Online Feedback Optimization

Updated 8 July 2026

Online Feedback Optimization is a closed-loop control paradigm that directly drives measured outputs toward solving steady-state constrained optimization problems using local sensitivity data.
It employs methods like scaled projected gradient descent and primal-dual dynamics to adaptively update inputs and manage uncertainties without requiring a complete process model.
Practical applications in process control and power systems demonstrate OFO's ability to improve convergence speed, reduce tuning iterations, and ensure robust constraint satisfaction.

Online Feedback Optimization (OFO) is a control paradigm that embeds a steady-state optimization problem into a real-time feedback controller. Rather than first identifying a detailed process model and then solving an open-loop optimal control problem, OFO directly drives measured plant outputs toward the solution of a constrained optimization by iteratively updating manipulated inputs from measurements and local sensitivity information. In its standard form, OFO targets problems such as

$\min_{u,y}\;\Phi(u,y)\quad\text{s.t.}\quad y=h(u),\;Au\le b,\;Cy\le d,$

and is motivated by settings with limited model availability, online measurements, and hard input/output constraints (Zagorowska et al., 14 Apr 2026). Closely related formulations use measurement-feedback primal-dual dynamics to track time-varying convex optima of dynamical or networked systems, emphasizing dynamic regret, tracking, and robustness to exogenous disturbances (Bernstein et al., 2018).

1. Conceptual scope and operating assumptions

The defining feature of OFO is the replacement of a purely feedforward optimizer by a closed-loop optimization algorithm that interacts continuously or iteratively with the physical system. In the steady-state setting, the plant is represented by an input-output map $y=h(u)$ , the controller measures $y^k$ , uses local sensitivity information such as $\nabla h(u^k)$ , computes an optimization step, and applies the updated input to the plant. The appeal is strongest when only bounds, online measurements, and local sensitivities are available, while a full model is unavailable or too inaccurate for reliable open-loop optimization (Zagorowska et al., 14 Apr 2026).

Most classical formulations assume that the plant approximately reaches steady state between controller iterations, or that a sampling interval $\Delta T$ induces a suitable timescale separation. Under that assumption, no explicit state $x$ is required in the optimization layer; the controller can operate directly on the measured steady-state outputs. This steady-state viewpoint underlies process-control applications such as gas lift and continuously-stirred tank reactor tuning, as well as power-system applications in distribution and subtransmission grids (Zagorowska et al., 14 Apr 2026).

A broader strand of literature extends the same feedback principle to dynamical settings in which the optimization problem itself varies over time. In those formulations, the optimizer is not merely solving a static steady-state problem repeatedly, but tracking a moving solution trajectory through measurement feedback. This is the setting of feedback-based online primal-dual methods and continuous-time primal-dual controllers for time-varying convex programs over networked systems (Bernstein et al., 2018).

2. Mathematical structure and update laws

A canonical discrete-time OFO controller based on scaled projected gradient descent computes, at iteration $k$ , a search direction by solving a small quadratic program:

$w^k = \arg\min_{w\in\mathbb{R}^{n_u}} \left\|\,w + S^k H(u^k)^\top\nabla\Phi(u^k,y^k)^\top\right\|^2_{(S^k)^{-1}}$

subject to linearized input and output constraints, and then updates the manipulated variables according to

$u^{k+1}=u^k+\alpha^k\,w^k,\qquad \alpha^k\in[\alpha_{\min},\alpha_{\max}],$

with

$H(u):=\begin{bmatrix}I_{n_u}&\nabla h(u)^\top\end{bmatrix}.$

Here, the scalar step size $y=h(u)$ 0 controls responsiveness versus stability, while the positive-definite scaling matrix $y=h(u)$ 1 acts as a pre-conditioner and changes both the direction and magnitude of the effective gradient step (Zagorowska et al., 14 Apr 2026).

An equivalent geometric interpretation is a projected gradient step onto a tangent cone or linearized feasible set. In formulations with unknown gradients, the descent direction is obtained by projecting the negative gradient onto the set of feasible directions that preserve input and output feasibility under a local linearization of $y=h(u)$ 2 (Gude et al., 26 May 2025). Power-system implementations often instantiate the same idea as a QP or MIQP with voltage, line-flow, power-injection, and actuator-rate constraints, using pre-computed sensitivities such as voltage-to-reactive-power or PTDF-like maps (Ortmann et al., 2022).

An alternative but closely related line of work uses primal-dual projected-gradient dynamics on a regularized Lagrangian, with the model output replaced by a real-time measurement $y=h(u)$ 3. In that setting, measurement feedback explicitly compensates model mismatch, avoids pervasive measurements of exogenous inputs, and naturally lends itself to distributed implementation over a communication graph (Bernstein et al., 2018).

3. Sensitivity estimation, tuning, and adaptation

Practical OFO performance depends strongly on tuning. Manual selection of step sizes and scaling matrices is difficult because the best values depend on unknown curvature, nonlinearities, and active constraints, while repeated plant experiments are time-consuming and laborious. Adaptive tuning therefore focuses on using online sensitivity information to update algorithm parameters directly inside the loop (Zagorowska et al., 14 Apr 2026).

For scaled projected-gradient OFO, one adaptive strategy computes the sensitivity of the objective $y=h(u)$ 4 with respect to the algorithm parameters. The sensitivity with respect to the scaling matrix is obtained through the chain rule and online differentiation of the QP solution, after which an auxiliary semidefinite program or a diagonal backtracking heuristic updates $y=h(u)$ 5 so as to enforce a local monotonic-descent condition. The step size is then adapted by fitting a local quadratic model of

$y=h(u)$ 6

and minimizing that quadratic over $y=h(u)$ 7. In the diagonal case, the operator-tunable parameters reduce to the five scalar bounds $y=h(u)$ 8 plus $y=h(u)$ 9, and no repeated plant experiments are required (Zagorowska et al., 14 Apr 2026).

When the plant Jacobian is not directly available, OFO must estimate sensitivities online. A standard approach uses recursive least squares and persistent excitation. The main difficulty is to guarantee a well-conditioned regression matrix without degrading the descent direction. A bilevel persistently exciting OFO controller addresses this by perturbing the nominal descent direction only if needed to restore full rank, replacing random Gaussian dithers by a minimal perturbation chosen through a lower-level convex QP (Gude et al., 26 May 2025).

Model adaptation can also be integrated at the plant-model level rather than at the optimization-parameter level. In compressor-station load sharing, Gaussian Process regression is used to learn the mismatch between nominal and true compressor efficiency maps, update the estimated power gradients, and feed the improved sensitivities back into the OFO loop. The reported effect is to reduce a power-consumption increase from $y^k$ 0 to $y^k$ 1 relative to a perfect-knowledge benchmark (Zagorowska et al., 2022).

Sensitivity analysis has also been used diagnostically. Closed-form expressions for $y^k$ 2, $y^k$ 3, and sensitivity to time-varying model mismatch show that early gradient errors affect many subsequent iterates, but that their marginal effect on $y^k$ 4 decreases for long operation times as later steps wash out earlier errors near the optimum (Zagorowska et al., 10 Mar 2025). This suggests a useful design principle: long-horizon OFO can tolerate imperfect online sensitivities more readily than short transient phases.

4. Stability, convergence, and tracking theory

The classical convergence picture is local and perturbation-based. For scaled projected-gradient OFO, existing results imply that under a sufficiently small constant step size and positive-definite scaling $y^k$ 5, the closed-loop dynamics converge to a KKT point of the underlying constrained steady-state optimization problem. In adaptive-tuning variants, local descent conditions and quadratic step-size models are intended to improve monotonicity and local convergence, while formal proofs of global convergence and constraint satisfaction under time-varying parameters remain ongoing work (Zagorowska et al., 14 Apr 2026).

For feedback-based online primal-dual methods on time-varying convex programs, the theory is sharper. Under suitable assumptions, analytical convergence claims are established in terms of dynamic regret, and when the controller is based on a regularized Lagrangian the tracking error converges Q-linearly to a neighborhood whose radius depends on measurement errors and optimizer drift (Bernstein et al., 2018). Continuous-time formulations for LTI plants use integral quadratic constraints and linear matrix inequalities to certify global exponential stability and bounded tracking error, with time-scale separation providing a sufficient route to feasibility of the stability conditions (Colombino et al., 2018).

A central theoretical issue is whether OFO truly requires the optimizer to evolve more slowly than the plant. Recent work addresses this directly. For monotone systems, OFO can achieve an optimal operating point regardless of the time constants of controller and plant; the sufficient conditions depend only on the steady-state behavior of the plant and are entirely independent of the transient dynamics (Bianchi et al., 19 Jun 2025). A complementary result proves a scaling-invariant stability condition without any timescale separation, based on a composite Lyapunov function defined as the $y^k$ 6 of plant-related and controller-related components (Bianchi et al., 2024).

Other extensions broaden the uncertainty model. In stochastic OFO for networks of non-compliant agents, the actual implemented input is a random variable depending on the commanded setpoint, and the resulting closed loop admits explicit mean-square error bounds under strong convexity, Lipschitzness, compactness, and bounded-noise assumptions (Lauand et al., 29 Aug 2025). In fully distributed, model-free OFO over networks, each agent runs local projected updates with stochastic gradient estimates and consensus-based tracking of the global objective; the analysis quantifies how performance depends on network size, iteration count, and consensus accuracy (Wang et al., 2024).

5. Representative application domains and reported results

The empirical literature is concentrated in process systems and electric-power applications, where measured outputs, operational limits, and real-time disturbances make closed-loop optimization attractive. A recurring pattern is that OFO replaces repeated nonlinear solves or brittle open-loop optimization by a sequence of small QPs, MIQPs, or projected-gradient steps embedded in the plant feedback loop.

Domain	OFO variant	Reported outcome
Gas lift and CSTR process control	Adaptive tuning of scaled projected gradient	Gas lift: manual tuning $y^k$ 7 iterations, adaptive OFO $y^k$ 8; CSTR: $y^k$ 9 lower mean squared error on tuning setpoint and $\nabla h(u^k)$ 0 lower error on validation setpoint (Zagorowska et al., 14 Apr 2026)
Curative distribution-grid flexibility	OFO dispatch of active/reactive power setpoints	A $\nabla h(u^k)$ 1 kW PCC-flexibility request was reached within $\nabla h(u^k)$ 2 iterations $\nabla h(u^k)$ 3 with steady-state error $\nabla h(u^k)$ 4 kW; EV charging transient $\nabla h(u^k)$ 5 kW, recovered within $\nabla h(u^k)$ 6– $\nabla h(u^k)$ 7 cycles (Klein-Helmkamp et al., 2023)
Real distribution-grid deployment	Reactive-power-flow OFO in 24/7 operation	Successful deployment in an operational environment corresponding to Technology Readiness Level (TRL) 7; convergence after a cost-function change was measured at $\nabla h(u^k)$ 8 min; no over/undervoltage occurred during tests (Ortmann et al., 2023)
Subtransmission-grid control	OFO with discrete taps and MIQP projection	$\nabla h(u^k)$ 9 curtailed at wind peak versus $\Delta T$ 0 with static rules; tracking errors $\Delta T$ 1 of limits; zero steady-state violations and $\Delta T$ 2 transient overshoot at worst (Ortmann et al., 2022)
Compressor-station load sharing	OFO with Gaussian Process model adaptation	Imperfect knowledge led to a $\Delta T$ 3 increase in power consumption; OFO with model adaptation reduced this increase to $\Delta T$ 4 (Zagorowska et al., 2022)

These case studies emphasize several themes. First, OFO is often deployed where only local sensitivities are available but constraint satisfaction on the real system is critical. Second, the online loop can compensate model mismatch and unmodeled disturbances by direct measurement feedback rather than by re-solving a detailed model-based optimization at every step. Third, discrete actuators, actuator-rate limits, and communication delays are not peripheral details but central elements of practical implementations, especially in power grids (Ortmann et al., 2022).

Beyond centralized plant control, OFO has also been extended to distributed and network-aware settings. In a network-aware recommender system based on Friedkin–Johnsen opinion dynamics, the controller trades off click-through rate and polarization reduction using only online click data, and the reported steady-state comparison shows $\Delta T$ 5 lower polarization cost with only $\Delta T$ 6 loss in CTR relative to a local OFO that treats users in isolation (Chandrasekaran et al., 2024).

Although OFO originated as a control-theoretic method for steady-state optimization of physical or networked systems, the phrase has broadened in adjacent areas. In advertising-text generation, a two-stage method described as “CTR-driven ad Text generation via Online feedback Preference optimization” first samples diverse candidate texts and then performs weighted Direct Preference Optimization from online A/B/n feedback. On $\Delta T$ 7 million held-out items, the reported result is a $\Delta T$ 8 win rate and a $\Delta T$ 9 relative CTR improvement, with long-term full deployment yielding a $x$ 0 CTR lift and a $x$ 1 RPM lift (Chen et al., 27 Jul 2025).

A closely related alignment formulation is Online AI Feedback, in which two responses are sampled from the current LLM at each training iteration and an LLM annotator provides an online preference label. Human evaluation on TL;DR reports Online DPO preferred $x$ 2 of the time, and prompt-controlled feedback can change the average response length from $x$ 3 tokens to $x$ 4 and $x$ 5 while preserving alignment gains over the supervised baseline (Guo et al., 2024).

These uses share the ideas of iterative improvement and online feedback, but their mathematical objects differ from classical OFO. Instead of optimizing $x$ 6 over plant inputs subject to physical constraints, they optimize preference-based or click-based learning objectives over a generative policy. A closer bridge to the classical setting appears in real-time preference optimization for human-in-the-loop control, where binary pairwise comparisons are used to estimate a descent direction for a controller coupled to a nonlinear plant, and explicit stability criteria and sub-optimality bounds are derived under assumptions on both plant dynamics and latent utility (Wang et al., 2 Jun 2025). This suggests that “Online Feedback Optimization” now denotes a family of feedback-driven optimization schemes whose unifying principle is online corrective use of measurements or preferences, but whose technical content depends strongly on whether the controlled object is a physical plant, a networked infrastructure, or a learned policy.