Research analysis · Control theory

Spike-only control with a certificate under firing-rate caps

A controller that talks to a physical system only through fixed-amplitude spikes, no more often than a hard rate limit allows, sounds like a toy model of a neuromorphic chip. It is also an uncomfortably accurate model of what it means to steer living neural tissue, where every command is a current pulse and cells impose their own refractory ceilings. Nader Meskin of Qatar University shows that such a controller need not be heuristic: by lifting the system into coordinates where spike-induced jumps are exactly linear, he builds a spiking controller for nonlinear plants that comes with a stability proof, an explicit ultimate bound, and a data-driven design recipe, then swings an inverted pendulum up from 99.94 percent of initial states under actuation too weak to hold it statically.

Source: Koopman Greedy Spiking Control for Nonlinear Systems under Firing-Rate Limits, arXiv:2610.41530, October 2026, preprint submitted to Automatica. Primary source. Read: the full 23-page PDF including the stability theorems, the synthesis linear program, and the pendulum case study with Table 1.

What the work claims

This is a single-author theory paper with one simulation case study, and it should be weighted as such: the mathematics is the contribution, the pendulum is a demonstration rig. The claim has three layers. First, for observable dictionaries invariant under differentiation along the input direction, a spike-induced state translation admits an exact finite-dimensional linear representation in Koopman lifted coordinates, so the jump dynamics that make spiking control awkward become exact matrix exponentials rather than approximations. Second, with firing decisions taken greedily by comparing quadratic Lyapunov functions, a clocked controller with at most one spike per tick and a hard rate bound of one spike per period is practically stable, with an explicit ultimate bound derived in Theorem 1 and restated as practical stability in Corollary 1. Third, the Lyapunov function itself can be built from Koopman eigenfunctions learned from unforced data alone, by generator-based regression, with the weights synthesised by a linear program that enforces a sampled relaxation of the rate-limited control-Lyapunov condition.1

The numbers that make it land: on a pendulum with gravity constant 9.81 per second squared and weak damping 0.1 per second, spikes inject plus or minus 0.06 radians per second of velocity at a 100 Hz clock, capping average actuation at 6 radians per second squared, or 0.61 of what gravity demands; static balancing is therefore impossible beyond 0.66 radians from upright. The learned neutral eigenfunction recovers the pendulum's mechanical energy with 0.02 percent error in the gravity constant, the estimated unstable eigenvalue 3.0825 matches the true 3.08 to four digits, and the lifted jump model matches direct evaluation to 2.8 times ten to the minus fourteen. Against a linear greedy controller tuned on the linearised system, the Koopman controller succeeds from 1680 of 1681 grid initial conditions, 99.94 percent, against 73.9 percent, and uses fewer spikes, 137 versus 148 on average over the 1243 states both solve. An ablation that zeroes the energy eigenfunction's weight collapses the success rate to 42.5 percent.

How it works

The Koopman operator watches a nonlinear system through observables, functions of the state, and evolves those observables linearly even when the underlying dynamics are nonlinear. The trick this paper exploits is that spikes do not flow: they translate the state by a fixed jump, and for a suitable dictionary of observables that translation is exactly the matrix exponential of a known generator, computable in closed form. Spiking control then becomes a sequence of exact discrete moves in a lifted space where firing decisions reduce to comparing the values of a quadratic Lyapunov function before and after the candidate jump. Greedy firing, spike only if the jump decreases the Lyapunov function enough, inherits a proof because the clock enforces a minimum inter-spike interval and the rate limit bounds average input, so between spikes the drift contracts the same function and the whole loop behaves like a contractive discrete-time system with a quantified error floor.

The design recipe is deliberately data-light. From 300 unforced trajectories of two seconds each, 30,000 samples with a one millisecond finite-difference step, the method fits the eigenfunction associated with the neutral mode, which for the pendulum is the total energy, and the eigenfunction of the unstable mode, whose growth the controller must actively cancel. Stable modes are left alone because they decay on their own; Proposition 3 makes the point that penalising the unstable eigenfunction drives the state toward the stable manifold where drift finishes the job without further spikes. A convex program then searches the eigenfunction weights for the largest certified sublevel set, and a bisection on a witness constant trades certificate size against basin size. The certificates are honest but conservative: the certified clock period comes out at the order of one millisecond while the controller demonstrably works at the ten millisecond clock actually used, a roughly tenfold conservatism the author reports openly, and the certificate-hungry synthesised designs shrink the basin from 99.94 to between 97.4 and 99.9 percent depending on the constraint settings, while growing the certified region by 30 to 80 percent.

Where a skeptic should push

The single most load-bearing assumption is that the lifted model is valid where the controller operates. For the jump map this is exact by construction, but the flow model is only approximate, because the span of the chosen dictionary is not invariant under the pendulum's drift; the paper says so. Everything downstream, the Lyapunov inequalities, the ultimate bound, the certified sets, inherits that approximation, and the proof machinery handles it through inequalities with slack rather than exact invariance. On a pendulum the approximation error is small and well behaved. On a system whose unmodelled dynamics are not small, the certificate degrades gracefully in theory and silently in practice.

The second push is on evidence breadth. There is one plant, one regime, simulation only, no hardware, no noise beyond what the nominal model contains, and full state feedback throughout: the controller can see the exact pendulum angle and velocity at every instant. Real actuation channels give you partial, delayed, quantized observations. The third push is definitional: practical stability is the right guarantee class for spiking actuation, convergence to a bounded neighbourhood rather than a point, and readers who skim the abstract as tracking or optimal control will overread it. The author does not overclaim; the abstract says practical stability with an explicit ultimate bound, and that is exactly what is delivered.

Rate-limited control and the organoid closed loop

For organoid intelligence this paper matters because it is the first clean statement of the control problem the field actually faces, written in the field's own actuation language. Every closed-loop organoid experiment is a spiking control problem: commands are fixed-amplitude current pulses, the number of stimulation channels is tiny, tissue safety imposes hard rate and charge limits, and the plant is nonlinear. What has been missing is not stimulation hardware but a policy with any guarantee attached. Meskin's framework shows the guarantee is achievable under conditions that can be written down, and the recipe is startlingly well matched to what a wetware lab can already do: learn the eigenstructure from unforced recordings, which is precisely the spontaneous-activity data every organoid on an array generates for free, and only then close the loop.

The non-obvious implication is about the Lyapunov function, not the spikes. The controller works because its energy-like function is the plant's own conserved quantity, recovered to 0.02 percent from data; stabilising means driving the tissue toward a target dynamical regime, energy-low or energy-high, defined in the tissue's own coordinates. That reframing travels: an organoid trainer should think of stimulation policy as regulating a computable activity-energy landscape, hill-climbing or valley-keeping in a measured coordinate, rather than replaying patterns. The spike-efficiency result, 137 versus 148 spikes on average, and the steady state near five spikes per second, also quantify how cheap event-based regulation can be compared with clocked dense stimulation, which is the only style of closed-loop stimulation most current setups know.

The threat is equally specific. The theory assumes the plant's vector field is fixed, so its Koopman eigenfunctions are eternal; living tissue rewrites its own vector field through plasticity on exactly the timescales of a training run, which means learned eigenfunctions expire, and a certificate computed on Monday does not cover Wednesday's substrate. The tenfold gap between certified and working clock periods on a toy plant is a warning about how conservative the machinery gets when the model is merely approximate, and tissue is worse than approximate, it is non-stationary. And full state feedback has no organoid analogue: you observe spikes through an electrode bottleneck, not a state vector. So the honest transfer is a floor plan: regulate an energy-like observable with event-based spikes under a rate cap, re-learn the coordinates as the tissue drifts, and settle for practical stability, a bounded neighbourhood of a target regime, as the strongest claim any wetware closed loop can currently aspire to. That reframing alone is worth the paper's weight to the field.

The bottom line

Established, in simulation with proofs: exact lifted models of spike-induced jumps, a greedy clocked spiking controller with a hard firing-rate limit that is provably practically stable with an explicit ultimate bound, and a data-driven design from unforced trajectories that swings up and stabilises an inverted pendulum from 99.94 percent of initial states under actuation too weak for static holding. Not established: hardware validation, robustness to model error beyond the pendulum, partial observation, or any plant that learns. The confirming result for the organoid programme would be a demonstration that energy-like eigenfunctions learned from an organoid's spontaneous activity survive long enough to regulate evoked activity with event-based stimulation under a rate cap, with the bound verified. The breaking result would be eigenfunction drift on developmental timescales that outpaces any re-learning schedule, which would reduce this framework to a silicon-only tool.

Frequently asked questions

What is greedy spiking control?

A control scheme in which the actuator emits fixed-amplitude spikes, and each firing decision is made by greedily checking whether the spike's jump would decrease a chosen Lyapunov function by enough to be worth its cost. There is no continuous control signal, only spike timing.

What does the Koopman operator contribute?

It provides lifted coordinates in which nonlinear flow becomes linear and, for suitable observable dictionaries, spike-induced state jumps become exact matrix exponentials. This turns firing decisions into exact comparisons of quadratic functions in the lifted space, which is what makes a stability proof possible.

What does practical stability mean here?

Convergence to a bounded neighbourhood of the target rather than to the target itself, with the bound derived explicitly. It is the natural guarantee class for spiking actuation, since fixed-amplitude discrete kicks cannot pin a continuous system to a point.

How data-hungry is the design?

Surprisingly not: the pendulum design uses 300 unforced trajectories of two seconds each, with no actuated data at all. The eigenfunctions are fit by generator-based regression on the drift alone, and the Lyapunov weights come from a linear program.

Why does the energy eigenfunction matter so much?

The ablation answers this: zeroing the weight of the learned energy eigenfunction collapses the success rate from 99.94 percent to 42.5 percent. Swing-up is energy regulation first and unstable-mode cancellation second; without the energy coordinate the controller cannot tell whether the pendulum has been pumped high enough.

Can this be used to control an organoid?

Not yet directly. The proofs assume a fixed plant, full state feedback, and an exact jump model. Living tissue violates the first by plasticity, the second because only spikes through a limited electrode count are observable, and the third because stimulation effects are stochastic. The transferable part is the architecture: regulate an energy-like observable with rate-capped event-based stimulation, re-learned as the tissue drifts.

References

  1. N Meskin. Koopman Greedy Spiking Control for Nonlinear Systems under Firing-Rate Limits. arXiv:2610.41530. 2026. https://arxiv.org/abs/2610.41530. Accessed 2026-11-06.