A related QP-based OFO appears in time-varying AC-OPF with cross-layer estimation. There, the controller operator $$\min_{u,y}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u).$$9 solves a convex QP built from first-order approximations of cost and constraints, box constraints on controllable injections, and a positive-definite regularization matrix $u$0. The measured or estimated system state replaces exact model evaluation, so the QP is solved around the current operating point rather than around an offline forecast (Picallo et al., 2021).

Primal–dual OFO is another recurrent realization. In voltage regulation, one introduces dual variables for upper and lower voltage bounds and updates them from measured voltages, while reactive power setpoints are updated by projected primal steps. The optimizer therefore consists of coupled primal and dual dynamics running continuously around the grid, rather than a centralized OPF solved to completion at each disturbance realization (Zhan et al., 2024).

3. Stability theory, timescale separation, and its removal

A large part of the OFO literature analyzes the plant–optimizer interconnection as a two-timescale system. In contraction-based singular perturbation formulations, the plant is the fast subsystem and the optimizer is the slow subsystem, with controller dynamics scaled by a small parameter $u$1. For linear time-invariant plants

$u$2

the reduced optimization flow is obtained by replacing the plant state with its steady state $u$3, and explicit tracking error bounds are derived when $u$4 is sufficiently small. In this regime, the admissible controller speed is coupled to the plant contraction rate, and stability relies on the optimizer evolving more slowly than the plant (Cothren et al., 2023).

The principal limitation of that analysis is that small controller gains improve stability margins at the expense of transient performance and responsiveness. This issue is treated directly by a stability theory that removes timescale separation. For the continuous-time OFO system

$u$5

a sufficient condition is formulated through Lyapunov-like functions $u$6 and $u$7 satisfying coupled dissipation inequalities,

$u$8

together with a dominance condition

$u$9

Using the composite Lyapunov function

$x=s(u)$0

one obtains exponential decay for any $x=s(u)$1. Under Assumptions 1, 2, and 3 in that framework, the closed-loop system is globally exponentially stable and converges to the unique optimal steady state, with a stability condition that is independent of the plant time constant and therefore scaling-invariant (Bianchi et al., 2024).

A parallel route to removing timescale separation is developed for monotone systems. For plants

$x=s(u)$2

with monotone dynamics and a well-defined steady-state map $x=s(u)$3, the OFO controller

$x=s(u)$4

is analyzed through a small-gain theorem for monotone systems. The sufficient conditions depend on the steady-state map and strong-convexity/Lipschitz properties of the induced optimization problem, not on transient plant dynamics. The resulting convergence theorem guarantees that the closed-loop converges to the unique optimizer for any gain $x=s(u)$5 (Bianchi et al., 19 Jun 2025).

These two lines of work establish a common point: the need for a slow optimizer is not fundamental. One approach uses a max-type composite Lyapunov function; the other uses monotone small-gain arguments. In both cases, the stabilizing conditions are expressed in terms of structural inequalities rather than controller–plant timescale separation (Bianchi et al., 2024).

4. Adaptive, estimation-based, and model-free OFO

Many OFO designs require only steady-state sensitivities rather than a full dynamical model, which has motivated adaptive and model-free variants. In adaptive real-time grid operation, the sensitivity matrix

$x=s(u)$6

is treated as an unknown time-varying quantity, vectorized as $x=s(u)$7, and estimated online through a Kalman filter / recursive least-squares scheme. The process model is

$x=s(u)$8

and the measurement equation is derived from

$x=s(u)$9

The OFO controller then uses $h(u)=g(s(u))$0 instead of a fixed sensitivity:

$h(u)=g(s(u))$1

Under persistent excitation and strong monotonicity/Lipschitz assumptions, the sensitivity estimate converges in expectation and the input sequence tracks the time-varying AC-OPF optimizer with an explicit bound that depends on excitation amplitude, optimizer drift, and steady-state estimation error (Picallo et al., 2021).

A related cross-layer design combines OFO with dynamic estimation and an online power-flow solver. Disturbances are estimated recursively from measurements, and the power-flow state is updated with a sensitivity-conditioned Newton-like step,

$h(u)=g(s(u))$2

so that only one prediction–correction step is carried out per sampling instant. The OFO controller then solves a regularized QP using the estimated state rather than an exact AC power-flow solve at each step (Picallo et al., 2021).

Model-free gradient learning also appears in OFO with persistency of excitation. A recent construction designs perturbations through a bilevel optimization program. The lower level computes a perturbation $h(u)=g(s(u))$3 that satisfies a full-rank condition on the recent input increments while minimizing

$h(u)=g(s(u))$4

so that excitation is generated with minimal deviation from the descent direction. In the reported case study, this persistently exciting OFO achieved the same profit as OFO with random input perturbations, and $h(u)=g(s(u))$5 higher profit than OFO without input perturbations (Gude et al., 26 May 2025).

These developments expand OFO from “measurement-based” to “measurement-driven and self-calibrating.” A plausible implication is that the practical boundary between model-based real-time optimization and adaptive feedback control is increasingly determined by how sensitivities are learned and regularized online, rather than by whether a first-principles model exists.

5. Distributed, networked, and multi-agent OFO

When the plant consists of interconnected subsystems, centralized OFO can become impractical because of communication bottlenecks and global sensitivity computation. A fully distributed model-free OFO over networks addresses this by combining stochastic zeroth-order gradient estimates with consensus-based tracking of the global objective value. For the steady-state problem

$h(u)=g(s(u))$6

each agent uses local measurements, random perturbations, and neighbor communication to approximate a global gradient, and then performs a projected local update. The resulting error bounds decompose in terms of network size, the number of consensus iterations, and consensus accuracy; in the unconstrained case, the expected distance to the optimizer decays linearly up to a residual term determined by gradient smoothing and consensus error (Wang et al., 2024).

A different distributed construction appears in real-time distribution-system voltage regulation. There, the outer loop performs a scaled primal–dual update using the sparsity of $h(u)=g(s(u))$7, while an inner feedback loop solves the non-Euclidean projection associated with the $h(u)=g(s(u))$8-norm by using voltage differences as gradient surrogates:

$h(u)=g(s(u))$9

With short-range communication between physical neighbours, this nested strategy converges under the linearized DistFlow model and achieves effective distributed voltage regulation (Zhan et al., 2024).

In multi-area transmission grids, OFO takes the form of online feedback equilibrium seeking rather than centralized optimization. Each area updates its local controls against a cost $\nabla h(u)$0 while the areas are dynamically coupled through the shared network, so the stacked pseudo-gradient defines a variational inequality whose solutions are Nash equilibria. Under convexity and cocoercivity assumptions, the projected feedback iteration converges to a Nash equilibrium of the multi-area congestion-management game, highlighting the distinction between decentralized stability and centralized social optimality (Belgioioso et al., 2023).

The same OFO logic has also been extended to networked socio-technical systems. In a recommender-system formulation, the plant is a social network with opinion dynamics, recommendations are the control input, and the objective combines click-through-rate and polarization:

$\nabla h(u)$1

Perception modules estimate opinions from click data, a Kalman filter estimates the sensitivity $\nabla h(u)$2, and a projected-gradient OFO controller updates recommendation positions. The numerical study reports that the network-aware OFO achieves approximately $\nabla h(u)$3 higher final mean CTR than a naive OFO baseline and approximately $\nabla h(u)$4 lower polarization cost (Chandrasekaran et al., 2024).

6. Applications, empirical demonstrations, and performance

Power systems remain the dominant application area. In reactive power flow optimization for a real Swiss distribution grid, OFO was deployed with off-the-shelf hardware and software, sending discrete power-factor setpoints to 16 PV inverters. The controller ran once per minute using measured voltages and substation reactive power, interacted with existing SCADA and legacy inverter logic, and was reported as a successful 24/7 deployment in an operational environment corresponding to Technology Readiness Level 7 (Ortmann et al., 2023).

In distribution-grid enhancement via coordinated Volt/VAr control, OFO was compared against droop-based methods on a low-voltage CIGRÉ grid and in a real feeder experiment. The year-long numerical study suggested that OFO can enable another $\nabla h(u)$5 of maximum active power injections compared to droop control, and the field experiment reported another $\nabla h(u)$6 enhancement compared to droop, while using only voltage magnitude measurements, minimal model knowledge, and communication with reactive power sources (Matt et al., 2023).

For real-time curative actions in transmission grids, OFO was used after a contingency to reduce the voltage difference across an open breaker so that reclosing becomes feasible. The controller sampled every 5 seconds, updated generator active power and voltage setpoints from measurements of voltages, line flows, and angle difference, and in the reported IEEE 39-bus scenario reduced the breaker mismatch within a few iterations while keeping setpoints inside the specified bounds (Ortmann et al., 2022).

OFO has also been studied as a flexibility controller in high-voltage distribution and sub-transmission grids. In simulations of transitions to vertices of the Feasible Operating Region, stability was found to be sensitive to gain and sensitivity approximations: with $\nabla h(u)$7, trajectories to all 36 vertices converged stably, whereas with $\nabla h(u)$8 some trajectories violated constraints or exhibited oscillatory behaviour. This work emphasizes that even when OFO is operationally attractive, parameter selection remains decisive for full-range safe operation (Klein-Helmkamp et al., 16 Jun 2025).

Outside power systems, OFO has been tuned for pressure control in centrifugal compressors by jointly optimizing the effective step size $\nabla h(u)$9 and the sampling time $$\min_{u\in\mathbb{R}^m,\;y\in\mathbb{R}^p}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u),$$0. The reported validation showed that simultaneous tuning of the OFO parameters and sampling time yields up to $$\min_{u\in\mathbb{R}^m,\;y\in\mathbb{R}^p}\ \Phi(u,y)\quad \text{s.t.}\quad y=h(u),$$1 better tracking performance than manual tuning based on steady state, and reduced the integral tracking error markedly for both step and sinusoidal validation trajectories (Zagorowska et al., 2023).

The breadth of application is therefore not restricted to electrical networks. OFO has been instantiated in social recommendation systems, gene-expression models, and compressor control, while preserving the same structural idea: optimization is performed through the physical process rather than around it (Chandrasekaran et al., 2024).

7. Limitations, misconceptions, and current research directions

A common misconception is that OFO is merely “online optimization with measurements.” In the cited literature, it is more specific: the optimization algorithm is the controller, the plant is part of the algorithmic loop, and convergence or tracking properties are statements about the closed-loop interconnection rather than about an optimizer run in isolation (Bianchi et al., 2024).

Another misconception is that measurement feedback alone eliminates all stability concerns. Several works show the opposite. Emergency-operation studies explicitly note that formal closed-loop stability for nonlinear power-system dynamics with few-second updates remains open, and that classical timescale-separation results can be too conservative for time-critical applications (Ortmann et al., 2022). Flexibility studies similarly show that improper tuning can lead to oscillatory or unstable behavior even when the controller structure is theoretically well motivated (Klein-Helmkamp et al., 16 Jun 2025).

The current limitations stated across the literature are comparatively consistent. Existing analyses often assume smooth costs and strong convexity, while extensions to nonconvex or merely convex costs are nontrivial; output constraints can require more delicate handling than box-constrained inputs; and many continuous-time guarantees do not immediately transfer to sampled-data or discrete-time implementations (Bianchi et al., 2024). In adaptive and distributed variants, persistent excitation, communication delays, missing data, and measurement noise remain major practical issues (Picallo et al., 2021).

Open research directions are likewise explicit. One line aims to design OFO algorithms that satisfy scaling-invariant stability conditions under weaker assumptions, potentially through regularization, adaptive gains, or second-order information (Bianchi et al., 2024). Another studies fully distributed or networked OFO with stronger robustness guarantees against limited communication and consensus error (Wang et al., 2024). In power systems, joint active–reactive control, unbalanced and meshed networks, and integration with legacy control hierarchies remain active topics (Zhan et al., 2024). In socio-technical settings, partial observability and learned perception models raise further questions about gradient quality, exploration, and local versus global optimality (Chandrasekaran et al., 2024).

Taken together, these works portray OFO as a unifying framework for steady-state optimal control under feedback, rather than a single algorithm. Its defining problem is the same across domains: how to use measurements, sensitivities, and feasible optimization dynamics to make a plant converge to, or track, an optimal operating point without requiring exact plant models. The main research frontier is no longer whether such interconnections can work in principle, but how broadly they can be stabilized, decentralized, and adapted without sacrificing responsiveness or constraint satisfaction.