The filter, not the spike generator, decides which burster you get
Bursting is the dominant signal phenotype of neural organoids, and it is usually treated as a unit of activity: more bursts, healthier tissue. Fyon, Brandoit, Mendolia, Sacre, Drion, Franci and colleagues at the University of Liege prove something sharper with a deliberately minimal model: in a spiking primitive built from one saturating nonlinearity in positive feedback around a linear frequency selector, the filter topology alone decides whether rest and repetitive firing can coexist, and therefore which of the classical bursting classes the system can produce at all. The nonlinearity only sets the threshold.
Source: Types of Bursting with a Two-Block Spiking Primitive, arXiv:2611.08045, 5 November 2026. Primary source. Read: the full 8-page PDF including all seven propositions, the two selector constructions, the numerical validations, and the discussion.
What the work claims
This is a theory paper in the control-theoretic tradition of Sepulchre, Drion and Franci: no new device, no new dataset, only a constructive classification, and it should be weighted accordingly. The object of study is a "two-block" excitable primitive: a static, monotone, saturating nonlinearity, whose slope at the origin is a gain k, closed in positive feedback around a stable linear filter H(s) that blocks constant signals (its DC gain is zero). Earlier work by some of the same authors showed this motif is sufficient for excitability;2 this paper asks which bursting classes it can reach, using two elementary selectors: a first-order high-pass filter and a second-order resonant band-pass filter.1
The central claim is a division of labour that holds across both selectors. The nonlinearity decides whether the system fires and where the rheobase sits: rest loses local stability exactly where the slope of the nonlinearity at the operating point equals one, independently of the selector. The selector decides whether rest and repetitive firing can coexist, and that coexistence is precisely what separates the burst classes. The high-pass filter never makes rest and firing coexist, whatever nonlinearity is used, so it can only support bursters that are swept across their own firing band, of which the parabolic class is the canonical case. The resonant band-pass filter does support coexistence, above a computable gain, and a single additional slow linear feedback loop then produces elliptic bursting. Square-wave bursting, which needs a bistable fast subsystem, is structurally unavailable in this architecture because the resting point is always unique.
How it works
The analysis is frequency-domain throughout. Because the filter blocks DC, the feedback contributes nothing at steady state, so the system has a unique resting point set by the input alone, and the feedback signal has zero mean both at rest and on any periodic orbit. The primitive is a Lur'e system, so global stability is read off a Nyquist plot with the circle criterion: below unit gain at the origin the resting point is globally asymptotically stable and nothing can happen at any input. The two selectors have Nyquist loci that travel the same circle but in characteristically different ways. The high-pass locus approaches unit gain only as frequency goes to infinity; the resonant band-pass reaches it at its centre frequency. That single geometric fact has a dynamical consequence: with the band-pass selector, spiking onsets at a finite frequency fixed by the filter alone, the defining signature of Hodgkin type II excitability, whereas the high-pass with a piecewise-linear saturation onsets at zero rate, which is type I.
Whether the onset is soft or hard is decided by the third derivative of the nonlinearity at threshold. For a tanh nonlinearity the Hopf bifurcation is supercritical for gains between 1 and 3/2 and subcritical above 3/2, and only the subcritical branch opens a window of inputs in which rest and a full-amplitude limit cycle coexist. That bistability window is what a slow feedback loop needs in order to burst: a first-order low-pass driven by the output and returned negatively to the input sweeps the operating point back and forth across the window, producing elliptic bursting in which the spike amplitude within a burst grows and collapses while the rate stays near the filter's centre frequency. The authors note that this bursting regime occupies a thin diagonal strip of parameter space whose width is inherited from the bistability window itself.
For the high-pass selector the route to bursting is different and, frankly, more surprising. No single additional linear slow branch can produce bursting at all: because rest and firing never coexist, the averaged slow flow is a scalar equation with a single globally attracting fixed point, so the system either rests or fires tonically. What works is a second copy of the same primitive, run on a slow timescale and biased inside its own firing band, whose relaxation cycle sweeps the fast copy across its firing threshold. The position of a switch inside this cascade selects the burst shape: reading the slow filter state gives the smooth ramp of a parabolic burst, while reading the slow output with the right bias produces abrupt onset and offset with a flat intra-burst rate, a trace that looks exactly like square-wave bursting even though the underlying mechanism is a threshold crossing on each side. The authors draw the moral explicitly: bursters cannot be classified from traces alone.
Where a skeptic should push
The most load-bearing assumption is that a linear selector plus one static nonlinearity is a fair skeleton of an excitable element. The paper is honest about the cost of that reduction. Everything is deterministic and low-dimensional; there is no noise, no channel kinetics, and no spatial extent. Real bursting neurons carry slow nonlinear currents, calcium-dependent feedback chief among them, and those escape the negative results: the proof that one linear slow branch cannot burst applies to linear branches, and a rectified or otherwise nonlinear branch can tell rest from firing even where a linear one is blind. The uniqueness of the resting point rules out square-wave bursting within this architecture, but adding a nonmonotone nonlinearity would restore the fast-subsystem bistability the architecture lacks, so "structurally unavailable" means unavailable here, not in nature.
There is also a fragility worth flagging. The elliptic regime lives in a thin strip of parameter space, and the numerical validations, while clean, are demonstrations of consistency rather than exploration of robustness. The cascade construction, meanwhile, is minimal in block count but arguably not in insight: a textbook parabolic burster is two-dimensional, and the two-copy cascade, with its switch and two gains, is not obviously cheaper to build or to tune once analogue imperfections enter. Finally, the tantalizing remark that neuromodulation retunes excitability type and burst class in biology, and that such retuning "remains to be investigated" in the primitive, is exactly the gap a sceptic would want closed before calling this a design procedure rather than a taxonomy.
Why burst shape misleads organoid readouts
The non-obvious implication for organoid intelligence is that burst shape is a filter property, not a firing property. A cortical organoid's spontaneous activity is read almost exclusively through bursts: burst rate as a maturation marker, burst duration as a health metric, burst synchrony as a network coupling measure. This paper says those traces carry a hidden variable. Which burst classes a tissue can express is decided by its subthreshold frequency-selective dynamics, that is, by the ion-channel composition that sets membrane resonance and adaptation, not by how excitable the cells are. Two organoids can emit statistically indistinguishable burst trains while sitting in different bifurcation regimes, doing different computations, and responding differently to the same stimulation. The paper's own cautionary example is the cleanest: a burster that looks square-wave in the trace can be a simple threshold-crossing device underneath. Any decoder, plasticity rule, or closed-loop stimulation protocol that assumes a bifurcation structure the tissue does not have is training on a fiction.
The opportunity is that excitability type becomes a measurable, and potentially controllable, specification for organoid computing. The paper hands over concrete diagnostics: measure the rate-input onset and see whether firing starts at arbitrarily low rate (type I) or at a finite frequency set by the substrate (type II); measure hysteresis in the rate-input curve to detect a bistability window; check whether intra-burst modulation lives in spike rate or in spike amplitude, which the paper shows is the visible fingerprint of threshold-crossing versus Hopf mechanisms. These are exactly the measurements multi-electrode arrays and patch clamp can deliver, and they convert "is this organoid a good computer" from a vibes question into a classification. The further opportunity the authors leave open is the one organoid systems are best placed to exploit: biology retunes burst class with neuromodulation, and a tissue-plus-closed-loop-pharmacology rig is a natural instrument for doing the same deliberately, steering a substrate between rate-coding and timing-coding regimes.
The threat is quieter. Reservoir-style computing on organoids trains a readout on whatever dynamics the tissue happens to produce; if developmental variability across batches silently swaps the effective filter bank, the computation the readout learned stops being the computation the tissue is doing, with no visible change in the activity anyone monitors. The paper also undercuts a lazy equivalence that pervades the field: "bursting organoid" is not one kind of machine, and treating it as one kind of machine will produce irreproducible results that get blamed on biology when the real culprit is an unmeasured bifurcation parameter.
The bottom line
Established: within a two-block spiking primitive, the selector topology provably decides the reachable burst classes, with closed-form thresholds, firing rates, and criticality conditions, validated numerically. Hypothesis, untested anywhere: that real excitable tissue, organoid tissue included, is well-described by this skeleton. The claim would be confirmed for a given preparation if onset frequency, rate-input hysteresis, and intra-burst amplitude versus rate modulation line up with the predicted selector signature; it would be broken by robust square-wave-like hysteresis in tissue, which would imply nonlinear slow feedback of a kind the architecture excludes. For organoid intelligence the paper's practical legacy is a discipline: classify the bifurcation before you train the decoder.
Frequently asked questions
What is a two-block spiking primitive?
It is a minimal excitable circuit: one static saturating nonlinearity, whose slope at the origin sets a gain k, closed in positive feedback around one stable linear filter whose DC gain is zero, so constant components are blocked. The filter transmits only variations of its input, and the feedback loop amplifies a selected band of them until the nonlinearity saturates, producing an all-or-none excursion, a spike, followed by autonomous return to rest.
Which bursting classes can the primitive produce?
Two of the three standard classes. With a first-order high-pass selector the system is monostable and supports parabolic bursting, where the intra-burst spike rate follows a smooth rise and fall, generated either by an external slow drive or by a second slow copy of the primitive. With a second-order resonant band-pass selector and a gain above 3/2 with a tanh nonlinearity, rest and firing coexist and a single slow linear feedback loop produces elliptic bursting, where spike amplitude rather than rate is modulated within the burst. Square-wave bursting is provably unavailable in this architecture.
What decides between type I and type II excitability here?
The selector decides. The high-pass filter reaches unit loop gain only at infinite frequency, so repetitive firing can onset at an arbitrarily low rate, which is type I behaviour with a piecewise-linear saturation. The resonant band-pass reaches unit gain at its centre frequency, so spiking onsets abruptly at that finite frequency, the signature of type II. The nonlinearity, meanwhile, sets where threshold sits, and its third derivative at threshold decides whether the onset is continuous or subcritical.
Why is square-wave bursting impossible in this architecture?
Square-wave bursting requires the fast subsystem to be bistable over a range of the slow variable: two stable states, rest and firing, coexisting. In the two-block primitive the filter blocks DC, so at any constant input there is exactly one equilibrium, set by the input alone. The fast subsystem can therefore never be bistable, and no slow feedback, linear or otherwise, can manufacture the hysteresis square-wave bursting needs.
Can this be tested on an organoid?
Yes, in principle with existing tools. Patch clamp or dense microelectrode arrays can measure rate-input curves: a continuous onset at zero rate indicates type I, an abrupt onset at a fixed frequency indicates type II, and a hysteresis loop between the up and down sweeps indicates a bistability window. Intra-burst analysis, checking whether the modulation lives in spike rate or spike amplitude, then discriminates threshold-crossing from Hopf mechanisms. No organoid study we know of reports these measurements as a classification.
What are the theory's main limits?
It is deterministic and low-dimensional, with no noise, no spatial structure, and no channel-level detail. Several striking results, including the impossibility of bursting from a single linear slow branch, hold only for linear slow feedback; the slow currents of real neurons are nonlinear and can escape them. And the elliptic bursting regime occupies a thin strip of parameter space, so its practical robustness is an open question the paper flags itself.
References
- A. Fyon, J. Brandoit, L. Mendolia, P. Sacre, G. Drion, A. Franci, C. Fernandez Lorden, and G. Bainier. Types of Bursting with a Two-Block Spiking Primitive. arXiv:2611.08045 [eess.SY]. 2026. https://arxiv.org/abs/2611.08045. Accessed 2026-11-07.
- G. Bainier, A. Fyon, C. Fernandez Lorden, and A. Franci. Excitability of high-frequency positive feedback. arXiv:2611.06822 [eess.SY]. 2026. https://arxiv.org/abs/2611.06822. Accessed 2026-11-07.