Closed-loop identification and the stim-record chain

What the work claims

This is a methods paper with simulation experiments, from EPFL's Institute of Mechanical Engineering (Meroi, Boroujeni, Massai, Ferrari-Trecate), supported by the Swiss National Science Foundation. It addresses a genuine difficulty: when a system runs under feedback, the input is correlated with the measurement noise, because the controller generates the input from the noisy measurements. Standard estimators fit that correlation into the dynamics. Worse, if the underlying system is unstable, open-loop data collection is impossible, and a learned model that is slightly wrong can turn the closed loop unstable in simulation or in deployment.¹

The claim has two parts. First, the authors derive a complete parameterization of all nonlinear systems stabilized by a given controller, generalizing the classical dual-Youla parameterization from linear control theory; any model drawn from this family is stabilized by the data-collection controller by construction, so closed-loop simulation rollouts cannot diverge. Second, they add a PAC-Bayesian training objective that, given an informative prior, delivers a finite-sample, non-vacuous bound on the model's expected prediction error without spending held-out data, optimized with Stein variational gradient descent. Context variables let one certified model identify a family of related systems and generalize to unseen members.¹

How it works

The classical dual-Youla parameterization represents every plant stabilized by a known controller as a fixed stable factorization plus a free stable parameter; searching over that free parameter can never leave the stabilized set. The paper generalizes this idea to nonlinear systems: it fixes a nominal model and the known controller, and parameterizes the discrepancy through a learnable operator constrained to be stable (in the experiments, a contractive state-space model), so that every candidate model inherits the stabilization guarantee. The indirect variant of the method then reduces the closed-loop identification to an open-loop problem whose regressors are, by construction, independent of the measurement noise, removing the bias that direct methods suffer under feedback.¹

Four simulation studies support the claims. On an open-loop unstable scalar plant (state update 3x plus tanh(x) plus input, stabilized by a known nonlinear controller), across 40 training runs at the highest noise level, every indirect and direct ICI model kept the closed loop stable, while a direct output-error baseline returned models the controller failed to stabilize in 25% of runs and a closed-loop output-error baseline in 2.5%. On a partially observed point-mass robot under colored measurement noise, indirect ICI held a closed-loop FIT of 89.6% at noise level 1 where the direct baseline fell to 72.4%; at noise level 2 only indirect ICI stayed positive at 82.7% against 43.5% and 30.3% for the baselines, and the authors measured a significant Pearson correlation (0.28) between the direct models' simulation error and the actual measurement noise, confirming the noise was absorbed into the learned dynamics. A context-conditioned variant reached 86.6% open-loop FIT on friction fields never seen in training versus 42.2% without context, an improvement of 44 percentage points that 15x more parameters could not replicate without the context signal. Under data scarcity, ERM training memorized noise (deployment FIT 67.7% against the pretrained prior's 86.6% at 4 trajectories) while the PAC-Bayesian posterior held about 85% FIT from a single trajectory upward, with the generalization certificate becoming non-vacuous from 4 trajectories at confidence 90%.¹

Where a skeptic should push

The most load-bearing assumption is that the plant is time-invariant and that the controller is known exactly. The certificate is with respect to the specific data-collection controller, not to any controller you might later deploy; swap the stimulation policy and the stability guarantee does not transfer. That matters because the most interesting use case, adapting the controller online, is precisely where the certificate is silent.¹

Second, everything is simulated. The benchmark plants are a scalar unstable system and a point-mass robot with synthetic colored noise; no physical plant, and certainly no biological one, appears in the experiments. Real disturbance structure, unmodeled drift, and partial observability far nastier than hidden velocities will stress the informative-prior requirement, and the paper's own data-splitting result shows the certificate is forfeited entirely without one. Third, the FIT metric on bounded rollouts can hide one-step prediction behavior that accumulates error slowly; the authors do probe out-of-distribution input waveforms, a good-faith check, but on synthetic excitation shapes. Treat the framework as a correct and useful formalization of a known trap, not as evidence the trap has been solved on hardware.

What this changes for closed-loop MEA rigs

Closed-loop microelectrode array experiments, where stimulation is triggered by detected activity or delivered on a feedback schedule, produce exactly the data regime this paper analyzes. The controller is the stimulation policy; the plant is the living preparation; the measurement noise is electrode noise plus the stimulation artifact, which in voltage recordings can be orders of magnitude larger than the spikes of interest. The non-obvious implication: because the stimulation is a function of the recorded activity, the artifact is feedback-correlated by construction, so it leaks into any model of tissue dynamics fit directly to stim-and-record data. This is the paper's drawback D.2 with an instrumentation name. Any lab fitting a recurrent network to predict organoid or culture responses to closed-loop stimuli, then using that model to design a better stimulus, is running the direct method the experiments show absorbing noise into its learned dynamics (correlation 0.28) and, at the unstable extreme, risking a fitted model whose interconnection with the stimulation policy is not stable.¹

The opportunity is a design spec for defensible closed-loop rigs: record the stimulation policy explicitly, keep an excitation component in the stimulus train that the feedback does not determine (the paper shows direct methods failing below minus 28% FIT when excitation is weak and recovering only above moderate levels), fit indirect-style models whose regressors are noise-independent, and demand a stability argument for model-plus-controller before closing the loop on living tissue. For seizure-suppression or destabilization-avoidance experiments on organoids, where the preparation itself can exhibit runaway bursting, that last requirement is not academic.

The threat cuts two ways. Instrumentation vendors selling adaptive stimulation toolchains inherit the liability: a prettier adaptive controller fitted to closed-loop data without excitation guarantees is a biased model with a stability certificate attached to the wrong controller. And the tissue fights back in ways the theory assumes away: plasticity makes the plant time-varying within a session, the controller's own artifact blanking corrupts the measurements it conditions on, and the guarantee that matters most, that the fitted dynamics of a living system are true rather than merely stable, is exactly the one no parameterization can supply. The defensible position for the field is to adopt the framework's discipline (known policy, deliberate excitation, noise-independent regressors, certified model class) while treating the certificate as necessary, never sufficient.

The bottom line

Established, in simulation: a complete nonlinear generalization of dual-Youla parameterization gives closed-loop stability by construction; indirect identification avoids the feedback-correlated bias that measurably contaminates direct fits (0.28 correlation with the injected noise); context conditioning generalized to unseen environments where 15x parameters without context could not; and a PAC-Bayesian certificate became non-vacuous from 4 trajectories. Unproven: anything on physical or biological plants, certificate transfer across controllers, and time-varying tissue. For microelectrode array work the paper's practical gift is diagnostic: if your stimulation is a function of your recordings, your system identification dataset is biased by design, and the fix is architectural, a matter of how the rig excites, regresses, and certifies, not of more data. What would confirm the claim: a demonstration on a physical unstable plant, and an adaptive-stimulation MEA study that audits its own fitted models for noise absorption before trusting them.