Controllers that steer a physical plant toward an optimal operating point often run on a stability proof that is far more pessimistic than the hardware. Online feedback optimization is one example: a controller nudges the inputs of a running plant, such as a building's heating or a power converter, along a gradient computed from live measurements. Its guarantees typically demand a tiny step size because they must hold for every plant the designer cannot rule out. Meanwhile the controller needs a model of how outputs respond to inputs (the plant's sensitivity, a Jacobian matrix), and that estimate is usually kept in check by parameter projection, which snaps any estimate that leaves a fixed "plausible" set back onto it. Both pieces of pessimism come from the same place, a large uncertainty set fixed before any data arrive.
We propose to make that set the object that shrinks. As measurements accumulate, we maintain the set of sensitivity matrices consistent with the data. We project the running estimate onto the subset of estimates that provably yield a convergent closed loop for every plant still in that set, and we recompute the largest step size that a stability certificate allows over the same set. The step size is therefore earned: it starts at the conservative worst-case value and grows toward the plant's empirical stability limit as the data rule out bad plants. The claim we will test is that this closes most of the gap between certified and empirically safe gains, a gap that can reach an order of magnitude (nested multi-timescale Lyapunov analysis of converter grids reports bounds off by a factor of ten). It should do so while never running a gain the current data cannot certify.
Core Claim
Take a plant whose steady-state output responds to inputs $u$ through an unknown sensitivity $H$. Online feedback optimization updates $u_{k+1} = \Pi_U\big(u_k - \eta_k(\nabla f(u_k) + \hat H_k^\top \nabla g(y_k))\big)$, where $f$ and $g$ are input and output costs, $y_k$ is the measured output, $\Pi_U$ projects onto the input constraints, $\hat H_k$ is the sensitivity estimate, and $\eta_k$ is the step size. When $\hat H_k \neq H$, convergence requires that the estimated gradient direction still correlate positively with the true one. For quadratic costs with weights $R$ (input) and $Q$ (output), this is a matrix inequality $\mathrm{sym}(R + \hat H^\top Q H) \succ 0$, and it is linear in $\hat H$ for fixed $H$ and linear in $H$ for fixed $\hat H$. That bilinear structure is the technical hook. If the data-consistent set $S_k$ of plants is a polytope, the set of robustly safe estimates
$$\mathcal{C}_k = \{\hat H : \mathrm{sym}(R + \hat H^\top Q H) \succeq \mu I \ \text{for all vertices } H \text{ of } S_k\}$$
is convex, where $\mu > 0$ is a margin, and it grows as $S_k$ shrinks. Projecting onto $\mathcal{C}_k$ is a small semidefinite program. The certified step size is then
$$\eta_k^\star = \max\{\eta : \text{a discrete-time max-type Lyapunov decrease condition holds for all } H \in S_k, \text{ given } \hat H_k\},$$
which is also a vertex-wise check. The new element is not projection, and not set-membership identification. It is the use of one shrinking set for two jobs: it defines where the estimate may live, and it defines the gain the certificate allows. As a result, conservatism falls monotonically with data rather than being fixed at design time.
Relation to Prior Work
Three papers motivate the gap. The analysis of online feedback optimization without timescale separation replaces quadratic Lyapunov functions with a max-type function $V=\max\{\xi V_x, V_u\}$. It certifies gains $\alpha=100$ where singular-perturbation arguments required $\alpha<10.1$, which shows that much of the conservatism comes from proof technique. That analysis assumes the plant Jacobian is known, works only in continuous time, and explicitly leaves online Jacobian estimation open. We supply both missing pieces: a discrete-time sampled analog, and the estimation loop. The sampled-data feedback equilibrium seeking work gives the discrete-time setting and the building benchmark, but its small-gain bounds require slow updates; it is the conservative sampled-data reference we aim to beat. The nested multi-timescale Lyapunov framework documents the size of the conservatism gap in practice.
Two papers supply methods. Robust adaptive MPC with moving-window set-membership shrinks a parameter set online, but it restricts the set to an axis-aligned hypercube and uses it only to size MPC tubes, never to relax a Lyapunov gain bound. We reuse its set-update idea with polytopes and test its hypercube simplification as an ablation. The Lur'e contractivity analysis of projection controllers shows that the non-expansiveness of metric projections enters stability certificates exactly rather than as a worst-case bound. We rely on the same property for $\Pi_U$ inside the discrete-time certificate.
Two papers constrain the design. Adaptive economic MPC with projected LMS projects estimates onto a static physical set and needs an open-loop stable plant; it is the static-projection baseline in our comparisons. Projection to State Stability shrinks uncertainty only along a Lyapunov derivative and relies on known Lipschitz constants. We need something weaker: a known bound on measurement noise. That assumption is still a real risk, which the falsification test below addresses. The RL projection results in the target-network deadly-triad analysis note the same pathology from the learning side: theoretical projection radii are far larger than needed. This suggests the idea is not specific to control, but we do not pursue RL here.
Research Program
The theory has three parts. First, we prove a discrete-time version of the max-type parameter-dominance condition for a sampled plant under zero-order hold. Second, we show that for a fixed $\hat H$ it is affine in $H$, so vertex checks over a polytope $S_k$ suffice. Third, we prove that while the true $H \in S_k$, projecting onto $\mathcal{C}_k$ and stepping with $\eta_k \le \eta_k^\star$ yields input-to-state stability with respect to plant transients and noise. Monotonicity is immediate: $S_{k+1}\subseteq S_k$ implies $\eta^\star_{k+1}\ge\eta^\star_k$.
The algorithm intersects $S_k$ with the slab of plants consistent with the newest input–output increment under a bounded-error model. To keep cost fixed, it uses a moving window and an outer polytope with a capped vertex count, following the complexity concern raised in robust adaptive MPC. It then runs the projection SDP and a scalar line search for $\eta_k^\star$. Excitation comes from the controller's own moves plus a small dither whose energy is budgeted and charged against performance.
Safety under misspecification is built in. If $S_k$ becomes empty, which happens when the plant changes or the noise bound was wrong, the set is reinflated to the prior and $\eta$ falls back to the conservative value. This makes the scheme fail safe rather than fail silent.
Baselines
The naive reference is the classical design. It uses a fixed nominal Jacobian with the step size certified by timescale separation, the regime that the max-type OFO paper reports as limited to $\alpha<10.1$ on its LTI plant.
The strongest existing competitor depends on the task. For theory-backed gain on the LTI benchmark, it is the max-type certificate itself, applied with the gain computed once over the full prior uncertainty set. This is the best that paper's method can do without learning, and it also has a reported stable operation at $\alpha=100$ when the Jacobian is exact. For the building benchmark, it is the sampled-data feedback equilibrium seeking controller, which reports a 55.15% reduction in constraint violations and a 17.99% reduction in energy cost against hysteresis control over 96 hours on the five-room BRCM model.
A further baseline is static-box projection in the style of projected-LMS adaptive economic MPC. It uses the same estimator, but projects onto a fixed physical set with a gain certified over that set; no reported numeric cost is found on Lacuna for it. We also include an oracle that certifies with the true $H$, which gives the ceiling on achievable certified gain.
Two ablations isolate the key ingredient. One shrinks $S_k$ but keeps projection onto the static physical set, which separates the gain-scheduling role of the set from its projection role. The other replaces polytopes with the hypercube sets of robust adaptive MPC, which tests whether set geometry matters.
All methods share the same plant realizations, noise sequences, dither budget, sampling interval, and solver. Every hand-tuned baseline gets the same 50-configuration tuning budget on held-out plants. Beating a baseline means two things: lower time-averaged suboptimality and constraint violation at equal dither energy, and zero divergent runs. On the building task, beating means improving on the feedback equilibrium seeking controller's violation and cost reductions measured against the same hysteresis controller in our own reimplementation.
Why not just…?
Why not just run certainty-equivalent estimation with a large, empirically tuned gain? This is the real threat, because on well-excited plants it may match our tracking. Its failure mode is the early transient and low-excitation periods, when the estimate can have the wrong sign structure. Unprojected updates are known to fail recursive feasibility in robust adaptive MPC. We will measure the fraction of 500 randomly drawn plants on which tuned certainty equivalence diverges or violates constraints. If that fraction is near zero and its performance is equal to ours, our contribution reduces to a certificate with no performance benefit, and we will report it that way.
Why not just certify once over a carefully chosen tight prior set? A tight prior trades conservatism for the risk of excluding the true plant, which is the bias problem of static projection. The static-box baseline, run with both a generous and a tight prior, shows this trade directly. The tight prior should either match our late-time gain while being wrong on some sampled plants, or be safe and slow.
Why not just use the max-type certificate with slow, decoupled estimation? Slow estimation reintroduces the timescale separation that the max-type OFO paper was written to remove, and the gain stays pinned to the worst case until the estimate settles. The first ablation, which shrinks the set but leaves projection static, quantifies how much of our gain comes from coupling the two roles.
Proposed Experiments
The first benchmark is the high-gain LTI plant from the max-type OFO paper, run with a randomized unknown sensitivity and bounded measurement noise. The primary metric is the conservatism ratio: the empirical maximal stable gain, found by bisection on the true plant, divided by the certified $\eta_k^\star$, tracked over time. The decisive check is whether this ratio falls from its prior-set value to within a factor of two of one inside a few hundred steps, without any run exceeding its certified gain.
The second benchmark is the five-room BRCM building model used by sampled-data feedback equilibrium seeking, with the room sensitivities treated as unknown. We report constraint violations and energy cost over 96 hours against hysteresis control.
The third check is falsification. Midway through a run we change the plant parameters, and separately we underestimate the noise bound by half. Then we verify three things: the empty-set test fires, the gain falls back, and stability is preserved. If the conservatism ratio does not shrink, or if fallback events are frequent at realistic noise levels, the premise that data can tighten certificates cheaply is wrong for this class. That result would still be worth publishing.
References
- A Lyapunov Framework for Nested Dynamical Systems on Multiple Time Scales With Application to Converter-Based Power Systems
- A stability condition for online feedback optimization without timescale separation
- Sampled-Data Online Feedback Equilibrium Seeking: Stability and Tracking
- Linear robust adaptive model predictive control: Computational complexity and conservatism
- Exponential Stability of Parametric Optimization-Based Controllers via Lur'e Contractivity
- Adaptive Economic Model Predictive Control for linear systems with performance guarantees
- A Control Lyapunov Perspective on Episodic Learning via Projection to State Stability
- Breaking the Deadly Triad with a Target Network