A controller that talks to its plant over a network has to decide two things. The first is when to send a fresh measurement. The second is how precisely to encode it. Event-triggered control, meaning feedback that transmits only when the state has drifted enough to matter, has made the "when" question very efficient. A recent refinement replaces the usual requirement that the system's energy fall at every instant with a looser one: the energy only has to stay below a decaying ceiling. Under this rule, vehicle-platoon controllers needed about 9.94 updates instead of 198.28 (Performance-Barrier-Based Event-Triggered Control With Applications to Network Systems). Every one of those updates, however, still carries a full-precision state vector. The "how precisely" question, which is a compression-accuracy trade-off, has been left out.
This note proposes treating the gap between the energy ceiling and the current energy, called the performance residual, as one budget that can be spent on two things: waiting longer before the next transmission, or sending fewer bits when one is made. When the residual is large, the system can afford a coarse 2–4 bit message. As the energy approaches the ceiling, precision rises. The decaying ceiling is a known function of time, so both ends of the link can shrink the quantizer's range along with it without exchanging any extra information. The working claim is that this keeps the ceiling guarantee exactly. It should also cut total transmitted bits substantially below the best fixed-precision version of the same trigger, and that gap is the claim the experiments must establish.
Core Claim
Two concepts need defining. A Lyapunov function $V(x)$ is a scalar "energy" of the state $x$ whose decrease certifies stability. A performance-barrier trigger enforces $V(x(t)) \le S(t)$, where $S(t) = (V(0)-V_\infty)e^{-ct} + V_\infty$ is an exponentially decaying ceiling with rate $c$ and floor $V_\infty$. Between events the actuator holds the last received value. The trigger therefore tolerates a measurement error $e(t)$ (the held value minus the true state) as long as a condition of the following form holds:
$$\beta(\|e(t)\|) \le \sigma\,\alpha(\|x(t)\|) + \gamma\,\big(S(t) - V(x(t))\big).$$
Here $\alpha$ is the nominal decay rate of $V$, $\beta$ bounds how much the error increases $V$, $\sigma\in(0,1)$ is the usual safety fraction, and $\gamma>0$ sets how much of the residual $S-V$ the error may consume. The proposal adds one ingredient. At an event the state is not sent exactly. It is quantized with $b$ bits on a range proportional to $\sqrt{S(t)}$, so the error immediately after the event is at most $\varepsilon_b(t) \propto 2^{-b}\sqrt{S(t)}$ rather than zero.
The bit depth $b$ is chosen at each event to minimize bits per unit time, $b/\hat\tau(b, r)$. Here $r = S-V$ is the current residual and $\hat\tau$ is a certified lower bound on the next inter-event time given that starting error. Two properties have to be proved. First, as long as $\varepsilon_b$ leaves the trigger inequality strictly satisfied right after the event, the barrier stays invariant by the same argument (Nagumo's invariance theorem) used by Ong and Cortés. Second, a positive minimum inter-event time survives, so there is no Zeno behavior, meaning infinitely many events in finite time. Because the quantizer range shrinks with $S(t)$, the number of bits per event stays bounded even as the state converges.
Research Program
The work builds on the performance-barrier analysis of Ong and Cortés, which proves envelope invariance and a positive minimum inter-event time for exact transmissions under globally Lipschitz dynamics. Its silence on message precision is the gap addressed here. A second methodological basis is Hertneck, Linsenmayer and Allgöwer. They show that average-decrease, non-monotonic certificates enlarge the admissible sampling period of periodic event-triggered control by at least $9/4$ and tolerate packet loss. Their periodic checking schedule is the right implementation for a digital quantizer, because it avoids continuous sensor monitoring.
The first deliverable is a theorem for the periodic, sampled version. It gives a certified $\hat\tau(b,r)$, a minimum inter-event time that depends on the bit floor, and invariance of $V \le S$ under up to $m$ consecutive dropouts, with the dropout bound inherited from Hertneck et al.
The second deliverable is the allocation rule. In its simplest form it is a lookup over $b \in \{2,\dots,16\}$ evaluated at each event. The project will also examine whether a greedy per-event choice is far from the offline optimum computed by dynamic programming on the pendulum.
The third deliverable carries the same accounting over to learned controllers, which is where the compression-accuracy literature becomes relevant. Neural Event-Triggered Control with Optimal Scheduling co-trains a neural controller and Lyapunov function with strict pointwise descent at full 32-bit precision. The plan is to replace its descent constraint with the barrier condition above. As a secondary, explicitly exploratory step, the controller itself will be compressed using 8-bit weights and 2:4 sparsity with block-diagonal correctors in the style of ARMOR. The resulting policy error is bounded empirically and charged to the same residual. This step is secondary because ARMOR's evidence comes from language-model perplexity, not closed-loop dynamics. Whether its corrector matrices preserve the Lipschitz behavior a control certificate needs is an open question that this step tests rather than assumes.
Baselines
The naive reference point is periodic transmission at the fixed bit depth needed to certify stability in the worst case. On the inverted pendulum, Hertneck et al. report that a periodic scheme requires sampling at $h = 1.38\times10^{-5}$ s, while their event-triggered scheme transmits on average every 0.48 s. Both use unquantized values, so bit counts follow directly from multiplying by the word length. This baseline shows how large the total problem is.
The strongest existing method is the performance-barrier trigger of Ong and Cortés, sending full-precision 32-bit states. Their reported result on the five-vehicle platoon over 50 initial conditions is 9.94 average controller updates, against 198.28 for a derivative-based trigger. That method will be reimplemented on the identical platoon setup, graph, and initial conditions, with the same $c$, $\sigma$, and $V_\infty$. Beating it means matching its envelope compliance (zero violations of $V \le S$) and its convergence time within 5%, while cutting total transmitted bits by at least 3×. Fewer updates alone would not count. No reported result was found on Lacuna for bit counts under any performance-barrier trigger, so these figures are produced here.
The ablation that isolates the key ingredient is a fixed-$b$ performance-barrier trigger, robustified against quantization error, swept over every $b$ from 2 to 16. Each $b$ gets the same tuning budget for $\sigma$ and $\gamma$, and the best $b$ is selected per system by an oracle with hindsight. A second ablation keeps adaptive $b$ but quantizes on a fixed range instead of the $S(t)$-scaled range. This separates the benefit of adaptivity from the benefit of the shared shrinking range.
For neural controllers, the comparator is Neural ETC with optimal scheduling run at full precision. On the 2D gene regulatory network, Lorenz system, and 100-dimensional cell model it reports as few as 6 to 13 triggers versus 1,999 for standard neural Lyapunov control, and an MSE of 5e-8 on the cell model. It will be retrained with the same number of optimizer steps as the proposed variant.
Why not just…?
Why not just use the best fixed bit depth with the existing barrier trigger? This is the real threat. If quantization error at the right fixed $b$ is already small relative to the residual for most of a trajectory, adaptivity buys little. The oracle-best-fixed-$b$ sweep settles the question. The project's claim survives only if adaptive allocation beats that oracle by a clear margin, roughly 1.5× fewer bits at equal envelope compliance. The expectation that it will rests on one asymmetry: a fixed $b$ has to be sized for the moments when $V$ is near $S$, while most time is spent well below it. If the margin comes out small, the honest conclusion is that the contribution reduces to the $S(t)$-scaled range, and the note would report exactly that.
Why not just use dynamic event triggering with an internal filter variable, plus quantization? Dynamic triggering, introduced by Postoyan et al. and made data-driven in Data-Driven Dynamic Event-Triggered Control, also allows temporary energy increases. It does not, however, provide an explicit ceiling that both ends of the link know. That shared ceiling is precisely what lets the quantizer range shrink without extra signaling, and what makes the residual a readable budget. A dynamic-trigger variant will still be run with the identical quantizer on the pendulum and platoon. If it matches on bits, the envelope adds nothing beyond guaranteed transient bounds.
Why not just send full precision less often, since events are already rare? With 6 to 13 events per run in Neural ETC, bits per event might appear negligible. On lossy, low-rate links, though, each message's size determines its latency and its chance of being dropped. Coarser messages also tend to arrive sooner, which interacts with the computational-delay concerns raised in Event-triggered Control and Online Learning for Networked Systems under Computational Delays. The experiments therefore report total bits and also convergence under a rate-limited channel with packet loss. If full precision wins under that channel, the idea fails on its own terms.
Proposed Experiments
The analytical systems are the Boubaker inverted pendulum from Hertneck et al. and the five-vehicle platoon from Ong and Cortés. The learned systems are the gene regulatory network, Lorenz, and 100-dimensional Michaelis–Menten benchmarks from Neural ETC. Metrics are total bits to reach a fixed neighborhood of the origin, number of events, empirical minimum inter-event time, envelope violations (which must be zero), and convergence time. Each is reported under three conditions: an ideal channel, a channel with up to $m=1$ consecutive dropout, and a rate-limited channel where message duration is proportional to $b$.
The first decisive check compares adaptive allocation with the oracle fixed-$b$ sweep. It falsifies the core claim if the bit reduction is under 1.5×.
The second decisive check stresses the Zeno-freedom proof. It drives $b$ to its floor and verifies that the measured minimum inter-event time stays above the certified $\hat\tau$, using Monte Carlo over 1,000 initial conditions.
The third check is exploratory and concerns the compressed neural controller. It asks whether an 8-bit, 2:4-sparse policy with charged policy error still converges on the cell model within an order of magnitude of the 5e-8 MSE that Neural ETC reports at full precision. Where the network is piecewise linear, the barrier condition is verified formally with the reachability tools of Reachable Polyhedral Marching.
A clean negative result on the first check would still be informative. It would show that, under barrier triggering, the time dimension already absorbs nearly all of the available margin.
References
- Performance-Barrier-Based Event-Triggered Control with Applications to Network Systems
- Nonlinear Dynamic Periodic Event-Triggered Control with Robustness to Packet Loss Based on Non-Monotonic Lyapunov Functions
- Neural Event-Triggered Control with Optimal Scheduling
- ARMOR: High-Performance Semi-Structured Pruning via Adaptive Matrix Factorization
- A Framework for the Event-Triggered Stabilization of Nonlinear Systems
- Data-Driven Dynamic Event-Triggered Control
- Event-triggered Control and Online Learning for Networked Systems under Computational Delays
- Reachable Polyhedral Marching (RPM): A Safety Verification Algorithm for Robotic Systems with Deep Neural Network Components