The spike-amplitude clause in closed-loop organoid control
A hybrid-systems control paper proves that one integrate-and-fire-like neuron, emitting spikes whose amplitude is proportional to the plant output at firing time, can asymptotically stabilize any static-output-feedback-stabilizable linear single-input single-output plant. Fixed-amplitude neuromorphic controllers, the kind every spiking hardware platform implements, can only guarantee practical stability: a persistent oscillation inside a bounded error zone. The actuator choice, it turns out, is a theorem-level decision, not an implementation detail.
Source: Proportional Neuro-Impulsive Control of SISO LTI Systems, arXiv, 2026. Primary source. Read: full text, 8-page PDF including proofs and appendix.
What the work claims
This is a theory paper with a simulation demonstration, from Christian Fernandez Lorden and Alessio Franci at the University of Liege, Elena Petri and Maurice Heemels at Eindhoven University of Technology, with Franci also at the WEL Research Institute. The authors define what they call a SISO LTI neuro-impulsive control system: a plant with state x and output y = Cx, closed in feedback with a single neural controller whose internal variable xi integrates the plant output plus a constant bias, xi-dot = Cx + I_b, until it hits a threshold Delta, at which point the plant state jumps to x-plus = (I minus alpha B C)x and the neuron resets to zero1.
The central claim is Theorem 11: for any such system whose plant can be stabilized by static output feedback, and for every threshold Delta greater than zero, there exist bias current and spike gain values such that the closed loop is semi-globally asymptotically stable. Semi-global means stability holds for every compact set of initial conditions, with controller parameters allowed to depend on that set. This is strictly stronger than what fixed-amplitude neuromorphic controllers deliver: those provably cap out at practical stability, an ultimate bounded error that oscillation makes unavoidable, because near the target a fixed-size correction always overshoots and far from the target the same small correction must be repeated at ever higher frequency1.
The price is stated honestly in the abstract and again in the conclusion: the spikes must be graded, amplitude-modulated, which the authors concede gives up the main practical advantage of neuromorphic control, its fixed-pulse simplicity. A second claim is constructive, an algorithm that turns a stabilizing continuous feedback gain and a set of admissible initial conditions into explicit controller parameters, tested on an inverted pendulum linearized around its upright equilibrium1.
How it works
The proof strategy is emulation. A static output-feedback loop x-dot = (A minus BKC)x with stabilizing gain K defines a Lyapunov function; the spiking controller is designed so that each flow-jump cycle imitates one step of that continuous loop. The link is the effective gain alpha over T(x), where T(x) is the dwell time between spikes: the ratio acts as the static gain the spiking controller emulates while flowing from a post-jump state. The bias current I_b bounds the maximum dwell time by Delta over I_b, so the controller can never go silent, and the first-order analysis shows the flow-jump transition matrix agrees with the continuous feedback exponential when T(x) is small1.
Why is this nontrivial? Because neither ingredient is contractive on its own. The jump map (I minus alpha B C) always has 1 as an eigenvalue when the plant order exceeds one, since BC is rank one, so the jump can never be a contraction; between spikes the plant follows its open-loop flow, which is unstable by assumption. Stability exists only in the combined flow-jump cycle, and the authors analyze it through the impact map, which takes the state immediately after one jump to the state after the next. A sequence of lemmas bounds the dwell time from above and below, bounds the deviation of the emulated gain from the reference gain K, and exhibits a range of (alpha, T) pairs in which the transition matrix contracts. Theorem 11 assembles these into parameter existence, and the figure-1 algorithm makes the construction mechanical: choose a gain K, solve a Lyapunov equation, find the worst-case initial condition, then pick I_b, Delta and alpha in that order1.
One elegant detail: because the proportional jump already carries the sign of the output y, a single neuron suffices for two-sided control. Fixed-amplitude controllers need at least two neurons, one per sign, since every spike pushes the same way1.
The numerical example makes the recipe concrete. The pendulum parameters are p1 = 1, p2 = 2, p3 = 1, p4 = 1, p5 = 2, p6 = 1, the stabilizing gains form the interval (2, 18), and the design uses K = 2.8 over a set whose largest element is (0.5, 0.5, 0). The algorithm returns I_b = 1251.82, Delta = 3.55 and alpha = 0.00794. Simulated from (2.5, 2.5, 0), five times outside the design set, the trajectories still converge, closely tracking the continuous loop they emulate. A relaxed controller that divides I_b by 1500 and Delta by 2 emulates the same gain at the equilibrium but violates the proof bounds, and it diverges quickly; it does lower the actuation frequency at the origin by a factor of 750, which is precisely the temptation the theorem warns against1.
Where a skeptic should push
The load-bearing assumption is the plant class. Everything is linear, time-invariant, single-input single-output, and stabilizable by static output feedback, a class real biological tissue is not a member of. Organoid dynamics are nonlinear, time-varying, and only locally linearizable, and the proof's semi-global guarantee explicitly depends on knowing a compact set of admissible initial conditions, which for living tissue is not a design input but an empirical unknown. The authors do not overclaim here, but a reader looking for a stability certificate to staple onto a wetware loop should notice that the theorem's hypotheses exclude exactly the properties that make tissue interesting.
Second, the graded spike is the entire contribution and the entire problem. The paper itself flags that amplitude-proportional pulses sacrifice the main practical advantage of neuromorphic control, and its proposed fallback, arrays of fixed-amplitude neurons working at different output scales so corrections quantize but keep shrinking, is an existence argument resting on motor-unit recruitment rate coding, not a demonstrated actuator. Third, the evaluation is simulation-only: there is no hardware pendulum, no noisy channel, no model error beyond the relaxed-parameter probe. And fourth, a disclosure worth noting in a peer-reviewed reading: the acknowledgments state that Claude Opus 5 and 5.5 were used to edit phrasing throughout and to draft initial versions of some passages in Section I, with the authors asserting that all technical content, results and proofs are their own. The mathematics here is checkable, and the proof structure reads as human-assembled, but the disclosure is a fact about the paper's provenance that a skeptical reader should hold.
Graded spikes and the organoid control contract
For organoid intelligence the paper reads as an actuator audit for closed-loop tissue control. Every serious organoid-computing program eventually wants the same loop: sense population activity, compare it to a desired regime, stimulate to correct. That loop's actuator is almost always fixed-amplitude, electrical pulses of set duration or optogenetic flashes of set intensity. This theorem states, in a precisely bounded setting, what that choice buys and what it forecloses: fixed-amplitude pulse feedback can at best hold tissue inside an error band, oscillating around the target, while proportional feedback can truly converge. Anyone designing a closed-loop stimulation protocol to pin an organoid's firing rate or suppress an emerging burst should read practical stability as a formal statement of the chasing-the-setpoint behavior these systems exhibit, not as a tuning failure1.
The constructive half of the paper transfers more directly than the theorem. The design recipe, pick a continuous feedback law, solve a Lyapunov equation, then choose threshold, bias and gain so that each flow-jump cycle emulates that law with bounded dwell time, is a template for converting a desired firing-rate regulation law into a pulsed stimulation protocol whose parameters come with a derivation rather than a grid search. Two ingredients deserve attention in wetware: the dwell-time bound Delta over I_b is a guaranteed minimum stimulation cadence, which in tissue maps onto charge-injection and duty-cycle safety constraints; and the sign-carrying proportional jump means a single stimulation channel implements two-sided control, halving electrode count relative to fixed-pulse designs1.
The threat is twofold. The shallow one is hype-correction: a stability certificate for a linear plant is not a stability certificate for tissue, and the field's appetite for borrowed authority could easily outrun the theorem's hypotheses. The deeper one is the paper's own trade-off made visible in biological terms. Tissue already implements graded output through recruitment, populations of neurons with heterogeneous thresholds contributing incrementally, the motor-unit principle the authors cite as their workaround. So the result quietly undercuts a design orthodoxy of the neuromorphic side, that binary fixed-amplitude spikes are the efficient primitive, while vindicating a property living substrates have natively. If graded amplitude is what asymptotic guarantees require, the interface question for organoid computing shifts from how many spikes to how many distinguishable amplitudes, and pulse-width or intensity resolution becomes a first-class specification of the acquisition chain, not an afterthought.
The bottom line
Established: within the hybrid-systems framework, a single integrate-and-fire-like neuron with output-proportional spikes renders any static-output-feedback-stabilizable SISO LTI plant semi-globally asymptotically stable, with a constructive parameter-selection algorithm demonstrated on a linearized inverted pendulum using I_b = 1251.82, Delta = 3.55, alpha = 0.00794, and simulated convergence from five times outside the design set. Also established, by the same authors' analysis: fixed-amplitude spiking controllers provably cannot do better than practical stability for the same class. Not established: hardware validation, robustness to noise or model error, any extension beyond linear SISO plants, or a practical graded-amplitude actuator. For organoid intelligence, the durable contribution is a clean separation of what your stimulation actuator mathematically permits: bounded chasing with fixed pulses, true convergence with graded ones. What would break the claim is a demonstration, in simulation or hardware, that the stability margin survives the moment the plant stops being linear.
Frequently asked questions
What is a neuro-impulsive controller?
A hybrid controller combining neuromorphic and impulsive-control ideas: a single integrate-and-fire-like neural variable integrates the plant output plus a bias current until a threshold is reached, then the plant state is updated by a jump proportional to the output at firing time and the neuron resets to zero.
What does semi-global asymptotic stability mean here?
That for every compact set of initial conditions, controller parameters exist making the closed loop converge to the origin, with the origin attracting all solutions started inside that set. It is weaker than global stability, which the paper does not claim, because the parameters may depend on the initial-condition set.
Why can fixed-amplitude spike controllers not match this?
Near the target, a fixed-size correction always overshoots, so oscillation inside an error bound is unavoidable, giving only practical stability. Far from the target the same small correction must fire ever faster to matter, and those two requirements pull against each other. Proportional spikes remove the first problem by shrinking with the output.
What did the pendulum example show?
For a linearized inverted pendulum with stabilizing gains in (2, 18), the design algorithm returned bias 1251.82, threshold 3.55 and gain 0.00794. Simulations from (2.5, 2.5, 0), five times outside the design set, converged and tracked the emulated continuous loop, while a relaxed parameter set outside the proof bounds diverged despite emulating the same equilibrium gain.
What is the catch for biological substrates?
The proof covers linear time-invariant single-input single-output plants only. Living neural tissue is nonlinear and time-varying, so the theorem is a design principle for stimulation protocols, with its dwell-time cadence bounds and sign-carrying single-channel control, rather than a transferable stability certificate.
References
- Fernandez Lorden, C., Petri, E., Heemels, W. P. M. H., Franci, A. Proportional Neuro-Impulsive Control of SISO LTI Systems. arXiv:2611.13043 [eess.SY], 2026. https://arxiv.org/abs/2611.13043. Code stated available at github.com/ChristianFernandezLorden/ECC27-FernandezLorden-SISONeuroImpulsive. Accessed 2026-11-18.