Reading Silence Inside Early-Stopped Moving Horizon Estimators

A sensor that transmits only when its reading changes by more than a set amount (a "send-on-delta" rule) saves bandwidth. It also tells the receiver something every time it stays quiet: the current reading is still within that amount of the last one sent. Today's optimization-based estimators for nonlinear systems ignore this. During quiet periods they simply run the model forward. This proposal writes the meaning of silence directly into the estimator's objective as a penalty that is zero inside the known band and grows only when the estimate implies a reading the sensor would have reported. It then proves that the estimator stays stable even when its non-convex optimization is stopped after one or two iterations, at whatever local stationary point it happens to reach.

The estimator is a moving horizon estimator (MHE): at every time step it solves a small optimization over a sliding window of recent data to find the state trajectory that best explains that data. For nonlinear models this optimization is non-convex. Embedded hardware cannot afford to solve it to global optimality. The central observation is that, under a send-on-delta rule, silence is equivalent to a measurement corrupted by bounded noise of size at most the threshold. That equivalence lets two existing stability results be combined: one for event-triggered MHE and one for MHE solved suboptimally. The result is a single estimator whose error bound grows with the threshold in an explicit way. It squeezes more accuracy out of each transmitted packet without paying for full convergence of the solver.

Core Claim

Event-triggered nonlinear MHE currently throws away the information in silent intervals. Adding a dead-zone penalty, one that ignores discrepancies inside the non-triggering band and penalizes only the excess, should give lower estimation error at a fixed transmission rate. Equivalently, it should give the same error with fewer transmissions. It achieves this with a solver capped at one or two iterations, and it keeps a provable robust exponential stability guarantee that holds at any stationary point the capped solver reaches. The new pieces are three: silence treated as an active term in the MHE objective, a stability proof that does not assume the event-driven problem is solved optimally, and an empirical account of whether the flat, kinked silence penalty creates or removes spurious local minima.

Where This Sits

Two event-triggered MHE papers form the closest work, and both motivate the gap. Event-triggered moving horizon estimation for nonlinear systems transmits and re-optimizes only when an output deviation crosses a threshold. Between events it predicts open-loop, and it proves robust global exponential stability under incremental input/output-to-state stability (i-IOSS). i-IOSS is the nonlinear analogue of detectability: two trajectories with matching inputs and outputs must converge. Robust stability of event-triggered nonlinear moving horizon estimation sends only one measurement per event and adds a constraint keeping estimated outputs near earlier predictions. Both assume the nonlinear program at each event is solved reliably, and neither uses the silence band as information during quiet steps. We differ on both counts.

Suboptimal Nonlinear Moving Horizon Estimation supplies the method for handling early termination. It requires only that the solver's output cost no more than a candidate produced by a simple stable auxiliary observer. Under that condition, stability survives stopping after one or two iterations. It assumes a measurement arrives at every step. We extend it to event-triggered sensing and to the dead-zone silence term.

Event-Triggered Particle Filter With One-Step Randomly Delayed Measurement already decodes silence in a nonlinear setting. It integrates the particle likelihood over the non-triggering zone, at a reported 7% to 44% computational overhead and with no deterministic stability guarantee. Our dead-zone penalty is the optimization-based counterpart of that integral.

Work on discrete and scalar systems shows that reasoning about silence is worth the effort. Remote Estimation of Markov Processes over Costly Channels: On Implicit Information Benefits reports roughly half the channel use at 90% accuracy compared with a randomized policy. Fundamental limits of remote estimation of autoregressive Markov processes under communication constraints proves threshold rules optimal for scalar autoregressive processes. Its proof relies on the error resetting to zero after each transmission, which fails in nonlinear multivariable systems. We therefore do not claim threshold optimality. We treat the threshold as a design parameter and characterize the error-versus-rate trade-off empirically.

Research Program

Take the sensor to be send-on-delta, applied to each output component separately. It transmits $y_t$ whenever some component differs from the last transmitted value $\bar y_t$ by more than $\xi$. Silence at step $t$ therefore certifies that $|y_t - \bar y_t| \le \xi$ componentwise. Both ends know this set without the sensor running a copy of the estimator.

At time $t$ the estimator decides a window of states $x$ and process disturbances $w$ by minimizing

$$ J_t = \Gamma(x_{t-N}) + \sum_{k} \|w_k\|_Q^2 + \sum_{k \in \mathcal{T}} \|y_k - h(x_k)\|_R^2 + \rho \sum_{k \in \mathcal{S}} \big\|\max(0,\ |\bar y_k - h(x_k)| - \xi)\big\|^2 . $$

Here $\Gamma$ is the prior cost on the first state in the window, $N$ is the window length, $h$ is the output map, $\mathcal{T}$ and $\mathcal{S}$ are the triggered and silent steps in the window, $Q$ and $R$ are weights, and $\rho$ weights the silence term. The last term is zero whenever the estimate is consistent with silence. It is the negative log-likelihood of bounded noise, smoothed to be continuously differentiable. We use a penalty rather than a hard constraint so that model mismatch cannot make the problem infeasible.

When the open-loop prediction already satisfies every band, the gradient of the silence term vanishes and no iteration is needed. Computation is spent only when the prediction starts to contradict what the silence implies.

The theory has three parts. First, we build the auxiliary observer candidate by feeding a nonlinear Luenberger-type observer the held value $\bar y_k$ on silent steps. A Luenberger-type observer corrects a model prediction with a gain times the output error. The held value is a genuine measurement with error at most $\xi$, so the i-IOSS machinery of Nonlinear Full Information and Moving Horizon Estimation: Robust Global Asymptotic Stability applies with $\xi$ entering as bounded measurement noise.

Second, we show that any iterate whose cost is no higher than this candidate's inherits robust global exponential stability. This holds for any number of iterations, zero included, and for any stationary point reached. The asymptotic error bound is a class-K function of $\xi$ plus the process-noise bound; a class-K function is increasing and zero at zero. We expect this dependence to be close to linear near zero, following the disturbance-to-error gain in the i-IOSS bound. The proof must handle the non-smooth kink in the penalty and the time-varying split into $\mathcal{T}$ and $\mathcal{S}$, which is where it departs from the suboptimal MHE analysis.

Third, we study the optimization landscape empirically. For each benchmark we run many random restarts of the full-convergence solve, with and without the silence term, and count distinct local minima and how far their costs fall from the best one found. This tests the concern that flat regions and band edges create stalling points.

A secondary thread handles the known weak point: a silent step that is really a dropped packet. The sensor sends a one-bit heartbeat at a low periodic rate. We measure how RMSE degrades when a fraction of transmissions are erased and the estimator wrongly reads the erasures as silence.

Baselines

The naive reference points are two, both run at the same average transmission rate as our method. The first is pure model prediction reset at each received packet. The second is periodic MHE with the sampling period stretched to match the rate. The second is informative on its own: Sample-based Moving Horizon Estimation reports RMSE rising from 0.1105 with data every two hours to 0.3622 with data every third day on the hypothalamic–pituitary–thyroid model. That gives us a published sparsity-versus-error curve against which to place event-triggered schemes on the same model.

The strongest existing competitors are the two event-triggered MHE schemes, Event-triggered moving horizon estimation for nonlinear systems and Robust stability of event-triggered nonlinear moving horizon estimation. They are the only methods that combine threshold transmission, nonlinear constrained estimation and stability guarantees. On the batch reactor, the first reports triggering the solver 14% of the time, an 86% reduction in computation, with error "similar to continuous MHE." No reported RMSE value for ET-MHE on the batch reactor was found on Lacuna. On the two-link arm, the second reports that its varying-horizon version beats its fixed-horizon version in 80% of cases, with accuracy gains up to 21%. No absolute error value for the two-link arm was found on Lacuna.

We will reimplement both with the same model, noise realizations, window length, trigger rule and threshold grid as our method. Each gets a fully converged interior-point solve, which favors them computationally. Tuning of weights and horizon gets an equal grid-search budget for every method on a held-out noise seed. Beating them means lower RMSE at every matched transmission rate across the threshold sweep, or equal RMSE at a rate at least a fifth lower. Our method must meet either criterion with its solver capped at two iterations.

For silence decoding in nonlinear tracking, the strongest competitor is Event-Triggered Particle Filter With One-Step Randomly Delayed Measurement. It reports keeping performance with communication cut to roughly 68%, at 7% to 44% extra computation. No scalar RMSE value was found on Lacuna for it. We run it on the five-state radar scenario with particle counts chosen to match our wall-clock time. Beating it means lower RMSE at matched rate and compute.

Two ablations isolate the key ingredients. The first keeps everything but replaces the dead-zone term with an ordinary quadratic penalty pulling $h(x_k)$ toward the held value $\bar y_k$. This tests whether the benefit comes from using silence correctly or merely from using the held value at all. The second keeps the silence term but removes the iteration cap, running to convergence. This measures what early termination costs.

Why not just…?

Why not just keep open-loop prediction between events, as existing ET-MHE does? It is stable, and it is free on silent steps. But it discards a certificate about the output at every quiet step. Its error drifts within the bound until an event fires, and Event-triggered moving horizon estimation for nonlinear systems itself notes both larger error at large thresholds and vulnerability to model-plant mismatch. Our lazy update keeps most of its cost advantage, since it iterates only when the prediction exits a band. The head-to-head threshold sweep will show whether the accuracy gain is worth the extra iterations. If the curves coincide, the silence term is not buying anything and we will report that.

Why not just treat the held value as an ordinary noisy measurement? This is the most serious simple alternative, because it is trivial to implement and already fits the suboptimal MHE analysis. Its weakness is bias. While the output drifts slowly inside the band, a quadratic pull toward the stale value drags the estimate backward, whereas the dead-zone term is indifferent anywhere inside the band. This is exactly the first ablation. If the quadratic version matches ours, the simpler method should be adopted, and the contribution shrinks to the stability proof for early-stopped event-triggered MHE.

Why not just use the event-triggered particle filter? It already reads silence correctly. However, its reported 7% to 44% overhead comes from computing integrals over the non-triggering zone, it degrades as state dimension grows, and it gives no deterministic stability guarantee. The compute-matched radar comparison, plus the multivariable reactor and arm benchmarks where particle counts must grow, will show where each wins.

Why not just run a mature solver to convergence? On a desktop that is often fine. The second ablation measures exactly how much accuracy convergence buys. The case for the cap is that event bursts coincide with the fast transients where solve time is scarcest. Suboptimal Nonlinear Moving Horizon Estimation reports that two iterations already clearly beat the auxiliary observer, with accuracy comparable to specialized solvers.

Proposed Experiments

We use four benchmarks from the file. The batch reactor tests constrained multivariable chemistry with non-negativity limits. The two-link arm tests coupled mechanical nonlinearities. The hypothalamic–pituitary–thyroid model tests sparse clinical sampling. The radar aircraft tracking scenario provides the particle-filter comparison. The modified Chua circuit, on which Nonlinear moving horizon estimation for robust state and parameter estimation reports parameter RMSE of 0.0209 for adaptive MHE, serves as a stress test for model mismatch during unexcited phases.

On each benchmark we sweep the threshold $\xi$ and plot RMSE against realized transmission rate, averaged over at least fifty noise seeds. We also record iterations and wall-clock time per step, and the fraction of silent steps that needed any iteration.

Three checks settle the core claim. The first is whether the capped silence-aware estimator's error-versus-rate curve lies strictly below both ET-MHE curves and the quadratic-hold ablation. The second is whether measured asymptotic error grows with $\xi$ within the proven bound, and roughly linearly near zero. The third is whether multi-start counts show the silence term leaves the number of distinct local minima unchanged or reduced. The erasure experiment then reports the packet-loss rate at which reading silence starts to hurt rather than help. That is the regime where this idea stops being useful.