An exact firing-rate curve for a minimal spiking neuron, with the excitability type decided by the sigmoid's edges
The frequency-current curve, steady firing rate versus constant input, is the oldest characterization in electrophysiology and the first thing an engineer wants for any spiking element. For a recently proposed minimal neuron built from a sigmoid in feedback with a high-pass filter, that curve previously required stiff numerical simulation. This paper derives it in closed form and proves that whether the neuron is Type I or Type II depends only on the local shape of the sigmoid at the edges of its high-gain zone.
Source: Characterization of the Frequency-Current Curve of First-Order High-Pass Neurons, arXiv:2611.09508, November 2026. Primary source. Read: full HTML text including theorems, proofs, and the compute-time table.
What the work claims
This is a mathematical analysis paper, theorems and proofs with numerical validation, not an experimental study. Its subject is the High-Pass (HP) neuron model, a spiking primitive introduced earlier in 2026: a static sigmoid activation placed in positive feedback with a high-pass filter. The neuron spikes tonically whenever its constant input lies inside the sigmoid's high-gain zone, where the sigmoid's slope exceeds 1, and rests otherwise. The paper's main theorem characterizes the tonic spiking limit cycle completely: outside the high-gain interval (a, b) there is a unique stable resting point, and inside it a unique stable limit cycle whose period can be written as an explicit integral expression1.
The practical form of the result is a computation recipe: the period, and hence the full frequency-current curve, can be evaluated by solving two equations that do not depend on the input and evaluating two input-dependent integrals, with no simulation of the differential equations at all. In the authors' Julia implementation, computing a 1000-point curve this way takes about 89 milliseconds for a Type I example and 75 milliseconds for a Type II example, versus about 156 and 191 milliseconds respectively to simulate only 10 points, making each computed point more than two orders of magnitude cheaper1. Two structural results fall out of the formula. First, the high-pass filter's time constant multiplies the entire period, so it rescales the frequency axis without changing the curve's shape, which is determined by the sigmoid alone. Second, the frequency-current curve is always smooth, strictly positive on the spiking interval, unimodal with at most one peak, and bounded, so the model cannot spike arbitrarily fast1.
The classification result is the elegant one. A spiking element is Type I if tonic firing starts at zero frequency as input crosses threshold, and Type II if it starts at a nonzero frequency, the classification that descends from Hodgkin's studies of repetitive firing in crab axons2. For the HP model, the type at each edge of the high-gain zone depends only on the sigmoid's local behavior just outside that edge: a sharp corner where the derivative approaches a value strictly below 1 gives Type I, while a twice-differentiable edge with bounded second derivative and derivative approaching 1 gives Type II1.
How it works
The analysis exploits a separation that the HP model is built to have: all the dynamics live in a single linear high-pass filter, and all the nonlinearity lives in a static sigmoid. In a suitable change of variables, a tonic spike is a trajectory that travels along two outer branches of the sigmoid's inverse S-curve, and the period splits into the time spent on the left branch and the time spent on the right branch. Each branch time is an integral of the branch's velocity, which is simple enough to write down exactly. The spike's turning points, where the trajectory jumps between branches, are fixed by two algebraic equations; once they are known, the period follows by quadrature1.
Two consequences matter for anyone using this as a component. The separation of timescale and shape means the filter time constant is a pure speed knob: halving it doubles every frequency on the curve without altering its form. And because the non-integral part of the period formula is always strictly positive, the firing rate has a hard ceiling set by the sigmoid's outer slopes, which bounds the model's bandwidth before any hardware is built. The Type I and Type II conditions then follow from how the integrals diverge or converge as the input approaches an edge: near a Type I edge the period grows logarithmically, so the frequency collapses to zero, while a smooth edge yields a finite limiting period and a nonzero onset frequency1. Validation is done against simulations of a regularized version of the model with a small regularization time constant of 10-8 times the filter time constant, and the analytical curves match, with the caveat that near Type I edges the simulations themselves become unstable precisely because the true period is diverging1.
Where a skeptic should push
The most load-bearing assumption is that a two-dimensional caricature deserves the authority the frequency-current curve carries in real electrophysiology. The HP model is deliberately minimal, essentially a practical rewriting of a FitzHugh-Nagumo-type excitable system, with one linear filter and no adaptation currents, no slow variables, no noise. Biological neurons near threshold have multiple interacting conductances, and their measured frequency-current curves bend, saturate, and change type with neuromodulation. Nothing here is validated against a recorded cell; the ground truth is numerical simulation of the same equations.
Second, the Type I and Type II conditions are sufficient conditions, not a complete classification: sigmoids that are neither cleanly sharp-cornered nor cleanly smooth at the edge fall outside the proposition. And the dramatic speedup, while real for this model, buys less than it appears: modern stiff solvers, adjoint methods, or event-based neuromorphic emulation were not the comparison target, only direct time-stepped simulation. Demonstrated: exact curves for broad sigmoid classes, the timescale-shape separation, bounded unimodal frequency, edge-based type classification. Asserted but not demonstrated: that this primitive, or its closed-form curve, maps onto the response of any physical neuron, silicon or biological.
A transfer-function spec for driving organoid tissue
Anyone building an interface to living neural tissue faces the same question this paper answers for a model: given an input current delivered through an electrode, what firing rate comes out, and how does the answer depend on the element's parameters? The paper's decomposition is the right mental toolkit for that problem. It shows that an excitable element's input-output curve factorizes into a shape, set by the nonlinearity, and a timescale, set by a single filter; that the shape is smooth, single-peaked, and bounded; and that the excitability type, which in networks governs synchronization at onset, is a local property of the nonlinearity's edges rather than a global one. For organoid intelligence, where closed-loop stimulation is the proposed training channel, this matters twice over: the stimulation protocol must respect that a living network's transfer saturates and that near a Type I-like onset the gain diverges, making tiny input errors produce large rate errors exactly where precision stimulation wants to operate.
The opportunity is a calibration discipline. A minimal closed-form curve means you can fit a culture's response to constant current steps with a small number of interpretable parameters, nonlinearity shape plus effective timescale, and get an honest input-output model that flags when the tissue leaves the fitted regime, something black-box neural surrogates do not do gracefully. The threat is subtler: if the minimal model's lesson is that excitability type is set by microscopic nonlinearity shape, then type is a hidden, drifting property of biological tissue. Ion channel expression, maturation state, and pharmacology can move a culture between smooth-onset and finite-onset behavior without any change in gross morphology, which would silently change what a fixed stimulation protocol does. A digital twin of the interface built with the wrong excitability type will systematically mispredict closed-loop responses, and the paper's logarithmic-period divergence near Type I onsets shows the misprediction is worst exactly at threshold. The paper also cuts at hardware claims: it shows a single static nonlinearity plus one linear filter is all a spiking element needs for a well-behaved, exactly characterizable transfer curve, which is a low bar for any competitor substrate claiming the same territory.
The bottom line
Established, within the model class: an exact, cheaply computable frequency-current curve for first-order high-pass neurons; timescale-shape factorization; bounded unimodal firing rates; and edge-local conditions for Type I versus Type II onset. Open: any correspondence to biological or silicon neurons with more than one dynamical degree of freedom, and behavior under time-varying input, which is the regime closed-loop organoid stimulation actually uses. What would confirm relevance: recovering the paper's curve shape and type classification from fits to constant-current step responses of real neurons or neuron analog circuits. What would break it: physical neurons whose measured curves are neither smooth-unimodal nor edge-classifiable, which would mean the minimal primitive omits exactly the dynamics that matter.
Frequently asked questions
What is the high-pass neuron model?
A minimal spiking element: a static sigmoid activation function placed in positive feedback with a high-pass filter. When the constant input sits where the sigmoid's slope exceeds 1, the feedback amplifies and the element spikes tonically; otherwise it rests at a stable fixed point.
What exactly did the paper compute in closed form?
The period of the tonic spiking limit cycle, and therefore the entire frequency-current curve, for any sigmoid in a broad defined class. The recipe is two input-independent equations for the spike's turning points plus two input-dependent integrals, with no simulation of the differential equations.
How much cheaper is the formula than simulation?
In the authors' Julia benchmarks, computing a 1000-point curve with the formula took roughly 75 to 89 milliseconds, versus roughly 156 to 191 milliseconds for direct simulation of just 10 points, so each computed point is over two orders of magnitude cheaper.
What decides Type I versus Type II firing?
Only the sigmoid's local shape just outside each edge of the high-gain zone. A sharp corner with limiting derivative strictly below 1 gives zero-frequency onset, Type I. A smooth, twice-differentiable edge with derivative approaching 1 gives finite-frequency onset, Type II.
Why is the firing rate bounded?
The period formula contains a strictly positive non-integral term that does not vanish as input varies, so the period never reaches zero and the frequency never diverges. The ceiling is set by the sigmoid's shape, not by any added constraint.
Does this apply to real neurons?
Not demonstrated. The model has a single linear filter and one static nonlinearity, no adaptation or noise, and is validated only against simulation. Its value for biological work is as a compact, interpretable template for fitting and reasoning about measured input-output curves.
References
- C. Fernandez Lorden. Characterization of the Frequency-Current Curve of First-Order High-Pass Neurons. arXiv:2611.09508. 2026. https://arxiv.org/abs/2611.09508. Accessed 2026-11-10.
- A. L. Hodgkin. The local electric changes associated with repetitive action in a non-medullated axon. Journal of Physiology 107(2), 1948. https://doi.org/10.1113/jphysiol.1948.sp004260. Accessed 2026-11-10.