A solvable model of a stochastic spiking neuron, with caveats
The FitzHugh-Nagumo model is the minimal sketch of an excitable neuron: one fast variable that spikes, one slow variable that recovers, and noise that decides when the jump happens. It is usually simulated and rarely solved. Maroulakos, Wal, Mishra, Toklikishvili and Chotorlishvili find an exact spectral solution for a frozen slow variable, using a supersymmetric factorisation borrowed from quantum mechanics, and extract both the relaxation rate of the noisy bistable neuron and a two-state coherence diagnostic. Just as instructive is what the same analysis breaks: during a genuine spike the trajectory leaves the bistable domain entirely, and the two-state picture, the coherence measure, and any quantum-flavoured reading all stop applying, a boundary the authors police carefully.
Source: Supersymmetric Neuromorphic FitzHugh-Nagumo model, arXiv:2611.03312, 3 November 2026. Primary source. Read: the full 12-page PDF including the normal-form reduction, the spectral construction, both illustrative examples, and the conclusions.
What the work claims
This is a theory paper, purely analytical with illustrative calculations, motivated by an effective model of antiferromagnetic spintronic devices, and it should be read as mathematics about neurons rather than neuroscience. The central claim is that a local normal-form reduction of stochastic FitzHugh-Nagumo dynamics, with the slow recovery variable frozen, yields a one-dimensional Fokker-Planck equation in a quartic potential whose generator is related by a similarity transformation to a positive semidefinite supersymmetric Hamiltonian. That factorisation has a concrete payoff: the first nonzero relaxation eigenvalue, the rate at which a bistable neuron forgets which well it started in, equals the ground-state eigenvalue of a partner operator, and in the weak-noise bistable regime it reduces to the sum of the two Kramers escape rates, with each rate set by the curvatures of its well and barrier and by the barrier height.1
The second claim is diagnostic. Projecting the similarity-transformed probability density onto its two lowest modes defines a normalised auxiliary two-state amplitude, and the authors derive a coherence-like quantity in a basis localised on the two wells, expressed through the basin populations and their equilibrium weights. Its maximum value of one over root two is not reached at equal populations in general, but when the population ratio equals the square root of the ratio of equilibrium weights; at equilibrium the localised-basis value is root two times the square root of the product of the equilibrium weights, while the spectral-basis value vanishes. The authors are explicit that this coherence characterises an auxiliary amplitude representation of classical stochastic dynamics and does not by itself establish microscopic quantum coherence.
How it works
The construction starts from the dynamics of a Néel vector, the antiferromagnetic order parameter, driven by spin torque in an antiferromagnet-heavy-metal bilayer. A tangent-space reduction maps the magnetic control parameters onto the canonical FitzHugh-Nagumo coordinates: the fast activation variable and the slow recovery variable, with the drive current acting as a bias that tilts the cubic-nullcline landscape. Thermal and interfacial fluctuations enter as additive noise, producing a stochastic differential equation whose probability density obeys a Fokker-Planck equation. Because the full two-dimensional drift is non-gradient, the factorisation is applied only after freezing the recovery variable, which reduces the problem to motion in a quartic potential that depends parametrically on the frozen slow coordinate and the bias.
The supersymmetric step writes the Fokker-Planck generator as a similarity transform of a Hamiltonian that factorises into a product of a supercharge and its adjoint, the same structure as supersymmetric quantum mechanics. All nonzero eigenvalues then occur in pairs, and the slowest relaxation mode, the one that governs interwell equilibration, is obtained from the partner Hamiltonian's ground state. In the weak-noise bistable regime this gives the Kramers form: each direction's rate is the square root of the product of well and barrier curvatures over 2 pi, times the exponential of minus the barrier height in units of the noise strength, and the total relaxation rate is the sum of the two directions. The two-state amplitude and its coherence diagnostic follow from projecting the transformed density onto the lowest pair of modes and rotating to a basis localised on the two wells, an orthonormal rotation whose interpretation as left-well and right-well states is itself a weak-noise approximation.
Where a skeptic should push
The most load-bearing assumption is the adiabatic freezing of the recovery variable, and the paper itself tells you where it strains. In the fixed-recovery example, monotonic relaxation between wells produces a pronounced coherence pulse while the ensemble mean settles without a single spike: the diagnostic fires on a non-event. In the spiking example, the trajectory repeatedly crosses the folds at two-thirds offset from the bias, leaves the bistable domain, and passes through regions where the wells are not separated by any meaningful barrier, so the localised two-state description does not apply uniformly along the cycle, and a deterministic waveform alone cannot supply the ensemble populations the diagnostic needs. A coherence reading is therefore invalid exactly during the spiking excursions that are the model's reason to exist. Extending the framework to an evolving recovery variable requires fast intrawell equilibration, weak recovery noise, and well-separated metastable modes, a regime real neurons and organoid tissue may rarely occupy.
Second, everything quantitative lives in the weak-noise bistable regime; the Kramers form is an asymptotic approximation, not a general identity. Third, the antiferromagnetic motivation is an effective-model bridge: the map from spin-torque parameters to FitzHugh-Nagumo parameters is established by local normal-form optimisation, with two parameter sets tabulated, but no device data appear anywhere in the paper. The work supplies exact spectra for an idealised neuron and an honest delimitation of that idealisation; it does not supply anything measured.
Spectral diagnostics for organoid intelligence
For organoid intelligence the useful idea is that excitability has a spectral signature, and that the slowest relaxation eigenvalue of a perturbed network encodes barrier heights in its landscape of activity states. The practical picture: stimulate an organoid, watch the evoked activity decay, and fit the slowest relaxation rates; if the underlying dynamics are even approximately bistable with weak effective noise, the rate structure is a Kramers form, and its temperature-like parameter and barrier heights become a non-invasive readout of the tissue's excitability state, sensitive to maturation, pharmacology, or inhibitory balance without fixing a single cell. Organoid labs already collect exactly this kind of perturb-and-decay data; the paper hands them a theoretical dictionary for reading it, plus a warning about when the dictionary pages are blank.
The threat is the paper's second lesson, and it cuts at current practice. Reduced binary descriptions of organoid dynamics, quiescent versus bursting, well one versus well two, are the working currency of reservoir-style readouts and state decoders. The analysis shows that such a two-state picture is least valid precisely during the transitions that carry the most information, because a spiking trajectory exits the bistable domain and the barriers separating the states cease to exist along the path. Any decoder that assumes the organoid is always in one of two metastable states will misread the escape events themselves, which is where computation-by-transitions schemes actually live. The correction is not to abandon reduction but to confine it to the quasi-stationary segments and treat crossings with a different model class, or with direct stochastic simulation as the authors recommend.
There is also a quiet governance value in the paper's restraint. In a literature where coherence language around biological and neuromorphic substrates invites quantum interpretations, the authors state plainly that their coherence is a classical auxiliary amplitude, basis-dependent, with the spectral-basis value vanishing at equilibrium while the localised-basis value does not. That separation of representation from physics is a template worth importing into organoid intelligence, where measures computed on spike trains or calcium traces are routinely reified into claims about the tissue's internal states. The opportunity is a spectral thermometer for excitable tissue and a rigorous map of where reduction fails. The trap the paper warns against is measuring the thermometer during the storm it cannot describe.
The bottom line
Established, analytically: for the frozen-recovery stochastic FitzHugh-Nagumo model, the relaxation spectrum factorises supersymmetrically, the slowest nonzero rate equals a partner-operator ground-state eigenvalue and takes Kramers form in the weak-noise bistable limit, and a basis-dependent two-state coherence diagnostic follows with its optima characterised. Not established: any of this during genuine spiking cycles, with an evolving recovery variable, outside the weak-noise regime, or in connection to measured device or tissue behaviour. The confirming result for the organoid programme would be perturbation-decay recordings from organoid networks whose fitted slow relaxation rates obey the predicted curvature-and-barrier structure across pharmacologically controlled excitability states. The breaking result would be relaxation spectra dominated by single events that exit the bistable domain, which is what the theory itself predicts spiking tissue will do.
Frequently asked questions
What is the FitzHugh-Nagumo model?
The minimal qualitative model of an excitable neuron: a fast activation variable with cubic nonlinearity that generates spikes, and a slow recovery variable that resets it. Noise decides when an excursion fires, which makes it the standard reduced model for studying stochastic spiking.
What does supersymmetric mean in this paper?
It is borrowed from quantum mechanics: the Fokker-Planck generator, after a similarity transformation, factorises into a supercharge times its adjoint, making the effective Hamiltonian positive semidefinite and pairing all nonzero eigenvalues. The slowest relaxation rate of the noisy neuron is then a partner operator's ground-state eigenvalue.
Does the paper claim quantum coherence in neurons?
No, and it says so explicitly. The derived coherence is an off-diagonal norm of a classical auxiliary two-state amplitude, it depends on the chosen basis, and it does not by itself establish microscopic quantum coherence in the magnetic device or in neurons.
Where does the two-state description fail?
During genuine spiking. A spiking trajectory repeatedly leaves the bistable domain through the folds, where wells are not separated by meaningful barriers, so the localised two-state approximation does not hold along the excursion, and ensemble populations, which the diagnostic needs, cannot come from a deterministic waveform alone.
How could this be used on real organoids?
As a spectral reading of perturb-and-decay experiments: the slowest relaxation rates of evoked activity would encode effective barrier heights of the tissue's activity landscape, giving a non-invasive monitor of excitability state, provided the tissue sits near a weak-noise bistable regime during the measurement.
Why is the coherence pulse without a spike important?
It demonstrates that the diagnostic can register a large coherence value during plain monotonic relaxation between wells, with no spiking excursion at all. A coherence reading therefore cannot be interpreted as evidence of spiking dynamics without independent population information.
References
- D Maroulakos, A Wal, S K Mishra, Z Toklikishvili, L Chotorlishvili. Supersymmetric Neuromorphic FitzHugh-Nagumo model. arXiv:2611.03312. 2026. https://arxiv.org/abs/2611.03312. Accessed 2026-11-06.