A small-gain theorem for spiking feedback loops
When an excitable system - a neuron, an organoid, an event-based sensor - is wired in feedback with another, does the closed loop settle into a stable rhythm, and can you predict that rhythm without simulating the full nonlinear dynamics? Cecconi, Bin, Marconi, and Sepulchre answer yes for the event-based regime: if the two systems' timing gains sum to less than one, a unique one-spike-per-event rhythm exists, is robust to jitter, and its period is the solution of a scalar equation you can measure open loop.
Source: Event-based analysis of neuromorphic feedback systems, arXiv preprint 2611.13832, submitted 10 Nov 2026. Primary source. Read the full 8-page preprint PDF.
What the work claims
This is a theory paper, submitted to the European Control Conference 2027, and it should be weighted as such: the contribution is a theorem plus an illustrating simulation, not a hardware demonstration. The setting is a feedback interconnection of two systems that each evolve by continuous internal dynamics but interact only through discrete events - spikes, threshold crossings, actuator switches. The running example is a conductance-based neuron coupled to a damped pendulum through an excitatory and an inhibitory synapse: two systems that are each non-oscillatory in isolation, yet whose interconnection can sustain a limit cycle.1
The central claim has three parts. First, each subsystem can be represented as an event operator: a causal, time-invariant map from the history of incoming inter-event intervals to the delay of its next outgoing event. Second, if both operators have fading memory - the influence of old inputs decays, quantified by summable weights - then a small-gain condition, the sum of the two systems' incremental timing gains being less than one, guarantees that the closed loop has a unique rhythm that is robust to additive timing perturbations such as transmission jitter.1 Third, in the unperturbed case that rhythm is periodic with one event from each system alternating, and its period is determined by a scalar balance equation whose ingredients are measurable by forcing each subsystem in open loop.1
A useful converse comes free: because each system's periodic response curve has slope bounded by its gain, if the magnitudes of the two curves' slopes at the crossing sum to at least one, the small-gain condition cannot hold - the rhythm is provably outside the theorem's reach.1
How it works
The trick is to cut the loop at the synapses. A synapse is modeled as an event readout, which extracts a discrete event from a continuous presynaptic signal, followed by a synaptic filter whose state resets at each event and which generates the physical postsynaptic input. Across that interface, all information is carried by event times.1 Time invariance then implies that the delay of an output event depends only on the relative timing of past inputs: writing Tk for the k-th inter-event interval, the delay obeys a memory functional Δk = ψ(Tk-1, Tk-2, ...). The analysis lives entirely in this event space - no state-space model of the coupled nonlinear system is ever constructed.1
The regime is one-to-one entrainment: input events arrive strictly sparsely enough that each triggers exactly one output event before the next input arrives. Within this regime the authors assume fading memory in the quantitative sense that the difference in delays between two histories is bounded by a weighted sum of interval differences with summable weights; the incremental gain γ of the operator - the worst-case amplification of a timing difference - is at most the sum of those weights.1
Crucially, the gain is not a fixed constant of the hardware: it is shaped by event sparsity. For a first-order lag with time constant τ followed by a threshold readout, the memory functional is ψ(v) = τ ln of (b over yth divided by the sum of exponentials of past interval ages), the periodic response is Δ(T) = τ ln[(b/yth)/(1 - e^(-T/τ))], and the local gain is γloc(T) = 1/(e^(T/τ) - 1).1 That gain decays exponentially as events become sparser than the system's recovery time: space your events and the system forgets faster than the loop can amplify, which is precisely what makes the small-gain condition achievable. The same logic extends to sparsely triggered excitable systems, which are contractive between well-separated events.
The main result, the event small-gain theorem, states that if γp + γn < 1, the loop equation has exactly one solution for any admissible timing perturbation, that solution depends causally on the perturbation, and the deviation caused by a perturbation is bounded by max of the two gains divided by (1 - γp - γn) times the perturbation size.1 With no perturbation, shift invariance forces the rhythm to be periodic, with a period T* solving the scalar balance T* = Δp(T*) + Δn(T*): the period must equal the sum of the delays that same period induces in the two subsystems. Each Δi is obtained by periodically forcing that subsystem in open loop - the closed-loop prediction falls out of two open-loop measurements and one scalar root find.1
In the validation simulation, a neuron firing by post-inhibitory rebound coupled to the pendulum settles within a few cycles onto the predicted one-to-one rhythm. The balance equation gives T* = 6.039 (in the paper's time units), split into a neuronal delay of 0.945 and a pendulum delay of 5.094; the response-curve slopes at the crossing are 1.1 x 10^-4 for the neuron and 0.29 for the pendulum, summing to 0.29, comfortably below one and consistent with the necessary condition.1
Where a skeptic should push
The load-bearing assumption is one-to-one entrainment itself. Everything is proved inside the regime where every input event produces exactly one output event and the intervals stay within a compact band; the theorem says nothing about what happens when the loop drifts outside it, which is where real systems spend much of their lives. A burst - several output events per input - a skipped response, or an interval outside the admissible band voids the premises silently. The one-to-one regime is assumed, not derived, for the general cascade; only for specific systems like the first-order lag does the paper derive the dwell-time condition that keeps the regime valid (the input period must exceed the time constant by a logarithmic factor).1
Second, fading memory and bounded gain are assumptions on the operator, and tissue violates them cheerfully. Synaptic plasticity, adaptation currents, and slow metabolic modulation all create history dependence with timescales of minutes to hours; nothing in the summable-weights condition forbids a very long tail, but the gain bound becomes correspondingly weak, and with it the robustness guarantee, since the jitter bound degrades as gains approach the critical sum of one. The analysis is also noise-free: the perturbations are bounded and adversarial, not stochastic, and event detection in biological settings is timing-jittered at the millisecond scale by physiology itself.
Third, the validation is a single simulation of one synthetic loop. The neuron-pendulum example confirms the balance equation reproduces the simulated period, which is the theorem's least surprising consequence. There is no experimental test on neuromorphic silicon, let alone on living tissue, and the paper does not claim one. As with any small-gain result, the practical value will stand or fall on whether the gains of real systems can be identified from data well enough to check γp + γn < 1 in the first place.
When organoid-silicon loops lock or break
The non-obvious implication for biological computing is that this theorem quietly describes the architecture of every bidirectional brain-computer interface, including the ones aimed at organoid intelligence. A dish of living neural tissue coupled to a silicon controller through an MEA is exactly the paper's object: two excitable systems with fading memory, interacting only through discrete events (stimulus pulses in one direction, detected spikes in the other). The field has been closing such loops empirically and hoping. What this paper contributes is a design grammar for when the loop has a stable identity at all.1
Three consequences are concrete. First, a tunability result: because each side's timing gain decays exponentially with event spacing relative to its recovery time constant (γloc = 1/(e^(T/τ) - 1) for the lag case), sparse stimulation is not merely gentler on tissue, it is the mechanism that keeps the loop in the contractive regime where a stable entrainment rhythm exists and is robust to jitter. Pushing the interface to high event rates does not just risk habituation - it raises the loop gain and can destroy the unique-rhythm property the system designer was counting on.1 Second, a measurement recipe: the entrained period, and hence the phase relationship the controller will settle into, can be predicted from two open-loop characterisations plus a scalar root solve, which is a tractable system-identification program on real MEA hardware. Third, a falsification test: if the sum of the measured response-curve slopes at the crossing is at least one, no stable one-to-one rhythm of this kind can exist - a negative result obtainable before anyone builds the closed loop.1
The threats are honest ones. Organoids, unlike the paper's sparsely triggered neuron, fire dense bursts in which the one-to-one assumption fails by construction; the theorem then delineates - rigorously, for once - the regime the system has left, rather than describing it. Plasticity violates fading memory on long timescales, so any entrainment a tissue-computer loop achieves is leased, not owned, and must be re-qualified as the substrate adapts. And there is a subtle hype-correction here: the theorem guarantees a rhythm's existence and robustness, not its usefulness - a locked loop that exchanges one spike per cycle at a stable period is the biological-computing equivalent of a clock signal, not a computation. The value of the framework is that it makes that distinction provable.
The bottom line
Established, within the paper's assumptions: a rigorous small-gain theorem in the event domain guaranteeing existence, uniqueness, and robustness of a one-to-one periodic rhythm for two event-coupled systems whose timing gains sum below one, with the rhythm's period fixed by a scalar balance equation built from open-loop measurements, and a slope-based necessary condition that rules the regime out.1 Not established: behavior outside the one-to-one entrainment regime, applicability to bursty or plastic biological tissue, noise performance beyond bounded adversarial jitter, or any experimental validation beyond one simulation. What would confirm relevance: identified timing gains and describing functions from a real neuron-silicon or organoid-silicon loop that satisfy the small-gain condition and reproduce a measured entrained period. What would break it: real interfaces whose event statistics force them outside one-to-one entrainment, where the theory as stated is silent.
Frequently asked questions
What is an event operator?
A causal, time-invariant map that describes a physical system purely in terms of its discrete events: it takes the history of incoming inter-event intervals and outputs the delay until the next outgoing event. It discards the continuous internal dynamics and keeps only the timing information that crosses the interface.
What does the small-gain condition mean here?
Each system amplifies timing differences by at most its incremental gain γ. If the two gains sum to less than one, a timing perturbation shrinks each trip around the loop, so the loop converges to a single rhythm. If they sum to one or more, the theorem gives no guarantee and its necessary slope condition says the rhythm regime is unattainable.
How can the loop period be predicted without simulating it?
Periodically force each subsystem in open loop and record its steady-state response delay as a function of forcing period - its event describing function. The closed-loop period is the value where the sum of the two describing functions crosses the diagonal: T* = Δp(T*) + Δn(T*). The authors' neuron-pendulum example reproduces the simulated period (6.039) this way.
What is one-to-one entrainment?
The regime where every input event triggers exactly one output event before the next input arrives, with inter-event intervals bounded in a compact band. The theorem is proved inside this regime; bursts, skipped responses, or out-of-band intervals fall outside it and are not covered.
Why does event spacing matter so much?
For a first-order lag with time constant τ, the timing gain is 1/(e^(T/τ) - 1), which decays exponentially as the interval T grows past τ. Sparse events let each system fully relax between events, killing memory of the past and lowering the loop gain - sparsity is what makes the stability condition reachable.
Does this apply to organoids on MEAs?
Conceptually the organoid-silicon loop is the paper's object: two excitable systems coupled by discrete events. But organoids fire dense bursts rather than one spike per event, violating one-to-one entrainment, and plasticity strains fading memory. The framework applies most cleanly to sparse, stereotyped interfaces, and otherwise marks out the regime the real system has left.
Is a locked rhythm a computation?
No. The theorem guarantees a stable periodic exchange of events - a clock, not a calculation. Its usefulness for biological computing is as systems theory: telling designers when a stable closed loop exists, what period it will settle at, and when no such loop is possible, before hardware is committed.
References
- A. Cecconi, M. Bin, L. Marconi, and R. Sepulchre. Event-based analysis of neuromorphic feedback systems. arXiv preprint 2611.13832. 2026. https://arxiv.org/abs/2611.13832. Accessed 2026-11-15.